svn commit: r350005 - head/sys/kern
Bryan Drewery
bdrewery at FreeBSD.org
Wed Jul 31 02:19:08 UTC 2019
On 7/30/2019 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]
What is the expected "working" test for the umtxq_check_susp() change? I
get the 100% CPU and kill -9 does not kill it.
> ~ # procstat -kk 19150
> PID TID COMM TDNAME KSTACK
> 19150 101524 python2.7 - do_sem_wait+0x150 __umtx_op_sem_wait+0x6e amd64_syscall+0x2bb fast_syscall_common+0x101
> 19150 101553 python2.7 - mi_switch+0x174 sleepq_switch+0x110 sleepq_catch_signals+0x417 sleepq_wait_sig+0xf _sleep+0x2d0 umtxq_sleep+0x153 do_sem_wait+0x41c __umtx_op_sem_wait+0x6e amd64_syscall+0x2bb fast_syscall_common+0x101
> 19150 101554 python2.7 - __mtx_lock_sleep+0x118 __mtx_lock_flags+0x102 _sleep+0x334 umtxq_busy+0xb7 do_sem_wait+0x150 __umtx_op_sem_wait+0x6e amd64_syscall+0x2bb fast_syscall_common+0x101
>
> 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);
>
--
Regards,
Bryan Drewery
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 618 bytes
Desc: OpenPGP digital signature
URL: <http://lists.freebsd.org/pipermail/svn-src-all/attachments/20190730/7c99578a/attachment.sig>
More information about the svn-src-all
mailing list