drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c

   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 <[email protected]>
  26  */
  27 #include <linux/pagemap.h>
  28 #include <drm/drmP.h>
  29 #include <drm/amdgpu_drm.h>
  30 #include "amdgpu.h"
  31 #include "amdgpu_trace.h"
  32
  33 int amdgpu_cs_get_ring(struct amdgpu_device *adev, u32 ip_type,
  34                        u32 ip_instance, u32 ring,
  35                        struct amdgpu_ring **out_ring)
  36 {
  37         /* Right now all IPs have only one instance - multiple rings. */
  38         if (ip_instance != 0) {
  39                 DRM_ERROR("invalid ip instance: %d\n", ip_instance);
  40                 return -EINVAL;
  41         }
  42
  43         switch (ip_type) {
  44         default:
  45                 DRM_ERROR("unknown ip type: %d\n", ip_type);
  46                 return -EINVAL;
  47         case AMDGPU_HW_IP_GFX:
  48                 if (ring < adev->gfx.num_gfx_rings) {
  49                         *out_ring = &adev->gfx.gfx_ring[ring];
  50                 } else {
  51                         DRM_ERROR("only %d gfx rings are supported now\n",
  52                                   adev->gfx.num_gfx_rings);
  53                         return -EINVAL;
  54                 }
  55                 break;
  56         case AMDGPU_HW_IP_COMPUTE:
  57                 if (ring < adev->gfx.num_compute_rings) {
  58                         *out_ring = &adev->gfx.compute_ring[ring];
  59                 } else {
  60                         DRM_ERROR("only %d compute rings are supported now\n",
  61                                   adev->gfx.num_compute_rings);
  62                         return -EINVAL;
  63                 }
  64                 break;
  65         case AMDGPU_HW_IP_DMA:
  66                 if (ring < adev->sdma.num_instances) {
  67                         *out_ring = &adev->sdma.instance[ring].ring;
  68                 } else {
  69                         DRM_ERROR("only %d SDMA rings are supported\n",
  70                                   adev->sdma.num_instances);
  71                         return -EINVAL;
  72                 }
  73                 break;
  74         case AMDGPU_HW_IP_UVD:
  75                 *out_ring = &adev->uvd.ring;
  76                 break;
  77         case AMDGPU_HW_IP_VCE:
  78                 if (ring < adev->vce.num_rings){
  79                         *out_ring = &adev->vce.ring[ring];
  80                 } else {
  81                         DRM_ERROR("only %d VCE rings are supported\n", adev->vce.num_rings);
  82                         return -EINVAL;
  83                 }
  84                 break;
  85         }
  86
  87         if (!(*out_ring && (*out_ring)->adev)) {
  88                 DRM_ERROR("Ring %d is not initialized on IP %d\n",
  89                           ring, ip_type);
  90                 return -EINVAL;
  91         }
  92
  93         return 0;
  94 }
  95
  96 static int amdgpu_cs_user_fence_chunk(struct amdgpu_cs_parser *p,
  97                                       struct drm_amdgpu_cs_chunk_fence *data,
  98                                       uint32_t *offset)
  99 {
 100         struct drm_gem_object *gobj;
 101         unsigned long size;
 102
 103         gobj = drm_gem_object_lookup(p->filp, data->handle);
 104         if (gobj == NULL)
 105                 return -EINVAL;
 106
 107         p->uf_entry.robj = amdgpu_bo_ref(gem_to_amdgpu_bo(gobj));
 108         p->uf_entry.priority = 0;
 109         p->uf_entry.tv.bo = &p->uf_entry.robj->tbo;
 110         p->uf_entry.tv.shared = true;
 111         p->uf_entry.user_pages = NULL;
 112
 113         size = amdgpu_bo_size(p->uf_entry.robj);
 114         if (size != PAGE_SIZE || (data->offset + 8) > size)
 115                 return -EINVAL;
 116
 117         *offset = data->offset;
 118
 119         drm_gem_object_unreference_unlocked(gobj);
 120
 121         if (amdgpu_ttm_tt_get_usermm(p->uf_entry.robj->tbo.ttm)) {
 122                 amdgpu_bo_unref(&p->uf_entry.robj);
 123                 return -EINVAL;
 124         }
 125
 126         return 0;
 127 }
 128
 129 int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data)
 130 {
 131         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 132         struct amdgpu_vm *vm = &fpriv->vm;
 133         union drm_amdgpu_cs *cs = data;
 134         uint64_t *chunk_array_user;
 135         uint64_t *chunk_array;
 136         unsigned size, num_ibs = 0;
 137         uint32_t uf_offset = 0;
 138         int i;
 139         int ret;
 140
 141         if (cs->in.num_chunks == 0)
 142                 return 0;
 143
 144         chunk_array = kmalloc_array(cs->in.num_chunks, sizeof(uint64_t), GFP_KERNEL);
 145         if (!chunk_array)
 146                 return -ENOMEM;
 147
 148         p->ctx = amdgpu_ctx_get(fpriv, cs->in.ctx_id);
 149         if (!p->ctx) {
 150                 ret = -EINVAL;
 151                 goto free_chunk;
 152         }
 153
 154         /* get chunks */
 155         chunk_array_user = (uint64_t __user *)(unsigned long)(cs->in.chunks);
 156         if (copy_from_user(chunk_array, chunk_array_user,
 157                            sizeof(uint64_t)*cs->in.num_chunks)) {
 158                 ret = -EFAULT;
 159                 goto put_ctx;
 160         }
 161
 162         p->nchunks = cs->in.num_chunks;
 163         p->chunks = kmalloc_array(p->nchunks, sizeof(struct amdgpu_cs_chunk),
 164                             GFP_KERNEL);
 165         if (!p->chunks) {
 166                 ret = -ENOMEM;
 167                 goto put_ctx;
 168         }
 169
 170         for (i = 0; i < p->nchunks; i++) {
 171                 struct drm_amdgpu_cs_chunk __user **chunk_ptr = NULL;
 172                 struct drm_amdgpu_cs_chunk user_chunk;
 173                 uint32_t __user *cdata;
 174
 175                 chunk_ptr = (void __user *)(unsigned long)chunk_array[i];
 176                 if (copy_from_user(&user_chunk, chunk_ptr,
 177                                        sizeof(struct drm_amdgpu_cs_chunk))) {
 178                         ret = -EFAULT;
 179                         i--;
 180                         goto free_partial_kdata;
 181                 }
 182                 p->chunks[i].chunk_id = user_chunk.chunk_id;
 183                 p->chunks[i].length_dw = user_chunk.length_dw;
 184
