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 
33 static enum drm_gpu_sched_stat amdgpu_job_timedout(struct drm_sched_job *s_job)
34 {
35 	struct amdgpu_ring *ring = to_amdgpu_ring(s_job->sched);
36 	struct amdgpu_job *job = to_amdgpu_job(s_job);
37 	struct amdgpu_task_info ti;
38 	struct amdgpu_device *adev = ring->adev;
39 	int idx;
40 	int r;
41 
42 	if (!drm_dev_enter(adev_to_drm(adev), &idx)) {
43 		DRM_INFO("%s - device unplugged skipping recovery on scheduler:%s",
44 			 __func__, s_job->sched->name);
45 
46 		/* Effectively the job is aborted as the device is gone */
47 		return DRM_GPU_SCHED_STAT_ENODEV;
48 	}
49 
50 	memset(&ti, 0, sizeof(struct amdgpu_task_info));
51 
52 	if (amdgpu_gpu_recovery &&
53 	    amdgpu_ring_soft_recovery(ring, job->vmid, s_job->s_fence->parent)) {
54 		DRM_ERROR("ring %s timeout, but soft recovered\n",
55 			  s_job->sched->name);
56 		goto exit;
57 	}
58 
59 	amdgpu_vm_get_task_info(ring->adev, job->pasid, &ti);
60 	DRM_ERROR("ring %s timeout, signaled seq=%u, emitted seq=%u\n",
61 		  job->base.sched->name, atomic_read(&ring->fence_drv.last_seq),
62 		  ring->fence_drv.sync_seq);
63 	DRM_ERROR("Process information: process %s pid %d thread %s pid %d\n",
64 		  ti.process_name, ti.tgid, ti.task_name, ti.pid);
65 
66 	if (amdgpu_device_should_recover_gpu(ring->adev)) {
67 		r = amdgpu_device_gpu_recover_imp(ring->adev, job);
68 		if (r)
69 			DRM_ERROR("GPU Recovery Failed: %d\n", r);
70 	} else {
71 		drm_sched_suspend_timeout(&ring->sched);
72 		if (amdgpu_sriov_vf(adev))
73 			adev->virt.tdr_debug = true;
74 	}
75 
76 exit:
77 	drm_dev_exit(idx);
78 	return DRM_GPU_SCHED_STAT_NOMINAL;
79 }
80 
81 int amdgpu_job_alloc(struct amdgpu_device *adev, unsigned num_ibs,
82 		     struct amdgpu_job **job, struct amdgpu_vm *vm)
83 {
84 	if (num_ibs == 0)
85 		return -EINVAL;
86 
87 	*job = kzalloc(struct_size(*job, ibs, num_ibs), GFP_KERNEL);
88 	if (!*job)
89 		return -ENOMEM;
90 
91 	/*
92 	 * Initialize the scheduler to at least some ring so that we always
93 	 * have a pointer to adev.
94 	 */
95 	(*job)->base.sched = &adev->rings[0]->sched;
96 	(*job)->vm = vm;
97 	(*job)->num_ibs = num_ibs;
98 
99 	amdgpu_sync_create(&(*job)->sync);
100 	amdgpu_sync_create(&(*job)->sched_sync);
101 	(*job)->vram_lost_counter = atomic_read(&adev->vram_lost_counter);
102 	(*job)->vm_pd_addr = AMDGPU_BO_INVALID_OFFSET;
103 
104 	return 0;
105 }
106 
107 int amdgpu_job_alloc_with_ib(struct amdgpu_device *adev, unsigned size,
108 		enum amdgpu_ib_pool_type pool_type,
109 		struct amdgpu_job **job)
110 {
111 	int r;
112 
113 	r = amdgpu_job_alloc(adev, 1, job, NULL);
114 	if (r)
115 		return r;
116 
117 	r = amdgpu_ib_get(adev, NULL, size, pool_type, &(*job)->ibs[0]);
118 	if (r)
119 		kfree(*job);
120 
121 	return r;
122 }
123 
124 void amdgpu_job_free_resources(struct amdgpu_job *job)
125 {
126 	struct amdgpu_ring *ring = to_amdgpu_ring(job->base.sched);
127 	struct dma_fence *f;
128 	struct dma_fence *hw_fence;
129 	unsigned i;
130 
131 	if (job->hw_fence.ops == NULL)
132 		hw_fence = job->external_hw_fence;
133 	else
134 		hw_fence = &job->hw_fence;
135 
136 	/* use sched fence if available */
137 	f = job->base.s_fence ? &job->base.s_fence->finished : hw_fence;
138 	for (i = 0; i < job->num_ibs; ++i)
139 		amdgpu_ib_free(ring->adev, &job->ibs[i], f);
140 }
141 
142 static void amdgpu_job_free_cb(struct drm_sched_job *s_job)
143 {
144 	struct amdgpu_job *job = to_amdgpu_job(s_job);
145 
146 	drm_sched_job_cleanup(s_job);
147 
148 	amdgpu_sync_free(&job->sync);
149 	amdgpu_sync_free(&job->sched_sync);
150 
151     /* only put the hw fence if has embedded fence */
152 	if (job->hw_fence.ops != NULL)
153 		dma_fence_put(&job->hw_fence);
154 	else
155 		kfree(job);
156 }
157 
158 void amdgpu_job_free(struct amdgpu_job *job)
159 {
160 	amdgpu_job_free_resources(job);
161 	amdgpu_sync_free(&job->sync);
162 	amdgpu_sync_free(&job->sched_sync);
163 
164 	/* only put the hw fence if has embedded fence */
165 	if (job->hw_fence.ops != NULL)
166 		dma_fence_put(&job->hw_fence);
167 	else
168 		kfree(job);
169 }
170 
171 int amdgpu_job_submit(struct amdgpu_job *job, struct drm_sched_entity *entity,
172 		      void *owner, struct dma_fence **f)
173 {
174 	int r;
175 
176 	if (!f)
177 		return -EINVAL;
