Skip to contentSkip to conversation

f2fs: use rwlock instead of rwsem for journal

PATCH In review Jaegeuk Kim 3 participants 8 messages Jul 22, 2021 – Jul 23, 2021 F2FS file system

lore
Share & tools
lore

Jaegeuk Kim

author maintainer
#0 lore raw

PATCH 1/1 f2fs: use rwlock instead of rwsem for journal

This tries to fix priority inversion in the below condition resulting in long checkpoint delay.

f2fs_get_node_info()
- nat_tree_lock
-> sleep in journal_rwsem
checkpoint
- waiting for nat_tree_lock

---
 fs/f2fs/node.c    | 16 ++++++++--------
 fs/f2fs/segment.c | 22 +++++++++++-----------
 fs/f2fs/segment.h |  2 +-
 3 files changed, 20 insertions(+), 20 deletions(-)
fs/f2fs/node.c +8 -8
index 0be9e2d7120e..821a40e02fb4 100644
@@ -567,13 +567,13 @@ int f2fs_get_node_info(struct f2fs_sb_info *sbi, nid_t nid,
 	memset(&ne, 0, sizeof(struct f2fs_nat_entry));
 
 	/* Check current segment summary */
-	down_read(&curseg->journal_rwsem);
+	read_lock(&curseg->journal_rwlock);
 	i = f2fs_lookup_journal_in_cursum(journal, NAT_JOURNAL, nid, 0);
 	if (i >= 0) {
 		ne = nat_in_journal(journal, i);
 		node_info_from_raw_nat(ni, &ne);
 	}
-	up_read(&curseg->journal_rwsem);
+	read_unlock(&curseg->journal_rwlock);
 	if (i >= 0) {
 		up_read(&nm_i->nat_tree_lock);
 		goto cache;
@@ -2338,7 +2338,7 @@ static void scan_curseg_cache(struct f2fs_sb_info *sbi)
 	struct f2fs_journal *journal = curseg->journal;
 	int i;
 
-	down_read(&curseg->journal_rwsem);
+	read_lock(&curseg->journal_rwlock);
 	for (i = 0; i < nats_in_cursum(journal); i++) {
 		block_t addr;
 		nid_t nid;
@@ -2350,7 +2350,7 @@ static void scan_curseg_cache(struct f2fs_sb_info *sbi)
 		else
 			remove_free_nid(sbi, nid);
 	}
-	up_read(&curseg->journal_rwsem);
+	read_unlock(&curseg->journal_rwlock);
 }
 
 static void scan_free_nid_bits(struct f2fs_sb_info *sbi)
@@ -2799,7 +2799,7 @@ static void remove_nats_in_journal(struct f2fs_sb_info *sbi)
 	struct f2fs_journal *journal = curseg->journal;
 	int i;
 
-	down_write(&curseg->journal_rwsem);
+	write_lock(&curseg->journal_rwlock);
 	for (i = 0; i < nats_in_cursum(journal); i++) {
 		struct nat_entry *ne;
 		struct f2fs_nat_entry raw_ne;
@@ -2831,7 +2831,7 @@ static void remove_nats_in_journal(struct f2fs_sb_info *sbi)
 		__set_nat_cache_dirty(nm_i, ne);
 	}
 	update_nats_in_cursum(journal, -i);
-	up_write(&curseg->journal_rwsem);
+	write_unlock(&curseg->journal_rwlock);
 }
 
 static void __adjust_nat_entry_set(struct nat_entry_set *nes,
@@ -2906,7 +2906,7 @@ static int __flush_nat_entry_set(struct f2fs_sb_info *sbi,
 		to_journal = false;
 
 	if (to_journal) {
-		down_write(&curseg->journal_rwsem);
+		write_lock(&curseg->journal_rwlock);
 	} else {
 		page = get_next_nat_page(sbi, start_nid);
 		if (IS_ERR(page))
@@ -2946,7 +2946,7 @@ static int __flush_nat_entry_set(struct f2fs_sb_info *sbi,
 	}
 
 	if (to_journal) {
-		up_write(&curseg->journal_rwsem);
+		write_unlock(&curseg->journal_rwlock);
 	} else {
 		__update_nat_bits(sbi, start_nid, page);
 		f2fs_put_page(page, 1);
fs/f2fs/segment.c +11 -11
index f9b7fb785e1d..c397c1c271ec 100644
@@ -2442,9 +2442,9 @@ static void write_current_sum_page(struct f2fs_sb_info *sbi,
 
 	mutex_lock(&curseg->curseg_mutex);
 
