On Tue, 2007-10-09 at 07:59 -0700, Trent Piepho wrote:
On Tue, 9 Oct 2007, Dave Dillow wrote:
It looks like my prepare() gets called by snd_pcm_action_nonatomic() in sound/core/pcm_native.c, and that takes a read lock on the link semaphore, so the group won't be changed, but I don't see anything that would prevent prepare() being called twice on the same substream concurrently. So, I'll probably need to keep some locking to avoid an issue there, unless I'm missing a lock upstream.
You're right about that, prepare() can be called concurrently on the same substream, as can all the non-atomic functions like hw_params(). To be sure, I wrote a multi-threaded test program to execute concurrent ioctls.
There is a per-device mutex for open/close, so they can't run concurrently.
Note that the ALSA code does *not* handle concurrent calls to hw_params() or prepare() correctly. Your driver may have a lock, but the alsa core doesn't, and will read and write to the runtime parameters concurrently. See http://article.gmane.org/gmane.linux.alsa.devel/48846
Good deal, so I'm not crazy. And to implement some sanity checking, I'll need to copy all the data I need out of the runtime with some memory barriers (or use the ones implicit in spin_locks) and then sanity check that, as otherwise there's no guarantee that hw_parms() won't be changing them while I'm looking at them....
At this point, I may just punt on the locking in prepare(), since we're all a bit screwed here, and once the driver gets merged look at fixing the locks upstream to close the issue.