[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <20070817205812.GA15575@linux.vnet.ibm.com>
Date: Fri, 17 Aug 2007 13:58:12 -0700
From: "Paul E. McKenney" <paulmck@...ux.vnet.ibm.com>
To: ego@...ibm.com
Cc: linux-kernel@...r.kernel.org, mingo@...e.hu, dvhltc@...ibm.com,
tytso@...ibm.com, akpm@...ux-foundation.org
Subject: Re: [PATCH RFC] CPU hotplug support for preemptible RCU
On Thu, Aug 16, 2007 at 06:26:18PM -0700, Paul E. McKenney wrote:
> On Tue, Aug 14, 2007 at 04:41:16PM -0700, Paul E. McKenney wrote:
> > Hello!
> >
> > Work in progress, not for inclusion.
> >
> > The attached patch passes multiple hours of rcutorture while hotplugging
> > CPUs every ten seconds on 64-bit PPC and x86_64. It fails miserably on
> > 32-bit i386 after a few hotplugs, but then again, so does stock 2.6.22
> > even without running rcutorture simultaneously.
> >
> > Is there some extra patch or hardware dependency for CPU hotplug on
> > 32-bit i386?
>
> Progress report: turns out to be at least partly a hardware dependency.
> I have found a couple of x86 machines that survive a goodly number of
> CPU-hotplug operations on stock 2.6.22. They also have some other
> problems, in one case, only booting at odd intervals regardless of
> what kernel you are using. Though when it does boot, it does seems to
> run reliably. :-/
>
> In contrast, doing even a small number of CPU hotplug operations on the
> other machines eventually causes themselves to NMI themselves to death
> regardless of what kernel I boot on them. One of these machines ends up
> writing LOCK_PREFIX and nop into kernel text on CPU hotplug operations.
> While other CPUs are running. I cannot believe that this is a good
> thing, but suppressing this with the smp-alt-boot (see the bootonly()
> function) doesn't seem to improve matters. Even so, seems to me that
> smp-alt-boot should be enabled by default. Unless there are objections,
> I will send a patch to do this.
And I finally located a couple of i386 machines that tolerate CPU hotplug
operations. They both passed an overnight CPU-hotplug-plus-rcutorture
run. ;-)
Thanx, Paul
> > Signed-off-by: Paul E. McKenney <paulmck@...ux.vnet.ibm.com>
> > ---
> >
> > include/linux/rcuclassic.h | 2
> > include/linux/rcupreempt.h | 2
> > kernel/rcuclassic.c | 8 +++
> > kernel/rcupreempt.c | 93 +++++++++++++++++++++++++++++++++++++++++++--
> > 4 files changed, 100 insertions(+), 5 deletions(-)
> >
> > diff -urpNa -X dontdiff linux-2.6.22-d-schedclassic/include/linux/rcuclassic.h linux-2.6.22-e-hotplugcpu/include/linux/rcuclassic.h
> > --- linux-2.6.22-d-schedclassic/include/linux/rcuclassic.h 2007-08-06 10:16:16.000000000 -0700
> > +++ linux-2.6.22-e-hotplugcpu/include/linux/rcuclassic.h 2007-08-07 18:15:08.000000000 -0700
> > @@ -83,6 +83,8 @@ static inline void rcu_bh_qsctr_inc(int
> > #define rcu_check_callbacks_rt(cpu, user)
> > #define rcu_init_rt()
> > #define rcu_needs_cpu_rt(cpu) 0
> > +#define rcu_offline_cpu_rt(cpu)
> > +#define rcu_online_cpu_rt(cpu)
> > #define rcu_pending_rt(cpu) 0
> > #define rcu_process_callbacks_rt(unused)
> >
> > diff -urpNa -X dontdiff linux-2.6.22-d-schedclassic/include/linux/rcupreempt.h linux-2.6.22-e-hotplugcpu/include/linux/rcupreempt.h
> > --- linux-2.6.22-d-schedclassic/include/linux/rcupreempt.h 2007-08-06 14:56:00.000000000 -0700
