If bio_add_folio() fails (because it is full),
erofs_fileio_scan_folio() needs to submit the I/O request via
erofs_fileio_rq_submit() and allocate a new I/O request with an empty
`struct bio`.  Then it retries the bio_add_folio() call.

However, at this point, erofs_onlinefolio_split() has already been
called which increments `folio->private`; the retry will call
erofs_onlinefolio_split() again, but there will never be a matching
erofs_onlinefolio_end() call.  This leaves the folio locked forever
and all waiters will be stuck in folio_wait_bit_common().

This bug has been added by commit ce63cb62d794 ("erofs: support
unencoded inodes for fileio"), but was practically unreachable because
there was room for 256 folios in the `struct bio` - until commit
9f74ae8c9ac9 ("erofs: shorten bvecs[] for file-backed mounts") which
reduced the array capacity to 16 folios.

It was now trivial to trigger the bug by manually invoking readahead
from userspace, e.g.:

 posix_fadvise(fd, 0, st.st_size, POSIX_FADV_WILLNEED);

This should be fixed by invoking erofs_onlinefolio_split() only after
bio_add_folio() has succeeded.  This is safe: asynchronous completions
invoking erofs_onlinefolio_end() will not unlock the folio because
erofs_fileio_scan_folio() is still holding a reference to be released
by erofs_onlinefolio_end() at the end.

Fixes: ce63cb62d794 ("erofs: support unencoded inodes for fileio")
Fixes: 9f74ae8c9ac9 ("erofs: shorten bvecs[] for file-backed mounts")
Cc: stable@vger.kernel.org
Signed-off-by: Max Kellermann <max.kellermann@ionos.com>
---
 fs/erofs/fileio.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/fs/erofs/fileio.c b/fs/erofs/fileio.c
index 4fa0a0121288..60c7cc4c105c 100644
--- a/fs/erofs/fileio.c
+++ b/fs/erofs/fileio.c
@@ -150,10 +150,10 @@ static int erofs_fileio_scan_folio(struct erofs_fileio *io, struct folio *folio)
 				io->rq->bio.bi_iter.bi_sector = io->dev.m_pa >> 9;
 				attached = 0;
 			}
-			if (!attached++)
-				erofs_onlinefolio_split(folio);
 			if (!bio_add_folio(&io->rq->bio, folio, len, cur))
 				goto io_retry;
+			if (!attached++)
+				erofs_onlinefolio_split(folio);
 			io->dev.m_pa += len;
 		}
 		cur += len;
-- 
2.47.2

On Tue, Apr 29, 2025 at 01:09:33AM +0200, Max Kellermann wrote:
> If bio_add_folio() fails (because it is full),
> erofs_fileio_scan_folio() needs to submit the I/O request via
> erofs_fileio_rq_submit() and allocate a new I/O request with an empty
> `struct bio`.  Then it retries the bio_add_folio() call.
> 
> However, at this point, erofs_onlinefolio_split() has already been
> called which increments `folio->private`; the retry will call
> erofs_onlinefolio_split() again, but there will never be a matching
> erofs_onlinefolio_end() call.  This leaves the folio locked forever
> and all waiters will be stuck in folio_wait_bit_common().
> 
> This bug has been added by commit ce63cb62d794 ("erofs: support
> unencoded inodes for fileio"), but was practically unreachable because
> there was room for 256 folios in the `struct bio` - until commit
> 9f74ae8c9ac9 ("erofs: shorten bvecs[] for file-backed mounts") which
> reduced the array capacity to 16 folios.
> 
> It was now trivial to trigger the bug by manually invoking readahead
> from userspace, e.g.:
> 
>  posix_fadvise(fd, 0, st.st_size, POSIX_FADV_WILLNEED);
> 
> This should be fixed by invoking erofs_onlinefolio_split() only after
> bio_add_folio() has succeeded.  This is safe: asynchronous completions
> invoking erofs_onlinefolio_end() will not unlock the folio because
> erofs_fileio_scan_folio() is still holding a reference to be released
> by erofs_onlinefolio_end() at the end.
> 
> Fixes: ce63cb62d794 ("erofs: support unencoded inodes for fileio")
> Fixes: 9f74ae8c9ac9 ("erofs: shorten bvecs[] for file-backed mounts")
> Cc: stable@vger.kernel.org
> Signed-off-by: Max Kellermann <max.kellermann@ionos.com>

Thanks for catching this! LGTM:
Reviewed-by: Gao Xiang <xiang@kernel.org>

Thanks,
Gao Xiang