Re: [PATCH] statx: optimize copy of struct statx to userspace

[David Howells <dhowells@redhat.com>]

Eric Biggers <ebiggers3@gmail.com> wrote:

> From: Eric Biggers <ebiggers@google.com>
> 
> I found that statx() was significantly slower than stat().  As a
> microbenchmark, I compared 10,000,000 invocations of fstat() on a tmpfs
> file to the same with statx() passed a NULL path:
> 
> 	$ time ./stat_benchmark
> 
> 	real	0m1.464s
> 	user	0m0.275s
> 	sys	0m1.187s
> 
> 	$ time ./statx_benchmark
> 
> 	real	0m5.530s
> 	user	0m0.281s
> 	sys	0m5.247s
> 
> statx is expected to be a little slower than stat because struct statx
> is larger than struct stat, but not by *that* much.  It turns out that
> most of the overhead was in copying struct statx to userspace,
> apparently mostly in all the stac/clac instructions that got generated
> for each __put_user() call.  (This was on x86_64, but some other
> architectures, e.g. arm64, have something similar now too.)
> 
> stat() instead initializes its struct on the stack and copies it to
> userspace with a single call to copy_to_user().  This turns out to be
> much faster, and changing statx to do this makes it almost as fast as
> stat:
> 
> 	$ time ./statx_benchmark
> 
> 	real	0m1.573s
> 	user	0m0.229s
> 	sys	0m1.344s
> 
> Signed-off-by: Eric Biggers <ebiggers@google.com>

Acked-by: David Howells <dhowells@redhat.com>