[PATCH] ocfs2: fix orphan recovery deadlock

Orphan dir recovery can deadlock with another process in
ocfs2_delete_inode() in some corner cases. Fix this by tracking recovery
state more closely and allowing it to handle inode wipes which might
deadlock.

Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>

5 files changed, 154 insertions, 32 deletions

diff --git a/fs/ocfs2/heartbeat.c b/fs/ocfs2/heartbeat.c
index 0bbd22f..cbfd45a 100644
--- a/fs/ocfs2/heartbeat.c
+++ b/fs/ocfs2/heartbeat.c
@@ -67,6 +67,7 @@ void ocfs2_init_node_maps(struct ocfs2_super *osb)
 	ocfs2_node_map_init(&osb->mounted_map);
 	ocfs2_node_map_init(&osb->recovery_map);
 	ocfs2_node_map_init(&osb->umount_map);
+	ocfs2_node_map_init(&osb->osb_recovering_orphan_dirs);
 }
 
 static void ocfs2_do_node_down(int node_num,
diff --git a/fs/ocfs2/inode.c b/fs/ocfs2/inode.c
index 8122489..315472a 100644
--- a/fs/ocfs2/inode.c
+++ b/fs/ocfs2/inode.c
@@ -41,6 +41,7 @@
 #include "dlmglue.h"
 #include "extent_map.h"
 #include "file.h"
+#include "heartbeat.h"
 #include "inode.h"
 #include "journal.h"
 #include "namei.h"
@@ -544,6 +545,42 @@ bail:
 	return status;
 }
 
+/* 
+ * Serialize with orphan dir recovery. If the process doing
+ * recovery on this orphan dir does an iget() with the dir
+ * i_mutex held, we'll deadlock here. Instead we detect this
+ * and exit early - recovery will wipe this inode for us.
+ */
+static int ocfs2_check_orphan_recovery_state(struct ocfs2_super *osb,
+					     int slot)
+{
+	int ret = 0;
+
+	spin_lock(&osb->osb_lock);
+	if (ocfs2_node_map_test_bit(osb, &osb->osb_recovering_orphan_dirs, slot)) {
+		mlog(0, "Recovery is happening on orphan dir %d, will skip "
+		     "this inode\n", slot);
+		ret = -EDEADLK;
+		goto out;
+	}
+	/* This signals to the orphan recovery process that it should
+	 * wait for us to handle the wipe. */
+	osb->osb_orphan_wipes[slot]++;
+out:
+	spin_unlock(&osb->osb_lock);
+	return ret;
+}
+
+static void ocfs2_signal_wipe_completion(struct ocfs2_super *osb,
+					 int slot)
+{
+	spin_lock(&osb->osb_lock);
+	osb->osb_orphan_wipes[slot]--;
+	spin_unlock(&osb->osb_lock);
+
+	wake_up(&osb->osb_wipe_event);
+}
+
 static int ocfs2_wipe_inode(struct inode *inode,
 			    struct buffer_head *di_bh)
 {
@@ -555,6 +592,11 @@ static int ocfs2_wipe_inode(struct inode *inode,
 	/* We've already voted on this so it should be readonly - no
 	 * spinlock needed. */
 	orphaned_slot = OCFS2_I(inode)->ip_orphaned_slot;
+
+	status = ocfs2_check_orphan_recovery_state(osb, orphaned_slot);
+	if (status)
+		return status;
+
 	orphan_dir_inode = ocfs2_get_system_file_inode(osb,
 						       ORPHAN_DIR_SYSTEM_INODE,
 						       orphaned_slot);
@@ -597,6 +639,7 @@ bail_unlock_dir:
 	brelse(orphan_dir_bh);
 bail:
 	iput(orphan_dir_inode);
+	ocfs2_signal_wipe_completion(osb, orphaned_slot);
 
 	return status;
 }
@@ -822,7 +865,8 @@ void ocfs2_delete_inode(struct inode *inode)
 
 	status = ocfs2_wipe_inode(inode, di_bh);
 	if (status < 0) {
-		mlog_errno(status);
+		if (status != -EDEADLK)
+			mlog_errno(status);
 		goto bail_unlock_inode;
 	}
 
diff --git a/fs/ocfs2/journal.c b/fs/ocfs2/journal.c
index d329c9d..4be801f 100644
--- a/fs/ocfs2/journal.c
+++ b/fs/ocfs2/journal.c
@@ -1408,21 +1408,17 @@ bail:
 	return status;
 }
 
