drivers/gpu/drm/i915/intel_pm.c

   1 /*
   2  * Copyright © 2012 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  *
  23  * Authors:
  24  *    Eugeni Dodonov <eugeni.dodonov@intel.com>
  25  *
  26  */
  27
  28 #include <linux/module.h>
  29 #include <linux/pm_runtime.h>
  30
  31 #include <drm/drm_atomic_helper.h>
  32 #include <drm/drm_fourcc.h>
  33 #include <drm/drm_plane_helper.h>
  34
  35 #include "display/intel_atomic.h"
  36 #include "display/intel_atomic_plane.h"
  37 #include "display/intel_bw.h"
  38 #include "display/intel_display_types.h"
  39 #include "display/intel_fbc.h"
  40 #include "display/intel_sprite.h"
  41
  42 #include "gt/intel_llc.h"
  43
  44 #include "i915_drv.h"
  45 #include "i915_fixed.h"
  46 #include "i915_irq.h"
  47 #include "i915_trace.h"
  48 #include "intel_pm.h"
  49 #include "intel_sideband.h"
  50 #include "../../../platform/x86/intel_ips.h"
  51
  52 /* Stores plane specific WM parameters */
  53 struct skl_wm_params {
  54         bool x_tiled, y_tiled;
  55         bool rc_surface;
  56         bool is_planar;
  57         u32 width;
  58         u8 cpp;
  59         u32 plane_pixel_rate;
  60         u32 y_min_scanlines;
  61         u32 plane_bytes_per_line;
  62         uint_fixed_16_16_t plane_blocks_per_line;
  63         uint_fixed_16_16_t y_tile_minimum;
  64         u32 linetime_us;
  65         u32 dbuf_block_size;
  66 };
  67
  68 /* used in computing the new watermarks state */
  69 struct intel_wm_config {
  70         unsigned int num_pipes_active;
  71         bool sprites_enabled;
  72         bool sprites_scaled;
  73 };
  74
  75 static void gen9_init_clock_gating(struct drm_i915_private *dev_priv)
  76 {
  77         if (HAS_LLC(dev_priv)) {
  78                 /*
  79                  * WaCompressedResourceDisplayNewHashMode:skl,kbl
  80                  * Display WA #0390: skl,kbl
  81                  *
  82                  * Must match Sampler, Pixel Back End, and Media. See
  83                  * WaCompressedResourceSamplerPbeMediaNewHashMode.
  84                  */
  85                 intel_uncore_write(&dev_priv->uncore, CHICKEN_PAR1_1,
  86                            intel_uncore_read(&dev_priv->uncore, CHICKEN_PAR1_1) |
  87                            SKL_DE_COMPRESSED_HASH_MODE);
  88         }
  89
  90         /* See Bspec note for PSR2_CTL bit 31, Wa#828:skl,bxt,kbl,cfl */
  91         intel_uncore_write(&dev_priv->uncore, CHICKEN_PAR1_1,
  92                    intel_uncore_read(&dev_priv->uncore, CHICKEN_PAR1_1) | SKL_EDP_PSR_FIX_RDWRAP);
  93
  94         /* WaEnableChickenDCPR:skl,bxt,kbl,glk,cfl */
  95         intel_uncore_write(&dev_priv->uncore, GEN8_CHICKEN_DCPR_1,
  96                    intel_uncore_read(&dev_priv->uncore, GEN8_CHICKEN_DCPR_1) | MASK_WAKEMEM);
  97
  98         /*
  99          * WaFbcWakeMemOn:skl,bxt,kbl,glk,cfl
 100          * Display WA #0859: skl,bxt,kbl,glk,cfl
 101          */
 102         intel_uncore_write(&dev_priv->uncore, DISP_ARB_CTL, intel_uncore_read(&dev_priv->uncore, DISP_ARB_CTL) |
 103                    DISP_FBC_MEMORY_WAKE);
 104 }
 105
 106 static void bxt_init_clock_gating(struct drm_i915_private *dev_priv)
 107 {
 108         gen9_init_clock_gating(dev_priv);
 109
 110         /* WaDisableSDEUnitClockGating:bxt */
 111         intel_uncore_write(&dev_priv->uncore, GEN8_UCGCTL6, intel_uncore_read(&dev_priv->uncore, GEN8_UCGCTL6) |
 112                    GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
 113
 114         /*
 115          * FIXME:
 116          * GEN8_HDCUNIT_CLOCK_GATE_DISABLE_HDCREQ applies on 3x6 GT SKUs only.
 117          */
 118         intel_uncore_write(&dev_priv->uncore, GEN8_UCGCTL6, intel_uncore_read(&dev_priv->uncore, GEN8_UCGCTL6) |
 119                    GEN8_HDCUNIT_CLOCK_GATE_DISABLE_HDCREQ);
 120
 121         /*
 122          * Wa: Backlight PWM may stop in the asserted state, causing backlight
 123          * to stay fully on.
 124          */
 125         intel_uncore_write(&dev_priv->uncore, GEN9_CLKGATE_DIS_0, intel_uncore_read(&dev_priv->uncore, GEN9_CLKGATE_DIS_0) |
 126                    PWM1_GATING_DIS | PWM2_GATING_DIS);
 127
 128         /*
 129          * Lower the display internal timeout.
 130          * This is needed to avoid any hard hangs when DSI port PLL
 131          * is off and a MMIO access is attempted by any privilege
 132          * application, using batch buffers or any other means.
 133          */
 134         intel_uncore_write(&dev_priv->uncore, RM_TIMEOUT, MMIO_TIMEOUT_US(950));
 135
 136         /*
 137          * WaFbcTurnOffFbcWatermark:bxt
 138          * Display WA #0562: bxt
 139          */
 140         intel_uncore_write(&dev_priv->uncore, DISP_ARB_CTL, intel_uncore_read(&dev_priv->uncore, DISP_ARB_CTL) |
 141                    DISP_FBC_WM_DIS);
 142
 143         /*
 144          * WaFbcHighMemBwCorruptionAvoidance:bxt
 145          * Display WA #0883: bxt
 146          */
 147         intel_uncore_write(&dev_priv->uncore, ILK_DPFC_CHICKEN, intel_uncore_read(&dev_priv->uncore, ILK_DPFC_CHICKEN) |
 148                    ILK_DPFC_DISABLE_DUMMY0);
 149 }
 150
 151 static void glk_init_clock_gating(struct drm_i915_private *dev_priv)
 152 {
 153         gen9_init_clock_gating(dev_priv);
 154
 155         /*
 156          * WaDisablePWMClockGating:glk
 157          * Backlight PWM may stop in the asserted state, causing backlight
 158          * to stay fully on.
 159          */
 160         intel_uncore_write(&dev_priv->uncore, GEN9_CLKGATE_DIS_0, intel_uncore_read(&dev_priv->uncore, GEN9_CLKGATE_DIS_0) |
 161                    PWM1_GATING_DIS | PWM2_GATING_DIS);
 162 }
 163
 164 static void pnv_get_mem_freq(struct drm_i915_private *dev_priv)
 165 {
 166         u32 tmp;
 167
 168         tmp = intel_uncore_read(&dev_priv->uncore, CLKCFG);
 169
 170         switch (tmp & CLKCFG_FSB_MASK) {
 171         case CLKCFG_FSB_533:
 172                 dev_priv->fsb_freq = 533; /* 133*4 */
 173                 break;
 174         case CLKCFG_FSB_800:
 175                 dev_priv->fsb_freq = 800; /* 200*4 */
 176                 break;
 177         case CLKCFG_FSB_667:
 178                 dev_priv->fsb_freq =  667; /* 167*4 */
 179                 break;
 180         case CLKCFG_FSB_400:
 181                 dev_priv->fsb_freq = 400; /* 100*4 */
 182                 break;
 183         }
 184
 185         switch (tmp & CLKCFG_MEM_MASK) {
 186         case CLKCFG_MEM_533:
 187                 dev_priv->mem_freq = 533;
 188                 break;
 189         case CLKCFG_MEM_667:
 190                 dev_priv->mem_freq = 667;
 191                 break;
 192         case CLKCFG_MEM_800:
 193                 dev_priv->mem_freq = 800;
 194                 break;
 195         }
 196
 197         /* detect pineview DDR3 setting */
 198         tmp = intel_uncore_read(&dev_priv->uncore, CSHRDDR3CTL);
 199         dev_priv->is_ddr3 = (tmp & CSHRDDR3CTL_DDR3) ? 1 : 0;
 200 }
 201
 202 static void ilk_get_mem_freq(struct drm_i915_private *dev_priv)
 203 {
 204         u16 ddrpll, csipll;
 205
 206         ddrpll = intel_uncore_read16(&dev_priv->uncore, DDRMPLL1);
 207         csipll = intel_uncore_read16(&dev_priv->uncore, CSIPLL0);
 208
 209         switch (ddrpll & 0xff) {
 210         case 0xc:
 211                 dev_priv->mem_freq = 800;
 212                 break;
 213         case 0x10:
 214                 dev_priv->mem_freq = 1066;
 215                 break;
 216         case 0x14:
 217                 dev_priv->mem_freq = 1333;
 218                 break;
 219         case 0x18:
 220                 dev_priv->mem_freq = 1600;
 221                 break;
 222         default:
 223                 drm_dbg(&dev_priv->drm, "unknown memory frequency 0x%02x\n",
 224                         ddrpll & 0xff);
 225                 dev_priv->mem_freq = 0;
 226                 break;
 227         }
 228
 229         switch (csipll & 0x3ff) {
 230         case 0x00c:
 231                 dev_priv->fsb_freq = 3200;
 232                 break;
 233         case 0x00e:
 234                 dev_priv->fsb_freq = 3733;
 235                 break;
 236         case 0x010:
 237                 dev_priv->fsb_freq = 4266;
 238                 break;
 239         case 0x012:
 240                 dev_priv->fsb_freq = 4800;
 241                 break;
 242         case 0x014:
 243                 dev_priv->fsb_freq = 5333;
 244                 break;
 245         case 0x016:
 246                 dev_priv->fsb_freq = 5866;
 247                 break;
 248         case 0x018:
 249                 dev_priv->fsb_freq = 6400;
 250                 break;
 251         default:
 252                 drm_dbg(&dev_priv->drm, "unknown fsb frequency 0x%04x\n",
 253                         csipll & 0x3ff);
 254                 dev_priv->fsb_freq = 0;
 255                 break;
 256         }
 257 }
 258
 259 static const struct cxsr_latency cxsr_latency_table[] = {
 260         {1, 0, 800, 400, 3382, 33382, 3983, 33983},    /* DDR2-400 SC */
 261         {1, 0, 800, 667, 3354, 33354, 3807, 33807},    /* DDR2-667 SC */
 262         {1, 0, 800, 800, 3347, 33347, 3763, 33763},    /* DDR2-800 SC */
 263         {1, 1, 800, 667, 6420, 36420, 6873, 36873},    /* DDR3-667 SC */
 264         {1, 1, 800, 800, 5902, 35902, 6318, 36318},    /* DDR3-800 SC */
 265
 266         {1, 0, 667, 400, 3400, 33400, 4021, 34021},    /* DDR2-400 SC */
 267         {1, 0, 667, 667, 3372, 33372, 3845, 33845},    /* DDR2-667 SC */
 268         {1, 0, 667, 800, 3386, 33386, 3822, 33822},    /* DDR2-800 SC */
 269         {1, 1, 667, 667, 6438, 36438, 6911, 36911},    /* DDR3-667 SC */
 270         {1, 1, 667, 800, 5941, 35941, 6377, 36377},    /* DDR3-800 SC */
 271
 272         {1, 0, 400, 400, 3472, 33472, 4173, 34173},    /* DDR2-400 SC */
 273         {1, 0, 400, 667, 3443, 33443, 3996, 33996},    /* DDR2-667 SC */
 274         {1, 0, 400, 800, 3430, 33430, 3946, 33946},    /* DDR2-800 SC */
 275         {1, 1, 400, 667, 6509, 36509, 7062, 37062},    /* DDR3-667 SC */
 276         {1, 1, 400, 800, 5985, 35985, 6501, 36501},    /* DDR3-800 SC */
 277
 278         {0, 0, 800, 400, 3438, 33438, 4065, 34065},    /* DDR2-400 SC */
 279         {0, 0, 800, 667, 3410, 33410, 3889, 33889},    /* DDR2-667 SC */
 280         {0, 0, 800, 800, 3403, 33403, 3845, 33845},    /* DDR2-800 SC */
 281         {0, 1, 800, 667, 6476, 36476, 6955, 36955},    /* DDR3-667 SC */
 282         {0, 1, 800, 800, 5958, 35958, 6400, 36400},    /* DDR3-800 SC */
 283
 284         {0, 0, 667, 400, 3456, 33456, 4103, 34106},    /* DDR2-400 SC */
 285         {0, 0, 667, 667, 3428, 33428, 3927, 33927},    /* DDR2-667 SC */
 286         {0, 0, 667, 800, 3443, 33443, 3905, 33905},    /* DDR2-800 SC */
 287         {0, 1, 667, 667, 6494, 36494, 6993, 36993},    /* DDR3-667 SC */
 288         {0, 1, 667, 800, 5998, 35998, 6460, 36460},    /* DDR3-800 SC */
 289
 290         {0, 0, 400, 400, 3528, 33528, 4255, 34255},    /* DDR2-400 SC */
 291         {0, 0, 400, 667, 3500, 33500, 4079, 34079},    /* DDR2-667 SC */
 292         {0, 0, 400, 800, 3487, 33487, 4029, 34029},    /* DDR2-800 SC */
 293         {0, 1, 400, 667, 6566, 36566, 7145, 37145},    /* DDR3-667 SC */
 294         {0, 1, 400, 800, 6042, 36042, 6584, 36584},    /* DDR3-800 SC */
 295 };
 296
 297 static const struct cxsr_latency *intel_get_cxsr_latency(bool is_desktop,
 298                                                          bool is_ddr3,
 299                                                          int fsb,
 300                                                          int mem)
 301 {
 302         const struct cxsr_latency *latency;
 303         int i;
 304
 305         if (fsb == 0 || mem == 0)
 306                 return NULL;
 307
 308         for (i = 0; i < ARRAY_SIZE(cxsr_latency_table); i++) {
 309                 latency = &cxsr_latency_table[i];
 310                 if (is_desktop == latency->is_desktop &&
 311                     is_ddr3 == latency->is_ddr3 &&
 312                     fsb == latency->fsb_freq && mem == latency->mem_freq)
 313                         return latency;
 314         }
 315
 316         DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
 317
 318         return NULL;
 319 }
 320
 321 static void chv_set_memory_dvfs(struct drm_i915_private *dev_priv, bool enable)
 322 {
 323         u32 val;
 324
 325         vlv_punit_get(dev_priv);
 326
 327         val = vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2);
 328         if (enable)
 329                 val &= ~FORCE_DDR_HIGH_FREQ;
 330         else
 331                 val |= FORCE_DDR_HIGH_FREQ;
 332         val &= ~FORCE_DDR_LOW_FREQ;
 333         val |= FORCE_DDR_FREQ_REQ_ACK;
 334         vlv_punit_write(dev_priv, PUNIT_REG_DDR_SETUP2, val);
 335
 336         if (wait_for((vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2) &
 337                       FORCE_DDR_FREQ_REQ_ACK) == 0, 3))
 338                 drm_err(&dev_priv->drm,
 339                         "timed out waiting for Punit DDR DVFS request\n");
 340
 341         vlv_punit_put(dev_priv);
 342 }
 343
 344 static void chv_set_memory_pm5(struct drm_i915_private *dev_priv, bool enable)
 345 {
 346         u32 val;
 347
 348         vlv_punit_get(dev_priv);
 349
 350         val = vlv_punit_read(dev_priv, PUNIT_REG_DSPSSPM);
 351         if (enable)
 352                 val |= DSP_MAXFIFO_PM5_ENABLE;
 353         else
 354                 val &= ~DSP_MAXFIFO_PM5_ENABLE;
 355         vlv_punit_write(dev_priv, PUNIT_REG_DSPSSPM, val);
 356
 357         vlv_punit_put(dev_priv);
 358 }
 359
 360 #define FW_WM(value, plane) \
 361         (((value) << DSPFW_ ## plane ## _SHIFT) & DSPFW_ ## plane ## _MASK)
 362
 363 static bool _intel_set_memory_cxsr(struct drm_i915_private *dev_priv, bool enable)
 364 {
 365         bool was_enabled;
 366         u32 val;
 367
 368         if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv)) {
 369                 was_enabled = intel_uncore_read(&dev_priv->uncore, FW_BLC_SELF_VLV) & FW_CSPWRDWNEN;
 370                 intel_uncore_write(&dev_priv->uncore, FW_BLC_SELF_VLV, enable ? FW_CSPWRDWNEN : 0);
 371                 intel_uncore_posting_read(&dev_priv->uncore, FW_BLC_SELF_VLV);
 372         } else if (IS_G4X(dev_priv) || IS_I965GM(dev_priv)) {
 373                 was_enabled = intel_uncore_read(&dev_priv->uncore, FW_BLC_SELF) & FW_BLC_SELF_EN;
 374                 intel_uncore_write(&dev_priv->uncore, FW_BLC_SELF, enable ? FW_BLC_SELF_EN : 0);
 375                 intel_uncore_posting_read(&dev_priv->uncore, FW_BLC_SELF);
 376         } else if (IS_PINEVIEW(dev_priv)) {
 377                 val = intel_uncore_read(&dev_priv->uncore, DSPFW3);
 378                 was_enabled = val & PINEVIEW_SELF_REFRESH_EN;
 379                 if (enable)
 380                         val |= PINEVIEW_SELF_REFRESH_EN;
 381                 else
 382                         val &= ~PINEVIEW_SELF_REFRESH_EN;
 383                 intel_uncore_write(&dev_priv->uncore, DSPFW3, val);
 384                 intel_uncore_posting_read(&dev_priv->uncore, DSPFW3);
 385         } else if (IS_I945G(dev_priv) || IS_I945GM(dev_priv)) {
 386                 was_enabled = intel_uncore_read(&dev_priv->uncore, FW_BLC_SELF) & FW_BLC_SELF_EN;
 387                 val = enable ? _MASKED_BIT_ENABLE(FW_BLC_SELF_EN) :
 388                                _MASKED_BIT_DISABLE(FW_BLC_SELF_EN);
 389                 intel_uncore_write(&dev_priv->uncore, FW_BLC_SELF, val);
 390                 intel_uncore_posting_read(&dev_priv->uncore, FW_BLC_SELF);
 391         } else if (IS_I915GM(dev_priv)) {
 392                 /*
 393                  * FIXME can't find a bit like this for 915G, and
 394                  * and yet it does have the related watermark in
 395                  * FW_BLC_SELF. What's going on?
 396                  */
 397                 was_enabled = intel_uncore_read(&dev_priv->uncore, INSTPM) & INSTPM_SELF_EN;
 398                 val = enable ? _MASKED_BIT_ENABLE(INSTPM_SELF_EN) :
 399                                _MASKED_BIT_DISABLE(INSTPM_SELF_EN);
 400                 intel_uncore_write(&dev_priv->uncore, INSTPM, val);
 401                 intel_uncore_posting_read(&dev_priv->uncore, INSTPM);
 402         } else {
 403                 return false;
 404         }
 405
 406         trace_intel_memory_cxsr(dev_priv, was_enabled, enable);
 407
 408         drm_dbg_kms(&dev_priv->drm, "memory self-refresh is %s (was %s)\n",
 409                     enableddisabled(enable),
 410                     enableddisabled(was_enabled));
 411
 412         return was_enabled;
 413 }
 414
 415 /**
 416  * intel_set_memory_cxsr - Configure CxSR state
 417  * @dev_priv: i915 device
 418  * @enable: Allow vs. disallow CxSR
 419  *
 420  * Allow or disallow the system to enter a special CxSR
 421  * (C-state self refresh) state. What typically happens in CxSR mode
 422  * is that several display FIFOs may get combined into a single larger
 423  * FIFO for a particular plane (so called max FIFO mode) to allow the
 424  * system to defer memory fetches longer, and the memory will enter
 425  * self refresh.
 426  *
 427  * Note that enabling CxSR does not guarantee that the system enter
 428  * this special mode, nor does it guarantee that the system stays
 429  * in that mode once entered. So this just allows/disallows the system
 430  * to autonomously utilize the CxSR mode. Other factors such as core
 431  * C-states will affect when/if the system actually enters/exits the
 432  * CxSR mode.
 433  *
 434  * Note that on VLV/CHV this actually only controls the max FIFO mode,
 435  * and the system is free to enter/exit memory self refresh at any time
 436  * even when the use of CxSR has been disallowed.
 437  *
 438  * While the system is actually in the CxSR/max FIFO mode, some plane
 439  * control registers will not get latched on vblank. Thus in order to
 440  * guarantee the system will respond to changes in the plane registers
 441  * we must always disallow CxSR prior to making changes to those registers.
 442  * Unfortunately the system will re-evaluate the CxSR conditions at
 443  * frame start which happens after vblank start (which is when the plane
 444  * registers would get latched), so we can't proceed with the plane update
 445  * during the same frame where we disallowed CxSR.
 446  *
 447  * Certain platforms also have a deeper HPLL SR mode. Fortunately the
 448  * HPLL SR mode depends on CxSR itself, so we don't have to hand hold
 449  * the hardware w.r.t. HPLL SR when writing to plane registers.
 450  * Disallowing just CxSR is sufficient.
 451  */
 452 bool intel_set_memory_cxsr(struct drm_i915_private *dev_priv, bool enable)
 453 {
 454         bool ret;
 455
 456         mutex_lock(&dev_priv->wm.wm_mutex);
 457         ret = _intel_set_memory_cxsr(dev_priv, enable);
 458         if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv))
 459                 dev_priv->wm.vlv.cxsr = enable;
 460         else if (IS_G4X(dev_priv))
 461                 dev_priv->wm.g4x.cxsr = enable;
 462         mutex_unlock(&dev_priv->wm.wm_mutex);
 463
 464         return ret;
 465 }
 466
 467 /*
 468  * Latency for FIFO fetches is dependent on several factors:
 469  *   - memory configuration (speed, channels)
 470  *   - chipset
 471  *   - current MCH state
 472  * It can be fairly high in some situations, so here we assume a fairly
 473  * pessimal value.  It's a tradeoff between extra memory fetches (if we
 474  * set this value too high, the FIFO will fetch frequently to stay full)
 475  * and power consumption (set it too low to save power and we might see
 476  * FIFO underruns and display "flicker").
 477  *
 478  * A value of 5us seems to be a good balance; safe for very low end
 479  * platforms but not overly aggressive on lower latency configs.
 480  */
 481 static const int pessimal_latency_ns = 5000;
 482
 483 #define VLV_FIFO_START(dsparb, dsparb2, lo_shift, hi_shift) \
 484         ((((dsparb) >> (lo_shift)) & 0xff) | ((((dsparb2) >> (hi_shift)) & 0x1) << 8))
 485
 486 static void vlv_get_fifo_size(struct intel_crtc_state *crtc_state)
 487 {
 488         struct intel_crtc *crtc = to_intel_crtc(crtc_state->uapi.crtc);
 489         struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
 490         struct vlv_fifo_state *fifo_state = &crtc_state->wm.vlv.fifo_state;
 491         enum pipe pipe = crtc->pipe;
 492         int sprite0_start, sprite1_start;
 493         u32 dsparb, dsparb2, dsparb3;
 494
 495         switch (pipe) {
 496         case PIPE_A:
 497                 dsparb = intel_uncore_read(&dev_priv->uncore, DSPARB);
 498                 dsparb2 = intel_uncore_read(&dev_priv->uncore, DSPARB2);
 499                 sprite0_start = VLV_FIFO_START(dsparb, dsparb2, 0, 0);
 500                 sprite1_start = VLV_FIFO_START(dsparb, dsparb2, 8, 4);
 501                 break;
 502         case PIPE_B:
 503                 dsparb = intel_uncore_read(&dev_priv->uncore, DSPARB);
 504                 dsparb2 = intel_uncore_read(&dev_priv->uncore, DSPARB2);
 505                 sprite0_start = VLV_FIFO_START(dsparb, dsparb2, 16, 8);
 506                 sprite1_start = VLV_FIFO_START(dsparb, dsparb2, 24, 12);
 507                 break;
 508         case PIPE_C:
 509                 dsparb2 = intel_uncore_read(&dev_priv->uncore, DSPARB2);
 510                 dsparb3 = intel_uncore_read(&dev_priv->uncore, DSPARB3);
 511                 sprite0_start = VLV_FIFO_START(dsparb3, dsparb2, 0, 16);
 512                 sprite1_start = VLV_FIFO_START(dsparb3, dsparb2, 8, 20);
 513                 break;
 514         default:
 515                 MISSING_CASE(pipe);
 516                 return;
 517         }
 518
 519         fifo_state->plane[PLANE_PRIMARY] = sprite0_start;
 520         fifo_state->plane[PLANE_SPRITE0] = sprite1_start - sprite0_start;
 521         fifo_state->plane[PLANE_SPRITE1] = 511 - sprite1_start;
 522         fifo_state->plane[PLANE_CURSOR] = 63;
 523 }
 524
 525 static int i9xx_get_fifo_size(struct drm_i915_private *dev_priv,
 526                               enum i9xx_plane_id i9xx_plane)
 527 {
 528         u32 dsparb = intel_uncore_read(&dev_priv->uncore, DSPARB);
 529         int size;
 530
 531         size = dsparb & 0x7f;
 532         if (i9xx_plane == PLANE_B)
 533                 size = ((dsparb >> DSPARB_CSTART_SHIFT) & 0x7f) - size;
 534
 535         drm_dbg_kms(&dev_priv->drm, "FIFO size - (0x%08x) %c: %d\n",
 536                     dsparb, plane_name(i9xx_plane), size);
 537
 538         return size;
 539 }
 540
 541 static int i830_get_fifo_size(struct drm_i915_private *dev_priv,
 542                               enum i9xx_plane_id i9xx_plane)
 543 {
 544         u32 dsparb = intel_uncore_read(&dev_priv->uncore, DSPARB);
 545         int size;
 546
 547         size = dsparb & 0x1ff;
 548         if (i9xx_plane == PLANE_B)
 549                 size = ((dsparb >> DSPARB_BEND_SHIFT) & 0x1ff) - size;
 550         size >>= 1; /* Convert to cachelines */
 551
 552         drm_dbg_kms(&dev_priv->drm, "FIFO size - (0x%08x) %c: %d\n",
 553                     dsparb, plane_name(i9xx_plane), size);
 554
 555         return size;
 556 }
 557
 558 static int i845_get_fifo_size(struct drm_i915_private *dev_priv,
 559                               enum i9xx_plane_id i9xx_plane)
 560 {
 561         u32 dsparb = intel_uncore_read(&dev_priv->uncore, DSPARB);
 562         int size;
 563
 564         size = dsparb & 0x7f;
 565         size >>= 2; /* Convert to cachelines */
 566
 567         drm_dbg_kms(&dev_priv->drm, "FIFO size - (0x%08x) %c: %d\n",
 568                     dsparb, plane_name(i9xx_plane), size);
 569
 570         return size;
 571 }
 572
 573 /* Pineview has different values for various configs */
 574 static const struct intel_watermark_params pnv_display_wm = {
 575         .fifo_size = PINEVIEW_DISPLAY_FIFO,
 576         .max_wm = PINEVIEW_MAX_WM,
 577         .default_wm = PINEVIEW_DFT_WM,
 578         .guard_size = PINEVIEW_GUARD_WM,
 579         .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
 580 };
 581
 582 static const struct intel_watermark_params pnv_display_hplloff_wm = {
 583         .fifo_size = PINEVIEW_DISPLAY_FIFO,
 584         .max_wm = PINEVIEW_MAX_WM,
 585         .default_wm = PINEVIEW_DFT_HPLLOFF_WM,
 586         .guard_size = PINEVIEW_GUARD_WM,
 587         .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
 588 };
 589
 590 static const struct intel_watermark_params pnv_cursor_wm = {
 591         .fifo_size = PINEVIEW_CURSOR_FIFO,
 592         .max_wm = PINEVIEW_CURSOR_MAX_WM,
 593         .default_wm = PINEVIEW_CURSOR_DFT_WM,
 594         .guard_size = PINEVIEW_CURSOR_GUARD_WM,
 595         .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
 596 };
 597
 598 static const struct intel_watermark_params pnv_cursor_hplloff_wm = {
 599         .fifo_size = PINEVIEW_CURSOR_FIFO,
 600         .max_wm = PINEVIEW_CURSOR_MAX_WM,
 601         .default_wm = PINEVIEW_CURSOR_DFT_WM,
 602         .guard_size = PINEVIEW_CURSOR_GUARD_WM,
 603         .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
 604 };
 605
 606 static const struct intel_watermark_params i965_cursor_wm_info = {
 607         .fifo_size = I965_CURSOR_FIFO,
 608         .max_wm = I965_CURSOR_MAX_WM,
 609         .default_wm = I965_CURSOR_DFT_WM,
 610         .guard_size = 2,
 611         .cacheline_size = I915_FIFO_LINE_SIZE,
 612 };
 613
 614 static const struct intel_watermark_params i945_wm_info = {
 615         .fifo_size = I945_FIFO_SIZE,
 616         .max_wm = I915_MAX_WM,
 617         .default_wm = 1,
 618         .guard_size = 2,
 619         .cacheline_size = I915_FIFO_LINE_SIZE,
 620 };
 621
 622 static const struct intel_watermark_params i915_wm_info = {
 623         .fifo_size = I915_FIFO_SIZE,
 624         .max_wm = I915_MAX_WM,
 625         .default_wm = 1,
 626         .guard_size = 2,
 627         .cacheline_size = I915_FIFO_LINE_SIZE,
 628 };
 629
 630 static const struct intel_watermark_params i830_a_wm_info = {
 631         .fifo_size = I855GM_FIFO_SIZE,
 632         .max_wm = I915_MAX_WM,
 633         .default_wm = 1,
 634         .guard_size = 2,
 635         .cacheline_size = I830_FIFO_LINE_SIZE,
 636 };
 637
 638 static const struct intel_watermark_params i830_bc_wm_info = {
 639         .fifo_size = I855GM_FIFO_SIZE,
 640         .max_wm = I915_MAX_WM/2,
 641         .default_wm = 1,
 642         .guard_size = 2,
 643         .cacheline_size = I830_FIFO_LINE_SIZE,
 644 };
 645
 646 static const struct intel_watermark_params i845_wm_info = {
 647         .fifo_size = I830_FIFO_SIZE,
 648         .max_wm = I915_MAX_WM,
 649         .default_wm = 1,
 650         .guard_size = 2,
 651         .cacheline_size = I830_FIFO_LINE_SIZE,
 652 };
 653
 654 /**
 655  * intel_wm_method1 - Method 1 / "small buffer" watermark formula
 656  * @pixel_rate: Pipe pixel rate in kHz
 657  * @cpp: Plane bytes per pixel
 658  * @latency: Memory wakeup latency in 0.1us units
 659  *
 660  * Compute the watermark using the method 1 or "small buffer"
 661  * formula. The caller may additonally add extra cachelines
 662  * to account for TLB misses and clock crossings.
 663  *
 664  * This method is concerned with the short term drain rate
 665  * of the FIFO, ie. it does not account for blanking periods
 666  * which would effectively reduce the average drain rate across
 667  * a longer period. The name "small" refers to the fact the
 668  * FIFO is relatively small compared to the amount of data
 669  * fetched.
 670  *
 671  * The FIFO level vs. time graph might look something like:
 672  *
 673  *   |\   |\
 674  *   | \  | \
 675  * __---__---__ (- plane active, _ blanking)
 676  * -> time
 677  *
 678  * or perhaps like this:
 679  *
 680  *   |\|\  |\|\
 681  * __----__----__ (- plane active, _ blanking)
 682  * -> time
 683  *
 684  * Returns:
 685  * The watermark in bytes
 686  */
 687 static unsigned int intel_wm_method1(unsigned int pixel_rate,
 688                                      unsigned int cpp,
 689                                      unsigned int latency)
 690 {
 691         u64 ret;
 692
 693         ret = mul_u32_u32(pixel_rate, cpp * latency);
 694         ret = DIV_ROUND_UP_ULL(ret, 10000);
 695
 696         return ret;
 697 }
 698
 699 /**
 700  * intel_wm_method2 - Method 2 / "large buffer" watermark formula
 701  * @pixel_rate: Pipe pixel rate in kHz
 702  * @htotal: Pipe horizontal total
 703  * @width: Plane width in pixels
 704  * @cpp: Plane bytes per pixel
 705  * @latency: Memory wakeup latency in 0.1us units
 706  *
 707  * Compute the watermark using the method 2 or "large buffer"
 708  * formula. The caller may additonally add extra cachelines
 709  * to account for TLB misses and clock crossings.
 710  *
 711  * This method is concerned with the long term drain rate
 712  * of the FIFO, ie. it does account for blanking periods
 713  * which effectively reduce the average drain rate across
 714  * a longer period. The name "large" refers to the fact the
 715  * FIFO is relatively large compared to the amount of data
 716  * fetched.
 717  *
 718  * The FIFO level vs. time graph might look something like:
 719  *
 720  *    |\___       |\___
 721  *    |    \___   |    \___
 722  *    |        \  |        \
 723  * __ --__--__--__--__--__--__ (- plane active, _ blanking)
 724  * -> time
 725  *
 726  * Returns:
 727  * The watermark in bytes
 728  */
 729 static unsigned int intel_wm_method2(unsigned int pixel_rate,
 730                                      unsigned int htotal,
 731                                      unsigned int width,
 732                                      unsigned int cpp,
 733                                      unsigned int latency)
 734 {
 735         unsigned int ret;
 736
 737         /*
 738          * FIXME remove once all users are computing
 739          * watermarks in the correct place.
 740          */
 741         if (WARN_ON_ONCE(htotal == 0))
 742                 htotal = 1;
 743
 744         ret = (latency * pixel_rate) / (htotal * 10000);
 745         ret = (ret + 1) * width * cpp;
 746
 747         return ret;
 748 }
 749
 750 /**
 751  * intel_calculate_wm - calculate watermark level
 752  * @pixel_rate: pixel clock
 753  * @wm: chip FIFO params
 754  * @fifo_size: size of the FIFO buffer
 755  * @cpp: bytes per pixel
 756  * @latency_ns: memory latency for the platform
 757  *
 758  * Calculate the watermark level (the level at which the display plane will
 759  * start fetching from memory again).  Each chip has a different display
 760  * FIFO size and allocation, so the caller needs to figure that out and pass
 761  * in the correct intel_watermark_params structure.
 762  *
 763  * As the pixel clock runs, the FIFO will be drained at a rate that depends
 764  * on the pixel size.  When it reaches the watermark level, it'll start
 765  * fetching FIFO line sized based chunks from memory until the FIFO fills
 766  * past the watermark point.  If the FIFO drains completely, a FIFO underrun
 767  * will occur, and a display engine hang could result.
 768  */
 769 static unsigned int intel_calculate_wm(int pixel_rate,
 770                                        const struct intel_watermark_params *wm,
 771                                        int fifo_size, int cpp,
 772                                        unsigned int latency_ns)
 773 {
 774         int entries, wm_size;
 775
 776         /*
 777          * Note: we need to make sure we don't overflow for various clock &
 778          * latency values.
 779          * clocks go from a few thousand to several hundred thousand.
 780          * latency is usually a few thousand
 781          */
 782         entries = intel_wm_method1(pixel_rate, cpp,
 783                                    latency_ns / 100);
 784         entries = DIV_ROUND_UP(entries, wm->cacheline_size) +
 785                 wm->guard_size;
 786         DRM_DEBUG_KMS("FIFO entries required for mode: %d\n", entries);
 787
 788         wm_size = fifo_size - entries;
 789         DRM_DEBUG_KMS("FIFO watermark level: %d\n", wm_size);
 790
 791         /* Don't promote wm_size to unsigned... */
 792         if (wm_size > wm->max_wm)
 793                 wm_size = wm->max_wm;
 794         if (wm_size <= 0)
 795                 wm_size = wm->default_wm;
 796
 797         /*
 798          * Bspec seems to indicate that the value shouldn't be lower than
 799          * 'burst size + 1'. Certainly 830 is quite unhappy with low values.
 800          * Lets go for 8 which is the burst size since certain platforms
 801          * already use a hardcoded 8 (which is what the spec says should be
 802          * done).
 803          */
 804         if (wm_size <= 8)
 805                 wm_size = 8;
 806
 807         return wm_size;
 808 }
 809
 810 static bool is_disabling(int old, int new, int threshold)
 811 {
 812         return old >= threshold && new < threshold;
 813 }
 814
 815 static bool is_enabling(int old, int new, int threshold)
 816 {
 817         return old < threshold && new >= threshold;
 818 }
 819
 820 static int intel_wm_num_levels(struct drm_i915_private *dev_priv)
 821 {
 822         return dev_priv->wm.max_level + 1;
 823 }
 824
 825 static bool intel_wm_plane_visible(const struct intel_crtc_state *crtc_state,
 826                                    const struct intel_plane_state *plane_state)
 827 {
 828         struct intel_plane *plane = to_intel_plane(plane_state->uapi.plane);
 829
 830         /* FIXME check the 'enable' instead */
 831         if (!crtc_state->hw.active)
 832                 return false;
 833
 834         /*
 835          * Treat cursor with fb as always visible since cursor updates
 836          * can happen faster than the vrefresh rate, and the current
 837          * watermark code doesn't handle that correctly. Cursor updates
 838          * which set/clear the fb or change the cursor size are going
 839          * to get throttled by intel_legacy_cursor_update() to work
 840          * around this problem with the watermark code.
 841          */
 842         if (plane->id == PLANE_CURSOR)
 843                 return plane_state->hw.fb != NULL;
 844         else
 845                 return plane_state->uapi.visible;
 846 }
 847
 848 static bool intel_crtc_active(struct intel_crtc *crtc)
 849 {
 850         /* Be paranoid as we can arrive here with only partial
 851          * state retrieved from the hardware during setup.
 852          *
 853          * We can ditch the adjusted_mode.crtc_clock check as soon
 854          * as Haswell has gained clock readout/fastboot support.
 855          *
 856          * We can ditch the crtc->primary->state->fb check as soon as we can
 857          * properly reconstruct framebuffers.
 858          *
 859          * FIXME: The intel_crtc->active here should be switched to
 860          * crtc->state->active once we have proper CRTC states wired up
 861          * for atomic.
 862          */
 863         return crtc->active && crtc->base.primary->state->fb &&
 864                 crtc->config->hw.adjusted_mode.crtc_clock;
 865 }
 866
 867 static struct intel_crtc *single_enabled_crtc(struct drm_i915_private *dev_priv)
 868 {
 869         struct intel_crtc *crtc, *enabled = NULL;
 870
 871         for_each_intel_crtc(&dev_priv->drm, crtc) {
 872                 if (intel_crtc_active(crtc)) {
 873                         if (enabled)
 874                                 return NULL;
 875                         enabled = crtc;
 876                 }
 877         }
 878
 879         return enabled;
 880 }
 881
 882 static void pnv_update_wm(struct intel_crtc *unused_crtc)
 883 {
 884         struct drm_i915_private *dev_priv = to_i915(unused_crtc->base.dev);
 885         struct intel_crtc *crtc;
 886         const struct cxsr_latency *latency;
 887         u32 reg;
 888         unsigned int wm;
 889
 890         latency = intel_get_cxsr_latency(!IS_MOBILE(dev_priv),
 891                                          dev_priv->is_ddr3,
 892                                          dev_priv->fsb_freq,
 893                                          dev_priv->mem_freq);
 894         if (!latency) {
 895                 drm_dbg_kms(&dev_priv->drm,
 896                             "Unknown FSB/MEM found, disable CxSR\n");
 897                 intel_set_memory_cxsr(dev_priv, false);
 898                 return;
 899         }
 900
 901         crtc = single_enabled_crtc(dev_priv);
 902         if (crtc) {
 903                 const struct drm_display_mode *pipe_mode =
 904                         &crtc->config->hw.pipe_mode;
 905                 const struct drm_framebuffer *fb =
 906                         crtc->base.primary->state->fb;
 907                 int cpp = fb->format->cpp[0];
 908                 int clock = pipe_mode->crtc_clock;
 909
 910                 /* Display SR */
 911                 wm = intel_calculate_wm(clock, &pnv_display_wm,
 912                                         pnv_display_wm.fifo_size,
 913                                         cpp, latency->display_sr);
 914                 reg = intel_uncore_read(&dev_priv->uncore, DSPFW1);
 915                 reg &= ~DSPFW_SR_MASK;
 916                 reg |= FW_WM(wm, SR);
 917                 intel_uncore_write(&dev_priv->uncore, DSPFW1, reg);
 918                 drm_dbg_kms(&dev_priv->drm, "DSPFW1 register is %x\n", reg);
 919
 920                 /* cursor SR */
 921                 wm = intel_calculate_wm(clock, &pnv_cursor_wm,
 922                                         pnv_display_wm.fifo_size,
 923                                         4, latency->cursor_sr);
 924                 reg = intel_uncore_read(&dev_priv->uncore, DSPFW3);
 925                 reg &= ~DSPFW_CURSOR_SR_MASK;
 926                 reg |= FW_WM(wm, CURSOR_SR);
 927                 intel_uncore_write(&dev_priv->uncore, DSPFW3, reg);
 928
 929                 /* Display HPLL off SR */
 930                 wm = intel_calculate_wm(clock, &pnv_display_hplloff_wm,
 931                                         pnv_display_hplloff_wm.fifo_size,
 932                                         cpp, latency->display_hpll_disable);
 933                 reg = intel_uncore_read(&dev_priv->uncore, DSPFW3);
 934                 reg &= ~DSPFW_HPLL_SR_MASK;
 935                 reg |= FW_WM(wm, HPLL_SR);
 936                 intel_uncore_write(&dev_priv->uncore, DSPFW3, reg);
 937
 938                 /* cursor HPLL off SR */
 939                 wm = intel_calculate_wm(clock, &pnv_cursor_hplloff_wm,
 940                                         pnv_display_hplloff_wm.fifo_size,
 941                                         4, latency->cursor_hpll_disable);
 942                 reg = intel_uncore_read(&dev_priv->uncore, DSPFW3);
 943                 reg &= ~DSPFW_HPLL_CURSOR_MASK;
 944                 reg |= FW_WM(wm, HPLL_CURSOR);
 945                 intel_uncore_write(&dev_priv->uncore, DSPFW3, reg);
 946                 drm_dbg_kms(&dev_priv->drm, "DSPFW3 register is %x\n", reg);
 947
 948                 intel_set_memory_cxsr(dev_priv, true);
 949         } else {
 950                 intel_set_memory_cxsr(dev_priv, false);
 951         }
 952 }
 953
 954 /*
 955  * Documentation says:
 956  * "If the line size is small, the TLB fetches can get in the way of the
 957  *  data fetches, causing some lag in the pixel data return which is not
 958  *  accounted for in the above formulas. The following adjustment only
 959  *  needs to be applied if eight whole lines fit in the buffer at once.
 960  *  The WM is adjusted upwards by the difference between the FIFO size
 961  *  and the size of 8 whole lines. This adjustment is always performed
 962  *  in the actual pixel depth regardless of whether FBC is enabled or not."
