Re: [PATCH v4 31/32] sound: usb: card: Allow for rediscovery of connected USB SND devices

[Takashi Iwai <tiwai@suse.de>]

On Mon, 28 Aug 2023 23:25:14 +0200,
Wesley Cheng wrote:
> 
> Hi Takashi,
> 
> On 7/25/2023 2:27 AM, Takashi Iwai wrote:
> > On Tue, 25 Jul 2023 11:15:11 +0200,
> > Pierre-Louis Bossart wrote:
> >> 
> >> 
> >> 
> >> On 7/25/23 04:34, Wesley Cheng wrote:
> >>> In case of notifying SND platform drivers of connection events, some of
> >>> these use cases, such as offloading, require an ASoC USB backend device to
> >>> be initialized before the events can be handled.  If the USB backend device
> >>> has not yet been probed, this leads to missing initial USB audio device
> >>> connection events.
> >>> 
> >>> Expose an API that traverses the usb_chip array for connected devices, and
> >>> to call the respective connection callback registered to the SND platform
> >>> driver.
> >>> 
> >>> Signed-off-by: Wesley Cheng <quic_wcheng@quicinc.com>
> >>> ---
> >>>   sound/usb/card.c | 19 +++++++++++++++++++
> >>>   sound/usb/card.h |  2 ++
> >>>   2 files changed, 21 insertions(+)
> >>> 
> >>> diff --git a/sound/usb/card.c b/sound/usb/card.c
> >>> index 365f6d978608..27a89aaa0bf3 100644
> >>> --- a/sound/usb/card.c
> >>> +++ b/sound/usb/card.c
> >>> @@ -170,6 +170,25 @@ struct snd_usb_stream *snd_usb_find_suppported_substream(int card_idx,
> >>>   }
> >>>   EXPORT_SYMBOL_GPL(snd_usb_find_suppported_substream);
> >>>   +/*
> >>> + * in case the platform driver was not ready at the time of USB SND
> >>> + * device connect, expose an API to discover all connected USB devices
> >>> + * so it can populate any dependent resources/structures.
> >>> + */
> >>> +void snd_usb_rediscover_devices(void)
> >>> +{
> >>> +	int i;
> >>> +
> >>> +	mutex_lock(&register_mutex);
> >>> +	for (i = 0; i < SNDRV_CARDS; i++) {
> >>> +		if (usb_chip[i])
> >>> +			if (platform_ops && platform_ops->connect_cb)
> >>> +				platform_ops->connect_cb(usb_chip[i]);
> >> 
> >> what happens if the USB device is removed while the platform device adds
> >> a port?
> > 
> > That should be protected by the register_mutex.  But there can be
> > other races (see below :)
> > 
> >> This sounds super-racy to me. It's the same set of problems we're having
> >> between audio and display/DRM, I would be surprised if this function
> >> dealt with all corner cases of insertion/removal, bind/unbind.
> > 
> > Yes, we need to be more careful about binding.
> > 
> > For example, in the current patch set, I see no way to prevent
> > unloading snd-usb-audio-qmi module, and it allows user to cut off the
> > stuff during operation, which may break things while the kernel is
> > running the code of the unloaded module.  You need to have a proper
> > module refcount management for avoiding such a scenario.  Most of
> > drivers don't need it because ALSA core part already takes care of
> > it.  But in this case, it requires a manual adjustment.
> > 
> 
> Sorry for the delayed response...this was routed to another folder,
> and missed it.
> 
> I see, that is a valid situation, so I think the best way to address
> it is to do something like the following:
> 
> static void __exit qc_usb_audio_offload_exit(void)
> {
> ...
> 	snd_usb_unregister_platform_ops();
> 	for (idx = 0; idx < SNDRV_CARDS; idx++)
> 		qc_usb_audio_offload_disconnect(uadev[idx].chip);
> 
> We'll first unregister the platform ops, so that we get no further
> connect/disconnect CBs, and then we'll forcefully ensure that all
> devices are removed/cleaned.  Part of the USB offload disconnect
> sequence will also forcefully stop the offload path on the external
> DSP by issuing a disconnect indication QMI message.
> 
> Then we can safely clean up the rest of the resources.  We do have
> refcounting in place for several of the other structures, but I think
> in the module exit case, we need to ensure the offload path is stopped
> fully.

Yes, the proper disconnection procedure is mandatory.
In addition, you can have a module refcount increment upon the card
binding, too.  This may prevent the unexpected exit (although it's
often still possible to unbind via sysfs).


thanks,

Takashi