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 #include <drm/i915_drm.h>
  31
  32 #include "gem/i915_gem_pm.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 #define CLK_RGB8I_MASK          0xffffff
 104
 105 #define RGB16_TO_COLORKEY(c) \
 106         (((c & 0xF800) << 8) | ((c & 0x07E0) << 5) | ((c & 0x001F) << 3))
 107 #define RGB15_TO_COLORKEY(c) \
 108         (((c & 0x7c00) << 9) | ((c & 0x03E0) << 6) | ((c & 0x001F) << 3))
 109
 110 /* overlay flip addr flag */
 111 #define OFC_UPDATE              0x1
 112
 113 /* polyphase filter coefficients */
 114 #define N_HORIZ_Y_TAPS          5
 115 #define N_VERT_Y_TAPS           3
 116 #define N_HORIZ_UV_TAPS         3
 117 #define N_VERT_UV_TAPS          3
 118 #define N_PHASES                17
 119 #define MAX_TAPS                5
 120
 121 /* memory bufferd overlay registers */
 122 struct overlay_registers {
 123         u32 OBUF_0Y;
 124         u32 OBUF_1Y;
 125         u32 OBUF_0U;
 126         u32 OBUF_0V;
 127         u32 OBUF_1U;
 128         u32 OBUF_1V;
 129         u32 OSTRIDE;
 130         u32 YRGB_VPH;
 131         u32 UV_VPH;
 132         u32 HORZ_PH;
 133         u32 INIT_PHS;
 134         u32 DWINPOS;
 135         u32 DWINSZ;
 136         u32 SWIDTH;
 137         u32 SWIDTHSW;
 138         u32 SHEIGHT;
 139         u32 YRGBSCALE;
 140         u32 UVSCALE;
 141         u32 OCLRC0;
 142         u32 OCLRC1;
 143         u32 DCLRKV;
 144         u32 DCLRKM;
 145         u32 SCLRKVH;
 146         u32 SCLRKVL;
 147         u32 SCLRKEN;
 148         u32 OCONFIG;
 149         u32 OCMD;
 150         u32 RESERVED1; /* 0x6C */
 151         u32 OSTART_0Y;
 152         u32 OSTART_1Y;
 153         u32 OSTART_0U;
 154         u32 OSTART_0V;
 155         u32 OSTART_1U;
 156         u32 OSTART_1V;
 157         u32 OTILEOFF_0Y;
 158         u32 OTILEOFF_1Y;
 159         u32 OTILEOFF_0U;
 160         u32 OTILEOFF_0V;
 161         u32 OTILEOFF_1U;
 162         u32 OTILEOFF_1V;
 163         u32 FASTHSCALE; /* 0xA0 */
 164         u32 UVSCALEV; /* 0xA4 */
 165         u32 RESERVEDC[(0x200 - 0xA8) / 4]; /* 0xA8 - 0x1FC */
 166         u16 Y_VCOEFS[N_VERT_Y_TAPS * N_PHASES]; /* 0x200 */
 167         u16 RESERVEDD[0x100 / 2 - N_VERT_Y_TAPS * N_PHASES];
 168         u16 Y_HCOEFS[N_HORIZ_Y_TAPS * N_PHASES]; /* 0x300 */
 169         u16 RESERVEDE[0x200 / 2 - N_HORIZ_Y_TAPS * N_PHASES];
 170         u16 UV_VCOEFS[N_VERT_UV_TAPS * N_PHASES]; /* 0x500 */
 171         u16 RESERVEDF[0x100 / 2 - N_VERT_UV_TAPS * N_PHASES];
 172         u16 UV_HCOEFS[N_HORIZ_UV_TAPS * N_PHASES]; /* 0x600 */
 173         u16 RESERVEDG[0x100 / 2 - N_HORIZ_UV_TAPS * N_PHASES];
 174 };
 175
 176 struct intel_overlay {
 177         struct drm_i915_private *i915;
 178         struct intel_context *context;
 179         struct intel_crtc *crtc;
 180         struct i915_vma *vma;
 181         struct i915_vma *old_vma;
 182         bool active;
 183         bool pfit_active;
 184         u32 pfit_vscale_ratio; /* shifted-point number, (1<<12) == 1.0 */
 185         u32 color_key:24;
 186         u32 color_key_enabled:1;
 187         u32 brightness, contrast, saturation;
 188         u32 old_xscale, old_yscale;
 189         /* register access */
 190         struct drm_i915_gem_object *reg_bo;
 191         struct overlay_registers __iomem *regs;
 192         u32 flip_addr;
 193         /* flip handling */
 194         struct i915_active last_flip;
 195         void (*flip_complete)(struct intel_overlay *ovl);
 196 };
 197
 198 static void i830_overlay_clock_gating(struct drm_i915_private *dev_priv,
 199                                       bool enable)
 200 {
 201         struct pci_dev *pdev = dev_priv->drm.pdev;
 202         u8 val;
 203
 204         /* WA_OVERLAY_CLKGATE:alm */
 205         if (enable)
 206                 I915_WRITE(DSPCLK_GATE_D, 0);
 207         else
 208                 I915_WRITE(DSPCLK_GATE_D, OVRUNIT_CLOCK_GATE_DISABLE);
 209
 210         /* WA_DISABLE_L2CACHE_CLOCK_GATING:alm */
 211         pci_bus_read_config_byte(pdev->bus,
 212                                  PCI_DEVFN(0, 0), I830_CLOCK_GATE, &val);
 213         if (enable)
 214                 val &= ~I830_L2_CACHE_CLOCK_GATE_DISABLE;
 215         else
 216                 val |= I830_L2_CACHE_CLOCK_GATE_DISABLE;
 217         pci_bus_write_config_byte(pdev->bus,
 218                                   PCI_DEVFN(0, 0), I830_CLOCK_GATE, val);
 219 }
 220
 221 static struct i915_request *
 222 alloc_request(struct intel_overlay *overlay, void (*fn)(struct intel_overlay *))
 223 {
 224         struct i915_request *rq;
 225         int err;
 226
 227         overlay->flip_complete = fn;
 228
 229         rq = i915_request_create(overlay->context);
 230         if (IS_ERR(rq))
 231                 return rq;
 232
 233         err = i915_active_add_request(&overlay->last_flip, rq);
 234         if (err) {
 235                 i915_request_add(rq);
 236                 return ERR_PTR(err);
 237         }
 238
 239         return rq;
 240 }
 241
 242 /* overlay needs to be disable in OCMD reg */
 243 static int intel_overlay_on(struct intel_overlay *overlay)
 244 {
 245         struct drm_i915_private *dev_priv = overlay->i915;
 246         struct i915_request *rq;
 247         u32 *cs;
 248
 249         WARN_ON(overlay->active);
 250
 251         rq = alloc_request(overlay, NULL);
 252         if (IS_ERR(rq))
 253                 return PTR_ERR(rq);
 254
 255         cs = intel_ring_begin(rq, 4);
 256         if (IS_ERR(cs)) {
 257                 i915_request_add(rq);
 258                 return PTR_ERR(cs);
 259         }
 260
 261         overlay->active = true;
 262
 263         if (IS_I830(dev_priv))
 264                 i830_overlay_clock_gating(dev_priv, false);
 265
 266         *cs++ = MI_OVERLAY_FLIP | MI_OVERLAY_ON;
 267         *cs++ = overlay->flip_addr | OFC_UPDATE;
 268         *cs++ = MI_WAIT_FOR_EVENT | MI_WAIT_FOR_OVERLAY_FLIP;
 269         *cs++ = MI_NOOP;
 270         intel_ring_advance(rq, cs);
 271
 272         i915_request_add(rq);
 273
