Re: [2.6.33-rc5] starting emacs makes lockdep warning

[AmÃrico Wang]

On Tue, Jan 26, 2010 at 5:14 PM, Eric W. Biederman
<ebiederm@xxxxxxxxxxxx> wrote:
> AmÃrico Wang <xiyou.wangcong@xxxxxxxxx> writes:
>
>> On Tue, Jan 26, 2010 at 3:45 PM, KOSAKI Motohiro
>> <kosaki.motohiro@xxxxxxxxxxxxxx> wrote:
>>> Hi
>>>
>>>> On Tue, Jan 26, 2010 at 02:01:12PM +0800, Am??rico Wang wrote:
>>>>
>>>> > I agree, it seems that patch is useless, since we already
>>>> > do lock_kernel() before calling __f_setown()...
>>>>
>>>> What's to prevent pid from being freed under us? ÂBKL won't...
>>>
>>> I don't understand this issue at all. so, this is stupid dumb question.
>>> Why can't we write following code?
>>>
>>>
>>> Â Â Â Â Â Â Â Âenum pid_type type;
>>> Â Â Â Â Â Â Â Âstruct pid *pid;
>>> Â Â Â Â Â Â Â Âif (!waitqueue_active(&tty->read_wait))
>>> Â Â Â Â Â Â Â Â Â Â Â Âtty->minimum_to_wake = 1;
>>> Â Â Â Â Â Â Â Âspin_lock_irqsave(&tty->ctrl_lock, flags);
>>> Â Â Â Â Â Â Â Âif (tty->pgrp) {
>>> Â Â Â Â Â Â Â Â Â Â Â Âpid = tty->pgrp;
>>> Â Â Â Â Â Â Â Â Â Â Â Âtype = PIDTYPE_PGID;
>>> Â Â Â Â Â Â Â Â} else {
>>> Â Â Â Â Â Â Â Â Â Â Â Âpid = task_pid(current);
>>> Â Â Â Â Â Â Â Â Â Â Â Âtype = PIDTYPE_PID;
>>> Â Â Â Â Â Â Â Â}
>>> Â Â Â Â Â Â Â Âget_pid(pid) Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â// insert here
>>> Â Â Â Â Â Â Â Âspin_unlock_irqrestore(&tty->ctrl_lock, flags);
>>> Â Â Â Â Â Â Â Âretval = __f_setown(filp, pid, type, 0);
>>> Â Â Â Â Â Â Â Âput_pid(pid) Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â// insert here
>>>
>>
>> Yeah, this seems reasonable for me, but not sure if this is the best fix.
>
> That or tweak __f_setown to use irqsave/irqrestore variants for it's
> locks, __f_setown is already atomic. ÂI prefer that direction because the
> code is just a little simpler.
>

Oh, very good advice!

Patch is below.

-------------->
Commit 703625118 causes a lockdep warning:

[ INFO: possible irq lock inversion dependency detected ]
2.6.33-rc5 #77
---------------------------------------------------------
emacs/1609 just changed the state of lock:
 (&(&tty->ctrl_lock)->rlock){+.....}, at: [<ffffffff8127c648>]
tty_fasync+0xe8/0x190
but this lock took another, HARDIRQ-unsafe lock in the past:
 (&(&sighand->siglock)->rlock){-.....}

This is due to we use write_lock_irq() in __f_setown() which turns
the IRQ on in write_unlock_irq(), causes this warning.

Switch it ot write_lock_irqsave() and write_unlock_irqrestore(),
as suggested by Eric.

Reported-by: KOSAKI Motohiro <kosaki.motohiro@xxxxxxxxxxxxxx>
Signed-off-by: WANG Cong <xiyou.wangcong@xxxxxxxxx>

----
diff --git a/fs/fcntl.c b/fs/fcntl.c
index 97e01dc..556b404 100644
--- a/fs/fcntl.c
+++ b/fs/fcntl.c
@@ -199,7 +199,8 @@ static int setfl(int fd, struct file * filp, unsigned long arg)
 static void f_modown(struct file *filp, struct pid *pid, enum pid_type type,
                      int force)
 {
-	write_lock_irq(&filp->f_owner.lock);
+	int flags;
+	write_lock_irqsave(&filp->f_owner.lock, flags);
 	if (force || !filp->f_owner.pid) {
 		put_pid(filp->f_owner.pid);
 		filp->f_owner.pid = get_pid(pid);
@@ -211,7 +212,7 @@ static void f_modown(struct file *filp, struct pid *pid, enum pid_type type,
 			filp->f_owner.euid = cred->euid;
 		}
 	}
-	write_unlock_irq(&filp->f_owner.lock);
+	write_unlock_irqrestore(&filp->f_owner.lock, flags);
 }
 
 int __f_setown(struct file *filp, struct pid *pid, enum pid_type type,