1 /* 2 * Copyright 2008 Jerome Glisse. 3 * All Rights Reserved. 4 * 5 * Permission is hereby granted, free of charge, to any person obtaining a 6 * copy of this software and associated documentation files (the "Software"), 7 * to deal in the Software without restriction, including without limitation 8 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 9 * and/or sell copies of the Software, and to permit persons to whom the 10 * Software is furnished to do so, subject to the following conditions: 11 * 12 * The above copyright notice and this permission notice (including the next 13 * paragraph) shall be included in all copies or substantial portions of the 14 * Software. 15 * 16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 19 * PRECISION INSIGHT AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR 20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER 22 * DEALINGS IN THE SOFTWARE. 23 * 24 * Authors: 25 * Jerome Glisse <glisse@freedesktop.org> 26 */ 27 #include <linux/list_sort.h> 28 #include <drm/drmP.h> 29 #include <uapi_drm/radeon_drm.h> 30 #include "radeon_reg.h" 31 #include "radeon.h" 32 #ifdef TRACE_TODO 33 #include "radeon_trace.h" 34 #endif 35 36 #define RADEON_CS_MAX_PRIORITY 32u 37 #define RADEON_CS_NUM_BUCKETS (RADEON_CS_MAX_PRIORITY + 1) 38 39 /* This is based on the bucket sort with O(n) time complexity. 40 * An item with priority "i" is added to bucket[i]. The lists are then 41 * concatenated in descending order. 42 */ 43 struct radeon_cs_buckets { 44 struct list_head bucket[RADEON_CS_NUM_BUCKETS]; 45 }; 46 47 static void radeon_cs_buckets_init(struct radeon_cs_buckets *b) 48 { 49 unsigned i; 50 51 for (i = 0; i < RADEON_CS_NUM_BUCKETS; i++) 52 INIT_LIST_HEAD(&b->bucket[i]); 53 } 54 55 static void radeon_cs_buckets_add(struct radeon_cs_buckets *b, 56 struct list_head *item, unsigned priority) 57 { 58 /* Since buffers which appear sooner in the relocation list are 59 * likely to be used more often than buffers which appear later 60 * in the list, the sort mustn't change the ordering of buffers 61 * with the same priority, i.e. it must be stable. 62 */ 63 list_add_tail(item, &b->bucket[min(priority, RADEON_CS_MAX_PRIORITY)]); 64 } 65 66 static void radeon_cs_buckets_get_list(struct radeon_cs_buckets *b, 67 struct list_head *out_list) 68 { 69 unsigned i; 70 71 /* Connect the sorted buckets in the output list. */ 72 for (i = 0; i < RADEON_CS_NUM_BUCKETS; i++) { 73 list_splice(&b->bucket[i], out_list); 74 } 75 } 76 77 static int radeon_cs_parser_relocs(struct radeon_cs_parser *p) 78 { 79 struct radeon_cs_chunk *chunk; 80 struct radeon_cs_buckets buckets; 81 unsigned i, j; 82 bool duplicate; 83 84 if (p->chunk_relocs_idx == -1) { 85 return 0; 86 } 87 chunk = &p->chunks[p->chunk_relocs_idx]; 88 p->dma_reloc_idx = 0; 89 /* FIXME: we assume that each relocs use 4 dwords */ 90 p->nrelocs = chunk->length_dw / 4; 91 p->relocs_ptr = kcalloc(p->nrelocs, sizeof(void *), GFP_KERNEL); 92 if (p->relocs_ptr == NULL) { 93 return -ENOMEM; 94 } 95 p->relocs = kcalloc(p->nrelocs, sizeof(struct