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 <drm/drmP.h>
  29 #include <drm/amdgpu_drm.h>
  30 #include "amdgpu.h"
  31 #include "amdgpu_trace.h"
  32
  33 int amdgpu_cs_get_ring(struct amdgpu_device *adev, u32 ip_type,
  34                        u32 ip_instance, u32 ring,
  35                        struct amdgpu_ring **out_ring)
  36 {
  37         /* Right now all IPs have only one instance - multiple rings. */
  38         if (ip_instance != 0) {
  39                 DRM_ERROR("invalid ip instance: %d\n", ip_instance);
  40                 return -EINVAL;
  41         }
  42
  43         switch (ip_type) {
  44         default:
  45                 DRM_ERROR("unknown ip type: %d\n", ip_type);
  46                 return -EINVAL;
  47         case AMDGPU_HW_IP_GFX:
  48                 if (ring < adev->gfx.num_gfx_rings) {
  49                         *out_ring = &adev->gfx.gfx_ring[ring];
  50                 } else {
  51                         DRM_ERROR("only %d gfx rings are supported now\n",
  52                                   adev->gfx.num_gfx_rings);
  53                         return -EINVAL;
  54                 }
  55                 break;
  56         case AMDGPU_HW_IP_COMPUTE:
  57                 if (ring < adev->gfx.num_compute_rings) {
  58                         *out_ring = &adev->gfx.compute_ring[ring];
  59                 } else {
  60                         DRM_ERROR("only %d compute rings are supported now\n",
  61                                   adev->gfx.num_compute_rings);
  62                         return -EINVAL;
  63                 }
  64                 break;
  65         case AMDGPU_HW_IP_DMA:
  66                 if (ring < adev->sdma.num_instances) {
  67                         *out_ring = &adev->sdma.instance[ring].ring;
  68                 } else {
  69                         DRM_ERROR("only %d SDMA rings are supported\n",
  70                                   adev->sdma.num_instances);
  71                         return -EINVAL;
  72                 }
  73                 break;
  74         case AMDGPU_HW_IP_UVD:
  75                 *out_ring = &adev->uvd.ring;
  76                 break;
  77         case AMDGPU_HW_IP_VCE:
  78                 if (ring < adev->vce.num_rings){
  79                         *out_ring = &adev->vce.ring[ring];
  80                 } else {
  81                         DRM_ERROR("only %d VCE rings are supported\n", adev->vce.num_rings);
  82                         return -EINVAL;
  83                 }
  84                 break;
  85         case AMDGPU_HW_IP_UVD_ENC:
  86                 if (ring < adev->uvd.num_enc_rings){
  87                         *out_ring = &adev->uvd.ring_enc[ring];
  88                 } else {
  89                         DRM_ERROR("only %d UVD ENC rings are supported\n",
  90                                 adev->uvd.num_enc_rings);
  91                         return -EINVAL;
  92                 }
  93                 break;
  94         }
  95
  96         if (!(*out_ring && (*out_ring)->adev)) {
  97                 DRM_ERROR("Ring %d is not initialized on IP %d\n",
  98                           ring, ip_type);
  99                 return -EINVAL;
 100         }
 101
 102         return 0;
 103 }
 104
 105 static int amdgpu_cs_user_fence_chunk(struct amdgpu_cs_parser *p,
 106                                       struct drm_amdgpu_cs_chunk_fence *data,
 107                                       uint32_t *offset)
 108 {
 109         struct drm_gem_object *gobj;
 110         unsigned long size;
 111
 112         gobj = drm_gem_object_lookup(p->filp, data->handle);
 113         if (gobj == NULL)
 114                 return -EINVAL;
 115
 116         p->uf_entry.robj = amdgpu_bo_ref(gem_to_amdgpu_bo(gobj));
 117         p->uf_entry.priority = 0;
 118         p->uf_entry.tv.bo = &p->uf_entry.robj->tbo;
 119         p->uf_entry.tv.shared = true;
 120         p->uf_entry.user_pages = NULL;
 121
 122         size = amdgpu_bo_size(p->uf_entry.robj);
 123         if (size != PAGE_SIZE || (data->offset + 8) > size)
 124                 return -EINVAL;
 125
 126         *offset = data->offset;
 127
 128         drm_gem_object_unreference_unlocked(gobj);
 129
 130         if (amdgpu_ttm_tt_get_usermm(p->uf_entry.robj->tbo.ttm)) {
 131                 amdgpu_bo_unref(&p->uf_entry.robj);
 132                 return -EINVAL;
 133         }
 134
 135         return 0;
 136 }
 137
 138 int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data)
 139 {
 140         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 141         struct amdgpu_vm *vm = &fpriv->vm;
 142         union drm_amdgpu_cs *cs = data;
 143         uint64_t *chunk_array_user;
 144         uint64_t *chunk_array;
 145         unsigned size, num_ibs = 0;
 146         uint32_t uf_offset = 0;
 147         int i;
 148         int ret;
 149
 150         if (cs->in.num_chunks == 0)
 151                 return 0;
 152
 153         chunk_array = kmalloc_array(cs->in.num_chunks, sizeof(uint64_t), GFP_KERNEL);
 154         if (!chunk_array)
 155                 return -ENOMEM;
 156
 157         p->ctx = amdgpu_ctx_get(fpriv, cs->in.ctx_id);
 158         if (!p->ctx) {
 159                 ret = -EINVAL;
 160                 goto free_chunk;
 161         }
 162
 163         /* get chunks */
 164         chunk_array_user = (uint64_t __user *)(uintptr_t)(cs->in.chunks);
 165         if (copy_from_user(chunk_array, chunk_array_user,
 166                            sizeof(uint64_t)*cs->in.num_chunks)) {
 167                 ret = -EFAULT;
 168                 goto put_ctx;
 169         }
 170
 171         p->nchunks = cs->in.num_chunks;
 172         p->chunks = kmalloc_array(p->nchunks, sizeof(struct amdgpu_cs_chunk),
 173                             GFP_KERNEL);
 174         if (!p->chunks) {
 175                 ret = -ENOMEM;
 176                 goto put_ctx;
 177         }
 178
 179         for (i = 0; i < p->nchunks; i++) {
 180                 struct drm_amdgpu_cs_chunk __user **chunk_ptr = NULL;
 181                 struct drm_amdgpu_cs_chunk user_chunk;
 182                 uint32_t __user *cdata;
 183
 184                 chunk_ptr = (void __user *)(uintptr_t)chunk_array[i];
 185                 if (copy_from_user(&user_chunk, chunk_ptr,
 186                                        sizeof(struct drm_amdgpu_cs_chunk))) {
 187                         ret = -EFAULT;
 188                         i--;
 189                         goto free_partial_kdata;
 190                 }
