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 #include <linux/dma-buf.h>
  32
  33 #include <drm/amdgpu_drm.h>
  34 #include <drm/drm_syncobj.h>
  35 #include "amdgpu_cs.h"
  36 #include "amdgpu.h"
  37 #include "amdgpu_trace.h"
  38 #include "amdgpu_gmc.h"
  39 #include "amdgpu_gem.h"
  40 #include "amdgpu_ras.h"
  41
  42 static int amdgpu_cs_user_fence_chunk(struct amdgpu_cs_parser *p,
  43                                       struct drm_amdgpu_cs_chunk_fence *data,
  44                                       uint32_t *offset)
  45 {
  46         struct drm_gem_object *gobj;
  47         struct amdgpu_bo *bo;
  48         unsigned long size;
  49         int r;
  50
  51         gobj = drm_gem_object_lookup(p->filp, data->handle);
  52         if (gobj == NULL)
  53                 return -EINVAL;
  54
  55         bo = amdgpu_bo_ref(gem_to_amdgpu_bo(gobj));
  56         p->uf_entry.priority = 0;
  57         p->uf_entry.tv.bo = &bo->tbo;
  58         /* One for TTM and two for the CS job */
  59         p->uf_entry.tv.num_shared = 3;
  60
  61         drm_gem_object_put(gobj);
  62
  63         size = amdgpu_bo_size(bo);
  64         if (size != PAGE_SIZE || (data->offset + 8) > size) {
  65                 r = -EINVAL;
  66                 goto error_unref;
  67         }
  68
  69         if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
  70                 r = -EINVAL;
  71                 goto error_unref;
  72         }
  73
  74         *offset = data->offset;
  75
  76         return 0;
  77
  78 error_unref:
  79         amdgpu_bo_unref(&bo);
  80         return r;
  81 }
  82
  83 static int amdgpu_cs_bo_handles_chunk(struct amdgpu_cs_parser *p,
  84                                       struct drm_amdgpu_bo_list_in *data)
  85 {
  86         int r;
  87         struct drm_amdgpu_bo_list_entry *info = NULL;
  88
  89         r = amdgpu_bo_create_list_entry_array(data, &info);
  90         if (r)
  91                 return r;
  92
  93         r = amdgpu_bo_list_create(p->adev, p->filp, info, data->bo_number,
  94                                   &p->bo_list);
  95         if (r)
  96                 goto error_free;
  97
  98         kvfree(info);
  99         return 0;
 100
 101 error_free:
 102         kvfree(info);
 103
 104         return r;
 105 }
 106
 107 static int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, union drm_amdgpu_cs *cs)
 108 {
 109         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 110         struct amdgpu_vm *vm = &fpriv->vm;
 111         uint64_t *chunk_array_user;
 112         uint64_t *chunk_array;
 113         unsigned size, num_ibs = 0;
 114         uint32_t uf_offset = 0;
 115         int i;
 116         int ret;
 117
 118         if (cs->in.num_chunks == 0)
 119                 return -EINVAL;
 120
 121         chunk_array = kvmalloc_array(cs->in.num_chunks, sizeof(uint64_t), GFP_KERNEL);
 122         if (!chunk_array)
 123                 return -ENOMEM;
 124
 125         p->ctx = amdgpu_ctx_get(fpriv, cs->in.ctx_id);
 126         if (!p->ctx) {
 127                 ret = -EINVAL;
 128                 goto free_chunk;
 129         }
 130
 131         mutex_lock(&p->ctx->lock);
 132
 133         /* skip guilty context job */
 134         if (atomic_read(&p->ctx->guilty) == 1) {
 135                 ret = -ECANCELED;
 136                 goto free_chunk;
 137         }
 138
 139         /* get chunks */
 140         chunk_array_user = u64_to_user_ptr(cs->in.chunks);
 141         if (copy_from_user(chunk_array, chunk_array_user,
 142                            sizeof(uint64_t)*cs->in.num_chunks)) {
 143                 ret = -EFAULT;
 144                 goto free_chunk;
 145         }
 146
 147         p->nchunks = cs->in.num_chunks;
 148         p->chunks = kvmalloc_array(p->nchunks, sizeof(struct amdgpu_cs_chunk),
 149                             GFP_KERNEL);
 150         if (!p->chunks) {
 151                 ret = -ENOMEM;
 152                 goto free_chunk;
 153         }
 154
 155         for (i = 0; i < p->nchunks; i++) {
 156                 struct drm_amdgpu_cs_chunk __user **chunk_ptr = NULL;
 157                 struct drm_amdgpu_cs_chunk user_chunk;
 158                 uint32_t __user *cdata;
 159
 160                 chunk_ptr = u64_to_user_ptr(chunk_array[i]);
 161                 if (copy_from_user(&user_chunk, chunk_ptr,
 162                                        sizeof(struct drm_amdgpu_cs_chunk))) {
 163                         ret = -EFAULT;
 164                         i--;
 165                         goto free_partial_kdata;
 166                 }
 167                 p->chunks[i].chunk_id = user_chunk.chunk_id;
 168                 p->chunks[i].length_dw = user_chunk.length_dw;
 169
 170                 size = p->chunks[i].length_dw;
 171                 cdata = u64_to_user_ptr(user_chunk.chunk_data);
 172
 173                 p->chunks[i].kdata = kvmalloc_array(size, sizeof(uint32_t), GFP_KERNEL);
 174                 if (p->chunks[i].kdata == NULL) {
 175                         ret = -ENOMEM;
 176                         i--;
 177                         goto free_partial_kdata;
 178                 }
 179                 size *= sizeof(uint32_t);
 180                 if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
 181                         ret = -EFAULT;
 182                         goto free_partial_kdata;
 183                 }
 184
 185                 switch (p->chunks[i].chunk_id) {
 186                 case AMDGPU_CHUNK_ID_IB:
 187                         ++num_ibs;
 188                         break;
 189
 190                 case AMDGPU_CHUNK_ID_FENCE:
 191                         size = sizeof(struct drm_amdgpu_cs_chunk_fence);
 192                         if (p->chunks[i].length_dw * sizeof(uint32_t) < size) {
 193                                 ret = -EINVAL;
 194                                 goto free_partial_kdata;
 195                         }
 196
 197                         ret = amdgpu_cs_user_fence_chunk(p, p->chunks[i].kdata,
 198                                                          &uf_offset);
 199                         if (ret)
 200                                 goto free_partial_kdata;
 201
 202                         break;
 203
 204                 case AMDGPU_CHUNK_ID_BO_HANDLES:
 205                         size = sizeof(struct drm_amdgpu_bo_list_in);
 206                         if (p->chunks[i].length_dw * sizeof(uint32_t) < size) {
 207                                 ret = -EINVAL;
 208                                 goto free_partial_kdata;
 209                         }
 210
 211                         ret = amdgpu_cs_bo_handles_chunk(p, p->chunks[i].kdata);
 212                         if (ret)
 213                                 goto free_partial_kdata;
 214
 215                         break;
 216
 217                 case AMDGPU_CHUNK_ID_DEPENDENCIES:
 218                 case AMDGPU_CHUNK_ID_SYNCOBJ_IN:
 219                 case AMDGPU_CHUNK_ID_SYNCOBJ_OUT:
 220                 case AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES:
 221                 case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_WAIT:
 222                 case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_SIGNAL:
 223                         break;
 224
 225                 default:
 226                         ret = -EINVAL;
 227                         goto free_partial_kdata;
 228                 }
 229         }
 230
 231         ret = amdgpu_job_alloc(p->adev, num_ibs, &p->job, vm);
 232         if (ret)
 233                 goto free_all_kdata;
 234
 235         if (p->ctx->vram_lost_counter != p->job->vram_lost_counter) {
 236                 ret = -ECANCELED;
 237                 goto free_all_kdata;
 238         }
 239
 240         if (p->uf_entry.tv.bo)
 241                 p->job->uf_addr = uf_offset;
 242         kvfree(chunk_array);
 243
 244         /* Use this opportunity to fill in task info for the vm */
 245         amdgpu_vm_set_task_info(vm);
 246
 247         return 0;
 248
 249 free_all_kdata:
 250         i = p->nchunks - 1;
 251 free_partial_kdata:
 252         for (; i >= 0; i--)
 253                 kvfree(p->chunks[i].kdata);
 254         kvfree(p->chunks);
 255         p->chunks = NULL;
 256         p->nchunks = 0;
 257 free_chunk:
 258         kvfree(chunk_array);
 259
 260         return ret;
 261 }
 262
 263 /* Convert microseconds to bytes. */
 264 static u64 us_to_bytes(struct amdgpu_device *adev, s64 us)
 265 {
 266         if (us <= 0 || !adev->mm_stats.log2_max_MBps)
 267                 return 0;
 268
 269         /* Since accum_us is incremented by a million per second, just
 270          * multiply it by the number of MB/s to get the number of bytes.
