Skip to contentSkip to conversation

f2fs-tools: optimize the addrs_per_block/inode loop

PATCH Superseded Wu Bo via Linux-f2fs-devel 1 participant 2 messages Jan 12 Library code

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 f2fs-tools: optimize the addrs_per_block/inode loop

ADDRS_PER_INODE and ADDRS_PER_BLOCK are macros of a function, and if the compiler does not optimize the code (-O0 for example), they would be called in every single loop, which could cause significant performance drops in some cases.

This scenario has been found before, refer to:
Commit: 1bb669e ("fsck.f2fs: avoid unnecessary recalculation")

To avoid this performance drop as much as possible, I changed the function to 'inline' definition, which could mitigate such cases in the future.

---
 fsck/dump.c       |  5 +++--
 fsck/fsck.c       | 15 ++++++++++-----
 fsck/mount.c      |  4 +++-
 include/f2fs_fs.h | 12 +++++++++++-
 lib/libf2fs.c     | 11 -----------
 5 files changed, 27 insertions(+), 20 deletions(-)
fsck/dump.c +3 -2
index f5c95de..a5f344b 100644
@@ -495,7 +495,7 @@ static int dump_inode_blk(struct f2fs_sb_info *sbi, u32 nid,
 {
 	u32 i = 0;
 	u64 ofs = 0;
-	u32 addr_per_block;
+	u32 addr_per_block, addr_per_inode;
 	u16 type = le16_to_cpu(node_blk->i.i_mode);
 	int ret = 0;
 
@@ -543,9 +543,10 @@ static int dump_inode_blk(struct f2fs_sb_info *sbi, u32 nid,
 	}
 
 	addr_per_block = ADDRS_PER_BLOCK(&node_blk->i);
+	addr_per_inode = ADDRS_PER_INODE(&node_blk->i);
 
 	/* check data blocks in inode */
-	for (i = 0; i < ADDRS_PER_INODE(&node_blk->i); i++, ofs++)
+	for (i = 0; i < addr_per_inode; i++, ofs++)
 		dump_data_blk(sbi, ofs * F2FS_BLKSIZE, le32_to_cpu(
 			node_blk->i.i_addr[get_extra_isize(node_blk) + i]), type);
 
fsck/fsck.c +10 -5
index db44f9d..1627bd0 100644
@@ -1516,8 +1516,9 @@ int fsck_chk_dnode_blk(struct f2fs_sb_info *sbi, struct f2fs_inode *inode,
 	bool compressed = i_flags & F2FS_COMPR_FL;
 	bool compr_rel = inode->i_inline & F2FS_COMPRESS_RELEASED;
 	u32 cluster_size = 1 << inode->i_log_cluster_size;
+	u32 addrs = ADDRS_PER_BLOCK(inode);
 
-	for (idx = 0; idx < ADDRS_PER_BLOCK(inode); idx++, child->pgofs++) {
+	for (idx = 0; idx < addrs; idx++, child->pgofs++) {
 		block_t blkaddr = le32_to_cpu(node_blk->dn.addr[idx]);
 
 		check_extent_info(child, blkaddr, 0);
@@ -1617,6 +1618,7 @@ int fsck_chk_didnode_blk(struct f2fs_sb_info *sbi, struct f2fs_inode *inode,
 {
 	int i = 0;
 	int need_fix = 0, ret = 0;
+	u32 addrs = ADDRS_PER_BLOCK(inode);
 
 	fsck_reada_all_direct_node_blocks(sbi, node_blk);
 
@@ -1637,7 +1639,7 @@ int fsck_chk_didnode_blk(struct f2fs_sb_info *sbi, struct f2fs_inode *inode,
 				FIX_MSG("Set double indirect node 0x%x -> 0", i);
 			}
 skip:
-			child->pgofs += ADDRS_PER_BLOCK(inode) * NIDS_PER_BLOCK;
+			child->pgofs += addrs * NIDS_PER_BLOCK;
 		}
 	}
 
@@ -3173,7 +3175,7 @@ static void fsck_disconnect_file_dnode(struct f2fs_sb_info *sbi,
 {
 	struct f2fs_node *node;
 	struct node_info ni;
-	u32 addr;
+	u32 addr, addr_per_block;
 	int i, err;
 
 	node = calloc(F2FS_BLKSIZE, 1);
@@ -3187,7 +3189,8 @@ static void fsck_disconnect_file_dnode(struct f2fs_sb_info *sbi,
 	release_block_cnt(sbi, dealloc);
 	release_block(sbi, ni.blk_addr, dealloc);
 