 191                 p->chunks[i].chunk_id = user_chunk.chunk_id;
 192                 p->chunks[i].length_dw = user_chunk.length_dw;
 193
 194                 size = p->chunks[i].length_dw;
 195                 cdata = (void __user *)(uintptr_t)user_chunk.chunk_data;
 196
 197                 p->chunks[i].kdata = kvmalloc_array(size, sizeof(uint32_t), GFP_KERNEL);
 198                 if (p->chunks[i].kdata == NULL) {
 199                         ret = -ENOMEM;
 200                         i--;
 201                         goto free_partial_kdata;
 202                 }
 203                 size *= sizeof(uint32_t);
 204                 if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
 205                         ret = -EFAULT;
 206                         goto free_partial_kdata;
 207                 }
 208
 209                 switch (p->chunks[i].chunk_id) {
 210                 case AMDGPU_CHUNK_ID_IB:
 211                         ++num_ibs;
 212                         break;
 213
 214                 case AMDGPU_CHUNK_ID_FENCE:
 215                         size = sizeof(struct drm_amdgpu_cs_chunk_fence);
 216                         if (p->chunks[i].length_dw * sizeof(uint32_t) < size) {
 217                                 ret = -EINVAL;
 218                                 goto free_partial_kdata;
 219                         }
 220
 221                         ret = amdgpu_cs_user_fence_chunk(p, p->chunks[i].kdata,
 222                                                          &uf_offset);
 223                         if (ret)
 224                                 goto free_partial_kdata;
 225
 226                         break;
 227
 228                 case AMDGPU_CHUNK_ID_DEPENDENCIES:
 229                         break;
 230
 231                 default:
 232                         ret = -EINVAL;
 233                         goto free_partial_kdata;
 234                 }
 235         }
 236
 237         ret = amdgpu_job_alloc(p->adev, num_ibs, &p->job, vm);
 238         if (ret)
 239                 goto free_all_kdata;
 240
 241         if (p->uf_entry.robj)
 242                 p->job->uf_addr = uf_offset;
 243         kfree(chunk_array);
 244         return 0;
 245
 246 free_all_kdata:
 247         i = p->nchunks - 1;
 248 free_partial_kdata:
 249         for (; i >= 0; i--)
 250                 kvfree(p->chunks[i].kdata);
 251         kfree(p->chunks);
 252         p->chunks = NULL;
 253         p->nchunks = 0;
 254 put_ctx:
 255         amdgpu_ctx_put(p->ctx);
 256 free_chunk:
 257         kfree(chunk_array);
 258
 259         return ret;
 260 }
 261
 262 /* Convert microseconds to bytes. */
 263 static u64 us_to_bytes(struct amdgpu_device *adev, s64 us)
 264 {
 265         if (us <= 0 || !adev->mm_stats.log2_max_MBps)
 266                 return 0;
 267
 268         /* Since accum_us is incremented by a million per second, just
 269          * multiply it by the number of MB/s to get the number of bytes.
 270          */
 271         return us << adev->mm_stats.log2_max_MBps;
 272 }
 273
 274 static s64 bytes_to_us(struct amdgpu_device *adev, u64 bytes)
 275 {
 276         if (!adev->mm_stats.log2_max_MBps)
 277                 return 0;
 278
 279         return bytes >> adev->mm_stats.log2_max_MBps;
 280 }
 281
 282 /* Returns how many bytes TTM can move right now. If no bytes can be moved,
 283  * it returns 0. If it returns non-zero, it's OK to move at least one buffer,
 284  * which means it can go over the threshold once. If that happens, the driver
 285  * will be in debt and no other buffer migrations can be done until that debt
 286  * is repaid.
 287  *
 288  * This approach allows moving a buffer of any size (it's important to allow
 289  * that).
 290  *
 291  * The currency is simply time in microseconds and it increases as the clock
 292  * ticks. The accumulated microseconds (us) are converted to bytes and
 293  * returned.
 294  */
 295 static u64 amdgpu_cs_get_threshold_for_moves(struct amdgpu_device *adev)
 296 {
 297         s64 time_us, increment_us;
 298         u64 max_bytes;
 299         u64 free_vram, total_vram, used_vram;
 300
 301         /* Allow a maximum of 200 accumulated ms. This is basically per-IB
 302          * throttling.
 303          *
 304          * It means that in order to get full max MBps, at least 5 IBs per
 305          * second must be submitted and not more than 200ms apart from each
 306          * other.
 307          */
 308         const s64 us_upper_bound = 200000;
 309
 310         if (!adev->mm_stats.log2_max_MBps)
 311                 return 0;
 312
 313         total_vram = adev->mc.real_vram_size - adev->vram_pin_size;
 314         used_vram = atomic64_read(&adev->vram_usage);
 315         free_vram = used_vram >= total_vram ? 0 : total_vram - used_vram;
 316
 317         spin_lock(&adev->mm_stats.lock);
 318
 319         /* Increase the amount of accumulated us. */
 320         time_us = ktime_to_us(ktime_get());
 321         increment_us = time_us - adev->mm_stats.last_update_us;
 322         adev->mm_stats.last_update_us = time_us;
 323         adev->mm_stats.accum_us = min(adev->mm_stats.accum_us + increment_us,
 324                                       us_upper_bound);
 325
 326         /* This prevents the short period of low performance when the VRAM
 327          * usage is low and the driver is in debt or doesn't have enough
 328          * accumulated us to fill VRAM quickly.
 329          *
 330          * The situation can occur in these cases:
 331          * - a lot of VRAM is freed by userspace
 332          * - the presence of a big buffer causes a lot of evictions
 333          *   (solution: split buffers into smaller ones)
 334          *
 335          * If 128 MB or 1/8th of VRAM is free, start filling it now by setting
 336          * accum_us to a positive number.
 337          */
 338         if (free_vram >= 128 * 1024 * 1024 || free_vram >= total_vram / 8) {
 339                 s64 min_us;
 340
 341                 /* Be more aggresive on dGPUs. Try to fill a portion of free
 342                  * VRAM now.
 343                  */
 344                 if (!(adev->flags & AMD_IS_APU))
 345                         min_us = bytes_to_us(adev, free_vram / 4);
 346                 else
 347                         min_us = 0; /* Reset accum_us on APUs. */
 348
 349                 adev->mm_stats.accum_us = max(min_us, adev->mm_stats.accum_us);
 350         }
 351
 352         /* This returns 0 if the driver is in debt to disallow (optional)
 353          * buffer moves.
