thread_httpd.pl -- Threaded HTTP server

Most code doesn't need to use this directly; instead use library(http/http_server), which combines this library with the typical HTTP libraries that most servers need.

This library defines the HTTP server frontend of choice for SWI-Prolog. It is based on the multi-threading capabilities of SWI-Prolog and thus exploits multiple cores to serve requests concurrently. The server scales well and can cooperate with library(thread_pool) to control the number of concurrent requests of a given type. For example, it can be configured to handle 200 file download requests concurrently, 2 requests that potentially uses a lot of memory and 8 requests that use a lot of CPU resources.

On Unix systems, this library can be combined with library(http/http_unix_daemon) to realise a proper Unix service process that creates a web server at port 80, runs under a specific account, optionally detaches from the controlling terminal, etc.

Combined with library(http/http_ssl_plugin) from the SSL package, this library can be used to create an HTTPS server. See <plbase>/doc/packages/examples/ssl/https for an example server using a self-signed SSL certificate.

http_server(:Goal, :Options) is det

Create a server at Port that calls Goal for each parsed request. Options provide a list of options. Defined options are

port(?Address): Port to bind to. Address is either a port or a term Host:Port. The port may be a variable, causing the system to select a free port. See tcp_bind/2.
unix_socket(+Path): Instead of binding to a TCP port, bind to a Unix Domain Socket at Path.
entry_page(+URI): Affects the message printed while the server is started. Interpreted as a URI relative to the server root.
tcp_socket(+Socket): If provided, use this socket instead of the creating one and binding it to an address. The socket must be bound to an address. Note that this also allows binding an HTTP server to a Unix domain socket (AF_UNIX). See socket_create/2.
workers(+Count): Determine the number of worker threads. Default is 5. This is fine for small scale usage. Public servers typically need a higher number.
timeout(+Seconds): Maximum time of inactivity trying to read the request after a connection has been opened. Default is 60 seconds. See set_stream/1 using the timeout option.
keep_alive_timeout(+Seconds): Time to keep `Keep alive' connections alive. Default is 2 seconds.
stack_limit(+Bytes): Stack limit to use for the workers. The default is inherited from the main thread. If you need to control resource usage you may consider the spawn option of http_handler/3 and library(thread_pool).
silent(Bool): If true (default false), do not print an informational message that the server was started.

A typical initialization for an HTTP server that uses http_dispatch/1 to relay requests to predicates is:

:- use_module(library(http/thread_httpd)).
:- use_module(library(http/http_dispatch)).

start_server(Port) :-
    http_server(http_dispatch, [port(Port)]).

Note that multiple servers can coexist in the same Prolog process. A notable application of this is to have both an HTTP and HTTPS server, where the HTTP server redirects to the HTTPS server for handling sensitive requests.

make_socket(+Address, :OptionsIn, -OptionsOut) is det[private]

Create the HTTP server socket and worker pool queue. OptionsOut is guaranteed to hold the option queue(QueueId).

Arguments:

OptionsIn

- is qualified to allow passing the module-sensitive ssl option argument.

make_addr_atom(+Scheme, +Address, -Atom) is det[private]

Create an atom that identifies the server's queue and thread resources.

create_server(:Goal, +Address, +Options) is det[private]

Create the main server thread that runs accept_server/2 to listen to new requests.

http_current_server(:Goal, ?Port) is nondet

True if Goal is the goal of a server at Port.

deprecated: - Use http_server_property(Port, goal(Goal))

http_server_property(?Port, ?Property) is nondet

True if Property is a property of the HTTP server running at Port. Defined properties are:

goal(:Goal): Goal used to start the server. This is often http_dispatch/1.
scheme(-Scheme): Scheme is one of http or https.
start_time(?Time): Time-stamp when the server was created.

http_workers(?Port, -Workers) is nondet

http_workers(+Port, +Workers:int) is det

Query or set the number of workers for the server at this port. The number of workers is dynamically modified. Setting it to 1 (one) can be used to profile the worker using tprofile/1.

See also: - library(http/http_dyn_workers) implements dynamic management of the worker pool depending on usage.

http_add_worker(+Port, +Options) is det

Add a new worker to the HTTP server for port Port. Options overrule the default queue options. The following additional options are processed:

max_idle_time(+Seconds): The created worker will automatically terminate if there is no new work within Seconds.

http_current_worker(?Port, ?ThreadID) is nondet

True if ThreadID is the identifier of a Prolog thread serving Port. This predicate is motivated to allow for the use of arbitrary interaction with the worker thread for development and statistics.

accept_server(:Goal, +Initiator, +Options)[private]

The goal of a small server-thread accepting new requests and posting them to the queue of workers.

close_server_socket(+Options)[private]

Close the server socket.

close_pending_accepts(+Queue)[private]

http_stop_server(+Port, +Options)

Stop the indicated HTTP server gracefully. First stops all workers, then stops the server.

