Re: [PATCH v2] cpufreq, powernv: Disable preemption while checking CPU throttling state throttling state

On Tue, Nov 22, 2016 at 2:07 PM, Denis Kirjanov [off-list ref] wrote:

With preemption turned on we can read incorrect throttling state while being switched
to CPU on a different chip.
The following BUG_() was hit while running the 4.8-rc5 kernel compiled
with CONFIG_PREEMPT_DEBUG on a POWER machine.


    [   67.700897] BUG: using smp_processor_id() in preemptible [00000000] code: cat/7343
    [   67.700988] caller is .powernv_cpufreq_throttle_check+0x2c/0x710
    [   67.700998] CPU: 13 PID: 7343 Comm: cat Not tainted 4.8.0-rc5-dirty #1
    [   67.701038] Call Trace:
    [   67.701066] [c0000007d25b75b0] [c000000000971378] .dump_stack+0xe4/0x150 (unreliable)
    [   67.701153] [c0000007d25b7640] [c0000000005162e4] .check_preemption_disabled+0x134/0x150
    [   67.701238] [c0000007d25b76e0] [c0000000007b63ac] .powernv_cpufreq_throttle_check+0x2c/0x710
    [   67.701322] [c0000007d25b7790] [c0000000007b6d18] .powernv_cpufreq_target_index+0x288/0x360
    [   67.701407] [c0000007d25b7870] [c0000000007acee4] .__cpufreq_driver_target+0x394/0x8c0
    [   67.701491] [c0000007d25b7920] [c0000000007b22ac] .cpufreq_set+0x7c/0xd0
    [   67.701565] [c0000007d25b79b0] [c0000000007adf50] .store_scaling_setspeed+0x80/0xc0
    [   67.701650] [c0000007d25b7a40] [c0000000007ae270] .store+0xa0/0x100
    [   67.701723] [c0000007d25b7ae0] [c0000000003566e8] .sysfs_kf_write+0x88/0xb0
    [   67.701796] [c0000007d25b7b70] [c0000000003553b8] .kernfs_fop_write+0x178/0x260
    [   67.701881] [c0000007d25b7c10] [c0000000002ac3cc] .__vfs_write+0x3c/0x1c0
    [   67.701954] [c0000007d25b7cf0] [c0000000002ad584] .vfs_write+0xc4/0x230
    [   67.702027] [c0000007d25b7d90] [c0000000002aeef8] .SyS_write+0x58/0x100
    [   67.702101] [c0000007d25b7e30] [c00000000000bfec] system_call+0x38/0xfc

The task which updates the cpu-frequency checks if the chip was
throttled. However, it does so without protecting itself against
preemption. Hence, if the task is preempted and woken up on a CPU on
another chip, it can end up updating the throttle statistics of the
incorrect chip.

This patch fixes the issue by ensuring that the throttle_check is
performed with preemption disabled.

Fixes: 09a972d16209 ("cpufreq: powernv: Report cpu frequency throttling")
Reviewed-by: Gautham R. Shenoy <redacted>
Signed-off-by: Denis Kirjanov <redacted>

That's already queued up for v4.10, you don't have to resend it.

Thanks,
Rafael