Re: [PATCH v3 0/4] cifs: follow-on fixes after fscache_resize_cookie() consolidation

On Sat, Aug 1, 2026 at 2:47 AM Frank Sorenson <[email protected]> wrote:
>
> After commit fa724e235cfd ("cifs: add fscache_resize_cookie() to
> cifs_setsize()") consolidated fscache cookie resizing into cifs_setsize(),
> two further fixes are needed and two dead-code blocks became removable.
>
> Patch 1 fixes cifs_do_truncate() (O_TRUNC via cifs_open()).  That path
> runs without i_rwsem, so it cannot use the locked resize path.
> cifs_invalidate_cache() is the correct alternative: it works without the
> lock and ensures stale cached data is not served once the cookie is later
> activated.
>
> Patch 2 addresses a race identified during review: cifs_setsize() calls
> fscache_resize_cookie() without i_rwsem, but another concurrent open may
> already have the cookie active (IS_CACHING set), making fscache_resize_cookie()
> a real operation requiring the lock.  The fix strips fscache_resize_cookie()
> from cifs_setsize() and introduces cifs_resize_file_locked(), which
> temporarily activates the cookie with fscache_use_cookie(), performs the
> resize under i_rwsem, then deactivates it.  Callers that already hold
> i_rwsem (cifs_file_set_size, smb2_duplicate_extents, smb3_simple_falloc)
> switch to the wrapper; cifs_do_truncate() continues to use cifs_setsize()
> followed by cifs_invalidate_cache() as established in patch 1.
>
> Patches 3 and 4 remove dead code.  Patch 3 removes a caller-side
> truncate_setsize() + fscache_resize_cookie() block from
> cifs_remap_file_range() that became redundant once smb2_duplicate_extents()
> started performing the full size update via cifs_setsize() under the
> i_rwsem held by lock_two_nondirectories().  Patch 4 removes equivalent
> dead blocks from cifs_setattr_unix() and cifs_setattr_nounix(): since
> cifs_file_set_size() calls cifs_setsize() on success, i_size always equals
> attrs->ia_size on the success path, making the subsequent size-inequality
> blocks unreachable.
>
> v3: Fixed broken patch
> v2:
>   - Added patch 2 (cifs_resize_file_locked): based on review feedback
>     (Huiwen He) that the fscache cookie is not guaranteed to be quiescent
>     in cifs_do_truncate() - another concurrent open may already have it
>     active.  The fix strips fscache_resize_cookie() from cifs_setsize() and
>     adds a locked wrapper for callers that hold i_rwsem.
>   - Patch 1 commit message updated: the original rationale (cookie is
>     always quiescent in cifs_do_truncate) was incorrect; the correct reason
>     is that i_rwsem cannot be held in that path.
>   - Patches 2-3 from v1 renumbered to 3-4 to place fixes before cleanups.
>   - Patches 3-4: add Reviewed-by from Huiwen He; otherwise unchanged from v1.
>   - All patches: add Reviewed-by from Paulo Alcantara.
>
> Frank Sorenson (4):
>   cifs: use cifs_invalidate_cache() in cifs_do_truncate() for O_TRUNC
>   cifs: add cifs_resize_file_locked() to guard fscache_resize_cookie()
>     under i_rwsem
>   cifs: remove redundant size-update block in cifs_remap_file_range()
>   cifs: remove dead size-update blocks in cifs_setattr_unix/nounix
I will apply this patch-set to #for-next.
Thanks!