 271          */
 272         return us << adev->mm_stats.log2_max_MBps;
 273 }
 274
 275 static s64 bytes_to_us(struct amdgpu_device *adev, u64 bytes)
 276 {
 277         if (!adev->mm_stats.log2_max_MBps)
 278                 return 0;
 279
 280         return bytes >> adev->mm_stats.log2_max_MBps;
 281 }
 282
 283 /* Returns how many bytes TTM can move right now. If no bytes can be moved,
 284  * it returns 0. If it returns non-zero, it's OK to move at least one buffer,
 285  * which means it can go over the threshold once. If that happens, the driver
 286  * will be in debt and no other buffer migrations can be done until that debt
 287  * is repaid.
 288  *
 289  * This approach allows moving a buffer of any size (it's important to allow
 290  * that).
 291  *
 292  * The currency is simply time in microseconds and it increases as the clock
 293  * ticks. The accumulated microseconds (us) are converted to bytes and
 294  * returned.
 295  */
 296 static void amdgpu_cs_get_threshold_for_moves(struct amdgpu_device *adev,
 297                                               u64 *max_bytes,
 298                                               u64 *max_vis_bytes)
 299 {
 300         s64 time_us, increment_us;
 301         u64 free_vram, total_vram, used_vram;
 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 = ttm_resource_manager_usage(&adev->mman.vram_mgr.manager);
 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 aggressive 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.vram_mgr);
 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(void *param, struct amdgpu_bo *bo)
 400 {
 401         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
 402         struct amdgpu_cs_parser *p = param;
 403         struct ttm_operation_ctx ctx = {
 404                 .interruptible = true,
 405                 .no_wait_gpu = false,
 406                 .resv = bo->tbo.base.resv
 407         };
 408         uint32_t domain;
 409         int r;
 410
 411         if (bo->tbo.pin_count)
 412                 return 0;
 413
 414         /* Don't move this buffer if we have depleted our allowance
 415          * to move it. Don't move anything if the threshold is zero.
 416          */
 417         if (p->bytes_moved < p->bytes_moved_threshold &&
 418             (!bo->tbo.base.dma_buf ||
 419             list_empty(&bo->tbo.base.dma_buf->attachments))) {
 420                 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
 421                     (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)) {
 422                         /* And don't move a CPU_ACCESS_REQUIRED BO to limited
 423                          * visible VRAM if we've depleted our allowance to do
 424                          * that.
 425                          */
 426                         if (p->bytes_moved_vis < p->bytes_moved_vis_threshold)
 427                                 domain = bo->preferred_domains;
 428                         else
 429                                 domain = bo->allowed_domains;
 430                 } else {
 431                         domain = bo->preferred_domains;
 432                 }
 433         } else {
 434                 domain = bo->allowed_domains;
 435         }
 436
 437 retry:
 438         amdgpu_bo_placement_from_domain(bo, domain);
 439         r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 440
 441         p->bytes_moved += ctx.bytes_moved;
 442         if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
 443             amdgpu_bo_in_cpu_visible_vram(bo))
 444                 p->bytes_moved_vis += ctx.bytes_moved;
 445
 446         if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
 447                 domain = bo->allowed_domains;
 448                 goto retry;
 449         }
 450
 451         return r;
 452 }
 453
 454 static int amdgpu_cs_list_validate(struct amdgpu_cs_parser *p,
 455                             struct list_head *validated)
 456 {
 457         struct ttm_operation_ctx ctx = { true, false };
 458         struct amdgpu_bo_list_entry *lobj;
 459         int r;
 460
 461         list_for_each_entry(lobj, validated, tv.head) {
 462                 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(lobj->tv.bo);
 463                 struct mm_struct *usermm;
 464
 465                 usermm = amdgpu_ttm_tt_get_usermm(bo->tbo.ttm);
 466                 if (usermm && usermm != current->mm)
 467                         return -EPERM;
 468
 469                 if (amdgpu_ttm_tt_is_userptr(bo->tbo.ttm) &&
 470                     lobj->user_invalidated && lobj->user_pages) {
 471                         amdgpu_bo_placement_from_domain(bo,
 472                                                         AMDGPU_GEM_DOMAIN_CPU);
 473                         r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 474                         if (r)
 475                                 return r;
 476
 477                         amdgpu_ttm_tt_set_user_pages(bo->tbo.ttm,
 478                                                      lobj->user_pages);
 479                 }
 480
 481                 r = amdgpu_cs_bo_validate(p, bo);
 482                 if (r)
 483                         return r;
 484
 485                 kvfree(lobj->user_pages);
 486                 lobj->user_pages = NULL;
 487         }
 488         return 0;
 489 }
 490
 491 static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
 492                                 union drm_amdgpu_cs *cs)
 493 {
 494         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 495         struct amdgpu_vm *vm = &fpriv->vm;
 496         struct amdgpu_bo_list_entry *e;
 497         struct list_head duplicates;
 498         struct amdgpu_bo *gds;
 499         struct amdgpu_bo *gws;
 500         struct amdgpu_bo *oa;
 501         int r;
 502
 503         INIT_LIST_HEAD(&p->validated);
 504
 505         /* p->bo_list could already be assigned if AMDGPU_CHUNK_ID_BO_HANDLES is present */
 506         if (cs->in.bo_list_handle) {
 507                 if (p->bo_list)
 508                         return -EINVAL;
 509
 510                 r = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle,
 511                                        &p->bo_list);
 512                 if (r)
 513                         return r;
 514         } else if (!p->bo_list) {
 515                 /* Create a empty bo_list when no handle is provided */
 516                 r = amdgpu_bo_list_create(p->adev, p->filp, NULL, 0,
 517                                           &p->bo_list);
 518                 if (r)
 519                         return r;
 520         }
 521
 522         /* One for TTM and one for the CS job */
 523         amdgpu_bo_list_for_each_entry(e, p->bo_list)
