1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 #define DEBUG_SUBSYSTEM S_RPC
39 #include <liblustre.h>
41 #include <obd_support.h>
42 #include <obd_class.h>
43 #include <lustre_net.h>
44 #include <lu_object.h>
45 #include <lnet/types.h>
46 #include "ptlrpc_internal.h"
48 /* The following are visible and mutable through /sys/module/ptlrpc */
49 int test_req_buffer_pressure = 0;
50 CFS_MODULE_PARM(test_req_buffer_pressure, "i", int, 0444,
51 "set non-zero to put pressure on request buffer pools");
52 unsigned int at_min = 0;
53 CFS_MODULE_PARM(at_min, "i", int, 0644,
54 "Adaptive timeout minimum (sec)");
55 unsigned int at_max = 600;
56 EXPORT_SYMBOL(at_max);
57 CFS_MODULE_PARM(at_max, "i", int, 0644,
58 "Adaptive timeout maximum (sec)");
59 unsigned int at_history = 600;
60 CFS_MODULE_PARM(at_history, "i", int, 0644,
61 "Adaptive timeouts remember the slowest event that took place "
62 "within this period (sec)");
63 static int at_early_margin = 5;
64 CFS_MODULE_PARM(at_early_margin, "i", int, 0644,
65 "How soon before an RPC deadline to send an early reply");
66 static int at_extra = 30;
67 CFS_MODULE_PARM(at_extra, "i", int, 0644,
68 "How much extra time to give with each early reply");
72 static int ptlrpc_server_post_idle_rqbds (struct ptlrpc_service *svc);
74 static CFS_LIST_HEAD(ptlrpc_all_services);
75 spinlock_t ptlrpc_all_services_lock;
78 ptlrpc_alloc_request_buffer (int size)
82 if (size > SVC_BUF_VMALLOC_THRESHOLD)
83 OBD_VMALLOC(ptr, size);
91 ptlrpc_free_request_buffer (char *ptr, int size)
93 if (size > SVC_BUF_VMALLOC_THRESHOLD)
99 struct ptlrpc_request_buffer_desc *
100 ptlrpc_alloc_rqbd (struct ptlrpc_service *svc)
102 struct ptlrpc_request_buffer_desc *rqbd;
108 rqbd->rqbd_service = svc;
109 rqbd->rqbd_refcount = 0;
110 rqbd->rqbd_cbid.cbid_fn = request_in_callback;
111 rqbd->rqbd_cbid.cbid_arg = rqbd;
112 CFS_INIT_LIST_HEAD(&rqbd->rqbd_reqs);
113 rqbd->rqbd_buffer = ptlrpc_alloc_request_buffer(svc->srv_buf_size);
115 if (rqbd->rqbd_buffer == NULL) {
120 spin_lock(&svc->srv_lock);
121 list_add(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
123 spin_unlock(&svc->srv_lock);
129 ptlrpc_free_rqbd (struct ptlrpc_request_buffer_desc *rqbd)
131 struct ptlrpc_service *svc = rqbd->rqbd_service;
133 LASSERT (rqbd->rqbd_refcount == 0);
134 LASSERT (list_empty(&rqbd->rqbd_reqs));
136 spin_lock(&svc->srv_lock);
137 list_del(&rqbd->rqbd_list);
139 spin_unlock(&svc->srv_lock);
141 ptlrpc_free_request_buffer (rqbd->rqbd_buffer, svc->srv_buf_size);
146 ptlrpc_grow_req_bufs(struct ptlrpc_service *svc)
148 struct ptlrpc_request_buffer_desc *rqbd;
151 CDEBUG(D_RPCTRACE, "%s: allocate %d new %d-byte reqbufs (%d/%d left)\n",
152 svc->srv_name, svc->srv_nbuf_per_group, svc->srv_buf_size,
153 svc->srv_nrqbd_receiving, svc->srv_nbufs);
154 for (i = 0; i < svc->srv_nbuf_per_group; i++) {
155 rqbd = ptlrpc_alloc_rqbd(svc);
158 CERROR ("%s: Can't allocate request buffer\n",
163 if (ptlrpc_server_post_idle_rqbds(svc) < 0)
171 ptlrpc_save_lock (struct ptlrpc_request *req,
172 struct lustre_handle *lock, int mode)
174 struct ptlrpc_reply_state *rs = req->rq_reply_state;
178 LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
180 idx = rs->rs_nlocks++;
181 rs->rs_locks[idx] = *lock;
182 rs->rs_modes[idx] = mode;
183 rs->rs_difficult = 1;
187 ptlrpc_schedule_difficult_reply (struct ptlrpc_reply_state *rs)
189 struct ptlrpc_service *svc = rs->rs_service;
192 LASSERT (spin_is_locked (&svc->srv_lock));
194 LASSERT (rs->rs_difficult);
195 rs->rs_scheduled_ever = 1; /* flag any notification attempt */
197 if (rs->rs_scheduled) /* being set up or already notified */
200 rs->rs_scheduled = 1;
201 list_del (&rs->rs_list);
202 list_add (&rs->rs_list, &svc->srv_reply_queue);
203 cfs_waitq_signal (&svc->srv_waitq);
207 ptlrpc_commit_replies (struct obd_device *obd)
209 struct list_head *tmp;
210 struct list_head *nxt;
212 /* Find any replies that have been committed and get their service
213 * to attend to complete them. */
215 /* CAVEAT EMPTOR: spinlock ordering!!! */
216 spin_lock(&obd->obd_uncommitted_replies_lock);
218 list_for_each_safe (tmp, nxt, &obd->obd_uncommitted_replies) {
219 struct ptlrpc_reply_state *rs =
220 list_entry(tmp, struct ptlrpc_reply_state, rs_obd_list);
222 LASSERT (rs->rs_difficult);
224 if (rs->rs_transno <= obd->obd_last_committed) {
225 struct ptlrpc_service *svc = rs->rs_service;
227 spin_lock (&svc->srv_lock);
228 list_del_init (&rs->rs_obd_list);
229 ptlrpc_schedule_difficult_reply (rs);
230 spin_unlock (&svc->srv_lock);
234 spin_unlock(&obd->obd_uncommitted_replies_lock);
238 ptlrpc_server_post_idle_rqbds (struct ptlrpc_service *svc)
240 struct ptlrpc_request_buffer_desc *rqbd;
245 spin_lock(&svc->srv_lock);
247 if (list_empty (&svc->srv_idle_rqbds)) {
248 spin_unlock(&svc->srv_lock);
252 rqbd = list_entry(svc->srv_idle_rqbds.next,
253 struct ptlrpc_request_buffer_desc,
255 list_del (&rqbd->rqbd_list);
257 /* assume we will post successfully */
258 svc->srv_nrqbd_receiving++;
259 list_add (&rqbd->rqbd_list, &svc->srv_active_rqbds);
261 spin_unlock(&svc->srv_lock);
263 rc = ptlrpc_register_rqbd(rqbd);
270 spin_lock(&svc->srv_lock);
272 svc->srv_nrqbd_receiving--;
273 list_del(&rqbd->rqbd_list);
274 list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
276 /* Don't complain if no request buffers are posted right now; LNET
277 * won't drop requests because we set the portal lazy! */
279 spin_unlock(&svc->srv_lock);
