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

   1 /*
   2  * Copyright 2015 Advanced Micro Devices, Inc.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20  * OTHER DEALINGS IN THE SOFTWARE.
  21  *
  22  *
  23  */
  24 #include <linux/kthread.h>
  25 #include <linux/wait.h>
  26 #include <linux/sched.h>
  27
  28 #include <drm/drm_drv.h>
  29
  30 #include "amdgpu.h"
  31 #include "amdgpu_trace.h"
  32 #include "amdgpu_reset.h"
  33 #include "amdgpu_dev_coredump.h"
  34 #include "amdgpu_xgmi.h"
  35
  36 static void amdgpu_job_do_core_dump(struct amdgpu_device *adev,
  37                                     struct amdgpu_job *job)
  38 {
  39         int i;
  40
  41         dev_info(adev->dev, "Dumping IP State\n");
  42         for (i = 0; i < adev->num_ip_blocks; i++)
  43                 if (adev->ip_blocks[i].version->funcs->dump_ip_state)
  44                         adev->ip_blocks[i].version->funcs
  45                                 ->dump_ip_state((void *)&adev->ip_blocks[i]);
  46         dev_info(adev->dev, "Dumping IP State Completed\n");
  47
  48         amdgpu_coredump(adev, true, false, job);
  49 }
  50
  51 static void amdgpu_job_core_dump(struct amdgpu_device *adev,
  52                                  struct amdgpu_job *job)
  53 {
  54         struct list_head device_list, *device_list_handle =  NULL;
  55         struct amdgpu_device *tmp_adev = NULL;
  56         struct amdgpu_hive_info *hive = NULL;
  57
  58         if (!amdgpu_sriov_vf(adev))
  59                 hive = amdgpu_get_xgmi_hive(adev);
  60         if (hive)
  61                 mutex_lock(&hive->hive_lock);
  62         /*
  63          * Reuse the logic in amdgpu_device_gpu_recover() to build list of
  64          * devices for code dump
  65          */
  66         INIT_LIST_HEAD(&device_list);
  67         if (!amdgpu_sriov_vf(adev) && (adev->gmc.xgmi.num_physical_nodes > 1) && hive) {
  68                 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head)
  69                         list_add_tail(&tmp_adev->reset_list, &device_list);
  70                 if (!list_is_first(&adev->reset_list, &device_list))
  71                         list_rotate_to_front(&adev->reset_list, &device_list);
  72                 device_list_handle = &device_list;
  73         } else {
  74                 list_add_tail(&adev->reset_list, &device_list);
  75                 device_list_handle = &device_list;
  76         }
  77
  78         /* Do the coredump for each device */
  79         list_for_each_entry(tmp_adev, device_list_handle, reset_list)
  80                 amdgpu_job_do_core_dump(tmp_adev, job);
  81
  82         if (hive) {
  83                 mutex_unlock(&hive->hive_lock);
  84                 amdgpu_put_xgmi_hive(hive);
  85         }
  86 }
  87
  88 static enum drm_gpu_sched_stat amdgpu_job_timedout(struct drm_sched_job *s_job)
  89 {
  90         struct amdgpu_ring *ring = to_amdgpu_ring(s_job->sched);
  91         struct amdgpu_job *job = to_amdgpu_job(s_job);
  92         struct amdgpu_task_info *ti;
  93         struct amdgpu_device *adev = ring->adev;
  94         int idx;
  95         int r;
  96
  97         if (!drm_dev_enter(adev_to_drm(adev), &idx)) {
  98                 dev_info(adev->dev, "%s - device unplugged skipping recovery on scheduler:%s",
  99                          __func__, s_job->sched->name);
 100
 101                 /* Effectively the job is aborted as the device is gone */
 102                 return DRM_GPU_SCHED_STAT_ENODEV;
 103         }
 104
 105         /*
 106          * Do the coredump immediately after a job timeout to get a very
 107          * close dump/snapshot/representation of GPU's current error status
 108          * Skip it for SRIOV, since VF FLR will be triggered by host driver
 109          * before job timeout
 110          */
 111         if (!amdgpu_sriov_vf(adev))
 112                 amdgpu_job_core_dump(adev, job);
 113
 114         if (amdgpu_gpu_recovery &&
 115             amdgpu_ring_soft_recovery(ring, job->vmid, s_job->s_fence->parent)) {
 116                 dev_err(adev->dev, "ring %s timeout, but soft recovered\n",
 117                         s_job->sched->name);
 118                 goto exit;
 119         }
 120
 121         dev_err(adev->dev, "ring %s timeout, signaled seq=%u, emitted seq=%u\n",
 122                 job->base.sched->name, atomic_read(&ring->fence_drv.last_seq),
 123                 ring->fence_drv.sync_seq);
 124
 125         ti = amdgpu_vm_get_task_info_pasid(ring->adev, job->pasid);
 126         if (ti) {
 127                 dev_err(adev->dev,
 128                         "Process information: process %s pid %d thread %s pid %d\n",
 129                         ti->process_name, ti->tgid, ti->task_name, ti->pid);
 130                 amdgpu_vm_put_task_info(ti);
 131         }
 132
 133         dma_fence_set_error(&s_job->s_fence->finished, -ETIME);
 134
 135         /* attempt a per ring reset */
 136         if (amdgpu_gpu_recovery &&
 137             ring->funcs->reset) {
 138                 dev_err(adev->dev, "Starting %s ring reset\n", s_job->sched->name);
 139                 /* stop the scheduler, but don't mess with the
 140                  * bad job yet because if ring reset fails
 141                  * we'll fall back to full GPU reset.
