drm/amd/amdgpu embed hw_fence into amdgpu_job
Subsystems (5)
PATCH 1/2 drm/amd/amdgpu embed hw_fence into amdgpu_job
From: Jack Zhang <[email protected]>
Why: Previously hw fence is alloced separately with job. It caused historical lifetime issues and corner cases. The ideal situation is to take fence to manage both job and fence's lifetime, and simplify the design of gpu-scheduler.
How:
We propose to embed hw_fence into amdgpu_job.
1. We cover the normal job submission by this method.
2. For ib_test, and submit without a parent job keep the legacy way to create a hw fence separately.
--- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c | 1 - drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 62 ++++++++++++++++----- drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 35 ++++++++---- drivers/gpu/drm/amd/amdgpu/amdgpu_job.h | 4 +- drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 2 +- 8 files changed, 80 insertions(+), 30 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c +0 -1
@@ -714,7 +714,6 @@ int amdgpu_amdkfd_submit_ib(struct kgd_dev *kgd, enum kgd_engine_type engine, ret =dma_fence_wait (f, false); err_ib_sched: -dma_fence_put (f);amdgpu_job_free (job); err:return ret;
drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c +1 -1
@@ -1414,7 +1414,7 @@ static void amdgpu_ib_preempt_mark_partial_job(struct amdgpu_ring *ring)continue ; } job =to_amdgpu_job (s_job); -if (preempted && job->fence == fence) +if (preempted && (&job->hw_fence) == fence)/* mark the job as preempted */ job->preemption_status |=AMDGPU_IB_PREEMPTED ; }
drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +49 -13
105 lines not on this page · raw message
drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c +1 -1
@@ -262,7 +262,7 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs, fence_flags | AMDGPU_FENCE_FLAG_64BIT); } - r =amdgpu_fence_emit (ring, f, fence_flags); + r =amdgpu_fence_emit (ring, f, job, fence_flags);if (r) {dev_err (adev->dev,"failed to emit fence (%d)\n" , r);if (job && job->vmid)
drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +24 -11
@@ -114,11 +114,16 @@ void amdgpu_job_free_resources(struct amdgpu_job *job) {struct amdgpu_ring *ring =to_amdgpu_ring (job->base.sched);struct dma_fence *f; +struct dma_fence *hw_fence;unsigned i; -/* use sched fence if available */ - f = job->base.s_fence ? &job->base.s_fence->finished : job->fence; +if (job->hw_fence.ops ==NULL ) + hw_fence = job->external_hw_fence; +else + hw_fence = &job->hw_fence; +/* use sched fence if available */ + f = job->base.s_fence ? &job->base.s_fence->finished : hw_fence;for (i =0 ; i < job->num_ibs; ++i)amdgpu_ib_free (ring->adev, &job->ibs[i], f); }
@@ -129,20 +134,27 @@ static void amdgpu_job_free_cb(struct drm_sched_job *s_job)drm_sched_job_cleanup (s_job); -dma_fence_put (job->fence);amdgpu_sync_free (&job->sync);amdgpu_sync_free (&job->sched_sync); -kfree (job); + +/* only put the hw fence if has embedded fence */ +if (job->hw_fence.ops !=NULL ) +dma_fence_put (&job->hw_fence); +else +kfree (job); }void amdgpu_job_free (struct amdgpu_job *job) {amdgpu_job_free_resources (job); - -dma_fence_put (job->fence);amdgpu_sync_free (&job->sync);amdgpu_sync_free (&job->sched_sync); -kfree (job); + +/* only put the hw fence if has embedded fence */ +if (job->hw_fence.ops !=NULL ) +dma_fence_put (&job->hw_fence); +else +kfree (job); } int amdgpu_job_submit(struct amdgpu_job *job, struct drm_sched_entity *entity,
@@ -171,11 +183,14 @@ int amdgpu_job_submit_direct(struct amdgpu_job *job, struct amdgpu_ring *ring, job->base.sched = &ring->sched; r =amdgpu_ib_schedule (ring, job->num_ibs, job->ibs,NULL , fence); - job->fence =dma_fence_get (*fence); +/* record external_hw_fence for direct submit */ + job->external_hw_fence =dma_fence_get (*fence);if (r)return r;amdgpu_job_free (job); +dma_fence_put (*fence); +return 0 ; }
@@ -233,10 +248,8 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job)if (r)DRM_ERROR ("Error scheduling IBs (%d)\n" , r); } -/* if gpu reset, hw fence will be replaced here */ -dma_fence_put (job->fence); - job->fence =dma_fence_get (fence); +dma_fence_get (fence);amdgpu_job_free_resources (job); fence = r ?ERR_PTR (r) : fence;
drivers/gpu/drm/amd/amdgpu/amdgpu_job.h +3 -1
@@ -46,7 +46,9 @@ struct amdgpu_job {struct amdgpu_sync sync;struct amdgpu_sync sched_sync;struct amdgpu_ib *ibs; -struct dma_fence *fence;/* the hw fence */ +struct dma_fence hw_fence; +struct amdgpu_ring *ring; +struct dma_fence *external_hw_fence;uint32_t preamble_status;uint32_t preemption_status;uint32_t num_ibs;
drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h +1 -1
@@ -118,7 +118,7 @@ int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring, unsigned irq_type);void amdgpu_fence_driver_suspend (struct amdgpu_device *adev);void amdgpu_fence_driver_resume (struct amdgpu_device *adev); -int amdgpu_fence_emit (struct amdgpu_ring *ring,struct dma_fence **fence, +int amdgpu_fence_emit (struct amdgpu_ring *ring,struct dma_fence **fence,struct amdgpu_job *job,unsigned flags);int amdgpu_fence_emit_polling (struct amdgpu_ring *ring,uint32_t *s,uint32_t timeout);
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +1 -1
@@ -1216,7 +1216,7 @@ int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job,amdgpu_gmc_emit_pasid_mapping (ring, job->vmid, job->pasid);if (vm_flush_needed || pasid_mapping_needed) { - r =amdgpu_fence_emit (ring, &fence,0 ); + r =amdgpu_fence_emit (ring, &fence,NULL ,0 );if (r)return r; }
Signature · 6 lines
2.25.1 _______________________________________________ amd-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/amd-gfx
PATCH 2/2 drm: add tdr support for embeded hw_fence
[Why]
After embeded hw_fence to amdgpu_job, we need to add tdr support for this feature.
[How]
1. Add a resubmit_flag for resubmit jobs.
2. Clear job fence from RCU and force complete vm flush fences in pre_asic_reset
3. skip dma_fence_get for resubmit jobs and add a dma_fence_put for guilty jobs.
--- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 12 +++++++++++- drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 16 +++++++++++----- drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 4 +++- drivers/gpu/drm/scheduler/sched_main.c | 1 + include/drm/gpu_scheduler.h | 1 + 5 files changed, 27 insertions(+), 7 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +11 -1
@@ -4382,7 +4382,7 @@ int amdgpu_device_mode1_reset(struct amdgpu_device *adev)int amdgpu_device_pre_asic_reset (struct amdgpu_device *adev,struct amdgpu_reset_context *reset_context) { -int i, r =0 ; +int i, j, r =0 ;struct amdgpu_job *job =NULL ;bool need_full_reset =test_bit (AMDGPU_NEED_FULL_RESET , &reset_context->flags);
@@ -4406,6 +4406,16 @@ int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev,if (!ring || !ring->sched.thread)continue ; +/*clear job fence from fence drv to avoid force_completion +*leave NULL and vm flush fence in fence drv */ +for (j =0 ; j <= ring->fence_drv.num_fences_mask; j ++) { +struct dma_fence *old,**ptr; + ptr = &ring->fence_drv.fences[j]; + old =rcu_dereference_protected (*ptr,1 ); + if (old && test_bit(DMA_FENCE_FLAG_USER_BITS, &old->flags))) { +RCU_INIT_POINTER (*ptr,NULL ); + } + }/* after all hw jobs are reset, hw fence is meaningless, so force_completion */ amdgpu_fence_driver_force_completion (ring); }
drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +11 -5
@@ -156,11 +156,17 @@ int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f, struct amd job->ring = ring; } - seq = ++ring->fence_drv.sync_seq; -dma_fence_init (fence, &amdgpu_fence_ops, - &ring->fence_drv.lock, - adev->fence_context + ring->idx, - seq); +if (job !=NULL && job->base.resubmit_flag ==1 ) { +/* reinit seq for resubmitted jobs */ + seq = ++ring->fence_drv.sync_seq; + fence->seqno = seq; + }else { + seq = ++ring->fence_drv.sync_seq; +dma_fence_init (fence, &amdgpu_fence_ops, + &ring->fence_drv.lock, + adev->fence_context + ring->idx, + seq); + }if (job !=NULL ) {/* mark this fence has a parent job */
drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +3 -1
@@ -241,6 +241,7 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job)dma_fence_set_error (finished, -ECANCELED );/* skip IB as well if VRAM lost */ if (finished->error <0 ) { +dma_fence_put (&job->hw_fence);DRM_INFO ("Skip scheduling IBs!\n" ); }else { r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs, job,
@@ -249,7 +250,8 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job)DRM_ERROR ("Error scheduling IBs (%d)\n" , r); } -dma_fence_get (fence); +if (!job->base.resubmit_flag) +dma_fence_get (fence);amdgpu_job_free_resources (job); fence = r ?ERR_PTR (r) : fence;
drivers/gpu/drm/scheduler/sched_main.c +1 -0
@@ -544,6 +544,7 @@ void drm_sched_resubmit_jobs_ext(struct drm_gpu_scheduler *sched, int max)dma_fence_set_error (&s_fence->finished, -ECANCELED );dma_fence_put (s_job->s_fence->parent); + s_job->resubmit_flag =1 ; fence = sched->ops->run_job (s_job); i++;
include/drm/gpu_scheduler.h +1 -0
@@ -198,6 +198,7 @@ struct drm_sched_job {enum drm_sched_priority s_priority;struct drm_sched_entity *entity;struct dma_fence_cb cb; +int resubmit_flag; }; static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job,
Signature · 6 lines
2.25.1 _______________________________________________ amd-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Re: [PATCH 2/2] drm: add tdr support for embeded hw_fence
Quoted 37 lines · Jingwen Chen
[Why]
After embededhw_fencetoamdgpu_job, we need to add tdr support for this feature.[How]
1. Add aresubmit_flagfor resubmit jobs.
2. Clear job fence from RCU and force complete vm flush fences inpre_asic_reset
3. skipdma_fence_getfor resubmit jobs and add adma_fence_putfor guilty jobs.Signed-off-by Jack Zhang <[email protected]>Signed-off-by Jingwen Chen <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 12 +++++++++++- drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 16 +++++++++++----- drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 4 +++- drivers/gpu/drm/scheduler/sched_main.c | 1 + include/drm/gpu_scheduler.h | 1 + 5 files changed, 27 insertions(+), 7 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1
-1@@ -4382,7 +4382,7 @@ int amdgpu_device_mode1_reset(struct amdgpu_device *adev)int amdgpu_device_pre_asic_reset (struct amdgpu_device *adev,struct amdgpu_reset_context *reset_context) { -int i, r =0 ; +int i, j, r =0 ;struct amdgpu_job *job =NULL ;bool need_full_reset =test_bit (AMDGPU_NEED_FULL_RESET , &reset_context->flags);@@ -4406,6 +4406,16 @@ int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev,if (!ring || !ring->sched.thread)continue ;
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev, · line 4417
Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +9
-0+/*clear job fence from fence drv to avoid force_completion +*leave NULL and vm flush fence in fence drv */ +for (j =0 ; j <= ring->fence_drv.num_fences_mask; j ++) { +struct dma_fence *old,**ptr; + ptr = &ring->fence_drv.fences[j]; + old =rcu_dereference_protected (*ptr,1 ); + if (old && test_bit(DMA_FENCE_FLAG_USER_BITS, &old->flags))) { +RCU_INIT_POINTER (*ptr,NULL ); + }
Is this to avoid premature job free because of dma_fence_put inside amdgpu_fence_process ? I can't currently remember why but we probably want all the HW fences currently in the ring to be forced signaled - maybe better to test for DMA_FENCE_FLAG_USER_BITS inside amdgpu_fence_process and still do the signaling but not the dma_fence_put part
Andrey
Quoted 13 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1
-0+ }/* after all hw jobs are reset, hw fence is meaningless, so force_completion */ amdgpu_fence_driver_force_completion (ring); }drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +0
-1@@ -156,11 +156,17 @@ int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f, struct amd job->ring = ring; } - seq = ++ring->fence_drv.sync_seq;
drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f, struct amd · line 164
Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +6
-4-dma_fence_init (fence, &amdgpu_fence_ops, - &ring->fence_drv.lock, - adev->fence_context + ring->idx, - seq); +if (job !=NULL && job->base.resubmit_flag ==1 ) { +/* reinit seq for resubmitted jobs */ + seq = ++ring->fence_drv.sync_seq; + fence->seqno = seq; + }else { + seq = ++ring->fence_drv.sync_seq;
Seems like you could do the above line only once above if-else as it was before
Quoted 47 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +5
-0+dma_fence_init (fence, &amdgpu_fence_ops, + &ring->fence_drv.lock, + adev->fence_context + ring->idx, + seq); + }if (job !=NULL ) {/* mark this fence has a parent job */ drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +3
-1@@ -241,6 +241,7 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job)dma_fence_set_error (finished, -ECANCELED );/* skip IB as well if VRAM lost */ if (finished->error <0 ) { +dma_fence_put (&job->hw_fence);DRM_INFO ("Skip scheduling IBs!\n" ); }else { r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs, job,@@ -249,7 +250,8 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job)DRM_ERROR ("Error scheduling IBs (%d)\n" , r); } -dma_fence_get (fence); +if (!job->base.resubmit_flag) +dma_fence_get (fence);amdgpu_job_free_resources (job); fence = r ?ERR_PTR (r) : fence;drivers/gpu/drm/scheduler/sched_main.c +1
-0@@ -544,6 +544,7 @@ void drm_sched_resubmit_jobs_ext(struct drm_gpu_scheduler *sched, int max)dma_fence_set_error (&s_fence->finished, -ECANCELED );dma_fence_put (s_job->s_fence->parent); + s_job->resubmit_flag =1 ; fence = sched->ops->run_job (s_job); i++;include/drm/gpu_scheduler.h +0
-0
include/drm/gpu_scheduler.h struct drm_sched_job { · line 204
Quoted 7 lines
include/drm/gpu_scheduler.h +1
-0@@ -198,6 +198,7 @@ struct drm_sched_job {enum drm_sched_priority s_priority;struct drm_sched_entity *entity;struct dma_fence_cb cb; +int resubmit_flag; }; static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job,
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Re: [PATCH 2/2] drm: add tdr support for embeded hw_fence
Quoted 58 lines · Andrey Grodzovsky
Quoted 36 lines · Jingwen Chen
[Why]
After embededhw_fencetoamdgpu_job, we need to add tdr support for this feature.[How]
1. Add aresubmit_flagfor resubmit jobs.
2. Clear job fence from RCU and force complete vm flush fences inpre_asic_reset
3. skipdma_fence_getfor resubmit jobs and add adma_fence_putfor guilty jobs.Signed-off-by Jack Zhang <[email protected]>Signed-off-by Jingwen Chen <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 12 +++++++++++- drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 16 +++++++++++----- drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 4 +++- drivers/gpu/drm/scheduler/sched_main.c | 1 + include/drm/gpu_scheduler.h | 1 + 5 files changed, 27 insertions(+), 7 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1
-1@@ -4382,7 +4382,7 @@ int amdgpu_device_mode1_reset(struct amdgpu_device *adev)int amdgpu_device_pre_asic_reset (struct amdgpu_device *adev,struct amdgpu_reset_context *reset_context) { -int i, r =0 ; +int i, j, r =0 ;struct amdgpu_job *job =NULL ;bool need_full_reset =test_bit (AMDGPU_NEED_FULL_RESET , &reset_context->flags);@@ -4406,6 +4406,16 @@ int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev,if (!ring || !ring->sched.thread)drivers/gpu/drm/amd/amdgpu/amdgpu_device.c int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev, · line 4416
Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +9
-0continue ; +/*clear job fence from fence drv to avoid force_completion +*leave NULL and vm flush fence in fence drv */ +for (j =0 ; j <= ring->fence_drv.num_fences_mask; j ++) { +struct dma_fence *old,**ptr; + ptr = &ring->fence_drv.fences[j]; + old =rcu_dereference_protected (*ptr,1 ); + if (old && test_bit(DMA_FENCE_FLAG_USER_BITS, &old->flags))) { +RCU_INIT_POINTER (*ptr,NULL ); + }Is this to avoid premature job free because of
dma_fence_putinsideamdgpu_fence_process?