-	for (i = 0; i < ADDRS_PER_BLOCK(inode); i++) {
+	addr_per_block = ADDRS_PER_BLOCK(inode);
+	for (i = 0; i < addr_per_block; i++) {
 		addr = le32_to_cpu(node->dn.addr[i]);
 		if (!addr)
 			continue;
@@ -3302,8 +3305,10 @@ static void fsck_disconnect_file(struct f2fs_sb_info *sbi, nid_t ino,
 
 	/* clear data counters */
 	if (!(node->i.i_inline & (F2FS_INLINE_DATA | F2FS_INLINE_DENTRY))) {
+		u32 addrs = ADDRS_PER_INODE(&node->i);
+
 		ofs = get_extra_isize(node);
-		for (i = 0; i < ADDRS_PER_INODE(&node->i); i++) {
+		for (i = 0; i < addrs; i++) {
 			block_t addr = le32_to_cpu(node->i.i_addr[ofs + i]);
 			if (!addr)
 				continue;
fsck/mount.c +3 -1
index 6f640a0..990f1e0 100644
@@ -281,6 +281,7 @@ void print_inode_info(struct f2fs_sb_info *sbi,
 	u32 namelen = le32_to_cpu(inode->i_namelen);
 	int enc_name = file_enc_name(inode);
 	int ofs = get_extra_isize(node);
+	u32 addrs;
 
 	pretty_print_filename(inode->i_name, namelen, en, enc_name);
 	if (name && en[0]) {
@@ -350,7 +351,8 @@ void print_inode_info(struct f2fs_sb_info *sbi,
 		}
 	}
 
-	for (i = 0; i < ADDRS_PER_INODE(inode); i++) {
+	addrs = ADDRS_PER_INODE(inode);
+	for (i = 0; i < addrs; i++) {
 		block_t blkaddr;
 		char *flag = "";
 
include/f2fs_fs.h +11 -1
index 3095b59..d0877b9 100644
@@ -1660,7 +1660,6 @@ struct f2fs_configuration {
 extern int utf8_to_utf16(char *, const char *, size_t, size_t);
 extern int utf16_to_utf8(char *, const char *, size_t, size_t);
 extern int log_base_2(uint32_t);
-extern unsigned int addrs_per_page(struct f2fs_inode *, bool);
 extern u64 f2fs_max_file_offset(struct f2fs_inode *);
 extern __u32 f2fs_inode_chksum(struct f2fs_node *);
 extern __u32 f2fs_checkpoint_chksum(struct f2fs_checkpoint *);
@@ -2212,4 +2211,15 @@ static inline bool __time_to_inject(int type, const char *func,
 	return false;
 }
 
+static inline unsigned int addrs_per_page(struct f2fs_inode *i, bool is_inode)
+{
+	unsigned int addrs = is_inode ? CUR_ADDRS_PER_INODE(i) -
+		get_inline_xattr_addrs(i) : DEF_ADDRS_PER_BLOCK;
+
+	if (!LINUX_S_ISREG(le16_to_cpu(i->i_mode)) ||
+			!(le32_to_cpu(i->i_flags) & F2FS_COMPR_FL))
+		return addrs;
+	return ALIGN_DOWN(addrs, 1 << i->i_log_cluster_size);
+}
+
 #endif	/*__F2FS_FS_H */
lib/libf2fs.c +0 -11
index 1a496b7..13e4d0d 100644
@@ -516,17 +516,6 @@ opaque_seq:
 	return __f2fs_dentry_hash(name, len);
 }
 
-unsigned int addrs_per_page(struct f2fs_inode *i, bool is_inode)
-{
-	unsigned int addrs = is_inode ? CUR_ADDRS_PER_INODE(i) -
-		get_inline_xattr_addrs(i) : DEF_ADDRS_PER_BLOCK;
-
-	if (!LINUX_S_ISREG(le16_to_cpu(i->i_mode)) ||
-			!(le32_to_cpu(i->i_flags) & F2FS_COMPR_FL))
-		return addrs;
-	return ALIGN_DOWN(addrs, 1 << i->i_log_cluster_size);
-}
-
 u64 f2fs_max_file_offset(struct f2fs_inode *i)
 {
 	if (!LINUX_S_ISREG(le16_to_cpu(i->i_mode)) ||
Signature · 8 lines
2.51.0



_______________________________________________
Linux-f2fs-devel mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/linux-f2fs-devel
#1 lore raw
Quoted 70 lines · Wu Bo

ADDRS_PER_INODE and ADDRS_PER_BLOCK are macros of a function, and if the compiler does not optimize the code (-O0 for example), they would be called in every single loop, which could cause significant performance drops in some cases.

