Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2017 Etnaviv Project
4 */
5
6#include <linux/moduleparam.h>
7
8#include "etnaviv_drv.h"
9#include "etnaviv_dump.h"
10#include "etnaviv_gem.h"
11#include "etnaviv_gpu.h"
12#include "etnaviv_sched.h"
13#include "state.xml.h"
14
15static int etnaviv_job_hang_limit = 0;
16module_param_named(job_hang_limit, etnaviv_job_hang_limit, int , 0444);
17static int etnaviv_hw_jobs_limit = 4;
18module_param_named(hw_job_limit, etnaviv_hw_jobs_limit, int , 0444);
19
20static struct dma_fence *
21etnaviv_sched_dependency(struct drm_sched_job *sched_job,
22 struct drm_sched_entity *entity)
23{
24 struct etnaviv_gem_submit *submit = to_etnaviv_submit(sched_job);
25 struct dma_fence *fence;
26 int i;
27
28 if (unlikely(submit->in_fence)) {
29 fence = submit->in_fence;
30 submit->in_fence = NULL;
31
32 if (!dma_fence_is_signaled(fence))
33 return fence;
34
35 dma_fence_put(fence);
36 }
37
38 for (i = 0; i < submit->nr_bos; i++) {
39 struct etnaviv_gem_submit_bo *bo = &submit->bos[i];
40 int j;
41
42 if (bo->excl) {
43 fence = bo->excl;
44 bo->excl = NULL;
45
46 if (!dma_fence_is_signaled(fence))
47 return fence;
48
49 dma_fence_put(fence);
50 }
51
52 for (j = 0; j < bo->nr_shared; j++) {
53 if (!bo->shared[j])
54 continue;
55
56 fence = bo->shared[j];
57 bo->shared[j] = NULL;
58
59 if (!dma_fence_is_signaled(fence))
60 return fence;
61
62 dma_fence_put(fence);
63 }
64 kfree(bo->shared);
65 bo->nr_shared = 0;
66 bo->shared = NULL;
67 }
68
69 return NULL;
70}
71
72static struct dma_fence *etnaviv_sched_run_job(struct drm_sched_job *sched_job)
73{
74 struct etnaviv_gem_submit *submit = to_etnaviv_submit(sched_job);
75 struct dma_fence *fence = NULL;
76
77 if (likely(!sched_job->s_fence->finished.error))
78 fence = etnaviv_gpu_submit(submit);
79 else
80 dev_dbg(submit->gpu->dev, "skipping bad job\n");
81
82 return fence;
83}
84
85static void etnaviv_sched_timedout_job(struct drm_sched_job *sched_job)
86{
87 struct etnaviv_gem_submit *submit = to_etnaviv_submit(sched_job);
88 struct etnaviv_gpu *gpu = submit->gpu;
89 u32 dma_addr;
90 int change;
91
92 /*
93 * If the GPU managed to complete this jobs fence, the timout is
94 * spurious. Bail out.
95 */
96 if (dma_fence_is_signaled(submit->out_fence))
97 return;
98
99 /*
100 * If the GPU is still making forward progress on the front-end (which
101 * should never loop) we shift out the timeout to give it a chance to
102 * finish the job.
