cifs: Release folio lock on fscache read hit.
Subsystems (2)
cover cifs: Release folio lock on fscache read hit.
Here's a patch that fixes what seems to be a showstopping bug in cifs-fscache. *Any time* that a fscache read hit occurs on a mounted cifs filesystem (and therefore cifs_readpage_worker is called), the page lock on the read page is leaked, causing subsequent accesses to the page to deadlock.
I've been running my machine for the past week with this change and it seems to work, but I'm a little concerned that this may mean cifs-fscache has no users (besides me). Up to you (maintainers) whether that means you should accept this patch... or delete the fscache support.
Russell Harmon (1):
cifs: Release folio lock on fscache read hit.
fs/smb/client/file.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
Signature · 1 line
2.34.1
PATCH 1/1 cifs: Release folio lock on fscache read hit.
Under the current code, when cifs_readpage_worker is called, the call contract is that the callee should unlock the page. This is documented in the read_folio section of Documentation/filesystems/vfs.rst as:
Quoted 2 lines
The filesystem should unlock the folio once the read has completed, whether it was successful or not.
Without this change, when fscache is in use and cache hit occurs during a read, the page lock is leaked, producing the following stack on subsequent reads (via mmap) to the page:
$ cat /proc/3890/task/12864/stack [<0>] folio_wait_bit_common+0x124/0x350 [<0>] filemap_read_folio+0xad/0xf0 [<0>] filemap_fault+0x8b1/0xab0 [<0>] __do_fault+0x39/0x150 [<0>] do_fault+0x25c/0x3e0 [<0>] __handle_mm_fault+0x6ca/0xc70 [<0>] handle_mm_fault+0xe9/0x350 [<0>] do_user_addr_fault+0x225/0x6c0 [<0>] exc_page_fault+0x84/0x1b0 [<0>] asm_exc_page_fault+0x27/0x30
This requires a reboot to resolve; it is a deadlock.
Note however that the call to cifs_readpage_from_fscache does mark the page clean, but does not free the folio lock. This happens in __cifs_readpage_from_fscache on success. Releasing the lock at that point however is not appropriate as cifs_readahead also calls cifs_readpage_from_fscache and *does* unconditionally release the lock after its return. This change therefore effectively makes cifs_readpage_worker work like cifs_readahead.
--- fs/smb/client/file.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
fs/smb/client/file.c +1 -1
@@ -4681,9 +4681,9 @@ static int cifs_readpage_worker(struct file *file, struct page *page, io_error:kunmap (page); -unlock_page (page); read_complete: +unlock_page (page);return rc; }
Signature · 1 line
2.34.1