Re: [syzbot] [xfs?] INFO: task hung in __fdget_pos (4)

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]


On Sun, Sep 3, 2023 at 10:04 PM Mateusz Guzik <mjguzik@xxxxxxxxx> wrote:
> On 9/3/23, Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:
> > On Sun, Sep 03, 2023 at 08:57:23PM +0200, Mateusz Guzik wrote:
> >
> >> This does not dump backtraces, just a list of tasks + some stats.
> >>
> >> The closest to useful here I found are 'w' ("Dumps tasks that are in
> >> uninterruptable (blocked) state.") and 'l' ("Shows a stack backtrace
> >> for all active CPUs."), both of which can miss the task which matters
> >> (e.g., stuck in a very much *interruptible* state with f_pos_lock
> >> held).
> >>
> >> Unless someone can point at a way to get all these stacks, I'm going
> >> to hack something up in the upcoming week, if only for immediate
> >> syzbot usage.
> >
> > Huh?  Sample of output here:
> > 2023-09-03T15:34:36.271833-04:00 duke kernel: [87367.574459] task:ssh-agent
> >      state:S stack:0     pid:3949  ppid:3947   flags:0x
< .. >
> > 2023-09-03T15:34:36.457355-04:00 duke kernel: [87367.759981]  </TASK>
> >
> > Looks like a stack trace to me; seeing one of the callers of fdget_pos()
> > in that would tell you who's currently holding *some* ->f_pos_lock.
> >
> > That - on 6.1.42, with fairly bland .config (minimal debugging;
> > I need that box for fast builds, among other things).  Enable
> > lockdep and you'll get who's holding which logs in addition
> > to those stack traces...
> >
> That's my bad, the only content which survived for me in terminal
> buffer was a spew for the scheduler and a ps-like task list.
> Stacktraces are indeed there. Sorry for the noise on this one. I
> verified stack dumps are not gated by any defines either btw, so it's
> all good on this front.
> All this brainos aside, I added Aleksandr Nogikh to To: in my previous
> e-mail. From short poking around I found he was responsive to some
> queries concerning Linux vs syzbot and I figured would a good person
> to ask.
> So, Aleksandr, tl;dr would it be a problem to add the 't' sysrq to
> syzbot in order to dump backtraces from all threads? Either break t
> over a serial port or t written to /proc/sysrq-trigger. Lack of this
> info blocks progress on the issue reported here
> (

That's a good suggestion, thanks!
I think it should be doable and would indeed simplify debugging of our
reports (even if not for this particular one, judging by the
conversation below :) ).

I've filed


> --
> Mateusz Guzik <mjguzik>

[Index of Archives]     [XFS Filesystem Development (older mail)]     [Linux Filesystem Development]     [Linux Audio Users]     [Yosemite Trails]     [Linux Kernel]     [Linux RAID]     [Linux SCSI]

  Powered by Linux