 185                 size = p->chunks[i].length_dw;
 186                 cdata = (void __user *)(unsigned long)user_chunk.chunk_data;
 187
 188                 p->chunks[i].kdata = drm_malloc_ab(size, sizeof(uint32_t));
 189                 if (p->chunks[i].kdata == NULL) {
 190                         ret = -ENOMEM;
 191                         i--;
 192                         goto free_partial_kdata;
 193                 }
 194                 size *= sizeof(uint32_t);
 195                 if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
 196                         ret = -EFAULT;
 197                         goto free_partial_kdata;
 198                 }
 199
 200                 switch (p->chunks[i].chunk_id) {
 201                 case AMDGPU_CHUNK_ID_IB:
 202                         ++num_ibs;
 203                         break;
 204
 205                 case AMDGPU_CHUNK_ID_FENCE:
 206                         size = sizeof(struct drm_amdgpu_cs_chunk_fence);
 207                         if (p->chunks[i].length_dw * sizeof(uint32_t) < size) {
 208                                 ret = -EINVAL;
 209                                 goto free_partial_kdata;
 210                         }
 211
 212                         ret = amdgpu_cs_user_fence_chunk(p, p->chunks[i].kdata,
 213                                                          &uf_offset);
 214                         if (ret)
 215                                 goto free_partial_kdata;
 216
 217                         break;
 218
 219                 case AMDGPU_CHUNK_ID_DEPENDENCIES:
 220                         break;
 221
 222                 default:
 223                         ret = -EINVAL;
 224                         goto free_partial_kdata;
 225                 }
 226         }
 227
 228         ret = amdgpu_job_alloc(p->adev, num_ibs, &p->job, vm);
 229         if (ret)
 230                 goto free_all_kdata;
 231
 232         if (p->uf_entry.robj)
 233                 p->job->uf_addr = uf_offset;
 234         kfree(chunk_array);
 235         return 0;
 236
 237 free_all_kdata:
 238         i = p->nchunks - 1;
 239 free_partial_kdata:
 240         for (; i >= 0; i--)
 241                 drm_free_large(p->chunks[i].kdata);
 242         kfree(p->chunks);
 243         p->chunks = NULL;
 244         p->nchunks = 0;
 245 put_ctx:
 246         amdgpu_ctx_put(p->ctx);
 247 free_chunk:
 248         kfree(chunk_array);
 249
 250         return ret;
 251 }
 252
 253 /* Convert microseconds to bytes. */
 254 static u64 us_to_bytes(struct amdgpu_device *adev, s64 us)
 255 {
 256         if (us <= 0 || !adev->mm_stats.log2_max_MBps)
 257                 return 0;
 258
 259         /* Since accum_us is incremented by a million per second, just
 260          * multiply it by the number of MB/s to get the number of bytes.
 261          */
 262         return us << adev->mm_stats.log2_max_MBps;
 263 }
 264
 265 static s64 bytes_to_us(struct amdgpu_device *adev, u64 bytes)
 266 {
 267         if (!adev->mm_stats.log2_max_MBps)
 268                 return 0;
 269
 270         return bytes >> adev->mm_stats.log2_max_MBps;
 271 }
 272
 273 /* Returns how many bytes TTM can move right now. If no bytes can be moved,
 274  * it returns 0. If it returns non-zero, it's OK to move at least one buffer,
 275  * which means it can go over the threshold once. If that happens, the driver
 276  * will be in debt and no other buffer migrations can be done until that debt
 277  * is repaid.
 278  *
 279  * This approach allows moving a buffer of any size (it's important to allow
 280  * that).
 281  *
 282  * The currency is simply time in microseconds and it increases as the clock
 283  * ticks. The accumulated microseconds (us) are converted to bytes and
 284  * returned.
 285  */
 286 static u64 amdgpu_cs_get_threshold_for_moves(struct amdgpu_device *adev)
 287 {
 288         s64 time_us, increment_us;
 289         u64 max_bytes;
 290         u64 free_vram, total_vram, used_vram;
 291
 292         /* Allow a maximum of 200 accumulated ms. This is basically per-IB
 293          * throttling.
 294          *
 295          * It means that in order to get full max MBps, at least 5 IBs per
 296          * second must be submitted and not more than 200ms apart from each
 297          * other.
 298          */
 299         const s64 us_upper_bound = 200000;
 300
 301         if (!adev->mm_stats.log2_max_MBps)
 302                 return 0;
 303
 304         total_vram = adev->mc.real_vram_size - adev->vram_pin_size;
 305         used_vram = atomic64_read(&adev->vram_usage);
 306         free_vram = used_vram >= total_vram ? 0 : total_vram - used_vram;
 307
 308         spin_lock(&adev->mm_stats.lock);
 309
 310         /* Increase the amount of accumulated us. */
 311         time_us = ktime_to_us(ktime_get());
 312         increment_us = time_us - adev->mm_stats.last_update_us;
 313         adev->mm_stats.last_update_us = time_us;
 314         adev->mm_stats.accum_us = min(adev->mm_stats.accum_us + increment_us,
 315                                       us_upper_bound);
 316
 317         /* This prevents the short period of low performance when the VRAM
 318          * usage is low and the driver is in debt or doesn't have enough
 319          * accumulated us to fill VRAM quickly.
 320          *
 321          * The situation can occur in these cases:
 322          * - a lot of VRAM is freed by userspace
 323          * - the presence of a big buffer causes a lot of evictions
 324          *   (solution: split buffers into smaller ones)
 325          *
 326          * If 128 MB or 1/8th of VRAM is free, start filling it now by setting
 327          * accum_us to a positive number.
 328          */
 329         if (free_vram >= 128 * 1024 * 1024 || free_vram >= total_vram / 8) {
 330                 s64 min_us;
 331
 332                 /* Be more aggresive on dGPUs. Try to fill a portion of free
 333                  * VRAM now.
 334                  */
 335                 if (!(adev->flags & AMD_IS_APU))
 336                         min_us = bytes_to_us(adev, free_vram / 4);
 337                 else
 338                         min_us = 0; /* Reset accum_us on APUs. */
 339
 340                 adev->mm_stats.accum_us = max(min_us, adev->mm_stats.accum_us);
 341         }
 342
 343         /* This returns 0 if the driver is in debt to disallow (optional)
 344          * buffer moves.