This scenario has been found before, refer to:
Commit: 1bb669e ("fsck.f2fs: avoid unnecessary recalculation")

To avoid this performance drop as much as possible, I changed the function to 'inline' definition, which could mitigate such cases in the future.

---
  fsck/dump.c       |  5 +++--
  fsck/fsck.c       | 15 ++++++++++-----
  fsck/mount.c      |  4 +++-
  include/f2fs_fs.h | 12 +++++++++++-
  lib/libf2fs.c     | 11 -----------
  5 files changed, 27 insertions(+), 20 deletions(-)
fsck/dump.c +3 -2
index f5c95de..a5f344b 100644
@@ -495,7 +495,7 @@ static int dump_inode_blk(struct f2fs_sb_info *sbi, u32 nid,
  {
  	u32 i = 0;
  	u64 ofs = 0;
-	u32 addr_per_block;
+	u32 addr_per_block, addr_per_inode;
  	u16 type = le16_to_cpu(node_blk->i.i_mode);
  	int ret = 0;
 
@@ -543,9 +543,10 @@ static int dump_inode_blk(struct f2fs_sb_info *sbi, u32 nid,
  	}
 
  	addr_per_block = ADDRS_PER_BLOCK(&node_blk->i);
+	addr_per_inode = ADDRS_PER_INODE(&node_blk->i);
 
  	/* check data blocks in inode */
-	for (i = 0; i < ADDRS_PER_INODE(&node_blk->i); i++, ofs++)
+	for (i = 0; i < addr_per_inode; i++, ofs++)
  		dump_data_blk(sbi, ofs * F2FS_BLKSIZE, le32_to_cpu(
  			node_blk->i.i_addr[get_extra_isize(node_blk) + i]), type);
 
fsck/fsck.c +3 -1
index db44f9d..1627bd0 100644
@@ -1516,8 +1516,9 @@ int fsck_chk_dnode_blk(struct f2fs_sb_info *sbi, struct f2fs_inode *inode,
  	bool compressed = i_flags & F2FS_COMPR_FL;
  	bool compr_rel = inode->i_inline & F2FS_COMPRESS_RELEASED;
  	u32 cluster_size = 1 << inode->i_log_cluster_size;
+	u32 addrs = ADDRS_PER_BLOCK(inode);
 
-	for (idx = 0; idx < ADDRS_PER_BLOCK(inode); idx++, child->pgofs++) {
+	for (idx = 0; idx < addrs; idx++, child->pgofs++) {
  		block_t blkaddr = le32_to_cpu(node_blk->dn.addr[idx]);
 
  		check_extent_info(child, blkaddr, 0);
@@ -1617,6 +1618,7 @@ int fsck_chk_didnode_blk(struct f2fs_sb_info *sbi, struct f2fs_inode *inode,
  {
  	int i = 0;
  	int need_fix = 0, ret = 0;
+	u32 addrs = ADDRS_PER_BLOCK(inode);
 
  	fsck_reada_all_direct_node_blocks(sbi, node_blk);
 

fsck/fsck.c int fsck_chk_didnode_blk(struct f2fs_sb_info *sbi, struct f2fs_inode *inode, · line 1642

Quoted 5 lines
fsck/fsck.c +1 -1
@@ -1637,7 +1639,7 @@ int fsck_chk_didnode_blk(struct f2fs_sb_info *sbi, struct f2fs_inode *inode,
  				FIX_MSG("Set double indirect node 0x%x -> 0", i);
  			}
  skip:
-			child->pgofs += ADDRS_PER_BLOCK(inode) * NIDS_PER_BLOCK;
+			child->pgofs += addrs * NIDS_PER_BLOCK;

I'm worried about that ADDRS_PER_BLOCK(inode) will change after we fixed some fields of inode block in fsck_chk_node_blk().

