Re: [PATCH] usb-audio: not submit urb for stopped endpoint

[Takashi Iwai]

On Tue, 12 Nov 2019 07:51:06 +0100,
Henry Lin wrote:
> 
> While output urb's snd_complete_urb() is executing, calling
> prepare_outbound_urb() may cause endpoint stopped before
> prepare_outbound_urb() returns and result in next urb submitted
> to stopped endpoint. usb-audio driver cannot re-use it afterwards as
> the urb is still hold by usb stack.
> 
> This change checks EP_FLAG_RUNNING flag after prepare_outbound_urb() again
> to let snd_complete_urb() know the endpoint already stopped and does not
> submit next urb.

OK, this part looks good and understandable.


> We observed two scenario have this issue:
> 1. While executing snd_complete_urb() to complete an output urb, calling
>    prepare_outbound_urb() let deactive_urbs() get called to unlink all
>    active urbs.
> 
> [  268.097066] [<ffffffc000af7638>] deactivate_urbs+0xd4/0x108
> [  268.102633] [<ffffffc000af87fc>] snd_usb_endpoint_stop+0x30/0x58
> [  268.108636] [<ffffffc000b0272c>] snd_usb_substream_playback_trigger+0xa4/0xf4
> [  268.115765] [<ffffffc000acdbd0>] snd_pcm_do_stop+0x4c/0x58
> [  268.121245] [<ffffffc000acda24>] snd_pcm_action_single+0x40/0x88
> [  268.127245] [<ffffffc000ace984>] snd_pcm_action+0x30/0xf0
> [  268.132632] [<ffffffc000acea68>] snd_pcm_stop+0x24/0x2c
> [  268.137851] [<ffffffc000ad5e14>] xrun+0x60/0x6c
> [  268.142374] [<ffffffc000ad7a98>] snd_pcm_update_state+0xa8/0x10c
> [  268.148374] [<ffffffc000ad7e24>] snd_pcm_update_hw_ptr0+0x328/0x344
> [  268.154635] [<ffffffc000ad7ed8>] snd_pcm_period_elapsed+0x98/0xb0
> [  268.160723] [<ffffffc000b02510>] prepare_playback_urb+0x46c/0x488
> [  268.166810] [<ffffffc000af7d60>] prepare_outbound_urb+0x60/0x1d4
> [  268.172805] [<ffffffc000af8d60>] snd_complete_urb+0x244/0x264
> [  268.178548] [<ffffffc00081fb38>] __usb_hcd_giveback_urb+0x94/0x104
> [  268.184721] [<ffffffc00081fbe4>] usb_hcd_giveback_urb+0x3c/0x114
> [  268.190724] [<ffffffc00084d4b4>] handle_tx_event+0x1304/0x1434
> [  268.196552] [<ffffffc00084dbc0>] xhci_handle_event+0x5dc/0x788
> [  268.202378] [<ffffffc00084dee4>] xhci_irq+0x178/0x280
> 
> 2. Userspace application stops playback from sound subsystem with below
>    call stack:
> 
> [   28.506477] CPU: 5 PID: 1274 Comm: AudioOut_25 Not tainted 4.4.38-tegra #31
> [   28.513430] Hardware name: quill (DT)
> [   28.517085] Call trace:
> [   28.519531] [<ffffffc000089a84>] dump_backtrace+0x0/0xf8
> [   28.524837] [<ffffffc000089c44>] show_stack+0x14/0x1c
> [   28.529885] [<ffffffc000401c54>] dump_stack+0xac/0xe0
> [   28.534931] [<ffffffc000b35f94>] deactivate_urbs+0x148/0x180
> [   28.540578] [<ffffffc000b37160>] snd_usb_endpoint_stop+0x30/0x58
> [   28.546571] [<ffffffc000b410d8>] snd_usb_substream_playback_trigger+0xa4/0xf4
> [   28.553699] [<ffffffc000b0c160>] snd_pcm_do_stop+0x4c/0x58
> [   28.559179] [<ffffffc000b0bfb4>] snd_pcm_action_single+0x40/0x88
> [   28.565178] [<ffffffc000b0cf14>] snd_pcm_action+0x30/0xf0
> [   28.570568] [<ffffffc000b0fbc8>] snd_pcm_drop+0xac/0x140
> [   28.575873] [<ffffffc000b0fc84>] snd_pcm_release_substream+0x28/0xb0
> [   28.582212] [<ffffffc000b0fd48>] snd_pcm_release+0x3c/0x98
> [   28.587686] [<ffffffc0001e3210>] __fput+0xe0/0x1ac
> [   28.592469] [<ffffffc0001e3334>] ____fput+0xc/0x14
> [   28.597253] [<ffffffc0000c2904>] task_work_run+0xa0/0xc0
> [   28.602558] [<ffffffc0000897bc>] do_notify_resume+0x48/0x60
> [   28.608123] [<ffffffc000084ee8>] work_pending+0x1c/0x20
> 
> In the call path, snd_pcm_stream spinlock has been acquired in
> snd_pcm_drop(). If an output urb is completed between the spinlock
> acquired and deactivate_urbs() clears EP_FLAG_RUNNING for the endpoint,
> its executing of snd_complete_urb() will be blocked for acquiring
> snd_pcm_stream spinlock in snd_pcm_period_elapsed() until the lock is
> released in snd_pcm_drop(). When snd_complete_urb() continues, all jobs
> for deactivate_urbs() are finished.

... but this part is unclear to me.  Do you mean that we have a
deadlock in these two concurrent calls without your patch?


Thanks!

Takashi