Ying Han wrote:
>
> -static int sigkill_pending(struct task_struct *tsk)
> +int sigkill_pending(struct task_struct *tsk)
> {
> return sigismember(&tsk->pending.signal, SIGKILL) ||
> sigismember(&tsk->signal->shared_pending.signal, SIGKILL);
> diff -puN mm/memory.c~make-get_user_pages-interruptible mm/memory.c
> --- a/mm/memory.c~make-get_user_pages-interruptible
> +++ a/mm/memory.c
> @@ -1219,12 +1219,12 @@ int __get_user_pages(struct task_struct
> struct page *page;
>
> /*
> - * If tsk is ooming, cut off its access to large memory
> - * allocations. It has a pending SIGKILL, but it can't
> - * be processed until returning to user space.
> + * If we have a pending SIGKILL, don't keep
> + * allocating memory.
> */
> - if (unlikely(test_tsk_thread_flag(tsk, TIF_MEMDIE)))
> - return i ? i : -ENOMEM;
> + if (unlikely(sigkill_pending(current) ||
> + sigkill_pending(tsk)))
Please do not export/use sigkill_pending(). It is "private" for ptrace_stop()
(and actually should die imho).
We have fatal_signal_pending() for that.
Oleg.
Thanks Oleg, i looked at the code again and this is a reasonable
change. I will make the change on the patch.
Andrew,
should i make a patch based on current change or make the same
patch as [V6]?
thanks
--Ying
On Thu, Nov 27, 2008 at 12:04 PM, Oleg Nesterov <[email protected]> wrote:
> Ying Han wrote:
>>
>> -static int sigkill_pending(struct task_struct *tsk)
>> +int sigkill_pending(struct task_struct *tsk)
>> {
>> return sigismember(&tsk->pending.signal, SIGKILL) ||
>> sigismember(&tsk->signal->shared_pending.signal, SIGKILL);
>> diff -puN mm/memory.c~make-get_user_pages-interruptible mm/memory.c
>> --- a/mm/memory.c~make-get_user_pages-interruptible
>> +++ a/mm/memory.c
>> @@ -1219,12 +1219,12 @@ int __get_user_pages(struct task_struct
>> struct page *page;
>>
>> /*
>> - * If tsk is ooming, cut off its access to large memory
>> - * allocations. It has a pending SIGKILL, but it can't
>> - * be processed until returning to user space.
>> + * If we have a pending SIGKILL, don't keep
>> + * allocating memory.
>> */
>> - if (unlikely(test_tsk_thread_flag(tsk, TIF_MEMDIE)))
>> - return i ? i : -ENOMEM;
>> + if (unlikely(sigkill_pending(current) ||
>> + sigkill_pending(tsk)))
>
> Please do not export/use sigkill_pending(). It is "private" for ptrace_stop()
> (and actually should die imho).
>
> We have fatal_signal_pending() for that.
>
> Oleg.
>
>
On Mon, 1 Dec 2008 15:00:14 -0800
Ying Han <[email protected]> wrote:
> On Thu, Nov 27, 2008 at 12:04 PM, Oleg Nesterov <[email protected]> wrote:
> > Ying Han wrote:
> >>
> >> -static int sigkill_pending(struct task_struct *tsk)
> >> +int sigkill_pending(struct task_struct *tsk)
> >> {
> >> return sigismember(&tsk->pending.signal, SIGKILL) ||
> >> sigismember(&tsk->signal->shared_pending.signal, SIGKILL);
> >> diff -puN mm/memory.c~make-get_user_pages-interruptible mm/memory.c
> >> --- a/mm/memory.c~make-get_user_pages-interruptible
> >> +++ a/mm/memory.c
> >> @@ -1219,12 +1219,12 @@ int __get_user_pages(struct task_struct
> >> struct page *page;
> >>
> >> /*
> >> - * If tsk is ooming, cut off its access to large memory
> >> - * allocations. It has a pending SIGKILL, but it can't
> >> - * be processed until returning to user space.
> >> + * If we have a pending SIGKILL, don't keep
> >> + * allocating memory.
> >> */
> >> - if (unlikely(test_tsk_thread_flag(tsk, TIF_MEMDIE)))
> >> - return i ? i : -ENOMEM;
> >> + if (unlikely(sigkill_pending(current) ||
> >> + sigkill_pending(tsk)))
> >
> > Please do not export/use sigkill_pending(). It is "private" for ptrace_stop()
> > (and actually should die imho).
> >
> > We have fatal_signal_pending() for that.
