[PATCH v8 0/3] exfat: speed up file creation in large directories

From: Yang Wen

Date: Wed Oct 14 2026 - 11:31:59 EST


Negative name lookups and empty-entry allocation can repeatedly scan a
directory from the beginning. Bulk creation in a large directory therefore
approaches O(N^2).

This series separates the optimization into three independently reviewable
steps. Patch 1 adds the Bloom filter used to reject definite name misses.
Patch 2 retains and correctly invalidates the next-empty-entry hint.
Patch 3 adds the LRU and shrinker used to reclaim filters under memory
pressure.

Test environment:
QEMU TCG multi-thread, 4 vCPUs, 6 GiB RAM
4 GiB exFAT image, 32 KiB clusters

Previous baseline Patched v8
real 589.48 s 30.27 s
user 4.72 s 3.40 s
sys 584.63 s 26.83 s

The unpatched baseline is a previously recorded result and was not rerun.
The patched run created 20,000 files; fsck.exfat reported a clean image.

Changes in v8:
- Rebase the series onto exfat/dev at 8e431c3bce0b.
- Drop unrelated blank-line changes from the Bloom-filter patch.
- Discard empty-entry hints collected by positive name lookups, since a
lookup can start at hint_stat and return before scanning the directory
prefix.
- Track the remaining root-directory chain length during volume-label scans
so their empty-entry hints remain valid across cluster boundaries.
- Remove a reclaimed filter from the LRU before clearing the inode pointer,
so the lockless free fast path cannot observe a linked but pointerless
filter during teardown.

Changes in v7:
- Size each Bloom filter from the allocated directory size, using four bits
per on-disk entry and clamping the bitmap between 512 bytes and 4 MiB.
- Track set bits and discard a filter when one third of its bits are set or
directory growth requires a larger bitmap. Rebuild it lazily after a
later complete negative lookup.
- Account reclaimable filters in pages so shrinker accounting follows their
dynamic allocation sizes.
- Preserve the shrinker core's scan budget and return SHRINK_STOP when the
LRU is already empty, preventing repeated no-progress scans.

Changes in v6:
- Invalidate the name filter when committing a new entry set fails, before
attempting rollback. Despite the reported write error, the new name may
already have reached disk.
- Move rename/move error invalidation into the Bloom filter patch so that
each patch remains correct when applied independently.

Changes in v5:
- Rebase the series onto the exFAT maintainer's dev branch.
- Treat every non-negative exfat_find_empty_entry() return value as a
successful allocation in the volume-label path.
- Record the minimum entry-set size for which a saved empty-entry hint is
valid. A shorter entry set now rescans from the beginning and can reuse a
smaller hole that an earlier, longer entry set could not use.

Changes in v4:
- Publish the next-empty-entry hint only after the directory entry set is
successfully committed, so post-allocation failures cannot skip an unused
slot.
- Invalidate the destination name filter when rename or move fails because
the new entry may already exist on disk.

Changes in v3:
- Split the change into Bloom filter, empty-entry hint, and shrinker
patches.
- Accept filenames containing exactly 255 UTF-16 code units while building
the Bloom filter.
- Invalidate the empty-entry hint in every path that can free directory
entries, preventing stale hints from skipping earlier holes.

Changes in v2:
- Move exfat_name_filter_free() to exfat_evict_inode() because
->free_inode() may run from an RCU callback in softirq context.

Yang Wen (3):
exfat: add a Bloom filter for negative name lookups
exfat: retain the next empty directory entry hint
exfat: reclaim name filters under memory pressure

fs/exfat/dir.c | 363 +++++++++++++++++++++++++++++++++++++++++++-
fs/exfat/exfat_fs.h | 30 +++-
fs/exfat/inode.c | 1 +
fs/exfat/namei.c | 87 ++++++++++-
fs/exfat/super.c | 9 ++
5 files changed, 478 insertions(+), 12 deletions(-)

base-commit: 8e431c3bce0bb51802fc79b10225cc90a15f15d1
--
2.34.1