target/linux/brcm2708/patches-4.4/0117-drm-vc4-Add-support-for-MSAA-rendering.patch

   1 From b66efe927216251ae27f9213e6b92b3a49deb73e Mon Sep 17 00:00:00 2001
   2 From: Eric Anholt <eric@anholt.net>
   3 Date: Fri, 17 Jul 2015 13:15:50 -0700
   4 Subject: [PATCH 117/127] drm/vc4: Add support for MSAA rendering.
   5
   6 For MSAA, you set a bit in the binner that halves the size of tiles in
   7 each direction, so you can pack 4 samples per pixel in the tile
   8 buffer.  During rendering, you can load and store raw tile buffer
   9 contents (to save the per-sample MSAA contents), or you can load/store
  10 resolved tile buffer contents (loads spam the pixel value to all 4
  11 samples, and stores either average the 4 color samples, or store the
  12 first sample for Z/S).
  13
  14 Signed-off-by: Eric Anholt <eric@anholt.net>
  15 ---
  16  drivers/gpu/drm/vc4/vc4_packet.h    |  23 ++-
  17  drivers/gpu/drm/vc4/vc4_render_cl.c | 274 ++++++++++++++++++++++++++++++------
  18  drivers/gpu/drm/vc4/vc4_validate.c  |   5 +-
  19  include/uapi/drm/vc4_drm.h          |  11 +-
  20  4 files changed, 258 insertions(+), 55 deletions(-)
  21
  22 --- a/drivers/gpu/drm/vc4/vc4_packet.h
  23 +++ b/drivers/gpu/drm/vc4/vc4_packet.h
  24 @@ -123,6 +123,11 @@ enum vc4_packet {
  25  #define VC4_PACKET_TILE_COORDINATES_SIZE                               3
  26  #define VC4_PACKET_GEM_HANDLES_SIZE                                    9
  27
  28 +/* Number of multisamples supported. */
  29 +#define VC4_MAX_SAMPLES                                                        4
  30 +/* Size of a full resolution color or Z tile buffer load/store. */
  31 +#define VC4_TILE_BUFFER_SIZE                   (64 * 64 * 4)
  32 +
  33  /** @{
  34   * Bits used by packets like VC4_PACKET_STORE_TILE_BUFFER_GENERAL and
  35   * VC4_PACKET_TILE_RENDERING_MODE_CONFIG.
  36 @@ -137,10 +142,20 @@ enum vc4_packet {
  37   * low bits of VC4_PACKET_STORE_FULL_RES_TILE_BUFFER and
  38   * VC4_PACKET_LOAD_FULL_RES_TILE_BUFFER.
  39   */
  40 -#define VC4_LOADSTORE_FULL_RES_EOF                     (1 << 3)
  41 -#define VC4_LOADSTORE_FULL_RES_DISABLE_CLEAR_ALL       (1 << 2)
  42 -#define VC4_LOADSTORE_FULL_RES_DISABLE_ZS              (1 << 1)
  43 -#define VC4_LOADSTORE_FULL_RES_DISABLE_COLOR           (1 << 0)
  44 +#define VC4_LOADSTORE_FULL_RES_EOF                     BIT(3)
  45 +#define VC4_LOADSTORE_FULL_RES_DISABLE_CLEAR_ALL       BIT(2)
  46 +#define VC4_LOADSTORE_FULL_RES_DISABLE_ZS              BIT(1)
  47 +#define VC4_LOADSTORE_FULL_RES_DISABLE_COLOR           BIT(0)
  48 +
  49 +/** @{
  50 + *
  51 + * low bits of VC4_PACKET_STORE_FULL_RES_TILE_BUFFER and
  52 + * VC4_PACKET_LOAD_FULL_RES_TILE_BUFFER.