 963  */
 964 static unsigned int g4x_tlb_miss_wa(int fifo_size, int width, int cpp)
 965 {
 966         int tlb_miss = fifo_size * 64 - width * cpp * 8;
 967
 968         return max(0, tlb_miss);
 969 }
 970
 971 static void g4x_write_wm_values(struct drm_i915_private *dev_priv,
 972                                 const struct g4x_wm_values *wm)
 973 {
 974         enum pipe pipe;
 975
 976         for_each_pipe(dev_priv, pipe)
 977                 trace_g4x_wm(intel_get_crtc_for_pipe(dev_priv, pipe), wm);
 978
 979         intel_uncore_write(&dev_priv->uncore, DSPFW1,
 980                    FW_WM(wm->sr.plane, SR) |
 981                    FW_WM(wm->pipe[PIPE_B].plane[PLANE_CURSOR], CURSORB) |
 982                    FW_WM(wm->pipe[PIPE_B].plane[PLANE_PRIMARY], PLANEB) |
 983                    FW_WM(wm->pipe[PIPE_A].plane[PLANE_PRIMARY], PLANEA));
 984         intel_uncore_write(&dev_priv->uncore, DSPFW2,
 985                    (wm->fbc_en ? DSPFW_FBC_SR_EN : 0) |
 986                    FW_WM(wm->sr.fbc, FBC_SR) |
 987                    FW_WM(wm->hpll.fbc, FBC_HPLL_SR) |
 988                    FW_WM(wm->pipe[PIPE_B].plane[PLANE_SPRITE0], SPRITEB) |
 989                    FW_WM(wm->pipe[PIPE_A].plane[PLANE_CURSOR], CURSORA) |
 990                    FW_WM(wm->pipe[PIPE_A].plane[PLANE_SPRITE0], SPRITEA));
 991         intel_uncore_write(&dev_priv->uncore, DSPFW3,
 992                    (wm->hpll_en ? DSPFW_HPLL_SR_EN : 0) |
 993                    FW_WM(wm->sr.cursor, CURSOR_SR) |
 994                    FW_WM(wm->hpll.cursor, HPLL_CURSOR) |
 995                    FW_WM(wm->hpll.plane, HPLL_SR));
 996
 997         intel_uncore_posting_read(&dev_priv->uncore, DSPFW1);
 998 }
 999
1000 #define FW_WM_VLV(value, plane) \
1001         (((value) << DSPFW_ ## plane ## _SHIFT) & DSPFW_ ## plane ## _MASK_VLV)
1002
1003 static void vlv_write_wm_values(struct drm_i915_private *dev_priv,
1004                                 const struct vlv_wm_values *wm)
1005 {
1006         enum pipe pipe;
1007
1008         for_each_pipe(dev_priv, pipe) {
1009                 trace_vlv_wm(intel_get_crtc_for_pipe(dev_priv, pipe), wm);
1010
1011                 intel_uncore_write(&dev_priv->uncore, VLV_DDL(pipe),
1012                            (wm->ddl[pipe].plane[PLANE_CURSOR] << DDL_CURSOR_SHIFT) |
1013                            (wm->ddl[pipe].plane[PLANE_SPRITE1] << DDL_SPRITE_SHIFT(1)) |
1014                            (wm->ddl[pipe].plane[PLANE_SPRITE0] << DDL_SPRITE_SHIFT(0)) |
1015                            (wm->ddl[pipe].plane[PLANE_PRIMARY] << DDL_PLANE_SHIFT));
1016         }
1017
1018         /*
1019          * Zero the (unused) WM1 watermarks, and also clear all the
1020          * high order bits so that there are no out of bounds values
1021          * present in the registers during the reprogramming.
1022          */
1023         intel_uncore_write(&dev_priv->uncore, DSPHOWM, 0);
1024         intel_uncore_write(&dev_priv->uncore, DSPHOWM1, 0);
1025         intel_uncore_write(&dev_priv->uncore, DSPFW4, 0);
1026         intel_uncore_write(&dev_priv->uncore, DSPFW5, 0);
1027         intel_uncore_write(&dev_priv->uncore, DSPFW6, 0);
1028
1029         intel_uncore_write(&dev_priv->uncore, DSPFW1,
1030                    FW_WM(wm->sr.plane, SR) |
1031                    FW_WM(wm->pipe[PIPE_B].plane[PLANE_CURSOR], CURSORB) |
1032                    FW_WM_VLV(wm->pipe[PIPE_B].plane[PLANE_PRIMARY], PLANEB) |
1033                    FW_WM_VLV(wm->pipe[PIPE_A].plane[PLANE_PRIMARY], PLANEA));
1034         intel_uncore_write(&dev_priv->uncore, DSPFW2,
1035                    FW_WM_VLV(wm->pipe[PIPE_A].plane[PLANE_SPRITE1], SPRITEB) |
1036                    FW_WM(wm->pipe[PIPE_A].plane[PLANE_CURSOR], CURSORA) |
1037                    FW_WM_VLV(wm->pipe[PIPE_A].plane[PLANE_SPRITE0], SPRITEA));
1038         intel_uncore_write(&dev_priv->uncore, DSPFW3,
1039                    FW_WM(wm->sr.cursor, CURSOR_SR));
1040
1041         if (IS_CHERRYVIEW(dev_priv)) {
1042                 intel_uncore_write(&dev_priv->uncore, DSPFW7_CHV,
1043                            FW_WM_VLV(wm->pipe[PIPE_B].plane[PLANE_SPRITE1], SPRITED) |
1044                            FW_WM_VLV(wm->pipe[PIPE_B].plane[PLANE_SPRITE0], SPRITEC));
1045                 intel_uncore_write(&dev_priv->uncore, DSPFW8_CHV,
1046                            FW_WM_VLV(wm->pipe[PIPE_C].plane[PLANE_SPRITE1], SPRITEF) |
1047                            FW_WM_VLV(wm->pipe[PIPE_C].plane[PLANE_SPRITE0], SPRITEE));
1048                 intel_uncore_write(&dev_priv->uncore, DSPFW9_CHV,
1049                            FW_WM_VLV(wm->pipe[PIPE_C].plane[PLANE_PRIMARY], PLANEC) |
1050                            FW_WM(wm->pipe[PIPE_C].plane[PLANE_CURSOR], CURSORC));
1051                 intel_uncore_write(&dev_priv->uncore, DSPHOWM,
1052                            FW_WM(wm->sr.plane >> 9, SR_HI) |
1053                            FW_WM(wm->pipe[PIPE_C].plane[PLANE_SPRITE1] >> 8, SPRITEF_HI) |
1054                            FW_WM(wm->pipe[PIPE_C].plane[PLANE_SPRITE0] >> 8, SPRITEE_HI) |
1055                            FW_WM(wm->pipe[PIPE_C].plane[PLANE_PRIMARY] >> 8, PLANEC_HI) |
1056                            FW_WM(wm->pipe[PIPE_B].plane[PLANE_SPRITE1] >> 8, SPRITED_HI) |
1057                            FW_WM(wm->pipe[PIPE_B].plane[PLANE_SPRITE0] >> 8, SPRITEC_HI) |
1058                            FW_WM(wm->pipe[PIPE_B].plane[PLANE_PRIMARY] >> 8, PLANEB_HI) |
1059                            FW_WM(wm->pipe[PIPE_A].plane[PLANE_SPRITE1] >> 8, SPRITEB_HI) |
1060                            FW_WM(wm->pipe[PIPE_A].plane[PLANE_SPRITE0] >> 8, SPRITEA_HI) |
1061                            FW_WM(wm->pipe[PIPE_A].plane[PLANE_PRIMARY] >> 8, PLANEA_HI));
1062         } else {
1063                 intel_uncore_write(&dev_priv->uncore, DSPFW7,
1064                            FW_WM_VLV(wm->pipe[PIPE_B].plane[PLANE_SPRITE1], SPRITED) |
1065                            FW_WM_VLV(wm->pipe[PIPE_B].plane[PLANE_SPRITE0], SPRITEC));
1066                 intel_uncore_write(&dev_priv->uncore, DSPHOWM,
1067                            FW_WM(wm->sr.plane >> 9, SR_HI) |
1068                            FW_WM(wm->pipe[PIPE_B].plane[PLANE_SPRITE1] >> 8, SPRITED_HI) |
1069                            FW_WM(wm->pipe[PIPE_B].plane[PLANE_SPRITE0] >> 8, SPRITEC_HI) |
1070                            FW_WM(wm->pipe[PIPE_B].plane[PLANE_PRIMARY] >> 8, PLANEB_HI) |
1071                            FW_WM(wm->pipe[PIPE_A].plane[PLANE_SPRITE1] >> 8, SPRITEB_HI) |
1072                            FW_WM(wm->pipe[PIPE_A].plane[PLANE_SPRITE0] >> 8, SPRITEA_HI) |
1073                            FW_WM(wm->pipe[PIPE_A].plane[PLANE_PRIMARY] >> 8, PLANEA_HI));
1074         }
1075
1076         intel_uncore_posting_read(&dev_priv->uncore, DSPFW1);
1077 }
1078
1079 #undef FW_WM_VLV
1080
1081 static void g4x_setup_wm_latency(struct drm_i915_private *dev_priv)
1082 {
1083         /* all latencies in usec */
1084         dev_priv->wm.pri_latency[G4X_WM_LEVEL_NORMAL] = 5;
1085         dev_priv->wm.pri_latency[G4X_WM_LEVEL_SR] = 12;
1086         dev_priv->wm.pri_latency[G4X_WM_LEVEL_HPLL] = 35;
1087
1088         dev_priv->wm.max_level = G4X_WM_LEVEL_HPLL;
1089 }
1090
1091 static int g4x_plane_fifo_size(enum plane_id plane_id, int level)
1092 {
1093         /*
1094          * DSPCNTR[13] supposedly controls whether the
1095          * primary plane can use the FIFO space otherwise
1096          * reserved for the sprite plane. It's not 100% clear
1097          * what the actual FIFO size is, but it looks like we
1098          * can happily set both primary and sprite watermarks
1099          * up to 127 cachelines. So that would seem to mean
1100          * that either DSPCNTR[13] doesn't do anything, or that
1101          * the total FIFO is >= 256 cachelines in size. Either
1102          * way, we don't seem to have to worry about this
1103          * repartitioning as the maximum watermark value the
1104          * register can hold for each plane is lower than the
1105          * minimum FIFO size.
1106          */
1107         switch (plane_id) {
1108         case PLANE_CURSOR:
1109                 return 63;
1110         case PLANE_PRIMARY:
1111                 return level == G4X_WM_LEVEL_NORMAL ? 127 : 511;
1112         case PLANE_SPRITE0:
1113                 return level == G4X_WM_LEVEL_NORMAL ? 127 : 0;
1114         default:
1115                 MISSING_CASE(plane_id);
1116                 return 0;
1117         }
1118 }
1119
1120 static int g4x_fbc_fifo_size(int level)
1121 {
1122         switch (level) {
1123         case G4X_WM_LEVEL_SR:
1124                 return 7;
1125         case G4X_WM_LEVEL_HPLL:
1126                 return 15;
1127         default:
1128                 MISSING_CASE(level);
1129                 return 0;
1130         }
1131 }
1132
1133 static u16 g4x_compute_wm(const struct intel_crtc_state *crtc_state,
1134                           const struct intel_plane_state *plane_state,
1135                           int level)
1136 {
1137         struct intel_plane *plane = to_intel_plane(plane_state->uapi.plane);
1138         struct drm_i915_private *dev_priv = to_i915(plane->base.dev);
1139         const struct drm_display_mode *pipe_mode =
1140                 &crtc_state->hw.pipe_mode;
1141         unsigned int latency = dev_priv->wm.pri_latency[level] * 10;
1142         unsigned int clock, htotal, cpp, width, wm;
1143
1144         if (latency == 0)
1145                 return USHRT_MAX;
1146
1147         if (!intel_wm_plane_visible(crtc_state, plane_state))
1148                 return 0;
1149
1150         cpp = plane_state->hw.fb->format->cpp[0];
1151
1152         /*
1153          * Not 100% sure which way ELK should go here as the
1154          * spec only says CL/CTG should assume 32bpp and BW
1155          * doesn't need to. But as these things followed the
1156          * mobile vs. desktop lines on gen3 as well, let's
1157          * assume ELK doesn't need this.
1158          *
1159          * The spec also fails to list such a restriction for
1160          * the HPLL watermark, which seems a little strange.
1161          * Let's use 32bpp for the HPLL watermark as well.
1162          */
1163         if (IS_GM45(dev_priv) && plane->id == PLANE_PRIMARY &&
1164             level != G4X_WM_LEVEL_NORMAL)
1165                 cpp = max(cpp, 4u);
1166
1167         clock = pipe_mode->crtc_clock;
1168         htotal = pipe_mode->crtc_htotal;
1169
1170         width = drm_rect_width(&plane_state->uapi.dst);
1171
1172         if (plane->id == PLANE_CURSOR) {
1173                 wm = intel_wm_method2(clock, htotal, width, cpp, latency);
1174         } else if (plane->id == PLANE_PRIMARY &&
1175                    level == G4X_WM_LEVEL_NORMAL) {
1176                 wm = intel_wm_method1(clock, cpp, latency);
1177         } else {
1178                 unsigned int small, large;
1179
1180                 small = intel_wm_method1(clock, cpp, latency);
1181                 large = intel_wm_method2(clock, htotal, width, cpp, latency);
1182
1183                 wm = min(small, large);
1184         }
1185
1186         wm += g4x_tlb_miss_wa(g4x_plane_fifo_size(plane->id, level),
1187                               width, cpp);
1188
1189         wm = DIV_ROUND_UP(wm, 64) + 2;
1190
1191         return min_t(unsigned int, wm, USHRT_MAX);
1192 }
1193
1194 static bool g4x_raw_plane_wm_set(struct intel_crtc_state *crtc_state,
1195                                  int level, enum plane_id plane_id, u16 value)
1196 {
1197         struct drm_i915_private *dev_priv = to_i915(crtc_state->uapi.crtc->dev);
1198         bool dirty = false;
1199
1200         for (; level < intel_wm_num_levels(dev_priv); level++) {
1201                 struct g4x_pipe_wm *raw = &crtc_state->wm.g4x.raw[level];
1202
1203                 dirty |= raw->plane[plane_id] != value;
1204                 raw->plane[plane_id] = value;
1205         }
1206
1207         return dirty;
1208 }
1209
1210 static bool g4x_raw_fbc_wm_set(struct intel_crtc_state *crtc_state,
1211                                int level, u16 value)
1212 {
1213         struct drm_i915_private *dev_priv = to_i915(crtc_state->uapi.crtc->dev);
1214         bool dirty = false;
1215
1216         /* NORMAL level doesn't have an FBC watermark */
1217         level = max(level, G4X_WM_LEVEL_SR);
1218
1219         for (; level < intel_wm_num_levels(dev_priv); level++) {
1220                 struct g4x_pipe_wm *raw = &crtc_state->wm.g4x.raw[level];
1221
1222                 dirty |= raw->fbc != value;
1223                 raw->fbc = value;
1224         }
1225
1226         return dirty;
1227 }
1228
1229 static u32 ilk_compute_fbc_wm(const struct intel_crtc_state *crtc_state,
1230                               const struct intel_plane_state *plane_state,
1231                               u32 pri_val);
1232
1233 static bool g4x_raw_plane_wm_compute(struct intel_crtc_state *crtc_state,
1234                                      const struct intel_plane_state *plane_state)
1235 {
1236         struct intel_plane *plane = to_intel_plane(plane_state->uapi.plane);
1237         struct drm_i915_private *dev_priv = to_i915(crtc_state->uapi.crtc->dev);
1238         int num_levels = intel_wm_num_levels(to_i915(plane->base.dev));
1239         enum plane_id plane_id = plane->id;
1240         bool dirty = false;
1241         int level;
1242
1243         if (!intel_wm_plane_visible(crtc_state, plane_state)) {
1244                 dirty |= g4x_raw_plane_wm_set(crtc_state, 0, plane_id, 0);
1245                 if (plane_id == PLANE_PRIMARY)
1246                         dirty |= g4x_raw_fbc_wm_set(crtc_state, 0, 0);
1247                 goto out;
1248         }
1249
1250         for (level = 0; level < num_levels; level++) {
1251                 struct g4x_pipe_wm *raw = &crtc_state->wm.g4x.raw[level];
1252                 int wm, max_wm;
1253
1254                 wm = g4x_compute_wm(crtc_state, plane_state, level);
1255                 max_wm = g4x_plane_fifo_size(plane_id, level);
1256
1257                 if (wm > max_wm)
1258                         break;
1259
1260                 dirty |= raw->plane[plane_id] != wm;
1261                 raw->plane[plane_id] = wm;
1262
1263                 if (plane_id != PLANE_PRIMARY ||
1264                     level == G4X_WM_LEVEL_NORMAL)
1265                         continue;
1266
1267                 wm = ilk_compute_fbc_wm(crtc_state, plane_state,
1268                                         raw->plane[plane_id]);
1269                 max_wm = g4x_fbc_fifo_size(level);
1270
1271                 /*
1272                  * FBC wm is not mandatory as we
1273                  * can always just disable its use.
1274                  */
1275                 if (wm > max_wm)
1276                         wm = USHRT_MAX;
1277
1278                 dirty |= raw->fbc != wm;
1279                 raw->fbc = wm;
1280         }
1281
1282         /* mark watermarks as invalid */
1283         dirty |= g4x_raw_plane_wm_set(crtc_state, level, plane_id, USHRT_MAX);
1284
1285         if (plane_id == PLANE_PRIMARY)
1286                 dirty |= g4x_raw_fbc_wm_set(crtc_state, level, USHRT_MAX);
1287
1288  out:
1289         if (dirty) {
1290                 drm_dbg_kms(&dev_priv->drm,
1291                             "%s watermarks: normal=%d, SR=%d, HPLL=%d\n",
1292                             plane->base.name,
1293                             crtc_state->wm.g4x.raw[G4X_WM_LEVEL_NORMAL].plane[plane_id],
1294                             crtc_state->wm.g4x.raw[G4X_WM_LEVEL_SR].plane[plane_id],
1295                             crtc_state->wm.g4x.raw[G4X_WM_LEVEL_HPLL].plane[plane_id]);
1296
1297                 if (plane_id == PLANE_PRIMARY)
1298                         drm_dbg_kms(&dev_priv->drm,
1299                                     "FBC watermarks: SR=%d, HPLL=%d\n",
1300                                     crtc_state->wm.g4x.raw[G4X_WM_LEVEL_SR].fbc,
1301                                     crtc_state->wm.g4x.raw[G4X_WM_LEVEL_HPLL].fbc);
1302         }
1303
1304         return dirty;
1305 }
1306
1307 static bool g4x_raw_plane_wm_is_valid(const struct intel_crtc_state *crtc_state,
1308                                       enum plane_id plane_id, int level)
1309 {
1310         const struct g4x_pipe_wm *raw = &crtc_state->wm.g4x.raw[level];
1311
1312         return raw->plane[plane_id] <= g4x_plane_fifo_size(plane_id, level);
1313 }
1314
1315 static bool g4x_raw_crtc_wm_is_valid(const struct intel_crtc_state *crtc_state,
1316                                      int level)
1317 {
1318         struct drm_i915_private *dev_priv = to_i915(crtc_state->uapi.crtc->dev);
1319
1320         if (level > dev_priv->wm.max_level)
1321                 return false;
1322
1323         return g4x_raw_plane_wm_is_valid(crtc_state, PLANE_PRIMARY, level) &&
1324                 g4x_raw_plane_wm_is_valid(crtc_state, PLANE_SPRITE0, level) &&
1325                 g4x_raw_plane_wm_is_valid(crtc_state, PLANE_CURSOR, level);
1326 }
1327
1328 /* mark all levels starting from 'level' as invalid */
1329 static void g4x_invalidate_wms(struct intel_crtc *crtc,
1330                                struct g4x_wm_state *wm_state, int level)
1331 {
1332         if (level <= G4X_WM_LEVEL_NORMAL) {
1333                 enum plane_id plane_id;
1334
1335                 for_each_plane_id_on_crtc(crtc, plane_id)
1336                         wm_state->wm.plane[plane_id] = USHRT_MAX;
1337         }
1338
1339         if (level <= G4X_WM_LEVEL_SR) {
1340                 wm_state->cxsr = false;
1341                 wm_state->sr.cursor = USHRT_MAX;
1342                 wm_state->sr.plane = USHRT_MAX;
1343                 wm_state->sr.fbc = USHRT_MAX;
1344         }
1345
1346         if (level <= G4X_WM_LEVEL_HPLL) {
1347                 wm_state->hpll_en = false;
1348                 wm_state->hpll.cursor = USHRT_MAX;
1349                 wm_state->hpll.plane = USHRT_MAX;
1350                 wm_state->hpll.fbc = USHRT_MAX;
1351         }
1352 }
1353
1354 static bool g4x_compute_fbc_en(const struct g4x_wm_state *wm_state,
1355                                int level)
1356 {
1357         if (level < G4X_WM_LEVEL_SR)
1358                 return false;
1359
1360         if (level >= G4X_WM_LEVEL_SR &&
1361             wm_state->sr.fbc > g4x_fbc_fifo_size(G4X_WM_LEVEL_SR))
1362                 return false;
1363
1364         if (level >= G4X_WM_LEVEL_HPLL &&
1365             wm_state->hpll.fbc > g4x_fbc_fifo_size(G4X_WM_LEVEL_HPLL))
1366                 return false;
1367
1368         return true;
1369 }
1370
1371 static int g4x_compute_pipe_wm(struct intel_crtc_state *crtc_state)
1372 {
1373         struct intel_crtc *crtc = to_intel_crtc(crtc_state->uapi.crtc);
1374         struct intel_atomic_state *state =
1375                 to_intel_atomic_state(crtc_state->uapi.state);
1376         struct g4x_wm_state *wm_state = &crtc_state->wm.g4x.optimal;
1377         int num_active_planes = hweight8(crtc_state->active_planes &
1378                                          ~BIT(PLANE_CURSOR));
1379         const struct g4x_pipe_wm *raw;
1380         const struct intel_plane_state *old_plane_state;
1381         const struct intel_plane_state *new_plane_state;
1382         struct intel_plane *plane;
1383         enum plane_id plane_id;
1384         int i, level;
1385         unsigned int dirty = 0;
1386
1387         for_each_oldnew_intel_plane_in_state(state, plane,
1388                                              old_plane_state,
1389                                              new_plane_state, i) {
1390                 if (new_plane_state->hw.crtc != &crtc->base &&
1391                     old_plane_state->hw.crtc != &crtc->base)
1392                         continue;
1393
1394                 if (g4x_raw_plane_wm_compute(crtc_state, new_plane_state))
1395                         dirty |= BIT(plane->id);
1396         }
1397
1398         if (!dirty)
1399                 return 0;
1400
1401         level = G4X_WM_LEVEL_NORMAL;
1402         if (!g4x_raw_crtc_wm_is_valid(crtc_state, level))
1403                 goto out;
1404
1405         raw = &crtc_state->wm.g4x.raw[level];
1406         for_each_plane_id_on_crtc(crtc, plane_id)
1407                 wm_state->wm.plane[plane_id] = raw->plane[plane_id];
1408
1409         level = G4X_WM_LEVEL_SR;
1410         if (!g4x_raw_crtc_wm_is_valid(crtc_state, level))
1411                 goto out;
1412
1413         raw = &crtc_state->wm.g4x.raw[level];
1414         wm_state->sr.plane = raw->plane[PLANE_PRIMARY];
1415         wm_state->sr.cursor = raw->plane[PLANE_CURSOR];
1416         wm_state->sr.fbc = raw->fbc;
1417
1418         wm_state->cxsr = num_active_planes == BIT(PLANE_PRIMARY);
1419
1420         level = G4X_WM_LEVEL_HPLL;
1421         if (!g4x_raw_crtc_wm_is_valid(crtc_state, level))
1422                 goto out;
1423
1424         raw = &crtc_state->wm.g4x.raw[level];
1425         wm_state->hpll.plane = raw->plane[PLANE_PRIMARY];
1426         wm_state->hpll.cursor = raw->plane[PLANE_CURSOR];
1427         wm_state->hpll.fbc = raw->fbc;
1428
1429         wm_state->hpll_en = wm_state->cxsr;
1430
1431         level++;
1432
1433  out:
1434         if (level == G4X_WM_LEVEL_NORMAL)
1435                 return -EINVAL;
1436
1437         /* invalidate the higher levels */
1438         g4x_invalidate_wms(crtc, wm_state, level);
1439
1440         /*
1441          * Determine if the FBC watermark(s) can be used. IF
1442          * this isn't the case we prefer to disable the FBC
1443          * watermark(s) rather than disable the SR/HPLL
1444          * level(s) entirely. 'level-1' is the highest valid
1445          * level here.
1446          */
1447         wm_state->fbc_en = g4x_compute_fbc_en(wm_state, level - 1);
1448
1449         return 0;
1450 }
1451
1452 static int g4x_compute_intermediate_wm(struct intel_crtc_state *new_crtc_state)
1453 {
1454         struct intel_crtc *crtc = to_intel_crtc(new_crtc_state->uapi.crtc);
1455         struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
1456         struct g4x_wm_state *intermediate = &new_crtc_state->wm.g4x.intermediate;
1457         const struct g4x_wm_state *optimal = &new_crtc_state->wm.g4x.optimal;
1458         struct intel_atomic_state *intel_state =
1459                 to_intel_atomic_state(new_crtc_state->uapi.state);
1460         const struct intel_crtc_state *old_crtc_state =
1461                 intel_atomic_get_old_crtc_state(intel_state, crtc);
1462         const struct g4x_wm_state *active = &old_crtc_state->wm.g4x.optimal;
1463         enum plane_id plane_id;
1464
1465         if (!new_crtc_state->hw.active || drm_atomic_crtc_needs_modeset(&new_crtc_state->uapi)) {
1466                 *intermediate = *optimal;
1467
1468                 intermediate->cxsr = false;
1469                 intermediate->hpll_en = false;
1470                 goto out;
1471         }
1472
1473         intermediate->cxsr = optimal->cxsr && active->cxsr &&
1474                 !new_crtc_state->disable_cxsr;
1475         intermediate->hpll_en = optimal->hpll_en && active->hpll_en &&
1476                 !new_crtc_state->disable_cxsr;
1477         intermediate->fbc_en = optimal->fbc_en && active->fbc_en;
1478
1479         for_each_plane_id_on_crtc(crtc, plane_id) {
1480                 intermediate->wm.plane[plane_id] =
1481                         max(optimal->wm.plane[plane_id],
1482                             active->wm.plane[plane_id]);
1483
1484                 drm_WARN_ON(&dev_priv->drm, intermediate->wm.plane[plane_id] >
1485                             g4x_plane_fifo_size(plane_id, G4X_WM_LEVEL_NORMAL));
1486         }
1487
1488         intermediate->sr.plane = max(optimal->sr.plane,
1489                                      active->sr.plane);
1490         intermediate->sr.cursor = max(optimal->sr.cursor,
1491                                       active->sr.cursor);
1492         intermediate->sr.fbc = max(optimal->sr.fbc,
1493                                    active->sr.fbc);
1494
1495         intermediate->hpll.plane = max(optimal->hpll.plane,
1496                                        active->hpll.plane);
1497         intermediate->hpll.cursor = max(optimal->hpll.cursor,
1498                                         active->hpll.cursor);
1499         intermediate->hpll.fbc = max(optimal->hpll.fbc,
1500                                      active->hpll.fbc);
1501
1502         drm_WARN_ON(&dev_priv->drm,
1503                     (intermediate->sr.plane >
1504                      g4x_plane_fifo_size(PLANE_PRIMARY, G4X_WM_LEVEL_SR) ||
1505                      intermediate->sr.cursor >
1506                      g4x_plane_fifo_size(PLANE_CURSOR, G4X_WM_LEVEL_SR)) &&
1507                     intermediate->cxsr);
1508         drm_WARN_ON(&dev_priv->drm,
1509                     (intermediate->sr.plane >
1510                      g4x_plane_fifo_size(PLANE_PRIMARY, G4X_WM_LEVEL_HPLL) ||
1511                      intermediate->sr.cursor >
1512                      g4x_plane_fifo_size(PLANE_CURSOR, G4X_WM_LEVEL_HPLL)) &&
1513                     intermediate->hpll_en);
1514
1515         drm_WARN_ON(&dev_priv->drm,
1516                     intermediate->sr.fbc > g4x_fbc_fifo_size(1) &&
1517                     intermediate->fbc_en && intermediate->cxsr);
1518         drm_WARN_ON(&dev_priv->drm,
1519                     intermediate->hpll.fbc > g4x_fbc_fifo_size(2) &&
1520                     intermediate->fbc_en && intermediate->hpll_en);
1521
1522 out:
1523         /*
1524          * If our intermediate WM are identical to the final WM, then we can
1525          * omit the post-vblank programming; only update if it's different.
1526          */
1527         if (memcmp(intermediate, optimal, sizeof(*intermediate)) != 0)
1528                 new_crtc_state->wm.need_postvbl_update = true;
1529
1530         return 0;
1531 }
1532
1533 static void g4x_merge_wm(struct drm_i915_private *dev_priv,
1534                          struct g4x_wm_values *wm)
1535 {
1536         struct intel_crtc *crtc;
1537         int num_active_pipes = 0;
1538
1539         wm->cxsr = true;
1540         wm->hpll_en = true;
1541         wm->fbc_en = true;
1542
1543         for_each_intel_crtc(&dev_priv->drm, crtc) {
1544                 const struct g4x_wm_state *wm_state = &crtc->wm.active.g4x;
1545
1546                 if (!crtc->active)
1547                         continue;
1548
1549                 if (!wm_state->cxsr)
1550                         wm->cxsr = false;
1551                 if (!wm_state->hpll_en)
1552                         wm->hpll_en = false;
1553                 if (!wm_state->fbc_en)
1554                         wm->fbc_en = false;
1555
1556                 num_active_pipes++;
1557         }
1558
1559         if (num_active_pipes != 1) {
1560                 wm->cxsr = false;
1561                 wm->hpll_en = false;
1562                 wm->fbc_en = false;
1563         }
1564
1565         for_each_intel_crtc(&dev_priv->drm, crtc) {
1566                 const struct g4x_wm_state *wm_state = &crtc->wm.active.g4x;
1567                 enum pipe pipe = crtc->pipe;
1568
1569                 wm->pipe[pipe] = wm_state->wm;
1570                 if (crtc->active && wm->cxsr)
1571                         wm->sr = wm_state->sr;
1572                 if (crtc->active && wm->hpll_en)
1573                         wm->hpll = wm_state->hpll;
1574         }
1575 }
1576
1577 static void g4x_program_watermarks(struct drm_i915_private *dev_priv)
1578 {
1579         struct g4x_wm_values *old_wm = &dev_priv->wm.g4x;
1580         struct g4x_wm_values new_wm = {};
1581
1582         g4x_merge_wm(dev_priv, &new_wm);
1583
1584         if (memcmp(old_wm, &new_wm, sizeof(new_wm)) == 0)
1585                 return;
1586
1587         if (is_disabling(old_wm->cxsr, new_wm.cxsr, true))
1588                 _intel_set_memory_cxsr(dev_priv, false);
1589
1590         g4x_write_wm_values(dev_priv, &new_wm);
1591
1592         if (is_enabling(old_wm->cxsr, new_wm.cxsr, true))
1593                 _intel_set_memory_cxsr(dev_priv, true);
1594
1595         *old_wm = new_wm;
1596 }
1597
1598 static void g4x_initial_watermarks(struct intel_atomic_state *state,
1599                                    struct intel_crtc *crtc)
1600 {
1601         struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
1602         const struct intel_crtc_state *crtc_state =
1603                 intel_atomic_get_new_crtc_state(state, crtc);
1604
1605         mutex_lock(&dev_priv->wm.wm_mutex);
1606         crtc->wm.active.g4x = crtc_state->wm.g4x.intermediate;
1607         g4x_program_watermarks(dev_priv);
1608         mutex_unlock(&dev_priv->wm.wm_mutex);
1609 }
1610
1611 static void g4x_optimize_watermarks(struct intel_atomic_state *state,
1612                                     struct intel_crtc *crtc)
1613 {
1614         struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
1615         const struct intel_crtc_state *crtc_state =
1616                 intel_atomic_get_new_crtc_state(state, crtc);
1617
1618         if (!crtc_state->wm.need_postvbl_update)
1619                 return;
1620
1621         mutex_lock(&dev_priv->wm.wm_mutex);
1622         crtc->wm.active.g4x = crtc_state->wm.g4x.optimal;
1623         g4x_program_watermarks(dev_priv);
1624         mutex_unlock(&dev_priv->wm.wm_mutex);
1625 }
1626
1627 /* latency must be in 0.1us units. */
1628 static unsigned int vlv_wm_method2(unsigned int pixel_rate,
1629                                    unsigned int htotal,
1630                                    unsigned int width,
1631                                    unsigned int cpp,
1632                                    unsigned int latency)
1633 {
1634         unsigned int ret;
1635
1636         ret = intel_wm_method2(pixel_rate, htotal,
1637                                width, cpp, latency);
1638         ret = DIV_ROUND_UP(ret, 64);
1639
1640         return ret;
1641 }
1642
1643 static void vlv_setup_wm_latency(struct drm_i915_private *dev_priv)
1644 {
1645         /* all latencies in usec */
1646         dev_priv->wm.pri_latency[VLV_WM_LEVEL_PM2] = 3;
1647
1648         dev_priv->wm.max_level = VLV_WM_LEVEL_PM2;
1649
1650         if (IS_CHERRYVIEW(dev_priv)) {
1651                 dev_priv->wm.pri_latency[VLV_WM_LEVEL_PM5] = 12;
1652                 dev_priv->wm.pri_latency[VLV_WM_LEVEL_DDR_DVFS] = 33;
1653
1654                 dev_priv->wm.max_level = VLV_WM_LEVEL_DDR_DVFS;
1655         }
1656 }
1657
1658 static u16 vlv_compute_wm_level(const struct intel_crtc_state *crtc_state,
1659                                 const struct intel_plane_state *plane_state,
1660                                 int level)
1661 {
1662         struct intel_plane *plane = to_intel_plane(plane_state->uapi.plane);
1663         struct drm_i915_private *dev_priv = to_i915(plane->base.dev);
1664         const struct drm_display_mode *pipe_mode =
1665                 &crtc_state->hw.pipe_mode;
1666         unsigned int clock, htotal, cpp, width, wm;
1667
1668         if (dev_priv->wm.pri_latency[level] == 0)
1669                 return USHRT_MAX;
1670
1671         if (!intel_wm_plane_visible(crtc_state, plane_state))
1672                 return 0;
1673
1674         cpp = plane_state->hw.fb->format->cpp[0];
1675         clock = pipe_mode->crtc_clock;
1676         htotal = pipe_mode->crtc_htotal;
1677         width = crtc_state->pipe_src_w;
1678
1679         if (plane->id == PLANE_CURSOR) {
1680                 /*
1681                  * FIXME the formula gives values that are
1682                  * too big for the cursor FIFO, and hence we
1683                  * would never be able to use cursors. For
1684                  * now just hardcode the watermark.
1685                  */
1686                 wm = 63;
1687         } else {
1688                 wm = vlv_wm_method2(clock, htotal, width, cpp,
1689                                     dev_priv->wm.pri_latency[level] * 10);
1690         }
1691
1692         return min_t(unsigned int, wm, USHRT_MAX);
1693 }
1694
1695 static bool vlv_need_sprite0_fifo_workaround(unsigned int active_planes)
1696 {
1697         return (active_planes & (BIT(PLANE_SPRITE0) |
1698                                  BIT(PLANE_SPRITE1))) == BIT(PLANE_SPRITE1);
1699 }
1700
1701 static int vlv_compute_fifo(struct intel_crtc_state *crtc_state)
1702 {
1703         struct intel_crtc *crtc = to_intel_crtc(crtc_state->uapi.crtc);
1704         struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
1705         const struct g4x_pipe_wm *raw =
1706                 &crtc_state->wm.vlv.raw[VLV_WM_LEVEL_PM2];
1707         struct vlv_fifo_state *fifo_state = &crtc_state->wm.vlv.fifo_state;
1708         unsigned int active_planes = crtc_state->active_planes & ~BIT(PLANE_CURSOR);
1709         int num_active_planes = hweight8(active_planes);
1710         const int fifo_size = 511;
1711         int fifo_extra, fifo_left = fifo_size;
1712         int sprite0_fifo_extra = 0;
1713         unsigned int total_rate;
1714         enum plane_id plane_id;
1715
1716         /*
1717          * When enabling sprite0 after sprite1 has already been enabled
1718          * we tend to get an underrun unless sprite0 already has some
1719          * FIFO space allcoated. Hence we always allocate at least one
1720          * cacheline for sprite0 whenever sprite1 is enabled.
1721          *
1722          * All other plane enable sequences appear immune to this problem.
1723          */
1724         if (vlv_need_sprite0_fifo_workaround(active_planes))
1725                 sprite0_fifo_extra = 1;
1726
1727         total_rate = raw->plane[PLANE_PRIMARY] +
1728                 raw->plane[PLANE_SPRITE0] +
1729                 raw->plane[PLANE_SPRITE1] +
1730                 sprite0_fifo_extra;
1731
1732         if (total_rate > fifo_size)
1733                 return -EINVAL;
1734
1735         if (total_rate == 0)
1736                 total_rate = 1;
1737
1738         for_each_plane_id_on_crtc(crtc, plane_id) {
1739                 unsigned int rate;
1740
1741                 if ((active_planes & BIT(plane_id)) == 0) {
1742                         fifo_state->plane[plane_id] = 0;
1743                         continue;
1744                 }
1745
1746                 rate = raw->plane[plane_id];
1747                 fifo_state->plane[plane_id] = fifo_size * rate / total_rate;
1748                 fifo_left -= fifo_state->plane[plane_id];
1749         }
1750
1751         fifo_state->plane[PLANE_SPRITE0] += sprite0_fifo_extra;
1752         fifo_left -= sprite0_fifo_extra;
1753
1754         fifo_state->plane[PLANE_CURSOR] = 63;
1755
1756         fifo_extra = DIV_ROUND_UP(fifo_left, num_active_planes ?: 1);
1757
1758         /* spread the remainder evenly */
1759         for_each_plane_id_on_crtc(crtc, plane_id) {
1760                 int plane_extra;
1761
1762                 if (fifo_left == 0)
1763                         break;
1764
1765                 if ((active_planes & BIT(plane_id)) == 0)
1766                         continue;
1767
1768                 plane_extra = min(fifo_extra, fifo_left);
1769                 fifo_state->plane[plane_id] += plane_extra;
1770                 fifo_left -= plane_extra;
1771         }
1772
1773         drm_WARN_ON(&dev_priv->drm, active_planes != 0 && fifo_left != 0);
1774
1775         /* give it all to the first plane if none are active */
1776         if (active_planes == 0) {
1777                 drm_WARN_ON(&dev_priv->drm, fifo_left != fifo_size);
1778                 fifo_state->plane[PLANE_PRIMARY] = fifo_left;
1779         }
1780
1781         return 0;
1782 }
1783
1784 /* mark all levels starting from 'level' as invalid */
1785 static void vlv_invalidate_wms(struct intel_crtc *crtc,
1786                                struct vlv_wm_state *wm_state, int level)
1787 {
1788         struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
1789
1790         for (; level < intel_wm_num_levels(dev_priv); level++) {
1791                 enum plane_id plane_id;
1792
1793                 for_each_plane_id_on_crtc(crtc, plane_id)
1794                         wm_state->wm[level].plane[plane_id] = USHRT_MAX;
1795
1796                 wm_state->sr[level].cursor = USHRT_MAX;
1797                 wm_state->sr[level].plane = USHRT_MAX;
1798         }
1799 }
1800
1801 static u16 vlv_invert_wm_value(u16 wm, u16 fifo_size)
1802 {
1803         if (wm > fifo_size)
1804                 return USHRT_MAX;
1805         else
1806                 return fifo_size - wm;
1807 }
1808
1809 /*
1810  * Starting from 'level' set all higher
1811  * levels to 'value' in the "raw" watermarks.
1812  */
1813 static bool vlv_raw_plane_wm_set(struct intel_crtc_state *crtc_state,
1814                                  int level, enum plane_id plane_id, u16 value)
1815 {
1816         struct drm_i915_private *dev_priv = to_i915(crtc_state->uapi.crtc->dev);
1817         int num_levels = intel_wm_num_levels(dev_priv);
1818         bool dirty = false;
1819
1820         for (; level < num_levels; level++) {
1821                 struct g4x_pipe_wm *raw = &crtc_state->wm.vlv.raw[level];
1822
1823                 dirty |= raw->plane[plane_id] != value;
1824                 raw->plane[plane_id] = value;
1825         }
1826
1827         return dirty;
1828 }
1829
1830 static bool vlv_raw_plane_wm_compute(struct intel_crtc_state *crtc_state,
1831                                      const struct intel_plane_state *plane_state)
1832 {
1833         struct intel_plane *plane = to_intel_plane(plane_state->uapi.plane);
1834         struct drm_i915_private *dev_priv = to_i915(crtc_state->uapi.crtc->dev);
1835         enum plane_id plane_id = plane->id;
1836         int num_levels = intel_wm_num_levels(to_i915(plane->base.dev));
1837         int level;
1838         bool dirty = false;
1839
1840         if (!intel_wm_plane_visible(crtc_state, plane_state)) {
1841                 dirty |= vlv_raw_plane_wm_set(crtc_state, 0, plane_id, 0);
1842                 goto out;
1843         }
1844
1845         for (level = 0; level < num_levels; level++) {
1846                 struct g4x_pipe_wm *raw = &crtc_state->wm.vlv.raw[level];
1847                 int wm = vlv_compute_wm_level(crtc_state, plane_state, level);
1848                 int max_wm = plane_id == PLANE_CURSOR ? 63 : 511;
1849
1850                 if (wm > max_wm)
1851                         break;
1852
1853                 dirty |= raw->plane[plane_id] != wm;
1854                 raw->plane[plane_id] = wm;
1855         }
1856
1857         /* mark all higher levels as invalid */
1858         dirty |= vlv_raw_plane_wm_set(crtc_state, level, plane_id, USHRT_MAX);
1859
1860 out:
1861         if (dirty)
1862                 drm_dbg_kms(&dev_priv->drm,
1863                             "%s watermarks: PM2=%d, PM5=%d, DDR DVFS=%d\n",
1864                             plane->base.name,
1865                             crtc_state->wm.vlv.raw[VLV_WM_LEVEL_PM2].plane[plane_id],
1866                             crtc_state->wm.vlv.raw[VLV_WM_LEVEL_PM5].plane[plane_id],
1867                             crtc_state->wm.vlv.raw[VLV_WM_LEVEL_DDR_DVFS].plane[plane_id]);
1868
1869         return dirty;
1870 }
1871
1872 static bool vlv_raw_plane_wm_is_valid(const struct intel_crtc_state *crtc_state,
1873                                       enum plane_id plane_id, int level)
1874 {
1875         const struct g4x_pipe_wm *raw =
1876                 &crtc_state->wm.vlv.raw[level];
1877         const struct vlv_fifo_state *fifo_state =
1878                 &crtc_state->wm.vlv.fifo_state;
1879
1880         return raw->plane[plane_id] <= fifo_state->plane[plane_id];
1881 }
1882
1883 static bool vlv_raw_crtc_wm_is_valid(const struct intel_crtc_state *crtc_state, int level)
1884 {
1885         return vlv_raw_plane_wm_is_valid(crtc_state, PLANE_PRIMARY, level) &&
1886                 vlv_raw_plane_wm_is_valid(crtc_state, PLANE_SPRITE0, level) &&
1887                 vlv_raw_plane_wm_is_valid(crtc_state, PLANE_SPRITE1, level) &&
1888                 vlv_raw_plane_wm_is_valid(crtc_state, PLANE_CURSOR, level);
1889 }
1890
1891 static int vlv_compute_pipe_wm(struct intel_crtc_state *crtc_state)
1892 {
1893         struct intel_crtc *crtc = to_intel_crtc(crtc_state->uapi.crtc);
1894         struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
1895         struct intel_atomic_state *state =
1896                 to_intel_atomic_state(crtc_state->uapi.state);
1897         struct vlv_wm_state *wm_state = &crtc_state->wm.vlv.optimal;
1898         const struct vlv_fifo_state *fifo_state =
1899                 &crtc_state->wm.vlv.fifo_state;
1900         int num_active_planes = hweight8(crtc_state->active_planes &
1901                                          ~BIT(PLANE_CURSOR));
1902         bool needs_modeset = drm_atomic_crtc_needs_modeset(&crtc_state->uapi);
1903         const struct intel_plane_state *old_plane_state;
1904         const struct intel_plane_state *new_plane_state;
1905         struct intel_plane *plane;
1906         enum plane_id plane_id;
1907         int level, ret, i;
1908         unsigned int dirty = 0;
1909
1910         for_each_oldnew_intel_plane_in_state(state, plane,
1911                                              old_plane_state,
1912                                              new_plane_state, i) {
1913                 if (new_plane_state->hw.crtc != &crtc->base &&
1914                     old_plane_state->hw.crtc != &crtc->base)
1915                         continue;
1916
1917                 if (vlv_raw_plane_wm_compute(crtc_state, new_plane_state))
1918                         dirty |= BIT(plane->id);
1919         }
1920
1921         /*
1922          * DSPARB registers may have been reset due to the
1923          * power well being turned off. Make sure we restore
1924          * them to a consistent state even if no primary/sprite
1925          * planes are initially active.
1926          */
1927         if (needs_modeset)
1928                 crtc_state->fifo_changed = true;
1929
1930         if (!dirty)
1931                 return 0;
1932
1933         /* cursor changes don't warrant a FIFO recompute */
1934         if (dirty & ~BIT(PLANE_CURSOR)) {
1935                 const struct intel_crtc_state *old_crtc_state =
1936                         intel_atomic_get_old_crtc_state(state, crtc);
1937                 const struct vlv_fifo_state *old_fifo_state =
1938                         &old_crtc_state->wm.vlv.fifo_state;
1939
1940                 ret = vlv_compute_fifo(crtc_state);
1941                 if (ret)
1942                         return ret;
1943
1944                 if (needs_modeset ||
1945                     memcmp(old_fifo_state, fifo_state,
1946                            sizeof(*fifo_state)) != 0)
1947                         crtc_state->fifo_changed = true;
1948         }
1949
1950         /* initially allow all levels */
1951         wm_state->num_levels = intel_wm_num_levels(dev_priv);
1952         /*
1953          * Note that enabling cxsr with no primary/sprite planes
1954          * enabled can wedge the pipe. Hence we only allow cxsr
1955          * with exactly one enabled primary/sprite plane.
