f2fs: avoid fi->i_gc_rwsem[WRITE] lock in f2fs_gc

The f2fs_gc() called by f2fs_balance_fs() requires to be called outside of
fi->i_gc_rwsem[WRITE], since f2fs_gc() can try to grab it in a loop.

If it hits the miximum retrials in GC, let's give a chance to release
gc_mutex for a short time in order not to go into live lock in the worst
case.
Reviewed-by: Chao Yu <yuchao0@huawei.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>

fs/f2fs/data.c

@@ -2217,14 +2217,14 @@ static void f2fs_write_failed(struct address_space *mapping, loff_t to)
 	loff_t i_size = i_size_read(inode);
 
 	if (to > i_size) {
-		down_write(&F2FS_I(inode)->i_mmap_sem);
 		down_write(&F2FS_I(inode)->i_gc_rwsem[WRITE]);
+		down_write(&F2FS_I(inode)->i_mmap_sem);
 
 		truncate_pagecache(inode, i_size);
 		f2fs_truncate_blocks(inode, i_size, true);
 
-		up_write(&F2FS_I(inode)->i_gc_rwsem[WRITE]);
 		up_write(&F2FS_I(inode)->i_mmap_sem);
+		up_write(&F2FS_I(inode)->i_gc_rwsem[WRITE]);
 	}
 }
 

fs/f2fs/f2fs.h

@@ -1243,6 +1243,7 @@ struct f2fs_sb_info {
 	unsigned int gc_mode;			/* current GC state */
 	/* for skip statistic */
 	unsigned long long skipped_atomic_files[2];	/* FG_GC and BG_GC */
+	unsigned long long skipped_gc_rwsem;		/* FG_GC only */
 
 	/* threshold for gc trials on pinned files */
 	u64 gc_pin_file_threshold;

fs/f2fs/gc.c

@@ -882,6 +882,7 @@ static void gc_data_segment(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
 			if (!down_write_trylock(
 				&F2FS_I(inode)->i_gc_rwsem[WRITE])) {
 				iput(inode);
+				sbi->skipped_gc_rwsem++;
 				continue;
 			}
 
@@ -911,6 +912,7 @@ static void gc_data_segment(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
 					continue;
 				if (!down_write_trylock(
 						&fi->i_gc_rwsem[WRITE])) {
+					sbi->skipped_gc_rwsem++;
 					up_write(&fi->i_gc_rwsem[READ]);
 					continue;
 				}
@@ -1048,6 +1050,7 @@ int f2fs_gc(struct f2fs_sb_info *sbi, bool sync,
 		.iroot = RADIX_TREE_INIT(gc_list.iroot, GFP_NOFS),
 	};
 	unsigned long long last_skipped = sbi->skipped_atomic_files[FG_GC];
+	unsigned long long first_skipped;
 	unsigned int skipped_round = 0, round = 0;
 
 	trace_f2fs_gc_begin(sbi->sb, sync, background,
@@ -1060,6 +1063,8 @@ int f2fs_gc(struct f2fs_sb_info *sbi, bool sync,
 				prefree_segments(sbi));
 
 	cpc.reason = __get_cp_reason(sbi);
+	sbi->skipped_gc_rwsem = 0;
+	first_skipped = last_skipped;
 gc_more:
 	if (unlikely(!(sbi->sb->s_flags & SB_ACTIVE))) {
 		ret = -EINVAL;
@@ -1101,7 +1106,8 @@ int f2fs_gc(struct f2fs_sb_info *sbi, bool sync,
 	total_freed += seg_freed;
 
 	if (gc_type == FG_GC) {
-		if (sbi->skipped_atomic_files[FG_GC] > last_skipped)
+		if (sbi->skipped_atomic_files[FG_GC] > last_skipped ||
+						sbi->skipped_gc_rwsem)
 			skipped_round++;
 		last_skipped = sbi->skipped_atomic_files[FG_GC];
 		round++;
@@ -1110,15 +1116,23 @@ int f2fs_gc(struct f2fs_sb_info *sbi, bool sync,
 	if (gc_type == FG_GC)
 		sbi->cur_victim_sec = NULL_SEGNO;
 
-	if (!sync) {
+	if (sync)
+		goto stop;
+
 	if (has_not_enough_free_secs(sbi, sec_freed, 0)) {
-			if (skipped_round > MAX_SKIP_ATOMIC_COUNT &&
-				skipped_round * 2 >= round)
-				f2fs_drop_inmem_pages_all(sbi, true);
+		if (skipped_round <= MAX_SKIP_GC_COUNT ||
+					skipped_round * 2 < round) {
 			segno = NULL_SEGNO;
 			goto gc_more;
 		}
 
+		if (first_skipped < last_skipped &&
+				(last_skipped - first_skipped) >
+						sbi->skipped_gc_rwsem) {
+			f2fs_drop_inmem_pages_all(sbi, true);
+			segno = NULL_SEGNO;
+			goto gc_more;
+		}
 		if (gc_type == FG_GC)
 			ret = f2fs_write_checkpoint(sbi, &cpc);
 	}

fs/f2fs/segment.c

@@ -445,8 +445,10 @@ int f2fs_commit_inmem_pages(struct inode *inode)
 	int err;
 
 	f2fs_balance_fs(sbi, true);
-	f2fs_lock_op(sbi);
 
+	down_write(&fi->i_gc_rwsem[WRITE]);
+
+	f2fs_lock_op(sbi);
 	set_inode_flag(inode, FI_ATOMIC_COMMIT);
 
 	mutex_lock(&fi->inmem_lock);
@@ -461,6 +463,8 @@ int f2fs_commit_inmem_pages(struct inode *inode)
 	clear_inode_flag(inode, FI_ATOMIC_COMMIT);
 
 	f2fs_unlock_op(sbi);
+	up_write(&fi->i_gc_rwsem[WRITE]);
+
 	return err;
 }
 

fs/f2fs/segment.h

@@ -215,7 +215,7 @@ struct segment_allocation {
 #define IS_DUMMY_WRITTEN_PAGE(page)			\
 		(page_private(page) == (unsigned long)DUMMY_WRITTEN_PAGE)
 
-#define MAX_SKIP_ATOMIC_COUNT			16
+#define MAX_SKIP_GC_COUNT			16
 
 struct inmem_pages {
 	struct list_head list;