drm/amdgpu: Avoid accessing job->entity after the job is scheduled.
Subsystems (2)
PATCH 1/3 drm/amdgpu: Avoid accessing job->entity after the job is scheduled.
Bug: amdgpu_job_free_cb was accessing s_job->s_entity when the allocated amdgpu_ctx (and the entity inside it) were already deallocated from amdgpu_cs_parser_fini.
Fix: Save job's priority on it's creation instead of accessing it from s_entity later on.
--- drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 3 +-- drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 5 ++--- drivers/gpu/drm/amd/scheduler/gpu_scheduler.c | 1 + drivers/gpu/drm/amd/scheduler/gpu_scheduler.h | 32 ++++++++++++--------------- 4 files changed, 18 insertions(+), 23 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +1 -2
@@ -1192,8 +1192,7 @@ static int amdgpu_cs_submit(struct amdgpu_cs_parser *p, job->uf_sequence = seq;amdgpu_job_free_resources (job); -amdgpu_ring_priority_get (job->ring, -amd_sched_get_job_priority (&job->base)); +amdgpu_ring_priority_get (job->ring, job->base.s_priority);trace_amdgpu_cs_ioctl (job);amd_sched_entity_push_job (&job->base);
drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +2 -3
@@ -104,7 +104,7 @@ static void amdgpu_job_free_cb(struct amd_sched_job *s_job) { struct amdgpu_job *job = container_of(s_job, struct amdgpu_job, base); -amdgpu_ring_priority_put (job->ring,amd_sched_get_job_priority (s_job)); +amdgpu_ring_priority_put (job->ring, s_job->s_priority);dma_fence_put (job->fence);amdgpu_sync_free (&job->sync);amdgpu_sync_free (&job->dep_sync);
@@ -141,8 +141,7 @@ int amdgpu_job_submit(struct amdgpu_job *job, struct amdgpu_ring *ring, job->fence_ctx = entity->fence_context; *f =dma_fence_get (&job->base.s_fence->finished);amdgpu_job_free_resources (job); -amdgpu_ring_priority_get (job->ring, -amd_sched_get_job_priority (&job->base)); +amdgpu_ring_priority_get (job->ring, job->base.s_priority);amd_sched_entity_push_job (&job->base);return 0 ;
drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +1 -0
@@ -529,6 +529,7 @@ int amd_sched_job_init(struct amd_sched_job *job, { job->sched = sched; job->s_entity = entity; + job->s_priority = entity->rq - sched->sched_rq; job->s_fence =amd_sched_fence_create (entity, owner);if (!job->s_fence)return -ENOMEM ;
drivers/gpu/drm/amd/scheduler/gpu_scheduler.h +14 -18
@@ -30,6 +30,19 @@struct amd_gpu_scheduler ;struct amd_sched_rq ; +enum amd_sched_priority { +AMD_SCHED_PRIORITY_MIN , +AMD_SCHED_PRIORITY_LOW =AMD_SCHED_PRIORITY_MIN , +AMD_SCHED_PRIORITY_NORMAL , +AMD_SCHED_PRIORITY_HIGH_SW , +AMD_SCHED_PRIORITY_HIGH_HW , +AMD_SCHED_PRIORITY_KERNEL , +AMD_SCHED_PRIORITY_MAX , +AMD_SCHED_PRIORITY_INVALID =-1 , +AMD_SCHED_PRIORITY_UNSET =-2 +}; + +/** * A scheduler entity is a wrapper around a job queue or a group * of other entities. Entities take turns emitting jobs from their
@@ -83,6 +96,7 @@ struct amd_sched_job {struct delayed_work work_tdr;uint64_t id;atomic_t karma; +enum amd_sched_priority s_priority; };extern const struct dma_fence_ops amd_sched_fence_ops_scheduled;
@@ -114,18 +128,6 @@ struct amd_sched_backend_ops {void (*free_job)(struct amd_sched_job *sched_job); }; -enum amd_sched_priority { -AMD_SCHED_PRIORITY_MIN , -AMD_SCHED_PRIORITY_LOW =AMD_SCHED_PRIORITY_MIN , -AMD_SCHED_PRIORITY_NORMAL , -AMD_SCHED_PRIORITY_HIGH_SW , -AMD_SCHED_PRIORITY_HIGH_HW , -AMD_SCHED_PRIORITY_KERNEL , -AMD_SCHED_PRIORITY_MAX , -AMD_SCHED_PRIORITY_INVALID =-1 , -AMD_SCHED_PRIORITY_UNSET =-2 -}; -/** * One scheduler is implemented for each hardware ring */
@@ -176,10 +178,4 @@ bool amd_sched_dependency_optimized(struct dma_fence* fence, struct amd_sched_entity *entity);void amd_sched_job_kickout (struct amd_sched_job *s_job); -static inline enum amd_sched_priority -amd_sched_get_job_priority (struct amd_sched_job *job) -{ -return (job->s_entity->rq - job->sched->sched_rq); -} -#endif
Signature · 6 lines
2.7.4 _______________________________________________ amd-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/amd-gfx
A message that is not in the archive (sent off-list or not yet received): 1508506327-1084-1-git-send-email-Andrey.Grodzovsky-5C7GfCeVMHo@public.gmane.org
PATCH 2/3 drm/amdgpu: Add SPSC queue to scheduler.
It is intended to sabstitute the bounded fifo we are currently using.
--- drivers/gpu/drm/amd/scheduler/spsc_queue.h | 120 +++++++++++++++++++++++++++++ 1 file changed, 120 insertions(+) create mode 100644 drivers/gpu/drm/amd/scheduler/spsc_queue.h
drivers/gpu/drm/amd/scheduler/spsc_queue.h new file +120 -0
@@ -0,0 +1,120 @@ +/* +* Copyright 2017 Advanced Micro Devices, Inc. +* +* Permission is hereby granted, free of charge, to any person obtaining a +* copy of this software and associated documentation files (the "Software"), +* to deal in the Software without restriction, including without limitation +* the rights to use, copy, modify, merge, publish, distribute, sublicense, +* and/or sell copies of the Software, and to permit persons to whom the +* Software is furnished to do so, subject to the following conditions: +* +* The above copyright notice and this permission notice shall be included in +* all copies or substantial portions of the Software. +* +* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL +* THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR +* OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, +* ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR +* OTHER DEALINGS IN THE SOFTWARE. +* +*/ + +#ifndef AMD_SCHEDULER_SPSC_QUEUE_H_ +#define AMD_SCHEDULER_SPSC_QUEUE_H_ + +#include <linux/atomic.h> + +/** SPSC lockless queue */ + +struct spsc_node { + +/* Stores spsc_node* */ +struct spsc_node *next; +}; + +struct spsc_queue { + +struct spsc_node *head; + +/* atomic pointer to struct spsc_node* */ +atomic_long_t tail; + +atomic_t job_count; +}; + +static inline void spsc_queue_init (struct spsc_queue *queue) +{ + queue->head =NULL ; +atomic_long_set (&queue->tail, (long )&queue->head); +atomic_set (&queue->job_count,0 ); +} + +static inline struct spsc_node *spsc_queue_peek (struct spsc_queue *queue) +{ +return queue->head; +} + +static inline int spsc_queue_count (struct spsc_queue *queue) { +return atomic_read (&queue->job_count); +} + +static inline bool spsc_queue_push (struct spsc_queue *queue,struct spsc_node *node) +{ +struct spsc_node **tail; + + node->next =NULL ; + +preempt_disable (); + + tail = (struct spsc_node **)atomic_long_xchg (&queue->tail, (long )&node->next); +WRITE_ONCE (*tail, node); +atomic_inc (&queue->job_count); + +/* +* In case of first element verify new node will be visible to the consumer +* thread when we ping the kernel thread that there is new work to do. +*/ +smp_wmb (); + +preempt_enable (); + +return tail == &queue->head; +} + + +static inline struct spsc_node *spsc_queue_pop (struct spsc_queue *queue) +{ +struct spsc_node *next, *node; + +/* Verify reading from memory and not the cache */ +smp_rmb (); + + node =READ_ONCE (queue->head); + +if (!node) +return NULL ; + + next =READ_ONCE (node->next); +WRITE_ONCE (queue->head, next); + +if (unlikely (!next)) { +/* slowpath for the last element in the queue */ + +if (atomic_long_cmpxchg (&queue->tail, + (long )&node->next,(long ) &queue->head) != (long )&node->next) { +/* Updating tail failed wait for new next to appear */ +do { +smp_rmb (); + }while (unlikely (!(queue->head =READ_ONCE (node->next)))); + } + } + +atomic_dec (&queue->job_count); +return node; +} + + + +#endif /* AMD_SCHEDULER_SPSC_QUEUE_H_ */
Signature · 6 lines
2.7.4 _______________________________________________ amd-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/amd-gfx
A message that is not in the archive (sent off-list or not yet received): 1508506327-1084-2-git-send-email-Andrey.Grodzovsky-5C7GfCeVMHo@public.gmane.org
RE: [PATCH 2/3] drm/amdgpu: Add SPSC queue to scheduler.
If the deadlock issue could be solved I don't see why we give up kfifo and switch to SPSC ......
-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Andrey Grodzovsky
Sent: 2017年10月20日 21:32
To: [email protected]
Cc: Grodzovsky, Andrey <[email protected]>; Koenig, Christian <[email protected]>
Subject: [PATCH 2/3] drm/amdgpu: Add SPSC queue to scheduler.
It is intended to sabstitute the bounded fifo we are currently using.
--- drivers/gpu/drm/amd/scheduler/spsc_queue.h | 120 +++++++++++++++++++++++++++++ 1 file changed, 120 insertions(+) create mode 100644 drivers/gpu/drm/amd/scheduler/spsc_queue.h
drivers/gpu/drm/amd/scheduler/spsc_queue.h new file +120 -0
@@ -0,0 +1,120 @@ +/* +* Copyright 2017 Advanced Micro Devices, Inc. +* +* Permission is hereby granted, free of charge, to any person +obtaining a +* copy of this software and associated documentation files (the +"Software"), +* to deal in the Software without restriction, including without +limitation +* the rights to use, copy, modify, merge, publish, distribute, +sublicense, +* and/or sell copies of the Software, and to permit persons to whom +the +* Software is furnished to do so, subject to the following conditions: +* +* The above copyright notice and this permission notice shall be +included in +* all copies or substantial portions of the Software. +* +* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, +EXPRESS OR +* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF +MERCHANTABILITY, +* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT +SHALL +* THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, +DAMAGES OR +* OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR +OTHERWISE, +* ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE +OR +* OTHER DEALINGS IN THE SOFTWARE. +* +*/ + +#ifndef AMD_SCHEDULER_SPSC_QUEUE_H_ +#define AMD_SCHEDULER_SPSC_QUEUE_H_ + +#include <linux/atomic.h> + +/** SPSC lockless queue */ + +struct spsc_node { + +/* Stores spsc_node* */ +struct spsc_node *next; +}; + +struct spsc_queue { + +struct spsc_node *head; + +/* atomic pointer to struct spsc_node* */ +atomic_long_t tail; + +atomic_t job_count; +}; + +static inline void spsc_queue_init (struct spsc_queue *queue) { + queue->head =NULL ; +atomic_long_set (&queue->tail, (long )&queue->head); +atomic_set (&queue->job_count,0 ); +} + +static inline struct spsc_node *spsc_queue_peek (struct spsc_queue +*queue) { +return queue->head; +} + +static inline int spsc_queue_count (struct spsc_queue *queue) { +return atomic_read (&queue->job_count); } + +static inline bool spsc_queue_push (struct spsc_queue *queue,struct +spsc_node *node) { +struct spsc_node **tail; + + node->next =NULL ; + +preempt_disable (); + + tail = (struct spsc_node **)atomic_long_xchg (&queue->tail, (long )&node->next); +WRITE_ONCE (*tail, node); +atomic_inc (&queue->job_count); + +/* +* In case of first element verify new node will be visible to the consumer +* thread when we ping the kernel thread that there is new work to do. +*/ +smp_wmb (); + +preempt_enable (); + +return tail == &queue->head; +} + + +static inline struct spsc_node *spsc_queue_pop (struct spsc_queue +*queue) { +struct spsc_node *next, *node; + +/* Verify reading from memory and not the cache */ +smp_rmb (); + + node =READ_ONCE (queue->head); + +if (!node) +return NULL ; + + next =READ_ONCE (node->next); +WRITE_ONCE (queue->head, next); + +if (unlikely (!next)) { +/* slowpath for the last element in the queue */ + +if (atomic_long_cmpxchg (&queue->tail, + (long )&node->next,(long ) &queue->head) != (long )&node->next) { +/* Updating tail failed wait for new next to appear */ +do { +smp_rmb (); + }while (unlikely (!(queue->head =READ_ONCE (node->next))));
+ }
+ }
+
+ atomic_dec(&queue->job_count);
+ return node;
+}
+
+
+
+#endif /* AMD_SCHEDULER_SPSC_QUEUE_H_ */
--
2.7.4
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
A message that is not in the archive (sent off-list or not yet received): BLUPR12MB04499B9F742E2CF76139F6BC84460-7LeqcoF/[email protected]
Re: [PATCH 2/3] drm/amdgpu: Add SPSC queue to scheduler.
Quoted 1 line · Monk
If the deadlock issue could be solved I don't see why we give up kfifo and switch to SPSC ......
The deadlock is solved because we don't block anymore waiting for consumer to dequeue items from the queue - which can only be achieved with not bounded container.
Thanks,
Andrey
Quoted 155 lines
-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Andrey Grodzovsky
Sent: 2017年10月20日 21:32
To: [email protected]
Cc: Grodzovsky, Andrey <[email protected]>; Koenig, Christian <[email protected]>
Subject: [PATCH 2/3] drm/amdgpu: Add SPSC queue to scheduler.It is intended to sabstitute the bounded fifo we are currently using.
Signed-off-by Andrey Grodzovsky <[email protected]>--- drivers/gpu/drm/amd/scheduler/spsc_queue.h | 120 +++++++++++++++++++++++++++++ 1 file changed, 120 insertions(+)create mode 100644
drivers/gpu/drm/amd/scheduler/spsc_queue.hdrivers/gpu/drm/amd/scheduler/spsc_queue.h new file +130
-1@@ -0,0 +1,120 @@ +/* +* Copyright 2017 Advanced Micro Devices, Inc. +* +* Permission is hereby granted, free of charge, to any person +obtaining a +* copy of this software and associated documentation files (the +"Software"), +* to deal in the Software without restriction, including without +limitation +* the rights to use, copy, modify, merge, publish, distribute, +sublicense, +* and/or sell copies of the Software, and to permit persons to whom +the +* Software is furnished to do so, subject to the following conditions: +* +* The above copyright notice and this permission notice shall be +included in +* all copies or substantial portions of the Software. +* +* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, +EXPRESS OR +* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF +MERCHANTABILITY, +* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT +SHALL +* THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, +DAMAGES OR +* OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR +OTHERWISE, +* ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE +OR +* OTHER DEALINGS IN THE SOFTWARE. +* +*/ + +#ifndef AMD_SCHEDULER_SPSC_QUEUE_H_ +#define AMD_SCHEDULER_SPSC_QUEUE_H_ + +#include <linux/atomic.h> + +/** SPSC lockless queue */ + +struct spsc_node { + +/* Stores spsc_node* */ +struct spsc_node *next; +}; + +struct spsc_queue { + +struct spsc_node *head; + +/* atomic pointer to struct spsc_node* */ +atomic_long_t tail; + +atomic_t job_count; +}; + +static inline void spsc_queue_init (struct spsc_queue *queue) { + queue->head =NULL ; +atomic_long_set (&queue->tail, (long )&queue->head); +atomic_set (&queue->job_count,0 ); +} + +static inline struct spsc_node *spsc_queue_peek (struct spsc_queue +*queue) { +return queue->head; +} + +static inline int spsc_queue_count (struct spsc_queue *queue) { +return atomic_read (&queue->job_count); } + +static inline bool spsc_queue_push (struct spsc_queue *queue,struct +spsc_node *node) { +struct spsc_node **tail; + + node->next =NULL ; + +preempt_disable (); + + tail = (struct spsc_node **)atomic_long_xchg (&queue->tail, (long )&node->next); +WRITE_ONCE (*tail, node); +atomic_inc (&queue->job_count); + +/* +* In case of first element verify new node will be visible to the consumer +* thread when we ping the kernel thread that there is new work to do. +*/ +smp_wmb (); + +preempt_enable (); + +return tail == &queue->head; +} + + +static inline struct spsc_node *spsc_queue_pop (struct spsc_queue +*queue) { +struct spsc_node *next, *node; + +/* Verify reading from memory and not the cache */ +smp_rmb (); + + node =READ_ONCE (queue->head); + +if (!node) +return NULL ; + + next =READ_ONCE (node->next); +WRITE_ONCE (queue->head, next); + +if (unlikely (!next)) { +/* slowpath for the last element in the queue */ + +if (atomic_long_cmpxchg (&queue->tail, + (long )&node->next,(long ) &queue->head) != (long )&node->next) { +/* Updating tail failed wait for new next to appear */ +do { +smp_rmb (); + }while (unlikely (!(queue->head =READ_ONCE (node->next)))); + } + } + +atomic_dec (&queue->job_count); +return node; +} + + + +#endif /* AMD_SCHEDULER_SPSC_QUEUE_H_ */ --2.7.4
Quoted 4 lines
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Re: [PATCH 2/3] drm/amdgpu: Add SPSC queue to scheduler.
In reply to a message not in the archiveAm 20.10.2017 um 15:32 schrieb Andrey Grodzovsky:
Quoted 4 lines
It is intended to sabstitute the bounded fifo we are currently using.
