Re: [patch 1/2] nohz: try to avoid IPI when configuring per-CPU posix timer

2020-09-03 Thread Marcelo Tosatti 
On Wed, Sep 02, 2020 at 01:38:59AM +0200, Frederic Weisbecker wrote:
> On Tue, Aug 25, 2020 at 03:41:48PM -0300, Marcelo Tosatti wrote:
> > When enabling per-CPU posix timers, an IPI to nohz_full CPUs might be
> > performed (to re-read the dependencies and possibly not re-enter
> > nohz_full on a given CPU).
> > 
> > A common case is for applications that run on nohz_full= CPUs 
> > to not use POSIX timers (eg DPDK). This patch skips the IPI 
> > in case the task allowed mask does not intersect with nohz_full= CPU mask,
> > when going through tick_nohz_dep_set_signal.
> > 
> > This reduces interruptions to nohz_full= CPUs.
> > 
> > Signed-off-by: Marcelo Tosatti 
> [...]
> >  /*
> > + * Set bit on nohz full dependency, kicking all cpus
> > + * only if task can run on nohz full CPUs.
> > + */
> > +static void tick_nohz_dep_set_all_cond(struct task_struct *tsk,
> > +  atomic_t *dep,
> > +  enum tick_dep_bits bit)
> > +{
> > +   int prev;
> > +   unsigned long flags;
> > +
> > +   prev = atomic_fetch_or(BIT(bit), dep);
> > +   if (prev)
> > +   return;
> > +
> > +   raw_spin_lock_irqsave(&tsk->pi_lock, flags);
> > +   if (cpumask_intersects(&tsk->cpus_mask, tick_nohz_full_mask))
> > +   tick_nohz_full_kick_all();
> 
> So that's for one task but what about the other threads in that
> process? We are setting the tick dependency on all tasks sharing that
> struct signal.

Hi Frederic,

Yep, fixing in -v2, thanks.

Re: [patch 1/2] nohz: try to avoid IPI when configuring per-CPU posix timer

2020-09-01 Thread Frederic Weisbecker 
On Tue, Aug 25, 2020 at 03:41:48PM -0300, Marcelo Tosatti wrote:
> When enabling per-CPU posix timers, an IPI to nohz_full CPUs might be
> performed (to re-read the dependencies and possibly not re-enter
> nohz_full on a given CPU).
> 
> A common case is for applications that run on nohz_full= CPUs 
> to not use POSIX timers (eg DPDK). This patch skips the IPI 
> in case the task allowed mask does not intersect with nohz_full= CPU mask,
> when going through tick_nohz_dep_set_signal.
> 
> This reduces interruptions to nohz_full= CPUs.
> 
> Signed-off-by: Marcelo Tosatti 
[...]
>  /*
> + * Set bit on nohz full dependency, kicking all cpus
> + * only if task can run on nohz full CPUs.
> + */
> +static void tick_nohz_dep_set_all_cond(struct task_struct *tsk,
> +atomic_t *dep,
> +enum tick_dep_bits bit)
> +{
> + int prev;
> + unsigned long flags;
> +
> + prev = atomic_fetch_or(BIT(bit), dep);
> + if (prev)
> + return;
> +
> + raw_spin_lock_irqsave(&tsk->pi_lock, flags);
> + if (cpumask_intersects(&tsk->cpus_mask, tick_nohz_full_mask))
> + tick_nohz_full_kick_all();

So that's for one task but what about the other threads in that
process? We are setting the tick dependency on all tasks sharing that
struct signal.

Thanks.

> + raw_spin_unlock_irqrestore(&tsk->pi_lock, flags);
> +}
> +

[patch 1/2] nohz: try to avoid IPI when configuring per-CPU posix timer

2020-08-25 Thread Marcelo Tosatti 
When enabling per-CPU posix timers, an IPI to nohz_full CPUs might be
performed (to re-read the dependencies and possibly not re-enter
nohz_full on a given CPU).

A common case is for applications that run on nohz_full= CPUs 
to not use POSIX timers (eg DPDK). This patch skips the IPI 
in case the task allowed mask does not intersect with nohz_full= CPU mask,
when going through tick_nohz_dep_set_signal.

This reduces interruptions to nohz_full= CPUs.