  53 + */
  54 +#define VC4_LOADSTORE_FULL_RES_EOF                     BIT(3)
  55 +#define VC4_LOADSTORE_FULL_RES_DISABLE_CLEAR_ALL       BIT(2)
  56 +#define VC4_LOADSTORE_FULL_RES_DISABLE_ZS              BIT(1)
  57 +#define VC4_LOADSTORE_FULL_RES_DISABLE_COLOR           BIT(0)
  58
  59  /** @{
  60   *
  61 --- a/drivers/gpu/drm/vc4/vc4_render_cl.c
  62 +++ b/drivers/gpu/drm/vc4/vc4_render_cl.c
  63 @@ -37,9 +37,11 @@
  64
  65  struct vc4_rcl_setup {
  66         struct drm_gem_cma_object *color_read;
  67 -       struct drm_gem_cma_object *color_ms_write;
  68 +       struct drm_gem_cma_object *color_write;
  69         struct drm_gem_cma_object *zs_read;
  70         struct drm_gem_cma_object *zs_write;
  71 +       struct drm_gem_cma_object *msaa_color_write;
  72 +       struct drm_gem_cma_object *msaa_zs_write;
  73
  74         struct drm_gem_cma_object *rcl;
  75         u32 next_offset;
  76 @@ -82,6 +84,22 @@ static void vc4_store_before_load(struct
  77  }
  78
  79  /*
  80 + * Calculates the physical address of the start of a tile in a RCL surface.
  81 + *
  82 + * Unlike the other load/store packets,
  83 + * VC4_PACKET_LOAD/STORE_FULL_RES_TILE_BUFFER don't look at the tile
  84 + * coordinates packet, and instead just store to the address given.
  85 + */
  86 +static uint32_t vc4_full_res_offset(struct vc4_exec_info *exec,
  87 +                                   struct drm_gem_cma_object *bo,
  88 +                                   struct drm_vc4_submit_rcl_surface *surf,
  89 +                                   uint8_t x, uint8_t y)
  90 +{
  91 +       return bo->paddr + surf->offset + VC4_TILE_BUFFER_SIZE *
  92 +               (DIV_ROUND_UP(exec->args->width, 32) * y + x);
  93 +}
  94 +
  95 +/*
  96   * Emits a PACKET_TILE_COORDINATES if one isn't already pending.
  97   *
  98   * The tile coordinates packet triggers a pending load if there is one, are
  99 @@ -108,22 +126,41 @@ static void emit_tile(struct vc4_exec_in
 100          * may be outstanding at a time.
 101          */
 102         if (setup->color_read) {
 103 -               rcl_u8(setup, VC4_PACKET_LOAD_TILE_BUFFER_GENERAL);
 104 -               rcl_u16(setup, args->color_read.bits);
 105 -               rcl_u32(setup,
 106 -                       setup->color_read->paddr + args->color_read.offset);
 107 +               if (args->color_read.flags &
 108 +                   VC4_SUBMIT_RCL_SURFACE_READ_IS_FULL_RES) {
 109 +                       rcl_u8(setup, VC4_PACKET_LOAD_FULL_RES_TILE_BUFFER);
 110 +                       rcl_u32(setup,
 111 +                               vc4_full_res_offset(exec, setup->color_read,
 112 +                                                   &args->color_read, x, y) |
 113 +                               VC4_LOADSTORE_FULL_RES_DISABLE_ZS);
 114 +               } else {
 115 +                       rcl_u8(setup, VC4_PACKET_LOAD_TILE_BUFFER_GENERAL);
 116 +                       rcl_u16(setup, args->color_read.bits);
 117 +                       rcl_u32(setup, setup->color_read->paddr +
 118 +                               args->color_read.offset);
 119 +               }
 120         }
 121
 122         if (setup->zs_read) {
 123 -               if (setup->color_read) {
 124 -                       /* Exec previous load. */
 125 -                       vc4_tile_coordinates(setup, x, y);
 126 -                       vc4_store_before_load(setup);
 127 +               if (args->zs_read.flags &
 128 +                   VC4_SUBMIT_RCL_SURFACE_READ_IS_FULL_RES) {
 129 +                       rcl_u8(setup, VC4_PACKET_LOAD_FULL_RES_TILE_BUFFER);
 130 +                       rcl_u32(setup,
 131 +                               vc4_full_res_offset(exec, setup->zs_read,
 132 +                                                   &args->zs_read, x, y) |