Signed-off-by Andrey Grodzovsky <[email protected]>
Two style nit picks below, with those fixed the patch is Reviewed-by: Christian König <[email protected]>
Quoted 60 lines
--- drivers/gpu/drm/amd/scheduler/spsc_queue.h | 120 +++++++++++++++++++++++++++++ 1 file changed, 120 insertions(+)create mode 100644
drivers/gpu/drm/amd/scheduler/spsc_queue.hdrivers/gpu/drm/amd/scheduler/spsc_queue.h new file +49
-0@@ -0,0 +1,120 @@ +/* +* Copyright 2017 Advanced Micro Devices, Inc. +* +* Permission is hereby granted, free of charge, to any person obtaining a +* copy of this software and associated documentation files (the "Software"), +* to deal in the Software without restriction, including without limitation +* the rights to use, copy, modify, merge, publish, distribute, sublicense, +* and/or sell copies of the Software, and to permit persons to whom the +* Software is furnished to do so, subject to the following conditions: +* +* The above copyright notice and this permission notice shall be included in +* all copies or substantial portions of the Software. +* +* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL +* THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR +* OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, +* ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR +* OTHER DEALINGS IN THE SOFTWARE. +* +*/ + +#ifndef AMD_SCHEDULER_SPSC_QUEUE_H_ +#define AMD_SCHEDULER_SPSC_QUEUE_H_ + +#include <linux/atomic.h> + +/** SPSC lockless queue */ + +struct spsc_node { + +/* Stores spsc_node* */ +struct spsc_node *next; +}; + +struct spsc_queue { + +struct spsc_node *head; + +/* atomic pointer to struct spsc_node* */ +atomic_long_t tail; + +atomic_t job_count; +}; + +static inline void spsc_queue_init (struct spsc_queue *queue) +{ + queue->head =NULL ;
drivers/gpu/drm/amd/scheduler/spsc_queue.h line 59
Quoted 10 lines
drivers/gpu/drm/amd/scheduler/spsc_queue.h new file +10
-0+atomic_long_set (&queue->tail, (long )&queue->head); +atomic_set (&queue->job_count,0 ); +} + +static inline struct spsc_node *spsc_queue_peek(struct spsc_queue *queue +{ +return queue->head; +} + +static inline int spsc_queue_count (struct spsc_queue *queue) {
The "{" should be on a new line.
Quoted 41 lines
drivers/gpu/drm/amd/scheduler/spsc_queue.h new file +41
-0+return atomic_read (&queue->job_count); +} + +static inline bool spsc_queue_push (struct spsc_queue *queue,struct spsc_node *node) +{ +struct spsc_node **tail; + + node->next =NULL ; + +preempt_disable (); + + tail = (struct spsc_node **)atomic_long_xchg (&queue->tail, (long )&node->next); +WRITE_ONCE (*tail, node); +atomic_inc (&queue->job_count); + +/* +* In case of first element verify new node will be visible to the consumer +* thread when we ping the kernel thread that there is new work to do. +*/ +smp_wmb (); + +preempt_enable (); + +return tail == &queue->head; +} + + +static inline struct spsc_node *spsc_queue_pop (struct spsc_queue *queue) +{ +struct spsc_node *next, *node; + +/* Verify reading from memory and not the cache */ +smp_rmb (); + + node =READ_ONCE (queue->head); + +if (!node) +return NULL ; + + next =READ_ONCE (node->next); +WRITE_ONCE (queue->head, next);
drivers/gpu/drm/amd/scheduler/spsc_queue.h line 110
Quoted 10 lines
drivers/gpu/drm/amd/scheduler/spsc_queue.h new file +10
-0+ +if (unlikely (!next)) { +/* slowpath for the last element in the queue */ + +if (atomic_long_cmpxchg (&queue->tail, + (long )&node->next,(long ) &queue->head) != (long )&node->next) { +/* Updating tail failed wait for new next to appear */ +do { +smp_rmb (); + }while (unlikely (!(queue->head =READ_ONCE (node->next))));
Missing space between } and while.
Regards,
Christian.
drivers/gpu/drm/amd/scheduler/spsc_queue.h line 120
Quoted 10 lines
drivers/gpu/drm/amd/scheduler/spsc_queue.h new file +10
-0+ } + } + +atomic_dec (&queue->job_count); +return node; +} + + + +#endif /* AMD_SCHEDULER_SPSC_QUEUE_H_ */
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
PATCH 3/3 drm/amdgpu: Fix deadlock during GPU reset.
In reply to a message not in the archiveSwitch from kfifo to SPSC queue.
Bug:
Kfifo is limited at size, during GPU reset it would fill up to limit and the pushing thread (producer) would wait for the scheduler worker to consume the items in the fifo while holding reservation lock on a BO. The gpu reset thread on the other hand blocks the scheduler during reset. Before it unblocks the sceduler it might want to recover VRAM and so will try to reserve the same BO the producer thread is already holding creating a deadlock.
Fix:
Switch from kfifo to SPSC queue which is unlimited in size.
--- drivers/gpu/drm/amd/scheduler/gpu_sched_trace.h | 4 +- drivers/gpu/drm/amd/scheduler/gpu_scheduler.c | 51 ++++++++++--------------- drivers/gpu/drm/amd/scheduler/gpu_scheduler.h | 4 +- 3 files changed, 26 insertions(+), 33 deletions(-)
drivers/gpu/drm/amd/scheduler/gpu_sched_trace.h +2 -2
@@ -28,8 +28,8 @@ TRACE_EVENT(amd_sched_job, __entry->id = sched_job->id; __entry->fence = &sched_job->s_fence->finished; __entry->name = sched_job->sched->name; - __entry->job_count =kfifo_len ( - &sched_job->s_entity->job_queue) /sizeof (sched_job); + __entry->job_count =spsc_queue_count ( + &sched_job->s_entity->job_queue); __entry->hw_job_count =atomic_read ( &sched_job->sched->hw_rq_count); ),
drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +21 -30
154 lines not on this page · raw message
drivers/gpu/drm/amd/scheduler/gpu_scheduler.h +3 -1
@@ -26,6 +26,7 @@#include <linux/kfifo.h> #include <linux/dma-fence.h> +#include "spsc_queue.h" struct amd_gpu_scheduler ;struct amd_sched_rq ;
@@ -56,7 +57,7 @@ struct amd_sched_entity {struct amd_gpu_scheduler *sched;spinlock_t queue_lock; -struct kfifo job_queue; +struct spsc_queue job_queue;atomic_t fence_seq;uint64_t fence_context;
@@ -87,6 +88,7 @@ struct amd_sched_fence { };struct amd_sched_job { +struct spsc_node queue_node;struct amd_gpu_scheduler *sched;struct amd_sched_entity *s_entity;struct amd_sched_fence *s_fence;
Signature · 6 lines
2.7.4 _______________________________________________ amd-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/amd-gfx
A message that is not in the archive (sent off-list or not yet received): 1508506327-1084-3-git-send-email-Andrey.Grodzovsky-5C7GfCeVMHo@public.gmane.org
RE: [PATCH 3/3] drm/amdgpu: Fix deadlock during GPU reset.
Why not use a more simple way ?
Like moving ttm_eu_fence_buffer_objects() to before amd_sched_entity_push_job() ? That could solve the deadlock from your description
And the push order is already guaranteed by context->mutex (which is also a patch from you)
BR Monk
-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Andrey Grodzovsky
Sent: 2017年10月20日 21:32
To: [email protected]
Cc: Grodzovsky, Andrey <[email protected]>; Koenig, Christian <[email protected]>
Subject: [PATCH 3/3] drm/amdgpu: Fix deadlock during GPU reset.
Switch from kfifo to SPSC queue.
Bug:
Kfifo is limited at size, during GPU reset it would fill up to limit and the pushing thread (producer) would wait for the scheduler worker to consume the items in the fifo while holding reservation lock on a BO. The gpu reset thread on the other hand blocks the scheduler during reset. Before it unblocks the sceduler it might want to recover VRAM and so will try to reserve the same BO the producer thread is already holding creating a deadlock.
Fix:
Switch from kfifo to SPSC queue which is unlimited in size.
--- drivers/gpu/drm/amd/scheduler/gpu_sched_trace.h | 4 +- drivers/gpu/drm/amd/scheduler/gpu_scheduler.c | 51 ++++++++++--------------- drivers/gpu/drm/amd/scheduler/gpu_scheduler.h | 4 +- 3 files changed, 26 insertions(+), 33 deletions(-)
drivers/gpu/drm/amd/scheduler/gpu_sched_trace.h +2 -2
@@ -28,8 +28,8 @@ TRACE_EVENT(amd_sched_job, __entry->id = sched_job->id; __entry->fence = &sched_job->s_fence->finished; __entry->name = sched_job->sched->name; - __entry->job_count =kfifo_len ( - &sched_job->s_entity->job_queue) /sizeof (sched_job); + __entry->job_count =spsc_queue_count ( + &sched_job->s_entity->job_queue); __entry->hw_job_count =atomic_read ( &sched_job->sched->hw_rq_count); ),
drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +5 -0
@@ -28,9 +28,14 @@#include <drm/drmP.h> #include "gpu_scheduler.h" +#include "spsc_queue.h" +#define CREATE_TRACE_POINTS #include "gpu_sched_trace.h" +#define to_amd_sched_job (sched_job) \ + container_of((sched_job), struct amd_sched_job, queue_node) + static bool amd_sched_entity_is_ready(struct amd_sched_entity *entity); static void amd_sched_wakeup(struct amd_gpu_scheduler *sched); static void amd_sched_process_job(struct dma_fence *f, struct dma_fence_cb *cb); @@ -123,8 +128,6 @@ int amd_sched_entity_init(struct amd_gpu_scheduler *sched,struct amd_sched_rq *rq,uint32_t jobs)
{
- int r;
-
if (!(sched && entity && rq))
return -EINVAL;
@@ -135,9 +138,7 @@ int amd_sched_entity_init(struct amd_gpu_scheduler *sched,
spin_lock_init(&entity->rq_lock);
spin_lock_init(&entity->queue_lock);
- r = kfifo_alloc(&entity->job_queue, jobs * sizeof(void *), GFP_KERNEL);
- if (r)
- return r;
+ spsc_queue_init(&entity->job_queue);
atomic_set(&entity->fence_seq, 0);
entity->fence_context = dma_fence_context_alloc(2); @@ -170,7 +171,7 @@ static bool amd_sched_entity_is_initialized(struct amd_gpu_scheduler *sched, static bool amd_sched_entity_is_idle(struct amd_sched_entity *entity) { rmb();
- if (kfifo_is_empty(&entity->job_queue))
+ if (spsc_queue_peek(&entity->job_queue) == NULL)
return true;
return false;
@@ -185,7 +186,7 @@ static bool amd_sched_entity_is_idle(struct amd_sched_entity *entity)
*/
static bool amd_sched_entity_is_ready(struct amd_sched_entity *entity) {
- if (kfifo_is_empty(&entity->job_queue))
+ if (spsc_queue_peek(&entity->job_queue) == NULL)
return false;
if (ACCESS_ONCE(entity->dependency))
@@ -227,7 +228,7 @@ void amd_sched_entity_fini(struct amd_gpu_scheduler *sched,
*/
kthread_park(sched->thread);
kthread_unpark(sched->thread);
- while (kfifo_out(&entity->job_queue, &job, sizeof(job))) {
+ while ((job = to_amd_sched_job(spsc_queue_pop(&entity->job_queue))))
+{
struct amd_sched_fence *s_fence = job->s_fence;
amd_sched_fence_scheduled(s_fence);
dma_fence_set_error(&s_fence->finished, -ESRCH); @@ -235,9 +236,7 @@ void amd_sched_entity_fini(struct amd_gpu_scheduler *sched, dma_fence_put(&s_fence->finished);
sched->ops->free_job(job);
}
-
}
- kfifo_free(&entity->job_queue);
}
static void amd_sched_entity_wakeup(struct dma_fence *f, struct dma_fence_cb *cb) @@ -332,18 +331,21 @@ static bool amd_sched_entity_add_dependency_cb(struct amd_sched_entity *entity) }
static struct amd_sched_job *
-amd_sched_entity_peek_job(struct amd_sched_entity *entity) +amd_sched_entity_pop_job(struct amd_sched_entity *entity)
{
struct amd_gpu_scheduler *sched = entity->sched;
- struct amd_sched_job *sched_job;
+ struct amd_sched_job *sched_job = to_amd_sched_job(
+ spsc_queue_peek(&entity->job_queue));
- if (!kfifo_out_peek(&entity->job_queue, &sched_job, sizeof(sched_job)))
+ if (!sched_job)
return NULL;
while ((entity->dependency = sched->ops->dependency(sched_job)))
if (amd_sched_entity_add_dependency_cb(entity))
return NULL;
+ sched_job->s_entity = NULL;
+ spsc_queue_pop(&entity->job_queue);
return sched_job;
}
@@ -354,18 +356,14 @@ amd_sched_entity_peek_job(struct amd_sched_entity *entity)
*
* Returns true if we could submit the job.
*/
-static bool amd_sched_entity_in(struct amd_sched_job *sched_job) +static void amd_sched_entity_in(struct amd_sched_job *sched_job)
{
struct amd_gpu_scheduler *sched = sched_job->sched;
struct amd_sched_entity *entity = sched_job->s_entity;
- bool added, first = false;
+ bool first = false;
spin_lock(&entity->queue_lock);
- added = kfifo_in(&entity->job_queue, &sched_job,
- sizeof(sched_job)) == sizeof(sched_job);
-
- if (added && kfifo_len(&entity->job_queue) == sizeof(sched_job))
- first = true;
+ first = spsc_queue_push(&entity->job_queue, &sched_job->queue_node);
spin_unlock(&entity->queue_lock);
@@ -377,7 +375,6 @@ static bool amd_sched_entity_in(struct amd_sched_job *sched_job)
spin_unlock(&entity->rq_lock); amd_sched_wakeup(sched);
}
- return added;
}
/* job_finish is called after hw fence signaled @@ -514,11 +511,8 @@ void amd_sched_job_recovery(struct amd_gpu_scheduler *sched)
*/
void amd_sched_entity_push_job(struct amd_sched_job *sched_job) {
- struct amd_sched_entity *entity = sched_job->s_entity;
-
trace_amd_sched_job(sched_job);
- wait_event(entity->sched->job_scheduled,
- amd_sched_entity_in(sched_job));
+ amd_sched_entity_in(sched_job);
}
/* init a sched_job with basic field */ @@ -611,7 +605,7 @@ static int amd_sched_main(void *param) {
struct sched_param sparam = {.sched_priority = 1};
struct amd_gpu_scheduler *sched = (struct amd_gpu_scheduler *)param;
- int r, count;
+ int r;
sched_setscheduler(current, SCHED_FIFO, &sparam);
@@ -629,7 +623,7 @@ static int amd_sched_main(void *param)
if (!entity)
continue;
- sched_job = amd_sched_entity_peek_job(entity);
+ sched_job = amd_sched_entity_pop_job(entity);
if (!sched_job)
continue;
@@ -656,9 +650,6 @@ static int amd_sched_main(void *param)
amd_sched_process_job(NULL, &s_fence->cb);
}
- count = kfifo_out(&entity->job_queue, &sched_job,
- sizeof(sched_job));
- WARN_ON(count != sizeof(sched_job));
wake_up(&sched->job_scheduled);
}
return 0;
drivers/gpu/drm/amd/scheduler/gpu_scheduler.h +3 -2
@@ -26,6 +26,7 @@#include <linux/kfifo.h> #include <linux/dma-fence.h> +#include "spsc_queue.h" struct amd_gpu_scheduler ;struct amd_sched_rq ;
@@ -56,7 +57,7 @@ struct amd_sched_entity {struct amd_gpu_scheduler *sched;spinlock_t queue_lock; -struct kfifo job_queue; +struct spsc_queue job_queue;atomic_t fence_seq;uint64_t fence_context;
@@ -87,6 +88,7 @@ struct amd_sched_fence { };struct amd_sched_job { +struct spsc_node queue_node;struct amd_gpu_scheduler *sched;struct amd_sched_entity *s_entity;struct amd_sched_fence *s_fence; --
2.7.4
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
A message that is not in the archive (sent off-list or not yet received): BLUPR12MB04491AF9EAD2623B5EA9D7B784460-7LeqcoF/[email protected]
Re: [PATCH 3/3] drm/amdgpu: Fix deadlock during GPU reset.
We discussed that as well, problem is that this won't be sufficient.
We push to the kfifo not only during command submission, but also for VM updates and TTM buffers moves.
So we can still deadlock because of them.
Regards,
Christian.
Am 23.10.2017 um 05:03 schrieb Liu, Monk:
Quoted 237 lines
Why not use a more simple way ?
Like moving
ttm_eu_fence_buffer_objects()to beforeamd_sched_entity_push_job()? That could solve the deadlock from your descriptionAnd the push order is already guaranteed by
context->mutex(which is also a patch from you)BR Monk
-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Andrey Grodzovsky
Sent: 2017年10月20日 21:32
To: [email protected]
Cc: Grodzovsky, Andrey <[email protected]>; Koenig, Christian <[email protected]>
Subject: [PATCH 3/3] drm/amdgpu: Fix deadlock during GPU reset.Switch from kfifo to SPSC queue.
