powerpc: Use enhanced touch instructions in POWER7 copy_to_user/copy_from_user
authorAnton Blanchard <anton@samba.org>
Mon, 28 May 2012 22:14:32 +0000 (22:14 +0000)
committerBenjamin Herrenschmidt <benh@kernel.crashing.org>
Tue, 3 Jul 2012 04:14:42 +0000 (14:14 +1000)
commita9514dc69d5c4f5d6d9e4b8eed40172abd150c61
tree5f71e45a5dc49a40cf2172964888db064183b090
parente250d4bca6cb91471e0757179a152c0943ecce4a
powerpc: Use enhanced touch instructions in POWER7 copy_to_user/copy_from_user

Version 2.06 of the POWER ISA introduced enhanced touch instructions,
allowing us to specify a number of attributes including the length of
a stream.

This patch adds a software stream for both loads and stores in the
POWER7 copy_tofrom_user loop. Since the setup is quite complicated
and we have to use an eieio to ensure correct ordering of the "GO"
command we only do this for copies above 4kB.

To quantify any performance improvements we need a working set
bigger than the caches so we operate on a 1GB file:

# dd if=/dev/zero of=/tmp/foo bs=1M count=1024

And we compare how fast we can read the file:

# dd if=/tmp/foo of=/dev/null bs=1M

before: 7.7 GB/s
after:  9.6 GB/s

A 25% improvement.

The worst case for this patch will be a completely L1 cache contained
copy of just over 4kB. We can test this with the copy_to_user
testcase we used to tune copy_tofrom_user originally:

http://ozlabs.org/~anton/junkcode/copy_to_user.c

# time ./copy_to_user2 -l 4224 -i 10000000

before: 6.807 s
after:  6.946 s

A 2% slowdown, which seems reasonable considering our data is unlikely
to be completely L1 contained.

Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
arch/powerpc/lib/copyuser_power7.S