Most of the time, an application wants to perform some amount of data buffering in addition to just responding to events. When we want to write data, for example, the usual pattern runs something like:

This buffered IO pattern is common enough that Libevent provides a generic mechanism for it. A "bufferevent" consists of an underlying transport (like a socket), a read buffer, and a write buffer. Instead of regular events, which give callbacks when the underlying transport is ready to be read or written, a bufferevent invokes its user-supplied callbacks when it has read or written enough data.

There are multiple types of bufferevent that all share a common interface. As of this writing, the following types exist:

As of Libevent 2.0.2-alpha, the bufferevents interfaces here are still not fully orthogonal across all bufferevent types. In other words, not every interface described below will work on all bufferevent types. The Libevent developers intend to correct this in future versions.

Bufferevents currently only work for stream-oriented protocols like TCP. There may in the future be support for datagram-oriented protocols like UDP.

All of the functions and types in this section are declared in event2/bufferevent.h. Functions specifically related to evbuffers are declared in event2/buffer.h; see the next chapter for information on those.

Every bufferevent has an input buffer and an output buffer. These are of type "struct evbuffer". When you have data to write on a bufferevent, you add it to the output buffer; when a bufferevent has data for you to read, you drain it from the input buffer.

The evbuffer interface supports many operations; we discuss them in a later section.

Every bufferevent has two data-related callbacks: a read callback and a write callback. By default, the read callback is called whenever any data is read from the underlying transport, and the write callback is called whenever enough data from the output buffer is emptied to the underlying transport. You can override the behavior of these functions by adjusting the read and write "watermarks" of the bufferevent.

Every bufferevent has four watermarks:

A bufferevent also has an "error" or "event" callback that gets invoked to tell the application about non-data-oriented events, like when a connection is closed or an error occurs. The following event flags are defined:

(The above event names are new in Libevent 2.0.2-alpha.)

By default, a bufferevent callbacks are executed immediately when the corresponding condition happens. (This is true of evbuffer callbacks too; we’ll get to those later.) This immediate invocation can make trouble when dependencies get complex. For example, suppose that there is a callback that moves data into evbuffer A when it grows empty, and another callback that processes data out of evbuffer A when it grows full. Since these calls are all happening on the stack, you might risk a stack overflow if the dependency grows nasty enough.

To solve this, you can tell a bufferevent (or an evbuffer) that its callbacks should be deferred. When the conditions are met for a deferred callback, rather than invoking it immediately, it is queued as part of the event_loop() call, and invoked after the regular events' callbacks.

(Deferred callbacks were introduced in Libevent 2.0.1-alpha.)

You can use one or more flags when creating a bufferevent to alter its behavior. Recognized flags are:

(Libevent 2.0.5-beta introduced BEV_OPT_UNLOCK_CALLBACKS. The other options above were new in Libevent 2.0.1-alpha.)

The simplest bufferevents to work with is the socket-based type. A socket-based bufferevent uses Libevent’s underlying event mechanism to detect when an underlying network socket is ready for read and/or write operations, and uses underlying network calls (like readv, writev, WSASend, or WSARecv) to transmit and receive data.

You can create a socket-based bufferevent using bufferevent_socket_new():

The base is an event_base, and options is a bitmask of bufferevent options (BEV_OPT_CLOSE_ON_FREE, etc). The fd argument is an optional file descriptor for a socket. You can set fd to -1 if you want to set the file descriptor later.

This function returns a bufferevent on success, and NULL on failure.

The bufferevent_socket_new() function was introduced in Libevent 2.0.1-alpha.

If the bufferevent’s socket is not yet connected, you can launch a new connection.

The address and addrlen arguments are as for the standard call connect(). If the bufferevent does not already have a socket set, calling this function allocates a new stream socket for it, and makes it nonblocking.

If the bufferevent does have a socket already, calling bufferevent_socket_connect() tells Libevent that the socket is not connected, and no reads or writes should be done on the socket until the connect operation has succeeded.

It is okay to add data to the output buffer before the connect is done.

This function returns 0 if the connect was successfully launched, and -1 if an error occurred.

The bufferevent_base_connect() function was introduced in Libevent-2.0.2-alpha. Before then, you had to manually call connect() on your socket yourself, and when the connection was complete, the bufferevent would report it as a write.

Note that you only get a BEV_EVENT_CONNECTED event if you launch the connect() attempt using bufferevent_socket_connect(). If you call connect() on your own, the connection gets reported as a write.

If you want to call connect() yourself, but still get receive a BEV_EVENT_CONNECTED event when the connection succeeds, call bufferevent_socket_connect(bev, NULL, 0) after connect() returns -1 with errno equal to EAGAIN or EINPROGRESS.

This function was introduced in Libevent 2.0.2-alpha.

Quite often, you’d like to combine resolving a hostname and connecting to it into a single operation. There’s an interface for that:

This function resolves the DNS name hostname, looking for addresses of type family. (Allowable family types are AF_INET, AF_INET6, and AF_UNSPEC.) If the name resolution fails, it invokes the event callback with an error event. If it succeeds, it launches a connection attempt just as bufferevent_connect would.

