RCU idle/no-CB changes for 3.9

index e441b77..ac6a75d 100644
@@ -305,17 +305,27 @@ cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
 }
 
 /*
- * Does the current CPU require a yet-as-unscheduled grace period?
+ * Does the current CPU require a not-yet-started grace period?
+ * The caller must have disabled interrupts to prevent races with
+ * normal callback registry.
  */
 static int
 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
 {
-	struct rcu_head **ntp;
+	int i;
 
-	ntp = rdp->nxttail[RCU_DONE_TAIL +
-			   (ACCESS_ONCE(rsp->completed) != rdp->completed)];
-	return rdp->nxttail[RCU_DONE_TAIL] && ntp && *ntp &&
-	       !rcu_gp_in_progress(rsp);
+	if (rcu_gp_in_progress(rsp))
+		return 0;  /* No, a grace period is already in progress. */
+	if (!rdp->nxttail[RCU_NEXT_TAIL])
+		return 0;  /* No, this is a no-CBs (or offline) CPU. */
+	if (*rdp->nxttail[RCU_NEXT_READY_TAIL])
+		return 1;  /* Yes, this CPU has newly registered callbacks. */
+	for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++)
+		if (rdp->nxttail[i - 1] != rdp->nxttail[i] &&
+		    ULONG_CMP_LT(ACCESS_ONCE(rsp->completed),
+				 rdp->nxtcompleted[i]))
+			return 1;  /* Yes, CBs for future grace period. */
+	return 0; /* No grace period needed. */
 }
 
 /*

@@ -1071,6 +1081,139 @@ static void init_callback_list(struct rcu_data *rdp)
 }
 
 /*
+ * Determine the value that ->completed will have at the end of the
+ * next subsequent grace period.  This is used to tag callbacks so that
+ * a CPU can invoke callbacks in a timely fashion even if that CPU has
+ * been dyntick-idle for an extended period with callbacks under the
+ * influence of RCU_FAST_NO_HZ.
+ *
+ * The caller must hold rnp->lock with interrupts disabled.
+ */
+static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
+				       struct rcu_node *rnp)
+{
+	/*
+	 * If RCU is idle, we just wait for the next grace period.
+	 * But we can only be sure that RCU is idle if we are looking
+	 * at the root rcu_node structure -- otherwise, a new grace
+	 * period might have started, but just not yet gotten around
+	 * to initializing the current non-root rcu_node structure.
+	 */
+	if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
+		return rnp->completed + 1;
+
+	/*
+	 * Otherwise, wait for a possible partial grace period and
+	 * then the subsequent full grace period.
+	 */
+	return rnp->completed + 2;
+}
+
+/*
+ * If there is room, assign a ->completed number to any callbacks on
+ * this CPU that have not already been assigned.  Also accelerate any
+ * callbacks that were previously assigned a ->completed number that has
+ * since proven to be too conservative, which can happen if callbacks get
+ * assigned a ->completed number while RCU is idle, but with reference to
+ * a non-root rcu_node structure.  This function is idempotent, so it does
+ * not hurt to call it repeatedly.
+ *
+ * The caller must hold rnp->lock with interrupts disabled.
+ */
+static void rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
+			       struct rcu_data *rdp)
+{
+	unsigned long c;
+	int i;
+
+	/* If the CPU has no callbacks, nothing to do. */
+	if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
+		return;
+
+	/*
+	 * Starting from the sublist containing the callbacks most
+	 * recently assigned a ->completed number and working down, find the
+	 * first sublist that is not assignable to an upcoming grace period.
+	 * Such a sublist has something in it (first two tests) and has
+	 * a ->completed number assigned that will complete sooner than
+	 * the ->completed number for newly arrived callbacks (last test).
+	 *
+	 * The key point is that any later sublist can be assigned the
+	 * same ->completed number as the newly arrived callbacks, which
+	 * means that the callbacks in any of these later sublist can be
+	 * grouped into a single sublist, whether or not they have already
+	 * been assigned a ->completed number.
+	 */
+	c = rcu_cbs_completed(rsp, rnp);
+	for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--)
+		if (rdp->nxttail[i] != rdp->nxttail[i - 1] &&
+		    !ULONG_CMP_GE(rdp->nxtcompleted[i], c))
+			break;
+
+	/*
+	 * If there are no sublist for unassigned callbacks, leave.
+	 * At the same time, advance "i" one sublist, so that "i" will
+	 * index into the sublist where all the remaining callbacks should
+	 * be grouped into.
+	 */
+	if (++i >= RCU_NEXT_TAIL)
+		return;
+
+	/*
+	 * Assign all subsequent callbacks' ->completed number to the next
+	 * full grace period and group them all in the sublist initially
+	 * indexed by "i".
+	 */
+	for (; i <= RCU_NEXT_TAIL; i++) {
+		rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL];
+		rdp->nxtcompleted[i] = c;
+	}
+}
+
+/*
+ * Move any callbacks whose grace period has completed to the
+ * RCU_DONE_TAIL sublist, then compact the remaining sublists and
+ * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
+ * sublist.  This function is idempotent, so it does not hurt to
+ * invoke it repeatedly.  As long as it is not invoked -too- often...
+ *
+ * The caller must hold rnp->lock with interrupts disabled.
+ */
+static void rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
+			    struct rcu_data *rdp)
+{
+	int i, j;
+
+	/* If the CPU has no callbacks, nothing to do. */
+	if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
+		return;
+
+	/*
+	 * Find all callbacks whose ->completed numbers indicate that they
+	 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
+	 */
+	for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) {
+		if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i]))
+			break;
+		rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i];
+	}
+	/* Clean up any sublist tail pointers that were misordered above. */
+	for (j = RCU_WAIT_TAIL; j < i; j++)
+		rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL];
+
+	/* Copy down callbacks to fill in empty sublists. */
+	for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) {
+		if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL])
+			break;
+		rdp->nxttail[j] = rdp->nxttail[i];
+		rdp->nxtcompleted[j] = rdp->nxtcompleted[i];
+	}
+
+	/* Classify any remaining callbacks. */
+	rcu_accelerate_cbs(rsp, rnp, rdp);
+}
+
+/*
  * Advance this CPU's callbacks, but only if the current grace period
  * has ended.  This may be called only from the CPU to whom the rdp
  * belongs.  In addition, the corresponding leaf rcu_node structure's

@@ -1080,12 +1223,15 @@ static void
 __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
 {
 	/* Did another grace period end? */
-	if (rdp->completed != rnp->completed) {
+	if (rdp->completed == rnp->completed) {
+
+		/* No, so just accelerate recent callbacks. */
+		rcu_accelerate_cbs(rsp, rnp, rdp);
 
-		/* Advance callbacks.  No harm if list empty. */
-		rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
-		rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
-		rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
+	} else {
+
+		/* Advance callbacks. */
+		rcu_advance_cbs(rsp, rnp, rdp);
 
 		/* Remember that we saw this grace-period completion. */
 		rdp->completed = rnp->completed;