178 
179 	r = drm_sched_job_init(&job->base, entity, owner);
180 	if (r)
181 		return r;
182 
183 	drm_sched_job_arm(&job->base);
184 
185 	*f = dma_fence_get(&job->base.s_fence->finished);
186 	amdgpu_job_free_resources(job);
187 	drm_sched_entity_push_job(&job->base);
188 
189 	return 0;
190 }
191 
192 int amdgpu_job_submit_direct(struct amdgpu_job *job, struct amdgpu_ring *ring,
193 			     struct dma_fence **fence)
194 {
195 	int r;
196 
197 	job->base.sched = &ring->sched;
198 	r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs, NULL, fence);
199 	/* record external_hw_fence for direct submit */
200 	job->external_hw_fence = dma_fence_get(*fence);
201 	if (r)
202 		return r;
203 
204 	amdgpu_job_free(job);
205 	dma_fence_put(*fence);
206 
207 	return 0;
208 }
209 
210 static struct dma_fence *amdgpu_job_dependency(struct drm_sched_job *sched_job,
211 					       struct drm_sched_entity *s_entity)
212 {
213 	struct amdgpu_ring *ring = to_amdgpu_ring(s_entity->rq->sched);
214 	struct amdgpu_job *job = to_amdgpu_job(sched_job);
215 	struct amdgpu_vm *vm = job->vm;
216 	struct dma_fence *fence;
217 	int r;
218 
219 	fence = amdgpu_sync_get_fence(&job->sync);
220 	if (fence && drm_sched_dependency_optimized(fence, s_entity)) {
221 		r = amdgpu_sync_fence(&job->sched_sync, fence);
222 		if (r)
223 			DRM_ERROR("Error adding fence (%d)\n", r);
224 	}
225 
226 	while (fence == NULL && vm && !job->vmid) {
227 		r = amdgpu_vmid_grab(vm, ring, &job->sync,
228 				     &job->base.s_fence->finished,
229 				     job);
230 		if (r)
231 			DRM_ERROR("Error getting VM ID (%d)\n", r);
232 
233 		fence = amdgpu_sync_get_fence(&job->sync);
234 	}
235 
236 	return fence;
237 }
238 
239 static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job)
240 {
241 	struct amdgpu_ring *ring = to_amdgpu_ring(sched_job->sched);
242 	struct dma_fence *fence = NULL, *finished;
243 	struct amdgpu_job *job;
244 	int r = 0;
245 
246 	job = to_amdgpu_job(sched_job);
247 	finished = &job->base.s_fence->finished;
248 
249 	BUG_ON(amdgpu_sync_peek_fence(&job->sync, NULL));
250 
251 	trace_amdgpu_sched_run_job(job);
252 
253 	if (job->vram_lost_counter != atomic_read(&ring->adev->vram_lost_counter))
254 		dma_fence_set_error(finished, -ECANCELED);/* skip IB as well if VRAM lost */
255 
256 	if (finished->error < 0) {
257 		DRM_INFO("Skip scheduling IBs!\n");
258 	} else {
259 		r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs, job,
260 				       &fence);
261 		if (r)
262 			DRM_ERROR("Error scheduling IBs (%d)\n", r);
263 	}
264 
265 	if (!job->job_run_counter)
266 		dma_fence_get(fence);
267 	else if (finished->error < 0)
268 		dma_fence_put(&job->hw_fence);
269 	job->job_run_counter++;
270 	amdgpu_job_free_resources(job);
271 
272 	fence = r ? ERR_PTR(r) : fence;
273 	return fence;
274 }
275 
276 #define to_drm_sched_job(sched_job)		\
277 		container_of((sched_job), struct drm_sched_job, queue_node)
278 
279 void amdgpu_job_stop_all_jobs_on_sched(struct drm_gpu_scheduler *sched)
280 {
281 	struct drm_sched_job *s_job;
282 	struct drm_sched_entity *s_entity = NULL;
283 	int i;
284 
285 	/* Signal all jobs not yet scheduled */
286 	for (i = DRM_SCHED_PRIORITY_COUNT - 1; i >= DRM_SCHED_PRIORITY_MIN; i--) {
287 		struct drm_sched_rq *rq = &sched->sched_rq[i];
288 
289 		if (!rq)
290 			continue;
291 
292 		spin_lock(&rq->lock);
293 		list_for_each_entry(s_entity, &rq->entities, list) {
294 			while ((s_job = to_drm_sched_job(spsc_queue_pop(&s_entity->job_queue)))) {
295 				struct drm_sched_fence *s_fence = s_job->s_fence;
296 
297 				dma_fence_signal(&s_fence->scheduled);
298 				dma_fence_set_error(&s_fence->finished, -EHWPOISON);
299 				dma_fence_signal(&s_fence->finished);
300 			}
301 		}
302 		spin_unlock(&rq->lock);
303 	}
304 
305 	/* Signal all jobs already scheduled to HW */
306 	list_for_each_entry(s_job, &sched->pending_list, list) {
307 		struct drm_sched_fence *s_fence = s_job->s_fence;
308 
309 		dma_fence_set_error(&s_fence->finished, -EHWPOISON);
310 		dma_fence_signal(&s_fence->finished);
311 	}
312 }
313 
314 const struct drm_sched_backend_ops amdgpu_sched_ops = {
315 	.dependency = amdgpu_job_dependency,
316 	.run_job = amdgpu_job_run,
317 	.timedout_job = amdgpu_job_timedout,
318 	.free_job = amdgpu_job_free_cb
319 };
320