drm/scheduler: Unwrap job dependencies
Subsystems (4)
RFC 1/1 drm/scheduler: Unwrap job dependencies
From: Rob Clark <[email protected]>
Container fences have burner contexts, which makes the trick to store at most one fence per context somewhat useless if we don't unwrap array or chain fences.
---
tbh, I'm not sure why we weren't doing this already, unless there is something I'm overlooking
drivers/gpu/drm/scheduler/sched_main.c | 43 +++++++++++++++++--------- 1 file changed, 28 insertions(+), 15 deletions(-)
drivers/gpu/drm/scheduler/sched_main.c +28 -15
@@ -41,20 +41,21 @@* 4. Entities themselves maintain a queue of jobs that will be scheduled on * the hardware. * * The jobs in a entity are always scheduled in the order that they were pushed. */ #include <linux/kthread.h> #include <linux/wait.h> #include <linux/sched.h> #include <linux/completion.h> +#include <linux/dma-fence-unwrap.h> #include <linux/dma-resv.h> #include <uapi/linux/sched/types.h> #include <drm/drm_print.h> #include <drm/drm_gem.h> #include <drm/gpu_scheduler.h> #include <drm/spsc_queue.h> #define CREATE_TRACE_POINTS #include "gpu_scheduler_trace.h"
@@ -665,41 +666,27 @@ void drm_sched_job_arm(struct drm_sched_job *job) sched = entity->rq->sched; job->sched = sched; job->s_priority = entity->rq - sched->sched_rq; job->id =atomic64_inc_return (&sched->job_id_count);drm_sched_fence_init (job->s_fence, job->entity); }EXPORT_SYMBOL (drm_sched_job_arm); -/** -* drm_sched_job_add_dependency - adds the fence as a job dependency -* @job: scheduler job to add the dependencies to -* @fence: the dma_fence to add to the list of dependencies. -* -* Note that @fence is consumed in both the success and error cases. -* -* Returns: -* 0 on success, or an error on failing to expand the array. -*/ -int drm_sched_job_add_dependency (struct drm_sched_job *job, -struct dma_fence *fence) +static int _add_dependency (struct drm_sched_job *job,struct dma_fence *fence) {struct dma_fence *entry;unsigned long index;u32 id =0 ;int ret; -if (!fence) -return 0 ; -/* Deduplicate if we already depend on a fence from the same context. * This lets the size of the array of deps scale with the number of * engines involved, rather than the number of BOs. */ xa_for_each (&job->dependencies, index, entry) {if (entry->context != fence->context)continue ;if (dma_fence_is_later (fence, entry)) {dma_fence_put (entry);
@@ -709,20 +696,46 @@ int drm_sched_job_add_dependency(struct drm_sched_job *job, }return 0 ; } ret =xa_alloc (&job->dependencies, &id, fence, xa_limit_32b,GFP_KERNEL );if (ret !=0 )dma_fence_put (fence);return ret; } + +/** +* drm_sched_job_add_dependency - adds the fence as a job dependency +* @job: scheduler job to add the dependencies to +* @fence: the dma_fence to add to the list of dependencies. +* +* Note that @fence is consumed in both the success and error cases. +* +* Returns: +* 0 on success, or an error on failing to expand the array. +*/ +int drm_sched_job_add_dependency (struct drm_sched_job *job, +struct dma_fence *fence) +{ +struct dma_fence_unwrap iter; +struct dma_fence *f; +int ret =0 ; + + dma_fence_unwrap_for_each (f, &iter, fence) { + ret =_add_dependency (job, f); +if (ret) +break ; + } + +return ret; +}EXPORT_SYMBOL (drm_sched_job_add_dependency);/** * drm_sched_job_add_resv_dependencies - add all fences from the resv to the job * @job: scheduler job to add the dependencies to * @resv: the dma_resv object to get the fences from * @usage: the dma_resv_usage to use to filter the fences * * This adds all fences matching the given usage from @resv to @job. * Must be called with the @resv lock held.
Signature · 1 line
2.39.2
Am 22.03.23 um 23:44 schrieb Rob Clark:
Quoted 5 lines
From: Rob Clark <[email protected]>
Container fences have burner contexts, which makes the trick to store at most one fence per context somewhat useless if we don't unwrap array or chain fences.
Mhm, we intentionally kept them not unwrapped since this way they only occupy one fence slot.
But it might be better to unwrap them if you add many of those dependencies.
Quoted 49 lines
Signed-off-by Rob Clark <[email protected]>---
tbh, I'm not sure why we weren't doing this already, unless there is something I'm overlookingdrivers/gpu/drm/scheduler/sched_main.c | 43 +++++++++++++++++--------- 1 file changed, 28 insertions(+), 15 deletions(-)drivers/gpu/drm/scheduler/sched_main.c +1
-3@@ -41,20 +41,21 @@* 4. Entities themselves maintain a queue of jobs that will be scheduled on * the hardware. * * The jobs in a entity are always scheduled in the order that they were pushed. */ #include <linux/kthread.h> #include <linux/wait.h> #include <linux/sched.h> #include <linux/completion.h> +#include <linux/dma-fence-unwrap.h> #include <linux/dma-resv.h> #include <uapi/linux/sched/types.h> #include <drm/drm_print.h> #include <drm/drm_gem.h> #include <drm/gpu_scheduler.h> #include <drm/spsc_queue.h> #define CREATE_TRACE_POINTS #include "gpu_scheduler_trace.h" @@ -665,41 +666,27 @@ void drm_sched_job_arm(struct drm_sched_job *job) sched = entity->rq->sched; job->sched = sched; job->s_priority = entity->rq - sched->sched_rq; job->id =atomic64_inc_return (&sched->job_id_count);drm_sched_fence_init (job->s_fence, job->entity); }EXPORT_SYMBOL (drm_sched_job_arm); -/** -* drm_sched_job_add_dependency - adds the fence as a job dependency -* @job: scheduler job to add the dependencies to
drivers/gpu/drm/scheduler/sched_main.c void drm_sched_job_arm(struct drm_sched_job *job) · line 676
Quoted 10 lines
drivers/gpu/drm/scheduler/sched_main.c +1
-9-* @fence: the dma_fence to add to the list of dependencies. -* -* Note that @fence is consumed in both the success and error cases. -* -* Returns: -* 0 on success, or an error on failing to expand the array. -*/ -int drm_sched_job_add_dependency (struct drm_sched_job *job, -struct dma_fence *fence) +static int _add_dependency (struct drm_sched_job *job,struct dma_fence *fence)
Please keep the drm_sched_job_ prefix here even for static functions. The symbol _add_dependency just sucks in a backtrace, especially when it's tail optimized.