I can't currently remember why but we probably want all the HW fences currently in the ring to
be forced signaled - maybe better to test forDMA_FENCE_FLAG_USER_BITSinsideamdgpu_fence_process
and still do the signaling but not thedma_fence_putpart
Quoted 1 line
Andrey
Hi Andrey,
This is to avoid signaling the same fence twice. If we still do the signaling, then the job in the pending list will be signaled first in force_completion, and later be signaled in resubmit. This will go to BUG() in amdgpu_fence_process.
Quoted 19 lines
Quoted 19 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1
-0+ }/* after all hw jobs are reset, hw fence is meaningless, so force_completion */ amdgpu_fence_driver_force_completion (ring); }drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +2
-5@@ -156,11 +156,17 @@ int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f, struct amd job->ring = ring; } - seq = ++ring->fence_drv.sync_seq; -dma_fence_init (fence, &amdgpu_fence_ops, - &ring->fence_drv.lock, - adev->fence_context + ring->idx, - seq); +if (job !=NULL && job->base.resubmit_flag ==1 ) { +/* reinit seq for resubmitted jobs */
Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f, struct amd · line 164
Quoted 4 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +4
-0+ seq = ++ring->fence_drv.sync_seq; + fence->seqno = seq; + }else { + seq = ++ring->fence_drv.sync_seq;Seems like you could do the above line only once above if-else as it was before
Sure, I will modify this.
Best Regards,
JingWen Chen
Quoted 40 lines
Quoted 40 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +5
-0+dma_fence_init (fence, &amdgpu_fence_ops, + &ring->fence_drv.lock, + adev->fence_context + ring->idx, + seq); + }if (job !=NULL ) {/* mark this fence has a parent job */ drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +3
-1@@ -241,6 +241,7 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job)dma_fence_set_error (finished, -ECANCELED );/* skip IB as well if VRAM lost */ if (finished->error <0 ) { +dma_fence_put (&job->hw_fence);DRM_INFO ("Skip scheduling IBs!\n" ); }else { r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs, job,@@ -249,7 +250,8 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job)DRM_ERROR ("Error scheduling IBs (%d)\n" , r); } -dma_fence_get (fence); +if (!job->base.resubmit_flag) +dma_fence_get (fence);amdgpu_job_free_resources (job); fence = r ?ERR_PTR (r) : fence;drivers/gpu/drm/scheduler/sched_main.c +1
-0@@ -544,6 +544,7 @@ void drm_sched_resubmit_jobs_ext(struct drm_gpu_scheduler *sched, int max)dma_fence_set_error (&s_fence->finished, -ECANCELED );dma_fence_put (s_job->s_fence->parent); + s_job->resubmit_flag =1 ; fence = sched->ops->run_job (s_job); i++;include/drm/gpu_scheduler.h +0
-1--- a/include/drm/gpu_scheduler.h
Quoted 8 lines
include/drm/gpu_scheduler.h struct drm_sched_job { · line 203
Quoted 8 lines
+++
b/include/drm/gpu_scheduler.hinclude/drm/gpu_scheduler.h +1
-0@@ -198,6 +198,7 @@ struct drm_sched_job {enum drm_sched_priority s_priority;struct drm_sched_entity *entity;struct dma_fence_cb cb; +int resubmit_flag; }; static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job,
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Re: [PATCH 2/2] drm: add tdr support for embeded hw_fence
Quoted 57 lines · Jingwen Chen
Quoted 56 lines · Andrey Grodzovsky
Quoted 36 lines · Jingwen Chen
[Why]
After embededhw_fencetoamdgpu_job, we need to add tdr support for this feature.[How]
1. Add aresubmit_flagfor resubmit jobs.
2. Clear job fence from RCU and force complete vm flush fences inpre_asic_reset
3. skipdma_fence_getfor resubmit jobs and add adma_fence_putfor guilty jobs.Signed-off-by Jack Zhang <[email protected]>Signed-off-by Jingwen Chen <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 12 +++++++++++- drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 16 +++++++++++----- drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 4 +++- drivers/gpu/drm/scheduler/sched_main.c | 1 + include/drm/gpu_scheduler.h | 1 + 5 files changed, 27 insertions(+), 7 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1
-1@@ -4382,7 +4382,7 @@ int amdgpu_device_mode1_reset(struct amdgpu_device *adev)int amdgpu_device_pre_asic_reset (struct amdgpu_device *adev,struct amdgpu_reset_context *reset_context) { -int i, r =0 ; +int i, j, r =0 ;struct amdgpu_job *job =NULL ;bool need_full_reset =test_bit (AMDGPU_NEED_FULL_RESET , &reset_context->flags);@@ -4406,6 +4406,16 @@ int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev,if (!ring || !ring->sched.thread)drivers/gpu/drm/amd/amdgpu/amdgpu_device.c int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev, · line 4416
Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +9
-0continue ; +/*clear job fence from fence drv to avoid force_completion +*leave NULL and vm flush fence in fence drv */ +for (j =0 ; j <= ring->fence_drv.num_fences_mask; j ++) { +struct dma_fence *old,**ptr; + ptr = &ring->fence_drv.fences[j]; + old =rcu_dereference_protected (*ptr,1 ); + if (old && test_bit(DMA_FENCE_FLAG_USER_BITS, &old->flags))) { +RCU_INIT_POINTER (*ptr,NULL ); + }Is this to avoid premature job free because of
dma_fence_putinsideamdgpu_fence_process?
I can't currently remember why but we probably want all the HW fences currently in the ring to
be forced signaled - maybe better to test forDMA_FENCE_FLAG_USER_BITSinsideamdgpu_fence_process
and still do the signaling but not thedma_fence_putpart
Quoted 7 lines
Quoted 1 line
Andrey
Hi Andrey,
This is to avoid signaling the same fence twice. If we still do the signaling, then the job in the pending list will be signaled first in
force_completion, and later be signaled in resubmit. This will go toBUG()inamdgpu_fence_process.
Oh, i see, how about just adding 'skip' flag to amdgpu_ring and setting it before calling amdgpu_fence_driver_force_completion and resetting it after, then inside amdgpu_fence_driver_force_completion you can just skip the signaling part with this flag for fences with DMA_FENCE_FLAG_USER_BITS set Less lines of code at least.
Andrey
Quoted 70 lines
Quoted 18 lines
Quoted 18 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1
-0+ }/* after all hw jobs are reset, hw fence is meaningless, so force_completion */ amdgpu_fence_driver_force_completion (ring); }drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +1
-5@@ -156,11 +156,17 @@ int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f, struct amd job->ring = ring; } - seq = ++ring->fence_drv.sync_seq; -dma_fence_init (fence, &amdgpu_fence_ops, - &ring->fence_drv.lock, - adev->fence_context + ring->idx, - seq); +if (job !=NULL && job->base.resubmit_flag ==1 ) {Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f, struct amd · line 164
Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +5
-0+/* reinit seq for resubmitted jobs */ + seq = ++ring->fence_drv.sync_seq; + fence->seqno = seq; + }else { + seq = ++ring->fence_drv.sync_seq;Seems like you could do the above line only once above if-else as it was before
Sure, I will modify this.
Best Regards,
JingWen ChenQuoted 39 lines
Quoted 39 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +5
-0+dma_fence_init (fence, &amdgpu_fence_ops, + &ring->fence_drv.lock, + adev->fence_context + ring->idx, + seq); + }if (job !=NULL ) {/* mark this fence has a parent job */ drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +3
-1@@ -241,6 +241,7 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job)dma_fence_set_error (finished, -ECANCELED );/* skip IB as well if VRAM lost */ if (finished->error <0 ) { +dma_fence_put (&job->hw_fence);DRM_INFO ("Skip scheduling IBs!\n" ); }else { r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs, job,@@ -249,7 +250,8 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job)DRM_ERROR ("Error scheduling IBs (%d)\n" , r); } -dma_fence_get (fence); +if (!job->base.resubmit_flag) +dma_fence_get (fence);amdgpu_job_free_resources (job); fence = r ?ERR_PTR (r) : fence;drivers/gpu/drm/scheduler/sched_main.c +1
-0@@ -544,6 +544,7 @@ void drm_sched_resubmit_jobs_ext(struct drm_gpu_scheduler *sched, int max)dma_fence_set_error (&s_fence->finished, -ECANCELED );dma_fence_put (s_job->s_fence->parent); + s_job->resubmit_flag =1 ; fence = sched->ops->run_job (s_job); i++;include/drm/gpu_scheduler.h +0
-1--- a/include/drm/gpu_scheduler.h
Quoted 8 lines
Quoted 8 lines
include/drm/gpu_scheduler.h struct drm_sched_job { · line 203
Quoted 8 lines
+++
b/include/drm/gpu_scheduler.hinclude/drm/gpu_scheduler.h +1
-0@@ -198,6 +198,7 @@ struct drm_sched_job {enum drm_sched_priority s_priority;struct drm_sched_entity *entity;struct dma_fence_cb cb; +int resubmit_flag; }; static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job,
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Re: [PATCH 2/2] drm: add tdr support for embeded hw_fence
Quoted 76 lines · Andrey Grodzovsky
Quoted 57 lines · Jingwen Chen
Quoted 56 lines · Andrey Grodzovsky
Quoted 36 lines · Jingwen Chen
[Why]
After embededhw_fencetoamdgpu_job, we need to add tdr support for this feature.[How]
1. Add aresubmit_flagfor resubmit jobs.
2. Clear job fence from RCU and force complete vm flush fences inpre_asic_reset
3. skipdma_fence_getfor resubmit jobs and add adma_fence_putfor guilty jobs.Signed-off-by Jack Zhang <[email protected]>Signed-off-by Jingwen Chen <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 12 +++++++++++- drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 16 +++++++++++----- drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 4 +++- drivers/gpu/drm/scheduler/sched_main.c | 1 + include/drm/gpu_scheduler.h | 1 + 5 files changed, 27 insertions(+), 7 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1
-1@@ -4382,7 +4382,7 @@ int amdgpu_device_mode1_reset(struct amdgpu_device *adev)int amdgpu_device_pre_asic_reset (struct amdgpu_device *adev,struct amdgpu_reset_context *reset_context) { -int i, r =0 ; +int i, j, r =0 ;struct amdgpu_job *job =NULL ;bool need_full_reset =test_bit (AMDGPU_NEED_FULL_RESET , &reset_context->flags);@@ -4406,6 +4406,16 @@ int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev,if (!ring || !ring->sched.thread)drivers/gpu/drm/amd/amdgpu/amdgpu_device.c int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev, · line 4416
Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +9
-0continue ; +/*clear job fence from fence drv to avoid force_completion +*leave NULL and vm flush fence in fence drv */ +for (j =0 ; j <= ring->fence_drv.num_fences_mask; j ++) { +struct dma_fence *old,**ptr; + ptr = &ring->fence_drv.fences[j]; + old =rcu_dereference_protected (*ptr,1 ); + if (old && test_bit(DMA_FENCE_FLAG_USER_BITS, &old->flags))) { +RCU_INIT_POINTER (*ptr,NULL ); + }Is this to avoid premature job free because of
dma_fence_putinsideamdgpu_fence_process?
I can't currently remember why but we probably want all the HW fences currently in the ring to
be forced signaled - maybe better to test forDMA_FENCE_FLAG_USER_BITSinsideamdgpu_fence_process
and still do the signaling but not thedma_fence_putpartQuoted 7 lines
Quoted 1 line
Andrey
Hi Andrey,
This is to avoid signaling the same fence twice. If we still do the signaling, then the job in the pending list will be signaled first in
force_completion, and later be signaled in resubmit. This will go toBUG()inamdgpu_fence_process.Oh, i see, how about just adding 'skip' flag to
amdgpu_ringand setting it before callingamdgpu_fence_driver_force_completionand resetting it after, then insideamdgpu_fence_driver_force_completion
you can just skip the signaling part with this flag for fences withDMA_FENCE_FLAG_USER_BITSset
Less lines of code at least.
Quoted 1 line
Andrey
Hi Andrey,
In this way, this issue still exists. If we just skip it in the force_completion, these fences are still in the RCU fence array. So when the FLR finishes, the eop interrupt function of ib_test will call amdgpu_fence_process, the skipped fences will be signaled along with ib_test fences and signaled again during resubmission.
Best Regards,
JingWen Chen
Quoted 71 lines
Quoted 71 lines
Quoted 19 lines
Quoted 19 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1
-0+ }/* after all hw jobs are reset, hw fence is meaningless, so force_completion */ amdgpu_fence_driver_force_completion (ring); }drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +1
-5@@ -156,11 +156,17 @@ int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f, struct amd job->ring = ring; } - seq = ++ring->fence_drv.sync_seq; -dma_fence_init (fence, &amdgpu_fence_ops, - &ring->fence_drv.lock, - adev->fence_context + ring->idx, - seq); +if (job !=NULL && job->base.resubmit_flag ==1 ) {Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f, struct amd · line 164
Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +5
-0+/* reinit seq for resubmitted jobs */ + seq = ++ring->fence_drv.sync_seq; + fence->seqno = seq; + }else { + seq = ++ring->fence_drv.sync_seq;Seems like you could do the above line only once above if-else as it was before
Sure, I will modify this.
Best Regards,
JingWen ChenQuoted 39 lines
Quoted 39 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +5
-0+dma_fence_init (fence, &amdgpu_fence_ops, + &ring->fence_drv.lock, + adev->fence_context + ring->idx, + seq); + }if (job !=NULL ) {/* mark this fence has a parent job */ drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +3
-1@@ -241,6 +241,7 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job)dma_fence_set_error (finished, -ECANCELED );/* skip IB as well if VRAM lost */ if (finished->error <0 ) { +dma_fence_put (&job->hw_fence);DRM_INFO ("Skip scheduling IBs!\n" ); }else { r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs, job,@@ -249,7 +250,8 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job)DRM_ERROR ("Error scheduling IBs (%d)\n" , r); } -dma_fence_get (fence); +if (!job->base.resubmit_flag) +dma_fence_get (fence);amdgpu_job_free_resources (job); fence = r ?ERR_PTR (r) : fence;drivers/gpu/drm/scheduler/sched_main.c +1
-0@@ -544,6 +544,7 @@ void drm_sched_resubmit_jobs_ext(struct drm_gpu_scheduler *sched, int max)dma_fence_set_error (&s_fence->finished, -ECANCELED );dma_fence_put (s_job->s_fence->parent); + s_job->resubmit_flag =1 ; fence = sched->ops->run_job (s_job); i++;include/drm/gpu_scheduler.h +0
-1--- a/include/drm/gpu_scheduler.h
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
include/drm/gpu_scheduler.h struct drm_sched_job { · line 203
Quoted 8 lines
+++
b/include/drm/gpu_scheduler.hinclude/drm/gpu_scheduler.h +1
-0@@ -198,6 +198,7 @@ struct drm_sched_job {enum drm_sched_priority s_priority;struct drm_sched_entity *entity;struct dma_fence_cb cb; +int resubmit_flag; }; static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job,
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Christian König
reviewerRe: [PATCH 2/2] drm: add tdr support for embeded hw_fence
In reply to Andrey GrodzovskyAm 22.07.21 um 16:45 schrieb Andrey Grodzovsky:
Quoted 71 lines · Jingwen Chen
Quoted 60 lines · Andrey Grodzovsky
Quoted 59 lines · Jingwen Chen
Quoted 39 lines
[Why]
After embededhw_fencetoamdgpu_job, we need to add tdr support for this feature.[How]
1. Add aresubmit_flagfor resubmit jobs.
2. Clear job fence from RCU and force complete vm flush fences inpre_asic_reset
3. skipdma_fence_getfor resubmit jobs and add adma_fence_putfor guilty jobs.Signed-off-by Jack Zhang <[email protected]>Signed-off-by Jingwen Chen <[email protected]>---
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c| 12 +++++++++++- drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 16 +++++++++++----- drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 4 +++- drivers/gpu/drm/scheduler/sched_main.c | 1 + include/drm/gpu_scheduler.h | 1 +5 files changed, 27 insertions(+), 7 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1
-1@@ -4382,7 +4382,7 @@ int amdgpu_device_mode1_reset(struct amdgpu_device *adev) int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev, struct amdgpu_reset_context *reset_context) { - int i, r = 0; + int i, j, r = 0; struct amdgpu_job *job = NULL;bool
need_full_reset=test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0
-0@@ -4406,6 +4406,16 @@ int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev, if (!ring || !ring->sched.thread) continue;drivers/gpu/drm/amd/amdgpu/amdgpu_device.c int amdgpu_device_pre_asic_reset(struct · line 4418
Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +9
-0+ /*clear job fence from fence drv to avoid force_completion + *leave NULL and vm flush fence in fence drv */ + for (j = 0; j <= ring->fence_drv.num_fences_mask; j ++) { + struct dma_fence *old,**ptr; + ptr = &ring->fence_drv.fences[j]; + old = rcu_dereference_protected(*ptr, 1); + if (old && test_bit(DMA_FENCE_FLAG_USER_BITS, &old->flags))) { + RCU_INIT_POINTER(*ptr, NULL); + }Is this to avoid premature job free because of
dma_fence_putinsideamdgpu_fence_process?
