| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Message-ID: <ZC2kSfNUXKK4PfpM@google.com>
Date: Wed, 5 Apr 2023 09:39:37 -0700
From: Jaegeuk Kim <jaegeuk@...nel.org>
To: Matthew Wilcox <willy@...radead.org>
Cc: Chao Yu <chao@...nel.org>, linux-kernel@...r.kernel.org,
linux-f2fs-devel@...ts.sourceforge.net, stable@...r.kernel.org,
linux-mm@...ck.org, linux-fsdevel@...r.kernel.org
Subject: Re: [f2fs-dev] [PATCH] f2fs: get out of a repeat loop when getting a
locked data page
On 03/27, Matthew Wilcox wrote:
> On Mon, Mar 27, 2023 at 08:30:33AM -0700, Jaegeuk Kim wrote:
> > On 03/26, Chao Yu wrote:
> > > On 2023/3/24 5:39, Jaegeuk Kim wrote:
> > > > https://bugzilla.kernel.org/show_bug.cgi?id=216050
> > > >
> > > > Somehow we're getting a page which has a different mapping.
> > > > Let's avoid the infinite loop.
> > > >
> > > > Cc: <stable@...r.kernel.org>
> > > > Signed-off-by: Jaegeuk Kim <jaegeuk@...nel.org>
> > > > ---
> > > > fs/f2fs/data.c | 8 ++------
> > > > 1 file changed, 2 insertions(+), 6 deletions(-)
> > > >
> > > > diff --git a/fs/f2fs/data.c b/fs/f2fs/data.c
> > > > index bf51e6e4eb64..80702c93e885 100644
> > > > --- a/fs/f2fs/data.c
> > > > +++ b/fs/f2fs/data.c
> > > > @@ -1329,18 +1329,14 @@ struct page *f2fs_get_lock_data_page(struct inode *inode, pgoff_t index,
> > > > {
> > > > struct address_space *mapping = inode->i_mapping;
> > > > struct page *page;
> > > > -repeat:
> > > > +
> > > > page = f2fs_get_read_data_page(inode, index, 0, for_write, NULL);
> > > > if (IS_ERR(page))
> > > > return page;
> > > > /* wait for read completion */
> > > > lock_page(page);
> > > > - if (unlikely(page->mapping != mapping)) {
> > >
> > > How about using such logic only for move_data_page() to limit affect for
> > > other paths?
> >
> > Why move_data_page() only? If this happens, we'll fall into a loop in anywhere?
> >
> > >
> > > Jaegeuk, any thoughts about why mapping is mismatch in between page's one and
> > > inode->i_mapping?
> >
> > >
> > > After several times code review, I didn't get any clue about why f2fs always
> > > get the different mapping in a loop.
> >
> > I couldn't find the path to happen this. So weird. Please check the history in the
> > bug.
> >
> > >
> > > Maybe we can loop MM guys to check whether below folio_file_page() may return
> > > page which has different mapping?
> >
> > Matthew may have some idea on this?
>
> There's a lot of comments in the bug ... hard to come into this one
> cold.
>
> I did notice this one (#119):
> : Interestingly, ref count is 514, which looks suspiciously as a binary
> : flag 1000000010. Is it possible that during 5.17/5.18 implementation
> : of a "pin", somehow binary flag was written to ref count, or something
> : like '1 << ...' happens?
>
> That indicates to me that somehow you've got hold of a THP that is in
> the page cache. Probably shmem/tmpfs. That indicate to me a refcount
> problem that looks something like this:
>
> f2fs allocates a page
> f2fs adds the page to the page cache
> f2fs puts the reference to the page without removing it from the
> page cache (how?)
Is it somewhat related to setting a bit in private field?
When we migrate the blocks, we do:
1) get_lock_page()
2) submit read
3) lock_page()
3) set_page_dirty()
4) set_page_private_gcing(page)
--- in fs/f2fs/f2fs.h
1409 #define PAGE_PRIVATE_SET_FUNC(name, flagname) \
1410 static inline void set_page_private_##name(struct page *page) \
1411 { \
1412 if (!PagePrivate(page)) { \
1413 get_page(page); \
1414 SetPagePrivate(page); \
1415 set_page_private(page, 0); \
1416 } \
1417 set_bit(PAGE_PRIVATE_NOT_POINTER, &page_private(page)); \
1418 set_bit(PAGE_PRIVATE_##flagname, &page_private(page)); \
1419 }
5) set_page_writebac()
6) submit write
7) unlock_page()
8) put_page(page)
Later, f2fs_invalidate_folio will do put_page again by:
clear_page_private_gcing(&folio->page);
--- in fs/f2fs/f2fs.h
1421 #define PAGE_PRIVATE_CLEAR_FUNC(name, flagname) \
1422 static inline void clear_page_private_##name(struct page *page) \
1423 { \
1424 clear_bit(PAGE_PRIVATE_##flagname, &page_private(page)); \
1425 if (page_private(page) == BIT(PAGE_PRIVATE_NOT_POINTER)) { \
1426 set_page_private(page, 0); \
1427 if (PagePrivate(page)) { \
1428 ClearPagePrivate(page); \
1429 put_page(page); \
1430 }\
1431 } \
1432 }
> page is now free, gets reallocated into a THP
> lookup from the f2fs file finds the new THP
> things explode messily
>
> Checking page->mapping is going to avoid the messy explosion, but
> you'll still have a page in the page cache which doesn't actually
> belong to you, and that's going to lead to subtle data corruption.
>
> This should be caught by page_expected_state(), called from
> free_page_is_bad(), called from free_pages_prepare(). Do your testers
> have CONFIG_DEBUG_VM enabled? That might give you a fighting chance at
> finding the last place which called put_page(). It won't necessarily be
> the _wrong_ place to call put_page() (that may have happened earlier),
> but it may give you a clue.
>
> > >
> > > struct page *pagecache_get_page(struct address_space *mapping, pgoff_t index,
> > > int fgp_flags, gfp_t gfp)
> > > {
> > > struct folio *folio;
> > >
> > > folio = __filemap_get_folio(mapping, index, fgp_flags, gfp);
> > > if (IS_ERR(folio))
> > > return NULL;
> > > return folio_file_page(folio, index);
> > > }
> > >
> > > Thanks,
> > >
> > > > - f2fs_put_page(page, 1);
> > > > - goto repeat;
> > > > - }
> > > > - if (unlikely(!PageUptodate(page))) {
> > > > + if (unlikely(page->mapping != mapping || !PageUptodate(page))) {
> > > > f2fs_put_page(page, 1);
> > > > return ERR_PTR(-EIO);
> > > > }
Powered by blists - more mailing lists