drm/amdkfd: Asynchronously free events
Subsystems (3)
Felix Kuehling
maintainerPATCH 1/1 drm/amdkfd: Asynchronously free events
The synchronize_rcu call in destroy_events can take several ms, which noticeably slows down applications destroying many events. Use kfree_rcu to free the event structure asynchronously and eliminate the synchronize_rcu call in the user thread.
--- drivers/gpu/drm/amd/amdkfd/kfd_events.c | 3 +-- drivers/gpu/drm/amd/amdkfd/kfd_events.h | 2 ++ 2 files changed, 3 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_events.c +1 -2
@@ -262,8 +262,7 @@ static void destroy_event(struct kfd_process *p, struct kfd_event *ev) p->signal_event_count--;idr_remove (&p->event_idr, ev->event_id); -synchronize_rcu (); -kfree (ev); +kfree_rcu (ev, rcu); }static void destroy_events (struct kfd_process *p)
drivers/gpu/drm/amd/amdkfd/kfd_events.h +2 -0
@@ -70,6 +70,8 @@ struct kfd_event {struct kfd_hsa_memory_exception_data memory_exception_data;struct kfd_hsa_hw_exception_data hw_exception_data; }; + +struct rcu_head rcu;/* for asynchronous kfree_rcu */ };#define KFD_EVENT_TIMEOUT_IMMEDIATE 0
Signature · 1 line
2.32.0
On 2022-04-07 22:39, Felix Kuehling wrote:
The synchronize_rcu call in destroy_events can take several ms, which noticeably slows down applications destroying many events. Use kfree_rcu to free the event structure asynchronously and eliminate the synchronize_rcu call in the user thread. Will need similar fix in kfd_smi_ev_release, as I noticed sometime it is slower to exit the app receiving SMI event.
Felix Kuehling
maintainerAm 2022-04-11 um 08:52 schrieb philip yang:
Quoted 8 lines · Felix Kuehling
Quoted 4 lines
The
synchronize_rcucall indestroy_eventscan take several ms, which noticeably slows down applications destroying many events. Usekfree_rcuto free the event structure asynchronously and eliminate thesynchronize_rcucall in the user thread.
Quoted 4 lines
Will need similar fix in
kfd_smi_ev_release, as I noticed sometime it is slower to exit the app receiving SMI event.Reviewed-by Philip Yang <[email protected]>
Thanks. In the case of kfd_smi_ev_release you'd need to use call_rcu because it's more than a simple kfree when you're freeing the client->fifo. That would have some implications for module unloading, too. Maybe you can simplify it by using DECLARE_KFIFO in struct kfd_smi_client, since we create the fifo with a fixed size anway. That way the cleanup is just a kfree_rcu(client).
Regards,
Felix
Quoted 33 lines
Quoted 33 lines
Signed-off-by Felix Kuehling<[email protected]>--- drivers/gpu/drm/amd/amdkfd/kfd_events.c | 3 +-- drivers/gpu/drm/amd/amdkfd/kfd_events.h | 2 ++ 2 files changed, 3 insertions(+), 2 deletions(-)drivers/gpu/drm/amd/amdkfd/kfd_events.c +1
-2@@ -262,8 +262,7 @@ static void destroy_event(struct kfd_process *p, struct kfd_event *ev) p->signal_event_count--;idr_remove (&p->event_idr, ev->event_id); -synchronize_rcu (); -kfree (ev); +kfree_rcu (ev, rcu); }static void destroy_events (struct kfd_process *p)drivers/gpu/drm/amd/amdkfd/kfd_events.h +2
-0@@ -70,6 +70,8 @@ struct kfd_event {struct kfd_hsa_memory_exception_data memory_exception_data;struct kfd_hsa_hw_exception_data hw_exception_data; }; + +struct rcu_head rcu;/* for asynchronous kfree_rcu */ };#define KFD_EVENT_TIMEOUT_IMMEDIATE 0