-static int ocfs2_recover_orphans(struct ocfs2_super *osb,
-				 int slot)
+static int ocfs2_queue_orphans(struct ocfs2_super *osb,
+			       int slot,
+			       struct inode **head)
 {
-	int status = 0;
-	int have_disk_lock = 0;
-	struct inode *inode = NULL;
-	struct inode *iter;
+	int status;
 	struct inode *orphan_dir_inode = NULL;
+	struct inode *iter;
 	unsigned long offset, blk, local;
 	struct buffer_head *bh = NULL;
 	struct ocfs2_dir_entry *de;
 	struct super_block *sb = osb->sb;
-	struct ocfs2_inode_info *oi;
-
-	mlog(0, "Recover inodes from orphan dir in slot %d\n", slot);
 
 	orphan_dir_inode = ocfs2_get_system_file_inode(osb,
 						       ORPHAN_DIR_SYSTEM_INODE,
@@ -1430,17 +1426,15 @@ static int ocfs2_recover_orphans(struct ocfs2_super *osb,
 	if  (!orphan_dir_inode) {
 		status = -ENOENT;
 		mlog_errno(status);
-		goto out;
-	}
+		return status;
+	}	
 
 	mutex_lock(&orphan_dir_inode->i_mutex);
 	status = ocfs2_meta_lock(orphan_dir_inode, NULL, NULL, 0);
 	if (status < 0) {
-		mutex_unlock(&orphan_dir_inode->i_mutex);
 		mlog_errno(status);
 		goto out;
 	}
-	have_disk_lock = 1;
 
 	offset = 0;
 	iter = NULL;
@@ -1451,11 +1445,10 @@ static int ocfs2_recover_orphans(struct ocfs2_super *osb,
 		if (!bh)
 			status = -EINVAL;
 		if (status < 0) {
-			mutex_unlock(&orphan_dir_inode->i_mutex);
 			if (bh)
 				brelse(bh);
 			mlog_errno(status);
-			goto out;
+			goto out_unlock;
 		}
 
 		local = 0;
@@ -1465,11 +1458,10 @@ static int ocfs2_recover_orphans(struct ocfs2_super *osb,
 
 			if (!ocfs2_check_dir_entry(orphan_dir_inode,
 						  de, bh, local)) {
-				mutex_unlock(&orphan_dir_inode->i_mutex);
 				status = -EINVAL;
 				mlog_errno(status);
 				brelse(bh);
-				goto out;
+				goto out_unlock;
 			}
 
 			local += le16_to_cpu(de->rec_len);
@@ -1504,18 +1496,95 @@ static int ocfs2_recover_orphans(struct ocfs2_super *osb,
 
 			mlog(0, "queue orphan %"MLFu64"\n",
 			     OCFS2_I(iter)->ip_blkno);
-			OCFS2_I(iter)->ip_next_orphan = inode;
-			inode = iter;
+			/* No locking is required for the next_orphan
+			 * queue as there is only ever a single
+			 * process doing orphan recovery. */
+			OCFS2_I(iter)->ip_next_orphan = *head;
+			*head = iter;
 		}
 		brelse(bh);
 	}
-	mutex_unlock(&orphan_dir_inode->i_mutex);
 