Bug:
Kfifo is limited at size, during GPU reset it would fill up to limit and the pushing thread (producer) would wait for the scheduler worker to consume the items in the fifo while holding reservation lock on a BO. The gpu reset thread on the other hand blocks the scheduler during reset. Before it unblocks the sceduler it might want to recover VRAM and so will try to reserve the same BO the producer thread is already holding creating a deadlock.Fix:
Switch from kfifo to SPSC queue which is unlimited in size.Signed-off-by Andrey Grodzovsky <[email protected]>--- drivers/gpu/drm/amd/scheduler/gpu_sched_trace.h | 4 +- drivers/gpu/drm/amd/scheduler/gpu_scheduler.c | 51 ++++++++++--------------- drivers/gpu/drm/amd/scheduler/gpu_scheduler.h | 4 +- 3 files changed, 26 insertions(+), 33 deletions(-)drivers/gpu/drm/amd/scheduler/gpu_sched_trace.h +2
-2@@ -28,8 +28,8 @@ TRACE_EVENT(amd_sched_job, __entry->id = sched_job->id; __entry->fence = &sched_job->s_fence->finished; __entry->name = sched_job->sched->name; - __entry->job_count =kfifo_len ( - &sched_job->s_entity->job_queue) /sizeof (sched_job); + __entry->job_count =spsc_queue_count ( + &sched_job->s_entity->job_queue); __entry->hw_job_count =atomic_read ( &sched_job->sched->hw_rq_count); ),drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +5
-0@@ -28,9 +28,14 @@#include <drm/drmP.h> #include "gpu_scheduler.h" +#include "spsc_queue.h" +#define CREATE_TRACE_POINTS #include "gpu_sched_trace.h" +#define to_amd_sched_job (sched_job) \ + container_of((sched_job), struct amd_sched_job, queue_node) + static bool amd_sched_entity_is_ready(struct amd_sched_entity *entity); static void amd_sched_wakeup(struct amd_gpu_scheduler *sched); static void amd_sched_process_job(struct dma_fence *f, struct dma_fence_cb *cb); @@ -123,8 +128,6 @@ int amd_sched_entity_init(struct amd_gpu_scheduler *sched,struct amd_sched_rq *rq,uint32_t jobs){drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +1
-5-int r; -if (!(sched && entity && rq))return -EINVAL ;@@ -135,9 +138,7 @@ int amd_sched_entity_init(struct amd_gpu_scheduler *sched,spin_lock_init (&entity->rq_lock);spin_lock_init (&entity->queue_lock); - r =kfifo_alloc (&entity->job_queue, jobs *sizeof (void *),GFP_KERNEL ); -if (r) -return r; +spsc_queue_init (&entity->job_queue);atomic_set (&entity->fence_seq,0 ); entity->fence_context = dma_fence_context_alloc(2); @@ -170,7 +171,7 @@ static bool amd_sched_entity_is_initialized(struct amd_gpu_scheduler *sched, static bool amd_sched_entity_is_idle(struct amd_sched_entity *entity) {rmb();drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +13
-16-if (kfifo_is_empty (&entity->job_queue)) +if (spsc_queue_peek (&entity->job_queue) ==NULL )return true;return false;@@ -185,7 +186,7 @@ static bool amd_sched_entity_is_idle(struct amd_sched_entity *entity)*/ static bool amd_sched_entity_is_ready (struct amd_sched_entity *entity) { -if (kfifo_is_empty (&entity->job_queue)) +if (spsc_queue_peek (&entity->job_queue) ==NULL )return false;if (ACCESS_ONCE (entity->dependency))@@ -227,7 +228,7 @@ void amd_sched_entity_fini(struct amd_gpu_scheduler *sched,*/ kthread_park (sched->thread);kthread_unpark (sched->thread); -while (kfifo_out (&entity->job_queue, &job,sizeof (job))) { +while ((job =to_amd_sched_job (spsc_queue_pop (&entity->job_queue)))) +{struct amd_sched_fence *s_fence = job->s_fence;amd_sched_fence_scheduled (s_fence); dma_fence_set_error(&s_fence->finished, -ESRCH); @@ -235,9 +236,7 @@ void amd_sched_entity_fini(struct amd_gpu_scheduler *sched, dma_fence_put(&s_fence->finished); sched->ops->free_job (job); } - } -kfifo_free (&entity->job_queue); } static void amd_sched_entity_wakeup(struct dma_fence *f, struct dma_fence_cb *cb) @@ -332,18 +331,21 @@ static bool amd_sched_entity_add_dependency_cb(struct amd_sched_entity *entity) }static struct amd_sched_job * -amd_sched_entity_peek_job (struct amd_sched_entity *entity) +amd_sched_entity_pop_job (struct amd_sched_entity *entity) {struct amd_gpu_scheduler *sched = entity->sched; -struct amd_sched_job *sched_job; +struct amd_sched_job *sched_job =to_amd_sched_job ( +spsc_queue_peek (&entity->job_queue)); -if (!kfifo_out_peek (&entity->job_queue, &sched_job,sizeof (sched_job))) +if (!sched_job)return NULL ;while ((entity->dependency = sched->ops->dependency (sched_job)))if (amd_sched_entity_add_dependency_cb (entity))return NULL ; + sched_job->s_entity =NULL ; +spsc_queue_pop (&entity->job_queue);return sched_job; }@@ -354,18 +356,14 @@ amd_sched_entity_peek_job(struct amd_sched_entity *entity)* * Returns true if we could submit the job. */ -static bool amd_sched_entity_in (struct amd_sched_job *sched_job) +static void amd_sched_entity_in (struct amd_sched_job *sched_job) {struct amd_gpu_scheduler *sched = sched_job->sched;struct amd_sched_entity *entity = sched_job->s_entity; -bool added, first = false; +bool first = false;spin_lock (&entity->queue_lock); - added =kfifo_in (&entity->job_queue, &sched_job, -sizeof (sched_job)) ==sizeof (sched_job); - -if (added &&kfifo_len (&entity->job_queue) ==sizeof (sched_job)) - first = true; + first =spsc_queue_push (&entity->job_queue, &sched_job->queue_node);spin_unlock (&entity->queue_lock);@@ -377,7 +375,6 @@ static bool amd_sched_entity_in(struct amd_sched_job *sched_job)spin_unlock (&entity->rq_lock);amd_sched_wakeup (sched); } -return added; }/* job_finish is called after hw fence signaled @@ -514,11 +511,8 @@ void amd_sched_job_recovery(struct amd_gpu_scheduler *sched) */ void amd_sched_entity_push_job(struct amd_sched_job *sched_job) {drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +3
-9-struct amd_sched_entity *entity = sched_job->s_entity; -trace_amd_sched_job (sched_job); -wait_event (entity->sched->job_scheduled, -amd_sched_entity_in (sched_job)); +amd_sched_entity_in (sched_job); } /* init a sched_job with basic field */ @@ -611,7 +605,7 @@ static int amd_sched_main(void *param) {struct sched_param sparam = {.sched_priority =1 };struct amd_gpu_scheduler *sched = (struct amd_gpu_scheduler *)param; -int r, count; +int r;sched_setscheduler (current,SCHED_FIFO , &sparam);@@ -629,7 +623,7 @@ static int amd_sched_main(void *param)if (!entity)continue ; - sched_job =amd_sched_entity_peek_job (entity); + sched_job =amd_sched_entity_pop_job (entity);if (!sched_job)continue ;@@ -656,9 +650,6 @@ static int amd_sched_main(void *param)amd_sched_process_job (NULL , &s_fence->cb); } - count =kfifo_out (&entity->job_queue, &sched_job, -sizeof (sched_job)); -WARN_ON (count !=sizeof (sched_job));wake_up (&sched->job_scheduled); }return 0 ;drivers/gpu/drm/amd/scheduler/gpu_scheduler.h +3
-2@@ -26,6 +26,7 @@#include <linux/kfifo.h> #include <linux/dma-fence.h> +#include "spsc_queue.h" struct amd_gpu_scheduler ;struct amd_sched_rq ;@@ -56,7 +57,7 @@ struct amd_sched_entity {struct amd_gpu_scheduler *sched;spinlock_t queue_lock; -struct kfifo job_queue; +struct spsc_queue job_queue;atomic_t fence_seq;uint64_t fence_context;@@ -87,6 +88,7 @@ struct amd_sched_fence { };struct amd_sched_job { +struct spsc_node queue_node;struct amd_gpu_scheduler *sched;struct amd_sched_entity *s_entity;struct amd_sched_fence *s_fence; --2.7.4
Quoted 8 lines
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
A message that is not in the archive (sent off-list or not yet received): 8153c0a9-8509-c2c3-c53e-e25f6f7e83f1-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org
RE: [PATCH 3/3] drm/amdgpu: Fix deadlock during GPU reset.
For VM updates, the situation is same as commands submission, deadlock can also be solved by moving ttm_eu_fence_buffer_objects() ahead of push_job() For TTM buffers moves can we pardon gpu_reset routine and wait till the TTM moves complete ?
-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: 2017年10月23日 15:09
To: Liu, Monk <[email protected]>; Grodzovsky, Andrey <[email protected]>; [email protected]
Cc: Koenig, Christian <[email protected]>
Subject: Re: [PATCH 3/3] drm/amdgpu: Fix deadlock during GPU reset.
We discussed that as well, problem is that this won't be sufficient.
We push to the kfifo not only during command submission, but also for VM updates and TTM buffers moves.
So we can still deadlock because of them.
Regards,
Christian.
Am 23.10.2017 um 05:03 schrieb Liu, Monk:
Quoted 247 lines
Why not use a more simple way ?
Like moving
ttm_eu_fence_buffer_objects()to beforeamd_sched_entity_push_job()? That could solve the deadlock from your descriptionAnd the push order is already guaranteed by
context->mutex(which is also a patch from you)BR Monk
-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Andrey Grodzovsky
Sent: 2017年10月20日 21:32
To: [email protected]
Cc: Grodzovsky, Andrey <[email protected]>; Koenig, Christian <[email protected]>
Subject: [PATCH 3/3] drm/amdgpu: Fix deadlock during GPU reset.Switch from kfifo to SPSC queue.
Bug:
Kfifo is limited at size, during GPU reset it would fill up to limit and the pushing thread (producer) would wait for the scheduler worker to consume the items in the fifo while holding reservation lock on a BO. The gpu reset thread on the other hand blocks the scheduler during reset. Before it unblocks the sceduler it might want to recover VRAM and so will try to reserve the same BO the producer thread is already holding creating a deadlock.Fix:
Switch from kfifo to SPSC queue which is unlimited in size.Signed-off-by Andrey Grodzovsky <[email protected]>--- drivers/gpu/drm/amd/scheduler/gpu_sched_trace.h | 4 +- drivers/gpu/drm/amd/scheduler/gpu_scheduler.c | 51 ++++++++++--------------- drivers/gpu/drm/amd/scheduler/gpu_scheduler.h | 4 +- 3 files changed, 26 insertions(+), 33 deletions(-)drivers/gpu/drm/amd/scheduler/gpu_sched_trace.h +2
-2@@ -28,8 +28,8 @@ TRACE_EVENT(amd_sched_job, __entry->id = sched_job->id; __entry->fence = &sched_job->s_fence->finished; __entry->name = sched_job->sched->name; - __entry->job_count =kfifo_len ( - &sched_job->s_entity->job_queue) /sizeof (sched_job); + __entry->job_count =spsc_queue_count ( + &sched_job->s_entity->job_queue); __entry->hw_job_count =atomic_read ( &sched_job->sched->hw_rq_count); ),drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +5
-0@@ -28,9 +28,14 @@#include <drm/drmP.h> #include "gpu_scheduler.h" +#include "spsc_queue.h" +#define CREATE_TRACE_POINTS #include "gpu_sched_trace.h" +#define to_amd_sched_job (sched_job) \ + container_of((sched_job), struct amd_sched_job, queue_node) + static bool amd_sched_entity_is_ready(struct amd_sched_entity *entity); static void amd_sched_wakeup(struct amd_gpu_scheduler *sched); static void amd_sched_process_job(struct dma_fence *f, struct dma_fence_cb *cb); @@ -123,8 +128,6 @@ int amd_sched_entity_init(struct amd_gpu_scheduler *sched,struct amd_sched_rq *rq,uint32_t jobs){drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +1
-5-int r; -if (!(sched && entity && rq))return -EINVAL ;@@ -135,9 +138,7 @@ int amd_sched_entity_init(struct amd_gpu_scheduler *sched,spin_lock_init (&entity->rq_lock);spin_lock_init (&entity->queue_lock); - r =kfifo_alloc (&entity->job_queue, jobs *sizeof (void *),GFP_KERNEL ); -if (r) -return r; +spsc_queue_init (&entity->job_queue);atomic_set (&entity->fence_seq,0 );entity->fence_context = dma_fence_context_alloc(2); @@ -170,7 +171,7 @@ static bool amd_sched_entity_is_initialized(struct amd_gpu_scheduler *sched, static bool amd_sched_entity_is_idle(struct amd_sched_entity *entity) { rmb();drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +2
-2-if (kfifo_is_empty (&entity->job_queue)) +if (spsc_queue_peek (&entity->job_queue) ==NULL )return true;return false;@@ -185,7 +186,7 @@ static bool amd_sched_entity_is_idle(struct amd_sched_entity *entity)*/ static bool amd_sched_entity_is_ready (struct amd_sched_entity *entity) { -if (kfifo_is_empty (&entity->job_queue)) +if (spsc_queue_peek (&entity->job_queue) ==NULL )return false; if (ACCESS_ONCE(entity->dependency)) @@ -227,7 +228,7 @@ void amd_sched_entity_fini(struct amd_gpu_scheduler *sched,*/ kthread_park(sched->thread); kthread_unpark(sched->thread);drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +3
-3-while (kfifo_out (&entity->job_queue, &job,sizeof (job))) { +while ((job = +to_amd_sched_job (spsc_queue_pop (&entity->job_queue)))) +{struct amd_sched_fence *s_fence = job->s_fence;amd_sched_fence_scheduled (s_fence); dma_fence_set_error(&s_fence->finished, -ESRCH); @@ -235,9 +236,7 @@ void amd_sched_entity_fini(struct amd_gpu_scheduler *sched, dma_fence_put(&s_fence->finished); sched->ops->free_job (job); } - } -kfifo_free (&entity->job_queue); } static void amd_sched_entity_wakeup(struct dma_fence *f, struct
dma_fence_cb*cb) @@ -332,18 +331,21 @@ static boolamd_sched_entity_add_dependency_cb(struct amd_sched_entity *entity)}static struct
amd_sched_job*drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +9
-11-amd_sched_entity_peek_job(struct amd_sched_entity *entity) +amd_sched_entity_pop_job(struct amd_sched_entity *entity) {struct amd_gpu_scheduler *sched = entity->sched; -struct amd_sched_job *sched_job; +struct amd_sched_job *sched_job =to_amd_sched_job ( +spsc_queue_peek (&entity->job_queue)); -if (!kfifo_out_peek (&entity->job_queue, &sched_job,sizeof (sched_job))) +if (!sched_job)return NULL ;while ((entity->dependency = sched->ops->dependency (sched_job)))if (amd_sched_entity_add_dependency_cb (entity))return NULL ; + sched_job->s_entity =NULL ; +spsc_queue_pop (&entity->job_queue);return sched_job; }@@ -354,18 +356,14 @@ amd_sched_entity_peek_job(struct amd_sched_entity *entity)* * Returns true if we could submit the job. */ -static bool amd_sched_entity_in (struct amd_sched_job *sched_job) +static void amd_sched_entity_in (struct amd_sched_job *sched_job) {struct amd_gpu_scheduler *sched = sched_job->sched;struct amd_sched_entity *entity = sched_job->s_entity; -bool added, first = false; +bool first = false;spin_lock (&entity->queue_lock); - added =kfifo_in (&entity->job_queue, &sched_job, -sizeof (sched_job)) ==sizeof (sched_job); - -if (added &&kfifo_len (&entity->job_queue) ==sizeof (sched_job)) - first = true; + first =spsc_queue_push (&entity->job_queue, &sched_job->queue_node);spin_unlock (&entity->queue_lock);@@ -377,7 +375,6 @@ static bool amd_sched_entity_in(struct amd_sched_job *sched_job)spin_unlock (&entity->rq_lock);amd_sched_wakeup (sched); } -return added; }/* job_finish is called after hw fence signaled @@ -514,11 +511,8 @@ void amd_sched_job_recovery(struct amd_gpu_scheduler *sched) */ void amd_sched_entity_push_job(struct amd_sched_job *sched_job) {drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +3
-9-struct amd_sched_entity *entity = sched_job->s_entity; -trace_amd_sched_job (sched_job); -wait_event (entity->sched->job_scheduled, -amd_sched_entity_in (sched_job)); +amd_sched_entity_in (sched_job); } /* init a sched_job with basic field */ @@ -611,7 +605,7 @@ static int amd_sched_main(void *param) {struct sched_param sparam = {.sched_priority =1 };struct amd_gpu_scheduler *sched = (struct amd_gpu_scheduler *)param; -int r, count; +int r;sched_setscheduler (current,SCHED_FIFO , &sparam);@@ -629,7 +623,7 @@ static int amd_sched_main(void *param)if (!entity)continue ; - sched_job =amd_sched_entity_peek_job (entity); + sched_job =amd_sched_entity_pop_job (entity);if (!sched_job)continue ;@@ -656,9 +650,6 @@ static int amd_sched_main(void *param)amd_sched_process_job (NULL , &s_fence->cb); } - count =kfifo_out (&entity->job_queue, &sched_job, -sizeof (sched_job)); -WARN_ON (count !=sizeof (sched_job));wake_up (&sched->job_scheduled); }return 0 ;drivers/gpu/drm/amd/scheduler/gpu_scheduler.h +3
-2@@ -26,6 +26,7 @@#include <linux/kfifo.h> #include <linux/dma-fence.h> +#include "spsc_queue.h" struct amd_gpu_scheduler ;struct amd_sched_rq ;@@ -56,7 +57,7 @@ struct amd_sched_entity {struct amd_gpu_scheduler *sched;spinlock_t queue_lock; -struct kfifo job_queue; +struct spsc_queue job_queue;atomic_t fence_seq;uint64_t fence_context;@@ -87,6 +88,7 @@ struct amd_sched_fence { };struct amd_sched_job { +struct spsc_node queue_node;struct amd_gpu_scheduler *sched;struct amd_sched_entity *s_entity;struct amd_sched_fence *s_fence; --2.7.4
Quoted 8 lines
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
A message that is not in the archive (sent off-list or not yet received): BLUPR12MB04490DF904B056086D164BDD84460-7LeqcoF/[email protected]
Re: [PATCH 3/3] drm/amdgpu: Fix deadlock during GPU reset.
