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
  28 #include <linux/file.h>
  29 #include <linux/pagemap.h>
  30 #include <linux/sync_file.h>
  31
  32 #include <drm/amdgpu_drm.h>
  33 #include <drm/drm_syncobj.h>
  34 #include "amdgpu.h"
  35 #include "amdgpu_trace.h"
  36 #include "amdgpu_gmc.h"
  37 #include "amdgpu_gem.h"
  38 #include "amdgpu_ras.h"
  39
  40 static int amdgpu_cs_user_fence_chunk(struct amdgpu_cs_parser *p,
  41                                       struct drm_amdgpu_cs_chunk_fence *data,
  42                                       uint32_t *offset)
  43 {
  44         struct drm_gem_object *gobj;
  45         struct amdgpu_bo *bo;
  46         unsigned long size;
  47         int r;
  48
  49         gobj = drm_gem_object_lookup(p->filp, data->handle);
  50         if (gobj == NULL)
  51                 return -EINVAL;
  52
  53         bo = amdgpu_bo_ref(gem_to_amdgpu_bo(gobj));
  54         p->uf_entry.priority = 0;
  55         p->uf_entry.tv.bo = &bo->tbo;
  56         /* One for TTM and one for the CS job */
  57         p->uf_entry.tv.num_shared = 2;
  58
  59         drm_gem_object_put_unlocked(gobj);
  60
  61         size = amdgpu_bo_size(bo);
  62         if (size != PAGE_SIZE || (data->offset + 8) > size) {
  63                 r = -EINVAL;
  64                 goto error_unref;
  65         }
  66
  67         if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
  68                 r = -EINVAL;
  69                 goto error_unref;
  70         }
  71
  72         *offset = data->offset;
  73
  74         return 0;
  75
  76 error_unref:
  77         amdgpu_bo_unref(&bo);
  78         return r;
  79 }
  80
  81 static int amdgpu_cs_bo_handles_chunk(struct amdgpu_cs_parser *p,
  82                                       struct drm_amdgpu_bo_list_in *data)
  83 {
  84         int r;
  85         struct drm_amdgpu_bo_list_entry *info = NULL;
  86
  87         r = amdgpu_bo_create_list_entry_array(data, &info);
  88         if (r)
  89                 return r;
  90
  91         r = amdgpu_bo_list_create(p->adev, p->filp, info, data->bo_number,
  92                                   &p->bo_list);
  93         if (r)
  94                 goto error_free;
  95
  96         kvfree(info);
  97         return 0;
  98
  99 error_free:
 100         if (info)
 101                 kvfree(info);
 102
 103         return r;
 104 }
 105
 106 static int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, union drm_amdgpu_cs *cs)
 107 {
 108         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 109         struct amdgpu_vm *vm = &fpriv->vm;
 110         uint64_t *chunk_array_user;
 111         uint64_t *chunk_array;
 112         unsigned size, num_ibs = 0;
 113         uint32_t uf_offset = 0;
 114         int i;
 115         int ret;
 116
 117         if (cs->in.num_chunks == 0)
 118                 return 0;
 119
 120         chunk_array = kmalloc_array(cs->in.num_chunks, sizeof(uint64_t), GFP_KERNEL);
 121         if (!chunk_array)
 122                 return -ENOMEM;
 123
 124         p->ctx = amdgpu_ctx_get(fpriv, cs->in.ctx_id);
 125         if (!p->ctx) {
 126                 ret = -EINVAL;
 127                 goto free_chunk;
 128         }
 129
 130         mutex_lock(&p->ctx->lock);
 131
 132         /* skip guilty context job */
 133         if (atomic_read(&p->ctx->guilty) == 1) {
 134                 ret = -ECANCELED;
 135                 goto free_chunk;
 136         }
 137
 138         /* get chunks */
 139         chunk_array_user = u64_to_user_ptr(cs->in.chunks);
 140         if (copy_from_user(chunk_array, chunk_array_user,
 141                            sizeof(uint64_t)*cs->in.num_chunks)) {
 142                 ret = -EFAULT;
 143                 goto free_chunk;
 144         }
 145
 146         p->nchunks = cs->in.num_chunks;
 147         p->chunks = kmalloc_array(p->nchunks, sizeof(struct amdgpu_cs_chunk),
 148                             GFP_KERNEL);
 149         if (!p->chunks) {
 150                 ret = -ENOMEM;
 151                 goto free_chunk;
 152         }
 153
 154         for (i = 0; i < p->nchunks; i++) {
 155                 struct drm_amdgpu_cs_chunk __user **chunk_ptr = NULL;
 156                 struct drm_amdgpu_cs_chunk user_chunk;
 157                 uint32_t __user *cdata;
 158
 159                 chunk_ptr = u64_to_user_ptr(chunk_array[i]);
 160                 if (copy_from_user(&user_chunk, chunk_ptr,
 161                                        sizeof(struct drm_amdgpu_cs_chunk))) {
 162                         ret = -EFAULT;
 163                         i--;
 164                         goto free_partial_kdata;
 165                 }
 166                 p->chunks[i].chunk_id = user_chunk.chunk_id;
 167                 p->chunks[i].length_dw = user_chunk.length_dw;
 168
 169                 size = p->chunks[i].length_dw;
 170                 cdata = u64_to_user_ptr(user_chunk.chunk_data);
 171
 172                 p->chunks[i].kdata = kvmalloc_array(size, sizeof(uint32_t), GFP_KERNEL);
 173                 if (p->chunks[i].kdata == NULL) {
 174                         ret = -ENOMEM;
 175                         i--;
 176                         goto free_partial_kdata;
 177                 }
 178                 size *= sizeof(uint32_t);
 179                 if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
 180                         ret = -EFAULT;
 181                         goto free_partial_kdata;
 182                 }
 183
 184                 switch (p->chunks[i].chunk_id) {
 185                 case AMDGPU_CHUNK_ID_IB:
 186                         ++num_ibs;
 187                         break;
 188
 189                 case AMDGPU_CHUNK_ID_FENCE:
 190                         size = sizeof(struct drm_amdgpu_cs_chunk_fence);
 191                         if (p->chunks[i].length_dw * sizeof(uint32_t) < size) {
 192                                 ret = -EINVAL;
 193                                 goto free_partial_kdata;
 194                         }
 195
 196                         ret = amdgpu_cs_user_fence_chunk(p, p->chunks[i].kdata,
 197                                                          &uf_offset);
 198                         if (ret)
 199                                 goto free_partial_kdata;
 200
 201                         break;
 202
 203                 case AMDGPU_CHUNK_ID_BO_HANDLES:
 204                         size = sizeof(struct drm_amdgpu_bo_list_in);
 205                         if (p->chunks[i].length_dw * sizeof(uint32_t) < size) {
 206                                 ret = -EINVAL;
 207                                 goto free_partial_kdata;
 208                         }
 209
 210                         ret = amdgpu_cs_bo_handles_chunk(p, p->chunks[i].kdata);
 211                         if (ret)
 212                                 goto free_partial_kdata;
 213
 214                         break;
 215
 216                 case AMDGPU_CHUNK_ID_DEPENDENCIES:
 217                 case AMDGPU_CHUNK_ID_SYNCOBJ_IN:
 218                 case AMDGPU_CHUNK_ID_SYNCOBJ_OUT:
 219                 case AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES:
 220                 case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_WAIT:
 221                 case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_SIGNAL:
 222                         break;
 223
 224                 default:
 225                         ret = -EINVAL;
 226                         goto free_partial_kdata;
 227                 }
 228         }
 229
 230         ret = amdgpu_job_alloc(p->adev, num_ibs, &p->job, vm);
 231         if (ret)
 232                 goto free_all_kdata;
 233
 234         if (p->ctx->vram_lost_counter != p->job->vram_lost_counter) {
 235                 ret = -ECANCELED;
 236                 goto free_all_kdata;
 237         }
 238
 239         if (p->uf_entry.tv.bo)
 240                 p->job->uf_addr = uf_offset;
 241         kfree(chunk_array);
 242
 243         /* Use this opportunity to fill in task info for the vm */
 244         amdgpu_vm_set_task_info(vm);
 245
 246         return 0;
 247
 248 free_all_kdata:
 249         i = p->nchunks - 1;
 250 free_partial_kdata:
 251         for (; i >= 0; i--)
 252                 kvfree(p->chunks[i].kdata);
 253         kfree(p->chunks);
 254         p->chunks = NULL;
 255         p->nchunks = 0;
 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 void amdgpu_cs_get_threshold_for_moves(struct amdgpu_device *adev,
 296                                               u64 *max_bytes,
 297                                               u64 *max_vis_bytes)
 298 {
 299         s64 time_us, increment_us;
 300         u64 free_vram, total_vram, used_vram;
 301
 302         /* Allow a maximum of 200 accumulated ms. This is basically per-IB
 303          * throttling.
