2 * Copyright © 2012 Intel Corporation
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:
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
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
24 * Eugeni Dodonov <eugeni.dodonov@intel.com>
28 #include <linux/module.h>
29 #include <linux/pm_runtime.h>
31 #include <drm/drm_atomic_helper.h>
32 #include <drm/drm_fourcc.h>
33 #include <drm/drm_plane_helper.h>
35 #include "display/intel_atomic.h"
36 #include "display/intel_display_types.h"
37 #include "display/intel_fbc.h"
38 #include "display/intel_sprite.h"
40 #include "gt/intel_llc.h"
44 #include "i915_trace.h"
46 #include "intel_sideband.h"
47 #include "../../../platform/x86/intel_ips.h"
49 static void gen9_init_clock_gating(struct drm_i915_private *dev_priv)
51 if (HAS_LLC(dev_priv)) {
53 * WaCompressedResourceDisplayNewHashMode:skl,kbl
54 * Display WA #0390: skl,kbl
56 * Must match Sampler, Pixel Back End, and Media. See
57 * WaCompressedResourceSamplerPbeMediaNewHashMode.
59 I915_WRITE(CHICKEN_PAR1_1,
60 I915_READ(CHICKEN_PAR1_1) |
61 SKL_DE_COMPRESSED_HASH_MODE);
64 /* See Bspec note for PSR2_CTL bit 31, Wa#828:skl,bxt,kbl,cfl */
65 I915_WRITE(CHICKEN_PAR1_1,
66 I915_READ(CHICKEN_PAR1_1) | SKL_EDP_PSR_FIX_RDWRAP);
68 /* WaEnableChickenDCPR:skl,bxt,kbl,glk,cfl */
69 I915_WRITE(GEN8_CHICKEN_DCPR_1,
70 I915_READ(GEN8_CHICKEN_DCPR_1) | MASK_WAKEMEM);
72 /* WaFbcTurnOffFbcWatermark:skl,bxt,kbl,cfl */
73 /* WaFbcWakeMemOn:skl,bxt,kbl,glk,cfl */
74 I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) |
76 DISP_FBC_MEMORY_WAKE);
78 /* WaFbcHighMemBwCorruptionAvoidance:skl,bxt,kbl,cfl */
79 I915_WRITE(ILK_DPFC_CHICKEN, I915_READ(ILK_DPFC_CHICKEN) |
80 ILK_DPFC_DISABLE_DUMMY0);
82 if (IS_SKYLAKE(dev_priv)) {
83 /* WaDisableDopClockGating */
84 I915_WRITE(GEN7_MISCCPCTL, I915_READ(GEN7_MISCCPCTL)
85 & ~GEN7_DOP_CLOCK_GATE_ENABLE);
89 static void bxt_init_clock_gating(struct drm_i915_private *dev_priv)
91 gen9_init_clock_gating(dev_priv);
93 /* WaDisableSDEUnitClockGating:bxt */
94 I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
95 GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
99 * GEN8_HDCUNIT_CLOCK_GATE_DISABLE_HDCREQ applies on 3x6 GT SKUs only.
101 I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
102 GEN8_HDCUNIT_CLOCK_GATE_DISABLE_HDCREQ);
105 * Wa: Backlight PWM may stop in the asserted state, causing backlight
108 I915_WRITE(GEN9_CLKGATE_DIS_0, I915_READ(GEN9_CLKGATE_DIS_0) |
109 PWM1_GATING_DIS | PWM2_GATING_DIS);
112 * Lower the display internal timeout.
113 * This is needed to avoid any hard hangs when DSI port PLL
114 * is off and a MMIO access is attempted by any privilege
115 * application, using batch buffers or any other means.
117 I915_WRITE(RM_TIMEOUT, MMIO_TIMEOUT_US(950));
120 static void glk_init_clock_gating(struct drm_i915_private *dev_priv)
122 gen9_init_clock_gating(dev_priv);
125 * WaDisablePWMClockGating:glk
126 * Backlight PWM may stop in the asserted state, causing backlight
129 I915_WRITE(GEN9_CLKGATE_DIS_0, I915_READ(GEN9_CLKGATE_DIS_0) |
130 PWM1_GATING_DIS | PWM2_GATING_DIS);
133 static void pnv_get_mem_freq(struct drm_i915_private *dev_priv)
137 tmp = I915_READ(CLKCFG);
139 switch (tmp & CLKCFG_FSB_MASK) {
141 dev_priv->fsb_freq = 533; /* 133*4 */
144 dev_priv->fsb_freq = 800; /* 200*4 */
147 dev_priv->fsb_freq = 667; /* 167*4 */
150 dev_priv->fsb_freq = 400; /* 100*4 */
154 switch (tmp & CLKCFG_MEM_MASK) {
156 dev_priv->mem_freq = 533;
159 dev_priv->mem_freq = 667;
162 dev_priv->mem_freq = 800;
166 /* detect pineview DDR3 setting */
167 tmp = I915_READ(CSHRDDR3CTL);
168 dev_priv->is_ddr3 = (tmp & CSHRDDR3CTL_DDR3) ? 1 : 0;
171 static void ilk_get_mem_freq(struct drm_i915_private *dev_priv)
175 ddrpll = intel_uncore_read16(&dev_priv->uncore, DDRMPLL1);
176 csipll = intel_uncore_read16(&dev_priv->uncore, CSIPLL0);
178 switch (ddrpll & 0xff) {
180 dev_priv->mem_freq = 800;
183 dev_priv->mem_freq = 1066;
186 dev_priv->mem_freq = 1333;
189 dev_priv->mem_freq = 1600;
192 drm_dbg(&dev_priv->drm, "unknown memory frequency 0x%02x\n",
194 dev_priv->mem_freq = 0;
198 switch (csipll & 0x3ff) {
200 dev_priv->fsb_freq = 3200;
203 dev_priv->fsb_freq = 3733;
206 dev_priv->fsb_freq = 4266;
209 dev_priv->fsb_freq = 4800;
212 dev_priv->fsb_freq = 5333;
215 dev_priv->fsb_freq = 5866;
218 dev_priv->fsb_freq = 6400;
221 drm_dbg(&dev_priv->drm, "unknown fsb frequency 0x%04x\n",
223 dev_priv->fsb_freq = 0;
228 static const struct cxsr_latency cxsr_latency_table[] = {
229 {1, 0, 800, 400, 3382, 33382, 3983, 33983}, /* DDR2-400 SC */
230 {1, 0, 800, 667, 3354, 33354, 3807, 33807}, /* DDR2-667 SC */
231 {1, 0, 800, 800, 3347, 33347, 3763, 33763}, /* DDR2-800 SC */
232 {1, 1, 800, 667, 6420, 36420, 6873, 36873}, /* DDR3-667 SC */
233 {1, 1, 800, 800, 5902, 35902, 6318, 36318}, /* DDR3-800 SC */
235 {1, 0, 667, 400, 3400, 33400, 4021, 34021}, /* DDR2-400 SC */
236 {1, 0, 667, 667, 3372, 33372, 3845, 33845}, /* DDR2-667 SC */
237 {1, 0, 667, 800, 3386, 33386, 3822, 33822}, /* DDR2-800 SC */
238 {1, 1, 667, 667, 6438, 36438, 6911, 36911}, /* DDR3-667 SC */
239 {1, 1, 667, 800, 5941, 35941, 6377, 36377}, /* DDR3-800 SC */
241 {1, 0, 400, 400, 3472, 33472, 4173, 34173}, /* DDR2-400 SC */
242 {1, 0, 400, 667, 3443, 33443, 3996, 33996}, /* DDR2-667 SC */
243 {1, 0, 400, 800, 3430, 33430, 3946, 33946}, /* DDR2-800 SC */
244 {1, 1, 400, 667, 6509, 36509, 7062, 37062}, /* DDR3-667 SC */
245 {1, 1, 400, 800, 5985, 35985, 6501, 36501}, /* DDR3-800 SC */
247 {0, 0, 800, 400, 3438, 33438, 4065, 34065}, /* DDR2-400 SC */
248 {0, 0, 800, 667, 3410, 33410, 3889, 33889}, /* DDR2-667 SC */
249 {0, 0, 800, 800, 3403, 33403, 3845, 33845}, /* DDR2-800 SC */
250 {0, 1, 800, 667, 6476, 36476, 6955, 36955}, /* DDR3-667 SC */
251 {0, 1, 800, 800, 5958, 35958, 6400, 36400}, /* DDR3-800 SC */
253 {0, 0, 667, 400, 3456, 33456, 4103, 34106}, /* DDR2-400 SC */
254 {0, 0, 667, 667, 3428, 33428, 3927, 33927}, /* DDR2-667 SC */
255 {0, 0, 667, 800, 3443, 33443, 3905, 33905}, /* DDR2-800 SC */
256 {0, 1, 667, 667, 6494, 36494, 6993, 36993}, /* DDR3-667 SC */
257 {0, 1, 667, 800, 5998, 35998, 6460, 36460}, /* DDR3-800 SC */
259 {0, 0, 400, 400, 3528, 33528, 4255, 34255}, /* DDR2-400 SC */
260 {0, 0, 400, 667, 3500, 33500, 4079, 34079}, /* DDR2-667 SC */
261 {0, 0, 400, 800, 3487, 33487, 4029, 34029}, /* DDR2-800 SC */
262 {0, 1, 400, 667, 6566, 36566, 7145, 37145}, /* DDR3-667 SC */
263 {0, 1, 400, 800, 6042, 36042, 6584, 36584}, /* DDR3-800 SC */
266 static const struct cxsr_latency *intel_get_cxsr_latency(bool is_desktop,
271 const struct cxsr_latency *latency;
274 if (fsb == 0 || mem == 0)
277 for (i = 0; i < ARRAY_SIZE(cxsr_latency_table); i++) {
278 latency = &cxsr_latency_table[i];
279 if (is_desktop == latency->is_desktop &&
280 is_ddr3 == latency->is_ddr3 &&
281 fsb == latency->fsb_freq && mem == latency->mem_freq)
285 DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
290 static void chv_set_memory_dvfs(struct drm_i915_private *dev_priv, bool enable)
294 vlv_punit_get(dev_priv);
296 val = vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2);
298 val &= ~FORCE_DDR_HIGH_FREQ;
300 val |= FORCE_DDR_HIGH_FREQ;
301 val &= ~FORCE_DDR_LOW_FREQ;
302 val |= FORCE_DDR_FREQ_REQ_ACK;
303 vlv_punit_write(dev_priv, PUNIT_REG_DDR_SETUP2, val);
305 if (wait_for((vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2) &
306 FORCE_DDR_FREQ_REQ_ACK) == 0, 3))
307 drm_err(&dev_priv->drm,
308 "timed out waiting for Punit DDR DVFS request\n");
310 vlv_punit_put(dev_priv);
313 static void chv_set_memory_pm5(struct drm_i915_private *dev_priv, bool enable)
317 vlv_punit_get(dev_priv);
319 val = vlv_punit_read(dev_priv, PUNIT_REG_DSPSSPM);
321 val |= DSP_MAXFIFO_PM5_ENABLE;
323 val &= ~DSP_MAXFIFO_PM5_ENABLE;
324 vlv_punit_write(dev_priv, PUNIT_REG_DSPSSPM, val);
326 vlv_punit_put(dev_priv);
329 #define FW_WM(value, plane) \
330 (((value) << DSPFW_ ## plane ## _SHIFT) & DSPFW_ ## plane ## _MASK)
332 static bool _intel_set_memory_cxsr(struct drm_i915_private *dev_priv, bool enable)
337 if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv)) {
338 was_enabled = I915_READ(FW_BLC_SELF_VLV) & FW_CSPWRDWNEN;
339 I915_WRITE(FW_BLC_SELF_VLV, enable ? FW_CSPWRDWNEN : 0);
340 POSTING_READ(FW_BLC_SELF_VLV);
341 } else if (IS_G4X(dev_priv) || IS_I965GM(dev_priv)) {
342 was_enabled = I915_READ(FW_BLC_SELF) & FW_BLC_SELF_EN;
343 I915_WRITE(FW_BLC_SELF, enable ? FW_BLC_SELF_EN : 0);
344 POSTING_READ(FW_BLC_SELF);
345 } else if (IS_PINEVIEW(dev_priv)) {
346 val = I915_READ(DSPFW3);
347 was_enabled = val & PINEVIEW_SELF_REFRESH_EN;
349 val |= PINEVIEW_SELF_REFRESH_EN;
351 val &= ~PINEVIEW_SELF_REFRESH_EN;
352 I915_WRITE(DSPFW3, val);
353 POSTING_READ(DSPFW3);
354 } else if (IS_I945G(dev_priv) || IS_I945GM(dev_priv)) {
355 was_enabled = I915_READ(FW_BLC_SELF) & FW_BLC_SELF_EN;
356 val = enable ? _MASKED_BIT_ENABLE(FW_BLC_SELF_EN) :
357 _MASKED_BIT_DISABLE(FW_BLC_SELF_EN);
358 I915_WRITE(FW_BLC_SELF, val);
359 POSTING_READ(FW_BLC_SELF);
360 } else if (IS_I915GM(dev_priv)) {
362 * FIXME can't find a bit like this for 915G, and
363 * and yet it does have the related watermark in
364 * FW_BLC_SELF. What's going on?
366 was_enabled = I915_READ(INSTPM) & INSTPM_SELF_EN;
367 val = enable ? _MASKED_BIT_ENABLE(INSTPM_SELF_EN) :
368 _MASKED_BIT_DISABLE(INSTPM_SELF_EN);
369 I915_WRITE(INSTPM, val);
370 POSTING_READ(INSTPM);
375 trace_intel_memory_cxsr(dev_priv, was_enabled, enable);
377 drm_dbg_kms(&dev_priv->drm, "memory self-refresh is %s (was %s)\n",
378 enableddisabled(enable),
379 enableddisabled(was_enabled));
385 * intel_set_memory_cxsr - Configure CxSR state
386 * @dev_priv: i915 device
387 * @enable: Allow vs. disallow CxSR
389 * Allow or disallow the system to enter a special CxSR
390 * (C-state self refresh) state. What typically happens in CxSR mode
391 * is that several display FIFOs may get combined into a single larger
392 * FIFO for a particular plane (so called max FIFO mode) to allow the
393 * system to defer memory fetches longer, and the memory will enter
396 * Note that enabling CxSR does not guarantee that the system enter
397 * this special mode, nor does it guarantee that the system stays
398 * in that mode once entered. So this just allows/disallows the system
399 * to autonomously utilize the CxSR mode. Other factors such as core
400 * C-states will affect when/if the system actually enters/exits the
403 * Note that on VLV/CHV this actually only controls the max FIFO mode,
404 * and the system is free to enter/exit memory self refresh at any time
405 * even when the use of CxSR has been disallowed.
407 * While the system is actually in the CxSR/max FIFO mode, some plane
408 * control registers will not get latched on vblank. Thus in order to
409 * guarantee the system will respond to changes in the plane registers
410 * we must always disallow CxSR prior to making changes to those registers.
411 * Unfortunately the system will re-evaluate the CxSR conditions at
412 * frame start which happens after vblank start (which is when the plane
413 * registers would get latched), so we can't proceed with the plane update
414 * during the same frame where we disallowed CxSR.
416 * Certain platforms also have a deeper HPLL SR mode. Fortunately the
417 * HPLL SR mode depends on CxSR itself, so we don't have to hand hold
418 * the hardware w.r.t. HPLL SR when writing to plane registers.
419 * Disallowing just CxSR is sufficient.
421 bool intel_set_memory_cxsr(struct drm_i915_private *dev_priv, bool enable)
425 mutex_lock(&dev_priv->wm.wm_mutex);
426 ret = _intel_set_memory_cxsr(dev_priv, enable);
427 if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv))
428 dev_priv->wm.vlv.cxsr = enable;
429 else if (IS_G4X(dev_priv))
430 dev_priv->wm.g4x.cxsr = enable;
431 mutex_unlock(&dev_priv->wm.wm_mutex);
437 * Latency for FIFO fetches is dependent on several factors:
438 * - memory configuration (speed, channels)
440 * - current MCH state
441 * It can be fairly high in some situations, so here we assume a fairly
442 * pessimal value. It's a tradeoff between extra memory fetches (if we
443 * set this value too high, the FIFO will fetch frequently to stay full)
444 * and power consumption (set it too low to save power and we might see
445 * FIFO underruns and display "flicker").
447 * A value of 5us seems to be a good balance; safe for very low end
448 * platforms but not overly aggressive on lower latency configs.
450 static const int pessimal_latency_ns = 5000;
452 #define VLV_FIFO_START(dsparb, dsparb2, lo_shift, hi_shift) \
453 ((((dsparb) >> (lo_shift)) & 0xff) | ((((dsparb2) >> (hi_shift)) & 0x1) << 8))
455 static void vlv_get_fifo_size(struct intel_crtc_state *crtc_state)
457 struct intel_crtc *crtc = to_intel_crtc(crtc_state->uapi.crtc);
458 struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
459 struct vlv_fifo_state *fifo_state = &crtc_state->wm.vlv.fifo_state;
460 enum pipe pipe = crtc->pipe;
461 int sprite0_start, sprite1_start;
462 u32 dsparb, dsparb2, dsparb3;
466 dsparb = I915_READ(DSPARB);
467 dsparb2 = I915_READ(DSPARB2);
468 sprite0_start = VLV_FIFO_START(dsparb, dsparb2, 0, 0);
469 sprite1_start = VLV_FIFO_START(dsparb, dsparb2, 8, 4);
472 dsparb = I915_READ(DSPARB);
473 dsparb2 = I915_READ(DSPARB2);
474 sprite0_start = VLV_FIFO_START(dsparb, dsparb2, 16, 8);
475 sprite1_start = VLV_FIFO_START(dsparb, dsparb2, 24, 12);
478 dsparb2 = I915_READ(DSPARB2);
479 dsparb3 = I915_READ(DSPARB3);
480 sprite0_start = VLV_FIFO_START(dsparb3, dsparb2, 0, 16);
481 sprite1_start = VLV_FIFO_START(dsparb3, dsparb2, 8, 20);
488 fifo_state->plane[PLANE_PRIMARY] = sprite0_start;
489 fifo_state->plane[PLANE_SPRITE0] = sprite1_start - sprite0_start;
490 fifo_state->plane[PLANE_SPRITE1] = 511 - sprite1_start;
491 fifo_state->plane[PLANE_CURSOR] = 63;
494 static int i9xx_get_fifo_size(struct drm_i915_private *dev_priv,
495 enum i9xx_plane_id i9xx_plane)
497 u32 dsparb = I915_READ(DSPARB);
500 size = dsparb & 0x7f;
501 if (i9xx_plane == PLANE_B)
502 size = ((dsparb >> DSPARB_CSTART_SHIFT) & 0x7f) - size;
504 drm_dbg_kms(&dev_priv->drm, "FIFO size - (0x%08x) %c: %d\n",
505 dsparb, plane_name(i9xx_plane), size);
510 static int i830_get_fifo_size(struct drm_i915_private *dev_priv,
511 enum i9xx_plane_id i9xx_plane)
513 u32 dsparb = I915_READ(DSPARB);
516 size = dsparb & 0x1ff;
517 if (i9xx_plane == PLANE_B)
518 size = ((dsparb >> DSPARB_BEND_SHIFT) & 0x1ff) - size;
519 size >>= 1; /* Convert to cachelines */
521 drm_dbg_kms(&dev_priv->drm, "FIFO size - (0x%08x) %c: %d\n",
522 dsparb, plane_name(i9xx_plane), size);
527 static int i845_get_fifo_size(struct drm_i915_private *dev_priv,
528 enum i9xx_plane_id i9xx_plane)
530 u32 dsparb = I915_READ(DSPARB);
533 size = dsparb & 0x7f;
534 size >>= 2; /* Convert to cachelines */
536 drm_dbg_kms(&dev_priv->drm, "FIFO size - (0x%08x) %c: %d\n",
537 dsparb, plane_name(i9xx_plane), size);
542 /* Pineview has different values for various configs */
543 static const struct intel_watermark_params pnv_display_wm = {
544 .fifo_size = PINEVIEW_DISPLAY_FIFO,
545 .max_wm = PINEVIEW_MAX_WM,
546 .default_wm = PINEVIEW_DFT_WM,
547 .guard_size = PINEVIEW_GUARD_WM,
548 .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
551 static const struct intel_watermark_params pnv_display_hplloff_wm = {
552 .fifo_size = PINEVIEW_DISPLAY_FIFO,
553 .max_wm = PINEVIEW_MAX_WM,
554 .default_wm = PINEVIEW_DFT_HPLLOFF_WM,
555 .guard_size = PINEVIEW_GUARD_WM,
556 .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
559 static const struct intel_watermark_params pnv_cursor_wm = {
560 .fifo_size = PINEVIEW_CURSOR_FIFO,
561 .max_wm = PINEVIEW_CURSOR_MAX_WM,
562 .default_wm = PINEVIEW_CURSOR_DFT_WM,
563 .guard_size = PINEVIEW_CURSOR_GUARD_WM,
564 .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
567 static const struct intel_watermark_params pnv_cursor_hplloff_wm = {
568 .fifo_size = PINEVIEW_CURSOR_FIFO,
569 .max_wm = PINEVIEW_CURSOR_MAX_WM,
570 .default_wm = PINEVIEW_CURSOR_DFT_WM,
571 .guard_size = PINEVIEW_CURSOR_GUARD_WM,
572 .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
575 static const struct intel_watermark_params i965_cursor_wm_info = {
576 .fifo_size = I965_CURSOR_FIFO,
577 .max_wm = I965_CURSOR_MAX_WM,
578 .default_wm = I965_CURSOR_DFT_WM,
580 .cacheline_size = I915_FIFO_LINE_SIZE,
583 static const struct intel_watermark_params i945_wm_info = {
584 .fifo_size = I945_FIFO_SIZE,
585 .max_wm = I915_MAX_WM,
588 .cacheline_size = I915_FIFO_LINE_SIZE,
591 static const struct intel_watermark_params i915_wm_info = {
592 .fifo_size = I915_FIFO_SIZE,
593 .max_wm = I915_MAX_WM,
596 .cacheline_size = I915_FIFO_LINE_SIZE,
599 static const struct intel_watermark_params i830_a_wm_info = {
600 .fifo_size = I855GM_FIFO_SIZE,
601 .max_wm = I915_MAX_WM,
604 .cacheline_size = I830_FIFO_LINE_SIZE,
607 static const struct intel_watermark_params i830_bc_wm_info = {
608 .fifo_size = I855GM_FIFO_SIZE,
609 .max_wm = I915_MAX_WM/2,
612 .cacheline_size = I830_FIFO_LINE_SIZE,
615 static const struct intel_watermark_params i845_wm_info = {
616 .fifo_size = I830_FIFO_SIZE,
617 .max_wm = I915_MAX_WM,
620 .cacheline_size = I830_FIFO_LINE_SIZE,
624 * intel_wm_method1 - Method 1 / "small buffer" watermark formula
625 * @pixel_rate: Pipe pixel rate in kHz
626 * @cpp: Plane bytes per pixel
627 * @latency: Memory wakeup latency in 0.1us units
629 * Compute the watermark using the method 1 or "small buffer"
630 * formula. The caller may additonally add extra cachelines
631 * to account for TLB misses and clock crossings.
633 * This method is concerned with the short term drain rate
634 * of the FIFO, ie. it does not account for blanking periods
635 * which would effectively reduce the average drain rate across
636 * a longer period. The name "small" refers to the fact the
637 * FIFO is relatively small compared to the amount of data
640 * The FIFO level vs. time graph might look something like:
644 * __---__---__ (- plane active, _ blanking)
647 * or perhaps like this:
650 * __----__----__ (- plane active, _ blanking)
654 * The watermark in bytes
656 static unsigned int intel_wm_method1(unsigned int pixel_rate,
658 unsigned int latency)
662 ret = mul_u32_u32(pixel_rate, cpp * latency);
663 ret = DIV_ROUND_UP_ULL(ret, 10000);
669 * intel_wm_method2 - Method 2 / "large buffer" watermark formula
670 * @pixel_rate: Pipe pixel rate in kHz
671 * @htotal: Pipe horizontal total
672 * @width: Plane width in pixels
673 * @cpp: Plane bytes per pixel
674 * @latency: Memory wakeup latency in 0.1us units
676 * Compute the watermark using the method 2 or "large buffer"
677 * formula. The caller may additonally add extra cachelines
678 * to account for TLB misses and clock crossings.
680 * This method is concerned with the long term drain rate
681 * of the FIFO, ie. it does account for blanking periods
682 * which effectively reduce the average drain rate across
683 * a longer period. The name "large" refers to the fact the
684 * FIFO is relatively large compared to the amount of data
687 * The FIFO level vs. time graph might look something like:
692 * __ --__--__--__--__--__--__ (- plane active, _ blanking)
696 * The watermark in bytes
698 static unsigned int intel_wm_method2(unsigned int pixel_rate,
702 unsigned int latency)
707 * FIXME remove once all users are computing
708 * watermarks in the correct place.
710 if (WARN_ON_ONCE(htotal == 0))
713 ret = (latency * pixel_rate) / (htotal * 10000);
714 ret = (ret + 1) * width * cpp;
720 * intel_calculate_wm - calculate watermark level
721 * @pixel_rate: pixel clock
722 * @wm: chip FIFO params
723 * @fifo_size: size of the FIFO buffer
724 * @cpp: bytes per pixel
725 * @latency_ns: memory latency for the platform
727 * Calculate the watermark level (the level at which the display plane will
728 * start fetching from memory again). Each chip has a different display
729 * FIFO size and allocation, so the caller needs to figure that out and pass
730 * in the correct intel_watermark_params structure.
732 * As the pixel clock runs, the FIFO will be drained at a rate that depends
733 * on the pixel size. When it reaches the watermark level, it'll start
734 * fetching FIFO line sized based chunks from memory until the FIFO fills
735 * past the watermark point. If the FIFO drains completely, a FIFO underrun
736 * will occur, and a display engine hang could result.
738 static unsigned int intel_calculate_wm(int pixel_rate,
739 const struct intel_watermark_params *wm,
740 int fifo_size, int cpp,
741 unsigned int latency_ns)
743 int entries, wm_size;
746 * Note: we need to make sure we don't overflow for various clock &
748 * clocks go from a few thousand to several hundred thousand.
749 * latency is usually a few thousand
751 entries = intel_wm_method1(pixel_rate, cpp,
753 entries = DIV_ROUND_UP(entries, wm->cacheline_size) +
755 DRM_DEBUG_KMS("FIFO entries required for mode: %d\n", entries);
757 wm_size = fifo_size - entries;
758 DRM_DEBUG_KMS("FIFO watermark level: %d\n", wm_size);
760 /* Don't promote wm_size to unsigned... */
761 if (wm_size > wm->max_wm)
762 wm_size = wm->max_wm;
764 wm_size = wm->default_wm;
767 * Bspec seems to indicate that the value shouldn't be lower than
768 * 'burst size + 1'. Certainly 830 is quite unhappy with low values.
769 * Lets go for 8 which is the burst size since certain platforms
770 * already use a hardcoded 8 (which is what the spec says should be
779 static bool is_disabling(int old, int new, int threshold)
781 return old >= threshold && new < threshold;
784 static bool is_enabling(int old, int new, int threshold)
786 return old < threshold && new >= threshold;
789 static int intel_wm_num_levels(struct drm_i915_private *dev_priv)
791 return dev_priv->wm.max_level + 1;
794 static bool intel_wm_plane_visible(const struct intel_crtc_state *crtc_state,
795 const struct intel_plane_state *plane_state)
797 struct intel_plane *plane = to_intel_plane(plane_state->uapi.plane);
799 /* FIXME check the 'enable' instead */
800 if (!crtc_state->hw.active)
804 * Treat cursor with fb as always visible since cursor updates
805 * can happen faster than the vrefresh rate, and the current
806 * watermark code doesn't handle that correctly. Cursor updates
807 * which set/clear the fb or change the cursor size are going
808 * to get throttled by intel_legacy_cursor_update() to work
809 * around this problem with the watermark code.
811 if (plane->id == PLANE_CURSOR)
812 return plane_state->hw.fb != NULL;
814 return plane_state->uapi.visible;
817 static bool intel_crtc_active(struct intel_crtc *crtc)
819 /* Be paranoid as we can arrive here with only partial
820 * state retrieved from the hardware during setup.
822 * We can ditch the adjusted_mode.crtc_clock check as soon
823 * as Haswell has gained clock readout/fastboot support.
825 * We can ditch the crtc->primary->state->fb check as soon as we can
826 * properly reconstruct framebuffers.
828 * FIXME: The intel_crtc->active here should be switched to
829 * crtc->state->active once we have proper CRTC states wired up
832 return crtc->active && crtc->base.primary->state->fb &&
833 crtc->config->hw.adjusted_mode.crtc_clock;
836 static struct intel_crtc *single_enabled_crtc(struct drm_i915_private *dev_priv)
838 struct intel_crtc *crtc, *enabled = NULL;
840 for_each_intel_crtc(&dev_priv->drm, crtc) {
841 if (intel_crtc_active(crtc)) {
851 static void pnv_update_wm(struct intel_crtc *unused_crtc)
853 struct drm_i915_private *dev_priv = to_i915(unused_crtc->base.dev);
854 struct intel_crtc *crtc;
855 const struct cxsr_latency *latency;
859 latency = intel_get_cxsr_latency(!IS_MOBILE(dev_priv),
864 drm_dbg_kms(&dev_priv->drm,
865 "Unknown FSB/MEM found, disable CxSR\n");
866 intel_set_memory_cxsr(dev_priv, false);
870 crtc = single_enabled_crtc(dev_priv);
872 const struct drm_display_mode *adjusted_mode =
873 &crtc->config->hw.adjusted_mode;
874 const struct drm_framebuffer *fb =
875 crtc->base.primary->state->fb;
876 int cpp = fb->format->cpp[0];
877 int clock = adjusted_mode->crtc_clock;
880 wm = intel_calculate_wm(clock, &pnv_display_wm,
881 pnv_display_wm.fifo_size,
882 cpp, latency->display_sr);
883 reg = I915_READ(DSPFW1);
884 reg &= ~DSPFW_SR_MASK;
885 reg |= FW_WM(wm, SR);
886 I915_WRITE(DSPFW1, reg);
887 drm_dbg_kms(&dev_priv->drm, "DSPFW1 register is %x\n", reg);
890 wm = intel_calculate_wm(clock, &pnv_cursor_wm,
891 pnv_display_wm.fifo_size,
892 4, latency->cursor_sr);
893 reg = I915_READ(DSPFW3);
894 reg &= ~DSPFW_CURSOR_SR_MASK;
895 reg |= FW_WM(wm, CURSOR_SR);
896 I915_WRITE(DSPFW3, reg);
898 /* Display HPLL off SR */
899 wm = intel_calculate_wm(clock, &pnv_display_hplloff_wm,
900 pnv_display_hplloff_wm.fifo_size,
901 cpp, latency->display_hpll_disable);
902 reg = I915_READ(DSPFW3);
903 reg &= ~DSPFW_HPLL_SR_MASK;
904 reg |= FW_WM(wm, HPLL_SR);
905 I915_WRITE(DSPFW3, reg);
907 /* cursor HPLL off SR */
908 wm = intel_calculate_wm(clock, &pnv_cursor_hplloff_wm,
909 pnv_display_hplloff_wm.fifo_size,
910 4, latency->cursor_hpll_disable);
911 reg = I915_READ(DSPFW3);
912 reg &= ~DSPFW_HPLL_CURSOR_MASK;
913 reg |= FW_WM(wm, HPLL_CURSOR);
914 I915_WRITE(DSPFW3, reg);
915 drm_dbg_kms(&dev_priv->drm, "DSPFW3 register is %x\n", reg);
917 intel_set_memory_cxsr(dev_priv, true);
919 intel_set_memory_cxsr(dev_priv, false);
924 * Documentation says:
925 * "If the line size is small, the TLB fetches can get in the way of the
926 * data fetches, causing some lag in the pixel data return which is not
927 * accounted for in the above formulas. The following adjustment only
928 * needs to be applied if eight whole lines fit in the buffer at once.
929 * The WM is adjusted upwards by the difference between the FIFO size
930 * and the size of 8 whole lines. This adjustment is always performed
931 * in the actual pixel depth regardless of whether FBC is enabled or not."
933 static unsigned int g4x_tlb_miss_wa(int fifo_size, int width, int cpp)
935 int tlb_miss = fifo_size * 64 - width * cpp * 8;
937 return max(0, tlb_miss);
940 static void g4x_write_wm_values(struct drm_i915_private *dev_priv,
941 const struct g4x_wm_values *wm)
945 for_each_pipe(dev_priv, pipe)
946 trace_g4x_wm(intel_get_crtc_for_pipe(dev_priv, pipe), wm);
949 FW_WM(wm->sr.plane, SR) |
950 FW_WM(wm->pipe[PIPE_B].plane[PLANE_CURSOR], CURSORB) |
951 FW_WM(wm->pipe[PIPE_B].plane[PLANE_PRIMARY], PLANEB) |
952 FW_WM(wm->pipe[PIPE_A].plane[PLANE_PRIMARY], PLANEA));
954 (wm->fbc_en ? DSPFW_FBC_SR_EN : 0) |
955 FW_WM(wm->sr.fbc, FBC_SR) |
956 FW_WM(wm->hpll.fbc, FBC_HPLL_SR) |
957 FW_WM(wm->pipe[PIPE_B].plane[PLANE_SPRITE0], SPRITEB) |
958 FW_WM(wm->pipe[PIPE_A].plane[PLANE_CURSOR], CURSORA) |
959 FW_WM(wm->pipe[PIPE_A].plane[PLANE_SPRITE0], SPRITEA));
961 (wm->hpll_en ? DSPFW_HPLL_SR_EN : 0) |
962 FW_WM(wm->sr.cursor, CURSOR_SR) |
963 FW_WM(wm->hpll.cursor, HPLL_CURSOR) |
964 FW_WM(wm->hpll.plane, HPLL_SR));
966 POSTING_READ(DSPFW1);
969 #define FW_WM_VLV(value, plane) \
970 (((value) << DSPFW_ ## plane ## _SHIFT) & DSPFW_ ## plane ## _MASK_VLV)
972 static void vlv_write_wm_values(struct drm_i915_private *dev_priv,
973 const struct vlv_wm_values *wm)
977 for_each_pipe(dev_priv, pipe) {
978 trace_vlv_wm(intel_get_crtc_for_pipe(dev_priv, pipe), wm);
980 I915_WRITE(VLV_DDL(pipe),
981 (wm->ddl[pipe].plane[PLANE_CURSOR] << DDL_CURSOR_SHIFT) |
982 (wm->ddl[pipe].plane[PLANE_SPRITE1] << DDL_SPRITE_SHIFT(1)) |
983 (wm->ddl[pipe].plane[PLANE_SPRITE0] << DDL_SPRITE_SHIFT(0)) |
984 (wm->ddl[pipe].plane[PLANE_PRIMARY] << DDL_PLANE_SHIFT));
988 * Zero the (unused) WM1 watermarks, and also clear all the
989 * high order bits so that there are no out of bounds values
990 * present in the registers during the reprogramming.
