[PATCH v2 0/3] x86: Prefer MWAIT over HALT on AMD processors

From: Wyes Karny
Date: Thu May 05 2022 - 06:49:27 EST


Currently in the absence of the cpuidle driver (eg: when global
C-States are disabled in the BIOS or when cpuidle is driver is not
compiled in), the default idle state on AMD Zen processors uses the
HALT instruction even though there is support for MWAIT instruction
which is more efficient than HALT.

HPC customers who want to optimize for lower latency are known to
disable Global C-States in the BIOS. In this scenario, the cpuidle driver
will not be loaded and the kernel will continue with the default idle state
chosen at boot time. On AMD systems currently the default idle state
is HALT which has a higher exit latency compared to MWAIT.

The reason for this is

1. Families prior to 10h didn't support MWAIT
2. Families 10h-15h supported MWAIT, but not MWAIT C1. Hence it was
preferable to use HALT as the default state on these systems.

However, AMD Family 17h onwards supports MWAIT as well as MWAIT
C1. And it is preferable to use MWAIT as the default idle state on
these systems, as it has lower exit latencies.

The below table represents the exit latency for HALT and MWAIT on AMD
Zen 3 system.
Exit latency is measured by issuing a wakeup (IPI) to other
CPU and measuring how many clock cycles it took to wakeup.
Each iteration measures 10K wakeups by pinning source and
destination.

HALT:

25.0000th percentile : 1900 ns
50.0000th percentile : 2000 ns
75.0000th percentile : 2300 ns
90.0000th percentile : 2500 ns
95.0000th percentile : 2600 ns
99.0000th percentile : 2800 ns
99.5000th percentile : 3000 ns
99.9000th percentile : 3400 ns
99.9500th percentile : 3600 ns
99.9900th percentile : 5900 ns
Min latency : 1700 ns
Max latency : 5900 ns
Total Samples 9999

MWAIT:

25.0000th percentile : 1400 ns
50.0000th percentile : 1500 ns
75.0000th percentile : 1700 ns
90.0000th percentile : 1800 ns
95.0000th percentile : 1900 ns
99.0000th percentile : 2300 ns
99.5000th percentile : 2500 ns
99.9000th percentile : 3200 ns
99.9500th percentile : 3500 ns
99.9900th percentile : 4600 ns
Min latency : 1200 ns
Max latency : 4600 ns
Total Samples 9997

Improvement (99th percentile): 21.74%

Below is another result for context_switch2 micro-benchmark,
which brings out the impact of improved wakeup latency through
increased context-switches per second.

Link: https://ozlabs.org/~anton/junkcode/context_switch2.c

with HALT:
-------------------------------
50.0000th percentile : 190184
75.0000th percentile : 191032
90.0000th percentile : 192314
95.0000th percentile : 192520
99.0000th percentile : 192844
MIN : 190148
MAX : 192852

with MWAIT:
-------------------------------
50.0000th percentile : 277444
75.0000th percentile : 278268
90.0000th percentile : 278888
95.0000th percentile : 279164
99.0000th percentile : 280504
MIN : 273278
MAX : 281410


Improvement(99th percentile): ~ 45.46%

A similar trend is observed on older Zen processors also.

Here we enable MWAIT instruction as the default idle call for AMD
Zen processors which support MWAIT. We retain the existing behaviour
for older processors which depend on HALT.

This patchset restores the decision tree that was present in the kernel
earlier due to Thomas Gleixner's patch:
commit 09fd4b4ef5bc ("x86: use cpuid to check MWAIT support for C1")

NOTE: This change only impacts the default idle behaviour in the
absence of cpuidle driver. If the cpuidle driver is present, it
controls the processor idle behaviour.

Fixes: commit b253149b843f ("sched/idle/x86: Restore mwait_idle() to fix boot hangs, to improve power savings and to improve performance")

Changelog:
v2:
- Remove vendor checks, fix idle=nomwait condition, fix documentation

Wyes Karny (3):
x86: Use HALT in default_idle when idle=nomwait cmdline arg is passed
x86: Remove vendor checks from prefer_mwait_c1_over_halt
x86: Fix comment for X86_FEATURE_ZEN

arch/x86/include/asm/cpufeatures.h | 2 +-
arch/x86/include/asm/mwait.h | 1 +
arch/x86/kernel/process.c | 39 ++++++++++++++++++++++--------
3 files changed, 31 insertions(+), 11 deletions(-)

--
2.27.0