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

   1 /*
   2  * Copyright 2008 Jerome Glisse.
   3  * All Rights Reserved.
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the "Software"),
   7  * to deal in the Software without restriction, including without limitation
   8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   9  * and/or sell copies of the Software, and to permit persons to whom the
  10  * Software is furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice (including the next
  13  * paragraph) shall be included in all copies or substantial portions of the
  14  * Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  19  * PRECISION INSIGHT AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
  20  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  21  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
  22  * DEALINGS IN THE SOFTWARE.
  23  *
  24  * Authors:
  25  *    Jerome Glisse <glisse@freedesktop.org>
  26  */
  27 #include <linux/pagemap.h>
  28 #include <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 < 2){
  79                         *out_ring = &adev->vce.ring[ring];
  80                 } else {
  81                         DRM_ERROR("only two VCE rings are supported\n");
  82                         return -EINVAL;
  83                 }
  84                 break;
  85         }
  86         return 0;
  87 }
  88
  89 static int amdgpu_cs_user_fence_chunk(struct amdgpu_cs_parser *p,
  90                                       struct drm_amdgpu_cs_chunk_fence *data,
  91                                       uint32_t *offset)
  92 {
  93         struct drm_gem_object *gobj;
  94         unsigned long size;
  95
  96         gobj = drm_gem_object_lookup(p->filp, data->handle);
  97         if (gobj == NULL)
  98                 return -EINVAL;
  99
 100         p->uf_entry.robj = amdgpu_bo_ref(gem_to_amdgpu_bo(gobj));
 101         p->uf_entry.priority = 0;
 102         p->uf_entry.tv.bo = &p->uf_entry.robj->tbo;
 103         p->uf_entry.tv.shared = true;
 104         p->uf_entry.user_pages = NULL;
 105
 106         size = amdgpu_bo_size(p->uf_entry.robj);
 107         if (size != PAGE_SIZE || (data->offset + 8) > size)
 108                 return -EINVAL;
 109
 110         *offset = data->offset;
 111
 112         drm_gem_object_unreference_unlocked(gobj);
 113
 114         if (amdgpu_ttm_tt_get_usermm(p->uf_entry.robj->tbo.ttm)) {
 115                 amdgpu_bo_unref(&p->uf_entry.robj);
 116                 return -EINVAL;
 117         }
 118
 119         return 0;
 120 }
 121
 122 int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data)
 123 {
 124         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 125         struct amdgpu_vm *vm = &fpriv->vm;
 126         union drm_amdgpu_cs *cs = data;
 127         uint64_t *chunk_array_user;
 128         uint64_t *chunk_array;
 129         unsigned size, num_ibs = 0;
 130         uint32_t uf_offset = 0;
 131         int i;
 132         int ret;
 133
 134         if (cs->in.num_chunks == 0)
 135                 return 0;
 136
 137         chunk_array = kmalloc_array(cs->in.num_chunks, sizeof(uint64_t), GFP_KERNEL);
 138         if (!chunk_array)
 139                 return -ENOMEM;
 140
 141         p->ctx = amdgpu_ctx_get(fpriv, cs->in.ctx_id);
 142         if (!p->ctx) {
 143                 ret = -EINVAL;
 144                 goto free_chunk;
 145         }
 146
 147         /* get chunks */
 148         chunk_array_user = (uint64_t __user *)(unsigned long)(cs->in.chunks);
 149         if (copy_from_user(chunk_array, chunk_array_user,
 150                            sizeof(uint64_t)*cs->in.num_chunks)) {
 151                 ret = -EFAULT;
 152                 goto put_ctx;
 153         }
 154
 155         p->nchunks = cs->in.num_chunks;
 156         p->chunks = kmalloc_array(p->nchunks, sizeof(struct amdgpu_cs_chunk),
 157                             GFP_KERNEL);
 158         if (!p->chunks) {
 159                 ret = -ENOMEM;
 160                 goto put_ctx;
 161         }
 162
 163         for (i = 0; i < p->nchunks; i++) {
 164                 struct drm_amdgpu_cs_chunk __user **chunk_ptr = NULL;
 165                 struct drm_amdgpu_cs_chunk user_chunk;
 166                 uint32_t __user *cdata;
 167
 168                 chunk_ptr = (void __user *)(unsigned long)chunk_array[i];
 169                 if (copy_from_user(&user_chunk, chunk_ptr,
 170                                        sizeof(struct drm_amdgpu_cs_chunk))) {
 171                         ret = -EFAULT;
 172                         i--;
 173                         goto free_partial_kdata;
 174                 }
 175                 p->chunks[i].chunk_id = user_chunk.chunk_id;
 176                 p->chunks[i].length_dw = user_chunk.length_dw;
 177
 178                 size = p->chunks[i].length_dw;
 179                 cdata = (void __user *)(unsigned long)user_chunk.chunk_data;
 180
 181                 p->chunks[i].kdata = drm_malloc_ab(size, sizeof(uint32_t));
 182                 if (p->chunks[i].kdata == NULL) {
 183                         ret = -ENOMEM;
 184                         i--;
 185                         goto free_partial_kdata;
 186                 }
 187                 size *= sizeof(uint32_t);
 188                 if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
