On Tue, Feb 02, 2010 at 07:56:36AM -0800, Linus Torvalds wrote:
> On Tue, 2 Feb 2010, Matthew Wilcox wrote:
> >
> > How about doing this:
> >
> > struct qstr {
> > - const unsigned char *name;
> > + const unsigned char name[0];
> > }
> >
> > struct dentry {
> > - struct qstr d_name;
> > + struct qstr *d_name;
> > - unsigned char d_iname[DNAME_INLINE_LEN_MIN]; /* small names */
> > + union {
> > + struct qstr d_iname;
> > + char pad[DNAME_INLINE_LEN_MIN];
> > + };
> > }
> >
> > Doesn't increase the size of struct dentry, and puts the hash and len
> > with the name. Increases long name allocations by 8 bytes each.
>
> Conceptually nice, but in practice that's absolutely horrible.
>
> Why? Because now the dentry lookup logic has to follow an additional
> pointer just to verify the hash and the length of the name. That's some of
> the hottest code we have, and the _last_ thing we want is another pointer
> dereference and cache access in the path that looks up the dentry hash
> chains.
I'd thought it was in the same cacheline ... but that's not generally
true since dentries are so large. And the d_name is at the other end
of the dentry, so they're guaranteed to be in different cachelines. Bum.
--
Matthew Wilcox Intel Open Source Technology Centre
"Bill, look, we understand that you're interested in selling us this
operating system, but compare it to ours. We can't possibly take such
a retrograde step."
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
