Skip to content

Commit dad5559

Browse files
kentrussellkentrussell
committed
Revert "drm/sched: Convert the GPU scheduler to variable number ..."
Revert submission 1033322 Reason for revert: Breaks GPU reset Reverted Changes: I518055d36:drm/amdgpu: drop setting buffer funcs in sdma442 I3161b3ba3:drm/amdgpu: fix buffer funcs setting order on susp... If96a48811:drm/amdklc: fix a intree build error. I370135be6:drm/sched: Don't disturb the entity when in RR-mod... I051ca6e0b:drm/sched: Move free worker re-queuing out of the ... I25fb2785c:drm/sched: Rename drm_sched_get_cleanup_job to be ... I666125051:drm/sched: Rename drm_sched_run_job_queue_if_ready... I9a32472a9:drm/sched: Rename drm_sched_free_job_queue to be m... I4c2c9bed1:drm/sched: Re-queue run job worker when drm_sched_... Ie513a85c7:drm/sched: Drain all entities in DRM sched run job... I855c12a4e:drm/sched: Split free_job into own work item I7a116284b:drm/sched: Convert drm scheduler to use a work que... Iffc1aa4cd:drm/amdgpu: move buffer funcs setting up a level I3753e9dd4:drm/sched: Convert the GPU scheduler to variable n... Change-Id: I8fa3facd7111e26a08ff77e6345b4437631e9ac0
1 parent cf8e343 commit dad5559

File tree

10 files changed

+32
-99
lines changed

10 files changed

+32
-99
lines changed

‎drivers/gpu/drm/amd/amdgpu/amdgpu_device.c‎

Lines changed: 0 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -2661,7 +2661,6 @@ static int amdgpu_device_init_schedulers(struct amdgpu_device *adev)
26612661
}
26622662

26632663
r = drm_sched_init(&ring->sched, &amdgpu_sched_ops,
2664-
DRM_SCHED_PRIORITY_COUNT,
26652664
ring->num_hw_submission, 0,
26662665
timeout, adev->reset_domain->wq,
26672666
ring->sched_score, ring->name,

‎drivers/gpu/drm/amd/amdgpu/amdgpu_job.c‎

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -326,8 +326,8 @@ void amdgpu_job_stop_all_jobs_on_sched(struct drm_gpu_scheduler *sched)
326326
int i;
327327

328328
/* Signal all jobs not yet scheduled */
329-
for (i = DRM_SCHED_PRIORITY_KERNEL; i < sched->num_rqs; i++) {
330-
struct drm_sched_rq *rq = sched->sched_rq[i];
329+
for (i = DRM_SCHED_PRIORITY_COUNT - 1; i >= DRM_SCHED_PRIORITY_KERNEL; i--) {
330+
struct drm_sched_rq *rq = &sched->sched_rq[i];
331331
spin_lock(&rq->lock);
332332
list_for_each_entry(s_entity, &rq->entities, list) {
333333
while ((s_job = to_drm_sched_job(spsc_queue_pop(&s_entity->job_queue)))) {

‎drivers/gpu/drm/etnaviv/etnaviv_sched.c‎

Lines changed: 0 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -135,7 +135,6 @@ int etnaviv_sched_init(struct etnaviv_gpu *gpu)
135135
int ret;
136136

137137
ret = drm_sched_init(&gpu->sched, &etnaviv_sched_ops,
138-
DRM_SCHED_PRIORITY_COUNT,
139138
etnaviv_hw_jobs_limit, etnaviv_job_hang_limit,
140139
msecs_to_jiffies(500), NULL, NULL,
141140
dev_name(gpu->dev), gpu->dev);

‎drivers/gpu/drm/lima/lima_sched.c‎

Lines changed: 0 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -489,7 +489,6 @@ int lima_sched_pipe_init(struct lima_sched_pipe *pipe, const char *name)
489489
INIT_WORK(&pipe->recover_work, lima_sched_recover_work);
490490

491491
return drm_sched_init(&pipe->base, &lima_sched_ops,
492-
DRM_SCHED_PRIORITY_COUNT,
493492
1,
494493
lima_job_hang_limit,
495494
msecs_to_jiffies(timeout), NULL,