Quoted 1 line
For VM updates, the situation is same as commands submission, deadlock can also be solved by moving
ttm_eu_fence_buffer_objects()ahead ofpush_job()
That won't work. All VM updates must be completed while the reservation object lock is still held, otherwise you run into possible concurrent VM updates.
Keep in mind that you can have multiple context per VM, so that isn't covered by the per context lock.
Quoted 1 line
For TTM buffers moves can we pardon
gpu_resetroutine and wait till the TTM moves complete ?
No, TTM callbacks into our driver to do the move. So we don't have an opportunity to drop the reservation lock.
Additional to that kfifo has some other drawbacks as well. For example we waste around 20k of memory for each context for multi media ring buffers which are mostly unused.
Regards,
Christian.
Am 23.10.2017 um 09:24 schrieb Liu, Monk:
Quoted 267 lines
For VM updates, the situation is same as commands submission, deadlock can also be solved by moving
ttm_eu_fence_buffer_objects()ahead ofpush_job()For TTM buffers moves can we pardongpu_resetroutine and wait till the TTM moves complete ?-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: 2017年10月23日 15:09
To: Liu, Monk <[email protected]>; Grodzovsky, Andrey <[email protected]>; [email protected]
Cc: Koenig, Christian <[email protected]>
Subject: Re: [PATCH 3/3] drm/amdgpu: Fix deadlock during GPU reset.We discussed that as well, problem is that this won't be sufficient.
We push to the kfifo not only during command submission, but also for VM updates and TTM buffers moves.
So we can still deadlock because of them.
Regards,
Christian.Am 23.10.2017 um 05:03 schrieb Liu, Monk:
Quoted 247 lines
Why not use a more simple way ?
Like moving
ttm_eu_fence_buffer_objects()to beforeamd_sched_entity_push_job()? That could solve the deadlock from your descriptionAnd the push order is already guaranteed by
context->mutex(which is also a patch from you)BR Monk
-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Andrey Grodzovsky
Sent: 2017年10月20日 21:32
To: [email protected]
Cc: Grodzovsky, Andrey <[email protected]>; Koenig, Christian <[email protected]>
Subject: [PATCH 3/3] drm/amdgpu: Fix deadlock during GPU reset.Switch from kfifo to SPSC queue.
Bug:
Kfifo is limited at size, during GPU reset it would fill up to limit and the pushing thread (producer) would wait for the scheduler worker to consume the items in the fifo while holding reservation lock on a BO. The gpu reset thread on the other hand blocks the scheduler during reset. Before it unblocks the sceduler it might want to recover VRAM and so will try to reserve the same BO the producer thread is already holding creating a deadlock.Fix:
Switch from kfifo to SPSC queue which is unlimited in size.Signed-off-by Andrey Grodzovsky <[email protected]>--- drivers/gpu/drm/amd/scheduler/gpu_sched_trace.h | 4 +- drivers/gpu/drm/amd/scheduler/gpu_scheduler.c | 51 ++++++++++--------------- drivers/gpu/drm/amd/scheduler/gpu_scheduler.h | 4 +- 3 files changed, 26 insertions(+), 33 deletions(-)drivers/gpu/drm/amd/scheduler/gpu_sched_trace.h +2
-2@@ -28,8 +28,8 @@ TRACE_EVENT(amd_sched_job, __entry->id = sched_job->id; __entry->fence = &sched_job->s_fence->finished; __entry->name = sched_job->sched->name; - __entry->job_count =kfifo_len ( - &sched_job->s_entity->job_queue) /sizeof (sched_job); + __entry->job_count =spsc_queue_count ( + &sched_job->s_entity->job_queue); __entry->hw_job_count =atomic_read ( &sched_job->sched->hw_rq_count); ),drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +5
-0@@ -28,9 +28,14 @@#include <drm/drmP.h> #include "gpu_scheduler.h" +#include "spsc_queue.h" +#define CREATE_TRACE_POINTS #include "gpu_sched_trace.h" +#define to_amd_sched_job (sched_job) \ + container_of((sched_job), struct amd_sched_job, queue_node) + static bool amd_sched_entity_is_ready(struct amd_sched_entity *entity); static void amd_sched_wakeup(struct amd_gpu_scheduler *sched); static void amd_sched_process_job(struct dma_fence *f, struct dma_fence_cb *cb); @@ -123,8 +128,6 @@ int amd_sched_entity_init(struct amd_gpu_scheduler *sched,struct amd_sched_rq *rq,uint32_t jobs){drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +1
-5-int r; -if (!(sched && entity && rq))return -EINVAL ;@@ -135,9 +138,7 @@ int amd_sched_entity_init(struct amd_gpu_scheduler *sched,spin_lock_init (&entity->rq_lock);spin_lock_init (&entity->queue_lock); - r =kfifo_alloc (&entity->job_queue, jobs *sizeof (void *),GFP_KERNEL ); -if (r) -return r; +spsc_queue_init (&entity->job_queue);atomic_set (&entity->fence_seq,0 );entity->fence_context = dma_fence_context_alloc(2); @@ -170,7 +171,7 @@ static bool amd_sched_entity_is_initialized(struct amd_gpu_scheduler *sched, static bool amd_sched_entity_is_idle(struct amd_sched_entity *entity) { rmb();drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +2
-2-if (kfifo_is_empty (&entity->job_queue)) +if (spsc_queue_peek (&entity->job_queue) ==NULL )return true;return false;@@ -185,7 +186,7 @@ static bool amd_sched_entity_is_idle(struct amd_sched_entity *entity)*/ static bool amd_sched_entity_is_ready (struct amd_sched_entity *entity) { -if (kfifo_is_empty (&entity->job_queue)) +if (spsc_queue_peek (&entity->job_queue) ==NULL )return false; if (ACCESS_ONCE(entity->dependency)) @@ -227,7 +228,7 @@ void amd_sched_entity_fini(struct amd_gpu_scheduler *sched,*/ kthread_park(sched->thread); kthread_unpark(sched->thread);drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +3
-3-while (kfifo_out (&entity->job_queue, &job,sizeof (job))) { +while ((job = +to_amd_sched_job (spsc_queue_pop (&entity->job_queue)))) +{struct amd_sched_fence *s_fence = job->s_fence;amd_sched_fence_scheduled (s_fence); dma_fence_set_error(&s_fence->finished, -ESRCH); @@ -235,9 +236,7 @@ void amd_sched_entity_fini(struct amd_gpu_scheduler *sched, dma_fence_put(&s_fence->finished); sched->ops->free_job (job); } - } -kfifo_free (&entity->job_queue); } static void amd_sched_entity_wakeup(struct dma_fence *f, struct
dma_fence_cb*cb) @@ -332,18 +331,21 @@ static boolamd_sched_entity_add_dependency_cb(struct amd_sched_entity *entity)}static struct
amd_sched_job*drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +9
-11-amd_sched_entity_peek_job(struct amd_sched_entity *entity) +amd_sched_entity_pop_job(struct amd_sched_entity *entity) {struct amd_gpu_scheduler *sched = entity->sched; -struct amd_sched_job *sched_job; +struct amd_sched_job *sched_job =to_amd_sched_job ( +spsc_queue_peek (&entity->job_queue)); -if (!kfifo_out_peek (&entity->job_queue, &sched_job,sizeof (sched_job))) +if (!sched_job)return NULL ;while ((entity->dependency = sched->ops->dependency (sched_job)))if (amd_sched_entity_add_dependency_cb (entity))return NULL ; + sched_job->s_entity =NULL ; +spsc_queue_pop (&entity->job_queue);return sched_job; }@@ -354,18 +356,14 @@ amd_sched_entity_peek_job(struct amd_sched_entity *entity)* * Returns true if we could submit the job. */ -static bool amd_sched_entity_in (struct amd_sched_job *sched_job) +static void amd_sched_entity_in (struct amd_sched_job *sched_job) {struct amd_gpu_scheduler *sched = sched_job->sched;struct amd_sched_entity *entity = sched_job->s_entity; -bool added, first = false; +bool first = false;spin_lock (&entity->queue_lock); - added =kfifo_in (&entity->job_queue, &sched_job, -sizeof (sched_job)) ==sizeof (sched_job); - -if (added &&kfifo_len (&entity->job_queue) ==sizeof (sched_job)) - first = true; + first =spsc_queue_push (&entity->job_queue, &sched_job->queue_node);spin_unlock (&entity->queue_lock);@@ -377,7 +375,6 @@ static bool amd_sched_entity_in(struct amd_sched_job *sched_job)spin_unlock (&entity->rq_lock);amd_sched_wakeup (sched); } -return added; }/* job_finish is called after hw fence signaled @@ -514,11 +511,8 @@ void amd_sched_job_recovery(struct amd_gpu_scheduler *sched) */void
amd_sched_entity_push_job(struct amd_sched_job *sched_job){drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +3
-9-struct amd_sched_entity *entity = sched_job->s_entity; -trace_amd_sched_job (sched_job); -wait_event (entity->sched->job_scheduled, -amd_sched_entity_in (sched_job)); +amd_sched_entity_in (sched_job); } /* init a sched_job with basic field */ @@ -611,7 +605,7 @@ static int amd_sched_main(void *param) {struct sched_param sparam = {.sched_priority =1 };struct amd_gpu_scheduler *sched = (struct amd_gpu_scheduler *)param; -int r, count; +int r;sched_setscheduler (current,SCHED_FIFO , &sparam);@@ -629,7 +623,7 @@ static int amd_sched_main(void *param)if (!entity)continue ; - sched_job =amd_sched_entity_peek_job (entity); + sched_job =amd_sched_entity_pop_job (entity);if (!sched_job)continue ;@@ -656,9 +650,6 @@ static int amd_sched_main(void *param)amd_sched_process_job (NULL , &s_fence->cb); } - count =kfifo_out (&entity->job_queue, &sched_job, -sizeof (sched_job)); -WARN_ON (count !=sizeof (sched_job));wake_up (&sched->job_scheduled); }return 0 ;drivers/gpu/drm/amd/scheduler/gpu_scheduler.h +3
-2@@ -26,6 +26,7 @@#include <linux/kfifo.h> #include <linux/dma-fence.h> +#include "spsc_queue.h" struct amd_gpu_scheduler ;struct amd_sched_rq ;@@ -56,7 +57,7 @@ struct amd_sched_entity {struct amd_gpu_scheduler *sched;spinlock_t queue_lock; -struct kfifo job_queue; +struct spsc_queue job_queue;atomic_t fence_seq;uint64_t fence_context;@@ -87,6 +88,7 @@ struct amd_sched_fence { };struct amd_sched_job { +struct spsc_node queue_node;struct amd_gpu_scheduler *sched;struct amd_sched_entity *s_entity;struct amd_sched_fence *s_fence; --2.7.4
Quoted 13 lines
Quoted 8 lines
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
A message that is not in the archive (sent off-list or not yet received): a9ed61a0-6e0b-6093-5c0b-4459cfbb2282-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org
RE: [PATCH 3/3] drm/amdgpu: Fix deadlock during GPU reset.
Let's make it clear enough:
Quoted 1 line
That won't work. All VM updates must be completed while the reservation object lock is still held, otherwise you run into possible concurrent VM updates.
All VM updates will be represented as fence which will be hook in the sync object, and after we push job to scheduler, we immediately call ttm_eu_fence_buffer_objects() So that means the reservation object lock is *not* hold anymore, right ? but keep in mind that this point scheduler may not even begin to work, so I don't believe This sentence " All VM updates must be completed while the reservation object lock is still held"
Quoted 1 line
No, TTM callbacks into our driver to do the move. So we don't have an opportunity to drop the reservation lock.
For TTM callbacks, can you help illustrate how the deadlock triggered ???
BR Monk
-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: 2017年10月23日 15:31
To: Liu, Monk <[email protected]>; Koenig, Christian <[email protected]>; Grodzovsky, Andrey <[email protected]>; [email protected]
Subject: Re: [PATCH 3/3] drm/amdgpu: Fix deadlock during GPU reset.
Quoted 3 lines
For VM updates, the situation is same as commands submission, deadlock can also be solved by moving
ttm_eu_fence_buffer_objects()ahead ofpush_job()
That won't work. All VM updates must be completed while the reservation object lock is still held, otherwise you run into possible concurrent VM updates.
Keep in mind that you can have multiple context per VM, so that isn't covered by the per context lock.
Quoted 1 line
For TTM buffers moves can we pardon
gpu_resetroutine and wait till the TTM moves complete ?
No, TTM callbacks into our driver to do the move. So we don't have an opportunity to drop the reservation lock.
Additional to that kfifo has some other drawbacks as well. For example we waste around 20k of memory for each context for multi media ring buffers which are mostly unused.
Regards,
Christian.
Am 23.10.2017 um 09:24 schrieb Liu, Monk:
Quoted 270 lines
For VM updates, the situation is same as commands submission, deadlock
can also be solved by movingttm_eu_fence_buffer_objects()ahead ofpush_job()For TTM buffers moves can we pardongpu_resetroutine and wait till the TTM moves complete ?-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: 2017年10月23日 15:09
To: Liu, Monk <[email protected]>; Grodzovsky, Andrey <[email protected]>; [email protected]
Cc: Koenig, Christian <[email protected]>
Subject: Re: [PATCH 3/3] drm/amdgpu: Fix deadlock during GPU reset.We discussed that as well, problem is that this won't be sufficient.
We push to the kfifo not only during command submission, but also for VM updates and TTM buffers moves.
So we can still deadlock because of them.
Regards,
Christian.Am 23.10.2017 um 05:03 schrieb Liu, Monk:
Quoted 249 lines
Why not use a more simple way ?
Like moving
ttm_eu_fence_buffer_objects()to beforeamd_sched_entity_push_job()? That could solve the deadlock from your descriptionAnd the push order is already guaranteed by
context->mutex(which is also a patch from you)BR Monk
-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Andrey Grodzovsky
Sent: 2017年10月20日 21:32
To: [email protected]
Cc: Grodzovsky, Andrey <[email protected]>; Koenig, Christian <[email protected]>
Subject: [PATCH 3/3] drm/amdgpu: Fix deadlock during GPU reset.Switch from kfifo to SPSC queue.