 133 +                               VC4_LOADSTORE_FULL_RES_DISABLE_COLOR);
 134 +               } else {
 135 +                       if (setup->color_read) {
 136 +                               /* Exec previous load. */
 137 +                               vc4_tile_coordinates(setup, x, y);
 138 +                               vc4_store_before_load(setup);
 139 +                       }
 140 +
 141 +                       rcl_u8(setup, VC4_PACKET_LOAD_TILE_BUFFER_GENERAL);
 142 +                       rcl_u16(setup, args->zs_read.bits);
 143 +                       rcl_u32(setup, setup->zs_read->paddr +
 144 +                               args->zs_read.offset);
 145                 }
 146 -
 147 -               rcl_u8(setup, VC4_PACKET_LOAD_TILE_BUFFER_GENERAL);
 148 -               rcl_u16(setup, args->zs_read.bits);
 149 -               rcl_u32(setup, setup->zs_read->paddr + args->zs_read.offset);
 150         }
 151
 152         /* Clipping depends on tile coordinates having been
 153 @@ -144,20 +181,60 @@ static void emit_tile(struct vc4_exec_in
 154                                 (y * exec->bin_tiles_x + x) * 32));
 155         }
 156
 157 +       if (setup->msaa_color_write) {
 158 +               bool last_tile_write = (!setup->msaa_zs_write &&
 159 +                                       !setup->zs_write &&
 160 +                                       !setup->color_write);
 161 +               uint32_t bits = VC4_LOADSTORE_FULL_RES_DISABLE_ZS;
 162 +
 163 +               if (!last_tile_write)
 164 +                       bits |= VC4_LOADSTORE_FULL_RES_DISABLE_CLEAR_ALL;
 165 +               else if (last)
 166 +                       bits |= VC4_LOADSTORE_FULL_RES_EOF;
 167 +               rcl_u8(setup, VC4_PACKET_STORE_FULL_RES_TILE_BUFFER);
 168 +               rcl_u32(setup,
 169 +                       vc4_full_res_offset(exec, setup->msaa_color_write,
 170 +                                           &args->msaa_color_write, x, y) |
 171 +                       bits);
 172 +       }
 173 +
 174 +       if (setup->msaa_zs_write) {
 175 +               bool last_tile_write = (!setup->zs_write &&
 176 +                                       !setup->color_write);
 177 +               uint32_t bits = VC4_LOADSTORE_FULL_RES_DISABLE_COLOR;
 178 +
 179 +               if (setup->msaa_color_write)
 180 +                       vc4_tile_coordinates(setup, x, y);
 181 +               if (!last_tile_write)
 182 +                       bits |= VC4_LOADSTORE_FULL_RES_DISABLE_CLEAR_ALL;
 183 +               else if (last)
 184 +                       bits |= VC4_LOADSTORE_FULL_RES_EOF;
 185 +               rcl_u8(setup, VC4_PACKET_STORE_FULL_RES_TILE_BUFFER);
 186 +               rcl_u32(setup,
 187 +                       vc4_full_res_offset(exec, setup->msaa_zs_write,
 188 +                                           &args->msaa_zs_write, x, y) |
 189 +                       bits);
 190 +       }
 191 +
 192         if (setup->zs_write) {
 193 +               bool last_tile_write = !setup->color_write;
 194 +
 195 +               if (setup->msaa_color_write || setup->msaa_zs_write)
 196 +                       vc4_tile_coordinates(setup, x, y);
 197 +
 198                 rcl_u8(setup, VC4_PACKET_STORE_TILE_BUFFER_GENERAL);
 199                 rcl_u16(setup, args->zs_write.bits |
 200 -                       (setup->color_ms_write ?
 201 -                        VC4_STORE_TILE_BUFFER_DISABLE_COLOR_CLEAR : 0));
 202 +                       (last_tile_write ?
 203 +                        0 : VC4_STORE_TILE_BUFFER_DISABLE_COLOR_CLEAR));
 204                 rcl_u32(setup,
 205                         (setup->zs_write->paddr + args->zs_write.offset) |
 206 -                       ((last && !setup->color_ms_write) ?
 207 +                       ((last && last_tile_write) ?
