When collapse_file() calls try_to_release_page(), it has already
isolated the page: so if releasing buffers happens to fail (as it
sometimes does), remember to putback_lru_page(): otherwise that page is
left unreclaimable and unfreeable, and the file extent uncollapsible.
Signed-off-by: Hugh Dickins <[email protected]>
Fixes: 99cb0dbd47a1 ("mm,thp: add read-only THP support for (non-shmem) FS")
Cc: [email protected] # v5.4+
---
mm/khugepaged.c | 1 +
1 file changed, 1 insertion(+)
--- 5.7-rc6/mm/khugepaged.c 2020-04-12 16:24:37.710999073 -0700
+++ linux/mm/khugepaged.c 2020-05-10 17:06:21.788398646 -0700
@@ -1692,6 +1692,7 @@ static void collapse_file(struct mm_stru
if (page_has_private(page) &&
!try_to_release_page(page, GFP_KERNEL)) {
result = SCAN_PAGE_HAS_PRIVATE;
+ putback_lru_page(page);
goto out_unlock;
}
> On May 23, 2020, at 6:50 PM, Hugh Dickins <[email protected]> wrote:
>
> When collapse_file() calls try_to_release_page(), it has already
> isolated the page: so if releasing buffers happens to fail (as it
> sometimes does), remember to putback_lru_page(): otherwise that page is
> left unreclaimable and unfreeable, and the file extent uncollapsible.
>
> Signed-off-by: Hugh Dickins <[email protected]>
> Fixes: 99cb0dbd47a1 ("mm,thp: add read-only THP support for (non-shmem) FS")
> Cc: [email protected] # v5.4+
Acked-by: Song Liu <[email protected]>
Thanks for the fix!
> ---
>
> mm/khugepaged.c | 1 +
> 1 file changed, 1 insertion(+)
>
> --- 5.7-rc6/mm/khugepaged.c 2020-04-12 16:24:37.710999073 -0700
> +++ linux/mm/khugepaged.c 2020-05-10 17:06:21.788398646 -0700
> @@ -1692,6 +1692,7 @@ static void collapse_file(struct mm_stru
> if (page_has_private(page) &&
> !try_to_release_page(page, GFP_KERNEL)) {
> result = SCAN_PAGE_HAS_PRIVATE;
> + putback_lru_page(page);
> goto out_unlock;
> }
>
On Sat, May 23, 2020 at 06:50:15PM -0700, Hugh Dickins wrote:
> When collapse_file() calls try_to_release_page(), it has already
> isolated the page: so if releasing buffers happens to fail (as it
> sometimes does), remember to putback_lru_page(): otherwise that page is
> left unreclaimable and unfreeable, and the file extent uncollapsible.
>
> Signed-off-by: Hugh Dickins <[email protected]>
> Fixes: 99cb0dbd47a1 ("mm,thp: add read-only THP support for (non-shmem) FS")
> Cc: [email protected] # v5.4+
Acked-by: Kirill A. Shutemov <[email protected]>
--
Kirill A. Shutemov
On Sat, May 23, 2020 at 06:50:15PM -0700, Hugh Dickins wrote:
> When collapse_file() calls try_to_release_page(), it has already
> isolated the page: so if releasing buffers happens to fail (as it
> sometimes does), remember to putback_lru_page(): otherwise that page is
> left unreclaimable and unfreeable, and the file extent uncollapsible.
Oof, I could imagine that was painful to debug (unless you already
suspected file THP due to a targeted test or similar). Kudos.
> Signed-off-by: Hugh Dickins <[email protected]>
> Fixes: 99cb0dbd47a1 ("mm,thp: add read-only THP support for (non-shmem) FS")
> Cc: [email protected] # v5.4+
Acked-by: Johannes Weiner <[email protected]>
On Tue, 26 May 2020, Johannes Weiner wrote:
> On Sat, May 23, 2020 at 06:50:15PM -0700, Hugh Dickins wrote:
> > When collapse_file() calls try_to_release_page(), it has already
> > isolated the page: so if releasing buffers happens to fail (as it
> > sometimes does), remember to putback_lru_page(): otherwise that page is
> > left unreclaimable and unfreeable, and the file extent uncollapsible.
>
> Oof, I could imagine that was painful to debug (unless you already
> suspected file THP due to a targeted test or similar). Kudos.
Thanks, but I have to refuse both your admiration and sympathy:
mercifully, it was just something I noticed by source inspection
when working there.
