On 7/30/19 4:27 PM, Konstantin Belousov wrote: > On Wed, Jul 31, 2019 at 02:13:02AM +0300, Konstantin Belousov wrote: >> On Tue, Jul 30, 2019 at 03:04:57PM -0700, Bryan Drewery wrote: >>> On 7/15/2019 12:18 PM, Konstantin Belousov wrote: >>>> Author: kib >>>> Date: Mon Jul 15 19:18:25 2019 >>>> New Revision: 350005 >>>> URL: https://svnweb.freebsd.org/changeset/base/350005 >>>> >>>> Log: >>>> In do_sem2_wait(), balance umtx_key_get() with umtx_key_release() on >>>> retry. >>>> >>> >>> Is this also needed in do_sem_wait()? A similar pattern seems to be there. >> No, I do not think do_sem_wait() has similar issue, because the again label >> does not re-get the key. >> >>> >>> I ask because of what I referenced on IRC. I have some processes stuck >>> in here from a 10.4 jail. >>> >>>> ~/git/poudriere # procstat -kk 1498 >>>> PID TID COMM TDNAME KSTACK >>>> 1498 100710 python2.7 - mi_switch+0x174 >>>> sleepq_switch+0x110 sleepq_catch_signals+0x417 sleepq_wait_sig+0xf >>>> _sleep+0x2d0 umtxq_sleep+0x153 do_sem_wait+0x42c __umtx_op_sem_wait+0x6e >>>> amd64_syscall+0x2bb fast_syscall_common+0x101 >>>> 1498 101575 python2.7 - mi_switch+0x174 >>>> sleepq_switch+0x110 sleepq_catch_signals+0x417 sleepq_wait_sig+0xf >>>> _sleep+0x2d0 umtxq_sleep+0x153 do_sem_wait+0x42c __umtx_op_sem_wait+0x6e >>>> amd64_syscall+0x2bb fast_syscall_common+0x101 >>>> 1498 101657 python2.7 - <running> >>> ... >>>> ~/git/poudriere # procstat -kk 1498 >>>> PID TID COMM TDNAME KSTACK >>>> 1498 100710 python2.7 - mi_switch+0x174 >>>> sleepq_switch+0x110 sleepq_catch_signals+0x417 sleepq_wait_sig+0xf >>>> _sleep+0x2d0 umtxq_sleep+0x153 do_sem_wait+0x42c __umtx_op_sem_wait+0x6e >>>> amd64_syscall+0x2bb fast_syscall_common+0x101 >>>> 1498 101575 python2.7 - mi_switch+0x174 >>>> sleepq_switch+0x110 sleepq_catch_signals+0x417 sleepq_wait_sig+0xf >>>> _sleep+0x2d0 umtxq_sleep+0x153 do_sem_wait+0x42c __umtx_op_sem_wait+0x6e >>>> amd64_syscall+0x2bb fast_syscall_common+0x101 >>>> 1498 101657 python2.7 - do_sem_wait+0x1b6 >>>> __umtx_op_sem_wait+0x6e amd64_syscall+0x2bb fast_syscall_common+0x101 >>> ... >>>> ~/git/poudriere # procstat -kk 94392 >>>> PID TID COMM TDNAME KSTACK >>>> 94392 101815 python2.7 - mi_switch+0x174 >>>> sleepq_switch+0x110 sleepq_catch_signals+0x417 sleepq_wait_sig+0xf >>>> _sleep+0x2d0 umtxq_sleep+0x153 do_sem_wait+0x42c __umtx_op_sem_wait+0x6e >>>> amd64_syscall+0x2bb fast_syscall_common+0x101 >>>> 94392 101816 python2.7 - >>>> __mtx_lock_sleep+0x118 __mtx_lock_flags+0x102 _sleep+0x334 umtxq_busy+0xb7 >>>> do_sem_wait+0x161 __umtx_op_sem_wait+0x6e amd64_syscall+0x2bb >>>> fast_syscall_common+0x101 >>>> 94392 102076 python2.7 - __mtx_lock_flags+0x94 >>>> do_sem_wait+0x228 __umtx_op_sem_wait+0x6e amd64_syscall+0x2bb >>>> fast_syscall_common+0x101 >> >> Try this. We should only retry casueword if it failed spuriously. >> >> diff --git a/sys/kern/kern_umtx.c b/sys/kern/kern_umtx.c >> index bb998457975..6c914ab6f3e 100644 >> --- a/sys/kern/kern_umtx.c >> +++ b/sys/kern/kern_umtx.c >> @@ -3229,7 +3229,8 @@ do_sem_wait(struct thread *td, struct _usem *sem, >> struct _umtx_time *timeout) >> rv = casueword32(&sem->_has_waiters, 0, &count1, 1); >> if (rv == 0) >> rv1 = fueword32(&sem->_count, &count); >> - if (rv == -1 || (rv == 0 && (rv1 == -1 || count != 0)) || rv == 1) { >> + if (rv == -1 || (rv == 0 && (rv1 == -1 || count != 0)) || >> + (rv == 1 && count1 == 0)) { >> umtxq_lock(&uq->uq_key); >> umtxq_unbusy(&uq->uq_key); >> umtxq_remove(uq); > > I think there is another problem, since even despite our intent of looping > just because of casueword returned 1, the umtxq_check_susp() should have > terminated the loop. I believe the following update would fix that. > > If you have time, can you please apply only the umtxq_check_susp() chunk > and see if it helps as well ? [Both chunks are needed for correctness, but > the umtxq_check_susp() is almost impossible to test in combination] > > diff --git a/sys/kern/kern_umtx.c b/sys/kern/kern_umtx.c > index bb998457975..08bdd1a1a9a 100644 > --- a/sys/kern/kern_umtx.c > +++ b/sys/kern/kern_umtx.c > @@ -723,13 +723,11 @@ umtxq_check_susp(struct thread *td, bool sleep) > error = 0; > p = td->td_proc; > PROC_LOCK(p); > - if (P_SHOULDSTOP(p) || > - ((p->p_flag & P_TRACED) && (td->td_dbgflags & TDB_SUSPEND))) { > - if (p->p_flag & P_SINGLE_EXIT) > - error = EINTR; > - else > - error = sleep ? thread_suspend_check(0) : ERESTART; > - } > + if (p->p_flag & P_SINGLE_EXIT) > + error = EINTR; > + else if (P_SHOULDSTOP(p) || > + ((p->p_flag & P_TRACED) && (td->td_dbgflags & TDB_SUSPEND))) > + error = sleep ? thread_suspend_check(0) : ERESTART; > PROC_UNLOCK(p); > return (error); > } > @@ -3229,7 +3227,8 @@ do_sem_wait(struct thread *td, struct _usem *sem, > struct _umtx_time *timeout) > rv = casueword32(&sem->_has_waiters, 0, &count1, 1); > if (rv == 0) > rv1 = fueword32(&sem->_count, &count); > - if (rv == -1 || (rv == 0 && (rv1 == -1 || count != 0)) || rv == 1) { > + if (rv == -1 || (rv == 0 && (rv1 == -1 || count != 0)) || > + (rv == 1 && count1 == 0)) { > umtxq_lock(&uq->uq_key); > umtxq_unbusy(&uq->uq_key); > umtxq_remove(uq); >
This 2nd change alone (&& count1 == 0) was sufficient to fix the endless loop problem. I am not sure how to test the umtxq_check_susp() change. Do I just need to ptrace the process? -- Regards, Bryan Drewery
signature.asc
Description: OpenPGP digital signature