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
amdgpu_job_timedout(struct drm_sched_job * s_job)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
amdgpu_job_alloc(struct amdgpu_device * adev,struct amdgpu_vm * vm,struct drm_sched_entity * entity,void * owner,unsigned int num_ibs,struct amdgpu_job ** job)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
amdgpu_job_alloc_with_ib(struct amdgpu_device * adev,struct drm_sched_entity * entity,void * owner,size_t size,enum amdgpu_ib_pool_type pool_type,struct amdgpu_job ** job)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
amdgpu_job_set_resources(struct amdgpu_job * job,struct amdgpu_bo * gds,struct amdgpu_bo * gws,struct amdgpu_bo * oa)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
amdgpu_job_free_resources(struct amdgpu_job * job)160 void amdgpu_job_free_resources(struct amdgpu_job *job)
161 {
162 struct amdgpu_ring *ring = to_amdgpu_ring(job->base.sched);
163 struct dma_fence *f;
164 unsigned i;
165
166 /* Check if any fences where initialized */
167 if (job->base.s_fence && job->base.s_fence->finished.ops)
168 f = &job->base.s_fence->finished;
169 else if (job->hw_fence.ops)
170 f = &job->hw_fence;
171 else
172 f = NULL;
173
174 for (i = 0; i < job->num_ibs; ++i)
175 amdgpu_ib_free(ring->adev, &job->ibs[i], f);
176 }
177
amdgpu_job_free_cb(struct drm_sched_job * s_job)178 static void amdgpu_job_free_cb(struct drm_sched_job *s_job)
179 {
180 struct amdgpu_job *job = to_amdgpu_job(s_job);
181
182 drm_sched_job_cleanup(s_job);
183
184 amdgpu_sync_free(&job->explicit_sync);
185
186 /* only put the hw fence if has embedded fence */
187 if (!job->hw_fence.ops)
188 kfree(job);
189 else
190 dma_fence_put(&job->hw_fence);
191 }
192
amdgpu_job_set_gang_leader(struct amdgpu_job * job,struct amdgpu_job * leader)193 void amdgpu_job_set_gang_leader(struct amdgpu_job *job,
194 struct amdgpu_job *leader)
195 {
196 struct dma_fence *fence = &leader->base.s_fence->scheduled;
197
198 WARN_ON(job->gang_submit);
199
200 /*
201 * Don't add a reference when we are the gang leader to avoid circle
202 * dependency.
203 */
204 if (job != leader)
205 dma_fence_get(fence);
206 job->gang_submit = fence;
207 }
208
amdgpu_job_free(struct amdgpu_job * job)209 void amdgpu_job_free(struct amdgpu_job *job)
210 {
211 if (job->base.entity)
212 drm_sched_job_cleanup(&job->base);
213
214 amdgpu_job_free_resources(job);
215 amdgpu_sync_free(&job->explicit_sync);
216 if (job->gang_submit != &job->base.s_fence->scheduled)
217 dma_fence_put(job->gang_submit);
218
219 if (!job->hw_fence.ops)
220 kfree(job);
221 else
222 dma_fence_put(&job->hw_fence);
223 }
224
amdgpu_job_submit(struct amdgpu_job * job)225 struct dma_fence *amdgpu_job_submit(struct amdgpu_job *job)
226 {
227 struct dma_fence *f;
228
229 drm_sched_job_arm(&job->base);
230 f = dma_fence_get(&job->base.s_fence->finished);
231 amdgpu_job_free_resources(job);
232 drm_sched_entity_push_job(&job->base);
233
234 return f;
235 }
236
amdgpu_job_submit_direct(struct amdgpu_job * job,struct amdgpu_ring * ring,struct dma_fence ** fence)237 int amdgpu_job_submit_direct(struct amdgpu_job *job, struct amdgpu_ring *ring,
238 struct dma_fence **fence)
239 {
240 int r;
241
242 job->base.sched = &ring->sched;
243 r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs, job, fence);
244
245 if (r)
246 return r;
247
248 amdgpu_job_free(job);
249 return 0;
250 }
251
252 static struct dma_fence *
