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

   1 /*
   2  * Copyright 2008 Jerome Glisse.
   3  * All Rights Reserved.
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the "Software"),
   7  * to deal in the Software without restriction, including without limitation
   8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   9  * and/or sell copies of the Software, and to permit persons to whom the
  10  * Software is furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice (including the next
  13  * paragraph) shall be included in all copies or substantial portions of the
  14  * Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  19  * PRECISION INSIGHT AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
  20  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  21  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
  22  * DEALINGS IN THE SOFTWARE.
  23  *
  24  * Authors:
  25  *    Jerome Glisse <[email protected]>
  26  */
  27 #include <linux/pagemap.h>
  28 #include <linux/sync_file.h>
  29 #include <drm/drmP.h>
  30 #include <drm/amdgpu_drm.h>
  31 #include <drm/drm_syncobj.h>
  32 #include "amdgpu.h"
  33 #include "amdgpu_trace.h"
  34
  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                 other = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
 413
 414                 /* Check if this BO is in one of the domains we need space for */
 415                 if (!(other & domain))
 416                         continue;
 417
 418                 /* Check if we can move this BO somewhere else */
 419                 other = bo->allowed_domains & ~domain;
 420                 if (!other)
 421                         continue;
 422
 423                 /* Good we can try to move this BO somewhere else */
 424                 amdgpu_ttm_placement_from_domain(bo, other);
 425                 update_bytes_moved_vis =
 426                         adev->mc.visible_vram_size < adev->mc.real_vram_size &&
 427                         bo->tbo.mem.mem_type == TTM_PL_VRAM &&
 428                         bo->tbo.mem.start < adev->mc.visible_vram_size >> PAGE_SHIFT;
 429                 initial_bytes_moved = atomic64_read(&adev->num_bytes_moved);
 430                 r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
 431                 bytes_moved = atomic64_read(&adev->num_bytes_moved) -
 432                         initial_bytes_moved;
 433                 p->bytes_moved += bytes_moved;
 434                 if (update_bytes_moved_vis)
 435                         p->bytes_moved_vis += bytes_moved;
 436
 437                 if (unlikely(r))
 438                         break;
 439
 440                 p->evictable = list_prev_entry(p->evictable, tv.head);
 441                 list_move(&candidate->tv.head, &p->validated);
 442
 443                 return true;
 444         }
 445
 446         return false;
 447 }
 448
 449 static int amdgpu_cs_validate(void *param, struct amdgpu_bo *bo)
 450 {
 451         struct amdgpu_cs_parser *p = param;
 452         int r;
 453
 454         do {
 455                 r = amdgpu_cs_bo_validate(p, bo);
 456         } while (r == -ENOMEM && amdgpu_cs_try_evict(p, bo));
 457         if (r)
 458                 return r;
 459
 460         if (bo->shadow)
 461                 r = amdgpu_cs_bo_validate(p, bo->shadow);
 462
 463         return r;
 464 }
 465
 466 static int amdgpu_cs_list_validate(struct amdgpu_cs_parser *p,
 467                             struct list_head *validated)
 468 {
 469         struct amdgpu_bo_list_entry *lobj;
 470         int r;
 471
 472         list_for_each_entry(lobj, validated, tv.head) {
 473                 struct amdgpu_bo *bo = lobj->robj;
 474                 bool binding_userptr = false;
 475                 struct mm_struct *usermm;
 476
 477                 usermm = amdgpu_ttm_tt_get_usermm(bo->tbo.ttm);
 478                 if (usermm && usermm != current->mm)
 479                         return -EPERM;
 480
 481                 /* Check if we have user pages and nobody bound the BO already */
 482                 if (amdgpu_ttm_tt_userptr_needs_pages(bo->tbo.ttm) &&
 483                     lobj->user_pages) {
 484                         amdgpu_ttm_placement_from_domain(bo,
 485                                                          AMDGPU_GEM_DOMAIN_CPU);
 486                         r = ttm_bo_validate(&bo->tbo, &bo->placement, true,
 487                                             false);
 488                         if (r)
 489                                 return r;
 490                         amdgpu_ttm_tt_set_user_pages(bo->tbo.ttm,
 491                                                      lobj->user_pages);
 492                         binding_userptr = true;
 493                 }
 494
 495                 if (p->evictable == lobj)
 496                         p->evictable = NULL;
 497
 498                 r = amdgpu_cs_validate(p, bo);
 499                 if (r)
 500                         return r;
 501
 502                 if (binding_userptr) {
 503                         kvfree(lobj->user_pages);
 504                         lobj->user_pages = NULL;
 505                 }
 506         }
 507         return 0;
 508 }
 509
 510 static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
 511                                 union drm_amdgpu_cs *cs)
 512 {
 513         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 514         struct amdgpu_bo_list_entry *e;
 515         struct list_head duplicates;
 516         unsigned i, tries = 10;
 517         int r;
 518
 519         INIT_LIST_HEAD(&p->validated);
 520
 521         p->bo_list = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle);
 522         if (p->bo_list) {
 523                 amdgpu_bo_list_get_list(p->bo_list, &p->validated);
 524                 if (p->bo_list->first_userptr != p->bo_list->num_entries)
 525                         p->mn = amdgpu_mn_get(p->adev);
 526         }
 527
 528         INIT_LIST_HEAD(&duplicates);
 529         amdgpu_vm_get_pd_bo(&fpriv->vm, &p->validated, &p->vm_pd);
 530
 531         if (p->uf_entry.robj)
 532                 list_add(&p->uf_entry.tv.head, &p->validated);
 533
 534         while (1) {
 535                 struct list_head need_pages;
 536                 unsigned i;
 537
 538                 r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true,
 539                                            &duplicates);
 540                 if (unlikely(r != 0)) {
 541                         if (r != -ERESTARTSYS)
 542                                 DRM_ERROR("ttm_eu_reserve_buffers failed.\n");
 543                         goto error_free_pages;
 544                 }
 545
 546                 /* Without a BO list we don't have userptr BOs */
 547                 if (!p->bo_list)
 548                         break;
 549
 550                 INIT_LIST_HEAD(&need_pages);
 551                 for (i = p->bo_list->first_userptr;
 552                      i < p->bo_list->num_entries; ++i) {
 553                         struct amdgpu_bo *bo;
 554
