2 * Copyright (c) 2003-2004, 2007, 2009 Sendmail, Inc. and its suppliers.
5 * By using this file, you agree to the terms and conditions set
6 * forth in the LICENSE file which can be found at the top level of
7 * the sendmail distribution.
9 * Contributed by Jose Marcio Martins da Cruz - Ecole des Mines de Paris
10 * Jose-Marcio.Martins@ensmp.fr
14 SM_RCSID("@(#)$Id: worker.c,v 8.17 2009/06/15 15:34:54 ca Exp $")
16 #include "libmilter.h"
20 typedef struct taskmgr_S taskmgr_T;
22 #define TM_SIGNATURE 0x23021957
26 long tm_signature; /* has the controller been initialized */
27 sthread_t tm_tid; /* thread id of controller */
28 smfi_hd_T tm_ctx_head; /* head of the linked list of contexts */
30 int tm_nb_workers; /* number of workers in the pool */
31 int tm_nb_idle; /* number of workers waiting */
33 int tm_p[2]; /* poll control pipe */
35 smutex_t tm_w_mutex; /* linked list access mutex */
36 scond_t tm_w_cond; /* */
39 static taskmgr_T Tskmgr = {0};
41 #define WRK_CTX_HEAD Tskmgr.tm_ctx_head
43 #define RD_PIPE (Tskmgr.tm_p[0])
44 #define WR_PIPE (Tskmgr.tm_p[1])
46 #define PIPE_SEND_SIGNAL() \
51 if (write(fd, &evt, sizeof(evt)) != sizeof(evt)) \
52 smi_log(SMI_LOG_ERR, \
53 "Error writing to event pipe: %s", \
54 sm_errstring(errno)); \
57 #ifndef USE_PIPE_WAKE_POLL
58 # define USE_PIPE_WAKE_POLL 1
59 #endif /* USE_PIPE_WAKE_POLL */
61 /* poll check periodicity (default 10000 - 10 s) */
62 #define POLL_TIMEOUT 10000
64 /* worker conditional wait timeout (default 10 s) */
65 #define COND_TIMEOUT 10
68 static int mi_close_session __P((SMFICTX_PTR));
70 static void *mi_worker __P((void *));
71 static void *mi_pool_controller __P((void *));
73 static int mi_list_add_ctx __P((SMFICTX_PTR));
74 static int mi_list_del_ctx __P((SMFICTX_PTR));
77 ** periodicity of cleaning up old sessions (timedout)
78 ** sessions list will be checked to find old inactive
79 ** sessions each DT_CHECK_OLD_SESSIONS sec
82 #define DT_CHECK_OLD_SESSIONS 600
84 #ifndef OLD_SESSION_TIMEOUT
85 # define OLD_SESSION_TIMEOUT ctx->ctx_timeout
86 #endif /* OLD_SESSION_TIMEOUT */
88 /* session states - with respect to the pool of workers */
89 #define WKST_INIT 0 /* initial state */
90 #define WKST_READY_TO_RUN 1 /* command ready do be read */
91 #define WKST_RUNNING 2 /* session running on a worker */
92 #define WKST_READY_TO_WAIT 3 /* session just finished by a worker */
93 #define WKST_WAITING 4 /* waiting for new command */
94 #define WKST_CLOSING 5 /* session finished */
97 # define MIN_WORKERS 2 /* minimum number of threads to keep around */
100 #define MIN_IDLE 1 /* minimum number of idle threads */
104 ** Macros for threads and mutex management
107 #define TASKMGR_LOCK() \
110 if (!smutex_lock(&Tskmgr.tm_w_mutex)) \
111 smi_log(SMI_LOG_ERR, "TASKMGR_LOCK error"); \
114 #define TASKMGR_UNLOCK() \
117 if (!smutex_unlock(&Tskmgr.tm_w_mutex)) \
118 smi_log(SMI_LOG_ERR, "TASKMGR_UNLOCK error"); \
121 #define TASKMGR_COND_WAIT() \
122 scond_timedwait(&Tskmgr.tm_w_cond, &Tskmgr.tm_w_mutex, COND_TIMEOUT)
124 #define TASKMGR_COND_SIGNAL() \
127 if (scond_signal(&Tskmgr.tm_w_cond) != 0) \
