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