| Message ID | 20240510101424.6279-1-tiwai@suse.de (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | ALSA: Fix deadlocks with kctl removals at disconnection | expand |
On 10. 05. 24 12:14, Takashi Iwai wrote: > In snd_card_disconnect(), we set card->shutdown flag at the beginning, > call callbacks and do sync for card->power_ref_sleep waiters at the > end. The callback may delete a kctl element, and this can lead to a > deadlock when the device was in the suspended state. Namely: > > * A process waits for the power up at snd_power_ref_and_wait() in > snd_ctl_info() or read/write() inside card->controls_rwsem. > > * The system gets disconnected meanwhile, and the driver tries to > delete a kctl via snd_ctl_remove*(); it tries to take > card->controls_rwsem again, but this is already locked by the > above. Since the sleeper isn't woken up, this deadlocks. > > An easy fix is to wake up sleepers before processing the driver > disconnect callbacks but right after setting the card->shutdown flag. > Then all sleepers will abort immediately, and the code flows again. > > So, basically this patch moves the wait_event() call at the right > timing. While we're at it, just to be sure, call wait_event_all() > instead of wait_event(), although we don't use exclusive events on > this queue for now. > > Link: https://bugzilla.kernel.org/show_bug.cgi?id=218816 > Cc: <stable@vger.kernel.org> > Signed-off-by: Takashi Iwai <tiwai@suse.de> LGTM Reviewed-by: Jaroslav Kysela <perex@perex.cz>
diff --git a/sound/core/init.c b/sound/core/init.c index 4ed5037d8693..89c8354862c4 100644 --- a/sound/core/init.c +++ b/sound/core/init.c @@ -516,6 +516,14 @@ void snd_card_disconnect(struct snd_card *card) } } +#ifdef CONFIG_PM + /* wake up sleepers here before other callbacks for avoiding potential + * deadlocks with other locks (e.g. in kctls); + * then this notifies the shutdown and sleepers would abort immediately + */ + wake_up_all(&card->power_sleep); +#endif + /* notify all connected devices about disconnection */ /* at this point, they cannot respond to any calls except release() */ @@ -543,7 +551,6 @@ void snd_card_disconnect(struct snd_card *card) } #ifdef CONFIG_PM - wake_up(&card->power_sleep); snd_power_sync_ref(card); #endif }
In snd_card_disconnect(), we set card->shutdown flag at the beginning, call callbacks and do sync for card->power_ref_sleep waiters at the end. The callback may delete a kctl element, and this can lead to a deadlock when the device was in the suspended state. Namely: * A process waits for the power up at snd_power_ref_and_wait() in snd_ctl_info() or read/write() inside card->controls_rwsem. * The system gets disconnected meanwhile, and the driver tries to delete a kctl via snd_ctl_remove*(); it tries to take card->controls_rwsem again, but this is already locked by the above. Since the sleeper isn't woken up, this deadlocks. An easy fix is to wake up sleepers before processing the driver disconnect callbacks but right after setting the card->shutdown flag. Then all sleepers will abort immediately, and the code flows again. So, basically this patch moves the wait_event() call at the right timing. While we're at it, just to be sure, call wait_event_all() instead of wait_event(), although we don't use exclusive events on this queue for now. Link: https://bugzilla.kernel.org/show_bug.cgi?id=218816 Cc: <stable@vger.kernel.org> Signed-off-by: Takashi Iwai <tiwai@suse.de> --- sound/core/init.c | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-)