992 I915_WRITE(DSPHOWM, 0);
993 I915_WRITE(DSPHOWM1, 0);
994 I915_WRITE(DSPFW4, 0);
995 I915_WRITE(DSPFW5, 0);
996 I915_WRITE(DSPFW6, 0);
999 FW_WM(wm->sr.plane, SR) |
1000 FW_WM(wm->pipe[PIPE_B].plane[PLANE_CURSOR], CURSORB) |
1001 FW_WM_VLV(wm->pipe[PIPE_B].plane[PLANE_PRIMARY], PLANEB) |
1002 FW_WM_VLV(wm->pipe[PIPE_A].plane[PLANE_PRIMARY], PLANEA));
1004 FW_WM_VLV(wm->pipe[PIPE_A].plane[PLANE_SPRITE1], SPRITEB) |
1005 FW_WM(wm->pipe[PIPE_A].plane[PLANE_CURSOR], CURSORA) |
1006 FW_WM_VLV(wm->pipe[PIPE_A].plane[PLANE_SPRITE0], SPRITEA));
1008 FW_WM(wm->sr.cursor, CURSOR_SR));
1010 if (IS_CHERRYVIEW(dev_priv)) {
1011 I915_WRITE(DSPFW7_CHV,
1012 FW_WM_VLV(wm->pipe[PIPE_B].plane[PLANE_SPRITE1], SPRITED) |
1013 FW_WM_VLV(wm->pipe[PIPE_B].plane[PLANE_SPRITE0], SPRITEC));
1014 I915_WRITE(DSPFW8_CHV,
1015 FW_WM_VLV(wm->pipe[PIPE_C].plane[PLANE_SPRITE1], SPRITEF) |
1016 FW_WM_VLV(wm->pipe[PIPE_C].plane[PLANE_SPRITE0], SPRITEE));
1017 I915_WRITE(DSPFW9_CHV,
1018 FW_WM_VLV(wm->pipe[PIPE_C].plane[PLANE_PRIMARY], PLANEC) |
1019 FW_WM(wm->pipe[PIPE_C].plane[PLANE_CURSOR], CURSORC));
1021 FW_WM(wm->sr.plane >> 9, SR_HI) |
1022 FW_WM(wm->pipe[PIPE_C].plane[PLANE_SPRITE1] >> 8, SPRITEF_HI) |
1023 FW_WM(wm->pipe[PIPE_C].plane[PLANE_SPRITE0] >> 8, SPRITEE_HI) |
1024 FW_WM(wm->pipe[PIPE_C].plane[PLANE_PRIMARY] >> 8, PLANEC_HI) |
1025 FW_WM(wm->pipe[PIPE_B].plane[PLANE_SPRITE1] >> 8, SPRITED_HI) |
1026 FW_WM(wm->pipe[PIPE_B].plane[PLANE_SPRITE0] >> 8, SPRITEC_HI) |
1027 FW_WM(wm->pipe[PIPE_B].plane[PLANE_PRIMARY] >> 8, PLANEB_HI) |
1028 FW_WM(wm->pipe[PIPE_A].plane[PLANE_SPRITE1] >> 8, SPRITEB_HI) |
1029 FW_WM(wm->pipe[PIPE_A].plane[PLANE_SPRITE0] >> 8, SPRITEA_HI) |
1030 FW_WM(wm->pipe[PIPE_A].plane[PLANE_PRIMARY] >> 8, PLANEA_HI));
1033 FW_WM_VLV(wm->pipe[PIPE_B].plane[PLANE_SPRITE1], SPRITED) |
1034 FW_WM_VLV(wm->pipe[PIPE_B].plane[PLANE_SPRITE0], SPRITEC));
1036 FW_WM(wm->sr.plane >> 9, SR_HI) |
1037 FW_WM(wm->pipe[PIPE_B].plane[PLANE_SPRITE1] >> 8, SPRITED_HI) |
1038 FW_WM(wm->pipe[PIPE_B].plane[PLANE_SPRITE0] >> 8, SPRITEC_HI) |
1039 FW_WM(wm->pipe[PIPE_B].plane[PLANE_PRIMARY] >> 8, PLANEB_HI) |
1040 FW_WM(wm->pipe[PIPE_A].plane[PLANE_SPRITE1] >> 8, SPRITEB_HI) |
1041 FW_WM(wm->pipe[PIPE_A].plane[PLANE_SPRITE0] >> 8, SPRITEA_HI) |
1042 FW_WM(wm->pipe[PIPE_A].plane[PLANE_PRIMARY] >> 8, PLANEA_HI));
1045 POSTING_READ(DSPFW1);
1050 static void g4x_setup_wm_latency(struct drm_i915_private *dev_priv)
1052 /* all latencies in usec */
1053 dev_priv->wm.pri_latency[G4X_WM_LEVEL_NORMAL] = 5;
1054 dev_priv->wm.pri_latency[G4X_WM_LEVEL_SR] = 12;
1055 dev_priv->wm.pri_latency[G4X_WM_LEVEL_HPLL] = 35;
1057 dev_priv->wm.max_level = G4X_WM_LEVEL_HPLL;
1060 static int g4x_plane_fifo_size(enum plane_id plane_id, int level)
1063 * DSPCNTR[13] supposedly controls whether the
1064 * primary plane can use the FIFO space otherwise
1065 * reserved for the sprite plane. It's not 100% clear
1066 * what the actual FIFO size is, but it looks like we
1067 * can happily set both primary and sprite watermarks
1068 * up to 127 cachelines. So that would seem to mean
1069 * that either DSPCNTR[13] doesn't do anything, or that
1070 * the total FIFO is >= 256 cachelines in size. Either
1071 * way, we don't seem to have to worry about this
1072 * repartitioning as the maximum watermark value the
1073 * register can hold for each plane is lower than the
1074 * minimum FIFO size.
1080 return level == G4X_WM_LEVEL_NORMAL ? 127 : 511;
1082 return level == G4X_WM_LEVEL_NORMAL ? 127 : 0;
1084 MISSING_CASE(plane_id);
1089 static int g4x_fbc_fifo_size(int level)
1092 case G4X_WM_LEVEL_SR:
1094 case G4X_WM_LEVEL_HPLL:
1097 MISSING_CASE(level);
1102 static u16 g4x_compute_wm(const struct intel_crtc_state *crtc_state,
1103 const struct intel_plane_state *plane_state,
1106 struct intel_plane *plane = to_intel_plane(plane_state->uapi.plane);
1107 struct drm_i915_private *dev_priv = to_i915(plane->base.dev);
1108 const struct drm_display_mode *adjusted_mode =
1109 &crtc_state->hw.adjusted_mode;
1110 unsigned int latency = dev_priv->wm.pri_latency[level] * 10;
1111 unsigned int clock, htotal, cpp, width, wm;
1116 if (!intel_wm_plane_visible(crtc_state, plane_state))
1119 cpp = plane_state->hw.fb->format->cpp[0];
1122 * Not 100% sure which way ELK should go here as the
1123 * spec only says CL/CTG should assume 32bpp and BW
1124 * doesn't need to. But as these things followed the
1125 * mobile vs. desktop lines on gen3 as well, let's
1126 * assume ELK doesn't need this.
1128 * The spec also fails to list such a restriction for
1129 * the HPLL watermark, which seems a little strange.
1130 * Let's use 32bpp for the HPLL watermark as well.
1132 if (IS_GM45(dev_priv) && plane->id == PLANE_PRIMARY &&
1133 level != G4X_WM_LEVEL_NORMAL)
1136 clock = adjusted_mode->crtc_clock;
1137 htotal = adjusted_mode->crtc_htotal;
1139 width = drm_rect_width(&plane_state->uapi.dst);
1141 if (plane->id == PLANE_CURSOR) {
1142 wm = intel_wm_method2(clock, htotal, width, cpp, latency);
1143 } else if (plane->id == PLANE_PRIMARY &&
1144 level == G4X_WM_LEVEL_NORMAL) {
1145 wm = intel_wm_method1(clock, cpp, latency);
1147 unsigned int small, large;
1149 small = intel_wm_method1(clock, cpp, latency);
1150 large = intel_wm_method2(clock, htotal, width, cpp, latency);
1152 wm = min(small, large);
1155 wm += g4x_tlb_miss_wa(g4x_plane_fifo_size(plane->id, level),
1158 wm = DIV_ROUND_UP(wm, 64) + 2;
1160 return min_t(unsigned int, wm, USHRT_MAX);
1163 static bool g4x_raw_plane_wm_set(struct intel_crtc_state *crtc_state,
1164 int level, enum plane_id plane_id, u16 value)
1166 struct drm_i915_private *dev_priv = to_i915(crtc_state->uapi.crtc->dev);
1169 for (; level < intel_wm_num_levels(dev_priv); level++) {
1170 struct g4x_pipe_wm *raw = &crtc_state->wm.g4x.raw[level];
1172 dirty |= raw->plane[plane_id] != value;
1173 raw->plane[plane_id] = value;
1179 static bool g4x_raw_fbc_wm_set(struct intel_crtc_state *crtc_state,
1180 int level, u16 value)
1182 struct drm_i915_private *dev_priv = to_i915(crtc_state->uapi.crtc->dev);
1185 /* NORMAL level doesn't have an FBC watermark */
1186 level = max(level, G4X_WM_LEVEL_SR);
1188 for (; level < intel_wm_num_levels(dev_priv); level++) {
1189 struct g4x_pipe_wm *raw = &crtc_state->wm.g4x.raw[level];
1191 dirty |= raw->fbc != value;
1198 static u32 ilk_compute_fbc_wm(const struct intel_crtc_state *crtc_state,
1199 const struct intel_plane_state *plane_state,
1202 static bool g4x_raw_plane_wm_compute(struct intel_crtc_state *crtc_state,
1203 const struct intel_plane_state *plane_state)
1205 struct intel_plane *plane = to_intel_plane(plane_state->uapi.plane);
1206 struct drm_i915_private *dev_priv = to_i915(crtc_state->uapi.crtc->dev);
1207 int num_levels = intel_wm_num_levels(to_i915(plane->base.dev));
1208 enum plane_id plane_id = plane->id;
1212 if (!intel_wm_plane_visible(crtc_state, plane_state)) {
1213 dirty |= g4x_raw_plane_wm_set(crtc_state, 0, plane_id, 0);
1214 if (plane_id == PLANE_PRIMARY)
1215 dirty |= g4x_raw_fbc_wm_set(crtc_state, 0, 0);
1219 for (level = 0; level < num_levels; level++) {
1220 struct g4x_pipe_wm *raw = &crtc_state->wm.g4x.raw[level];
1223 wm = g4x_compute_wm(crtc_state, plane_state, level);
1224 max_wm = g4x_plane_fifo_size(plane_id, level);
1229 dirty |= raw->plane[plane_id] != wm;
1230 raw->plane[plane_id] = wm;
1232 if (plane_id != PLANE_PRIMARY ||
1233 level == G4X_WM_LEVEL_NORMAL)
1236 wm = ilk_compute_fbc_wm(crtc_state, plane_state,
1237 raw->plane[plane_id]);
1238 max_wm = g4x_fbc_fifo_size(level);
1241 * FBC wm is not mandatory as we
1242 * can always just disable its use.
1247 dirty |= raw->fbc != wm;
1251 /* mark watermarks as invalid */
1252 dirty |= g4x_raw_plane_wm_set(crtc_state, level, plane_id, USHRT_MAX);
1254 if (plane_id == PLANE_PRIMARY)
1255 dirty |= g4x_raw_fbc_wm_set(crtc_state, level, USHRT_MAX);
1259 drm_dbg_kms(&dev_priv->drm,
1260 "%s watermarks: normal=%d, SR=%d, HPLL=%d\n",
1262 crtc_state->wm.g4x.raw[G4X_WM_LEVEL_NORMAL].plane[plane_id],
1263 crtc_state->wm.g4x.raw[G4X_WM_LEVEL_SR].plane[plane_id],
1264 crtc_state->wm.g4x.raw[G4X_WM_LEVEL_HPLL].plane[plane_id]);
1266 if (plane_id == PLANE_PRIMARY)
1267 drm_dbg_kms(&dev_priv->drm,
1268 "FBC watermarks: SR=%d, HPLL=%d\n",
1269 crtc_state->wm.g4x.raw[G4X_WM_LEVEL_SR].fbc,
1270 crtc_state->wm.g4x.raw[G4X_WM_LEVEL_HPLL].fbc);
1276 static bool g4x_raw_plane_wm_is_valid(const struct intel_crtc_state *crtc_state,
1277 enum plane_id plane_id, int level)
1279 const struct g4x_pipe_wm *raw = &crtc_state->wm.g4x.raw[level];
1281 return raw->plane[plane_id] <= g4x_plane_fifo_size(plane_id, level);
1284 static bool g4x_raw_crtc_wm_is_valid(const struct intel_crtc_state *crtc_state,
1287 struct drm_i915_private *dev_priv = to_i915(crtc_state->uapi.crtc->dev);
1289 if (level > dev_priv->wm.max_level)
1292 return g4x_raw_plane_wm_is_valid(crtc_state, PLANE_PRIMARY, level) &&
1293 g4x_raw_plane_wm_is_valid(crtc_state, PLANE_SPRITE0, level) &&
1294 g4x_raw_plane_wm_is_valid(crtc_state, PLANE_CURSOR, level);
1297 /* mark all levels starting from 'level' as invalid */
1298 static void g4x_invalidate_wms(struct intel_crtc *crtc,
1299 struct g4x_wm_state *wm_state, int level)
1301 if (level <= G4X_WM_LEVEL_NORMAL) {
1302 enum plane_id plane_id;
1304 for_each_plane_id_on_crtc(crtc, plane_id)
1305 wm_state->wm.plane[plane_id] = USHRT_MAX;
1308 if (level <= G4X_WM_LEVEL_SR) {
1309 wm_state->cxsr = false;
1310 wm_state->sr.cursor = USHRT_MAX;
1311 wm_state->sr.plane = USHRT_MAX;
1312 wm_state->sr.fbc = USHRT_MAX;
1315 if (level <= G4X_WM_LEVEL_HPLL) {
1316 wm_state->hpll_en = false;
1317 wm_state->hpll.cursor = USHRT_MAX;
1318 wm_state->hpll.plane = USHRT_MAX;
1319 wm_state->hpll.fbc = USHRT_MAX;
1323 static int g4x_compute_pipe_wm(struct intel_crtc_state *crtc_state)
1325 struct intel_crtc *crtc = to_intel_crtc(crtc_state->uapi.crtc);
1326 struct intel_atomic_state *state =
1327 to_intel_atomic_state(crtc_state->uapi.state);
1328 struct g4x_wm_state *wm_state = &crtc_state->wm.g4x.optimal;
1329 int num_active_planes = hweight8(crtc_state->active_planes &
1330 ~BIT(PLANE_CURSOR));
1331 const struct g4x_pipe_wm *raw;
1332 const struct intel_plane_state *old_plane_state;
1333 const struct intel_plane_state *new_plane_state;
1334 struct intel_plane *plane;
1335 enum plane_id plane_id;
1337 unsigned int dirty = 0;
1339 for_each_oldnew_intel_plane_in_state(state, plane,
1341 new_plane_state, i) {
1342 if (new_plane_state->hw.crtc != &crtc->base &&
1343 old_plane_state->hw.crtc != &crtc->base)
1346 if (g4x_raw_plane_wm_compute(crtc_state, new_plane_state))
1347 dirty |= BIT(plane->id);
1353 level = G4X_WM_LEVEL_NORMAL;
1354 if (!g4x_raw_crtc_wm_is_valid(crtc_state, level))
1357 raw = &crtc_state->wm.g4x.raw[level];
1358 for_each_plane_id_on_crtc(crtc, plane_id)
1359 wm_state->wm.plane[plane_id] = raw->plane[plane_id];
1361 level = G4X_WM_LEVEL_SR;
1363 if (!g4x_raw_crtc_wm_is_valid(crtc_state, level))
1366 raw = &crtc_state->wm.g4x.raw[level];
1367 wm_state->sr.plane = raw->plane[PLANE_PRIMARY];
1368 wm_state->sr.cursor = raw->plane[PLANE_CURSOR];
1369 wm_state->sr.fbc = raw->fbc;
1371 wm_state->cxsr = num_active_planes == BIT(PLANE_PRIMARY);
1373 level = G4X_WM_LEVEL_HPLL;
1375 if (!g4x_raw_crtc_wm_is_valid(crtc_state, level))
1378 raw = &crtc_state->wm.g4x.raw[level];
1379 wm_state->hpll.plane = raw->plane[PLANE_PRIMARY];
1380 wm_state->hpll.cursor = raw->plane[PLANE_CURSOR];
1381 wm_state->hpll.fbc = raw->fbc;
1383 wm_state->hpll_en = wm_state->cxsr;
1388 if (level == G4X_WM_LEVEL_NORMAL)
1391 /* invalidate the higher levels */
1392 g4x_invalidate_wms(crtc, wm_state, level);
1395 * Determine if the FBC watermark(s) can be used. IF
1396 * this isn't the case we prefer to disable the FBC
1397 ( watermark(s) rather than disable the SR/HPLL
1398 * level(s) entirely.
1400 wm_state->fbc_en = level > G4X_WM_LEVEL_NORMAL;
1402 if (level >= G4X_WM_LEVEL_SR &&
1403 wm_state->sr.fbc > g4x_fbc_fifo_size(G4X_WM_LEVEL_SR))
1404 wm_state->fbc_en = false;
1405 else if (level >= G4X_WM_LEVEL_HPLL &&
1406 wm_state->hpll.fbc > g4x_fbc_fifo_size(G4X_WM_LEVEL_HPLL))
1407 wm_state->fbc_en = false;
1412 static int g4x_compute_intermediate_wm(struct intel_crtc_state *new_crtc_state)
1414 struct intel_crtc *crtc = to_intel_crtc(new_crtc_state->uapi.crtc);
1415 struct g4x_wm_state *intermediate = &new_crtc_state->wm.g4x.intermediate;
1416 const struct g4x_wm_state *optimal = &new_crtc_state->wm.g4x.optimal;
1417 struct intel_atomic_state *intel_state =
1418 to_intel_atomic_state(new_crtc_state->uapi.state);
1419 const struct intel_crtc_state *old_crtc_state =
1420 intel_atomic_get_old_crtc_state(intel_state, crtc);
1421 const struct g4x_wm_state *active = &old_crtc_state->wm.g4x.optimal;
1422 enum plane_id plane_id;
1424 if (!new_crtc_state->hw.active || drm_atomic_crtc_needs_modeset(&new_crtc_state->uapi)) {
1425 *intermediate = *optimal;
1427 intermediate->cxsr = false;
1428 intermediate->hpll_en = false;
1432 intermediate->cxsr = optimal->cxsr && active->cxsr &&
1433 !new_crtc_state->disable_cxsr;
1434 intermediate->hpll_en = optimal->hpll_en && active->hpll_en &&
1435 !new_crtc_state->disable_cxsr;
1436 intermediate->fbc_en = optimal->fbc_en && active->fbc_en;
1438 for_each_plane_id_on_crtc(crtc, plane_id) {
1439 intermediate->wm.plane[plane_id] =
1440 max(optimal->wm.plane[plane_id],
1441 active->wm.plane[plane_id]);
1443 WARN_ON(intermediate->wm.plane[plane_id] >
1444 g4x_plane_fifo_size(plane_id, G4X_WM_LEVEL_NORMAL));
1447 intermediate->sr.plane = max(optimal->sr.plane,
1449 intermediate->sr.cursor = max(optimal->sr.cursor,
1451 intermediate->sr.fbc = max(optimal->sr.fbc,
1454 intermediate->hpll.plane = max(optimal->hpll.plane,
1455 active->hpll.plane);
1456 intermediate->hpll.cursor = max(optimal->hpll.cursor,
1457 active->hpll.cursor);
1458 intermediate->hpll.fbc = max(optimal->hpll.fbc,
1461 WARN_ON((intermediate->sr.plane >
1462 g4x_plane_fifo_size(PLANE_PRIMARY, G4X_WM_LEVEL_SR) ||
1463 intermediate->sr.cursor >
1464 g4x_plane_fifo_size(PLANE_CURSOR, G4X_WM_LEVEL_SR)) &&
1465 intermediate->cxsr);
1466 WARN_ON((intermediate->sr.plane >
1467 g4x_plane_fifo_size(PLANE_PRIMARY, G4X_WM_LEVEL_HPLL) ||
1468 intermediate->sr.cursor >
1469 g4x_plane_fifo_size(PLANE_CURSOR, G4X_WM_LEVEL_HPLL)) &&
1470 intermediate->hpll_en);
1472 WARN_ON(intermediate->sr.fbc > g4x_fbc_fifo_size(1) &&
1473 intermediate->fbc_en && intermediate->cxsr);
1474 WARN_ON(intermediate->hpll.fbc > g4x_fbc_fifo_size(2) &&
1475 intermediate->fbc_en && intermediate->hpll_en);
1479 * If our intermediate WM are identical to the final WM, then we can
1480 * omit the post-vblank programming; only update if it's different.
1482 if (memcmp(intermediate, optimal, sizeof(*intermediate)) != 0)
1483 new_crtc_state->wm.need_postvbl_update = true;
1488 static void g4x_merge_wm(struct drm_i915_private *dev_priv,
1489 struct g4x_wm_values *wm)
1491 struct intel_crtc *crtc;
1492 int num_active_pipes = 0;
1498 for_each_intel_crtc(&dev_priv->drm, crtc) {
1499 const struct g4x_wm_state *wm_state = &crtc->wm.active.g4x;
1504 if (!wm_state->cxsr)
1506 if (!wm_state->hpll_en)
1507 wm->hpll_en = false;
1508 if (!wm_state->fbc_en)
1514 if (num_active_pipes != 1) {
1516 wm->hpll_en = false;
1520 for_each_intel_crtc(&dev_priv->drm, crtc) {
1521 const struct g4x_wm_state *wm_state = &crtc->wm.active.g4x;
1522 enum pipe pipe = crtc->pipe;
1524 wm->pipe[pipe] = wm_state->wm;
1525 if (crtc->active && wm->cxsr)
1526 wm->sr = wm_state->sr;
1527 if (crtc->active && wm->hpll_en)
1528 wm->hpll = wm_state->hpll;
1532 static void g4x_program_watermarks(struct drm_i915_private *dev_priv)
1534 struct g4x_wm_values *old_wm = &dev_priv->wm.g4x;
1535 struct g4x_wm_values new_wm = {};
1537 g4x_merge_wm(dev_priv, &new_wm);
1539 if (memcmp(old_wm, &new_wm, sizeof(new_wm)) == 0)
1542 if (is_disabling(old_wm->cxsr, new_wm.cxsr, true))
1543 _intel_set_memory_cxsr(dev_priv, false);
1545 g4x_write_wm_values(dev_priv, &new_wm);
1547 if (is_enabling(old_wm->cxsr, new_wm.cxsr, true))
1548 _intel_set_memory_cxsr(dev_priv, true);
1553 static void g4x_initial_watermarks(struct intel_atomic_state *state,
1554 struct intel_crtc *crtc)
1556 struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
1557 const struct intel_crtc_state *crtc_state =
1558 intel_atomic_get_new_crtc_state(state, crtc);
1560 mutex_lock(&dev_priv->wm.wm_mutex);
1561 crtc->wm.active.g4x = crtc_state->wm.g4x.intermediate;
1562 g4x_program_watermarks(dev_priv);
1563 mutex_unlock(&dev_priv->wm.wm_mutex);
1566 static void g4x_optimize_watermarks(struct intel_atomic_state *state,
1567 struct intel_crtc *crtc)
1569 struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
1570 const struct intel_crtc_state *crtc_state =
1571 intel_atomic_get_new_crtc_state(state, crtc);
1573 if (!crtc_state->wm.need_postvbl_update)
1576 mutex_lock(&dev_priv->wm.wm_mutex);
1577 crtc->wm.active.g4x = crtc_state->wm.g4x.optimal;
1578 g4x_program_watermarks(dev_priv);
1579 mutex_unlock(&dev_priv->wm.wm_mutex);
1582 /* latency must be in 0.1us units. */
1583 static unsigned int vlv_wm_method2(unsigned int pixel_rate,
1584 unsigned int htotal,
1587 unsigned int latency)
1591 ret = intel_wm_method2(pixel_rate, htotal,
1592 width, cpp, latency);
1593 ret = DIV_ROUND_UP(ret, 64);
1598 static void vlv_setup_wm_latency(struct drm_i915_private *dev_priv)
1600 /* all latencies in usec */
1601 dev_priv->wm.pri_latency[VLV_WM_LEVEL_PM2] = 3;
1603 dev_priv->wm.max_level = VLV_WM_LEVEL_PM2;
1605 if (IS_CHERRYVIEW(dev_priv)) {
1606 dev_priv->wm.pri_latency[VLV_WM_LEVEL_PM5] = 12;
1607 dev_priv->wm.pri_latency[VLV_WM_LEVEL_DDR_DVFS] = 33;
1609 dev_priv->wm.max_level = VLV_WM_LEVEL_DDR_DVFS;
1613 static u16 vlv_compute_wm_level(const struct intel_crtc_state *crtc_state,
1614 const struct intel_plane_state *plane_state,
1617 struct intel_plane *plane = to_intel_plane(plane_state->uapi.plane);
1618 struct drm_i915_private *dev_priv = to_i915(plane->base.dev);
1619 const struct drm_display_mode *adjusted_mode =
1620 &crtc_state->hw.adjusted_mode;
1621 unsigned int clock, htotal, cpp, width, wm;
1623 if (dev_priv->wm.pri_latency[level] == 0)
1626 if (!intel_wm_plane_visible(crtc_state, plane_state))
1629 cpp = plane_state->hw.fb->format->cpp[0];
1630 clock = adjusted_mode->crtc_clock;
1631 htotal = adjusted_mode->crtc_htotal;
1632 width = crtc_state->pipe_src_w;
1634 if (plane->id == PLANE_CURSOR) {
1636 * FIXME the formula gives values that are
1637 * too big for the cursor FIFO, and hence we
1638 * would never be able to use cursors. For
1639 * now just hardcode the watermark.
1643 wm = vlv_wm_method2(clock, htotal, width, cpp,
1644 dev_priv->wm.pri_latency[level] * 10);
1647 return min_t(unsigned int, wm, USHRT_MAX);
1650 static bool vlv_need_sprite0_fifo_workaround(unsigned int active_planes)
1652 return (active_planes & (BIT(PLANE_SPRITE0) |
1653 BIT(PLANE_SPRITE1))) == BIT(PLANE_SPRITE1);
1656 static int vlv_compute_fifo(struct intel_crtc_state *crtc_state)
1658 struct intel_crtc *crtc = to_intel_crtc(crtc_state->uapi.crtc);
1659 const struct g4x_pipe_wm *raw =
1660 &crtc_state->wm.vlv.raw[VLV_WM_LEVEL_PM2];
1661 struct vlv_fifo_state *fifo_state = &crtc_state->wm.vlv.fifo_state;
1662 unsigned int active_planes = crtc_state->active_planes & ~BIT(PLANE_CURSOR);
1663 int num_active_planes = hweight8(active_planes);
1664 const int fifo_size = 511;
1665 int fifo_extra, fifo_left = fifo_size;
1666 int sprite0_fifo_extra = 0;
1667 unsigned int total_rate;
1668 enum plane_id plane_id;
1671 * When enabling sprite0 after sprite1 has already been enabled
1672 * we tend to get an underrun unless sprite0 already has some
1673 * FIFO space allcoated. Hence we always allocate at least one
1674 * cacheline for sprite0 whenever sprite1 is enabled.
1676 * All other plane enable sequences appear immune to this problem.
1678 if (vlv_need_sprite0_fifo_workaround(active_planes))
1679 sprite0_fifo_extra = 1;
1681 total_rate = raw->plane[PLANE_PRIMARY] +
1682 raw->plane[PLANE_SPRITE0] +
1683 raw->plane[PLANE_SPRITE1] +
1686 if (total_rate > fifo_size)
1689 if (total_rate == 0)
1692 for_each_plane_id_on_crtc(crtc, plane_id) {
1695 if ((active_planes & BIT(plane_id)) == 0) {
1696 fifo_state->plane[plane_id] = 0;
1700 rate = raw->plane[plane_id];
1701 fifo_state->plane[plane_id] = fifo_size * rate / total_rate;
1702 fifo_left -= fifo_state->plane[plane_id];
1705 fifo_state->plane[PLANE_SPRITE0] += sprite0_fifo_extra;
1706 fifo_left -= sprite0_fifo_extra;
1708 fifo_state->plane[PLANE_CURSOR] = 63;
1710 fifo_extra = DIV_ROUND_UP(fifo_left, num_active_planes ?: 1);
1712 /* spread the remainder evenly */
1713 for_each_plane_id_on_crtc(crtc, plane_id) {
1719 if ((active_planes & BIT(plane_id)) == 0)
1722 plane_extra = min(fifo_extra, fifo_left);
1723 fifo_state->plane[plane_id] += plane_extra;
1724 fifo_left -= plane_extra;
1727 WARN_ON(active_planes != 0 && fifo_left != 0);
1729 /* give it all to the first plane if none are active */
1730 if (active_planes == 0) {
1731 WARN_ON(fifo_left != fifo_size);
1732 fifo_state->plane[PLANE_PRIMARY] = fifo_left;
1738 /* mark all levels starting from 'level' as invalid */
1739 static void vlv_invalidate_wms(struct intel_crtc *crtc,
1740 struct vlv_wm_state *wm_state, int level)
1742 struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
1744 for (; level < intel_wm_num_levels(dev_priv); level++) {
1745 enum plane_id plane_id;
1747 for_each_plane_id_on_crtc(crtc, plane_id)
1748 wm_state->wm[level].plane[plane_id] = USHRT_MAX;
1750 wm_state->sr[level].cursor = USHRT_MAX;
1751 wm_state->sr[level].plane = USHRT_MAX;
1755 static u16 vlv_invert_wm_value(u16 wm, u16 fifo_size)
1760 return fifo_size - wm;
1764 * Starting from 'level' set all higher
1765 * levels to 'value' in the "raw" watermarks.
1767 static bool vlv_raw_plane_wm_set(struct intel_crtc_state *crtc_state,
1768 int level, enum plane_id plane_id, u16 value)
1770 struct drm_i915_private *dev_priv = to_i915(crtc_state->uapi.crtc->dev);
1771 int num_levels = intel_wm_num_levels(dev_priv);
1774 for (; level < num_levels; level++) {
1775 struct g4x_pipe_wm *raw = &crtc_state->wm.vlv.raw[level];
1777 dirty |= raw->plane[plane_id] != value;
1778 raw->plane[plane_id] = value;
1784 static bool vlv_raw_plane_wm_compute(struct intel_crtc_state *crtc_state,
1785 const struct intel_plane_state *plane_state)
1787 struct intel_plane *plane = to_intel_plane(plane_state->uapi.plane);
1788 struct drm_i915_private *dev_priv = to_i915(crtc_state->uapi.crtc->dev);
1789 enum plane_id plane_id = plane->id;
1790 int num_levels = intel_wm_num_levels(to_i915(plane->base.dev));
1794 if (!intel_wm_plane_visible(crtc_state, plane_state)) {
1795 dirty |= vlv_raw_plane_wm_set(crtc_state, 0, plane_id, 0);
1799 for (level = 0; level < num_levels; level++) {
1800 struct g4x_pipe_wm *raw = &crtc_state->wm.vlv.raw[level];
1801 int wm = vlv_compute_wm_level(crtc_state, plane_state, level);
1802 int max_wm = plane_id == PLANE_CURSOR ? 63 : 511;
1807 dirty |= raw->plane[plane_id] != wm;
1808 raw->plane[plane_id] = wm;
1811 /* mark all higher levels as invalid */
1812 dirty |= vlv_raw_plane_wm_set(crtc_state, level, plane_id, USHRT_MAX);
1816 drm_dbg_kms(&dev_priv->drm,
1817 "%s watermarks: PM2=%d, PM5=%d, DDR DVFS=%d\n",
1819 crtc_state->wm.vlv.raw[VLV_WM_LEVEL_PM2].plane[plane_id],
1820 crtc_state->wm.vlv.raw[VLV_WM_LEVEL_PM5].plane[plane_id],
1821 crtc_state->wm.vlv.raw[VLV_WM_LEVEL_DDR_DVFS].plane[plane_id]);
1826 static bool vlv_raw_plane_wm_is_valid(const struct intel_crtc_state *crtc_state,
1827 enum plane_id plane_id, int level)
1829 const struct g4x_pipe_wm *raw =
1830 &crtc_state->wm.vlv.raw[level];
1831 const struct vlv_fifo_state *fifo_state =
1832 &crtc_state->wm.vlv.fifo_state;
1834 return raw->plane[plane_id] <= fifo_state->plane[plane_id];
1837 static bool vlv_raw_crtc_wm_is_valid(const struct intel_crtc_state *crtc_state, int level)
1839 return vlv_raw_plane_wm_is_valid(crtc_state, PLANE_PRIMARY, level) &&
1840 vlv_raw_plane_wm_is_valid(crtc_state, PLANE_SPRITE0, level) &&
1841 vlv_raw_plane_wm_is_valid(crtc_state, PLANE_SPRITE1, level) &&
1842 vlv_raw_plane_wm_is_valid(crtc_state, PLANE_CURSOR, level);
1845 static int vlv_compute_pipe_wm(struct intel_crtc_state *crtc_state)
1847 struct intel_crtc *crtc = to_intel_crtc(crtc_state->uapi.crtc);
1848 struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
1849 struct intel_atomic_state *state =
1850 to_intel_atomic_state(crtc_state->uapi.state);
1851 struct vlv_wm_state *wm_state = &crtc_state->wm.vlv.optimal;
1852 const struct vlv_fifo_state *fifo_state =
1853 &crtc_state->wm.vlv.fifo_state;
1854 int num_active_planes = hweight8(crtc_state->active_planes &
1855 ~BIT(PLANE_CURSOR));
1856 bool needs_modeset = drm_atomic_crtc_needs_modeset(&crtc_state->uapi);
1857 const struct intel_plane_state *old_plane_state;
1858 const struct intel_plane_state *new_plane_state;
1859 struct intel_plane *plane;
1860 enum plane_id plane_id;
1862 unsigned int dirty = 0;
1864 for_each_oldnew_intel_plane_in_state(state, plane,
1866 new_plane_state, i) {
1867 if (new_plane_state->hw.crtc != &crtc->base &&
1868 old_plane_state->hw.crtc != &crtc->base)
1871 if (vlv_raw_plane_wm_compute(crtc_state, new_plane_state))
1872 dirty |= BIT(plane->id);
1876 * DSPARB registers may have been reset due to the
1877 * power well being turned off. Make sure we restore
1878 * them to a consistent state even if no primary/sprite
1879 * planes are initially active.
1882 crtc_state->fifo_changed = true;
1887 /* cursor changes don't warrant a FIFO recompute */
1888 if (dirty & ~BIT(PLANE_CURSOR)) {
1889 const struct intel_crtc_state *old_crtc_state =
1890 intel_atomic_get_old_crtc_state(state, crtc);
1891 const struct vlv_fifo_state *old_fifo_state =
1892 &old_crtc_state->wm.vlv.fifo_state;
1894 ret = vlv_compute_fifo(crtc_state);
1898 if (needs_modeset ||
1899 memcmp(old_fifo_state, fifo_state,
1900 sizeof(*fifo_state)) != 0)
1901 crtc_state->fifo_changed = true;
1904 /* initially allow all levels */
1905 wm_state->num_levels = intel_wm_num_levels(dev_priv);
1907 * Note that enabling cxsr with no primary/sprite planes
1908 * enabled can wedge the pipe. Hence we only allow cxsr
1909 * with exactly one enabled primary/sprite plane.
