Messages in this thread Patch in this message |  | | Date | Sun, 23 May 2021 15:26:39 +0200 | From | Frederic Weisbecker <> | Subject | Re: [entry] 47b8ff194c: will-it-scale.per_process_ops -3.0% regression |
| |
On Wed, May 19, 2021 at 01:27:00PM +0800, Oliver Sang wrote: > hi, Peter Zijlstra, > > On Fri, May 14, 2021 at 12:13:22PM +0200, Peter Zijlstra wrote: > > On Fri, May 14, 2021 at 01:11:27AM +0200, Frederic Weisbecker wrote: > > > > > That sounds good, but two things: > > > > > > 1) Even if context tracking is not running, we still need to handle > > > deferred wakeups on idle. But all user/guest/idle currently use the > > > same function. > > > > > > 2) Context tracking may be running even when full nohz is not. But here only > > > full nohz is concerned. > > > > > > So the change should rather be as follows (completely untested!). > > > I rather put the static key check in tick.h in order not to involve > > > deep dependencies inside rcupdate.h (especially rcupdate.h -> tick.h -> sched.h) > > > > Compiles and boots for me, 0day folks, could you please test this makes > > the reported regression go away? > > sorry for late. > > on which base should we apply below patch? > we tried to apply upon 4ae7dc97f726e ("entry/kvm: Explicitly flush pending rcuog > wakeup before last rescheduling point") and v5.12, but both build failed like below: > > ERROR: modpost: "tick_nohz_full_mask" [arch/x86/kvm/kvm.ko] undefined! > > or is there any particular kconfig need to enable? > (the kconfig which we used to build 47b8ff194c is in original report > which maybe you could help have a look what we may miss)
Ah indeed I need to export tick_nohz_full_mask. Too bad but that's the fate of many cpu masks anyway. Can you try the following patch instead? It boots and seem to behave:
--- From: Frederic Weisbecker <frederic@kernel.org> Date: Sun, 23 May 2021 14:23:30 +0200 Subject: [PATCH] tick/nohz: Only check for RCU deferred wakeup on user/guest entry when needed
Checking for and processing RCU-nocb deferred wakeup upon user/guest entry is only relevant when nohz_full runs on the local CPU, otherwise the periodic tick should take care of it.
Make sure we don't needlessly pollute these fast-paths.
Reported-by: kernel test robot <oliver.sang@intel.com> Signed-off-by: Frederic Weisbecker <frederic@kernel.org> --- include/linux/entry-kvm.h | 3 ++- include/linux/tick.h | 7 +++++++ kernel/entry/common.c | 5 +++-- kernel/time/tick-sched.c | 1 + 4 files changed, 13 insertions(+), 3 deletions(-)
diff --git a/include/linux/entry-kvm.h b/include/linux/entry-kvm.h index 8b2b1d68b954..136b8d97d8c0 100644 --- a/include/linux/entry-kvm.h +++ b/include/linux/entry-kvm.h @@ -3,6 +3,7 @@ #define __LINUX_ENTRYKVM_H #include <linux/entry-common.h> +#include <linux/tick.h> /* Transfer to guest mode work */ #ifdef CONFIG_KVM_XFER_TO_GUEST_WORK @@ -57,7 +58,7 @@ int xfer_to_guest_mode_handle_work(struct kvm_vcpu *vcpu); static inline void xfer_to_guest_mode_prepare(void) { lockdep_assert_irqs_disabled(); - rcu_nocb_flush_deferred_wakeup(); + tick_nohz_user_enter_prepare(); } /** diff --git a/include/linux/tick.h b/include/linux/tick.h index 7340613c7eff..1a0ff88fa107 100644 --- a/include/linux/tick.h +++ b/include/linux/tick.h @@ -11,6 +11,7 @@ #include <linux/context_tracking_state.h> #include <linux/cpumask.h> #include <linux/sched.h> +#include <linux/rcupdate.h> #ifdef CONFIG_GENERIC_CLOCKEVENTS extern void __init tick_init(void); @@ -300,4 +301,10 @@ static inline void tick_nohz_task_switch(void) __tick_nohz_task_switch(); } +static inline void tick_nohz_user_enter_prepare(void) +{ + if (tick_nohz_full_cpu(smp_processor_id())) + rcu_nocb_flush_deferred_wakeup(); +} + #endif diff --git a/kernel/entry/common.c b/kernel/entry/common.c index a0b3b04fb596..bf16395b9e13 100644 --- a/kernel/entry/common.c +++ b/kernel/entry/common.c @@ -5,6 +5,7 @@ #include <linux/highmem.h> #include <linux/livepatch.h> #include <linux/audit.h> +#include <linux/tick.h> #include "common.h" @@ -186,7 +187,7 @@ static unsigned long exit_to_user_mode_loop(struct pt_regs *regs, local_irq_disable_exit_to_user(); /* Check if any of the above work has queued a deferred wakeup */ - rcu_nocb_flush_deferred_wakeup(); + tick_nohz_user_enter_prepare(); ti_work = READ_ONCE(current_thread_info()->flags); } @@ -202,7 +203,7 @@ static void exit_to_user_mode_prepare(struct pt_regs *regs) lockdep_assert_irqs_disabled(); /* Flush pending rcuog wakeup before the last need_resched() check */ - rcu_nocb_flush_deferred_wakeup(); + tick_nohz_user_enter_prepare(); if (unlikely(ti_work & EXIT_TO_USER_MODE_WORK)) ti_work = exit_to_user_mode_loop(regs, ti_work); diff --git a/kernel/time/tick-sched.c b/kernel/time/tick-sched.c index 828b091501ca..6784f27a3099 100644 --- a/kernel/time/tick-sched.c +++ b/kernel/time/tick-sched.c @@ -230,6 +230,7 @@ static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs) #ifdef CONFIG_NO_HZ_FULL cpumask_var_t tick_nohz_full_mask; +EXPORT_SYMBOL_GPL(tick_nohz_full_mask); bool tick_nohz_full_running; EXPORT_SYMBOL_GPL(tick_nohz_full_running); static atomic_t tick_dep_mask; -- 2.25.1
|  |