 555                         e = &p->bo_list->array[i];
 556                         bo = e->robj;
 557
 558                         if (amdgpu_ttm_tt_userptr_invalidated(bo->tbo.ttm,
 559                                  &e->user_invalidated) && e->user_pages) {
 560
 561                                 /* We acquired a page array, but somebody
 562                                  * invalidated it. Free it and try again
 563                                  */
 564                                 release_pages(e->user_pages,
 565                                               bo->tbo.ttm->num_pages);
 566                                 kvfree(e->user_pages);
 567                                 e->user_pages = NULL;
 568                         }
 569
 570                         if (amdgpu_ttm_tt_userptr_needs_pages(bo->tbo.ttm) &&
 571                             !e->user_pages) {
 572                                 list_del(&e->tv.head);
 573                                 list_add(&e->tv.head, &need_pages);
 574
 575                                 amdgpu_bo_unreserve(e->robj);
 576                         }
 577                 }
 578
 579                 if (list_empty(&need_pages))
 580                         break;
 581
 582                 /* Unreserve everything again. */
 583                 ttm_eu_backoff_reservation(&p->ticket, &p->validated);
 584
 585                 /* We tried too many times, just abort */
 586                 if (!--tries) {
 587                         r = -EDEADLK;
 588                         DRM_ERROR("deadlock in %s\n", __func__);
 589                         goto error_free_pages;
 590                 }
 591
 592                 /* Fill the page arrays for all userptrs. */
 593                 list_for_each_entry(e, &need_pages, tv.head) {
 594                         struct ttm_tt *ttm = e->robj->tbo.ttm;
 595
 596                         e->user_pages = kvmalloc_array(ttm->num_pages,
 597                                                          sizeof(struct page*),
 598                                                          GFP_KERNEL | __GFP_ZERO);
 599                         if (!e->user_pages) {
 600                                 r = -ENOMEM;
 601                                 DRM_ERROR("calloc failure in %s\n", __func__);
 602                                 goto error_free_pages;
 603                         }
 604
 605                         r = amdgpu_ttm_tt_get_user_pages(ttm, e->user_pages);
 606                         if (r) {
 607                                 DRM_ERROR("amdgpu_ttm_tt_get_user_pages failed.\n");
 608                                 kvfree(e->user_pages);
 609                                 e->user_pages = NULL;
 610                                 goto error_free_pages;
 611                         }
 612                 }
 613
 614                 /* And try again. */
 615                 list_splice(&need_pages, &p->validated);
 616         }
 617
 618         amdgpu_cs_get_threshold_for_moves(p->adev, &p->bytes_moved_threshold,
 619                                           &p->bytes_moved_vis_threshold);
 620         p->bytes_moved = 0;
 621         p->bytes_moved_vis = 0;
 622         p->evictable = list_last_entry(&p->validated,
 623                                        struct amdgpu_bo_list_entry,
 624                                        tv.head);
 625
 626         r = amdgpu_vm_validate_pt_bos(p->adev, &fpriv->vm,
 627                                       amdgpu_cs_validate, p);
 628         if (r) {
 629                 DRM_ERROR("amdgpu_vm_validate_pt_bos() failed.\n");
 630                 goto error_validate;
 631         }
 632
 633         r = amdgpu_cs_list_validate(p, &duplicates);
 634         if (r) {
 635                 DRM_ERROR("amdgpu_cs_list_validate(duplicates) failed.\n");
 636                 goto error_validate;
 637         }
 638
 639         r = amdgpu_cs_list_validate(p, &p->validated);
 640         if (r) {
 641                 DRM_ERROR("amdgpu_cs_list_validate(validated) failed.\n");
 642                 goto error_validate;
 643         }
 644
 645         amdgpu_cs_report_moved_bytes(p->adev, p->bytes_moved,
 646                                      p->bytes_moved_vis);
 647         if (p->bo_list) {
 648                 struct amdgpu_bo *gds = p->bo_list->gds_obj;
 649                 struct amdgpu_bo *gws = p->bo_list->gws_obj;
 650                 struct amdgpu_bo *oa = p->bo_list->oa_obj;
 651                 struct amdgpu_vm *vm = &fpriv->vm;
 652                 unsigned i;
 653
 654                 for (i = 0; i < p->bo_list->num_entries; i++) {
 655                         struct amdgpu_bo *bo = p->bo_list->array[i].robj;
 656
 657                         p->bo_list->array[i].bo_va = amdgpu_vm_bo_find(vm, bo);
 658                 }
 659
 660                 if (gds) {
 661                         p->job->gds_base = amdgpu_bo_gpu_offset(gds);
 662                         p->job->gds_size = amdgpu_bo_size(gds);
 663                 }
 664                 if (gws) {
 665                         p->job->gws_base = amdgpu_bo_gpu_offset(gws);
 666                         p->job->gws_size = amdgpu_bo_size(gws);
 667                 }
 668                 if (oa) {
 669                         p->job->oa_base = amdgpu_bo_gpu_offset(oa);
 670                         p->job->oa_size = amdgpu_bo_size(oa);
 671                 }
 672         }
 673
 674         if (!r && p->uf_entry.robj) {
 675                 struct amdgpu_bo *uf = p->uf_entry.robj;
 676
 677                 r = amdgpu_ttm_bind(&uf->tbo, &uf->tbo.mem);
 678                 p->job->uf_addr += amdgpu_bo_gpu_offset(uf);
 679         }
 680
 681 error_validate:
 682         if (r)
 683                 ttm_eu_backoff_reservation(&p->ticket, &p->validated);
 684
 685 error_free_pages:
 686
 687         if (p->bo_list) {
 688                 for (i = p->bo_list->first_userptr;
 689                      i < p->bo_list->num_entries; ++i) {
 690                         e = &p->bo_list->array[i];
 691
 692                         if (!e->user_pages)
 693                                 continue;
 694
 695                         release_pages(e->user_pages,
 696                                       e->robj->tbo.ttm->num_pages);
 697                         kvfree(e->user_pages);
 698                 }
 699         }
 700
 701         return r;
 702 }
 703
 704 static int amdgpu_cs_sync_rings(struct amdgpu_cs_parser *p)
 705 {
 706         struct amdgpu_bo_list_entry *e;
 707         int r;
 708
 709         list_for_each_entry(e, &p->validated, tv.head) {
 710                 struct reservation_object *resv = e->robj->tbo.resv;
 711                 r = amdgpu_sync_resv(p->adev, &p->job->sync, resv, p->filp,
 712                                      amdgpu_bo_explicit_sync(e->robj));
 713
 714                 if (r)
 715                         return r;
 716         }
 717         return 0;
 718 }
 719
 720 /**
 721  * cs_parser_fini() - clean parser states
 722  * @parser:     parser structure holding parsing context.
