2 Unix SMB/CIFS implementation.
4 process model: prefork (n client connections per process)
6 Copyright (C) Andrew Tridgell 1992-2005
7 Copyright (C) James J Myers 2003 <myersjj@samba.org>
8 Copyright (C) Stefan (metze) Metzmacher 2004
9 Copyright (C) Andrew Bartlett 2008 <abartlet@samba.org>
10 Copyright (C) David Disseldorp 2008 <ddiss@sgi.com>
12 This program is free software; you can redistribute it and/or modify
13 it under the terms of the GNU General Public License as published by
14 the Free Software Foundation; either version 3 of the License, or
15 (at your option) any later version.
17 This program is distributed in the hope that it will be useful,
18 but WITHOUT ANY WARRANTY; without even the implied warranty of
19 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
20 GNU General Public License for more details.
22 You should have received a copy of the GNU General Public License
23 along with this program. If not, see <http://www.gnu.org/licenses/>.
26 * The pre-fork process model distributes the server workload amongst several
27 * designated worker threads (e.g. 'prefork-worker-ldap-0',
28 * 'prefork-worker-ldap-1', etc). The number of worker threads is controlled
29 * by the 'prefork children' conf setting. The worker threads are controlled
30 * by a prefork master process (e.g. 'prefork-master-ldap'). The prefork master
31 * doesn't handle the server workload (i.e. processing messages) itself, but is
32 * responsible for restarting workers if they exit unexpectedly. The top-level
33 * samba process is responsible for restarting the master process if it exits.
38 #include "lib/events/events.h"
39 #include "lib/messaging/messaging.h"
40 #include "lib/socket/socket.h"
41 #include "smbd/process_model.h"
42 #include "cluster/cluster.h"
43 #include "param/param.h"
45 #include "lib/util/tfork.h"
46 #include "lib/messaging/irpc.h"
48 #define min(a, b) (((a) < (b)) ? (a) : (b))
50 NTSTATUS process_model_prefork_init(void);
51 static void prefork_new_task(
52 struct tevent_context *ev,
53 struct loadparm_context *lp_ctx,
54 const char *service_name,
55 struct task_server *(*new_task_fn)(struct tevent_context *,
56 struct loadparm_context *lp_ctx,
61 const struct service_details *service_details,
63 static void prefork_fork_worker(struct task_server *task,
64 struct tevent_context *ev,
65 struct tevent_context *ev2,
66 struct loadparm_context *lp_ctx,
67 const struct service_details *service_details,
68 const char *service_name,
70 unsigned restart_delay,
71 struct process_details *pd);
72 static void prefork_child_pipe_handler(struct tevent_context *ev,
73 struct tevent_fd *fde,
76 static void setup_handlers(struct tevent_context *ev,
77 struct loadparm_context *lp_ctx,
81 * State needed to restart the master process or a worker process if they
84 struct master_restart_context {
85 struct task_server *(*new_task_fn)(struct tevent_context *,
86 struct loadparm_context *lp_ctx,
93 struct worker_restart_context {
94 unsigned int instance;
95 struct task_server *task;
96 struct tevent_context *ev2;
100 struct restart_context {
101 struct loadparm_context *lp_ctx;
104 const struct service_details *service_details;
105 const char *service_name;
106 unsigned restart_delay;
107 struct master_restart_context *master;
108 struct worker_restart_context *worker;
111 static void sighup_signal_handler(struct tevent_context *ev,
112 struct tevent_signal *se,
113 int signum, int count, void *siginfo,
116 debug_schedule_reopen_logs();
119 static void sigterm_signal_handler(struct tevent_context *ev,
120 struct tevent_signal *se,
121 int signum, int count, void *siginfo,
125 if (getpgrp() == getpid()) {
127 * We're the process group leader, send
128 * SIGTERM to our process group.
