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_parser_init(struct amdgpu_cs_parser *p,
  43                                  struct amdgpu_device *adev,
  44                                  struct drm_file *filp,
  45                                  union drm_amdgpu_cs *cs)
  46 {
  47         struct amdgpu_fpriv *fpriv = filp->driver_priv;
  48
  49         if (cs->in.num_chunks == 0)
  50                 return -EINVAL;
  51
  52         memset(p, 0, sizeof(*p));
  53         p->adev = adev;
  54         p->filp = filp;
  55
  56         p->ctx = amdgpu_ctx_get(fpriv, cs->in.ctx_id);
  57         if (!p->ctx)
  58                 return -EINVAL;
  59
  60         if (atomic_read(&p->ctx->guilty)) {
  61                 amdgpu_ctx_put(p->ctx);
  62                 return -ECANCELED;
  63         }
  64         return 0;
  65 }
  66
  67 static int amdgpu_cs_job_idx(struct amdgpu_cs_parser *p,
  68                              struct drm_amdgpu_cs_chunk_ib *chunk_ib)
  69 {
  70         struct drm_sched_entity *entity;
  71         unsigned int i;
  72         int r;
  73
  74         r = amdgpu_ctx_get_entity(p->ctx, chunk_ib->ip_type,
  75                                   chunk_ib->ip_instance,
  76                                   chunk_ib->ring, &entity);
  77         if (r)
  78                 return r;
  79
  80         /*
  81          * Abort if there is no run queue associated with this entity.
  82          * Possibly because of disabled HW IP.
  83          */
  84         if (entity->rq == NULL)
  85                 return -EINVAL;
  86
  87         /* Check if we can add this IB to some existing job */
  88         for (i = 0; i < p->gang_size; ++i)
  89                 if (p->entities[i] == entity)
  90                         return i;
  91
  92         /* If not increase the gang size if possible */
  93         if (i == AMDGPU_CS_GANG_SIZE)
  94                 return -EINVAL;
  95
  96         p->entities[i] = entity;
  97         p->gang_size = i + 1;
  98         return i;
  99 }
 100
 101 static int amdgpu_cs_p1_ib(struct amdgpu_cs_parser *p,
 102                            struct drm_amdgpu_cs_chunk_ib *chunk_ib,
 103                            unsigned int *num_ibs)
 104 {
 105         int r;
 106
 107         r = amdgpu_cs_job_idx(p, chunk_ib);
 108         if (r < 0)
 109                 return r;
 110
 111         ++(num_ibs[r]);
 112         return 0;
 113 }
 114
 115 static int amdgpu_cs_p1_user_fence(struct amdgpu_cs_parser *p,
 116                                    struct drm_amdgpu_cs_chunk_fence *data,
 117                                    uint32_t *offset)
 118 {
 119         struct drm_gem_object *gobj;
 120         struct amdgpu_bo *bo;
 121         unsigned long size;
 122         int r;
 123
 124         gobj = drm_gem_object_lookup(p->filp, data->handle);
 125         if (gobj == NULL)
 126                 return -EINVAL;
 127
 128         bo = amdgpu_bo_ref(gem_to_amdgpu_bo(gobj));
 129         p->uf_entry.priority = 0;
 130         p->uf_entry.tv.bo = &bo->tbo;
 131         /* One for TTM and two for the CS job */
 132         p->uf_entry.tv.num_shared = 3;
 133
 134         drm_gem_object_put(gobj);
 135
 136         size = amdgpu_bo_size(bo);
 137         if (size != PAGE_SIZE || (data->offset + 8) > size) {
 138                 r = -EINVAL;
 139                 goto error_unref;
 140         }
 141
 142         if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
 143                 r = -EINVAL;
 144                 goto error_unref;
 145         }
 146
 147         *offset = data->offset;
 148
 149         return 0;
 150
 151 error_unref:
 152         amdgpu_bo_unref(&bo);
 153         return r;
 154 }
 155
 156 static int amdgpu_cs_p1_bo_handles(struct amdgpu_cs_parser *p,
 157                                    struct drm_amdgpu_bo_list_in *data)
 158 {
 159         struct drm_amdgpu_bo_list_entry *info;
 160         int r;
 161
 162         r = amdgpu_bo_create_list_entry_array(data, &info);
 163         if (r)
 164                 return r;
 165
 166         r = amdgpu_bo_list_create(p->adev, p->filp, info, data->bo_number,
 167                                   &p->bo_list);
 168         if (r)
 169                 goto error_free;
 170
 171         kvfree(info);
 172         return 0;
 173
 174 error_free:
 175         kvfree(info);
 176
 177         return r;
 178 }
 179
 180 /* Copy the data from userspace and go over it the first time */
 181 static int amdgpu_cs_pass1(struct amdgpu_cs_parser *p,
 182                            union drm_amdgpu_cs *cs)
 183 {
 184         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 185         unsigned int num_ibs[AMDGPU_CS_GANG_SIZE] = { };
 186         struct amdgpu_vm *vm = &fpriv->vm;
 187         uint64_t *chunk_array_user;
 188         uint64_t *chunk_array;
 189         uint32_t uf_offset = 0;
 190         unsigned int size;
 191         int ret;
 192         int i;
 193
 194         chunk_array = kvmalloc_array(cs->in.num_chunks, sizeof(uint64_t),
 195                                      GFP_KERNEL);
 196         if (!chunk_array)
 197                 return -ENOMEM;
 198
 199         /* get chunks */
 200         chunk_array_user = u64_to_user_ptr(cs->in.chunks);
 201         if (copy_from_user(chunk_array, chunk_array_user,
 202                            sizeof(uint64_t)*cs->in.num_chunks)) {
 203                 ret = -EFAULT;
 204                 goto free_chunk;
 205         }
 206
 207         p->nchunks = cs->in.num_chunks;
 208         p->chunks = kvmalloc_array(p->nchunks, sizeof(struct amdgpu_cs_chunk),
 209                             GFP_KERNEL);
 210         if (!p->chunks) {
 211                 ret = -ENOMEM;
 212                 goto free_chunk;
 213         }
 214
 215         for (i = 0; i < p->nchunks; i++) {
 216                 struct drm_amdgpu_cs_chunk __user **chunk_ptr = NULL;
 217                 struct drm_amdgpu_cs_chunk user_chunk;
 218                 uint32_t __user *cdata;
 219
 220                 chunk_ptr = u64_to_user_ptr(chunk_array[i]);
 221                 if (copy_from_user(&user_chunk, chunk_ptr,
 222                                        sizeof(struct drm_amdgpu_cs_chunk))) {
 223                         ret = -EFAULT;
 224                         i--;
 225                         goto free_partial_kdata;
 226                 }
 227                 p->chunks[i].chunk_id = user_chunk.chunk_id;
 228                 p->chunks[i].length_dw = user_chunk.length_dw;
 229
 230                 size = p->chunks[i].length_dw;
 231                 cdata = u64_to_user_ptr(user_chunk.chunk_data);
 232
 233                 p->chunks[i].kdata = kvmalloc_array(size, sizeof(uint32_t),
 234                                                     GFP_KERNEL);
 235                 if (p->chunks[i].kdata == NULL) {
 236                         ret = -ENOMEM;
 237                         i--;
 238                         goto free_partial_kdata;
 239                 }
 240                 size *= sizeof(uint32_t);
 241                 if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
 242                         ret = -EFAULT;
 243                         goto free_partial_kdata;
 244                 }
 245
 246                 /* Assume the worst on the following checks */
 247                 ret = -EINVAL;
 248                 switch (p->chunks[i].chunk_id) {
 249                 case AMDGPU_CHUNK_ID_IB:
 250                         if (size < sizeof(struct drm_amdgpu_cs_chunk_ib))
 251                                 goto free_partial_kdata;
 252
 253                         ret = amdgpu_cs_p1_ib(p, p->chunks[i].kdata, num_ibs);
 254                         if (ret)
 255                                 goto free_partial_kdata;
 256                         break;
 257
 258                 case AMDGPU_CHUNK_ID_FENCE:
 259                         if (size < sizeof(struct drm_amdgpu_cs_chunk_fence))
