btree: Implement faster binary search algorithm

This implements a binary search algorithm for B-Trees that reduces
branching to the absolute minimum necessary for a binary search
algorithm. It also enables the compiler to inlines the comparator to
ensure that the only slowdown when doing binary search is from waiting
for memory accesses.

Consumers must opt into using the faster algorithm.

Signed-off-by: Richard Yao <[email protected]>

Author: ryao

include/sys/btree.h

@@ -105,8 +105,12 @@ typedef struct zfs_btree_index {
 	boolean_t	bti_before;
 } zfs_btree_index_t;
 
-typedef struct btree {
+typedef struct btree zfs_btree_t;
+
+struct btree {
 	int (*bt_compar) (const void *, const void *);
+	void * (*bt_find_in_buf) (zfs_btree_t *, uint8_t *, uint32_t,
+	    const void *, zfs_btree_index_t *);
 	size_t			bt_elem_size;
 	size_t			bt_leaf_size;
 	uint32_t		bt_leaf_cap;
@@ -115,7 +119,41 @@ typedef struct btree {
 	uint64_t		bt_num_nodes;
 	zfs_btree_hdr_t		*bt_root;
 	zfs_btree_leaf_t	*bt_bulk; // non-null if bulk loading
-} zfs_btree_t;
+};
+
+/*
+ * Implementation of Shar's algorithm designed to accelerate binary search by
+ * eliminating impossible to predict branches. For optimality, this should be
+ * used to generate the search function in the same file as the comparator and
+ * the comparator must be marked `__attribute__((always_inline) inline` so that
+ * the compiler can inline the search function.
+ *
+ * Consumers must opt into using this function by setting ->bt_find_in_buf
+ * after creating the btree.
+ */
+#define	ZFS_BTREE_FIND_IN_BUF_FUNC(NAME, T, COMP)			\
+static void *								\
+NAME(zfs_btree_t *tree, uint8_t *buf, uint32_t nelems,			\
+    const void *value, zfs_btree_index_t *where)			\
+{									\
+	T *i = (T *)buf;						\
+	(void) tree;							\
+	while (nelems > 1) {						\
+		uint32_t half = nelems / 2;				\
+		nelems -= half;						\
+		i += (COMP(&i[half - 1], value) < 0) * half;		\
+	}								\
+									\
+	int comp = COMP(i, value);					\
+	where->bti_offset = (i - (T *)buf) + (comp < 0);		\
+	where->bti_before = (comp != 0);				\
+									\
+	if (comp == 0) {						\
+		return (i);						\
+	}								\
+									\
+	return (NULL);							\
+}
 
 /*
  * Allocate and deallocate caches for btree nodes.

module/Kbuild.in

@@ -34,6 +34,17 @@ ifeq ($(CONFIG_KASAN),y)
 ZFS_MODULE_CFLAGS += -Wno-error=frame-larger-than=
 endif
 
+# Generated binary search code is particularly bad with this optimization.
+# Oddly, range_tree.c is not affected.
+# Disable it on the other two until the following upstream issue is resolved:
+# https://github.com/llvm/llvm-project/issues/62790
+ifeq ($(CONFIG_X86_64),y)
+ifeq ($(CONFIG_CC_IS_CLANG),y)
+CFLAGS_zfs/dsl_scan.o += -mllvm -x86-cmov-converter=false
+CFLAGS_zfs/metaslab.o += -mllvm -x86-cmov-converter=false
+endif
+endif
+
 ifneq ($(KBUILD_EXTMOD),)
 @CONFIG_QAT_TRUE@ZFS_MODULE_CFLAGS += -I@QAT_SRC@/include
 @CONFIG_QAT_TRUE@KBUILD_EXTRA_SYMBOLS += @QAT_SYMBOLS@

module/Makefile.bsd

@@ -400,6 +400,15 @@ beforeinstall:
 