 304          *
 305          * It means that in order to get full max MBps, at least 5 IBs per
 306          * second must be submitted and not more than 200ms apart from each
 307          * other.
 308          */
 309         const s64 us_upper_bound = 200000;
 310
 311         if (!adev->mm_stats.log2_max_MBps) {
 312                 *max_bytes = 0;
 313                 *max_vis_bytes = 0;
 314                 return;
 315         }
 316
 317         total_vram = adev->gmc.real_vram_size - atomic64_read(&adev->vram_pin_size);
 318         used_vram = amdgpu_vram_mgr_usage(&adev->mman.bdev.man[TTM_PL_VRAM]);
 319         free_vram = used_vram >= total_vram ? 0 : total_vram - used_vram;
 320
 321         spin_lock(&adev->mm_stats.lock);
 322
 323         /* Increase the amount of accumulated us. */
 324         time_us = ktime_to_us(ktime_get());
 325         increment_us = time_us - adev->mm_stats.last_update_us;
 326         adev->mm_stats.last_update_us = time_us;
 327         adev->mm_stats.accum_us = min(adev->mm_stats.accum_us + increment_us,
 328                                       us_upper_bound);
 329
 330         /* This prevents the short period of low performance when the VRAM
 331          * usage is low and the driver is in debt or doesn't have enough
 332          * accumulated us to fill VRAM quickly.
 333          *
 334          * The situation can occur in these cases:
 335          * - a lot of VRAM is freed by userspace
 336          * - the presence of a big buffer causes a lot of evictions
 337          *   (solution: split buffers into smaller ones)
 338          *
 339          * If 128 MB or 1/8th of VRAM is free, start filling it now by setting
 340          * accum_us to a positive number.
 341          */
 342         if (free_vram >= 128 * 1024 * 1024 || free_vram >= total_vram / 8) {
 343                 s64 min_us;
 344
 345                 /* Be more aggresive on dGPUs. Try to fill a portion of free
 346                  * VRAM now.
 347                  */
 348                 if (!(adev->flags & AMD_IS_APU))
 349                         min_us = bytes_to_us(adev, free_vram / 4);
 350                 else
 351                         min_us = 0; /* Reset accum_us on APUs. */
 352
 353                 adev->mm_stats.accum_us = max(min_us, adev->mm_stats.accum_us);
 354         }
 355
 356         /* This is set to 0 if the driver is in debt to disallow (optional)
 357          * buffer moves.
 358          */
 359         *max_bytes = us_to_bytes(adev, adev->mm_stats.accum_us);
 360
 361         /* Do the same for visible VRAM if half of it is free */
 362         if (!amdgpu_gmc_vram_full_visible(&adev->gmc)) {
 363                 u64 total_vis_vram = adev->gmc.visible_vram_size;
 364                 u64 used_vis_vram =
 365                         amdgpu_vram_mgr_vis_usage(&adev->mman.bdev.man[TTM_PL_VRAM]);
 366
 367                 if (used_vis_vram < total_vis_vram) {
 368                         u64 free_vis_vram = total_vis_vram - used_vis_vram;
 369                         adev->mm_stats.accum_us_vis = min(adev->mm_stats.accum_us_vis +
 370                                                           increment_us, us_upper_bound);
 371
 372                         if (free_vis_vram >= total_vis_vram / 2)
 373                                 adev->mm_stats.accum_us_vis =
 374                                         max(bytes_to_us(adev, free_vis_vram / 2),
 375                                             adev->mm_stats.accum_us_vis);
 376                 }
 377
 378                 *max_vis_bytes = us_to_bytes(adev, adev->mm_stats.accum_us_vis);
 379         } else {
 380                 *max_vis_bytes = 0;
 381         }
 382
 383         spin_unlock(&adev->mm_stats.lock);
 384 }
 385
 386 /* Report how many bytes have really been moved for the last command
 387  * submission. This can result in a debt that can stop buffer migrations
 388  * temporarily.
 389  */
 390 void amdgpu_cs_report_moved_bytes(struct amdgpu_device *adev, u64 num_bytes,
 391                                   u64 num_vis_bytes)
 392 {
 393         spin_lock(&adev->mm_stats.lock);
 394         adev->mm_stats.accum_us -= bytes_to_us(adev, num_bytes);
 395         adev->mm_stats.accum_us_vis -= bytes_to_us(adev, num_vis_bytes);
 396         spin_unlock(&adev->mm_stats.lock);
 397 }
 398
 399 static int amdgpu_cs_bo_validate(struct amdgpu_cs_parser *p,
 400                                  struct amdgpu_bo *bo)
 401 {
 402         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
 403         struct ttm_operation_ctx ctx = {
 404                 .interruptible = true,
 405                 .no_wait_gpu = false,
 406                 .resv = bo->tbo.base.resv,
 407                 .flags = 0
 408         };
 409         uint32_t domain;
 410         int r;
 411
 412         if (bo->pin_count)
 413                 return 0;
 414
 415         /* Don't move this buffer if we have depleted our allowance
 416          * to move it. Don't move anything if the threshold is zero.
 417          */
 418         if (p->bytes_moved < p->bytes_moved_threshold) {
 419                 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
 420                     (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)) {
 421                         /* And don't move a CPU_ACCESS_REQUIRED BO to limited
 422                          * visible VRAM if we've depleted our allowance to do
 423                          * that.
