Node Monitoring

This document describes the feature of the BeOS kernel to monitor nodes. First, there is an explanation of what kind of functionality we have to reproduce (along with the higher level API), then we will present the implementation in Haiku.

Requirements - Exported Functionality in BeOS

From user-level, BeOS exports the following API as found in the storage/NodeMonitor.h header file:

	status_t watch_node(const node_ref *node, 
uint32 flags,
BMessenger target);

status_t watch_node(const node_ref *node,
uint32 flags,
const BHandler *handler,
const BLooper *looper = NULL);

status_t stop_watching(BMessenger target);

status_t stop_watching(const BHandler *handler,
const BLooper *looper = NULL);

The kernel also exports two other functions to be used from file system add-ons that causes the kernel to send out notification messages:

	int notify_listener(int op, nspace_id nsid,
vnode_id vnida, vnode_id vnidb,
vnode_id vnidc, const char *name);

int send_notification(port_id port, long token,
ulong what, long op, nspace_id nsida,
nspace_id nsidb, vnode_id vnida,
vnode_id vnidb, vnode_id vnidc,
const char *name);

The latter is only used for live query updates, but is obviously called by the former. The port/token pair identify a unique BLooper/BHandler pair, and is used internally to address those high-level objects from the kernel.

When a file system calls the notify_listener() function, it will have a look if there are monitors for that node which meet the specified constraints - and it will call send_notification() for every single message to be send.

Each of the parameters vnida - vnidc has a dedicated meaning:

  • vnida: the parent directory of the "main" node
  • vnidb: the target parent directory for a move
  • vnidc: the node that has triggered the notification to be send

The flags parameter in watch_node() understands the following constants:

  • B_STOP_WATCHING
    watch_node() will stop to watch the specified node.
  • B_WATCH_NAME
    name changes are notified through a B_ENTRY_MOVED opcode.
  • B_WATCH_STAT
    changes to the node's stat structure are notified with a B_STAT_CHANGED code.
  • B_WATCH_ATTR
    attribute changes will cause a B_ATTR_CHANGED to be send.
  • B_WATCH_DIRECTORY

    notifies on changes made to the specified directory, i.e. B_ENTRY_REMOVED, B_ENTRY_CREATED
  • B_WATCH_ALL
    is a short-hand for the flags above.
  • B_WATCH_MOUNT
    causes B_DEVICE_MOUNTED and B_DEVICE_UNMOUNTED to be send.

Node monitors are maintained per team - every team can have up to 4096 monitors, although there exists a private kernel call to raise this limit (for example, Tracker is using it intensively).

The kernel is able to send the BMessages directly to the specified BLooper and BHandler; it achieves this using the application kit's token mechanism. The message is constructed manually in the kernel, it doesn't use any application kit services.

Meeting the Requirements in an Optimal Way - Implementation in Haiku

If you assume that every file operation could trigger a notification message to be send, it's clear that the node monitoring system must be optimized for sending messages. For every call to notify_listener(), the kernel must check if there are any monitors for the node that was updated.

Those monitors are put into a hash table which has the device number and the vnode ID as keys. Each of the monitors maintains a list of listeners which specify which port/token pair should be notified for what change. Since the vnodes are created and deleted as needed by the kernel, the node monitor is maintained independently from them; a simple pointer from a vnode to its monitor is not possible.

The main structures that are involved in providing the node monitoring functionality look like this:

	struct monitor_listener {
monitor_listener *next;
monitor_listener *prev;
list_link monitor_link;
port_id port;
int32 token;
uint32 flags;
node_monitor *monitor;
};

struct node_monitor {
node_monitor *next;
mount_id device;
vnode_id node;
struct list listeners;
};

The relevant part of the I/O context structure is this:

	struct io_context {
...
struct list node_monitors;
uint32 num_monitors;
uint32 max_monitors;
};

If you call watch_node() on a file with a flags parameter unequal to B_STOP_WATCHING, the following will happen in the node monitor:

  1. The add_node_monitor() function does a hash lookup for the device/vnode pair. If there is no node_monitor yet for this pair, a new one will be created.
  2. The list of listeners is scanned for the provided port/token pair (the BLooper/BHandler pointer will already be translated in user-space), and the new flag is or'd to the old field, or a new monitor_listener is created if necessary - in the latter case, the team's node monitor counter is incremented.

If it's called with B_STOP_WATCHING defined, the reverse operation take effect, and the monitor field is used to see if this monitor doesn't have any listeners anymore, in which case it will be removed.

Note the presence of the max_monitors - there is no hard limit the kernel exposes to userland applications; the listeners are maintained in a doubly-linked list.

If a team is shut down, all listeners from its I/O context will be removed - since every listener stores a pointer to its monitor, determining the monitors that can be removed because of this operation is very cheap.

The notify_listener() also only does a hash lookup for the device/node pair it got from the file system, and sends out as many notifications as specified by the listeners of the monitor that belong to that node.

If a node is deleted from the disk, the corresponding node_monitor and its listeners will be removed as well, to prevent watching a new file that accidently happen to have the same device/node pair (as is possible with BFS, for example).

Differences Between Both Implementations

Although the aim was to create a completely compatible monitoring implementation, there are some notable differences between the two.

BeOS reserves a certain number of slots for calls to watch_node() - each call to that function will use one slot, even if you call it twice for the same node. Haiku, however, will always use one slot per node - you could call watch_node() several times, but you would waste only one slot.

While this is an implementational detail, it also causes a change in behaviour for applications; in BeOS, applications will get one message for every watch_node() call, in Haiku, you'll get only one message per node. If an application relies on this strange behaviour of the BeOS kernel, it will no longer work correctly.

The other difference is that Haiku exports its node monitoring functionality to kernel modules as well, and provides an extra plain C API for them to use.

And Beyond?

The current implementation directly iterates over all listeners and sends out notifications as required synchronously in the context of the thread that triggered the notification to be sent.

If a node monitor needs to send out several messages, this could theoretically greatly decrease file system performance. To optimize for this case, the required data of the notification could be put into a queue and be sent by a dedicated worker thread. Since this requires an additional copy operation and a reserved address space for this queue, this optimization could be more expensive than the current implementation, depending on the usage pattern of the node monitoring mechanism.

With BFS, it would be possible to introduce the possibility to automatically watch all files in a specified directory. While this would be very convenient at application level, it comes with several disadvantages:

  1. This feature might not be easily accomplishable for many file systems; a file system must be able to retrieve a node by ID only - it might not be feasible to find out about the parent directory for many file systems.
  2. Although it could potentially safe node monitors, it might cause the kernel to send out a lot more messages to the application than it needs. With the restriction the kernel imposes to the number of watched nodes for a team, the application's designer might try to be much stricter with the number of monitors his application will consume.

While 1) might be a real show stopper, 2) is almost invalidated because of Tracker's usage of node monitors; it consumes a monitor for every entry it displays, which might be several thousands. Implementing this feature would not only greatly speed up maintaining this massive need of monitors, and cut down memory usage, but also ease the implementation at application level.

Even 1) could be solved if the kernel could query a file system if it can support this particular feature; it could then automatically monitor all files in that directory without adding complexity to the application using this feature. Of course, the effort to provide this functionality is much larger then - but for applications like Tracker, the complexity would be removed from the application without extra cost.

However, at the time of writing this, none of the discussed feature extensions have been implemented for the currently developed version R1 of Haiku.