Quoted 33 lines
drivers/gpu/drm/scheduler/sched_main.c +3
-3{struct dma_fence *entry;unsigned long index;u32 id =0 ;int ret; -if (!fence) -return 0 ; -/* Deduplicate if we already depend on a fence from the same context. * This lets the size of the array of deps scale with the number of * engines involved, rather than the number of BOs. */ xa_for_each (&job->dependencies, index, entry) {if (entry->context != fence->context)continue ;if (dma_fence_is_later (fence, entry)) {dma_fence_put (entry);@@ -709,20 +696,46 @@ int drm_sched_job_add_dependency(struct drm_sched_job *job, }return 0 ; } ret =xa_alloc (&job->dependencies, &id, fence, xa_limit_32b,GFP_KERNEL );if (ret !=0 )dma_fence_put (fence);return ret; } + +/** +* drm_sched_job_add_dependency - adds the fence as a job dependency
drivers/gpu/drm/scheduler/sched_main.c int drm_sched_job_add_dependency(struct drm_sched_job *job, · line 718
Quoted 10 lines
drivers/gpu/drm/scheduler/sched_main.c +10
-0+* @job: scheduler job to add the dependencies to +* @fence: the dma_fence to add to the list of dependencies. +* +* Note that @fence is consumed in both the success and error cases. +* +* Returns: +* 0 on success, or an error on failing to expand the array. +*/ +int drm_sched_job_add_dependency (struct drm_sched_job *job, +struct dma_fence *fence)
Maybe name the new function drm_sched_job_unwrap_add_dependency or something like this.
I need to double check, but I think for some cases we don't need or don't even want this in the driver.
Christian.
Quoted 23 lines
drivers/gpu/drm/scheduler/sched_main.c +13
-0+{ +struct dma_fence_unwrap iter; +struct dma_fence *f; +int ret =0 ; + + dma_fence_unwrap_for_each (f, &iter, fence) { + ret =_add_dependency (job, f); +if (ret) +break ; + } + +return ret; +}EXPORT_SYMBOL (drm_sched_job_add_dependency);/** * drm_sched_job_add_resv_dependencies - add all fences from the resv to the job * @job: scheduler job to add the dependencies to * @resv: the dma_resv object to get the fences from * @usage: the dma_resv_usage to use to filter the fences * * This adds all fences matching the given usage from @resv to @job. * Must be called with the @resv lock held.
Quoted 124 lines · Christian König
Am 22.03.23 um 23:44 schrieb Rob Clark:
Quoted 5 lines
From: Rob Clark <[email protected]>
Container fences have burner contexts, which makes the trick to store at most one fence per context somewhat useless if we don't unwrap array or chain fences.
Mhm, we intentionally kept them not unwrapped since this way they only occupy one fence slot.
But it might be better to unwrap them if you add many of those dependencies.
Quoted 49 lines
Signed-off-by Rob Clark <[email protected]>---
tbh, I'm not sure why we weren't doing this already, unless there is something I'm overlookingdrivers/gpu/drm/scheduler/sched_main.c | 43 +++++++++++++++++--------- 1 file changed, 28 insertions(+), 15 deletions(-)drivers/gpu/drm/scheduler/sched_main.c +1
-3@@ -41,20 +41,21 @@* 4. Entities themselves maintain a queue of jobs that will be scheduled on * the hardware. * * The jobs in a entity are always scheduled in the order that they were pushed. */ #include <linux/kthread.h> #include <linux/wait.h> #include <linux/sched.h> #include <linux/completion.h> +#include <linux/dma-fence-unwrap.h> #include <linux/dma-resv.h> #include <uapi/linux/sched/types.h> #include <drm/drm_print.h> #include <drm/drm_gem.h> #include <drm/gpu_scheduler.h> #include <drm/spsc_queue.h> #define CREATE_TRACE_POINTS #include "gpu_scheduler_trace.h" @@ -665,41 +666,27 @@ void drm_sched_job_arm(struct drm_sched_job *job) sched = entity->rq->sched; job->sched = sched; job->s_priority = entity->rq - sched->sched_rq; job->id =atomic64_inc_return (&sched->job_id_count);drm_sched_fence_init (job->s_fence, job->entity); }EXPORT_SYMBOL (drm_sched_job_arm); -/** -* drm_sched_job_add_dependency - adds the fence as a job dependency -* @job: scheduler job to add the dependencies to drivers/gpu/drm/scheduler/sched_main.c void drm_sched_job_arm(struct drm_sched_job *job) · line 676
Quoted 10 lines
drivers/gpu/drm/scheduler/sched_main.c +1
-9-* @fence: the dma_fence to add to the list of dependencies. -* -* Note that @fence is consumed in both the success and error cases. -* -* Returns: -* 0 on success, or an error on failing to expand the array. -*/ -int drm_sched_job_add_dependency (struct drm_sched_job *job, -struct dma_fence *fence) +static int _add_dependency (struct drm_sched_job *job,struct dma_fence *fence)Please keep the
drm_sched_job_prefix here even for static functions. The symbol_add_dependencyjust sucks in a backtrace, especially when it's tail optimized.Quoted 33 lines
drivers/gpu/drm/scheduler/sched_main.c +3
-3{struct dma_fence *entry;unsigned long index;u32 id =0 ;int ret; -if (!fence) -return 0 ; -/* Deduplicate if we already depend on a fence from the same context. * This lets the size of the array of deps scale with the number of * engines involved, rather than the number of BOs. */ xa_for_each (&job->dependencies, index, entry) {if (entry->context != fence->context)continue ;if (dma_fence_is_later (fence, entry)) {dma_fence_put (entry);@@ -709,20 +696,46 @@ int drm_sched_job_add_dependency(struct drm_sched_job *job, }return 0 ; } ret =xa_alloc (&job->dependencies, &id, fence, xa_limit_32b,GFP_KERNEL );if (ret !=0 )dma_fence_put (fence);return ret; } + +/** +* drm_sched_job_add_dependency - adds the fence as a job dependency drivers/gpu/drm/scheduler/sched_main.c int drm_sched_job_add_dependency(struct drm_sched_job *job, · line 718
Quoted 10 lines
drivers/gpu/drm/scheduler/sched_main.c +10
-0+* @job: scheduler job to add the dependencies to +* @fence: the dma_fence to add to the list of dependencies. +* +* Note that @fence is consumed in both the success and error cases. +* +* Returns: +* 0 on success, or an error on failing to expand the array. +*/ +int drm_sched_job_add_dependency (struct drm_sched_job *job, +struct dma_fence *fence)Maybe name the new function
drm_sched_job_unwrap_add_dependencyor something like this.
