bcachefs: New varints

Previous varint implementation used by the inode code was not nearly as
fast as it could have been; partly because it was attempting to encode
integers up to 96 bits (for timestamps) but this meant that encoding and
decoding the length required a table lookup.

Instead, we'll just encode timestamps greater than 64 bits as two
separate varints; this will make decoding/encoding of inodes
significantly faster overall.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

fs/bcachefs/Makefile

@@ -57,4 +57,5 @@ bcachefs-y		:=	\
 	tests.o			\
 	trace.o			\
 	util.o			\
+	varint.o		\
 	xattr.o

fs/bcachefs/bcachefs_format.h

@@ -673,10 +673,10 @@ struct bch_inode_generation {
 } __attribute__((packed, aligned(8)));
 
 #define BCH_INODE_FIELDS()			\
-	x(bi_atime,			64)	\
-	x(bi_ctime,			64)	\
-	x(bi_mtime,			64)	\
-	x(bi_otime,			64)	\
+	x(bi_atime,			96)	\
+	x(bi_ctime,			96)	\
+	x(bi_mtime,			96)	\
+	x(bi_otime,			96)	\
 	x(bi_size,			64)	\
 	x(bi_sectors,			64)	\
 	x(bi_uid,			32)	\
@@ -743,7 +743,8 @@ enum {
 #define BCH_INODE_UNLINKED	(1 << __BCH_INODE_UNLINKED)
 
 LE32_BITMASK(INODE_STR_HASH,	struct bch_inode, bi_flags, 20, 24);
-LE32_BITMASK(INODE_NR_FIELDS,	struct bch_inode, bi_flags, 24, 32);
+LE32_BITMASK(INODE_NR_FIELDS,	struct bch_inode, bi_flags, 24, 31);
+LE32_BITMASK(INODE_NEW_VARINT,	struct bch_inode, bi_flags, 31, 32);
 
 /* Dirents */
 
@@ -1334,13 +1335,15 @@ LE64_BITMASK(BCH_SB_ERASURE_CODE,	struct bch_sb, flags[3],  0, 16);
 	x(btree_ptr_v2,			11)	\
 	x(extents_above_btree_updates,	12)	\
 	x(btree_updates_journalled,	13)	\
-	x(reflink_inline_data,		14)
+	x(reflink_inline_data,		14)	\
+	x(new_varint,			15)
 
 #define BCH_SB_FEATURES_ALL				\
 	((1ULL << BCH_FEATURE_new_siphash)|		\
 	 (1ULL << BCH_FEATURE_new_extent_overwrite)|	\
 	 (1ULL << BCH_FEATURE_btree_ptr_v2)|		\
-	 (1ULL << BCH_FEATURE_extents_above_btree_updates))
+	 (1ULL << BCH_FEATURE_extents_above_btree_updates)|\
+	 (1ULL << BCH_FEATURE_new_varint))\
 
 enum bch_sb_feature {
 #define x(f, n) BCH_FEATURE_##f,

fs/bcachefs/fsck.c

@@ -537,7 +537,7 @@ static int check_extents(struct bch_fs *c)
 
 			bch2_trans_unlock(&trans);
 
-			bch2_inode_pack(&p, &w.inode);
+			bch2_inode_pack(c, &p, &w.inode);
 
 			ret = bch2_btree_insert(c, BTREE_ID_INODES,
 						&p.inode.k_i, NULL, NULL,
@@ -808,7 +808,7 @@ static int check_root(struct bch_fs *c, struct bch_inode_unpacked *root_inode)
 			0, NULL);
 	root_inode->bi_inum = BCACHEFS_ROOT_INO;
 
-	bch2_inode_pack(&packed, root_inode);
+	bch2_inode_pack(c, &packed, root_inode);
 
 	return bch2_btree_insert(c, BTREE_ID_INODES, &packed.inode.k_i,
 				 NULL, NULL,
@@ -1326,7 +1326,7 @@ static int check_inode(struct btree_trans *trans,
 	if (do_update) {
 		struct bkey_inode_buf p;
 
