drivers/gpu/drm/i915/display/intel_overlay.c

   1 /*
   2  * Copyright © 2009
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  20  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  21  * SOFTWARE.
  22  *
  23  * Authors:
  24  *    Daniel Vetter <daniel@ffwll.ch>
  25  *
  26  * Derived from Xorg ddx, xf86-video-intel, src/i830_video.c
  27  */
  28
  29 #include <drm/drm_fourcc.h>
  30
  31 #include "gem/i915_gem_pm.h"
  32 #include "gt/intel_ring.h"
  33
  34 #include "i915_drv.h"
  35 #include "i915_reg.h"
  36 #include "intel_display_types.h"
  37 #include "intel_frontbuffer.h"
  38 #include "intel_overlay.h"
  39
  40 /* Limits for overlay size. According to intel doc, the real limits are:
  41  * Y width: 4095, UV width (planar): 2047, Y height: 2047,
  42  * UV width (planar): * 1023. But the xorg thinks 2048 for height and width. Use
  43  * the mininum of both.  */
  44 #define IMAGE_MAX_WIDTH         2048
  45 #define IMAGE_MAX_HEIGHT        2046 /* 2 * 1023 */
  46 /* on 830 and 845 these large limits result in the card hanging */
  47 #define IMAGE_MAX_WIDTH_LEGACY  1024
  48 #define IMAGE_MAX_HEIGHT_LEGACY 1088
  49
  50 /* overlay register definitions */
  51 /* OCMD register */
  52 #define OCMD_TILED_SURFACE      (0x1<<19)
  53 #define OCMD_MIRROR_MASK        (0x3<<17)
  54 #define OCMD_MIRROR_MODE        (0x3<<17)
  55 #define OCMD_MIRROR_HORIZONTAL  (0x1<<17)
  56 #define OCMD_MIRROR_VERTICAL    (0x2<<17)
  57 #define OCMD_MIRROR_BOTH        (0x3<<17)
  58 #define OCMD_BYTEORDER_MASK     (0x3<<14) /* zero for YUYV or FOURCC YUY2 */
  59 #define OCMD_UV_SWAP            (0x1<<14) /* YVYU */
  60 #define OCMD_Y_SWAP             (0x2<<14) /* UYVY or FOURCC UYVY */
  61 #define OCMD_Y_AND_UV_SWAP      (0x3<<14) /* VYUY */
  62 #define OCMD_SOURCE_FORMAT_MASK (0xf<<10)
  63 #define OCMD_RGB_888            (0x1<<10) /* not in i965 Intel docs */
  64 #define OCMD_RGB_555            (0x2<<10) /* not in i965 Intel docs */
  65 #define OCMD_RGB_565            (0x3<<10) /* not in i965 Intel docs */
  66 #define OCMD_YUV_422_PACKED     (0x8<<10)
  67 #define OCMD_YUV_411_PACKED     (0x9<<10) /* not in i965 Intel docs */
  68 #define OCMD_YUV_420_PLANAR     (0xc<<10)
  69 #define OCMD_YUV_422_PLANAR     (0xd<<10)
  70 #define OCMD_YUV_410_PLANAR     (0xe<<10) /* also 411 */
  71 #define OCMD_TVSYNCFLIP_PARITY  (0x1<<9)
  72 #define OCMD_TVSYNCFLIP_ENABLE  (0x1<<7)
  73 #define OCMD_BUF_TYPE_MASK      (0x1<<5)
  74 #define OCMD_BUF_TYPE_FRAME     (0x0<<5)
  75 #define OCMD_BUF_TYPE_FIELD     (0x1<<5)
  76 #define OCMD_TEST_MODE          (0x1<<4)
  77 #define OCMD_BUFFER_SELECT      (0x3<<2)
  78 #define OCMD_BUFFER0            (0x0<<2)
  79 #define OCMD_BUFFER1            (0x1<<2)
  80 #define OCMD_FIELD_SELECT       (0x1<<2)
  81 #define OCMD_FIELD0             (0x0<<1)
  82 #define OCMD_FIELD1             (0x1<<1)
  83 #define OCMD_ENABLE             (0x1<<0)
  84
  85 /* OCONFIG register */
  86 #define OCONF_PIPE_MASK         (0x1<<18)
  87 #define OCONF_PIPE_A            (0x0<<18)
  88 #define OCONF_PIPE_B            (0x1<<18)
  89 #define OCONF_GAMMA2_ENABLE     (0x1<<16)
  90 #define OCONF_CSC_MODE_BT601    (0x0<<5)
  91 #define OCONF_CSC_MODE_BT709    (0x1<<5)
  92 #define OCONF_CSC_BYPASS        (0x1<<4)
  93 #define OCONF_CC_OUT_8BIT       (0x1<<3)
  94 #define OCONF_TEST_MODE         (0x1<<2)
  95 #define OCONF_THREE_LINE_BUFFER (0x1<<0)
  96 #define OCONF_TWO_LINE_BUFFER   (0x0<<0)
  97
  98 /* DCLRKM (dst-key) register */
  99 #define DST_KEY_ENABLE          (0x1<<31)
 100 #define CLK_RGB24_MASK          0x0
 101 #define CLK_RGB16_MASK          0x070307
 102 #define CLK_RGB15_MASK          0x070707
 103
 104 #define RGB30_TO_COLORKEY(c) \
 105         ((((c) & 0x3fc00000) >> 6) | (((c) & 0x000ff000) >> 4) | (((c) & 0x000003fc) >> 2))
 106 #define RGB16_TO_COLORKEY(c) \
 107         ((((c) & 0xf800) << 8) | (((c) & 0x07e0) << 5) | (((c) & 0x001f) << 3))
 108 #define RGB15_TO_COLORKEY(c) \
 109         ((((c) & 0x7c00) << 9) | (((c) & 0x03e0) << 6) | (((c) & 0x001f) << 3))
 110 #define RGB8I_TO_COLORKEY(c) \
 111         ((((c) & 0xff) << 16) | (((c) & 0xff) << 8) | (((c) & 0xff) << 0))
 112
 113 /* overlay flip addr flag */
 114 #define OFC_UPDATE              0x1
 115
 116 /* polyphase filter coefficients */
 117 #define N_HORIZ_Y_TAPS          5
 118 #define N_VERT_Y_TAPS           3
 119 #define N_HORIZ_UV_TAPS         3
 120 #define N_VERT_UV_TAPS          3
 121 #define N_PHASES                17
 122 #define MAX_TAPS                5
 123
 124 /* memory bufferd overlay registers */
 125 struct overlay_registers {
 126         u32 OBUF_0Y;
 127         u32 OBUF_1Y;
 128         u32 OBUF_0U;
 129         u32 OBUF_0V;
 130         u32 OBUF_1U;
 131         u32 OBUF_1V;
 132         u32 OSTRIDE;
 133         u32 YRGB_VPH;
 134         u32 UV_VPH;
 135         u32 HORZ_PH;
 136         u32 INIT_PHS;
 137         u32 DWINPOS;
 138         u32 DWINSZ;
 139         u32 SWIDTH;
 140         u32 SWIDTHSW;
 141         u32 SHEIGHT;
 142         u32 YRGBSCALE;
 143         u32 UVSCALE;
 144         u32 OCLRC0;
 145         u32 OCLRC1;
 146         u32 DCLRKV;
 147         u32 DCLRKM;
 148         u32 SCLRKVH;
 149         u32 SCLRKVL;
 150         u32 SCLRKEN;
 151         u32 OCONFIG;
 152         u32 OCMD;
 153         u32 RESERVED1; /* 0x6C */
 154         u32 OSTART_0Y;
 155         u32 OSTART_1Y;
 156         u32 OSTART_0U;
 157         u32 OSTART_0V;
 158         u32 OSTART_1U;
 159         u32 OSTART_1V;
 160         u32 OTILEOFF_0Y;
 161         u32 OTILEOFF_1Y;
 162         u32 OTILEOFF_0U;
 163         u32 OTILEOFF_0V;
 164         u32 OTILEOFF_1U;
 165         u32 OTILEOFF_1V;
 166         u32 FASTHSCALE; /* 0xA0 */
 167         u32 UVSCALEV; /* 0xA4 */
 168         u32 RESERVEDC[(0x200 - 0xA8) / 4]; /* 0xA8 - 0x1FC */
 169         u16 Y_VCOEFS[N_VERT_Y_TAPS * N_PHASES]; /* 0x200 */
 170         u16 RESERVEDD[0x100 / 2 - N_VERT_Y_TAPS * N_PHASES];
 171         u16 Y_HCOEFS[N_HORIZ_Y_TAPS * N_PHASES]; /* 0x300 */
 172         u16 RESERVEDE[0x200 / 2 - N_HORIZ_Y_TAPS * N_PHASES];
 173         u16 UV_VCOEFS[N_VERT_UV_TAPS * N_PHASES]; /* 0x500 */
 174         u16 RESERVEDF[0x100 / 2 - N_VERT_UV_TAPS * N_PHASES];
 175         u16 UV_HCOEFS[N_HORIZ_UV_TAPS * N_PHASES]; /* 0x600 */
 176         u16 RESERVEDG[0x100 / 2 - N_HORIZ_UV_TAPS * N_PHASES];
 177 };
 178
 179 struct intel_overlay {