 260                                 goto free_partial_kdata;
 261
 262                         ret = amdgpu_cs_p1_user_fence(p, p->chunks[i].kdata,
 263                                                       &uf_offset);
 264                         if (ret)
 265                                 goto free_partial_kdata;
 266                         break;
 267
 268                 case AMDGPU_CHUNK_ID_BO_HANDLES:
 269                         if (size < sizeof(struct drm_amdgpu_bo_list_in))
 270                                 goto free_partial_kdata;
 271
 272                         ret = amdgpu_cs_p1_bo_handles(p, p->chunks[i].kdata);
 273                         if (ret)
 274                                 goto free_partial_kdata;
 275                         break;
 276
 277                 case AMDGPU_CHUNK_ID_DEPENDENCIES:
 278                 case AMDGPU_CHUNK_ID_SYNCOBJ_IN:
 279                 case AMDGPU_CHUNK_ID_SYNCOBJ_OUT:
 280                 case AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES:
 281                 case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_WAIT:
 282                 case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_SIGNAL:
 283                         break;
 284
 285                 default:
 286                         goto free_partial_kdata;
 287                 }
 288         }
 289
 290         if (!p->gang_size)
 291                 return -EINVAL;
 292
 293         for (i = 0; i < p->gang_size; ++i) {
 294                 ret = amdgpu_job_alloc(p->adev, num_ibs[i], &p->jobs[i], vm);
 295                 if (ret)
 296                         goto free_all_kdata;
 297
 298                 ret = drm_sched_job_init(&p->jobs[i]->base, p->entities[i],
 299                                          &fpriv->vm);
 300                 if (ret)
 301                         goto free_all_kdata;
 302         }
 303         p->gang_leader = p->jobs[p->gang_size - 1];
 304
 305         if (p->ctx->vram_lost_counter != p->gang_leader->vram_lost_counter) {
 306                 ret = -ECANCELED;
 307                 goto free_all_kdata;
 308         }
 309
 310         if (p->uf_entry.tv.bo)
 311                 p->gang_leader->uf_addr = uf_offset;
 312         kvfree(chunk_array);
 313
 314         /* Use this opportunity to fill in task info for the vm */
 315         amdgpu_vm_set_task_info(vm);
 316
 317         return 0;
 318
 319 free_all_kdata:
 320         i = p->nchunks - 1;
 321 free_partial_kdata:
 322         for (; i >= 0; i--)
 323                 kvfree(p->chunks[i].kdata);
 324         kvfree(p->chunks);
 325         p->chunks = NULL;
 326         p->nchunks = 0;
 327 free_chunk:
 328         kvfree(chunk_array);
 329
 330         return ret;
 331 }
 332
 333 static int amdgpu_cs_p2_ib(struct amdgpu_cs_parser *p,
 334                            struct amdgpu_cs_chunk *chunk,
 335                            unsigned int *ce_preempt,
 336                            unsigned int *de_preempt)
 337 {
 338         struct drm_amdgpu_cs_chunk_ib *chunk_ib = chunk->kdata;
 339         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 340         struct amdgpu_vm *vm = &fpriv->vm;
 341         struct amdgpu_ring *ring;
 342         struct amdgpu_job *job;
 343         struct amdgpu_ib *ib;
 344         int r;
 345
 346         r = amdgpu_cs_job_idx(p, chunk_ib);
 347         if (r < 0)
 348                 return r;
 349
 350         job = p->jobs[r];
 351         ring = amdgpu_job_ring(job);
 352         ib = &job->ibs[job->num_ibs++];
 353
 354         /* MM engine doesn't support user fences */
 355         if (p->uf_entry.tv.bo && ring->funcs->no_user_fence)
 356                 return -EINVAL;
 357
 358         if (chunk_ib->ip_type == AMDGPU_HW_IP_GFX &&
 359             chunk_ib->flags & AMDGPU_IB_FLAG_PREEMPT) {
 360                 if (chunk_ib->flags & AMDGPU_IB_FLAG_CE)
 361                         (*ce_preempt)++;
 362                 else
 363                         (*de_preempt)++;
 364
 365                 /* Each GFX command submit allows only 1 IB max
 366                  * preemptible for CE & DE */
 367                 if (*ce_preempt > 1 || *de_preempt > 1)
 368                         return -EINVAL;
 369         }
 370
 371         if (chunk_ib->flags & AMDGPU_IB_FLAG_PREAMBLE)
 372                 job->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT;
 373
 374         r =  amdgpu_ib_get(p->adev, vm, ring->funcs->parse_cs ?
 375                            chunk_ib->ib_bytes : 0,
 376                            AMDGPU_IB_POOL_DELAYED, ib);
 377         if (r) {
 378                 DRM_ERROR("Failed to get ib !\n");
 379                 return r;
 380         }
 381
 382         ib->gpu_addr = chunk_ib->va_start;
 383         ib->length_dw = chunk_ib->ib_bytes / 4;
 384         ib->flags = chunk_ib->flags;
 385         return 0;
 386 }
 387
 388 static int amdgpu_cs_p2_dependencies(struct amdgpu_cs_parser *p,
 389                                      struct amdgpu_cs_chunk *chunk)
 390 {
 391         struct drm_amdgpu_cs_chunk_dep *deps = chunk->kdata;
 392         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 393         unsigned num_deps;
 394         int i, r;
 395
 396         num_deps = chunk->length_dw * 4 /
 397                 sizeof(struct drm_amdgpu_cs_chunk_dep);
 398
 399         for (i = 0; i < num_deps; ++i) {
 400                 struct amdgpu_ctx *ctx;
 401                 struct drm_sched_entity *entity;
 402                 struct dma_fence *fence;
 403
 404                 ctx = amdgpu_ctx_get(fpriv, deps[i].ctx_id);
 405                 if (ctx == NULL)
 406                         return -EINVAL;
 407
 408                 r = amdgpu_ctx_get_entity(ctx, deps[i].ip_type,
 409                                           deps[i].ip_instance,
 410                                           deps[i].ring, &entity);
 411                 if (r) {
 412                         amdgpu_ctx_put(ctx);
 413                         return r;
 414                 }
 415
 416                 fence = amdgpu_ctx_get_fence(ctx, entity, deps[i].handle);
 417                 amdgpu_ctx_put(ctx);
 418
 419                 if (IS_ERR(fence))
 420                         return PTR_ERR(fence);
 421                 else if (!fence)
 422                         continue;
 423
 424                 if (chunk->chunk_id == AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES) {
 425                         struct drm_sched_fence *s_fence;
 426                         struct dma_fence *old = fence;
 427
 428                         s_fence = to_drm_sched_fence(fence);
 429                         fence = dma_fence_get(&s_fence->scheduled);
 430                         dma_fence_put(old);
 431                 }
 432
 433                 r = amdgpu_sync_fence(&p->gang_leader->sync, fence);
 434                 dma_fence_put(fence);
 435                 if (r)
 436                         return r;
 437         }
 438         return 0;
 439 }
 440
 441 static int amdgpu_syncobj_lookup_and_add(struct amdgpu_cs_parser *p,
 442                                          uint32_t handle, u64 point,
 443                                          u64 flags)
 444 {
 445         struct dma_fence *fence;
 446         int r;
 447
 448         r = drm_syncobj_find_fence(p->filp, handle, point, flags, &fence);
 449         if (r) {
 450                 DRM_ERROR("syncobj %u failed to find fence @ %llu (%d)!\n",
 451                           handle, point, r);
 452                 return r;
 453         }
 454
 455         r = amdgpu_sync_fence(&p->gang_leader->sync, fence);
 456         dma_fence_put(fence);
 457
 458         return r;
 459 }
 460
 461 static int amdgpu_cs_p2_syncobj_in(struct amdgpu_cs_parser *p,
 462                                    struct amdgpu_cs_chunk *chunk)
 463 {
 464         struct drm_amdgpu_cs_chunk_sem *deps = chunk->kdata;
 465         unsigned num_deps;
