Skip to contentSkip to conversation

f2fs: support data flush in checkpoint

PATCH In review Chao Yu 2 participants 3 messages Dec 15, 2015 – Dec 16, 2015 F2FS file system

lore
Share & tools
lore

Chao Yu

author
#0 lore raw

PATCH 5/8 f2fs: support data flush in checkpoint

Previously, when finishing a checkpoint, we only keep consistency of all fs meta info including meta inode, node inode, dentry page of directory inode, so, after a sudden power cut, f2fs can recover from last checkpoint with full directory structure.

But during checkpoint, we didn't flush dirty pages of regular and symlink inode, so such dirty datas still in memory will be lost in that moment of power off.

In order to reduce the chance of lost data, this patch tries to flush user data before starting a checkpoint. So user's data written between two checkpoint which may not be fsynced could be saved.

---
 fs/f2fs/checkpoint.c | 11 +++++++++++
 fs/f2fs/f2fs.h       |  5 +++++
 2 files changed, 16 insertions(+)
fs/f2fs/checkpoint.c +11 -0
index f33c4d7..217571f 100644
@@ -849,6 +849,17 @@ static int block_operations(struct f2fs_sb_info *sbi)
 
 	blk_start_plug(&plug);
 
+retry_flush_datas:
+	/* write all the dirty data pages */
+	if (get_pages(sbi, F2FS_DIRTY_DATAS)) {
+		sync_dirty_inodes(sbi, FILE_INODE);
+		if (unlikely(f2fs_cp_error(sbi))) {
+			err = -EIO;
+			goto out;
+		}
+		goto retry_flush_datas;
+	}
+
 retry_flush_dents:
 	f2fs_lock_all(sbi);
 	/* write all the dirty dentry pages */
fs/f2fs/f2fs.h +5 -0
index d8bef3c..2477b2f5 100644
@@ -648,6 +648,7 @@ struct f2fs_sm_info {
 enum count_type {
 	F2FS_WRITEBACK,
 	F2FS_DIRTY_DENTS,
+	F2FS_DIRTY_DATAS,
 	F2FS_DIRTY_NODES,
 	F2FS_DIRTY_META,
 	F2FS_INMEM_PAGES,
@@ -1068,6 +1069,8 @@ static inline void inode_inc_dirty_pages(struct inode *inode)
 	atomic_inc(&F2FS_I(inode)->dirty_pages);
 	if (S_ISDIR(inode->i_mode))
 		inc_page_count(F2FS_I_SB(inode), F2FS_DIRTY_DENTS);
+	else
+		inc_page_count(F2FS_I_SB(inode), F2FS_DIRTY_DATAS);
 }
 
 static inline void dec_page_count(struct f2fs_sb_info *sbi, int count_type)
@@ -1085,6 +1088,8 @@ static inline void inode_dec_dirty_pages(struct inode *inode)
 
 	if (S_ISDIR(inode->i_mode))
 		dec_page_count(F2FS_I_SB(inode), F2FS_DIRTY_DENTS);
+	else
+		dec_page_count(F2FS_I_SB(inode), F2FS_DIRTY_DATAS);
 }
 
 static inline int get_pages(struct f2fs_sb_info *sbi, int count_type)
Signature · 1 line
2.6.3
maintainer
#1 lore raw

Hi Chao,

Quoted 28 lines · Chao Yu

Previously, when finishing a checkpoint, we only keep consistency of all fs meta info including meta inode, node inode, dentry page of directory inode, so, after a sudden power cut, f2fs can recover from last checkpoint with full directory structure.

But during checkpoint, we didn't flush dirty pages of regular and symlink inode, so such dirty datas still in memory will be lost in that moment of power off.

In order to reduce the chance of lost data, this patch tries to flush user data before starting a checkpoint. So user's data written between two checkpoint which may not be fsynced could be saved.

