Re: [PATCH] [8/18] BKL-removal: Remove BKL from remote_llseek

[Andi Kleen <ak@suse.de>]

On Monday 28 January 2008 00:08:56 Trond Myklebust wrote:
> 
> On Sun, 2008-01-27 at 16:18 -0600, Steve French wrote:
> > If two seeks overlap, can't you end up with an f_pos value that is
> > different than what either thread seeked to? or if you have a seek and
> > a read overlap can't you end up with the read occurring in the midst
> > of an update of f_pos (which takes more than one instruction on
> > various architectures), e.g. reading an f_pos, which has only the
> > lower half of a 64 bit field updated?   I agree that you shouldn't
> > have seeks racing in parallel but I think it is preferable to get
> > either the updated f_pos or the earlier f_pos not something 1/2
> > updated.
> 
> Why? The threads are doing something inherently liable to corrupt data
> anyway. If they can race over the seek, why wouldn't they race over the
> read or write too?
> The race in lseek() should probably be the least of your worries in this
> case.

The problem is that it's not a race in who gets to do its thing first, but a 
parallel reader can actually see a corrupted value from the two independent 
words on 32bit (e.g. during a 4GB). And this could actually completely corrupt 
f_pos when it happens with two racing relative seeks or read/write()s

I would consider that a bug.

Fixes would be either to always take a spinlock to update this (nasty on 
platforms where spinlocks are expensive like P4) or define some architecture
specific way to read/write 64bit values consistently. In theory also some
lazy locking seqlock like mechanism could be used, but that had the disadvantage
of being theoretically starvable.

-Andi