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 34 static enum drm_gpu_sched_stat amdgpu_job_timedout(struct drm_sched_job *s_job) 35 { 36 struct amdgpu_ring *ring = to_amdgpu_ring(s_job->sched); 37 struct amdgpu_job *job = to_amdgpu_job(s_job); 38 struct amdgpu_task_info ti; 39 struct amdgpu_device *adev = ring->adev; 40 int idx; 41 int r; 42 43 if (!drm_dev_enter(adev_to_drm(adev), &idx)) { 44 DRM_INFO("%s - device unplugged skipping recovery on scheduler:%s", 45 __func__, s_job->sched->name); 46 47 /* Effectively the job is aborted as the device is gone */ 48 return DRM_GPU_SCHED_STAT_ENODEV; 49 } 50 51 memset(&ti, 0, sizeof(struct amdgpu_task_info)); 52 adev->job_hang = true; 53 54 if (amdgpu_gpu_recovery && 55 amdgpu_ring_soft_recovery(ring, job->vmid, s_job->s_fence->parent)) { 56 DRM_ERROR("ring %s timeout, but soft recovered\n", 57 s_job->sched->name); 58 goto exit; 59 } 60 61 amdgpu_vm_get_task_info(ring->adev, job->pasid, &ti); 62 DRM_ERROR("ring %s timeout, signaled seq=%u, emitted seq=%u\n", 63 job->base.sched->name, atomic_read(&ring->fence_drv.last_seq), 64 ring->fence_drv.sync_seq); 65 DRM_ERROR("Process information: process %s pid %d thread %s pid %d\n", 66 ti.process_name, ti.tgid, ti.task_name, ti.pid); 67 68 dma_fence_set_error(&s_job->s_fence->finished, -ETIME); 69 70 if (amdgpu_device_should_recover_gpu(ring->adev)) { 71 struct amdgpu_reset_context reset_context; 72 memset(&reset_context, 0, sizeof(reset_context)); 73 74 reset_context.method = AMD_RESET_METHOD_NONE; 75 reset_context.reset_req_dev = adev; 76 clear_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags); 77 78 r = amdgpu_device_gpu_recover(ring->adev, job, &reset_context); 79 if (r) 80 DRM_ERROR("GPU Recovery Failed: %d\n", r); 81 } else { 82 drm_sched_suspend_timeout(&ring->sched); 83 if (amdgpu_sriov_vf(adev)) 84 adev->virt.tdr_debug = true; 85 } 86 87 exit: 88 adev->job_hang = false; 89 drm_dev_exit(idx); 90 return DRM_GPU_SCHED_STAT_NOMINAL; 91 } 92 93 int amdgpu_job_alloc(struct amdgpu_device *adev, struct amdgpu_vm *vm, 94 struct drm_sched_entity *entity, void *owner, 95 unsigned int num_ibs, struct amdgpu_job **job) 96 { 97 if (num_ibs == 0) 98 return -EINVAL; 99 100 *job = kzalloc(struct_size(*job, ibs, num_ibs), GFP_KERNEL); 101 if (!*job) 102 return -ENOMEM; 103 104 /* 105 * Initialize the scheduler to at least some ring so that we always 106 * have a pointer to adev. 107 */ 108 (*job)->base.sched = &adev->rings[0]->sched; 109 (*job)->vm = vm; 110 111 amdgpu_sync_create(&(*job)->explicit_sync); 112 (*job)->generation = amdgpu_vm_generation(adev, vm); 113 (*job)->vm_pd_addr = AMDGPU_BO_INVALID_OFFSET; 114 115 if (!entity) 116 return 0; 117 118 return drm_sched_job_init(&(*job)->base, entity, owner); 119 } 120 121 int amdgpu_job_alloc_with_ib(struct amdgpu_device *adev, 122 struct drm_sched_entity *entity, void *owner, 123 size_t size, enum amdgpu_ib_pool_type pool_type, 124 struct amdgpu_job **job) 125 { 126 int r; 127 128 r = amdgpu_job_alloc(adev, NULL, entity, owner, 1, job); 129 if (r) 130 return r; 131 132 (*job)->num_ibs = 1; 133 r = amdgpu_ib_get(adev, NULL, size, pool_type, &(*job)->ibs[0]); 134 if (r) { 135 if (entity) 136 drm_sched_job_cleanup(&(*job)->base); 137 kfree(*job); 138 } 139 140 return r; 141 } 142 143 void amdgpu_job_set_resources(struct amdgpu_job *job, struct amdgpu_bo *gds, 144 struct amdgpu_bo *gws, struct amdgpu_bo *oa) 145 { 146 if (gds) { 147 job->gds_base = amdgpu_bo_gpu_offset(gds) >> PAGE_SHIFT; 148 job->gds_size = amdgpu_bo_size(gds) >> PAGE_SHIFT; 149 } 150 if (gws) { 151 job->gws_base = amdgpu_bo_gpu_offset(gws) >> PAGE_SHIFT; 152 job->gws_size = amdgpu_bo_size(gws) >> PAGE_SHIFT; 153 } 154 if (oa) { 155 job->oa_base = amdgpu_bo_gpu_offset(oa) >> PAGE_SHIFT; 156 job->oa_size = amdgpu_bo_size(oa) >> PAGE_SHIFT; 157 } 158 } 159 160 void amdgpu_job_free_resources(struct amdgpu_job *job) 161 { 162 struct dma_fence *f; 163 unsigned i; 164 165 /* Check if any fences where initialized */ 166 if (job->base.s_fence && job->base.s_fence->finished.ops) 167 f = &job->base.s_fence->finished; 168 else if (job->hw_fence.ops) 169 f = &job->hw_fence; 170 else 171 f = NULL; 172 173 for (i = 0; i < job->num_ibs; ++i) 174 amdgpu_ib_free(NULL, &job->ibs[i], f); 175 } 176 177 static void amdgpu_job_free_cb(struct drm_sched_job *s_job) 178 { 179 struct amdgpu_job *job = to_amdgpu_job(s_job); 180 181 drm_sched_job_cleanup(s_job); 182 183 amdgpu_sync_free(&job->explicit_sync); 184 185 /* only put the hw fence if has embedded fence */ 186 if (!job->hw_fence.ops) 187 kfree(job); 188 else 189 dma_fence_put(&job->hw_fence); 190 } 191 192 void amdgpu_job_set_gang_leader(struct amdgpu_job *job, 193 struct amdgpu_job *leader) 194 { 195 struct dma_fence *fence = &leader->base.s_fence->scheduled; 196 197 WARN_ON(job->gang_submit); 198 199 /* 200 * Don't add a reference when we are the gang leader to avoid circle 201 * dependency. 202 */ 203 if (job != leader) 204 dma_fence_get(fence); 205 job->gang_submit = fence; 206 } 207 208 void amdgpu_job_free(struct amdgpu_job *job) 209 { 210 if (job->base.entity) 211 drm_sched_job_cleanup(&job->base); 212 213 amdgpu_job_free_resources(job); 214 amdgpu_sync_free(&job->explicit_sync); 215 if (job->gang_submit != &job->base.s_fence->scheduled) 216 dma_fence_put(job->gang_submit); 217 218 if (!job->hw_fence.ops) 219 kfree(job); 220 else 221 dma_fence_put(&job->hw_fence); 222 } 223 224 struct dma_fence *amdgpu_job_submit(struct amdgpu_job *job) 225 { 226 struct dma_fence *f; 227 228 drm_sched_job_arm(&job->base); 229 f = dma_fence_get(&job->base.s_fence->finished); 230 amdgpu_job_free_resources(job); 231 drm_sched_entity_push_job(&job->base); 232 233 return f; 234 } 235 236 int amdgpu_job_submit_direct(struct amdgpu_job *job, struct amdgpu_ring *ring, 237 struct dma_fence **fence) 238 { 239 int r; 240 241 job->base.sched = &ring->sched; 242 r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs, job, fence); 243 244 if (r) 245 return r; 246 247 amdgpu_job_free(job); 248 return 0; 249 } 250 251 static struct dma_fence * 252 amdgpu_job_prepare_job(struct drm_sched_job *sched_job, 253 struct drm_sched_entity *s_entity) 254 { 255 struct