@@ -1392,17 +1538,10 @@ rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
 	/*
 	 * Because there is no grace period in progress right now,
 	 * any callbacks we have up to this point will be satisfied
-	 * by the next grace period.  So promote all callbacks to be
-	 * handled after the end of the next grace period.  If the
-	 * CPU is not yet aware of the end of the previous grace period,
-	 * we need to allow for the callback advancement that will
-	 * occur when it does become aware.  Deadlock prevents us from
-	 * making it aware at this point: We cannot acquire a leaf
-	 * rcu_node ->lock while holding the root rcu_node ->lock.
+	 * by the next grace period.  So this is a good place to
+	 * assign a grace period number to recently posted callbacks.
 	 */
-	rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
-	if (rdp->completed == rsp->completed)
-		rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
+	rcu_accelerate_cbs(rsp, rnp, rdp);
 
 	rsp->gp_flags = RCU_GP_FLAG_INIT;
 	raw_spin_unlock(&rnp->lock); /* Interrupts remain disabled. */

@@ -1527,7 +1666,7 @@ rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
 		 * This GP can't end until cpu checks in, so all of our
 		 * callbacks can be processed during the next GP.
 		 */
-		rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
+		rcu_accelerate_cbs(rsp, rnp, rdp);
 
 		rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
 	}

@@ -1779,7 +1918,7 @@ static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
 	long bl, count, count_lazy;
 	int i;
 
-	/* If no callbacks are ready, just return.*/
+	/* If no callbacks are ready, just return. */
 	if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
 		trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
 		trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),

@@ -2008,19 +2147,19 @@ __rcu_process_callbacks(struct rcu_state *rsp)
 
 	WARN_ON_ONCE(rdp->beenonline == 0);
 
-	/*
-	 * Advance callbacks in response to end of earlier grace
-	 * period that some other CPU ended.
-	 */
+	/* Handle the end of a grace period that some other CPU ended.  */
 	rcu_process_gp_end(rsp, rdp);
 
 	/* Update RCU state based on any recent quiescent states. */
 	rcu_check_quiescent_state(rsp, rdp);
 
 	/* Does this CPU require a not-yet-started grace period? */
+	local_irq_save(flags);
 	if (cpu_needs_another_gp(rsp, rdp)) {
-		raw_spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
+		raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */
 		rcu_start_gp(rsp, flags);  /* releases above lock */
+	} else {
+		local_irq_restore(flags);
 	}
 
 	/* If there are callbacks ready, invoke them. */

index 4b69291..c9f362e 100644
@@ -282,6 +282,8 @@ struct rcu_data {
 	 */
 	struct rcu_head *nxtlist;
 	struct rcu_head **nxttail[RCU_NEXT_SIZE];
+	unsigned long	nxtcompleted[RCU_NEXT_SIZE];
+					/* grace periods for sublists. */
 	long		qlen_lazy;	/* # of lazy queued callbacks */
 	long		qlen;		/* # of queued callbacks, incl lazy */
 	long		qlen_last_fqs_check;

index d4f559b..5678114 100644
@@ -44,8 +44,10 @@ TRACE_EVENT(rcu_utilization,
  * of a new grace period or the end of an old grace period ("cpustart"
  * and "cpuend", respectively), a CPU passing through a quiescent
  * state ("cpuqs"), a CPU coming online or going offline ("cpuonl"
- * and "cpuofl", respectively), and a CPU being kicked for being too
- * long in dyntick-idle mode ("kick").
+ * and "cpuofl", respectively), a CPU being kicked for being too
+ * long in dyntick-idle mode ("kick"), a CPU accelerating its new
+ * callbacks to RCU_NEXT_READY_TAIL ("AccReadyCB"), and a CPU
+ * accelerating its new callbacks to RCU_WAIT_TAIL ("AccWaitCB").
  */
 TRACE_EVENT(rcu_grace_period,
 

index ac6a75d..e9dce4f 100644
@@ -1168,6 +1168,12 @@ static void rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
 		rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL];
 		rdp->nxtcompleted[i] = c;
 	}
+
+	/* Trace depending on how much we were able to accelerate. */
+	if (!*rdp->nxttail[RCU_WAIT_TAIL])
+		trace_rcu_grace_period(rsp->name, rdp->gpnum, "AccWaitCB");
+	else
+		trace_rcu_grace_period(rsp->name, rdp->gpnum, "AccReadyCB");
 }
 
 /*

index 5678114..ef0bf31 100644
@@ -72,6 +72,58 @@ TRACE_EVENT(rcu_grace_period,
 );
 
 /*
+ * Tracepoint for no-callbacks grace-period events.  The caller should
+ * pull the data from the rcu_node structure, other than rcuname, which
+ * comes from the rcu_state structure, and event, which is one of the
+ * following:
+ *
+ * "Startleaf": Request a nocb grace period based on leaf-node data.
+ * "Startedleaf": Leaf-node start proved sufficient.
+ * "Startedleafroot": Leaf-node start proved sufficient after checking root.
+ * "Startedroot": Requested a nocb grace period based on root-node data.
+ * "StartWait": Start waiting for the requested grace period.
+ * "ResumeWait": Resume waiting after signal.
+ * "EndWait": Complete wait.
+ * "Cleanup": Clean up rcu_node structure after previous GP.
+ * "CleanupMore": Clean up, and another no-CB GP is needed.
+ */
+TRACE_EVENT(rcu_nocb_grace_period,
+
+	TP_PROTO(char *rcuname, unsigned long gpnum, unsigned long completed,
+		 unsigned long c, u8 level, int grplo, int grphi,
+		 char *gpevent),
+
+	TP_ARGS(rcuname, gpnum, completed, c, level, grplo, grphi, gpevent),
+
+	TP_STRUCT__entry(
+		__field(char *, rcuname)
+		__field(unsigned long, gpnum)
+		__field(unsigned long, completed)
+		__field(unsigned long, c)
+		__field(u8, level)
+		__field(int, grplo)
+		__field(int, grphi)
+		__field(char *, gpevent)
+	),
+
+	TP_fast_assign(
+		__entry->rcuname = rcuname;
+		__entry->gpnum = gpnum;
+		__entry->completed = completed;
+		__entry->c = c;
+		__entry->level = level;
+		__entry->grplo = grplo;
+		__entry->grphi = grphi;
+		__entry->gpevent = gpevent;
+	),
+
+	TP_printk("%s %lu %lu %lu %u %d %d %s",
+		  __entry->rcuname, __entry->gpnum, __entry->completed,
+		  __entry->c, __entry->level, __entry->grplo, __entry->grphi,
+		  __entry->gpevent)
+);
+
+/*
  * Tracepoint for grace-period-initialization events.  These are
  * distinguished by the type of RCU, the new grace-period number, the
  * rcu_node structure level, the starting and ending CPU covered by the