Quoted 2 lines
I need to double check, but I think for some cases we don't need or don't even want this in the driver.
I'd be curious to know the cases where you don't want this.. one thing I was thinking about, what if you have a container fence with two contained fences. One is on the same ctx as the job, one is not but signals sooner. You end up artificially waiting on both, which seems sub-optimal.
Anyways, I can make this a new entrypoint which unwraps, and/or rename the internal static function, if we think this is a good idea.
BR,
-R
Quoted 26 lines
Christian.
Quoted 23 lines
drivers/gpu/drm/scheduler/sched_main.c +13
-0+{ +struct dma_fence_unwrap iter; +struct dma_fence *f; +int ret =0 ; + + dma_fence_unwrap_for_each (f, &iter, fence) { + ret =_add_dependency (job, f); +if (ret) +break ; + } + +return ret; +}EXPORT_SYMBOL (drm_sched_job_add_dependency);/** * drm_sched_job_add_resv_dependencies - add all fences from the resv to the job * @job: scheduler job to add the dependencies to * @resv: the dma_resv object to get the fences from * @usage: the dma_resv_usage to use to filter the fences * * This adds all fences matching the given usage from @resv to @job. * Must be called with the @resv lock held.
Am 23.03.23 um 14:54 schrieb Rob Clark:
Quoted 121 lines
Quoted 119 lines · Christian König
Am 22.03.23 um 23:44 schrieb Rob Clark:
Quoted 5 lines
From: Rob Clark <[email protected]>
Container fences have burner contexts, which makes the trick to store at most one fence per context somewhat useless if we don't unwrap array or chain fences.
Mhm, we intentionally kept them not unwrapped since this way they only occupy one fence slot.
But it might be better to unwrap them if you add many of those dependencies.
Quoted 48 lines
Signed-off-by Rob Clark <[email protected]>---
tbh, I'm not sure why we weren't doing this already, unless there is something I'm overlookingdrivers/gpu/drm/scheduler/sched_main.c | 43 +++++++++++++++++--------- 1 file changed, 28 insertions(+), 15 deletions(-)drivers/gpu/drm/scheduler/sched_main.c +1
-3@@ -41,20 +41,21 @@* 4. Entities themselves maintain a queue of jobs that will be scheduled on * the hardware. * * The jobs in a entity are always scheduled in the order that they were pushed. */ #include <linux/kthread.h> #include <linux/wait.h> #include <linux/sched.h> #include <linux/completion.h> +#include <linux/dma-fence-unwrap.h> #include <linux/dma-resv.h> #include <uapi/linux/sched/types.h> #include <drm/drm_print.h> #include <drm/drm_gem.h> #include <drm/gpu_scheduler.h> #include <drm/spsc_queue.h> #define CREATE_TRACE_POINTS #include "gpu_scheduler_trace.h" @@ -665,41 +666,27 @@ void drm_sched_job_arm(struct drm_sched_job *job) sched = entity->rq->sched; job->sched = sched; job->s_priority = entity->rq - sched->sched_rq; job->id =atomic64_inc_return (&sched->job_id_count);drm_sched_fence_init (job->s_fence, job->entity); }EXPORT_SYMBOL (drm_sched_job_arm); -/** -* drm_sched_job_add_dependency - adds the fence as a job dependency -* @job: scheduler job to add the dependencies to drivers/gpu/drm/scheduler/sched_main.c void drm_sched_job_arm(struct drm_sched_job *job) · line 676
Quoted 10 lines
drivers/gpu/drm/scheduler/sched_main.c +1
-9-* @fence: the dma_fence to add to the list of dependencies. -* -* Note that @fence is consumed in both the success and error cases. -* -* Returns: -* 0 on success, or an error on failing to expand the array. -*/ -int drm_sched_job_add_dependency (struct drm_sched_job *job, -struct dma_fence *fence) +static int _add_dependency (struct drm_sched_job *job,struct dma_fence *fence)Please keep the
drm_sched_job_prefix here even for static functions. The symbol_add_dependencyjust sucks in a backtrace, especially when it's tail optimized.Quoted 33 lines
drivers/gpu/drm/scheduler/sched_main.c +3
-3{struct dma_fence *entry;unsigned long index;u32 id =0 ;int ret; -if (!fence) -return 0 ; -/* Deduplicate if we already depend on a fence from the same context. * This lets the size of the array of deps scale with the number of * engines involved, rather than the number of BOs. */ xa_for_each (&job->dependencies, index, entry) {if (entry->context != fence->context)continue ;if (dma_fence_is_later (fence, entry)) {dma_fence_put (entry);@@ -709,20 +696,46 @@ int drm_sched_job_add_dependency(struct drm_sched_job *job, }return 0 ; } ret =xa_alloc (&job->dependencies, &id, fence, xa_limit_32b,GFP_KERNEL );if (ret !=0 )dma_fence_put (fence);return ret; } + +/** +* drm_sched_job_add_dependency - adds the fence as a job dependency drivers/gpu/drm/scheduler/sched_main.c int drm_sched_job_add_dependency(struct drm_sched_job *job, · line 718
Quoted 10 lines
drivers/gpu/drm/scheduler/sched_main.c +10
-0+* @job: scheduler job to add the dependencies to +* @fence: the dma_fence to add to the list of dependencies. +* +* Note that @fence is consumed in both the success and error cases. +* +* Returns: +* 0 on success, or an error on failing to expand the array. +*/ +int drm_sched_job_add_dependency (struct drm_sched_job *job, +struct dma_fence *fence)Maybe name the new function
drm_sched_job_unwrap_add_dependencyor something like this.
