[PATCH RFC] fs/buffer.c: update per-CPU bh_lru cache via RCU

[Marcelo Tosatti <mtosatti@redhat.com>]

umount causes invalidate_bh_lrus which calls an IPI on each 
CPU that has non empty per-CPU cache:

	on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1);

This interrupts CPUs which might be executing code sensitive 
to interferences.

To avoid the IPI, free the per-CPU caches remotely via RCU.
Two bh_lrus structures for each CPU are allocated: one is being
used (assigned to per-CPU bh_lru pointer), and the other is
being freed (or idle).

Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>

diff --git a/fs/buffer.c b/fs/buffer.c
index d9c6d1fbb6dd..9f9ed7fffec8 100644
--- a/fs/buffer.c
+++ b/fs/buffer.c
@@ -1203,7 +1203,21 @@ struct bh_lru {
 	struct buffer_head *bhs[BH_LRU_SIZE];
 };
 
-static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
+
+/*
+ * Allocate two bh_lrus structures for each CPU. bh_lru points to the
+ * one that is currently in use, and the update path does
+ * (consider cpu->bh_lru = bh_lrus[0]).
+ *
+ * cpu->bh_lru = bh_lrus[1]
+ * synchronize_rcu()
+ * free bh's in bh_lrus[0]
+ */
+unsigned int bh_lru_idx;
+static DEFINE_PER_CPU(struct bh_lru, bh_lrus[2]) = {{{ NULL }}, {{NULL}}};
+static DEFINE_PER_CPU(struct bh_lru, *bh_lru);
+
+static DEFINE_MUTEX(bh_lru_invalidate_mutex);
 
 #ifdef CONFIG_SMP
 #define bh_lru_lock()	local_irq_disable()
@@ -1245,16 +1259,19 @@ static void bh_lru_install(struct buffer_head *bh)
 		return;
 	}
 
-	b = this_cpu_ptr(&bh_lrus);
+	rcu_read_lock();
+	b = rcu_dereference(per_cpu(bh_lru, smp_processor_id()));
 	for (i = 0; i < BH_LRU_SIZE; i++) {
 		swap(evictee, b->bhs[i]);
 		if (evictee == bh) {
+			rcu_read_unlock();
 			bh_lru_unlock();
 			return;
 		}
 	}
 
 	get_bh(bh);
+	rcu_read_unlock();
 	bh_lru_unlock();
 	brelse(evictee);
 }
@@ -1266,28 +1283,32 @@ static struct buffer_head *
 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
 {
 	struct buffer_head *ret = NULL;
+	struct bh_lru *lru;
 	unsigned int i;
 
 	check_irqs_on();
 	bh_lru_lock();
+	rcu_read_lock();
+
+	lru = rcu_dereference(per_cpu(bh_lru, smp_processor_id()));
 	for (i = 0; i < BH_LRU_SIZE; i++) {
-		struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
+		struct buffer_head *bh = lru->bhs[i];
 
 		if (bh && bh->b_blocknr == block && bh->b_bdev == bdev &&
 		    bh->b_size == size) {
 			if (i) {
 				while (i) {
-					__this_cpu_write(bh_lrus.bhs[i],
-						__this_cpu_read(bh_lrus.bhs[i - 1]));
+					lru->bhs[i] = lru->bhs[i - 1];
 					i--;
 				}
-				__this_cpu_write(bh_lrus.bhs[0], bh);
+				lru->bhs[0] = bh;
 			}
 			get_bh(bh);
 			ret = bh;
 			break;
 		}
 	}
+	rcu_read_unlock();
 	bh_lru_unlock();
 	return ret;
 }
@@ -1381,35 +1402,56 @@ static void __invalidate_bh_lrus(struct bh_lru *b)
 		b->bhs[i] = NULL;
 	}
 }
-/*
- * invalidate_bh_lrus() is called rarely - but not only at unmount.
- * This doesn't race because it runs in each cpu either in irq
- * or with preempt disabled.
- */
-static void invalidate_bh_lru(void *arg)
-{
-	struct bh_lru *b = &get_cpu_var(bh_lrus);
-
-	__invalidate_bh_lrus(b);
-	put_cpu_var(bh_lrus);
-}
 