 466         int i, r;
 467
 468         num_deps = chunk->length_dw * 4 /
 469                 sizeof(struct drm_amdgpu_cs_chunk_sem);
 470         for (i = 0; i < num_deps; ++i) {
 471                 r = amdgpu_syncobj_lookup_and_add(p, deps[i].handle, 0, 0);
 472                 if (r)
 473                         return r;
 474         }
 475
 476         return 0;
 477 }
 478
 479 static int amdgpu_cs_p2_syncobj_timeline_wait(struct amdgpu_cs_parser *p,
 480                                               struct amdgpu_cs_chunk *chunk)
 481 {
 482         struct drm_amdgpu_cs_chunk_syncobj *syncobj_deps = chunk->kdata;
 483         unsigned num_deps;
 484         int i, r;
 485
 486         num_deps = chunk->length_dw * 4 /
 487                 sizeof(struct drm_amdgpu_cs_chunk_syncobj);
 488         for (i = 0; i < num_deps; ++i) {
 489                 r = amdgpu_syncobj_lookup_and_add(p, syncobj_deps[i].handle,
 490                                                   syncobj_deps[i].point,
 491                                                   syncobj_deps[i].flags);
 492                 if (r)
 493                         return r;
 494         }
 495
 496         return 0;
 497 }
 498
 499 static int amdgpu_cs_p2_syncobj_out(struct amdgpu_cs_parser *p,
 500                                     struct amdgpu_cs_chunk *chunk)
 501 {
 502         struct drm_amdgpu_cs_chunk_sem *deps = chunk->kdata;
 503         unsigned num_deps;
 504         int i;
 505
 506         num_deps = chunk->length_dw * 4 /
 507                 sizeof(struct drm_amdgpu_cs_chunk_sem);
 508
 509         if (p->post_deps)
 510                 return -EINVAL;
 511
 512         p->post_deps = kmalloc_array(num_deps, sizeof(*p->post_deps),
 513                                      GFP_KERNEL);
 514         p->num_post_deps = 0;
 515
 516         if (!p->post_deps)
 517                 return -ENOMEM;
 518
 519
 520         for (i = 0; i < num_deps; ++i) {
 521                 p->post_deps[i].syncobj =
 522                         drm_syncobj_find(p->filp, deps[i].handle);
 523                 if (!p->post_deps[i].syncobj)
 524                         return -EINVAL;
 525                 p->post_deps[i].chain = NULL;
 526                 p->post_deps[i].point = 0;
 527                 p->num_post_deps++;
 528         }
 529
 530         return 0;
 531 }
 532
 533 static int amdgpu_cs_p2_syncobj_timeline_signal(struct amdgpu_cs_parser *p,
 534                                                 struct amdgpu_cs_chunk *chunk)
 535 {
 536         struct drm_amdgpu_cs_chunk_syncobj *syncobj_deps = chunk->kdata;
 537         unsigned num_deps;
 538         int i;
 539
 540         num_deps = chunk->length_dw * 4 /
 541                 sizeof(struct drm_amdgpu_cs_chunk_syncobj);
 542
 543         if (p->post_deps)
 544                 return -EINVAL;
 545
 546         p->post_deps = kmalloc_array(num_deps, sizeof(*p->post_deps),
 547                                      GFP_KERNEL);
 548         p->num_post_deps = 0;
 549
 550         if (!p->post_deps)
 551                 return -ENOMEM;
 552
 553         for (i = 0; i < num_deps; ++i) {
 554                 struct amdgpu_cs_post_dep *dep = &p->post_deps[i];
 555
 556                 dep->chain = NULL;
 557                 if (syncobj_deps[i].point) {
 558                         dep->chain = dma_fence_chain_alloc();
 559                         if (!dep->chain)
 560                                 return -ENOMEM;
 561                 }
 562
 563                 dep->syncobj = drm_syncobj_find(p->filp,
 564                                                 syncobj_deps[i].handle);
 565                 if (!dep->syncobj) {
 566                         dma_fence_chain_free(dep->chain);
 567                         return -EINVAL;
 568                 }
 569                 dep->point = syncobj_deps[i].point;
 570                 p->num_post_deps++;
 571         }
 572
 573         return 0;
 574 }
 575
 576 static int amdgpu_cs_pass2(struct amdgpu_cs_parser *p)
 577 {
 578         unsigned int ce_preempt = 0, de_preempt = 0;
 579         int i, r;
 580
 581         for (i = 0; i < p->nchunks; ++i) {
 582                 struct amdgpu_cs_chunk *chunk;
 583
 584                 chunk = &p->chunks[i];
 585
 586                 switch (chunk->chunk_id) {
 587                 case AMDGPU_CHUNK_ID_IB:
 588                         r = amdgpu_cs_p2_ib(p, chunk, &ce_preempt, &de_preempt);
 589                         if (r)
 590                                 return r;
 591                         break;
 592                 case AMDGPU_CHUNK_ID_DEPENDENCIES:
 593                 case AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES:
 594                         r = amdgpu_cs_p2_dependencies(p, chunk);
 595                         if (r)
 596                                 return r;
 597                         break;
 598                 case AMDGPU_CHUNK_ID_SYNCOBJ_IN:
 599                         r = amdgpu_cs_p2_syncobj_in(p, chunk);
 600                         if (r)
 601                                 return r;
 602                         break;
 603                 case AMDGPU_CHUNK_ID_SYNCOBJ_OUT:
 604                         r = amdgpu_cs_p2_syncobj_out(p, chunk);
 605                         if (r)
 606                                 return r;
 607                         break;
 608                 case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_WAIT:
 609                         r = amdgpu_cs_p2_syncobj_timeline_wait(p, chunk);
 610                         if (r)
 611                                 return r;
 612                         break;
 613                 case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_SIGNAL:
 614                         r = amdgpu_cs_p2_syncobj_timeline_signal(p, chunk);
 615                         if (r)
 616                                 return r;
 617                         break;
 618                 }
 619         }
 620
 621         return 0;
 622 }
 623
 624 /* Convert microseconds to bytes. */
 625 static u64 us_to_bytes(struct amdgpu_device *adev, s64 us)
 626 {
 627         if (us <= 0 || !adev->mm_stats.log2_max_MBps)
 628                 return 0;
 629
 630         /* Since accum_us is incremented by a million per second, just
 631          * multiply it by the number of MB/s to get the number of bytes.
 632          */
 633         return us << adev->mm_stats.log2_max_MBps;
 634 }
 635
 636 static s64 bytes_to_us(struct amdgpu_device *adev, u64 bytes)
 637 {
 638         if (!adev->mm_stats.log2_max_MBps)
 639                 return 0;
 640
 641         return bytes >> adev->mm_stats.log2_max_MBps;
 642 }
 643
 644 /* Returns how many bytes TTM can move right now. If no bytes can be moved,
 645  * it returns 0. If it returns non-zero, it's OK to move at least one buffer,
 646  * which means it can go over the threshold once. If that happens, the driver
 647  * will be in debt and no other buffer migrations can be done until that debt
 648  * is repaid.
 649  *
 650  * This approach allows moving a buffer of any size (it's important to allow
 651  * that).
 652  *
 653  * The currency is simply time in microseconds and it increases as the clock
 654  * ticks. The accumulated microseconds (us) are converted to bytes and
 655  * returned.
 656  */
 657 static void amdgpu_cs_get_threshold_for_moves(struct amdgpu_device *adev,
 658                                               u64 *max_bytes,
 659                                               u64 *max_vis_bytes)
 660 {
 661         s64 time_us, increment_us;
 662         u64 free_vram, total_vram, used_vram;
 663         /* Allow a maximum of 200 accumulated ms. This is basically per-IB
 664          * throttling.
 665          *
 666          * It means that in order to get full max MBps, at least 5 IBs per
 667          * second must be submitted and not more than 200ms apart from each
 668          * other.