-	down_read(&curseg->journal_rwsem);
+	read_lock(&curseg->journal_rwlock);
 	memcpy(&dst->journal, curseg->journal, SUM_JOURNAL_SIZE);
-	up_read(&curseg->journal_rwsem);
+	read_unlock(&curseg->journal_rwlock);
 
 	memcpy(dst->entries, src->entries, SUM_ENTRY_SIZE);
 	memcpy(&dst->footer, &src->footer, SUM_FOOTER_SIZE);
@@ -3876,9 +3876,9 @@ static int read_normal_summaries(struct f2fs_sb_info *sbi, int type)
 	mutex_lock(&curseg->curseg_mutex);
 
 	/* update journal info */
-	down_write(&curseg->journal_rwsem);
+	write_lock(&curseg->journal_rwlock);
 	memcpy(curseg->journal, &sum->journal, SUM_JOURNAL_SIZE);
-	up_write(&curseg->journal_rwsem);
+	write_unlock(&curseg->journal_rwlock);
 
 	memcpy(curseg->sum_blk->entries, sum->entries, SUM_ENTRY_SIZE);
 	memcpy(&curseg->sum_blk->footer, &sum->footer, SUM_FOOTER_SIZE);
@@ -4136,7 +4136,7 @@ static void remove_sits_in_journal(struct f2fs_sb_info *sbi)
 	struct f2fs_journal *journal = curseg->journal;
 	int i;
 
-	down_write(&curseg->journal_rwsem);
+	write_lock(&curseg->journal_rwlock);
 	for (i = 0; i < sits_in_cursum(journal); i++) {
 		unsigned int segno;
 		bool dirtied;
@@ -4148,7 +4148,7 @@ static void remove_sits_in_journal(struct f2fs_sb_info *sbi)
 			add_sit_entry(segno, &SM_I(sbi)->sit_entry_set);
 	}
 	update_sits_in_cursum(journal, -i);
-	up_write(&curseg->journal_rwsem);
+	write_unlock(&curseg->journal_rwlock);
 }
 
 /*
@@ -4204,7 +4204,7 @@ void f2fs_flush_sit_entries(struct f2fs_sb_info *sbi, struct cp_control *cpc)
 			to_journal = false;
 
 		if (to_journal) {
-			down_write(&curseg->journal_rwsem);
+			write_lock(&curseg->journal_rwlock);
 		} else {
 			page = get_next_sit_page(sbi, start_segno);
 			raw_sit = page_address(page);
@@ -4251,7 +4251,7 @@ void f2fs_flush_sit_entries(struct f2fs_sb_info *sbi, struct cp_control *cpc)
 		}
 
 		if (to_journal)
-			up_write(&curseg->journal_rwsem);
+			write_unlock(&curseg->journal_rwlock);
 		else
 			f2fs_put_page(page, 1);
 
@@ -4432,7 +4432,7 @@ static int build_curseg(struct f2fs_sb_info *sbi)
 		array[i].sum_blk = f2fs_kzalloc(sbi, PAGE_SIZE, GFP_KERNEL);
 		if (!array[i].sum_blk)
 			return -ENOMEM;
-		init_rwsem(&array[i].journal_rwsem);
+		rwlock_init(&array[i].journal_rwlock);
 		array[i].journal = f2fs_kzalloc(sbi,
 				sizeof(struct f2fs_journal), GFP_KERNEL);
 		if (!array[i].journal)
@@ -4509,7 +4509,7 @@ static int build_sit_entries(struct f2fs_sb_info *sbi)
 		start_blk += readed;
 	} while (start_blk < sit_blk_cnt);
 