 723  * @error:      error number
 724  *
 725  * If error is set than unvalidate buffer, otherwise just free memory
 726  * used by parsing context.
 727  **/
 728 static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error,
 729                                   bool backoff)
 730 {
 731         unsigned i;
 732
 733         if (error && backoff)
 734                 ttm_eu_backoff_reservation(&parser->ticket,
 735                                            &parser->validated);
 736
 737         for (i = 0; i < parser->num_post_dep_syncobjs; i++)
 738                 drm_syncobj_put(parser->post_dep_syncobjs[i]);
 739         kfree(parser->post_dep_syncobjs);
 740
 741         dma_fence_put(parser->fence);
 742
 743         if (parser->ctx) {
 744                 mutex_unlock(&parser->ctx->lock);
 745                 amdgpu_ctx_put(parser->ctx);
 746         }
 747         if (parser->bo_list)
 748                 amdgpu_bo_list_put(parser->bo_list);
 749
 750         for (i = 0; i < parser->nchunks; i++)
 751                 kvfree(parser->chunks[i].kdata);
 752         kfree(parser->chunks);
 753         if (parser->job)
 754                 amdgpu_job_free(parser->job);
 755         amdgpu_bo_unref(&parser->uf_entry.robj);
 756 }
 757
 758 static int amdgpu_bo_vm_update_pte(struct amdgpu_cs_parser *p)
 759 {
 760         struct amdgpu_device *adev = p->adev;
 761         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 762         struct amdgpu_vm *vm = &fpriv->vm;
 763         struct amdgpu_bo_va *bo_va;
 764         struct amdgpu_bo *bo;
 765         int i, r;
 766
 767         r = amdgpu_vm_update_directories(adev, vm);
 768         if (r)
 769                 return r;
 770
 771         r = amdgpu_vm_clear_freed(adev, vm, NULL);
 772         if (r)
 773                 return r;
 774
 775         r = amdgpu_vm_bo_update(adev, fpriv->prt_va, false);
 776         if (r)
 777                 return r;
 778
 779         r = amdgpu_sync_fence(adev, &p->job->sync,
 780                               fpriv->prt_va->last_pt_update);
 781         if (r)
 782                 return r;
 783
 784         if (amdgpu_sriov_vf(adev)) {
 785                 struct dma_fence *f;
 786
 787                 bo_va = fpriv->csa_va;
 788                 BUG_ON(!bo_va);
 789                 r = amdgpu_vm_bo_update(adev, bo_va, false);
 790                 if (r)
 791                         return r;
 792
 793                 f = bo_va->last_pt_update;
 794                 r = amdgpu_sync_fence(adev, &p->job->sync, f);
 795                 if (r)
 796                         return r;
 797         }
 798
 799         if (p->bo_list) {
 800                 for (i = 0; i < p->bo_list->num_entries; i++) {
 801                         struct dma_fence *f;
 802
 803                         /* ignore duplicates */
 804                         bo = p->bo_list->array[i].robj;
 805                         if (!bo)
 806                                 continue;
 807
 808                         bo_va = p->bo_list->array[i].bo_va;
 809                         if (bo_va == NULL)
 810                                 continue;
 811
 812                         r = amdgpu_vm_bo_update(adev, bo_va, false);
 813                         if (r)
 814                                 return r;
 815
 816                         f = bo_va->last_pt_update;
 817                         r = amdgpu_sync_fence(adev, &p->job->sync, f);
 818                         if (r)
 819                                 return r;
 820                 }
 821
 822         }
 823
 824         r = amdgpu_vm_handle_moved(adev, vm);
 825         if (r)
 826                 return r;
 827
 828         r = amdgpu_sync_fence(adev, &p->job->sync, vm->last_update);
 829         if (r)
 830                 return r;
 831
 832         if (amdgpu_vm_debug && p->bo_list) {
 833                 /* Invalidate all BOs to test for userspace bugs */
 834                 for (i = 0; i < p->bo_list->num_entries; i++) {
 835                         /* ignore duplicates */
 836                         bo = p->bo_list->array[i].robj;
 837                         if (!bo)
 838                                 continue;
 839
 840                         amdgpu_vm_bo_invalidate(adev, bo, false);
 841                 }
 842         }
 843
 844         return r;
 845 }
 846
 847 static int amdgpu_cs_ib_vm_chunk(struct amdgpu_device *adev,
 848                                  struct amdgpu_cs_parser *p)
 849 {
 850         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 851         struct amdgpu_vm *vm = &fpriv->vm;
 852         struct amdgpu_ring *ring = p->job->ring;
 853         int r;
 854
 855         /* Only for UVD/VCE VM emulation */
 856         if (p->job->ring->funcs->parse_cs) {
 857                 unsigned i, j;
 858
 859                 for (i = 0, j = 0; i < p->nchunks && j < p->job->num_ibs; i++) {
 860                         struct drm_amdgpu_cs_chunk_ib *chunk_ib;
 861                         struct amdgpu_bo_va_mapping *m;
 862                         struct amdgpu_bo *aobj = NULL;
 863                         struct amdgpu_cs_chunk *chunk;
 864                         struct amdgpu_ib *ib;