 208                          VC4_LOADSTORE_TILE_BUFFER_EOF : 0));
 209         }
 210
 211 -       if (setup->color_ms_write) {
 212 -               if (setup->zs_write) {
 213 -                       /* Reset after previous store */
 214 +       if (setup->color_write) {
 215 +               if (setup->msaa_color_write || setup->msaa_zs_write ||
 216 +                   setup->zs_write) {
 217                         vc4_tile_coordinates(setup, x, y);
 218                 }
 219
 220 @@ -192,14 +269,26 @@ static int vc4_create_rcl_bo(struct drm_
 221         }
 222
 223         if (setup->color_read) {
 224 -               loop_body_size += (VC4_PACKET_LOAD_TILE_BUFFER_GENERAL_SIZE);
 225 +               if (args->color_read.flags &
 226 +                   VC4_SUBMIT_RCL_SURFACE_READ_IS_FULL_RES) {
 227 +                       loop_body_size += VC4_PACKET_LOAD_FULL_RES_TILE_BUFFER_SIZE;
 228 +               } else {
 229 +                       loop_body_size += VC4_PACKET_LOAD_TILE_BUFFER_GENERAL_SIZE;
 230 +               }
 231         }
 232         if (setup->zs_read) {
 233 -               if (setup->color_read) {
 234 -                       loop_body_size += VC4_PACKET_TILE_COORDINATES_SIZE;
 235 -                       loop_body_size += VC4_PACKET_STORE_TILE_BUFFER_GENERAL_SIZE;
 236 +               if (args->zs_read.flags &
 237 +                   VC4_SUBMIT_RCL_SURFACE_READ_IS_FULL_RES) {
 238 +                       loop_body_size += VC4_PACKET_LOAD_FULL_RES_TILE_BUFFER_SIZE;
 239 +               } else {
 240 +                       if (setup->color_read &&
 241 +                           !(args->color_read.flags &
 242 +                             VC4_SUBMIT_RCL_SURFACE_READ_IS_FULL_RES)) {
 243 +                               loop_body_size += VC4_PACKET_TILE_COORDINATES_SIZE;
 244 +                               loop_body_size += VC4_PACKET_STORE_TILE_BUFFER_GENERAL_SIZE;
 245 +                       }
 246 +                       loop_body_size += VC4_PACKET_LOAD_TILE_BUFFER_GENERAL_SIZE;
 247                 }
 248 -               loop_body_size += VC4_PACKET_LOAD_TILE_BUFFER_GENERAL_SIZE;
 249         }
 250
 251         if (has_bin) {
 252 @@ -207,13 +296,23 @@ static int vc4_create_rcl_bo(struct drm_
 253                 loop_body_size += VC4_PACKET_BRANCH_TO_SUB_LIST_SIZE;
 254         }
 255
 256 +       if (setup->msaa_color_write)
 257 +               loop_body_size += VC4_PACKET_STORE_FULL_RES_TILE_BUFFER_SIZE;
 258 +       if (setup->msaa_zs_write)
 259 +               loop_body_size += VC4_PACKET_STORE_FULL_RES_TILE_BUFFER_SIZE;
 260 +
 261         if (setup->zs_write)
 262                 loop_body_size += VC4_PACKET_STORE_TILE_BUFFER_GENERAL_SIZE;
 263 -       if (setup->color_ms_write) {
 264 -               if (setup->zs_write)
 265 -                       loop_body_size += VC4_PACKET_TILE_COORDINATES_SIZE;
 266 +       if (setup->color_write)
 267                 loop_body_size += VC4_PACKET_STORE_MS_TILE_BUFFER_SIZE;
 268 -       }
 269 +
 270 +       /* We need a VC4_PACKET_TILE_COORDINATES in between each store. */
 271 +       loop_body_size += VC4_PACKET_TILE_COORDINATES_SIZE *
 272 +               ((setup->msaa_color_write != NULL) +
 273 +                (setup->msaa_zs_write != NULL) +
 274 +                (setup->color_write != NULL) +
 275 +                (setup->zs_write != NULL) - 1);
 276 +
 277         size += xtiles * ytiles * loop_body_size;
 278
 279         setup->rcl = &vc4_bo_create(dev, size, true)->base;
 280 @@ -224,13 +323,12 @@ static int vc4_create_rcl_bo(struct drm_
 281
 282         rcl_u8(setup, VC4_PACKET_TILE_RENDERING_MODE_CONFIG);
 283         rcl_u32(setup,
 284 -               (setup->color_ms_write ?
