* [PATCH] f2fs: skip checkpoint for compressed file write
@ 2025-08-21 2:40 Jiucheng Xu via B4 Relay
2025-08-25 2:16 ` Chao Yu
0 siblings, 1 reply; 2+ messages in thread
From: Jiucheng Xu via B4 Relay @ 2025-08-21 2:40 UTC (permalink / raw)
To: Jaegeuk Kim, Chao Yu
Cc: linux-f2fs-devel, linux-kernel, Tao Zeng, Jianxin Pan, Tuan Zhang,
Jiucheng Xu
From: Tao Zeng <tao.zeng@amlogic.com>
Always do checkpoint is a heavy behavior for compressed file.
But for contiguous writing of a file, checkpoint need to be
skipped to help improve performance.
Tested with iozone for always do check point on compressed data,
results are:
File stride size set to 17 * record size.
random random
KB reclen write rewrite read reread read write
102400 4 1314 35488 234231 1683793 1212394 35334
102400 8 2387 54133 244584 1871789 1644952 52478
102400 16 5060 7059 298052 1901792 1894929 6600
102400 32 9315 13118 424723 1825565 1924235 12041
102400 64 17028 22258 491181 1844443 1968247 22115
102400 128 30551 38008 445192 1796615 1917466 38527
102400 256 46944 55006 509625 1630910 1715586 56201
102400 512 63355 70432 434639 1406089 1487569 72718
102400 1024 83036 86742 447141 1420505 1503320 88913
102400 2048 98577 101971 450287 1434918 1522294 106374
102400 4096 113300 116994 451286 1435321 1527404 119579
102400 8192 132532 133044 488503 1458688 1540595 141167
102400 16384 143246 143857 489376 1469878 1556530 151362
We can see that writing speed of small pieces of data(less than 16KB) is
very slow.
With this change, iozone data are on the same hardware:
random random
KB reclen write rewrite read reread read write
102400 4 14658 34796 232797 1985764 1219513 34509
102400 8 25980 53695 233218 2419198 1788989 51927
102400 16 49556 50325 266754 2780871 2256746 50593
102400 32 79947 80783 393452 2755413 2467949 77681
102400 64 104866 105830 531853 2816504 2596812 106223
102400 128 142097 142034 651876 2885805 2728473 143296
102400 256 146972 144822 535727 2542080 2450922 157390
102400 512 126591 152480 571581 2055442 2052839 156512
102400 1024 135164 143667 654547 2052594 2045214 130488
102400 2048 127587 124889 491258 2058457 2059454 141273
102400 4096 124280 123959 660713 2067394 2056889 131967
102400 8192 138240 136233 509709 2102040 2090773 149215
102400 16384 146524 145333 576519 2118162 2096482 158704
We can see that speed of write with small pieces of data increased a lot.
Signed-off-by: Tao Zeng <tao.zeng@amlogic.com>
Signed-off-by: Jiucheng Xu <jiucheng.xu@amlogic.com>
---
fs/f2fs/f2fs.h | 1 +
fs/f2fs/file.c | 18 ++++++++++++++----
2 files changed, 15 insertions(+), 4 deletions(-)
diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h
index d6a49de1b7e919eda12354c074b8b253b2a9ea3f..2f820d531cdb32c0fc050aca05ffd3d00395a618 100644
--- a/fs/f2fs/f2fs.h
+++ b/fs/f2fs/f2fs.h
@@ -1531,6 +1531,7 @@ enum compress_algorithm_type {
enum compress_flag {
COMPRESS_CHKSUM,
+ COMPRESS_SKIP_WRITE_CP,
COMPRESS_MAX_FLAG,
};
diff --git a/fs/f2fs/file.c b/fs/f2fs/file.c
index 42faaed6a02da01f2bd117a5c55e1761beaffde6..3561b407f45e7aa97c7dcf911d4dddbc01ec2ca4 100644
--- a/fs/f2fs/file.c
+++ b/fs/f2fs/file.c
@@ -212,7 +212,9 @@ static inline enum cp_reason_type need_do_checkpoint(struct inode *inode)
if (!S_ISREG(inode->i_mode))
cp_reason = CP_NON_REGULAR;
- else if (f2fs_compressed_file(inode))
+ else if (f2fs_compressed_file(inode) &&
+ !(F2FS_I(inode)->i_compress_flag &
+ BIT(COMPRESS_SKIP_WRITE_CP)))
cp_reason = CP_COMPRESSED;
else if (inode->i_nlink != 1)
cp_reason = CP_HARDLINK;
@@ -5234,6 +5236,11 @@ static ssize_t f2fs_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
f2fs_dio_write_iter(iocb, from, &may_need_sync) :