 669          */
 670         const s64 us_upper_bound = 200000;
 671
 672         if (!adev->mm_stats.log2_max_MBps) {
 673                 *max_bytes = 0;
 674                 *max_vis_bytes = 0;
 675                 return;
 676         }
 677
 678         total_vram = adev->gmc.real_vram_size - atomic64_read(&adev->vram_pin_size);
 679         used_vram = ttm_resource_manager_usage(&adev->mman.vram_mgr.manager);
 680         free_vram = used_vram >= total_vram ? 0 : total_vram - used_vram;
 681
 682         spin_lock(&adev->mm_stats.lock);
 683
 684         /* Increase the amount of accumulated us. */
 685         time_us = ktime_to_us(ktime_get());
 686         increment_us = time_us - adev->mm_stats.last_update_us;
 687         adev->mm_stats.last_update_us = time_us;
 688         adev->mm_stats.accum_us = min(adev->mm_stats.accum_us + increment_us,
 689                                       us_upper_bound);
 690
 691         /* This prevents the short period of low performance when the VRAM
 692          * usage is low and the driver is in debt or doesn't have enough
 693          * accumulated us to fill VRAM quickly.
 694          *
 695          * The situation can occur in these cases:
 696          * - a lot of VRAM is freed by userspace
 697          * - the presence of a big buffer causes a lot of evictions
 698          *   (solution: split buffers into smaller ones)
 699          *
 700          * If 128 MB or 1/8th of VRAM is free, start filling it now by setting
 701          * accum_us to a positive number.
 702          */
 703         if (free_vram >= 128 * 1024 * 1024 || free_vram >= total_vram / 8) {
 704                 s64 min_us;
 705
 706                 /* Be more aggressive on dGPUs. Try to fill a portion of free
 707                  * VRAM now.
 708                  */
 709                 if (!(adev->flags & AMD_IS_APU))
 710                         min_us = bytes_to_us(adev, free_vram / 4);
 711                 else
 712                         min_us = 0; /* Reset accum_us on APUs. */
 713
 714                 adev->mm_stats.accum_us = max(min_us, adev->mm_stats.accum_us);
 715         }
 716
 717         /* This is set to 0 if the driver is in debt to disallow (optional)
 718          * buffer moves.
 719          */
 720         *max_bytes = us_to_bytes(adev, adev->mm_stats.accum_us);
 721
 722         /* Do the same for visible VRAM if half of it is free */
 723         if (!amdgpu_gmc_vram_full_visible(&adev->gmc)) {
 724                 u64 total_vis_vram = adev->gmc.visible_vram_size;
 725                 u64 used_vis_vram =
 726                   amdgpu_vram_mgr_vis_usage(&adev->mman.vram_mgr);
 727
 728                 if (used_vis_vram < total_vis_vram) {
 729                         u64 free_vis_vram = total_vis_vram - used_vis_vram;
 730                         adev->mm_stats.accum_us_vis = min(adev->mm_stats.accum_us_vis +
 731                                                           increment_us, us_upper_bound);
 732
 733                         if (free_vis_vram >= total_vis_vram / 2)
 734                                 adev->mm_stats.accum_us_vis =
 735                                         max(bytes_to_us(adev, free_vis_vram / 2),
 736                                             adev->mm_stats.accum_us_vis);
 737                 }
 738
 739                 *max_vis_bytes = us_to_bytes(adev, adev->mm_stats.accum_us_vis);
 740         } else {
 741                 *max_vis_bytes = 0;
 742         }
 743
 744         spin_unlock(&adev->mm_stats.lock);
 745 }
 746
 747 /* Report how many bytes have really been moved for the last command
 748  * submission. This can result in a debt that can stop buffer migrations
 749  * temporarily.
 750  */
 751 void amdgpu_cs_report_moved_bytes(struct amdgpu_device *adev, u64 num_bytes,
 752                                   u64 num_vis_bytes)
 753 {
 754         spin_lock(&adev->mm_stats.lock);
 755         adev->mm_stats.accum_us -= bytes_to_us(adev, num_bytes);
 756         adev->mm_stats.accum_us_vis -= bytes_to_us(adev, num_vis_bytes);
 757         spin_unlock(&adev->mm_stats.lock);
 758 }
 759
 760 static int amdgpu_cs_bo_validate(void *param, struct amdgpu_bo *bo)
 761 {
 762         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
 763         struct amdgpu_cs_parser *p = param;
 764         struct ttm_operation_ctx ctx = {
 765                 .interruptible = true,
 766                 .no_wait_gpu = false,
 767                 .resv = bo->tbo.base.resv
 768         };
 769         uint32_t domain;
 770         int r;
 771
 772         if (bo->tbo.pin_count)
 773                 return 0;
 774
 775         /* Don't move this buffer if we have depleted our allowance
 776          * to move it. Don't move anything if the threshold is zero.
 777          */
 778         if (p->bytes_moved < p->bytes_moved_threshold &&
 779             (!bo->tbo.base.dma_buf ||
 780             list_empty(&bo->tbo.base.dma_buf->attachments))) {
 781                 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
 782                     (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)) {
 783                         /* And don't move a CPU_ACCESS_REQUIRED BO to limited
 784                          * visible VRAM if we've depleted our allowance to do
 785                          * that.
 786                          */
 787                         if (p->bytes_moved_vis < p->bytes_moved_vis_threshold)
 788                                 domain = bo->preferred_domains;
 789                         else
 790                                 domain = bo->allowed_domains;
 791                 } else {
 792                         domain = bo->preferred_domains;
 793                 }
 794         } else {
 795                 domain = bo->allowed_domains;
 796         }
 797
 798 retry:
 799         amdgpu_bo_placement_from_domain(bo, domain);
 800         r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 801
 802         p->bytes_moved += ctx.bytes_moved;
 803         if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
 804             amdgpu_bo_in_cpu_visible_vram(bo))
 805                 p->bytes_moved_vis += ctx.bytes_moved;
 806
 807         if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
 808                 domain = bo->allowed_domains;
 809                 goto retry;
 810         }
 811
 812         return r;
 813 }
 814
 815 static int amdgpu_cs_list_validate(struct amdgpu_cs_parser *p,
 816                             struct list_head *validated)
 817 {
 818         struct ttm_operation_ctx ctx = { true, false };
 819         struct amdgpu_bo_list_entry *lobj;
 820         int r;
 821
 822         list_for_each_entry(lobj, validated, tv.head) {
 823                 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(lobj->tv.bo);
 824                 struct mm_struct *usermm;
 825
 826                 usermm = amdgpu_ttm_tt_get_usermm(bo->tbo.ttm);
 827                 if (usermm && usermm != current->mm)
 828                         return -EPERM;
 829
 830                 if (amdgpu_ttm_tt_is_userptr(bo->tbo.ttm) &&
 831                     lobj->user_invalidated && lobj->user_pages) {
 832                         amdgpu_bo_placement_from_domain(bo,
 833                                                         AMDGPU_GEM_DOMAIN_CPU);
 834                         r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 835                         if (r)
 836                                 return r;
 837
 838                         amdgpu_ttm_tt_set_user_pages(bo->tbo.ttm,
 839                                                      lobj->user_pages);
 840                 }
 841
 842                 r = amdgpu_cs_bo_validate(p, bo);
 843                 if (r)
 844                         return r;
 845
 846                 kvfree(lobj->user_pages);
 847                 lobj->user_pages = NULL;
