Recently, we encounter a deadlock case where there are many tasks
hung on agi and agf xfs_buf. It end up with a deadlock between
agf and inode xfs_buf as following,
xfs_ifree() xfs_bmap_btalloc()
xfs_iunlink_remove() xfs_alloc_vextent()
hold inode bp hold agf
xfs_difree() xfs_trans_log_inode()
require agf require inode bp
The task requires inode bp with agf held block other tasks which
want to require agf with agi held. Then the whole filesystem
looks like a agi/agf deadlock. To fix this issue, get agf in
xfs_iunlink_remove() after it get agi and before get indoe bp.
Signed-off-by: Wang Jianchao (Kuaishou) <jianchao.wan9@xxxxxxxxx>
---
fs/xfs/xfs_inode.c | 10 ++++++++++
1 file changed, 10 insertions(+)
diff --git a/fs/xfs/xfs_inode.c b/fs/xfs/xfs_inode.c
index 9de6205fe134..c0a29fd00b0e 100644
--- a/fs/xfs/xfs_inode.c
+++ b/fs/xfs/xfs_inode.c
@@ -36,6 +36,7 @@
#include "xfs_reflink.h"
#include "xfs_ag.h"
#include "xfs_log_priv.h"
+#include "xfs_alloc.h"
struct kmem_cache *xfs_inode_cache;
@@ -2337,6 +2338,7 @@ xfs_iunlink_remove(
{
struct xfs_mount *mp = tp->t_mountp;
struct xfs_agi *agi;
+ struct xfs_buf *agfbp;
struct xfs_buf *agibp;
struct xfs_buf *last_ibp;
struct xfs_dinode *last_dip = NULL;
@@ -2352,6 +2354,14 @@ xfs_iunlink_remove(
error = xfs_read_agi(mp, tp, pag->pag_agno, &agibp);
if (error)
return error;
+
+ /*
+ * Get the agf buffer first to ensure the lock ordering against inode bp
+ */
+ error = xfs_read_agf(mp, tp, pag->pag_agno, 0, &agfbp);
+ if (error)
+ return error;
+
agi = agibp->b_addr;
/*
--
2.17.1
