On Sun, 9 Nov 2025 at 14:44, Linus Torvalds
<[email protected]> wrote:
>
> And again: that patch will slow things down [..]
Having done a profile just to see, the regular allocation path
(getname_flags -> kmem_cache_alloc_noprof) certainly shows up.
But on that test set (kernel build), I never hit the 128-byte limit,
and interestingly putname() shows up a tiny bit more than
getname_flags().
At least one reason seems to be that the
if (refcnt != 1) {
thing in putname() is mispredicted, and without the auditing code, the
"refcnt == 1" case is obviously the common case.
Anyway. Not a great test, and this is all the "good behavior", and all
the *real* costs in that particular path are in strncpy_from_user(),
kmem_cache_free() and kmem_cache_alloc_noprof().
And yeah, STAC/CLAC is expensive on my old Zen 2 machine, as are the
cmpxchg16b in the slab alloc/free paths.
And in the end, the actual path walking is more expensive than all of
this, as is selinux_inode_permission(). So the actual filename copying
isn't *really* all that noticeable - you have to look for it.
Linus