 345          */
 346         max_bytes = us_to_bytes(adev, adev->mm_stats.accum_us);
 347
 348         spin_unlock(&adev->mm_stats.lock);
 349         return max_bytes;
 350 }
 351
 352 /* Report how many bytes have really been moved for the last command
 353  * submission. This can result in a debt that can stop buffer migrations
 354  * temporarily.
 355  */
 356 void amdgpu_cs_report_moved_bytes(struct amdgpu_device *adev, u64 num_bytes)
 357 {
 358         spin_lock(&adev->mm_stats.lock);
 359         adev->mm_stats.accum_us -= bytes_to_us(adev, num_bytes);
 360         spin_unlock(&adev->mm_stats.lock);
 361 }
 362
 363 static int amdgpu_cs_bo_validate(struct amdgpu_cs_parser *p,
 364                                  struct amdgpu_bo *bo)
 365 {
 366         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
 367         u64 initial_bytes_moved;
 368         uint32_t domain;
 369         int r;
 370
 371         if (bo->pin_count)
 372                 return 0;
 373
 374         /* Don't move this buffer if we have depleted our allowance
 375          * to move it. Don't move anything if the threshold is zero.
 376          */
 377         if (p->bytes_moved < p->bytes_moved_threshold)
 378                 domain = bo->prefered_domains;
 379         else
 380                 domain = bo->allowed_domains;
 381
 382 retry:
 383         amdgpu_ttm_placement_from_domain(bo, domain);
 384         initial_bytes_moved = atomic64_read(&adev->num_bytes_moved);
 385         r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
 386         p->bytes_moved += atomic64_read(&adev->num_bytes_moved) -
 387                 initial_bytes_moved;
 388
 389         if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
 390                 domain = bo->allowed_domains;
 391                 goto retry;
 392         }
 393
 394         return r;
 395 }
 396
 397 /* Last resort, try to evict something from the current working set */
 398 static bool amdgpu_cs_try_evict(struct amdgpu_cs_parser *p,
 399                                 struct amdgpu_bo *validated)
 400 {
 401         uint32_t domain = validated->allowed_domains;
 402         int r;
 403
 404         if (!p->evictable)
 405                 return false;
 406
 407         for (;&p->evictable->tv.head != &p->validated;
 408              p->evictable = list_prev_entry(p->evictable, tv.head)) {
 409
 410                 struct amdgpu_bo_list_entry *candidate = p->evictable;
 411                 struct amdgpu_bo *bo = candidate->robj;
 412                 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
 413                 u64 initial_bytes_moved;
 414                 uint32_t other;
 415
 416                 /* If we reached our current BO we can forget it */
 417                 if (candidate->robj == validated)
 418                         break;
 419
 420                 other = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
 421
 422                 /* Check if this BO is in one of the domains we need space for */
 423                 if (!(other & domain))
 424                         continue;
 425
 426                 /* Check if we can move this BO somewhere else */
 427                 other = bo->allowed_domains & ~domain;
 428                 if (!other)
 429                         continue;
 430
 431                 /* Good we can try to move this BO somewhere else */
 432                 amdgpu_ttm_placement_from_domain(bo, other);
 433                 initial_bytes_moved = atomic64_read(&adev->num_bytes_moved);
 434                 r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
 435                 p->bytes_moved += atomic64_read(&adev->num_bytes_moved) -
 436                         initial_bytes_moved;
 437
 438                 if (unlikely(r))
 439                         break;
 440
 441                 p->evictable = list_prev_entry(p->evictable, tv.head);
 442                 list_move(&candidate->tv.head, &p->validated);
 443
 444                 return true;
 445         }
 446
 447         return false;
 448 }
 449
 450 static int amdgpu_cs_validate(void *param, struct amdgpu_bo *bo)
 451 {
 452         struct amdgpu_cs_parser *p = param;
 453         int r;
 454
 455         do {
 456                 r = amdgpu_cs_bo_validate(p, bo);
 457         } while (r == -ENOMEM && amdgpu_cs_try_evict(p, bo));
 458         if (r)
 459                 return r;
 460
 461         if (bo->shadow)
 462                 r = amdgpu_cs_bo_validate(p, bo->shadow);
 463
 464         return r;
 465 }
 466
 467 static int amdgpu_cs_list_validate(struct amdgpu_cs_parser *p,
 468                             struct list_head *validated)
 469 {
 470         struct amdgpu_bo_list_entry *lobj;
 471         int r;
 472
 473         list_for_each_entry(lobj, validated, tv.head) {
 474                 struct amdgpu_bo *bo = lobj->robj;
 475                 bool binding_userptr = false;
 476                 struct mm_struct *usermm;
 477
 478                 usermm = amdgpu_ttm_tt_get_usermm(bo->tbo.ttm);
 479                 if (usermm && usermm != current->mm)
 480                         return -EPERM;
 481
 482                 /* Check if we have user pages and nobody bound the BO already */
 483                 if (lobj->user_pages && bo->tbo.ttm->state != tt_bound) {
 484                         size_t size = sizeof(struct page *);
 485
 486                         size *= bo->tbo.ttm->num_pages;
 487                         memcpy(bo->tbo.ttm->pages, lobj->user_pages, size);
 488                         binding_userptr = true;
 489                 }
 490
 491                 if (p->evictable == lobj)
 492                         p->evictable = NULL;
 493
 494                 r = amdgpu_cs_validate(p, bo);
 495                 if (r)
 496                         return r;
 497
 498                 if (binding_userptr) {
 499                         drm_free_large(lobj->user_pages);
 500                         lobj->user_pages = NULL;
 501                 }
 502         }
 503         return 0;
 504 }
 505
 506 static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
 507                                 union drm_amdgpu_cs *cs)
 508 {
 509         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 510         struct amdgpu_bo_list_entry *e;
 511         struct list_head duplicates;
 512         bool need_mmap_lock = false;
 513         unsigned i, tries = 10;
 514         int r;
 515
 516         INIT_LIST_HEAD(&p->validated);
 517
 518         p->bo_list = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle);
 519         if (p->bo_list) {
 520                 need_mmap_lock = p->bo_list->first_userptr !=
 521                         p->bo_list->num_entries;
 522                 amdgpu_bo_list_get_list(p->bo_list, &p->validated);
 523         }
 524
 525         INIT_LIST_HEAD(&duplicates);
 526         amdgpu_vm_get_pd_bo(&fpriv->vm, &p->validated, &p->vm_pd);
 527
 528         if (p->uf_entry.robj)
 529                 list_add(&p->uf_entry.tv.head, &p->validated);
 530
 531         if (need_mmap_lock)
 532                 down_read(&current->mm->mmap_sem);
 533
 534         while (1) {
 535                 struct list_head need_pages;
 536                 unsigned i;
 537
 538                 r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true,
 539                                            &duplicates);
 540                 if (unlikely(r != 0)) {
 541                         if (r != -ERESTARTSYS)
 542                                 DRM_ERROR("ttm_eu_reserve_buffers failed.\n");