To be done: - Realise non-graceful stop

http_enough_workers(+Queue, +Why, +Peer) is det

Check that we have enough workers in our queue. If not, call the hook schedule_workers/1 to extend the worker pool. This predicate can be used by accept_hook/2.

http:schedule_workers(+Data:dict) is semidet[multifile]

Hook called if a new connection or a keep-alive connection cannot be scheduled immediately to a worker. Dict contains the following keys:

port:Port: Port number that identifies the server.
reason:Reason: One of accept for a new connection or keep_alive if a worker tries to reschedule itself.
peer:Peer: Identify the other end of the connection
waiting:Size: Number of messages waiting in the queue.
queue:Queue: Message queue used to dispatch accepted messages.

Note that, when called with reason:accept, we are called in the time critical main accept loop. An implementation of this hook shall typically send the event to thread dedicated to dynamic worker-pool management.

See also: - http_add_worker/2 may be used to create (temporary) extra workers.

create_workers(+Options)[private]

Create the pool of HTTP worker-threads. Each worker has the alias http_worker_N.

resize_pool(+Queue, +Workers) is det[private]

Create or destroy workers. If workers are destroyed, the call waits until the desired number of waiters is reached.

http_worker(+Options)[private]

Run HTTP worker main loop. Workers simply wait until they are passed an accepted socket to process a client.

If the message quit(Sender) is read from the queue, the worker stops.

open_client(+Message, +Queue, -Goal, -In, -Out, +Options, -ClientOptions) is semidet[private]

Opens the connection to the client in a worker from the message sent to the queue by accept_server/2.

open_client(+Message, +Goal, -In, -Out, -ClientOptions, +Options) is det[private]

check_keep_alive_connection(+In, +TimeOut, +Peer, +In, +Out) is semidet[private]

Wait for the client for at most TimeOut seconds. Succeed if the client starts a new request within this time. Otherwise close the connection and fail.

done_worker[private]

Called when worker is terminated due to http_workers/2 or a (debugging) exception. In the latter case, recreate_worker/2 creates a new worker.

recreate_worker(+Status, +Queue) is semidet[private]

Deal with the possibility that threads are, during development, killed with abort/0. We recreate the worker to avoid that eventually we run out of workers. If we are aborted due to a halt/0 call, thread_create/3 will raise a permission error.

The first clause deals with the possibility that we cannot write to user_error. This is possible when Prolog is started as a service using some service managers. Would be nice if we could write an error, but where?

thread_httpd:message_level(+Exception, -Level)[multifile]

Determine the message stream used for exceptions that may occur during server_loop/5. Being multifile, clauses can be added by the application to refine error handling. See also message_hook/3 for further programming error handling.

read_remaining_request(+StartBody, +Request) is semidet[private]

If our handler did not read the complete request we must read the remainder if we are dealing with a Keep-alive connection.

http_requeue(+Header)

Re-queue a connection to the worker pool. This deals with processing additional requests on keep-alive connections.

http_process(Message, Queue, +Options)[private]

Handle a single client message on the given stream.

http_close_connection(+Request)

Close connection associated to Request. See also http_requeue/1.

close_connection(+Peer, +In, +Out)[private]

Closes the connection from the server to the client. Errors are currently silently ignored.

http_spawn(:Goal, +Options) is det

Continue this connection on a new thread. A handler may call http_spawn/2 to start a new thread that continues processing the current request using Goal. The original thread returns to the worker pool for processing new requests. Options are passed to thread_create/3, except for:

pool(+Pool): Interfaces to library(thread_pool), starting the thread on the given pool.

If a pool does not exist, this predicate calls the multifile hook create_pool/1 to create it. If this predicate succeeds the operation is retried.

create_pool(+Pool)[private]

Lazy creation of worker-pools for the HTTP server. This predicate calls the hook create_pool/1. If the hook fails it creates a default pool of size 10. This should suffice most typical usecases. Note that we get a permission error if the pool is already created. We can ignore this.

thread_repeat_wait(:Goal) is multi[private]

Acts as repeat, thread_idle(Goal), choosing whether to use a long or short idle time based on the average firing rate.

new_rate_mma(+N, +Resolution, -State) is det[private]

update_rate_mma(!State, -MMA) is det[private]

Implement Modified Moving Average computing the average time between requests as an exponential moving average with alpha=1/N.

Arguments:

Resolution

- is the time resolution in 1/Resolution seconds. All storage is done in integers to avoid the need for stack freezing in nb_setarg/3.