From a2d8bc6cb4a3024661baf877242f123787d0c054 Mon Sep 17 00:00:00 2001 From: Eric Paris Date: Thu, 21 May 2009 17:01:37 -0400 Subject: fsnotify: generic notification queue and waitq inotify needs to do asyc notification in which event information is stored on a queue until the listener is ready to receive it. This patch implements a generic notification queue for inotify (and later fanotify) to store events to be sent at a later time. Signed-off-by: Eric Paris Acked-by: Al Viro Cc: Christoph Hellwig --- fs/notify/notification.c | 230 +++++++++++++++++++++++++++++++++++++++++++++-- 1 file changed, 223 insertions(+), 7 deletions(-) (limited to 'fs/notify/notification.c') diff --git a/fs/notify/notification.c b/fs/notify/notification.c index b8e9a87..dddecc7 100644 --- a/fs/notify/notification.c +++ b/fs/notify/notification.c @@ -16,6 +16,21 @@ * the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA. */ +/* + * Basic idea behind the notification queue: An fsnotify group (like inotify) + * sends the userspace notification about events asyncronously some time after + * the event happened. When inotify gets an event it will need to add that + * event to the group notify queue. Since a single event might need to be on + * multiple group's notification queues we can't add the event directly to each + * queue and instead add a small "event_holder" to each queue. This event_holder + * has a pointer back to the original event. Since the majority of events are + * going to end up on one, and only one, notification queue we embed one + * event_holder into each event. This means we have a single allocation instead + * of always needing two. If the embedded event_holder is already in use by + * another group a new event_holder (from fsnotify_event_holder_cachep) will be + * allocated and used. + */ + #include #include #include @@ -33,6 +48,21 @@ #include "fsnotify.h" static struct kmem_cache *fsnotify_event_cachep; +static struct kmem_cache *fsnotify_event_holder_cachep; +/* + * This is a magic event we send when the q is too full. Since it doesn't + * hold real event information we just keep one system wide and use it any time + * it is needed. It's refcnt is set 1 at kernel init time and will never + * get set to 0 so it will never get 'freed' + */ +static struct fsnotify_event q_overflow_event; + +/* return true if the notify queue is empty, false otherwise */ +bool fsnotify_notify_queue_is_empty(struct fsnotify_group *group) +{ + BUG_ON(!mutex_is_locked(&group->notification_mutex)); + return list_empty(&group->notification_list) ? true : false; +} void fsnotify_get_event(struct fsnotify_event *event) { @@ -52,19 +82,176 @@ void fsnotify_put_event(struct fsnotify_event *event) } } +struct fsnotify_event_holder *fsnotify_alloc_event_holder(void) +{ + return kmem_cache_alloc(fsnotify_event_holder_cachep, GFP_KERNEL); +} + +void fsnotify_destroy_event_holder(struct fsnotify_event_holder *holder) +{ + kmem_cache_free(fsnotify_event_holder_cachep, holder); +} + +/* + * check if 2 events contain the same information. + */ +static bool event_compare(struct fsnotify_event *old, struct fsnotify_event *new) +{ + if ((old->mask == new->mask) && + (old->to_tell == new->to_tell) && + (old->data_type == new->data_type)) { + switch (old->data_type) { + case (FSNOTIFY_EVENT_INODE): + if (old->inode == new->inode) + return true; + break; + case (FSNOTIFY_EVENT_PATH): + if ((old->path.mnt == new->path.mnt) && + (old->path.dentry == new->path.dentry)) + return true; + case (FSNOTIFY_EVENT_NONE): + return true; + }; + } + return false; +} + /* - * Allocate a new event which will be sent to each group's handle_event function - * if the group was interested in this particular event. + * Add an event to the group notification queue. The group can later pull this + * event off the queue to deal with. If the event is successfully added to the + * group's notification queue, a reference is taken on event. */ -struct fsnotify_event *fsnotify_create_event(struct inode *to_tell, __u32 mask, - void *data, int data_type) +int fsnotify_add_notify_event(struct fsnotify_group *group, struct fsnotify_event *event) +{ + struct fsnotify_event_holder *holder = NULL; + struct list_head *list = &group->notification_list; + struct fsnotify_event_holder *last_holder; + struct fsnotify_event *last_event; + + /* + * There is one fsnotify_event_holder embedded inside each fsnotify_event. + * Check if we expect to be able to use that holder. If not alloc a new + * holder. + * For the overflow event it's possible that something will use the in + * event holder before we get the lock so we may need to jump back and + * alloc a new holder, this can't happen for most events... + */ + if (!list_empty(&event->holder.event_list)) { +alloc_holder: + holder = fsnotify_alloc_event_holder(); + if (!holder) + return -ENOMEM; + } + + mutex_lock(&group->notification_mutex); + + if (group->q_len >= group->max_events) + event = &q_overflow_event; + + spin_lock(&event->lock); + + if (list_empty(&event->holder.event_list)) { + if (unlikely(holder)) + fsnotify_destroy_event_holder(holder); + holder = &event->holder; + } else if (unlikely(!holder)) { + /* between the time we checked above and got the lock the in + * event holder was used, go back and get a new one */ + spin_unlock(&event->lock); + mutex_unlock(&group->notification_mutex); + goto alloc_holder; + } + + if (!list_empty(list)) { + last_holder = list_entry(list->prev, struct fsnotify_event_holder, event_list); + last_event = last_holder->event; + if (event_compare(last_event, event)) { + spin_unlock(&event->lock); + mutex_unlock(&group->notification_mutex); + if (holder != &event->holder) + fsnotify_destroy_event_holder(holder); + return 0; + } + } + + group->q_len++; + holder->event = event; + + fsnotify_get_event(event); + list_add_tail(&holder->event_list, list); + spin_unlock(&event->lock); + mutex_unlock(&group->notification_mutex); + + wake_up(&group->notification_waitq); + return 0; +} + +/* + * Remove and return the first event from the notification list. There is a + * reference held on this event since it was on the list. It is the responsibility + * of the caller to drop this reference. + */ +struct fsnotify_event *fsnotify_remove_notify_event(struct fsnotify_group *group) { struct fsnotify_event *event; + struct fsnotify_event_holder *holder; - event = kmem_cache_alloc(fsnotify_event_cachep, GFP_KERNEL); - if (!event) - return NULL; + BUG_ON(!mutex_is_locked(&group->notification_mutex)); + holder = list_first_entry(&group->notification_list, struct fsnotify_event_holder, event_list); + + event = holder->event; + + spin_lock(&event->lock); + holder->event = NULL; + list_del_init(&holder->event_list); + spin_unlock(&event->lock); + + /* event == holder means we are referenced through the in event holder */ + if (holder != &event->holder) + fsnotify_destroy_event_holder(holder); + + group->q_len--; + + return event; +} + +/* + * This will not remove the event, that must be done with fsnotify_remove_notify_event() + */ +struct fsnotify_event *fsnotify_peek_notify_event(struct fsnotify_group *group) +{ + struct fsnotify_event *event; + struct fsnotify_event_holder *holder; + + BUG_ON(!mutex_is_locked(&group->notification_mutex)); + + holder = list_first_entry(&group->notification_list, struct fsnotify_event_holder, event_list); + event = holder->event; + + return event; +} + +/* + * Called when a group is being torn down to clean up any outstanding + * event notifications. + */ +void fsnotify_flush_notify(struct fsnotify_group *group) +{ + struct fsnotify_event *event; + + mutex_lock(&group->notification_mutex); + while (!fsnotify_notify_queue_is_empty(group)) { + event = fsnotify_remove_notify_event(group); + fsnotify_put_event(event); /* matches fsnotify_add_notify_event */ + } + mutex_unlock(&group->notification_mutex); +} + +static void initialize_event(struct fsnotify_event *event) +{ + event->holder.event = NULL; + INIT_LIST_HEAD(&event->holder.event_list); atomic_set(&event->refcnt, 1); spin_lock_init(&event->lock); @@ -72,7 +259,32 @@ struct fsnotify_event *fsnotify_create_event(struct inode *to_tell, __u32 mask, event->path.dentry = NULL; event->path.mnt = NULL; event->inode = NULL; + event->data_type = FSNOTIFY_EVENT_NONE; + event->to_tell = NULL; +} + +/* + * fsnotify_create_event - Allocate a new event which will be sent to each + * group's handle_event function if the group was interested in this + * particular event. + * + * @to_tell the inode which is supposed to receive the event (sometimes a + * parent of the inode to which the event happened. + * @mask what actually happened. + * @data pointer to the object which was actually affected + * @data_type flag indication if the data is a file, path, inode, nothing... + */ +struct fsnotify_event *fsnotify_create_event(struct inode *to_tell, __u32 mask, + void *data, int data_type) +{ + struct fsnotify_event *event; + + event = kmem_cache_alloc(fsnotify_event_cachep, GFP_KERNEL); + if (!event) + return NULL; + + initialize_event(event); event->to_tell = to_tell; switch (data_type) { @@ -114,6 +326,10 @@ struct fsnotify_event *fsnotify_create_event(struct inode *to_tell, __u32 mask, __init int fsnotify_notification_init(void) { fsnotify_event_cachep = KMEM_CACHE(fsnotify_event, SLAB_PANIC); + fsnotify_event_holder_cachep = KMEM_CACHE(fsnotify_event_holder, SLAB_PANIC); + + initialize_event(&q_overflow_event); + q_overflow_event.mask = FS_Q_OVERFLOW; return 0; } -- cgit v1.1