I can't currently remember why but we probably want all the HW fences currently in the ring to
be forced signaled - maybe better to test forDMA_FENCE_FLAG_USER_BITSinsideamdgpu_fence_process
and still do the signaling but not thedma_fence_putpartQuoted 7 lines
Quoted 1 line
Andrey
Hi Andrey,
This is to avoid signaling the same fence twice. If we still do the signaling, then the job in the pending list will be signaled first in
force_completion, and later be signaled in resubmit. This will go toBUG()inamdgpu_fence_process.
Quoted 4 lines
Oh, i see, how about just adding 'skip' flag to
amdgpu_ringand setting it before callingamdgpu_fence_driver_force_completionand resetting it after, then insideamdgpu_fence_driver_force_completionyou can just skip the signaling part with this flag for fences withDMA_FENCE_FLAG_USER_BITSset Less lines of code at least.
Still sounds quite a bit hacky.
I would rather suggest to completely drop the approach with amdgpu_fence_driver_force_completion(). I could never see why we would want that in the first place.
Regards,
Christian.
Quoted 78 lines
Andrey
Quoted 74 lines
Quoted 19 lines
Quoted 19 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1
-0+ } /* after all hw jobs are reset, hw fence is meaningless, so force_completion */ amdgpu_fence_driver_force_completion(ring);}
drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +1
-5@@ -156,11 +156,17 @@ int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f, struct amd job->ring = ring; } - seq = ++ring->fence_drv.sync_seq; - dma_fence_init(fence, &amdgpu_fence_ops, - &ring->fence_drv.lock, - adev->fence_context + ring->idx, - seq); + if (job != NULL && job->base.resubmit_flag == 1) {Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f, struct amd · line 164
Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +5
-0+ /* reinit seq for resubmitted jobs */ + seq = ++ring->fence_drv.sync_seq; + fence->seqno = seq; + } else { + seq = ++ring->fence_drv.sync_seq;Seems like you could do the above line only once above if-else as it was before
Sure, I will modify this.
Best Regards,
JingWen ChenQuoted 42 lines
Quoted 42 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +5
-0+ dma_fence_init(fence, &amdgpu_fence_ops, + &ring->fence_drv.lock, + adev->fence_context + ring->idx, + seq); + } if (job != NULL) { /* mark this fence has a parent job */drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +0
-0@@ -241,6 +241,7 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job)
dma_fence_set_error(finished, -ECANCELED);/* skip IB as well if VRAM lost */ if (finished->error< 0) {
+dma_fence_put(&job->hw_fence);
DRM_INFO("Skip scheduling IBs!\n");
} else {
r = amdgpu_ib_schedule(ring,job->num_ibs,job->ibs, job,drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +2
-1@@ -249,7 +250,8 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job) DRM_ERROR("Error scheduling IBs (%d)\n", r); } - dma_fence_get(fence); + if (!job->base.resubmit_flag) + dma_fence_get(fence); amdgpu_job_free_resources(job); fence = r ? ERR_PTR(r) : fence;drivers/gpu/drm/scheduler/sched_main.c +1
-0@@ -544,6 +544,7 @@ void drm_sched_resubmit_jobs_ext(struct drm_gpu_scheduler *sched, int max) dma_fence_set_error(&s_fence->finished, -ECANCELED); dma_fence_put(s_job->s_fence->parent); + s_job->resubmit_flag = 1; fence = sched->ops->run_job(s_job); i++;include/drm/gpu_scheduler.h +0
-1--- a/include/drm/gpu_scheduler.h
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
include/drm/gpu_scheduler.h struct drm_sched_job { · line 203
Quoted 8 lines
+++
b/include/drm/gpu_scheduler.hinclude/drm/gpu_scheduler.h +1
-0@@ -198,6 +198,7 @@ struct drm_sched_job { enum drm_sched_priority s_priority; struct drm_sched_entity *entity; struct dma_fence_cb cb; + int resubmit_flag; }; static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job,
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Re: [PATCH 2/2] drm: add tdr support for embeded hw_fence
Quoted 81 lines · Christian König
Am 22.07.21 um 16:45 schrieb Andrey Grodzovsky:
Quoted 72 lines · Jingwen Chen
Quoted 61 lines · Andrey Grodzovsky
Quoted 60 lines · Jingwen Chen
Quoted 40 lines
[Why]
After embededhw_fencetoamdgpu_job, we need to add tdr support for this feature.[How]
1. Add aresubmit_flagfor resubmit jobs.
2. Clear job fence from RCU and force complete vm flush fences inpre_asic_reset
3. skipdma_fence_getfor resubmit jobs and add adma_fence_putfor guilty jobs.Signed-off-by Jack Zhang <[email protected]>Signed-off-by Jingwen Chen <[email protected]>---
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c| 12 +++++++++++- drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 16 +++++++++++----- drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 4 +++- drivers/gpu/drm/scheduler/sched_main.c | 1 + include/drm/gpu_scheduler.h | 1 +5 files changed, 27 insertions(+), 7 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1
-1@@ -4382,7 +4382,7 @@ int amdgpu_device_mode1_reset(struct amdgpu_device *adev) int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev, struct amdgpu_reset_context *reset_context) { - int i, r = 0; + int i, j, r = 0; struct amdgpu_job *job = NULL;bool
need_full_reset=test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0
-0@@ -4406,6 +4406,16 @@ int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev, if (!ring || !ring->sched.thread) continue;drivers/gpu/drm/amd/amdgpu/amdgpu_device.c int · line 4418
Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +9
-0+ /*clear job fence from fence drv to avoid force_completion + *leave NULL and vm flush fence in fence drv */ + for (j = 0; j <= ring->fence_drv.num_fences_mask; j ++) { + struct dma_fence *old,**ptr; + ptr = &ring->fence_drv.fences[j]; + old = rcu_dereference_protected(*ptr, 1); + if (old && test_bit(DMA_FENCE_FLAG_USER_BITS, &old->flags))) { + RCU_INIT_POINTER(*ptr, NULL); + }Is this to avoid premature job free because of
dma_fence_putinsideamdgpu_fence_process?
I can't currently remember why but we probably want all the HW fences currently in the ring to
be forced signaled - maybe better to test forDMA_FENCE_FLAG_USER_BITSinsideamdgpu_fence_process
and still do the signaling but not thedma_fence_putpartQuoted 7 lines
Quoted 1 line
Andrey
Hi Andrey,
This is to avoid signaling the same fence twice. If we still do the signaling, then the job in the pending list will be signaled first in
force_completion, and later be signaled in resubmit. This will go toBUG()inamdgpu_fence_process.Quoted 7 lines
Oh, i see, how about just adding 'skip' flag to
amdgpu_ringand setting it before callingamdgpu_fence_driver_force_completionand resetting it after, then insideamdgpu_fence_driver_force_completion
you can just skip the signaling part with this flag for fences withDMA_FENCE_FLAG_USER_BITSset
Less lines of code at least.
Quoted 9 lines
Still sounds quite a bit hacky.
I would rather suggest to completely drop the approach with
amdgpu_fence_driver_force_completion(). I could never see why we would want that in the first place.Regards,
Christian.
Hi Christian,
I keep the amdgpu_fence_driver_force_completion here to make sure the vm flush fence is signaled and put.
So the key question is whether the fence of ib test should be the first fence to be signaled after reset.
If it should be, it means not only fences with DMA_FENCE_FLAG_USER_BITS but also vm flush fences should be removed from RCU fence array before ib_test. Then we must do the force_completion here for vm flush fence, otherwise there will be a memory leak here as no one will signal and put it after we remove it from RCU.
If not, then the first fence to signal could be vm flush fence. And I'm OK with drop amdgpu_fence_driver_force_completion here.
Best Regards,
JingWen Chen
Quoted 85 lines
Quoted 85 lines
Andrey
Quoted 81 lines
Quoted 22 lines
Quoted 22 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1
-0+ } /* after all hw jobs are reset, hw fence ismeaningless, so
force_completion*/amdgpu_fence_driver_force_completion(ring); }drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +0
-0@@ -156,11 +156,17 @@ int amdgpu_fence_emit(struct
amdgpu_ring*ring, structdma_fence**f, struct amdjob->ring= ring;
}
- seq = ++ring->fence_drv.sync_seq;
- dma_fence_init(fence,&amdgpu_fence_ops, -&ring->fence_drv.lock,
-adev->fence_context+ring->idx, - seq);
+ if (job != NULL &&job->base.resubmit_flag== 1) { + /* reinit seq for resubmitted jobs */Quoted 8 lines
Quoted 4 lines
+ seq = ++ring->fence_drv.sync_seq;
+fence->seqno= seq;
+ } else {
+ seq = ++ring->fence_drv.sync_seq;Seems like you could do the above line only once above if-else as it was
beforeSure, I will modify this.
Best Regards,
JingWen ChenQuoted 46 lines
Quoted 46 lines
+ dma_fence_init(fence,
&amdgpu_fence_ops,
+&ring->fence_drv.lock,
+adev->fence_context+ring->idx,
+ seq);
+ }
if (job != NULL) {
/* mark this fence has a parent job */drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +0
-0@@ -241,6 +241,7 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job)
dma_fence_set_error(finished, -ECANCELED);/* skip IB as well if VRAM lost */
if (finished->error< 0) {
+dma_fence_put(&job->hw_fence);
DRM_INFO("Skip scheduling IBs!\n");
} else {
r = amdgpu_ib_schedule(ring,job->num_ibs,job->ibs, job,drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +2
-1@@ -249,7 +250,8 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job) DRM_ERROR("Error scheduling IBs (%d)\n", r); } - dma_fence_get(fence); + if (!job->base.resubmit_flag) + dma_fence_get(fence); amdgpu_job_free_resources(job); fence = r ? ERR_PTR(r) : fence;drivers/gpu/drm/scheduler/sched_main.c +1
-0@@ -544,6 +544,7 @@ void drm_sched_resubmit_jobs_ext(struct drm_gpu_scheduler *sched, int max) dma_fence_set_error(&s_fence->finished, -ECANCELED); dma_fence_put(s_job->s_fence->parent); + s_job->resubmit_flag = 1; fence = sched->ops->run_job(s_job); i++;include/drm/gpu_scheduler.h +0
-0
Quoted 9 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
include/drm/gpu_scheduler.h +1
-0@@ -198,6 +198,7 @@ struct drm_sched_job { enum drm_sched_priority s_priority; struct drm_sched_entity *entity; struct dma_fence_cb cb; + int resubmit_flag; };static inline bool drm_sched_invalidate_job(struct
drm_sched_job*s_job,
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Re: [PATCH 2/2] drm: add tdr support for embeded hw_fence
Quoted 84 lines · Jingwen Chen
Quoted 83 lines · Christian König
Am 22.07.21 um 16:45 schrieb Andrey Grodzovsky:
Quoted 69 lines · Jingwen Chen
Quoted 60 lines · Andrey Grodzovsky
Quoted 59 lines · Jingwen Chen
Quoted 40 lines
[Why]
After embededhw_fencetoamdgpu_job, we need to add tdr support for this feature.[How]
1. Add aresubmit_flagfor resubmit jobs.
2. Clear job fence from RCU and force complete vm flush fences inpre_asic_reset
3. skipdma_fence_getfor resubmit jobs and add adma_fence_putfor guilty jobs.Signed-off-by Jack Zhang <[email protected]>Signed-off-by Jingwen Chen <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 12 +++++++++++- drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 16 +++++++++++----- drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 4 +++- drivers/gpu/drm/scheduler/sched_main.c | 1 + include/drm/gpu_scheduler.h | 1 + 5 files changed, 27 insertions(+), 7 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1
-1@@ -4382,7 +4382,7 @@ int amdgpu_device_mode1_reset(struct amdgpu_device *adev) int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev, struct amdgpu_reset_context *reset_context) { - int i, r = 0; + int i, j, r = 0; struct amdgpu_job *job = NULL; bool need_full_reset =
test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0
-0@@ -4406,6 +4406,16 @@ int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev, if (!ring || !ring->sched.thread) continue;drivers/gpu/drm/amd/amdgpu/amdgpu_device.c int · line 4418
Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +9
-0+ /*clear job fence from fence drv to avoid force_completion + *leave NULL and vm flush fence in fence drv */ + for (j = 0; j <= ring->fence_drv.num_fences_mask; j ++) { + struct dma_fence *old,**ptr; + ptr = &ring->fence_drv.fences[j]; + old = rcu_dereference_protected(*ptr, 1); + if (old && test_bit(DMA_FENCE_FLAG_USER_BITS, &old->flags))) { + RCU_INIT_POINTER(*ptr, NULL); + }Is this to avoid premature job free because of
dma_fence_putinsideamdgpu_fence_process?
I can't currently remember why but we probably want all the HW fences currently in the ring to
be forced signaled - maybe better to test forDMA_FENCE_FLAG_USER_BITSinsideamdgpu_fence_process
and still do the signaling but not thedma_fence_putpartQuoted 7 lines
Quoted 1 line
Andrey
Hi Andrey,
This is to avoid signaling the same fence twice. If we still do the signaling, then the job in the pending list will be signaled first in
force_completion, and later be signaled in resubmit. This will go toBUG()inamdgpu_fence_process.Quoted 7 lines
Oh, i see, how about just adding 'skip' flag to
amdgpu_ringand setting it before callingamdgpu_fence_driver_force_completionand resetting it after, then insideamdgpu_fence_driver_force_completion
you can just skip the signaling part with this flag for fences withDMA_FENCE_FLAG_USER_BITSset
Less lines of code at least.Still sounds quite a bit hacky.
I would rather suggest to completely drop the approach with
amdgpu_fence_driver_force_completion(). I could never see why we would want that in the first place.
Quoted 7 lines
Quoted 3 lines
Regards,
Christian.Hi Christian,
I keep the
amdgpu_fence_driver_force_completionhere to make sure the vm flush fence is signaled and put.
Right, so we need to do force completion for the sake of all the fences without parent jobs since there is code which wait directly on them.
Quoted 2 lines
So the key question is whether the fence of ib test should be the first fence to be signaled after reset.
What do you mean by 'after reset' - at this point in the code there was no ASIC reset yet, we just stopped the schedulers and detached the HW fences from their parent jobs (sched_fence)
Quoted 3 lines
If it should be, it means not only fences with
DMA_FENCE_FLAG_USER_BITSbut also vm flush fences should be removed from RCU fence array beforeib_test.
Which ib_test is it, the one after ASIC reset ?
Andrey
Quoted 95 lines
Then we must do the
force_completionhere for vm flush
fence, otherwise there will be a memory leak here as no one will signal and put it after we remove it from RCU.
If not, then the first fence to signal could be vm flush fence. And I'm OK with dropamdgpu_fence_driver_force_completionhere.Best Regards,
JingWen ChenQuoted 86 lines
Quoted 86 lines
Andrey
Quoted 83 lines
Quoted 20 lines
Quoted 20 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1
-0+ } /* after all hw jobs are reset, hw fence ismeaningless, so
force_completion*/
amdgpu_fence_driver_force_completion(ring); }drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +0
-0@@ -156,11 +156,17 @@ int amdgpu_fence_emit(struct
amdgpu_ring*ring, structdma_fence**f, struct amd
job->ring= ring;
}
- seq = ++ring->fence_drv.sync_seq;
- dma_fence_init(fence,&amdgpu_fence_ops, -&ring->fence_drv.lock,
-adev->fence_context+ring->idx, - seq);
+ if (job != NULL &&job->base.resubmit_flag== 1) {Quoted 8 lines
Quoted 5 lines
+ /* reinit seq for resubmitted jobs */
+ seq = ++ring->fence_drv.sync_seq;
+fence->seqno= seq;
+ } else {
+ seq = ++ring->fence_drv.sync_seq;Seems like you could do the above line only once above if-else as it was
beforeSure, I will modify this.
