Merge pull request #931 from nchaigne/3.0.x-fb4-rc
[freeradius.git] / src / main / process.c
1 /*
2  *   This program is free software; you can redistribute it and/or modify
3  *   it under the terms of the GNU General Public License as published by
4  *   the Free Software Foundation; either version 2 of the License, or
5  *   (at your option) any later version.
6  *
7  *   This program is distributed in the hope that it will be useful,
8  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
9  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
10  *   GNU General Public License for more details.
11  *
12  *   You should have received a copy of the GNU General Public License
13  *   along with this program; if not, write to the Free Software
14  *   Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
15  */
16
17 /**
18  * $Id$
19  *
20  * @file process.c
21  * @brief Defines the state machines that control how requests are processed.
22  *
23  * @copyright 2012  The FreeRADIUS server project
24  * @copyright 2012  Alan DeKok <aland@deployingradius.com>
25  */
26
27 RCSID("$Id$")
28
29 #include <freeradius-devel/radiusd.h>
30 #include <freeradius-devel/process.h>
31 #include <freeradius-devel/modules.h>
32 #include <freeradius-devel/state.h>
33
34 #include <freeradius-devel/rad_assert.h>
35
36 #ifdef WITH_DETAIL
37 #include <freeradius-devel/detail.h>
38 #endif
39
40 #include <signal.h>
41 #include <fcntl.h>
42
43 #ifdef HAVE_SYS_WAIT_H
44 #       include <sys/wait.h>
45 #endif
46
47 extern pid_t radius_pid;
48 extern fr_cond_t *debug_condition;
49
50 static bool spawn_flag = false;
51 static bool just_started = true;
52 time_t fr_start_time = (time_t)-1;
53 static rbtree_t *pl = NULL;
54 static fr_event_list_t *el = NULL;
55
56 fr_event_list_t *radius_event_list_corral(UNUSED event_corral_t hint) {
57         /* Currently we do not run a second event loop for modules. */
58         return el;
59 }
60
61 static char const *action_codes[] = {
62         "INVALID",
63         "run",
64         "done",
65         "dup",
66         "timer",
67 #ifdef WITH_PROXY
68         "proxy-reply"
69 #endif
70 };
71
72 #ifdef DEBUG_STATE_MACHINE
73 #  define TRACE_STATE_MACHINE \
74 if (debug_flag) do { \
75         struct timeval debug_tv; \
76         gettimeofday(&debug_tv, NULL); \
77         debug_tv.tv_sec -= fr_start_time; \
78         printf("(%u) %d.%06d ********\tSTATE %s action %s live M-%s C-%s\t********\n",\
79                request->number, (int) debug_tv.tv_sec, (int) debug_tv.tv_usec, \
80                __FUNCTION__, action_codes[action], master_state_names[request->master_state], \
81                child_state_names[request->child_state]); \
82 } while (0)
83
84 static char const *master_state_names[REQUEST_MASTER_NUM_STATES] = {
85         "?",
86         "active",
87         "stop-processing",
88         "counted"
89 };
90
91 static char const *child_state_names[REQUEST_CHILD_NUM_STATES] = {
92         "?",
93         "queued",
94         "running",
95         "proxied",
96         "reject-delay",
97         "cleanup-delay",
98         "done"
99 };
100
101 #else
102 #  define TRACE_STATE_MACHINE {}
103 #endif
104
105 static NEVER_RETURNS void _rad_panic(char const *file, unsigned int line, char const *msg)
106 {
107         ERROR("%s[%u]: %s", file, line, msg);
108         fr_exit_now(1);
109 }
110
111 #define rad_panic(x) _rad_panic(__FILE__, __LINE__, x)
112
113 /** Declare a state in the state machine
114  *
115  * Expands to the start of a function definition for a given state.
116  *
117  * @param _x the name of the state.
118  */
119 #define STATE_MACHINE_DECL(_x) static void _x(REQUEST *request, int action)
120
121 static void request_timer(void *ctx);
122
123 /** Insert #REQUEST back into the event heap, to continue executing at a future time
124  *
125  * @param file the state machine timer call occurred in.
126  * @param line the state machine timer call occurred on.
127  * @param request to set add the timer event for.
128  * @param when the event should fine.
129  * @param action to perform when we resume processing the request.
130  */
131 static inline void state_machine_timer(char const *file, int line, REQUEST *request,
132                                        struct timeval *when, fr_state_action_t action)
133 {
134         request->timer_action = action;
135         if (!fr_event_insert(el, request_timer, request, when, &request->ev)) {
136                 _rad_panic(file, line, "Failed to insert event");
137         }
138 }
139
140 /** @copybrief state_machine_timer
141  *
142  * @param _x the action to perform when we resume processing the request.
143  */
144 #define STATE_MACHINE_TIMER(_x) state_machine_timer(__FILE__, __LINE__, request, &when, _x)
145
146 /*
147  *      We need a different VERIFY_REQUEST macro in process.c
148  *      To avoid the race conditions with the master thread
149  *      checking the REQUEST whilst it's being worked on by
150  *      the child.
151  */
152 #if defined(WITH_VERIFY_PTR) && defined(HAVE_PTHREAD_H)
153 #  undef VERIFY_REQUEST
154 #  define VERIFY_REQUEST(_x) if (pthread_equal(pthread_self(), _x->child_pid) != 0) verify_request(__FILE__, __LINE__, _x)
155 #endif
156
157 /**
158  * @section request_timeline
159  *
160  *      Time sequence of a request
161  * @code
162  *
163  *      RQ-----------------P=============================Y-J-C
164  *       ::::::::::::::::::::::::::::::::::::::::::::::::::::::::M
165  * @endcode
166  *
167  * -    R: received.  Duplicate detection is done, and request is
168  *         cached.
169  *
170  * -    Q: Request is placed onto a queue for child threads to pick up.
171  *         If there are no child threads, the request goes immediately
172  *         to P.
173  *
174  * -    P: Processing the request through the modules.
175  *
176  * -    Y: Reply is ready.  Rejects MAY be delayed here.  All other
177  *         replies are sent immediately.
178  *
179  * -    J: Reject is sent "response_delay" after the reply is ready.
180  *
181  * -    C: For Access-Requests, After "cleanup_delay", the request is
182  *         deleted.  Accounting-Request packets go directly from Y to C.
183  *
184  * -    M: Max request time.  If the request hits this timer, it is
185  *         forcibly stopped.
186  *
187  *      Other considerations include duplicate and conflicting
188  *      packets.  When a dupicate packet is received, it is ignored
189  *      until we've reached Y, as no response is ready.  If the reply
190  *      is a reject, duplicates are ignored until J, when we're ready
191  *      to send the reply.  In between the reply being sent (Y or J),
192  *      and C, the server responds to duplicates by sending the cached
193  *      reply.
194  *
195  *      Conflicting packets are sent in 2 situations.
196  *
197  *      The first is in between R and Y.  In that case, we consider
198  *      it as a hint that we're taking too long, and the NAS has given
199  *      up on the request.  We then behave just as if the M timer was
200  *      reached, and we discard the current request.  This allows us
201  *      to process the new one.
202  *
203  *      The second case is when we're at Y, but we haven't yet
204  *      finished processing the request.  This is a race condition in
205  *      the threading code (avoiding locks is faster).  It means that
206  *      a thread has actually encoded and sent the reply, and that the
207  *      NAS has responded with a new packet.  The server can then
208  *      safely mark the current request as "OK to delete", and behaves
209  *      just as if the M timer was reached.  This usually happens only
210  *      in high-load situations.
211  *
212  *      Duplicate packets are sent when the NAS thinks we're taking
213  *      too long, and wants a reply.  From R-Y, duplicates are
214  *      ignored.  From Y-J (for Access-Rejects), duplicates are also
215  *      ignored.  From Y-C, duplicates get a duplicate reply.  *And*,
216  *      they cause the "cleanup_delay" time to be extended.  This
217  *      extension means that we're more likely to send a duplicate
218  *      reply (if we have one), or to suppress processing the packet
219  *      twice if we didn't reply to it.
220  *
221  *      All functions in this file should be thread-safe, and should
222  *      assume thet the REQUEST structure is being accessed
223  *      simultaneously by the main thread, and by the child worker
224  *      threads.  This means that timers, etc. cannot be updated in
225  *      the child thread.
226  *
227  *      Instead, the master thread periodically calls request->process
228  *      with action TIMER.  It's up to the individual functions to
229  *      determine how to handle that.  They need to check if they're
230  *      being called from a child thread or the master, and then do
231  *      different things based on that.
232  */
233 #ifdef WITH_PROXY
234 static fr_packet_list_t *proxy_list = NULL;
235 static TALLOC_CTX *proxy_ctx = NULL;
236 #endif
237
238 #ifdef HAVE_PTHREAD_H
239 #  ifdef WITH_PROXY
240 static pthread_mutex_t proxy_mutex;
241 static bool proxy_no_new_sockets = false;
242 #  endif
243
244 #  define PTHREAD_MUTEX_LOCK if (spawn_flag) pthread_mutex_lock
245 #  define PTHREAD_MUTEX_UNLOCK if (spawn_flag) pthread_mutex_unlock
246
247 static pthread_t NO_SUCH_CHILD_PID;
248 #  define NO_CHILD_THREAD request->child_pid = NO_SUCH_CHILD_PID
249
250 #else
251 /*
252  *      This is easier than ifdef's throughout the code.
253  */
254 #  define PTHREAD_MUTEX_LOCK(_x)
255 #  define PTHREAD_MUTEX_UNLOCK(_x)
256 #  define NO_CHILD_THREAD
257 #endif
258
259 #ifdef HAVE_PTHREAD_H
260 static bool we_are_master(void)
261 {
262         if (spawn_flag &&
263             (pthread_equal(pthread_self(), NO_SUCH_CHILD_PID) == 0)) {
264                 return false;
265         }
266
267         return true;
268 }
269
270 /*
271  *      Assertions are debug checks.
272  */
273 #  ifndef NDEBUG
274 #    define ASSERT_MASTER       if (!we_are_master()) rad_panic("We are not master")
275 #    endif
276 #else
277
278 /*
279  *      No threads: we're always master.
280  */
281 #  define we_are_master(_x) (1)
282 #endif  /* HAVE_PTHREAD_H */
283
284 #ifndef ASSERT_MASTER
285 #  define ASSERT_MASTER
286 #endif
287
288 static int event_new_fd(rad_listen_t *this);
289
290 /*
291  *      We need mutexes around the event FD list *only* in certain
292  *      cases.
293  */
294 #if defined (HAVE_PTHREAD_H) && (defined(WITH_PROXY) || defined(WITH_TCP))
295 static rad_listen_t *new_listeners = NULL;
296
297 static pthread_mutex_t  fd_mutex;
298 #  define FD_MUTEX_LOCK if (spawn_flag) pthread_mutex_lock
299 #  define FD_MUTEX_UNLOCK if (spawn_flag) pthread_mutex_unlock
300
301 void radius_update_listener(rad_listen_t *this)
302 {
303         /*
304          *      Just do it ourselves.
305          */
306         if (we_are_master()) {
307                 event_new_fd(this);
308                 return;
309         }
310
311         FD_MUTEX_LOCK(&fd_mutex);
312
313         /*
314          *      If it's already in the list, don't add it again.
315          */
316         if (this->next) {
317                 FD_MUTEX_UNLOCK(&fd_mutex);
318                 return;
319         }
320
321         /*
322          *      Otherwise, add it to the list
323          */
324         this->next = new_listeners;
325         new_listeners = this;
326         FD_MUTEX_UNLOCK(&fd_mutex);
327         radius_signal_self(RADIUS_SIGNAL_SELF_NEW_FD);
328 }
329 #else
330 void radius_update_listener(rad_listen_t *this)
331 {
332         /*
333          *      No threads.  Just insert it.
334          */
335         event_new_fd(this);
336 }
337 /*
338  *      This is easier than ifdef's throughout the code.
339  */
340 #  define FD_MUTEX_LOCK(_x)
341 #  define FD_MUTEX_UNLOCK(_x)
342 #endif
343
344 static int request_num_counter = 1;
345 #ifdef WITH_PROXY
346 static int request_will_proxy(REQUEST *request) CC_HINT(nonnull);
347 static int request_proxy(REQUEST *request, int retransmit) CC_HINT(nonnull);
348 STATE_MACHINE_DECL(request_ping) CC_HINT(nonnull);
349
350 STATE_MACHINE_DECL(request_response_delay) CC_HINT(nonnull);
351 STATE_MACHINE_DECL(request_cleanup_delay) CC_HINT(nonnull);
352 STATE_MACHINE_DECL(request_running) CC_HINT(nonnull);
353 STATE_MACHINE_DECL(request_done) CC_HINT(nonnull);
354
355 STATE_MACHINE_DECL(proxy_no_reply) CC_HINT(nonnull);
356 STATE_MACHINE_DECL(proxy_running) CC_HINT(nonnull);
357 STATE_MACHINE_DECL(proxy_wait_for_reply) CC_HINT(nonnull);
358
359 static int process_proxy_reply(REQUEST *request, RADIUS_PACKET *reply) CC_HINT(nonnull (1));
360 static void remove_from_proxy_hash(REQUEST *request) CC_HINT(nonnull);
361 static void remove_from_proxy_hash_nl(REQUEST *request, bool yank) CC_HINT(nonnull);
362 static int insert_into_proxy_hash(REQUEST *request) CC_HINT(nonnull);
363 #endif
364
365 static REQUEST *request_setup(TALLOC_CTX *ctx, rad_listen_t *listener, RADIUS_PACKET *packet,
366                               RADCLIENT *client, RAD_REQUEST_FUNP fun);
367 static int request_pre_handler(REQUEST *request, UNUSED int action) CC_HINT(nonnull);
368
369 #ifdef WITH_COA
370 static void request_coa_originate(REQUEST *request) CC_HINT(nonnull);
371 STATE_MACHINE_DECL(coa_wait_for_reply) CC_HINT(nonnull);
372 STATE_MACHINE_DECL(coa_no_reply) CC_HINT(nonnull);
373 STATE_MACHINE_DECL(coa_running) CC_HINT(nonnull);
374 static void coa_separate(REQUEST *request) CC_HINT(nonnull);
375 #  define COA_SEPARATE if (request->coa) coa_separate(request->coa);
376 #else
377 #  define COA_SEPARATE
378 #endif
379
380 #define CHECK_FOR_STOP do { if (request->master_state == REQUEST_STOP_PROCESSING) {request_done(request, FR_ACTION_DONE);return;}} while (0)
381
382 #undef USEC
383 #define USEC (1000000)
384
385 #define INSERT_EVENT(_function, _ctx) if (!fr_event_insert(el, _function, _ctx, &((_ctx)->when), &((_ctx)->ev))) { _rad_panic(__FILE__, __LINE__, "Failed to insert event"); }
386
387 static void tv_add(struct timeval *tv, int usec_delay)
388 {
389         if (usec_delay >= USEC) {
390                 tv->tv_sec += usec_delay / USEC;
391                 usec_delay %= USEC;
392         }
393         tv->tv_usec += usec_delay;
394
395         if (tv->tv_usec >= USEC) {
396                 tv->tv_sec += tv->tv_usec / USEC;
397                 tv->tv_usec %= USEC;
398         }
399 }
400
401 /*
402  *      Debug the packet if requested.
403  */
404 static void debug_packet(REQUEST *request, RADIUS_PACKET *packet, bool received)
405 {
406         char src_ipaddr[128];
407         char dst_ipaddr[128];
408
409         if (!packet) return;
410         if (!RDEBUG_ENABLED) return;
411
412         /*
413          *      Client-specific debugging re-prints the input
414          *      packet into the client log.
415          *
416          *      This really belongs in a utility library
417          */
418         if (is_radius_code(packet->code)) {
419                 RDEBUG("%s %s Id %i from %s:%i to %s:%i length %zu",
420                        received ? "Received" : "Sent",
421                        fr_packet_codes[packet->code],
422                        packet->id,
423                        inet_ntop(packet->src_ipaddr.af,
424                                  &packet->src_ipaddr.ipaddr,
425                                  src_ipaddr, sizeof(src_ipaddr)),
426                        packet->src_port,
427                        inet_ntop(packet->dst_ipaddr.af,
428                                  &packet->dst_ipaddr.ipaddr,
429                                  dst_ipaddr, sizeof(dst_ipaddr)),
430                        packet->dst_port,
431                        packet->data_len);
432         } else {
433                 RDEBUG("%s code %i Id %i from %s:%i to %s:%i length %zu",
434                        received ? "Received" : "Sent",
435                        packet->code,
436                        packet->id,
437                        inet_ntop(packet->src_ipaddr.af,
438                                  &packet->src_ipaddr.ipaddr,
439                                  src_ipaddr, sizeof(src_ipaddr)),
440                        packet->src_port,
441                        inet_ntop(packet->dst_ipaddr.af,
442                                  &packet->dst_ipaddr.ipaddr,
443                                  dst_ipaddr, sizeof(dst_ipaddr)),
444                        packet->dst_port,
445                        packet->data_len);
446         }
447
448         if (received) {
449                 rdebug_pair_list(L_DBG_LVL_1, request, packet->vps, NULL);
450         } else {
451                 rdebug_proto_pair_list(L_DBG_LVL_1, request, packet->vps);
452         }
453 }
454
455
456 /***********************************************************************
457  *
458  *      Start of RADIUS server state machine.
459  *
460  ***********************************************************************/
461
462 static struct timeval *request_response_window(REQUEST *request)
463 {
464         VERIFY_REQUEST(request);
465
466         if (request->client) {
467                 /*
468                  *      The client hasn't set the response window.  Return
469                  *      either the home server one, if set, or the global one.
470                  */
471                 if (!timerisset(&request->client->response_window)) {
472                         return &request->home_server->response_window;
473                 }
474
475                 if (timercmp(&request->client->response_window,
476                              &request->home_server->response_window, <)) {
477                         return &request->client->response_window;
478                 }
479         }
480
481         rad_assert(request->home_server != NULL);
482         return &request->home_server->response_window;
483 }
484
485 /*
486  * Determine initial request processing delay.
487  */
488 static int request_init_delay(REQUEST *request)
489 {
490         struct timeval half_response_window;
491
492         VERIFY_REQUEST(request);
493
494         /* Allow client response window to lower initial delay */
495         if (timerisset(&request->client->response_window)) {
496                 half_response_window.tv_sec = request->client->response_window.tv_sec >> 1;
497                 half_response_window.tv_usec =
498                         ((request->client->response_window.tv_sec & 1) * USEC +
499                                 request->client->response_window.tv_usec) >> 1;
500                 if (timercmp(&half_response_window, &request->root->init_delay, <))
501                         return (int)half_response_window.tv_sec * USEC +
502                                 (int)half_response_window.tv_usec;
503         }
504
505         return (int)request->root->init_delay.tv_sec * USEC +
506                 (int)request->root->init_delay.tv_usec;
507 }
508
509 /*
510  *      Callback for ALL timer events related to the request.
511  */
512 static void request_timer(void *ctx)
513 {
514         REQUEST *request = talloc_get_type_abort(ctx, REQUEST);
515         int action;
516
517         action = request->timer_action;
518
519         TRACE_STATE_MACHINE;
520
521         request->process(request, action);
522 }
523
524 /*
525  *      Wrapper for talloc pools.  If there's no parent, just free the
526  *      request.  If there is a parent, free the parent INSTEAD of the
527  *      request.
528  */
529 static void request_free(REQUEST *request)
530 {
531         void *ptr;
532
533         rad_assert(request->ev == NULL);
534         rad_assert(!request->in_request_hash);
535         rad_assert(!request->in_proxy_hash);
536
537         if ((request->options & RAD_REQUEST_OPTION_CTX) == 0) {
538                 talloc_free(request);
539                 return;
540         }
541
542         ptr = talloc_parent(request);
543         rad_assert(ptr != NULL);
544         talloc_free(ptr);
545 }
546
547
548 #ifdef WITH_PROXY
549 static void proxy_reply_too_late(REQUEST *request)
550 {
551         char buffer[128];
552
553         RDEBUG2("Reply from home server %s port %d  - ID: %d arrived too late.  Try increasing 'retry_delay' or 'max_request_time'",
554                 inet_ntop(request->proxy->dst_ipaddr.af,
555                           &request->proxy->dst_ipaddr.ipaddr,
556                           buffer, sizeof(buffer)),
557                 request->proxy->dst_port, request->proxy->id);
558 }
559 #endif
560
561
562 /** Mark a request DONE and clean it up.
563  *
564  *  When a request is DONE, it can have ties to a number of other
565  *  portions of the server.  The request hash, proxy hash, events,
566  *  child threads, etc.  This function takes care of either cleaning
567  *  up the request, or managing the timers to wait for the ties to be
568  *  removed.
569  *
570  *  \dot
571  *      digraph done {
572  *              done -> done [ label = "still running" ];
573  *      }
574  *  \enddot
575  */
576 static void request_done(REQUEST *request, int action)
577 {
578         struct timeval now, when;
579
580         VERIFY_REQUEST(request);
581
582         TRACE_STATE_MACHINE;
583
584         /*
585          *      Force this no matter what.
586          */
587         request->process = request_done;
588
589 #ifdef WITH_DETAIL
590         /*
591          *      Tell the detail listener that we're done.
592          */
593         if (request->listener &&
594             (request->listener->type == RAD_LISTEN_DETAIL) &&
595             (request->simul_max != 1)) {
596                 request->simul_max = 1;
597                 request->listener->send(request->listener,
598                                         request);
599         }
600 #endif
601
602 #ifdef HAVE_PTHREAD_H
603         /*
604          *      If called from a child thread, mark ourselves as done,
605          *      and wait for the master thread timer to clean us up.
606          */
607         if (!we_are_master()) {
608                 NO_CHILD_THREAD;
609                 request->child_state = REQUEST_DONE;
610                 return;
611         }
612 #endif
613
614         /*
615          *      Mark the request as STOP.
616          */
617         request->master_state = REQUEST_STOP_PROCESSING;
618
619 #ifdef WITH_COA
620         /*
621          *      Move the CoA request to its own handler.
622          */
623         if (request->coa) {
624                 coa_separate(request->coa);
625         } else if (request->parent && (request->parent->coa == request)) {
626                 coa_separate(request);
627         }
628 #endif
629
630         /*
631          *      It doesn't hurt to send duplicate replies.  All other
632          *      signals are ignored, as the request will be cleaned up
633          *      soon anyways.
634          */
635         switch (action) {
636         case FR_ACTION_DUP:
637 #ifdef WITH_DETAIL
638                 rad_assert(request->listener != NULL);
639 #endif
640                 if (request->reply->code != 0) {
641                         request->listener->send(request->listener, request);
642                         return;
643                 } else {
644                         RDEBUG("No reply.  Ignoring retransmit");
645                 }
646                 break;
647
648                 /*
649                  *      Mark the request as done.
650                  */
651         case FR_ACTION_DONE:
652 #ifdef HAVE_PTHREAD_H
653                 /*
654                  *      If the child is still running, leave it alone.
655                  */
656                 if (spawn_flag && (request->child_state <= REQUEST_RUNNING)) {
657                         break;
658                 }
659 #endif
660
661 #ifdef DEBUG_STATE_MACHINE
662                 if (debug_flag) printf("(%u) ********\tSTATE %s C-%s -> C-%s\t********\n",
663                                        request->number, __FUNCTION__,
664                                        child_state_names[request->child_state],
665                                        child_state_names[REQUEST_DONE]);
666 #endif
667                 request->child_state = REQUEST_DONE;
668                 break;
669
670                 /*
671                  *      Called when the child is taking too long to
672                  *      finish.  We've already marked it "please
673                  *      stop", so we don't complain any more.
