drivers/gpu/drm/i915/gem/selftests/i915_gem_client_blt.c

   1 // SPDX-License-Identifier: MIT
   2 /*
   3  * Copyright © 2019 Intel Corporation
   4  */
   5
   6 #include "i915_selftest.h"
   7
   8 #include "gt/intel_context.h"
   9 #include "gt/intel_engine_regs.h"
  10 #include "gt/intel_engine_user.h"
  11 #include "gt/intel_gpu_commands.h"
  12 #include "gt/intel_gt.h"
  13 #include "gt/intel_gt_regs.h"
  14 #include "gem/i915_gem_lmem.h"
  15
  16 #include "selftests/igt_flush_test.h"
  17 #include "selftests/mock_drm.h"
  18 #include "selftests/i915_random.h"
  19 #include "huge_gem_object.h"
  20 #include "mock_context.h"
  21
  22 #define OW_SIZE 16                      /* in bytes */
  23 #define F_SUBTILE_SIZE 64               /* in bytes */
  24 #define F_TILE_WIDTH 128                /* in bytes */
  25 #define F_TILE_HEIGHT 32                /* in pixels */
  26 #define F_SUBTILE_WIDTH  OW_SIZE        /* in bytes */
  27 #define F_SUBTILE_HEIGHT 4              /* in pixels */
  28
  29 static int linear_x_y_to_ftiled_pos(int x, int y, u32 stride, int bpp)
  30 {
  31         int tile_base;
  32         int tile_x, tile_y;
  33         int swizzle, subtile;
  34         int pixel_size = bpp / 8;
  35         int pos;
  36
  37         /*
  38          * Subtile remapping for F tile. Note that map[a]==b implies map[b]==a
  39          * so we can use the same table to tile and until.
  40          */
  41         static const u8 f_subtile_map[] = {
  42                  0,  1,  2,  3,  8,  9, 10, 11,
  43                  4,  5,  6,  7, 12, 13, 14, 15,
  44                 16, 17, 18, 19, 24, 25, 26, 27,
  45                 20, 21, 22, 23, 28, 29, 30, 31,
  46                 32, 33, 34, 35, 40, 41, 42, 43,
  47                 36, 37, 38, 39, 44, 45, 46, 47,
  48                 48, 49, 50, 51, 56, 57, 58, 59,
  49                 52, 53, 54, 55, 60, 61, 62, 63
  50         };
  51
  52         x *= pixel_size;
  53         /*
  54          * Where does the 4k tile start (in bytes)?  This is the same for Y and
  55          * F so we can use the Y-tile algorithm to get to that point.
  56          */
  57         tile_base =
  58                 y / F_TILE_HEIGHT * stride * F_TILE_HEIGHT +
  59                 x / F_TILE_WIDTH * 4096;
  60
  61         /* Find pixel within tile */
  62         tile_x = x % F_TILE_WIDTH;
  63         tile_y = y % F_TILE_HEIGHT;
  64
  65         /* And figure out the subtile within the 4k tile */
  66         subtile = tile_y / F_SUBTILE_HEIGHT * 8 + tile_x / F_SUBTILE_WIDTH;
  67
  68         /* Swizzle the subtile number according to the bspec diagram */
  69         swizzle = f_subtile_map[subtile];
  70
  71         /* Calculate new position */
  72         pos = tile_base +
  73                 swizzle * F_SUBTILE_SIZE +
  74                 tile_y % F_SUBTILE_HEIGHT * OW_SIZE +
  75                 tile_x % F_SUBTILE_WIDTH;
  76
  77         GEM_BUG_ON(!IS_ALIGNED(pos, pixel_size));
  78
  79         return pos / pixel_size * 4;
  80 }
  81
  82 enum client_tiling {
  83         CLIENT_TILING_LINEAR,
  84         CLIENT_TILING_X,
  85         CLIENT_TILING_Y,
  86         CLIENT_TILING_4,
  87         CLIENT_NUM_TILING_TYPES
  88 };
  89
  90 #define WIDTH 512
  91 #define HEIGHT 32
  92
  93 struct blit_buffer {
  94         struct i915_vma *vma;
  95         u32 start_val;
  96         enum client_tiling tiling;
  97 };
  98
  99 struct tiled_blits {
 100         struct intel_context *ce;
 101         struct blit_buffer buffers[3];
 102         struct blit_buffer scratch;
 103         struct i915_vma *batch;
 104         u64 hole;
 105         u64 align;
 106         u32 width;
 107         u32 height;
 108 };
 109
 110 static bool supports_x_tiling(const struct drm_i915_private *i915)
 111 {
 112         int gen = GRAPHICS_VER(i915);
 113
 114         if (gen < 12)
 115                 return true;
 116
