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 <linux/sync_file.h>
  29 #include <drm/drmP.h>
  30 #include <drm/amdgpu_drm.h>
  31 #include <drm/drm_syncobj.h>
  32 #include "amdgpu.h"
  33 #include "amdgpu_trace.h"
  34 #include "amdgpu_gmc.h"
  35
  36 static int amdgpu_cs_user_fence_chunk(struct amdgpu_cs_parser *p,
  37                                       struct drm_amdgpu_cs_chunk_fence *data,
  38                                       uint32_t *offset)
  39 {
  40         struct drm_gem_object *gobj;
  41         unsigned long size;
  42         int r;
  43
  44         gobj = drm_gem_object_lookup(p->filp, data->handle);
  45         if (gobj == NULL)
  46                 return -EINVAL;
  47
  48         p->uf_entry.robj = amdgpu_bo_ref(gem_to_amdgpu_bo(gobj));
  49         p->uf_entry.priority = 0;
  50         p->uf_entry.tv.bo = &p->uf_entry.robj->tbo;
  51         p->uf_entry.tv.shared = true;
  52         p->uf_entry.user_pages = NULL;
  53
  54         drm_gem_object_put_unlocked(gobj);
  55
  56         size = amdgpu_bo_size(p->uf_entry.robj);
  57         if (size != PAGE_SIZE || (data->offset + 8) > size) {
  58                 r = -EINVAL;
  59                 goto error_unref;
  60         }
  61
  62         if (amdgpu_ttm_tt_get_usermm(p->uf_entry.robj->tbo.ttm)) {
  63                 r = -EINVAL;
  64                 goto error_unref;
  65         }
  66
  67         *offset = data->offset;
  68
  69         return 0;
  70
  71 error_unref:
  72         amdgpu_bo_unref(&p->uf_entry.robj);
  73         return r;
  74 }
  75
  76 static int amdgpu_cs_bo_handles_chunk(struct amdgpu_cs_parser *p,
  77                                       struct drm_amdgpu_bo_list_in *data)
  78 {
  79         int r;
  80         struct drm_amdgpu_bo_list_entry *info = NULL;
  81
  82         r = amdgpu_bo_create_list_entry_array(data, &info);
  83         if (r)
  84                 return r;
  85
  86         r = amdgpu_bo_list_create(p->adev, p->filp, info, data->bo_number,
  87                                   &p->bo_list);
  88         if (r)
  89                 goto error_free;
  90
  91         kvfree(info);
  92         return 0;
  93
  94 error_free:
  95         if (info)
  96                 kvfree(info);
  97
  98         return r;
  99 }
 100
 101 static int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, union drm_amdgpu_cs *cs)
 102 {
 103         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 104         struct amdgpu_vm *vm = &fpriv->vm;
 105         uint64_t *chunk_array_user;
 106         uint64_t *chunk_array;
 107         unsigned size, num_ibs = 0;
 108         uint32_t uf_offset = 0;
 109         int i;
 110         int ret;
 111
 112         if (cs->in.num_chunks == 0)
 113                 return 0;
 114
 115         chunk_array = kmalloc_array(cs->in.num_chunks, sizeof(uint64_t), GFP_KERNEL);
 116         if (!chunk_array)
 117                 return -ENOMEM;
 118
 119         p->ctx = amdgpu_ctx_get(fpriv, cs->in.ctx_id);
 120         if (!p->ctx) {
 121                 ret = -EINVAL;
 122                 goto free_chunk;
 123         }
 124
 125         /* skip guilty context job */
 126         if (atomic_read(&p->ctx->guilty) == 1) {
 127                 ret = -ECANCELED;
 128                 goto free_chunk;
 129         }
 130
 131         mutex_lock(&p->ctx->lock);
 132
 133         /* get chunks */
 134         chunk_array_user = u64_to_user_ptr(cs->in.chunks);
 135         if (copy_from_user(chunk_array, chunk_array_user,
 136                            sizeof(uint64_t)*cs->in.num_chunks)) {
 137                 ret = -EFAULT;
 138                 goto free_chunk;
 139         }
 140
 141         p->nchunks = cs->in.num_chunks;
 142         p->chunks = kmalloc_array(p->nchunks, sizeof(struct amdgpu_cs_chunk),
 143                             GFP_KERNEL);
 144         if (!p->chunks) {
 145                 ret = -ENOMEM;
 146                 goto free_chunk;
 147         }
 148
 149         for (i = 0; i < p->nchunks; i++) {
 150                 struct drm_amdgpu_cs_chunk __user **chunk_ptr = NULL;
 151                 struct drm_amdgpu_cs_chunk user_chunk;
 152                 uint32_t __user *cdata;
 153
 154                 chunk_ptr = u64_to_user_ptr(chunk_array[i]);
 155                 if (copy_from_user(&user_chunk, chunk_ptr,
 156                                        sizeof(struct drm_amdgpu_cs_chunk))) {
 157                         ret = -EFAULT;
 158                         i--;
 159                         goto free_partial_kdata;
 160                 }
 161                 p->chunks[i].chunk_id = user_chunk.chunk_id;
 162                 p->chunks[i].length_dw = user_chunk.length_dw;
 163
 164                 size = p->chunks[i].length_dw;
 165                 cdata = u64_to_user_ptr(user_chunk.chunk_data);
 166
 167                 p->chunks[i].kdata = kvmalloc_array(size, sizeof(uint32_t), GFP_KERNEL);
 168                 if (p->chunks[i].kdata == NULL) {
 169                         ret = -ENOMEM;
 170                         i--;
 171                         goto free_partial_kdata;
 172                 }
 173                 size *= sizeof(uint32_t);
 174                 if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
 175                         ret = -EFAULT;
 176                         goto free_partial_kdata;
 177                 }
 178
 179                 switch (p->chunks[i].chunk_id) {
 180                 case AMDGPU_CHUNK_ID_IB:
 181                         ++num_ibs;
 182                         break;
 183
 184                 case AMDGPU_CHUNK_ID_FENCE:
 185                         size = sizeof(struct drm_amdgpu_cs_chunk_fence);
 186                         if (p->chunks[i].length_dw * sizeof(uint32_t) < size) {
 187                                 ret = -EINVAL;
 188                                 goto free_partial_kdata;
 189                         }
 190
 191                         ret = amdgpu_cs_user_fence_chunk(p, p->chunks[i].kdata,
 192                                                          &uf_offset);
 193                         if (ret)
 194                                 goto free_partial_kdata;
 195
 196                         break;
 197
 198                 case AMDGPU_CHUNK_ID_BO_HANDLES:
 199                         size = sizeof(struct drm_amdgpu_bo_list_in);
 200                         if (p->chunks[i].length_dw * sizeof(uint32_t) < size) {
 201                                 ret = -EINVAL;
 202                                 goto free_partial_kdata;
 203                         }
 204
 205                         ret = amdgpu_cs_bo_handles_chunk(p, p->chunks[i].kdata);
 206                         if (ret)
 207                                 goto free_partial_kdata;
 208
 209                         break;
 210
 211                 case AMDGPU_CHUNK_ID_DEPENDENCIES:
 212                 case AMDGPU_CHUNK_ID_SYNCOBJ_IN:
 213                 case AMDGPU_CHUNK_ID_SYNCOBJ_OUT:
 214                         break;
 215
 216                 default:
 217                         ret = -EINVAL;
 218                         goto free_partial_kdata;
 219                 }
 220         }
 221
 222         ret = amdgpu_job_alloc(p->adev, num_ibs, &p->job, vm);
 223         if (ret)
 224                 goto free_all_kdata;
 225
 226         if (p->ctx->vram_lost_counter != p->job->vram_lost_counter) {
 227                 ret = -ECANCELED;
 228                 goto free_all_kdata;
 229         }
 230
 231         if (p->uf_entry.robj)
 232                 p->job->uf_addr = uf_offset;
 233         kfree(chunk_array);
 234
 235         /* Use this opportunity to fill in task info for the vm */
 236         amdgpu_vm_set_task_info(vm);
 237
 238         return 0;
 239
 240 free_all_kdata:
 241         i = p->nchunks - 1;
 242 free_partial_kdata:
 243         for (; i >= 0; i--)
 244                 kvfree(p->chunks[i].kdata);
 245         kfree(p->chunks);
 246         p->chunks = NULL;
 247         p->nchunks = 0;
 248 free_chunk:
 249         kfree(chunk_array);
 250
 251         return ret;
 252 }
 253
 254 /* Convert microseconds to bytes. */
 255 static u64 us_to_bytes(struct amdgpu_device *adev, s64 us)
 256 {
 257         if (us <= 0 || !adev->mm_stats.log2_max_MBps)
 258                 return 0;
 259
 260         /* Since accum_us is incremented by a million per second, just
 261          * multiply it by the number of MB/s to get the number of bytes.
