Re: [BUG] I think sleepqueue need to be protected in sleepq_broadcast

On Mon, Sep 8, 2008 at 12:56 PM, John Baldwin <jhb_at_freebsd.org> wrote:
> On Tuesday 02 September 2008 09:40:49 pm Peter Wemm wrote:
[..]
>> I don't know if it is the same problem, but mx2.freebsd.org, running
>> today's 6.4-PRERELEASE just died with:
>> Sep  3 00:20:11 mx2 sshd[15333]: fatal: Read from socket failed: Connection
>> resr panic: Assertion td->td_flags & TDF_SINTR failed at
>> ../../../kern/subr_sleepque5 cpuid = 2
>> KDB: enter: panic
>> FreeBSD 6.4-PRERELEASE #7: Tue Sep  2 19:43:27 UTC 2008
>> This was after about 3 hours of uptime.  It has previously run happily
>> for months at a time before today's rebuild.
>
> So I think what happened is that the thread was woken up while the sleepq
> chain was unlocked while the thread unlocks the sx lock.  The code handles
> this fine already since the same race can happen when dropping the lock while
> checking for signals.  However, in this case TDF_SINTR won't be true anymore.
> The assertion just needs to be updated.  Try this:
>
> Index: subr_sleepqueue.c
> ===================================================================
> --- subr_sleepqueue.c   (revision 182874)
> +++ subr_sleepqueue.c   (working copy)
> _at__at_ -382,7 +382,7 _at__at_
>        CTR3(KTR_PROC, "sleepq catching signals: thread %p (pid %ld, %s)",
>                (void *)td, (long)p->p_pid, p->p_comm);
>
> -       MPASS(td->td_flags & TDF_SINTR);
> +       MPASS((td->td_sleepqueue != NULL) ^ (td->td_flags & TDF_SINTR));
>        mtx_unlock_spin(&sc->sc_lock);
>
>        /* See if there are any pending signals for this thread. */

This is running on mx2 right now.

-- 
Peter Wemm - peter_at_wemm.org; peter_at_FreeBSD.org; peter_at_yahoo-inc.com; KI6FJV
"All of this is for nothing if we don't go to the stars" - JMS/B5
"If Java had true garbage collection, most programs would delete
themselves upon execution." -- Robert Sewell