674                  */
675         case FR_ACTION_TIMER:
676                 break;
677
678 #ifdef WITH_PROXY
679         case FR_ACTION_PROXY_REPLY:
680                 proxy_reply_too_late(request);
681                 break;
682 #endif
683
684         default:
685                 break;
686         }
687
688         /*
689          *      Remove it from the request hash.
690          */
691         if (request->in_request_hash) {
692                 if (!rbtree_deletebydata(pl, &request->packet)) {
693                         rad_assert(0 == 1);
694                 }
695                 request->in_request_hash = false;
696         }
697
698 #ifdef WITH_PROXY
699         /*
700          *      Wait for the proxy ID to expire.  This allows us to
701          *      avoid re-use of proxy IDs for a while.
702          */
703         if (request->in_proxy_hash) {
704                 rad_assert(request->proxy != NULL);
705
706                 fr_event_now(el, &now);
707                 when = request->proxy->timestamp;
708
709 #ifdef WITH_COA
710                 if (((request->proxy->code == PW_CODE_COA_REQUEST) ||
711                      (request->proxy->code == PW_CODE_DISCONNECT_REQUEST)) &&
712                     (request->packet->code != request->proxy->code)) {
713                         when.tv_sec += request->home_server->coa_mrd;
714                 } else
715 #endif
716                         timeradd(&when, request_response_window(request), &when);
717
718                 /*
719                  *      We haven't received all responses, AND there's still
720                  *      time to wait.  Do so.
721                  */
722                 if ((request->num_proxied_requests > request->num_proxied_responses) &&
723 #ifdef WITH_TCP
724                     (request->home_server->proto != IPPROTO_TCP) &&
725 #endif
726                     timercmp(&now, &when, <)) {
727                         RDEBUG("Waiting for more responses from the home server");
728                         goto wait_some_more;
729                 }
730
731                 /*
732                  *      Time to remove it.
733                  */
734                 remove_from_proxy_hash(request);
735         }
736 #endif
737
738 #ifdef HAVE_PTHREAD_H
739         /*
740          *      If there's no children, we can mark the request as done.
741          */
742         if (!spawn_flag) request->child_state = REQUEST_DONE;
743 #endif
744
745         /*
746          *      If the child is still running, wait for it to be finished.
747          */
748         if (request->child_state <= REQUEST_RUNNING) {
749                 gettimeofday(&now, NULL);
750 #ifdef WITH_PROXY
751         wait_some_more:
752 #endif
753                 when = now;
754                 if (request->delay < (USEC / 3)) request->delay = USEC / 3;
755                 tv_add(&when, request->delay);
756                 request->delay += request->delay >> 1;
757                 if (request->delay > (10 * USEC)) request->delay = 10 * USEC;
758
759                 STATE_MACHINE_TIMER(FR_ACTION_TIMER);
760                 return;
761         }
762
763 #ifdef HAVE_PTHREAD_H
764         rad_assert(request->child_pid == NO_SUCH_CHILD_PID);
765 #endif
766
767         /*
768          *      @todo: do final states for TCP sockets, too?
769          */
770         request_stats_final(request);
771 #ifdef WITH_TCP
772         if (request->listener) {
773                 request->listener->count--;
774
775                 /*
776                  *      If we're the last one, remove the listener now.
777                  */
778                 if ((request->listener->count == 0) &&
779                     (request->listener->status >= RAD_LISTEN_STATUS_FROZEN)) {
780                         event_new_fd(request->listener);
781                 }
782         }
783 #endif
784
785         if (request->packet) {
786                 RDEBUG2("Cleaning up request packet ID %u with timestamp +%d",
787                         request->packet->id,
788                         (unsigned int) (request->timestamp - fr_start_time));
789         } /* else don't print anything */
790
791         ASSERT_MASTER;
792         fr_event_delete(el, &request->ev);
793         request_free(request);
794 }
795
796
797 static void request_cleanup_delay_init(REQUEST *request)
798 {
799         struct timeval now, when;
800
801         VERIFY_REQUEST(request);
802
803         /*
804          *      Do cleanup delay ONLY for RADIUS packets from a real
805          *      client.  Everything else just gets cleaned up
806          *      immediately.
807          */
808         if (!(request->packet->code == PW_CODE_ACCESS_REQUEST)
809 #ifdef WITH_COA
810             || (request->packet->code == PW_CODE_COA_REQUEST)
811             || (request->packet->code == PW_CODE_DISCONNECT_REQUEST)
812 #endif
813                 ) goto done;
814
815         if (!request->root->cleanup_delay) goto done;
816
817         gettimeofday(&now, NULL);
818
819         rad_assert(request->reply->timestamp.tv_sec != 0);
820         when = request->reply->timestamp;
821
822         request->delay = request->root->cleanup_delay;
823         when.tv_sec += request->delay;
824
825         /*
826          *      Set timer for when we need to clean it up.
827          */
828         if (timercmp(&when, &now, >)) {
829 #ifdef DEBUG_STATE_MACHINE
830                 if (debug_flag) printf("(%u) ********\tNEXT-STATE %s -> %s\n", request->number, __FUNCTION__, "request_cleanup_delay");
831 #endif
832                 request->process = request_cleanup_delay;
833                 request->child_state = REQUEST_CLEANUP_DELAY;
834
835                 /*
836                  *      Update this if we can, otherwise let the timers pick it up.
837                  */
838                 if (we_are_master()) {
839                         STATE_MACHINE_TIMER(FR_ACTION_TIMER);
840                 } else {
841                         NO_CHILD_THREAD;
842                 }
843                 return;
844         }
845
846         /*
847          *      Otherwise just clean it up.
848          */
849 done:
850         request_done(request, FR_ACTION_DONE);
851 }
852
853
854 /*
855  *      Enforce max_request_time.
856  */
857 static void request_max_time(REQUEST *request)
858 {
859         struct timeval now, when;
860         rad_assert(request->magic == REQUEST_MAGIC);
861 #ifdef DEBUG_STATE_MACHINE
862         int action = FR_ACTION_TIMER;
863 #endif
864
865         VERIFY_REQUEST(request);
866
867         TRACE_STATE_MACHINE;
868         ASSERT_MASTER;
869
870         /*
871          *      The child thread has acknowledged it's done.
872          *      Transition to the DONE state.
873          *
874          *      If the request was marked STOP, then the "check for
875          *      stop" macro already took care of it.
876          */
877         if (request->child_state == REQUEST_DONE) {
878         done:
879                 request_done(request, FR_ACTION_DONE);
880                 return;
881         }
882
883         /*
884          *      The request is still running.  Enforce max_request_time.
885          */
886         fr_event_now(el, &now);
887         when = request->packet->timestamp;
888         when.tv_sec += request->root->max_request_time;
889
890         /*
891          *      Taking too long: tell it to die.
892          */
893         if (timercmp(&now, &when, >=)) {
894 #ifdef HAVE_PTHREAD_H
895                 /*
896                  *      If there's a child thread processing it,
897                  *      complain.
898                  */
899                 if (spawn_flag &&
900                     (pthread_equal(request->child_pid, NO_SUCH_CHILD_PID) == 0)) {
901                         ERROR("Unresponsive child for request %u, in component %s module %s",
902                               request->number,
903                               request->component ? request->component : "<core>",
904                               request->module ? request->module : "<core>");
905                         exec_trigger(request, NULL, "server.thread.unresponsive", true);
906                 }
907 #endif
908                 /*
909                  *      Tell the request that it's done.
910                  */
911                 goto done;
912         }
913
914         /*
915          *      Sleep for some more.  We HOPE that the child will
916          *      become responsive at some point in the future.  We do
917          *      this by adding 50% to the current timer.
918          */
919         when = now;
920         tv_add(&when, request->delay);
921         request->delay += request->delay >> 1;
922         STATE_MACHINE_TIMER(FR_ACTION_TIMER);
923 }
924
925 static void request_queue_or_run(REQUEST *request,
926                                  fr_request_process_t process)
927 {
928 #ifdef DEBUG_STATE_MACHINE
929         int action = FR_ACTION_TIMER;
930 #endif
931
932         VERIFY_REQUEST(request);
933
934         TRACE_STATE_MACHINE;
935
936         /*
937          *      Do this here so that fewer other functions need to do
938          *      it.
939          */
940         if (request->master_state == REQUEST_STOP_PROCESSING) {
941 #ifdef DEBUG_STATE_MACHINE
942                 if (debug_flag) printf("(%u) ********\tSTATE %s M-%s causes C-%s-> C-%s\t********\n",
943                                        request->number, __FUNCTION__,
944                                        master_state_names[request->master_state],
945                                        child_state_names[request->child_state],
946                                        child_state_names[REQUEST_DONE]);
947 #endif
948                 request_done(request, FR_ACTION_DONE);
949                 return;
950         }
951
952         request->process = process;
953
954         if (we_are_master()) {
955                 struct timeval when;
956
957                 /*
958                  *      (re) set the initial delay.
959                  */
960                 request->delay = request_init_delay(request);
961                 if (request->delay > USEC) request->delay = USEC;
962                 gettimeofday(&when, NULL);
963                 tv_add(&when, request->delay);
964                 request->delay += request->delay >> 1;
965
966                 STATE_MACHINE_TIMER(FR_ACTION_TIMER);
967
968 #ifdef HAVE_PTHREAD_H
969                 if (spawn_flag) {
970                         /*
971                          *      A child thread will eventually pick it up.
972                          */
973                         if (request_enqueue(request)) return;
974
975                         /*
976                          *      Otherwise we're not going to do anything with
977                          *      it...
978                          */
979                         request_done(request, FR_ACTION_DONE);
980                         return;
981                 }
982 #endif
983         }
984
985         request->child_state = REQUEST_RUNNING;
986         request->process(request, FR_ACTION_RUN);
987
988 #ifdef WNOHANG
989         /*
990          *      Requests that care about child process exit
991          *      codes have already either called
992          *      rad_waitpid(), or they've given up.
993          */
994         while (waitpid(-1, NULL, WNOHANG) > 0);
995 #endif
996 }
997
998
999 static void request_dup(REQUEST *request)
1000 {
1001         ERROR("(%u) Ignoring duplicate packet from "
1002               "client %s port %d - ID: %u due to unfinished request "
1003               "in component %s module %s",
1004               request->number, request->client->shortname,
1005               request->packet->src_port,request->packet->id,
1006               request->component, request->module);
1007 }
1008
1009
1010 /** Sit on a request until it's time to clean it up.
1011  *
1012  *  A NAS may not see a response from the server.  When the NAS
1013  *  retransmits, we want to be able to send a cached reply back.  The
1014  *  alternative is to re-process the packet, which does bad things for
1015  *  EAP, among others.
1016  *
1017  *  IF we do see a NAS retransmit, we extend the cleanup delay,
1018  *  because the NAS might miss our cached reply.
1019  *
1020  *  Otherwise, once we reach cleanup_delay, we transition to DONE.
1021  *
1022  *  \dot
1023  *      digraph cleanup_delay {
1024  *              cleanup_delay;
1025  *              send_reply [ label = "send_reply\nincrease cleanup delay" ];
1026  *
1027  *              cleanup_delay -> send_reply [ label = "DUP" ];
1028  *              send_reply -> cleanup_delay;
1029  *              cleanup_delay -> proxy_reply_too_late [ label = "PROXY_REPLY", arrowhead = "none" ];
1030  *              cleanup_delay -> cleanup_delay [ label = "TIMER < timeout" ];
1031  *              cleanup_delay -> done [ label = "TIMER >= timeout" ];
1032  *      }
1033  *  \enddot
1034  */
1035 static void request_cleanup_delay(REQUEST *request, int action)
1036 {
1037         struct timeval when, now;
1038
1039         VERIFY_REQUEST(request);
1040
1041         TRACE_STATE_MACHINE;
1042         ASSERT_MASTER;
1043         COA_SEPARATE;
1044         CHECK_FOR_STOP;
1045
1046         switch (action) {
1047         case FR_ACTION_DUP:
1048                 if (request->reply->code != 0) {
1049                         request->listener->send(request->listener, request);
1050                 } else {
1051                         RDEBUG("No reply.  Ignoring retransmit");
1052                 }
1053
1054                 /*
1055                  *      Double the cleanup_delay to catch retransmits.
1056                  */
1057                 when = request->reply->timestamp;
1058                 request->delay += request->delay;
1059                 when.tv_sec += request->delay;
1060
1061                 STATE_MACHINE_TIMER(FR_ACTION_TIMER);
1062                 break;
1063
1064 #ifdef WITH_PROXY
1065         case FR_ACTION_PROXY_REPLY:
1066                 proxy_reply_too_late(request);
1067                 break;
1068 #endif
1069
1070         case FR_ACTION_TIMER:
1071                 fr_event_now(el, &now);
1072
1073                 rad_assert(request->root->cleanup_delay > 0);
1074
1075                 when = request->reply->timestamp;
1076                 when.tv_sec += request->root->cleanup_delay;
1077
1078                 if (timercmp(&when, &now, >)) {
1079 #ifdef DEBUG_STATE_MACHINE
1080                         if (debug_flag) printf("(%u) ********\tNEXT-STATE %s -> %s\n", request->number, __FUNCTION__, "request_cleanup_delay");
1081 #endif
1082                         STATE_MACHINE_TIMER(FR_ACTION_TIMER);
1083                         return;
1084                 } /* else it's time to clean up */
1085
1086                 request_done(request, REQUEST_DONE);
1087                 break;
1088
1089         default:
1090                 RDEBUG3("%s: Ignoring action %s", __FUNCTION__, action_codes[action]);
1091                 break;
1092         }
1093 }
1094
1095
1096 /** Sit on a request until it's time to respond to it.
1097  *
1098  *  For security reasons, rejects (and maybe some other) packets are
1099  *  delayed for a while before we respond.  This delay means that
1100  *  badly behaved NASes don't hammer the server with authentication
1101  *  attempts.
1102  *
1103  *  Otherwise, once we reach response_delay, we send the reply, and
1104  *  transition to cleanup_delay.
1105  *
1106  *  \dot
1107  *      digraph response_delay {
1108  *              response_delay -> proxy_reply_too_late [ label = "PROXY_REPLY", arrowhead = "none" ];
1109  *              response_delay -> response_delay [ label = "DUP, TIMER < timeout" ];
1110  *              response_delay -> send_reply [ label = "TIMER >= timeout" ];
1111  *              send_reply -> cleanup_delay;
1112  *      }
1113  *  \enddot
1114  */
1115 static void request_response_delay(REQUEST *request, int action)
1116 {
1117         struct timeval when, now;
1118
1119         VERIFY_REQUEST(request);
1120
1121         TRACE_STATE_MACHINE;
1122         ASSERT_MASTER;
1123         COA_SEPARATE;
1124         CHECK_FOR_STOP;
1125
1126         switch (action) {
1127         case FR_ACTION_DUP:
1128                 ERROR("(%u) Discarding duplicate request from "
1129                       "client %s port %d - ID: %u due to delayed response",
1130                       request->number, request->client->shortname,
1131                       request->packet->src_port,request->packet->id);
1132                 break;
1133
1134 #ifdef WITH_PROXY
1135         case FR_ACTION_PROXY_REPLY:
1136                 proxy_reply_too_late(request);
1137                 break;
1138 #endif
1139
1140         case FR_ACTION_TIMER:
1141                 fr_event_now(el, &now);
1142
1143                 rad_assert(request->response_delay.tv_sec > 0);
1144
1145                 /*
1146                  *      See if it's time to send the reply.  If not,
1147                  *      we wait some more.
1148                  */
1149                 when = request->reply->timestamp;
1150
1151                 tv_add(&when, request->response_delay.tv_sec * USEC);
1152                 tv_add(&when, request->response_delay.tv_usec);
1153
1154                 if (timercmp(&when, &now, >)) {
1155 #ifdef DEBUG_STATE_MACHINE
1156                         if (debug_flag) printf("(%u) ********\tNEXT-STATE %s -> %s\n", request->number, __FUNCTION__, "request_response_delay");
1157 #endif
1158                         STATE_MACHINE_TIMER(FR_ACTION_TIMER);
1159                         return;
1160                 } /* else it's time to send the reject */
1161
1162                 RDEBUG2("Sending delayed response");
1163                 debug_packet(request, request->reply, false);
1164                 request->listener->send(request->listener, request);
1165
1166                 /*
1167                  *      Clean up the request.
1168                  */
1169                 request_cleanup_delay_init(request);
1170                 break;
1171
1172         default:
1173                 RDEBUG3("%s: Ignoring action %s", __FUNCTION__, action_codes[action]);
1174                 break;
1175         }
1176 }
1177
1178
1179 static int request_pre_handler(REQUEST *request, UNUSED int action)
1180 {
1181         int rcode;
1182
1183         VERIFY_REQUEST(request);
1184
1185         TRACE_STATE_MACHINE;
1186
1187         if (request->master_state == REQUEST_STOP_PROCESSING) return 0;
1188
1189         /*
1190          *      Don't decode the packet if it's an internal "fake"
1191          *      request.  Instead, just return so that the caller can
1192          *      process it.
1193          */
1194         if (request->packet->dst_port == 0) {
1195                 request->username = pairfind(request->packet->vps, PW_USER_NAME, 0, TAG_ANY);
1196                 request->password = pairfind(request->packet->vps, PW_USER_PASSWORD, 0, TAG_ANY);
1197                 return 1;
1198         }
1199
1200         if (!request->packet->vps) { /* FIXME: check for correct state */
1201                 rcode = request->listener->decode(request->listener, request);
1202
1203 #ifdef WITH_UNLANG
1204                 if (debug_condition) {
1205                         /*
1206                          *      Ignore parse errors.
1207                          */
1208                         if (radius_evaluate_cond(request, RLM_MODULE_OK, 0, debug_condition)) {
1209                                 request->log.lvl = L_DBG_LVL_2;
1210                                 request->log.func = vradlog_request;
1211                         }
1212                 }
1213 #endif
1214
1215                 debug_packet(request, request->packet, true);
1216         } else {
1217                 rcode = 0;
1218         }
1219
1220         if (rcode < 0) {
1221                 RATE_LIMIT(INFO("Dropping packet without response because of error: %s", fr_strerror()));
1222                 request->reply->offset = -2; /* bad authenticator */
1223                 return 0;
1224         }
1225
1226         if (!request->username) {
1227                 request->username = pairfind(request->packet->vps, PW_USER_NAME, 0, TAG_ANY);
1228         }
1229
1230         return 1;
1231 }
1232
1233
1234 /**  Do the final processing of a request before we reply to the NAS.
1235  *
1236  *  Various cleanups, suppress responses, copy Proxy-State, and set
1237  *  response_delay or cleanup_delay;
1238  */
1239 static void request_finish(REQUEST *request, int action)
1240 {
1241         VALUE_PAIR *vp;
1242
1243         VERIFY_REQUEST(request);
1244
1245         TRACE_STATE_MACHINE;
1246         CHECK_FOR_STOP;
1247
1248         (void) action;  /* -Wunused */
1249
1250 #ifdef WITH_COA
1251         /*
1252          *      Don't do post-auth if we're a CoA request originated
1253          *      from an Access-Request.  See request_alloc_coa() for
1254          *      details.
1255          */
1256         if ((request->options & RAD_REQUEST_OPTION_COA) != 0) goto done;
1257 #endif
1258
1259         /*
1260          *      Override the response code if a control:Response-Packet-Type attribute is present.
1261          */
1262         vp = pairfind(request->config, PW_RESPONSE_PACKET_TYPE, 0, TAG_ANY);
1263         if (vp) {
1264                 if (vp->vp_integer == 256) {
1265                         RDEBUG2("Not responding to request");
1266                         request->reply->code = 0;
1267                 } else {
1268                         request->reply->code = vp->vp_integer;
1269                 }
1270         }
1271         /*
1272          *      Catch Auth-Type := Reject BEFORE proxying the packet.
1273          */
1274         else if (request->packet->code == PW_CODE_ACCESS_REQUEST) {
1275                 if (request->reply->code == 0) {
1276                         vp = pairfind(request->config, PW_AUTH_TYPE, 0, TAG_ANY);
1277                         if (!vp || (vp->vp_integer != 5)) {
1278                                 RDEBUG2("There was no response configured: "
1279                                         "rejecting request");
1280                         }
1281
1282                         request->reply->code = PW_CODE_ACCESS_REJECT;
1283                 }
1284         }
1285
1286         /*
1287          *      Copy Proxy-State from the request to the reply.
1288          */
1289         vp = paircopy_by_num(request->reply, request->packet->vps,
1290                        PW_PROXY_STATE, 0, TAG_ANY);
1291         if (vp) pairadd(&request->reply->vps, vp);
1292
1293         /*
1294          *      Call Post-Auth for Access-Request packets.
1295          */
1296         if (request->packet->code == PW_CODE_ACCESS_REQUEST) {
1297                 rad_postauth(request);
1298         }
1299
1300
1301         /*
1302          *      Clean up.  These are no longer needed.
1303          */
1304         gettimeofday(&request->reply->timestamp, NULL);
1305
1306         /*
1307          *      Fake packets get marked as "done", and have the
1308          *      proxy-reply section deal with the reply attributes.
1309          *      We therefore don't free the reply attributes.
1310          */
1311         if (request->packet->dst_port == 0) {
1312                 RDEBUG("Finished internally proxied request.");
1313                 NO_CHILD_THREAD;
1314                 request->child_state = REQUEST_DONE;
1315                 return;
1316         }
1317
1318 #ifdef WITH_DETAIL
1319         /*
1320          *      Always send the reply to the detail listener.
1321          */
1322         if (request->listener->type == RAD_LISTEN_DETAIL) {
1323                 request->simul_max = 1;
1324
1325                 /*
1326                  *      But only print the reply if there is one.
1327                  */
1328                 if (request->reply->code != 0) {
1329                         debug_packet(request, request->reply, false);
1330                 }
1331
1332                 request->listener->send(request->listener, request);
1333                 goto done;
1334         }
1335 #endif
1336
1337         /*
1338          *      Ignore all "do not respond" packets.
1339          *      Except for the detail ones, which need to ping
1340          *      the detail file reader so that it will retransmit.
1341          */
1342         if (!request->reply->code) {
1343                 RDEBUG("Not sending reply to client.");
1344                 goto done;
1345         }
1346
1347         /*
1348          *      If it's not in the request hash, we MIGHT not want to
1349          *      send a reply.
1350          *
1351          *      If duplicate packets are allowed, then then only
1352          *      reason to NOT be in the request hash is because we
1353          *      don't want to send a reply.
1354          *
1355          *      FIXME: this is crap.  The rest of the state handling
1356          *      should use a different field so that we don't have two
1357          *      meanings for it.
1358          *
1359          *      Otherwise duplicates are forbidden, and the request is
1360          *      SUPPOSED to avoid the request hash.
1361          *
1362          *      In that case, we need to send a reply.
1363          */
1364         if (!request->in_request_hash &&
1365             !request->listener->nodup) {
1366                 RDEBUG("Suppressing reply to client.");
1367                 goto done;
1368         }
1369
1370         /*
1371          *      See if we need to delay an Access-Reject packet.
1372          */
1373         if ((request->reply->code == PW_CODE_ACCESS_REJECT) &&
1374             (request->root->reject_delay.tv_sec > 0)) {
1375                 request->response_delay = request->root->reject_delay;
1376
1377 #ifdef WITH_PROXY
1378                 /*
1379                  *      If we timed out a proxy packet, don't delay
1380                  *      the reject any more.
1381                  */
1382                 if (request->proxy && !request->proxy_reply) {
1383                         request->response_delay.tv_sec = 0;
1384                         request->response_delay.tv_usec = 0;
1385                 }
1386 #endif
1387         }
1388
1389         /*
1390          *      Send the reply.
1391          */
1392         if (request->response_delay.tv_sec == 0) {
1393                 rad_assert(request->response_delay.tv_usec == 0);
1394
1395                 /*
1396                  *      Don't print a reply if there's none to send.