 117         if (!HAS_LMEM(i915) || IS_DG1(i915))
 118                 return false;
 119
 120         return true;
 121 }
 122
 123 static bool fast_blit_ok(const struct blit_buffer *buf)
 124 {
 125         int gen = GRAPHICS_VER(buf->vma->vm->i915);
 126
 127         if (gen < 9)
 128                 return false;
 129
 130         if (gen < 12)
 131                 return true;
 132
 133         /* filter out platforms with unsupported X-tile support in fastblit */
 134         if (buf->tiling == CLIENT_TILING_X && !supports_x_tiling(buf->vma->vm->i915))
 135                 return false;
 136
 137         return true;
 138 }
 139
 140 static int prepare_blit(const struct tiled_blits *t,
 141                         struct blit_buffer *dst,
 142                         struct blit_buffer *src,
 143                         struct drm_i915_gem_object *batch)
 144 {
 145         const int ver = GRAPHICS_VER(to_i915(batch->base.dev));
 146         bool use_64b_reloc = ver >= 8;
 147         u32 src_pitch, dst_pitch;
 148         u32 cmd, *cs;
 149
 150         cs = i915_gem_object_pin_map_unlocked(batch, I915_MAP_WC);
 151         if (IS_ERR(cs))
 152                 return PTR_ERR(cs);
 153
 154         if (fast_blit_ok(dst) && fast_blit_ok(src)) {
 155                 struct intel_gt *gt = t->ce->engine->gt;
 156                 u32 src_tiles = 0, dst_tiles = 0;
 157                 u32 src_4t = 0, dst_4t = 0;
 158
 159                 /* Need to program BLIT_CCTL if it is not done previously
 160                  * before using XY_FAST_COPY_BLT
 161                  */
 162                 *cs++ = MI_LOAD_REGISTER_IMM(1);
 163                 *cs++ = i915_mmio_reg_offset(BLIT_CCTL(t->ce->engine->mmio_base));
 164                 *cs++ = (BLIT_CCTL_SRC_MOCS(gt->mocs.uc_index) |
 165                          BLIT_CCTL_DST_MOCS(gt->mocs.uc_index));
 166
 167                 src_pitch = t->width; /* in dwords */
 168                 if (src->tiling == CLIENT_TILING_4) {
 169                         src_tiles = XY_FAST_COPY_BLT_D0_SRC_TILE_MODE(YMAJOR);
 170                         src_4t = XY_FAST_COPY_BLT_D1_SRC_TILE4;
 171                 } else if (src->tiling == CLIENT_TILING_Y) {
 172                         src_tiles = XY_FAST_COPY_BLT_D0_SRC_TILE_MODE(YMAJOR);
 173                 } else if (src->tiling == CLIENT_TILING_X) {
 174                         src_tiles = XY_FAST_COPY_BLT_D0_SRC_TILE_MODE(TILE_X);
 175                 } else {
 176                         src_pitch *= 4; /* in bytes */
 177                 }
 178
 179                 dst_pitch = t->width; /* in dwords */
 180                 if (dst->tiling == CLIENT_TILING_4) {
 181                         dst_tiles = XY_FAST_COPY_BLT_D0_DST_TILE_MODE(YMAJOR);
 182                         dst_4t = XY_FAST_COPY_BLT_D1_DST_TILE4;
 183                 } else if (dst->tiling == CLIENT_TILING_Y) {
 184                         dst_tiles = XY_FAST_COPY_BLT_D0_DST_TILE_MODE(YMAJOR);
 185                 } else if (dst->tiling == CLIENT_TILING_X) {
 186                         dst_tiles = XY_FAST_COPY_BLT_D0_DST_TILE_MODE(TILE_X);
 187                 } else {
 188                         dst_pitch *= 4; /* in bytes */
 189                 }
 190
 191                 *cs++ = GEN9_XY_FAST_COPY_BLT_CMD | (10 - 2) |
 192                         src_tiles | dst_tiles;
 193                 *cs++ = src_4t | dst_4t | BLT_DEPTH_32 | dst_pitch;
 194                 *cs++ = 0;
 195                 *cs++ = t->height << 16 | t->width;
 196                 *cs++ = lower_32_bits(dst->vma->node.start);
 197                 *cs++ = upper_32_bits(dst->vma->node.start);
 198                 *cs++ = 0;
 199                 *cs++ = src_pitch;
 200                 *cs++ = lower_32_bits(src->vma->node.start);
 201                 *cs++ = upper_32_bits(src->vma->node.start);
 202         } else {
 203                 if (ver >= 6) {