‎drivers/gpu/drm/msm/msm_ringbuffer.c‎

Lines changed: 0 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -95,7 +95,6 @@ struct msm_ringbuffer *msm_ringbuffer_new(struct msm_gpu *gpu, int id,
9595
sched_timeout = MAX_SCHEDULE_TIMEOUT;
9696

9797
ret = drm_sched_init(&ring->sched, &msm_sched_ops,
98-
DRM_SCHED_PRIORITY_COUNT,
9998
num_hw_submissions, 0, sched_timeout,
10099
NULL, NULL, to_msm_bo(ring->bo)->name, gpu->dev->dev);
101100
if (ret) {

‎drivers/gpu/drm/panfrost/panfrost_job.c‎

Lines changed: 0 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -871,7 +871,6 @@ int panfrost_job_init(struct panfrost_device *pfdev)
871871

872872
ret = drm_sched_init(&js->queue[j].sched,
873873
&panfrost_sched_ops,
874-
DRM_SCHED_PRIORITY_COUNT,
875874
nentries, 0,
876875
msecs_to_jiffies(JOB_TIMEOUT_MS),
877876
pfdev->reset.wq,

‎drivers/gpu/drm/scheduler/sched_entity.c‎

Lines changed: 3 additions & 15 deletions
Original file line numberDiff line numberDiff line change
@@ -75,20 +75,8 @@ int drm_sched_entity_init(struct drm_sched_entity *entity,
7575
RCU_INIT_POINTER(entity->last_scheduled, NULL);
7676
RB_CLEAR_NODE(&entity->rb_tree_node);
7777

78-
if (!sched_list[0]->sched_rq) {
79-
/* Warn drivers not to do this and to fix their DRM
80-
* calling order.
81-
*/
82-
pr_warn("%s: called with uninitialized scheduler\n", __func__);
83-
} else if (num_sched_list) {
84-
/* The "priority" of an entity cannot exceed the number
85-
* of run-queues of a scheduler.
86-
*/
87-
if (entity->priority >= sched_list[0]->num_rqs)
88-
entity->priority = max_t(u32, sched_list[0]->num_rqs,
89-
DRM_SCHED_PRIORITY_KERNEL);
90-
entity->rq = sched_list[0]->sched_rq[entity->priority];
91-
}
78+
if(num_sched_list)
79+
entity->rq = &sched_list[0]->sched_rq[entity->priority];
9280

9381
init_completion(&entity->entity_idle);
9482

@@ -646,7 +634,7 @@ void drm_sched_entity_select_rq(struct drm_sched_entity *entity)
646634

647635
spin_lock(&entity->rq_lock);
648636
sched = drm_sched_pick_best(entity->sched_list, entity->num_sched_list);
649-
rq = sched ? sched->sched_rq[entity->priority] : NULL;
637+
rq = sched ? &sched->sched_rq[entity->priority] : NULL;
650638
if (rq != entity->rq) {
651639
drm_sched_rq_remove_entity(entity->rq, entity);
652640
entity->rq = rq;

‎drivers/gpu/drm/scheduler/sched_main.c‎

Lines changed: 25 additions & 68 deletions
Original file line numberDiff line numberDiff line change
@@ -732,14 +732,8 @@ int drm_sched_job_init(struct drm_sched_job *job,
732732
struct drm_sched_entity *entity,
733733
u32 credits, void *owner)
734734
{
735-
if (!entity->rq) {
736-
/* This will most likely be followed by missing frames
737-
* or worse--a blank screen--leave a trail in the
738-
* logs, so this can be debugged easier.
739-
*/
740-
drm_err(job->sched, "%s: entity has no rq!\n", __func__);
735+
if (!entity->rq)
741736
return -ENOENT;
742-
}
743737

744738
if (unlikely(!credits)) {
745739
pr_err("*ERROR* %s: credits cannot be 0!\n", __func__);
@@ -784,7 +778,7 @@ void drm_sched_job_arm(struct drm_sched_job *job)
784778
sched = entity->rq->sched;
785779

786780
job->sched = sched;
787-
job->s_priority = entity->priority;
781+
job->s_priority = entity->rq - sched->sched_rq;
788782
job->id = atomic64_inc_return(&sched->job_id_count);
789783

790784
drm_sched_fence_init(job->s_fence, job->entity);
@@ -999,11 +993,10 @@ drm_sched_select_entity(struct drm_gpu_scheduler *sched)
999993
int i;
1000994