amdgpu_ring *ring = to_amdgpu_ring(s_entity->rq->sched); 256 struct amdgpu_job *job = to_amdgpu_job(sched_job); 257 struct dma_fence *fence = NULL; 258 int r; 259 260 r = drm_sched_entity_error(s_entity); 261 if (r) 262 goto error; 263 264 if (!fence && job->gang_submit) 265 fence = amdgpu_device_switch_gang(ring->adev, job->gang_submit); 266 267 while (!fence && job->vm && !job->vmid) { 268 r = amdgpu_vmid_grab(job->vm, ring, job, &fence); 269 if (r) { 270 DRM_ERROR("Error getting VM ID (%d)\n", r); 271 goto error; 272 } 273 } 274 275 return fence; 276 277 error: 278 dma_fence_set_error(&job->base.s_fence->finished, r); 279 return NULL; 280 } 281 282 static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job) 283 { 284 struct amdgpu_ring *ring = to_amdgpu_ring(sched_job->sched); 285 struct amdgpu_device *adev = ring->adev; 286 struct dma_fence *fence = NULL, *finished; 287 struct amdgpu_job *job; 288 int r = 0; 289 290 job = to_amdgpu_job(sched_job); 291 finished = &job->base.s_fence->finished; 292 293 trace_amdgpu_sched_run_job(job); 294 295 /* Skip job if VRAM is lost and never resubmit gangs */ 296 if (job->generation != amdgpu_vm_generation(adev, job->vm) || 297 (job->job_run_counter && job->gang_submit)) 298 dma_fence_set_error(finished, -ECANCELED); 299 300 if (finished->error < 0) { 301 dev_dbg(adev->dev, "Skip scheduling IBs in ring(%s)", 302 ring->name); 303 } else { 304 r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs, job, 305 &fence); 306 if (r) 307 dev_err(adev->dev, 308 "Error scheduling IBs (%d) in ring(%s)", r, 309 ring->name); 310 } 311 312 job->job_run_counter++; 313 amdgpu_job_free_resources(job); 314 315 fence = r ? ERR_PTR(r) : fence; 316 return fence; 317 } 318 319 #define to_drm_sched_job(sched_job) \ 320 container_of((sched_job), struct drm_sched_job, queue_node) 321 322 void amdgpu_job_stop_all_jobs_on_sched(struct drm_gpu_scheduler *sched) 323 { 324 struct drm_sched_job *s_job; 325 struct drm_sched_entity *s_entity = NULL; 326 int i; 327 328 /* Signal all jobs not yet scheduled */ 329 for (i = DRM_SCHED_PRIORITY_COUNT - 1; i >= DRM_SCHED_PRIORITY_MIN; i--) { 330 struct drm_sched_rq *rq = &sched->sched_rq[i]; 331 spin_lock(&rq->lock); 332 list_for_each_entry(s_entity, &rq->entities, list) { 333 while ((s_job = to_drm_sched_job(spsc_queue_pop(&s_entity->job_queue)))) { 334 struct drm_sched_fence *s_fence = s_job->s_fence; 335 336 dma_fence_signal(&s_fence->scheduled); 337 dma_fence_set_error(&s_fence->finished, -EHWPOISON); 338 dma_fence_signal(&s_fence->finished); 339 } 340 } 341 spin_unlock(&rq->lock); 342 } 343 344 /* Signal all jobs already scheduled to HW */ 345 list_for_each_entry(s_job, &sched->pending_list, list) { 346 struct drm_sched_fence *s_fence = s_job->s_fence; 347 348 dma_fence_set_error(&s_fence->finished, -EHWPOISON); 349 dma_fence_signal(&s_fence->finished); 350 } 351 } 352 353 const struct drm_sched_backend_ops amdgpu_sched_ops = { 354 .prepare_job = amdgpu_job_prepare_job, 355 .run_job = amdgpu_job_run, 356 .timedout_job = amdgpu_job_timedout, 357 .free_job = amdgpu_job_free_cb 358 }; 359