Quoted 7 lines
Quoted 2 lines
I need to double check, but I think for some cases we don't need or don't even want this in the driver.
I'd be curious to know the cases where you don't want this.. one thing I was thinking about, what if you have a container fence with two contained fences. One is on the same ctx as the job, one is not but signals sooner. You end up artificially waiting on both, which seems sub-optimal.
Well resv objects don't contain other containers for example.
Then we also have an use case in amdgpu where fence need to be explicitly waited for even when they are from the same ctx as the job because otherwise we wouldn't see everything cache coherent.
On the other hand we currently handle that amdgpu use case differently and the extra overhead of unwrapping fences even if they can't be containers is probably negligible.
Quoted 2 lines
Anyways, I can make this a new entrypoint which unwraps, and/or rename the internal static function, if we think this is a good idea.
If you think that's unnecessary keeping your original approach is fine with me as well.
Regards,
Christian.
Quoted 29 lines
BR,
-RQuoted 25 lines
Christian.
Quoted 23 lines
drivers/gpu/drm/scheduler/sched_main.c +13
-0+{ +struct dma_fence_unwrap iter; +struct dma_fence *f; +int ret =0 ; + + dma_fence_unwrap_for_each (f, &iter, fence) { + ret =_add_dependency (job, f); +if (ret) +break ; + } + +return ret; +}EXPORT_SYMBOL (drm_sched_job_add_dependency);/** * drm_sched_job_add_resv_dependencies - add all fences from the resv to the job * @job: scheduler job to add the dependencies to * @resv: the dma_resv object to get the fences from * @usage: the dma_resv_usage to use to filter the fences * * This adds all fences matching the given usage from @resv to @job. * Must be called with the @resv lock held.
Quoted 125 lines · Christian König
Am 23.03.23 um 14:54 schrieb Rob Clark:
Quoted 123 lines
Quoted 121 lines · Christian König
Am 22.03.23 um 23:44 schrieb Rob Clark:
Quoted 5 lines
From: Rob Clark <[email protected]>
Container fences have burner contexts, which makes the trick to store at most one fence per context somewhat useless if we don't unwrap array or chain fences.
Mhm, we intentionally kept them not unwrapped since this way they only occupy one fence slot.
But it might be better to unwrap them if you add many of those dependencies.
Quoted 48 lines
Signed-off-by Rob Clark <[email protected]>---
tbh, I'm not sure why we weren't doing this already, unless there is something I'm overlookingdrivers/gpu/drm/scheduler/sched_main.c | 43 +++++++++++++++++--------- 1 file changed, 28 insertions(+), 15 deletions(-)drivers/gpu/drm/scheduler/sched_main.c +1
-3@@ -41,20 +41,21 @@* 4. Entities themselves maintain a queue of jobs that will be scheduled on * the hardware. * * The jobs in a entity are always scheduled in the order that they were pushed. */ #include <linux/kthread.h> #include <linux/wait.h> #include <linux/sched.h> #include <linux/completion.h> +#include <linux/dma-fence-unwrap.h> #include <linux/dma-resv.h> #include <uapi/linux/sched/types.h> #include <drm/drm_print.h> #include <drm/drm_gem.h> #include <drm/gpu_scheduler.h> #include <drm/spsc_queue.h> #define CREATE_TRACE_POINTS #include "gpu_scheduler_trace.h" @@ -665,41 +666,27 @@ void drm_sched_job_arm(struct drm_sched_job *job) sched = entity->rq->sched; job->sched = sched; job->s_priority = entity->rq - sched->sched_rq; job->id =atomic64_inc_return (&sched->job_id_count);drm_sched_fence_init (job->s_fence, job->entity); }EXPORT_SYMBOL (drm_sched_job_arm); -/** -* drm_sched_job_add_dependency - adds the fence as a job dependency -* @job: scheduler job to add the dependencies to drivers/gpu/drm/scheduler/sched_main.c void drm_sched_job_arm(struct drm_sched_job *job) · line 676
Quoted 10 lines
drivers/gpu/drm/scheduler/sched_main.c +1
-9-* @fence: the dma_fence to add to the list of dependencies. -* -* Note that @fence is consumed in both the success and error cases. -* -* Returns: -* 0 on success, or an error on failing to expand the array. -*/ -int drm_sched_job_add_dependency (struct drm_sched_job *job, -struct dma_fence *fence) +static int _add_dependency (struct drm_sched_job *job,struct dma_fence *fence)Please keep the
drm_sched_job_prefix here even for static functions. The symbol_add_dependencyjust sucks in a backtrace, especially when it's tail optimized.Quoted 33 lines
drivers/gpu/drm/scheduler/sched_main.c +3
-3{struct dma_fence *entry;unsigned long index;u32 id =0 ;int ret; -if (!fence) -return 0 ; -/* Deduplicate if we already depend on a fence from the same context. * This lets the size of the array of deps scale with the number of * engines involved, rather than the number of BOs. */ xa_for_each (&job->dependencies, index, entry) {if (entry->context != fence->context)continue ;if (dma_fence_is_later (fence, entry)) {dma_fence_put (entry);@@ -709,20 +696,46 @@ int drm_sched_job_add_dependency(struct drm_sched_job *job, }return 0 ; } ret =xa_alloc (&job->dependencies, &id, fence, xa_limit_32b,GFP_KERNEL );if (ret !=0 )dma_fence_put (fence);return ret; } + +/** +* drm_sched_job_add_dependency - adds the fence as a job dependency drivers/gpu/drm/scheduler/sched_main.c int drm_sched_job_add_dependency(struct drm_sched_job *job, · line 718
Quoted 10 lines
drivers/gpu/drm/scheduler/sched_main.c +10
-0+* @job: scheduler job to add the dependencies to +* @fence: the dma_fence to add to the list of dependencies. +* +* Note that @fence is consumed in both the success and error cases. +* +* Returns: +* 0 on success, or an error on failing to expand the array. +*/ +int drm_sched_job_add_dependency (struct drm_sched_job *job, +struct dma_fence *fence)Maybe name the new function
drm_sched_job_unwrap_add_dependencyor something like this.I need to double check, but I think for some cases we don't need or don't even want this in the driver.
