fanotify: Avoid softlockups when reading many events
authorJan Kara <jack@suse.cz>
Wed, 15 Jul 2020 12:06:21 +0000 (14:06 +0200)
committerJan Kara <jack@suse.cz>
Wed, 15 Jul 2020 13:23:28 +0000 (15:23 +0200)
When user provides large buffer for events and there are lots of events
available, we can try to copy them all to userspace without scheduling
which can softlockup the kernel (furthermore exacerbated by the
contention on notification_lock). Add a scheduling point after copying
each event.

Note that usually the real underlying problem is the cost of fanotify
event merging and the resulting contention on notification_lock but this
is a cheap way to somewhat reduce the problem until we can properly
address that.

Reported-by: Francesco Ruggeri <fruggeri@arista.com>
Link: https://lore.kernel.org/lkml/20200714025417.A25EB95C0339@us180.sjc.aristanetworks.com
Reviewed-by: Amir Goldstein <amir73il@gmail.com>
Signed-off-by: Jan Kara <jack@suse.cz>
fs/notify/fanotify/fanotify_user.c

index 63b5dff..d7f63ae 100644 (file)
@@ -412,6 +412,11 @@ static ssize_t fanotify_read(struct file *file, char __user *buf,
 
        add_wait_queue(&group->notification_waitq, &wait);
        while (1) {
+               /*
+                * User can supply arbitrarily large buffer. Avoid softlockups
+                * in case there are lots of available events.
+                */
+               cond_resched();
                event = get_one_event(group, count);
                if (IS_ERR(event)) {
                        ret = PTR_ERR(event);