threads.c

radius server在linux下的源码

/*
 * threads.c	request threading support
 *
 * Version:	$Id: threads.c,v 1.77.2.2 2005/04/11 23:45:22 aland Exp $
 *
 *   This program is free software; you can redistribute it and/or modify
 *   it under the terms of the GNU General Public License as published by
 *   the Free Software Foundation; either version 2 of the License, or
 *   (at your option) any later version.
 *
 *   This program is distributed in the hope that it will be useful,
 *   but WITHOUT ANY WARRANTY; without even the implied warranty of
 *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 *   GNU General Public License for more details.
 *
 *   You should have received a copy of the GNU General Public License
 *   along with this program; if not, write to the Free Software
 *   Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
 *
 * Copyright 2000  The FreeRADIUS server project
 * Copyright 2000  Alan DeKok <[e-mail]>
 */

#include "libradius.h"

#ifdef HAVE_PTHREAD_H

#include <stdlib.h>
#include <string.h>
#include <semaphore.h>
#include <signal.h>

#ifdef HAVE_SYS_WAIT_H
#include <sys/wait.h>
#endif

#include "radiusd.h"
#include "rad_assert.h"
#include "conffile.h"

static const char rcsid[] =
"$Id: threads.c,v 1.77.2.2 2005/04/11 23:45:22 aland Exp $";

#define SEMAPHORE_LOCKED	(0)
#define SEMAPHORE_UNLOCKED	(1)

#define THREAD_RUNNING		(1)
#define THREAD_CANCELLED	(2)
#define THREAD_EXITED		(3)

/*
 *  A data structure which contains the information about
 *  the current thread.
 *
 *  pthread_id     pthread id
 *  thread_num     server thread number, 1...number of threads
 *  semaphore     used to block the thread until a request comes in
 *  status        is the thread running or exited?
 *  request_count the number of requests that this thread has handled
 *  timestamp     when the thread started executing.
 */
typedef struct THREAD_HANDLE {
	struct THREAD_HANDLE *prev;
	struct THREAD_HANDLE *next;
	pthread_t            pthread_id;
	int                  thread_num;
	int                  status;
	unsigned int         request_count;
	time_t               timestamp;
	REQUEST		     *request;
} THREAD_HANDLE;

/*
 *	For the request queue.
 */
typedef struct request_queue_t {
	REQUEST	    	  *request;
	RAD_REQUEST_FUNP  fun;
} request_queue_t;


#define MAX_WAITERS (256)
/*
 *	A data structure to manage the thread pool.  There's no real
 *	need for a data structure, but it makes things conceptually
 *	easier.
 */
typedef struct THREAD_POOL {
	THREAD_HANDLE *head;
	THREAD_HANDLE *tail;

	int total_threads;
	int max_thread_num;
	int start_threads;
	int max_threads;
	int min_spare_threads;
	int max_spare_threads;
	unsigned int max_requests_per_thread;
	unsigned long request_count;
	time_t time_last_spawned;
	int cleanup_delay;

	/*
	 *	If threaded, we have to pay more attention to
	 *	child PID's when we fork...
	 */
	pthread_mutex_t	wait_mutex;
	int		wait_head;
	int		wait_tail;
	pid_t		wait[MAX_WAITERS];

	/*
	 *	All threads wait on this semaphore, for requests
	 *	to enter the queue.
	 */
	sem_t		semaphore;

	/*
	 *	To ensure only one thread at a time touches the queue.
	 */
	pthread_mutex_t	mutex;

	int		active_threads;
	int		queue_head; /* first filled entry */
	int		queue_tail; /* first empty entry */
	int		queue_size;
	request_queue_t *queue;
} THREAD_POOL;

static THREAD_POOL thread_pool;
static int pool_initialized = FALSE;


