[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

[Xen-devel] [PATCH v3 3/6] xen: RCU/x86/ARM: discount CPUs that were idle when grace period started.

Xen is a tickless (micro-)kernel, i.e., when a CPU becomes
idle there is no timer tick that will periodically wake the
CPU up.
OTOH, when we imported RCU from Linux, Linux was (on x86) a
ticking kernel, i.e., there was a periodic timer tick always
running, even on idle CPUs. This was bad for power consumption,
but, for instance, made it easy to monitor the quiescent states
of all the CPUs, and hence tell when RCU grace periods ended.

In Xen, that is impossible, and that's particularly problematic
when the system is very lightly loaded, as some CPUs may never
have the chance to tell the RCU core logic about their quiescence,
and grace periods could extend indefinitely!

This has led, on x86, to long (and unpredictable) delays between
RCU callbacks queueing and their actual invokation. On ARM, we've
even seen infinite grace periods (e.g., complate_domain_destroy()
never being actually invoked!). See here:


The first step for fixing this situation is for RCU to record,
at the beginning of a grace period, which CPUs are already idle.
In fact, being idle, they can't be in the middle of any read-side
critical section, and we don't have to wait for their quiescence.

This is tracked in a cpumask, in a similar way to how it was also
done in Linux (on s390, which was tickless already). It is also
basically the same approach used for making Linux x86 tickless,
in 2.6.21 on (see commit 79bf2bb3 "tick-management: dyntick /
highres functionality").

For correctness, wee also add barriers. One is also present in
Linux, (see commit c3f59023, "Fix RCU race in access of nohz_cpu_mask",
although, we change the code comment to something that makes better
sense for us). The other (which is its pair), is put in the newly
introduced function rcu_idle_enter(), right after updating the
cpumask. They prevent races between CPUs going idle during the
beginning of a grace period.

Signed-off-by: Dario Faggioli <dario.faggioli@xxxxxxxxxx>
Cc: Andrew Cooper <andrew.cooper3@xxxxxxxxxx>
Cc: George Dunlap <George.Dunlap@xxxxxxxxxxxxx>
Cc: Ian Jackson <ian.jackson@xxxxxxxxxxxxx>
Cc: Jan Beulich <jbeulich@xxxxxxxx>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@xxxxxxxxxx>
Cc: Stefano Stabellini <sstabellini@xxxxxxxxxx>
Cc: Tim Deegan <tim@xxxxxxx>
Cc: Wei Liu <wei.liu2@xxxxxxxxxx>
Cc: Julien Grall <julien.grall@xxxxxxx>
Changes from v2:
* initialize idle_cpumask to "all clear", i.e., all the CPUs are busy;
  they'll clear their bit out themselves as soon as the run the idle
  loop (pretty soon anyway).

Changes from v1:
* call rcu_idle_{enter,exit}() from tick suspension/restarting logic.  This
  widen the window during which a CPU has its bit set in the idle cpumask.
  During review, it was suggested to do the opposite (narrow it), and that's
  what I did first. But then, I changed my mind, as doing things as they look
  now (wide window), cures another pre-existing (and independent) raca which
  Tim discovered, still during v1 review;
* add a barrier in rcu_idle_enter() too, to properly deal with the race Tim
  pointed out during review;
* mark CPU where RCU initialization happens, at boot, as non-idle.
 xen/common/rcupdate.c      |   41 +++++++++++++++++++++++++++++++++++++++--
 xen/common/schedule.c      |    2 ++
 xen/include/xen/rcupdate.h |    3 +++
 3 files changed, 44 insertions(+), 2 deletions(-)

diff --git a/xen/common/rcupdate.c b/xen/common/rcupdate.c
index 8cc5a82..12ae7da 100644
--- a/xen/common/rcupdate.c
+++ b/xen/common/rcupdate.c
@@ -52,7 +52,8 @@ static struct rcu_ctrlblk {
     int  next_pending;  /* Is the next batch already waiting?         */
     spinlock_t  lock __cacheline_aligned;
-    cpumask_t   cpumask; /* CPUs that need to switch in order    */
+    cpumask_t   cpumask; /* CPUs that need to switch in order ... */
+    cpumask_t   idle_cpumask; /* ... unless they are already idle */
     /* for current batch to proceed.        */
 } __cacheline_aligned rcu_ctrlblk = {
     .cur = -300,
@@ -248,7 +249,16 @@ static void rcu_start_batch(struct rcu_ctrlblk *rcp)
-        cpumask_copy(&rcp->cpumask, &cpu_online_map);
+       /*
+        * Make sure the increment of rcp->cur is visible so, even if a
+        * CPU that is about to go idle, is captured inside rcp->cpumask,
+        * rcu_pending() will return false, which then means cpu_quiet()
+        * will be invoked, before the CPU would actually enter idle.
+        *
+        * This barrier is paired with the one in rcu_idle_enter().
+        */
+        smp_mb();
+        cpumask_andnot(&rcp->cpumask, &cpu_online_map, &rcp->idle_cpumask);
@@ -474,7 +484,34 @@ static struct notifier_block cpu_nfb = {
 void __init rcu_init(void)
     void *cpu = (void *)(long)smp_processor_id();
+    cpumask_clear(&rcu_ctrlblk.idle_cpumask);
     cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);
     open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
+ * The CPU is becoming idle, so no more read side critical
+ * sections, and one more step toward grace period.
+ */
+void rcu_idle_enter(unsigned int cpu)
+    ASSERT(!cpumask_test_cpu(cpu, &rcu_ctrlblk.idle_cpumask));
+    cpumask_set_cpu(cpu, &rcu_ctrlblk.idle_cpumask);
+    /*
+     * If some other CPU is starting a new grace period, we'll notice that
+     * by seeing a new value in rcp->cur (different than our quiescbatch).
+     * That will force us all the way until cpu_quiet(), clearing our bit
+     * in rcp->cpumask, even in case we managed to get in there.
+     *
+     * Se the comment before cpumask_andnot() in  rcu_start_batch().
+     */
+    smp_mb();
+void rcu_idle_exit(unsigned int cpu)
+    ASSERT(cpumask_test_cpu(cpu, &rcu_ctrlblk.idle_cpumask));
+    cpumask_clear_cpu(cpu, &rcu_ctrlblk.idle_cpumask);
diff --git a/xen/common/schedule.c b/xen/common/schedule.c
index e83f4c7..c6f4817 100644
--- a/xen/common/schedule.c
+++ b/xen/common/schedule.c
@@ -1903,6 +1903,7 @@ void sched_tick_suspend(void)
     sched = per_cpu(scheduler, cpu);
     SCHED_OP(sched, tick_suspend, cpu);
+    rcu_idle_enter(cpu);
 void sched_tick_resume(void)
@@ -1910,6 +1911,7 @@ void sched_tick_resume(void)
     struct scheduler *sched;
     unsigned int cpu = smp_processor_id();
+    rcu_idle_exit(cpu);
     sched = per_cpu(scheduler, cpu);
     SCHED_OP(sched, tick_resume, cpu);
diff --git a/xen/include/xen/rcupdate.h b/xen/include/xen/rcupdate.h
index 557a7b1..561ac43 100644
--- a/xen/include/xen/rcupdate.h
+++ b/xen/include/xen/rcupdate.h
@@ -146,4 +146,7 @@ void call_rcu(struct rcu_head *head,
 int rcu_barrier(void);
+void rcu_idle_enter(unsigned int cpu);
+void rcu_idle_exit(unsigned int cpu);
 #endif /* __XEN_RCUPDATE_H */

Xen-devel mailing list



Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.