128 smi_log(SMI_LOG_ERR, "TASKMGR_COND_SIGNAL error"); \
131 #define LAUNCH_WORKER(ctx) \
137 if ((r = thread_create(&tid, mi_worker, ctx)) != 0) \
138 smi_log(SMI_LOG_ERR, "LAUNCH_WORKER error: %s",\
143 # define POOL_LEV_DPRINTF(lev, x) \
145 if ((lev) < ctx->ctx_dbg) \
148 #else /* POOL_DEBUG */
149 # define POOL_LEV_DPRINTF(lev, x)
150 #endif /* POOL_DEBUG */
153 ** MI_START_SESSION -- Start a session in the pool of workers
156 ** ctx -- context structure
159 ** MI_SUCCESS/MI_FAILURE
163 mi_start_session(ctx)
168 SM_ASSERT(Tskmgr.tm_signature == TM_SIGNATURE);
169 SM_ASSERT(ctx != NULL);
170 POOL_LEV_DPRINTF(4, ("PIPE r=[%d] w=[%d]", RD_PIPE, WR_PIPE));
173 if (mi_list_add_ctx(ctx) != MI_SUCCESS)
181 /* if there is an idle worker, signal it, otherwise start new worker */
182 if (Tskmgr.tm_nb_idle > 0)
184 ctx->ctx_wstate = WKST_READY_TO_RUN;
185 TASKMGR_COND_SIGNAL();
189 ctx->ctx_wstate = WKST_RUNNING;
197 ** MI_CLOSE_SESSION -- Close a session and clean up data structures
200 ** ctx -- context structure
203 ** MI_SUCCESS/MI_FAILURE
207 mi_close_session(ctx)
210 SM_ASSERT(ctx != NULL);
212 (void) mi_list_del_ctx(ctx);
219 ** MI_POOL_CONTROLER_INIT -- Launch the worker pool controller
220 ** Must be called before starting sessions.
226 ** MI_SUCCESS/MI_FAILURE
230 mi_pool_controller_init()
235 if (Tskmgr.tm_signature == TM_SIGNATURE)
238 SM_TAILQ_INIT(&WRK_CTX_HEAD);
239 Tskmgr.tm_tid = (sthread_t) -1;
240 Tskmgr.tm_nb_workers = 0;
241 Tskmgr.tm_nb_idle = 0;
243 if (pipe(Tskmgr.tm_p) != 0)
245 smi_log(SMI_LOG_ERR, "can't create event pipe: %s",
246 sm_errstring(errno));
250 (void) smutex_init(&Tskmgr.tm_w_mutex);
251 (void) scond_init(&Tskmgr.tm_w_cond);
253 /* Launch the pool controller */
254 if ((r = thread_create(&tid, mi_pool_controller, (void *) NULL)) != 0)
256 smi_log(SMI_LOG_ERR, "can't create controller thread: %s",
261 Tskmgr.tm_signature = TM_SIGNATURE;
263 /* Create the pool of workers */
264 for (i = 0; i < MIN_WORKERS; i++)
266 if ((r = thread_create(&tid, mi_worker, (void *) NULL)) != 0)
268 smi_log(SMI_LOG_ERR, "can't create workers crew: %s",
278 ** MI_POOL_CONTROLLER -- manage the pool of workers
279 ** This thread must be running when listener begins
290 ** Look for timed out sessions
291 ** Select sessions to wait for sendmail command
292 ** Poll set of file descriptors
295 ** For each file descriptor ready
296 ** launch new thread if no worker available
298 ** signal waiting worker
301 /* Poll structure array (pollfd) size step */
304 #define WAIT_FD(i) (pfd[i].fd)
305 #define WAITFN "POLL"
308 mi_pool_controller(arg)
311 struct pollfd *pfd = NULL;
313 bool rebuild_set = true;
314 int pcnt = 0; /* error count for poll() failures */
317 Tskmgr.tm_tid = sthread_get_id();
318 if (pthread_detach(Tskmgr.tm_tid) != 0)
320 smi_log(SMI_LOG_ERR, "Failed to detach pool controller thread");
324 pfd = (struct pollfd *) malloc(PFD_STEP * sizeof(struct pollfd));
327 smi_log(SMI_LOG_ERR, "Failed to malloc pollfd array: %s",
328 sm_errstring(errno));
333 lastcheck = time(NULL);
340 POOL_LEV_DPRINTF(4, ("Let's %s again...", WAITFN));
342 if (mi_stop() != MILTER_CONT)
