]> git.kernelconcepts.de Git - karo-tx-linux.git/commit
powerpc: Fix accounting of softirq time when idle
authorAnton Blanchard <anton@samba.org>
Sun, 20 Mar 2011 15:28:03 +0000 (15:28 +0000)
committerGreg Kroah-Hartman <gregkh@suse.de>
Thu, 14 Apr 2011 20:02:05 +0000 (13:02 -0700)
commit81dacc2d64030457bfd85ba901deb17a682e6196
tree1f8e1e01d4681aba1b9fb57630340bf8644e7682
parentf7b5886892a642bfdc39619a63794d17ec333896
powerpc: Fix accounting of softirq time when idle

commit ad5d1c888e556bc00c4e86f452cad4a3a87d22c1 upstream.

commit cf9efce0ce31 (powerpc: Account time using timebase rather
than PURR) used in_irq() to detect if the time was spent in
interrupt processing. This only catches hardirq context so if we
are in softirq context and in the idle loop we end up accounting it
as idle time. If we instead use in_interrupt() we catch both softirq
and hardirq time.

The issue was found when running a network intensive workload. top
showed the following:

0.0%us,  1.1%sy,  0.0%ni, 85.7%id,  0.0%wa,  9.9%hi,  3.3%si,  0.0%st

85.7% idle. But this was wildly different to the perf events data.
To confirm the suspicion I ran something to keep the core busy:

# yes > /dev/null &

8.2%us,  0.0%sy,  0.0%ni,  0.0%id,  0.0%wa, 10.3%hi, 81.4%si,  0.0%st

We only got 8.2% of the CPU for the userspace task and softirq has
shot up to 81.4%.

With the patch below top shows the correct stats:

0.0%us,  0.0%sy,  0.0%ni,  5.3%id,  0.0%wa, 13.3%hi, 81.3%si,  0.0%st

Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
arch/powerpc/kernel/time.c