1001995
/* Kernel run queue has higher priority than normal run queue*/
1002-
for (i = DRM_SCHED_PRIORITY_KERNEL; i < sched->num_rqs; i++) {
1003-
entity = drm_sched_policy == DRM_SCHED_POLICY_FIFO ?
1004-
drm_sched_rq_select_entity_fifo(sched, sched->sched_rq[i]) :
1005-
drm_sched_rq_select_entity_rr(sched, sched->sched_rq[i]);
1006-
996+
for (i = DRM_SCHED_PRIORITY_COUNT - 1; i >= DRM_SCHED_PRIORITY_KERNEL; i--) {
997+
entity = drm_sched_policy == DRM_SCHED_POLICY_FIFO ?
998+
drm_sched_rq_select_entity_fifo(sched, &sched->sched_rq[i]) :
999+
drm_sched_rq_select_entity_rr(sched, &sched->sched_rq[i]);
10071000
if (entity)
10081001
break;
10091002
}
@@ -1183,9 +1176,7 @@ static int drm_sched_main(void *param)
11831176
*
11841177
* @sched: scheduler instance
11851178
* @ops: backend operations for this scheduler
1186-
* @num_rqs: number of runqueues, one for each priority, up to DRM_SCHED_PRIORITY_COUNT
11871179
* @credit_limit: the number of credits this scheduler can hold from all jobs
1188-
* @num_rqs: number of runqueues, one for each priority, up to DRM_SCHED_PRIORITY_COUNT
11891180
* @hang_limit: number of times to allow a job to hang before dropping it
11901181
* @timeout: timeout value in jiffies for the scheduler
11911182
* @timeout_wq: workqueue to use for timeout work. If NULL, the system_wq is
@@ -1198,12 +1189,11 @@ static int drm_sched_main(void *param)
11981189
*/
11991190
int drm_sched_init(struct drm_gpu_scheduler *sched,
12001191
const struct drm_sched_backend_ops *ops,
1201-
u32 num_rqs, u32 credit_limit, unsigned int hang_limit,
1192+
u32 credit_limit, unsigned int hang_limit,
12021193
long timeout, struct workqueue_struct *timeout_wq,
12031194
atomic_t *score, const char *name, struct device *dev)
12041195
{
12051196
int i, ret;
1206-
12071197
sched->ops = ops;
12081198
sched->credit_limit = credit_limit;
12091199
sched->name = name;
@@ -1213,35 +1203,8 @@ int drm_sched_init(struct drm_gpu_scheduler *sched,
12131203
sched->score = score ? score : &sched->_score;
12141204
sched->dev = dev;
12151205

1216-
if (num_rqs > DRM_SCHED_PRIORITY_COUNT) {
1217-
/* This is a gross violation--tell drivers what the problem is.
1218-
*/
1219-
drm_err(sched, "%s: num_rqs cannot be greater than DRM_SCHED_PRIORITY_COUNT\n",
1220-
__func__);
1221-
return -EINVAL;
1222-
} else if (sched->sched_rq) {
1223-
/* Not an error, but warn anyway so drivers can
1224-
* fine-tune their DRM calling order, and return all
1225-
* is good.
1226-
*/
1227-
drm_warn(sched, "%s: scheduler already initialized!\n", __func__);
1228-
return 0;
1229-
}
1230-
1231-
sched->sched_rq = kmalloc_array(num_rqs, sizeof(*sched->sched_rq),
1232-
GFP_KERNEL | __GFP_ZERO);
1233-
if (!sched->sched_rq) {
1234-
drm_err(sched, "%s: out of memory for sched_rq\n", __func__);
1235-
return -ENOMEM;
1236-
}
1237-
sched->num_rqs = num_rqs;
1238-
ret = -ENOMEM;
1239-
for (i = DRM_SCHED_PRIORITY_KERNEL; i < sched->num_rqs; i++) {
1240-
sched->sched_rq[i] = kzalloc(sizeof(*sched->sched_rq[i]), GFP_KERNEL);
1241-
if (!sched->sched_rq[i])
1242-
goto Out_unroll;
1243-
drm_sched_rq_init(sched, sched->sched_rq[i]);
1244-
}
1206+
for (i = DRM_SCHED_PRIORITY_KERNEL; i < DRM_SCHED_PRIORITY_COUNT; i++)
1207+
drm_sched_rq_init(sched, &sched->sched_rq[i]);
12451208

