On Mon, Apr 09, 2018 at 08:20:32AM -0700, Matthew Wilcox wrote:
> On Mon, Apr 09, 2018 at 11:49:58PM +0900, Minchan Kim wrote:
> > On Mon, Apr 09, 2018 at 08:25:06PM +0800, Chao Yu wrote:
> > > On 2018/4/9 19:25, Minchan Kim wrote:
> > > > On Mon, Apr 09, 2018 at 04:14:03AM -0700, Matthew Wilcox wrote:
> > > >> On Mon, Apr 09, 2018 at 12:09:30PM +0900, Minchan Kim wrote:
> > > >>> Look at fs/f2fs/inode.c
> > > >>> mapping_set_gfp_mask(inode->i_mapping, GFP_F2FS_ZERO);
> > > >>>
> > > >>> __add_to_page_cache_locked
> > > >>> radix_tree_maybe_preload
> > > >>>
> > > >>> add_to_page_cache_lru
> > >
> > > No, sometimes, we need to write meta data to new allocated block address,
> > > then we will allocate a zeroed page in inner inode's address space, and
> > > fill partial data in it, and leave other place with zero value which means
> > > some fields are initial status.
> >
> > Thanks for the explaining.
> >
> > > There are two inner inodes (meta inode and node inode) setting __GFP_ZERO,
> > > I have just checked them, for both of them, we can avoid using __GFP_ZERO,
> > > and do initialization by ourselves to avoid unneeded/redundant zeroing
> > > from mm.
> >
> > Yub, it would be desirable for f2fs. Please go ahead for f2fs side.
> > However, I think current problem is orthgonal. Now, the problem is
> > radix_tree_node allocation is bind to page cache allocation.
> > Why does FS cannot allocate page cache with __GFP_ZERO?
> > I agree if the concern is only performance matter as Matthew mentioned.
> > But it is beyond that because it shouldn't do due to limitation
> > of workingset shadow entry implementation. I think such coupling is
> > not a good idea.
> >
> > I think right approach to abstract shadow entry in radix_tree is
> > to mask off __GFP_ZERO in radix_tree's allocation APIs.
>
> I don't think this is something the radix tree should know about.
Because shadow entry implementation is hidden by radix tree implemetation.
IOW, radix tree user cannot know how it works.
> SLAB should be checking for it (the patch I posted earlier in this
I don't think it's right approach. SLAB constructor can initialize
some metadata for slab page populated as well as page zeroing.
However, __GFP_ZERO means only clearing pages, not metadata.
So it's different semantic. No need to mix out.
> thread), but the right place to filter this out is in the caller of
> radix_tree_maybe_preload -- it's already filtering out HIGHMEM pages,
> and should filter out GFP_ZERO too.
radix_tree_[maybe]_preload is exported API, which are error-prone
for out of modules or upcoming customers.
More proper place is __radix_tree_preload.
>
> diff --git a/mm/filemap.c b/mm/filemap.c
> index c2147682f4c3..a87a523eea8e 100644
> --- a/mm/filemap.c
> +++ b/mm/filemap.c
> @@ -785,7 +785,7 @@ int replace_page_cache_page(struct page *old, struct page *new, gfp_t gfp_mask)
> VM_BUG_ON_PAGE(!PageLocked(new), new);
> VM_BUG_ON_PAGE(new->mapping, new);
>
> - error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
> + error = radix_tree_preload(gfp_mask & ~(__GFP_HIGHMEM | __GFP_ZERO));
> if (!error) {
> struct address_space *mapping = old->mapping;
> void (*freepage)(struct page *);
> @@ -841,7 +841,8 @@ static int __add_to_page_cache_locked(struct page *page,
> return error;
> }
>
> - error = radix_tree_maybe_preload(gfp_mask & ~__GFP_HIGHMEM);
> + error = radix_tree_maybe_preload(gfp_mask &
> + ~(__GFP_HIGHMEM | __GFP_ZERO));
> if (error) {
> if (!huge)
> mem_cgroup_cancel_charge(page, memcg, false);