 180         struct drm_i915_private *i915;
 181         struct intel_context *context;
 182         struct intel_crtc *crtc;
 183         struct i915_vma *vma;
 184         struct i915_vma *old_vma;
 185         bool active;
 186         bool pfit_active;
 187         u32 pfit_vscale_ratio; /* shifted-point number, (1<<12) == 1.0 */
 188         u32 color_key:24;
 189         u32 color_key_enabled:1;
 190         u32 brightness, contrast, saturation;
 191         u32 old_xscale, old_yscale;
 192         /* register access */
 193         struct drm_i915_gem_object *reg_bo;
 194         struct overlay_registers __iomem *regs;
 195         u32 flip_addr;
 196         /* flip handling */
 197         struct i915_active last_flip;
 198         void (*flip_complete)(struct intel_overlay *ovl);
 199 };
 200
 201 static void i830_overlay_clock_gating(struct drm_i915_private *dev_priv,
 202                                       bool enable)
 203 {
 204         struct pci_dev *pdev = dev_priv->drm.pdev;
 205         u8 val;
 206
 207         /* WA_OVERLAY_CLKGATE:alm */
 208         if (enable)
 209                 intel_de_write(dev_priv, DSPCLK_GATE_D, 0);
 210         else
 211                 intel_de_write(dev_priv, DSPCLK_GATE_D,
 212                                OVRUNIT_CLOCK_GATE_DISABLE);
 213
 214         /* WA_DISABLE_L2CACHE_CLOCK_GATING:alm */
 215         pci_bus_read_config_byte(pdev->bus,
 216                                  PCI_DEVFN(0, 0), I830_CLOCK_GATE, &val);
 217         if (enable)
 218                 val &= ~I830_L2_CACHE_CLOCK_GATE_DISABLE;
 219         else
 220                 val |= I830_L2_CACHE_CLOCK_GATE_DISABLE;
 221         pci_bus_write_config_byte(pdev->bus,
 222                                   PCI_DEVFN(0, 0), I830_CLOCK_GATE, val);
 223 }
 224
 225 static struct i915_request *
 226 alloc_request(struct intel_overlay *overlay, void (*fn)(struct intel_overlay *))
 227 {
 228         struct i915_request *rq;
 229         int err;
 230
 231         overlay->flip_complete = fn;
 232
 233         rq = i915_request_create(overlay->context);
 234         if (IS_ERR(rq))
 235                 return rq;
 236
 237         err = i915_active_add_request(&overlay->last_flip, rq);
 238         if (err) {
 239                 i915_request_add(rq);
 240                 return ERR_PTR(err);
 241         }
 242
 243         return rq;
 244 }
 245
 246 /* overlay needs to be disable in OCMD reg */
 247 static int intel_overlay_on(struct intel_overlay *overlay)
 248 {
 249         struct drm_i915_private *dev_priv = overlay->i915;
 250         struct i915_request *rq;
 251         u32 *cs;
 252
 253         drm_WARN_ON(&dev_priv->drm, overlay->active);
 254
 255         rq = alloc_request(overlay, NULL);
 256         if (IS_ERR(rq))
 257                 return PTR_ERR(rq);
 258
 259         cs = intel_ring_begin(rq, 4);
 260         if (IS_ERR(cs)) {
 261                 i915_request_add(rq);
 262                 return PTR_ERR(cs);
 263         }
 264
 265         overlay->active = true;
 266
 267         if (IS_I830(dev_priv))
 268                 i830_overlay_clock_gating(dev_priv, false);
 269
 270         *cs++ = MI_OVERLAY_FLIP | MI_OVERLAY_ON;
 271         *cs++ = overlay->flip_addr | OFC_UPDATE;
 272         *cs++ = MI_WAIT_FOR_EVENT | MI_WAIT_FOR_OVERLAY_FLIP;
 273         *cs++ = MI_NOOP;
 274         intel_ring_advance(rq, cs);
 275
 276         i915_request_add(rq);
 277
 278         return i915_active_wait(&overlay->last_flip);
 279 }
 280
 281 static void intel_overlay_flip_prepare(struct intel_overlay *overlay,
 282                                        struct i915_vma *vma)
 283 {
 284         enum pipe pipe = overlay->crtc->pipe;
 285         struct intel_frontbuffer *from = NULL, *to = NULL;
 286
 287         drm_WARN_ON(&overlay->i915->drm, overlay->old_vma);
 288
 289         if (overlay->vma)
 290                 from = intel_frontbuffer_get(overlay->vma->obj);
 291         if (vma)
 292                 to = intel_frontbuffer_get(vma->obj);
 293
 294         intel_frontbuffer_track(from, to, INTEL_FRONTBUFFER_OVERLAY(pipe));
 295
 296         if (to)
 297                 intel_frontbuffer_put(to);
 298         if (from)
 299                 intel_frontbuffer_put(from);
 300
 301         intel_frontbuffer_flip_prepare(overlay->i915,
 302                                        INTEL_FRONTBUFFER_OVERLAY(pipe));
 303
 304         overlay->old_vma = overlay->vma;
 305         if (vma)
 306                 overlay->vma = i915_vma_get(vma);
 307         else
 308                 overlay->vma = NULL;
 309 }
 310
 311 /* overlay needs to be enabled in OCMD reg */
 312 static int intel_overlay_continue(struct intel_overlay *overlay,
 313                                   struct i915_vma *vma,
 314                                   bool load_polyphase_filter)
 315 {
 316         struct drm_i915_private *dev_priv = overlay->i915;
 317         struct i915_request *rq;
 318         u32 flip_addr = overlay->flip_addr;
 319         u32 tmp, *cs;
 320
 321         drm_WARN_ON(&dev_priv->drm, !overlay->active);
 322
 323         if (load_polyphase_filter)
 324                 flip_addr |= OFC_UPDATE;
 325
 326         /* check for underruns */
 327         tmp = intel_de_read(dev_priv, DOVSTA);
 328         if (tmp & (1 << 17))
 329                 drm_dbg(&dev_priv->drm, "overlay underrun, DOVSTA: %x\n", tmp);
 330
 331         rq = alloc_request(overlay, NULL);
 332         if (IS_ERR(rq))
 333                 return PTR_ERR(rq);
 334
 335         cs = intel_ring_begin(rq, 2);
 336         if (IS_ERR(cs)) {
 337                 i915_request_add(rq);
 338                 return PTR_ERR(cs);
 339         }
 340
 341         *cs++ = MI_OVERLAY_FLIP | MI_OVERLAY_CONTINUE;
 342         *cs++ = flip_addr;
 343         intel_ring_advance(rq, cs);
 344
 345         intel_overlay_flip_prepare(overlay, vma);
 346         i915_request_add(rq);
 347
 348         return 0;
 349 }
 350
 351 static void intel_overlay_release_old_vma(struct intel_overlay *overlay)
 352 {
 353         struct i915_vma *vma;
 354
 355         vma = fetch_and_zero(&overlay->old_vma);
 356         if (drm_WARN_ON(&overlay->i915->drm, !vma))
 357                 return;
 358
 359         intel_frontbuffer_flip_complete(overlay->i915,
 360                                         INTEL_FRONTBUFFER_OVERLAY(overlay->crtc->pipe));