 524                 e->tv.num_shared = 2;
 525
 526         amdgpu_bo_list_get_list(p->bo_list, &p->validated);
 527
 528         INIT_LIST_HEAD(&duplicates);
 529         amdgpu_vm_get_pd_bo(&fpriv->vm, &p->validated, &p->vm_pd);
 530
 531         if (p->uf_entry.tv.bo && !ttm_to_amdgpu_bo(p->uf_entry.tv.bo)->parent)
 532                 list_add(&p->uf_entry.tv.head, &p->validated);
 533
 534         /* Get userptr backing pages. If pages are updated after registered
 535          * in amdgpu_gem_userptr_ioctl(), amdgpu_cs_list_validate() will do
 536          * amdgpu_ttm_backend_bind() to flush and invalidate new pages
 537          */
 538         amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
 539                 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
 540                 bool userpage_invalidated = false;
 541                 int i;
 542
 543                 e->user_pages = kvmalloc_array(bo->tbo.ttm->num_pages,
 544                                         sizeof(struct page *),
 545                                         GFP_KERNEL | __GFP_ZERO);
 546                 if (!e->user_pages) {
 547                         DRM_ERROR("kvmalloc_array failure\n");
 548                         r = -ENOMEM;
 549                         goto out_free_user_pages;
 550                 }
 551
 552                 r = amdgpu_ttm_tt_get_user_pages(bo, e->user_pages);
 553                 if (r) {
 554                         kvfree(e->user_pages);
 555                         e->user_pages = NULL;
 556                         goto out_free_user_pages;
 557                 }
 558
 559                 for (i = 0; i < bo->tbo.ttm->num_pages; i++) {
 560                         if (bo->tbo.ttm->pages[i] != e->user_pages[i]) {
 561                                 userpage_invalidated = true;
 562                                 break;
 563                         }
 564                 }
 565                 e->user_invalidated = userpage_invalidated;
 566         }
 567
 568         r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true,
 569                                    &duplicates);
 570         if (unlikely(r != 0)) {
 571                 if (r != -ERESTARTSYS)
 572                         DRM_ERROR("ttm_eu_reserve_buffers failed.\n");
 573                 goto out_free_user_pages;
 574         }
 575
 576         amdgpu_bo_list_for_each_entry(e, p->bo_list) {
 577                 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
 578
 579                 e->bo_va = amdgpu_vm_bo_find(vm, bo);
 580         }
 581
 582         /* Move fence waiting after getting reservation lock of
 583          * PD root. Then there is no need on a ctx mutex lock.
 584          */
 585         r = amdgpu_ctx_wait_prev_fence(p->ctx, p->entity);
 586         if (unlikely(r != 0)) {
 587                 if (r != -ERESTARTSYS)
 588                         DRM_ERROR("amdgpu_ctx_wait_prev_fence failed.\n");
 589                 goto error_validate;
 590         }
 591
 592         amdgpu_cs_get_threshold_for_moves(p->adev, &p->bytes_moved_threshold,
 593                                           &p->bytes_moved_vis_threshold);
 594         p->bytes_moved = 0;
 595         p->bytes_moved_vis = 0;
 596
 597         r = amdgpu_vm_validate_pt_bos(p->adev, &fpriv->vm,
 598                                       amdgpu_cs_bo_validate, p);
 599         if (r) {
 600                 DRM_ERROR("amdgpu_vm_validate_pt_bos() failed.\n");
 601                 goto error_validate;
 602         }
 603
 604         r = amdgpu_cs_list_validate(p, &duplicates);
 605         if (r)
 606                 goto error_validate;
 607
 608         r = amdgpu_cs_list_validate(p, &p->validated);
 609         if (r)
 610                 goto error_validate;
 611
 612         amdgpu_cs_report_moved_bytes(p->adev, p->bytes_moved,
 613                                      p->bytes_moved_vis);
 614
 615         gds = p->bo_list->gds_obj;
 616         gws = p->bo_list->gws_obj;
 617         oa = p->bo_list->oa_obj;
 618
 619         if (gds) {
 620                 p->job->gds_base = amdgpu_bo_gpu_offset(gds) >> PAGE_SHIFT;
 621                 p->job->gds_size = amdgpu_bo_size(gds) >> PAGE_SHIFT;
 622         }
 623         if (gws) {
 624                 p->job->gws_base = amdgpu_bo_gpu_offset(gws) >> PAGE_SHIFT;
 625                 p->job->gws_size = amdgpu_bo_size(gws) >> PAGE_SHIFT;
 626         }
 627         if (oa) {
 628                 p->job->oa_base = amdgpu_bo_gpu_offset(oa) >> PAGE_SHIFT;
 629                 p->job->oa_size = amdgpu_bo_size(oa) >> PAGE_SHIFT;
 630         }
 631
 632         if (!r && p->uf_entry.tv.bo) {
 633                 struct amdgpu_bo *uf = ttm_to_amdgpu_bo(p->uf_entry.tv.bo);
 634
 635                 r = amdgpu_ttm_alloc_gart(&uf->tbo);
 636                 p->job->uf_addr += amdgpu_bo_gpu_offset(uf);
 637         }
 638
 639 error_validate:
 640         if (r)
 641                 ttm_eu_backoff_reservation(&p->ticket, &p->validated);
 642
 643 out_free_user_pages:
 644         if (r) {
 645                 amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
 646                         struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
 647
 648                         if (!e->user_pages)
 649                                 continue;
 650                         amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm);
 651                         kvfree(e->user_pages);
 652                         e->user_pages = NULL;
 653                 }
 654         }
 655         return r;
 656 }
 657
 658 static int amdgpu_cs_sync_rings(struct amdgpu_cs_parser *p)
 659 {
 660         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 661         struct amdgpu_bo_list_entry *e;
 662         int r;
 663
 664         list_for_each_entry(e, &p->validated, tv.head) {
 665                 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
 666                 struct dma_resv *resv = bo->tbo.base.resv;
 667                 enum amdgpu_sync_mode sync_mode;
 668
 669                 sync_mode = amdgpu_bo_explicit_sync(bo) ?
 670                         AMDGPU_SYNC_EXPLICIT : AMDGPU_SYNC_NE_OWNER;
 671                 r = amdgpu_sync_resv(p->adev, &p->job->sync, resv, sync_mode,
 672                                      &fpriv->vm);
 673                 if (r)
 674                         return r;
 675         }
 676         return 0;
 677 }
 678
 679 /**
 680  * amdgpu_cs_parser_fini() - clean parser states
 681  * @parser:     parser structure holding parsing context.
 682  * @error:      error number
 683  * @backoff:    indicator to backoff the reservation
 684  *
 685  * If error is set then unvalidate buffer, otherwise just free memory
 686  * used by parsing context.