 285 -                (setup->color_ms_write->paddr +
 286 -                 args->color_ms_write.offset) :
 287 +               (setup->color_write ? (setup->color_write->paddr +
 288 +                                      args->color_write.offset) :
 289                  0));
 290         rcl_u16(setup, args->width);
 291         rcl_u16(setup, args->height);
 292 -       rcl_u16(setup, args->color_ms_write.bits);
 293 +       rcl_u16(setup, args->color_write.bits);
 294
 295         /* The tile buffer gets cleared when the previous tile is stored.  If
 296          * the clear values changed between frames, then the tile buffer has
 297 @@ -267,6 +365,56 @@ static int vc4_create_rcl_bo(struct drm_
 298         return 0;
 299  }
 300
 301 +static int vc4_full_res_bounds_check(struct vc4_exec_info *exec,
 302 +                                    struct drm_gem_cma_object *obj,
 303 +                                    struct drm_vc4_submit_rcl_surface *surf)
 304 +{
 305 +       struct drm_vc4_submit_cl *args = exec->args;
 306 +       u32 render_tiles_stride = DIV_ROUND_UP(exec->args->width, 32);
 307 +
 308 +       if (surf->offset > obj->base.size) {
 309 +               DRM_ERROR("surface offset %d > BO size %zd\n",
 310 +                         surf->offset, obj->base.size);
 311 +               return -EINVAL;
 312 +       }
 313 +
 314 +       if ((obj->base.size - surf->offset) / VC4_TILE_BUFFER_SIZE <
 315 +           render_tiles_stride * args->max_y_tile + args->max_x_tile) {
 316 +               DRM_ERROR("MSAA tile %d, %d out of bounds "
 317 +                         "(bo size %zd, offset %d).\n",
 318 +                         args->max_x_tile, args->max_y_tile,
 319 +                         obj->base.size,
 320 +                         surf->offset);
 321 +               return -EINVAL;
 322 +       }
 323 +
 324 +       return 0;
 325 +}
 326 +
 327 +static int vc4_rcl_msaa_surface_setup(struct vc4_exec_info *exec,
 328 +                                     struct drm_gem_cma_object **obj,
 329 +                                     struct drm_vc4_submit_rcl_surface *surf)
 330 +{
 331 +       if (surf->flags != 0 || surf->bits != 0) {
 332 +               DRM_ERROR("MSAA surface had nonzero flags/bits\n");
 333 +               return -EINVAL;
 334 +       }
 335 +
 336 +       if (surf->hindex == ~0)
 337 +               return 0;
 338 +
 339 +       *obj = vc4_use_bo(exec, surf->hindex);
 340 +       if (!*obj)
 341 +               return -EINVAL;
 342 +
 343 +       if (surf->offset & 0xf) {
 344 +               DRM_ERROR("MSAA write must be 16b aligned.\n");
 345 +               return -EINVAL;
 346 +       }
 347 +
 348 +       return vc4_full_res_bounds_check(exec, *obj, surf);
 349 +}
 350 +
 351  static int vc4_rcl_surface_setup(struct vc4_exec_info *exec,
 352                                  struct drm_gem_cma_object **obj,
 353                                  struct drm_vc4_submit_rcl_surface *surf)
 354 @@ -278,9 +426,10 @@ static int vc4_rcl_surface_setup(struct
 355         uint8_t format = VC4_GET_FIELD(surf->bits,
 356                                        VC4_LOADSTORE_TILE_BUFFER_FORMAT);
 357         int cpp;
 358 +       int ret;
 359
 360 -       if (surf->pad != 0) {
 361 -               DRM_ERROR("Padding unset\n");
 362 +       if (surf->flags & ~VC4_SUBMIT_RCL_SURFACE_READ_IS_FULL_RES) {
 363 +               DRM_ERROR("Extra flags set\n");
 364                 return -EINVAL;
 365         }
 366