 361
 362         i915_gem_object_unpin_from_display_plane(vma);
 363         i915_vma_put(vma);
 364 }
 365
 366 static void
 367 intel_overlay_release_old_vid_tail(struct intel_overlay *overlay)
 368 {
 369         intel_overlay_release_old_vma(overlay);
 370 }
 371
 372 static void intel_overlay_off_tail(struct intel_overlay *overlay)
 373 {
 374         struct drm_i915_private *dev_priv = overlay->i915;
 375
 376         intel_overlay_release_old_vma(overlay);
 377
 378         overlay->crtc->overlay = NULL;
 379         overlay->crtc = NULL;
 380         overlay->active = false;
 381
 382         if (IS_I830(dev_priv))
 383                 i830_overlay_clock_gating(dev_priv, true);
 384 }
 385
 386 static void
 387 intel_overlay_last_flip_retire(struct i915_active *active)
 388 {
 389         struct intel_overlay *overlay =
 390                 container_of(active, typeof(*overlay), last_flip);
 391
 392         if (overlay->flip_complete)
 393                 overlay->flip_complete(overlay);
 394 }
 395
 396 /* overlay needs to be disabled in OCMD reg */
 397 static int intel_overlay_off(struct intel_overlay *overlay)
 398 {
 399         struct i915_request *rq;
 400         u32 *cs, flip_addr = overlay->flip_addr;
 401
 402         drm_WARN_ON(&overlay->i915->drm, !overlay->active);
 403
 404         /* According to intel docs the overlay hw may hang (when switching
 405          * off) without loading the filter coeffs. It is however unclear whether
 406          * this applies to the disabling of the overlay or to the switching off
 407          * of the hw. Do it in both cases */
 408         flip_addr |= OFC_UPDATE;
 409
 410         rq = alloc_request(overlay, intel_overlay_off_tail);
 411         if (IS_ERR(rq))
 412                 return PTR_ERR(rq);
 413
 414         cs = intel_ring_begin(rq, 6);
 415         if (IS_ERR(cs)) {
 416                 i915_request_add(rq);
 417                 return PTR_ERR(cs);
 418         }
 419
 420         /* wait for overlay to go idle */
 421         *cs++ = MI_OVERLAY_FLIP | MI_OVERLAY_CONTINUE;
 422         *cs++ = flip_addr;
 423         *cs++ = MI_WAIT_FOR_EVENT | MI_WAIT_FOR_OVERLAY_FLIP;
 424
 425         /* turn overlay off */
 426         *cs++ = MI_OVERLAY_FLIP | MI_OVERLAY_OFF;
 427         *cs++ = flip_addr;
 428         *cs++ = MI_WAIT_FOR_EVENT | MI_WAIT_FOR_OVERLAY_FLIP;
 429
 430         intel_ring_advance(rq, cs);
 431
 432         intel_overlay_flip_prepare(overlay, NULL);
 433         i915_request_add(rq);
 434
 435         return i915_active_wait(&overlay->last_flip);
 436 }
 437
 438 /* recover from an interruption due to a signal
 439  * We have to be careful not to repeat work forever an make forward progess. */
 440 static int intel_overlay_recover_from_interrupt(struct intel_overlay *overlay)
 441 {
 442         return i915_active_wait(&overlay->last_flip);
 443 }
 444
 445 /* Wait for pending overlay flip and release old frame.
 446  * Needs to be called before the overlay register are changed
 447  * via intel_overlay_(un)map_regs
 448  */
 449 static int intel_overlay_release_old_vid(struct intel_overlay *overlay)
 450 {
 451         struct drm_i915_private *dev_priv = overlay->i915;
 452         struct i915_request *rq;
 453         u32 *cs;
 454
 455         /*
 456          * Only wait if there is actually an old frame to release to
 457          * guarantee forward progress.
 458          */
 459         if (!overlay->old_vma)
 460                 return 0;
 461
 462         if (!(intel_de_read(dev_priv, GEN2_ISR) & I915_OVERLAY_PLANE_FLIP_PENDING_INTERRUPT)) {
 463                 intel_overlay_release_old_vid_tail(overlay);
 464                 return 0;
 465         }
 466
 467         rq = alloc_request(overlay, intel_overlay_release_old_vid_tail);
 468         if (IS_ERR(rq))
 469                 return PTR_ERR(rq);
 470
 471         cs = intel_ring_begin(rq, 2);
 472         if (IS_ERR(cs)) {
 473                 i915_request_add(rq);
 474                 return PTR_ERR(cs);
 475         }
 476
 477         *cs++ = MI_WAIT_FOR_EVENT | MI_WAIT_FOR_OVERLAY_FLIP;
 478         *cs++ = MI_NOOP;
 479         intel_ring_advance(rq, cs);
 480
 481         i915_request_add(rq);
 482
 483         return i915_active_wait(&overlay->last_flip);
 484 }
 485
 486 void intel_overlay_reset(struct drm_i915_private *dev_priv)
 487 {
 488         struct intel_overlay *overlay = dev_priv->overlay;
 489
 490         if (!overlay)
 491                 return;
 492
 493         overlay->old_xscale = 0;
 494         overlay->old_yscale = 0;
 495         overlay->crtc = NULL;
 496         overlay->active = false;
 497 }
 498
 499 static int packed_depth_bytes(u32 format)
 500 {
 501         switch (format & I915_OVERLAY_DEPTH_MASK) {
 502         case I915_OVERLAY_YUV422:
 503                 return 4;
 504         case I915_OVERLAY_YUV411:
 505                 /* return 6; not implemented */
 506         default:
 507                 return -EINVAL;
 508         }
 509 }
 510
 511 static int packed_width_bytes(u32 format, short width)
 512 {
 513         switch (format & I915_OVERLAY_DEPTH_MASK) {
 514         case I915_OVERLAY_YUV422:
 515                 return width << 1;
 516         default:
 517                 return -EINVAL;
 518         }
 519 }
 520
 521 static int uv_hsubsampling(u32 format)
 522 {
 523         switch (format & I915_OVERLAY_DEPTH_MASK) {
 524         case I915_OVERLAY_YUV422:
 525         case I915_OVERLAY_YUV420:
 526                 return 2;
 527         case I915_OVERLAY_YUV411:
 528         case I915_OVERLAY_YUV410:
 529                 return 4;
 530         default:
 531                 return -EINVAL;
 532         }
 533 }
 534
 535 static int uv_vsubsampling(u32 format)
 536 {
 537         switch (format & I915_OVERLAY_DEPTH_MASK) {
 538         case I915_OVERLAY_YUV420:
 539         case I915_OVERLAY_YUV410:
 540                 return 2;
 541         case I915_OVERLAY_YUV422:
 542         case I915_OVERLAY_YUV411:
 543                 return 1;
 544         default:
 545                 return -EINVAL;
 546         }
 547 }
 548
 549 static u32 calc_swidthsw(struct drm_i915_private *dev_priv, u32 offset, u32 width)
 550 {
 551         u32 sw;
 552
 553         if (IS_GEN(dev_priv, 2))
 554                 sw = ALIGN((offset & 31) + width, 32);
 555         else
