fsnotify: srcu to protect read side of inode and vfsmount locks

Currently reading the inode->i_fsnotify_marks or
vfsmount->mnt_fsnotify_marks lists are protected by a spinlock on both the
read and the write side.  This patch protects the read side of those lists
with a new single srcu.

Signed-off-by: Eric Paris <eparis@redhat.com>

fs/notify/fsnotify.c

@@ -144,29 +144,30 @@ void __fsnotify_flush_ignored_mask(struct inode *inode, void *data, int data_is)
 {
 	struct fsnotify_mark *mark;
 	struct hlist_node *node;
+	int idx;
+
+	idx = srcu_read_lock(&fsnotify_mark_srcu);
 
 	if (!hlist_empty(&inode->i_fsnotify_marks)) {
-		spin_lock(&inode->i_lock);
-		hlist_for_each_entry(mark, node, &inode->i_fsnotify_marks, i.i_list) {
+		hlist_for_each_entry_rcu(mark, node, &inode->i_fsnotify_marks, i.i_list) {
 			if (!(mark->flags & FSNOTIFY_MARK_FLAG_IGNORED_SURV_MODIFY))
 				mark->ignored_mask = 0;
 		}
-		spin_unlock(&inode->i_lock);
 	}
 
 	if (data_is == FSNOTIFY_EVENT_FILE) {
 		struct vfsmount *mnt;
 
 		mnt = ((struct file *)data)->f_path.mnt;
 		if (mnt && !hlist_empty(&mnt->mnt_fsnotify_marks)) {
-			spin_lock(&mnt->mnt_root->d_lock);
-			hlist_for_each_entry(mark, node, &mnt->mnt_fsnotify_marks, m.m_list) {
+			hlist_for_each_entry_rcu(mark, node, &mnt->mnt_fsnotify_marks, m.m_list) {
 				if (!(mark->flags & FSNOTIFY_MARK_FLAG_IGNORED_SURV_MODIFY))
 					mark->ignored_mask = 0;
 			}
-			spin_unlock(&mnt->mnt_root->d_lock);
 		}
 	}
+
+	srcu_read_unlock(&fsnotify_mark_srcu, idx);
 }
 
 static int send_to_group(struct fsnotify_group *group, struct inode *to_tell,
@@ -208,8 +209,10 @@ static bool needed_by_vfsmount(__u32 test_mask, struct vfsmount *mnt)
 int fsnotify(struct inode *to_tell, __u32 mask, void *data, int data_is,
 	     const unsigned char *file_name, u32 cookie)
 {
+	struct fsnotify_mark *mark;
 	struct fsnotify_group *group;
 	struct fsnotify_event *event = NULL;
+	struct hlist_node *node;
 	struct vfsmount *mnt = NULL;
 	int idx, ret = 0;
 	/* global tests shouldn't care about events on child only the specific event */
@@ -237,35 +240,47 @@ int fsnotify(struct inode *to_tell, __u32 mask, void *data, int data_is,
 	    !needed_by_vfsmount(test_mask, mnt))
 		return 0;
 
-	/*
-	 * SRCU!!  the groups list is very very much read only and the path is
-	 * very hot.  The VAST majority of events are not going to need to do
-	 * anything other than walk the list so it's crazy to pre-allocate.
-	 */
-	idx = srcu_read_lock(&fsnotify_grp_srcu);
+	idx = srcu_read_lock(&fsnotify_mark_srcu);
 
 	if (test_mask & to_tell->i_fsnotify_mask) {
-		list_for_each_entry_rcu(group, &fsnotify_inode_groups, inode_group_list) {
-			if (test_mask & group->mask) {
-				ret = send_to_group(group, to_tell, NULL, mask, data, data_is,
-						    cookie, file_name, &event);
+		hlist_for_each_entry_rcu(mark, node, &to_tell->i_fsnotify_marks, i.i_list) {
+
+			pr_debug("%s: inode_loop: mark=%p mark->mask=%x mark->ignored_mask=%x\n",
+				 __func__, mark, mark->mask, mark->ignored_mask);
+
+			if (test_mask & mark->mask & ~mark->ignored_mask) {
+				group = mark->group;
+				if (!group)
+					continue;
+				ret = send_to_group(group, to_tell, NULL, mask,
+						    data, data_is, cookie, file_name,
+						    &event);
 				if (ret)
 					goto out;
 			}
 		}
 	}
-	if (needed_by_vfsmount(test_mask, mnt)) {
-		list_for_each_entry_rcu(group, &fsnotify_vfsmount_groups, vfsmount_group_list) {
-			if (test_mask & group->mask) {
-				ret = send_to_group(group, to_tell, mnt, mask, data, data_is,
-						    cookie, file_name, &event);
+
+	if (mnt && (test_mask & mnt->mnt_fsnotify_mask)) {
+		hlist_for_each_entry_rcu(mark, node, &mnt->mnt_fsnotify_marks, m.m_list) {
+
+			pr_debug("%s: mnt_loop: mark=%p mark->mask=%x mark->ignored_mask=%x\n",
+				 __func__, mark, mark->mask, mark->ignored_mask);
+
+			if (test_mask & mark->mask & ~mark->ignored_mask)  {
+				group = mark->group;
+				if (!group)
+					continue;
+				ret = send_to_group(group, to_tell, mnt, mask,
+						    data, data_is, cookie, file_name,
+						    &event);
 				if (ret)
 					goto out;
 			}
 		}
 	}
 out:
-	srcu_read_unlock(&fsnotify_grp_srcu, idx);
+	srcu_read_unlock(&fsnotify_mark_srcu, idx);
 	/*
 	 * fsnotify_create_event() took a reference so the event can't be cleaned
 	 * up while we are still trying to add it to lists, drop that one.
@@ -279,8 +294,14 @@ EXPORT_SYMBOL_GPL(fsnotify);
 