 204                         *cs++ = MI_LOAD_REGISTER_IMM(1);
 205                         *cs++ = i915_mmio_reg_offset(BCS_SWCTRL);
 206                         cmd = (BCS_SRC_Y | BCS_DST_Y) << 16;
 207                         if (src->tiling == CLIENT_TILING_Y)
 208                                 cmd |= BCS_SRC_Y;
 209                         if (dst->tiling == CLIENT_TILING_Y)
 210                                 cmd |= BCS_DST_Y;
 211                         *cs++ = cmd;
 212
 213                         cmd = MI_FLUSH_DW;
 214                         if (ver >= 8)
 215                                 cmd++;
 216                         *cs++ = cmd;
 217                         *cs++ = 0;
 218                         *cs++ = 0;
 219                         *cs++ = 0;
 220                 }
 221
 222                 cmd = XY_SRC_COPY_BLT_CMD | BLT_WRITE_RGBA | (8 - 2);
 223                 if (ver >= 8)
 224                         cmd += 2;
 225
 226                 src_pitch = t->width * 4;
 227                 if (src->tiling) {
 228                         cmd |= XY_SRC_COPY_BLT_SRC_TILED;
 229                         src_pitch /= 4;
 230                 }
 231
 232                 dst_pitch = t->width * 4;
 233                 if (dst->tiling) {
 234                         cmd |= XY_SRC_COPY_BLT_DST_TILED;
 235                         dst_pitch /= 4;
 236                 }
 237
 238                 *cs++ = cmd;
 239                 *cs++ = BLT_DEPTH_32 | BLT_ROP_SRC_COPY | dst_pitch;
 240                 *cs++ = 0;
 241                 *cs++ = t->height << 16 | t->width;
 242                 *cs++ = lower_32_bits(dst->vma->node.start);
 243                 if (use_64b_reloc)
 244                         *cs++ = upper_32_bits(dst->vma->node.start);
 245                 *cs++ = 0;
 246                 *cs++ = src_pitch;
 247                 *cs++ = lower_32_bits(src->vma->node.start);
 248                 if (use_64b_reloc)
 249                         *cs++ = upper_32_bits(src->vma->node.start);
 250         }
 251
 252         *cs++ = MI_BATCH_BUFFER_END;
 253
 254         i915_gem_object_flush_map(batch);
 255         i915_gem_object_unpin_map(batch);
 256
 257         return 0;
 258 }
 259
 260 static void tiled_blits_destroy_buffers(struct tiled_blits *t)
 261 {
 262         int i;
 263
 264         for (i = 0; i < ARRAY_SIZE(t->buffers); i++)
 265                 i915_vma_put(t->buffers[i].vma);
 266
 267         i915_vma_put(t->scratch.vma);
 268         i915_vma_put(t->batch);
 269 }
 270
 271 static struct i915_vma *
 272 __create_vma(struct tiled_blits *t, size_t size, bool lmem)
 273 {
 274         struct drm_i915_private *i915 = t->ce->vm->i915;
 275         struct drm_i915_gem_object *obj;
 276         struct i915_vma *vma;
 277
 278         if (lmem)
 279                 obj = i915_gem_object_create_lmem(i915, size, 0);
 280         else
 281                 obj = i915_gem_object_create_shmem(i915, size);
 282         if (IS_ERR(obj))
 283                 return ERR_CAST(obj);
 284
 285         vma = i915_vma_instance(obj, t->ce->vm, NULL);
 286         if (IS_ERR(vma))
 287                 i915_gem_object_put(obj);
 288
 289         return vma;
 290 }
 291
 292 static struct i915_vma *create_vma(struct tiled_blits *t, bool lmem)
 293 {
 294         return __create_vma(t, PAGE_ALIGN(t->width * t->height * 4), lmem);
 295 }
 296
 297 static int tiled_blits_create_buffers(struct tiled_blits *t,
 298                                       int width, int height,
 299                                       struct rnd_state *prng)
 300 {
 301         struct drm_i915_private *i915 = t->ce->engine->i915;
 302         int i;
 303
 304         t->width = width;
 305         t->height = height;
 306
 307         t->batch = __create_vma(t, PAGE_SIZE, false);
 308         if (IS_ERR(t->batch))
 309                 return PTR_ERR(t->batch);
 310