-		bch2_inode_pack(&p, &u);
+		bch2_inode_pack(c, &p, &u);
 
 		ret = __bch2_trans_do(trans, NULL, NULL,
 				      BTREE_INSERT_NOFAIL|

fs/bcachefs/inode.c

@@ -8,6 +8,7 @@
 #include "extents.h"
 #include "inode.h"
 #include "str_hash.h"
+#include "varint.h"
 
 #include <linux/random.h>
 
@@ -89,22 +90,17 @@ static int inode_decode_field(const u8 *in, const u8 *end,
 	return bytes;
 }
 
-void bch2_inode_pack(struct bkey_inode_buf *packed,
-		     const struct bch_inode_unpacked *inode)
+static noinline void bch2_inode_pack_v1(struct bkey_inode_buf *packed,
+					const struct bch_inode_unpacked *inode)
 {
-	u8 *out = packed->inode.v.fields;
+	struct bkey_i_inode *k = &packed->inode;
+	u8 *out = k->v.fields;
 	u8 *end = (void *) &packed[1];
 	u8 *last_nonzero_field = out;
 	unsigned nr_fields = 0, last_nonzero_fieldnr = 0;
 	unsigned bytes;
 
-	bkey_inode_init(&packed->inode.k_i);
-	packed->inode.k.p.offset	= inode->bi_inum;
-	packed->inode.v.bi_hash_seed	= inode->bi_hash_seed;
-	packed->inode.v.bi_flags	= cpu_to_le32(inode->bi_flags);
-	packed->inode.v.bi_mode		= cpu_to_le16(inode->bi_mode);
-
-#define x(_name, _bits)					\
+#define x(_name, _bits)							\
 	out += inode_encode_field(out, end, 0, inode->_name);		\
 	nr_fields++;							\
 									\
@@ -123,7 +119,69 @@ void bch2_inode_pack(struct bkey_inode_buf *packed,
 	set_bkey_val_bytes(&packed->inode.k, bytes);
 	memset_u64s_tail(&packed->inode.v, 0, bytes);
 
-	SET_INODE_NR_FIELDS(&packed->inode.v, nr_fields);
+	SET_INODE_NR_FIELDS(&k->v, nr_fields);
+}
+
+static void bch2_inode_pack_v2(struct bkey_inode_buf *packed,
+			       const struct bch_inode_unpacked *inode)
+{
+	struct bkey_i_inode *k = &packed->inode;
+	u8 *out = k->v.fields;
+	u8 *end = (void *) &packed[1];
+	u8 *last_nonzero_field = out;
+	unsigned nr_fields = 0, last_nonzero_fieldnr = 0;
+	unsigned bytes;
+	int ret;
+
+#define x(_name, _bits)							\
+	nr_fields++;							\
+									\
+	if (inode->_name) {						\
+		ret = bch2_varint_encode(out, inode->_name);		\
+		out += ret;						\
+									\
+		if (_bits > 64)						\
+			*out++ = 0;					\
+									\
+		last_nonzero_field = out;				\
+		last_nonzero_fieldnr = nr_fields;			\
+	} else {							\
+		*out++ = 0;						\
+									\
+		if (_bits > 64)						\
+			*out++ = 0;					\
+	}
+
+	BCH_INODE_FIELDS()
+#undef  x
+	BUG_ON(out > end);
+
+	out = last_nonzero_field;
+	nr_fields = last_nonzero_fieldnr;
+
+	bytes = out - (u8 *) &packed->inode.v;
+	set_bkey_val_bytes(&packed->inode.k, bytes);
+	memset_u64s_tail(&packed->inode.v, 0, bytes);
+
+	SET_INODE_NR_FIELDS(&k->v, nr_fields);
+}
+
+void bch2_inode_pack(struct bch_fs *c,
+		     struct bkey_inode_buf *packed,
+		     const struct bch_inode_unpacked *inode)
+{
+	bkey_inode_init(&packed->inode.k_i);
+	packed->inode.k.p.offset	= inode->bi_inum;
+	packed->inode.v.bi_hash_seed	= inode->bi_hash_seed;
+	packed->inode.v.bi_flags	= cpu_to_le32(inode->bi_flags);
+	packed->inode.v.bi_mode		= cpu_to_le16(inode->bi_mode);
+
+	if (c->sb.features & (1ULL << BCH_FEATURE_new_varint)) {
+		SET_INODE_NEW_VARINT(&packed->inode.v, true);
+		bch2_inode_pack_v2(packed, inode);
+	} else {
+		bch2_inode_pack_v1(packed, inode);
+	}
 
