f2fs: write small sized IO to hot log

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 <[email protected]>

---
 fs/f2fs/data.c  | 15 +++++++++++++--
 fs/f2fs/f2fs.h  |  3 +++
 fs/f2fs/super.c |  2 ++
 3 files changed, 18 insertions(+), 2 deletions(-)

fs/f2fs/data.c +13 -2

index 8f36080b47c4..b1cac6d85bcb 100644
@@ -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.

fs/f2fs/f2fs.h +3 -0

index 32d6f674c114..fd39db681226 100644
@@ -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;
 

fs/f2fs/super.c +2 -0

index 2d78f3c76d18..cb65e6d0d275 100644
@@ -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]);

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 <[email protected]>

---
 fs/f2fs/data.c  | 15 +++++++++++++--
 fs/f2fs/f2fs.h  |  3 +++
 fs/f2fs/super.c |  2 ++
 3 files changed, 18 insertions(+), 2 deletions(-)

fs/f2fs/data.c +13 -2

index 8f36080b47c4..b1cac6d85bcb 100644
@@ -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.

fs/f2fs/f2fs.h +3 -0

index 32d6f674c114..fd39db681226 100644
@@ -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;
 

fs/f2fs/super.c +2 -0

index 2d78f3c76d18..cb65e6d0d275 100644
@@ -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]);

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.

From d9f00695c5e56c48611ade3ced89432ef2b59a27 Mon Sep 17 00:00:00 2001

It would better split small and large IOs separately in order to get more consecutive big writes.

The default threshold is set to 64KB, but configurable by sysfs/min_hot_blocks.

Signed-off-by Jaegeuk Kim <[email protected]>

---
 fs/f2fs/data.c    |  9 +++++++++
 fs/f2fs/f2fs.h    |  2 ++
 fs/f2fs/segment.c | 13 ++++++-------
 fs/f2fs/segment.h |  1 +
 fs/f2fs/super.c   |  2 ++
 5 files changed, 20 insertions(+), 7 deletions(-)

fs/f2fs/data.c +0 -0

index 090413236b27..8f36080b47c4 100644

fs/f2fs/data.c +7 -0

 	err = -EAGAIN;
 	if (f2fs_has_inline_data(inode)) {

@@ -1457,6 +1459,7 @@ static int __write_data_page(struct page *page, bool *submitted,
 	if (wbc->for_reclaim) {
 		f2fs_submit_merged_bio_cond(sbi, inode, 0, page->index,
 						DATA, WRITE);
+		clear_inode_flag(inode, FI_HOT_DATA);
 		remove_dirty_inode(inode);
 		submitted = NULL;
 	}

@@ -1511,6 +1514,12 @@ static int f2fs_write_cache_pages(struct address_space *mapping,
 
 	pagevec_init(&pvec, 0);
 
+	if (get_dirty_pages(mapping->host) <=
+				SM_I(F2FS_M_SB(mapping))->min_hot_blocks)
+		set_inode_flag(mapping->host, FI_HOT_DATA);
+	else
+		clear_inode_flag(mapping->host, FI_HOT_DATA);
+
 	if (wbc->range_cyclic) {
 		writeback_index = mapping->writeback_index; /* prev offset */
 		index = writeback_index;

fs/f2fs/f2fs.h +2 -0

index 5a49518ee786..32d6f674c114 100644
@@ -678,6 +678,7 @@ struct f2fs_sm_info {
 	unsigned int ipu_policy;	/* in-place-update policy */
 	unsigned int min_ipu_util;	/* in-place-update threshold */
 	unsigned int min_fsync_blocks;	/* threshold for fsync */
+	unsigned int min_hot_blocks;	/* threshold for hot block allocation */
 
 	/* for flush command control */
 	struct flush_cmd_control *fcc_info;

@@ -1717,6 +1718,7 @@ enum {
 	FI_DO_DEFRAG,		/* indicate defragment is running */
 	FI_DIRTY_FILE,		/* indicate regular/symlink has dirty pages */
 	FI_NO_PREALLOC,		/* indicate skipped preallocated blocks */
+	FI_HOT_DATA,		/* indicate file is hot */
 };
 
 static inline void __mark_inode_dirty_flag(struct inode *inode,

fs/f2fs/segment.c +6 -7

index b5b2a4745328..bff3f3bc7827 100644
@@ -1841,18 +1841,16 @@ static int __get_segment_type_6(struct page *page, enum page_type p_type)
 	if (p_type == DATA) {
 		struct inode *inode = page->mapping->host;
 
-		if (S_ISDIR(inode->i_mode))
-			return CURSEG_HOT_DATA;
-		else if (is_cold_data(page) || file_is_cold(inode))
+		if (is_cold_data(page) || file_is_cold(inode))
 			return CURSEG_COLD_DATA;
-		else
-			return CURSEG_WARM_DATA;
+		if (is_inode_flag_set(inode, FI_HOT_DATA))
+			return CURSEG_HOT_DATA;
+		return CURSEG_WARM_DATA;
 	} else {
 		if (IS_DNODE(page))
 			return is_cold_node(page) ? CURSEG_WARM_NODE :
 						CURSEG_HOT_NODE;
-		else
-			return CURSEG_COLD_NODE;
+		return CURSEG_COLD_NODE;
 	}
 }
 

