More aggsum optimizations.

 - Avoid atomic_add() when updating as_lower_bound/as_upper_bound.
Previous code was excessively strong on 64bit systems while not
strong enough on 32bit ones.  Instead introduce and use real
atomic_load() and atomic_store() operations, just an assignments
on 64bit machines, but using proper atomics on 32bit ones to avoid
torn reads/writes.

 - Reduce number of buckets on large systems.  Extra buckets not as
much improve add speed, as hurt reads.  Unlike wmsum for aggsum
reads are still important.

Signed-off-by: Alexander Motin <[email protected]>
Sponsored-By: iXsystems, Inc.

include/os/linux/spl/sys/atomic.h

@@ -48,6 +48,8 @@
 #define	atomic_sub_32_nv(v, i)	atomic_sub_return((i), (atomic_t *)(v))
 #define	atomic_cas_32(v, x, y)	atomic_cmpxchg((atomic_t *)(v), x, y)
 #define	atomic_swap_32(v, x)	atomic_xchg((atomic_t *)(v), x)
+#define	atomic_load_32(v)	atomic_read((atomic_t *)(v))
+#define	atomic_store_32(v, x)	atomic_set((atomic_t *)(v), x)
 #define	atomic_inc_64(v)	atomic64_inc((atomic64_t *)(v))
 #define	atomic_dec_64(v)	atomic64_dec((atomic64_t *)(v))
 #define	atomic_add_64(v, i)	atomic64_add((i), (atomic64_t *)(v))
@@ -58,6 +60,8 @@
 #define	atomic_sub_64_nv(v, i)	atomic64_sub_return((i), (atomic64_t *)(v))
 #define	atomic_cas_64(v, x, y)	atomic64_cmpxchg((atomic64_t *)(v), x, y)
 #define	atomic_swap_64(v, x)	atomic64_xchg((atomic64_t *)(v), x)
+#define	atomic_load_64(v)	atomic64_read((atomic64_t *)(v))
+#define	atomic_store_64(v, x)	atomic64_set((atomic64_t *)(v), x)
 
 #ifdef _LP64
 static __inline__ void *

include/sys/aggsum.h

@@ -39,15 +39,16 @@ struct aggsum_bucket {
 typedef struct aggsum {
 	kmutex_t as_lock;
 	int64_t as_lower_bound;
-	int64_t as_upper_bound;
+	uint64_t as_upper_bound;
+	aggsum_bucket_t *as_buckets ____cacheline_aligned;
 	uint_t as_numbuckets;
-	aggsum_bucket_t *as_buckets;
+	uint_t as_bucketshift;
 } aggsum_t;
 
 void aggsum_init(aggsum_t *, uint64_t);
 void aggsum_fini(aggsum_t *);
 int64_t aggsum_lower_bound(aggsum_t *);
-int64_t aggsum_upper_bound(aggsum_t *);
+uint64_t aggsum_upper_bound(aggsum_t *);
 int aggsum_compare(aggsum_t *, uint64_t);
 uint64_t aggsum_value(aggsum_t *);
 void aggsum_add(aggsum_t *, int64_t);

lib/libspl/atomic.c

@@ -313,6 +313,19 @@ atomic_swap_ptr(volatile void *target, void *bits)
 	return (__atomic_exchange_n((void **)target, bits, __ATOMIC_SEQ_CST));
 }
 
+#ifndef _LP64
+uint64_t
+atomic_load_64(volatile uint64_t *target)
+{
+	return (__atomic_load_n(target, __ATOMIC_RELAXED));
+}
+
+void
+atomic_store_64(volatile uint64_t *target, uint64_t bits)
+{
+	return (__atomic_store_n(target, bits, __ATOMIC_RELAXED));
+}
+#endif
 
 int
 atomic_set_long_excl(volatile ulong_t *target, uint_t value)

lib/libspl/include/atomic.h

@@ -245,6 +245,49 @@ extern ulong_t atomic_swap_ulong(volatile ulong_t *, ulong_t);
 extern uint64_t atomic_swap_64(volatile uint64_t *, uint64_t);
 #endif
 