Bug:
Kfifo is limited at size, during GPU reset it would fill up to limit and the pushing thread (producer) would wait for the scheduler worker to consume the items in the fifo while holding reservation lock on a BO. The gpu reset thread on the other hand blocks the scheduler during reset. Before it unblocks the sceduler it might want to recover VRAM and so will try to reserve the same BO the producer thread is already holding creating a deadlock.Fix:
Switch from kfifo to SPSC queue which is unlimited in size.Signed-off-by Andrey Grodzovsky <[email protected]>--- drivers/gpu/drm/amd/scheduler/gpu_sched_trace.h | 4 +- drivers/gpu/drm/amd/scheduler/gpu_scheduler.c | 51 ++++++++++--------------- drivers/gpu/drm/amd/scheduler/gpu_scheduler.h | 4 +- 3 files changed, 26 insertions(+), 33 deletions(-)drivers/gpu/drm/amd/scheduler/gpu_sched_trace.h +2
-2@@ -28,8 +28,8 @@ TRACE_EVENT(amd_sched_job, __entry->id = sched_job->id; __entry->fence = &sched_job->s_fence->finished; __entry->name = sched_job->sched->name; - __entry->job_count =kfifo_len ( - &sched_job->s_entity->job_queue) /sizeof (sched_job); + __entry->job_count =spsc_queue_count ( + &sched_job->s_entity->job_queue); __entry->hw_job_count =atomic_read ( &sched_job->sched->hw_rq_count); ),drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +5
-0@@ -28,9 +28,14 @@#include <drm/drmP.h> #include "gpu_scheduler.h" +#include "spsc_queue.h" +#define CREATE_TRACE_POINTS #include "gpu_sched_trace.h" +#define to_amd_sched_job (sched_job) \ + container_of((sched_job), struct amd_sched_job, queue_node) + static bool amd_sched_entity_is_ready(struct amd_sched_entity *entity); static void amd_sched_wakeup(struct amd_gpu_scheduler *sched); static void amd_sched_process_job(struct dma_fence *f, struct dma_fence_cb *cb); @@ -123,8 +128,6 @@ int amd_sched_entity_init(struct amd_gpu_scheduler *sched,struct amd_sched_rq *rq,uint32_t jobs){drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +1
-5-int r; -if (!(sched && entity && rq))return -EINVAL ;@@ -135,9 +138,7 @@ int amd_sched_entity_init(struct amd_gpu_scheduler *sched,spin_lock_init (&entity->rq_lock);spin_lock_init (&entity->queue_lock); - r =kfifo_alloc (&entity->job_queue, jobs *sizeof (void *),GFP_KERNEL ); -if (r) -return r; +spsc_queue_init (&entity->job_queue);atomic_set (&entity->fence_seq,0 );entity->fence_context = dma_fence_context_alloc(2); @@ -170,7 +171,7 @@ static bool amd_sched_entity_is_initialized(struct amd_gpu_scheduler *sched, static bool amd_sched_entity_is_idle(struct amd_sched_entity *entity) { rmb();drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +2
-2-if (kfifo_is_empty (&entity->job_queue)) +if (spsc_queue_peek (&entity->job_queue) ==NULL )return true;return false;@@ -185,7 +186,7 @@ static bool amd_sched_entity_is_idle(struct amd_sched_entity *entity)*/ static bool amd_sched_entity_is_ready (struct amd_sched_entity *entity) { -if (kfifo_is_empty (&entity->job_queue)) +if (spsc_queue_peek (&entity->job_queue) ==NULL )return false; if (ACCESS_ONCE(entity->dependency)) @@ -227,7 +228,7 @@ void amd_sched_entity_fini(struct amd_gpu_scheduler *sched,*/ kthread_park(sched->thread); kthread_unpark(sched->thread);drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +3
-3-while (kfifo_out (&entity->job_queue, &job,sizeof (job))) { +while ((job = +to_amd_sched_job (spsc_queue_pop (&entity->job_queue)))) +{struct amd_sched_fence *s_fence = job->s_fence;amd_sched_fence_scheduled (s_fence); dma_fence_set_error(&s_fence->finished, -ESRCH); @@ -235,9 +236,7 @@ void amd_sched_entity_fini(struct amd_gpu_scheduler *sched, dma_fence_put(&s_fence->finished); sched->ops->free_job (job); } - } -kfifo_free (&entity->job_queue); } static void amd_sched_entity_wakeup(struct dma_fence *f, struct
dma_fence_cb*cb) @@ -332,18 +331,21 @@ static boolamd_sched_entity_add_dependency_cb(struct amd_sched_entity *entity)}static struct
amd_sched_job*drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +10
-11-amd_sched_entity_peek_job(struct amd_sched_entity *entity) +amd_sched_entity_pop_job(struct amd_sched_entity *entity) {struct amd_gpu_scheduler *sched = entity->sched; -struct amd_sched_job *sched_job; +struct amd_sched_job *sched_job =to_amd_sched_job ( +spsc_queue_peek (&entity->job_queue)); -if (!kfifo_out_peek (&entity->job_queue, &sched_job,sizeof (sched_job))) +if (!sched_job)return NULL ;while ((entity->dependency = sched->ops->dependency (sched_job)))if (amd_sched_entity_add_dependency_cb (entity))return NULL ; + sched_job->s_entity =NULL ; +spsc_queue_pop (&entity->job_queue);return sched_job; }@@ -354,18 +356,14 @@ amd_sched_entity_peek_job(struct amd_sched_entity *entity)* * Returns true if we could submit the job. */ -static bool amd_sched_entity_in (struct amd_sched_job *sched_job) +static void amd_sched_entity_in (struct amd_sched_job *sched_job) {struct amd_gpu_scheduler *sched = sched_job->sched;struct amd_sched_entity *entity = sched_job->s_entity; -bool added, first = false; +bool first = false;spin_lock (&entity->queue_lock); - added =kfifo_in (&entity->job_queue, &sched_job, -sizeof (sched_job)) ==sizeof (sched_job); - -if (added &&kfifo_len (&entity->job_queue) ==sizeof (sched_job)) - first = true; + first =spsc_queue_push (&entity->job_queue, +&sched_job->queue_node);spin_unlock (&entity->queue_lock);@@ -377,7 +375,6 @@ static bool amd_sched_entity_in(struct amd_sched_job *sched_job)spin_unlock (&entity->rq_lock);amd_sched_wakeup (sched); } -return added; }/* job_finish is called after hw fence signaled @@ -514,11 +511,8 @@ void amd_sched_job_recovery(struct amd_gpu_scheduler *sched) */void
amd_sched_entity_push_job(struct amd_sched_job *sched_job){drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +3
-9-struct amd_sched_entity *entity = sched_job->s_entity; -trace_amd_sched_job (sched_job); -wait_event (entity->sched->job_scheduled, -amd_sched_entity_in (sched_job)); +amd_sched_entity_in (sched_job); } /* init a sched_job with basic field */ @@ -611,7 +605,7 @@ static int amd_sched_main(void *param) {struct sched_param sparam = {.sched_priority =1 };struct amd_gpu_scheduler *sched = (struct amd_gpu_scheduler *)param; -int r, count; +int r;sched_setscheduler (current,SCHED_FIFO , &sparam);@@ -629,7 +623,7 @@ static int amd_sched_main(void *param)if (!entity)continue ; - sched_job =amd_sched_entity_peek_job (entity); + sched_job =amd_sched_entity_pop_job (entity);if (!sched_job)continue ;@@ -656,9 +650,6 @@ static int amd_sched_main(void *param)amd_sched_process_job (NULL , &s_fence->cb); } - count =kfifo_out (&entity->job_queue, &sched_job, -sizeof (sched_job)); -WARN_ON (count !=sizeof (sched_job));wake_up (&sched->job_scheduled); }return 0 ;drivers/gpu/drm/amd/scheduler/gpu_scheduler.h +3
-2@@ -26,6 +26,7 @@#include <linux/kfifo.h> #include <linux/dma-fence.h> +#include "spsc_queue.h" struct amd_gpu_scheduler ;struct amd_sched_rq ;@@ -56,7 +57,7 @@ struct amd_sched_entity {struct amd_gpu_scheduler *sched;spinlock_t queue_lock; -struct kfifo job_queue; +struct spsc_queue job_queue;atomic_t fence_seq;uint64_t fence_context;@@ -87,6 +88,7 @@ struct amd_sched_fence { };struct amd_sched_job { +struct spsc_node queue_node;struct amd_gpu_scheduler *sched;struct amd_sched_entity *s_entity;struct amd_sched_fence *s_fence; --2.7.4
Quoted 13 lines
Quoted 8 lines
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
A message that is not in the archive (sent off-list or not yet received): BLUPR12MB0449E8F1F121B277B598430684460-7LeqcoF/[email protected]
RE: [PATCH 3/3] drm/amdgpu: Fix deadlock during GPU reset.
I read the source again, looks like I misunderstand your point, let me clear it again:
1, the cs_submit deadlock can be resolved by moving eu_fence_buffer()
2, in GEM_VA IOCTL, the reservation lock is hold in the very beginning, and vm_update_directories() need that lock held till all related JOBs submitted to VM entity, so if VM entity is full the deadlock could hit, and we can not move "eu_backoff_reservaton" to before push_job() under job_submit() due to concurrent concerns 3, for TTM moves, e.g. ttm_bo_moves(), it will take the reserve lock on the bo from swap_lru list and then call into amdgpu ttm part, so the deadlock could also hit in job_sumbit() when KFIFO is FULL.
You guys are right, the the second and third deadlock is still there
BR Monk
-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Liu, Monk
Sent: 2017年10月23日 15:39
To: Koenig, Christian <[email protected]>; Grodzovsky, Andrey <[email protected]>; [email protected]
Subject: RE: [PATCH 3/3] drm/amdgpu: Fix deadlock during GPU reset.
Let's make it clear enough:
Quoted 1 line
That won't work. All VM updates must be completed while the reservation object lock is still held, otherwise you run into possible concurrent VM updates.
All VM updates will be represented as fence which will be hook in the sync object, and after we push job to scheduler, we immediately call ttm_eu_fence_buffer_objects() So that means the reservation object lock is *not* hold anymore, right ? but keep in mind that this point scheduler may not even begin to work, so I don't believe This sentence " All VM updates must be completed while the reservation object lock is still held"
Quoted 1 line
No, TTM callbacks into our driver to do the move. So we don't have an opportunity to drop the reservation lock.
For TTM callbacks, can you help illustrate how the deadlock triggered ???
BR Monk
-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: 2017年10月23日 15:31
To: Liu, Monk <[email protected]>; Koenig, Christian <[email protected]>; Grodzovsky, Andrey <[email protected]>; [email protected]
Subject: Re: [PATCH 3/3] drm/amdgpu: Fix deadlock during GPU reset.
Quoted 3 lines
For VM updates, the situation is same as commands submission, deadlock can also be solved by moving
ttm_eu_fence_buffer_objects()ahead ofpush_job()
That won't work. All VM updates must be completed while the reservation object lock is still held, otherwise you run into possible concurrent VM updates.
Keep in mind that you can have multiple context per VM, so that isn't covered by the per context lock.
Quoted 1 line
For TTM buffers moves can we pardon
gpu_resetroutine and wait till the TTM moves complete ?
No, TTM callbacks into our driver to do the move. So we don't have an opportunity to drop the reservation lock.
Additional to that kfifo has some other drawbacks as well. For example we waste around 20k of memory for each context for multi media ring buffers which are mostly unused.
Regards,
Christian.
Am 23.10.2017 um 09:24 schrieb Liu, Monk:
Quoted 270 lines
For VM updates, the situation is same as commands submission, deadlock
can also be solved by movingttm_eu_fence_buffer_objects()ahead ofpush_job()For TTM buffers moves can we pardongpu_resetroutine and wait till the TTM moves complete ?-----Original Message-----
From: Christian König [mailto:[email protected]]
Sent: 2017年10月23日 15:09
To: Liu, Monk <[email protected]>; Grodzovsky, Andrey <[email protected]>; [email protected]
Cc: Koenig, Christian <[email protected]>
Subject: Re: [PATCH 3/3] drm/amdgpu: Fix deadlock during GPU reset.We discussed that as well, problem is that this won't be sufficient.
We push to the kfifo not only during command submission, but also for VM updates and TTM buffers moves.
So we can still deadlock because of them.
Regards,
Christian.Am 23.10.2017 um 05:03 schrieb Liu, Monk:
Quoted 249 lines
Why not use a more simple way ?
Like moving
ttm_eu_fence_buffer_objects()to beforeamd_sched_entity_push_job()? That could solve the deadlock from your descriptionAnd the push order is already guaranteed by
context->mutex(which is also a patch from you)BR Monk
-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Andrey Grodzovsky
Sent: 2017年10月20日 21:32
To: [email protected]
Cc: Grodzovsky, Andrey <[email protected]>; Koenig, Christian <[email protected]>
Subject: [PATCH 3/3] drm/amdgpu: Fix deadlock during GPU reset.Switch from kfifo to SPSC queue.
Bug:
Kfifo is limited at size, during GPU reset it would fill up to limit and the pushing thread (producer) would wait for the scheduler worker to consume the items in the fifo while holding reservation lock on a BO. The gpu reset thread on the other hand blocks the scheduler during reset. Before it unblocks the sceduler it might want to recover VRAM and so will try to reserve the same BO the producer thread is already holding creating a deadlock.Fix:
Switch from kfifo to SPSC queue which is unlimited in size.Signed-off-by Andrey Grodzovsky <[email protected]>--- drivers/gpu/drm/amd/scheduler/gpu_sched_trace.h | 4 +- drivers/gpu/drm/amd/scheduler/gpu_scheduler.c | 51 ++++++++++--------------- drivers/gpu/drm/amd/scheduler/gpu_scheduler.h | 4 +- 3 files changed, 26 insertions(+), 33 deletions(-)drivers/gpu/drm/amd/scheduler/gpu_sched_trace.h +2
-2@@ -28,8 +28,8 @@ TRACE_EVENT(amd_sched_job, __entry->id = sched_job->id; __entry->fence = &sched_job->s_fence->finished; __entry->name = sched_job->sched->name; - __entry->job_count =kfifo_len ( - &sched_job->s_entity->job_queue) /sizeof (sched_job); + __entry->job_count =spsc_queue_count ( + &sched_job->s_entity->job_queue); __entry->hw_job_count =atomic_read ( &sched_job->sched->hw_rq_count); ),drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +5
-0@@ -28,9 +28,14 @@#include <drm/drmP.h> #include "gpu_scheduler.h" +#include "spsc_queue.h" +#define CREATE_TRACE_POINTS #include "gpu_sched_trace.h" +#define to_amd_sched_job (sched_job) \ + container_of((sched_job), struct amd_sched_job, queue_node) + static bool amd_sched_entity_is_ready(struct amd_sched_entity *entity); static void amd_sched_wakeup(struct amd_gpu_scheduler *sched); static void amd_sched_process_job(struct dma_fence *f, struct dma_fence_cb *cb); @@ -123,8 +128,6 @@ int amd_sched_entity_init(struct amd_gpu_scheduler *sched,struct amd_sched_rq *rq,uint32_t jobs){drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +1
-5-int r; -if (!(sched && entity && rq))return -EINVAL ;@@ -135,9 +138,7 @@ int amd_sched_entity_init(struct amd_gpu_scheduler *sched,spin_lock_init (&entity->rq_lock);spin_lock_init (&entity->queue_lock); - r =kfifo_alloc (&entity->job_queue, jobs *sizeof (void *),GFP_KERNEL ); -if (r) -return r; +spsc_queue_init (&entity->job_queue);atomic_set (&entity->fence_seq,0 );entity->fence_context = dma_fence_context_alloc(2); @@ -170,7 +171,7 @@ static bool amd_sched_entity_is_initialized(struct amd_gpu_scheduler *sched, static bool amd_sched_entity_is_idle(struct amd_sched_entity *entity) { rmb();drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +2
-2-if (kfifo_is_empty (&entity->job_queue)) +if (spsc_queue_peek (&entity->job_queue) ==NULL )return true;return false;@@ -185,7 +186,7 @@ static bool amd_sched_entity_is_idle(struct amd_sched_entity *entity)*/ static bool amd_sched_entity_is_ready (struct amd_sched_entity *entity) { -if (kfifo_is_empty (&entity->job_queue)) +if (spsc_queue_peek (&entity->job_queue) ==NULL )return false; if (ACCESS_ONCE(entity->dependency)) @@ -227,7 +228,7 @@ void amd_sched_entity_fini(struct amd_gpu_scheduler *sched,*/ kthread_park(sched->thread); kthread_unpark(sched->thread);drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +3
-3-while (kfifo_out (&entity->job_queue, &job,sizeof (job))) { +while ((job = +to_amd_sched_job (spsc_queue_pop (&entity->job_queue)))) +{struct amd_sched_fence *s_fence = job->s_fence;amd_sched_fence_scheduled (s_fence); dma_fence_set_error(&s_fence->finished, -ESRCH); @@ -235,9 +236,7 @@ void amd_sched_entity_fini(struct amd_gpu_scheduler *sched, dma_fence_put(&s_fence->finished); sched->ops->free_job (job); } - } -kfifo_free (&entity->job_queue); } static void amd_sched_entity_wakeup(struct dma_fence *f, struct
dma_fence_cb*cb) @@ -332,18 +331,21 @@ static boolamd_sched_entity_add_dependency_cb(struct amd_sched_entity *entity)}static struct
amd_sched_job*drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +10
-11-amd_sched_entity_peek_job(struct amd_sched_entity *entity) +amd_sched_entity_pop_job(struct amd_sched_entity *entity) {struct amd_gpu_scheduler *sched = entity->sched; -struct amd_sched_job *sched_job; +struct amd_sched_job *sched_job =to_amd_sched_job ( +spsc_queue_peek (&entity->job_queue)); -if (!kfifo_out_peek (&entity->job_queue, &sched_job,sizeof (sched_job))) +if (!sched_job)return NULL ;while ((entity->dependency = sched->ops->dependency (sched_job)))if (amd_sched_entity_add_dependency_cb (entity))return NULL ; + sched_job->s_entity =NULL ; +spsc_queue_pop (&entity->job_queue);return sched_job; }@@ -354,18 +356,14 @@ amd_sched_entity_peek_job(struct amd_sched_entity *entity)* * Returns true if we could submit the job. */ -static bool amd_sched_entity_in (struct amd_sched_job *sched_job) +static void amd_sched_entity_in (struct amd_sched_job *sched_job) {struct amd_gpu_scheduler *sched = sched_job->sched;struct amd_sched_entity *entity = sched_job->s_entity; -bool added, first = false; +bool first = false;spin_lock (&entity->queue_lock); - added =kfifo_in (&entity->job_queue, &sched_job, -sizeof (sched_job)) ==sizeof (sched_job); - -if (added &&kfifo_len (&entity->job_queue) ==sizeof (sched_job)) - first = true; + first =spsc_queue_push (&entity->job_queue, +&sched_job->queue_node);spin_unlock (&entity->queue_lock);@@ -377,7 +375,6 @@ static bool amd_sched_entity_in(struct amd_sched_job *sched_job)spin_unlock (&entity->rq_lock);amd_sched_wakeup (sched); } -return added; }/* job_finish is called after hw fence signaled @@ -514,11 +511,8 @@ void amd_sched_job_recovery(struct amd_gpu_scheduler *sched) */void
amd_sched_entity_push_job(struct amd_sched_job *sched_job){drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +3
-9-struct amd_sched_entity *entity = sched_job->s_entity; -trace_amd_sched_job (sched_job); -wait_event (entity->sched->job_scheduled, -amd_sched_entity_in (sched_job)); +amd_sched_entity_in (sched_job); } /* init a sched_job with basic field */ @@ -611,7 +605,7 @@ static int amd_sched_main(void *param) {struct sched_param sparam = {.sched_priority =1 };struct amd_gpu_scheduler *sched = (struct amd_gpu_scheduler *)param; -int r, count; +int r;sched_setscheduler (current,SCHED_FIFO , &sparam);@@ -629,7 +623,7 @@ static int amd_sched_main(void *param)if (!entity)continue ; - sched_job =amd_sched_entity_peek_job (entity); + sched_job =amd_sched_entity_pop_job (entity);if (!sched_job)continue ;@@ -656,9 +650,6 @@ static int amd_sched_main(void *param)amd_sched_process_job (NULL , &s_fence->cb); } - count =kfifo_out (&entity->job_queue, &sched_job, -sizeof (sched_job)); -WARN_ON (count !=sizeof (sched_job));wake_up (&sched->job_scheduled); }return 0 ;drivers/gpu/drm/amd/scheduler/gpu_scheduler.h +3
-2@@ -26,6 +26,7 @@#include <linux/kfifo.h> #include <linux/dma-fence.h> +#include "spsc_queue.h" struct amd_gpu_scheduler ;struct amd_sched_rq ;@@ -56,7 +57,7 @@ struct amd_sched_entity {struct amd_gpu_scheduler *sched;spinlock_t queue_lock; -struct kfifo job_queue; +struct spsc_queue job_queue;atomic_t fence_seq;uint64_t fence_context;@@ -87,6 +88,7 @@ struct amd_sched_fence { };struct amd_sched_job { +struct spsc_node queue_node;struct amd_gpu_scheduler *sched;struct amd_sched_entity *s_entity;struct amd_sched_fence *s_fence; --2.7.4
Quoted 13 lines
Quoted 8 lines
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Re: [PATCH 3/3] drm/amdgpu: Fix deadlock during GPU reset.