Best Regards,
JingWen ChenQuoted 50 lines
Quoted 50 lines
+ dma_fence_init(fence,
&amdgpu_fence_ops, +&ring->fence_drv.lock,
+adev->fence_context+ring->idx, + seq);
+ }
if (job != NULL) {
/* mark this fence has a parent job */drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +0
-0@@ -241,6 +241,7 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job) dma_fence_set_error(finished, -ECANCELED);/* skipIB as well if VRAM lost */
if (finished->error< 0) {
+dma_fence_put(&job->hw_fence);
DRM_INFO("Skip scheduling IBs!\n");
} else {
r = amdgpu_ib_schedule(ring,job->num_ibs,job->ibs, job,drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +2
-1@@ -249,7 +250,8 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job) DRM_ERROR("Error scheduling IBs (%d)\n", r); } - dma_fence_get(fence); + if (!job->base.resubmit_flag) + dma_fence_get(fence); amdgpu_job_free_resources(job); fence = r ? ERR_PTR(r) : fence;drivers/gpu/drm/scheduler/sched_main.c +1
-0@@ -544,6 +544,7 @@ void drm_sched_resubmit_jobs_ext(struct drm_gpu_scheduler *sched, int max) dma_fence_set_error(&s_fence->finished, -ECANCELED); dma_fence_put(s_job->s_fence->parent); + s_job->resubmit_flag = 1; fence = sched->ops->run_job(s_job); i++;include/drm/gpu_scheduler.h +0
-0@@ -198,6 +198,7 @@ struct drm_sched_job { enum drm_sched_priority s_priority; struct drm_sched_entity *entity; struct dma_fence_cb cb;
Quoted 8 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
include/drm/gpu_scheduler.h struct drm_sched_job {
Quoted 4 lines
include/drm/gpu_scheduler.h +1
-0+ int resubmit_flag; }; static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job,_______________________________________________
amd-gfx mailing list
[email protected]
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&data=04%7C01%7Candrey.grodzovsky%40amd.com%7C9749ced7ce6645bd832408d94d305aaa%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637625692355112825%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=OhLndCUDlWcrhg%2FA0QQ%2FoONxdmQ46CT7P%2F8uvSTGnQ8%3D&reserved=0
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Re: [PATCH 2/2] drm: add tdr support for embeded hw_fence
Quoted 99 lines · Andrey Grodzovsky
Quoted 84 lines · Jingwen Chen
Quoted 83 lines · Christian König
Am 22.07.21 um 16:45 schrieb Andrey Grodzovsky:
Quoted 69 lines · Jingwen Chen
Quoted 60 lines · Andrey Grodzovsky
Quoted 59 lines · Jingwen Chen
Quoted 40 lines
[Why]
After embededhw_fencetoamdgpu_job, we need to add tdr support for this feature.[How]
1. Add aresubmit_flagfor resubmit jobs.
2. Clear job fence from RCU and force complete vm flush fences inpre_asic_reset
3. skipdma_fence_getfor resubmit jobs and add adma_fence_putfor guilty jobs.Signed-off-by Jack Zhang <[email protected]>Signed-off-by Jingwen Chen <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 12 +++++++++++- drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 16 +++++++++++----- drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 4 +++- drivers/gpu/drm/scheduler/sched_main.c | 1 + include/drm/gpu_scheduler.h | 1 + 5 files changed, 27 insertions(+), 7 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1
-1@@ -4382,7 +4382,7 @@ int amdgpu_device_mode1_reset(struct amdgpu_device *adev) int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev, struct amdgpu_reset_context *reset_context) { - int i, r = 0; + int i, j, r = 0; struct amdgpu_job *job = NULL; bool need_full_reset =
test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0
-0@@ -4406,6 +4406,16 @@ int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev, if (!ring || !ring->sched.thread) continue;drivers/gpu/drm/amd/amdgpu/amdgpu_device.c int · line 4418
Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +9
-0+ /*clear job fence from fence drv to avoid force_completion + *leave NULL and vm flush fence in fence drv */ + for (j = 0; j <= ring->fence_drv.num_fences_mask; j ++) { + struct dma_fence *old,**ptr; + ptr = &ring->fence_drv.fences[j]; + old = rcu_dereference_protected(*ptr, 1); + if (old && test_bit(DMA_FENCE_FLAG_USER_BITS, &old->flags))) { + RCU_INIT_POINTER(*ptr, NULL); + }Is this to avoid premature job free because of
dma_fence_putinsideamdgpu_fence_process?
I can't currently remember why but we probably want all the HW fences currently in the ring to
be forced signaled - maybe better to test forDMA_FENCE_FLAG_USER_BITSinsideamdgpu_fence_process
and still do the signaling but not thedma_fence_putpartQuoted 7 lines
Quoted 1 line
Andrey
Hi Andrey,
This is to avoid signaling the same fence twice. If we still do the signaling, then the job in the pending list will be signaled first in
force_completion, and later be signaled in resubmit. This will go toBUG()inamdgpu_fence_process.Quoted 7 lines
Oh, i see, how about just adding 'skip' flag to
amdgpu_ringand setting it before callingamdgpu_fence_driver_force_completionand resetting it after, then insideamdgpu_fence_driver_force_completion
you can just skip the signaling part with this flag for fences withDMA_FENCE_FLAG_USER_BITSset
Less lines of code at least.Still sounds quite a bit hacky.
I would rather suggest to completely drop the approach with
amdgpu_fence_driver_force_completion(). I could never see why we would want that in the first place.Quoted 7 lines
Quoted 3 lines
Regards,
Christian.Hi Christian,
I keep the
amdgpu_fence_driver_force_completionhere to make sure the vm flush fence is signaled and put.Right, so we need to do force completion for the sake of all the fences without parent jobs since there is code which wait directly on them.
Quoted 7 lines
Quoted 2 lines
So the key question is whether the fence of ib test should be the first fence to be signaled after reset.
What do you mean by 'after reset' - at this point in the code there was no ASIC reset yet, we just stopped the schedulers and detached the HW fences from their parent jobs (
sched_fence)
I mean the ASIC reset. There will be a ib_test for each ring after ASIC reset.
Quoted 9 lines
Quoted 5 lines
If it should be, it means not only fences with
DMA_FENCE_FLAG_USER_BITSbut also vm flush fences should be removed from RCU fence array beforeib_test.Which
ib_testis it, the one after ASIC reset ?
Yes
Best Regards,
JingWen Chen
Quoted 105 lines
Andrey
Quoted 103 lines
Then we must do the
force_completionhere for vm flush
fence, otherwise there will be a memory leak here as no one will signal and put it after we remove it from RCU.
If not, then the first fence to signal could be vm flush fence. And I'm OK with dropamdgpu_fence_driver_force_completionhere.Best Regards,
JingWen ChenQuoted 82 lines
Quoted 82 lines
Andrey
Quoted 79 lines
Quoted 20 lines
Quoted 20 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1
-0+ } /* after all hw jobs are reset, hw fence ismeaningless, so
force_completion*/
amdgpu_fence_driver_force_completion(ring); }drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +0
-0@@ -156,11 +156,17 @@ int amdgpu_fence_emit(struct
amdgpu_ring*ring, structdma_fence**f, struct amd
job->ring= ring;
}
- seq = ++ring->fence_drv.sync_seq;
- dma_fence_init(fence,&amdgpu_fence_ops, -&ring->fence_drv.lock,
-adev->fence_context+ring->idx, - seq);
+ if (job != NULL &&job->base.resubmit_flag== 1) {Quoted 8 lines
Quoted 5 lines
+ /* reinit seq for resubmitted jobs */
+ seq = ++ring->fence_drv.sync_seq;
+fence->seqno= seq;
+ } else {
+ seq = ++ring->fence_drv.sync_seq;Seems like you could do the above line only once above if-else as it was
beforeSure, I will modify this.
Best Regards,
JingWen ChenQuoted 46 lines
Quoted 46 lines
+ dma_fence_init(fence,
&amdgpu_fence_ops, +&ring->fence_drv.lock,
+adev->fence_context+ring->idx, + seq);
+ }
if (job != NULL) {
/* mark this fence has a parent job */drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +0
-0@@ -241,6 +241,7 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job) dma_fence_set_error(finished, -ECANCELED);/* skipIB as well if VRAM lost */
if (finished->error< 0) {
+dma_fence_put(&job->hw_fence);
DRM_INFO("Skip scheduling IBs!\n");
} else {
r = amdgpu_ib_schedule(ring,job->num_ibs,job->ibs, job,drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +2
-1@@ -249,7 +250,8 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job) DRM_ERROR("Error scheduling IBs (%d)\n", r); } - dma_fence_get(fence); + if (!job->base.resubmit_flag) + dma_fence_get(fence); amdgpu_job_free_resources(job); fence = r ? ERR_PTR(r) : fence;drivers/gpu/drm/scheduler/sched_main.c +1
-0@@ -544,6 +544,7 @@ void drm_sched_resubmit_jobs_ext(struct drm_gpu_scheduler *sched, int max) dma_fence_set_error(&s_fence->finished, -ECANCELED); dma_fence_put(s_job->s_fence->parent); + s_job->resubmit_flag = 1; fence = sched->ops->run_job(s_job); i++;include/drm/gpu_scheduler.h +0
-0Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
include/drm/gpu_scheduler.h struct drm_sched_job { · line 204
Quoted 8 lines
include/drm/gpu_scheduler.h +1
-0@@ -198,6 +198,7 @@ struct drm_sched_job { enum drm_sched_priority s_priority; struct drm_sched_entity *entity; struct dma_fence_cb cb; + int resubmit_flag; }; static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job,_______________________________________________
amd-gfx mailing list
[email protected]
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&data=04%7C01%7Candrey.grodzovsky%40amd.com%7C9749ced7ce6645bd832408d94d305aaa%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637625692355112825%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=OhLndCUDlWcrhg%2FA0QQ%2FoONxdmQ46CT7P%2F8uvSTGnQ8%3D&reserved=0
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Re: [PATCH 2/2] drm: add tdr support for embeded hw_fence
Quoted 97 lines · Jingwen Chen
Quoted 96 lines · Andrey Grodzovsky
Quoted 83 lines · Jingwen Chen
Quoted 82 lines · Christian König
Am 22.07.21 um 16:45 schrieb Andrey Grodzovsky:
Quoted 68 lines · Jingwen Chen
Quoted 60 lines · Andrey Grodzovsky
Quoted 59 lines · Jingwen Chen
Quoted 40 lines
[Why]
After embededhw_fencetoamdgpu_job, we need to add tdr support for this feature.[How]
1. Add aresubmit_flagfor resubmit jobs.
2. Clear job fence from RCU and force complete vm flush fences inpre_asic_reset
3. skipdma_fence_getfor resubmit jobs and add adma_fence_putfor guilty jobs.Signed-off-by Jack Zhang <[email protected]>Signed-off-by Jingwen Chen <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 12 +++++++++++- drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 16 +++++++++++----- drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 4 +++- drivers/gpu/drm/scheduler/sched_main.c | 1 + include/drm/gpu_scheduler.h | 1 + 5 files changed, 27 insertions(+), 7 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1
-1@@ -4382,7 +4382,7 @@ int amdgpu_device_mode1_reset(struct amdgpu_device *adev) int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev, struct amdgpu_reset_context *reset_context) { - int i, r = 0; + int i, j, r = 0; struct amdgpu_job *job = NULL; bool need_full_reset =test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0
-0@@ -4406,6 +4406,16 @@ int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev, if (!ring || !ring->sched.thread) continue;drivers/gpu/drm/amd/amdgpu/amdgpu_device.c int · line 4418
Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +9
-0+ /*clear job fence from fence drv to avoid force_completion + *leave NULL and vm flush fence in fence drv */ + for (j = 0; j <= ring->fence_drv.num_fences_mask; j ++) { + struct dma_fence *old,**ptr; + ptr = &ring->fence_drv.fences[j]; + old = rcu_dereference_protected(*ptr, 1); + if (old && test_bit(DMA_FENCE_FLAG_USER_BITS, &old->flags))) { + RCU_INIT_POINTER(*ptr, NULL); + }Is this to avoid premature job free because of
dma_fence_putinsideamdgpu_fence_process?
I can't currently remember why but we probably want all the HW fences currently in the ring to
be forced signaled - maybe better to test forDMA_FENCE_FLAG_USER_BITSinsideamdgpu_fence_process
and still do the signaling but not thedma_fence_putpartQuoted 7 lines
Quoted 1 line
Andrey
Hi Andrey,
This is to avoid signaling the same fence twice. If we still do the signaling, then the job in the pending list will be signaled first in
force_completion, and later be signaled in resubmit. This will go toBUG()inamdgpu_fence_process.Quoted 7 lines
Oh, i see, how about just adding 'skip' flag to
amdgpu_ringand setting it before callingamdgpu_fence_driver_force_completionand resetting it after, then insideamdgpu_fence_driver_force_completion
you can just skip the signaling part with this flag for fences withDMA_FENCE_FLAG_USER_BITSset
Less lines of code at least.Still sounds quite a bit hacky.
I would rather suggest to completely drop the approach with
amdgpu_fence_driver_force_completion(). I could never see why we would want that in the first place.Quoted 7 lines
Quoted 3 lines
Regards,
Christian.Hi Christian,
I keep the
amdgpu_fence_driver_force_completionhere to make sure the vm flush fence is signaled and put.Right, so we need to do force completion for the sake of all the fences without parent jobs since there is code which wait directly on them.
Quoted 8 lines
Quoted 6 lines
Quoted 2 lines
So the key question is whether the fence of ib test should be the first fence to be signaled after reset.
What do you mean by 'after reset' - at this point in the code there was no ASIC reset yet, we just stopped the schedulers and detached the HW fences from their parent jobs (
sched_fence)I mean the ASIC reset. There will be a
ib_testfor each ring after ASIC reset.
Then why wouldn't they be the first ? They will emit new fences into the ring which will be signaled right away because the ASIC went through reset and is not hung anymore. All the previous fences, including VM flush once from before the reset will be already signaled by this time from amdgpu_fence_driver_force_completion.
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
If it should be, it means not only fences with
DMA_FENCE_FLAG_USER_BITSbut also vm flush fences should be removed from RCU fence array beforeib_test.
As I mentioned above, not clear to me why VM fences should get special treatment here.
Andrey
Quoted 113 lines
Quoted 3 lines
Which
ib_testis it, the one after ASIC reset ?Yes
Best Regards,
JingWen ChenQuoted 105 lines
Andrey
Quoted 103 lines
Then we must do the
force_completionhere for vm flush
fence, otherwise there will be a memory leak here as no one will signal and put it after we remove it from RCU.
If not, then the first fence to signal could be vm flush fence. And I'm OK with dropamdgpu_fence_driver_force_completionhere.Best Regards,
JingWen ChenQuoted 82 lines
Quoted 82 lines
Andrey
Quoted 79 lines
Quoted 20 lines
Quoted 20 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1
-0+ } /* after all hw jobs are reset, hw fence ismeaningless, so
force_completion*/
amdgpu_fence_driver_force_completion(ring); }drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +0
-0@@ -156,11 +156,17 @@ int amdgpu_fence_emit(struct
amdgpu_ring*ring, structdma_fence**f, struct amd
job->ring= ring;
}
- seq = ++ring->fence_drv.sync_seq;
- dma_fence_init(fence,&amdgpu_fence_ops, -&ring->fence_drv.lock,
-adev->fence_context+ring->idx, - seq);
+ if (job != NULL &&job->base.resubmit_flag== 1) {Quoted 8 lines
Quoted 5 lines
+ /* reinit seq for resubmitted jobs */
+ seq = ++ring->fence_drv.sync_seq;
+fence->seqno= seq;
+ } else {
+ seq = ++ring->fence_drv.sync_seq;Seems like you could do the above line only once above if-else as it was
beforeSure, I will modify this.