130 DBG_NOTICE("SIGTERM: killing children\n");
131 kill(-getpgrp(), SIGTERM);
134 DBG_NOTICE("Exiting pid %d on SIGTERM\n", getpid());
140 called when the process model is selected
142 static void prefork_model_init(void)
146 static void prefork_reload_after_fork(void)
150 ldb_wrap_fork_hook();
151 /* Must be done after a fork() to reset messaging contexts. */
152 status = imessaging_reinit_all();
153 if (!NT_STATUS_IS_OK(status)) {
154 smb_panic("Failed to re-initialise imessaging after fork");
159 * clean up any messaging associated with the old process.
162 static void irpc_cleanup(
163 struct loadparm_context *lp_ctx,
164 struct tevent_context *ev,
167 TALLOC_CTX *mem_ctx = talloc_new(NULL);
168 struct imessaging_context *msg_ctx = NULL;
169 NTSTATUS status = NT_STATUS_OK;
171 if (mem_ctx == NULL) {
172 DBG_ERR("OOM cleaning up irpc\n");
175 msg_ctx = imessaging_client_init(mem_ctx, lp_ctx, ev);
176 if (msg_ctx == NULL) {
177 DBG_ERR("Unable to create imessaging_context\n");
178 TALLOC_FREE(mem_ctx);
181 status = imessaging_process_cleanup(msg_ctx, pid);
182 if (!NT_STATUS_IS_OK(status)) {
183 DBG_ERR("imessaging_process_cleanup returned (%s)\n",
185 TALLOC_FREE(mem_ctx);
189 TALLOC_FREE(mem_ctx);
193 * handle EOF on the parent-to-all-children pipe in the child, i.e.
194 * the parent has died and its end of the pipe has been closed.
195 * The child handles this by exiting as well.
197 static void prefork_pipe_handler(struct tevent_context *event_ctx,
198 struct tevent_fd *fde, uint16_t flags,
201 struct loadparm_context *lp_ctx = NULL;
205 * free the fde which removes the event and stops it firing again
210 * Clean up any irpc end points this process had.
213 lp_ctx = talloc_get_type_abort(private_data, struct loadparm_context);
214 irpc_cleanup(lp_ctx, event_ctx, pid);
216 DBG_NOTICE("Child %d exiting\n", getpid());
217 TALLOC_FREE(event_ctx);
223 * Called by the top-level samba process to create a new prefork master process
225 static void prefork_fork_master(
226 struct tevent_context *ev,
227 struct loadparm_context *lp_ctx,
228 const char *service_name,
229 struct task_server *(*new_task_fn)(struct tevent_context *,
230 struct loadparm_context *lp_ctx,
235 const struct service_details *service_details,
236 unsigned restart_delay,
240 struct tfork* t = NULL;
243 struct tevent_context *ev2;
244 struct task_server *task = NULL;
245 struct process_details pd = initial_process_details;
250 smb_panic("failure in tfork\n");
253 DBG_NOTICE("Forking [%s] pre-fork master process\n", service_name);
254 pid = tfork_child_pid(t);
256 struct tevent_fd *fde = NULL;
257 int fd = tfork_event_fd(t);
258 struct restart_context *rc = NULL;
260 /* Register a pipe handler that gets called when the prefork
261 * master process terminates.