In reply to a message not in the archiveAm 20.10.2017 um 15:32 schrieb Andrey Grodzovsky:
Quoted 168 lines
Switch from kfifo to SPSC queue.
Bug:
Kfifo is limited at size, during GPU reset it would fill up to limit and the pushing thread (producer) would wait for the scheduler worker to consume the items in the fifo while holding reservation lock on a BO. The gpu reset thread on the other hand blocks the scheduler during reset. Before it unblocks the sceduler it might want to recover VRAM and so will try to reserve the same BO the producer thread is already holding creating a deadlock.Fix:
Switch from kfifo to SPSC queue which is unlimited in size.Signed-off-by Andrey Grodzovsky <[email protected]>--- drivers/gpu/drm/amd/scheduler/gpu_sched_trace.h | 4 +- drivers/gpu/drm/amd/scheduler/gpu_scheduler.c | 51 ++++++++++--------------- drivers/gpu/drm/amd/scheduler/gpu_scheduler.h | 4 +- 3 files changed, 26 insertions(+), 33 deletions(-)drivers/gpu/drm/amd/scheduler/gpu_sched_trace.h +2
-2@@ -28,8 +28,8 @@ TRACE_EVENT(amd_sched_job, __entry->id = sched_job->id; __entry->fence = &sched_job->s_fence->finished; __entry->name = sched_job->sched->name; - __entry->job_count =kfifo_len ( - &sched_job->s_entity->job_queue) /sizeof (sched_job); + __entry->job_count =spsc_queue_count ( + &sched_job->s_entity->job_queue); __entry->hw_job_count =atomic_read ( &sched_job->sched->hw_rq_count); ),drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +18
-21@@ -28,9 +28,14 @@#include <drm/drmP.h> #include "gpu_scheduler.h" +#include "spsc_queue.h" +#define CREATE_TRACE_POINTS #include "gpu_sched_trace.h" +#define to_amd_sched_job (sched_job) \ + container_of((sched_job), struct amd_sched_job, queue_node) +static bool amd_sched_entity_is_ready (struct amd_sched_entity *entity);static void amd_sched_wakeup (struct amd_gpu_scheduler *sched);static void amd_sched_process_job (struct dma_fence *f,struct dma_fence_cb *cb);@@ -123,8 +128,6 @@ int amd_sched_entity_init(struct amd_gpu_scheduler *sched, struct amd_sched_rq *rq, uint32_t jobs) { - int r; -if (!(sched && entity && rq))return -EINVAL ;@@ -135,9 +138,7 @@ int amd_sched_entity_init(struct amd_gpu_scheduler *sched,spin_lock_init (&entity->rq_lock);spin_lock_init (&entity->queue_lock); - r =kfifo_alloc (&entity->job_queue, jobs *sizeof (void *),GFP_KERNEL ); -if (r) -return r; +spsc_queue_init (&entity->job_queue);atomic_set (&entity->fence_seq,0 ); entity->fence_context =dma_fence_context_alloc (2 );@@ -170,7 +171,7 @@ static bool amd_sched_entity_is_initialized(struct amd_gpu_scheduler *sched,static bool amd_sched_entity_is_idle (struct amd_sched_entity *entity) {rmb (); -if (kfifo_is_empty (&entity->job_queue)) +if (spsc_queue_peek (&entity->job_queue) ==NULL )return true;return false;@@ -185,7 +186,7 @@ static bool amd_sched_entity_is_idle(struct amd_sched_entity *entity)*/ static bool amd_sched_entity_is_ready (struct amd_sched_entity *entity) { -if (kfifo_is_empty (&entity->job_queue)) +if (spsc_queue_peek (&entity->job_queue) ==NULL )return false;if (ACCESS_ONCE (entity->dependency))@@ -227,7 +228,7 @@ void amd_sched_entity_fini(struct amd_gpu_scheduler *sched,*/ kthread_park (sched->thread);kthread_unpark (sched->thread); -while (kfifo_out (&entity->job_queue, &job,sizeof (job))) { +while ((job =to_amd_sched_job (spsc_queue_pop (&entity->job_queue)))) {struct amd_sched_fence *s_fence = job->s_fence;amd_sched_fence_scheduled (s_fence);dma_fence_set_error (&s_fence->finished, -ESRCH );@@ -235,9 +236,7 @@ void amd_sched_entity_fini(struct amd_gpu_scheduler *sched,dma_fence_put (&s_fence->finished); sched->ops->free_job (job); } - } -kfifo_free (&entity->job_queue); }static void amd_sched_entity_wakeup (struct dma_fence *f,struct dma_fence_cb *cb)@@ -332,18 +331,21 @@ static bool amd_sched_entity_add_dependency_cb(struct amd_sched_entity *entity) }static struct amd_sched_job * -amd_sched_entity_peek_job (struct amd_sched_entity *entity) +amd_sched_entity_pop_job (struct amd_sched_entity *entity) {struct amd_gpu_scheduler *sched = entity->sched; -struct amd_sched_job *sched_job; +struct amd_sched_job *sched_job =to_amd_sched_job ( +spsc_queue_peek (&entity->job_queue)); -if (!kfifo_out_peek (&entity->job_queue, &sched_job,sizeof (sched_job))) +if (!sched_job)return NULL ;while ((entity->dependency = sched->ops->dependency (sched_job)))if (amd_sched_entity_add_dependency_cb (entity))return NULL ; + sched_job->s_entity =NULL ; +spsc_queue_pop (&entity->job_queue);return sched_job; }@@ -354,18 +356,14 @@ amd_sched_entity_peek_job(struct amd_sched_entity *entity)* * Returns true if we could submit the job. */ -static bool amd_sched_entity_in (struct amd_sched_job *sched_job) +static void amd_sched_entity_in (struct amd_sched_job *sched_job) {struct amd_gpu_scheduler *sched = sched_job->sched;struct amd_sched_entity *entity = sched_job->s_entity; -bool added, first = false; +bool first = false;spin_lock (&entity->queue_lock); - added =kfifo_in (&entity->job_queue, &sched_job, -sizeof (sched_job)) ==sizeof (sched_job); - -if (added &&kfifo_len (&entity->job_queue) ==sizeof (sched_job)) - first = true; + first =spsc_queue_push (&entity->job_queue, &sched_job->queue_node);spin_unlock (&entity->queue_lock);@@ -377,7 +375,6 @@ static bool amd_sched_entity_in(struct amd_sched_job *sched_job)spin_unlock (&entity->rq_lock);amd_sched_wakeup (sched); } -return added; }/* job_finish is called after hw fence signaled
drivers/gpu/drm/amd/scheduler/gpu_scheduler.c void amd_sched_job_recovery(struct amd_gpu_scheduler *sched) · line 515
Quoted 9 lines
drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +1
-4@@ -514,11 +511,8 @@ void amd_sched_job_recovery(struct amd_gpu_scheduler *sched)*/ void amd_sched_entity_push_job (struct amd_sched_job *sched_job) { -struct amd_sched_entity *entity = sched_job->s_entity; -trace_amd_sched_job (sched_job); -wait_event (entity->sched->job_scheduled, -amd_sched_entity_in (sched_job)); +amd_sched_entity_in (sched_job);
To save a few loc amd_sched_entity_in() can now be completely merged into this function.
Apart from that the patch looks really good to me and is Reviewed-by: Christian König <[email protected]>.
Ping me internally for new work if you now have time again.
Thanks,
Christian.
Quoted 53 lines
drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +2
-5}/* init a sched_job with basic field */ @@ -611,7 +605,7 @@ static int amd_sched_main(void *param) {struct sched_param sparam = {.sched_priority =1 };struct amd_gpu_scheduler *sched = (struct amd_gpu_scheduler *)param; -int r, count; +int r;sched_setscheduler (current,SCHED_FIFO , &sparam);@@ -629,7 +623,7 @@ static int amd_sched_main(void *param)if (!entity)continue ; - sched_job =amd_sched_entity_peek_job (entity); + sched_job =amd_sched_entity_pop_job (entity);if (!sched_job)continue ;@@ -656,9 +650,6 @@ static int amd_sched_main(void *param)amd_sched_process_job (NULL , &s_fence->cb); } - count =kfifo_out (&entity->job_queue, &sched_job, -sizeof (sched_job)); -WARN_ON (count !=sizeof (sched_job));wake_up (&sched->job_scheduled); }return 0 ;drivers/gpu/drm/amd/scheduler/gpu_scheduler.h +2
-1@@ -26,6 +26,7 @@#include <linux/kfifo.h> #include <linux/dma-fence.h> +#include "spsc_queue.h" struct amd_gpu_scheduler ;struct amd_sched_rq ;@@ -56,7 +57,7 @@ struct amd_sched_entity {struct amd_gpu_scheduler *sched;spinlock_t queue_lock; -struct kfifo job_queue; +struct spsc_queue job_queue;atomic_t fence_seq;uint64_t fence_context;
drivers/gpu/drm/amd/scheduler/gpu_scheduler.h struct amd_sched_fence { · line 94
Quoted 7 lines
drivers/gpu/drm/amd/scheduler/gpu_scheduler.h +1
-0@@ -87,6 +88,7 @@ struct amd_sched_fence { };struct amd_sched_job { +struct spsc_node queue_node;struct amd_gpu_scheduler *sched;struct amd_sched_entity *s_entity;struct amd_sched_fence *s_fence;
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Quoted 6 lines · Andrey Grodzovsky
Bug:
amdgpu_job_free_cbwas accessings_job->s_entitywhen the allocatedamdgpu_ctx(and the entity inside it) were already deallocated fromamdgpu_cs_parser_fini.Fix: Save job's priority on it's creation instead of accessing it from
s_entitylater on.
I'm not sure if this is the correct approach for a fix.
Keeping s_entity as a dangling pointer could result in a similar bugs being reintroduced. For example, there would still be a race condition between amdgpu_cs_parser_fini() and amdgpu_job_dependency().
Instead, it might be better for the job to grab a reference to the context during job_init(), and put that reference on job free.
Regards,
Andres
Quoted 111 lines
Signed-off-by Andrey Grodzovsky <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 3 +-- drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 5 ++--- drivers/gpu/drm/amd/scheduler/gpu_scheduler.c | 1 + drivers/gpu/drm/amd/scheduler/gpu_scheduler.h | 32 ++++++++++++--------------- 4 files changed, 18 insertions(+), 23 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +1
-2@@ -1192,8 +1192,7 @@ static int amdgpu_cs_submit(struct amdgpu_cs_parser *p, job->uf_sequence = seq;amdgpu_job_free_resources (job); -amdgpu_ring_priority_get (job->ring, -amd_sched_get_job_priority (&job->base)); +amdgpu_ring_priority_get (job->ring, job->base.s_priority);trace_amdgpu_cs_ioctl (job);amd_sched_entity_push_job (&job->base);drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +2
-3@@ -104,7 +104,7 @@ static void amdgpu_job_free_cb(struct amd_sched_job *s_job) { struct amdgpu_job *job = container_of(s_job, struct amdgpu_job, base); -amdgpu_ring_priority_put (job->ring,amd_sched_get_job_priority (s_job)); +amdgpu_ring_priority_put (job->ring, s_job->s_priority);dma_fence_put (job->fence);amdgpu_sync_free (&job->sync);amdgpu_sync_free (&job->dep_sync);@@ -141,8 +141,7 @@ int amdgpu_job_submit(struct amdgpu_job *job, struct amdgpu_ring *ring, job->fence_ctx = entity->fence_context; *f =dma_fence_get (&job->base.s_fence->finished);amdgpu_job_free_resources (job); -amdgpu_ring_priority_get (job->ring, -amd_sched_get_job_priority (&job->base)); +amdgpu_ring_priority_get (job->ring, job->base.s_priority);amd_sched_entity_push_job (&job->base);return 0 ;drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +1
-0@@ -529,6 +529,7 @@ int amd_sched_job_init(struct amd_sched_job *job, { job->sched = sched; job->s_entity = entity; + job->s_priority = entity->rq - sched->sched_rq; job->s_fence =amd_sched_fence_create (entity, owner);if (!job->s_fence)return -ENOMEM ;drivers/gpu/drm/amd/scheduler/gpu_scheduler.h +14
-12@@ -30,6 +30,19 @@struct amd_gpu_scheduler ;struct amd_sched_rq ; +enum amd_sched_priority { +AMD_SCHED_PRIORITY_MIN , +AMD_SCHED_PRIORITY_LOW =AMD_SCHED_PRIORITY_MIN , +AMD_SCHED_PRIORITY_NORMAL , +AMD_SCHED_PRIORITY_HIGH_SW , +AMD_SCHED_PRIORITY_HIGH_HW , +AMD_SCHED_PRIORITY_KERNEL , +AMD_SCHED_PRIORITY_MAX , +AMD_SCHED_PRIORITY_INVALID =-1 , +AMD_SCHED_PRIORITY_UNSET =-2 +}; + +/** * A scheduler entity is a wrapper around a job queue or a group * of other entities. Entities take turns emitting jobs from their @@ -83,6 +96,7 @@ struct amd_sched_job {struct delayed_work work_tdr;uint64_t id;atomic_t karma; +enum amd_sched_priority s_priority; };extern const struct dma_fence_ops amd_sched_fence_ops_scheduled;@@ -114,18 +128,6 @@ struct amd_sched_backend_ops {void (*free_job)(struct amd_sched_job *sched_job); }; -enum amd_sched_priority { -AMD_SCHED_PRIORITY_MIN , -AMD_SCHED_PRIORITY_LOW =AMD_SCHED_PRIORITY_MIN , -AMD_SCHED_PRIORITY_NORMAL , -AMD_SCHED_PRIORITY_HIGH_SW , -AMD_SCHED_PRIORITY_HIGH_HW , -AMD_SCHED_PRIORITY_KERNEL , -AMD_SCHED_PRIORITY_MAX , -AMD_SCHED_PRIORITY_INVALID =-1 , -AMD_SCHED_PRIORITY_UNSET =-2 -}; -/** * One scheduler is implemented for each hardware ring */
drivers/gpu/drm/amd/scheduler/gpu_scheduler.h bool amd_sched_dependency_optimized(struct dma_fence* fence, · line 181
Quoted 10 lines
drivers/gpu/drm/amd/scheduler/gpu_scheduler.h +0
-6@@ -176,10 +178,4 @@ bool amd_sched_dependency_optimized(struct dma_fence* fence, struct amd_sched_entity *entity);void amd_sched_job_kickout (struct amd_sched_job *s_job); -static inline enum amd_sched_priority -amd_sched_get_job_priority (struct amd_sched_job *job) -{ -return (job->s_entity->rq - job->sched->sched_rq); -} -#endif
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
A message that is not in the archive (sent off-list or not yet received): ea764df5-6c80-e1cd-a4cd-562c86c7d553-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org
Quoted 7 lines · Andres Rodriguez
Quoted 4 lines · Andrey Grodzovsky
Bug:
amdgpu_job_free_cbwas accessings_job->s_entitywhen the allocatedamdgpu_ctx(and the entity inside it) were already deallocated fromamdgpu_cs_parser_fini.
Quoted 6 lines
Quoted 2 lines
Fix: Save job's priority on it's creation instead of accessing it from
s_entitylater on.I'm not sure if this is the correct approach for a fix.
Keeping
s_entityas a dangling pointer could result in a similar bugs being reintroduced. For example, there would still be a race condition betweenamdgpu_cs_parser_fini()andamdgpu_job_dependency().
.dependency hook is called only in once place amd_sched_entity_pop_job, amdgpu_cs_parser_fini will wait (from amd_sched_entity_fini) for wake_up(&sched->job_scheduled) from amd_sched_main so I don't see a race here.
Quoted 3 lines
Instead, it might be better for the job to grab a reference to the context during
job_init(), and put that reference on job free.
Originally it was my thinkimg to, but I consulted Christian and he advised that quote - "it's not the best idea since
the problem is that when we terminate a process we need to make sure that all resources are released or at least not hold for much longer. When we keep the ctx alive with the job we need to also keep the VM alive and that means we need to keep all the VM page tables alive".