 848         }
 849         return 0;
 850 }
 851
 852 static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
 853                                 union drm_amdgpu_cs *cs)
 854 {
 855         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 856         struct amdgpu_vm *vm = &fpriv->vm;
 857         struct amdgpu_bo_list_entry *e;
 858         struct list_head duplicates;
 859         unsigned int i;
 860         int r;
 861
 862         INIT_LIST_HEAD(&p->validated);
 863
 864         /* p->bo_list could already be assigned if AMDGPU_CHUNK_ID_BO_HANDLES is present */
 865         if (cs->in.bo_list_handle) {
 866                 if (p->bo_list)
 867                         return -EINVAL;
 868
 869                 r = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle,
 870                                        &p->bo_list);
 871                 if (r)
 872                         return r;
 873         } else if (!p->bo_list) {
 874                 /* Create a empty bo_list when no handle is provided */
 875                 r = amdgpu_bo_list_create(p->adev, p->filp, NULL, 0,
 876                                           &p->bo_list);
 877                 if (r)
 878                         return r;
 879         }
 880
 881         mutex_lock(&p->bo_list->bo_list_mutex);
 882
 883         /* One for TTM and one for the CS job */
 884         amdgpu_bo_list_for_each_entry(e, p->bo_list)
 885                 e->tv.num_shared = 2;
 886
 887         amdgpu_bo_list_get_list(p->bo_list, &p->validated);
 888
 889         INIT_LIST_HEAD(&duplicates);
 890         amdgpu_vm_get_pd_bo(&fpriv->vm, &p->validated, &p->vm_pd);
 891
 892         if (p->uf_entry.tv.bo && !ttm_to_amdgpu_bo(p->uf_entry.tv.bo)->parent)
 893                 list_add(&p->uf_entry.tv.head, &p->validated);
 894
 895         /* Get userptr backing pages. If pages are updated after registered
 896          * in amdgpu_gem_userptr_ioctl(), amdgpu_cs_list_validate() will do
 897          * amdgpu_ttm_backend_bind() to flush and invalidate new pages
 898          */
 899         amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
 900                 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
 901                 bool userpage_invalidated = false;
 902                 int i;
 903
 904                 e->user_pages = kvmalloc_array(bo->tbo.ttm->num_pages,
 905                                         sizeof(struct page *),
 906                                         GFP_KERNEL | __GFP_ZERO);
 907                 if (!e->user_pages) {
 908                         DRM_ERROR("kvmalloc_array failure\n");
 909                         r = -ENOMEM;
 910                         goto out_free_user_pages;
 911                 }
 912
 913                 r = amdgpu_ttm_tt_get_user_pages(bo, e->user_pages);
 914                 if (r) {
 915                         kvfree(e->user_pages);
 916                         e->user_pages = NULL;
 917                         goto out_free_user_pages;
 918                 }
 919
 920                 for (i = 0; i < bo->tbo.ttm->num_pages; i++) {
 921                         if (bo->tbo.ttm->pages[i] != e->user_pages[i]) {
 922                                 userpage_invalidated = true;
 923                                 break;
 924                         }
 925                 }
 926                 e->user_invalidated = userpage_invalidated;
 927         }
 928
 929         r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true,
 930                                    &duplicates);
 931         if (unlikely(r != 0)) {
 932                 if (r != -ERESTARTSYS)
 933                         DRM_ERROR("ttm_eu_reserve_buffers failed.\n");
 934                 goto out_free_user_pages;
 935         }
 936
 937         amdgpu_bo_list_for_each_entry(e, p->bo_list) {
 938                 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
 939
 940                 e->bo_va = amdgpu_vm_bo_find(vm, bo);
 941         }
 942
 943         amdgpu_cs_get_threshold_for_moves(p->adev, &p->bytes_moved_threshold,
 944                                           &p->bytes_moved_vis_threshold);
 945         p->bytes_moved = 0;
 946         p->bytes_moved_vis = 0;
 947
 948         r = amdgpu_vm_validate_pt_bos(p->adev, &fpriv->vm,
 949                                       amdgpu_cs_bo_validate, p);
 950         if (r) {
 951                 DRM_ERROR("amdgpu_vm_validate_pt_bos() failed.\n");
 952                 goto error_validate;
 953         }
 954
 955         r = amdgpu_cs_list_validate(p, &duplicates);
 956         if (r)
 957                 goto error_validate;
 958
 959         r = amdgpu_cs_list_validate(p, &p->validated);
 960         if (r)
 961                 goto error_validate;
 962
 963         if (p->uf_entry.tv.bo) {
 964                 struct amdgpu_bo *uf = ttm_to_amdgpu_bo(p->uf_entry.tv.bo);
 965
 966                 r = amdgpu_ttm_alloc_gart(&uf->tbo);
 967                 if (r)
 968                         goto error_validate;
 969
 970                 p->gang_leader->uf_addr += amdgpu_bo_gpu_offset(uf);
 971         }
 972
 973         amdgpu_cs_report_moved_bytes(p->adev, p->bytes_moved,
 974                                      p->bytes_moved_vis);
 975
 976         for (i = 0; i < p->gang_size; ++i)
 977                 amdgpu_job_set_resources(p->jobs[i], p->bo_list->gds_obj,
 978                                          p->bo_list->gws_obj,
 979                                          p->bo_list->oa_obj);
 980         return 0;
 981
 982 error_validate:
 983         ttm_eu_backoff_reservation(&p->ticket, &p->validated);
 984
 985 out_free_user_pages:
 986         amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
 987                 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
 988
 989                 if (!e->user_pages)
 990                         continue;
 991                 amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm);
 992                 kvfree(e->user_pages);
 993                 e->user_pages = NULL;
 994         }
 995         return r;
 996 }
 997
 998 static void trace_amdgpu_cs_ibs(struct amdgpu_cs_parser *p)
 999 {
1000         int i, j;
1001
1002         if (!trace_amdgpu_cs_enabled())
1003                 return;
1004
1005         for (i = 0; i < p->gang_size; ++i) {
1006                 struct amdgpu_job *job = p->jobs[i];
1007
1008                 for (j = 0; j < job->num_ibs; ++j)
1009                         trace_amdgpu_cs(p, job, &job->ibs[j]);
1010         }
1011 }
1012
1013 static int amdgpu_cs_patch_ibs(struct amdgpu_cs_parser *p,
1014                                struct amdgpu_job *job)
1015 {
1016         struct amdgpu_ring *ring = amdgpu_job_ring(job);
1017         unsigned int i;
1018         int r;
1019
1020         /* Only for UVD/VCE VM emulation */
1021         if (!ring->funcs->parse_cs && !ring->funcs->patch_cs_in_place)
1022                 return 0;
1023
1024         for (i = 0; i < job->num_ibs; ++i) {
1025                 struct amdgpu_ib *ib = &job->ibs[i];
1026                 struct amdgpu_bo_va_mapping *m;
1027                 struct amdgpu_bo *aobj;
1028                 uint64_t va_start;
1029                 uint8_t *kptr;
1030
1031                 va_start = ib->gpu_addr & AMDGPU_GMC_HOLE_MASK;
1032                 r = amdgpu_cs_find_mapping(p, va_start, &aobj, &m);
1033                 if (r) {
1034                         DRM_ERROR("IB va_start is invalid\n");
1035                         return r;
1036                 }
1037
1038                 if ((va_start + ib->length_dw * 4) >
1039                     (m->last + 1) * AMDGPU_GPU_PAGE_SIZE) {
1040                         DRM_ERROR("IB va_start+ib_bytes is invalid\n");
1041                         return -EINVAL;
1042                 }
1043
1044                 /* the IB should be reserved at this point */
