[PATCH v3 06/16] fsnotify: create helpers for group mark_mutex lock

[Amir Goldstein <amir73il@gmail.com>]

Create helpers to take and release the group mark_mutex lock.

Define a flag FSNOTIFY_GROUP_NOFS in fsnotify_group that determines
if the mark_mutex lock is fs reclaim safe or not.  If not safe, the
lock helpers take the lock and disable direct fs reclaim.

In that case we annotate the mutex with a different lockdep class to
express to lockdep that an allocation of mark of an fs reclaim safe group
may take the group lock of another "NOFS" group to evict inodes.

For now, converted only the callers in common code and no backend
defines the NOFS flag.  It is intended to be set by fanotify for
evictable marks support.

Suggested-by: Jan Kara <jack@suse.cz>
Link: https://lore.kernel.org/r/20220321112310.vpr7oxro2xkz5llh@quack3.lan/
Signed-off-by: Amir Goldstein <amir73il@gmail.com>
---
 fs/notify/fdinfo.c               |  4 ++--
 fs/notify/group.c                | 11 +++++++++++
 fs/notify/mark.c                 | 24 +++++++++++-------------
 include/linux/fsnotify_backend.h | 28 ++++++++++++++++++++++++++++
 4 files changed, 52 insertions(+), 15 deletions(-)

diff --git a/fs/notify/fdinfo.c b/fs/notify/fdinfo.c
index 3451708fd035..1f34c5c29fdb 100644
--- a/fs/notify/fdinfo.c
+++ b/fs/notify/fdinfo.c
@@ -28,13 +28,13 @@ static void show_fdinfo(struct seq_file *m, struct file *f,
 	struct fsnotify_group *group = f->private_data;
 	struct fsnotify_mark *mark;
 
-	mutex_lock(&group->mark_mutex);
+	fsnotify_group_lock(group);
 	list_for_each_entry(mark, &group->marks_list, g_list) {
 		show(m, mark);
 		if (seq_has_overflowed(m))
 			break;
 	}
-	mutex_unlock(&group->mark_mutex);
+	fsnotify_group_unlock(group);
 }
 
 #if defined(CONFIG_EXPORTFS)
diff --git a/fs/notify/group.c b/fs/notify/group.c
index 18446b7b0d49..3664029af5ae 100644
--- a/fs/notify/group.c
+++ b/fs/notify/group.c
@@ -115,6 +115,7 @@ static struct fsnotify_group *__fsnotify_alloc_group(
 				const struct fsnotify_ops *ops,
 				int flags, gfp_t gfp)
 {
+	static struct lock_class_key nofs_marks_lock;
 	struct fsnotify_group *group;
 
 	group = kzalloc(sizeof(struct fsnotify_group), gfp);
@@ -135,6 +136,16 @@ static struct fsnotify_group *__fsnotify_alloc_group(
 
 	group->ops = ops;
 	group->flags = flags;
+	/*
+	 * For most backends, eviction of inode with a mark is not expected,
+	 * because marks hold a refcount on the inode against eviction.
+	 *
+	 * Use a different lockdep class for groups that support evictable
+	 * inode marks, because with evictable marks, mark_mutex is NOT
+	 * fs-reclaim safe - the mutex is taken when evicting inodes.
+	 */
+	if (FSNOTIFY_GROUP_FLAG(group, NOFS))
+		lockdep_set_class(&group->mark_mutex, &nofs_marks_lock);
 
 	return group;
 }
diff --git a/fs/notify/mark.c b/fs/notify/mark.c
index ea8f557881b1..7120918d8251 100644
--- a/fs/notify/mark.c
+++ b/fs/notify/mark.c
@@ -398,9 +398,7 @@ void fsnotify_finish_user_wait(struct fsnotify_iter_info *iter_info)
  */
 void fsnotify_detach_mark(struct fsnotify_mark *mark)
 {
-	struct fsnotify_group *group = mark->group;
-
-	WARN_ON_ONCE(!mutex_is_locked(&group->mark_mutex));
+	fsnotify_group_assert_locked(mark->group);
 	WARN_ON_ONCE(!srcu_read_lock_held(&fsnotify_mark_srcu) &&
 		     refcount_read(&mark->refcnt) < 1 +
 			!!(mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED));
@@ -452,9 +450,9 @@ void fsnotify_free_mark(struct fsnotify_mark *mark)
 void fsnotify_destroy_mark(struct fsnotify_mark *mark,
 			   struct fsnotify_group *group)
 {
-	mutex_lock(&group->mark_mutex);
+	fsnotify_group_lock(group);
 	fsnotify_detach_mark(mark);
-	mutex_unlock(&group->mark_mutex);
+	fsnotify_group_unlock(group);
 	fsnotify_free_mark(mark);
 }
 EXPORT_SYMBOL_GPL(fsnotify_destroy_mark);
@@ -673,7 +671,7 @@ int fsnotify_add_mark_locked(struct fsnotify_mark *mark,
 	struct fsnotify_group *group = mark->group;
 	int ret = 0;
 