 189                         ret = -EFAULT;
 190                         goto free_partial_kdata;
 191                 }
 192
 193                 switch (p->chunks[i].chunk_id) {
 194                 case AMDGPU_CHUNK_ID_IB:
 195                         ++num_ibs;
 196                         break;
 197
 198                 case AMDGPU_CHUNK_ID_FENCE:
 199                         size = sizeof(struct drm_amdgpu_cs_chunk_fence);
 200                         if (p->chunks[i].length_dw * sizeof(uint32_t) < size) {
 201                                 ret = -EINVAL;
 202                                 goto free_partial_kdata;
 203                         }
 204
 205                         ret = amdgpu_cs_user_fence_chunk(p, p->chunks[i].kdata,
 206                                                          &uf_offset);
 207                         if (ret)
 208                                 goto free_partial_kdata;
 209
 210                         break;
 211
 212                 case AMDGPU_CHUNK_ID_DEPENDENCIES:
 213                         break;
 214
 215                 default:
 216                         ret = -EINVAL;
 217                         goto free_partial_kdata;
 218                 }
 219         }
 220
 221         ret = amdgpu_job_alloc(p->adev, num_ibs, &p->job, vm);
 222         if (ret)
 223                 goto free_all_kdata;
 224
 225         if (p->uf_entry.robj)
 226                 p->job->uf_addr = uf_offset;
 227         kfree(chunk_array);
 228         return 0;
 229
 230 free_all_kdata:
 231         i = p->nchunks - 1;
 232 free_partial_kdata:
 233         for (; i >= 0; i--)
 234                 drm_free_large(p->chunks[i].kdata);
 235         kfree(p->chunks);
 236 put_ctx:
 237         amdgpu_ctx_put(p->ctx);
 238 free_chunk:
 239         kfree(chunk_array);
 240
 241         return ret;
 242 }
 243
 244 /* Convert microseconds to bytes. */
 245 static u64 us_to_bytes(struct amdgpu_device *adev, s64 us)
 246 {
 247         if (us <= 0 || !adev->mm_stats.log2_max_MBps)
 248                 return 0;
 249
 250         /* Since accum_us is incremented by a million per second, just
 251          * multiply it by the number of MB/s to get the number of bytes.
 252          */
 253         return us << adev->mm_stats.log2_max_MBps;
 254 }
 255
 256 static s64 bytes_to_us(struct amdgpu_device *adev, u64 bytes)
 257 {
 258         if (!adev->mm_stats.log2_max_MBps)
 259                 return 0;
 260
 261         return bytes >> adev->mm_stats.log2_max_MBps;
 262 }
 263
 264 /* Returns how many bytes TTM can move right now. If no bytes can be moved,
 265  * it returns 0. If it returns non-zero, it's OK to move at least one buffer,
 266  * which means it can go over the threshold once. If that happens, the driver
 267  * will be in debt and no other buffer migrations can be done until that debt
 268  * is repaid.
 269  *
 270  * This approach allows moving a buffer of any size (it's important to allow
 271  * that).
 272  *
 273  * The currency is simply time in microseconds and it increases as the clock
 274  * ticks. The accumulated microseconds (us) are converted to bytes and
 275  * returned.
 276  */
 277 static u64 amdgpu_cs_get_threshold_for_moves(struct amdgpu_device *adev)
 278 {
 279         s64 time_us, increment_us;
 280         u64 max_bytes;
 281         u64 free_vram, total_vram, used_vram;
 282
 283         /* Allow a maximum of 200 accumulated ms. This is basically per-IB
 284          * throttling.
 285          *
 286          * It means that in order to get full max MBps, at least 5 IBs per
 287          * second must be submitted and not more than 200ms apart from each
 288          * other.
 289          */
 290         const s64 us_upper_bound = 200000;
 291
 292         if (!adev->mm_stats.log2_max_MBps)
 293                 return 0;
 294
 295         total_vram = adev->mc.real_vram_size - adev->vram_pin_size;
 296         used_vram = atomic64_read(&adev->vram_usage);
 297         free_vram = used_vram >= total_vram ? 0 : total_vram - used_vram;
 298
 299         spin_lock(&adev->mm_stats.lock);
 300
 301         /* Increase the amount of accumulated us. */
 302         time_us = ktime_to_us(ktime_get());
 303         increment_us = time_us - adev->mm_stats.last_update_us;
 304         adev->mm_stats.last_update_us = time_us;
 305         adev->mm_stats.accum_us = min(adev->mm_stats.accum_us + increment_us,
 306                                       us_upper_bound);
 307
 308         /* This prevents the short period of low performance when the VRAM
 309          * usage is low and the driver is in debt or doesn't have enough
 310          * accumulated us to fill VRAM quickly.
 311          *
 312          * The situation can occur in these cases:
 313          * - a lot of VRAM is freed by userspace
 314          * - the presence of a big buffer causes a lot of evictions
 315          *   (solution: split buffers into smaller ones)
 316          *
 317          * If 128 MB or 1/8th of VRAM is free, start filling it now by setting
 318          * accum_us to a positive number.
 319          */
 320         if (free_vram >= 128 * 1024 * 1024 || free_vram >= total_vram / 8) {
 321                 s64 min_us;
 322
 323                 /* Be more aggresive on dGPUs. Try to fill a portion of free
 324                  * VRAM now.