 311         t->scratch.vma = create_vma(t, false);
 312         if (IS_ERR(t->scratch.vma)) {
 313                 i915_vma_put(t->batch);
 314                 return PTR_ERR(t->scratch.vma);
 315         }
 316
 317         for (i = 0; i < ARRAY_SIZE(t->buffers); i++) {
 318                 struct i915_vma *vma;
 319
 320                 vma = create_vma(t, HAS_LMEM(i915) && i % 2);
 321                 if (IS_ERR(vma)) {
 322                         tiled_blits_destroy_buffers(t);
 323                         return PTR_ERR(vma);
 324                 }
 325
 326                 t->buffers[i].vma = vma;
 327                 t->buffers[i].tiling =
 328                         i915_prandom_u32_max_state(CLIENT_NUM_TILING_TYPES, prng);
 329
 330                 /* Platforms support either TileY or Tile4, not both */
 331                 if (HAS_4TILE(i915) && t->buffers[i].tiling == CLIENT_TILING_Y)
 332                         t->buffers[i].tiling = CLIENT_TILING_4;
 333                 else if (!HAS_4TILE(i915) && t->buffers[i].tiling == CLIENT_TILING_4)
 334                         t->buffers[i].tiling = CLIENT_TILING_Y;
 335         }
 336
 337         return 0;
 338 }
 339
 340 static void fill_scratch(struct tiled_blits *t, u32 *vaddr, u32 val)
 341 {
 342         int i;
 343
 344         t->scratch.start_val = val;
 345         for (i = 0; i < t->width * t->height; i++)
 346                 vaddr[i] = val++;
 347
 348         i915_gem_object_flush_map(t->scratch.vma->obj);
 349 }
 350
 351 static u64 swizzle_bit(unsigned int bit, u64 offset)
 352 {
 353         return (offset & BIT_ULL(bit)) >> (bit - 6);
 354 }
 355
 356 static u64 tiled_offset(const struct intel_gt *gt,
 357                         u64 v,
 358                         unsigned int stride,
 359                         enum client_tiling tiling,
 360                         int x_pos, int y_pos)
 361 {
 362         unsigned int swizzle;
 363         u64 x, y;
 364
 365         if (tiling == CLIENT_TILING_LINEAR)
 366                 return v;
 367
 368         y = div64_u64_rem(v, stride, &x);
 369
 370         if (tiling == CLIENT_TILING_4) {
 371                 v = linear_x_y_to_ftiled_pos(x_pos, y_pos, stride, 32);
 372
 373                 /* no swizzling for f-tiling */
 374                 swizzle = I915_BIT_6_SWIZZLE_NONE;
 375         } else if (tiling == CLIENT_TILING_X) {
 376                 v = div64_u64_rem(y, 8, &y) * stride * 8;
 377                 v += y * 512;
 378                 v += div64_u64_rem(x, 512, &x) << 12;
 379                 v += x;
 380
 381                 swizzle = gt->ggtt->bit_6_swizzle_x;
 382         } else {
 383                 const unsigned int ytile_span = 16;
 384                 const unsigned int ytile_height = 512;
 385
 386                 v = div64_u64_rem(y, 32, &y) * stride * 32;
 387                 v += y * ytile_span;
 388                 v += div64_u64_rem(x, ytile_span, &x) * ytile_height;
 389                 v += x;
 390
 391                 swizzle = gt->ggtt->bit_6_swizzle_y;
 392         }
 393
 394         switch (swizzle) {
 395         case I915_BIT_6_SWIZZLE_9:
 396                 v ^= swizzle_bit(9, v);
 397                 break;
 398         case I915_BIT_6_SWIZZLE_9_10:
 399                 v ^= swizzle_bit(9, v) ^ swizzle_bit(10, v);
 400                 break;
 401         case I915_BIT_6_SWIZZLE_9_11:
 402                 v ^= swizzle_bit(9, v) ^ swizzle_bit(11, v);
 403                 break;
 404         case I915_BIT_6_SWIZZLE_9_10_11:
 405                 v ^= swizzle_bit(9, v) ^ swizzle_bit(10, v) ^ swizzle_bit(11, v);
 406                 break;
 407         }
 408
 409         return v;
 410 }
 411
 412 static const char *repr_tiling(enum client_tiling tiling)
 413 {
 414         switch (tiling) {
 415         case CLIENT_TILING_LINEAR: return "linear";
 416         case CLIENT_TILING_X: return "X";