 424                          */
 425                         if (p->bytes_moved_vis < p->bytes_moved_vis_threshold)
 426                                 domain = bo->preferred_domains;
 427                         else
 428                                 domain = bo->allowed_domains;
 429                 } else {
 430                         domain = bo->preferred_domains;
 431                 }
 432         } else {
 433                 domain = bo->allowed_domains;
 434         }
 435
 436 retry:
 437         amdgpu_bo_placement_from_domain(bo, domain);
 438         r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 439
 440         p->bytes_moved += ctx.bytes_moved;
 441         if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
 442             amdgpu_bo_in_cpu_visible_vram(bo))
 443                 p->bytes_moved_vis += ctx.bytes_moved;
 444
 445         if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
 446                 domain = bo->allowed_domains;
 447                 goto retry;
 448         }
 449
 450         return r;
 451 }
 452
 453 static int amdgpu_cs_validate(void *param, struct amdgpu_bo *bo)
 454 {
 455         struct amdgpu_cs_parser *p = param;
 456         int r;
 457
 458         r = amdgpu_cs_bo_validate(p, bo);
 459         if (r)
 460                 return r;
 461
 462         if (bo->shadow)
 463                 r = amdgpu_cs_bo_validate(p, bo->shadow);
 464
 465         return r;
 466 }
 467
 468 static int amdgpu_cs_list_validate(struct amdgpu_cs_parser *p,
 469                             struct list_head *validated)
 470 {
 471         struct ttm_operation_ctx ctx = { true, false };
 472         struct amdgpu_bo_list_entry *lobj;
 473         int r;
 474
 475         list_for_each_entry(lobj, validated, tv.head) {
 476                 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(lobj->tv.bo);
 477                 struct mm_struct *usermm;
 478
 479                 usermm = amdgpu_ttm_tt_get_usermm(bo->tbo.ttm);
 480                 if (usermm && usermm != current->mm)
 481                         return -EPERM;
 482
 483                 if (amdgpu_ttm_tt_is_userptr(bo->tbo.ttm) &&
 484                     lobj->user_invalidated && lobj->user_pages) {
 485                         amdgpu_bo_placement_from_domain(bo,
 486                                                         AMDGPU_GEM_DOMAIN_CPU);
 487                         r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 488                         if (r)
 489                                 return r;
 490
 491                         amdgpu_ttm_tt_set_user_pages(bo->tbo.ttm,
 492                                                      lobj->user_pages);
 493                 }
 494
 495                 r = amdgpu_cs_validate(p, bo);
 496                 if (r)
 497                         return r;
 498
 499                 kvfree(lobj->user_pages);
 500                 lobj->user_pages = NULL;
 501         }
 502         return 0;
 503 }
 504
 505 static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
 506                                 union drm_amdgpu_cs *cs)
 507 {
 508         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 509         struct amdgpu_vm *vm = &fpriv->vm;
 510         struct amdgpu_bo_list_entry *e;
 511         struct list_head duplicates;
 512         struct amdgpu_bo *gds;
 513         struct amdgpu_bo *gws;
 514         struct amdgpu_bo *oa;
 515         int r;
 516
 517         INIT_LIST_HEAD(&p->validated);
 518
 519         /* p->bo_list could already be assigned if AMDGPU_CHUNK_ID_BO_HANDLES is present */
 520         if (cs->in.bo_list_handle) {
 521                 if (p->bo_list)
 522                         return -EINVAL;
 523
 524                 r = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle,
 525                                        &p->bo_list);
 526                 if (r)
 527                         return r;
 528         } else if (!p->bo_list) {
 529                 /* Create a empty bo_list when no handle is provided */
 530                 r = amdgpu_bo_list_create(p->adev, p->filp, NULL, 0,
 531                                           &p->bo_list);
 532                 if (r)
 533                         return r;
 534         }
 535
 536         /* One for TTM and one for the CS job */
 537         amdgpu_bo_list_for_each_entry(e, p->bo_list)
 538                 e->tv.num_shared = 2;
 539
 540         amdgpu_bo_list_get_list(p->bo_list, &p->validated);
 541
 542         INIT_LIST_HEAD(&duplicates);
 543         amdgpu_vm_get_pd_bo(&fpriv->vm, &p->validated, &p->vm_pd);
 544
 545         if (p->uf_entry.tv.bo && !ttm_to_amdgpu_bo(p->uf_entry.tv.bo)->parent)
 546                 list_add(&p->uf_entry.tv.head, &p->validated);
 547
 548         /* Get userptr backing pages. If pages are updated after registered
 549          * in amdgpu_gem_userptr_ioctl(), amdgpu_cs_list_validate() will do
 550          * amdgpu_ttm_backend_bind() to flush and invalidate new pages
 551          */
 552         amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
 553                 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
 554                 bool userpage_invalidated = false;
 555                 int i;
 556
 557                 e->user_pages = kvmalloc_array(bo->tbo.ttm->num_pages,
 558                                         sizeof(struct page *),
 559                                         GFP_KERNEL | __GFP_ZERO);
 560                 if (!e->user_pages) {
 561                         DRM_ERROR("calloc failure\n");
 562                         return -ENOMEM;
 563                 }
 564
 565                 r = amdgpu_ttm_tt_get_user_pages(bo, e->user_pages);
 566                 if (r) {
 567                         kvfree(e->user_pages);
 568                         e->user_pages = NULL;
 569                         return r;
 570                 }
 571
 572                 for (i = 0; i < bo->tbo.ttm->num_pages; i++) {
 573                         if (bo->tbo.ttm->pages[i] != e->user_pages[i]) {
 574                                 userpage_invalidated = true;
 575                                 break;
 576                         }
 577                 }
 578                 e->user_invalidated = userpage_invalidated;
 579         }
 580
 581         r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true,
 582                                    &duplicates);
 583         if (unlikely(r != 0)) {
 584                 if (r != -ERESTARTSYS)
 585                         DRM_ERROR("ttm_eu_reserve_buffers failed.\n");
 586                 goto out;
 587         }
 588
 589         amdgpu_cs_get_threshold_for_moves(p->adev, &p->bytes_moved_threshold,
 590                                           &p->bytes_moved_vis_threshold);
 591         p->bytes_moved = 0;
 592         p->bytes_moved_vis = 0;
 593
 594         r = amdgpu_vm_validate_pt_bos(p->adev, &fpriv->vm,
 595                                       amdgpu_cs_validate, p);
 596         if (r) {
 597                 DRM_ERROR("amdgpu_vm_validate_pt_bos() failed.\n");
 598                 goto error_validate;
 599         }
 600
 601         r = amdgpu_cs_list_validate(p, &duplicates);
 602         if (r)
 603                 goto error_validate;
 604
 605         r = amdgpu_cs_list_validate(p, &p->validated);
 606         if (r)
 607                 goto error_validate;
 608
 609         amdgpu_cs_report_moved_bytes(p->adev, p->bytes_moved,
 610                                      p->bytes_moved_vis);
 611
 612         gds = p->bo_list->gds_obj;
 613         gws = p->bo_list->gws_obj;
 614         oa = p->bo_list->oa_obj;
 615
 616         amdgpu_bo_list_for_each_entry(e, p->bo_list) {
 617                 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
 618
 619                 /* Make sure we use the exclusive slot for shared BOs */
 620                 if (bo->prime_shared_count)
 621                         e->tv.num_shared = 0;
 622                 e->bo_va = amdgpu_vm_bo_find(vm, bo);
 623         }
 624
 625         if (gds) {
 626                 p->job->gds_base = amdgpu_bo_gpu_offset(gds) >> PAGE_SHIFT;
 627                 p->job->gds_size = amdgpu_bo_size(gds) >> PAGE_SHIFT;
 628         }
 629         if (gws) {
 630                 p->job->gws_base = amdgpu_bo_gpu_offset(gws) >> PAGE_SHIFT;
 631                 p->job->gws_size = amdgpu_bo_size(gws) >> PAGE_SHIFT;
 632         }
 633         if (oa) {
 634                 p->job->oa_base = amdgpu_bo_gpu_offset(oa) >> PAGE_SHIFT;
 635                 p->job->oa_size = amdgpu_bo_size(oa) >> PAGE_SHIFT;
 636         }
 637
 638         if (!r && p->uf_entry.tv.bo) {
 639                 struct amdgpu_bo *uf = ttm_to_amdgpu_bo(p->uf_entry.tv.bo);
 640
 641                 r = amdgpu_ttm_alloc_gart(&uf->tbo);
 642                 p->job->uf_addr += amdgpu_bo_gpu_offset(uf);
 643         }
 644
 645 error_validate:
 646         if (r)
 647                 ttm_eu_backoff_reservation(&p->ticket, &p->validated);
 648 out:
 649         return r;
 650 }
 651
 652 static int amdgpu_cs_sync_rings(struct amdgpu_cs_parser *p)
 653 {
 654         struct amdgpu_bo_list_entry *e;
 655         int r;
 656
 657         list_for_each_entry(e, &p->validated, tv.head) {
 658                 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
 659                 struct dma_resv *resv = bo->tbo.base.resv;
 660
 661                 r = amdgpu_sync_resv(p->adev, &p->job->sync, resv, p->filp,
 662                                      amdgpu_bo_explicit_sync(bo));
 663
 664                 if (r)
 665                         return r;
 666         }
 667         return 0;
 668 }
 669
 670 /**
 671  * cs_parser_fini() - clean parser states
 672  * @parser:     parser structure holding parsing context.