Best Regards,
JingWen ChenQuoted 46 lines
Quoted 46 lines
+ dma_fence_init(fence,
&amdgpu_fence_ops, +&ring->fence_drv.lock,
+adev->fence_context+ring->idx, + seq);
+ }
if (job != NULL) {
/* mark this fence has a parent job */drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +0
-0@@ -241,6 +241,7 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job) dma_fence_set_error(finished, -ECANCELED);/* skipIB as well if VRAM lost */
if (finished->error< 0) {
+dma_fence_put(&job->hw_fence);
DRM_INFO("Skip scheduling IBs!\n");
} else {
r = amdgpu_ib_schedule(ring,job->num_ibs,job->ibs, job,drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +2
-1@@ -249,7 +250,8 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job) DRM_ERROR("Error scheduling IBs (%d)\n", r); } - dma_fence_get(fence); + if (!job->base.resubmit_flag) + dma_fence_get(fence); amdgpu_job_free_resources(job); fence = r ? ERR_PTR(r) : fence;drivers/gpu/drm/scheduler/sched_main.c +1
-0@@ -544,6 +544,7 @@ void drm_sched_resubmit_jobs_ext(struct drm_gpu_scheduler *sched, int max) dma_fence_set_error(&s_fence->finished, -ECANCELED); dma_fence_put(s_job->s_fence->parent); + s_job->resubmit_flag = 1; fence = sched->ops->run_job(s_job); i++;include/drm/gpu_scheduler.h +0
-0Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
include/drm/gpu_scheduler.h struct drm_sched_job { · line 204
Quoted 8 lines
include/drm/gpu_scheduler.h +1
-0@@ -198,6 +198,7 @@ struct drm_sched_job { enum drm_sched_priority s_priority; struct drm_sched_entity *entity; struct dma_fence_cb cb; + int resubmit_flag; }; static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job,_______________________________________________
amd-gfx mailing list
[email protected]
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&data=04%7C01%7Candrey.grodzovsky%40amd.com%7C9749ced7ce6645bd832408d94d305aaa%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637625692355112825%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=OhLndCUDlWcrhg%2FA0QQ%2FoONxdmQ46CT7P%2F8uvSTGnQ8%3D&reserved=0
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Re: [PATCH 2/2] drm: add tdr support for embeded hw_fence
Quoted 109 lines · Andrey Grodzovsky
Quoted 97 lines · Jingwen Chen
Quoted 96 lines · Andrey Grodzovsky
Quoted 83 lines · Jingwen Chen
Quoted 82 lines · Christian König
Am 22.07.21 um 16:45 schrieb Andrey Grodzovsky:
Quoted 68 lines · Jingwen Chen
Quoted 60 lines · Andrey Grodzovsky
Quoted 59 lines · Jingwen Chen
Quoted 40 lines
[Why]
After embededhw_fencetoamdgpu_job, we need to add tdr support for this feature.[How]
1. Add aresubmit_flagfor resubmit jobs.
2. Clear job fence from RCU and force complete vm flush fences inpre_asic_reset
3. skipdma_fence_getfor resubmit jobs and add adma_fence_putfor guilty jobs.Signed-off-by Jack Zhang <[email protected]>Signed-off-by Jingwen Chen <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 12 +++++++++++- drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 16 +++++++++++----- drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 4 +++- drivers/gpu/drm/scheduler/sched_main.c | 1 + include/drm/gpu_scheduler.h | 1 + 5 files changed, 27 insertions(+), 7 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1
-1@@ -4382,7 +4382,7 @@ int amdgpu_device_mode1_reset(struct amdgpu_device *adev) int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev, struct amdgpu_reset_context *reset_context) { - int i, r = 0; + int i, j, r = 0; struct amdgpu_job *job = NULL; bool need_full_reset =test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0
-0@@ -4406,6 +4406,16 @@ int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev, if (!ring || !ring->sched.thread) continue;drivers/gpu/drm/amd/amdgpu/amdgpu_device.c int · line 4418
Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +9
-0+ /*clear job fence from fence drv to avoid force_completion + *leave NULL and vm flush fence in fence drv */ + for (j = 0; j <= ring->fence_drv.num_fences_mask; j ++) { + struct dma_fence *old,**ptr; + ptr = &ring->fence_drv.fences[j]; + old = rcu_dereference_protected(*ptr, 1); + if (old && test_bit(DMA_FENCE_FLAG_USER_BITS, &old->flags))) { + RCU_INIT_POINTER(*ptr, NULL); + }Is this to avoid premature job free because of
dma_fence_putinsideamdgpu_fence_process?
I can't currently remember why but we probably want all the HW fences currently in the ring to
be forced signaled - maybe better to test forDMA_FENCE_FLAG_USER_BITSinsideamdgpu_fence_process
and still do the signaling but not thedma_fence_putpartQuoted 7 lines
Quoted 1 line
Andrey
Hi Andrey,
This is to avoid signaling the same fence twice. If we still do the signaling, then the job in the pending list will be signaled first in
force_completion, and later be signaled in resubmit. This will go toBUG()inamdgpu_fence_process.Quoted 7 lines
Oh, i see, how about just adding 'skip' flag to
amdgpu_ringand setting it before callingamdgpu_fence_driver_force_completionand resetting it after, then insideamdgpu_fence_driver_force_completion
you can just skip the signaling part with this flag for fences withDMA_FENCE_FLAG_USER_BITSset
Less lines of code at least.Still sounds quite a bit hacky.
I would rather suggest to completely drop the approach with
amdgpu_fence_driver_force_completion(). I could never see why we would want that in the first place.Quoted 7 lines
Quoted 3 lines
Regards,
Christian.Hi Christian,
I keep the
amdgpu_fence_driver_force_completionhere to make sure the vm flush fence is signaled and put.Right, so we need to do force completion for the sake of all the fences without parent jobs since there is code which wait directly on them.
Quoted 8 lines
Quoted 6 lines
Quoted 2 lines
So the key question is whether the fence of ib test should be the first fence to be signaled after reset.
What do you mean by 'after reset' - at this point in the code there was no ASIC reset yet, we just stopped the schedulers and detached the HW fences from their parent jobs (
sched_fence)I mean the ASIC reset. There will be a
ib_testfor each ring after ASIC reset.
Quoted 9 lines
Then why wouldn't they be the first ? They will emit new fences into the ring
which will be signaled right away because the ASIC went through reset and is not
hung anymore. All the previous fences, including VM flush once from before the
reset will be already signaled by this time fromamdgpu_fence_driver_force_completion.
Hi Andrey,
Sorry I didn't make it clear. I mean if we drop force_completion here, and keep other code unchanged, then the ib_test wouldn't be the first to be signaled.
Best Regards,
JingWen Chen
Quoted 7 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
If it should be, it means not only fences with
DMA_FENCE_FLAG_USER_BITSbut also vm flush fences should be removed from RCU fence array beforeib_test.
Quoted 119 lines
As I mentioned above, not clear to me why VM fences should get special treatment here.
Andrey
Quoted 113 lines
Quoted 3 lines
Which
ib_testis it, the one after ASIC reset ?Yes
Best Regards,
JingWen ChenQuoted 105 lines
Andrey
Quoted 103 lines
Then we must do the
force_completionhere for vm flush
fence, otherwise there will be a memory leak here as no one will signal and put it after we remove it from RCU.
If not, then the first fence to signal could be vm flush fence. And I'm OK with dropamdgpu_fence_driver_force_completionhere.Best Regards,
JingWen ChenQuoted 82 lines
Quoted 82 lines
Andrey
Quoted 79 lines
Quoted 20 lines
Quoted 20 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1
-0+ } /* after all hw jobs are reset, hw fence ismeaningless, so
force_completion*/
amdgpu_fence_driver_force_completion(ring); }drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +0
-0@@ -156,11 +156,17 @@ int amdgpu_fence_emit(struct
amdgpu_ring*ring, structdma_fence**f, struct amd
job->ring= ring;
}
- seq = ++ring->fence_drv.sync_seq;
- dma_fence_init(fence,&amdgpu_fence_ops, -&ring->fence_drv.lock,
-adev->fence_context+ring->idx, - seq);
+ if (job != NULL &&job->base.resubmit_flag== 1) {Quoted 8 lines
Quoted 5 lines
+ /* reinit seq for resubmitted jobs */
+ seq = ++ring->fence_drv.sync_seq;
+fence->seqno= seq;
+ } else {
+ seq = ++ring->fence_drv.sync_seq;Seems like you could do the above line only once above if-else as it was
beforeSure, I will modify this.
Best Regards,
JingWen ChenQuoted 46 lines
Quoted 46 lines
+ dma_fence_init(fence,
&amdgpu_fence_ops, +&ring->fence_drv.lock,
+adev->fence_context+ring->idx, + seq);
+ }
if (job != NULL) {
/* mark this fence has a parent job */drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +0
-0@@ -241,6 +241,7 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job) dma_fence_set_error(finished, -ECANCELED);/* skipIB as well if VRAM lost */
if (finished->error< 0) {
+dma_fence_put(&job->hw_fence);
DRM_INFO("Skip scheduling IBs!\n");
} else {
r = amdgpu_ib_schedule(ring,job->num_ibs,job->ibs, job,drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +2
-1@@ -249,7 +250,8 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job) DRM_ERROR("Error scheduling IBs (%d)\n", r); } - dma_fence_get(fence); + if (!job->base.resubmit_flag) + dma_fence_get(fence); amdgpu_job_free_resources(job); fence = r ? ERR_PTR(r) : fence;drivers/gpu/drm/scheduler/sched_main.c +1
-0@@ -544,6 +544,7 @@ void drm_sched_resubmit_jobs_ext(struct drm_gpu_scheduler *sched, int max) dma_fence_set_error(&s_fence->finished, -ECANCELED); dma_fence_put(s_job->s_fence->parent); + s_job->resubmit_flag = 1; fence = sched->ops->run_job(s_job); i++;include/drm/gpu_scheduler.h +0
-0Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
include/drm/gpu_scheduler.h struct drm_sched_job { · line 204
Quoted 8 lines
include/drm/gpu_scheduler.h +1
-0@@ -198,6 +198,7 @@ struct drm_sched_job { enum drm_sched_priority s_priority; struct drm_sched_entity *entity; struct dma_fence_cb cb; + int resubmit_flag; }; static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job,_______________________________________________
amd-gfx mailing list
[email protected]
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&data=04%7C01%7Candrey.grodzovsky%40amd.com%7C9749ced7ce6645bd832408d94d305aaa%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637625692355112825%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=OhLndCUDlWcrhg%2FA0QQ%2FoONxdmQ46CT7P%2F8uvSTGnQ8%3D&reserved=0
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Re: [PATCH 2/2] drm: add tdr support for embeded hw_fence
Quoted 115 lines · Jingwen Chen
Quoted 105 lines · Andrey Grodzovsky
Quoted 96 lines · Jingwen Chen
Quoted 95 lines · Andrey Grodzovsky
Quoted 83 lines · Jingwen Chen
Quoted 82 lines · Christian König
Am 22.07.21 um 16:45 schrieb Andrey Grodzovsky:
Quoted 68 lines · Jingwen Chen
Quoted 60 lines · Andrey Grodzovsky
Quoted 59 lines · Jingwen Chen
Quoted 40 lines
[Why]
After embededhw_fencetoamdgpu_job, we need to add tdr support for this feature.[How]
1. Add aresubmit_flagfor resubmit jobs.
2. Clear job fence from RCU and force complete vm flush fences inpre_asic_reset
3. skipdma_fence_getfor resubmit jobs and add adma_fence_putfor guilty jobs.Signed-off-by Jack Zhang <[email protected]>Signed-off-by Jingwen Chen <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 12 +++++++++++- drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 16 +++++++++++----- drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 4 +++- drivers/gpu/drm/scheduler/sched_main.c | 1 + include/drm/gpu_scheduler.h | 1 + 5 files changed, 27 insertions(+), 7 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1
-1@@ -4382,7 +4382,7 @@ int amdgpu_device_mode1_reset(struct amdgpu_device *adev) int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev, struct amdgpu_reset_context *reset_context) { - int i, r = 0; + int i, j, r = 0; struct amdgpu_job *job = NULL; bool need_full_reset =test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0
-0@@ -4406,6 +4406,16 @@ int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev, if (!ring || !ring->sched.thread) continue;drivers/gpu/drm/amd/amdgpu/amdgpu_device.c int · line 4418
Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +9
-0+ /*clear job fence from fence drv to avoid force_completion + *leave NULL and vm flush fence in fence drv */ + for (j = 0; j <= ring->fence_drv.num_fences_mask; j ++) { + struct dma_fence *old,**ptr; + ptr = &ring->fence_drv.fences[j]; + old = rcu_dereference_protected(*ptr, 1); + if (old && test_bit(DMA_FENCE_FLAG_USER_BITS, &old->flags))) { + RCU_INIT_POINTER(*ptr, NULL); + }Is this to avoid premature job free because of
dma_fence_putinsideamdgpu_fence_process?
I can't currently remember why but we probably want all the HW fences currently in the ring to
be forced signaled - maybe better to test forDMA_FENCE_FLAG_USER_BITSinsideamdgpu_fence_process
and still do the signaling but not thedma_fence_putpartQuoted 7 lines
Quoted 1 line
Andrey
Hi Andrey,
This is to avoid signaling the same fence twice. If we still do the signaling, then the job in the pending list will be signaled first in
force_completion, and later be signaled in resubmit. This will go toBUG()inamdgpu_fence_process.Quoted 7 lines
Oh, i see, how about just adding 'skip' flag to
amdgpu_ringand setting it before callingamdgpu_fence_driver_force_completionand resetting it after, then insideamdgpu_fence_driver_force_completion
you can just skip the signaling part with this flag for fences withDMA_FENCE_FLAG_USER_BITSset
Less lines of code at least.Still sounds quite a bit hacky.
I would rather suggest to completely drop the approach with
amdgpu_fence_driver_force_completion(). I could never see why we would want that in the first place.Quoted 7 lines
Quoted 3 lines
Regards,
Christian.Hi Christian,
I keep the
amdgpu_fence_driver_force_completionhere to make sure the vm flush fence is signaled and put.Right, so we need to do force completion for the sake of all the fences without parent jobs since there is code which wait directly on them.
Quoted 7 lines
Quoted 5 lines
Quoted 2 lines
So the key question is whether the fence of ib test should be the first fence to be signaled after reset.
What do you mean by 'after reset' - at this point in the code there was no ASIC reset yet, we just stopped the schedulers and detached the HW fences from their parent jobs (
sched_fence)I mean the ASIC reset. There will be a
ib_testfor each ring after ASIC reset.Quoted 9 lines
Then why wouldn't they be the first ? They will emit new fences into the ring
which will be signaled right away because the ASIC went through reset and is not
hung anymore. All the previous fences, including VM flush once from before the
reset will be already signaled by this time fromamdgpu_fence_driver_force_completion.
Quoted 5 lines
Hi Andrey,
Sorry I didn't make it clear. I mean if we drop
force_completionhere, and keep other code unchanged, then theib_testwouldn't be the first to be signaled.
At least in my opinion the order is not important of who will be signaled first. I do think it's important to force signal all the old fences before HW reset for 2 reasons: 1st - it's conceptually wrong to leave them unsignaled and let them be signaled post ASIC reset because it's not them who actually completed whatever they were doing but rather ASIC reset allowed later fences (with higher sync_seq) to signal (e.g. ring test) and those earlier remaining fences (e.g. VM_FLUSH) just were signaled as part of this. 2nd - Possible deadlock where ring->fence_drv.fences is full with fences before ASIC reset because we don't force signal and so don't empty it and then after reset we call amdgpu_fence_emit as part of ib_test but it de-references old yet unsignaled fence and does dma_fence_wait to make sure it completed which it didn't and will not because EOP ISR will be called only after actual fence emission which is blocked by the wait.
Andrey
Quoted 126 lines
Best Regards,
JingWen ChenQuoted 122 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
If it should be, it means not only fences with
DMA_FENCE_FLAG_USER_BITSbut also vm flush fences should be removed from RCU fence array beforeib_test.As I mentioned above, not clear to me why VM fences should get special treatment here.
Andrey
Quoted 112 lines
Quoted 2 lines
Which
ib_testis it, the one after ASIC reset ?Yes
Best Regards,
JingWen ChenQuoted 105 lines
Andrey
Quoted 103 lines
Then we must do the
force_completionhere for vm flush
fence, otherwise there will be a memory leak here as no one will signal and put it after we remove it from RCU.
If not, then the first fence to signal could be vm flush fence. And I'm OK with dropamdgpu_fence_driver_force_completionhere.Best Regards,
JingWen ChenQuoted 82 lines
Quoted 82 lines
Andrey
Quoted 79 lines
Quoted 20 lines
Quoted 20 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1
-0+ } /* after all hw jobs are reset, hw fence ismeaningless, so
force_completion*/amdgpu_fence_driver_force_completion(ring); }drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +0
-0@@ -156,11 +156,17 @@ int amdgpu_fence_emit(struct
amdgpu_ring*ring, structdma_fence**f, struct amdjob->ring = ring; }- seq = ++ring->fence_drv.sync_seq;
- dma_fence_init(fence,&amdgpu_fence_ops, -&ring->fence_drv.lock,
-adev->fence_context+ring->idx, - seq);
+ if (job != NULL &&job->base.resubmit_flag== 1) {Quoted 8 lines
Quoted 5 lines
+ /* reinit seq for resubmitted jobs */
+ seq = ++ring->fence_drv.sync_seq;
+fence->seqno= seq;
+ } else {
+ seq = ++ring->fence_drv.sync_seq;Seems like you could do the above line only once above if-else as it was
beforeSure, I will modify this.