+out_unlock:
 	ocfs2_meta_unlock(orphan_dir_inode, 0);
-	have_disk_lock = 0;
-
+out:
+	mutex_unlock(&orphan_dir_inode->i_mutex);
 	iput(orphan_dir_inode);
-	orphan_dir_inode = NULL;
+	return status;
+}
+
+static int ocfs2_orphan_recovery_can_continue(struct ocfs2_super *osb,
+					      int slot)
+{
+	int ret;
+
+	spin_lock(&osb->osb_lock);
+	ret = !osb->osb_orphan_wipes[slot];
+	spin_unlock(&osb->osb_lock);
+	return ret;
+}
+
+static void ocfs2_mark_recovering_orphan_dir(struct ocfs2_super *osb,
+					     int slot)
+{
+	spin_lock(&osb->osb_lock);
+	/* Mark ourselves such that new processes in delete_inode()
+	 * know to quit early. */
+	ocfs2_node_map_set_bit(osb, &osb->osb_recovering_orphan_dirs, slot);
+	while (osb->osb_orphan_wipes[slot]) {
+		/* If any processes are already in the middle of an
+		 * orphan wipe on this dir, then we need to wait for
+		 * them. */
+		spin_unlock(&osb->osb_lock);
+		wait_event_interruptible(osb->osb_wipe_event,
+					 ocfs2_orphan_recovery_can_continue(osb, slot));
+		spin_lock(&osb->osb_lock);
+	}
+	spin_unlock(&osb->osb_lock);
+}
+
+static void ocfs2_clear_recovering_orphan_dir(struct ocfs2_super *osb,
+					      int slot)
+{
+	ocfs2_node_map_clear_bit(osb, &osb->osb_recovering_orphan_dirs, slot);
+}
+
+/*
+ * Orphan recovery. Each mounted node has it's own orphan dir which we
+ * must run during recovery. Our strategy here is to build a list of
+ * the inodes in the orphan dir and iget/iput them. The VFS does
+ * (most) of the rest of the work.
+ *
+ * Orphan recovery can happen at any time, not just mount so we have a
+ * couple of extra considerations.
+ *
+ * - We grab as many inodes as we can under the orphan dir lock -
+ *   doing iget() outside the orphan dir risks getting a reference on
+ *   an invalid inode.
+ * - We must be sure not to deadlock with other processes on the
+ *   system wanting to run delete_inode(). This can happen when they go
+ *   to lock the orphan dir and the orphan recovery process attempts to
+ *   iget() inside the orphan dir lock. This can be avoided by
+ *   advertising our state to ocfs2_delete_inode().
+ */
+static int ocfs2_recover_orphans(struct ocfs2_super *osb,
+				 int slot)
+{
+	int ret = 0;
+	struct inode *inode = NULL;
+	struct inode *iter;
+	struct ocfs2_inode_info *oi;
+
+	mlog(0, "Recover inodes from orphan dir in slot %d\n", slot);
+
+	ocfs2_mark_recovering_orphan_dir(osb, slot);
+	ret = ocfs2_queue_orphans(osb, slot, &inode);
+	ocfs2_clear_recovering_orphan_dir(osb, slot);
+
+	/* Error here should be noted, but we want to continue with as
+	 * many queued inodes as we've got. */
+	if (ret)
+		mlog_errno(ret);
 
 	while (inode) {
 		oi = OCFS2_I(inode);
@@ -1541,14 +1610,7 @@ static int ocfs2_recover_orphans(struct ocfs2_super *osb,
 		inode = iter;
 	}
 
-out:
-	if (have_disk_lock)
-		ocfs2_meta_unlock(orphan_dir_inode, 0);
-
-	if (orphan_dir_inode)
-		iput(orphan_dir_inode);
-
-	return status;
+	return ret;
 }
 
 static int ocfs2_wait_on_mount(struct ocfs2_super *osb)
diff --git a/fs/ocfs2/ocfs2.h b/fs/ocfs2/ocfs2.h
index 19360e3..e89de9b 100644
--- a/fs/ocfs2/ocfs2.h
+++ b/fs/ocfs2/ocfs2.h
@@ -287,6 +287,10 @@ struct ocfs2_super
 	struct inode			*osb_tl_inode;
 	struct buffer_head		*osb_tl_bh;
 	struct work_struct		osb_truncate_log_wq;
+
+	struct ocfs2_node_map		osb_recovering_orphan_dirs;
+	unsigned int			*osb_orphan_wipes;
+	wait_queue_head_t		osb_wipe_event;
 };
 
 #define OCFS2_SB(sb)	    ((struct ocfs2_super *)(sb)->s_fs_info)
diff --git a/fs/ocfs2/super.c b/fs/ocfs2/super.c
index 046824b..8dd3aaf 100644
--- a/fs/ocfs2/super.c
+++ b/fs/ocfs2/super.c
@@ -1325,6 +1325,16 @@ static int ocfs2_initialize_super(struct super_block *sb,
 	}
 	mlog(ML_NOTICE, "max_slots for this device: %u\n", osb->max_slots);
 
+	init_waitqueue_head(&osb->osb_wipe_event);
+	osb->osb_orphan_wipes = kcalloc(osb->max_slots,
+					sizeof(*osb->osb_orphan_wipes),
+					GFP_KERNEL);
+	if (!osb->osb_orphan_wipes) {
+		status = -ENOMEM;
+		mlog_errno(status);
+		goto bail;
+	}
+
 	osb->s_feature_compat =
 		le32_to_cpu(OCFS2_RAW_SB(di)->s_feature_compat);
 	osb->s_feature_ro_compat =
@@ -1638,6 +1648,7 @@ static void ocfs2_delete_osb(struct ocfs2_super *osb)
 	if (osb->slot_info)
 		ocfs2_free_slot_info(osb->slot_info);
 
+	kfree(osb->osb_orphan_wipes);
 	/* FIXME
 	 * This belongs in journal shutdown, but because we have to
 	 * allocate osb->journal at the start of ocfs2_initalize_osb(),