2.6.28, limiting cpu time doesn't work
Subsystems (3)
Hello,
after upgrade to 2.6.28 ulimit -t doesn't work. for example:
bash# ulimit -t 3; cpuhog
(where cpuhog is any program that continuously use cpu)
with 2.6.27.20 cpuhog gets killed after 3sec as expected.
with 2.6.28, 2.6.28.8, 2.6.29-rc8-git4 it's keep running indefinitely. ulimit -a and /proc/<pid>/limits show that cputime limit was set correctly.
have i missed some global api changes? recompilation of recent glibc with new kernel headers and shell doesn't help...
ps please cc me, i'm not on the list
Quoted 9 lines · Peter Lojkin
after upgrade to 2.6.28 ulimit -t doesn't work. for example:
bash# ulimit -t 3; cpuhog
(where cpuhog is any program that continuously use cpu)
with 2.6.27.20 cpuhog gets killed after 3sec as expected.
with 2.6.28, 2.6.28.8, 2.6.29-rc8-git4 it's keep running indefinitely. ulimit -a and /proc/<pid>/limits show that cputime limit was set correctly.
Found this message on http://bugzilla.kernel.org/show_bug.cgi?id=12911 ...
I _think_ posix_cpu_timers_init_group() is not right, it should copy cputime_expires->prof_exp.
Peter, any chance you can test the (uncompiled/untested) patch below?
Also, I assume that something like
$ ulimit -t 3 $ while true; do true; done
kills the shell correctly, yes? IOW, I suspect that ulimit works, but cpuhog never check RLIMIT_CPU because fastpath_timer_check() always returns 0 due to task_cputime_zero(&sig->cputime_expires) == T.
I'm afraid we need the fix fo 2.6.29 as well, but I am looking at rc3.
Hmm. check_process_timers() updates ->cputime_expires, but it never clears (say) cputime_expires.prof_exp, why? Can't we just do
if (cputime_gt(sig->cputime_expires.prof_exp, prof_expires)) sig->cputime_expires.prof_exp = prof_expires;
at the end?
Oleg.
kernel/fork.c +1 -3
@@ -790,9 +790,7 @@ static void posix_cpu_timers_init_group( sig->it_prof_incr = cputime_zero;/* Cached expiration times. */ - sig->cputime_expires.prof_exp = cputime_zero; - sig->cputime_expires.virt_exp = cputime_zero; - sig->cputime_expires.sched_exp =0 ; + sig->cputime_expires = current->signal->cputime_expires;/* The timer lists. */ INIT_LIST_HEAD (&sig->cpu_timers[0 ]);
Quoted 6 lines · Oleg Nesterov
Found this message on http://bugzilla.kernel.org/show_bug.cgi?id=12911 ...
I _think_
posix_cpu_timers_init_group()is not right, it should copycputime_expires->prof_exp.Peter, any chance you can test the (uncompiled/untested) patch below?
yes, with this patch 2.6.28.8 works as expected, thank you!
Quoted 6 lines
Also, I assume that something like
$ ulimit -t 3 $ while true; do true; donekills the shell correctly, yes?
yes, i've found it after original message and posted it in follow up:
http://marc.info/?l=linux-kernel&m=123770019023095&w=4
Quoted 4 lines
IOW, I suspect that ulimit works, but
cpuhog never checkRLIMIT_CPUbecausefastpath_timer_check()always returns 0 due totask_cputime_zero(&sig->cputime_expires)== T.
Quoted 8 lines
I'm afraid we need the fix fo 2.6.29 as well, but I am looking at rc3.
Hmm.
check_process_timers()updates ->cputime_expires, but it never clears (say) cputime_expires.prof_exp, why? Can't we just doif (cputime_gt(sig->cputime_expires.prof_exp, prof_expires)) sig->cputime_expires.prof_exp = prof_expires;at the end?
if you need to test any more patches on the subject i'm ready to do it. regression test system for our project depends on ability to limit cpu time, so it's major problem for us...
