Re: [PATCH 1.1/2] xfs: abstract PF_FSTRANS to PF_MEMALLOC_NOFS

[Michal Hocko]

On Wed 27-04-16 11:41:51, Andreas Dilger wrote:
> On Apr 27, 2016, at 5:54 AM, Michal Hocko <mhocko@xxxxxxxxxx> wrote:
[...]
> > --- a/fs/xfs/kmem.c
> > +++ b/fs/xfs/kmem.c
> > @@ -80,13 +80,13 @@ kmem_zalloc_large(size_t size, xfs_km_flags_t flags)
> > 	 * context via PF_MEMALLOC_NOIO to prevent memory reclaim re-entering
> > 	 * the filesystem here and potentially deadlocking.
> > 	 */
> > -	if ((current->flags & PF_FSTRANS) || (flags & KM_NOFS))
> > +	if ((current->flags & PF_MEMALLOC_NOFS) || (flags & KM_NOFS))
> > 		noio_flag = memalloc_noio_save();
> > 
> > 	lflags = kmem_flags_convert(flags);
> > 	ptr = __vmalloc(size, lflags | __GFP_HIGHMEM | __GFP_ZERO, PAGE_KERNEL);
> > 
> > -	if ((current->flags & PF_FSTRANS) || (flags & KM_NOFS))
> > +	if ((current->flags & PF_MEMALLOC_NOFS) || (flags & KM_NOFS))
> > 		memalloc_noio_restore(noio_flag);
> 
> Not really the fault of this patch, but it brings this nasty bit of code into
> the light.  Is all of this machinery still needed given that __vmalloc() can
> accept GFP flags?  If yes, wouldn't it be better to fix __vmalloc() to honor
> the GFP flags instead of working around it in the filesystem code?

This is not that easy. __vmalloc can accept gfp flags but it doesn't
honor __GFP_IO 100%. IIRC some paths like page table allocations are
hardcoded GFP_KERNEL. Besides that I would like to have GFP_NOIO used
via memalloc_noio_{save,restore} API as well for the similar reasons as
GFP_NOFS - it is just easier to explain scope than particular code paths
which might be shared.
-- 
Michal Hocko
SUSE Labs