1956          */
1957         wm_state->cxsr = crtc->pipe != PIPE_C && num_active_planes == 1;
1958
1959         for (level = 0; level < wm_state->num_levels; level++) {
1960                 const struct g4x_pipe_wm *raw = &crtc_state->wm.vlv.raw[level];
1961                 const int sr_fifo_size = INTEL_NUM_PIPES(dev_priv) * 512 - 1;
1962
1963                 if (!vlv_raw_crtc_wm_is_valid(crtc_state, level))
1964                         break;
1965
1966                 for_each_plane_id_on_crtc(crtc, plane_id) {
1967                         wm_state->wm[level].plane[plane_id] =
1968                                 vlv_invert_wm_value(raw->plane[plane_id],
1969                                                     fifo_state->plane[plane_id]);
1970                 }
1971
1972                 wm_state->sr[level].plane =
1973                         vlv_invert_wm_value(max3(raw->plane[PLANE_PRIMARY],
1974                                                  raw->plane[PLANE_SPRITE0],
1975                                                  raw->plane[PLANE_SPRITE1]),
1976                                             sr_fifo_size);
1977
1978                 wm_state->sr[level].cursor =
1979                         vlv_invert_wm_value(raw->plane[PLANE_CURSOR],
1980                                             63);
1981         }
1982
1983         if (level == 0)
1984                 return -EINVAL;
1985
1986         /* limit to only levels we can actually handle */
1987         wm_state->num_levels = level;
1988
1989         /* invalidate the higher levels */
1990         vlv_invalidate_wms(crtc, wm_state, level);
1991
1992         return 0;
1993 }
1994
1995 #define VLV_FIFO(plane, value) \
1996         (((value) << DSPARB_ ## plane ## _SHIFT_VLV) & DSPARB_ ## plane ## _MASK_VLV)
1997
1998 static void vlv_atomic_update_fifo(struct intel_atomic_state *state,
1999                                    struct intel_crtc *crtc)
2000 {
2001         struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
2002         struct intel_uncore *uncore = &dev_priv->uncore;
2003         const struct intel_crtc_state *crtc_state =
2004                 intel_atomic_get_new_crtc_state(state, crtc);
2005         const struct vlv_fifo_state *fifo_state =
2006                 &crtc_state->wm.vlv.fifo_state;
2007         int sprite0_start, sprite1_start, fifo_size;
2008         u32 dsparb, dsparb2, dsparb3;
2009
2010         if (!crtc_state->fifo_changed)
2011                 return;
2012
2013         sprite0_start = fifo_state->plane[PLANE_PRIMARY];
2014         sprite1_start = fifo_state->plane[PLANE_SPRITE0] + sprite0_start;
2015         fifo_size = fifo_state->plane[PLANE_SPRITE1] + sprite1_start;
2016
2017         drm_WARN_ON(&dev_priv->drm, fifo_state->plane[PLANE_CURSOR] != 63);
2018         drm_WARN_ON(&dev_priv->drm, fifo_size != 511);
2019
2020         trace_vlv_fifo_size(crtc, sprite0_start, sprite1_start, fifo_size);
2021
2022         /*
2023          * uncore.lock serves a double purpose here. It allows us to
2024          * use the less expensive I915_{READ,WRITE}_FW() functions, and
2025          * it protects the DSPARB registers from getting clobbered by
2026          * parallel updates from multiple pipes.
2027          *
2028          * intel_pipe_update_start() has already disabled interrupts
2029          * for us, so a plain spin_lock() is sufficient here.
2030          */
2031         spin_lock(&uncore->lock);
2032
2033         switch (crtc->pipe) {
2034         case PIPE_A:
2035                 dsparb = intel_uncore_read_fw(uncore, DSPARB);
2036                 dsparb2 = intel_uncore_read_fw(uncore, DSPARB2);
2037
2038                 dsparb &= ~(VLV_FIFO(SPRITEA, 0xff) |
2039                             VLV_FIFO(SPRITEB, 0xff));
2040                 dsparb |= (VLV_FIFO(SPRITEA, sprite0_start) |
2041                            VLV_FIFO(SPRITEB, sprite1_start));
2042
2043                 dsparb2 &= ~(VLV_FIFO(SPRITEA_HI, 0x1) |
2044                              VLV_FIFO(SPRITEB_HI, 0x1));
2045                 dsparb2 |= (VLV_FIFO(SPRITEA_HI, sprite0_start >> 8) |
2046                            VLV_FIFO(SPRITEB_HI, sprite1_start >> 8));
2047
2048                 intel_uncore_write_fw(uncore, DSPARB, dsparb);
2049                 intel_uncore_write_fw(uncore, DSPARB2, dsparb2);
2050                 break;
2051         case PIPE_B:
2052                 dsparb = intel_uncore_read_fw(uncore, DSPARB);
2053                 dsparb2 = intel_uncore_read_fw(uncore, DSPARB2);
2054
2055                 dsparb &= ~(VLV_FIFO(SPRITEC, 0xff) |
2056                             VLV_FIFO(SPRITED, 0xff));
2057                 dsparb |= (VLV_FIFO(SPRITEC, sprite0_start) |
2058                            VLV_FIFO(SPRITED, sprite1_start));
2059
2060                 dsparb2 &= ~(VLV_FIFO(SPRITEC_HI, 0xff) |
2061                              VLV_FIFO(SPRITED_HI, 0xff));
2062                 dsparb2 |= (VLV_FIFO(SPRITEC_HI, sprite0_start >> 8) |
2063                            VLV_FIFO(SPRITED_HI, sprite1_start >> 8));
2064
2065                 intel_uncore_write_fw(uncore, DSPARB, dsparb);
2066                 intel_uncore_write_fw(uncore, DSPARB2, dsparb2);
2067                 break;
2068         case PIPE_C:
2069                 dsparb3 = intel_uncore_read_fw(uncore, DSPARB3);
2070                 dsparb2 = intel_uncore_read_fw(uncore, DSPARB2);
2071
2072                 dsparb3 &= ~(VLV_FIFO(SPRITEE, 0xff) |
2073                              VLV_FIFO(SPRITEF, 0xff));
2074                 dsparb3 |= (VLV_FIFO(SPRITEE, sprite0_start) |
2075                             VLV_FIFO(SPRITEF, sprite1_start));
2076
2077                 dsparb2 &= ~(VLV_FIFO(SPRITEE_HI, 0xff) |
2078                              VLV_FIFO(SPRITEF_HI, 0xff));
2079                 dsparb2 |= (VLV_FIFO(SPRITEE_HI, sprite0_start >> 8) |
2080                            VLV_FIFO(SPRITEF_HI, sprite1_start >> 8));
2081
2082                 intel_uncore_write_fw(uncore, DSPARB3, dsparb3);
2083                 intel_uncore_write_fw(uncore, DSPARB2, dsparb2);
2084                 break;
2085         default:
2086                 break;
2087         }
2088
2089         intel_uncore_posting_read_fw(uncore, DSPARB);
2090
2091         spin_unlock(&uncore->lock);
2092 }
2093
2094 #undef VLV_FIFO
2095
2096 static int vlv_compute_intermediate_wm(struct intel_crtc_state *new_crtc_state)
2097 {
2098         struct intel_crtc *crtc = to_intel_crtc(new_crtc_state->uapi.crtc);
2099         struct vlv_wm_state *intermediate = &new_crtc_state->wm.vlv.intermediate;
2100         const struct vlv_wm_state *optimal = &new_crtc_state->wm.vlv.optimal;
2101         struct intel_atomic_state *intel_state =
2102                 to_intel_atomic_state(new_crtc_state->uapi.state);
2103         const struct intel_crtc_state *old_crtc_state =
2104                 intel_atomic_get_old_crtc_state(intel_state, crtc);
2105         const struct vlv_wm_state *active = &old_crtc_state->wm.vlv.optimal;
2106         int level;
2107
2108         if (!new_crtc_state->hw.active || drm_atomic_crtc_needs_modeset(&new_crtc_state->uapi)) {
2109                 *intermediate = *optimal;
2110
2111                 intermediate->cxsr = false;
2112                 goto out;
2113         }
2114
2115         intermediate->num_levels = min(optimal->num_levels, active->num_levels);
2116         intermediate->cxsr = optimal->cxsr && active->cxsr &&
2117                 !new_crtc_state->disable_cxsr;
2118
2119         for (level = 0; level < intermediate->num_levels; level++) {
2120                 enum plane_id plane_id;
2121
2122                 for_each_plane_id_on_crtc(crtc, plane_id) {
2123                         intermediate->wm[level].plane[plane_id] =
2124                                 min(optimal->wm[level].plane[plane_id],
2125                                     active->wm[level].plane[plane_id]);
2126                 }
2127
2128                 intermediate->sr[level].plane = min(optimal->sr[level].plane,
2129                                                     active->sr[level].plane);
2130                 intermediate->sr[level].cursor = min(optimal->sr[level].cursor,
2131                                                      active->sr[level].cursor);
2132         }
2133
2134         vlv_invalidate_wms(crtc, intermediate, level);
2135
2136 out:
2137         /*
2138          * If our intermediate WM are identical to the final WM, then we can
2139          * omit the post-vblank programming; only update if it's different.
2140          */
2141         if (memcmp(intermediate, optimal, sizeof(*intermediate)) != 0)
2142                 new_crtc_state->wm.need_postvbl_update = true;
2143
2144         return 0;
2145 }
2146
2147 static void vlv_merge_wm(struct drm_i915_private *dev_priv,
2148                          struct vlv_wm_values *wm)
2149 {
2150         struct intel_crtc *crtc;
2151         int num_active_pipes = 0;
2152
2153         wm->level = dev_priv->wm.max_level;
2154         wm->cxsr = true;
2155
2156         for_each_intel_crtc(&dev_priv->drm, crtc) {
2157                 const struct vlv_wm_state *wm_state = &crtc->wm.active.vlv;
2158
2159                 if (!crtc->active)
2160                         continue;
2161
2162                 if (!wm_state->cxsr)
2163                         wm->cxsr = false;
2164
2165                 num_active_pipes++;
2166                 wm->level = min_t(int, wm->level, wm_state->num_levels - 1);
2167         }
2168
2169         if (num_active_pipes != 1)
2170                 wm->cxsr = false;
2171
2172         if (num_active_pipes > 1)
2173                 wm->level = VLV_WM_LEVEL_PM2;
2174
2175         for_each_intel_crtc(&dev_priv->drm, crtc) {
2176                 const struct vlv_wm_state *wm_state = &crtc->wm.active.vlv;
2177                 enum pipe pipe = crtc->pipe;
2178
2179                 wm->pipe[pipe] = wm_state->wm[wm->level];
2180                 if (crtc->active && wm->cxsr)
2181                         wm->sr = wm_state->sr[wm->level];
2182
2183                 wm->ddl[pipe].plane[PLANE_PRIMARY] = DDL_PRECISION_HIGH | 2;
2184                 wm->ddl[pipe].plane[PLANE_SPRITE0] = DDL_PRECISION_HIGH | 2;
2185                 wm->ddl[pipe].plane[PLANE_SPRITE1] = DDL_PRECISION_HIGH | 2;
2186                 wm->ddl[pipe].plane[PLANE_CURSOR] = DDL_PRECISION_HIGH | 2;
2187         }
2188 }
2189
2190 static void vlv_program_watermarks(struct drm_i915_private *dev_priv)
2191 {
2192         struct vlv_wm_values *old_wm = &dev_priv->wm.vlv;
2193         struct vlv_wm_values new_wm = {};
2194
2195         vlv_merge_wm(dev_priv, &new_wm);
2196
2197         if (memcmp(old_wm, &new_wm, sizeof(new_wm)) == 0)
2198                 return;
2199
2200         if (is_disabling(old_wm->level, new_wm.level, VLV_WM_LEVEL_DDR_DVFS))
2201                 chv_set_memory_dvfs(dev_priv, false);
2202
2203         if (is_disabling(old_wm->level, new_wm.level, VLV_WM_LEVEL_PM5))
2204                 chv_set_memory_pm5(dev_priv, false);
2205
2206         if (is_disabling(old_wm->cxsr, new_wm.cxsr, true))
2207                 _intel_set_memory_cxsr(dev_priv, false);
2208
2209         vlv_write_wm_values(dev_priv, &new_wm);
2210
2211         if (is_enabling(old_wm->cxsr, new_wm.cxsr, true))
2212                 _intel_set_memory_cxsr(dev_priv, true);
2213
2214         if (is_enabling(old_wm->level, new_wm.level, VLV_WM_LEVEL_PM5))
2215                 chv_set_memory_pm5(dev_priv, true);
2216
2217         if (is_enabling(old_wm->level, new_wm.level, VLV_WM_LEVEL_DDR_DVFS))
2218                 chv_set_memory_dvfs(dev_priv, true);
2219
2220         *old_wm = new_wm;
2221 }
2222
2223 static void vlv_initial_watermarks(struct intel_atomic_state *state,
2224                                    struct intel_crtc *crtc)
2225 {
2226         struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
2227         const struct intel_crtc_state *crtc_state =
2228                 intel_atomic_get_new_crtc_state(state, crtc);
2229
2230         mutex_lock(&dev_priv->wm.wm_mutex);
2231         crtc->wm.active.vlv = crtc_state->wm.vlv.intermediate;
2232         vlv_program_watermarks(dev_priv);
2233         mutex_unlock(&dev_priv->wm.wm_mutex);
2234 }
2235
2236 static void vlv_optimize_watermarks(struct intel_atomic_state *state,
2237                                     struct intel_crtc *crtc)
2238 {
2239         struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
2240         const struct intel_crtc_state *crtc_state =
2241                 intel_atomic_get_new_crtc_state(state, crtc);
2242
2243         if (!crtc_state->wm.need_postvbl_update)
2244                 return;
2245
2246         mutex_lock(&dev_priv->wm.wm_mutex);
2247         crtc->wm.active.vlv = crtc_state->wm.vlv.optimal;
2248         vlv_program_watermarks(dev_priv);
2249         mutex_unlock(&dev_priv->wm.wm_mutex);
2250 }
2251
2252 static void i965_update_wm(struct intel_crtc *unused_crtc)
2253 {
2254         struct drm_i915_private *dev_priv = to_i915(unused_crtc->base.dev);
2255         struct intel_crtc *crtc;
2256         int srwm = 1;
2257         int cursor_sr = 16;
2258         bool cxsr_enabled;
2259
2260         /* Calc sr entries for one plane configs */
2261         crtc = single_enabled_crtc(dev_priv);
2262         if (crtc) {
2263                 /* self-refresh has much higher latency */
2264                 static const int sr_latency_ns = 12000;
2265                 const struct drm_display_mode *pipe_mode =
2266                         &crtc->config->hw.pipe_mode;
2267                 const struct drm_framebuffer *fb =
2268                         crtc->base.primary->state->fb;
2269                 int clock = pipe_mode->crtc_clock;
2270                 int htotal = pipe_mode->crtc_htotal;
2271                 int hdisplay = crtc->config->pipe_src_w;
2272                 int cpp = fb->format->cpp[0];
2273                 int entries;
2274
2275                 entries = intel_wm_method2(clock, htotal,
2276                                            hdisplay, cpp, sr_latency_ns / 100);
2277                 entries = DIV_ROUND_UP(entries, I915_FIFO_LINE_SIZE);
2278                 srwm = I965_FIFO_SIZE - entries;
2279                 if (srwm < 0)
2280                         srwm = 1;
2281                 srwm &= 0x1ff;
2282                 drm_dbg_kms(&dev_priv->drm,
2283                             "self-refresh entries: %d, wm: %d\n",
2284                             entries, srwm);
2285
2286                 entries = intel_wm_method2(clock, htotal,
2287                                            crtc->base.cursor->state->crtc_w, 4,
2288                                            sr_latency_ns / 100);
2289                 entries = DIV_ROUND_UP(entries,
2290                                        i965_cursor_wm_info.cacheline_size) +
2291                         i965_cursor_wm_info.guard_size;
2292
2293                 cursor_sr = i965_cursor_wm_info.fifo_size - entries;
2294                 if (cursor_sr > i965_cursor_wm_info.max_wm)
2295                         cursor_sr = i965_cursor_wm_info.max_wm;
2296
2297                 drm_dbg_kms(&dev_priv->drm,
2298                             "self-refresh watermark: display plane %d "
2299                             "cursor %d\n", srwm, cursor_sr);
2300
2301                 cxsr_enabled = true;
2302         } else {
2303                 cxsr_enabled = false;
2304                 /* Turn off self refresh if both pipes are enabled */
2305                 intel_set_memory_cxsr(dev_priv, false);
2306         }
2307
2308         drm_dbg_kms(&dev_priv->drm,
2309                     "Setting FIFO watermarks - A: 8, B: 8, C: 8, SR %d\n",
2310                     srwm);
2311
2312         /* 965 has limitations... */
2313         intel_uncore_write(&dev_priv->uncore, DSPFW1, FW_WM(srwm, SR) |
2314                    FW_WM(8, CURSORB) |
2315                    FW_WM(8, PLANEB) |
2316                    FW_WM(8, PLANEA));
2317         intel_uncore_write(&dev_priv->uncore, DSPFW2, FW_WM(8, CURSORA) |
2318                    FW_WM(8, PLANEC_OLD));
2319         /* update cursor SR watermark */
2320         intel_uncore_write(&dev_priv->uncore, DSPFW3, FW_WM(cursor_sr, CURSOR_SR));
2321
2322         if (cxsr_enabled)
2323                 intel_set_memory_cxsr(dev_priv, true);
2324 }
2325
2326 #undef FW_WM
2327
2328 static void i9xx_update_wm(struct intel_crtc *unused_crtc)
2329 {
2330         struct drm_i915_private *dev_priv = to_i915(unused_crtc->base.dev);
2331         const struct intel_watermark_params *wm_info;
2332         u32 fwater_lo;
2333         u32 fwater_hi;
2334         int cwm, srwm = 1;
2335         int fifo_size;
2336         int planea_wm, planeb_wm;
2337         struct intel_crtc *crtc, *enabled = NULL;
2338
2339         if (IS_I945GM(dev_priv))
2340                 wm_info = &i945_wm_info;
2341         else if (!IS_GEN(dev_priv, 2))
2342                 wm_info = &i915_wm_info;
2343         else
2344                 wm_info = &i830_a_wm_info;
2345
2346         fifo_size = dev_priv->display.get_fifo_size(dev_priv, PLANE_A);
2347         crtc = intel_get_crtc_for_plane(dev_priv, PLANE_A);
2348         if (intel_crtc_active(crtc)) {
2349                 const struct drm_display_mode *pipe_mode =
2350                         &crtc->config->hw.pipe_mode;
2351                 const struct drm_framebuffer *fb =
2352                         crtc->base.primary->state->fb;
2353                 int cpp;
2354
2355                 if (IS_GEN(dev_priv, 2))
2356                         cpp = 4;
2357                 else
2358                         cpp = fb->format->cpp[0];
2359
2360                 planea_wm = intel_calculate_wm(pipe_mode->crtc_clock,
2361                                                wm_info, fifo_size, cpp,
2362                                                pessimal_latency_ns);
2363                 enabled = crtc;
2364         } else {
2365                 planea_wm = fifo_size - wm_info->guard_size;
2366                 if (planea_wm > (long)wm_info->max_wm)
2367                         planea_wm = wm_info->max_wm;
2368         }
2369
2370         if (IS_GEN(dev_priv, 2))
2371                 wm_info = &i830_bc_wm_info;
2372
2373         fifo_size = dev_priv->display.get_fifo_size(dev_priv, PLANE_B);
2374         crtc = intel_get_crtc_for_plane(dev_priv, PLANE_B);
2375         if (intel_crtc_active(crtc)) {
2376                 const struct drm_display_mode *pipe_mode =
2377                         &crtc->config->hw.pipe_mode;
2378                 const struct drm_framebuffer *fb =
2379                         crtc->base.primary->state->fb;
2380                 int cpp;
2381
2382                 if (IS_GEN(dev_priv, 2))
2383                         cpp = 4;
2384                 else
2385                         cpp = fb->format->cpp[0];
2386
2387                 planeb_wm = intel_calculate_wm(pipe_mode->crtc_clock,
2388                                                wm_info, fifo_size, cpp,
2389                                                pessimal_latency_ns);
2390                 if (enabled == NULL)
2391                         enabled = crtc;
2392                 else
2393                         enabled = NULL;
2394         } else {
2395                 planeb_wm = fifo_size - wm_info->guard_size;
2396                 if (planeb_wm > (long)wm_info->max_wm)
2397                         planeb_wm = wm_info->max_wm;
2398         }
2399
2400         drm_dbg_kms(&dev_priv->drm,
2401                     "FIFO watermarks - A: %d, B: %d\n", planea_wm, planeb_wm);
2402
2403         if (IS_I915GM(dev_priv) && enabled) {
2404                 struct drm_i915_gem_object *obj;
2405
2406                 obj = intel_fb_obj(enabled->base.primary->state->fb);
2407
2408                 /* self-refresh seems busted with untiled */
2409                 if (!i915_gem_object_is_tiled(obj))
2410                         enabled = NULL;
2411         }
2412
2413         /*
2414          * Overlay gets an aggressive default since video jitter is bad.
2415          */
2416         cwm = 2;
2417
2418         /* Play safe and disable self-refresh before adjusting watermarks. */
2419         intel_set_memory_cxsr(dev_priv, false);
2420
2421         /* Calc sr entries for one plane configs */
2422         if (HAS_FW_BLC(dev_priv) && enabled) {
2423                 /* self-refresh has much higher latency */
2424                 static const int sr_latency_ns = 6000;
2425                 const struct drm_display_mode *pipe_mode =
2426                         &enabled->config->hw.pipe_mode;
2427                 const struct drm_framebuffer *fb =
2428                         enabled->base.primary->state->fb;
2429                 int clock = pipe_mode->crtc_clock;
2430                 int htotal = pipe_mode->crtc_htotal;
2431                 int hdisplay = enabled->config->pipe_src_w;
2432                 int cpp;
2433                 int entries;
2434
2435                 if (IS_I915GM(dev_priv) || IS_I945GM(dev_priv))
2436                         cpp = 4;
2437                 else
2438                         cpp = fb->format->cpp[0];
2439
2440                 entries = intel_wm_method2(clock, htotal, hdisplay, cpp,
2441                                            sr_latency_ns / 100);
2442                 entries = DIV_ROUND_UP(entries, wm_info->cacheline_size);
2443                 drm_dbg_kms(&dev_priv->drm,
2444                             "self-refresh entries: %d\n", entries);
2445                 srwm = wm_info->fifo_size - entries;
2446                 if (srwm < 0)
2447                         srwm = 1;
2448
2449                 if (IS_I945G(dev_priv) || IS_I945GM(dev_priv))
2450                         intel_uncore_write(&dev_priv->uncore, FW_BLC_SELF,
2451                                    FW_BLC_SELF_FIFO_MASK | (srwm & 0xff));
2452                 else
2453                         intel_uncore_write(&dev_priv->uncore, FW_BLC_SELF, srwm & 0x3f);
2454         }
2455
2456         drm_dbg_kms(&dev_priv->drm,
2457                     "Setting FIFO watermarks - A: %d, B: %d, C: %d, SR %d\n",
2458                      planea_wm, planeb_wm, cwm, srwm);
2459
2460         fwater_lo = ((planeb_wm & 0x3f) << 16) | (planea_wm & 0x3f);
2461         fwater_hi = (cwm & 0x1f);
2462
2463         /* Set request length to 8 cachelines per fetch */
2464         fwater_lo = fwater_lo | (1 << 24) | (1 << 8);
2465         fwater_hi = fwater_hi | (1 << 8);
2466
2467         intel_uncore_write(&dev_priv->uncore, FW_BLC, fwater_lo);
2468         intel_uncore_write(&dev_priv->uncore, FW_BLC2, fwater_hi);
2469
2470         if (enabled)
2471                 intel_set_memory_cxsr(dev_priv, true);
2472 }
2473
2474 static void i845_update_wm(struct intel_crtc *unused_crtc)
2475 {
2476         struct drm_i915_private *dev_priv = to_i915(unused_crtc->base.dev);
2477         struct intel_crtc *crtc;
2478         const struct drm_display_mode *pipe_mode;
2479         u32 fwater_lo;
2480         int planea_wm;
2481
2482         crtc = single_enabled_crtc(dev_priv);
2483         if (crtc == NULL)
2484                 return;
2485
2486         pipe_mode = &crtc->config->hw.pipe_mode;
2487         planea_wm = intel_calculate_wm(pipe_mode->crtc_clock,
2488                                        &i845_wm_info,
2489                                        dev_priv->display.get_fifo_size(dev_priv, PLANE_A),
2490                                        4, pessimal_latency_ns);
2491         fwater_lo = intel_uncore_read(&dev_priv->uncore, FW_BLC) & ~0xfff;
2492         fwater_lo |= (3<<8) | planea_wm;
2493
2494         drm_dbg_kms(&dev_priv->drm,
2495                     "Setting FIFO watermarks - A: %d\n", planea_wm);
2496
2497         intel_uncore_write(&dev_priv->uncore, FW_BLC, fwater_lo);
2498 }
2499
2500 /* latency must be in 0.1us units. */
2501 static unsigned int ilk_wm_method1(unsigned int pixel_rate,
2502                                    unsigned int cpp,
2503                                    unsigned int latency)
2504 {
2505         unsigned int ret;
2506
2507         ret = intel_wm_method1(pixel_rate, cpp, latency);
2508         ret = DIV_ROUND_UP(ret, 64) + 2;
2509
2510         return ret;
2511 }
2512
2513 /* latency must be in 0.1us units. */
2514 static unsigned int ilk_wm_method2(unsigned int pixel_rate,
2515                                    unsigned int htotal,
2516                                    unsigned int width,
2517                                    unsigned int cpp,
2518                                    unsigned int latency)
2519 {
2520         unsigned int ret;
2521
2522         ret = intel_wm_method2(pixel_rate, htotal,
2523                                width, cpp, latency);
2524         ret = DIV_ROUND_UP(ret, 64) + 2;
2525
2526         return ret;
2527 }
2528
2529 static u32 ilk_wm_fbc(u32 pri_val, u32 horiz_pixels, u8 cpp)
2530 {
2531         /*
2532          * Neither of these should be possible since this function shouldn't be
2533          * called if the CRTC is off or the plane is invisible.  But let's be
2534          * extra paranoid to avoid a potential divide-by-zero if we screw up
2535          * elsewhere in the driver.
2536          */
2537         if (WARN_ON(!cpp))
2538                 return 0;
2539         if (WARN_ON(!horiz_pixels))
2540                 return 0;
2541
2542         return DIV_ROUND_UP(pri_val * 64, horiz_pixels * cpp) + 2;
2543 }
2544
2545 struct ilk_wm_maximums {
2546         u16 pri;
2547         u16 spr;
2548         u16 cur;
2549         u16 fbc;
2550 };
2551
2552 /*
2553  * For both WM_PIPE and WM_LP.
2554  * mem_value must be in 0.1us units.
2555  */
2556 static u32 ilk_compute_pri_wm(const struct intel_crtc_state *crtc_state,
2557                               const struct intel_plane_state *plane_state,
2558                               u32 mem_value, bool is_lp)
2559 {
2560         u32 method1, method2;
2561         int cpp;
2562
2563         if (mem_value == 0)
2564                 return U32_MAX;
2565
2566         if (!intel_wm_plane_visible(crtc_state, plane_state))
2567                 return 0;
2568
2569         cpp = plane_state->hw.fb->format->cpp[0];
2570
2571         method1 = ilk_wm_method1(crtc_state->pixel_rate, cpp, mem_value);
2572
2573         if (!is_lp)
2574                 return method1;
2575
2576         method2 = ilk_wm_method2(crtc_state->pixel_rate,
2577                                  crtc_state->hw.pipe_mode.crtc_htotal,
2578                                  drm_rect_width(&plane_state->uapi.dst),
2579                                  cpp, mem_value);
2580
2581         return min(method1, method2);
2582 }
2583
2584 /*
2585  * For both WM_PIPE and WM_LP.
2586  * mem_value must be in 0.1us units.
2587  */
2588 static u32 ilk_compute_spr_wm(const struct intel_crtc_state *crtc_state,
2589                               const struct intel_plane_state *plane_state,
2590                               u32 mem_value)
2591 {
2592         u32 method1, method2;
2593         int cpp;
2594
2595         if (mem_value == 0)
2596                 return U32_MAX;
2597
2598         if (!intel_wm_plane_visible(crtc_state, plane_state))
2599                 return 0;
2600
2601         cpp = plane_state->hw.fb->format->cpp[0];
2602
2603         method1 = ilk_wm_method1(crtc_state->pixel_rate, cpp, mem_value);
2604         method2 = ilk_wm_method2(crtc_state->pixel_rate,
2605                                  crtc_state->hw.pipe_mode.crtc_htotal,
2606                                  drm_rect_width(&plane_state->uapi.dst),
2607                                  cpp, mem_value);
2608         return min(method1, method2);
2609 }
2610
2611 /*
2612  * For both WM_PIPE and WM_LP.
2613  * mem_value must be in 0.1us units.
2614  */
2615 static u32 ilk_compute_cur_wm(const struct intel_crtc_state *crtc_state,
2616                               const struct intel_plane_state *plane_state,
2617                               u32 mem_value)
2618 {
2619         int cpp;
2620
2621         if (mem_value == 0)
2622                 return U32_MAX;
2623
2624         if (!intel_wm_plane_visible(crtc_state, plane_state))
2625                 return 0;
2626
2627         cpp = plane_state->hw.fb->format->cpp[0];
2628
2629         return ilk_wm_method2(crtc_state->pixel_rate,
2630                               crtc_state->hw.pipe_mode.crtc_htotal,
2631                               drm_rect_width(&plane_state->uapi.dst),
2632                               cpp, mem_value);
2633 }
2634
2635 /* Only for WM_LP. */
2636 static u32 ilk_compute_fbc_wm(const struct intel_crtc_state *crtc_state,
2637                               const struct intel_plane_state *plane_state,
2638                               u32 pri_val)
2639 {
2640         int cpp;
2641
2642         if (!intel_wm_plane_visible(crtc_state, plane_state))
2643                 return 0;
2644
2645         cpp = plane_state->hw.fb->format->cpp[0];
2646
2647         return ilk_wm_fbc(pri_val, drm_rect_width(&plane_state->uapi.dst),
2648                           cpp);
2649 }
2650
2651 static unsigned int
2652 ilk_display_fifo_size(const struct drm_i915_private *dev_priv)
2653 {
2654         if (INTEL_GEN(dev_priv) >= 8)
2655                 return 3072;
2656         else if (INTEL_GEN(dev_priv) >= 7)
2657                 return 768;
2658         else
2659                 return 512;
2660 }
2661
2662 static unsigned int
2663 ilk_plane_wm_reg_max(const struct drm_i915_private *dev_priv,
2664                      int level, bool is_sprite)
2665 {
2666         if (INTEL_GEN(dev_priv) >= 8)
2667                 /* BDW primary/sprite plane watermarks */
2668                 return level == 0 ? 255 : 2047;
2669         else if (INTEL_GEN(dev_priv) >= 7)
2670                 /* IVB/HSW primary/sprite plane watermarks */
2671                 return level == 0 ? 127 : 1023;
2672         else if (!is_sprite)
2673                 /* ILK/SNB primary plane watermarks */
2674                 return level == 0 ? 127 : 511;
2675         else
2676                 /* ILK/SNB sprite plane watermarks */
2677                 return level == 0 ? 63 : 255;
2678 }
2679
2680 static unsigned int
2681 ilk_cursor_wm_reg_max(const struct drm_i915_private *dev_priv, int level)
2682 {
2683         if (INTEL_GEN(dev_priv) >= 7)
2684                 return level == 0 ? 63 : 255;
2685         else
2686                 return level == 0 ? 31 : 63;
2687 }
2688
2689 static unsigned int ilk_fbc_wm_reg_max(const struct drm_i915_private *dev_priv)
2690 {
2691         if (INTEL_GEN(dev_priv) >= 8)
2692                 return 31;
2693         else
2694                 return 15;
2695 }
2696
2697 /* Calculate the maximum primary/sprite plane watermark */
2698 static unsigned int ilk_plane_wm_max(const struct drm_i915_private *dev_priv,
2699                                      int level,
2700                                      const struct intel_wm_config *config,
2701                                      enum intel_ddb_partitioning ddb_partitioning,
2702                                      bool is_sprite)
2703 {
2704         unsigned int fifo_size = ilk_display_fifo_size(dev_priv);
2705
2706         /* if sprites aren't enabled, sprites get nothing */
2707         if (is_sprite && !config->sprites_enabled)
2708                 return 0;
2709
2710         /* HSW allows LP1+ watermarks even with multiple pipes */
2711         if (level == 0 || config->num_pipes_active > 1) {
2712                 fifo_size /= INTEL_NUM_PIPES(dev_priv);
2713
2714                 /*
2715                  * For some reason the non self refresh
2716                  * FIFO size is only half of the self
2717                  * refresh FIFO size on ILK/SNB.
2718                  */
2719                 if (INTEL_GEN(dev_priv) <= 6)
2720                         fifo_size /= 2;
2721         }
2722
2723         if (config->sprites_enabled) {
2724                 /* level 0 is always calculated with 1:1 split */
2725                 if (level > 0 && ddb_partitioning == INTEL_DDB_PART_5_6) {
2726                         if (is_sprite)
2727                                 fifo_size *= 5;
2728                         fifo_size /= 6;
2729                 } else {
2730                         fifo_size /= 2;
2731                 }
2732         }
2733
2734         /* clamp to max that the registers can hold */
2735         return min(fifo_size, ilk_plane_wm_reg_max(dev_priv, level, is_sprite));
2736 }
2737
2738 /* Calculate the maximum cursor plane watermark */
2739 static unsigned int ilk_cursor_wm_max(const struct drm_i915_private *dev_priv,
2740                                       int level,
2741                                       const struct intel_wm_config *config)
2742 {
2743         /* HSW LP1+ watermarks w/ multiple pipes */
2744         if (level > 0 && config->num_pipes_active > 1)
2745                 return 64;
2746
2747         /* otherwise just report max that registers can hold */
2748         return ilk_cursor_wm_reg_max(dev_priv, level);
2749 }
2750
2751 static void ilk_compute_wm_maximums(const struct drm_i915_private *dev_priv,
2752                                     int level,
2753                                     const struct intel_wm_config *config,
2754                                     enum intel_ddb_partitioning ddb_partitioning,
2755                                     struct ilk_wm_maximums *max)
2756 {
2757         max->pri = ilk_plane_wm_max(dev_priv, level, config, ddb_partitioning, false);
2758         max->spr = ilk_plane_wm_max(dev_priv, level, config, ddb_partitioning, true);
2759         max->cur = ilk_cursor_wm_max(dev_priv, level, config);
2760         max->fbc = ilk_fbc_wm_reg_max(dev_priv);
2761 }
2762
2763 static void ilk_compute_wm_reg_maximums(const struct drm_i915_private *dev_priv,
2764                                         int level,
2765                                         struct ilk_wm_maximums *max)
2766 {
2767         max->pri = ilk_plane_wm_reg_max(dev_priv, level, false);
2768         max->spr = ilk_plane_wm_reg_max(dev_priv, level, true);
2769         max->cur = ilk_cursor_wm_reg_max(dev_priv, level);
2770         max->fbc = ilk_fbc_wm_reg_max(dev_priv);
2771 }
2772
2773 static bool ilk_validate_wm_level(int level,
2774                                   const struct ilk_wm_maximums *max,
2775                                   struct intel_wm_level *result)
2776 {
2777         bool ret;
2778
2779         /* already determined to be invalid? */
2780         if (!result->enable)
2781                 return false;
2782
2783         result->enable = result->pri_val <= max->pri &&
2784                          result->spr_val <= max->spr &&
2785                          result->cur_val <= max->cur;
2786
2787         ret = result->enable;
2788
2789         /*
2790          * HACK until we can pre-compute everything,
2791          * and thus fail gracefully if LP0 watermarks
2792          * are exceeded...
2793          */
2794         if (level == 0 && !result->enable) {
2795                 if (result->pri_val > max->pri)
2796                         DRM_DEBUG_KMS("Primary WM%d too large %u (max %u)\n",
2797                                       level, result->pri_val, max->pri);
2798                 if (result->spr_val > max->spr)
2799                         DRM_DEBUG_KMS("Sprite WM%d too large %u (max %u)\n",
2800                                       level, result->spr_val, max->spr);
2801                 if (result->cur_val > max->cur)
2802                         DRM_DEBUG_KMS("Cursor WM%d too large %u (max %u)\n",
2803                                       level, result->cur_val, max->cur);
2804
2805                 result->pri_val = min_t(u32, result->pri_val, max->pri);
2806                 result->spr_val = min_t(u32, result->spr_val, max->spr);
2807                 result->cur_val = min_t(u32, result->cur_val, max->cur);
2808                 result->enable = true;
2809         }
2810
2811         return ret;
2812 }
2813
2814 static void ilk_compute_wm_level(const struct drm_i915_private *dev_priv,
2815                                  const struct intel_crtc *crtc,
2816                                  int level,
2817                                  struct intel_crtc_state *crtc_state,
2818                                  const struct intel_plane_state *pristate,
2819                                  const struct intel_plane_state *sprstate,
2820                                  const struct intel_plane_state *curstate,
2821                                  struct intel_wm_level *result)
2822 {
2823         u16 pri_latency = dev_priv->wm.pri_latency[level];
2824         u16 spr_latency = dev_priv->wm.spr_latency[level];
2825         u16 cur_latency = dev_priv->wm.cur_latency[level];
2826
2827         /* WM1+ latency values stored in 0.5us units */
2828         if (level > 0) {
2829                 pri_latency *= 5;
2830                 spr_latency *= 5;
2831                 cur_latency *= 5;
2832         }
2833
2834         if (pristate) {
2835                 result->pri_val = ilk_compute_pri_wm(crtc_state, pristate,
2836                                                      pri_latency, level);
2837                 result->fbc_val = ilk_compute_fbc_wm(crtc_state, pristate, result->pri_val);
2838         }
2839
2840         if (sprstate)
2841                 result->spr_val = ilk_compute_spr_wm(crtc_state, sprstate, spr_latency);
2842
2843         if (curstate)
2844                 result->cur_val = ilk_compute_cur_wm(crtc_state, curstate, cur_latency);
2845
2846         result->enable = true;
2847 }
2848
2849 static void intel_read_wm_latency(struct drm_i915_private *dev_priv,
2850                                   u16 wm[8])
2851 {
2852         struct intel_uncore *uncore = &dev_priv->uncore;
2853
2854         if (INTEL_GEN(dev_priv) >= 9) {
2855                 u32 val;
2856                 int ret, i;
2857                 int level, max_level = ilk_wm_max_level(dev_priv);
2858
2859                 /* read the first set of memory latencies[0:3] */
2860                 val = 0; /* data0 to be programmed to 0 for first set */
2861                 ret = sandybridge_pcode_read(dev_priv,
2862                                              GEN9_PCODE_READ_MEM_LATENCY,
2863                                              &val, NULL);
2864
2865                 if (ret) {
2866                         drm_err(&dev_priv->drm,
2867                                 "SKL Mailbox read error = %d\n", ret);
2868                         return;
2869                 }
2870
2871                 wm[0] = val & GEN9_MEM_LATENCY_LEVEL_MASK;
2872                 wm[1] = (val >> GEN9_MEM_LATENCY_LEVEL_1_5_SHIFT) &
2873                                 GEN9_MEM_LATENCY_LEVEL_MASK;
2874                 wm[2] = (val >> GEN9_MEM_LATENCY_LEVEL_2_6_SHIFT) &
2875                                 GEN9_MEM_LATENCY_LEVEL_MASK;
2876                 wm[3] = (val >> GEN9_MEM_LATENCY_LEVEL_3_7_SHIFT) &
2877                                 GEN9_MEM_LATENCY_LEVEL_MASK;
2878
2879                 /* read the second set of memory latencies[4:7] */
2880                 val = 1; /* data0 to be programmed to 1 for second set */
2881                 ret = sandybridge_pcode_read(dev_priv,
2882                                              GEN9_PCODE_READ_MEM_LATENCY,
2883                                              &val, NULL);
2884                 if (ret) {
2885                         drm_err(&dev_priv->drm,
2886                                 "SKL Mailbox read error = %d\n", ret);
2887                         return;
2888                 }
2889
2890                 wm[4] = val & GEN9_MEM_LATENCY_LEVEL_MASK;
2891                 wm[5] = (val >> GEN9_MEM_LATENCY_LEVEL_1_5_SHIFT) &
2892                                 GEN9_MEM_LATENCY_LEVEL_MASK;
2893                 wm[6] = (val >> GEN9_MEM_LATENCY_LEVEL_2_6_SHIFT) &
2894                                 GEN9_MEM_LATENCY_LEVEL_MASK;
2895                 wm[7] = (val >> GEN9_MEM_LATENCY_LEVEL_3_7_SHIFT) &
2896                                 GEN9_MEM_LATENCY_LEVEL_MASK;
2897
2898                 /*
2899                  * If a level n (n > 1) has a 0us latency, all levels m (m >= n)
2900                  * need to be disabled. We make sure to sanitize the values out
2901                  * of the punit to satisfy this requirement.
2902                  */
2903                 for (level = 1; level <= max_level; level++) {
2904                         if (wm[level] == 0) {
2905                                 for (i = level + 1; i <= max_level; i++)
2906                                         wm[i] = 0;
2907                                 break;
2908                         }
2909                 }
2910
2911                 /*
2912                  * WaWmMemoryReadLatency:skl+,glk
2913                  *
2914                  * punit doesn't take into account the read latency so we need
2915                  * to add 2us to the various latency levels we retrieve from the
2916                  * punit when level 0 response data us 0us.
2917                  */
2918                 if (wm[0] == 0) {
2919                         wm[0] += 2;
2920                         for (level = 1; level <= max_level; level++) {
2921                                 if (wm[level] == 0)
2922                                         break;
2923                                 wm[level] += 2;
2924                         }
2925                 }
2926
2927                 /*
2928                  * WA Level-0 adjustment for 16GB DIMMs: SKL+
2929                  * If we could not get dimm info enable this WA to prevent from
2930                  * any underrun. If not able to get Dimm info assume 16GB dimm
2931                  * to avoid any underrun.
2932                  */
2933                 if (dev_priv->dram_info.wm_lv_0_adjust_needed)
2934                         wm[0] += 1;
2935
2936         } else if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv)) {
2937                 u64 sskpd = intel_uncore_read64(uncore, MCH_SSKPD);
2938
2939                 wm[0] = (sskpd >> 56) & 0xFF;
2940                 if (wm[0] == 0)
2941                         wm[0] = sskpd & 0xF;
2942                 wm[1] = (sskpd >> 4) & 0xFF;
2943                 wm[2] = (sskpd >> 12) & 0xFF;
2944                 wm[3] = (sskpd >> 20) & 0x1FF;
2945                 wm[4] = (sskpd >> 32) & 0x1FF;
2946         } else if (INTEL_GEN(dev_priv) >= 6) {
2947                 u32 sskpd = intel_uncore_read(uncore, MCH_SSKPD);
2948
2949                 wm[0] = (sskpd >> SSKPD_WM0_SHIFT) & SSKPD_WM_MASK;
2950                 wm[1] = (sskpd >> SSKPD_WM1_SHIFT) & SSKPD_WM_MASK;
2951                 wm[2] = (sskpd >> SSKPD_WM2_SHIFT) & SSKPD_WM_MASK;
2952                 wm[3] = (sskpd >> SSKPD_WM3_SHIFT) & SSKPD_WM_MASK;
2953         } else if (INTEL_GEN(dev_priv) >= 5) {
2954                 u32 mltr = intel_uncore_read(uncore, MLTR_ILK);
2955
2956                 /* ILK primary LP0 latency is 700 ns */
2957                 wm[0] = 7;
2958                 wm[1] = (mltr >> MLTR_WM1_SHIFT) & ILK_SRLT_MASK;
2959                 wm[2] = (mltr >> MLTR_WM2_SHIFT) & ILK_SRLT_MASK;
2960         } else {
2961                 MISSING_CASE(INTEL_DEVID(dev_priv));
2962         }
2963 }
2964
2965 static void intel_fixup_spr_wm_latency(struct drm_i915_private *dev_priv,
2966                                        u16 wm[5])
2967 {
2968         /* ILK sprite LP0 latency is 1300 ns */
2969         if (IS_GEN(dev_priv, 5))
2970                 wm[0] = 13;
2971 }
2972
2973 static void intel_fixup_cur_wm_latency(struct drm_i915_private *dev_priv,
2974                                        u16 wm[5])
2975 {
2976         /* ILK cursor LP0 latency is 1300 ns */
2977         if (IS_GEN(dev_priv, 5))
2978                 wm[0] = 13;
2979 }
2980
2981 int ilk_wm_max_level(const struct drm_i915_private *dev_priv)
2982 {
2983         /* how many WM levels are we expecting */
2984         if (INTEL_GEN(dev_priv) >= 9)
2985                 return 7;
2986         else if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
2987                 return 4;
2988         else if (INTEL_GEN(dev_priv) >= 6)
2989                 return 3;
2990         else
2991                 return 2;
2992 }
2993
2994 static void intel_print_wm_latency(struct drm_i915_private *dev_priv,
2995                                    const char *name,
2996                                    const u16 wm[8])
2997 {
2998         int level, max_level = ilk_wm_max_level(dev_priv);
2999
3000         for (level = 0; level <= max_level; level++) {
3001                 unsigned int latency = wm[level];
3002
3003                 if (latency == 0) {
3004                         drm_dbg_kms(&dev_priv->drm,
3005                                     "%s WM%d latency not provided\n",
3006                                     name, level);
3007                         continue;
3008                 }
3009
3010                 /*
3011                  * - latencies are in us on gen9.