 865                         uint64_t offset;
 866                         uint8_t *kptr;
 867
 868                         chunk = &p->chunks[i];
 869                         ib = &p->job->ibs[j];
 870                         chunk_ib = chunk->kdata;
 871
 872                         if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
 873                                 continue;
 874
 875                         r = amdgpu_cs_find_mapping(p, chunk_ib->va_start,
 876                                                    &aobj, &m);
 877                         if (r) {
 878                                 DRM_ERROR("IB va_start is invalid\n");
 879                                 return r;
 880                         }
 881
 882                         if ((chunk_ib->va_start + chunk_ib->ib_bytes) >
 883                             (m->last + 1) * AMDGPU_GPU_PAGE_SIZE) {
 884                                 DRM_ERROR("IB va_start+ib_bytes is invalid\n");
 885                                 return -EINVAL;
 886                         }
 887
 888                         /* the IB should be reserved at this point */
 889                         r = amdgpu_bo_kmap(aobj, (void **)&kptr);
 890                         if (r) {
 891                                 return r;
 892                         }
 893
 894                         offset = m->start * AMDGPU_GPU_PAGE_SIZE;
 895                         kptr += chunk_ib->va_start - offset;
 896
 897                         memcpy(ib->ptr, kptr, chunk_ib->ib_bytes);
 898                         amdgpu_bo_kunmap(aobj);
 899
 900                         r = amdgpu_ring_parse_cs(ring, p, j);
 901                         if (r)
 902                                 return r;
 903
 904                         j++;
 905                 }
 906         }
 907
 908         if (p->job->vm) {
 909                 p->job->vm_pd_addr = amdgpu_bo_gpu_offset(vm->root.base.bo);
 910
 911                 r = amdgpu_bo_vm_update_pte(p);
 912                 if (r)
 913                         return r;
 914         }
 915
 916         return amdgpu_cs_sync_rings(p);
 917 }
 918
 919 static int amdgpu_cs_ib_fill(struct amdgpu_device *adev,
 920                              struct amdgpu_cs_parser *parser)
 921 {
 922         struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
 923         struct amdgpu_vm *vm = &fpriv->vm;
 924         int i, j;
 925         int r, ce_preempt = 0, de_preempt = 0;
 926
 927         for (i = 0, j = 0; i < parser->nchunks && j < parser->job->num_ibs; i++) {
 928                 struct amdgpu_cs_chunk *chunk;
 929                 struct amdgpu_ib *ib;
 930                 struct drm_amdgpu_cs_chunk_ib *chunk_ib;
 931                 struct amdgpu_ring *ring;
 932
 933                 chunk = &parser->chunks[i];
 934                 ib = &parser->job->ibs[j];
 935                 chunk_ib = (struct drm_amdgpu_cs_chunk_ib *)chunk->kdata;
 936
 937                 if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
 938                         continue;
 939
 940                 if (chunk_ib->ip_type == AMDGPU_HW_IP_GFX && amdgpu_sriov_vf(adev)) {
 941                         if (chunk_ib->flags & AMDGPU_IB_FLAG_PREEMPT) {
 942                                 if (chunk_ib->flags & AMDGPU_IB_FLAG_CE)
 943                                         ce_preempt++;
 944                                 else
 945                                         de_preempt++;
 946                         }
 947
 948                         /* each GFX command submit allows 0 or 1 IB preemptible for CE & DE */
 949                         if (ce_preempt > 1 || de_preempt > 1)
 950                                 return -EINVAL;
 951                 }
 952
 953                 r = amdgpu_queue_mgr_map(adev, &parser->ctx->queue_mgr, chunk_ib->ip_type,
 954                                          chunk_ib->ip_instance, chunk_ib->ring, &ring);
 955                 if (r)
 956                         return r;
 957
 958                 if (chunk_ib->flags & AMDGPU_IB_FLAG_PREAMBLE) {
 959                         parser->job->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT;
 960                         if (!parser->ctx->preamble_presented) {
 961                                 parser->job->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT_FIRST;
 962                                 parser->ctx->preamble_presented = true;
 963                         }
 964                 }
 965
 966                 if (parser->job->ring && parser->job->ring != ring)
 967                         return -EINVAL;
 968
 969                 parser->job->ring = ring;
 970
 971                 r =  amdgpu_ib_get(adev, vm,
 972                                         ring->funcs->parse_cs ? chunk_ib->ib_bytes : 0,
 973                                         ib);
 974                 if (r) {
 975                         DRM_ERROR("Failed to get ib !\n");
 976                         return r;
 977                 }
 978
 979                 ib->gpu_addr = chunk_ib->va_start;
 980                 ib->length_dw = chunk_ib->ib_bytes / 4;
 981                 ib->flags = chunk_ib->flags;
 982