-	down_read(&curseg->journal_rwsem);
+	read_lock(&curseg->journal_rwlock);
 	for (i = 0; i < sits_in_cursum(journal); i++) {
 		unsigned int old_valid_blocks;
 
@@ -4551,7 +4551,7 @@ static int build_sit_entries(struct f2fs_sb_info *sbi)
 							old_valid_blocks;
 		}
 	}
-	up_read(&curseg->journal_rwsem);
+	read_unlock(&curseg->journal_rwlock);
 
 	if (!err && total_node_blocks != valid_node_count(sbi)) {
 		f2fs_err(sbi, "SIT is corrupted node# %u vs %u",
fs/f2fs/segment.h +1 -1
index 050230c70a53..afb9041dc2e6 100644
@@ -306,7 +306,7 @@ struct victim_selection {
 struct curseg_info {
 	struct mutex curseg_mutex;		/* lock for consistency */
 	struct f2fs_summary_block *sum_blk;	/* cached summary block */
-	struct rw_semaphore journal_rwsem;	/* protect journal area */
+	rwlock_t journal_rwlock;		/* protect journal area */
 	struct f2fs_journal *journal;		/* cached journal info */
 	unsigned char alloc_type;		/* current allocation type */
 	unsigned short seg_type;		/* segment type like CURSEG_XXX_TYPE */
Signature · 1 line
2.32.0.402.g57bb445576-goog
#1 lore raw
Quoted 3 lines · On Wed, 21 Jul 2021 18:41:49 -0700 Jaegeuk Kim wrote:

This tries to fix priority inversion in the below condition resulting in long checkpoint delay.

If you are right and the patch makes real sense, feel free to specify a bit more on the priority inversion you are trying to fix, and how the fix helps shorten the checkpoint delay. And test data is welcome.

Quoted 6 lines

f2fs_get_node_info()
- nat_tree_lock
-> sleep in journal_rwsem
checkpoint
- waiting for nat_tree_lock

_______________________________________________
Linux-f2fs-devel mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/linux-f2fs-devel

maintainer
#2 lore raw
In reply to Jaegeuk Kim
Quoted 21 lines · Jaegeuk Kim

This tries to fix priority inversion in the below condition resulting in long checkpoint delay.

f2fs_get_node_info()
- nat_tree_lock
-> sleep in journal_rwsem
checkpoint
- waiting for nat_tree_lock

---
  fs/f2fs/node.c    | 16 ++++++++--------
  fs/f2fs/segment.c | 22 +++++++++++-----------
  fs/f2fs/segment.h |  2 +-
  3 files changed, 20 insertions(+), 20 deletions(-)
fs/f2fs/node.c +0 -0
index 0be9e2d7120e..821a40e02fb4 100644

fs/f2fs/node.c int f2fs_get_node_info(struct f2fs_sb_info *sbi, nid_t nid, · line 570

Quoted 5 lines
fs/f2fs/node.c +1 -1
@@ -567,13 +567,13 @@ int f2fs_get_node_info(struct f2fs_sb_info *sbi, nid_t nid,
  	memset(&ne, 0, sizeof(struct f2fs_nat_entry));
 
  	/* Check current segment summary */
-	down_read(&curseg->journal_rwsem);
+	read_lock(&curseg->journal_rwlock);

It seem journal_rwsem covers too many operations, e.g.

- scan_curseg_cache
- add_free_nid
- f2fs_kmem_cache_alloc
- kmem_cache_alloc(__GFP_NOFAIL)

Thanks,

Quoted 161 lines
fs/f2fs/node.c +7 -7
  	i = f2fs_lookup_journal_in_cursum(journal, NAT_JOURNAL, nid, 0);
  	if (i >= 0) {
  		ne = nat_in_journal(journal, i);
  		node_info_from_raw_nat(ni, &ne);
  	}
-	up_read(&curseg->journal_rwsem);
+	read_unlock(&curseg->journal_rwlock);
  	if (i >= 0) {
  		up_read(&nm_i->nat_tree_lock);
  		goto cache;
@@ -2338,7 +2338,7 @@ static void scan_curseg_cache(struct f2fs_sb_info *sbi)
  	struct f2fs_journal *journal = curseg->journal;
  	int i;
 