radeon_cs_reloc), GFP_KERNEL); 96 if (p->relocs == NULL) { 97 return -ENOMEM; 98 } 99 100 radeon_cs_buckets_init(&buckets); 101 102 for (i = 0; i < p->nrelocs; i++) { 103 struct drm_radeon_cs_reloc *r; 104 unsigned priority; 105 106 duplicate = false; 107 r = (struct drm_radeon_cs_reloc *)&chunk->kdata[i*4]; 108 for (j = 0; j < i; j++) { 109 if (r->handle == p->relocs[j].handle) { 110 p->relocs_ptr[i] = &p->relocs[j]; 111 duplicate = true; 112 break; 113 } 114 } 115 if (duplicate) { 116 p->relocs[i].handle = 0; 117 continue; 118 } 119 120 p->relocs[i].gobj = drm_gem_object_lookup(p->filp, r->handle); 121 if (p->relocs[i].gobj == NULL) { 122 DRM_ERROR("gem object lookup failed 0x%x\n", 123 r->handle); 124 return -ENOENT; 125 } 126 p->relocs_ptr[i] = &p->relocs[i]; 127 p->relocs[i].robj = gem_to_radeon_bo(p->relocs[i].gobj); 128 129 /* The userspace buffer priorities are from 0 to 15. A higher 130 * number means the buffer is more important. 131 * Also, the buffers used for write have a higher priority than 132 * the buffers used for read only, which doubles the range 133 * to 0 to 31. 32 is reserved for the kernel driver. 134 */ 135 priority = (r->flags & RADEON_RELOC_PRIO_MASK) * 2 136 + !!r->write_domain; 137 138 /* the first reloc of an UVD job is the msg and that must be in 139 VRAM, also but everything into VRAM on AGP cards and older 140 IGP chips to avoid image corruptions */ 141 if (p->ring == R600_RING_TYPE_UVD_INDEX && 142 (i == 0 || (p->rdev->flags & RADEON_IS_AGP) || 143 p->rdev->family == CHIP_RS780 || 144 p->rdev->family == CHIP_RS880)) { 145 146 /* TODO: is this still needed for NI+ ? */ 147 p->relocs[i].prefered_domains = 148 RADEON_GEM_DOMAIN_VRAM; 149 150 p->relocs[i].allowed_domains = 151 RADEON_GEM_DOMAIN_VRAM; 152 153 /* prioritize this over any other relocation */ 154 priority = RADEON_CS_MAX_PRIORITY; 155 } else { 156 uint32_t domain = r->write_domain ? 157 r->write_domain : r->read_domains; 158 159 if (domain & RADEON_GEM_DOMAIN_CPU) { 160 DRM_ERROR("RADEON_GEM_DOMAIN_CPU is not valid " 161 "for command submission\n"); 162 return -EINVAL; 163 } 164 165 p->relocs[i].prefered_domains = domain; 166 if (domain == RADEON_GEM_DOMAIN_VRAM) 167 domain |= RADEON_GEM_DOMAIN_GTT; 168 p->relocs[i].allowed_domains = domain; 169 } 170 171 p->relocs[i].tv.bo = &p->relocs[i].robj->tbo; 172 p->relocs[i].handle = r->handle; 173 174 radeon_cs_buckets_add(&buckets, &p->relocs[i].tv.head, 175 priority); 176 } 177 178 radeon_cs_buckets_get_list(&buckets, &p->validated); 179 180 if (p->cs_flags & RADEON_CS_USE_VM) 181 p->vm_bos = radeon_vm_get_bos(p->rdev, p->ib.vm, 182 &p->validated); 183 184 return radeon_bo_list_validate(p->rdev, &p->ticket, &p->validated, p->ring); 185 } 186 187 static int radeon_cs_get_ring(struct radeon_cs_parser *p, u32 ring, s32 priority) 188 { 189 p->priority = priority; 190 191 switch (ring) { 192 default: 193 DRM_ERROR("unknown ring id: %d\n", ring); 194 return -EINVAL; 195 case RADEON_CS_RING_GFX: 196 p->ring = RADEON_RING_TYPE_GFX_INDEX; 197 break; 198 case RADEON_CS_RING_COMPUTE: 