Quoted 14 lines
kernel/fork.c +1
-3@@ -790,9 +790,7 @@ static void posix_cpu_timers_init_group( sig->it_prof_incr = cputime_zero;/* Cached expiration times. */ - sig->cputime_expires.prof_exp = cputime_zero; - sig->cputime_expires.virt_exp = cputime_zero; - sig->cputime_expires.sched_exp =0 ; + sig->cputime_expires = current->signal->cputime_expires;/* The timer lists. */ INIT_LIST_HEAD (&sig->cpu_timers[0 ]);
Quoted 11 lines · Peter Lojkin
Quoted 6 lines · Oleg Nesterov
Found this message on http://bugzilla.kernel.org/show_bug.cgi?id=12911 ...
I _think_
posix_cpu_timers_init_group()is not right, it should copycputime_expires->prof_exp.Peter, any chance you can test the (uncompiled/untested) patch below?
yes, with this patch 2.6.28.8 works as expected, thank you!
Great, thanks!
Quoted 3 lines
if you need to test any more patches on the subject i'm ready to do it. regression test system for our project depends on ability to limit cpu time, so it's major problem for us...
I am not sure what should we do, this needs more discussion.
Probably the most simple patch for -stable and 2.6.29 is below. (with this patch we don't even need update_rlimit_cpu(), afaics).
Oleg.
kernel/posix-cpu-timers.c +2 -1
@@ -1263,7 +1263,8 @@ static inline int fastpath_timer_check(sif (task_cputime_expired (&group_sample, &sig->cputime_expires))return 1 ; } -return 0 ; + +return sig->rlim[RLIMIT_CPU ].rlim_cur !=RLIM_INFINITY ; }/*
Quoted 29 lines · Oleg Nesterov
Quoted 11 lines · Peter Lojkin
Quoted 6 lines · Oleg Nesterov
Found this message on http://bugzilla.kernel.org/show_bug.cgi?id=12911 ...
I _think_
posix_cpu_timers_init_group()is not right, it should copycputime_expires->prof_exp.Peter, any chance you can test the (uncompiled/untested) patch below?
yes, with this patch 2.6.28.8 works as expected, thank you!
Great, thanks!
Quoted 3 lines
if you need to test any more patches on the subject i'm ready to do it. regression test system for our project depends on ability to limit cpu time, so it's major problem for us...
I am not sure what should we do, this needs more discussion.
Probably the most simple patch for -stable and 2.6.29 is below. (with this patch we don't even need
update_rlimit_cpu(), afaics).Oleg.
kernel/posix-cpu-timers.c static inline int fastpath_timer_check(s · line 1267
Quoted 6 lines
kernel/posix-cpu-timers.c +2
-1@@ -1263,7 +1263,8 @@ static inline int fastpath_timer_check(sif (task_cputime_expired (&group_sample, &sig->cputime_expires))return 1 ; } -return 0 ; + +return sig->rlim[RLIMIT_CPU ].rlim_cur !=RLIM_INFINITY ;
Once this version of the patch is confirmed to fix the bug, please send a full patch with a full changelog, SOB, Reported-by and Tested-by lines.
Thanks,
Ingo
PATCH 1/1 BUG 12911: fix RLIMIT_CPU && fork()
See http://bugzilla.kernel.org/show_bug.cgi?id=12911
copy_signal() copies signal->rlim, but RLIMIT_CPU is "lost". Because posix_cpu_timers_init_group() sets cputime_expires.prof_exp = 0 and thus fastpath_timer_check() returns false unless we have other cpu timers.
This is the minimal fix for 2.6.29 (tested) and 2.6.28. The patch is not optimal, we need further cleanups here. With this patch update_rlimit_cpu() is not really needed, but I don't think it should be removed.
