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
  28 #include <linux/file.h>
  29 #include <linux/pagemap.h>
  30 #include <linux/sync_file.h>
  31 #include <linux/dma-buf.h>
  32
  33 #include <drm/amdgpu_drm.h>
  34 #include <drm/drm_syncobj.h>
  35 #include "amdgpu_cs.h"
  36 #include "amdgpu.h"
  37 #include "amdgpu_trace.h"
  38 #include "amdgpu_gmc.h"
  39 #include "amdgpu_gem.h"
  40 #include "amdgpu_ras.h"
  41
  42 static int amdgpu_cs_user_fence_chunk(struct amdgpu_cs_parser *p,
  43                                       struct drm_amdgpu_cs_chunk_fence *data,
  44                                       uint32_t *offset)
  45 {
  46         struct drm_gem_object *gobj;
  47         struct amdgpu_bo *bo;
  48         unsigned long size;
  49         int r;
  50
  51         gobj = drm_gem_object_lookup(p->filp, data->handle);
  52         if (gobj == NULL)
  53                 return -EINVAL;
  54
  55         bo = amdgpu_bo_ref(gem_to_amdgpu_bo(gobj));
  56         p->uf_entry.priority = 0;
  57         p->uf_entry.tv.bo = &bo->tbo;
  58         /* One for TTM and two for the CS job */
  59         p->uf_entry.tv.num_shared = 3;
  60
  61         drm_gem_object_put(gobj);
  62
  63         size = amdgpu_bo_size(bo);
  64         if (size != PAGE_SIZE || (data->offset + 8) > size) {
  65                 r = -EINVAL;
  66                 goto error_unref;
  67         }
  68
  69         if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
  70                 r = -EINVAL;
  71                 goto error_unref;
  72         }
  73
  74         *offset = data->offset;
  75
  76         return 0;
  77
  78 error_unref:
  79         amdgpu_bo_unref(&bo);
  80         return r;
  81 }
  82
  83 static int amdgpu_cs_bo_handles_chunk(struct amdgpu_cs_parser *p,
  84                                       struct drm_amdgpu_bo_list_in *data)
  85 {
  86         int r;
  87         struct drm_amdgpu_bo_list_entry *info = NULL;
  88
  89         r = amdgpu_bo_create_list_entry_array(data, &info);
  90         if (r)
  91                 return r;
  92
  93         r = amdgpu_bo_list_create(p->adev, p->filp, info, data->bo_number,
  94                                   &p->bo_list);
  95         if (r)
  96                 goto error_free;
  97
  98         kvfree(info);
  99         return 0;
 100
 101 error_free:
 102         kvfree(info);
 103
 104         return r;
 105 }
 106
 107 static int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, union drm_amdgpu_cs *cs)
 108 {
 109         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 110         struct amdgpu_vm *vm = &fpriv->vm;
 111         uint64_t *chunk_array_user;
 112         uint64_t *chunk_array;
 113         unsigned size, num_ibs = 0;
 114         uint32_t uf_offset = 0;
 115         int i;
 116         int ret;
 117
 118         if (cs->in.num_chunks == 0)
 119                 return -EINVAL;
 120
 121         chunk_array = kvmalloc_array(cs->in.num_chunks, sizeof(uint64_t), GFP_KERNEL);
 122         if (!chunk_array)
 123                 return -ENOMEM;
 124
 125         p->ctx = amdgpu_ctx_get(fpriv, cs->in.ctx_id);
 126         if (!p->ctx) {
 127                 ret = -EINVAL;
 128                 goto free_chunk;
 129         }
 130
 131         mutex_lock(&p->ctx->lock);
 132
 133         /* skip guilty context job */
 134         if (atomic_read(&p->ctx->guilty) == 1) {
 135                 ret = -ECANCELED;
 136                 goto free_chunk;
 137         }
 138
 139         /* get chunks */
 140         chunk_array_user = u64_to_user_ptr(cs->in.chunks);
 141         if (copy_from_user(chunk_array, chunk_array_user,
 142                            sizeof(uint64_t)*cs->in.num_chunks)) {
 143                 ret = -EFAULT;
 144                 goto free_chunk;
 145         }
 146
 147         p->nchunks = cs->in.num_chunks;
 148         p->chunks = kvmalloc_array(p->nchunks, sizeof(struct amdgpu_cs_chunk),
 149                             GFP_KERNEL);
 150         if (!p->chunks) {
 151                 ret = -ENOMEM;
 152                 goto free_chunk;
 153         }
 154
 155         for (i = 0; i < p->nchunks; i++) {
 156                 struct drm_amdgpu_cs_chunk __user **chunk_ptr = NULL;
 157                 struct drm_amdgpu_cs_chunk user_chunk;
 158                 uint32_t __user *cdata;
 159
 160                 chunk_ptr = u64_to_user_ptr(chunk_array[i]);
 161                 if (copy_from_user(&user_chunk, chunk_ptr,
 162                                        sizeof(struct drm_amdgpu_cs_chunk))) {
 163                         ret = -EFAULT;
 164                         i--;
 165                         goto free_partial_kdata;
 166                 }
 167                 p->chunks[i].chunk_id = user_chunk.chunk_id;
 168                 p->chunks[i].length_dw = user_chunk.length_dw;
 169
 170                 size = p->chunks[i].length_dw;
 171                 cdata = u64_to_user_ptr(user_chunk.chunk_data);
 172
 173                 p->chunks[i].kdata = kvmalloc_array(size, sizeof(uint32_t), GFP_KERNEL);
 174                 if (p->chunks[i].kdata == NULL) {
 175                         ret = -ENOMEM;
 176                         i--;
 177                         goto free_partial_kdata;
 178                 }
 179                 size *= sizeof(uint32_t);
 180                 if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
 181                         ret = -EFAULT;
 182                         goto free_partial_kdata;
 183                 }
 184
 185                 switch (p->chunks[i].chunk_id) {
 186                 case AMDGPU_CHUNK_ID_IB:
 187                         ++num_ibs;
 188                         break;
 189
 190                 case AMDGPU_CHUNK_ID_FENCE:
 191                         size = sizeof(struct drm_amdgpu_cs_chunk_fence);
 192                         if (p->chunks[i].length_dw * sizeof(uint32_t) < size) {
 193                                 ret = -EINVAL;
 194                                 goto free_partial_kdata;
 195                         }
 196
 197                         ret = amdgpu_cs_user_fence_chunk(p, p->chunks[i].kdata,
 198                                                          &uf_offset);
 199                         if (ret)
 200                                 goto free_partial_kdata;
 201
 202                         break;
 203
 204                 case AMDGPU_CHUNK_ID_BO_HANDLES:
 205                         size = sizeof(struct drm_amdgpu_bo_list_in);
 206                         if (p->chunks[i].length_dw * sizeof(uint32_t) < size) {
 207                                 ret = -EINVAL;
 208                                 goto free_partial_kdata;
 209                         }
 210
 211                         ret = amdgpu_cs_bo_handles_chunk(p, p->chunks[i].kdata);
 212                         if (ret)
 213                                 goto free_partial_kdata;
 214
 215                         break;
 216
 217                 case AMDGPU_CHUNK_ID_DEPENDENCIES:
 218                 case AMDGPU_CHUNK_ID_SYNCOBJ_IN:
 219                 case AMDGPU_CHUNK_ID_SYNCOBJ_OUT:
 220                 case AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES:
 221                 case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_WAIT:
 222                 case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_SIGNAL:
 223                         break;
 224
 225                 default:
 226                         ret = -EINVAL;
 227                         goto free_partial_kdata;
 228                 }
 229         }
 230
 231         ret = amdgpu_job_alloc(p->adev, num_ibs, &p->job, vm);
 232         if (ret)
 233                 goto free_all_kdata;
 234
 235         if (p->ctx->vram_lost_counter != p->job->vram_lost_counter) {
 236                 ret = -ECANCELED;
 237                 goto free_all_kdata;
 238         }
 239
 240         if (p->uf_entry.tv.bo)
 241                 p->job->uf_addr = uf_offset;
 242         kvfree(chunk_array);
 243
 244         /* Use this opportunity to fill in task info for the vm */
 245         amdgpu_vm_set_task_info(vm);
 246
 247         return 0;
 248
 249 free_all_kdata:
 250         i = p->nchunks - 1;
 251 free_partial_kdata:
 252         for (; i >= 0; i--)
 253                 kvfree(p->chunks[i].kdata);
 254         kvfree(p->chunks);
 255         p->chunks = NULL;
 256         p->nchunks = 0;
 257 free_chunk:
 258         kvfree(chunk_array);
 259
 260         return ret;
 261 }
 262
 263 /* Convert microseconds to bytes. */
 264 static u64 us_to_bytes(struct amdgpu_device *adev, s64 us)
 265 {
 266         if (us <= 0 || !adev->mm_stats.log2_max_MBps)
 267                 return 0;
 268
 269         /* Since accum_us is incremented by a million per second, just
 270          * multiply it by the number of MB/s to get the number of bytes.
