Re: Sleeping thread (tid 100033, pid 16): panic in FreeBSD 10.0-CURRENT/amd64 r228662

On Tuesday, December 20, 2011 9:20:09 am Attilio Rao wrote:
> 2011/12/20 John Baldwin <jhb_at_freebsd.org>:
> > On Saturday, December 17, 2011 10:41:15 pm mdf_at_freebsd.org wrote:
> >> On Sat, Dec 17, 2011 at 5:45 PM, Alexander Kabaev <kabaev_at_gmail.com> wrote:
> >> > On Sun, 18 Dec 2011 01:09:00 +0100
> >> > "O. Hartmann" <ohartman_at_zedat.fu-berlin.de> wrote:
> >> >
> >> >> Sleeping thread (tid 100033, pid 16) owns a non sleepable lock
> >> >> panic: sleeping thread
> >> >> cpuid = 0
> >> >>
> >> >> PID 16 is always USB on my box.
> >> >
> >> > You really need to give us a backtrace when you quote panics. It is
> >> > impossible to make any sense of the above panic message without more
> >> > context.
> >>
> >> In the case of this panic, the stack of the thread which panics is
> >> useless; it's someone trying to propagate priority that discovered it.
> >>  A backtrace on tid 100033 would be useful.
> >>
> >> With WITNESS enabled, it's possible to have this panic display the
> >> stack of the incorrectly sleeping thread at the time it acquired the
> >> lock, as well, but this code isn't in CURRENT or any release.  I have
> >> a patch at $WORK I can dig up on Monday.
> >
> > Huh?  The stock kernel dumps a stack trace of the offending thread if you have
> > DDB enabled:
> >
> >                /*
> >                 * If the thread is asleep, then we are probably about
> >                 * to deadlock.  To make debugging this easier, just
> >                 * panic and tell the user which thread misbehaved so
> >                 * they can hopefully get a stack trace from the truly
> >                 * misbehaving thread.
> >                 */
> >                if (TD_IS_SLEEPING(td)) {
> >                        printf(
> >                "Sleeping thread (tid %d, pid %d) owns a non-sleepable lock\n",
> >                            td->td_tid, td->td_proc->p_pid);
> > #ifdef DDB
> >                        db_trace_thread(td, -1);
> > #endif
> >                        panic("sleeping thread");
> >                }
> >
> > It may be that we can make use of the STACK API here instead to output this
> > trace even when DDB isn't enabled.  The patch below tries to do that
> > (untested).  It does some odd thigns though since it is effectively running
> > from a panic context already, so it uses a statically allocated 'struct stack'
> > rather than using stack_create() and uses stack_print_ddb() since it is
> > holding spin locks and can't possibly grab an sx lock:
> >
> > Index: subr_turnstile.c
> > ===================================================================
> > --- subr_turnstile.c    (revision 228534)
> > +++ subr_turnstile.c    (working copy)
> > _at__at_ -72,6 +72,7 _at__at_ __FBSDID("$FreeBSD$");
> >  #include <sys/proc.h>
> >  #include <sys/queue.h>
> >  #include <sys/sched.h>
> > +#include <sys/stack.h>
> >  #include <sys/sysctl.h>
> >  #include <sys/turnstile.h>
> >
> > _at__at_ -175,6 +176,7 _at__at_ static void turnstile_fini(void *mem, int size);
> >  static void
> >  propagate_priority(struct thread *td)
> >  {
> > +       static struct stack st;
> >        struct turnstile *ts;
> >        int pri;
> >
> > _at__at_ -217,8 +219,10 _at__at_ propagate_priority(struct thread *td)
> >                        printf(
> >                "Sleeping thread (tid %d, pid %d) owns a non-sleepable lock\n",
> >                            td->td_tid, td->td_proc->p_pid);
> > -#ifdef DDB
> > -                       db_trace_thread(td, -1);
> > +#ifdef STACK
> > +                       stack_zero(&st);
> > +                       stack_save_td(&st, td);
> > +                       stack_print_ddb(&st);
> >  #endif
> >                        panic("sleeping thread");
> >                }
> >
> > --
> 
> I'm not sure it is a wise idea to trimm the DDB part, because it is
> much more common than STACK enabled. Note that stack(9) is working if
> you define DDB too, so I'd say to do that for both.
> Also, I don't think you need the stack_zero() prior to set it.

Err, STACK is enabled in GENERIC in release kernels but DDB is not, so I think
STACK is the more common one.  As far as stack_zero(), I was just being paranoid.

> As we are here, however, I have a question for Robert here: do you
> think we should support the _ddb() variant of options even in the case
> DDB is not enabled in the kernel?
> Probabilly the way it is nowadays is easier to have stack(9) both
> defined for DDB and STACK, but in general I wouldn't advertise that.

The _ddb variants are always enabled by my reading.  They just use different
entry points into the linker that don't use locking.

-- 
John Baldwin