---
 fs/f2fs/checkpoint.c | 11 +++++++++++
 fs/f2fs/f2fs.h       |  5 +++++
 2 files changed, 16 insertions(+)
fs/f2fs/checkpoint.c +1 -0
index f33c4d7..217571f 100644
@@ -849,6 +849,17 @@ static int block_operations(struct f2fs_sb_info *sbi)
 
 	blk_start_plug(&plug);
 
+retry_flush_datas:

fs/f2fs/checkpoint.c static int block_operations(struct f2fs_sb_info *sbi) · line 862

Quoted 10 lines
fs/f2fs/checkpoint.c +10 -0
+	/* write all the dirty data pages */
+	if (get_pages(sbi, F2FS_DIRTY_DATAS)) {
+		sync_dirty_inodes(sbi, FILE_INODE);
+		if (unlikely(f2fs_cp_error(sbi))) {
+			err = -EIO;
+			goto out;
+		}
+		goto retry_flush_datas;
+	}
+

Please don't do like this; our checkpoint should be different from system sync. I don't want to increase the checkpoint latency at all even in f2fs_gc as well.

I think that something like this would be better.

In f2fs_balance_fs_bg(),

	if (!avaliable_free_memory() || ...) {
		if (test_opt(DATA_FLUSH))
			sync_dirty_inodes(FILE_INODE);
		f2fs_sync_fs();
	}

Please add its mount option with a disabled one first.

fs/f2fs/f2fs.h struct f2fs_sm_info { · line 651

Quoted 12 lines
fs/f2fs/checkpoint.c +0 -0
 retry_flush_dents:
 	f2fs_lock_all(sbi);
 	/* write all the dirty dentry pages */
fs/f2fs/f2fs.h +1 -0
index d8bef3c..2477b2f5 100644
@@ -648,6 +648,7 @@ struct f2fs_sm_info {
 enum count_type {
 	F2FS_WRITEBACK,
 	F2FS_DIRTY_DENTS,
+	F2FS_DIRTY_DATAS,

This should be F2FS_DIRTY_DATA.

And, it'd better to merge this part into your previous patch:
"f2fs: record dirty status of regular/symlink inodes"

I'll do that, so please check it out later from dev-test.

fs/f2fs/f2fs.h static inline void inode_inc_dirty_pages(struct inode *inode) · line 1073

Quoted 9 lines
fs/f2fs/f2fs.h +2 -0
 	F2FS_DIRTY_NODES,
 	F2FS_DIRTY_META,
 	F2FS_INMEM_PAGES,
@@ -1068,6 +1069,8 @@ static inline void inode_inc_dirty_pages(struct inode *inode)
 	atomic_inc(&F2FS_I(inode)->dirty_pages);
 	if (S_ISDIR(inode->i_mode))
 		inc_page_count(F2FS_I_SB(inode), F2FS_DIRTY_DENTS);
+	else
+		inc_page_count(F2FS_I_SB(inode), F2FS_DIRTY_DATAS);

ditto.

fs/f2fs/f2fs.h static inline void inode_dec_dirty_pages(struct inode *inode) · line 1092

Quoted 9 lines
fs/f2fs/f2fs.h +2 -0
 }
 
 static inline void dec_page_count(struct f2fs_sb_info *sbi, int count_type)
@@ -1085,6 +1088,8 @@ static inline void inode_dec_dirty_pages(struct inode *inode)
 
 	if (S_ISDIR(inode->i_mode))
 		dec_page_count(F2FS_I_SB(inode), F2FS_DIRTY_DENTS);
+	else
+		dec_page_count(F2FS_I_SB(inode), F2FS_DIRTY_DATAS);

ditto.

Quoted 6 lines
fs/f2fs/f2fs.h +0 -0
 }
 
 static inline int get_pages(struct f2fs_sb_info *sbi, int count_type)
Signature · 1 line
2.6.3
author
#2 lore raw

Hi Jaegeuk,

Quoted 37 lines

-----Original Message-----
From: Jaegeuk Kim [mailto:[email protected]]
Sent: Wednesday, December 16, 2015 5:59 AM
To: Chao Yu
Cc: [email protected]; [email protected]
Subject: Re: [PATCH 5/8] f2fs: support data flush in checkpoint

Hi Chao,

Quoted 27 lines · Chao Yu

Previously, when finishing a checkpoint, we only keep consistency of all fs meta info including meta inode, node inode, dentry page of directory inode, so, after a sudden power cut, f2fs can recover from last checkpoint with full directory structure.

But during checkpoint, we didn't flush dirty pages of regular and symlink inode, so such dirty datas still in memory will be lost in that moment of power off.

In order to reduce the chance of lost data, this patch tries to flush user data before starting a checkpoint. So user's data written between two checkpoint which may not be fsynced could be saved.