Thanks,

Quoted 96 lines
fsck/fsck.c +6 -3
  		}
  	}
 
@@ -3173,7 +3175,7 @@ static void fsck_disconnect_file_dnode(struct f2fs_sb_info *sbi,
  {
  	struct f2fs_node *node;
  	struct node_info ni;
-	u32 addr;
+	u32 addr, addr_per_block;
  	int i, err;
 
  	node = calloc(F2FS_BLKSIZE, 1);
@@ -3187,7 +3189,8 @@ static void fsck_disconnect_file_dnode(struct f2fs_sb_info *sbi,
  	release_block_cnt(sbi, dealloc);
  	release_block(sbi, ni.blk_addr, dealloc);
 
-	for (i = 0; i < ADDRS_PER_BLOCK(inode); i++) {
+	addr_per_block = ADDRS_PER_BLOCK(inode);
+	for (i = 0; i < addr_per_block; i++) {
  		addr = le32_to_cpu(node->dn.addr[i]);
  		if (!addr)
  			continue;
@@ -3302,8 +3305,10 @@ static void fsck_disconnect_file(struct f2fs_sb_info *sbi, nid_t ino,
 
  	/* clear data counters */
  	if (!(node->i.i_inline & (F2FS_INLINE_DATA | F2FS_INLINE_DENTRY))) {
+		u32 addrs = ADDRS_PER_INODE(&node->i);
+
  		ofs = get_extra_isize(node);
-		for (i = 0; i < ADDRS_PER_INODE(&node->i); i++) {
+		for (i = 0; i < addrs; i++) {
  			block_t addr = le32_to_cpu(node->i.i_addr[ofs + i]);
  			if (!addr)
  				continue;
fsck/mount.c +3 -1
index 6f640a0..990f1e0 100644
@@ -281,6 +281,7 @@ void print_inode_info(struct f2fs_sb_info *sbi,
  	u32 namelen = le32_to_cpu(inode->i_namelen);
  	int enc_name = file_enc_name(inode);
  	int ofs = get_extra_isize(node);
+	u32 addrs;
 
  	pretty_print_filename(inode->i_name, namelen, en, enc_name);
  	if (name && en[0]) {
@@ -350,7 +351,8 @@ void print_inode_info(struct f2fs_sb_info *sbi,
  		}
  	}
 
-	for (i = 0; i < ADDRS_PER_INODE(inode); i++) {
+	addrs = ADDRS_PER_INODE(inode);
+	for (i = 0; i < addrs; i++) {
  		block_t blkaddr;
  		char *flag = "";
 
include/f2fs_fs.h +11 -1
index 3095b59..d0877b9 100644
@@ -1660,7 +1660,6 @@ struct f2fs_configuration {
  extern int utf8_to_utf16(char *, const char *, size_t, size_t);
  extern int utf16_to_utf8(char *, const char *, size_t, size_t);
  extern int log_base_2(uint32_t);
-extern unsigned int addrs_per_page(struct f2fs_inode *, bool);
  extern u64 f2fs_max_file_offset(struct f2fs_inode *);
  extern __u32 f2fs_inode_chksum(struct f2fs_node *);
  extern __u32 f2fs_checkpoint_chksum(struct f2fs_checkpoint *);
@@ -2212,4 +2211,15 @@ static inline bool __time_to_inject(int type, const char *func,
  	return false;
  }
 
+static inline unsigned int addrs_per_page(struct f2fs_inode *i, bool is_inode)
+{
+	unsigned int addrs = is_inode ? CUR_ADDRS_PER_INODE(i) -
+		get_inline_xattr_addrs(i) : DEF_ADDRS_PER_BLOCK;
+
+	if (!LINUX_S_ISREG(le16_to_cpu(i->i_mode)) ||
+			!(le32_to_cpu(i->i_flags) & F2FS_COMPR_FL))
+		return addrs;
+	return ALIGN_DOWN(addrs, 1 << i->i_log_cluster_size);
+}
+
  #endif	/*__F2FS_FS_H */
lib/libf2fs.c +0 -4
index 1a496b7..13e4d0d 100644
@@ -516,17 +516,6 @@ opaque_seq:
  	return __f2fs_dentry_hash(name, len);
  }
 
-unsigned int addrs_per_page(struct f2fs_inode *i, bool is_inode)
-{
-	unsigned int addrs = is_inode ? CUR_ADDRS_PER_INODE(i) -
-		get_inline_xattr_addrs(i) : DEF_ADDRS_PER_BLOCK;

lib/libf2fs.c opaque_seq: · line 521

Quoted 10 lines
lib/libf2fs.c +0 -7
-
-	if (!LINUX_S_ISREG(le16_to_cpu(i->i_mode)) ||
-			!(le32_to_cpu(i->i_flags) & F2FS_COMPR_FL))
-		return addrs;
-	return ALIGN_DOWN(addrs, 1 << i->i_log_cluster_size);
-}
-
  u64 f2fs_max_file_offset(struct f2fs_inode *i)
  {
  	if (!LINUX_S_ISREG(le16_to_cpu(i->i_mode)) ||

_______________________________________________
Linux-f2fs-devel mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/linux-f2fs-devel