linux-kernel.vger.kernel.org archive mirror
 help / color / mirror / Atom feed
From: Oleg Nesterov <oleg@redhat.com>
To: Kirill Tkhai <ktkhai@parallels.com>
Cc: linux-kernel@vger.kernel.org,
	Peter Zijlstra <peterz@infradead.org>,
	Ingo Molnar <mingo@redhat.com>,
	Vladimir Davydov <vdavydov@parallels.com>,
	Kirill Tkhai <tkhai@yandex.ru>
Subject: Re: [PATCH RFC] sched: Revert delayed_put_task_struct() and fix use after free
Date: Wed, 15 Oct 2014 21:40:44 +0200	[thread overview]
Message-ID: <20141015194044.GA4557@redhat.com> (raw)
In-Reply-To: <20141015150641.GA2755@redhat.com>

On 10/15, Oleg Nesterov wrote:
>
> On 10/15, Kirill Tkhai wrote:
> >
> > Regarding to scheduler this may be a reason of use-after-free.
> >
> >     task_numa_compare()                    schedule()
> >         rcu_read_lock()                        ...
> >         cur = ACCESS_ONCE(dst_rq->curr)        ...
> >             ...                                rq->curr = next;
> >             ...                                    context_switch()
> >             ...                                        finish_task_switch()
> >             ...                                            put_task_struct()
> >             ...                                                __put_task_struct()
> >             ...                                                    free_task_struct()
> >             task_numa_assign()                                     ...
> >                 get_task_struct()                                  ...
>
> Agreed. I don't understand this code (will try to take another look later),
> but at first glance this looks wrong.
>
> At least the code like
>
> 	rcu_read_lock();
> 	get_task_struct(foreign_rq->curr);
> 	rcu_read_unlock();
>
> is certainly wrong. And _probably_ the problem should be fixed here. Perhaps
> we can add try_to_get_task_struct() which does atomic_inc_not_zero() ...

Yes, but perhaps in this particular case another simple fix makes more
sense. The patch below needs a comment to explain that we check PF_EXITING
because:

	1. It doesn't make sense to migrate the exiting task. Although perhaps
	   we could check ->mm == NULL instead.

	   But let me repeat that I do not understand this code, I am not sure
	   we can equally treat is_idle_task() and PF_EXITING here...

	2. If PF_EXITING is not set (or ->mm != NULL) then delayed_put_task_struct()
	   won't be called until we drop rcu_read_lock(), and thus get_task_struct()
	   is safe.

And. it seems that there is another problem? Can't task_h_load(cur) race
with itself if 2 CPU's call task_numa_migrate() and inspect the same rq
in parallel? Again, I don't understand this code, but update_cfs_rq_h_load()
doesn't look "atomic". In fact I am not even sure about task_h_load(env->p),
p == current but we do not disable preemption.

What do you think?

Oleg.

--- x/kernel/sched/fair.c
+++ x/kernel/sched/fair.c
@@ -1165,7 +1165,7 @@ static void task_numa_compare(struct tas
 
 	rcu_read_lock();
 	cur = ACCESS_ONCE(dst_rq->curr);
-	if (cur->pid == 0) /* idle */
+	if (is_idle_task(cur) || (curr->flags & PF_EXITING))
 		cur = NULL;
 
 	/*


  reply	other threads:[~2014-10-15 19:44 UTC|newest]

Thread overview: 32+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2014-10-15 12:31 [PATCH RFC] sched: Revert delayed_put_task_struct() and fix use after free Kirill Tkhai
2014-10-15 15:06 ` Oleg Nesterov
2014-10-15 19:40   ` Oleg Nesterov [this message]
2014-10-15 21:46     ` Kirill Tkhai
2014-10-15 22:02       ` Kirill Tkhai
2014-10-16  7:59       ` Peter Zijlstra
2014-10-16  8:16         ` Kirill Tkhai
2014-10-16  9:43           ` Peter Zijlstra
2014-10-16  9:50             ` Kirill Tkhai
2014-10-16  9:51               ` Kirill Tkhai
2014-10-16 10:04                 ` Kirill Tkhai
2014-10-17 21:34       ` Oleg Nesterov
2014-10-16  7:56     ` Peter Zijlstra
2014-10-16  8:01   ` Peter Zijlstra
2014-10-16 22:05     ` Oleg Nesterov
2014-10-17 21:36 ` [PATCH] sched/numa: fix unsafe get_task_struct() in task_numa_assign() Oleg Nesterov
2014-10-18  8:15   ` Kirill Tkhai
2014-10-18  8:33     ` Kirill Tkhai
2014-10-18 19:36       ` Peter Zijlstra
2014-10-18 21:18         ` Oleg Nesterov
2014-10-19  8:20         ` Kirill Tkhai
2014-10-18 20:56     ` Oleg Nesterov
2014-10-18 23:13       ` Kirill Tkhai
2014-10-19 19:24         ` Oleg Nesterov
2014-10-19 19:37           ` Oleg Nesterov
2014-10-19 19:43             ` Oleg Nesterov
2014-10-20  9:03               ` Kirill Tkhai
2014-10-20  9:13             ` Peter Zijlstra
2014-10-20 10:36               ` Kirill Tkhai
2014-10-20  9:00           ` Kirill Tkhai
2014-10-19 21:38         ` Peter Zijlstra
2014-10-20  8:56           ` Kirill Tkhai

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=20141015194044.GA4557@redhat.com \
    --to=oleg@redhat.com \
    --cc=ktkhai@parallels.com \
    --cc=linux-kernel@vger.kernel.org \
    --cc=mingo@redhat.com \
    --cc=peterz@infradead.org \
    --cc=tkhai@yandex.ru \
    --cc=vdavydov@parallels.com \
    /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).