The proper fix (I think) is:
- set_process_cpu_timer() should just start the cputimer->running logic (it does), no need to change cputime_expires.xxx_exp
- posix_cpu_timers_init_group() should set ->running when needed
- fastpath_timer_check() can check ->running instead of task_cputime_zero(signal->cputime_expires)
kernel/posix-cpu-timers.c +2 -1
@@ -1370,7 +1370,8 @@ static inline int fastpath_timer_check(sif (task_cputime_expired (&group_sample, &sig->cputime_expires))return 1 ; } -return 0 ; + +return sig->rlim[RLIMIT_CPU ].rlim_cur !=RLIM_INFINITY ; }/*
[tip:timers/urgent] posix timers: fix RLIMIT_CPU && fork()
Commit-ID: 37bebc70d7ad4144c571d74500db3bb26ec0c0eb Gitweb: http://git.kernel.org/tip/37bebc70d7ad4144c571d74500db3bb26ec0c0eb Author: Oleg Nesterov <[email protected]> AuthorDate: Mon, 23 Mar 2009 20:34:11 +0100 Committer: Ingo Molnar <[email protected]> CommitDate: Mon, 23 Mar 2009 20:43:35 +0100
posix timers: fix RLIMIT_CPU && fork()
See http://bugzilla.kernel.org/show_bug.cgi?id=12911
copy_signal() copies signal->rlim, but RLIMIT_CPU is "lost". Because posix_cpu_timers_init_group() sets cputime_expires.prof_exp = 0 and thus fastpath_timer_check() returns false unless we have other cpu timers.
This is the minimal fix for 2.6.29 (tested) and 2.6.28. The patch is not optimal, we need further cleanups here. With this patch update_rlimit_cpu() is not really needed, but I don't think it should be removed.
The proper fix (I think) is:
- set_process_cpu_timer() should just start the cputimer->running logic (it does), no need to change cputime_expires.xxx_exp
- posix_cpu_timers_init_group() should set ->running when needed
- fastpath_timer_check() can check ->running instead of task_cputime_zero(signal->cputime_expires)
LKML-Reference: <[email protected]>
--- kernel/posix-cpu-timers.c | 3 ++- 1 files changed, 2 insertions(+), 1 deletions(-)
kernel/posix-cpu-timers.c +2 -1
@@ -1370,7 +1370,8 @@ static inline int fastpath_timer_check(struct task_struct *tsk)if (task_cputime_expired (&group_sample, &sig->cputime_expires))return 1 ; } -return 0 ; + +return sig->rlim[RLIMIT_CPU ].rlim_cur !=RLIM_INFINITY ; }/*
Re: [tip:timers/urgent] posix timers: fix RLIMIT_CPU && fork()
Quoted 10 lines · Oleg Nesterov
Commit-ID: 37bebc70d7ad4144c571d74500db3bb26ec0c0eb Gitweb: http://git.kernel.org/tip/37bebc70d7ad4144c571d74500db3bb26ec0c0eb Author: Oleg Nesterov <[email protected]> AuthorDate: Mon, 23 Mar 2009 20:34:11 +0100 Committer: Ingo Molnar <[email protected]> CommitDate: Mon, 23 Mar 2009 20:43:35 +0100posix timers: fix
RLIMIT_CPU&&fork()
Quoted 8 lines
See http://bugzilla.kernel.org/show_bug.cgi?id=12911
copy_signal()copiessignal->rlim, butRLIMIT_CPUis "lost". Becauseposix_cpu_timers_init_group()sets cputime_expires.prof_exp = 0 and thusfastpath_timer_check()returns false unless we have other cpu timers.This is the minimal fix for 2.6.29 (tested) and 2.6.28. The patch is not optimal,
Ingo, please drop this patch, it is very suboptimal.
My intent was to make the obviously correct patch for 2.6.29, but since it was already released I'll send another one.