 687  **/
 688 static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error,
 689                                   bool backoff)
 690 {
 691         unsigned i;
 692
 693         if (error && backoff)
 694                 ttm_eu_backoff_reservation(&parser->ticket,
 695                                            &parser->validated);
 696
 697         for (i = 0; i < parser->num_post_deps; i++) {
 698                 drm_syncobj_put(parser->post_deps[i].syncobj);
 699                 kfree(parser->post_deps[i].chain);
 700         }
 701         kfree(parser->post_deps);
 702
 703         dma_fence_put(parser->fence);
 704
 705         if (parser->ctx) {
 706                 mutex_unlock(&parser->ctx->lock);
 707                 amdgpu_ctx_put(parser->ctx);
 708         }
 709         if (parser->bo_list)
 710                 amdgpu_bo_list_put(parser->bo_list);
 711
 712         for (i = 0; i < parser->nchunks; i++)
 713                 kvfree(parser->chunks[i].kdata);
 714         kvfree(parser->chunks);
 715         if (parser->job)
 716                 amdgpu_job_free(parser->job);
 717         if (parser->uf_entry.tv.bo) {
 718                 struct amdgpu_bo *uf = ttm_to_amdgpu_bo(parser->uf_entry.tv.bo);
 719
 720                 amdgpu_bo_unref(&uf);
 721         }
 722 }
 723
 724 static int amdgpu_cs_vm_handling(struct amdgpu_cs_parser *p)
 725 {
 726         struct amdgpu_ring *ring = to_amdgpu_ring(p->entity->rq->sched);
 727         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 728         struct amdgpu_device *adev = p->adev;
 729         struct amdgpu_vm *vm = &fpriv->vm;
 730         struct amdgpu_bo_list_entry *e;
 731         struct amdgpu_bo_va *bo_va;
 732         struct amdgpu_bo *bo;
 733         int r;
 734
 735         /* Only for UVD/VCE VM emulation */
 736         if (ring->funcs->parse_cs || ring->funcs->patch_cs_in_place) {
 737                 unsigned i, j;
 738
 739                 for (i = 0, j = 0; i < p->nchunks && j < p->job->num_ibs; i++) {
 740                         struct drm_amdgpu_cs_chunk_ib *chunk_ib;
 741                         struct amdgpu_bo_va_mapping *m;
 742                         struct amdgpu_bo *aobj = NULL;
 743                         struct amdgpu_cs_chunk *chunk;
 744                         uint64_t offset, va_start;
 745                         struct amdgpu_ib *ib;
 746                         uint8_t *kptr;
 747
 748                         chunk = &p->chunks[i];
 749                         ib = &p->job->ibs[j];
 750                         chunk_ib = chunk->kdata;
 751
 752                         if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
 753                                 continue;
 754
 755                         va_start = chunk_ib->va_start & AMDGPU_GMC_HOLE_MASK;
 756                         r = amdgpu_cs_find_mapping(p, va_start, &aobj, &m);
 757                         if (r) {
 758                                 DRM_ERROR("IB va_start is invalid\n");
 759                                 return r;
 760                         }
 761
 762                         if ((va_start + chunk_ib->ib_bytes) >
 763                             (m->last + 1) * AMDGPU_GPU_PAGE_SIZE) {
 764                                 DRM_ERROR("IB va_start+ib_bytes is invalid\n");
 765                                 return -EINVAL;
 766                         }
 767
 768                         /* the IB should be reserved at this point */
 769                         r = amdgpu_bo_kmap(aobj, (void **)&kptr);
 770                         if (r) {
 771                                 return r;
 772                         }
 773
 774                         offset = m->start * AMDGPU_GPU_PAGE_SIZE;
 775                         kptr += va_start - offset;
 776
 777                         if (ring->funcs->parse_cs) {
 778                                 memcpy(ib->ptr, kptr, chunk_ib->ib_bytes);
 779                                 amdgpu_bo_kunmap(aobj);
 780
 781                                 r = amdgpu_ring_parse_cs(ring, p, p->job, ib);
 782                                 if (r)
 783                                         return r;
 784                         } else {
 785                                 ib->ptr = (uint32_t *)kptr;
 786                                 r = amdgpu_ring_patch_cs_in_place(ring, p, p->job, ib);
 787                                 amdgpu_bo_kunmap(aobj);
 788                                 if (r)
 789                                         return r;
 790                         }
 791
 792                         j++;
 793                 }
 794         }
 795
 796         if (!p->job->vm)
 797                 return amdgpu_cs_sync_rings(p);
 798
 799
 800         r = amdgpu_vm_clear_freed(adev, vm, NULL);
 801         if (r)
 802                 return r;
 803
 804         r = amdgpu_vm_bo_update(adev, fpriv->prt_va, false);
 805         if (r)
 806                 return r;
 807
 808         r = amdgpu_sync_fence(&p->job->sync, fpriv->prt_va->last_pt_update);
 809         if (r)
 810                 return r;
 811
 812         if (amdgpu_mcbp || amdgpu_sriov_vf(adev)) {
 813                 bo_va = fpriv->csa_va;
 814                 BUG_ON(!bo_va);
 815                 r = amdgpu_vm_bo_update(adev, bo_va, false);
 816                 if (r)
 817                         return r;
 818
 819                 r = amdgpu_sync_fence(&p->job->sync, bo_va->last_pt_update);
 820                 if (r)
 821                         return r;
 822         }
 823
 824         amdgpu_bo_list_for_each_entry(e, p->bo_list) {
 825                 /* ignore duplicates */
 826                 bo = ttm_to_amdgpu_bo(e->tv.bo);
 827                 if (!bo)
 828                         continue;
 829
 830                 bo_va = e->bo_va;
 831                 if (bo_va == NULL)
 832                         continue;
 833
 834                 r = amdgpu_vm_bo_update(adev, bo_va, false);
 835                 if (r)
 836                         return r;
 837
 838                 r = amdgpu_sync_fence(&p->job->sync, bo_va->last_pt_update);
 839                 if (r)
 840                         return r;
 841         }
 842
 843         r = amdgpu_vm_handle_moved(adev, vm);
 844         if (r)
 845                 return r;
 846
 847         r = amdgpu_vm_update_pdes(adev, vm, false);
 848         if (r)
 849                 return r;
 850
 851         r = amdgpu_sync_fence(&p->job->sync, vm->last_update);
 852         if (r)
 853                 return r;
 854
 855         p->job->vm_pd_addr = amdgpu_gmc_pd_addr(vm->root.bo);
 856
 857         if (amdgpu_vm_debug) {
 858                 /* Invalidate all BOs to test for userspace bugs */
 859                 amdgpu_bo_list_for_each_entry(e, p->bo_list) {
 860                         struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
 861
 862                         /* ignore duplicates */
 863                         if (!bo)
 864                                 continue;
 865
 866                         amdgpu_vm_bo_invalidate(adev, bo, false);
 867                 }
 868         }
 869
 870         return amdgpu_cs_sync_rings(p);
 871 }
 872
 873 static int amdgpu_cs_ib_fill(struct amdgpu_device *adev,
 874                              struct amdgpu_cs_parser *parser)
 875 {
 876         struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
 877         struct amdgpu_vm *vm = &fpriv->vm;
 878         int r, ce_preempt = 0, de_preempt = 0;
 879         struct amdgpu_ring *ring;
 880         int i, j;
 881
 882         for (i = 0, j = 0; i < parser->nchunks && j < parser->job->num_ibs; i++) {
 883                 struct amdgpu_cs_chunk *chunk;
 884                 struct amdgpu_ib *ib;
 885                 struct drm_amdgpu_cs_chunk_ib *chunk_ib;
 886                 struct drm_sched_entity *entity;
 887
 888                 chunk = &parser->chunks[i];
 889                 ib = &parser->job->ibs[j];
 890                 chunk_ib = (struct drm_amdgpu_cs_chunk_ib *)chunk->kdata;
 891
 892                 if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
 893                         continue;
 894
 895                 if (chunk_ib->ip_type == AMDGPU_HW_IP_GFX &&
 896                     (amdgpu_mcbp || amdgpu_sriov_vf(adev))) {
 897                         if (chunk_ib->flags & AMDGPU_IB_FLAG_PREEMPT) {
 898                                 if (chunk_ib->flags & AMDGPU_IB_FLAG_CE)
 899                                         ce_preempt++;
 900                                 else
 901                                         de_preempt++;
 902                         }
 903
 904                         /* each GFX command submit allows 0 or 1 IB preemptible for CE & DE */
 905                         if (ce_preempt > 1 || de_preempt > 1)
 906                                 return -EINVAL;
 907                 }
 908
 909                 r = amdgpu_ctx_get_entity(parser->ctx, chunk_ib->ip_type,
 910                                           chunk_ib->ip_instance, chunk_ib->ring,
 911                                           &entity);
 912                 if (r)
 913                         return r;
 914
 915                 if (chunk_ib->flags & AMDGPU_IB_FLAG_PREAMBLE)
 916                         parser->job->preamble_status |=
 917                                 AMDGPU_PREAMBLE_IB_PRESENT;
 918
 919                 if (parser->entity && parser->entity != entity)
 920                         return -EINVAL;
 921
 922                 /* Return if there is no run queue associated with this entity.