-	BUG_ON(!mutex_is_locked(&group->mark_mutex));
+	fsnotify_group_assert_locked(group);
 
 	/*
 	 * LOCKING ORDER!!!!
@@ -713,9 +711,9 @@ int fsnotify_add_mark(struct fsnotify_mark *mark, fsnotify_connp_t *connp,
 	int ret;
 	struct fsnotify_group *group = mark->group;
 
-	mutex_lock(&group->mark_mutex);
+	fsnotify_group_lock(group);
 	ret = fsnotify_add_mark_locked(mark, connp, obj_type, fsid);
-	mutex_unlock(&group->mark_mutex);
+	fsnotify_group_unlock(group);
 	return ret;
 }
 EXPORT_SYMBOL_GPL(fsnotify_add_mark);
@@ -769,24 +767,24 @@ void fsnotify_clear_marks_by_group(struct fsnotify_group *group,
 	 * move marks to free to to_free list in one go and then free marks in
 	 * to_free list one by one.
 	 */
-	mutex_lock(&group->mark_mutex);
+	fsnotify_group_lock(group);
 	list_for_each_entry_safe(mark, lmark, &group->marks_list, g_list) {
 		if (mark->connector->type == obj_type)
 			list_move(&mark->g_list, &to_free);
 	}
-	mutex_unlock(&group->mark_mutex);
+	fsnotify_group_unlock(group);
 
 clear:
 	while (1) {
-		mutex_lock(&group->mark_mutex);
+		fsnotify_group_lock(group);
 		if (list_empty(head)) {
-			mutex_unlock(&group->mark_mutex);
+			fsnotify_group_unlock(group);
 			break;
 		}
 		mark = list_first_entry(head, struct fsnotify_mark, g_list);
 		fsnotify_get_mark(mark);
 		fsnotify_detach_mark(mark);
-		mutex_unlock(&group->mark_mutex);
+		fsnotify_group_unlock(group);
 		fsnotify_free_mark(mark);
 		fsnotify_put_mark(mark);
 	}
diff --git a/include/linux/fsnotify_backend.h b/include/linux/fsnotify_backend.h
index 2057ae4bf8e9..fc6fdbaac797 100644
--- a/include/linux/fsnotify_backend.h
+++ b/include/linux/fsnotify_backend.h
@@ -20,6 +20,7 @@
 #include <linux/user_namespace.h>
 #include <linux/refcount.h>
 #include <linux/mempool.h>
+#include <linux/sched/mm.h>
 
 /*
  * IN_* from inotfy.h lines up EXACTLY with FS_*, this is so we can easily
@@ -211,9 +212,11 @@ struct fsnotify_group {
 	bool shutdown;		/* group is being shut down, don't queue more events */
 
 #define FSNOTIFY_GROUP_USER	0x01 /* user allocated group */
+#define FSNOTIFY_GROUP_NOFS	0x02 /* group lock is not direct reclaim safe */
 #define FSNOTIFY_GROUP_FLAG(group, flag) \
 	((group)->flags & FSNOTIFY_GROUP_ ## flag)
 	int flags;
+	unsigned int owner_flags;	/* stored flags of mark_mutex owner */
 
 	/* stores all fastpath marks assoc with this group so they can be cleaned on unregister */
 	struct mutex mark_mutex;	/* protect marks_list */
@@ -255,6 +258,31 @@ struct fsnotify_group {
 	};
 };
 
+/*
+ * These helpers are used to prevent deadlock when reclaiming inodes with
+ * evictable marks of the same group that is allocating a new mark.
+ */
+static inline void fsnotify_group_lock(struct fsnotify_group *group)
+{
+	mutex_lock(&group->mark_mutex);
+	if (FSNOTIFY_GROUP_FLAG(group, NOFS))
+		group->owner_flags = memalloc_nofs_save();
+}
+
+static inline void fsnotify_group_unlock(struct fsnotify_group *group)
+{
+	if (FSNOTIFY_GROUP_FLAG(group, NOFS))
+		memalloc_nofs_restore(group->owner_flags);
+	mutex_unlock(&group->mark_mutex);
+}
+
+static inline void fsnotify_group_assert_locked(struct fsnotify_group *group)
+{
+	WARN_ON_ONCE(!mutex_is_locked(&group->mark_mutex));
+	if (FSNOTIFY_GROUP_FLAG(group, NOFS))
+		WARN_ON_ONCE(!(current->flags & PF_MEMALLOC_NOFS));
+}
+
 /* When calling fsnotify tell it if the data is a path or inode */
 enum fsnotify_data_type {
 	FSNOTIFY_EVENT_NONE,
-- 
2.35.1