199 if (p->rdev->family >= CHIP_TAHITI) { 200 if (p->priority > 0) 201 p->ring = CAYMAN_RING_TYPE_CP1_INDEX; 202 else 203 p->ring = CAYMAN_RING_TYPE_CP2_INDEX; 204 } else 205 p->ring = RADEON_RING_TYPE_GFX_INDEX; 206 break; 207 case RADEON_CS_RING_DMA: 208 if (p->rdev->family >= CHIP_CAYMAN) { 209 if (p->priority > 0) 210 p->ring = R600_RING_TYPE_DMA_INDEX; 211 else 212 p->ring = CAYMAN_RING_TYPE_DMA1_INDEX; 213 } else if (p->rdev->family >= CHIP_RV770) { 214 p->ring = R600_RING_TYPE_DMA_INDEX; 215 } else { 216 return -EINVAL; 217 } 218 break; 219 case RADEON_CS_RING_UVD: 220 p->ring = R600_RING_TYPE_UVD_INDEX; 221 break; 222 case RADEON_CS_RING_VCE: 223 /* TODO: only use the low priority ring for now */ 224 p->ring = TN_RING_TYPE_VCE1_INDEX; 225 break; 226 } 227 return 0; 228 } 229 230 static void radeon_cs_sync_rings(struct radeon_cs_parser *p) 231 { 232 int i; 233 234 for (i = 0; i < p->nrelocs; i++) { 235 if (!p->relocs[i].robj) 236 continue; 237 238 radeon_semaphore_sync_to(p->ib.semaphore, 239 p->relocs[i].robj->tbo.sync_obj); 240 } 241 } 242 243 /* XXX: note that this is called from the legacy UMS CS ioctl as well */ 244 int radeon_cs_parser_init(struct radeon_cs_parser *p, void *data) 245 { 246 struct drm_radeon_cs *cs = data; 247 uint64_t *chunk_array_ptr; 248 unsigned size, i; 249 u32 ring = RADEON_CS_RING_GFX; 250 s32 priority = 0; 251 252 if (!cs->num_chunks) { 253 return 0; 254 } 255 /* get chunks */ 256 INIT_LIST_HEAD(&p->validated); 257 p->idx = 0; 258 p->ib.sa_bo = NULL; 259 p->ib.semaphore = NULL; 260 p->const_ib.sa_bo = NULL; 261 p->const_ib.semaphore = NULL; 262 p->chunk_ib_idx = -1; 263 p->chunk_relocs_idx = -1; 264 p->chunk_flags_idx = -1; 265 p->chunk_const_ib_idx = -1; 266 p->chunks_array = kcalloc(cs->num_chunks, sizeof(uint64_t), GFP_KERNEL); 267 if (p->chunks_array == NULL) { 268 return -ENOMEM; 269 } 270 chunk_array_ptr = (uint64_t *)(unsigned long)(cs->chunks); 271 if (copy_from_user(p->chunks_array, chunk_array_ptr, 272 sizeof(uint64_t)*cs->num_chunks)) { 273 return -EFAULT; 274 } 275 p->cs_flags = 0; 276 p->nchunks = cs->num_chunks; 277 p->chunks = kcalloc(p->nchunks, sizeof(struct radeon_cs_chunk), GFP_KERNEL); 278 if (p->chunks == NULL) { 279 return -ENOMEM; 280 } 281 for (i = 0; i < p->nchunks; i++) { 282 struct drm_radeon_cs_chunk __user **chunk_ptr = NULL; 283 struct drm_radeon_cs_chunk user_chunk; 284 uint32_t __user *cdata; 285 286 chunk_ptr = (void __user*)(unsigned long)p->chunks_array[i]; 287 if (copy_from_user(&user_chunk, chunk_ptr, 288 sizeof(struct drm_radeon_cs_chunk))) { 289 return -EFAULT; 290 } 291 p->chunks[i].length_dw = user_chunk.length_dw; 292 p->chunks[i].chunk_id = user_chunk.chunk_id; 293 if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_RELOCS) { 294 p->chunk_relocs_idx = i; 295 } 296 if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_IB) { 297 p->chunk_ib_idx = i; 298 /* zero length IB isn't useful */ 299 if (p->chunks[i].length_dw == 0) 300 return -EINVAL; 301 } 302 if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_CONST_IB) { 303 