 271          */
 272         return us << adev->mm_stats.log2_max_MBps;
 273 }
 274
 275 static s64 bytes_to_us(struct amdgpu_device *adev, u64 bytes)
 276 {
 277         if (!adev->mm_stats.log2_max_MBps)
 278                 return 0;
 279
 280         return bytes >> adev->mm_stats.log2_max_MBps;
 281 }
 282
 283 /* Returns how many bytes TTM can move right now. If no bytes can be moved,
 284  * it returns 0. If it returns non-zero, it's OK to move at least one buffer,
 285  * which means it can go over the threshold once. If that happens, the driver
 286  * will be in debt and no other buffer migrations can be done until that debt
 287  * is repaid.
 288  *
 289  * This approach allows moving a buffer of any size (it's important to allow
 290  * that).
 291  *
 292  * The currency is simply time in microseconds and it increases as the clock
 293  * ticks. The accumulated microseconds (us) are converted to bytes and
 294  * returned.
 295  */
 296 static void amdgpu_cs_get_threshold_for_moves(struct amdgpu_device *adev,
 297                                               u64 *max_bytes,
 298                                               u64 *max_vis_bytes)
 299 {
 300         s64 time_us, increment_us;
 301         u64 free_vram, total_vram, used_vram;
 302         /* Allow a maximum of 200 accumulated ms. This is basically per-IB
 303          * throttling.
 304          *
 305          * It means that in order to get full max MBps, at least 5 IBs per
 306          * second must be submitted and not more than 200ms apart from each
 307          * other.
 308          */
 309         const s64 us_upper_bound = 200000;
 310
 311         if (!adev->mm_stats.log2_max_MBps) {
 312                 *max_bytes = 0;
 313                 *max_vis_bytes = 0;
 314                 return;
 315         }
 316
 317         total_vram = adev->gmc.real_vram_size - atomic64_read(&adev->vram_pin_size);
 318         used_vram = ttm_resource_manager_usage(&adev->mman.vram_mgr.manager);
 319         free_vram = used_vram >= total_vram ? 0 : total_vram - used_vram;
 320
 321         spin_lock(&adev->mm_stats.lock);
 322
 323         /* Increase the amount of accumulated us. */
 324         time_us = ktime_to_us(ktime_get());
 325         increment_us = time_us - adev->mm_stats.last_update_us;
 326         adev->mm_stats.last_update_us = time_us;
 327         adev->mm_stats.accum_us = min(adev->mm_stats.accum_us + increment_us,
 328                                       us_upper_bound);
 329
 330         /* This prevents the short period of low performance when the VRAM
 331          * usage is low and the driver is in debt or doesn't have enough
 332          * accumulated us to fill VRAM quickly.
 333          *
 334          * The situation can occur in these cases:
 335          * - a lot of VRAM is freed by userspace
 336          * - the presence of a big buffer causes a lot of evictions
 337          *   (solution: split buffers into smaller ones)
 338          *
 339          * If 128 MB or 1/8th of VRAM is free, start filling it now by setting
 340          * accum_us to a positive number.
 341          */
 342         if (free_vram >= 128 * 1024 * 1024 || free_vram >= total_vram / 8) {
 343                 s64 min_us;
 344
 345                 /* Be more aggressive on dGPUs. Try to fill a portion of free
 346                  * VRAM now.
 347                  */
 348                 if (!(adev->flags & AMD_IS_APU))
 349                         min_us = bytes_to_us(adev, free_vram / 4);
 350                 else
 351                         min_us = 0; /* Reset accum_us on APUs. */
 352
 353                 adev->mm_stats.accum_us = max(min_us, adev->mm_stats.accum_us);
 354         }
 355
 356         /* This is set to 0 if the driver is in debt to disallow (optional)
 357          * buffer moves.
 358          */
 359         *max_bytes = us_to_bytes(adev, adev->mm_stats.accum_us);
 360
 361         /* Do the same for visible VRAM if half of it is free */
 362         if (!amdgpu_gmc_vram_full_visible(&adev->gmc)) {
 363                 u64 total_vis_vram = adev->gmc.visible_vram_size;
 364                 u64 used_vis_vram =
 365                   amdgpu_vram_mgr_vis_usage(&adev->mman.vram_mgr);
 366
 367                 if (used_vis_vram < total_vis_vram) {
 368                         u64 free_vis_vram = total_vis_vram - used_vis_vram;
 369                         adev->mm_stats.accum_us_vis = min(adev->mm_stats.accum_us_vis +
 370                                                           increment_us, us_upper_bound);
 371
 372                         if (free_vis_vram >= total_vis_vram / 2)
 373                                 adev->mm_stats.accum_us_vis =
 374                                         max(bytes_to_us(adev, free_vis_vram / 2),
 375                                             adev->mm_stats.accum_us_vis);
 376                 }
 377
 378                 *max_vis_bytes = us_to_bytes(adev, adev->mm_stats.accum_us_vis);
 379         } else {
 380                 *max_vis_bytes = 0;
 381         }
 382
 383         spin_unlock(&adev->mm_stats.lock);
 384 }
 385
 386 /* Report how many bytes have really been moved for the last command
 387  * submission. This can result in a debt that can stop buffer migrations
 388  * temporarily.
 389  */
 390 void amdgpu_cs_report_moved_bytes(struct amdgpu_device *adev, u64 num_bytes,
 391                                   u64 num_vis_bytes)
 392 {
 393         spin_lock(&adev->mm_stats.lock);
 394         adev->mm_stats.accum_us -= bytes_to_us(adev, num_bytes);
 395         adev->mm_stats.accum_us_vis -= bytes_to_us(adev, num_vis_bytes);
 396         spin_unlock(&adev->mm_stats.lock);
 397 }
 398
 399 static int amdgpu_cs_bo_validate(void *param, struct amdgpu_bo *bo)
 400 {
 401         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
 402         struct amdgpu_cs_parser *p = param;
 403         struct ttm_operation_ctx ctx = {
 404                 .interruptible = true,
 405                 .no_wait_gpu = false,
 406                 .resv = bo->tbo.base.resv
 407         };
 408         uint32_t domain;
 409         int r;
 410
 411         if (bo->tbo.pin_count)
 412                 return 0;
 413
 414         /* Don't move this buffer if we have depleted our allowance
 415          * to move it. Don't move anything if the threshold is zero.
 416          */
 417         if (p->bytes_moved < p->bytes_moved_threshold &&
 418             (!bo->tbo.base.dma_buf ||
 419             list_empty(&bo->tbo.base.dma_buf->attachments))) {
 420                 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
 421                     (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)) {
 422                         /* And don't move a CPU_ACCESS_REQUIRED BO to limited
 423                          * visible VRAM if we've depleted our allowance to do
 424                          * that.