 274         return i915_active_wait(&overlay->last_flip);
 275 }
 276
 277 static void intel_overlay_flip_prepare(struct intel_overlay *overlay,
 278                                        struct i915_vma *vma)
 279 {
 280         enum pipe pipe = overlay->crtc->pipe;
 281
 282         WARN_ON(overlay->old_vma);
 283
 284         intel_frontbuffer_track(overlay->vma ? overlay->vma->obj->frontbuffer : NULL,
 285                                 vma ? vma->obj->frontbuffer : NULL,
 286                                 INTEL_FRONTBUFFER_OVERLAY(pipe));
 287
 288         intel_frontbuffer_flip_prepare(overlay->i915,
 289                                        INTEL_FRONTBUFFER_OVERLAY(pipe));
 290
 291         overlay->old_vma = overlay->vma;
 292         if (vma)
 293                 overlay->vma = i915_vma_get(vma);
 294         else
 295                 overlay->vma = NULL;
 296 }
 297
 298 /* overlay needs to be enabled in OCMD reg */
 299 static int intel_overlay_continue(struct intel_overlay *overlay,
 300                                   struct i915_vma *vma,
 301                                   bool load_polyphase_filter)
 302 {
 303         struct drm_i915_private *dev_priv = overlay->i915;
 304         struct i915_request *rq;
 305         u32 flip_addr = overlay->flip_addr;
 306         u32 tmp, *cs;
 307
 308         WARN_ON(!overlay->active);
 309
 310         if (load_polyphase_filter)
 311                 flip_addr |= OFC_UPDATE;
 312
 313         /* check for underruns */
 314         tmp = I915_READ(DOVSTA);
 315         if (tmp & (1 << 17))
 316                 DRM_DEBUG("overlay underrun, DOVSTA: %x\n", tmp);
 317
 318         rq = alloc_request(overlay, NULL);
 319         if (IS_ERR(rq))
 320                 return PTR_ERR(rq);
 321
 322         cs = intel_ring_begin(rq, 2);
 323         if (IS_ERR(cs)) {
 324                 i915_request_add(rq);
 325                 return PTR_ERR(cs);
 326         }
 327
 328         *cs++ = MI_OVERLAY_FLIP | MI_OVERLAY_CONTINUE;
 329         *cs++ = flip_addr;
 330         intel_ring_advance(rq, cs);
 331
 332         intel_overlay_flip_prepare(overlay, vma);
 333         i915_request_add(rq);
 334
 335         return 0;
 336 }
 337
 338 static void intel_overlay_release_old_vma(struct intel_overlay *overlay)
 339 {
 340         struct i915_vma *vma;
 341
 342         vma = fetch_and_zero(&overlay->old_vma);
 343         if (WARN_ON(!vma))
 344                 return;
 345
 346         intel_frontbuffer_flip_complete(overlay->i915,
 347                                         INTEL_FRONTBUFFER_OVERLAY(overlay->crtc->pipe));
 348
 349         i915_gem_object_unpin_from_display_plane(vma);
 350         i915_vma_put(vma);
 351 }
 352
 353 static void
 354 intel_overlay_release_old_vid_tail(struct intel_overlay *overlay)
 355 {
 356         intel_overlay_release_old_vma(overlay);
 357 }
 358
 359 static void intel_overlay_off_tail(struct intel_overlay *overlay)
 360 {
 361         struct drm_i915_private *dev_priv = overlay->i915;
 362
 363         intel_overlay_release_old_vma(overlay);
 364
 365         overlay->crtc->overlay = NULL;
 366         overlay->crtc = NULL;
 367         overlay->active = false;
 368
 369         if (IS_I830(dev_priv))
 370                 i830_overlay_clock_gating(dev_priv, true);
 371 }
 372
 373 static void
 374 intel_overlay_last_flip_retire(struct i915_active *active)
 375 {
 376         struct intel_overlay *overlay =
 377                 container_of(active, typeof(*overlay), last_flip);
 378
 379         if (overlay->flip_complete)
 380                 overlay->flip_complete(overlay);
 381 }
 382
 383 /* overlay needs to be disabled in OCMD reg */
 384 static int intel_overlay_off(struct intel_overlay *overlay)
 385 {
 386         struct i915_request *rq;
 387         u32 *cs, flip_addr = overlay->flip_addr;
 388
 389         WARN_ON(!overlay->active);
 390
 391         /* According to intel docs the overlay hw may hang (when switching
 392          * off) without loading the filter coeffs. It is however unclear whether
 393          * this applies to the disabling of the overlay or to the switching off
 394          * of the hw. Do it in both cases */
 395         flip_addr |= OFC_UPDATE;
 396
 397         rq = alloc_request(overlay, intel_overlay_off_tail);
 398         if (IS_ERR(rq))
 399                 return PTR_ERR(rq);
 400
 401         cs = intel_ring_begin(rq, 6);
 402         if (IS_ERR(cs)) {
 403                 i915_request_add(rq);
 404                 return PTR_ERR(cs);
 405         }
 406
 407         /* wait for overlay to go idle */
 408         *cs++ = MI_OVERLAY_FLIP | MI_OVERLAY_CONTINUE;
 409         *cs++ = flip_addr;
 410         *cs++ = MI_WAIT_FOR_EVENT | MI_WAIT_FOR_OVERLAY_FLIP;
 411
 412         /* turn overlay off */
 413         *cs++ = MI_OVERLAY_FLIP | MI_OVERLAY_OFF;
 414         *cs++ = flip_addr;
 415         *cs++ = MI_WAIT_FOR_EVENT | MI_WAIT_FOR_OVERLAY_FLIP;
 416
 417         intel_ring_advance(rq, cs);
 418
 419         intel_overlay_flip_prepare(overlay, NULL);
 420         i915_request_add(rq);
 421
 422         return i915_active_wait(&overlay->last_flip);
 423 }
 424
 425 /* recover from an interruption due to a signal
 426  * We have to be careful not to repeat work forever an make forward progess. */
 427 static int intel_overlay_recover_from_interrupt(struct intel_overlay *overlay)
 428 {
 429         return i915_active_wait(&overlay->last_flip);
 430 }
 431
 432 /* Wait for pending overlay flip and release old frame.
 433  * Needs to be called before the overlay register are changed
 434  * via intel_overlay_(un)map_regs
 435  */
 436 static int intel_overlay_release_old_vid(struct intel_overlay *overlay)
 437 {
 438         struct drm_i915_private *dev_priv = overlay->i915;
 439         struct i915_request *rq;
 440         u32 *cs;
 441
 442         lockdep_assert_held(&dev_priv->drm.struct_mutex);
 443
 444         /*
 445          * Only wait if there is actually an old frame to release to
 446          * guarantee forward progress.