---
 fs/f2fs/checkpoint.c | 11 +++++++++++
 fs/f2fs/f2fs.h       |  5 +++++
 2 files changed, 16 insertions(+)
fs/f2fs/checkpoint.c +0 -0
index f33c4d7..217571f 100644
@@ -849,6 +849,17 @@ static int block_operations(struct f2fs_sb_info *sbi)
 
 	blk_start_plug(&plug);
 
Quoted 14 lines

fs/f2fs/checkpoint.c static int block_operations(struct f2fs_sb_info *sbi) · line 862

Quoted 11 lines
fs/f2fs/checkpoint.c +11 -0
+retry_flush_datas:
+	/* write all the dirty data pages */
+	if (get_pages(sbi, F2FS_DIRTY_DATAS)) {
+		sync_dirty_inodes(sbi, FILE_INODE);
+		if (unlikely(f2fs_cp_error(sbi))) {
+			err = -EIO;
+			goto out;
+		}
+		goto retry_flush_datas;
+	}
+

Please don't do like this; our checkpoint should be different from system sync. I don't want to increase the checkpoint latency at all even in f2fs_gc as well.

Alright.

Since flushing dirty data definitely increasing our latency of the interface, so what I thought was that let user take responsibility for risk of potential longer latency once user choose to mount with data_flush option.

Now, look into this implementation, it seems that adding flushing operation into checkpoint looks advancing rashly, it spreads random potential long latency everywhere.

Quoted 2 lines

I think that something like this would be better.

I will follow that currently. :)

Quoted 9 lines

In f2fs_balance_fs_bg(),

	if (!avaliable_free_memory() || ...) {
		if (test_opt(DATA_FLUSH))
			sync_dirty_inodes(FILE_INODE);
		f2fs_sync_fs();
	}

Please add its mount option with a disabled one first.

OK, I will do that.

Quoted 16 lines

fs/f2fs/f2fs.h struct f2fs_sm_info { · line 651

Quoted 13 lines
fs/f2fs/checkpoint.c +0 -0
 retry_flush_dents:
 	f2fs_lock_all(sbi);
 	/* write all the dirty dentry pages */
fs/f2fs/f2fs.h +1 -0
index d8bef3c..2477b2f5 100644
@@ -648,6 +648,7 @@ struct f2fs_sm_info {
 enum count_type {
 	F2FS_WRITEBACK,
 	F2FS_DIRTY_DENTS,
+	F2FS_DIRTY_DATAS,

This should be F2FS_DIRTY_DATA.

Quoted 4 lines

And, it'd better to merge this part into your previous patch:
"f2fs: record dirty status of regular/symlink inodes"

I'll do that, so please check it out later from dev-test.

Thanks for doing this, still I will resend the patches for the people who wants to comment on them.

Thanks,

Quoted 31 lines

fs/f2fs/f2fs.h static inline void inode_inc_dirty_pages(struct inode *inode) · line 1073

Quoted 10 lines
fs/f2fs/f2fs.h +2 -0
 	F2FS_DIRTY_NODES,
 	F2FS_DIRTY_META,
 	F2FS_INMEM_PAGES,
@@ -1068,6 +1069,8 @@ static inline void inode_inc_dirty_pages(struct inode *inode)
 	atomic_inc(&F2FS_I(inode)->dirty_pages);
 	if (S_ISDIR(inode->i_mode))
 		inc_page_count(F2FS_I_SB(inode), F2FS_DIRTY_DENTS);
+	else
+		inc_page_count(F2FS_I_SB(inode), F2FS_DIRTY_DATAS);

ditto.

fs/f2fs/f2fs.h static inline void inode_dec_dirty_pages(struct inode *inode) · line 1092

Quoted 9 lines
fs/f2fs/f2fs.h +2 -0
 }
 
 static inline void dec_page_count(struct f2fs_sb_info *sbi, int count_type)
@@ -1085,6 +1088,8 @@ static inline void inode_dec_dirty_pages(struct inode *inode)
 
 	if (S_ISDIR(inode->i_mode))
 		dec_page_count(F2FS_I_SB(inode), F2FS_DIRTY_DENTS);
+	else
+		dec_page_count(F2FS_I_SB(inode), F2FS_DIRTY_DATAS);

ditto.

Quoted 6 lines
fs/f2fs/f2fs.h +0 -1
 }
 
 static inline int get_pages(struct f2fs_sb_info *sbi, int count_type)
--

2.6.3