284 struct ptlrpc_service *ptlrpc_init_svc_conf(struct ptlrpc_service_conf *c,
285 svc_handler_t h, char *name,
286 struct proc_dir_entry *proc_entry,
287 svcreq_printfn_t prntfn,
290 return ptlrpc_init_svc(c->psc_nbufs, c->psc_bufsize,
291 c->psc_max_req_size, c->psc_max_reply_size,
292 c->psc_req_portal, c->psc_rep_portal,
293 c->psc_watchdog_factor,
295 prntfn, c->psc_min_threads, c->psc_max_threads,
296 threadname, c->psc_ctx_tags);
298 EXPORT_SYMBOL(ptlrpc_init_svc_conf);
300 static void ptlrpc_at_timer(unsigned long castmeharder)
302 struct ptlrpc_service *svc = (struct ptlrpc_service *)castmeharder;
303 svc->srv_at_check = 1;
304 svc->srv_at_checktime = cfs_time_current();
305 cfs_waitq_signal(&svc->srv_waitq);
308 /* @threadname should be 11 characters or less - 3 will be added on */
309 struct ptlrpc_service *
310 ptlrpc_init_svc(int nbufs, int bufsize, int max_req_size, int max_reply_size,
311 int req_portal, int rep_portal, int watchdog_factor,
312 svc_handler_t handler, char *name,
313 cfs_proc_dir_entry_t *proc_entry,
314 svcreq_printfn_t svcreq_printfn,
315 int min_threads, int max_threads,
316 char *threadname, __u32 ctx_tags)
319 struct ptlrpc_service *service;
323 LASSERT (bufsize >= max_req_size + SPTLRPC_MAX_PAYLOAD);
324 LASSERT (ctx_tags != 0);
326 OBD_ALLOC_PTR(service);
330 /* First initialise enough for early teardown */
332 service->srv_name = name;
333 spin_lock_init(&service->srv_lock);
334 CFS_INIT_LIST_HEAD(&service->srv_threads);
335 cfs_waitq_init(&service->srv_waitq);
337 service->srv_nbuf_per_group = test_req_buffer_pressure ? 1 : nbufs;
338 service->srv_max_req_size = max_req_size + SPTLRPC_MAX_PAYLOAD;
339 service->srv_buf_size = bufsize;
340 service->srv_rep_portal = rep_portal;
341 service->srv_req_portal = req_portal;
342 service->srv_watchdog_factor = watchdog_factor;
343 service->srv_handler = handler;
344 service->srv_request_history_print_fn = svcreq_printfn;
345 service->srv_request_seq = 1; /* valid seq #s start at 1 */
346 service->srv_request_max_cull_seq = 0;
347 service->srv_threads_min = min_threads;
348 service->srv_threads_max = max_threads;
349 service->srv_thread_name = threadname;
350 service->srv_ctx_tags = ctx_tags;
352 rc = LNetSetLazyPortal(service->srv_req_portal);
355 CFS_INIT_LIST_HEAD(&service->srv_request_queue);
356 CFS_INIT_LIST_HEAD(&service->srv_idle_rqbds);
357 CFS_INIT_LIST_HEAD(&service->srv_active_rqbds);
358 CFS_INIT_LIST_HEAD(&service->srv_history_rqbds);
359 CFS_INIT_LIST_HEAD(&service->srv_request_history);
360 CFS_INIT_LIST_HEAD(&service->srv_active_replies);
361 CFS_INIT_LIST_HEAD(&service->srv_reply_queue);
362 CFS_INIT_LIST_HEAD(&service->srv_free_rs_list);
363 cfs_waitq_init(&service->srv_free_rs_waitq);
365 spin_lock_init(&service->srv_at_lock);
366 CFS_INIT_LIST_HEAD(&service->srv_req_in_queue);
367 CFS_INIT_LIST_HEAD(&service->srv_at_list);
368 cfs_timer_init(&service->srv_at_timer, ptlrpc_at_timer, service);
369 /* At SOW, service time should be quick; 10s seems generous. If client
370 timeout is less than this, we'll be sending an early reply. */
371 at_init(&service->srv_at_estimate, 10, 0);
373 spin_lock (&ptlrpc_all_services_lock);
374 list_add (&service->srv_list, &ptlrpc_all_services);
375 spin_unlock (&ptlrpc_all_services_lock);
377 /* Now allocate the request buffers */
378 rc = ptlrpc_grow_req_bufs(service);
379 /* We shouldn't be under memory pressure at startup, so
380 * fail if we can't post all our buffers at this time. */
384 /* Now allocate pool of reply buffers */
385 /* Increase max reply size to next power of two */
386 service->srv_max_reply_size = 1;
387 while (service->srv_max_reply_size <
388 max_reply_size + SPTLRPC_MAX_PAYLOAD)
389 service->srv_max_reply_size <<= 1;
391 if (proc_entry != NULL)
392 ptlrpc_lprocfs_register_service(proc_entry, service);
394 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
395 service->srv_name, service->srv_req_portal);
399 ptlrpc_unregister_service(service);
404 * to actually free the request, must be called without holding svc_lock.
405 * note it's caller's responsibility to unlink req->rq_list.
407 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
409 LASSERT(atomic_read(&req->rq_refcount) == 0);
410 LASSERT(list_empty(&req->rq_timed_list));
412 /* DEBUG_REQ() assumes the reply state of a request with a valid
413 * ref will not be destroyed until that reference is dropped. */
414 ptlrpc_req_drop_rs(req);
416 sptlrpc_svc_ctx_decref(req);
418 if (req != &req->rq_rqbd->rqbd_req) {
419 /* NB request buffers use an embedded
420 * req if the incoming req unlinked the
421 * MD; this isn't one of them! */
422 OBD_FREE(req, sizeof(*req));
427 * drop a reference count of the request. if it reaches 0, we either
428 * put it into history list, or free it immediately.
430 static void ptlrpc_server_drop_request(struct ptlrpc_request *req)
432 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
433 struct ptlrpc_service *svc = rqbd->rqbd_service;
435 struct list_head *tmp;
436 struct list_head *nxt;
438 if (!atomic_dec_and_test(&req->rq_refcount))
441 spin_lock(&svc->srv_lock);
443 svc->srv_n_active_reqs--;
444 list_add(&req->rq_list, &rqbd->rqbd_reqs);
446 refcount = --(rqbd->rqbd_refcount);
448 /* request buffer is now idle: add to history */
449 list_del(&rqbd->rqbd_list);
450 list_add_tail(&rqbd->rqbd_list, &svc->srv_history_rqbds);
451 svc->srv_n_history_rqbds++;
453 /* cull some history?