 447          */
 448         if (!overlay->old_vma)
 449                 return 0;
 450
 451         if (!(I915_READ(GEN2_ISR) & I915_OVERLAY_PLANE_FLIP_PENDING_INTERRUPT)) {
 452                 intel_overlay_release_old_vid_tail(overlay);
 453                 return 0;
 454         }
 455
 456         rq = alloc_request(overlay, intel_overlay_release_old_vid_tail);
 457         if (IS_ERR(rq))
 458                 return PTR_ERR(rq);
 459
 460         cs = intel_ring_begin(rq, 2);
 461         if (IS_ERR(cs)) {
 462                 i915_request_add(rq);
 463                 return PTR_ERR(cs);
 464         }
 465
 466         *cs++ = MI_WAIT_FOR_EVENT | MI_WAIT_FOR_OVERLAY_FLIP;
 467         *cs++ = MI_NOOP;
 468         intel_ring_advance(rq, cs);
 469
 470         i915_request_add(rq);
 471
 472         return i915_active_wait(&overlay->last_flip);
 473 }
 474
 475 void intel_overlay_reset(struct drm_i915_private *dev_priv)
 476 {
 477         struct intel_overlay *overlay = dev_priv->overlay;
 478
 479         if (!overlay)
 480                 return;
 481
 482         overlay->old_xscale = 0;
 483         overlay->old_yscale = 0;
 484         overlay->crtc = NULL;
 485         overlay->active = false;
 486 }
 487
 488 static int packed_depth_bytes(u32 format)
 489 {
 490         switch (format & I915_OVERLAY_DEPTH_MASK) {
 491         case I915_OVERLAY_YUV422:
 492                 return 4;
 493         case I915_OVERLAY_YUV411:
 494                 /* return 6; not implemented */
 495         default:
 496                 return -EINVAL;
 497         }
 498 }
 499
 500 static int packed_width_bytes(u32 format, short width)
 501 {
 502         switch (format & I915_OVERLAY_DEPTH_MASK) {
 503         case I915_OVERLAY_YUV422:
 504                 return width << 1;
 505         default:
 506                 return -EINVAL;
 507         }
 508 }
 509
 510 static int uv_hsubsampling(u32 format)
 511 {
 512         switch (format & I915_OVERLAY_DEPTH_MASK) {
 513         case I915_OVERLAY_YUV422:
 514         case I915_OVERLAY_YUV420:
 515                 return 2;
 516         case I915_OVERLAY_YUV411:
 517         case I915_OVERLAY_YUV410:
 518                 return 4;
 519         default:
 520                 return -EINVAL;
 521         }
 522 }
 523
 524 static int uv_vsubsampling(u32 format)
 525 {
 526         switch (format & I915_OVERLAY_DEPTH_MASK) {
 527         case I915_OVERLAY_YUV420:
 528         case I915_OVERLAY_YUV410:
 529                 return 2;
 530         case I915_OVERLAY_YUV422:
 531         case I915_OVERLAY_YUV411:
 532                 return 1;
 533         default:
 534                 return -EINVAL;
 535         }
 536 }
 537
 538 static u32 calc_swidthsw(struct drm_i915_private *dev_priv, u32 offset, u32 width)
 539 {
 540         u32 sw;
 541
 542         if (IS_GEN(dev_priv, 2))
 543                 sw = ALIGN((offset & 31) + width, 32);
 544         else
 545                 sw = ALIGN((offset & 63) + width, 64);
 546
 547         if (sw == 0)
 548                 return 0;
 549
 550         return (sw - 32) >> 3;
 551 }
 552
 553 static const u16 y_static_hcoeffs[N_PHASES][N_HORIZ_Y_TAPS] = {
 554         [ 0] = { 0x3000, 0xb4a0, 0x1930, 0x1920, 0xb4a0, },
 555         [ 1] = { 0x3000, 0xb500, 0x19d0, 0x1880, 0xb440, },
 556         [ 2] = { 0x3000, 0xb540, 0x1a88, 0x2f80, 0xb3e0, },
 557         [ 3] = { 0x3000, 0xb580, 0x1b30, 0x2e20, 0xb380, },
 558         [ 4] = { 0x3000, 0xb5c0, 0x1bd8, 0x2cc0, 0xb320, },
 559         [ 5] = { 0x3020, 0xb5e0, 0x1c60, 0x2b80, 0xb2c0, },
 560         [ 6] = { 0x3020, 0xb5e0, 0x1cf8, 0x2a20, 0xb260, },
 561         [ 7] = { 0x3020, 0xb5e0, 0x1d80, 0x28e0, 0xb200, },
 562         [ 8] = { 0x3020, 0xb5c0, 0x1e08, 0x3f40, 0xb1c0, },
 563         [ 9] = { 0x3020, 0xb580, 0x1e78, 0x3ce0, 0xb160, },
 564         [10] = { 0x3040, 0xb520, 0x1ed8, 0x3aa0, 0xb120, },
 565         [11] = { 0x3040, 0xb4a0, 0x1f30, 0x3880, 0xb0e0, },
 566         [12] = { 0x3040, 0xb400, 0x1f78, 0x3680, 0xb0a0, },
 567         [13] = { 0x3020, 0xb340, 0x1fb8, 0x34a0, 0xb060, },
 568         [14] = { 0x3020, 0xb240, 0x1fe0, 0x32e0, 0xb040, },
 569         [15] = { 0x3020, 0xb140, 0x1ff8, 0x3160, 0xb020, },
 570         [16] = { 0xb000, 0x3000, 0x0800, 0x3000, 0xb000, },
 571 };
 572
 573 static const u16 uv_static_hcoeffs[N_PHASES][N_HORIZ_UV_TAPS] = {
 574         [ 0] = { 0x3000, 0x1800, 0x1800, },
 575         [ 1] = { 0xb000, 0x18d0, 0x2e60, },
 576         [ 2] = { 0xb000, 0x1990, 0x2ce0, },
 577         [ 3] = { 0xb020, 0x1a68, 0x2b40, },
 578         [ 4] = { 0xb040, 0x1b20, 0x29e0, },
 579         [ 5] = { 0xb060, 0x1bd8, 0x2880, },
 580         [ 6] = { 0xb080, 0x1c88, 0x3e60, },
 581         [ 7] = { 0xb0a0, 0x1d28, 0x3c00, },
 582         [ 8] = { 0xb0c0, 0x1db8, 0x39e0, },
 583         [ 9] = { 0xb0e0, 0x1e40, 0x37e0, },
 584         [10] = { 0xb100, 0x1eb8, 0x3620, },
 585         [11] = { 0xb100, 0x1f18, 0x34a0, },
 586         [12] = { 0xb100, 0x1f68, 0x3360, },
 587         [13] = { 0xb0e0, 0x1fa8, 0x3240, },
 588         [14] = { 0xb0c0, 0x1fe0, 0x3140, },
 589         [15] = { 0xb060, 0x1ff0, 0x30a0, },
 590         [16] = { 0x3000, 0x0800, 0x3000, },
 591 };
 592
 593 static void update_polyphase_filter(struct overlay_registers __iomem *regs)
 594 {
 595         memcpy_toio(regs->Y_HCOEFS, y_static_hcoeffs, sizeof(y_static_hcoeffs));
 596         memcpy_toio(regs->UV_HCOEFS, uv_static_hcoeffs,
 597                     sizeof(uv_static_hcoeffs));
 598 }
 599
 600 static bool update_scaling_factors(struct intel_overlay *overlay,
 601                                    struct overlay_registers __iomem *regs,
 602                                    struct drm_intel_overlay_put_image *params)
 603 {
 604         /* fixed point with a 12 bit shift */
 605         u32 xscale, yscale, xscale_UV, yscale_UV;
 606 #define FP_SHIFT 12
 607 #define FRACT_MASK 0xfff
 608         bool scale_changed = false;
 609         int uv_hscale = uv_hsubsampling(params->flags);
 610         int uv_vscale = uv_vsubsampling(params->flags);
 611
 612         if (params->dst_width > 1)
 613                 xscale = ((params->src_scan_width - 1) << FP_SHIFT) /
 614                         params->dst_width;
 615         else
 616                 xscale = 1 << FP_SHIFT;
 617
 618         if (params->dst_height > 1)