 983                 j++;
 984         }
 985
 986         /* UVD & VCE fw doesn't support user fences */
 987         if (parser->job->uf_addr && (
 988             parser->job->ring->funcs->type == AMDGPU_RING_TYPE_UVD ||
 989             parser->job->ring->funcs->type == AMDGPU_RING_TYPE_VCE))
 990                 return -EINVAL;
 991
 992         return amdgpu_ctx_wait_prev_fence(parser->ctx, parser->job->ring->idx);
 993 }
 994
 995 static int amdgpu_cs_process_fence_dep(struct amdgpu_cs_parser *p,
 996                                        struct amdgpu_cs_chunk *chunk)
 997 {
 998         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 999         unsigned num_deps;
1000         int i, r;
1001         struct drm_amdgpu_cs_chunk_dep *deps;
1002
1003         deps = (struct drm_amdgpu_cs_chunk_dep *)chunk->kdata;
1004         num_deps = chunk->length_dw * 4 /
1005                 sizeof(struct drm_amdgpu_cs_chunk_dep);
1006
1007         for (i = 0; i < num_deps; ++i) {
1008                 struct amdgpu_ring *ring;
1009                 struct amdgpu_ctx *ctx;
1010                 struct dma_fence *fence;
1011
1012                 ctx = amdgpu_ctx_get(fpriv, deps[i].ctx_id);
1013                 if (ctx == NULL)
1014                         return -EINVAL;
1015
1016                 r = amdgpu_queue_mgr_map(p->adev, &ctx->queue_mgr,
1017                                          deps[i].ip_type,
1018                                          deps[i].ip_instance,
1019                                          deps[i].ring, &ring);
1020                 if (r) {
1021                         amdgpu_ctx_put(ctx);
1022                         return r;
1023                 }
1024
1025                 fence = amdgpu_ctx_get_fence(ctx, ring,
1026                                              deps[i].handle);
1027                 if (IS_ERR(fence)) {
1028                         r = PTR_ERR(fence);
1029                         amdgpu_ctx_put(ctx);
1030                         return r;
1031                 } else if (fence) {
1032                         r = amdgpu_sync_fence(p->adev, &p->job->sync,
1033                                               fence);
1034                         dma_fence_put(fence);
1035                         amdgpu_ctx_put(ctx);
1036                         if (r)
1037                                 return r;
1038                 }
1039         }
1040         return 0;
1041 }
1042
1043 static int amdgpu_syncobj_lookup_and_add_to_sync(struct amdgpu_cs_parser *p,
1044                                                  uint32_t handle)
1045 {
1046         int r;
1047         struct dma_fence *fence;
1048         r = drm_syncobj_find_fence(p->filp, handle, &fence);
1049         if (r)
1050                 return r;
1051
1052         r = amdgpu_sync_fence(p->adev, &p->job->sync, fence);
1053         dma_fence_put(fence);
1054
1055         return r;
1056 }
1057
1058 static int amdgpu_cs_process_syncobj_in_dep(struct amdgpu_cs_parser *p,
1059                                             struct amdgpu_cs_chunk *chunk)
1060 {
1061         unsigned num_deps;
1062         int i, r;
1063         struct drm_amdgpu_cs_chunk_sem *deps;
1064
1065         deps = (struct drm_amdgpu_cs_chunk_sem *)chunk->kdata;
1066         num_deps = chunk->length_dw * 4 /
1067                 sizeof(struct drm_amdgpu_cs_chunk_sem);
1068
1069         for (i = 0; i < num_deps; ++i) {
1070                 r = amdgpu_syncobj_lookup_and_add_to_sync(p, deps[i].handle);
1071                 if (r)
1072                         return r;
1073         }
1074         return 0;
1075 }
1076
1077 static int amdgpu_cs_process_syncobj_out_dep(struct amdgpu_cs_parser *p,
1078                                              struct amdgpu_cs_chunk *chunk)
1079 {
1080         unsigned num_deps;
1081         int i;
1082         struct drm_amdgpu_cs_chunk_sem *deps;
1083         deps = (struct drm_amdgpu_cs_chunk_sem *)chunk->kdata;
1084         num_deps = chunk->length_dw * 4 /
1085                 sizeof(struct drm_amdgpu_cs_chunk_sem);
1086
1087         p->post_dep_syncobjs = kmalloc_array(num_deps,
1088                                              sizeof(struct drm_syncobj *),
1089                                              GFP_KERNEL);
1090         p->num_post_dep_syncobjs = 0;
1091
1092         if (!p->post_dep_syncobjs)
1093                 return -ENOMEM;
1094
1095         for (i = 0; i < num_deps; ++i) {
1096                 p->post_dep_syncobjs[i] = drm_syncobj_find(p->filp, deps[i].handle);
1097                 if (!p->post_dep_syncobjs[i])
1098                         return -EINVAL;
1099                 p->num_post_dep_syncobjs++;
1100         }
1101         return 0;
1102 }
1103
1104 static int amdgpu_cs_dependencies(struct amdgpu_device *adev,
1105                                   struct amdgpu_cs_parser *p)
1106 {
1107         int i, r;
1108
1109         for (i = 0; i < p->nchunks; ++i) {
1110                 struct amdgpu_cs_chunk *chunk;
1111
1112                 chunk = &p->chunks[i];
1113
1114                 if (chunk->chunk_id == AMDGPU_CHUNK_ID_DEPENDENCIES) {
1115                         r = amdgpu_cs_process_fence_dep(p, chunk);
