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_at_array *array;
471 struct ptlrpc_service *service;
472 unsigned int size, index;
476 LASSERT (bufsize >= max_req_size + SPTLRPC_MAX_PAYLOAD);
477 LASSERT (ctx_tags != 0);
479 OBD_ALLOC_PTR(service);
483 /* First initialise enough for early teardown */
485 service->srv_name = name;
486 spin_lock_init(&service->srv_lock);
487 CFS_INIT_LIST_HEAD(&service->srv_threads);
488 cfs_waitq_init(&service->srv_waitq);
490 service->srv_nbuf_per_group = test_req_buffer_pressure ? 1 : nbufs;
491 service->srv_max_req_size = max_req_size + SPTLRPC_MAX_PAYLOAD;
492 service->srv_buf_size = bufsize;
493 service->srv_rep_portal = rep_portal;
494 service->srv_req_portal = req_portal;
495 service->srv_watchdog_factor = watchdog_factor;
496 service->srv_handler = handler;
497 service->srv_request_history_print_fn = svcreq_printfn;
498 service->srv_request_seq = 1; /* valid seq #s start at 1 */
499 service->srv_request_max_cull_seq = 0;
500 service->srv_threads_min = min_threads;
501 service->srv_threads_max = max_threads;
502 service->srv_thread_name = threadname;
503 service->srv_ctx_tags = ctx_tags;
504 service->srv_hpreq_handler = hp_handler;
505 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
506 service->srv_hpreq_count = 0;
507 service->srv_n_hpreq = 0;
509 rc = LNetSetLazyPortal(service->srv_req_portal);
512 CFS_INIT_LIST_HEAD(&service->srv_request_queue);
513 CFS_INIT_LIST_HEAD(&service->srv_request_hpq);
514 CFS_INIT_LIST_HEAD(&service->srv_idle_rqbds);
515 CFS_INIT_LIST_HEAD(&service->srv_active_rqbds);
516 CFS_INIT_LIST_HEAD(&service->srv_history_rqbds);
517 CFS_INIT_LIST_HEAD(&service->srv_request_history);
518 CFS_INIT_LIST_HEAD(&service->srv_active_replies);
520 CFS_INIT_LIST_HEAD(&service->srv_reply_queue);
522 CFS_INIT_LIST_HEAD(&service->srv_free_rs_list);
523 cfs_waitq_init(&service->srv_free_rs_waitq);
524 atomic_set(&service->srv_n_difficult_replies, 0);
526 spin_lock_init(&service->srv_at_lock);
527 CFS_INIT_LIST_HEAD(&service->srv_req_in_queue);
529 array = &service->srv_at_array;
530 size = at_est2timeout(at_max);
531 array->paa_size = size;
532 array->paa_count = 0;
533 array->paa_deadline = -1;
535 /* allocate memory for srv_at_array (ptlrpc_at_array) */
536 OBD_ALLOC(array->paa_reqs_array, sizeof(struct list_head) * size);
537 if (array->paa_reqs_array == NULL)
540 for (index = 0; index < size; index++)
541 CFS_INIT_LIST_HEAD(&array->paa_reqs_array[index]);
543 OBD_ALLOC(array->paa_reqs_count, sizeof(__u32) * size);
544 if (array->paa_reqs_count == NULL)
547 cfs_timer_init(&service->srv_at_timer, ptlrpc_at_timer, service);
548 /* At SOW, service time should be quick; 10s seems generous. If client
549 timeout is less than this, we'll be sending an early reply. */
550 at_init(&service->srv_at_estimate, 10, 0);
552 spin_lock (&ptlrpc_all_services_lock);
553 list_add (&service->srv_list, &ptlrpc_all_services);
554 spin_unlock (&ptlrpc_all_services_lock);
556 /* Now allocate the request buffers */
557 rc = ptlrpc_grow_req_bufs(service);
558 /* We shouldn't be under memory pressure at startup, so
559 * fail if we can't post all our buffers at this time. */
563 /* Now allocate pool of reply buffers */
564 /* Increase max reply size to next power of two */
565 service->srv_max_reply_size = 1;
566 while (service->srv_max_reply_size <
567 max_reply_size + SPTLRPC_MAX_PAYLOAD)
568 service->srv_max_reply_size <<= 1;
570 if (proc_entry != NULL)
571 ptlrpc_lprocfs_register_service(proc_entry, service);
573 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
574 service->srv_name, service->srv_req_portal);
578 ptlrpc_unregister_service(service);
583 * to actually free the request, must be called without holding svc_lock.
584 * note it's caller's responsibility to unlink req->rq_list.
586 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
588 LASSERT(atomic_read(&req->rq_refcount) == 0);
589 LASSERT(list_empty(&req->rq_timed_list));
591 /* DEBUG_REQ() assumes the reply state of a request with a valid
592 * ref will not be destroyed until that reference is dropped. */
593 ptlrpc_req_drop_rs(req);
595 sptlrpc_svc_ctx_decref(req);
597 if (req != &req->rq_rqbd->rqbd_req) {
598 /* NB request buffers use an embedded
599 * req if the incoming req unlinked the
600 * MD; this isn't one of them! */
601 OBD_FREE(req, sizeof(*req));
606 * drop a reference count of the request. if it reaches 0, we either
607 * put it into history list, or free it immediately.