3012                  * - before then, WM1+ latency values are in 0.5us units
3013                  */
3014                 if (INTEL_GEN(dev_priv) >= 9)
3015                         latency *= 10;
3016                 else if (level > 0)
3017                         latency *= 5;
3018
3019                 drm_dbg_kms(&dev_priv->drm,
3020                             "%s WM%d latency %u (%u.%u usec)\n", name, level,
3021                             wm[level], latency / 10, latency % 10);
3022         }
3023 }
3024
3025 static bool ilk_increase_wm_latency(struct drm_i915_private *dev_priv,
3026                                     u16 wm[5], u16 min)
3027 {
3028         int level, max_level = ilk_wm_max_level(dev_priv);
3029
3030         if (wm[0] >= min)
3031                 return false;
3032
3033         wm[0] = max(wm[0], min);
3034         for (level = 1; level <= max_level; level++)
3035                 wm[level] = max_t(u16, wm[level], DIV_ROUND_UP(min, 5));
3036
3037         return true;
3038 }
3039
3040 static void snb_wm_latency_quirk(struct drm_i915_private *dev_priv)
3041 {
3042         bool changed;
3043
3044         /*
3045          * The BIOS provided WM memory latency values are often
3046          * inadequate for high resolution displays. Adjust them.
3047          */
3048         changed = ilk_increase_wm_latency(dev_priv, dev_priv->wm.pri_latency, 12) |
3049                 ilk_increase_wm_latency(dev_priv, dev_priv->wm.spr_latency, 12) |
3050                 ilk_increase_wm_latency(dev_priv, dev_priv->wm.cur_latency, 12);
3051
3052         if (!changed)
3053                 return;
3054
3055         drm_dbg_kms(&dev_priv->drm,
3056                     "WM latency values increased to avoid potential underruns\n");
3057         intel_print_wm_latency(dev_priv, "Primary", dev_priv->wm.pri_latency);
3058         intel_print_wm_latency(dev_priv, "Sprite", dev_priv->wm.spr_latency);
3059         intel_print_wm_latency(dev_priv, "Cursor", dev_priv->wm.cur_latency);
3060 }
3061
3062 static void snb_wm_lp3_irq_quirk(struct drm_i915_private *dev_priv)
3063 {
3064         /*
3065          * On some SNB machines (Thinkpad X220 Tablet at least)
3066          * LP3 usage can cause vblank interrupts to be lost.
3067          * The DEIIR bit will go high but it looks like the CPU
3068          * never gets interrupted.
3069          *
3070          * It's not clear whether other interrupt source could
3071          * be affected or if this is somehow limited to vblank
3072          * interrupts only. To play it safe we disable LP3
3073          * watermarks entirely.
3074          */
3075         if (dev_priv->wm.pri_latency[3] == 0 &&
3076             dev_priv->wm.spr_latency[3] == 0 &&
3077             dev_priv->wm.cur_latency[3] == 0)
3078                 return;
3079
3080         dev_priv->wm.pri_latency[3] = 0;
3081         dev_priv->wm.spr_latency[3] = 0;
3082         dev_priv->wm.cur_latency[3] = 0;
3083
3084         drm_dbg_kms(&dev_priv->drm,
3085                     "LP3 watermarks disabled due to potential for lost interrupts\n");
3086         intel_print_wm_latency(dev_priv, "Primary", dev_priv->wm.pri_latency);
3087         intel_print_wm_latency(dev_priv, "Sprite", dev_priv->wm.spr_latency);
3088         intel_print_wm_latency(dev_priv, "Cursor", dev_priv->wm.cur_latency);
3089 }
3090
3091 static void ilk_setup_wm_latency(struct drm_i915_private *dev_priv)
3092 {
3093         intel_read_wm_latency(dev_priv, dev_priv->wm.pri_latency);
3094
3095         memcpy(dev_priv->wm.spr_latency, dev_priv->wm.pri_latency,
3096                sizeof(dev_priv->wm.pri_latency));
3097         memcpy(dev_priv->wm.cur_latency, dev_priv->wm.pri_latency,
3098                sizeof(dev_priv->wm.pri_latency));
3099
3100         intel_fixup_spr_wm_latency(dev_priv, dev_priv->wm.spr_latency);
3101         intel_fixup_cur_wm_latency(dev_priv, dev_priv->wm.cur_latency);
3102
3103         intel_print_wm_latency(dev_priv, "Primary", dev_priv->wm.pri_latency);
3104         intel_print_wm_latency(dev_priv, "Sprite", dev_priv->wm.spr_latency);
3105         intel_print_wm_latency(dev_priv, "Cursor", dev_priv->wm.cur_latency);
3106
3107         if (IS_GEN(dev_priv, 6)) {
3108                 snb_wm_latency_quirk(dev_priv);
3109                 snb_wm_lp3_irq_quirk(dev_priv);
3110         }
3111 }
3112
3113 static void skl_setup_wm_latency(struct drm_i915_private *dev_priv)
3114 {
3115         intel_read_wm_latency(dev_priv, dev_priv->wm.skl_latency);
3116         intel_print_wm_latency(dev_priv, "Gen9 Plane", dev_priv->wm.skl_latency);
3117 }
3118
3119 static bool ilk_validate_pipe_wm(const struct drm_i915_private *dev_priv,
3120                                  struct intel_pipe_wm *pipe_wm)
3121 {
3122         /* LP0 watermark maximums depend on this pipe alone */
3123         const struct intel_wm_config config = {
3124                 .num_pipes_active = 1,
3125                 .sprites_enabled = pipe_wm->sprites_enabled,
3126                 .sprites_scaled = pipe_wm->sprites_scaled,
3127         };
3128         struct ilk_wm_maximums max;
3129
3130         /* LP0 watermarks always use 1/2 DDB partitioning */
3131         ilk_compute_wm_maximums(dev_priv, 0, &config, INTEL_DDB_PART_1_2, &max);
3132
3133         /* At least LP0 must be valid */
3134         if (!ilk_validate_wm_level(0, &max, &pipe_wm->wm[0])) {
3135                 drm_dbg_kms(&dev_priv->drm, "LP0 watermark invalid\n");
3136                 return false;
3137         }
3138
3139         return true;
3140 }
3141
3142 /* Compute new watermarks for the pipe */
3143 static int ilk_compute_pipe_wm(struct intel_crtc_state *crtc_state)
3144 {
3145         struct drm_i915_private *dev_priv = to_i915(crtc_state->uapi.crtc->dev);
3146         struct intel_crtc *crtc = to_intel_crtc(crtc_state->uapi.crtc);
3147         struct intel_pipe_wm *pipe_wm;
3148         struct intel_plane *plane;
3149         const struct intel_plane_state *plane_state;
3150         const struct intel_plane_state *pristate = NULL;
3151         const struct intel_plane_state *sprstate = NULL;
3152         const struct intel_plane_state *curstate = NULL;
3153         int level, max_level = ilk_wm_max_level(dev_priv), usable_level;
3154         struct ilk_wm_maximums max;
3155
3156         pipe_wm = &crtc_state->wm.ilk.optimal;
3157
3158         intel_atomic_crtc_state_for_each_plane_state(plane, plane_state, crtc_state) {
3159                 if (plane->base.type == DRM_PLANE_TYPE_PRIMARY)
3160                         pristate = plane_state;
3161                 else if (plane->base.type == DRM_PLANE_TYPE_OVERLAY)
3162                         sprstate = plane_state;
3163                 else if (plane->base.type == DRM_PLANE_TYPE_CURSOR)
3164                         curstate = plane_state;
3165         }
3166
3167         pipe_wm->pipe_enabled = crtc_state->hw.active;
3168         if (sprstate) {
3169                 pipe_wm->sprites_enabled = sprstate->uapi.visible;
3170                 pipe_wm->sprites_scaled = sprstate->uapi.visible &&
3171                         (drm_rect_width(&sprstate->uapi.dst) != drm_rect_width(&sprstate->uapi.src) >> 16 ||
3172                          drm_rect_height(&sprstate->uapi.dst) != drm_rect_height(&sprstate->uapi.src) >> 16);
3173         }
3174
3175         usable_level = max_level;
3176
3177         /* ILK/SNB: LP2+ watermarks only w/o sprites */
3178         if (INTEL_GEN(dev_priv) <= 6 && pipe_wm->sprites_enabled)
3179                 usable_level = 1;
3180
3181         /* ILK/SNB/IVB: LP1+ watermarks only w/o scaling */
3182         if (pipe_wm->sprites_scaled)
3183                 usable_level = 0;
3184
3185         memset(&pipe_wm->wm, 0, sizeof(pipe_wm->wm));
3186         ilk_compute_wm_level(dev_priv, crtc, 0, crtc_state,
3187                              pristate, sprstate, curstate, &pipe_wm->wm[0]);
3188
3189         if (!ilk_validate_pipe_wm(dev_priv, pipe_wm))
3190                 return -EINVAL;
3191
3192         ilk_compute_wm_reg_maximums(dev_priv, 1, &max);
3193
3194         for (level = 1; level <= usable_level; level++) {
3195                 struct intel_wm_level *wm = &pipe_wm->wm[level];
3196
3197                 ilk_compute_wm_level(dev_priv, crtc, level, crtc_state,
3198                                      pristate, sprstate, curstate, wm);
3199
3200                 /*
3201                  * Disable any watermark level that exceeds the
3202                  * register maximums since such watermarks are
3203                  * always invalid.
3204                  */
3205                 if (!ilk_validate_wm_level(level, &max, wm)) {
3206                         memset(wm, 0, sizeof(*wm));
3207                         break;
3208                 }
3209         }
3210
3211         return 0;
3212 }
3213
3214 /*
3215  * Build a set of 'intermediate' watermark values that satisfy both the old
3216  * state and the new state.  These can be programmed to the hardware
3217  * immediately.
3218  */
3219 static int ilk_compute_intermediate_wm(struct intel_crtc_state *newstate)
3220 {
3221         struct intel_crtc *intel_crtc = to_intel_crtc(newstate->uapi.crtc);
3222         struct drm_i915_private *dev_priv = to_i915(intel_crtc->base.dev);
3223         struct intel_pipe_wm *a = &newstate->wm.ilk.intermediate;
3224         struct intel_atomic_state *intel_state =
3225                 to_intel_atomic_state(newstate->uapi.state);
3226         const struct intel_crtc_state *oldstate =
3227                 intel_atomic_get_old_crtc_state(intel_state, intel_crtc);
3228         const struct intel_pipe_wm *b = &oldstate->wm.ilk.optimal;
3229         int level, max_level = ilk_wm_max_level(dev_priv);
3230
3231         /*
3232          * Start with the final, target watermarks, then combine with the
3233          * currently active watermarks to get values that are safe both before
3234          * and after the vblank.
3235          */
3236         *a = newstate->wm.ilk.optimal;
3237         if (!newstate->hw.active || drm_atomic_crtc_needs_modeset(&newstate->uapi) ||
3238             intel_state->skip_intermediate_wm)
3239                 return 0;
3240
3241         a->pipe_enabled |= b->pipe_enabled;
3242         a->sprites_enabled |= b->sprites_enabled;
3243         a->sprites_scaled |= b->sprites_scaled;
3244
3245         for (level = 0; level <= max_level; level++) {
3246                 struct intel_wm_level *a_wm = &a->wm[level];
3247                 const struct intel_wm_level *b_wm = &b->wm[level];
3248
3249                 a_wm->enable &= b_wm->enable;
3250                 a_wm->pri_val = max(a_wm->pri_val, b_wm->pri_val);
3251                 a_wm->spr_val = max(a_wm->spr_val, b_wm->spr_val);
3252                 a_wm->cur_val = max(a_wm->cur_val, b_wm->cur_val);
3253                 a_wm->fbc_val = max(a_wm->fbc_val, b_wm->fbc_val);
3254         }
3255
3256         /*
3257          * We need to make sure that these merged watermark values are
3258          * actually a valid configuration themselves.  If they're not,
3259          * there's no safe way to transition from the old state to
3260          * the new state, so we need to fail the atomic transaction.
3261          */
3262         if (!ilk_validate_pipe_wm(dev_priv, a))
3263                 return -EINVAL;
3264
3265         /*
3266          * If our intermediate WM are identical to the final WM, then we can
3267          * omit the post-vblank programming; only update if it's different.
3268          */
3269         if (memcmp(a, &newstate->wm.ilk.optimal, sizeof(*a)) != 0)
3270                 newstate->wm.need_postvbl_update = true;
3271
3272         return 0;
3273 }
3274
3275 /*
3276  * Merge the watermarks from all active pipes for a specific level.
3277  */
3278 static void ilk_merge_wm_level(struct drm_i915_private *dev_priv,
3279                                int level,
3280                                struct intel_wm_level *ret_wm)
3281 {
3282         const struct intel_crtc *intel_crtc;
3283
3284         ret_wm->enable = true;
3285
3286         for_each_intel_crtc(&dev_priv->drm, intel_crtc) {
3287                 const struct intel_pipe_wm *active = &intel_crtc->wm.active.ilk;
3288                 const struct intel_wm_level *wm = &active->wm[level];
3289
3290                 if (!active->pipe_enabled)
3291                         continue;
3292
3293                 /*
3294                  * The watermark values may have been used in the past,
3295                  * so we must maintain them in the registers for some
3296                  * time even if the level is now disabled.
3297                  */
3298                 if (!wm->enable)
3299                         ret_wm->enable = false;
3300
3301                 ret_wm->pri_val = max(ret_wm->pri_val, wm->pri_val);
3302                 ret_wm->spr_val = max(ret_wm->spr_val, wm->spr_val);
3303                 ret_wm->cur_val = max(ret_wm->cur_val, wm->cur_val);
3304                 ret_wm->fbc_val = max(ret_wm->fbc_val, wm->fbc_val);
3305         }
3306 }
3307
3308 /*
3309  * Merge all low power watermarks for all active pipes.
3310  */
3311 static void ilk_wm_merge(struct drm_i915_private *dev_priv,
3312                          const struct intel_wm_config *config,
3313                          const struct ilk_wm_maximums *max,
3314                          struct intel_pipe_wm *merged)
3315 {
3316         int level, max_level = ilk_wm_max_level(dev_priv);
3317         int last_enabled_level = max_level;
3318
3319         /* ILK/SNB/IVB: LP1+ watermarks only w/ single pipe */
3320         if ((INTEL_GEN(dev_priv) <= 6 || IS_IVYBRIDGE(dev_priv)) &&
3321             config->num_pipes_active > 1)
3322                 last_enabled_level = 0;
3323
3324         /* ILK: FBC WM must be disabled always */
3325         merged->fbc_wm_enabled = INTEL_GEN(dev_priv) >= 6;
3326
3327         /* merge each WM1+ level */
3328         for (level = 1; level <= max_level; level++) {
3329                 struct intel_wm_level *wm = &merged->wm[level];
3330
3331                 ilk_merge_wm_level(dev_priv, level, wm);
3332
3333                 if (level > last_enabled_level)
3334                         wm->enable = false;
3335                 else if (!ilk_validate_wm_level(level, max, wm))
3336                         /* make sure all following levels get disabled */
3337                         last_enabled_level = level - 1;
3338
3339                 /*
3340                  * The spec says it is preferred to disable
3341                  * FBC WMs instead of disabling a WM level.
3342                  */
3343                 if (wm->fbc_val > max->fbc) {
3344                         if (wm->enable)
3345                                 merged->fbc_wm_enabled = false;
3346                         wm->fbc_val = 0;
3347                 }
3348         }
3349
3350         /* ILK: LP2+ must be disabled when FBC WM is disabled but FBC enabled */
3351         /*
3352          * FIXME this is racy. FBC might get enabled later.
3353          * What we should check here is whether FBC can be
3354          * enabled sometime later.
3355          */
3356         if (IS_GEN(dev_priv, 5) && !merged->fbc_wm_enabled &&
3357             intel_fbc_is_active(dev_priv)) {
3358                 for (level = 2; level <= max_level; level++) {
3359                         struct intel_wm_level *wm = &merged->wm[level];
3360
3361                         wm->enable = false;
3362                 }
3363         }
3364 }
3365
3366 static int ilk_wm_lp_to_level(int wm_lp, const struct intel_pipe_wm *pipe_wm)
3367 {
3368         /* LP1,LP2,LP3 levels are either 1,2,3 or 1,3,4 */
3369         return wm_lp + (wm_lp >= 2 && pipe_wm->wm[4].enable);
3370 }
3371
3372 /* The value we need to program into the WM_LPx latency field */
3373 static unsigned int ilk_wm_lp_latency(struct drm_i915_private *dev_priv,
3374                                       int level)
3375 {
3376         if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
3377                 return 2 * level;
3378         else
3379                 return dev_priv->wm.pri_latency[level];
3380 }
3381
3382 static void ilk_compute_wm_results(struct drm_i915_private *dev_priv,
3383                                    const struct intel_pipe_wm *merged,
3384                                    enum intel_ddb_partitioning partitioning,
3385                                    struct ilk_wm_values *results)
3386 {
3387         struct intel_crtc *intel_crtc;
3388         int level, wm_lp;
3389
3390         results->enable_fbc_wm = merged->fbc_wm_enabled;
3391         results->partitioning = partitioning;
3392
3393         /* LP1+ register values */
3394         for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
3395                 const struct intel_wm_level *r;
3396
3397                 level = ilk_wm_lp_to_level(wm_lp, merged);
3398
3399                 r = &merged->wm[level];
3400
3401                 /*
3402                  * Maintain the watermark values even if the level is
3403                  * disabled. Doing otherwise could cause underruns.
3404                  */
3405                 results->wm_lp[wm_lp - 1] =
3406                         (ilk_wm_lp_latency(dev_priv, level) << WM1_LP_LATENCY_SHIFT) |
3407                         (r->pri_val << WM1_LP_SR_SHIFT) |
3408                         r->cur_val;
3409
3410                 if (r->enable)
3411                         results->wm_lp[wm_lp - 1] |= WM1_LP_SR_EN;
3412
3413                 if (INTEL_GEN(dev_priv) >= 8)
3414                         results->wm_lp[wm_lp - 1] |=
3415                                 r->fbc_val << WM1_LP_FBC_SHIFT_BDW;
3416                 else
3417                         results->wm_lp[wm_lp - 1] |=
3418                                 r->fbc_val << WM1_LP_FBC_SHIFT;
3419
3420                 /*
3421                  * Always set WM1S_LP_EN when spr_val != 0, even if the
3422                  * level is disabled. Doing otherwise could cause underruns.
3423                  */
3424                 if (INTEL_GEN(dev_priv) <= 6 && r->spr_val) {
3425                         drm_WARN_ON(&dev_priv->drm, wm_lp != 1);
3426                         results->wm_lp_spr[wm_lp - 1] = WM1S_LP_EN | r->spr_val;
3427                 } else
3428                         results->wm_lp_spr[wm_lp - 1] = r->spr_val;
3429         }
3430
3431         /* LP0 register values */
3432         for_each_intel_crtc(&dev_priv->drm, intel_crtc) {
3433                 enum pipe pipe = intel_crtc->pipe;
3434                 const struct intel_pipe_wm *pipe_wm = &intel_crtc->wm.active.ilk;
3435                 const struct intel_wm_level *r = &pipe_wm->wm[0];
3436
3437                 if (drm_WARN_ON(&dev_priv->drm, !r->enable))
3438                         continue;
3439
3440                 results->wm_pipe[pipe] =
3441                         (r->pri_val << WM0_PIPE_PLANE_SHIFT) |
3442                         (r->spr_val << WM0_PIPE_SPRITE_SHIFT) |
3443                         r->cur_val;
3444         }
3445 }
3446
3447 /* Find the result with the highest level enabled. Check for enable_fbc_wm in
3448  * case both are at the same level. Prefer r1 in case they're the same. */
3449 static struct intel_pipe_wm *
3450 ilk_find_best_result(struct drm_i915_private *dev_priv,
3451                      struct intel_pipe_wm *r1,
3452                      struct intel_pipe_wm *r2)
3453 {
3454         int level, max_level = ilk_wm_max_level(dev_priv);
3455         int level1 = 0, level2 = 0;
3456
3457         for (level = 1; level <= max_level; level++) {
3458                 if (r1->wm[level].enable)
3459                         level1 = level;
3460                 if (r2->wm[level].enable)
3461                         level2 = level;
3462         }
3463
3464         if (level1 == level2) {
3465                 if (r2->fbc_wm_enabled && !r1->fbc_wm_enabled)
3466                         return r2;
3467                 else
3468                         return r1;
3469         } else if (level1 > level2) {
3470                 return r1;
3471         } else {
3472                 return r2;
3473         }
3474 }
3475
3476 /* dirty bits used to track which watermarks need changes */
3477 #define WM_DIRTY_PIPE(pipe) (1 << (pipe))
3478 #define WM_DIRTY_LP(wm_lp) (1 << (15 + (wm_lp)))
3479 #define WM_DIRTY_LP_ALL (WM_DIRTY_LP(1) | WM_DIRTY_LP(2) | WM_DIRTY_LP(3))
3480 #define WM_DIRTY_FBC (1 << 24)
3481 #define WM_DIRTY_DDB (1 << 25)
3482
3483 static unsigned int ilk_compute_wm_dirty(struct drm_i915_private *dev_priv,
3484                                          const struct ilk_wm_values *old,
3485                                          const struct ilk_wm_values *new)
3486 {
3487         unsigned int dirty = 0;
3488         enum pipe pipe;
3489         int wm_lp;
3490
3491         for_each_pipe(dev_priv, pipe) {
3492                 if (old->wm_pipe[pipe] != new->wm_pipe[pipe]) {
3493                         dirty |= WM_DIRTY_PIPE(pipe);
3494                         /* Must disable LP1+ watermarks too */
3495                         dirty |= WM_DIRTY_LP_ALL;
3496                 }
3497         }
3498
3499         if (old->enable_fbc_wm != new->enable_fbc_wm) {
3500                 dirty |= WM_DIRTY_FBC;
3501                 /* Must disable LP1+ watermarks too */
3502                 dirty |= WM_DIRTY_LP_ALL;
3503         }
3504
3505         if (old->partitioning != new->partitioning) {
3506                 dirty |= WM_DIRTY_DDB;
3507                 /* Must disable LP1+ watermarks too */
3508                 dirty |= WM_DIRTY_LP_ALL;
3509         }
3510
3511         /* LP1+ watermarks already deemed dirty, no need to continue */
3512         if (dirty & WM_DIRTY_LP_ALL)
3513                 return dirty;
3514
3515         /* Find the lowest numbered LP1+ watermark in need of an update... */
3516         for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
3517                 if (old->wm_lp[wm_lp - 1] != new->wm_lp[wm_lp - 1] ||
3518                     old->wm_lp_spr[wm_lp - 1] != new->wm_lp_spr[wm_lp - 1])
3519                         break;
3520         }
3521
3522         /* ...and mark it and all higher numbered LP1+ watermarks as dirty */
3523         for (; wm_lp <= 3; wm_lp++)
3524                 dirty |= WM_DIRTY_LP(wm_lp);
3525
3526         return dirty;
3527 }
3528
3529 static bool _ilk_disable_lp_wm(struct drm_i915_private *dev_priv,
3530                                unsigned int dirty)
3531 {
3532         struct ilk_wm_values *previous = &dev_priv->wm.hw;
3533         bool changed = false;
3534
3535         if (dirty & WM_DIRTY_LP(3) && previous->wm_lp[2] & WM1_LP_SR_EN) {
3536                 previous->wm_lp[2] &= ~WM1_LP_SR_EN;
3537                 intel_uncore_write(&dev_priv->uncore, WM3_LP_ILK, previous->wm_lp[2]);
3538                 changed = true;
3539         }
3540         if (dirty & WM_DIRTY_LP(2) && previous->wm_lp[1] & WM1_LP_SR_EN) {
3541                 previous->wm_lp[1] &= ~WM1_LP_SR_EN;
3542                 intel_uncore_write(&dev_priv->uncore, WM2_LP_ILK, previous->wm_lp[1]);
3543                 changed = true;
3544         }
3545         if (dirty & WM_DIRTY_LP(1) && previous->wm_lp[0] & WM1_LP_SR_EN) {
3546                 previous->wm_lp[0] &= ~WM1_LP_SR_EN;
3547                 intel_uncore_write(&dev_priv->uncore, WM1_LP_ILK, previous->wm_lp[0]);
3548                 changed = true;
3549         }
3550
3551         /*
3552          * Don't touch WM1S_LP_EN here.
3553          * Doing so could cause underruns.
3554          */
3555
3556         return changed;
3557 }
3558
3559 /*
3560  * The spec says we shouldn't write when we don't need, because every write
3561  * causes WMs to be re-evaluated, expending some power.
3562  */
3563 static void ilk_write_wm_values(struct drm_i915_private *dev_priv,
3564                                 struct ilk_wm_values *results)
3565 {
3566         struct ilk_wm_values *previous = &dev_priv->wm.hw;
3567         unsigned int dirty;
3568         u32 val;
3569
3570         dirty = ilk_compute_wm_dirty(dev_priv, previous, results);
3571         if (!dirty)
3572                 return;
3573
3574         _ilk_disable_lp_wm(dev_priv, dirty);
3575
3576         if (dirty & WM_DIRTY_PIPE(PIPE_A))
3577                 intel_uncore_write(&dev_priv->uncore, WM0_PIPE_ILK(PIPE_A), results->wm_pipe[0]);
3578         if (dirty & WM_DIRTY_PIPE(PIPE_B))
3579                 intel_uncore_write(&dev_priv->uncore, WM0_PIPE_ILK(PIPE_B), results->wm_pipe[1]);
3580         if (dirty & WM_DIRTY_PIPE(PIPE_C))
3581                 intel_uncore_write(&dev_priv->uncore, WM0_PIPE_ILK(PIPE_C), results->wm_pipe[2]);
3582
3583         if (dirty & WM_DIRTY_DDB) {
3584                 if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv)) {
3585                         val = intel_uncore_read(&dev_priv->uncore, WM_MISC);
3586                         if (results->partitioning == INTEL_DDB_PART_1_2)
3587                                 val &= ~WM_MISC_DATA_PARTITION_5_6;
3588                         else
3589                                 val |= WM_MISC_DATA_PARTITION_5_6;
3590                         intel_uncore_write(&dev_priv->uncore, WM_MISC, val);
3591                 } else {
3592                         val = intel_uncore_read(&dev_priv->uncore, DISP_ARB_CTL2);
3593                         if (results->partitioning == INTEL_DDB_PART_1_2)
3594                                 val &= ~DISP_DATA_PARTITION_5_6;
3595                         else
3596                                 val |= DISP_DATA_PARTITION_5_6;
3597                         intel_uncore_write(&dev_priv->uncore, DISP_ARB_CTL2, val);
3598                 }
3599         }
3600
3601         if (dirty & WM_DIRTY_FBC) {
3602                 val = intel_uncore_read(&dev_priv->uncore, DISP_ARB_CTL);
3603                 if (results->enable_fbc_wm)
3604                         val &= ~DISP_FBC_WM_DIS;
3605                 else
3606                         val |= DISP_FBC_WM_DIS;
3607                 intel_uncore_write(&dev_priv->uncore, DISP_ARB_CTL, val);
3608         }
3609
3610         if (dirty & WM_DIRTY_LP(1) &&
3611             previous->wm_lp_spr[0] != results->wm_lp_spr[0])
3612                 intel_uncore_write(&dev_priv->uncore, WM1S_LP_ILK, results->wm_lp_spr[0]);
3613
3614         if (INTEL_GEN(dev_priv) >= 7) {
3615                 if (dirty & WM_DIRTY_LP(2) && previous->wm_lp_spr[1] != results->wm_lp_spr[1])
3616                         intel_uncore_write(&dev_priv->uncore, WM2S_LP_IVB, results->wm_lp_spr[1]);
3617                 if (dirty & WM_DIRTY_LP(3) && previous->wm_lp_spr[2] != results->wm_lp_spr[2])
3618                         intel_uncore_write(&dev_priv->uncore, WM3S_LP_IVB, results->wm_lp_spr[2]);
3619         }
3620
3621         if (dirty & WM_DIRTY_LP(1) && previous->wm_lp[0] != results->wm_lp[0])
3622                 intel_uncore_write(&dev_priv->uncore, WM1_LP_ILK, results->wm_lp[0]);
3623         if (dirty & WM_DIRTY_LP(2) && previous->wm_lp[1] != results->wm_lp[1])
3624                 intel_uncore_write(&dev_priv->uncore, WM2_LP_ILK, results->wm_lp[1]);
3625         if (dirty & WM_DIRTY_LP(3) && previous->wm_lp[2] != results->wm_lp[2])
3626                 intel_uncore_write(&dev_priv->uncore, WM3_LP_ILK, results->wm_lp[2]);
3627
3628         dev_priv->wm.hw = *results;
3629 }
3630
3631 bool ilk_disable_lp_wm(struct drm_i915_private *dev_priv)
3632 {
3633         return _ilk_disable_lp_wm(dev_priv, WM_DIRTY_LP_ALL);
3634 }
3635
3636 u8 intel_enabled_dbuf_slices_mask(struct drm_i915_private *dev_priv)
3637 {
3638         int i;
3639         int max_slices = INTEL_INFO(dev_priv)->num_supported_dbuf_slices;
3640         u8 enabled_slices_mask = 0;
3641
3642         for (i = 0; i < max_slices; i++) {
3643                 if (intel_uncore_read(&dev_priv->uncore, DBUF_CTL_S(i)) & DBUF_POWER_STATE)
3644                         enabled_slices_mask |= BIT(i);
3645         }
3646
3647         return enabled_slices_mask;
3648 }
3649
3650 /*
3651  * FIXME: We still don't have the proper code detect if we need to apply the WA,
3652  * so assume we'll always need it in order to avoid underruns.
3653  */
3654 static bool skl_needs_memory_bw_wa(struct drm_i915_private *dev_priv)
3655 {
3656         return IS_GEN9_BC(dev_priv) || IS_BROXTON(dev_priv);
3657 }
3658
3659 static bool
3660 intel_has_sagv(struct drm_i915_private *dev_priv)
3661 {
3662         return (IS_GEN9_BC(dev_priv) || INTEL_GEN(dev_priv) >= 10) &&
3663                 dev_priv->sagv_status != I915_SAGV_NOT_CONTROLLED;
3664 }
3665
3666 static void
3667 skl_setup_sagv_block_time(struct drm_i915_private *dev_priv)
3668 {
3669         if (INTEL_GEN(dev_priv) >= 12) {
3670                 u32 val = 0;
3671                 int ret;
3672
3673                 ret = sandybridge_pcode_read(dev_priv,
3674                                              GEN12_PCODE_READ_SAGV_BLOCK_TIME_US,
3675                                              &val, NULL);
3676                 if (!ret) {
3677                         dev_priv->sagv_block_time_us = val;
3678                         return;
3679                 }
3680
3681                 drm_dbg(&dev_priv->drm, "Couldn't read SAGV block time!\n");
3682         } else if (IS_GEN(dev_priv, 11)) {
3683                 dev_priv->sagv_block_time_us = 10;
3684                 return;
3685         } else if (IS_GEN(dev_priv, 10)) {
3686                 dev_priv->sagv_block_time_us = 20;
3687                 return;
3688         } else if (IS_GEN(dev_priv, 9)) {
3689                 dev_priv->sagv_block_time_us = 30;
3690                 return;
3691         } else {
3692                 MISSING_CASE(INTEL_GEN(dev_priv));
3693         }
3694
3695         /* Default to an unusable block time */
3696         dev_priv->sagv_block_time_us = -1;
3697 }
3698
3699 /*
3700  * SAGV dynamically adjusts the system agent voltage and clock frequencies
3701  * depending on power and performance requirements. The display engine access
3702  * to system memory is blocked during the adjustment time. Because of the
3703  * blocking time, having this enabled can cause full system hangs and/or pipe
3704  * underruns if we don't meet all of the following requirements:
3705  *
3706  *  - <= 1 pipe enabled
3707  *  - All planes can enable watermarks for latencies >= SAGV engine block time
3708  *  - We're not using an interlaced display configuration
3709  */
3710 static int
3711 intel_enable_sagv(struct drm_i915_private *dev_priv)
3712 {
3713         int ret;
3714
3715         if (!intel_has_sagv(dev_priv))
3716                 return 0;
3717
3718         if (dev_priv->sagv_status == I915_SAGV_ENABLED)
3719                 return 0;
3720
3721         drm_dbg_kms(&dev_priv->drm, "Enabling SAGV\n");
3722         ret = sandybridge_pcode_write(dev_priv, GEN9_PCODE_SAGV_CONTROL,
3723                                       GEN9_SAGV_ENABLE);
3724
3725         /* We don't need to wait for SAGV when enabling */
3726
3727         /*
3728          * Some skl systems, pre-release machines in particular,
3729          * don't actually have SAGV.
3730          */
3731         if (IS_SKYLAKE(dev_priv) && ret == -ENXIO) {
3732                 drm_dbg(&dev_priv->drm, "No SAGV found on system, ignoring\n");
3733                 dev_priv->sagv_status = I915_SAGV_NOT_CONTROLLED;
3734                 return 0;
3735         } else if (ret < 0) {
3736                 drm_err(&dev_priv->drm, "Failed to enable SAGV\n");
3737                 return ret;
3738         }
3739
3740         dev_priv->sagv_status = I915_SAGV_ENABLED;
3741         return 0;
3742 }
3743
3744 static int
3745 intel_disable_sagv(struct drm_i915_private *dev_priv)
3746 {
3747         int ret;
3748
3749         if (!intel_has_sagv(dev_priv))
3750                 return 0;
3751
3752         if (dev_priv->sagv_status == I915_SAGV_DISABLED)
3753                 return 0;
3754
3755         drm_dbg_kms(&dev_priv->drm, "Disabling SAGV\n");
3756         /* bspec says to keep retrying for at least 1 ms */
3757         ret = skl_pcode_request(dev_priv, GEN9_PCODE_SAGV_CONTROL,
3758                                 GEN9_SAGV_DISABLE,
3759                                 GEN9_SAGV_IS_DISABLED, GEN9_SAGV_IS_DISABLED,
3760                                 1);
3761         /*
3762          * Some skl systems, pre-release machines in particular,
3763          * don't actually have SAGV.
3764          */
3765         if (IS_SKYLAKE(dev_priv) && ret == -ENXIO) {
3766                 drm_dbg(&dev_priv->drm, "No SAGV found on system, ignoring\n");
3767                 dev_priv->sagv_status = I915_SAGV_NOT_CONTROLLED;
3768                 return 0;
3769         } else if (ret < 0) {
3770                 drm_err(&dev_priv->drm, "Failed to disable SAGV (%d)\n", ret);
3771                 return ret;
3772         }
3773
3774         dev_priv->sagv_status = I915_SAGV_DISABLED;
3775         return 0;
3776 }
3777
3778 void intel_sagv_pre_plane_update(struct intel_atomic_state *state)
3779 {
3780         struct drm_i915_private *dev_priv = to_i915(state->base.dev);
3781         const struct intel_bw_state *new_bw_state;
3782         const struct intel_bw_state *old_bw_state;
3783         u32 new_mask = 0;
3784
3785         /*
3786          * Just return if we can't control SAGV or don't have it.
3787          * This is different from situation when we have SAGV but just can't
3788          * afford it due to DBuf limitation - in case if SAGV is completely
3789          * disabled in a BIOS, we are not even allowed to send a PCode request,
3790          * as it will throw an error. So have to check it here.
3791          */
3792         if (!intel_has_sagv(dev_priv))
3793                 return;
3794
3795         new_bw_state = intel_atomic_get_new_bw_state(state);
3796         if (!new_bw_state)
3797                 return;
3798
3799         if (INTEL_GEN(dev_priv) < 11 && !intel_can_enable_sagv(dev_priv, new_bw_state)) {
3800                 intel_disable_sagv(dev_priv);
3801                 return;
3802         }
3803
3804         old_bw_state = intel_atomic_get_old_bw_state(state);
3805         /*
3806          * Nothing to mask
3807          */
3808         if (new_bw_state->qgv_points_mask == old_bw_state->qgv_points_mask)
3809                 return;
3810
3811         new_mask = old_bw_state->qgv_points_mask | new_bw_state->qgv_points_mask;
3812
3813         /*
3814          * If new mask is zero - means there is nothing to mask,
3815          * we can only unmask, which should be done in unmask.
3816          */
3817         if (!new_mask)
3818                 return;
3819
3820         /*
3821          * Restrict required qgv points before updating the configuration.
3822          * According to BSpec we can't mask and unmask qgv points at the same
3823          * time. Also masking should be done before updating the configuration
3824          * and unmasking afterwards.
3825          */
3826         icl_pcode_restrict_qgv_points(dev_priv, new_mask);
3827 }
3828
3829 void intel_sagv_post_plane_update(struct intel_atomic_state *state)
3830 {
3831         struct drm_i915_private *dev_priv = to_i915(state->base.dev);
3832         const struct intel_bw_state *new_bw_state;
3833         const struct intel_bw_state *old_bw_state;
3834         u32 new_mask = 0;
3835
3836         /*
3837          * Just return if we can't control SAGV or don't have it.
3838          * This is different from situation when we have SAGV but just can't
3839          * afford it due to DBuf limitation - in case if SAGV is completely
3840          * disabled in a BIOS, we are not even allowed to send a PCode request,
3841          * as it will throw an error. So have to check it here.
3842          */
3843         if (!intel_has_sagv(dev_priv))
3844                 return;
3845
3846         new_bw_state = intel_atomic_get_new_bw_state(state);
3847         if (!new_bw_state)
3848                 return;
3849
3850         if (INTEL_GEN(dev_priv) < 11 && intel_can_enable_sagv(dev_priv, new_bw_state)) {
3851                 intel_enable_sagv(dev_priv);
3852                 return;
3853         }
3854
3855         old_bw_state = intel_atomic_get_old_bw_state(state);
3856         /*
3857          * Nothing to unmask
3858          */
3859         if (new_bw_state->qgv_points_mask == old_bw_state->qgv_points_mask)
3860                 return;
3861
3862         new_mask = new_bw_state->qgv_points_mask;
3863
3864         /*
3865          * Allow required qgv points after updating the configuration.
3866          * According to BSpec we can't mask and unmask qgv points at the same
3867          * time. Also masking should be done before updating the configuration
3868          * and unmasking afterwards.
3869          */
3870         icl_pcode_restrict_qgv_points(dev_priv, new_mask);
3871 }
3872
3873 static bool skl_crtc_can_enable_sagv(const struct intel_crtc_state *crtc_state)
3874 {
3875         struct intel_crtc *crtc = to_intel_crtc(crtc_state->uapi.crtc);
3876         struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
3877         enum plane_id plane_id;
3878
3879         if (!intel_has_sagv(dev_priv))
3880                 return false;
3881
3882         if (!crtc_state->hw.active)
3883                 return true;
3884
3885         if (crtc_state->hw.pipe_mode.flags & DRM_MODE_FLAG_INTERLACE)
3886                 return false;
3887
3888         for_each_plane_id_on_crtc(crtc, plane_id) {
3889                 const struct skl_plane_wm *wm =
3890                         &crtc_state->wm.skl.optimal.planes[plane_id];
3891                 int level;
3892
3893                 /* Skip this plane if it's not enabled */
3894                 if (!wm->wm[0].plane_en)
3895                         continue;
3896
3897                 /* Find the highest enabled wm level for this plane */
3898                 for (level = ilk_wm_max_level(dev_priv);
3899                      !wm->wm[level].plane_en; --level)
3900                      { }
3901
3902                 /*
3903                  * If any of the planes on this pipe don't enable wm levels that
3904                  * incur memory latencies higher than sagv_block_time_us we
3905                  * can't enable SAGV.
3906                  */
3907                 if (!wm->wm[level].can_sagv)
3908                         return false;
3909         }
3910
3911         return true;
3912 }
3913
3914 static bool tgl_crtc_can_enable_sagv(const struct intel_crtc_state *crtc_state)
3915 {
3916         struct intel_crtc *crtc = to_intel_crtc(crtc_state->uapi.crtc);
3917         enum plane_id plane_id;
3918
3919         if (!crtc_state->hw.active)
3920                 return true;
3921
3922         for_each_plane_id_on_crtc(crtc, plane_id) {
3923                 const struct skl_ddb_entry *plane_alloc =
3924                         &crtc_state->wm.skl.plane_ddb_y[plane_id];
3925                 const struct skl_plane_wm *wm =
3926                         &crtc_state->wm.skl.optimal.planes[plane_id];
3927
3928                 if (skl_ddb_entry_size(plane_alloc) < wm->sagv_wm0.min_ddb_alloc)
3929                         return false;
3930         }
3931
3932         return true;
3933 }
3934
3935 static bool intel_crtc_can_enable_sagv(const struct intel_crtc_state *crtc_state)
3936 {
3937         struct intel_crtc *crtc = to_intel_crtc(crtc_state->uapi.crtc);
3938         struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
3939
3940         if (INTEL_GEN(dev_priv) >= 12)
3941                 return tgl_crtc_can_enable_sagv(crtc_state);
3942         else
3943                 return skl_crtc_can_enable_sagv(crtc_state);
3944 }
3945
3946 bool intel_can_enable_sagv(struct drm_i915_private *dev_priv,
3947                            const struct intel_bw_state *bw_state)
3948 {
3949         if (INTEL_GEN(dev_priv) < 11 &&
3950             bw_state->active_pipes && !is_power_of_2(bw_state->active_pipes))
3951                 return false;
3952
3953         return bw_state->pipe_sagv_reject == 0;
3954 }
3955
3956 static int intel_compute_sagv_mask(struct intel_atomic_state *state)
3957 {
3958         struct drm_i915_private *dev_priv = to_i915(state->base.dev);
3959         int ret;
3960         struct intel_crtc *crtc;
3961         struct intel_crtc_state *new_crtc_state;
3962         struct intel_bw_state *new_bw_state = NULL;
3963         const struct intel_bw_state *old_bw_state = NULL;
3964         int i;
3965
3966         for_each_new_intel_crtc_in_state(state, crtc,
3967                                          new_crtc_state, i) {
3968                 new_bw_state = intel_atomic_get_bw_state(state);
3969                 if (IS_ERR(new_bw_state))
3970                         return PTR_ERR(new_bw_state);
3971
3972                 old_bw_state = intel_atomic_get_old_bw_state(state);
3973
3974                 if (intel_crtc_can_enable_sagv(new_crtc_state))
3975                         new_bw_state->pipe_sagv_reject &= ~BIT(crtc->pipe);
3976                 else
3977                         new_bw_state->pipe_sagv_reject |= BIT(crtc->pipe);
3978         }
3979
3980         if (!new_bw_state)
3981                 return 0;
3982
3983         new_bw_state->active_pipes =
3984                 intel_calc_active_pipes(state, old_bw_state->active_pipes);
3985
3986         if (new_bw_state->active_pipes != old_bw_state->active_pipes) {
3987                 ret = intel_atomic_lock_global_state(&new_bw_state->base);
3988                 if (ret)
3989                         return ret;
3990         }
3991
3992         for_each_new_intel_crtc_in_state(state, crtc,
3993                                          new_crtc_state, i) {
3994                 struct skl_pipe_wm *pipe_wm = &new_crtc_state->wm.skl.optimal;
3995
3996                 /*
3997                  * We store use_sagv_wm in the crtc state rather than relying on
3998                  * that bw state since we have no convenient way to get at the
3999                  * latter from the plane commit hooks (especially in the legacy
4000                  * cursor case)
4001                  */
4002                 pipe_wm->use_sagv_wm = INTEL_GEN(dev_priv) >= 12 &&
4003                                        intel_can_enable_sagv(dev_priv, new_bw_state);
4004         }
4005
4006         if (intel_can_enable_sagv(dev_priv, new_bw_state) !=
4007             intel_can_enable_sagv(dev_priv, old_bw_state)) {
4008                 ret = intel_atomic_serialize_global_state(&new_bw_state->base);
4009                 if (ret)
4010                         return ret;
4011         } else if (new_bw_state->pipe_sagv_reject != old_bw_state->pipe_sagv_reject) {
4012                 ret = intel_atomic_lock_global_state(&new_bw_state->base);
4013                 if (ret)
4014                         return ret;
4015         }
4016
4017         return 0;
4018 }
4019
4020 static int intel_dbuf_size(struct drm_i915_private *dev_priv)
4021 {
4022         int ddb_size = INTEL_INFO(dev_priv)->ddb_size;
4023
4024         drm_WARN_ON(&dev_priv->drm, ddb_size == 0);
4025
4026         if (INTEL_GEN(dev_priv) < 11)
4027                 return ddb_size - 4; /* 4 blocks for bypass path allocation */
4028
4029         return ddb_size;
4030 }
4031
4032 static int intel_dbuf_slice_size(struct drm_i915_private *dev_priv)
4033 {
4034         return intel_dbuf_size(dev_priv) /
4035                 INTEL_INFO(dev_priv)->num_supported_dbuf_slices;
4036 }
4037
4038 static void
4039 skl_ddb_entry_for_slices(struct drm_i915_private *dev_priv, u8 slice_mask,
4040                          struct skl_ddb_entry *ddb)
4041 {
4042         int slice_size = intel_dbuf_slice_size(dev_priv);
4043
4044         if (!slice_mask) {
4045                 ddb->start = 0;
4046                 ddb->end = 0;
4047                 return;
4048         }
4049
4050         ddb->start = (ffs(slice_mask) - 1) * slice_size;
4051         ddb->end = fls(slice_mask) * slice_size;
4052
4053         WARN_ON(ddb->start >= ddb->end);
4054         WARN_ON(ddb->end > intel_dbuf_size(dev_priv));
4055 }
4056
4057 u32 skl_ddb_dbuf_slice_mask(struct drm_i915_private *dev_priv,
4058                             const struct skl_ddb_entry *entry)
4059 {
4060         u32 slice_mask = 0;
4061         u16 ddb_size = intel_dbuf_size(dev_priv);
4062         u16 num_supported_slices = INTEL_INFO(dev_priv)->num_supported_dbuf_slices;
4063         u16 slice_size = ddb_size / num_supported_slices;
4064         u16 start_slice;
4065         u16 end_slice;
4066
4067         if (!skl_ddb_entry_size(entry))
4068                 return 0;
4069
4070         start_slice = entry->start / slice_size;
4071         end_slice = (entry->end - 1) / slice_size;
4072
4073         /*
4074          * Per plane DDB entry can in a really worst case be on multiple slices
4075          * but single entry is anyway contigious.
