fsnotify: use hash table for faster events merge

In order to improve event merge performance, hash events in a 128 size
hash table by the event merge key.

The fanotify_event size grows by two pointers, but we just reduced its
size by removing the objectid member, so overall its size is increased
by one pointer.

Permission events and overflow event are not merged so they are also
not hashed.

Link: https://lore.kernel.org/r/20210304104826.3993892-5-amir73il@gmail.com
Signed-off-by: Amir Goldstein <amir73il@gmail.com>
Signed-off-by: Jan Kara <jack@suse.cz>

3 files changed, 97 insertions, 7 deletions

diff --git a/fs/notify/fanotify/fanotify.c b/fs/notify/fanotify/fanotify.c
index 43a606f15370..50b3abc06215 100644
--- a/fs/notify/fanotify/fanotify.c
+++ b/fs/notify/fanotify/fanotify.c
@@ -149,12 +149,15 @@ static bool fanotify_should_merge(struct fanotify_event *old,
 }
 
 /* and the list better be locked by something too! */
-static int fanotify_merge(struct list_head *list, struct fsnotify_event *event)
+static int fanotify_merge(struct fsnotify_group *group,
+			  struct fsnotify_event *event)
 {
-	struct fsnotify_event *test_event;
 	struct fanotify_event *old, *new = FANOTIFY_E(event);
+	unsigned int bucket = fanotify_event_hash_bucket(group, new);
+	struct hlist_head *hlist = &group->fanotify_data.merge_hash[bucket];
 
-	pr_debug("%s: list=%p event=%p\n", __func__, list, event);
+	pr_debug("%s: group=%p event=%p bucket=%u\n", __func__,
+		 group, event, bucket);
 
 	/*
 	 * Don't merge a permission event with any other event so that we know
@@ -164,8 +167,7 @@ static int fanotify_merge(struct list_head *list, struct fsnotify_event *event)
 	if (fanotify_is_perm_event(new->mask))
 		return 0;
 
-	list_for_each_entry_reverse(test_event, list, list) {
-		old = FANOTIFY_E(test_event);
+	hlist_for_each_entry(old, hlist, merge_list) {
 		if (fanotify_should_merge(old, new)) {
 			old->mask |= new->mask;
 			return 1;
@@ -203,8 +205,11 @@ static int fanotify_get_response(struct fsnotify_group *group,
 			return ret;
 		}
 		/* Event not yet reported? Just remove it. */
-		if (event->state == FAN_EVENT_INIT)
+		if (event->state == FAN_EVENT_INIT) {
 			fsnotify_remove_queued_event(group, &event->fae.fse);
+			/* Permission events are not supposed to be hashed */
+			WARN_ON_ONCE(!hlist_unhashed(&event->fae.merge_list));
+		}
 		/*
 		 * Event may be also answered in case signal delivery raced
 		 * with wakeup. In that case we have nothing to do besides
@@ -679,6 +684,24 @@ static __kernel_fsid_t fanotify_get_fsid(struct fsnotify_iter_info *iter_info)
 	return fsid;
 }
 
+/*
+ * Add an event to hash table for faster merge.
+ */
+static void fanotify_insert_event(struct fsnotify_group *group,
+				  struct fsnotify_event *fsn_event)
+{
+	struct fanotify_event *event = FANOTIFY_E(fsn_event);
+	unsigned int bucket = fanotify_event_hash_bucket(group, event);
+	struct hlist_head *hlist = &group->fanotify_data.merge_hash[bucket];
+
+	assert_spin_locked(&group->notification_lock);
+
+	pr_debug("%s: group=%p event=%p bucket=%u\n", __func__,
+		 group, event, bucket);
+
+	hlist_add_head(&event->merge_list, hlist);
+}
+
 static int fanotify_handle_event(struct fsnotify_group *group, u32 mask,
 				 const void *data, int data_type,
 				 struct inode *dir,
@@ -749,7 +772,9 @@ static int fanotify_handle_event(struct fsnotify_group *group, u32 mask,
 	}
 
 	fsn_event = &event->fse;
-	ret = fsnotify_add_event(group, fsn_event, fanotify_merge);
+	ret = fsnotify_add_event(group, fsn_event, fanotify_merge,
+				 fanotify_is_hashed_event(mask) ?
+				 fanotify_insert_event : NULL);
 	if (ret) {
 		/* Permission events shouldn't be merged */
 		BUG_ON(ret == 1 && mask & FANOTIFY_PERM_EVENTS);
@@ -772,6 +797,7 @@ static void fanotify_free_group_priv(struct fsnotify_group *group)
 {
 	struct user_struct *user;
 