 354          */
 355         max_bytes = us_to_bytes(adev, adev->mm_stats.accum_us);
 356
 357         spin_unlock(&adev->mm_stats.lock);
 358         return max_bytes;
 359 }
 360
 361 /* Report how many bytes have really been moved for the last command
 362  * submission. This can result in a debt that can stop buffer migrations
 363  * temporarily.
 364  */
 365 void amdgpu_cs_report_moved_bytes(struct amdgpu_device *adev, u64 num_bytes)
 366 {
 367         spin_lock(&adev->mm_stats.lock);
 368         adev->mm_stats.accum_us -= bytes_to_us(adev, num_bytes);
 369         spin_unlock(&adev->mm_stats.lock);
 370 }
 371
 372 static int amdgpu_cs_bo_validate(struct amdgpu_cs_parser *p,
 373                                  struct amdgpu_bo *bo)
 374 {
 375         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
 376         u64 initial_bytes_moved;
 377         uint32_t domain;
 378         int r;
 379
 380         if (bo->pin_count)
 381                 return 0;
 382
 383         /* Don't move this buffer if we have depleted our allowance
 384          * to move it. Don't move anything if the threshold is zero.
 385          */
 386         if (p->bytes_moved < p->bytes_moved_threshold)
 387                 domain = bo->prefered_domains;
 388         else
 389                 domain = bo->allowed_domains;
 390
 391 retry:
 392         amdgpu_ttm_placement_from_domain(bo, domain);
 393         initial_bytes_moved = atomic64_read(&adev->num_bytes_moved);
 394         r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
 395         p->bytes_moved += atomic64_read(&adev->num_bytes_moved) -
 396                 initial_bytes_moved;
 397
 398         if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
 399                 domain = bo->allowed_domains;
 400                 goto retry;
 401         }
 402
 403         return r;
 404 }
 405
 406 /* Last resort, try to evict something from the current working set */
 407 static bool amdgpu_cs_try_evict(struct amdgpu_cs_parser *p,
 408                                 struct amdgpu_bo *validated)
 409 {
 410         uint32_t domain = validated->allowed_domains;
 411         int r;
 412
 413         if (!p->evictable)
 414                 return false;
 415
 416         for (;&p->evictable->tv.head != &p->validated;
 417              p->evictable = list_prev_entry(p->evictable, tv.head)) {
 418
 419                 struct amdgpu_bo_list_entry *candidate = p->evictable;
 420                 struct amdgpu_bo *bo = candidate->robj;
 421                 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
 422                 u64 initial_bytes_moved;
 423                 uint32_t other;
 424
 425                 /* If we reached our current BO we can forget it */
 426                 if (candidate->robj == validated)
 427                         break;
 428
 429                 other = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
 430
 431                 /* Check if this BO is in one of the domains we need space for */
 432                 if (!(other & domain))
 433                         continue;
 434
 435                 /* Check if we can move this BO somewhere else */
 436                 other = bo->allowed_domains & ~domain;
 437                 if (!other)
 438                         continue;
 439
 440                 /* Good we can try to move this BO somewhere else */
 441                 amdgpu_ttm_placement_from_domain(bo, other);
 442                 initial_bytes_moved = atomic64_read(&adev->num_bytes_moved);
 443                 r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
 444                 p->bytes_moved += atomic64_read(&adev->num_bytes_moved) -
 445                         initial_bytes_moved;
 446
 447                 if (unlikely(r))
 448                         break;
 449
 450                 p->evictable = list_prev_entry(p->evictable, tv.head);
 451                 list_move(&candidate->tv.head, &p->validated);
 452
 453                 return true;
 454         }
 455
 456         return false;
 457 }
 458
 459 static int amdgpu_cs_validate(void *param, struct amdgpu_bo *bo)
 460 {
 461         struct amdgpu_cs_parser *p = param;
 462         int r;
 463
 464         do {
 465                 r = amdgpu_cs_bo_validate(p, bo);
 466         } while (r == -ENOMEM && amdgpu_cs_try_evict(p, bo));
 467         if (r)
 468                 return r;
 469
 470         if (bo->shadow)
 471                 r = amdgpu_cs_bo_validate(p, bo->shadow);
 472
 473         return r;
 474 }
 475
 476 static int amdgpu_cs_list_validate(struct amdgpu_cs_parser *p,
 477                             struct list_head *validated)
 478 {
 479         struct amdgpu_bo_list_entry *lobj;
 480         int r;
 481
 482         list_for_each_entry(lobj, validated, tv.head) {
 483                 struct amdgpu_bo *bo = lobj->robj;
 484                 bool binding_userptr = false;
 485                 struct mm_struct *usermm;
 486
 487                 usermm = amdgpu_ttm_tt_get_usermm(bo->tbo.ttm);
 488                 if (usermm && usermm != current->mm)
 489                         return -EPERM;
 490
 491                 /* Check if we have user pages and nobody bound the BO already */
 492                 if (lobj->user_pages && bo->tbo.ttm->state != tt_bound) {
 493                         size_t size = sizeof(struct page *);
 494
 495                         size *= bo->tbo.ttm->num_pages;
 496                         memcpy(bo->tbo.ttm->pages, lobj->user_pages, size);
 497                         binding_userptr = true;
 498                 }
 499
 500                 if (p->evictable == lobj)
 501                         p->evictable = NULL;
 502
 503                 r = amdgpu_cs_validate(p, bo);
 504                 if (r)
 505                         return r;
 506
 507                 if (binding_userptr) {
 508                         kvfree(lobj->user_pages);
 509                         lobj->user_pages = NULL;
 510                 }
 511         }
 512         return 0;
 513 }
 514
 515 static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
 516                                 union drm_amdgpu_cs *cs)
 517 {
 518         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 519         struct amdgpu_bo_list_entry *e;
 520         struct list_head duplicates;
 521         bool need_mmap_lock = false;
 522         unsigned i, tries = 10;
 523         int r;
 524
 525         INIT_LIST_HEAD(&p->validated);
 526
 527         p->bo_list = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle);
 528         if (p->bo_list) {
 529                 need_mmap_lock = p->bo_list->first_userptr !=
 530                         p->bo_list->num_entries;
 531                 amdgpu_bo_list_get_list(p->bo_list, &p->validated);
 532         }
 533
 534         INIT_LIST_HEAD(&duplicates);
 535         amdgpu_vm_get_pd_bo(&fpriv->vm, &p->validated, &p->vm_pd);
 536
 537         if (p->uf_entry.robj)
 538                 list_add(&p->uf_entry.tv.head, &p->validated);
 539
 540         if (need_mmap_lock)
 541                 down_read(&current->mm->mmap_sem);
 542
 543         while (1) {
 544                 struct list_head need_pages;
 545                 unsigned i;
 546
 547                 r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true,
 548                                            &duplicates);
 549                 if (unlikely(r != 0)) {
 550                         if (r != -ERESTARTSYS)
 551                                 DRM_ERROR("ttm_eu_reserve_buffers failed.\n");