263 rc = talloc_zero(ev, struct restart_context);
265 smb_panic("OOM allocating restart context\n");
269 rc->service_name = service_name;
270 rc->service_details = service_details;
271 rc->from_parent_fd = from_parent_fd;
272 rc->restart_delay = restart_delay;
273 rc->master = talloc_zero(rc, struct master_restart_context);
274 if (rc->master == NULL) {
275 smb_panic("OOM allocating master restart context\n");
278 rc->master->new_task_fn = new_task_fn;
279 rc->master->private_data = private_data;
282 ev, ev, fd, TEVENT_FD_READ, prefork_child_pipe_handler, rc);
284 smb_panic("Failed to add child pipe handler, "
287 tevent_fd_set_auto_close(fde);
292 setproctitle("task[%s] pre-fork master", service_name);
295 * this will free all the listening sockets and all state that
296 * is not associated with this new connection
298 if (tevent_re_initialise(ev) != 0) {
299 smb_panic("Failed to re-initialise tevent after fork");
301 prefork_reload_after_fork();
302 setup_handlers(ev, lp_ctx, from_parent_fd);
304 if (service_details->inhibit_pre_fork) {
306 ev, lp_ctx, cluster_id(pid, 0), private_data, NULL);
308 * The task does not support pre-fork
310 if (task != NULL && service_details->post_fork != NULL) {
311 service_details->post_fork(task, &pd);
313 tevent_loop_wait(ev);
319 * This is now the child code. We need a completely new event_context
322 ev2 = s4_event_context_init(NULL);
324 /* setup this new connection: process will bind to it's sockets etc
326 * While we can use ev for the child, which has been re-initialised
327 * above we must run the new task under ev2 otherwise the children would
328 * be listening on the sockets. Also we don't want the top level
329 * process accepting and handling requests, it's responsible for
330 * monitoring and controlling the child work processes.
332 task = new_task_fn(ev2, lp_ctx, cluster_id(pid, 0), private_data, NULL);
340 * Register an irpc name that can be used by the samba-tool processes
344 struct talloc_ctx *ctx = talloc_new(NULL);
347 DBG_ERR("Out of memory");
350 name = talloc_asprintf(ctx, "prefork-master-%s", service_name);
351 irpc_add_name(task->msg_ctx, name);
356 int default_children;
357 default_children = lpcfg_prefork_children(lp_ctx);
358 num_children = lpcfg_parm_int(lp_ctx, NULL, "prefork children",
359 service_name, default_children);
361 if (num_children == 0) {
362 DBG_WARNING("Number of pre-fork children for %s is zero, "
363 "NO worker processes will be started for %s\n",
364 service_name, service_name);
366 DBG_NOTICE("Forking %d %s worker processes\n",
367 num_children, service_name);
370 * the prefork master creates its own control pipe, so the prefork
371 * workers can detect if the master exits (in which case an EOF gets
372 * written). (Whereas from_parent_fd is the control pipe from the
373 * top-level process that the prefork master listens on)
377 ret = pipe(control_pipe);
379 smb_panic("Unable to create worker control pipe\n");
381 smb_set_close_on_exec(control_pipe[0]);
382 smb_set_close_on_exec(control_pipe[1]);
386 * We are now free to spawn some worker processes
388 for (i=0; i < num_children; i++) {
389 prefork_fork_worker(task,
401 /* Don't listen on the sockets we just gave to the children */
402 tevent_loop_wait(ev);
404 /* We need to keep ev2 until we're finished for the messaging to work */
410 * Restarts a child process if it exits unexpectedly
412 static void prefork_restart(struct tevent_context *ev,
413 struct restart_context *rc)
415 unsigned max_backoff = 0;
416 unsigned backoff = 0;
417 unsigned restart_delay = rc->restart_delay;
418 unsigned default_value = 0;
421 * If the child process is constantly exiting, then restarting it can
422 * consume a lot of resources. In which case, we want to backoff a bit
423 * before respawning it
425 default_value = lpcfg_prefork_backoff_increment(rc->lp_ctx);
426 backoff = lpcfg_parm_int(rc->lp_ctx,
428 "prefork backoff increment",
432 default_value = lpcfg_prefork_maximum_backoff(rc->lp_ctx);
433 max_backoff = lpcfg_parm_int(rc->lp_ctx,
435 "prefork maximum backoff",
439 if (restart_delay > 0) {
440 DBG_ERR("Restarting [%s] pre-fork %s in (%d) seconds\n",
442 (rc->master == NULL) ? "worker" : "master",
444 sleep(restart_delay);
446 restart_delay += backoff;
447 restart_delay = min(restart_delay, max_backoff);
449 if (rc->master != NULL) {
450 DBG_ERR("Restarting [%s] pre-fork master\n", rc->service_name);
451 prefork_fork_master(ev,
454 rc->master->new_task_fn,
455 rc->master->private_data,
459 } else if (rc->worker != NULL) {
460 struct process_details pd = initial_process_details;
461 DBG_ERR("Restarting [%s] pre-fork worker(%d)\n",
463 rc->worker->instance);
464 pd.instances = rc->worker->instance;
465 prefork_fork_worker(rc->worker->task,
471 rc->worker->control_pipe,
478 handle EOF on the child pipe in the parent, so we know when a
479 process terminates without using SIGCHLD or waiting on all possible pids.