 543                         goto error_free_pages;
 544                 }
 545
 546                 /* Without a BO list we don't have userptr BOs */
 547                 if (!p->bo_list)
 548                         break;
 549
 550                 INIT_LIST_HEAD(&need_pages);
 551                 for (i = p->bo_list->first_userptr;
 552                      i < p->bo_list->num_entries; ++i) {
 553
 554                         e = &p->bo_list->array[i];
 555
 556                         if (amdgpu_ttm_tt_userptr_invalidated(e->robj->tbo.ttm,
 557                                  &e->user_invalidated) && e->user_pages) {
 558
 559                                 /* We acquired a page array, but somebody
 560                                  * invalidated it. Free it an try again
 561                                  */
 562                                 release_pages(e->user_pages,
 563                                               e->robj->tbo.ttm->num_pages,
 564                                               false);
 565                                 drm_free_large(e->user_pages);
 566                                 e->user_pages = NULL;
 567                         }
 568
 569                         if (e->robj->tbo.ttm->state != tt_bound &&
 570                             !e->user_pages) {
 571                                 list_del(&e->tv.head);
 572                                 list_add(&e->tv.head, &need_pages);
 573
 574                                 amdgpu_bo_unreserve(e->robj);
 575                         }
 576                 }
 577
 578                 if (list_empty(&need_pages))
 579                         break;
 580
 581                 /* Unreserve everything again. */
 582                 ttm_eu_backoff_reservation(&p->ticket, &p->validated);
 583
 584                 /* We tried too many times, just abort */
 585                 if (!--tries) {
 586                         r = -EDEADLK;
 587                         DRM_ERROR("deadlock in %s\n", __func__);
 588                         goto error_free_pages;
 589                 }
 590
 591                 /* Fill the page arrays for all useptrs. */
 592                 list_for_each_entry(e, &need_pages, tv.head) {
 593                         struct ttm_tt *ttm = e->robj->tbo.ttm;
 594
 595                         e->user_pages = drm_calloc_large(ttm->num_pages,
 596                                                          sizeof(struct page*));
 597                         if (!e->user_pages) {
 598                                 r = -ENOMEM;
 599                                 DRM_ERROR("calloc failure in %s\n", __func__);
 600                                 goto error_free_pages;
 601                         }
 602
 603                         r = amdgpu_ttm_tt_get_user_pages(ttm, e->user_pages);
 604                         if (r) {
 605                                 DRM_ERROR("amdgpu_ttm_tt_get_user_pages failed.\n");
 606                                 drm_free_large(e->user_pages);
 607                                 e->user_pages = NULL;
 608                                 goto error_free_pages;
 609                         }
 610                 }
 611
 612                 /* And try again. */
 613                 list_splice(&need_pages, &p->validated);
 614         }
 615
 616         p->bytes_moved_threshold = amdgpu_cs_get_threshold_for_moves(p->adev);
 617         p->bytes_moved = 0;
 618         p->evictable = list_last_entry(&p->validated,
 619                                        struct amdgpu_bo_list_entry,
 620                                        tv.head);
 621
 622         r = amdgpu_vm_validate_pt_bos(p->adev, &fpriv->vm,
 623                                       amdgpu_cs_validate, p);
 624         if (r) {
 625                 DRM_ERROR("amdgpu_vm_validate_pt_bos() failed.\n");
 626                 goto error_validate;
 627         }
 628
 629         r = amdgpu_cs_list_validate(p, &duplicates);
 630         if (r) {
 631                 DRM_ERROR("amdgpu_cs_list_validate(duplicates) failed.\n");
 632                 goto error_validate;
 633         }
 634
 635         r = amdgpu_cs_list_validate(p, &p->validated);
 636         if (r) {
 637                 DRM_ERROR("amdgpu_cs_list_validate(validated) failed.\n");
 638                 goto error_validate;
 639         }
 640
 641         amdgpu_cs_report_moved_bytes(p->adev, p->bytes_moved);
 642
 643         fpriv->vm.last_eviction_counter =
 644                 atomic64_read(&p->adev->num_evictions);
 645
 646         if (p->bo_list) {
 647                 struct amdgpu_bo *gds = p->bo_list->gds_obj;
 648                 struct amdgpu_bo *gws = p->bo_list->gws_obj;
 649                 struct amdgpu_bo *oa = p->bo_list->oa_obj;
 650                 struct amdgpu_vm *vm = &fpriv->vm;
 651                 unsigned i;
 652
 653                 for (i = 0; i < p->bo_list->num_entries; i++) {
 654                         struct amdgpu_bo *bo = p->bo_list->array[i].robj;
 655
 656                         p->bo_list->array[i].bo_va = amdgpu_vm_bo_find(vm, bo);
 657                 }
 658
 659                 if (gds) {
 660                         p->job->gds_base = amdgpu_bo_gpu_offset(gds);
 661                         p->job->gds_size = amdgpu_bo_size(gds);
 662                 }
 663                 if (gws) {
 664                         p->job->gws_base = amdgpu_bo_gpu_offset(gws);
 665                         p->job->gws_size = amdgpu_bo_size(gws);
 666                 }
 667                 if (oa) {
 668                         p->job->oa_base = amdgpu_bo_gpu_offset(oa);
 669                         p->job->oa_size = amdgpu_bo_size(oa);
 670                 }
 671         }
 672
 673         if (!r && p->uf_entry.robj) {
 674                 struct amdgpu_bo *uf = p->uf_entry.robj;
 675
 676                 r = amdgpu_ttm_bind(&uf->tbo, &uf->tbo.mem);
 677                 p->job->uf_addr += amdgpu_bo_gpu_offset(uf);
 678         }
 679
 680 error_validate:
 681         if (r) {
 682                 amdgpu_vm_move_pt_bos_in_lru(p->adev, &fpriv->vm);
 683                 ttm_eu_backoff_reservation(&p->ticket, &p->validated);
 684         }
 685
 686 error_free_pages:
 687
 688         if (need_mmap_lock)
 689                 up_read(&current->mm->mmap_sem);
 690
 691         if (p->bo_list) {
 692                 for (i = p->bo_list->first_userptr;
 693                      i < p->bo_list->num_entries; ++i) {
 694                         e = &p->bo_list->array[i];
 695
 696                         if (!e->user_pages)
 697                                 continue;
 698
 699                         release_pages(e->user_pages,
 700                                       e->robj->tbo.ttm->num_pages,
 701                                       false);
 702                         drm_free_large(e->user_pages);
 703                 }
 704         }
 705
 706         return r;
 707 }
 708
 709 static int amdgpu_cs_sync_rings(struct amdgpu_cs_parser *p)
 710 {
 711         struct amdgpu_bo_list_entry *e;
 712         int r;
 713
 714         list_for_each_entry(e, &p->validated, tv.head) {
 715                 struct reservation_object *resv = e->robj->tbo.resv;
 716                 r = amdgpu_sync_resv(p->adev, &p->job->sync, resv, p->filp);
 717
 718                 if (r)
 719                         return r;
 720         }
 721         return 0;
 722 }
 723
 724 /**
 725  * cs_parser_fini() - clean parser states
 726  * @parser:     parser structure holding parsing context.
