]> git.kernelconcepts.de Git - karo-tx-linux.git/commitdiff
rcu: Better hotplug handling for synchronize_sched_expedited()
authorPaul E. McKenney <paulmck@linux.vnet.ibm.com>
Thu, 3 Sep 2015 07:45:02 +0000 (00:45 -0700)
committerPaul E. McKenney <paulmck@linux.vnet.ibm.com>
Wed, 7 Oct 2015 23:02:50 +0000 (16:02 -0700)
Earlier versions of synchronize_sched_expedited() can prematurely end
grace periods due to the fact that a CPU marked as cpu_is_offline()
can still be using RCU read-side critical sections during the time that
CPU makes its last pass through the scheduler and into the idle loop
and during the time that a given CPU is in the process of coming online.
This commit therefore eliminates this window by adding additional
interaction with the CPU-hotplug operations.

Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
kernel/rcu/tree.c

index b4f71ffa299ce630fc9208a0d54b2bfe27d6b179..42b317e13776ab4462ddfac09a9d6f475bc46f1b 100644 (file)
@@ -246,17 +246,23 @@ static int rcu_gp_in_progress(struct rcu_state *rsp)
  */
 void rcu_sched_qs(void)
 {
+       unsigned long flags;
+
        if (__this_cpu_read(rcu_sched_data.cpu_no_qs.s)) {
                trace_rcu_grace_period(TPS("rcu_sched"),
                                       __this_cpu_read(rcu_sched_data.gpnum),
                                       TPS("cpuqs"));
                __this_cpu_write(rcu_sched_data.cpu_no_qs.b.norm, false);
+               if (!__this_cpu_read(rcu_sched_data.cpu_no_qs.b.exp))
+                       return;
+               local_irq_save(flags);
                if (__this_cpu_read(rcu_sched_data.cpu_no_qs.b.exp)) {
                        __this_cpu_write(rcu_sched_data.cpu_no_qs.b.exp, false);
                        rcu_report_exp_rdp(&rcu_sched_state,
                                           this_cpu_ptr(&rcu_sched_data),
                                           true);
                }
+               local_irq_restore(flags);
        }
 }
 
@@ -3553,7 +3559,10 @@ static void rcu_report_exp_cpu_mult(struct rcu_state *rsp, struct rcu_node *rnp,
 
        raw_spin_lock_irqsave(&rnp->lock, flags);
        smp_mb__after_unlock_lock();
-       WARN_ON_ONCE((rnp->expmask & mask) != mask);
+       if (!(rnp->expmask & mask)) {
+               raw_spin_unlock_irqrestore(&rnp->lock, flags);
+               return;
+       }
        rnp->expmask &= ~mask;
        __rcu_report_exp_rnp(rsp, rnp, wake, flags); /* Releases rnp->lock. */
 }
@@ -3644,12 +3653,37 @@ static struct rcu_node *exp_funnel_lock(struct rcu_state *rsp, unsigned long s)
 }
 
 /* Invoked on each online non-idle CPU for expedited quiescent state. */
-static void synchronize_sched_expedited_cpu_stop(void *data)
+static void sync_sched_exp_handler(void *data)
 {
+       struct rcu_data *rdp;
+       struct rcu_node *rnp;
+       struct rcu_state *rsp = data;
+
+       rdp = this_cpu_ptr(rsp->rda);
+       rnp = rdp->mynode;
+       if (!(READ_ONCE(rnp->expmask) & rdp->grpmask) ||
+           __this_cpu_read(rcu_sched_data.cpu_no_qs.b.exp))
+               return;
        __this_cpu_write(rcu_sched_data.cpu_no_qs.b.exp, true);
        resched_cpu(smp_processor_id());
 }
 
+/* Send IPI for expedited cleanup if needed at end of CPU-hotplug operation. */
+static void sync_sched_exp_online_cleanup(int cpu)
+{
+       struct rcu_data *rdp;
+       int ret;
+       struct rcu_node *rnp;
+       struct rcu_state *rsp = &rcu_sched_state;
+
+       rdp = per_cpu_ptr(rsp->rda, cpu);
+       rnp = rdp->mynode;
+       if (!(READ_ONCE(rnp->expmask) & rdp->grpmask))
+               return;
+       ret = smp_call_function_single(cpu, sync_sched_exp_handler, rsp, 0);
+       WARN_ON_ONCE(ret);
+}
+
 /*
  * Select the nodes that the upcoming expedited grace period needs
  * to wait for.
@@ -3677,7 +3711,6 @@ static void sync_rcu_exp_select_cpus(struct rcu_state *rsp,
                        struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
 
                        if (raw_smp_processor_id() == cpu ||
-                           cpu_is_offline(cpu) ||
                            !(atomic_add_return(0, &rdtp->dynticks) & 0x1))
                                mask_ofl_test |= rdp->grpmask;
                }
@@ -3697,9 +3730,28 @@ static void sync_rcu_exp_select_cpus(struct rcu_state *rsp,
                for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask <<= 1) {
                        if (!(mask_ofl_ipi & mask))
                                continue;
+retry_ipi:
                        ret = smp_call_function_single(cpu, func, rsp, 0);
-                       if (!ret)
+                       if (!ret) {
                                mask_ofl_ipi &= ~mask;
+                       } else {
+                               /* Failed, raced with offline. */
+                               raw_spin_lock_irqsave(&rnp->lock, flags);
+                               if (cpu_online(cpu) &&
+                                   (rnp->expmask & mask)) {
+                                       raw_spin_unlock_irqrestore(&rnp->lock,
+                                                                  flags);
+                                       schedule_timeout_uninterruptible(1);
+                                       if (cpu_online(cpu) &&
+                                           (rnp->expmask & mask))
+                                               goto retry_ipi;
+                                       raw_spin_lock_irqsave(&rnp->lock,
+                                                             flags);
+                               }
+                               if (!(rnp->expmask & mask))
+                                       mask_ofl_ipi &= ~mask;
+                               raw_spin_unlock_irqrestore(&rnp->lock, flags);
+                       }
                }
                /* Report quiescent states for those that went offline. */
                mask_ofl_test |= mask_ofl_ipi;
@@ -3796,7 +3848,7 @@ void synchronize_sched_expedited(void)
                return;  /* Someone else did our work for us. */
 
        rcu_exp_gp_seq_start(rsp);
-       sync_rcu_exp_select_cpus(rsp, synchronize_sched_expedited_cpu_stop);
+       sync_rcu_exp_select_cpus(rsp, sync_sched_exp_handler);
        synchronize_sched_expedited_wait(rsp);
 
        rcu_exp_gp_seq_end(rsp);
@@ -4183,6 +4235,7 @@ int rcu_cpu_notify(struct notifier_block *self,
                break;
        case CPU_ONLINE:
        case CPU_DOWN_FAILED:
+               sync_sched_exp_online_cleanup(cpu);
                rcu_boost_kthread_setaffinity(rnp, -1);
                break;
        case CPU_DOWN_PREPARE:
@@ -4195,7 +4248,10 @@ int rcu_cpu_notify(struct notifier_block *self,
                break;
        case CPU_DYING_IDLE:
                /* QS for any half-done expedited RCU-sched GP. */
-               rcu_sched_qs();
+               preempt_disable();
+               rcu_report_exp_rdp(&rcu_sched_state,
+                                  this_cpu_ptr(rcu_sched_state.rda), true);
+               preempt_enable();
 
                for_each_rcu_flavor(rsp) {
                        rcu_cleanup_dying_idle_cpu(cpu, rsp);