On Mon, Jul 26, 2021 at 03:03:14PM +0200, Andreas Gruenbacher wrote:
> On Mon, Jul 26, 2021 at 2:33 PM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
> > On Mon, Jul 26, 2021 at 01:06:11PM +0200, Andreas Gruenbacher wrote:
> > > @@ -671,11 +683,11 @@ static size_t iomap_write_end_inline(struct inode *inode, struct page *page,
> > > void *addr;
> > >
> > > WARN_ON_ONCE(!PageUptodate(page));
> > > - BUG_ON(pos + copied > PAGE_SIZE - offset_in_page(iomap->inline_data));
> > > + BUG_ON(!iomap_inline_data_size_valid(iomap));
> > >
> > > flush_dcache_page(page);
> > > addr = kmap_atomic(page);
> > > - memcpy(iomap->inline_data + pos, addr + pos, copied);
> > > + memcpy(iomap_inline_data(iomap, pos), addr + pos, copied);
> > > kunmap_atomic(addr);
> > >
> > > mark_inode_dirty(inode);
> >
> > Only tangentially related ... why do we memcpy the data into the tail
> > at write_end() time instead of at writepage() time? I see there's a
> > workaround for that in gfs2's page_mkwrite():
> >
> > if (gfs2_is_stuffed(ip)) {
> > err = gfs2_unstuff_dinode(ip);
> >
> > (an mmap store cannot change the size of the file, so this would be
> > unnecessary)
>
> Not sure if an additional __set_page_dirty_nobuffers is needed in that
> case, but doing the writeback at writepage time should work just as
> well. It's just that gfs2 did it at write time historically. The
> un-inlining in gfs2_page_mkwrite() could probably also be removed.
>
> I can give this a try, but I'll unfortunately be AFK for the next
> couple of days.
I tend to leave it as another new story and can be resolved with
another patch to improve it (or I will stuck in this, I need to do
my own development stuff instead of spinning with this iomap patch
since I can see this already work well for gfs2 and erofs), I will
update the patch Andreas posted with Christoph's comments.
Thanks,
Gao Xiang