-	down_read(&curseg->journal_rwsem);
+	read_lock(&curseg->journal_rwlock);
  	for (i = 0; i < nats_in_cursum(journal); i++) {
  		block_t addr;
  		nid_t nid;
@@ -2350,7 +2350,7 @@ static void scan_curseg_cache(struct f2fs_sb_info *sbi)
  		else
  			remove_free_nid(sbi, nid);
  	}
-	up_read(&curseg->journal_rwsem);
+	read_unlock(&curseg->journal_rwlock);
  }
 
  static void scan_free_nid_bits(struct f2fs_sb_info *sbi)
@@ -2799,7 +2799,7 @@ static void remove_nats_in_journal(struct f2fs_sb_info *sbi)
  	struct f2fs_journal *journal = curseg->journal;
  	int i;
 
-	down_write(&curseg->journal_rwsem);
+	write_lock(&curseg->journal_rwlock);
  	for (i = 0; i < nats_in_cursum(journal); i++) {
  		struct nat_entry *ne;
  		struct f2fs_nat_entry raw_ne;
@@ -2831,7 +2831,7 @@ static void remove_nats_in_journal(struct f2fs_sb_info *sbi)
  		__set_nat_cache_dirty(nm_i, ne);
  	}
  	update_nats_in_cursum(journal, -i);
-	up_write(&curseg->journal_rwsem);
+	write_unlock(&curseg->journal_rwlock);
  }
 
  static void __adjust_nat_entry_set(struct nat_entry_set *nes,
@@ -2906,7 +2906,7 @@ static int __flush_nat_entry_set(struct f2fs_sb_info *sbi,
  		to_journal = false;
 
  	if (to_journal) {
-		down_write(&curseg->journal_rwsem);
+		write_lock(&curseg->journal_rwlock);
  	} else {
  		page = get_next_nat_page(sbi, start_nid);
  		if (IS_ERR(page))
@@ -2946,7 +2946,7 @@ static int __flush_nat_entry_set(struct f2fs_sb_info *sbi,
  	}
 
  	if (to_journal) {
-		up_write(&curseg->journal_rwsem);
+		write_unlock(&curseg->journal_rwlock);
  	} else {
  		__update_nat_bits(sbi, start_nid, page);
  		f2fs_put_page(page, 1);
fs/f2fs/segment.c +11 -11
index f9b7fb785e1d..c397c1c271ec 100644
@@ -2442,9 +2442,9 @@ static void write_current_sum_page(struct f2fs_sb_info *sbi,
 
  	mutex_lock(&curseg->curseg_mutex);
 
-	down_read(&curseg->journal_rwsem);
+	read_lock(&curseg->journal_rwlock);
  	memcpy(&dst->journal, curseg->journal, SUM_JOURNAL_SIZE);
-	up_read(&curseg->journal_rwsem);
+	read_unlock(&curseg->journal_rwlock);
 
  	memcpy(dst->entries, src->entries, SUM_ENTRY_SIZE);
  	memcpy(&dst->footer, &src->footer, SUM_FOOTER_SIZE);
@@ -3876,9 +3876,9 @@ static int read_normal_summaries(struct f2fs_sb_info *sbi, int type)
  	mutex_lock(&curseg->curseg_mutex);
 
  	/* update journal info */
-	down_write(&curseg->journal_rwsem);
+	write_lock(&curseg->journal_rwlock);
  	memcpy(curseg->journal, &sum->journal, SUM_JOURNAL_SIZE);
-	up_write(&curseg->journal_rwsem);
+	write_unlock(&curseg->journal_rwlock);
 
  	memcpy(curseg->sum_blk->entries, sum->entries, SUM_ENTRY_SIZE);
  	memcpy(&curseg->sum_blk->footer, &sum->footer, SUM_FOOTER_SIZE);
@@ -4136,7 +4136,7 @@ static void remove_sits_in_journal(struct f2fs_sb_info *sbi)
  	struct f2fs_journal *journal = curseg->journal;
  	int i;
 