 673  * @error:      error number
 674  *
 675  * If error is set than unvalidate buffer, otherwise just free memory
 676  * used by parsing context.
 677  **/
 678 static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error,
 679                                   bool backoff)
 680 {
 681         unsigned i;
 682
 683         if (error && backoff)
 684                 ttm_eu_backoff_reservation(&parser->ticket,
 685                                            &parser->validated);
 686
 687         for (i = 0; i < parser->num_post_deps; i++) {
 688                 drm_syncobj_put(parser->post_deps[i].syncobj);
 689                 kfree(parser->post_deps[i].chain);
 690         }
 691         kfree(parser->post_deps);
 692
 693         dma_fence_put(parser->fence);
 694
 695         if (parser->ctx) {
 696                 mutex_unlock(&parser->ctx->lock);
 697                 amdgpu_ctx_put(parser->ctx);
 698         }
 699         if (parser->bo_list)
 700                 amdgpu_bo_list_put(parser->bo_list);
 701
 702         for (i = 0; i < parser->nchunks; i++)
 703                 kvfree(parser->chunks[i].kdata);
 704         kfree(parser->chunks);
 705         if (parser->job)
 706                 amdgpu_job_free(parser->job);
 707         if (parser->uf_entry.tv.bo) {
 708                 struct amdgpu_bo *uf = ttm_to_amdgpu_bo(parser->uf_entry.tv.bo);
 709
 710                 amdgpu_bo_unref(&uf);
 711         }
 712 }
 713
 714 static int amdgpu_cs_vm_handling(struct amdgpu_cs_parser *p)
 715 {
 716         struct amdgpu_ring *ring = to_amdgpu_ring(p->entity->rq->sched);
 717         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 718         struct amdgpu_device *adev = p->adev;
 719         struct amdgpu_vm *vm = &fpriv->vm;
 720         struct amdgpu_bo_list_entry *e;
 721         struct amdgpu_bo_va *bo_va;
 722         struct amdgpu_bo *bo;
 723         int r;
 724
 725         /* Only for UVD/VCE VM emulation */
 726         if (ring->funcs->parse_cs || ring->funcs->patch_cs_in_place) {
 727                 unsigned i, j;
 728
 729                 for (i = 0, j = 0; i < p->nchunks && j < p->job->num_ibs; i++) {
 730                         struct drm_amdgpu_cs_chunk_ib *chunk_ib;
 731                         struct amdgpu_bo_va_mapping *m;
 732                         struct amdgpu_bo *aobj = NULL;
 733                         struct amdgpu_cs_chunk *chunk;
 734                         uint64_t offset, va_start;
 735                         struct amdgpu_ib *ib;
 736                         uint8_t *kptr;
 737
 738                         chunk = &p->chunks[i];
 739                         ib = &p->job->ibs[j];
 740                         chunk_ib = chunk->kdata;
 741
 742                         if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
 743                                 continue;
 744
 745                         va_start = chunk_ib->va_start & AMDGPU_GMC_HOLE_MASK;
 746                         r = amdgpu_cs_find_mapping(p, va_start, &aobj, &m);
 747                         if (r) {
 748                                 DRM_ERROR("IB va_start is invalid\n");
 749                                 return r;
 750                         }
 751
 752                         if ((va_start + chunk_ib->ib_bytes) >
 753                             (m->last + 1) * AMDGPU_GPU_PAGE_SIZE) {
 754                                 DRM_ERROR("IB va_start+ib_bytes is invalid\n");
 755                                 return -EINVAL;
 756                         }
 757
 758                         /* the IB should be reserved at this point */
 759                         r = amdgpu_bo_kmap(aobj, (void **)&kptr);
 760                         if (r) {
 761                                 return r;
 762                         }
 763
 764                         offset = m->start * AMDGPU_GPU_PAGE_SIZE;
 765                         kptr += va_start - offset;
 766
 767                         if (ring->funcs->parse_cs) {
 768                                 memcpy(ib->ptr, kptr, chunk_ib->ib_bytes);
 769                                 amdgpu_bo_kunmap(aobj);
 770
 771                                 r = amdgpu_ring_parse_cs(ring, p, j);
 772                                 if (r)
 773                                         return r;
 774                         } else {
 775                                 ib->ptr = (uint32_t *)kptr;
 776                                 r = amdgpu_ring_patch_cs_in_place(ring, p, j);
 777                                 amdgpu_bo_kunmap(aobj);
 778                                 if (r)
 779                                         return r;
 780                         }
 781
 782                         j++;
 783                 }
 784         }
 785
 786         if (!p->job->vm)
 787                 return amdgpu_cs_sync_rings(p);
 788
 789
 790         r = amdgpu_vm_clear_freed(adev, vm, NULL);
 791         if (r)
 792                 return r;
 793
 794         r = amdgpu_vm_bo_update(adev, fpriv->prt_va, false);
 795         if (r)
 796                 return r;
 797
 798         r = amdgpu_sync_vm_fence(&p->job->sync, fpriv->prt_va->last_pt_update);
 799         if (r)
 800                 return r;
 801
 802         if (amdgpu_mcbp || amdgpu_sriov_vf(adev)) {
 803                 bo_va = fpriv->csa_va;
 804                 BUG_ON(!bo_va);
 805                 r = amdgpu_vm_bo_update(adev, bo_va, false);
 806                 if (r)
 807                         return r;
 808
 809                 r = amdgpu_sync_vm_fence(&p->job->sync, bo_va->last_pt_update);
 810                 if (r)
 811                         return r;
 812         }
 813
 814         amdgpu_bo_list_for_each_entry(e, p->bo_list) {
 815                 /* ignore duplicates */
 816                 bo = ttm_to_amdgpu_bo(e->tv.bo);
 817                 if (!bo)
 818                         continue;
 819
 820                 bo_va = e->bo_va;
 821                 if (bo_va == NULL)
 822                         continue;
 823
 824                 r = amdgpu_vm_bo_update(adev, bo_va, false);
 825                 if (r)
 826                         return r;
 827
 828                 r = amdgpu_sync_vm_fence(&p->job->sync, bo_va->last_pt_update);
 829                 if (r)
 830                         return r;
 831         }
 832
 833         r = amdgpu_vm_handle_moved(adev, vm);
 834         if (r)
 835                 return r;
 836
 837         r = amdgpu_vm_update_pdes(adev, vm, false);
 838         if (r)
 839                 return r;
 840
 841         r = amdgpu_sync_vm_fence(&p->job->sync, vm->last_update);
 842         if (r)
 843                 return r;
 844
 845         p->job->vm_pd_addr = amdgpu_gmc_pd_addr(vm->root.base.bo);
 846
 847         if (amdgpu_vm_debug) {
 848                 /* Invalidate all BOs to test for userspace bugs */
 849                 amdgpu_bo_list_for_each_entry(e, p->bo_list) {
 850                         struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
 851
 852                         /* ignore duplicates */
 853                         if (!bo)
 854                                 continue;
 855
 856                         amdgpu_vm_bo_invalidate(adev, bo, false);
 857                 }
 858         }
 859
 860         return amdgpu_cs_sync_rings(p);
 861 }
 862
