Re: [BUG] ALSA: usb-audio: drain may fail with multi-port close race

[Takashi Iwai <tiwai@suse.de>]

On Tue, 18 Feb 2025 11:56:19 +0100,
John Keeping wrote:
> 
> On Tue, Feb 18, 2025 at 09:07:26AM +0100, Takashi Iwai wrote:
> > On Mon, 17 Feb 2025 19:38:31 +0100,
> > John Keeping wrote:
> > > 
> > > On Mon, Feb 17, 2025 at 06:06:16PM +0100, Takashi Iwai wrote:
> > > > On Mon, 17 Feb 2025 12:16:46 +0100,
> > > > John Keeping wrote:
> > > > > 
> > > > > I'm seeing a bug where data sometimes fails to send on USB MIDI devices
> > > > > with multiple ports which seems to be a result of a race around closing
> > > > > ports introduced by commit 0125de38122f0 ("ALSA: usb-audio: Cancel
> > > > > pending work at closing a MIDI substream").
> > > > > 
> > > > > The scenario is essentially this program:
> > > > > 
> > > > > 	snd_rawmidi_t *port0, *port1;
> > > > > 	snd_rawmidi_open(NULL, &port0, "hw:0,0,0", 0);
> > > > > 	snd_rawmidi_open(NULL, &port1, "hw:0,0,1", 0);
> > > > > 
> > > > > 	snd_rawmidi_write(port0, data, len);
> > > > > 
> > > > > 	snd_rawmidi_close(port1);
> > > > > 	snd_rawmidi_close(port0);
> > > > > 
> > > > > What happens seems to be the following:
> > > > > 
> > > > > 	write(port0)
> > > > > 	`- snd_usbmidi_output_trigger
> > > > > 	   `- queue_work()
> > > > >         close(port1)
> > > > > 	`- snd_usbmidi_output_close
> > > > > 	   `- cancel_work_sync()	# Work has not yet started here
> > > > > 	close(port0)
> > > > > 	`- snd_rawmidi_drain_output
> > > > > 	   # Times out because nothing is processing outbound data!
> > > > > 
> > > > > The two ports interact like this because they are on the same endpoint,
> > > > > so should the work only be cancelled when the last endpoint is closed?
> > > > 
> > > > How about the following patch work?
> > > > It's a band-aid, but should suffice.  The callback is already
> > > > protected with rawmidi open_mutex.
> > > 
> > > Yes, this patch fixes it and is
> > > 
> > > Tested-by: John Keeping <jkeeping@inmusicbrands.com>
> > 
> > Thank you for quick testing!
> > 
> > Looking at the code again, I think the suggested fix isn't right.
> > It still allows some pending work accessing the freed object.
> > 
> > Could you test the following one-liner instead?
> 
> Tested-by: John Keeping <jkeeping@inmusicbrands.com>
> 
> The patch below also fixes the issue.  Thanks!

OK, I'm going to submit and merge the proper patch.


thanks,

Takashi