p->chunk_const_ib_idx = i; 304 /* zero length CONST IB isn't useful */ 305 if (p->chunks[i].length_dw == 0) 306 return -EINVAL; 307 } 308 if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_FLAGS) { 309 p->chunk_flags_idx = i; 310 /* zero length flags aren't useful */ 311 if (p->chunks[i].length_dw == 0) 312 return -EINVAL; 313 } 314 315 size = p->chunks[i].length_dw; 316 cdata = (void __user *)(unsigned long)user_chunk.chunk_data; 317 p->chunks[i].user_ptr = cdata; 318 if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_CONST_IB) 319 continue; 320 321 if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_IB) { 322 if (!p->rdev || !(p->rdev->flags & RADEON_IS_AGP)) 323 continue; 324 } 325 326 p->chunks[i].kdata = drm_malloc_ab(size, sizeof(uint32_t)); 327 size *= sizeof(uint32_t); 328 if (p->chunks[i].kdata == NULL) { 329 return -ENOMEM; 330 } 331 if (copy_from_user(p->chunks[i].kdata, cdata, size)) { 332 return -EFAULT; 333 } 334 if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_FLAGS) { 335 p->cs_flags = p->chunks[i].kdata[0]; 336 if (p->chunks[i].length_dw > 1) 337 ring = p->chunks[i].kdata[1]; 338 if (p->chunks[i].length_dw > 2) 339 priority = (s32)p->chunks[i].kdata[2]; 340 } 341 } 342 343 /* these are KMS only */ 344 if (p->rdev) { 345 if ((p->cs_flags & RADEON_CS_USE_VM) && 346 !p->rdev->vm_manager.enabled) { 347 DRM_ERROR("VM not active on asic!\n"); 348 return -EINVAL; 349 } 350 351 if (radeon_cs_get_ring(p, ring, priority)) 352 return -EINVAL; 353 354 /* we only support VM on some SI+ rings */ 355 if ((p->cs_flags & RADEON_CS_USE_VM) == 0) { 356 if (p->rdev->asic->ring[p->ring]->cs_parse == NULL) { 357 DRM_ERROR("Ring %d requires VM!\n", p->ring); 358 return -EINVAL; 359 } 360 } else { 361 if (p->rdev->asic->ring[p->ring]->ib_parse == NULL) { 362 DRM_ERROR("VM not supported on ring %d!\n", 363 p->ring); 364 return -EINVAL; 365 } 366 } 367 } 368 369 return 0; 370 } 371 372 static int cmp_size_smaller_first(void *priv, struct list_head *a, 373 struct list_head *b) 374 { 375 struct radeon_cs_reloc *la = list_entry(a, struct radeon_cs_reloc, tv.head); 376 struct radeon_cs_reloc *lb = list_entry(b, struct radeon_cs_reloc, tv.head); 377 378 /* Sort A before B if A is smaller. */ 379 return (int)la->robj->tbo.num_pages - (int)lb->robj->tbo.num_pages; 380 } 381 382 /** 383 * cs_parser_fini() - clean parser states 384 * @parser: parser structure holding parsing context. 385 * @error: error number 386 * 387 * If error is set than unvalidate buffer, otherwise just free memory 388 * used by parsing context. 389 **/ 390 static void radeon_cs_parser_fini(struct radeon_cs_parser *parser, int error, bool backoff) 391 { 392 unsigned i; 393 394 if (!error) { 395 /* Sort the buffer list from the smallest to largest buffer, 396 * which affects the order of buffers in the LRU list. 397 * This assures that the smallest buffers are added first 398 * to the LRU list, so they are likely to be later evicted 399 * first, instead of large buffers whose eviction is more 400 * expensive. 