@@ -2959,6 +2957,7 @@ int build_segment_manager(struct f2fs_sb_info *sbi)
 		sm_info->ipu_policy = 1 << F2FS_IPU_FSYNC;
 	sm_info->min_ipu_util = DEF_MIN_IPU_UTIL;
 	sm_info->min_fsync_blocks = DEF_MIN_FSYNC_BLOCKS;
+	sm_info->min_hot_blocks = DEF_MIN_HOT_BLOCKS;
 
 	sm_info->trim_sections = DEF_BATCHED_TRIM_SECTIONS;
 

fs/f2fs/segment.h +1 -0

index 31846b0fcb95..57e36c1ce7bd 100644
@@ -540,6 +540,7 @@ static inline int utilization(struct f2fs_sb_info *sbi)
  */
 #define DEF_MIN_IPU_UTIL	70
 #define DEF_MIN_FSYNC_BLOCKS	8
+#define DEF_MIN_HOT_BLOCKS	16
 
 enum {
 	F2FS_IPU_FORCE,

fs/f2fs/super.c +2 -0

index b4c5c6298698..2d78f3c76d18 100644
@@ -296,6 +296,7 @@ F2FS_RW_ATTR(SM_INFO, f2fs_sm_info, batched_trim_sections, trim_sections);
 F2FS_RW_ATTR(SM_INFO, f2fs_sm_info, ipu_policy, ipu_policy);
 F2FS_RW_ATTR(SM_INFO, f2fs_sm_info, min_ipu_util, min_ipu_util);
 F2FS_RW_ATTR(SM_INFO, f2fs_sm_info, min_fsync_blocks, min_fsync_blocks);
+F2FS_RW_ATTR(SM_INFO, f2fs_sm_info, min_hot_blocks, min_hot_blocks);
 F2FS_RW_ATTR(NM_INFO, f2fs_nm_info, ram_thresh, ram_thresh);
 F2FS_RW_ATTR(NM_INFO, f2fs_nm_info, ra_nid_pages, ra_nid_pages);
 F2FS_RW_ATTR(NM_INFO, f2fs_nm_info, dirty_nats_ratio, dirty_nats_ratio);

@@ -321,6 +322,7 @@ static struct attribute *f2fs_attrs[] = {
 	ATTR_LIST(ipu_policy),
 	ATTR_LIST(min_ipu_util),
 	ATTR_LIST(min_fsync_blocks),
+	ATTR_LIST(min_hot_blocks),
 	ATTR_LIST(max_victim_search),
 	ATTR_LIST(dir_level),
 	ATTR_LIST(ram_thresh),

Hi Jaegeuk,

Quoted 14 lines · Jaegeuk Kim

It would better split small and large IOs separately in order to get more consecutive big writes.

The default threshold is set to 64KB, but configurable by sysfs/min_hot_blocks.

Signed-off-by Jaegeuk Kim <[email protected]>

---
 fs/f2fs/data.c    |  9 +++++++++
 fs/f2fs/f2fs.h    |  2 ++
 fs/f2fs/segment.c | 13 ++++++-------
 fs/f2fs/segment.h |  1 +
 fs/f2fs/super.c   |  2 ++
 5 files changed, 20 insertions(+), 7 deletions(-)

fs/f2fs/data.c static int __write_data_page(struct page *page, bool *submitted, · line 1436

Quoted 10 lines

fs/f2fs/data.c +2 -0

index 090413236b27..8f36080b47c4 100644
@@ -1432,6 +1432,8 @@ static int __write_data_page(struct page *page, bool *submitted,
 		need_balance_fs = true;
 	else if (has_not_enough_free_secs(sbi, 0, 0))
 		goto redirty_out;
+	else
+		set_inode_flag(inode, FI_HOT_DATA);

Why here we need this, can you explain more about this?

Thanks.

Quoted 113 lines

fs/f2fs/data.c +7 -0

 	err = -EAGAIN;
 	if (f2fs_has_inline_data(inode)) {

@@ -1457,6 +1459,7 @@ static int __write_data_page(struct page *page, bool *submitted,
 	if (wbc->for_reclaim) {
 		f2fs_submit_merged_bio_cond(sbi, inode, 0, page->index,
 						DATA, WRITE);
+		clear_inode_flag(inode, FI_HOT_DATA);
 		remove_dirty_inode(inode);
 		submitted = NULL;
 	}

@@ -1511,6 +1514,12 @@ static int f2fs_write_cache_pages(struct address_space *mapping,
 
 	pagevec_init(&pvec, 0);
 
