| Message ID | 1547206493-105303-1-git-send-email-phil@raspberrypi.org (mailing list archive) |
|---|---|
| State | Mainlined, archived |
| Commit | a50c4c9a65779ca835746b5fd79d3d5278afbdbe |
| Headers | show |
| Series | staging: vchiq: Fix local event signalling | expand |
> Phil Elwell <phil@raspberrypi.org> hat am 11. Januar 2019 um 12:34 geschrieben: > > > Prior to the recent event reworking (see Fixes), thread synchronisation > was implemented using completions, the worker thread being woken with > a call to complete(). The replacement uses waitqueues, which are more > like condition variables in that the waiting thread is only woken if > the condition is true. > > When the VPU signals the ARM, it first sets the event's fired flag to > indicate which event is being signalled, but the places in the > ARM-side code where the worker thread is being woken - > remote_event_signal_local via request_poll - did not do so as it > wasn't previously necessary, and since the armed flag was being > cleared this lead to a deadlock. > > Fixes: 852b2876a8a8 ("staging: vchiq: rework remove_event handling") > Signed-off-by: Phil Elwell <phil@raspberrypi.org> Tested-by: Stefan Wahren <stefan.wahren@i2se.com> Thanks
Hi Phil, On Fri, Jan 11, 2019 at 11:34:53AM +0000, Phil Elwell wrote: > Prior to the recent event reworking (see Fixes), thread synchronisation > was implemented using completions, the worker thread being woken with > a call to complete(). The replacement uses waitqueues, which are more > like condition variables in that the waiting thread is only woken if > the condition is true. > > When the VPU signals the ARM, it first sets the event's fired flag to > indicate which event is being signalled, but the places in the > ARM-side code where the worker thread is being woken - > remote_event_signal_local via request_poll - did not do so as it > wasn't previously necessary, and since the armed flag was being > cleared this lead to a deadlock. This fixes an issue I've had on linux 5.0-pre to 5.0-rc2+: the bcm2835-audio driver would block on close, and then nothing would work until the process was killed. Sample log: bcm2835_audio bcm2835_audio: failed to close VCHI service connection (status=1) > > Fixes: 852b2876a8a8 ("staging: vchiq: rework remove_event handling") > Signed-off-by: Phil Elwell <phil@raspberrypi.org> Tested-by: Anisse Astier <anisse@astier.eu> Regards, Anisse
diff --git a/drivers/staging/vc04_services/interface/vchiq_arm/vchiq_core.c b/drivers/staging/vc04_services/interface/vchiq_arm/vchiq_core.c index 9e17ec6..53f5a1c 100644 --- a/drivers/staging/vc04_services/interface/vchiq_arm/vchiq_core.c +++ b/drivers/staging/vc04_services/interface/vchiq_arm/vchiq_core.c @@ -446,6 +446,7 @@ remote_event_wait(wait_queue_head_t *wq, struct remote_event *event) static inline void remote_event_signal_local(wait_queue_head_t *wq, struct remote_event *event) { + event->fired = 1; event->armed = 0; wake_up_all(wq); }
Prior to the recent event reworking (see Fixes), thread synchronisation was implemented using completions, the worker thread being woken with a call to complete(). The replacement uses waitqueues, which are more like condition variables in that the waiting thread is only woken if the condition is true. When the VPU signals the ARM, it first sets the event's fired flag to indicate which event is being signalled, but the places in the ARM-side code where the worker thread is being woken - remote_event_signal_local via request_poll - did not do so as it wasn't previously necessary, and since the armed flag was being cleared this lead to a deadlock. Fixes: 852b2876a8a8 ("staging: vchiq: rework remove_event handling") Signed-off-by: Phil Elwell <phil@raspberrypi.org> --- drivers/staging/vc04_services/interface/vchiq_arm/vchiq_core.c | 1 + 1 file changed, 1 insertion(+)