 619                 yscale = ((params->src_scan_height - 1) << FP_SHIFT) /
 620                         params->dst_height;
 621         else
 622                 yscale = 1 << FP_SHIFT;
 623
 624         /*if (params->format & I915_OVERLAY_YUV_PLANAR) {*/
 625         xscale_UV = xscale/uv_hscale;
 626         yscale_UV = yscale/uv_vscale;
 627         /* make the Y scale to UV scale ratio an exact multiply */
 628         xscale = xscale_UV * uv_hscale;
 629         yscale = yscale_UV * uv_vscale;
 630         /*} else {
 631           xscale_UV = 0;
 632           yscale_UV = 0;
 633           }*/
 634
 635         if (xscale != overlay->old_xscale || yscale != overlay->old_yscale)
 636                 scale_changed = true;
 637         overlay->old_xscale = xscale;
 638         overlay->old_yscale = yscale;
 639
 640         iowrite32(((yscale & FRACT_MASK) << 20) |
 641                   ((xscale >> FP_SHIFT)  << 16) |
 642                   ((xscale & FRACT_MASK) << 3),
 643                  &regs->YRGBSCALE);
 644
 645         iowrite32(((yscale_UV & FRACT_MASK) << 20) |
 646                   ((xscale_UV >> FP_SHIFT)  << 16) |
 647                   ((xscale_UV & FRACT_MASK) << 3),
 648                  &regs->UVSCALE);
 649
 650         iowrite32((((yscale    >> FP_SHIFT) << 16) |
 651                    ((yscale_UV >> FP_SHIFT) << 0)),
 652                  &regs->UVSCALEV);
 653
 654         if (scale_changed)
 655                 update_polyphase_filter(regs);
 656
 657         return scale_changed;
 658 }
 659
 660 static void update_colorkey(struct intel_overlay *overlay,
 661                             struct overlay_registers __iomem *regs)
 662 {
 663         const struct intel_plane_state *state =
 664                 to_intel_plane_state(overlay->crtc->base.primary->state);
 665         u32 key = overlay->color_key;
 666         u32 format = 0;
 667         u32 flags = 0;
 668
 669         if (overlay->color_key_enabled)
 670                 flags |= DST_KEY_ENABLE;
 671
 672         if (state->base.visible)
 673                 format = state->base.fb->format->format;
 674
 675         switch (format) {
 676         case DRM_FORMAT_C8:
 677                 key = 0;
 678                 flags |= CLK_RGB8I_MASK;
 679                 break;
 680         case DRM_FORMAT_XRGB1555:
 681                 key = RGB15_TO_COLORKEY(key);
 682                 flags |= CLK_RGB15_MASK;
 683                 break;
 684         case DRM_FORMAT_RGB565:
 685                 key = RGB16_TO_COLORKEY(key);
 686                 flags |= CLK_RGB16_MASK;
 687                 break;
 688         default:
 689                 flags |= CLK_RGB24_MASK;
 690                 break;
 691         }
 692
 693         iowrite32(key, &regs->DCLRKV);
 694         iowrite32(flags, &regs->DCLRKM);
 695 }
 696
 697 static u32 overlay_cmd_reg(struct drm_intel_overlay_put_image *params)
 698 {
 699         u32 cmd = OCMD_ENABLE | OCMD_BUF_TYPE_FRAME | OCMD_BUFFER0;
 700
 701         if (params->flags & I915_OVERLAY_YUV_PLANAR) {
 702                 switch (params->flags & I915_OVERLAY_DEPTH_MASK) {
 703                 case I915_OVERLAY_YUV422:
 704                         cmd |= OCMD_YUV_422_PLANAR;
 705                         break;
 706                 case I915_OVERLAY_YUV420:
 707                         cmd |= OCMD_YUV_420_PLANAR;
 708                         break;
 709                 case I915_OVERLAY_YUV411:
 710                 case I915_OVERLAY_YUV410:
 711                         cmd |= OCMD_YUV_410_PLANAR;
 712                         break;
 713                 }
 714         } else { /* YUV packed */
 715                 switch (params->flags & I915_OVERLAY_DEPTH_MASK) {
 716                 case I915_OVERLAY_YUV422:
 717                         cmd |= OCMD_YUV_422_PACKED;
 718                         break;
 719                 case I915_OVERLAY_YUV411:
 720                         cmd |= OCMD_YUV_411_PACKED;
 721                         break;
 722                 }
 723
 724                 switch (params->flags & I915_OVERLAY_SWAP_MASK) {
 725                 case I915_OVERLAY_NO_SWAP:
 726                         break;
 727                 case I915_OVERLAY_UV_SWAP:
 728                         cmd |= OCMD_UV_SWAP;
 729                         break;
 730                 case I915_OVERLAY_Y_SWAP:
 731                         cmd |= OCMD_Y_SWAP;
 732                         break;
 733                 case I915_OVERLAY_Y_AND_UV_SWAP:
 734                         cmd |= OCMD_Y_AND_UV_SWAP;
 735                         break;
 736                 }
 737         }
 738
 739         return cmd;
 740 }
 741
 742 static int intel_overlay_do_put_image(struct intel_overlay *overlay,
 743                                       struct drm_i915_gem_object *new_bo,
 744                                       struct drm_intel_overlay_put_image *params)
 745 {
 746         struct overlay_registers __iomem *regs = overlay->regs;
 747         struct drm_i915_private *dev_priv = overlay->i915;
 748         u32 swidth, swidthsw, sheight, ostride;
 749         enum pipe pipe = overlay->crtc->pipe;
 750         bool scale_changed = false;
 751         struct i915_vma *vma;
 752         int ret, tmp_width;
 753
 754         lockdep_assert_held(&dev_priv->drm.struct_mutex);
 755         WARN_ON(!drm_modeset_is_locked(&dev_priv->drm.mode_config.connection_mutex));
 756
 757         ret = intel_overlay_release_old_vid(overlay);
 758         if (ret != 0)
 759                 return ret;
 760
 761         atomic_inc(&dev_priv->gpu_error.pending_fb_pin);
 762
 763         i915_gem_object_lock(new_bo);
 764         vma = i915_gem_object_pin_to_display_plane(new_bo,
 765                                                    0, NULL, PIN_MAPPABLE);
 766         i915_gem_object_unlock(new_bo);
 767         if (IS_ERR(vma)) {
 768                 ret = PTR_ERR(vma);
 769                 goto out_pin_section;
 770         }
 771         intel_frontbuffer_flush(new_bo->frontbuffer, ORIGIN_DIRTYFB);
 772
 773         if (!overlay->active) {
 774                 u32 oconfig;
 775
 776                 oconfig = OCONF_CC_OUT_8BIT;
 777                 if (IS_GEN(dev_priv, 4))
 778                         oconfig |= OCONF_CSC_MODE_BT709;
 779                 oconfig |= pipe == 0 ?