/*
 *	A mapping of configuration file names to internal integers
 */
static const CONF_PARSER thread_config[] = {
	{ "start_servers",           PW_TYPE_INTEGER, 0, &thread_pool.start_threads,           "5" },
	{ "max_servers",             PW_TYPE_INTEGER, 0, &thread_pool.max_threads,             "32" },
	{ "min_spare_servers",       PW_TYPE_INTEGER, 0, &thread_pool.min_spare_threads,       "3" },
	{ "max_spare_servers",       PW_TYPE_INTEGER, 0, &thread_pool.max_spare_threads,       "10" },
	{ "max_requests_per_server", PW_TYPE_INTEGER, 0, &thread_pool.max_requests_per_thread, "0" },
	{ "cleanup_delay",           PW_TYPE_INTEGER, 0, &thread_pool.cleanup_delay,           "5" },
	{ NULL, -1, 0, NULL, NULL }
};


/*
 *	We don't want to catch SIGCHLD for a host of reasons.
 *
 *	- exec_wait means that someone, somewhere, somewhen, will
 *	call waitpid(), and catch the child.
 *
 *	- SIGCHLD is delivered to a random thread, not the one that
 *	forked.
 *
 *	- if another thread catches the child, we have to coordinate
 *	with the thread doing the waiting.
 *
 *	- if we don't waitpid() for non-wait children, they'll be zombies,
 *	and will hang around forever.
 *
 */
static void reap_children(void)
{
	if (thread_pool.wait_head != thread_pool.wait_tail) {
		int num;
		
		pthread_mutex_lock(&thread_pool.wait_mutex);
		for (num = ((thread_pool.wait_tail + MAX_WAITERS) - thread_pool.wait_head) % MAX_WAITERS;
		     num != 0;
		     num--) {
			pid_t pid = thread_pool.wait[thread_pool.wait_head];
			
			thread_pool.wait_head++;
			thread_pool.wait_head %= MAX_WAITERS;
			
			/*
			 *	Child is still alive: move it to the tail.
			 */
			if (waitpid(pid, NULL, WNOHANG) == 0) {
				if (((thread_pool.wait_tail + 1) % MAX_WAITERS)
				    == thread_pool.wait_head) {
					rad_assert(0 == 1);
				}

				thread_pool.wait[thread_pool.wait_tail] = pid;
				thread_pool.wait_tail++;
				thread_pool.wait_tail %= MAX_WAITERS;
			} /* else no child, or was already reaped */
		}
		pthread_mutex_unlock(&thread_pool.wait_mutex);
	}
}


/*
 *	Add a request to the list of waiting requests.
 *	This function gets called ONLY from the main handler thread...
 *
 *	This function should never fail.
 */
static void request_enqueue(REQUEST *request, RAD_REQUEST_FUNP fun)
{
	int num_entries;

	pthread_mutex_lock(&thread_pool.mutex);

	thread_pool.request_count++;

	/*
	 *	If the queue is empty, re-set the indices to zero,
	 *	for no particular reason...
	 */
	if ((thread_pool.queue_head == thread_pool.queue_tail) &&
	    (thread_pool.queue_head != 0)) {
		thread_pool.queue_head = thread_pool.queue_tail = 0;
	}

	/*
	 *	If the queue is full, die.
	 *
	 *	The math is to take into account the fact that it's a
	 *	circular queue.
	 */
	num_entries = ((thread_pool.queue_tail + thread_pool.queue_size) -
		       thread_pool.queue_head) % thread_pool.queue_size;
	if (num_entries == (thread_pool.queue_size - 1)) {
		request_queue_t *new_queue;

		/*
		 *	If the queue becomes larger than 65536,
		 *	there's a serious problem.
		 */
		if (thread_pool.queue_size >= 65536) {
			pthread_mutex_unlock(&thread_pool.mutex);

			/*
			 *	Mark the request as done.
			 */
			radlog(L_ERR|L_CONS, "!!! ERROR !!! The server is blocked: discarding new request %d", request->number);
			request->finished = TRUE;
			return;
		}