+/*
+ * Atomically read variable.
+ */
+#define	atomic_load_char(p)	(*(volatile uchar_t *)(p))
+#define	atomic_load_short(p)	(*(volatile ushort_t *)(p))
+#define	atomic_load_int(p)	(*(volatile uint_t *)(p))
+#define	atomic_load_long(p)	(*(volatile ulong_t *)(p))
+#define	atomic_load_ptr(p)	(*(volatile __typeof(*p) *)(p))
+#define	atomic_load_8(p)	(*(volatile uint8_t *)(p))
+#define	atomic_load_16(p)	(*(volatile uint16_t *)(p))
+#define	atomic_load_32(p)	(*(volatile uint32_t *)(p))
+#ifdef _LP64
+#define	atomic_load_64(p)	(*(volatile uint64_t *)(p))
+#elif defined(_INT64_TYPE)
+extern uint64_t atomic_load_64(volatile uint64_t *);
+#endif
+
+/*
+ * Atomically write variable.
+ */
+#define	atomic_store_char(p, v)		\
+	(*(volatile uchar_t *)(p) = (uchar_t)(v))
+#define	atomic_store_short(p, v)	\
+	(*(volatile ushort_t *)(p) = (ushort_t)(v))
+#define	atomic_store_int(p, v)		\
+	(*(volatile uint_t *)(p) = (uint_t)(v))
+#define	atomic_store_long(p, v)		\
+	(*(volatile ulong_t *)(p) = (ulong_t)(v))
+#define	atomic_store_ptr(p, v)		\
+	(*(volatile __typeof(*p) *)(p) = (v))
+#define	atomic_store_8(p, v)		\
+	(*(volatile uint8_t *)(p) = (uint8_t)(v))
+#define	atomic_store_16(p, v)		\
+	(*(volatile uint16_t *)(p) = (uint16_t)(v))
+#define	atomic_store_32(p, v)		\
+	(*(volatile uint32_t *)(p) = (uint32_t)(v))
+#ifdef _LP64
+#define	atomic_store_64(p, v)		\
+	(*(volatile uint64_t *)(p) = (uint64_t)(v))
+#elif defined(_INT64_TYPE)
+extern void atomic_store_64(volatile uint64_t *, uint64_t);
+#endif
+
 /*
  * Perform an exclusive atomic bit set/clear on a target.
  * Returns 0 if bit was successfully set/cleared, or -1

module/zfs/aggsum.c

@@ -78,21 +78,26 @@
  */
 
 /*
- * We will borrow aggsum_borrow_multiplier times the current request, so we will
- * have to get the as_lock approximately every aggsum_borrow_multiplier calls to
- * aggsum_delta().
+ * We will borrow 2^aggsum_borrow_shift times the current request, so we will
+ * have to get the as_lock approximately every 2^aggsum_borrow_shift calls to
+ * aggsum_add().
  */
-static uint_t aggsum_borrow_multiplier = 10;
+static uint_t aggsum_borrow_shift = 4;
 
 void
 aggsum_init(aggsum_t *as, uint64_t value)
 {
 	bzero(as, sizeof (*as));
 	as->as_lower_bound = as->as_upper_bound = value;
 	mutex_init(&as->as_lock, NULL, MUTEX_DEFAULT, NULL);
-	as->as_numbuckets = boot_ncpus;
-	as->as_buckets = kmem_zalloc(boot_ncpus * sizeof (aggsum_bucket_t),
-	    KM_SLEEP);
+	/*
+	 * Too many buckets may hurt read performance without improving
+	 * write.  From 12 CPUs use bucket per 2 CPUs, from 48 per 4, etc.
+	 */
+	as->as_bucketshift = highbit64(boot_ncpus / 6) / 2;
+	as->as_numbuckets = ((boot_ncpus - 1) >> as->as_bucketshift) + 1;
+	as->as_buckets = kmem_zalloc(as->as_numbuckets *
+	    sizeof (aggsum_bucket_t), KM_SLEEP);
 	for (int i = 0; i < as->as_numbuckets; i++) {
 		mutex_init(&as->as_buckets[i].asc_lock,
 		    NULL, MUTEX_DEFAULT, NULL);
@@ -111,59 +116,49 @@ aggsum_fini(aggsum_t *as)
 int64_t
 aggsum_lower_bound(aggsum_t *as)
 {
-	return (as->as_lower_bound);
+	return (atomic_load_64((volatile uint64_t *)&as->as_lower_bound));
 }
 
