Right after a vPE is made resident, the code starts polling the
GICR_VPENDBASER.Dirty bit until it becomes 0, where the delay_us
is set to 10. But in our measurement, it takes only hundreds of
nanoseconds, or 1~2 microseconds, to finish parsing the VPT in most
cases. What's more, we found that the MMIO delay on GICv4.1 system
(HiSilicon) is about 10 times higher than that on GICv4.0 system in
kvm-unit-tests (the specific data is as follows).
| GICv4.1 emulator | GICv4.0 emulator
mmio_read_user (ns) | 12811 | 1598
After analysis, this is mainly caused by the 10 delay_us, so it might
really hurt performance.
To avoid this, we can set the delay_us to 1, which is more appropriate
in this situation and universal. Besides, we can delay the execution
of the polling, giving the GIC a chance to work in parallel with the CPU
on the entry path.
Shenming Lu (2):
irqchip/gic-v4.1: Reduce the delay time of the poll on the
GICR_VPENDBASER.Dirty bit
KVM: arm64: Delay the execution of the polling on the
GICR_VPENDBASER.Dirty bit
arch/arm64/kvm/vgic/vgic-v4.c | 16 ++++++++++++++++
arch/arm64/kvm/vgic/vgic.c | 3 +++
drivers/irqchip/irq-gic-v3-its.c | 18 +++++++++++++-----
drivers/irqchip/irq-gic-v4.c | 11 +++++++++++
include/kvm/arm_vgic.h | 3 +++
include/linux/irqchip/arm-gic-v4.h | 4 ++++
6 files changed, 50 insertions(+), 5 deletions(-)
--
2.23.0
