2022-02-08 19:19:54

by Chao Yu

Subject: [PATCH] f2fs: avoid unneeded preflush during checkpoint()

During checkpoint, we have committed preflush command via f2fs_flush_device_cache()
to persist all metadata pages except last #2 CP pack page, so we don't need to
commit another preflush command in commit_checkpoint(), remove it to avoid unneeded
write cache overhead.

Signed-off-by: Chao Yu <[email protected]>
---
fs/f2fs/data.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/fs/f2fs/data.c b/fs/f2fs/data.c
index 8c417864c66a..15ac18bbbc8e 100644
--- a/fs/f2fs/data.c
+++ b/fs/f2fs/data.c
@@ -598,7 +598,7 @@ static void __f2fs_submit_merged_write(struct f2fs_sb_info *sbi,
io->fio.op = REQ_OP_WRITE;
io->fio.op_flags = REQ_META | REQ_PRIO | REQ_SYNC;
if (!test_opt(sbi, NOBARRIER))
- io->fio.op_flags |= REQ_PREFLUSH | REQ_FUA;
+ io->fio.op_flags |= REQ_FUA;
}
__submit_merged_bio(io);
up_write(&io->io_rwsem);
--
2.32.0

2022-02-08 22:37:18

by Chao Yu

Subject: Re: [PATCH] f2fs: avoid unneeded preflush during checkpoint()

On 2022/2/8 3:14, Jaegeuk Kim wrote:
> On 02/07, Chao Yu wrote:
>> During checkpoint, we have committed preflush command via f2fs_flush_device_cache()
>
> 787 int f2fs_flush_device_cache(struct f2fs_sb_info *sbi)
> 788 {
> 789 int ret = 0, i;
> 790
> 791 if (!f2fs_is_multi_device(sbi))
> 792 return 0;
>
> Seems a wrong assumption.

Oh, I missed that f2fs_flush_device_cache was introduced to flush devices except
first device when multi-device feature is on.

for (i = 1; i < sbi->s_ndevs; i++) {
int count = DEFAULT_RETRY_IO_COUNT;

So, this patch seems wrong...

Instead, below case looks wrong since it needs to trigger flush during checkpoint()
no matter nobarrier is on or off.

if (test_opt(sbi, NOBARRIER))
return 0;

Thanks,

>
>> to persist all metadata pages except last #2 CP pack page, so we don't need to
>> commit another preflush command in commit_checkpoint(), remove it to avoid unneeded
>> write cache overhead.
>>
>> Signed-off-by: Chao Yu <[email protected]>
>> ---
>> fs/f2fs/data.c | 2 +-
>> 1 file changed, 1 insertion(+), 1 deletion(-)
>>
>> diff --git a/fs/f2fs/data.c b/fs/f2fs/data.c
>> index 8c417864c66a..15ac18bbbc8e 100644
>> --- a/fs/f2fs/data.c
>> +++ b/fs/f2fs/data.c
>> @@ -598,7 +598,7 @@ static void __f2fs_submit_merged_write(struct f2fs_sb_info *sbi,
>> io->fio.op = REQ_OP_WRITE;
>> io->fio.op_flags = REQ_META | REQ_PRIO | REQ_SYNC;
>> if (!test_opt(sbi, NOBARRIER))
>> - io->fio.op_flags |= REQ_PREFLUSH | REQ_FUA;
>> + io->fio.op_flags |= REQ_FUA;
>> }
>> __submit_merged_bio(io);
>> up_write(&io->io_rwsem);
>> --
>> 2.32.0

2022-02-09 10:44:31

by Jaegeuk Kim

Subject: Re: [PATCH] f2fs: avoid unneeded preflush during checkpoint()

On 02/07, Chao Yu wrote:
> During checkpoint, we have committed preflush command via f2fs_flush_device_cache()

787 int f2fs_flush_device_cache(struct f2fs_sb_info *sbi)
788 {
789 int ret = 0, i;
790
791 if (!f2fs_is_multi_device(sbi))
792 return 0;

Seems a wrong assumption.

> to persist all metadata pages except last #2 CP pack page, so we don't need to
> commit another preflush command in commit_checkpoint(), remove it to avoid unneeded
> write cache overhead.
>
> Signed-off-by: Chao Yu <[email protected]>
> ---
> fs/f2fs/data.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/fs/f2fs/data.c b/fs/f2fs/data.c
> index 8c417864c66a..15ac18bbbc8e 100644
> --- a/fs/f2fs/data.c
> +++ b/fs/f2fs/data.c
> @@ -598,7 +598,7 @@ static void __f2fs_submit_merged_write(struct f2fs_sb_info *sbi,
> io->fio.op = REQ_OP_WRITE;
> io->fio.op_flags = REQ_META | REQ_PRIO | REQ_SYNC;
> if (!test_opt(sbi, NOBARRIER))
> - io->fio.op_flags |= REQ_PREFLUSH | REQ_FUA;
> + io->fio.op_flags |= REQ_FUA;
> }
> __submit_merged_bio(io);
> up_write(&io->io_rwsem);
> --
> 2.32.0