Re: kernel locks due to USB I/O

[Alan Stern <stern@rowland.harvard.edu>]

On Thu, Nov 19, 2020 at 01:50:13PM -0500, Alberto Sentieri wrote:
> > > However, just before the lock, I got this sequence:
> > > 
> > > [  387.326795] usb 1-1.5.3.1: usbdev_do_ioctl: SUBMITURB
> > > [  387.326801] usb 1-1.5.3.1: userurb 0000000000000000, ep1 int-in, length
> > > 64
> > > [  387.327531] usb 1-1.5.3.1: urb complete
> > > [  387.327539] usb 1-1.5.3.1: userurb         pK-error, ep1 int-out,
> > > actual_length 64 status 0
> > > [  387.327582] usb 1-1.5.3.1: usbdev_do_ioctl: REAPURBNDELAY
> > > [  387.327590] usb 1-1.5.3.1: reap 0000000000000000
> > > [  387.327598] usb 1-1.5.3.1: usbdev_do_ioctl: REAPURBNDELAY
> > > [  388.109400] usb 3-1.5.7.1: usbdev_do_ioctl: REAPURBNDELAY
> > > [  388.110032] usb 3-1.6.7.1: usbdev_do_ioctl: REAPURBNDELAY
> > > [  388.110169] usb 3-1.5.5.1: usbdev_do_ioctl: REAPURBNDELAY
> > > [  388.110605] usb 3-1.6.5.1: usbdev_do_ioctl: REAPURBNDELAY
> > > [  388.111394] usb 3-1.1.1: usbdev_do_ioctl: REAPURBNDELAY
> > > [  388.403278] usb 3-1.6.4.1: usbdev_do_ioctl: REAPURBNDELAY
> > > [  388.404011] usb 3-1.5.1.1: usbdev_do_ioctl: REAPURBNDELAY
> > > [  388.404344] usb 3-1.6.3.1: usbdev_do_ioctl: REAPURBNDELAY
> > > [  388.404649] usb 3-1.4.1: usbdev_do_ioctl: REAPURBNDELAY
> > > [  388.404790] usb 3-1.3.1: usbdev_do_ioctl: REAPURBNDELAY
> > > [  388.405283] usb 3-1.5.3.1: usbdev_do_ioctl: REAPURBNDELAY
> > > [  388.407828] usb 3-1.6.1.1: usbdev_do_ioctl: REAPURBNDELAY
> > > [  388.408771] usb 3-1.5.4.1: usbdev_do_ioctl: REAPURBNDELAY
> > > [  388.409155] usb 3-1.5.2.1: usbdev_do_ioctl: REAPURBNDELAY
> > > [  388.409452] usb 3-1.5.6.1: usbdev_do_ioctl: REAPURBNDELAY
> > > [  388.409743] usb 3-1.6.6.1: usbdev_do_ioctl: REAPURBNDELAY
> > > [  388.409848] usb 3-1.2.1: usbdev_do_ioctl: REAPURBNDELAY
> > > [  388.410840] usb 3-1.6.2.1: usbdev_do_ioctl: REAPURBNDELAY
> > This rapid sequence of reap requests with no completions or actual
> > reaps is odd.  Is this the sort of thing you expect your program to
> > do?  Maybe you can add some debugging output to the program.
> What apparently happened before that (but unfortunately I did not capture it
> this time) is that the particular device return errno -71 in some URBs,
> without de-enumerating.

That's actually quite bad.  -71 (-EPROTO) indicates a serious error in
the USB protocol.  Most often it means that the kernel did not receive
a reply packet from the device -- this could be caused by the device's
firmware crashing or a USB cable failing.  In a well-functioning
system it should never occur.

> However, I am only seeing userurbs with status 0 in
> my the netconsole capture, before the ones with status -108 (just before the
> lock). So, if error -71 happened, all URBs are reaped and the devices are
> then closed. Usually I see error -71 in the program log happening for a few
> devices at the same time.

That probably means there's something wrong with a hub they are all
connected to.  But of course, a problem in a hub wouldn't explain why
the whole system freezes.

>  After closing the interface, the program opens it
> again. The return from open should be a valid file descriptor, otherwise it
> would fail. After opening the device successfully it starts submitting urbs,
> and error -108 is coming back, telling me that the device just opened is not
> open.

It means that the device is no longer present.  Perhaps it was
logically (if not physically) disconnected after the open but before
the URBs were submitted.

> It just came to my attention now: maybe something to be investigated is what
> the kernel does if a interface is closed due to a URB error and immediately
> opened again.

In itself, that should work perfectly.  But whatever caused the USB
error could go on to cause other things to go wrong.

It might help to see a usbmon trace.  You merely have to send the data 
from /sys/kernel/debug/usb/usbmon/1u (or 3u, or even both -- the number 
refers to the USB bus being monitored) through a UDP pipe, using netcat 
or something similar, to another computer which could safely store the 
output.

Alan Stern