Quoted 7 lines
Quoted 5 lines
I'd be curious to know the cases where you don't want this.. one thing I was thinking about, what if you have a container fence with two contained fences. One is on the same ctx as the job, one is not but signals sooner. You end up artificially waiting on both, which seems sub-optimal.
Well resv objects don't contain other containers for example.
I suppose I have the explicit sync case more in mind, where the dependent fence ends up being a chain or array (if userspace is merging fence fd's).
Quoted 3 lines
Then we also have an use case in amdgpu where fence need to be explicitly waited for even when they are from the same ctx as the job because otherwise we wouldn't see everything cache coherent.
This was the kinda weird case I wanted to make sure I wasn't breaking. I remember seeing something fly by for this, but can't find it now or remember what amdgpu's solution was..
Quoted 9 lines
On the other hand we currently handle that amdgpu use case differently and the extra overhead of unwrapping fences even if they can't be containers is probably negligible.
Quoted 2 lines
Anyways, I can make this a new entrypoint which unwraps, and/or rename the internal static function, if we think this is a good idea.
If you think that's unnecessary keeping your original approach is fine with me as well.
I'm going to assume unnecessary until someone speaks up with their weird specific case ;-)
BR,
-R
Quoted 33 lines
Regards,
Christian.Quoted 29 lines
BR,
-RQuoted 25 lines
Christian.
Quoted 23 lines
drivers/gpu/drm/scheduler/sched_main.c +13
-0+{ +struct dma_fence_unwrap iter; +struct dma_fence *f; +int ret =0 ; + + dma_fence_unwrap_for_each (f, &iter, fence) { + ret =_add_dependency (job, f); +if (ret) +break ; + } + +return ret; +}EXPORT_SYMBOL (drm_sched_job_add_dependency);/** * drm_sched_job_add_resv_dependencies - add all fences from the resv to the job * @job: scheduler job to add the dependencies to * @resv: the dma_resv object to get the fences from * @usage: the dma_resv_usage to use to filter the fences * * This adds all fences matching the given usage from @resv to @job. * Must be called with the @resv lock held.
Quoted 152 lines · Rob Clark
Quoted 125 lines · Christian König
Am 23.03.23 um 14:54 schrieb Rob Clark:
Quoted 123 lines
Quoted 121 lines · Christian König
Am 22.03.23 um 23:44 schrieb Rob Clark:
Quoted 5 lines
From: Rob Clark <[email protected]>
Container fences have burner contexts, which makes the trick to store at most one fence per context somewhat useless if we don't unwrap array or chain fences.
Mhm, we intentionally kept them not unwrapped since this way they only occupy one fence slot.
But it might be better to unwrap them if you add many of those dependencies.
Quoted 48 lines
Signed-off-by Rob Clark <[email protected]>---
tbh, I'm not sure why we weren't doing this already, unless there is something I'm overlookingdrivers/gpu/drm/scheduler/sched_main.c | 43 +++++++++++++++++--------- 1 file changed, 28 insertions(+), 15 deletions(-)drivers/gpu/drm/scheduler/sched_main.c +1
-3@@ -41,20 +41,21 @@* 4. Entities themselves maintain a queue of jobs that will be scheduled on * the hardware. * * The jobs in a entity are always scheduled in the order that they were pushed. */ #include <linux/kthread.h> #include <linux/wait.h> #include <linux/sched.h> #include <linux/completion.h> +#include <linux/dma-fence-unwrap.h> #include <linux/dma-resv.h> #include <uapi/linux/sched/types.h> #include <drm/drm_print.h> #include <drm/drm_gem.h> #include <drm/gpu_scheduler.h> #include <drm/spsc_queue.h> #define CREATE_TRACE_POINTS #include "gpu_scheduler_trace.h" @@ -665,41 +666,27 @@ void drm_sched_job_arm(struct drm_sched_job *job) sched = entity->rq->sched; job->sched = sched; job->s_priority = entity->rq - sched->sched_rq; job->id =atomic64_inc_return (&sched->job_id_count);drm_sched_fence_init (job->s_fence, job->entity); }EXPORT_SYMBOL (drm_sched_job_arm); -/** -* drm_sched_job_add_dependency - adds the fence as a job dependency -* @job: scheduler job to add the dependencies to drivers/gpu/drm/scheduler/sched_main.c void drm_sched_job_arm(struct drm_sched_job *job) · line 676
Quoted 10 lines
drivers/gpu/drm/scheduler/sched_main.c +1
-9-* @fence: the dma_fence to add to the list of dependencies. -* -* Note that @fence is consumed in both the success and error cases. -* -* Returns: -* 0 on success, or an error on failing to expand the array. -*/ -int drm_sched_job_add_dependency (struct drm_sched_job *job, -struct dma_fence *fence) +static int _add_dependency (struct drm_sched_job *job,struct dma_fence *fence)Please keep the
drm_sched_job_prefix here even for static functions. The symbol_add_dependencyjust sucks in a backtrace, especially when it's tail optimized.Quoted 33 lines
drivers/gpu/drm/scheduler/sched_main.c +3
-3{struct dma_fence *entry;unsigned long index;u32 id =0 ;int ret; -if (!fence) -return 0 ; -/* Deduplicate if we already depend on a fence from the same context. * This lets the size of the array of deps scale with the number of * engines involved, rather than the number of BOs. */ xa_for_each (&job->dependencies, index, entry) {if (entry->context != fence->context)continue ;if (dma_fence_is_later (fence, entry)) {dma_fence_put (entry);@@ -709,20 +696,46 @@ int drm_sched_job_add_dependency(struct drm_sched_job *job, }return 0 ; } ret =xa_alloc (&job->dependencies, &id, fence, xa_limit_32b,GFP_KERNEL );if (ret !=0 )dma_fence_put (fence);return ret; } + +/** +* drm_sched_job_add_dependency - adds the fence as a job dependency drivers/gpu/drm/scheduler/sched_main.c int drm_sched_job_add_dependency(struct drm_sched_job *job, · line 718
Quoted 10 lines
drivers/gpu/drm/scheduler/sched_main.c +10
-0+* @job: scheduler job to add the dependencies to +* @fence: the dma_fence to add to the list of dependencies. +* +* Note that @fence is consumed in both the success and error cases. +* +* Returns: +* 0 on success, or an error on failing to expand the array. +*/ +int drm_sched_job_add_dependency (struct drm_sched_job *job, +struct dma_fence *fence)Maybe name the new function
drm_sched_job_unwrap_add_dependencyor something like this.I need to double check, but I think for some cases we don't need or don't even want this in the driver.