 863 static int amdgpu_cs_ib_fill(struct amdgpu_device *adev,
 864                              struct amdgpu_cs_parser *parser)
 865 {
 866         struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
 867         struct amdgpu_vm *vm = &fpriv->vm;
 868         int r, ce_preempt = 0, de_preempt = 0;
 869         struct amdgpu_ring *ring;
 870         int i, j;
 871
 872         for (i = 0, j = 0; i < parser->nchunks && j < parser->job->num_ibs; i++) {
 873                 struct amdgpu_cs_chunk *chunk;
 874                 struct amdgpu_ib *ib;
 875                 struct drm_amdgpu_cs_chunk_ib *chunk_ib;
 876                 struct drm_sched_entity *entity;
 877
 878                 chunk = &parser->chunks[i];
 879                 ib = &parser->job->ibs[j];
 880                 chunk_ib = (struct drm_amdgpu_cs_chunk_ib *)chunk->kdata;
 881
 882                 if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
 883                         continue;
 884
 885                 if (chunk_ib->ip_type == AMDGPU_HW_IP_GFX &&
 886                     (amdgpu_mcbp || amdgpu_sriov_vf(adev))) {
 887                         if (chunk_ib->flags & AMDGPU_IB_FLAG_PREEMPT) {
 888                                 if (chunk_ib->flags & AMDGPU_IB_FLAG_CE)
 889                                         ce_preempt++;
 890                                 else
 891                                         de_preempt++;
 892                         }
 893
 894                         /* each GFX command submit allows 0 or 1 IB preemptible for CE & DE */
 895                         if (ce_preempt > 1 || de_preempt > 1)
 896                                 return -EINVAL;
 897                 }
 898
 899                 r = amdgpu_ctx_get_entity(parser->ctx, chunk_ib->ip_type,
 900                                           chunk_ib->ip_instance, chunk_ib->ring,
 901                                           &entity);
 902                 if (r)
 903                         return r;
 904
 905                 if (chunk_ib->flags & AMDGPU_IB_FLAG_PREAMBLE)
 906                         parser->job->preamble_status |=
 907                                 AMDGPU_PREAMBLE_IB_PRESENT;
 908
 909                 if (parser->entity && parser->entity != entity)
 910                         return -EINVAL;
 911
 912                 /* Return if there is no run queue associated with this entity.
 913                  * Possibly because of disabled HW IP*/
 914                 if (entity->rq == NULL)
 915                         return -EINVAL;
 916
 917                 parser->entity = entity;
 918
 919                 ring = to_amdgpu_ring(entity->rq->sched);
 920                 r =  amdgpu_ib_get(adev, vm, ring->funcs->parse_cs ?
 921                                    chunk_ib->ib_bytes : 0, ib);
 922                 if (r) {
 923                         DRM_ERROR("Failed to get ib !\n");
 924                         return r;
 925                 }
 926
 927                 ib->gpu_addr = chunk_ib->va_start;
 928                 ib->length_dw = chunk_ib->ib_bytes / 4;
 929                 ib->flags = chunk_ib->flags;
 930
 931                 j++;
 932         }
 933
 934         /* MM engine doesn't support user fences */
 935         ring = to_amdgpu_ring(parser->entity->rq->sched);
 936         if (parser->job->uf_addr && ring->funcs->no_user_fence)
 937                 return -EINVAL;
 938
 939         return amdgpu_ctx_wait_prev_fence(parser->ctx, parser->entity);
 940 }
 941
 942 static int amdgpu_cs_process_fence_dep(struct amdgpu_cs_parser *p,
 943                                        struct amdgpu_cs_chunk *chunk)
 944 {
 945         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 946         unsigned num_deps;
 947         int i, r;
 948         struct drm_amdgpu_cs_chunk_dep *deps;
 949
 950         deps = (struct drm_amdgpu_cs_chunk_dep *)chunk->kdata;
 951         num_deps = chunk->length_dw * 4 /
 952                 sizeof(struct drm_amdgpu_cs_chunk_dep);
 953
 954         for (i = 0; i < num_deps; ++i) {
 955                 struct amdgpu_ctx *ctx;
 956                 struct drm_sched_entity *entity;
 957                 struct dma_fence *fence;
 958
 959                 ctx = amdgpu_ctx_get(fpriv, deps[i].ctx_id);
 960                 if (ctx == NULL)
 961                         return -EINVAL;
 962
 963                 r = amdgpu_ctx_get_entity(ctx, deps[i].ip_type,
 964                                           deps[i].ip_instance,
 965                                           deps[i].ring, &entity);
 966                 if (r) {
 967                         amdgpu_ctx_put(ctx);
 968                         return r;
 969                 }
 970
 971                 fence = amdgpu_ctx_get_fence(ctx, entity, deps[i].handle);
 972                 amdgpu_ctx_put(ctx);
 973
 974                 if (IS_ERR(fence))
 975                         return PTR_ERR(fence);
 976                 else if (!fence)
 977                         continue;
 978
 979                 if (chunk->chunk_id == AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES) {
 980                         struct drm_sched_fence *s_fence;
 981                         struct dma_fence *old = fence;
 982
 983                         s_fence = to_drm_sched_fence(fence);
 984                         fence = dma_fence_get(&s_fence->scheduled);
 985                         dma_fence_put(old);
 986                 }
 987
 988                 r = amdgpu_sync_fence(&p->job->sync, fence, true);
 989                 dma_fence_put(fence);
 990                 if (r)
 991                         return r;
 992         }
 993         return 0;
 994 }
 995
 996 static int amdgpu_syncobj_lookup_and_add_to_sync(struct amdgpu_cs_parser *p,
 997                                                  uint32_t handle, u64 point,
 998                                                  u64 flags)
 999 {
1000         struct dma_fence *fence;
1001         int r;
1002
1003         r = drm_syncobj_find_fence(p->filp, handle, point, flags, &fence);
1004         if (r) {
1005                 DRM_ERROR("syncobj %u failed to find fence @ %llu (%d)!\n",
1006                           handle, point, r);
1007                 return r;
1008         }
1009
1010         r = amdgpu_sync_fence(&p->job->sync, fence, true);
1011         dma_fence_put(fence);
1012
1013         return r;
1014 }
1015
1016 static int amdgpu_cs_process_syncobj_in_dep(struct amdgpu_cs_parser *p,
1017                                             struct amdgpu_cs_chunk *chunk)
1018 {
1019         struct drm_amdgpu_cs_chunk_sem *deps;
1020         unsigned num_deps;
1021         int i, r;
1022
1023         deps = (struct drm_amdgpu_cs_chunk_sem *)chunk->kdata;
1024         num_deps = chunk->length_dw * 4 /
1025                 sizeof(struct drm_amdgpu_cs_chunk_sem);
1026         for (i = 0; i < num_deps; ++i) {
1027                 r = amdgpu_syncobj_lookup_and_add_to_sync(p, deps[i].handle,
1028                                                           0, 0);
1029                 if (r)
1030                         return r;
1031         }
1032
1033         return 0;
1034 }
1035
1036
1037 static int amdgpu_cs_process_syncobj_timeline_in_dep(struct amdgpu_cs_parser *p,
1038                                                      struct amdgpu_cs_chunk *chunk)
1039 {
1040         struct drm_amdgpu_cs_chunk_syncobj *syncobj_deps;