Thanks,
Andrey
Quoted 116 lines
Regards,
AndresQuoted 112 lines
Signed-off-by Andrey Grodzovsky <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 3 +-- drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 5 ++--- drivers/gpu/drm/amd/scheduler/gpu_scheduler.c | 1 + drivers/gpu/drm/amd/scheduler/gpu_scheduler.h | 32 ++++++++++++--------------- 4 files changed, 18 insertions(+), 23 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +1
-2@@ -1192,8 +1192,7 @@ static int amdgpu_cs_submit(struct amdgpu_cs_parser *p, job->uf_sequence = seq;amdgpu_job_free_resources (job); -amdgpu_ring_priority_get (job->ring, -amd_sched_get_job_priority (&job->base)); +amdgpu_ring_priority_get (job->ring, job->base.s_priority);trace_amdgpu_cs_ioctl (job);amd_sched_entity_push_job (&job->base);drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +2
-2@@ -104,7 +104,7 @@ static void amdgpu_job_free_cb(struct amd_sched_job *s_job) { struct amdgpu_job *job = container_of(s_job, struct amdgpu_job, base); -amdgpu_ring_priority_put (job->ring,amd_sched_get_job_priority (s_job)); +amdgpu_ring_priority_put (job->ring, s_job->s_priority);dma_fence_put (job->fence);amdgpu_sync_free (&job->sync);amdgpu_sync_free (&job->dep_sync);@@ -141,8 +141,7 @@ int amdgpu_job_submit(struct amdgpu_job *job,struct amdgpu_ring *ring, job->fence_ctx = entity->fence_context; *f =dma_fence_get (&job->base.s_fence->finished);amdgpu_job_free_resources (job); -amdgpu_ring_priority_get (job->ring, -amd_sched_get_job_priority (&job->base)); +amdgpu_ring_priority_get (job->ring, job->base.s_priority);amd_sched_entity_push_job (&job->base);return 0 ;drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +1
-0@@ -529,6 +529,7 @@ int amd_sched_job_init(struct amd_sched_job *job, { job->sched = sched; job->s_entity = entity; + job->s_priority = entity->rq - sched->sched_rq; job->s_fence =amd_sched_fence_create (entity, owner);if (!job->s_fence)return -ENOMEM ;drivers/gpu/drm/amd/scheduler/gpu_scheduler.h +13
-11@@ -30,6 +30,19 @@struct amd_gpu_scheduler ;struct amd_sched_rq ; +enum amd_sched_priority { +AMD_SCHED_PRIORITY_MIN , +AMD_SCHED_PRIORITY_LOW =AMD_SCHED_PRIORITY_MIN , +AMD_SCHED_PRIORITY_NORMAL , +AMD_SCHED_PRIORITY_HIGH_SW , +AMD_SCHED_PRIORITY_HIGH_HW , +AMD_SCHED_PRIORITY_KERNEL , +AMD_SCHED_PRIORITY_MAX , +AMD_SCHED_PRIORITY_INVALID =-1 , +AMD_SCHED_PRIORITY_UNSET =-2 +}; + +/** * A scheduler entity is a wrapper around a job queue or a group * of other entities. Entities take turns emitting jobs from their @@ -83,6 +96,7 @@ struct amd_sched_job {struct delayed_work work_tdr;uint64_t id;atomic_t karma; +enum amd_sched_priority s_priority; };extern const struct dma_fence_ops amd_sched_fence_ops_scheduled;@@ -114,18 +128,6 @@ struct amd_sched_backend_ops {void (*free_job)(struct amd_sched_job *sched_job); }; -enum amd_sched_priority { -AMD_SCHED_PRIORITY_MIN , -AMD_SCHED_PRIORITY_LOW =AMD_SCHED_PRIORITY_MIN , -AMD_SCHED_PRIORITY_NORMAL , -AMD_SCHED_PRIORITY_HIGH_SW , -AMD_SCHED_PRIORITY_HIGH_HW , -AMD_SCHED_PRIORITY_KERNEL , -AMD_SCHED_PRIORITY_MAX , -AMD_SCHED_PRIORITY_INVALID =-1 , -AMD_SCHED_PRIORITY_UNSET =-2 -}; -/** * One scheduler is implemented for each hardware ring */ @@ -176,10 +178,4 @@ bool amd_sched_dependency_optimized(structdma_fence * fence, struct amd_sched_entity *entity);void amd_sched_job_kickout (struct amd_sched_job *s_job);
Quoted 8 lines
drivers/gpu/drm/amd/scheduler/gpu_scheduler.h bool amd_sched_dependency_optimized(struct dma_fence* fence, · line 181
Quoted 8 lines
drivers/gpu/drm/amd/scheduler/gpu_scheduler.h +0
-5-static inline enum amd_sched_priority -amd_sched_get_job_priority (struct amd_sched_job *job) -{ -return (job->s_entity->rq - job->sched->sched_rq); -} -#endif
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
A message that is not in the archive (sent off-list or not yet received): ecb8a9d2-ba94-c7a4-f5d7-156b061245d8-5C7GfCeVMHo@public.gmane.org
Quoted 17 lines · Andrey Grodzovsky
Quoted 7 lines · Andres Rodriguez
Quoted 4 lines · Andrey Grodzovsky
Bug:
amdgpu_job_free_cbwas accessings_job->s_entitywhen the allocatedamdgpu_ctx(and the entity inside it) were already deallocated fromamdgpu_cs_parser_fini.Quoted 6 lines
Quoted 2 lines
Fix: Save job's priority on it's creation instead of accessing it from
s_entitylater on.I'm not sure if this is the correct approach for a fix.
Keeping
s_entityas a dangling pointer could result in a similar bugs being reintroduced. For example, there would still be a race condition betweenamdgpu_cs_parser_fini()andamdgpu_job_dependency().
Quoted 9 lines
.dependency hook is called only in once place
amd_sched_entity_pop_job,amdgpu_cs_parser_finiwill wait (fromamd_sched_entity_fini) forwake_up(&sched->job_scheduled)fromamd_sched_mainso I don't see a race here.Quoted 3 lines
Instead, it might be better for the job to grab a reference to the context during
job_init(), and put that reference on job free.Originally it was my thinkimg to, but I consulted Christian and he advised that quote - "it's not the best idea since
the problem is that when we terminate a process we need to make sure that all resources are released or at least not hold for much longer. When we keep the ctx alive with the job we need to also keep the VM alive and that means we need to keep all the VM page tables alive".
That makes sense.
Since s_entity is tied to the context reference held by the parser, can you set it to NULL when you drop the context reference there?
At least that way we can easily detect misuse of s_entity after it enters a "possibly deleted" state.
Regards,
Andres
Quoted 117 lines
Thanks,
AndreyQuoted 114 lines
Regards,
AndresQuoted 110 lines
Signed-off-by Andrey Grodzovsky <[email protected]>---
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 3 +-- drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 5 ++---drivers/gpu/drm/amd/scheduler/gpu_scheduler.c| 1 +drivers/gpu/drm/amd/scheduler/gpu_scheduler.h| 32
++++++++++++---------------4 files changed, 18 insertions(+), 23 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +1
-2@@ -1192,8 +1192,7 @@ static int amdgpu_cs_submit(structamdgpu_cs_parser *p, job->uf_sequence = seq; amdgpu_job_free_resources(job); - amdgpu_ring_priority_get(job->ring, -amd_sched_get_job_priority (&job->base)); + amdgpu_ring_priority_get(job->ring, job->base.s_priority); trace_amdgpu_cs_ioctl(job); amd_sched_entity_push_job(&job->base);drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +1
-0@@ -104,7 +104,7 @@ static void amdgpu_job_free_cb(struct amd_sched_job *s_job) { struct amdgpu_job *job = container_of(s_job, struct amdgpu_job, base); - amdgpu_ring_priority_put(job->ring, amd_sched_get_job_priority(s_job)); + amdgpu_ring_priority_put(job->ring, s_job->s_priority); dma_fence_put(job->fence); amdgpu_sync_free(&job->sync);
amdgpu_sync_free(&job->dep_sync);drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +1
-2@@ -141,8 +141,7 @@ int amdgpu_job_submit(struct amdgpu_job *job,struct amdgpu_ring *ring, job->fence_ctx = entity->fence_context; *f = dma_fence_get(&job->base.s_fence->finished); amdgpu_job_free_resources(job); - amdgpu_ring_priority_get(job->ring, -amd_sched_get_job_priority (&job->base)); + amdgpu_ring_priority_get(job->ring, job->base.s_priority); amd_sched_entity_push_job(&job->base); return 0;drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +1
-0@@ -529,6 +529,7 @@ int amd_sched_job_init(struct amd_sched_job *job, { job->sched = sched; job->s_entity = entity; + job->s_priority = entity->rq - sched->sched_rq; job->s_fence = amd_sched_fence_create(entity, owner); if (!job->s_fence) return -ENOMEM;drivers/gpu/drm/amd/scheduler/gpu_scheduler.h +13
-11@@ -30,6 +30,19 @@ struct amd_gpu_scheduler; struct amd_sched_rq; +enum amd_sched_priority { + AMD_SCHED_PRIORITY_MIN, + AMD_SCHED_PRIORITY_LOW = AMD_SCHED_PRIORITY_MIN, + AMD_SCHED_PRIORITY_NORMAL, + AMD_SCHED_PRIORITY_HIGH_SW, + AMD_SCHED_PRIORITY_HIGH_HW, + AMD_SCHED_PRIORITY_KERNEL, + AMD_SCHED_PRIORITY_MAX, + AMD_SCHED_PRIORITY_INVALID = -1, + AMD_SCHED_PRIORITY_UNSET = -2 +}; + + /*** A scheduler entity is a wrapper around a job queue or a group * of other entities. Entities take turns emitting jobs from their @@ -83,6 +96,7 @@ struct amd_sched_job { struct delayed_work work_tdr; uint64_t id; atomic_t karma; + enum amd_sched_priority s_priority; }; extern const struct dma_fence_ops amd_sched_fence_ops_scheduled;@@ -114,18 +128,6 @@ struct amd_sched_backend_ops { void (*free_job)(struct amd_sched_job *sched_job); }; -enum amd_sched_priority { - AMD_SCHED_PRIORITY_MIN, - AMD_SCHED_PRIORITY_LOW = AMD_SCHED_PRIORITY_MIN, - AMD_SCHED_PRIORITY_NORMAL, - AMD_SCHED_PRIORITY_HIGH_SW, - AMD_SCHED_PRIORITY_HIGH_HW, - AMD_SCHED_PRIORITY_KERNEL, - AMD_SCHED_PRIORITY_MAX, - AMD_SCHED_PRIORITY_INVALID = -1, - AMD_SCHED_PRIORITY_UNSET = -2 -}; - /*** One scheduler is implemented for each hardware ring */ @@ -176,10 +178,4 @@ bool amd_sched_dependency_optimized(struct dma_fence* fence, struct amd_sched_entity *entity); void amd_sched_job_kickout(struct amd_sched_job *s_job);
Quoted 9 lines
Quoted 8 lines
Quoted 8 lines
-static inline enum
amd_sched_prioritydrivers/gpu/drm/amd/scheduler/gpu_scheduler.h +0
-5-amd_sched_get_job_priority(struct amd_sched_job *job) -{ - return (job->s_entity->rq - job->sched->sched_rq); -} -#endif
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
A message that is not in the archive (sent off-list or not yet received): 86853e29-07c1-700b-8aa2-2183fbc64d2f-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org
Quoted 26 lines · Andres Rodriguez
Quoted 23 lines · Andrey Grodzovsky
Quoted 7 lines · Andres Rodriguez
Quoted 4 lines · Andrey Grodzovsky
Bug:
amdgpu_job_free_cbwas accessings_job->s_entitywhen the allocatedamdgpu_ctx(and the entity inside it) were already deallocated fromamdgpu_cs_parser_fini.Quoted 6 lines
Quoted 2 lines
Fix: Save job's priority on it's creation instead of accessing it from
s_entitylater on.I'm not sure if this is the correct approach for a fix.
Keeping
s_entityas a dangling pointer could result in a similar bugs being reintroduced. For example, there would still be a race condition betweenamdgpu_cs_parser_fini()andamdgpu_job_dependency()..dependency hook is called only in once place
amd_sched_entity_pop_job,amdgpu_cs_parser_finiwill wait (fromamd_sched_entity_fini) forwake_up(&sched->job_scheduled)fromamd_sched_mainso I don't see a race here.Quoted 3 lines
Instead, it might be better for the job to grab a reference to the context during
job_init(), and put that reference on job free.
Quoted 7 lines
Quoted 3 lines
Originally it was my thinkimg to, but I consulted Christian and he advised that quote - "it's not the best idea since
the problem is that when we terminate a process we need to make sure that all resources are released or at least not hold for much longer. When we keep the ctx alive with the job we need to also keep the VM alive and that means we need to keep all the VM page tables alive".That makes sense.
Since
s_entityis tied to the context reference held by the parser, can you set it to NULL when you drop the context reference there?
O am not sure i understand - you want to send s_job->s_entity to NULL in amd_sched_entity_fini for each remaining job in the queue ? But all the jobs remaining in the queue are destroyed there anyway.
Thanks,
Andrey
Quoted 134 lines
At least that way we can easily detect misuse of
s_entityafter it enters a "possibly deleted" state.Regards,
AndresQuoted 128 lines
Thanks,
AndreyQuoted 124 lines
Regards,
AndresQuoted 120 lines
Signed-off-by Andrey Grodzovsky <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 3 +-- drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 5 ++--- drivers/gpu/drm/amd/scheduler/gpu_scheduler.c | 1 + drivers/gpu/drm/amd/scheduler/gpu_scheduler.h | 32 ++++++++++++--------------- 4 files changed, 18 insertions(+), 23 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +1
-2@@ -1192,8 +1192,7 @@ static int amdgpu_cs_submit(struct amdgpu_cs_parser *p, job->uf_sequence = seq;amdgpu_job_free_resources (job); -amdgpu_ring_priority_get (job->ring, -amd_sched_get_job_priority (&job->base)); +amdgpu_ring_priority_get (job->ring, job->base.s_priority);trace_amdgpu_cs_ioctl (job);amd_sched_entity_push_job (&job->base);drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +2
-2@@ -104,7 +104,7 @@ static void amdgpu_job_free_cb(struct amd_sched_job *s_job) { struct amdgpu_job *job = container_of(s_job, struct amdgpu_job, base); -amdgpu_ring_priority_put (job->ring,amd_sched_get_job_priority (s_job)); +amdgpu_ring_priority_put (job->ring, s_job->s_priority);dma_fence_put (job->fence);amdgpu_sync_free (&job->sync);amdgpu_sync_free (&job->dep_sync);@@ -141,8 +141,7 @@ int amdgpu_job_submit(struct amdgpu_job *job,struct amdgpu_ring *ring, job->fence_ctx = entity->fence_context; *f =dma_fence_get (&job->base.s_fence->finished);amdgpu_job_free_resources (job); -amdgpu_ring_priority_get (job->ring, -amd_sched_get_job_priority (&job->base)); +amdgpu_ring_priority_get (job->ring, job->base.s_priority);amd_sched_entity_push_job (&job->base);return 0 ;drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +1
-0@@ -529,6 +529,7 @@ int amd_sched_job_init(struct amd_sched_job *job, { job->sched = sched; job->s_entity = entity; + job->s_priority = entity->rq - sched->sched_rq; job->s_fence =amd_sched_fence_create (entity, owner);if (!job->s_fence)return -ENOMEM ;drivers/gpu/drm/amd/scheduler/gpu_scheduler.h +13
-16@@ -30,6 +30,19 @@struct amd_gpu_scheduler ;struct amd_sched_rq ; +enum amd_sched_priority { +AMD_SCHED_PRIORITY_MIN , +AMD_SCHED_PRIORITY_LOW =AMD_SCHED_PRIORITY_MIN , +AMD_SCHED_PRIORITY_NORMAL , +AMD_SCHED_PRIORITY_HIGH_SW , +AMD_SCHED_PRIORITY_HIGH_HW , +AMD_SCHED_PRIORITY_KERNEL , +AMD_SCHED_PRIORITY_MAX , +AMD_SCHED_PRIORITY_INVALID =-1 , +AMD_SCHED_PRIORITY_UNSET =-2 +}; + +/** * A scheduler entity is a wrapper around a job queue or a group * of other entities. Entities take turns emitting jobs from their @@ -83,6 +96,7 @@ struct amd_sched_job {struct delayed_work work_tdr;uint64_t id;atomic_t karma; +enum amd_sched_priority s_priority; };extern const struct dma_fence_ops amd_sched_fence_ops_scheduled;@@ -114,18 +128,6 @@ struct amd_sched_backend_ops {void (*free_job)(struct amd_sched_job *sched_job); }; -enum amd_sched_priority { -AMD_SCHED_PRIORITY_MIN , -AMD_SCHED_PRIORITY_LOW =AMD_SCHED_PRIORITY_MIN , -AMD_SCHED_PRIORITY_NORMAL , -AMD_SCHED_PRIORITY_HIGH_SW , -AMD_SCHED_PRIORITY_HIGH_HW , -AMD_SCHED_PRIORITY_KERNEL , -AMD_SCHED_PRIORITY_MAX , -AMD_SCHED_PRIORITY_INVALID =-1 , -AMD_SCHED_PRIORITY_UNSET =-2 -}; -/** * One scheduler is implemented for each hardware ring */ @@ -176,10 +178,4 @@ bool amd_sched_dependency_optimized(structdma_fence * fence, struct amd_sched_entity *entity);void amd_sched_job_kickout (struct amd_sched_job *s_job); -static inline enum amd_sched_priority -amd_sched_get_job_priority (struct amd_sched_job *job) -{ -return (job->s_entity->rq - job->sched->sched_rq); -} -#endif
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
A message that is not in the archive (sent off-list or not yet received): 184ba414-981a-02eb-1a6f-7bd878b4f3b1-5C7GfCeVMHo@public.gmane.org
Am 20.10.2017 um 18:51 schrieb Andrey Grodzovsky:
Quoted 33 lines · Andres Rodriguez
Quoted 30 lines · Andrey Grodzovsky
Quoted 26 lines · Andres Rodriguez
Quoted 7 lines · Andrey Grodzovsky
Quoted 4 lines
Bug:
amdgpu_job_free_cbwas accessings_job->s_entitywhen the allocatedamdgpu_ctx(and the entity inside it) were already deallocated fromamdgpu_cs_parser_fini.Quoted 6 lines
Quoted 2 lines
Fix: Save job's priority on it's creation instead of accessing it from
s_entitylater on.I'm not sure if this is the correct approach for a fix.
Keeping
s_entityas a dangling pointer could result in a similar bugs being reintroduced. For example, there would still be a race condition betweenamdgpu_cs_parser_fini()andamdgpu_job_dependency()..dependency hook is called only in once place
amd_sched_entity_pop_job,amdgpu_cs_parser_finiwill wait (fromamd_sched_entity_fini) forwake_up(&sched->job_scheduled)fromamd_sched_mainso I don't see a race here.Quoted 3 lines
Instead, it might be better for the job to grab a reference to the context during
job_init(), and put that reference on job free.Originally it was my thinkimg to, but I consulted Christian and he advised that quote - "it's not the best idea since
the problem is that when we terminate a process we need to make sure that all resources are released or at least not hold for much longer. When we keep the ctx alive with the job we need to also keep the VM alive and that means we need to keep all the VM page tables alive".
Quoted 5 lines
Quoted 3 lines
That makes sense.
Since
s_entityis tied to the context reference held by the parser, can you set it to NULL when you drop the context reference there?O am not sure i understand - you want to send
s_job->s_entityto NULL inamd_sched_entity_finifor each remaining job in the queue ? But all the jobs remaining in the queue are destroyed there anyway.