401 * 402 * This slightly lowers the number of bytes moved by TTM 403 * per frame under memory pressure. 404 */ 405 list_sort(NULL, &parser->validated, cmp_size_smaller_first); 406 407 ttm_eu_fence_buffer_objects(&parser->ticket, 408 &parser->validated, 409 parser->ib.fence); 410 } else if (backoff) { 411 ttm_eu_backoff_reservation(&parser->ticket, 412 &parser->validated); 413 } 414 415 if (parser->relocs != NULL) { 416 for (i = 0; i < parser->nrelocs; i++) { 417 if (parser->relocs[i].gobj) 418 drm_gem_object_unreference_unlocked(parser->relocs[i].gobj); 419 } 420 } 421 kfree(parser->track); 422 kfree(parser->relocs); 423 kfree(parser->relocs_ptr); 424 drm_free_large(parser->vm_bos); 425 for (i = 0; i < parser->nchunks; i++) 426 drm_free_large(parser->chunks[i].kdata); 427 kfree(parser->chunks); 428 kfree(parser->chunks_array); 429 radeon_ib_free(parser->rdev, &parser->ib); 430 radeon_ib_free(parser->rdev, &parser->const_ib); 431 } 432 433 static int radeon_cs_ib_chunk(struct radeon_device *rdev, 434 struct radeon_cs_parser *parser) 435 { 436 int r; 437 438 if (parser->chunk_ib_idx == -1) 439 return 0; 440 441 if (parser->cs_flags & RADEON_CS_USE_VM) 442 return 0; 443 444 r = radeon_cs_parse(rdev, parser->ring, parser); 445 if (r || parser->parser_error) { 446 DRM_ERROR("Invalid command stream !\n"); 447 return r; 448 } 449 450 if (parser->ring == R600_RING_TYPE_UVD_INDEX) 451 radeon_uvd_note_usage(rdev); 452 else if ((parser->ring == TN_RING_TYPE_VCE1_INDEX) || 453 (parser->ring == TN_RING_TYPE_VCE2_INDEX)) 454 radeon_vce_note_usage(rdev); 455 456 radeon_cs_sync_rings(parser); 457 r = radeon_ib_schedule(rdev, &parser->ib, NULL, true); 458 if (r) { 459 DRM_ERROR("Failed to schedule IB !\n"); 460 } 461 return r; 462 } 463 464 static int radeon_bo_vm_update_pte(struct radeon_cs_parser *p, 465 struct radeon_vm *vm) 466 { 467 struct radeon_device *rdev = p->rdev; 468 struct radeon_bo_va *bo_va; 469 int i, r; 470 471 r = radeon_vm_update_page_directory(rdev, vm); 472 if (r) 473 return r; 474 475 r = radeon_vm_clear_freed(rdev, vm); 476 if (r) 477 return r; 478 479 if (vm->ib_bo_va == NULL) { 480 DRM_ERROR("Tmp BO not in VM!\n"); 481 return -EINVAL; 482 } 483 484 r = radeon_vm_bo_update(rdev, vm->ib_bo_va, 485 &rdev->ring_tmp_bo.bo->tbo.mem); 486 if (r) 487 return r; 488 489 for (i = 0; i < p->nrelocs; i++) { 490 struct radeon_bo *bo; 491 492 /* ignore duplicates */ 493 if (p->relocs_ptr[i] != &p->relocs[i]) 494 continue; 495 496 bo = p->relocs[i].robj; 497 bo_va = radeon_vm_bo_find(vm, bo); 498 if (bo_va == NULL) { 499 dev_err(rdev->dev, "bo %p not in vm %p\n", bo, vm); 500 return -EINVAL; 501 } 502 503 r = radeon_vm_bo_update(rdev, bo_va, &bo->tbo.mem); 504 if (r) 505 return r; 506 } 507 508 return radeon_vm_clear_invalids(rdev, vm); 509 } 510 511 static int radeon_cs_ib_vm_chunk(struct radeon_device *rdev, 512 struct radeon_cs_parser *parser) 513 { 514 struct radeon_fpriv *fpriv = parser->filp->driver_priv; 515 struct radeon_vm *vm = &fpriv->vm; 516 int r; 517 518 if (parser->chunk_ib_idx == -1) 519 return 0; 520 if ((parser->cs_flags & RADEON_CS_USE_VM) == 0) 521 return 0; 522 523 if (parser->const_ib.length_dw) { 524 r = radeon_ring_ib_parse(rdev, parser->ring, &parser->const_ib); 525 if (r) { 526 return r; 527 } 528 } 529 530 r = radeon_ring_ib_parse(rdev, parser->ring, &parser->ib); 531 if (r) { 532 return r; 533 } 534 535 if (parser->ring == R600_RING_TYPE_UVD_INDEX) 536 radeon_uvd_note_usage(rdev); 537 538 lockmgr(&vm->mutex, LK_EXCLUSIVE); 539 r = radeon_bo_vm_update_pte(parser, vm); 540 if (r) { 541 goto out; 542 } 543 radeon_cs_sync_rings(parser); 544 radeon_semaphore_sync_to(parser->ib.semaphore, vm->fence); 545 546 if ((rdev->family >= CHIP_TAHITI) && 547 (parser->chunk_const_ib_idx != -1)) { 548 r = radeon_ib_schedule(rdev, &parser->ib, &parser->const_ib, true); 549 } else { 550 r = radeon_ib_schedule(rdev, &parser->ib, NULL, true); 551 } 552 553 out: 554 lockmgr(&vm->mutex, LK_RELEASE); 555 return r; 556 } 557 558 static int radeon_cs_handle_lockup(struct radeon_device *rdev, int r) 559 { 560 if (r == -EDEADLK) { 561 r = radeon_gpu_reset(rdev); 562 if (!r) 563 r = -EAGAIN; 564 } 565 return r; 566 } 567 568 static int radeon_cs_ib_fill(struct radeon_device *rdev, struct radeon_cs_parser *parser) 569 { 570 struct radeon_cs_chunk *ib_chunk; 571 struct radeon_vm *vm = NULL; 572 int r; 573 574 if (parser->chunk_ib_idx == -1) 575 return 0; 576 577 if (parser->cs_flags & RADEON_CS_USE_VM) { 578 struct radeon_fpriv *fpriv = parser->filp->driver_priv; 579 vm = &fpriv->vm; 580 581 if ((rdev->family >= CHIP_TAHITI) && 582 (parser->chunk_const_ib_idx != -1)) { 583 ib_chunk = &parser->chunks[parser->chunk_const_ib_idx]; 584 if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) { 585 DRM_ERROR("cs IB CONST too big: %d\n", ib_chunk->length_dw); 586 return -EINVAL; 587 } 588 r = radeon_ib_get(rdev, parser->ring, &parser->const_ib, 589 vm, ib_chunk->length_dw * 4); 590 if (r) { 591 DRM_ERROR("Failed to get const ib !\n"); 592 return r; 593 } 594 parser->const_ib.is_const_ib = true; 595 parser->const_ib.length_dw = ib_chunk->length_dw; 596 if (copy_from_user(parser->const_ib.ptr, 597 ib_chunk->user_ptr, 598 ib_chunk->length_dw * 4)) 599 return -EFAULT; 600 } 601 602 ib_chunk = &parser->chunks[parser->chunk_ib_idx]; 603 if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) { 604 DRM_ERROR("cs IB too big: %d\n", ib_chunk->length_dw); 605 return -EINVAL; 606 } 607 } 608 ib_chunk = &parser->chunks[parser->chunk_ib_idx]; 609 610 r = radeon_ib_get(rdev, parser->ring, &parser->ib, 611 vm, ib_chunk->length_dw * 4); 612 if (r) { 613 DRM_ERROR("Failed to get ib !\n"); 614 return r; 615 } 616 parser->ib.length_dw = ib_chunk->length_dw; 617 if (ib_chunk->kdata) 618 memcpy(parser->ib.ptr, ib_chunk->kdata, ib_chunk->length_dw * 4); 619 else if (copy_from_user(parser->ib.ptr, ib_chunk->user_ptr, ib_chunk->length_dw * 4)) 620 return -EFAULT; 621 return 0; 622 } 623 624 int radeon_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp) 625 { 626 struct radeon_device *rdev = dev->dev_private; 627 struct radeon_cs_parser parser; 628 int r; 629 630 