Message ID | 20170611235835.7400-11-npiggin@gmail.com (mailing list archive) |
---|---|
State | Superseded |
Headers | show |
On Mon, Jun 12, 2017 at 09:58:31AM +1000, Nicholas Piggin wrote: > local_irq_enable can cause interrupts to be taken which could > take significant amount of processing time. The idle process > should set its polling flag before this, so another process that > wakes it during this time will not have to send an IPI. > > Expand the TIF_POLLING_NRFLAG coverage to as large as possible. > > Signed-off-by: Nicholas Piggin <npiggin@gmail.com> Looks good. Were you able to see this make a difference in any of the tests ? Reviewed-by: Gautham R. Shenoy <ego@linux.vnet.ibm.com> > --- > drivers/cpuidle/cpuidle-powernv.c | 4 +++- > drivers/cpuidle/cpuidle-pseries.c | 3 ++- > 2 files changed, 5 insertions(+), 2 deletions(-) > > diff --git a/drivers/cpuidle/cpuidle-powernv.c b/drivers/cpuidle/cpuidle-powernv.c > index 79152676f62b..50b3c2e0306f 100644 > --- a/drivers/cpuidle/cpuidle-powernv.c > +++ b/drivers/cpuidle/cpuidle-powernv.c > @@ -51,9 +51,10 @@ static int snooze_loop(struct cpuidle_device *dev, > { > u64 snooze_exit_time; > > - local_irq_enable(); > set_thread_flag(TIF_POLLING_NRFLAG); > > + local_irq_enable(); > + > snooze_exit_time = get_tb() + snooze_timeout; > ppc64_runlatch_off(); > HMT_very_low(); > @@ -66,6 +67,7 @@ static int snooze_loop(struct cpuidle_device *dev, > ppc64_runlatch_on(); > clear_thread_flag(TIF_POLLING_NRFLAG); > smp_mb(); > + > return index; > } > > diff --git a/drivers/cpuidle/cpuidle-pseries.c b/drivers/cpuidle/cpuidle-pseries.c > index 166ccd711ec9..7b12bb2ea70f 100644 > --- a/drivers/cpuidle/cpuidle-pseries.c > +++ b/drivers/cpuidle/cpuidle-pseries.c > @@ -62,9 +62,10 @@ static int snooze_loop(struct cpuidle_device *dev, > unsigned long in_purr; > u64 snooze_exit_time; > > + set_thread_flag(TIF_POLLING_NRFLAG); > + > idle_loop_prolog(&in_purr); > local_irq_enable(); > - set_thread_flag(TIF_POLLING_NRFLAG); > snooze_exit_time = get_tb() + snooze_timeout; > > while (!need_resched()) { > -- > 2.11.0 >
On Mon, 12 Jun 2017 20:40:25 +0530 Gautham R Shenoy <ego@linux.vnet.ibm.com> wrote: > On Mon, Jun 12, 2017 at 09:58:31AM +1000, Nicholas Piggin wrote: > > local_irq_enable can cause interrupts to be taken which could > > take significant amount of processing time. The idle process > > should set its polling flag before this, so another process that > > wakes it during this time will not have to send an IPI. > > > > Expand the TIF_POLLING_NRFLAG coverage to as large as possible. > > > > Signed-off-by: Nicholas Piggin <npiggin@gmail.com> > > Looks good. Were you able to see this make a difference in any of the > tests ? No I didn't measure a difference or have a test case where this was noticable. I think on a workload with some IO interrupts as well as cross-CPU wakeups, then statistically we should see some improvement in IPI rates with this patch. Thanks, Nick
diff --git a/drivers/cpuidle/cpuidle-powernv.c b/drivers/cpuidle/cpuidle-powernv.c index 79152676f62b..50b3c2e0306f 100644 --- a/drivers/cpuidle/cpuidle-powernv.c +++ b/drivers/cpuidle/cpuidle-powernv.c @@ -51,9 +51,10 @@ static int snooze_loop(struct cpuidle_device *dev, { u64 snooze_exit_time; - local_irq_enable(); set_thread_flag(TIF_POLLING_NRFLAG); + local_irq_enable(); + snooze_exit_time = get_tb() + snooze_timeout; ppc64_runlatch_off(); HMT_very_low(); @@ -66,6 +67,7 @@ static int snooze_loop(struct cpuidle_device *dev, ppc64_runlatch_on(); clear_thread_flag(TIF_POLLING_NRFLAG); smp_mb(); + return index; } diff --git a/drivers/cpuidle/cpuidle-pseries.c b/drivers/cpuidle/cpuidle-pseries.c index 166ccd711ec9..7b12bb2ea70f 100644 --- a/drivers/cpuidle/cpuidle-pseries.c +++ b/drivers/cpuidle/cpuidle-pseries.c @@ -62,9 +62,10 @@ static int snooze_loop(struct cpuidle_device *dev, unsigned long in_purr; u64 snooze_exit_time; + set_thread_flag(TIF_POLLING_NRFLAG); + idle_loop_prolog(&in_purr); local_irq_enable(); - set_thread_flag(TIF_POLLING_NRFLAG); snooze_exit_time = get_tb() + snooze_timeout; while (!need_resched()) {
local_irq_enable can cause interrupts to be taken which could take significant amount of processing time. The idle process should set its polling flag before this, so another process that wakes it during this time will not have to send an IPI. Expand the TIF_POLLING_NRFLAG coverage to as large as possible. Signed-off-by: Nicholas Piggin <npiggin@gmail.com> --- drivers/cpuidle/cpuidle-powernv.c | 4 +++- drivers/cpuidle/cpuidle-pseries.c | 3 ++- 2 files changed, 5 insertions(+), 2 deletions(-)