Quoted 7 lines
Quoted 5 lines
I'd be curious to know the cases where you don't want this.. one thing I was thinking about, what if you have a container fence with two contained fences. One is on the same ctx as the job, one is not but signals sooner. You end up artificially waiting on both, which seems sub-optimal.
Well resv objects don't contain other containers for example.
I suppose I have the explicit sync case more in mind, where the dependent fence ends up being a chain or array (if userspace is merging fence fd's).
Quoted 3 lines
Then we also have an use case in amdgpu where fence need to be explicitly waited for even when they are from the same ctx as the job because otherwise we wouldn't see everything cache coherent.
This was the kinda weird case I wanted to make sure I wasn't breaking. I remember seeing something fly by for this, but can't find it now or remember what amdgpu's solution was..
Quoted 4 lines
On the other hand we currently handle that amdgpu use case differently and the extra overhead of unwrapping fences even if they can't be containers is probably negligible.
Quoted 8 lines
Quoted 5 lines
Quoted 2 lines
Anyways, I can make this a new entrypoint which unwraps, and/or rename the internal static function, if we think this is a good idea.
If you think that's unnecessary keeping your original approach is fine with me as well.
I'm going to assume unnecessary until someone speaks up with their weird specific case ;-)
So, this patch turns out to blow up spectacularly with dma_fence refcnt underflows when I enable DRIVER_SYNCOBJ_TIMELINE .. I think, because it starts unwrapping fence chains, possibly in parallel with fence signaling on the retire path. Is it supposed to be permissible to unwrap a fence chain concurrently?
BR,
-R
Quoted 36 lines
BR,
-RQuoted 33 lines
Regards,
Christian.Quoted 29 lines
BR,
-RQuoted 25 lines
Christian.
Quoted 23 lines
drivers/gpu/drm/scheduler/sched_main.c +13
-0+{ +struct dma_fence_unwrap iter; +struct dma_fence *f; +int ret =0 ; + + dma_fence_unwrap_for_each (f, &iter, fence) { + ret =_add_dependency (job, f); +if (ret) +break ; + } + +return ret; +}EXPORT_SYMBOL (drm_sched_job_add_dependency);/** * drm_sched_job_add_resv_dependencies - add all fences from the resv to the job * @job: scheduler job to add the dependencies to * @resv: the dma_resv object to get the fences from * @usage: the dma_resv_usage to use to filter the fences * * This adds all fences matching the given usage from @resv to @job. * Must be called with the @resv lock held.
Am 04.12.23 um 22:54 schrieb Rob Clark:
Quoted 7 lines · Rob Clark
Quoted 1 line
[SNIP]
So, this patch turns out to blow up spectacularly with
dma_fencerefcnt underflows when I enableDRIVER_SYNCOBJ_TIMELINE.. I think, because it starts unwrapping fence chains, possibly in parallel with fence signaling on the retire path. Is it supposed to be permissible to unwrap a fence chain concurrently?
The DMA-fence chain object and helper functions were designed so that concurrent accesses to all elements are always possible.
See dma_fence_chain_walk() and dma_fence_chain_get_prev() for example. dma_fence_chain_walk() starts with a reference to the current fence (the anchor of the walk) and tries to grab an up to date reference on the previous fence in the chain. Only after that reference is successfully acquired we drop the reference to the anchor where we started.
Same for dma_fence_array_first(), dma_fence_array_next(). Here we hold a reference to the array which in turn holds references to each fence inside the array until it is destroyed itself.
When this blows up we have somehow mixed up the references somewhere.
Regards,
Christian.
Quoted 3 lines
BR,
-R
Quoted 19 lines · Christian König
Am 04.12.23 um 22:54 schrieb Rob Clark:
Quoted 7 lines · Rob Clark
Quoted 1 line
[SNIP]
So, this patch turns out to blow up spectacularly with
dma_fencerefcnt underflows when I enableDRIVER_SYNCOBJ_TIMELINE.. I think, because it starts unwrapping fence chains, possibly in parallel with fence signaling on the retire path. Is it supposed to be permissible to unwrap a fence chain concurrently?The DMA-fence chain object and helper functions were designed so that concurrent accesses to all elements are always possible.
See
dma_fence_chain_walk()anddma_fence_chain_get_prev()for example.dma_fence_chain_walk()starts with a reference to the current fence (the anchor of the walk) and tries to grab an up to date reference on the previous fence in the chain. Only after that reference is successfully acquired we drop the reference to the anchor where we started.
Quoted 5 lines
Same for
dma_fence_array_first(),dma_fence_array_next(). Here we hold a reference to the array which in turn holds references to each fence inside the array until it is destroyed itself.When this blows up we have somehow mixed up the references somewhere.
That's what it looked like to me, but wanted to make sure I wasn't overlooking something subtle. And in this case, the fence actually should be the syncobj timeline point fence, not the fence chain. Virtgpu has essentially the same logic (there we really do want to unwrap fences so we can pass host fences back to host rather than waiting in guest), I'm not sure if it would blow up in the same way.
