At Thu, 18 Oct 2012 17:39:05 +0200, Matthieu CASTET wrote:
Takashi Iwai a écrit :
At Tue, 16 Oct 2012 18:01:13 +0200, Matthieu CASTET wrote:
Takashi Iwai a écrit :
At Mon, 15 Oct 2012 19:41:40 +0200, Matthieu CASTET wrote:
Hi Takashi,
But I believe I found other races in the alsa char device handling. With the attached patch, if you disconnect the usb audio device between "msleep o" and "msleep o+", you will free the card resource (snd_card_do_free) and then use it [1].
I did in in snd_ctl_open, but the same thing could be done in snd_pcm_open, ...
OK, we'd need a generic open/close protection. For PCM, there is already a fix in my last patchset, so it should work, but for other devices, the paths are still uncovered.
I don't think it will work for pcm : the begin of snd_pcm_open is not protected by any lock even with your patch In snd_open we release sound_mutex before calling file->f_op->open.
The snd_lookup_minor_data and snd_card_file_add should be protected by a lock. Attached a patch (pcm.crash) that help to trigger [1].
Check out my latest patches in topic/usb-disconnect-fix branch. It's been fixed there.
Thanks,
the first tests seems ok. We will do more strees test.
Do you have any good/bad test result? If no regression is seen, I'm going to apply the patches for 3.7-rc4.
thanks,
Takashi