 417         case CLIENT_TILING_Y: return "Y";
 418         case CLIENT_TILING_4: return "F";
 419         default: return "unknown";
 420         }
 421 }
 422
 423 static int verify_buffer(const struct tiled_blits *t,
 424                          struct blit_buffer *buf,
 425                          struct rnd_state *prng)
 426 {
 427         const u32 *vaddr;
 428         int ret = 0;
 429         int x, y, p;
 430
 431         x = i915_prandom_u32_max_state(t->width, prng);
 432         y = i915_prandom_u32_max_state(t->height, prng);
 433         p = y * t->width + x;
 434
 435         vaddr = i915_gem_object_pin_map_unlocked(buf->vma->obj, I915_MAP_WC);
 436         if (IS_ERR(vaddr))
 437                 return PTR_ERR(vaddr);
 438
 439         if (vaddr[0] != buf->start_val) {
 440                 ret = -EINVAL;
 441         } else {
 442                 u64 v = tiled_offset(buf->vma->vm->gt,
 443                                      p * 4, t->width * 4,
 444                                      buf->tiling, x, y);
 445
 446                 if (vaddr[v / sizeof(*vaddr)] != buf->start_val + p)
 447                         ret = -EINVAL;
 448         }
 449         if (ret) {
 450                 pr_err("Invalid %s tiling detected at (%d, %d), start_val %x\n",
 451                        repr_tiling(buf->tiling),
 452                        x, y, buf->start_val);
 453                 igt_hexdump(vaddr, 4096);
 454         }
 455
 456         i915_gem_object_unpin_map(buf->vma->obj);
 457         return ret;
 458 }
 459
 460 static int move_to_active(struct i915_vma *vma,
 461                           struct i915_request *rq,
 462                           unsigned int flags)
 463 {
 464         int err;
 465
 466         i915_vma_lock(vma);
 467         err = i915_request_await_object(rq, vma->obj, false);
 468         if (err == 0)
 469                 err = i915_vma_move_to_active(vma, rq, flags);
 470         i915_vma_unlock(vma);
 471
 472         return err;
 473 }
 474
 475 static int pin_buffer(struct i915_vma *vma, u64 addr)
 476 {
 477         int err;
 478
 479         if (drm_mm_node_allocated(&vma->node) && vma->node.start != addr) {
 480                 err = i915_vma_unbind_unlocked(vma);
 481                 if (err)
 482                         return err;
 483         }
 484
 485         err = i915_vma_pin(vma, 0, 0, PIN_USER | PIN_OFFSET_FIXED | addr);
 486         if (err)
 487                 return err;
 488
 489         return 0;
 490 }
 491
 492 static int
 493 tiled_blit(struct tiled_blits *t,
 494            struct blit_buffer *dst, u64 dst_addr,
 495            struct blit_buffer *src, u64 src_addr)
 496 {
 497         struct i915_request *rq;
 498         int err;
 499
 500         err = pin_buffer(src->vma, src_addr);
 501         if (err) {
 502                 pr_err("Cannot pin src @ %llx\n", src_addr);
 503                 return err;
 504         }
 505
 506         err = pin_buffer(dst->vma, dst_addr);
 507         if (err) {
 508                 pr_err("Cannot pin dst @ %llx\n", dst_addr);
 509                 goto err_src;
 510         }
 511
 512         err = i915_vma_pin(t->batch, 0, 0, PIN_USER | PIN_HIGH);
 513         if (err) {
 514                 pr_err("cannot pin batch\n");
 515                 goto err_dst;
 516         }
 517
 518         err = prepare_blit(t, dst, src, t->batch->obj);
 519         if (err)
 520                 goto err_bb;
 521
 522         rq = intel_context_create_request(t->ce);
 523         if (IS_ERR(rq)) {
 524                 err = PTR_ERR(rq);
 525                 goto err_bb;
 526         }
 527
 528         err = move_to_active(t->batch, rq, 0);
 529         if (!err)
 530                 err = move_to_active(src->vma, rq, 0);
 531         if (!err)
 532                 err = move_to_active(dst->vma, rq, 0);
 533         if (!err)
 534                 err = rq->engine->emit_bb_start(rq,