 static __init int fsnotify_init(void)
 {
+	int ret;
+
 	BUG_ON(hweight32(ALL_FSNOTIFY_EVENTS) != 23);
 
-	return init_srcu_struct(&fsnotify_grp_srcu);
+	ret = init_srcu_struct(&fsnotify_mark_srcu);
+	if (ret)
+		panic("initializing fsnotify_mark_srcu");
+
+	return 0;
 }
-subsys_initcall(fsnotify_init);
+core_initcall(fsnotify_init);

fs/notify/fsnotify.h

@@ -6,8 +6,6 @@
 #include <linux/srcu.h>
 #include <linux/types.h>
 
-/* protects reads of fsnotify_groups */
-extern struct srcu_struct fsnotify_grp_srcu;
 /* all groups which receive inode fsnotify events */
 extern struct list_head fsnotify_inode_groups;
 /* all groups which receive vfsmount fsnotify events */
@@ -20,6 +18,9 @@ extern __u32 fsnotify_vfsmount_mask;
 /* destroy all events sitting in this groups notification queue */
 extern void fsnotify_flush_notify(struct fsnotify_group *group);
 
+/* protects reads of inode and vfsmount marks list */
+extern struct srcu_struct fsnotify_mark_srcu;
+
 extern void fsnotify_set_inode_mark_mask_locked(struct fsnotify_mark *fsn_mark,
 						__u32 mask);
 /* add a mark to an inode */

fs/notify/group.c

@@ -30,8 +30,6 @@
 
 /* protects writes to fsnotify_groups and fsnotify_mask */
 static DEFINE_MUTEX(fsnotify_grp_mutex);
-/* protects reads while running the fsnotify_groups list */
-struct srcu_struct fsnotify_grp_srcu;
 /* all groups registered to receive inode filesystem notifications */
 LIST_HEAD(fsnotify_inode_groups);
 /* all groups registered to receive mount point filesystem notifications */
@@ -50,18 +48,17 @@ void fsnotify_recalc_global_mask(void)
 	struct fsnotify_group *group;
 	__u32 inode_mask = 0;
 	__u32 vfsmount_mask = 0;
-	int idx;
 
-	idx = srcu_read_lock(&fsnotify_grp_srcu);
+	mutex_lock(&fsnotify_grp_mutex);
 	list_for_each_entry_rcu(group, &fsnotify_inode_groups, inode_group_list)
 		inode_mask |= group->mask;
 	list_for_each_entry_rcu(group, &fsnotify_vfsmount_groups, vfsmount_group_list)
 		vfsmount_mask |= group->mask;
-
-	srcu_read_unlock(&fsnotify_grp_srcu, idx);
 
 	fsnotify_inode_mask = inode_mask;
 	fsnotify_vfsmount_mask = vfsmount_mask;
+
+	mutex_unlock(&fsnotify_grp_mutex);
 }
 
 /*
@@ -168,6 +165,8 @@ static void fsnotify_destroy_group(struct fsnotify_group *group)
 	/* clear all inode marks for this group */
 	fsnotify_clear_marks_by_group(group);
 
+	synchronize_srcu(&fsnotify_mark_srcu);
+
 	/* past the point of no return, matches the initial value of 1 */
 	if (atomic_dec_and_test(&group->num_marks))
 		fsnotify_final_destroy_group(group);
@@ -216,12 +215,7 @@ void fsnotify_put_group(struct fsnotify_group *group)
 	 */
 	__fsnotify_evict_group(group);
 