f2fs_buffered_write_iter(iocb, from);
+ /* skip checkpoint for normal write compress file */
+ if (f2fs_compressed_file(inode))
+ F2FS_I(inode)->i_compress_flag |=
+ BIT(COMPRESS_SKIP_WRITE_CP);
+
trace_f2fs_datawrite_end(inode, orig_pos, ret);
}
@@ -5250,14 +5257,17 @@ static ssize_t f2fs_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
}
clear_inode_flag(inode, FI_PREALLOCATED_ALL);
+
+ if (ret > 0 && may_need_sync)
+ ret = generic_write_sync(iocb, ret);
+
+ if (f2fs_compressed_file(inode))
+ F2FS_I(inode)->i_compress_flag &= ~BIT(COMPRESS_SKIP_WRITE_CP);
out_unlock:
inode_unlock(inode);
out:
trace_f2fs_file_write_iter(inode, orig_pos, orig_count, ret);
- if (ret > 0 && may_need_sync)
- ret = generic_write_sync(iocb, ret);
-
/* If buffered IO was forced, flush and drop the data from
* the page cache to preserve O_DIRECT semantics
*/
---
base-commit: 3ea4ad0a1df0bcbfd5ccdcea56d57ca4678ae2a8
change-id: 20250820-dev-31b792e8e1fb
Best regards,
--
Jiucheng Xu <jiucheng.xu@amlogic.com>
^ permalink raw reply related [flat|nested] 2+ messages in thread
* Re: [PATCH] f2fs: skip checkpoint for compressed file write
2025-08-21 2:40 [PATCH] f2fs: skip checkpoint for compressed file write Jiucheng Xu via B4 Relay
@ 2025-08-25 2:16 ` Chao Yu
0 siblings, 0 replies; 2+ messages in thread
From: Chao Yu @ 2025-08-25 2:16 UTC (permalink / raw)
To: jiucheng.xu, Jaegeuk Kim
Cc: chao, linux-f2fs-devel, linux-kernel, Tao Zeng, Jianxin Pan,
Tuan Zhang
On 8/21/25 10:40, Jiucheng Xu via B4 Relay wrote:
> From: Tao Zeng <tao.zeng@amlogic.com>
>
> Always do checkpoint is a heavy behavior for compressed file.
> But for contiguous writing of a file, checkpoint need to be
> skipped to help improve performance.
>
> Tested with iozone for always do check point on compressed data,
> results are:
>
> File stride size set to 17 * record size.
> random random
> KB reclen write rewrite read reread read write
> 102400 4 1314 35488 234231 1683793 1212394 35334
> 102400 8 2387 54133 244584 1871789 1644952 52478
> 102400 16 5060 7059 298052 1901792 1894929 6600
> 102400 32 9315 13118 424723 1825565 1924235 12041
> 102400 64 17028 22258 491181 1844443 1968247 22115
> 102400 128 30551 38008 445192 1796615 1917466 38527
> 102400 256 46944 55006 509625 1630910 1715586 56201
> 102400 512 63355 70432 434639 1406089 1487569 72718
> 102400 1024 83036 86742 447141 1420505 1503320 88913
> 102400 2048 98577 101971 450287 1434918 1522294 106374
> 102400 4096 113300 116994 451286 1435321 1527404 119579
> 102400 8192 132532 133044 488503 1458688 1540595 141167
> 102400 16384 143246 143857 489376 1469878 1556530 151362
>
> We can see that writing speed of small pieces of data(less than 16KB) is
> very slow.
>
> With this change, iozone data are on the same hardware:
> random random
> KB reclen write rewrite read reread read write
> 102400 4 14658 34796 232797 1985764 1219513 34509
> 102400 8 25980 53695 233218 2419198 1788989 51927
> 102400 16 49556 50325 266754 2780871 2256746 50593
> 102400 32 79947 80783 393452 2755413 2467949 77681
> 102400 64 104866 105830 531853 2816504 2596812 106223
> 102400 128 142097 142034 651876 2885805 2728473 143296
> 102400 256 146972 144822 535727 2542080 2450922 157390
> 102400 512 126591 152480 571581 2055442 2052839 156512
> 102400 1024 135164 143667 654547 2052594 2045214 130488
> 102400 2048 127587 124889 491258 2058457 2059454 141273
> 102400 4096 124280 123959 660713 2067394 2056889 131967
> 102400 8192 138240 136233 509709 2102040 2090773 149215
> 102400 16384 146524 145333 576519 2118162 2096482 158704
>
> We can see that speed of write with small pieces of data increased a lot.