1045                 r = amdgpu_bo_kmap(aobj, (void **)&kptr);
1046                 if (r) {
1047                         return r;
1048                 }
1049
1050                 kptr += va_start - (m->start * AMDGPU_GPU_PAGE_SIZE);
1051
1052                 if (ring->funcs->parse_cs) {
1053                         memcpy(ib->ptr, kptr, ib->length_dw * 4);
1054                         amdgpu_bo_kunmap(aobj);
1055
1056                         r = amdgpu_ring_parse_cs(ring, p, job, ib);
1057                         if (r)
1058                                 return r;
1059                 } else {
1060                         ib->ptr = (uint32_t *)kptr;
1061                         r = amdgpu_ring_patch_cs_in_place(ring, p, job, ib);
1062                         amdgpu_bo_kunmap(aobj);
1063                         if (r)
1064                                 return r;
1065                 }
1066         }
1067
1068         return 0;
1069 }
1070
1071 static int amdgpu_cs_patch_jobs(struct amdgpu_cs_parser *p)
1072 {
1073         unsigned int i;
1074         int r;
1075
1076         for (i = 0; i < p->gang_size; ++i) {
1077                 r = amdgpu_cs_patch_ibs(p, p->jobs[i]);
1078                 if (r)
1079                         return r;
1080         }
1081         return 0;
1082 }
1083
1084 static int amdgpu_cs_vm_handling(struct amdgpu_cs_parser *p)
1085 {
1086         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
1087         struct amdgpu_job *job = p->gang_leader;
1088         struct amdgpu_device *adev = p->adev;
1089         struct amdgpu_vm *vm = &fpriv->vm;
1090         struct amdgpu_bo_list_entry *e;
1091         struct amdgpu_bo_va *bo_va;
1092         struct amdgpu_bo *bo;
1093         unsigned int i;
1094         int r;
1095
1096         r = amdgpu_vm_clear_freed(adev, vm, NULL);
1097         if (r)
1098                 return r;
1099
1100         r = amdgpu_vm_bo_update(adev, fpriv->prt_va, false);
1101         if (r)
1102                 return r;
1103
1104         r = amdgpu_sync_fence(&job->sync, fpriv->prt_va->last_pt_update);
1105         if (r)
1106                 return r;
1107
1108         if (fpriv->csa_va) {
1109                 bo_va = fpriv->csa_va;
1110                 BUG_ON(!bo_va);
1111                 r = amdgpu_vm_bo_update(adev, bo_va, false);
1112                 if (r)
1113                         return r;
1114
1115                 r = amdgpu_sync_fence(&job->sync, bo_va->last_pt_update);
1116                 if (r)
1117                         return r;
1118         }
1119
1120         amdgpu_bo_list_for_each_entry(e, p->bo_list) {
1121                 /* ignore duplicates */
1122                 bo = ttm_to_amdgpu_bo(e->tv.bo);
1123                 if (!bo)
1124                         continue;
1125
1126                 bo_va = e->bo_va;
1127                 if (bo_va == NULL)
1128                         continue;
1129
1130                 r = amdgpu_vm_bo_update(adev, bo_va, false);
1131                 if (r)
1132                         return r;
1133
1134                 r = amdgpu_sync_fence(&job->sync, bo_va->last_pt_update);
1135                 if (r)
1136                         return r;
1137         }
1138
1139         r = amdgpu_vm_handle_moved(adev, vm);
1140         if (r)
1141                 return r;
1142
1143         r = amdgpu_vm_update_pdes(adev, vm, false);
1144         if (r)
1145                 return r;
1146
1147         r = amdgpu_sync_fence(&job->sync, vm->last_update);
1148         if (r)
1149                 return r;
1150
1151         for (i = 0; i < p->gang_size; ++i) {
1152                 job = p->jobs[i];
1153
1154                 if (!job->vm)
1155                         continue;
1156
1157                 job->vm_pd_addr = amdgpu_gmc_pd_addr(vm->root.bo);
1158         }
1159
1160         if (amdgpu_vm_debug) {
1161                 /* Invalidate all BOs to test for userspace bugs */
1162                 amdgpu_bo_list_for_each_entry(e, p->bo_list) {
1163                         struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
1164
1165                         /* ignore duplicates */
1166                         if (!bo)
1167                                 continue;
1168
1169                         amdgpu_vm_bo_invalidate(adev, bo, false);
1170                 }
1171         }
1172
1173         return 0;
1174 }
1175
1176 static int amdgpu_cs_sync_rings(struct amdgpu_cs_parser *p)
1177 {
1178         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
1179         struct amdgpu_job *leader = p->gang_leader;
1180         struct amdgpu_bo_list_entry *e;
1181         unsigned int i;
1182         int r;
1183
1184         list_for_each_entry(e, &p->validated, tv.head) {
1185                 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
1186                 struct dma_resv *resv = bo->tbo.base.resv;
1187                 enum amdgpu_sync_mode sync_mode;
1188
1189                 sync_mode = amdgpu_bo_explicit_sync(bo) ?
1190                         AMDGPU_SYNC_EXPLICIT : AMDGPU_SYNC_NE_OWNER;
1191                 r = amdgpu_sync_resv(p->adev, &leader->sync, resv, sync_mode,
1192                                      &fpriv->vm);
1193                 if (r)
1194                         return r;
1195         }
1196
1197         for (i = 0; i < p->gang_size - 1; ++i) {
1198                 r = amdgpu_sync_clone(&leader->sync, &p->jobs[i]->sync);
1199                 if (r)
1200                         return r;
1201         }
1202
1203         r = amdgpu_ctx_wait_prev_fence(p->ctx, p->entities[p->gang_size - 1]);
1204         if (r && r != -ERESTARTSYS)
1205                 DRM_ERROR("amdgpu_ctx_wait_prev_fence failed.\n");
1206
1207         return r;
1208 }
1209
1210 static void amdgpu_cs_post_dependencies(struct amdgpu_cs_parser *p)
1211 {
1212         int i;
1213
1214         for (i = 0; i < p->num_post_deps; ++i) {
1215                 if (p->post_deps[i].chain && p->post_deps[i].point) {
1216                         drm_syncobj_add_point(p->post_deps[i].syncobj,
1217                                               p->post_deps[i].chain,
1218                                               p->fence, p->post_deps[i].point);
1219                         p->post_deps[i].chain = NULL;
1220                 } else {
1221                         drm_syncobj_replace_fence(p->post_deps[i].syncobj,
1222                                                   p->fence);
1223                 }
1224         }
1225 }
1226
1227 static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
1228                             union drm_amdgpu_cs *cs)
1229 {
1230         struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
1231         struct amdgpu_job *leader = p->gang_leader;
1232         struct amdgpu_bo_list_entry *e;
1233         unsigned int i;
1234         uint64_t seq;
1235         int r;
1236
1237         for (i = 0; i < p->gang_size; ++i)
1238                 drm_sched_job_arm(&p->jobs[i]->base);
1239
1240         for (i = 0; i < (p->gang_size - 1); ++i) {
1241                 struct dma_fence *fence;
1242
1243                 fence = &p->jobs[i]->base.s_fence->scheduled;
1244                 r = amdgpu_sync_fence(&leader->sync, fence);
1245                 if (r)
1246                         goto error_cleanup;
1247         }
1248
1249         if (p->gang_size > 1) {
1250                 for (i = 0; i < p->gang_size; ++i)
1251                         amdgpu_job_set_gang_leader(p->jobs[i], leader);
1252         }
1253
1254         /* No memory allocation is allowed while holding the notifier lock.
1255          * The lock is held until amdgpu_cs_submit is finished and fence is
1256          * added to BOs.
1257          */
1258         mutex_lock(&p->adev->notifier_lock);
1259
1260         /* If userptr are invalidated after amdgpu_cs_parser_bos(), return
1261          * -EAGAIN, drmIoctl in libdrm will restart the amdgpu_cs_ioctl.