-int64_t
+uint64_t
 aggsum_upper_bound(aggsum_t *as)
 {
-	return (as->as_upper_bound);
-}
-
-static void
-aggsum_flush_bucket(aggsum_t *as, struct aggsum_bucket *asb)
-{
-	ASSERT(MUTEX_HELD(&as->as_lock));
-	ASSERT(MUTEX_HELD(&asb->asc_lock));
-
-	/*
-	 * We use atomic instructions for this because we read the upper and
-	 * lower bounds without the lock, so we need stores to be atomic.
-	 */
-	atomic_add_64((volatile uint64_t *)&as->as_lower_bound,
-	    asb->asc_delta + asb->asc_borrowed);
-	atomic_add_64((volatile uint64_t *)&as->as_upper_bound,
-	    asb->asc_delta - asb->asc_borrowed);
-	asb->asc_delta = 0;
-	asb->asc_borrowed = 0;
+	return (atomic_load_64(&as->as_upper_bound));
 }
 
 uint64_t
 aggsum_value(aggsum_t *as)
 {
-	int64_t rv;
+	int64_t lb;
+	uint64_t ub;
 
 	mutex_enter(&as->as_lock);
-	if (as->as_lower_bound == as->as_upper_bound) {
-		rv = as->as_lower_bound;
+	lb = as->as_lower_bound;
+	ub = as->as_upper_bound;
+	if (lb == ub) {
 		for (int i = 0; i < as->as_numbuckets; i++) {
 			ASSERT0(as->as_buckets[i].asc_delta);
 			ASSERT0(as->as_buckets[i].asc_borrowed);
 		}
 		mutex_exit(&as->as_lock);
-		return (rv);
+		return (lb);
 	}
 	for (int i = 0; i < as->as_numbuckets; i++) {
 		struct aggsum_bucket *asb = &as->as_buckets[i];
+		if (asb->asc_borrowed == 0)
+			continue;
 		mutex_enter(&asb->asc_lock);
-		aggsum_flush_bucket(as, asb);
+		lb += asb->asc_delta + asb->asc_borrowed;
+		ub += asb->asc_delta - asb->asc_borrowed;
+		asb->asc_delta = 0;
+		asb->asc_borrowed = 0;
 		mutex_exit(&asb->asc_lock);
 	}
-	VERIFY3U(as->as_lower_bound, ==, as->as_upper_bound);
-	rv = as->as_lower_bound;
+	ASSERT3U(lb, ==, ub);
+	atomic_store_64((volatile uint64_t *)&as->as_lower_bound, lb);
+	atomic_store_64(&as->as_upper_bound, lb);
 	mutex_exit(&as->as_lock);
 
-	return (rv);
+	return (lb);
 }
 
 void
@@ -172,7 +167,8 @@ aggsum_add(aggsum_t *as, int64_t delta)
 	struct aggsum_bucket *asb;
 	int64_t borrow;
 
-	asb = &as->as_buckets[CPU_SEQID_UNSTABLE % as->as_numbuckets];
+	asb = &as->as_buckets[(CPU_SEQID_UNSTABLE >> as->as_bucketshift) %
+	    as->as_numbuckets];
 
