linux-kernel.vger.kernel.org archive mirror
 help / color / mirror / Atom feed
From: Oleg Nesterov <oleg@redhat.com>
To: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
Cc: Andrew Morton <akpm@linux-foundation.org>,
	Don Zickus <dzickus@redhat.com>,
	Frederic Weisbecker <fweisbec@gmail.com>,
	Ingo Molnar <mingo@elte.hu>,
	Jerome Marchand <jmarchan@redhat.com>,
	Mandeep Singh Baines <msb@google.com>,
	Roland McGrath <roland@redhat.com>,
	linux-kernel@vger.kernel.org, stable@kernel.org,
	"Eric W. Biederman" <ebiederm@xmission.com>
Subject: Re: while_each_thread() under rcu_read_lock() is broken?
Date: Wed, 23 Jun 2010 17:24:21 +0200	[thread overview]
Message-ID: <20100623152421.GA8445@redhat.com> (raw)
In-Reply-To: <20100622221226.GP2290@linux.vnet.ibm.com>

On 06/22, Paul E. McKenney wrote:
>
> On Tue, Jun 22, 2010 at 11:23:57PM +0200, Oleg Nesterov wrote:
> > On 06/21, Paul E. McKenney wrote:
> > >
> > > Yeah, would be good to avoid this.  Not sure it can be avoided, though.
> >
> > Why? I think next_thread_careful() from
> > 	http://marc.info/?l=linux-kernel&m=127714242731448
> > should work.
> >
> > If the caller holds tasklist or siglock, this change has no effect.
> >
> > If the caller does while_each_thread() under rcu_read_lock(), then
> > it is OK to break the loop earlier than we do now. The lockless
> > while_each_thread() works in a "best effort" manner anyway, if it
> > races with exit_group() or exec() it can miss some/most/all sub-threads
> > (including the new leader) with or without this change.
> >
> > Yes, zap_threads() needs additional fixes. But I think it is better
> > to complicate a couple of lockless callers (or just change them
> > to take tasklist) which must not miss an "interesting" thread.
>
> Is it the case that creating a new group leader from an existing group
> always destroys the old group?  It certainly is the case for exec().

Yes. And only exec() can change the leader.

> Anyway, if creating a new thread group implies destroying the old one,
> and if the thread group leader cannot be concurrently creating a new
> thread group and traversing the old one, then yes, I do believe your
> code at http://marc.info/?l=linux-kernel&m=127714242731448 will work.
>
> Assuming that the call to next_thread_careful(t) in the definition of
> while_each_thread() is replaced with next_thread_careful(g,t).

Great.

> And give or take memory barriers.
>
> The implied memory barrier mentioned in the comment in your example code
> is the spin_lock_irqsave() and spin_unlock_irqrestore()

Argh. No, no, no. I meant unlock + lock, not lock + unlock. But when
I look at __change_pid() now I do not see the 2nd lock() after
unlock(pidmap_lock). I was wrong, thanks.

And, even if I was correct it is not nice to rely on the internals
of detach_pid(). If we use next_thread_careful(), __unhash_process()
needs wmb.

Thanks!

> > And. Whatever we do with de_thread(), this can't fix the lockless
> > while_each_thread(not_a_group_leader, t). I do not know if there is
> > any user which does this though.
> > fastpath_timer_check()->thread_group_cputimer() does this, but this
> > is wrong and we already have the patch which removes it.
>
> Indeed.  Suppose that the starting task is the one immediately preceding
> the task group leader.  You get a pointer to the task in question
> and traverse to the next task (the task group leader), during which
> time the thread group leader does exec() and maybe a pthread_create()
> or two.  Oops!  You are now now traversing the wrong thread group!

Yes, but there is another more simple and much more feasible race.
while_each_thread(non_leader, t) will loop forever if non_leader
simply exits and passes __unhash_process(). After that next_thread(t)
can never return non_leader.

> There are ways of fixing this, but all the ones I know of require more
> fields in the task structure,

Just in case, I hope that next_thread_careful() handles this case too.

> so best if we don't need to start somewhere
> other than a group leader.

(I assume, you mean the lockless case)

I am not sure. Because it is not trivial to enforce this rule even if
we add a fat comment. Please look at check_hung_uninterruptible_tasks().
It does do_each_thread/while_each_thread and thus it always starts at
the leader. But in fact this is not true due to rcu_lock_break().