454 * I expect only about 1 or 2 rqbds need to be recycled here */
455 while (svc->srv_n_history_rqbds > svc->srv_max_history_rqbds) {
456 rqbd = list_entry(svc->srv_history_rqbds.next,
457 struct ptlrpc_request_buffer_desc,
460 list_del(&rqbd->rqbd_list);
461 svc->srv_n_history_rqbds--;
463 /* remove rqbd's reqs from svc's req history while
464 * I've got the service lock */
465 list_for_each(tmp, &rqbd->rqbd_reqs) {
466 req = list_entry(tmp, struct ptlrpc_request,
468 /* Track the highest culled req seq */
469 if (req->rq_history_seq >
470 svc->srv_request_max_cull_seq)
471 svc->srv_request_max_cull_seq =
473 list_del(&req->rq_history_list);
476 spin_unlock(&svc->srv_lock);
478 list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
479 req = list_entry(rqbd->rqbd_reqs.next,
480 struct ptlrpc_request,
482 list_del(&req->rq_list);
483 ptlrpc_server_free_request(req);
486 spin_lock(&svc->srv_lock);
488 * now all reqs including the embedded req has been
489 * disposed, schedule request buffer for re-use.
491 LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) == 0);
492 list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
495 spin_unlock(&svc->srv_lock);
496 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
497 /* If we are low on memory, we are not interested in history */
498 list_del(&req->rq_list);
499 list_del_init(&req->rq_history_list);
500 spin_unlock(&svc->srv_lock);
502 ptlrpc_server_free_request(req);
504 spin_unlock(&svc->srv_lock);
509 * to finish a request: stop sending more early replies, and release
510 * the request. should be called after we finished handling the request.
512 static void ptlrpc_server_finish_request(struct ptlrpc_request *req)
514 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
516 if (req->rq_phase != RQ_PHASE_NEW) /* incorrect message magic */
517 DEBUG_REQ(D_INFO, req, "free req");
519 spin_lock(&svc->srv_at_lock);
520 req->rq_sent_final = 1;
521 list_del_init(&req->rq_timed_list);
522 spin_unlock(&svc->srv_at_lock);
524 ptlrpc_server_drop_request(req);
527 /* This function makes sure dead exports are evicted in a timely manner.
528 This function is only called when some export receives a message (i.e.,
529 the network is up.) */
530 static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
532 struct obd_export *oldest_exp;
539 /* Compensate for slow machines, etc, by faking our request time
540 into the future. Although this can break the strict time-ordering
541 of the list, we can be really lazy here - we don't have to evict
542 at the exact right moment. Eventually, all silent exports
543 will make it to the top of the list. */
544 exp->exp_last_request_time = max(exp->exp_last_request_time,
545 cfs_time_current_sec() + extra_delay);
547 CDEBUG(D_HA, "updating export %s at "CFS_TIME_T" exp %p\n",
548 exp->exp_client_uuid.uuid,
549 exp->exp_last_request_time, exp);
551 /* exports may get disconnected from the chain even though the
552 export has references, so we must keep the spin lock while
553 manipulating the lists */
554 spin_lock(&exp->exp_obd->obd_dev_lock);
556 if (list_empty(&exp->exp_obd_chain_timed)) {
557 /* this one is not timed */
558 spin_unlock(&exp->exp_obd->obd_dev_lock);
563 list_move_tail(&exp->exp_obd_chain_timed,
564 &exp->exp_obd->obd_exports_timed);
566 oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
567 struct obd_export, exp_obd_chain_timed);
568 oldest_time = oldest_exp->exp_last_request_time;
569 spin_unlock(&exp->exp_obd->obd_dev_lock);
571 if (exp->exp_obd->obd_recovering) {
572 /* be nice to everyone during recovery */
577 /* Note - racing to start/reset the obd_eviction timer is safe */
578 if (exp->exp_obd->obd_eviction_timer == 0) {
579 /* Check if the oldest entry is expired. */
580 if (cfs_time_current_sec() > (oldest_time + PING_EVICT_TIMEOUT +
582 /* We need a second timer, in case the net was down and
583 * it just came back. Since the pinger may skip every
584 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
585 * we better wait for 3. */
586 exp->exp_obd->obd_eviction_timer =
587 cfs_time_current_sec() + 3 * PING_INTERVAL;
588 CDEBUG(D_HA, "%s: Think about evicting %s from "CFS_TIME_T"\n",
589 exp->exp_obd->obd_name, obd_export_nid2str(exp),
593 if (cfs_time_current_sec() >
594 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
595 /* The evictor won't evict anyone who we've heard from
596 * recently, so we don't have to check before we start
598 if (!ping_evictor_wake(exp))
599 exp->exp_obd->obd_eviction_timer = 0;
606 static int ptlrpc_check_req(struct ptlrpc_request *req)
608 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
609 req->rq_export->exp_conn_cnt)) {
610 DEBUG_REQ(D_ERROR, req,
611 "DROPPING req from old connection %d < %d",
612 lustre_msg_get_conn_cnt(req->rq_reqmsg),
613 req->rq_export->exp_conn_cnt);
616 if (unlikely(req->rq_export->exp_obd &&
617 req->rq_export->exp_obd->obd_fail)) {
618 /* Failing over, don't handle any more reqs, send
619 error response instead. */
620 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
621 req, req->rq_export->exp_obd->obd_name);
622 req->rq_status = -ENODEV;
630 static void ptlrpc_at_set_timer(struct ptlrpc_service *svc)
632 struct ptlrpc_request *rq;
635 spin_lock(&svc->srv_at_lock);
636 if (list_empty(&svc->srv_at_list)) {
637 cfs_timer_disarm(&svc->srv_at_timer);
638 spin_unlock(&svc->srv_at_lock);
642 /* Set timer for closest deadline */
643 rq = list_entry(svc->srv_at_list.next, struct ptlrpc_request,
645 next = (__s32)(rq->rq_deadline - cfs_time_current_sec() -
648 ptlrpc_at_timer((unsigned long)svc);
650 cfs_timer_arm(&svc->srv_at_timer, cfs_time_shift(next));
651 spin_unlock(&svc->srv_at_lock);
652 CDEBUG(D_INFO, "armed %s at %+ds\n", svc->srv_name, next);
655 /* Add rpc to early reply check list */
656 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
658 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
659 struct ptlrpc_request *rq;
665 if (req->rq_no_reply)
668 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
671 spin_lock(&svc->srv_at_lock);
673 if (unlikely(req->rq_sent_final)) {
674 spin_unlock(&svc->srv_at_lock);
678 LASSERT(list_empty(&req->rq_timed_list));
679 /* Add to sorted list. Presumably latest rpcs will have the latest
680 deadlines, so search backward. */
681 list_for_each_entry_reverse(rq, &svc->srv_at_list, rq_timed_list) {
682 if (req->rq_deadline >= rq->rq_deadline) {
683 list_add(&req->rq_timed_list, &rq->rq_timed_list);
689 /* Add to front if shortest deadline or list empty */
690 list_add(&req->rq_timed_list, &svc->srv_at_list);
692 /* Check if we're the head of the list */
693 found = (svc->srv_at_list.next == &req->rq_timed_list);
695 spin_unlock(&svc->srv_at_lock);
698 ptlrpc_at_set_timer(svc);
703 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req,
706 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
707 struct ptlrpc_request *reqcopy;
708 struct lustre_msg *reqmsg;
709 cfs_duration_t olddl = req->rq_deadline - cfs_time_current_sec();
714 /* deadline is when the client expects us to reply, margin is the
715 difference between clients' and servers' expectations */
716 DEBUG_REQ(D_ADAPTTO, req,
717 "%ssending early reply (deadline %+lds, margin %+lds) for "
718 "%d+%d", AT_OFF ? "AT off - not " : "",
719 olddl, olddl - at_get(&svc->srv_at_estimate),
720 at_get(&svc->srv_at_estimate), extra_time);
726 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), "
727 "not sending early reply. Consider increasing "