BR,
-R
Quoted 6 lines
Regards,
Christian.Quoted 3 lines
BR,
-R
Am 05.12.23 um 16:41 schrieb Rob Clark:
Quoted 23 lines
Quoted 21 lines · Christian König
Am 04.12.23 um 22:54 schrieb Rob Clark:
Quoted 7 lines · Rob Clark
Quoted 1 line
[SNIP]
So, this patch turns out to blow up spectacularly with
dma_fencerefcnt underflows when I enableDRIVER_SYNCOBJ_TIMELINE.. I think, because it starts unwrapping fence chains, possibly in parallel with fence signaling on the retire path. Is it supposed to be permissible to unwrap a fence chain concurrently?The DMA-fence chain object and helper functions were designed so that concurrent accesses to all elements are always possible.
See
dma_fence_chain_walk()anddma_fence_chain_get_prev()for example.dma_fence_chain_walk()starts with a reference to the current fence (the anchor of the walk) and tries to grab an up to date reference on the previous fence in the chain. Only after that reference is successfully acquired we drop the reference to the anchor where we started.Same for
dma_fence_array_first(),dma_fence_array_next(). Here we hold a reference to the array which in turn holds references to each fence inside the array until it is destroyed itself.
Quoted 7 lines
Quoted 1 line
When this blows up we have somehow mixed up the references somewhere.
That's what it looked like to me, but wanted to make sure I wasn't overlooking something subtle. And in this case, the fence actually should be the syncobj timeline point fence, not the fence chain. Virtgpu has essentially the same logic (there we really do want to unwrap fences so we can pass host fences back to host rather than waiting in guest), I'm not sure if it would blow up in the same way.
Well do you have a backtrace of what exactly happens?
Maybe we have some _put() before _get() or something like this.
Thanks,
Christian.
Quoted 9 lines
BR,
-RQuoted 5 lines
Regards,
Christian.Quoted 2 lines
BR,
-R
Quoted 33 lines · Christian König
Am 05.12.23 um 16:41 schrieb Rob Clark:
Quoted 23 lines
Quoted 21 lines · Christian König
Am 04.12.23 um 22:54 schrieb Rob Clark:
Quoted 7 lines · Rob Clark
Quoted 1 line
[SNIP]
So, this patch turns out to blow up spectacularly with
dma_fencerefcnt underflows when I enableDRIVER_SYNCOBJ_TIMELINE.. I think, because it starts unwrapping fence chains, possibly in parallel with fence signaling on the retire path. Is it supposed to be permissible to unwrap a fence chain concurrently?The DMA-fence chain object and helper functions were designed so that concurrent accesses to all elements are always possible.
See
dma_fence_chain_walk()anddma_fence_chain_get_prev()for example.dma_fence_chain_walk()starts with a reference to the current fence (the anchor of the walk) and tries to grab an up to date reference on the previous fence in the chain. Only after that reference is successfully acquired we drop the reference to the anchor where we started.Same for
dma_fence_array_first(),dma_fence_array_next(). Here we hold a reference to the array which in turn holds references to each fence inside the array until it is destroyed itself.Quoted 7 lines
Quoted 1 line
When this blows up we have somehow mixed up the references somewhere.
That's what it looked like to me, but wanted to make sure I wasn't overlooking something subtle. And in this case, the fence actually should be the syncobj timeline point fence, not the fence chain. Virtgpu has essentially the same logic (there we really do want to unwrap fences so we can pass host fences back to host rather than waiting in guest), I'm not sure if it would blow up in the same way.
Quoted 3 lines
Well do you have a backtrace of what exactly happens?
Maybe we have some
_put()before_get()or something like this.
I hacked up something to store the backtrace in dma_fence_release() (and leak the block so the backtrace would still be around later when dma_fence_get/put was later called) and ended up with:
[ 152.811360] freed at: [ 152.813718] dma_fence_release+0x30/0x134 [ 152.817865] dma_fence_put+0x38/0x98 [gpu_sched] [ 152.822657] drm_sched_job_add_dependency+0x160/0x18c [gpu_sched] [ 152.828948] drm_sched_job_add_syncobj_dependency+0x58/0x88 [gpu_sched] [ 152.835770] msm_ioctl_gem_submit+0x580/0x1160 [msm] [ 152.841070] drm_ioctl_kernel+0xec/0x16c [ 152.845132] drm_ioctl+0x2e8/0x3f4 [ 152.848646] vfs_ioctl+0x30/0x50 [ 152.851982] __arm64_sys_ioctl+0x80/0xb4 [ 152.856039] invoke_syscall+0x8c/0x120 [ 152.859919] el0_svc_common.constprop.0+0xc0/0xdc [ 152.864777] do_el0_svc+0x24/0x30 [ 152.868207] el0_svc+0x8c/0xd8 [ 152.871365] el0t_64_sync_handler+0x84/0x12c [ 152.875771] el0t_64_sync+0x190/0x194
I suppose that doesn't guarantee that this was the problematic put. But dropping this patch to unwrap the fence makes the problem go away..
BR,
-R
Quoted 13 lines
Thanks,
Christian.Quoted 9 lines
BR,
-RQuoted 5 lines
Regards,
Christian.Quoted 2 lines
BR,
-R
Quoted 60 lines · Rob Clark
Quoted 33 lines · Christian König
Am 05.12.23 um 16:41 schrieb Rob Clark:
Quoted 23 lines
Quoted 21 lines · Christian König
Am 04.12.23 um 22:54 schrieb Rob Clark:
Quoted 7 lines · Rob Clark
Quoted 1 line
[SNIP]
So, this patch turns out to blow up spectacularly with
dma_fencerefcnt underflows when I enableDRIVER_SYNCOBJ_TIMELINE.. I think, because it starts unwrapping fence chains, possibly in parallel with fence signaling on the retire path. Is it supposed to be permissible to unwrap a fence chain concurrently?The DMA-fence chain object and helper functions were designed so that concurrent accesses to all elements are always possible.
See
dma_fence_chain_walk()anddma_fence_chain_get_prev()for example.dma_fence_chain_walk()starts with a reference to the current fence (the anchor of the walk) and tries to grab an up to date reference on the previous fence in the chain. Only after that reference is successfully acquired we drop the reference to the anchor where we started.Same for
dma_fence_array_first(),dma_fence_array_next(). Here we hold a reference to the array which in turn holds references to each fence inside the array until it is destroyed itself.Quoted 7 lines
Quoted 1 line
When this blows up we have somehow mixed up the references somewhere.