		/*
		 *	Malloc a new queue, doubled in size, copy the
		 *	data from the current queue over to it, zero
		 *	out the second half of the queue, free the old
		 *	one, and replace thread_pool.queue with the
		 *	new one.
		 */
		new_queue = rad_malloc(sizeof(*new_queue) * thread_pool.queue_size * 2);
		memcpy(new_queue, thread_pool.queue,
		       sizeof(*new_queue) * thread_pool.queue_size);
		memset(new_queue + sizeof(*new_queue) * thread_pool.queue_size,
		       0, sizeof(*new_queue) * thread_pool.queue_size);

		free(thread_pool.queue);
		thread_pool.queue = new_queue;
		thread_pool.queue_size *= 2;
	}

	/*
	 *	Add the data to the queue tail, increment the tail,
	 *	and signal the semaphore that there's another request
	 *	in the queue.
	 */
	thread_pool.queue[thread_pool.queue_tail].request = request;
	thread_pool.queue[thread_pool.queue_tail].fun = fun;
	thread_pool.queue_tail++;
	thread_pool.queue_tail &= (thread_pool.queue_size - 1);

	pthread_mutex_unlock(&thread_pool.mutex);

	/*
	 *	There's one more request in the queue.
	 *
	 *	Note that we're not touching the queue any more, so
	 *	the semaphore post is outside of the mutex.  This also
	 *	means that when the thread wakes up and tries to lock
	 *	the mutex, it will be unlocked, and there won't be
	 *	contention.
	 */

	sem_post(&thread_pool.semaphore);

	return;
}

/*
 *	Remove a request from the queue.
 */
static void request_dequeue(REQUEST **request, RAD_REQUEST_FUNP *fun)
{
	reap_children();

	pthread_mutex_lock(&thread_pool.mutex);

	/*
	 *	Head & tail are the same.  There's nothing in
	 *	the queue.
	 */
	if (thread_pool.queue_head == thread_pool.queue_tail) {
		pthread_mutex_unlock(&thread_pool.mutex);
		*request = NULL;
		*fun = NULL;
		return;
	}

	*request = thread_pool.queue[thread_pool.queue_head].request;
	*fun = thread_pool.queue[thread_pool.queue_head].fun;

	rad_assert(*request != NULL);
	rad_assert((*request)->magic == REQUEST_MAGIC);
	rad_assert(*fun != NULL);

	thread_pool.queue_head++;
	thread_pool.queue_head &= (thread_pool.queue_size - 1);

	/*
	 *	FIXME: Check the request timestamp.  If it's more than
	 *	"clean_delay" seconds old, then discard the request,
	 *	log an error, and try to de-queue another request.
	 *
	 *	The main clean-up code won't delete the request from
	 *	the request list, because it's not marked "finished"
	 */

	/*
	 *	The thread is currently processing a request.
	 */
	thread_pool.active_threads++;

	pthread_mutex_unlock(&thread_pool.mutex);

	/*
	 *	If the request is currently being processed, then that
	 *	MAY be OK, if it's a proxy reply.  In that case,
	 *	sending the packet may result in a reply being
	 *	received before that thread clears the child_pid.
	 *
	 *	In that case, we busy-wait for the request to be free.
	 *
	 *	We COULD push it onto the queue and try to grab
	 *	another request, but what if this is the only request?
	 *	What if there are multiple such packets with race
	 *	conditions?  We don't want to thrash the queue...
	 *
	 *	This busy-wait is less than optimal, but it's simple,
	 *	fail-safe, and it works.
	 */
	if ((*request)->child_pid != NO_SUCH_CHILD_PID) {
		int count, ok;
		struct timeval tv;
#ifdef HAVE_PTHREAD_SIGMASK
		sigset_t set, old_set;

		/*
		 *	Block a large number of signals which could
		 *	cause the select to return EINTR
		 */
		sigemptyset(&set);
		sigaddset(&set, SIGPIPE);
		sigaddset(&set, SIGCONT);
		sigaddset(&set, SIGSTOP);
		sigaddset(&set, SIGCHLD);
		pthread_sigmask(SIG_BLOCK, &set, &old_set);
#endif

		rad_assert((*request)->proxy_reply != NULL);

		ok = FALSE;