@@ -593,6 +645,9 @@ TRACE_EVENT(rcu_barrier,
 #define trace_rcu_grace_period(rcuname, gpnum, gpevent) do { } while (0)
 #define trace_rcu_grace_period_init(rcuname, gpnum, level, grplo, grphi, \
 				    qsmask) do { } while (0)
+#define trace_rcu_nocb_grace_period(rcuname, gpnum, completed, c, \
+				    level, grplo, grphi, event) \
+				    do { } while (0)
 #define trace_rcu_preempt_task(rcuname, pid, gpnum) do { } while (0)
 #define trace_rcu_unlock_preempted_task(rcuname, gpnum, pid) do { } while (0)
 #define trace_rcu_quiescent_state_report(rcuname, gpnum, mask, qsmask, level, \

index e9dce4f..8b110fa 100644
@@ -316,6 +316,8 @@ cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
 
 	if (rcu_gp_in_progress(rsp))
 		return 0;  /* No, a grace period is already in progress. */
+	if (rcu_nocb_needs_gp(rsp))
+		return 1;  /* Yes, a no-CBs CPU needs one. */
 	if (!rdp->nxttail[RCU_NEXT_TAIL])
 		return 0;  /* No, this is a no-CBs (or offline) CPU. */
 	if (*rdp->nxttail[RCU_NEXT_READY_TAIL])

@@ -1400,6 +1402,7 @@ int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in)
 static void rcu_gp_cleanup(struct rcu_state *rsp)
 {
 	unsigned long gp_duration;
+	int nocb = 0;
 	struct rcu_data *rdp;
 	struct rcu_node *rnp = rcu_get_root(rsp);
 

@@ -1430,11 +1433,13 @@ static void rcu_gp_cleanup(struct rcu_state *rsp)
 	rcu_for_each_node_breadth_first(rsp, rnp) {
 		raw_spin_lock_irq(&rnp->lock);
 		rnp->completed = rsp->gpnum;
+		nocb += rcu_nocb_gp_cleanup(rsp, rnp);
 		raw_spin_unlock_irq(&rnp->lock);
 		cond_resched();
 	}
 	rnp = rcu_get_root(rsp);
 	raw_spin_lock_irq(&rnp->lock);
+	rcu_nocb_gp_set(rnp, nocb);
 
 	rsp->completed = rsp->gpnum; /* Declare grace period done. */
 	trace_rcu_grace_period(rsp->name, rsp->completed, "end");

@@ -2951,7 +2956,6 @@ static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
 	struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
 	struct rcu_node *rnp = rdp->mynode;
 	struct rcu_state *rsp;
-	int ret = NOTIFY_OK;
 
 	trace_rcu_utilization("Start CPU hotplug");
 	switch (action) {

@@ -2965,10 +2969,7 @@ static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
 		rcu_boost_kthread_setaffinity(rnp, -1);
 		break;
 	case CPU_DOWN_PREPARE:
-		if (nocb_cpu_expendable(cpu))
-			rcu_boost_kthread_setaffinity(rnp, cpu);
-		else
-			ret = NOTIFY_BAD;
+		rcu_boost_kthread_setaffinity(rnp, cpu);
 		break;
 	case CPU_DYING:
 	case CPU_DYING_FROZEN:

@@ -2992,7 +2993,7 @@ static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
 		break;
 	}
 	trace_rcu_utilization("End CPU hotplug");
-	return ret;
+	return NOTIFY_OK;
 }
 
 /*

@@ -3123,6 +3124,7 @@ static void __init rcu_init_one(struct rcu_state *rsp,
 			}
 			rnp->level = i;
 			INIT_LIST_HEAD(&rnp->blkd_tasks);
+			rcu_init_one_nocb(rnp);
 		}
 	}
 

@@ -3208,7 +3210,6 @@ void __init rcu_init(void)
 	rcu_init_one(&rcu_sched_state, &rcu_sched_data);
 	rcu_init_one(&rcu_bh_state, &rcu_bh_data);
 	__rcu_init_preempt();
-	rcu_init_nocb();
 	 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
 
 	/*

index c9f362e..ef26eab 100644
@@ -200,6 +200,12 @@ struct rcu_node {
 				/* Refused to boost: not sure why, though. */
 				/*  This can happen due to race conditions. */
 #endif /* #ifdef CONFIG_RCU_BOOST */
+#ifdef CONFIG_RCU_NOCB_CPU
+	wait_queue_head_t nocb_gp_wq[2];
+				/* Place for rcu_nocb_kthread() to wait GP. */
+	int n_nocb_gp_requests[2];
+				/* Counts of upcoming no-CB GP requests. */
+#endif /* #ifdef CONFIG_RCU_NOCB_CPU */
 	raw_spinlock_t fqslock ____cacheline_internodealigned_in_smp;
 } ____cacheline_internodealigned_in_smp;
 

@@ -384,12 +390,6 @@ struct rcu_state {
 	struct rcu_data __percpu *rda;		/* pointer of percu rcu_data. */
 	void (*call)(struct rcu_head *head,	/* call_rcu() flavor. */
 		     void (*func)(struct rcu_head *head));
-#ifdef CONFIG_RCU_NOCB_CPU
-	void (*call_remote)(struct rcu_head *head,
-		     void (*func)(struct rcu_head *head));
-						/* call_rcu() flavor, but for */
-						/*  placing on remote CPU. */
-#endif /* #ifdef CONFIG_RCU_NOCB_CPU */
 
 	/* The following fields are guarded by the root rcu_node's lock. */
 

@@ -538,16 +538,18 @@ static void print_cpu_stall_info(struct rcu_state *rsp, int cpu);
 static void print_cpu_stall_info_end(void);
 static void zero_cpu_stall_ticks(struct rcu_data *rdp);
 static void increment_cpu_stall_ticks(void);
+static int rcu_nocb_needs_gp(struct rcu_state *rsp);
+static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq);
+static int rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp);
+static void rcu_init_one_nocb(struct rcu_node *rnp);
 static bool is_nocb_cpu(int cpu);
 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
 			    bool lazy);
 static bool rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
 				      struct rcu_data *rdp);
-static bool nocb_cpu_expendable(int cpu);
 static void rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp);
 static void rcu_spawn_nocb_kthreads(struct rcu_state *rsp);
 static void init_nocb_callback_list(struct rcu_data *rdp);
-static void __init rcu_init_nocb(void);
 
 #endif /* #ifndef RCU_TREE_NONCORE */
 

339 lines not on this page · raw message

index 37750bc..eb9b473 100644
@@ -86,6 +86,19 @@ static void __init rcu_bootup_announce_oddness(void)
 	if (nr_cpu_ids != NR_CPUS)
 		printk(KERN_INFO "\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%d.\n", NR_CPUS, nr_cpu_ids);
 #ifdef CONFIG_RCU_NOCB_CPU
+#if CONFIG_RCU_NOCB_CPU_DEFAULT != 0
+	if (!have_rcu_nocb_mask) {
+		alloc_bootmem_cpumask_var(&rcu_nocb_mask);
+		have_rcu_nocb_mask = true;
+	}
+#if CONFIG_RCU_NOCB_CPU_DEFAULT == 1
+	pr_info("\tExperimental no-CBs CPU 0\n");
+	cpumask_set_cpu(0, rcu_nocb_mask);
+#else /* #if CONFIG_RCU_NOCB_CPU_DEFAULT == 1 */
+	pr_info("\tExperimental no-CBs for all CPUs\n");
+	cpumask_setall(rcu_nocb_mask);
+#endif /* #else #if CONFIG_RCU_NOCB_CPU_DEFAULT == 1 */
+#endif /* #if CONFIG_RCU_NOCB_CPU_DEFAULT != 0 */
 	if (have_rcu_nocb_mask) {
 		cpulist_scnprintf(nocb_buf, sizeof(nocb_buf), rcu_nocb_mask);
 		pr_info("\tExperimental no-CBs CPUs: %s.\n", nocb_buf);

index 8b110fa..4ec797e 100644
@@ -64,7 +64,7 @@
 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
 