 556                 sw = ALIGN((offset & 63) + width, 64);
 557
 558         if (sw == 0)
 559                 return 0;
 560
 561         return (sw - 32) >> 3;
 562 }
 563
 564 static const u16 y_static_hcoeffs[N_PHASES][N_HORIZ_Y_TAPS] = {
 565         [ 0] = { 0x3000, 0xb4a0, 0x1930, 0x1920, 0xb4a0, },
 566         [ 1] = { 0x3000, 0xb500, 0x19d0, 0x1880, 0xb440, },
 567         [ 2] = { 0x3000, 0xb540, 0x1a88, 0x2f80, 0xb3e0, },
 568         [ 3] = { 0x3000, 0xb580, 0x1b30, 0x2e20, 0xb380, },
 569         [ 4] = { 0x3000, 0xb5c0, 0x1bd8, 0x2cc0, 0xb320, },
 570         [ 5] = { 0x3020, 0xb5e0, 0x1c60, 0x2b80, 0xb2c0, },
 571         [ 6] = { 0x3020, 0xb5e0, 0x1cf8, 0x2a20, 0xb260, },
 572         [ 7] = { 0x3020, 0xb5e0, 0x1d80, 0x28e0, 0xb200, },
 573         [ 8] = { 0x3020, 0xb5c0, 0x1e08, 0x3f40, 0xb1c0, },
 574         [ 9] = { 0x3020, 0xb580, 0x1e78, 0x3ce0, 0xb160, },
 575         [10] = { 0x3040, 0xb520, 0x1ed8, 0x3aa0, 0xb120, },
 576         [11] = { 0x3040, 0xb4a0, 0x1f30, 0x3880, 0xb0e0, },
 577         [12] = { 0x3040, 0xb400, 0x1f78, 0x3680, 0xb0a0, },
 578         [13] = { 0x3020, 0xb340, 0x1fb8, 0x34a0, 0xb060, },
 579         [14] = { 0x3020, 0xb240, 0x1fe0, 0x32e0, 0xb040, },
 580         [15] = { 0x3020, 0xb140, 0x1ff8, 0x3160, 0xb020, },
 581         [16] = { 0xb000, 0x3000, 0x0800, 0x3000, 0xb000, },
 582 };
 583
 584 static const u16 uv_static_hcoeffs[N_PHASES][N_HORIZ_UV_TAPS] = {
 585         [ 0] = { 0x3000, 0x1800, 0x1800, },
 586         [ 1] = { 0xb000, 0x18d0, 0x2e60, },
 587         [ 2] = { 0xb000, 0x1990, 0x2ce0, },
 588         [ 3] = { 0xb020, 0x1a68, 0x2b40, },
 589         [ 4] = { 0xb040, 0x1b20, 0x29e0, },
 590         [ 5] = { 0xb060, 0x1bd8, 0x2880, },
 591         [ 6] = { 0xb080, 0x1c88, 0x3e60, },
 592         [ 7] = { 0xb0a0, 0x1d28, 0x3c00, },
 593         [ 8] = { 0xb0c0, 0x1db8, 0x39e0, },
 594         [ 9] = { 0xb0e0, 0x1e40, 0x37e0, },
 595         [10] = { 0xb100, 0x1eb8, 0x3620, },
 596         [11] = { 0xb100, 0x1f18, 0x34a0, },
 597         [12] = { 0xb100, 0x1f68, 0x3360, },
 598         [13] = { 0xb0e0, 0x1fa8, 0x3240, },
 599         [14] = { 0xb0c0, 0x1fe0, 0x3140, },
 600         [15] = { 0xb060, 0x1ff0, 0x30a0, },
 601         [16] = { 0x3000, 0x0800, 0x3000, },
 602 };
 603
 604 static void update_polyphase_filter(struct overlay_registers __iomem *regs)
 605 {
 606         memcpy_toio(regs->Y_HCOEFS, y_static_hcoeffs, sizeof(y_static_hcoeffs));
 607         memcpy_toio(regs->UV_HCOEFS, uv_static_hcoeffs,
 608                     sizeof(uv_static_hcoeffs));
 609 }
 610
 611 static bool update_scaling_factors(struct intel_overlay *overlay,
 612                                    struct overlay_registers __iomem *regs,
 613                                    struct drm_intel_overlay_put_image *params)
 614 {
 615         /* fixed point with a 12 bit shift */
 616         u32 xscale, yscale, xscale_UV, yscale_UV;
 617 #define FP_SHIFT 12
 618 #define FRACT_MASK 0xfff
 619         bool scale_changed = false;
 620         int uv_hscale = uv_hsubsampling(params->flags);
 621         int uv_vscale = uv_vsubsampling(params->flags);
 622
 623         if (params->dst_width > 1)
 624                 xscale = ((params->src_scan_width - 1) << FP_SHIFT) /
 625                         params->dst_width;
 626         else
 627                 xscale = 1 << FP_SHIFT;
 628
 629         if (params->dst_height > 1)
 630                 yscale = ((params->src_scan_height - 1) << FP_SHIFT) /
 631                         params->dst_height;
 632         else
 633                 yscale = 1 << FP_SHIFT;
 634
 635         /*if (params->format & I915_OVERLAY_YUV_PLANAR) {*/
 636         xscale_UV = xscale/uv_hscale;
 637         yscale_UV = yscale/uv_vscale;
 638         /* make the Y scale to UV scale ratio an exact multiply */
 639         xscale = xscale_UV * uv_hscale;
 640         yscale = yscale_UV * uv_vscale;
 641         /*} else {
 642           xscale_UV = 0;
 643           yscale_UV = 0;
 644           }*/
 645
 646         if (xscale != overlay->old_xscale || yscale != overlay->old_yscale)
 647                 scale_changed = true;
 648         overlay->old_xscale = xscale;
 649         overlay->old_yscale = yscale;
 650
 651         iowrite32(((yscale & FRACT_MASK) << 20) |
 652                   ((xscale >> FP_SHIFT)  << 16) |
 653                   ((xscale & FRACT_MASK) << 3),
 654                  &regs->YRGBSCALE);
 655
 656         iowrite32(((yscale_UV & FRACT_MASK) << 20) |
 657                   ((xscale_UV >> FP_SHIFT)  << 16) |
 658                   ((xscale_UV & FRACT_MASK) << 3),
 659                  &regs->UVSCALE);
 660
 661         iowrite32((((yscale    >> FP_SHIFT) << 16) |
 662                    ((yscale_UV >> FP_SHIFT) << 0)),
 663                  &regs->UVSCALEV);
 664
 665         if (scale_changed)
 666                 update_polyphase_filter(regs);
 667
 668         return scale_changed;
 669 }
 670
 671 static void update_colorkey(struct intel_overlay *overlay,
 672                             struct overlay_registers __iomem *regs)
 673 {
 674         const struct intel_plane_state *state =
 675                 to_intel_plane_state(overlay->crtc->base.primary->state);
 676         u32 key = overlay->color_key;
 677         u32 format = 0;
 678         u32 flags = 0;
 679
 680         if (overlay->color_key_enabled)
 681                 flags |= DST_KEY_ENABLE;
 682
 683         if (state->uapi.visible)
 684                 format = state->hw.fb->format->format;
 685
 686         switch (format) {
 687         case DRM_FORMAT_C8:
 688                 key = RGB8I_TO_COLORKEY(key);
 689                 flags |= CLK_RGB24_MASK;
 690                 break;
 691         case DRM_FORMAT_XRGB1555:
 692                 key = RGB15_TO_COLORKEY(key);
 693                 flags |= CLK_RGB15_MASK;
 694                 break;
 695         case DRM_FORMAT_RGB565:
 696                 key = RGB16_TO_COLORKEY(key);
 697                 flags |= CLK_RGB16_MASK;
 698                 break;
 699         case DRM_FORMAT_XRGB2101010:
 700         case DRM_FORMAT_XBGR2101010:
 701                 key = RGB30_TO_COLORKEY(key);
 702                 flags |= CLK_RGB24_MASK;
 703                 break;
 704         default:
 705                 flags |= CLK_RGB24_MASK;
 706                 break;
 707         }
 708
 709         iowrite32(key, &regs->DCLRKV);
