Re: USB Disk Stalls on -current

On Sun, Feb 6, 2022 at 12:02 PM Sean Bruno <sbruno@freebsd.org> wrote:

>
>
> >
> >
> > So there's some tools you can use. For usb, there's usbdump that can
> > get you the USB transactions. I've not used it enough to give more
> details
> > here. This will let you know what's going on, and when, on the USB
> endpoint.
> >
> > You can also enable the CAM_IOSCHED stuff. This will allow you to get
> > latency
> > measurements for 'requests in the sim' which basically will tell you
> > what your
> > latency spread is for the drives. This will tell you if things are
> > getting caught
> > up in the USB layer, or after CAM's da driver completes the I/O request
> > (granted, that's almost certainly not happening, but it will help you
> > figure out
> > what's going on and put numbers to the oddities you are seeing).
> >
> > Also, make sure you have good cables. I've had lots of hicups over the
> > years from dodgy USB cables. Also make sure you have good, high quality
> > enclosures. Many from the USB2 time-period are sketchy at best and I
> > went through several at one point trying to find a good one. I'd be
> > tempted to
> > get USB 3 enclosures. I've had better luck with USB3 gear than USB2 gear
> > here, but you need a USB-3 controller to get USB-3 speeds which might not
> > be compatible with the NUC's built-in stuff (though my NUC has one USB3
> > port, there's lots of different models).
> >
> > Usually, though, I see weirdness associated with dmesg messages from
> > usb, cam, etc when the hardware is on the sketch end.
> >
> > Warner
>
> I'm assuming that I have a fairly dodgy USB device, as the pauses seem
> to correspond to this from CAM being emitted:
>
> Feb  6 11:56:43 alice kernel: (da0:umass-sim1:1:0:0): READ(10). CDB: 28
> 00 36 69 02 6e 00 00 80 00
> Feb  6 11:56:43 alice kernel: (da0:umass-sim1:1:0:0): CAM status: CCB
> request completed with an error
> Feb  6 11:56:43 alice kernel: (da0:umass-sim1:1:0:0): Retrying command,
> 2 more tries remain
>
>
> Things resume after this is emitted, but there is a substantial
> (multiple minutes) pause here.  I would assume that timeouts would fire
> much quicker.
>

The default timeout is 60s.

You can reduce that substantially by setting kern.cam.da.default_timeout
to a smaller level. Disk operations completed within 5s these days,
except spin ups. Heck, nearly all complete within 500ms. You
might try setting this value to maybe 3 or 5 or 10 to see if that helps the
hiccups without introducing extra retries when the load is heavy. The
smaller values give a faster recovery, but too small a number may result
in timeouts and errors under load. I think you need to set this as a
tuneable.

Warner