Re: [RFC] Bug in journal_commit_transaction?

[Jan Kara <jack@suse.cz>]

  Hello,

> On Thu, 2006-04-20 at 10:47 +0200, Jan Kara wrote:
> 
> >   I've got no answer but I still think my argument was sound ;) Below
> > is a patch also with a verbose comment what it fixes and how.
> 
> Looks good to me, but it complicates the logic for this case and removes
> some useful debug checks for conditions which are almost always true.
> 
> Wouldn't it be better and safer just to make those almost-always true
> conditions always true?  We can do that pretty simply, just by
> processing the t_forget list in two passes, always doing the bitmaps
> last.  (Though this will need care, as the list is potentially moving
> under our feet...)
> 
> Actually, scratch that, because journal_unmap_buffer() can return a
> buffer to the committing transaction's forget list at any time, so we
> really cannot guarantee to do all non-bitmaps first: new non-bitmaps
> might arrive after we've started doing the bitmaps.  Ugh.
  Thanks for looking into it.  Yes, I agree that keeping that safety
check would be nicer. But I was thinking for some time if we cannot
somehow make sure that bitmaps are processed last but I didn't come up
with anything useful. One possibility might be to do some magic in
do_get_write_access() and get_create_access() if we find out we are
reallocating buffer on forget list but I'm afraid that simple solutions
(e.g. wait for commit) would harm performance which might be even
worse...

								Honza

-- 
Jan Kara <jack@suse.cz>
SuSE CR Labs