 710         iowrite32(flags, &regs->DCLRKM);
 711 }
 712
 713 static u32 overlay_cmd_reg(struct drm_intel_overlay_put_image *params)
 714 {
 715         u32 cmd = OCMD_ENABLE | OCMD_BUF_TYPE_FRAME | OCMD_BUFFER0;
 716
 717         if (params->flags & I915_OVERLAY_YUV_PLANAR) {
 718                 switch (params->flags & I915_OVERLAY_DEPTH_MASK) {
 719                 case I915_OVERLAY_YUV422:
 720                         cmd |= OCMD_YUV_422_PLANAR;
 721                         break;
 722                 case I915_OVERLAY_YUV420:
 723                         cmd |= OCMD_YUV_420_PLANAR;
 724                         break;
 725                 case I915_OVERLAY_YUV411:
 726                 case I915_OVERLAY_YUV410:
 727                         cmd |= OCMD_YUV_410_PLANAR;
 728                         break;
 729                 }
 730         } else { /* YUV packed */
 731                 switch (params->flags & I915_OVERLAY_DEPTH_MASK) {
 732                 case I915_OVERLAY_YUV422:
 733                         cmd |= OCMD_YUV_422_PACKED;
 734                         break;
 735                 case I915_OVERLAY_YUV411:
 736                         cmd |= OCMD_YUV_411_PACKED;
 737                         break;
 738                 }
 739
 740                 switch (params->flags & I915_OVERLAY_SWAP_MASK) {
 741                 case I915_OVERLAY_NO_SWAP:
 742                         break;
 743                 case I915_OVERLAY_UV_SWAP:
 744                         cmd |= OCMD_UV_SWAP;
 745                         break;
 746                 case I915_OVERLAY_Y_SWAP:
 747                         cmd |= OCMD_Y_SWAP;
 748                         break;
 749                 case I915_OVERLAY_Y_AND_UV_SWAP:
 750                         cmd |= OCMD_Y_AND_UV_SWAP;
 751                         break;
 752                 }
 753         }
 754
 755         return cmd;
 756 }
 757
 758 static int intel_overlay_do_put_image(struct intel_overlay *overlay,
 759                                       struct drm_i915_gem_object *new_bo,
 760                                       struct drm_intel_overlay_put_image *params)
 761 {
 762         struct overlay_registers __iomem *regs = overlay->regs;
 763         struct drm_i915_private *dev_priv = overlay->i915;
 764         u32 swidth, swidthsw, sheight, ostride;
 765         enum pipe pipe = overlay->crtc->pipe;
 766         bool scale_changed = false;
 767         struct i915_vma *vma;
 768         int ret, tmp_width;
 769
 770         drm_WARN_ON(&dev_priv->drm,
 771                     !drm_modeset_is_locked(&dev_priv->drm.mode_config.connection_mutex));
 772
 773         ret = intel_overlay_release_old_vid(overlay);
 774         if (ret != 0)
 775                 return ret;
 776
 777         atomic_inc(&dev_priv->gpu_error.pending_fb_pin);
 778
 779         vma = i915_gem_object_pin_to_display_plane(new_bo,
 780                                                    0, NULL, PIN_MAPPABLE);
 781         if (IS_ERR(vma)) {
 782                 ret = PTR_ERR(vma);
 783                 goto out_pin_section;
 784         }
 785         i915_gem_object_flush_frontbuffer(new_bo, ORIGIN_DIRTYFB);
 786
 787         if (!overlay->active) {
 788                 const struct intel_crtc_state *crtc_state =
 789                         overlay->crtc->config;
 790                 u32 oconfig = 0;
 791
 792                 if (crtc_state->gamma_enable &&
 793                     crtc_state->gamma_mode == GAMMA_MODE_MODE_8BIT)
 794                         oconfig |= OCONF_CC_OUT_8BIT;
 795                 if (crtc_state->gamma_enable)
 796                         oconfig |= OCONF_GAMMA2_ENABLE;
 797                 if (IS_GEN(dev_priv, 4))
 798                         oconfig |= OCONF_CSC_MODE_BT709;
 799                 oconfig |= pipe == 0 ?
 800                         OCONF_PIPE_A : OCONF_PIPE_B;
 801                 iowrite32(oconfig, &regs->OCONFIG);
 802
 803                 ret = intel_overlay_on(overlay);
 804                 if (ret != 0)
 805                         goto out_unpin;
 806         }
 807
 808         iowrite32(params->dst_y << 16 | params->dst_x, &regs->DWINPOS);
 809         iowrite32(params->dst_height << 16 | params->dst_width, &regs->DWINSZ);
 810
 811         if (params->flags & I915_OVERLAY_YUV_PACKED)
 812                 tmp_width = packed_width_bytes(params->flags,
 813                                                params->src_width);
 814         else
 815                 tmp_width = params->src_width;
 816
 817         swidth = params->src_width;
 818         swidthsw = calc_swidthsw(dev_priv, params->offset_Y, tmp_width);
 819         sheight = params->src_height;
 820         iowrite32(i915_ggtt_offset(vma) + params->offset_Y, &regs->OBUF_0Y);
 821         ostride = params->stride_Y;
 822
 823         if (params->flags & I915_OVERLAY_YUV_PLANAR) {
 824                 int uv_hscale = uv_hsubsampling(params->flags);
 825                 int uv_vscale = uv_vsubsampling(params->flags);
 826                 u32 tmp_U, tmp_V;
 827
 828                 swidth |= (params->src_width / uv_hscale) << 16;
 829                 sheight |= (params->src_height / uv_vscale) << 16;
 830
 831                 tmp_U = calc_swidthsw(dev_priv, params->offset_U,
 832                                       params->src_width / uv_hscale);
 833                 tmp_V = calc_swidthsw(dev_priv, params->offset_V,
 834                                       params->src_width / uv_hscale);
 835                 swidthsw |= max(tmp_U, tmp_V) << 16;
 836
 837                 iowrite32(i915_ggtt_offset(vma) + params->offset_U,
 838                           &regs->OBUF_0U);
 839                 iowrite32(i915_ggtt_offset(vma) + params->offset_V,
 840                           &regs->OBUF_0V);
 841
 842                 ostride |= params->stride_UV << 16;
 843         }
 844
 845         iowrite32(swidth, &regs->SWIDTH);
 846         iowrite32(swidthsw, &regs->SWIDTHSW);
 847         iowrite32(sheight, &regs->SHEIGHT);
 848         iowrite32(ostride, &regs->OSTRIDE);
 849
 850         scale_changed = update_scaling_factors(overlay, regs, params);
 851
 852         update_colorkey(overlay, regs);
 853
 854         iowrite32(overlay_cmd_reg(params), &regs->OCMD);
 855
 856         ret = intel_overlay_continue(overlay, vma, scale_changed);
 857         if (ret)
 858                 goto out_unpin;