4076          */
4077         while (start_slice <= end_slice) {
4078                 slice_mask |= BIT(start_slice);
4079                 start_slice++;
4080         }
4081
4082         return slice_mask;
4083 }
4084
4085 static unsigned int intel_crtc_ddb_weight(const struct intel_crtc_state *crtc_state)
4086 {
4087         const struct drm_display_mode *pipe_mode = &crtc_state->hw.pipe_mode;
4088         int hdisplay, vdisplay;
4089
4090         if (!crtc_state->hw.active)
4091                 return 0;
4092
4093         /*
4094          * Watermark/ddb requirement highly depends upon width of the
4095          * framebuffer, So instead of allocating DDB equally among pipes
4096          * distribute DDB based on resolution/width of the display.
4097          */
4098         drm_mode_get_hv_timing(pipe_mode, &hdisplay, &vdisplay);
4099
4100         return hdisplay;
4101 }
4102
4103 static void intel_crtc_dbuf_weights(const struct intel_dbuf_state *dbuf_state,
4104                                     enum pipe for_pipe,
4105                                     unsigned int *weight_start,
4106                                     unsigned int *weight_end,
4107                                     unsigned int *weight_total)
4108 {
4109         struct drm_i915_private *dev_priv =
4110                 to_i915(dbuf_state->base.state->base.dev);
4111         enum pipe pipe;
4112
4113         *weight_start = 0;
4114         *weight_end = 0;
4115         *weight_total = 0;
4116
4117         for_each_pipe(dev_priv, pipe) {
4118                 int weight = dbuf_state->weight[pipe];
4119
4120                 /*
4121                  * Do not account pipes using other slice sets
4122                  * luckily as of current BSpec slice sets do not partially
4123                  * intersect(pipes share either same one slice or same slice set
4124                  * i.e no partial intersection), so it is enough to check for
4125                  * equality for now.
4126                  */
4127                 if (dbuf_state->slices[pipe] != dbuf_state->slices[for_pipe])
4128                         continue;
4129
4130                 *weight_total += weight;
4131                 if (pipe < for_pipe) {
4132                         *weight_start += weight;
4133                         *weight_end += weight;
4134                 } else if (pipe == for_pipe) {
4135                         *weight_end += weight;
4136                 }
4137         }
4138 }
4139
4140 static int
4141 skl_crtc_allocate_ddb(struct intel_atomic_state *state, struct intel_crtc *crtc)
4142 {
4143         struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
4144         unsigned int weight_total, weight_start, weight_end;
4145         const struct intel_dbuf_state *old_dbuf_state =
4146                 intel_atomic_get_old_dbuf_state(state);
4147         struct intel_dbuf_state *new_dbuf_state =
4148                 intel_atomic_get_new_dbuf_state(state);
4149         struct intel_crtc_state *crtc_state;
4150         struct skl_ddb_entry ddb_slices;
4151         enum pipe pipe = crtc->pipe;
4152         u32 ddb_range_size;
4153         u32 dbuf_slice_mask;
4154         u32 start, end;
4155         int ret;
4156
4157         if (new_dbuf_state->weight[pipe] == 0) {
4158                 new_dbuf_state->ddb[pipe].start = 0;
4159                 new_dbuf_state->ddb[pipe].end = 0;
4160                 goto out;
4161         }
4162
4163         dbuf_slice_mask = new_dbuf_state->slices[pipe];
4164
4165         skl_ddb_entry_for_slices(dev_priv, dbuf_slice_mask, &ddb_slices);
4166         ddb_range_size = skl_ddb_entry_size(&ddb_slices);
4167
4168         intel_crtc_dbuf_weights(new_dbuf_state, pipe,
4169                                 &weight_start, &weight_end, &weight_total);
4170
4171         start = ddb_range_size * weight_start / weight_total;
4172         end = ddb_range_size * weight_end / weight_total;
4173
4174         new_dbuf_state->ddb[pipe].start = ddb_slices.start + start;
4175         new_dbuf_state->ddb[pipe].end = ddb_slices.start + end;
4176
4177 out:
4178         if (skl_ddb_entry_equal(&old_dbuf_state->ddb[pipe],
4179                                 &new_dbuf_state->ddb[pipe]))
4180                 return 0;
4181
4182         ret = intel_atomic_lock_global_state(&new_dbuf_state->base);
4183         if (ret)
4184                 return ret;
4185
4186         crtc_state = intel_atomic_get_crtc_state(&state->base, crtc);
4187         if (IS_ERR(crtc_state))
4188                 return PTR_ERR(crtc_state);
4189
4190         crtc_state->wm.skl.ddb = new_dbuf_state->ddb[pipe];
4191
4192         drm_dbg_kms(&dev_priv->drm,
4193                     "[CRTC:%d:%s] dbuf slices 0x%x -> 0x%x, ddb (%d - %d) -> (%d - %d), active pipes 0x%x -> 0x%x\n",
4194                     crtc->base.base.id, crtc->base.name,
4195                     old_dbuf_state->slices[pipe], new_dbuf_state->slices[pipe],
4196                     old_dbuf_state->ddb[pipe].start, old_dbuf_state->ddb[pipe].end,
4197                     new_dbuf_state->ddb[pipe].start, new_dbuf_state->ddb[pipe].end,
4198                     old_dbuf_state->active_pipes, new_dbuf_state->active_pipes);
4199
4200         return 0;
4201 }
4202
4203 static int skl_compute_wm_params(const struct intel_crtc_state *crtc_state,
4204                                  int width, const struct drm_format_info *format,
4205                                  u64 modifier, unsigned int rotation,
4206                                  u32 plane_pixel_rate, struct skl_wm_params *wp,
4207                                  int color_plane);
4208 static void skl_compute_plane_wm(const struct intel_crtc_state *crtc_state,
4209                                  int level,
4210                                  unsigned int latency,
4211                                  const struct skl_wm_params *wp,
4212                                  const struct skl_wm_level *result_prev,
4213                                  struct skl_wm_level *result /* out */);
4214
4215 static unsigned int
4216 skl_cursor_allocation(const struct intel_crtc_state *crtc_state,
4217                       int num_active)
4218 {
4219         struct drm_i915_private *dev_priv = to_i915(crtc_state->uapi.crtc->dev);
4220         int level, max_level = ilk_wm_max_level(dev_priv);
4221         struct skl_wm_level wm = {};
4222         int ret, min_ddb_alloc = 0;
4223         struct skl_wm_params wp;
4224
4225         ret = skl_compute_wm_params(crtc_state, 256,
4226                                     drm_format_info(DRM_FORMAT_ARGB8888),
4227                                     DRM_FORMAT_MOD_LINEAR,
4228                                     DRM_MODE_ROTATE_0,
4229                                     crtc_state->pixel_rate, &wp, 0);
4230         drm_WARN_ON(&dev_priv->drm, ret);
4231
4232         for (level = 0; level <= max_level; level++) {
4233                 unsigned int latency = dev_priv->wm.skl_latency[level];
4234
4235                 skl_compute_plane_wm(crtc_state, level, latency, &wp, &wm, &wm);
4236                 if (wm.min_ddb_alloc == U16_MAX)
4237                         break;
4238
4239                 min_ddb_alloc = wm.min_ddb_alloc;
4240         }
4241
4242         return max(num_active == 1 ? 32 : 8, min_ddb_alloc);
4243 }
4244
4245 static void skl_ddb_entry_init_from_hw(struct drm_i915_private *dev_priv,
4246                                        struct skl_ddb_entry *entry, u32 reg)
4247 {
4248
4249         entry->start = reg & DDB_ENTRY_MASK;
4250         entry->end = (reg >> DDB_ENTRY_END_SHIFT) & DDB_ENTRY_MASK;
4251
4252         if (entry->end)
4253                 entry->end += 1;
4254 }
4255
4256 static void
4257 skl_ddb_get_hw_plane_state(struct drm_i915_private *dev_priv,
4258                            const enum pipe pipe,
4259                            const enum plane_id plane_id,
4260                            struct skl_ddb_entry *ddb_y,
4261                            struct skl_ddb_entry *ddb_uv)
4262 {
4263         u32 val, val2;
4264         u32 fourcc = 0;
4265
4266         /* Cursor doesn't support NV12/planar, so no extra calculation needed */
4267         if (plane_id == PLANE_CURSOR) {
4268                 val = intel_uncore_read(&dev_priv->uncore, CUR_BUF_CFG(pipe));
4269                 skl_ddb_entry_init_from_hw(dev_priv, ddb_y, val);
4270                 return;
4271         }
4272
4273         val = intel_uncore_read(&dev_priv->uncore, PLANE_CTL(pipe, plane_id));
4274
4275         /* No DDB allocated for disabled planes */
4276         if (val & PLANE_CTL_ENABLE)
4277                 fourcc = skl_format_to_fourcc(val & PLANE_CTL_FORMAT_MASK,
4278                                               val & PLANE_CTL_ORDER_RGBX,
4279                                               val & PLANE_CTL_ALPHA_MASK);
4280
4281         if (INTEL_GEN(dev_priv) >= 11) {
4282                 val = intel_uncore_read(&dev_priv->uncore, PLANE_BUF_CFG(pipe, plane_id));
4283                 skl_ddb_entry_init_from_hw(dev_priv, ddb_y, val);
4284         } else {
4285                 val = intel_uncore_read(&dev_priv->uncore, PLANE_BUF_CFG(pipe, plane_id));
4286                 val2 = intel_uncore_read(&dev_priv->uncore, PLANE_NV12_BUF_CFG(pipe, plane_id));
4287
4288                 if (fourcc &&
4289                     drm_format_info_is_yuv_semiplanar(drm_format_info(fourcc)))
4290                         swap(val, val2);
4291
4292                 skl_ddb_entry_init_from_hw(dev_priv, ddb_y, val);
4293                 skl_ddb_entry_init_from_hw(dev_priv, ddb_uv, val2);
4294         }
4295 }
4296
4297 void skl_pipe_ddb_get_hw_state(struct intel_crtc *crtc,
4298                                struct skl_ddb_entry *ddb_y,
4299                                struct skl_ddb_entry *ddb_uv)
4300 {
4301         struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
4302         enum intel_display_power_domain power_domain;
4303         enum pipe pipe = crtc->pipe;
4304         intel_wakeref_t wakeref;
4305         enum plane_id plane_id;
4306
4307         power_domain = POWER_DOMAIN_PIPE(pipe);
4308         wakeref = intel_display_power_get_if_enabled(dev_priv, power_domain);
4309         if (!wakeref)
4310                 return;
4311
4312         for_each_plane_id_on_crtc(crtc, plane_id)
4313                 skl_ddb_get_hw_plane_state(dev_priv, pipe,
4314                                            plane_id,
4315                                            &ddb_y[plane_id],
4316                                            &ddb_uv[plane_id]);
4317
4318         intel_display_power_put(dev_priv, power_domain, wakeref);
4319 }
4320
4321 /*
4322  * Determines the downscale amount of a plane for the purposes of watermark calculations.
4323  * The bspec defines downscale amount as:
4324  *
4325  * """
4326  * Horizontal down scale amount = maximum[1, Horizontal source size /
4327  *                                           Horizontal destination size]
4328  * Vertical down scale amount = maximum[1, Vertical source size /
4329  *                                         Vertical destination size]
4330  * Total down scale amount = Horizontal down scale amount *
4331  *                           Vertical down scale amount
4332  * """
4333  *
4334  * Return value is provided in 16.16 fixed point form to retain fractional part.
4335  * Caller should take care of dividing & rounding off the value.
4336  */
4337 static uint_fixed_16_16_t
4338 skl_plane_downscale_amount(const struct intel_crtc_state *crtc_state,
4339                            const struct intel_plane_state *plane_state)
4340 {
4341         struct drm_i915_private *dev_priv = to_i915(crtc_state->uapi.crtc->dev);
4342         u32 src_w, src_h, dst_w, dst_h;
4343         uint_fixed_16_16_t fp_w_ratio, fp_h_ratio;
4344         uint_fixed_16_16_t downscale_h, downscale_w;
4345
4346         if (drm_WARN_ON(&dev_priv->drm,
4347                         !intel_wm_plane_visible(crtc_state, plane_state)))
4348                 return u32_to_fixed16(0);
4349
4350         /*
4351          * Src coordinates are already rotated by 270 degrees for
4352          * the 90/270 degree plane rotation cases (to match the
4353          * GTT mapping), hence no need to account for rotation here.
4354          *
4355          * n.b., src is 16.16 fixed point, dst is whole integer.
4356          */
4357         src_w = drm_rect_width(&plane_state->uapi.src) >> 16;
4358         src_h = drm_rect_height(&plane_state->uapi.src) >> 16;
4359         dst_w = drm_rect_width(&plane_state->uapi.dst);
4360         dst_h = drm_rect_height(&plane_state->uapi.dst);
4361
4362         fp_w_ratio = div_fixed16(src_w, dst_w);
4363         fp_h_ratio = div_fixed16(src_h, dst_h);
4364         downscale_w = max_fixed16(fp_w_ratio, u32_to_fixed16(1));
4365         downscale_h = max_fixed16(fp_h_ratio, u32_to_fixed16(1));
4366
4367         return mul_fixed16(downscale_w, downscale_h);
4368 }
4369
4370 struct dbuf_slice_conf_entry {
4371         u8 active_pipes;
4372         u8 dbuf_mask[I915_MAX_PIPES];
4373 };
4374
4375 /*
4376  * Table taken from Bspec 12716
4377  * Pipes do have some preferred DBuf slice affinity,
4378  * plus there are some hardcoded requirements on how
4379  * those should be distributed for multipipe scenarios.
4380  * For more DBuf slices algorithm can get even more messy
4381  * and less readable, so decided to use a table almost
4382  * as is from BSpec itself - that way it is at least easier
4383  * to compare, change and check.
4384  */
4385 static const struct dbuf_slice_conf_entry icl_allowed_dbufs[] =
4386 /* Autogenerated with igt/tools/intel_dbuf_map tool: */
4387 {
4388         {
4389                 .active_pipes = BIT(PIPE_A),
4390                 .dbuf_mask = {
4391                         [PIPE_A] = BIT(DBUF_S1),
4392                 },
4393         },
4394         {
4395                 .active_pipes = BIT(PIPE_B),
4396                 .dbuf_mask = {
4397                         [PIPE_B] = BIT(DBUF_S1),
4398                 },
4399         },
4400         {
4401                 .active_pipes = BIT(PIPE_A) | BIT(PIPE_B),
4402                 .dbuf_mask = {
4403                         [PIPE_A] = BIT(DBUF_S1),
4404                         [PIPE_B] = BIT(DBUF_S2),
4405                 },
4406         },
4407         {
4408                 .active_pipes = BIT(PIPE_C),
4409                 .dbuf_mask = {
4410                         [PIPE_C] = BIT(DBUF_S2),
4411                 },
4412         },
4413         {
4414                 .active_pipes = BIT(PIPE_A) | BIT(PIPE_C),
4415                 .dbuf_mask = {
4416                         [PIPE_A] = BIT(DBUF_S1),
4417                         [PIPE_C] = BIT(DBUF_S2),
4418                 },
4419         },
4420         {
4421                 .active_pipes = BIT(PIPE_B) | BIT(PIPE_C),
4422                 .dbuf_mask = {
4423                         [PIPE_B] = BIT(DBUF_S1),
4424                         [PIPE_C] = BIT(DBUF_S2),
4425                 },
4426         },
4427         {
4428                 .active_pipes = BIT(PIPE_A) | BIT(PIPE_B) | BIT(PIPE_C),
4429                 .dbuf_mask = {
4430                         [PIPE_A] = BIT(DBUF_S1),
4431                         [PIPE_B] = BIT(DBUF_S1),
4432                         [PIPE_C] = BIT(DBUF_S2),
4433                 },
4434         },
4435         {}
4436 };
4437
4438 /*
4439  * Table taken from Bspec 49255
4440  * Pipes do have some preferred DBuf slice affinity,
4441  * plus there are some hardcoded requirements on how
4442  * those should be distributed for multipipe scenarios.
4443  * For more DBuf slices algorithm can get even more messy
4444  * and less readable, so decided to use a table almost
4445  * as is from BSpec itself - that way it is at least easier
4446  * to compare, change and check.
4447  */
4448 static const struct dbuf_slice_conf_entry tgl_allowed_dbufs[] =
4449 /* Autogenerated with igt/tools/intel_dbuf_map tool: */
4450 {
4451         {
4452                 .active_pipes = BIT(PIPE_A),
4453                 .dbuf_mask = {
4454                         [PIPE_A] = BIT(DBUF_S1) | BIT(DBUF_S2),
4455                 },
4456         },
4457         {
4458                 .active_pipes = BIT(PIPE_B),
4459                 .dbuf_mask = {
4460                         [PIPE_B] = BIT(DBUF_S1) | BIT(DBUF_S2),
4461                 },
4462         },
4463         {
4464                 .active_pipes = BIT(PIPE_A) | BIT(PIPE_B),
4465                 .dbuf_mask = {
4466                         [PIPE_A] = BIT(DBUF_S2),
4467                         [PIPE_B] = BIT(DBUF_S1),
4468                 },
4469         },
4470         {
4471                 .active_pipes = BIT(PIPE_C),
4472                 .dbuf_mask = {
4473                         [PIPE_C] = BIT(DBUF_S2) | BIT(DBUF_S1),
4474                 },
4475         },
4476         {
4477                 .active_pipes = BIT(PIPE_A) | BIT(PIPE_C),
4478                 .dbuf_mask = {
4479                         [PIPE_A] = BIT(DBUF_S1),
4480                         [PIPE_C] = BIT(DBUF_S2),
4481                 },
4482         },
4483         {
4484                 .active_pipes = BIT(PIPE_B) | BIT(PIPE_C),
4485                 .dbuf_mask = {
4486                         [PIPE_B] = BIT(DBUF_S1),
4487                         [PIPE_C] = BIT(DBUF_S2),
4488                 },
4489         },
4490         {
4491                 .active_pipes = BIT(PIPE_A) | BIT(PIPE_B) | BIT(PIPE_C),
4492                 .dbuf_mask = {
4493                         [PIPE_A] = BIT(DBUF_S1),
4494                         [PIPE_B] = BIT(DBUF_S1),
4495                         [PIPE_C] = BIT(DBUF_S2),
4496                 },
4497         },
4498         {
4499                 .active_pipes = BIT(PIPE_D),
4500                 .dbuf_mask = {
4501                         [PIPE_D] = BIT(DBUF_S2) | BIT(DBUF_S1),
4502                 },
4503         },
4504         {
4505                 .active_pipes = BIT(PIPE_A) | BIT(PIPE_D),
4506                 .dbuf_mask = {
4507                         [PIPE_A] = BIT(DBUF_S1),
4508                         [PIPE_D] = BIT(DBUF_S2),
4509                 },
4510         },
4511         {
4512                 .active_pipes = BIT(PIPE_B) | BIT(PIPE_D),
4513                 .dbuf_mask = {
4514                         [PIPE_B] = BIT(DBUF_S1),
4515                         [PIPE_D] = BIT(DBUF_S2),
4516                 },
4517         },
4518         {
4519                 .active_pipes = BIT(PIPE_A) | BIT(PIPE_B) | BIT(PIPE_D),
4520                 .dbuf_mask = {
4521                         [PIPE_A] = BIT(DBUF_S1),
4522                         [PIPE_B] = BIT(DBUF_S1),
4523                         [PIPE_D] = BIT(DBUF_S2),
4524                 },
4525         },
4526         {
4527                 .active_pipes = BIT(PIPE_C) | BIT(PIPE_D),
4528                 .dbuf_mask = {
4529                         [PIPE_C] = BIT(DBUF_S1),
4530                         [PIPE_D] = BIT(DBUF_S2),
4531                 },
4532         },
4533         {
4534                 .active_pipes = BIT(PIPE_A) | BIT(PIPE_C) | BIT(PIPE_D),
4535                 .dbuf_mask = {
4536                         [PIPE_A] = BIT(DBUF_S1),
4537                         [PIPE_C] = BIT(DBUF_S2),
4538                         [PIPE_D] = BIT(DBUF_S2),
4539                 },
4540         },
4541         {
4542                 .active_pipes = BIT(PIPE_B) | BIT(PIPE_C) | BIT(PIPE_D),
4543                 .dbuf_mask = {
4544                         [PIPE_B] = BIT(DBUF_S1),
4545                         [PIPE_C] = BIT(DBUF_S2),
4546                         [PIPE_D] = BIT(DBUF_S2),
4547                 },
4548         },
4549         {
4550                 .active_pipes = BIT(PIPE_A) | BIT(PIPE_B) | BIT(PIPE_C) | BIT(PIPE_D),
4551                 .dbuf_mask = {
4552                         [PIPE_A] = BIT(DBUF_S1),
4553                         [PIPE_B] = BIT(DBUF_S1),
4554                         [PIPE_C] = BIT(DBUF_S2),
4555                         [PIPE_D] = BIT(DBUF_S2),
4556                 },
4557         },
4558         {}
4559 };
4560
4561 static u8 compute_dbuf_slices(enum pipe pipe, u8 active_pipes,
4562                               const struct dbuf_slice_conf_entry *dbuf_slices)
4563 {
4564         int i;
4565
4566         for (i = 0; i < dbuf_slices[i].active_pipes; i++) {
4567                 if (dbuf_slices[i].active_pipes == active_pipes)
4568                         return dbuf_slices[i].dbuf_mask[pipe];
4569         }
4570         return 0;
4571 }
4572
4573 /*
4574  * This function finds an entry with same enabled pipe configuration and
4575  * returns correspondent DBuf slice mask as stated in BSpec for particular
4576  * platform.
4577  */
4578 static u8 icl_compute_dbuf_slices(enum pipe pipe, u8 active_pipes)
4579 {
4580         /*
4581          * FIXME: For ICL this is still a bit unclear as prev BSpec revision
4582          * required calculating "pipe ratio" in order to determine
4583          * if one or two slices can be used for single pipe configurations
4584          * as additional constraint to the existing table.
4585          * However based on recent info, it should be not "pipe ratio"
4586          * but rather ratio between pixel_rate and cdclk with additional
4587          * constants, so for now we are using only table until this is
4588          * clarified. Also this is the reason why crtc_state param is
4589          * still here - we will need it once those additional constraints
4590          * pop up.
4591          */
4592         return compute_dbuf_slices(pipe, active_pipes, icl_allowed_dbufs);
4593 }
4594
4595 static u8 tgl_compute_dbuf_slices(enum pipe pipe, u8 active_pipes)
4596 {
4597         return compute_dbuf_slices(pipe, active_pipes, tgl_allowed_dbufs);
4598 }
4599
4600 static u8 skl_compute_dbuf_slices(struct intel_crtc *crtc, u8 active_pipes)
4601 {
4602         struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
4603         enum pipe pipe = crtc->pipe;
4604
4605         if (IS_GEN(dev_priv, 12))
4606                 return tgl_compute_dbuf_slices(pipe, active_pipes);
4607         else if (IS_GEN(dev_priv, 11))
4608                 return icl_compute_dbuf_slices(pipe, active_pipes);
4609         /*
4610          * For anything else just return one slice yet.
4611          * Should be extended for other platforms.
4612          */
4613         return active_pipes & BIT(pipe) ? BIT(DBUF_S1) : 0;
4614 }
4615
4616 static u64
4617 skl_plane_relative_data_rate(const struct intel_crtc_state *crtc_state,
4618                              const struct intel_plane_state *plane_state,
4619                              int color_plane)
4620 {
4621         struct intel_plane *plane = to_intel_plane(plane_state->uapi.plane);
4622         const struct drm_framebuffer *fb = plane_state->hw.fb;
4623         u32 data_rate;
4624         u32 width = 0, height = 0;
4625         uint_fixed_16_16_t down_scale_amount;
4626         u64 rate;
4627
4628         if (!plane_state->uapi.visible)
4629                 return 0;
4630
4631         if (plane->id == PLANE_CURSOR)
4632                 return 0;
4633
4634         if (color_plane == 1 &&
4635             !intel_format_info_is_yuv_semiplanar(fb->format, fb->modifier))
4636                 return 0;
4637
4638         /*
4639          * Src coordinates are already rotated by 270 degrees for
4640          * the 90/270 degree plane rotation cases (to match the
4641          * GTT mapping), hence no need to account for rotation here.
4642          */
4643         width = drm_rect_width(&plane_state->uapi.src) >> 16;
4644         height = drm_rect_height(&plane_state->uapi.src) >> 16;
4645
4646         /* UV plane does 1/2 pixel sub-sampling */
4647         if (color_plane == 1) {
4648                 width /= 2;
4649                 height /= 2;
4650         }
4651
4652         data_rate = width * height;
4653
4654         down_scale_amount = skl_plane_downscale_amount(crtc_state, plane_state);
4655
4656         rate = mul_round_up_u32_fixed16(data_rate, down_scale_amount);
4657
4658         rate *= fb->format->cpp[color_plane];
4659         return rate;
4660 }
4661
4662 static u64
4663 skl_get_total_relative_data_rate(struct intel_atomic_state *state,
4664                                  struct intel_crtc *crtc)
4665 {
4666         struct intel_crtc_state *crtc_state =
4667                 intel_atomic_get_new_crtc_state(state, crtc);
4668         const struct intel_plane_state *plane_state;
4669         struct intel_plane *plane;
4670         u64 total_data_rate = 0;
4671         enum plane_id plane_id;
4672         int i;
4673
4674         /* Calculate and cache data rate for each plane */
4675         for_each_new_intel_plane_in_state(state, plane, plane_state, i) {
4676                 if (plane->pipe != crtc->pipe)
4677                         continue;
4678
4679                 plane_id = plane->id;
4680
4681                 /* packed/y */
4682                 crtc_state->plane_data_rate[plane_id] =
4683                         skl_plane_relative_data_rate(crtc_state, plane_state, 0);
4684
4685                 /* uv-plane */
4686                 crtc_state->uv_plane_data_rate[plane_id] =
4687                         skl_plane_relative_data_rate(crtc_state, plane_state, 1);
4688         }
4689
4690         for_each_plane_id_on_crtc(crtc, plane_id) {
4691                 total_data_rate += crtc_state->plane_data_rate[plane_id];
4692                 total_data_rate += crtc_state->uv_plane_data_rate[plane_id];
4693         }
4694
4695         return total_data_rate;
4696 }
4697
4698 static u64
4699 icl_get_total_relative_data_rate(struct intel_atomic_state *state,
4700                                  struct intel_crtc *crtc)
4701 {
4702         struct intel_crtc_state *crtc_state =
4703                 intel_atomic_get_new_crtc_state(state, crtc);
4704         const struct intel_plane_state *plane_state;
4705         struct intel_plane *plane;
4706         u64 total_data_rate = 0;
4707         enum plane_id plane_id;
4708         int i;
4709
4710         /* Calculate and cache data rate for each plane */
4711         for_each_new_intel_plane_in_state(state, plane, plane_state, i) {
4712                 if (plane->pipe != crtc->pipe)
4713                         continue;
4714
4715                 plane_id = plane->id;
4716
4717                 if (!plane_state->planar_linked_plane) {
4718                         crtc_state->plane_data_rate[plane_id] =
4719                                 skl_plane_relative_data_rate(crtc_state, plane_state, 0);
4720                 } else {
4721                         enum plane_id y_plane_id;
4722
4723                         /*
4724                          * The slave plane might not iterate in
4725                          * intel_atomic_crtc_state_for_each_plane_state(),
4726                          * and needs the master plane state which may be
4727                          * NULL if we try get_new_plane_state(), so we
4728                          * always calculate from the master.
4729                          */
4730                         if (plane_state->planar_slave)
4731                                 continue;
4732
4733                         /* Y plane rate is calculated on the slave */
4734                         y_plane_id = plane_state->planar_linked_plane->id;
4735                         crtc_state->plane_data_rate[y_plane_id] =
4736                                 skl_plane_relative_data_rate(crtc_state, plane_state, 0);
4737
4738                         crtc_state->plane_data_rate[plane_id] =
4739                                 skl_plane_relative_data_rate(crtc_state, plane_state, 1);
4740                 }
4741         }
4742
4743         for_each_plane_id_on_crtc(crtc, plane_id)
4744                 total_data_rate += crtc_state->plane_data_rate[plane_id];
4745
4746         return total_data_rate;
4747 }
4748
4749 static const struct skl_wm_level *
4750 skl_plane_wm_level(const struct intel_crtc_state *crtc_state,
4751                    enum plane_id plane_id,
4752                    int level)
4753 {
4754         const struct skl_pipe_wm *pipe_wm = &crtc_state->wm.skl.optimal;
4755         const struct skl_plane_wm *wm = &pipe_wm->planes[plane_id];
4756
4757         if (level == 0 && pipe_wm->use_sagv_wm)
4758                 return &wm->sagv_wm0;
4759
4760         return &wm->wm[level];
4761 }
4762
4763 static int
4764 skl_allocate_plane_ddb(struct intel_atomic_state *state,
4765                        struct intel_crtc *crtc)
4766 {
4767         struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
4768         struct intel_crtc_state *crtc_state =
4769                 intel_atomic_get_new_crtc_state(state, crtc);
4770         const struct intel_dbuf_state *dbuf_state =
4771                 intel_atomic_get_new_dbuf_state(state);
4772         const struct skl_ddb_entry *alloc = &dbuf_state->ddb[crtc->pipe];
4773         int num_active = hweight8(dbuf_state->active_pipes);
4774         u16 alloc_size, start = 0;
4775         u16 total[I915_MAX_PLANES] = {};
4776         u16 uv_total[I915_MAX_PLANES] = {};
4777         u64 total_data_rate;
4778         enum plane_id plane_id;
4779         u32 blocks;
4780         int level;
4781
4782         /* Clear the partitioning for disabled planes. */
4783         memset(crtc_state->wm.skl.plane_ddb_y, 0, sizeof(crtc_state->wm.skl.plane_ddb_y));
4784         memset(crtc_state->wm.skl.plane_ddb_uv, 0, sizeof(crtc_state->wm.skl.plane_ddb_uv));
4785
4786         if (!crtc_state->hw.active)
4787                 return 0;
4788
4789         if (INTEL_GEN(dev_priv) >= 11)
4790                 total_data_rate =
4791                         icl_get_total_relative_data_rate(state, crtc);
4792         else
4793                 total_data_rate =
4794                         skl_get_total_relative_data_rate(state, crtc);
4795
4796         alloc_size = skl_ddb_entry_size(alloc);
4797         if (alloc_size == 0)
4798                 return 0;
4799
4800         /* Allocate fixed number of blocks for cursor. */
4801         total[PLANE_CURSOR] = skl_cursor_allocation(crtc_state, num_active);
4802         alloc_size -= total[PLANE_CURSOR];
4803         crtc_state->wm.skl.plane_ddb_y[PLANE_CURSOR].start =
4804                 alloc->end - total[PLANE_CURSOR];
4805         crtc_state->wm.skl.plane_ddb_y[PLANE_CURSOR].end = alloc->end;
4806
4807         if (total_data_rate == 0)
4808                 return 0;
4809
4810         /*
4811          * Find the highest watermark level for which we can satisfy the block
4812          * requirement of active planes.
4813          */
4814         for (level = ilk_wm_max_level(dev_priv); level >= 0; level--) {
4815                 blocks = 0;
4816                 for_each_plane_id_on_crtc(crtc, plane_id) {
4817                         const struct skl_plane_wm *wm =
4818                                 &crtc_state->wm.skl.optimal.planes[plane_id];
4819
4820                         if (plane_id == PLANE_CURSOR) {
4821                                 if (wm->wm[level].min_ddb_alloc > total[PLANE_CURSOR]) {
4822                                         drm_WARN_ON(&dev_priv->drm,
4823                                                     wm->wm[level].min_ddb_alloc != U16_MAX);
4824                                         blocks = U32_MAX;
4825                                         break;
4826                                 }
4827                                 continue;
4828                         }
4829
4830                         blocks += wm->wm[level].min_ddb_alloc;
4831                         blocks += wm->uv_wm[level].min_ddb_alloc;
4832                 }
4833
4834                 if (blocks <= alloc_size) {
4835                         alloc_size -= blocks;
4836                         break;
4837                 }
4838         }
4839
4840         if (level < 0) {
4841                 drm_dbg_kms(&dev_priv->drm,
4842                             "Requested display configuration exceeds system DDB limitations");
4843                 drm_dbg_kms(&dev_priv->drm, "minimum required %d/%d\n",
4844                             blocks, alloc_size);
4845                 return -EINVAL;
4846         }
4847
4848         /*
4849          * Grant each plane the blocks it requires at the highest achievable
4850          * watermark level, plus an extra share of the leftover blocks
4851          * proportional to its relative data rate.
4852          */
4853         for_each_plane_id_on_crtc(crtc, plane_id) {
4854                 const struct skl_plane_wm *wm =
4855                         &crtc_state->wm.skl.optimal.planes[plane_id];
4856                 u64 rate;
4857                 u16 extra;
4858
4859                 if (plane_id == PLANE_CURSOR)
4860                         continue;
4861
4862                 /*
4863                  * We've accounted for all active planes; remaining planes are
4864                  * all disabled.
4865                  */
4866                 if (total_data_rate == 0)
4867                         break;
4868
4869                 rate = crtc_state->plane_data_rate[plane_id];
4870                 extra = min_t(u16, alloc_size,
4871                               DIV64_U64_ROUND_UP(alloc_size * rate,
4872                                                  total_data_rate));
4873                 total[plane_id] = wm->wm[level].min_ddb_alloc + extra;
4874                 alloc_size -= extra;
4875                 total_data_rate -= rate;
4876
4877                 if (total_data_rate == 0)
4878                         break;
4879
4880                 rate = crtc_state->uv_plane_data_rate[plane_id];
4881                 extra = min_t(u16, alloc_size,
4882                               DIV64_U64_ROUND_UP(alloc_size * rate,
4883                                                  total_data_rate));
4884                 uv_total[plane_id] = wm->uv_wm[level].min_ddb_alloc + extra;
4885                 alloc_size -= extra;
4886                 total_data_rate -= rate;
4887         }
4888         drm_WARN_ON(&dev_priv->drm, alloc_size != 0 || total_data_rate != 0);
4889
4890         /* Set the actual DDB start/end points for each plane */
4891         start = alloc->start;
4892         for_each_plane_id_on_crtc(crtc, plane_id) {
4893                 struct skl_ddb_entry *plane_alloc =
4894                         &crtc_state->wm.skl.plane_ddb_y[plane_id];
4895                 struct skl_ddb_entry *uv_plane_alloc =
4896                         &crtc_state->wm.skl.plane_ddb_uv[plane_id];
4897
4898                 if (plane_id == PLANE_CURSOR)
4899                         continue;
4900
4901                 /* Gen11+ uses a separate plane for UV watermarks */
4902                 drm_WARN_ON(&dev_priv->drm,
4903                             INTEL_GEN(dev_priv) >= 11 && uv_total[plane_id]);
4904
4905                 /* Leave disabled planes at (0,0) */
4906                 if (total[plane_id]) {
4907                         plane_alloc->start = start;
4908                         start += total[plane_id];
4909                         plane_alloc->end = start;
4910                 }
4911
4912                 if (uv_total[plane_id]) {
4913                         uv_plane_alloc->start = start;
4914                         start += uv_total[plane_id];
4915                         uv_plane_alloc->end = start;
4916                 }
4917         }
4918
4919         /*
4920          * When we calculated watermark values we didn't know how high
4921          * of a level we'd actually be able to hit, so we just marked
4922          * all levels as "enabled."  Go back now and disable the ones
4923          * that aren't actually possible.
4924          */
4925         for (level++; level <= ilk_wm_max_level(dev_priv); level++) {
4926                 for_each_plane_id_on_crtc(crtc, plane_id) {
4927                         struct skl_plane_wm *wm =
4928                                 &crtc_state->wm.skl.optimal.planes[plane_id];
4929
4930                         /*
4931                          * We only disable the watermarks for each plane if
4932                          * they exceed the ddb allocation of said plane. This
4933                          * is done so that we don't end up touching cursor
4934                          * watermarks needlessly when some other plane reduces
4935                          * our max possible watermark level.
4936                          *
4937                          * Bspec has this to say about the PLANE_WM enable bit:
4938                          * "All the watermarks at this level for all enabled
4939                          *  planes must be enabled before the level will be used."
4940                          * So this is actually safe to do.
4941                          */
4942                         if (wm->wm[level].min_ddb_alloc > total[plane_id] ||
4943                             wm->uv_wm[level].min_ddb_alloc > uv_total[plane_id])
4944                                 memset(&wm->wm[level], 0, sizeof(wm->wm[level]));
4945
4946                         /*
4947                          * Wa_1408961008:icl, ehl
4948                          * Underruns with WM1+ disabled
4949                          */
4950                         if (IS_GEN(dev_priv, 11) &&
4951                             level == 1 && wm->wm[0].plane_en) {
4952                                 wm->wm[level].plane_res_b = wm->wm[0].plane_res_b;
4953                                 wm->wm[level].plane_res_l = wm->wm[0].plane_res_l;
4954                                 wm->wm[level].ignore_lines = wm->wm[0].ignore_lines;
4955                         }
4956                 }
4957         }
4958
4959         /*
4960          * Go back and disable the transition watermark if it turns out we
4961          * don't have enough DDB blocks for it.
4962          */
4963         for_each_plane_id_on_crtc(crtc, plane_id) {
4964                 struct skl_plane_wm *wm =
4965                         &crtc_state->wm.skl.optimal.planes[plane_id];
4966
4967                 if (wm->trans_wm.plane_res_b >= total[plane_id])
4968                         memset(&wm->trans_wm, 0, sizeof(wm->trans_wm));
4969         }
4970
4971         return 0;
4972 }
4973
4974 /*
4975  * The max latency should be 257 (max the punit can code is 255 and we add 2us
4976  * for the read latency) and cpp should always be <= 8, so that
4977  * should allow pixel_rate up to ~2 GHz which seems sufficient since max
4978  * 2xcdclk is 1350 MHz and the pixel rate should never exceed that.
4979 */
4980 static uint_fixed_16_16_t
4981 skl_wm_method1(const struct drm_i915_private *dev_priv, u32 pixel_rate,
4982                u8 cpp, u32 latency, u32 dbuf_block_size)
4983 {
4984         u32 wm_intermediate_val;
4985         uint_fixed_16_16_t ret;
4986
4987         if (latency == 0)
4988                 return FP_16_16_MAX;
4989
4990         wm_intermediate_val = latency * pixel_rate * cpp;
4991         ret = div_fixed16(wm_intermediate_val, 1000 * dbuf_block_size);
4992
4993         if (INTEL_GEN(dev_priv) >= 10 || IS_GEMINILAKE(dev_priv))
4994                 ret = add_fixed16_u32(ret, 1);
4995
4996         return ret;
4997 }
4998
4999 static uint_fixed_16_16_t
5000 skl_wm_method2(u32 pixel_rate, u32 pipe_htotal, u32 latency,
5001                uint_fixed_16_16_t plane_blocks_per_line)
5002 {
5003         u32 wm_intermediate_val;
5004         uint_fixed_16_16_t ret;
5005
5006         if (latency == 0)
5007                 return FP_16_16_MAX;
5008
5009         wm_intermediate_val = latency * pixel_rate;
5010         wm_intermediate_val = DIV_ROUND_UP(wm_intermediate_val,
5011                                            pipe_htotal * 1000);
5012         ret = mul_u32_fixed16(wm_intermediate_val, plane_blocks_per_line);
5013         return ret;
5014 }
5015
5016 static uint_fixed_16_16_t
5017 intel_get_linetime_us(const struct intel_crtc_state *crtc_state)
5018 {
5019         struct drm_i915_private *dev_priv = to_i915(crtc_state->uapi.crtc->dev);
5020         u32 pixel_rate;
5021         u32 crtc_htotal;
5022         uint_fixed_16_16_t linetime_us;
5023
5024         if (!crtc_state->hw.active)
5025                 return u32_to_fixed16(0);
5026
5027         pixel_rate = crtc_state->pixel_rate;
5028
5029         if (drm_WARN_ON(&dev_priv->drm, pixel_rate == 0))
5030                 return u32_to_fixed16(0);
5031
5032         crtc_htotal = crtc_state->hw.pipe_mode.crtc_htotal;
5033         linetime_us = div_fixed16(crtc_htotal * 1000, pixel_rate);
5034
5035         return linetime_us;
5036 }
5037
5038 static int
5039 skl_compute_wm_params(const struct intel_crtc_state *crtc_state,
5040                       int width, const struct drm_format_info *format,
5041                       u64 modifier, unsigned int rotation,
5042                       u32 plane_pixel_rate, struct skl_wm_params *wp,
5043                       int color_plane)
5044 {
5045         struct intel_crtc *crtc = to_intel_crtc(crtc_state->uapi.crtc);
5046         struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
5047         u32 interm_pbpl;
5048
5049         /* only planar format has two planes */
5050         if (color_plane == 1 &&
5051             !intel_format_info_is_yuv_semiplanar(format, modifier)) {
5052                 drm_dbg_kms(&dev_priv->drm,
5053                             "Non planar format have single plane\n");
5054                 return -EINVAL;
5055         }
5056
5057         wp->y_tiled = modifier == I915_FORMAT_MOD_Y_TILED ||
5058                       modifier == I915_FORMAT_MOD_Yf_TILED ||
5059                       modifier == I915_FORMAT_MOD_Y_TILED_CCS ||
5060                       modifier == I915_FORMAT_MOD_Yf_TILED_CCS;
5061         wp->x_tiled = modifier == I915_FORMAT_MOD_X_TILED;
5062         wp->rc_surface = modifier == I915_FORMAT_MOD_Y_TILED_CCS ||
5063                          modifier == I915_FORMAT_MOD_Yf_TILED_CCS;
5064         wp->is_planar = intel_format_info_is_yuv_semiplanar(format, modifier);
5065
5066         wp->width = width;
5067         if (color_plane == 1 && wp->is_planar)
5068                 wp->width /= 2;
5069
5070         wp->cpp = format->cpp[color_plane];
5071         wp->plane_pixel_rate = plane_pixel_rate;
5072
5073         if (INTEL_GEN(dev_priv) >= 11 &&
5074             modifier == I915_FORMAT_MOD_Yf_TILED  && wp->cpp == 1)
5075                 wp->dbuf_block_size = 256;
5076         else
5077                 wp->dbuf_block_size = 512;
5078
5079         if (drm_rotation_90_or_270(rotation)) {
5080                 switch (wp->cpp) {
5081                 case 1:
5082                         wp->y_min_scanlines = 16;
5083                         break;
5084                 case 2:
5085                         wp->y_min_scanlines = 8;
5086                         break;
5087                 case 4:
5088                         wp->y_min_scanlines = 4;
5089                         break;
5090                 default:
5091                         MISSING_CASE(wp->cpp);
5092                         return -EINVAL;
5093                 }
5094         } else {
5095                 wp->y_min_scanlines = 4;
5096         }
5097
5098         if (skl_needs_memory_bw_wa(dev_priv))
5099                 wp->y_min_scanlines *= 2;
5100
5101         wp->plane_bytes_per_line = wp->width * wp->cpp;
5102         if (wp->y_tiled) {
5103                 interm_pbpl = DIV_ROUND_UP(wp->plane_bytes_per_line *
5104                                            wp->y_min_scanlines,
5105                                            wp->dbuf_block_size);
5106
5107                 if (INTEL_GEN(dev_priv) >= 10 || IS_GEMINILAKE(dev_priv))
5108                         interm_pbpl++;
5109
5110                 wp->plane_blocks_per_line = div_fixed16(interm_pbpl,
5111                                                         wp->y_min_scanlines);
5112         } else {
5113                 interm_pbpl = DIV_ROUND_UP(wp->plane_bytes_per_line,
5114                                            wp->dbuf_block_size);
5115
5116                 if (!wp->x_tiled ||
5117                     INTEL_GEN(dev_priv) >= 10 || IS_GEMINILAKE(dev_priv))
5118                         interm_pbpl++;
5119
5120                 wp->plane_blocks_per_line = u32_to_fixed16(interm_pbpl);
5121         }
5122
5123         wp->y_tile_minimum = mul_u32_fixed16(wp->y_min_scanlines,
5124                                              wp->plane_blocks_per_line);
5125
5126         wp->linetime_us = fixed16_to_u32_round_up(
5127                                         intel_get_linetime_us(crtc_state));
5128
5129         return 0;
5130 }
5131
5132 static int
5133 skl_compute_plane_wm_params(const struct intel_crtc_state *crtc_state,
5134                             const struct intel_plane_state *plane_state,
5135                             struct skl_wm_params *wp, int color_plane)
5136 {
5137         const struct drm_framebuffer *fb = plane_state->hw.fb;
5138         int width;
5139
5140         /*
5141          * Src coordinates are already rotated by 270 degrees for
5142          * the 90/270 degree plane rotation cases (to match the
5143          * GTT mapping), hence no need to account for rotation here.