 	if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG)) {
 		struct bch_inode_unpacked unpacked;
@@ -135,26 +193,23 @@ void bch2_inode_pack(struct bkey_inode_buf *packed,
 		BUG_ON(unpacked.bi_hash_seed	!= inode->bi_hash_seed);
 		BUG_ON(unpacked.bi_mode		!= inode->bi_mode);
 
-#define x(_name, _bits)	BUG_ON(unpacked._name != inode->_name);
+#define x(_name, _bits)	if (unpacked._name != inode->_name)		\
+			panic("unpacked %llu should be %llu",		\
+			      (u64) unpacked._name, (u64) inode->_name);
 		BCH_INODE_FIELDS()
 #undef  x
 	}
 }
 
-int bch2_inode_unpack(struct bkey_s_c_inode inode,
-		      struct bch_inode_unpacked *unpacked)
+static noinline int bch2_inode_unpack_v1(struct bkey_s_c_inode inode,
+				struct bch_inode_unpacked *unpacked)
 {
 	const u8 *in = inode.v->fields;
-	const u8 *end = (void *) inode.v + bkey_val_bytes(inode.k);
+	const u8 *end = bkey_val_end(inode);
 	u64 field[2];
 	unsigned fieldnr = 0, field_bits;
 	int ret;
 
-	unpacked->bi_inum	= inode.k->p.offset;
-	unpacked->bi_hash_seed	= inode.v->bi_hash_seed;
-	unpacked->bi_flags	= le32_to_cpu(inode.v->bi_flags);
-	unpacked->bi_mode	= le16_to_cpu(inode.v->bi_mode);
-
 #define x(_name, _bits)					\
 	if (fieldnr++ == INODE_NR_FIELDS(inode.v)) {			\
 		unsigned offset = offsetof(struct bch_inode_unpacked, _name);\
@@ -177,6 +232,62 @@ int bch2_inode_unpack(struct bkey_s_c_inode inode,
 #undef  x
 
 	/* XXX: signal if there were more fields than expected? */
+	return 0;
+}
+
+static int bch2_inode_unpack_v2(struct bkey_s_c_inode inode,
+				struct bch_inode_unpacked *unpacked)
+{
+	const u8 *in = inode.v->fields;
+	const u8 *end = bkey_val_end(inode);
+	unsigned fieldnr = 0;
+	int ret;
+	u64 v[2];
+
+#define x(_name, _bits)							\
+	if (fieldnr < INODE_NR_FIELDS(inode.v)) {			\
+		ret = bch2_varint_decode(in, end, &v[0]);		\
+		if (ret < 0)						\
+			return ret;					\
+		in += ret;						\
+									\
+		if (_bits > 64) {					\
+			ret = bch2_varint_decode(in, end, &v[1]);	\
+			if (ret < 0)					\
+				return ret;				\
+			in += ret;					\
+		} else {						\
+			v[1] = 0;					\
+		}							\
+	} else {							\
+		v[0] = v[1] = 0;					\
+	}								\
+									\
+	unpacked->_name = v[0];						\
+	if (v[1] || v[0] != unpacked->_name)				\
+		return -1;						\
+	fieldnr++;
+
+	BCH_INODE_FIELDS()
+#undef  x
+
+	/* XXX: signal if there were more fields than expected? */
+	return 0;
+}
+
+int bch2_inode_unpack(struct bkey_s_c_inode inode,
+		      struct bch_inode_unpacked *unpacked)
+{
+	unpacked->bi_inum	= inode.k->p.offset;
+	unpacked->bi_hash_seed	= inode.v->bi_hash_seed;
+	unpacked->bi_flags	= le32_to_cpu(inode.v->bi_flags);
+	unpacked->bi_mode	= le16_to_cpu(inode.v->bi_mode);
+
+	if (INODE_NEW_VARINT(inode.v)) {
+		return bch2_inode_unpack_v2(inode, unpacked);
+	} else {
+		return bch2_inode_unpack_v1(inode, unpacked);
+	}
 
