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.
Signed-off-by: Chao Yu <[email protected]>
---
fs/f2fs/checkpoint.c | 11 +++++++++++
fs/f2fs/f2fs.h | 5 +++++
2 files changed, 16 insertions(+)
diff --git a/fs/f2fs/checkpoint.c b/fs/f2fs/checkpoint.c
index f33c4d7..217571f 100644
--- a/fs/f2fs/checkpoint.c
+++ b/fs/f2fs/checkpoint.c
@@ -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 */
diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h
index d8bef3c..2477b2f5 100644
--- a/fs/f2fs/f2fs.h
+++ b/fs/f2fs/f2fs.h
@@ -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)
--
2.6.3
Hi Chao,
On Tue, Dec 15, 2015 at 01:33:18PM +0800, Chao Yu wrote:
> 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.
>
> Signed-off-by: Chao Yu <[email protected]>
> ---
> fs/f2fs/checkpoint.c | 11 +++++++++++
> fs/f2fs/f2fs.h | 5 +++++
> 2 files changed, 16 insertions(+)
>
> diff --git a/fs/f2fs/checkpoint.c b/fs/f2fs/checkpoint.c
> index f33c4d7..217571f 100644
> --- a/fs/f2fs/checkpoint.c
> +++ b/fs/f2fs/checkpoint.c
> @@ -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;
> + }
> +
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.
> retry_flush_dents:
> f2fs_lock_all(sbi);
> /* write all the dirty dentry pages */
> diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h
> index d8bef3c..2477b2f5 100644
> --- a/fs/f2fs/f2fs.h
> +++ b/fs/f2fs/f2fs.h
> @@ -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.
> 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.
> }
>
> 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.
> }
>
> static inline int get_pages(struct f2fs_sb_info *sbi, int count_type)
> --
> 2.6.3
>
Hi Jaegeuk,
> -----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,
>
> On Tue, Dec 15, 2015 at 01:33:18PM +0800, Chao Yu wrote:
> > 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.
> >
> > Signed-off-by: Chao Yu <[email protected]>
> > ---
> > fs/f2fs/checkpoint.c | 11 +++++++++++
> > fs/f2fs/f2fs.h | 5 +++++
> > 2 files changed, 16 insertions(+)
> >
> > diff --git a/fs/f2fs/checkpoint.c b/fs/f2fs/checkpoint.c
> > index f33c4d7..217571f 100644
> > --- a/fs/f2fs/checkpoint.c
> > +++ b/fs/f2fs/checkpoint.c
> > @@ -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;
> > + }
> > +
>
> 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.
>
> I think that something like this would be better.
I will follow that currently. :)
>
> 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.
>
> > retry_flush_dents:
> > f2fs_lock_all(sbi);
> > /* write all the dirty dentry pages */
> > diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h
> > index d8bef3c..2477b2f5 100644
> > --- a/fs/f2fs/f2fs.h
> > +++ b/fs/f2fs/f2fs.h
> > @@ -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.
Thanks for doing this, still I will resend the patches for the people
who wants to comment on them.
Thanks,
>
> > 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.
>
> > }
> >
> > 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.
>
> > }
> >
> > static inline int get_pages(struct f2fs_sb_info *sbi, int count_type)
> > --
> > 2.6.3
> >