lockmgr(&rdev->exclusive_lock, LK_EXCLUSIVE); 631 if (!rdev->accel_working) { 632 lockmgr(&rdev->exclusive_lock, LK_RELEASE); 633 return -EBUSY; 634 } 635 if (rdev->in_reset) { 636 lockmgr(&rdev->exclusive_lock, LK_RELEASE); 637 r = radeon_gpu_reset(rdev); 638 if (!r) 639 r = -EAGAIN; 640 return r; 641 } 642 /* initialize parser */ 643 memset(&parser, 0, sizeof(struct radeon_cs_parser)); 644 parser.filp = filp; 645 parser.rdev = rdev; 646 parser.dev = rdev->dev; 647 parser.family = rdev->family; 648 r = radeon_cs_parser_init(&parser, data); 649 if (r) { 650 DRM_ERROR("Failed to initialize parser !\n"); 651 radeon_cs_parser_fini(&parser, r, false); 652 lockmgr(&rdev->exclusive_lock, LK_RELEASE); 653 r = radeon_cs_handle_lockup(rdev, r); 654 return r; 655 } 656 657 r = radeon_cs_ib_fill(rdev, &parser); 658 if (!r) { 659 r = radeon_cs_parser_relocs(&parser); 660 if (r && r != -ERESTARTSYS) 661 DRM_ERROR("Failed to parse relocation %d!\n", r); 662 } 663 664 if (r) { 665 radeon_cs_parser_fini(&parser, r, false); 666 lockmgr(&rdev->exclusive_lock, LK_RELEASE); 667 r = radeon_cs_handle_lockup(rdev, r); 668 return r; 669 } 670 671 #if TRACE_TODO 672 trace_radeon_cs(&parser); 673 #endif 674 675 r = radeon_cs_ib_chunk(rdev, &parser); 676 if (r) { 677 goto out; 678 } 679 r = radeon_cs_ib_vm_chunk(rdev, &parser); 680 if (r) { 681 goto out; 682 } 683 out: 684 radeon_cs_parser_fini(&parser, r, true); 685 lockmgr(&rdev->exclusive_lock, LK_RELEASE); 686 r = radeon_cs_handle_lockup(rdev, r); 687 return r; 688 } 689 690 /** 691 * radeon_cs_packet_parse() - parse cp packet and point ib index to next packet 692 * @parser: parser structure holding parsing context. 693 * @pkt: where to store packet information 694 * 695 * Assume that chunk_ib_index is properly set. Will return -EINVAL 696 * if packet is bigger than remaining ib size. or if packets is unknown. 697 **/ 698 int radeon_cs_packet_parse(struct radeon_cs_parser *p, 699 struct radeon_cs_packet *pkt, 700 unsigned idx) 701 { 702 struct radeon_cs_chunk *ib_chunk = &p->chunks[p->chunk_ib_idx]; 703 struct radeon_device *rdev = p->rdev; 704 uint32_t header; 705 706 if (idx >= ib_chunk->length_dw) { 707 DRM_ERROR("Can not parse packet at %d after CS end %d !\n", 708 idx, ib_chunk->length_dw); 709 return -EINVAL; 710 } 711 header = radeon_get_ib_value(p, idx); 712 pkt->idx = idx; 713 pkt->type = RADEON_CP_PACKET_GET_TYPE(header); 714 pkt->count = RADEON_CP_PACKET_GET_COUNT(header); 715 pkt->one_reg_wr = 0; 716 switch (pkt->type) { 717 case RADEON_PACKET_TYPE0: 718 if (rdev->family < CHIP_R600) { 719 pkt->reg = R100_CP_PACKET0_GET_REG(header); 720 pkt->one_reg_wr = 721 RADEON_CP_PACKET0_GET_ONE_REG_WR(header); 722 } else 723 pkt->reg = R600_CP_PACKET0_GET_REG(header); 724 break; 725 case RADEON_PACKET_TYPE3: 726 pkt->opcode = RADEON_CP_PACKET3_GET_OPCODE(header); 727 break; 728 case RADEON_PACKET_TYPE2: 729 pkt->count = -1; 730 break; 731 default: 732 DRM_ERROR("Unknown packet type %d at %d !\n", pkt->type, idx); 733 return -EINVAL; 734 } 735 if ((pkt->count + 1 + pkt->idx) >= ib_chunk->length_dw) { 736 DRM_ERROR("Packet (%d:%d:%d) end after CS buffer (%d) !