-#define RCU_STATE_INITIALIZER(sname, cr) { \
+#define RCU_STATE_INITIALIZER(sname, sabbr, cr) { \
 	.level = { &sname##_state.node[0] }, \
 	.call = cr, \
 	.fqs_state = RCU_GP_IDLE, \

@@ -76,13 +76,14 @@ static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
 	.barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
 	.onoff_mutex = __MUTEX_INITIALIZER(sname##_state.onoff_mutex), \
 	.name = #sname, \
+	.abbr = sabbr, \
 }
 
 struct rcu_state rcu_sched_state =
-	RCU_STATE_INITIALIZER(rcu_sched, call_rcu_sched);
+	RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched);
 DEFINE_PER_CPU(struct rcu_data, rcu_sched_data);
 
-struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh, call_rcu_bh);
+struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh);
 DEFINE_PER_CPU(struct rcu_data, rcu_bh_data);
 
 static struct rcu_state *rcu_state;

index ef26eab..c865117 100644
@@ -452,6 +452,7 @@ struct rcu_state {
 	unsigned long gp_max;			/* Maximum GP duration in */
 						/*  jiffies. */
 	char *name;				/* Name of structure. */
+	char abbr;				/* Abbreviated name. */
 	struct list_head flavors;		/* List of RCU flavors. */
 };
 

index eb9b473..ab1bdde 100644
@@ -111,7 +111,7 @@ static void __init rcu_bootup_announce_oddness(void)
 #ifdef CONFIG_TREE_PREEMPT_RCU
 
 struct rcu_state rcu_preempt_state =
-	RCU_STATE_INITIALIZER(rcu_preempt, call_rcu);
+	RCU_STATE_INITIALIZER(rcu_preempt, 'p', call_rcu);
 DEFINE_PER_CPU(struct rcu_data, rcu_preempt_data);
 static struct rcu_state *rcu_state = &rcu_preempt_state;
 

@@ -2510,7 +2510,8 @@ static void __init rcu_spawn_nocb_kthreads(struct rcu_state *rsp)
 		return;
 	for_each_cpu(cpu, rcu_nocb_mask) {
 		rdp = per_cpu_ptr(rsp->rda, cpu);
-		t = kthread_run(rcu_nocb_kthread, rdp, "rcuo%d", cpu);
+		t = kthread_run(rcu_nocb_kthread, rdp,
+				"rcuo%d%c", cpu, rsp->abbr);
 		BUG_ON(IS_ERR(t));
 		ACCESS_ONCE(rdp->nocb_kthread) = t;
 	}

index ab1bdde..0997e9f 100644
@@ -1617,6 +1617,15 @@ static void rcu_idle_count_callbacks_posted(void)
 #define RCU_IDLE_GP_DELAY 4		/* Roughly one grace period. */
 #define RCU_IDLE_LAZY_GP_DELAY (6 * HZ)	/* Roughly six seconds. */
 
+static int rcu_idle_flushes = RCU_IDLE_FLUSHES;
+module_param(rcu_idle_flushes, int, 0644);
+static int rcu_idle_opt_flushes = RCU_IDLE_OPT_FLUSHES;
+module_param(rcu_idle_opt_flushes, int, 0644);
+static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY;
+module_param(rcu_idle_gp_delay, int, 0644);
+static int rcu_idle_lazy_gp_delay = RCU_IDLE_LAZY_GP_DELAY;
+module_param(rcu_idle_lazy_gp_delay, int, 0644);
+
 extern int tick_nohz_enabled;
 
 /*

@@ -1696,10 +1705,10 @@ int rcu_needs_cpu(int cpu, unsigned long *delta_jiffies)
 	}
 	/* Set up for the possibility that RCU will post a timer. */
 	if (rcu_cpu_has_nonlazy_callbacks(cpu)) {
-		*delta_jiffies = round_up(RCU_IDLE_GP_DELAY + jiffies,
-					  RCU_IDLE_GP_DELAY) - jiffies;
+		*delta_jiffies = round_up(rcu_idle_gp_delay + jiffies,
+					  rcu_idle_gp_delay) - jiffies;
 	} else {
-		*delta_jiffies = jiffies + RCU_IDLE_LAZY_GP_DELAY;
+		*delta_jiffies = jiffies + rcu_idle_lazy_gp_delay;
 		*delta_jiffies = round_jiffies(*delta_jiffies) - jiffies;
 	}
 	return 0;

@@ -1805,11 +1814,11 @@ static void rcu_prepare_for_idle(int cpu)
 		if (rcu_cpu_has_nonlazy_callbacks(cpu)) {
 			trace_rcu_prep_idle("User dyntick with callbacks");
 			rdtp->idle_gp_timer_expires =
-				round_up(jiffies + RCU_IDLE_GP_DELAY,
-					 RCU_IDLE_GP_DELAY);
+				round_up(jiffies + rcu_idle_gp_delay,
+					 rcu_idle_gp_delay);
 		} else if (rcu_cpu_has_callbacks(cpu)) {
 			rdtp->idle_gp_timer_expires =
-				round_jiffies(jiffies + RCU_IDLE_LAZY_GP_DELAY);
+				round_jiffies(jiffies + rcu_idle_lazy_gp_delay);
 			trace_rcu_prep_idle("User dyntick with lazy callbacks");
 		} else {
 			return;

@@ -1861,8 +1870,8 @@ static void rcu_prepare_for_idle(int cpu)
 	/* Check and update the ->dyntick_drain sequencing. */
 	if (rdtp->dyntick_drain <= 0) {
 		/* First time through, initialize the counter. */
-		rdtp->dyntick_drain = RCU_IDLE_FLUSHES;
-	} else if (rdtp->dyntick_drain <= RCU_IDLE_OPT_FLUSHES &&
+		rdtp->dyntick_drain = rcu_idle_flushes;
+	} else if (rdtp->dyntick_drain <= rcu_idle_opt_flushes &&
 		   !rcu_pending(cpu) &&
 		   !local_softirq_pending()) {
 		/* Can we go dyntick-idle despite still having callbacks? */