 142                  */
 143                 drm_sched_wqueue_stop(&ring->sched);
 144                 r = amdgpu_ring_reset(ring, job->vmid);
 145                 if (!r) {
 146                         if (amdgpu_ring_sched_ready(ring))
 147                                 drm_sched_stop(&ring->sched, s_job);
 148                         atomic_inc(&ring->adev->gpu_reset_counter);
 149                         amdgpu_fence_driver_force_completion(ring);
 150                         if (amdgpu_ring_sched_ready(ring))
 151                                 drm_sched_start(&ring->sched, 0);
 152                         goto exit;
 153                 }
 154                 dev_err(adev->dev, "Ring %s reset failure\n", ring->sched.name);
 155         }
 156
 157         if (amdgpu_device_should_recover_gpu(ring->adev)) {
 158                 struct amdgpu_reset_context reset_context;
 159                 memset(&reset_context, 0, sizeof(reset_context));
 160
 161                 reset_context.method = AMD_RESET_METHOD_NONE;
 162                 reset_context.reset_req_dev = adev;
 163                 reset_context.src = AMDGPU_RESET_SRC_JOB;
 164                 clear_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags);
 165
 166                 /*
 167                  * To avoid an unnecessary extra coredump, as we have already
 168                  * got the very close representation of GPU's error status
 169                  */
 170                 set_bit(AMDGPU_SKIP_COREDUMP, &reset_context.flags);
 171
 172                 r = amdgpu_device_gpu_recover(ring->adev, job, &reset_context);
 173                 if (r)
 174                         dev_err(adev->dev, "GPU Recovery Failed: %d\n", r);
 175         } else {
 176                 drm_sched_suspend_timeout(&ring->sched);
 177                 if (amdgpu_sriov_vf(adev))
 178                         adev->virt.tdr_debug = true;
 179         }
 180
 181 exit:
 182         drm_dev_exit(idx);
 183         return DRM_GPU_SCHED_STAT_NOMINAL;
 184 }
 185
 186 int amdgpu_job_alloc(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 187                      struct drm_sched_entity *entity, void *owner,
 188                      unsigned int num_ibs, struct amdgpu_job **job)
 189 {
 190         if (num_ibs == 0)
 191                 return -EINVAL;
 192
 193         *job = kzalloc(struct_size(*job, ibs, num_ibs), GFP_KERNEL);
 194         if (!*job)
 195                 return -ENOMEM;
 196
 197         (*job)->vm = vm;
 198
 199         amdgpu_sync_create(&(*job)->explicit_sync);
 200         (*job)->generation = amdgpu_vm_generation(adev, vm);
 201         (*job)->vm_pd_addr = AMDGPU_BO_INVALID_OFFSET;
 202
 203         if (!entity)
 204                 return 0;
 205
 206         return drm_sched_job_init(&(*job)->base, entity, 1, owner);
 207 }
 208
 209 int amdgpu_job_alloc_with_ib(struct amdgpu_device *adev,
 210                              struct drm_sched_entity *entity, void *owner,
 211                              size_t size, enum amdgpu_ib_pool_type pool_type,
 212                              struct amdgpu_job **job)
 213 {
 214         int r;
 215
 216         r = amdgpu_job_alloc(adev, NULL, entity, owner, 1, job);
 217         if (r)
 218                 return r;
 219
 220         (*job)->num_ibs = 1;
 221         r = amdgpu_ib_get(adev, NULL, size, pool_type, &(*job)->ibs[0]);
 222         if (r) {
 223                 if (entity)
 224                         drm_sched_job_cleanup(&(*job)->base);
 225                 kfree(*job);
 226         }
 227
 228         return r;
 229 }
 230
 231 void amdgpu_job_set_resources(struct amdgpu_job *job, struct amdgpu_bo *gds,
 232                               struct amdgpu_bo *gws, struct amdgpu_bo *oa)
 233 {
 234         if (gds) {
 235                 job->gds_base = amdgpu_bo_gpu_offset(gds) >> PAGE_SHIFT;
 236                 job->gds_size = amdgpu_bo_size(gds) >> PAGE_SHIFT;
 237         }
 238         if (gws) {
 239                 job->gws_base = amdgpu_bo_gpu_offset(gws) >> PAGE_SHIFT;
 240                 job->gws_size = amdgpu_bo_size(gws) >> PAGE_SHIFT;
 241         }
 242         if (oa) {