 425                          */
 426                         if (p->bytes_moved_vis < p->bytes_moved_vis_threshold)
 427                                 domain = bo->preferred_domains;
 428                         else
 429                                 domain = bo->allowed_domains;
 430                 } else {
 431                         domain = bo->preferred_domains;
 432                 }
 433         } else {
 434                 domain = bo->allowed_domains;
 435         }
 436
 437 retry:
 438         amdgpu_bo_placement_from_domain(bo, domain);
 439         r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 440
 441         p->bytes_moved += ctx.bytes_moved;
 442         if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
 443             amdgpu_bo_in_cpu_visible_vram(bo))
 444                 p->bytes_moved_vis += ctx.bytes_moved;
 445
 446         if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
 447                 domain = bo->allowed_domains;
 448                 goto retry;
 449         }
 450
 451         return r;
 452 }
 453
 454 static int amdgpu_cs_list_validate(struct amdgpu_cs_parser *p,
 455                             struct list_head *validated)
 456 {
 457         struct ttm_operation_ctx ctx = { true, false };
 458         struct amdgpu_bo_list_entry *lobj;
 459         int r;
 460
 461         list_for_each_entry(lobj, validated, tv.head) {
 462                 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(lobj->tv.bo);
 463                 struct mm_struct *usermm;
 464
 465                 usermm = amdgpu_ttm_tt_get_usermm(bo->tbo.ttm);
 466                 if (usermm && usermm != current->mm)
 467                         return -EPERM;
 468
 469                 if (amdgpu_ttm_tt_is_userptr(bo->tbo.ttm) &&
 470                     lobj->user_invalidated && lobj->user_pages) {
 471                         amdgpu_bo_placement_from_domain(bo,
 472                                                         AMDGPU_GEM_DOMAIN_CPU);
 473                         r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 474                         if (r)
 475                                 return r;
 476
 477                         amdgpu_ttm_tt_set_user_pages(bo->tbo.ttm,
 478                                                      lobj->user_pages);
 479                 }
 480
 481                 r = amdgpu_cs_bo_validate(p, bo);
 482                 if (r)
 483                         return r;
 484
 485                 kvfree(lobj->user_pages);
 486                 lobj->user_pages = NULL;
 487         }
 488         return 0;
 489 }
 490
 491 static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
 492                                 union drm_amdgpu_cs *cs)
 493 {
 494         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 495         struct amdgpu_vm *vm = &fpriv->vm;
 496         struct amdgpu_bo_list_entry *e;
 497         struct list_head duplicates;
 498         struct amdgpu_bo *gds;
 499         struct amdgpu_bo *gws;
 500         struct amdgpu_bo *oa;
 501         int r;
 502
 503         INIT_LIST_HEAD(&p->validated);
 504
 505         /* p->bo_list could already be assigned if AMDGPU_CHUNK_ID_BO_HANDLES is present */
 506         if (cs->in.bo_list_handle) {
 507                 if (p->bo_list)
 508                         return -EINVAL;
 509
 510                 r = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle,
 511                                        &p->bo_list);
 512                 if (r)
 513                         return r;
 514         } else if (!p->bo_list) {
 515                 /* Create a empty bo_list when no handle is provided */
 516                 r = amdgpu_bo_list_create(p->adev, p->filp, NULL, 0,
 517                                           &p->bo_list);
 518                 if (r)
 519                         return r;
 520         }
 521
 522         mutex_lock(&p->bo_list->bo_list_mutex);
 523
 524         /* One for TTM and one for the CS job */
 525         amdgpu_bo_list_for_each_entry(e, p->bo_list)
 526                 e->tv.num_shared = 2;
 527
 528         amdgpu_bo_list_get_list(p->bo_list, &p->validated);
 529
 530         INIT_LIST_HEAD(&duplicates);
 531         amdgpu_vm_get_pd_bo(&fpriv->vm, &p->validated, &p->vm_pd);
 532
 533         if (p->uf_entry.tv.bo && !ttm_to_amdgpu_bo(p->uf_entry.tv.bo)->parent)
 534                 list_add(&p->uf_entry.tv.head, &p->validated);
 535
 536         /* Get userptr backing pages. If pages are updated after registered
 537          * in amdgpu_gem_userptr_ioctl(), amdgpu_cs_list_validate() will do
 538          * amdgpu_ttm_backend_bind() to flush and invalidate new pages
 539          */
 540         amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
 541                 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
 542                 bool userpage_invalidated = false;
 543                 int i;
 544
 545                 e->user_pages = kvmalloc_array(bo->tbo.ttm->num_pages,
 546                                         sizeof(struct page *),
 547                                         GFP_KERNEL | __GFP_ZERO);
 548                 if (!e->user_pages) {
 549                         DRM_ERROR("kvmalloc_array failure\n");
 550                         r = -ENOMEM;
 551                         goto out_free_user_pages;
 552                 }
 553
 554                 r = amdgpu_ttm_tt_get_user_pages(bo, e->user_pages);
 555                 if (r) {
 556                         kvfree(e->user_pages);
 557                         e->user_pages = NULL;
 558                         goto out_free_user_pages;
 559                 }
 560
 561                 for (i = 0; i < bo->tbo.ttm->num_pages; i++) {
 562                         if (bo->tbo.ttm->pages[i] != e->user_pages[i]) {
 563                                 userpage_invalidated = true;
 564                                 break;
 565                         }
 566                 }
 567                 e->user_invalidated = userpage_invalidated;
 568         }
 569
 570         r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true,
 571                                    &duplicates);
 572         if (unlikely(r != 0)) {
 573                 if (r != -ERESTARTSYS)
 574                         DRM_ERROR("ttm_eu_reserve_buffers failed.\n");
 575                 goto out_free_user_pages;
 576         }
 577
 578         amdgpu_bo_list_for_each_entry(e, p->bo_list) {
 579                 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
 580
 581                 e->bo_va = amdgpu_vm_bo_find(vm, bo);
 582         }
 583
 584         /* Move fence waiting after getting reservation lock of
 585          * PD root. Then there is no need on a ctx mutex lock.
 586          */
 587         r = amdgpu_ctx_wait_prev_fence(p->ctx, p->entity);
 588         if (unlikely(r != 0)) {
 589                 if (r != -ERESTARTSYS)
 590                         DRM_ERROR("amdgpu_ctx_wait_prev_fence failed.\n");
 591                 goto error_validate;
 592         }
 593
 594         amdgpu_cs_get_threshold_for_moves(p->adev, &p->bytes_moved_threshold,
 595                                           &p->bytes_moved_vis_threshold);
 596         p->bytes_moved = 0;
 597         p->bytes_moved_vis = 0;
 598
 599         r = amdgpu_vm_validate_pt_bos(p->adev, &fpriv->vm,
 600                                       amdgpu_cs_bo_validate, p);
 601         if (r) {
 602                 DRM_ERROR("amdgpu_vm_validate_pt_bos() failed.\n");
 603                 goto error_validate;
 604         }
 605
 606         r = amdgpu_cs_list_validate(p, &duplicates);
 607         if (r)
 608                 goto error_validate;
 609
 610         r = amdgpu_cs_list_validate(p, &p->validated);
 611         if (r)
 612                 goto error_validate;
 613
 614         amdgpu_cs_report_moved_bytes(p->adev, p->bytes_moved,
 615                                      p->bytes_moved_vis);
 616
 617         gds = p->bo_list->gds_obj;
 618         gws = p->bo_list->gws_obj;
 619         oa = p->bo_list->oa_obj;
 620
 621         if (gds) {
 622                 p->job->gds_base = amdgpu_bo_gpu_offset(gds) >> PAGE_SHIFT;
 623                 p->job->gds_size = amdgpu_bo_size(gds) >> PAGE_SHIFT;
 624         }
 625         if (gws) {
 626                 p->job->gws_base = amdgpu_bo_gpu_offset(gws) >> PAGE_SHIFT;
 627                 p->job->gws_size = amdgpu_bo_size(gws) >> PAGE_SHIFT;
 628         }
 629         if (oa) {
 630                 p->job->oa_base = amdgpu_bo_gpu_offset(oa) >> PAGE_SHIFT;
 631                 p->job->oa_size = amdgpu_bo_size(oa) >> PAGE_SHIFT;
 632         }
 633
 634         if (!r && p->uf_entry.tv.bo) {
 635                 struct amdgpu_bo *uf = ttm_to_amdgpu_bo(p->uf_entry.tv.bo);
 636
 637                 r = amdgpu_ttm_alloc_gart(&uf->tbo);
 638                 p->job->uf_addr += amdgpu_bo_gpu_offset(uf);
 639         }
 640
 641 error_validate:
 642         if (r)
 643                 ttm_eu_backoff_reservation(&p->ticket, &p->validated);
 644
 645 out_free_user_pages:
 646         if (r) {
 647                 amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
 648                         struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
 649
 650                         if (!e->user_pages)
 651                                 continue;
 652                         amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm);
 653                         kvfree(e->user_pages);
 654                         e->user_pages = NULL;
 655                 }
 656                 mutex_unlock(&p->bo_list->bo_list_mutex);
 657         }
 658         return r;
 659 }
 660
 661 static int amdgpu_cs_sync_rings(struct amdgpu_cs_parser *p)
 662 {
 663         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 664         struct amdgpu_bo_list_entry *e;
 665         int r;
 666
 667         list_for_each_entry(e, &p->validated, tv.head) {
 668                 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
 669                 struct dma_resv *resv = bo->tbo.base.resv;
 670                 enum amdgpu_sync_mode sync_mode;
 671
 672                 sync_mode = amdgpu_bo_explicit_sync(bo) ?
