Re: [PATCH] dcache: Translating dentry into pathname without taking rename_lock

[Linus Torvalds <torvalds@linux-foundation.org>]

On Wed, Sep 4, 2013 at 6:49 PM, Waiman Long <waiman.long@hp.com> wrote:
>
> So what I am going to do is to use memchr() to locate a null
> byte within the given length. If one is found, the string must be invalid
> and there is no point in doing the copying. Instead, EINVAL will be returned
> and the code will check the sequence number then. The memchr() function can
> be fast if an architecture-specific version exists.

We don't have an architecture-specific fast version of memchr, because
nobody sane has ever cared about that abomination of a function. Also,
if rename() overwrites the pathname (switching inline names), I guess
a zero could disappear between your memchr and the copy. Although I
think we always have an ending NUL byte for the inline case, so that
should make sure that memchr would find it *eventually*.

But regardless, that's really not what you want to do.

You should do:

 - use the name length as a maximum

 - do a byte-at-a-time copy, stopping at a zero (it's going to be
faster than memchr anyway)

Then, later on, we can do one that does a word-at-a-time using the
CONFIG_DCACHE_WORD_ACCESS magic: we know dentry names are always
word-aligned, and we have an efficient "has_zero()" function for
finding zero bytes in a word.

Really.

            Linus