diff options
| author | Christian Brauner <brauner@kernel.org> | 2026-11-05 15:51:17 +0100 |
|---|---|---|
| committer | Christian Brauner <brauner@kernel.org> | 2026-11-05 15:52:42 +0100 |
| commit | 66a52b662e5c09c7e784d2656e0b7ab14b396a8f (patch) | |
| tree | 7a561da87f998abd7699b712ce1cb9fe954d3278 /include/linux/timerqueue.h | |
| parent | 05441d697f7f9a927264f4058cbbc6a03dfcac9d (diff) | |
| parent | 45eb9737058fa78a73f17a5da55c8cb482c6d5bb (diff) | |
Merge patch series "libfs: fix readdir of offset directories"
Christian Brauner <brauner@kernel.org> says:
readdir on tmpfs walks d_children and only uses the offset tree to map
the file position back to a dentry. d_children and offset order
disagree after a rename over a name, an exchange or a wrap of the
offset allocator. An unlink between two getdents64() calls then makes
the walk skip entries or return the whole directory again. So walk the
offset tree itself in descending offset order.
v1 did that with a full tree descent and a reference for every entry
and was quite a bit slower. Now the walk keeps one ma_state for the
whole call and takes no references at all. That relies on all writers
holding i_rwsem exclusive while readdir holds it shared, and the maple
tree only allows that if i_rwsem is the lock of the tree. The first
patch makes it so, and lockdep now checks the writers. The last patch
stops an lseek() past the end from returning the whole directory again.
1M entries on tmpfs, getdents64() with a 32k buffer: 146ms before,
about 230ms with v1, 114ms now. Four processes listing the directory at
the same time: 470ms per listing before, 113ms now.
Tested with Nickolai's reproducer, generic/736 and a seek-after-closedir
test, on tmpfs and on tmpfs exported over NFSv3 and NFSv4.2, on KASAN
and lockdep kernels. I also wrote a TLA+ model of the walk (cf. [1]).
[1]: https://github.com/brauner/linux-formal/tree/main/kernel/libfs/offset-readdir
* patches from https://patch.msgid.link/20261101-work-libfs-offset_readdir-v2-0-c621125b56d1@kernel.org:
libfs: end offset readdir at positions past DIR_OFFSET_EOD
libfs: walk offset directories in offset order
libfs: make i_rwsem the lock of the offset tree
Link: https://patch.msgid.link/20261101-work-libfs-offset_readdir-v2-0-c621125b56d1@kernel.org
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Diffstat (limited to 'include/linux/timerqueue.h')
0 files changed, 0 insertions, 0 deletions
