Re: [PATCH v3 1/2] libfs: reduce the number of memory allocations in generic_ci_match

[Eric Biggers <ebiggers@kernel.org>]

On Fri, Jul 04, 2025 at 10:43:57AM +0800, Yuwen Chen wrote:
> During path traversal, the generic_ci_match function may be called
> multiple times. The number of memory allocations and releases
> in it accounts for a relatively high proportion in the flamegraph.
> This patch significantly reduces the number of memory allocations
> in generic_ci_match through pre - allocation.
> 
> Signed-off-by: Yuwen Chen <ywen.chen@foxmail.com>
> ---
>  fs/ext4/namei.c    |  2 +-
>  fs/f2fs/dir.c      |  2 +-
>  fs/libfs.c         | 33 ++++++++++++++++++++++++++++++---
>  include/linux/fs.h |  8 +++++++-
>  4 files changed, 39 insertions(+), 6 deletions(-)
> 

The reason the allocation is needed at all is because generic_ci_match() has to
decrypt the encrypted on-disk filename from the dentry that it's matching
against.  It can't decrypt in-place, since the source buffer is in the pagecache
which must not be modified.  Hence, a separate destination buffer is needed.

Filenames have a maximum length of NAME_MAX, i.e. 255, bytes.

It would be *much* simpler to just allocate that on the stack.

And we almost can.  255 bytes is on the high end of what can be acceptable to
allocate on the stack in the kernel.  However, here it would give a lot of
benefit and would always occur close to the leaves in the call graph.  So the
size is not a barrier here, IMO.

The real problem is, once again, the legacy crypto_skcipher API, which requires
that the source/destination buffers be provided as scatterlists.  In Linux, the
kernel stack can be in the vmalloc area.  Thus, the buffers passed to
crypto_skcipher cannot be stack buffers unless the caller actually is aware of
how to turn a vmalloc'ed buffer into a scatterlist, which is hard to do.  (See
verity_ahash_update() in drivers/md/dm-verity-target.c for an example.)

Fortunately, I'm currently in the process of introducing library APIs that will
supersede these legacy crypto APIs.  They'll be simpler and faster and won't
have these silly limitations like not working on virtual addresses...  I plan to
make fscrypt use the library APIs instead of the legacy crypto API.

It will take some time to land everything, though.  We can consider this
patchset as a workaround in the mean time.  But it's sad to see the legacy
crypto API continue to cause problems and more time be wasted on these problems.

I do wonder if the "turn a vmalloc'ed buffer into a scatterlist" trick that some
code in the kernel uses is something that would be worth adopting for now in
fname_decrypt().  As I mentioned above, it's hard to do (you have to go page by
page), but it's possible.  That would allow immediately moving
generic_ci_match() to use a stack allocation, which would avoid adding all the
complexity of the preallocation that you have in this patchset.

- Eric