 325                  */
 326                 if (!(adev->flags & AMD_IS_APU))
 327                         min_us = bytes_to_us(adev, free_vram / 4);
 328                 else
 329                         min_us = 0; /* Reset accum_us on APUs. */
 330
 331                 adev->mm_stats.accum_us = max(min_us, adev->mm_stats.accum_us);
 332         }
 333
 334         /* This returns 0 if the driver is in debt to disallow (optional)
 335          * buffer moves.
 336          */
 337         max_bytes = us_to_bytes(adev, adev->mm_stats.accum_us);
 338
 339         spin_unlock(&adev->mm_stats.lock);
 340         return max_bytes;
 341 }
 342
 343 /* Report how many bytes have really been moved for the last command
 344  * submission. This can result in a debt that can stop buffer migrations
 345  * temporarily.
 346  */
 347 static void amdgpu_cs_report_moved_bytes(struct amdgpu_device *adev,
 348                                          u64 num_bytes)
 349 {
 350         spin_lock(&adev->mm_stats.lock);
 351         adev->mm_stats.accum_us -= bytes_to_us(adev, num_bytes);
 352         spin_unlock(&adev->mm_stats.lock);
 353 }
 354
 355 static int amdgpu_cs_bo_validate(struct amdgpu_cs_parser *p,
 356                                  struct amdgpu_bo *bo)
 357 {
 358         u64 initial_bytes_moved;
 359         uint32_t domain;
 360         int r;
 361
 362         if (bo->pin_count)
 363                 return 0;
 364
 365         /* Don't move this buffer if we have depleted our allowance
 366          * to move it. Don't move anything if the threshold is zero.
 367          */
 368         if (p->bytes_moved < p->bytes_moved_threshold)
 369                 domain = bo->prefered_domains;
 370         else
 371                 domain = bo->allowed_domains;
 372
 373 retry:
 374         amdgpu_ttm_placement_from_domain(bo, domain);
 375         initial_bytes_moved = atomic64_read(&bo->adev->num_bytes_moved);
 376         r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
 377         p->bytes_moved += atomic64_read(&bo->adev->num_bytes_moved) -
 378                 initial_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_list_entry *lobj)
 391 {
 392         uint32_t domain = lobj->robj->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                 u64 initial_bytes_moved;
 404                 uint32_t other;
 405
 406                 /* If we reached our current BO we can forget it */
 407                 if (candidate == lobj)
 408                         break;
 409
 410                 other = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
 411
 412                 /* Check if this BO is in one of the domains we need space for */
 413                 if (!(other & domain))
 414                         continue;
 415
 416                 /* Check if we can move this BO somewhere else */
 417                 other = bo->allowed_domains & ~domain;
 418                 if (!other)
 419                         continue;
 420
 421                 /* Good we can try to move this BO somewhere else */
 422                 amdgpu_ttm_placement_from_domain(bo, other);
 423                 initial_bytes_moved = atomic64_read(&bo->adev->num_bytes_moved);
 424                 r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
 425                 p->bytes_moved += atomic64_read(&bo->adev->num_bytes_moved) -
 426                         initial_bytes_moved;
 427
 428                 if (unlikely(r))
 429                         break;
 430
 431                 p->evictable = list_prev_entry(p->evictable, tv.head);
 432                 list_move(&candidate->tv.head, &p->validated);
 433
 434                 return true;
 435         }
 436
 437         return false;
 438 }
 439
 440 static int amdgpu_cs_list_validate(struct amdgpu_cs_parser *p,
 441                             struct list_head *validated)
 442 {
 443         struct amdgpu_bo_list_entry *lobj;
 444         int r;
 445
 446         list_for_each_entry(lobj, validated, tv.head) {
 447                 struct amdgpu_bo *bo = lobj->robj;
 448                 bool binding_userptr = false;
 449                 struct mm_struct *usermm;
 450
 451                 usermm = amdgpu_ttm_tt_get_usermm(bo->tbo.ttm);
 452                 if (usermm && usermm != current->mm)
 453                         return -EPERM;
 454
 455                 /* Check if we have user pages and nobody bound the BO already */
 456                 if (lobj->user_pages && bo->tbo.ttm->state != tt_bound) {
 457                         size_t size = sizeof(struct page *);
 458
 459                         size *= bo->tbo.ttm->num_pages;
 460                         memcpy(bo->tbo.ttm->pages, lobj->user_pages, size);
 461                         binding_userptr = true;
 462                 }
 463
 464                 if (p->evictable == lobj)
 465                         p->evictable = NULL;
 466
 467                 do {
 468                         r = amdgpu_cs_bo_validate(p, bo);
 469                 } while (r == -ENOMEM && amdgpu_cs_try_evict(p, lobj));
 470                 if (r)
 471                         return r;
 472
 473                 if (bo->shadow) {
 474                         r = amdgpu_cs_bo_validate(p, bo);
 475                         if (r)
 476                                 return r;
 477                 }
 478
 479                 if (binding_userptr) {
 480                         drm_free_large(lobj->user_pages);
 481                         lobj->user_pages = NULL;
 482                 }
 483         }
 484         return 0;
 485 }
 486
 487 static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
 488                                 union drm_amdgpu_cs *cs)
 489 {
 490         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 491         struct amdgpu_bo_list_entry *e;