 262          */
 263         return us << adev->mm_stats.log2_max_MBps;
 264 }
 265
 266 static s64 bytes_to_us(struct amdgpu_device *adev, u64 bytes)
 267 {
 268         if (!adev->mm_stats.log2_max_MBps)
 269                 return 0;
 270
 271         return bytes >> adev->mm_stats.log2_max_MBps;
 272 }
 273
 274 /* Returns how many bytes TTM can move right now. If no bytes can be moved,
 275  * it returns 0. If it returns non-zero, it's OK to move at least one buffer,
 276  * which means it can go over the threshold once. If that happens, the driver
 277  * will be in debt and no other buffer migrations can be done until that debt
 278  * is repaid.
 279  *
 280  * This approach allows moving a buffer of any size (it's important to allow
 281  * that).
 282  *
 283  * The currency is simply time in microseconds and it increases as the clock
 284  * ticks. The accumulated microseconds (us) are converted to bytes and
 285  * returned.
 286  */
 287 static void amdgpu_cs_get_threshold_for_moves(struct amdgpu_device *adev,
 288                                               u64 *max_bytes,
 289                                               u64 *max_vis_bytes)
 290 {
 291         s64 time_us, increment_us;
 292         u64 free_vram, total_vram, used_vram;
 293
 294         /* Allow a maximum of 200 accumulated ms. This is basically per-IB
 295          * throttling.
 296          *
 297          * It means that in order to get full max MBps, at least 5 IBs per
 298          * second must be submitted and not more than 200ms apart from each
 299          * other.
 300          */
 301         const s64 us_upper_bound = 200000;
 302
 303         if (!adev->mm_stats.log2_max_MBps) {
 304                 *max_bytes = 0;
 305                 *max_vis_bytes = 0;
 306                 return;
 307         }
 308
 309         total_vram = adev->gmc.real_vram_size - atomic64_read(&adev->vram_pin_size);
 310         used_vram = amdgpu_vram_mgr_usage(&adev->mman.bdev.man[TTM_PL_VRAM]);
 311         free_vram = used_vram >= total_vram ? 0 : total_vram - used_vram;
 312
 313         spin_lock(&adev->mm_stats.lock);
 314
 315         /* Increase the amount of accumulated us. */
 316         time_us = ktime_to_us(ktime_get());
 317         increment_us = time_us - adev->mm_stats.last_update_us;
 318         adev->mm_stats.last_update_us = time_us;
 319         adev->mm_stats.accum_us = min(adev->mm_stats.accum_us + increment_us,
 320                                       us_upper_bound);
 321
 322         /* This prevents the short period of low performance when the VRAM
 323          * usage is low and the driver is in debt or doesn't have enough
 324          * accumulated us to fill VRAM quickly.
 325          *
 326          * The situation can occur in these cases:
 327          * - a lot of VRAM is freed by userspace
 328          * - the presence of a big buffer causes a lot of evictions
 329          *   (solution: split buffers into smaller ones)
 330          *
 331          * If 128 MB or 1/8th of VRAM is free, start filling it now by setting
 332          * accum_us to a positive number.
 333          */
 334         if (free_vram >= 128 * 1024 * 1024 || free_vram >= total_vram / 8) {
 335                 s64 min_us;
 336
 337                 /* Be more aggresive on dGPUs. Try to fill a portion of free
 338                  * VRAM now.
 339                  */
 340                 if (!(adev->flags & AMD_IS_APU))
 341                         min_us = bytes_to_us(adev, free_vram / 4);
 342                 else
 343                         min_us = 0; /* Reset accum_us on APUs. */
 344
 345                 adev->mm_stats.accum_us = max(min_us, adev->mm_stats.accum_us);
 346         }
 347
 348         /* This is set to 0 if the driver is in debt to disallow (optional)
 349          * buffer moves.
 350          */
 351         *max_bytes = us_to_bytes(adev, adev->mm_stats.accum_us);
 352
 353         /* Do the same for visible VRAM if half of it is free */
 354         if (!amdgpu_gmc_vram_full_visible(&adev->gmc)) {
 355                 u64 total_vis_vram = adev->gmc.visible_vram_size;
 356                 u64 used_vis_vram =
 357                         amdgpu_vram_mgr_vis_usage(&adev->mman.bdev.man[TTM_PL_VRAM]);
 358
 359                 if (used_vis_vram < total_vis_vram) {
 360                         u64 free_vis_vram = total_vis_vram - used_vis_vram;
 361                         adev->mm_stats.accum_us_vis = min(adev->mm_stats.accum_us_vis +
 362                                                           increment_us, us_upper_bound);
 363
 364                         if (free_vis_vram >= total_vis_vram / 2)
 365                                 adev->mm_stats.accum_us_vis =
 366                                         max(bytes_to_us(adev, free_vis_vram / 2),
 367                                             adev->mm_stats.accum_us_vis);
 368                 }
 369
 370                 *max_vis_bytes = us_to_bytes(adev, adev->mm_stats.accum_us_vis);
 371         } else {
 372                 *max_vis_bytes = 0;
 373         }
 374
 375         spin_unlock(&adev->mm_stats.lock);
 376 }
 377
 378 /* Report how many bytes have really been moved for the last command
 379  * submission. This can result in a debt that can stop buffer migrations
 380  * temporarily.
 381  */
 382 void amdgpu_cs_report_moved_bytes(struct amdgpu_device *adev, u64 num_bytes,
 383                                   u64 num_vis_bytes)
 384 {
 385         spin_lock(&adev->mm_stats.lock);
 386         adev->mm_stats.accum_us -= bytes_to_us(adev, num_bytes);
 387         adev->mm_stats.accum_us_vis -= bytes_to_us(adev, num_vis_bytes);
 388         spin_unlock(&adev->mm_stats.lock);
 389 }
 390
 391 static int amdgpu_cs_bo_validate(struct amdgpu_cs_parser *p,
 392                                  struct amdgpu_bo *bo)
 393 {
 394         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
 395         struct ttm_operation_ctx ctx = {
 396                 .interruptible = true,
 397                 .no_wait_gpu = false,
 398                 .resv = bo->tbo.resv,
 399                 .flags = 0
 400         };
 401         uint32_t domain;
 402         int r;
 403
 404         if (bo->pin_count)
 405                 return 0;
 406
 407         /* Don't move this buffer if we have depleted our allowance
 408          * to move it. Don't move anything if the threshold is zero.
 409          */
 410         if (p->bytes_moved < p->bytes_moved_threshold) {
 411                 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
 412                     (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)) {
 413                         /* And don't move a CPU_ACCESS_REQUIRED BO to limited
 414                          * visible VRAM if we've depleted our allowance to do
 415                          * that.