> > +++ linux-2.6.22-e-hotplugcpu/include/linux/rcupreempt.h 2007-08-07 18:15:10.000000000 -0700
> > @@ -59,6 +59,8 @@ extern void rcu_advance_callbacks_rt(int
> > extern void rcu_check_callbacks_rt(int cpu, int user);
> > extern void rcu_init_rt(void);
> > extern int rcu_needs_cpu_rt(int cpu);
> > +extern void rcu_offline_cpu_rt(int cpu);
> > +extern void rcu_online_cpu_rt(int cpu);
> > extern int rcu_pending_rt(int cpu);
> > struct softirq_action;
> > extern void rcu_process_callbacks_rt(struct softirq_action *unused);
> > diff -urpNa -X dontdiff linux-2.6.22-d-schedclassic/kernel/rcuclassic.c linux-2.6.22-e-hotplugcpu/kernel/rcuclassic.c
> > --- linux-2.6.22-d-schedclassic/kernel/rcuclassic.c 2007-08-06 14:07:26.000000000 -0700
> > +++ linux-2.6.22-e-hotplugcpu/kernel/rcuclassic.c 2007-08-11 08:25:55.000000000 -0700
> > @@ -404,14 +404,19 @@ static void __rcu_offline_cpu(struct rcu
> > static void rcu_offline_cpu(int cpu)
> > {
> > struct rcu_data *this_rdp = &get_cpu_var(rcu_data);
> > +#ifdef CONFIG_CLASSIC_RCU
> > struct rcu_data *this_bh_rdp = &get_cpu_var(rcu_bh_data);
> > +#endif /* #ifdef CONFIG_CLASSIC_RCU */
> >
> > __rcu_offline_cpu(this_rdp, &rcu_ctrlblk,
> > &per_cpu(rcu_data, cpu));
> > +#ifdef CONFIG_CLASSIC_RCU
> > __rcu_offline_cpu(this_bh_rdp, &rcu_bh_ctrlblk,
> > &per_cpu(rcu_bh_data, cpu));
> > - put_cpu_var(rcu_data);
> > put_cpu_var(rcu_bh_data);
> > +#endif /* #ifdef CONFIG_CLASSIC_RCU */
> > + put_cpu_var(rcu_data);
> > + rcu_offline_cpu_rt(cpu);
> > }
> >
> > #else
> > @@ -561,6 +566,7 @@ static void __devinit rcu_online_cpu(int
> > rdp->passed_quiesc = &per_cpu(rcu_data_passed_quiesc, cpu);
> > rcu_init_percpu_data(cpu, &rcu_bh_ctrlblk, bh_rdp);
> > bh_rdp->passed_quiesc = &per_cpu(rcu_data_bh_passed_quiesc, cpu);
> > + rcu_online_cpu_rt(cpu);
> > }
> >
> > static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
> > diff -urpNa -X dontdiff linux-2.6.22-d-schedclassic/kernel/rcupreempt.c linux-2.6.22-e-hotplugcpu/kernel/rcupreempt.c
> > --- linux-2.6.22-d-schedclassic/kernel/rcupreempt.c 2007-08-06 14:58:07.000000000 -0700
> > +++ linux-2.6.22-e-hotplugcpu/kernel/rcupreempt.c 2007-08-11 04:02:10.000000000 -0700
> > @@ -125,6 +125,8 @@ enum rcu_mb_flag_values {
> > };
> > static DEFINE_PER_CPU(enum rcu_mb_flag_values, rcu_mb_flag) = rcu_mb_done;
> >
> > +static cpumask_t rcu_cpu_online_map = CPU_MASK_NONE;
> > +
> > /*
> > * Macro that prevents the compiler from reordering accesses, but does
> > * absolutely -nothing- to prevent CPUs from reordering. This is used
> > @@ -400,7 +402,7 @@ rcu_try_flip_idle(void)
> >
> > /* Now ask each CPU for acknowledgement of the flip. */
> >
> > - for_each_possible_cpu(cpu)
> > + for_each_cpu_mask(cpu, rcu_cpu_online_map)
> > per_cpu(rcu_flip_flag, cpu) = rcu_flipped;
> >
> > return 1;
> > @@ -416,7 +418,7 @@ rcu_try_flip_waitack(void)
> > int cpu;
> >
> > RCU_TRACE_ME(rcupreempt_trace_try_flip_a1);
> > - for_each_possible_cpu(cpu)
> > + for_each_cpu_mask(cpu, rcu_cpu_online_map)
> > if (per_cpu(rcu_flip_flag, cpu) != rcu_flip_seen) {
> > RCU_TRACE_ME(rcupreempt_trace_try_flip_ae1);
> > return 0;
> > @@ -460,7 +462,7 @@ rcu_try_flip_waitzero(void)
> >
> > /* Call for a memory barrier from each CPU. */
> >
> > - for_each_possible_cpu(cpu)
> > + for_each_cpu_mask(cpu, rcu_cpu_online_map)
> > per_cpu(rcu_mb_flag, cpu) = rcu_mb_needed;
> >