1911 wm_state->cxsr = crtc->pipe != PIPE_C && num_active_planes == 1;
1913 for (level = 0; level < wm_state->num_levels; level++) {
1914 const struct g4x_pipe_wm *raw = &crtc_state->wm.vlv.raw[level];
1915 const int sr_fifo_size = INTEL_NUM_PIPES(dev_priv) * 512 - 1;
1917 if (!vlv_raw_crtc_wm_is_valid(crtc_state, level))
1920 for_each_plane_id_on_crtc(crtc, plane_id) {
1921 wm_state->wm[level].plane[plane_id] =
1922 vlv_invert_wm_value(raw->plane[plane_id],
1923 fifo_state->plane[plane_id]);
1926 wm_state->sr[level].plane =
1927 vlv_invert_wm_value(max3(raw->plane[PLANE_PRIMARY],
1928 raw->plane[PLANE_SPRITE0],
1929 raw->plane[PLANE_SPRITE1]),
1932 wm_state->sr[level].cursor =
1933 vlv_invert_wm_value(raw->plane[PLANE_CURSOR],
1940 /* limit to only levels we can actually handle */
1941 wm_state->num_levels = level;
1943 /* invalidate the higher levels */
1944 vlv_invalidate_wms(crtc, wm_state, level);
1949 #define VLV_FIFO(plane, value) \
1950 (((value) << DSPARB_ ## plane ## _SHIFT_VLV) & DSPARB_ ## plane ## _MASK_VLV)
1952 static void vlv_atomic_update_fifo(struct intel_atomic_state *state,
1953 struct intel_crtc *crtc)
1955 struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
1956 struct intel_uncore *uncore = &dev_priv->uncore;
1957 const struct intel_crtc_state *crtc_state =
1958 intel_atomic_get_new_crtc_state(state, crtc);
1959 const struct vlv_fifo_state *fifo_state =
1960 &crtc_state->wm.vlv.fifo_state;
1961 int sprite0_start, sprite1_start, fifo_size;
1962 u32 dsparb, dsparb2, dsparb3;
1964 if (!crtc_state->fifo_changed)
1967 sprite0_start = fifo_state->plane[PLANE_PRIMARY];
1968 sprite1_start = fifo_state->plane[PLANE_SPRITE0] + sprite0_start;
1969 fifo_size = fifo_state->plane[PLANE_SPRITE1] + sprite1_start;
1971 drm_WARN_ON(&dev_priv->drm, fifo_state->plane[PLANE_CURSOR] != 63);
1972 drm_WARN_ON(&dev_priv->drm, fifo_size != 511);
1974 trace_vlv_fifo_size(crtc, sprite0_start, sprite1_start, fifo_size);
1977 * uncore.lock serves a double purpose here. It allows us to
1978 * use the less expensive I915_{READ,WRITE}_FW() functions, and
1979 * it protects the DSPARB registers from getting clobbered by
1980 * parallel updates from multiple pipes.
1982 * intel_pipe_update_start() has already disabled interrupts
1983 * for us, so a plain spin_lock() is sufficient here.
1985 spin_lock(&uncore->lock);
1987 switch (crtc->pipe) {
1989 dsparb = intel_uncore_read_fw(uncore, DSPARB);
1990 dsparb2 = intel_uncore_read_fw(uncore, DSPARB2);
1992 dsparb &= ~(VLV_FIFO(SPRITEA, 0xff) |
1993 VLV_FIFO(SPRITEB, 0xff));
1994 dsparb |= (VLV_FIFO(SPRITEA, sprite0_start) |
1995 VLV_FIFO(SPRITEB, sprite1_start));
1997 dsparb2 &= ~(VLV_FIFO(SPRITEA_HI, 0x1) |
1998 VLV_FIFO(SPRITEB_HI, 0x1));
1999 dsparb2 |= (VLV_FIFO(SPRITEA_HI, sprite0_start >> 8) |
2000 VLV_FIFO(SPRITEB_HI, sprite1_start >> 8));
2002 intel_uncore_write_fw(uncore, DSPARB, dsparb);
2003 intel_uncore_write_fw(uncore, DSPARB2, dsparb2);
2006 dsparb = intel_uncore_read_fw(uncore, DSPARB);
2007 dsparb2 = intel_uncore_read_fw(uncore, DSPARB2);
2009 dsparb &= ~(VLV_FIFO(SPRITEC, 0xff) |
2010 VLV_FIFO(SPRITED, 0xff));
2011 dsparb |= (VLV_FIFO(SPRITEC, sprite0_start) |
2012 VLV_FIFO(SPRITED, sprite1_start));
2014 dsparb2 &= ~(VLV_FIFO(SPRITEC_HI, 0xff) |
2015 VLV_FIFO(SPRITED_HI, 0xff));
2016 dsparb2 |= (VLV_FIFO(SPRITEC_HI, sprite0_start >> 8) |
2017 VLV_FIFO(SPRITED_HI, sprite1_start >> 8));
2019 intel_uncore_write_fw(uncore, DSPARB, dsparb);
2020 intel_uncore_write_fw(uncore, DSPARB2, dsparb2);
2023 dsparb3 = intel_uncore_read_fw(uncore, DSPARB3);
2024 dsparb2 = intel_uncore_read_fw(uncore, DSPARB2);
2026 dsparb3 &= ~(VLV_FIFO(SPRITEE, 0xff) |
2027 VLV_FIFO(SPRITEF, 0xff));
2028 dsparb3 |= (VLV_FIFO(SPRITEE, sprite0_start) |
2029 VLV_FIFO(SPRITEF, sprite1_start));
2031 dsparb2 &= ~(VLV_FIFO(SPRITEE_HI, 0xff) |
2032 VLV_FIFO(SPRITEF_HI, 0xff));
2033 dsparb2 |= (VLV_FIFO(SPRITEE_HI, sprite0_start >> 8) |
2034 VLV_FIFO(SPRITEF_HI, sprite1_start >> 8));
2036 intel_uncore_write_fw(uncore, DSPARB3, dsparb3);
2037 intel_uncore_write_fw(uncore, DSPARB2, dsparb2);
2043 intel_uncore_posting_read_fw(uncore, DSPARB);
2045 spin_unlock(&uncore->lock);
2050 static int vlv_compute_intermediate_wm(struct intel_crtc_state *new_crtc_state)
2052 struct intel_crtc *crtc = to_intel_crtc(new_crtc_state->uapi.crtc);
2053 struct vlv_wm_state *intermediate = &new_crtc_state->wm.vlv.intermediate;
2054 const struct vlv_wm_state *optimal = &new_crtc_state->wm.vlv.optimal;
2055 struct intel_atomic_state *intel_state =
2056 to_intel_atomic_state(new_crtc_state->uapi.state);
2057 const struct intel_crtc_state *old_crtc_state =
2058 intel_atomic_get_old_crtc_state(intel_state, crtc);
2059 const struct vlv_wm_state *active = &old_crtc_state->wm.vlv.optimal;
2062 if (!new_crtc_state->hw.active || drm_atomic_crtc_needs_modeset(&new_crtc_state->uapi)) {
2063 *intermediate = *optimal;
2065 intermediate->cxsr = false;
2069 intermediate->num_levels = min(optimal->num_levels, active->num_levels);
2070 intermediate->cxsr = optimal->cxsr && active->cxsr &&
2071 !new_crtc_state->disable_cxsr;
2073 for (level = 0; level < intermediate->num_levels; level++) {
2074 enum plane_id plane_id;
2076 for_each_plane_id_on_crtc(crtc, plane_id) {
2077 intermediate->wm[level].plane[plane_id] =
2078 min(optimal->wm[level].plane[plane_id],
2079 active->wm[level].plane[plane_id]);
2082 intermediate->sr[level].plane = min(optimal->sr[level].plane,
2083 active->sr[level].plane);
2084 intermediate->sr[level].cursor = min(optimal->sr[level].cursor,
2085 active->sr[level].cursor);
2088 vlv_invalidate_wms(crtc, intermediate, level);
2092 * If our intermediate WM are identical to the final WM, then we can
2093 * omit the post-vblank programming; only update if it's different.
2095 if (memcmp(intermediate, optimal, sizeof(*intermediate)) != 0)
2096 new_crtc_state->wm.need_postvbl_update = true;
2101 static void vlv_merge_wm(struct drm_i915_private *dev_priv,
2102 struct vlv_wm_values *wm)
2104 struct intel_crtc *crtc;
2105 int num_active_pipes = 0;
2107 wm->level = dev_priv->wm.max_level;
2110 for_each_intel_crtc(&dev_priv->drm, crtc) {
2111 const struct vlv_wm_state *wm_state = &crtc->wm.active.vlv;
2116 if (!wm_state->cxsr)
2120 wm->level = min_t(int, wm->level, wm_state->num_levels - 1);
2123 if (num_active_pipes != 1)
2126 if (num_active_pipes > 1)
2127 wm->level = VLV_WM_LEVEL_PM2;
2129 for_each_intel_crtc(&dev_priv->drm, crtc) {
2130 const struct vlv_wm_state *wm_state = &crtc->wm.active.vlv;
2131 enum pipe pipe = crtc->pipe;
2133 wm->pipe[pipe] = wm_state->wm[wm->level];
2134 if (crtc->active && wm->cxsr)
2135 wm->sr = wm_state->sr[wm->level];
2137 wm->ddl[pipe].plane[PLANE_PRIMARY] = DDL_PRECISION_HIGH | 2;
2138 wm->ddl[pipe].plane[PLANE_SPRITE0] = DDL_PRECISION_HIGH | 2;
2139 wm->ddl[pipe].plane[PLANE_SPRITE1] = DDL_PRECISION_HIGH | 2;
2140 wm->ddl[pipe].plane[PLANE_CURSOR] = DDL_PRECISION_HIGH | 2;
2144 static void vlv_program_watermarks(struct drm_i915_private *dev_priv)
2146 struct vlv_wm_values *old_wm = &dev_priv->wm.vlv;
2147 struct vlv_wm_values new_wm = {};
2149 vlv_merge_wm(dev_priv, &new_wm);
2151 if (memcmp(old_wm, &new_wm, sizeof(new_wm)) == 0)
2154 if (is_disabling(old_wm->level, new_wm.level, VLV_WM_LEVEL_DDR_DVFS))
2155 chv_set_memory_dvfs(dev_priv, false);
2157 if (is_disabling(old_wm->level, new_wm.level, VLV_WM_LEVEL_PM5))
2158 chv_set_memory_pm5(dev_priv, false);
2160 if (is_disabling(old_wm->cxsr, new_wm.cxsr, true))
2161 _intel_set_memory_cxsr(dev_priv, false);
2163 vlv_write_wm_values(dev_priv, &new_wm);
2165 if (is_enabling(old_wm->cxsr, new_wm.cxsr, true))
2166 _intel_set_memory_cxsr(dev_priv, true);
2168 if (is_enabling(old_wm->level, new_wm.level, VLV_WM_LEVEL_PM5))
2169 chv_set_memory_pm5(dev_priv, true);
2171 if (is_enabling(old_wm->level, new_wm.level, VLV_WM_LEVEL_DDR_DVFS))
2172 chv_set_memory_dvfs(dev_priv, true);
2177 static void vlv_initial_watermarks(struct intel_atomic_state *state,
2178 struct intel_crtc *crtc)
2180 struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
2181 const struct intel_crtc_state *crtc_state =
2182 intel_atomic_get_new_crtc_state(state, crtc);
2184 mutex_lock(&dev_priv->wm.wm_mutex);
2185 crtc->wm.active.vlv = crtc_state->wm.vlv.intermediate;
2186 vlv_program_watermarks(dev_priv);
2187 mutex_unlock(&dev_priv->wm.wm_mutex);
2190 static void vlv_optimize_watermarks(struct intel_atomic_state *state,
2191 struct intel_crtc *crtc)
2193 struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
2194 const struct intel_crtc_state *crtc_state =
2195 intel_atomic_get_new_crtc_state(state, crtc);
2197 if (!crtc_state->wm.need_postvbl_update)
2200 mutex_lock(&dev_priv->wm.wm_mutex);
2201 crtc->wm.active.vlv = crtc_state->wm.vlv.optimal;
2202 vlv_program_watermarks(dev_priv);
2203 mutex_unlock(&dev_priv->wm.wm_mutex);
2206 static void i965_update_wm(struct intel_crtc *unused_crtc)
2208 struct drm_i915_private *dev_priv = to_i915(unused_crtc->base.dev);
2209 struct intel_crtc *crtc;
2214 /* Calc sr entries for one plane configs */
2215 crtc = single_enabled_crtc(dev_priv);
2217 /* self-refresh has much higher latency */
2218 static const int sr_latency_ns = 12000;
2219 const struct drm_display_mode *adjusted_mode =
2220 &crtc->config->hw.adjusted_mode;
2221 const struct drm_framebuffer *fb =
2222 crtc->base.primary->state->fb;
2223 int clock = adjusted_mode->crtc_clock;
2224 int htotal = adjusted_mode->crtc_htotal;
2225 int hdisplay = crtc->config->pipe_src_w;
2226 int cpp = fb->format->cpp[0];
2229 entries = intel_wm_method2(clock, htotal,
2230 hdisplay, cpp, sr_latency_ns / 100);
2231 entries = DIV_ROUND_UP(entries, I915_FIFO_LINE_SIZE);
2232 srwm = I965_FIFO_SIZE - entries;
2236 drm_dbg_kms(&dev_priv->drm,
2237 "self-refresh entries: %d, wm: %d\n",
2240 entries = intel_wm_method2(clock, htotal,
2241 crtc->base.cursor->state->crtc_w, 4,
2242 sr_latency_ns / 100);
2243 entries = DIV_ROUND_UP(entries,
2244 i965_cursor_wm_info.cacheline_size) +
2245 i965_cursor_wm_info.guard_size;
2247 cursor_sr = i965_cursor_wm_info.fifo_size - entries;
2248 if (cursor_sr > i965_cursor_wm_info.max_wm)
2249 cursor_sr = i965_cursor_wm_info.max_wm;
2251 drm_dbg_kms(&dev_priv->drm,
2252 "self-refresh watermark: display plane %d "
2253 "cursor %d\n", srwm, cursor_sr);
2255 cxsr_enabled = true;
2257 cxsr_enabled = false;
2258 /* Turn off self refresh if both pipes are enabled */
2259 intel_set_memory_cxsr(dev_priv, false);
2262 drm_dbg_kms(&dev_priv->drm,
2263 "Setting FIFO watermarks - A: 8, B: 8, C: 8, SR %d\n",
2266 /* 965 has limitations... */
2267 I915_WRITE(DSPFW1, FW_WM(srwm, SR) |
2271 I915_WRITE(DSPFW2, FW_WM(8, CURSORA) |
2272 FW_WM(8, PLANEC_OLD));
2273 /* update cursor SR watermark */
2274 I915_WRITE(DSPFW3, FW_WM(cursor_sr, CURSOR_SR));
2277 intel_set_memory_cxsr(dev_priv, true);
2282 static void i9xx_update_wm(struct intel_crtc *unused_crtc)
2284 struct drm_i915_private *dev_priv = to_i915(unused_crtc->base.dev);
2285 const struct intel_watermark_params *wm_info;
2290 int planea_wm, planeb_wm;
2291 struct intel_crtc *crtc, *enabled = NULL;
2293 if (IS_I945GM(dev_priv))
2294 wm_info = &i945_wm_info;
2295 else if (!IS_GEN(dev_priv, 2))
2296 wm_info = &i915_wm_info;
2298 wm_info = &i830_a_wm_info;
2300 fifo_size = dev_priv->display.get_fifo_size(dev_priv, PLANE_A);
2301 crtc = intel_get_crtc_for_plane(dev_priv, PLANE_A);
2302 if (intel_crtc_active(crtc)) {
2303 const struct drm_display_mode *adjusted_mode =
2304 &crtc->config->hw.adjusted_mode;
2305 const struct drm_framebuffer *fb =
2306 crtc->base.primary->state->fb;
2309 if (IS_GEN(dev_priv, 2))
2312 cpp = fb->format->cpp[0];
2314 planea_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
2315 wm_info, fifo_size, cpp,
2316 pessimal_latency_ns);
2319 planea_wm = fifo_size - wm_info->guard_size;
2320 if (planea_wm > (long)wm_info->max_wm)
2321 planea_wm = wm_info->max_wm;
2324 if (IS_GEN(dev_priv, 2))
2325 wm_info = &i830_bc_wm_info;
2327 fifo_size = dev_priv->display.get_fifo_size(dev_priv, PLANE_B);
2328 crtc = intel_get_crtc_for_plane(dev_priv, PLANE_B);
2329 if (intel_crtc_active(crtc)) {
2330 const struct drm_display_mode *adjusted_mode =
2331 &crtc->config->hw.adjusted_mode;
2332 const struct drm_framebuffer *fb =
2333 crtc->base.primary->state->fb;
2336 if (IS_GEN(dev_priv, 2))
2339 cpp = fb->format->cpp[0];
2341 planeb_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
2342 wm_info, fifo_size, cpp,
2343 pessimal_latency_ns);
2344 if (enabled == NULL)
2349 planeb_wm = fifo_size - wm_info->guard_size;
2350 if (planeb_wm > (long)wm_info->max_wm)
2351 planeb_wm = wm_info->max_wm;
2354 drm_dbg_kms(&dev_priv->drm,
2355 "FIFO watermarks - A: %d, B: %d\n", planea_wm, planeb_wm);
2357 if (IS_I915GM(dev_priv) && enabled) {
2358 struct drm_i915_gem_object *obj;
2360 obj = intel_fb_obj(enabled->base.primary->state->fb);
2362 /* self-refresh seems busted with untiled */
2363 if (!i915_gem_object_is_tiled(obj))
2368 * Overlay gets an aggressive default since video jitter is bad.
2372 /* Play safe and disable self-refresh before adjusting watermarks. */
2373 intel_set_memory_cxsr(dev_priv, false);
2375 /* Calc sr entries for one plane configs */
2376 if (HAS_FW_BLC(dev_priv) && enabled) {
2377 /* self-refresh has much higher latency */
2378 static const int sr_latency_ns = 6000;
2379 const struct drm_display_mode *adjusted_mode =
2380 &enabled->config->hw.adjusted_mode;
2381 const struct drm_framebuffer *fb =
2382 enabled->base.primary->state->fb;
2383 int clock = adjusted_mode->crtc_clock;
2384 int htotal = adjusted_mode->crtc_htotal;
2385 int hdisplay = enabled->config->pipe_src_w;
2389 if (IS_I915GM(dev_priv) || IS_I945GM(dev_priv))
2392 cpp = fb->format->cpp[0];
2394 entries = intel_wm_method2(clock, htotal, hdisplay, cpp,
2395 sr_latency_ns / 100);
2396 entries = DIV_ROUND_UP(entries, wm_info->cacheline_size);
2397 drm_dbg_kms(&dev_priv->drm,
2398 "self-refresh entries: %d\n", entries);
2399 srwm = wm_info->fifo_size - entries;
2403 if (IS_I945G(dev_priv) || IS_I945GM(dev_priv))
2404 I915_WRITE(FW_BLC_SELF,
2405 FW_BLC_SELF_FIFO_MASK | (srwm & 0xff));
2407 I915_WRITE(FW_BLC_SELF, srwm & 0x3f);
2410 drm_dbg_kms(&dev_priv->drm,
2411 "Setting FIFO watermarks - A: %d, B: %d, C: %d, SR %d\n",
2412 planea_wm, planeb_wm, cwm, srwm);
2414 fwater_lo = ((planeb_wm & 0x3f) << 16) | (planea_wm & 0x3f);
2415 fwater_hi = (cwm & 0x1f);
2417 /* Set request length to 8 cachelines per fetch */
2418 fwater_lo = fwater_lo | (1 << 24) | (1 << 8);
2419 fwater_hi = fwater_hi | (1 << 8);
2421 I915_WRITE(FW_BLC, fwater_lo);
2422 I915_WRITE(FW_BLC2, fwater_hi);
2425 intel_set_memory_cxsr(dev_priv, true);
2428 static void i845_update_wm(struct intel_crtc *unused_crtc)
2430 struct drm_i915_private *dev_priv = to_i915(unused_crtc->base.dev);
2431 struct intel_crtc *crtc;
2432 const struct drm_display_mode *adjusted_mode;
2436 crtc = single_enabled_crtc(dev_priv);
2440 adjusted_mode = &crtc->config->hw.adjusted_mode;
2441 planea_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
2443 dev_priv->display.get_fifo_size(dev_priv, PLANE_A),
2444 4, pessimal_latency_ns);
2445 fwater_lo = I915_READ(FW_BLC) & ~0xfff;
2446 fwater_lo |= (3<<8) | planea_wm;
2448 drm_dbg_kms(&dev_priv->drm,
2449 "Setting FIFO watermarks - A: %d\n", planea_wm);
2451 I915_WRITE(FW_BLC, fwater_lo);
2454 /* latency must be in 0.1us units. */
2455 static unsigned int ilk_wm_method1(unsigned int pixel_rate,
2457 unsigned int latency)
2461 ret = intel_wm_method1(pixel_rate, cpp, latency);
2462 ret = DIV_ROUND_UP(ret, 64) + 2;
2467 /* latency must be in 0.1us units. */
2468 static unsigned int ilk_wm_method2(unsigned int pixel_rate,
2469 unsigned int htotal,
2472 unsigned int latency)
2476 ret = intel_wm_method2(pixel_rate, htotal,
2477 width, cpp, latency);
2478 ret = DIV_ROUND_UP(ret, 64) + 2;
2483 static u32 ilk_wm_fbc(u32 pri_val, u32 horiz_pixels, u8 cpp)
2486 * Neither of these should be possible since this function shouldn't be
2487 * called if the CRTC is off or the plane is invisible. But let's be
2488 * extra paranoid to avoid a potential divide-by-zero if we screw up
2489 * elsewhere in the driver.
2493 if (WARN_ON(!horiz_pixels))
2496 return DIV_ROUND_UP(pri_val * 64, horiz_pixels * cpp) + 2;
2499 struct ilk_wm_maximums {
2507 * For both WM_PIPE and WM_LP.
2508 * mem_value must be in 0.1us units.
2510 static u32 ilk_compute_pri_wm(const struct intel_crtc_state *crtc_state,
2511 const struct intel_plane_state *plane_state,
2512 u32 mem_value, bool is_lp)
2514 u32 method1, method2;
2520 if (!intel_wm_plane_visible(crtc_state, plane_state))
2523 cpp = plane_state->hw.fb->format->cpp[0];
2525 method1 = ilk_wm_method1(crtc_state->pixel_rate, cpp, mem_value);
2530 method2 = ilk_wm_method2(crtc_state->pixel_rate,
2531 crtc_state->hw.adjusted_mode.crtc_htotal,
2532 drm_rect_width(&plane_state->uapi.dst),
2535 return min(method1, method2);
2539 * For both WM_PIPE and WM_LP.
2540 * mem_value must be in 0.1us units.
2542 static u32 ilk_compute_spr_wm(const struct intel_crtc_state *crtc_state,
2543 const struct intel_plane_state *plane_state,
2546 u32 method1, method2;
2552 if (!intel_wm_plane_visible(crtc_state, plane_state))
2555 cpp = plane_state->hw.fb->format->cpp[0];
2557 method1 = ilk_wm_method1(crtc_state->pixel_rate, cpp, mem_value);
2558 method2 = ilk_wm_method2(crtc_state->pixel_rate,
2559 crtc_state->hw.adjusted_mode.crtc_htotal,
2560 drm_rect_width(&plane_state->uapi.dst),
2562 return min(method1, method2);
2566 * For both WM_PIPE and WM_LP.
2567 * mem_value must be in 0.1us units.
2569 static u32 ilk_compute_cur_wm(const struct intel_crtc_state *crtc_state,
2570 const struct intel_plane_state *plane_state,
2578 if (!intel_wm_plane_visible(crtc_state, plane_state))
2581 cpp = plane_state->hw.fb->format->cpp[0];
2583 return ilk_wm_method2(crtc_state->pixel_rate,
2584 crtc_state->hw.adjusted_mode.crtc_htotal,
2585 drm_rect_width(&plane_state->uapi.dst),
2589 /* Only for WM_LP. */
2590 static u32 ilk_compute_fbc_wm(const struct intel_crtc_state *crtc_state,
2591 const struct intel_plane_state *plane_state,
2596 if (!intel_wm_plane_visible(crtc_state, plane_state))
2599 cpp = plane_state->hw.fb->format->cpp[0];
2601 return ilk_wm_fbc(pri_val, drm_rect_width(&plane_state->uapi.dst),
2606 ilk_display_fifo_size(const struct drm_i915_private *dev_priv)
2608 if (INTEL_GEN(dev_priv) >= 8)
2610 else if (INTEL_GEN(dev_priv) >= 7)
2617 ilk_plane_wm_reg_max(const struct drm_i915_private *dev_priv,
2618 int level, bool is_sprite)
2620 if (INTEL_GEN(dev_priv) >= 8)
2621 /* BDW primary/sprite plane watermarks */
2622 return level == 0 ? 255 : 2047;
2623 else if (INTEL_GEN(dev_priv) >= 7)
2624 /* IVB/HSW primary/sprite plane watermarks */
2625 return level == 0 ? 127 : 1023;
2626 else if (!is_sprite)
2627 /* ILK/SNB primary plane watermarks */
2628 return level == 0 ? 127 : 511;
2630 /* ILK/SNB sprite plane watermarks */
2631 return level == 0 ? 63 : 255;
2635 ilk_cursor_wm_reg_max(const struct drm_i915_private *dev_priv, int level)
2637 if (INTEL_GEN(dev_priv) >= 7)
2638 return level == 0 ? 63 : 255;
2640 return level == 0 ? 31 : 63;
2643 static unsigned int ilk_fbc_wm_reg_max(const struct drm_i915_private *dev_priv)
2645 if (INTEL_GEN(dev_priv) >= 8)
2651 /* Calculate the maximum primary/sprite plane watermark */
2652 static unsigned int ilk_plane_wm_max(const struct drm_i915_private *dev_priv,
2654 const struct intel_wm_config *config,
2655 enum intel_ddb_partitioning ddb_partitioning,
2658 unsigned int fifo_size = ilk_display_fifo_size(dev_priv);
2660 /* if sprites aren't enabled, sprites get nothing */
2661 if (is_sprite && !config->sprites_enabled)
2664 /* HSW allows LP1+ watermarks even with multiple pipes */
2665 if (level == 0 || config->num_pipes_active > 1) {
2666 fifo_size /= INTEL_NUM_PIPES(dev_priv);
2669 * For some reason the non self refresh
2670 * FIFO size is only half of the self
2671 * refresh FIFO size on ILK/SNB.
2673 if (INTEL_GEN(dev_priv) <= 6)
2677 if (config->sprites_enabled) {
2678 /* level 0 is always calculated with 1:1 split */
2679 if (level > 0 && ddb_partitioning == INTEL_DDB_PART_5_6) {
2688 /* clamp to max that the registers can hold */
2689 return min(fifo_size, ilk_plane_wm_reg_max(dev_priv, level, is_sprite));
2692 /* Calculate the maximum cursor plane watermark */
2693 static unsigned int ilk_cursor_wm_max(const struct drm_i915_private *dev_priv,
2695 const struct intel_wm_config *config)
2697 /* HSW LP1+ watermarks w/ multiple pipes */
2698 if (level > 0 && config->num_pipes_active > 1)
2701 /* otherwise just report max that registers can hold */
2702 return ilk_cursor_wm_reg_max(dev_priv, level);
2705 static void ilk_compute_wm_maximums(const struct drm_i915_private *dev_priv,
2707 const struct intel_wm_config *config,
2708 enum intel_ddb_partitioning ddb_partitioning,
2709 struct ilk_wm_maximums *max)
2711 max->pri = ilk_plane_wm_max(dev_priv, level, config, ddb_partitioning, false);
2712 max->spr = ilk_plane_wm_max(dev_priv, level, config, ddb_partitioning, true);
2713 max->cur = ilk_cursor_wm_max(dev_priv, level, config);
2714 max->fbc = ilk_fbc_wm_reg_max(dev_priv);
2717 static void ilk_compute_wm_reg_maximums(const struct drm_i915_private *dev_priv,
2719 struct ilk_wm_maximums *max)
2721 max->pri = ilk_plane_wm_reg_max(dev_priv, level, false);
2722 max->spr = ilk_plane_wm_reg_max(dev_priv, level, true);
2723 max->cur = ilk_cursor_wm_reg_max(dev_priv, level);
2724 max->fbc = ilk_fbc_wm_reg_max(dev_priv);
2727 static bool ilk_validate_wm_level(int level,
2728 const struct ilk_wm_maximums *max,
2729 struct intel_wm_level *result)
2733 /* already determined to be invalid? */
2734 if (!result->enable)
2737 result->enable = result->pri_val <= max->pri &&
2738 result->spr_val <= max->spr &&
2739 result->cur_val <= max->cur;
2741 ret = result->enable;
2744 * HACK until we can pre-compute everything,
2745 * and thus fail gracefully if LP0 watermarks
2748 if (level == 0 && !result->enable) {
2749 if (result->pri_val > max->pri)
2750 DRM_DEBUG_KMS("Primary WM%d too large %u (max %u)\n",
2751 level, result->pri_val, max->pri);
2752 if (result->spr_val > max->spr)
2753 DRM_DEBUG_KMS("Sprite WM%d too large %u (max %u)\n",
2754 level, result->spr_val, max->spr);
2755 if (result->cur_val > max->cur)
2756 DRM_DEBUG_KMS("Cursor WM%d too large %u (max %u)\n",
2757 level, result->cur_val, max->cur);
2759 result->pri_val = min_t(u32, result->pri_val, max->pri);
2760 result->spr_val = min_t(u32, result->spr_val, max->spr);
2761 result->cur_val = min_t(u32, result->cur_val, max->cur);
2762 result->enable = true;
2768 static void ilk_compute_wm_level(const struct drm_i915_private *dev_priv,
2769 const struct intel_crtc *crtc,
2771 struct intel_crtc_state *crtc_state,
2772 const struct intel_plane_state *pristate,
2773 const struct intel_plane_state *sprstate,
2774 const struct intel_plane_state *curstate,
2775 struct intel_wm_level *result)
2777 u16 pri_latency = dev_priv->wm.pri_latency[level];
2778 u16 spr_latency = dev_priv->wm.spr_latency[level];
2779 u16 cur_latency = dev_priv->wm.cur_latency[level];
2781 /* WM1+ latency values stored in 0.5us units */
2789 result->pri_val = ilk_compute_pri_wm(crtc_state, pristate,
2790 pri_latency, level);
2791 result->fbc_val = ilk_compute_fbc_wm(crtc_state, pristate, result->pri_val);
2795 result->spr_val = ilk_compute_spr_wm(crtc_state, sprstate, spr_latency);
2798 result->cur_val = ilk_compute_cur_wm(crtc_state, curstate, cur_latency);
2800 result->enable = true;
2803 static void intel_read_wm_latency(struct drm_i915_private *dev_priv,
2806 struct intel_uncore *uncore = &dev_priv->uncore;
2808 if (INTEL_GEN(dev_priv) >= 9) {
2811 int level, max_level = ilk_wm_max_level(dev_priv);
2813 /* read the first set of memory latencies[0:3] */
2814 val = 0; /* data0 to be programmed to 0 for first set */
2815 ret = sandybridge_pcode_read(dev_priv,
2816 GEN9_PCODE_READ_MEM_LATENCY,
2820 drm_err(&dev_priv->drm,
2821 "SKL Mailbox read error = %d\n", ret);
2825 wm[0] = val & GEN9_MEM_LATENCY_LEVEL_MASK;
2826 wm[1] = (val >> GEN9_MEM_LATENCY_LEVEL_1_5_SHIFT) &
2827 GEN9_MEM_LATENCY_LEVEL_MASK;
2828 wm[2] = (val >> GEN9_MEM_LATENCY_LEVEL_2_6_SHIFT) &
2829 GEN9_MEM_LATENCY_LEVEL_MASK;
2830 wm[3] = (val >> GEN9_MEM_LATENCY_LEVEL_3_7_SHIFT) &
2831 GEN9_MEM_LATENCY_LEVEL_MASK;
2833 /* read the second set of memory latencies[4:7] */
2834 val = 1; /* data0 to be programmed to 1 for second set */
2835 ret = sandybridge_pcode_read(dev_priv,
2836 GEN9_PCODE_READ_MEM_LATENCY,
2839 drm_err(&dev_priv->drm,
2840 "SKL Mailbox read error = %d\n", ret);
2844 wm[4] = val & GEN9_MEM_LATENCY_LEVEL_MASK;
2845 wm[5] = (val >> GEN9_MEM_LATENCY_LEVEL_1_5_SHIFT) &
2846 GEN9_MEM_LATENCY_LEVEL_MASK;
2847 wm[6] = (val >> GEN9_MEM_LATENCY_LEVEL_2_6_SHIFT) &
2848 GEN9_MEM_LATENCY_LEVEL_MASK;
2849 wm[7] = (val >> GEN9_MEM_LATENCY_LEVEL_3_7_SHIFT) &
2850 GEN9_MEM_LATENCY_LEVEL_MASK;
2853 * If a level n (n > 1) has a 0us latency, all levels m (m >= n)
2854 * need to be disabled. We make sure to sanitize the values out
2855 * of the punit to satisfy this requirement.
2857 for (level = 1; level <= max_level; level++) {
2858 if (wm[level] == 0) {
2859 for (i = level + 1; i <= max_level; i++)
2866 * WaWmMemoryReadLatency:skl+,glk
2868 * punit doesn't take into account the read latency so we need
2869 * to add 2us to the various latency levels we retrieve from the
2870 * punit when level 0 response data us 0us.
2874 for (level = 1; level <= max_level; level++) {
2882 * WA Level-0 adjustment for 16GB DIMMs: SKL+
2883 * If we could not get dimm info enable this WA to prevent from
2884 * any underrun. If not able to get Dimm info assume 16GB dimm
2885 * to avoid any underrun.
2887 if (dev_priv->dram_info.is_16gb_dimm)
2890 } else if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv)) {
2891 u64 sskpd = intel_uncore_read64(uncore, MCH_SSKPD);
2893 wm[0] = (sskpd >> 56) & 0xFF;
2895 wm[0] = sskpd & 0xF;
2896 wm[1] = (sskpd >> 4) & 0xFF;
2897 wm[2] = (sskpd >> 12) & 0xFF;
2898 wm[3] = (sskpd >> 20) & 0x1FF;
2899 wm[4] = (sskpd >> 32) & 0x1FF;
2900 } else if (INTEL_GEN(dev_priv) >= 6) {
2901 u32 sskpd = intel_uncore_read(uncore, MCH_SSKPD);
2903 wm[0] = (sskpd >> SSKPD_WM0_SHIFT) & SSKPD_WM_MASK;
2904 wm[1] = (sskpd >> SSKPD_WM1_SHIFT) & SSKPD_WM_MASK;
2905 wm[2] = (sskpd >> SSKPD_WM2_SHIFT) & SSKPD_WM_MASK;
2906 wm[3] = (sskpd >> SSKPD_WM3_SHIFT) & SSKPD_WM_MASK;
2907 } else if (INTEL_GEN(dev_priv) >= 5) {
2908 u32 mltr = intel_uncore_read(uncore, MLTR_ILK);
2910 /* ILK primary LP0 latency is 700 ns */
2912 wm[1] = (mltr >> MLTR_WM1_SHIFT) & ILK_SRLT_MASK;
2913 wm[2] = (mltr >> MLTR_WM2_SHIFT) & ILK_SRLT_MASK;
2915 MISSING_CASE(INTEL_DEVID(dev_priv));
2919 static void intel_fixup_spr_wm_latency(struct drm_i915_private *dev_priv,
2922 /* ILK sprite LP0 latency is 1300 ns */
2923 if (IS_GEN(dev_priv, 5))
2927 static void intel_fixup_cur_wm_latency(struct drm_i915_private *dev_priv,
2930 /* ILK cursor LP0 latency is 1300 ns */
2931 if (IS_GEN(dev_priv, 5))
2935 int ilk_wm_max_level(const struct drm_i915_private *dev_priv)
2937 /* how many WM levels are we expecting */
2938 if (INTEL_GEN(dev_priv) >= 9)
2940 else if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
2942 else if (INTEL_GEN(dev_priv) >= 6)
2948 static void intel_print_wm_latency(struct drm_i915_private *dev_priv,
2952 int level, max_level = ilk_wm_max_level(dev_priv);
2954 for (level = 0; level <= max_level; level++) {
2955 unsigned int latency = wm[level];
2958 drm_dbg_kms(&dev_priv->drm,
2959 "%s WM%d latency not provided\n",
2965 * - latencies are in us on gen9.