 416                          */
 417                         if (p->bytes_moved_vis < p->bytes_moved_vis_threshold)
 418                                 domain = bo->preferred_domains;
 419                         else
 420                                 domain = bo->allowed_domains;
 421                 } else {
 422                         domain = bo->preferred_domains;
 423                 }
 424         } else {
 425                 domain = bo->allowed_domains;
 426         }
 427
 428 retry:
 429         amdgpu_bo_placement_from_domain(bo, domain);
 430         r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 431
 432         p->bytes_moved += ctx.bytes_moved;
 433         if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
 434             amdgpu_bo_in_cpu_visible_vram(bo))
 435                 p->bytes_moved_vis += ctx.bytes_moved;
 436
 437         if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
 438                 domain = bo->allowed_domains;
 439                 goto retry;
 440         }
 441
 442         return r;
 443 }
 444
 445 /* Last resort, try to evict something from the current working set */
 446 static bool amdgpu_cs_try_evict(struct amdgpu_cs_parser *p,
 447                                 struct amdgpu_bo *validated)
 448 {
 449         uint32_t domain = validated->allowed_domains;
 450         struct ttm_operation_ctx ctx = { true, false };
 451         int r;
 452
 453         if (!p->evictable)
 454                 return false;
 455
 456         for (;&p->evictable->tv.head != &p->validated;
 457              p->evictable = list_prev_entry(p->evictable, tv.head)) {
 458
 459                 struct amdgpu_bo_list_entry *candidate = p->evictable;
 460                 struct amdgpu_bo *bo = candidate->robj;
 461                 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
 462                 bool update_bytes_moved_vis;
 463                 uint32_t other;
 464
 465                 /* If we reached our current BO we can forget it */
 466                 if (candidate->robj == validated)
 467                         break;
 468
 469                 /* We can't move pinned BOs here */
 470                 if (bo->pin_count)
 471                         continue;
 472
 473                 other = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
 474
 475                 /* Check if this BO is in one of the domains we need space for */
 476                 if (!(other & domain))
 477                         continue;
 478
 479                 /* Check if we can move this BO somewhere else */
 480                 other = bo->allowed_domains & ~domain;
 481                 if (!other)
 482                         continue;
 483
 484                 /* Good we can try to move this BO somewhere else */
 485                 update_bytes_moved_vis =
 486                                 !amdgpu_gmc_vram_full_visible(&adev->gmc) &&
 487                                 amdgpu_bo_in_cpu_visible_vram(bo);
 488                 amdgpu_bo_placement_from_domain(bo, other);
 489                 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 490                 p->bytes_moved += ctx.bytes_moved;
 491                 if (update_bytes_moved_vis)
 492                         p->bytes_moved_vis += ctx.bytes_moved;
 493
 494                 if (unlikely(r))
 495                         break;
 496
 497                 p->evictable = list_prev_entry(p->evictable, tv.head);
 498                 list_move(&candidate->tv.head, &p->validated);
 499
 500                 return true;
 501         }
 502
 503         return false;
 504 }
 505
 506 static int amdgpu_cs_validate(void *param, struct amdgpu_bo *bo)
 507 {
 508         struct amdgpu_cs_parser *p = param;
 509         int r;
 510
 511         do {
 512                 r = amdgpu_cs_bo_validate(p, bo);
 513         } while (r == -ENOMEM && amdgpu_cs_try_evict(p, bo));
 514         if (r)
 515                 return r;
 516
 517         if (bo->shadow)
 518                 r = amdgpu_cs_bo_validate(p, bo->shadow);
 519
 520         return r;
 521 }
 522
 523 static int amdgpu_cs_list_validate(struct amdgpu_cs_parser *p,
 524                             struct list_head *validated)
 525 {
 526         struct ttm_operation_ctx ctx = { true, false };
 527         struct amdgpu_bo_list_entry *lobj;
 528         int r;
 529
 530         list_for_each_entry(lobj, validated, tv.head) {
 531                 struct amdgpu_bo *bo = lobj->robj;
 532                 bool binding_userptr = false;
 533                 struct mm_struct *usermm;
 534
 535                 usermm = amdgpu_ttm_tt_get_usermm(bo->tbo.ttm);
 536                 if (usermm && usermm != current->mm)
 537                         return -EPERM;
 538
 539                 /* Check if we have user pages and nobody bound the BO already */
 540                 if (amdgpu_ttm_tt_userptr_needs_pages(bo->tbo.ttm) &&
 541                     lobj->user_pages) {
 542                         amdgpu_bo_placement_from_domain(bo,
 543                                                         AMDGPU_GEM_DOMAIN_CPU);
 544                         r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 545                         if (r)
 546                                 return r;
 547                         amdgpu_ttm_tt_set_user_pages(bo->tbo.ttm,
 548                                                      lobj->user_pages);
 549                         binding_userptr = true;
 550                 }
 551
 552                 if (p->evictable == lobj)
 553                         p->evictable = NULL;
 554
 555                 r = amdgpu_cs_validate(p, bo);
 556                 if (r)
 557                         return r;
 558
 559                 if (binding_userptr) {
 560                         kvfree(lobj->user_pages);
 561                         lobj->user_pages = NULL;
 562                 }
 563         }
 564         return 0;
 565 }
 566
 567 static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
 568                                 union drm_amdgpu_cs *cs)
 569 {
 570         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 571         struct amdgpu_vm *vm = &fpriv->vm;
 572         struct amdgpu_bo_list_entry *e;
 573         struct list_head duplicates;
 574         struct amdgpu_bo *gds;
 575         struct amdgpu_bo *gws;
 576         struct amdgpu_bo *oa;
 577         unsigned tries = 10;
 578         int r;
 579
 580         INIT_LIST_HEAD(&p->validated);
 581
 582         /* p->bo_list could already be assigned if AMDGPU_CHUNK_ID_BO_HANDLES is present */
 583         if (cs->in.bo_list_handle) {
 584                 if (p->bo_list)
 585                         return -EINVAL;
 586
 587                 r = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle,
 588                                        &p->bo_list);
 589                 if (r)
 590                         return r;
 591         } else if (!p->bo_list) {
 592                 /* Create a empty bo_list when no handle is provided */
 593                 r = amdgpu_bo_list_create(p->adev, p->filp, NULL, 0,
 594                                           &p->bo_list);
 595                 if (r)
 596                         return r;
 597         }
 598
 599         amdgpu_bo_list_get_list(p->bo_list, &p->validated);
 600         if (p->bo_list->first_userptr != p->bo_list->num_entries)
 601                 p->mn = amdgpu_mn_get(p->adev, AMDGPU_MN_TYPE_GFX);