 492         struct list_head duplicates;
 493         bool need_mmap_lock = false;
 494         unsigned i, tries = 10;
 495         int r;
 496
 497         INIT_LIST_HEAD(&p->validated);
 498
 499         p->bo_list = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle);
 500         if (p->bo_list) {
 501                 need_mmap_lock = p->bo_list->first_userptr !=
 502                         p->bo_list->num_entries;
 503                 amdgpu_bo_list_get_list(p->bo_list, &p->validated);
 504         }
 505
 506         INIT_LIST_HEAD(&duplicates);
 507         amdgpu_vm_get_pd_bo(&fpriv->vm, &p->validated, &p->vm_pd);
 508
 509         if (p->uf_entry.robj)
 510                 list_add(&p->uf_entry.tv.head, &p->validated);
 511
 512         if (need_mmap_lock)
 513                 down_read(&current->mm->mmap_sem);
 514
 515         while (1) {
 516                 struct list_head need_pages;
 517                 unsigned i;
 518
 519                 r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true,
 520                                            &duplicates);
 521                 if (unlikely(r != 0)) {
 522                         DRM_ERROR("ttm_eu_reserve_buffers failed.\n");
 523                         goto error_free_pages;
 524                 }
 525
 526                 /* Without a BO list we don't have userptr BOs */
 527                 if (!p->bo_list)
 528                         break;
 529
 530                 INIT_LIST_HEAD(&need_pages);
 531                 for (i = p->bo_list->first_userptr;
 532                      i < p->bo_list->num_entries; ++i) {
 533
 534                         e = &p->bo_list->array[i];
 535
 536                         if (amdgpu_ttm_tt_userptr_invalidated(e->robj->tbo.ttm,
 537                                  &e->user_invalidated) && e->user_pages) {
 538
 539                                 /* We acquired a page array, but somebody
 540                                  * invalidated it. Free it an try again
 541                                  */
 542                                 release_pages(e->user_pages,
 543                                               e->robj->tbo.ttm->num_pages,
 544                                               false);
 545                                 drm_free_large(e->user_pages);
 546                                 e->user_pages = NULL;
 547                         }
 548
 549                         if (e->robj->tbo.ttm->state != tt_bound &&
 550                             !e->user_pages) {
 551                                 list_del(&e->tv.head);
 552                                 list_add(&e->tv.head, &need_pages);
 553
 554                                 amdgpu_bo_unreserve(e->robj);
 555                         }
 556                 }
 557
 558                 if (list_empty(&need_pages))
 559                         break;
 560
 561                 /* Unreserve everything again. */
 562                 ttm_eu_backoff_reservation(&p->ticket, &p->validated);
 563
 564                 /* We tried too many times, just abort */
 565                 if (!--tries) {
 566                         r = -EDEADLK;
 567                         DRM_ERROR("deadlock in %s\n", __func__);
 568                         goto error_free_pages;
 569                 }
 570
 571                 /* Fill the page arrays for all useptrs. */
 572                 list_for_each_entry(e, &need_pages, tv.head) {
 573                         struct ttm_tt *ttm = e->robj->tbo.ttm;
 574
 575                         e->user_pages = drm_calloc_large(ttm->num_pages,
 576                                                          sizeof(struct page*));
 577                         if (!e->user_pages) {
 578                                 r = -ENOMEM;
 579                                 DRM_ERROR("calloc failure in %s\n", __func__);
 580                                 goto error_free_pages;
 581                         }
 582
 583                         r = amdgpu_ttm_tt_get_user_pages(ttm, e->user_pages);
 584                         if (r) {
 585                                 DRM_ERROR("amdgpu_ttm_tt_get_user_pages failed.\n");
 586                                 drm_free_large(e->user_pages);
 587                                 e->user_pages = NULL;
 588                                 goto error_free_pages;
 589                         }
 590                 }
 591
 592                 /* And try again. */
 593                 list_splice(&need_pages, &p->validated);
 594         }
 595
 596         amdgpu_vm_get_pt_bos(p->adev, &fpriv->vm, &duplicates);
 597
 598         p->bytes_moved_threshold = amdgpu_cs_get_threshold_for_moves(p->adev);
 599         p->bytes_moved = 0;
 600         p->evictable = list_last_entry(&p->validated,
 601                                        struct amdgpu_bo_list_entry,
 602                                        tv.head);
 603
 604         r = amdgpu_cs_list_validate(p, &duplicates);
 605         if (r) {
 606                 DRM_ERROR("amdgpu_cs_list_validate(duplicates) failed.\n");
 607                 goto error_validate;
 608         }
 609
 610         r = amdgpu_cs_list_validate(p, &p->validated);
 611         if (r) {
 612                 DRM_ERROR("amdgpu_cs_list_validate(validated) failed.\n");
 613                 goto error_validate;
 614         }
 615
 616         amdgpu_cs_report_moved_bytes(p->adev, p->bytes_moved);
 617
 618         fpriv->vm.last_eviction_counter =
 619                 atomic64_read(&p->adev->num_evictions);
 620
 621         if (p->bo_list) {
 622                 struct amdgpu_bo *gds = p->bo_list->gds_obj;
 623                 struct amdgpu_bo *gws = p->bo_list->gws_obj;
 624                 struct amdgpu_bo *oa = p->bo_list->oa_obj;
 625                 struct amdgpu_vm *vm = &fpriv->vm;
 626                 unsigned i;
 627