I think what Andres means here is exactly what we planned to do anyway. Set job->s_entity to NULL as soon as we know that the entity is not used any more and might be released.
In the long term we should target towards making s_job->s_entity as well as job->vm superfluous. This way we could even push remaining jobs on a graveyard entity when we destroy one and timeout.
Alternatively we could look into why wait_event_killable is sometimes not killable as the name says :)
Maybe we can get to a point where we can finally reboot the system cleanly even when the GPU is stuck.
Regards,
Christian.
Quoted 137 lines
Thanks,
AndreyQuoted 132 lines
At least that way we can easily detect misuse of
s_entityafter it enters a "possibly deleted" state.Regards,
AndresQuoted 126 lines
Thanks,
AndreyQuoted 122 lines
Regards,
AndresQuoted 118 lines
Signed-off-by Andrey Grodzovsky <[email protected]>---
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 3 +-- drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 5 ++---drivers/gpu/drm/amd/scheduler/gpu_scheduler.c| 1 +drivers/gpu/drm/amd/scheduler/gpu_scheduler.h| 32
++++++++++++---------------4 files changed, 18 insertions(+), 23 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +1
-2@@ -1192,8 +1192,7 @@ static int amdgpu_cs_submit(structamdgpu_cs_parser *p, job->uf_sequence = seq; amdgpu_job_free_resources(job); - amdgpu_ring_priority_get(job->ring, -amd_sched_get_job_priority (&job->base)); + amdgpu_ring_priority_get(job->ring, job->base.s_priority); trace_amdgpu_cs_ioctl(job); amd_sched_entity_push_job(&job->base);drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +1
-0@@ -104,7 +104,7 @@ static void amdgpu_job_free_cb(struct amd_sched_job *s_job) { struct amdgpu_job *job = container_of(s_job, struct amdgpu_job, base); - amdgpu_ring_priority_put(job->ring, amd_sched_get_job_priority(s_job)); + amdgpu_ring_priority_put(job->ring, s_job->s_priority); dma_fence_put(job->fence); amdgpu_sync_free(&job->sync);
amdgpu_sync_free(&job->dep_sync);drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +1
-2@@ -141,8 +141,7 @@ int amdgpu_job_submit(struct amdgpu_job *job,struct amdgpu_ring *ring, job->fence_ctx = entity->fence_context; *f = dma_fence_get(&job->base.s_fence->finished); amdgpu_job_free_resources(job); - amdgpu_ring_priority_get(job->ring, -amd_sched_get_job_priority (&job->base)); + amdgpu_ring_priority_get(job->ring, job->base.s_priority); amd_sched_entity_push_job(&job->base); return 0;drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +1
-0@@ -529,6 +529,7 @@ int amd_sched_job_init(struct amd_sched_job *job, { job->sched = sched; job->s_entity = entity; + job->s_priority = entity->rq - sched->sched_rq; job->s_fence = amd_sched_fence_create(entity, owner); if (!job->s_fence) return -ENOMEM;drivers/gpu/drm/amd/scheduler/gpu_scheduler.h +13
-11@@ -30,6 +30,19 @@ struct amd_gpu_scheduler; struct amd_sched_rq; +enum amd_sched_priority { + AMD_SCHED_PRIORITY_MIN, + AMD_SCHED_PRIORITY_LOW = AMD_SCHED_PRIORITY_MIN, + AMD_SCHED_PRIORITY_NORMAL, + AMD_SCHED_PRIORITY_HIGH_SW, + AMD_SCHED_PRIORITY_HIGH_HW, + AMD_SCHED_PRIORITY_KERNEL, + AMD_SCHED_PRIORITY_MAX, + AMD_SCHED_PRIORITY_INVALID = -1, + AMD_SCHED_PRIORITY_UNSET = -2 +}; + + /*** A scheduler entity is a wrapper around a job queue or a group * of other entities. Entities take turns emitting jobs from their @@ -83,6 +96,7 @@ struct amd_sched_job { struct delayed_work work_tdr; uint64_t id; atomic_t karma; + enum amd_sched_priority s_priority; }; extern const struct dma_fence_ops amd_sched_fence_ops_scheduled;@@ -114,18 +128,6 @@ struct amd_sched_backend_ops { void (*free_job)(struct amd_sched_job *sched_job); }; -enum amd_sched_priority { - AMD_SCHED_PRIORITY_MIN, - AMD_SCHED_PRIORITY_LOW = AMD_SCHED_PRIORITY_MIN, - AMD_SCHED_PRIORITY_NORMAL, - AMD_SCHED_PRIORITY_HIGH_SW, - AMD_SCHED_PRIORITY_HIGH_HW, - AMD_SCHED_PRIORITY_KERNEL, - AMD_SCHED_PRIORITY_MAX, - AMD_SCHED_PRIORITY_INVALID = -1, - AMD_SCHED_PRIORITY_UNSET = -2 -}; - /*** One scheduler is implemented for each hardware ring */ @@ -176,10 +178,4 @@ bool amd_sched_dependency_optimized(struct dma_fence* fence, struct amd_sched_entity *entity); void amd_sched_job_kickout(struct amd_sched_job *s_job);-static inline enum
amd_sched_prioritydrivers/gpu/drm/amd/scheduler/gpu_scheduler.h +0
-5-amd_sched_get_job_priority(struct amd_sched_job *job) -{ - return (job->s_entity->rq - job->sched->sched_rq); -} -#endif
Quoted 4 lines
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
A message that is not in the archive (sent off-list or not yet received): 79fd05fa-711c-c454-36a1-54b31adda225-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org
Andres Rodriguez
Reviewed-byQuoted 34 lines · Christian König
Am 20.10.2017 um 18:51 schrieb Andrey Grodzovsky:
Quoted 33 lines · Andres Rodriguez
Quoted 30 lines · Andrey Grodzovsky
Quoted 26 lines · Andres Rodriguez
Quoted 7 lines · Andrey Grodzovsky
Quoted 4 lines
Bug:
amdgpu_job_free_cbwas accessings_job->s_entitywhen the allocatedamdgpu_ctx(and the entity inside it) were already deallocated fromamdgpu_cs_parser_fini.Quoted 6 lines
Quoted 2 lines
Fix: Save job's priority on it's creation instead of accessing it from
s_entitylater on.I'm not sure if this is the correct approach for a fix.
Keeping
s_entityas a dangling pointer could result in a similar bugs being reintroduced. For example, there would still be a race condition betweenamdgpu_cs_parser_fini()andamdgpu_job_dependency()..dependency hook is called only in once place
amd_sched_entity_pop_job,amdgpu_cs_parser_finiwill wait (fromamd_sched_entity_fini) forwake_up(&sched->job_scheduled)fromamd_sched_mainso I don't see a race here.Quoted 3 lines
Instead, it might be better for the job to grab a reference to the context during
job_init(), and put that reference on job free.Originally it was my thinkimg to, but I consulted Christian and he advised that quote - "it's not the best idea since
the problem is that when we terminate a process we need to make sure that all resources are released or at least not hold for much longer. When we keep the ctx alive with the job we need to also keep the VM alive and that means we need to keep all the VM page tables alive".
Quoted 7 lines
Quoted 5 lines
Quoted 3 lines
That makes sense.
Since
s_entityis tied to the context reference held by the parser, can you set it to NULL when you drop the context reference there?O am not sure i understand - you want to send
s_job->s_entityto NULL inamd_sched_entity_finifor each remaining job in the queue ? But all the jobs remaining in the queue are destroyed there anyway.I think what Andres means here is exactly what we planned to do anyway. Set
job->s_entityto NULL as soon as we know that the entity is not used any more and might be released.
Yeah this is what I would like to see. If you already have discussed it and have a plan to address it, then this patch looks good to me for static and dynamic priorities.
Feel free to add:
Quoted 153 lines
In the long term we should target towards making
s_job->s_entityas well asjob->vmsuperfluous. This way we could even push remaining jobs on a graveyard entity when we destroy one and timeout.Alternatively we could look into why
wait_event_killableis sometimes not killable as the name says :)Maybe we can get to a point where we can finally reboot the system cleanly even when the GPU is stuck.
Regards,
Christian.Quoted 141 lines
Thanks,
AndreyQuoted 132 lines
At least that way we can easily detect misuse of
s_entityafter it enters a "possibly deleted" state.Regards,
AndresQuoted 126 lines
Thanks,
AndreyQuoted 122 lines
Regards,
AndresQuoted 118 lines
Signed-off-by Andrey Grodzovsky <[email protected]>---
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 3 +-- drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 5 ++---drivers/gpu/drm/amd/scheduler/gpu_scheduler.c| 1 +drivers/gpu/drm/amd/scheduler/gpu_scheduler.h| 32
++++++++++++---------------4 files changed, 18 insertions(+), 23 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +1
-2@@ -1192,8 +1192,7 @@ static int amdgpu_cs_submit(structamdgpu_cs_parser *p, job->uf_sequence = seq; amdgpu_job_free_resources(job); - amdgpu_ring_priority_get(job->ring, -amd_sched_get_job_priority (&job->base)); + amdgpu_ring_priority_get(job->ring, job->base.s_priority); trace_amdgpu_cs_ioctl(job); amd_sched_entity_push_job(&job->base);drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +1
-0@@ -104,7 +104,7 @@ static void amdgpu_job_free_cb(struct amd_sched_job *s_job) { struct amdgpu_job *job = container_of(s_job, struct amdgpu_job, base); - amdgpu_ring_priority_put(job->ring, amd_sched_get_job_priority(s_job)); + amdgpu_ring_priority_put(job->ring, s_job->s_priority); dma_fence_put(job->fence); amdgpu_sync_free(&job->sync);
amdgpu_sync_free(&job->dep_sync);drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +1
-2@@ -141,8 +141,7 @@ int amdgpu_job_submit(struct amdgpu_job *job,struct amdgpu_ring *ring, job->fence_ctx = entity->fence_context; *f = dma_fence_get(&job->base.s_fence->finished); amdgpu_job_free_resources(job); - amdgpu_ring_priority_get(job->ring, -amd_sched_get_job_priority (&job->base)); + amdgpu_ring_priority_get(job->ring, job->base.s_priority); amd_sched_entity_push_job(&job->base); return 0;drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +1
-0@@ -529,6 +529,7 @@ int amd_sched_job_init(struct amd_sched_job *job, { job->sched = sched; job->s_entity = entity; + job->s_priority = entity->rq - sched->sched_rq; job->s_fence = amd_sched_fence_create(entity, owner); if (!job->s_fence) return -ENOMEM;drivers/gpu/drm/amd/scheduler/gpu_scheduler.h +13
-11@@ -30,6 +30,19 @@ struct amd_gpu_scheduler; struct amd_sched_rq; +enum amd_sched_priority { + AMD_SCHED_PRIORITY_MIN, + AMD_SCHED_PRIORITY_LOW = AMD_SCHED_PRIORITY_MIN, + AMD_SCHED_PRIORITY_NORMAL, + AMD_SCHED_PRIORITY_HIGH_SW, + AMD_SCHED_PRIORITY_HIGH_HW, + AMD_SCHED_PRIORITY_KERNEL, + AMD_SCHED_PRIORITY_MAX, + AMD_SCHED_PRIORITY_INVALID = -1, + AMD_SCHED_PRIORITY_UNSET = -2 +}; + + /*** A scheduler entity is a wrapper around a job queue or a group * of other entities. Entities take turns emitting jobs from their @@ -83,6 +96,7 @@ struct amd_sched_job { struct delayed_work work_tdr; uint64_t id; atomic_t karma; + enum amd_sched_priority s_priority; }; extern const struct dma_fence_ops amd_sched_fence_ops_scheduled;@@ -114,18 +128,6 @@ struct amd_sched_backend_ops { void (*free_job)(struct amd_sched_job *sched_job); }; -enum amd_sched_priority { - AMD_SCHED_PRIORITY_MIN, - AMD_SCHED_PRIORITY_LOW = AMD_SCHED_PRIORITY_MIN, - AMD_SCHED_PRIORITY_NORMAL, - AMD_SCHED_PRIORITY_HIGH_SW, - AMD_SCHED_PRIORITY_HIGH_HW, - AMD_SCHED_PRIORITY_KERNEL, - AMD_SCHED_PRIORITY_MAX, - AMD_SCHED_PRIORITY_INVALID = -1, - AMD_SCHED_PRIORITY_UNSET = -2 -}; - /*** One scheduler is implemented for each hardware ring */ @@ -176,10 +178,4 @@ bool amd_sched_dependency_optimized(struct dma_fence* fence, struct amd_sched_entity *entity); void amd_sched_job_kickout(struct amd_sched_job *s_job);-static inline enum
amd_sched_prioritydrivers/gpu/drm/amd/scheduler/gpu_scheduler.h +0
-5-amd_sched_get_job_priority(struct amd_sched_job *job) -{ - return (job->s_entity->rq - job->sched->sched_rq); -} -#endif
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Christian König
Reviewed-byAm 20.10.2017 um 15:32 schrieb Andrey Grodzovsky:
Quoted 8 lines
Bug:
amdgpu_job_free_cbwas accessings_job->s_entitywhen the allocatedamdgpu_ctx(and the entity inside it) were already deallocated fromamdgpu_cs_parser_fini.Fix: Save job's priority on it's creation instead of accessing it from
s_entitylater on.Signed-off-by Andrey Grodzovsky <[email protected]>
Quoted 108 lines
--- drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 3 +-- drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 5 ++--- drivers/gpu/drm/amd/scheduler/gpu_scheduler.c | 1 + drivers/gpu/drm/amd/scheduler/gpu_scheduler.h | 32 ++++++++++++--------------- 4 files changed, 18 insertions(+), 23 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +1
-2@@ -1192,8 +1192,7 @@ static int amdgpu_cs_submit(struct amdgpu_cs_parser *p, job->uf_sequence = seq;amdgpu_job_free_resources (job); -amdgpu_ring_priority_get (job->ring, -amd_sched_get_job_priority (&job->base)); +amdgpu_ring_priority_get (job->ring, job->base.s_priority);trace_amdgpu_cs_ioctl (job);amd_sched_entity_push_job (&job->base);drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +2
-3@@ -104,7 +104,7 @@ static void amdgpu_job_free_cb(struct amd_sched_job *s_job) { struct amdgpu_job *job = container_of(s_job, struct amdgpu_job, base); -amdgpu_ring_priority_put (job->ring,amd_sched_get_job_priority (s_job)); +amdgpu_ring_priority_put (job->ring, s_job->s_priority);dma_fence_put (job->fence);amdgpu_sync_free (&job->sync);amdgpu_sync_free (&job->dep_sync);@@ -141,8 +141,7 @@ int amdgpu_job_submit(struct amdgpu_job *job, struct amdgpu_ring *ring, job->fence_ctx = entity->fence_context; *f =dma_fence_get (&job->base.s_fence->finished);amdgpu_job_free_resources (job); -amdgpu_ring_priority_get (job->ring, -amd_sched_get_job_priority (&job->base)); +amdgpu_ring_priority_get (job->ring, job->base.s_priority);amd_sched_entity_push_job (&job->base);return 0 ;drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +1
-0@@ -529,6 +529,7 @@ int amd_sched_job_init(struct amd_sched_job *job, { job->sched = sched; job->s_entity = entity; + job->s_priority = entity->rq - sched->sched_rq; job->s_fence =amd_sched_fence_create (entity, owner);if (!job->s_fence)return -ENOMEM ;drivers/gpu/drm/amd/scheduler/gpu_scheduler.h +14
-12@@ -30,6 +30,19 @@struct amd_gpu_scheduler ;struct amd_sched_rq ; +enum amd_sched_priority { +AMD_SCHED_PRIORITY_MIN , +AMD_SCHED_PRIORITY_LOW =AMD_SCHED_PRIORITY_MIN , +AMD_SCHED_PRIORITY_NORMAL , +AMD_SCHED_PRIORITY_HIGH_SW , +AMD_SCHED_PRIORITY_HIGH_HW , +AMD_SCHED_PRIORITY_KERNEL , +AMD_SCHED_PRIORITY_MAX , +AMD_SCHED_PRIORITY_INVALID =-1 , +AMD_SCHED_PRIORITY_UNSET =-2 +}; + +/** * A scheduler entity is a wrapper around a job queue or a group * of other entities. Entities take turns emitting jobs from their @@ -83,6 +96,7 @@ struct amd_sched_job {struct delayed_work work_tdr;uint64_t id;atomic_t karma; +enum amd_sched_priority s_priority; };extern const struct dma_fence_ops amd_sched_fence_ops_scheduled;@@ -114,18 +128,6 @@ struct amd_sched_backend_ops {void (*free_job)(struct amd_sched_job *sched_job); }; -enum amd_sched_priority { -AMD_SCHED_PRIORITY_MIN , -AMD_SCHED_PRIORITY_LOW =AMD_SCHED_PRIORITY_MIN , -AMD_SCHED_PRIORITY_NORMAL , -AMD_SCHED_PRIORITY_HIGH_SW , -AMD_SCHED_PRIORITY_HIGH_HW , -AMD_SCHED_PRIORITY_KERNEL , -AMD_SCHED_PRIORITY_MAX , -AMD_SCHED_PRIORITY_INVALID =-1 , -AMD_SCHED_PRIORITY_UNSET =-2 -}; -/** * One scheduler is implemented for each hardware ring */
drivers/gpu/drm/amd/scheduler/gpu_scheduler.h bool amd_sched_dependency_optimized(struct dma_fence* fence, · line 181
Quoted 10 lines
drivers/gpu/drm/amd/scheduler/gpu_scheduler.h +0
-6@@ -176,10 +178,4 @@ bool amd_sched_dependency_optimized(struct dma_fence* fence, struct amd_sched_entity *entity);void amd_sched_job_kickout (struct amd_sched_job *s_job); -static inline enum amd_sched_priority -amd_sched_get_job_priority (struct amd_sched_job *job) -{ -return (job->s_entity->rq - job->sched->sched_rq); -} -#endif
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx