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, int no_ack)
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;
184 rs->rs_no_ack = !!no_ack;
189 #define HRT_RUNNING 0
190 #define HRT_STOPPING 1
192 struct ptlrpc_hr_thread {
194 unsigned long hrt_flags;
195 cfs_waitq_t hrt_wait;
196 struct list_head hrt_queue;
197 struct completion hrt_completion;
200 struct ptlrpc_hr_service {
204 struct ptlrpc_hr_thread hr_threads[0];
208 struct list_head rsb_replies;
209 struct ptlrpc_service *rsb_svc;
210 unsigned int rsb_n_replies;
214 * A pointer to per-node reply handling service.
216 static struct ptlrpc_hr_service *ptlrpc_hr = NULL;
219 * maximum mumber of replies scheduled in one batch
221 #define MAX_SCHEDULED 256
224 * Initialize a reply batch.
228 static void rs_batch_init(struct rs_batch *b)
230 memset(b, 0, sizeof *b);
231 CFS_INIT_LIST_HEAD(&b->rsb_replies);
235 * Dispatch all replies accumulated in the batch to one from
236 * dedicated reply handing threads.
240 static void rs_batch_dispatch(struct rs_batch *b)
242 if (b->rsb_n_replies != 0) {
243 struct ptlrpc_hr_service *hr = ptlrpc_hr;
246 idx = hr->hr_index++;
247 if (hr->hr_index >= hr->hr_n_threads)
250 spin_lock(&hr->hr_threads[idx].hrt_lock);
251 list_splice_init(&b->rsb_replies,
252 &hr->hr_threads[idx].hrt_queue);
253 spin_unlock(&hr->hr_threads[idx].hrt_lock);
254 cfs_waitq_signal(&hr->hr_threads[idx].hrt_wait);
255 b->rsb_n_replies = 0;
260 * Add a reply to a batch.
261 * Add one reply object to a batch, schedule batched replies if overload.
266 static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs)
268 struct ptlrpc_service *svc = rs->rs_service;
270 if (svc != b->rsb_svc || b->rsb_n_replies >= MAX_SCHEDULED) {
271 if (b->rsb_svc != NULL) {
272 rs_batch_dispatch(b);
273 spin_unlock(&b->rsb_svc->srv_lock);
275 spin_lock(&svc->srv_lock);
278 spin_lock(&rs->rs_lock);
279 rs->rs_scheduled_ever = 1;
280 if (rs->rs_scheduled == 0) {
281 list_move(&rs->rs_list, &b->rsb_replies);
282 rs->rs_scheduled = 1;
285 spin_unlock(&rs->rs_lock);
289 * Reply batch finalization.
290 * Dispatch remaining replies from the batch
291 * and release remaining spinlock.
295 static void rs_batch_fini(struct rs_batch *b)
297 if (b->rsb_svc != 0) {
298 rs_batch_dispatch(b);
299 spin_unlock(&b->rsb_svc->srv_lock);
303 #define DECLARE_RS_BATCH(b) struct rs_batch b
305 #else /* __KERNEL__ */
307 #define rs_batch_init(b) do{}while(0)
308 #define rs_batch_fini(b) do{}while(0)
309 #define rs_batch_add(b, r) ptlrpc_schedule_difficult_reply(r)
310 #define DECLARE_RS_BATCH(b)
312 #endif /* __KERNEL__ */
314 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
317 struct ptlrpc_hr_service *hr = ptlrpc_hr;
321 LASSERT(list_empty(&rs->rs_list));
323 idx = hr->hr_index++;
324 if (hr->hr_index >= hr->hr_n_threads)
326 spin_lock(&hr->hr_threads[idx].hrt_lock);
327 list_add_tail(&rs->rs_list, &hr->hr_threads[idx].hrt_queue);
328 spin_unlock(&hr->hr_threads[idx].hrt_lock);
329 cfs_waitq_signal(&hr->hr_threads[idx].hrt_wait);
332 list_add_tail(&rs->rs_list, &rs->rs_service->srv_reply_queue);
337 ptlrpc_schedule_difficult_reply (struct ptlrpc_reply_state *rs)
341 LASSERT_SPIN_LOCKED(&rs->rs_service->srv_lock);
342 LASSERT_SPIN_LOCKED(&rs->rs_lock);
343 LASSERT (rs->rs_difficult);
344 rs->rs_scheduled_ever = 1; /* flag any notification attempt */
346 if (rs->rs_scheduled) { /* being set up or already notified */
351 rs->rs_scheduled = 1;
352 list_del_init(&rs->rs_list);
353 ptlrpc_dispatch_difficult_reply(rs);
358 ptlrpc_commit_replies (struct obd_device *obd)
360 struct list_head *tmp;
361 struct list_head *nxt;
362 DECLARE_RS_BATCH(batch);
365 rs_batch_init(&batch);
366 /* Find any replies that have been committed and get their service
367 * to attend to complete them. */
369 /* CAVEAT EMPTOR: spinlock ordering!!! */
370 spin_lock(&obd->obd_uncommitted_replies_lock);
371 list_for_each_safe (tmp, nxt, &obd->obd_uncommitted_replies) {
372 struct ptlrpc_reply_state *rs =
373 list_entry(tmp, struct ptlrpc_reply_state, rs_obd_list);
375 LASSERT (rs->rs_difficult);
377 if (rs->rs_transno <= obd->obd_last_committed) {
378 list_del_init(&rs->rs_obd_list);
379 rs_batch_add(&batch, rs);
382 spin_unlock(&obd->obd_uncommitted_replies_lock);
383 rs_batch_fini(&batch);
388 ptlrpc_server_post_idle_rqbds (struct ptlrpc_service *svc)
390 struct ptlrpc_request_buffer_desc *rqbd;
395 spin_lock(&svc->srv_lock);
397 if (list_empty (&svc->srv_idle_rqbds)) {
398 spin_unlock(&svc->srv_lock);
402 rqbd = list_entry(svc->srv_idle_rqbds.next,
403 struct ptlrpc_request_buffer_desc,
405 list_del (&rqbd->rqbd_list);
407 /* assume we will post successfully */
408 svc->srv_nrqbd_receiving++;
409 list_add (&rqbd->rqbd_list, &svc->srv_active_rqbds);
411 spin_unlock(&svc->srv_lock);
413 rc = ptlrpc_register_rqbd(rqbd);
420 spin_lock(&svc->srv_lock);
422 svc->srv_nrqbd_receiving--;
423 list_del(&rqbd->rqbd_list);
424 list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
426 /* Don't complain if no request buffers are posted right now; LNET
427 * won't drop requests because we set the portal lazy! */
429 spin_unlock(&svc->srv_lock);
434 struct ptlrpc_service *ptlrpc_init_svc_conf(struct ptlrpc_service_conf *c,
435 svc_handler_t h, char *name,
436 struct proc_dir_entry *proc_entry,
437 svcreq_printfn_t prntfn,
440 return ptlrpc_init_svc(c->psc_nbufs, c->psc_bufsize,
441 c->psc_max_req_size, c->psc_max_reply_size,
442 c->psc_req_portal, c->psc_rep_portal,
443 c->psc_watchdog_factor,
445 prntfn, c->psc_min_threads, c->psc_max_threads,
446 threadname, c->psc_ctx_tags, NULL);
448 EXPORT_SYMBOL(ptlrpc_init_svc_conf);
450 static void ptlrpc_at_timer(unsigned long castmeharder)
452 struct ptlrpc_service *svc = (struct ptlrpc_service *)castmeharder;
453 svc->srv_at_check = 1;
454 svc->srv_at_checktime = cfs_time_current();
455 cfs_waitq_signal(&svc->srv_waitq);
458 /* @threadname should be 11 characters or less - 3 will be added on */
459 struct ptlrpc_service *
460 ptlrpc_init_svc(int nbufs, int bufsize, int max_req_size, int max_reply_size,
461 int req_portal, int rep_portal, int watchdog_factor,
462 svc_handler_t handler, char *name,
463 cfs_proc_dir_entry_t *proc_entry,
464 svcreq_printfn_t svcreq_printfn,
465 int min_threads, int max_threads,
466 char *threadname, __u32 ctx_tags,
467 svc_hpreq_handler_t hp_handler)
470 struct ptlrpc_service *service;
474 LASSERT (bufsize >= max_req_size + SPTLRPC_MAX_PAYLOAD);
475 LASSERT (ctx_tags != 0);
477 OBD_ALLOC_PTR(service);
481 /* First initialise enough for early teardown */
483 service->srv_name = name;
484 spin_lock_init(&service->srv_lock);
485 CFS_INIT_LIST_HEAD(&service->srv_threads);
486 cfs_waitq_init(&service->srv_waitq);
488 service->srv_nbuf_per_group = test_req_buffer_pressure ? 1 : nbufs;
489 service->srv_max_req_size = max_req_size + SPTLRPC_MAX_PAYLOAD;
490 service->srv_buf_size = bufsize;
491 service->srv_rep_portal = rep_portal;
492 service->srv_req_portal = req_portal;
493 service->srv_watchdog_factor = watchdog_factor;
494 service->srv_handler = handler;
495 service->srv_request_history_print_fn = svcreq_printfn;
496 service->srv_request_seq = 1; /* valid seq #s start at 1 */
497 service->srv_request_max_cull_seq = 0;
498 service->srv_threads_min = min_threads;
499 service->srv_threads_max = max_threads;
500 service->srv_thread_name = threadname;
501 service->srv_ctx_tags = ctx_tags;
502 service->srv_hpreq_handler = hp_handler;
503 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
504 service->srv_hpreq_count = 0;
505 service->srv_n_hpreq = 0;
507 rc = LNetSetLazyPortal(service->srv_req_portal);
510 CFS_INIT_LIST_HEAD(&service->srv_request_queue);
511 CFS_INIT_LIST_HEAD(&service->srv_request_hpq);
512 CFS_INIT_LIST_HEAD(&service->srv_idle_rqbds);
513 CFS_INIT_LIST_HEAD(&service->srv_active_rqbds);
514 CFS_INIT_LIST_HEAD(&service->srv_history_rqbds);
515 CFS_INIT_LIST_HEAD(&service->srv_request_history);
516 CFS_INIT_LIST_HEAD(&service->srv_active_replies);
518 CFS_INIT_LIST_HEAD(&service->srv_reply_queue);
520 CFS_INIT_LIST_HEAD(&service->srv_free_rs_list);
521 cfs_waitq_init(&service->srv_free_rs_waitq);
522 atomic_set(&service->srv_n_difficult_replies, 0);
524 spin_lock_init(&service->srv_at_lock);
525 CFS_INIT_LIST_HEAD(&service->srv_req_in_queue);
526 CFS_INIT_LIST_HEAD(&service->srv_at_list);
527 cfs_timer_init(&service->srv_at_timer, ptlrpc_at_timer, service);
528 /* At SOW, service time should be quick; 10s seems generous. If client
529 timeout is less than this, we'll be sending an early reply. */
530 at_init(&service->srv_at_estimate, 10, 0);
532 spin_lock (&ptlrpc_all_services_lock);
533 list_add (&service->srv_list, &ptlrpc_all_services);
534 spin_unlock (&ptlrpc_all_services_lock);
536 /* Now allocate the request buffers */
537 rc = ptlrpc_grow_req_bufs(service);
538 /* We shouldn't be under memory pressure at startup, so
539 * fail if we can't post all our buffers at this time. */
543 /* Now allocate pool of reply buffers */
544 /* Increase max reply size to next power of two */
545 service->srv_max_reply_size = 1;
546 while (service->srv_max_reply_size <
547 max_reply_size + SPTLRPC_MAX_PAYLOAD)
548 service->srv_max_reply_size <<= 1;
550 if (proc_entry != NULL)
551 ptlrpc_lprocfs_register_service(proc_entry, service);
553 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
554 service->srv_name, service->srv_req_portal);
558 ptlrpc_unregister_service(service);
563 * to actually free the request, must be called without holding svc_lock.
564 * note it's caller's responsibility to unlink req->rq_list.
566 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
568 LASSERT(atomic_read(&req->rq_refcount) == 0);
569 LASSERT(list_empty(&req->rq_timed_list));
571 /* DEBUG_REQ() assumes the reply state of a request with a valid
572 * ref will not be destroyed until that reference is dropped. */
573 ptlrpc_req_drop_rs(req);
575 sptlrpc_svc_ctx_decref(req);
577 if (req != &req->rq_rqbd->rqbd_req) {
578 /* NB request buffers use an embedded
579 * req if the incoming req unlinked the
580 * MD; this isn't one of them! */
581 OBD_FREE(req, sizeof(*req));
586 * drop a reference count of the request. if it reaches 0, we either
587 * put it into history list, or free it immediately.
589 static void ptlrpc_server_drop_request(struct ptlrpc_request *req)
591 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
592 struct ptlrpc_service *svc = rqbd->rqbd_service;
594 struct list_head *tmp;
595 struct list_head *nxt;
597 if (!atomic_dec_and_test(&req->rq_refcount))
600 spin_lock(&svc->srv_lock);
602 svc->srv_n_active_reqs--;
603 list_add(&req->rq_list, &rqbd->rqbd_reqs);
605 refcount = --(rqbd->rqbd_refcount);
607 /* request buffer is now idle: add to history */
608 list_del(&rqbd->rqbd_list);
609 list_add_tail(&rqbd->rqbd_list, &svc->srv_history_rqbds);
610 svc->srv_n_history_rqbds++;
612 /* cull some history?
613 * I expect only about 1 or 2 rqbds need to be recycled here */
614 while (svc->srv_n_history_rqbds > svc->srv_max_history_rqbds) {
615 rqbd = list_entry(svc->srv_history_rqbds.next,
616 struct ptlrpc_request_buffer_desc,
619 list_del(&rqbd->rqbd_list);
620 svc->srv_n_history_rqbds--;
622 /* remove rqbd's reqs from svc's req history while
623 * I've got the service lock */
624 list_for_each(tmp, &rqbd->rqbd_reqs) {
625 req = list_entry(tmp, struct ptlrpc_request,
627 /* Track the highest culled req seq */
628 if (req->rq_history_seq >
629 svc->srv_request_max_cull_seq)
630 svc->srv_request_max_cull_seq =
632 list_del(&req->rq_history_list);
635 spin_unlock(&svc->srv_lock);
637 list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
638 req = list_entry(rqbd->rqbd_reqs.next,
639 struct ptlrpc_request,
641 list_del(&req->rq_list);
642 ptlrpc_server_free_request(req);
645 spin_lock(&svc->srv_lock);
647 * now all reqs including the embedded req has been
648 * disposed, schedule request buffer for re-use.
650 LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) == 0);
651 list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
654 spin_unlock(&svc->srv_lock);
655 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
656 /* If we are low on memory, we are not interested in history */
657 list_del(&req->rq_list);
658 list_del_init(&req->rq_history_list);
659 spin_unlock(&svc->srv_lock);
661 ptlrpc_server_free_request(req);
663 spin_unlock(&svc->srv_lock);
668 * to finish a request: stop sending more early replies, and release
669 * the request. should be called after we finished handling the request.
671 static void ptlrpc_server_finish_request(struct ptlrpc_request *req)
673 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
675 if (req->rq_export) {
676 class_export_put(req->rq_export);
677 req->rq_export = NULL;
680 if (req->rq_phase != RQ_PHASE_NEW) /* incorrect message magic */
681 DEBUG_REQ(D_INFO, req, "free req");
683 spin_lock(&svc->srv_at_lock);
684 req->rq_sent_final = 1;
685 list_del_init(&req->rq_timed_list);
686 spin_unlock(&svc->srv_at_lock);
688 ptlrpc_server_drop_request(req);
691 /* This function makes sure dead exports are evicted in a timely manner.
692 This function is only called when some export receives a message (i.e.,
693 the network is up.) */
694 static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
696 struct obd_export *oldest_exp;
697 time_t oldest_time, new_time;
703 /* Compensate for slow machines, etc, by faking our request time
704 into the future. Although this can break the strict time-ordering
705 of the list, we can be really lazy here - we don't have to evict
706 at the exact right moment. Eventually, all silent exports
707 will make it to the top of the list. */
709 /* Do not pay attention on 1sec or smaller renewals. */
710 new_time = cfs_time_current_sec() + extra_delay;
711 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
714 exp->exp_last_request_time = new_time;
715 CDEBUG(D_HA, "updating export %s at "CFS_TIME_T" exp %p\n",
716 exp->exp_client_uuid.uuid,
717 exp->exp_last_request_time, exp);
719 /* exports may get disconnected from the chain even though the
720 export has references, so we must keep the spin lock while
721 manipulating the lists */
722 spin_lock(&exp->exp_obd->obd_dev_lock);
724 if (list_empty(&exp->exp_obd_chain_timed)) {
725 /* this one is not timed */
726 spin_unlock(&exp->exp_obd->obd_dev_lock);
730 list_move_tail(&exp->exp_obd_chain_timed,
731 &exp->exp_obd->obd_exports_timed);
733 oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
734 struct obd_export, exp_obd_chain_timed);
735 oldest_time = oldest_exp->exp_last_request_time;
736 spin_unlock(&exp->exp_obd->obd_dev_lock);
738 if (exp->exp_obd->obd_recovering) {
739 /* be nice to everyone during recovery */
744 /* Note - racing to start/reset the obd_eviction timer is safe */
745 if (exp->exp_obd->obd_eviction_timer == 0) {
746 /* Check if the oldest entry is expired. */
747 if (cfs_time_current_sec() > (oldest_time + PING_EVICT_TIMEOUT +
749 /* We need a second timer, in case the net was down and
750 * it just came back. Since the pinger may skip every
751 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
752 * we better wait for 3. */
753 exp->exp_obd->obd_eviction_timer =
754 cfs_time_current_sec() + 3 * PING_INTERVAL;
755 CDEBUG(D_HA, "%s: Think about evicting %s from "CFS_TIME_T"\n",
756 exp->exp_obd->obd_name, obd_export_nid2str(exp),
760 if (cfs_time_current_sec() >
761 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
762 /* The evictor won't evict anyone who we've heard from
763 * recently, so we don't have to check before we start
765 if (!ping_evictor_wake(exp))
766 exp->exp_obd->obd_eviction_timer = 0;
773 static int ptlrpc_check_req(struct ptlrpc_request *req)
775 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
776 req->rq_export->exp_conn_cnt)) {
777 DEBUG_REQ(D_ERROR, req,
778 "DROPPING req from old connection %d < %d",
779 lustre_msg_get_conn_cnt(req->rq_reqmsg),
780 req->rq_export->exp_conn_cnt);
783 if (unlikely(req->rq_export->exp_obd &&
784 req->rq_export->exp_obd->obd_fail)) {
785 /* Failing over, don't handle any more reqs, send
786 error response instead. */
787 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
788 req, req->rq_export->exp_obd->obd_name);
789 req->rq_status = -ENODEV;
797 static void ptlrpc_at_set_timer(struct ptlrpc_service *svc)
799 struct ptlrpc_request *rq;
802 spin_lock(&svc->srv_at_lock);
803 if (list_empty(&svc->srv_at_list)) {
804 cfs_timer_disarm(&svc->srv_at_timer);
805 spin_unlock(&svc->srv_at_lock);
809 /* Set timer for closest deadline */
810 rq = list_entry(svc->srv_at_list.next, struct ptlrpc_request,
812 next = (__s32)(rq->rq_deadline - cfs_time_current_sec() -
815 ptlrpc_at_timer((unsigned long)svc);
817 cfs_timer_arm(&svc->srv_at_timer, cfs_time_shift(next));
818 spin_unlock(&svc->srv_at_lock);
819 CDEBUG(D_INFO, "armed %s at %+ds\n", svc->srv_name, next);
822 /* Add rpc to early reply check list */
823 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
825 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
826 struct ptlrpc_request *rq;
832 if (req->rq_no_reply)
835 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
838 spin_lock(&svc->srv_at_lock);
840 if (unlikely(req->rq_sent_final)) {
841 spin_unlock(&svc->srv_at_lock);
845 LASSERT(list_empty(&req->rq_timed_list));
846 /* Add to sorted list. Presumably latest rpcs will have the latest
847 deadlines, so search backward. */
848 list_for_each_entry_reverse(rq, &svc->srv_at_list, rq_timed_list) {
849 if (req->rq_deadline >= rq->rq_deadline) {
850 list_add(&req->rq_timed_list, &rq->rq_timed_list);
856 /* Add to front if shortest deadline or list empty */
857 list_add(&req->rq_timed_list, &svc->srv_at_list);
859 /* Check if we're the head of the list */
860 found = (svc->srv_at_list.next == &req->rq_timed_list);
862 spin_unlock(&svc->srv_at_lock);
865 ptlrpc_at_set_timer(svc);
870 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req,
873 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
874 struct ptlrpc_request *reqcopy;
875 struct lustre_msg *reqmsg;
876 cfs_duration_t olddl = req->rq_deadline - cfs_time_current_sec();
881 /* deadline is when the client expects us to reply, margin is the
882 difference between clients' and servers' expectations */
883 DEBUG_REQ(D_ADAPTTO, req,
884 "%ssending early reply (deadline %+lds, margin %+lds) for "
885 "%d+%d", AT_OFF ? "AT off - not " : "",
886 olddl, olddl - at_get(&svc->srv_at_estimate),
887 at_get(&svc->srv_at_estimate), extra_time);
893 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), "
894 "not sending early reply. Consider increasing "
895 "at_early_margin (%d)?", olddl, at_early_margin);
897 /* Return an error so we're not re-added to the timed list. */
901 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
902 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
903 "but no AT support");
907 if (req->rq_export && req->rq_export->exp_in_recovery) {
908 /* don't increase server estimates during recovery, and give
909 clients the full recovery time. */
910 newdl = cfs_time_current_sec() +
911 req->rq_export->exp_obd->obd_recovery_timeout;
914 /* Fake our processing time into the future to ask the
915 clients for some extra amount of time */
916 extra_time += cfs_time_current_sec() -
917 req->rq_arrival_time.tv_sec;
918 at_add(&svc->srv_at_estimate, extra_time);
920 newdl = req->rq_arrival_time.tv_sec +
921 at_get(&svc->srv_at_estimate);
923 if (req->rq_deadline >= newdl) {
924 /* We're not adding any time, no need to send an early reply
925 (e.g. maybe at adaptive_max) */
926 DEBUG_REQ(D_WARNING, req, "Couldn't add any time ("
927 CFS_DURATION_T"/"CFS_DURATION_T"), "
928 "not sending early reply\n", olddl,
929 cfs_time_sub(newdl, cfs_time_current_sec()));
933 OBD_ALLOC(reqcopy, sizeof *reqcopy);
936 OBD_ALLOC(reqmsg, req->rq_reqlen);
938 OBD_FREE(reqcopy, sizeof *reqcopy);
943 reqcopy->rq_reply_state = NULL;
944 reqcopy->rq_rep_swab_mask = 0;
945 reqcopy->rq_pack_bulk = 0;
946 reqcopy->rq_pack_udesc = 0;
947 reqcopy->rq_packed_final = 0;
948 sptlrpc_svc_ctx_addref(reqcopy);
949 /* We only need the reqmsg for the magic */
950 reqcopy->rq_reqmsg = reqmsg;
951 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
953 if (req->rq_sent_final) {
954 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
955 "abort sending early reply\n");
960 reqcopy->rq_export = class_conn2export(
961 lustre_msg_get_handle(reqcopy->rq_reqmsg));
962 if (reqcopy->rq_export == NULL)
963 GOTO(out, rc = -ENODEV);
966 class_export_rpc_get(reqcopy->rq_export);
967 if (reqcopy->rq_export->exp_obd &&
968 reqcopy->rq_export->exp_obd->obd_fail)
969 GOTO(out_put, rc = -ENODEV);
971 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
975 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
978 /* Adjust our own deadline to what we told the client */
979 req->rq_deadline = newdl;
980 req->rq_early_count++; /* number sent, server side */
982 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
985 /* Free the (early) reply state from lustre_pack_reply.
986 (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
987 ptlrpc_req_drop_rs(reqcopy);
990 class_export_rpc_put(reqcopy->rq_export);
991 class_export_put(reqcopy->rq_export);
993 sptlrpc_svc_ctx_decref(reqcopy);
994 OBD_FREE(reqmsg, req->rq_reqlen);
995 OBD_FREE(reqcopy, sizeof *reqcopy);
999 /* Send early replies to everybody expiring within at_early_margin
1000 asking for at_extra time */
1001 static int ptlrpc_at_check_timed(struct ptlrpc_service *svc)
1003 struct ptlrpc_request *rq, *n;
1004 struct list_head work_list;
1005 time_t now = cfs_time_current_sec();
1006 cfs_duration_t delay;
1007 int first, counter = 0;
1010 spin_lock(&svc->srv_at_lock);
1011 if (svc->srv_at_check == 0) {
1012 spin_unlock(&svc->srv_at_lock);
1015 delay = cfs_time_sub(cfs_time_current(), svc->srv_at_checktime);
1016 svc->srv_at_check = 0;
1018 if (list_empty(&svc->srv_at_list)) {
1019 spin_unlock(&svc->srv_at_lock);
1023 /* The timer went off, but maybe the nearest rpc already completed. */
1024 rq = list_entry(svc->srv_at_list.next, struct ptlrpc_request,
1026 first = (int)(rq->rq_deadline - now);
1027 if (first > at_early_margin) {
1028 /* We've still got plenty of time. Reset the timer. */
1029 spin_unlock(&svc->srv_at_lock);
1030 ptlrpc_at_set_timer(svc);
1034 /* We're close to a timeout, and we don't know how much longer the
1035 server will take. Send early replies to everyone expiring soon. */
1036 CFS_INIT_LIST_HEAD(&work_list);
1037 list_for_each_entry_safe(rq, n, &svc->srv_at_list, rq_timed_list) {
1038 if (rq->rq_deadline <= now + at_early_margin) {
1039 list_move_tail(&rq->rq_timed_list, &work_list);
1046 spin_unlock(&svc->srv_at_lock);
1048 /* we have a new earliest deadline, restart the timer */
1049 ptlrpc_at_set_timer(svc);
1051 CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
1052 "replies\n", first, at_extra, counter);
1054 /* We're already past request deadlines before we even get a
1055 chance to send early replies */
1056 LCONSOLE_WARN("%s: This server is not able to keep up with "
1057 "request traffic (cpu-bound).\n", svc->srv_name);
1058 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, "
1059 "delay="CFS_DURATION_T"(jiff)\n",
1060 counter, svc->srv_n_queued_reqs, svc->srv_n_active_reqs,
1061 at_get(&svc->srv_at_estimate), delay);
1064 /* ptlrpc_server_finish_request may delete an entry out of
1066 spin_lock(&svc->srv_at_lock);
1067 while (!list_empty(&work_list)) {
1068 rq = list_entry(work_list.next, struct ptlrpc_request,
1070 list_del_init(&rq->rq_timed_list);
1071 /* if the entry is still in the worklist, it hasn't been
1072 deleted, and is safe to take a ref to keep the req around */
1073 atomic_inc(&rq->rq_refcount);
1074 spin_unlock(&svc->srv_at_lock);
1076 if (ptlrpc_at_send_early_reply(rq, at_extra) == 0)
1077 ptlrpc_at_add_timed(rq);
1079 ptlrpc_server_drop_request(rq);
1080 spin_lock(&svc->srv_at_lock);
1082 spin_unlock(&svc->srv_at_lock);
1088 * Put the request to the export list if the request may become
1089 * a high priority one.
1091 static int ptlrpc_hpreq_init(struct ptlrpc_service *svc,
1092 struct ptlrpc_request *req)
1097 if (svc->srv_hpreq_handler) {
1098 rc = svc->srv_hpreq_handler(req);
1102 if (req->rq_export && req->rq_ops) {
1103 spin_lock(&req->rq_export->exp_lock);
1104 list_add(&req->rq_exp_list, &req->rq_export->exp_queued_rpc);
1105 spin_unlock(&req->rq_export->exp_lock);
1111 /** Remove the request from the export list. */
1112 static void ptlrpc_hpreq_fini(struct ptlrpc_request *req)
1115 if (req->rq_export && req->rq_ops) {
1116 spin_lock(&req->rq_export->exp_lock);
1117 list_del_init(&req->rq_exp_list);
1118 spin_unlock(&req->rq_export->exp_lock);
1124 * Make the request a high priority one.
1126 * All the high priority requests are queued in a separate FIFO
1127 * ptlrpc_service::srv_request_hpq list which is parallel to
1128 * ptlrpc_service::srv_request_queue list but has a higher priority
1131 * \see ptlrpc_server_handle_request().
1133 static void ptlrpc_hpreq_reorder_nolock(struct ptlrpc_service *svc,
1134 struct ptlrpc_request *req)
1137 LASSERT(svc != NULL);
1138 spin_lock(&req->rq_lock);
1139 if (req->rq_hp == 0) {
1140 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1142 /* Add to the high priority queue. */
1143 list_move_tail(&req->rq_list, &svc->srv_request_hpq);
1145 if (opc != OBD_PING)
1146 DEBUG_REQ(D_NET, req, "high priority req");
1148 spin_unlock(&req->rq_lock);
1152 void ptlrpc_hpreq_reorder(struct ptlrpc_request *req)
1154 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
1157 spin_lock(&svc->srv_lock);
1158 /* It may happen that the request is already taken for the processing
1159 * but still in the export list, do not re-add it into the HP list. */
1160 if (req->rq_phase == RQ_PHASE_NEW)
1161 ptlrpc_hpreq_reorder_nolock(svc, req);
1162 spin_unlock(&svc->srv_lock);
1166 /** Check if the request if a high priority one. */
1167 static int ptlrpc_server_hpreq_check(struct ptlrpc_request *req)
1172 /* Check by request opc. */
1173 opc = lustre_msg_get_opc(req->rq_reqmsg);
1174 if (opc == OBD_PING)
1177 /* Perform request specific check. */
1178 if (req->rq_ops && req->rq_ops->hpreq_check)
1179 rc = req->rq_ops->hpreq_check(req);
1183 /** Check if a request is a high priority one. */
1184 static int ptlrpc_server_request_add(struct ptlrpc_service *svc,
1185 struct ptlrpc_request *req)
1190 rc = ptlrpc_server_hpreq_check(req);
1194 spin_lock(&svc->srv_lock);
1195 /* Before inserting the request into the queue, check if it is not
1196 * inserted yet, or even already handled -- it may happen due to
1197 * a racing ldlm_server_blocking_ast(). */
1198 if (req->rq_phase == RQ_PHASE_NEW && list_empty(&req->rq_list)) {
1200 ptlrpc_hpreq_reorder_nolock(svc, req);
1202 list_add_tail(&req->rq_list, &svc->srv_request_queue);
1204 spin_unlock(&svc->srv_lock);
1209 /* Only allow normal priority requests on a service that has a high-priority
1210 * queue if forced (i.e. cleanup), if there are other high priority requests
1211 * already being processed (i.e. those threads can service more high-priority
1212 * requests), or if there are enough idle threads that a later thread can do
1213 * a high priority request. */
1214 static int ptlrpc_server_allow_normal(struct ptlrpc_service *svc, int force)
1216 return force || !svc->srv_hpreq_handler || svc->srv_n_hpreq > 0 ||
1217 svc->srv_n_active_reqs < svc->srv_threads_running - 2;
1220 static struct ptlrpc_request *
1221 ptlrpc_server_request_get(struct ptlrpc_service *svc, int force)
1223 struct ptlrpc_request *req = NULL;
1226 if (ptlrpc_server_allow_normal(svc, force) &&
1227 !list_empty(&svc->srv_request_queue) &&
1228 (list_empty(&svc->srv_request_hpq) ||
1229 svc->srv_hpreq_count >= svc->srv_hpreq_ratio)) {
1230 req = list_entry(svc->srv_request_queue.next,
1231 struct ptlrpc_request, rq_list);
1232 svc->srv_hpreq_count = 0;
1233 } else if (!list_empty(&svc->srv_request_hpq)) {
1234 req = list_entry(svc->srv_request_hpq.next,
1235 struct ptlrpc_request, rq_list);
1236 svc->srv_hpreq_count++;
1241 static int ptlrpc_server_request_pending(struct ptlrpc_service *svc, int force)
1243 return ((ptlrpc_server_allow_normal(svc, force) &&
1244 !list_empty(&svc->srv_request_queue)) ||
1245 !list_empty(&svc->srv_request_hpq));
1248 /* Handle freshly incoming reqs, add to timed early reply list,
1249 pass on to regular request queue */
1251 ptlrpc_server_handle_req_in(struct ptlrpc_service *svc)
1253 struct ptlrpc_request *req;
1260 spin_lock(&svc->srv_lock);
1261 if (list_empty(&svc->srv_req_in_queue)) {
1262 spin_unlock(&svc->srv_lock);
1266 req = list_entry(svc->srv_req_in_queue.next,
1267 struct ptlrpc_request, rq_list);
1268 list_del_init (&req->rq_list);
1269 /* Consider this still a "queued" request as far as stats are
1271 spin_unlock(&svc->srv_lock);
1273 /* go through security check/transform */
1274 rc = sptlrpc_svc_unwrap_request(req);
1278 case SECSVC_COMPLETE:
1279 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1287 /* Clear request swab mask; this is a new request */
1288 req->rq_req_swab_mask = 0;
1290 rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
1292 CERROR("error unpacking request: ptl %d from %s x"LPU64"\n",
1293 svc->srv_req_portal, libcfs_id2str(req->rq_peer),
1298 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1300 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
1301 LPU64"\n", svc->srv_req_portal,
1302 libcfs_id2str(req->rq_peer), req->rq_xid);
1307 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1308 CERROR("wrong packet type received (type=%u) from %s\n",
1309 lustre_msg_get_type(req->rq_reqmsg),
1310 libcfs_id2str(req->rq_peer));
1314 switch(lustre_msg_get_opc(req->rq_reqmsg)) {
1317 req->rq_bulk_write = 1;
1321 req->rq_bulk_read = 1;
1325 CDEBUG(D_NET, "got req "LPD64"\n", req->rq_xid);
1327 req->rq_export = class_conn2export(
1328 lustre_msg_get_handle(req->rq_reqmsg));
1329 if (req->rq_export) {
1330 rc = ptlrpc_check_req(req);
1332 rc = sptlrpc_target_export_check(req->rq_export, req);
1334 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
1335 "illegal security flavor,");
1340 ptlrpc_update_export_timer(req->rq_export, 0);
1343 /* req_in handling should/must be fast */
1344 if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5)
1345 DEBUG_REQ(D_WARNING, req, "Slow req_in handling "CFS_DURATION_T"s",
1346 cfs_time_sub(cfs_time_current_sec(),
1347 req->rq_arrival_time.tv_sec));
1349 /* Set rpc server deadline and add it to the timed list */
1350 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1351 MSGHDR_AT_SUPPORT) ?
1352 /* The max time the client expects us to take */
1353 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1354 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1355 if (unlikely(deadline == 0)) {
1356 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1360 ptlrpc_at_add_timed(req);
1361 rc = ptlrpc_hpreq_init(svc, req);
1365 /* Move it over to the request processing queue */
1366 rc = ptlrpc_server_request_add(svc, req);
1369 cfs_waitq_signal(&svc->srv_waitq);
1373 spin_lock(&svc->srv_lock);
1374 svc->srv_n_queued_reqs--;
1375 svc->srv_n_active_reqs++;
1376 spin_unlock(&svc->srv_lock);
1377 ptlrpc_server_finish_request(req);
1383 ptlrpc_server_handle_request(struct ptlrpc_service *svc,
1384 struct ptlrpc_thread *thread)
1386 struct obd_export *export = NULL;
1387 struct ptlrpc_request *request;
1388 struct timeval work_start;
1389 struct timeval work_end;
1397 spin_lock(&svc->srv_lock);
1398 if (unlikely(!ptlrpc_server_request_pending(svc, 0) ||
1401 /* !@%$# liblustre only has 1 thread */
1402 atomic_read(&svc->srv_n_difficult_replies) != 0 &&
1404 svc->srv_n_active_reqs >= (svc->srv_threads_running - 1)))) {
1405 /* Don't handle regular requests in the last thread, in order * re
1406 * to handle difficult replies (which might block other threads)
1407 * as well as handle any incoming reqs, early replies, etc.
1408 * That means we always need at least 2 service threads. */
1409 spin_unlock(&svc->srv_lock);
1413 request = ptlrpc_server_request_get(svc, 0);
1414 if (request == NULL) {
1415 spin_unlock(&svc->srv_lock);
1419 opc = lustre_msg_get_opc(request->rq_reqmsg);
1420 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
1421 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
1422 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
1423 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
1425 if (unlikely(fail_opc)) {
1426 if (request->rq_export && request->rq_ops) {
1427 spin_unlock(&svc->srv_lock);
1428 OBD_FAIL_TIMEOUT(fail_opc, 4);
1429 spin_lock(&svc->srv_lock);
1430 request = ptlrpc_server_request_get(svc, 0);
1431 if (request == NULL) {
1432 spin_unlock(&svc->srv_lock);
1435 LASSERT(ptlrpc_server_request_pending(svc, 0));
1439 list_del_init(&request->rq_list);
1440 svc->srv_n_queued_reqs--;
1441 svc->srv_n_active_reqs++;
1445 /* The phase is changed under the lock here because we need to know
1446 * the request is under processing (see ptlrpc_hpreq_reorder()). */
1447 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
1448 spin_unlock(&svc->srv_lock);
1450 ptlrpc_hpreq_fini(request);
1452 if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
1453 libcfs_debug_dumplog();
1455 do_gettimeofday(&work_start);
1456 timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
1457 if (likely(svc->srv_stats != NULL)) {
1458 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
1460 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
1461 svc->srv_n_queued_reqs);
1462 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
1463 svc->srv_n_active_reqs);
1464 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
1465 at_get(&svc->srv_at_estimate));
1468 rc = lu_context_init(&request->rq_session,
1469 LCT_SESSION|LCT_REMEMBER|LCT_NOREF);
1471 CERROR("Failure to initialize session: %d\n", rc);
1474 request->rq_session.lc_thread = thread;
1475 request->rq_session.lc_cookie = 0x5;
1476 lu_context_enter(&request->rq_session);
1478 CDEBUG(D_NET, "got req "LPU64"\n", request->rq_xid);
1480 request->rq_svc_thread = thread;
1482 request->rq_svc_thread->t_env->le_ses = &request->rq_session;
1484 if (likely(request->rq_export)) {
1485 if (unlikely(ptlrpc_check_req(request)))
1487 ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
1488 export = class_export_rpc_get(request->rq_export);
1491 /* Discard requests queued for longer than the deadline.
1492 The deadline is increased if we send an early reply. */
1493 if (cfs_time_current_sec() > request->rq_deadline) {
1494 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s"
1495 ": deadline "CFS_DURATION_T":"CFS_DURATION_T"s ago\n",
1496 libcfs_id2str(request->rq_peer),
1497 cfs_time_sub(request->rq_deadline,
1498 request->rq_arrival_time.tv_sec),
1499 cfs_time_sub(cfs_time_current_sec(),
1500 request->rq_deadline));
1501 goto put_rpc_export;
1504 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
1505 "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1506 (request->rq_export ?
1507 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1508 (request->rq_export ?
1509 atomic_read(&request->rq_export->exp_refcount) : -99),
1510 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1511 libcfs_id2str(request->rq_peer),
1512 lustre_msg_get_opc(request->rq_reqmsg));
1514 OBD_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, obd_fail_val);
1516 rc = svc->srv_handler(request);
1518 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
1520 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
1521 "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1522 (request->rq_export ?
1523 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1524 (request->rq_export ?
1525 atomic_read(&request->rq_export->exp_refcount) : -99),
1526 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1527 libcfs_id2str(request->rq_peer),
1528 lustre_msg_get_opc(request->rq_reqmsg));
1532 class_export_rpc_put(export);
1534 lu_context_exit(&request->rq_session);
1535 lu_context_fini(&request->rq_session);
1537 if (unlikely(cfs_time_current_sec() > request->rq_deadline)) {
1538 DEBUG_REQ(D_WARNING, request, "Request x"LPU64" took longer "
1539 "than estimated ("CFS_DURATION_T":"CFS_DURATION_T"s);"
1540 " client may timeout.",
1541 request->rq_xid, cfs_time_sub(request->rq_deadline,
1542 request->rq_arrival_time.tv_sec),
1543 cfs_time_sub(cfs_time_current_sec(),
1544 request->rq_deadline));
1547 do_gettimeofday(&work_end);
1548 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1549 CDEBUG(D_RPCTRACE, "request x"LPU64" opc %u from %s processed in "
1550 "%ldus (%ldus total) trans "LPU64" rc %d/%d\n",
1551 request->rq_xid, lustre_msg_get_opc(request->rq_reqmsg),
1552 libcfs_id2str(request->rq_peer), timediff,
1553 cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL),
1554 request->rq_repmsg ? lustre_msg_get_transno(request->rq_repmsg) :
1555 request->rq_transno, request->rq_status,
1556 request->rq_repmsg ? lustre_msg_get_status(request->rq_repmsg):
1558 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
1559 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
1560 int opc = opcode_offset(op);
1561 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
1562 LASSERT(opc < LUSTRE_MAX_OPCODES);
1563 lprocfs_counter_add(svc->srv_stats,
1564 opc + EXTRA_MAX_OPCODES,
1568 if (unlikely(request->rq_early_count)) {
1569 DEBUG_REQ(D_ADAPTTO, request,
1570 "sent %d early replies before finishing in "
1572 request->rq_early_count,
1573 cfs_time_sub(work_end.tv_sec,
1574 request->rq_arrival_time.tv_sec));
1578 spin_lock(&svc->srv_lock);
1581 spin_unlock(&svc->srv_lock);
1582 ptlrpc_server_finish_request(request);
1588 * An internal function to process a single reply state object.
1591 ptlrpc_handle_rs (struct ptlrpc_reply_state *rs)
1593 struct ptlrpc_service *svc = rs->rs_service;
1594 struct obd_export *exp;
1595 struct obd_device *obd;
1600 exp = rs->rs_export;
1603 LASSERT (rs->rs_difficult);
1604 LASSERT (rs->rs_scheduled);
1605 LASSERT (list_empty(&rs->rs_list));
1607 spin_lock (&exp->exp_lock);
1608 /* Noop if removed already */
1609 list_del_init (&rs->rs_exp_list);
1610 spin_unlock (&exp->exp_lock);
1612 /* Avoid obd_uncommitted_replies_lock contention if we 100% sure that
1613 * rs has been removed from the list already */
1614 if (!list_empty_careful(&rs->rs_obd_list)) {
1615 spin_lock(&obd->obd_uncommitted_replies_lock);
1616 list_del_init(&rs->rs_obd_list);
1617 spin_unlock(&obd->obd_uncommitted_replies_lock);
1620 spin_lock(&rs->rs_lock);
1622 been_handled = rs->rs_handled;
1625 nlocks = rs->rs_nlocks; /* atomic "steal", but */
1626 rs->rs_nlocks = 0; /* locks still on rs_locks! */
1628 if (nlocks == 0 && !been_handled) {
1629 /* If we see this, we should already have seen the warning
1630 * in mds_steal_ack_locks() */
1631 CWARN("All locks stolen from rs %p x"LPD64".t"LPD64
1634 rs->rs_xid, rs->rs_transno,
1635 lustre_msg_get_opc(rs->rs_msg),
1636 libcfs_nid2str(exp->exp_connection->c_peer.nid));
1639 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
1640 spin_unlock(&rs->rs_lock);
1642 if (!been_handled && rs->rs_on_net) {
1643 LNetMDUnlink(rs->rs_md_h);
1644 /* Ignore return code; we're racing with
1648 while (nlocks-- > 0)
1649 ldlm_lock_decref(&rs->rs_locks[nlocks],
1650 rs->rs_modes[nlocks]);
1652 spin_lock(&rs->rs_lock);
1655 rs->rs_scheduled = 0;
1657 if (!rs->rs_on_net) {
1659 spin_unlock(&rs->rs_lock);
1661 class_export_put (exp);
1662 rs->rs_export = NULL;
1663 ptlrpc_rs_decref (rs);
1664 atomic_dec (&svc->srv_outstanding_replies);
1665 if (atomic_dec_and_test(&svc->srv_n_difficult_replies) &&
1666 svc->srv_is_stopping)
1667 cfs_waitq_broadcast(&svc->srv_waitq);
1671 /* still on the net; callback will schedule */
1672 spin_unlock(&rs->rs_lock);
1679 * Check whether given service has a reply available for processing
1682 * \param svc a ptlrpc service
1683 * \retval 0 no replies processes
1684 * \retval 1 one reply processed
1687 ptlrpc_server_handle_reply(struct ptlrpc_service *svc)
1689 struct ptlrpc_reply_state *rs = NULL;
1692 spin_lock(&svc->srv_lock);
1693 if (!list_empty(&svc->srv_reply_queue)) {
1694 rs = list_entry(svc->srv_reply_queue.prev,
1695 struct ptlrpc_reply_state,
1697 list_del_init(&rs->rs_list);
1699 spin_unlock(&svc->srv_lock);
1701 ptlrpc_handle_rs(rs);
1705 /* FIXME make use of timeout later */
1707 liblustre_check_services (void *arg)
1709 int did_something = 0;
1711 struct list_head *tmp, *nxt;
1714 /* I'm relying on being single threaded, not to have to lock
1715 * ptlrpc_all_services etc */
1716 list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
1717 struct ptlrpc_service *svc =
1718 list_entry (tmp, struct ptlrpc_service, srv_list);
1720 if (svc->srv_threads_running != 0) /* I've recursed */
1723 /* service threads can block for bulk, so this limits us
1724 * (arbitrarily) to recursing 1 stack frame per service.
1725 * Note that the problem with recursion is that we have to
1726 * unwind completely before our caller can resume. */
1728 svc->srv_threads_running++;
1731 rc = ptlrpc_server_handle_req_in(svc);
1732 rc |= ptlrpc_server_handle_reply(svc);
1733 rc |= ptlrpc_at_check_timed(svc);
1734 rc |= ptlrpc_server_handle_request(svc, NULL);
1735 rc |= (ptlrpc_server_post_idle_rqbds(svc) > 0);
1736 did_something |= rc;
1739 svc->srv_threads_running--;
1742 RETURN(did_something);
1744 #define ptlrpc_stop_all_threads(s) do {} while (0)
1746 #else /* __KERNEL__ */
1748 /* Don't use daemonize, it removes fs struct from new thread (bug 418) */
1749 void ptlrpc_daemonize(char *name)
1751 struct fs_struct *fs = current->fs;
1753 atomic_inc(&fs->count);
1754 cfs_daemonize(name);
1755 exit_fs(cfs_current());
1757 ll_set_fs_pwd(current->fs, init_task.fs->pwdmnt, init_task.fs->pwd);
1761 ptlrpc_check_rqbd_pool(struct ptlrpc_service *svc)
1763 int avail = svc->srv_nrqbd_receiving;
1764 int low_water = test_req_buffer_pressure ? 0 :
1765 svc->srv_nbuf_per_group/2;
1767 /* NB I'm not locking; just looking. */
1769 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
1770 * allowed the request history to grow out of control. We could put a
1771 * sanity check on that here and cull some history if we need the
1774 if (avail <= low_water)
1775 ptlrpc_grow_req_bufs(svc);
1778 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQBUF_AVAIL_CNTR,
1783 ptlrpc_retry_rqbds(void *arg)
1785 struct ptlrpc_service *svc = (struct ptlrpc_service *)arg;
1787 svc->srv_rqbd_timeout = 0;
1788 return (-ETIMEDOUT);
1791 static int ptlrpc_main(void *arg)
1793 struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
1794 struct ptlrpc_service *svc = data->svc;
1795 struct ptlrpc_thread *thread = data->thread;
1796 struct obd_device *dev = data->dev;
1797 struct ptlrpc_reply_state *rs;
1798 #ifdef WITH_GROUP_INFO
1799 struct group_info *ginfo = NULL;
1802 int counter = 0, rc = 0;
1805 ptlrpc_daemonize(data->name);
1807 #if defined(HAVE_NODE_TO_CPUMASK) && defined(CONFIG_NUMA)
1808 /* we need to do this before any per-thread allocation is done so that
1809 * we get the per-thread allocations on local node. bug 7342 */
1810 if (svc->srv_cpu_affinity) {
1813 for (cpu = 0, num_cpu = 0; cpu < num_possible_cpus(); cpu++) {
1814 if (!cpu_online(cpu))
1816 if (num_cpu == thread->t_id % num_online_cpus())
1820 set_cpus_allowed(cfs_current(), node_to_cpumask(cpu_to_node(cpu)));
1824 #ifdef WITH_GROUP_INFO
1825 ginfo = groups_alloc(0);
1831 set_current_groups(ginfo);
1832 put_group_info(ginfo);
1835 if (svc->srv_init != NULL) {
1836 rc = svc->srv_init(thread);
1841 rc = lu_context_init(&env.le_ctx,
1842 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
1846 thread->t_env = &env;
1847 env.le_ctx.lc_thread = thread;
1848 env.le_ctx.lc_cookie = 0x6;
1850 /* Alloc reply state structure for this one */
1851 OBD_ALLOC_GFP(rs, svc->srv_max_reply_size, CFS_ALLOC_STD);
1857 /* Record that the thread is running */
1858 thread->t_flags = SVC_RUNNING;
1860 * wake up our creator. Note: @data is invalid after this point,
1861 * because it's allocated on ptlrpc_start_thread() stack.
1863 cfs_waitq_signal(&thread->t_ctl_waitq);
1865 thread->t_watchdog = lc_watchdog_add(max_t(int, obd_timeout, AT_OFF ? 0 :
1866 at_get(&svc->srv_at_estimate))
1867 * svc->srv_watchdog_factor,
1870 spin_lock(&svc->srv_lock);
1871 svc->srv_threads_running++;
1872 list_add(&rs->rs_list, &svc->srv_free_rs_list);
1873 spin_unlock(&svc->srv_lock);
1874 cfs_waitq_signal(&svc->srv_free_rs_waitq);
1876 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
1877 svc->srv_threads_running);
1879 /* XXX maintain a list of all managed devices: insert here */
1881 while ((thread->t_flags & SVC_STOPPING) == 0) {
1882 /* Don't exit while there are replies to be handled */
1883 struct l_wait_info lwi = LWI_TIMEOUT(svc->srv_rqbd_timeout,
1884 ptlrpc_retry_rqbds, svc);
1886 lc_watchdog_disable(thread->t_watchdog);
1890 l_wait_event_exclusive (svc->srv_waitq,
1891 ((thread->t_flags & SVC_STOPPING) != 0) ||
1892 (!list_empty(&svc->srv_idle_rqbds) &&
1893 svc->srv_rqbd_timeout == 0) ||
1894 !list_empty(&svc->srv_req_in_queue) ||
1895 (ptlrpc_server_request_pending(svc, 0) &&
1896 (svc->srv_n_active_reqs <
1897 (svc->srv_threads_running - 1))) ||
1901 lc_watchdog_touch_ms(thread->t_watchdog, max_t(int, obd_timeout,
1903 at_get(&svc->srv_at_estimate)) *
1904 svc->srv_watchdog_factor);
1906 ptlrpc_check_rqbd_pool(svc);
1908 if ((svc->srv_threads_started < svc->srv_threads_max) &&
1909 (svc->srv_n_active_reqs >= (svc->srv_threads_started - 1))){
1910 /* Ignore return code - we tried... */
1911 ptlrpc_start_thread(dev, svc);
1914 if (!list_empty(&svc->srv_req_in_queue)) {
1915 /* Process all incoming reqs before handling any */
1916 ptlrpc_server_handle_req_in(svc);
1917 /* but limit ourselves in case of flood */
1918 if (counter++ < 1000)
1923 if (svc->srv_at_check)
1924 ptlrpc_at_check_timed(svc);
1926 /* don't handle requests in the last thread */
1927 if (ptlrpc_server_request_pending(svc, 0) &&
1928 (svc->srv_n_active_reqs < (svc->srv_threads_running - 1))) {
1929 lu_context_enter(&env.le_ctx);
1930 ptlrpc_server_handle_request(svc, thread);
1931 lu_context_exit(&env.le_ctx);
1934 if (!list_empty(&svc->srv_idle_rqbds) &&
1935 ptlrpc_server_post_idle_rqbds(svc) < 0) {
1936 /* I just failed to repost request buffers. Wait
1937 * for a timeout (unless something else happens)
1938 * before I try again */
1939 svc->srv_rqbd_timeout = cfs_time_seconds(1)/10;
1940 CDEBUG(D_RPCTRACE,"Posted buffers: %d\n",
1941 svc->srv_nrqbd_receiving);
1945 lc_watchdog_delete(thread->t_watchdog);
1946 thread->t_watchdog = NULL;
1950 * deconstruct service specific state created by ptlrpc_start_thread()
1952 if (svc->srv_done != NULL)
1953 svc->srv_done(thread);
1955 lu_context_fini(&env.le_ctx);
1957 CDEBUG(D_NET, "service thread %d exiting: rc %d\n", thread->t_id, rc);
1959 spin_lock(&svc->srv_lock);
1960 svc->srv_threads_running--; /* must know immediately */
1962 thread->t_flags = SVC_STOPPED;
1964 cfs_waitq_signal(&thread->t_ctl_waitq);
1965 spin_unlock(&svc->srv_lock);
1970 struct ptlrpc_hr_args {
1973 struct ptlrpc_hr_service *hrs;
1976 static int hrt_dont_sleep(struct ptlrpc_hr_thread *t,
1977 struct list_head *replies)
1981 spin_lock(&t->hrt_lock);
1982 list_splice_init(&t->hrt_queue, replies);
1983 result = test_bit(HRT_STOPPING, &t->hrt_flags) ||
1984 !list_empty(replies);
1985 spin_unlock(&t->hrt_lock);
1989 static int ptlrpc_hr_main(void *arg)
1991 struct ptlrpc_hr_args * hr_args = arg;
1992 struct ptlrpc_hr_service *hr = hr_args->hrs;
1993 struct ptlrpc_hr_thread *t = &hr->hr_threads[hr_args->thread_index];
1994 char threadname[20];
1995 CFS_LIST_HEAD(replies);
1997 snprintf(threadname, sizeof(threadname),
1998 "ptlrpc_hr_%d", hr_args->thread_index);
2000 ptlrpc_daemonize(threadname);
2001 #if defined(HAVE_NODE_TO_CPUMASK)
2002 set_cpus_allowed(cfs_current(),
2003 node_to_cpumask(cpu_to_node(hr_args->cpu_index)));
2005 set_bit(HRT_RUNNING, &t->hrt_flags);
2006 cfs_waitq_signal(&t->hrt_wait);
2008 while (!test_bit(HRT_STOPPING, &t->hrt_flags)) {
2010 cfs_wait_event(t->hrt_wait, hrt_dont_sleep(t, &replies));
2011 while (!list_empty(&replies)) {
2012 struct ptlrpc_reply_state *rs;
2014 rs = list_entry(replies.prev,
2015 struct ptlrpc_reply_state,
2017 list_del_init(&rs->rs_list);
2018 ptlrpc_handle_rs(rs);
2022 clear_bit(HRT_RUNNING, &t->hrt_flags);
2023 complete(&t->hrt_completion);
2028 static int ptlrpc_start_hr_thread(struct ptlrpc_hr_service *hr, int n, int cpu)
2030 struct ptlrpc_hr_thread *t = &hr->hr_threads[n];
2031 struct ptlrpc_hr_args args;
2035 args.thread_index = n;
2036 args.cpu_index = cpu;
2039 rc = cfs_kernel_thread(ptlrpc_hr_main, (void*)&args,
2040 CLONE_VM|CLONE_FILES);
2042 complete(&t->hrt_completion);
2045 cfs_wait_event(t->hrt_wait, test_bit(HRT_RUNNING, &t->hrt_flags));
2051 static void ptlrpc_stop_hr_thread(struct ptlrpc_hr_thread *t)
2055 set_bit(HRT_STOPPING, &t->hrt_flags);
2056 cfs_waitq_signal(&t->hrt_wait);
2057 wait_for_completion(&t->hrt_completion);
2062 static void ptlrpc_stop_hr_threads(struct ptlrpc_hr_service *hrs)
2067 for (n = 0; n < hrs->hr_n_threads; n++)
2068 ptlrpc_stop_hr_thread(&hrs->hr_threads[n]);
2073 static int ptlrpc_start_hr_threads(struct ptlrpc_hr_service *hr)
2076 int n, cpu, threads_started = 0;
2079 LASSERT(hr != NULL);
2080 LASSERT(hr->hr_n_threads > 0);
2082 for (n = 0, cpu = 0; n < hr->hr_n_threads; n++) {
2083 #if defined(HAVE_NODE_TO_CPUMASK)
2084 while(!cpu_online(cpu)) {
2086 if (cpu >= num_possible_cpus())
2090 rc = ptlrpc_start_hr_thread(hr, n, cpu);
2096 if (threads_started == 0) {
2097 CERROR("No reply handling threads started\n");
2100 if (threads_started < hr->hr_n_threads) {
2101 CWARN("Started only %d reply handling threads from %d\n",
2102 threads_started, hr->hr_n_threads);
2103 hr->hr_n_threads = threads_started;
2108 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
2109 struct ptlrpc_thread *thread)
2111 struct l_wait_info lwi = { 0 };
2114 CDEBUG(D_RPCTRACE, "Stopping thread %p\n", thread);
2115 spin_lock(&svc->srv_lock);
2116 thread->t_flags = SVC_STOPPING;
2117 spin_unlock(&svc->srv_lock);
2119 cfs_waitq_broadcast(&svc->srv_waitq);
2120 l_wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED),
2123 spin_lock(&svc->srv_lock);
2124 list_del(&thread->t_link);
2125 spin_unlock(&svc->srv_lock);
2127 OBD_FREE_PTR(thread);
2131 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2133 struct ptlrpc_thread *thread;
2136 spin_lock(&svc->srv_lock);
2137 while (!list_empty(&svc->srv_threads)) {
2138 thread = list_entry(svc->srv_threads.next,
2139 struct ptlrpc_thread, t_link);
2141 spin_unlock(&svc->srv_lock);
2142 ptlrpc_stop_thread(svc, thread);
2143 spin_lock(&svc->srv_lock);
2146 spin_unlock(&svc->srv_lock);
2150 int ptlrpc_start_threads(struct obd_device *dev, struct ptlrpc_service *svc)
2155 /* We require 2 threads min - see note in
2156 ptlrpc_server_handle_request */
2157 LASSERT(svc->srv_threads_min >= 2);
2158 for (i = 0; i < svc->srv_threads_min; i++) {
2159 rc = ptlrpc_start_thread(dev, svc);
2160 /* We have enough threads, don't start more. b=15759 */
2164 CERROR("cannot start %s thread #%d: rc %d\n",
2165 svc->srv_thread_name, i, rc);
2166 ptlrpc_stop_all_threads(svc);
2172 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc)
2174 struct l_wait_info lwi = { 0 };
2175 struct ptlrpc_svc_data d;
2176 struct ptlrpc_thread *thread;
2181 CDEBUG(D_RPCTRACE, "%s started %d min %d max %d running %d\n",
2182 svc->srv_name, svc->srv_threads_started, svc->srv_threads_min,
2183 svc->srv_threads_max, svc->srv_threads_running);
2184 if (unlikely(svc->srv_threads_started >= svc->srv_threads_max) ||
2185 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
2186 svc->srv_threads_started == svc->srv_threads_min - 1))
2189 OBD_ALLOC_PTR(thread);
2192 cfs_waitq_init(&thread->t_ctl_waitq);
2194 spin_lock(&svc->srv_lock);
2195 if (svc->srv_threads_started >= svc->srv_threads_max) {
2196 spin_unlock(&svc->srv_lock);
2197 OBD_FREE_PTR(thread);
2200 list_add(&thread->t_link, &svc->srv_threads);
2201 id = svc->srv_threads_started++;
2202 spin_unlock(&svc->srv_lock);
2205 sprintf(name, "%s_%02d", svc->srv_thread_name, id);
2211 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", name);
2213 /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
2214 * just drop the VM and FILES in ptlrpc_daemonize() right away.