+	if (get_dirty_pages(mapping->host) <=
+				SM_I(F2FS_M_SB(mapping))->min_hot_blocks)
+		set_inode_flag(mapping->host, FI_HOT_DATA);
+	else
+		clear_inode_flag(mapping->host, FI_HOT_DATA);
+
 	if (wbc->range_cyclic) {
 		writeback_index = mapping->writeback_index; /* prev offset */
 		index = writeback_index;

fs/f2fs/f2fs.h +2 -0

index 5a49518ee786..32d6f674c114 100644
@@ -678,6 +678,7 @@ struct f2fs_sm_info {
 	unsigned int ipu_policy;	/* in-place-update policy */
 	unsigned int min_ipu_util;	/* in-place-update threshold */
 	unsigned int min_fsync_blocks;	/* threshold for fsync */
+	unsigned int min_hot_blocks;	/* threshold for hot block allocation */
 
 	/* for flush command control */
 	struct flush_cmd_control *fcc_info;

@@ -1717,6 +1718,7 @@ enum {
 	FI_DO_DEFRAG,		/* indicate defragment is running */
 	FI_DIRTY_FILE,		/* indicate regular/symlink has dirty pages */
 	FI_NO_PREALLOC,		/* indicate skipped preallocated blocks */
+	FI_HOT_DATA,		/* indicate file is hot */
 };
 
 static inline void __mark_inode_dirty_flag(struct inode *inode,

fs/f2fs/segment.c +6 -7

index b5b2a4745328..bff3f3bc7827 100644
@@ -1841,18 +1841,16 @@ static int __get_segment_type_6(struct page *page, enum page_type p_type)
 	if (p_type == DATA) {
 		struct inode *inode = page->mapping->host;
 
-		if (S_ISDIR(inode->i_mode))
-			return CURSEG_HOT_DATA;
-		else if (is_cold_data(page) || file_is_cold(inode))
+		if (is_cold_data(page) || file_is_cold(inode))
 			return CURSEG_COLD_DATA;
-		else
-			return CURSEG_WARM_DATA;
+		if (is_inode_flag_set(inode, FI_HOT_DATA))
+			return CURSEG_HOT_DATA;
+		return CURSEG_WARM_DATA;
 	} else {
 		if (IS_DNODE(page))
 			return is_cold_node(page) ? CURSEG_WARM_NODE :
 						CURSEG_HOT_NODE;
-		else
-			return CURSEG_COLD_NODE;
+		return CURSEG_COLD_NODE;
 	}
 }
 

@@ -2959,6 +2957,7 @@ int build_segment_manager(struct f2fs_sb_info *sbi)
 		sm_info->ipu_policy = 1 << F2FS_IPU_FSYNC;
 	sm_info->min_ipu_util = DEF_MIN_IPU_UTIL;
 	sm_info->min_fsync_blocks = DEF_MIN_FSYNC_BLOCKS;
+	sm_info->min_hot_blocks = DEF_MIN_HOT_BLOCKS;
 
 	sm_info->trim_sections = DEF_BATCHED_TRIM_SECTIONS;
 

fs/f2fs/segment.h +1 -0

index 31846b0fcb95..57e36c1ce7bd 100644
@@ -540,6 +540,7 @@ static inline int utilization(struct f2fs_sb_info *sbi)
  */
 #define DEF_MIN_IPU_UTIL	70
 #define DEF_MIN_FSYNC_BLOCKS	8
+#define DEF_MIN_HOT_BLOCKS	16
 
 enum {
 	F2FS_IPU_FORCE,

fs/f2fs/super.c +2 -0

index b4c5c6298698..2d78f3c76d18 100644
@@ -296,6 +296,7 @@ F2FS_RW_ATTR(SM_INFO, f2fs_sm_info, batched_trim_sections, trim_sections);
 F2FS_RW_ATTR(SM_INFO, f2fs_sm_info, ipu_policy, ipu_policy);
 F2FS_RW_ATTR(SM_INFO, f2fs_sm_info, min_ipu_util, min_ipu_util);
 F2FS_RW_ATTR(SM_INFO, f2fs_sm_info, min_fsync_blocks, min_fsync_blocks);
+F2FS_RW_ATTR(SM_INFO, f2fs_sm_info, min_hot_blocks, min_hot_blocks);
 F2FS_RW_ATTR(NM_INFO, f2fs_nm_info, ram_thresh, ram_thresh);
 F2FS_RW_ATTR(NM_INFO, f2fs_nm_info, ra_nid_pages, ra_nid_pages);
 F2FS_RW_ATTR(NM_INFO, f2fs_nm_info, dirty_nats_ratio, dirty_nats_ratio);

@@ -321,6 +322,7 @@ static struct attribute *f2fs_attrs[] = {
 	ATTR_LIST(ipu_policy),
 	ATTR_LIST(min_ipu_util),
 	ATTR_LIST(min_fsync_blocks),
+	ATTR_LIST(min_hot_blocks),
 	ATTR_LIST(max_victim_search),
 	ATTR_LIST(dir_level),
 	ATTR_LIST(ram_thresh),