On Mon, Jul 25, 2011 at 8:23 PM, Linus Torvalds
<torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
>
> What I *did* consider was to have the bits just be "slow case" bits.
> And just set them all (blindly) at inode allocation time. Then, the
> users would simply end up doing [ ... ]
I ended up with this, after all. Except rather than be "slow case"
bits, they are "fast case" bits. The code seemed more natural that
way: clear the bits initially, and then set a bit for the default case
we care about.
And I use inode->i_lock to protect the setting, to avoid the need for
atomic ops on an "unsigned short". It's going to happen once per bit
and per inode lifetime, it didn't seem worth it to try to be clever
about it.
I considered using i_state, which is already a proper "unsigned long"
exactly so that the bitops will work on it. But since the aim was to
be dense in the D$, it seemed more natural to use the 16-bit slot we
have next to i_mode.
My timings still showed roughly a 2% improvement (well 1.8% on the
"repeat ten times, do the average") but I didn't really do a lot of
deviation testing. The noise in the numbers is quite noticeable
between reboots, but I've seen that 1-2% several times over various
different implementations of the same "try to avoid touching
inode->i_op->xyz", so I'm pretty convinced it's not just random noise.
Linus
--
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