728 "at_early_margin (%d)?", olddl, at_early_margin);
730 /* Return an error so we're not re-added to the timed list. */
734 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
735 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
736 "but no AT support");
740 if (req->rq_export && req->rq_export->exp_in_recovery) {
741 /* don't increase server estimates during recovery, and give
742 clients the full recovery time. */
743 newdl = cfs_time_current_sec() +
744 req->rq_export->exp_obd->obd_recovery_timeout;
747 /* Fake our processing time into the future to ask the
748 clients for some extra amount of time */
749 extra_time += cfs_time_current_sec() -
750 req->rq_arrival_time.tv_sec;
751 at_add(&svc->srv_at_estimate, extra_time);
753 newdl = req->rq_arrival_time.tv_sec +
754 at_get(&svc->srv_at_estimate);
756 if (req->rq_deadline >= newdl) {
757 /* We're not adding any time, no need to send an early reply
758 (e.g. maybe at adaptive_max) */
759 DEBUG_REQ(D_WARNING, req, "Couldn't add any time ("
760 CFS_DURATION_T"/"CFS_DURATION_T"), "
761 "not sending early reply\n", olddl,
762 cfs_time_sub(newdl, cfs_time_current_sec()));
766 OBD_ALLOC(reqcopy, sizeof *reqcopy);
769 OBD_ALLOC(reqmsg, req->rq_reqlen);
771 OBD_FREE(reqcopy, sizeof *reqcopy);
776 reqcopy->rq_reply_state = NULL;
777 reqcopy->rq_rep_swab_mask = 0;
778 reqcopy->rq_pack_bulk = 0;
779 reqcopy->rq_pack_udesc = 0;
780 reqcopy->rq_packed_final = 0;
781 sptlrpc_svc_ctx_addref(reqcopy);
782 /* We only need the reqmsg for the magic */
783 reqcopy->rq_reqmsg = reqmsg;
784 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
786 if (req->rq_sent_final) {
787 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
788 "abort sending early reply\n");
793 reqcopy->rq_export = class_conn2export(
794 lustre_msg_get_handle(reqcopy->rq_reqmsg));
795 if (reqcopy->rq_export == NULL)
796 GOTO(out, rc = -ENODEV);
799 class_export_rpc_get(reqcopy->rq_export);
800 if (reqcopy->rq_export->exp_obd &&
801 reqcopy->rq_export->exp_obd->obd_fail)
802 GOTO(out_put, rc = -ENODEV);
804 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
808 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
811 /* Adjust our own deadline to what we told the client */
812 req->rq_deadline = newdl;
813 req->rq_early_count++; /* number sent, server side */
815 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
818 /* Free the (early) reply state from lustre_pack_reply.
819 (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
820 ptlrpc_req_drop_rs(reqcopy);
823 class_export_rpc_put(reqcopy->rq_export);
824 class_export_put(reqcopy->rq_export);
826 sptlrpc_svc_ctx_decref(reqcopy);
827 OBD_FREE(reqmsg, req->rq_reqlen);
828 OBD_FREE(reqcopy, sizeof *reqcopy);
832 /* Send early replies to everybody expiring within at_early_margin
833 asking for at_extra time */
834 static int ptlrpc_at_check_timed(struct ptlrpc_service *svc)
836 struct ptlrpc_request *rq, *n;
837 struct list_head work_list;
838 time_t now = cfs_time_current_sec();
839 cfs_duration_t delay;
840 int first, counter = 0;
843 spin_lock(&svc->srv_at_lock);
844 if (svc->srv_at_check == 0) {
845 spin_unlock(&svc->srv_at_lock);
848 delay = cfs_time_sub(cfs_time_current(), svc->srv_at_checktime);
849 svc->srv_at_check = 0;
851 if (list_empty(&svc->srv_at_list)) {
852 spin_unlock(&svc->srv_at_lock);
856 /* The timer went off, but maybe the nearest rpc already completed. */
857 rq = list_entry(svc->srv_at_list.next, struct ptlrpc_request,
859 first = (int)(rq->rq_deadline - now);
860 if (first > at_early_margin) {
861 /* We've still got plenty of time. Reset the timer. */
862 spin_unlock(&svc->srv_at_lock);
863 ptlrpc_at_set_timer(svc);
867 /* We're close to a timeout, and we don't know how much longer the
868 server will take. Send early replies to everyone expiring soon. */
869 CFS_INIT_LIST_HEAD(&work_list);
870 list_for_each_entry_safe(rq, n, &svc->srv_at_list, rq_timed_list) {
871 if (rq->rq_deadline <= now + at_early_margin) {
872 list_move_tail(&rq->rq_timed_list, &work_list);
879 spin_unlock(&svc->srv_at_lock);
881 /* we have a new earliest deadline, restart the timer */
882 ptlrpc_at_set_timer(svc);
884 CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
885 "replies\n", first, at_extra, counter);
887 /* We're already past request deadlines before we even get a
888 chance to send early replies */
889 LCONSOLE_WARN("%s: This server is not able to keep up with "
890 "request traffic (cpu-bound).\n", svc->srv_name);
891 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, "
892 "delay="CFS_DURATION_T"(jiff)\n",
893 counter, svc->srv_n_queued_reqs, svc->srv_n_active_reqs,
894 at_get(&svc->srv_at_estimate), delay);
897 /* ptlrpc_server_finish_request may delete an entry out of
899 spin_lock(&svc->srv_at_lock);
900 while (!list_empty(&work_list)) {
901 rq = list_entry(work_list.next, struct ptlrpc_request,
903 list_del_init(&rq->rq_timed_list);
904 /* if the entry is still in the worklist, it hasn't been
905 deleted, and is safe to take a ref to keep the req around */
906 atomic_inc(&rq->rq_refcount);
907 spin_unlock(&svc->srv_at_lock);
909 if (ptlrpc_at_send_early_reply(rq, at_extra) == 0)
910 ptlrpc_at_add_timed(rq);
912 ptlrpc_server_drop_request(rq);
913 spin_lock(&svc->srv_at_lock);
915 spin_unlock(&svc->srv_at_lock);
920 /* Handle freshly incoming reqs, add to timed early reply list,
921 pass on to regular request queue */
923 ptlrpc_server_handle_req_in(struct ptlrpc_service *svc)
925 struct ptlrpc_request *req;
932 spin_lock(&svc->srv_lock);
933 if (list_empty(&svc->srv_req_in_queue)) {
934 spin_unlock(&svc->srv_lock);
938 req = list_entry(svc->srv_req_in_queue.next,
939 struct ptlrpc_request, rq_list);
940 list_del_init (&req->rq_list);
941 /* Consider this still a "queued" request as far as stats are
943 spin_unlock(&svc->srv_lock);
945 /* go through security check/transform */
946 rc = sptlrpc_svc_unwrap_request(req);
950 case SECSVC_COMPLETE:
951 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
959 /* Clear request swab mask; this is a new request */
960 req->rq_req_swab_mask = 0;
962 rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
964 CERROR("error unpacking request: ptl %d from %s x"LPU64"\n",
965 svc->srv_req_portal, libcfs_id2str(req->rq_peer),
970 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
972 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
973 LPU64"\n", svc->srv_req_portal,
974 libcfs_id2str(req->rq_peer), req->rq_xid);
979 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
980 CERROR("wrong packet type received (type=%u) from %s\n",
981 lustre_msg_get_type(req->rq_reqmsg),
982 libcfs_id2str(req->rq_peer));
986 CDEBUG(D_NET, "got req "LPD64"\n", req->rq_xid);
988 req->rq_export = class_conn2export(
989 lustre_msg_get_handle(req->rq_reqmsg));
990 if (req->rq_export) {
991 rc = ptlrpc_check_req(req);
993 rc = sptlrpc_target_export_check(req->rq_export, req);
995 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
996 "illegal security flavor,");
999 class_export_put(req->rq_export);
1000 req->rq_export = NULL;
1005 /* req_in handling should/must be fast */
1006 if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5)
1007 DEBUG_REQ(D_WARNING, req, "Slow req_in handling "CFS_DURATION_T"s",
1008 cfs_time_sub(cfs_time_current_sec(),
1009 req->rq_arrival_time.tv_sec));
1011 /* Set rpc server deadline and add it to the timed list */
1012 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1013 MSGHDR_AT_SUPPORT) ?