 673                         AMDGPU_SYNC_EXPLICIT : AMDGPU_SYNC_NE_OWNER;
 674                 r = amdgpu_sync_resv(p->adev, &p->job->sync, resv, sync_mode,
 675                                      &fpriv->vm);
 676                 if (r)
 677                         return r;
 678         }
 679         return 0;
 680 }
 681
 682 /**
 683  * amdgpu_cs_parser_fini() - clean parser states
 684  * @parser:     parser structure holding parsing context.
 685  * @error:      error number
 686  * @backoff:    indicator to backoff the reservation
 687  *
 688  * If error is set then unvalidate buffer, otherwise just free memory
 689  * used by parsing context.
 690  **/
 691 static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error,
 692                                   bool backoff)
 693 {
 694         unsigned i;
 695
 696         if (error && backoff) {
 697                 ttm_eu_backoff_reservation(&parser->ticket,
 698                                            &parser->validated);
 699                 mutex_unlock(&parser->bo_list->bo_list_mutex);
 700         }
 701
 702         for (i = 0; i < parser->num_post_deps; i++) {
 703                 drm_syncobj_put(parser->post_deps[i].syncobj);
 704                 kfree(parser->post_deps[i].chain);
 705         }
 706         kfree(parser->post_deps);
 707
 708         dma_fence_put(parser->fence);
 709
 710         if (parser->ctx) {
 711                 mutex_unlock(&parser->ctx->lock);
 712                 amdgpu_ctx_put(parser->ctx);
 713         }
 714         if (parser->bo_list)
 715                 amdgpu_bo_list_put(parser->bo_list);
 716
 717         for (i = 0; i < parser->nchunks; i++)
 718                 kvfree(parser->chunks[i].kdata);
 719         kvfree(parser->chunks);
 720         if (parser->job)
 721                 amdgpu_job_free(parser->job);
 722         if (parser->uf_entry.tv.bo) {
 723                 struct amdgpu_bo *uf = ttm_to_amdgpu_bo(parser->uf_entry.tv.bo);
 724
 725                 amdgpu_bo_unref(&uf);
 726         }
 727 }
 728
 729 static int amdgpu_cs_vm_handling(struct amdgpu_cs_parser *p)
 730 {
 731         struct amdgpu_ring *ring = to_amdgpu_ring(p->entity->rq->sched);
 732         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 733         struct amdgpu_device *adev = p->adev;
 734         struct amdgpu_vm *vm = &fpriv->vm;
 735         struct amdgpu_bo_list_entry *e;
 736         struct amdgpu_bo_va *bo_va;
 737         struct amdgpu_bo *bo;
 738         int r;
 739
 740         /* Only for UVD/VCE VM emulation */
 741         if (ring->funcs->parse_cs || ring->funcs->patch_cs_in_place) {
 742                 unsigned i, j;
 743
 744                 for (i = 0, j = 0; i < p->nchunks && j < p->job->num_ibs; i++) {
 745                         struct drm_amdgpu_cs_chunk_ib *chunk_ib;
 746                         struct amdgpu_bo_va_mapping *m;
 747                         struct amdgpu_bo *aobj = NULL;
 748                         struct amdgpu_cs_chunk *chunk;
 749                         uint64_t offset, va_start;
 750                         struct amdgpu_ib *ib;
 751                         uint8_t *kptr;
 752
 753                         chunk = &p->chunks[i];
 754                         ib = &p->job->ibs[j];
 755                         chunk_ib = chunk->kdata;
 756
 757                         if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
 758                                 continue;
 759
 760                         va_start = chunk_ib->va_start & AMDGPU_GMC_HOLE_MASK;
 761                         r = amdgpu_cs_find_mapping(p, va_start, &aobj, &m);
 762                         if (r) {
 763                                 DRM_ERROR("IB va_start is invalid\n");
 764                                 return r;
 765                         }
 766
 767                         if ((va_start + chunk_ib->ib_bytes) >
 768                             (m->last + 1) * AMDGPU_GPU_PAGE_SIZE) {
 769                                 DRM_ERROR("IB va_start+ib_bytes is invalid\n");
 770                                 return -EINVAL;
 771                         }
 772
 773                         /* the IB should be reserved at this point */
 774                         r = amdgpu_bo_kmap(aobj, (void **)&kptr);
 775                         if (r) {
 776                                 return r;
 777                         }
 778
 779                         offset = m->start * AMDGPU_GPU_PAGE_SIZE;
 780                         kptr += va_start - offset;
 781
 782                         if (ring->funcs->parse_cs) {
 783                                 memcpy(ib->ptr, kptr, chunk_ib->ib_bytes);
 784                                 amdgpu_bo_kunmap(aobj);
 785
 786                                 r = amdgpu_ring_parse_cs(ring, p, p->job, ib);
 787                                 if (r)
 788                                         return r;
 789                         } else {
 790                                 ib->ptr = (uint32_t *)kptr;
 791                                 r = amdgpu_ring_patch_cs_in_place(ring, p, p->job, ib);
 792                                 amdgpu_bo_kunmap(aobj);
 793                                 if (r)
 794                                         return r;
 795                         }
 796
 797                         j++;
 798                 }
 799         }
 800
 801         if (!p->job->vm)
 802                 return amdgpu_cs_sync_rings(p);
 803
 804
 805         r = amdgpu_vm_clear_freed(adev, vm, NULL);
 806         if (r)
 807                 return r;
 808
 809         r = amdgpu_vm_bo_update(adev, fpriv->prt_va, false);
 810         if (r)
 811                 return r;
 812
 813         r = amdgpu_sync_fence(&p->job->sync, fpriv->prt_va->last_pt_update);
 814         if (r)
 815                 return r;
 816
 817         if (amdgpu_mcbp || amdgpu_sriov_vf(adev)) {
 818                 bo_va = fpriv->csa_va;
 819                 BUG_ON(!bo_va);
 820                 r = amdgpu_vm_bo_update(adev, bo_va, false);
 821                 if (r)
 822                         return r;
 823
 824                 r = amdgpu_sync_fence(&p->job->sync, bo_va->last_pt_update);
 825                 if (r)
 826                         return r;
 827         }
 828
 829         amdgpu_bo_list_for_each_entry(e, p->bo_list) {
 830                 /* ignore duplicates */
 831                 bo = ttm_to_amdgpu_bo(e->tv.bo);
 832                 if (!bo)
 833                         continue;
 834
 835                 bo_va = e->bo_va;
 836                 if (bo_va == NULL)
 837                         continue;
 838
 839                 r = amdgpu_vm_bo_update(adev, bo_va, false);
 840                 if (r) {
 841                         mutex_unlock(&p->bo_list->bo_list_mutex);
 842                         return r;
 843                 }
 844
 845                 r = amdgpu_sync_fence(&p->job->sync, bo_va->last_pt_update);
 846                 if (r) {
 847                         mutex_unlock(&p->bo_list->bo_list_mutex);
 848                         return r;
 849                 }
 850         }
 851
 852         r = amdgpu_vm_handle_moved(adev, vm);
 853         if (r)
 854                 return r;
 855
 856         r = amdgpu_vm_update_pdes(adev, vm, false);
 857         if (r)
 858                 return r;
 859
 860         r = amdgpu_sync_fence(&p->job->sync, vm->last_update);
 861         if (r)
 862                 return r;
 863
 864         p->job->vm_pd_addr = amdgpu_gmc_pd_addr(vm->root.bo);
 865
 866         if (amdgpu_vm_debug) {
 867                 /* Invalidate all BOs to test for userspace bugs */
