irqchip/gic-v4.1: Optimize the delay time of the poll on the GICR_VENPENDBASER.Dirty bit
Subsystems (2)
PATCH 1/1 irqchip/gic-v4.1: Optimize the delay time of the poll on the GICR_VENPENDBASER.Dirty bit
Every time the vPE is scheduled, the code starts polling the GICR_VPENDBASER.Dirty bit until it becomes 0. It's OK. But the delay_us of the poll function is directly set to 10, which is a long time for most interrupts. In our measurement, it takes only 1~2 microseconds, or even hundreds of nanoseconds, to finish parsing the VPT in most cases. However, in the current implementation, if the GICR_VPENDBASER.Dirty bit is not 0 immediately after the vPE is scheduled, it will directly wait for 10 microseconds, resulting in meaningless waiting.
In order to avoid meaningless waiting, we can set the delay_us to 0, which can exit the poll function immediately when the Dirty bit becomes 0.
--- drivers/irqchip/irq-gic-v3-its.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/irqchip/irq-gic-v3-its.c +1 -1
@@ -3803,7 +3803,7 @@ static void its_wait_vpt_parse_complete(void)WARN_ON_ONCE (readq_relaxed_poll_timeout_atomic (vlpi_base +GICR_VPENDBASER , val, !(val & GICR_VPENDBASER_Dirty), -10 ,500 )); +0 ,500 )); }static void its_vpe_schedule (struct its_vpe *vpe)
Signature · 1 line
2.19.1
Marc Zyngier
maintainerQuoted 10 lines · Shenming Lu
Every time the vPE is scheduled, the code starts polling the GICR_VPENDBASER.Dirty bit until it becomes 0. It's OK. But the
delay_usof the poll function is directly set to 10, which is a long time for most interrupts. In our measurement, it takes only 1~2 microseconds, or even hundreds of nanoseconds, to finish parsing the VPT in most cases. However, in the current implementation, if the GICR_VPENDBASER.Dirty bit is not 0 immediately after the vPE is scheduled, it will directly wait for 10 microseconds, resulting in meaningless waiting.
Quoted 3 lines
In order to avoid meaningless waiting, we can set the
delay_usto 0, which can exit the poll function immediately when the Dirty bit becomes 0.
We clearly have a difference in interpretation of the word "meaningless".
With this, you are busy-waiting on the register, adding even more overhead at the RD level. How is that better? The whole point is to back off and let the RD do its stuff in the background. This is also based on a massive sample of *one* implementation. How is that representative?
It would be a lot more convincing if you measured the difference it makes on the total scheduling latency of a vcpu. Assuming it makes *any* observable difference.
Thanks,
M.
Quoted 20 lines
Signed-off-by Shenming Lu <[email protected]>--- drivers/irqchip/irq-gic-v3-its.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)drivers/irqchip/irq-gic-v3-its.c +1
-1@@ -3803,7 +3803,7 @@ static void its_wait_vpt_parse_complete(void)WARN_ON_ONCE (readq_relaxed_poll_timeout_atomic (vlpi_base +GICR_VPENDBASER , val, !(val & GICR_VPENDBASER_Dirty), -10 ,500 )); +0 ,500 )); }static void
its_vpe_schedule(struct its_vpe *vpe)
Signature · 1 line
Jazz is not dead. It just smells funny...