> > RCU_TRACE_ME(rcupreempt_trace_try_flip_z2);
> > @@ -478,7 +480,7 @@ rcu_try_flip_waitmb(void)
> > int cpu;
> >
> > RCU_TRACE_ME(rcupreempt_trace_try_flip_m1);
> > - for_each_possible_cpu(cpu)
> > + for_each_cpu_mask(cpu, rcu_cpu_online_map)
> > if (per_cpu(rcu_mb_flag, cpu) != rcu_mb_done) {
> > RCU_TRACE_ME(rcupreempt_trace_try_flip_me1);
> > return 0;
> > @@ -583,6 +585,89 @@ void rcu_advance_callbacks_rt(int cpu, i
> > spin_unlock_irqrestore(&rdp->lock, oldirq);
> > }
> >
> > +#ifdef CONFIG_HOTPLUG_CPU
> > +
> > +#define rcu_offline_cpu_rt_enqueue(srclist, srctail, dstlist, dsttail) do { \
> > + *dsttail = srclist; \
> > + if (srclist != NULL) { \
> > + dsttail = srctail; \
> > + srclist = NULL; \
> > + srctail = &srclist;\
> > + } \
> > + } while (0)
> > +
> > +
> > +void rcu_offline_cpu_rt(int cpu)
> > +{
> > + int i;
> > + struct rcu_head *list = NULL;
> > + unsigned long oldirq;
> > + struct rcu_data *rdp = RCU_DATA_CPU(cpu);
> > + struct rcu_head **tail = &list;
> > +
> > + /* Remove all callbacks from the newly dead CPU, retaining order. */
> > +
> > + spin_lock_irqsave(&rdp->lock, oldirq);
> > + rcu_offline_cpu_rt_enqueue(rdp->donelist, rdp->donetail, list, tail);
> > + for (i = GP_STAGES - 1; i >= 0; i--)
> > + rcu_offline_cpu_rt_enqueue(rdp->waitlist[i], rdp->waittail[i],
> > + list, tail);
> > + rcu_offline_cpu_rt_enqueue(rdp->nextlist, rdp->nexttail, list, tail);
> > + spin_unlock_irqrestore(&rdp->lock, oldirq);
> > + rdp->waitlistcount = 0;
> > +
> > + /* Disengage the newly dead CPU from grace-period computation. */
> > +
> > + spin_lock_irqsave(&rcu_ctrlblk.fliplock, oldirq);
> > + rcu_check_mb(cpu);
> > + if (per_cpu(rcu_flip_flag, cpu) == rcu_flipped) {
> > + smp_mb(); /* Subsequent counter accesses must see new value */
> > + per_cpu(rcu_flip_flag, cpu) = rcu_flip_seen;
> > + smp_mb(); /* Subsequent RCU read-side critical sections */
> > + /* seen -after- acknowledgement. */
> > + }
> > + cpu_clear(cpu, rcu_cpu_online_map);
> > + spin_unlock_irqrestore(&rcu_ctrlblk.fliplock, oldirq);
> > +
> > + /*
> > + * Place the removed callbacks on the current CPU's queue.
> > + * Make them all start a new grace period: simple approach,
> > + * in theory could starve a given set of callbacks, but
> > + * you would need to be doing some serious CPU hotplugging
> > + * to make this happen. If this becomes a problem, adding
> > + * a synchronize_rcu() to the hotplug path would be a simple
> > + * fix.
> > + */
> > +
> > + rdp = RCU_DATA_ME();
> > + spin_lock_irqsave(&rdp->lock, oldirq);
> > + *rdp->nexttail = list;
> > + if (list)
> > + rdp->nexttail = tail;
> > + spin_unlock_irqrestore(&rdp->lock, oldirq);
> > +}
> > +
> > +void __devinit rcu_online_cpu_rt(int cpu)
> > +{
> > + unsigned long oldirq;
> > +
> > + spin_lock_irqsave(&rcu_ctrlblk.fliplock, oldirq);
> > + cpu_set(cpu, rcu_cpu_online_map);
> > + spin_unlock_irqrestore(&rcu_ctrlblk.fliplock, oldirq);
> > +}
> > +
> > +#else /* #ifdef CONFIG_HOTPLUG_CPU */
> > +
> > +void rcu_offline_cpu(int cpu)
> > +{
> > +}
> > +
> > +void __devinit rcu_online_cpu_rt(int cpu)
> > +{
> > +}
> > +
> > +#endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
> > +
> > void rcu_process_callbacks_rt(struct softirq_action *unused)
> > {
> > unsigned long flags;
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
Powered by blists - more mailing lists