 .include <bsd.kmod.mk>
 
+# Generated binary search code is particularly bad with this optimization.
+# Oddly, range_tree.c is not affected.
+# Disable it on the other two until the following upstream issue is resolved:
+# https://github.com/llvm/llvm-project/issues/62790
+.if ${CC} == "clang"
+CFLAGS.dsl_scan.c= -mllvm -x86-cmov-converter=false
+CFLAGS.metaslab.c= -mllvm -x86-cmov-converter=false
+.endif
+
 CFLAGS.sysctl_os.c= -include ../zfs_config.h
 CFLAGS.xxhash.c+= -include ${SYSDIR}/sys/_null.h
 

module/zfs/btree.c

@@ -198,6 +198,10 @@ zfs_btree_create(zfs_btree_t *tree, int (*compar) (const void *, const void *),
 	zfs_btree_create_custom(tree, compar, size, BTREE_LEAF_SIZE);
 }
 
+static void *
+zfs_btree_find_in_buf(zfs_btree_t *tree, uint8_t *buf, uint32_t nelems,
+    const void *value, zfs_btree_index_t *where);
+
 void
 zfs_btree_create_custom(zfs_btree_t *tree,
     int (*compar) (const void *, const void *),
@@ -208,6 +212,7 @@ zfs_btree_create_custom(zfs_btree_t *tree,
 	ASSERT3U(size, <=, esize / 2);
 	memset(tree, 0, sizeof (*tree));
 	tree->bt_compar = compar;
+	tree->bt_find_in_buf = zfs_btree_find_in_buf;
 	tree->bt_elem_size = size;
 	tree->bt_leaf_size = lsize;
 	tree->bt_leaf_cap = P2ALIGN(esize / size, 2);
@@ -303,7 +308,7 @@ zfs_btree_find(zfs_btree_t *tree, const void *value, zfs_btree_index_t *where)
 			 * element in the last leaf, it's in the last leaf or
 			 * it's not in the tree.
 			 */
-			void *d = zfs_btree_find_in_buf(tree,
+			void *d = tree->bt_find_in_buf(tree,
 			    last_leaf->btl_elems +
 			    last_leaf->btl_hdr.bth_first * size,
 			    last_leaf->btl_hdr.bth_count, value, &idx);
@@ -327,7 +332,7 @@ zfs_btree_find(zfs_btree_t *tree, const void *value, zfs_btree_index_t *where)
 	for (node = (zfs_btree_core_t *)tree->bt_root; depth < tree->bt_height;
 	    node = (zfs_btree_core_t *)node->btc_children[child], depth++) {
 		ASSERT3P(node, !=, NULL);
-		void *d = zfs_btree_find_in_buf(tree, node->btc_elems,
+		void *d = tree->bt_find_in_buf(tree, node->btc_elems,
 		    node->btc_hdr.bth_count, value, &idx);
 		EQUIV(d != NULL, !idx.bti_before);
 		if (d != NULL) {
@@ -347,7 +352,7 @@ zfs_btree_find(zfs_btree_t *tree, const void *value, zfs_btree_index_t *where)
 	 */
 	zfs_btree_leaf_t *leaf = (depth == 0 ?
 	    (zfs_btree_leaf_t *)tree->bt_root : (zfs_btree_leaf_t *)node);
-	void *d = zfs_btree_find_in_buf(tree, leaf->btl_elems +
+	void *d = tree->bt_find_in_buf(tree, leaf->btl_elems +
 	    leaf->btl_hdr.bth_first * size,
 	    leaf->btl_hdr.bth_count, value, &idx);
 
@@ -671,7 +676,7 @@ zfs_btree_insert_into_parent(zfs_btree_t *tree, zfs_btree_hdr_t *old_node,
 	zfs_btree_hdr_t *par_hdr = &parent->btc_hdr;
 	zfs_btree_index_t idx;
 	ASSERT(zfs_btree_is_core(par_hdr));
-	VERIFY3P(zfs_btree_find_in_buf(tree, parent->btc_elems,
+	VERIFY3P(tree->bt_find_in_buf(tree, parent->btc_elems,
 	    par_hdr->bth_count, buf, &idx), ==, NULL);
 	ASSERT(idx.bti_before);
 	uint32_t offset = idx.bti_offset;
@@ -897,7 +902,7 @@ zfs_btree_find_parent_idx(zfs_btree_t *tree, zfs_btree_hdr_t *hdr)
 	}
 	zfs_btree_index_t idx;
 	zfs_btree_core_t *parent = hdr->bth_parent;
-	VERIFY3P(zfs_btree_find_in_buf(tree, parent->btc_elems,
+	VERIFY3P(tree->bt_find_in_buf(tree, parent->btc_elems,
 	    parent->btc_hdr.bth_count, buf, &idx), ==, NULL);
 	ASSERT(idx.bti_before);
 	ASSERT3U(idx.bti_offset, <=, parent->btc_hdr.bth_count);