The dns_base argument is optional. If it is NULL, then Libevent blocks while waiting for the name lookup to finish, which usually isn’t what you want. If it is provided, then Libevent uses it to look up the hostname asynchronously. See chapter R9 for more info on DNS.

As with bufferevent_socket_connect(), this function tells Libevent that any existing socket on the bufferevent is not connected, and no reads or writes should be done on the socket until the resolve is finished and the connect operation has succeeded.

If an error occurs, it might be a DNS hostname lookup error. You can find out what the most recent error was by calling bufferevent_socket_get_dns_error(). If the returned error code is 0, no DNS error was detected.

The bufferevent_socket_connect_hostname() function was new in Libevent 2.0.3-alpha; bufferevent_socket_get_dns_error() was new in 2.0.5-beta.

The functions in this section work with multiple bufferevent implementations.

This function frees a bufferevent. Bufferevents are internally reference-counted, so if the bufferevent has pending deferred callbacks when you free it, it won’t be deleted until the callbacks are done.

The bufferevent_free() function does, however, try to free the bufferevent as soon as possible. If there is pending data to write on the bufferevent, it probably won’t be flushed before the bufferevent is freed.

If the BEV_OPT_CLOSE_ON_FREE flag was set, and this bufferevent has a socket or underlying bufferevent associated with it as its transport, that transport is closed when you free the bufferevent.

This function was introduced in Libevent 0.8.

The bufferevent_setcb() function changes one or more of the callbacks of a bufferevent. The readcb, writecb, and eventcb functions are called (respectively) when enough data is read, when enough data is written, or when an event occurs. The first argument of each is the bufferevent that has had the event happen. The last argument is the value provided by the user in the cbarg parameter of bufferevent_callcb(): You can use this to pass data to your callbacks. The events argument of the event callback is a bitmask of event flags: see "callbacks and watermarks" above.

You can disable a callback by passing NULL instead of the callback function. Note all the callback functions on a bufferevent share a single cbarg value, so changing it will affect all of them.

You can retrieve the currently set callbacks for a bufferevent by passing pointers to bufferevent_getcb(), which sets *readcb_ptr to the current read callback, *writecb_ptr to the current write callback, *eventcb_ptr to the current event callback, and *cbarg_ptr to the current callback argument field. Any of these pointers set to NULL will be ignored.

The bufferevent_setcb() function was introduced in Libevent 1.4.4. The type names "bufferevent_data_cb" and "bufferevent_event_cb" were new in Libevent 2.0.2-alpha. The bufferevent_getcb() function was added in 2.1.1-alpha.

You can enable or disable the events EV_READ, EV_WRITE, or EV_READ|EV_WRITE on a bufferevent. When reading or writing is not enabled, the bufferevent will not try to read or write data.

There is no need to disable writing when the output buffer is empty: the bufferevent automatically stops writing, and restarts again when there is data to write.

Similarly, there is no need to disable reading when the input buffer is up to its high-water mark: the bufferevent automatically stops reading, and restarts again when there is space to read.

By default, a newly created bufferevent has writing enabled, but not reading.

You can call bufferevent_get_enabled() to see which events are currently enabled on the bufferevent.

These functions were introduced in Libevent 0.8, except for bufferevent_get_enabled(), which was introduced in version 2.0.3-alpha.

The bufferevent_setwatermark() function adjusts the read watermarks, the write watermarks, or both, of a single bufferevent. (If EV_READ is set in the events field, the read watermarks are adjusted. If EV_WRITE is set in the events field, the write watermarks are adjusted.)

A high-water mark of 0 is equivalent to "unlimited".

This function was first exposed in Libevent 1.4.4.

Reading and writing data from the network does you no good if you can’t look at it. Bufferevents give you these methods to give them data to write, and to get the data to read:

These two functions are very powerful fundamental: they return the input and output buffers respectively. For full information on all the operations you can perform on an evbuffer type, see the next chapter.

Note that the application may only remove (not add) data on the input buffer, and may only add (not remove) data from the output buffer.

If writing on the bufferevent was stalled because of too little data (or if reading was stalled because of too much), then adding data to the output buffer (or removing data from the input buffer) will automatically restart it.

These functions were introduced in Libevent 2.0.1-alpha.

These functions add data to a bufferevent’s output buffer. Calling bufferevent_write() adds size bytes from the memory at data to the end of the output buffer. Calling bufferevent_write_buffer() removes the entire contents of buf and puts them at the end of the output buffer. Both return 0 if successful, or -1 if an error occurred.

These functions have existed since Libevent 0.8.

These functions remove data from a bufferevent’s input buffer. The bufferevent_read() function removes up to size bytes from the input buffer, storing them into the memory at data. It returns the number of bytes actually removed. The bufferevent_read_buffer() function drains the entire contents of the input buffer and places them into buf; it returns 0 on success and -1 on failure.

Note that with bufferevent_read(), the memory chunk at data must actually have enough space to hold size bytes of data.