1397                  */
1398                 if (request->reply->code != 0) {
1399                         debug_packet(request, request->reply, false);
1400                         request->listener->send(request->listener, request);
1401                 }
1402
1403         done:
1404                 RDEBUG2("Finished request");
1405                 request->component = "<core>";
1406                 request->module = "<done>";
1407
1408                 request_cleanup_delay_init(request);
1409
1410         } else {
1411                 /*
1412                  *      Encode and sign it here, so that the master
1413                  *      thread can just send the encoded data, which
1414                  *      means it does less work.
1415                  */
1416                 RDEBUG2("Delaying response for %d.%06d seconds",
1417                         (int) request->response_delay.tv_sec, (int) request->response_delay.tv_usec);
1418                 request->listener->encode(request->listener, request);
1419                 request->component = "<core>";
1420                 request->module = "<delay>";
1421                 request->process = request_response_delay;
1422                 NO_CHILD_THREAD;
1423                 request->child_state = REQUEST_RESPONSE_DELAY;
1424         }
1425 }
1426
1427 /** Process a request from a client.
1428  *
1429  *  The outcome might be that the request is proxied.
1430  *
1431  *  \dot
1432  *      digraph running {
1433  *              running -> running [ label = "TIMER < max_request_time" ];
1434  *              running -> done [ label = "TIMER >= max_request_time" ];
1435  *              running -> proxy [ label = "proxied" ];
1436  *              running -> dup [ label = "DUP", arrowhead = "none" ];
1437  *      }
1438  *  \enddot
1439  */
1440 static void request_running(REQUEST *request, int action)
1441 {
1442         VERIFY_REQUEST(request);
1443
1444         TRACE_STATE_MACHINE;
1445         CHECK_FOR_STOP;
1446
1447         switch (action) {
1448         case FR_ACTION_TIMER:
1449                 COA_SEPARATE;
1450                 request_max_time(request);
1451                 break;
1452
1453         case FR_ACTION_DUP:
1454                 request_dup(request);
1455                 break;
1456
1457         case FR_ACTION_RUN:
1458                 if (!request_pre_handler(request, action)) {
1459 #ifdef DEBUG_STATE_MACHINE
1460                         if (debug_flag) printf("(%u) ********\tSTATE %s failed in pre-handler C-%s -> C-%s\t********\n",
1461                                                request->number, __FUNCTION__,
1462                                                child_state_names[request->child_state],
1463                                                child_state_names[REQUEST_DONE]);
1464 #endif
1465
1466                         NO_CHILD_THREAD;
1467                         request->child_state = REQUEST_DONE;
1468                         break;
1469                 }
1470
1471                 rad_assert(request->handle != NULL);
1472                 request->handle(request);
1473
1474 #ifdef WITH_PROXY
1475                 /*
1476                  *      We may need to send a proxied request.
1477                  */
1478                 if ((action == FR_ACTION_RUN) &&
1479                     request_will_proxy(request)) {
1480 #ifdef DEBUG_STATE_MACHINE
1481                         if (debug_flag) printf("(%u) ********\tWill Proxy\t********\n", request->number);
1482 #endif
1483                         /*
1484                          *      If this fails, it
1485                          *      takes care of setting
1486                          *      up the post proxy fail
1487                          *      handler.
1488                          */
1489                         if (request_proxy(request, 0) < 0) goto req_finished;
1490                 } else
1491 #endif
1492                 {
1493 #ifdef DEBUG_STATE_MACHINE
1494                         if (debug_flag) printf("(%u) ********\tFinished\t********\n", request->number);
1495 #endif
1496
1497 #ifdef WITH_COA
1498                         /*
1499                          *      Maybe originate a CoA request.
1500                          */
1501                         if ((action == FR_ACTION_RUN) && request->coa) {
1502                                 request_coa_originate(request);
1503                         }
1504 #endif
1505
1506 #ifdef WITH_PROXY
1507                 req_finished:
1508 #endif
1509                         request_finish(request, action);
1510                 }
1511                 break;
1512
1513         default:
1514                 RDEBUG3("%s: Ignoring action %s", __FUNCTION__, action_codes[action]);
1515                 break;
1516         }
1517 }
1518
1519 int request_receive(TALLOC_CTX *ctx, rad_listen_t *listener, RADIUS_PACKET *packet,
1520                     RADCLIENT *client, RAD_REQUEST_FUNP fun)
1521 {
1522         uint32_t count;
1523         RADIUS_PACKET **packet_p;
1524         REQUEST *request = NULL;
1525         struct timeval now;
1526         listen_socket_t *sock = NULL;
1527
1528         VERIFY_PACKET(packet);
1529
1530         /*
1531          *      Set the last packet received.
1532          */
1533         gettimeofday(&now, NULL);
1534
1535         packet->timestamp = now;
1536
1537 #ifdef WITH_ACCOUNTING
1538         if (listener->type != RAD_LISTEN_DETAIL)
1539 #endif
1540         {
1541                 sock = listener->data;
1542                 sock->last_packet = now.tv_sec;
1543         }
1544
1545         /*
1546          *      Skip everything if required.
1547          */
1548         if (listener->nodup) goto skip_dup;
1549
1550         packet_p = rbtree_finddata(pl, &packet);
1551         if (packet_p) {
1552                 rad_child_state_t child_state;
1553
1554                 request = fr_packet2myptr(REQUEST, packet, packet_p);
1555                 rad_assert(request->in_request_hash);
1556                 child_state = request->child_state;
1557
1558                 /*
1559                  *      Same src/dst ip/port, length, and
1560                  *      authentication vector: must be a duplicate.
1561                  */
1562                 if ((request->packet->data_len == packet->data_len) &&
1563                     (memcmp(request->packet->vector, packet->vector,
1564                             sizeof(packet->vector)) == 0)) {
1565
1566 #ifdef WITH_STATS
1567                         switch (packet->code) {
1568                         case PW_CODE_ACCESS_REQUEST:
1569                                 FR_STATS_INC(auth, total_dup_requests);
1570                                 break;
1571
1572 #ifdef WITH_ACCOUNTING
1573                         case PW_CODE_ACCOUNTING_REQUEST:
1574                                 FR_STATS_INC(acct, total_dup_requests);
1575                                 break;
1576 #endif
1577 #ifdef WITH_COA
1578                         case PW_CODE_COA_REQUEST:
1579                                 FR_STATS_INC(coa, total_dup_requests);
1580                                 break;
1581
1582                         case PW_CODE_DISCONNECT_REQUEST:
1583                                 FR_STATS_INC(dsc, total_dup_requests);
1584                                 break;
1585 #endif
1586
1587                         default:
1588                                 break;
1589                         }
1590 #endif  /* WITH_STATS */
1591
1592                         /*
1593                          *      Tell the state machine that there's a
1594                          *      duplicate request.
1595                          */
1596                         request->process(request, FR_ACTION_DUP);
1597                         return 0; /* duplicate of live request */
1598                 }
1599
1600                 /*
1601                  *      Mark the request as done ASAP, and before we
1602                  *      log anything.  The child may stop processing
1603                  *      the request just as we're logging the
1604                  *      complaint.
1605                  */
1606                 request_done(request, FR_ACTION_DONE);
1607                 request = NULL;
1608
1609                 /*
1610                  *      It's a new request, not a duplicate.  If the
1611                  *      old one is done, then we can clean it up.
1612                  */
1613                 if (child_state <= REQUEST_RUNNING) {
1614                         /*
1615                          *      The request is still QUEUED or RUNNING.  That's a problem.
1616                          */
1617                         ERROR("Received conflicting packet from "
1618                               "client %s port %d - ID: %u due to "
1619                               "unfinished request.  Giving up on old request.",
1620                               client->shortname,
1621                               packet->src_port, packet->id);
1622                 }
1623
1624                 /*
1625                  *      Mark the old request as done.  If there's no
1626                  *      child, the request will be cleaned up
1627                  *      immediately.  If there is a child, we'll set a
1628                  *      timer to go clean up the request.
1629                  */
1630         } /* else the new packet is unique */
1631
1632         /*
1633          *      Quench maximum number of outstanding requests.
1634          */
1635         if (main_config.max_requests &&
1636             ((count = rbtree_num_elements(pl)) > main_config.max_requests)) {
1637                 RATE_LIMIT(ERROR("Dropping request (%d is too many): from client %s port %d - ID: %d", count,
1638                                  client->shortname,
1639                                  packet->src_port, packet->id);
1640                            WARN("Please check the configuration file.\n"
1641                                 "\tThe value for 'max_requests' is probably set too low.\n"));
1642
1643                 exec_trigger(NULL, NULL, "server.max_requests", true);
1644                 return 0;
1645         }
1646
1647 skip_dup:
1648         /*
1649          *      Rate-limit the incoming packets
1650          */
1651         if (sock && sock->max_rate) {
1652                 uint32_t pps;
1653
1654                 pps = rad_pps(&sock->rate_pps_old, &sock->rate_pps_now, &sock->rate_time, &now);
1655                 if (pps > sock->max_rate) {
1656                         DEBUG("Dropping request due to rate limiting");
1657                         return 0;
1658                 }
1659                 sock->rate_pps_now++;
1660         }
1661
1662         /*
1663          *      Allocate a pool for the request.
1664          */
1665         if (!ctx) {
1666                 ctx = talloc_pool(NULL, main_config.talloc_pool_size);
1667                 if (!ctx) return 0;
1668
1669                 /*
1670                  *      The packet is still allocated from a different
1671                  *      context, but oh well.
1672                  */
1673                 (void) talloc_steal(ctx, packet);
1674         }
1675
1676         request = request_setup(ctx, listener, packet, client, fun);
1677         if (!request) {
1678                 talloc_free(ctx);
1679                 return 1;
1680         }
1681
1682         /*
1683          *      Mark it as a "real" request with a context.
1684          */
1685         request->options |= RAD_REQUEST_OPTION_CTX;
1686
1687         /*
1688          *      Remember the request in the list.
1689          */
1690         if (!listener->nodup) {
1691                 if (!rbtree_insert(pl, &request->packet)) {
1692                         RERROR("Failed to insert request in the list of live requests: discarding it");
1693                         request_done(request, FR_ACTION_DONE);
1694                         return 1;
1695                 }
1696
1697                 request->in_request_hash = true;
1698         }
1699
1700         /*
1701          *      Process it.  Send a response, and free it.
1702          */
1703         if (listener->synchronous) {
1704 #ifdef WITH_DETAIL
1705                 rad_assert(listener->type != RAD_LISTEN_DETAIL);
1706 #endif
1707
1708                 request->listener->decode(request->listener, request);
1709                 request->username = pairfind(request->packet->vps, PW_USER_NAME, 0, TAG_ANY);
1710                 request->password = pairfind(request->packet->vps, PW_USER_PASSWORD, 0, TAG_ANY);
1711
1712                 fun(request);
1713
1714                 if (request->reply->code != 0) {
1715                         request->listener->send(request->listener, request);
1716                 } else {
1717                         RDEBUG("Not sending reply");
1718                 }
1719
1720                 /*
1721                  *      Don't do delayed reject.  Oh well.
1722                  */
1723                 request_free(request);
1724                 return 1;
1725         }
1726
1727         /*
1728          *      Otherwise, insert it into the state machine.
1729          *      The child threads will take care of processing it.
1730          */
1731         request_queue_or_run(request, request_running);
1732
1733         return 1;
1734 }
1735
1736
1737 static REQUEST *request_setup(TALLOC_CTX *ctx, rad_listen_t *listener, RADIUS_PACKET *packet,
1738                               RADCLIENT *client, RAD_REQUEST_FUNP fun)
1739 {
1740         REQUEST *request;
1741
1742         /*
1743          *      Create and initialize the new request.
1744          */
1745         request = request_alloc(ctx);
1746         if (!request) {
1747                 ERROR("No memory");
1748                 return NULL;
1749         }
1750         request->reply = rad_alloc(request, false);
1751         if (!request->reply) {
1752                 ERROR("No memory");
1753                 talloc_free(request);
1754                 return NULL;
1755         }
1756
1757         request->listener = listener;
1758         request->client = client;
1759         request->packet = talloc_steal(request, packet);
1760         request->number = request_num_counter++;
1761         request->priority = listener->type;
1762         request->master_state = REQUEST_ACTIVE;
1763         request->child_state = REQUEST_RUNNING;
1764 #ifdef DEBUG_STATE_MACHINE
1765         if (debug_flag) printf("(%u) ********\tSTATE %s C-%s -> C-%s\t********\n",
1766                                request->number, __FUNCTION__,
1767                                child_state_names[request->child_state],
1768                                child_state_names[REQUEST_RUNNING]);
1769 #endif
1770 #ifdef HAVE_PTHREAD_H
1771         request->child_pid = NO_SUCH_CHILD_PID;
1772 #endif
1773         request->handle = fun;
1774         NO_CHILD_THREAD;
1775
1776 #ifdef WITH_STATS
1777         request->listener->stats.last_packet = request->packet->timestamp.tv_sec;
1778         if (packet->code == PW_CODE_ACCESS_REQUEST) {
1779                 request->client->auth.last_packet = request->packet->timestamp.tv_sec;
1780                 radius_auth_stats.last_packet = request->packet->timestamp.tv_sec;
1781 #ifdef WITH_ACCOUNTING
1782         } else if (packet->code == PW_CODE_ACCOUNTING_REQUEST) {
1783                 request->client->acct.last_packet = request->packet->timestamp.tv_sec;
1784                 radius_acct_stats.last_packet = request->packet->timestamp.tv_sec;
1785 #endif
1786         }
1787 #endif  /* WITH_STATS */
1788
1789         /*
1790          *      Status-Server packets go to the head of the queue.
1791          */
1792         if (request->packet->code == PW_CODE_STATUS_SERVER) request->priority = 0;
1793
1794         /*
1795          *      Set virtual server identity
1796          */
1797         if (client->server) {
1798                 request->server = client->server;
1799         } else if (listener->server) {
1800                 request->server = listener->server;
1801         } else {
1802                 request->server = NULL;
1803         }
1804
1805         request->root = &main_config;
1806 #ifdef WITH_TCP
1807         request->listener->count++;
1808 #endif
1809
1810         /*
1811          *      The request passes many of our sanity checks.
1812          *      From here on in, if anything goes wrong, we
1813          *      send a reject message, instead of dropping the
1814          *      packet.
1815          */
1816
1817         /*
1818          *      Build the reply template from the request.
1819          */
1820
1821         request->reply->sockfd = request->packet->sockfd;
1822         request->reply->dst_ipaddr = request->packet->src_ipaddr;
1823         request->reply->src_ipaddr = request->packet->dst_ipaddr;
1824         request->reply->dst_port = request->packet->src_port;
1825         request->reply->src_port = request->packet->dst_port;
1826         request->reply->id = request->packet->id;
1827         request->reply->code = 0; /* UNKNOWN code */
1828         memcpy(request->reply->vector, request->packet->vector,
1829                sizeof(request->reply->vector));
1830         request->reply->vps = NULL;
1831         request->reply->data = NULL;
1832         request->reply->data_len = 0;
1833
1834         return request;
1835 }
1836
1837 #ifdef WITH_TCP
1838 /***********************************************************************
1839  *
1840  *      TCP Handlers.
1841  *
1842  ***********************************************************************/
1843
1844 /*
1845  *      Timer function for all TCP sockets.
1846  */
1847 static void tcp_socket_timer(void *ctx)
1848 {
1849         rad_listen_t *listener = talloc_get_type_abort(ctx, rad_listen_t);
1850         listen_socket_t *sock = listener->data;
1851         struct timeval end, now;
1852         char buffer[256];
1853         fr_socket_limit_t *limit;
1854
1855         ASSERT_MASTER;
1856
1857         if (listener->status != RAD_LISTEN_STATUS_KNOWN) return;
1858
1859         fr_event_now(el, &now);
1860
1861         switch (listener->type) {
1862 #ifdef WITH_PROXY
1863         case RAD_LISTEN_PROXY:
1864                 limit = &sock->home->limit;
1865                 break;
1866 #endif
1867
1868         case RAD_LISTEN_AUTH:
1869 #ifdef WITH_ACCOUNTING
1870         case RAD_LISTEN_ACCT:
1871 #endif
1872                 limit = &sock->limit;
1873                 break;
1874
1875         default:
1876                 return;
1877         }
1878
1879         /*
1880          *      If we enforce a lifetime, do it now.
1881          */
1882         if (limit->lifetime > 0) {
1883                 end.tv_sec = sock->opened + limit->lifetime;
1884                 end.tv_usec = 0;
1885
1886                 if (timercmp(&end, &now, <=)) {
1887                         listener->print(listener, buffer, sizeof(buffer));
1888                         DEBUG("Reached maximum lifetime on socket %s", buffer);
1889
1890                 do_close:
1891
1892 #ifdef WITH_PROXY
1893                         /*
1894                          *      Proxy sockets get frozen, so that we don't use
1895                          *      them for new requests.  But we do keep them
1896                          *      open to listen for replies to requests we had
1897                          *      previously sent.
1898                          */
1899                         if (listener->type == RAD_LISTEN_PROXY) {
1900                                 PTHREAD_MUTEX_LOCK(&proxy_mutex);
1901                                 if (!fr_packet_list_socket_freeze(proxy_list,
1902                                                                   listener->fd)) {
1903                                         ERROR("Fatal error freezing socket: %s", fr_strerror());
1904                                         fr_exit(1);
1905                                 }
1906                                 PTHREAD_MUTEX_UNLOCK(&proxy_mutex);
1907                         }
1908 #endif
1909
1910                         /*
1911                          *      Mark the socket as "don't use if at all possible".
1912                          */
1913                         listener->status = RAD_LISTEN_STATUS_FROZEN;
1914                         event_new_fd(listener);
1915                         return;
1916                 }
1917         } else {
1918                 end = now;
1919                 end.tv_sec += 3600;
1920         }
1921
1922         /*
1923          *      Enforce an idle timeout.
1924          */
1925         if (limit->idle_timeout > 0) {
1926                 struct timeval idle;
1927
1928                 rad_assert(sock->last_packet != 0);
1929                 idle.tv_sec = sock->last_packet + limit->idle_timeout;
1930                 idle.tv_usec = 0;
1931
1932                 if (timercmp(&idle, &now, <=)) {
1933                         listener->print(listener, buffer, sizeof(buffer));
1934                         DEBUG("Reached idle timeout on socket %s", buffer);
1935                         goto do_close;
1936                 }
1937
1938                 /*
1939                  *      Enforce the minimum of idle timeout or lifetime.
1940                  */
1941                 if (timercmp(&idle, &end, <)) {
1942                         end = idle;
1943                 }
1944         }
1945
1946         /*
1947          *      Wake up at t + 0.5s.  The code above checks if the timers
1948          *      are <= t.  This addition gives us a bit of leeway.
1949          */
1950         end.tv_usec = USEC / 2;
1951
1952         ASSERT_MASTER;
1953         if (!fr_event_insert(el, tcp_socket_timer, listener, &end, &sock->ev)) {
1954                 rad_panic("Failed to insert event");
1955         }
1956 }
1957
1958
1959 #ifdef WITH_PROXY
1960 /*
1961  *      Add +/- 2s of jitter, as suggested in RFC 3539
1962  *      and in RFC 5080.
1963  */
1964 static void add_jitter(struct timeval *when)
1965 {
1966         uint32_t jitter;
1967
1968         when->tv_sec -= 2;
1969
1970         jitter = fr_rand();
1971         jitter ^= (jitter >> 10);
1972         jitter &= ((1 << 22) - 1); /* 22 bits of 1 */
1973
1974         /*
1975          *      Add in ~ (4 * USEC) of jitter.
1976          */
1977         tv_add(when, jitter);
1978 }
1979
1980 /*
1981  *      Called by socket_del to remove requests with this socket
1982  */
1983 static int eol_proxy_listener(void *ctx, void *data)
1984 {
1985         rad_listen_t *this = talloc_get_type_abort(ctx, rad_listen_t);
1986         RADIUS_PACKET **proxy_p = data;
1987         REQUEST *request;
1988
1989         request = fr_packet2myptr(REQUEST, proxy, proxy_p);
1990         if (request->proxy_listener != this) return 0;
1991
1992         /*
1993          *      The normal "remove_from_proxy_hash" tries to grab the
1994          *      proxy mutex.  We already have it held, so grabbing it
1995          *      again will cause a deadlock.  Instead, call the "no
1996          *      lock" version of the function.
1997          */
1998         rad_assert(request->in_proxy_hash == true);
1999         remove_from_proxy_hash_nl(request, false);
2000
2001         /*
2002          *      Don't mark it as DONE.  The client can retransmit, and
2003          *      the packet SHOULD be re-proxied somewhere else.
2004          *
2005          *      Return "2" means that the rbtree code will remove it
2006          *      from the tree, and we don't need to do it ourselves.
2007          */
2008         return 2;
2009 }
2010 #endif  /* WITH_PROXY */
2011
2012 static int eol_listener(void *ctx, void *data)
2013 {
2014         rad_listen_t *this = talloc_get_type_abort(ctx, rad_listen_t);
2015         RADIUS_PACKET **packet_p = data;
2016         REQUEST *request;
2017
2018         request = fr_packet2myptr(REQUEST, packet, packet_p);
2019         if (request->listener != this) return 0;
2020
2021         request->master_state = REQUEST_STOP_PROCESSING;
2022         request->process = request_done;
2023
2024         return 0;
2025 }
2026 #endif  /* WITH_TCP */
2027
2028 #ifdef WITH_PROXY
2029 /***********************************************************************
2030  *
2031  *      Proxy handlers for the state machine.
2032  *
2033  ***********************************************************************/
2034
2035 /*
2036  *      Called with the proxy mutex held
2037  */
2038 static void remove_from_proxy_hash_nl(REQUEST *request, bool yank)
2039 {
2040         VERIFY_REQUEST(request);
2041
2042         if (!request->in_proxy_hash) return;
2043
2044         fr_packet_list_id_free(proxy_list, request->proxy, yank);
2045         request->in_proxy_hash = false;
2046
2047         /*
2048          *      On the FIRST reply, decrement the count of outstanding
2049          *      requests.  Note that this is NOT the count of sent
2050          *      packets, but whether or not the home server has
2051          *      responded at all.
2052          */
2053         if (request->home_server &&
2054             request->home_server->currently_outstanding) {
2055                 request->home_server->currently_outstanding--;
2056
2057                 /*
2058                  *      If we're NOT sending it packets, AND it's been
2059                  *      a while since we got a response, then we don't
2060                  *      know if it's alive or dead.
2061                  */
2062                 if ((request->home_server->currently_outstanding == 0) &&
2063                     (request->home_server->state == HOME_STATE_ALIVE)) {
2064                         struct timeval when, now;
2065
2066                         when.tv_sec = request->home_server->last_packet_recv ;
2067                         when.tv_usec = 0;
2068
2069                         timeradd(&when, request_response_window(request), &when);
2070                         gettimeofday(&now, NULL);
2071
2072                         /*
2073                          *      last_packet + response_window
2074                          *
2075                          *      We *administratively* mark the home
2076                          *      server as "unknown" state, because we
2077                          *      haven't seen a packet for a while.
2078                          */
2079                         if (timercmp(&now, &when, >)) {
2080                                 request->home_server->state = HOME_STATE_UNKNOWN;
2081                                 request->home_server->last_packet_sent = 0;
2082                                 request->home_server->last_packet_recv = 0;
2083                         }
2084                 }
2085         }
2086
2087 #ifdef WITH_TCP
2088         rad_assert(request->proxy_listener != NULL);
2089         request->proxy_listener->count--;
2090 #endif
2091         request->proxy_listener = NULL;
2092
2093         /*
2094          *      Got from YES in hash, to NO, not in hash while we hold
2095          *      the mutex.  This guarantees that when another thread
2096          *      grabs the mutex, the "not in hash" flag is correct.