103 */
104 dma_addr = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
105 change = dma_addr - gpu->hangcheck_dma_addr;
106 if (change < 0 || change > 16) {
107 gpu->hangcheck_dma_addr = dma_addr;
108 return;
109 }
110
111 /* block scheduler */
112 drm_sched_stop(&gpu->sched, sched_job);
113
114 if(sched_job)
115 drm_sched_increase_karma(sched_job);
116
117 /* get the GPU back into the init state */
118 etnaviv_core_dump(submit);
119 etnaviv_gpu_recover_hang(gpu);
120
121 drm_sched_resubmit_jobs(&gpu->sched);
122
123 /* restart scheduler after GPU is usable again */
124 drm_sched_start(&gpu->sched, true);
125}
126
127static void etnaviv_sched_free_job(struct drm_sched_job *sched_job)
128{
129 struct etnaviv_gem_submit *submit = to_etnaviv_submit(sched_job);
130
131 drm_sched_job_cleanup(sched_job);
132
133 etnaviv_submit_put(submit);
134}
135
136static const struct drm_sched_backend_ops etnaviv_sched_ops = {
137 .dependency = etnaviv_sched_dependency,
138 .run_job = etnaviv_sched_run_job,
139 .timedout_job = etnaviv_sched_timedout_job,
140 .free_job = etnaviv_sched_free_job,
141};
142
143int etnaviv_sched_push_job(struct drm_sched_entity *sched_entity,
144 struct etnaviv_gem_submit *submit)
145{
146 int ret = 0;
147
148 /*
149 * Hold the fence lock across the whole operation to avoid jobs being
150 * pushed out of order with regard to their sched fence seqnos as
151 * allocated in drm_sched_job_init.
152 */
153 mutex_lock(&submit->gpu->fence_lock);
154
155 ret = drm_sched_job_init(&submit->sched_job, sched_entity,
156 submit->ctx);
157 if (ret)
158 goto out_unlock;
159
160 submit->out_fence = dma_fence_get(&submit->sched_job.s_fence->finished);
161 submit->out_fence_id = idr_alloc_cyclic(&submit->gpu->fence_idr,
162 submit->out_fence, 0,
163 INT_MAX, GFP_KERNEL);
164 if (submit->out_fence_id < 0) {
165 drm_sched_job_cleanup(&submit->sched_job);
166 ret = -ENOMEM;
167 goto out_unlock;
168 }
169
170 /* the scheduler holds on to the job now */
171 kref_get(&submit->refcount);
172
173 drm_sched_entity_push_job(&submit->sched_job, sched_entity);
174
175out_unlock:
176 mutex_unlock(&submit->gpu->fence_lock);
177
178 return ret;
179}
180
181int etnaviv_sched_init(struct etnaviv_gpu *gpu)
182{
183 int ret;
184
185 ret = drm_sched_init(&gpu->sched, &etnaviv_sched_ops,
186 etnaviv_hw_jobs_limit, etnaviv_job_hang_limit,
187 msecs_to_jiffies(500), dev_name(gpu->dev));
188 if (ret)
189 return ret;
190
191 return 0;
192}
193
194void etnaviv_sched_fini(struct etnaviv_gpu *gpu)
195{
196 drm_sched_fini(&gpu->sched);
197}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2017 Etnaviv Project
4 */
5
6#include <linux/moduleparam.h>
7
8#include "etnaviv_drv.h"
9#include "etnaviv_dump.h"
10#include "etnaviv_gem.h"
11#include "etnaviv_gpu.h"
12#include "etnaviv_sched.h"
13#include "state.xml.h"
14
15static int etnaviv_job_hang_limit = 0;
16module_param_named(job_hang_limit, etnaviv_job_hang_limit, int , 0444);
17static int etnaviv_hw_jobs_limit = 4;
18module_param_named(hw_job_limit, etnaviv_hw_jobs_limit, int , 0444);
19
20static struct dma_fence *etnaviv_sched_run_job(struct drm_sched_job *sched_job)
21{
22 struct etnaviv_gem_submit *submit = to_etnaviv_submit(sched_job);
23 struct dma_fence *fence = NULL;
24
25 if (likely(!sched_job->s_fence->finished.error))
26 fence = etnaviv_gpu_submit(submit);
27 else
28 dev_dbg(submit->gpu->dev, "skipping bad job\n");
29
30 return fence;
31}
32
33static enum drm_gpu_sched_stat etnaviv_sched_timedout_job(struct drm_sched_job
34 *sched_job)
35{
36 struct etnaviv_gem_submit *submit = to_etnaviv_submit(sched_job);
37 struct etnaviv_gpu *gpu = submit->gpu;
38 u32 dma_addr;
39 int change;
40
41 /* block scheduler */
42 drm_sched_stop(&gpu->sched, sched_job);
43
44 /*
45 * If the GPU managed to complete this jobs fence, the timout is
46 * spurious. Bail out.