@@ -1871,11 +1880,11 @@ static void rcu_prepare_for_idle(int cpu)
 		if (rcu_cpu_has_nonlazy_callbacks(cpu)) {
 			trace_rcu_prep_idle("Dyntick with callbacks");
 			rdtp->idle_gp_timer_expires =
-				round_up(jiffies + RCU_IDLE_GP_DELAY,
-					 RCU_IDLE_GP_DELAY);
+				round_up(jiffies + rcu_idle_gp_delay,
+					 rcu_idle_gp_delay);
 		} else {
 			rdtp->idle_gp_timer_expires =
-				round_jiffies(jiffies + RCU_IDLE_LAZY_GP_DELAY);
+				round_jiffies(jiffies + rcu_idle_lazy_gp_delay);
 			trace_rcu_prep_idle("Dyntick with lazy callbacks");
 		}
 		tp = &rdtp->idle_gp_timer;

index 4ec797e..392c977 100644
@@ -1434,6 +1434,9 @@ static void rcu_gp_cleanup(struct rcu_state *rsp)
 	rcu_for_each_node_breadth_first(rsp, rnp) {
 		raw_spin_lock_irq(&rnp->lock);
 		rnp->completed = rsp->gpnum;
+		rdp = this_cpu_ptr(rsp->rda);
+		if (rnp == rdp->mynode)
+			__rcu_process_gp_end(rsp, rnp, rdp);
 		nocb += rcu_nocb_gp_cleanup(rsp, rnp);
 		raw_spin_unlock_irq(&rnp->lock);
 		cond_resched();

@@ -1446,6 +1449,7 @@ static void rcu_gp_cleanup(struct rcu_state *rsp)
 	trace_rcu_grace_period(rsp->name, rsp->completed, "end");
 	rsp->fqs_state = RCU_GP_IDLE;
 	rdp = this_cpu_ptr(rsp->rda);
+	rcu_advance_cbs(rsp, rnp, rdp);  /* Reduce false positives below. */
 	if (cpu_needs_another_gp(rsp, rdp))
 		rsp->gp_flags = 1;
 	raw_spin_unlock_irq(&rnp->lock);

@@ -1535,6 +1539,15 @@ rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
 	struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
 	struct rcu_node *rnp = rcu_get_root(rsp);
 
+	/*
+	 * If there is no grace period in progress right now, any
+	 * callbacks we have up to this point will be satisfied by the
+	 * next grace period.  Also, advancing the callbacks reduces the
+	 * probability of false positives from cpu_needs_another_gp()
+	 * resulting in pointless grace periods.  So, advance callbacks!
+	 */
+	rcu_advance_cbs(rsp, rnp, rdp);
+
 	if (!rsp->gp_kthread ||
 	    !cpu_needs_another_gp(rsp, rdp)) {
 		/*

@@ -1547,14 +1560,6 @@ rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
 		return;
 	}
 
-	/*
-	 * Because there is no grace period in progress right now,
-	 * any callbacks we have up to this point will be satisfied
-	 * by the next grace period.  So this is a good place to
-	 * assign a grace period number to recently posted callbacks.
-	 */
-	rcu_accelerate_cbs(rsp, rnp, rdp);
-
 	rsp->gp_flags = RCU_GP_FLAG_INIT;
 	raw_spin_unlock(&rnp->lock); /* Interrupts remain disabled. */
 

index 275aa3f..4b37b50 100644
@@ -73,6 +73,7 @@ extern void do_trace_rcu_torture_read(char *rcutorturename,
 #define UINT_CMP_LT(a, b)	(UINT_MAX / 2 < (a) - (b))
 #define ULONG_CMP_GE(a, b)	(ULONG_MAX / 2 >= (a) - (b))
 #define ULONG_CMP_LT(a, b)	(ULONG_MAX / 2 < (a) - (b))
+#define ulong2long(a)		(*(long *)(&(a)))
 
 /* Exported common interfaces */
 

index 392c977..4b8d91c 100644
@@ -2678,19 +2678,27 @@ static int rcu_pending(int cpu)
 }
 
 /*
- * Check to see if any future RCU-related work will need to be done
- * by the current CPU, even if none need be done immediately, returning
- * 1 if so.
+ * Return true if the specified CPU has any callback.  If all_lazy is
+ * non-NULL, store an indication of whether all callbacks are lazy.
+ * (If there are no callbacks, all of them are deemed to be lazy.)
  */
-static int rcu_cpu_has_callbacks(int cpu)
+static int rcu_cpu_has_callbacks(int cpu, bool *all_lazy)
 {
+	bool al = true;
+	bool hc = false;
+	struct rcu_data *rdp;
 	struct rcu_state *rsp;
 
-	/* RCU callbacks either ready or pending? */
-	for_each_rcu_flavor(rsp)
-		if (per_cpu_ptr(rsp->rda, cpu)->nxtlist)
-			return 1;
-	return 0;
+	for_each_rcu_flavor(rsp) {
+		rdp = per_cpu_ptr(rsp->rda, cpu);
+		if (rdp->qlen != rdp->qlen_lazy)
+			al = false;
+		if (rdp->nxtlist)
+			hc = true;
+	}
+	if (all_lazy)
+		*all_lazy = al;
+	return hc;
 }
 
 /*

@@ -2912,7 +2920,6 @@ rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
 	rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
 	atomic_set(&rdp->dynticks->dynticks,
 		   (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
-	rcu_prepare_for_idle_init(cpu);
 	raw_spin_unlock(&rnp->lock);		/* irqs remain disabled. */
 
 	/* Add CPU to rcu_node bitmasks. */

@@ -2986,7 +2993,6 @@ static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
 		 */
 		for_each_rcu_flavor(rsp)
 			rcu_cleanup_dying_cpu(rsp);
-		rcu_cleanup_after_idle(cpu);
 		break;
 	case CPU_DEAD:
 	case CPU_DEAD_FROZEN:

index c865117..282b1d7 100644
@@ -88,18 +88,13 @@ struct rcu_dynticks {
 	int dynticks_nmi_nesting;   /* Track NMI nesting level. */
 	atomic_t dynticks;	    /* Even value for idle, else odd. */
 #ifdef CONFIG_RCU_FAST_NO_HZ
-	int dyntick_drain;	    /* Prepare-for-idle state variable. */
-	unsigned long dyntick_holdoff;
-				    /* No retries for the jiffy of failure. */
-	struct timer_list idle_gp_timer;
-				    /* Wake up CPU sleeping with callbacks. */
-	unsigned long idle_gp_timer_expires;
-				    /* When to wake up CPU (for repost). */
-	bool idle_first_pass;	    /* First pass of attempt to go idle? */
+	bool all_lazy;		    /* Are all CPU's CBs lazy? */
 	unsigned long nonlazy_posted;
 				    /* # times non-lazy CBs posted to CPU. */
 	unsigned long nonlazy_posted_snap;
 				    /* idle-period nonlazy_posted snapshot. */
+	unsigned long last_accelerate;
+				    /* Last jiffy CBs were accelerated. */
 	int tick_nohz_enabled_snap; /* Previously seen value from sysfs. */
 #endif /* #ifdef CONFIG_RCU_FAST_NO_HZ */
 #ifdef CONFIG_RCU_USER_QS

@@ -530,7 +525,6 @@ static int __cpuinit rcu_spawn_one_boost_kthread(struct rcu_state *rsp,
 						 struct rcu_node *rnp);
 #endif /* #ifdef CONFIG_RCU_BOOST */
 static void __cpuinit rcu_prepare_kthreads(int cpu);
-static void rcu_prepare_for_idle_init(int cpu);
 static void rcu_cleanup_after_idle(int cpu);
 static void rcu_prepare_for_idle(int cpu);
 static void rcu_idle_count_callbacks_posted(void);