 628                 for (i = 0; i < p->bo_list->num_entries; i++) {
 629                         struct amdgpu_bo *bo = p->bo_list->array[i].robj;
 630
 631                         p->bo_list->array[i].bo_va = amdgpu_vm_bo_find(vm, bo);
 632                 }
 633
 634                 if (gds) {
 635                         p->job->gds_base = amdgpu_bo_gpu_offset(gds);
 636                         p->job->gds_size = amdgpu_bo_size(gds);
 637                 }
 638                 if (gws) {
 639                         p->job->gws_base = amdgpu_bo_gpu_offset(gws);
 640                         p->job->gws_size = amdgpu_bo_size(gws);
 641                 }
 642                 if (oa) {
 643                         p->job->oa_base = amdgpu_bo_gpu_offset(oa);
 644                         p->job->oa_size = amdgpu_bo_size(oa);
 645                 }
 646         }
 647
 648         if (!r && p->uf_entry.robj) {
 649                 struct amdgpu_bo *uf = p->uf_entry.robj;
 650
 651                 r = amdgpu_ttm_bind(&uf->tbo, &uf->tbo.mem);
 652                 p->job->uf_addr += amdgpu_bo_gpu_offset(uf);
 653         }
 654
 655 error_validate:
 656         if (r) {
 657                 amdgpu_vm_move_pt_bos_in_lru(p->adev, &fpriv->vm);
 658                 ttm_eu_backoff_reservation(&p->ticket, &p->validated);
 659         }
 660
 661 error_free_pages:
 662
 663         if (need_mmap_lock)
 664                 up_read(&current->mm->mmap_sem);
 665
 666         if (p->bo_list) {
 667                 for (i = p->bo_list->first_userptr;
 668                      i < p->bo_list->num_entries; ++i) {
 669                         e = &p->bo_list->array[i];
 670
 671                         if (!e->user_pages)
 672                                 continue;
 673
 674                         release_pages(e->user_pages,
 675                                       e->robj->tbo.ttm->num_pages,
 676                                       false);
 677                         drm_free_large(e->user_pages);
 678                 }
 679         }
 680
 681         return r;
 682 }
 683
 684 static int amdgpu_cs_sync_rings(struct amdgpu_cs_parser *p)
 685 {
 686         struct amdgpu_bo_list_entry *e;
 687         int r;
 688
 689         list_for_each_entry(e, &p->validated, tv.head) {
 690                 struct reservation_object *resv = e->robj->tbo.resv;
 691                 r = amdgpu_sync_resv(p->adev, &p->job->sync, resv, p->filp);
 692
 693                 if (r)
 694                         return r;
 695         }
 696         return 0;
 697 }
 698
 699 /**
 700  * cs_parser_fini() - clean parser states
 701  * @parser:     parser structure holding parsing context.
 702  * @error:      error number
 703  *
 704  * If error is set than unvalidate buffer, otherwise just free memory
 705  * used by parsing context.
 706  **/
 707 static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error, bool backoff)
 708 {
 709         struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
 710         unsigned i;
 711
 712         if (!error) {
 713                 amdgpu_vm_move_pt_bos_in_lru(parser->adev, &fpriv->vm);
 714
 715                 ttm_eu_fence_buffer_objects(&parser->ticket,
 716                                             &parser->validated,
 717                                             parser->fence);
 718         } else if (backoff) {
 719                 ttm_eu_backoff_reservation(&parser->ticket,
 720                                            &parser->validated);
 721         }
 722         fence_put(parser->fence);
 723
 724         if (parser->ctx)
 725                 amdgpu_ctx_put(parser->ctx);
 726         if (parser->bo_list)
 727                 amdgpu_bo_list_put(parser->bo_list);
 728
 729         for (i = 0; i < parser->nchunks; i++)
 730                 drm_free_large(parser->chunks[i].kdata);
 731         kfree(parser->chunks);
 732         if (parser->job)
 733                 amdgpu_job_free(parser->job);
 734         amdgpu_bo_unref(&parser->uf_entry.robj);
 735 }
 736
 737 static int amdgpu_bo_vm_update_pte(struct amdgpu_cs_parser *p,
 738                                    struct amdgpu_vm *vm)
 739 {
 740         struct amdgpu_device *adev = p->adev;
 741         struct amdgpu_bo_va *bo_va;
 742         struct amdgpu_bo *bo;
 743         int i, r;
 744
 745         r = amdgpu_vm_update_page_directory(adev, vm);
 746         if (r)
 747                 return r;
 748
 749         r = amdgpu_sync_fence(adev, &p->job->sync, vm->page_directory_fence);
 750         if (r)
 751                 return r;
 752
 753         r = amdgpu_vm_clear_freed(adev, vm);
 754         if (r)
 755                 return r;
 756
 757         if (p->bo_list) {
 758                 for (i = 0; i < p->bo_list->num_entries; i++) {
 759                         struct fence *f;
 760
 761                         /* ignore duplicates */
 762                         bo = p->bo_list->array[i].robj;
 763                         if (!bo)
 764                                 continue;
 765
 766                         bo_va = p->bo_list->array[i].bo_va;
 767                         if (bo_va == NULL)
 768                                 continue;
 769
 770                         r = amdgpu_vm_bo_update(adev, bo_va, false);
 771                         if (r)
 772                                 return r;
 773
 774                         f = bo_va->last_pt_update;
 775                         r = amdgpu_sync_fence(adev, &p->job->sync, f);
 776                         if (r)
 777                                 return r;
 778                 }
 779
 780         }
 781
 782         r = amdgpu_vm_clear_invalids(adev, vm, &p->job->sync);
 783
 784         if (amdgpu_vm_debug && p->bo_list) {
 785                 /* Invalidate all BOs to test for userspace bugs */