1041         unsigned num_deps;
1042         int i, r;
1043
1044         syncobj_deps = (struct drm_amdgpu_cs_chunk_syncobj *)chunk->kdata;
1045         num_deps = chunk->length_dw * 4 /
1046                 sizeof(struct drm_amdgpu_cs_chunk_syncobj);
1047         for (i = 0; i < num_deps; ++i) {
1048                 r = amdgpu_syncobj_lookup_and_add_to_sync(p,
1049                                                           syncobj_deps[i].handle,
1050                                                           syncobj_deps[i].point,
1051                                                           syncobj_deps[i].flags);
1052                 if (r)
1053                         return r;
1054         }
1055
1056         return 0;
1057 }
1058
1059 static int amdgpu_cs_process_syncobj_out_dep(struct amdgpu_cs_parser *p,
1060                                              struct amdgpu_cs_chunk *chunk)
1061 {
1062         struct drm_amdgpu_cs_chunk_sem *deps;
1063         unsigned num_deps;
1064         int i;
1065
1066         deps = (struct drm_amdgpu_cs_chunk_sem *)chunk->kdata;
1067         num_deps = chunk->length_dw * 4 /
1068                 sizeof(struct drm_amdgpu_cs_chunk_sem);
1069
1070         if (p->post_deps)
1071                 return -EINVAL;
1072
1073         p->post_deps = kmalloc_array(num_deps, sizeof(*p->post_deps),
1074                                      GFP_KERNEL);
1075         p->num_post_deps = 0;
1076
1077         if (!p->post_deps)
1078                 return -ENOMEM;
1079
1080
1081         for (i = 0; i < num_deps; ++i) {
1082                 p->post_deps[i].syncobj =
1083                         drm_syncobj_find(p->filp, deps[i].handle);
1084                 if (!p->post_deps[i].syncobj)
1085                         return -EINVAL;
1086                 p->post_deps[i].chain = NULL;
1087                 p->post_deps[i].point = 0;
1088                 p->num_post_deps++;
1089         }
1090
1091         return 0;
1092 }
1093
1094
1095 static int amdgpu_cs_process_syncobj_timeline_out_dep(struct amdgpu_cs_parser *p,
1096                                                       struct amdgpu_cs_chunk *chunk)
1097 {
1098         struct drm_amdgpu_cs_chunk_syncobj *syncobj_deps;
1099         unsigned num_deps;
1100         int i;
1101
1102         syncobj_deps = (struct drm_amdgpu_cs_chunk_syncobj *)chunk->kdata;
1103         num_deps = chunk->length_dw * 4 /
1104                 sizeof(struct drm_amdgpu_cs_chunk_syncobj);
1105
1106         if (p->post_deps)
1107                 return -EINVAL;
1108
1109         p->post_deps = kmalloc_array(num_deps, sizeof(*p->post_deps),
1110                                      GFP_KERNEL);
1111         p->num_post_deps = 0;
1112
1113         if (!p->post_deps)
1114                 return -ENOMEM;
1115
1116         for (i = 0; i < num_deps; ++i) {
1117                 struct amdgpu_cs_post_dep *dep = &p->post_deps[i];
1118
1119                 dep->chain = NULL;
1120                 if (syncobj_deps[i].point) {
1121                         dep->chain = kmalloc(sizeof(*dep->chain), GFP_KERNEL);
1122                         if (!dep->chain)
1123                                 return -ENOMEM;
1124                 }
1125
1126                 dep->syncobj = drm_syncobj_find(p->filp,
1127                                                 syncobj_deps[i].handle);
1128                 if (!dep->syncobj) {
1129                         kfree(dep->chain);
1130                         return -EINVAL;
1131                 }
1132                 dep->point = syncobj_deps[i].point;
1133                 p->num_post_deps++;
1134         }
1135
1136         return 0;
1137 }
1138
1139 static int amdgpu_cs_dependencies(struct amdgpu_device *adev,
1140                                   struct amdgpu_cs_parser *p)
1141 {
1142         int i, r;
1143
1144         for (i = 0; i < p->nchunks; ++i) {
1145                 struct amdgpu_cs_chunk *chunk;
1146
1147                 chunk = &p->chunks[i];
1148
1149                 switch (chunk->chunk_id) {
1150                 case AMDGPU_CHUNK_ID_DEPENDENCIES:
1151                 case AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES:
1152                         r = amdgpu_cs_process_fence_dep(p, chunk);
1153                         if (r)
1154                                 return r;
1155                         break;
1156                 case AMDGPU_CHUNK_ID_SYNCOBJ_IN:
1157                         r = amdgpu_cs_process_syncobj_in_dep(p, chunk);
1158                         if (r)
1159                                 return r;
1160                         break;
1161                 case AMDGPU_CHUNK_ID_SYNCOBJ_OUT:
1162                         r = amdgpu_cs_process_syncobj_out_dep(p, chunk);
1163                         if (r)
1164                                 return r;
1165                         break;
1166                 case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_WAIT:
1167                         r = amdgpu_cs_process_syncobj_timeline_in_dep(p, chunk);
1168                         if (r)
1169                                 return r;
1170                         break;
1171                 case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_SIGNAL:
1172                         r = amdgpu_cs_process_syncobj_timeline_out_dep(p, chunk);
1173                         if (r)
1174                                 return r;
1175                         break;
1176                 }
1177         }
1178
1179         return 0;
1180 }
1181
1182 static void amdgpu_cs_post_dependencies(struct amdgpu_cs_parser *p)
1183 {
1184         int i;
1185
1186         for (i = 0; i < p->num_post_deps; ++i) {
1187                 if (p->post_deps[i].chain && p->post_deps[i].point) {
1188                         drm_syncobj_add_point(p->post_deps[i].syncobj,
1189                                               p->post_deps[i].chain,
1190                                               p->fence, p->post_deps[i].point);
1191                         p->post_deps[i].chain = NULL;
1192                 } else {
1193                         drm_syncobj_replace_fence(p->post_deps[i].syncobj,
1194                                                   p->fence);
1195                 }
1196         }
1197 }
1198
1199 static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
1200                             union drm_amdgpu_cs *cs)
1201 {
1202         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
1203         struct drm_sched_entity *entity = p->entity;
1204         enum drm_sched_priority priority;
1205         struct amdgpu_ring *ring;
1206         struct amdgpu_bo_list_entry *e;
1207         struct amdgpu_job *job;
1208         uint64_t seq;
1209         int r;
1210
1211         job = p->job;
1212         p->job = NULL;
1213
1214         r = drm_sched_job_init(&job->base, entity, p->filp);
1215         if (r)
1216                 goto error_unlock;
1217
1218         /* No memory allocation is allowed while holding the notifier lock.
1219          * The lock is held until amdgpu_cs_submit is finished and fence is
1220          * added to BOs.
1221          */
1222         mutex_lock(&p->adev->notifier_lock);
1223
1224         /* If userptr are invalidated after amdgpu_cs_parser_bos(), return
1225          * -EAGAIN, drmIoctl in libdrm will restart the amdgpu_cs_ioctl.