12461209
init_waitqueue_head(&sched->wake_up_worker);
12471210
init_waitqueue_head(&sched->job_scheduled);
@@ -1252,24 +1215,18 @@ int drm_sched_init(struct drm_gpu_scheduler *sched,
12521215
atomic_set(&sched->_score, 0);
12531216
atomic64_set(&sched->job_id_count, 0);
12541217

1255-
/* Each scheduler will run on a seperate kernel thread */
1256-
sched->thread = kthread_run(drm_sched_main, sched, sched->name);
1257-
if (IS_ERR(sched->thread)) {
1258-
ret = PTR_ERR(sched->thread);
1259-
sched->thread = NULL;
1260-
DRM_DEV_ERROR(sched->dev, "Failed to create scheduler for %s.\n", name);
1261-
goto Out_unroll;
1262-
}
1218+
1219+
/* Each scheduler will run on a seperate kernel thread */
1220+
sched->thread = kthread_run(drm_sched_main, sched, sched->name);
1221+
if (IS_ERR(sched->thread)) {
1222+
ret = PTR_ERR(sched->thread);
1223+
sched->thread = NULL;
1224+
DRM_DEV_ERROR(sched->dev, "Failed to create scheduler for %s.\n", name);
1225+
return ret;
1226+
}
12631227

12641228
sched->ready = true;
12651229
return 0;
1266-
Out_unroll:
1267-
for (--i ; i >= DRM_SCHED_PRIORITY_KERNEL; i--)
1268-
kfree(sched->sched_rq[i]);
1269-
kfree(sched->sched_rq);
1270-
sched->sched_rq = NULL;
1271-
drm_err(sched, "%s: Failed to setup GPU scheduler--out of memory\n", __func__);
1272-
return ret;
12731230
}
12741231
EXPORT_SYMBOL(drm_sched_init);
12751232

@@ -1288,8 +1245,8 @@ void drm_sched_fini(struct drm_gpu_scheduler *sched)
12881245
if (sched->thread)
12891246
kthread_stop(sched->thread);
12901247

1291-
for (i = DRM_SCHED_PRIORITY_KERNEL; i < sched->num_rqs; i++) {
1292-
struct drm_sched_rq *rq = sched->sched_rq[i];
1248+
for (i = DRM_SCHED_PRIORITY_COUNT - 1; i >= DRM_SCHED_PRIORITY_KERNEL; i--) {
1249+
struct drm_sched_rq *rq = &sched->sched_rq[i];
12931250

12941251
spin_lock(&rq->lock);
12951252
list_for_each_entry(s_entity, &rq->entities, list)
@@ -1300,7 +1257,7 @@ void drm_sched_fini(struct drm_gpu_scheduler *sched)
13001257
*/
13011258
s_entity->stopped = true;
13021259
spin_unlock(&rq->lock);
1303-
kfree(sched->sched_rq[i]);
1260+
13041261
}
13051262

13061263
/* Wakeup everyone stuck in drm_sched_entity_flush for this scheduler */
@@ -1310,8 +1267,6 @@ void drm_sched_fini(struct drm_gpu_scheduler *sched)
13101267
cancel_delayed_work_sync(&sched->work_tdr);
13111268

13121269
sched->ready = false;
1313-
kfree(sched->sched_rq);
1314-
sched->sched_rq = NULL;
13151270
}
13161271
EXPORT_SYMBOL(drm_sched_fini);
13171272

@@ -1337,8 +1292,10 @@ void drm_sched_increase_karma(struct drm_sched_job *bad)
13371292
*/
13381293
if (bad->s_priority != DRM_SCHED_PRIORITY_KERNEL) {
13391294
atomic_inc(&bad->karma);
1340-
for (i = DRM_SCHED_PRIORITY_HIGH; i < sched->num_rqs; i++) {
1341-
struct drm_sched_rq *rq = sched->sched_rq[i];
1295+
1296+
for (i = DRM_SCHED_PRIORITY_KERNEL; i < DRM_SCHED_PRIORITY_COUNT;
1297+
i++) {
1298+
struct drm_sched_rq *rq = &sched->sched_rq[i];
13421299

13431300
spin_lock(&rq->lock);
13441301
list_for_each_entry_safe(entity, tmp, &rq->entities, list) {