Or proc_task_readdir(). It finds the leader, does get_task_struct(leader)
and drops rcu lock. After that first_tid() takes rcu_read_lock() again
and does the open coded while_each_thread(). At first glance this case
looks fine, "for (pos = leader; nr > 0; --nr)" can't loop forever.
But in fact it is not:

	- proc_task_readdir() finds the leader L, does get_task_struct()
	  and drops rcu lock.

	  Suppose that filp->f_pos >= 2.
          Suppose also that the previous tid cached in filp->f_version
          is no longer valid.

	  The caller is preempted.

	- a sub-thread T does exec, and does release_task(L).

	  But L->thread_group->next == T, so far everything is good

	- T spawns other sub-threads (so that get_nr_threads() > nr)
	  and exits.

	- grace period passes, T's task_struct is freed/unmapped/reused

	- proc_task_readdir() resumes and calls first_tid(L).

	  next_thread(L) returns T == nowhere

It is very possible that I missed something here, my only point is
that I think it would be safer to assume nothing about the leaderness.

Oleg.


  reply	other threads:[~2010-06-23 15:26 UTC|newest]

Thread overview: 45+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2010-06-18 19:02 [PATCH] fix the racy check_hung_uninterruptible_tasks()->rcu_lock_break() logic Oleg Nesterov
2010-06-18 19:34 ` while_each_thread() under rcu_read_lock() is broken? Oleg Nesterov
2010-06-18 21:08   ` Roland McGrath
2010-06-18 22:37     ` Oleg Nesterov
2010-06-18 22:33   ` Paul E. McKenney
2010-06-21 17:09     ` Oleg Nesterov
2010-06-21 17:44       ` Oleg Nesterov
2010-06-21 18:00         ` Oleg Nesterov
2010-06-21 19:02         ` Roland McGrath
2010-06-21 20:06           ` Oleg Nesterov
2010-06-21 21:19             ` Eric W. Biederman
2010-06-22 14:34               ` Oleg Nesterov
2010-07-08 23:59             ` Roland McGrath
2010-07-09  0:41               ` Paul E. McKenney
2010-07-09  1:01                 ` Roland McGrath
2010-07-09 16:18                   ` Paul E. McKenney
2010-06-21 20:51       ` Paul E. McKenney
2010-06-21 21:22         ` Eric W. Biederman
2010-06-21 21:38           ` Paul E. McKenney
2010-06-22 21:23         ` Oleg Nesterov
2010-06-22 22:12           ` Paul E. McKenney
2010-06-23 15:24             ` Oleg Nesterov [this message]
2010-06-24 18:07               ` Paul E. McKenney
2010-06-24 18:50                 ` Chris Friesen
2010-06-24 22:00                   ` Oleg Nesterov
2010-06-25  0:08                     ` Eric W. Biederman
2010-06-25  3:42                       ` Paul E. McKenney
2010-06-25 10:08                       ` Oleg Nesterov
2010-07-09  0:52                       ` Roland McGrath
2010-06-24 21:14                 ` Roland McGrath
2010-06-25  3:37                   ` Paul E. McKenney
2010-07-09  0:41                     ` Roland McGrath
2010-06-24 21:57                 ` Oleg Nesterov
2010-06-25  3:41                   ` Paul E. McKenney
2010-06-25  9:55                     ` Oleg Nesterov
2010-06-28 23:43                       ` Paul E. McKenney
2010-06-29 13:05                         ` Oleg Nesterov
2010-06-29 15:34                           ` Paul E. McKenney
2010-06-29 17:54                             ` Oleg Nesterov
2010-06-19  5:00   ` Mandeep Baines
2010-06-19  5:35     ` Frederic Weisbecker
2010-06-19 15:44       ` Mandeep Baines
2010-06-19 19:19     ` Oleg Nesterov
2010-06-18 20:11 ` [PATCH] fix the racy check_hung_uninterruptible_tasks()->rcu_lock_break() logic Frederic Weisbecker
2010-06-18 20:38 ` Mandeep Singh Baines

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20100623152421.GA8445@redhat.com \
    --to=oleg@redhat.com \
    --cc=akpm@linux-foundation.org \
    --cc=dzickus@redhat.com \
    --cc=ebiederm@xmission.com \
    --cc=fweisbec@gmail.com \
    --cc=jmarchan@redhat.com \
    --cc=linux-kernel@vger.kernel.org \
    --cc=mingo@elte.hu \
    --cc=msb@google.com \
    --cc=paulmck@linux.vnet.ibm.com \
    --cc=roland@redhat.com \
    --cc=stable@kernel.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).