 786                 for (i = 0; i < p->bo_list->num_entries; i++) {
 787                         /* ignore duplicates */
 788                         bo = p->bo_list->array[i].robj;
 789                         if (!bo)
 790                                 continue;
 791
 792                         amdgpu_vm_bo_invalidate(adev, bo);
 793                 }
 794         }
 795
 796         return r;
 797 }
 798
 799 static int amdgpu_cs_ib_vm_chunk(struct amdgpu_device *adev,
 800                                  struct amdgpu_cs_parser *p)
 801 {
 802         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 803         struct amdgpu_vm *vm = &fpriv->vm;
 804         struct amdgpu_ring *ring = p->job->ring;
 805         int i, r;
 806
 807         /* Only for UVD/VCE VM emulation */
 808         if (ring->funcs->parse_cs) {
 809                 p->job->vm = NULL;
 810                 for (i = 0; i < p->job->num_ibs; i++) {
 811                         r = amdgpu_ring_parse_cs(ring, p, i);
 812                         if (r)
 813                                 return r;
 814                 }
 815         } else {
 816                 p->job->vm_pd_addr = amdgpu_bo_gpu_offset(vm->page_directory);
 817
 818                 r = amdgpu_bo_vm_update_pte(p, vm);
 819                 if (r)
 820                         return r;
 821         }
 822
 823         return amdgpu_cs_sync_rings(p);
 824 }
 825
 826 static int amdgpu_cs_handle_lockup(struct amdgpu_device *adev, int r)
 827 {
 828         if (r == -EDEADLK) {
 829                 r = amdgpu_gpu_reset(adev);
 830                 if (!r)
 831                         r = -EAGAIN;
 832         }
 833         return r;
 834 }
 835
 836 static int amdgpu_cs_ib_fill(struct amdgpu_device *adev,
 837                              struct amdgpu_cs_parser *parser)
 838 {
 839         struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
 840         struct amdgpu_vm *vm = &fpriv->vm;
 841         int i, j;
 842         int r;
 843
 844         for (i = 0, j = 0; i < parser->nchunks && j < parser->job->num_ibs; i++) {
 845                 struct amdgpu_cs_chunk *chunk;
 846                 struct amdgpu_ib *ib;
 847                 struct drm_amdgpu_cs_chunk_ib *chunk_ib;
 848                 struct amdgpu_ring *ring;
 849
 850                 chunk = &parser->chunks[i];
 851                 ib = &parser->job->ibs[j];
 852                 chunk_ib = (struct drm_amdgpu_cs_chunk_ib *)chunk->kdata;
 853
 854                 if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
 855                         continue;
 856
 857                 r = amdgpu_cs_get_ring(adev, chunk_ib->ip_type,
 858                                        chunk_ib->ip_instance, chunk_ib->ring,
 859                                        &ring);
 860                 if (r)
 861                         return r;
 862
 863                 if (ib->flags & AMDGPU_IB_FLAG_PREAMBLE) {
 864                         parser->job->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT;
 865                         if (!parser->ctx->preamble_presented) {
 866                                 parser->job->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT_FIRST;
 867                                 parser->ctx->preamble_presented = true;
 868                         }
 869                 }
 870
 871                 if (parser->job->ring && parser->job->ring != ring)
 872                         return -EINVAL;
 873
 874                 parser->job->ring = ring;
 875
 876                 if (ring->funcs->parse_cs) {
 877                         struct amdgpu_bo_va_mapping *m;
 878                         struct amdgpu_bo *aobj = NULL;
 879                         uint64_t offset;
 880                         uint8_t *kptr;
 881
 882                         m = amdgpu_cs_find_mapping(parser, chunk_ib->va_start,
 883                                                    &aobj);
 884                         if (!aobj) {
 885                                 DRM_ERROR("IB va_start is invalid\n");
 886                                 return -EINVAL;
 887                         }
 888
 889                         if ((chunk_ib->va_start + chunk_ib->ib_bytes) >
 890                             (m->it.last + 1) * AMDGPU_GPU_PAGE_SIZE) {
 891                                 DRM_ERROR("IB va_start+ib_bytes is invalid\n");
 892                                 return -EINVAL;
 893                         }
 894
 895                         /* the IB should be reserved at this point */
 896                         r = amdgpu_bo_kmap(aobj, (void **)&kptr);
 897                         if (r) {
 898                                 return r;
 899                         }
 900
 901                         offset = ((uint64_t)m->it.start) * AMDGPU_GPU_PAGE_SIZE;
 902                         kptr += chunk_ib->va_start - offset;
 903
 904                         r =  amdgpu_ib_get(adev, NULL, chunk_ib->ib_bytes, ib);
 905                         if (r) {
 906                                 DRM_ERROR("Failed to get ib !\n");
 907                                 return r;
 908                         }
 909
 910                         memcpy(ib->ptr, kptr, chunk_ib->ib_bytes);
 911                         amdgpu_bo_kunmap(aobj);
 912                 } else {
 913                         r =  amdgpu_ib_get(adev, vm, 0, ib);
 914                         if (r) {
 915                                 DRM_ERROR("Failed to get ib !\n");
 916                                 return r;
 917                         }
 918
 919                         ib->gpu_addr = chunk_ib->va_start;
 920                 }
 921
 922                 ib->length_dw = chunk_ib->ib_bytes / 4;
 923                 ib->flags = chunk_ib->flags;
 924                 j++;
 925         }
 926
 927         /* UVD & VCE fw doesn't support user fences */