‎drivers/gpu/drm/v3d/v3d_sched.c‎

Lines changed: 0 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -733,7 +733,6 @@ v3d_sched_init(struct v3d_dev *v3d)
733733

734734
ret = drm_sched_init(&v3d->queue[V3D_BIN].sched,
735735
&v3d_bin_sched_ops,
736-
DRM_SCHED_PRIORITY_COUNT,
737736
hw_jobs_limit, job_hang_limit,
738737
msecs_to_jiffies(hang_limit_ms), NULL,
739738
NULL, "v3d_bin", v3d->drm.dev);
@@ -742,7 +741,6 @@ v3d_sched_init(struct v3d_dev *v3d)
742741

743742
ret = drm_sched_init(&v3d->queue[V3D_RENDER].sched,
744743
&v3d_render_sched_ops,
745-
DRM_SCHED_PRIORITY_COUNT,
746744
hw_jobs_limit, job_hang_limit,
747745
msecs_to_jiffies(hang_limit_ms), NULL,
748746
NULL, "v3d_render", v3d->drm.dev);
@@ -751,7 +749,6 @@ v3d_sched_init(struct v3d_dev *v3d)
751749

752750
ret = drm_sched_init(&v3d->queue[V3D_TFU].sched,
753751
&v3d_tfu_sched_ops,
754-
DRM_SCHED_PRIORITY_COUNT,
755752
hw_jobs_limit, job_hang_limit,
756753
msecs_to_jiffies(hang_limit_ms), NULL,
757754
NULL, "v3d_tfu", v3d->drm.dev);
@@ -761,7 +758,6 @@ v3d_sched_init(struct v3d_dev *v3d)
761758
if (v3d_has_csd(v3d)) {
762759
ret = drm_sched_init(&v3d->queue[V3D_CSD].sched,
763760
&v3d_csd_sched_ops,
764-
DRM_SCHED_PRIORITY_COUNT,
765761
hw_jobs_limit, job_hang_limit,
766762
msecs_to_jiffies(hang_limit_ms), NULL,
767763
NULL, "v3d_csd", v3d->drm.dev);
@@ -770,7 +766,6 @@ v3d_sched_init(struct v3d_dev *v3d)
770766

771767
ret = drm_sched_init(&v3d->queue[V3D_CACHE_CLEAN].sched,
772768
&v3d_cache_clean_sched_ops,
773-
DRM_SCHED_PRIORITY_COUNT,
774769
hw_jobs_limit, job_hang_limit,
775770
msecs_to_jiffies(hang_limit_ms), NULL,
776771
NULL, "v3d_cache_clean", v3d->drm.dev);

‎include/drm/gpu_scheduler.h‎

Lines changed: 2 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -490,7 +490,6 @@ struct drm_sched_backend_ops {
490490
* @credit_count: the current credit count of this scheduler
491491
* @timeout: the time after which a job is removed from the scheduler.
492492
* @name: name of the ring for which this scheduler is being used.
493-
* @num_rqs: Number of run-queues. This is at most DRM_SCHED_PRIORITY_COUNT,
494493
* as there's usually one run-queue per priority, but could be less.
495494
* @sched_rq: An allocated array of run-queues of size @num_rqs;
496495
* @wake_up_worker: the wait queue on which the scheduler sleeps until a job
@@ -521,8 +520,7 @@ struct drm_gpu_scheduler {
521520
atomic_t credit_count;
522521
long timeout;
523522
const char *name;
524-
u32 num_rqs;
525-
struct drm_sched_rq **sched_rq;
523+
struct drm_sched_rq sched_rq[DRM_SCHED_PRIORITY_COUNT];
526524
wait_queue_head_t wake_up_worker;
527525
wait_queue_head_t job_scheduled;
528526
atomic64_t job_id_count;
@@ -541,7 +539,7 @@ struct drm_gpu_scheduler {
541539

542540
int drm_sched_init(struct drm_gpu_scheduler *sched,
543541
const struct drm_sched_backend_ops *ops,
544-
u32 num_rqs, u32 credit_limit, unsigned int hang_limit,
542+
u32 credit_limit, unsigned int hang_limit,
545543
long timeout, struct workqueue_struct *timeout_wq,
546544
atomic_t *score, const char *name, struct device *dev);
547545

0 commit comments

Comments
 (0)