Best Regards,
JingWen ChenQuoted 46 lines
Quoted 46 lines
+ dma_fence_init(fence,
&amdgpu_fence_ops,
+&ring->fence_drv.lock,
+adev->fence_context+ring->idx,
+ seq);
+ }if (job != NULL) { /* mark this fence has a parent job */drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +0
-0@@ -241,6 +241,7 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job) dma_fence_set_error(finished, -ECANCELED);/* skipIB as well if VRAM lost */
if (finished->error< 0) {+ dma_fence_put(&job->hw_fence); DRM_INFO("Skip scheduling IBs!\n"); } else { r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs, job,drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +2
-1@@ -249,7 +250,8 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job) DRM_ERROR("Error scheduling IBs (%d)\n", r); } - dma_fence_get(fence); + if (!job->base.resubmit_flag) + dma_fence_get(fence); amdgpu_job_free_resources(job); fence = r ? ERR_PTR(r) : fence;drivers/gpu/drm/scheduler/sched_main.c +1
-0@@ -544,6 +544,7 @@ void drm_sched_resubmit_jobs_ext(struct drm_gpu_scheduler *sched, int max) dma_fence_set_error(&s_fence->finished, -ECANCELED); dma_fence_put(s_job->s_fence->parent); + s_job->resubmit_flag = 1; fence = sched->ops->run_job(s_job); i++;include/drm/gpu_scheduler.h +0
-0Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
include/drm/gpu_scheduler.h struct drm_sched_job { · line 204
Quoted 8 lines
include/drm/gpu_scheduler.h +1
-0@@ -198,6 +198,7 @@ struct drm_sched_job { enum drm_sched_priority s_priority; struct drm_sched_entity *entity; struct dma_fence_cb cb; + int resubmit_flag; }; static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job,_______________________________________________
amd-gfx mailing list
[email protected]
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&data=04%7C01%7Candrey.grodzovsky%40amd.com%7C9749ced7ce6645bd832408d94d305aaa%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637625692355112825%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=OhLndCUDlWcrhg%2FA0QQ%2FoONxdmQ46CT7P%2F8uvSTGnQ8%3D&reserved=0
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Re: [PATCH 2/2] drm: add tdr support for embeded hw_fence
Quoted 142 lines · Andrey Grodzovsky
Quoted 115 lines · Jingwen Chen
Quoted 105 lines · Andrey Grodzovsky
Quoted 96 lines · Jingwen Chen
Quoted 95 lines · Andrey Grodzovsky
Quoted 83 lines · Jingwen Chen
Quoted 82 lines · Christian König
Am 22.07.21 um 16:45 schrieb Andrey Grodzovsky:
Quoted 68 lines · Jingwen Chen
Quoted 60 lines · Andrey Grodzovsky
Quoted 59 lines · Jingwen Chen
Quoted 40 lines
[Why]
After embededhw_fencetoamdgpu_job, we need to add tdr support for this feature.[How]
1. Add aresubmit_flagfor resubmit jobs.
2. Clear job fence from RCU and force complete vm flush fences inpre_asic_reset
3. skipdma_fence_getfor resubmit jobs and add adma_fence_putfor guilty jobs.Signed-off-by Jack Zhang <[email protected]>Signed-off-by Jingwen Chen <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 12 +++++++++++- drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 16 +++++++++++----- drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 4 +++- drivers/gpu/drm/scheduler/sched_main.c | 1 + include/drm/gpu_scheduler.h | 1 + 5 files changed, 27 insertions(+), 7 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1
-1@@ -4382,7 +4382,7 @@ int amdgpu_device_mode1_reset(struct amdgpu_device *adev) int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev, struct amdgpu_reset_context *reset_context) { - int i, r = 0; + int i, j, r = 0; struct amdgpu_job *job = NULL; bool need_full_reset =test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0
-0@@ -4406,6 +4406,16 @@ int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev, if (!ring || !ring->sched.thread) continue;drivers/gpu/drm/amd/amdgpu/amdgpu_device.c int · line 4418
Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +9
-0+ /*clear job fence from fence drv to avoid force_completion + *leave NULL and vm flush fence in fence drv */ + for (j = 0; j <= ring->fence_drv.num_fences_mask; j ++) { + struct dma_fence *old,**ptr; + ptr = &ring->fence_drv.fences[j]; + old = rcu_dereference_protected(*ptr, 1); + if (old && test_bit(DMA_FENCE_FLAG_USER_BITS, &old->flags))) { + RCU_INIT_POINTER(*ptr, NULL); + }Is this to avoid premature job free because of
dma_fence_putinsideamdgpu_fence_process?
I can't currently remember why but we probably want all the HW fences currently in the ring to
be forced signaled - maybe better to test forDMA_FENCE_FLAG_USER_BITSinsideamdgpu_fence_process
and still do the signaling but not thedma_fence_putpartQuoted 7 lines
Quoted 1 line
Andrey
Hi Andrey,
This is to avoid signaling the same fence twice. If we still do the signaling, then the job in the pending list will be signaled first in
force_completion, and later be signaled in resubmit. This will go toBUG()inamdgpu_fence_process.Quoted 7 lines
Oh, i see, how about just adding 'skip' flag to
amdgpu_ringand setting it before callingamdgpu_fence_driver_force_completionand resetting it after, then insideamdgpu_fence_driver_force_completion
you can just skip the signaling part with this flag for fences withDMA_FENCE_FLAG_USER_BITSset
Less lines of code at least.Still sounds quite a bit hacky.
I would rather suggest to completely drop the approach with
amdgpu_fence_driver_force_completion(). I could never see why we would want that in the first place.Quoted 7 lines
Quoted 3 lines
Regards,
Christian.Hi Christian,
I keep the
amdgpu_fence_driver_force_completionhere to make sure the vm flush fence is signaled and put.Right, so we need to do force completion for the sake of all the fences without parent jobs since there is code which wait directly on them.
Quoted 7 lines
Quoted 5 lines
Quoted 2 lines
So the key question is whether the fence of ib test should be the first fence to be signaled after reset.
What do you mean by 'after reset' - at this point in the code there was no ASIC reset yet, we just stopped the schedulers and detached the HW fences from their parent jobs (
sched_fence)I mean the ASIC reset. There will be a
ib_testfor each ring after ASIC reset.Quoted 9 lines
Then why wouldn't they be the first ? They will emit new fences into the ring
which will be signaled right away because the ASIC went through reset and is not
hung anymore. All the previous fences, including VM flush once from before the
reset will be already signaled by this time fromamdgpu_fence_driver_force_completion.Quoted 5 lines
Hi Andrey,
Sorry I didn't make it clear. I mean if we drop
force_completionhere, and keep other code unchanged, then theib_testwouldn't be the first to be signaled.At least in my opinion the order is not important of who will be signaled first. I do think it's important
to force signal all the old fences before HW reset for 2 reasons: 1st - it's conceptually wrong to leave them
unsignaled and let them be signaled post ASIC reset because it's not them who actually completed whatever
they were doing but rather ASIC reset allowed later fences (with highersync_seq) to signal (e.g. ring test) and
those earlier remaining fences (e.g. VM_FLUSH) just were signaled as part of this. 2nd - Possible deadlock wherering->fence_drv.fencesis full with fences before ASIC reset because we don't force signal and so don't empty it and then
after reset we callamdgpu_fence_emitas part ofib_testbut it de-references old yet unsignaled fence and doesdma_fence_waitto make sure it completed which it didn't and will not because EOP ISR will be called only after actual
fence emission which is blocked by the wait.
Quoted 1 line
Andrey
Hi Andrey
Yes, I forget to talk about the deadlock previously. Actually I have encountered a deadlock during debug already which is exactly the same reason you mentioned.
I totally agree with not dropping force_completion.
Quoted 65 lines
Quoted 65 lines
Best Regards,
JingWen ChenQuoted 59 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
If it should be, it means not only fences with
DMA_FENCE_FLAG_USER_BITSbut also vm flush fences should be removed from RCU fence array beforeib_test.As I mentioned above, not clear to me why VM fences should get special treatment here.
Andrey
Quoted 49 lines
Quoted 2 lines
Which
ib_testis it, the one after ASIC reset ?Yes
Best Regards,
JingWen ChenQuoted 42 lines
Andrey
Quoted 40 lines
Then we must do the
force_completionhere for vm flush
fence, otherwise there will be a memory leak here as no one will signal and put it after we remove it from RCU.
If not, then the first fence to signal could be vm flush fence. And I'm OK with dropamdgpu_fence_driver_force_completionhere.Best Regards,
JingWen ChenQuoted 31 lines
Quoted 31 lines
Andrey
Quoted 28 lines
Quoted 28 lines
Quoted 14 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1
-0+ } /* after all hw jobs are reset, hw fence ismeaningless, so
force_completion*/amdgpu_fence_driver_force_completion(ring); }drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +0
-0@@ -156,11 +156,17 @@ int amdgpu_fence_emit(struct
amdgpu_ring*ring, structdma_fence**f, struct amdjob->ring = ring; }Quoted 11 lines
- seq = ++ring->fence_drv.sync_seq;
- dma_fence_init(fence,&amdgpu_fence_ops, -&ring->fence_drv.lock,
-adev->fence_context+ring->idx, - seq);
+ if (job != NULL &&job->base.resubmit_flag== 1) { + /* reinit seq for resubmitted jobs */ + seq = ++ring->fence_drv.sync_seq;
+fence->seqno= seq;
+ } else {
+ seq = ++ring->fence_drv.sync_seq;Seems like you could do the above line only once above if-else as it was
before
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Sure, I will modify this.
I will modify this part and upload a v2 patch. Can you help add a ACK or review by?
Best Regards,
JingWen Chen
Quoted 52 lines
Quoted 52 lines
Quoted 52 lines
Quoted 52 lines
Quoted 52 lines
Quoted 52 lines
Quoted 52 lines
Quoted 52 lines
Quoted 52 lines
Best Regards,
JingWen ChenQuoted 50 lines
Quoted 50 lines
+ dma_fence_init(fence,
&amdgpu_fence_ops,
+&ring->fence_drv.lock,
+adev->fence_context+ring->idx,
+ seq);
+ }if (job != NULL) { /* mark this fence has a parent job */drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +0
-0@@ -241,6 +241,7 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job) dma_fence_set_error(finished, -ECANCELED);/* skipIB as well if VRAM lost */
if (finished->error< 0) {+ dma_fence_put(&job->hw_fence); DRM_INFO("Skip scheduling IBs!\n"); } else { r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs, job,drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +2
-1@@ -249,7 +250,8 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job) DRM_ERROR("Error scheduling IBs (%d)\n", r); } - dma_fence_get(fence); + if (!job->base.resubmit_flag) + dma_fence_get(fence); amdgpu_job_free_resources(job); fence = r ? ERR_PTR(r) : fence;drivers/gpu/drm/scheduler/sched_main.c +1
-0@@ -544,6 +544,7 @@ void drm_sched_resubmit_jobs_ext(struct drm_gpu_scheduler *sched, int max) dma_fence_set_error(&s_fence->finished, -ECANCELED); dma_fence_put(s_job->s_fence->parent); + s_job->resubmit_flag = 1; fence = sched->ops->run_job(s_job); i++;include/drm/gpu_scheduler.h +0
-0@@ -198,6 +198,7 @@ struct drm_sched_job { enum drm_sched_priority s_priority; struct drm_sched_entity *entity; struct dma_fence_cb cb;
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
include/drm/gpu_scheduler.h struct drm_sched_job { · line 204
Quoted 4 lines
include/drm/gpu_scheduler.h +1
-0+ int resubmit_flag; }; static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job,_______________________________________________
amd-gfx mailing list
[email protected]
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&data=04%7C01%7Candrey.grodzovsky%40amd.com%7C9749ced7ce6645bd832408d94d305aaa%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637625692355112825%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=OhLndCUDlWcrhg%2FA0QQ%2FoONxdmQ46CT7P%2F8uvSTGnQ8%3D&reserved=0
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Christian König
reviewerRe: [PATCH 2/2] drm: add tdr support for embeded hw_fence
In reply to Jingwen ChenAm 22.07.21 um 18:47 schrieb Jingwen Chen:
Quoted 89 lines · Christian König
Quoted 77 lines
Am 22.07.21 um 16:45 schrieb Andrey Grodzovsky:
Quoted 69 lines · Jingwen Chen
Quoted 60 lines · Andrey Grodzovsky
Quoted 59 lines · Jingwen Chen
Quoted 40 lines
[Why]
After embededhw_fencetoamdgpu_job, we need to add tdr support for this feature.[How]
1. Add aresubmit_flagfor resubmit jobs.
2. Clear job fence from RCU and force complete vm flush fences inpre_asic_reset
3. skipdma_fence_getfor resubmit jobs and add adma_fence_putfor guilty jobs.Signed-off-by Jack Zhang <[email protected]>Signed-off-by Jingwen Chen <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 12 +++++++++++- drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 16 +++++++++++----- drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 4 +++- drivers/gpu/drm/scheduler/sched_main.c | 1 + include/drm/gpu_scheduler.h | 1 + 5 files changed, 27 insertions(+), 7 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1
-1@@ -4382,7 +4382,7 @@ int amdgpu_device_mode1_reset(struct amdgpu_device *adev) int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev, struct amdgpu_reset_context *reset_context) { - int i, r = 0; + int i, j, r = 0; struct amdgpu_job *job = NULL; bool need_full_reset =
test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0
-0@@ -4406,6 +4406,16 @@ int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev, if (!ring || !ring->sched.thread) continue;drivers/gpu/drm/amd/amdgpu/amdgpu_device.c int · line 4418
Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +9
-0+ /*clear job fence from fence drv to avoid force_completion + *leave NULL and vm flush fence in fence drv */ + for (j = 0; j <= ring->fence_drv.num_fences_mask; j ++) { + struct dma_fence *old,**ptr; + ptr = &ring->fence_drv.fences[j]; + old = rcu_dereference_protected(*ptr, 1); + if (old && test_bit(DMA_FENCE_FLAG_USER_BITS, &old->flags))) { + RCU_INIT_POINTER(*ptr, NULL); + }Is this to avoid premature job free because of
dma_fence_putinsideamdgpu_fence_process?
I can't currently remember why but we probably want all the HW fences currently in the ring to
be forced signaled - maybe better to test forDMA_FENCE_FLAG_USER_BITSinsideamdgpu_fence_process
and still do the signaling but not thedma_fence_putpartQuoted 7 lines
Quoted 1 line
Andrey
Hi Andrey,
This is to avoid signaling the same fence twice. If we still do the signaling, then the job in the pending list will be signaled first in
force_completion, and later be signaled in resubmit. This will go toBUG()inamdgpu_fence_process.Quoted 7 lines
Oh, i see, how about just adding 'skip' flag to
amdgpu_ringand setting it before callingamdgpu_fence_driver_force_completionand resetting it after, then insideamdgpu_fence_driver_force_completion
you can just skip the signaling part with this flag for fences withDMA_FENCE_FLAG_USER_BITSset
Less lines of code at least.Quoted 9 lines
Still sounds quite a bit hacky.
I would rather suggest to completely drop the approach with
amdgpu_fence_driver_force_completion(). I could never see why we would want that in the first place.Regards,
Christian.Hi Christian,
Quoted 11 lines
I keep the
amdgpu_fence_driver_force_completionhere to make sure the vm flush fence is signaled and put.
So the key question is whether the fence of ib test should be the first fence to be signaled after reset.
If it should be, it means not only fences withDMA_FENCE_FLAG_USER_BITSbut also vm flush fences should be removed from RCU fence array beforeib_test. Then we must do theforce_completionhere for vm flush fence, otherwise there will be a memory leak here as no one will signal and put it after we remove it from RCU.
If not, then the first fence to signal could be vm flush fence. And I'm OK with dropamdgpu_fence_driver_force_completionhere.
The key problem is that amdgpu_fence_driver_force_completion() goes over the RCU array and signals everything there.
What we should do instead is to go over the jobs and cleanup the ones we don't want to re-submit and keep the rest.
And please drop any usage of DMA_FENCE_FLAG_USER_BITS. That is not something which should be abused for reset handling.
Regards,
Christian.
Quoted 86 lines
Best Regards,
JingWen ChenQuoted 82 lines
Quoted 82 lines
Andrey
Quoted 79 lines
Quoted 20 lines
Quoted 20 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1
-0+ } /* after all hw jobs are reset, hw fence ismeaningless, so
force_completion*/
amdgpu_fence_driver_force_completion(ring); }drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +0
-0@@ -156,11 +156,17 @@ int amdgpu_fence_emit(struct
amdgpu_ring*ring, structdma_fence**f, struct amd
job->ring= ring;
}
- seq = ++ring->fence_drv.sync_seq;
- dma_fence_init(fence,&amdgpu_fence_ops, -&ring->fence_drv.lock,
-adev->fence_context+ring->idx, - seq);
+ if (job != NULL &&job->base.resubmit_flag== 1) {Quoted 8 lines
Quoted 5 lines
+ /* reinit seq for resubmitted jobs */
+ seq = ++ring->fence_drv.sync_seq;
+fence->seqno= seq;
+ } else {
+ seq = ++ring->fence_drv.sync_seq;Seems like you could do the above line only once above if-else as it was
beforeSure, I will modify this.