 367 @@ -290,6 +439,25 @@ static int vc4_rcl_surface_setup(struct
 368         if (!vc4_use_bo(exec, surf->hindex, VC4_MODE_RENDER, obj))
 369                 return -EINVAL;
 370
 371 +       if (surf->flags & VC4_SUBMIT_RCL_SURFACE_READ_IS_FULL_RES) {
 372 +               if (surf == &exec->args->zs_write) {
 373 +                       DRM_ERROR("general zs write may not be a full-res.\n");
 374 +                       return -EINVAL;
 375 +               }
 376 +
 377 +               if (surf->bits != 0) {
 378 +                       DRM_ERROR("load/store general bits set with "
 379 +                                 "full res load/store.\n");
 380 +                       return -EINVAL;
 381 +               }
 382 +
 383 +               ret = vc4_full_res_bounds_check(exec, *obj, surf);
 384 +               if (!ret)
 385 +                       return ret;
 386 +
 387 +               return 0;
 388 +       }
 389 +
 390         if (surf->bits & ~(VC4_LOADSTORE_TILE_BUFFER_TILING_MASK |
 391                            VC4_LOADSTORE_TILE_BUFFER_BUFFER_MASK |
 392                            VC4_LOADSTORE_TILE_BUFFER_FORMAT_MASK)) {
 393 @@ -341,9 +509,10 @@ static int vc4_rcl_surface_setup(struct
 394  }
 395
 396  static int
 397 -vc4_rcl_ms_surface_setup(struct vc4_exec_info *exec,
 398 -                        struct drm_gem_cma_object **obj,
 399 -                        struct drm_vc4_submit_rcl_surface *surf)
 400 +vc4_rcl_render_config_surface_setup(struct vc4_exec_info *exec,
 401 +                                   struct vc4_rcl_setup *setup,
 402 +                                   struct drm_gem_cma_object **obj,
 403 +                                   struct drm_vc4_submit_rcl_surface *surf)
 404  {
 405         uint8_t tiling = VC4_GET_FIELD(surf->bits,
 406                                        VC4_RENDER_CONFIG_MEMORY_FORMAT);
 407 @@ -351,13 +520,15 @@ vc4_rcl_ms_surface_setup(struct vc4_exec
 408                                        VC4_RENDER_CONFIG_FORMAT);
 409         int cpp;
 410
 411 -       if (surf->pad != 0) {
 412 -               DRM_ERROR("Padding unset\n");
 413 +       if (surf->flags != 0) {
 414 +               DRM_ERROR("No flags supported on render config.\n");
 415                 return -EINVAL;
 416         }
 417
 418         if (surf->bits & ~(VC4_RENDER_CONFIG_MEMORY_FORMAT_MASK |
 419 -                          VC4_RENDER_CONFIG_FORMAT_MASK)) {
 420 +                          VC4_RENDER_CONFIG_FORMAT_MASK |
 421 +                          VC4_RENDER_CONFIG_MS_MODE_4X |
 422 +                          VC4_RENDER_CONFIG_DECIMATE_MODE_4X)) {
 423                 DRM_ERROR("Unknown bits in render config: 0x%04x\n",
 424                           surf->bits);
 425                 return -EINVAL;
 426 @@ -413,18 +584,20 @@ int vc4_get_rcl(struct drm_device *dev,
 427         if (has_bin &&
 428             (args->max_x_tile > exec->bin_tiles_x ||
 429              args->max_y_tile > exec->bin_tiles_y)) {
 430 -               DRM_ERROR("Render tiles (%d,%d) outside of bin config (%d,%d)\n",
 431 +               DRM_ERROR("Render tiles (%d,%d) outside of bin config "
 432 +                         "(%d,%d)\n",
 433                           args->max_x_tile, args->max_y_tile,
 434                           exec->bin_tiles_x, exec->bin_tiles_y);
 435                 return -EINVAL;
 436         }
 437
 438 -       ret = vc4_rcl_surface_setup(exec, &setup.color_read, &args->color_read);
 439 +       ret = vc4_rcl_render_config_surface_setup(exec, &setup,
