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 #include <drm/drmP.h>
28 #include "amdgpu.h"
29 #include "amdgpu_trace.h"
30 
31 static void amdgpu_job_timedout(struct drm_sched_job *s_job)
32 {
33 	struct amdgpu_ring *ring = to_amdgpu_ring(s_job->sched);
34 	struct amdgpu_job *job = to_amdgpu_job(s_job);
35 
36 	DRM_ERROR("ring %s timeout, signaled seq=%u, emitted seq=%u\n",
37 		  job->base.sched->name, atomic_read(&ring->fence_drv.last_seq),
38 		  ring->fence_drv.sync_seq);
39 
40 	amdgpu_device_gpu_recover(ring->adev, job, false);
41 }
42 
43 int amdgpu_job_alloc(struct amdgpu_device *adev, unsigned num_ibs,
44 		     struct amdgpu_job **job, struct amdgpu_vm *vm)
45 {
46 	size_t size = sizeof(struct amdgpu_job);
47 
48 	if (num_ibs == 0)
49 		return -EINVAL;
50 
51 	size += sizeof(struct amdgpu_ib) * num_ibs;
52 
53 	*job = kzalloc(size, GFP_KERNEL);
54 	if (!*job)
55 		return -ENOMEM;
56 
57 	/*
58 	 * Initialize the scheduler to at least some ring so that we always
59 	 * have a pointer to adev.
60 	 */
61 	(*job)->base.sched = &adev->rings[0]->sched;
62 	(*job)->vm = vm;
63 	(*job)->ibs = (void *)&(*job)[1];
64 	(*job)->num_ibs = num_ibs;
65 
66 	amdgpu_sync_create(&(*job)->sync);
67 	amdgpu_sync_create(&(*job)->sched_sync);
68 	(*job)->vram_lost_counter = atomic_read(&adev->vram_lost_counter);
69 
70 	return 0;
71 }
72 
73 int amdgpu_job_alloc_with_ib(struct amdgpu_device *adev, unsigned size,
74 			     struct amdgpu_job **job)
75 {
76 	int r;
77 
78 	r = amdgpu_job_alloc(adev, 1, job, NULL);
79 	if (r)
80 		return r;
81 
82 	r = amdgpu_ib_get(adev, NULL, size, &(*job)->ibs[0]);
83 	if (r)
84 		kfree(*job);
85 	else
86 		(*job)->vm_pd_addr = adev->gart.table_addr;
87 
88 	return r;
89 }
90 
91 void amdgpu_job_free_resources(struct amdgpu_job *job)
92 {
93 	struct amdgpu_ring *ring = to_amdgpu_ring(job->base.sched);
94 	struct dma_fence *f;
95 	unsigned i;
96 
97 	/* use sched fence if available */
98 	f = job->base.s_fence ? &job->base.s_fence->finished : job->fence;
99 
100 	for (i = 0; i < job->num_ibs; ++i)
101 		amdgpu_ib_free(ring->adev, &job->ibs[i], f);
102 }
103 
104 static void amdgpu_job_free_cb(struct drm_sched_job *s_job)
105 {
106 	struct amdgpu_ring *ring = to_amdgpu_ring(s_job->sched);
107 	struct amdgpu_job *job = to_amdgpu_job(s_job);
108 
109 	amdgpu_ring_priority_put(ring, s_job->s_priority);
110 	dma_fence_put(job->fence);
111 	amdgpu_sync_free(&job->sync);
112 	amdgpu_sync_free(&job->sched_sync);
113 	kfree(job);
114 }
115 
116 void amdgpu_job_free(struct amdgpu_job *job)
117 {
118 	amdgpu_job_free_resources(job);
119 
120 	dma_fence_put(job->fence);
121 	amdgpu_sync_free(&job->sync);
122 	amdgpu_sync_free(&job->sched_sync);
123 	kfree(job);
124 }
125 
126 int amdgpu_job_submit(struct amdgpu_job *job, struct drm_sched_entity *entity,
127 		      void *owner, struct dma_fence **f)
128 {
129 	enum drm_sched_priority priority;
130 	struct amdgpu_ring *ring;
131 	int r;
132 
133 	if (!f)
134 		return -EINVAL;
135 
136 	r = drm_sched_job_init(&job->base, entity, owner);