That's what it looked like to me, but wanted to make sure I wasn't overlooking something subtle. And in this case, the fence actually should be the syncobj timeline point fence, not the fence chain. Virtgpu has essentially the same logic (there we really do want to unwrap fences so we can pass host fences back to host rather than waiting in guest), I'm not sure if it would blow up in the same way.
Quoted 3 lines
Well do you have a backtrace of what exactly happens?
Maybe we have some
_put()before_get()or something like this.I hacked up something to store the backtrace in
dma_fence_release()(and leak the block so the backtrace would still be around later when dma_fence_get/put was later called) and ended up with:[ 152.811360] freed at: [ 152.813718] dma_fence_release+0x30/0x134 [ 152.817865] dma_fence_put+0x38/0x98 [gpu_sched] [ 152.822657] drm_sched_job_add_dependency+0x160/0x18c [gpu_sched] [ 152.828948] drm_sched_job_add_syncobj_dependency+0x58/0x88 [gpu_sched] [ 152.835770] msm_ioctl_gem_submit+0x580/0x1160 [msm] [ 152.841070] drm_ioctl_kernel+0xec/0x16c [ 152.845132] drm_ioctl+0x2e8/0x3f4 [ 152.848646] vfs_ioctl+0x30/0x50 [ 152.851982] __arm64_sys_ioctl+0x80/0xb4 [ 152.856039] invoke_syscall+0x8c/0x120 [ 152.859919] el0_svc_common.constprop.0+0xc0/0xdc [ 152.864777] do_el0_svc+0x24/0x30 [ 152.868207] el0_svc+0x8c/0xd8 [ 152.871365] el0t_64_sync_handler+0x84/0x12c [ 152.875771] el0t_64_sync+0x190/0x194
Quoted 3 lines
I suppose that doesn't guarantee that this was the problematic put. But dropping this patch to unwrap the fence makes the problem go away..
Oh, hmm, _add_dependency() is consuming the fence reference
BR,
-R
Quoted 16 lines
BR,
-RQuoted 13 lines
Thanks,
Christian.Quoted 9 lines
BR,
-RQuoted 5 lines
Regards,
Christian.Quoted 2 lines
BR,
-R
Christian König
reviewerAm 05.12.23 um 18:14 schrieb Rob Clark:
Quoted 57 lines · Rob Clark
Quoted 56 lines · Christian König
Quoted 31 lines
Am 05.12.23 um 16:41 schrieb Rob Clark:
Quoted 23 lines
Quoted 21 lines · Christian König
Am 04.12.23 um 22:54 schrieb Rob Clark:
Quoted 7 lines · Rob Clark
Quoted 1 line
[SNIP]
So, this patch turns out to blow up spectacularly with
dma_fencerefcnt underflows when I enableDRIVER_SYNCOBJ_TIMELINE.. I think, because it starts unwrapping fence chains, possibly in parallel with fence signaling on the retire path. Is it supposed to be permissible to unwrap a fence chain concurrently?The DMA-fence chain object and helper functions were designed so that concurrent accesses to all elements are always possible.
See
dma_fence_chain_walk()anddma_fence_chain_get_prev()for example.dma_fence_chain_walk()starts with a reference to the current fence (the anchor of the walk) and tries to grab an up to date reference on the previous fence in the chain. Only after that reference is successfully acquired we drop the reference to the anchor where we started.Same for
dma_fence_array_first(),dma_fence_array_next(). Here we hold a reference to the array which in turn holds references to each fence inside the array until it is destroyed itself.Quoted 7 lines
Quoted 1 line
When this blows up we have somehow mixed up the references somewhere.
That's what it looked like to me, but wanted to make sure I wasn't overlooking something subtle. And in this case, the fence actually should be the syncobj timeline point fence, not the fence chain. Virtgpu has essentially the same logic (there we really do want to unwrap fences so we can pass host fences back to host rather than waiting in guest), I'm not sure if it would blow up in the same way.
Quoted 3 lines
Well do you have a backtrace of what exactly happens?
Maybe we have some
_put()before_get()or something like this.I hacked up something to store the backtrace in
dma_fence_release()(and leak the block so the backtrace would still be around later when dma_fence_get/put was later called) and ended up with:[ 152.811360] freed at: [ 152.813718] dma_fence_release+0x30/0x134 [ 152.817865] dma_fence_put+0x38/0x98 [gpu_sched] [ 152.822657] drm_sched_job_add_dependency+0x160/0x18c [gpu_sched] [ 152.828948] drm_sched_job_add_syncobj_dependency+0x58/0x88 [gpu_sched] [ 152.835770] msm_ioctl_gem_submit+0x580/0x1160 [msm] [ 152.841070] drm_ioctl_kernel+0xec/0x16c [ 152.845132] drm_ioctl+0x2e8/0x3f4 [ 152.848646] vfs_ioctl+0x30/0x50 [ 152.851982] __arm64_sys_ioctl+0x80/0xb4 [ 152.856039] invoke_syscall+0x8c/0x120 [ 152.859919] el0_svc_common.constprop.0+0xc0/0xdc [ 152.864777] do_el0_svc+0x24/0x30 [ 152.868207] el0_svc+0x8c/0xd8 [ 152.871365] el0t_64_sync_handler+0x84/0x12c [ 152.875771] el0t_64_sync+0x190/0x194
Quoted 4 lines
Quoted 3 lines
I suppose that doesn't guarantee that this was the problematic put. But dropping this patch to unwrap the fence makes the problem go away..
Oh, hmm,
_add_dependency()is consuming the fence reference
Yeah, I was just about to point that out as well :)
Should be trivial to fix,
Christian
Quoted 21 lines
BR,
-RQuoted 14 lines
BR,
-RQuoted 11 lines
Thanks,
Christian.Quoted 8 lines
BR,
-RQuoted 5 lines
Regards,
Christian.Quoted 2 lines
BR,
-R_______________________________________________
Linaro-mm-sig mailing list -- [email protected] To unsubscribe send an email to [email protected]