 bool has_bh_in_lru(int cpu, void *dummy)
 {
-	struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
+	struct bh_lru *b;
 	int i;
-	
+
+	rcu_read_lock();
+	b = rcu_dereference(per_cpu(bh_lru, cpu));
 	for (i = 0; i < BH_LRU_SIZE; i++) {
-		if (b->bhs[i])
+               if (b->bhs[i]) {
+			rcu_read_unlock();
 			return true;
+		}
 	}
 
+	rcu_read_unlock();
 	return false;
 }
 
+/*
+ * invalidate_bh_lrus() is called rarely - but not only at unmount.
+ */
 void invalidate_bh_lrus(void)
 {
-	on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1);
+	int cpu, oidx, nidx;
+
+	mutex_lock(&bh_lru_invalidate_mutex);
+	oidx = bh_lru_idx;
+	bh_lru_idx++;
+	if (bh_lru_idx >= 2)
+		bh_lru_idx = 0;
+
+	nidx = bh_lru_idx;
+	/* Assign the per-CPU bh_lru pointer */
+	cpus_read_lock();
+	for_each_online_cpu(cpu)
+		rcu_assign_pointer(per_cpu(bh_lru, cpu), per_cpu_ptr(&bh_lrus[nidx], cpu));
+	cpus_read_unlock();
+	synchronize_rcu_expedited();
+
+	cpus_read_lock();
+	for_each_online_cpu(cpu) {
+		struct bh_lru *b = per_cpu_ptr(&bh_lrus[oidx], cpu);
+
+		bh_lru_lock();
+		__invalidate_bh_lrus(b);
+		bh_lru_unlock();
+	}
+	cpus_read_unlock();
+	mutex_unlock(&bh_lru_invalidate_mutex);
 }
 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
 
@@ -1422,8 +1464,10 @@ void invalidate_bh_lrus_cpu(void)
 	struct bh_lru *b;
 
 	bh_lru_lock();
-	b = this_cpu_ptr(&bh_lrus);
+	rcu_read_lock();
+	b = rcu_dereference(per_cpu(bh_lru, smp_processor_id()));
 	__invalidate_bh_lrus(b);
+	rcu_read_unlock();
 	bh_lru_unlock();
 }
 
@@ -2923,12 +2967,15 @@ EXPORT_SYMBOL(free_buffer_head);
 static int buffer_exit_cpu_dead(unsigned int cpu)
 {
 	int i;
-	struct bh_lru *b = &per_cpu(bh_lrus, cpu);
+	struct bh_lru *b;
 
+	rcu_read_lock();
+	b = rcu_dereference(per_cpu(bh_lru, cpu));
 	for (i = 0; i < BH_LRU_SIZE; i++) {
 		brelse(b->bhs[i]);
 		b->bhs[i] = NULL;
 	}
+	rcu_read_unlock();
 	this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
 	per_cpu(bh_accounting, cpu).nr = 0;
 	return 0;
@@ -3021,7 +3068,7 @@ EXPORT_SYMBOL(__bh_read_batch);
 void __init buffer_init(void)
 {
 	unsigned long nrpages;
-	int ret;
+	int ret, cpu;
 
 	bh_cachep = kmem_cache_create("buffer_head",
 			sizeof(struct buffer_head), 0,
@@ -3029,6 +3076,11 @@ void __init buffer_init(void)
 				SLAB_MEM_SPREAD),
 				NULL);
 
+	cpus_read_lock();
+	for_each_online_cpu(cpu)
+		per_cpu(bh_lru, cpu) = per_cpu_ptr(&bh_lrus[0], cpu);
+	cpus_read_unlock();
+
 	/*
 	 * Limit the bh occupancy to 10% of ZONE_NORMAL
 	 */