 928         if (parser->job->uf_addr && (
 929             parser->job->ring->type == AMDGPU_RING_TYPE_UVD ||
 930             parser->job->ring->type == AMDGPU_RING_TYPE_VCE))
 931                 return -EINVAL;
 932
 933         return 0;
 934 }
 935
 936 static int amdgpu_cs_dependencies(struct amdgpu_device *adev,
 937                                   struct amdgpu_cs_parser *p)
 938 {
 939         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 940         int i, j, r;
 941
 942         for (i = 0; i < p->nchunks; ++i) {
 943                 struct drm_amdgpu_cs_chunk_dep *deps;
 944                 struct amdgpu_cs_chunk *chunk;
 945                 unsigned num_deps;
 946
 947                 chunk = &p->chunks[i];
 948
 949                 if (chunk->chunk_id != AMDGPU_CHUNK_ID_DEPENDENCIES)
 950                         continue;
 951
 952                 deps = (struct drm_amdgpu_cs_chunk_dep *)chunk->kdata;
 953                 num_deps = chunk->length_dw * 4 /
 954                         sizeof(struct drm_amdgpu_cs_chunk_dep);
 955
 956                 for (j = 0; j < num_deps; ++j) {
 957                         struct amdgpu_ring *ring;
 958                         struct amdgpu_ctx *ctx;
 959                         struct fence *fence;
 960
 961                         r = amdgpu_cs_get_ring(adev, deps[j].ip_type,
 962                                                deps[j].ip_instance,
 963                                                deps[j].ring, &ring);
 964                         if (r)
 965                                 return r;
 966
 967                         ctx = amdgpu_ctx_get(fpriv, deps[j].ctx_id);
 968                         if (ctx == NULL)
 969                                 return -EINVAL;
 970
 971                         fence = amdgpu_ctx_get_fence(ctx, ring,
 972                                                      deps[j].handle);
 973                         if (IS_ERR(fence)) {
 974                                 r = PTR_ERR(fence);
 975                                 amdgpu_ctx_put(ctx);
 976                                 return r;
 977
 978                         } else if (fence) {
 979                                 r = amdgpu_sync_fence(adev, &p->job->sync,
 980                                                       fence);
 981                                 fence_put(fence);
 982                                 amdgpu_ctx_put(ctx);
 983                                 if (r)
 984                                         return r;
 985                         }
 986                 }
 987         }
 988
 989         return 0;
 990 }
 991
 992 static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
 993                             union drm_amdgpu_cs *cs)
 994 {
 995         struct amdgpu_ring *ring = p->job->ring;
 996         struct amd_sched_entity *entity = &p->ctx->rings[ring->idx].entity;
 997         struct amdgpu_job *job;
 998         int r;
 999
1000         job = p->job;
1001         p->job = NULL;
1002
1003         r = amd_sched_job_init(&job->base, &ring->sched, entity, p->filp);
1004         if (r) {
1005                 amdgpu_job_free(job);
1006                 return r;
1007         }
1008
1009         job->owner = p->filp;
1010         job->fence_ctx = entity->fence_context;
1011         p->fence = fence_get(&job->base.s_fence->finished);
1012         cs->out.handle = amdgpu_ctx_add_fence(p->ctx, ring, p->fence);
1013         job->uf_sequence = cs->out.handle;
1014         amdgpu_job_free_resources(job);
1015
1016         trace_amdgpu_cs_ioctl(job);
1017         amd_sched_entity_push_job(&job->base);
1018
1019         return 0;
1020 }
1021
1022 int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
1023 {
1024         struct amdgpu_device *adev = dev->dev_private;
1025         union drm_amdgpu_cs *cs = data;
1026         struct amdgpu_cs_parser parser = {};
1027         bool reserved_buffers = false;
1028         int i, r;
1029
1030         if (!adev->accel_working)
1031                 return -EBUSY;
1032
1033         parser.adev = adev;
1034         parser.filp = filp;
1035
1036         r = amdgpu_cs_parser_init(&parser, data);
1037         if (r) {
1038                 DRM_ERROR("Failed to initialize parser !\n");
1039                 amdgpu_cs_parser_fini(&parser, r, false);
1040                 r = amdgpu_cs_handle_lockup(adev, r);
1041                 return r;
1042         }
1043         r = amdgpu_cs_parser_bos(&parser, data);
1044         if (r == -ENOMEM)
1045                 DRM_ERROR("Not enough memory for command submission!\n");
1046         else if (r && r != -ERESTARTSYS)
1047                 DRM_ERROR("Failed to process the buffer list %d!\n", r);
1048         else if (!r) {
1049                 reserved_buffers = true;
1050                 r = amdgpu_cs_ib_fill(adev, &parser);
1051         }
1052
1053         if (!r) {
1054                 r = amdgpu_cs_dependencies(adev, &parser);
1055                 if (r)
1056                         DRM_ERROR("Failed in the dependencies handling %d!\n", r);
1057         }
1058
1059         if (r)
1060                 goto out;
1061
1062         for (i = 0; i < parser.job->num_ibs; i++)
1063                 trace_amdgpu_cs(&parser, i);
1064
1065         r = amdgpu_cs_ib_vm_chunk(adev, &parser);
1066         if (r)
1067                 goto out;
1068
1069         r = amdgpu_cs_submit(&parser, cs);
1070
1071 out:
1072         amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
1073         r = amdgpu_cs_handle_lockup(adev, r);
1074         return r;
1075 }
1076
1077 /**
1078  * amdgpu_cs_wait_ioctl - wait for a command submission to finish
1079  *
1080  * @dev: drm device
1081  * @data: data from userspace
1082  * @filp: file private
1083  *
1084  * Wait for the command submission identified by handle to finish.