> >
> > Oleg.
> >
(top-posting repaired..)
> Thanks Oleg, i looked at the code again and this is a reasonable
> change. I will make the change on the patch.
> Andrew,
> should i make a patch based on current change or make the same
> patch as [V6]?
I don't mind either way. I guess a fresh new patch would simplify
review for everyone.
Note that fatal_signal_pending() is not an exact replacement for
sigkill_pending() - fatal_signal_pending() doesn't test the shared
pending signals.
thanks Andrew.
On Mon, Dec 1, 2008 at 3:09 PM, Andrew Morton <[email protected]> wrote:
> On Mon, 1 Dec 2008 15:00:14 -0800
> Ying Han <[email protected]> wrote:
>
>> On Thu, Nov 27, 2008 at 12:04 PM, Oleg Nesterov <[email protected]> wrote:
>> > Ying Han wrote:
>> >>
>> >> -static int sigkill_pending(struct task_struct *tsk)
>> >> +int sigkill_pending(struct task_struct *tsk)
>> >> {
>> >> return sigismember(&tsk->pending.signal, SIGKILL) ||
>> >> sigismember(&tsk->signal->shared_pending.signal, SIGKILL);
>> >> diff -puN mm/memory.c~make-get_user_pages-interruptible mm/memory.c
>> >> --- a/mm/memory.c~make-get_user_pages-interruptible
>> >> +++ a/mm/memory.c
>> >> @@ -1219,12 +1219,12 @@ int __get_user_pages(struct task_struct
>> >> struct page *page;
>> >>
>> >> /*
>> >> - * If tsk is ooming, cut off its access to large memory
>> >> - * allocations. It has a pending SIGKILL, but it can't
>> >> - * be processed until returning to user space.
>> >> + * If we have a pending SIGKILL, don't keep
>> >> + * allocating memory.
>> >> */
>> >> - if (unlikely(test_tsk_thread_flag(tsk, TIF_MEMDIE)))
>> >> - return i ? i : -ENOMEM;
>> >> + if (unlikely(sigkill_pending(current) ||
>> >> + sigkill_pending(tsk)))
>> >
>> > Please do not export/use sigkill_pending(). It is "private" for ptrace_stop()
>> > (and actually should die imho).
>> >
>> > We have fatal_signal_pending() for that.
>> >
>> > Oleg.
>> >
>
> (top-posting repaired..)
>
>> Thanks Oleg, i looked at the code again and this is a reasonable
>> change. I will make the change on the patch.
>> Andrew,
>> should i make a patch based on current change or make the same
>> patch as [V6]?
>
> I don't mind either way. I guess a fresh new patch would simplify
> review for everyone.
>
> Note that fatal_signal_pending() is not an exact replacement for
> sigkill_pending() - fatal_signal_pending() doesn't test the shared
> pending signals.
>
yeah, i noticed that and that is why i used the sigkill_pending
initially. But after a deeper look
at the signal code, i found for SIGKILL ( signal we care about in this
patch ), tsk->pending is a
superset of tsk->signal->shared_pending.
--Ying
>
>
On 12/01, Ying Han wrote:
>
> On Mon, Dec 1, 2008 at 3:09 PM, Andrew Morton <[email protected]> wrote:
> >
> > Note that fatal_signal_pending() is not an exact replacement for
> > sigkill_pending() - fatal_signal_pending() doesn't test the shared
> > pending signals.
> >
> yeah, i noticed that and that is why i used the sigkill_pending
> initially. But after a deeper look
> at the signal code, i found for SIGKILL ( signal we care about in this
> patch ), tsk->pending is a
> superset of tsk->signal->shared_pending.
Yes. If we have SIGKILL in ->shared_pending, we must have it in
in ->pending too.
The only exception is when we send SIGKILL to the already dying task
which has already dequeued SIGKILL from ->pending. And we do have a
special case: coredumping. elf_core_dump() does get_user_pages()
when the process is "almost" dead, it has SIGNAL_GROUP_EXIT.
So, if get_user_pages() calls sigkill_pending() instead of
fatal_signal_pending(), this can help. But:
- this relies on the fact that we always queue SIGKILL
even if the task is dead. This behaviour can be changed.
- this doesn't stop the coredumping, it will continue
with DUMP_SEEK(). Yes, this is better if we want to
stop this thread to populate the memory, but I think
we should just fix the coredumping - it should be
interruptible.
Actually, the patch exists:
http://marc.info/?l=linux-kernel&m=121665710711931
Oleg.