 602
 603         INIT_LIST_HEAD(&duplicates);
 604         amdgpu_vm_get_pd_bo(&fpriv->vm, &p->validated, &p->vm_pd);
 605
 606         if (p->uf_entry.robj && !p->uf_entry.robj->parent)
 607                 list_add(&p->uf_entry.tv.head, &p->validated);
 608
 609         while (1) {
 610                 struct list_head need_pages;
 611
 612                 r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true,
 613                                            &duplicates);
 614                 if (unlikely(r != 0)) {
 615                         if (r != -ERESTARTSYS)
 616                                 DRM_ERROR("ttm_eu_reserve_buffers failed.\n");
 617                         goto error_free_pages;
 618                 }
 619
 620                 INIT_LIST_HEAD(&need_pages);
 621                 amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
 622                         struct amdgpu_bo *bo = e->robj;
 623
 624                         if (amdgpu_ttm_tt_userptr_invalidated(bo->tbo.ttm,
 625                                  &e->user_invalidated) && e->user_pages) {
 626
 627                                 /* We acquired a page array, but somebody
 628                                  * invalidated it. Free it and try again
 629                                  */
 630                                 release_pages(e->user_pages,
 631                                               bo->tbo.ttm->num_pages);
 632                                 kvfree(e->user_pages);
 633                                 e->user_pages = NULL;
 634                         }
 635
 636                         if (amdgpu_ttm_tt_userptr_needs_pages(bo->tbo.ttm) &&
 637                             !e->user_pages) {
 638                                 list_del(&e->tv.head);
 639                                 list_add(&e->tv.head, &need_pages);
 640
 641                                 amdgpu_bo_unreserve(e->robj);
 642                         }
 643                 }
 644
 645                 if (list_empty(&need_pages))
 646                         break;
 647
 648                 /* Unreserve everything again. */
 649                 ttm_eu_backoff_reservation(&p->ticket, &p->validated);
 650
 651                 /* We tried too many times, just abort */
 652                 if (!--tries) {
 653                         r = -EDEADLK;
 654                         DRM_ERROR("deadlock in %s\n", __func__);
 655                         goto error_free_pages;
 656                 }
 657
 658                 /* Fill the page arrays for all userptrs. */
 659                 list_for_each_entry(e, &need_pages, tv.head) {
 660                         struct ttm_tt *ttm = e->robj->tbo.ttm;
 661
 662                         e->user_pages = kvmalloc_array(ttm->num_pages,
 663                                                          sizeof(struct page*),
 664                                                          GFP_KERNEL | __GFP_ZERO);
 665                         if (!e->user_pages) {
 666                                 r = -ENOMEM;
 667                                 DRM_ERROR("calloc failure in %s\n", __func__);
 668                                 goto error_free_pages;
 669                         }
 670
 671                         r = amdgpu_ttm_tt_get_user_pages(ttm, e->user_pages);
 672                         if (r) {
 673                                 DRM_ERROR("amdgpu_ttm_tt_get_user_pages failed.\n");
 674                                 kvfree(e->user_pages);
 675                                 e->user_pages = NULL;
 676                                 goto error_free_pages;
 677                         }
 678                 }
 679
 680                 /* And try again. */
 681                 list_splice(&need_pages, &p->validated);
 682         }
 683
 684         amdgpu_cs_get_threshold_for_moves(p->adev, &p->bytes_moved_threshold,
 685                                           &p->bytes_moved_vis_threshold);
 686         p->bytes_moved = 0;
 687         p->bytes_moved_vis = 0;
 688         p->evictable = list_last_entry(&p->validated,
 689                                        struct amdgpu_bo_list_entry,
 690                                        tv.head);
 691
 692         r = amdgpu_vm_validate_pt_bos(p->adev, &fpriv->vm,
 693                                       amdgpu_cs_validate, p);
 694         if (r) {
 695                 DRM_ERROR("amdgpu_vm_validate_pt_bos() failed.\n");
 696                 goto error_validate;
 697         }
 698
 699         r = amdgpu_cs_list_validate(p, &duplicates);
 700         if (r) {
 701                 DRM_ERROR("amdgpu_cs_list_validate(duplicates) failed.\n");
 702                 goto error_validate;
 703         }
 704
 705         r = amdgpu_cs_list_validate(p, &p->validated);
 706         if (r) {
 707                 DRM_ERROR("amdgpu_cs_list_validate(validated) failed.\n");
 708                 goto error_validate;
 709         }
 710
 711         amdgpu_cs_report_moved_bytes(p->adev, p->bytes_moved,
 712                                      p->bytes_moved_vis);
 713
 714         gds = p->bo_list->gds_obj;
 715         gws = p->bo_list->gws_obj;
 716         oa = p->bo_list->oa_obj;
 717
 718         amdgpu_bo_list_for_each_entry(e, p->bo_list)
 719                 e->bo_va = amdgpu_vm_bo_find(vm, e->robj);
 720
 721         if (gds) {
 722                 p->job->gds_base = amdgpu_bo_gpu_offset(gds);
 723                 p->job->gds_size = amdgpu_bo_size(gds);
 724         }
 725         if (gws) {
 726                 p->job->gws_base = amdgpu_bo_gpu_offset(gws);
 727                 p->job->gws_size = amdgpu_bo_size(gws);
 728         }
 729         if (oa) {
 730                 p->job->oa_base = amdgpu_bo_gpu_offset(oa);
 731                 p->job->oa_size = amdgpu_bo_size(oa);
 732         }
 733
 734         if (!r && p->uf_entry.robj) {
 735                 struct amdgpu_bo *uf = p->uf_entry.robj;
 736
 737                 r = amdgpu_ttm_alloc_gart(&uf->tbo);
 738                 p->job->uf_addr += amdgpu_bo_gpu_offset(uf);
 739         }
 740
 741 error_validate:
 742         if (r)
 743                 ttm_eu_backoff_reservation(&p->ticket, &p->validated);
 744
 745 error_free_pages:
 746
 747         amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
 748                 if (!e->user_pages)
 749                         continue;
 750
 751                 release_pages(e->user_pages,
 752                               e->robj->tbo.ttm->num_pages);
 753                 kvfree(e->user_pages);
 754         }
 755
 756         return r;
 757 }
 758
 759 static int amdgpu_cs_sync_rings(struct amdgpu_cs_parser *p)
 760 {
 761         struct amdgpu_bo_list_entry *e;
 762         int r;
 763
 764         list_for_each_entry(e, &p->validated, tv.head) {
 765                 struct reservation_object *resv = e->robj->tbo.resv;
 766                 r = amdgpu_sync_resv(p->adev, &p->job->sync, resv, p->filp,
 767                                      amdgpu_bo_explicit_sync(e->robj));
 768
 769                 if (r)
 770                         return r;
 771         }
 772         return 0;
 773 }
 774
 775 /**
 776  * cs_parser_fini() - clean parser states
 777  * @parser:     parser structure holding parsing context.
 778  * @error:      error number
 779  *
 780  * If error is set than unvalidate buffer, otherwise just free memory
 781  * used by parsing context.
