Re: [PATCH 1/5] ext4: Avoid races caused by on-line resizing andSMP memory reordering

[Jan Kara]

On Tue 28-04-09 13:14:45, Theodore Tso wrote:
> On Tue, Apr 28, 2009 at 06:23:59PM +0200, Jan Kara wrote:
> >   Ouch... Hmm, smp_rmb() isn't completely free and mainly it's a bit
> > ugly and prone to errors (I'm afraid next time someone changes the
> > allocation code, we miss some barriers again)... so.. Maybe a stupid
> > idea but wouldn't it be easier to solve the online resize like: freeze
> > the filesystem, do all the changes required for extend, unfreeze the
> > filesystem?
> 
> Eric suggested a helper function when reading from s_groups_count.
> That would take care of the code maintenance headache.  One problem
> with using freeze/thaw is it won't work without a journal, and we do
> support filesystems without journals in ext4.  (Probably the best
> solution for netbooks with crapola SSD's.)
  Well, in non-journalling case, we could introduce a rw semaphore
(read acquired / released in journal_start / journal_stop, write acquired
when the fs is frozen). This might be useful for other rare cases where
freezing the fs would be beneficial. But yes, if wrapping into a helper
function works then that might be the easiest way to go.

> As far as smb_rmb() not being free, it's essentially free for
> x86/x86_64 platforms.  Is it really that costly on other
> architectures?
  I had a feeling that it's not that expensive but not quite free either on
x86/x86_64 (I know even less about other archs) - it has to lock the bus,
writes in local CPU caches have to be flushed, no? Probably we don't care
given the size of the functions doing allocation... As an excercise I was
trying to google some numbers but was not really successful, just some
comments about tens of cycles in some emails...

									Honza
-- 
Jan Kara <jack@xxxxxxx>
SUSE Labs, CR
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/