 440 +                                                 &setup.color_write,
 441 +                                                 &args->color_write);
 442         if (ret)
 443                 return ret;
 444
 445 -       ret = vc4_rcl_ms_surface_setup(exec, &setup.color_ms_write,
 446 -                                      &args->color_ms_write);
 447 +       ret = vc4_rcl_surface_setup(exec, &setup.color_read, &args->color_read);
 448         if (ret)
 449                 return ret;
 450
 451 @@ -436,10 +609,21 @@ int vc4_get_rcl(struct drm_device *dev,
 452         if (ret)
 453                 return ret;
 454
 455 +       ret = vc4_rcl_msaa_surface_setup(exec, &setup.msaa_color_write,
 456 +                                        &args->msaa_color_write);
 457 +       if (ret)
 458 +               return ret;
 459 +
 460 +       ret = vc4_rcl_msaa_surface_setup(exec, &setup.msaa_zs_write,
 461 +                                        &args->msaa_zs_write);
 462 +       if (ret)
 463 +               return ret;
 464 +
 465         /* We shouldn't even have the job submitted to us if there's no
 466          * surface to write out.
 467          */
 468 -       if (!setup.color_ms_write && !setup.zs_write) {
 469 +       if (!setup.color_write && !setup.zs_write &&
 470 +           !setup.msaa_color_write && !setup.msaa_zs_write) {
 471                 DRM_ERROR("RCL requires color or Z/S write\n");
 472                 return -EINVAL;
 473         }
 474 --- a/drivers/gpu/drm/vc4/vc4_validate.c
 475 +++ b/drivers/gpu/drm/vc4/vc4_validate.c
 476 @@ -400,9 +400,8 @@ validate_tile_binning_config(VALIDATE_AR
 477         }
 478
 479         if (flags & (VC4_BIN_CONFIG_DB_NON_MS |
 480 -                    VC4_BIN_CONFIG_TILE_BUFFER_64BIT |
 481 -                    VC4_BIN_CONFIG_MS_MODE_4X)) {
 482 -               DRM_ERROR("unsupported bining config flags 0x%02x\n", flags);
 483 +                    VC4_BIN_CONFIG_TILE_BUFFER_64BIT)) {
 484 +               DRM_ERROR("unsupported binning config flags 0x%02x\n", flags);
 485                 return -EINVAL;
 486         }
 487
 488 --- a/include/uapi/drm/vc4_drm.h
 489 +++ b/include/uapi/drm/vc4_drm.h
 490 @@ -46,10 +46,13 @@ struct drm_vc4_submit_rcl_surface {
 491         uint32_t hindex; /* Handle index, or ~0 if not present. */
 492         uint32_t offset; /* Offset to start of buffer. */
 493         /*
 494 -         * Bits for either render config (color_ms_write) or load/store packet.
 495 +         * Bits for either render config (color_write) or load/store packet.
 496 +         * Bits should all be 0 for MSAA load/stores.
 497          */
 498         uint16_t bits;
 499 -       uint16_t pad;
 500 +
 501 +#define VC4_SUBMIT_RCL_SURFACE_READ_IS_FULL_RES                (1 << 0)
 502 +       uint16_t flags;
 503  };
 504
 505  /**
 506 @@ -128,9 +131,11 @@ struct drm_vc4_submit_cl {
 507         uint8_t max_x_tile;
 508         uint8_t max_y_tile;
 509         struct drm_vc4_submit_rcl_surface color_read;
 510 -       struct drm_vc4_submit_rcl_surface color_ms_write;
 511 +       struct drm_vc4_submit_rcl_surface color_write;
 512         struct drm_vc4_submit_rcl_surface zs_read;
 513         struct drm_vc4_submit_rcl_surface zs_write;
 514 +       struct drm_vc4_submit_rcl_surface msaa_color_write;
 515 +       struct drm_vc4_submit_rcl_surface msaa_zs_write;
 516         uint32_t clear_color[2];
 517         uint32_t clear_z;
 518         uint8_t clear_s;