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