 727  * @error:      error number
 728  *
 729  * If error is set than unvalidate buffer, otherwise just free memory
 730  * used by parsing context.
 731  **/
 732 static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error, bool backoff)
 733 {
 734         struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
 735         unsigned i;
 736
 737         if (!error) {
 738                 amdgpu_vm_move_pt_bos_in_lru(parser->adev, &fpriv->vm);
 739
 740                 ttm_eu_fence_buffer_objects(&parser->ticket,
 741                                             &parser->validated,
 742                                             parser->fence);
 743         } else if (backoff) {
 744                 ttm_eu_backoff_reservation(&parser->ticket,
 745                                            &parser->validated);
 746         }
 747         dma_fence_put(parser->fence);
 748
 749         if (parser->ctx)
 750                 amdgpu_ctx_put(parser->ctx);
 751         if (parser->bo_list)
 752                 amdgpu_bo_list_put(parser->bo_list);
 753
 754         for (i = 0; i < parser->nchunks; i++)
 755                 drm_free_large(parser->chunks[i].kdata);
 756         kfree(parser->chunks);
 757         if (parser->job)
 758                 amdgpu_job_free(parser->job);
 759         amdgpu_bo_unref(&parser->uf_entry.robj);
 760 }
 761
 762 static int amdgpu_bo_vm_update_pte(struct amdgpu_cs_parser *p,
 763                                    struct amdgpu_vm *vm)
 764 {
 765         struct amdgpu_device *adev = p->adev;
 766         struct amdgpu_bo_va *bo_va;
 767         struct amdgpu_bo *bo;
 768         int i, r;
 769
 770         r = amdgpu_vm_update_page_directory(adev, vm);
 771         if (r)
 772                 return r;
 773
 774         r = amdgpu_sync_fence(adev, &p->job->sync, vm->page_directory_fence);
 775         if (r)
 776                 return r;
 777
 778         r = amdgpu_vm_clear_freed(adev, vm);
 779         if (r)
 780                 return r;
 781
 782         if (amdgpu_sriov_vf(adev)) {
 783                 struct dma_fence *f;
 784                 bo_va = vm->csa_bo_va;
 785                 BUG_ON(!bo_va);
 786                 r = amdgpu_vm_bo_update(adev, bo_va, false);
 787                 if (r)
 788                         return r;
 789
 790                 f = bo_va->last_pt_update;
 791                 r = amdgpu_sync_fence(adev, &p->job->sync, f);
 792                 if (r)
 793                         return r;
 794         }
 795
 796         if (p->bo_list) {
 797                 for (i = 0; i < p->bo_list->num_entries; i++) {
 798                         struct dma_fence *f;
 799
 800                         /* ignore duplicates */
 801                         bo = p->bo_list->array[i].robj;
 802                         if (!bo)
 803                                 continue;
 804
 805                         bo_va = p->bo_list->array[i].bo_va;
 806                         if (bo_va == NULL)
 807                                 continue;
 808
 809                         r = amdgpu_vm_bo_update(adev, bo_va, false);
 810                         if (r)
 811                                 return r;
 812
 813                         f = bo_va->last_pt_update;
 814                         r = amdgpu_sync_fence(adev, &p->job->sync, f);
 815                         if (r)
 816                                 return r;
 817                 }
 818
 819         }
 820
 821         r = amdgpu_vm_clear_invalids(adev, vm, &p->job->sync);
 822
 823         if (amdgpu_vm_debug && p->bo_list) {
 824                 /* Invalidate all BOs to test for userspace bugs */
 825                 for (i = 0; i < p->bo_list->num_entries; i++) {
 826                         /* ignore duplicates */
 827                         bo = p->bo_list->array[i].robj;
 828                         if (!bo)
 829                                 continue;
 830
 831                         amdgpu_vm_bo_invalidate(adev, bo);
 832                 }
 833         }
 834
 835         return r;
 836 }
 837
 838 static int amdgpu_cs_ib_vm_chunk(struct amdgpu_device *adev,
 839                                  struct amdgpu_cs_parser *p)
 840 {
 841         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 842         struct amdgpu_vm *vm = &fpriv->vm;
 843         struct amdgpu_ring *ring = p->job->ring;
 844         int i, r;
 845
 846         /* Only for UVD/VCE VM emulation */
 847         if (ring->funcs->parse_cs) {
 848                 for (i = 0; i < p->job->num_ibs; i++) {
 849                         r = amdgpu_ring_parse_cs(ring, p, i);
 850                         if (r)
 851                                 return r;
 852                 }
 853         }
 854
 855         if (p->job->vm) {
 856                 p->job->vm_pd_addr = amdgpu_bo_gpu_offset(vm->page_directory);
 857
 858                 r = amdgpu_bo_vm_update_pte(p, vm);
 859                 if (r)
 860                         return r;
 861         }
 862
 863         return amdgpu_cs_sync_rings(p);
 864 }
 865
 866 static int amdgpu_cs_ib_fill(struct amdgpu_device *adev,
 867                              struct amdgpu_cs_parser *parser)
 868 {
 869         struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
 870         struct amdgpu_vm *vm = &fpriv->vm;
 871         int i, j;
 872         int r;
 873
 874         for (i = 0, j = 0; i < parser->nchunks && j < parser->job->num_ibs; i++) {
 875                 struct amdgpu_cs_chunk *chunk;
 876                 struct amdgpu_ib *ib;
 877                 struct drm_amdgpu_cs_chunk_ib *chunk_ib;
 878                 struct amdgpu_ring *ring;
 879
 880                 chunk = &parser->chunks[i];
 881                 ib = &parser->job->ibs[j];
 882                 chunk_ib = (struct drm_amdgpu_cs_chunk_ib *)chunk->kdata;
 883
 884                 if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
 885                         continue;
 886
 887                 r = amdgpu_cs_get_ring(adev, chunk_ib->ip_type,
 888                                        chunk_ib->ip_instance, chunk_ib->ring,
 889                                        &ring);
 890                 if (r)
 891                         return r;
 892
 893                 if (ib->flags & AMDGPU_IB_FLAG_PREAMBLE) {
 894                         parser->job->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT;