2966 * - before then, WM1+ latency values are in 0.5us units
2968 if (INTEL_GEN(dev_priv) >= 9)
2973 drm_dbg_kms(&dev_priv->drm,
2974 "%s WM%d latency %u (%u.%u usec)\n", name, level,
2975 wm[level], latency / 10, latency % 10);
2979 static bool ilk_increase_wm_latency(struct drm_i915_private *dev_priv,
2982 int level, max_level = ilk_wm_max_level(dev_priv);
2987 wm[0] = max(wm[0], min);
2988 for (level = 1; level <= max_level; level++)
2989 wm[level] = max_t(u16, wm[level], DIV_ROUND_UP(min, 5));
2994 static void snb_wm_latency_quirk(struct drm_i915_private *dev_priv)
2999 * The BIOS provided WM memory latency values are often
3000 * inadequate for high resolution displays. Adjust them.
3002 changed = ilk_increase_wm_latency(dev_priv, dev_priv->wm.pri_latency, 12) |
3003 ilk_increase_wm_latency(dev_priv, dev_priv->wm.spr_latency, 12) |
3004 ilk_increase_wm_latency(dev_priv, dev_priv->wm.cur_latency, 12);
3009 drm_dbg_kms(&dev_priv->drm,
3010 "WM latency values increased to avoid potential underruns\n");
3011 intel_print_wm_latency(dev_priv, "Primary", dev_priv->wm.pri_latency);
3012 intel_print_wm_latency(dev_priv, "Sprite", dev_priv->wm.spr_latency);
3013 intel_print_wm_latency(dev_priv, "Cursor", dev_priv->wm.cur_latency);
3016 static void snb_wm_lp3_irq_quirk(struct drm_i915_private *dev_priv)
3019 * On some SNB machines (Thinkpad X220 Tablet at least)
3020 * LP3 usage can cause vblank interrupts to be lost.
3021 * The DEIIR bit will go high but it looks like the CPU
3022 * never gets interrupted.
3024 * It's not clear whether other interrupt source could
3025 * be affected or if this is somehow limited to vblank
3026 * interrupts only. To play it safe we disable LP3
3027 * watermarks entirely.
3029 if (dev_priv->wm.pri_latency[3] == 0 &&
3030 dev_priv->wm.spr_latency[3] == 0 &&
3031 dev_priv->wm.cur_latency[3] == 0)
3034 dev_priv->wm.pri_latency[3] = 0;
3035 dev_priv->wm.spr_latency[3] = 0;
3036 dev_priv->wm.cur_latency[3] = 0;
3038 drm_dbg_kms(&dev_priv->drm,
3039 "LP3 watermarks disabled due to potential for lost interrupts\n");
3040 intel_print_wm_latency(dev_priv, "Primary", dev_priv->wm.pri_latency);
3041 intel_print_wm_latency(dev_priv, "Sprite", dev_priv->wm.spr_latency);
3042 intel_print_wm_latency(dev_priv, "Cursor", dev_priv->wm.cur_latency);
3045 static void ilk_setup_wm_latency(struct drm_i915_private *dev_priv)
3047 intel_read_wm_latency(dev_priv, dev_priv->wm.pri_latency);
3049 memcpy(dev_priv->wm.spr_latency, dev_priv->wm.pri_latency,
3050 sizeof(dev_priv->wm.pri_latency));
3051 memcpy(dev_priv->wm.cur_latency, dev_priv->wm.pri_latency,
3052 sizeof(dev_priv->wm.pri_latency));
3054 intel_fixup_spr_wm_latency(dev_priv, dev_priv->wm.spr_latency);
3055 intel_fixup_cur_wm_latency(dev_priv, dev_priv->wm.cur_latency);
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);
3061 if (IS_GEN(dev_priv, 6)) {
3062 snb_wm_latency_quirk(dev_priv);
3063 snb_wm_lp3_irq_quirk(dev_priv);
3067 static void skl_setup_wm_latency(struct drm_i915_private *dev_priv)
3069 intel_read_wm_latency(dev_priv, dev_priv->wm.skl_latency);
3070 intel_print_wm_latency(dev_priv, "Gen9 Plane", dev_priv->wm.skl_latency);
3073 static bool ilk_validate_pipe_wm(const struct drm_i915_private *dev_priv,
3074 struct intel_pipe_wm *pipe_wm)
3076 /* LP0 watermark maximums depend on this pipe alone */
3077 const struct intel_wm_config config = {
3078 .num_pipes_active = 1,
3079 .sprites_enabled = pipe_wm->sprites_enabled,
3080 .sprites_scaled = pipe_wm->sprites_scaled,
3082 struct ilk_wm_maximums max;
3084 /* LP0 watermarks always use 1/2 DDB partitioning */
3085 ilk_compute_wm_maximums(dev_priv, 0, &config, INTEL_DDB_PART_1_2, &max);
3087 /* At least LP0 must be valid */
3088 if (!ilk_validate_wm_level(0, &max, &pipe_wm->wm[0])) {
3089 drm_dbg_kms(&dev_priv->drm, "LP0 watermark invalid\n");
3096 /* Compute new watermarks for the pipe */
3097 static int ilk_compute_pipe_wm(struct intel_crtc_state *crtc_state)
3099 struct drm_i915_private *dev_priv = to_i915(crtc_state->uapi.crtc->dev);
3100 struct intel_crtc *crtc = to_intel_crtc(crtc_state->uapi.crtc);
3101 struct intel_pipe_wm *pipe_wm;
3102 struct intel_plane *plane;
3103 const struct intel_plane_state *plane_state;
3104 const struct intel_plane_state *pristate = NULL;
3105 const struct intel_plane_state *sprstate = NULL;
3106 const struct intel_plane_state *curstate = NULL;
3107 int level, max_level = ilk_wm_max_level(dev_priv), usable_level;
3108 struct ilk_wm_maximums max;
3110 pipe_wm = &crtc_state->wm.ilk.optimal;
3112 intel_atomic_crtc_state_for_each_plane_state(plane, plane_state, crtc_state) {
3113 if (plane->base.type == DRM_PLANE_TYPE_PRIMARY)
3114 pristate = plane_state;
3115 else if (plane->base.type == DRM_PLANE_TYPE_OVERLAY)
3116 sprstate = plane_state;
3117 else if (plane->base.type == DRM_PLANE_TYPE_CURSOR)
3118 curstate = plane_state;
3121 pipe_wm->pipe_enabled = crtc_state->hw.active;
3123 pipe_wm->sprites_enabled = sprstate->uapi.visible;
3124 pipe_wm->sprites_scaled = sprstate->uapi.visible &&
3125 (drm_rect_width(&sprstate->uapi.dst) != drm_rect_width(&sprstate->uapi.src) >> 16 ||
3126 drm_rect_height(&sprstate->uapi.dst) != drm_rect_height(&sprstate->uapi.src) >> 16);
3129 usable_level = max_level;
3131 /* ILK/SNB: LP2+ watermarks only w/o sprites */
3132 if (INTEL_GEN(dev_priv) <= 6 && pipe_wm->sprites_enabled)
3135 /* ILK/SNB/IVB: LP1+ watermarks only w/o scaling */
3136 if (pipe_wm->sprites_scaled)
3139 memset(&pipe_wm->wm, 0, sizeof(pipe_wm->wm));
3140 ilk_compute_wm_level(dev_priv, crtc, 0, crtc_state,
3141 pristate, sprstate, curstate, &pipe_wm->wm[0]);
3143 if (!ilk_validate_pipe_wm(dev_priv, pipe_wm))
3146 ilk_compute_wm_reg_maximums(dev_priv, 1, &max);
3148 for (level = 1; level <= usable_level; level++) {
3149 struct intel_wm_level *wm = &pipe_wm->wm[level];
3151 ilk_compute_wm_level(dev_priv, crtc, level, crtc_state,
3152 pristate, sprstate, curstate, wm);
3155 * Disable any watermark level that exceeds the
3156 * register maximums since such watermarks are
3159 if (!ilk_validate_wm_level(level, &max, wm)) {
3160 memset(wm, 0, sizeof(*wm));
3169 * Build a set of 'intermediate' watermark values that satisfy both the old
3170 * state and the new state. These can be programmed to the hardware
3173 static int ilk_compute_intermediate_wm(struct intel_crtc_state *newstate)
3175 struct intel_crtc *intel_crtc = to_intel_crtc(newstate->uapi.crtc);
3176 struct drm_i915_private *dev_priv = to_i915(intel_crtc->base.dev);
3177 struct intel_pipe_wm *a = &newstate->wm.ilk.intermediate;
3178 struct intel_atomic_state *intel_state =
3179 to_intel_atomic_state(newstate->uapi.state);
3180 const struct intel_crtc_state *oldstate =
3181 intel_atomic_get_old_crtc_state(intel_state, intel_crtc);
3182 const struct intel_pipe_wm *b = &oldstate->wm.ilk.optimal;
3183 int level, max_level = ilk_wm_max_level(dev_priv);
3186 * Start with the final, target watermarks, then combine with the
3187 * currently active watermarks to get values that are safe both before
3188 * and after the vblank.
3190 *a = newstate->wm.ilk.optimal;
3191 if (!newstate->hw.active || drm_atomic_crtc_needs_modeset(&newstate->uapi) ||
3192 intel_state->skip_intermediate_wm)
3195 a->pipe_enabled |= b->pipe_enabled;
3196 a->sprites_enabled |= b->sprites_enabled;
3197 a->sprites_scaled |= b->sprites_scaled;
3199 for (level = 0; level <= max_level; level++) {
3200 struct intel_wm_level *a_wm = &a->wm[level];
3201 const struct intel_wm_level *b_wm = &b->wm[level];
3203 a_wm->enable &= b_wm->enable;
3204 a_wm->pri_val = max(a_wm->pri_val, b_wm->pri_val);
3205 a_wm->spr_val = max(a_wm->spr_val, b_wm->spr_val);
3206 a_wm->cur_val = max(a_wm->cur_val, b_wm->cur_val);
3207 a_wm->fbc_val = max(a_wm->fbc_val, b_wm->fbc_val);
3211 * We need to make sure that these merged watermark values are
3212 * actually a valid configuration themselves. If they're not,
3213 * there's no safe way to transition from the old state to
3214 * the new state, so we need to fail the atomic transaction.
3216 if (!ilk_validate_pipe_wm(dev_priv, a))
3220 * If our intermediate WM are identical to the final WM, then we can
3221 * omit the post-vblank programming; only update if it's different.
3223 if (memcmp(a, &newstate->wm.ilk.optimal, sizeof(*a)) != 0)
3224 newstate->wm.need_postvbl_update = true;
3230 * Merge the watermarks from all active pipes for a specific level.
3232 static void ilk_merge_wm_level(struct drm_i915_private *dev_priv,
3234 struct intel_wm_level *ret_wm)
3236 const struct intel_crtc *intel_crtc;
3238 ret_wm->enable = true;
3240 for_each_intel_crtc(&dev_priv->drm, intel_crtc) {
3241 const struct intel_pipe_wm *active = &intel_crtc->wm.active.ilk;
3242 const struct intel_wm_level *wm = &active->wm[level];
3244 if (!active->pipe_enabled)
3248 * The watermark values may have been used in the past,
3249 * so we must maintain them in the registers for some
3250 * time even if the level is now disabled.
3253 ret_wm->enable = false;
3255 ret_wm->pri_val = max(ret_wm->pri_val, wm->pri_val);
3256 ret_wm->spr_val = max(ret_wm->spr_val, wm->spr_val);
3257 ret_wm->cur_val = max(ret_wm->cur_val, wm->cur_val);
3258 ret_wm->fbc_val = max(ret_wm->fbc_val, wm->fbc_val);
3263 * Merge all low power watermarks for all active pipes.
3265 static void ilk_wm_merge(struct drm_i915_private *dev_priv,
3266 const struct intel_wm_config *config,
3267 const struct ilk_wm_maximums *max,
3268 struct intel_pipe_wm *merged)
3270 int level, max_level = ilk_wm_max_level(dev_priv);
3271 int last_enabled_level = max_level;
3273 /* ILK/SNB/IVB: LP1+ watermarks only w/ single pipe */
3274 if ((INTEL_GEN(dev_priv) <= 6 || IS_IVYBRIDGE(dev_priv)) &&
3275 config->num_pipes_active > 1)
3276 last_enabled_level = 0;
3278 /* ILK: FBC WM must be disabled always */
3279 merged->fbc_wm_enabled = INTEL_GEN(dev_priv) >= 6;
3281 /* merge each WM1+ level */
3282 for (level = 1; level <= max_level; level++) {
3283 struct intel_wm_level *wm = &merged->wm[level];
3285 ilk_merge_wm_level(dev_priv, level, wm);
3287 if (level > last_enabled_level)
3289 else if (!ilk_validate_wm_level(level, max, wm))
3290 /* make sure all following levels get disabled */
3291 last_enabled_level = level - 1;
3294 * The spec says it is preferred to disable
3295 * FBC WMs instead of disabling a WM level.
3297 if (wm->fbc_val > max->fbc) {
3299 merged->fbc_wm_enabled = false;
3304 /* ILK: LP2+ must be disabled when FBC WM is disabled but FBC enabled */
3306 * FIXME this is racy. FBC might get enabled later.
3307 * What we should check here is whether FBC can be
3308 * enabled sometime later.
3310 if (IS_GEN(dev_priv, 5) && !merged->fbc_wm_enabled &&
3311 intel_fbc_is_active(dev_priv)) {
3312 for (level = 2; level <= max_level; level++) {
3313 struct intel_wm_level *wm = &merged->wm[level];
3320 static int ilk_wm_lp_to_level(int wm_lp, const struct intel_pipe_wm *pipe_wm)
3322 /* LP1,LP2,LP3 levels are either 1,2,3 or 1,3,4 */
3323 return wm_lp + (wm_lp >= 2 && pipe_wm->wm[4].enable);
3326 /* The value we need to program into the WM_LPx latency field */
3327 static unsigned int ilk_wm_lp_latency(struct drm_i915_private *dev_priv,
3330 if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
3333 return dev_priv->wm.pri_latency[level];
3336 static void ilk_compute_wm_results(struct drm_i915_private *dev_priv,
3337 const struct intel_pipe_wm *merged,
3338 enum intel_ddb_partitioning partitioning,
3339 struct ilk_wm_values *results)
3341 struct intel_crtc *intel_crtc;
3344 results->enable_fbc_wm = merged->fbc_wm_enabled;
3345 results->partitioning = partitioning;
3347 /* LP1+ register values */
3348 for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
3349 const struct intel_wm_level *r;
3351 level = ilk_wm_lp_to_level(wm_lp, merged);
3353 r = &merged->wm[level];
3356 * Maintain the watermark values even if the level is
3357 * disabled. Doing otherwise could cause underruns.
3359 results->wm_lp[wm_lp - 1] =
3360 (ilk_wm_lp_latency(dev_priv, level) << WM1_LP_LATENCY_SHIFT) |
3361 (r->pri_val << WM1_LP_SR_SHIFT) |
3365 results->wm_lp[wm_lp - 1] |= WM1_LP_SR_EN;
3367 if (INTEL_GEN(dev_priv) >= 8)
3368 results->wm_lp[wm_lp - 1] |=
3369 r->fbc_val << WM1_LP_FBC_SHIFT_BDW;
3371 results->wm_lp[wm_lp - 1] |=
3372 r->fbc_val << WM1_LP_FBC_SHIFT;
3375 * Always set WM1S_LP_EN when spr_val != 0, even if the
3376 * level is disabled. Doing otherwise could cause underruns.
3378 if (INTEL_GEN(dev_priv) <= 6 && r->spr_val) {
3379 drm_WARN_ON(&dev_priv->drm, wm_lp != 1);
3380 results->wm_lp_spr[wm_lp - 1] = WM1S_LP_EN | r->spr_val;
3382 results->wm_lp_spr[wm_lp - 1] = r->spr_val;
3385 /* LP0 register values */
3386 for_each_intel_crtc(&dev_priv->drm, intel_crtc) {
3387 enum pipe pipe = intel_crtc->pipe;
3388 const struct intel_pipe_wm *pipe_wm = &intel_crtc->wm.active.ilk;
3389 const struct intel_wm_level *r = &pipe_wm->wm[0];
3391 if (drm_WARN_ON(&dev_priv->drm, !r->enable))
3394 results->wm_pipe[pipe] =
3395 (r->pri_val << WM0_PIPE_PLANE_SHIFT) |
3396 (r->spr_val << WM0_PIPE_SPRITE_SHIFT) |
3401 /* Find the result with the highest level enabled. Check for enable_fbc_wm in
3402 * case both are at the same level. Prefer r1 in case they're the same. */
3403 static struct intel_pipe_wm *
3404 ilk_find_best_result(struct drm_i915_private *dev_priv,
3405 struct intel_pipe_wm *r1,
3406 struct intel_pipe_wm *r2)
3408 int level, max_level = ilk_wm_max_level(dev_priv);
3409 int level1 = 0, level2 = 0;
3411 for (level = 1; level <= max_level; level++) {
3412 if (r1->wm[level].enable)
3414 if (r2->wm[level].enable)
3418 if (level1 == level2) {
3419 if (r2->fbc_wm_enabled && !r1->fbc_wm_enabled)
3423 } else if (level1 > level2) {
3430 /* dirty bits used to track which watermarks need changes */
3431 #define WM_DIRTY_PIPE(pipe) (1 << (pipe))
3432 #define WM_DIRTY_LP(wm_lp) (1 << (15 + (wm_lp)))
3433 #define WM_DIRTY_LP_ALL (WM_DIRTY_LP(1) | WM_DIRTY_LP(2) | WM_DIRTY_LP(3))
3434 #define WM_DIRTY_FBC (1 << 24)
3435 #define WM_DIRTY_DDB (1 << 25)
3437 static unsigned int ilk_compute_wm_dirty(struct drm_i915_private *dev_priv,
3438 const struct ilk_wm_values *old,
3439 const struct ilk_wm_values *new)
3441 unsigned int dirty = 0;
3445 for_each_pipe(dev_priv, pipe) {
3446 if (old->wm_pipe[pipe] != new->wm_pipe[pipe]) {
3447 dirty |= WM_DIRTY_PIPE(pipe);
3448 /* Must disable LP1+ watermarks too */
3449 dirty |= WM_DIRTY_LP_ALL;
3453 if (old->enable_fbc_wm != new->enable_fbc_wm) {
3454 dirty |= WM_DIRTY_FBC;
3455 /* Must disable LP1+ watermarks too */
3456 dirty |= WM_DIRTY_LP_ALL;
3459 if (old->partitioning != new->partitioning) {
3460 dirty |= WM_DIRTY_DDB;
3461 /* Must disable LP1+ watermarks too */
3462 dirty |= WM_DIRTY_LP_ALL;
3465 /* LP1+ watermarks already deemed dirty, no need to continue */
3466 if (dirty & WM_DIRTY_LP_ALL)
3469 /* Find the lowest numbered LP1+ watermark in need of an update... */
3470 for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
3471 if (old->wm_lp[wm_lp - 1] != new->wm_lp[wm_lp - 1] ||
3472 old->wm_lp_spr[wm_lp - 1] != new->wm_lp_spr[wm_lp - 1])
3476 /* ...and mark it and all higher numbered LP1+ watermarks as dirty */
3477 for (; wm_lp <= 3; wm_lp++)
3478 dirty |= WM_DIRTY_LP(wm_lp);
3483 static bool _ilk_disable_lp_wm(struct drm_i915_private *dev_priv,
3486 struct ilk_wm_values *previous = &dev_priv->wm.hw;
3487 bool changed = false;
3489 if (dirty & WM_DIRTY_LP(3) && previous->wm_lp[2] & WM1_LP_SR_EN) {
3490 previous->wm_lp[2] &= ~WM1_LP_SR_EN;
3491 I915_WRITE(WM3_LP_ILK, previous->wm_lp[2]);
3494 if (dirty & WM_DIRTY_LP(2) && previous->wm_lp[1] & WM1_LP_SR_EN) {
3495 previous->wm_lp[1] &= ~WM1_LP_SR_EN;
3496 I915_WRITE(WM2_LP_ILK, previous->wm_lp[1]);
3499 if (dirty & WM_DIRTY_LP(1) && previous->wm_lp[0] & WM1_LP_SR_EN) {
3500 previous->wm_lp[0] &= ~WM1_LP_SR_EN;
3501 I915_WRITE(WM1_LP_ILK, previous->wm_lp[0]);
3506 * Don't touch WM1S_LP_EN here.
3507 * Doing so could cause underruns.
3514 * The spec says we shouldn't write when we don't need, because every write
3515 * causes WMs to be re-evaluated, expending some power.
3517 static void ilk_write_wm_values(struct drm_i915_private *dev_priv,
3518 struct ilk_wm_values *results)
3520 struct ilk_wm_values *previous = &dev_priv->wm.hw;
3524 dirty = ilk_compute_wm_dirty(dev_priv, previous, results);
3528 _ilk_disable_lp_wm(dev_priv, dirty);
3530 if (dirty & WM_DIRTY_PIPE(PIPE_A))
3531 I915_WRITE(WM0_PIPEA_ILK, results->wm_pipe[0]);
3532 if (dirty & WM_DIRTY_PIPE(PIPE_B))
3533 I915_WRITE(WM0_PIPEB_ILK, results->wm_pipe[1]);
3534 if (dirty & WM_DIRTY_PIPE(PIPE_C))
3535 I915_WRITE(WM0_PIPEC_IVB, results->wm_pipe[2]);
3537 if (dirty & WM_DIRTY_DDB) {
3538 if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv)) {
3539 val = I915_READ(WM_MISC);
3540 if (results->partitioning == INTEL_DDB_PART_1_2)
3541 val &= ~WM_MISC_DATA_PARTITION_5_6;
3543 val |= WM_MISC_DATA_PARTITION_5_6;
3544 I915_WRITE(WM_MISC, val);
3546 val = I915_READ(DISP_ARB_CTL2);
3547 if (results->partitioning == INTEL_DDB_PART_1_2)
3548 val &= ~DISP_DATA_PARTITION_5_6;
3550 val |= DISP_DATA_PARTITION_5_6;
3551 I915_WRITE(DISP_ARB_CTL2, val);
3555 if (dirty & WM_DIRTY_FBC) {
3556 val = I915_READ(DISP_ARB_CTL);
3557 if (results->enable_fbc_wm)
3558 val &= ~DISP_FBC_WM_DIS;
3560 val |= DISP_FBC_WM_DIS;
3561 I915_WRITE(DISP_ARB_CTL, val);
3564 if (dirty & WM_DIRTY_LP(1) &&
3565 previous->wm_lp_spr[0] != results->wm_lp_spr[0])
3566 I915_WRITE(WM1S_LP_ILK, results->wm_lp_spr[0]);
3568 if (INTEL_GEN(dev_priv) >= 7) {
3569 if (dirty & WM_DIRTY_LP(2) && previous->wm_lp_spr[1] != results->wm_lp_spr[1])
3570 I915_WRITE(WM2S_LP_IVB, results->wm_lp_spr[1]);
3571 if (dirty & WM_DIRTY_LP(3) && previous->wm_lp_spr[2] != results->wm_lp_spr[2])
3572 I915_WRITE(WM3S_LP_IVB, results->wm_lp_spr[2]);
3575 if (dirty & WM_DIRTY_LP(1) && previous->wm_lp[0] != results->wm_lp[0])
3576 I915_WRITE(WM1_LP_ILK, results->wm_lp[0]);
3577 if (dirty & WM_DIRTY_LP(2) && previous->wm_lp[1] != results->wm_lp[1])
3578 I915_WRITE(WM2_LP_ILK, results->wm_lp[1]);
3579 if (dirty & WM_DIRTY_LP(3) && previous->wm_lp[2] != results->wm_lp[2])
3580 I915_WRITE(WM3_LP_ILK, results->wm_lp[2]);
3582 dev_priv->wm.hw = *results;
3585 bool ilk_disable_lp_wm(struct drm_i915_private *dev_priv)
3587 return _ilk_disable_lp_wm(dev_priv, WM_DIRTY_LP_ALL);
3590 u8 intel_enabled_dbuf_slices_mask(struct drm_i915_private *dev_priv)
3593 int max_slices = INTEL_INFO(dev_priv)->num_supported_dbuf_slices;
3594 u8 enabled_slices_mask = 0;
3596 for (i = 0; i < max_slices; i++) {
3597 if (I915_READ(DBUF_CTL_S(i)) & DBUF_POWER_STATE)
3598 enabled_slices_mask |= BIT(i);
3601 return enabled_slices_mask;
3605 * FIXME: We still don't have the proper code detect if we need to apply the WA,
3606 * so assume we'll always need it in order to avoid underruns.
3608 static bool skl_needs_memory_bw_wa(struct drm_i915_private *dev_priv)
3610 return IS_GEN9_BC(dev_priv) || IS_BROXTON(dev_priv);
3614 intel_has_sagv(struct drm_i915_private *dev_priv)
3617 if (IS_GEN(dev_priv, 12))
3620 return (IS_GEN9_BC(dev_priv) || INTEL_GEN(dev_priv) >= 10) &&
3621 dev_priv->sagv_status != I915_SAGV_NOT_CONTROLLED;
3625 skl_setup_sagv_block_time(struct drm_i915_private *dev_priv)
3627 if (INTEL_GEN(dev_priv) >= 12) {
3631 ret = sandybridge_pcode_read(dev_priv,
3632 GEN12_PCODE_READ_SAGV_BLOCK_TIME_US,
3635 dev_priv->sagv_block_time_us = val;
3639 drm_dbg(&dev_priv->drm, "Couldn't read SAGV block time!\n");
3640 } else if (IS_GEN(dev_priv, 11)) {
3641 dev_priv->sagv_block_time_us = 10;
3643 } else if (IS_GEN(dev_priv, 10)) {
3644 dev_priv->sagv_block_time_us = 20;
3646 } else if (IS_GEN(dev_priv, 9)) {
3647 dev_priv->sagv_block_time_us = 30;
3650 MISSING_CASE(INTEL_GEN(dev_priv));
3653 /* Default to an unusable block time */
3654 dev_priv->sagv_block_time_us = -1;
3658 * SAGV dynamically adjusts the system agent voltage and clock frequencies
3659 * depending on power and performance requirements. The display engine access
3660 * to system memory is blocked during the adjustment time. Because of the
3661 * blocking time, having this enabled can cause full system hangs and/or pipe
3662 * underruns if we don't meet all of the following requirements:
3664 * - <= 1 pipe enabled
3665 * - All planes can enable watermarks for latencies >= SAGV engine block time
3666 * - We're not using an interlaced display configuration
3669 intel_enable_sagv(struct drm_i915_private *dev_priv)
3673 if (!intel_has_sagv(dev_priv))
3676 if (dev_priv->sagv_status == I915_SAGV_ENABLED)
3679 drm_dbg_kms(&dev_priv->drm, "Enabling SAGV\n");
3680 ret = sandybridge_pcode_write(dev_priv, GEN9_PCODE_SAGV_CONTROL,
3683 /* We don't need to wait for SAGV when enabling */
3686 * Some skl systems, pre-release machines in particular,
3687 * don't actually have SAGV.
3689 if (IS_SKYLAKE(dev_priv) && ret == -ENXIO) {
3690 drm_dbg(&dev_priv->drm, "No SAGV found on system, ignoring\n");
3691 dev_priv->sagv_status = I915_SAGV_NOT_CONTROLLED;
3693 } else if (ret < 0) {
3694 drm_err(&dev_priv->drm, "Failed to enable SAGV\n");
3698 dev_priv->sagv_status = I915_SAGV_ENABLED;
3703 intel_disable_sagv(struct drm_i915_private *dev_priv)
3707 if (!intel_has_sagv(dev_priv))
3710 if (dev_priv->sagv_status == I915_SAGV_DISABLED)
3713 drm_dbg_kms(&dev_priv->drm, "Disabling SAGV\n");
3714 /* bspec says to keep retrying for at least 1 ms */
3715 ret = skl_pcode_request(dev_priv, GEN9_PCODE_SAGV_CONTROL,
3717 GEN9_SAGV_IS_DISABLED, GEN9_SAGV_IS_DISABLED,
3720 * Some skl systems, pre-release machines in particular,
3721 * don't actually have SAGV.
3723 if (IS_SKYLAKE(dev_priv) && ret == -ENXIO) {
3724 drm_dbg(&dev_priv->drm, "No SAGV found on system, ignoring\n");
3725 dev_priv->sagv_status = I915_SAGV_NOT_CONTROLLED;
3727 } else if (ret < 0) {
3728 drm_err(&dev_priv->drm, "Failed to disable SAGV (%d)\n", ret);
3732 dev_priv->sagv_status = I915_SAGV_DISABLED;
3736 bool intel_can_enable_sagv(struct intel_atomic_state *state)
3738 struct drm_device *dev = state->base.dev;
3739 struct drm_i915_private *dev_priv = to_i915(dev);
3740 struct intel_crtc *crtc;
3741 struct intel_plane *plane;
3742 struct intel_crtc_state *crtc_state;
3746 if (!intel_has_sagv(dev_priv))
3750 * If there are no active CRTCs, no additional checks need be performed
3752 if (hweight8(state->active_pipes) == 0)
3756 * SKL+ workaround: bspec recommends we disable SAGV when we have
3757 * more then one pipe enabled
3759 if (hweight8(state->active_pipes) > 1)
3762 /* Since we're now guaranteed to only have one active CRTC... */
3763 pipe = ffs(state->active_pipes) - 1;
3764 crtc = intel_get_crtc_for_pipe(dev_priv, pipe);
3765 crtc_state = to_intel_crtc_state(crtc->base.state);
3767 if (crtc_state->hw.adjusted_mode.flags & DRM_MODE_FLAG_INTERLACE)
3770 for_each_intel_plane_on_crtc(dev, crtc, plane) {
3771 struct skl_plane_wm *wm =
3772 &crtc_state->wm.skl.optimal.planes[plane->id];
3774 /* Skip this plane if it's not enabled */
3775 if (!wm->wm[0].plane_en)
3778 /* Find the highest enabled wm level for this plane */
3779 for (level = ilk_wm_max_level(dev_priv);
3780 !wm->wm[level].plane_en; --level)
3783 latency = dev_priv->wm.skl_latency[level];
3785 if (skl_needs_memory_bw_wa(dev_priv) &&
3786 plane->base.state->fb->modifier ==
3787 I915_FORMAT_MOD_X_TILED)
3791 * If any of the planes on this pipe don't enable wm levels that
3792 * incur memory latencies higher than sagv_block_time_us we
3793 * can't enable SAGV.
3795 if (latency < dev_priv->sagv_block_time_us)
3803 * Calculate initial DBuf slice offset, based on slice size
3804 * and mask(i.e if slice size is 1024 and second slice is enabled
3805 * offset would be 1024)
3808 icl_get_first_dbuf_slice_offset(u32 dbuf_slice_mask,
3812 unsigned int offset = 0;
3814 if (!dbuf_slice_mask)
3817 offset = (ffs(dbuf_slice_mask) - 1) * slice_size;
3819 WARN_ON(offset >= ddb_size);
3823 static u16 intel_get_ddb_size(struct drm_i915_private *dev_priv)
3825 u16 ddb_size = INTEL_INFO(dev_priv)->ddb_size;
3827 drm_WARN_ON(&dev_priv->drm, ddb_size == 0);
3829 if (INTEL_GEN(dev_priv) < 11)
3830 return ddb_size - 4; /* 4 blocks for bypass path allocation */
3835 static u8 skl_compute_dbuf_slices(const struct intel_crtc_state *crtc_state,
3839 skl_ddb_get_pipe_allocation_limits(struct drm_i915_private *dev_priv,
3840 const struct intel_crtc_state *crtc_state,
3841 const u64 total_data_rate,
3842 struct skl_ddb_entry *alloc, /* out */
3843 int *num_active /* out */)
3845 struct drm_atomic_state *state = crtc_state->uapi.state;
3846 struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
3847 struct drm_crtc *for_crtc = crtc_state->uapi.crtc;
3848 const struct intel_crtc *crtc;
3849 u32 pipe_width = 0, total_width_in_range = 0, width_before_pipe_in_range = 0;
3850 enum pipe for_pipe = to_intel_crtc(for_crtc)->pipe;
3854 u32 dbuf_slice_mask;
3858 u32 total_slice_mask;
3861 if (drm_WARN_ON(&dev_priv->drm, !state) || !crtc_state->hw.active) {
3864 *num_active = hweight8(dev_priv->active_pipes);
3868 if (intel_state->active_pipe_changes)
3869 active_pipes = intel_state->active_pipes;
3871 active_pipes = dev_priv->active_pipes;
3873 *num_active = hweight8(active_pipes);
3875 ddb_size = intel_get_ddb_size(dev_priv);
3877 slice_size = ddb_size / INTEL_INFO(dev_priv)->num_supported_dbuf_slices;
3880 * If the state doesn't change the active CRTC's or there is no
3881 * modeset request, then there's no need to recalculate;
3882 * the existing pipe allocation limits should remain unchanged.
3883 * Note that we're safe from racing commits since any racing commit
3884 * that changes the active CRTC list or do modeset would need to
3885 * grab _all_ crtc locks, including the one we currently hold.
3887 if (!intel_state->active_pipe_changes && !intel_state->modeset) {
3889 * alloc may be cleared by clear_intel_crtc_state,
3890 * copy from old state to be sure
3892 *alloc = to_intel_crtc_state(for_crtc->state)->wm.skl.ddb;
3897 * Get allowed DBuf slices for correspondent pipe and platform.
3899 dbuf_slice_mask = skl_compute_dbuf_slices(crtc_state, active_pipes);
3901 DRM_DEBUG_KMS("DBuf slice mask %x pipe %c active pipes %x\n",
3903 pipe_name(for_pipe), active_pipes);
3906 * Figure out at which DBuf slice we start, i.e if we start at Dbuf S2
3907 * and slice size is 1024, the offset would be 1024
3909 offset = icl_get_first_dbuf_slice_offset(dbuf_slice_mask,
3910 slice_size, ddb_size);
3913 * Figure out total size of allowed DBuf slices, which is basically
3914 * a number of allowed slices for that pipe multiplied by slice size.
3916 * range ddb entries are still allocated in proportion to display width.