1116                         if (r)
1117                                 return r;
1118                 } else if (chunk->chunk_id == AMDGPU_CHUNK_ID_SYNCOBJ_IN) {
1119                         r = amdgpu_cs_process_syncobj_in_dep(p, chunk);
1120                         if (r)
1121                                 return r;
1122                 } else if (chunk->chunk_id == AMDGPU_CHUNK_ID_SYNCOBJ_OUT) {
1123                         r = amdgpu_cs_process_syncobj_out_dep(p, chunk);
1124                         if (r)
1125                                 return r;
1126                 }
1127         }
1128
1129         return 0;
1130 }
1131
1132 static void amdgpu_cs_post_dependencies(struct amdgpu_cs_parser *p)
1133 {
1134         int i;
1135
1136         for (i = 0; i < p->num_post_dep_syncobjs; ++i)
1137                 drm_syncobj_replace_fence(p->post_dep_syncobjs[i], p->fence);
1138 }
1139
1140 static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
1141                             union drm_amdgpu_cs *cs)
1142 {
1143         struct amdgpu_ring *ring = p->job->ring;
1144         struct amd_sched_entity *entity = &p->ctx->rings[ring->idx].entity;
1145         struct amdgpu_job *job;
1146         unsigned i;
1147         uint64_t seq;
1148
1149         int r;
1150
1151         amdgpu_mn_lock(p->mn);
1152         if (p->bo_list) {
1153                 for (i = p->bo_list->first_userptr;
1154                      i < p->bo_list->num_entries; ++i) {
1155                         struct amdgpu_bo *bo = p->bo_list->array[i].robj;
1156
1157                         if (amdgpu_ttm_tt_userptr_needs_pages(bo->tbo.ttm)) {
1158                                 amdgpu_mn_unlock(p->mn);
1159                                 return -ERESTARTSYS;
1160                         }
1161                 }
1162         }
1163
1164         job = p->job;
1165         p->job = NULL;
1166
1167         r = amd_sched_job_init(&job->base, &ring->sched, entity, p->filp);
1168         if (r) {
1169                 amdgpu_job_free(job);
1170                 amdgpu_mn_unlock(p->mn);
1171                 return r;
1172         }
1173
1174         job->owner = p->filp;
1175         job->fence_ctx = entity->fence_context;
1176         p->fence = dma_fence_get(&job->base.s_fence->finished);
1177
1178         r = amdgpu_ctx_add_fence(p->ctx, ring, p->fence, &seq);
1179         if (r) {
1180                 dma_fence_put(p->fence);
1181                 dma_fence_put(&job->base.s_fence->finished);
1182                 amdgpu_job_free(job);
1183                 amdgpu_mn_unlock(p->mn);
1184                 return r;
1185         }
1186
1187         amdgpu_cs_post_dependencies(p);
1188
1189         cs->out.handle = seq;
1190         job->uf_sequence = seq;
1191
1192         amdgpu_job_free_resources(job);
1193         amdgpu_ring_priority_get(job->ring,
1194                                  amd_sched_get_job_priority(&job->base));
1195
1196         trace_amdgpu_cs_ioctl(job);
1197         amd_sched_entity_push_job(&job->base);
1198
1199         ttm_eu_fence_buffer_objects(&p->ticket, &p->validated, p->fence);
1200         amdgpu_mn_unlock(p->mn);
1201
1202         return 0;
1203 }
1204
1205 int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
1206 {
1207         struct amdgpu_device *adev = dev->dev_private;
1208         union drm_amdgpu_cs *cs = data;
1209         struct amdgpu_cs_parser parser = {};
1210         bool reserved_buffers = false;
1211         int i, r;
1212
1213         if (!adev->accel_working)
1214                 return -EBUSY;
1215
1216         parser.adev = adev;
1217         parser.filp = filp;
1218
1219         r = amdgpu_cs_parser_init(&parser, data);
1220         if (r) {
1221                 DRM_ERROR("Failed to initialize parser !\n");
1222                 goto out;
1223         }
1224
1225         r = amdgpu_cs_ib_fill(adev, &parser);
1226         if (r)
1227                 goto out;
1228
1229         r = amdgpu_cs_parser_bos(&parser, data);
1230         if (r) {
1231                 if (r == -ENOMEM)
1232                         DRM_ERROR("Not enough memory for command submission!\n");
1233                 else if (r != -ERESTARTSYS)
1234                         DRM_ERROR("Failed to process the buffer list %d!\n", r);
1235                 goto out;
1236         }
1237
1238         reserved_buffers = true;
1239
1240         r = amdgpu_cs_dependencies(adev, &parser);
1241         if (r) {
1242                 DRM_ERROR("Failed in the dependencies handling %d!\n", r);
1243                 goto out;
1244         }
1245
1246         for (i = 0; i < parser.job->num_ibs; i++)
1247                 trace_amdgpu_cs(&parser, i);
1248
1249         r = amdgpu_cs_ib_vm_chunk(adev, &parser);
1250         if (r)
1251                 goto out;
1252
1253         r = amdgpu_cs_submit(&parser, cs);
1254
1255 out:
1256         amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
1257         return r;
1258 }
1259
1260 /**
1261  * amdgpu_cs_wait_ioctl - wait for a command submission to finish
1262  *
1263  * @dev: drm device
1264  * @data: data from userspace
1265  * @filp: file private
1266  *
1267  * Wait for the command submission identified by handle to finish.