 782  **/
 783 static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error,
 784                                   bool backoff)
 785 {
 786         unsigned i;
 787
 788         if (error && backoff)
 789                 ttm_eu_backoff_reservation(&parser->ticket,
 790                                            &parser->validated);
 791
 792         for (i = 0; i < parser->num_post_dep_syncobjs; i++)
 793                 drm_syncobj_put(parser->post_dep_syncobjs[i]);
 794         kfree(parser->post_dep_syncobjs);
 795
 796         dma_fence_put(parser->fence);
 797
 798         if (parser->ctx) {
 799                 mutex_unlock(&parser->ctx->lock);
 800                 amdgpu_ctx_put(parser->ctx);
 801         }
 802         if (parser->bo_list)
 803                 amdgpu_bo_list_put(parser->bo_list);
 804
 805         for (i = 0; i < parser->nchunks; i++)
 806                 kvfree(parser->chunks[i].kdata);
 807         kfree(parser->chunks);
 808         if (parser->job)
 809                 amdgpu_job_free(parser->job);
 810         amdgpu_bo_unref(&parser->uf_entry.robj);
 811 }
 812
 813 static int amdgpu_bo_vm_update_pte(struct amdgpu_cs_parser *p)
 814 {
 815         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 816         struct amdgpu_device *adev = p->adev;
 817         struct amdgpu_vm *vm = &fpriv->vm;
 818         struct amdgpu_bo_list_entry *e;
 819         struct amdgpu_bo_va *bo_va;
 820         struct amdgpu_bo *bo;
 821         int r;
 822
 823         r = amdgpu_vm_clear_freed(adev, vm, NULL);
 824         if (r)
 825                 return r;
 826
 827         r = amdgpu_vm_bo_update(adev, fpriv->prt_va, false);
 828         if (r)
 829                 return r;
 830
 831         r = amdgpu_sync_fence(adev, &p->job->sync,
 832                               fpriv->prt_va->last_pt_update, false);
 833         if (r)
 834                 return r;
 835
 836         if (amdgpu_sriov_vf(adev)) {
 837                 struct dma_fence *f;
 838
 839                 bo_va = fpriv->csa_va;
 840                 BUG_ON(!bo_va);
 841                 r = amdgpu_vm_bo_update(adev, bo_va, false);
 842                 if (r)
 843                         return r;
 844
 845                 f = bo_va->last_pt_update;
 846                 r = amdgpu_sync_fence(adev, &p->job->sync, f, false);
 847                 if (r)
 848                         return r;
 849         }
 850
 851         amdgpu_bo_list_for_each_entry(e, p->bo_list) {
 852                 struct dma_fence *f;
 853
 854                 /* ignore duplicates */
 855                 bo = e->robj;
 856                 if (!bo)
 857                         continue;
 858
 859                 bo_va = e->bo_va;
 860                 if (bo_va == NULL)
 861                         continue;
 862
 863                 r = amdgpu_vm_bo_update(adev, bo_va, false);
 864                 if (r)
 865                         return r;
 866
 867                 f = bo_va->last_pt_update;
 868                 r = amdgpu_sync_fence(adev, &p->job->sync, f, false);
 869                 if (r)
 870                         return r;
 871         }
 872
 873         r = amdgpu_vm_handle_moved(adev, vm);
 874         if (r)
 875                 return r;
 876
 877         r = amdgpu_vm_update_directories(adev, vm);
 878         if (r)
 879                 return r;
 880
 881         r = amdgpu_sync_fence(adev, &p->job->sync, vm->last_update, false);
 882         if (r)
 883                 return r;
 884
 885         if (amdgpu_vm_debug) {
 886                 /* Invalidate all BOs to test for userspace bugs */
 887                 amdgpu_bo_list_for_each_entry(e, p->bo_list) {
 888                         /* ignore duplicates */
 889                         if (!e->robj)
 890                                 continue;
 891
 892                         amdgpu_vm_bo_invalidate(adev, e->robj, false);
 893                 }
 894         }
 895
 896         return r;
 897 }
 898
 899 static int amdgpu_cs_ib_vm_chunk(struct amdgpu_device *adev,
 900                                  struct amdgpu_cs_parser *p)
 901 {
 902         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 903         struct amdgpu_vm *vm = &fpriv->vm;
 904         struct amdgpu_ring *ring = p->ring;
 905         int r;
 906
 907         /* Only for UVD/VCE VM emulation */
 908         if (p->ring->funcs->parse_cs || p->ring->funcs->patch_cs_in_place) {
 909                 unsigned i, j;
 910
 911                 for (i = 0, j = 0; i < p->nchunks && j < p->job->num_ibs; i++) {
 912                         struct drm_amdgpu_cs_chunk_ib *chunk_ib;
 913                         struct amdgpu_bo_va_mapping *m;
 914                         struct amdgpu_bo *aobj = NULL;
 915                         struct amdgpu_cs_chunk *chunk;
 916                         uint64_t offset, va_start;
 917                         struct amdgpu_ib *ib;
 918                         uint8_t *kptr;
 919
 920                         chunk = &p->chunks[i];
 921                         ib = &p->job->ibs[j];
 922                         chunk_ib = chunk->kdata;
 923
 924                         if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
 925                                 continue;
 926
 927                         va_start = chunk_ib->va_start & AMDGPU_VA_HOLE_MASK;
 928                         r = amdgpu_cs_find_mapping(p, va_start, &aobj, &m);
 929                         if (r) {
 930                                 DRM_ERROR("IB va_start is invalid\n");
 931                                 return r;
 932                         }
 933
 934                         if ((va_start + chunk_ib->ib_bytes) >
 935                             (m->last + 1) * AMDGPU_GPU_PAGE_SIZE) {
 936                                 DRM_ERROR("IB va_start+ib_bytes is invalid\n");
 937                                 return -EINVAL;
 938                         }
 939
 940                         /* the IB should be reserved at this point */
 941                         r = amdgpu_bo_kmap(aobj, (void **)&kptr);
 942                         if (r) {
 943                                 return r;
 944                         }
 945
 946                         offset = m->start * AMDGPU_GPU_PAGE_SIZE;
 947                         kptr += va_start - offset;
 948
 949                         if (p->ring->funcs->parse_cs) {
 950                                 memcpy(ib->ptr, kptr, chunk_ib->ib_bytes);
 951                                 amdgpu_bo_kunmap(aobj);
 952
 953                                 r = amdgpu_ring_parse_cs(ring, p, j);
 954                                 if (r)
 955                                         return r;
 956                         } else {
 957                                 ib->ptr = (uint32_t *)kptr;
 958                                 r = amdgpu_ring_patch_cs_in_place(ring, p, j);
 959                                 amdgpu_bo_kunmap(aobj);
 960                                 if (r)
 961                                         return r;
 962                         }
 963
 964                         j++;
 965                 }
 966         }
 967
 968         if (p->job->vm) {
 969                 p->job->vm_pd_addr = amdgpu_bo_gpu_offset(vm->root.base.bo);
 970
 971                 r = amdgpu_bo_vm_update_pte(p);
 972                 if (r)
 973                         return r;
 974
 975                 r = reservation_object_reserve_shared(vm->root.base.bo->tbo.resv);
 976                 if (r)
 977                         return r;
 978         }
 979
 980         return amdgpu_cs_sync_rings(p);
 981 }
 982
 983 static int amdgpu_cs_ib_fill(struct amdgpu_device *adev,
 984                              struct amdgpu_cs_parser *parser)
 985 {
 986         struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
 987         struct amdgpu_vm *vm = &fpriv->vm;
 988         int i, j;
 989         int r, ce_preempt = 0, de_preempt = 0;
 990
 991         for (i = 0, j = 0; i < parser->nchunks && j < parser->job->num_ibs; i++) {
 992                 struct amdgpu_cs_chunk *chunk;
 993                 struct amdgpu_ib *ib;
 994                 struct drm_amdgpu_cs_chunk_ib *chunk_ib;
 995                 struct amdgpu_ring *ring;
 996
 997                 chunk = &parser->chunks[i];