 535                                                 t->batch->node.start,
 536                                                 t->batch->node.size,
 537                                                 0);
 538         i915_request_get(rq);
 539         i915_request_add(rq);
 540         if (i915_request_wait(rq, 0, HZ / 2) < 0)
 541                 err = -ETIME;
 542         i915_request_put(rq);
 543
 544         dst->start_val = src->start_val;
 545 err_bb:
 546         i915_vma_unpin(t->batch);
 547 err_dst:
 548         i915_vma_unpin(dst->vma);
 549 err_src:
 550         i915_vma_unpin(src->vma);
 551         return err;
 552 }
 553
 554 static struct tiled_blits *
 555 tiled_blits_create(struct intel_engine_cs *engine, struct rnd_state *prng)
 556 {
 557         struct drm_mm_node hole;
 558         struct tiled_blits *t;
 559         u64 hole_size;
 560         int err;
 561
 562         t = kzalloc(sizeof(*t), GFP_KERNEL);
 563         if (!t)
 564                 return ERR_PTR(-ENOMEM);
 565
 566         t->ce = intel_context_create(engine);
 567         if (IS_ERR(t->ce)) {
 568                 err = PTR_ERR(t->ce);
 569                 goto err_free;
 570         }
 571
 572         t->align = i915_vm_min_alignment(t->ce->vm, INTEL_MEMORY_LOCAL);
 573         t->align = max(t->align,
 574                        i915_vm_min_alignment(t->ce->vm, INTEL_MEMORY_SYSTEM));
 575
 576         hole_size = 2 * round_up(WIDTH * HEIGHT * 4, t->align);
 577         hole_size *= 2; /* room to maneuver */
 578         hole_size += 2 * t->align; /* padding on either side */
 579
 580         mutex_lock(&t->ce->vm->mutex);
 581         memset(&hole, 0, sizeof(hole));
 582         err = drm_mm_insert_node_in_range(&t->ce->vm->mm, &hole,
 583                                           hole_size, t->align,
 584                                           I915_COLOR_UNEVICTABLE,
 585                                           0, U64_MAX,
 586                                           DRM_MM_INSERT_BEST);
 587         if (!err)
 588                 drm_mm_remove_node(&hole);
 589         mutex_unlock(&t->ce->vm->mutex);
 590         if (err) {
 591                 err = -ENODEV;
 592                 goto err_put;
 593         }
 594
 595         t->hole = hole.start + t->align;
 596         pr_info("Using hole at %llx\n", t->hole);
 597
 598         err = tiled_blits_create_buffers(t, WIDTH, HEIGHT, prng);
 599         if (err)
 600                 goto err_put;
 601
 602         return t;
 603
 604 err_put:
 605         intel_context_put(t->ce);
 606 err_free:
 607         kfree(t);
 608         return ERR_PTR(err);
 609 }
 610
 611 static void tiled_blits_destroy(struct tiled_blits *t)
 612 {
 613         tiled_blits_destroy_buffers(t);
 614
 615         intel_context_put(t->ce);
 616         kfree(t);
 617 }
 618
 619 static int tiled_blits_prepare(struct tiled_blits *t,
 620                                struct rnd_state *prng)
 621 {
 622         u64 offset = round_up(t->width * t->height * 4, t->align);
 623         u32 *map;
 624         int err;
 625         int i;
 626
 627         map = i915_gem_object_pin_map_unlocked(t->scratch.vma->obj, I915_MAP_WC);
 628         if (IS_ERR(map))
 629                 return PTR_ERR(map);
 630
 631         /* Use scratch to fill objects */
 632         for (i = 0; i < ARRAY_SIZE(t->buffers); i++) {
 633                 fill_scratch(t, map, prandom_u32_state(prng));
 634                 GEM_BUG_ON(verify_buffer(t, &t->scratch, prng));
 635
 636                 err = tiled_blit(t,
 637                                  &t->buffers[i], t->hole + offset,
 638                                  &t->scratch, t->hole);
 639                 if (err == 0)
 640                         err = verify_buffer(t, &t->buffers[i], prng);
 641                 if (err) {
 642                         pr_err("Failed to create buffer %d\n", i);
 643                         break;