 780                         OCONF_PIPE_A : OCONF_PIPE_B;
 781                 iowrite32(oconfig, &regs->OCONFIG);
 782
 783                 ret = intel_overlay_on(overlay);
 784                 if (ret != 0)
 785                         goto out_unpin;
 786         }
 787
 788         iowrite32(params->dst_y << 16 | params->dst_x, &regs->DWINPOS);
 789         iowrite32(params->dst_height << 16 | params->dst_width, &regs->DWINSZ);
 790
 791         if (params->flags & I915_OVERLAY_YUV_PACKED)
 792                 tmp_width = packed_width_bytes(params->flags,
 793                                                params->src_width);
 794         else
 795                 tmp_width = params->src_width;
 796
 797         swidth = params->src_width;
 798         swidthsw = calc_swidthsw(dev_priv, params->offset_Y, tmp_width);
 799         sheight = params->src_height;
 800         iowrite32(i915_ggtt_offset(vma) + params->offset_Y, &regs->OBUF_0Y);
 801         ostride = params->stride_Y;
 802
 803         if (params->flags & I915_OVERLAY_YUV_PLANAR) {
 804                 int uv_hscale = uv_hsubsampling(params->flags);
 805                 int uv_vscale = uv_vsubsampling(params->flags);
 806                 u32 tmp_U, tmp_V;
 807
 808                 swidth |= (params->src_width / uv_hscale) << 16;
 809                 sheight |= (params->src_height / uv_vscale) << 16;
 810
 811                 tmp_U = calc_swidthsw(dev_priv, params->offset_U,
 812                                       params->src_width / uv_hscale);
 813                 tmp_V = calc_swidthsw(dev_priv, params->offset_V,
 814                                       params->src_width / uv_hscale);
 815                 swidthsw |= max(tmp_U, tmp_V) << 16;
 816
 817                 iowrite32(i915_ggtt_offset(vma) + params->offset_U,
 818                           &regs->OBUF_0U);
 819                 iowrite32(i915_ggtt_offset(vma) + params->offset_V,
 820                           &regs->OBUF_0V);
 821
 822                 ostride |= params->stride_UV << 16;
 823         }
 824
 825         iowrite32(swidth, &regs->SWIDTH);
 826         iowrite32(swidthsw, &regs->SWIDTHSW);
 827         iowrite32(sheight, &regs->SHEIGHT);
 828         iowrite32(ostride, &regs->OSTRIDE);
 829
 830         scale_changed = update_scaling_factors(overlay, regs, params);
 831
 832         update_colorkey(overlay, regs);
 833
 834         iowrite32(overlay_cmd_reg(params), &regs->OCMD);
 835
 836         ret = intel_overlay_continue(overlay, vma, scale_changed);
 837         if (ret)
 838                 goto out_unpin;
 839
 840         return 0;
 841
 842 out_unpin:
 843         i915_gem_object_unpin_from_display_plane(vma);
 844 out_pin_section:
 845         atomic_dec(&dev_priv->gpu_error.pending_fb_pin);
 846
 847         return ret;
 848 }
 849
 850 int intel_overlay_switch_off(struct intel_overlay *overlay)
 851 {
 852         struct drm_i915_private *dev_priv = overlay->i915;
 853         int ret;
 854
 855         lockdep_assert_held(&dev_priv->drm.struct_mutex);
 856         WARN_ON(!drm_modeset_is_locked(&dev_priv->drm.mode_config.connection_mutex));
 857
 858         ret = intel_overlay_recover_from_interrupt(overlay);
 859         if (ret != 0)
 860                 return ret;
 861
 862         if (!overlay->active)
 863                 return 0;
 864
 865         ret = intel_overlay_release_old_vid(overlay);
 866         if (ret != 0)
 867                 return ret;
 868
 869         iowrite32(0, &overlay->regs->OCMD);
 870
 871         return intel_overlay_off(overlay);
 872 }
 873
 874 static int check_overlay_possible_on_crtc(struct intel_overlay *overlay,
 875                                           struct intel_crtc *crtc)
 876 {
 877         if (!crtc->active)
 878                 return -EINVAL;
 879
 880         /* can't use the overlay with double wide pipe */
 881         if (crtc->config->double_wide)
 882                 return -EINVAL;
 883
 884         return 0;
 885 }
 886
 887 static void update_pfit_vscale_ratio(struct intel_overlay *overlay)
 888 {
 889         struct drm_i915_private *dev_priv = overlay->i915;
 890         u32 pfit_control = I915_READ(PFIT_CONTROL);
 891         u32 ratio;
 892
 893         /* XXX: This is not the same logic as in the xorg driver, but more in
 894          * line with the intel documentation for the i965
 895          */
 896         if (INTEL_GEN(dev_priv) >= 4) {
 897                 /* on i965 use the PGM reg to read out the autoscaler values */
 898                 ratio = I915_READ(PFIT_PGM_RATIOS) >> PFIT_VERT_SCALE_SHIFT_965;
 899         } else {
 900                 if (pfit_control & VERT_AUTO_SCALE)
 901                         ratio = I915_READ(PFIT_AUTO_RATIOS);
 902                 else
 903                         ratio = I915_READ(PFIT_PGM_RATIOS);
 904                 ratio >>= PFIT_VERT_SCALE_SHIFT;
 905         }
 906
 907         overlay->pfit_vscale_ratio = ratio;
 908 }
 909
 910 static int check_overlay_dst(struct intel_overlay *overlay,
 911                              struct drm_intel_overlay_put_image *rec)
 912 {
 913         const struct intel_crtc_state *pipe_config =
 914                 overlay->crtc->config;
 915
 916         if (rec->dst_x < pipe_config->pipe_src_w &&
 917             rec->dst_x + rec->dst_width <= pipe_config->pipe_src_w &&
 918             rec->dst_y < pipe_config->pipe_src_h &&
 919             rec->dst_y + rec->dst_height <= pipe_config->pipe_src_h)
 920                 return 0;
 921         else
 922                 return -EINVAL;
 923 }
 924
 925 static int check_overlay_scaling(struct drm_intel_overlay_put_image *rec)
 926 {
 927         u32 tmp;
 928
 929         /* downscaling limit is 8.0 */
 930         tmp = ((rec->src_scan_height << 16) / rec->dst_height) >> 16;
 931         if (tmp > 7)
 932                 return -EINVAL;
 933
 934         tmp = ((rec->src_scan_width << 16) / rec->dst_width) >> 16;
 935         if (tmp > 7)
 936                 return -EINVAL;
 937
 938         return 0;
 939 }
 940
 941 static int check_overlay_src(struct drm_i915_private *dev_priv,
 942                              struct drm_intel_overlay_put_image *rec,
 943                              struct drm_i915_gem_object *new_bo)
 944 {
 945         int uv_hscale = uv_hsubsampling(rec->flags);
 946         int uv_vscale = uv_vsubsampling(rec->flags);
 947         u32 stride_mask;
 948         int depth;