2097          */
2098         RDEBUG3("proxy: request is no longer in proxy hash");
2099 }
2100
2101 static void remove_from_proxy_hash(REQUEST *request)
2102 {
2103         VERIFY_REQUEST(request);
2104
2105         /*
2106          *      Check this without grabbing the mutex because it's a
2107          *      lot faster that way.
2108          */
2109         if (!request->in_proxy_hash) return;
2110
2111         /*
2112          *      The "not in hash" flag is definitive.  However, if the
2113          *      flag says that it IS in the hash, there might still be
2114          *      a race condition where it isn't.
2115          */
2116         PTHREAD_MUTEX_LOCK(&proxy_mutex);
2117
2118         if (!request->in_proxy_hash) {
2119                 PTHREAD_MUTEX_UNLOCK(&proxy_mutex);
2120                 return;
2121         }
2122
2123         remove_from_proxy_hash_nl(request, true);
2124
2125         PTHREAD_MUTEX_UNLOCK(&proxy_mutex);
2126 }
2127
2128 static int insert_into_proxy_hash(REQUEST *request)
2129 {
2130         char buf[128];
2131         int tries;
2132         bool success = false;
2133         void *proxy_listener;
2134
2135         VERIFY_REQUEST(request);
2136
2137         rad_assert(request->proxy != NULL);
2138         rad_assert(request->home_server != NULL);
2139         rad_assert(proxy_list != NULL);
2140
2141
2142         PTHREAD_MUTEX_LOCK(&proxy_mutex);
2143         proxy_listener = NULL;
2144         request->num_proxied_requests = 1;
2145         request->num_proxied_responses = 0;
2146
2147         for (tries = 0; tries < 2; tries++) {
2148                 rad_listen_t *this;
2149                 listen_socket_t *sock;
2150
2151                 RDEBUG3("proxy: Trying to allocate ID (%d/2)", tries);
2152                 success = fr_packet_list_id_alloc(proxy_list,
2153                                                 request->home_server->proto,
2154                                                 &request->proxy, &proxy_listener);
2155                 if (success) break;
2156
2157                 if (tries > 0) continue; /* try opening new socket only once */
2158
2159 #ifdef HAVE_PTHREAD_H
2160                 if (proxy_no_new_sockets) break;
2161 #endif
2162
2163                 RDEBUG3("proxy: Trying to open a new listener to the home server");
2164                 this = proxy_new_listener(proxy_ctx, request->home_server, 0);
2165                 if (!this) {
2166                         PTHREAD_MUTEX_UNLOCK(&proxy_mutex);
2167                         goto fail;
2168                 }
2169
2170                 request->proxy->src_port = 0; /* Use any new socket */
2171                 proxy_listener = this;
2172
2173                 sock = this->data;
2174                 if (!fr_packet_list_socket_add(proxy_list, this->fd,
2175                                                sock->proto,
2176                                                &sock->other_ipaddr, sock->other_port,
2177                                                this)) {
2178
2179 #ifdef HAVE_PTHREAD_H
2180                         proxy_no_new_sockets = true;
2181 #endif
2182                         PTHREAD_MUTEX_UNLOCK(&proxy_mutex);
2183
2184                         /*
2185                          *      This is bad.  However, the
2186                          *      packet list now supports 256
2187                          *      open sockets, which should
2188                          *      minimize this problem.
2189                          */
2190                         ERROR("Failed adding proxy socket: %s",
2191                               fr_strerror());
2192                         goto fail;
2193                 }
2194
2195                 /*
2196                  *      Add it to the event loop.  Ensure that we have
2197                  *      only one mutex locked at a time.
2198                  */
2199                 PTHREAD_MUTEX_UNLOCK(&proxy_mutex);
2200                 radius_update_listener(this);
2201                 PTHREAD_MUTEX_LOCK(&proxy_mutex);
2202         }
2203
2204         if (!proxy_listener || !success) {
2205                 PTHREAD_MUTEX_UNLOCK(&proxy_mutex);
2206                 REDEBUG2("proxy: Failed allocating Id for proxied request");
2207         fail:
2208                 request->proxy_listener = NULL;
2209                 request->in_proxy_hash = false;
2210                 return 0;
2211         }
2212
2213         rad_assert(request->proxy->id >= 0);
2214
2215         request->proxy_listener = proxy_listener;
2216         request->in_proxy_hash = true;
2217         RDEBUG3("proxy: request is now in proxy hash");
2218
2219         /*
2220          *      Keep track of maximum outstanding requests to a
2221          *      particular home server.  'max_outstanding' is
2222          *      enforced in home_server_ldb(), in realms.c.
2223          */
2224         request->home_server->currently_outstanding++;
2225
2226 #ifdef WITH_TCP
2227         request->proxy_listener->count++;
2228 #endif
2229
2230         PTHREAD_MUTEX_UNLOCK(&proxy_mutex);
2231
2232         RDEBUG3("proxy: allocating destination %s port %d - Id %d",
2233                inet_ntop(request->proxy->dst_ipaddr.af,
2234                          &request->proxy->dst_ipaddr.ipaddr, buf, sizeof(buf)),
2235                request->proxy->dst_port,
2236                request->proxy->id);
2237
2238         return 1;
2239 }
2240
2241 static int process_proxy_reply(REQUEST *request, RADIUS_PACKET *reply)
2242 {
2243         int rcode;
2244         int post_proxy_type = 0;
2245         VALUE_PAIR *vp;
2246
2247         VERIFY_REQUEST(request);
2248
2249         /*
2250          *      There may be a proxy reply, but it may be too late.
2251          */
2252         if (!request->home_server->server && !request->proxy_listener) return 0;
2253
2254         /*
2255          *      Delete any reply we had accumulated until now.
2256          */
2257         RDEBUG2("Clearing existing &reply: attributes");
2258         pairfree(&request->reply->vps);
2259
2260         /*
2261          *      Run the packet through the post-proxy stage,
2262          *      BEFORE playing games with the attributes.
2263          */
2264         vp = pairfind(request->config, PW_POST_PROXY_TYPE, 0, TAG_ANY);
2265         if (vp) {
2266                 post_proxy_type = vp->vp_integer;
2267         /*
2268          *      If we have a proxy_reply, and it was a reject, or a NAK
2269          *      setup Post-Proxy <type>.
2270          *
2271          *      If the <type> doesn't have a section, then the Post-Proxy
2272          *      section is ignored.
2273          */
2274         } else if (reply) {
2275                 DICT_VALUE *dval = NULL;
2276
2277                 switch (reply->code) {
2278                 case PW_CODE_ACCESS_REJECT:
2279                         dval = dict_valbyname(PW_POST_PROXY_TYPE, 0, "Reject");
2280                         if (dval) post_proxy_type = dval->value;
2281                         break;
2282
2283                 case PW_CODE_DISCONNECT_NAK:
2284                         dval = dict_valbyname(PW_POST_PROXY_TYPE, 0, fr_packet_codes[reply->code]);
2285                         if (dval) post_proxy_type = dval->value;
2286                         break;
2287
2288                 case PW_CODE_COA_NAK:
2289                         dval = dict_valbyname(PW_POST_PROXY_TYPE, 0, fr_packet_codes[reply->code]);
2290                         if (dval) post_proxy_type = dval->value;
2291                         break;
2292
2293                 default:
2294                         break;
2295                 }
2296
2297                 /*
2298                  *      Create config:Post-Proxy-Type
2299                  */
2300                 if (dval) {
2301                         vp = radius_paircreate(request, &request->config, PW_POST_PROXY_TYPE, 0);
2302                         vp->vp_integer = dval->value;
2303                 }
2304         }
2305
2306         if (post_proxy_type > 0) RDEBUG2("Found Post-Proxy-Type %s",
2307                                          dict_valnamebyattr(PW_POST_PROXY_TYPE, 0, post_proxy_type));
2308
2309         if (reply) {
2310                 VERIFY_PACKET(reply);
2311
2312                 /*
2313                  *      Decode the packet if required.
2314                  */
2315                 if (request->proxy_listener) {
2316                         rcode = request->proxy_listener->decode(request->proxy_listener, request);
2317                         debug_packet(request, reply, true);
2318
2319                         /*
2320                          *      Pro-actively remove it from the proxy hash.
2321                          *      This is later than in 2.1.x, but it means that
2322                          *      the replies are authenticated before being
2323                          *      removed from the hash.
2324                          */
2325                         if ((rcode == 0) &&
2326                             (request->num_proxied_requests <= request->num_proxied_responses)) {
2327                                 remove_from_proxy_hash(request);
2328                         }
2329                 } else {
2330                         rad_assert(!request->in_proxy_hash);
2331                 }
2332         } else if (request->in_proxy_hash) {
2333                 remove_from_proxy_hash(request);
2334         }
2335
2336         if (request->home_pool && request->home_pool->virtual_server) {
2337                 char const *old_server = request->server;
2338
2339                 request->server = request->home_pool->virtual_server;
2340                 RDEBUG2("server %s {", request->server);
2341                 RINDENT();
2342                 rcode = process_post_proxy(post_proxy_type, request);
2343                 REXDENT();
2344                 RDEBUG2("}");
2345                 request->server = old_server;
2346         } else {
2347                 rcode = process_post_proxy(post_proxy_type, request);
2348         }
2349
2350 #ifdef WITH_COA
2351         if (request->packet->code == request->proxy->code)
2352           /*
2353            *    Don't run the next bit if we originated a CoA
2354            *    packet, after receiving an Access-Request or
2355            *    Accounting-Request.
2356            */
2357 #endif
2358
2359         /*
2360          *      There may NOT be a proxy reply, as we may be
2361          *      running Post-Proxy-Type = Fail.
2362          */
2363         if (reply) {
2364                 pairadd(&request->reply->vps, paircopy(request->reply, reply->vps));
2365
2366                 /*
2367                  *      Delete the Proxy-State Attributes from
2368                  *      the reply.  These include Proxy-State
2369                  *      attributes from us and remote server.
2370                  */
2371                 pairdelete(&request->reply->vps, PW_PROXY_STATE, 0, TAG_ANY);
2372         }
2373
2374         switch (rcode) {
2375         default:  /* Don't do anything */
2376                 break;
2377         case RLM_MODULE_FAIL:
2378                 return 0;
2379
2380         case RLM_MODULE_HANDLED:
2381                 return 0;
2382         }
2383
2384         return 1;
2385 }
2386
2387 int request_proxy_reply(RADIUS_PACKET *packet)
2388 {
2389         RADIUS_PACKET **proxy_p;
2390         REQUEST *request;
2391         struct timeval now;
2392         char buffer[128];
2393
2394         VERIFY_PACKET(packet);
2395
2396         PTHREAD_MUTEX_LOCK(&proxy_mutex);
2397         proxy_p = fr_packet_list_find_byreply(proxy_list, packet);
2398
2399         if (!proxy_p) {
2400                 PTHREAD_MUTEX_UNLOCK(&proxy_mutex);
2401                 PROXY("No outstanding request was found for reply from host %s port %d - ID %u",
2402                        inet_ntop(packet->src_ipaddr.af,
2403                                  &packet->src_ipaddr.ipaddr,
2404                                  buffer, sizeof(buffer)),
2405                        packet->src_port, packet->id);
2406                 return 0;
2407         }
2408
2409         request = fr_packet2myptr(REQUEST, proxy, proxy_p);
2410         request->num_proxied_responses++; /* needs to be protected by lock */
2411
2412         PTHREAD_MUTEX_UNLOCK(&proxy_mutex);
2413
2414         /*
2415          *      No reply, BUT the current packet fails verification:
2416          *      ignore it.  This does the MD5 calculations in the
2417          *      server core, but I guess we can fix that later.
2418          */
2419         if (!request->proxy_reply &&
2420             (rad_verify(packet, request->proxy,
2421                         request->home_server->secret) != 0)) {
2422                 DEBUG("Ignoring spoofed proxy reply.  Signature is invalid");
2423                 return 0;
2424         }
2425
2426         /*
2427          *      The home server sent us a packet which doesn't match
2428          *      something we have: ignore it.  This is done only to
2429          *      catch the case of broken systems.
2430          */
2431         if (request->proxy_reply &&
2432             (memcmp(request->proxy_reply->vector,
2433                     packet->vector,
2434                     sizeof(request->proxy_reply->vector)) != 0)) {
2435                 RDEBUG2("Ignoring conflicting proxy reply");
2436                 return 0;
2437         }
2438
2439         gettimeofday(&now, NULL);
2440
2441         /*
2442          *      Status-Server packets don't count as real packets.
2443          */
2444         if (request->proxy->code != PW_CODE_STATUS_SERVER) {
2445                 listen_socket_t *sock = request->proxy_listener->data;
2446
2447                 request->home_server->last_packet_recv = now.tv_sec;
2448                 sock->last_packet = now.tv_sec;
2449         }
2450
2451         /*
2452          *      If we have previously seen a reply, ignore the
2453          *      duplicate.
2454          */
2455         if (request->proxy_reply) {
2456                 RDEBUG2("Discarding duplicate reply from host %s port %d  - ID: %d",
2457                         inet_ntop(packet->src_ipaddr.af,
2458                                   &packet->src_ipaddr.ipaddr,
2459                                   buffer, sizeof(buffer)),
2460                         packet->src_port, packet->id);
2461                 return 0;
2462         }
2463
2464         /*
2465          *      Call the state machine to do something useful with the
2466          *      request.
2467          */
2468         request->proxy_reply = talloc_steal(request, packet);
2469         packet->timestamp = now;
2470         request->priority = RAD_LISTEN_PROXY;
2471
2472 #ifdef WITH_STATS
2473         /*
2474          *      Update the proxy listener stats here, because only one
2475          *      thread accesses that at a time.  The home_server and
2476          *      main proxy_*_stats structures are updated once the
2477          *      request is cleaned up.
2478          */
2479         request->proxy_listener->stats.total_responses++;
2480
2481         request->home_server->stats.last_packet = packet->timestamp.tv_sec;
2482         request->proxy_listener->stats.last_packet = packet->timestamp.tv_sec;
2483
2484         switch (request->proxy->code) {
2485         case PW_CODE_ACCESS_REQUEST:
2486                 proxy_auth_stats.last_packet = packet->timestamp.tv_sec;
2487
2488                 if (request->proxy_reply->code == PW_CODE_ACCESS_ACCEPT) {
2489                         request->proxy_listener->stats.total_access_accepts++;
2490
2491                 } else if (request->proxy_reply->code == PW_CODE_ACCESS_REJECT) {
2492                         request->proxy_listener->stats.total_access_rejects++;
2493
2494                 } else if (request->proxy_reply->code == PW_CODE_ACCESS_CHALLENGE) {
2495                         request->proxy_listener->stats.total_access_challenges++;
2496                 }
2497                 break;
2498
2499 #ifdef WITH_ACCOUNTING
2500         case PW_CODE_ACCOUNTING_REQUEST:
2501                 request->proxy_listener->stats.total_responses++;
2502                 proxy_acct_stats.last_packet = packet->timestamp.tv_sec;
2503                 break;
2504
2505 #endif
2506
2507 #ifdef WITH_COA
2508         case PW_CODE_COA_REQUEST:
2509                 request->proxy_listener->stats.total_responses++;
2510                 proxy_coa_stats.last_packet = packet->timestamp.tv_sec;
2511                 break;
2512
2513         case PW_CODE_DISCONNECT_REQUEST:
2514                 request->proxy_listener->stats.total_responses++;
2515                 proxy_dsc_stats.last_packet = packet->timestamp.tv_sec;
2516                 break;
2517
2518 #endif
2519         default:
2520                 break;
2521         }
2522 #endif
2523
2524         /*
2525          *      We've received a reply.  If we hadn't been sending it
2526          *      packets for a while, just mark it alive.
2527          */
2528         if (request->home_server->state == HOME_STATE_UNKNOWN) {
2529                 request->home_server->state = HOME_STATE_ALIVE;
2530                 request->home_server->response_timeouts = 0;
2531         }
2532
2533         /*
2534          *      Tell the request state machine that we have a proxy
2535          *      reply.  Depending on the function, this should either
2536          *      ignore it, or process it.
2537          */
2538         request->process(request, FR_ACTION_PROXY_REPLY);
2539
2540         return 1;
2541 }
2542
2543
2544 static int setup_post_proxy_fail(REQUEST *request)
2545 {
2546         DICT_VALUE const *dval = NULL;
2547         VALUE_PAIR *vp;
2548
2549         VERIFY_REQUEST(request);
2550
2551         if (request->proxy->code == PW_CODE_ACCESS_REQUEST) {
2552                 dval = dict_valbyname(PW_POST_PROXY_TYPE, 0,
2553                                       "Fail-Authentication");
2554         } else if (request->proxy->code == PW_CODE_ACCOUNTING_REQUEST) {
2555                 dval = dict_valbyname(PW_POST_PROXY_TYPE, 0,
2556                                       "Fail-Accounting");
2557 #ifdef WITH_COA
2558         } else if (request->proxy->code == PW_CODE_COA_REQUEST) {
2559                 dval = dict_valbyname(PW_POST_PROXY_TYPE, 0, "Fail-CoA");
2560
2561         } else if (request->proxy->code == PW_CODE_DISCONNECT_REQUEST) {
2562                 dval = dict_valbyname(PW_POST_PROXY_TYPE, 0, "Fail-Disconnect");
2563 #endif
2564         } else {
2565                 WARN("Unknown packet type in Post-Proxy-Type Fail: ignoring");
2566                 return 0;
2567         }
2568
2569         if (!dval) dval = dict_valbyname(PW_POST_PROXY_TYPE, 0, "Fail");
2570
2571         if (!dval) {
2572                 pairdelete(&request->config, PW_POST_PROXY_TYPE, 0, TAG_ANY);
2573                 return 0;
2574         }
2575
2576         vp = pairfind(request->config, PW_POST_PROXY_TYPE, 0, TAG_ANY);
2577         if (!vp) vp = radius_paircreate(request, &request->config,
2578                                         PW_POST_PROXY_TYPE, 0);
2579         vp->vp_integer = dval->value;
2580
2581         return 1;
2582 }
2583
2584
2585 /** Process a request after the proxy has timed out.
2586  *
2587  *  Run the packet through Post-Proxy-Type Fail
2588  *
2589  *  \dot
2590  *      digraph proxy_no_reply {
2591  *              proxy_no_reply;
2592  *
2593  *              proxy_no_reply -> dup [ label = "DUP", arrowhead = "none" ];
2594  *              proxy_no_reply -> timer [ label = "TIMER < max_request_time" ];
2595  *              proxy_no_reply -> proxy_reply_too_late [ label = "PROXY_REPLY" arrowhead = "none"];
2596  *              proxy_no_reply -> process_proxy_reply [ label = "RUN" ];
2597  *              proxy_no_reply -> done [ label = "TIMER >= timeout" ];
2598  *      }
2599  *  \enddot
2600  */
2601 static void proxy_no_reply(REQUEST *request, int action)
2602 {
2603         VERIFY_REQUEST(request);
2604
2605         TRACE_STATE_MACHINE;
2606         CHECK_FOR_STOP;
2607
2608         switch (action) {
2609         case FR_ACTION_DUP:
2610                 request_dup(request);
2611                 break;
2612
2613         case FR_ACTION_TIMER:
2614                 request_max_time(request);
2615                 break;
2616
2617         case FR_ACTION_PROXY_REPLY:
2618                 proxy_reply_too_late(request);
2619                 break;
2620
2621         case FR_ACTION_RUN:
2622                 if (process_proxy_reply(request, NULL)) {
2623                         request->handle(request);
2624                 }
2625                 request_finish(request, action);
2626                 break;
2627
2628         default:
2629                 RDEBUG3("%s: Ignoring action %s", __FUNCTION__, action_codes[action]);
2630                 break;
2631         }
2632 }
2633
2634 /** Process the request after receiving a proxy reply.
2635  *
2636  *  Throught the post-proxy section, and the through the handler
2637  *  function.
2638  *
2639  *  \dot
2640  *      digraph proxy_running {
2641  *              proxy_running;
2642  *
2643  *              proxy_running -> dup [ label = "DUP", arrowhead = "none" ];
2644  *              proxy_running -> timer [ label = "TIMER < max_request_time" ];
2645  *              proxy_running -> process_proxy_reply [ label = "RUN" ];
2646  *              proxy_running -> done [ label = "TIMER >= timeout" ];
2647  *      }
2648  *  \enddot
2649  */
2650 static void proxy_running(REQUEST *request, int action)
2651 {
2652         VERIFY_REQUEST(request);
2653
2654         TRACE_STATE_MACHINE;
2655         CHECK_FOR_STOP;
2656
2657         switch (action) {
2658         case FR_ACTION_DUP:
2659                 request_dup(request);
2660                 break;
2661
2662         case FR_ACTION_TIMER:
2663                 request_max_time(request);
2664                 break;
2665
2666         case FR_ACTION_RUN:
2667                 if (process_proxy_reply(request, request->proxy_reply)) {
2668                         request->handle(request);
2669                 }
2670                 request_finish(request, action);
2671                 break;
2672
2673         default:                /* duplicate proxy replies are suppressed */
2674                 RDEBUG3("%s: Ignoring action %s", __FUNCTION__, action_codes[action]);
2675                 break;
2676         }
2677 }
2678
2679 /** Determine if a #REQUEST needs to be proxied, and perform pre-proxy operations
2680  *
2681  * Whether a request will be proxied is determined by the attributes present
2682  * in request->config. If any of the following attributes are found, the
2683  * request may be proxied.
2684  *
2685  * The key attributes are:
2686  *   - PW_PROXY_TO_REALM          - Specifies a realm the request should be proxied to.
2687  *   - PW_HOME_SERVER_POOL        - Specifies a specific home server pool to proxy to.
2688  *   - PW_PACKET_DST_IP_ADDRESS   - Specifies a specific IPv4 home server to proxy to.
2689  *   - PW_PACKET_DST_IPV6_ADDRESS - Specifies a specific IPv6 home server to proxy to.
2690  *
2691  * Certain packet types such as #PW_CODE_STATUS_SERVER will never be proxied.
2692  *
2693  * If request should be proxied, will:
2694  *   - Add request:Proxy-State
2695  *   - Strip the current username value of its realm (depending on config)
2696  *   - Create a CHAP-Challenge from the original request vector, if one doesn't already
2697  *     exist.
2698  *   - Call the pre-process section in the current server, or in the virtual server
2699  *     associated with the home server pool we're proxying to.
2700  *
2701  * @todo A lot of this logic is RADIUS specific, and should be moved out into a protocol
2702  *      specific function.
2703  *
2704  * @param request The #REQUEST to evaluate for proxying.
2705  * @return 0 if not proxying, 1 if request should be proxied, -1 on error.
2706  */
2707 static int request_will_proxy(REQUEST *request)
2708 {
2709         int rcode, pre_proxy_type = 0;
2710         char const *realmname = NULL;
2711         VALUE_PAIR *vp, *strippedname;
2712         home_server_t *home;
2713         REALM *realm = NULL;
2714         home_pool_t *pool = NULL;
2715
2716         VERIFY_REQUEST(request);
2717
2718         if (!request->root->proxy_requests) return 0;
2719         if (request->packet->dst_port == 0) return 0;
2720         if (request->packet->code == PW_CODE_STATUS_SERVER) return 0;
2721         if (request->in_proxy_hash) return 0;
2722
2723         /*
2724          *      FIXME: for 3.0, allow this only for rejects?
2725          */
2726         if (request->reply->code != 0) return 0;
2727
2728         vp = pairfind(request->config, PW_PROXY_TO_REALM, 0, TAG_ANY);
2729         if (vp) {
2730                 realm = realm_find2(vp->vp_strvalue);
2731                 if (!realm) {
2732                         REDEBUG2("Cannot proxy to unknown realm %s",
2733                                 vp->vp_strvalue);
2734                         return 0;
2735                 }
2736
2737                 realmname = vp->vp_strvalue;
2738
2739                 /*
2740                  *      Figure out which pool to use.
