cputimers/proc: do_task_stat()->thread_group_times() is racy and O(n) under ->siglock
RFC PATCH 2/2 cputimers/proc: do_task_stat()->thread_group_times() is racy and O(n) under ->siglock
Nowadays ->siglock is overloaded, it would be really nice to change do_task_stat() to walk through the list of threads lockless. And note that we are doing while_each_thread() twice!
while_each_thread() is rcu-safe, but thread_group_times() also needs ->siglock to serialize the modifications of signal_struct->prev_Xtime members.
(however, please note that currently do_task_stat() can race with
wait_task_zombie() which calls thread_group_times() without siglock).
This patch changes the code back to use thread_group_cputime(), as we did before 0cf55e1e "sched, cputime: Introduce thread_group_times()".
Of course, this makes the output from /proc/pid/stat less accurate, but otoh this allows us to make do_task_stat() (apart from ->tty bits).
---
fs/proc/array.c | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-)
fs/proc/array.c renamed from fs/proc/array.c~PROC_5_DTS_TGTS_IS_PITA +5 -1
@@ -403,6 +403,7 @@ static int do_task_stat(struct seq_fileif (lock_task_sighand (task, &flags)) {struct signal_struct *sig = task->signal; +struct task_cputime cputime;if (sig->tty) {struct pid *pgrp =tty_get_pgrp (sig->tty);
@@ -433,8 +434,11 @@ static int do_task_stat(struct seq_file min_flt += sig->min_flt; maj_flt += sig->maj_flt; -thread_group_times (task, &utime, &stime); + gtime =cputime_add (gtime, sig->gtime); +thread_group_cputime (task, &cputime); + utime = cputime.utime; + stime = cputime.stime; } sid =task_session_nr_ns (task, ns);
Quoted 9 lines · Oleg Nesterov
Nowadays ->siglock is overloaded, it would be really nice to change
do_task_stat()to walk through the list of threads lockless. And note that we are doingwhile_each_thread()twice!
while_each_thread()is rcu-safe, butthread_group_times()also needs ->siglock to serialize the modifications ofsignal_struct->prev_Xtimemembers.
Right, so from what I remember the issue is that, yes top et al rely on that monotonicity, but more importantly I think clock_gettime(CLOCK_PROCESS_CPUTIME_ID) should indeed use ->siglock to ensure it serializes against do_exit() so that either we iterate the thread or get the accumulated runtime from signal_struct but not both (or neither).
Quoted 9 lines · Peter Zijlstra
Quoted 7 lines · Oleg Nesterov
Nowadays ->siglock is overloaded, it would be really nice to change
do_task_stat()to walk through the list of threads lockless. And note that we are doingwhile_each_thread()twice!
while_each_thread()is rcu-safe, butthread_group_times()also needs ->siglock to serialize the modifications ofsignal_struct->prev_Xtimemembers.
First of all, let me reply to myself. I see that I wasn't clear at all.
This patch does the first step to remove one reason for ->siglock (modification of ->prev_Xtime). But this is very minor, I guess we could change thread_group_times() to take signal->cputimer->lock.
The goal was to call thread_group_cputime() lockless under rcu lock (either directly, or via thread_group_times(), this doesn't matter) to avoid while_each_thread() under ->siglock.
And in this case /proc/pid/stat can't report utime/stime atomically. Whatever we do we can race with exit, so it doesn't make sense to play with ->prev_Xtime.
Quoted 2 lines
Right, so from what I remember the issue is that, yes top et al rely on that monotonicity,
Really? So, do you think the change above will break user-space?
How sad :/
Quoted 5 lines
but more importantly I think
clock_gettime(CLOCK_PROCESS_CPUTIME_ID)should indeed use ->siglock to ensure it serializes againstdo_exit()so that either we iterate the thread or get the accumulated runtime fromsignal_structbut not both (or neither).
Oh. I forgot everything I knew about posix-cpu-timers... But, it seems, posix_cpu_clock_get() calls thread_group_cputime() under tasklist and thus can't race with exit.
Oleg.
Quoted 6 lines · Oleg Nesterov
Quoted 2 lines
Right, so from what I remember the issue is that, yes top et al rely on that monotonicity,
Really? So, do you think the change above will break user-space?
How sad :/
IIRC top can give very funny results if you break it hard enough, it likes to give 9999% cputime if the thing goes backwards over the sample interval.
But I'm not sure your race is large enough to ever show up like that, so it might all just work out, but in general it does tend to require monotonic times.
Quoted 9 lines · Oleg Nesterov
Quoted 5 lines
but more importantly I think
clock_gettime(CLOCK_PROCESS_CPUTIME_ID)should indeed use ->siglock to ensure it serializes againstdo_exit()so that either we iterate the thread or get the accumulated runtime fromsignal_structbut not both (or neither).Oh. I forgot everything I knew about posix-cpu-timers... But, it seems,
posix_cpu_clock_get()callsthread_group_cputime()under tasklist and thus can't race with exit.
We assure thread_group_cputime() is called with one of: tasklist_lock or ->siglock to avoid races with __exit_signal. Except oom killer and elf core-dump code where is no lock, where we assume exit is not called or we don't care of inaccurate results.
Stanislaw