-	down_write(&curseg->journal_rwsem);
+	write_lock(&curseg->journal_rwlock);
  	for (i = 0; i < sits_in_cursum(journal); i++) {
  		unsigned int segno;
  		bool dirtied;
@@ -4148,7 +4148,7 @@ static void remove_sits_in_journal(struct f2fs_sb_info *sbi)
  			add_sit_entry(segno, &SM_I(sbi)->sit_entry_set);
  	}
  	update_sits_in_cursum(journal, -i);
-	up_write(&curseg->journal_rwsem);
+	write_unlock(&curseg->journal_rwlock);
  }
 
  /*
@@ -4204,7 +4204,7 @@ void f2fs_flush_sit_entries(struct f2fs_sb_info *sbi, struct cp_control *cpc)
  			to_journal = false;
 
  		if (to_journal) {
-			down_write(&curseg->journal_rwsem);
+			write_lock(&curseg->journal_rwlock);
  		} else {
  			page = get_next_sit_page(sbi, start_segno);
  			raw_sit = page_address(page);
@@ -4251,7 +4251,7 @@ void f2fs_flush_sit_entries(struct f2fs_sb_info *sbi, struct cp_control *cpc)
  		}
 
  		if (to_journal)
-			up_write(&curseg->journal_rwsem);
+			write_unlock(&curseg->journal_rwlock);
  		else
  			f2fs_put_page(page, 1);
 
@@ -4432,7 +4432,7 @@ static int build_curseg(struct f2fs_sb_info *sbi)
  		array[i].sum_blk = f2fs_kzalloc(sbi, PAGE_SIZE, GFP_KERNEL);
  		if (!array[i].sum_blk)
  			return -ENOMEM;
-		init_rwsem(&array[i].journal_rwsem);
+		rwlock_init(&array[i].journal_rwlock);
  		array[i].journal = f2fs_kzalloc(sbi,
  				sizeof(struct f2fs_journal), GFP_KERNEL);
  		if (!array[i].journal)
@@ -4509,7 +4509,7 @@ static int build_sit_entries(struct f2fs_sb_info *sbi)
  		start_blk += readed;
  	} while (start_blk < sit_blk_cnt);
 
-	down_read(&curseg->journal_rwsem);
+	read_lock(&curseg->journal_rwlock);
  	for (i = 0; i < sits_in_cursum(journal); i++) {
  		unsigned int old_valid_blocks;
 
@@ -4551,7 +4551,7 @@ static int build_sit_entries(struct f2fs_sb_info *sbi)
  							old_valid_blocks;
  		}
  	}
-	up_read(&curseg->journal_rwsem);
+	read_unlock(&curseg->journal_rwlock);
 
  	if (!err && total_node_blocks != valid_node_count(sbi)) {
  		f2fs_err(sbi, "SIT is corrupted node# %u vs %u",
fs/f2fs/segment.h +0 -0
index 050230c70a53..afb9041dc2e6 100644

fs/f2fs/segment.h struct victim_selection { · line 312

Quoted 8 lines
fs/f2fs/segment.h +1 -1
@@ -306,7 +306,7 @@ struct victim_selection {
  struct curseg_info {
  	struct mutex curseg_mutex;		/* lock for consistency */
  	struct f2fs_summary_block *sum_blk;	/* cached summary block */
-	struct rw_semaphore journal_rwsem;	/* protect journal area */
+	rwlock_t journal_rwlock;		/* protect journal area */
  	struct f2fs_journal *journal;		/* cached journal info */
  	unsigned char alloc_type;		/* current allocation type */
  	unsigned short seg_type;		/* segment type like CURSEG_XXX_TYPE */

_______________________________________________
Linux-f2fs-devel mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/linux-f2fs-devel

author maintainer
#3 lore raw

Re: [PATCH v2] f2fs: don't sleep while grabing nat_tree_lock

In reply to Jaegeuk Kim

This tries to fix priority inversion in the below condition resulting in long checkpoint delay.

f2fs_get_node_info()
- nat_tree_lock
-> sleep to grab journal_rwsem by contention

checkpoint
- waiting for nat_tree_lock

In order to let checkpoint go, let's release nat_tree_lock, if there's a journal_rwsem contention.