module/zfs/dsl_scan.c

@@ -4195,6 +4195,7 @@ scan_exec_io(dsl_pool_t *dp, const blkptr_t *bp, int zio_flags,
  * with single operation.  Plus it makes scrubs more sequential and reduces
  * chances that minor extent change move it within the B-tree.
  */
+__attribute__((always_inline)) inline
 static int
 ext_size_compare(const void *x, const void *y)
 {
@@ -4203,13 +4204,17 @@ ext_size_compare(const void *x, const void *y)
 	return (TREE_CMP(*a, *b));
 }
 
+ZFS_BTREE_FIND_IN_BUF_FUNC(ext_size_find_in_buf, uint64_t,
+    ext_size_compare)
+
 static void
 ext_size_create(range_tree_t *rt, void *arg)
 {
 	(void) rt;
 	zfs_btree_t *size_tree = arg;
 
 	zfs_btree_create(size_tree, ext_size_compare, sizeof (uint64_t));
+	size_tree->bt_find_in_buf = ext_size_find_in_buf;
 }
 
 static void

module/zfs/metaslab.c

@@ -1342,6 +1342,7 @@ metaslab_group_allocatable(metaslab_group_t *mg, metaslab_group_t *rotor,
  * Comparison function for the private size-ordered tree using 32-bit
  * ranges. Tree is sorted by size, larger sizes at the end of the tree.
  */
+__attribute__((always_inline)) inline
 static int
 metaslab_rangesize32_compare(const void *x1, const void *x2)
 {
@@ -1352,16 +1353,15 @@ metaslab_rangesize32_compare(const void *x1, const void *x2)
 	uint64_t rs_size2 = r2->rs_end - r2->rs_start;
 
 	int cmp = TREE_CMP(rs_size1, rs_size2);
-	if (likely(cmp))
-		return (cmp);
 
-	return (TREE_CMP(r1->rs_start, r2->rs_start));
+	return (cmp + !cmp * TREE_CMP(r1->rs_start, r2->rs_start));
 }
 
 /*
  * Comparison function for the private size-ordered tree using 64-bit
  * ranges. Tree is sorted by size, larger sizes at the end of the tree.
  */
+__attribute__((always_inline)) inline
 static int
 metaslab_rangesize64_compare(const void *x1, const void *x2)
 {
@@ -1372,11 +1372,10 @@ metaslab_rangesize64_compare(const void *x1, const void *x2)
 	uint64_t rs_size2 = r2->rs_end - r2->rs_start;
 
 	int cmp = TREE_CMP(rs_size1, rs_size2);
-	if (likely(cmp))
-		return (cmp);
 
-	return (TREE_CMP(r1->rs_start, r2->rs_start));
+	return (cmp + !cmp * TREE_CMP(r1->rs_start, r2->rs_start));
 }
+
 typedef struct metaslab_rt_arg {
 	zfs_btree_t *mra_bt;
 	uint32_t mra_floor_shift;
@@ -1412,6 +1411,13 @@ metaslab_size_tree_full_load(range_tree_t *rt)
 	range_tree_walk(rt, metaslab_size_sorted_add, &arg);
 }
 
