On Tue, Mar 14, 2023 at 06:36:21PM -0400, Theodore Ts'o wrote:
> On Thu, Jan 26, 2023 at 08:23:54PM +0000, Matthew Wilcox (Oracle) wrote:
> > Saves a number of calls to compound_head().
>
> Is this left over from an earlier version of this patch series? There
> are no changes to calls to compound_head() that I can find in this
> patch.
They're hidden. Here are the ones from this patch:
- if (!PageUptodate(page)) {
- unlock_page(page);
- put_page(page);
- unlock_page(page);
- put_page(page);
That's five. I may have missed some.
> > @@ -565,10 +564,9 @@ static int ext4_convert_inline_data_to_extent(struct address_space *mapping,
> >
> > /* We cannot recurse into the filesystem as the transaction is already
> > * started */
> > - flags = memalloc_nofs_save();
> > - page = grab_cache_page_write_begin(mapping, 0);
> > - memalloc_nofs_restore(flags);
> > - if (!page) {
> > + folio = __filemap_get_folio(mapping, 0, FGP_WRITEBEGIN | FGP_NOFS,
> > + mapping_gfp_mask(mapping));
> > + if (!folio) {
> > ret = -ENOMEM;
> > goto out;
> > }
>
> Is there a reason why to use FGP_NOFS as opposed to using
> memalloc_nofs_{save,restore}()?
>
> I thought using memalloc_nofs_save() is considered the perferred
> approach by mm-folks.
Ideally, yes, we'd use memalloc_nofs_save(), but not like this! The way
it's supposed to be used is at the point where you do something which
makes the fs non-reentrant. ie when you start the transaction, you should
be calling memalloc_nofs_save() and when you finish the transaction,
you should be calling memalloc_nofs_restore(). That way, you don't
need to adorn the entire filesystem with GFP_NOFS/FGP_NOFS/whatever,
you have one place where you mark yourself non-reentrant and you're done.
Once ext4 does this every time it starts a transaction, we can drop
the FGP_NOFS flag usage in ext4, and once every filesystem does it,
we can drop the entire flag, and that will make me happy. It's a long
road, though.