>
> Signed-off-by: Tao Zeng <tao.zeng@amlogic.com>
> Signed-off-by: Jiucheng Xu <jiucheng.xu@amlogic.com>
> ---
> fs/f2fs/f2fs.h | 1 +
> fs/f2fs/file.c | 18 ++++++++++++++----
> 2 files changed, 15 insertions(+), 4 deletions(-)
>
> diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h
> index d6a49de1b7e919eda12354c074b8b253b2a9ea3f..2f820d531cdb32c0fc050aca05ffd3d00395a618 100644
> --- a/fs/f2fs/f2fs.h
> +++ b/fs/f2fs/f2fs.h
> @@ -1531,6 +1531,7 @@ enum compress_algorithm_type {
>
> enum compress_flag {
> COMPRESS_CHKSUM,
> + COMPRESS_SKIP_WRITE_CP,
> COMPRESS_MAX_FLAG,
> };
>
> diff --git a/fs/f2fs/file.c b/fs/f2fs/file.c
> index 42faaed6a02da01f2bd117a5c55e1761beaffde6..3561b407f45e7aa97c7dcf911d4dddbc01ec2ca4 100644
> --- a/fs/f2fs/file.c
> +++ b/fs/f2fs/file.c
> @@ -212,7 +212,9 @@ static inline enum cp_reason_type need_do_checkpoint(struct inode *inode)
>
> if (!S_ISREG(inode->i_mode))
> cp_reason = CP_NON_REGULAR;
> - else if (f2fs_compressed_file(inode))
> + else if (f2fs_compressed_file(inode) &&
> + !(F2FS_I(inode)->i_compress_flag &
> + BIT(COMPRESS_SKIP_WRITE_CP)))
> cp_reason = CP_COMPRESSED;
IIRC, we can not avoid checkpoint for fsync on compressed file w/ this way,
since we haven't supported compressed file recovery yet.
You can check this w/ the way as below:
write -> fsync -> shutdown -> recovery -> check data&meta
Thanks,
> else if (inode->i_nlink != 1)
> cp_reason = CP_HARDLINK;
> @@ -5234,6 +5236,11 @@ static ssize_t f2fs_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
> f2fs_dio_write_iter(iocb, from, &may_need_sync) :
> f2fs_buffered_write_iter(iocb, from);
>
> + /* skip checkpoint for normal write compress file */
> + if (f2fs_compressed_file(inode))
> + F2FS_I(inode)->i_compress_flag |=
> + BIT(COMPRESS_SKIP_WRITE_CP);
> +
> trace_f2fs_datawrite_end(inode, orig_pos, ret);
> }
>
> @@ -5250,14 +5257,17 @@ static ssize_t f2fs_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
> }
>
> clear_inode_flag(inode, FI_PREALLOCATED_ALL);
> +
> + if (ret > 0 && may_need_sync)
> + ret = generic_write_sync(iocb, ret);
> +
> + if (f2fs_compressed_file(inode))
> + F2FS_I(inode)->i_compress_flag &= ~BIT(COMPRESS_SKIP_WRITE_CP);
> out_unlock:
> inode_unlock(inode);
> out:
> trace_f2fs_file_write_iter(inode, orig_pos, orig_count, ret);
>
> - if (ret > 0 && may_need_sync)
> - ret = generic_write_sync(iocb, ret);
> -
> /* If buffered IO was forced, flush and drop the data from
> * the page cache to preserve O_DIRECT semantics
> */
>
> ---
> base-commit: 3ea4ad0a1df0bcbfd5ccdcea56d57ca4678ae2a8
> change-id: 20250820-dev-31b792e8e1fb
>
> Best regards,
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2025-08-25 2:16 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2025-08-21 2:40 [PATCH] f2fs: skip checkpoint for compressed file write Jiucheng Xu via B4 Relay
2025-08-25 2:16 ` Chao Yu
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).