1268  */
1269 int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data,
1270                          struct drm_file *filp)
1271 {
1272         union drm_amdgpu_wait_cs *wait = data;
1273         struct amdgpu_device *adev = dev->dev_private;
1274         unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
1275         struct amdgpu_ring *ring = NULL;
1276         struct amdgpu_ctx *ctx;
1277         struct dma_fence *fence;
1278         long r;
1279
1280         ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
1281         if (ctx == NULL)
1282                 return -EINVAL;
1283
1284         r = amdgpu_queue_mgr_map(adev, &ctx->queue_mgr,
1285                                  wait->in.ip_type, wait->in.ip_instance,
1286                                  wait->in.ring, &ring);
1287         if (r) {
1288                 amdgpu_ctx_put(ctx);
1289                 return r;
1290         }
1291
1292         fence = amdgpu_ctx_get_fence(ctx, ring, wait->in.handle);
1293         if (IS_ERR(fence))
1294                 r = PTR_ERR(fence);
1295         else if (fence) {
1296                 r = dma_fence_wait_timeout(fence, true, timeout);
1297                 if (r > 0 && fence->error)
1298                         r = fence->error;
1299                 dma_fence_put(fence);
1300         } else
1301                 r = 1;
1302
1303         amdgpu_ctx_put(ctx);
1304         if (r < 0)
1305                 return r;
1306
1307         memset(wait, 0, sizeof(*wait));
1308         wait->out.status = (r == 0);
1309
1310         return 0;
1311 }
1312
1313 /**
1314  * amdgpu_cs_get_fence - helper to get fence from drm_amdgpu_fence
1315  *
1316  * @adev: amdgpu device
1317  * @filp: file private
1318  * @user: drm_amdgpu_fence copied from user space
1319  */
1320 static struct dma_fence *amdgpu_cs_get_fence(struct amdgpu_device *adev,
1321                                              struct drm_file *filp,
1322                                              struct drm_amdgpu_fence *user)
1323 {
1324         struct amdgpu_ring *ring;
1325         struct amdgpu_ctx *ctx;
1326         struct dma_fence *fence;
1327         int r;
1328
1329         ctx = amdgpu_ctx_get(filp->driver_priv, user->ctx_id);
1330         if (ctx == NULL)
1331                 return ERR_PTR(-EINVAL);
1332
1333         r = amdgpu_queue_mgr_map(adev, &ctx->queue_mgr, user->ip_type,
1334                                  user->ip_instance, user->ring, &ring);
1335         if (r) {
1336                 amdgpu_ctx_put(ctx);
1337                 return ERR_PTR(r);
1338         }
1339
1340         fence = amdgpu_ctx_get_fence(ctx, ring, user->seq_no);
1341         amdgpu_ctx_put(ctx);
1342
1343         return fence;
1344 }
1345
1346 int amdgpu_cs_fence_to_handle_ioctl(struct drm_device *dev, void *data,
1347                                     struct drm_file *filp)
1348 {
1349         struct amdgpu_device *adev = dev->dev_private;
1350         union drm_amdgpu_fence_to_handle *info = data;
1351         struct dma_fence *fence;
1352         struct drm_syncobj *syncobj;
1353         struct sync_file *sync_file;
1354         int fd, r;
1355
1356         fence = amdgpu_cs_get_fence(adev, filp, &info->in.fence);
1357         if (IS_ERR(fence))
1358                 return PTR_ERR(fence);
1359
1360         switch (info->in.what) {
1361         case AMDGPU_FENCE_TO_HANDLE_GET_SYNCOBJ:
1362                 r = drm_syncobj_create(&syncobj, 0, fence);
1363                 dma_fence_put(fence);
1364                 if (r)
1365                         return r;
1366                 r = drm_syncobj_get_handle(filp, syncobj, &info->out.handle);
1367                 drm_syncobj_put(syncobj);
1368                 return r;
1369
1370         case AMDGPU_FENCE_TO_HANDLE_GET_SYNCOBJ_FD:
1371                 r = drm_syncobj_create(&syncobj, 0, fence);
1372                 dma_fence_put(fence);
1373                 if (r)
1374                         return r;
1375                 r = drm_syncobj_get_fd(syncobj, (int*)&info->out.handle);
1376                 drm_syncobj_put(syncobj);
1377                 return r;
1378
1379         case AMDGPU_FENCE_TO_HANDLE_GET_SYNC_FILE_FD:
1380                 fd = get_unused_fd_flags(O_CLOEXEC);
1381                 if (fd < 0) {
1382                         dma_fence_put(fence);
1383                         return fd;
1384                 }
1385
1386                 sync_file = sync_file_create(fence);
1387                 dma_fence_put(fence);
1388                 if (!sync_file) {
1389                         put_unused_fd(fd);
1390                         return -ENOMEM;
1391                 }
1392
1393                 fd_install(fd, sync_file->file);
1394                 info->out.handle = fd;
1395                 return 0;
1396
1397         default:
1398                 return -EINVAL;
1399         }
1400 }
1401
1402 /**
1403  * amdgpu_cs_wait_all_fence - wait on all fences to signal
1404  *
1405  * @adev: amdgpu device
1406  * @filp: file private
1407  * @wait: wait parameters
1408  * @fences: array of drm_amdgpu_fence
1409  */
1410 static int amdgpu_cs_wait_all_fences(struct amdgpu_device *adev,
1411                                      struct drm_file *filp,
1412                                      union drm_amdgpu_wait_fences *wait,
1413                                      struct drm_amdgpu_fence *fences)
1414 {
1415         uint32_t fence_count = wait->in.fence_count;
1416         unsigned int i;
1417         long r = 1;
1418
1419         for (i = 0; i < fence_count; i++) {
1420                 struct dma_fence *fence;
1421                 unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1422
1423                 fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1424                 if (IS_ERR(fence))
1425                         return PTR_ERR(fence);
1426                 else if (!fence)
1427                         continue;
1428
1429                 r = dma_fence_wait_timeout(fence, true, timeout);