 998                 ib = &parser->job->ibs[j];
 999                 chunk_ib = (struct drm_amdgpu_cs_chunk_ib *)chunk->kdata;
1000
1001                 if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
1002                         continue;
1003
1004                 if (chunk_ib->ip_type == AMDGPU_HW_IP_GFX && amdgpu_sriov_vf(adev)) {
1005                         if (chunk_ib->flags & AMDGPU_IB_FLAG_PREEMPT) {
1006                                 if (chunk_ib->flags & AMDGPU_IB_FLAG_CE)
1007                                         ce_preempt++;
1008                                 else
1009                                         de_preempt++;
1010                         }
1011
1012                         /* each GFX command submit allows 0 or 1 IB preemptible for CE & DE */
1013                         if (ce_preempt > 1 || de_preempt > 1)
1014                                 return -EINVAL;
1015                 }
1016
1017                 r = amdgpu_queue_mgr_map(adev, &parser->ctx->queue_mgr, chunk_ib->ip_type,
1018                                          chunk_ib->ip_instance, chunk_ib->ring, &ring);
1019                 if (r)
1020                         return r;
1021
1022                 if (chunk_ib->flags & AMDGPU_IB_FLAG_PREAMBLE)
1023                         parser->job->preamble_status |=
1024                                 AMDGPU_PREAMBLE_IB_PRESENT;
1025
1026                 if (parser->ring && parser->ring != ring)
1027                         return -EINVAL;
1028
1029                 parser->ring = ring;
1030
1031                 r =  amdgpu_ib_get(adev, vm,
1032                                         ring->funcs->parse_cs ? chunk_ib->ib_bytes : 0,
1033                                         ib);
1034                 if (r) {
1035                         DRM_ERROR("Failed to get ib !\n");
1036                         return r;
1037                 }
1038
1039                 ib->gpu_addr = chunk_ib->va_start;
1040                 ib->length_dw = chunk_ib->ib_bytes / 4;
1041                 ib->flags = chunk_ib->flags;
1042
1043                 j++;
1044         }
1045
1046         /* UVD & VCE fw doesn't support user fences */
1047         if (parser->job->uf_addr && (
1048             parser->ring->funcs->type == AMDGPU_RING_TYPE_UVD ||
1049             parser->ring->funcs->type == AMDGPU_RING_TYPE_VCE))
1050                 return -EINVAL;
1051
1052         return amdgpu_ctx_wait_prev_fence(parser->ctx, parser->ring->idx);
1053 }
1054
1055 static int amdgpu_cs_process_fence_dep(struct amdgpu_cs_parser *p,
1056                                        struct amdgpu_cs_chunk *chunk)
1057 {
1058         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
1059         unsigned num_deps;
1060         int i, r;
1061         struct drm_amdgpu_cs_chunk_dep *deps;
1062
1063         deps = (struct drm_amdgpu_cs_chunk_dep *)chunk->kdata;
1064         num_deps = chunk->length_dw * 4 /
1065                 sizeof(struct drm_amdgpu_cs_chunk_dep);
1066
1067         for (i = 0; i < num_deps; ++i) {
1068                 struct amdgpu_ring *ring;
1069                 struct amdgpu_ctx *ctx;
1070                 struct dma_fence *fence;
1071
1072                 ctx = amdgpu_ctx_get(fpriv, deps[i].ctx_id);
1073                 if (ctx == NULL)
1074                         return -EINVAL;
1075
1076                 r = amdgpu_queue_mgr_map(p->adev, &ctx->queue_mgr,
1077                                          deps[i].ip_type,
1078                                          deps[i].ip_instance,
1079                                          deps[i].ring, &ring);
1080                 if (r) {
1081                         amdgpu_ctx_put(ctx);
1082                         return r;
1083                 }
1084
1085                 fence = amdgpu_ctx_get_fence(ctx, ring,
1086                                              deps[i].handle);
1087                 if (IS_ERR(fence)) {
1088                         r = PTR_ERR(fence);
1089                         amdgpu_ctx_put(ctx);
1090                         return r;
1091                 } else if (fence) {
1092                         r = amdgpu_sync_fence(p->adev, &p->job->sync, fence,
1093                                         true);
1094                         dma_fence_put(fence);
1095                         amdgpu_ctx_put(ctx);
1096                         if (r)
1097                                 return r;
1098                 }
1099         }
1100         return 0;
1101 }
1102
1103 static int amdgpu_syncobj_lookup_and_add_to_sync(struct amdgpu_cs_parser *p,
1104                                                  uint32_t handle)
1105 {
1106         int r;
1107         struct dma_fence *fence;
1108         r = drm_syncobj_find_fence(p->filp, handle, &fence);
1109         if (r)
1110                 return r;
1111
1112         r = amdgpu_sync_fence(p->adev, &p->job->sync, fence, true);
1113         dma_fence_put(fence);
1114
1115         return r;
1116 }
1117
1118 static int amdgpu_cs_process_syncobj_in_dep(struct amdgpu_cs_parser *p,
1119                                             struct amdgpu_cs_chunk *chunk)
1120 {
1121         unsigned num_deps;
1122         int i, r;
1123         struct drm_amdgpu_cs_chunk_sem *deps;
1124
1125         deps = (struct drm_amdgpu_cs_chunk_sem *)chunk->kdata;
1126         num_deps = chunk->length_dw * 4 /
1127                 sizeof(struct drm_amdgpu_cs_chunk_sem);
1128
1129         for (i = 0; i < num_deps; ++i) {
1130                 r = amdgpu_syncobj_lookup_and_add_to_sync(p, deps[i].handle);
1131                 if (r)
1132                         return r;
1133         }
1134         return 0;
1135 }
1136
1137 static int amdgpu_cs_process_syncobj_out_dep(struct amdgpu_cs_parser *p,
1138                                              struct amdgpu_cs_chunk *chunk)
1139 {
1140         unsigned num_deps;
1141         int i;
1142         struct drm_amdgpu_cs_chunk_sem *deps;
1143         deps = (struct drm_amdgpu_cs_chunk_sem *)chunk->kdata;
1144         num_deps = chunk->length_dw * 4 /
1145                 sizeof(struct drm_amdgpu_cs_chunk_sem);
1146
1147         p->post_dep_syncobjs = kmalloc_array(num_deps,
1148                                              sizeof(struct drm_syncobj *),
1149                                              GFP_KERNEL);
1150         p->num_post_dep_syncobjs = 0;
1151
1152         if (!p->post_dep_syncobjs)
1153                 return -ENOMEM;
1154
1155         for (i = 0; i < num_deps; ++i) {
1156                 p->post_dep_syncobjs[i] = drm_syncobj_find(p->filp, deps[i].handle);
1157                 if (!p->post_dep_syncobjs[i])
1158                         return -EINVAL;
1159                 p->num_post_dep_syncobjs++;
1160         }
1161         return 0;
1162 }
1163
1164 static int amdgpu_cs_dependencies(struct amdgpu_device *adev,
1165                                   struct amdgpu_cs_parser *p)
1166 {
1167         int i, r;
1168
1169         for (i = 0; i < p->nchunks; ++i) {
1170                 struct amdgpu_cs_chunk *chunk;
1171
1172                 chunk = &p->chunks[i];
1173
1174                 if (chunk->chunk_id == AMDGPU_CHUNK_ID_DEPENDENCIES) {
1175                         r = amdgpu_cs_process_fence_dep(p, chunk);
1176                         if (r)
1177                                 return r;
1178                 } else if (chunk->chunk_id == AMDGPU_CHUNK_ID_SYNCOBJ_IN) {
1179                         r = amdgpu_cs_process_syncobj_in_dep(p, chunk);
1180                         if (r)
1181                                 return r;
1182                 } else if (chunk->chunk_id == AMDGPU_CHUNK_ID_SYNCOBJ_OUT) {
1183                         r = amdgpu_cs_process_syncobj_out_dep(p, chunk);
1184                         if (r)
1185                                 return r;
1186                 }
1187         }
1188
1189         return 0;
1190 }
1191
1192 static void amdgpu_cs_post_dependencies(struct amdgpu_cs_parser *p)
1193 {
1194         int i;
1195
1196         for (i = 0; i < p->num_post_dep_syncobjs; ++i)
1197                 drm_syncobj_replace_fence(p->post_dep_syncobjs[i], p->fence);
1198 }
1199
1200 static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
1201                             union drm_amdgpu_cs *cs)
1202 {
