Re: [f2fs-dev] [PATCH 2/2] f2fs: avoid IO split due to mixed WB_SYNC_ALL and WB_SYNC_NONE

[heyunlei]

Hi Jaegeuk,

Can we split in place update bios into single sbi->f2fs_bio_info for more page
merged in out place update? This case can be show as below:

in place update submit a bio with one page
        out place update submit a bio with one page
in place update submit a bio with one page
        out place update submit a bio with one page
... ...

just like WB_SYNC_ALL and WB_SYNC_NONE case.

Thanks.

On 2017/3/30 4:48, Jaegeuk Kim wrote:

If two threads try to flush dirty pages in different inodes respectively,
f2fs_write_data_pages() will produce WRITE and WRITE_SYNC one at a time,
resulting in a lot of 4KB seperated IOs.

So, this patch gives higher priority to WB_SYNC_ALL IOs and gathers write
IOs with a big WRITE_SYNC'ed bio.

Signed-off-by: Jaegeuk Kim <jaeg...@kernel.org>
---
 fs/f2fs/data.c  | 15 +++++++++++++--
 fs/f2fs/f2fs.h  |  3 +++
 fs/f2fs/super.c |  2 ++
 3 files changed, 18 insertions(+), 2 deletions(-)

diff --git a/fs/f2fs/data.c b/fs/f2fs/data.c
index 8f36080b47c4..b1cac6d85bcb 100644
--- a/fs/f2fs/data.c
+++ b/fs/f2fs/data.c
@@ -1605,8 +1605,10 @@ static int f2fs_write_cache_pages(struct address_space 
*mapping,
                                last_idx = page->index;
                        }

-                       if (--wbc->nr_to_write <= 0 &&
-                           wbc->sync_mode == WB_SYNC_NONE) {
+                       /* give a priority to WB_SYNC threads */
+                       if ((atomic_read(&F2FS_M_SB(mapping)->wb_sync_req) ||
+                                       --wbc->nr_to_write <= 0) &&
+                                       wbc->sync_mode == WB_SYNC_NONE) {
                                done = 1;
                                break;
                        }
@@ -1662,9 +1664,18 @@ static int f2fs_write_data_pages(struct address_space 
*mapping,

        trace_f2fs_writepages(mapping->host, wbc, DATA);

+       /* to avoid spliting IOs due to mixed WB_SYNC_ALL and WB_SYNC_NONE */
+       if (wbc->sync_mode == WB_SYNC_ALL)
+               atomic_inc(&sbi->wb_sync_req);
+       else if (atomic_read(&sbi->wb_sync_req))
+               goto skip_write;
+
        blk_start_plug(&plug);
        ret = f2fs_write_cache_pages(mapping, wbc);
        blk_finish_plug(&plug);
+
+       if (wbc->sync_mode == WB_SYNC_ALL)
+               atomic_dec(&sbi->wb_sync_req);
        /*
         * if some pages were truncated, we cannot guarantee its mapping->host
         * to detect pending bios.
diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h
index 32d6f674c114..fd39db681226 100644
--- a/fs/f2fs/f2fs.h
+++ b/fs/f2fs/f2fs.h
@@ -888,6 +888,9 @@ struct f2fs_sb_info {
        /* # of allocated blocks */
        struct percpu_counter alloc_valid_block_count;

+       /* writeback control */
+       atomic_t wb_sync_req;                   /* count # of WB_SYNC threads */
+
        /* valid inode count */
        struct percpu_counter total_valid_inode_count;

diff --git a/fs/f2fs/super.c b/fs/f2fs/super.c
index 2d78f3c76d18..cb65e6d0d275 100644
--- a/fs/f2fs/super.c
+++ b/fs/f2fs/super.c
@@ -1566,6 +1566,8 @@ static void init_sb_info(struct f2fs_sb_info *sbi)
        for (i = 0; i < NR_COUNT_TYPE; i++)
                atomic_set(&sbi->nr_pages[i], 0);

+       atomic_set(&sbi->wb_sync_req, 0);
+
        INIT_LIST_HEAD(&sbi->s_list);
        mutex_init(&sbi->umount_mutex);
        mutex_init(&sbi->wio_mutex[NODE]);