2741                  */
2742                 if (request->packet->code == PW_CODE_ACCESS_REQUEST) {
2743                         pool = realm->auth_pool;
2744
2745 #ifdef WITH_ACCOUNTING
2746                 } else if (request->packet->code == PW_CODE_ACCOUNTING_REQUEST) {
2747                         pool = realm->acct_pool;
2748 #endif
2749
2750 #ifdef WITH_COA
2751                 } else if ((request->packet->code == PW_CODE_COA_REQUEST) ||
2752                            (request->packet->code == PW_CODE_DISCONNECT_REQUEST)) {
2753                         pool = realm->coa_pool;
2754 #endif
2755
2756                 } else {
2757                         return 0;
2758                 }
2759
2760         } else if ((vp = pairfind(request->config, PW_HOME_SERVER_POOL, 0, TAG_ANY)) != NULL) {
2761                 int pool_type;
2762
2763                 switch (request->packet->code) {
2764                 case PW_CODE_ACCESS_REQUEST:
2765                         pool_type = HOME_TYPE_AUTH;
2766                         break;
2767
2768 #ifdef WITH_ACCOUNTING
2769                 case PW_CODE_ACCOUNTING_REQUEST:
2770                         pool_type = HOME_TYPE_ACCT;
2771                         break;
2772 #endif
2773
2774 #ifdef WITH_COA
2775                 case PW_CODE_COA_REQUEST:
2776                 case PW_CODE_DISCONNECT_REQUEST:
2777                         pool_type = HOME_TYPE_COA;
2778                         break;
2779 #endif
2780
2781                 default:
2782                         return 0;
2783                 }
2784
2785                 pool = home_pool_byname(vp->vp_strvalue, pool_type);
2786
2787         /*
2788          *      Send it directly to a home server (i.e. NAS)
2789          */
2790         } else if (((vp = pairfind(request->config, PW_PACKET_DST_IP_ADDRESS, 0, TAG_ANY)) != NULL) ||
2791                    ((vp = pairfind(request->config, PW_PACKET_DST_IPV6_ADDRESS, 0, TAG_ANY)) != NULL)) {
2792                 VALUE_PAIR *port;
2793                 uint16_t dst_port;
2794                 fr_ipaddr_t dst_ipaddr;
2795
2796                 memset(&dst_ipaddr, 0, sizeof(dst_ipaddr));
2797
2798                 if (vp->da->attr == PW_PACKET_DST_IP_ADDRESS) {
2799                         dst_ipaddr.af = AF_INET;
2800                         dst_ipaddr.ipaddr.ip4addr.s_addr = vp->vp_ipaddr;
2801                         dst_ipaddr.prefix = 32;
2802                 } else {
2803                         dst_ipaddr.af = AF_INET6;
2804                         memcpy(&dst_ipaddr.ipaddr.ip6addr, &vp->vp_ipv6addr, sizeof(vp->vp_ipv6addr));
2805                         dst_ipaddr.prefix = 128;
2806                 }
2807
2808                 port = pairfind(request->config, PW_PACKET_DST_PORT, 0, TAG_ANY);
2809                 if (!port) {
2810                 dst_port = PW_COA_UDP_PORT;
2811                 } else {
2812                         dst_port = vp->vp_integer;
2813                 }
2814
2815                 /*
2816                  *      Nothing does CoA over TCP.
2817                  */
2818                 home = home_server_find(&dst_ipaddr, dst_port, IPPROTO_UDP);
2819                 if (!home) {
2820                         char buffer[256];
2821
2822                         WARN("No such CoA home server %s port %u",
2823                              inet_ntop(dst_ipaddr.af, &dst_ipaddr.ipaddr, buffer, sizeof(buffer)),
2824                              (unsigned int) dst_port);
2825                         return 0;
2826                 }
2827
2828                 goto do_home;
2829
2830         } else {
2831                 return 0;
2832         }
2833
2834         if (!pool) {
2835                 RWDEBUG2("Cancelling proxy as no home pool exists");
2836                 return 0;
2837         }
2838
2839         if (request->listener->synchronous) {
2840                 WARN("Cannot proxy a request which is from a 'synchronous' socket");
2841                 return 0;
2842         }
2843
2844         request->home_pool = pool;
2845
2846         home = home_server_ldb(realmname, pool, request);
2847
2848         if (!home) {
2849                 REDEBUG2("Failed to find live home server: Cancelling proxy");
2850                 return 0;
2851         }
2852
2853 do_home:
2854         home_server_update_request(home, request);
2855
2856 #ifdef WITH_COA
2857         /*
2858          *      Once we've decided to proxy a request, we cannot send
2859          *      a CoA packet.  So we free up any CoA packet here.
2860          */
2861         if (request->coa) request_done(request->coa, FR_ACTION_DONE);
2862 #endif
2863
2864         /*
2865          *      Remember that we sent the request to a Realm.
2866          */
2867         if (realmname) pairmake_packet("Realm", realmname, T_OP_EQ);
2868
2869         /*
2870          *      Strip the name, if told to.
2871          *
2872          *      Doing it here catches the case of proxied tunneled
2873          *      requests.
2874          */
2875         if (realm && (realm->strip_realm == true) &&
2876            (strippedname = pairfind(request->proxy->vps, PW_STRIPPED_USER_NAME, 0, TAG_ANY)) != NULL) {
2877                 /*
2878                  *      If there's a Stripped-User-Name attribute in
2879                  *      the request, then use THAT as the User-Name
2880                  *      for the proxied request, instead of the
2881                  *      original name.
2882                  *
2883                  *      This is done by making a copy of the
2884                  *      Stripped-User-Name attribute, turning it into
2885                  *      a User-Name attribute, deleting the
2886                  *      Stripped-User-Name and User-Name attributes
2887                  *      from the vps list, and making the new
2888                  *      User-Name the head of the vps list.
2889                  */
2890                 vp = pairfind(request->proxy->vps, PW_USER_NAME, 0, TAG_ANY);
2891                 if (!vp) {
2892                         vp_cursor_t cursor;
2893                         vp = radius_paircreate(NULL, NULL,
2894                                                PW_USER_NAME, 0);
2895                         rad_assert(vp != NULL); /* handled by above function */
2896                         /* Insert at the START of the list */
2897                         /* FIXME: Can't make assumptions about ordering */
2898                         fr_cursor_init(&cursor, &vp);
2899                         fr_cursor_merge(&cursor, request->proxy->vps);
2900                         request->proxy->vps = vp;
2901                 }
2902                 pairstrcpy(vp, strippedname->vp_strvalue);
2903
2904                 /*
2905                  *      Do NOT delete Stripped-User-Name.
2906                  */
2907         }
2908
2909         /*
2910          *      If there is no PW_CHAP_CHALLENGE attribute but
2911          *      there is a PW_CHAP_PASSWORD we need to add it
2912          *      since we can't use the request authenticator
2913          *      anymore - we changed it.
2914          */
2915         if ((request->packet->code == PW_CODE_ACCESS_REQUEST) &&
2916             pairfind(request->proxy->vps, PW_CHAP_PASSWORD, 0, TAG_ANY) &&
2917             pairfind(request->proxy->vps, PW_CHAP_CHALLENGE, 0, TAG_ANY) == NULL) {
2918                 vp = radius_paircreate(request->proxy, &request->proxy->vps, PW_CHAP_CHALLENGE, 0);
2919                 pairmemcpy(vp, request->packet->vector, sizeof(request->packet->vector));
2920         }
2921
2922         /*
2923          *      The RFC's say we have to do this, but FreeRADIUS
2924          *      doesn't need it.
2925          */
2926         vp = radius_paircreate(request->proxy, &request->proxy->vps, PW_PROXY_STATE, 0);
2927         pairsprintf(vp, "%u", request->packet->id);
2928
2929         /*
2930          *      Should be done BEFORE inserting into proxy hash, as
2931          *      pre-proxy may use this information, or change it.
2932          */
2933         request->proxy->code = request->packet->code;
2934
2935         /*
2936          *      Call the pre-proxy routines.
2937          */
2938         vp = pairfind(request->config, PW_PRE_PROXY_TYPE, 0, TAG_ANY);
2939         if (vp) {
2940                 DICT_VALUE const *dval = dict_valbyattr(vp->da->attr, vp->da->vendor, vp->vp_integer);
2941                 /* Must be a validation issue */
2942                 rad_assert(dval);
2943                 RDEBUG2("Found Pre-Proxy-Type %s", dval->name);
2944                 pre_proxy_type = vp->vp_integer;
2945         }
2946
2947         /*
2948          *      home_pool may be NULL when originating CoA packets,
2949          *      because they go directly to an IP address.
2950          */
2951         if (request->home_pool && request->home_pool->virtual_server) {
2952                 char const *old_server = request->server;
2953
2954                 request->server = request->home_pool->virtual_server;
2955
2956                 RDEBUG2("server %s {", request->server);
2957                 RINDENT();
2958                 rcode = process_pre_proxy(pre_proxy_type, request);
2959                 REXDENT();
2960                 RDEBUG2("}");
2961
2962                 request->server = old_server;
2963         } else {
2964                 rcode = process_pre_proxy(pre_proxy_type, request);
2965         }
2966
2967         switch (rcode) {
2968         case RLM_MODULE_FAIL:
2969         case RLM_MODULE_INVALID:
2970         case RLM_MODULE_NOTFOUND:
2971         case RLM_MODULE_USERLOCK:
2972         default:
2973                 /* FIXME: debug print failed stuff */
2974                 return -1;
2975
2976         case RLM_MODULE_REJECT:
2977         case RLM_MODULE_HANDLED:
2978                 return 0;
2979
2980         /*
2981          *      Only proxy the packet if the pre-proxy code succeeded.
2982          */
2983         case RLM_MODULE_NOOP:
2984         case RLM_MODULE_OK:
2985         case RLM_MODULE_UPDATED:
2986                 return 1;
2987         }
2988 }
2989
2990 static int request_proxy(REQUEST *request, int retransmit)
2991 {
2992         char buffer[128];
2993
2994         VERIFY_REQUEST(request);
2995
2996         rad_assert(request->parent == NULL);
2997         rad_assert(request->home_server != NULL);
2998
2999         if (request->master_state == REQUEST_STOP_PROCESSING) return 0;
3000
3001 #ifdef WITH_COA
3002         if (request->coa) {
3003                 RWDEBUG("Cannot proxy and originate CoA packets at the same time.  Cancelling CoA request");
3004                 request_done(request->coa, FR_ACTION_DONE);
3005         }
3006 #endif
3007
3008         /*
3009          *      The request may need sending to a virtual server.
3010          *      This code is more than a little screwed up.  The rest
3011          *      of the state machine doesn't handle parent / child
3012          *      relationships well.  i.e. if the child request takes
3013          *      too long, the core will mark the *parent* as "stop
3014          *      processing".  And the child will continue without
3015          *      knowing anything...
3016          *
3017          *      So, we have some horrible hacks to get around that.
3018          */
3019         if (request->home_server->server) {
3020                 REQUEST *fake;
3021
3022                 if (request->packet->dst_port == 0) {
3023                         WARN("Cannot proxy an internal request");
3024                         return 0;
3025                 }
3026
3027                 DEBUG("Proxying to virtual server %s",
3028                       request->home_server->server);
3029
3030                 /*
3031                  *      Packets to virtual serrers don't get
3032                  *      retransmissions sent to them.  And the virtual
3033                  *      server is run ONLY if we have no child
3034                  *      threads, or we're running in a child thread.
3035                  */
3036                 rad_assert(retransmit == 0);
3037                 rad_assert(!spawn_flag || !we_are_master());
3038
3039                 fake = request_alloc_fake(request);
3040
3041                 fake->packet->vps = paircopy(fake->packet, request->packet->vps);
3042                 talloc_free(request->proxy);
3043
3044                 fake->server = request->home_server->server;
3045                 fake->handle = request->handle;
3046                 fake->process = NULL; /* should never be run for anything */
3047
3048                 /*
3049                  *      Run the virtual server.
3050                  */
3051                 request_running(fake, FR_ACTION_RUN);
3052
3053                 request->proxy = talloc_steal(request, fake->packet);
3054                 fake->packet = NULL;
3055                 request->proxy_reply = talloc_steal(request, fake->reply);
3056                 fake->reply = NULL;
3057
3058                 talloc_free(fake);
3059
3060                 /*
3061                  *      No reply code, toss the reply we have,
3062                  *      and do post-proxy-type Fail.
3063                  */
3064                 if (!request->proxy_reply->code) {
3065                         TALLOC_FREE(request->proxy_reply);
3066                         setup_post_proxy_fail(request);
3067                 }
3068
3069                 /*
3070                  *      Do the proxy reply (if any)
3071                  */
3072                 if (process_proxy_reply(request, request->proxy_reply)) {
3073                         request->handle(request);
3074                 }
3075
3076                 return -1;      /* so we call request_finish */
3077         }
3078
3079         /*
3080          *      We're actually sending a proxied packet.  Do that now.
3081          */
3082         if (!request->in_proxy_hash && !insert_into_proxy_hash(request)) {
3083                 RPROXY("Failed to insert request into the proxy list");
3084                 return -1;
3085         }
3086
3087         rad_assert(request->proxy->id >= 0);
3088
3089         if (debug_flag) {
3090                 struct timeval *response_window;
3091
3092                 response_window = request_response_window(request);
3093
3094 #ifdef WITH_TLS
3095                 if (request->home_server->tls) {
3096                         RDEBUG2("Proxying request to home server %s port %d (TLS) timeout %d.%06d",
3097                                 inet_ntop(request->proxy->dst_ipaddr.af,
3098                                           &request->proxy->dst_ipaddr.ipaddr,
3099                                           buffer, sizeof(buffer)),
3100                                 request->proxy->dst_port,
3101                                 (int) response_window->tv_sec, (int) response_window->tv_usec);
3102                 } else
3103 #endif
3104                         RDEBUG2("Proxying request to home server %s port %d timeout %d.%06d",
3105                                 inet_ntop(request->proxy->dst_ipaddr.af,
3106                                           &request->proxy->dst_ipaddr.ipaddr,
3107                                           buffer, sizeof(buffer)),
3108                                 request->proxy->dst_port,
3109                                 (int) response_window->tv_sec, (int) response_window->tv_usec);
3110
3111
3112         }
3113
3114         gettimeofday(&request->proxy_retransmit, NULL);
3115         if (!retransmit) {
3116                 request->proxy->timestamp = request->proxy_retransmit;
3117         }
3118         request->home_server->last_packet_sent = request->proxy_retransmit.tv_sec;
3119
3120         /*
3121          *      Encode the packet before we do anything else.
3122          */
3123         request->proxy_listener->encode(request->proxy_listener, request);
3124         debug_packet(request, request->proxy, false);
3125
3126         /*
3127          *      Set the state function, then the state, no child, and
3128          *      send the packet.
3129          */
3130         request->process = proxy_wait_for_reply;
3131         request->child_state = REQUEST_PROXIED;
3132         NO_CHILD_THREAD;
3133
3134         /*
3135          *      And send the packet.
3136          */
3137         request->proxy_listener->send(request->proxy_listener, request);
3138         return 1;
3139 }
3140
3141 /*
3142  *      Proxy the packet as if it was new.
3143  */
3144 static int request_proxy_anew(REQUEST *request)
3145 {
3146         home_server_t *home;
3147
3148         VERIFY_REQUEST(request);
3149
3150         /*
3151          *      Delete the request from the proxy list.
3152          *
3153          *      The packet list code takes care of ensuring that IDs
3154          *      aren't reused until all 256 IDs have been used.  So
3155          *      there's a 1/256 chance of re-using the same ID when
3156          *      we're sending to the same home server.  Which is
3157          *      acceptable.
3158          */
3159         remove_from_proxy_hash(request);
3160
3161         /*
3162          *      Find a live home server for the request.
3163          */
3164         home = home_server_ldb(NULL, request->home_pool, request);
3165         if (!home) {
3166                 REDEBUG2("Failed to find live home server for request");
3167         post_proxy_fail:
3168                 if (setup_post_proxy_fail(request)) {
3169                         request_queue_or_run(request, proxy_running);
3170                 } else {
3171                         gettimeofday(&request->reply->timestamp, NULL);
3172                         request_cleanup_delay_init(request);
3173                 }
3174                 return 0;
3175         }
3176         home_server_update_request(home, request);
3177
3178         if (!insert_into_proxy_hash(request)) {
3179                 RPROXY("Failed to insert retransmission into the proxy list");
3180                 goto post_proxy_fail;
3181         }
3182
3183         /*
3184          *      Free the old packet, to force re-encoding
3185          */
3186         talloc_free(request->proxy->data);
3187         request->proxy->data = NULL;
3188         request->proxy->data_len = 0;
3189
3190 #ifdef WITH_ACCOUNTING
3191         /*
3192          *      Update the Acct-Delay-Time attribute.
3193          */
3194         if (request->packet->code == PW_CODE_ACCOUNTING_REQUEST) {
3195                 VALUE_PAIR *vp;
3196
3197                 vp = pairfind(request->proxy->vps, PW_ACCT_DELAY_TIME, 0, TAG_ANY);
3198                 if (!vp) vp = radius_paircreate(request->proxy,
3199                                                 &request->proxy->vps,
3200                                                 PW_ACCT_DELAY_TIME, 0);
3201                 if (vp) {
3202                         struct timeval now;
3203
3204                         gettimeofday(&now, NULL);
3205                         vp->vp_integer += now.tv_sec - request->proxy_retransmit.tv_sec;
3206                 }
3207         }
3208 #endif
3209
3210         if (request_proxy(request, 1) != 1) goto post_proxy_fail;
3211
3212         return 1;
3213 }
3214
3215
3216 /** Ping a home server.
3217  *
3218  */
3219 static void request_ping(REQUEST *request, int action)
3220 {
3221         home_server_t *home = request->home_server;
3222         char buffer[128];
3223
3224         VERIFY_REQUEST(request);
3225
3226         TRACE_STATE_MACHINE;
3227         ASSERT_MASTER;
3228
3229         switch (action) {
3230         case FR_ACTION_TIMER:
3231                 ERROR("No response to status check %d for home server %s port %d",
3232                        request->number,
3233                        inet_ntop(request->proxy->dst_ipaddr.af,
3234                                  &request->proxy->dst_ipaddr.ipaddr,
3235                                  buffer, sizeof(buffer)),
3236                        request->proxy->dst_port);
3237                 break;
3238
3239         case FR_ACTION_PROXY_REPLY:
3240                 rad_assert(request->in_proxy_hash);
3241
3242                 request->home_server->num_received_pings++;
3243                 RPROXY("Received response to status check %d (%d in current sequence)",
3244                        request->number, home->num_received_pings);
3245
3246                 /*
3247                  *      Remove the request from any hashes
3248                  */
3249                 fr_event_delete(el, &request->ev);
3250                 remove_from_proxy_hash(request);
3251
3252                 /*
3253                  *      The control socket may have marked the home server as
3254                  *      alive.  OR, it may have suddenly started responding to
3255                  *      requests again.  If so, don't re-do the "make alive"
3256                  *      work.
3257                  */
3258                 if (home->state == HOME_STATE_ALIVE) break;
3259
3260                 /*
3261                  *      It's dead, and we haven't received enough ping
3262                  *      responses to mark it "alive".  Wait a bit.
3263                  *
3264                  *      If it's zombie, we mark it alive immediately.
3265                  */
3266                 if ((home->state == HOME_STATE_IS_DEAD) &&
3267                     (home->num_received_pings < home->num_pings_to_alive)) {
3268                         return;
3269                 }
3270
3271                 /*
3272                  *      Mark it alive and delete any outstanding
3273                  *      pings.
3274                  */
3275                 home->state = HOME_STATE_ALIVE;
3276                 home->response_timeouts = 0;
3277                 exec_trigger(request, home->cs, "home_server.alive", false);
3278                 home->currently_outstanding = 0;
3279                 home->num_sent_pings = 0;
3280                 home->num_received_pings = 0;
3281                 gettimeofday(&home->revive_time, NULL);
3282
3283                 fr_event_delete(el, &home->ev);
3284
3285                 RPROXY("Marking home server %s port %d alive",
3286                        inet_ntop(request->proxy->dst_ipaddr.af,
3287                                  &request->proxy->dst_ipaddr.ipaddr,
3288                                  buffer, sizeof(buffer)),
3289                        request->proxy->dst_port);
3290                 break;
3291
3292         default:
3293                 RDEBUG3("%s: Ignoring action %s", __FUNCTION__, action_codes[action]);
3294                 break;
3295         }
3296
3297         rad_assert(!request->in_request_hash);
3298         rad_assert(request->ev == NULL);
3299         NO_CHILD_THREAD;
3300         request_done(request, FR_ACTION_DONE);
3301 }
3302
3303 /*
3304  *      Called from start of zombie period, OR after control socket
3305  *      marks the home server dead.
3306  */
3307 static void ping_home_server(void *ctx)
3308 {
3309         home_server_t *home = talloc_get_type_abort(ctx, home_server_t);
3310         REQUEST *request;
3311         VALUE_PAIR *vp;
3312         struct timeval when, now;
3313
3314         if ((home->state == HOME_STATE_ALIVE) ||
3315 #ifdef WITH_TCP
3316             (home->proto == IPPROTO_TCP) ||
3317 #endif
3318             (home->ev != NULL)) {
3319                 return;
3320         }
3321
3322         gettimeofday(&now, NULL);
3323         ASSERT_MASTER;
3324
3325         /*
3326          *      We've run out of zombie time.  Mark it dead.
3327          */
3328         if (home->state == HOME_STATE_ZOMBIE) {
3329                 when = home->zombie_period_start;
3330                 when.tv_sec += home->zombie_period;
3331
3332                 if (timercmp(&when, &now, <)) {
3333                         DEBUG("PING: Zombie period is over for home server %s", home->log_name);
3334                         mark_home_server_dead(home, &now);
3335                 }
3336         }
3337
3338         /*
3339          *      We're not supposed to be pinging it.  Just wake up
3340          *      when we're supposed to mark it dead.
3341          */
3342         if (home->ping_check == HOME_PING_CHECK_NONE) {
3343                 if (home->state == HOME_STATE_ZOMBIE) {
3344                         when = home->zombie_period_start;
3345                         when.tv_sec += home->zombie_period;
3346                         INSERT_EVENT(ping_home_server, home);
3347                 }
3348
3349                 /*
3350                  *      Else mark_home_server_dead will set a timer
3351                  *      for revive_interval.
