The tick is active on idle adaptive-tick CPUs when /dev/cpu_dma_latency is used

From: Thomas Lindroth <thomas.lindroth@gmail.com>
To: linux-pm@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Subject: The tick is active on idle adaptive-tick CPUs when /dev/cpu_dma_latency is used
Date: Fri, 5 Jul 2019 19:22:35 +0200	[thread overview]
Message-ID: <79b247b3-e056-610e-9a07-e685dfdaa6c9@gmail.com> (raw)

On recent kernels the tick remains active on idle adaptive-tick CPUs when a small value is written to /dev/cpu_dma_latency to restrict the highest C-state. Before the idle loop redesign in 4.17 idle CPUs had the tick disabled even when C-state were restricted. Is this change intentional or a regression?

I use an x86_64 system built with CONFIG_NO_HZ_FULL that I recently upgraded to the 4.19 series from the 4.14 series. I noticed that adaptive-tick CPUs (nohz_full=1-7) still fire timer interrupts about 1000 times/s (CONFIG_HZ_1000=y) even when they are mostly idle. Some debugging showed that this only happens when a program is writing to /dev/cpu_dma_latency to restrict C-states. The old 4.14 kernel only have around 10 timer interrupts per second on idle adaptive-tick CPU even when C-states are restricted that way.

I would expect an adaptive-tick CPU to turn off the tick when it has 0 or 1 processes to run and enable the tick for >2 processes. Kernels after 4.17 instead have the tick on when 0 or >2 processes are running and the tick off in the 1 process case. Since the tick is off when a single process is running that workload isn't directly harmed by the change but if the CPU use hyperthreading the constant wakeups on an idle HT sibling will reduce performance on the other sibling.

They way I look for timer interrupts is by comparing the LOC line in /proc/interrupts or using the hrtimer_expire_entry tracepoint when function=tick_sched_timer. Both methods seem to give the same results.

I can reproduce the problem using an i7-4790K CPU with /sys/devices/system/cpu/cpuidle/current_driver:intel_idle. I can also reproduce the problem on an old core2duo laptop with current_driver:acpi_idle but I can't reproduce the problem in a virtual machine where current_driver:none. I also can't reproduce the problem if C-states are restricted using the intel_idle.max_cstate=1 kernel argument instead of /dev/cpu_dma_latency.

The commit that introduced the change is 554c8aa8ec "sched: idle: Select idle state before stopping the tick" in v4.17 and the problem exists at least up to kernel 5.1 using the menu cpuidle governor.