Best Regards,
JingWen ChenQuoted 46 lines
Quoted 46 lines
+ dma_fence_init(fence,
&amdgpu_fence_ops, +&ring->fence_drv.lock,
+adev->fence_context+ring->idx, + seq);
+ }
if (job != NULL) {
/* mark this fence has a parent job */drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +0
-0@@ -241,6 +241,7 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job) dma_fence_set_error(finished, -ECANCELED);/* skipIB as well if VRAM lost */
if (finished->error< 0) {
+dma_fence_put(&job->hw_fence);
DRM_INFO("Skip scheduling IBs!\n");
} else {
r = amdgpu_ib_schedule(ring,job->num_ibs,job->ibs, job,drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +2
-1@@ -249,7 +250,8 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job) DRM_ERROR("Error scheduling IBs (%d)\n", r); } - dma_fence_get(fence); + if (!job->base.resubmit_flag) + dma_fence_get(fence); amdgpu_job_free_resources(job); fence = r ? ERR_PTR(r) : fence;drivers/gpu/drm/scheduler/sched_main.c +1
-0@@ -544,6 +544,7 @@ void drm_sched_resubmit_jobs_ext(struct drm_gpu_scheduler *sched, int max) dma_fence_set_error(&s_fence->finished, -ECANCELED); dma_fence_put(s_job->s_fence->parent); + s_job->resubmit_flag = 1; fence = sched->ops->run_job(s_job); i++;include/drm/gpu_scheduler.h +0
-0
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
include/drm/gpu_scheduler.h struct drm_sched_job { · line 204
Quoted 8 lines
include/drm/gpu_scheduler.h +1
-0@@ -198,6 +198,7 @@ struct drm_sched_job { enum drm_sched_priority s_priority; struct drm_sched_entity *entity; struct dma_fence_cb cb; + int resubmit_flag; }; static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job,
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Re: [PATCH 2/2] drm: add tdr support for embeded hw_fence
Quoted 102 lines · Christian König
Am 22.07.21 um 18:47 schrieb Jingwen Chen:
Quoted 89 lines · Christian König
Quoted 77 lines
Am 22.07.21 um 16:45 schrieb Andrey Grodzovsky:
Quoted 69 lines · Jingwen Chen
Quoted 60 lines · Andrey Grodzovsky
Quoted 59 lines · Jingwen Chen
Quoted 40 lines
[Why]
After embededhw_fencetoamdgpu_job, we need to add tdr support for this feature.[How]
1. Add aresubmit_flagfor resubmit jobs.
2. Clear job fence from RCU and force complete vm flush fences inpre_asic_reset
3. skipdma_fence_getfor resubmit jobs and add adma_fence_putfor guilty jobs.Signed-off-by Jack Zhang <[email protected]>Signed-off-by Jingwen Chen <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 12 +++++++++++- drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 16 +++++++++++----- drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 4 +++- drivers/gpu/drm/scheduler/sched_main.c | 1 + include/drm/gpu_scheduler.h | 1 + 5 files changed, 27 insertions(+), 7 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1
-1@@ -4382,7 +4382,7 @@ int amdgpu_device_mode1_reset(struct amdgpu_device *adev) int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev, struct amdgpu_reset_context *reset_context) { - int i, r = 0; + int i, j, r = 0; struct amdgpu_job *job = NULL; bool need_full_reset =
test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0
-0@@ -4406,6 +4406,16 @@ int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev, if (!ring || !ring->sched.thread) continue;drivers/gpu/drm/amd/amdgpu/amdgpu_device.c int · line 4418
Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +9
-0+ /*clear job fence from fence drv to avoid force_completion + *leave NULL and vm flush fence in fence drv */ + for (j = 0; j <= ring->fence_drv.num_fences_mask; j ++) { + struct dma_fence *old,**ptr; + ptr = &ring->fence_drv.fences[j]; + old = rcu_dereference_protected(*ptr, 1); + if (old && test_bit(DMA_FENCE_FLAG_USER_BITS, &old->flags))) { + RCU_INIT_POINTER(*ptr, NULL); + }Is this to avoid premature job free because of
dma_fence_putinsideamdgpu_fence_process?
I can't currently remember why but we probably want all the HW fences currently in the ring to
be forced signaled - maybe better to test forDMA_FENCE_FLAG_USER_BITSinsideamdgpu_fence_process
and still do the signaling but not thedma_fence_putpartQuoted 7 lines
Quoted 1 line
Andrey
Hi Andrey,
This is to avoid signaling the same fence twice. If we still do the signaling, then the job in the pending list will be signaled first in
force_completion, and later be signaled in resubmit. This will go toBUG()inamdgpu_fence_process.Quoted 7 lines
Oh, i see, how about just adding 'skip' flag to
amdgpu_ringand setting it before callingamdgpu_fence_driver_force_completionand resetting it after, then insideamdgpu_fence_driver_force_completion
you can just skip the signaling part with this flag for fences withDMA_FENCE_FLAG_USER_BITSset
Less lines of code at least.Quoted 9 lines
Still sounds quite a bit hacky.
I would rather suggest to completely drop the approach with
amdgpu_fence_driver_force_completion(). I could never see why we would want that in the first place.Regards,
Christian.Hi Christian,
Quoted 11 lines
I keep the
amdgpu_fence_driver_force_completionhere to make sure the vm flush fence is signaled and put.
So the key question is whether the fence of ib test should be the first fence to be signaled after reset.
If it should be, it means not only fences withDMA_FENCE_FLAG_USER_BITSbut also vm flush fences should be removed from RCU fence array beforeib_test. Then we must do theforce_completionhere for vm flush fence, otherwise there will be a memory leak here as no one will signal and put it after we remove it from RCU.
If not, then the first fence to signal could be vm flush fence. And I'm OK with dropamdgpu_fence_driver_force_completionhere.
Quoted 6 lines
The key problem is that
amdgpu_fence_driver_force_completion()goes over the RCU array and signals everything there.What we should do instead is to go over the jobs and cleanup the ones we don't want to re-submit and keep the rest.
Hi Christian,
The thing is vm flush fence has no job passed to amdgpu_fence_emit, so go through the jobs cannot help find the vm flush fence. And keep the rest fences in the RCU array will lead to signaling them in the ib_test right after ASIC reset. While they will be signaled again during resubmission. What I'm doning here is just trying to cleanup the fences without a parent job and make sure the rest fences won't be signaled twice.
Quoted 3 lines
And please drop any usage of
DMA_FENCE_FLAG_USER_BITS. That is not something which should be abused for reset handling.
The DMA_FENCE_FLAG_USER_BITS here is to mark this fence has a parent job. If this is not a proper usage here, do you have any suggestions about how to identify whether the fence has a parent job?
Best Regards,
JingWen Chen
Quoted 89 lines
Regards,
Christian.Quoted 86 lines
Best Regards,
JingWen ChenQuoted 82 lines
Quoted 82 lines
Andrey
Quoted 79 lines
Quoted 20 lines
Quoted 20 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1
-0+ } /* after all hw jobs are reset, hw fence ismeaningless, so
force_completion*/
amdgpu_fence_driver_force_completion(ring); }drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +0
-0@@ -156,11 +156,17 @@ int amdgpu_fence_emit(struct
amdgpu_ring*ring, structdma_fence**f, struct amd
job->ring= ring;
}
- seq = ++ring->fence_drv.sync_seq;
- dma_fence_init(fence,&amdgpu_fence_ops, -&ring->fence_drv.lock,
-adev->fence_context+ring->idx, - seq);
+ if (job != NULL &&job->base.resubmit_flag== 1) {Quoted 8 lines
Quoted 5 lines
+ /* reinit seq for resubmitted jobs */
+ seq = ++ring->fence_drv.sync_seq;
+fence->seqno= seq;
+ } else {
+ seq = ++ring->fence_drv.sync_seq;Seems like you could do the above line only once above if-else as it was
beforeSure, I will modify this.
Best Regards,
JingWen ChenQuoted 46 lines
Quoted 46 lines
+ dma_fence_init(fence,
&amdgpu_fence_ops, +&ring->fence_drv.lock,
+adev->fence_context+ring->idx, + seq);
+ }
if (job != NULL) {
/* mark this fence has a parent job */drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +0
-0@@ -241,6 +241,7 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job) dma_fence_set_error(finished, -ECANCELED);/* skipIB as well if VRAM lost */
if (finished->error< 0) {
+dma_fence_put(&job->hw_fence);
DRM_INFO("Skip scheduling IBs!\n");
} else {
r = amdgpu_ib_schedule(ring,job->num_ibs,job->ibs, job,drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +2
-1@@ -249,7 +250,8 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job) DRM_ERROR("Error scheduling IBs (%d)\n", r); } - dma_fence_get(fence); + if (!job->base.resubmit_flag) + dma_fence_get(fence); amdgpu_job_free_resources(job); fence = r ? ERR_PTR(r) : fence;drivers/gpu/drm/scheduler/sched_main.c +1
-0@@ -544,6 +544,7 @@ void drm_sched_resubmit_jobs_ext(struct drm_gpu_scheduler *sched, int max) dma_fence_set_error(&s_fence->finished, -ECANCELED); dma_fence_put(s_job->s_fence->parent); + s_job->resubmit_flag = 1; fence = sched->ops->run_job(s_job); i++;include/drm/gpu_scheduler.h +0
-0
Quoted 9 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
include/drm/gpu_scheduler.h struct drm_sched_job { · line 204
Quoted 8 lines
include/drm/gpu_scheduler.h +1
-0@@ -198,6 +198,7 @@ struct drm_sched_job { enum drm_sched_priority s_priority; struct drm_sched_entity *entity; struct dma_fence_cb cb; + int resubmit_flag; }; static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job,
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
RE: [PATCH 2/2] drm: add tdr support for embeded hw_fence
[AMD Official Use Only]
Hi Christian
How about this way:
#define AMDGPU_DMA_FENCE_FLAG_INSIDE_JOB DMA_FENCE_FLAG_USER_BITS
And we can use AMDGPU_DMA_FENCE_FLAG_INSIDE_JOB instead of DMA_FENCE_FLAG_USER_BITS everywhere
Thanks
------------------------------------------
Monk Liu | Cloud-GPU Core team
------------------------------------------
-----Original Message-----
From: Jingwen Chen <[email protected]>
Sent: Friday, July 23, 2021 3:07 PM
To: Christian König <[email protected]>; Grodzovsky, Andrey <[email protected]>; [email protected]
Cc: Chen, Horace <[email protected]>; Liu, Monk <[email protected]>; Zhang, Jack (Jian) <[email protected]>
Subject: Re: [PATCH 2/2] drm: add tdr support for embeded hw_fence
Quoted 110 lines · Christian König
Am 22.07.21 um 18:47 schrieb Jingwen Chen:
Quoted 97 lines · Christian König
Quoted 85 lines
Am 22.07.21 um 16:45 schrieb Andrey Grodzovsky:
Quoted 78 lines · Jingwen Chen
Quoted 68 lines · Andrey Grodzovsky
Quoted 67 lines · Jingwen Chen
Quoted 55 lines
[Why]
After embededhw_fencetoamdgpu_job, we need to add tdr support for this feature.[How]
1. Add aresubmit_flagfor resubmit jobs.
2. Clear job fence from RCU and force complete vm flush fences in
pre_asic_reset
3. skipdma_fence_getfor resubmit jobs and add adma_fence_put
for guilty jobs.Signed-off-by Jack Zhang <[email protected]>Signed-off-by Jingwen Chen <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 12 +++++++++++- drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 16 +++++++++++----- drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 4 +++- drivers/gpu/drm/scheduler/sched_main.c | 1 + include/drm/gpu_scheduler.h | 1 + 5 files changed, 27 insertions(+), 7 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1
-1@@ -4382,7 +4382,7 @@ int amdgpu_device_mode1_reset(struct amdgpu_device *adev) int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev, struct amdgpu_reset_context *reset_context) { - int i, r = 0; + int i, j, r = 0;struct
amdgpu_job*job = NULL;
boolneed_full_reset= test_bit(AMDGPU_NEED_FULL_RESET,&reset_context->flags); @@ -4406,6 +4406,16 @@ int amdgpu_device_pre_asic_reset(structamdgpu_device*adev,
if (!ring || !ring->sched.thread) continue;
+ /*clear job fence from fence drv to avoiddrivers/gpu/drm/amd/amdgpu/amdgpu_device.c +8
-0+force_completion +*leave NULL and vm flush fence in fence drv */ + for (j = 0; j <= ring->fence_drv.num_fences_mask; +j ++) { + struct dma_fence *old,**ptr; + ptr = &ring->fence_drv.fences[j]; + old = rcu_dereference_protected(*ptr, 1); + if (old && test_bit(DMA_FENCE_FLAG_USER_BITS,Quoted 3 lines
&old->flags))) {
+RCU_INIT_POINTER(*ptr, NULL); + }Is this to avoid premature job free because of
dma_fence_putinsideamdgpu_fence_process?
I can't currently remember why but we probably want all the HW fences currently in the ring to be forced signaled - maybe better to test forDMA_FENCE_FLAG_USER_BITSinsideamdgpu_fence_processand still do the signaling but not thedma_fence_putpartQuoted 8 lines
Quoted 1 line
Andrey
Hi Andrey,
This is to avoid signaling the same fence twice. If we still do the signaling, then the job in the pending list will be signaled first in
force_completion, and later be signaled in resubmit. This will go toBUG()inamdgpu_fence_process.Quoted 6 lines
Oh, i see, how about just adding 'skip' flag to
amdgpu_ringand setting it before callingamdgpu_fence_driver_force_completionand resetting it after, then insideamdgpu_fence_driver_force_completion
you can just skip the signaling part with this flag for fences withDMA_FENCE_FLAG_USER_BITSset Less lines of code at least.Quoted 9 lines
Still sounds quite a bit hacky.
I would rather suggest to completely drop the approach with
amdgpu_fence_driver_force_completion(). I could never see why we would want that in the first place.Regards,
Christian.Hi Christian,
Quoted 11 lines
I keep the
amdgpu_fence_driver_force_completionhere to make sure the vm flush fence is signaled and put.
So the key question is whether the fence of ib test should be the first fence to be signaled after reset.
If it should be, it means not only fences withDMA_FENCE_FLAG_USER_BITSbut also vm flush fences should be removed from RCU fence array beforeib_test. Then we must do theforce_completionhere for vm flush fence, otherwise there will be a memory leak here as no one will signal and put it after we remove it from RCU. If not, then the first fence to signal could be vm flush fence. And I'm OK with dropamdgpu_fence_driver_force_completionhere.
Quoted 6 lines
The key problem is that
amdgpu_fence_driver_force_completion()goes over the RCU array and signals everything there.What we should do instead is to go over the jobs and cleanup the ones we don't want to re-submit and keep the rest.
Hi Christian,
The thing is vm flush fence has no job passed to amdgpu_fence_emit, so go through the jobs cannot help find the vm flush fence. And keep the rest fences in the RCU array will lead to signaling them in the ib_test right after ASIC reset. While they will be signaled again during resubmission. What I'm doning here is just trying to cleanup the fences without a parent job and make sure the rest fences won't be signaled twice.
Quoted 3 lines
And please drop any usage of
DMA_FENCE_FLAG_USER_BITS. That is not something which should be abused for reset handling.
The DMA_FENCE_FLAG_USER_BITS here is to mark this fence has a parent job. If this is not a proper usage here, do you have any suggestions about how to identify whether the fence has a parent job?
Best Regards,
JingWen Chen
Quoted 88 lines
Regards,
Christian.Quoted 85 lines
Best Regards,
JingWen ChenQuoted 81 lines
Quoted 81 lines
Andrey
Quoted 78 lines
Quoted 19 lines
Quoted 19 lines
+ }
/* after all hw jobs are reset, hw fence is
meaningless, soforce_completion*/
amdgpu_fence_driver_force_completion(ring); }drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +0
-0@@ -156,11 +156,17 @@ int amdgpu_fence_emit(struct
amdgpu_ring*ring, structdma_fence**f, struct amd
job->ring= ring;
}
- seq = ++ring->fence_drv.sync_seq;
- dma_fence_init(fence,&amdgpu_fence_ops, -&ring->fence_drv.lock,
-adev->fence_context+ring->idx, - seq);Quoted 8 lines
Quoted 6 lines
+ if (job != NULL &&
job->base.resubmit_flag== 1) { + /* reinit seq for resubmitted jobs */ + seq = ++ring->fence_drv.sync_seq;
+fence->seqno= seq;
+ } else {
+ seq = ++ring->fence_drv.sync_seq;Seems like you could do the above line only once above if-else as it was before
Sure, I will modify this.