 868                 amdgpu_bo_list_for_each_entry(e, p->bo_list) {
 869                         struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
 870
 871                         /* ignore duplicates */
 872                         if (!bo)
 873                                 continue;
 874
 875                         amdgpu_vm_bo_invalidate(adev, bo, false);
 876                 }
 877         }
 878
 879         return amdgpu_cs_sync_rings(p);
 880 }
 881
 882 static int amdgpu_cs_ib_fill(struct amdgpu_device *adev,
 883                              struct amdgpu_cs_parser *parser)
 884 {
 885         struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
 886         struct amdgpu_vm *vm = &fpriv->vm;
 887         int r, ce_preempt = 0, de_preempt = 0;
 888         struct amdgpu_ring *ring;
 889         int i, j;
 890
 891         for (i = 0, j = 0; i < parser->nchunks && j < parser->job->num_ibs; i++) {
 892                 struct amdgpu_cs_chunk *chunk;
 893                 struct amdgpu_ib *ib;
 894                 struct drm_amdgpu_cs_chunk_ib *chunk_ib;
 895                 struct drm_sched_entity *entity;
 896
 897                 chunk = &parser->chunks[i];
 898                 ib = &parser->job->ibs[j];
 899                 chunk_ib = (struct drm_amdgpu_cs_chunk_ib *)chunk->kdata;
 900
 901                 if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
 902                         continue;
 903
 904                 if (chunk_ib->ip_type == AMDGPU_HW_IP_GFX &&
 905                     (amdgpu_mcbp || amdgpu_sriov_vf(adev))) {
 906                         if (chunk_ib->flags & AMDGPU_IB_FLAG_PREEMPT) {
 907                                 if (chunk_ib->flags & AMDGPU_IB_FLAG_CE)
 908                                         ce_preempt++;
 909                                 else
 910                                         de_preempt++;
 911                         }
 912
 913                         /* each GFX command submit allows 0 or 1 IB preemptible for CE & DE */
 914                         if (ce_preempt > 1 || de_preempt > 1)
 915                                 return -EINVAL;
 916                 }
 917
 918                 r = amdgpu_ctx_get_entity(parser->ctx, chunk_ib->ip_type,
 919                                           chunk_ib->ip_instance, chunk_ib->ring,
 920                                           &entity);
 921                 if (r)
 922                         return r;
 923
 924                 if (chunk_ib->flags & AMDGPU_IB_FLAG_PREAMBLE)
 925                         parser->job->preamble_status |=
 926                                 AMDGPU_PREAMBLE_IB_PRESENT;
 927
 928                 if (parser->entity && parser->entity != entity)
 929                         return -EINVAL;
 930
 931                 /* Return if there is no run queue associated with this entity.
 932                  * Possibly because of disabled HW IP*/
 933                 if (entity->rq == NULL)
 934                         return -EINVAL;
 935
 936                 parser->entity = entity;
 937
 938                 ring = to_amdgpu_ring(entity->rq->sched);
 939                 r =  amdgpu_ib_get(adev, vm, ring->funcs->parse_cs ?
 940                                    chunk_ib->ib_bytes : 0,
 941                                    AMDGPU_IB_POOL_DELAYED, ib);
 942                 if (r) {
 943                         DRM_ERROR("Failed to get ib !\n");
 944                         return r;
 945                 }
 946
 947                 ib->gpu_addr = chunk_ib->va_start;
 948                 ib->length_dw = chunk_ib->ib_bytes / 4;
 949                 ib->flags = chunk_ib->flags;
 950
 951                 j++;
 952         }
 953
 954         /* MM engine doesn't support user fences */
 955         ring = to_amdgpu_ring(parser->entity->rq->sched);
 956         if (parser->job->uf_addr && ring->funcs->no_user_fence)
 957                 return -EINVAL;
 958
 959         return 0;
 960 }
 961
 962 static int amdgpu_cs_process_fence_dep(struct amdgpu_cs_parser *p,
 963                                        struct amdgpu_cs_chunk *chunk)
 964 {
 965         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 966         unsigned num_deps;
 967         int i, r;
 968         struct drm_amdgpu_cs_chunk_dep *deps;
 969
 970         deps = (struct drm_amdgpu_cs_chunk_dep *)chunk->kdata;
 971         num_deps = chunk->length_dw * 4 /
 972                 sizeof(struct drm_amdgpu_cs_chunk_dep);
 973
 974         for (i = 0; i < num_deps; ++i) {
 975                 struct amdgpu_ctx *ctx;
 976                 struct drm_sched_entity *entity;
 977                 struct dma_fence *fence;
 978
 979                 ctx = amdgpu_ctx_get(fpriv, deps[i].ctx_id);
 980                 if (ctx == NULL)
 981                         return -EINVAL;
 982
 983                 r = amdgpu_ctx_get_entity(ctx, deps[i].ip_type,
 984                                           deps[i].ip_instance,
 985                                           deps[i].ring, &entity);
 986                 if (r) {
 987                         amdgpu_ctx_put(ctx);
 988                         return r;
 989                 }
 990
 991                 fence = amdgpu_ctx_get_fence(ctx, entity, deps[i].handle);
 992                 amdgpu_ctx_put(ctx);
 993
 994                 if (IS_ERR(fence))
 995                         return PTR_ERR(fence);
 996                 else if (!fence)
 997                         continue;
 998
 999                 if (chunk->chunk_id == AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES) {
1000                         struct drm_sched_fence *s_fence;
1001                         struct dma_fence *old = fence;
1002
1003                         s_fence = to_drm_sched_fence(fence);
1004                         fence = dma_fence_get(&s_fence->scheduled);
1005                         dma_fence_put(old);
1006                 }
1007
1008                 r = amdgpu_sync_fence(&p->job->sync, fence);
1009                 dma_fence_put(fence);
1010                 if (r)
1011                         return r;
1012         }
1013         return 0;
1014 }
1015
1016 static int amdgpu_syncobj_lookup_and_add_to_sync(struct amdgpu_cs_parser *p,
1017                                                  uint32_t handle, u64 point,
1018                                                  u64 flags)
1019 {
1020         struct dma_fence *fence;
1021         int r;
1022
1023         r = drm_syncobj_find_fence(p->filp, handle, point, flags, &fence);
1024         if (r) {
1025                 DRM_ERROR("syncobj %u failed to find fence @ %llu (%d)!\n",
1026                           handle, point, r);
1027                 return r;
1028         }
1029
1030         r = amdgpu_sync_fence(&p->job->sync, fence);
1031         dma_fence_put(fence);
1032
1033         return r;
1034 }
1035
1036 static int amdgpu_cs_process_syncobj_in_dep(struct amdgpu_cs_parser *p,
1037                                             struct amdgpu_cs_chunk *chunk)
1038 {
1039         struct drm_amdgpu_cs_chunk_sem *deps;
1040         unsigned num_deps;
1041         int i, r;
1042
1043         deps = (struct drm_amdgpu_cs_chunk_sem *)chunk->kdata;
1044         num_deps = chunk->length_dw * 4 /
1045                 sizeof(struct drm_amdgpu_cs_chunk_sem);
1046         for (i = 0; i < num_deps; ++i) {
1047                 r = amdgpu_syncobj_lookup_and_add_to_sync(p, deps[i].handle,
1048                                                           0, 0);
1049                 if (r)
1050                         return r;
1051         }
1052
1053         return 0;
1054 }
1055
1056
1057 static int amdgpu_cs_process_syncobj_timeline_in_dep(struct amdgpu_cs_parser *p,