1262          */
1263         r = 0;
1264         amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
1265                 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
1266
1267                 r |= !amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm);
1268         }
1269         if (r) {
1270                 r = -EAGAIN;
1271                 goto error_unlock;
1272         }
1273
1274         p->fence = dma_fence_get(&leader->base.s_fence->finished);
1275         list_for_each_entry(e, &p->validated, tv.head) {
1276
1277                 /* Everybody except for the gang leader uses READ */
1278                 for (i = 0; i < (p->gang_size - 1); ++i) {
1279                         dma_resv_add_fence(e->tv.bo->base.resv,
1280                                            &p->jobs[i]->base.s_fence->finished,
1281                                            DMA_RESV_USAGE_READ);
1282                 }
1283
1284                 /* The gang leader is remembered as writer */
1285                 e->tv.num_shared = 0;
1286         }
1287
1288         seq = amdgpu_ctx_add_fence(p->ctx, p->entities[p->gang_size - 1],
1289                                    p->fence);
1290         amdgpu_cs_post_dependencies(p);
1291
1292         if ((leader->preamble_status & AMDGPU_PREAMBLE_IB_PRESENT) &&
1293             !p->ctx->preamble_presented) {
1294                 leader->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT_FIRST;
1295                 p->ctx->preamble_presented = true;
1296         }
1297
1298         cs->out.handle = seq;
1299         leader->uf_sequence = seq;
1300
1301         amdgpu_vm_bo_trace_cs(&fpriv->vm, &p->ticket);
1302         for (i = 0; i < p->gang_size; ++i) {
1303                 amdgpu_job_free_resources(p->jobs[i]);
1304                 trace_amdgpu_cs_ioctl(p->jobs[i]);
1305                 drm_sched_entity_push_job(&p->jobs[i]->base);
1306                 p->jobs[i] = NULL;
1307         }
1308
1309         amdgpu_vm_move_to_lru_tail(p->adev, &fpriv->vm);
1310         ttm_eu_fence_buffer_objects(&p->ticket, &p->validated, p->fence);
1311
1312         mutex_unlock(&p->adev->notifier_lock);
1313         mutex_unlock(&p->bo_list->bo_list_mutex);
1314         return 0;
1315
1316 error_unlock:
1317         mutex_unlock(&p->adev->notifier_lock);
1318
1319 error_cleanup:
1320         for (i = 0; i < p->gang_size; ++i)
1321                 drm_sched_job_cleanup(&p->jobs[i]->base);
1322         return r;
1323 }
1324
1325 /* Cleanup the parser structure */
1326 static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser)
1327 {
1328         unsigned i;
1329
1330         for (i = 0; i < parser->num_post_deps; i++) {
1331                 drm_syncobj_put(parser->post_deps[i].syncobj);
1332                 kfree(parser->post_deps[i].chain);
1333         }
1334         kfree(parser->post_deps);
1335
1336         dma_fence_put(parser->fence);
1337
1338         if (parser->ctx)
1339                 amdgpu_ctx_put(parser->ctx);
1340         if (parser->bo_list)
1341                 amdgpu_bo_list_put(parser->bo_list);
1342
1343         for (i = 0; i < parser->nchunks; i++)
1344                 kvfree(parser->chunks[i].kdata);
1345         kvfree(parser->chunks);
1346         for (i = 0; i < parser->gang_size; ++i) {
1347                 if (parser->jobs[i])
1348                         amdgpu_job_free(parser->jobs[i]);
1349         }
1350         if (parser->uf_entry.tv.bo) {
1351                 struct amdgpu_bo *uf = ttm_to_amdgpu_bo(parser->uf_entry.tv.bo);
1352
1353                 amdgpu_bo_unref(&uf);
1354         }
1355 }
1356
1357 int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
1358 {
1359         struct amdgpu_device *adev = drm_to_adev(dev);
1360         struct amdgpu_cs_parser parser;
1361         int r;
1362
1363         if (amdgpu_ras_intr_triggered())
1364                 return -EHWPOISON;
1365
1366         if (!adev->accel_working)
1367                 return -EBUSY;
1368
1369         r = amdgpu_cs_parser_init(&parser, adev, filp, data);
1370         if (r) {
1371                 if (printk_ratelimit())
1372                         DRM_ERROR("Failed to initialize parser %d!\n", r);
1373                 return r;
1374         }
1375
1376         r = amdgpu_cs_pass1(&parser, data);
1377         if (r)
1378                 goto error_fini;
1379
1380         r = amdgpu_cs_pass2(&parser);
1381         if (r)
1382                 goto error_fini;
1383
1384         r = amdgpu_cs_parser_bos(&parser, data);
1385         if (r) {
1386                 if (r == -ENOMEM)
1387                         DRM_ERROR("Not enough memory for command submission!\n");
1388                 else if (r != -ERESTARTSYS && r != -EAGAIN)
1389                         DRM_ERROR("Failed to process the buffer list %d!\n", r);
1390                 goto error_fini;
1391         }
1392
1393         r = amdgpu_cs_patch_jobs(&parser);
1394         if (r)
1395                 goto error_backoff;
1396
1397         r = amdgpu_cs_vm_handling(&parser);
1398         if (r)
1399                 goto error_backoff;
1400
1401         r = amdgpu_cs_sync_rings(&parser);
1402         if (r)
1403                 goto error_backoff;
1404
1405         trace_amdgpu_cs_ibs(&parser);
1406
1407         r = amdgpu_cs_submit(&parser, data);
1408         if (r)
1409                 goto error_backoff;
1410
1411         amdgpu_cs_parser_fini(&parser);
1412         return 0;
1413
1414 error_backoff:
1415         ttm_eu_backoff_reservation(&parser.ticket, &parser.validated);
1416         mutex_unlock(&parser.bo_list->bo_list_mutex);
1417
1418 error_fini:
1419         amdgpu_cs_parser_fini(&parser);
1420         return r;
1421 }
1422
1423 /**
1424  * amdgpu_cs_wait_ioctl - wait for a command submission to finish
1425  *
1426  * @dev: drm device
1427  * @data: data from userspace
1428  * @filp: file private
1429  *
1430  * Wait for the command submission identified by handle to finish.
1431  */
1432 int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data,
1433                          struct drm_file *filp)
1434 {
1435         union drm_amdgpu_wait_cs *wait = data;
1436         unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
1437         struct drm_sched_entity *entity;
1438         struct amdgpu_ctx *ctx;
1439         struct dma_fence *fence;
1440         long r;
1441
1442         ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
1443         if (ctx == NULL)
1444                 return -EINVAL;
1445
1446         r = amdgpu_ctx_get_entity(ctx, wait->in.ip_type, wait->in.ip_instance,
1447                                   wait->in.ring, &entity);
1448         if (r) {
1449                 amdgpu_ctx_put(ctx);
1450                 return r;
1451         }
1452
1453         fence = amdgpu_ctx_get_fence(ctx, entity, wait->in.handle);
1454         if (IS_ERR(fence))
1455                 r = PTR_ERR(fence);
1456         else if (fence) {
1457                 r = dma_fence_wait_timeout(fence, true, timeout);
1458                 if (r > 0 && fence->error)
1459                         r = fence->error;
1460                 dma_fence_put(fence);
1461         } else
1462                 r = 1;
1463
1464         amdgpu_ctx_put(ctx);
1465         if (r < 0)
1466                 return r;
1467
1468         memset(wait, 0, sizeof(*wait));
1469         wait->out.status = (r == 0);
1470
1471         return 0;
1472 }
1473
1474 /**
1475  * amdgpu_cs_get_fence - helper to get fence from drm_amdgpu_fence
1476  *
1477  * @adev: amdgpu device
1478  * @filp: file private
1479  * @user: drm_amdgpu_fence copied from user space
1480  */
1481 static struct dma_fence *amdgpu_cs_get_fence(struct amdgpu_device *adev,
1482                                              struct drm_file *filp,
1483                                              struct drm_amdgpu_fence *user)
1484 {
1485         struct drm_sched_entity *entity;
1486         struct amdgpu_ctx *ctx;
1487         struct dma_fence *fence;
1488         int r;
1489
1490         ctx = amdgpu_ctx_get(filp->driver_priv, user->ctx_id);
1491         if (ctx == NULL)
1492                 return ERR_PTR(-EINVAL);
1493
1494         r = amdgpu_ctx_get_entity(ctx, user->ip_type, user->ip_instance,
1495                                   user->ring, &entity);
1496         if (r) {
1497                 amdgpu_ctx_put(ctx);
1498                 return ERR_PTR(r);
1499         }
1500
1501         fence = amdgpu_ctx_get_fence(ctx, entity, user->seq_no);
1502         amdgpu_ctx_put(ctx);
1503
1504         return fence;
1505 }
1506
1507 int amdgpu_cs_fence_to_handle_ioctl(struct drm_device *dev, void *data,
1508                                     struct drm_file *filp)
1509 {
1510         struct amdgpu_device *adev = drm_to_adev(dev);
1511         union drm_amdgpu_fence_to_handle *info = data;
1512         struct dma_fence *fence;
1513         struct drm_syncobj *syncobj;
1514         struct sync_file *sync_file;
1515         int fd, r;
1516