137 	if (r)
138 		return r;
139 
140 	job->owner = owner;
141 	*f = dma_fence_get(&job->base.s_fence->finished);
142 	amdgpu_job_free_resources(job);
143 	priority = job->base.s_priority;
144 	drm_sched_entity_push_job(&job->base, entity);
145 
146 	ring = to_amdgpu_ring(entity->rq->sched);
147 	amdgpu_ring_priority_get(ring, priority);
148 
149 	return 0;
150 }
151 
152 int amdgpu_job_submit_direct(struct amdgpu_job *job, struct amdgpu_ring *ring,
153 			     struct dma_fence **fence)
154 {
155 	int r;
156 
157 	job->base.sched = &ring->sched;
158 	r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs, NULL, fence);
159 	job->fence = dma_fence_get(*fence);
160 	if (r)
161 		return r;
162 
163 	amdgpu_job_free(job);
164 	return 0;
165 }
166 
167 static struct dma_fence *amdgpu_job_dependency(struct drm_sched_job *sched_job,
168 					       struct drm_sched_entity *s_entity)
169 {
170 	struct amdgpu_ring *ring = to_amdgpu_ring(s_entity->rq->sched);
171 	struct amdgpu_job *job = to_amdgpu_job(sched_job);
172 	struct amdgpu_vm *vm = job->vm;
173 	struct dma_fence *fence;
174 	bool explicit = false;
175 	int r;
176 
177 	fence = amdgpu_sync_get_fence(&job->sync, &explicit);
178 	if (fence && explicit) {
179 		if (drm_sched_dependency_optimized(fence, s_entity)) {
180 			r = amdgpu_sync_fence(ring->adev, &job->sched_sync,
181 					      fence, false);
182 			if (r)
183 				DRM_ERROR("Error adding fence (%d)\n", r);
184 		}
185 	}
186 
187 	while (fence == NULL && vm && !job->vmid) {
188 		r = amdgpu_vmid_grab(vm, ring, &job->sync,
189 				     &job->base.s_fence->finished,
190 				     job);
191 		if (r)
192 			DRM_ERROR("Error getting VM ID (%d)\n", r);
193 
194 		fence = amdgpu_sync_get_fence(&job->sync, NULL);
195 	}
196 
197 	return fence;
198 }
199 
200 static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job)
201 {
202 	struct amdgpu_ring *ring = to_amdgpu_ring(sched_job->sched);
203 	struct dma_fence *fence = NULL, *finished;
204 	struct amdgpu_job *job;
205 	int r;
206 
207 	job = to_amdgpu_job(sched_job);
208 	finished = &job->base.s_fence->finished;
209 
210 	BUG_ON(amdgpu_sync_peek_fence(&job->sync, NULL));
211 
212 	trace_amdgpu_sched_run_job(job);
213 
214 	if (job->vram_lost_counter != atomic_read(&ring->adev->vram_lost_counter))
215 		dma_fence_set_error(finished, -ECANCELED);/* skip IB as well if VRAM lost */
216 
217 	if (finished->error < 0) {
218 		DRM_INFO("Skip scheduling IBs!\n");
219 	} else {
220 		r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs, job,
221 				       &fence);
222 		if (r)
223 			DRM_ERROR("Error scheduling IBs (%d)\n", r);
224 	}
225 	/* if gpu reset, hw fence will be replaced here */
226 	dma_fence_put(job->fence);
227 	job->fence = dma_fence_get(fence);
228 
229 	amdgpu_job_free_resources(job);
230 	return fence;
231 }
232 
233 const struct drm_sched_backend_ops amdgpu_sched_ops = {
234 	.dependency = amdgpu_job_dependency,
235 	.run_job = amdgpu_job_run,
236 	.timedout_job = amdgpu_job_timedout,
237 	.free_job = amdgpu_job_free_cb
238 };
239