349 /* check for timed out sessions? */
350 if (lastcheck + DT_CHECK_OLD_SESSIONS < now)
352 ctx = SM_TAILQ_FIRST(&WRK_CTX_HEAD);
353 while (ctx != SM_TAILQ_END(&WRK_CTX_HEAD))
357 ctx_nxt = SM_TAILQ_NEXT(ctx, ctx_link);
358 if (ctx->ctx_wstate == WKST_WAITING)
360 if (ctx->ctx_wait == 0)
362 else if (ctx->ctx_wait + OLD_SESSION_TIMEOUT
365 /* if session timed out, close it */
366 sfsistat (*fi_close) __P((SMFICTX *));
369 ("Closing old connection: sd=%d id=%d",
373 if ((fi_close = ctx->ctx_smfi->xxfi_close) != NULL)
374 (void) (*fi_close)(ctx);
376 mi_close_session(ctx);
387 ** Initialize poll set.
388 ** Insert into the poll set the file descriptors of
389 ** all sessions waiting for a command from sendmail.
394 /* begin with worker pipe */
395 pfd[nfd].fd = RD_PIPE;
396 pfd[nfd].events = MI_POLL_RD_FLAGS;
397 pfd[nfd].revents = 0;
400 SM_TAILQ_FOREACH(ctx, &WRK_CTX_HEAD, ctx_link)
403 ** update ctx_wait - start of wait moment -
407 if (ctx->ctx_wstate == WKST_READY_TO_WAIT)
410 /* add the session to the pollfd array? */
411 if ((ctx->ctx_wstate == WKST_READY_TO_WAIT) ||
412 (ctx->ctx_wstate == WKST_WAITING))
415 ** Resize the pollfd array if it
416 ** isn't large enough.
424 new = (dim_pfd + PFD_STEP) *
426 tpfd = (struct pollfd *)
436 "Failed to realloc pollfd array:%s",
437 sm_errstring(errno));
441 /* add the session to pollfd array */
444 ctx->ctx_wstate = WKST_WAITING;
445 pfd[nfd].fd = ctx->ctx_sd;
446 pfd[nfd].events = MI_POLL_RD_FLAGS;
447 pfd[nfd].revents = 0;
457 /* Everything is ready, let's wait for an event */
458 rfd = poll(pfd, nfd, POLL_TIMEOUT);
460 POOL_LEV_DPRINTF(4, ("%s returned: at epoch %d value %d",
476 "%s() failed (%s), %s",
477 WAITFN, sm_errstring(errno),
478 pcnt >= MAX_FAILS_S ? "abort" : "try again");
480 if (pcnt >= MAX_FAILS_S)
485 /* something happened */
486 for (i = 0; i < nfd; i++)
488 if (pfd[i].revents == 0)
491 POOL_LEV_DPRINTF(4, ("%s event on pfd[%d/%d]=%d ",
495 /* has a worker signaled an end of task ? */
496 if (WAIT_FD(i) == RD_PIPE)
502 ("PIPE WILL READ evt = %08X %08X",
503 pfd[i].events, pfd[i].revents));
505 if ((pfd[i].revents & MI_POLL_RD_FLAGS) != 0)
507 r = read(RD_PIPE, &evt, sizeof(evt));
508 if (r == sizeof(evt))
515 ("PIPE DONE READ i=[%d] fd=[%d] r=[%d] evt=[%d]",
516 i, RD_PIPE, r, evt));
518 if ((pfd[i].revents & ~MI_POLL_RD_FLAGS) != 0)
520 /* Exception handling */
525 /* no ! sendmail wants to send a command */
526 SM_TAILQ_FOREACH(ctx, &WRK_CTX_HEAD, ctx_link)
528 if (ctx->ctx_wstate != WKST_WAITING)
532 ("Checking context sd=%d - fd=%d ",
533 ctx->ctx_sd , WAIT_FD(i)));
535 if (ctx->ctx_sd == pfd[i].fd)
540 ("TASK: found %d for fd[%d]=%d",
541 ctx->ctx_sid, i, WAIT_FD(i)));
543 if (Tskmgr.tm_nb_idle > 0)
545 ctx->ctx_wstate = WKST_READY_TO_RUN;
546 TASKMGR_COND_SIGNAL();
550 ctx->ctx_wstate = WKST_RUNNING;
559 ("TASK %s FOUND - Checking PIPE for fd[%d]",
560 ctx != NULL ? "" : "NOT", WAIT_FD(i)));
568 Tskmgr.tm_signature = 0;
573 ctx = SM_TAILQ_FIRST(&WRK_CTX_HEAD);
576 mi_close_session(ctx);
579 (void) smutex_destroy(&Tskmgr.tm_w_mutex);
580 (void) scond_destroy(&Tskmgr.tm_w_cond);
586 ** Look for a task ready to run.