481 We need to ensure we do not ignore SIGCHLD because we need it to
482 work to get a valid error code from samba_runcmd_*().
484 static void prefork_child_pipe_handler(struct tevent_context *ev,
485 struct tevent_fd *fde,
489 struct restart_context *rc = NULL;
493 /* free the fde which removes the event and stops it firing again */
496 /* the child has closed the pipe, assume its dead */
498 rc = talloc_get_type_abort(private_data, struct restart_context);
499 pid = tfork_child_pid(rc->t);
502 irpc_cleanup(rc->lp_ctx, ev, pid);
503 status = tfork_status(&rc->t, false);
505 DBG_ERR("Parent %d, Child %d terminated, "
506 "unable to get status code from tfork\n",
508 prefork_restart(ev, rc);
509 } else if (WIFEXITED(status)) {
510 status = WEXITSTATUS(status);
511 DBG_ERR("Parent %d, Child %d exited with status %d\n",
512 getpid(), pid, status);
514 prefork_restart(ev, rc);
516 } else if (WIFSIGNALED(status)) {
517 status = WTERMSIG(status);
518 DBG_ERR("Parent %d, Child %d terminated with signal %d\n",
519 getpid(), pid, status);
520 if (status == SIGABRT || status == SIGBUS || status == SIGFPE ||
521 status == SIGILL || status == SIGSYS || status == SIGSEGV) {
523 prefork_restart(ev, rc);
526 /* tfork allocates tfork structures with malloc */
527 tfork_destroy(&rc->t);
534 called when a listening socket becomes readable.
536 static void prefork_accept_connection(
537 struct tevent_context *ev,
538 struct loadparm_context *lp_ctx,
539 struct socket_context *listen_socket,
540 void (*new_conn)(struct tevent_context *,
541 struct loadparm_context *,
542 struct socket_context *,
547 void *process_context)
550 struct socket_context *connected_socket;
551 pid_t pid = getpid();
553 /* accept an incoming connection. */
554 status = socket_accept(listen_socket, &connected_socket);
555 if (!NT_STATUS_IS_OK(status)) {
557 * For prefork we can ignore STATUS_MORE_ENTRIES, as once a
558 * connection becomes available all waiting processes are
559 * woken, but only one gets work to process.
560 * AKA the thundering herd.
