After a few more weeks of digging, I have come to the tentative
conclusion that either the XHCI driver, or the USB sound driver,
or both, fail to handle USB errors correctly.

I have some questions at the bottom, after a (brief-ish) explanation
of exactly what seems to go wrong.

TL;DR: arecord from a misbehaving device can hang forever
after a USB error, due to poll on /dev/snd/timer never returning.

The details: under some mysterious circumstances, the PCM290x
family sound chips can send more data than expected during an
isochronous transfer, leading to a babble error. Those
circumstances seem to in part depend on the USB host controller
and/or the electrical environment, since the chips work just
fine for most people.

Receiving data past the end of the isochronous transfer window
scheduled for a device results in the XHCI controller throwing
a babble error, which moves the endpoint into halted state.

This is followed by the host controller software sending a
reset endpoint command, and moving the endpoint into stopped
state, as specified on pages 164-165 of the XHCI specification.

However, the USB sound driver seems to have no idea that this
error happened. The function retire_capture_urb looks at the
status of each isochronous frame, but seems to be under the
assumption that the sound device just keeps on running.

The function snd_complete_urb seems to only detect that the
device is not running if usb_submit_urb returns a failure.

        err = usb_submit_urb(urb, GFP_ATOMIC);
        if (err == 0)
                return;

        usb_audio_err(ep->chip, "cannot submit urb (err = %d)\n", err);

        if (ep->data_subs && ep->data_subs->pcm_substream) {
                substream = ep->data_subs->pcm_substream;
                snd_pcm_stop_xrun(substream);
        }

However, the XHCI driver will happily submit an URB to a
stopped device. Looking at the call trace usb_submit_urb ->
xhci_urb_enqueue -> xhci_queue_isoc_tx_prepare -> prepare_ring,
you can see this code:

        /* Make sure the endpoint has been added to xHC schedule */
        switch (ep_state) {
...
        case EP_STATE_HALTED:
                xhci_dbg(xhci, "WARN halted endpoint, queueing URB anyway.\n");
        case EP_STATE_STOPPED:
        case EP_STATE_RUNNING:
                break;

This leads me to a few questions:
- should retire_capture_urb call snd_pcm_stop_xrun,
  or another function like it, if it sees certain
  errors in the iso frame in the URB?
- should snd_complete_urb do something with these
  errors, too, in case they happen on the sync frames
  and not the data frames?
- does the XHCI code need to ring the doorbell when
  submitting an URB to a stopped device, or is it
  always up to the higher-level driver to fully reset
  the device before it can do anything useful?
- if a device in stopped state does not do anything
  useful, should usb_submit_urb return an error?
- how should the USB sound driver recover from these
  occasional and/or one-off errors? stop the sound
  stream, or try to reinitialize the device and start
  recording again?

I am willing to write patches and can test with my
setup, but both the sound code and the USB code are
new to me so I would like to know what direction I
should go in :)

-- 
All Rights Reversed.