1058                                                      struct amdgpu_cs_chunk *chunk)
1059 {
1060         struct drm_amdgpu_cs_chunk_syncobj *syncobj_deps;
1061         unsigned num_deps;
1062         int i, r;
1063
1064         syncobj_deps = (struct drm_amdgpu_cs_chunk_syncobj *)chunk->kdata;
1065         num_deps = chunk->length_dw * 4 /
1066                 sizeof(struct drm_amdgpu_cs_chunk_syncobj);
1067         for (i = 0; i < num_deps; ++i) {
1068                 r = amdgpu_syncobj_lookup_and_add_to_sync(p,
1069                                                           syncobj_deps[i].handle,
1070                                                           syncobj_deps[i].point,
1071                                                           syncobj_deps[i].flags);
1072                 if (r)
1073                         return r;
1074         }
1075
1076         return 0;
1077 }
1078
1079 static int amdgpu_cs_process_syncobj_out_dep(struct amdgpu_cs_parser *p,
1080                                              struct amdgpu_cs_chunk *chunk)
1081 {
1082         struct drm_amdgpu_cs_chunk_sem *deps;
1083         unsigned num_deps;
1084         int i;
1085
1086         deps = (struct drm_amdgpu_cs_chunk_sem *)chunk->kdata;
1087         num_deps = chunk->length_dw * 4 /
1088                 sizeof(struct drm_amdgpu_cs_chunk_sem);
1089
1090         if (p->post_deps)
1091                 return -EINVAL;
1092
1093         p->post_deps = kmalloc_array(num_deps, sizeof(*p->post_deps),
1094                                      GFP_KERNEL);
1095         p->num_post_deps = 0;
1096
1097         if (!p->post_deps)
1098                 return -ENOMEM;
1099
1100
1101         for (i = 0; i < num_deps; ++i) {
1102                 p->post_deps[i].syncobj =
1103                         drm_syncobj_find(p->filp, deps[i].handle);
1104                 if (!p->post_deps[i].syncobj)
1105                         return -EINVAL;
1106                 p->post_deps[i].chain = NULL;
1107                 p->post_deps[i].point = 0;
1108                 p->num_post_deps++;
1109         }
1110
1111         return 0;
1112 }
1113
1114
1115 static int amdgpu_cs_process_syncobj_timeline_out_dep(struct amdgpu_cs_parser *p,
1116                                                       struct amdgpu_cs_chunk *chunk)
1117 {
1118         struct drm_amdgpu_cs_chunk_syncobj *syncobj_deps;
1119         unsigned num_deps;
1120         int i;
1121
1122         syncobj_deps = (struct drm_amdgpu_cs_chunk_syncobj *)chunk->kdata;
1123         num_deps = chunk->length_dw * 4 /
1124                 sizeof(struct drm_amdgpu_cs_chunk_syncobj);
1125
1126         if (p->post_deps)
1127                 return -EINVAL;
1128
1129         p->post_deps = kmalloc_array(num_deps, sizeof(*p->post_deps),
1130                                      GFP_KERNEL);
1131         p->num_post_deps = 0;
1132
1133         if (!p->post_deps)
1134                 return -ENOMEM;
1135
1136         for (i = 0; i < num_deps; ++i) {
1137                 struct amdgpu_cs_post_dep *dep = &p->post_deps[i];
1138
1139                 dep->chain = NULL;
1140                 if (syncobj_deps[i].point) {
1141                         dep->chain = dma_fence_chain_alloc();
1142                         if (!dep->chain)
1143                                 return -ENOMEM;
1144                 }
1145
1146                 dep->syncobj = drm_syncobj_find(p->filp,
1147                                                 syncobj_deps[i].handle);
1148                 if (!dep->syncobj) {
1149                         dma_fence_chain_free(dep->chain);
1150                         return -EINVAL;
1151                 }
1152                 dep->point = syncobj_deps[i].point;
1153                 p->num_post_deps++;
1154         }
1155
1156         return 0;
1157 }
1158
1159 static int amdgpu_cs_dependencies(struct amdgpu_device *adev,
1160                                   struct amdgpu_cs_parser *p)
1161 {
1162         int i, r;
1163
1164         /* TODO: Investigate why we still need the context lock */
1165         mutex_unlock(&p->ctx->lock);
1166
1167         for (i = 0; i < p->nchunks; ++i) {
1168                 struct amdgpu_cs_chunk *chunk;
1169
1170                 chunk = &p->chunks[i];
1171
1172                 switch (chunk->chunk_id) {
1173                 case AMDGPU_CHUNK_ID_DEPENDENCIES:
1174                 case AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES:
1175                         r = amdgpu_cs_process_fence_dep(p, chunk);
1176                         if (r)
1177                                 goto out;
1178                         break;
1179                 case AMDGPU_CHUNK_ID_SYNCOBJ_IN:
1180                         r = amdgpu_cs_process_syncobj_in_dep(p, chunk);
1181                         if (r)
1182                                 goto out;
1183                         break;
1184                 case AMDGPU_CHUNK_ID_SYNCOBJ_OUT:
1185                         r = amdgpu_cs_process_syncobj_out_dep(p, chunk);
1186                         if (r)
1187                                 goto out;
1188                         break;
1189                 case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_WAIT:
1190                         r = amdgpu_cs_process_syncobj_timeline_in_dep(p, chunk);
1191                         if (r)
1192                                 goto out;
1193                         break;
1194                 case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_SIGNAL:
1195                         r = amdgpu_cs_process_syncobj_timeline_out_dep(p, chunk);
1196                         if (r)
1197                                 goto out;
1198                         break;
1199                 }
1200         }
1201
1202 out:
1203         mutex_lock(&p->ctx->lock);
1204         return r;
1205 }
1206
1207 static void amdgpu_cs_post_dependencies(struct amdgpu_cs_parser *p)
1208 {
1209         int i;
1210
1211         for (i = 0; i < p->num_post_deps; ++i) {
1212                 if (p->post_deps[i].chain && p->post_deps[i].point) {
1213                         drm_syncobj_add_point(p->post_deps[i].syncobj,
1214                                               p->post_deps[i].chain,
1215                                               p->fence, p->post_deps[i].point);
1216                         p->post_deps[i].chain = NULL;
1217                 } else {
1218                         drm_syncobj_replace_fence(p->post_deps[i].syncobj,
1219                                                   p->fence);
1220                 }
1221         }
1222 }
1223
1224 static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
1225                             union drm_amdgpu_cs *cs)
1226 {
1227         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
1228         struct drm_sched_entity *entity = p->entity;
1229         struct amdgpu_bo_list_entry *e;
1230         struct amdgpu_job *job;
1231         uint64_t seq;
1232         int r;
1233
1234         job = p->job;
1235         p->job = NULL;
1236
1237         r = drm_sched_job_init(&job->base, entity, &fpriv->vm);
1238         if (r)
1239                 goto error_unlock;
1240
1241         drm_sched_job_arm(&job->base);
1242
1243         /* No memory allocation is allowed while holding the notifier lock.
1244          * The lock is held until amdgpu_cs_submit is finished and fence is
1245          * added to BOs.
1246          */
1247         mutex_lock(&p->adev->notifier_lock);
1248
1249         /* If userptr are invalidated after amdgpu_cs_parser_bos(), return
1250          * -EAGAIN, drmIoctl in libdrm will restart the amdgpu_cs_ioctl.