---
Change log from v1:
- drop rwlock but take retry logic to reduce lock-holding time

 fs/f2fs/node.c | 14 ++++++++++----
 1 file changed, 10 insertions(+), 4 deletions(-)
fs/f2fs/node.c +10 -4
index 0be9e2d7120e..c60ba4179bb2 100644
@@ -552,7 +552,7 @@ int f2fs_get_node_info(struct f2fs_sb_info *sbi, nid_t nid,
 	int i;
 
 	ni->nid = nid;
-
+retry:
 	/* Check nat cache */
 	down_read(&nm_i->nat_tree_lock);
 	e = __lookup_nat_cache(nm_i, nid);
@@ -564,10 +564,16 @@ int f2fs_get_node_info(struct f2fs_sb_info *sbi, nid_t nid,
 		return 0;
 	}
 
-	memset(&ne, 0, sizeof(struct f2fs_nat_entry));
+	/*
+	 * Check current segment summary by trying to grab journal_rwsem first.
+	 * This sem is on the critical path on the checkpoint requiring the above
+	 * nat_tree_lock. Therefore, we should retry, if we failed to grab here.
+	 */
+	if (!down_read_trylock(&curseg->journal_rwsem)) {
+		up_read(&nm_i->nat_tree_lock);
+		goto retry;
+	}
 
-	/* Check current segment summary */
-	down_read(&curseg->journal_rwsem);
 	i = f2fs_lookup_journal_in_cursum(journal, NAT_JOURNAL, nid, 0);
 	if (i >= 0) {
 		ne = nat_in_journal(journal, i);
Signature · 1 line
2.32.0.432.gabb21c7263-goog
maintainer
#4 lore raw

Re: [f2fs-dev] [PATCH v2] f2fs: don't sleep while grabing nat_tree_lock

Quoted 7 lines · Jaegeuk Kim

This tries to fix priority inversion in the below condition resulting in long checkpoint delay.

f2fs_get_node_info()
- nat_tree_lock
-> sleep to grab journal_rwsem by contention

Quoted 5 lines

checkpoint
- waiting for nat_tree_lock

In order to let checkpoint go, let's release nat_tree_lock, if there's a journal_rwsem contention.

Write lock of nat_tree_lock is held from many places, how about just retrying unlock/lock only if checkpoint() is flushing nat blocks?

---
 fs/f2fs/f2fs.h | 1 +
 fs/f2fs/node.c | 8 ++++++--
 2 files changed, 7 insertions(+), 2 deletions(-)
fs/f2fs/f2fs.h +1 -0
index ad8f99d7235f..05f41a15fda4 100644
@@ -877,6 +877,7 @@ struct f2fs_nm_info {
 	spinlock_t nat_list_lock;	/* protect clean nat entry list */
 	unsigned int nat_cnt[MAX_NAT_STATE]; /* the # of cached nat entries */
 	unsigned int nat_blocks;	/* # of nat blocks */
+	bool flushing_nat;		/* indicate checkpoint() is flushing nat blocks */
 
 	/* free node ids management */
 	struct radix_tree_root free_nid_root;/* root of the free_nid cache */
fs/f2fs/node.c +6 -2
index c60ba4179bb2..2caa171a68f8 100644
@@ -569,7 +569,7 @@ int f2fs_get_node_info(struct f2fs_sb_info *sbi, nid_t nid,
 	 * This sem is on the critical path on the checkpoint requiring the above
 	 * nat_tree_lock. Therefore, we should retry, if we failed to grab here.
 	 */
-	if (!down_read_trylock(&curseg->journal_rwsem)) {
+	if (!down_read_trylock(&curseg->journal_rwsem) && nm_i->flushing_nat) {
 		up_read(&nm_i->nat_tree_lock);
 		goto retry;
 	}
@@ -2981,6 +2981,8 @@ int f2fs_flush_nat_entries(struct f2fs_sb_info *sbi, struct cp_control *cpc)
 	LIST_HEAD(sets);
 	int err = 0;
 