amdgpu_job_prepare_job(struct drm_sched_job * sched_job,struct drm_sched_entity * s_entity)253 amdgpu_job_prepare_job(struct drm_sched_job *sched_job,
254 struct drm_sched_entity *s_entity)
255 {
256 struct amdgpu_ring *ring = to_amdgpu_ring(s_entity->rq->sched);
257 struct amdgpu_job *job = to_amdgpu_job(sched_job);
258 struct dma_fence *fence = NULL;
259 int r;
260
261 r = drm_sched_entity_error(s_entity);
262 if (r)
263 goto error;
264
265 if (!fence && job->gang_submit)
266 fence = amdgpu_device_switch_gang(ring->adev, job->gang_submit);
267
268 while (!fence && job->vm && !job->vmid) {
269 r = amdgpu_vmid_grab(job->vm, ring, job, &fence);
270 if (r) {
271 DRM_ERROR("Error getting VM ID (%d)\n", r);
272 goto error;
273 }
274 }
275
276 return fence;
277
278 error:
279 dma_fence_set_error(&job->base.s_fence->finished, r);
280 return NULL;
281 }
282
amdgpu_job_run(struct drm_sched_job * sched_job)283 static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job)
284 {
285 struct amdgpu_ring *ring = to_amdgpu_ring(sched_job->sched);
286 struct amdgpu_device *adev = ring->adev;
287 struct dma_fence *fence = NULL, *finished;
288 struct amdgpu_job *job;
289 int r = 0;
290
291 job = to_amdgpu_job(sched_job);
292 finished = &job->base.s_fence->finished;
293
294 trace_amdgpu_sched_run_job(job);
295
296 /* Skip job if VRAM is lost and never resubmit gangs */
297 if (job->generation != amdgpu_vm_generation(adev, job->vm) ||
298 (job->job_run_counter && job->gang_submit))
299 dma_fence_set_error(finished, -ECANCELED);
300
301 if (finished->error < 0) {
302 dev_dbg(adev->dev, "Skip scheduling IBs in ring(%s)",
303 ring->name);
304 } else {
305 r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs, job,
306 &fence);
307 if (r)
308 dev_err(adev->dev,
309 "Error scheduling IBs (%d) in ring(%s)", r,
310 ring->name);
311 }
312
313 job->job_run_counter++;
314 amdgpu_job_free_resources(job);
315
316 fence = r ? ERR_PTR(r) : fence;
317 return fence;
318 }
319
320 #define to_drm_sched_job(sched_job) \
321 container_of((sched_job), struct drm_sched_job, queue_node)
322
amdgpu_job_stop_all_jobs_on_sched(struct drm_gpu_scheduler * sched)323 void amdgpu_job_stop_all_jobs_on_sched(struct drm_gpu_scheduler *sched)
324 {
325 struct drm_sched_job *s_job;
326 struct drm_sched_entity *s_entity = NULL;
327 int i;
328
329 /* Signal all jobs not yet scheduled */
330 for (i = DRM_SCHED_PRIORITY_COUNT - 1; i >= DRM_SCHED_PRIORITY_MIN; i--) {
331 struct drm_sched_rq *rq = &sched->sched_rq[i];
332 spin_lock(&rq->lock);
333 list_for_each_entry(s_entity, &rq->entities, list) {
334 while ((s_job = to_drm_sched_job(spsc_queue_pop(&s_entity->job_queue)))) {
335 struct drm_sched_fence *s_fence = s_job->s_fence;
336
337 dma_fence_signal(&s_fence->scheduled);
338 dma_fence_set_error(&s_fence->finished, -EHWPOISON);
339 dma_fence_signal(&s_fence->finished);
340 }
341 }
342 spin_unlock(&rq->lock);
343 }
344
345 /* Signal all jobs already scheduled to HW */
346 list_for_each_entry(s_job, &sched->pending_list, list) {
347 struct drm_sched_fence *s_fence = s_job->s_fence;
348
349 dma_fence_set_error(&s_fence->finished, -EHWPOISON);
350 dma_fence_signal(&s_fence->finished);
351 }
352 }
353
354 const struct drm_sched_backend_ops amdgpu_sched_ops = {
355 .prepare_job = amdgpu_job_prepare_job,
356 .run_job = amdgpu_job_run,
357 .timedout_job = amdgpu_job_timedout,
358 .free_job = amdgpu_job_free_cb
359 };
360