3918 ddb_range_size = hweight8(dbuf_slice_mask) * slice_size;
3921 * Watermark/ddb requirement highly depends upon width of the
3922 * framebuffer, So instead of allocating DDB equally among pipes
3923 * distribute DDB based on resolution/width of the display.
3925 total_slice_mask = dbuf_slice_mask;
3926 for_each_new_intel_crtc_in_state(intel_state, crtc, crtc_state, i) {
3927 const struct drm_display_mode *adjusted_mode =
3928 &crtc_state->hw.adjusted_mode;
3929 enum pipe pipe = crtc->pipe;
3930 int hdisplay, vdisplay;
3931 u32 pipe_dbuf_slice_mask;
3933 if (!crtc_state->hw.active)
3936 pipe_dbuf_slice_mask = skl_compute_dbuf_slices(crtc_state,
3940 * According to BSpec pipe can share one dbuf slice with another
3941 * pipes or pipe can use multiple dbufs, in both cases we
3942 * account for other pipes only if they have exactly same mask.
3943 * However we need to account how many slices we should enable
3946 total_slice_mask |= pipe_dbuf_slice_mask;
3949 * Do not account pipes using other slice sets
3950 * luckily as of current BSpec slice sets do not partially
3951 * intersect(pipes share either same one slice or same slice set
3952 * i.e no partial intersection), so it is enough to check for
3955 if (dbuf_slice_mask != pipe_dbuf_slice_mask)
3958 drm_mode_get_hv_timing(adjusted_mode, &hdisplay, &vdisplay);
3960 total_width_in_range += hdisplay;
3962 if (pipe < for_pipe)
3963 width_before_pipe_in_range += hdisplay;
3964 else if (pipe == for_pipe)
3965 pipe_width = hdisplay;
3969 * FIXME: For now we always enable slice S1 as per
3970 * the Bspec display initialization sequence.
3972 intel_state->enabled_dbuf_slices_mask = total_slice_mask | BIT(DBUF_S1);
3974 start = ddb_range_size * width_before_pipe_in_range / total_width_in_range;
3975 end = ddb_range_size *
3976 (width_before_pipe_in_range + pipe_width) / total_width_in_range;
3978 alloc->start = offset + start;
3979 alloc->end = offset + end;
3981 DRM_DEBUG_KMS("Pipe %d ddb %d-%d\n", for_pipe,
3982 alloc->start, alloc->end);
3983 DRM_DEBUG_KMS("Enabled ddb slices mask %x num supported %d\n",
3984 intel_state->enabled_dbuf_slices_mask,
3985 INTEL_INFO(dev_priv)->num_supported_dbuf_slices);
3988 static int skl_compute_wm_params(const struct intel_crtc_state *crtc_state,
3989 int width, const struct drm_format_info *format,
3990 u64 modifier, unsigned int rotation,
3991 u32 plane_pixel_rate, struct skl_wm_params *wp,
3993 static void skl_compute_plane_wm(const struct intel_crtc_state *crtc_state,
3995 const struct skl_wm_params *wp,
3996 const struct skl_wm_level *result_prev,
3997 struct skl_wm_level *result /* out */);
4000 skl_cursor_allocation(const struct intel_crtc_state *crtc_state,
4003 struct drm_i915_private *dev_priv = to_i915(crtc_state->uapi.crtc->dev);
4004 int level, max_level = ilk_wm_max_level(dev_priv);
4005 struct skl_wm_level wm = {};
4006 int ret, min_ddb_alloc = 0;
4007 struct skl_wm_params wp;
4009 ret = skl_compute_wm_params(crtc_state, 256,
4010 drm_format_info(DRM_FORMAT_ARGB8888),
4011 DRM_FORMAT_MOD_LINEAR,
4013 crtc_state->pixel_rate, &wp, 0);
4014 drm_WARN_ON(&dev_priv->drm, ret);
4016 for (level = 0; level <= max_level; level++) {
4017 skl_compute_plane_wm(crtc_state, level, &wp, &wm, &wm);
4018 if (wm.min_ddb_alloc == U16_MAX)
4021 min_ddb_alloc = wm.min_ddb_alloc;
4024 return max(num_active == 1 ? 32 : 8, min_ddb_alloc);
4027 static void skl_ddb_entry_init_from_hw(struct drm_i915_private *dev_priv,
4028 struct skl_ddb_entry *entry, u32 reg)
4031 entry->start = reg & DDB_ENTRY_MASK;
4032 entry->end = (reg >> DDB_ENTRY_END_SHIFT) & DDB_ENTRY_MASK;
4039 skl_ddb_get_hw_plane_state(struct drm_i915_private *dev_priv,
4040 const enum pipe pipe,
4041 const enum plane_id plane_id,
4042 struct skl_ddb_entry *ddb_y,
4043 struct skl_ddb_entry *ddb_uv)
4048 /* Cursor doesn't support NV12/planar, so no extra calculation needed */
4049 if (plane_id == PLANE_CURSOR) {
4050 val = I915_READ(CUR_BUF_CFG(pipe));
4051 skl_ddb_entry_init_from_hw(dev_priv, ddb_y, val);
4055 val = I915_READ(PLANE_CTL(pipe, plane_id));
4057 /* No DDB allocated for disabled planes */
4058 if (val & PLANE_CTL_ENABLE)
4059 fourcc = skl_format_to_fourcc(val & PLANE_CTL_FORMAT_MASK,
4060 val & PLANE_CTL_ORDER_RGBX,
4061 val & PLANE_CTL_ALPHA_MASK);
4063 if (INTEL_GEN(dev_priv) >= 11) {
4064 val = I915_READ(PLANE_BUF_CFG(pipe, plane_id));
4065 skl_ddb_entry_init_from_hw(dev_priv, ddb_y, val);
4067 val = I915_READ(PLANE_BUF_CFG(pipe, plane_id));
4068 val2 = I915_READ(PLANE_NV12_BUF_CFG(pipe, plane_id));
4071 drm_format_info_is_yuv_semiplanar(drm_format_info(fourcc)))
4074 skl_ddb_entry_init_from_hw(dev_priv, ddb_y, val);
4075 skl_ddb_entry_init_from_hw(dev_priv, ddb_uv, val2);
4079 void skl_pipe_ddb_get_hw_state(struct intel_crtc *crtc,
4080 struct skl_ddb_entry *ddb_y,
4081 struct skl_ddb_entry *ddb_uv)
4083 struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
4084 enum intel_display_power_domain power_domain;
4085 enum pipe pipe = crtc->pipe;
4086 intel_wakeref_t wakeref;
4087 enum plane_id plane_id;
4089 power_domain = POWER_DOMAIN_PIPE(pipe);
4090 wakeref = intel_display_power_get_if_enabled(dev_priv, power_domain);
4094 for_each_plane_id_on_crtc(crtc, plane_id)
4095 skl_ddb_get_hw_plane_state(dev_priv, pipe,
4100 intel_display_power_put(dev_priv, power_domain, wakeref);
4103 void skl_ddb_get_hw_state(struct drm_i915_private *dev_priv)
4105 dev_priv->enabled_dbuf_slices_mask =
4106 intel_enabled_dbuf_slices_mask(dev_priv);
4110 * Determines the downscale amount of a plane for the purposes of watermark calculations.
4111 * The bspec defines downscale amount as:
4114 * Horizontal down scale amount = maximum[1, Horizontal source size /
4115 * Horizontal destination size]
4116 * Vertical down scale amount = maximum[1, Vertical source size /
4117 * Vertical destination size]
4118 * Total down scale amount = Horizontal down scale amount *
4119 * Vertical down scale amount
4122 * Return value is provided in 16.16 fixed point form to retain fractional part.
4123 * Caller should take care of dividing & rounding off the value.
4125 static uint_fixed_16_16_t
4126 skl_plane_downscale_amount(const struct intel_crtc_state *crtc_state,
4127 const struct intel_plane_state *plane_state)
4129 u32 src_w, src_h, dst_w, dst_h;
4130 uint_fixed_16_16_t fp_w_ratio, fp_h_ratio;
4131 uint_fixed_16_16_t downscale_h, downscale_w;
4133 if (WARN_ON(!intel_wm_plane_visible(crtc_state, plane_state)))
4134 return u32_to_fixed16(0);
4137 * Src coordinates are already rotated by 270 degrees for
4138 * the 90/270 degree plane rotation cases (to match the
4139 * GTT mapping), hence no need to account for rotation here.
4141 * n.b., src is 16.16 fixed point, dst is whole integer.
4143 src_w = drm_rect_width(&plane_state->uapi.src) >> 16;
4144 src_h = drm_rect_height(&plane_state->uapi.src) >> 16;
4145 dst_w = drm_rect_width(&plane_state->uapi.dst);
4146 dst_h = drm_rect_height(&plane_state->uapi.dst);
4148 fp_w_ratio = div_fixed16(src_w, dst_w);
4149 fp_h_ratio = div_fixed16(src_h, dst_h);
4150 downscale_w = max_fixed16(fp_w_ratio, u32_to_fixed16(1));
4151 downscale_h = max_fixed16(fp_h_ratio, u32_to_fixed16(1));
4153 return mul_fixed16(downscale_w, downscale_h);
4156 struct dbuf_slice_conf_entry {
4158 u8 dbuf_mask[I915_MAX_PIPES];
4162 * Table taken from Bspec 12716
4163 * Pipes do have some preferred DBuf slice affinity,
4164 * plus there are some hardcoded requirements on how
4165 * those should be distributed for multipipe scenarios.
4166 * For more DBuf slices algorithm can get even more messy
4167 * and less readable, so decided to use a table almost
4168 * as is from BSpec itself - that way it is at least easier
4169 * to compare, change and check.
4171 static const struct dbuf_slice_conf_entry icl_allowed_dbufs[] =
4172 /* Autogenerated with igt/tools/intel_dbuf_map tool: */
4175 .active_pipes = BIT(PIPE_A),
4177 [PIPE_A] = BIT(DBUF_S1),
4181 .active_pipes = BIT(PIPE_B),
4183 [PIPE_B] = BIT(DBUF_S1),
4187 .active_pipes = BIT(PIPE_A) | BIT(PIPE_B),
4189 [PIPE_A] = BIT(DBUF_S1),
4190 [PIPE_B] = BIT(DBUF_S2),
4194 .active_pipes = BIT(PIPE_C),
4196 [PIPE_C] = BIT(DBUF_S2),
4200 .active_pipes = BIT(PIPE_A) | BIT(PIPE_C),
4202 [PIPE_A] = BIT(DBUF_S1),
4203 [PIPE_C] = BIT(DBUF_S2),
4207 .active_pipes = BIT(PIPE_B) | BIT(PIPE_C),
4209 [PIPE_B] = BIT(DBUF_S1),
4210 [PIPE_C] = BIT(DBUF_S2),
4214 .active_pipes = BIT(PIPE_A) | BIT(PIPE_B) | BIT(PIPE_C),
4216 [PIPE_A] = BIT(DBUF_S1),
4217 [PIPE_B] = BIT(DBUF_S1),
4218 [PIPE_C] = BIT(DBUF_S2),
4225 * Table taken from Bspec 49255
4226 * Pipes do have some preferred DBuf slice affinity,
4227 * plus there are some hardcoded requirements on how
4228 * those should be distributed for multipipe scenarios.
4229 * For more DBuf slices algorithm can get even more messy
4230 * and less readable, so decided to use a table almost
4231 * as is from BSpec itself - that way it is at least easier
4232 * to compare, change and check.
4234 static const struct dbuf_slice_conf_entry tgl_allowed_dbufs[] =
4235 /* Autogenerated with igt/tools/intel_dbuf_map tool: */
4238 .active_pipes = BIT(PIPE_A),
4240 [PIPE_A] = BIT(DBUF_S1) | BIT(DBUF_S2),
4244 .active_pipes = BIT(PIPE_B),
4246 [PIPE_B] = BIT(DBUF_S1) | BIT(DBUF_S2),
4250 .active_pipes = BIT(PIPE_A) | BIT(PIPE_B),
4252 [PIPE_A] = BIT(DBUF_S2),
4253 [PIPE_B] = BIT(DBUF_S1),
4257 .active_pipes = BIT(PIPE_C),
4259 [PIPE_C] = BIT(DBUF_S2) | BIT(DBUF_S1),
4263 .active_pipes = BIT(PIPE_A) | BIT(PIPE_C),
4265 [PIPE_A] = BIT(DBUF_S1),
4266 [PIPE_C] = BIT(DBUF_S2),
4270 .active_pipes = BIT(PIPE_B) | BIT(PIPE_C),
4272 [PIPE_B] = BIT(DBUF_S1),
4273 [PIPE_C] = BIT(DBUF_S2),
4277 .active_pipes = BIT(PIPE_A) | BIT(PIPE_B) | BIT(PIPE_C),
4279 [PIPE_A] = BIT(DBUF_S1),
4280 [PIPE_B] = BIT(DBUF_S1),
4281 [PIPE_C] = BIT(DBUF_S2),
4285 .active_pipes = BIT(PIPE_D),
4287 [PIPE_D] = BIT(DBUF_S2) | BIT(DBUF_S1),
4291 .active_pipes = BIT(PIPE_A) | BIT(PIPE_D),
4293 [PIPE_A] = BIT(DBUF_S1),
4294 [PIPE_D] = BIT(DBUF_S2),
4298 .active_pipes = BIT(PIPE_B) | BIT(PIPE_D),
4300 [PIPE_B] = BIT(DBUF_S1),
4301 [PIPE_D] = BIT(DBUF_S2),
4305 .active_pipes = BIT(PIPE_A) | BIT(PIPE_B) | BIT(PIPE_D),
4307 [PIPE_A] = BIT(DBUF_S1),
4308 [PIPE_B] = BIT(DBUF_S1),
4309 [PIPE_D] = BIT(DBUF_S2),
4313 .active_pipes = BIT(PIPE_C) | BIT(PIPE_D),
4315 [PIPE_C] = BIT(DBUF_S1),
4316 [PIPE_D] = BIT(DBUF_S2),
4320 .active_pipes = BIT(PIPE_A) | BIT(PIPE_C) | BIT(PIPE_D),
4322 [PIPE_A] = BIT(DBUF_S1),
4323 [PIPE_C] = BIT(DBUF_S2),
4324 [PIPE_D] = BIT(DBUF_S2),
4328 .active_pipes = BIT(PIPE_B) | BIT(PIPE_C) | BIT(PIPE_D),
4330 [PIPE_B] = BIT(DBUF_S1),
4331 [PIPE_C] = BIT(DBUF_S2),
4332 [PIPE_D] = BIT(DBUF_S2),
4336 .active_pipes = BIT(PIPE_A) | BIT(PIPE_B) | BIT(PIPE_C) | BIT(PIPE_D),
4338 [PIPE_A] = BIT(DBUF_S1),
4339 [PIPE_B] = BIT(DBUF_S1),
4340 [PIPE_C] = BIT(DBUF_S2),
4341 [PIPE_D] = BIT(DBUF_S2),
4347 static u8 compute_dbuf_slices(enum pipe pipe, u8 active_pipes,
4348 const struct dbuf_slice_conf_entry *dbuf_slices)
4352 for (i = 0; i < dbuf_slices[i].active_pipes; i++) {
4353 if (dbuf_slices[i].active_pipes == active_pipes)
4354 return dbuf_slices[i].dbuf_mask[pipe];
4360 * This function finds an entry with same enabled pipe configuration and
4361 * returns correspondent DBuf slice mask as stated in BSpec for particular
4364 static u8 icl_compute_dbuf_slices(enum pipe pipe, u8 active_pipes)
4367 * FIXME: For ICL this is still a bit unclear as prev BSpec revision
4368 * required calculating "pipe ratio" in order to determine
4369 * if one or two slices can be used for single pipe configurations
4370 * as additional constraint to the existing table.
4371 * However based on recent info, it should be not "pipe ratio"
4372 * but rather ratio between pixel_rate and cdclk with additional
4373 * constants, so for now we are using only table until this is
4374 * clarified. Also this is the reason why crtc_state param is
4375 * still here - we will need it once those additional constraints
4378 return compute_dbuf_slices(pipe, active_pipes, icl_allowed_dbufs);
4381 static u8 tgl_compute_dbuf_slices(enum pipe pipe, u8 active_pipes)
4383 return compute_dbuf_slices(pipe, active_pipes, tgl_allowed_dbufs);
4386 static u8 skl_compute_dbuf_slices(const struct intel_crtc_state *crtc_state,
4389 struct intel_crtc *crtc = to_intel_crtc(crtc_state->uapi.crtc);
4390 struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
4391 enum pipe pipe = crtc->pipe;
4393 if (IS_GEN(dev_priv, 12))
4394 return tgl_compute_dbuf_slices(pipe, active_pipes);
4395 else if (IS_GEN(dev_priv, 11))
4396 return icl_compute_dbuf_slices(pipe, active_pipes);
4398 * For anything else just return one slice yet.
4399 * Should be extended for other platforms.
4401 return BIT(DBUF_S1);
4405 skl_plane_relative_data_rate(const struct intel_crtc_state *crtc_state,
4406 const struct intel_plane_state *plane_state,
4409 struct intel_plane *plane = to_intel_plane(plane_state->uapi.plane);
4410 const struct drm_framebuffer *fb = plane_state->hw.fb;
4412 u32 width = 0, height = 0;
4413 uint_fixed_16_16_t down_scale_amount;
4416 if (!plane_state->uapi.visible)
4419 if (plane->id == PLANE_CURSOR)
4422 if (color_plane == 1 &&
4423 !intel_format_info_is_yuv_semiplanar(fb->format, fb->modifier))
4427 * Src coordinates are already rotated by 270 degrees for
4428 * the 90/270 degree plane rotation cases (to match the
4429 * GTT mapping), hence no need to account for rotation here.
4431 width = drm_rect_width(&plane_state->uapi.src) >> 16;
4432 height = drm_rect_height(&plane_state->uapi.src) >> 16;
4434 /* UV plane does 1/2 pixel sub-sampling */
4435 if (color_plane == 1) {
4440 data_rate = width * height;
4442 down_scale_amount = skl_plane_downscale_amount(crtc_state, plane_state);
4444 rate = mul_round_up_u32_fixed16(data_rate, down_scale_amount);
4446 rate *= fb->format->cpp[color_plane];
4451 skl_get_total_relative_data_rate(struct intel_crtc_state *crtc_state,
4452 u64 *plane_data_rate,
4453 u64 *uv_plane_data_rate)
4455 struct intel_plane *plane;
4456 const struct intel_plane_state *plane_state;
4457 u64 total_data_rate = 0;
4459 /* Calculate and cache data rate for each plane */
4460 intel_atomic_crtc_state_for_each_plane_state(plane, plane_state, crtc_state) {
4461 enum plane_id plane_id = plane->id;
4465 rate = skl_plane_relative_data_rate(crtc_state, plane_state, 0);
4466 plane_data_rate[plane_id] = rate;
4467 total_data_rate += rate;
4470 rate = skl_plane_relative_data_rate(crtc_state, plane_state, 1);
4471 uv_plane_data_rate[plane_id] = rate;
4472 total_data_rate += rate;
4475 return total_data_rate;
4479 icl_get_total_relative_data_rate(struct intel_crtc_state *crtc_state,
4480 u64 *plane_data_rate)
4482 struct intel_plane *plane;
4483 const struct intel_plane_state *plane_state;
4484 u64 total_data_rate = 0;
4486 /* Calculate and cache data rate for each plane */
4487 intel_atomic_crtc_state_for_each_plane_state(plane, plane_state, crtc_state) {
4488 enum plane_id plane_id = plane->id;
4491 if (!plane_state->planar_linked_plane) {
4492 rate = skl_plane_relative_data_rate(crtc_state, plane_state, 0);
4493 plane_data_rate[plane_id] = rate;
4494 total_data_rate += rate;
4496 enum plane_id y_plane_id;
4499 * The slave plane might not iterate in
4500 * intel_atomic_crtc_state_for_each_plane_state(),
4501 * and needs the master plane state which may be
4502 * NULL if we try get_new_plane_state(), so we
4503 * always calculate from the master.
4505 if (plane_state->planar_slave)
4508 /* Y plane rate is calculated on the slave */
4509 rate = skl_plane_relative_data_rate(crtc_state, plane_state, 0);
4510 y_plane_id = plane_state->planar_linked_plane->id;
4511 plane_data_rate[y_plane_id] = rate;
4512 total_data_rate += rate;
4514 rate = skl_plane_relative_data_rate(crtc_state, plane_state, 1);
4515 plane_data_rate[plane_id] = rate;
4516 total_data_rate += rate;
4520 return total_data_rate;
4524 skl_allocate_pipe_ddb(struct intel_crtc_state *crtc_state)
4526 struct intel_crtc *crtc = to_intel_crtc(crtc_state->uapi.crtc);
4527 struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
4528 struct skl_ddb_entry *alloc = &crtc_state->wm.skl.ddb;
4529 u16 alloc_size, start = 0;
4530 u16 total[I915_MAX_PLANES] = {};
4531 u16 uv_total[I915_MAX_PLANES] = {};
4532 u64 total_data_rate;
4533 enum plane_id plane_id;
4535 u64 plane_data_rate[I915_MAX_PLANES] = {};
4536 u64 uv_plane_data_rate[I915_MAX_PLANES] = {};
4540 /* Clear the partitioning for disabled planes. */
4541 memset(crtc_state->wm.skl.plane_ddb_y, 0, sizeof(crtc_state->wm.skl.plane_ddb_y));
4542 memset(crtc_state->wm.skl.plane_ddb_uv, 0, sizeof(crtc_state->wm.skl.plane_ddb_uv));
4544 if (!crtc_state->hw.active) {
4545 alloc->start = alloc->end = 0;
4549 if (INTEL_GEN(dev_priv) >= 11)
4551 icl_get_total_relative_data_rate(crtc_state,
4555 skl_get_total_relative_data_rate(crtc_state,
4557 uv_plane_data_rate);
4560 skl_ddb_get_pipe_allocation_limits(dev_priv, crtc_state, total_data_rate,
4561 alloc, &num_active);
4562 alloc_size = skl_ddb_entry_size(alloc);
4563 if (alloc_size == 0)
4566 /* Allocate fixed number of blocks for cursor. */
4567 total[PLANE_CURSOR] = skl_cursor_allocation(crtc_state, num_active);
4568 alloc_size -= total[PLANE_CURSOR];
4569 crtc_state->wm.skl.plane_ddb_y[PLANE_CURSOR].start =
4570 alloc->end - total[PLANE_CURSOR];
4571 crtc_state->wm.skl.plane_ddb_y[PLANE_CURSOR].end = alloc->end;
4573 if (total_data_rate == 0)
4577 * Find the highest watermark level for which we can satisfy the block
4578 * requirement of active planes.
4580 for (level = ilk_wm_max_level(dev_priv); level >= 0; level--) {
4582 for_each_plane_id_on_crtc(crtc, plane_id) {
4583 const struct skl_plane_wm *wm =
4584 &crtc_state->wm.skl.optimal.planes[plane_id];
4586 if (plane_id == PLANE_CURSOR) {
4587 if (wm->wm[level].min_ddb_alloc > total[PLANE_CURSOR]) {
4588 drm_WARN_ON(&dev_priv->drm,
4589 wm->wm[level].min_ddb_alloc != U16_MAX);
4596 blocks += wm->wm[level].min_ddb_alloc;
4597 blocks += wm->uv_wm[level].min_ddb_alloc;
4600 if (blocks <= alloc_size) {
4601 alloc_size -= blocks;
4607 drm_dbg_kms(&dev_priv->drm,
4608 "Requested display configuration exceeds system DDB limitations");
4609 drm_dbg_kms(&dev_priv->drm, "minimum required %d/%d\n",
4610 blocks, alloc_size);
4615 * Grant each plane the blocks it requires at the highest achievable
4616 * watermark level, plus an extra share of the leftover blocks
4617 * proportional to its relative data rate.
4619 for_each_plane_id_on_crtc(crtc, plane_id) {
4620 const struct skl_plane_wm *wm =
4621 &crtc_state->wm.skl.optimal.planes[plane_id];
4625 if (plane_id == PLANE_CURSOR)
4629 * We've accounted for all active planes; remaining planes are
4632 if (total_data_rate == 0)
4635 rate = plane_data_rate[plane_id];
4636 extra = min_t(u16, alloc_size,
4637 DIV64_U64_ROUND_UP(alloc_size * rate,
4639 total[plane_id] = wm->wm[level].min_ddb_alloc + extra;
4640 alloc_size -= extra;
4641 total_data_rate -= rate;
4643 if (total_data_rate == 0)
4646 rate = uv_plane_data_rate[plane_id];
4647 extra = min_t(u16, alloc_size,
4648 DIV64_U64_ROUND_UP(alloc_size * rate,
4650 uv_total[plane_id] = wm->uv_wm[level].min_ddb_alloc + extra;
4651 alloc_size -= extra;
4652 total_data_rate -= rate;
4654 drm_WARN_ON(&dev_priv->drm, alloc_size != 0 || total_data_rate != 0);
4656 /* Set the actual DDB start/end points for each plane */
4657 start = alloc->start;
4658 for_each_plane_id_on_crtc(crtc, plane_id) {
4659 struct skl_ddb_entry *plane_alloc =
4660 &crtc_state->wm.skl.plane_ddb_y[plane_id];
4661 struct skl_ddb_entry *uv_plane_alloc =
4662 &crtc_state->wm.skl.plane_ddb_uv[plane_id];
4664 if (plane_id == PLANE_CURSOR)
4667 /* Gen11+ uses a separate plane for UV watermarks */
4668 drm_WARN_ON(&dev_priv->drm,
4669 INTEL_GEN(dev_priv) >= 11 && uv_total[plane_id]);
4671 /* Leave disabled planes at (0,0) */
4672 if (total[plane_id]) {
4673 plane_alloc->start = start;
4674 start += total[plane_id];
4675 plane_alloc->end = start;
4678 if (uv_total[plane_id]) {
4679 uv_plane_alloc->start = start;
4680 start += uv_total[plane_id];
4681 uv_plane_alloc->end = start;
4686 * When we calculated watermark values we didn't know how high
4687 * of a level we'd actually be able to hit, so we just marked
4688 * all levels as "enabled." Go back now and disable the ones
4689 * that aren't actually possible.
4691 for (level++; level <= ilk_wm_max_level(dev_priv); level++) {
4692 for_each_plane_id_on_crtc(crtc, plane_id) {
4693 struct skl_plane_wm *wm =
4694 &crtc_state->wm.skl.optimal.planes[plane_id];
4697 * We only disable the watermarks for each plane if
4698 * they exceed the ddb allocation of said plane. This
4699 * is done so that we don't end up touching cursor
4700 * watermarks needlessly when some other plane reduces
4701 * our max possible watermark level.
4703 * Bspec has this to say about the PLANE_WM enable bit:
4704 * "All the watermarks at this level for all enabled
4705 * planes must be enabled before the level will be used."
4706 * So this is actually safe to do.
4708 if (wm->wm[level].min_ddb_alloc > total[plane_id] ||
4709 wm->uv_wm[level].min_ddb_alloc > uv_total[plane_id])
4710 memset(&wm->wm[level], 0, sizeof(wm->wm[level]));
4713 * Wa_1408961008:icl, ehl
4714 * Underruns with WM1+ disabled
4716 if (IS_GEN(dev_priv, 11) &&
4717 level == 1 && wm->wm[0].plane_en) {
4718 wm->wm[level].plane_res_b = wm->wm[0].plane_res_b;
4719 wm->wm[level].plane_res_l = wm->wm[0].plane_res_l;
4720 wm->wm[level].ignore_lines = wm->wm[0].ignore_lines;
4726 * Go back and disable the transition watermark if it turns out we
4727 * don't have enough DDB blocks for it.
4729 for_each_plane_id_on_crtc(crtc, plane_id) {
4730 struct skl_plane_wm *wm =
4731 &crtc_state->wm.skl.optimal.planes[plane_id];
4733 if (wm->trans_wm.plane_res_b >= total[plane_id])
4734 memset(&wm->trans_wm, 0, sizeof(wm->trans_wm));
4741 * The max latency should be 257 (max the punit can code is 255 and we add 2us
4742 * for the read latency) and cpp should always be <= 8, so that
4743 * should allow pixel_rate up to ~2 GHz which seems sufficient since max
4744 * 2xcdclk is 1350 MHz and the pixel rate should never exceed that.
4746 static uint_fixed_16_16_t
4747 skl_wm_method1(const struct drm_i915_private *dev_priv, u32 pixel_rate,
4748 u8 cpp, u32 latency, u32 dbuf_block_size)
4750 u32 wm_intermediate_val;
4751 uint_fixed_16_16_t ret;
4754 return FP_16_16_MAX;
4756 wm_intermediate_val = latency * pixel_rate * cpp;
4757 ret = div_fixed16(wm_intermediate_val, 1000 * dbuf_block_size);
4759 if (INTEL_GEN(dev_priv) >= 10)
4760 ret = add_fixed16_u32(ret, 1);
4765 static uint_fixed_16_16_t
4766 skl_wm_method2(u32 pixel_rate, u32 pipe_htotal, u32 latency,
4767 uint_fixed_16_16_t plane_blocks_per_line)
4769 u32 wm_intermediate_val;
4770 uint_fixed_16_16_t ret;
4773 return FP_16_16_MAX;
4775 wm_intermediate_val = latency * pixel_rate;
4776 wm_intermediate_val = DIV_ROUND_UP(wm_intermediate_val,
4777 pipe_htotal * 1000);
4778 ret = mul_u32_fixed16(wm_intermediate_val, plane_blocks_per_line);
4782 static uint_fixed_16_16_t
4783 intel_get_linetime_us(const struct intel_crtc_state *crtc_state)
4787 uint_fixed_16_16_t linetime_us;
4789 if (!crtc_state->hw.active)
4790 return u32_to_fixed16(0);
4792 pixel_rate = crtc_state->pixel_rate;
4794 if (WARN_ON(pixel_rate == 0))
4795 return u32_to_fixed16(0);
4797 crtc_htotal = crtc_state->hw.adjusted_mode.crtc_htotal;
4798 linetime_us = div_fixed16(crtc_htotal * 1000, pixel_rate);
4804 skl_adjusted_plane_pixel_rate(const struct intel_crtc_state *crtc_state,
4805 const struct intel_plane_state *plane_state)
4807 u64 adjusted_pixel_rate;
4808 uint_fixed_16_16_t downscale_amount;
4810 /* Shouldn't reach here on disabled planes... */
4811 if (WARN_ON(!intel_wm_plane_visible(crtc_state, plane_state)))
4815 * Adjusted plane pixel rate is just the pipe's adjusted pixel rate
4816 * with additional adjustments for plane-specific scaling.
4818 adjusted_pixel_rate = crtc_state->pixel_rate;
4819 downscale_amount = skl_plane_downscale_amount(crtc_state, plane_state);
4821 return mul_round_up_u32_fixed16(adjusted_pixel_rate,
4826 skl_compute_wm_params(const struct intel_crtc_state *crtc_state,
4827 int width, const struct drm_format_info *format,
4828 u64 modifier, unsigned int rotation,
4829 u32 plane_pixel_rate, struct skl_wm_params *wp,
4832 struct intel_crtc *crtc = to_intel_crtc(crtc_state->uapi.crtc);
4833 struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
4836 /* only planar format has two planes */
4837 if (color_plane == 1 &&
4838 !intel_format_info_is_yuv_semiplanar(format, modifier)) {
4839 drm_dbg_kms(&dev_priv->drm,
4840 "Non planar format have single plane\n");
4844 wp->y_tiled = modifier == I915_FORMAT_MOD_Y_TILED ||
4845 modifier == I915_FORMAT_MOD_Yf_TILED ||
4846 modifier == I915_FORMAT_MOD_Y_TILED_CCS ||
4847 modifier == I915_FORMAT_MOD_Yf_TILED_CCS;
4848 wp->x_tiled = modifier == I915_FORMAT_MOD_X_TILED;
4849 wp->rc_surface = modifier == I915_FORMAT_MOD_Y_TILED_CCS ||
4850 modifier == I915_FORMAT_MOD_Yf_TILED_CCS;
4851 wp->is_planar = intel_format_info_is_yuv_semiplanar(format, modifier);
4854 if (color_plane == 1 && wp->is_planar)
4857 wp->cpp = format->cpp[color_plane];
4858 wp->plane_pixel_rate = plane_pixel_rate;
4860 if (INTEL_GEN(dev_priv) >= 11 &&
4861 modifier == I915_FORMAT_MOD_Yf_TILED && wp->cpp == 1)
4862 wp->dbuf_block_size = 256;
4864 wp->dbuf_block_size = 512;
4866 if (drm_rotation_90_or_270(rotation)) {
4869 wp->y_min_scanlines = 16;
4872 wp->y_min_scanlines = 8;
4875 wp->y_min_scanlines = 4;
4878 MISSING_CASE(wp->cpp);
4882 wp->y_min_scanlines = 4;
4885 if (skl_needs_memory_bw_wa(dev_priv))
4886 wp->y_min_scanlines *= 2;
4888 wp->plane_bytes_per_line = wp->width * wp->cpp;
4890 interm_pbpl = DIV_ROUND_UP(wp->plane_bytes_per_line *
4891 wp->y_min_scanlines,
4892 wp->dbuf_block_size);
4894 if (INTEL_GEN(dev_priv) >= 10)
4897 wp->plane_blocks_per_line = div_fixed16(interm_pbpl,
4898 wp->y_min_scanlines);
4899 } else if (wp->x_tiled && IS_GEN(dev_priv, 9)) {
4900 interm_pbpl = DIV_ROUND_UP(wp->plane_bytes_per_line,
4901 wp->dbuf_block_size);
4902 wp->plane_blocks_per_line = u32_to_fixed16(interm_pbpl);
4904 interm_pbpl = DIV_ROUND_UP(wp->plane_bytes_per_line,
4905 wp->dbuf_block_size) + 1;
4906 wp->plane_blocks_per_line = u32_to_fixed16(interm_pbpl);
4909 wp->y_tile_minimum = mul_u32_fixed16(wp->y_min_scanlines,
4910 wp->plane_blocks_per_line);
4912 wp->linetime_us = fixed16_to_u32_round_up(
4913 intel_get_linetime_us(crtc_state));
4919 skl_compute_plane_wm_params(const struct intel_crtc_state *crtc_state,
4920 const struct intel_plane_state *plane_state,
4921 struct skl_wm_params *wp, int color_plane)
4923 const struct drm_framebuffer *fb = plane_state->hw.fb;
4927 * Src coordinates are already rotated by 270 degrees for
4928 * the 90/270 degree plane rotation cases (to match the
4929 * GTT mapping), hence no need to account for rotation here.