 923                  * Possibly because of disabled HW IP*/
 924                 if (entity->rq == NULL)
 925                         return -EINVAL;
 926
 927                 parser->entity = entity;
 928
 929                 ring = to_amdgpu_ring(entity->rq->sched);
 930                 r =  amdgpu_ib_get(adev, vm, ring->funcs->parse_cs ?
 931                                    chunk_ib->ib_bytes : 0,
 932                                    AMDGPU_IB_POOL_DELAYED, ib);
 933                 if (r) {
 934                         DRM_ERROR("Failed to get ib !\n");
 935                         return r;
 936                 }
 937
 938                 ib->gpu_addr = chunk_ib->va_start;
 939                 ib->length_dw = chunk_ib->ib_bytes / 4;
 940                 ib->flags = chunk_ib->flags;
 941
 942                 j++;
 943         }
 944
 945         /* MM engine doesn't support user fences */
 946         ring = to_amdgpu_ring(parser->entity->rq->sched);
 947         if (parser->job->uf_addr && ring->funcs->no_user_fence)
 948                 return -EINVAL;
 949
 950         return 0;
 951 }
 952
 953 static int amdgpu_cs_process_fence_dep(struct amdgpu_cs_parser *p,
 954                                        struct amdgpu_cs_chunk *chunk)
 955 {
 956         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 957         unsigned num_deps;
 958         int i, r;
 959         struct drm_amdgpu_cs_chunk_dep *deps;
 960
 961         deps = (struct drm_amdgpu_cs_chunk_dep *)chunk->kdata;
 962         num_deps = chunk->length_dw * 4 /
 963                 sizeof(struct drm_amdgpu_cs_chunk_dep);
 964
 965         for (i = 0; i < num_deps; ++i) {
 966                 struct amdgpu_ctx *ctx;
 967                 struct drm_sched_entity *entity;
 968                 struct dma_fence *fence;
 969
 970                 ctx = amdgpu_ctx_get(fpriv, deps[i].ctx_id);
 971                 if (ctx == NULL)
 972                         return -EINVAL;
 973
 974                 r = amdgpu_ctx_get_entity(ctx, deps[i].ip_type,
 975                                           deps[i].ip_instance,
 976                                           deps[i].ring, &entity);
 977                 if (r) {
 978                         amdgpu_ctx_put(ctx);
 979                         return r;
 980                 }
 981
 982                 fence = amdgpu_ctx_get_fence(ctx, entity, deps[i].handle);
 983                 amdgpu_ctx_put(ctx);
 984
 985                 if (IS_ERR(fence))
 986                         return PTR_ERR(fence);
 987                 else if (!fence)
 988                         continue;
 989
 990                 if (chunk->chunk_id == AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES) {
 991                         struct drm_sched_fence *s_fence;
 992                         struct dma_fence *old = fence;
 993
 994                         s_fence = to_drm_sched_fence(fence);
 995                         fence = dma_fence_get(&s_fence->scheduled);
 996                         dma_fence_put(old);
 997                 }
 998
 999                 r = amdgpu_sync_fence(&p->job->sync, fence);
1000                 dma_fence_put(fence);
1001                 if (r)
1002                         return r;
1003         }
1004         return 0;
1005 }
1006
1007 static int amdgpu_syncobj_lookup_and_add_to_sync(struct amdgpu_cs_parser *p,
1008                                                  uint32_t handle, u64 point,
1009                                                  u64 flags)
1010 {
1011         struct dma_fence *fence;
1012         int r;
1013
1014         r = drm_syncobj_find_fence(p->filp, handle, point, flags, &fence);
1015         if (r) {
1016                 DRM_ERROR("syncobj %u failed to find fence @ %llu (%d)!\n",
1017                           handle, point, r);
1018                 return r;
1019         }
1020
1021         r = amdgpu_sync_fence(&p->job->sync, fence);
1022         dma_fence_put(fence);
1023
1024         return r;
1025 }
1026
1027 static int amdgpu_cs_process_syncobj_in_dep(struct amdgpu_cs_parser *p,
1028                                             struct amdgpu_cs_chunk *chunk)
1029 {
1030         struct drm_amdgpu_cs_chunk_sem *deps;
1031         unsigned num_deps;
1032         int i, r;
1033
1034         deps = (struct drm_amdgpu_cs_chunk_sem *)chunk->kdata;
1035         num_deps = chunk->length_dw * 4 /
1036                 sizeof(struct drm_amdgpu_cs_chunk_sem);
1037         for (i = 0; i < num_deps; ++i) {
1038                 r = amdgpu_syncobj_lookup_and_add_to_sync(p, deps[i].handle,
1039                                                           0, 0);
1040                 if (r)
1041                         return r;
1042         }
1043
1044         return 0;
1045 }
1046
1047
1048 static int amdgpu_cs_process_syncobj_timeline_in_dep(struct amdgpu_cs_parser *p,
1049                                                      struct amdgpu_cs_chunk *chunk)
1050 {
1051         struct drm_amdgpu_cs_chunk_syncobj *syncobj_deps;
1052         unsigned num_deps;
1053         int i, r;
1054
1055         syncobj_deps = (struct drm_amdgpu_cs_chunk_syncobj *)chunk->kdata;
1056         num_deps = chunk->length_dw * 4 /
1057                 sizeof(struct drm_amdgpu_cs_chunk_syncobj);
1058         for (i = 0; i < num_deps; ++i) {
1059                 r = amdgpu_syncobj_lookup_and_add_to_sync(p,
1060                                                           syncobj_deps[i].handle,
1061                                                           syncobj_deps[i].point,
1062                                                           syncobj_deps[i].flags);
1063                 if (r)
1064                         return r;
1065         }
1066
1067         return 0;
1068 }
1069
1070 static int amdgpu_cs_process_syncobj_out_dep(struct amdgpu_cs_parser *p,
1071                                              struct amdgpu_cs_chunk *chunk)
1072 {
1073         struct drm_amdgpu_cs_chunk_sem *deps;
1074         unsigned num_deps;
1075         int i;
1076
1077         deps = (struct drm_amdgpu_cs_chunk_sem *)chunk->kdata;
1078         num_deps = chunk->length_dw * 4 /
1079                 sizeof(struct drm_amdgpu_cs_chunk_sem);
1080
1081         if (p->post_deps)
