Re: [PATCH] mm: slub: Ensure that slab_unlock() is atomic

From: Vineet Gupta
Date: Wed Mar 09 2016 - 01:43:53 EST

+CC linux-arch, parisc folks, PeterZ

On Wednesday 09 March 2016 02:10 AM, Christoph Lameter wrote:
> On Tue, 8 Mar 2016, Vineet Gupta wrote:
>> # set the bit
>> 80543b8e: ld_s r2,[r13,0] <--- (A) Finds PG_locked is set
>> 80543b90: or r3,r2,1 <--- (B) other core unlocks right here
>> 80543b94: st_s r3,[r13,0] <--- (C) sets PG_locked (overwrites unlock)
> Duh. Guess you need to take the spinlock also in the arch specific
> implementation of __bit_spin_unlock(). This is certainly not the only case
> in which we use the __ op to unlock.

__bit_spin_lock() by definition is *not* required to be atomic, bit_spin_lock() is
- so I don't think we need a spinlock there.

There is clearly a problem in slub code that it is pairing a test_and_set_bit()
with a __clear_bit(). Latter can obviously clobber former if they are not a single
instruction each unlike x86 or they use llock/scond kind of instructions where the
interim store from other core is detected and causes a retry of whole llock/scond

BTW ARC is not the only arch which suffers from this - other arches potentially
also are. AFAIK PARISC also doesn't have atomic r-m-w and also uses a set of
external hashed spinlocks to protect the r-m-w sequences.

So there also we have the same race because the outer spin lock is not taken for
slab_unlock() -> __bit_spin_lock() -> __clear_bit.

Arguably I can fix the ARC !LLSC variant of test_and_set_bit() to not set the bit
unconditionally but only if it was clear (PARISC does the same). That would be a
slight micro-optimization as we won't need another snoop transaction to make line
writable and that would also elide this problem, but I think there is a
fundamental problem here in slub which is pairing atomic and non atomic ops - for
performance reasons. It doesn't work on all arches and/or configurations.

> You need a true atomic op or you need to take the "spinlock" in all
> cases where you modify the bit.

No we don't in __bit_spin_lock and we already do in bit_spin_lock.

> If you take the lock in __bit_spin_unlock
> then the race cannot happen.

Of course it won't but that means we penalize all non atomic callers of the API
with a superfluous spinlock which is not require din first place given the
definition of API.

>> Are you convinced now !
> Yes, please fix your arch specific code.