 	return 0;
 }
@@ -223,7 +334,7 @@ int bch2_inode_write(struct btree_trans *trans,
 	if (IS_ERR(inode_p))
 		return PTR_ERR(inode_p);
 
-	bch2_inode_pack(inode_p, inode);
+	bch2_inode_pack(trans->c, inode_p, inode);
 	bch2_trans_update(trans, iter, &inode_p->inode.k_i, 0);
 	return 0;
 }
@@ -426,10 +537,7 @@ int bch2_inode_create(struct btree_trans *trans,
 	inode_u->bi_inum	= k.k->p.offset;
 	inode_u->bi_generation	= bkey_generation(k);
 
-	bch2_inode_pack(inode_p, inode_u);
-	bch2_trans_update(trans, iter, &inode_p->inode.k_i, 0);
-	bch2_trans_iter_put(trans, iter);
-	return 0;
+	return bch2_inode_write(trans, iter, inode_u);
 }
 
 int bch2_inode_rm(struct bch_fs *c, u64 inode_nr)
@@ -553,32 +661,3 @@ int bch2_inode_find_by_inum(struct bch_fs *c, u64 inode_nr,
 	return bch2_trans_do(c, NULL, NULL, 0,
 		bch2_inode_find_by_inum_trans(&trans, inode_nr, inode));
 }
-
-#ifdef CONFIG_BCACHEFS_DEBUG
-void bch2_inode_pack_test(void)
-{
-	struct bch_inode_unpacked *u, test_inodes[] = {
-		{
-			.bi_atime	= U64_MAX,
-			.bi_ctime	= U64_MAX,
-			.bi_mtime	= U64_MAX,
-			.bi_otime	= U64_MAX,
-			.bi_size	= U64_MAX,
-			.bi_sectors	= U64_MAX,
-			.bi_uid		= U32_MAX,
-			.bi_gid		= U32_MAX,
-			.bi_nlink	= U32_MAX,
-			.bi_generation	= U32_MAX,
-			.bi_dev		= U32_MAX,
-		},
-	};
-
-	for (u = test_inodes;
-	     u < test_inodes + ARRAY_SIZE(test_inodes);
-	     u++) {
-		struct bkey_inode_buf p;
-
-		bch2_inode_pack(&p, u);
-	}
-}
-#endif

fs/bcachefs/inode.h

@@ -24,6 +24,14 @@ void bch2_inode_generation_to_text(struct printbuf *, struct bch_fs *,
 	.val_to_text	= bch2_inode_generation_to_text,	\
 }
 
+#if 0
+typedef struct {
+	u64			lo;
+	u32			hi;
+} __packed __aligned(4) u96;
+#endif
+typedef u64 u96;
+
 struct bch_inode_unpacked {
 	u64			bi_inum;
 	__le64			bi_hash_seed;
@@ -43,7 +51,8 @@ struct bkey_inode_buf {
 #undef  x
 } __attribute__((packed, aligned(8)));
 