\n", 737 pkt->idx, pkt->type, pkt->count, ib_chunk->length_dw); 738 return -EINVAL; 739 } 740 return 0; 741 } 742 743 /** 744 * radeon_cs_packet_next_is_pkt3_nop() - test if the next packet is P3 NOP 745 * @p: structure holding the parser context. 746 * 747 * Check if the next packet is NOP relocation packet3. 748 **/ 749 bool radeon_cs_packet_next_is_pkt3_nop(struct radeon_cs_parser *p) 750 { 751 struct radeon_cs_packet p3reloc; 752 int r; 753 754 r = radeon_cs_packet_parse(p, &p3reloc, p->idx); 755 if (r) 756 return false; 757 if (p3reloc.type != RADEON_PACKET_TYPE3) 758 return false; 759 if (p3reloc.opcode != RADEON_PACKET3_NOP) 760 return false; 761 return true; 762 } 763 764 /** 765 * radeon_cs_dump_packet() - dump raw packet context 766 * @p: structure holding the parser context. 767 * @pkt: structure holding the packet. 768 * 769 * Used mostly for debugging and error reporting. 770 **/ 771 void radeon_cs_dump_packet(struct radeon_cs_parser *p, 772 struct radeon_cs_packet *pkt) 773 { 774 volatile uint32_t *ib; 775 unsigned i; 776 unsigned idx; 777 778 ib = p->ib.ptr; 779 idx = pkt->idx; 780 for (i = 0; i <= (pkt->count + 1); i++, idx++) 781 DRM_INFO("ib[%d]=0x%08X\n", idx, ib[idx]); 782 } 783 784 /** 785 * radeon_cs_packet_next_reloc() - parse next (should be reloc) packet 786 * @parser: parser structure holding parsing context. 787 * @data: pointer to relocation data 788 * @offset_start: starting offset 789 * @offset_mask: offset mask (to align start offset on) 790 * @reloc: reloc informations 791 * 792 * Check if next packet is relocation packet3, do bo validation and compute 793 * GPU offset using the provided start. 794 **/ 795 int radeon_cs_packet_next_reloc(struct radeon_cs_parser *p, 796 struct radeon_cs_reloc **cs_reloc, 797 int nomm) 798 { 799 struct radeon_cs_chunk *relocs_chunk; 800 struct radeon_cs_packet p3reloc; 801 unsigned idx; 802 int r; 803 804 if (p->chunk_relocs_idx == -1) { 805 DRM_ERROR("No relocation chunk !\n"); 806 return -EINVAL; 807 } 808 *cs_reloc = NULL; 809 relocs_chunk = &p->chunks[p->chunk_relocs_idx]; 810 r = radeon_cs_packet_parse(p, &p3reloc, p->idx); 811 if (r) 812 return r; 813 p->idx += p3reloc.count + 2; 814 if (p3reloc.type != RADEON_PACKET_TYPE3 || 815 p3reloc.opcode != RADEON_PACKET3_NOP) { 816 DRM_ERROR("No packet3 for relocation for packet at %d.\n", 817 p3reloc.idx); 818 radeon_cs_dump_packet(p, &p3reloc); 819 return -EINVAL; 820 } 821 idx = radeon_get_ib_value(p, p3reloc.idx + 1); 822 if (idx >= relocs_chunk->length_dw) { 823 DRM_ERROR("Relocs at %d after relocations chunk end %d !\n", 824 idx, relocs_chunk->length_dw); 825 radeon_cs_dump_packet(p, &p3reloc); 826 return -EINVAL; 827 } 828 /* FIXME: we assume reloc size is 4 dwords */ 829 if (nomm) { 830 *cs_reloc = p->relocs; 831 (*cs_reloc)->gpu_offset = 832 (u64)relocs_chunk->kdata[idx + 3] << 32; 833 (*cs_reloc)->gpu_offset |= relocs_chunk->kdata[idx + 0]; 834 } else 835 *cs_reloc = p->relocs_ptr[(idx / 4)]; 836 return 0; 837 } 838