4931 width = drm_rect_width(&plane_state->uapi.src) >> 16;
4933 return skl_compute_wm_params(crtc_state, width,
4934 fb->format, fb->modifier,
4935 plane_state->hw.rotation,
4936 skl_adjusted_plane_pixel_rate(crtc_state, plane_state),
4940 static bool skl_wm_has_lines(struct drm_i915_private *dev_priv, int level)
4942 if (INTEL_GEN(dev_priv) >= 10 || IS_GEMINILAKE(dev_priv))
4945 /* The number of lines are ignored for the level 0 watermark. */
4949 static void skl_compute_plane_wm(const struct intel_crtc_state *crtc_state,
4951 const struct skl_wm_params *wp,
4952 const struct skl_wm_level *result_prev,
4953 struct skl_wm_level *result /* out */)
4955 struct drm_i915_private *dev_priv = to_i915(crtc_state->uapi.crtc->dev);
4956 u32 latency = dev_priv->wm.skl_latency[level];
4957 uint_fixed_16_16_t method1, method2;
4958 uint_fixed_16_16_t selected_result;
4959 u32 res_blocks, res_lines, min_ddb_alloc = 0;
4963 result->min_ddb_alloc = U16_MAX;
4968 * WaIncreaseLatencyIPCEnabled: kbl,cfl
4969 * Display WA #1141: kbl,cfl
4971 if ((IS_KABYLAKE(dev_priv) || IS_COFFEELAKE(dev_priv)) ||
4972 dev_priv->ipc_enabled)
4975 if (skl_needs_memory_bw_wa(dev_priv) && wp->x_tiled)
4978 method1 = skl_wm_method1(dev_priv, wp->plane_pixel_rate,
4979 wp->cpp, latency, wp->dbuf_block_size);
4980 method2 = skl_wm_method2(wp->plane_pixel_rate,
4981 crtc_state->hw.adjusted_mode.crtc_htotal,
4983 wp->plane_blocks_per_line);
4986 selected_result = max_fixed16(method2, wp->y_tile_minimum);
4988 if ((wp->cpp * crtc_state->hw.adjusted_mode.crtc_htotal /
4989 wp->dbuf_block_size < 1) &&
4990 (wp->plane_bytes_per_line / wp->dbuf_block_size < 1)) {
4991 selected_result = method2;
4992 } else if (latency >= wp->linetime_us) {
4993 if (IS_GEN(dev_priv, 9) &&
4994 !IS_GEMINILAKE(dev_priv))
4995 selected_result = min_fixed16(method1, method2);
4997 selected_result = method2;
4999 selected_result = method1;
5003 res_blocks = fixed16_to_u32_round_up(selected_result) + 1;
5004 res_lines = div_round_up_fixed16(selected_result,
5005 wp->plane_blocks_per_line);
5007 if (IS_GEN9_BC(dev_priv) || IS_BROXTON(dev_priv)) {
5008 /* Display WA #1125: skl,bxt,kbl */
5009 if (level == 0 && wp->rc_surface)
5011 fixed16_to_u32_round_up(wp->y_tile_minimum);
5013 /* Display WA #1126: skl,bxt,kbl */
5014 if (level >= 1 && level <= 7) {
5017 fixed16_to_u32_round_up(wp->y_tile_minimum);
5018 res_lines += wp->y_min_scanlines;
5024 * Make sure result blocks for higher latency levels are
5025 * atleast as high as level below the current level.
5026 * Assumption in DDB algorithm optimization for special
5027 * cases. Also covers Display WA #1125 for RC.
5029 if (result_prev->plane_res_b > res_blocks)
5030 res_blocks = result_prev->plane_res_b;
5034 if (INTEL_GEN(dev_priv) >= 11) {
5038 if (res_lines % wp->y_min_scanlines == 0)
5039 extra_lines = wp->y_min_scanlines;
5041 extra_lines = wp->y_min_scanlines * 2 -
5042 res_lines % wp->y_min_scanlines;
5044 min_ddb_alloc = mul_round_up_u32_fixed16(res_lines + extra_lines,
5045 wp->plane_blocks_per_line);
5047 min_ddb_alloc = res_blocks +
5048 DIV_ROUND_UP(res_blocks, 10);
5052 if (!skl_wm_has_lines(dev_priv, level))
5055 if (res_lines > 31) {
5057 result->min_ddb_alloc = U16_MAX;
5062 * If res_lines is valid, assume we can use this watermark level
5063 * for now. We'll come back and disable it after we calculate the
5064 * DDB allocation if it turns out we don't actually have enough
5065 * blocks to satisfy it.
5067 result->plane_res_b = res_blocks;
5068 result->plane_res_l = res_lines;
5069 /* Bspec says: value >= plane ddb allocation -> invalid, hence the +1 here */
5070 result->min_ddb_alloc = max(min_ddb_alloc, res_blocks) + 1;
5071 result->plane_en = true;
5075 skl_compute_wm_levels(const struct intel_crtc_state *crtc_state,
5076 const struct skl_wm_params *wm_params,
5077 struct skl_wm_level *levels)
5079 struct drm_i915_private *dev_priv = to_i915(crtc_state->uapi.crtc->dev);
5080 int level, max_level = ilk_wm_max_level(dev_priv);
5081 struct skl_wm_level *result_prev = &levels[0];
5083 for (level = 0; level <= max_level; level++) {
5084 struct skl_wm_level *result = &levels[level];
5086 skl_compute_plane_wm(crtc_state, level, wm_params,
5087 result_prev, result);
5089 result_prev = result;
5093 static void skl_compute_transition_wm(const struct intel_crtc_state *crtc_state,
5094 const struct skl_wm_params *wp,
5095 struct skl_plane_wm *wm)
5097 struct drm_device *dev = crtc_state->uapi.crtc->dev;
5098 const struct drm_i915_private *dev_priv = to_i915(dev);
5099 u16 trans_min, trans_y_tile_min;
5100 const u16 trans_amount = 10; /* This is configurable amount */
5101 u16 wm0_sel_res_b, trans_offset_b, res_blocks;
5103 /* Transition WM are not recommended by HW team for GEN9 */
5104 if (INTEL_GEN(dev_priv) <= 9)
5107 /* Transition WM don't make any sense if ipc is disabled */
5108 if (!dev_priv->ipc_enabled)
5112 if (INTEL_GEN(dev_priv) >= 11)
5115 trans_offset_b = trans_min + trans_amount;
5118 * The spec asks for Selected Result Blocks for wm0 (the real value),
5119 * not Result Blocks (the integer value). Pay attention to the capital
5120 * letters. The value wm_l0->plane_res_b is actually Result Blocks, but
5121 * since Result Blocks is the ceiling of Selected Result Blocks plus 1,
5122 * and since we later will have to get the ceiling of the sum in the
5123 * transition watermarks calculation, we can just pretend Selected
5124 * Result Blocks is Result Blocks minus 1 and it should work for the
5125 * current platforms.
5127 wm0_sel_res_b = wm->wm[0].plane_res_b - 1;
5131 (u16)mul_round_up_u32_fixed16(2, wp->y_tile_minimum);
5132 res_blocks = max(wm0_sel_res_b, trans_y_tile_min) +
5135 res_blocks = wm0_sel_res_b + trans_offset_b;
5137 /* WA BUG:1938466 add one block for non y-tile planes */
5138 if (IS_CNL_REVID(dev_priv, CNL_REVID_A0, CNL_REVID_A0))
5144 * Just assume we can enable the transition watermark. After
5145 * computing the DDB we'll come back and disable it if that
5146 * assumption turns out to be false.
5148 wm->trans_wm.plane_res_b = res_blocks + 1;
5149 wm->trans_wm.plane_en = true;
5152 static int skl_build_plane_wm_single(struct intel_crtc_state *crtc_state,
5153 const struct intel_plane_state *plane_state,
5154 enum plane_id plane_id, int color_plane)
5156 struct skl_plane_wm *wm = &crtc_state->wm.skl.optimal.planes[plane_id];
5157 struct skl_wm_params wm_params;
5160 ret = skl_compute_plane_wm_params(crtc_state, plane_state,
5161 &wm_params, color_plane);
5165 skl_compute_wm_levels(crtc_state, &wm_params, wm->wm);
5166 skl_compute_transition_wm(crtc_state, &wm_params, wm);
5171 static int skl_build_plane_wm_uv(struct intel_crtc_state *crtc_state,
5172 const struct intel_plane_state *plane_state,
5173 enum plane_id plane_id)
5175 struct skl_plane_wm *wm = &crtc_state->wm.skl.optimal.planes[plane_id];
5176 struct skl_wm_params wm_params;
5179 wm->is_planar = true;
5181 /* uv plane watermarks must also be validated for NV12/Planar */
5182 ret = skl_compute_plane_wm_params(crtc_state, plane_state,
5187 skl_compute_wm_levels(crtc_state, &wm_params, wm->uv_wm);
5192 static int skl_build_plane_wm(struct intel_crtc_state *crtc_state,
5193 const struct intel_plane_state *plane_state)
5195 struct intel_plane *plane = to_intel_plane(plane_state->uapi.plane);
5196 const struct drm_framebuffer *fb = plane_state->hw.fb;
5197 enum plane_id plane_id = plane->id;
5200 if (!intel_wm_plane_visible(crtc_state, plane_state))
5203 ret = skl_build_plane_wm_single(crtc_state, plane_state,
5208 if (fb->format->is_yuv && fb->format->num_planes > 1) {
5209 ret = skl_build_plane_wm_uv(crtc_state, plane_state,
5218 static int icl_build_plane_wm(struct intel_crtc_state *crtc_state,
5219 const struct intel_plane_state *plane_state)
5221 enum plane_id plane_id = to_intel_plane(plane_state->uapi.plane)->id;
5224 /* Watermarks calculated in master */
5225 if (plane_state->planar_slave)
5228 if (plane_state->planar_linked_plane) {
5229 const struct drm_framebuffer *fb = plane_state->hw.fb;
5230 enum plane_id y_plane_id = plane_state->planar_linked_plane->id;
5232 WARN_ON(!intel_wm_plane_visible(crtc_state, plane_state));
5233 WARN_ON(!fb->format->is_yuv ||
5234 fb->format->num_planes == 1);
5236 ret = skl_build_plane_wm_single(crtc_state, plane_state,
5241 ret = skl_build_plane_wm_single(crtc_state, plane_state,
5245 } else if (intel_wm_plane_visible(crtc_state, plane_state)) {
5246 ret = skl_build_plane_wm_single(crtc_state, plane_state,
5255 static int skl_build_pipe_wm(struct intel_crtc_state *crtc_state)
5257 struct drm_i915_private *dev_priv = to_i915(crtc_state->uapi.crtc->dev);
5258 struct skl_pipe_wm *pipe_wm = &crtc_state->wm.skl.optimal;
5259 struct intel_plane *plane;
5260 const struct intel_plane_state *plane_state;
5264 * We'll only calculate watermarks for planes that are actually
5265 * enabled, so make sure all other planes are set as disabled.
5267 memset(pipe_wm->planes, 0, sizeof(pipe_wm->planes));
5269 intel_atomic_crtc_state_for_each_plane_state(plane, plane_state,
5272 if (INTEL_GEN(dev_priv) >= 11)
5273 ret = icl_build_plane_wm(crtc_state, plane_state);
5275 ret = skl_build_plane_wm(crtc_state, plane_state);
5283 static void skl_ddb_entry_write(struct drm_i915_private *dev_priv,
5285 const struct skl_ddb_entry *entry)
5288 intel_de_write_fw(dev_priv, reg,
5289 (entry->end - 1) << 16 | entry->start);
5291 intel_de_write_fw(dev_priv, reg, 0);
5294 static void skl_write_wm_level(struct drm_i915_private *dev_priv,
5296 const struct skl_wm_level *level)
5300 if (level->plane_en)
5302 if (level->ignore_lines)
5303 val |= PLANE_WM_IGNORE_LINES;
5304 val |= level->plane_res_b;
5305 val |= level->plane_res_l << PLANE_WM_LINES_SHIFT;
5307 intel_de_write_fw(dev_priv, reg, val);
5310 void skl_write_plane_wm(struct intel_plane *plane,
5311 const struct intel_crtc_state *crtc_state)
5313 struct drm_i915_private *dev_priv = to_i915(plane->base.dev);
5314 int level, max_level = ilk_wm_max_level(dev_priv);
5315 enum plane_id plane_id = plane->id;
5316 enum pipe pipe = plane->pipe;
5317 const struct skl_plane_wm *wm =
5318 &crtc_state->wm.skl.optimal.planes[plane_id];
5319 const struct skl_ddb_entry *ddb_y =
5320 &crtc_state->wm.skl.plane_ddb_y[plane_id];
5321 const struct skl_ddb_entry *ddb_uv =
5322 &crtc_state->wm.skl.plane_ddb_uv[plane_id];
5324 for (level = 0; level <= max_level; level++) {
5325 skl_write_wm_level(dev_priv, PLANE_WM(pipe, plane_id, level),
5328 skl_write_wm_level(dev_priv, PLANE_WM_TRANS(pipe, plane_id),
5331 if (INTEL_GEN(dev_priv) >= 11) {
5332 skl_ddb_entry_write(dev_priv,
5333 PLANE_BUF_CFG(pipe, plane_id), ddb_y);
5338 swap(ddb_y, ddb_uv);
5340 skl_ddb_entry_write(dev_priv,
5341 PLANE_BUF_CFG(pipe, plane_id), ddb_y);
5342 skl_ddb_entry_write(dev_priv,
5343 PLANE_NV12_BUF_CFG(pipe, plane_id), ddb_uv);
5346 void skl_write_cursor_wm(struct intel_plane *plane,
5347 const struct intel_crtc_state *crtc_state)
5349 struct drm_i915_private *dev_priv = to_i915(plane->base.dev);
5350 int level, max_level = ilk_wm_max_level(dev_priv);
5351 enum plane_id plane_id = plane->id;
5352 enum pipe pipe = plane->pipe;
5353 const struct skl_plane_wm *wm =
5354 &crtc_state->wm.skl.optimal.planes[plane_id];
5355 const struct skl_ddb_entry *ddb =
5356 &crtc_state->wm.skl.plane_ddb_y[plane_id];
5358 for (level = 0; level <= max_level; level++) {
5359 skl_write_wm_level(dev_priv, CUR_WM(pipe, level),
5362 skl_write_wm_level(dev_priv, CUR_WM_TRANS(pipe), &wm->trans_wm);
5364 skl_ddb_entry_write(dev_priv, CUR_BUF_CFG(pipe), ddb);
5367 bool skl_wm_level_equals(const struct skl_wm_level *l1,
5368 const struct skl_wm_level *l2)
5370 return l1->plane_en == l2->plane_en &&
5371 l1->ignore_lines == l2->ignore_lines &&
5372 l1->plane_res_l == l2->plane_res_l &&
5373 l1->plane_res_b == l2->plane_res_b;
5376 static bool skl_plane_wm_equals(struct drm_i915_private *dev_priv,
5377 const struct skl_plane_wm *wm1,
5378 const struct skl_plane_wm *wm2)
5380 int level, max_level = ilk_wm_max_level(dev_priv);
5382 for (level = 0; level <= max_level; level++) {
5383 if (!skl_wm_level_equals(&wm1->wm[level], &wm2->wm[level]) ||
5384 !skl_wm_level_equals(&wm1->uv_wm[level], &wm2->uv_wm[level]))
5388 return skl_wm_level_equals(&wm1->trans_wm, &wm2->trans_wm);
5391 static inline bool skl_ddb_entries_overlap(const struct skl_ddb_entry *a,
5392 const struct skl_ddb_entry *b)
5394 return a->start < b->end && b->start < a->end;
5397 bool skl_ddb_allocation_overlaps(const struct skl_ddb_entry *ddb,
5398 const struct skl_ddb_entry *entries,
5399 int num_entries, int ignore_idx)
5403 for (i = 0; i < num_entries; i++) {
5404 if (i != ignore_idx &&
5405 skl_ddb_entries_overlap(ddb, &entries[i]))
5413 skl_ddb_add_affected_planes(const struct intel_crtc_state *old_crtc_state,
5414 struct intel_crtc_state *new_crtc_state)
5416 struct intel_atomic_state *state = to_intel_atomic_state(new_crtc_state->uapi.state);
5417 struct intel_crtc *crtc = to_intel_crtc(new_crtc_state->uapi.crtc);
5418 struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
5419 struct intel_plane *plane;
5421 for_each_intel_plane_on_crtc(&dev_priv->drm, crtc, plane) {
5422 struct intel_plane_state *plane_state;
5423 enum plane_id plane_id = plane->id;
5425 if (skl_ddb_entry_equal(&old_crtc_state->wm.skl.plane_ddb_y[plane_id],
5426 &new_crtc_state->wm.skl.plane_ddb_y[plane_id]) &&
5427 skl_ddb_entry_equal(&old_crtc_state->wm.skl.plane_ddb_uv[plane_id],
5428 &new_crtc_state->wm.skl.plane_ddb_uv[plane_id]))
5431 plane_state = intel_atomic_get_plane_state(state, plane);
5432 if (IS_ERR(plane_state))
5433 return PTR_ERR(plane_state);
5435 new_crtc_state->update_planes |= BIT(plane_id);
5442 skl_compute_ddb(struct intel_atomic_state *state)
5444 struct drm_i915_private *dev_priv = to_i915(state->base.dev);
5445 struct intel_crtc_state *old_crtc_state;
5446 struct intel_crtc_state *new_crtc_state;
5447 struct intel_crtc *crtc;
5450 state->enabled_dbuf_slices_mask = dev_priv->enabled_dbuf_slices_mask;
5452 for_each_oldnew_intel_crtc_in_state(state, crtc, old_crtc_state,
5453 new_crtc_state, i) {
5454 ret = skl_allocate_pipe_ddb(new_crtc_state);
5458 ret = skl_ddb_add_affected_planes(old_crtc_state,
5467 static char enast(bool enable)
5469 return enable ? '*' : ' ';
5473 skl_print_wm_changes(struct intel_atomic_state *state)
5475 struct drm_i915_private *dev_priv = to_i915(state->base.dev);
5476 const struct intel_crtc_state *old_crtc_state;
5477 const struct intel_crtc_state *new_crtc_state;
5478 struct intel_plane *plane;
5479 struct intel_crtc *crtc;
5482 if (!drm_debug_enabled(DRM_UT_KMS))
5485 for_each_oldnew_intel_crtc_in_state(state, crtc, old_crtc_state,
5486 new_crtc_state, i) {
5487 const struct skl_pipe_wm *old_pipe_wm, *new_pipe_wm;
5489 old_pipe_wm = &old_crtc_state->wm.skl.optimal;
5490 new_pipe_wm = &new_crtc_state->wm.skl.optimal;
5492 for_each_intel_plane_on_crtc(&dev_priv->drm, crtc, plane) {
5493 enum plane_id plane_id = plane->id;
5494 const struct skl_ddb_entry *old, *new;
5496 old = &old_crtc_state->wm.skl.plane_ddb_y[plane_id];
5497 new = &new_crtc_state->wm.skl.plane_ddb_y[plane_id];
5499 if (skl_ddb_entry_equal(old, new))
5502 drm_dbg_kms(&dev_priv->drm,
5503 "[PLANE:%d:%s] ddb (%4d - %4d) -> (%4d - %4d), size %4d -> %4d\n",
5504 plane->base.base.id, plane->base.name,
5505 old->start, old->end, new->start, new->end,
5506 skl_ddb_entry_size(old), skl_ddb_entry_size(new));
5509 for_each_intel_plane_on_crtc(&dev_priv->drm, crtc, plane) {
5510 enum plane_id plane_id = plane->id;
5511 const struct skl_plane_wm *old_wm, *new_wm;
5513 old_wm = &old_pipe_wm->planes[plane_id];
5514 new_wm = &new_pipe_wm->planes[plane_id];
5516 if (skl_plane_wm_equals(dev_priv, old_wm, new_wm))
5519 drm_dbg_kms(&dev_priv->drm,
5520 "[PLANE:%d:%s] level %cwm0,%cwm1,%cwm2,%cwm3,%cwm4,%cwm5,%cwm6,%cwm7,%ctwm"
5521 " -> %cwm0,%cwm1,%cwm2,%cwm3,%cwm4,%cwm5,%cwm6,%cwm7,%ctwm\n",
5522 plane->base.base.id, plane->base.name,
5523 enast(old_wm->wm[0].plane_en), enast(old_wm->wm[1].plane_en),
5524 enast(old_wm->wm[2].plane_en), enast(old_wm->wm[3].plane_en),
5525 enast(old_wm->wm[4].plane_en), enast(old_wm->wm[5].plane_en),
5526 enast(old_wm->wm[6].plane_en), enast(old_wm->wm[7].plane_en),
5527 enast(old_wm->trans_wm.plane_en),
5528 enast(new_wm->wm[0].plane_en), enast(new_wm->wm[1].plane_en),
5529 enast(new_wm->wm[2].plane_en), enast(new_wm->wm[3].plane_en),
5530 enast(new_wm->wm[4].plane_en), enast(new_wm->wm[5].plane_en),
5531 enast(new_wm->wm[6].plane_en), enast(new_wm->wm[7].plane_en),
5532 enast(new_wm->trans_wm.plane_en));
5534 drm_dbg_kms(&dev_priv->drm,
5535 "[PLANE:%d:%s] lines %c%3d,%c%3d,%c%3d,%c%3d,%c%3d,%c%3d,%c%3d,%c%3d,%c%3d"
5536 " -> %c%3d,%c%3d,%c%3d,%c%3d,%c%3d,%c%3d,%c%3d,%c%3d,%c%3d\n",
5537 plane->base.base.id, plane->base.name,
5538 enast(old_wm->wm[0].ignore_lines), old_wm->wm[0].plane_res_l,
5539 enast(old_wm->wm[1].ignore_lines), old_wm->wm[1].plane_res_l,
5540 enast(old_wm->wm[2].ignore_lines), old_wm->wm[2].plane_res_l,
5541 enast(old_wm->wm[3].ignore_lines), old_wm->wm[3].plane_res_l,
5542 enast(old_wm->wm[4].ignore_lines), old_wm->wm[4].plane_res_l,
5543 enast(old_wm->wm[5].ignore_lines), old_wm->wm[5].plane_res_l,
5544 enast(old_wm->wm[6].ignore_lines), old_wm->wm[6].plane_res_l,
5545 enast(old_wm->wm[7].ignore_lines), old_wm->wm[7].plane_res_l,
5546 enast(old_wm->trans_wm.ignore_lines), old_wm->trans_wm.plane_res_l,
5548 enast(new_wm->wm[0].ignore_lines), new_wm->wm[0].plane_res_l,
5549 enast(new_wm->wm[1].ignore_lines), new_wm->wm[1].plane_res_l,
5550 enast(new_wm->wm[2].ignore_lines), new_wm->wm[2].plane_res_l,
5551 enast(new_wm->wm[3].ignore_lines), new_wm->wm[3].plane_res_l,
5552 enast(new_wm->wm[4].ignore_lines), new_wm->wm[4].plane_res_l,
5553 enast(new_wm->wm[5].ignore_lines), new_wm->wm[5].plane_res_l,
5554 enast(new_wm->wm[6].ignore_lines), new_wm->wm[6].plane_res_l,
5555 enast(new_wm->wm[7].ignore_lines), new_wm->wm[7].plane_res_l,
5556 enast(new_wm->trans_wm.ignore_lines), new_wm->trans_wm.plane_res_l);
5558 drm_dbg_kms(&dev_priv->drm,
5559 "[PLANE:%d:%s] blocks %4d,%4d,%4d,%4d,%4d,%4d,%4d,%4d,%4d"
5560 " -> %4d,%4d,%4d,%4d,%4d,%4d,%4d,%4d,%4d\n",
5561 plane->base.base.id, plane->base.name,
5562 old_wm->wm[0].plane_res_b, old_wm->wm[1].plane_res_b,
5563 old_wm->wm[2].plane_res_b, old_wm->wm[3].plane_res_b,
5564 old_wm->wm[4].plane_res_b, old_wm->wm[5].plane_res_b,
5565 old_wm->wm[6].plane_res_b, old_wm->wm[7].plane_res_b,
5566 old_wm->trans_wm.plane_res_b,
5567 new_wm->wm[0].plane_res_b, new_wm->wm[1].plane_res_b,
5568 new_wm->wm[2].plane_res_b, new_wm->wm[3].plane_res_b,
5569 new_wm->wm[4].plane_res_b, new_wm->wm[5].plane_res_b,
5570 new_wm->wm[6].plane_res_b, new_wm->wm[7].plane_res_b,
5571 new_wm->trans_wm.plane_res_b);
5573 drm_dbg_kms(&dev_priv->drm,
5574 "[PLANE:%d:%s] min_ddb %4d,%4d,%4d,%4d,%4d,%4d,%4d,%4d,%4d"
5575 " -> %4d,%4d,%4d,%4d,%4d,%4d,%4d,%4d,%4d\n",
5576 plane->base.base.id, plane->base.name,
5577 old_wm->wm[0].min_ddb_alloc, old_wm->wm[1].min_ddb_alloc,
5578 old_wm->wm[2].min_ddb_alloc, old_wm->wm[3].min_ddb_alloc,
5579 old_wm->wm[4].min_ddb_alloc, old_wm->wm[5].min_ddb_alloc,
5580 old_wm->wm[6].min_ddb_alloc, old_wm->wm[7].min_ddb_alloc,
5581 old_wm->trans_wm.min_ddb_alloc,
5582 new_wm->wm[0].min_ddb_alloc, new_wm->wm[1].min_ddb_alloc,
5583 new_wm->wm[2].min_ddb_alloc, new_wm->wm[3].min_ddb_alloc,
5584 new_wm->wm[4].min_ddb_alloc, new_wm->wm[5].min_ddb_alloc,
5585 new_wm->wm[6].min_ddb_alloc, new_wm->wm[7].min_ddb_alloc,
5586 new_wm->trans_wm.min_ddb_alloc);
5591 static int intel_add_all_pipes(struct intel_atomic_state *state)
5593 struct drm_i915_private *dev_priv = to_i915(state->base.dev);
5594 struct intel_crtc *crtc;
5596 for_each_intel_crtc(&dev_priv->drm, crtc) {
5597 struct intel_crtc_state *crtc_state;
5599 crtc_state = intel_atomic_get_crtc_state(&state->base, crtc);
5600 if (IS_ERR(crtc_state))
5601 return PTR_ERR(crtc_state);
5608 skl_ddb_add_affected_pipes(struct intel_atomic_state *state)
5610 struct drm_i915_private *dev_priv = to_i915(state->base.dev);
5614 * If this is our first atomic update following hardware readout,
5615 * we can't trust the DDB that the BIOS programmed for us. Let's
5616 * pretend that all pipes switched active status so that we'll
5617 * ensure a full DDB recompute.
5619 if (dev_priv->wm.distrust_bios_wm) {
5620 ret = drm_modeset_lock(&dev_priv->drm.mode_config.connection_mutex,
5621 state->base.acquire_ctx);
5625 state->active_pipe_changes = INTEL_INFO(dev_priv)->pipe_mask;
5628 * We usually only initialize state->active_pipes if we
5629 * we're doing a modeset; make sure this field is always
5630 * initialized during the sanitization process that happens
5631 * on the first commit too.
5633 if (!state->modeset)
5634 state->active_pipes = dev_priv->active_pipes;
5638 * If the modeset changes which CRTC's are active, we need to
5639 * recompute the DDB allocation for *all* active pipes, even
5640 * those that weren't otherwise being modified in any way by this
5641 * atomic commit. Due to the shrinking of the per-pipe allocations
5642 * when new active CRTC's are added, it's possible for a pipe that
5643 * we were already using and aren't changing at all here to suddenly
5644 * become invalid if its DDB needs exceeds its new allocation.
5646 * Note that if we wind up doing a full DDB recompute, we can't let
5647 * any other display updates race with this transaction, so we need
5648 * to grab the lock on *all* CRTC's.
5650 if (state->active_pipe_changes || state->modeset) {
5651 ret = intel_add_all_pipes(state);
5660 * To make sure the cursor watermark registers are always consistent
5661 * with our computed state the following scenario needs special
5665 * 2. move cursor entirely offscreen
5668 * Step 2. does call .disable_plane() but does not zero the watermarks
5669 * (since we consider an offscreen cursor still active for the purposes
5670 * of watermarks). Step 3. would not normally call .disable_plane()
5671 * because the actual plane visibility isn't changing, and we don't
5672 * deallocate the cursor ddb until the pipe gets disabled. So we must
5673 * force step 3. to call .disable_plane() to update the watermark
5674 * registers properly.
5676 * Other planes do not suffer from this issues as their watermarks are
5677 * calculated based on the actual plane visibility. The only time this
5678 * can trigger for the other planes is during the initial readout as the
5679 * default value of the watermarks registers is not zero.
5681 static int skl_wm_add_affected_planes(struct intel_atomic_state *state,
5682 struct intel_crtc *crtc)
5684 struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
5685 const struct intel_crtc_state *old_crtc_state =
5686 intel_atomic_get_old_crtc_state(state, crtc);
5687 struct intel_crtc_state *new_crtc_state =
5688 intel_atomic_get_new_crtc_state(state, crtc);
5689 struct intel_plane *plane;
5691 for_each_intel_plane_on_crtc(&dev_priv->drm, crtc, plane) {
5692 struct intel_plane_state *plane_state;
5693 enum plane_id plane_id = plane->id;
5696 * Force a full wm update for every plane on modeset.
5697 * Required because the reset value of the wm registers
5698 * is non-zero, whereas we want all disabled planes to
5699 * have zero watermarks. So if we turn off the relevant
5700 * power well the hardware state will go out of sync
5701 * with the software state.
5703 if (!drm_atomic_crtc_needs_modeset(&new_crtc_state->uapi) &&
5704 skl_plane_wm_equals(dev_priv,
5705 &old_crtc_state->wm.skl.optimal.planes[plane_id],
5706 &new_crtc_state->wm.skl.optimal.planes[plane_id]))
5709 plane_state = intel_atomic_get_plane_state(state, plane);
5710 if (IS_ERR(plane_state))
5711 return PTR_ERR(plane_state);
5713 new_crtc_state->update_planes |= BIT(plane_id);
5720 skl_compute_wm(struct intel_atomic_state *state)
5722 struct intel_crtc *crtc;
5723 struct intel_crtc_state *new_crtc_state;
5724 struct intel_crtc_state *old_crtc_state;
5727 ret = skl_ddb_add_affected_pipes(state);
5732 * Calculate WM's for all pipes that are part of this transaction.
5733 * Note that skl_ddb_add_affected_pipes may have added more CRTC's that
5734 * weren't otherwise being modified if pipe allocations had to change.
5736 for_each_oldnew_intel_crtc_in_state(state, crtc, old_crtc_state,
5737 new_crtc_state, i) {
5738 ret = skl_build_pipe_wm(new_crtc_state);
5742 ret = skl_wm_add_affected_planes(state, crtc);
5747 ret = skl_compute_ddb(state);
5751 skl_print_wm_changes(state);
5756 static void ilk_compute_wm_config(struct drm_i915_private *dev_priv,
5757 struct intel_wm_config *config)
5759 struct intel_crtc *crtc;
5761 /* Compute the currently _active_ config */
5762 for_each_intel_crtc(&dev_priv->drm, crtc) {
5763 const struct intel_pipe_wm *wm = &crtc->wm.active.ilk;
5765 if (!wm->pipe_enabled)
5768 config->sprites_enabled |= wm->sprites_enabled;
5769 config->sprites_scaled |= wm->sprites_scaled;
5770 config->num_pipes_active++;
5774 static void ilk_program_watermarks(struct drm_i915_private *dev_priv)
5776 struct intel_pipe_wm lp_wm_1_2 = {}, lp_wm_5_6 = {}, *best_lp_wm;
5777 struct ilk_wm_maximums max;
5778 struct intel_wm_config config = {};
5779 struct ilk_wm_values results = {};
5780 enum intel_ddb_partitioning partitioning;
5782 ilk_compute_wm_config(dev_priv, &config);
5784 ilk_compute_wm_maximums(dev_priv, 1, &config, INTEL_DDB_PART_1_2, &max);
5785 ilk_wm_merge(dev_priv, &config, &max, &lp_wm_1_2);
5787 /* 5/6 split only in single pipe config on IVB+ */
5788 if (INTEL_GEN(dev_priv) >= 7 &&
5789 config.num_pipes_active == 1 && config.sprites_enabled) {
5790 ilk_compute_wm_maximums(dev_priv, 1, &config, INTEL_DDB_PART_5_6, &max);
5791 ilk_wm_merge(dev_priv, &config, &max, &lp_wm_5_6);
5793 best_lp_wm = ilk_find_best_result(dev_priv, &lp_wm_1_2, &lp_wm_5_6);
5795 best_lp_wm = &lp_wm_1_2;
5798 partitioning = (best_lp_wm == &lp_wm_1_2) ?
5799 INTEL_DDB_PART_1_2 : INTEL_DDB_PART_5_6;
5801 ilk_compute_wm_results(dev_priv, best_lp_wm, partitioning, &results);
5803 ilk_write_wm_values(dev_priv, &results);
5806 static void ilk_initial_watermarks(struct intel_atomic_state *state,
5807 struct intel_crtc *crtc)
5809 struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
5810 const struct intel_crtc_state *crtc_state =
5811 intel_atomic_get_new_crtc_state(state, crtc);
5813 mutex_lock(&dev_priv->wm.wm_mutex);
5814 crtc->wm.active.ilk = crtc_state->wm.ilk.intermediate;
5815 ilk_program_watermarks(dev_priv);
5816 mutex_unlock(&dev_priv->wm.wm_mutex);
5819 static void ilk_optimize_watermarks(struct intel_atomic_state *state,
5820 struct intel_crtc *crtc)
5822 struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
5823 const struct intel_crtc_state *crtc_state =
5824 intel_atomic_get_new_crtc_state(state, crtc);
5826 if (!crtc_state->wm.need_postvbl_update)
5829 mutex_lock(&dev_priv->wm.wm_mutex);
5830 crtc->wm.active.ilk = crtc_state->wm.ilk.optimal;
5831 ilk_program_watermarks(dev_priv);
5832 mutex_unlock(&dev_priv->wm.wm_mutex);
5835 static inline void skl_wm_level_from_reg_val(u32 val,
5836 struct skl_wm_level *level)
5838 level->plane_en = val & PLANE_WM_EN;
5839 level->ignore_lines = val & PLANE_WM_IGNORE_LINES;
5840 level->plane_res_b = val & PLANE_WM_BLOCKS_MASK;
5841 level->plane_res_l = (val >> PLANE_WM_LINES_SHIFT) &
5842 PLANE_WM_LINES_MASK;
5845 void skl_pipe_wm_get_hw_state(struct intel_crtc *crtc,
5846 struct skl_pipe_wm *out)
5848 struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
5849 enum pipe pipe = crtc->pipe;
5850 int level, max_level;
5851 enum plane_id plane_id;
5854 max_level = ilk_wm_max_level(dev_priv);
5856 for_each_plane_id_on_crtc(crtc, plane_id) {
5857 struct skl_plane_wm *wm = &out->planes[plane_id];
5859 for (level = 0; level <= max_level; level++) {
5860 if (plane_id != PLANE_CURSOR)
5861 val = I915_READ(PLANE_WM(pipe, plane_id, level));
5863 val = I915_READ(CUR_WM(pipe, level));
5865 skl_wm_level_from_reg_val(val, &wm->wm[level]);
5868 if (plane_id != PLANE_CURSOR)
5869 val = I915_READ(PLANE_WM_TRANS(pipe, plane_id));
5871 val = I915_READ(CUR_WM_TRANS(pipe));
5873 skl_wm_level_from_reg_val(val, &wm->trans_wm);
5880 void skl_wm_get_hw_state(struct drm_i915_private *dev_priv)
5882 struct intel_crtc *crtc;
5883 struct intel_crtc_state *crtc_state;
5885 skl_ddb_get_hw_state(dev_priv);
5886 for_each_intel_crtc(&dev_priv->drm, crtc) {
5887 crtc_state = to_intel_crtc_state(crtc->base.state);
5889 skl_pipe_wm_get_hw_state(crtc, &crtc_state->wm.skl.optimal);
5892 if (dev_priv->active_pipes) {
5893 /* Fully recompute DDB on first atomic commit */
5894 dev_priv->wm.distrust_bios_wm = true;
5898 static void ilk_pipe_wm_get_hw_state(struct intel_crtc *crtc)
5900 struct drm_device *dev = crtc->base.dev;
5901 struct drm_i915_private *dev_priv = to_i915(dev);
5902 struct ilk_wm_values *hw = &dev_priv->wm.hw;
5903 struct intel_crtc_state *crtc_state = to_intel_crtc_state(crtc->base.state);
5904 struct intel_pipe_wm *active = &crtc_state->wm.ilk.optimal;
5905 enum pipe pipe = crtc->pipe;
5906 static const i915_reg_t wm0_pipe_reg[] = {
5907 [PIPE_A] = WM0_PIPEA_ILK,
5908 [PIPE_B] = WM0_PIPEB_ILK,
5909 [PIPE_C] = WM0_PIPEC_IVB,
5912 hw->wm_pipe[pipe] = I915_READ(wm0_pipe_reg[pipe]);
5914 memset(active, 0, sizeof(*active));
5916 active->pipe_enabled = crtc->active;
5918 if (active->pipe_enabled) {
5919 u32 tmp = hw->wm_pipe[pipe];
5922 * For active pipes LP0 watermark is marked as
5923 * enabled, and LP1+ watermaks as disabled since
5924 * we can't really reverse compute them in case
5925 * multiple pipes are active.