1014 /* The max time the client expects us to take */
1015 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1016 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1017 if (unlikely(deadline == 0)) {
1018 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1022 ptlrpc_at_add_timed(req);
1024 /* Move it over to the request processing queue */
1025 spin_lock(&svc->srv_lock);
1026 list_add_tail(&req->rq_list, &svc->srv_request_queue);
1027 cfs_waitq_signal(&svc->srv_waitq);
1028 spin_unlock(&svc->srv_lock);
1032 spin_lock(&svc->srv_lock);
1033 svc->srv_n_queued_reqs--;
1034 svc->srv_n_active_reqs++;
1035 spin_unlock(&svc->srv_lock);
1036 ptlrpc_server_finish_request(req);
1042 ptlrpc_server_handle_request(struct ptlrpc_service *svc,
1043 struct ptlrpc_thread *thread)
1045 struct obd_export *export = NULL;
1046 struct ptlrpc_request *request;
1047 struct timeval work_start;
1048 struct timeval work_end;
1055 spin_lock(&svc->srv_lock);
1056 if (unlikely(list_empty (&svc->srv_request_queue) ||
1059 /* !@%$# liblustre only has 1 thread */
1060 svc->srv_n_difficult_replies != 0 &&
1062 svc->srv_n_active_reqs >= (svc->srv_threads_running - 1)))) {
1063 /* Don't handle regular requests in the last thread, in order * re
1064 * to handle difficult replies (which might block other threads)
1065 * as well as handle any incoming reqs, early replies, etc.
1066 * That means we always need at least 2 service threads. */
1067 spin_unlock(&svc->srv_lock);
1071 request = list_entry (svc->srv_request_queue.next,
1072 struct ptlrpc_request, rq_list);
1073 list_del_init (&request->rq_list);
1074 svc->srv_n_queued_reqs--;
1075 svc->srv_n_active_reqs++;
1077 spin_unlock(&svc->srv_lock);
1079 if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
1080 libcfs_debug_dumplog();
1082 do_gettimeofday(&work_start);
1083 timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
1084 if (likely(svc->srv_stats != NULL)) {
1085 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
1087 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
1088 svc->srv_n_queued_reqs);
1089 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
1090 svc->srv_n_active_reqs);
1091 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
1092 at_get(&svc->srv_at_estimate));
1095 rc = lu_context_init(&request->rq_session, LCT_SESSION);
1097 CERROR("Failure to initialize session: %d\n", rc);
1100 request->rq_session.lc_thread = thread;
1101 lu_context_enter(&request->rq_session);
1103 CDEBUG(D_NET, "got req "LPU64"\n", request->rq_xid);
1105 request->rq_svc_thread = thread;
1107 request->rq_svc_thread->t_env->le_ses = &request->rq_session;
1109 request->rq_export = class_conn2export(
1110 lustre_msg_get_handle(request->rq_reqmsg));
1112 if (likely(request->rq_export)) {
1113 if (unlikely(ptlrpc_check_req(request)))
1115 ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
1116 export = class_export_rpc_get(request->rq_export);
1119 /* Discard requests queued for longer than the deadline.
1120 The deadline is increased if we send an early reply. */
1121 if (cfs_time_current_sec() > request->rq_deadline) {
1122 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s"
1123 ": deadline "CFS_DURATION_T":"CFS_DURATION_T"s ago\n",
1124 libcfs_id2str(request->rq_peer),
1125 cfs_time_sub(request->rq_deadline,
1126 request->rq_arrival_time.tv_sec),
1127 cfs_time_sub(cfs_time_current_sec(),
1128 request->rq_deadline));
1129 goto put_rpc_export;
1132 request->rq_phase = RQ_PHASE_INTERPRET;
1134 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
1135 "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1136 (request->rq_export ?
1137 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1138 (request->rq_export ?
1139 atomic_read(&request->rq_export->exp_refcount) : -99),
1140 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1141 libcfs_id2str(request->rq_peer),
1142 lustre_msg_get_opc(request->rq_reqmsg));
1144 OBD_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, obd_fail_val);
1146 rc = svc->srv_handler(request);
1148 request->rq_phase = RQ_PHASE_COMPLETE;
1150 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
1151 "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1152 (request->rq_export ?
1153 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1154 (request->rq_export ?