1082                 return -EINVAL;
1083
1084         p->post_deps = kmalloc_array(num_deps, sizeof(*p->post_deps),
1085                                      GFP_KERNEL);
1086         p->num_post_deps = 0;
1087
1088         if (!p->post_deps)
1089                 return -ENOMEM;
1090
1091
1092         for (i = 0; i < num_deps; ++i) {
1093                 p->post_deps[i].syncobj =
1094                         drm_syncobj_find(p->filp, deps[i].handle);
1095                 if (!p->post_deps[i].syncobj)
1096                         return -EINVAL;
1097                 p->post_deps[i].chain = NULL;
1098                 p->post_deps[i].point = 0;
1099                 p->num_post_deps++;
1100         }
1101
1102         return 0;
1103 }
1104
1105
1106 static int amdgpu_cs_process_syncobj_timeline_out_dep(struct amdgpu_cs_parser *p,
1107                                                       struct amdgpu_cs_chunk *chunk)
1108 {
1109         struct drm_amdgpu_cs_chunk_syncobj *syncobj_deps;
1110         unsigned num_deps;
1111         int i;
1112
1113         syncobj_deps = (struct drm_amdgpu_cs_chunk_syncobj *)chunk->kdata;
1114         num_deps = chunk->length_dw * 4 /
1115                 sizeof(struct drm_amdgpu_cs_chunk_syncobj);
1116
1117         if (p->post_deps)
1118                 return -EINVAL;
1119
1120         p->post_deps = kmalloc_array(num_deps, sizeof(*p->post_deps),
1121                                      GFP_KERNEL);
1122         p->num_post_deps = 0;
1123
1124         if (!p->post_deps)
1125                 return -ENOMEM;
1126
1127         for (i = 0; i < num_deps; ++i) {
1128                 struct amdgpu_cs_post_dep *dep = &p->post_deps[i];
1129
1130                 dep->chain = NULL;
1131                 if (syncobj_deps[i].point) {
1132                         dep->chain = dma_fence_chain_alloc();
1133                         if (!dep->chain)
1134                                 return -ENOMEM;
1135                 }
1136
1137                 dep->syncobj = drm_syncobj_find(p->filp,
1138                                                 syncobj_deps[i].handle);
1139                 if (!dep->syncobj) {
1140                         dma_fence_chain_free(dep->chain);
1141                         return -EINVAL;
1142                 }
1143                 dep->point = syncobj_deps[i].point;
1144                 p->num_post_deps++;
1145         }
1146
1147         return 0;
1148 }
1149
1150 static int amdgpu_cs_dependencies(struct amdgpu_device *adev,
1151                                   struct amdgpu_cs_parser *p)
1152 {
1153         int i, r;
1154
1155         /* TODO: Investigate why we still need the context lock */
1156         mutex_unlock(&p->ctx->lock);
1157
1158         for (i = 0; i < p->nchunks; ++i) {
1159                 struct amdgpu_cs_chunk *chunk;
1160
1161                 chunk = &p->chunks[i];
1162
1163                 switch (chunk->chunk_id) {
1164                 case AMDGPU_CHUNK_ID_DEPENDENCIES:
1165                 case AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES:
1166                         r = amdgpu_cs_process_fence_dep(p, chunk);
1167                         if (r)
1168                                 goto out;
1169                         break;
1170                 case AMDGPU_CHUNK_ID_SYNCOBJ_IN:
1171                         r = amdgpu_cs_process_syncobj_in_dep(p, chunk);
1172                         if (r)
1173                                 goto out;
1174                         break;
1175                 case AMDGPU_CHUNK_ID_SYNCOBJ_OUT:
1176                         r = amdgpu_cs_process_syncobj_out_dep(p, chunk);
1177                         if (r)
1178                                 goto out;
1179                         break;
1180                 case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_WAIT:
1181                         r = amdgpu_cs_process_syncobj_timeline_in_dep(p, chunk);
1182                         if (r)
1183                                 goto out;
1184                         break;
1185                 case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_SIGNAL:
1186                         r = amdgpu_cs_process_syncobj_timeline_out_dep(p, chunk);
1187                         if (r)
1188                                 goto out;
1189                         break;
1190                 }
1191         }
1192
1193 out:
1194         mutex_lock(&p->ctx->lock);
1195         return r;
1196 }
1197
1198 static void amdgpu_cs_post_dependencies(struct amdgpu_cs_parser *p)
1199 {
1200         int i;
1201
1202         for (i = 0; i < p->num_post_deps; ++i) {
1203                 if (p->post_deps[i].chain && p->post_deps[i].point) {
1204                         drm_syncobj_add_point(p->post_deps[i].syncobj,
1205                                               p->post_deps[i].chain,
1206                                               p->fence, p->post_deps[i].point);
1207                         p->post_deps[i].chain = NULL;
1208                 } else {
1209                         drm_syncobj_replace_fence(p->post_deps[i].syncobj,
1210                                                   p->fence);
1211                 }
1212         }
1213 }
1214
1215 static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
1216                             union drm_amdgpu_cs *cs)
1217 {
1218         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
1219         struct drm_sched_entity *entity = p->entity;
1220         struct amdgpu_bo_list_entry *e;
1221         struct amdgpu_job *job;
1222         uint64_t seq;
1223         int r;
1224
1225         job = p->job;
1226         p->job = NULL;
1227
1228         r = drm_sched_job_init(&job->base, entity, &fpriv->vm);
1229         if (r)
1230                 goto error_unlock;
1231
1232         drm_sched_job_arm(&job->base);
1233
1234         /* No memory allocation is allowed while holding the notifier lock.
1235          * The lock is held until amdgpu_cs_submit is finished and fence is
1236          * added to BOs.
1237          */
1238         mutex_lock(&p->adev->notifier_lock);
1239
1240         /* If userptr are invalidated after amdgpu_cs_parser_bos(), return
1241          * -EAGAIN, drmIoctl in libdrm will restart the amdgpu_cs_ioctl.