-void bch2_inode_pack(struct bkey_inode_buf *, const struct bch_inode_unpacked *);
+void bch2_inode_pack(struct bch_fs *, struct bkey_inode_buf *,
+		     const struct bch_inode_unpacked *);
 int bch2_inode_unpack(struct bkey_s_c_inode, struct bch_inode_unpacked *);
 
 struct btree_iter *bch2_inode_peek(struct btree_trans *,
@@ -166,10 +175,4 @@ static inline void bch2_inode_nlink_set(struct bch_inode_unpacked *bi,
 	}
 }
 
-#ifdef CONFIG_BCACHEFS_DEBUG
-void bch2_inode_pack_test(void);
-#else
-static inline void bch2_inode_pack_test(void) {}
-#endif
-
 #endif /* _BCACHEFS_INODE_H */

fs/bcachefs/io.c

@@ -310,7 +310,7 @@ int bch2_extent_update(struct btree_trans *trans,
 		inode_u.bi_sectors += delta;
 
 		if (delta || new_i_size) {
-			bch2_inode_pack(&inode_p, &inode_u);
+			bch2_inode_pack(trans->c, &inode_p, &inode_u);
 			bch2_trans_update(trans, inode_iter,
 					  &inode_p.inode.k_i, 0);
 		}

fs/bcachefs/recovery.c

@@ -1320,7 +1320,7 @@ int bch2_fs_initialize(struct bch_fs *c)
 	bch2_inode_init(c, &root_inode, 0, 0,
 			S_IFDIR|S_IRWXU|S_IRUGO|S_IXUGO, 0, NULL);
 	root_inode.bi_inum = BCACHEFS_ROOT_INO;
-	bch2_inode_pack(&packed_inode, &root_inode);
+	bch2_inode_pack(c, &packed_inode, &root_inode);
 
 	err = "error creating root directory";
 	ret = bch2_btree_insert(c, BTREE_ID_INODES,

fs/bcachefs/super.c

@@ -2027,7 +2027,6 @@ static void bcachefs_exit(void)
 static int __init bcachefs_init(void)
 {
 	bch2_bkey_pack_test();
-	bch2_inode_pack_test();
 
 	if (!(bcachefs_kset = kset_create_and_add("bcachefs", NULL, fs_kobj)) ||
 	    bch2_chardev_init() ||

fs/bcachefs/varint.c

+// SPDX-License-Identifier: GPL-2.0
+
+#include <linux/bitops.h>
+#include <linux/math.h>
+#include <asm/unaligned.h>
+
+#include "varint.h"
+
+int bch2_varint_encode(u8 *out, u64 v)
+{
+	unsigned bits = fls64(v|1);
+	unsigned bytes = DIV_ROUND_UP(bits, 7);
+
+	if (likely(bytes < 9)) {
+		v <<= bytes;
+		v |= ~(~0 << (bytes - 1));
+	} else {
+		*out++ = 255;
+		bytes = 9;
+	}
+
+	put_unaligned_le64(v, out);
+	return bytes;
+}
+
+int bch2_varint_decode(const u8 *in, const u8 *end, u64 *out)
+{
+	u64 v = get_unaligned_le64(in);
+	unsigned bytes = ffz(v & 255) + 1;
+
+	if (unlikely(in + bytes > end))
+		return -1;
+
+	if (likely(bytes < 9)) {
+		v >>= bytes;
+		v &= ~(~0ULL << (7 * bytes));
+	} else {
+		v = get_unaligned_le64(++in);
+	}
+
+	*out = v;
+	return bytes;
+}

fs/bcachefs/varint.h

+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _BCACHEFS_VARINT_H
+#define _BCACHEFS_VARINT_H
+
+int bch2_varint_encode(u8 *, u64);
+int bch2_varint_decode(const u8 *, const u8 *, u64 *);
+
+#endif /* _BCACHEFS_VARINT_H */