561 * In the short term this should not be an issue as the number
562 * of workers should be a small multiple of the number of cpus
563 * In the longer term socket_accept needs to implement a
564 * mutex/semaphore (like apache does) to serialise the accepts
566 if (!NT_STATUS_EQUAL(status, STATUS_MORE_ENTRIES)) {
567 DBG_ERR("Worker process (%d), error in accept [%s]\n",
568 getpid(), nt_errstr(status));
573 talloc_steal(private_data, connected_socket);
575 new_conn(ev, lp_ctx, connected_socket,
576 cluster_id(pid, socket_get_fd(connected_socket)),
577 private_data, process_context);
580 static void setup_handlers(
581 struct tevent_context *ev,
582 struct loadparm_context *lp_ctx,
585 struct tevent_fd *fde = NULL;
586 struct tevent_signal *se = NULL;
588 fde = tevent_add_fd(ev, ev, from_parent_fd, TEVENT_FD_READ,
589 prefork_pipe_handler, lp_ctx);
591 smb_panic("Failed to add fd handler after fork");
594 se = tevent_add_signal(ev,
598 sighup_signal_handler,
601 smb_panic("Failed to add SIGHUP handler after fork");
604 se = tevent_add_signal(ev,
608 sigterm_signal_handler,
611 smb_panic("Failed to add SIGTERM handler after fork");
616 * Called by the prefork master to create a new prefork worker process
618 static void prefork_fork_worker(struct task_server *task,
619 struct tevent_context *ev,
620 struct tevent_context *ev2,
621 struct loadparm_context *lp_ctx,
622 const struct service_details *service_details,
623 const char *service_name,
625 unsigned restart_delay,
626 struct process_details *pd)
628 struct tfork *w = NULL;
633 smb_panic("failure in tfork\n");
636 pid = tfork_child_pid(w);
638 struct tevent_fd *fde = NULL;
639 int fd = tfork_event_fd(w);
640 struct restart_context *rc = NULL;
643 * we're the parent (prefork master), so store enough info to
644 * restart the worker/child if it exits unexpectedly
646 rc = talloc_zero(ev, struct restart_context);
648 smb_panic("OOM allocating restart context\n");
652 rc->service_name = service_name;
653 rc->service_details = service_details;
654 rc->restart_delay = restart_delay;
656 rc->worker = talloc_zero(rc, struct worker_restart_context);
657 if (rc->worker == NULL) {
658 smb_panic("OOM allocating master restart context\n");
660 rc->worker->ev2 = ev2;
661 rc->worker->instance = pd->instances;
662 rc->worker->task = task;
663 rc->worker->control_pipe[0] = control_pipe[0];
664 rc->worker->control_pipe[1] = control_pipe[1];
667 ev, ev, fd, TEVENT_FD_READ, prefork_child_pipe_handler, rc);
669 smb_panic("Failed to add child pipe handler, "
672 tevent_fd_set_auto_close(fde);
676 * we're the child (prefork-worker). We never write to the
677 * control pipe, but listen on the read end in case our parent
678 * (the pre-fork master) exits
680 close(control_pipe[1]);
681 setup_handlers(ev2, lp_ctx, control_pipe[0]);
689 setproctitle("task[%s] pre-forked worker(%d)",
692 prefork_reload_after_fork();
693 if (service_details->post_fork != NULL) {
694 service_details->post_fork(task, pd);
697 struct talloc_ctx *ctx = talloc_new(NULL);
700 smb_panic("OOM allocating talloc context\n");
702 name = talloc_asprintf(ctx,
703 "prefork-worker-%s-%d",
706 irpc_add_name(task->msg_ctx, name);
709 tevent_loop_wait(ev2);
715 * called to create a new server task
717 static void prefork_new_task(
718 struct tevent_context *ev,
719 struct loadparm_context *lp_ctx,
720 const char *service_name,
721 struct task_server *(*new_task_fn)(struct tevent_context *,
722 struct loadparm_context *lp_ctx,
723 struct server_id , void *, void *),
725 const struct service_details *service_details,
728 prefork_fork_master(ev,
740 * called when a task terminates
742 static void prefork_terminate_task(struct tevent_context *ev,
743 struct loadparm_context *lp_ctx,
746 void *process_context)
748 DBG_DEBUG("called with reason[%s]\n", reason);
758 * called when a connection completes
760 static void prefork_terminate_connection(struct tevent_context *ev,
761 struct loadparm_context *lp_ctx,
763 void *process_context)
767 /* called to set a title of a task or connection */
768 static void prefork_set_title(struct tevent_context *ev, const char *title)
772 static const struct model_ops prefork_ops = {
774 .model_init = prefork_model_init,
775 .accept_connection = prefork_accept_connection,
776 .new_task = prefork_new_task,
777 .terminate_task = prefork_terminate_task,
778 .terminate_connection = prefork_terminate_connection,
779 .set_title = prefork_set_title,
783 * initialise the prefork process model, registering ourselves with the
784 * process model subsystem
786 NTSTATUS process_model_prefork_init(void)
788 return register_process_model(&prefork_ops);