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

From: Vineet Gupta
Date: Wed Mar 09 2016 - 06:13:24 EST

On Wednesday 09 March 2016 04:01 PM, Peter Zijlstra wrote:
> On Wed, Mar 09, 2016 at 11:13:49AM +0100, Peter Zijlstra wrote:
>> ---
>> Subject: bitops: Do not default to __clear_bit() for __clear_bit_unlock()
>> __clear_bit_unlock() is a special little snowflake. While it carries the
>> non-atomic '__' prefix, it is specifically documented to pair with
>> test_and_set_bit() and therefore should be 'somewhat' atomic.
>> Therefore the generic implementation of __clear_bit_unlock() cannot use
>> the fully non-atomic __clear_bit() as a default.
>> If an arch is able to do better; is must provide an implementation of
>> __clear_bit_unlock() itself.
> FWIW, we could probably undo this if we unified all the spinlock based
> atomic ops implementations (there's a whole bunch doing essentially the
> same), and special cased __clear_bit_unlock() for that.
> Collapsing them is probably a good idea anyway, just a fair bit of
> non-trivial work to figure out all the differences and if they matter
> etc..

Indeed I thought about this when we first did the SMP port. The only issue was
somewhat more generated code with the hashed spinlocks (vs. my dumb 2 spin locks -
which as I see now will also cause false sharing - likely ending up in the same
cache line), but I was more of a micro-optimization freak then than I'm now :-)

So yeah I agree !