I did then put in a debug count to see if it ever got hit in practice,
and checked after big multi-testing runs: it was sometimes hit, but
certainly not often, and I don't know what it was racing with when
it happened - would depend on filesystem anyway (ext4 in our case).
Saying "source inspection" reminds me: there is another funny in there,
but I don't think it matters very much in practice, and might need
rather a lot of testing to justify any particular patch: where
page_cache_sync_readahead() asks for PAGE_SIZE pages!
"end - index" seems a more reasonable number to me: but then we
might find that reading ahead into the next huge extent had actually
been a useful optimization (and those readahead functions impose
their own caps, so PAGE_SIZE shouldn't work out too outrageously).
Hugh
>
> > Signed-off-by: Hugh Dickins <[email protected]>
> > Fixes: 99cb0dbd47a1 ("mm,thp: add read-only THP support for (non-shmem) FS")
> > Cc: [email protected] # v5.4+
>
> Acked-by: Johannes Weiner <[email protected]>
On Thu, 28 May 2020 at 02:58, Hugh Dickins <[email protected]> wrote:
>
> On Tue, 26 May 2020, Johannes Weiner wrote:
> > On Sat, May 23, 2020 at 06:50:15PM -0700, Hugh Dickins wrote:
> > > When collapse_file() calls try_to_release_page(), it has already
> > > isolated the page: so if releasing buffers happens to fail (as it
> > > sometimes does), remember to putback_lru_page(): otherwise that page is
> > > left unreclaimable and unfreeable, and the file extent uncollapsible.
> >
> > Oof, I could imagine that was painful to debug (unless you already
> > suspected file THP due to a targeted test or similar). Kudos.
>
> Thanks, but I have to refuse both your admiration and sympathy:
> mercifully, it was just something I noticed by source inspection
> when working there.
>
> I did then put in a debug count to see if it ever got hit in practice,
> and checked after big multi-testing runs: it was sometimes hit, but
> certainly not often, and I don't know what it was racing with when
> it happened - would depend on filesystem anyway (ext4 in our case).
>
> Saying "source inspection" reminds me: there is another funny in there,
> but I don't think it matters very much in practice, and might need
> rather a lot of testing to justify any particular patch: where
> page_cache_sync_readahead() asks for PAGE_SIZE pages!
>
> "end - index" seems a more reasonable number to me: but then we
> might find that reading ahead into the next huge extent had actually
> been a useful optimization (and those readahead functions impose
> their own caps, so PAGE_SIZE shouldn't work out too outrageously).
My two cents,
Do you have a test case / test steps to validate this patch ?
- Naresh
On Wed, May 27, 2020 at 02:28:10PM -0700, Hugh Dickins wrote:
> Saying "source inspection" reminds me: there is another funny in there,
> but I don't think it matters very much in practice, and might need
> rather a lot of testing to justify any particular patch: where
> page_cache_sync_readahead() asks for PAGE_SIZE pages!
>
> "end - index" seems a more reasonable number to me: but then we
> might find that reading ahead into the next huge extent had actually
> been a useful optimization (and those readahead functions impose
> their own caps, so PAGE_SIZE shouldn't work out too outrageously).
That readahead was only added in 99cb0dbd47a15d395bf3faa78dc122bc5efe3fc0
so it probably hasn't really been performance tested yet.
On Wed, 3 Jun 2020, Matthew Wilcox wrote:
> On Wed, May 27, 2020 at 02:28:10PM -0700, Hugh Dickins wrote:
> > Saying "source inspection" reminds me: there is another funny in there,
> > but I don't think it matters very much in practice, and might need
> > rather a lot of testing to justify any particular patch: where
> > page_cache_sync_readahead() asks for PAGE_SIZE pages!
> >
> > "end - index" seems a more reasonable number to me: but then we
> > might find that reading ahead into the next huge extent had actually
> > been a useful optimization (and those readahead functions impose
> > their own caps, so PAGE_SIZE shouldn't work out too outrageously).
>
> That readahead was only added in 99cb0dbd47a15d395bf3faa78dc122bc5efe3fc0
> so it probably hasn't really been performance tested yet.
I can well imagine that an entirely new body of code, exercised
in the background by khugepaged, will not be a leading candidate for
performance comparisons - rightly so; but arguing that by the commit
being in the tree for only eight months seems... odd :)
Hugh