 644                 }
 645         }
 646
 647         i915_gem_object_unpin_map(t->scratch.vma->obj);
 648         return err;
 649 }
 650
 651 static int tiled_blits_bounce(struct tiled_blits *t, struct rnd_state *prng)
 652 {
 653         u64 offset = round_up(t->width * t->height * 4, 2 * t->align);
 654         int err;
 655
 656         /* We want to check position invariant tiling across GTT eviction */
 657
 658         err = tiled_blit(t,
 659                          &t->buffers[1], t->hole + offset / 2,
 660                          &t->buffers[0], t->hole + 2 * offset);
 661         if (err)
 662                 return err;
 663
 664         /* Simulating GTT eviction of the same buffer / layout */
 665         t->buffers[2].tiling = t->buffers[0].tiling;
 666
 667         /* Reposition so that we overlap the old addresses, and slightly off */
 668         err = tiled_blit(t,
 669                          &t->buffers[2], t->hole + t->align,
 670                          &t->buffers[1], t->hole + 3 * offset / 2);
 671         if (err)
 672                 return err;
 673
 674         err = verify_buffer(t, &t->buffers[2], prng);
 675         if (err)
 676                 return err;
 677
 678         return 0;
 679 }
 680
 681 static int __igt_client_tiled_blits(struct intel_engine_cs *engine,
 682                                     struct rnd_state *prng)
 683 {
 684         struct tiled_blits *t;
 685         int err;
 686
 687         t = tiled_blits_create(engine, prng);
 688         if (IS_ERR(t))
 689                 return PTR_ERR(t);
 690
 691         err = tiled_blits_prepare(t, prng);
 692         if (err)
 693                 goto out;
 694
 695         err = tiled_blits_bounce(t, prng);
 696         if (err)
 697                 goto out;
 698
 699 out:
 700         tiled_blits_destroy(t);
 701         return err;
 702 }
 703
 704 static bool has_bit17_swizzle(int sw)
 705 {
 706         return (sw == I915_BIT_6_SWIZZLE_9_10_17 ||
 707                 sw == I915_BIT_6_SWIZZLE_9_17);
 708 }
 709
 710 static bool bad_swizzling(struct drm_i915_private *i915)
 711 {
 712         struct i915_ggtt *ggtt = to_gt(i915)->ggtt;
 713
 714         if (i915->gem_quirks & GEM_QUIRK_PIN_SWIZZLED_PAGES)
 715                 return true;
 716
 717         if (has_bit17_swizzle(ggtt->bit_6_swizzle_x) ||
 718             has_bit17_swizzle(ggtt->bit_6_swizzle_y))
 719                 return true;
 720
 721         return false;
 722 }
 723
 724 static int igt_client_tiled_blits(void *arg)
 725 {
 726         struct drm_i915_private *i915 = arg;
 727         I915_RND_STATE(prng);
 728         int inst = 0;
 729
 730         /* Test requires explicit BLT tiling controls */
 731         if (GRAPHICS_VER(i915) < 4)
 732                 return 0;
 733
 734         if (bad_swizzling(i915)) /* Requires sane (sub-page) swizzling */
 735                 return 0;
 736
 737         do {
 738                 struct intel_engine_cs *engine;
 739                 int err;
 740
 741                 engine = intel_engine_lookup_user(i915,
 742                                                   I915_ENGINE_CLASS_COPY,
 743                                                   inst++);
 744                 if (!engine)
 745                         return 0;
 746
 747                 err = __igt_client_tiled_blits(engine, &prng);
 748                 if (err == -ENODEV)
 749                         err = 0;
 750                 if (err)
 751                         return err;
 752         } while (1);
 753 }
 754
 755 int i915_gem_client_blt_live_selftests(struct drm_i915_private *i915)
 756 {
 757         static const struct i915_subtest tests[] = {
 758                 SUBTEST(igt_client_tiled_blits),
 759         };
 760
 761         if (intel_gt_is_wedged(to_gt(i915)))
 762                 return 0;
 763
 764         return i915_live_subtests(tests, i915);
 765 }