 552                         goto error_free_pages;
 553                 }
 554
 555                 /* Without a BO list we don't have userptr BOs */
 556                 if (!p->bo_list)
 557                         break;
 558
 559                 INIT_LIST_HEAD(&need_pages);
 560                 for (i = p->bo_list->first_userptr;
 561                      i < p->bo_list->num_entries; ++i) {
 562
 563                         e = &p->bo_list->array[i];
 564
 565                         if (amdgpu_ttm_tt_userptr_invalidated(e->robj->tbo.ttm,
 566                                  &e->user_invalidated) && e->user_pages) {
 567
 568                                 /* We acquired a page array, but somebody
 569                                  * invalidated it. Free it an try again
 570                                  */
 571                                 release_pages(e->user_pages,
 572                                               e->robj->tbo.ttm->num_pages,
 573                                               false);
 574                                 kvfree(e->user_pages);
 575                                 e->user_pages = NULL;
 576                         }
 577
 578                         if (e->robj->tbo.ttm->state != tt_bound &&
 579                             !e->user_pages) {
 580                                 list_del(&e->tv.head);
 581                                 list_add(&e->tv.head, &need_pages);
 582
 583                                 amdgpu_bo_unreserve(e->robj);
 584                         }
 585                 }
 586
 587                 if (list_empty(&need_pages))
 588                         break;
 589
 590                 /* Unreserve everything again. */
 591                 ttm_eu_backoff_reservation(&p->ticket, &p->validated);
 592
 593                 /* We tried too many times, just abort */
 594                 if (!--tries) {
 595                         r = -EDEADLK;
 596                         DRM_ERROR("deadlock in %s\n", __func__);
 597                         goto error_free_pages;
 598                 }
 599
 600                 /* Fill the page arrays for all useptrs. */
 601                 list_for_each_entry(e, &need_pages, tv.head) {
 602                         struct ttm_tt *ttm = e->robj->tbo.ttm;
 603
 604                         e->user_pages = kvmalloc_array(ttm->num_pages,
 605                                                          sizeof(struct page*),
 606                                                          GFP_KERNEL | __GFP_ZERO);
 607                         if (!e->user_pages) {
 608                                 r = -ENOMEM;
 609                                 DRM_ERROR("calloc failure in %s\n", __func__);
 610                                 goto error_free_pages;
 611                         }
 612
 613                         r = amdgpu_ttm_tt_get_user_pages(ttm, e->user_pages);
 614                         if (r) {
 615                                 DRM_ERROR("amdgpu_ttm_tt_get_user_pages failed.\n");
 616                                 kvfree(e->user_pages);
 617                                 e->user_pages = NULL;
 618                                 goto error_free_pages;
 619                         }
 620                 }
 621
 622                 /* And try again. */
 623                 list_splice(&need_pages, &p->validated);
 624         }
 625
 626         p->bytes_moved_threshold = amdgpu_cs_get_threshold_for_moves(p->adev);
 627         p->bytes_moved = 0;
 628         p->evictable = list_last_entry(&p->validated,
 629                                        struct amdgpu_bo_list_entry,
 630                                        tv.head);
 631
 632         r = amdgpu_vm_validate_pt_bos(p->adev, &fpriv->vm,
 633                                       amdgpu_cs_validate, p);
 634         if (r) {
 635                 DRM_ERROR("amdgpu_vm_validate_pt_bos() failed.\n");
 636                 goto error_validate;
 637         }
 638
 639         r = amdgpu_cs_list_validate(p, &duplicates);
 640         if (r) {
 641                 DRM_ERROR("amdgpu_cs_list_validate(duplicates) failed.\n");
 642                 goto error_validate;
 643         }
 644
 645         r = amdgpu_cs_list_validate(p, &p->validated);
 646         if (r) {
 647                 DRM_ERROR("amdgpu_cs_list_validate(validated) failed.\n");
 648                 goto error_validate;
 649         }
 650
 651         amdgpu_cs_report_moved_bytes(p->adev, p->bytes_moved);
 652
 653         fpriv->vm.last_eviction_counter =
 654                 atomic64_read(&p->adev->num_evictions);
 655
 656         if (p->bo_list) {
 657                 struct amdgpu_bo *gds = p->bo_list->gds_obj;
 658                 struct amdgpu_bo *gws = p->bo_list->gws_obj;
 659                 struct amdgpu_bo *oa = p->bo_list->oa_obj;
 660                 struct amdgpu_vm *vm = &fpriv->vm;
 661                 unsigned i;
 662
 663                 for (i = 0; i < p->bo_list->num_entries; i++) {
 664                         struct amdgpu_bo *bo = p->bo_list->array[i].robj;
 665
 666                         p->bo_list->array[i].bo_va = amdgpu_vm_bo_find(vm, bo);
 667                 }
 668
 669                 if (gds) {
 670                         p->job->gds_base = amdgpu_bo_gpu_offset(gds);
 671                         p->job->gds_size = amdgpu_bo_size(gds);
 672                 }
 673                 if (gws) {
 674                         p->job->gws_base = amdgpu_bo_gpu_offset(gws);
 675                         p->job->gws_size = amdgpu_bo_size(gws);
 676                 }
 677                 if (oa) {
 678                         p->job->oa_base = amdgpu_bo_gpu_offset(oa);
 679                         p->job->oa_size = amdgpu_bo_size(oa);
 680                 }
 681         }
 682
 683         if (!r && p->uf_entry.robj) {
 684                 struct amdgpu_bo *uf = p->uf_entry.robj;
 685
 686                 r = amdgpu_ttm_bind(&uf->tbo, &uf->tbo.mem);
 687                 p->job->uf_addr += amdgpu_bo_gpu_offset(uf);
 688         }
 689
 690 error_validate:
 691         if (r) {
 692                 amdgpu_vm_move_pt_bos_in_lru(p->adev, &fpriv->vm);
 693                 ttm_eu_backoff_reservation(&p->ticket, &p->validated);
 694         }
 695
 696 error_free_pages:
 697
 698         if (need_mmap_lock)
 699                 up_read(&current->mm->mmap_sem);
 700
 701         if (p->bo_list) {
 702                 for (i = p->bo_list->first_userptr;
 703                      i < p->bo_list->num_entries; ++i) {
 704                         e = &p->bo_list->array[i];
 705
 706                         if (!e->user_pages)
 707                                 continue;
 708
 709                         release_pages(e->user_pages,
 710                                       e->robj->tbo.ttm->num_pages,
 711                                       false);
 712                         kvfree(e->user_pages);
 713                 }
 714         }
 715
 716         return r;
 717 }
 718
 719 static int amdgpu_cs_sync_rings(struct amdgpu_cs_parser *p)
 720 {
 721         struct amdgpu_bo_list_entry *e;
 722         int r;
 723
 724         list_for_each_entry(e, &p->validated, tv.head) {
 725                 struct reservation_object *resv = e->robj->tbo.resv;
 726                 r = amdgpu_sync_resv(p->adev, &p->job->sync, resv, p->filp);
 727
 728                 if (r)
 729                         return r;
 730         }
 731         return 0;
 732 }
 733
 734 /**
 735  * cs_parser_fini() - clean parser states
 736  * @parser:     parser structure holding parsing context.