Best Regards,
JingWen ChenQuoted 46 lines
Quoted 46 lines
+ dma_fence_init(fence,
&amdgpu_fence_ops, +&ring->fence_drv.lock,
+adev->fence_context+ring->idx, + seq);
+ }
if (job != NULL) {
/* mark this fence has a parent job */drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +0
-0@@ -241,6 +241,7 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job) dma_fence_set_error(finished, -ECANCELED);/* skipIB as well if VRAM lost */
if (finished->error< 0) {
+dma_fence_put(&job->hw_fence);
DRM_INFO("Skip scheduling IBs!\n");
} else {
r = amdgpu_ib_schedule(ring,job->num_ibs,job->ibs, job,drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +2
-1@@ -249,7 +250,8 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job) DRM_ERROR("Error scheduling IBs (%d)\n", r); } - dma_fence_get(fence); + if (!job->base.resubmit_flag) + dma_fence_get(fence); amdgpu_job_free_resources(job); fence = r ? ERR_PTR(r) : fence;drivers/gpu/drm/scheduler/sched_main.c +1
-0@@ -544,6 +544,7 @@ void drm_sched_resubmit_jobs_ext(struct drm_gpu_scheduler *sched, int max) dma_fence_set_error(&s_fence->finished, -ECANCELED); dma_fence_put(s_job->s_fence->parent); + s_job->resubmit_flag = 1; fence = sched->ops->run_job(s_job); i++;include/drm/gpu_scheduler.h +0
-0
Quoted 9 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
include/drm/gpu_scheduler.h struct drm_sched_job { · line 204
Quoted 8 lines
include/drm/gpu_scheduler.h +1
-0@@ -198,6 +198,7 @@ struct drm_sched_job { enum drm_sched_priority s_priority; struct drm_sched_entity *entity; struct dma_fence_cb cb; + int resubmit_flag; }; static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job,
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Christian König
reviewerRe: [PATCH 2/2] drm: add tdr support for embeded hw_fence
In reply to Jingwen ChenAm 23.07.21 um 09:07 schrieb Jingwen Chen:
Quoted 10 lines
[SNIP]
Hi Christian,The thing is vm flush fence has no job passed to
amdgpu_fence_emit, so go through the jobs cannot help find the vm flush fence. And keep the rest fences in the RCU array will lead to signaling them in theib_testright after ASIC reset. While they will be signaled again during resubmission. What I'm doning here is just trying to cleanup the fences without a parent job and make sure the rest fences won't be signaled twice.
It took me a moment to realize what you are talking about here.
This is for the KIQ! You need different handling for the KIQ than for the scheduler controlled rings.
It is not only the flush jobs, but the KIQ needs a complete reset because of the register writes pushed there as well.
Quoted 6 lines
Quoted 3 lines
And please drop any usage of
DMA_FENCE_FLAG_USER_BITS. That is not something which should be abused for reset handling.The
DMA_FENCE_FLAG_USER_BITShere is to mark this fence has a parent job. If this is not a proper usage here, do you have any suggestions about how to identify whether the fence has a parent job?
You don't need to mark the fences at all. Everything on the KIQ ring needs to be force completed since none of the fences on that ring have a parent job.
Christian.
Quoted 91 lines
Best Regards,
JingWen ChenQuoted 88 lines
Regards,
Christian.Quoted 85 lines
Best Regards,
JingWen ChenQuoted 82 lines
Quoted 82 lines
Andrey
Quoted 79 lines
Quoted 20 lines
Quoted 20 lines
+ }
/* after all hw jobs are reset, hw fence is
meaningless, soforce_completion*/
amdgpu_fence_driver_force_completion(ring); }drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +0
-0@@ -156,11 +156,17 @@ int amdgpu_fence_emit(struct
amdgpu_ring*ring, structdma_fence**f, struct amd
job->ring= ring;
}
- seq = ++ring->fence_drv.sync_seq;
- dma_fence_init(fence,&amdgpu_fence_ops, -&ring->fence_drv.lock,
-adev->fence_context+ring->idx, - seq);
+ if (job != NULL &&job->base.resubmit_flag== 1) {Quoted 8 lines
Quoted 5 lines
+ /* reinit seq for resubmitted jobs */
+ seq = ++ring->fence_drv.sync_seq;
+fence->seqno= seq;
+ } else {
+ seq = ++ring->fence_drv.sync_seq;Seems like you could do the above line only once above if-else as it was
beforeSure, I will modify this.
Best Regards,
JingWen ChenQuoted 46 lines
Quoted 46 lines
+ dma_fence_init(fence,
&amdgpu_fence_ops, +&ring->fence_drv.lock,
+adev->fence_context+ring->idx, + seq);
+ }
if (job != NULL) {
/* mark this fence has a parent job */drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +0
-0@@ -241,6 +241,7 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job) dma_fence_set_error(finished, -ECANCELED);/* skipIB as well if VRAM lost */
if (finished->error< 0) {
+dma_fence_put(&job->hw_fence);
DRM_INFO("Skip scheduling IBs!\n");
} else {
r = amdgpu_ib_schedule(ring,job->num_ibs,job->ibs, job,drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +2
-1@@ -249,7 +250,8 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job) DRM_ERROR("Error scheduling IBs (%d)\n", r); } - dma_fence_get(fence); + if (!job->base.resubmit_flag) + dma_fence_get(fence); amdgpu_job_free_resources(job); fence = r ? ERR_PTR(r) : fence;drivers/gpu/drm/scheduler/sched_main.c +1
-0@@ -544,6 +544,7 @@ void drm_sched_resubmit_jobs_ext(struct drm_gpu_scheduler *sched, int max) dma_fence_set_error(&s_fence->finished, -ECANCELED); dma_fence_put(s_job->s_fence->parent); + s_job->resubmit_flag = 1; fence = sched->ops->run_job(s_job); i++;include/drm/gpu_scheduler.h +0
-0
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
include/drm/gpu_scheduler.h struct drm_sched_job { · line 204
Quoted 8 lines
include/drm/gpu_scheduler.h +1
-0@@ -198,6 +198,7 @@ struct drm_sched_job { enum drm_sched_priority s_priority; struct drm_sched_entity *entity; struct dma_fence_cb cb; + int resubmit_flag; }; static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job,
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Re: [PATCH 2/2] drm: add tdr support for embeded hw_fence
Quoted 27 lines · Christian König
Am 23.07.21 um 09:07 schrieb Jingwen Chen:
Quoted 10 lines
[SNIP]
Hi Christian,The thing is vm flush fence has no job passed to
amdgpu_fence_emit, so go through the jobs cannot help find the vm flush fence. And keep the rest fences in the RCU array will lead to signaling them in theib_testright after ASIC reset. While they will be signaled again during resubmission. What I'm doning here is just trying to cleanup the fences without a parent job and make sure the rest fences won't be signaled twice.It took me a moment to realize what you are talking about here.
This is for the KIQ! You need different handling for the KIQ than for the scheduler controlled rings.
It is not only the flush jobs, but the KIQ needs a complete reset because of the register writes pushed there as well.
Quoted 6 lines
Quoted 3 lines
And please drop any usage of
DMA_FENCE_FLAG_USER_BITS. That is not something which should be abused for reset handling.The
DMA_FENCE_FLAG_USER_BITShere is to mark this fence has a parent job. If this is not a proper usage here, do you have any suggestions about how to identify whether the fence has a parent job?
Quoted 6 lines
You don't need to mark the fences at all. Everything on the KIQ ring needs to be force completed since none of the fences on that ring have a parent job.
Christian.
Hi Christian
Yes KIQ ring fences all need force_completion. But we do need to mark the fences. Say we have a gfx ring job with vm_flush=1 sending to amdgpu_ib_schedule, then in amdgpu_vm_flush, after the amdgpu_ring_emit_vm_flush is done, we will create a vm flush fence on gfx ring with amdgpu_fence_emit(ring, &fence, NULL, 0).
Then this vm flush fence we create here has no parent job but it's on gfx ring.
If we only do force_completion for KIQ ring and just clear the RCU array for the scheduler controlled rings, nobody will signal and put this gfx ring vm_flush fence again. When this job is resubmitted, it will just create a new vm_flush fence. This is a memleak and I have seen this memleak during my test.
Best Regards,
JingWen Chen
Quoted 91 lines
Quoted 91 lines
Best Regards,
JingWen ChenQuoted 88 lines
Regards,
Christian.Quoted 85 lines
Best Regards,
JingWen ChenQuoted 82 lines
Quoted 82 lines
Andrey
Quoted 79 lines
Quoted 20 lines
Quoted 20 lines
+ }
/* after all hw jobs are reset, hw fence is
meaningless, soforce_completion*/
amdgpu_fence_driver_force_completion(ring); }drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +0
-0@@ -156,11 +156,17 @@ int amdgpu_fence_emit(struct
amdgpu_ring*ring, structdma_fence**f, struct amd
job->ring= ring;
}
- seq = ++ring->fence_drv.sync_seq;
- dma_fence_init(fence,&amdgpu_fence_ops, -&ring->fence_drv.lock,
-adev->fence_context+ring->idx, - seq);
+ if (job != NULL &&job->base.resubmit_flag== 1) {Quoted 8 lines
Quoted 5 lines
+ /* reinit seq for resubmitted jobs */
+ seq = ++ring->fence_drv.sync_seq;
+fence->seqno= seq;
+ } else {
+ seq = ++ring->fence_drv.sync_seq;Seems like you could do the above line only once above if-else as it was
beforeSure, I will modify this.
Best Regards,
JingWen ChenQuoted 46 lines
Quoted 46 lines
+ dma_fence_init(fence,
&amdgpu_fence_ops, +&ring->fence_drv.lock,
+adev->fence_context+ring->idx, + seq);
+ }
if (job != NULL) {
/* mark this fence has a parent job */drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +0
-0@@ -241,6 +241,7 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job) dma_fence_set_error(finished, -ECANCELED);/* skipIB as well if VRAM lost */
if (finished->error< 0) {
+dma_fence_put(&job->hw_fence);
DRM_INFO("Skip scheduling IBs!\n");
} else {
r = amdgpu_ib_schedule(ring,job->num_ibs,job->ibs, job,drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +2
-1@@ -249,7 +250,8 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job) DRM_ERROR("Error scheduling IBs (%d)\n", r); } - dma_fence_get(fence); + if (!job->base.resubmit_flag) + dma_fence_get(fence); amdgpu_job_free_resources(job); fence = r ? ERR_PTR(r) : fence;drivers/gpu/drm/scheduler/sched_main.c +1
-0@@ -544,6 +544,7 @@ void drm_sched_resubmit_jobs_ext(struct drm_gpu_scheduler *sched, int max) dma_fence_set_error(&s_fence->finished, -ECANCELED); dma_fence_put(s_job->s_fence->parent); + s_job->resubmit_flag = 1; fence = sched->ops->run_job(s_job); i++;include/drm/gpu_scheduler.h +0
-0
Quoted 9 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
include/drm/gpu_scheduler.h struct drm_sched_job { · line 204
Quoted 8 lines
include/drm/gpu_scheduler.h +1
-0@@ -198,6 +198,7 @@ struct drm_sched_job { enum drm_sched_priority s_priority; struct drm_sched_entity *entity; struct dma_fence_cb cb; + int resubmit_flag; }; static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job,
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Christian König
reviewerRe: [PATCH 2/2] drm: add tdr support for embeded hw_fence
Am 23.07.21 um 11:25 schrieb Jingwen Chen:
Quoted 32 lines · Christian König
Quoted 22 lines
Am 23.07.21 um 09:07 schrieb Jingwen Chen:
Quoted 10 lines
[SNIP]
Hi Christian,The thing is vm flush fence has no job passed to
amdgpu_fence_emit, so go through the jobs cannot help find the vm flush fence. And keep the rest fences in the RCU array will lead to signaling them in theib_testright after ASIC reset. While they will be signaled again during resubmission. What I'm doning here is just trying to cleanup the fences without a parent job and make sure the rest fences won't be signaled twice.It took me a moment to realize what you are talking about here.
This is for the KIQ! You need different handling for the KIQ than for the scheduler controlled rings.
It is not only the flush jobs, but the KIQ needs a complete reset because of the register writes pushed there as well.
Quoted 3 lines
Quoted 3 lines
And please drop any usage of
DMA_FENCE_FLAG_USER_BITS. That is not something which should be abused for reset handling.Quoted 9 lines
Quoted 3 lines
The
DMA_FENCE_FLAG_USER_BITShere is to mark this fence has a parent job. If this is not a proper usage here, do you have any suggestions about how to identify whether the fence has a parent job?You don't need to mark the fences at all. Everything on the KIQ ring needs to be force completed since none of the fences on that ring have a parent job.
Christian.
Quoted 7 lines
Hi Christian
Yes KIQ ring fences all need
force_completion. But we do need to mark the fences. Say we have a gfx ring job with vm_flush=1 sending toamdgpu_ib_schedule, then inamdgpu_vm_flush, after theamdgpu_ring_emit_vm_flushis done, we will create a vm flush fence on gfx ring withamdgpu_fence_emit(ring, &fence, NULL, 0).
That is illegal and needs to be fixed as well.
Quoted 7 lines
Then this vm flush fence we create here has no parent job but it's on gfx ring.
If we only doforce_completionfor KIQ ring and just clear the RCU array for the scheduler controlled rings, nobody will signal and put this gfx ringvm_flushfence again. When this job is resubmitted, it will just create a newvm_flushfence. This is a memleak and I have seen this memleak during my test.
At least I'm better understanding the problem now as well.
But you are just trying to circumvent symptoms and not fixing the root cause.
See any memory allocation during command submission must be prevented. This also includes the flush fence.
Regards,
Christian.
Quoted 92 lines
Best Regards,
JingWen ChenQuoted 89 lines
Quoted 89 lines
Best Regards,
JingWen ChenQuoted 87 lines
Regards,
Christian.Quoted 84 lines
Best Regards,
JingWen ChenQuoted 82 lines
Quoted 82 lines
Andrey
Quoted 79 lines
Quoted 20 lines
Quoted 20 lines
+ }
/* after all hw jobs are reset, hw fence is
meaningless, soforce_completion*/amdgpu_fence_driver_force_completion(ring); }drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +0
-0@@ -156,11 +156,17 @@ int amdgpu_fence_emit(struct
amdgpu_ring*ring, structdma_fence**f, struct amdjob->ring = ring; }- seq = ++ring->fence_drv.sync_seq;
- dma_fence_init(fence,&amdgpu_fence_ops, -&ring->fence_drv.lock,
-adev->fence_context+ring->idx, - seq);
+ if (job != NULL &&job->base.resubmit_flag== 1) {Quoted 8 lines
Quoted 5 lines
+ /* reinit seq for resubmitted jobs */
+ seq = ++ring->fence_drv.sync_seq;
+fence->seqno= seq;
+ } else {
+ seq = ++ring->fence_drv.sync_seq;Seems like you could do the above line only once above if-else as it was
beforeSure, I will modify this.
Best Regards,
JingWen ChenQuoted 46 lines
Quoted 46 lines
+ dma_fence_init(fence,
&amdgpu_fence_ops,
+&ring->fence_drv.lock,
+adev->fence_context+ring->idx,
+ seq);
+ }if (job != NULL) { /* mark this fence has a parent job */drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +0
-0@@ -241,6 +241,7 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job) dma_fence_set_error(finished, -ECANCELED);/* skipIB as well if VRAM lost */
if (finished->error< 0) {+ dma_fence_put(&job->hw_fence); DRM_INFO("Skip scheduling IBs!\n"); } else { r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs, job,drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +2
-1@@ -249,7 +250,8 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job) DRM_ERROR("Error scheduling IBs (%d)\n", r); } - dma_fence_get(fence); + if (!job->base.resubmit_flag) + dma_fence_get(fence); amdgpu_job_free_resources(job); fence = r ? ERR_PTR(r) : fence;drivers/gpu/drm/scheduler/sched_main.c +1
-0@@ -544,6 +544,7 @@ void drm_sched_resubmit_jobs_ext(struct drm_gpu_scheduler *sched, int max) dma_fence_set_error(&s_fence->finished, -ECANCELED); dma_fence_put(s_job->s_fence->parent); + s_job->resubmit_flag = 1; fence = sched->ops->run_job(s_job); i++;include/drm/gpu_scheduler.h +0
-0
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
include/drm/gpu_scheduler.h struct drm_sched_job { · line 204
Quoted 8 lines
include/drm/gpu_scheduler.h +1
-0@@ -198,6 +198,7 @@ struct drm_sched_job { enum drm_sched_priority s_priority; struct drm_sched_entity *entity; struct dma_fence_cb cb; + int resubmit_flag; }; static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job,
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx