Re: [git pull] d_revalidate pile

[Al Viro <viro@zeniv.linux.org.uk>]

On Mon, Jan 27, 2025 at 05:27:08PM -0800, Linus Torvalds wrote:
> On Mon, 27 Jan 2025 at 17:21, Al Viro <viro@zeniv.linux.org.uk> wrote:
> >
> > Umm...  On some architectures it does, but mostly that's the ones
> > where unaligned word loads are costly.  Which target do you have
> > in mind?
> 
> I was more thinking that we could just make the fallback case be a 'memcmp()'.
> 
> It's not like this particular place matters - as you say, that
> byte-at-a-time code is only used on architectures that don't enable
> the dcache word-at-a-time code (that requires the special "do loads
> that can fault" zeropad helper), but we've had some other places where
> we'd worry about the string length.
> 
> Look at d_path() for another example. That copy_from_kernel_nofault()
> in prepend_copy()...

Hmm...  So something like

/*
 * Returns a pointer to name and a length; length might be
 * inaccurate in case of race with dentry renaming, but
 * it will not exceed the distance from returned pointer
 * to the end of containing object.
 * Caller MUST hold rcu_read_lock().
 * Caller MUST NOT expect the contents of name to remain
 * stable - it can change at any time.
 */
const char *__d_name_rcu(struct dentry *dentry, int *p)
{
	const char *name = smp_load_acquire(&dentry->d_name.name);

	if (unlikely(name != &dentry->d_shortname.string))
		*p = container_of(name, struct external_name, name)->len;
	else if (unlikely((*p = dentry->d_name.name) >= DNAME_INLINE_LEN)
		*p = DNAME_INLINE_LEN - 1;
	return name;
}

with very limited accessibility (basically, dcache.c and d_path.c)

prepend_name() might be able to use that...