 895                         if (!parser->ctx->preamble_presented) {
 896                                 parser->job->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT_FIRST;
 897                                 parser->ctx->preamble_presented = true;
 898                         }
 899                 }
 900
 901                 if (parser->job->ring && parser->job->ring != ring)
 902                         return -EINVAL;
 903
 904                 parser->job->ring = ring;
 905
 906                 if (ring->funcs->parse_cs) {
 907                         struct amdgpu_bo_va_mapping *m;
 908                         struct amdgpu_bo *aobj = NULL;
 909                         uint64_t offset;
 910                         uint8_t *kptr;
 911
 912                         m = amdgpu_cs_find_mapping(parser, chunk_ib->va_start,
 913                                                    &aobj);
 914                         if (!aobj) {
 915                                 DRM_ERROR("IB va_start is invalid\n");
 916                                 return -EINVAL;
 917                         }
 918
 919                         if ((chunk_ib->va_start + chunk_ib->ib_bytes) >
 920                             (m->it.last + 1) * AMDGPU_GPU_PAGE_SIZE) {
 921                                 DRM_ERROR("IB va_start+ib_bytes is invalid\n");
 922                                 return -EINVAL;
 923                         }
 924
 925                         /* the IB should be reserved at this point */
 926                         r = amdgpu_bo_kmap(aobj, (void **)&kptr);
 927                         if (r) {
 928                                 return r;
 929                         }
 930
 931                         offset = ((uint64_t)m->it.start) * AMDGPU_GPU_PAGE_SIZE;
 932                         kptr += chunk_ib->va_start - offset;
 933
 934                         r =  amdgpu_ib_get(adev, vm, chunk_ib->ib_bytes, ib);
 935                         if (r) {
 936                                 DRM_ERROR("Failed to get ib !\n");
 937                                 return r;
 938                         }
 939
 940                         memcpy(ib->ptr, kptr, chunk_ib->ib_bytes);
 941                         amdgpu_bo_kunmap(aobj);
 942                 } else {
 943                         r =  amdgpu_ib_get(adev, vm, 0, ib);
 944                         if (r) {
 945                                 DRM_ERROR("Failed to get ib !\n");
 946                                 return r;
 947                         }
 948
 949                 }
 950
 951                 ib->gpu_addr = chunk_ib->va_start;
 952                 ib->length_dw = chunk_ib->ib_bytes / 4;
 953                 ib->flags = chunk_ib->flags;
 954                 j++;
 955         }
 956
 957         /* UVD & VCE fw doesn't support user fences */
 958         if (parser->job->uf_addr && (
 959             parser->job->ring->funcs->type == AMDGPU_RING_TYPE_UVD ||
 960             parser->job->ring->funcs->type == AMDGPU_RING_TYPE_VCE))
 961                 return -EINVAL;
 962
 963         return 0;
 964 }
 965
 966 static int amdgpu_cs_dependencies(struct amdgpu_device *adev,
 967                                   struct amdgpu_cs_parser *p)
 968 {
 969         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 970         int i, j, r;
 971
 972         for (i = 0; i < p->nchunks; ++i) {
 973                 struct drm_amdgpu_cs_chunk_dep *deps;
 974                 struct amdgpu_cs_chunk *chunk;
 975                 unsigned num_deps;
 976
 977                 chunk = &p->chunks[i];
 978
 979                 if (chunk->chunk_id != AMDGPU_CHUNK_ID_DEPENDENCIES)
 980                         continue;
 981
 982                 deps = (struct drm_amdgpu_cs_chunk_dep *)chunk->kdata;
 983                 num_deps = chunk->length_dw * 4 /
 984                         sizeof(struct drm_amdgpu_cs_chunk_dep);
 985
 986                 for (j = 0; j < num_deps; ++j) {
 987                         struct amdgpu_ring *ring;
 988                         struct amdgpu_ctx *ctx;
 989                         struct dma_fence *fence;
 990
 991                         r = amdgpu_cs_get_ring(adev, deps[j].ip_type,
 992                                                deps[j].ip_instance,
 993                                                deps[j].ring, &ring);
 994                         if (r)
 995                                 return r;
 996
 997                         ctx = amdgpu_ctx_get(fpriv, deps[j].ctx_id);
 998                         if (ctx == NULL)
 999                                 return -EINVAL;
1000
1001                         fence = amdgpu_ctx_get_fence(ctx, ring,
1002                                                      deps[j].handle);
1003                         if (IS_ERR(fence)) {
1004                                 r = PTR_ERR(fence);
1005                                 amdgpu_ctx_put(ctx);
1006                                 return r;
1007
1008                         } else if (fence) {
1009                                 r = amdgpu_sync_fence(adev, &p->job->sync,
1010                                                       fence);
1011                                 dma_fence_put(fence);
1012                                 amdgpu_ctx_put(ctx);
1013                                 if (r)
1014                                         return r;
1015                         }
1016                 }
1017         }
1018
1019         return 0;
1020 }
1021
1022 static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
1023                             union drm_amdgpu_cs *cs)
1024 {
1025         struct amdgpu_ring *ring = p->job->ring;
1026         struct amd_sched_entity *entity = &p->ctx->rings[ring->idx].entity;
1027         struct amdgpu_job *job;
1028         int r;
1029
1030         job = p->job;
1031         p->job = NULL;
1032
1033         r = amd_sched_job_init(&job->base, &ring->sched, entity, p->filp);
1034         if (r) {
1035                 amdgpu_job_free(job);
1036                 return r;
1037         }
1038
1039         job->owner = p->filp;
1040         job->fence_ctx = entity->fence_context;
1041         p->fence = dma_fence_get(&job->base.s_fence->finished);
1042         cs->out.handle = amdgpu_ctx_add_fence(p->ctx, ring, p->fence);
1043         job->uf_sequence = cs->out.handle;
