2007-08-12 17:03:23

by Oleg Nesterov

Subject: [PATCH 2/4] posix-timers: fix creation race

sys_timer_create() sets ->it_process and unlocks ->siglock, then checks
tmr->it_sigev_notify to define if get_task_struct() is needed.

We already passed ->it_id to the caller, another thread can delete this
timer and free its memory in between.

As a minimal fix, move this code under ->siglock, sys_timer_delete() takes
it too before calling release_posix_timer(). A proper serialization would
be to take ->it_lock, we add a partly initialized timer on posix_timers_id,
not good.

Signed-off-by: Oleg Nesterov <[email protected]>

--- t/kernel/posix-timers.c~2_CREATE 2007-08-12 17:59:17.000000000 +0400
+++ t/kernel/posix-timers.c 2007-08-12 18:11:33.000000000 +0400
@@ -547,13 +547,12 @@ sys_timer_create(const clockid_t which_c
new_timer->it_process = process;
list_add(&new_timer->list,
&process->signal->posix_timers);
- spin_unlock_irqrestore(&process->sighand->siglock, flags);
if (new_timer->it_sigev_notify == (SIGEV_SIGNAL|SIGEV_THREAD_ID))
get_task_struct(process);
} else {
- spin_unlock_irqrestore(&process->sighand->siglock, flags);
process = NULL;
}
+ spin_unlock_irqrestore(&process->sighand->siglock, flags);
}
read_unlock(&tasklist_lock);
if (!process) {

2007-08-12 19:06:32

by Thomas Gleixner

Subject: Re: [PATCH 2/4] posix-timers: fix creation race

On Sun, 2007-08-12 at 21:05 +0400, Oleg Nesterov wrote:
> sys_timer_create() sets ->it_process and unlocks ->siglock, then checks
> tmr->it_sigev_notify to define if get_task_struct() is needed.
>
> We already passed ->it_id to the caller, another thread can delete this
> timer and free its memory in between.
>
> As a minimal fix, move this code under ->siglock, sys_timer_delete() takes
> it too before calling release_posix_timer(). A proper serialization would
> be to take ->it_lock, we add a partly initialized timer on posix_timers_id,
> not good.
>
> Signed-off-by: Oleg Nesterov <[email protected]>

Acked-by: Thomas Gleixner <[email protected]>

> --- t/kernel/posix-timers.c~2_CREATE 2007-08-12 17:59:17.000000000 +0400
> +++ t/kernel/posix-timers.c 2007-08-12 18:11:33.000000000 +0400
> @@ -547,13 +547,12 @@ sys_timer_create(const clockid_t which_c
> new_timer->it_process = process;
> list_add(&new_timer->list,
> &process->signal->posix_timers);
> - spin_unlock_irqrestore(&process->sighand->siglock, flags);
> if (new_timer->it_sigev_notify == (SIGEV_SIGNAL|SIGEV_THREAD_ID))
> get_task_struct(process);
> } else {
> - spin_unlock_irqrestore(&process->sighand->siglock, flags);
> process = NULL;
> }
> + spin_unlock_irqrestore(&process->sighand->siglock, flags);
> }
> read_unlock(&tasklist_lock);
> if (!process) {
>

2007-08-12 19:44:50

by Oleg Nesterov

Subject: Re: [PATCH 2/4] posix-timers: fix creation race

On 08/12, Thomas Gleixner wrote:
>
> On Sun, 2007-08-12 at 21:05 +0400, Oleg Nesterov wrote:
> > sys_timer_create() sets ->it_process and unlocks ->siglock, then checks
> > tmr->it_sigev_notify to define if get_task_struct() is needed.
> >
> > We already passed ->it_id to the caller, another thread can delete this
> > timer and free its memory in between.
> >
> > As a minimal fix, move this code under ->siglock, sys_timer_delete() takes
> > it too before calling release_posix_timer(). A proper serialization would
> > be to take ->it_lock, we add a partly initialized timer on posix_timers_id,
> > not good.
> >
> > Signed-off-by: Oleg Nesterov <[email protected]>
>
> Acked-by: Thomas Gleixner <[email protected]>

Oooh, thanks, but...