 949         u32 tmp;
 950
 951         /* check src dimensions */
 952         if (IS_I845G(dev_priv) || IS_I830(dev_priv)) {
 953                 if (rec->src_height > IMAGE_MAX_HEIGHT_LEGACY ||
 954                     rec->src_width  > IMAGE_MAX_WIDTH_LEGACY)
 955                         return -EINVAL;
 956         } else {
 957                 if (rec->src_height > IMAGE_MAX_HEIGHT ||
 958                     rec->src_width  > IMAGE_MAX_WIDTH)
 959                         return -EINVAL;
 960         }
 961
 962         /* better safe than sorry, use 4 as the maximal subsampling ratio */
 963         if (rec->src_height < N_VERT_Y_TAPS*4 ||
 964             rec->src_width  < N_HORIZ_Y_TAPS*4)
 965                 return -EINVAL;
 966
 967         /* check alignment constraints */
 968         switch (rec->flags & I915_OVERLAY_TYPE_MASK) {
 969         case I915_OVERLAY_RGB:
 970                 /* not implemented */
 971                 return -EINVAL;
 972
 973         case I915_OVERLAY_YUV_PACKED:
 974                 if (uv_vscale != 1)
 975                         return -EINVAL;
 976
 977                 depth = packed_depth_bytes(rec->flags);
 978                 if (depth < 0)
 979                         return depth;
 980
 981                 /* ignore UV planes */
 982                 rec->stride_UV = 0;
 983                 rec->offset_U = 0;
 984                 rec->offset_V = 0;
 985                 /* check pixel alignment */
 986                 if (rec->offset_Y % depth)
 987                         return -EINVAL;
 988                 break;
 989
 990         case I915_OVERLAY_YUV_PLANAR:
 991                 if (uv_vscale < 0 || uv_hscale < 0)
 992                         return -EINVAL;
 993                 /* no offset restrictions for planar formats */
 994                 break;
 995
 996         default:
 997                 return -EINVAL;
 998         }
 999
1000         if (rec->src_width % uv_hscale)
1001                 return -EINVAL;
1002
1003         /* stride checking */
1004         if (IS_I830(dev_priv) || IS_I845G(dev_priv))
1005                 stride_mask = 255;
1006         else
1007                 stride_mask = 63;
1008
1009         if (rec->stride_Y & stride_mask || rec->stride_UV & stride_mask)
1010                 return -EINVAL;
1011         if (IS_GEN(dev_priv, 4) && rec->stride_Y < 512)
1012                 return -EINVAL;
1013
1014         tmp = (rec->flags & I915_OVERLAY_TYPE_MASK) == I915_OVERLAY_YUV_PLANAR ?
1015                 4096 : 8192;
1016         if (rec->stride_Y > tmp || rec->stride_UV > 2*1024)
1017                 return -EINVAL;
1018
1019         /* check buffer dimensions */
1020         switch (rec->flags & I915_OVERLAY_TYPE_MASK) {
1021         case I915_OVERLAY_RGB:
1022         case I915_OVERLAY_YUV_PACKED:
1023                 /* always 4 Y values per depth pixels */
1024                 if (packed_width_bytes(rec->flags, rec->src_width) > rec->stride_Y)
1025                         return -EINVAL;
1026
1027                 tmp = rec->stride_Y*rec->src_height;
1028                 if (rec->offset_Y + tmp > new_bo->base.size)
1029                         return -EINVAL;
1030                 break;
1031
1032         case I915_OVERLAY_YUV_PLANAR:
1033                 if (rec->src_width > rec->stride_Y)
1034                         return -EINVAL;
1035                 if (rec->src_width/uv_hscale > rec->stride_UV)
1036                         return -EINVAL;
1037
1038                 tmp = rec->stride_Y * rec->src_height;
1039                 if (rec->offset_Y + tmp > new_bo->base.size)
1040                         return -EINVAL;
1041
1042                 tmp = rec->stride_UV * (rec->src_height / uv_vscale);
1043                 if (rec->offset_U + tmp > new_bo->base.size ||
1044                     rec->offset_V + tmp > new_bo->base.size)
1045                         return -EINVAL;
1046                 break;
1047         }
1048
1049         return 0;
1050 }
1051
1052 int intel_overlay_put_image_ioctl(struct drm_device *dev, void *data,
1053                                   struct drm_file *file_priv)
1054 {
1055         struct drm_intel_overlay_put_image *params = data;
1056         struct drm_i915_private *dev_priv = to_i915(dev);
1057         struct intel_overlay *overlay;
1058         struct drm_crtc *drmmode_crtc;
1059         struct intel_crtc *crtc;
1060         struct drm_i915_gem_object *new_bo;
1061         int ret;
1062
1063         overlay = dev_priv->overlay;
1064         if (!overlay) {
1065                 DRM_DEBUG("userspace bug: no overlay\n");
1066                 return -ENODEV;
1067         }
1068
1069         if (!(params->flags & I915_OVERLAY_ENABLE)) {
1070                 drm_modeset_lock_all(dev);
1071                 mutex_lock(&dev->struct_mutex);
1072
1073                 ret = intel_overlay_switch_off(overlay);
1074
1075                 mutex_unlock(&dev->struct_mutex);
1076                 drm_modeset_unlock_all(dev);
1077
1078                 return ret;
1079         }
1080
1081         drmmode_crtc = drm_crtc_find(dev, file_priv, params->crtc_id);
1082         if (!drmmode_crtc)
1083                 return -ENOENT;
1084         crtc = to_intel_crtc(drmmode_crtc);
1085
1086         new_bo = i915_gem_object_lookup(file_priv, params->bo_handle);
1087         if (!new_bo)
1088                 return -ENOENT;
1089
1090         drm_modeset_lock_all(dev);
1091         mutex_lock(&dev->struct_mutex);
1092
1093         if (i915_gem_object_is_tiled(new_bo)) {
1094                 DRM_DEBUG_KMS("buffer used for overlay image can not be tiled\n");
1095                 ret = -EINVAL;
1096                 goto out_unlock;
1097         }
1098
1099         ret = intel_overlay_recover_from_interrupt(overlay);
1100         if (ret != 0)
1101                 goto out_unlock;
1102
1103         if (overlay->crtc != crtc) {
1104                 ret = intel_overlay_switch_off(overlay);
1105                 if (ret != 0)
1106                         goto out_unlock;
1107
1108                 ret = check_overlay_possible_on_crtc(overlay, crtc);
1109                 if (ret != 0)
1110                         goto out_unlock;
1111
1112                 overlay->crtc = crtc;
1113                 crtc->overlay = overlay;
1114
1115                 /* line too wide, i.e. one-line-mode */
1116                 if (crtc->config->pipe_src_w > 1024 &&