439 lines not on this page · raw message

index 0d095dc..49099e8 100644
@@ -46,8 +46,6 @@
 #define RCU_TREE_NONCORE
 #include "rcutree.h"
 
-#define ulong2long(a) (*(long *)(&(a)))
-
 static int r_open(struct inode *inode, struct file *file,
 					const struct seq_operations *op)
 {

index 4b8d91c..7207435 100644
@@ -1559,16 +1559,14 @@ rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
 		raw_spin_unlock_irqrestore(&rnp->lock, flags);
 		return;
 	}
-
 	rsp->gp_flags = RCU_GP_FLAG_INIT;
-	raw_spin_unlock(&rnp->lock); /* Interrupts remain disabled. */
 
 	/* Ensure that CPU is aware of completion of last grace period. */
-	rcu_process_gp_end(rsp, rdp);
-	local_irq_restore(flags);
+	__rcu_process_gp_end(rsp, rdp->mynode, rdp);
 
 	/* Wake up rcu_gp_kthread() to start the grace period. */
 	wake_up(&rsp->gp_wq);
+	raw_spin_unlock_irqrestore(&rnp->lock, flags);
 }
 
 /*

index ef0bf31..0dc0177 100644
@@ -72,10 +72,10 @@ TRACE_EVENT(rcu_grace_period,
 );
 
 /*
- * Tracepoint for no-callbacks grace-period events.  The caller should
- * pull the data from the rcu_node structure, other than rcuname, which
- * comes from the rcu_state structure, and event, which is one of the
- * following:
+ * Tracepoint for future grace-period events, including those for no-callbacks
+ * CPUs.  The caller should pull the data from the rcu_node structure,
+ * other than rcuname, which comes from the rcu_state structure, and event,
+ * which is one of the following:
  *
  * "Startleaf": Request a nocb grace period based on leaf-node data.
  * "Startedleaf": Leaf-node start proved sufficient.

@@ -87,7 +87,7 @@ TRACE_EVENT(rcu_grace_period,
  * "Cleanup": Clean up rcu_node structure after previous GP.
  * "CleanupMore": Clean up, and another no-CB GP is needed.
  */
-TRACE_EVENT(rcu_nocb_grace_period,
+TRACE_EVENT(rcu_future_grace_period,
 
 	TP_PROTO(char *rcuname, unsigned long gpnum, unsigned long completed,
 		 unsigned long c, u8 level, int grplo, int grphi,

@@ -645,9 +645,9 @@ TRACE_EVENT(rcu_barrier,
 #define trace_rcu_grace_period(rcuname, gpnum, gpevent) do { } while (0)
 #define trace_rcu_grace_period_init(rcuname, gpnum, level, grplo, grphi, \
 				    qsmask) do { } while (0)
-#define trace_rcu_nocb_grace_period(rcuname, gpnum, completed, c, \
-				    level, grplo, grphi, event) \
-				    do { } while (0)
+#define trace_rcu_future_grace_period(rcuname, gpnum, completed, c, \
+				      level, grplo, grphi, event) \
+				      do { } while (0)
 #define trace_rcu_preempt_task(rcuname, pid, gpnum) do { } while (0)
 #define trace_rcu_unlock_preempted_task(rcuname, gpnum, pid) do { } while (0)
 #define trace_rcu_quiescent_state_report(rcuname, gpnum, mask, qsmask, level, \

105 lines not on this page · raw message

index 7207435..8ca18ec 100644
@@ -1525,16 +1525,14 @@ static int __noreturn rcu_gp_kthread(void *arg)
 /*
  * Start a new RCU grace period if warranted, re-initializing the hierarchy
  * 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.
+ * the root node's ->lock and 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)
-	__releases(rcu_get_root(rsp)->lock)
+rcu_start_gp(struct rcu_state *rsp)
 {
 	struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
 	struct rcu_node *rnp = rcu_get_root(rsp);

@@ -1548,15 +1546,13 @@ rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
 	 */
 	rcu_advance_cbs(rsp, rnp, rdp);
 
-	if (!rsp->gp_kthread ||
-	    !cpu_needs_another_gp(rsp, rdp)) {
+	if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
 		/*
 		 * Either we have not yet spawned the grace-period
 		 * task, this CPU does not need another grace period,
 		 * or a grace period is already in progress.
 		 * Either way, don't start a new grace period.
 		 */
-		raw_spin_unlock_irqrestore(&rnp->lock, flags);
 		return;
 	}
 	rsp->gp_flags = RCU_GP_FLAG_INIT;

@@ -1566,15 +1562,14 @@ rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
 
 	/* Wake up rcu_gp_kthread() to start the grace period. */
 	wake_up(&rsp->gp_wq);
-	raw_spin_unlock_irqrestore(&rnp->lock, flags);
 }
 
 /*
  * Report a full set of quiescent states to the specified rcu_state
  * data structure.  This involves cleaning up after the prior grace
  * period and letting rcu_start_gp() start up the next grace period
- * if one is needed.  Note that the caller must hold rnp->lock, as
- * required by rcu_start_gp(), which will release it.
+ * if one is needed.  Note that the caller must hold rnp->lock, which
+ * is released before return.
  */
 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
 	__releases(rcu_get_root(rsp)->lock)

@@ -2172,7 +2167,8 @@ __rcu_process_callbacks(struct rcu_state *rsp)
 	local_irq_save(flags);
 	if (cpu_needs_another_gp(rsp, rdp)) {
 		raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */
-		rcu_start_gp(rsp, flags);  /* releases above lock */
+		rcu_start_gp(rsp);
+		raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
 	} else {
 		local_irq_restore(flags);
 	}

@@ -2252,11 +2248,11 @@ static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
 
 		/* Start a new grace period if one not already started. */
 		if (!rcu_gp_in_progress(rsp)) {
-			unsigned long nestflag;
 			struct rcu_node *rnp_root = rcu_get_root(rsp);
 
-			raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
-			rcu_start_gp(rsp, nestflag);  /* rlses rnp_root->lock */
+			raw_spin_lock(&rnp_root->lock);
+			rcu_start_gp(rsp);
+			raw_spin_lock(&rnp_root->lock);
 		} else {
 			/* Give the grace period a kick. */
 			rdp->blimit = LONG_MAX;

index d09acdf..736dd2c 100644
@@ -2213,7 +2213,6 @@ static void rcu_nocb_wait_gp(struct rcu_data *rdp)
 	unsigned long c;
 	bool d;
 	unsigned long flags;
-	unsigned long flags1;
 	struct rcu_node *rnp = rdp->mynode;
 	struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
 

@@ -2275,8 +2274,8 @@ static void rcu_nocb_wait_gp(struct rcu_data *rdp)
 						      c, rnp->level,
 						      rnp->grplo, rnp->grphi,
 						      "Startedroot");
-			local_save_flags(flags1);
-			rcu_start_gp(rdp->rsp, flags1); /* Rlses ->lock. */
+			rcu_start_gp(rdp->rsp);
+			raw_spin_unlock(&rnp->lock);
 		}
 
