Re: [PATCH] net: fix race in the receive/select

[Eric Dumazet <eric.dumazet@gmail.com>]

Davide Libenzi a écrit :
> On Fri, 26 Jun 2009, Eric Dumazet wrote:
> 
>> Davide Libenzi a écrit :
>>> On Thu, 25 Jun 2009, Oleg Nesterov wrote:
>>>
>>>> Can't really comment this patch, except this all looks reasonable to me.
>>>> Add more CCs.
>>> While this can work, IMO it'd be cleaner to have the smp_mb() moved from 
>>> fs/select.c to the ->poll() function.
>>> Having a barrier that matches another one in another susbsystem, because 
>>> of the special locking logic of such subsystem, is not too shiny IMHO.
>>>
>> Yes but barrier is necessary only if add_wait_queue() was actually called, and __pollwait()
>> does this call.
>>
>> Adding a plain smp_mb() in tcp_poll() for example would slowdown select()/poll() with NULL
>> timeout.
> 
> Do you think of it as good design adding an MB on a subsystem, because of 
> the special locking logic of another one?
> The (eventual) slowdown, IMO can be argued sideways, by saying that 
> non-socket users will pay the price for their polls.
> 

I wont argue with you David, just try to correct bugs.

fs/ext4/ioctl.c line 182

	set_current_state(TASK_INTERRUPTIBLE);
	add_wait_queue(&EXT4_SB(sb)->ro_wait_queue, &wait);
	if (timer_pending(&EXT4_SB(sb)->turn_ro_timer)) {
		schedule();

Another example of missing barrier after add_wait_queue()

Because add_wait_queue() misses a barrier, we have to add one after each call.

Maybe it would be safer to add barrier in add_wait_queue() itself, not in _pollwait().