Al Viro a écrit :

On Thu, Nov 27, 2008 at 12:32:59AM +0100, Eric Dumazet wrote:

quoted

This function arms a flag (MNT_SPECIAL) on the vfs, to avoid
refcounting on permanent system vfs.
Use this function for sockets, pipes, anonymous fds.

IMO that's pushing it past the point of usefulness; unless you can show
that this really gives considerable win on pipes et.al. *AND* that it
doesn't hurt other loads...

Well, if this is the last cache line that might be shared, then yes, numbers can talk.
But coming from 10 to 1 instead of 0 is OK I guess

dput() part: again, I want to see what happens on other loads; it's probably
fine (and win is certainly more than from mntput() change), but...  The
thing is, atomic_dec_and_lock() in there is often done on dentries with
d_count > 1 and that's fairly cheap (and doesn't involve contention on
dcache_lock on sane targets).

FWIW, unless there's a really good reason to do alpha atomic_dec_and_lock()
in a special way, I'd try to compare with

        if (atomic_add_unless(&dentry->d_count, -1, 1))
                return;

I dont know, but *reading* d_count before trying to write it is expensive
on modern cpus. Oprofile clearly show that on Intel Core2.

Then, *testing* the flag before doing the atomic_something() has the same
problem. Or we should put flag in a different cache line.

I am lazy (time for a sleep here), maybe we are smart here and use a trick like that already ?

atomic_t atomic_read_with_write_intent(atomic_t *v)
{
        int val = 0;
	/*
	 * No LOCK prefix here, we only give a write intent hint to cpu
	 */
        asm volatile("xaddl %0, %1"
                     : "+r" (val), "+m" (v->counter)
                     : : "memory");
        return val;
}

	if (your flag)
		sod off to special
	spin_lock(&dcache_lock);
	if (atomic_dec_and_test(&dentry->d_count)) {
		spin_unlock(&dcache_lock);
		return;
	}
	the rest as usual