Signed-off-by: Marcelo Tosatti 

---
 include/linux/tick.h   |   11 +++
 kernel/time/posix-cpu-timers.c |4 ++--
 kernel/time/tick-sched.c   |   27 +--
 3 files changed, 34 insertions(+), 8 deletions(-)

Index: linux-2.6/include/linux/tick.h
===
--- linux-2.6.orig/include/linux/tick.h
+++ linux-2.6/include/linux/tick.h
@@ -207,7 +207,8 @@ extern void tick_nohz_dep_set_task(struc
   enum tick_dep_bits bit);
 extern void tick_nohz_dep_clear_task(struct task_struct *tsk,
 enum tick_dep_bits bit);
-extern void tick_nohz_dep_set_signal(struct signal_struct *signal,
+extern void tick_nohz_dep_set_signal(struct task_struct *tsk,
+struct signal_struct *signal,
 enum tick_dep_bits bit);
 extern void tick_nohz_dep_clear_signal(struct signal_struct *signal,
   enum tick_dep_bits bit);
@@ -252,11 +253,12 @@ static inline void tick_dep_clear_task(s
if (tick_nohz_full_enabled())
tick_nohz_dep_clear_task(tsk, bit);
 }
-static inline void tick_dep_set_signal(struct signal_struct *signal,
+static inline void tick_dep_set_signal(struct task_struct *tsk,
+  struct signal_struct *signal,
   enum tick_dep_bits bit)
 {
if (tick_nohz_full_enabled())
-   tick_nohz_dep_set_signal(signal, bit);
+   tick_nohz_dep_set_signal(tsk, signal, bit);
 }
 static inline void tick_dep_clear_signal(struct signal_struct *signal,
 enum tick_dep_bits bit)
@@ -284,7 +286,8 @@ static inline void tick_dep_set_task(str
 enum tick_dep_bits bit) { }
 static inline void tick_dep_clear_task(struct task_struct *tsk,
   enum tick_dep_bits bit) { }
-static inline void tick_dep_set_signal(struct signal_struct *signal,
+static inline void tick_dep_set_signal(struct task_struct *tsk,
+  struct signal_struct *signal,
   enum tick_dep_bits bit) { }
 static inline void tick_dep_clear_signal(struct signal_struct *signal,
 enum tick_dep_bits bit) { }
Index: linux-2.6/kernel/time/posix-cpu-timers.c
===
--- linux-2.6.orig/kernel/time/posix-cpu-timers.c
+++ linux-2.6/kernel/time/posix-cpu-timers.c
@@ -523,7 +523,7 @@ static void arm_timer(struct k_itimer *t
if (CPUCLOCK_PERTHREAD(timer->it_clock))
tick_dep_set_task(p, TICK_DEP_BIT_POSIX_TIMER);
else
-   tick_dep_set_signal(p->signal, TICK_DEP_BIT_POSIX_TIMER);
+   tick_dep_set_signal(p, p->signal, TICK_DEP_BIT_POSIX_TIMER);
 }
 
 /*
@@ -1358,7 +1358,7 @@ void set_process_cpu_timer(struct task_s
if (*newval < *nextevt)
*nextevt = *newval;
 
-   tick_dep_set_signal(tsk->signal, TICK_DEP_BIT_POSIX_TIMER);
+   tick_dep_set_signal(tsk, tsk->signal, TICK_DEP_BIT_POSIX_TIMER);
 }
 
 static int do_cpu_nanosleep(const clockid_t which_clock, int flags,
Index: linux-2.6/kernel/time/tick-sched.c
===
--- linux-2.6.orig/kernel/time/tick-sched.c
+++ linux-2.6/kernel/time/tick-sched.c
@@ -302,6 +302,27 @@ static void tick_nohz_dep_set_all(atomic
 }
 
 /*
+ * Set bit on nohz full dependency, kicking all cpus
+ * only if task can run on nohz full CPUs.
+ */
+static void tick_nohz_dep_set_all_cond(struct task_struct *tsk,
+  atomic_t *dep,
+  enum tick_dep_bits bit)
+{
+   int prev;
+   unsigned long flags;
+
+   prev = atomic_fetch_or(BIT(bit), dep);
+   if (prev)
+   return;
+
+   raw_spin_lock_irqsave(&tsk->pi_lock, flags);
+   if (cpumask_intersects(&tsk->cpus_mask, tick_nohz_full_mask))
+   tick_nohz_full_kick_all();
+   raw_spin_unlock_irqrestore(&tsk->pi_lock, flags);
+}
+
+/*
  * Set a global tick dependency. Used by perf events that rely on freq and
  * by unstable clock.
  */
@@ -382,9 +403,11 @@ EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_ta
  * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to 
elap