Re: [PATCH][V3]Make get_user_pages interruptible

["Pekka Enberg" <penberg@cs.helsinki.fi>]

Hi Paul,

On Fri, Nov 21, 2008 at 5:31 PM, Ying Han <yinghan@google.com> wrote:
>>                         */
>> -                       if (unlikely(test_tsk_thread_flag(tsk, TIF_MEMDIE)))
>> -                               return i ? i : -ENOMEM;
>> +                       if (unlikely(sigkill_pending(tsk)))
>> +                               return i ? i : -ERESTARTSYS;

On Mon, Nov 24, 2008 at 10:02 PM, Paul Menage <menage@google.com> wrote:
> You've changed the check from sigkill_pending(current) to sigkill_pending(tsk).
>
> I originally made that sigkill_pending(current) since we want to avoid
> tasks entering an unkillable state just because they're doing
> get_user_pages() on a system that's short of memory. Admittedly for
> the main case that we care about, mlock() (or an mmap() with
> MCL_FUTURE set) then tsk==current, but philosophically it seems to me
> to be more correct to do the check against current than tsk, since
> current is the thing that's actually allocating the memory. But maybe
> it would be better to check both?

Well, most callers seem to pass 'current' to get_user_pages() but for
the out-of-tree revoke patches, for example, you certainly want to
check sigkill_pending(current) as well; otherwise the revoke operation
is unkillable while in get_user_pages().

Not that revoke() is going to hit mainline any time soon but it does
serve as an argument for checking both.

                        Pekka