1226          */
1227         amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
1228                 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
1229
1230                 r |= !amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm);
1231         }
1232         if (r) {
1233                 r = -EAGAIN;
1234                 goto error_abort;
1235         }
1236
1237         p->fence = dma_fence_get(&job->base.s_fence->finished);
1238
1239         amdgpu_ctx_add_fence(p->ctx, entity, p->fence, &seq);
1240         amdgpu_cs_post_dependencies(p);
1241
1242         if ((job->preamble_status & AMDGPU_PREAMBLE_IB_PRESENT) &&
1243             !p->ctx->preamble_presented) {
1244                 job->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT_FIRST;
1245                 p->ctx->preamble_presented = true;
1246         }
1247
1248         cs->out.handle = seq;
1249         job->uf_sequence = seq;
1250
1251         amdgpu_job_free_resources(job);
1252
1253         trace_amdgpu_cs_ioctl(job);
1254         amdgpu_vm_bo_trace_cs(&fpriv->vm, &p->ticket);
1255         priority = job->base.s_priority;
1256         drm_sched_entity_push_job(&job->base, entity);
1257
1258         ring = to_amdgpu_ring(entity->rq->sched);
1259         amdgpu_ring_priority_get(ring, priority);
1260
1261         amdgpu_vm_move_to_lru_tail(p->adev, &fpriv->vm);
1262
1263         ttm_eu_fence_buffer_objects(&p->ticket, &p->validated, p->fence);
1264         mutex_unlock(&p->adev->notifier_lock);
1265
1266         return 0;
1267
1268 error_abort:
1269         drm_sched_job_cleanup(&job->base);
1270         mutex_unlock(&p->adev->notifier_lock);
1271
1272 error_unlock:
1273         amdgpu_job_free(job);
1274         return r;
1275 }
1276
1277 int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
1278 {
1279         struct amdgpu_device *adev = dev->dev_private;
1280         union drm_amdgpu_cs *cs = data;
1281         struct amdgpu_cs_parser parser = {};
1282         bool reserved_buffers = false;
1283         int i, r;
1284
1285         if (amdgpu_ras_intr_triggered())
1286                 return -EHWPOISON;
1287
1288         if (!adev->accel_working)
1289                 return -EBUSY;
1290
1291         parser.adev = adev;
1292         parser.filp = filp;
1293
1294         r = amdgpu_cs_parser_init(&parser, data);
1295         if (r) {
1296                 DRM_ERROR("Failed to initialize parser %d!\n", r);
1297                 goto out;
1298         }
1299
1300         r = amdgpu_cs_ib_fill(adev, &parser);
1301         if (r)
1302                 goto out;
1303
1304         r = amdgpu_cs_dependencies(adev, &parser);
1305         if (r) {
1306                 DRM_ERROR("Failed in the dependencies handling %d!\n", r);
1307                 goto out;
1308         }
1309
1310         r = amdgpu_cs_parser_bos(&parser, data);
1311         if (r) {
1312                 if (r == -ENOMEM)
1313                         DRM_ERROR("Not enough memory for command submission!\n");
1314                 else if (r != -ERESTARTSYS && r != -EAGAIN)
1315                         DRM_ERROR("Failed to process the buffer list %d!\n", r);
1316                 goto out;
1317         }
1318
1319         reserved_buffers = true;
1320
1321         for (i = 0; i < parser.job->num_ibs; i++)
1322                 trace_amdgpu_cs(&parser, i);
1323
1324         r = amdgpu_cs_vm_handling(&parser);
1325         if (r)
1326                 goto out;
1327
1328         r = amdgpu_cs_submit(&parser, cs);
1329
1330 out:
1331         amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
1332
1333         return r;
1334 }
1335
1336 /**
1337  * amdgpu_cs_wait_ioctl - wait for a command submission to finish
1338  *
1339  * @dev: drm device
1340  * @data: data from userspace
1341  * @filp: file private
1342  *
1343  * Wait for the command submission identified by handle to finish.
1344  */
1345 int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data,
1346                          struct drm_file *filp)
1347 {
1348         union drm_amdgpu_wait_cs *wait = data;
1349         unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
1350         struct drm_sched_entity *entity;
1351         struct amdgpu_ctx *ctx;
1352         struct dma_fence *fence;
1353         long r;
1354
1355         ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
1356         if (ctx == NULL)
1357                 return -EINVAL;
1358
1359         r = amdgpu_ctx_get_entity(ctx, wait->in.ip_type, wait->in.ip_instance,
1360                                   wait->in.ring, &entity);
1361         if (r) {
1362                 amdgpu_ctx_put(ctx);
1363                 return r;
1364         }
1365
1366         fence = amdgpu_ctx_get_fence(ctx, entity, wait->in.handle);
1367         if (IS_ERR(fence))
1368                 r = PTR_ERR(fence);
1369         else if (fence) {
1370                 r = dma_fence_wait_timeout(fence, true, timeout);
1371                 if (r > 0 && fence->error)
1372                         r = fence->error;
1373                 dma_fence_put(fence);
1374         } else
1375                 r = 1;
1376
1377         amdgpu_ctx_put(ctx);
1378         if (r < 0)
1379                 return r;
1380
1381         memset(wait, 0, sizeof(*wait));
1382         wait->out.status = (r == 0);
1383
1384         return 0;
1385 }
1386
1387 /**
1388  * amdgpu_cs_get_fence - helper to get fence from drm_amdgpu_fence
1389  *
1390  * @adev: amdgpu device
1391  * @filp: file private
1392  * @user: drm_amdgpu_fence copied from user space
1393  */
1394 static struct dma_fence *amdgpu_cs_get_fence(struct amdgpu_device *adev,
1395                                              struct drm_file *filp,
1396                                              struct drm_amdgpu_fence *user)
1397 {
1398         struct drm_sched_entity *entity;
1399         struct amdgpu_ctx *ctx;
1400         struct dma_fence *fence;
1401         int r;
1402
1403         ctx = amdgpu_ctx_get(filp->driver_priv, user->ctx_id);
1404         if (ctx == NULL)
1405                 return ERR_PTR(-EINVAL);
1406
1407         r = amdgpu_ctx_get_entity(ctx, user->ip_type, user->ip_instance,
1408                                   user->ring, &entity);
1409         if (r) {
1410                 amdgpu_ctx_put(ctx);
1411                 return ERR_PTR(r);
1412         }
1413
1414         fence = amdgpu_ctx_get_fence(ctx, entity, user->seq_no);
1415         amdgpu_ctx_put(ctx);
1416
1417         return fence;
1418 }
1419
1420 int amdgpu_cs_fence_to_handle_ioctl(struct drm_device *dev, void *data,
1421                                     struct drm_file *filp)
1422 {
1423         struct amdgpu_device *adev = dev->dev_private;
1424         union drm_amdgpu_fence_to_handle *info = data;
1425         struct dma_fence *fence;
1426         struct drm_syncobj *syncobj;
1427         struct sync_file *sync_file;
1428         int fd, r;
1429
1430         fence = amdgpu_cs_get_fence(adev, filp, &info->in.fence);
1431         if (IS_ERR(fence))
1432                 return PTR_ERR(fence);
1433
1434         if (!fence)
1435                 fence = dma_fence_get_stub();
1436
1437         switch (info->in.what) {
1438         case AMDGPU_FENCE_TO_HANDLE_GET_SYNCOBJ:
1439                 r = drm_syncobj_create(&syncobj, 0, fence);
1440                 dma_fence_put(fence);
1441                 if (r)
1442                         return r;
1443                 r = drm_syncobj_get_handle(filp, syncobj, &info->out.handle);
1444                 drm_syncobj_put(syncobj);
1445                 return r;
1446
1447         case AMDGPU_FENCE_TO_HANDLE_GET_SYNCOBJ_FD:
1448                 r = drm_syncobj_create(&syncobj, 0, fence);
1449                 dma_fence_put(fence);
1450                 if (r)
1451                         return r;
1452                 r = drm_syncobj_get_fd(syncobj, (int*)&info->out.handle);
1453                 drm_syncobj_put(syncobj);