1044         amdgpu_job_free_resources(job);
1045
1046         trace_amdgpu_cs_ioctl(job);
1047         amd_sched_entity_push_job(&job->base);
1048
1049         return 0;
1050 }
1051
1052 int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
1053 {
1054         struct amdgpu_device *adev = dev->dev_private;
1055         union drm_amdgpu_cs *cs = data;
1056         struct amdgpu_cs_parser parser = {};
1057         bool reserved_buffers = false;
1058         int i, r;
1059
1060         if (!adev->accel_working)
1061                 return -EBUSY;
1062
1063         parser.adev = adev;
1064         parser.filp = filp;
1065
1066         r = amdgpu_cs_parser_init(&parser, data);
1067         if (r) {
1068                 DRM_ERROR("Failed to initialize parser !\n");
1069                 goto out;
1070         }
1071
1072         r = amdgpu_cs_parser_bos(&parser, data);
1073         if (r) {
1074                 if (r == -ENOMEM)
1075                         DRM_ERROR("Not enough memory for command submission!\n");
1076                 else if (r != -ERESTARTSYS)
1077                         DRM_ERROR("Failed to process the buffer list %d!\n", r);
1078                 goto out;
1079         }
1080
1081         reserved_buffers = true;
1082         r = amdgpu_cs_ib_fill(adev, &parser);
1083         if (r)
1084                 goto out;
1085
1086         r = amdgpu_cs_dependencies(adev, &parser);
1087         if (r) {
1088                 DRM_ERROR("Failed in the dependencies handling %d!\n", r);
1089                 goto out;
1090         }
1091
1092         for (i = 0; i < parser.job->num_ibs; i++)
1093                 trace_amdgpu_cs(&parser, i);
1094
1095         r = amdgpu_cs_ib_vm_chunk(adev, &parser);
1096         if (r)
1097                 goto out;
1098
1099         r = amdgpu_cs_submit(&parser, cs);
1100
1101 out:
1102         amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
1103         return r;
1104 }
1105
1106 /**
1107  * amdgpu_cs_wait_ioctl - wait for a command submission to finish
1108  *
1109  * @dev: drm device
1110  * @data: data from userspace
1111  * @filp: file private
1112  *
1113  * Wait for the command submission identified by handle to finish.
1114  */
1115 int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data,
1116                          struct drm_file *filp)
1117 {
1118         union drm_amdgpu_wait_cs *wait = data;
1119         struct amdgpu_device *adev = dev->dev_private;
1120         unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
1121         struct amdgpu_ring *ring = NULL;
1122         struct amdgpu_ctx *ctx;
1123         struct dma_fence *fence;
1124         long r;
1125
1126         r = amdgpu_cs_get_ring(adev, wait->in.ip_type, wait->in.ip_instance,
1127                                wait->in.ring, &ring);
1128         if (r)
1129                 return r;
1130
1131         ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
1132         if (ctx == NULL)
1133                 return -EINVAL;
1134
1135         fence = amdgpu_ctx_get_fence(ctx, ring, wait->in.handle);
1136         if (IS_ERR(fence))
1137                 r = PTR_ERR(fence);
1138         else if (fence) {
1139                 r = dma_fence_wait_timeout(fence, true, timeout);
1140                 dma_fence_put(fence);
1141         } else
1142                 r = 1;
1143
1144         amdgpu_ctx_put(ctx);
1145         if (r < 0)
1146                 return r;
1147
1148         memset(wait, 0, sizeof(*wait));
1149         wait->out.status = (r == 0);
1150
1151         return 0;
1152 }
1153
1154 /**
1155  * amdgpu_cs_get_fence - helper to get fence from drm_amdgpu_fence
1156  *
1157  * @adev: amdgpu device
1158  * @filp: file private
1159  * @user: drm_amdgpu_fence copied from user space
1160  */
1161 static struct dma_fence *amdgpu_cs_get_fence(struct amdgpu_device *adev,
1162                                              struct drm_file *filp,
1163                                              struct drm_amdgpu_fence *user)
1164 {
1165         struct amdgpu_ring *ring;
1166         struct amdgpu_ctx *ctx;
1167         struct dma_fence *fence;
1168         int r;
1169
1170         r = amdgpu_cs_get_ring(adev, user->ip_type, user->ip_instance,
1171                                user->ring, &ring);
1172         if (r)
1173                 return ERR_PTR(r);
1174
1175         ctx = amdgpu_ctx_get(filp->driver_priv, user->ctx_id);
1176         if (ctx == NULL)
1177                 return ERR_PTR(-EINVAL);
1178
1179         fence = amdgpu_ctx_get_fence(ctx, ring, user->seq_no);
1180         amdgpu_ctx_put(ctx);
1181
1182         return fence;
1183 }
1184
1185 /**
1186  * amdgpu_cs_wait_all_fence - wait on all fences to signal
1187  *
1188  * @adev: amdgpu device
1189  * @filp: file private
1190  * @wait: wait parameters
1191  * @fences: array of drm_amdgpu_fence
1192  */
1193 static int amdgpu_cs_wait_all_fences(struct amdgpu_device *adev,
1194                                      struct drm_file *filp,
1195                                      union drm_amdgpu_wait_fences *wait,
1196                                      struct drm_amdgpu_fence *fences)
1197 {
1198         uint32_t fence_count = wait->in.fence_count;
1199         unsigned int i;
1200         long r = 1;
1201
1202         for (i = 0; i < fence_count; i++) {
1203                 struct dma_fence *fence;
1204                 unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1205
1206                 fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1207                 if (IS_ERR(fence))
1208                         return PTR_ERR(fence);
1209                 else if (!fence)
1210                         continue;
1211
1212                 r = dma_fence_wait_timeout(fence, true, timeout);
1213                 if (r < 0)
1214                         return r;
1215
1216                 if (r == 0)
1217                         break;
1218         }
1219
1220         memset(wait, 0, sizeof(*wait));
1221         wait->out.status = (r > 0);
1222
1223         return 0;
1224 }
1225
1226 /**
1227  * amdgpu_cs_wait_any_fence - wait on any fence to signal
1228  *
1229  * @adev: amdgpu device
1230  * @filp: file private
1231  * @wait: wait parameters
1232  * @fences: array of drm_amdgpu_fence
1233  */
1234 static int amdgpu_cs_wait_any_fence(struct amdgpu_device *adev,
1235                                     struct drm_file *filp,