		/*
		 *	Sleep for 100 milliseconds.  If the other thread
		 *	doesn't get serviced in this time, to clear
		 *	the "child_pid" entry, then the server is too
		 *	busy, so we die.
		 */
		for (count = 0; count < 10; count++) {
			tv.tv_sec = 0;
			tv.tv_usec = 10000; /* sleep for 10 milliseconds */

			/*
			 *	Portable sleep that's thread-safe.
			 *
			 *	Don't worry about interrupts, as they're
			 *	blocked above.
			 */
			select(0, NULL, NULL, NULL, &tv);
			if ((*request)->child_pid == NO_SUCH_CHILD_PID) {
				ok = TRUE;
				break;
			}
		}

#ifdef HAVE_PTHREAD_SIGMASK
		/*
		 *	Restore the original thread signal mask.
		 */
		pthread_sigmask(SIG_SETMASK, &old_set, NULL);
#endif

		if (!ok) {
			radlog(L_ERR, "FATAL!  Server is too busy to process requests");
			exit(1);
		}
	}

	return;
}


/*
 *	The main thread handler for requests.
 *
 *	Wait on the semaphore until we have it, and process the request.
 */
static void *request_handler_thread(void *arg)
{
	RAD_REQUEST_FUNP  fun;
	THREAD_HANDLE	  *self = (THREAD_HANDLE *) arg;
#ifdef HAVE_PTHREAD_SIGMASK
	sigset_t set;

	/*
	 *	Block SIGHUP handling for the child threads.
	 *
	 *	This ensures that only the main server thread will
	 *	process HUP signals.
	 *
	 *	If we don't have sigprocmask, then it shouldn't be
	 *	a problem, either, as the sig_hup handler should check
	 *	for this condition.
	 */
	sigemptyset(&set);
	sigaddset(&set, SIGHUP);
	sigaddset(&set, SIGINT);
	sigaddset(&set, SIGQUIT);
	sigaddset(&set, SIGTERM);
	pthread_sigmask(SIG_BLOCK, &set, NULL);
#endif

	/*
	 *	Loop forever, until told to exit.
	 */
	do {
		/*
		 *	Wait to be signalled.
		 */
		DEBUG2("Thread %d waiting to be assigned a request",
		       self->thread_num);
	re_wait:
		if (sem_wait(&thread_pool.semaphore) != 0) {
			/*
			 *	Interrupted system call.  Go back to
			 *	waiting, but DON'T print out any more
			 *	text.
			 */
			if (errno == EINTR) {
				DEBUG2("Re-wait %d", self->thread_num);
				goto re_wait;
			}
			radlog(L_ERR, "Thread %d failed waiting for semaphore: %s: Exiting\n",
			       self->thread_num, strerror(errno));
			break;
		}

		DEBUG2("Thread %d got semaphore", self->thread_num);

		/*
		 *	Try to grab a request from the queue.
		 *
		 *	It may be empty, in which case we fail
		 *	gracefully.
		 */
		request_dequeue(&self->request, &fun);
		if (!self->request) continue;

		self->request->child_pid = self->pthread_id;
		self->request_count++;

		DEBUG2("Thread %d handling request %d, (%d handled so far)",
		       self->thread_num, self->request->number,
		       self->request_count);

		/*
		 *	Respond, and reset request->child_pid
		 */
		rad_respond(self->request, fun);
		self->request = NULL;

		/*
		 *	Update the active threads.
		 */
		pthread_mutex_lock(&thread_pool.mutex);
		rad_assert(thread_pool.active_threads > 0);
		thread_pool.active_threads--;
		pthread_mutex_unlock(&thread_pool.mutex);
	} while (self->status != THREAD_CANCELLED);

	DEBUG2("Thread %d exiting...", self->thread_num);

	/*
	 *  Do this as the LAST thing before exiting.