rcu: Simplify offline processing
authorPaul E. McKenney <paul.mckenney@linaro.org>
Sat, 7 Jan 2012 19:03:57 +0000 (11:03 -0800)
committerPaul E. McKenney <paulmck@linux.vnet.ibm.com>
Tue, 21 Feb 2012 17:03:34 +0000 (09:03 -0800)
Move ->qsmaskinit and blkd_tasks[] manipulation to the CPU_DYING
notifier.  This simplifies the code by eliminating a potential
deadlock and by reducing the responsibilities of force_quiescent_state().
Also rename functions to make their connection to the CPU-hotplug
stages explicit.

Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
kernel/rcutree.c
kernel/rcutree.h
kernel/rcutree_plugin.h

index acf2d67..575f91d 100644 (file)
@@ -943,6 +943,10 @@ rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_dat
  * in preparation for detecting the next grace period.  The caller must hold
  * the root node's ->lock, which is released before return.  Hard irqs must
  * be disabled.
+ *
+ * Note that it is legal for a dying CPU (which is marked as offline) to
+ * invoke this function.  This can happen when the dying CPU reports its
+ * quiescent state.
  */
 static void
 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
@@ -1245,118 +1249,101 @@ rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
 
 /*
  * Move a dying CPU's RCU callbacks to online CPU's callback list.
- * Synchronization is not required because this function executes
- * in stop_machine() context.
+ * Also record a quiescent state for this CPU for the current grace period.
+ * Synchronization and interrupt disabling are not required because
+ * this function executes in stop_machine() context.  Therefore, cleanup
+ * operations that might block must be done later from the CPU_DEAD
+ * notifier.
+ *
+ * Note that the outgoing CPU's bit has already been cleared in the
+ * cpu_online_mask.  This allows us to randomly pick a callback
+ * destination from the bits set in that mask.
  */
-static void rcu_send_cbs_to_online(struct rcu_state *rsp)
+static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
 {
+       unsigned long flags;
        int i;
-       /* current DYING CPU is cleared in the cpu_online_mask */
+       unsigned long mask;
+       int need_report;
        int receive_cpu = cpumask_any(cpu_online_mask);
        struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
        struct rcu_data *receive_rdp = per_cpu_ptr(rsp->rda, receive_cpu);
+       struct rcu_node *rnp = rdp->mynode; /* For dying CPU. */
+
+       /* Move callbacks to some other CPU. */
+       if (rdp->nxtlist != NULL) {
+               *receive_rdp->nxttail[RCU_NEXT_TAIL] = rdp->nxtlist;
+               receive_rdp->nxttail[RCU_NEXT_TAIL] =
+                               rdp->nxttail[RCU_NEXT_TAIL];
+               receive_rdp->qlen_lazy += rdp->qlen_lazy;
+               receive_rdp->qlen += rdp->qlen;
+               receive_rdp->n_cbs_adopted += rdp->qlen;
+               rdp->n_cbs_orphaned += rdp->qlen;
+
+               rdp->nxtlist = NULL;
+               for (i = 0; i < RCU_NEXT_SIZE; i++)
+                       rdp->nxttail[i] = &rdp->nxtlist;
+               rdp->qlen_lazy = 0;
+               rdp->qlen = 0;
+       }
 
-       if (rdp->nxtlist == NULL)
-               return;  /* irqs disabled, so comparison is stable. */
-
-       *receive_rdp->nxttail[RCU_NEXT_TAIL] = rdp->nxtlist;
-       receive_rdp->nxttail[RCU_NEXT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
-       receive_rdp->qlen_lazy += rdp->qlen_lazy;
-       receive_rdp->qlen += rdp->qlen;
-       receive_rdp->n_cbs_adopted += rdp->qlen;
-       rdp->n_cbs_orphaned += rdp->qlen;
-
-       rdp->nxtlist = NULL;
-       for (i = 0; i < RCU_NEXT_SIZE; i++)
-               rdp->nxttail[i] = &rdp->nxtlist;
-       rdp->qlen_lazy = 0;
-       rdp->qlen = 0;
-}
-
-/*
- * Remove the outgoing CPU from the bitmasks in the rcu_node hierarchy
- * and move all callbacks from the outgoing CPU to the current one.
- * There can only be one CPU hotplug operation at a time, so no other
- * CPU can be attempting to update rcu_cpu_kthread_task.
- */
-static void __rcu_offline_cpu(int cpu, struct rcu_state *rsp)
-{
-       unsigned long flags;
-       unsigned long mask;
-       int need_report = 0;
-       struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
-       struct rcu_node *rnp;
-
-       rcu_stop_cpu_kthread(cpu);
-
-       /* Exclude any attempts to start a new grace period. */
-       raw_spin_lock_irqsave(&rsp->onofflock, flags);
-
-       /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
-       rnp = rdp->mynode;      /* this is the outgoing CPU's rnp. */
+       /* Record a quiescent state for the dying CPU. */
        mask = rdp->grpmask;    /* rnp->grplo is constant. */
+       trace_rcu_grace_period(rsp->name,
+                              rnp->gpnum + 1 - !!(rnp->qsmask & mask),
+                              "cpuofl");
+       rcu_report_qs_rdp(smp_processor_id(), rsp, rdp, rsp->gpnum);
+       /* Note that rcu_report_qs_rdp() might call trace_rcu_grace_period(). */
+
+       /*
+        * Remove the dying CPU from the bitmasks in the rcu_node
+        * hierarchy.  Because we are in stop_machine() context, we
+        * automatically exclude ->onofflock critical sections.
+        */
        do {
-               raw_spin_lock(&rnp->lock);      /* irqs already disabled. */
+               raw_spin_lock_irqsave(&rnp->lock, flags);
                rnp->qsmaskinit &= ~mask;
                if (rnp->qsmaskinit != 0) {
-                       if (rnp != rdp->mynode)
-                               raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
-                       else
-                               trace_rcu_grace_period(rsp->name,
-                                                      rnp->gpnum + 1 -
-                                                      !!(rnp->qsmask & mask),
-                                                      "cpuofl");
+                       raw_spin_unlock_irqrestore(&rnp->lock, flags);
                        break;
                }
                if (rnp == rdp->mynode) {
-                       trace_rcu_grace_period(rsp->name,
-                                              rnp->gpnum + 1 -
-                                              !!(rnp->qsmask & mask),
-                                              "cpuofl");
                        need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
+                       if (need_report & RCU_OFL_TASKS_NORM_GP)
+                               rcu_report_unblock_qs_rnp(rnp, flags);
+                       else
+                               raw_spin_unlock_irqrestore(&rnp->lock, flags);
+                       if (need_report & RCU_OFL_TASKS_EXP_GP)
+                               rcu_report_exp_rnp(rsp, rnp, true);
                } else
-                       raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
+                       raw_spin_unlock_irqrestore(&rnp->lock, flags);
                mask = rnp->grpmask;
                rnp = rnp->parent;
        } while (rnp != NULL);
-
-       /*
-        * We still hold the leaf rcu_node structure lock here, and
-        * irqs are still disabled.  The reason for this subterfuge is
-        * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
-        * held leads to deadlock.
-        */
-       raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
-       rnp = rdp->mynode;
-       if (need_report & RCU_OFL_TASKS_NORM_GP)
-               rcu_report_unblock_qs_rnp(rnp, flags);
-       else
-               raw_spin_unlock_irqrestore(&rnp->lock, flags);
-       if (need_report & RCU_OFL_TASKS_EXP_GP)
-               rcu_report_exp_rnp(rsp, rnp, true);
-       rcu_node_kthread_setaffinity(rnp, -1);
 }
 
 /*
- * Remove the specified CPU from the RCU hierarchy and move any pending
- * callbacks that it might have to the current CPU.  This code assumes
- * that at least one CPU in the system will remain running at all times.
- * Any attempt to offline -all- CPUs is likely to strand RCU callbacks.
+ * The CPU has been completely removed, and some other CPU is reporting
+ * this fact from process context.  Do the remainder of the cleanup.
+ * There can only be one CPU hotplug operation at a time, so no other
+ * CPU can be attempting to update rcu_cpu_kthread_task.
  */
-static void rcu_offline_cpu(int cpu)
+static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
 {
-       __rcu_offline_cpu(cpu, &rcu_sched_state);
-       __rcu_offline_cpu(cpu, &rcu_bh_state);
-       rcu_preempt_offline_cpu(cpu);
+       struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
+       struct rcu_node *rnp = rdp->mynode;
+
+       rcu_stop_cpu_kthread(cpu);
+       rcu_node_kthread_setaffinity(rnp, -1);
 }
 
 #else /* #ifdef CONFIG_HOTPLUG_CPU */
 
-static void rcu_send_cbs_to_online(struct rcu_state *rsp)
+static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
 {
 }
 
-static void rcu_offline_cpu(int cpu)
+static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
 {
 }
 
@@ -1725,6 +1712,7 @@ __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
         * a quiescent state betweentimes.
         */
        local_irq_save(flags);
+       WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
        rdp = this_cpu_ptr(rsp->rda);
 
