[PATCH 1/2] fcntl: fix potential deadlocks for &fown_struct.lock

Jeff Layton jlayton at kernel.org
Fri Jul 2 11:44:48 UTC 2021


On Fri, 2021-07-02 at 17:18 +0800, Desmond Cheong Zhi Xi wrote:
> Syzbot reports a potential deadlock in do_fcntl:
> 
> ========================================================
> WARNING: possible irq lock inversion dependency detected
> 5.12.0-syzkaller #0 Not tainted
> --------------------------------------------------------
> syz-executor132/8391 just changed the state of lock:
> ffff888015967bf8 (&f->f_owner.lock){.+..}-{2:2}, at: f_getown_ex fs/fcntl.c:211 [inline]
> ffff888015967bf8 (&f->f_owner.lock){.+..}-{2:2}, at: do_fcntl+0x8b4/0x1200 fs/fcntl.c:395
> but this lock was taken by another, HARDIRQ-safe lock in the past:
>  (&dev->event_lock){-...}-{2:2}
> 
> and interrupts could create inverse lock ordering between them.
> 
> other info that might help us debug this:
> Chain exists of:
>   &dev->event_lock --> &new->fa_lock --> &f->f_owner.lock
> 
>  Possible interrupt unsafe locking scenario:
> 
>        CPU0                    CPU1
>        ----                    ----
>   lock(&f->f_owner.lock);
>                                local_irq_disable();
>                                lock(&dev->event_lock);
>                                lock(&new->fa_lock);
>   <Interrupt>
>     lock(&dev->event_lock);
> 
>  *** DEADLOCK ***
> 
> This happens because there is a lock hierarchy of
> &dev->event_lock --> &new->fa_lock --> &f->f_owner.lock
> from the following call chain:
> 
>   input_inject_event():
>     spin_lock_irqsave(&dev->event_lock,...);
>     input_handle_event():
>       input_pass_values():
>         input_to_handler():
>           evdev_events():
>             evdev_pass_values():
>               spin_lock(&client->buffer_lock);
>               __pass_event():
>                 kill_fasync():
>                   kill_fasync_rcu():
>                     read_lock(&fa->fa_lock);
>                     send_sigio():
>                       read_lock_irqsave(&fown->lock,...);
> 
> However, since &dev->event_lock is HARDIRQ-safe, interrupts have to be
> disabled while grabbing &f->f_owner.lock, otherwise we invert the lock
> hierarchy.
> 
> Hence, we replace calls to read_lock/read_unlock on &f->f_owner.lock,
> with read_lock_irq/read_unlock_irq.
> 

Patches look reasonable overall, but why does this one use read_lock_irq
and the other one use read_lock_irqsave? Don't we need to *_irqsasve in
both patches?


> Reported-and-tested-by: syzbot+e6d5398a02c516ce5e70 at syzkaller.appspotmail.com
> Signed-off-by: Desmond Cheong Zhi Xi <desmondcheongzx at gmail.com>
> ---
>  fs/fcntl.c | 13 +++++++------
>  1 file changed, 7 insertions(+), 6 deletions(-)
> 
> diff --git a/fs/fcntl.c b/fs/fcntl.c
> index dfc72f15be7f..cf9e81dfa615 100644
> --- a/fs/fcntl.c
> +++ b/fs/fcntl.c
> @@ -150,7 +150,8 @@ void f_delown(struct file *filp)
>  pid_t f_getown(struct file *filp)
>  {
>  	pid_t pid = 0;
> -	read_lock(&filp->f_owner.lock);
> +
> +	read_lock_irq(&filp->f_owner.lock);
>  	rcu_read_lock();
>  	if (pid_task(filp->f_owner.pid, filp->f_owner.pid_type)) {
>  		pid = pid_vnr(filp->f_owner.pid);
> @@ -158,7 +159,7 @@ pid_t f_getown(struct file *filp)
>  			pid = -pid;
>  	}
>  	rcu_read_unlock();
> -	read_unlock(&filp->f_owner.lock);
> +	read_unlock_irq(&filp->f_owner.lock);
>  	return pid;
>  }
>  
> @@ -208,7 +209,7 @@ static int f_getown_ex(struct file *filp, unsigned long arg)
>  	struct f_owner_ex owner = {};
>  	int ret = 0;
>  
> -	read_lock(&filp->f_owner.lock);
> +	read_lock_irq(&filp->f_owner.lock);
>  	rcu_read_lock();
>  	if (pid_task(filp->f_owner.pid, filp->f_owner.pid_type))
>  		owner.pid = pid_vnr(filp->f_owner.pid);
> @@ -231,7 +232,7 @@ static int f_getown_ex(struct file *filp, unsigned long arg)
>  		ret = -EINVAL;
>  		break;
>  	}
> -	read_unlock(&filp->f_owner.lock);
> +	read_unlock_irq(&filp->f_owner.lock);
>  
>  	if (!ret) {
>  		ret = copy_to_user(owner_p, &owner, sizeof(owner));
> @@ -249,10 +250,10 @@ static int f_getowner_uids(struct file *filp, unsigned long arg)
>  	uid_t src[2];
>  	int err;
>  
> -	read_lock(&filp->f_owner.lock);
> +	read_lock_irq(&filp->f_owner.lock);
>  	src[0] = from_kuid(user_ns, filp->f_owner.uid);
>  	src[1] = from_kuid(user_ns, filp->f_owner.euid);
> -	read_unlock(&filp->f_owner.lock);
> +	read_unlock_irq(&filp->f_owner.lock);
>  
>  	err  = put_user(src[0], &dst[0]);
>  	err |= put_user(src[1], &dst[1]);

-- 
Jeff Layton <jlayton at kernel.org>



More information about the Linux-kernel-mentees mailing list