1155 atomic_read(&request->rq_export->exp_refcount) : -99),
1156 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1157 libcfs_id2str(request->rq_peer),
1158 lustre_msg_get_opc(request->rq_reqmsg));
1162 class_export_rpc_put(export);
1164 if (likely(request->rq_export != NULL))
1165 class_export_put(request->rq_export);
1167 lu_context_exit(&request->rq_session);
1168 lu_context_fini(&request->rq_session);
1170 if (unlikely(cfs_time_current_sec() > request->rq_deadline)) {
1171 DEBUG_REQ(D_WARNING, request, "Request x"LPU64" took longer "
1172 "than estimated ("CFS_DURATION_T":"CFS_DURATION_T"s);"
1173 " client may timeout.",
1174 request->rq_xid, cfs_time_sub(request->rq_deadline,
1175 request->rq_arrival_time.tv_sec),
1176 cfs_time_sub(cfs_time_current_sec(),
1177 request->rq_deadline));
1180 do_gettimeofday(&work_end);
1181 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1182 CDEBUG(D_RPCTRACE, "request x"LPU64" opc %u from %s processed in "
1183 "%ldus (%ldus total) trans "LPU64" rc %d/%d\n",
1184 request->rq_xid, lustre_msg_get_opc(request->rq_reqmsg),
1185 libcfs_id2str(request->rq_peer), timediff,
1186 cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL),
1187 request->rq_repmsg ? lustre_msg_get_transno(request->rq_repmsg) :
1188 request->rq_transno, request->rq_status,
1189 request->rq_repmsg ? lustre_msg_get_status(request->rq_repmsg):
1191 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
1192 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
1193 int opc = opcode_offset(op);
1194 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
1195 LASSERT(opc < LUSTRE_MAX_OPCODES);
1196 lprocfs_counter_add(svc->srv_stats,
1197 opc + EXTRA_MAX_OPCODES,
1201 if (unlikely(request->rq_early_count)) {
1202 DEBUG_REQ(D_ADAPTTO, request,
1203 "sent %d early replies before finishing in "
1205 request->rq_early_count,
1206 cfs_time_sub(work_end.tv_sec,
1207 request->rq_arrival_time.tv_sec));
1211 ptlrpc_server_finish_request(request);
1217 ptlrpc_server_handle_reply (struct ptlrpc_service *svc)
1219 struct ptlrpc_reply_state *rs;
1220 struct obd_export *exp;
1221 struct obd_device *obd;
1226 spin_lock(&svc->srv_lock);
1227 if (list_empty (&svc->srv_reply_queue)) {
1228 spin_unlock(&svc->srv_lock);
1232 rs = list_entry (svc->srv_reply_queue.next,
1233 struct ptlrpc_reply_state, rs_list);
1235 exp = rs->rs_export;
1238 LASSERT (rs->rs_difficult);
1239 LASSERT (rs->rs_scheduled);
1241 list_del_init (&rs->rs_list);
1243 /* Disengage from notifiers carefully (lock order - irqrestore below!)*/
1244 spin_unlock(&svc->srv_lock);
1246 spin_lock (&obd->obd_uncommitted_replies_lock);
1247 /* Noop if removed already */
1248 list_del_init (&rs->rs_obd_list);
1249 spin_unlock (&obd->obd_uncommitted_replies_lock);
1251 spin_lock (&exp->exp_lock);
1252 /* Noop if removed already */
1253 list_del_init (&rs->rs_exp_list);
1254 spin_unlock (&exp->exp_lock);
1256 spin_lock(&svc->srv_lock);
1258 been_handled = rs->rs_handled;
1261 nlocks = rs->rs_nlocks; /* atomic "steal", but */
1262 rs->rs_nlocks = 0; /* locks still on rs_locks! */
1264 if (nlocks == 0 && !been_handled) {
1265 /* If we see this, we should already have seen the warning
1266 * in mds_steal_ack_locks() */
1267 CWARN("All locks stolen from rs %p x"LPD64".t"LPD64
1270 rs->rs_xid, rs->rs_transno,
1271 lustre_msg_get_opc(rs->rs_msg),
1272 libcfs_nid2str(exp->exp_connection->c_peer.nid));
1275 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
1276 spin_unlock(&svc->srv_lock);
1278 if (!been_handled && rs->rs_on_net) {
1279 LNetMDUnlink(rs->rs_md_h);
1280 /* Ignore return code; we're racing with
1284 while (nlocks-- > 0)
1285 ldlm_lock_decref(&rs->rs_locks[nlocks],
1286 rs->rs_modes[nlocks]);
1288 spin_lock(&svc->srv_lock);
1291 rs->rs_scheduled = 0;
1293 if (!rs->rs_on_net) {
1295 svc->srv_n_difficult_replies--;
1296 spin_unlock(&svc->srv_lock);
1298 class_export_put (exp);
1299 rs->rs_export = NULL;
1300 ptlrpc_rs_decref (rs);
1301 atomic_dec (&svc->srv_outstanding_replies);
1305 /* still on the net; callback will schedule */
1306 spin_unlock(&svc->srv_lock);
1311 /* FIXME make use of timeout later */
1313 liblustre_check_services (void *arg)
1315 int did_something = 0;
1317 struct list_head *tmp, *nxt;
1320 /* I'm relying on being single threaded, not to have to lock
1321 * ptlrpc_all_services etc */
1322 list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
1323 struct ptlrpc_service *svc =
1324 list_entry (tmp, struct ptlrpc_service, srv_list);
1326 if (svc->srv_threads_running != 0) /* I've recursed */
1329 /* service threads can block for bulk, so this limits us
1330 * (arbitrarily) to recursing 1 stack frame per service.
1331 * Note that the problem with recursion is that we have to
1332 * unwind completely before our caller can resume. */
1334 svc->srv_threads_running++;
1337 rc = ptlrpc_server_handle_req_in(svc);
1338 rc |= ptlrpc_server_handle_reply(svc);
1339 rc |= ptlrpc_at_check_timed(svc);
1340 rc |= ptlrpc_server_handle_request(svc, NULL);
1341 rc |= (ptlrpc_server_post_idle_rqbds(svc) > 0);
1342 did_something |= rc;
1345 svc->srv_threads_running--;
1348 RETURN(did_something);
1350 #define ptlrpc_stop_all_threads(s) do {} while (0)
1352 #else /* __KERNEL__ */
1354 /* Don't use daemonize, it removes fs struct from new thread (bug 418) */
1355 void ptlrpc_daemonize(char *name)
1357 struct fs_struct *fs = current->fs;
1359 atomic_inc(&fs->count);
1360 cfs_daemonize(name);
1361 exit_fs(cfs_current());
1363 ll_set_fs_pwd(current->fs, init_task.fs->pwdmnt, init_task.fs->pwd);
1367 ptlrpc_check_rqbd_pool(struct ptlrpc_service *svc)
1369 int avail = svc->srv_nrqbd_receiving;
1370 int low_water = test_req_buffer_pressure ? 0 :
1371 svc->srv_nbuf_per_group/2;
1373 /* NB I'm not locking; just looking. */
1375 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
1376 * allowed the request history to grow out of control. We could put a
1377 * sanity check on that here and cull some history if we need the
1380 if (avail <= low_water)
1381 ptlrpc_grow_req_bufs(svc);
1384 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQBUF_AVAIL_CNTR,
1389 ptlrpc_retry_rqbds(void *arg)
1391 struct ptlrpc_service *svc = (struct ptlrpc_service *)arg;
1393 svc->srv_rqbd_timeout = 0;
1394 return (-ETIMEDOUT);
1397 static int ptlrpc_main(void *arg)
1399 struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
1400 struct ptlrpc_service *svc = data->svc;
1401 struct ptlrpc_thread *thread = data->thread;
1402 struct obd_device *dev = data->dev;
1403 struct ptlrpc_reply_state *rs;
1404 struct lc_watchdog *watchdog;
1405 #ifdef WITH_GROUP_INFO
1406 struct group_info *ginfo = NULL;
1409 int counter = 0, rc = 0;
1412 ptlrpc_daemonize(data->name);
1414 #if defined(HAVE_NODE_TO_CPUMASK) && defined(CONFIG_NUMA)
1415 /* we need to do this before any per-thread allocation is done so that
1416 * we get the per-thread allocations on local node. bug 7342 */
1417 if (svc->srv_cpu_affinity) {
1420 for (cpu = 0, num_cpu = 0; cpu < num_possible_cpus(); cpu++) {
1421 if (!cpu_online(cpu))
1423 if (num_cpu == thread->t_id % num_online_cpus())
1427 set_cpus_allowed(cfs_current(), node_to_cpumask(cpu_to_node(cpu)));
1431 #ifdef WITH_GROUP_INFO
1432 ginfo = groups_alloc(0);
1438 set_current_groups(ginfo);
1439 put_group_info(ginfo);
1442 if (svc->srv_init != NULL) {
1443 rc = svc->srv_init(thread);
1448 rc = lu_context_init(&env.le_ctx, svc->srv_ctx_tags);
1452 thread->t_env = &env;
1453 env.le_ctx.lc_thread = thread;
1455 /* Alloc reply state structure for this one */
1456 OBD_ALLOC_GFP(rs, svc->srv_max_reply_size, CFS_ALLOC_STD);
1462 /* Record that the thread is running */
1463 thread->t_flags = SVC_RUNNING;
1465 * wake up our creator. Note: @data is invalid after this point,
1466 * because it's allocated on ptlrpc_start_thread() stack.