47 */
48 if (dma_fence_is_signaled(submit->out_fence))
49 goto out_no_timeout;
50
51 /*
52 * If the GPU is still making forward progress on the front-end (which
53 * should never loop) we shift out the timeout to give it a chance to
54 * finish the job.
55 */
56 dma_addr = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
57 change = dma_addr - gpu->hangcheck_dma_addr;
58 if (gpu->completed_fence != gpu->hangcheck_fence ||
59 change < 0 || change > 16) {
60 gpu->hangcheck_dma_addr = dma_addr;
61 gpu->hangcheck_fence = gpu->completed_fence;
62 goto out_no_timeout;
63 }
64
65 if(sched_job)
66 drm_sched_increase_karma(sched_job);
67
68 /* get the GPU back into the init state */
69 etnaviv_core_dump(submit);
70 etnaviv_gpu_recover_hang(submit);
71
72 drm_sched_resubmit_jobs(&gpu->sched);
73
74 drm_sched_start(&gpu->sched, true);
75 return DRM_GPU_SCHED_STAT_NOMINAL;
76
77out_no_timeout:
78 /* restart scheduler after GPU is usable again */
79 drm_sched_start(&gpu->sched, true);
80 return DRM_GPU_SCHED_STAT_NOMINAL;
81}
82
83static void etnaviv_sched_free_job(struct drm_sched_job *sched_job)
84{
85 struct etnaviv_gem_submit *submit = to_etnaviv_submit(sched_job);
86
87 drm_sched_job_cleanup(sched_job);
88
89 etnaviv_submit_put(submit);
90}
91
92static const struct drm_sched_backend_ops etnaviv_sched_ops = {
93 .run_job = etnaviv_sched_run_job,
94 .timedout_job = etnaviv_sched_timedout_job,
95 .free_job = etnaviv_sched_free_job,
96};
97
98int etnaviv_sched_push_job(struct etnaviv_gem_submit *submit)
99{
100 int ret = 0;
101
102 /*
103 * Hold the fence lock across the whole operation to avoid jobs being
104 * pushed out of order with regard to their sched fence seqnos as
105 * allocated in drm_sched_job_arm.
106 */
107 mutex_lock(&submit->gpu->fence_lock);
108
109 drm_sched_job_arm(&submit->sched_job);
110
111 submit->out_fence = dma_fence_get(&submit->sched_job.s_fence->finished);
112 submit->out_fence_id = idr_alloc_cyclic(&submit->gpu->fence_idr,
113 submit->out_fence, 0,
114 INT_MAX, GFP_KERNEL);
115 if (submit->out_fence_id < 0) {
116 drm_sched_job_cleanup(&submit->sched_job);
117 ret = -ENOMEM;
118 goto out_unlock;
119 }
120
121 /* the scheduler holds on to the job now */
122 kref_get(&submit->refcount);
123
124 drm_sched_entity_push_job(&submit->sched_job);
125
126out_unlock:
127 mutex_unlock(&submit->gpu->fence_lock);
128
129 return ret;
130}
131
132int etnaviv_sched_init(struct etnaviv_gpu *gpu)
133{
134 int ret;
135
136 ret = drm_sched_init(&gpu->sched, &etnaviv_sched_ops,
137 etnaviv_hw_jobs_limit, etnaviv_job_hang_limit,
138 msecs_to_jiffies(500), NULL, NULL,
139 dev_name(gpu->dev), gpu->dev);
140 if (ret)
141 return ret;
142
143 return 0;
144}
145
146void etnaviv_sched_fini(struct etnaviv_gpu *gpu)
147{
148 drm_sched_fini(&gpu->sched);
149}