And,
Quoted 7 lines
we need further cleanups here. With this patch
update_rlimit_cpu()is not really needed, but I don't think it should be removed.The proper fix (I think) is:
-
set_process_cpu_timer()should just start thecputimer->runninglogic (it does), no need to change cputime_expires.xxx_exp
I am stupid, of course we should set cputime_expires.xxx_exp to avoid the slow path in run_posix_cpu_timers().
Oleg.
Re: [tip:timers/urgent] posix timers: fix RLIMIT_CPU && fork()
Quoted 13 lines · Oleg Nesterov
Quoted 12 lines · Oleg Nesterov
Commit-ID: 37bebc70d7ad4144c571d74500db3bb26ec0c0eb Gitweb: http://git.kernel.org/tip/37bebc70d7ad4144c571d74500db3bb26ec0c0eb Author: Oleg Nesterov <[email protected]> AuthorDate: Mon, 23 Mar 2009 20:34:11 +0100 Committer: Ingo Molnar <[email protected]> CommitDate: Mon, 23 Mar 2009 20:43:35 +0100posix timers: fix
RLIMIT_CPU&&fork()
Quoted 8 lines
Quoted 6 lines
copy_signal()copiessignal->rlim, butRLIMIT_CPUis "lost". Becauseposix_cpu_timers_init_group()sets cputime_expires.prof_exp = 0 and thusfastpath_timer_check()returns false unless we have other cpu timers.This is the minimal fix for 2.6.29 (tested) and 2.6.28. The patch is not optimal,
Ingo, please drop this patch, it is very suboptimal.
suboptimal why?
Quoted 2 lines
My intent was to make the obviously correct patch for 2.6.29, but since it was already released I'll send another one.
Please send it - thanks.
Ingo
Re: [tip:timers/urgent] posix timers: fix RLIMIT_CPU && fork()
Quoted 20 lines · Ingo Molnar
Quoted 17 lines · Oleg Nesterov
Quoted 16 lines · Oleg Nesterov
Commit-ID: 37bebc70d7ad4144c571d74500db3bb26ec0c0eb Gitweb: http://git.kernel.org/tip/37bebc70d7ad4144c571d74500db3bb26ec0c0eb Author: Oleg Nesterov <[email protected]> AuthorDate: Mon, 23 Mar 2009 20:34:11 +0100 Committer: Ingo Molnar <[email protected]> CommitDate: Mon, 23 Mar 2009 20:43:35 +0100posix timers: fix
RLIMIT_CPU&&fork()See http://bugzilla.kernel.org/show_bug.cgi?id=12911
copy_signal()copiessignal->rlim, butRLIMIT_CPUis "lost". Becauseposix_cpu_timers_init_group()sets cputime_expires.prof_exp = 0 and thusfastpath_timer_check()returns false unless we have other cpu timers.
Quoted 6 lines
Quoted 4 lines
Quoted 2 lines
This is the minimal fix for 2.6.29 (tested) and 2.6.28. The patch is not optimal,
Ingo, please drop this patch, it is very suboptimal.
suboptimal why?
Because this patch provokes the slow path on every tick if this process has rlim[RLIMIT_CPU].rlim_cur != RLIM_INFINITY, even if RLIMIT_CPU is not expired yet.
So I think the initial patch I sent (which modifies copy_signal) is better, but first I'd like to re-check the code once again.
Oleg.
Quoted 4 lines · Oleg Nesterov
I am not sure what should we do, this needs more discussion.
Probably the most simple patch for -stable and 2.6.29 is below. (with this patch we don't even need
update_rlimit_cpu(), afaics).
tested on both 2.6.28.9 and 2.6.29 - RLIMIT_CPU works as expected
Quoted 12 lines
kernel/posix-cpu-timers.c +2
-1@@ -1263,7 +1263,8 @@ static inline int fastpath_timer_check(sif (task_cputime_expired (&group_sample, &sig->cputime_expires))return 1 ; } -return 0 ; + +return sig->rlim[RLIMIT_CPU ].rlim_cur !=RLIM_INFINITY ; }/*