 737  * @error:      error number
 738  *
 739  * If error is set than unvalidate buffer, otherwise just free memory
 740  * used by parsing context.
 741  **/
 742 static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error, bool backoff)
 743 {
 744         struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
 745         unsigned i;
 746
 747         if (!error) {
 748                 amdgpu_vm_move_pt_bos_in_lru(parser->adev, &fpriv->vm);
 749
 750                 ttm_eu_fence_buffer_objects(&parser->ticket,
 751                                             &parser->validated,
 752                                             parser->fence);
 753         } else if (backoff) {
 754                 ttm_eu_backoff_reservation(&parser->ticket,
 755                                            &parser->validated);
 756         }
 757         dma_fence_put(parser->fence);
 758
 759         if (parser->ctx)
 760                 amdgpu_ctx_put(parser->ctx);
 761         if (parser->bo_list)
 762                 amdgpu_bo_list_put(parser->bo_list);
 763
 764         for (i = 0; i < parser->nchunks; i++)
 765                 kvfree(parser->chunks[i].kdata);
 766         kfree(parser->chunks);
 767         if (parser->job)
 768                 amdgpu_job_free(parser->job);
 769         amdgpu_bo_unref(&parser->uf_entry.robj);
 770 }
 771
 772 static int amdgpu_bo_vm_update_pte(struct amdgpu_cs_parser *p)
 773 {
 774         struct amdgpu_device *adev = p->adev;
 775         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 776         struct amdgpu_vm *vm = &fpriv->vm;
 777         struct amdgpu_bo_va *bo_va;
 778         struct amdgpu_bo *bo;
 779         int i, r;
 780
 781         r = amdgpu_vm_update_directories(adev, vm);
 782         if (r)
 783                 return r;
 784
 785         r = amdgpu_sync_fence(adev, &p->job->sync, vm->last_dir_update);
 786         if (r)
 787                 return r;
 788
 789         r = amdgpu_vm_clear_freed(adev, vm, NULL);
 790         if (r)
 791                 return r;
 792
 793         r = amdgpu_vm_bo_update(adev, fpriv->prt_va, false);
 794         if (r)
 795                 return r;
 796
 797         r = amdgpu_sync_fence(adev, &p->job->sync,
 798                               fpriv->prt_va->last_pt_update);
 799         if (r)
 800                 return r;
 801
 802         if (amdgpu_sriov_vf(adev)) {
 803                 struct dma_fence *f;
 804                 bo_va = vm->csa_bo_va;
 805                 BUG_ON(!bo_va);
 806                 r = amdgpu_vm_bo_update(adev, bo_va, false);
 807                 if (r)
 808                         return r;
 809
 810                 f = bo_va->last_pt_update;
 811                 r = amdgpu_sync_fence(adev, &p->job->sync, f);
 812                 if (r)
 813                         return r;
 814         }
 815
 816         if (p->bo_list) {
 817                 for (i = 0; i < p->bo_list->num_entries; i++) {
 818                         struct dma_fence *f;
 819
 820                         /* ignore duplicates */
 821                         bo = p->bo_list->array[i].robj;
 822                         if (!bo)
 823                                 continue;
 824
 825                         bo_va = p->bo_list->array[i].bo_va;
 826                         if (bo_va == NULL)
 827                                 continue;
 828
 829                         r = amdgpu_vm_bo_update(adev, bo_va, false);
 830                         if (r)
 831                                 return r;
 832
 833                         f = bo_va->last_pt_update;
 834                         r = amdgpu_sync_fence(adev, &p->job->sync, f);
 835                         if (r)
 836                                 return r;
 837                 }
 838
 839         }
 840
 841         r = amdgpu_vm_clear_invalids(adev, vm, &p->job->sync);
 842
 843         if (amdgpu_vm_debug && p->bo_list) {
 844                 /* Invalidate all BOs to test for userspace bugs */
 845                 for (i = 0; i < p->bo_list->num_entries; i++) {
 846                         /* ignore duplicates */
 847                         bo = p->bo_list->array[i].robj;
 848                         if (!bo)
 849                                 continue;
 850
 851                         amdgpu_vm_bo_invalidate(adev, bo);
 852                 }
 853         }
 854
 855         return r;
 856 }
 857
 858 static int amdgpu_cs_ib_vm_chunk(struct amdgpu_device *adev,
 859                                  struct amdgpu_cs_parser *p)
 860 {
 861         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 862         struct amdgpu_vm *vm = &fpriv->vm;
 863         struct amdgpu_ring *ring = p->job->ring;
 864         int i, r;
 865
 866         /* Only for UVD/VCE VM emulation */
 867         if (ring->funcs->parse_cs) {
 868                 for (i = 0; i < p->job->num_ibs; i++) {
 869                         r = amdgpu_ring_parse_cs(ring, p, i);
 870                         if (r)
 871                                 return r;
 872                 }
 873         }
 874
 875         if (p->job->vm) {
 876                 p->job->vm_pd_addr = amdgpu_bo_gpu_offset(vm->root.bo);
 877
 878                 r = amdgpu_bo_vm_update_pte(p);
 879                 if (r)
 880                         return r;
 881         }
 882
 883         return amdgpu_cs_sync_rings(p);
 884 }
 885
 886 static int amdgpu_cs_ib_fill(struct amdgpu_device *adev,
 887                              struct amdgpu_cs_parser *parser)
 888 {
 889         struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
 890         struct amdgpu_vm *vm = &fpriv->vm;
 891         int i, j;
 892         int r, ce_preempt = 0, de_preempt = 0;
 893
 894         for (i = 0, j = 0; i < parser->nchunks && j < parser->job->num_ibs; i++) {
 895                 struct amdgpu_cs_chunk *chunk;
 896                 struct amdgpu_ib *ib;
 897                 struct drm_amdgpu_cs_chunk_ib *chunk_ib;
 898                 struct amdgpu_ring *ring;
 899
 900                 chunk = &parser->chunks[i];
 901                 ib = &parser->job->ibs[j];
 902                 chunk_ib = (struct drm_amdgpu_cs_chunk_ib *)chunk->kdata;
 903
 904                 if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
 905                         continue;
 906
 907                 if (chunk_ib->ip_type == AMDGPU_HW_IP_GFX && amdgpu_sriov_vf(adev)) {
 908                         if (chunk_ib->flags & AMDGPU_IB_FLAG_PREEMPT) {
 909                                 if (chunk_ib->flags & AMDGPU_IB_FLAG_CE)
 910                                         ce_preempt++;
