| Message ID | 150307946273.29525.14036690240810795204.stgit@Solace.fritz.box (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
>>> On 18.08.17 at 20:04, <dario.faggioli@citrix.com> wrote: > 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). Just to make sure I correctly understand this - you really mean "they'll set their bit themselves ..."? Jan
On 08/18/2017 07:04 PM, Dario Faggioli wrote: > 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: > > https://lists.xenproject.org/archives/html/xen-devel/2017-01/msg02454.html > > 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@citrix.com> Reviewed-by: George Dunlap <george.dunlap@citrix.com>
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) smp_wmb(); rcp->cur++; - 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); register_cpu_notifier(&cpu_nfb); 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 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: https://lists.xenproject.org/archives/html/xen-devel/2017-01/msg02454.html 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@citrix.com> --- Cc: Andrew Cooper <andrew.cooper3@citrix.com> Cc: George Dunlap <George.Dunlap@eu.citrix.com> Cc: Ian Jackson <ian.jackson@eu.citrix.com> Cc: Jan Beulich <jbeulich@suse.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Stefano Stabellini <sstabellini@kernel.org> Cc: Tim Deegan <tim@xen.org> Cc: Wei Liu <wei.liu2@citrix.com> Cc: Julien Grall <julien.grall@arm.com> --- 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(-)