1251          */
1252         amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
1253                 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
1254
1255                 r |= !amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm);
1256         }
1257         if (r) {
1258                 r = -EAGAIN;
1259                 goto error_abort;
1260         }
1261
1262         p->fence = dma_fence_get(&job->base.s_fence->finished);
1263
1264         seq = amdgpu_ctx_add_fence(p->ctx, entity, p->fence);
1265         amdgpu_cs_post_dependencies(p);
1266
1267         if ((job->preamble_status & AMDGPU_PREAMBLE_IB_PRESENT) &&
1268             !p->ctx->preamble_presented) {
1269                 job->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT_FIRST;
1270                 p->ctx->preamble_presented = true;
1271         }
1272
1273         cs->out.handle = seq;
1274         job->uf_sequence = seq;
1275
1276         amdgpu_job_free_resources(job);
1277
1278         trace_amdgpu_cs_ioctl(job);
1279         amdgpu_vm_bo_trace_cs(&fpriv->vm, &p->ticket);
1280         drm_sched_entity_push_job(&job->base);
1281
1282         amdgpu_vm_move_to_lru_tail(p->adev, &fpriv->vm);
1283
1284         /* Make sure all BOs are remembered as writers */
1285         amdgpu_bo_list_for_each_entry(e, p->bo_list)
1286                 e->tv.num_shared = 0;
1287
1288         ttm_eu_fence_buffer_objects(&p->ticket, &p->validated, p->fence);
1289         mutex_unlock(&p->adev->notifier_lock);
1290         mutex_unlock(&p->bo_list->bo_list_mutex);
1291
1292         return 0;
1293
1294 error_abort:
1295         drm_sched_job_cleanup(&job->base);
1296         mutex_unlock(&p->adev->notifier_lock);
1297
1298 error_unlock:
1299         amdgpu_job_free(job);
1300         return r;
1301 }
1302
1303 static void trace_amdgpu_cs_ibs(struct amdgpu_cs_parser *parser)
1304 {
1305         int i;
1306
1307         if (!trace_amdgpu_cs_enabled())
1308                 return;
1309
1310         for (i = 0; i < parser->job->num_ibs; i++)
1311                 trace_amdgpu_cs(parser, i);
1312 }
1313
1314 int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
1315 {
1316         struct amdgpu_device *adev = drm_to_adev(dev);
1317         union drm_amdgpu_cs *cs = data;
1318         struct amdgpu_cs_parser parser = {};
1319         bool reserved_buffers = false;
1320         int r;
1321
1322         if (amdgpu_ras_intr_triggered())
1323                 return -EHWPOISON;
1324
1325         if (!adev->accel_working)
1326                 return -EBUSY;
1327
1328         parser.adev = adev;
1329         parser.filp = filp;
1330
1331         r = amdgpu_cs_parser_init(&parser, data);
1332         if (r) {
1333                 if (printk_ratelimit())
1334                         DRM_ERROR("Failed to initialize parser %d!\n", r);
1335                 goto out;
1336         }
1337
1338         r = amdgpu_cs_ib_fill(adev, &parser);
1339         if (r)
1340                 goto out;
1341
1342         r = amdgpu_cs_dependencies(adev, &parser);
1343         if (r) {
1344                 DRM_ERROR("Failed in the dependencies handling %d!\n", r);
1345                 goto out;
1346         }
1347
1348         r = amdgpu_cs_parser_bos(&parser, data);
1349         if (r) {
1350                 if (r == -ENOMEM)
1351                         DRM_ERROR("Not enough memory for command submission!\n");
1352                 else if (r != -ERESTARTSYS && r != -EAGAIN)
1353                         DRM_ERROR("Failed to process the buffer list %d!\n", r);
1354                 goto out;
1355         }
1356
1357         reserved_buffers = true;
1358
1359         trace_amdgpu_cs_ibs(&parser);
1360
1361         r = amdgpu_cs_vm_handling(&parser);
1362         if (r)
1363                 goto out;
1364
1365         r = amdgpu_cs_submit(&parser, cs);
1366
1367 out:
1368         amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
1369
1370         return r;
1371 }
1372
1373 /**
1374  * amdgpu_cs_wait_ioctl - wait for a command submission to finish
1375  *
1376  * @dev: drm device
1377  * @data: data from userspace
1378  * @filp: file private
1379  *
1380  * Wait for the command submission identified by handle to finish.
1381  */
1382 int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data,
1383                          struct drm_file *filp)
1384 {
1385         union drm_amdgpu_wait_cs *wait = data;
1386         unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
1387         struct drm_sched_entity *entity;
1388         struct amdgpu_ctx *ctx;
1389         struct dma_fence *fence;
1390         long r;
1391
1392         ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
1393         if (ctx == NULL)
1394                 return -EINVAL;
1395
1396         r = amdgpu_ctx_get_entity(ctx, wait->in.ip_type, wait->in.ip_instance,
1397                                   wait->in.ring, &entity);
1398         if (r) {
1399                 amdgpu_ctx_put(ctx);
1400                 return r;
1401         }
1402
1403         fence = amdgpu_ctx_get_fence(ctx, entity, wait->in.handle);
1404         if (IS_ERR(fence))
1405                 r = PTR_ERR(fence);
1406         else if (fence) {
1407                 r = dma_fence_wait_timeout(fence, true, timeout);
1408                 if (r > 0 && fence->error)
1409                         r = fence->error;
1410                 dma_fence_put(fence);
1411         } else
1412                 r = 1;
1413
1414         amdgpu_ctx_put(ctx);
1415         if (r < 0)
1416                 return r;
1417
1418         memset(wait, 0, sizeof(*wait));
1419         wait->out.status = (r == 0);
1420
1421         return 0;
1422 }
1423
1424 /**
1425  * amdgpu_cs_get_fence - helper to get fence from drm_amdgpu_fence
1426  *
1427  * @adev: amdgpu device
1428  * @filp: file private
1429  * @user: drm_amdgpu_fence copied from user space
1430  */
1431 static struct dma_fence *amdgpu_cs_get_fence(struct amdgpu_device *adev,
1432                                              struct drm_file *filp,
1433                                              struct drm_amdgpu_fence *user)
1434 {
1435         struct drm_sched_entity *entity;
1436         struct amdgpu_ctx *ctx;
1437         struct dma_fence *fence;
1438         int r;
1439
1440         ctx = amdgpu_ctx_get(filp->driver_priv, user->ctx_id);
1441         if (ctx == NULL)
1442                 return ERR_PTR(-EINVAL);
1443
1444         r = amdgpu_ctx_get_entity(ctx, user->ip_type, user->ip_instance,
1445                                   user->ring, &entity);
1446         if (r) {
1447                 amdgpu_ctx_put(ctx);
1448                 return ERR_PTR(r);
1449         }
1450
1451         fence = amdgpu_ctx_get_fence(ctx, entity, user->seq_no);
1452         amdgpu_ctx_put(ctx);
1453
1454         return fence;
1455 }
1456
1457 int amdgpu_cs_fence_to_handle_ioctl(struct drm_device *dev, void *data,
1458                                     struct drm_file *filp)
1459 {
1460         struct amdgpu_device *adev = drm_to_adev(dev);
1461         union drm_amdgpu_fence_to_handle *info = data;
1462         struct dma_fence *fence;
1463         struct drm_syncobj *syncobj;
1464         struct sync_file *sync_file;
1465         int fd, r;
1466
1467         fence = amdgpu_cs_get_fence(adev, filp, &info->in.fence);
1468         if (IS_ERR(fence))
1469                 return PTR_ERR(fence);
1470
1471         if (!fence)
1472                 fence = dma_fence_get_stub();
1473
1474         switch (info->in.what) {
1475         case AMDGPU_FENCE_TO_HANDLE_GET_SYNCOBJ:
1476                 r = drm_syncobj_create(&syncobj, 0, fence);
1477                 dma_fence_put(fence);
1478                 if (r)
1479                         return r;
1480                 r = drm_syncobj_get_handle(filp, syncobj, &info->out.handle);
1481                 drm_syncobj_put(syncobj);
1482                 return r;
1483
1484         case AMDGPU_FENCE_TO_HANDLE_GET_SYNCOBJ_FD:
1485                 r = drm_syncobj_create(&syncobj, 0, fence);
1486                 dma_fence_put(fence);
1487                 if (r)
1488                         return r;