5144          */
5145         width = drm_rect_width(&plane_state->uapi.src) >> 16;
5146
5147         return skl_compute_wm_params(crtc_state, width,
5148                                      fb->format, fb->modifier,
5149                                      plane_state->hw.rotation,
5150                                      intel_plane_pixel_rate(crtc_state, plane_state),
5151                                      wp, color_plane);
5152 }
5153
5154 static bool skl_wm_has_lines(struct drm_i915_private *dev_priv, int level)
5155 {
5156         if (INTEL_GEN(dev_priv) >= 10 || IS_GEMINILAKE(dev_priv))
5157                 return true;
5158
5159         /* The number of lines are ignored for the level 0 watermark. */
5160         return level > 0;
5161 }
5162
5163 static void skl_compute_plane_wm(const struct intel_crtc_state *crtc_state,
5164                                  int level,
5165                                  unsigned int latency,
5166                                  const struct skl_wm_params *wp,
5167                                  const struct skl_wm_level *result_prev,
5168                                  struct skl_wm_level *result /* out */)
5169 {
5170         struct drm_i915_private *dev_priv = to_i915(crtc_state->uapi.crtc->dev);
5171         uint_fixed_16_16_t method1, method2;
5172         uint_fixed_16_16_t selected_result;
5173         u32 res_blocks, res_lines, min_ddb_alloc = 0;
5174
5175         if (latency == 0) {
5176                 /* reject it */
5177                 result->min_ddb_alloc = U16_MAX;
5178                 return;
5179         }
5180
5181         /*
5182          * WaIncreaseLatencyIPCEnabled: kbl,cfl
5183          * Display WA #1141: kbl,cfl
5184          */
5185         if ((IS_KABYLAKE(dev_priv) ||
5186              IS_COFFEELAKE(dev_priv) ||
5187              IS_COMETLAKE(dev_priv)) &&
5188             dev_priv->ipc_enabled)
5189                 latency += 4;
5190
5191         if (skl_needs_memory_bw_wa(dev_priv) && wp->x_tiled)
5192                 latency += 15;
5193
5194         method1 = skl_wm_method1(dev_priv, wp->plane_pixel_rate,
5195                                  wp->cpp, latency, wp->dbuf_block_size);
5196         method2 = skl_wm_method2(wp->plane_pixel_rate,
5197                                  crtc_state->hw.pipe_mode.crtc_htotal,
5198                                  latency,
5199                                  wp->plane_blocks_per_line);
5200
5201         if (wp->y_tiled) {
5202                 selected_result = max_fixed16(method2, wp->y_tile_minimum);
5203         } else {
5204                 if ((wp->cpp * crtc_state->hw.pipe_mode.crtc_htotal /
5205                      wp->dbuf_block_size < 1) &&
5206                      (wp->plane_bytes_per_line / wp->dbuf_block_size < 1)) {
5207                         selected_result = method2;
5208                 } else if (latency >= wp->linetime_us) {
5209                         if (IS_GEN(dev_priv, 9) &&
5210                             !IS_GEMINILAKE(dev_priv))
5211                                 selected_result = min_fixed16(method1, method2);
5212                         else
5213                                 selected_result = method2;
5214                 } else {
5215                         selected_result = method1;
5216                 }
5217         }
5218
5219         res_blocks = fixed16_to_u32_round_up(selected_result) + 1;
5220         res_lines = div_round_up_fixed16(selected_result,
5221                                          wp->plane_blocks_per_line);
5222
5223         if (IS_GEN9_BC(dev_priv) || IS_BROXTON(dev_priv)) {
5224                 /* Display WA #1125: skl,bxt,kbl */
5225                 if (level == 0 && wp->rc_surface)
5226                         res_blocks +=
5227                                 fixed16_to_u32_round_up(wp->y_tile_minimum);
5228
5229                 /* Display WA #1126: skl,bxt,kbl */
5230                 if (level >= 1 && level <= 7) {
5231                         if (wp->y_tiled) {
5232                                 res_blocks +=
5233                                     fixed16_to_u32_round_up(wp->y_tile_minimum);
5234                                 res_lines += wp->y_min_scanlines;
5235                         } else {
5236                                 res_blocks++;
5237                         }
5238
5239                         /*
5240                          * Make sure result blocks for higher latency levels are
5241                          * atleast as high as level below the current level.
5242                          * Assumption in DDB algorithm optimization for special
5243                          * cases. Also covers Display WA #1125 for RC.
5244                          */
5245                         if (result_prev->plane_res_b > res_blocks)
5246                                 res_blocks = result_prev->plane_res_b;
5247                 }
5248         }
5249
5250         if (INTEL_GEN(dev_priv) >= 11) {
5251                 if (wp->y_tiled) {
5252                         int extra_lines;
5253
5254                         if (res_lines % wp->y_min_scanlines == 0)
5255                                 extra_lines = wp->y_min_scanlines;
5256                         else
5257                                 extra_lines = wp->y_min_scanlines * 2 -
5258                                         res_lines % wp->y_min_scanlines;
5259
5260                         min_ddb_alloc = mul_round_up_u32_fixed16(res_lines + extra_lines,
5261                                                                  wp->plane_blocks_per_line);
5262                 } else {
5263                         min_ddb_alloc = res_blocks +
5264                                 DIV_ROUND_UP(res_blocks, 10);
5265                 }
5266         }
5267
5268         if (!skl_wm_has_lines(dev_priv, level))
5269                 res_lines = 0;
5270
5271         if (res_lines > 31) {
5272                 /* reject it */
5273                 result->min_ddb_alloc = U16_MAX;
5274                 return;
5275         }
5276
5277         /*
5278          * If res_lines is valid, assume we can use this watermark level
5279          * for now.  We'll come back and disable it after we calculate the
5280          * DDB allocation if it turns out we don't actually have enough
5281          * blocks to satisfy it.
5282          */
5283         result->plane_res_b = res_blocks;
5284         result->plane_res_l = res_lines;
5285         /* Bspec says: value >= plane ddb allocation -> invalid, hence the +1 here */
5286         result->min_ddb_alloc = max(min_ddb_alloc, res_blocks) + 1;
5287         result->plane_en = true;
5288
5289         if (INTEL_GEN(dev_priv) < 12)
5290                 result->can_sagv = latency >= dev_priv->sagv_block_time_us;
5291 }
5292
5293 static void
5294 skl_compute_wm_levels(const struct intel_crtc_state *crtc_state,
5295                       const struct skl_wm_params *wm_params,
5296                       struct skl_wm_level *levels)
5297 {
5298         struct drm_i915_private *dev_priv = to_i915(crtc_state->uapi.crtc->dev);
5299         int level, max_level = ilk_wm_max_level(dev_priv);
5300         struct skl_wm_level *result_prev = &levels[0];
5301
5302         for (level = 0; level <= max_level; level++) {
5303                 struct skl_wm_level *result = &levels[level];
5304                 unsigned int latency = dev_priv->wm.skl_latency[level];
5305
5306                 skl_compute_plane_wm(crtc_state, level, latency,
5307                                      wm_params, result_prev, result);
5308
5309                 result_prev = result;
5310         }
5311 }
5312
5313 static void tgl_compute_sagv_wm(const struct intel_crtc_state *crtc_state,
5314                                 const struct skl_wm_params *wm_params,
5315                                 struct skl_plane_wm *plane_wm)
5316 {
5317         struct drm_i915_private *dev_priv = to_i915(crtc_state->uapi.crtc->dev);
5318         struct skl_wm_level *sagv_wm = &plane_wm->sagv_wm0;
5319         struct skl_wm_level *levels = plane_wm->wm;
5320         unsigned int latency = dev_priv->wm.skl_latency[0] + dev_priv->sagv_block_time_us;
5321
5322         skl_compute_plane_wm(crtc_state, 0, latency,
5323                              wm_params, &levels[0],
5324                              sagv_wm);
5325 }
5326
5327 static void skl_compute_transition_wm(const struct intel_crtc_state *crtc_state,
5328                                       const struct skl_wm_params *wp,
5329                                       struct skl_plane_wm *wm)
5330 {
5331         struct drm_device *dev = crtc_state->uapi.crtc->dev;
5332         const struct drm_i915_private *dev_priv = to_i915(dev);
5333         u16 trans_min, trans_amount, trans_y_tile_min;
5334         u16 wm0_sel_res_b, trans_offset_b, res_blocks;
5335
5336         /* Transition WM don't make any sense if ipc is disabled */
5337         if (!dev_priv->ipc_enabled)
5338                 return;
5339
5340         /*
5341          * WaDisableTWM:skl,kbl,cfl,bxt
5342          * Transition WM are not recommended by HW team for GEN9
5343          */
5344         if (IS_GEN9_BC(dev_priv) || IS_BROXTON(dev_priv))
5345                 return;
5346
5347         if (INTEL_GEN(dev_priv) >= 11)
5348                 trans_min = 4;
5349         else
5350                 trans_min = 14;
5351
5352         /* Display WA #1140: glk,cnl */
5353         if (IS_CANNONLAKE(dev_priv) || IS_GEMINILAKE(dev_priv))
5354                 trans_amount = 0;
5355         else
5356                 trans_amount = 10; /* This is configurable amount */
5357
5358         trans_offset_b = trans_min + trans_amount;
5359
5360         /*
5361          * The spec asks for Selected Result Blocks for wm0 (the real value),
5362          * not Result Blocks (the integer value). Pay attention to the capital
5363          * letters. The value wm_l0->plane_res_b is actually Result Blocks, but
5364          * since Result Blocks is the ceiling of Selected Result Blocks plus 1,
5365          * and since we later will have to get the ceiling of the sum in the
5366          * transition watermarks calculation, we can just pretend Selected
5367          * Result Blocks is Result Blocks minus 1 and it should work for the
5368          * current platforms.
5369          */
5370         wm0_sel_res_b = wm->wm[0].plane_res_b - 1;
5371
5372         if (wp->y_tiled) {
5373                 trans_y_tile_min =
5374                         (u16)mul_round_up_u32_fixed16(2, wp->y_tile_minimum);
5375                 res_blocks = max(wm0_sel_res_b, trans_y_tile_min) +
5376                                 trans_offset_b;
5377         } else {
5378                 res_blocks = wm0_sel_res_b + trans_offset_b;
5379         }
5380
5381         /*
5382          * Just assume we can enable the transition watermark.  After
5383          * computing the DDB we'll come back and disable it if that
5384          * assumption turns out to be false.
5385          */
5386         wm->trans_wm.plane_res_b = res_blocks + 1;
5387         wm->trans_wm.plane_en = true;
5388 }
5389
5390 static int skl_build_plane_wm_single(struct intel_crtc_state *crtc_state,
5391                                      const struct intel_plane_state *plane_state,
5392                                      enum plane_id plane_id, int color_plane)
5393 {
5394         struct intel_crtc *crtc = to_intel_crtc(crtc_state->uapi.crtc);
5395         struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
5396         struct skl_plane_wm *wm = &crtc_state->wm.skl.raw.planes[plane_id];
5397         struct skl_wm_params wm_params;
5398         int ret;
5399
5400         ret = skl_compute_plane_wm_params(crtc_state, plane_state,
5401                                           &wm_params, color_plane);
5402         if (ret)
5403                 return ret;
5404
5405         skl_compute_wm_levels(crtc_state, &wm_params, wm->wm);
5406
5407         if (INTEL_GEN(dev_priv) >= 12)
5408                 tgl_compute_sagv_wm(crtc_state, &wm_params, wm);
5409
5410         skl_compute_transition_wm(crtc_state, &wm_params, wm);
5411
5412         return 0;
5413 }
5414
5415 static int skl_build_plane_wm_uv(struct intel_crtc_state *crtc_state,
5416                                  const struct intel_plane_state *plane_state,
5417                                  enum plane_id plane_id)
5418 {
5419         struct skl_plane_wm *wm = &crtc_state->wm.skl.raw.planes[plane_id];
5420         struct skl_wm_params wm_params;
5421         int ret;
5422
5423         wm->is_planar = true;
5424
5425         /* uv plane watermarks must also be validated for NV12/Planar */
5426         ret = skl_compute_plane_wm_params(crtc_state, plane_state,
5427                                           &wm_params, 1);
5428         if (ret)
5429                 return ret;
5430
5431         skl_compute_wm_levels(crtc_state, &wm_params, wm->uv_wm);
5432
5433         return 0;
5434 }
5435
5436 static int skl_build_plane_wm(struct intel_crtc_state *crtc_state,
5437                               const struct intel_plane_state *plane_state)
5438 {
5439         struct intel_plane *plane = to_intel_plane(plane_state->uapi.plane);
5440         enum plane_id plane_id = plane->id;
5441         struct skl_plane_wm *wm = &crtc_state->wm.skl.raw.planes[plane_id];
5442         const struct drm_framebuffer *fb = plane_state->hw.fb;
5443         int ret;
5444
5445         memset(wm, 0, sizeof(*wm));
5446
5447         if (!intel_wm_plane_visible(crtc_state, plane_state))
5448                 return 0;
5449
5450         ret = skl_build_plane_wm_single(crtc_state, plane_state,
5451                                         plane_id, 0);
5452         if (ret)
5453                 return ret;
5454
5455         if (fb->format->is_yuv && fb->format->num_planes > 1) {
5456                 ret = skl_build_plane_wm_uv(crtc_state, plane_state,
5457                                             plane_id);
5458                 if (ret)
5459                         return ret;
5460         }
5461
5462         return 0;
5463 }
5464
5465 static int icl_build_plane_wm(struct intel_crtc_state *crtc_state,
5466                               const struct intel_plane_state *plane_state)
5467 {
5468         struct intel_plane *plane = to_intel_plane(plane_state->uapi.plane);
5469         struct drm_i915_private *dev_priv = to_i915(plane->base.dev);
5470         enum plane_id plane_id = plane->id;
5471         struct skl_plane_wm *wm = &crtc_state->wm.skl.raw.planes[plane_id];
5472         int ret;
5473
5474         memset(wm, 0, sizeof(*wm));
5475
5476         /* Watermarks calculated in master */
5477         if (plane_state->planar_slave)
5478                 return 0;
5479
5480         if (plane_state->planar_linked_plane) {
5481                 const struct drm_framebuffer *fb = plane_state->hw.fb;
5482                 enum plane_id y_plane_id = plane_state->planar_linked_plane->id;
5483
5484                 drm_WARN_ON(&dev_priv->drm,
5485                             !intel_wm_plane_visible(crtc_state, plane_state));
5486                 drm_WARN_ON(&dev_priv->drm, !fb->format->is_yuv ||
5487                             fb->format->num_planes == 1);
5488
5489                 ret = skl_build_plane_wm_single(crtc_state, plane_state,
5490                                                 y_plane_id, 0);
5491                 if (ret)
5492                         return ret;
5493
5494                 ret = skl_build_plane_wm_single(crtc_state, plane_state,
5495                                                 plane_id, 1);
5496                 if (ret)
5497                         return ret;
5498         } else if (intel_wm_plane_visible(crtc_state, plane_state)) {
5499                 ret = skl_build_plane_wm_single(crtc_state, plane_state,
5500                                                 plane_id, 0);
5501                 if (ret)
5502                         return ret;
5503         }
5504
5505         return 0;
5506 }
5507
5508 static int skl_build_pipe_wm(struct intel_atomic_state *state,
5509                              struct intel_crtc *crtc)
5510 {
5511         struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
5512         struct intel_crtc_state *crtc_state =
5513                 intel_atomic_get_new_crtc_state(state, crtc);
5514         const struct intel_plane_state *plane_state;
5515         struct intel_plane *plane;
5516         int ret, i;
5517
5518         for_each_new_intel_plane_in_state(state, plane, plane_state, i) {
5519                 /*
5520                  * FIXME should perhaps check {old,new}_plane_crtc->hw.crtc
5521                  * instead but we don't populate that correctly for NV12 Y
5522                  * planes so for now hack this.
5523                  */
5524                 if (plane->pipe != crtc->pipe)
5525                         continue;
5526
5527                 if (INTEL_GEN(dev_priv) >= 11)
5528                         ret = icl_build_plane_wm(crtc_state, plane_state);
5529                 else
5530                         ret = skl_build_plane_wm(crtc_state, plane_state);
5531                 if (ret)
5532                         return ret;
5533         }
5534
5535         crtc_state->wm.skl.optimal = crtc_state->wm.skl.raw;
5536
5537         return 0;
5538 }
5539
5540 static void skl_ddb_entry_write(struct drm_i915_private *dev_priv,
5541                                 i915_reg_t reg,
5542                                 const struct skl_ddb_entry *entry)
5543 {
5544         if (entry->end)
5545                 intel_de_write_fw(dev_priv, reg,
5546                                   (entry->end - 1) << 16 | entry->start);
5547         else
5548                 intel_de_write_fw(dev_priv, reg, 0);
5549 }
5550
5551 static void skl_write_wm_level(struct drm_i915_private *dev_priv,
5552                                i915_reg_t reg,
5553                                const struct skl_wm_level *level)
5554 {
5555         u32 val = 0;
5556
5557         if (level->plane_en)
5558                 val |= PLANE_WM_EN;
5559         if (level->ignore_lines)
5560                 val |= PLANE_WM_IGNORE_LINES;
5561         val |= level->plane_res_b;
5562         val |= level->plane_res_l << PLANE_WM_LINES_SHIFT;
5563
5564         intel_de_write_fw(dev_priv, reg, val);
5565 }
5566
5567 void skl_write_plane_wm(struct intel_plane *plane,
5568                         const struct intel_crtc_state *crtc_state)
5569 {
5570         struct drm_i915_private *dev_priv = to_i915(plane->base.dev);
5571         int level, max_level = ilk_wm_max_level(dev_priv);
5572         enum plane_id plane_id = plane->id;
5573         enum pipe pipe = plane->pipe;
5574         const struct skl_plane_wm *wm =
5575                 &crtc_state->wm.skl.optimal.planes[plane_id];
5576         const struct skl_ddb_entry *ddb_y =
5577                 &crtc_state->wm.skl.plane_ddb_y[plane_id];
5578         const struct skl_ddb_entry *ddb_uv =
5579                 &crtc_state->wm.skl.plane_ddb_uv[plane_id];
5580
5581         for (level = 0; level <= max_level; level++) {
5582                 const struct skl_wm_level *wm_level;
5583
5584                 wm_level = skl_plane_wm_level(crtc_state, plane_id, level);
5585
5586                 skl_write_wm_level(dev_priv, PLANE_WM(pipe, plane_id, level),
5587                                    wm_level);
5588         }
5589         skl_write_wm_level(dev_priv, PLANE_WM_TRANS(pipe, plane_id),
5590                            &wm->trans_wm);
5591
5592         if (INTEL_GEN(dev_priv) >= 11) {
5593                 skl_ddb_entry_write(dev_priv,
5594                                     PLANE_BUF_CFG(pipe, plane_id), ddb_y);
5595                 return;
5596         }
5597
5598         if (wm->is_planar)
5599                 swap(ddb_y, ddb_uv);
5600
5601         skl_ddb_entry_write(dev_priv,
5602                             PLANE_BUF_CFG(pipe, plane_id), ddb_y);
5603         skl_ddb_entry_write(dev_priv,
5604                             PLANE_NV12_BUF_CFG(pipe, plane_id), ddb_uv);
5605 }
5606
5607 void skl_write_cursor_wm(struct intel_plane *plane,
5608                          const struct intel_crtc_state *crtc_state)
5609 {
5610         struct drm_i915_private *dev_priv = to_i915(plane->base.dev);
5611         int level, max_level = ilk_wm_max_level(dev_priv);
5612         enum plane_id plane_id = plane->id;
5613         enum pipe pipe = plane->pipe;
5614         const struct skl_plane_wm *wm =
5615                 &crtc_state->wm.skl.optimal.planes[plane_id];
5616         const struct skl_ddb_entry *ddb =
5617                 &crtc_state->wm.skl.plane_ddb_y[plane_id];
5618
5619         for (level = 0; level <= max_level; level++) {
5620                 const struct skl_wm_level *wm_level;
5621
5622                 wm_level = skl_plane_wm_level(crtc_state, plane_id, level);
5623
5624                 skl_write_wm_level(dev_priv, CUR_WM(pipe, level),
5625                                    wm_level);
5626         }
5627         skl_write_wm_level(dev_priv, CUR_WM_TRANS(pipe), &wm->trans_wm);
5628
5629         skl_ddb_entry_write(dev_priv, CUR_BUF_CFG(pipe), ddb);
5630 }
5631
5632 bool skl_wm_level_equals(const struct skl_wm_level *l1,
5633                          const struct skl_wm_level *l2)
5634 {
5635         return l1->plane_en == l2->plane_en &&
5636                 l1->ignore_lines == l2->ignore_lines &&
5637                 l1->plane_res_l == l2->plane_res_l &&
5638                 l1->plane_res_b == l2->plane_res_b;
5639 }
5640
5641 static bool skl_plane_wm_equals(struct drm_i915_private *dev_priv,
5642                                 const struct skl_plane_wm *wm1,
5643                                 const struct skl_plane_wm *wm2)
5644 {
5645         int level, max_level = ilk_wm_max_level(dev_priv);
5646
5647         for (level = 0; level <= max_level; level++) {
5648                 /*
5649                  * We don't check uv_wm as the hardware doesn't actually
5650                  * use it. It only gets used for calculating the required
5651                  * ddb allocation.
5652                  */
5653                 if (!skl_wm_level_equals(&wm1->wm[level], &wm2->wm[level]))
5654                         return false;
5655         }
5656
5657         return skl_wm_level_equals(&wm1->trans_wm, &wm2->trans_wm);
5658 }
5659
5660 static bool skl_ddb_entries_overlap(const struct skl_ddb_entry *a,
5661                                     const struct skl_ddb_entry *b)
5662 {
5663         return a->start < b->end && b->start < a->end;
5664 }
5665
5666 static void skl_ddb_entry_union(struct skl_ddb_entry *a,
5667                                 const struct skl_ddb_entry *b)
5668 {
5669         if (a->end && b->end) {
5670                 a->start = min(a->start, b->start);
5671                 a->end = max(a->end, b->end);
5672         } else if (b->end) {
5673                 a->start = b->start;
5674                 a->end = b->end;
5675         }
5676 }
5677
5678 bool skl_ddb_allocation_overlaps(const struct skl_ddb_entry *ddb,
5679                                  const struct skl_ddb_entry *entries,
5680                                  int num_entries, int ignore_idx)
5681 {
5682         int i;
5683
5684         for (i = 0; i < num_entries; i++) {
5685                 if (i != ignore_idx &&
5686                     skl_ddb_entries_overlap(ddb, &entries[i]))
5687                         return true;
5688         }
5689
5690         return false;
5691 }
5692
5693 static int
5694 skl_ddb_add_affected_planes(const struct intel_crtc_state *old_crtc_state,
5695                             struct intel_crtc_state *new_crtc_state)
5696 {
5697         struct intel_atomic_state *state = to_intel_atomic_state(new_crtc_state->uapi.state);
5698         struct intel_crtc *crtc = to_intel_crtc(new_crtc_state->uapi.crtc);
5699         struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
5700         struct intel_plane *plane;
5701
5702         for_each_intel_plane_on_crtc(&dev_priv->drm, crtc, plane) {
5703                 struct intel_plane_state *plane_state;
5704                 enum plane_id plane_id = plane->id;
5705
5706                 if (skl_ddb_entry_equal(&old_crtc_state->wm.skl.plane_ddb_y[plane_id],
5707                                         &new_crtc_state->wm.skl.plane_ddb_y[plane_id]) &&
5708                     skl_ddb_entry_equal(&old_crtc_state->wm.skl.plane_ddb_uv[plane_id],
5709                                         &new_crtc_state->wm.skl.plane_ddb_uv[plane_id]))
5710                         continue;
5711
5712                 plane_state = intel_atomic_get_plane_state(state, plane);
5713                 if (IS_ERR(plane_state))
5714                         return PTR_ERR(plane_state);
5715
5716                 new_crtc_state->update_planes |= BIT(plane_id);
5717         }
5718
5719         return 0;
5720 }
5721
5722 static u8 intel_dbuf_enabled_slices(const struct intel_dbuf_state *dbuf_state)
5723 {
5724         struct drm_i915_private *dev_priv = to_i915(dbuf_state->base.state->base.dev);
5725         u8 enabled_slices;
5726         enum pipe pipe;
5727
5728         /*
5729          * FIXME: For now we always enable slice S1 as per
5730          * the Bspec display initialization sequence.
5731          */
5732         enabled_slices = BIT(DBUF_S1);
5733
5734         for_each_pipe(dev_priv, pipe)
5735                 enabled_slices |= dbuf_state->slices[pipe];
5736
5737         return enabled_slices;
5738 }
5739
5740 static int
5741 skl_compute_ddb(struct intel_atomic_state *state)
5742 {
5743         struct drm_i915_private *dev_priv = to_i915(state->base.dev);
5744         const struct intel_dbuf_state *old_dbuf_state;
5745         struct intel_dbuf_state *new_dbuf_state = NULL;
5746         const struct intel_crtc_state *old_crtc_state;
5747         struct intel_crtc_state *new_crtc_state;
5748         struct intel_crtc *crtc;
5749         int ret, i;
5750
5751         for_each_new_intel_crtc_in_state(state, crtc, new_crtc_state, i) {
5752                 new_dbuf_state = intel_atomic_get_dbuf_state(state);
5753                 if (IS_ERR(new_dbuf_state))
5754                         return PTR_ERR(new_dbuf_state);
5755
5756                 old_dbuf_state = intel_atomic_get_old_dbuf_state(state);
5757                 break;
5758         }
5759
5760         if (!new_dbuf_state)
5761                 return 0;
5762
5763         new_dbuf_state->active_pipes =
5764                 intel_calc_active_pipes(state, old_dbuf_state->active_pipes);
5765
5766         if (old_dbuf_state->active_pipes != new_dbuf_state->active_pipes) {
5767                 ret = intel_atomic_lock_global_state(&new_dbuf_state->base);
5768                 if (ret)
5769                         return ret;
5770         }
5771
5772         for_each_intel_crtc(&dev_priv->drm, crtc) {
5773                 enum pipe pipe = crtc->pipe;
5774
5775                 new_dbuf_state->slices[pipe] =
5776                         skl_compute_dbuf_slices(crtc, new_dbuf_state->active_pipes);
5777
5778                 if (old_dbuf_state->slices[pipe] == new_dbuf_state->slices[pipe])
5779                         continue;
5780
5781                 ret = intel_atomic_lock_global_state(&new_dbuf_state->base);
5782                 if (ret)
5783                         return ret;
5784         }
5785
5786         new_dbuf_state->enabled_slices = intel_dbuf_enabled_slices(new_dbuf_state);
5787
5788         if (old_dbuf_state->enabled_slices != new_dbuf_state->enabled_slices) {
5789                 ret = intel_atomic_serialize_global_state(&new_dbuf_state->base);
5790                 if (ret)
5791                         return ret;
5792
5793                 drm_dbg_kms(&dev_priv->drm,
5794                             "Enabled dbuf slices 0x%x -> 0x%x (out of %d dbuf slices)\n",
5795                             old_dbuf_state->enabled_slices,
5796                             new_dbuf_state->enabled_slices,
5797                             INTEL_INFO(dev_priv)->num_supported_dbuf_slices);
5798         }
5799
5800         for_each_new_intel_crtc_in_state(state, crtc, new_crtc_state, i) {
5801                 enum pipe pipe = crtc->pipe;
5802
5803                 new_dbuf_state->weight[pipe] = intel_crtc_ddb_weight(new_crtc_state);
5804
5805                 if (old_dbuf_state->weight[pipe] == new_dbuf_state->weight[pipe])
5806                         continue;
5807
5808                 ret = intel_atomic_lock_global_state(&new_dbuf_state->base);
5809                 if (ret)
5810                         return ret;
5811         }
5812
5813         for_each_intel_crtc(&dev_priv->drm, crtc) {
5814                 ret = skl_crtc_allocate_ddb(state, crtc);
5815                 if (ret)
5816                         return ret;
5817         }
5818
5819         for_each_oldnew_intel_crtc_in_state(state, crtc, old_crtc_state,
5820                                             new_crtc_state, i) {
5821                 ret = skl_allocate_plane_ddb(state, crtc);
5822                 if (ret)
5823                         return ret;
5824
5825                 ret = skl_ddb_add_affected_planes(old_crtc_state,
5826                                                   new_crtc_state);
5827                 if (ret)
5828                         return ret;
5829         }
5830
5831         return 0;
5832 }
5833
5834 static char enast(bool enable)
5835 {
5836         return enable ? '*' : ' ';
5837 }
5838
5839 static void
5840 skl_print_wm_changes(struct intel_atomic_state *state)
5841 {
5842         struct drm_i915_private *dev_priv = to_i915(state->base.dev);
5843         const struct intel_crtc_state *old_crtc_state;
5844         const struct intel_crtc_state *new_crtc_state;
5845         struct intel_plane *plane;
5846         struct intel_crtc *crtc;
5847         int i;
5848
5849         if (!drm_debug_enabled(DRM_UT_KMS))
5850                 return;
5851
5852         for_each_oldnew_intel_crtc_in_state(state, crtc, old_crtc_state,
5853                                             new_crtc_state, i) {
5854                 const struct skl_pipe_wm *old_pipe_wm, *new_pipe_wm;
5855
5856                 old_pipe_wm = &old_crtc_state->wm.skl.optimal;
5857                 new_pipe_wm = &new_crtc_state->wm.skl.optimal;
5858
5859                 for_each_intel_plane_on_crtc(&dev_priv->drm, crtc, plane) {
5860                         enum plane_id plane_id = plane->id;
5861                         const struct skl_ddb_entry *old, *new;
5862
5863                         old = &old_crtc_state->wm.skl.plane_ddb_y[plane_id];
5864                         new = &new_crtc_state->wm.skl.plane_ddb_y[plane_id];
5865
5866                         if (skl_ddb_entry_equal(old, new))
5867                                 continue;
5868
5869                         drm_dbg_kms(&dev_priv->drm,
5870                                     "[PLANE:%d:%s] ddb (%4d - %4d) -> (%4d - %4d), size %4d -> %4d\n",
5871                                     plane->base.base.id, plane->base.name,
5872                                     old->start, old->end, new->start, new->end,
5873                                     skl_ddb_entry_size(old), skl_ddb_entry_size(new));
5874                 }
5875
5876                 for_each_intel_plane_on_crtc(&dev_priv->drm, crtc, plane) {
5877                         enum plane_id plane_id = plane->id;
5878                         const struct skl_plane_wm *old_wm, *new_wm;
5879
5880                         old_wm = &old_pipe_wm->planes[plane_id];
5881                         new_wm = &new_pipe_wm->planes[plane_id];
5882
5883                         if (skl_plane_wm_equals(dev_priv, old_wm, new_wm))
5884                                 continue;
5885
5886                         drm_dbg_kms(&dev_priv->drm,
5887                                     "[PLANE:%d:%s]   level %cwm0,%cwm1,%cwm2,%cwm3,%cwm4,%cwm5,%cwm6,%cwm7,%ctwm,%cswm"
5888                                     " -> %cwm0,%cwm1,%cwm2,%cwm3,%cwm4,%cwm5,%cwm6,%cwm7,%ctwm,%cswm\n",
5889                                     plane->base.base.id, plane->base.name,
5890                                     enast(old_wm->wm[0].plane_en), enast(old_wm->wm[1].plane_en),
5891                                     enast(old_wm->wm[2].plane_en), enast(old_wm->wm[3].plane_en),
5892                                     enast(old_wm->wm[4].plane_en), enast(old_wm->wm[5].plane_en),
5893                                     enast(old_wm->wm[6].plane_en), enast(old_wm->wm[7].plane_en),
5894                                     enast(old_wm->trans_wm.plane_en),
5895                                     enast(old_wm->sagv_wm0.plane_en),
5896                                     enast(new_wm->wm[0].plane_en), enast(new_wm->wm[1].plane_en),
5897                                     enast(new_wm->wm[2].plane_en), enast(new_wm->wm[3].plane_en),
5898                                     enast(new_wm->wm[4].plane_en), enast(new_wm->wm[5].plane_en),
5899                                     enast(new_wm->wm[6].plane_en), enast(new_wm->wm[7].plane_en),
5900                                     enast(new_wm->trans_wm.plane_en),
5901                                     enast(new_wm->sagv_wm0.plane_en));
5902
5903                         drm_dbg_kms(&dev_priv->drm,
5904                                     "[PLANE:%d:%s]   lines %c%3d,%c%3d,%c%3d,%c%3d,%c%3d,%c%3d,%c%3d,%c%3d,%c%3d,%c%3d"
5905                                       " -> %c%3d,%c%3d,%c%3d,%c%3d,%c%3d,%c%3d,%c%3d,%c%3d,%c%3d,%c%3d\n",
5906                                     plane->base.base.id, plane->base.name,
5907                                     enast(old_wm->wm[0].ignore_lines), old_wm->wm[0].plane_res_l,
5908                                     enast(old_wm->wm[1].ignore_lines), old_wm->wm[1].plane_res_l,
5909                                     enast(old_wm->wm[2].ignore_lines), old_wm->wm[2].plane_res_l,
5910                                     enast(old_wm->wm[3].ignore_lines), old_wm->wm[3].plane_res_l,
5911                                     enast(old_wm->wm[4].ignore_lines), old_wm->wm[4].plane_res_l,
5912                                     enast(old_wm->wm[5].ignore_lines), old_wm->wm[5].plane_res_l,
5913                                     enast(old_wm->wm[6].ignore_lines), old_wm->wm[6].plane_res_l,
5914                                     enast(old_wm->wm[7].ignore_lines), old_wm->wm[7].plane_res_l,
5915                                     enast(old_wm->trans_wm.ignore_lines), old_wm->trans_wm.plane_res_l,
5916                                     enast(old_wm->sagv_wm0.ignore_lines), old_wm->sagv_wm0.plane_res_l,
5917
5918                                     enast(new_wm->wm[0].ignore_lines), new_wm->wm[0].plane_res_l,
5919                                     enast(new_wm->wm[1].ignore_lines), new_wm->wm[1].plane_res_l,
5920                                     enast(new_wm->wm[2].ignore_lines), new_wm->wm[2].plane_res_l,
5921                                     enast(new_wm->wm[3].ignore_lines), new_wm->wm[3].plane_res_l,
5922                                     enast(new_wm->wm[4].ignore_lines), new_wm->wm[4].plane_res_l,
5923                                     enast(new_wm->wm[5].ignore_lines), new_wm->wm[5].plane_res_l,
5924                                     enast(new_wm->wm[6].ignore_lines), new_wm->wm[6].plane_res_l,
5925                                     enast(new_wm->wm[7].ignore_lines), new_wm->wm[7].plane_res_l,
5926                                     enast(new_wm->trans_wm.ignore_lines), new_wm->trans_wm.plane_res_l,
5927                                     enast(new_wm->sagv_wm0.ignore_lines), new_wm->sagv_wm0.plane_res_l);
5928
5929                         drm_dbg_kms(&dev_priv->drm,
5930                                     "[PLANE:%d:%s]  blocks %4d,%4d,%4d,%4d,%4d,%4d,%4d,%4d,%4d,%4d"
5931                                     " -> %4d,%4d,%4d,%4d,%4d,%4d,%4d,%4d,%4d,%4d\n",
5932                                     plane->base.base.id, plane->base.name,
5933                                     old_wm->wm[0].plane_res_b, old_wm->wm[1].plane_res_b,
5934                                     old_wm->wm[2].plane_res_b, old_wm->wm[3].plane_res_b,
5935                                     old_wm->wm[4].plane_res_b, old_wm->wm[5].plane_res_b,
5936                                     old_wm->wm[6].plane_res_b, old_wm->wm[7].plane_res_b,
5937                                     old_wm->trans_wm.plane_res_b,
5938                                     old_wm->sagv_wm0.plane_res_b,
5939                                     new_wm->wm[0].plane_res_b, new_wm->wm[1].plane_res_b,
5940                                     new_wm->wm[2].plane_res_b, new_wm->wm[3].plane_res_b,
5941                                     new_wm->wm[4].plane_res_b, new_wm->wm[5].plane_res_b,
5942                                     new_wm->wm[6].plane_res_b, new_wm->wm[7].plane_res_b,
5943                                     new_wm->trans_wm.plane_res_b,
5944                                     new_wm->sagv_wm0.plane_res_b);
5945
5946                         drm_dbg_kms(&dev_priv->drm,
5947                                     "[PLANE:%d:%s] min_ddb %4d,%4d,%4d,%4d,%4d,%4d,%4d,%4d,%4d,%4d"
5948                                     " -> %4d,%4d,%4d,%4d,%4d,%4d,%4d,%4d,%4d,%4d\n",
5949                                     plane->base.base.id, plane->base.name,
5950                                     old_wm->wm[0].min_ddb_alloc, old_wm->wm[1].min_ddb_alloc,
5951                                     old_wm->wm[2].min_ddb_alloc, old_wm->wm[3].min_ddb_alloc,
5952                                     old_wm->wm[4].min_ddb_alloc, old_wm->wm[5].min_ddb_alloc,
5953                                     old_wm->wm[6].min_ddb_alloc, old_wm->wm[7].min_ddb_alloc,
5954                                     old_wm->trans_wm.min_ddb_alloc,
5955                                     old_wm->sagv_wm0.min_ddb_alloc,
5956                                     new_wm->wm[0].min_ddb_alloc, new_wm->wm[1].min_ddb_alloc,
5957                                     new_wm->wm[2].min_ddb_alloc, new_wm->wm[3].min_ddb_alloc,
5958                                     new_wm->wm[4].min_ddb_alloc, new_wm->wm[5].min_ddb_alloc,
5959                                     new_wm->wm[6].min_ddb_alloc, new_wm->wm[7].min_ddb_alloc,
5960                                     new_wm->trans_wm.min_ddb_alloc,
5961                                     new_wm->sagv_wm0.min_ddb_alloc);
5962                 }
5963         }
5964 }
5965
5966 /*
5967  * To make sure the cursor watermark registers are always consistent
5968  * with our computed state the following scenario needs special
5969  * treatment:
5970  *
5971  * 1. enable cursor
5972  * 2. move cursor entirely offscreen
5973  * 3. disable cursor
5974  *
5975  * Step 2. does call .disable_plane() but does not zero the watermarks
5976  * (since we consider an offscreen cursor still active for the purposes
5977  * of watermarks). Step 3. would not normally call .disable_plane()
5978  * because the actual plane visibility isn't changing, and we don't
5979  * deallocate the cursor ddb until the pipe gets disabled. So we must
5980  * force step 3. to call .disable_plane() to update the watermark
5981  * registers properly.
5982  *
5983  * Other planes do not suffer from this issues as their watermarks are
5984  * calculated based on the actual plane visibility. The only time this
5985  * can trigger for the other planes is during the initial readout as the
5986  * default value of the watermarks registers is not zero.
5987  */
5988 static int skl_wm_add_affected_planes(struct intel_atomic_state *state,
5989                                       struct intel_crtc *crtc)
5990 {
5991         struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
5992         const struct intel_crtc_state *old_crtc_state =
5993                 intel_atomic_get_old_crtc_state(state, crtc);
5994         struct intel_crtc_state *new_crtc_state =
5995                 intel_atomic_get_new_crtc_state(state, crtc);
5996         struct intel_plane *plane;
5997
5998         for_each_intel_plane_on_crtc(&dev_priv->drm, crtc, plane) {
5999                 struct intel_plane_state *plane_state;
6000                 enum plane_id plane_id = plane->id;
6001
6002                 /*
6003                  * Force a full wm update for every plane on modeset.
6004                  * Required because the reset value of the wm registers
6005                  * is non-zero, whereas we want all disabled planes to
6006                  * have zero watermarks. So if we turn off the relevant
6007                  * power well the hardware state will go out of sync
6008                  * with the software state.
6009                  */
6010                 if (!drm_atomic_crtc_needs_modeset(&new_crtc_state->uapi) &&
6011                     skl_plane_wm_equals(dev_priv,
6012                                         &old_crtc_state->wm.skl.optimal.planes[plane_id],
6013                                         &new_crtc_state->wm.skl.optimal.planes[plane_id]))
6014                         continue;
6015
6016                 plane_state = intel_atomic_get_plane_state(state, plane);
6017                 if (IS_ERR(plane_state))
6018                         return PTR_ERR(plane_state);
6019
6020                 new_crtc_state->update_planes |= BIT(plane_id);
6021         }
6022
6023         return 0;
6024 }
6025
6026 static int
6027 skl_compute_wm(struct intel_atomic_state *state)
6028 {
6029         struct intel_crtc *crtc;
6030         struct intel_crtc_state *new_crtc_state;
6031         int ret, i;
6032
6033         for_each_new_intel_crtc_in_state(state, crtc, new_crtc_state, i) {
6034                 ret = skl_build_pipe_wm(state, crtc);
6035                 if (ret)
6036                         return ret;
6037         }
6038
6039         ret = skl_compute_ddb(state);
6040         if (ret)
6041                 return ret;
6042
6043         ret = intel_compute_sagv_mask(state);
6044         if (ret)
6045                 return ret;
6046
6047         /*
6048          * skl_compute_ddb() will have adjusted the final watermarks
6049          * based on how much ddb is available. Now we can actually
6050          * check if the final watermarks changed.
6051          */
6052         for_each_new_intel_crtc_in_state(state, crtc, new_crtc_state, i) {
6053                 ret = skl_wm_add_affected_planes(state, crtc);
6054                 if (ret)
6055                         return ret;
6056         }
6057
6058         skl_print_wm_changes(state);
6059
6060         return 0;
6061 }
6062
6063 static void ilk_compute_wm_config(struct drm_i915_private *dev_priv,
6064                                   struct intel_wm_config *config)
6065 {
6066         struct intel_crtc *crtc;
6067
6068         /* Compute the currently _active_ config */
6069         for_each_intel_crtc(&dev_priv->drm, crtc) {
6070                 const struct intel_pipe_wm *wm = &crtc->wm.active.ilk;
6071
6072                 if (!wm->pipe_enabled)
6073                         continue;
6074
6075                 config->sprites_enabled |= wm->sprites_enabled;
6076                 config->sprites_scaled |= wm->sprites_scaled;
6077                 config->num_pipes_active++;
6078         }
6079 }
6080
6081 static void ilk_program_watermarks(struct drm_i915_private *dev_priv)
6082 {
6083         struct intel_pipe_wm lp_wm_1_2 = {}, lp_wm_5_6 = {}, *best_lp_wm;
6084         struct ilk_wm_maximums max;
6085         struct intel_wm_config config = {};
6086         struct ilk_wm_values results = {};
6087         enum intel_ddb_partitioning partitioning;
6088
6089         ilk_compute_wm_config(dev_priv, &config);
6090
6091         ilk_compute_wm_maximums(dev_priv, 1, &config, INTEL_DDB_PART_1_2, &max);
6092         ilk_wm_merge(dev_priv, &config, &max, &lp_wm_1_2);
6093
6094         /* 5/6 split only in single pipe config on IVB+ */
6095         if (INTEL_GEN(dev_priv) >= 7 &&
6096             config.num_pipes_active == 1 && config.sprites_enabled) {
6097                 ilk_compute_wm_maximums(dev_priv, 1, &config, INTEL_DDB_PART_5_6, &max);
6098                 ilk_wm_merge(dev_priv, &config, &max, &lp_wm_5_6);
6099
6100                 best_lp_wm = ilk_find_best_result(dev_priv, &lp_wm_1_2, &lp_wm_5_6);
6101         } else {
6102                 best_lp_wm = &lp_wm_1_2;
6103         }
6104
6105         partitioning = (best_lp_wm == &lp_wm_1_2) ?