1117                     crtc->config->gmch_pfit.control & PFIT_ENABLE) {
1118                         overlay->pfit_active = true;
1119                         update_pfit_vscale_ratio(overlay);
1120                 } else
1121                         overlay->pfit_active = false;
1122         }
1123
1124         ret = check_overlay_dst(overlay, params);
1125         if (ret != 0)
1126                 goto out_unlock;
1127
1128         if (overlay->pfit_active) {
1129                 params->dst_y = (((u32)params->dst_y << 12) /
1130                                  overlay->pfit_vscale_ratio);
1131                 /* shifting right rounds downwards, so add 1 */
1132                 params->dst_height = (((u32)params->dst_height << 12) /
1133                                  overlay->pfit_vscale_ratio) + 1;
1134         }
1135
1136         if (params->src_scan_height > params->src_height ||
1137             params->src_scan_width > params->src_width) {
1138                 ret = -EINVAL;
1139                 goto out_unlock;
1140         }
1141
1142         ret = check_overlay_src(dev_priv, params, new_bo);
1143         if (ret != 0)
1144                 goto out_unlock;
1145
1146         /* Check scaling after src size to prevent a divide-by-zero. */
1147         ret = check_overlay_scaling(params);
1148         if (ret != 0)
1149                 goto out_unlock;
1150
1151         ret = intel_overlay_do_put_image(overlay, new_bo, params);
1152         if (ret != 0)
1153                 goto out_unlock;
1154
1155         mutex_unlock(&dev->struct_mutex);
1156         drm_modeset_unlock_all(dev);
1157         i915_gem_object_put(new_bo);
1158
1159         return 0;
1160
1161 out_unlock:
1162         mutex_unlock(&dev->struct_mutex);
1163         drm_modeset_unlock_all(dev);
1164         i915_gem_object_put(new_bo);
1165
1166         return ret;
1167 }
1168
1169 static void update_reg_attrs(struct intel_overlay *overlay,
1170                              struct overlay_registers __iomem *regs)
1171 {
1172         iowrite32((overlay->contrast << 18) | (overlay->brightness & 0xff),
1173                   &regs->OCLRC0);
1174         iowrite32(overlay->saturation, &regs->OCLRC1);
1175 }
1176
1177 static bool check_gamma_bounds(u32 gamma1, u32 gamma2)
1178 {
1179         int i;
1180
1181         if (gamma1 & 0xff000000 || gamma2 & 0xff000000)
1182                 return false;
1183
1184         for (i = 0; i < 3; i++) {
1185                 if (((gamma1 >> i*8) & 0xff) >= ((gamma2 >> i*8) & 0xff))
1186                         return false;
1187         }
1188
1189         return true;
1190 }
1191
1192 static bool check_gamma5_errata(u32 gamma5)
1193 {
1194         int i;
1195
1196         for (i = 0; i < 3; i++) {
1197                 if (((gamma5 >> i*8) & 0xff) == 0x80)
1198                         return false;
1199         }
1200
1201         return true;
1202 }
1203
1204 static int check_gamma(struct drm_intel_overlay_attrs *attrs)
1205 {
1206         if (!check_gamma_bounds(0, attrs->gamma0) ||
1207             !check_gamma_bounds(attrs->gamma0, attrs->gamma1) ||
1208             !check_gamma_bounds(attrs->gamma1, attrs->gamma2) ||
1209             !check_gamma_bounds(attrs->gamma2, attrs->gamma3) ||
1210             !check_gamma_bounds(attrs->gamma3, attrs->gamma4) ||
1211             !check_gamma_bounds(attrs->gamma4, attrs->gamma5) ||
1212             !check_gamma_bounds(attrs->gamma5, 0x00ffffff))
1213                 return -EINVAL;
1214
1215         if (!check_gamma5_errata(attrs->gamma5))
1216                 return -EINVAL;
1217
1218         return 0;
1219 }
1220
1221 int intel_overlay_attrs_ioctl(struct drm_device *dev, void *data,
1222                               struct drm_file *file_priv)
1223 {
1224         struct drm_intel_overlay_attrs *attrs = data;
1225         struct drm_i915_private *dev_priv = to_i915(dev);
1226         struct intel_overlay *overlay;
1227         int ret;
1228
1229         overlay = dev_priv->overlay;
1230         if (!overlay) {
1231                 DRM_DEBUG("userspace bug: no overlay\n");
1232                 return -ENODEV;
1233         }
1234
1235         drm_modeset_lock_all(dev);
1236         mutex_lock(&dev->struct_mutex);
1237
1238         ret = -EINVAL;
1239         if (!(attrs->flags & I915_OVERLAY_UPDATE_ATTRS)) {
1240                 attrs->color_key  = overlay->color_key;
1241                 attrs->brightness = overlay->brightness;
1242                 attrs->contrast   = overlay->contrast;
1243                 attrs->saturation = overlay->saturation;
1244
1245                 if (!IS_GEN(dev_priv, 2)) {
1246                         attrs->gamma0 = I915_READ(OGAMC0);
1247                         attrs->gamma1 = I915_READ(OGAMC1);
1248                         attrs->gamma2 = I915_READ(OGAMC2);
1249                         attrs->gamma3 = I915_READ(OGAMC3);
1250                         attrs->gamma4 = I915_READ(OGAMC4);
1251                         attrs->gamma5 = I915_READ(OGAMC5);
1252                 }
1253         } else {
1254                 if (attrs->brightness < -128 || attrs->brightness > 127)
1255                         goto out_unlock;
1256                 if (attrs->contrast > 255)
1257                         goto out_unlock;
1258                 if (attrs->saturation > 1023)
1259                         goto out_unlock;
1260
1261                 overlay->color_key  = attrs->color_key;
1262                 overlay->brightness = attrs->brightness;
1263                 overlay->contrast   = attrs->contrast;
1264                 overlay->saturation = attrs->saturation;
1265
1266                 update_reg_attrs(overlay, overlay->regs);
1267
1268                 if (attrs->flags & I915_OVERLAY_UPDATE_GAMMA) {
1269                         if (IS_GEN(dev_priv, 2))
1270                                 goto out_unlock;
1271
1272                         if (overlay->active) {
1273                                 ret = -EBUSY;
1274                                 goto out_unlock;
1275                         }
1276
1277                         ret = check_gamma(attrs);
1278                         if (ret)
1279                                 goto out_unlock;
1280
1281                         I915_WRITE(OGAMC0, attrs->gamma0);
1282                         I915_WRITE(OGAMC1, attrs->gamma1);
1283                         I915_WRITE(OGAMC2, attrs->gamma2);
1284                         I915_WRITE(OGAMC3, attrs->gamma3);