 	/* Try fast path if we already borrowed enough before. */
 	mutex_enter(&asb->asc_lock);
@@ -188,21 +184,22 @@ aggsum_add(aggsum_t *as, int64_t delta)
 	 * We haven't borrowed enough.  Take the global lock and borrow
 	 * considering what is requested now and what we borrowed before.
 	 */
-	borrow = (delta < 0 ? -delta : delta) * aggsum_borrow_multiplier;
+	borrow = (delta < 0 ? -delta : delta);
+	borrow <<= aggsum_borrow_shift + as->as_bucketshift;
 	mutex_enter(&as->as_lock);
-	mutex_enter(&asb->asc_lock);
-	delta += asb->asc_delta;
-	asb->asc_delta = 0;
 	if (borrow >= asb->asc_borrowed)
 		borrow -= asb->asc_borrowed;
 	else
 		borrow = (borrow - (int64_t)asb->asc_borrowed) / 4;
+	mutex_enter(&asb->asc_lock);
+	delta += asb->asc_delta;
+	asb->asc_delta = 0;
 	asb->asc_borrowed += borrow;
-	atomic_add_64((volatile uint64_t *)&as->as_lower_bound,
-	    delta - borrow);
-	atomic_add_64((volatile uint64_t *)&as->as_upper_bound,
-	    delta + borrow);
 	mutex_exit(&asb->asc_lock);
+	atomic_store_64((volatile uint64_t *)&as->as_lower_bound,
+	    as->as_lower_bound + delta - borrow);
+	atomic_store_64(&as->as_upper_bound,
+	    as->as_upper_bound + delta + borrow);
 	mutex_exit(&as->as_lock);
 }
 
@@ -214,27 +211,37 @@ aggsum_add(aggsum_t *as, int64_t delta)
 int
 aggsum_compare(aggsum_t *as, uint64_t target)
 {
-	if (as->as_upper_bound < target)
+	int64_t lb;
+	uint64_t ub;
+	int i;
+
+	if (atomic_load_64(&as->as_upper_bound) < target)
 		return (-1);
-	if (as->as_lower_bound > target)
+	lb = atomic_load_64((volatile uint64_t *)&as->as_lower_bound);
+	if (lb > 0 && (uint64_t)lb > target)
 		return (1);
 	mutex_enter(&as->as_lock);
-	for (int i = 0; i < as->as_numbuckets; i++) {
+	lb = as->as_lower_bound;
+	ub = as->as_upper_bound;
+	for (i = 0; i < as->as_numbuckets; i++) {
 		struct aggsum_bucket *asb = &as->as_buckets[i];
+		if (asb->asc_borrowed == 0)
+			continue;
 		mutex_enter(&asb->asc_lock);
-		aggsum_flush_bucket(as, asb);
+		lb += asb->asc_delta + asb->asc_borrowed;
+		ub += asb->asc_delta - asb->asc_borrowed;
+		asb->asc_delta = 0;
+		asb->asc_borrowed = 0;
 		mutex_exit(&asb->asc_lock);
-		if (as->as_upper_bound < target) {
-			mutex_exit(&as->as_lock);
-			return (-1);
-		}
-		if (as->as_lower_bound > target) {
-			mutex_exit(&as->as_lock);
-			return (1);
-		}
+		if (ub < target || (lb > 0 && (uint64_t)lb > target))
+			break;
+	}
+	if (i >= as->as_numbuckets) {
+		ASSERT3U(lb, ==, ub);
+		ASSERT3U(lb, ==, target);
 	}
-	VERIFY3U(as->as_lower_bound, ==, as->as_upper_bound);
-	ASSERT3U(as->as_lower_bound, ==, target);
+	atomic_store_64((volatile uint64_t *)&as->as_lower_bound, lb);
+	atomic_store_64(&as->as_upper_bound, ub);
 	mutex_exit(&as->as_lock);
-	return (0);
+	return (ub < target ? -1 : (uint64_t)lb > target ? 1 : 0);
 }