1468 cfs_waitq_signal(&thread->t_ctl_waitq);
1470 watchdog = lc_watchdog_add(max_t(int, obd_timeout, AT_OFF ? 0 :
1471 at_get(&svc->srv_at_estimate)) *
1472 svc->srv_watchdog_factor, NULL, NULL);
1474 spin_lock(&svc->srv_lock);
1475 svc->srv_threads_running++;
1476 list_add(&rs->rs_list, &svc->srv_free_rs_list);
1477 spin_unlock(&svc->srv_lock);
1478 cfs_waitq_signal(&svc->srv_free_rs_waitq);
1480 CDEBUG(D_NET, "service thread %d (#%d)started\n", thread->t_id,
1481 svc->srv_threads_running);
1483 /* XXX maintain a list of all managed devices: insert here */
1485 while ((thread->t_flags & SVC_STOPPING) == 0 ||
1486 svc->srv_n_difficult_replies != 0) {
1487 /* Don't exit while there are replies to be handled */
1488 struct l_wait_info lwi = LWI_TIMEOUT(svc->srv_rqbd_timeout,
1489 ptlrpc_retry_rqbds, svc);
1491 lc_watchdog_disable(watchdog);
1495 l_wait_event_exclusive (svc->srv_waitq,
1496 ((thread->t_flags & SVC_STOPPING) != 0 &&
1497 svc->srv_n_difficult_replies == 0) ||
1498 (!list_empty(&svc->srv_idle_rqbds) &&
1499 svc->srv_rqbd_timeout == 0) ||
1500 !list_empty(&svc->srv_req_in_queue) ||
1501 !list_empty(&svc->srv_reply_queue) ||
1502 (!list_empty(&svc->srv_request_queue) &&
1503 (svc->srv_n_active_reqs <
1504 (svc->srv_threads_running - 1))) ||
1508 lc_watchdog_touch_ms(watchdog, max_t(int, obd_timeout,
1510 at_get(&svc->srv_at_estimate)) *
1511 svc->srv_watchdog_factor);
1513 ptlrpc_check_rqbd_pool(svc);
1515 if ((svc->srv_threads_started < svc->srv_threads_max) &&
1516 (svc->srv_n_active_reqs >= (svc->srv_threads_started - 1))){
1517 /* Ignore return code - we tried... */
1518 ptlrpc_start_thread(dev, svc);
1521 if (!list_empty(&svc->srv_reply_queue))
1522 ptlrpc_server_handle_reply(svc);
1524 if (!list_empty(&svc->srv_req_in_queue)) {
1525 /* Process all incoming reqs before handling any */
1526 ptlrpc_server_handle_req_in(svc);
1527 /* but limit ourselves in case of flood */
1528 if (counter++ < 1000)
1533 if (svc->srv_at_check)
1534 ptlrpc_at_check_timed(svc);
1536 /* don't handle requests in the last thread */
1537 if (!list_empty (&svc->srv_request_queue) &&
1538 (svc->srv_n_active_reqs < (svc->srv_threads_running - 1))) {
1539 lu_context_enter(&env.le_ctx);
1540 ptlrpc_server_handle_request(svc, thread);
1541 lu_context_exit(&env.le_ctx);
1544 if (!list_empty(&svc->srv_idle_rqbds) &&
1545 ptlrpc_server_post_idle_rqbds(svc) < 0) {
1546 /* I just failed to repost request buffers. Wait
1547 * for a timeout (unless something else happens)
1548 * before I try again */
1549 svc->srv_rqbd_timeout = cfs_time_seconds(1)/10;
1550 CDEBUG(D_RPCTRACE,"Posted buffers: %d\n",
1551 svc->srv_nrqbd_receiving);
1555 lc_watchdog_delete(watchdog);
1559 * deconstruct service specific state created by ptlrpc_start_thread()
1561 if (svc->srv_done != NULL)
1562 svc->srv_done(thread);
1564 lu_context_fini(&env.le_ctx);
1566 CDEBUG(D_NET, "service thread %d exiting: rc %d\n", thread->t_id, rc);
1568 spin_lock(&svc->srv_lock);
1569 svc->srv_threads_running--; /* must know immediately */
1571 thread->t_flags = SVC_STOPPED;
1573 cfs_waitq_signal(&thread->t_ctl_waitq);
1574 spin_unlock(&svc->srv_lock);
1579 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
1580 struct ptlrpc_thread *thread)
1582 struct l_wait_info lwi = { 0 };
1584 spin_lock(&svc->srv_lock);
1585 thread->t_flags = SVC_STOPPING;
1586 spin_unlock(&svc->srv_lock);
1588 cfs_waitq_broadcast(&svc->srv_waitq);
1589 l_wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED),
1592 spin_lock(&svc->srv_lock);
1593 list_del(&thread->t_link);
1594 spin_unlock(&svc->srv_lock);
1596 OBD_FREE_PTR(thread);
1599 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
1601 struct ptlrpc_thread *thread;
1603 spin_lock(&svc->srv_lock);
1604 while (!list_empty(&svc->srv_threads)) {
1605 thread = list_entry(svc->srv_threads.next,
1606 struct ptlrpc_thread, t_link);
1608 spin_unlock(&svc->srv_lock);
1609 ptlrpc_stop_thread(svc, thread);
1610 spin_lock(&svc->srv_lock);
1613 spin_unlock(&svc->srv_lock);
1616 int ptlrpc_start_threads(struct obd_device *dev, struct ptlrpc_service *svc)
1621 /* We require 2 threads min - see note in
1622 ptlrpc_server_handle_request */
1623 LASSERT(svc->srv_threads_min >= 2);
1624 for (i = 0; i < svc->srv_threads_min; i++) {
1625 rc = ptlrpc_start_thread(dev, svc);
1626 /* We have enough threads, don't start more. b=15759 */
1630 CERROR("cannot start %s thread #%d: rc %d\n",
1631 svc->srv_thread_name, i, rc);
1632 ptlrpc_stop_all_threads(svc);
1638 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc)
1640 struct l_wait_info lwi = { 0 };
1641 struct ptlrpc_svc_data d;
1642 struct ptlrpc_thread *thread;
1647 CDEBUG(D_RPCTRACE, "%s started %d min %d max %d running %d\n",
1648 svc->srv_name, svc->srv_threads_started, svc->srv_threads_min,
1649 svc->srv_threads_max, svc->srv_threads_running);
1650 if (unlikely(svc->srv_threads_started >= svc->srv_threads_max) ||
1651 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
1652 svc->srv_threads_started == svc->srv_threads_min - 1))
1655 OBD_ALLOC_PTR(thread);
1658 cfs_waitq_init(&thread->t_ctl_waitq);
1660 spin_lock(&svc->srv_lock);
1661 if (svc->srv_threads_started >= svc->srv_threads_max) {
1662 spin_unlock(&svc->srv_lock);
1663 OBD_FREE_PTR(thread);
1666 list_add(&thread->t_link, &svc->srv_threads);
1667 id = svc->srv_threads_started++;
1668 spin_unlock(&svc->srv_lock);
1671 sprintf(name, "%s_%02d", svc->srv_thread_name, id);
1677 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", name);
1679 /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
1680 * just drop the VM and FILES in ptlrpc_daemonize() right away.