The bufferevent_read() function has existed since Libevent 0.8; bufferevent_read_buffer() was introduced in Libevent 2.0.1-alpha.

As with other events, you can have a timeout get invoked if a certain amount of time passes without any data having been successfully written or read by a bufferevent.

Setting a timeout to NULL is supposed to remove it; however before Libevent 2.1.2-alpha this wouldn’t work with all event types. (As a workaround for older versions, you can try setting the timeout to a multi-day interval and/or having your eventcb function ignore BEV_TIMEOUT events when you don’t want them.)

The read timeout will trigger if the bufferevent waits at least timeout_read seconds while trying to read read. The write timeout will trigger if the bufferevent waits at least timeout_write seconds while trying to write data.

Note that the timeouts only count when the bufferevent would like to read or write. In other words, the read timeout is not enabled if reading is disabled on the bufferevent, or if the input buffer is full (at its high-water mark). Similarly, the write timeout is not enabled if if writing is disabled, or if there is no data to write.

When a read or write timeout occurs, the corresponding read or write operation becomes disabled on the bufferevent. The event callback is then invoked with either BEV_EVENT_TIMEOUT|BEV_EVENT_READING or BEV_EVENT_TIMEOUT|BEV_EVENT_WRITING.

This functions has existed since Libevent 2.0.1-alpha. It didn’t behave consistently across bufferevent types until Libevent 2.0.4-alpha.

Flushing a bufferevent tells the bufferevent to force as many bytes as possible to be read to or written from the underlying transport, ignoring other restrictions that might otherwise keep them from being written. Its detailed function depends on the type of the bufferevent.

The iotype argument should be EV_READ, EV_WRITE, or EV_READ|EV_WRITE to indicate whether bytes being read, written, or both should be processed. The state argument may be one of BEV_NORMAL, BEV_FLUSH, or BEV_FINISHED. BEV_FINISHED indicates that the other side should be told that no more data will be sent; the distinction between BEV_NORMAL and BEV_FLUSH depends on the type of the bufferevent.

The bufferevent_flush() function returns -1 on failure, 0 if no data was flushed, or 1 if some data was flushed.

Currently (as of Libevent 2.0.5-beta), bufferevent_flush() is only implemented for some bufferevent types. In particular, socket-based bufferevents don’t have it.

These bufferevent functions are not supported on all bufferevent types.

This function adjusts the priority of the events used to implement bufev to pri. See event_priority_set() for more information on priorities.

This function returns 0 on success, and -1 on failure. It works on socket-based bufferevents only.

The bufferevent_priority_set() function was introduced in Libevent 1.0; bufferevent_get_priority() didn’t appear until Libevent 2.1.2-alpha.

These functions set or return the file descriptor for a fd-based event. Only socket-based bufferevents support setfd(). Both return -1 on failure; setfd() returns 0 on success.

The bufferevent_setfd() function was introduced in Libevent 1.4.4; the bufferevent_getfd() function was introduced in Libevent 2.0.2-alpha.

This function returns the event_base of a bufferevent. It was introduced in 2.0.9-rc.

This function returns the bufferevent that another bufferevent is using as a transport, if any. For information on when this situation would occur, see notes on filtering bufferevents.

This function was introduced in Libevent 2.0.2-alpha.

As with evbuffers, sometimes you want to ensure that a number of operations on a bufferevent are all performed atomically. Libevent exposes functions that you can use to manually lock and unlock a bufferevent.

Note that locking a bufferevent has no effect if the bufferevent was not given the BEV_OPT_THREADSAFE thread on creation, or if Libevent’s threading support wasn’t activated.

Locking the bufferevent with this function will lock its associated evbuffers as well. These functions are recursive: it is safe to lock a bufferevent for which you already hold the lock. You must, of course, call unlock once for every time that you locked the bufferevent.

These functions were introduced in Libevent 2.0.6-rc.

The bufferevent backend code underwent substantial revision between Libevent 1.4 and Libevent 2.0. In the old interface, it was sometimes normal to build with access to the internals of the struct bufferevent, and to use macros that relied on this access.

To make matters confusing, the old code sometimes used names for bufferevent functionality that were prefixed with "evbuffer".

Here’s a brief guideline of what things used to be called before Libevent 2.0:

The old functions were defined in event.h, not in event2/bufferevent.h.

If you still need access to the internals of the common parts of the bufferevent struct, you can include event2/bufferevent_struct.h. We recommend against it: the contents of struct bufferevent WILL change between versions of Libevent. The macros and names in this section are available if you include event2/bufferevent_compat.h.

The interface to set up a bufferevent differed in older versions:

The bufferevent_new() function creates a socket bufferevent only, and does so on the deprecated "default" event_base. Calling bufferevent_base_set adjusts the event_base of a socket bufferevent only.

Instead of setting timeouts as struct timeval, they were set as numbers of seconds:

Finally, note that the underlying evbuffer implementation for Libevent versions before 2.0 was pretty inefficient, to the point where using bufferevents for high-performance apps was kind of questionable.