3352                  */
3353                 return;
3354         }
3355
3356
3357         request = request_alloc(NULL);
3358         if (!request) return;
3359         request->number = request_num_counter++;
3360         NO_CHILD_THREAD;
3361
3362         request->proxy = rad_alloc(request, true);
3363         rad_assert(request->proxy != NULL);
3364
3365         if (home->ping_check == HOME_PING_CHECK_STATUS_SERVER) {
3366                 request->proxy->code = PW_CODE_STATUS_SERVER;
3367
3368                 pairmake(request->proxy, &request->proxy->vps,
3369                          "Message-Authenticator", "0x00", T_OP_SET);
3370
3371         } else if (home->type == HOME_TYPE_AUTH) {
3372                 request->proxy->code = PW_CODE_ACCESS_REQUEST;
3373
3374                 pairmake(request->proxy, &request->proxy->vps,
3375                          "User-Name", home->ping_user_name, T_OP_SET);
3376                 pairmake(request->proxy, &request->proxy->vps,
3377                          "User-Password", home->ping_user_password, T_OP_SET);
3378                 pairmake(request->proxy, &request->proxy->vps,
3379                          "Service-Type", "Authenticate-Only", T_OP_SET);
3380                 pairmake(request->proxy, &request->proxy->vps,
3381                          "Message-Authenticator", "0x00", T_OP_SET);
3382
3383         } else {
3384 #ifdef WITH_ACCOUNTING
3385                 request->proxy->code = PW_CODE_ACCOUNTING_REQUEST;
3386
3387                 pairmake(request->proxy, &request->proxy->vps,
3388                          "User-Name", home->ping_user_name, T_OP_SET);
3389                 pairmake(request->proxy, &request->proxy->vps,
3390                          "Acct-Status-Type", "Stop", T_OP_SET);
3391                 pairmake(request->proxy, &request->proxy->vps,
3392                          "Acct-Session-Id", "00000000", T_OP_SET);
3393                 vp = pairmake(request->proxy, &request->proxy->vps,
3394                               "Event-Timestamp", "0", T_OP_SET);
3395                 vp->vp_date = now.tv_sec;
3396 #else
3397                 rad_assert("Internal sanity check failed");
3398 #endif
3399         }
3400
3401         vp = pairmake(request->proxy, &request->proxy->vps,
3402                       "NAS-Identifier", "", T_OP_SET);
3403         if (vp) {
3404                 pairsprintf(vp, "Status Check %u. Are you alive?",
3405                             home->num_sent_pings);
3406         }
3407
3408         request->proxy->src_ipaddr = home->src_ipaddr;
3409         request->proxy->dst_ipaddr = home->ipaddr;
3410         request->proxy->dst_port = home->port;
3411         request->home_server = home;
3412 #ifdef DEBUG_STATE_MACHINE
3413         if (debug_flag) printf("(%u) ********\tSTATE %s C-%s -> C-%s\t********\n", request->number, __FUNCTION__,
3414                                child_state_names[request->child_state],
3415                                child_state_names[REQUEST_DONE]);
3416         if (debug_flag) printf("(%u) ********\tNEXT-STATE %s -> %s\n", request->number, __FUNCTION__, "request_ping");
3417 #endif
3418 #ifdef HAVE_PTHREAD_H
3419         rad_assert(request->child_pid == NO_SUCH_CHILD_PID);
3420 #endif
3421         request->child_state = REQUEST_PROXIED;
3422         request->process = request_ping;
3423
3424         rad_assert(request->proxy_listener == NULL);
3425
3426         if (!insert_into_proxy_hash(request)) {
3427                 RPROXY("Failed to insert status check %d into proxy list.  Discarding it.",
3428                        request->number);
3429
3430                 rad_assert(!request->in_request_hash);
3431                 rad_assert(!request->in_proxy_hash);
3432                 rad_assert(request->ev == NULL);
3433                 talloc_free(request);
3434                 return;
3435         }
3436
3437         /*
3438          *      Set up the timer callback.
3439          */
3440         when = now;
3441         when.tv_sec += home->ping_timeout;
3442
3443         DEBUG("PING: Waiting %u seconds for response to ping",
3444               home->ping_timeout);
3445
3446         STATE_MACHINE_TIMER(FR_ACTION_TIMER);
3447         home->num_sent_pings++;
3448
3449         rad_assert(request->proxy_listener != NULL);
3450         request->proxy_listener->send(request->proxy_listener,
3451                                       request);
3452
3453         /*
3454          *      Add +/- 2s of jitter, as suggested in RFC 3539
3455          *      and in the Issues and Fixes draft.
3456          */
3457         home->when = now;
3458         home->when.tv_sec += home->ping_interval;
3459
3460         add_jitter(&home->when);
3461
3462         DEBUG("PING: Next status packet in %u seconds", home->ping_interval);
3463         INSERT_EVENT(ping_home_server, home);
3464 }
3465
3466 static void home_trigger(home_server_t *home, char const *trigger)
3467 {
3468         REQUEST *my_request;
3469         RADIUS_PACKET *my_packet;
3470
3471         my_request = talloc_zero(NULL, REQUEST);
3472         my_packet = talloc_zero(my_request, RADIUS_PACKET);
3473         my_request->proxy = my_packet;
3474         my_packet->dst_ipaddr = home->ipaddr;
3475         my_packet->src_ipaddr = home->src_ipaddr;
3476
3477         exec_trigger(my_request, home->cs, trigger, false);
3478         talloc_free(my_request);
3479 }
3480
3481 static void mark_home_server_zombie(home_server_t *home, struct timeval *now, struct timeval *response_window)
3482 {
3483         time_t start;
3484         char buffer[128];
3485
3486         ASSERT_MASTER;
3487
3488         rad_assert((home->state == HOME_STATE_ALIVE) ||
3489                    (home->state == HOME_STATE_UNKNOWN));
3490
3491 #ifdef WITH_TCP
3492         if (home->proto == IPPROTO_TCP) {
3493                 WARN("Not marking TCP server %s zombie", home->log_name);
3494                 return;
3495         }
3496 #endif
3497
3498         /*
3499          *      We've received a real packet recently.  Don't mark the
3500          *      server as zombie until we've received NO packets for a
3501          *      while.  The "1/4" of zombie period was chosen rather
3502          *      arbitrarily.  It's a balance between too short, which
3503          *      gives quick fail-over and fail-back, or too long,
3504          *      where the proxy still sends packets to an unresponsive
3505          *      home server.
3506          */
3507         start = now->tv_sec - ((home->zombie_period + 3) / 4);
3508         if (home->last_packet_recv >= start) {
3509                 DEBUG("Recieved reply from home server %d seconds ago.  Might not be zombie.",
3510                       (int) (now->tv_sec - home->last_packet_recv));
3511                 return;
3512         }
3513
3514         home->state = HOME_STATE_ZOMBIE;
3515         home_trigger(home, "home_server.zombie");
3516
3517         /*
3518          *      Set the home server to "zombie", as of the time
3519          *      calculated above.
3520          */
3521         home->zombie_period_start.tv_sec = start;
3522         home->zombie_period_start.tv_usec = USEC / 2;
3523
3524         fr_event_delete(el, &home->ev);
3525
3526         home->num_sent_pings = 0;
3527         home->num_received_pings = 0;
3528
3529         PROXY( "Marking home server %s port %d as zombie (it has not responded in %d.%06d seconds).",
3530                inet_ntop(home->ipaddr.af, &home->ipaddr.ipaddr,
3531                          buffer, sizeof(buffer)),
3532                home->port, (int) response_window->tv_sec, (int) response_window->tv_usec);
3533
3534         ping_home_server(home);
3535 }
3536
3537
3538 void revive_home_server(void *ctx)
3539 {
3540         home_server_t *home = talloc_get_type_abort(ctx, home_server_t);
3541         char buffer[128];
3542
3543 #ifdef WITH_TCP
3544         rad_assert(home->proto != IPPROTO_TCP);
3545 #endif
3546
3547         home->state = HOME_STATE_ALIVE;
3548         home->response_timeouts = 0;
3549         home_trigger(home, "home_server.alive");
3550         home->currently_outstanding = 0;
3551         gettimeofday(&home->revive_time, NULL);
3552
3553         /*
3554          *      Delete any outstanding events.
3555          */
3556         ASSERT_MASTER;
3557         if (home->ev) fr_event_delete(el, &home->ev);
3558
3559         PROXY( "Marking home server %s port %d alive again... we have no idea if it really is alive or not.",
3560                inet_ntop(home->ipaddr.af, &home->ipaddr.ipaddr,
3561                          buffer, sizeof(buffer)),
3562                home->port);
3563 }
3564
3565 void mark_home_server_dead(home_server_t *home, struct timeval *when)
3566 {
3567         int previous_state = home->state;
3568         char buffer[128];
3569
3570 #ifdef WITH_TCP
3571         if (home->proto == IPPROTO_TCP) {
3572                 WARN("Not marking TCP server dead");
3573                 return;
3574         }
3575 #endif
3576
3577         PROXY( "Marking home server %s port %d as dead.",
3578                inet_ntop(home->ipaddr.af, &home->ipaddr.ipaddr,
3579                          buffer, sizeof(buffer)),
3580                home->port);
3581
3582         home->state = HOME_STATE_IS_DEAD;
3583         home_trigger(home, "home_server.dead");
3584
3585         if (home->ping_check != HOME_PING_CHECK_NONE) {
3586                 /*
3587                  *      If the control socket marks us dead, start
3588                  *      pinging.  Otherwise, we already started
3589                  *      pinging when it was marked "zombie".
3590                  */
3591                 if (previous_state == HOME_STATE_ALIVE) {
3592                         ping_home_server(home);
3593                 } else {
3594                         DEBUG("PING: Already pinging home server %s", home->log_name);
3595                 }
3596
3597         } else {
3598                 /*
3599                  *      Revive it after a fixed period of time.  This
3600                  *      is very, very, bad.
3601                  */
3602                 home->when = *when;
3603                 home->when.tv_sec += home->revive_interval;
3604
3605                 DEBUG("PING: Reviving home server %s in %u seconds", home->log_name, home->revive_interval);
3606                 ASSERT_MASTER;
3607                 INSERT_EVENT(revive_home_server, home);
3608         }
3609 }
3610
3611 /** Wait for a reply after proxying a request.
3612  *
3613  *  Retransmit the proxied packet, or time out and go to
3614  *  proxy_no_reply.  Mark the home server unresponsive, etc.
3615  *
3616  *  If we do receive a reply, we transition to proxy_running.
3617  *
3618  *  \dot
3619  *      digraph proxy_wait_for_reply {
3620  *              proxy_wait_for_reply;
3621  *
3622  *              proxy_wait_for_reply -> retransmit_proxied_request [ label = "DUP", arrowhead = "none" ];
3623  *              proxy_wait_for_reply -> proxy_no_reply [ label = "TIMER >= response_window" ];
3624  *              proxy_wait_for_reply -> timer [ label = "TIMER < max_request_time" ];
3625  *              proxy_wait_for_reply -> proxy_running [ label = "PROXY_REPLY" arrowhead = "none"];
3626  *              proxy_wait_for_reply -> done [ label = "TIMER >= max_request_time" ];
3627  *      }
3628  *  \enddot
3629  */
3630 static void proxy_wait_for_reply(REQUEST *request, int action)
3631 {
3632         struct timeval now, when;
3633         struct timeval *response_window = NULL;
3634         home_server_t *home = request->home_server;
3635         char buffer[128];
3636
3637         VERIFY_REQUEST(request);
3638
3639         TRACE_STATE_MACHINE;
3640         CHECK_FOR_STOP;
3641
3642         rad_assert(request->packet->code != PW_CODE_STATUS_SERVER);
3643         rad_assert(request->home_server != NULL);
3644
3645         gettimeofday(&now, NULL);
3646
3647         switch (action) {
3648         case FR_ACTION_DUP:
3649                 /*
3650                  *      We have a reply, ignore the retransmit.
3651                  */
3652                 if (request->proxy_reply) return;
3653
3654                 /*
3655                  *      The request was proxied to a virtual server.
3656                  *      Ignore the retransmit.
3657                  */
3658                 if (request->home_server->server) return;
3659
3660                 /*
3661                  *      Use a new connection when the home server is
3662                  *      dead, or when there's no proxy listener, or
3663                  *      when the listener is failed or dead.
3664                  *
3665                  *      If the listener is known or frozen, use it for
3666                  *      retransmits.
3667                  */
3668                 if ((home->state == HOME_STATE_IS_DEAD) ||
3669                     !request->proxy_listener ||
3670                     (request->proxy_listener->status >= RAD_LISTEN_STATUS_EOL)) {
3671                         request_proxy_anew(request);
3672                         return;
3673                 }
3674
3675 #ifdef WITH_TCP
3676                 if (home->proto == IPPROTO_TCP) {
3677                         DEBUG2("Suppressing duplicate proxied request (tcp) to home server %s port %d proto TCP - ID: %d",
3678                                inet_ntop(request->proxy->dst_ipaddr.af,
3679                                          &request->proxy->dst_ipaddr.ipaddr,
3680                                          buffer, sizeof(buffer)),
3681                                request->proxy->dst_port,
3682                                request->proxy->id);
3683                         return;
3684                 }
3685 #endif
3686
3687                 /*
3688                  *      More than one retransmit a second is stupid,
3689                  *      and should be suppressed by the proxy.
3690                  */
3691                 when = request->proxy_retransmit;
3692                 when.tv_sec++;
3693
3694                 if (timercmp(&now, &when, <)) {
3695                         DEBUG2("Suppressing duplicate proxied request (too fast) to home server %s port %d proto TCP - ID: %d",
3696                                inet_ntop(request->proxy->dst_ipaddr.af,
3697                                          &request->proxy->dst_ipaddr.ipaddr,
3698                                          buffer, sizeof(buffer)),
3699                                request->proxy->dst_port,
3700                                request->proxy->id);
3701                         return;
3702                 }
3703
3704 #ifdef WITH_ACCOUNTING
3705                 /*
3706                  *      If we update the Acct-Delay-Time, we need to
3707                  *      get a new ID.
3708                  */
3709                 if ((request->packet->code == PW_CODE_ACCOUNTING_REQUEST) &&
3710                     pairfind(request->proxy->vps, PW_ACCT_DELAY_TIME, 0, TAG_ANY)) {
3711                         request_proxy_anew(request);
3712                         return;
3713                 }
3714 #endif
3715
3716                 RDEBUG2("Sending duplicate proxied request to home server %s port %d - ID: %d",
3717                         inet_ntop(request->proxy->dst_ipaddr.af,
3718                                   &request->proxy->dst_ipaddr.ipaddr,
3719                                   buffer, sizeof(buffer)),
3720                         request->proxy->dst_port,
3721                         request->proxy->id);
3722                 request->num_proxied_requests++;
3723
3724                 rad_assert(request->proxy_listener != NULL);
3725                 FR_STATS_TYPE_INC(home->stats.total_requests);
3726                 home->last_packet_sent = now.tv_sec;
3727                 request->proxy_retransmit = now;
3728                 debug_packet(request, request->proxy, false);
3729                 request->proxy_listener->send(request->proxy_listener, request);
3730                 break;
3731
3732         case FR_ACTION_TIMER:
3733                 response_window = request_response_window(request);
3734
3735 #ifdef WITH_TCP
3736                 if (!request->proxy_listener ||
3737                     (request->proxy_listener->status >= RAD_LISTEN_STATUS_EOL)) {
3738                         remove_from_proxy_hash(request);
3739
3740                         when = request->packet->timestamp;
3741                         when.tv_sec += request->root->max_request_time;
3742
3743                         if (timercmp(&when, &now, >)) {
3744                                 RDEBUG("Waiting for client retransmission in order to do a proxy retransmit");
3745                                 STATE_MACHINE_TIMER(FR_ACTION_TIMER);
3746                                 return;
3747                         }
3748                 } else
3749 #endif
3750                 {
3751                         /*
3752                          *      Wake up "response_window" time in the future.
3753                          *      i.e. when MY packet hasn't received a response.
3754                          *
3755                          *      Note that we DO NOT mark the home server as
3756                          *      zombie if it doesn't respond to us.  It may be
3757                          *      responding to other (better looking) packets.
3758                          */
3759                         when = request->proxy->timestamp;
3760                         timeradd(&when, response_window, &when);
3761
3762                         /*
3763                          *      Not at the response window.  Set the timer for
3764                          *      that.
3765                          */
3766                         if (timercmp(&when, &now, >)) {
3767                                 struct timeval diff;
3768                                 timersub(&when, &now, &diff);
3769
3770                                 RDEBUG("Expecting proxy response no later than %d.%06d seconds from now",
3771                                        (int) diff.tv_sec, (int) diff.tv_usec);
3772                                 STATE_MACHINE_TIMER(FR_ACTION_TIMER);
3773                                 return;
3774                         }
3775                 }
3776
3777                 RDEBUG("No proxy response, giving up on request and marking it done");
3778
3779                 /*
3780                  *      If we haven't received any packets for
3781                  *      "response_window", then mark the home server
3782                  *      as zombie.
3783                  *
3784                  *      If the connection is TCP, then another
3785                  *      "watchdog timer" function takes care of pings,
3786                  *      etc.  So we don't need to do it here.
3787                  *
3788                  *      This check should really be part of a home
3789                  *      server state machine.
3790                  */
3791                 if (((home->state == HOME_STATE_ALIVE) ||
3792                      (home->state == HOME_STATE_UNKNOWN))
3793 #ifdef WITH_TCP
3794                     && (home->proto != IPPROTO_TCP)
3795 #endif
3796                         ) {
3797                         home->response_timeouts++;
3798                         if (home->response_timeouts >= home->max_response_timeouts)
3799                                 mark_home_server_zombie(home, &now, response_window);
3800                 }
3801
3802                 FR_STATS_TYPE_INC(home->stats.total_timeouts);
3803                 if (home->type == HOME_TYPE_AUTH) {
3804                         if (request->proxy_listener) FR_STATS_TYPE_INC(request->proxy_listener->stats.total_timeouts);
3805                         FR_STATS_TYPE_INC(proxy_auth_stats.total_timeouts);
3806                 }
3807 #ifdef WITH_ACCT
3808                 else if (home->type == HOME_TYPE_ACCT) {
3809                         if (request->proxy_listener) FR_STATS_TYPE_INC(request->proxy_listener->stats.total_timeouts);
3810                         FR_STATS_TYPE_INC(proxy_acct_stats.total_timeouts);
3811                 }
3812 #endif
3813
3814                 /*
3815                  *      There was no response within the window.  Stop
3816                  *      the request.  If the client retransmitted, it
3817                  *      may have failed over to another home server.
3818                  *      But that one may be dead, too.
3819                  *
3820                  *      The extra verbose message if we have a username,
3821                  *      is extremely useful if the proxy is part of a chain
3822                  *      and the final home server, is not the one we're
3823                  *      proxying to.
3824                  */
3825                 if (request->username) {
3826                         RERROR("Failing proxied request for user \"%s\", due to lack of any response from home "
3827                                "server %s port %d",
3828                                request->username->vp_strvalue,
3829                                inet_ntop(request->proxy->dst_ipaddr.af,
3830                                          &request->proxy->dst_ipaddr.ipaddr,
3831                                          buffer, sizeof(buffer)),
3832                                request->proxy->dst_port);
3833                 } else {
3834                         RERROR("Failing proxied request, due to lack of any response from home server %s port %d",
3835                                inet_ntop(request->proxy->dst_ipaddr.af,
3836                                          &request->proxy->dst_ipaddr.ipaddr,
3837                                          buffer, sizeof(buffer)),
3838                                request->proxy->dst_port);
3839                 }
3840
3841                 if (setup_post_proxy_fail(request)) {
3842                         request_queue_or_run(request, proxy_no_reply);
3843                 } else {
3844                         gettimeofday(&request->reply->timestamp, NULL);
3845                         request_cleanup_delay_init(request);
3846                 }
3847                 break;
3848
3849                 /*
3850                  *      We received a new reply.  Go process it.
3851                  */
3852         case FR_ACTION_PROXY_REPLY:
3853                 request_queue_or_run(request, proxy_running);
3854                 break;
3855
3856         default:
3857                 RDEBUG3("%s: Ignoring action %s", __FUNCTION__, action_codes[action]);
3858                 break;
3859         }
3860 }
3861 #endif  /* WITH_PROXY */
3862
3863
3864 /***********************************************************************
3865  *
3866  *  CoA code
3867  *
3868  ***********************************************************************/
3869 #ifdef WITH_COA
3870 static int null_handler(UNUSED REQUEST *request)
3871 {
3872         return 0;
3873 }
3874
3875 /*
3876  *      See if we need to originate a CoA request.
3877  */
3878 static void request_coa_originate(REQUEST *request)
3879 {
3880         int rcode, pre_proxy_type = 0;
3881         VALUE_PAIR *vp;
3882         REQUEST *coa;
3883         fr_ipaddr_t ipaddr;
3884         char buffer[256];
3885
3886         VERIFY_REQUEST(request);
3887
3888         rad_assert(request->coa != NULL);
3889         rad_assert(request->proxy == NULL);
3890         rad_assert(!request->in_proxy_hash);
3891         rad_assert(request->proxy_reply == NULL);
3892
3893         /*
3894          *      Check whether we want to originate one, or cancel one.
3895          */
3896         vp = pairfind(request->config, PW_SEND_COA_REQUEST, 0, TAG_ANY);
3897         if (!vp) {
3898                 vp = pairfind(request->coa->proxy->vps, PW_SEND_COA_REQUEST, 0, TAG_ANY);
3899         }
3900
3901         if (vp) {
3902                 if (vp->vp_integer == 0) {
3903                 fail:
3904                         TALLOC_FREE(request->coa);
3905                         return;
3906                 }
3907         }
3908
3909         coa = request->coa;
3910
3911         /*
3912          *      src_ipaddr will be set up in proxy_encode.
3913          */
3914         memset(&ipaddr, 0, sizeof(ipaddr));
3915         vp = pairfind(coa->proxy->vps, PW_PACKET_DST_IP_ADDRESS, 0, TAG_ANY);
3916         if (vp) {
3917                 ipaddr.af = AF_INET;
3918                 ipaddr.ipaddr.ip4addr.s_addr = vp->vp_ipaddr;
3919                 ipaddr.prefix = 32;
3920         } else if ((vp = pairfind(coa->proxy->vps, PW_PACKET_DST_IPV6_ADDRESS, 0, TAG_ANY)) != NULL) {
3921                 ipaddr.af = AF_INET6;
3922                 ipaddr.ipaddr.ip6addr = vp->vp_ipv6addr;
3923                 ipaddr.prefix = 128;
3924         } else if ((vp = pairfind(coa->proxy->vps, PW_HOME_SERVER_POOL, 0, TAG_ANY)) != NULL) {
3925                 coa->home_pool = home_pool_byname(vp->vp_strvalue,
3926                                                   HOME_TYPE_COA);
3927                 if (!coa->home_pool) {
3928                         RWDEBUG2("No such home_server_pool %s",
3929                                vp->vp_strvalue);
3930                         goto fail;
3931                 }
3932
3933                 /*
3934                  *      Prefer the pool to one server
3935                  */
3936         } else if (request->client->coa_pool) {
3937                 coa->home_pool = request->client->coa_pool;
3938
3939         } else if (request->client->coa_server) {
3940                 coa->home_server = request->client->coa_server;
3941
3942         } else {
3943                 /*
3944                  *      If all else fails, send it to the client that
3945                  *      originated this request.
3946                  */
3947                 memcpy(&ipaddr, &request->packet->src_ipaddr, sizeof(ipaddr));
3948         }
3949
3950         /*
3951          *      Use the pool, if it exists.
3952          */
3953         if (coa->home_pool) {
3954                 coa->home_server = home_server_ldb(NULL, coa->home_pool, coa);
3955                 if (!coa->home_server) {
3956                         RWDEBUG("No live home server for home_server_pool %s", coa->home_pool->name);
3957                         goto fail;
3958                 }
3959                 home_server_update_request(coa->home_server, coa);
3960
3961         } else if (!coa->home_server) {
3962                 uint16_t port = PW_COA_UDP_PORT;
3963
3964                 vp = pairfind(coa->proxy->vps, PW_PACKET_DST_PORT, 0, TAG_ANY);
3965                 if (vp) port = vp->vp_integer;
3966
3967                 coa->home_server = home_server_find(&ipaddr, port, IPPROTO_UDP);
3968                 if (!coa->home_server) {
3969                         RWDEBUG2("Unknown destination %s:%d for CoA request.",
3970                                inet_ntop(ipaddr.af, &ipaddr.ipaddr,
3971                                          buffer, sizeof(buffer)), port);
3972                         goto fail;
3973                 }
3974         }
3975
3976         vp = pairfind(coa->proxy->vps, PW_PACKET_TYPE, 0, TAG_ANY);
3977         if (vp) {
3978                 switch (vp->vp_integer) {
3979                 case PW_CODE_COA_REQUEST:
3980                 case PW_CODE_DISCONNECT_REQUEST:
3981                         coa->proxy->code = vp->vp_integer;
3982                         break;
3983
3984                 default:
3985                         DEBUG("Cannot set CoA Packet-Type to code %d",
3986                               vp->vp_integer);
3987                         goto fail;
3988                 }
3989         }
3990
3991         if (!coa->proxy->code) coa->proxy->code = PW_CODE_COA_REQUEST;
3992
3993         /*
3994          *      The rest of the server code assumes that
3995          *      request->packet && request->reply exist.  Copy them
3996          *      from the original request.