 859
 860         return 0;
 861
 862 out_unpin:
 863         i915_gem_object_unpin_from_display_plane(vma);
 864 out_pin_section:
 865         atomic_dec(&dev_priv->gpu_error.pending_fb_pin);
 866
 867         return ret;
 868 }
 869
 870 int intel_overlay_switch_off(struct intel_overlay *overlay)
 871 {
 872         struct drm_i915_private *dev_priv = overlay->i915;
 873         int ret;
 874
 875         drm_WARN_ON(&dev_priv->drm,
 876                     !drm_modeset_is_locked(&dev_priv->drm.mode_config.connection_mutex));
 877
 878         ret = intel_overlay_recover_from_interrupt(overlay);
 879         if (ret != 0)
 880                 return ret;
 881
 882         if (!overlay->active)
 883                 return 0;
 884
 885         ret = intel_overlay_release_old_vid(overlay);
 886         if (ret != 0)
 887                 return ret;
 888
 889         iowrite32(0, &overlay->regs->OCMD);
 890
 891         return intel_overlay_off(overlay);
 892 }
 893
 894 static int check_overlay_possible_on_crtc(struct intel_overlay *overlay,
 895                                           struct intel_crtc *crtc)
 896 {
 897         if (!crtc->active)
 898                 return -EINVAL;
 899
 900         /* can't use the overlay with double wide pipe */
 901         if (crtc->config->double_wide)
 902                 return -EINVAL;
 903
 904         return 0;
 905 }
 906
 907 static void update_pfit_vscale_ratio(struct intel_overlay *overlay)
 908 {
 909         struct drm_i915_private *dev_priv = overlay->i915;
 910         u32 pfit_control = intel_de_read(dev_priv, PFIT_CONTROL);
 911         u32 ratio;
 912
 913         /* XXX: This is not the same logic as in the xorg driver, but more in
 914          * line with the intel documentation for the i965
 915          */
 916         if (INTEL_GEN(dev_priv) >= 4) {
 917                 /* on i965 use the PGM reg to read out the autoscaler values */
 918                 ratio = intel_de_read(dev_priv, PFIT_PGM_RATIOS) >> PFIT_VERT_SCALE_SHIFT_965;
 919         } else {
 920                 if (pfit_control & VERT_AUTO_SCALE)
 921                         ratio = intel_de_read(dev_priv, PFIT_AUTO_RATIOS);
 922                 else
 923                         ratio = intel_de_read(dev_priv, PFIT_PGM_RATIOS);
 924                 ratio >>= PFIT_VERT_SCALE_SHIFT;
 925         }
 926
 927         overlay->pfit_vscale_ratio = ratio;
 928 }
 929
 930 static int check_overlay_dst(struct intel_overlay *overlay,
 931                              struct drm_intel_overlay_put_image *rec)
 932 {
 933         const struct intel_crtc_state *pipe_config =
 934                 overlay->crtc->config;
 935
 936         if (rec->dst_x < pipe_config->pipe_src_w &&
 937             rec->dst_x + rec->dst_width <= pipe_config->pipe_src_w &&
 938             rec->dst_y < pipe_config->pipe_src_h &&
 939             rec->dst_y + rec->dst_height <= pipe_config->pipe_src_h)
 940                 return 0;
 941         else
 942                 return -EINVAL;
 943 }
 944
 945 static int check_overlay_scaling(struct drm_intel_overlay_put_image *rec)
 946 {
 947         u32 tmp;
 948
 949         /* downscaling limit is 8.0 */
 950         tmp = ((rec->src_scan_height << 16) / rec->dst_height) >> 16;
 951         if (tmp > 7)
 952                 return -EINVAL;
 953
 954         tmp = ((rec->src_scan_width << 16) / rec->dst_width) >> 16;
 955         if (tmp > 7)
 956                 return -EINVAL;
 957
 958         return 0;
 959 }
 960
 961 static int check_overlay_src(struct drm_i915_private *dev_priv,
 962                              struct drm_intel_overlay_put_image *rec,
 963                              struct drm_i915_gem_object *new_bo)
 964 {
 965         int uv_hscale = uv_hsubsampling(rec->flags);
 966         int uv_vscale = uv_vsubsampling(rec->flags);
 967         u32 stride_mask;
 968         int depth;
 969         u32 tmp;
 970
 971         /* check src dimensions */
 972         if (IS_I845G(dev_priv) || IS_I830(dev_priv)) {
 973                 if (rec->src_height > IMAGE_MAX_HEIGHT_LEGACY ||
 974                     rec->src_width  > IMAGE_MAX_WIDTH_LEGACY)
 975                         return -EINVAL;
 976         } else {
 977                 if (rec->src_height > IMAGE_MAX_HEIGHT ||
 978                     rec->src_width  > IMAGE_MAX_WIDTH)
 979                         return -EINVAL;
 980         }
 981
 982         /* better safe than sorry, use 4 as the maximal subsampling ratio */
 983         if (rec->src_height < N_VERT_Y_TAPS*4 ||
 984             rec->src_width  < N_HORIZ_Y_TAPS*4)
 985                 return -EINVAL;
 986
 987         /* check alignment constraints */
 988         switch (rec->flags & I915_OVERLAY_TYPE_MASK) {
 989         case I915_OVERLAY_RGB:
 990                 /* not implemented */
 991                 return -EINVAL;
 992
 993         case I915_OVERLAY_YUV_PACKED:
 994                 if (uv_vscale != 1)
 995                         return -EINVAL;
 996
 997                 depth = packed_depth_bytes(rec->flags);
 998                 if (depth < 0)
 999                         return depth;
1000
1001                 /* ignore UV planes */
1002                 rec->stride_UV = 0;
1003                 rec->offset_U = 0;
1004                 rec->offset_V = 0;
1005                 /* check pixel alignment */
1006                 if (rec->offset_Y % depth)
1007                         return -EINVAL;
1008                 break;
1009
1010         case I915_OVERLAY_YUV_PLANAR:
1011                 if (uv_vscale < 0 || uv_hscale < 0)
1012                         return -EINVAL;
1013                 /* no offset restrictions for planar formats */
1014                 break;
1015
1016         default:
1017                 return -EINVAL;
1018         }
1019
1020         if (rec->src_width % uv_hscale)
1021                 return -EINVAL;
1022
1023         /* stride checking */
1024         if (IS_I830(dev_priv) || IS_I845G(dev_priv))
1025                 stride_mask = 255;
1026         else
1027                 stride_mask = 63;
1028
1029         if (rec->stride_Y & stride_mask || rec->stride_UV & stride_mask)
1030                 return -EINVAL;
1031         if (IS_GEN(dev_priv, 4) && rec->stride_Y < 512)
1032                 return -EINVAL;
1033
1034         tmp = (rec->flags & I915_OVERLAY_TYPE_MASK) == I915_OVERLAY_YUV_PLANAR ?