-	/*
-	 * now it's off the list, so the only thing we might care about is
-	 * srcu access....
-	 */
 	mutex_unlock(&fsnotify_grp_mutex);
-	synchronize_srcu(&fsnotify_grp_srcu);
 
 	/* and now it is really dead. _Nothing_ could be seeing it */
 	fsnotify_recalc_global_mask();

fs/notify/mark.c

@@ -85,17 +85,24 @@
 #include <linux/fs.h>
 #include <linux/init.h>
 #include <linux/kernel.h>
+#include <linux/kthread.h>
 #include <linux/module.h>
 #include <linux/mutex.h>
 #include <linux/slab.h>
 #include <linux/spinlock.h>
+#include <linux/srcu.h>
 #include <linux/writeback.h> /* for inode_lock */
 
 #include <asm/atomic.h>
 
 #include <linux/fsnotify_backend.h>
 #include "fsnotify.h"
 
+struct srcu_struct fsnotify_mark_srcu;
+static DEFINE_SPINLOCK(destroy_lock);
+static LIST_HEAD(destroy_list);
+static DECLARE_WAIT_QUEUE_HEAD(destroy_waitq);
+
 void fsnotify_get_mark(struct fsnotify_mark *mark)
 {
 	atomic_inc(&mark->refcnt);
@@ -144,11 +151,14 @@ void fsnotify_destroy_mark(struct fsnotify_mark *mark)
 
 	list_del_init(&mark->g_list);
 
-	fsnotify_put_mark(mark); /* for i_list and g_list */
-
 	spin_unlock(&group->mark_lock);
 	spin_unlock(&mark->lock);
 
+	spin_lock(&destroy_lock);
+	list_add(&mark->destroy_list, &destroy_list);
+	spin_unlock(&destroy_lock);
+	wake_up(&destroy_waitq);
+
 	/*
 	 * Some groups like to know that marks are being freed.  This is a
 	 * callback to the group function to let it know that this mark
@@ -263,12 +273,17 @@ int fsnotify_add_mark(struct fsnotify_mark *mark,
 err:
 	mark->flags &= ~FSNOTIFY_MARK_FLAG_ALIVE;
 	list_del_init(&mark->g_list);
+	mark->group = NULL;
 	atomic_dec(&group->num_marks);
-	fsnotify_put_mark(mark);
 
 	spin_unlock(&group->mark_lock);
 	spin_unlock(&mark->lock);
 
+	spin_lock(&destroy_lock);
+	list_add(&mark->destroy_list, &destroy_list);
+	spin_unlock(&destroy_lock);
+	wake_up(&destroy_waitq);
+
 	return ret;
 }
 
@@ -326,3 +341,42 @@ void fsnotify_init_mark(struct fsnotify_mark *mark,
 	atomic_set(&mark->refcnt, 1);
 	mark->free_mark = free_mark;
 }
+
+static int fsnotify_mark_destroy(void *ignored)
+{
+	struct fsnotify_mark *mark, *next;
+	LIST_HEAD(private_destroy_list);
+
+	for (;;) {
+		spin_lock(&destroy_lock);
+		list_for_each_entry_safe(mark, next, &destroy_list, destroy_list) {
+			list_del(&mark->destroy_list);
+			list_add(&mark->destroy_list, &private_destroy_list);
+		}
+		spin_unlock(&destroy_lock);
+
+		synchronize_srcu(&fsnotify_mark_srcu);
+
+		list_for_each_entry_safe(mark, next, &private_destroy_list, destroy_list) {
+			list_del_init(&mark->destroy_list);
+			fsnotify_put_mark(mark);
+		}
+
+		wait_event_interruptible(destroy_waitq, !list_empty(&destroy_list));
+	}
+
+	return 0;
+}
+
+static int __init fsnotify_mark_init(void)
+{
+	struct task_struct *thread;
+
+	thread = kthread_run(fsnotify_mark_destroy, NULL,
+			     "fsnotify_mark");
+	if (IS_ERR(thread))
+		panic("unable to start fsnotify mark destruction thread.");
+
+	return 0;
+}
+device_initcall(fsnotify_mark_init);

include/linux/fsnotify_backend.h

@@ -302,6 +302,7 @@ struct fsnotify_mark {
 #define FSNOTIFY_MARK_FLAG_IGNORED_SURV_MODIFY	0x08
 #define FSNOTIFY_MARK_FLAG_ALIVE		0x10
 	unsigned int flags;		/* vfsmount or inode mark? */
+	struct list_head destroy_list;
 	void (*free_mark)(struct fsnotify_mark *mark); /* called on final put+free */
 };