Re: 2.4, 2.5: SMP race: __sync_single_inode vs. __mark_inode_dirty

[Andrew Morton]

Mikulas Patocka <[e-mail]> wrote:
>
> Hi.
> 
> there's a SMP race condition between __sync_single_inode (or __sync_one on
> 2.4.20) and __mark_inode_dirty. __mark_inode_dirty doesn't take inode
> spinlock. As we know -- unless you take a spinlock or use barrier,
> processor can change order of instructions.
> 

Looks good to me, although my understanding of these memory ordering issues
is woeful.

We do want to avoid taking inode_lock in mark_inode_dirty() - that is called
very frequently.  I'm rather surprised that inode_lock contention has not
been a problem thus far.

Longer-term we should probably turn i_state into a ulong and only run atomic
bitops against it.