5927 active->wm[0].enable = true;
5928 active->wm[0].pri_val = (tmp & WM0_PIPE_PLANE_MASK) >> WM0_PIPE_PLANE_SHIFT;
5929 active->wm[0].spr_val = (tmp & WM0_PIPE_SPRITE_MASK) >> WM0_PIPE_SPRITE_SHIFT;
5930 active->wm[0].cur_val = tmp & WM0_PIPE_CURSOR_MASK;
5932 int level, max_level = ilk_wm_max_level(dev_priv);
5935 * For inactive pipes, all watermark levels
5936 * should be marked as enabled but zeroed,
5937 * which is what we'd compute them to.
5939 for (level = 0; level <= max_level; level++)
5940 active->wm[level].enable = true;
5943 crtc->wm.active.ilk = *active;
5946 #define _FW_WM(value, plane) \
5947 (((value) & DSPFW_ ## plane ## _MASK) >> DSPFW_ ## plane ## _SHIFT)
5948 #define _FW_WM_VLV(value, plane) \
5949 (((value) & DSPFW_ ## plane ## _MASK_VLV) >> DSPFW_ ## plane ## _SHIFT)
5951 static void g4x_read_wm_values(struct drm_i915_private *dev_priv,
5952 struct g4x_wm_values *wm)
5956 tmp = I915_READ(DSPFW1);
5957 wm->sr.plane = _FW_WM(tmp, SR);
5958 wm->pipe[PIPE_B].plane[PLANE_CURSOR] = _FW_WM(tmp, CURSORB);
5959 wm->pipe[PIPE_B].plane[PLANE_PRIMARY] = _FW_WM(tmp, PLANEB);
5960 wm->pipe[PIPE_A].plane[PLANE_PRIMARY] = _FW_WM(tmp, PLANEA);
5962 tmp = I915_READ(DSPFW2);
5963 wm->fbc_en = tmp & DSPFW_FBC_SR_EN;
5964 wm->sr.fbc = _FW_WM(tmp, FBC_SR);
5965 wm->hpll.fbc = _FW_WM(tmp, FBC_HPLL_SR);
5966 wm->pipe[PIPE_B].plane[PLANE_SPRITE0] = _FW_WM(tmp, SPRITEB);
5967 wm->pipe[PIPE_A].plane[PLANE_CURSOR] = _FW_WM(tmp, CURSORA);
5968 wm->pipe[PIPE_A].plane[PLANE_SPRITE0] = _FW_WM(tmp, SPRITEA);
5970 tmp = I915_READ(DSPFW3);
5971 wm->hpll_en = tmp & DSPFW_HPLL_SR_EN;
5972 wm->sr.cursor = _FW_WM(tmp, CURSOR_SR);
5973 wm->hpll.cursor = _FW_WM(tmp, HPLL_CURSOR);
5974 wm->hpll.plane = _FW_WM(tmp, HPLL_SR);
5977 static void vlv_read_wm_values(struct drm_i915_private *dev_priv,
5978 struct vlv_wm_values *wm)
5983 for_each_pipe(dev_priv, pipe) {
5984 tmp = I915_READ(VLV_DDL(pipe));
5986 wm->ddl[pipe].plane[PLANE_PRIMARY] =
5987 (tmp >> DDL_PLANE_SHIFT) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
5988 wm->ddl[pipe].plane[PLANE_CURSOR] =
5989 (tmp >> DDL_CURSOR_SHIFT) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
5990 wm->ddl[pipe].plane[PLANE_SPRITE0] =
5991 (tmp >> DDL_SPRITE_SHIFT(0)) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
5992 wm->ddl[pipe].plane[PLANE_SPRITE1] =
5993 (tmp >> DDL_SPRITE_SHIFT(1)) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
5996 tmp = I915_READ(DSPFW1);
5997 wm->sr.plane = _FW_WM(tmp, SR);
5998 wm->pipe[PIPE_B].plane[PLANE_CURSOR] = _FW_WM(tmp, CURSORB);
5999 wm->pipe[PIPE_B].plane[PLANE_PRIMARY] = _FW_WM_VLV(tmp, PLANEB);
6000 wm->pipe[PIPE_A].plane[PLANE_PRIMARY] = _FW_WM_VLV(tmp, PLANEA);
6002 tmp = I915_READ(DSPFW2);
6003 wm->pipe[PIPE_A].plane[PLANE_SPRITE1] = _FW_WM_VLV(tmp, SPRITEB);
6004 wm->pipe[PIPE_A].plane[PLANE_CURSOR] = _FW_WM(tmp, CURSORA);
6005 wm->pipe[PIPE_A].plane[PLANE_SPRITE0] = _FW_WM_VLV(tmp, SPRITEA);
6007 tmp = I915_READ(DSPFW3);
6008 wm->sr.cursor = _FW_WM(tmp, CURSOR_SR);
6010 if (IS_CHERRYVIEW(dev_priv)) {
6011 tmp = I915_READ(DSPFW7_CHV);
6012 wm->pipe[PIPE_B].plane[PLANE_SPRITE1] = _FW_WM_VLV(tmp, SPRITED);
6013 wm->pipe[PIPE_B].plane[PLANE_SPRITE0] = _FW_WM_VLV(tmp, SPRITEC);
6015 tmp = I915_READ(DSPFW8_CHV);
6016 wm->pipe[PIPE_C].plane[PLANE_SPRITE1] = _FW_WM_VLV(tmp, SPRITEF);
6017 wm->pipe[PIPE_C].plane[PLANE_SPRITE0] = _FW_WM_VLV(tmp, SPRITEE);
6019 tmp = I915_READ(DSPFW9_CHV);
6020 wm->pipe[PIPE_C].plane[PLANE_PRIMARY] = _FW_WM_VLV(tmp, PLANEC);
6021 wm->pipe[PIPE_C].plane[PLANE_CURSOR] = _FW_WM(tmp, CURSORC);
6023 tmp = I915_READ(DSPHOWM);
6024 wm->sr.plane |= _FW_WM(tmp, SR_HI) << 9;
6025 wm->pipe[PIPE_C].plane[PLANE_SPRITE1] |= _FW_WM(tmp, SPRITEF_HI) << 8;
6026 wm->pipe[PIPE_C].plane[PLANE_SPRITE0] |= _FW_WM(tmp, SPRITEE_HI) << 8;
6027 wm->pipe[PIPE_C].plane[PLANE_PRIMARY] |= _FW_WM(tmp, PLANEC_HI) << 8;
6028 wm->pipe[PIPE_B].plane[PLANE_SPRITE1] |= _FW_WM(tmp, SPRITED_HI) << 8;
6029 wm->pipe[PIPE_B].plane[PLANE_SPRITE0] |= _FW_WM(tmp, SPRITEC_HI) << 8;
6030 wm->pipe[PIPE_B].plane[PLANE_PRIMARY] |= _FW_WM(tmp, PLANEB_HI) << 8;
6031 wm->pipe[PIPE_A].plane[PLANE_SPRITE1] |= _FW_WM(tmp, SPRITEB_HI) << 8;
6032 wm->pipe[PIPE_A].plane[PLANE_SPRITE0] |= _FW_WM(tmp, SPRITEA_HI) << 8;
6033 wm->pipe[PIPE_A].plane[PLANE_PRIMARY] |= _FW_WM(tmp, PLANEA_HI) << 8;
6035 tmp = I915_READ(DSPFW7);
6036 wm->pipe[PIPE_B].plane[PLANE_SPRITE1] = _FW_WM_VLV(tmp, SPRITED);
6037 wm->pipe[PIPE_B].plane[PLANE_SPRITE0] = _FW_WM_VLV(tmp, SPRITEC);
6039 tmp = I915_READ(DSPHOWM);
6040 wm->sr.plane |= _FW_WM(tmp, SR_HI) << 9;
6041 wm->pipe[PIPE_B].plane[PLANE_SPRITE1] |= _FW_WM(tmp, SPRITED_HI) << 8;
6042 wm->pipe[PIPE_B].plane[PLANE_SPRITE0] |= _FW_WM(tmp, SPRITEC_HI) << 8;
6043 wm->pipe[PIPE_B].plane[PLANE_PRIMARY] |= _FW_WM(tmp, PLANEB_HI) << 8;
6044 wm->pipe[PIPE_A].plane[PLANE_SPRITE1] |= _FW_WM(tmp, SPRITEB_HI) << 8;
6045 wm->pipe[PIPE_A].plane[PLANE_SPRITE0] |= _FW_WM(tmp, SPRITEA_HI) << 8;
6046 wm->pipe[PIPE_A].plane[PLANE_PRIMARY] |= _FW_WM(tmp, PLANEA_HI) << 8;
6053 void g4x_wm_get_hw_state(struct drm_i915_private *dev_priv)
6055 struct g4x_wm_values *wm = &dev_priv->wm.g4x;
6056 struct intel_crtc *crtc;
6058 g4x_read_wm_values(dev_priv, wm);
6060 wm->cxsr = I915_READ(FW_BLC_SELF) & FW_BLC_SELF_EN;
6062 for_each_intel_crtc(&dev_priv->drm, crtc) {
6063 struct intel_crtc_state *crtc_state =
6064 to_intel_crtc_state(crtc->base.state);
6065 struct g4x_wm_state *active = &crtc->wm.active.g4x;
6066 struct g4x_pipe_wm *raw;
6067 enum pipe pipe = crtc->pipe;
6068 enum plane_id plane_id;
6069 int level, max_level;
6071 active->cxsr = wm->cxsr;
6072 active->hpll_en = wm->hpll_en;
6073 active->fbc_en = wm->fbc_en;
6075 active->sr = wm->sr;
6076 active->hpll = wm->hpll;
6078 for_each_plane_id_on_crtc(crtc, plane_id) {
6079 active->wm.plane[plane_id] =
6080 wm->pipe[pipe].plane[plane_id];
6083 if (wm->cxsr && wm->hpll_en)
6084 max_level = G4X_WM_LEVEL_HPLL;
6086 max_level = G4X_WM_LEVEL_SR;
6088 max_level = G4X_WM_LEVEL_NORMAL;
6090 level = G4X_WM_LEVEL_NORMAL;
6091 raw = &crtc_state->wm.g4x.raw[level];
6092 for_each_plane_id_on_crtc(crtc, plane_id)
6093 raw->plane[plane_id] = active->wm.plane[plane_id];
6095 if (++level > max_level)
6098 raw = &crtc_state->wm.g4x.raw[level];
6099 raw->plane[PLANE_PRIMARY] = active->sr.plane;
6100 raw->plane[PLANE_CURSOR] = active->sr.cursor;
6101 raw->plane[PLANE_SPRITE0] = 0;
6102 raw->fbc = active->sr.fbc;
6104 if (++level > max_level)
6107 raw = &crtc_state->wm.g4x.raw[level];
6108 raw->plane[PLANE_PRIMARY] = active->hpll.plane;
6109 raw->plane[PLANE_CURSOR] = active->hpll.cursor;
6110 raw->plane[PLANE_SPRITE0] = 0;
6111 raw->fbc = active->hpll.fbc;
6114 for_each_plane_id_on_crtc(crtc, plane_id)
6115 g4x_raw_plane_wm_set(crtc_state, level,
6116 plane_id, USHRT_MAX);
6117 g4x_raw_fbc_wm_set(crtc_state, level, USHRT_MAX);
6119 crtc_state->wm.g4x.optimal = *active;
6120 crtc_state->wm.g4x.intermediate = *active;
6122 drm_dbg_kms(&dev_priv->drm,
6123 "Initial watermarks: pipe %c, plane=%d, cursor=%d, sprite=%d\n",
6125 wm->pipe[pipe].plane[PLANE_PRIMARY],
6126 wm->pipe[pipe].plane[PLANE_CURSOR],
6127 wm->pipe[pipe].plane[PLANE_SPRITE0]);
6130 drm_dbg_kms(&dev_priv->drm,
6131 "Initial SR watermarks: plane=%d, cursor=%d fbc=%d\n",
6132 wm->sr.plane, wm->sr.cursor, wm->sr.fbc);
6133 drm_dbg_kms(&dev_priv->drm,
6134 "Initial HPLL watermarks: plane=%d, SR cursor=%d fbc=%d\n",
6135 wm->hpll.plane, wm->hpll.cursor, wm->hpll.fbc);
6136 drm_dbg_kms(&dev_priv->drm, "Initial SR=%s HPLL=%s FBC=%s\n",
6137 yesno(wm->cxsr), yesno(wm->hpll_en), yesno(wm->fbc_en));
6140 void g4x_wm_sanitize(struct drm_i915_private *dev_priv)
6142 struct intel_plane *plane;
6143 struct intel_crtc *crtc;
6145 mutex_lock(&dev_priv->wm.wm_mutex);
6147 for_each_intel_plane(&dev_priv->drm, plane) {
6148 struct intel_crtc *crtc =
6149 intel_get_crtc_for_pipe(dev_priv, plane->pipe);
6150 struct intel_crtc_state *crtc_state =
6151 to_intel_crtc_state(crtc->base.state);
6152 struct intel_plane_state *plane_state =
6153 to_intel_plane_state(plane->base.state);
6154 struct g4x_wm_state *wm_state = &crtc_state->wm.g4x.optimal;
6155 enum plane_id plane_id = plane->id;
6158 if (plane_state->uapi.visible)
6161 for (level = 0; level < 3; level++) {
6162 struct g4x_pipe_wm *raw =
6163 &crtc_state->wm.g4x.raw[level];
6165 raw->plane[plane_id] = 0;
6166 wm_state->wm.plane[plane_id] = 0;
6169 if (plane_id == PLANE_PRIMARY) {
6170 for (level = 0; level < 3; level++) {
6171 struct g4x_pipe_wm *raw =
6172 &crtc_state->wm.g4x.raw[level];
6176 wm_state->sr.fbc = 0;
6177 wm_state->hpll.fbc = 0;
6178 wm_state->fbc_en = false;
6182 for_each_intel_crtc(&dev_priv->drm, crtc) {
6183 struct intel_crtc_state *crtc_state =
6184 to_intel_crtc_state(crtc->base.state);
6186 crtc_state->wm.g4x.intermediate =
6187 crtc_state->wm.g4x.optimal;
6188 crtc->wm.active.g4x = crtc_state->wm.g4x.optimal;
6191 g4x_program_watermarks(dev_priv);
6193 mutex_unlock(&dev_priv->wm.wm_mutex);
6196 void vlv_wm_get_hw_state(struct drm_i915_private *dev_priv)
6198 struct vlv_wm_values *wm = &dev_priv->wm.vlv;
6199 struct intel_crtc *crtc;
6202 vlv_read_wm_values(dev_priv, wm);
6204 wm->cxsr = I915_READ(FW_BLC_SELF_VLV) & FW_CSPWRDWNEN;
6205 wm->level = VLV_WM_LEVEL_PM2;
6207 if (IS_CHERRYVIEW(dev_priv)) {
6208 vlv_punit_get(dev_priv);
6210 val = vlv_punit_read(dev_priv, PUNIT_REG_DSPSSPM);
6211 if (val & DSP_MAXFIFO_PM5_ENABLE)
6212 wm->level = VLV_WM_LEVEL_PM5;
6215 * If DDR DVFS is disabled in the BIOS, Punit
6216 * will never ack the request. So if that happens
6217 * assume we don't have to enable/disable DDR DVFS
6218 * dynamically. To test that just set the REQ_ACK
6219 * bit to poke the Punit, but don't change the
6220 * HIGH/LOW bits so that we don't actually change
6221 * the current state.
6223 val = vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2);
6224 val |= FORCE_DDR_FREQ_REQ_ACK;
6225 vlv_punit_write(dev_priv, PUNIT_REG_DDR_SETUP2, val);
6227 if (wait_for((vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2) &
6228 FORCE_DDR_FREQ_REQ_ACK) == 0, 3)) {
6229 drm_dbg_kms(&dev_priv->drm,
6230 "Punit not acking DDR DVFS request, "
6231 "assuming DDR DVFS is disabled\n");
6232 dev_priv->wm.max_level = VLV_WM_LEVEL_PM5;
6234 val = vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2);
6235 if ((val & FORCE_DDR_HIGH_FREQ) == 0)
6236 wm->level = VLV_WM_LEVEL_DDR_DVFS;
6239 vlv_punit_put(dev_priv);
6242 for_each_intel_crtc(&dev_priv->drm, crtc) {
6243 struct intel_crtc_state *crtc_state =
6244 to_intel_crtc_state(crtc->base.state);
6245 struct vlv_wm_state *active = &crtc->wm.active.vlv;
6246 const struct vlv_fifo_state *fifo_state =
6247 &crtc_state->wm.vlv.fifo_state;
6248 enum pipe pipe = crtc->pipe;
6249 enum plane_id plane_id;
6252 vlv_get_fifo_size(crtc_state);
6254 active->num_levels = wm->level + 1;
6255 active->cxsr = wm->cxsr;
6257 for (level = 0; level < active->num_levels; level++) {
6258 struct g4x_pipe_wm *raw =
6259 &crtc_state->wm.vlv.raw[level];
6261 active->sr[level].plane = wm->sr.plane;
6262 active->sr[level].cursor = wm->sr.cursor;
6264 for_each_plane_id_on_crtc(crtc, plane_id) {
6265 active->wm[level].plane[plane_id] =
6266 wm->pipe[pipe].plane[plane_id];
6268 raw->plane[plane_id] =
6269 vlv_invert_wm_value(active->wm[level].plane[plane_id],
6270 fifo_state->plane[plane_id]);
6274 for_each_plane_id_on_crtc(crtc, plane_id)
6275 vlv_raw_plane_wm_set(crtc_state, level,
6276 plane_id, USHRT_MAX);
6277 vlv_invalidate_wms(crtc, active, level);
6279 crtc_state->wm.vlv.optimal = *active;
6280 crtc_state->wm.vlv.intermediate = *active;
6282 drm_dbg_kms(&dev_priv->drm,
6283 "Initial watermarks: pipe %c, plane=%d, cursor=%d, sprite0=%d, sprite1=%d\n",
6285 wm->pipe[pipe].plane[PLANE_PRIMARY],
6286 wm->pipe[pipe].plane[PLANE_CURSOR],
6287 wm->pipe[pipe].plane[PLANE_SPRITE0],
6288 wm->pipe[pipe].plane[PLANE_SPRITE1]);
6291 drm_dbg_kms(&dev_priv->drm,
6292 "Initial watermarks: SR plane=%d, SR cursor=%d level=%d cxsr=%d\n",
6293 wm->sr.plane, wm->sr.cursor, wm->level, wm->cxsr);
6296 void vlv_wm_sanitize(struct drm_i915_private *dev_priv)
6298 struct intel_plane *plane;
6299 struct intel_crtc *crtc;
6301 mutex_lock(&dev_priv->wm.wm_mutex);
6303 for_each_intel_plane(&dev_priv->drm, plane) {
6304 struct intel_crtc *crtc =
6305 intel_get_crtc_for_pipe(dev_priv, plane->pipe);
6306 struct intel_crtc_state *crtc_state =
6307 to_intel_crtc_state(crtc->base.state);
6308 struct intel_plane_state *plane_state =
6309 to_intel_plane_state(plane->base.state);
6310 struct vlv_wm_state *wm_state = &crtc_state->wm.vlv.optimal;
6311 const struct vlv_fifo_state *fifo_state =
6312 &crtc_state->wm.vlv.fifo_state;
6313 enum plane_id plane_id = plane->id;
6316 if (plane_state->uapi.visible)
6319 for (level = 0; level < wm_state->num_levels; level++) {
6320 struct g4x_pipe_wm *raw =
6321 &crtc_state->wm.vlv.raw[level];
6323 raw->plane[plane_id] = 0;
6325 wm_state->wm[level].plane[plane_id] =
6326 vlv_invert_wm_value(raw->plane[plane_id],
6327 fifo_state->plane[plane_id]);
6331 for_each_intel_crtc(&dev_priv->drm, crtc) {
6332 struct intel_crtc_state *crtc_state =
6333 to_intel_crtc_state(crtc->base.state);
6335 crtc_state->wm.vlv.intermediate =
6336 crtc_state->wm.vlv.optimal;
6337 crtc->wm.active.vlv = crtc_state->wm.vlv.optimal;
6340 vlv_program_watermarks(dev_priv);
6342 mutex_unlock(&dev_priv->wm.wm_mutex);
6346 * FIXME should probably kill this and improve
6347 * the real watermark readout/sanitation instead
6349 static void ilk_init_lp_watermarks(struct drm_i915_private *dev_priv)
6351 I915_WRITE(WM3_LP_ILK, I915_READ(WM3_LP_ILK) & ~WM1_LP_SR_EN);
6352 I915_WRITE(WM2_LP_ILK, I915_READ(WM2_LP_ILK) & ~WM1_LP_SR_EN);
6353 I915_WRITE(WM1_LP_ILK, I915_READ(WM1_LP_ILK) & ~WM1_LP_SR_EN);
6356 * Don't touch WM1S_LP_EN here.
6357 * Doing so could cause underruns.
6361 void ilk_wm_get_hw_state(struct drm_i915_private *dev_priv)
6363 struct ilk_wm_values *hw = &dev_priv->wm.hw;
6364 struct intel_crtc *crtc;
6366 ilk_init_lp_watermarks(dev_priv);
6368 for_each_intel_crtc(&dev_priv->drm, crtc)
6369 ilk_pipe_wm_get_hw_state(crtc);
6371 hw->wm_lp[0] = I915_READ(WM1_LP_ILK);
6372 hw->wm_lp[1] = I915_READ(WM2_LP_ILK);
6373 hw->wm_lp[2] = I915_READ(WM3_LP_ILK);
6375 hw->wm_lp_spr[0] = I915_READ(WM1S_LP_ILK);
6376 if (INTEL_GEN(dev_priv) >= 7) {
6377 hw->wm_lp_spr[1] = I915_READ(WM2S_LP_IVB);
6378 hw->wm_lp_spr[2] = I915_READ(WM3S_LP_IVB);
6381 if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
6382 hw->partitioning = (I915_READ(WM_MISC) & WM_MISC_DATA_PARTITION_5_6) ?
6383 INTEL_DDB_PART_5_6 : INTEL_DDB_PART_1_2;
6384 else if (IS_IVYBRIDGE(dev_priv))
6385 hw->partitioning = (I915_READ(DISP_ARB_CTL2) & DISP_DATA_PARTITION_5_6) ?
6386 INTEL_DDB_PART_5_6 : INTEL_DDB_PART_1_2;
6389 !(I915_READ(DISP_ARB_CTL) & DISP_FBC_WM_DIS);
6393 * intel_update_watermarks - update FIFO watermark values based on current modes
6394 * @crtc: the #intel_crtc on which to compute the WM
6396 * Calculate watermark values for the various WM regs based on current mode
6397 * and plane configuration.
6399 * There are several cases to deal with here:
6400 * - normal (i.e. non-self-refresh)
6401 * - self-refresh (SR) mode
6402 * - lines are large relative to FIFO size (buffer can hold up to 2)
6403 * - lines are small relative to FIFO size (buffer can hold more than 2
6404 * lines), so need to account for TLB latency
6406 * The normal calculation is:
6407 * watermark = dotclock * bytes per pixel * latency
6408 * where latency is platform & configuration dependent (we assume pessimal
6411 * The SR calculation is:
6412 * watermark = (trunc(latency/line time)+1) * surface width *
6415 * line time = htotal / dotclock
6416 * surface width = hdisplay for normal plane and 64 for cursor
6417 * and latency is assumed to be high, as above.
6419 * The final value programmed to the register should always be rounded up,
6420 * and include an extra 2 entries to account for clock crossings.
6422 * We don't use the sprite, so we can ignore that. And on Crestline we have
6423 * to set the non-SR watermarks to 8.
6425 void intel_update_watermarks(struct intel_crtc *crtc)
6427 struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
6429 if (dev_priv->display.update_wm)
6430 dev_priv->display.update_wm(crtc);
6433 void intel_enable_ipc(struct drm_i915_private *dev_priv)
6437 if (!HAS_IPC(dev_priv))
6440 val = I915_READ(DISP_ARB_CTL2);
6442 if (dev_priv->ipc_enabled)
6443 val |= DISP_IPC_ENABLE;
6445 val &= ~DISP_IPC_ENABLE;
6447 I915_WRITE(DISP_ARB_CTL2, val);
6450 static bool intel_can_enable_ipc(struct drm_i915_private *dev_priv)
6452 /* Display WA #0477 WaDisableIPC: skl */
6453 if (IS_SKYLAKE(dev_priv))
6456 /* Display WA #1141: SKL:all KBL:all CFL */
6457 if (IS_KABYLAKE(dev_priv) || IS_COFFEELAKE(dev_priv))
6458 return dev_priv->dram_info.symmetric_memory;
6463 void intel_init_ipc(struct drm_i915_private *dev_priv)
6465 if (!HAS_IPC(dev_priv))
6468 dev_priv->ipc_enabled = intel_can_enable_ipc(dev_priv);
6470 intel_enable_ipc(dev_priv);
6473 static void ibx_init_clock_gating(struct drm_i915_private *dev_priv)
6476 * On Ibex Peak and Cougar Point, we need to disable clock
6477 * gating for the panel power sequencer or it will fail to
6478 * start up when no ports are active.
6480 I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE);
6483 static void g4x_disable_trickle_feed(struct drm_i915_private *dev_priv)
6487 for_each_pipe(dev_priv, pipe) {
6488 I915_WRITE(DSPCNTR(pipe),
6489 I915_READ(DSPCNTR(pipe)) |
6490 DISPPLANE_TRICKLE_FEED_DISABLE);
6492 I915_WRITE(DSPSURF(pipe), I915_READ(DSPSURF(pipe)));
6493 POSTING_READ(DSPSURF(pipe));
6497 static void ilk_init_clock_gating(struct drm_i915_private *dev_priv)
6499 u32 dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
6503 * WaFbcDisableDpfcClockGating:ilk
6505 dspclk_gate |= ILK_DPFCRUNIT_CLOCK_GATE_DISABLE |
6506 ILK_DPFCUNIT_CLOCK_GATE_DISABLE |
6507 ILK_DPFDUNIT_CLOCK_GATE_ENABLE;
6509 I915_WRITE(PCH_3DCGDIS0,
6510 MARIUNIT_CLOCK_GATE_DISABLE |
6511 SVSMUNIT_CLOCK_GATE_DISABLE);
6512 I915_WRITE(PCH_3DCGDIS1,
6513 VFMUNIT_CLOCK_GATE_DISABLE);
6516 * According to the spec the following bits should be set in
6517 * order to enable memory self-refresh
6518 * The bit 22/21 of 0x42004
6519 * The bit 5 of 0x42020
6520 * The bit 15 of 0x45000
6522 I915_WRITE(ILK_DISPLAY_CHICKEN2,
6523 (I915_READ(ILK_DISPLAY_CHICKEN2) |
6524 ILK_DPARB_GATE | ILK_VSDPFD_FULL));
6525 dspclk_gate |= ILK_DPARBUNIT_CLOCK_GATE_ENABLE;
6526 I915_WRITE(DISP_ARB_CTL,
6527 (I915_READ(DISP_ARB_CTL) |
6531 * Based on the document from hardware guys the following bits
6532 * should be set unconditionally in order to enable FBC.
6533 * The bit 22 of 0x42000
6534 * The bit 22 of 0x42004
6535 * The bit 7,8,9 of 0x42020.
6537 if (IS_IRONLAKE_M(dev_priv)) {
6538 /* WaFbcAsynchFlipDisableFbcQueue:ilk */
6539 I915_WRITE(ILK_DISPLAY_CHICKEN1,
6540 I915_READ(ILK_DISPLAY_CHICKEN1) |
6542 I915_WRITE(ILK_DISPLAY_CHICKEN2,
6543 I915_READ(ILK_DISPLAY_CHICKEN2) |
6547 I915_WRITE(ILK_DSPCLK_GATE_D, dspclk_gate);
6549 I915_WRITE(ILK_DISPLAY_CHICKEN2,
6550 I915_READ(ILK_DISPLAY_CHICKEN2) |
6551 ILK_ELPIN_409_SELECT);
6552 I915_WRITE(_3D_CHICKEN2,
6553 _3D_CHICKEN2_WM_READ_PIPELINED << 16 |
6554 _3D_CHICKEN2_WM_READ_PIPELINED);
6556 /* WaDisableRenderCachePipelinedFlush:ilk */
6557 I915_WRITE(CACHE_MODE_0,
6558 _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE));
6560 /* WaDisable_RenderCache_OperationalFlush:ilk */
6561 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
6563 g4x_disable_trickle_feed(dev_priv);
6565 ibx_init_clock_gating(dev_priv);
6568 static void cpt_init_clock_gating(struct drm_i915_private *dev_priv)
6574 * On Ibex Peak and Cougar Point, we need to disable clock
6575 * gating for the panel power sequencer or it will fail to
6576 * start up when no ports are active.
6578 I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE |
6579 PCH_DPLUNIT_CLOCK_GATE_DISABLE |
6580 PCH_CPUNIT_CLOCK_GATE_DISABLE);
6581 I915_WRITE(SOUTH_CHICKEN2, I915_READ(SOUTH_CHICKEN2) |
6582 DPLS_EDP_PPS_FIX_DIS);
6583 /* The below fixes the weird display corruption, a few pixels shifted
6584 * downward, on (only) LVDS of some HP laptops with IVY.
6586 for_each_pipe(dev_priv, pipe) {
6587 val = I915_READ(TRANS_CHICKEN2(pipe));
6588 val |= TRANS_CHICKEN2_TIMING_OVERRIDE;
6589 val &= ~TRANS_CHICKEN2_FDI_POLARITY_REVERSED;
6590 if (dev_priv->vbt.fdi_rx_polarity_inverted)
6591 val |= TRANS_CHICKEN2_FDI_POLARITY_REVERSED;
6592 val &= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_COUNTER;
6593 val &= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_MODESWITCH;
6594 I915_WRITE(TRANS_CHICKEN2(pipe), val);
6596 /* WADP0ClockGatingDisable */
6597 for_each_pipe(dev_priv, pipe) {
6598 I915_WRITE(TRANS_CHICKEN1(pipe),
6599 TRANS_CHICKEN1_DP0UNIT_GC_DISABLE);
6603 static void gen6_check_mch_setup(struct drm_i915_private *dev_priv)
6607 tmp = I915_READ(MCH_SSKPD);
6608 if ((tmp & MCH_SSKPD_WM0_MASK) != MCH_SSKPD_WM0_VAL)
6609 drm_dbg_kms(&dev_priv->drm,
6610 "Wrong MCH_SSKPD value: 0x%08x This can cause underruns.\n",
6614 static void gen6_init_clock_gating(struct drm_i915_private *dev_priv)
6616 u32 dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
6618 I915_WRITE(ILK_DSPCLK_GATE_D, dspclk_gate);
6620 I915_WRITE(ILK_DISPLAY_CHICKEN2,
6621 I915_READ(ILK_DISPLAY_CHICKEN2) |
6622 ILK_ELPIN_409_SELECT);
6624 /* WaDisableHiZPlanesWhenMSAAEnabled:snb */
6625 I915_WRITE(_3D_CHICKEN,
6626 _MASKED_BIT_ENABLE(_3D_CHICKEN_HIZ_PLANE_DISABLE_MSAA_4X_SNB));
6628 /* WaDisable_RenderCache_OperationalFlush:snb */
6629 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
6632 * BSpec recoomends 8x4 when MSAA is used,
6633 * however in practice 16x4 seems fastest.
6635 * Note that PS/WM thread counts depend on the WIZ hashing
6636 * disable bit, which we don't touch here, but it's good
6637 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
6639 I915_WRITE(GEN6_GT_MODE,
6640 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
6642 I915_WRITE(CACHE_MODE_0,
6643 _MASKED_BIT_DISABLE(CM0_STC_EVICT_DISABLE_LRA_SNB));
6645 I915_WRITE(GEN6_UCGCTL1,
6646 I915_READ(GEN6_UCGCTL1) |
6647 GEN6_BLBUNIT_CLOCK_GATE_DISABLE |
6648 GEN6_CSUNIT_CLOCK_GATE_DISABLE);
6650 /* According to the BSpec vol1g, bit 12 (RCPBUNIT) clock
6651 * gating disable must be set. Failure to set it results in
6652 * flickering pixels due to Z write ordering failures after
6653 * some amount of runtime in the Mesa "fire" demo, and Unigine
6654 * Sanctuary and Tropics, and apparently anything else with
6655 * alpha test or pixel discard.
6657 * According to the spec, bit 11 (RCCUNIT) must also be set,
6658 * but we didn't debug actual testcases to find it out.
6660 * WaDisableRCCUnitClockGating:snb
6661 * WaDisableRCPBUnitClockGating:snb
6663 I915_WRITE(GEN6_UCGCTL2,
6664 GEN6_RCPBUNIT_CLOCK_GATE_DISABLE |
6665 GEN6_RCCUNIT_CLOCK_GATE_DISABLE);
6667 /* WaStripsFansDisableFastClipPerformanceFix:snb */
6668 I915_WRITE(_3D_CHICKEN3,
6669 _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_FASTCLIP_CULL));
6673 * "This bit must be set if 3DSTATE_CLIP clip mode is set to normal and
6674 * 3DSTATE_SF number of SF output attributes is more than 16."
