On Thu, 22 Jul 2010, Ted Ts'o wrote:
> > I'll change this to
> >
> > do {
> > new_transaction = kzalloc(sizeof(*new_transaction),
> > GFP_NOFS);
> > } while (!new_transaction);
> >
> > in the next phase when I introduce __GFP_KILLABLE (that jbd and jbd2 can't
> > use because they are GFP_NOFS).
>
> OK, I can carry a patch which does this in the ext4 tree to push to
> linus when the merge window opens shortly, since the goal is you want
> to get rid of __GFP_NOFAIL altogether, right?
>
Yup, I was trying to do the removal in two phases.
First, remove __GFP_NOFAIL from callers that don't seem to need it. I
found that they were actually needed in some cases such as jbd, jbd2, and
sparc although the reason was specific to those subsystems at a higher
level and their error handling was actually unused code since __GFP_NOFAIL
cannot return NULL.
Second, replace __GFP_NOFAIL with __GFP_KILLABLE which converts existing
users of __GFP_NOFAIL into the do-while loop above and adding
__GFP_KILLABLE for allocations allowing __GFP_FS which does memory
compaction for order > 0, direct reclaim, and the oom killer but does not
retry the allocation. That would be the responsibility of the caller.
This ends up removing several branches from the page allocator.
I didn't think about converting the existing GFP_NOFS | __GFP_NOFAIL
callers into the do-while loop above until you mentioned it, thanks. I'll
send patches to do that shortly.
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxxx For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>