+
+ZFS_BTREE_FIND_IN_BUF_FUNC(metaslab_rt_find_rangesize32_in_buf,
+    range_seg32_t, metaslab_rangesize32_compare)
+
+ZFS_BTREE_FIND_IN_BUF_FUNC(metaslab_rt_find_rangesize64_in_buf,
+    range_seg64_t, metaslab_rangesize64_compare)
+
 /*
  * Create any block allocator specific components. The current allocators
  * rely on using both a size-ordered range_tree_t and an array of uint64_t's.
@@ -1428,15 +1434,20 @@ metaslab_rt_create(range_tree_t *rt, void *arg)
 	case RANGE_SEG32:
 		size = sizeof (range_seg32_t);
 		compare = metaslab_rangesize32_compare;
+		zfs_btree_create(size_tree, compare, size);
+		size_tree->bt_find_in_buf =
+		    metaslab_rt_find_rangesize32_in_buf;
 		break;
 	case RANGE_SEG64:
 		size = sizeof (range_seg64_t);
 		compare = metaslab_rangesize64_compare;
+		zfs_btree_create(size_tree, compare, size);
+		size_tree->bt_find_in_buf =
+		    metaslab_rt_find_rangesize64_in_buf;
 		break;
 	default:
 		panic("Invalid range seg type %d", rt->rt_type);
 	}
-	zfs_btree_create(size_tree, compare, size);
 	mrap->mra_floor_shift = metaslab_by_size_min_shift;
 }
 

module/zfs/range_tree.c

@@ -151,6 +151,7 @@ range_tree_stat_decr(range_tree_t *rt, range_seg_t *rs)
 	rt->rt_histogram[idx]--;
 }
 
+__attribute__((always_inline)) inline
 static int
 range_tree_seg32_compare(const void *x1, const void *x2)
 {
@@ -163,6 +164,7 @@ range_tree_seg32_compare(const void *x1, const void *x2)
 	return ((r1->rs_start >= r2->rs_end) - (r1->rs_end <= r2->rs_start));
 }
 
+__attribute__((always_inline)) inline
 static int
 range_tree_seg64_compare(const void *x1, const void *x2)
 {
@@ -175,6 +177,7 @@ range_tree_seg64_compare(const void *x1, const void *x2)
 	return ((r1->rs_start >= r2->rs_end) - (r1->rs_end <= r2->rs_start));
 }
 
+__attribute__((always_inline)) inline
 static int
 range_tree_seg_gap_compare(const void *x1, const void *x2)
 {
@@ -187,6 +190,15 @@ range_tree_seg_gap_compare(const void *x1, const void *x2)
 	return ((r1->rs_start >= r2->rs_end) - (r1->rs_end <= r2->rs_start));
 }
 
+ZFS_BTREE_FIND_IN_BUF_FUNC(range_tree_seg32_find_in_buf, range_seg32_t,
+    range_tree_seg32_compare)
+
+ZFS_BTREE_FIND_IN_BUF_FUNC(range_tree_seg64_find_in_buf, range_seg64_t,
+    range_tree_seg64_compare)
+
+ZFS_BTREE_FIND_IN_BUF_FUNC(range_tree_seg_gap_find_in_buf, range_seg_gap_t,
+    range_tree_seg_gap_compare)
+
 range_tree_t *
 range_tree_create_gap(const range_tree_ops_t *ops, range_seg_type_t type,
     void *arg, uint64_t start, uint64_t shift, uint64_t gap)
@@ -201,19 +213,24 @@ range_tree_create_gap(const range_tree_ops_t *ops, range_seg_type_t type,
 	case RANGE_SEG32:
 		size = sizeof (range_seg32_t);
 		compare = range_tree_seg32_compare;
+		zfs_btree_create(&rt->rt_root, compare, size);
+		rt->rt_root.bt_find_in_buf = range_tree_seg32_find_in_buf;
 		break;
 	case RANGE_SEG64:
 		size = sizeof (range_seg64_t);
 		compare = range_tree_seg64_compare;
+		zfs_btree_create(&rt->rt_root, compare, size);
+		rt->rt_root.bt_find_in_buf = range_tree_seg64_find_in_buf;
 		break;
 	case RANGE_SEG_GAP:
 		size = sizeof (range_seg_gap_t);
 		compare = range_tree_seg_gap_compare;
+		zfs_btree_create(&rt->rt_root, compare, size);
+		rt->rt_root.bt_find_in_buf = range_tree_seg_gap_find_in_buf;
 		break;
 	default:
 		panic("Invalid range seg type %d", type);
 	}
-	zfs_btree_create(&rt->rt_root, compare, size);
 
 	rt->rt_ops = ops;
 	rt->rt_gap = gap;