        /* Add the callback to our list. */
@@ -2155,16 +2143,18 @@ static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
                 * touch any data without introducing corruption. We send the
                 * dying CPU's callbacks to an arbitrarily chosen online CPU.
                 */
-               rcu_send_cbs_to_online(&rcu_bh_state);
-               rcu_send_cbs_to_online(&rcu_sched_state);
-               rcu_preempt_send_cbs_to_online();
+               rcu_cleanup_dying_cpu(&rcu_bh_state);
+               rcu_cleanup_dying_cpu(&rcu_sched_state);
+               rcu_preempt_cleanup_dying_cpu();
                rcu_cleanup_after_idle(cpu);
                break;
        case CPU_DEAD:
        case CPU_DEAD_FROZEN:
        case CPU_UP_CANCELED:
        case CPU_UP_CANCELED_FROZEN:
-               rcu_offline_cpu(cpu);
+               rcu_cleanup_dead_cpu(cpu, &rcu_bh_state);
+               rcu_cleanup_dead_cpu(cpu, &rcu_sched_state);
+               rcu_preempt_cleanup_dead_cpu(cpu);
                break;
        default:
                break;
index af2af3c..05e0367 100644 (file)
@@ -439,8 +439,8 @@ static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp);
 static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
                                     struct rcu_node *rnp,
                                     struct rcu_data *rdp);
-static void rcu_preempt_offline_cpu(int cpu);
 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
+static void rcu_preempt_cleanup_dead_cpu(int cpu);
 static void rcu_preempt_check_callbacks(int cpu);
 static void rcu_preempt_process_callbacks(void);
 void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu));
@@ -451,7 +451,7 @@ static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
 static int rcu_preempt_pending(int cpu);
 static int rcu_preempt_needs_cpu(int cpu);
 static void __cpuinit rcu_preempt_init_percpu_data(int cpu);
-static void rcu_preempt_send_cbs_to_online(void);
+static void rcu_preempt_cleanup_dying_cpu(void);
 static void __init __rcu_init_preempt(void);
 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags);
 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp);
index 7adf232..eeb2cc6 100644 (file)
@@ -618,16 +618,16 @@ static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
        return retval;
 }
 
+#endif /* #ifdef CONFIG_HOTPLUG_CPU */
+
 /*
  * Do CPU-offline processing for preemptible RCU.
  */
-static void rcu_preempt_offline_cpu(int cpu)
+static void rcu_preempt_cleanup_dead_cpu(int cpu)
 {
-       __rcu_offline_cpu(cpu, &rcu_preempt_state);
+       rcu_cleanup_dead_cpu(cpu, &rcu_preempt_state);
 }
 
-#endif /* #ifdef CONFIG_HOTPLUG_CPU */
-
 /*
  * Check for a quiescent state from the current CPU.  When a task blocks,
  * the task is recorded in the corresponding CPU's rcu_node structure,
@@ -912,11 +912,12 @@ static void __cpuinit rcu_preempt_init_percpu_data(int cpu)
 }
 
 /*
- * Move preemptible RCU's callbacks from dying CPU to other online CPU.
+ * Move preemptible RCU's callbacks from dying CPU to other online CPU
+ * and record a quiescent state.
  */
-static void rcu_preempt_send_cbs_to_online(void)
+static void rcu_preempt_cleanup_dying_cpu(void)
 {
-       rcu_send_cbs_to_online(&rcu_preempt_state);
+       rcu_cleanup_dying_cpu(&rcu_preempt_state);
 }
 
 /*
@@ -1052,16 +1053,16 @@ static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
        return 0;
 }
 
+#endif /* #ifdef CONFIG_HOTPLUG_CPU */
+
 /*
  * Because preemptible RCU does not exist, it never needs CPU-offline
  * processing.
  */
-static void rcu_preempt_offline_cpu(int cpu)
+static void rcu_preempt_cleanup_dead_cpu(int cpu)
 {
 }
 
-#endif /* #ifdef CONFIG_HOTPLUG_CPU */
-
 /*
  * Because preemptible RCU does not exist, it never has any callbacks
  * to check.
@@ -1153,9 +1154,9 @@ static void __cpuinit rcu_preempt_init_percpu_data(int cpu)
 }
 
 /*
- * Because there is no preemptible RCU, there are no callbacks to move.
+ * Because there is no preemptible RCU, there is no cleanup to do.
  */
-static void rcu_preempt_send_cbs_to_online(void)
+static void rcu_preempt_cleanup_dying_cpu(void)
 {
 }