1517         fence = amdgpu_cs_get_fence(adev, filp, &info->in.fence);
1518         if (IS_ERR(fence))
1519                 return PTR_ERR(fence);
1520
1521         if (!fence)
1522                 fence = dma_fence_get_stub();
1523
1524         switch (info->in.what) {
1525         case AMDGPU_FENCE_TO_HANDLE_GET_SYNCOBJ:
1526                 r = drm_syncobj_create(&syncobj, 0, fence);
1527                 dma_fence_put(fence);
1528                 if (r)
1529                         return r;
1530                 r = drm_syncobj_get_handle(filp, syncobj, &info->out.handle);
1531                 drm_syncobj_put(syncobj);
1532                 return r;
1533
1534         case AMDGPU_FENCE_TO_HANDLE_GET_SYNCOBJ_FD:
1535                 r = drm_syncobj_create(&syncobj, 0, fence);
1536                 dma_fence_put(fence);
1537                 if (r)
1538                         return r;
1539                 r = drm_syncobj_get_fd(syncobj, (int *)&info->out.handle);
1540                 drm_syncobj_put(syncobj);
1541                 return r;
1542
1543         case AMDGPU_FENCE_TO_HANDLE_GET_SYNC_FILE_FD:
1544                 fd = get_unused_fd_flags(O_CLOEXEC);
1545                 if (fd < 0) {
1546                         dma_fence_put(fence);
1547                         return fd;
1548                 }
1549
1550                 sync_file = sync_file_create(fence);
1551                 dma_fence_put(fence);
1552                 if (!sync_file) {
1553                         put_unused_fd(fd);
1554                         return -ENOMEM;
1555                 }
1556
1557                 fd_install(fd, sync_file->file);
1558                 info->out.handle = fd;
1559                 return 0;
1560
1561         default:
1562                 dma_fence_put(fence);
1563                 return -EINVAL;
1564         }
1565 }
1566
1567 /**
1568  * amdgpu_cs_wait_all_fences - wait on all fences to signal
1569  *
1570  * @adev: amdgpu device
1571  * @filp: file private
1572  * @wait: wait parameters
1573  * @fences: array of drm_amdgpu_fence
1574  */
1575 static int amdgpu_cs_wait_all_fences(struct amdgpu_device *adev,
1576                                      struct drm_file *filp,
1577                                      union drm_amdgpu_wait_fences *wait,
1578                                      struct drm_amdgpu_fence *fences)
1579 {
1580         uint32_t fence_count = wait->in.fence_count;
1581         unsigned int i;
1582         long r = 1;
1583
1584         for (i = 0; i < fence_count; i++) {
1585                 struct dma_fence *fence;
1586                 unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1587
1588                 fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1589                 if (IS_ERR(fence))
1590                         return PTR_ERR(fence);
1591                 else if (!fence)
1592                         continue;
1593
1594                 r = dma_fence_wait_timeout(fence, true, timeout);
1595                 dma_fence_put(fence);
1596                 if (r < 0)
1597                         return r;
1598
1599                 if (r == 0)
1600                         break;
1601
1602                 if (fence->error)
1603                         return fence->error;
1604         }
1605
1606         memset(wait, 0, sizeof(*wait));
1607         wait->out.status = (r > 0);
1608
1609         return 0;
1610 }
1611
1612 /**
1613  * amdgpu_cs_wait_any_fence - wait on any fence to signal
1614  *
1615  * @adev: amdgpu device
1616  * @filp: file private
1617  * @wait: wait parameters
1618  * @fences: array of drm_amdgpu_fence
1619  */
1620 static int amdgpu_cs_wait_any_fence(struct amdgpu_device *adev,
1621                                     struct drm_file *filp,
1622                                     union drm_amdgpu_wait_fences *wait,
1623                                     struct drm_amdgpu_fence *fences)
1624 {
1625         unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1626         uint32_t fence_count = wait->in.fence_count;
1627         uint32_t first = ~0;
1628         struct dma_fence **array;
1629         unsigned int i;
1630         long r;
1631
1632         /* Prepare the fence array */
1633         array = kcalloc(fence_count, sizeof(struct dma_fence *), GFP_KERNEL);
1634
1635         if (array == NULL)
1636                 return -ENOMEM;
1637
1638         for (i = 0; i < fence_count; i++) {
1639                 struct dma_fence *fence;
1640
1641                 fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1642                 if (IS_ERR(fence)) {
1643                         r = PTR_ERR(fence);
1644                         goto err_free_fence_array;
1645                 } else if (fence) {
1646                         array[i] = fence;
1647                 } else { /* NULL, the fence has been already signaled */
1648                         r = 1;
1649                         first = i;
1650                         goto out;
1651                 }
1652         }
1653
1654         r = dma_fence_wait_any_timeout(array, fence_count, true, timeout,
1655                                        &first);
1656         if (r < 0)
1657                 goto err_free_fence_array;
1658
1659 out:
1660         memset(wait, 0, sizeof(*wait));
1661         wait->out.status = (r > 0);
1662         wait->out.first_signaled = first;
1663
1664         if (first < fence_count && array[first])
1665                 r = array[first]->error;
1666         else
1667                 r = 0;
1668
1669 err_free_fence_array:
1670         for (i = 0; i < fence_count; i++)
1671                 dma_fence_put(array[i]);
1672         kfree(array);
1673
1674         return r;
1675 }
1676
1677 /**
1678  * amdgpu_cs_wait_fences_ioctl - wait for multiple command submissions to finish
1679  *
1680  * @dev: drm device
1681  * @data: data from userspace
1682  * @filp: file private
1683  */
1684 int amdgpu_cs_wait_fences_ioctl(struct drm_device *dev, void *data,
1685                                 struct drm_file *filp)
1686 {
1687         struct amdgpu_device *adev = drm_to_adev(dev);
1688         union drm_amdgpu_wait_fences *wait = data;
1689         uint32_t fence_count = wait->in.fence_count;
1690         struct drm_amdgpu_fence *fences_user;
1691         struct drm_amdgpu_fence *fences;
1692         int r;
1693
1694         /* Get the fences from userspace */
1695         fences = kmalloc_array(fence_count, sizeof(struct drm_amdgpu_fence),
1696                         GFP_KERNEL);
1697         if (fences == NULL)
1698                 return -ENOMEM;
1699
1700         fences_user = u64_to_user_ptr(wait->in.fences);
1701         if (copy_from_user(fences, fences_user,
1702                 sizeof(struct drm_amdgpu_fence) * fence_count)) {
1703                 r = -EFAULT;
1704                 goto err_free_fences;
1705         }
1706
1707         if (wait->in.wait_all)
1708                 r = amdgpu_cs_wait_all_fences(adev, filp, wait, fences);
1709         else
1710                 r = amdgpu_cs_wait_any_fence(adev, filp, wait, fences);
1711
1712 err_free_fences:
1713         kfree(fences);
1714
1715         return r;
1716 }
1717
1718 /**
1719  * amdgpu_cs_find_mapping - find bo_va for VM address
1720  *
1721  * @parser: command submission parser context
1722  * @addr: VM address
1723  * @bo: resulting BO of the mapping found
1724  * @map: Placeholder to return found BO mapping
1725  *
1726  * Search the buffer objects in the command submission context for a certain
1727  * virtual memory address. Returns allocation structure when found, NULL
1728  * otherwise.
1729  */
1730 int amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
1731                            uint64_t addr, struct amdgpu_bo **bo,
1732                            struct amdgpu_bo_va_mapping **map)
1733 {
1734         struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
1735         struct ttm_operation_ctx ctx = { false, false };
1736         struct amdgpu_vm *vm = &fpriv->vm;
1737         struct amdgpu_bo_va_mapping *mapping;
1738         int r;
1739
1740         addr /= AMDGPU_GPU_PAGE_SIZE;
1741
1742         mapping = amdgpu_vm_bo_lookup_mapping(vm, addr);
1743         if (!mapping || !mapping->bo_va || !mapping->bo_va->base.bo)
1744                 return -EINVAL;
1745
1746         *bo = mapping->bo_va->base.bo;
1747         *map = mapping;
1748
1749         /* Double check that the BO is reserved by this CS */
1750         if (dma_resv_locking_ctx((*bo)->tbo.base.resv) != &parser->ticket)
1751                 return -EINVAL;
1752
1753         if (!((*bo)->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)) {
1754                 (*bo)->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
1755                 amdgpu_bo_placement_from_domain(*bo, (*bo)->allowed_domains);
1756                 r = ttm_bo_validate(&(*bo)->tbo, &(*bo)->placement, &ctx);
1757                 if (r)
1758                         return r;
1759         }
1760
1761         return amdgpu_ttm_alloc_gart(&(*bo)->tbo);
1762 }