1035                 4096 : 8192;
1036         if (rec->stride_Y > tmp || rec->stride_UV > 2*1024)
1037                 return -EINVAL;
1038
1039         /* check buffer dimensions */
1040         switch (rec->flags & I915_OVERLAY_TYPE_MASK) {
1041         case I915_OVERLAY_RGB:
1042         case I915_OVERLAY_YUV_PACKED:
1043                 /* always 4 Y values per depth pixels */
1044                 if (packed_width_bytes(rec->flags, rec->src_width) > rec->stride_Y)
1045                         return -EINVAL;
1046
1047                 tmp = rec->stride_Y*rec->src_height;
1048                 if (rec->offset_Y + tmp > new_bo->base.size)
1049                         return -EINVAL;
1050                 break;
1051
1052         case I915_OVERLAY_YUV_PLANAR:
1053                 if (rec->src_width > rec->stride_Y)
1054                         return -EINVAL;
1055                 if (rec->src_width/uv_hscale > rec->stride_UV)
1056                         return -EINVAL;
1057
1058                 tmp = rec->stride_Y * rec->src_height;
1059                 if (rec->offset_Y + tmp > new_bo->base.size)
1060                         return -EINVAL;
1061
1062                 tmp = rec->stride_UV * (rec->src_height / uv_vscale);
1063                 if (rec->offset_U + tmp > new_bo->base.size ||
1064                     rec->offset_V + tmp > new_bo->base.size)
1065                         return -EINVAL;
1066                 break;
1067         }
1068
1069         return 0;
1070 }
1071
1072 int intel_overlay_put_image_ioctl(struct drm_device *dev, void *data,
1073                                   struct drm_file *file_priv)
1074 {
1075         struct drm_intel_overlay_put_image *params = data;
1076         struct drm_i915_private *dev_priv = to_i915(dev);
1077         struct intel_overlay *overlay;
1078         struct drm_crtc *drmmode_crtc;
1079         struct intel_crtc *crtc;
1080         struct drm_i915_gem_object *new_bo;
1081         int ret;
1082
1083         overlay = dev_priv->overlay;
1084         if (!overlay) {
1085                 drm_dbg(&dev_priv->drm, "userspace bug: no overlay\n");
1086                 return -ENODEV;
1087         }
1088
1089         if (!(params->flags & I915_OVERLAY_ENABLE)) {
1090                 drm_modeset_lock_all(dev);
1091                 ret = intel_overlay_switch_off(overlay);
1092                 drm_modeset_unlock_all(dev);
1093
1094                 return ret;
1095         }
1096
1097         drmmode_crtc = drm_crtc_find(dev, file_priv, params->crtc_id);
1098         if (!drmmode_crtc)
1099                 return -ENOENT;
1100         crtc = to_intel_crtc(drmmode_crtc);
1101
1102         new_bo = i915_gem_object_lookup(file_priv, params->bo_handle);
1103         if (!new_bo)
1104                 return -ENOENT;
1105
1106         drm_modeset_lock_all(dev);
1107
1108         if (i915_gem_object_is_tiled(new_bo)) {
1109                 drm_dbg_kms(&dev_priv->drm,
1110                             "buffer used for overlay image can not be tiled\n");
1111                 ret = -EINVAL;
1112                 goto out_unlock;
1113         }
1114
1115         ret = intel_overlay_recover_from_interrupt(overlay);
1116         if (ret != 0)
1117                 goto out_unlock;
1118
1119         if (overlay->crtc != crtc) {
1120                 ret = intel_overlay_switch_off(overlay);
1121                 if (ret != 0)
1122                         goto out_unlock;
1123
1124                 ret = check_overlay_possible_on_crtc(overlay, crtc);
1125                 if (ret != 0)
1126                         goto out_unlock;
1127
1128                 overlay->crtc = crtc;
1129                 crtc->overlay = overlay;
1130
1131                 /* line too wide, i.e. one-line-mode */
1132                 if (crtc->config->pipe_src_w > 1024 &&
1133                     crtc->config->gmch_pfit.control & PFIT_ENABLE) {
1134                         overlay->pfit_active = true;
1135                         update_pfit_vscale_ratio(overlay);
1136                 } else
1137                         overlay->pfit_active = false;
1138         }
1139
1140         ret = check_overlay_dst(overlay, params);
1141         if (ret != 0)
1142                 goto out_unlock;
1143
1144         if (overlay->pfit_active) {
1145                 params->dst_y = (((u32)params->dst_y << 12) /
1146                                  overlay->pfit_vscale_ratio);
1147                 /* shifting right rounds downwards, so add 1 */
1148                 params->dst_height = (((u32)params->dst_height << 12) /
1149                                  overlay->pfit_vscale_ratio) + 1;
1150         }
1151
1152         if (params->src_scan_height > params->src_height ||
1153             params->src_scan_width > params->src_width) {
1154                 ret = -EINVAL;
1155                 goto out_unlock;
1156         }
1157
1158         ret = check_overlay_src(dev_priv, params, new_bo);
1159         if (ret != 0)
1160                 goto out_unlock;
1161
1162         /* Check scaling after src size to prevent a divide-by-zero. */
1163         ret = check_overlay_scaling(params);
1164         if (ret != 0)
1165                 goto out_unlock;
1166
1167         ret = intel_overlay_do_put_image(overlay, new_bo, params);
1168         if (ret != 0)
1169                 goto out_unlock;
1170
1171         drm_modeset_unlock_all(dev);
1172         i915_gem_object_put(new_bo);
1173
1174         return 0;
1175
1176 out_unlock:
1177         drm_modeset_unlock_all(dev);
1178         i915_gem_object_put(new_bo);
1179
1180         return ret;
1181 }
1182
1183 static void update_reg_attrs(struct intel_overlay *overlay,
1184                              struct overlay_registers __iomem *regs)
1185 {
1186         iowrite32((overlay->contrast << 18) | (overlay->brightness & 0xff),
1187                   &regs->OCLRC0);
1188         iowrite32(overlay->saturation, &regs->OCLRC1);
1189 }
1190
1191 static bool check_gamma_bounds(u32 gamma1, u32 gamma2)
1192 {
1193         int i;
1194
1195         if (gamma1 & 0xff000000 || gamma2 & 0xff000000)
1196                 return false;
1197
1198         for (i = 0; i < 3; i++) {
1199                 if (((gamma1 >> i*8) & 0xff) >= ((gamma2 >> i*8) & 0xff))
1200                         return false;
1201         }
1202
1203         return true;
1204 }
1205
1206 static bool check_gamma5_errata(u32 gamma5)
1207 {
1208         int i;
1209
1210         for (i = 0; i < 3; i++) {
1211                 if (((gamma5 >> i*8) & 0xff) == 0x80)
1212                         return false;
1213         }
1214
1215         return true;
1216 }
1217
1218 static int check_gamma(struct drm_intel_overlay_attrs *attrs)
1219 {
1220         if (!check_gamma_bounds(0, attrs->gamma0) ||
1221             !check_gamma_bounds(attrs->gamma0, attrs->gamma1) ||
1222             !check_gamma_bounds(attrs->gamma1, attrs->gamma2) ||
1223             !check_gamma_bounds(attrs->gamma2, attrs->gamma3) ||
1224             !check_gamma_bounds(attrs->gamma3, attrs->gamma4) ||
1225             !check_gamma_bounds(attrs->gamma4, attrs->gamma5) ||
1226             !check_gamma_bounds(attrs->gamma5, 0x00ffffff))
1227                 return -EINVAL;
1228
1229         if (!check_gamma5_errata(attrs->gamma5))
1230                 return -EINVAL;
1231
1232         return 0;
1233 }
1234
1235 int intel_overlay_attrs_ioctl(struct drm_device *dev, void *data,
1236                               struct drm_file *file_priv)
1237 {
1238         struct drm_intel_overlay_attrs *attrs = data;
1239         struct drm_i915_private *dev_priv = to_i915(dev);
1240         struct intel_overlay *overlay;
1241         int ret;
1242
1243         overlay = dev_priv->overlay;
1244         if (!overlay) {
1245                 drm_dbg(&dev_priv->drm, "userspace bug: no overlay\n");
1246                 return -ENODEV;
1247         }
1248
1249         drm_modeset_lock_all(dev);
1250
1251         ret = -EINVAL;
1252         if (!(attrs->flags & I915_OVERLAY_UPDATE_ATTRS)) {
1253                 attrs->color_key  = overlay->color_key;
1254                 attrs->brightness = overlay->brightness;
1255                 attrs->contrast   = overlay->contrast;
1256                 attrs->saturation = overlay->saturation;
1257
1258                 if (!IS_GEN(dev_priv, 2)) {
1259                         attrs->gamma0 = intel_de_read(dev_priv, OGAMC0);