1242          */
1243         amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
1244                 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
1245
1246                 r |= !amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm);
1247         }
1248         if (r) {
1249                 r = -EAGAIN;
1250                 goto error_abort;
1251         }
1252
1253         p->fence = dma_fence_get(&job->base.s_fence->finished);
1254
1255         seq = amdgpu_ctx_add_fence(p->ctx, entity, p->fence);
1256         amdgpu_cs_post_dependencies(p);
1257
1258         if ((job->preamble_status & AMDGPU_PREAMBLE_IB_PRESENT) &&
1259             !p->ctx->preamble_presented) {
1260                 job->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT_FIRST;
1261                 p->ctx->preamble_presented = true;
1262         }
1263
1264         cs->out.handle = seq;
1265         job->uf_sequence = seq;
1266
1267         amdgpu_job_free_resources(job);
1268
1269         trace_amdgpu_cs_ioctl(job);
1270         amdgpu_vm_bo_trace_cs(&fpriv->vm, &p->ticket);
1271         drm_sched_entity_push_job(&job->base);
1272
1273         amdgpu_vm_move_to_lru_tail(p->adev, &fpriv->vm);
1274
1275         /* Make sure all BOs are remembered as writers */
1276         amdgpu_bo_list_for_each_entry(e, p->bo_list)
1277                 e->tv.num_shared = 0;
1278
1279         ttm_eu_fence_buffer_objects(&p->ticket, &p->validated, p->fence);
1280         mutex_unlock(&p->adev->notifier_lock);
1281
1282         return 0;
1283
1284 error_abort:
1285         drm_sched_job_cleanup(&job->base);
1286         mutex_unlock(&p->adev->notifier_lock);
1287
1288 error_unlock:
1289         amdgpu_job_free(job);
1290         return r;
1291 }
1292
1293 static void trace_amdgpu_cs_ibs(struct amdgpu_cs_parser *parser)
1294 {
1295         int i;
1296
1297         if (!trace_amdgpu_cs_enabled())
1298                 return;
1299
1300         for (i = 0; i < parser->job->num_ibs; i++)
1301                 trace_amdgpu_cs(parser, i);
1302 }
1303
1304 int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
1305 {
1306         struct amdgpu_device *adev = drm_to_adev(dev);
1307         union drm_amdgpu_cs *cs = data;
1308         struct amdgpu_cs_parser parser = {};
1309         bool reserved_buffers = false;
1310         int r;
1311
1312         if (amdgpu_ras_intr_triggered())
1313                 return -EHWPOISON;
1314
1315         if (!adev->accel_working)
1316                 return -EBUSY;
1317
1318         parser.adev = adev;
1319         parser.filp = filp;
1320
1321         r = amdgpu_cs_parser_init(&parser, data);
1322         if (r) {
1323                 if (printk_ratelimit())
1324                         DRM_ERROR("Failed to initialize parser %d!\n", r);
1325                 goto out;
1326         }
1327
1328         r = amdgpu_cs_ib_fill(adev, &parser);
1329         if (r)
1330                 goto out;
1331
1332         r = amdgpu_cs_dependencies(adev, &parser);
1333         if (r) {
1334                 DRM_ERROR("Failed in the dependencies handling %d!\n", r);
1335                 goto out;
1336         }
1337
1338         r = amdgpu_cs_parser_bos(&parser, data);
1339         if (r) {
1340                 if (r == -ENOMEM)
1341                         DRM_ERROR("Not enough memory for command submission!\n");
1342                 else if (r != -ERESTARTSYS && r != -EAGAIN)
1343                         DRM_ERROR("Failed to process the buffer list %d!\n", r);
1344                 goto out;
1345         }
1346
1347         reserved_buffers = true;
1348
1349         trace_amdgpu_cs_ibs(&parser);
1350
1351         r = amdgpu_cs_vm_handling(&parser);
1352         if (r)
1353                 goto out;
1354
1355         r = amdgpu_cs_submit(&parser, cs);
1356
1357 out:
1358         amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
1359
1360         return r;
1361 }
1362
1363 /**
1364  * amdgpu_cs_wait_ioctl - wait for a command submission to finish
1365  *
1366  * @dev: drm device
1367  * @data: data from userspace
1368  * @filp: file private
1369  *
1370  * Wait for the command submission identified by handle to finish.
1371  */
1372 int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data,
1373                          struct drm_file *filp)
1374 {
1375         union drm_amdgpu_wait_cs *wait = data;
1376         unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
1377         struct drm_sched_entity *entity;
1378         struct amdgpu_ctx *ctx;
1379         struct dma_fence *fence;
1380         long r;
1381
1382         ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
1383         if (ctx == NULL)
1384                 return -EINVAL;
1385
1386         r = amdgpu_ctx_get_entity(ctx, wait->in.ip_type, wait->in.ip_instance,
1387                                   wait->in.ring, &entity);
1388         if (r) {
1389                 amdgpu_ctx_put(ctx);
1390                 return r;
1391         }
1392
1393         fence = amdgpu_ctx_get_fence(ctx, entity, wait->in.handle);
1394         if (IS_ERR(fence))
1395                 r = PTR_ERR(fence);
1396         else if (fence) {
1397                 r = dma_fence_wait_timeout(fence, true, timeout);
1398                 if (r > 0 && fence->error)
1399                         r = fence->error;
1400                 dma_fence_put(fence);
1401         } else
1402                 r = 1;
1403
1404         amdgpu_ctx_put(ctx);
1405         if (r < 0)
1406                 return r;
1407
1408         memset(wait, 0, sizeof(*wait));
1409         wait->out.status = (r == 0);
1410
1411         return 0;
1412 }
1413
1414 /**
1415  * amdgpu_cs_get_fence - helper to get fence from drm_amdgpu_fence
1416  *
1417  * @adev: amdgpu device
1418  * @filp: file private
1419  * @user: drm_amdgpu_fence copied from user space
1420  */
1421 static struct dma_fence *amdgpu_cs_get_fence(struct amdgpu_device *adev,
1422                                              struct drm_file *filp,
1423                                              struct drm_amdgpu_fence *user)
1424 {
1425         struct drm_sched_entity *entity;
1426         struct amdgpu_ctx *ctx;
1427         struct dma_fence *fence;
1428         int r;
1429
1430         ctx = amdgpu_ctx_get(filp->driver_priv, user->ctx_id);
1431         if (ctx == NULL)
1432                 return ERR_PTR(-EINVAL);
1433
1434         r = amdgpu_ctx_get_entity(ctx, user->ip_type, user->ip_instance,
1435                                   user->ring, &entity);
1436         if (r) {
1437                 amdgpu_ctx_put(ctx);
1438                 return ERR_PTR(r);
1439         }
1440
1441         fence = amdgpu_ctx_get_fence(ctx, entity, user->seq_no);
1442         amdgpu_ctx_put(ctx);
1443
1444         return fence;
1445 }
1446
1447 int amdgpu_cs_fence_to_handle_ioctl(struct drm_device *dev, void *data,
1448                                     struct drm_file *filp)
1449 {
1450         struct amdgpu_device *adev = drm_to_adev(dev);
1451         union drm_amdgpu_fence_to_handle *info = data;
1452         struct dma_fence *fence;
1453         struct drm_syncobj *syncobj;
1454         struct sync_file *sync_file;
1455         int fd, r;
1456
1457         fence = amdgpu_cs_get_fence(adev, filp, &info->in.fence);
1458         if (IS_ERR(fence))
1459                 return PTR_ERR(fence);
1460
1461         if (!fence)
1462                 fence = dma_fence_get_stub();
1463
1464         switch (info->in.what) {
1465         case AMDGPU_FENCE_TO_HANDLE_GET_SYNCOBJ:
1466                 r = drm_syncobj_create(&syncobj, 0, fence);
1467                 dma_fence_put(fence);
1468                 if (r)
1469                         return r;
1470                 r = drm_syncobj_get_handle(filp, syncobj, &info->out.handle);
1471                 drm_syncobj_put(syncobj);
1472                 return r;
1473
1474         case AMDGPU_FENCE_TO_HANDLE_GET_SYNCOBJ_FD:
1475                 r = drm_syncobj_create(&syncobj, 0, fence);
1476                 dma_fence_put(fence);
1477                 if (r)
1478                         return r;