587 ** Value of ctx is NULL or a pointer to a task ready to run.
590 #define GET_TASK_READY_TO_RUN() \
591 SM_TAILQ_FOREACH(ctx, &WRK_CTX_HEAD, ctx_link) \
593 if (ctx->ctx_wstate == WKST_READY_TO_RUN) \
595 ctx->ctx_wstate = WKST_RUNNING; \
601 ** MI_WORKER -- worker thread
602 ** executes tasks distributed by the mi_pool_controller
603 ** or by mi_start_session
606 ** arg -- pointer to context structure
621 ctx = (SMFICTX_PTR) arg;
624 ctx->ctx_wstate = WKST_RUNNING;
626 t_id = sthread_get_id();
627 if (pthread_detach(t_id) != 0)
629 smi_log(SMI_LOG_ERR, "Failed to detach worker thread");
631 ctx->ctx_wstate = WKST_READY_TO_RUN;
636 Tskmgr.tm_nb_workers++;
641 if (mi_stop() != MILTER_CONT)
644 /* let's handle next task... */
650 ("worker %d: new task -> let's handle it",
652 res = mi_engine(ctx);
654 ("worker %d: mi_engine returned %d", t_id, res));
657 if (res != MI_CONTINUE)
659 ctx->ctx_wstate = WKST_CLOSING;
662 ** Delete context from linked list of
663 ** sessions and close session.
666 mi_close_session(ctx);
670 ctx->ctx_wstate = WKST_READY_TO_WAIT;
673 ("writing to event pipe..."));
676 ** Signal task controller to add new session
687 /* check if there is any task waiting to be served */
690 GET_TASK_READY_TO_RUN();
700 ** if not, let's check if there is enough idle workers
704 if (Tskmgr.tm_nb_workers > MIN_WORKERS &&
705 Tskmgr.tm_nb_idle > MIN_IDLE)
708 POOL_LEV_DPRINTF(4, ("worker %d: checking ... %d %d", t_id,
709 Tskmgr.tm_nb_workers, Tskmgr.tm_nb_idle + 1));
713 POOL_LEV_DPRINTF(4, ("worker %d: quitting... ", t_id));
714 Tskmgr.tm_nb_workers--;
720 ** if no task ready to run, wait for another one
727 /* look for a task */
728 GET_TASK_READY_TO_RUN();
736 ** MI_LIST_ADD_CTX -- add new session to linked list
739 ** ctx -- context structure
742 ** MI_FAILURE/MI_SUCCESS
749 SM_ASSERT(ctx != NULL);
750 SM_TAILQ_INSERT_TAIL(&WRK_CTX_HEAD, ctx, ctx_link);
755 ** MI_LIST_DEL_CTX -- remove session from linked list when finished
758 ** ctx -- context structure
761 ** MI_FAILURE/MI_SUCCESS
768 SM_ASSERT(ctx != NULL);
769 if (SM_TAILQ_EMPTY(&WRK_CTX_HEAD))
772 SM_TAILQ_REMOVE(&WRK_CTX_HEAD, ctx, ctx_link);
775 #endif /* _FFR_WORKERS_POOL */