+	nm_i->flushing_nat = true;
+
 	/*
 	 * during unmount, let's flush nat_bits before checking
 	 * nat_cnt[DIRTY_NAT].
@@ -2992,7 +2994,7 @@ int f2fs_flush_nat_entries(struct f2fs_sb_info *sbi, struct cp_control *cpc)
 	}
 
 	if (!nm_i->nat_cnt[DIRTY_NAT])
-		return 0;
+		goto out;
 
 	down_write(&nm_i->nat_tree_lock);
 
@@ -3026,6 +3028,8 @@ int f2fs_flush_nat_entries(struct f2fs_sb_info *sbi, struct cp_control *cpc)
 	up_write(&nm_i->nat_tree_lock);
 	/* Allow dirty nats by node block allocation in write_begin */
 
+out:
+	nm_i->flushing_nat = false;
 	return err;
 }
 
Signature · 1 line
2.22.1
author maintainer
#5 lore raw

Re: [PATCH v3] f2fs: don't sleep while grabing nat_tree_lock

In reply to Jaegeuk Kim

This tries to fix priority inversion in the below condition resulting in long checkpoint delay.

f2fs_get_node_info()
- nat_tree_lock
-> sleep to grab journal_rwsem by contention

checkpoint
- waiting for nat_tree_lock

In order to let checkpoint go, let's release nat_tree_lock, if there's a journal_rwsem contention.

---
Change log from v2:
- don't bother checkpoint

 fs/f2fs/node.c | 17 +++++++++++++----
 1 file changed, 13 insertions(+), 4 deletions(-)
fs/f2fs/node.c +13 -4
index 0be9e2d7120e..b26642daa3d2 100644
@@ -552,7 +552,7 @@ int f2fs_get_node_info(struct f2fs_sb_info *sbi, nid_t nid,
 	int i;
 
 	ni->nid = nid;
-
+retry:
 	/* Check nat cache */
 	down_read(&nm_i->nat_tree_lock);
 	e = __lookup_nat_cache(nm_i, nid);
@@ -564,10 +564,19 @@ int f2fs_get_node_info(struct f2fs_sb_info *sbi, nid_t nid,
 		return 0;
 	}
 
-	memset(&ne, 0, sizeof(struct f2fs_nat_entry));
+	/*
+	 * Check current segment summary by trying to grab journal_rwsem first.
+	 * This sem is on the critical path on the checkpoint requiring the above
+	 * nat_tree_lock. Therefore, we should retry, if we failed to grab here
+	 * while not bothering checkpoint.
+	 */
+	if (rwsem_is_locked(&sbi->cp_global_sem)) {
+		down_read(&curseg->journal_rwsem);
+	} else if (!down_read_trylock(&curseg->journal_rwsem)) {
+		up_read(&nm_i->nat_tree_lock);
+		goto retry;
+	}
 
-	/* Check current segment summary */
-	down_read(&curseg->journal_rwsem);
 	i = f2fs_lookup_journal_in_cursum(journal, NAT_JOURNAL, nid, 0);
 	if (i >= 0) {
 		ne = nat_in_journal(journal, i);
Signature · 1 line
2.32.0.432.gabb21c7263-goog
maintainer
#6 lore raw

Re: [f2fs-dev] [PATCH v3] f2fs: don't sleep while grabing nat_tree_lock

Quoted 37 lines · Jaegeuk Kim

This tries to fix priority inversion in the below condition resulting in long checkpoint delay.

f2fs_get_node_info()
- nat_tree_lock
-> sleep to grab journal_rwsem by contention

checkpoint
- waiting for nat_tree_lock

In order to let checkpoint go, let's release nat_tree_lock, if there's a journal_rwsem contention.