6676 I915_WRITE(_3D_CHICKEN3,
6677 _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_PIPELINED_ATTR_FETCH));
6680 * According to the spec the following bits should be
6681 * set in order to enable memory self-refresh and fbc:
6682 * The bit21 and bit22 of 0x42000
6683 * The bit21 and bit22 of 0x42004
6684 * The bit5 and bit7 of 0x42020
6685 * The bit14 of 0x70180
6686 * The bit14 of 0x71180
6688 * WaFbcAsynchFlipDisableFbcQueue:snb
6690 I915_WRITE(ILK_DISPLAY_CHICKEN1,
6691 I915_READ(ILK_DISPLAY_CHICKEN1) |
6692 ILK_FBCQ_DIS | ILK_PABSTRETCH_DIS);
6693 I915_WRITE(ILK_DISPLAY_CHICKEN2,
6694 I915_READ(ILK_DISPLAY_CHICKEN2) |
6695 ILK_DPARB_GATE | ILK_VSDPFD_FULL);
6696 I915_WRITE(ILK_DSPCLK_GATE_D,
6697 I915_READ(ILK_DSPCLK_GATE_D) |
6698 ILK_DPARBUNIT_CLOCK_GATE_ENABLE |
6699 ILK_DPFDUNIT_CLOCK_GATE_ENABLE);
6701 g4x_disable_trickle_feed(dev_priv);
6703 cpt_init_clock_gating(dev_priv);
6705 gen6_check_mch_setup(dev_priv);
6708 static void gen7_setup_fixed_func_scheduler(struct drm_i915_private *dev_priv)
6710 u32 reg = I915_READ(GEN7_FF_THREAD_MODE);
6713 * WaVSThreadDispatchOverride:ivb,vlv
6715 * This actually overrides the dispatch
6716 * mode for all thread types.
6718 reg &= ~GEN7_FF_SCHED_MASK;
6719 reg |= GEN7_FF_TS_SCHED_HW;
6720 reg |= GEN7_FF_VS_SCHED_HW;
6721 reg |= GEN7_FF_DS_SCHED_HW;
6723 I915_WRITE(GEN7_FF_THREAD_MODE, reg);
6726 static void lpt_init_clock_gating(struct drm_i915_private *dev_priv)
6729 * TODO: this bit should only be enabled when really needed, then
6730 * disabled when not needed anymore in order to save power.
6732 if (HAS_PCH_LPT_LP(dev_priv))
6733 I915_WRITE(SOUTH_DSPCLK_GATE_D,
6734 I915_READ(SOUTH_DSPCLK_GATE_D) |
6735 PCH_LP_PARTITION_LEVEL_DISABLE);
6737 /* WADPOClockGatingDisable:hsw */
6738 I915_WRITE(TRANS_CHICKEN1(PIPE_A),
6739 I915_READ(TRANS_CHICKEN1(PIPE_A)) |
6740 TRANS_CHICKEN1_DP0UNIT_GC_DISABLE);
6743 static void lpt_suspend_hw(struct drm_i915_private *dev_priv)
6745 if (HAS_PCH_LPT_LP(dev_priv)) {
6746 u32 val = I915_READ(SOUTH_DSPCLK_GATE_D);
6748 val &= ~PCH_LP_PARTITION_LEVEL_DISABLE;
6749 I915_WRITE(SOUTH_DSPCLK_GATE_D, val);
6753 static void gen8_set_l3sqc_credits(struct drm_i915_private *dev_priv,
6754 int general_prio_credits,
6755 int high_prio_credits)
6760 /* WaTempDisableDOPClkGating:bdw */
6761 misccpctl = I915_READ(GEN7_MISCCPCTL);
6762 I915_WRITE(GEN7_MISCCPCTL, misccpctl & ~GEN7_DOP_CLOCK_GATE_ENABLE);
6764 val = I915_READ(GEN8_L3SQCREG1);
6765 val &= ~L3_PRIO_CREDITS_MASK;
6766 val |= L3_GENERAL_PRIO_CREDITS(general_prio_credits);
6767 val |= L3_HIGH_PRIO_CREDITS(high_prio_credits);
6768 I915_WRITE(GEN8_L3SQCREG1, val);
6771 * Wait at least 100 clocks before re-enabling clock gating.
6772 * See the definition of L3SQCREG1 in BSpec.
6774 POSTING_READ(GEN8_L3SQCREG1);
6776 I915_WRITE(GEN7_MISCCPCTL, misccpctl);
6779 static void icl_init_clock_gating(struct drm_i915_private *dev_priv)
6781 /* This is not an Wa. Enable to reduce Sampler power */
6782 I915_WRITE(GEN10_DFR_RATIO_EN_AND_CHICKEN,
6783 I915_READ(GEN10_DFR_RATIO_EN_AND_CHICKEN) & ~DFR_DISABLE);
6785 /* WaEnable32PlaneMode:icl */
6786 I915_WRITE(GEN9_CSFE_CHICKEN1_RCS,
6787 _MASKED_BIT_ENABLE(GEN11_ENABLE_32_PLANE_MODE));
6790 * Wa_1408615072:icl,ehl (vsunit)
6791 * Wa_1407596294:icl,ehl (hsunit)
6793 intel_uncore_rmw(&dev_priv->uncore, UNSLICE_UNIT_LEVEL_CLKGATE,
6794 0, VSUNIT_CLKGATE_DIS | HSUNIT_CLKGATE_DIS);
6796 /* Wa_1407352427:icl,ehl */
6797 intel_uncore_rmw(&dev_priv->uncore, UNSLICE_UNIT_LEVEL_CLKGATE2,
6798 0, PSDUNIT_CLKGATE_DIS);
6800 /*Wa_14010594013:icl, ehl */
6801 intel_uncore_rmw(&dev_priv->uncore, GEN8_CHICKEN_DCPR_1,
6802 0, CNL_DELAY_PMRSP);
6805 static void tgl_init_clock_gating(struct drm_i915_private *dev_priv)
6807 u32 vd_pg_enable = 0;
6810 /* Wa_1408615072:tgl */
6811 intel_uncore_rmw(&dev_priv->uncore, UNSLICE_UNIT_LEVEL_CLKGATE2,
6812 0, VSUNIT_CLKGATE_DIS_TGL);
6814 /* This is not a WA. Enable VD HCP & MFX_ENC powergate */
6815 for (i = 0; i < I915_MAX_VCS; i++) {
6816 if (HAS_ENGINE(dev_priv, _VCS(i)))
6817 vd_pg_enable |= VDN_HCP_POWERGATE_ENABLE(i) |
6818 VDN_MFX_POWERGATE_ENABLE(i);
6821 I915_WRITE(POWERGATE_ENABLE,
6822 I915_READ(POWERGATE_ENABLE) | vd_pg_enable);
6824 /* Wa_1409825376:tgl (pre-prod)*/
6825 if (IS_TGL_REVID(dev_priv, TGL_REVID_A0, TGL_REVID_A0))
6826 I915_WRITE(GEN9_CLKGATE_DIS_3, I915_READ(GEN9_CLKGATE_DIS_3) |
6827 TGL_VRH_GATING_DIS);
6830 static void cnp_init_clock_gating(struct drm_i915_private *dev_priv)
6832 if (!HAS_PCH_CNP(dev_priv))
6835 /* Display WA #1181 WaSouthDisplayDisablePWMCGEGating: cnp */
6836 I915_WRITE(SOUTH_DSPCLK_GATE_D, I915_READ(SOUTH_DSPCLK_GATE_D) |
6837 CNP_PWM_CGE_GATING_DISABLE);
6840 static void cnl_init_clock_gating(struct drm_i915_private *dev_priv)
6843 cnp_init_clock_gating(dev_priv);
6845 /* This is not an Wa. Enable for better image quality */
6846 I915_WRITE(_3D_CHICKEN3,
6847 _MASKED_BIT_ENABLE(_3D_CHICKEN3_AA_LINE_QUALITY_FIX_ENABLE));
6849 /* WaEnableChickenDCPR:cnl */
6850 I915_WRITE(GEN8_CHICKEN_DCPR_1,
6851 I915_READ(GEN8_CHICKEN_DCPR_1) | MASK_WAKEMEM);
6853 /* WaFbcWakeMemOn:cnl */
6854 I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) |
6855 DISP_FBC_MEMORY_WAKE);
6857 val = I915_READ(SLICE_UNIT_LEVEL_CLKGATE);
6858 /* ReadHitWriteOnlyDisable:cnl */
6859 val |= RCCUNIT_CLKGATE_DIS;
6860 /* WaSarbUnitClockGatingDisable:cnl (pre-prod) */
6861 if (IS_CNL_REVID(dev_priv, CNL_REVID_A0, CNL_REVID_B0))
6862 val |= SARBUNIT_CLKGATE_DIS;
6863 I915_WRITE(SLICE_UNIT_LEVEL_CLKGATE, val);
6865 /* Wa_2201832410:cnl */
6866 val = I915_READ(SUBSLICE_UNIT_LEVEL_CLKGATE);
6867 val |= GWUNIT_CLKGATE_DIS;
6868 I915_WRITE(SUBSLICE_UNIT_LEVEL_CLKGATE, val);
6870 /* WaDisableVFclkgate:cnl */
6871 /* WaVFUnitClockGatingDisable:cnl */
6872 val = I915_READ(UNSLICE_UNIT_LEVEL_CLKGATE);
6873 val |= VFUNIT_CLKGATE_DIS;
6874 I915_WRITE(UNSLICE_UNIT_LEVEL_CLKGATE, val);
6877 static void cfl_init_clock_gating(struct drm_i915_private *dev_priv)
6879 cnp_init_clock_gating(dev_priv);
6880 gen9_init_clock_gating(dev_priv);
6882 /* WaFbcNukeOnHostModify:cfl */
6883 I915_WRITE(ILK_DPFC_CHICKEN, I915_READ(ILK_DPFC_CHICKEN) |
6884 ILK_DPFC_NUKE_ON_ANY_MODIFICATION);
6887 static void kbl_init_clock_gating(struct drm_i915_private *dev_priv)
6889 gen9_init_clock_gating(dev_priv);
6891 /* WaDisableSDEUnitClockGating:kbl */
6892 if (IS_KBL_REVID(dev_priv, 0, KBL_REVID_B0))
6893 I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
6894 GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
6896 /* WaDisableGamClockGating:kbl */
6897 if (IS_KBL_REVID(dev_priv, 0, KBL_REVID_B0))
6898 I915_WRITE(GEN6_UCGCTL1, I915_READ(GEN6_UCGCTL1) |
6899 GEN6_GAMUNIT_CLOCK_GATE_DISABLE);
6901 /* WaFbcNukeOnHostModify:kbl */
6902 I915_WRITE(ILK_DPFC_CHICKEN, I915_READ(ILK_DPFC_CHICKEN) |
6903 ILK_DPFC_NUKE_ON_ANY_MODIFICATION);
6906 static void skl_init_clock_gating(struct drm_i915_private *dev_priv)
6908 gen9_init_clock_gating(dev_priv);
6910 /* WAC6entrylatency:skl */
6911 I915_WRITE(FBC_LLC_READ_CTRL, I915_READ(FBC_LLC_READ_CTRL) |
6912 FBC_LLC_FULLY_OPEN);
6914 /* WaFbcNukeOnHostModify:skl */
6915 I915_WRITE(ILK_DPFC_CHICKEN, I915_READ(ILK_DPFC_CHICKEN) |
6916 ILK_DPFC_NUKE_ON_ANY_MODIFICATION);
6919 static void bdw_init_clock_gating(struct drm_i915_private *dev_priv)
6923 /* WaSwitchSolVfFArbitrationPriority:bdw */
6924 I915_WRITE(GAM_ECOCHK, I915_READ(GAM_ECOCHK) | HSW_ECOCHK_ARB_PRIO_SOL);
6926 /* WaPsrDPAMaskVBlankInSRD:bdw */
6927 I915_WRITE(CHICKEN_PAR1_1,
6928 I915_READ(CHICKEN_PAR1_1) | DPA_MASK_VBLANK_SRD);
6930 /* WaPsrDPRSUnmaskVBlankInSRD:bdw */
6931 for_each_pipe(dev_priv, pipe) {
6932 I915_WRITE(CHICKEN_PIPESL_1(pipe),
6933 I915_READ(CHICKEN_PIPESL_1(pipe)) |
6934 BDW_DPRS_MASK_VBLANK_SRD);
6937 /* WaVSRefCountFullforceMissDisable:bdw */
6938 /* WaDSRefCountFullforceMissDisable:bdw */
6939 I915_WRITE(GEN7_FF_THREAD_MODE,
6940 I915_READ(GEN7_FF_THREAD_MODE) &
6941 ~(GEN8_FF_DS_REF_CNT_FFME | GEN7_FF_VS_REF_CNT_FFME));
6943 I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL,
6944 _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE));
6946 /* WaDisableSDEUnitClockGating:bdw */
6947 I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
6948 GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
6950 /* WaProgramL3SqcReg1Default:bdw */
6951 gen8_set_l3sqc_credits(dev_priv, 30, 2);
6953 /* WaKVMNotificationOnConfigChange:bdw */
6954 I915_WRITE(CHICKEN_PAR2_1, I915_READ(CHICKEN_PAR2_1)
6955 | KVM_CONFIG_CHANGE_NOTIFICATION_SELECT);
6957 lpt_init_clock_gating(dev_priv);
6959 /* WaDisableDopClockGating:bdw
6961 * Also see the CHICKEN2 write in bdw_init_workarounds() to disable DOP
6964 I915_WRITE(GEN6_UCGCTL1,
6965 I915_READ(GEN6_UCGCTL1) | GEN6_EU_TCUNIT_CLOCK_GATE_DISABLE);
6968 static void hsw_init_clock_gating(struct drm_i915_private *dev_priv)
6970 /* L3 caching of data atomics doesn't work -- disable it. */
6971 I915_WRITE(HSW_SCRATCH1, HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE);
6972 I915_WRITE(HSW_ROW_CHICKEN3,
6973 _MASKED_BIT_ENABLE(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE));
6975 /* This is required by WaCatErrorRejectionIssue:hsw */
6976 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
6977 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
6978 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
6980 /* WaVSRefCountFullforceMissDisable:hsw */
6981 I915_WRITE(GEN7_FF_THREAD_MODE,
6982 I915_READ(GEN7_FF_THREAD_MODE) & ~GEN7_FF_VS_REF_CNT_FFME);
6984 /* WaDisable_RenderCache_OperationalFlush:hsw */
6985 I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
6987 /* enable HiZ Raw Stall Optimization */
6988 I915_WRITE(CACHE_MODE_0_GEN7,
6989 _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE));
6991 /* WaDisable4x2SubspanOptimization:hsw */
6992 I915_WRITE(CACHE_MODE_1,
6993 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
6996 * BSpec recommends 8x4 when MSAA is used,
6997 * however in practice 16x4 seems fastest.
6999 * Note that PS/WM thread counts depend on the WIZ hashing
7000 * disable bit, which we don't touch here, but it's good
7001 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
7003 I915_WRITE(GEN7_GT_MODE,
7004 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
7006 /* WaSampleCChickenBitEnable:hsw */
7007 I915_WRITE(HALF_SLICE_CHICKEN3,
7008 _MASKED_BIT_ENABLE(HSW_SAMPLE_C_PERFORMANCE));
7010 /* WaSwitchSolVfFArbitrationPriority:hsw */
7011 I915_WRITE(GAM_ECOCHK, I915_READ(GAM_ECOCHK) | HSW_ECOCHK_ARB_PRIO_SOL);
7013 lpt_init_clock_gating(dev_priv);
7016 static void ivb_init_clock_gating(struct drm_i915_private *dev_priv)
7020 I915_WRITE(ILK_DSPCLK_GATE_D, ILK_VRHUNIT_CLOCK_GATE_DISABLE);
7022 /* WaDisableEarlyCull:ivb */
7023 I915_WRITE(_3D_CHICKEN3,
7024 _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL));
7026 /* WaDisableBackToBackFlipFix:ivb */
7027 I915_WRITE(IVB_CHICKEN3,
7028 CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
7029 CHICKEN3_DGMG_DONE_FIX_DISABLE);
7031 /* WaDisablePSDDualDispatchEnable:ivb */
7032 if (IS_IVB_GT1(dev_priv))
7033 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
7034 _MASKED_BIT_ENABLE(GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
7036 /* WaDisable_RenderCache_OperationalFlush:ivb */
7037 I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7039 /* Apply the WaDisableRHWOOptimizationForRenderHang:ivb workaround. */
7040 I915_WRITE(GEN7_COMMON_SLICE_CHICKEN1,
7041 GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC);
7043 /* WaApplyL3ControlAndL3ChickenMode:ivb */
7044 I915_WRITE(GEN7_L3CNTLREG1,
7045 GEN7_WA_FOR_GEN7_L3_CONTROL);
7046 I915_WRITE(GEN7_L3_CHICKEN_MODE_REGISTER,
7047 GEN7_WA_L3_CHICKEN_MODE);
7048 if (IS_IVB_GT1(dev_priv))
7049 I915_WRITE(GEN7_ROW_CHICKEN2,
7050 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
7052 /* must write both registers */
7053 I915_WRITE(GEN7_ROW_CHICKEN2,
7054 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
7055 I915_WRITE(GEN7_ROW_CHICKEN2_GT2,
7056 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
7059 /* WaForceL3Serialization:ivb */
7060 I915_WRITE(GEN7_L3SQCREG4, I915_READ(GEN7_L3SQCREG4) &
7061 ~L3SQ_URB_READ_CAM_MATCH_DISABLE);
7064 * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
7065 * This implements the WaDisableRCZUnitClockGating:ivb workaround.
7067 I915_WRITE(GEN6_UCGCTL2,
7068 GEN6_RCZUNIT_CLOCK_GATE_DISABLE);
7070 /* This is required by WaCatErrorRejectionIssue:ivb */
7071 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
7072 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
7073 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
7075 g4x_disable_trickle_feed(dev_priv);
7077 gen7_setup_fixed_func_scheduler(dev_priv);
7079 if (0) { /* causes HiZ corruption on ivb:gt1 */
7080 /* enable HiZ Raw Stall Optimization */
7081 I915_WRITE(CACHE_MODE_0_GEN7,
7082 _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE));
7085 /* WaDisable4x2SubspanOptimization:ivb */
7086 I915_WRITE(CACHE_MODE_1,
7087 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
7090 * BSpec recommends 8x4 when MSAA is used,
7091 * however in practice 16x4 seems fastest.
7093 * Note that PS/WM thread counts depend on the WIZ hashing
7094 * disable bit, which we don't touch here, but it's good
7095 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
7097 I915_WRITE(GEN7_GT_MODE,
7098 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
7100 snpcr = I915_READ(GEN6_MBCUNIT_SNPCR);
7101 snpcr &= ~GEN6_MBC_SNPCR_MASK;
7102 snpcr |= GEN6_MBC_SNPCR_MED;
7103 I915_WRITE(GEN6_MBCUNIT_SNPCR, snpcr);
7105 if (!HAS_PCH_NOP(dev_priv))
7106 cpt_init_clock_gating(dev_priv);
7108 gen6_check_mch_setup(dev_priv);
7111 static void vlv_init_clock_gating(struct drm_i915_private *dev_priv)
7113 /* WaDisableEarlyCull:vlv */
7114 I915_WRITE(_3D_CHICKEN3,
7115 _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL));
7117 /* WaDisableBackToBackFlipFix:vlv */
7118 I915_WRITE(IVB_CHICKEN3,
7119 CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
7120 CHICKEN3_DGMG_DONE_FIX_DISABLE);
7122 /* WaPsdDispatchEnable:vlv */
7123 /* WaDisablePSDDualDispatchEnable:vlv */
7124 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
7125 _MASKED_BIT_ENABLE(GEN7_MAX_PS_THREAD_DEP |
7126 GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
7128 /* WaDisable_RenderCache_OperationalFlush:vlv */
7129 I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7131 /* WaForceL3Serialization:vlv */
7132 I915_WRITE(GEN7_L3SQCREG4, I915_READ(GEN7_L3SQCREG4) &
7133 ~L3SQ_URB_READ_CAM_MATCH_DISABLE);
7135 /* WaDisableDopClockGating:vlv */
7136 I915_WRITE(GEN7_ROW_CHICKEN2,
7137 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
7139 /* This is required by WaCatErrorRejectionIssue:vlv */
7140 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
7141 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
7142 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
7144 gen7_setup_fixed_func_scheduler(dev_priv);
7147 * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
7148 * This implements the WaDisableRCZUnitClockGating:vlv workaround.
7150 I915_WRITE(GEN6_UCGCTL2,
7151 GEN6_RCZUNIT_CLOCK_GATE_DISABLE);
7153 /* WaDisableL3Bank2xClockGate:vlv
7154 * Disabling L3 clock gating- MMIO 940c[25] = 1
7155 * Set bit 25, to disable L3_BANK_2x_CLK_GATING */
7156 I915_WRITE(GEN7_UCGCTL4,
7157 I915_READ(GEN7_UCGCTL4) | GEN7_L3BANK2X_CLOCK_GATE_DISABLE);
7160 * BSpec says this must be set, even though
7161 * WaDisable4x2SubspanOptimization isn't listed for VLV.
7163 I915_WRITE(CACHE_MODE_1,
7164 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
7167 * BSpec recommends 8x4 when MSAA is used,
7168 * however in practice 16x4 seems fastest.
7170 * Note that PS/WM thread counts depend on the WIZ hashing
7171 * disable bit, which we don't touch here, but it's good
7172 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
7174 I915_WRITE(GEN7_GT_MODE,
7175 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
7178 * WaIncreaseL3CreditsForVLVB0:vlv
7179 * This is the hardware default actually.
7181 I915_WRITE(GEN7_L3SQCREG1, VLV_B0_WA_L3SQCREG1_VALUE);
7184 * WaDisableVLVClockGating_VBIIssue:vlv
7185 * Disable clock gating on th GCFG unit to prevent a delay
7186 * in the reporting of vblank events.
7188 I915_WRITE(VLV_GUNIT_CLOCK_GATE, GCFG_DIS);
7191 static void chv_init_clock_gating(struct drm_i915_private *dev_priv)
7193 /* WaVSRefCountFullforceMissDisable:chv */
7194 /* WaDSRefCountFullforceMissDisable:chv */
7195 I915_WRITE(GEN7_FF_THREAD_MODE,
7196 I915_READ(GEN7_FF_THREAD_MODE) &
7197 ~(GEN8_FF_DS_REF_CNT_FFME | GEN7_FF_VS_REF_CNT_FFME));
7199 /* WaDisableSemaphoreAndSyncFlipWait:chv */
7200 I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL,
7201 _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE));
7203 /* WaDisableCSUnitClockGating:chv */
7204 I915_WRITE(GEN6_UCGCTL1, I915_READ(GEN6_UCGCTL1) |
7205 GEN6_CSUNIT_CLOCK_GATE_DISABLE);
7207 /* WaDisableSDEUnitClockGating:chv */
7208 I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
7209 GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
7212 * WaProgramL3SqcReg1Default:chv
7213 * See gfxspecs/Related Documents/Performance Guide/
7214 * LSQC Setting Recommendations.
7216 gen8_set_l3sqc_credits(dev_priv, 38, 2);
7219 static void g4x_init_clock_gating(struct drm_i915_private *dev_priv)
7223 I915_WRITE(RENCLK_GATE_D1, 0);
7224 I915_WRITE(RENCLK_GATE_D2, VF_UNIT_CLOCK_GATE_DISABLE |
7225 GS_UNIT_CLOCK_GATE_DISABLE |
7226 CL_UNIT_CLOCK_GATE_DISABLE);
7227 I915_WRITE(RAMCLK_GATE_D, 0);
7228 dspclk_gate = VRHUNIT_CLOCK_GATE_DISABLE |
7229 OVRUNIT_CLOCK_GATE_DISABLE |
7230 OVCUNIT_CLOCK_GATE_DISABLE;
7231 if (IS_GM45(dev_priv))
7232 dspclk_gate |= DSSUNIT_CLOCK_GATE_DISABLE;
7233 I915_WRITE(DSPCLK_GATE_D, dspclk_gate);
7235 /* WaDisableRenderCachePipelinedFlush */
7236 I915_WRITE(CACHE_MODE_0,
7237 _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE));
7239 /* WaDisable_RenderCache_OperationalFlush:g4x */
7240 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7242 g4x_disable_trickle_feed(dev_priv);
7245 static void i965gm_init_clock_gating(struct drm_i915_private *dev_priv)
7247 struct intel_uncore *uncore = &dev_priv->uncore;
7249 intel_uncore_write(uncore, RENCLK_GATE_D1, I965_RCC_CLOCK_GATE_DISABLE);
7250 intel_uncore_write(uncore, RENCLK_GATE_D2, 0);
7251 intel_uncore_write(uncore, DSPCLK_GATE_D, 0);
7252 intel_uncore_write(uncore, RAMCLK_GATE_D, 0);
7253 intel_uncore_write16(uncore, DEUC, 0);
7254 intel_uncore_write(uncore,
7256 _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
7258 /* WaDisable_RenderCache_OperationalFlush:gen4 */
7259 intel_uncore_write(uncore,
7261 _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7264 static void i965g_init_clock_gating(struct drm_i915_private *dev_priv)
7266 I915_WRITE(RENCLK_GATE_D1, I965_RCZ_CLOCK_GATE_DISABLE |
7267 I965_RCC_CLOCK_GATE_DISABLE |
7268 I965_RCPB_CLOCK_GATE_DISABLE |
7269 I965_ISC_CLOCK_GATE_DISABLE |
7270 I965_FBC_CLOCK_GATE_DISABLE);
7271 I915_WRITE(RENCLK_GATE_D2, 0);
7272 I915_WRITE(MI_ARB_STATE,
7273 _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
7275 /* WaDisable_RenderCache_OperationalFlush:gen4 */
7276 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7279 static void gen3_init_clock_gating(struct drm_i915_private *dev_priv)
7281 u32 dstate = I915_READ(D_STATE);
7283 dstate |= DSTATE_PLL_D3_OFF | DSTATE_GFX_CLOCK_GATING |
7284 DSTATE_DOT_CLOCK_GATING;
7285 I915_WRITE(D_STATE, dstate);
7287 if (IS_PINEVIEW(dev_priv))
7288 I915_WRITE(ECOSKPD, _MASKED_BIT_ENABLE(ECO_GATING_CX_ONLY));
7290 /* IIR "flip pending" means done if this bit is set */
7291 I915_WRITE(ECOSKPD, _MASKED_BIT_DISABLE(ECO_FLIP_DONE));
7293 /* interrupts should cause a wake up from C3 */
7294 I915_WRITE(INSTPM, _MASKED_BIT_ENABLE(INSTPM_AGPBUSY_INT_EN));
7296 /* On GEN3 we really need to make sure the ARB C3 LP bit is set */
7297 I915_WRITE(MI_ARB_STATE, _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE));
7299 I915_WRITE(MI_ARB_STATE,
7300 _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
7303 static void i85x_init_clock_gating(struct drm_i915_private *dev_priv)
7305 I915_WRITE(RENCLK_GATE_D1, SV_CLOCK_GATE_DISABLE);
7307 /* interrupts should cause a wake up from C3 */
7308 I915_WRITE(MI_STATE, _MASKED_BIT_ENABLE(MI_AGPBUSY_INT_EN) |
7309 _MASKED_BIT_DISABLE(MI_AGPBUSY_830_MODE));
7311 I915_WRITE(MEM_MODE,
7312 _MASKED_BIT_ENABLE(MEM_DISPLAY_TRICKLE_FEED_DISABLE));
7315 static void i830_init_clock_gating(struct drm_i915_private *dev_priv)
7317 I915_WRITE(MEM_MODE,
7318 _MASKED_BIT_ENABLE(MEM_DISPLAY_A_TRICKLE_FEED_DISABLE) |
7319 _MASKED_BIT_ENABLE(MEM_DISPLAY_B_TRICKLE_FEED_DISABLE));
7322 void intel_init_clock_gating(struct drm_i915_private *dev_priv)
7324 dev_priv->display.init_clock_gating(dev_priv);
7327 void intel_suspend_hw(struct drm_i915_private *dev_priv)
7329 if (HAS_PCH_LPT(dev_priv))
7330 lpt_suspend_hw(dev_priv);
7333 static void nop_init_clock_gating(struct drm_i915_private *dev_priv)
7335 drm_dbg_kms(&dev_priv->drm,
7336 "No clock gating settings or workarounds applied.\n");
7340 * intel_init_clock_gating_hooks - setup the clock gating hooks
7341 * @dev_priv: device private
7343 * Setup the hooks that configure which clocks of a given platform can be
7344 * gated and also apply various GT and display specific workarounds for these
7345 * platforms. Note that some GT specific workarounds are applied separately
7346 * when GPU contexts or batchbuffers start their execution.
7348 void intel_init_clock_gating_hooks(struct drm_i915_private *dev_priv)
7350 if (IS_GEN(dev_priv, 12))
7351 dev_priv->display.init_clock_gating = tgl_init_clock_gating;
7352 else if (IS_GEN(dev_priv, 11))
7353 dev_priv->display.init_clock_gating = icl_init_clock_gating;
7354 else if (IS_CANNONLAKE(dev_priv))
7355 dev_priv->display.init_clock_gating = cnl_init_clock_gating;
7356 else if (IS_COFFEELAKE(dev_priv))
7357 dev_priv->display.init_clock_gating = cfl_init_clock_gating;
7358 else if (IS_SKYLAKE(dev_priv))
7359 dev_priv->display.init_clock_gating = skl_init_clock_gating;
7360 else if (IS_KABYLAKE(dev_priv))
7361 dev_priv->display.init_clock_gating = kbl_init_clock_gating;
7362 else if (IS_BROXTON(dev_priv))
7363 dev_priv->display.init_clock_gating = bxt_init_clock_gating;
7364 else if (IS_GEMINILAKE(dev_priv))
7365 dev_priv->display.init_clock_gating = glk_init_clock_gating;
7366 else if (IS_BROADWELL(dev_priv))
7367 dev_priv->display.init_clock_gating = bdw_init_clock_gating;
7368 else if (IS_CHERRYVIEW(dev_priv))
7369 dev_priv->display.init_clock_gating = chv_init_clock_gating;
7370 else if (IS_HASWELL(dev_priv))
7371 dev_priv->display.init_clock_gating = hsw_init_clock_gating;
7372 else if (IS_IVYBRIDGE(dev_priv))
7373 dev_priv->display.init_clock_gating = ivb_init_clock_gating;
7374 else if (IS_VALLEYVIEW(dev_priv))
7375 dev_priv->display.init_clock_gating = vlv_init_clock_gating;
7376 else if (IS_GEN(dev_priv, 6))
7377 dev_priv->display.init_clock_gating = gen6_init_clock_gating;
7378 else if (IS_GEN(dev_priv, 5))
7379 dev_priv->display.init_clock_gating = ilk_init_clock_gating;
7380 else if (IS_G4X(dev_priv))
7381 dev_priv->display.init_clock_gating = g4x_init_clock_gating;
7382 else if (IS_I965GM(dev_priv))
7383 dev_priv->display.init_clock_gating = i965gm_init_clock_gating;
7384 else if (IS_I965G(dev_priv))
7385 dev_priv->display.init_clock_gating = i965g_init_clock_gating;
7386 else if (IS_GEN(dev_priv, 3))
7387 dev_priv->display.init_clock_gating = gen3_init_clock_gating;
7388 else if (IS_I85X(dev_priv) || IS_I865G(dev_priv))
7389 dev_priv->display.init_clock_gating = i85x_init_clock_gating;
7390 else if (IS_GEN(dev_priv, 2))
7391 dev_priv->display.init_clock_gating = i830_init_clock_gating;
7393 MISSING_CASE(INTEL_DEVID(dev_priv));
7394 dev_priv->display.init_clock_gating = nop_init_clock_gating;
7398 /* Set up chip specific power management-related functions */
7399 void intel_init_pm(struct drm_i915_private *dev_priv)
7402 if (IS_PINEVIEW(dev_priv))
7403 pnv_get_mem_freq(dev_priv);
7404 else if (IS_GEN(dev_priv, 5))
7405 ilk_get_mem_freq(dev_priv);
7407 if (intel_has_sagv(dev_priv))
7408 skl_setup_sagv_block_time(dev_priv);
7410 /* For FIFO watermark updates */
7411 if (INTEL_GEN(dev_priv) >= 9) {
7412 skl_setup_wm_latency(dev_priv);
7413 dev_priv->display.compute_global_watermarks = skl_compute_wm;
7414 } else if (HAS_PCH_SPLIT(dev_priv)) {
7415 ilk_setup_wm_latency(dev_priv);
7417 if ((IS_GEN(dev_priv, 5) && dev_priv->wm.pri_latency[1] &&
7418 dev_priv->wm.spr_latency[1] && dev_priv->wm.cur_latency[1]) ||
7419 (!IS_GEN(dev_priv, 5) && dev_priv->wm.pri_latency[0] &&
7420 dev_priv->wm.spr_latency[0] && dev_priv->wm.cur_latency[0])) {
7421 dev_priv->display.compute_pipe_wm = ilk_compute_pipe_wm;
7422 dev_priv->display.compute_intermediate_wm =
7423 ilk_compute_intermediate_wm;
7424 dev_priv->display.initial_watermarks =
7425 ilk_initial_watermarks;
7426 dev_priv->display.optimize_watermarks =
7427 ilk_optimize_watermarks;
7429 drm_dbg_kms(&dev_priv->drm,
7430 "Failed to read display plane latency. "
7433 } else if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv)) {
7434 vlv_setup_wm_latency(dev_priv);
7435 dev_priv->display.compute_pipe_wm = vlv_compute_pipe_wm;
7436 dev_priv->display.compute_intermediate_wm = vlv_compute_intermediate_wm;
7437 dev_priv->display.initial_watermarks = vlv_initial_watermarks;
7438 dev_priv->display.optimize_watermarks = vlv_optimize_watermarks;
7439 dev_priv->display.atomic_update_watermarks = vlv_atomic_update_fifo;
7440 } else if (IS_G4X(dev_priv)) {
7441 g4x_setup_wm_latency(dev_priv);
7442 dev_priv->display.compute_pipe_wm = g4x_compute_pipe_wm;
7443 dev_priv->display.compute_intermediate_wm = g4x_compute_intermediate_wm;
7444 dev_priv->display.initial_watermarks = g4x_initial_watermarks;
7445 dev_priv->display.optimize_watermarks = g4x_optimize_watermarks;
7446 } else if (IS_PINEVIEW(dev_priv)) {
7447 if (!intel_get_cxsr_latency(!IS_MOBILE(dev_priv),
7450 dev_priv->mem_freq)) {
7451 drm_info(&dev_priv->drm,
7452 "failed to find known CxSR latency "
7453 "(found ddr%s fsb freq %d, mem freq %d), "
7455 (dev_priv->is_ddr3 == 1) ? "3" : "2",
7456 dev_priv->fsb_freq, dev_priv->mem_freq);
7457 /* Disable CxSR and never update its watermark again */
7458 intel_set_memory_cxsr(dev_priv, false);
7459 dev_priv->display.update_wm = NULL;
7461 dev_priv->display.update_wm = pnv_update_wm;
7462 } else if (IS_GEN(dev_priv, 4)) {
7463 dev_priv->display.update_wm = i965_update_wm;
7464 } else if (IS_GEN(dev_priv, 3)) {
7465 dev_priv->display.update_wm = i9xx_update_wm;
7466 dev_priv->display.get_fifo_size = i9xx_get_fifo_size;
7467 } else if (IS_GEN(dev_priv, 2)) {
7468 if (INTEL_NUM_PIPES(dev_priv) == 1) {
7469 dev_priv->display.update_wm = i845_update_wm;
7470 dev_priv->display.get_fifo_size = i845_get_fifo_size;
7472 dev_priv->display.update_wm = i9xx_update_wm;
7473 dev_priv->display.get_fifo_size = i830_get_fifo_size;
7476 drm_err(&dev_priv->drm,
7477 "unexpected fall-through in %s\n", __func__);
7481 void intel_pm_setup(struct drm_i915_private *dev_priv)
7483 dev_priv->runtime_pm.suspended = false;
7484 atomic_set(&dev_priv->runtime_pm.wakeref_count, 0);