 911                                 else
 912                                         de_preempt++;
 913                         }
 914
 915                         /* each GFX command submit allows 0 or 1 IB preemptible for CE & DE */
 916                         if (ce_preempt > 1 || de_preempt > 1)
 917                                 return -EINVAL;
 918                 }
 919
 920                 r = amdgpu_cs_get_ring(adev, chunk_ib->ip_type,
 921                                        chunk_ib->ip_instance, chunk_ib->ring,
 922                                        &ring);
 923                 if (r)
 924                         return r;
 925
 926                 if (chunk_ib->flags & AMDGPU_IB_FLAG_PREAMBLE) {
 927                         parser->job->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT;
 928                         if (!parser->ctx->preamble_presented) {
 929                                 parser->job->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT_FIRST;
 930                                 parser->ctx->preamble_presented = true;
 931                         }
 932                 }
 933
 934                 if (parser->job->ring && parser->job->ring != ring)
 935                         return -EINVAL;
 936
 937                 parser->job->ring = ring;
 938
 939                 if (ring->funcs->parse_cs) {
 940                         struct amdgpu_bo_va_mapping *m;
 941                         struct amdgpu_bo *aobj = NULL;
 942                         uint64_t offset;
 943                         uint8_t *kptr;
 944
 945                         m = amdgpu_cs_find_mapping(parser, chunk_ib->va_start,
 946                                                    &aobj);
 947                         if (!aobj) {
 948                                 DRM_ERROR("IB va_start is invalid\n");
 949                                 return -EINVAL;
 950                         }
 951
 952                         if ((chunk_ib->va_start + chunk_ib->ib_bytes) >
 953                             (m->last + 1) * AMDGPU_GPU_PAGE_SIZE) {
 954                                 DRM_ERROR("IB va_start+ib_bytes is invalid\n");
 955                                 return -EINVAL;
 956                         }
 957
 958                         /* the IB should be reserved at this point */
 959                         r = amdgpu_bo_kmap(aobj, (void **)&kptr);
 960                         if (r) {
 961                                 return r;
 962                         }
 963
 964                         offset = m->start * AMDGPU_GPU_PAGE_SIZE;
 965                         kptr += chunk_ib->va_start - offset;
 966
 967                         r =  amdgpu_ib_get(adev, vm, chunk_ib->ib_bytes, ib);
 968                         if (r) {
 969                                 DRM_ERROR("Failed to get ib !\n");
 970                                 return r;
 971                         }
 972
 973                         memcpy(ib->ptr, kptr, chunk_ib->ib_bytes);
 974                         amdgpu_bo_kunmap(aobj);
 975                 } else {
 976                         r =  amdgpu_ib_get(adev, vm, 0, ib);
 977                         if (r) {
 978                                 DRM_ERROR("Failed to get ib !\n");
 979                                 return r;
 980                         }
 981
 982                 }
 983
 984                 ib->gpu_addr = chunk_ib->va_start;
 985                 ib->length_dw = chunk_ib->ib_bytes / 4;
 986                 ib->flags = chunk_ib->flags;
 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 0;
 997 }
 998
 999 static int amdgpu_cs_dependencies(struct amdgpu_device *adev,
1000                                   struct amdgpu_cs_parser *p)
1001 {
1002         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
1003         int i, j, r;
1004
1005         for (i = 0; i < p->nchunks; ++i) {
1006                 struct drm_amdgpu_cs_chunk_dep *deps;
1007                 struct amdgpu_cs_chunk *chunk;
1008                 unsigned num_deps;
1009
1010                 chunk = &p->chunks[i];
1011
1012                 if (chunk->chunk_id != AMDGPU_CHUNK_ID_DEPENDENCIES)
1013                         continue;
1014
1015                 deps = (struct drm_amdgpu_cs_chunk_dep *)chunk->kdata;
1016                 num_deps = chunk->length_dw * 4 /
1017                         sizeof(struct drm_amdgpu_cs_chunk_dep);
1018
1019                 for (j = 0; j < num_deps; ++j) {
1020                         struct amdgpu_ring *ring;
1021                         struct amdgpu_ctx *ctx;
1022                         struct dma_fence *fence;
1023
1024                         r = amdgpu_cs_get_ring(adev, deps[j].ip_type,
1025                                                deps[j].ip_instance,
1026                                                deps[j].ring, &ring);
1027                         if (r)
1028                                 return r;
1029
1030                         ctx = amdgpu_ctx_get(fpriv, deps[j].ctx_id);
1031                         if (ctx == NULL)
1032                                 return -EINVAL;
1033
1034                         fence = amdgpu_ctx_get_fence(ctx, ring,
1035                                                      deps[j].handle);
1036                         if (IS_ERR(fence)) {
1037                                 r = PTR_ERR(fence);
1038                                 amdgpu_ctx_put(ctx);
1039                                 return r;
1040
1041                         } else if (fence) {
1042                                 r = amdgpu_sync_fence(adev, &p->job->sync,
1043                                                       fence);
1044                                 dma_fence_put(fence);
1045                                 amdgpu_ctx_put(ctx);
1046                                 if (r)
1047                                         return r;
1048                         }
1049                 }
1050         }
1051
1052         return 0;
1053 }
1054
1055 static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
1056                             union drm_amdgpu_cs *cs)
1057 {
1058         struct amdgpu_ring *ring = p->job->ring;
1059         struct amd_sched_entity *entity = &p->ctx->rings[ring->idx].entity;
1060         struct amdgpu_job *job;
1061         int r;
1062
1063         job = p->job;
1064         p->job = NULL;
1065
1066         r = amd_sched_job_init(&job->base, &ring->sched, entity, p->filp);
1067         if (r) {
1068                 amdgpu_job_free(job);
1069                 return r;
1070         }
