Re: [PATCH] coredump: fix unfreezable coredumping task

[Michal Hocko]

On Tue 04-10-16 18:13:05, Oleg Nesterov wrote:
> On 10/04, Michal Hocko wrote:
> >
> > On Fri 30-09-16 14:47:41, Oleg Nesterov wrote:
> > > On 09/30, Andrey Ryabinin wrote:
> > > >
> > > > @@ -423,7 +424,9 @@ static int coredump_wait(int exit_code, struct core_state *core_state)
> > > >  	if (core_waiters > 0) {
> > > >  		struct core_thread *ptr;
> > > >
> > > > +		freezer_do_not_count();
> > > >  		wait_for_completion(&core_state->startup);
> > > > +		freezer_count();
> > >
> > > Agreed... we could probably even do
> > >
> > > 	--- x/fs/coredump.c
> > > 	+++ x/fs/coredump.c
> > > 	@@ -423,7 +423,13 @@ static int coredump_wait(int exit_code, 
> > > 		if (core_waiters > 0) {
> > > 			struct core_thread *ptr;
> > > 	 
> > > 	-		wait_for_completion(&core_state->startup);
> > > 	+		if (wait_for_completion_interruptible(&core_state->startup)) {
> > > 	+			/* see the comment in dump_interrupted() */
> > > 	+			down_write(&mm->mmap_sem);
> > > 	+			coredump_finish(mm, false);
> > > 	+			up_write(&mm->mmap_sem);
> > > 	+			return -EINTR;
> > > 	+		}
> > > 			/*
> > > 			 * Wait for all the threads to become inactive, so that
> > > 			 * all the thread context (extended register state, like
> >
> > This looks like a very good idea to me. We really want to make the whole
> > coredump_wait killable.
> 
> Well, it is already killable. 

Except wait_for_completion is not killable and the exiting tasks might
be blocked in a !killable state blocking this one to continue. But...

> And with the change above it can sleep
> in down_write(mmap_sem) and we really need this lock to abort, so it
> won't necessarily react to SIGKILL faster.

you are right that somebody might be holding mmap_sem and we cannot get
rid of it here.

> > I guess this should help us to remove the
> > hackish sig->flags & SIGNAL_GROUP_COREDUMP check from
> > __task_will_free_mem.
> 
> Why? This doesn't depend on "killable". __task_will_free_mem() checks
> this flag to detect the CLONE_VM processes which won't exit soon because
> they participate in the coredumping.

I just (wrongly) assumed that if we make this path killable completely
we can guarantee a forward progress and get rid of SIGNAL_GROUP_COREDUMP
check completely. But you are right this won't be sufficient.
-- 
Michal Hocko
SUSE Labs