On Wed, Nov 25, 2020 at 04:11:57PM -0800, Hugh Dickins wrote:
> The little fix definitely needed was shown by generic/083: each
> fsstress waiting for page lock, happens even without forcing huge
> pages. See below...
Huh ... I need to look into why my xfstests run is skipping generic/083:
0006 generic/083 3s ... run fstests generic/083 at 2020-11-26 12:11:52
0006 [not run] this test requires a valid $SCRATCH_MNT and unique
0006 Ran: generic/083
0006 Not run: generic/083
> > if (!unfalloc || !PageUptodate(page)) {
> > if (page_mapping(page) != mapping) {
> > /* Page was replaced by swap: retry */
> > unlock_page(page);
> > - index--;
> > + put_page(page);
> > break;
> > }
> > VM_BUG_ON_PAGE(PageWriteback(page), page);
> > - if (shmem_punch_compound(page, start, end))
> > - truncate_inode_page(mapping, page);
> > - else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
> > - /* Wipe the page and don't get stuck */
> > - clear_highpage(page);
> > - flush_dcache_page(page);
> > - set_page_dirty(page);
> > - if (index <
> > - round_up(start, HPAGE_PMD_NR))
> > - start = index + 1;
> > - }
> > + index = truncate_inode_partial_page(mapping,
> > + page, lstart, lend);
> > + if (index > end)
> > + end = indices[i] - 1;
> > }
> > - unlock_page(page);
>
> The fix needed is here: instead of deleting that unlock_page(page)
> line, it needs to be } else { unlock_page(page); }
It also needs a put_page(page);
That's now taken care of by truncate_inode_partial_page(), so if we're
not calling that, we need to put the page as well. ie this:
+++ b/mm/shmem.c
@@ -954,6 +954,9 @@ static void shmem_undo_range(struct inode *inode, loff_t lstart, loff_t lend,
page, lstart, lend);
if (index > end)
end = indices[i] - 1;
+ } else {
+ unlock_page(page);
+ put_page(page);
}
}
index = indices[i - 1] + 1;
> > }
> > + index = indices[i - 1] + 1;
> > pagevec_remove_exceptionals(&pvec);
> > - pagevec_release(&pvec);
> > - index++;
> > + pagevec_reinit(&pvec);