1203         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
1204         struct amdgpu_ring *ring = p->ring;
1205         struct drm_sched_entity *entity = &p->ctx->rings[ring->idx].entity;
1206         enum drm_sched_priority priority;
1207         struct amdgpu_bo_list_entry *e;
1208         struct amdgpu_job *job;
1209         uint64_t seq;
1210
1211         int r;
1212
1213         job = p->job;
1214         p->job = NULL;
1215
1216         r = drm_sched_job_init(&job->base, entity, p->filp);
1217         if (r)
1218                 goto error_unlock;
1219
1220         /* No memory allocation is allowed while holding the mn lock */
1221         amdgpu_mn_lock(p->mn);
1222         amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
1223                 struct amdgpu_bo *bo = e->robj;
1224
1225                 if (amdgpu_ttm_tt_userptr_needs_pages(bo->tbo.ttm)) {
1226                         r = -ERESTARTSYS;
1227                         goto error_abort;
1228                 }
1229         }
1230
1231         job->owner = p->filp;
1232         p->fence = dma_fence_get(&job->base.s_fence->finished);
1233
1234         r = amdgpu_ctx_add_fence(p->ctx, ring, p->fence, &seq);
1235         if (r) {
1236                 dma_fence_put(p->fence);
1237                 dma_fence_put(&job->base.s_fence->finished);
1238                 amdgpu_job_free(job);
1239                 amdgpu_mn_unlock(p->mn);
1240                 return r;
1241         }
1242
1243         amdgpu_cs_post_dependencies(p);
1244
1245         if ((job->preamble_status & AMDGPU_PREAMBLE_IB_PRESENT) &&
1246             !p->ctx->preamble_presented) {
1247                 job->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT_FIRST;
1248                 p->ctx->preamble_presented = true;
1249         }
1250
1251         cs->out.handle = seq;
1252         job->uf_sequence = seq;
1253
1254         amdgpu_job_free_resources(job);
1255
1256         trace_amdgpu_cs_ioctl(job);
1257         amdgpu_vm_bo_trace_cs(&fpriv->vm, &p->ticket);
1258         priority = job->base.s_priority;
1259         drm_sched_entity_push_job(&job->base, entity);
1260
1261         ring = to_amdgpu_ring(entity->rq->sched);
1262         amdgpu_ring_priority_get(ring, priority);
1263
1264         ttm_eu_fence_buffer_objects(&p->ticket, &p->validated, p->fence);
1265         amdgpu_mn_unlock(p->mn);
1266
1267         return 0;
1268
1269 error_abort:
1270         dma_fence_put(&job->base.s_fence->finished);
1271         job->base.s_fence = NULL;
1272         amdgpu_mn_unlock(p->mn);
1273
1274 error_unlock:
1275         amdgpu_job_free(job);
1276         return r;
1277 }
1278
1279 int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
1280 {
1281         struct amdgpu_device *adev = dev->dev_private;
1282         union drm_amdgpu_cs *cs = data;
1283         struct amdgpu_cs_parser parser = {};
1284         bool reserved_buffers = false;
1285         int i, r;
1286
1287         if (!adev->accel_working)
1288                 return -EBUSY;
1289
1290         parser.adev = adev;
1291         parser.filp = filp;
1292
1293         r = amdgpu_cs_parser_init(&parser, data);
1294         if (r) {
1295                 DRM_ERROR("Failed to initialize parser !\n");
1296                 goto out;
1297         }
1298
1299         r = amdgpu_cs_ib_fill(adev, &parser);
1300         if (r)
1301                 goto out;
1302
1303         r = amdgpu_cs_parser_bos(&parser, data);
1304         if (r) {
1305                 if (r == -ENOMEM)
1306                         DRM_ERROR("Not enough memory for command submission!\n");
1307                 else if (r != -ERESTARTSYS)
1308                         DRM_ERROR("Failed to process the buffer list %d!\n", r);
1309                 goto out;
1310         }
1311
1312         reserved_buffers = true;
1313
1314         r = amdgpu_cs_dependencies(adev, &parser);
1315         if (r) {
1316                 DRM_ERROR("Failed in the dependencies handling %d!\n", r);
1317                 goto out;
1318         }
1319
1320         for (i = 0; i < parser.job->num_ibs; i++)
1321                 trace_amdgpu_cs(&parser, i);
1322
1323         r = amdgpu_cs_ib_vm_chunk(adev, &parser);
1324         if (r)
1325                 goto out;
1326
1327         r = amdgpu_cs_submit(&parser, cs);
1328
1329 out:
1330         amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
1331         return r;
1332 }
1333
1334 /**
1335  * amdgpu_cs_wait_ioctl - wait for a command submission to finish
1336  *
1337  * @dev: drm device
1338  * @data: data from userspace
1339  * @filp: file private
1340  *
1341  * Wait for the command submission identified by handle to finish.
1342  */
1343 int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data,
1344                          struct drm_file *filp)
1345 {
1346         union drm_amdgpu_wait_cs *wait = data;
1347         struct amdgpu_device *adev = dev->dev_private;
1348         unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
1349         struct amdgpu_ring *ring = NULL;
1350         struct amdgpu_ctx *ctx;
1351         struct dma_fence *fence;
1352         long r;
1353
1354         ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
1355         if (ctx == NULL)
1356                 return -EINVAL;
1357
1358         r = amdgpu_queue_mgr_map(adev, &ctx->queue_mgr,
1359                                  wait->in.ip_type, wait->in.ip_instance,
1360                                  wait->in.ring, &ring);
1361         if (r) {
1362                 amdgpu_ctx_put(ctx);
1363                 return r;
1364         }
1365
1366         fence = amdgpu_ctx_get_fence(ctx, ring, wait->in.handle);
1367         if (IS_ERR(fence))
1368                 r = PTR_ERR(fence);
1369         else if (fence) {
1370                 r = dma_fence_wait_timeout(fence, true, timeout);
1371                 if (r > 0 && fence->error)
1372                         r = fence->error;
1373                 dma_fence_put(fence);
1374         } else
1375                 r = 1;
1376
1377         amdgpu_ctx_put(ctx);
1378         if (r < 0)
1379                 return r;
1380
1381         memset(wait, 0, sizeof(*wait));
1382         wait->out.status = (r == 0);
1383
1384         return 0;
1385 }
1386
1387 /**
1388  * amdgpu_cs_get_fence - helper to get fence from drm_amdgpu_fence
1389  *
1390  * @adev: amdgpu device
1391  * @filp: file private
1392  * @user: drm_amdgpu_fence copied from user space
1393  */
1394 static struct dma_fence *amdgpu_cs_get_fence(struct amdgpu_device *adev,
1395                                              struct drm_file *filp,
1396                                              struct drm_amdgpu_fence *user)
1397 {
1398         struct amdgpu_ring *ring;
1399         struct amdgpu_ctx *ctx;
1400         struct dma_fence *fence;
1401         int r;
1402
1403         ctx = amdgpu_ctx_get(filp->driver_priv, user->ctx_id);
1404         if (ctx == NULL)
1405                 return ERR_PTR(-EINVAL);
1406
1407         r = amdgpu_queue_mgr_map(adev, &ctx->queue_mgr, user->ip_type,
1408                                  user->ip_instance, user->ring, &ring);
1409         if (r) {
1410                 amdgpu_ctx_put(ctx);
1411                 return ERR_PTR(r);
1412         }
1413
1414         fence = amdgpu_ctx_get_fence(ctx, ring, user->seq_no);
1415         amdgpu_ctx_put(ctx);
1416
1417         return fence;
1418 }
1419
1420 int amdgpu_cs_fence_to_handle_ioctl(struct drm_device *dev, void *data,
1421                                     struct drm_file *filp)
1422 {
1423         struct amdgpu_device *adev = dev->dev_private;
1424         union drm_amdgpu_fence_to_handle *info = data;
1425         struct dma_fence *fence;
1426         struct drm_syncobj *syncobj;
1427         struct sync_file *sync_file;
1428         int fd, r;
1429
1430         fence = amdgpu_cs_get_fence(adev, filp, &info->in.fence);
1431         if (IS_ERR(fence))
1432                 return PTR_ERR(fence);
1433
1434         switch (info->in.what) {
1435         case AMDGPU_FENCE_TO_HANDLE_GET_SYNCOBJ:
1436                 r = drm_syncobj_create(&syncobj, 0, fence);
1437                 dma_fence_put(fence);
1438                 if (r)
1439                         return r;
1440                 r = drm_syncobj_get_handle(filp, syncobj, &info->out.handle);
1441                 drm_syncobj_put(syncobj);