 		/* Clean up locking and irq state. */

index 282b1d7..775d96c 100644
@@ -198,9 +198,9 @@ struct rcu_node {
 #ifdef CONFIG_RCU_NOCB_CPU
 	wait_queue_head_t nocb_gp_wq[2];
 				/* Place for rcu_nocb_kthread() to wait GP. */
-	int n_nocb_gp_requests[2];
-				/* Counts of upcoming no-CB GP requests. */
 #endif /* #ifdef CONFIG_RCU_NOCB_CPU */
+	int need_future_gp[2];
+				/* Counts of upcoming no-CB GP requests. */
 	raw_spinlock_t fqslock ____cacheline_internodealigned_in_smp;
 } ____cacheline_internodealigned_in_smp;
 

index 736dd2c..e4037bd 100644
@@ -2057,7 +2057,7 @@ static int rcu_nocb_needs_gp(struct rcu_state *rsp)
 {
 	struct rcu_node *rnp = rcu_get_root(rsp);
 
-	return rnp->n_nocb_gp_requests[(ACCESS_ONCE(rnp->completed) + 1) & 0x1];
+	return rnp->need_future_gp[(ACCESS_ONCE(rnp->completed) + 1) & 0x1];
 }
 
 /*

@@ -2071,8 +2071,8 @@ static int rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
 	int needmore;
 
 	wake_up_all(&rnp->nocb_gp_wq[c & 0x1]);
-	rnp->n_nocb_gp_requests[c & 0x1] = 0;
-	needmore = rnp->n_nocb_gp_requests[(c + 1) & 0x1];
+	rnp->need_future_gp[c & 0x1] = 0;
+	needmore = rnp->need_future_gp[(c + 1) & 0x1];
 	trace_rcu_future_grace_period(rsp->name, rnp->gpnum, rnp->completed,
 				      c, rnp->level, rnp->grplo, rnp->grphi,
 				      needmore ? "CleanupMore" : "Cleanup");

@@ -2080,7 +2080,7 @@ static int rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
 }
 
 /*
- * Set the root rcu_node structure's ->n_nocb_gp_requests field
+ * Set the root rcu_node structure's ->need_future_gp field
  * based on the sum of those of all rcu_node structures.  This does
  * double-count the root rcu_node structure's requests, but this
  * is necessary to handle the possibility of a rcu_nocb_kthread()

@@ -2089,7 +2089,7 @@ static int rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
  */
 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
 {
-	rnp->n_nocb_gp_requests[(rnp->completed + 1) & 0x1] += nrq;
+	rnp->need_future_gp[(rnp->completed + 1) & 0x1] += nrq;
 }
 
 static void rcu_init_one_nocb(struct rcu_node *rnp)

@@ -2220,7 +2220,7 @@ static void rcu_nocb_wait_gp(struct rcu_data *rdp)
 	c = rnp->completed + 2;
 
 	/* Count our request for a grace period. */
-	rnp->n_nocb_gp_requests[c & 0x1]++;
+	rnp->need_future_gp[c & 0x1]++;
 	trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
 				      rnp->completed, c, rnp->level,
 				      rnp->grplo, rnp->grphi, "Startleaf");

@@ -2264,10 +2264,10 @@ static void rcu_nocb_wait_gp(struct rcu_data *rdp)
 			 * Adjust counters accordingly and start the
 			 * needed grace period.
 			 */
-			rnp->n_nocb_gp_requests[c & 0x1]--;
+			rnp->need_future_gp[c & 0x1]--;
 			c = rnp_root->completed + 1;
-			rnp->n_nocb_gp_requests[c & 0x1]++;
-			rnp_root->n_nocb_gp_requests[c & 0x1]++;
+			rnp->need_future_gp[c & 0x1]++;
+			rnp_root->need_future_gp[c & 0x1]++;
 			trace_rcu_future_grace_period(rdp->rsp->name,
 						      rnp->gpnum,
 						      rnp->completed,

index 8ca18ec..bd42feb 100644
@@ -230,6 +230,7 @@ static ulong jiffies_till_next_fqs = RCU_JIFFIES_TILL_FORCE_QS;
 module_param(jiffies_till_first_fqs, ulong, 0644);
 module_param(jiffies_till_next_fqs, ulong, 0644);
 
+static void rcu_start_gp(struct rcu_state *rsp);
 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *));
 static void force_quiescent_state(struct rcu_state *rsp);
 static int rcu_pending(int cpu);

@@ -1113,6 +1114,120 @@ static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
 }
 
 /*
+ * Trace-event helper function for rcu_start_future_gp() and
+ * rcu_nocb_wait_gp().
+ */
+static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
+				unsigned long c, char *s)
+{
+	trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
+				      rnp->completed, c, rnp->level,
+				      rnp->grplo, rnp->grphi, s);
+}
+
+/*
+ * Start some future grace period, as needed to handle newly arrived
+ * callbacks.  The required future grace periods are recorded in each
+ * rcu_node structure's ->need_future_gp field.
+ *
+ * The caller must hold the specified rcu_node structure's ->lock.
+ */
+static unsigned long __maybe_unused
+rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp)
+{
+	unsigned long c;
+	int i;
+	struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
+
+	/*
+	 * Pick up grace-period number for new callbacks.  If this
+	 * grace period is already marked as needed, return to the caller.
+	 */
+	c = rcu_cbs_completed(rdp->rsp, rnp);
+	trace_rcu_future_gp(rnp, rdp, c, "Startleaf");
+	if (rnp->need_future_gp[c & 0x1]) {
+		trace_rcu_future_gp(rnp, rdp, c, "Prestartleaf");
+		return c;
+	}
+
+	/*
+	 * If either this rcu_node structure or the root rcu_node structure
+	 * believe that a grace period is in progress, then we must wait
+	 * for the one following, which is in "c".  Because our request
+	 * will be noticed at the end of the current grace period, we don't
+	 * need to explicitly start one.
+	 */
+	if (rnp->gpnum != rnp->completed ||
+	    ACCESS_ONCE(rnp->gpnum) != ACCESS_ONCE(rnp->completed)) {
+		rnp->need_future_gp[c & 0x1]++;
+		trace_rcu_future_gp(rnp, rdp, c, "Startedleaf");
+		return c;
+	}
+
+	/*
+	 * There might be no grace period in progress.  If we don't already
+	 * hold it, acquire the root rcu_node structure's lock in order to
+	 * start one (if needed).
+	 */
+	if (rnp != rnp_root)
+		raw_spin_lock(&rnp_root->lock);
+
+	/*
+	 * Get a new grace-period number.  If there really is no grace
+	 * period in progress, it will be smaller than the one we obtained
+	 * earlier.  Adjust callbacks as needed.  Note that even no-CBs
+	 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
+	 */
+	c = rcu_cbs_completed(rdp->rsp, rnp_root);
+	for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++)
+		if (ULONG_CMP_LT(c, rdp->nxtcompleted[i]))
+			rdp->nxtcompleted[i] = c;
+
+	/*
+	 * If the needed for the required grace period is already
+	 * recorded, trace and leave.
+	 */
+	if (rnp_root->need_future_gp[c & 0x1]) {
+		trace_rcu_future_gp(rnp, rdp, c, "Prestartedroot");
+		goto unlock_out;
+	}
+
+	/* Record the need for the future grace period. */
+	rnp_root->need_future_gp[c & 0x1]++;
+
+	/* If a grace period is not already in progress, start one. */
+	if (rnp_root->gpnum != rnp_root->completed) {
+		trace_rcu_future_gp(rnp, rdp, c, "Startedleafroot");
+	} else {
+		trace_rcu_future_gp(rnp, rdp, c, "Startedroot");
+		rcu_start_gp(rdp->rsp);
+	}
+unlock_out:
+	if (rnp != rnp_root)
+		raw_spin_unlock(&rnp_root->lock);
+	return c;
+}
+
+/*
+ * Clean up any old requests for the just-ended grace period.  Also return
+ * whether any additional grace periods have been requested.  Also invoke
+ * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
+ * waiting for this grace period to complete.
+ */
+static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
+{
+	int c = rnp->completed;
+	int needmore;
+	struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
+
+	rcu_nocb_gp_cleanup(rsp, rnp);
+	rnp->need_future_gp[c & 0x1] = 0;
+	needmore = rnp->need_future_gp[(c + 1) & 0x1];
+	trace_rcu_future_gp(rnp, rdp, c, needmore ? "CleanupMore" : "Cleanup");
+	return needmore;
+}
+
+/*
  * If there is room, assign a ->completed number to any callbacks on
  * this CPU that have not already been assigned.  Also accelerate any
  * callbacks that were previously assigned a ->completed number that has