 243                 job->oa_base = amdgpu_bo_gpu_offset(oa) >> PAGE_SHIFT;
 244                 job->oa_size = amdgpu_bo_size(oa) >> PAGE_SHIFT;
 245         }
 246 }
 247
 248 void amdgpu_job_free_resources(struct amdgpu_job *job)
 249 {
 250         struct dma_fence *f;
 251         unsigned i;
 252
 253         /* Check if any fences where initialized */
 254         if (job->base.s_fence && job->base.s_fence->finished.ops)
 255                 f = &job->base.s_fence->finished;
 256         else if (job->hw_fence.ops)
 257                 f = &job->hw_fence;
 258         else
 259                 f = NULL;
 260
 261         for (i = 0; i < job->num_ibs; ++i)
 262                 amdgpu_ib_free(&job->ibs[i], f);
 263 }
 264
 265 static void amdgpu_job_free_cb(struct drm_sched_job *s_job)
 266 {
 267         struct amdgpu_job *job = to_amdgpu_job(s_job);
 268
 269         drm_sched_job_cleanup(s_job);
 270
 271         amdgpu_sync_free(&job->explicit_sync);
 272
 273         /* only put the hw fence if has embedded fence */
 274         if (!job->hw_fence.ops)
 275                 kfree(job);
 276         else
 277                 dma_fence_put(&job->hw_fence);
 278 }
 279
 280 void amdgpu_job_set_gang_leader(struct amdgpu_job *job,
 281                                 struct amdgpu_job *leader)
 282 {
 283         struct dma_fence *fence = &leader->base.s_fence->scheduled;
 284
 285         WARN_ON(job->gang_submit);
 286
 287         /*
 288          * Don't add a reference when we are the gang leader to avoid circle
 289          * dependency.
 290          */
 291         if (job != leader)
 292                 dma_fence_get(fence);
 293         job->gang_submit = fence;
 294 }
 295
 296 void amdgpu_job_free(struct amdgpu_job *job)
 297 {
 298         if (job->base.entity)
 299                 drm_sched_job_cleanup(&job->base);
 300
 301         amdgpu_job_free_resources(job);
 302         amdgpu_sync_free(&job->explicit_sync);
 303         if (job->gang_submit != &job->base.s_fence->scheduled)
 304                 dma_fence_put(job->gang_submit);
 305
 306         if (!job->hw_fence.ops)
 307                 kfree(job);
 308         else
 309                 dma_fence_put(&job->hw_fence);
 310 }
 311
 312 struct dma_fence *amdgpu_job_submit(struct amdgpu_job *job)
 313 {
 314         struct dma_fence *f;
 315
 316         drm_sched_job_arm(&job->base);
 317         f = dma_fence_get(&job->base.s_fence->finished);
 318         amdgpu_job_free_resources(job);
 319         drm_sched_entity_push_job(&job->base);
 320
 321         return f;
 322 }
 323
 324 int amdgpu_job_submit_direct(struct amdgpu_job *job, struct amdgpu_ring *ring,
 325                              struct dma_fence **fence)
 326 {
 327         int r;
 328
 329         job->base.sched = &ring->sched;
 330         r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs, job, fence);
 331
 332         if (r)
 333                 return r;
 334
 335         amdgpu_job_free(job);
 336         return 0;
 337 }
 338
 339 static struct dma_fence *
 340 amdgpu_job_prepare_job(struct drm_sched_job *sched_job,
 341                       struct drm_sched_entity *s_entity)
 342 {
 343         struct amdgpu_ring *ring = to_amdgpu_ring(s_entity->rq->sched);
 344         struct amdgpu_job *job = to_amdgpu_job(sched_job);
 345         struct dma_fence *fence = NULL;
 346         int r;
 347
 348         r = drm_sched_entity_error(s_entity);
 349         if (r)
 350                 goto error;
 351
 352         if (job->gang_submit)
 353                 fence = amdgpu_device_switch_gang(ring->adev, job->gang_submit);
 354
 355         if (!fence && job->vm && !job->vmid) {
 356                 r = amdgpu_vmid_grab(job->vm, ring, job, &fence);
 357                 if (r) {
 358                         dev_err(ring->adev->dev, "Error getting VM ID (%d)\n", r);
 359                         goto error;
 360                 }
 361                 /*
 362                  * The VM structure might be released after the VMID is
 363                  * assigned, we had multiple problems with people trying to use
 364                  * the VM pointer so better set it to NULL.
