Re: [RESEND PATCH] kernfs: Avoid re-adding kernfs_node into kernfs_notify_list.

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



Hello Tejun,
Thanks for your feedback.

On 2/7/22 6:11 am, Tejun Heo wrote:
> Hello,
> 
> On Sat, Jul 02, 2022 at 01:46:04AM +1000, Imran Khan wrote:
>> @@ -992,9 +993,11 @@ void kernfs_notify(struct kernfs_node *kn)
>>  	rcu_read_unlock();
>>  
>>  	/* schedule work to kick fsnotify */
>> -	kernfs_get(kn);
>> -	llist_add(&kn->attr.notify_next, &kernfs_notify_list);
>> -	schedule_work(&kernfs_notify_work);
>> +	if (kn->attr.notify_next.next != NULL) {
>> +		kernfs_get(kn);
>> +		llist_add(&kn->attr.notify_next, &kernfs_notify_list);
>> +		schedule_work(&kernfs_notify_work);
>> +	}
> 
> Aren't you just narrowing the race window here? What prevents two
> threads simultaneously testing for non NULL and then entering the
> addition block together?
> 
Indeed that is possible.
> Looked at the llist code and it doesn't support multiple producers
> trying to add the same node, unfortunately, so I'm not sure llist is
> gonna work here. For now, the right thing to do prolly is reverting
> it.
> 

Can we use kernfs_notify_lock like below snippet to serialize producers
(kernfs_notify):

spin_lock_irqsave(&kernfs_notify_lock, flags);
if (kn->attr.notify_next.next != NULL) {
	kernfs_get(kn);
	llist_add(&kn->attr.notify_next, &kernfs_notify_list);
	schedule_work(&kernfs_notify_work);
}
spin_unlock_irqsave(&kernfs_notify_lock, flags);

As per following comments at the beginning of llist.h

 * Cases where locking is not needed:
 * If there are multiple producers and multiple consumers, llist_add can be
 * used in producers and llist_del_all can be used in consumers simultaneously
 * without locking. Also a single consumer can use llist_del_first while
 * multiple producers simultaneously use llist_add, without any locking.

Multiple producers and single consumer can work in parallel but as in our case
addition is dependent on kn->attr.notify_next.next != NULL, we may keep the
checking and list addition under kernfs_notify_lock and for consumer just lock
free->next = NULL under kernfs_notify_lock.

Having said this, I am okay with reverting the llist change as well, because
anyways it is not helping in the contentions that we are chasing here, but I
thought of sharing the above idea to see if it is reliable and better than
revert option.

Thanks
 -- Imran




[Index of Archives]     [Kernel Newbies]     [Security]     [Netfilter]     [Bugtraq]     [Linux PPP]     [Linux FS]     [Yosemite News]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Samba]     [Video 4 Linux]     [Linmodem]     [Device Mapper]     [Linux Kernel for ARM]

  Powered by Linux