1071
1072         job->owner = p->filp;
1073         job->fence_ctx = entity->fence_context;
1074         p->fence = dma_fence_get(&job->base.s_fence->finished);
1075         cs->out.handle = amdgpu_ctx_add_fence(p->ctx, ring, p->fence);
1076         job->uf_sequence = cs->out.handle;
1077         amdgpu_job_free_resources(job);
1078         amdgpu_cs_parser_fini(p, 0, true);
1079
1080         trace_amdgpu_cs_ioctl(job);
1081         amd_sched_entity_push_job(&job->base);
1082
1083         return 0;
1084 }
1085
1086 int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
1087 {
1088         struct amdgpu_device *adev = dev->dev_private;
1089         union drm_amdgpu_cs *cs = data;
1090         struct amdgpu_cs_parser parser = {};
1091         bool reserved_buffers = false;
1092         int i, r;
1093
1094         if (!adev->accel_working)
1095                 return -EBUSY;
1096
1097         parser.adev = adev;
1098         parser.filp = filp;
1099
1100         r = amdgpu_cs_parser_init(&parser, data);
1101         if (r) {
1102                 DRM_ERROR("Failed to initialize parser !\n");
1103                 goto out;
1104         }
1105
1106         r = amdgpu_cs_parser_bos(&parser, data);
1107         if (r) {
1108                 if (r == -ENOMEM)
1109                         DRM_ERROR("Not enough memory for command submission!\n");
1110                 else if (r != -ERESTARTSYS)
1111                         DRM_ERROR("Failed to process the buffer list %d!\n", r);
1112                 goto out;
1113         }
1114
1115         reserved_buffers = true;
1116         r = amdgpu_cs_ib_fill(adev, &parser);
1117         if (r)
1118                 goto out;
1119
1120         r = amdgpu_cs_dependencies(adev, &parser);
1121         if (r) {
1122                 DRM_ERROR("Failed in the dependencies handling %d!\n", r);
1123                 goto out;
1124         }
1125
1126         for (i = 0; i < parser.job->num_ibs; i++)
1127                 trace_amdgpu_cs(&parser, i);
1128
1129         r = amdgpu_cs_ib_vm_chunk(adev, &parser);
1130         if (r)
1131                 goto out;
1132
1133         r = amdgpu_cs_submit(&parser, cs);
1134         if (r)
1135                 goto out;
1136
1137         return 0;
1138 out:
1139         amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
1140         return r;
1141 }
1142
1143 /**
1144  * amdgpu_cs_wait_ioctl - wait for a command submission to finish
1145  *
1146  * @dev: drm device
1147  * @data: data from userspace
1148  * @filp: file private
1149  *
1150  * Wait for the command submission identified by handle to finish.
1151  */
1152 int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data,
1153                          struct drm_file *filp)
1154 {
1155         union drm_amdgpu_wait_cs *wait = data;
1156         struct amdgpu_device *adev = dev->dev_private;
1157         unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
1158         struct amdgpu_ring *ring = NULL;
1159         struct amdgpu_ctx *ctx;
1160         struct dma_fence *fence;
1161         long r;
1162
1163         r = amdgpu_cs_get_ring(adev, wait->in.ip_type, wait->in.ip_instance,
1164                                wait->in.ring, &ring);
1165         if (r)
1166                 return r;
1167
1168         ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
1169         if (ctx == NULL)
1170                 return -EINVAL;
1171
1172         fence = amdgpu_ctx_get_fence(ctx, ring, wait->in.handle);
1173         if (IS_ERR(fence))
1174                 r = PTR_ERR(fence);
1175         else if (fence) {
1176                 r = dma_fence_wait_timeout(fence, true, timeout);
1177                 dma_fence_put(fence);
1178         } else
1179                 r = 1;
1180
1181         amdgpu_ctx_put(ctx);
1182         if (r < 0)
1183                 return r;
1184
1185         memset(wait, 0, sizeof(*wait));
1186         wait->out.status = (r == 0);
1187
1188         return 0;
1189 }
1190
1191 /**
1192  * amdgpu_cs_get_fence - helper to get fence from drm_amdgpu_fence
1193  *
1194  * @adev: amdgpu device
1195  * @filp: file private
1196  * @user: drm_amdgpu_fence copied from user space
1197  */
1198 static struct dma_fence *amdgpu_cs_get_fence(struct amdgpu_device *adev,
1199                                              struct drm_file *filp,
1200                                              struct drm_amdgpu_fence *user)
1201 {
1202         struct amdgpu_ring *ring;
1203         struct amdgpu_ctx *ctx;
1204         struct dma_fence *fence;
1205         int r;
1206
1207         r = amdgpu_cs_get_ring(adev, user->ip_type, user->ip_instance,
1208                                user->ring, &ring);
1209         if (r)
1210                 return ERR_PTR(r);
1211
1212         ctx = amdgpu_ctx_get(filp->driver_priv, user->ctx_id);
1213         if (ctx == NULL)
1214                 return ERR_PTR(-EINVAL);
1215
1216         fence = amdgpu_ctx_get_fence(ctx, ring, user->seq_no);
1217         amdgpu_ctx_put(ctx);
1218
1219         return fence;
1220 }
1221
1222 /**
1223  * amdgpu_cs_wait_all_fence - wait on all fences to signal
1224  *
1225  * @adev: amdgpu device
1226  * @filp: file private
1227  * @wait: wait parameters
1228  * @fences: array of drm_amdgpu_fence
1229  */
1230 static int amdgpu_cs_wait_all_fences(struct amdgpu_device *adev,
1231                                      struct drm_file *filp,
1232                                      union drm_amdgpu_wait_fences *wait,
1233                                      struct drm_amdgpu_fence *fences)
1234 {
1235         uint32_t fence_count = wait->in.fence_count;
1236         unsigned int i;
1237         long r = 1;
1238
1239         for (i = 0; i < fence_count; i++) {
1240                 struct dma_fence *fence;
1241                 unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1242
1243                 fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1244                 if (IS_ERR(fence))
1245                         return PTR_ERR(fence);
1246                 else if (!fence)
1247                         continue;
1248
1249                 r = dma_fence_wait_timeout(fence, true, timeout);
1250                 dma_fence_put(fence);
1251                 if (r < 0)
1252                         return r;
1253
1254                 if (r == 0)
1255                         break;
1256         }
1257
1258         memset(wait, 0, sizeof(*wait));
1259         wait->out.status = (r > 0);
1260
1261         return 0;
1262 }
1263
1264 /**
1265  * amdgpu_cs_wait_any_fence - wait on any fence to signal
1266  *
1267  * @adev: amdgpu device
