]> git.baikalelectronics.ru Git - kernel.git/commit
proc: Use a dedicated lock in struct pid
authorEric W. Biederman <ebiederm@xmission.com>
Tue, 7 Apr 2020 14:43:04 +0000 (09:43 -0500)
committerEric W. Biederman <ebiederm@xmission.com>
Thu, 9 Apr 2020 17:15:35 +0000 (12:15 -0500)
commit63f818f46af9f8b3f17b9695501e8d08959feb60
treea035279059164c2805f6e56d4c2d18259448d403
parentd1e7fd6462ca9fc76650fbe6ca800e35b24267da
proc: Use a dedicated lock in struct pid

syzbot wrote:
> ========================================================
> WARNING: possible irq lock inversion dependency detected
> 5.6.0-syzkaller #0 Not tainted
> --------------------------------------------------------
> swapper/1/0 just changed the state of lock:
ffffffff898090d8 (tasklist_lock){.+.?}-{2:2}, at: send_sigurg+0x9f/0x320 fs/fcntl.c:840
> but this lock took another, SOFTIRQ-unsafe lock in the past:
>  (&pid->wait_pidfd){+.+.}-{2:2}
>
>
> and interrupts could create inverse lock ordering between them.
>
>
> other info that might help us debug this:
>  Possible interrupt unsafe locking scenario:
>
>        CPU0                    CPU1
>        ----                    ----
>   lock(&pid->wait_pidfd);
>                                local_irq_disable();
>                                lock(tasklist_lock);
>                                lock(&pid->wait_pidfd);
>   <Interrupt>
>     lock(tasklist_lock);
>
>  *** DEADLOCK ***
>
> 4 locks held by swapper/1/0:

The problem is that because wait_pidfd.lock is taken under the tasklist
lock.  It must always be taken with irqs disabled as tasklist_lock can be
taken from interrupt context and if wait_pidfd.lock was already taken this
would create a lock order inversion.

Oleg suggested just disabling irqs where I have added extra calls to
wait_pidfd.lock.  That should be safe and I think the code will eventually
do that.  It was rightly pointed out by Christian that sharing the
wait_pidfd.lock was a premature optimization.

It is also true that my pre-merge window testing was insufficient.  So
remove the premature optimization and give struct pid a dedicated lock of
it's own for struct pid things.  I have verified that lockdep sees all 3
paths where we take the new pid->lock and lockdep does not complain.

It is my current day dream that one day pid->lock can be used to guard the
task lists as well and then the tasklist_lock won't need to be held to
deliver signals.  That will require taking pid->lock with irqs disabled.

Acked-by: Christian Brauner <christian.brauner@ubuntu.com>
Link: https://lore.kernel.org/lkml/00000000000011d66805a25cd73f@google.com/
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Christian Brauner <christian.brauner@ubuntu.com>
Reported-by: syzbot+343f75cdeea091340956@syzkaller.appspotmail.com
Reported-by: syzbot+832aabf700bc3ec920b9@syzkaller.appspotmail.com
Reported-by: syzbot+f675f964019f884dbd0f@syzkaller.appspotmail.com
Reported-by: syzbot+a9fb1457d720a55d6dc5@syzkaller.appspotmail.com
Fixes: 7bc3e6e55acf ("proc: Use a list of inodes to flush from proc")
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
fs/proc/base.c
include/linux/pid.h
kernel/pid.c