@@ -1350,9 +1465,9 @@ static int rcu_gp_init(struct rcu_state *rsp)
 		rdp = this_cpu_ptr(rsp->rda);
 		rcu_preempt_check_blocked_tasks(rnp);
 		rnp->qsmask = rnp->qsmaskinit;
-		rnp->gpnum = rsp->gpnum;
+		ACCESS_ONCE(rnp->gpnum) = rsp->gpnum;
 		WARN_ON_ONCE(rnp->completed != rsp->completed);
-		rnp->completed = rsp->completed;
+		ACCESS_ONCE(rnp->completed) = rsp->completed;
 		if (rnp == rdp->mynode)
 			rcu_start_gp_per_cpu(rsp, rnp, rdp);
 		rcu_preempt_boost_start_gp(rnp);

@@ -1433,11 +1548,11 @@ static void rcu_gp_cleanup(struct rcu_state *rsp)
 	 */
 	rcu_for_each_node_breadth_first(rsp, rnp) {
 		raw_spin_lock_irq(&rnp->lock);
-		rnp->completed = rsp->gpnum;
+		ACCESS_ONCE(rnp->completed) = rsp->gpnum;
 		rdp = this_cpu_ptr(rsp->rda);
 		if (rnp == rdp->mynode)
 			__rcu_process_gp_end(rsp, rnp, rdp);
-		nocb += rcu_nocb_gp_cleanup(rsp, rnp);
+		nocb += rcu_future_gp_cleanup(rsp, rnp);
 		raw_spin_unlock_irq(&rnp->lock);
 		cond_resched();
 	}

index 775d96c..d6a043e 100644
@@ -535,7 +535,7 @@ static void zero_cpu_stall_ticks(struct rcu_data *rdp);
 static void increment_cpu_stall_ticks(void);
 static int rcu_nocb_needs_gp(struct rcu_state *rsp);
 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq);
-static int rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp);
+static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp);
 static void rcu_init_one_nocb(struct rcu_node *rnp);
 static bool is_nocb_cpu(int cpu);
 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,

139 lines not on this page · raw message

index bd42feb..f7399b4 100644
@@ -230,7 +230,8 @@ static ulong jiffies_till_next_fqs = RCU_JIFFIES_TILL_FORCE_QS;
 module_param(jiffies_till_first_fqs, ulong, 0644);
 module_param(jiffies_till_next_fqs, ulong, 0644);
 
-static void rcu_start_gp(struct rcu_state *rsp);
+static void rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
+				  struct rcu_data *rdp);
 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *));
 static void force_quiescent_state(struct rcu_state *rsp);
 static int rcu_pending(int cpu);

@@ -1200,7 +1201,7 @@ rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp)
 		trace_rcu_future_gp(rnp, rdp, c, "Startedleafroot");
 	} else {
 		trace_rcu_future_gp(rnp, rdp, c, "Startedroot");
-		rcu_start_gp(rdp->rsp);
+		rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp);
 	}
 unlock_out:
 	if (rnp != rnp_root)

@@ -1286,6 +1287,8 @@ static void rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
 		rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL];
 		rdp->nxtcompleted[i] = c;
 	}
+	/* Record any needed additional grace periods. */
+	rcu_start_future_gp(rnp, rdp);
 
 	/* Trace depending on how much we were able to accelerate. */
 	if (!*rdp->nxttail[RCU_WAIT_TAIL])

@@ -1647,20 +1650,9 @@ static int __noreturn rcu_gp_kthread(void *arg)
  * quiescent state.
  */
 static void
-rcu_start_gp(struct rcu_state *rsp)
+rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
+		      struct rcu_data *rdp)
 {
-	struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
-	struct rcu_node *rnp = rcu_get_root(rsp);
-
-	/*
-	 * If there is no grace period in progress right now, any
-	 * callbacks we have up to this point will be satisfied by the
-	 * next grace period.  Also, advancing the callbacks reduces the
-	 * probability of false positives from cpu_needs_another_gp()
-	 * resulting in pointless grace periods.  So, advance callbacks!
-	 */
-	rcu_advance_cbs(rsp, rnp, rdp);
-
 	if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
 		/*
 		 * Either we have not yet spawned the grace-period

@@ -1672,14 +1664,36 @@ rcu_start_gp(struct rcu_state *rsp)
 	}
 	rsp->gp_flags = RCU_GP_FLAG_INIT;
 
-	/* Ensure that CPU is aware of completion of last grace period. */
-	__rcu_process_gp_end(rsp, rdp->mynode, rdp);
-
 	/* Wake up rcu_gp_kthread() to start the grace period. */
 	wake_up(&rsp->gp_wq);
 }
 
 /*
+ * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
+ * callbacks.  Note that rcu_start_gp_advanced() cannot do this because it
+ * is invoked indirectly from rcu_advance_cbs(), which would result in
+ * endless recursion -- or would do so if it wasn't for the self-deadlock
+ * that is encountered beforehand.
+ */
+static void
+rcu_start_gp(struct rcu_state *rsp)
+{
+	struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
+	struct rcu_node *rnp = rcu_get_root(rsp);
+
+	/*
+	 * If there is no grace period in progress right now, any
+	 * callbacks we have up to this point will be satisfied by the
+	 * next grace period.  Also, advancing the callbacks reduces the
+	 * probability of false positives from cpu_needs_another_gp()
+	 * resulting in pointless grace periods.  So, advance callbacks
+	 * then start the grace period!
+	 */
+	rcu_advance_cbs(rsp, rnp, rdp);
+	rcu_start_gp_advanced(rsp, rnp, rdp);
+}
+
+/*
  * Report a full set of quiescent states to the specified rcu_state
  * data structure.  This involves cleaning up after the prior grace
  * period and letting rcu_start_gp() start up the next grace period