svn commit: r329448 - head/sys/kern
Konstantin Belousov
kib at freebsd.org
Sat Feb 17 16:26:45 UTC 2018
On Sat, Feb 17, 2018 at 05:07:07PM +0100, Mateusz Guzik wrote:
> On Sat, Feb 17, 2018 at 01:27:38PM +0200, Konstantin Belousov wrote:
> > On Sat, Feb 17, 2018 at 08:48:46AM +0000, Mateusz Guzik wrote:
> > > Author: mjg
> > > Date: Sat Feb 17 08:48:45 2018
> > > New Revision: 329448
> > > URL: https://svnweb.freebsd.org/changeset/base/329448
> > >
> > > Log:
> > > exit: get rid of PROC_SLOCK when checking a process to report
> > Was this tested ?
> >
>
> I was trussing multithreaded microbenchmarks, no issues.
>
> > In particular, are you aware of r309539 ?
> >
>
> So it looks like I misread the code - I have grepped
> thread_suspend_switch operating with the proc locked and misread
> thread_suspend_one's assert as PROC_LOCK_ASSERT.
>
> That said, I think this is harmless. Regardless of the lock the
> inspecting thread can race and check "too soon". Even for a case where
> it decides to report, I don't see anything which would depend on the
> suspending thread to finish.
It was definitely not harmless when I tried to avoid the spin lock there,
but I do not remember exact failure mode. Most likely, it was a missed
report of the traced child indeed, but I am not sure that truss triggered
it. Most likely, Peter Holm was the reporter, since he is listed in
the commit.
>
> However, locking can be employed in a way which is avoided in the common
> case:
>
> diff --git a/sys/kern/kern_exit.c b/sys/kern/kern_exit.c
> index b063bda5b7ff..4ae24bcd7059 100644
> --- a/sys/kern/kern_exit.c
> +++ b/sys/kern/kern_exit.c
> @@ -1174,6 +1174,7 @@ kern_wait6(struct thread *td, idtype_t idtype, id_t
> id, int *status,
> struct proc *p, *q;
> pid_t pid;
> int error, nfound, ret;
> + bool report;
>
> AUDIT_ARG_VALUE((int)idtype); /* XXX - This is likely wrong! */
> AUDIT_ARG_PID((pid_t)id); /* XXX - This may be wrong! */
> @@ -1226,27 +1227,36 @@ kern_wait6(struct thread *td, idtype_t idtype, id_t
> id, int *status,
> PROC_LOCK_ASSERT(p, MA_OWNED);
>
> if ((options & WTRAPPED) != 0 &&
> - (p->p_flag & P_TRACED) != 0 &&
> - (p->p_flag & (P_STOPPED_TRACE | P_STOPPED_SIG)) != 0 &&
> - p->p_suspcount == p->p_numthreads &&
> - (p->p_flag & P_WAITED) == 0) {
> + (p->p_flag & P_TRACED) != 0) {
> + PROC_SLOCK(p);
> + report =
> + ((p->p_flag & (P_STOPPED_TRACE |
> P_STOPPED_SIG)) &&
> + p->p_suspcount == p->p_numthreads &&
> + (p->p_flag & P_WAITED) == 0);
> + PROC_SUNLOCK(p);
> + if (report) {
> CTR4(KTR_PTRACE,
> "wait: returning trapped pid %d status %#x "
> "(xstat %d) xthread %d",
> p->p_pid, W_STOPCODE(p->p_xsig), p->p_xsig,
> p->p_xthread != NULL ?
> p->p_xthread->td_tid : -1);
> - report_alive_proc(td, p, siginfo, status, options,
> - CLD_TRAPPED);
> - return (0);
> + report_alive_proc(td, p, siginfo, status,
> + options, CLD_TRAPPED);
> + return (0);
> + }
> }
> if ((options & WUNTRACED) != 0 &&
> - (p->p_flag & P_STOPPED_SIG) != 0 &&
> - p->p_suspcount == p->p_numthreads &&
> - (p->p_flag & P_WAITED) == 0) {
> - report_alive_proc(td, p, siginfo, status, options,
> + report_alive_proc(td, p, siginfo, status,
> + options, CLD_TRAPPED);
> + return (0);
> + }
> }
> if ((options & WUNTRACED) != 0 &&
> - (p->p_flag & P_STOPPED_SIG) != 0 &&
> - p->p_suspcount == p->p_numthreads &&
> - (p->p_flag & P_WAITED) == 0) {
> - report_alive_proc(td, p, siginfo, status, options,
> - CLD_STOPPED);
> - return (0);
> + (p->p_flag & P_STOPPED_SIG) != 0) {
> + PROC_SLOCK(p);
> + report = (p->p_suspcount == p->p_numthreads &&
> + ((p->p_flag & P_WAITED) == 0));
> + PROC_SUNLOCK(p);
> + if (report) {
> + report_alive_proc(td, p, siginfo, status,
> + options, CLD_STOPPED);
> + return (0);
> + }
> }
> if ((options & WCONTINUED) != 0 &&
> (p->p_flag & P_CONTINUED) != 0) {
>
If this works, I am fine with the patch. Still, I would prefer to have
the original commit reverted for now, until the fix is tested. Then,
commit updated patch.
More information about the svn-src-all
mailing list