6106                        INTEL_DDB_PART_1_2 : INTEL_DDB_PART_5_6;
6107
6108         ilk_compute_wm_results(dev_priv, best_lp_wm, partitioning, &results);
6109
6110         ilk_write_wm_values(dev_priv, &results);
6111 }
6112
6113 static void ilk_initial_watermarks(struct intel_atomic_state *state,
6114                                    struct intel_crtc *crtc)
6115 {
6116         struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
6117         const struct intel_crtc_state *crtc_state =
6118                 intel_atomic_get_new_crtc_state(state, crtc);
6119
6120         mutex_lock(&dev_priv->wm.wm_mutex);
6121         crtc->wm.active.ilk = crtc_state->wm.ilk.intermediate;
6122         ilk_program_watermarks(dev_priv);
6123         mutex_unlock(&dev_priv->wm.wm_mutex);
6124 }
6125
6126 static void ilk_optimize_watermarks(struct intel_atomic_state *state,
6127                                     struct intel_crtc *crtc)
6128 {
6129         struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
6130         const struct intel_crtc_state *crtc_state =
6131                 intel_atomic_get_new_crtc_state(state, crtc);
6132
6133         if (!crtc_state->wm.need_postvbl_update)
6134                 return;
6135
6136         mutex_lock(&dev_priv->wm.wm_mutex);
6137         crtc->wm.active.ilk = crtc_state->wm.ilk.optimal;
6138         ilk_program_watermarks(dev_priv);
6139         mutex_unlock(&dev_priv->wm.wm_mutex);
6140 }
6141
6142 static void skl_wm_level_from_reg_val(u32 val, struct skl_wm_level *level)
6143 {
6144         level->plane_en = val & PLANE_WM_EN;
6145         level->ignore_lines = val & PLANE_WM_IGNORE_LINES;
6146         level->plane_res_b = val & PLANE_WM_BLOCKS_MASK;
6147         level->plane_res_l = (val >> PLANE_WM_LINES_SHIFT) &
6148                 PLANE_WM_LINES_MASK;
6149 }
6150
6151 void skl_pipe_wm_get_hw_state(struct intel_crtc *crtc,
6152                               struct skl_pipe_wm *out)
6153 {
6154         struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
6155         enum pipe pipe = crtc->pipe;
6156         int level, max_level;
6157         enum plane_id plane_id;
6158         u32 val;
6159
6160         max_level = ilk_wm_max_level(dev_priv);
6161
6162         for_each_plane_id_on_crtc(crtc, plane_id) {
6163                 struct skl_plane_wm *wm = &out->planes[plane_id];
6164
6165                 for (level = 0; level <= max_level; level++) {
6166                         if (plane_id != PLANE_CURSOR)
6167                                 val = intel_uncore_read(&dev_priv->uncore, PLANE_WM(pipe, plane_id, level));
6168                         else
6169                                 val = intel_uncore_read(&dev_priv->uncore, CUR_WM(pipe, level));
6170
6171                         skl_wm_level_from_reg_val(val, &wm->wm[level]);
6172                 }
6173
6174                 if (INTEL_GEN(dev_priv) >= 12)
6175                         wm->sagv_wm0 = wm->wm[0];
6176
6177                 if (plane_id != PLANE_CURSOR)
6178                         val = intel_uncore_read(&dev_priv->uncore, PLANE_WM_TRANS(pipe, plane_id));
6179                 else
6180                         val = intel_uncore_read(&dev_priv->uncore, CUR_WM_TRANS(pipe));
6181
6182                 skl_wm_level_from_reg_val(val, &wm->trans_wm);
6183         }
6184
6185         if (!crtc->active)
6186                 return;
6187 }
6188
6189 void skl_wm_get_hw_state(struct drm_i915_private *dev_priv)
6190 {
6191         struct intel_dbuf_state *dbuf_state =
6192                 to_intel_dbuf_state(dev_priv->dbuf.obj.state);
6193         struct intel_crtc *crtc;
6194
6195         for_each_intel_crtc(&dev_priv->drm, crtc) {
6196                 struct intel_crtc_state *crtc_state =
6197                         to_intel_crtc_state(crtc->base.state);
6198                 enum pipe pipe = crtc->pipe;
6199                 enum plane_id plane_id;
6200
6201                 skl_pipe_wm_get_hw_state(crtc, &crtc_state->wm.skl.optimal);
6202                 crtc_state->wm.skl.raw = crtc_state->wm.skl.optimal;
6203
6204                 memset(&dbuf_state->ddb[pipe], 0, sizeof(dbuf_state->ddb[pipe]));
6205
6206                 for_each_plane_id_on_crtc(crtc, plane_id) {
6207                         struct skl_ddb_entry *ddb_y =
6208                                 &crtc_state->wm.skl.plane_ddb_y[plane_id];
6209                         struct skl_ddb_entry *ddb_uv =
6210                                 &crtc_state->wm.skl.plane_ddb_uv[plane_id];
6211
6212                         skl_ddb_get_hw_plane_state(dev_priv, crtc->pipe,
6213                                                    plane_id, ddb_y, ddb_uv);
6214
6215                         skl_ddb_entry_union(&dbuf_state->ddb[pipe], ddb_y);
6216                         skl_ddb_entry_union(&dbuf_state->ddb[pipe], ddb_uv);
6217                 }
6218
6219                 dbuf_state->slices[pipe] =
6220                         skl_compute_dbuf_slices(crtc, dbuf_state->active_pipes);
6221
6222                 dbuf_state->weight[pipe] = intel_crtc_ddb_weight(crtc_state);
6223
6224                 crtc_state->wm.skl.ddb = dbuf_state->ddb[pipe];
6225
6226                 drm_dbg_kms(&dev_priv->drm,
6227                             "[CRTC:%d:%s] dbuf slices 0x%x, ddb (%d - %d), active pipes 0x%x\n",
6228                             crtc->base.base.id, crtc->base.name,
6229                             dbuf_state->slices[pipe], dbuf_state->ddb[pipe].start,
6230                             dbuf_state->ddb[pipe].end, dbuf_state->active_pipes);
6231         }
6232
6233         dbuf_state->enabled_slices = dev_priv->dbuf.enabled_slices;
6234 }
6235
6236 static void ilk_pipe_wm_get_hw_state(struct intel_crtc *crtc)
6237 {
6238         struct drm_device *dev = crtc->base.dev;
6239         struct drm_i915_private *dev_priv = to_i915(dev);
6240         struct ilk_wm_values *hw = &dev_priv->wm.hw;
6241         struct intel_crtc_state *crtc_state = to_intel_crtc_state(crtc->base.state);
6242         struct intel_pipe_wm *active = &crtc_state->wm.ilk.optimal;
6243         enum pipe pipe = crtc->pipe;
6244
6245         hw->wm_pipe[pipe] = intel_uncore_read(&dev_priv->uncore, WM0_PIPE_ILK(pipe));
6246
6247         memset(active, 0, sizeof(*active));
6248
6249         active->pipe_enabled = crtc->active;
6250
6251         if (active->pipe_enabled) {
6252                 u32 tmp = hw->wm_pipe[pipe];
6253
6254                 /*
6255                  * For active pipes LP0 watermark is marked as
6256                  * enabled, and LP1+ watermaks as disabled since
6257                  * we can't really reverse compute them in case
6258                  * multiple pipes are active.
6259                  */
6260                 active->wm[0].enable = true;
6261                 active->wm[0].pri_val = (tmp & WM0_PIPE_PLANE_MASK) >> WM0_PIPE_PLANE_SHIFT;
6262                 active->wm[0].spr_val = (tmp & WM0_PIPE_SPRITE_MASK) >> WM0_PIPE_SPRITE_SHIFT;
6263                 active->wm[0].cur_val = tmp & WM0_PIPE_CURSOR_MASK;
6264         } else {
6265                 int level, max_level = ilk_wm_max_level(dev_priv);
6266
6267                 /*
6268                  * For inactive pipes, all watermark levels
6269                  * should be marked as enabled but zeroed,
6270                  * which is what we'd compute them to.
6271                  */
6272                 for (level = 0; level <= max_level; level++)
6273                         active->wm[level].enable = true;
6274         }
6275
6276         crtc->wm.active.ilk = *active;
6277 }
6278
6279 #define _FW_WM(value, plane) \
6280         (((value) & DSPFW_ ## plane ## _MASK) >> DSPFW_ ## plane ## _SHIFT)
6281 #define _FW_WM_VLV(value, plane) \
6282         (((value) & DSPFW_ ## plane ## _MASK_VLV) >> DSPFW_ ## plane ## _SHIFT)
6283
6284 static void g4x_read_wm_values(struct drm_i915_private *dev_priv,
6285                                struct g4x_wm_values *wm)
6286 {
6287         u32 tmp;
6288
6289         tmp = intel_uncore_read(&dev_priv->uncore, DSPFW1);
6290         wm->sr.plane = _FW_WM(tmp, SR);
6291         wm->pipe[PIPE_B].plane[PLANE_CURSOR] = _FW_WM(tmp, CURSORB);
6292         wm->pipe[PIPE_B].plane[PLANE_PRIMARY] = _FW_WM(tmp, PLANEB);
6293         wm->pipe[PIPE_A].plane[PLANE_PRIMARY] = _FW_WM(tmp, PLANEA);
6294
6295         tmp = intel_uncore_read(&dev_priv->uncore, DSPFW2);
6296         wm->fbc_en = tmp & DSPFW_FBC_SR_EN;
6297         wm->sr.fbc = _FW_WM(tmp, FBC_SR);
6298         wm->hpll.fbc = _FW_WM(tmp, FBC_HPLL_SR);
6299         wm->pipe[PIPE_B].plane[PLANE_SPRITE0] = _FW_WM(tmp, SPRITEB);
6300         wm->pipe[PIPE_A].plane[PLANE_CURSOR] = _FW_WM(tmp, CURSORA);
6301         wm->pipe[PIPE_A].plane[PLANE_SPRITE0] = _FW_WM(tmp, SPRITEA);
6302
6303         tmp = intel_uncore_read(&dev_priv->uncore, DSPFW3);
6304         wm->hpll_en = tmp & DSPFW_HPLL_SR_EN;
6305         wm->sr.cursor = _FW_WM(tmp, CURSOR_SR);
6306         wm->hpll.cursor = _FW_WM(tmp, HPLL_CURSOR);
6307         wm->hpll.plane = _FW_WM(tmp, HPLL_SR);
6308 }
6309
6310 static void vlv_read_wm_values(struct drm_i915_private *dev_priv,
6311                                struct vlv_wm_values *wm)
6312 {
6313         enum pipe pipe;
6314         u32 tmp;
6315
6316         for_each_pipe(dev_priv, pipe) {
6317                 tmp = intel_uncore_read(&dev_priv->uncore, VLV_DDL(pipe));
6318
6319                 wm->ddl[pipe].plane[PLANE_PRIMARY] =
6320                         (tmp >> DDL_PLANE_SHIFT) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
6321                 wm->ddl[pipe].plane[PLANE_CURSOR] =
6322                         (tmp >> DDL_CURSOR_SHIFT) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
6323                 wm->ddl[pipe].plane[PLANE_SPRITE0] =
6324                         (tmp >> DDL_SPRITE_SHIFT(0)) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
6325                 wm->ddl[pipe].plane[PLANE_SPRITE1] =
6326                         (tmp >> DDL_SPRITE_SHIFT(1)) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
6327         }
6328
6329         tmp = intel_uncore_read(&dev_priv->uncore, DSPFW1);
6330         wm->sr.plane = _FW_WM(tmp, SR);
6331         wm->pipe[PIPE_B].plane[PLANE_CURSOR] = _FW_WM(tmp, CURSORB);
6332         wm->pipe[PIPE_B].plane[PLANE_PRIMARY] = _FW_WM_VLV(tmp, PLANEB);
6333         wm->pipe[PIPE_A].plane[PLANE_PRIMARY] = _FW_WM_VLV(tmp, PLANEA);
6334
6335         tmp = intel_uncore_read(&dev_priv->uncore, DSPFW2);
6336         wm->pipe[PIPE_A].plane[PLANE_SPRITE1] = _FW_WM_VLV(tmp, SPRITEB);
6337         wm->pipe[PIPE_A].plane[PLANE_CURSOR] = _FW_WM(tmp, CURSORA);
6338         wm->pipe[PIPE_A].plane[PLANE_SPRITE0] = _FW_WM_VLV(tmp, SPRITEA);
6339
6340         tmp = intel_uncore_read(&dev_priv->uncore, DSPFW3);
6341         wm->sr.cursor = _FW_WM(tmp, CURSOR_SR);
6342
6343         if (IS_CHERRYVIEW(dev_priv)) {
6344                 tmp = intel_uncore_read(&dev_priv->uncore, DSPFW7_CHV);
6345                 wm->pipe[PIPE_B].plane[PLANE_SPRITE1] = _FW_WM_VLV(tmp, SPRITED);
6346                 wm->pipe[PIPE_B].plane[PLANE_SPRITE0] = _FW_WM_VLV(tmp, SPRITEC);
6347
6348                 tmp = intel_uncore_read(&dev_priv->uncore, DSPFW8_CHV);
6349                 wm->pipe[PIPE_C].plane[PLANE_SPRITE1] = _FW_WM_VLV(tmp, SPRITEF);
6350                 wm->pipe[PIPE_C].plane[PLANE_SPRITE0] = _FW_WM_VLV(tmp, SPRITEE);
6351
6352                 tmp = intel_uncore_read(&dev_priv->uncore, DSPFW9_CHV);
6353                 wm->pipe[PIPE_C].plane[PLANE_PRIMARY] = _FW_WM_VLV(tmp, PLANEC);
6354                 wm->pipe[PIPE_C].plane[PLANE_CURSOR] = _FW_WM(tmp, CURSORC);
6355
6356                 tmp = intel_uncore_read(&dev_priv->uncore, DSPHOWM);
6357                 wm->sr.plane |= _FW_WM(tmp, SR_HI) << 9;
6358                 wm->pipe[PIPE_C].plane[PLANE_SPRITE1] |= _FW_WM(tmp, SPRITEF_HI) << 8;
6359                 wm->pipe[PIPE_C].plane[PLANE_SPRITE0] |= _FW_WM(tmp, SPRITEE_HI) << 8;
6360                 wm->pipe[PIPE_C].plane[PLANE_PRIMARY] |= _FW_WM(tmp, PLANEC_HI) << 8;
6361                 wm->pipe[PIPE_B].plane[PLANE_SPRITE1] |= _FW_WM(tmp, SPRITED_HI) << 8;
6362                 wm->pipe[PIPE_B].plane[PLANE_SPRITE0] |= _FW_WM(tmp, SPRITEC_HI) << 8;
6363                 wm->pipe[PIPE_B].plane[PLANE_PRIMARY] |= _FW_WM(tmp, PLANEB_HI) << 8;
6364                 wm->pipe[PIPE_A].plane[PLANE_SPRITE1] |= _FW_WM(tmp, SPRITEB_HI) << 8;
6365                 wm->pipe[PIPE_A].plane[PLANE_SPRITE0] |= _FW_WM(tmp, SPRITEA_HI) << 8;
6366                 wm->pipe[PIPE_A].plane[PLANE_PRIMARY] |= _FW_WM(tmp, PLANEA_HI) << 8;
6367         } else {
6368                 tmp = intel_uncore_read(&dev_priv->uncore, DSPFW7);
6369                 wm->pipe[PIPE_B].plane[PLANE_SPRITE1] = _FW_WM_VLV(tmp, SPRITED);
6370                 wm->pipe[PIPE_B].plane[PLANE_SPRITE0] = _FW_WM_VLV(tmp, SPRITEC);
6371
6372                 tmp = intel_uncore_read(&dev_priv->uncore, DSPHOWM);
6373                 wm->sr.plane |= _FW_WM(tmp, SR_HI) << 9;
6374                 wm->pipe[PIPE_B].plane[PLANE_SPRITE1] |= _FW_WM(tmp, SPRITED_HI) << 8;
6375                 wm->pipe[PIPE_B].plane[PLANE_SPRITE0] |= _FW_WM(tmp, SPRITEC_HI) << 8;
6376                 wm->pipe[PIPE_B].plane[PLANE_PRIMARY] |= _FW_WM(tmp, PLANEB_HI) << 8;
6377                 wm->pipe[PIPE_A].plane[PLANE_SPRITE1] |= _FW_WM(tmp, SPRITEB_HI) << 8;
6378                 wm->pipe[PIPE_A].plane[PLANE_SPRITE0] |= _FW_WM(tmp, SPRITEA_HI) << 8;
6379                 wm->pipe[PIPE_A].plane[PLANE_PRIMARY] |= _FW_WM(tmp, PLANEA_HI) << 8;
6380         }
6381 }
6382
6383 #undef _FW_WM
6384 #undef _FW_WM_VLV
6385
6386 void g4x_wm_get_hw_state(struct drm_i915_private *dev_priv)
6387 {
6388         struct g4x_wm_values *wm = &dev_priv->wm.g4x;
6389         struct intel_crtc *crtc;
6390
6391         g4x_read_wm_values(dev_priv, wm);
6392
6393         wm->cxsr = intel_uncore_read(&dev_priv->uncore, FW_BLC_SELF) & FW_BLC_SELF_EN;
6394
6395         for_each_intel_crtc(&dev_priv->drm, crtc) {
6396                 struct intel_crtc_state *crtc_state =
6397                         to_intel_crtc_state(crtc->base.state);
6398                 struct g4x_wm_state *active = &crtc->wm.active.g4x;
6399                 struct g4x_pipe_wm *raw;
6400                 enum pipe pipe = crtc->pipe;
6401                 enum plane_id plane_id;
6402                 int level, max_level;
6403
6404                 active->cxsr = wm->cxsr;
6405                 active->hpll_en = wm->hpll_en;
6406                 active->fbc_en = wm->fbc_en;
6407
6408                 active->sr = wm->sr;
6409                 active->hpll = wm->hpll;
6410
6411                 for_each_plane_id_on_crtc(crtc, plane_id) {
6412                         active->wm.plane[plane_id] =
6413                                 wm->pipe[pipe].plane[plane_id];
6414                 }
6415
6416                 if (wm->cxsr && wm->hpll_en)
6417                         max_level = G4X_WM_LEVEL_HPLL;
6418                 else if (wm->cxsr)
6419                         max_level = G4X_WM_LEVEL_SR;
6420                 else
6421                         max_level = G4X_WM_LEVEL_NORMAL;
6422
6423                 level = G4X_WM_LEVEL_NORMAL;
6424                 raw = &crtc_state->wm.g4x.raw[level];
6425                 for_each_plane_id_on_crtc(crtc, plane_id)
6426                         raw->plane[plane_id] = active->wm.plane[plane_id];
6427
6428                 if (++level > max_level)
6429                         goto out;
6430
6431                 raw = &crtc_state->wm.g4x.raw[level];
6432                 raw->plane[PLANE_PRIMARY] = active->sr.plane;
6433                 raw->plane[PLANE_CURSOR] = active->sr.cursor;
6434                 raw->plane[PLANE_SPRITE0] = 0;
6435                 raw->fbc = active->sr.fbc;
6436
6437                 if (++level > max_level)
6438                         goto out;
6439
6440                 raw = &crtc_state->wm.g4x.raw[level];
6441                 raw->plane[PLANE_PRIMARY] = active->hpll.plane;
6442                 raw->plane[PLANE_CURSOR] = active->hpll.cursor;
6443                 raw->plane[PLANE_SPRITE0] = 0;
6444                 raw->fbc = active->hpll.fbc;
6445
6446         out:
6447                 for_each_plane_id_on_crtc(crtc, plane_id)
6448                         g4x_raw_plane_wm_set(crtc_state, level,
6449                                              plane_id, USHRT_MAX);
6450                 g4x_raw_fbc_wm_set(crtc_state, level, USHRT_MAX);
6451
6452                 crtc_state->wm.g4x.optimal = *active;
6453                 crtc_state->wm.g4x.intermediate = *active;
6454
6455                 drm_dbg_kms(&dev_priv->drm,
6456                             "Initial watermarks: pipe %c, plane=%d, cursor=%d, sprite=%d\n",
6457                             pipe_name(pipe),
6458                             wm->pipe[pipe].plane[PLANE_PRIMARY],
6459                             wm->pipe[pipe].plane[PLANE_CURSOR],
6460                             wm->pipe[pipe].plane[PLANE_SPRITE0]);
6461         }
6462
6463         drm_dbg_kms(&dev_priv->drm,
6464                     "Initial SR watermarks: plane=%d, cursor=%d fbc=%d\n",
6465                     wm->sr.plane, wm->sr.cursor, wm->sr.fbc);
6466         drm_dbg_kms(&dev_priv->drm,
6467                     "Initial HPLL watermarks: plane=%d, SR cursor=%d fbc=%d\n",
6468                     wm->hpll.plane, wm->hpll.cursor, wm->hpll.fbc);
6469         drm_dbg_kms(&dev_priv->drm, "Initial SR=%s HPLL=%s FBC=%s\n",
6470                     yesno(wm->cxsr), yesno(wm->hpll_en), yesno(wm->fbc_en));
6471 }
6472
6473 void g4x_wm_sanitize(struct drm_i915_private *dev_priv)
6474 {
6475         struct intel_plane *plane;
6476         struct intel_crtc *crtc;
6477
6478         mutex_lock(&dev_priv->wm.wm_mutex);
6479
6480         for_each_intel_plane(&dev_priv->drm, plane) {
6481                 struct intel_crtc *crtc =
6482                         intel_get_crtc_for_pipe(dev_priv, plane->pipe);
6483                 struct intel_crtc_state *crtc_state =
6484                         to_intel_crtc_state(crtc->base.state);
6485                 struct intel_plane_state *plane_state =
6486                         to_intel_plane_state(plane->base.state);
6487                 struct g4x_wm_state *wm_state = &crtc_state->wm.g4x.optimal;
6488                 enum plane_id plane_id = plane->id;
6489                 int level;
6490
6491                 if (plane_state->uapi.visible)
6492                         continue;
6493
6494                 for (level = 0; level < 3; level++) {
6495                         struct g4x_pipe_wm *raw =
6496                                 &crtc_state->wm.g4x.raw[level];
6497
6498                         raw->plane[plane_id] = 0;
6499                         wm_state->wm.plane[plane_id] = 0;
6500                 }
6501
6502                 if (plane_id == PLANE_PRIMARY) {
6503                         for (level = 0; level < 3; level++) {
6504                                 struct g4x_pipe_wm *raw =
6505                                         &crtc_state->wm.g4x.raw[level];
6506                                 raw->fbc = 0;
6507                         }
6508
6509                         wm_state->sr.fbc = 0;
6510                         wm_state->hpll.fbc = 0;
6511                         wm_state->fbc_en = false;
6512                 }
6513         }
6514
6515         for_each_intel_crtc(&dev_priv->drm, crtc) {
6516                 struct intel_crtc_state *crtc_state =
6517                         to_intel_crtc_state(crtc->base.state);
6518
6519                 crtc_state->wm.g4x.intermediate =
6520                         crtc_state->wm.g4x.optimal;
6521                 crtc->wm.active.g4x = crtc_state->wm.g4x.optimal;
6522         }
6523
6524         g4x_program_watermarks(dev_priv);
6525
6526         mutex_unlock(&dev_priv->wm.wm_mutex);
6527 }
6528
6529 void vlv_wm_get_hw_state(struct drm_i915_private *dev_priv)
6530 {
6531         struct vlv_wm_values *wm = &dev_priv->wm.vlv;
6532         struct intel_crtc *crtc;
6533         u32 val;
6534
6535         vlv_read_wm_values(dev_priv, wm);
6536
6537         wm->cxsr = intel_uncore_read(&dev_priv->uncore, FW_BLC_SELF_VLV) & FW_CSPWRDWNEN;
6538         wm->level = VLV_WM_LEVEL_PM2;
6539
6540         if (IS_CHERRYVIEW(dev_priv)) {
6541                 vlv_punit_get(dev_priv);
6542
6543                 val = vlv_punit_read(dev_priv, PUNIT_REG_DSPSSPM);
6544                 if (val & DSP_MAXFIFO_PM5_ENABLE)
6545                         wm->level = VLV_WM_LEVEL_PM5;
6546
6547                 /*
6548                  * If DDR DVFS is disabled in the BIOS, Punit
6549                  * will never ack the request. So if that happens
6550                  * assume we don't have to enable/disable DDR DVFS
6551                  * dynamically. To test that just set the REQ_ACK
6552                  * bit to poke the Punit, but don't change the
6553                  * HIGH/LOW bits so that we don't actually change
6554                  * the current state.
6555                  */
6556                 val = vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2);
6557                 val |= FORCE_DDR_FREQ_REQ_ACK;
6558                 vlv_punit_write(dev_priv, PUNIT_REG_DDR_SETUP2, val);
6559
6560                 if (wait_for((vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2) &
6561                               FORCE_DDR_FREQ_REQ_ACK) == 0, 3)) {
6562                         drm_dbg_kms(&dev_priv->drm,
6563                                     "Punit not acking DDR DVFS request, "
6564                                     "assuming DDR DVFS is disabled\n");
6565                         dev_priv->wm.max_level = VLV_WM_LEVEL_PM5;
6566                 } else {
6567                         val = vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2);
6568                         if ((val & FORCE_DDR_HIGH_FREQ) == 0)
6569                                 wm->level = VLV_WM_LEVEL_DDR_DVFS;
6570                 }
6571
6572                 vlv_punit_put(dev_priv);
6573         }
6574
6575         for_each_intel_crtc(&dev_priv->drm, crtc) {
6576                 struct intel_crtc_state *crtc_state =
6577                         to_intel_crtc_state(crtc->base.state);
6578                 struct vlv_wm_state *active = &crtc->wm.active.vlv;
6579                 const struct vlv_fifo_state *fifo_state =
6580                         &crtc_state->wm.vlv.fifo_state;
6581                 enum pipe pipe = crtc->pipe;
6582                 enum plane_id plane_id;
6583                 int level;
6584
6585                 vlv_get_fifo_size(crtc_state);
6586
6587                 active->num_levels = wm->level + 1;
6588                 active->cxsr = wm->cxsr;
6589
6590                 for (level = 0; level < active->num_levels; level++) {
6591                         struct g4x_pipe_wm *raw =
6592                                 &crtc_state->wm.vlv.raw[level];
6593
6594                         active->sr[level].plane = wm->sr.plane;
6595                         active->sr[level].cursor = wm->sr.cursor;
6596
6597                         for_each_plane_id_on_crtc(crtc, plane_id) {
6598                                 active->wm[level].plane[plane_id] =
6599                                         wm->pipe[pipe].plane[plane_id];
6600
6601                                 raw->plane[plane_id] =
6602                                         vlv_invert_wm_value(active->wm[level].plane[plane_id],
6603                                                             fifo_state->plane[plane_id]);
6604                         }
6605                 }
6606
6607                 for_each_plane_id_on_crtc(crtc, plane_id)
6608                         vlv_raw_plane_wm_set(crtc_state, level,
6609                                              plane_id, USHRT_MAX);
6610                 vlv_invalidate_wms(crtc, active, level);
6611
6612                 crtc_state->wm.vlv.optimal = *active;
6613                 crtc_state->wm.vlv.intermediate = *active;
6614
6615                 drm_dbg_kms(&dev_priv->drm,
6616                             "Initial watermarks: pipe %c, plane=%d, cursor=%d, sprite0=%d, sprite1=%d\n",
6617                             pipe_name(pipe),
6618                             wm->pipe[pipe].plane[PLANE_PRIMARY],
6619                             wm->pipe[pipe].plane[PLANE_CURSOR],
6620                             wm->pipe[pipe].plane[PLANE_SPRITE0],
6621                             wm->pipe[pipe].plane[PLANE_SPRITE1]);
6622         }
6623
6624         drm_dbg_kms(&dev_priv->drm,
6625                     "Initial watermarks: SR plane=%d, SR cursor=%d level=%d cxsr=%d\n",
6626                     wm->sr.plane, wm->sr.cursor, wm->level, wm->cxsr);
6627 }
6628
6629 void vlv_wm_sanitize(struct drm_i915_private *dev_priv)
6630 {
6631         struct intel_plane *plane;
6632         struct intel_crtc *crtc;
6633
6634         mutex_lock(&dev_priv->wm.wm_mutex);
6635
6636         for_each_intel_plane(&dev_priv->drm, plane) {
6637                 struct intel_crtc *crtc =
6638                         intel_get_crtc_for_pipe(dev_priv, plane->pipe);
6639                 struct intel_crtc_state *crtc_state =
6640                         to_intel_crtc_state(crtc->base.state);
6641                 struct intel_plane_state *plane_state =
6642                         to_intel_plane_state(plane->base.state);
6643                 struct vlv_wm_state *wm_state = &crtc_state->wm.vlv.optimal;
6644                 const struct vlv_fifo_state *fifo_state =
6645                         &crtc_state->wm.vlv.fifo_state;
6646                 enum plane_id plane_id = plane->id;
6647                 int level;
6648
6649                 if (plane_state->uapi.visible)
6650                         continue;
6651
6652                 for (level = 0; level < wm_state->num_levels; level++) {
6653                         struct g4x_pipe_wm *raw =
6654                                 &crtc_state->wm.vlv.raw[level];
6655
6656                         raw->plane[plane_id] = 0;
6657
6658                         wm_state->wm[level].plane[plane_id] =
6659                                 vlv_invert_wm_value(raw->plane[plane_id],
6660                                                     fifo_state->plane[plane_id]);
6661                 }
6662         }
6663
6664         for_each_intel_crtc(&dev_priv->drm, crtc) {
6665                 struct intel_crtc_state *crtc_state =
6666                         to_intel_crtc_state(crtc->base.state);
6667
6668                 crtc_state->wm.vlv.intermediate =
6669                         crtc_state->wm.vlv.optimal;
6670                 crtc->wm.active.vlv = crtc_state->wm.vlv.optimal;
6671         }
6672
6673         vlv_program_watermarks(dev_priv);
6674
6675         mutex_unlock(&dev_priv->wm.wm_mutex);
6676 }
6677
6678 /*
6679  * FIXME should probably kill this and improve
6680  * the real watermark readout/sanitation instead
6681  */
6682 static void ilk_init_lp_watermarks(struct drm_i915_private *dev_priv)
6683 {
6684         intel_uncore_write(&dev_priv->uncore, WM3_LP_ILK, intel_uncore_read(&dev_priv->uncore, WM3_LP_ILK) & ~WM1_LP_SR_EN);
6685         intel_uncore_write(&dev_priv->uncore, WM2_LP_ILK, intel_uncore_read(&dev_priv->uncore, WM2_LP_ILK) & ~WM1_LP_SR_EN);
6686         intel_uncore_write(&dev_priv->uncore, WM1_LP_ILK, intel_uncore_read(&dev_priv->uncore, WM1_LP_ILK) & ~WM1_LP_SR_EN);
6687
6688         /*
6689          * Don't touch WM1S_LP_EN here.
6690          * Doing so could cause underruns.
6691          */
6692 }
6693
6694 void ilk_wm_get_hw_state(struct drm_i915_private *dev_priv)
6695 {
6696         struct ilk_wm_values *hw = &dev_priv->wm.hw;
6697         struct intel_crtc *crtc;
6698
6699         ilk_init_lp_watermarks(dev_priv);
6700
6701         for_each_intel_crtc(&dev_priv->drm, crtc)
6702                 ilk_pipe_wm_get_hw_state(crtc);
6703
6704         hw->wm_lp[0] = intel_uncore_read(&dev_priv->uncore, WM1_LP_ILK);
6705         hw->wm_lp[1] = intel_uncore_read(&dev_priv->uncore, WM2_LP_ILK);
6706         hw->wm_lp[2] = intel_uncore_read(&dev_priv->uncore, WM3_LP_ILK);
6707
6708         hw->wm_lp_spr[0] = intel_uncore_read(&dev_priv->uncore, WM1S_LP_ILK);
6709         if (INTEL_GEN(dev_priv) >= 7) {
6710                 hw->wm_lp_spr[1] = intel_uncore_read(&dev_priv->uncore, WM2S_LP_IVB);
6711                 hw->wm_lp_spr[2] = intel_uncore_read(&dev_priv->uncore, WM3S_LP_IVB);
6712         }
6713
6714         if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
6715                 hw->partitioning = (intel_uncore_read(&dev_priv->uncore, WM_MISC) & WM_MISC_DATA_PARTITION_5_6) ?
6716                         INTEL_DDB_PART_5_6 : INTEL_DDB_PART_1_2;
6717         else if (IS_IVYBRIDGE(dev_priv))
6718                 hw->partitioning = (intel_uncore_read(&dev_priv->uncore, DISP_ARB_CTL2) & DISP_DATA_PARTITION_5_6) ?
6719                         INTEL_DDB_PART_5_6 : INTEL_DDB_PART_1_2;
6720
6721         hw->enable_fbc_wm =
6722                 !(intel_uncore_read(&dev_priv->uncore, DISP_ARB_CTL) & DISP_FBC_WM_DIS);
6723 }
6724
6725 /**
6726  * intel_update_watermarks - update FIFO watermark values based on current modes
6727  * @crtc: the #intel_crtc on which to compute the WM
6728  *
6729  * Calculate watermark values for the various WM regs based on current mode
6730  * and plane configuration.
6731  *
6732  * There are several cases to deal with here:
6733  *   - normal (i.e. non-self-refresh)
6734  *   - self-refresh (SR) mode
6735  *   - lines are large relative to FIFO size (buffer can hold up to 2)
6736  *   - lines are small relative to FIFO size (buffer can hold more than 2
6737  *     lines), so need to account for TLB latency
6738  *
6739  *   The normal calculation is:
6740  *     watermark = dotclock * bytes per pixel * latency
6741  *   where latency is platform & configuration dependent (we assume pessimal
6742  *   values here).
6743  *
6744  *   The SR calculation is:
6745  *     watermark = (trunc(latency/line time)+1) * surface width *
6746  *       bytes per pixel
6747  *   where
6748  *     line time = htotal / dotclock
6749  *     surface width = hdisplay for normal plane and 64 for cursor
6750  *   and latency is assumed to be high, as above.
6751  *
6752  * The final value programmed to the register should always be rounded up,
6753  * and include an extra 2 entries to account for clock crossings.
6754  *
6755  * We don't use the sprite, so we can ignore that.  And on Crestline we have
6756  * to set the non-SR watermarks to 8.
6757  */
6758 void intel_update_watermarks(struct intel_crtc *crtc)
6759 {
6760         struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
6761
6762         if (dev_priv->display.update_wm)
6763                 dev_priv->display.update_wm(crtc);
6764 }
6765
6766 void intel_enable_ipc(struct drm_i915_private *dev_priv)
6767 {
6768         u32 val;
6769
6770         if (!HAS_IPC(dev_priv))
6771                 return;
6772
6773         val = intel_uncore_read(&dev_priv->uncore, DISP_ARB_CTL2);
6774
6775         if (dev_priv->ipc_enabled)
6776                 val |= DISP_IPC_ENABLE;
6777         else
6778                 val &= ~DISP_IPC_ENABLE;
6779
6780         intel_uncore_write(&dev_priv->uncore, DISP_ARB_CTL2, val);
6781 }
6782
6783 static bool intel_can_enable_ipc(struct drm_i915_private *dev_priv)
6784 {
6785         /* Display WA #0477 WaDisableIPC: skl */
6786         if (IS_SKYLAKE(dev_priv))
6787                 return false;
6788
6789         /* Display WA #1141: SKL:all KBL:all CFL */
6790         if (IS_KABYLAKE(dev_priv) ||
6791             IS_COFFEELAKE(dev_priv) ||
6792             IS_COMETLAKE(dev_priv))
6793                 return dev_priv->dram_info.symmetric_memory;
6794
6795         return true;
6796 }
6797
6798 void intel_init_ipc(struct drm_i915_private *dev_priv)
6799 {
6800         if (!HAS_IPC(dev_priv))
6801                 return;
6802
6803         dev_priv->ipc_enabled = intel_can_enable_ipc(dev_priv);
6804
6805         intel_enable_ipc(dev_priv);
6806 }
6807
6808 static void ibx_init_clock_gating(struct drm_i915_private *dev_priv)
6809 {
6810         /*
6811          * On Ibex Peak and Cougar Point, we need to disable clock
6812          * gating for the panel power sequencer or it will fail to
6813          * start up when no ports are active.
6814          */
6815         intel_uncore_write(&dev_priv->uncore, SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE);
6816 }
6817
6818 static void g4x_disable_trickle_feed(struct drm_i915_private *dev_priv)
6819 {
6820         enum pipe pipe;
6821
6822         for_each_pipe(dev_priv, pipe) {
6823                 intel_uncore_write(&dev_priv->uncore, DSPCNTR(pipe),
6824                            intel_uncore_read(&dev_priv->uncore, DSPCNTR(pipe)) |
6825                            DISPPLANE_TRICKLE_FEED_DISABLE);
6826
6827                 intel_uncore_write(&dev_priv->uncore, DSPSURF(pipe), intel_uncore_read(&dev_priv->uncore, DSPSURF(pipe)));
6828                 intel_uncore_posting_read(&dev_priv->uncore, DSPSURF(pipe));
6829         }
6830 }
6831
6832 static void ilk_init_clock_gating(struct drm_i915_private *dev_priv)
6833 {
6834         u32 dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
6835
6836         /*
6837          * Required for FBC
6838          * WaFbcDisableDpfcClockGating:ilk
6839          */
6840         dspclk_gate |= ILK_DPFCRUNIT_CLOCK_GATE_DISABLE |
6841                    ILK_DPFCUNIT_CLOCK_GATE_DISABLE |
6842                    ILK_DPFDUNIT_CLOCK_GATE_ENABLE;
6843
6844         intel_uncore_write(&dev_priv->uncore, PCH_3DCGDIS0,
6845                    MARIUNIT_CLOCK_GATE_DISABLE |
6846                    SVSMUNIT_CLOCK_GATE_DISABLE);
6847         intel_uncore_write(&dev_priv->uncore, PCH_3DCGDIS1,
6848                    VFMUNIT_CLOCK_GATE_DISABLE);
6849
6850         /*
6851          * According to the spec the following bits should be set in
6852          * order to enable memory self-refresh
6853          * The bit 22/21 of 0x42004
6854          * The bit 5 of 0x42020
6855          * The bit 15 of 0x45000
6856          */
6857         intel_uncore_write(&dev_priv->uncore, ILK_DISPLAY_CHICKEN2,
6858                    (intel_uncore_read(&dev_priv->uncore, ILK_DISPLAY_CHICKEN2) |
6859                     ILK_DPARB_GATE | ILK_VSDPFD_FULL));
6860         dspclk_gate |= ILK_DPARBUNIT_CLOCK_GATE_ENABLE;
6861         intel_uncore_write(&dev_priv->uncore, DISP_ARB_CTL,
6862                    (intel_uncore_read(&dev_priv->uncore, DISP_ARB_CTL) |
6863                     DISP_FBC_WM_DIS));
6864
6865         /*
6866          * Based on the document from hardware guys the following bits
6867          * should be set unconditionally in order to enable FBC.
6868          * The bit 22 of 0x42000
6869          * The bit 22 of 0x42004
6870          * The bit 7,8,9 of 0x42020.
6871          */
6872         if (IS_IRONLAKE_M(dev_priv)) {
6873                 /* WaFbcAsynchFlipDisableFbcQueue:ilk */
6874                 intel_uncore_write(&dev_priv->uncore, ILK_DISPLAY_CHICKEN1,
6875                            intel_uncore_read(&dev_priv->uncore, ILK_DISPLAY_CHICKEN1) |
6876                            ILK_FBCQ_DIS);
6877                 intel_uncore_write(&dev_priv->uncore, ILK_DISPLAY_CHICKEN2,
6878                            intel_uncore_read(&dev_priv->uncore, ILK_DISPLAY_CHICKEN2) |
6879                            ILK_DPARB_GATE);
6880         }
6881
6882         intel_uncore_write(&dev_priv->uncore, ILK_DSPCLK_GATE_D, dspclk_gate);
6883
6884         intel_uncore_write(&dev_priv->uncore, ILK_DISPLAY_CHICKEN2,
6885                    intel_uncore_read(&dev_priv->uncore, ILK_DISPLAY_CHICKEN2) |
6886                    ILK_ELPIN_409_SELECT);
6887
6888         g4x_disable_trickle_feed(dev_priv);
6889
6890         ibx_init_clock_gating(dev_priv);
6891 }
6892
6893 static void cpt_init_clock_gating(struct drm_i915_private *dev_priv)
6894 {
6895         enum pipe pipe;
6896         u32 val;
6897
6898         /*
6899          * On Ibex Peak and Cougar Point, we need to disable clock
6900          * gating for the panel power sequencer or it will fail to
6901          * start up when no ports are active.
6902          */
6903         intel_uncore_write(&dev_priv->uncore, SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE |
6904                    PCH_DPLUNIT_CLOCK_GATE_DISABLE |
6905                    PCH_CPUNIT_CLOCK_GATE_DISABLE);
6906         intel_uncore_write(&dev_priv->uncore, SOUTH_CHICKEN2, intel_uncore_read(&dev_priv->uncore, SOUTH_CHICKEN2) |
6907                    DPLS_EDP_PPS_FIX_DIS);
6908         /* The below fixes the weird display corruption, a few pixels shifted
6909          * downward, on (only) LVDS of some HP laptops with IVY.
6910          */
6911         for_each_pipe(dev_priv, pipe) {
6912                 val = intel_uncore_read(&dev_priv->uncore, TRANS_CHICKEN2(pipe));
6913                 val |= TRANS_CHICKEN2_TIMING_OVERRIDE;
6914                 val &= ~TRANS_CHICKEN2_FDI_POLARITY_REVERSED;
6915                 if (dev_priv->vbt.fdi_rx_polarity_inverted)
6916                         val |= TRANS_CHICKEN2_FDI_POLARITY_REVERSED;
6917                 val &= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_COUNTER;
6918                 val &= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_MODESWITCH;
6919                 intel_uncore_write(&dev_priv->uncore, TRANS_CHICKEN2(pipe), val);
6920         }
6921         /* WADP0ClockGatingDisable */
6922         for_each_pipe(dev_priv, pipe) {
6923                 intel_uncore_write(&dev_priv->uncore, TRANS_CHICKEN1(pipe),
6924                            TRANS_CHICKEN1_DP0UNIT_GC_DISABLE);
6925         }
6926 }
6927
6928 static void gen6_check_mch_setup(struct drm_i915_private *dev_priv)
6929 {
6930         u32 tmp;
6931
6932         tmp = intel_uncore_read(&dev_priv->uncore, MCH_SSKPD);
6933         if ((tmp & MCH_SSKPD_WM0_MASK) != MCH_SSKPD_WM0_VAL)
6934                 drm_dbg_kms(&dev_priv->drm,
6935                             "Wrong MCH_SSKPD value: 0x%08x This can cause underruns.\n",
6936                             tmp);
6937 }
6938
6939 static void gen6_init_clock_gating(struct drm_i915_private *dev_priv)
6940 {
6941         u32 dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
6942
6943         intel_uncore_write(&dev_priv->uncore, ILK_DSPCLK_GATE_D, dspclk_gate);
6944
6945         intel_uncore_write(&dev_priv->uncore, ILK_DISPLAY_CHICKEN2,
6946                    intel_uncore_read(&dev_priv->uncore, ILK_DISPLAY_CHICKEN2) |
6947                    ILK_ELPIN_409_SELECT);
6948
6949         intel_uncore_write(&dev_priv->uncore, GEN6_UCGCTL1,
6950                    intel_uncore_read(&dev_priv->uncore, GEN6_UCGCTL1) |
6951                    GEN6_BLBUNIT_CLOCK_GATE_DISABLE |
6952                    GEN6_CSUNIT_CLOCK_GATE_DISABLE);
6953
6954         /* According to the BSpec vol1g, bit 12 (RCPBUNIT) clock
6955          * gating disable must be set.  Failure to set it results in
6956          * flickering pixels due to Z write ordering failures after
6957          * some amount of runtime in the Mesa "fire" demo, and Unigine
6958          * Sanctuary and Tropics, and apparently anything else with
6959          * alpha test or pixel discard.