1430                 dma_fence_put(fence);
1431                 if (r < 0)
1432                         return r;
1433
1434                 if (r == 0)
1435                         break;
1436
1437                 if (fence->error)
1438                         return fence->error;
1439         }
1440
1441         memset(wait, 0, sizeof(*wait));
1442         wait->out.status = (r > 0);
1443
1444         return 0;
1445 }
1446
1447 /**
1448  * amdgpu_cs_wait_any_fence - wait on any fence to signal
1449  *
1450  * @adev: amdgpu device
1451  * @filp: file private
1452  * @wait: wait parameters
1453  * @fences: array of drm_amdgpu_fence
1454  */
1455 static int amdgpu_cs_wait_any_fence(struct amdgpu_device *adev,
1456                                     struct drm_file *filp,
1457                                     union drm_amdgpu_wait_fences *wait,
1458                                     struct drm_amdgpu_fence *fences)
1459 {
1460         unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1461         uint32_t fence_count = wait->in.fence_count;
1462         uint32_t first = ~0;
1463         struct dma_fence **array;
1464         unsigned int i;
1465         long r;
1466
1467         /* Prepare the fence array */
1468         array = kcalloc(fence_count, sizeof(struct dma_fence *), GFP_KERNEL);
1469
1470         if (array == NULL)
1471                 return -ENOMEM;
1472
1473         for (i = 0; i < fence_count; i++) {
1474                 struct dma_fence *fence;
1475
1476                 fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1477                 if (IS_ERR(fence)) {
1478                         r = PTR_ERR(fence);
1479                         goto err_free_fence_array;
1480                 } else if (fence) {
1481                         array[i] = fence;
1482                 } else { /* NULL, the fence has been already signaled */
1483                         r = 1;
1484                         first = i;
1485                         goto out;
1486                 }
1487         }
1488
1489         r = dma_fence_wait_any_timeout(array, fence_count, true, timeout,
1490                                        &first);
1491         if (r < 0)
1492                 goto err_free_fence_array;
1493
1494 out:
1495         memset(wait, 0, sizeof(*wait));
1496         wait->out.status = (r > 0);
1497         wait->out.first_signaled = first;
1498
1499         if (first < fence_count && array[first])
1500                 r = array[first]->error;
1501         else
1502                 r = 0;
1503
1504 err_free_fence_array:
1505         for (i = 0; i < fence_count; i++)
1506                 dma_fence_put(array[i]);
1507         kfree(array);
1508
1509         return r;
1510 }
1511
1512 /**
1513  * amdgpu_cs_wait_fences_ioctl - wait for multiple command submissions to finish
1514  *
1515  * @dev: drm device
1516  * @data: data from userspace
1517  * @filp: file private
1518  */
1519 int amdgpu_cs_wait_fences_ioctl(struct drm_device *dev, void *data,
1520                                 struct drm_file *filp)
1521 {
1522         struct amdgpu_device *adev = dev->dev_private;
1523         union drm_amdgpu_wait_fences *wait = data;
1524         uint32_t fence_count = wait->in.fence_count;
1525         struct drm_amdgpu_fence *fences_user;
1526         struct drm_amdgpu_fence *fences;
1527         int r;
1528
1529         /* Get the fences from userspace */
1530         fences = kmalloc_array(fence_count, sizeof(struct drm_amdgpu_fence),
1531                         GFP_KERNEL);
1532         if (fences == NULL)
1533                 return -ENOMEM;
1534
1535         fences_user = u64_to_user_ptr(wait->in.fences);
1536         if (copy_from_user(fences, fences_user,
1537                 sizeof(struct drm_amdgpu_fence) * fence_count)) {
1538                 r = -EFAULT;
1539                 goto err_free_fences;
1540         }
1541
1542         if (wait->in.wait_all)
1543                 r = amdgpu_cs_wait_all_fences(adev, filp, wait, fences);
1544         else
1545                 r = amdgpu_cs_wait_any_fence(adev, filp, wait, fences);
1546
1547 err_free_fences:
1548         kfree(fences);
1549
1550         return r;
1551 }
1552
1553 /**
1554  * amdgpu_cs_find_bo_va - find bo_va for VM address
1555  *
1556  * @parser: command submission parser context
1557  * @addr: VM address
1558  * @bo: resulting BO of the mapping found
1559  *
1560  * Search the buffer objects in the command submission context for a certain
1561  * virtual memory address. Returns allocation structure when found, NULL
1562  * otherwise.
1563  */
1564 int amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
1565                            uint64_t addr, struct amdgpu_bo **bo,
1566                            struct amdgpu_bo_va_mapping **map)
1567 {
1568         struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
1569         struct amdgpu_vm *vm = &fpriv->vm;
1570         struct amdgpu_bo_va_mapping *mapping;
1571         int r;
1572
1573         addr /= AMDGPU_GPU_PAGE_SIZE;
1574
1575         mapping = amdgpu_vm_bo_lookup_mapping(vm, addr);
1576         if (!mapping || !mapping->bo_va || !mapping->bo_va->base.bo)
1577                 return -EINVAL;
1578
1579         *bo = mapping->bo_va->base.bo;
1580         *map = mapping;
1581
1582         /* Double check that the BO is reserved by this CS */
1583         if (READ_ONCE((*bo)->tbo.resv->lock.ctx) != &parser->ticket)
1584                 return -EINVAL;
1585
1586         if (!((*bo)->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)) {
1587                 (*bo)->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
1588                 amdgpu_ttm_placement_from_domain(*bo, (*bo)->allowed_domains);
1589                 r = ttm_bo_validate(&(*bo)->tbo, &(*bo)->placement, false,
1590                                     false);
1591                 if (r)
1592                         return r;
1593         }
1594
1595         return amdgpu_ttm_bind(&(*bo)->tbo, &(*bo)->tbo.mem);
1596 }