3997          */
3998         rad_assert(coa->packet != NULL);
3999         rad_assert(coa->packet->vps == NULL);
4000
4001         coa->packet = rad_copy_packet(coa, request->packet);
4002         coa->reply = rad_copy_packet(coa, request->reply);
4003
4004         coa->config = paircopy(coa, request->config);
4005         coa->num_coa_requests = 0;
4006         coa->handle = null_handler;
4007         coa->number = request->number; /* it's associated with the same request */
4008
4009         /*
4010          *      Call the pre-proxy routines.
4011          */
4012         vp = pairfind(request->config, PW_PRE_PROXY_TYPE, 0, TAG_ANY);
4013         if (vp) {
4014                 DICT_VALUE const *dval = dict_valbyattr(vp->da->attr, vp->da->vendor, vp->vp_integer);
4015                 /* Must be a validation issue */
4016                 rad_assert(dval);
4017                 RDEBUG2("Found Pre-Proxy-Type %s", dval->name);
4018                 pre_proxy_type = vp->vp_integer;
4019         }
4020
4021         if (coa->home_pool && coa->home_pool->virtual_server) {
4022                 char const *old_server = coa->server;
4023
4024                 coa->server = coa->home_pool->virtual_server;
4025                 RDEBUG2("server %s {", coa->server);
4026                 RINDENT();
4027                 rcode = process_pre_proxy(pre_proxy_type, coa);
4028                 REXDENT();
4029                 RDEBUG2("}");
4030                 coa->server = old_server;
4031         } else {
4032                 rcode = process_pre_proxy(pre_proxy_type, coa);
4033         }
4034         switch (rcode) {
4035         default:
4036                 goto fail;
4037
4038         /*
4039          *      Only send the CoA packet if the pre-proxy code succeeded.
4040          */
4041         case RLM_MODULE_NOOP:
4042         case RLM_MODULE_OK:
4043         case RLM_MODULE_UPDATED:
4044                 break;
4045         }
4046
4047         /*
4048          *      Source IP / port is set when the proxy socket
4049          *      is chosen.
4050          */
4051         coa->proxy->dst_ipaddr = coa->home_server->ipaddr;
4052         coa->proxy->dst_port = coa->home_server->port;
4053
4054         if (!insert_into_proxy_hash(coa)) {
4055                 radlog_request(L_PROXY, 0, coa, "Failed to insert CoA request into proxy list");
4056                 goto fail;
4057         }
4058
4059         /*
4060          *      We CANNOT divorce the CoA request from the parent
4061          *      request.  This function is running in a child thread,
4062          *      and we need access to the main event loop in order to
4063          *      to add the timers for the CoA packet.
4064          *
4065          *      Instead, we wait for the timer on the parent request
4066          *      to fire.
4067          */
4068         gettimeofday(&coa->proxy->timestamp, NULL);
4069         coa->packet->timestamp = coa->proxy->timestamp; /* for max_request_time */
4070         coa->home_server->last_packet_sent = coa->proxy->timestamp.tv_sec;
4071         coa->delay = 0;         /* need to calculate a new delay */
4072
4073         /*
4074          *      If requested, put a State attribute into the packet,
4075          *      and cache the VPS.
4076          */
4077         fr_state_put_vps(coa, NULL, coa->packet);
4078
4079         /*
4080          *      Encode the packet before we do anything else.
4081          */
4082         coa->proxy_listener->encode(coa->proxy_listener, coa);
4083         debug_packet(coa, coa->proxy, false);
4084
4085 #ifdef DEBUG_STATE_MACHINE
4086         if (debug_flag) printf("(%u) ********\tSTATE %s C-%s -> C-%s\t********\n", request->number, __FUNCTION__,
4087                                child_state_names[request->child_state],
4088                                child_state_names[REQUEST_PROXIED]);
4089 #endif
4090
4091         /*
4092          *      Set the state function, then the state, no child, and
4093          *      send the packet.
4094          */
4095         coa->process = coa_wait_for_reply;
4096         coa->child_state = REQUEST_PROXIED;
4097
4098 #ifdef HAVE_PTHREAD_H
4099         coa->child_pid = NO_SUCH_CHILD_PID;
4100 #endif
4101
4102         /*
4103          *      And send the packet.
4104          */
4105         coa->proxy_listener->send(coa->proxy_listener, coa);
4106 }
4107
4108
4109 static void coa_retransmit(REQUEST *request)
4110 {
4111         uint32_t delay, frac;
4112         struct timeval now, when, mrd;
4113
4114         VERIFY_REQUEST(request);
4115
4116         fr_event_now(el, &now);
4117
4118         /*
4119          *      FIXME: Enforce max_request_time
4120          */
4121
4122         if (request->delay == 0) {
4123                 /*
4124                  *      Implement re-transmit algorithm as per RFC 5080
4125                  *      Section 2.2.1.
4126                  *
4127                  *      We want IRT + RAND*IRT
4128                  *      or 0.9 IRT + rand(0,.2) IRT
4129                  *
4130                  *      2^20 ~ USEC, and we want 2.
4131                  *      rand(0,0.2) USEC ~ (rand(0,2^21) / 10)
4132                  */
4133                 delay = (fr_rand() & ((1 << 22) - 1)) / 10;
4134                 request->delay = delay * request->home_server->coa_irt;
4135                 delay = request->home_server->coa_irt * USEC;
4136                 delay -= delay / 10;
4137                 delay += request->delay;
4138                 request->delay = delay;
4139
4140                 when = request->proxy->timestamp;
4141                 tv_add(&when, delay);
4142
4143                 if (timercmp(&when, &now, >)) {
4144                         STATE_MACHINE_TIMER(FR_ACTION_TIMER);
4145                         return;
4146                 }
4147         }
4148
4149         /*
4150          *      Retransmit CoA request.
4151          */
4152
4153         /*
4154          *      Cap count at MRC, if it is non-zero.
4155          */
4156         if (request->home_server->coa_mrc &&
4157             (request->num_coa_requests >= request->home_server->coa_mrc)) {
4158                 char buffer[128];
4159
4160                 RERROR("Failing request - originate-coa ID %u, due to lack of any response from coa server %s port %d",
4161                        request->proxy->id,
4162                                inet_ntop(request->proxy->dst_ipaddr.af,
4163                                          &request->proxy->dst_ipaddr.ipaddr,
4164                                          buffer, sizeof(buffer)),
4165                                request->proxy->dst_port);
4166
4167                 if (setup_post_proxy_fail(request)) {
4168                         request_queue_or_run(request, coa_no_reply);
4169                 } else {
4170                         request_done(request, FR_ACTION_DONE);
4171                 }
4172                 return;
4173         }
4174
4175         /*
4176          *      RFC 5080 Section 2.2.1
4177          *
4178          *      RT = 2*RTprev + RAND*RTprev
4179          *         = 1.9 * RTprev + rand(0,.2) * RTprev
4180          *         = 1.9 * RTprev + rand(0,1) * (RTprev / 5)
4181          */
4182         delay = fr_rand();
4183         delay ^= (delay >> 16);
4184         delay &= 0xffff;
4185         frac = request->delay / 5;
4186         delay = ((frac >> 16) * delay) + (((frac & 0xffff) * delay) >> 16);
4187
4188         delay += (2 * request->delay) - (request->delay / 10);
4189
4190         /*
4191          *      Cap delay at MRT, if MRT is non-zero.
4192          */
4193         if (request->home_server->coa_mrt &&
4194             (delay > (request->home_server->coa_mrt * USEC))) {
4195                 int mrt_usec = request->home_server->coa_mrt * USEC;
4196
4197                 /*
4198                  *      delay = MRT + RAND * MRT
4199                  *            = 0.9 MRT + rand(0,.2)  * MRT
4200                  */
4201                 delay = fr_rand();
4202                 delay ^= (delay >> 15);
4203                 delay &= 0x1ffff;
4204                 delay = ((mrt_usec >> 16) * delay) + (((mrt_usec & 0xffff) * delay) >> 16);
4205                 delay += mrt_usec - (mrt_usec / 10);
4206         }
4207
4208         request->delay = delay;
4209         when = now;
4210         tv_add(&when, request->delay);
4211         mrd = request->proxy->timestamp;
4212         mrd.tv_sec += request->home_server->coa_mrd;
4213
4214         /*
4215          *      Cap duration at MRD.
4216          */
4217         if (timercmp(&mrd, &when, <)) {
4218                 when = mrd;
4219         }
4220         STATE_MACHINE_TIMER(FR_ACTION_TIMER);
4221
4222         request->num_coa_requests++; /* is NOT reset by code 3 lines above! */
4223
4224         FR_STATS_TYPE_INC(request->home_server->stats.total_requests);
4225
4226         request->proxy_listener->send(request->proxy_listener,
4227                                       request);
4228 }
4229
4230
4231 /** Wait for a reply after originating a CoA a request.
4232  *
4233  *  Retransmit the proxied packet, or time out and go to
4234  *  coa_no_reply.  Mark the home server unresponsive, etc.
4235  *
4236  *  If we do receive a reply, we transition to coa_running.
4237  *
4238  *  \dot
4239  *      digraph coa_wait_for_reply {
4240  *              coa_wait_for_reply;
4241  *
4242  *              coa_wait_for_reply -> coa_no_reply [ label = "TIMER >= response_window" ];
4243  *              coa_wait_for_reply -> timer [ label = "TIMER < max_request_time" ];
4244  *              coa_wait_for_reply -> coa_running [ label = "PROXY_REPLY" arrowhead = "none"];
4245  *              coa_wait_for_reply -> done [ label = "TIMER >= max_request_time" ];
4246  *      }
4247  *  \enddot
4248  */
4249 static void coa_wait_for_reply(REQUEST *request, int action)
4250 {
4251         VERIFY_REQUEST(request);
4252
4253         TRACE_STATE_MACHINE;
4254         ASSERT_MASTER;
4255         CHECK_FOR_STOP;
4256
4257         switch (action) {
4258         case FR_ACTION_TIMER:
4259                 if (request->parent) coa_separate(request);
4260
4261                 coa_retransmit(request);
4262                 break;
4263
4264         case FR_ACTION_PROXY_REPLY:
4265                 if (request->parent) coa_separate(request);
4266
4267                 request_queue_or_run(request, coa_running);
4268                 break;
4269
4270         default:
4271                 RDEBUG3("%s: Ignoring action %s", __FUNCTION__, action_codes[action]);
4272                 break;
4273         }
4274 }
4275
4276 static void coa_separate(REQUEST *request)
4277 {
4278         VERIFY_REQUEST(request);
4279 #ifdef DEBUG_STATE_MACHINE
4280         int action = FR_ACTION_TIMER;
4281 #endif
4282
4283         TRACE_STATE_MACHINE;
4284         ASSERT_MASTER;
4285
4286         rad_assert(request->parent != NULL);
4287         rad_assert(request->parent->coa == request);
4288         rad_assert(request->ev == NULL);
4289         rad_assert(!request->in_request_hash);
4290         rad_assert(request->coa == NULL);
4291
4292         rad_assert(request->proxy_reply || request->proxy_listener);
4293
4294         (void) talloc_steal(NULL, request);
4295         request->parent->coa = NULL;
4296         request->parent = NULL;
4297 }
4298
4299
4300 /** Process a request after the CoA has timed out.
4301  *
4302  *  Run the packet through Post-Proxy-Type Fail
4303  *
4304  *  \dot
4305  *      digraph coa_no_reply {
4306  *              coa_no_reply;
4307  *
4308  *              coa_no_reply -> dup [ label = "DUP", arrowhead = "none" ];
4309  *              coa_no_reply -> timer [ label = "TIMER < max_request_time" ];
4310  *              coa_no_reply -> coa_reply_too_late [ label = "PROXY_REPLY" arrowhead = "none"];
4311  *              coa_no_reply -> process_proxy_reply [ label = "RUN" ];
4312  *              coa_no_reply -> done [ label = "TIMER >= timeout" ];
4313  *      }
4314  *  \enddot
4315  */
4316 static void coa_no_reply(REQUEST *request, int action)
4317 {
4318         char buffer[128];
4319
4320         VERIFY_REQUEST(request);
4321
4322         TRACE_STATE_MACHINE;
4323         CHECK_FOR_STOP;
4324
4325         switch (action) {
4326         case FR_ACTION_TIMER:
4327                 request_max_time(request);
4328                 break;
4329
4330         case FR_ACTION_PROXY_REPLY: /* too late! */
4331                 RDEBUG2("Reply from CoA server %s port %d  - ID: %d arrived too late.",
4332                         inet_ntop(request->proxy->src_ipaddr.af,
4333                                   &request->proxy->src_ipaddr.ipaddr,
4334                                   buffer, sizeof(buffer)),
4335                         request->proxy->dst_port, request->proxy->id);
4336                 break;
4337
4338         case FR_ACTION_RUN:
4339                 if (process_proxy_reply(request, NULL)) {
4340                         request->handle(request);
4341                 }
4342                 request_done(request, FR_ACTION_DONE);
4343                 break;
4344
4345         default:
4346                 RDEBUG3("%s: Ignoring action %s", __FUNCTION__, action_codes[action]);
4347                 break;
4348         }
4349 }
4350
4351
4352 /** Process the request after receiving a coa reply.
4353  *
4354  *  Throught the post-proxy section, and the through the handler
4355  *  function.
4356  *
4357  *  \dot
4358  *      digraph coa_running {
4359  *              coa_running;
4360  *
4361  *              coa_running -> timer [ label = "TIMER < max_request_time" ];
4362  *              coa_running -> process_proxy_reply [ label = "RUN" ];
4363  *              coa_running -> done [ label = "TIMER >= timeout" ];
4364  *      }
4365  *  \enddot
4366  */
4367 static void coa_running(REQUEST *request, int action)
4368 {
4369         VERIFY_REQUEST(request);
4370
4371         TRACE_STATE_MACHINE;
4372         CHECK_FOR_STOP;
4373
4374         switch (action) {
4375         case FR_ACTION_TIMER:
4376                 request_max_time(request);
4377                 break;
4378
4379         case FR_ACTION_RUN:
4380                 if (process_proxy_reply(request, request->proxy_reply)) {
4381                         request->handle(request);
4382                 }
4383                 request_done(request, FR_ACTION_DONE);
4384                 break;
4385
4386         default:
4387                 RDEBUG3("%s: Ignoring action %s", __FUNCTION__, action_codes[action]);
4388                 break;
4389         }
4390 }
4391 #endif  /* WITH_COA */
4392
4393 /***********************************************************************
4394  *
4395  *  End of the State machine.  Start of additional helper code.
4396  *
4397  ***********************************************************************/
4398
4399 /***********************************************************************
4400  *
4401  *      Event handlers.
4402  *
4403  ***********************************************************************/
4404 static void event_socket_handler(fr_event_list_t *xel, UNUSED int fd, void *ctx)
4405 {
4406         rad_listen_t *listener = talloc_get_type_abort(ctx, rad_listen_t);
4407
4408         rad_assert(xel == el);
4409
4410         if ((listener->fd < 0)
4411 #ifdef WITH_DETAIL
4412 #ifndef WITH_DETAIL_THREAD
4413             && (listener->type != RAD_LISTEN_DETAIL)
4414 #endif
4415 #endif
4416                 ) {
4417                 char buffer[256];
4418
4419                 listener->print(listener, buffer, sizeof(buffer));
4420                 ERROR("FATAL: Asked to read from closed socket: %s",
4421                        buffer);
4422
4423                 rad_panic("Socket was closed on us!");
4424                 fr_exit_now(1);
4425         }
4426
4427         listener->recv(listener);
4428 }
4429
4430 #ifdef WITH_DETAIL
4431 #ifdef WITH_DETAIL_THREAD
4432 #else
4433 /*
4434  *      This function is called periodically to see if this detail
4435  *      file is available for reading.
4436  */
4437 static void event_poll_detail(void *ctx)
4438 {
4439         int delay;
4440         rad_listen_t *this = talloc_get_type_abort(ctx, rad_listen_t);
4441         struct timeval when, now;
4442         listen_detail_t *detail = this->data;
4443
4444         rad_assert(this->type == RAD_LISTEN_DETAIL);
4445
4446  redo:
4447         event_socket_handler(el, this->fd, this);
4448
4449         fr_event_now(el, &now);
4450         when = now;
4451
4452         /*
4453          *      Backdoor API to get the delay until the next poll
4454          *      time.
4455          */
4456         delay = this->encode(this, NULL);
4457         if (delay == 0) goto redo;
4458
4459         tv_add(&when, delay);
4460
4461         ASSERT_MASTER;
4462         if (!fr_event_insert(el, event_poll_detail, this,
4463                              &when, &detail->ev)) {
4464                 ERROR("Failed creating handler");
4465                 fr_exit(1);
4466         }
4467 }
4468 #endif  /* WITH_DETAIL_THREAD */
4469 #endif  /* WITH_DETAIL */
4470
4471 static void event_status(struct timeval *wake)
4472 {
4473 #if !defined(HAVE_PTHREAD_H) && defined(WNOHANG)
4474         int argval;
4475 #endif
4476
4477         if (debug_flag == 0) {
4478                 if (just_started) {
4479                         INFO("Ready to process requests");
4480                         just_started = false;
4481                 }
4482                 return;
4483         }
4484
4485         if (!wake) {
4486                 INFO("Ready to process requests");
4487
4488         } else if ((wake->tv_sec != 0) ||
4489                    (wake->tv_usec >= 100000)) {
4490                 DEBUG("Waking up in %d.%01u seconds.",
4491                       (int) wake->tv_sec, (unsigned int) wake->tv_usec / 100000);
4492         }
4493
4494
4495         /*
4496          *      FIXME: Put this somewhere else, where it isn't called
4497          *      all of the time...
4498          */
4499
4500 #if !defined(HAVE_PTHREAD_H) && defined(WNOHANG)
4501         /*
4502          *      If there are no child threads, then there may
4503          *      be child processes.  In that case, wait for
4504          *      their exit status, and throw that exit status
4505          *      away.  This helps get rid of zxombie children.
4506          */
4507         while (waitpid(-1, &argval, WNOHANG) > 0) {
4508                 /* do nothing */
4509         }
4510 #endif
4511
4512 }
4513
4514 #ifdef WITH_TCP
4515 static void listener_free_cb(void *ctx)
4516 {
4517         rad_listen_t *this = talloc_get_type_abort(ctx, rad_listen_t);
4518         char buffer[1024];
4519
4520         if (this->count > 0) {
4521                 struct timeval when;
4522                 listen_socket_t *sock = this->data;
4523
4524                 fr_event_now(el, &when);
4525                 when.tv_sec += 3;
4526
4527                 ASSERT_MASTER;
4528                 if (!fr_event_insert(el, listener_free_cb, this, &when,
4529                                      &(sock->ev))) {
4530                         rad_panic("Failed to insert event");
4531                 }
4532
4533                 return;
4534         }
4535
4536         /*
4537          *      It's all free, close the socket.
4538          */
4539
4540         this->print(this, buffer, sizeof(buffer));
4541         DEBUG("... cleaning up socket %s", buffer);
4542         rad_assert(this->next == NULL);
4543         talloc_free(this);
4544 }
4545 #endif
4546
4547 #ifdef WITH_PROXY
4548 static int proxy_eol_cb(void *ctx, void *data)
4549 {
4550         struct timeval when;
4551         REQUEST *request = fr_packet2myptr(REQUEST, proxy, data);
4552
4553         if (request->proxy_listener != ctx) return 0;
4554
4555         /*
4556          *      We don't care if it's being processed in a child thread.
4557          */
4558
4559 #ifdef WITH_ACCOUNTING
4560         /*
4561          *      Accounting packets should be deleted immediately.
4562          *      They will never be retransmitted by the client.
4563          */
4564         if (request->proxy->code == PW_CODE_ACCOUNTING_REQUEST) {
4565                 RDEBUG("Stopping request due to failed connection to home server");
4566                 request->master_state = REQUEST_STOP_PROCESSING;
4567         }
4568 #endif
4569
4570         /*
4571          *      Reset the timer to be now, so that the request is
4572          *      quickly updated.  But spread the requests randomly
4573          *      over the next second, so that we don't overload the
4574          *      server.
4575          */
4576         fr_event_now(el, &when);
4577         tv_add(&when, fr_rand() % USEC);
4578         STATE_MACHINE_TIMER(FR_ACTION_TIMER);
4579
4580         /*
4581          *      Don't delete it from the list.
4582          */
4583         return 0;
4584 }
4585 #endif
4586
4587 static int event_new_fd(rad_listen_t *this)
4588 {
4589         char buffer[1024];
4590
4591         ASSERT_MASTER;
4592
4593         if (this->status == RAD_LISTEN_STATUS_KNOWN) return 1;
4594
4595         this->print(this, buffer, sizeof(buffer));
4596
4597         if (this->status == RAD_LISTEN_STATUS_INIT) {
4598                 listen_socket_t *sock = this->data;
4599
4600                 rad_assert(sock != NULL);
4601                 if (just_started) {
4602                         DEBUG("Listening on %s", buffer);
4603
4604 #ifdef WITH_PROXY
4605                 } else if (this->type == RAD_LISTEN_PROXY) {
4606                         home_server_t *home;
4607
4608                         home = sock->home;
4609                         if (!home || !home->limit.max_connections) {
4610                                 INFO(" ... adding new socket %s", buffer);
4611                         } else {
4612                                 INFO(" ... adding new socket %s (%u of %u)", buffer,
4613                                      home->limit.num_connections, home->limit.max_connections);
4614                         }
4615
4616 #endif
4617                 } else {
4618                         INFO(" ... adding new socket %s", buffer);
4619                 }
4620
4621                 switch (this->type) {
4622 #ifdef WITH_DETAIL
4623                 /*
4624                  *      Detail files are always known, and aren't
4625                  *      put into the socket event loop.
4626                  */
4627                 case RAD_LISTEN_DETAIL:
4628                         this->status = RAD_LISTEN_STATUS_KNOWN;
4629
4630 #ifndef WITH_DETAIL_THREAD
4631                         /*
4632                          *      Set up the first poll interval.
4633                          */
4634                         event_poll_detail(this);
4635                         return 1;
4636 #else
4637                         break;  /* add the FD to the list */
4638 #endif
4639 #endif  /* WITH_DETAIL */
4640
4641 #ifdef WITH_PROXY
4642                 /*
4643                  *      Add it to the list of sockets we can use.
4644                  *      Server sockets (i.e. auth/acct) are never
4645                  *      added to the packet list.
4646                  */
4647                 case RAD_LISTEN_PROXY:
4648 #ifdef WITH_TCP
4649                         rad_assert((sock->proto == IPPROTO_UDP) || (sock->home != NULL));
4650
4651                         /*
4652                          *      Add timers to outgoing child sockets, if necessary.
4653                          */
4654                         if (sock->proto == IPPROTO_TCP && sock->opened &&
4655                             (sock->home->limit.lifetime || sock->home->limit.idle_timeout)) {
4656                                 struct timeval when;
4657
4658                                 when.tv_sec = sock->opened + 1;
4659                                 when.tv_usec = 0;
4660
4661                                 ASSERT_MASTER;
4662                                 if (!fr_event_insert(el, tcp_socket_timer, this, &when,
4663                                                      &(sock->ev))) {
4664                                         rad_panic("Failed to insert event");
4665                                 }
4666                         }
4667 #endif
4668                         break;
4669 #endif  /* WITH_PROXY */
4670
4671                         /*
4672                          *      FIXME: put idle timers on command sockets.