1454                 return r;
1455
1456         case AMDGPU_FENCE_TO_HANDLE_GET_SYNC_FILE_FD:
1457                 fd = get_unused_fd_flags(O_CLOEXEC);
1458                 if (fd < 0) {
1459                         dma_fence_put(fence);
1460                         return fd;
1461                 }
1462
1463                 sync_file = sync_file_create(fence);
1464                 dma_fence_put(fence);
1465                 if (!sync_file) {
1466                         put_unused_fd(fd);
1467                         return -ENOMEM;
1468                 }
1469
1470                 fd_install(fd, sync_file->file);
1471                 info->out.handle = fd;
1472                 return 0;
1473
1474         default:
1475                 return -EINVAL;
1476         }
1477 }
1478
1479 /**
1480  * amdgpu_cs_wait_all_fence - wait on all fences to signal
1481  *
1482  * @adev: amdgpu device
1483  * @filp: file private
1484  * @wait: wait parameters
1485  * @fences: array of drm_amdgpu_fence
1486  */
1487 static int amdgpu_cs_wait_all_fences(struct amdgpu_device *adev,
1488                                      struct drm_file *filp,
1489                                      union drm_amdgpu_wait_fences *wait,
1490                                      struct drm_amdgpu_fence *fences)
1491 {
1492         uint32_t fence_count = wait->in.fence_count;
1493         unsigned int i;
1494         long r = 1;
1495
1496         for (i = 0; i < fence_count; i++) {
1497                 struct dma_fence *fence;
1498                 unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1499
1500                 fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1501                 if (IS_ERR(fence))
1502                         return PTR_ERR(fence);
1503                 else if (!fence)
1504                         continue;
1505
1506                 r = dma_fence_wait_timeout(fence, true, timeout);
1507                 dma_fence_put(fence);
1508                 if (r < 0)
1509                         return r;
1510
1511                 if (r == 0)
1512                         break;
1513
1514                 if (fence->error)
1515                         return fence->error;
1516         }
1517
1518         memset(wait, 0, sizeof(*wait));
1519         wait->out.status = (r > 0);
1520
1521         return 0;
1522 }
1523
1524 /**
1525  * amdgpu_cs_wait_any_fence - wait on any fence to signal
1526  *
1527  * @adev: amdgpu device
1528  * @filp: file private
1529  * @wait: wait parameters
1530  * @fences: array of drm_amdgpu_fence
1531  */
1532 static int amdgpu_cs_wait_any_fence(struct amdgpu_device *adev,
1533                                     struct drm_file *filp,
1534                                     union drm_amdgpu_wait_fences *wait,
1535                                     struct drm_amdgpu_fence *fences)
1536 {
1537         unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1538         uint32_t fence_count = wait->in.fence_count;
1539         uint32_t first = ~0;
1540         struct dma_fence **array;
1541         unsigned int i;
1542         long r;
1543
1544         /* Prepare the fence array */
1545         array = kcalloc(fence_count, sizeof(struct dma_fence *), GFP_KERNEL);
1546
1547         if (array == NULL)
1548                 return -ENOMEM;
1549
1550         for (i = 0; i < fence_count; i++) {
1551                 struct dma_fence *fence;
1552
1553                 fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1554                 if (IS_ERR(fence)) {
1555                         r = PTR_ERR(fence);
1556                         goto err_free_fence_array;
1557                 } else if (fence) {
1558                         array[i] = fence;
1559                 } else { /* NULL, the fence has been already signaled */
1560                         r = 1;
1561                         first = i;
1562                         goto out;
1563                 }
1564         }
1565
1566         r = dma_fence_wait_any_timeout(array, fence_count, true, timeout,
1567                                        &first);
1568         if (r < 0)
1569                 goto err_free_fence_array;
1570
1571 out:
1572         memset(wait, 0, sizeof(*wait));
1573         wait->out.status = (r > 0);
1574         wait->out.first_signaled = first;
1575
1576         if (first < fence_count && array[first])
1577                 r = array[first]->error;
1578         else
1579                 r = 0;
1580
1581 err_free_fence_array:
1582         for (i = 0; i < fence_count; i++)
1583                 dma_fence_put(array[i]);
1584         kfree(array);
1585
1586         return r;
1587 }
1588
1589 /**
1590  * amdgpu_cs_wait_fences_ioctl - wait for multiple command submissions to finish
1591  *
1592  * @dev: drm device
1593  * @data: data from userspace
1594  * @filp: file private
1595  */
1596 int amdgpu_cs_wait_fences_ioctl(struct drm_device *dev, void *data,
1597                                 struct drm_file *filp)
1598 {
1599         struct amdgpu_device *adev = dev->dev_private;
1600         union drm_amdgpu_wait_fences *wait = data;
1601         uint32_t fence_count = wait->in.fence_count;
1602         struct drm_amdgpu_fence *fences_user;
1603         struct drm_amdgpu_fence *fences;
1604         int r;
1605
1606         /* Get the fences from userspace */
1607         fences = kmalloc_array(fence_count, sizeof(struct drm_amdgpu_fence),
1608                         GFP_KERNEL);
1609         if (fences == NULL)
1610                 return -ENOMEM;
1611
1612         fences_user = u64_to_user_ptr(wait->in.fences);
1613         if (copy_from_user(fences, fences_user,
1614                 sizeof(struct drm_amdgpu_fence) * fence_count)) {
1615                 r = -EFAULT;
1616                 goto err_free_fences;
1617         }
1618
1619         if (wait->in.wait_all)
1620                 r = amdgpu_cs_wait_all_fences(adev, filp, wait, fences);
1621         else
1622                 r = amdgpu_cs_wait_any_fence(adev, filp, wait, fences);
1623
1624 err_free_fences:
1625         kfree(fences);
1626
1627         return r;
1628 }
1629
1630 /**
1631  * amdgpu_cs_find_bo_va - find bo_va for VM address
1632  *
1633  * @parser: command submission parser context
1634  * @addr: VM address
1635  * @bo: resulting BO of the mapping found
1636  *
1637  * Search the buffer objects in the command submission context for a certain
1638  * virtual memory address. Returns allocation structure when found, NULL
1639  * otherwise.
1640  */
1641 int amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
1642                            uint64_t addr, struct amdgpu_bo **bo,
1643                            struct amdgpu_bo_va_mapping **map)
1644 {
1645         struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
1646         struct ttm_operation_ctx ctx = { false, false };
1647         struct amdgpu_vm *vm = &fpriv->vm;
1648         struct amdgpu_bo_va_mapping *mapping;
1649         int r;
1650
1651         addr /= AMDGPU_GPU_PAGE_SIZE;
1652
1653         mapping = amdgpu_vm_bo_lookup_mapping(vm, addr);
1654         if (!mapping || !mapping->bo_va || !mapping->bo_va->base.bo)
1655                 return -EINVAL;
1656
1657         *bo = mapping->bo_va->base.bo;
1658         *map = mapping;
1659
1660         /* Double check that the BO is reserved by this CS */
1661         if (dma_resv_locking_ctx((*bo)->tbo.base.resv) != &parser->ticket)
1662                 return -EINVAL;
1663
1664         if (!((*bo)->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)) {
1665                 (*bo)->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
1666                 amdgpu_bo_placement_from_domain(*bo, (*bo)->allowed_domains);
1667                 r = ttm_bo_validate(&(*bo)->tbo, &(*bo)->placement, &ctx);
1668                 if (r)
1669                         return r;
1670         }
1671
1672         return amdgpu_ttm_alloc_gart(&(*bo)->tbo);
1673 }