+	kfree(group->fanotify_data.merge_hash);
 	user = group->fanotify_data.user;
 	atomic_dec(&user->fanotify_listeners);
 	free_uid(user);
diff --git a/fs/notify/fanotify/fanotify.h b/fs/notify/fanotify/fanotify.h
index 9871f76cd9c2..4a5e555dc3d2 100644
--- a/fs/notify/fanotify/fanotify.h
+++ b/fs/notify/fanotify/fanotify.h
@@ -3,6 +3,7 @@
 #include <linux/path.h>
 #include <linux/slab.h>
 #include <linux/exportfs.h>
+#include <linux/hashtable.h>
 
 extern struct kmem_cache *fanotify_mark_cache;
 extern struct kmem_cache *fanotify_fid_event_cachep;
@@ -150,6 +151,7 @@ enum fanotify_event_type {
 
 struct fanotify_event {
 	struct fsnotify_event fse;
+	struct hlist_node merge_list;	/* List for hashed merge */
 	u32 mask;
 	struct {
 		unsigned int type : FANOTIFY_EVENT_TYPE_BITS;
@@ -162,6 +164,7 @@ static inline void fanotify_init_event(struct fanotify_event *event,
 				       unsigned int hash, u32 mask)
 {
 	fsnotify_init_event(&event->fse);
+	INIT_HLIST_NODE(&event->merge_list);
 	event->hash = hash;
 	event->mask = mask;
 	event->pid = NULL;
@@ -299,3 +302,25 @@ static inline struct path *fanotify_event_path(struct fanotify_event *event)
 	else
 		return NULL;
 }
+
+/*
+ * Use 128 size hash table to speed up events merge.
+ */
+#define FANOTIFY_HTABLE_BITS	(7)
+#define FANOTIFY_HTABLE_SIZE	(1 << FANOTIFY_HTABLE_BITS)
+#define FANOTIFY_HTABLE_MASK	(FANOTIFY_HTABLE_SIZE - 1)
+
+/*
+ * Permission events and overflow event do not get merged - don't hash them.
+ */
+static inline bool fanotify_is_hashed_event(u32 mask)
+{
+	return !fanotify_is_perm_event(mask) && !(mask & FS_Q_OVERFLOW);
+}
+
+static inline unsigned int fanotify_event_hash_bucket(
+						struct fsnotify_group *group,
+						struct fanotify_event *event)
+{
+	return event->hash & FANOTIFY_HTABLE_MASK;
+}
diff --git a/fs/notify/fanotify/fanotify_user.c b/fs/notify/fanotify/fanotify_user.c
index 16162207e886..b89f332248bd 100644
--- a/fs/notify/fanotify/fanotify_user.c
+++ b/fs/notify/fanotify/fanotify_user.c
@@ -90,6 +90,23 @@ static int fanotify_event_info_len(unsigned int fid_mode,
 }
 
 /*
+ * Remove an hashed event from merge hash table.
+ */
+static void fanotify_unhash_event(struct fsnotify_group *group,
+				  struct fanotify_event *event)
+{
+	assert_spin_locked(&group->notification_lock);
+
+	pr_debug("%s: group=%p event=%p bucket=%u\n", __func__,
+		 group, event, fanotify_event_hash_bucket(group, event));
+
+	if (WARN_ON_ONCE(hlist_unhashed(&event->merge_list)))
+		return;
+
+	hlist_del_init(&event->merge_list);
+}
+
+/*
  * Get an fanotify notification event if one exists and is small
  * enough to fit in "count". Return an error pointer if the count
  * is not large enough. When permission event is dequeued, its state is
@@ -126,6 +143,8 @@ static struct fanotify_event *get_one_event(struct fsnotify_group *group,
 	fsnotify_remove_first_event(group);
 	if (fanotify_is_perm_event(event->mask))
 		FANOTIFY_PERM(event)->state = FAN_EVENT_REPORTED;
+	if (fanotify_is_hashed_event(event->mask))
+		fanotify_unhash_event(group, event);
 out:
 	spin_unlock(&group->notification_lock);
 	return event;
@@ -925,6 +944,20 @@ static struct fsnotify_event *fanotify_alloc_overflow_event(void)
 	return &oevent->fse;
 }
 
+static struct hlist_head *fanotify_alloc_merge_hash(void)
+{
+	struct hlist_head *hash;
+
+	hash = kmalloc(sizeof(struct hlist_head) << FANOTIFY_HTABLE_BITS,
+		       GFP_KERNEL_ACCOUNT);
+	if (!hash)
+		return NULL;
+
+	__hash_init(hash, FANOTIFY_HTABLE_SIZE);
+
+	return hash;
+}
+
 /* fanotify syscalls */
 SYSCALL_DEFINE2(fanotify_init, unsigned int, flags, unsigned int, event_f_flags)
 {
@@ -993,6 +1026,12 @@ SYSCALL_DEFINE2(fanotify_init, unsigned int, flags, unsigned int, event_f_flags)
 	atomic_inc(&user->fanotify_listeners);
 	group->memcg = get_mem_cgroup_from_mm(current->mm);
 
+	group->fanotify_data.merge_hash = fanotify_alloc_merge_hash();
+	if (!group->fanotify_data.merge_hash) {
+		fd = -ENOMEM;
+		goto out_destroy_group;
+	}
+
 	group->overflow_event = fanotify_alloc_overflow_event();
 	if (unlikely(!group->overflow_event)) {
 		fd = -ENOMEM;