On Tue, Apr 20, 2021 at 04:22:53PM -0400, Douglas Gilbert wrote:
> On 2021-04-20 2:52 a.m., Ming Lei wrote:
> > On Tue, Apr 20, 2021 at 12:54 PM Douglas Gilbert <dgilbert@xxxxxxxxxxxx> wrote:
> > >
> > > On 2021-04-19 11:22 p.m., Bart Van Assche wrote:
> > > > On 4/19/21 8:06 PM, Douglas Gilbert wrote:
> > > > > I have always suspected under extreme pressure the block layer (or scsi
> > > > > mid-level) does strange things, like an IO hang, attempts to prove that
> > > > > usually lead back to my own code :-). But I have one example recently
> > > > > where upwards of 10 commands had been submitted (blk_execute_rq_nowait())
> > > > > and the following one stalled (all on the same thread). Seconds later
> > > > > those 10 commands reported DID_TIME_OUT, the stalled thread awoke, and
> > > > > my dd variant went to its conclusion (reporting 10 errors). Following
> > > > > copies showed no ill effects.
> > > > >
> > > > > My weapons of choice are sg_dd, actually sgh_dd and sg_mrq_dd. Those last
> > > > > two monitor for stalls during the copy. Each submitted READ and WRITE
> > > > > command gets its pack_id from an incrementing atomic and a management
> > > > > thread in those copies checks every 300 milliseconds that that atomic
> > > > > value is greater than the previous check. If not, dump the state of the
> > > > > sg driver. The stalled request was in busy state with a timeout of 1
> > > > > nanosecond which indicated that blk_execute_rq_nowait() had not been
> > > > > called. So the chief suspect would be blk_get_request() followed by
> > > > > the bio setup calls IMO.
> > > > >
> > > > > So it certainly looked like an IO hang, not a locking, resource nor
> > > > > corruption issue IMO. That was with a branch off MKP's
> > > > > 5.13/scsi-staging branch taken a few weeks back. So basically
> > > > > lk 5.12.0-rc1 .
> > > >
> > > > Hi Doug,
> > > >
> > > > If it would be possible to develop a script that reproduces this hang and
> > > > if that script can be shared I will help with root-causing and fixing this
> > > > hang.
> > >
> > > Possible, but not very practical:
> > > 1) apply supplied 83 patches to sg driver
> > > 2) apply pending patch to scsi_debug driver
> > > 3) find a stable kernel platform (perhaps not lk 5.12.0-rc1)
> > > 4) run supplied scripts for three weeks
> > > 5) dig through the output and maybe find one case (there were lots
> > > of EAGAINs from blk_get_request() but they are expected when
> > > thrashing the storage layers)
> >
> > Or collecting the debugfs log after IO hang is triggered in your test:
> >
> > (cd /sys/kernel/debug/block/$SDEV && find . -type f -exec grep -aH . {} \;)
> >
> > $SDEV is the disk on which IO hang is observed.
>
> Thanks. I'll try adding that to my IO hang trigger code.
>
> My patches on the sg driver add debugfs support so these produce
> the same output:
> cat /proc/scsi/sg/debug
> cat /sys/kernel/debug/scsi_generic/snapshot
>
> There is also a /sys/kernel/debug/scsi_generic/snapped file whose
> contents reflect the driver's state when ioctl(<sg_fd>, SG_DEBUG, &one)
> was last called.
>
> When I test, the root file system is usually on a NVMe SSD so the
> state of all SCSI disks present should be dumped as they are part
> of my test. Also I find the netconsole module extremely useful and
> have an old laptop on my network running:
> socat udp-recv:6665 - > socat.txt
>
> picking up the UDP packets from netconsole on port 6665. Not quite as
> good as monitoring a hardware serial console, but less fiddly. And
> most modern laptops don't have access to a serial console so
> netconsole is the only option.
Yeah, years ago netconsole does help me much when serial console isn't
available, especially ssh doesn't work at that time, such as kernel
panic.
>
> Another observation: upper level issues seem to impact the submission
> side of request handling (e.g. the IO hang I was trying to describe)
> while error injection I can do (e.g. using the scsi_debug driver)
> impacts the completion side (obviously). Are there any tools to inject
> errors into the block layer submission code?
block layer supports fault injection in submission side, see
should_fail_bio() which is called from submit_bio_checks().
thanks,
Ming