1442                 return r;
1443
1444         case AMDGPU_FENCE_TO_HANDLE_GET_SYNCOBJ_FD:
1445                 r = drm_syncobj_create(&syncobj, 0, fence);
1446                 dma_fence_put(fence);
1447                 if (r)
1448                         return r;
1449                 r = drm_syncobj_get_fd(syncobj, (int*)&info->out.handle);
1450                 drm_syncobj_put(syncobj);
1451                 return r;
1452
1453         case AMDGPU_FENCE_TO_HANDLE_GET_SYNC_FILE_FD:
1454                 fd = get_unused_fd_flags(O_CLOEXEC);
1455                 if (fd < 0) {
1456                         dma_fence_put(fence);
1457                         return fd;
1458                 }
1459
1460                 sync_file = sync_file_create(fence);
1461                 dma_fence_put(fence);
1462                 if (!sync_file) {
1463                         put_unused_fd(fd);
1464                         return -ENOMEM;
1465                 }
1466
1467                 fd_install(fd, sync_file->file);
1468                 info->out.handle = fd;
1469                 return 0;
1470
1471         default:
1472                 return -EINVAL;
1473         }
1474 }
1475
1476 /**
1477  * amdgpu_cs_wait_all_fence - wait on all fences to signal
1478  *
1479  * @adev: amdgpu device
1480  * @filp: file private
1481  * @wait: wait parameters
1482  * @fences: array of drm_amdgpu_fence
1483  */
1484 static int amdgpu_cs_wait_all_fences(struct amdgpu_device *adev,
1485                                      struct drm_file *filp,
1486                                      union drm_amdgpu_wait_fences *wait,
1487                                      struct drm_amdgpu_fence *fences)
1488 {
1489         uint32_t fence_count = wait->in.fence_count;
1490         unsigned int i;
1491         long r = 1;
1492
1493         for (i = 0; i < fence_count; i++) {
1494                 struct dma_fence *fence;
1495                 unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1496
1497                 fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1498                 if (IS_ERR(fence))
1499                         return PTR_ERR(fence);
1500                 else if (!fence)
1501                         continue;
1502
1503                 r = dma_fence_wait_timeout(fence, true, timeout);
1504                 dma_fence_put(fence);
1505                 if (r < 0)
1506                         return r;
1507
1508                 if (r == 0)
1509                         break;
1510
1511                 if (fence->error)
1512                         return fence->error;
1513         }
1514
1515         memset(wait, 0, sizeof(*wait));
1516         wait->out.status = (r > 0);
1517
1518         return 0;
1519 }
1520
1521 /**
1522  * amdgpu_cs_wait_any_fence - wait on any fence to signal
1523  *
1524  * @adev: amdgpu device
1525  * @filp: file private
1526  * @wait: wait parameters
1527  * @fences: array of drm_amdgpu_fence
1528  */
1529 static int amdgpu_cs_wait_any_fence(struct amdgpu_device *adev,
1530                                     struct drm_file *filp,
1531                                     union drm_amdgpu_wait_fences *wait,
1532                                     struct drm_amdgpu_fence *fences)
1533 {
1534         unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1535         uint32_t fence_count = wait->in.fence_count;
1536         uint32_t first = ~0;
1537         struct dma_fence **array;
1538         unsigned int i;
1539         long r;
1540
1541         /* Prepare the fence array */
1542         array = kcalloc(fence_count, sizeof(struct dma_fence *), GFP_KERNEL);
1543
1544         if (array == NULL)
1545                 return -ENOMEM;
1546
1547         for (i = 0; i < fence_count; i++) {
1548                 struct dma_fence *fence;
1549
1550                 fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1551                 if (IS_ERR(fence)) {
1552                         r = PTR_ERR(fence);
1553                         goto err_free_fence_array;
1554                 } else if (fence) {
1555                         array[i] = fence;
1556                 } else { /* NULL, the fence has been already signaled */
1557                         r = 1;
1558                         first = i;
1559                         goto out;
1560                 }
1561         }
1562
1563         r = dma_fence_wait_any_timeout(array, fence_count, true, timeout,
1564                                        &first);
1565         if (r < 0)
1566                 goto err_free_fence_array;
1567
1568 out:
1569         memset(wait, 0, sizeof(*wait));
1570         wait->out.status = (r > 0);
1571         wait->out.first_signaled = first;
1572
1573         if (first < fence_count && array[first])
1574                 r = array[first]->error;
1575         else
1576                 r = 0;
1577
1578 err_free_fence_array:
1579         for (i = 0; i < fence_count; i++)
1580                 dma_fence_put(array[i]);
1581         kfree(array);
1582
1583         return r;
1584 }
1585
1586 /**
1587  * amdgpu_cs_wait_fences_ioctl - wait for multiple command submissions to finish
1588  *
1589  * @dev: drm device
1590  * @data: data from userspace
1591  * @filp: file private
1592  */
1593 int amdgpu_cs_wait_fences_ioctl(struct drm_device *dev, void *data,
1594                                 struct drm_file *filp)
1595 {
1596         struct amdgpu_device *adev = dev->dev_private;
1597         union drm_amdgpu_wait_fences *wait = data;
1598         uint32_t fence_count = wait->in.fence_count;
1599         struct drm_amdgpu_fence *fences_user;
1600         struct drm_amdgpu_fence *fences;
1601         int r;
1602
1603         /* Get the fences from userspace */
1604         fences = kmalloc_array(fence_count, sizeof(struct drm_amdgpu_fence),
1605                         GFP_KERNEL);
1606         if (fences == NULL)
1607                 return -ENOMEM;
1608
1609         fences_user = u64_to_user_ptr(wait->in.fences);
1610         if (copy_from_user(fences, fences_user,
1611                 sizeof(struct drm_amdgpu_fence) * fence_count)) {
1612                 r = -EFAULT;
1613                 goto err_free_fences;
1614         }
1615
1616         if (wait->in.wait_all)
1617                 r = amdgpu_cs_wait_all_fences(adev, filp, wait, fences);
1618         else
1619                 r = amdgpu_cs_wait_any_fence(adev, filp, wait, fences);
1620
1621 err_free_fences:
1622         kfree(fences);
1623
1624         return r;
1625 }
1626
1627 /**
1628  * amdgpu_cs_find_bo_va - find bo_va for VM address
1629  *
1630  * @parser: command submission parser context
1631  * @addr: VM address
1632  * @bo: resulting BO of the mapping found
1633  *
1634  * Search the buffer objects in the command submission context for a certain
1635  * virtual memory address. Returns allocation structure when found, NULL
1636  * otherwise.
1637  */
1638 int amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
1639                            uint64_t addr, struct amdgpu_bo **bo,
1640                            struct amdgpu_bo_va_mapping **map)
1641 {
1642         struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
1643         struct ttm_operation_ctx ctx = { false, false };
1644         struct amdgpu_vm *vm = &fpriv->vm;
1645         struct amdgpu_bo_va_mapping *mapping;
1646         int r;
1647
1648         addr /= AMDGPU_GPU_PAGE_SIZE;
1649
1650         mapping = amdgpu_vm_bo_lookup_mapping(vm, addr);
1651         if (!mapping || !mapping->bo_va || !mapping->bo_va->base.bo)
1652                 return -EINVAL;
1653
1654         *bo = mapping->bo_va->base.bo;
1655         *map = mapping;
1656
1657         /* Double check that the BO is reserved by this CS */
1658         if (READ_ONCE((*bo)->tbo.resv->lock.ctx) != &parser->ticket)
1659                 return -EINVAL;
1660
1661         if (!((*bo)->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)) {
1662                 (*bo)->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
1663                 amdgpu_bo_placement_from_domain(*bo, (*bo)->allowed_domains);
1664                 r = ttm_bo_validate(&(*bo)->tbo, &(*bo)->placement, &ctx);
1665                 if (r)
1666                         return r;
1667         }
1668
1669         return amdgpu_ttm_alloc_gart(&(*bo)->tbo);
1670 }