1285                         I915_WRITE(OGAMC4, attrs->gamma4);
1286                         I915_WRITE(OGAMC5, attrs->gamma5);
1287                 }
1288         }
1289         overlay->color_key_enabled = (attrs->flags & I915_OVERLAY_DISABLE_DEST_COLORKEY) == 0;
1290
1291         ret = 0;
1292 out_unlock:
1293         mutex_unlock(&dev->struct_mutex);
1294         drm_modeset_unlock_all(dev);
1295
1296         return ret;
1297 }
1298
1299 static int get_registers(struct intel_overlay *overlay, bool use_phys)
1300 {
1301         struct drm_i915_private *i915 = overlay->i915;
1302         struct drm_i915_gem_object *obj;
1303         struct i915_vma *vma;
1304         int err;
1305
1306         obj = i915_gem_object_create_stolen(i915, PAGE_SIZE);
1307         if (obj == NULL)
1308                 obj = i915_gem_object_create_internal(i915, PAGE_SIZE);
1309         if (IS_ERR(obj))
1310                 return PTR_ERR(obj);
1311
1312         vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0, PIN_MAPPABLE);
1313         if (IS_ERR(vma)) {
1314                 err = PTR_ERR(vma);
1315                 goto err_put_bo;
1316         }
1317
1318         if (use_phys)
1319                 overlay->flip_addr = sg_dma_address(obj->mm.pages->sgl);
1320         else
1321                 overlay->flip_addr = i915_ggtt_offset(vma);
1322         overlay->regs = i915_vma_pin_iomap(vma);
1323         i915_vma_unpin(vma);
1324
1325         if (IS_ERR(overlay->regs)) {
1326                 err = PTR_ERR(overlay->regs);
1327                 goto err_put_bo;
1328         }
1329
1330         overlay->reg_bo = obj;
1331         return 0;
1332
1333 err_put_bo:
1334         i915_gem_object_put(obj);
1335         return err;
1336 }
1337
1338 void intel_overlay_setup(struct drm_i915_private *dev_priv)
1339 {
1340         struct intel_overlay *overlay;
1341         int ret;
1342
1343         if (!HAS_OVERLAY(dev_priv))
1344                 return;
1345
1346         if (!HAS_ENGINE(dev_priv, RCS0))
1347                 return;
1348
1349         overlay = kzalloc(sizeof(*overlay), GFP_KERNEL);
1350         if (!overlay)
1351                 return;
1352
1353         overlay->i915 = dev_priv;
1354         overlay->context = dev_priv->engine[RCS0]->kernel_context;
1355         GEM_BUG_ON(!overlay->context);
1356
1357         overlay->color_key = 0x0101fe;
1358         overlay->color_key_enabled = true;
1359         overlay->brightness = -19;
1360         overlay->contrast = 75;
1361         overlay->saturation = 146;
1362
1363         i915_active_init(dev_priv,
1364                          &overlay->last_flip,
1365                          NULL, intel_overlay_last_flip_retire);
1366
1367         ret = get_registers(overlay, OVERLAY_NEEDS_PHYSICAL(dev_priv));
1368         if (ret)
1369                 goto out_free;
1370
1371         memset_io(overlay->regs, 0, sizeof(struct overlay_registers));
1372         update_polyphase_filter(overlay->regs);
1373         update_reg_attrs(overlay, overlay->regs);
1374
1375         dev_priv->overlay = overlay;
1376         DRM_INFO("Initialized overlay support.\n");
1377         return;
1378
1379 out_free:
1380         kfree(overlay);
1381 }
1382
1383 void intel_overlay_cleanup(struct drm_i915_private *dev_priv)
1384 {
1385         struct intel_overlay *overlay;
1386
1387         overlay = fetch_and_zero(&dev_priv->overlay);
1388         if (!overlay)
1389                 return;
1390
1391         /*
1392          * The bo's should be free'd by the generic code already.
1393          * Furthermore modesetting teardown happens beforehand so the
1394          * hardware should be off already.
1395          */
1396         WARN_ON(overlay->active);
1397
1398         i915_gem_object_put(overlay->reg_bo);
1399         i915_active_fini(&overlay->last_flip);
1400
1401         kfree(overlay);
1402 }
1403
1404 #if IS_ENABLED(CONFIG_DRM_I915_CAPTURE_ERROR)
1405
1406 struct intel_overlay_error_state {
1407         struct overlay_registers regs;
1408         unsigned long base;
1409         u32 dovsta;
1410         u32 isr;
1411 };
1412
1413 struct intel_overlay_error_state *
1414 intel_overlay_capture_error_state(struct drm_i915_private *dev_priv)
1415 {
1416         struct intel_overlay *overlay = dev_priv->overlay;
1417         struct intel_overlay_error_state *error;
1418
1419         if (!overlay || !overlay->active)
1420                 return NULL;
1421
1422         error = kmalloc(sizeof(*error), GFP_ATOMIC);
1423         if (error == NULL)
1424                 return NULL;
1425
1426         error->dovsta = I915_READ(DOVSTA);
1427         error->isr = I915_READ(GEN2_ISR);
1428         error->base = overlay->flip_addr;
1429
1430         memcpy_fromio(&error->regs, overlay->regs, sizeof(error->regs));
1431
1432         return error;
1433 }
1434
1435 void
1436 intel_overlay_print_error_state(struct drm_i915_error_state_buf *m,
1437                                 struct intel_overlay_error_state *error)
1438 {
1439         i915_error_printf(m, "Overlay, status: 0x%08x, interrupt: 0x%08x\n",
1440                           error->dovsta, error->isr);
1441         i915_error_printf(m, "  Register file at 0x%08lx:\n",
1442                           error->base);
1443
1444 #define P(x) i915_error_printf(m, "    " #x ":  0x%08x\n", error->regs.x)
1445         P(OBUF_0Y);
1446         P(OBUF_1Y);
1447         P(OBUF_0U);
1448         P(OBUF_0V);
1449         P(OBUF_1U);
1450         P(OBUF_1V);
1451         P(OSTRIDE);
1452         P(YRGB_VPH);
1453         P(UV_VPH);
1454         P(HORZ_PH);
1455         P(INIT_PHS);
1456         P(DWINPOS);
1457         P(DWINSZ);
1458         P(SWIDTH);
1459         P(SWIDTHSW);
1460         P(SHEIGHT);
1461         P(YRGBSCALE);
1462         P(UVSCALE);
1463         P(OCLRC0);
1464         P(OCLRC1);
1465         P(DCLRKV);
1466         P(DCLRKM);
1467         P(SCLRKVH);
1468         P(SCLRKVL);
1469         P(SCLRKEN);
1470         P(OCONFIG);
1471         P(OCMD);
1472         P(OSTART_0Y);
1473         P(OSTART_1Y);
1474         P(OSTART_0U);
1475         P(OSTART_0V);
1476         P(OSTART_1U);
1477         P(OSTART_1V);
1478         P(OTILEOFF_0Y);
1479         P(OTILEOFF_1Y);
1480         P(OTILEOFF_0U);
1481         P(OTILEOFF_0V);
1482         P(OTILEOFF_1U);
1483         P(OTILEOFF_1V);
1484         P(FASTHSCALE);
1485         P(UVSCALEV);
1486 #undef P
1487 }
1488
1489 #endif