1479                 r = drm_syncobj_get_fd(syncobj, (int *)&info->out.handle);
1480                 drm_syncobj_put(syncobj);
1481                 return r;
1482
1483         case AMDGPU_FENCE_TO_HANDLE_GET_SYNC_FILE_FD:
1484                 fd = get_unused_fd_flags(O_CLOEXEC);
1485                 if (fd < 0) {
1486                         dma_fence_put(fence);
1487                         return fd;
1488                 }
1489
1490                 sync_file = sync_file_create(fence);
1491                 dma_fence_put(fence);
1492                 if (!sync_file) {
1493                         put_unused_fd(fd);
1494                         return -ENOMEM;
1495                 }
1496
1497                 fd_install(fd, sync_file->file);
1498                 info->out.handle = fd;
1499                 return 0;
1500
1501         default:
1502                 dma_fence_put(fence);
1503                 return -EINVAL;
1504         }
1505 }
1506
1507 /**
1508  * amdgpu_cs_wait_all_fences - wait on all fences to signal
1509  *
1510  * @adev: amdgpu device
1511  * @filp: file private
1512  * @wait: wait parameters
1513  * @fences: array of drm_amdgpu_fence
1514  */
1515 static int amdgpu_cs_wait_all_fences(struct amdgpu_device *adev,
1516                                      struct drm_file *filp,
1517                                      union drm_amdgpu_wait_fences *wait,
1518                                      struct drm_amdgpu_fence *fences)
1519 {
1520         uint32_t fence_count = wait->in.fence_count;
1521         unsigned int i;
1522         long r = 1;
1523
1524         for (i = 0; i < fence_count; i++) {
1525                 struct dma_fence *fence;
1526                 unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1527
1528                 fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1529                 if (IS_ERR(fence))
1530                         return PTR_ERR(fence);
1531                 else if (!fence)
1532                         continue;
1533
1534                 r = dma_fence_wait_timeout(fence, true, timeout);
1535                 dma_fence_put(fence);
1536                 if (r < 0)
1537                         return r;
1538
1539                 if (r == 0)
1540                         break;
1541
1542                 if (fence->error)
1543                         return fence->error;
1544         }
1545
1546         memset(wait, 0, sizeof(*wait));
1547         wait->out.status = (r > 0);
1548
1549         return 0;
1550 }
1551
1552 /**
1553  * amdgpu_cs_wait_any_fence - wait on any fence to signal
1554  *
1555  * @adev: amdgpu device
1556  * @filp: file private
1557  * @wait: wait parameters
1558  * @fences: array of drm_amdgpu_fence
1559  */
1560 static int amdgpu_cs_wait_any_fence(struct amdgpu_device *adev,
1561                                     struct drm_file *filp,
1562                                     union drm_amdgpu_wait_fences *wait,
1563                                     struct drm_amdgpu_fence *fences)
1564 {
1565         unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1566         uint32_t fence_count = wait->in.fence_count;
1567         uint32_t first = ~0;
1568         struct dma_fence **array;
1569         unsigned int i;
1570         long r;
1571
1572         /* Prepare the fence array */
1573         array = kcalloc(fence_count, sizeof(struct dma_fence *), GFP_KERNEL);
1574
1575         if (array == NULL)
1576                 return -ENOMEM;
1577
1578         for (i = 0; i < fence_count; i++) {
1579                 struct dma_fence *fence;
1580
1581                 fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1582                 if (IS_ERR(fence)) {
1583                         r = PTR_ERR(fence);
1584                         goto err_free_fence_array;
1585                 } else if (fence) {
1586                         array[i] = fence;
1587                 } else { /* NULL, the fence has been already signaled */
1588                         r = 1;
1589                         first = i;
1590                         goto out;
1591                 }
1592         }
1593
1594         r = dma_fence_wait_any_timeout(array, fence_count, true, timeout,
1595                                        &first);
1596         if (r < 0)
1597                 goto err_free_fence_array;
1598
1599 out:
1600         memset(wait, 0, sizeof(*wait));
1601         wait->out.status = (r > 0);
1602         wait->out.first_signaled = first;
1603
1604         if (first < fence_count && array[first])
1605                 r = array[first]->error;
1606         else
1607                 r = 0;
1608
1609 err_free_fence_array:
1610         for (i = 0; i < fence_count; i++)
1611                 dma_fence_put(array[i]);
1612         kfree(array);
1613
1614         return r;
1615 }
1616
1617 /**
1618  * amdgpu_cs_wait_fences_ioctl - wait for multiple command submissions to finish
1619  *
1620  * @dev: drm device
1621  * @data: data from userspace
1622  * @filp: file private
1623  */
1624 int amdgpu_cs_wait_fences_ioctl(struct drm_device *dev, void *data,
1625                                 struct drm_file *filp)
1626 {
1627         struct amdgpu_device *adev = drm_to_adev(dev);
1628         union drm_amdgpu_wait_fences *wait = data;
1629         uint32_t fence_count = wait->in.fence_count;
1630         struct drm_amdgpu_fence *fences_user;
1631         struct drm_amdgpu_fence *fences;
1632         int r;
1633
1634         /* Get the fences from userspace */
1635         fences = kmalloc_array(fence_count, sizeof(struct drm_amdgpu_fence),
1636                         GFP_KERNEL);
1637         if (fences == NULL)
1638                 return -ENOMEM;
1639
1640         fences_user = u64_to_user_ptr(wait->in.fences);
1641         if (copy_from_user(fences, fences_user,
1642                 sizeof(struct drm_amdgpu_fence) * fence_count)) {
1643                 r = -EFAULT;
1644                 goto err_free_fences;
1645         }
1646
1647         if (wait->in.wait_all)
1648                 r = amdgpu_cs_wait_all_fences(adev, filp, wait, fences);
1649         else
1650                 r = amdgpu_cs_wait_any_fence(adev, filp, wait, fences);
1651
1652 err_free_fences:
1653         kfree(fences);
1654
1655         return r;
1656 }
1657
1658 /**
1659  * amdgpu_cs_find_mapping - find bo_va for VM address
1660  *
1661  * @parser: command submission parser context
1662  * @addr: VM address
1663  * @bo: resulting BO of the mapping found
1664  * @map: Placeholder to return found BO mapping
1665  *
1666  * Search the buffer objects in the command submission context for a certain
1667  * virtual memory address. Returns allocation structure when found, NULL
1668  * otherwise.
1669  */
1670 int amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
1671                            uint64_t addr, struct amdgpu_bo **bo,
1672                            struct amdgpu_bo_va_mapping **map)
1673 {
1674         struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
1675         struct ttm_operation_ctx ctx = { false, false };
1676         struct amdgpu_vm *vm = &fpriv->vm;
1677         struct amdgpu_bo_va_mapping *mapping;
1678         int r;
1679
1680         addr /= AMDGPU_GPU_PAGE_SIZE;
1681
1682         mapping = amdgpu_vm_bo_lookup_mapping(vm, addr);
1683         if (!mapping || !mapping->bo_va || !mapping->bo_va->base.bo)
1684                 return -EINVAL;
1685
1686         *bo = mapping->bo_va->base.bo;
1687         *map = mapping;
1688
1689         /* Double check that the BO is reserved by this CS */
1690         if (dma_resv_locking_ctx((*bo)->tbo.base.resv) != &parser->ticket)
1691                 return -EINVAL;
1692
1693         if (!((*bo)->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)) {
1694                 (*bo)->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
1695                 amdgpu_bo_placement_from_domain(*bo, (*bo)->allowed_domains);
1696                 r = ttm_bo_validate(&(*bo)->tbo, &(*bo)->placement, &ctx);
1697                 if (r)
1698                         return r;
1699         }
1700
1701         return amdgpu_ttm_alloc_gart(&(*bo)->tbo);
1702 }