1085  */
1086 int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data,
1087                          struct drm_file *filp)
1088 {
1089         union drm_amdgpu_wait_cs *wait = data;
1090         struct amdgpu_device *adev = dev->dev_private;
1091         unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
1092         struct amdgpu_ring *ring = NULL;
1093         struct amdgpu_ctx *ctx;
1094         struct fence *fence;
1095         long r;
1096
1097         r = amdgpu_cs_get_ring(adev, wait->in.ip_type, wait->in.ip_instance,
1098                                wait->in.ring, &ring);
1099         if (r)
1100                 return r;
1101
1102         ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
1103         if (ctx == NULL)
1104                 return -EINVAL;
1105
1106         fence = amdgpu_ctx_get_fence(ctx, ring, wait->in.handle);
1107         if (IS_ERR(fence))
1108                 r = PTR_ERR(fence);
1109         else if (fence) {
1110                 r = fence_wait_timeout(fence, true, timeout);
1111                 fence_put(fence);
1112         } else
1113                 r = 1;
1114
1115         amdgpu_ctx_put(ctx);
1116         if (r < 0)
1117                 return r;
1118
1119         memset(wait, 0, sizeof(*wait));
1120         wait->out.status = (r == 0);
1121
1122         return 0;
1123 }
1124
1125 /**
1126  * amdgpu_cs_find_bo_va - find bo_va for VM address
1127  *
1128  * @parser: command submission parser context
1129  * @addr: VM address
1130  * @bo: resulting BO of the mapping found
1131  *
1132  * Search the buffer objects in the command submission context for a certain
1133  * virtual memory address. Returns allocation structure when found, NULL
1134  * otherwise.
1135  */
1136 struct amdgpu_bo_va_mapping *
1137 amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
1138                        uint64_t addr, struct amdgpu_bo **bo)
1139 {
1140         struct amdgpu_bo_va_mapping *mapping;
1141         unsigned i;
1142
1143         if (!parser->bo_list)
1144                 return NULL;
1145
1146         addr /= AMDGPU_GPU_PAGE_SIZE;
1147
1148         for (i = 0; i < parser->bo_list->num_entries; i++) {
1149                 struct amdgpu_bo_list_entry *lobj;
1150
1151                 lobj = &parser->bo_list->array[i];
1152                 if (!lobj->bo_va)
1153                         continue;
1154
1155                 list_for_each_entry(mapping, &lobj->bo_va->valids, list) {
1156                         if (mapping->it.start > addr ||
1157                             addr > mapping->it.last)
1158                                 continue;
1159
1160                         *bo = lobj->bo_va->bo;
1161                         return mapping;
1162                 }
1163
1164                 list_for_each_entry(mapping, &lobj->bo_va->invalids, list) {
1165                         if (mapping->it.start > addr ||
1166                             addr > mapping->it.last)
1167                                 continue;
1168
1169                         *bo = lobj->bo_va->bo;
1170                         return mapping;
1171                 }
1172         }
1173
1174         return NULL;
1175 }
1176
1177 /**
1178  * amdgpu_cs_sysvm_access_required - make BOs accessible by the system VM
1179  *
1180  * @parser: command submission parser context
1181  *
1182  * Helper for UVD/VCE VM emulation, make sure BOs are accessible by the system VM.
1183  */
1184 int amdgpu_cs_sysvm_access_required(struct amdgpu_cs_parser *parser)
1185 {
1186         unsigned i;
1187         int r;
1188
1189         if (!parser->bo_list)
1190                 return 0;
1191
1192         for (i = 0; i < parser->bo_list->num_entries; i++) {
1193                 struct amdgpu_bo *bo = parser->bo_list->array[i].robj;
1194
1195                 r = amdgpu_ttm_bind(&bo->tbo, &bo->tbo.mem);
1196                 if (unlikely(r))
1197                         return r;
1198         }
1199
1200         return 0;
1201 }