1260                         attrs->gamma1 = intel_de_read(dev_priv, OGAMC1);
1261                         attrs->gamma2 = intel_de_read(dev_priv, OGAMC2);
1262                         attrs->gamma3 = intel_de_read(dev_priv, OGAMC3);
1263                         attrs->gamma4 = intel_de_read(dev_priv, OGAMC4);
1264                         attrs->gamma5 = intel_de_read(dev_priv, OGAMC5);
1265                 }
1266         } else {
1267                 if (attrs->brightness < -128 || attrs->brightness > 127)
1268                         goto out_unlock;
1269                 if (attrs->contrast > 255)
1270                         goto out_unlock;
1271                 if (attrs->saturation > 1023)
1272                         goto out_unlock;
1273
1274                 overlay->color_key  = attrs->color_key;
1275                 overlay->brightness = attrs->brightness;
1276                 overlay->contrast   = attrs->contrast;
1277                 overlay->saturation = attrs->saturation;
1278
1279                 update_reg_attrs(overlay, overlay->regs);
1280
1281                 if (attrs->flags & I915_OVERLAY_UPDATE_GAMMA) {
1282                         if (IS_GEN(dev_priv, 2))
1283                                 goto out_unlock;
1284
1285                         if (overlay->active) {
1286                                 ret = -EBUSY;
1287                                 goto out_unlock;
1288                         }
1289
1290                         ret = check_gamma(attrs);
1291                         if (ret)
1292                                 goto out_unlock;
1293
1294                         intel_de_write(dev_priv, OGAMC0, attrs->gamma0);
1295                         intel_de_write(dev_priv, OGAMC1, attrs->gamma1);
1296                         intel_de_write(dev_priv, OGAMC2, attrs->gamma2);
1297                         intel_de_write(dev_priv, OGAMC3, attrs->gamma3);
1298                         intel_de_write(dev_priv, OGAMC4, attrs->gamma4);
1299                         intel_de_write(dev_priv, OGAMC5, attrs->gamma5);
1300                 }
1301         }
1302         overlay->color_key_enabled = (attrs->flags & I915_OVERLAY_DISABLE_DEST_COLORKEY) == 0;
1303
1304         ret = 0;
1305 out_unlock:
1306         drm_modeset_unlock_all(dev);
1307
1308         return ret;
1309 }
1310
1311 static int get_registers(struct intel_overlay *overlay, bool use_phys)
1312 {
1313         struct drm_i915_private *i915 = overlay->i915;
1314         struct drm_i915_gem_object *obj;
1315         struct i915_vma *vma;
1316         int err;
1317
1318         obj = i915_gem_object_create_stolen(i915, PAGE_SIZE);
1319         if (IS_ERR(obj))
1320                 obj = i915_gem_object_create_internal(i915, PAGE_SIZE);
1321         if (IS_ERR(obj))
1322                 return PTR_ERR(obj);
1323
1324         vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0, PIN_MAPPABLE);
1325         if (IS_ERR(vma)) {
1326                 err = PTR_ERR(vma);
1327                 goto err_put_bo;
1328         }
1329
1330         if (use_phys)
1331                 overlay->flip_addr = sg_dma_address(obj->mm.pages->sgl);
1332         else
1333                 overlay->flip_addr = i915_ggtt_offset(vma);
1334         overlay->regs = i915_vma_pin_iomap(vma);
1335         i915_vma_unpin(vma);
1336
1337         if (IS_ERR(overlay->regs)) {
1338                 err = PTR_ERR(overlay->regs);
1339                 goto err_put_bo;
1340         }
1341
1342         overlay->reg_bo = obj;
1343         return 0;
1344
1345 err_put_bo:
1346         i915_gem_object_put(obj);
1347         return err;
1348 }
1349
1350 void intel_overlay_setup(struct drm_i915_private *dev_priv)
1351 {
1352         struct intel_overlay *overlay;
1353         struct intel_engine_cs *engine;
1354         int ret;
1355
1356         if (!HAS_OVERLAY(dev_priv))
1357                 return;
1358
1359         engine = dev_priv->gt.engine[RCS0];
1360         if (!engine || !engine->kernel_context)
1361                 return;
1362
1363         overlay = kzalloc(sizeof(*overlay), GFP_KERNEL);
1364         if (!overlay)
1365                 return;
1366
1367         overlay->i915 = dev_priv;
1368         overlay->context = engine->kernel_context;
1369         GEM_BUG_ON(!overlay->context);
1370
1371         overlay->color_key = 0x0101fe;
1372         overlay->color_key_enabled = true;
1373         overlay->brightness = -19;
1374         overlay->contrast = 75;
1375         overlay->saturation = 146;
1376
1377         i915_active_init(&overlay->last_flip,
1378                          NULL, intel_overlay_last_flip_retire);
1379
1380         ret = get_registers(overlay, OVERLAY_NEEDS_PHYSICAL(dev_priv));
1381         if (ret)
1382                 goto out_free;
1383
1384         memset_io(overlay->regs, 0, sizeof(struct overlay_registers));
1385         update_polyphase_filter(overlay->regs);
1386         update_reg_attrs(overlay, overlay->regs);
1387
1388         dev_priv->overlay = overlay;
1389         drm_info(&dev_priv->drm, "Initialized overlay support.\n");
1390         return;
1391
1392 out_free:
1393         kfree(overlay);
1394 }
1395
1396 void intel_overlay_cleanup(struct drm_i915_private *dev_priv)
1397 {
1398         struct intel_overlay *overlay;
1399
1400         overlay = fetch_and_zero(&dev_priv->overlay);
1401         if (!overlay)
1402                 return;
1403
1404         /*
1405          * The bo's should be free'd by the generic code already.
1406          * Furthermore modesetting teardown happens beforehand so the
1407          * hardware should be off already.
1408          */
1409         drm_WARN_ON(&dev_priv->drm, overlay->active);
1410
1411         i915_gem_object_put(overlay->reg_bo);
1412         i915_active_fini(&overlay->last_flip);
1413
1414         kfree(overlay);
1415 }
1416
1417 #if IS_ENABLED(CONFIG_DRM_I915_CAPTURE_ERROR)
1418
1419 struct intel_overlay_error_state {
1420         struct overlay_registers regs;
1421         unsigned long base;
1422         u32 dovsta;
1423         u32 isr;
1424 };
1425
1426 struct intel_overlay_error_state *
1427 intel_overlay_capture_error_state(struct drm_i915_private *dev_priv)
1428 {
1429         struct intel_overlay *overlay = dev_priv->overlay;
1430         struct intel_overlay_error_state *error;
1431
1432         if (!overlay || !overlay->active)
1433                 return NULL;
1434
1435         error = kmalloc(sizeof(*error), GFP_ATOMIC);
1436         if (error == NULL)
1437                 return NULL;
1438
1439         error->dovsta = intel_de_read(dev_priv, DOVSTA);
1440         error->isr = intel_de_read(dev_priv, GEN2_ISR);
1441         error->base = overlay->flip_addr;
1442
1443         memcpy_fromio(&error->regs, overlay->regs, sizeof(error->regs));
1444
1445         return error;
1446 }
1447
1448 void
1449 intel_overlay_print_error_state(struct drm_i915_error_state_buf *m,
1450                                 struct intel_overlay_error_state *error)
1451 {
1452         i915_error_printf(m, "Overlay, status: 0x%08x, interrupt: 0x%08x\n",
1453                           error->dovsta, error->isr);
1454         i915_error_printf(m, "  Register file at 0x%08lx:\n",
1455                           error->base);
1456
1457 #define P(x) i915_error_printf(m, "    " #x ":  0x%08x\n", error->regs.x)
1458         P(OBUF_0Y);
1459         P(OBUF_1Y);
1460         P(OBUF_0U);
1461         P(OBUF_0V);
1462         P(OBUF_1U);
1463         P(OBUF_1V);
1464         P(OSTRIDE);
1465         P(YRGB_VPH);
1466         P(UV_VPH);
1467         P(HORZ_PH);
1468         P(INIT_PHS);
1469         P(DWINPOS);
1470         P(DWINSZ);
1471         P(SWIDTH);
1472         P(SWIDTHSW);
1473         P(SHEIGHT);
1474         P(YRGBSCALE);
1475         P(UVSCALE);
1476         P(OCLRC0);
1477         P(OCLRC1);
1478         P(DCLRKV);
1479         P(DCLRKM);
1480         P(SCLRKVH);
1481         P(SCLRKVL);
1482         P(SCLRKEN);
1483         P(OCONFIG);
1484         P(OCMD);
1485         P(OSTART_0Y);
1486         P(OSTART_1Y);
1487         P(OSTART_0U);
1488         P(OSTART_0V);
1489         P(OSTART_1U);
1490         P(OSTART_1V);
1491         P(OTILEOFF_0Y);
1492         P(OTILEOFF_1Y);
1493         P(OTILEOFF_0U);
1494         P(OTILEOFF_0V);
1495         P(OTILEOFF_1U);
1496         P(OTILEOFF_1V);
1497         P(FASTHSCALE);
1498         P(UVSCALEV);
1499 #undef P
1500 }
1501
1502 #endif