1268  * @filp: file private
1269  * @wait: wait parameters
1270  * @fences: array of drm_amdgpu_fence
1271  */
1272 static int amdgpu_cs_wait_any_fence(struct amdgpu_device *adev,
1273                                     struct drm_file *filp,
1274                                     union drm_amdgpu_wait_fences *wait,
1275                                     struct drm_amdgpu_fence *fences)
1276 {
1277         unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1278         uint32_t fence_count = wait->in.fence_count;
1279         uint32_t first = ~0;
1280         struct dma_fence **array;
1281         unsigned int i;
1282         long r;
1283
1284         /* Prepare the fence array */
1285         array = kcalloc(fence_count, sizeof(struct dma_fence *), GFP_KERNEL);
1286
1287         if (array == NULL)
1288                 return -ENOMEM;
1289
1290         for (i = 0; i < fence_count; i++) {
1291                 struct dma_fence *fence;
1292
1293                 fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1294                 if (IS_ERR(fence)) {
1295                         r = PTR_ERR(fence);
1296                         goto err_free_fence_array;
1297                 } else if (fence) {
1298                         array[i] = fence;
1299                 } else { /* NULL, the fence has been already signaled */
1300                         r = 1;
1301                         goto out;
1302                 }
1303         }
1304
1305         r = dma_fence_wait_any_timeout(array, fence_count, true, timeout,
1306                                        &first);
1307         if (r < 0)
1308                 goto err_free_fence_array;
1309
1310 out:
1311         memset(wait, 0, sizeof(*wait));
1312         wait->out.status = (r > 0);
1313         wait->out.first_signaled = first;
1314         /* set return value 0 to indicate success */
1315         r = 0;
1316
1317 err_free_fence_array:
1318         for (i = 0; i < fence_count; i++)
1319                 dma_fence_put(array[i]);
1320         kfree(array);
1321
1322         return r;
1323 }
1324
1325 /**
1326  * amdgpu_cs_wait_fences_ioctl - wait for multiple command submissions to finish
1327  *
1328  * @dev: drm device
1329  * @data: data from userspace
1330  * @filp: file private
1331  */
1332 int amdgpu_cs_wait_fences_ioctl(struct drm_device *dev, void *data,
1333                                 struct drm_file *filp)
1334 {
1335         struct amdgpu_device *adev = dev->dev_private;
1336         union drm_amdgpu_wait_fences *wait = data;
1337         uint32_t fence_count = wait->in.fence_count;
1338         struct drm_amdgpu_fence *fences_user;
1339         struct drm_amdgpu_fence *fences;
1340         int r;
1341
1342         /* Get the fences from userspace */
1343         fences = kmalloc_array(fence_count, sizeof(struct drm_amdgpu_fence),
1344                         GFP_KERNEL);
1345         if (fences == NULL)
1346                 return -ENOMEM;
1347
1348         fences_user = (void __user *)(uintptr_t)(wait->in.fences);
1349         if (copy_from_user(fences, fences_user,
1350                 sizeof(struct drm_amdgpu_fence) * fence_count)) {
1351                 r = -EFAULT;
1352                 goto err_free_fences;
1353         }
1354
1355         if (wait->in.wait_all)
1356                 r = amdgpu_cs_wait_all_fences(adev, filp, wait, fences);
1357         else
1358                 r = amdgpu_cs_wait_any_fence(adev, filp, wait, fences);
1359
1360 err_free_fences:
1361         kfree(fences);
1362
1363         return r;
1364 }
1365
1366 /**
1367  * amdgpu_cs_find_bo_va - find bo_va for VM address
1368  *
1369  * @parser: command submission parser context
1370  * @addr: VM address
1371  * @bo: resulting BO of the mapping found
1372  *
1373  * Search the buffer objects in the command submission context for a certain
1374  * virtual memory address. Returns allocation structure when found, NULL
1375  * otherwise.
1376  */
1377 struct amdgpu_bo_va_mapping *
1378 amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
1379                        uint64_t addr, struct amdgpu_bo **bo)
1380 {
1381         struct amdgpu_bo_va_mapping *mapping;
1382         unsigned i;
1383
1384         if (!parser->bo_list)
1385                 return NULL;
1386
1387         addr /= AMDGPU_GPU_PAGE_SIZE;
1388
1389         for (i = 0; i < parser->bo_list->num_entries; i++) {
1390                 struct amdgpu_bo_list_entry *lobj;
1391
1392                 lobj = &parser->bo_list->array[i];
1393                 if (!lobj->bo_va)
1394                         continue;
1395
1396                 list_for_each_entry(mapping, &lobj->bo_va->valids, list) {
1397                         if (mapping->start > addr ||
1398                             addr > mapping->last)
1399                                 continue;
1400
1401                         *bo = lobj->bo_va->bo;
1402                         return mapping;
1403                 }
1404
1405                 list_for_each_entry(mapping, &lobj->bo_va->invalids, list) {
1406                         if (mapping->start > addr ||
1407                             addr > mapping->last)
1408                                 continue;
1409
1410                         *bo = lobj->bo_va->bo;
1411                         return mapping;
1412                 }
1413         }
1414
1415         return NULL;
1416 }
1417
1418 /**
1419  * amdgpu_cs_sysvm_access_required - make BOs accessible by the system VM
1420  *
1421  * @parser: command submission parser context
1422  *
1423  * Helper for UVD/VCE VM emulation, make sure BOs are accessible by the system VM.
1424  */
1425 int amdgpu_cs_sysvm_access_required(struct amdgpu_cs_parser *parser)
1426 {
1427         unsigned i;
1428         int r;
1429
1430         if (!parser->bo_list)
1431                 return 0;
1432
1433         for (i = 0; i < parser->bo_list->num_entries; i++) {
1434                 struct amdgpu_bo *bo = parser->bo_list->array[i].robj;
1435
1436                 r = amdgpu_ttm_bind(&bo->tbo, &bo->tbo.mem);
1437                 if (unlikely(r))
1438                         return r;
1439
1440                 if (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)
1441                         continue;
1442
1443                 bo->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
1444                 amdgpu_ttm_placement_from_domain(bo, bo->allowed_domains);
1445                 r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
1446                 if (unlikely(r))
1447                         return r;
1448         }
1449
1450         return 0;
1451 }