4673                          */
4674
4675                 default:
4676 #ifdef WITH_TCP
4677                         /*
4678                          *      Add timers to incoming child sockets, if necessary.
4679                          */
4680                         if (sock->proto == IPPROTO_TCP && sock->opened &&
4681                             (sock->limit.lifetime || sock->limit.idle_timeout)) {
4682                                 struct timeval when;
4683
4684                                 when.tv_sec = sock->opened + 1;
4685                                 when.tv_usec = 0;
4686
4687                                 ASSERT_MASTER;
4688                                 if (!fr_event_insert(el, tcp_socket_timer, this, &when,
4689                                                      &(sock->ev))) {
4690                                         ERROR("Failed adding timer for socket: %s", fr_strerror());
4691                                         fr_exit(1);
4692                                 }
4693                         }
4694 #endif
4695                         break;
4696                 } /* switch over listener types */
4697
4698                 /*
4699                  *      All sockets: add the FD to the event handler.
4700                  */
4701                 if (!fr_event_fd_insert(el, 0, this->fd,
4702                                         event_socket_handler, this)) {
4703                         ERROR("Failed adding event handler for socket: %s", fr_strerror());
4704                         fr_exit(1);
4705                 }
4706
4707                 this->status = RAD_LISTEN_STATUS_KNOWN;
4708                 return 1;
4709         } /* end of INIT */
4710
4711 #ifdef WITH_TCP
4712         /*
4713          *      The socket has reached a timeout.  Try to close it.
4714          */
4715         if (this->status == RAD_LISTEN_STATUS_FROZEN) {
4716                 /*
4717                  *      Requests are still using the socket.  Wait for
4718                  *      them to finish.
4719                  */
4720                 if (this->count > 0) {
4721                         struct timeval when;
4722                         listen_socket_t *sock = this->data;
4723
4724                         /*
4725                          *      Try again to clean up the socket in 30
4726                          *      seconds.
4727                          */
4728                         gettimeofday(&when, NULL);
4729                         when.tv_sec += 30;
4730
4731                         ASSERT_MASTER;
4732                         if (!fr_event_insert(el,
4733                                              (fr_event_callback_t) event_new_fd,
4734                                              this, &when, &sock->ev)) {
4735                                 rad_panic("Failed to insert event");
4736                         }
4737
4738                         return 1;
4739                 }
4740
4741                 fr_event_fd_delete(el, 0, this->fd);
4742                 this->status = RAD_LISTEN_STATUS_REMOVE_NOW;
4743         }
4744
4745         /*
4746          *      The socket has had a catastrophic error.  Close it.
4747          */
4748         if (this->status == RAD_LISTEN_STATUS_EOL) {
4749                 /*
4750                  *      Remove it from the list of live FD's.
4751                  */
4752                 fr_event_fd_delete(el, 0, this->fd);
4753
4754 #ifdef WITH_PROXY
4755                 /*
4756                  *      Tell all requests using this socket that the socket is dead.
4757                  */
4758                 if (this->type == RAD_LISTEN_PROXY) {
4759                         PTHREAD_MUTEX_LOCK(&proxy_mutex);
4760                         if (!fr_packet_list_socket_freeze(proxy_list,
4761                                                           this->fd)) {
4762                                 ERROR("Fatal error freezing socket: %s", fr_strerror());
4763                                 fr_exit(1);
4764                         }
4765
4766                         if (this->count > 0) {
4767                                 fr_packet_list_walk(proxy_list, this, proxy_eol_cb);
4768                         }
4769                         PTHREAD_MUTEX_UNLOCK(&proxy_mutex);
4770                 }
4771 #endif
4772
4773                 /*
4774                  *      Requests are still using the socket.  Wait for
4775                  *      them to finish.
4776                  */
4777                 if (this->count > 0) {
4778                         struct timeval when;
4779                         listen_socket_t *sock = this->data;
4780
4781                         /*
4782                          *      Try again to clean up the socket in 30
4783                          *      seconds.
4784                          */
4785                         gettimeofday(&when, NULL);
4786                         when.tv_sec += 30;
4787
4788                         ASSERT_MASTER;
4789                         if (!fr_event_insert(el,
4790                                              (fr_event_callback_t) event_new_fd,
4791                                              this, &when, &sock->ev)) {
4792                                 rad_panic("Failed to insert event");
4793                         }
4794
4795                         return 1;
4796                 }
4797
4798                 /*
4799                  *      No one is using the socket.  We can remove it now.
4800                  */
4801                 this->status = RAD_LISTEN_STATUS_REMOVE_NOW;
4802         } /* socket is at EOL */
4803 #endif
4804
4805         /*
4806          *      Nuke the socket.
4807          */
4808         if (this->status == RAD_LISTEN_STATUS_REMOVE_NOW) {
4809                 int devnull;
4810 #ifdef WITH_TCP
4811                 listen_socket_t *sock = this->data;
4812 #endif
4813                 struct timeval when;
4814
4815                 /*
4816                  *      Re-open the socket, pointing it to /dev/null.
4817                  *      This means that all writes proceed without
4818                  *      blocking, and all reads return "no data".
4819                  *
4820                  *      This leaves the socket active, so any child
4821                  *      threads won't go insane.  But it means that
4822                  *      they cannot send or receive any packets.
4823                  *
4824                  *      This is EXTRA work in the normal case, when
4825                  *      sockets are closed without error.  But it lets
4826                  *      us have one simple processing method for all
4827                  *      sockets.
4828                  */
4829                 devnull = open("/dev/null", O_RDWR);
4830                 if (devnull < 0) {
4831                         ERROR("FATAL failure opening /dev/null: %s",
4832                                fr_syserror(errno));
4833                         fr_exit(1);
4834                 }
4835                 if (dup2(devnull, this->fd) < 0) {
4836                         ERROR("FATAL failure closing socket: %s",
4837                                fr_syserror(errno));
4838                         fr_exit(1);
4839                 }
4840                 close(devnull);
4841
4842 #ifdef WITH_DETAIL
4843                 rad_assert(this->type != RAD_LISTEN_DETAIL);
4844 #endif
4845
4846 #ifdef WITH_TCP
4847 #ifdef WITH_PROXY
4848                 /*
4849                  *      The socket is dead.  Force all proxied packets
4850                  *      to stop using it.  And then remove it from the
4851                  *      list of outgoing sockets.
4852                  */
4853                 if (this->type == RAD_LISTEN_PROXY) {
4854                         home_server_t *home;
4855
4856                         home = sock->home;
4857                         if (!home || !home->limit.max_connections) {
4858                                 INFO(" ... shutting down socket %s", buffer);
4859                         } else {
4860                                 INFO(" ... shutting down socket %s (%u of %u)", buffer,
4861                                      home->limit.num_connections, home->limit.max_connections);
4862                         }
4863
4864                         PTHREAD_MUTEX_LOCK(&proxy_mutex);
4865                         fr_packet_list_walk(proxy_list, this, eol_proxy_listener);
4866
4867                         if (!fr_packet_list_socket_del(proxy_list, this->fd)) {
4868                                 ERROR("Fatal error removing socket %s: %s",
4869                                       buffer, fr_strerror());
4870                                 fr_exit(1);
4871                         }
4872                         PTHREAD_MUTEX_UNLOCK(&proxy_mutex);
4873                 } else
4874 #endif
4875                 {
4876                         INFO(" ... shutting down socket %s", buffer);
4877
4878                         /*
4879                          *      EOL all requests using this socket.
4880                          */
4881                         rbtree_walk(pl, RBTREE_DELETE_ORDER, eol_listener, this);
4882                 }
4883
4884                 /*
4885                  *      No child threads, clean it up now.
4886                  */
4887                 if (!spawn_flag) {
4888                         ASSERT_MASTER;
4889                         if (sock->ev) fr_event_delete(el, &sock->ev);
4890                         listen_free(&this);
4891                         return 1;
4892                 }
4893
4894                 /*
4895                  *      Wait until all requests using this socket are done.
4896                  */
4897                 gettimeofday(&when, NULL);
4898                 when.tv_sec += 3;
4899
4900                 ASSERT_MASTER;
4901                 if (!fr_event_insert(el, listener_free_cb, this, &when,
4902                                      &(sock->ev))) {
4903                         rad_panic("Failed to insert event");
4904                 }
4905         }
4906 #endif  /* WITH_TCP */
4907
4908         return 1;
4909 }
4910
4911 /***********************************************************************
4912  *
4913  *      Signal handlers.
4914  *
4915  ***********************************************************************/
4916
4917 static void handle_signal_self(int flag)
4918 {
4919         ASSERT_MASTER;
4920
4921         if ((flag & (RADIUS_SIGNAL_SELF_EXIT | RADIUS_SIGNAL_SELF_TERM)) != 0) {
4922                 if ((flag & RADIUS_SIGNAL_SELF_EXIT) != 0) {
4923                         INFO("Signalled to exit");
4924                         fr_event_loop_exit(el, 1);
4925                 } else {
4926                         INFO("Signalled to terminate");
4927                         fr_event_loop_exit(el, 2);
4928                 }
4929
4930                 return;
4931         } /* else exit/term flags weren't set */
4932
4933         /*
4934          *      Tell the even loop to stop processing.
4935          */
4936         if ((flag & RADIUS_SIGNAL_SELF_HUP) != 0) {
4937                 time_t when;
4938                 static time_t last_hup = 0;
4939
4940                 when = time(NULL);
4941                 if ((int) (when - last_hup) < 5) {
4942                         INFO("Ignoring HUP (less than 5s since last one)");
4943                         return;
4944                 }
4945
4946                 INFO("Received HUP signal");
4947
4948                 last_hup = when;
4949
4950                 exec_trigger(NULL, NULL, "server.signal.hup", true);
4951                 fr_event_loop_exit(el, 0x80);
4952         }
4953
4954 #if defined(WITH_DETAIL) && !defined(WITH_DETAIL_THREAD)
4955         if ((flag & RADIUS_SIGNAL_SELF_DETAIL) != 0) {
4956                 rad_listen_t *this;
4957
4958                 /*
4959                  *      FIXME: O(N) loops suck.
4960                  */
4961                 for (this = main_config.listen;
4962                      this != NULL;
4963                      this = this->next) {
4964                         if (this->type != RAD_LISTEN_DETAIL) continue;
4965
4966                         /*
4967                          *      This one didn't send the signal, skip
4968                          *      it.
4969                          */
4970                         if (!this->decode(this, NULL)) continue;
4971
4972                         /*
4973                          *      Go service the interrupt.
4974                          */
4975                         event_poll_detail(this);
4976                 }
4977         }
4978 #endif
4979
4980 #if defined(WITH_TCP) && defined(WITH_PROXY) && defined(HAVE_PTHREAD_H)
4981         /*
4982          *      There are new listeners in the list.  Run
4983          *      event_new_fd() on them.
4984          */
4985         if ((flag & RADIUS_SIGNAL_SELF_NEW_FD) != 0) {
4986                 rad_listen_t *this, *next;
4987
4988                 FD_MUTEX_LOCK(&fd_mutex);
4989
4990                 /*
4991                  *      FIXME: unlock the mutex before calling
4992                  *      event_new_fd()?
4993                  */
4994                 for (this = new_listeners; this != NULL; this = next) {
4995                         next = this->next;
4996                         this->next = NULL;
4997
4998                         event_new_fd(this);
4999                 }
5000
5001                 new_listeners = NULL;
5002                 FD_MUTEX_UNLOCK(&fd_mutex);
5003         }
5004 #endif
5005 }
5006
5007 #ifndef HAVE_PTHREAD_H
5008 void radius_signal_self(int flag)
5009 {
5010         return handle_signal_self(flag);
5011 }
5012
5013 #else
5014 static int self_pipe[2] = { -1, -1 };
5015
5016 /*
5017  *      Inform ourselves that we received a signal.
5018  */
5019 void radius_signal_self(int flag)
5020 {
5021         ssize_t rcode;
5022         uint8_t buffer[16];
5023
5024         /*
5025          *      The read MUST be non-blocking for this to work.
5026          */
5027         rcode = read(self_pipe[0], buffer, sizeof(buffer));
5028         if (rcode > 0) {
5029                 ssize_t i;
5030
5031                 for (i = 0; i < rcode; i++) {
5032                         buffer[0] |= buffer[i];
5033                 }
5034         } else {
5035                 buffer[0] = 0;
5036         }
5037
5038         buffer[0] |= flag;
5039
5040         if (write(self_pipe[1], buffer, 1) < 0) fr_exit(0);
5041 }
5042
5043
5044 static void event_signal_handler(UNUSED fr_event_list_t *xel,
5045                                  UNUSED int fd, UNUSED void *ctx)
5046 {
5047         ssize_t i, rcode;
5048         uint8_t buffer[32];
5049
5050         rcode = read(self_pipe[0], buffer, sizeof(buffer));
5051         if (rcode <= 0) return;
5052
5053         /*
5054          *      Merge pending signals.
5055          */
5056         for (i = 0; i < rcode; i++) {
5057                 buffer[0] |= buffer[i];
5058         }
5059
5060         handle_signal_self(buffer[0]);
5061 }
5062 #endif  /* HAVE_PTHREAD_H */
5063
5064 /***********************************************************************
5065  *
5066  *      Bootstrapping code.
5067  *
5068  ***********************************************************************/
5069
5070 /*
5071  *      Externally-visibly functions.
5072  */
5073 int radius_event_init(TALLOC_CTX *ctx) {
5074         el = fr_event_list_create(ctx, event_status);
5075         if (!el) return 0;
5076
5077         return 1;
5078 }
5079
5080 static int packet_entry_cmp(void const *one, void const *two)
5081 {
5082         RADIUS_PACKET const * const *a = one;
5083         RADIUS_PACKET const * const *b = two;
5084
5085         return fr_packet_cmp(*a, *b);
5086 }
5087
5088
5089 int radius_event_start(CONF_SECTION *cs, bool have_children)
5090 {
5091         rad_listen_t *head = NULL;
5092
5093         if (fr_start_time != (time_t)-1) return 0;
5094
5095         time(&fr_start_time);
5096
5097         if (!check_config) {
5098                 /*
5099                  *  radius_event_init() must be called first
5100                  */
5101                 rad_assert(el);
5102
5103                 pl = rbtree_create(NULL, packet_entry_cmp, NULL, 0);
5104                 if (!pl) return 0;      /* leak el */
5105         }
5106
5107         request_num_counter = 0;
5108
5109 #ifdef WITH_PROXY
5110         if (main_config.proxy_requests && !check_config) {
5111                 /*
5112                  *      Create the tree for managing proxied requests and
5113                  *      responses.
5114                  */
5115                 proxy_list = fr_packet_list_create(1);
5116                 if (!proxy_list) return 0;
5117
5118 #ifdef HAVE_PTHREAD_H
5119                 if (pthread_mutex_init(&proxy_mutex, NULL) != 0) {
5120                         ERROR("FATAL: Failed to initialize proxy mutex: %s",
5121                                fr_syserror(errno));
5122                         fr_exit(1);
5123                 }
5124 #endif
5125
5126                 /*
5127                  *      The "init_delay" is set to "response_window".
5128                  *      Reset it to half of "response_window" in order
5129                  *      to give the event loop enough time to service
5130                  *      the event before hitting "response_window".
5131                  */
5132                 main_config.init_delay.tv_usec += (main_config.init_delay.tv_sec & 0x01) * USEC;
5133                 main_config.init_delay.tv_usec >>= 1;
5134                 main_config.init_delay.tv_sec >>= 1;
5135
5136                 proxy_ctx = talloc_init("proxy");
5137         }
5138 #endif
5139
5140         /*
5141          *      Move all of the thread calls to this file?
5142          *
5143          *      It may be best for the mutexes to be in this file...
5144          */
5145         spawn_flag = have_children;
5146
5147 #ifdef HAVE_PTHREAD_H
5148         NO_SUCH_CHILD_PID = pthread_self(); /* not a child thread */
5149
5150         /*
5151          *      Initialize the threads ONLY if we're spawning, AND
5152          *      we're running normally.
5153          */
5154         if (have_children && !check_config &&
5155             (thread_pool_init(cs, &spawn_flag) < 0)) {
5156                 fr_exit(1);
5157         }
5158 #endif
5159
5160         if (check_config) {
5161                 DEBUG("%s: #### Skipping IP addresses and Ports ####",
5162                        main_config.name);
5163                 if (listen_init(cs, &head, spawn_flag) < 0) {
5164                         fflush(NULL);
5165                         fr_exit(1);
5166                 }
5167                 return 1;
5168         }
5169
5170 #ifdef HAVE_PTHREAD_H
5171         /*
5172          *      Child threads need a pipe to signal us, as do the
5173          *      signal handlers.
5174          */
5175         if (pipe(self_pipe) < 0) {
5176                 ERROR("Error opening internal pipe: %s", fr_syserror(errno));
5177                 fr_exit(1);
5178         }
5179         if ((fcntl(self_pipe[0], F_SETFL, O_NONBLOCK) < 0) ||
5180             (fcntl(self_pipe[0], F_SETFD, FD_CLOEXEC) < 0)) {
5181                 ERROR("Error setting internal flags: %s", fr_syserror(errno));
5182                 fr_exit(1);
5183         }
5184         if ((fcntl(self_pipe[1], F_SETFL, O_NONBLOCK) < 0) ||
5185             (fcntl(self_pipe[1], F_SETFD, FD_CLOEXEC) < 0)) {
5186                 ERROR("Error setting internal flags: %s", fr_syserror(errno));
5187                 fr_exit(1);
5188         }
5189         DEBUG4("Created signal pipe.  Read end FD %i, write end FD %i", self_pipe[0], self_pipe[1]);
5190
5191         if (!fr_event_fd_insert(el, 0, self_pipe[0], event_signal_handler, el)) {
5192                 ERROR("Failed creating signal pipe handler: %s", fr_strerror());
5193                 fr_exit(1);
5194         }
5195 #endif
5196
5197         DEBUG("%s: #### Opening IP addresses and Ports ####", main_config.name);
5198
5199         /*
5200          *      The server temporarily switches to an unprivileged
5201          *      user very early in the bootstrapping process.
5202          *      However, some sockets MAY require privileged access
5203          *      (bind to device, or to port < 1024, or to raw
5204          *      sockets).  Those sockets need to call suid up/down
5205          *      themselves around the functions that need a privileged
5206          *      uid.
5207          */
5208         if (listen_init(cs, &head, spawn_flag) < 0) {
5209                 fr_exit_now(1);
5210         }
5211
5212         main_config.listen = head;
5213
5214         /*
5215         *       At this point, no one has any business *ever* going
5216         *       back to root uid.
5217         */
5218         rad_suid_down_permanent();
5219
5220         return 1;
5221 }
5222
5223
5224 #ifdef WITH_PROXY
5225 static int proxy_delete_cb(UNUSED void *ctx, void *data)
5226 {
5227         REQUEST *request = fr_packet2myptr(REQUEST, proxy, data);
5228
5229         VERIFY_REQUEST(request);
5230
5231         request->master_state = REQUEST_STOP_PROCESSING;
5232
5233 #ifdef HAVE_PTHREAD_H
5234         if (pthread_equal(request->child_pid, NO_SUCH_CHILD_PID) == 0) return 0;
5235 #endif
5236
5237         /*
5238          *      If it's queued we can't delete it from the queue.
5239          *
5240          *      Otherwise, it's OK to delete it.  Even RUNNING, because
5241          *      that will get caught by the check above.
5242          */
5243         if (request->child_state == REQUEST_QUEUED) return 0;
5244
5245         request->in_proxy_hash = false;
5246
5247         if (!request->in_request_hash) {
5248                 request_done(request, FR_ACTION_DONE);
5249         }
5250
5251         /*
5252          *      Delete it from the list.
5253          */
5254         return 2;
5255 }
5256 #endif
5257
5258
5259 static int request_delete_cb(UNUSED void *ctx, void *data)
5260 {
5261         REQUEST *request = fr_packet2myptr(REQUEST, packet, data);
5262
5263         VERIFY_REQUEST(request);
5264
5265         request->master_state = REQUEST_STOP_PROCESSING;
5266
5267         /*
5268          *      Not done, or the child thread is still processing it.
5269          */
5270         if (request->child_state < REQUEST_RESPONSE_DELAY) return 0; /* continue */
5271
5272 #ifdef HAVE_PTHREAD_H
5273         if (pthread_equal(request->child_pid, NO_SUCH_CHILD_PID) == 0) return 0;
5274 #endif
5275
5276 #ifdef WITH_PROXY
5277         rad_assert(request->in_proxy_hash == false);
5278 #endif
5279
5280         request->in_request_hash = false;
5281         ASSERT_MASTER;
5282         if (request->ev) fr_event_delete(el, &request->ev);
5283
5284         if (main_config.memory_report) {
5285                 RDEBUG2("Cleaning up request packet ID %u with timestamp +%d",
5286                         request->packet->id,
5287                         (unsigned int) (request->timestamp - fr_start_time));
5288         }
5289
5290 #ifdef WITH_COA
5291         if (request->coa) {
5292                 rad_assert(!request->coa->in_proxy_hash);
5293         }
5294 #endif
5295
5296         request_free(request);
5297
5298         /*
5299          *      Delete it from the list, and continue;
5300          */
5301         return 2;
5302 }
5303
5304
5305 void radius_event_free(void)
5306 {
5307         ASSERT_MASTER;
5308
5309 #ifdef WITH_PROXY
5310         /*
5311          *      There are requests in the proxy hash that aren't
5312          *      referenced from anywhere else.  Remove them first.
5313          */
5314         if (proxy_list) {
5315                 fr_packet_list_walk(proxy_list, NULL, proxy_delete_cb);
5316         }
5317 #endif
5318
5319         rbtree_walk(pl, RBTREE_DELETE_ORDER,  request_delete_cb, NULL);
5320
5321         if (spawn_flag) {
5322                 /*
5323                  *      Now that all requests have been marked "please stop",
5324                  *      ensure that all of the threads have exited.
5325                  */
5326 #ifdef HAVE_PTHREAD_H
5327                 thread_pool_stop();
5328 #endif
5329
5330                 /*
5331                  *      Walk the lists again, ensuring that all
5332                  *      requests are done.
5333                  */
5334                 if (main_config.memory_report) {
5335                         int num;
5336
5337 #ifdef WITH_PROXY
5338                         if (proxy_list) {
5339                                 fr_packet_list_walk(proxy_list, NULL, proxy_delete_cb);
5340                                 num = fr_packet_list_num_elements(proxy_list);
5341                                 if (num > 0) {
5342                                         ERROR("Proxy list has %d requests still in it.", num);
5343                                 }
5344                         }
5345 #endif
5346
5347                         rbtree_walk(pl, RBTREE_DELETE_ORDER, request_delete_cb, NULL);
5348                         num = rbtree_num_elements(pl);
5349                         if (num > 0) {
5350                                 ERROR("Request list has %d requests still in it.", num);
5351                         }
5352                 }
5353         }
5354
5355         rbtree_free(pl);
5356         pl = NULL;
5357
5358 #ifdef WITH_PROXY
5359         fr_packet_list_free(proxy_list);
5360         proxy_list = NULL;
5361
5362         if (proxy_ctx) talloc_free(proxy_ctx);
5363 #endif
5364
5365         TALLOC_FREE(el);
5366
5367         if (debug_condition) talloc_free(debug_condition);
5368 }
5369
5370 int radius_event_process(void)
5371 {
5372         if (!el) return 0;
5373
5374         return fr_event_loop(el);
5375 }