> > --- t/kernel/posix-timers.c~2_CREATE 2007-08-12 17:59:17.000000000 +0400
> > +++ t/kernel/posix-timers.c 2007-08-12 18:11:33.000000000 +0400
> > @@ -547,13 +547,12 @@ sys_timer_create(const clockid_t which_c
> > new_timer->it_process = process;
> > list_add(&new_timer->list,
> > &process->signal->posix_timers);
> > - spin_unlock_irqrestore(&process->sighand->siglock, flags);
> > if (new_timer->it_sigev_notify == (SIGEV_SIGNAL|SIGEV_THREAD_ID))
> > get_task_struct(process);
> > } else {
> > - spin_unlock_irqrestore(&process->sighand->siglock, flags);
> > process = NULL;
> > }
> > + spin_unlock_irqrestore(&process->sighand->siglock, flags);

i'll resend this patch tomorrow. We can't do spin_unlock_irqrestore(&process->...)
if we set process = NULL above.

This all needs a cleanup anyway. The PF_EXITING check and related comment are
bogus.

Other patches do not depend on this one.

Oleg.

2007-08-12 19:59:25

by Oleg Nesterov

Subject: Re: [PATCH 2/4] posix-timers: fix creation race

On 08/12, Oleg Nesterov wrote:
>
> > > - spin_unlock_irqrestore(&process->sighand->siglock, flags);
> > > process = NULL;
> > > }
> > > + spin_unlock_irqrestore(&process->sighand->siglock, flags);
>
> i'll resend this patch tomorrow. We can't do spin_unlock_irqrestore(&process->...)
> if we set process = NULL above.

Another attempt, no cleanups, just a minimal fix for now.


[PATCH 2/4] posix-timers: fix creation race

sys_timer_create() sets ->it_process and unlocks ->siglock, then checks
tmr->it_sigev_notify to define if get_task_struct() is needed.

We already passed ->it_id to the caller, another thread can delete this
timer and free its memory in between.

As a minimal fix, move this code under ->siglock, sys_timer_delete() takes
it too before calling release_posix_timer(). A proper serialization would
be to take ->it_lock, we add a partly initialized timer on posix_timers_id,
not good.

Signed-off-by: Oleg Nesterov <[email protected]>

--- t/kernel/posix-timers.c~ 2007-08-12 23:50:46.000000000 +0400
+++ t/kernel/posix-timers.c 2007-08-12 23:52:07.000000000 +0400
@@ -547,9 +547,9 @@ sys_timer_create(const clockid_t which_c
new_timer->it_process = process;
list_add(&new_timer->list,
&process->signal->posix_timers);
- spin_unlock_irqrestore(&process->sighand->siglock, flags);
if (new_timer->it_sigev_notify == (SIGEV_SIGNAL|SIGEV_THREAD_ID))
get_task_struct(process);
+ spin_unlock_irqrestore(&process->sighand->siglock, flags);
} else {
spin_unlock_irqrestore(&process->sighand->siglock, flags);
process = NULL;

2007-08-13 12:50:24

by Thomas Gleixner

Subject: Re: [PATCH 2/4] posix-timers: fix creation race

On Sun, 2007-08-12 at 23:44 +0400, Oleg Nesterov wrote:
> > > --- t/kernel/posix-timers.c~2_CREATE 2007-08-12 17:59:17.000000000 +0400
> > > +++ t/kernel/posix-timers.c 2007-08-12 18:11:33.000000000 +0400
> > > @@ -547,13 +547,12 @@ sys_timer_create(const clockid_t which_c
> > > new_timer->it_process = process;
> > > list_add(&new_timer->list,
> > > &process->signal->posix_timers);
> > > - spin_unlock_irqrestore(&process->sighand->siglock, flags);
> > > if (new_timer->it_sigev_notify == (SIGEV_SIGNAL|SIGEV_THREAD_ID))
> > > get_task_struct(process);
> > > } else {
> > > - spin_unlock_irqrestore(&process->sighand->siglock, flags);
> > > process = NULL;
> > > }
> > > + spin_unlock_irqrestore(&process->sighand->siglock, flags);
>
> i'll resend this patch tomorrow. We can't do spin_unlock_irqrestore(&process->...)
> if we set process = NULL above.

Doh, half of my brain seems to be still on vacation.

tglx