1236                                     union drm_amdgpu_wait_fences *wait,
1237                                     struct drm_amdgpu_fence *fences)
1238 {
1239         unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1240         uint32_t fence_count = wait->in.fence_count;
1241         uint32_t first = ~0;
1242         struct dma_fence **array;
1243         unsigned int i;
1244         long r;
1245
1246         /* Prepare the fence array */
1247         array = kcalloc(fence_count, sizeof(struct dma_fence *), GFP_KERNEL);
1248
1249         if (array == NULL)
1250                 return -ENOMEM;
1251
1252         for (i = 0; i < fence_count; i++) {
1253                 struct dma_fence *fence;
1254
1255                 fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1256                 if (IS_ERR(fence)) {
1257                         r = PTR_ERR(fence);
1258                         goto err_free_fence_array;
1259                 } else if (fence) {
1260                         array[i] = fence;
1261                 } else { /* NULL, the fence has been already signaled */
1262                         r = 1;
1263                         goto out;
1264                 }
1265         }
1266
1267         r = dma_fence_wait_any_timeout(array, fence_count, true, timeout,
1268                                        &first);
1269         if (r < 0)
1270                 goto err_free_fence_array;
1271
1272 out:
1273         memset(wait, 0, sizeof(*wait));
1274         wait->out.status = (r > 0);
1275         wait->out.first_signaled = first;
1276         /* set return value 0 to indicate success */
1277         r = 0;
1278
1279 err_free_fence_array:
1280         for (i = 0; i < fence_count; i++)
1281                 dma_fence_put(array[i]);
1282         kfree(array);
1283
1284         return r;
1285 }
1286
1287 /**
1288  * amdgpu_cs_wait_fences_ioctl - wait for multiple command submissions to finish
1289  *
1290  * @dev: drm device
1291  * @data: data from userspace
1292  * @filp: file private
1293  */
1294 int amdgpu_cs_wait_fences_ioctl(struct drm_device *dev, void *data,
1295                                 struct drm_file *filp)
1296 {
1297         struct amdgpu_device *adev = dev->dev_private;
1298         union drm_amdgpu_wait_fences *wait = data;
1299         uint32_t fence_count = wait->in.fence_count;
1300         struct drm_amdgpu_fence *fences_user;
1301         struct drm_amdgpu_fence *fences;
1302         int r;
1303
1304         /* Get the fences from userspace */
1305         fences = kmalloc_array(fence_count, sizeof(struct drm_amdgpu_fence),
1306                         GFP_KERNEL);
1307         if (fences == NULL)
1308                 return -ENOMEM;
1309
1310         fences_user = (void __user *)(unsigned long)(wait->in.fences);
1311         if (copy_from_user(fences, fences_user,
1312                 sizeof(struct drm_amdgpu_fence) * fence_count)) {
1313                 r = -EFAULT;
1314                 goto err_free_fences;
1315         }
1316
1317         if (wait->in.wait_all)
1318                 r = amdgpu_cs_wait_all_fences(adev, filp, wait, fences);
1319         else
1320                 r = amdgpu_cs_wait_any_fence(adev, filp, wait, fences);
1321
1322 err_free_fences:
1323         kfree(fences);
1324
1325         return r;
1326 }
1327
1328 /**
1329  * amdgpu_cs_find_bo_va - find bo_va for VM address
1330  *
1331  * @parser: command submission parser context
1332  * @addr: VM address
1333  * @bo: resulting BO of the mapping found
1334  *
1335  * Search the buffer objects in the command submission context for a certain
1336  * virtual memory address. Returns allocation structure when found, NULL
1337  * otherwise.
1338  */
1339 struct amdgpu_bo_va_mapping *
1340 amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
1341                        uint64_t addr, struct amdgpu_bo **bo)
1342 {
1343         struct amdgpu_bo_va_mapping *mapping;
1344         unsigned i;
1345
1346         if (!parser->bo_list)
1347                 return NULL;
1348
1349         addr /= AMDGPU_GPU_PAGE_SIZE;
1350
1351         for (i = 0; i < parser->bo_list->num_entries; i++) {
1352                 struct amdgpu_bo_list_entry *lobj;
1353
1354                 lobj = &parser->bo_list->array[i];
1355                 if (!lobj->bo_va)
1356                         continue;
1357
1358                 list_for_each_entry(mapping, &lobj->bo_va->valids, list) {
1359                         if (mapping->it.start > addr ||
1360                             addr > mapping->it.last)
1361                                 continue;
1362
1363                         *bo = lobj->bo_va->bo;
1364                         return mapping;
1365                 }
1366
1367                 list_for_each_entry(mapping, &lobj->bo_va->invalids, list) {
1368                         if (mapping->it.start > addr ||
1369                             addr > mapping->it.last)
1370                                 continue;
1371
1372                         *bo = lobj->bo_va->bo;
1373                         return mapping;
1374                 }
1375         }
1376
1377         return NULL;
1378 }
1379
1380 /**
1381  * amdgpu_cs_sysvm_access_required - make BOs accessible by the system VM
1382  *
1383  * @parser: command submission parser context
1384  *
1385  * Helper for UVD/VCE VM emulation, make sure BOs are accessible by the system VM.
1386  */
1387 int amdgpu_cs_sysvm_access_required(struct amdgpu_cs_parser *parser)
1388 {
1389         unsigned i;
1390         int r;
1391
1392         if (!parser->bo_list)
1393                 return 0;
1394
1395         for (i = 0; i < parser->bo_list->num_entries; i++) {
1396                 struct amdgpu_bo *bo = parser->bo_list->array[i].robj;
1397
1398                 r = amdgpu_ttm_bind(&bo->tbo, &bo->tbo.mem);
1399                 if (unlikely(r))
1400                         return r;
1401
1402                 if (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)
1403                         continue;
1404
1405                 bo->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
1406                 amdgpu_ttm_placement_from_domain(bo, bo->allowed_domains);
1407                 r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
1408                 if (unlikely(r))
1409                         return r;
1410         }
1411
1412         return 0;
1413 }