From 19c7377b56a618c2a6ae1dfef1c22f19ad9d427c Mon Sep 17 00:00:00 2001
From: Chao Yu <chao2.yu@samsung.com>
Date: Fri, 26 Feb 2016 09:33:04 +0800
Subject: f2fs: fix to avoid deadlock when merging inline data

When testing with fsstress, kworker and user threads were both blocked:

INFO: task kworker/u16:1:16580 blocked for more than 120 seconds.
"echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
kworker/u16:1   D ffff8803f2595390     0 16580      2 0x00000000
Workqueue: writeback bdi_writeback_workfn (flush-251:0)
 ffff8802730e5760 0000000000000046 ffff880274729fc0 0000000000012440
 ffff8802730e5fd8 ffff8802730e4010 0000000000012440 0000000000012440
 ffff8802730e5fd8 0000000000012440 ffff880274729fc0 ffff88026eb50000
Call Trace:
 [<ffffffff816fe9d9>] schedule+0x29/0x70
 [<ffffffff816ff895>] rwsem_down_read_failed+0xa5/0xf9
 [<ffffffff81378584>] call_rwsem_down_read_failed+0x14/0x30
 [<ffffffffa0694feb>] f2fs_write_data_page+0x31b/0x420 [f2fs]
 [<ffffffffa0690f1a>] __f2fs_writepage+0x1a/0x50 [f2fs]
 [<ffffffffa06922a0>] f2fs_write_data_pages+0xe0/0x290 [f2fs]
 [<ffffffff811473b3>] do_writepages+0x23/0x40
 [<ffffffff811cc3ee>] __writeback_single_inode+0x4e/0x250
 [<ffffffff811cd4f1>] writeback_sb_inodes+0x2c1/0x470
 [<ffffffff811cd73e>] __writeback_inodes_wb+0x9e/0xd0
 [<ffffffff811cda0b>] wb_writeback+0x1fb/0x2d0
 [<ffffffff811cdb7c>] wb_do_writeback+0x9c/0x220
 [<ffffffff811ce232>] bdi_writeback_workfn+0x72/0x1c0
 [<ffffffff8106b74e>] process_one_work+0x1de/0x5b0
 [<ffffffff8106e78f>] worker_thread+0x11f/0x3e0
 [<ffffffff810750ce>] kthread+0xde/0xf0
 [<ffffffff817093f8>] ret_from_fork+0x58/0x90

fsstress thread stack:
 [<ffffffff81139f0e>] sleep_on_page+0xe/0x20
 [<ffffffff81139ef7>] __lock_page+0x67/0x70
 [<ffffffff8113b100>] find_lock_page+0x50/0x80
 [<ffffffff8113b24f>] find_or_create_page+0x3f/0xb0
 [<ffffffffa06983a9>] sync_node_pages+0x259/0x810 [f2fs]
 [<ffffffffa068d874>] write_checkpoint+0x1a4/0xce0 [f2fs]
 [<ffffffffa0686b0c>] f2fs_sync_fs+0x7c/0xd0 [f2fs]
 [<ffffffffa067c813>] f2fs_sync_file+0x143/0x5f0 [f2fs]
 [<ffffffff811d301b>] vfs_fsync_range+0x2b/0x40
 [<ffffffff811d304c>] vfs_fsync+0x1c/0x20
 [<ffffffff811d3291>] do_fsync+0x41/0x70
 [<ffffffff811d32d3>] SyS_fdatasync+0x13/0x20
 [<ffffffff817094a2>] system_call_fastpath+0x16/0x1b
 [<ffffffffffffffff>] 0xffffffffffffffff

The reason of this issue is:
CPU0:					CPU1:
 - f2fs_write_data_pages
					 - f2fs_sync_fs
					  - write_checkpoint
					   - block_operations
					    - f2fs_lock_all
					     - down_write(sbi->cp_rwsem)
  - lock_page(page)
  - f2fs_write_data_page
					    - sync_node_pages
					     - flush_inline_data
					      - pagecache_get_page(page, GFP_LOCK)
   - f2fs_lock_op
    - down_read(sbi->cp_rwsem)

This patch alters to use trylock_page in flush_inline_data to fix this ABBA
deadlock issue.

Signed-off-by: Chao Yu <chao2.yu@samsung.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
---
 fs/f2fs/node.c | 9 +++++++--
 1 file changed, 7 insertions(+), 2 deletions(-)

(limited to 'fs/f2fs')

diff --git a/fs/f2fs/node.c b/fs/f2fs/node.c
index 7b613dd16e77..d714b607bf46 100644
--- a/fs/f2fs/node.c
+++ b/fs/f2fs/node.c
@@ -1203,10 +1203,13 @@ static void flush_inline_data(struct f2fs_sb_info *sbi, nid_t ino)
 	if (!inode)
 		return;
 
-	page = pagecache_get_page(inode->i_mapping, 0, FGP_LOCK|FGP_NOWAIT, 0);
+	page = pagecache_get_page(inode->i_mapping, 0, FGP_NOWAIT, 0);
 	if (!page)
 		goto iput_out;
 
+	if (!trylock_page(page))
+		goto release_out;
+
 	if (!PageUptodate(page))
 		goto page_out;
 
@@ -1221,7 +1224,9 @@ static void flush_inline_data(struct f2fs_sb_info *sbi, nid_t ino)
 	else
 		set_page_dirty(page);
 page_out:
-	f2fs_put_page(page, 1);
+	unlock_page(page);
+release_out:
+	f2fs_put_page(page, 0);
 iput_out:
 	iput(inode);
 }
-- 
cgit v1.2.3