2216 rc = cfs_kernel_thread(ptlrpc_main, &d, CLONE_VM | CLONE_FILES);
2218 CERROR("cannot start thread '%s': rc %d\n", name, rc);
2220 spin_lock(&svc->srv_lock);
2221 list_del(&thread->t_link);
2222 --svc->srv_threads_started;
2223 spin_unlock(&svc->srv_lock);
2225 OBD_FREE(thread, sizeof(*thread));
2228 l_wait_event(thread->t_ctl_waitq,
2229 thread->t_flags & (SVC_RUNNING | SVC_STOPPED), &lwi);
2231 rc = (thread->t_flags & SVC_STOPPED) ? thread->t_id : 0;
2236 int ptlrpc_hr_init(void)
2239 int n_cpus = num_online_cpus();
2240 struct ptlrpc_hr_service *hr;
2244 LASSERT(ptlrpc_hr == NULL);
2246 size = offsetof(struct ptlrpc_hr_service, hr_threads[n_cpus]);
2247 OBD_ALLOC(hr, size);
2250 for (i = 0; i < n_cpus; i++) {
2251 struct ptlrpc_hr_thread *t = &hr->hr_threads[i];
2253 spin_lock_init(&t->hrt_lock);
2254 cfs_waitq_init(&t->hrt_wait);
2255 CFS_INIT_LIST_HEAD(&t->hrt_queue);
2256 init_completion(&t->hrt_completion);
2258 hr->hr_n_threads = n_cpus;
2262 RETURN(ptlrpc_start_hr_threads(hr));
2265 void ptlrpc_hr_fini(void)
2267 if (ptlrpc_hr != NULL) {
2268 ptlrpc_stop_hr_threads(ptlrpc_hr);
2269 OBD_FREE(ptlrpc_hr, ptlrpc_hr->hr_size);
2274 #endif /* __KERNEL__ */
2277 * Wait until all already scheduled replies are processed.
2279 static void ptlrpc_wait_replies(struct ptlrpc_service *svc)
2283 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
2285 rc = l_wait_event(svc->srv_waitq,
2286 atomic_read(&svc->srv_n_difficult_replies) == 0,
2290 CWARN("Unexpectedly long timeout %p\n", svc);
2294 int ptlrpc_unregister_service(struct ptlrpc_service *service)
2297 struct l_wait_info lwi;
2298 struct list_head *tmp;
2299 struct ptlrpc_reply_state *rs, *t;
2302 service->srv_is_stopping = 1;
2303 cfs_timer_disarm(&service->srv_at_timer);
2305 ptlrpc_stop_all_threads(service);
2306 LASSERT(list_empty(&service->srv_threads));
2308 spin_lock (&ptlrpc_all_services_lock);
2309 list_del_init (&service->srv_list);
2310 spin_unlock (&ptlrpc_all_services_lock);
2312 ptlrpc_lprocfs_unregister_service(service);
2314 /* All history will be culled when the next request buffer is
2316 service->srv_max_history_rqbds = 0;
2318 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
2320 rc = LNetClearLazyPortal(service->srv_req_portal);
2323 /* Unlink all the request buffers. This forces a 'final' event with
2324 * its 'unlink' flag set for each posted rqbd */
2325 list_for_each(tmp, &service->srv_active_rqbds) {
2326 struct ptlrpc_request_buffer_desc *rqbd =
2327 list_entry(tmp, struct ptlrpc_request_buffer_desc,
2330 rc = LNetMDUnlink(rqbd->rqbd_md_h);
2331 LASSERT (rc == 0 || rc == -ENOENT);
2334 /* Wait for the network to release any buffers it's currently
2337 spin_lock(&service->srv_lock);
2338 rc = service->srv_nrqbd_receiving;
2339 spin_unlock(&service->srv_lock);
2344 /* Network access will complete in finite time but the HUGE
2345 * timeout lets us CWARN for visibility of sluggish NALs */
2346 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2347 cfs_time_seconds(1), NULL, NULL);
2348 rc = l_wait_event(service->srv_waitq,
2349 service->srv_nrqbd_receiving == 0,
2351 if (rc == -ETIMEDOUT)
2352 CWARN("Service %s waiting for request buffers\n",
2356 /* schedule all outstanding replies to terminate them */
2357 spin_lock(&service->srv_lock);
2358 while (!list_empty(&service->srv_active_replies)) {
2359 struct ptlrpc_reply_state *rs =
2360 list_entry(service->srv_active_replies.next,
2361 struct ptlrpc_reply_state, rs_list);
2362 spin_lock(&rs->rs_lock);
2363 ptlrpc_schedule_difficult_reply(rs);
2364 spin_unlock(&rs->rs_lock);
2366 spin_unlock(&service->srv_lock);
2368 /* purge the request queue. NB No new replies (rqbds all unlinked)
2369 * and no service threads, so I'm the only thread noodling the
2370 * request queue now */
2371 while (!list_empty(&service->srv_req_in_queue)) {
2372 struct ptlrpc_request *req =
2373 list_entry(service->srv_req_in_queue.next,
2374 struct ptlrpc_request,
2377 list_del(&req->rq_list);
2378 service->srv_n_queued_reqs--;
2379 service->srv_n_active_reqs++;
2380 ptlrpc_server_finish_request(req);
2382 while (ptlrpc_server_request_pending(service, 1)) {
2383 struct ptlrpc_request *req;
2385 req = ptlrpc_server_request_get(service, 1);
2386 list_del(&req->rq_list);
2387 service->srv_n_queued_reqs--;
2388 service->srv_n_active_reqs++;
2389 ptlrpc_hpreq_fini(req);
2390 ptlrpc_server_finish_request(req);
2392 LASSERT(service->srv_n_queued_reqs == 0);
2393 LASSERT(service->srv_n_active_reqs == 0);
2394 LASSERT(service->srv_n_history_rqbds == 0);
2395 LASSERT(list_empty(&service->srv_active_rqbds));
2397 /* Now free all the request buffers since nothing references them
2399 while (!list_empty(&service->srv_idle_rqbds)) {
2400 struct ptlrpc_request_buffer_desc *rqbd =
2401 list_entry(service->srv_idle_rqbds.next,
2402 struct ptlrpc_request_buffer_desc,
2405 ptlrpc_free_rqbd(rqbd);
2408 ptlrpc_wait_replies(service);
2410 list_for_each_entry_safe(rs, t, &service->srv_free_rs_list, rs_list) {
2411 list_del(&rs->rs_list);
2412 OBD_FREE(rs, service->srv_max_reply_size);
2415 /* In case somebody rearmed this in the meantime */
2416 cfs_timer_disarm(&service->srv_at_timer);
2418 OBD_FREE_PTR(service);
2422 /* Returns 0 if the service is healthy.
2424 * Right now, it just checks to make sure that requests aren't languishing
2425 * in the queue. We'll use this health check to govern whether a node needs
2426 * to be shot, so it's intentionally non-aggressive. */
2427 int ptlrpc_service_health_check(struct ptlrpc_service *svc)
2429 struct ptlrpc_request *request;
2430 struct timeval right_now;
2436 do_gettimeofday(&right_now);
2438 spin_lock(&svc->srv_lock);
2439 if (!ptlrpc_server_request_pending(svc, 1)) {
2440 spin_unlock(&svc->srv_lock);
2444 /* How long has the next entry been waiting? */
2445 if (list_empty(&svc->srv_request_queue))
2446 request = list_entry(svc->srv_request_hpq.next,
2447 struct ptlrpc_request, rq_list);
2449 request = list_entry(svc->srv_request_queue.next,
2450 struct ptlrpc_request, rq_list);
2451 timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
2452 spin_unlock(&svc->srv_lock);
2454 if ((timediff / ONE_MILLION) > (AT_OFF ? obd_timeout * 3/2 :
2456 CERROR("%s: unhealthy - request has been waiting %lds\n",
2457 svc->srv_name, timediff / ONE_MILLION);