---
Change log from v2:
- don't bother checkpoint

  fs/f2fs/node.c | 17 +++++++++++++----
  1 file changed, 13 insertions(+), 4 deletions(-)
fs/f2fs/node.c +1 -1
index 0be9e2d7120e..b26642daa3d2 100644
@@ -552,7 +552,7 @@ int f2fs_get_node_info(struct f2fs_sb_info *sbi, nid_t nid,
  	int i;
 
  	ni->nid = nid;
-
+retry:
  	/* Check nat cache */
  	down_read(&nm_i->nat_tree_lock);
  	e = __lookup_nat_cache(nm_i, nid);
@@ -564,10 +564,19 @@ int f2fs_get_node_info(struct f2fs_sb_info *sbi, nid_t nid,
  		return 0;

fs/f2fs/node.c int f2fs_get_node_info(struct f2fs_sb_info *sbi, nid_t nid, · line 573

Quoted 10 lines
fs/f2fs/node.c +7 -1
  	}
 
-	memset(&ne, 0, sizeof(struct f2fs_nat_entry));
+	/*
+	 * Check current segment summary by trying to grab journal_rwsem first.
+	 * This sem is on the critical path on the checkpoint requiring the above
+	 * nat_tree_lock. Therefore, we should retry, if we failed to grab here
+	 * while not bothering checkpoint.
+	 */
+	if (rwsem_is_locked(&sbi->cp_global_sem)) {

You mean: if (!rwsem_is_locked()) ?

IMO, once CP is processing, all readers who needs read nat will be blocked...

Thanks,

Quoted 12 lines
fs/f2fs/node.c +5 -2
+		down_read(&curseg->journal_rwsem);
+	} else if (!down_read_trylock(&curseg->journal_rwsem)) {
+		up_read(&nm_i->nat_tree_lock);
+		goto retry;
+	}
 
-	/* Check current segment summary */
-	down_read(&curseg->journal_rwsem);
  	i = f2fs_lookup_journal_in_cursum(journal, NAT_JOURNAL, nid, 0);
  	if (i >= 0) {
  		ne = nat_in_journal(journal, i);
author maintainer
#7 lore raw

Re: [f2fs-dev] [PATCH v4] f2fs: don't sleep while grabing nat_tree_lock

In reply to Jaegeuk Kim

This tries to fix priority inversion in the below condition resulting in long checkpoint delay.

f2fs_get_node_info()
- nat_tree_lock
-> sleep to grab journal_rwsem by contention

checkpoint
- waiting for nat_tree_lock

In order to let checkpoint go, let's release nat_tree_lock, if there's a journal_rwsem contention.

---

Change log from v3:
- fix wrong condition check

 fs/f2fs/node.c | 17 +++++++++++++----
 1 file changed, 13 insertions(+), 4 deletions(-)
fs/f2fs/node.c +13 -4
index 0be9e2d7120e..c945a9730f3c 100644
@@ -552,7 +552,7 @@ int f2fs_get_node_info(struct f2fs_sb_info *sbi, nid_t nid,
 	int i;
 
 	ni->nid = nid;
-
+retry:
 	/* Check nat cache */
 	down_read(&nm_i->nat_tree_lock);
 	e = __lookup_nat_cache(nm_i, nid);
@@ -564,10 +564,19 @@ int f2fs_get_node_info(struct f2fs_sb_info *sbi, nid_t nid,
 		return 0;
 	}
 
-	memset(&ne, 0, sizeof(struct f2fs_nat_entry));
+	/*
+	 * Check current segment summary by trying to grab journal_rwsem first.
+	 * This sem is on the critical path on the checkpoint requiring the above
+	 * nat_tree_lock. Therefore, we should retry, if we failed to grab here
+	 * while not bothering checkpoint.
+	 */
+	if (!rwsem_is_locked(&sbi->cp_global_sem)) {
+		down_read(&curseg->journal_rwsem);
+	} else if (!down_read_trylock(&curseg->journal_rwsem)) {
+		up_read(&nm_i->nat_tree_lock);
+		goto retry;
+	}
 
-	/* Check current segment summary */
-	down_read(&curseg->journal_rwsem);
 	i = f2fs_lookup_journal_in_cursum(journal, NAT_JOURNAL, nid, 0);
 	if (i >= 0) {
 		ne = nat_in_journal(journal, i);
Signature · 1 line
2.32.0.432.gabb21c7263-goog