1682 rc = cfs_kernel_thread(ptlrpc_main, &d, CLONE_VM | CLONE_FILES);
1684 CERROR("cannot start thread '%s': rc %d\n", name, rc);
1686 spin_lock(&svc->srv_lock);
1687 list_del(&thread->t_link);
1688 --svc->srv_threads_started;
1689 spin_unlock(&svc->srv_lock);
1691 OBD_FREE(thread, sizeof(*thread));
1694 l_wait_event(thread->t_ctl_waitq,
1695 thread->t_flags & (SVC_RUNNING | SVC_STOPPED), &lwi);
1697 rc = (thread->t_flags & SVC_STOPPED) ? thread->t_id : 0;
1702 int ptlrpc_unregister_service(struct ptlrpc_service *service)
1705 struct l_wait_info lwi;
1706 struct list_head *tmp;
1707 struct ptlrpc_reply_state *rs, *t;
1709 cfs_timer_disarm(&service->srv_at_timer);
1711 ptlrpc_stop_all_threads(service);
1712 LASSERT(list_empty(&service->srv_threads));
1714 spin_lock (&ptlrpc_all_services_lock);
1715 list_del_init (&service->srv_list);
1716 spin_unlock (&ptlrpc_all_services_lock);
1718 ptlrpc_lprocfs_unregister_service(service);
1720 /* All history will be culled when the next request buffer is
1722 service->srv_max_history_rqbds = 0;
1724 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
1726 rc = LNetClearLazyPortal(service->srv_req_portal);
1729 /* Unlink all the request buffers. This forces a 'final' event with
1730 * its 'unlink' flag set for each posted rqbd */
1731 list_for_each(tmp, &service->srv_active_rqbds) {
1732 struct ptlrpc_request_buffer_desc *rqbd =
1733 list_entry(tmp, struct ptlrpc_request_buffer_desc,
1736 rc = LNetMDUnlink(rqbd->rqbd_md_h);
1737 LASSERT (rc == 0 || rc == -ENOENT);
1740 /* Wait for the network to release any buffers it's currently
1743 spin_lock(&service->srv_lock);
1744 rc = service->srv_nrqbd_receiving;
1745 spin_unlock(&service->srv_lock);
1750 /* Network access will complete in finite time but the HUGE
1751 * timeout lets us CWARN for visibility of sluggish NALs */
1752 lwi = LWI_TIMEOUT(cfs_time_seconds(LONG_UNLINK), NULL, NULL);
1753 rc = l_wait_event(service->srv_waitq,
1754 service->srv_nrqbd_receiving == 0,
1756 if (rc == -ETIMEDOUT)
1757 CWARN("Service %s waiting for request buffers\n",
1761 /* schedule all outstanding replies to terminate them */
1762 spin_lock(&service->srv_lock);
1763 while (!list_empty(&service->srv_active_replies)) {
1764 struct ptlrpc_reply_state *rs =
1765 list_entry(service->srv_active_replies.next,
1766 struct ptlrpc_reply_state, rs_list);
1767 ptlrpc_schedule_difficult_reply(rs);
1769 spin_unlock(&service->srv_lock);
1771 /* purge the request queue. NB No new replies (rqbds all unlinked)
1772 * and no service threads, so I'm the only thread noodling the
1773 * request queue now */
1774 while (!list_empty(&service->srv_req_in_queue)) {
1775 struct ptlrpc_request *req =
1776 list_entry(service->srv_req_in_queue.next,
1777 struct ptlrpc_request,
1780 list_del(&req->rq_list);
1781 service->srv_n_queued_reqs--;
1782 service->srv_n_active_reqs++;
1783 ptlrpc_server_finish_request(req);
1785 while (!list_empty(&service->srv_request_queue)) {
1786 struct ptlrpc_request *req =
1787 list_entry(service->srv_request_queue.next,
1788 struct ptlrpc_request,
1791 list_del(&req->rq_list);
1792 service->srv_n_queued_reqs--;
1793 service->srv_n_active_reqs++;
1795 ptlrpc_server_finish_request(req);
1797 LASSERT(service->srv_n_queued_reqs == 0);
1798 LASSERT(service->srv_n_active_reqs == 0);
1799 LASSERT(service->srv_n_history_rqbds == 0);
1800 LASSERT(list_empty(&service->srv_active_rqbds));
1802 /* Now free all the request buffers since nothing references them
1804 while (!list_empty(&service->srv_idle_rqbds)) {
1805 struct ptlrpc_request_buffer_desc *rqbd =
1806 list_entry(service->srv_idle_rqbds.next,
1807 struct ptlrpc_request_buffer_desc,
1810 ptlrpc_free_rqbd(rqbd);
1813 /* wait for all outstanding replies to complete (they were
1814 * scheduled having been flagged to abort above) */
1815 while (atomic_read(&service->srv_outstanding_replies) != 0) {
1816 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10), NULL, NULL);
1818 rc = l_wait_event(service->srv_waitq,
1819 !list_empty(&service->srv_reply_queue), &lwi);
1820 LASSERT(rc == 0 || rc == -ETIMEDOUT);
1823 ptlrpc_server_handle_reply(service);
1826 CWARN("Unexpectedly long timeout %p\n", service);
1829 list_for_each_entry_safe(rs, t, &service->srv_free_rs_list, rs_list) {
1830 list_del(&rs->rs_list);
1831 OBD_FREE(rs, service->srv_max_reply_size);
1834 /* In case somebody rearmed this in the meantime */
1835 cfs_timer_disarm(&service->srv_at_timer);
1837 OBD_FREE_PTR(service);
1841 /* Returns 0 if the service is healthy.
1843 * Right now, it just checks to make sure that requests aren't languishing
1844 * in the queue. We'll use this health check to govern whether a node needs
1845 * to be shot, so it's intentionally non-aggressive. */
1846 int ptlrpc_service_health_check(struct ptlrpc_service *svc)
1848 struct ptlrpc_request *request;
1849 struct timeval right_now;
1855 do_gettimeofday(&right_now);
1857 spin_lock(&svc->srv_lock);
1858 if (list_empty(&svc->srv_request_queue)) {
1859 spin_unlock(&svc->srv_lock);
1863 /* How long has the next entry been waiting? */
1864 request = list_entry(svc->srv_request_queue.next,
1865 struct ptlrpc_request, rq_list);
1866 timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
1867 spin_unlock(&svc->srv_lock);
1869 if ((timediff / ONE_MILLION) > (AT_OFF ? obd_timeout * 3/2 :
1871 CERROR("%s: unhealthy - request has been waiting %lds\n",
1872 svc->srv_name, timediff / ONE_MILLION);