 365                  */
 366                 if (!fence)
 367                         job->vm = NULL;
 368         }
 369
 370         return fence;
 371
 372 error:
 373         dma_fence_set_error(&job->base.s_fence->finished, r);
 374         return NULL;
 375 }
 376
 377 static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job)
 378 {
 379         struct amdgpu_ring *ring = to_amdgpu_ring(sched_job->sched);
 380         struct amdgpu_device *adev = ring->adev;
 381         struct dma_fence *fence = NULL, *finished;
 382         struct amdgpu_job *job;
 383         int r = 0;
 384
 385         job = to_amdgpu_job(sched_job);
 386         finished = &job->base.s_fence->finished;
 387
 388         trace_amdgpu_sched_run_job(job);
 389
 390         /* Skip job if VRAM is lost and never resubmit gangs */
 391         if (job->generation != amdgpu_vm_generation(adev, job->vm) ||
 392             (job->job_run_counter && job->gang_submit))
 393                 dma_fence_set_error(finished, -ECANCELED);
 394
 395         if (finished->error < 0) {
 396                 dev_dbg(adev->dev, "Skip scheduling IBs in ring(%s)",
 397                         ring->name);
 398         } else {
 399                 r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs, job,
 400                                        &fence);
 401                 if (r)
 402                         dev_err(adev->dev,
 403                                 "Error scheduling IBs (%d) in ring(%s)", r,
 404                                 ring->name);
 405         }
 406
 407         job->job_run_counter++;
 408         amdgpu_job_free_resources(job);
 409
 410         fence = r ? ERR_PTR(r) : fence;
 411         return fence;
 412 }
 413
 414 #define to_drm_sched_job(sched_job)             \
 415                 container_of((sched_job), struct drm_sched_job, queue_node)
 416
 417 void amdgpu_job_stop_all_jobs_on_sched(struct drm_gpu_scheduler *sched)
 418 {
 419         struct drm_sched_job *s_job;
 420         struct drm_sched_entity *s_entity = NULL;
 421         int i;
 422
 423         /* Signal all jobs not yet scheduled */
 424         for (i = DRM_SCHED_PRIORITY_KERNEL; i < sched->num_rqs; i++) {
 425                 struct drm_sched_rq *rq = sched->sched_rq[i];
 426                 spin_lock(&rq->lock);
 427                 list_for_each_entry(s_entity, &rq->entities, list) {
 428                         while ((s_job = to_drm_sched_job(spsc_queue_pop(&s_entity->job_queue)))) {
 429                                 struct drm_sched_fence *s_fence = s_job->s_fence;
 430
 431                                 dma_fence_signal(&s_fence->scheduled);
 432                                 dma_fence_set_error(&s_fence->finished, -EHWPOISON);
 433                                 dma_fence_signal(&s_fence->finished);
 434                         }
 435                 }
 436                 spin_unlock(&rq->lock);
 437         }
 438
 439         /* Signal all jobs already scheduled to HW */
 440         list_for_each_entry(s_job, &sched->pending_list, list) {
 441                 struct drm_sched_fence *s_fence = s_job->s_fence;
 442
 443                 dma_fence_set_error(&s_fence->finished, -EHWPOISON);
 444                 dma_fence_signal(&s_fence->finished);
 445         }
 446 }
 447
 448 const struct drm_sched_backend_ops amdgpu_sched_ops = {
 449         .prepare_job = amdgpu_job_prepare_job,
 450         .run_job = amdgpu_job_run,
 451         .timedout_job = amdgpu_job_timedout,
 452         .free_job = amdgpu_job_free_cb
 453 };