6960          *
6961          * According to the spec, bit 11 (RCCUNIT) must also be set,
6962          * but we didn't debug actual testcases to find it out.
6963          *
6964          * WaDisableRCCUnitClockGating:snb
6965          * WaDisableRCPBUnitClockGating:snb
6966          */
6967         intel_uncore_write(&dev_priv->uncore, GEN6_UCGCTL2,
6968                    GEN6_RCPBUNIT_CLOCK_GATE_DISABLE |
6969                    GEN6_RCCUNIT_CLOCK_GATE_DISABLE);
6970
6971         /*
6972          * According to the spec the following bits should be
6973          * set in order to enable memory self-refresh and fbc:
6974          * The bit21 and bit22 of 0x42000
6975          * The bit21 and bit22 of 0x42004
6976          * The bit5 and bit7 of 0x42020
6977          * The bit14 of 0x70180
6978          * The bit14 of 0x71180
6979          *
6980          * WaFbcAsynchFlipDisableFbcQueue:snb
6981          */
6982         intel_uncore_write(&dev_priv->uncore, ILK_DISPLAY_CHICKEN1,
6983                    intel_uncore_read(&dev_priv->uncore, ILK_DISPLAY_CHICKEN1) |
6984                    ILK_FBCQ_DIS | ILK_PABSTRETCH_DIS);
6985         intel_uncore_write(&dev_priv->uncore, ILK_DISPLAY_CHICKEN2,
6986                    intel_uncore_read(&dev_priv->uncore, ILK_DISPLAY_CHICKEN2) |
6987                    ILK_DPARB_GATE | ILK_VSDPFD_FULL);
6988         intel_uncore_write(&dev_priv->uncore, ILK_DSPCLK_GATE_D,
6989                    intel_uncore_read(&dev_priv->uncore, ILK_DSPCLK_GATE_D) |
6990                    ILK_DPARBUNIT_CLOCK_GATE_ENABLE  |
6991                    ILK_DPFDUNIT_CLOCK_GATE_ENABLE);
6992
6993         g4x_disable_trickle_feed(dev_priv);
6994
6995         cpt_init_clock_gating(dev_priv);
6996
6997         gen6_check_mch_setup(dev_priv);
6998 }
6999
7000 static void lpt_init_clock_gating(struct drm_i915_private *dev_priv)
7001 {
7002         /*
7003          * TODO: this bit should only be enabled when really needed, then
7004          * disabled when not needed anymore in order to save power.
7005          */
7006         if (HAS_PCH_LPT_LP(dev_priv))
7007                 intel_uncore_write(&dev_priv->uncore, SOUTH_DSPCLK_GATE_D,
7008                            intel_uncore_read(&dev_priv->uncore, SOUTH_DSPCLK_GATE_D) |
7009                            PCH_LP_PARTITION_LEVEL_DISABLE);
7010
7011         /* WADPOClockGatingDisable:hsw */
7012         intel_uncore_write(&dev_priv->uncore, TRANS_CHICKEN1(PIPE_A),
7013                    intel_uncore_read(&dev_priv->uncore, TRANS_CHICKEN1(PIPE_A)) |
7014                    TRANS_CHICKEN1_DP0UNIT_GC_DISABLE);
7015 }
7016
7017 static void lpt_suspend_hw(struct drm_i915_private *dev_priv)
7018 {
7019         if (HAS_PCH_LPT_LP(dev_priv)) {
7020                 u32 val = intel_uncore_read(&dev_priv->uncore, SOUTH_DSPCLK_GATE_D);
7021
7022                 val &= ~PCH_LP_PARTITION_LEVEL_DISABLE;
7023                 intel_uncore_write(&dev_priv->uncore, SOUTH_DSPCLK_GATE_D, val);
7024         }
7025 }
7026
7027 static void gen8_set_l3sqc_credits(struct drm_i915_private *dev_priv,
7028                                    int general_prio_credits,
7029                                    int high_prio_credits)
7030 {
7031         u32 misccpctl;
7032         u32 val;
7033
7034         /* WaTempDisableDOPClkGating:bdw */
7035         misccpctl = intel_uncore_read(&dev_priv->uncore, GEN7_MISCCPCTL);
7036         intel_uncore_write(&dev_priv->uncore, GEN7_MISCCPCTL, misccpctl & ~GEN7_DOP_CLOCK_GATE_ENABLE);
7037
7038         val = intel_uncore_read(&dev_priv->uncore, GEN8_L3SQCREG1);
7039         val &= ~L3_PRIO_CREDITS_MASK;
7040         val |= L3_GENERAL_PRIO_CREDITS(general_prio_credits);
7041         val |= L3_HIGH_PRIO_CREDITS(high_prio_credits);
7042         intel_uncore_write(&dev_priv->uncore, GEN8_L3SQCREG1, val);
7043
7044         /*
7045          * Wait at least 100 clocks before re-enabling clock gating.
7046          * See the definition of L3SQCREG1 in BSpec.
7047          */
7048         intel_uncore_posting_read(&dev_priv->uncore, GEN8_L3SQCREG1);
7049         udelay(1);
7050         intel_uncore_write(&dev_priv->uncore, GEN7_MISCCPCTL, misccpctl);
7051 }
7052
7053 static void icl_init_clock_gating(struct drm_i915_private *dev_priv)
7054 {
7055         /* Wa_1409120013:icl,ehl */
7056         intel_uncore_write(&dev_priv->uncore, ILK_DPFC_CHICKEN,
7057                    ILK_DPFC_CHICKEN_COMP_DUMMY_PIXEL);
7058
7059         /* This is not an Wa. Enable to reduce Sampler power */
7060         intel_uncore_write(&dev_priv->uncore, GEN10_DFR_RATIO_EN_AND_CHICKEN,
7061                    intel_uncore_read(&dev_priv->uncore, GEN10_DFR_RATIO_EN_AND_CHICKEN) & ~DFR_DISABLE);
7062
7063         /*Wa_14010594013:icl, ehl */
7064         intel_uncore_rmw(&dev_priv->uncore, GEN8_CHICKEN_DCPR_1,
7065                          0, CNL_DELAY_PMRSP);
7066 }
7067
7068 static void gen12lp_init_clock_gating(struct drm_i915_private *dev_priv)
7069 {
7070         /* Wa_1409120013:tgl,rkl,adl_s,dg1 */
7071         intel_uncore_write(&dev_priv->uncore, ILK_DPFC_CHICKEN,
7072                            ILK_DPFC_CHICKEN_COMP_DUMMY_PIXEL);
7073
7074         /* Wa_1409825376:tgl (pre-prod)*/
7075         if (IS_TGL_DISP_REVID(dev_priv, TGL_REVID_A0, TGL_REVID_B1))
7076                 intel_uncore_write(&dev_priv->uncore, GEN9_CLKGATE_DIS_3, intel_uncore_read(&dev_priv->uncore, GEN9_CLKGATE_DIS_3) |
7077                            TGL_VRH_GATING_DIS);
7078
7079         /* Wa_14011059788:tgl,rkl,adl_s,dg1 */
7080         intel_uncore_rmw(&dev_priv->uncore, GEN10_DFR_RATIO_EN_AND_CHICKEN,
7081                          0, DFR_DISABLE);
7082 }
7083
7084 static void dg1_init_clock_gating(struct drm_i915_private *dev_priv)
7085 {
7086         gen12lp_init_clock_gating(dev_priv);
7087
7088         /* Wa_1409836686:dg1[a0] */
7089         if (IS_DG1_REVID(dev_priv, DG1_REVID_A0, DG1_REVID_A0))
7090                 intel_uncore_write(&dev_priv->uncore, GEN9_CLKGATE_DIS_3, intel_uncore_read(&dev_priv->uncore, GEN9_CLKGATE_DIS_3) |
7091                            DPT_GATING_DIS);
7092 }
7093
7094 static void cnp_init_clock_gating(struct drm_i915_private *dev_priv)
7095 {
7096         if (!HAS_PCH_CNP(dev_priv))
7097                 return;
7098
7099         /* Display WA #1181 WaSouthDisplayDisablePWMCGEGating: cnp */
7100         intel_uncore_write(&dev_priv->uncore, SOUTH_DSPCLK_GATE_D, intel_uncore_read(&dev_priv->uncore, SOUTH_DSPCLK_GATE_D) |
7101                    CNP_PWM_CGE_GATING_DISABLE);
7102 }
7103
7104 static void cnl_init_clock_gating(struct drm_i915_private *dev_priv)
7105 {
7106         u32 val;
7107         cnp_init_clock_gating(dev_priv);
7108
7109         /* This is not an Wa. Enable for better image quality */
7110         intel_uncore_write(&dev_priv->uncore, _3D_CHICKEN3,
7111                    _MASKED_BIT_ENABLE(_3D_CHICKEN3_AA_LINE_QUALITY_FIX_ENABLE));
7112
7113         /* WaEnableChickenDCPR:cnl */
7114         intel_uncore_write(&dev_priv->uncore, GEN8_CHICKEN_DCPR_1,
7115                    intel_uncore_read(&dev_priv->uncore, GEN8_CHICKEN_DCPR_1) | MASK_WAKEMEM);
7116
7117         /*
7118          * WaFbcWakeMemOn:cnl
7119          * Display WA #0859: cnl
7120          */
7121         intel_uncore_write(&dev_priv->uncore, DISP_ARB_CTL, intel_uncore_read(&dev_priv->uncore, DISP_ARB_CTL) |
7122                    DISP_FBC_MEMORY_WAKE);
7123
7124         val = intel_uncore_read(&dev_priv->uncore, SLICE_UNIT_LEVEL_CLKGATE);
7125         /* ReadHitWriteOnlyDisable:cnl */
7126         val |= RCCUNIT_CLKGATE_DIS;
7127         intel_uncore_write(&dev_priv->uncore, SLICE_UNIT_LEVEL_CLKGATE, val);
7128
7129         /* Wa_2201832410:cnl */
7130         val = intel_uncore_read(&dev_priv->uncore, SUBSLICE_UNIT_LEVEL_CLKGATE);
7131         val |= GWUNIT_CLKGATE_DIS;
7132         intel_uncore_write(&dev_priv->uncore, SUBSLICE_UNIT_LEVEL_CLKGATE, val);
7133
7134         /* WaDisableVFclkgate:cnl */
7135         /* WaVFUnitClockGatingDisable:cnl */
7136         val = intel_uncore_read(&dev_priv->uncore, UNSLICE_UNIT_LEVEL_CLKGATE);
7137         val |= VFUNIT_CLKGATE_DIS;
7138         intel_uncore_write(&dev_priv->uncore, UNSLICE_UNIT_LEVEL_CLKGATE, val);
7139 }
7140
7141 static void cfl_init_clock_gating(struct drm_i915_private *dev_priv)
7142 {
7143         cnp_init_clock_gating(dev_priv);
7144         gen9_init_clock_gating(dev_priv);
7145
7146         /* WAC6entrylatency:cfl */
7147         intel_uncore_write(&dev_priv->uncore, FBC_LLC_READ_CTRL, intel_uncore_read(&dev_priv->uncore, FBC_LLC_READ_CTRL) |
7148                    FBC_LLC_FULLY_OPEN);
7149
7150         /*
7151          * WaFbcTurnOffFbcWatermark:cfl
7152          * Display WA #0562: cfl
7153          */
7154         intel_uncore_write(&dev_priv->uncore, DISP_ARB_CTL, intel_uncore_read(&dev_priv->uncore, DISP_ARB_CTL) |
7155                    DISP_FBC_WM_DIS);
7156
7157         /*
7158          * WaFbcNukeOnHostModify:cfl
7159          * Display WA #0873: cfl
7160          */
7161         intel_uncore_write(&dev_priv->uncore, ILK_DPFC_CHICKEN, intel_uncore_read(&dev_priv->uncore, ILK_DPFC_CHICKEN) |
7162                    ILK_DPFC_NUKE_ON_ANY_MODIFICATION);
7163 }
7164
7165 static void kbl_init_clock_gating(struct drm_i915_private *dev_priv)
7166 {
7167         gen9_init_clock_gating(dev_priv);
7168
7169         /* WAC6entrylatency:kbl */
7170         intel_uncore_write(&dev_priv->uncore, FBC_LLC_READ_CTRL, intel_uncore_read(&dev_priv->uncore, FBC_LLC_READ_CTRL) |
7171                    FBC_LLC_FULLY_OPEN);
7172
7173         /* WaDisableSDEUnitClockGating:kbl */
7174         if (IS_KBL_GT_REVID(dev_priv, 0, KBL_REVID_B0))
7175                 intel_uncore_write(&dev_priv->uncore, GEN8_UCGCTL6, intel_uncore_read(&dev_priv->uncore, GEN8_UCGCTL6) |
7176                            GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
7177
7178         /* WaDisableGamClockGating:kbl */
7179         if (IS_KBL_GT_REVID(dev_priv, 0, KBL_REVID_B0))
7180                 intel_uncore_write(&dev_priv->uncore, GEN6_UCGCTL1, intel_uncore_read(&dev_priv->uncore, GEN6_UCGCTL1) |
7181                            GEN6_GAMUNIT_CLOCK_GATE_DISABLE);
7182
7183         /*
7184          * WaFbcTurnOffFbcWatermark:kbl
7185          * Display WA #0562: kbl
7186          */
7187         intel_uncore_write(&dev_priv->uncore, DISP_ARB_CTL, intel_uncore_read(&dev_priv->uncore, DISP_ARB_CTL) |
7188                    DISP_FBC_WM_DIS);
7189
7190         /*
7191          * WaFbcNukeOnHostModify:kbl
7192          * Display WA #0873: kbl
7193          */
7194         intel_uncore_write(&dev_priv->uncore, ILK_DPFC_CHICKEN, intel_uncore_read(&dev_priv->uncore, ILK_DPFC_CHICKEN) |
7195                    ILK_DPFC_NUKE_ON_ANY_MODIFICATION);
7196 }
7197
7198 static void skl_init_clock_gating(struct drm_i915_private *dev_priv)
7199 {
7200         gen9_init_clock_gating(dev_priv);
7201
7202         /* WaDisableDopClockGating:skl */
7203         intel_uncore_write(&dev_priv->uncore, GEN7_MISCCPCTL, intel_uncore_read(&dev_priv->uncore, GEN7_MISCCPCTL) &
7204                    ~GEN7_DOP_CLOCK_GATE_ENABLE);
7205
7206         /* WAC6entrylatency:skl */
7207         intel_uncore_write(&dev_priv->uncore, FBC_LLC_READ_CTRL, intel_uncore_read(&dev_priv->uncore, FBC_LLC_READ_CTRL) |
7208                    FBC_LLC_FULLY_OPEN);
7209
7210         /*
7211          * WaFbcTurnOffFbcWatermark:skl
7212          * Display WA #0562: skl
7213          */
7214         intel_uncore_write(&dev_priv->uncore, DISP_ARB_CTL, intel_uncore_read(&dev_priv->uncore, DISP_ARB_CTL) |
7215                    DISP_FBC_WM_DIS);
7216
7217         /*
7218          * WaFbcNukeOnHostModify:skl
7219          * Display WA #0873: skl
7220          */
7221         intel_uncore_write(&dev_priv->uncore, ILK_DPFC_CHICKEN, intel_uncore_read(&dev_priv->uncore, ILK_DPFC_CHICKEN) |
7222                    ILK_DPFC_NUKE_ON_ANY_MODIFICATION);
7223
7224         /*
7225          * WaFbcHighMemBwCorruptionAvoidance:skl
7226          * Display WA #0883: skl
7227          */
7228         intel_uncore_write(&dev_priv->uncore, ILK_DPFC_CHICKEN, intel_uncore_read(&dev_priv->uncore, ILK_DPFC_CHICKEN) |
7229                    ILK_DPFC_DISABLE_DUMMY0);
7230 }
7231
7232 static void bdw_init_clock_gating(struct drm_i915_private *dev_priv)
7233 {
7234         enum pipe pipe;
7235
7236         /* WaFbcAsynchFlipDisableFbcQueue:hsw,bdw */
7237         intel_uncore_write(&dev_priv->uncore, CHICKEN_PIPESL_1(PIPE_A),
7238                    intel_uncore_read(&dev_priv->uncore, CHICKEN_PIPESL_1(PIPE_A)) |
7239                    HSW_FBCQ_DIS);
7240
7241         /* WaSwitchSolVfFArbitrationPriority:bdw */
7242         intel_uncore_write(&dev_priv->uncore, GAM_ECOCHK, intel_uncore_read(&dev_priv->uncore, GAM_ECOCHK) | HSW_ECOCHK_ARB_PRIO_SOL);
7243
7244         /* WaPsrDPAMaskVBlankInSRD:bdw */
7245         intel_uncore_write(&dev_priv->uncore, CHICKEN_PAR1_1,
7246                    intel_uncore_read(&dev_priv->uncore, CHICKEN_PAR1_1) | DPA_MASK_VBLANK_SRD);
7247
7248         /* WaPsrDPRSUnmaskVBlankInSRD:bdw */
7249         for_each_pipe(dev_priv, pipe) {
7250                 intel_uncore_write(&dev_priv->uncore, CHICKEN_PIPESL_1(pipe),
7251                            intel_uncore_read(&dev_priv->uncore, CHICKEN_PIPESL_1(pipe)) |
7252                            BDW_DPRS_MASK_VBLANK_SRD);
7253         }
7254
7255         /* WaVSRefCountFullforceMissDisable:bdw */
7256         /* WaDSRefCountFullforceMissDisable:bdw */
7257         intel_uncore_write(&dev_priv->uncore, GEN7_FF_THREAD_MODE,
7258                    intel_uncore_read(&dev_priv->uncore, GEN7_FF_THREAD_MODE) &
7259                    ~(GEN8_FF_DS_REF_CNT_FFME | GEN7_FF_VS_REF_CNT_FFME));
7260
7261         intel_uncore_write(&dev_priv->uncore, GEN6_RC_SLEEP_PSMI_CONTROL,
7262                    _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE));
7263
7264         /* WaDisableSDEUnitClockGating:bdw */
7265         intel_uncore_write(&dev_priv->uncore, GEN8_UCGCTL6, intel_uncore_read(&dev_priv->uncore, GEN8_UCGCTL6) |
7266                    GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
7267
7268         /* WaProgramL3SqcReg1Default:bdw */
7269         gen8_set_l3sqc_credits(dev_priv, 30, 2);
7270
7271         /* WaKVMNotificationOnConfigChange:bdw */
7272         intel_uncore_write(&dev_priv->uncore, CHICKEN_PAR2_1, intel_uncore_read(&dev_priv->uncore, CHICKEN_PAR2_1)
7273                    | KVM_CONFIG_CHANGE_NOTIFICATION_SELECT);
7274
7275         lpt_init_clock_gating(dev_priv);
7276
7277         /* WaDisableDopClockGating:bdw
7278          *
7279          * Also see the CHICKEN2 write in bdw_init_workarounds() to disable DOP
7280          * clock gating.
7281          */
7282         intel_uncore_write(&dev_priv->uncore, GEN6_UCGCTL1,
7283                    intel_uncore_read(&dev_priv->uncore, GEN6_UCGCTL1) | GEN6_EU_TCUNIT_CLOCK_GATE_DISABLE);
7284 }
7285
7286 static void hsw_init_clock_gating(struct drm_i915_private *dev_priv)
7287 {
7288         /* WaFbcAsynchFlipDisableFbcQueue:hsw,bdw */
7289         intel_uncore_write(&dev_priv->uncore, CHICKEN_PIPESL_1(PIPE_A),
7290                    intel_uncore_read(&dev_priv->uncore, CHICKEN_PIPESL_1(PIPE_A)) |
7291                    HSW_FBCQ_DIS);
7292
7293         /* This is required by WaCatErrorRejectionIssue:hsw */
7294         intel_uncore_write(&dev_priv->uncore, GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
7295                    intel_uncore_read(&dev_priv->uncore, GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
7296                    GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
7297
7298         /* WaSwitchSolVfFArbitrationPriority:hsw */
7299         intel_uncore_write(&dev_priv->uncore, GAM_ECOCHK, intel_uncore_read(&dev_priv->uncore, GAM_ECOCHK) | HSW_ECOCHK_ARB_PRIO_SOL);
7300
7301         lpt_init_clock_gating(dev_priv);
7302 }
7303
7304 static void ivb_init_clock_gating(struct drm_i915_private *dev_priv)
7305 {
7306         u32 snpcr;
7307
7308         intel_uncore_write(&dev_priv->uncore, ILK_DSPCLK_GATE_D, ILK_VRHUNIT_CLOCK_GATE_DISABLE);
7309
7310         /* WaFbcAsynchFlipDisableFbcQueue:ivb */
7311         intel_uncore_write(&dev_priv->uncore, ILK_DISPLAY_CHICKEN1,
7312                    intel_uncore_read(&dev_priv->uncore, ILK_DISPLAY_CHICKEN1) |
7313                    ILK_FBCQ_DIS);
7314
7315         /* WaDisableBackToBackFlipFix:ivb */
7316         intel_uncore_write(&dev_priv->uncore, IVB_CHICKEN3,
7317                    CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
7318                    CHICKEN3_DGMG_DONE_FIX_DISABLE);
7319
7320         if (IS_IVB_GT1(dev_priv))
7321                 intel_uncore_write(&dev_priv->uncore, GEN7_ROW_CHICKEN2,
7322                            _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
7323         else {
7324                 /* must write both registers */
7325                 intel_uncore_write(&dev_priv->uncore, GEN7_ROW_CHICKEN2,
7326                            _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
7327                 intel_uncore_write(&dev_priv->uncore, GEN7_ROW_CHICKEN2_GT2,
7328                            _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
7329         }
7330
7331         /*
7332          * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
7333          * This implements the WaDisableRCZUnitClockGating:ivb workaround.
7334          */
7335         intel_uncore_write(&dev_priv->uncore, GEN6_UCGCTL2,
7336                    GEN6_RCZUNIT_CLOCK_GATE_DISABLE);
7337
7338         /* This is required by WaCatErrorRejectionIssue:ivb */
7339         intel_uncore_write(&dev_priv->uncore, GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
7340                         intel_uncore_read(&dev_priv->uncore, GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
7341                         GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
7342
7343         g4x_disable_trickle_feed(dev_priv);
7344
7345         snpcr = intel_uncore_read(&dev_priv->uncore, GEN6_MBCUNIT_SNPCR);
7346         snpcr &= ~GEN6_MBC_SNPCR_MASK;
7347         snpcr |= GEN6_MBC_SNPCR_MED;
7348         intel_uncore_write(&dev_priv->uncore, GEN6_MBCUNIT_SNPCR, snpcr);
7349
7350         if (!HAS_PCH_NOP(dev_priv))
7351                 cpt_init_clock_gating(dev_priv);
7352
7353         gen6_check_mch_setup(dev_priv);
7354 }
7355
7356 static void vlv_init_clock_gating(struct drm_i915_private *dev_priv)
7357 {
7358         /* WaDisableBackToBackFlipFix:vlv */
7359         intel_uncore_write(&dev_priv->uncore, IVB_CHICKEN3,
7360                    CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
7361                    CHICKEN3_DGMG_DONE_FIX_DISABLE);
7362
7363         /* WaDisableDopClockGating:vlv */
7364         intel_uncore_write(&dev_priv->uncore, GEN7_ROW_CHICKEN2,
7365                    _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
7366
7367         /* This is required by WaCatErrorRejectionIssue:vlv */
7368         intel_uncore_write(&dev_priv->uncore, GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
7369                    intel_uncore_read(&dev_priv->uncore, GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
7370                    GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
7371
7372         /*
7373          * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
7374          * This implements the WaDisableRCZUnitClockGating:vlv workaround.
7375          */
7376         intel_uncore_write(&dev_priv->uncore, GEN6_UCGCTL2,
7377                    GEN6_RCZUNIT_CLOCK_GATE_DISABLE);
7378
7379         /* WaDisableL3Bank2xClockGate:vlv
7380          * Disabling L3 clock gating- MMIO 940c[25] = 1
7381          * Set bit 25, to disable L3_BANK_2x_CLK_GATING */
7382         intel_uncore_write(&dev_priv->uncore, GEN7_UCGCTL4,
7383                    intel_uncore_read(&dev_priv->uncore, GEN7_UCGCTL4) | GEN7_L3BANK2X_CLOCK_GATE_DISABLE);
7384
7385         /*
7386          * WaDisableVLVClockGating_VBIIssue:vlv
7387          * Disable clock gating on th GCFG unit to prevent a delay
7388          * in the reporting of vblank events.
7389          */
7390         intel_uncore_write(&dev_priv->uncore, VLV_GUNIT_CLOCK_GATE, GCFG_DIS);
7391 }
7392
7393 static void chv_init_clock_gating(struct drm_i915_private *dev_priv)
7394 {
7395         /* WaVSRefCountFullforceMissDisable:chv */
7396         /* WaDSRefCountFullforceMissDisable:chv */
7397         intel_uncore_write(&dev_priv->uncore, GEN7_FF_THREAD_MODE,
7398                    intel_uncore_read(&dev_priv->uncore, GEN7_FF_THREAD_MODE) &
7399                    ~(GEN8_FF_DS_REF_CNT_FFME | GEN7_FF_VS_REF_CNT_FFME));
7400
7401         /* WaDisableSemaphoreAndSyncFlipWait:chv */
7402         intel_uncore_write(&dev_priv->uncore, GEN6_RC_SLEEP_PSMI_CONTROL,
7403                    _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE));
7404
7405         /* WaDisableCSUnitClockGating:chv */
7406         intel_uncore_write(&dev_priv->uncore, GEN6_UCGCTL1, intel_uncore_read(&dev_priv->uncore, GEN6_UCGCTL1) |
7407                    GEN6_CSUNIT_CLOCK_GATE_DISABLE);
7408
7409         /* WaDisableSDEUnitClockGating:chv */
7410         intel_uncore_write(&dev_priv->uncore, GEN8_UCGCTL6, intel_uncore_read(&dev_priv->uncore, GEN8_UCGCTL6) |
7411                    GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
7412
7413         /*
7414          * WaProgramL3SqcReg1Default:chv
7415          * See gfxspecs/Related Documents/Performance Guide/
7416          * LSQC Setting Recommendations.
7417          */
7418         gen8_set_l3sqc_credits(dev_priv, 38, 2);
7419 }
7420
7421 static void g4x_init_clock_gating(struct drm_i915_private *dev_priv)
7422 {
7423         u32 dspclk_gate;
7424
7425         intel_uncore_write(&dev_priv->uncore, RENCLK_GATE_D1, 0);
7426         intel_uncore_write(&dev_priv->uncore, RENCLK_GATE_D2, VF_UNIT_CLOCK_GATE_DISABLE |
7427                    GS_UNIT_CLOCK_GATE_DISABLE |
7428                    CL_UNIT_CLOCK_GATE_DISABLE);
7429         intel_uncore_write(&dev_priv->uncore, RAMCLK_GATE_D, 0);
7430         dspclk_gate = VRHUNIT_CLOCK_GATE_DISABLE |
7431                 OVRUNIT_CLOCK_GATE_DISABLE |
7432                 OVCUNIT_CLOCK_GATE_DISABLE;
7433         if (IS_GM45(dev_priv))
7434                 dspclk_gate |= DSSUNIT_CLOCK_GATE_DISABLE;
7435         intel_uncore_write(&dev_priv->uncore, DSPCLK_GATE_D, dspclk_gate);
7436
7437         g4x_disable_trickle_feed(dev_priv);
7438 }
7439
7440 static void i965gm_init_clock_gating(struct drm_i915_private *dev_priv)
7441 {
7442         struct intel_uncore *uncore = &dev_priv->uncore;
7443
7444         intel_uncore_write(uncore, RENCLK_GATE_D1, I965_RCC_CLOCK_GATE_DISABLE);
7445         intel_uncore_write(uncore, RENCLK_GATE_D2, 0);
7446         intel_uncore_write(uncore, DSPCLK_GATE_D, 0);
7447         intel_uncore_write(uncore, RAMCLK_GATE_D, 0);
7448         intel_uncore_write16(uncore, DEUC, 0);
7449         intel_uncore_write(uncore,
7450                            MI_ARB_STATE,
7451                            _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
7452 }
7453
7454 static void i965g_init_clock_gating(struct drm_i915_private *dev_priv)
7455 {
7456         intel_uncore_write(&dev_priv->uncore, RENCLK_GATE_D1, I965_RCZ_CLOCK_GATE_DISABLE |
7457                    I965_RCC_CLOCK_GATE_DISABLE |
7458                    I965_RCPB_CLOCK_GATE_DISABLE |
7459                    I965_ISC_CLOCK_GATE_DISABLE |
7460                    I965_FBC_CLOCK_GATE_DISABLE);
7461         intel_uncore_write(&dev_priv->uncore, RENCLK_GATE_D2, 0);
7462         intel_uncore_write(&dev_priv->uncore, MI_ARB_STATE,
7463                    _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
7464 }
7465
7466 static void gen3_init_clock_gating(struct drm_i915_private *dev_priv)
7467 {
7468         u32 dstate = intel_uncore_read(&dev_priv->uncore, D_STATE);
7469
7470         dstate |= DSTATE_PLL_D3_OFF | DSTATE_GFX_CLOCK_GATING |
7471                 DSTATE_DOT_CLOCK_GATING;
7472         intel_uncore_write(&dev_priv->uncore, D_STATE, dstate);
7473
7474         if (IS_PINEVIEW(dev_priv))
7475                 intel_uncore_write(&dev_priv->uncore, ECOSKPD, _MASKED_BIT_ENABLE(ECO_GATING_CX_ONLY));
7476
7477         /* IIR "flip pending" means done if this bit is set */
7478         intel_uncore_write(&dev_priv->uncore, ECOSKPD, _MASKED_BIT_DISABLE(ECO_FLIP_DONE));
7479
7480         /* interrupts should cause a wake up from C3 */
7481         intel_uncore_write(&dev_priv->uncore, INSTPM, _MASKED_BIT_ENABLE(INSTPM_AGPBUSY_INT_EN));
7482
7483         /* On GEN3 we really need to make sure the ARB C3 LP bit is set */
7484         intel_uncore_write(&dev_priv->uncore, MI_ARB_STATE, _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE));
7485
7486         intel_uncore_write(&dev_priv->uncore, MI_ARB_STATE,
7487                    _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
7488 }
7489
7490 static void i85x_init_clock_gating(struct drm_i915_private *dev_priv)
7491 {
7492         intel_uncore_write(&dev_priv->uncore, RENCLK_GATE_D1, SV_CLOCK_GATE_DISABLE);
7493
7494         /* interrupts should cause a wake up from C3 */
7495         intel_uncore_write(&dev_priv->uncore, MI_STATE, _MASKED_BIT_ENABLE(MI_AGPBUSY_INT_EN) |
7496                    _MASKED_BIT_DISABLE(MI_AGPBUSY_830_MODE));
7497
7498         intel_uncore_write(&dev_priv->uncore, MEM_MODE,
7499                    _MASKED_BIT_ENABLE(MEM_DISPLAY_TRICKLE_FEED_DISABLE));
7500
7501         /*
7502          * Have FBC ignore 3D activity since we use software
7503          * render tracking, and otherwise a pure 3D workload
7504          * (even if it just renders a single frame and then does
7505          * abosultely nothing) would not allow FBC to recompress
7506          * until a 2D blit occurs.
7507          */
7508         intel_uncore_write(&dev_priv->uncore, SCPD0,
7509                    _MASKED_BIT_ENABLE(SCPD_FBC_IGNORE_3D));
7510 }
7511
7512 static void i830_init_clock_gating(struct drm_i915_private *dev_priv)
7513 {
7514         intel_uncore_write(&dev_priv->uncore, MEM_MODE,
7515                    _MASKED_BIT_ENABLE(MEM_DISPLAY_A_TRICKLE_FEED_DISABLE) |
7516                    _MASKED_BIT_ENABLE(MEM_DISPLAY_B_TRICKLE_FEED_DISABLE));
7517 }
7518
7519 void intel_init_clock_gating(struct drm_i915_private *dev_priv)
7520 {
7521         dev_priv->display.init_clock_gating(dev_priv);
7522 }
7523
7524 void intel_suspend_hw(struct drm_i915_private *dev_priv)
7525 {
7526         if (HAS_PCH_LPT(dev_priv))
7527                 lpt_suspend_hw(dev_priv);
7528 }
7529
7530 static void nop_init_clock_gating(struct drm_i915_private *dev_priv)
7531 {
7532         drm_dbg_kms(&dev_priv->drm,
7533                     "No clock gating settings or workarounds applied.\n");
7534 }
7535
7536 /**
7537  * intel_init_clock_gating_hooks - setup the clock gating hooks
7538  * @dev_priv: device private
7539  *
7540  * Setup the hooks that configure which clocks of a given platform can be
7541  * gated and also apply various GT and display specific workarounds for these
7542  * platforms. Note that some GT specific workarounds are applied separately
7543  * when GPU contexts or batchbuffers start their execution.
7544  */
7545 void intel_init_clock_gating_hooks(struct drm_i915_private *dev_priv)
7546 {
7547         if (IS_DG1(dev_priv))
7548                 dev_priv->display.init_clock_gating = dg1_init_clock_gating;
7549         else if (IS_GEN(dev_priv, 12))
7550                 dev_priv->display.init_clock_gating = gen12lp_init_clock_gating;
7551         else if (IS_GEN(dev_priv, 11))
7552                 dev_priv->display.init_clock_gating = icl_init_clock_gating;
7553         else if (IS_CANNONLAKE(dev_priv))
7554                 dev_priv->display.init_clock_gating = cnl_init_clock_gating;
7555         else if (IS_COFFEELAKE(dev_priv) || IS_COMETLAKE(dev_priv))
7556                 dev_priv->display.init_clock_gating = cfl_init_clock_gating;
7557         else if (IS_SKYLAKE(dev_priv))
7558                 dev_priv->display.init_clock_gating = skl_init_clock_gating;
7559         else if (IS_KABYLAKE(dev_priv))
7560                 dev_priv->display.init_clock_gating = kbl_init_clock_gating;
7561         else if (IS_BROXTON(dev_priv))
7562                 dev_priv->display.init_clock_gating = bxt_init_clock_gating;
7563         else if (IS_GEMINILAKE(dev_priv))
7564                 dev_priv->display.init_clock_gating = glk_init_clock_gating;
7565         else if (IS_BROADWELL(dev_priv))
7566                 dev_priv->display.init_clock_gating = bdw_init_clock_gating;
7567         else if (IS_CHERRYVIEW(dev_priv))
7568                 dev_priv->display.init_clock_gating = chv_init_clock_gating;
7569         else if (IS_HASWELL(dev_priv))
7570                 dev_priv->display.init_clock_gating = hsw_init_clock_gating;
7571         else if (IS_IVYBRIDGE(dev_priv))
7572                 dev_priv->display.init_clock_gating = ivb_init_clock_gating;
7573         else if (IS_VALLEYVIEW(dev_priv))
7574                 dev_priv->display.init_clock_gating = vlv_init_clock_gating;
7575         else if (IS_GEN(dev_priv, 6))
7576                 dev_priv->display.init_clock_gating = gen6_init_clock_gating;
7577         else if (IS_GEN(dev_priv, 5))
7578                 dev_priv->display.init_clock_gating = ilk_init_clock_gating;
7579         else if (IS_G4X(dev_priv))
7580                 dev_priv->display.init_clock_gating = g4x_init_clock_gating;
7581         else if (IS_I965GM(dev_priv))
7582                 dev_priv->display.init_clock_gating = i965gm_init_clock_gating;
7583         else if (IS_I965G(dev_priv))
7584                 dev_priv->display.init_clock_gating = i965g_init_clock_gating;
7585         else if (IS_GEN(dev_priv, 3))
7586                 dev_priv->display.init_clock_gating = gen3_init_clock_gating;
7587         else if (IS_I85X(dev_priv) || IS_I865G(dev_priv))
7588                 dev_priv->display.init_clock_gating = i85x_init_clock_gating;
7589         else if (IS_GEN(dev_priv, 2))
7590                 dev_priv->display.init_clock_gating = i830_init_clock_gating;
7591         else {
7592                 MISSING_CASE(INTEL_DEVID(dev_priv));
7593                 dev_priv->display.init_clock_gating = nop_init_clock_gating;
7594         }
7595 }
7596
7597 /* Set up chip specific power management-related functions */
7598 void intel_init_pm(struct drm_i915_private *dev_priv)
7599 {
7600         /* For cxsr */
7601         if (IS_PINEVIEW(dev_priv))
7602                 pnv_get_mem_freq(dev_priv);
7603         else if (IS_GEN(dev_priv, 5))
7604                 ilk_get_mem_freq(dev_priv);
7605
7606         if (intel_has_sagv(dev_priv))
7607                 skl_setup_sagv_block_time(dev_priv);
7608
7609         /* For FIFO watermark updates */
7610         if (INTEL_GEN(dev_priv) >= 9) {
7611                 skl_setup_wm_latency(dev_priv);
7612                 dev_priv->display.compute_global_watermarks = skl_compute_wm;
7613         } else if (HAS_PCH_SPLIT(dev_priv)) {
7614                 ilk_setup_wm_latency(dev_priv);
7615
7616                 if ((IS_GEN(dev_priv, 5) && dev_priv->wm.pri_latency[1] &&
7617                      dev_priv->wm.spr_latency[1] && dev_priv->wm.cur_latency[1]) ||
7618                     (!IS_GEN(dev_priv, 5) && dev_priv->wm.pri_latency[0] &&
7619                      dev_priv->wm.spr_latency[0] && dev_priv->wm.cur_latency[0])) {
7620                         dev_priv->display.compute_pipe_wm = ilk_compute_pipe_wm;
7621                         dev_priv->display.compute_intermediate_wm =
7622                                 ilk_compute_intermediate_wm;
7623                         dev_priv->display.initial_watermarks =
7624                                 ilk_initial_watermarks;
7625                         dev_priv->display.optimize_watermarks =
7626                                 ilk_optimize_watermarks;
7627                 } else {
7628                         drm_dbg_kms(&dev_priv->drm,
7629                                     "Failed to read display plane latency. "
7630                                     "Disable CxSR\n");
7631                 }
7632         } else if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv)) {
7633                 vlv_setup_wm_latency(dev_priv);
7634                 dev_priv->display.compute_pipe_wm = vlv_compute_pipe_wm;
7635                 dev_priv->display.compute_intermediate_wm = vlv_compute_intermediate_wm;
7636                 dev_priv->display.initial_watermarks = vlv_initial_watermarks;
7637                 dev_priv->display.optimize_watermarks = vlv_optimize_watermarks;
7638                 dev_priv->display.atomic_update_watermarks = vlv_atomic_update_fifo;
7639         } else if (IS_G4X(dev_priv)) {
7640                 g4x_setup_wm_latency(dev_priv);
7641                 dev_priv->display.compute_pipe_wm = g4x_compute_pipe_wm;
7642                 dev_priv->display.compute_intermediate_wm = g4x_compute_intermediate_wm;
7643                 dev_priv->display.initial_watermarks = g4x_initial_watermarks;
7644                 dev_priv->display.optimize_watermarks = g4x_optimize_watermarks;
7645         } else if (IS_PINEVIEW(dev_priv)) {
7646                 if (!intel_get_cxsr_latency(!IS_MOBILE(dev_priv),
7647                                             dev_priv->is_ddr3,
7648                                             dev_priv->fsb_freq,
7649                                             dev_priv->mem_freq)) {
7650                         drm_info(&dev_priv->drm,
7651                                  "failed to find known CxSR latency "
7652                                  "(found ddr%s fsb freq %d, mem freq %d), "
7653                                  "disabling CxSR\n",
7654                                  (dev_priv->is_ddr3 == 1) ? "3" : "2",
7655                                  dev_priv->fsb_freq, dev_priv->mem_freq);
7656                         /* Disable CxSR and never update its watermark again */
7657                         intel_set_memory_cxsr(dev_priv, false);
7658                         dev_priv->display.update_wm = NULL;
7659                 } else
7660                         dev_priv->display.update_wm = pnv_update_wm;
7661         } else if (IS_GEN(dev_priv, 4)) {
7662                 dev_priv->display.update_wm = i965_update_wm;
7663         } else if (IS_GEN(dev_priv, 3)) {
7664                 dev_priv->display.update_wm = i9xx_update_wm;
7665                 dev_priv->display.get_fifo_size = i9xx_get_fifo_size;
7666         } else if (IS_GEN(dev_priv, 2)) {
7667                 if (INTEL_NUM_PIPES(dev_priv) == 1) {
7668                         dev_priv->display.update_wm = i845_update_wm;
7669                         dev_priv->display.get_fifo_size = i845_get_fifo_size;
7670                 } else {
7671                         dev_priv->display.update_wm = i9xx_update_wm;
7672                         dev_priv->display.get_fifo_size = i830_get_fifo_size;
7673                 }
7674         } else {
7675                 drm_err(&dev_priv->drm,
7676                         "unexpected fall-through in %s\n", __func__);
7677         }
7678 }
7679
7680 void intel_pm_setup(struct drm_i915_private *dev_priv)
7681 {
7682         dev_priv->runtime_pm.suspended = false;
7683         atomic_set(&dev_priv->runtime_pm.wakeref_count, 0);
7684 }
7685
7686 static struct intel_global_state *intel_dbuf_duplicate_state(struct intel_global_obj *obj)
7687 {
7688         struct intel_dbuf_state *dbuf_state;
7689
7690         dbuf_state = kmemdup(obj->state, sizeof(*dbuf_state), GFP_KERNEL);
7691         if (!dbuf_state)
7692                 return NULL;
7693
7694         return &dbuf_state->base;
7695 }
7696
7697 static void intel_dbuf_destroy_state(struct intel_global_obj *obj,
7698                                      struct intel_global_state *state)
7699 {
7700         kfree(state);
7701 }
7702
7703 static const struct intel_global_state_funcs intel_dbuf_funcs = {
7704         .atomic_duplicate_state = intel_dbuf_duplicate_state,
7705         .atomic_destroy_state = intel_dbuf_destroy_state,
7706 };
7707
7708 struct intel_dbuf_state *
7709 intel_atomic_get_dbuf_state(struct intel_atomic_state *state)
7710 {
7711         struct drm_i915_private *dev_priv = to_i915(state->base.dev);
7712         struct intel_global_state *dbuf_state;
7713
7714         dbuf_state = intel_atomic_get_global_obj_state(state, &dev_priv->dbuf.obj);
7715         if (IS_ERR(dbuf_state))
7716                 return ERR_CAST(dbuf_state);
7717
7718         return to_intel_dbuf_state(dbuf_state);
7719 }
7720
7721 int intel_dbuf_init(struct drm_i915_private *dev_priv)
7722 {
7723         struct intel_dbuf_state *dbuf_state;
7724
7725         dbuf_state = kzalloc(sizeof(*dbuf_state), GFP_KERNEL);
7726         if (!dbuf_state)
7727                 return -ENOMEM;
7728
7729         intel_atomic_global_obj_init(dev_priv, &dev_priv->dbuf.obj,
7730                                      &dbuf_state->base, &intel_dbuf_funcs);
7731
7732         return 0;
7733 }
7734
7735 void intel_dbuf_pre_plane_update(struct intel_atomic_state *state)
7736 {
7737         struct drm_i915_private *dev_priv = to_i915(state->base.dev);
7738         const struct intel_dbuf_state *new_dbuf_state =
7739                 intel_atomic_get_new_dbuf_state(state);
7740         const struct intel_dbuf_state *old_dbuf_state =
7741                 intel_atomic_get_old_dbuf_state(state);
7742
7743         if (!new_dbuf_state ||
7744             new_dbuf_state->enabled_slices == old_dbuf_state->enabled_slices)
7745                 return;
7746
7747         WARN_ON(!new_dbuf_state->base.changed);
7748
7749         gen9_dbuf_slices_update(dev_priv,
7750                                 old_dbuf_state->enabled_slices |
7751                                 new_dbuf_state->enabled_slices);
7752 }
7753
7754 void intel_dbuf_post_plane_update(struct intel_atomic_state *state)
7755 {
7756         struct drm_i915_private *dev_priv = to_i915(state->base.dev);
7757         const struct intel_dbuf_state *new_dbuf_state =
7758                 intel_atomic_get_new_dbuf_state(state);
7759         const struct intel_dbuf_state *old_dbuf_state =
7760                 intel_atomic_get_old_dbuf_state(state);
7761
7762         if (!new_dbuf_state ||
7763             new_dbuf_state->enabled_slices == old_dbuf_state->enabled_slices)
7764                 return;
7765
7766         WARN_ON(!new_dbuf_state->base.changed);
7767
7768         gen9_dbuf_slices_update(dev_priv,
7769                                 new_dbuf_state->enabled_slices);
7770 }