1489                 r = drm_syncobj_get_fd(syncobj, (int *)&info->out.handle);
1490                 drm_syncobj_put(syncobj);
1491                 return r;
1492
1493         case AMDGPU_FENCE_TO_HANDLE_GET_SYNC_FILE_FD:
1494                 fd = get_unused_fd_flags(O_CLOEXEC);
1495                 if (fd < 0) {
1496                         dma_fence_put(fence);
1497                         return fd;
1498                 }
1499
1500                 sync_file = sync_file_create(fence);
1501                 dma_fence_put(fence);
1502                 if (!sync_file) {
1503                         put_unused_fd(fd);
1504                         return -ENOMEM;
1505                 }
1506
1507                 fd_install(fd, sync_file->file);
1508                 info->out.handle = fd;
1509                 return 0;
1510
1511         default:
1512                 dma_fence_put(fence);
1513                 return -EINVAL;
1514         }
1515 }
1516
1517 /**
1518  * amdgpu_cs_wait_all_fences - wait on all fences to signal
1519  *
1520  * @adev: amdgpu device
1521  * @filp: file private
1522  * @wait: wait parameters
1523  * @fences: array of drm_amdgpu_fence
1524  */
1525 static int amdgpu_cs_wait_all_fences(struct amdgpu_device *adev,
1526                                      struct drm_file *filp,
1527                                      union drm_amdgpu_wait_fences *wait,
1528                                      struct drm_amdgpu_fence *fences)
1529 {
1530         uint32_t fence_count = wait->in.fence_count;
1531         unsigned int i;
1532         long r = 1;
1533
1534         for (i = 0; i < fence_count; i++) {
1535                 struct dma_fence *fence;
1536                 unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1537
1538                 fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1539                 if (IS_ERR(fence))
1540                         return PTR_ERR(fence);
1541                 else if (!fence)
1542                         continue;
1543
1544                 r = dma_fence_wait_timeout(fence, true, timeout);
1545                 dma_fence_put(fence);
1546                 if (r < 0)
1547                         return r;
1548
1549                 if (r == 0)
1550                         break;
1551
1552                 if (fence->error)
1553                         return fence->error;
1554         }
1555
1556         memset(wait, 0, sizeof(*wait));
1557         wait->out.status = (r > 0);
1558
1559         return 0;
1560 }
1561
1562 /**
1563  * amdgpu_cs_wait_any_fence - wait on any fence to signal
1564  *
1565  * @adev: amdgpu device
1566  * @filp: file private
1567  * @wait: wait parameters
1568  * @fences: array of drm_amdgpu_fence
1569  */
1570 static int amdgpu_cs_wait_any_fence(struct amdgpu_device *adev,
1571                                     struct drm_file *filp,
1572                                     union drm_amdgpu_wait_fences *wait,
1573                                     struct drm_amdgpu_fence *fences)
1574 {
1575         unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1576         uint32_t fence_count = wait->in.fence_count;
1577         uint32_t first = ~0;
1578         struct dma_fence **array;
1579         unsigned int i;
1580         long r;
1581
1582         /* Prepare the fence array */
1583         array = kcalloc(fence_count, sizeof(struct dma_fence *), GFP_KERNEL);
1584
1585         if (array == NULL)
1586                 return -ENOMEM;
1587
1588         for (i = 0; i < fence_count; i++) {
1589                 struct dma_fence *fence;
1590
1591                 fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1592                 if (IS_ERR(fence)) {
1593                         r = PTR_ERR(fence);
1594                         goto err_free_fence_array;
1595                 } else if (fence) {
1596                         array[i] = fence;
1597                 } else { /* NULL, the fence has been already signaled */
1598                         r = 1;
1599                         first = i;
1600                         goto out;
1601                 }
1602         }
1603
1604         r = dma_fence_wait_any_timeout(array, fence_count, true, timeout,
1605                                        &first);
1606         if (r < 0)
1607                 goto err_free_fence_array;
1608
1609 out:
1610         memset(wait, 0, sizeof(*wait));
1611         wait->out.status = (r > 0);
1612         wait->out.first_signaled = first;
1613
1614         if (first < fence_count && array[first])
1615                 r = array[first]->error;
1616         else
1617                 r = 0;
1618
1619 err_free_fence_array:
1620         for (i = 0; i < fence_count; i++)
1621                 dma_fence_put(array[i]);
1622         kfree(array);
1623
1624         return r;
1625 }
1626
1627 /**
1628  * amdgpu_cs_wait_fences_ioctl - wait for multiple command submissions to finish
1629  *
1630  * @dev: drm device
1631  * @data: data from userspace
1632  * @filp: file private
1633  */
1634 int amdgpu_cs_wait_fences_ioctl(struct drm_device *dev, void *data,
1635                                 struct drm_file *filp)
1636 {
1637         struct amdgpu_device *adev = drm_to_adev(dev);
1638         union drm_amdgpu_wait_fences *wait = data;
1639         uint32_t fence_count = wait->in.fence_count;
1640         struct drm_amdgpu_fence *fences_user;
1641         struct drm_amdgpu_fence *fences;
1642         int r;
1643
1644         /* Get the fences from userspace */
1645         fences = kmalloc_array(fence_count, sizeof(struct drm_amdgpu_fence),
1646                         GFP_KERNEL);
1647         if (fences == NULL)
1648                 return -ENOMEM;
1649
1650         fences_user = u64_to_user_ptr(wait->in.fences);
1651         if (copy_from_user(fences, fences_user,
1652                 sizeof(struct drm_amdgpu_fence) * fence_count)) {
1653                 r = -EFAULT;
1654                 goto err_free_fences;
1655         }
1656
1657         if (wait->in.wait_all)
1658                 r = amdgpu_cs_wait_all_fences(adev, filp, wait, fences);
1659         else
1660                 r = amdgpu_cs_wait_any_fence(adev, filp, wait, fences);
1661
1662 err_free_fences:
1663         kfree(fences);
1664
1665         return r;
1666 }
1667
1668 /**
1669  * amdgpu_cs_find_mapping - find bo_va for VM address
1670  *
1671  * @parser: command submission parser context
1672  * @addr: VM address
1673  * @bo: resulting BO of the mapping found
1674  * @map: Placeholder to return found BO mapping
1675  *
1676  * Search the buffer objects in the command submission context for a certain
1677  * virtual memory address. Returns allocation structure when found, NULL
1678  * otherwise.
1679  */
1680 int amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
1681                            uint64_t addr, struct amdgpu_bo **bo,
1682                            struct amdgpu_bo_va_mapping **map)
1683 {
1684         struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
1685         struct ttm_operation_ctx ctx = { false, false };
1686         struct amdgpu_vm *vm = &fpriv->vm;
1687         struct amdgpu_bo_va_mapping *mapping;
1688         int r;
1689
1690         addr /= AMDGPU_GPU_PAGE_SIZE;
1691
1692         mapping = amdgpu_vm_bo_lookup_mapping(vm, addr);
1693         if (!mapping || !mapping->bo_va || !mapping->bo_va->base.bo)
1694                 return -EINVAL;
1695
1696         *bo = mapping->bo_va->base.bo;
1697         *map = mapping;
1698
1699         /* Double check that the BO is reserved by this CS */
1700         if (dma_resv_locking_ctx((*bo)->tbo.base.resv) != &parser->ticket)
1701                 return -EINVAL;
1702
1703         if (!((*bo)->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)) {
1704                 (*bo)->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
1705                 amdgpu_bo_placement_from_domain(*bo, (*bo)->allowed_domains);
1706                 r = ttm_bo_validate(&(*bo)->tbo, &(*bo)->placement, &ctx);
1707                 if (r)
1708                         return r;
1709         }
1710
1711         return amdgpu_ttm_alloc_gart(&(*bo)->tbo);
1712 }