609 static void ptlrpc_server_drop_request(struct ptlrpc_request *req)
611 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
612 struct ptlrpc_service *svc = rqbd->rqbd_service;
614 struct list_head *tmp;
615 struct list_head *nxt;
617 if (!atomic_dec_and_test(&req->rq_refcount))
620 spin_lock(&svc->srv_lock);
622 svc->srv_n_active_reqs--;
623 list_add(&req->rq_list, &rqbd->rqbd_reqs);
625 refcount = --(rqbd->rqbd_refcount);
627 /* request buffer is now idle: add to history */
628 list_del(&rqbd->rqbd_list);
629 list_add_tail(&rqbd->rqbd_list, &svc->srv_history_rqbds);
630 svc->srv_n_history_rqbds++;
632 /* cull some history?
633 * I expect only about 1 or 2 rqbds need to be recycled here */
634 while (svc->srv_n_history_rqbds > svc->srv_max_history_rqbds) {
635 rqbd = list_entry(svc->srv_history_rqbds.next,
636 struct ptlrpc_request_buffer_desc,
639 list_del(&rqbd->rqbd_list);
640 svc->srv_n_history_rqbds--;
642 /* remove rqbd's reqs from svc's req history while
643 * I've got the service lock */
644 list_for_each(tmp, &rqbd->rqbd_reqs) {
645 req = list_entry(tmp, struct ptlrpc_request,
647 /* Track the highest culled req seq */
648 if (req->rq_history_seq >
649 svc->srv_request_max_cull_seq)
650 svc->srv_request_max_cull_seq =
652 list_del(&req->rq_history_list);
655 spin_unlock(&svc->srv_lock);
657 list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
658 req = list_entry(rqbd->rqbd_reqs.next,
659 struct ptlrpc_request,
661 list_del(&req->rq_list);
662 ptlrpc_server_free_request(req);
665 spin_lock(&svc->srv_lock);
667 * now all reqs including the embedded req has been
668 * disposed, schedule request buffer for re-use.
670 LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) == 0);
671 list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
674 spin_unlock(&svc->srv_lock);
675 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
676 /* If we are low on memory, we are not interested in history */
677 list_del(&req->rq_list);
678 list_del_init(&req->rq_history_list);
679 spin_unlock(&svc->srv_lock);
681 ptlrpc_server_free_request(req);
683 spin_unlock(&svc->srv_lock);
688 * to finish a request: stop sending more early replies, and release
689 * the request. should be called after we finished handling the request.
691 static void ptlrpc_server_finish_request(struct ptlrpc_request *req)
693 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
695 if (req->rq_export) {
696 class_export_put(req->rq_export);
697 req->rq_export = NULL;
700 if (req->rq_phase != RQ_PHASE_NEW) /* incorrect message magic */
701 DEBUG_REQ(D_INFO, req, "free req");
703 spin_lock(&svc->srv_at_lock);
704 req->rq_sent_final = 1;
705 list_del_init(&req->rq_timed_list);
706 if (req->rq_at_linked) {
707 struct ptlrpc_at_array *array = &svc->srv_at_array;
708 __u32 index = req->rq_at_index;
710 req->rq_at_linked = 0;
711 array->paa_reqs_count[index]--;
714 spin_unlock(&svc->srv_at_lock);
716 ptlrpc_server_drop_request(req);
719 /* This function makes sure dead exports are evicted in a timely manner.
720 This function is only called when some export receives a message (i.e.,
721 the network is up.) */
722 static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
724 struct obd_export *oldest_exp;
725 time_t oldest_time, new_time;
731 /* Compensate for slow machines, etc, by faking our request time
732 into the future. Although this can break the strict time-ordering
733 of the list, we can be really lazy here - we don't have to evict
734 at the exact right moment. Eventually, all silent exports
735 will make it to the top of the list. */
737 /* Do not pay attention on 1sec or smaller renewals. */
738 new_time = cfs_time_current_sec() + extra_delay;
739 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
742 exp->exp_last_request_time = new_time;
743 CDEBUG(D_HA, "updating export %s at "CFS_TIME_T" exp %p\n",
744 exp->exp_client_uuid.uuid,
745 exp->exp_last_request_time, exp);
747 /* exports may get disconnected from the chain even though the
748 export has references, so we must keep the spin lock while
749 manipulating the lists */
750 spin_lock(&exp->exp_obd->obd_dev_lock);
752 if (list_empty(&exp->exp_obd_chain_timed)) {
753 /* this one is not timed */
754 spin_unlock(&exp->exp_obd->obd_dev_lock);
758 list_move_tail(&exp->exp_obd_chain_timed,
759 &exp->exp_obd->obd_exports_timed);
761 oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
762 struct obd_export, exp_obd_chain_timed);
763 oldest_time = oldest_exp->exp_last_request_time;
764 spin_unlock(&exp->exp_obd->obd_dev_lock);
766 if (exp->exp_obd->obd_recovering) {
767 /* be nice to everyone during recovery */
772 /* Note - racing to start/reset the obd_eviction timer is safe */
773 if (exp->exp_obd->obd_eviction_timer == 0) {
774 /* Check if the oldest entry is expired. */
775 if (cfs_time_current_sec() > (oldest_time + PING_EVICT_TIMEOUT +
777 /* We need a second timer, in case the net was down and
778 * it just came back. Since the pinger may skip every
779 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
780 * we better wait for 3. */
781 exp->exp_obd->obd_eviction_timer =
782 cfs_time_current_sec() + 3 * PING_INTERVAL;
783 CDEBUG(D_HA, "%s: Think about evicting %s from "CFS_TIME_T"\n",
784 exp->exp_obd->obd_name, obd_export_nid2str(exp),
788 if (cfs_time_current_sec() >
789 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
790 /* The evictor won't evict anyone who we've heard from
791 * recently, so we don't have to check before we start
793 if (!ping_evictor_wake(exp))
794 exp->exp_obd->obd_eviction_timer = 0;
801 static int ptlrpc_check_req(struct ptlrpc_request *req)
803 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
804 req->rq_export->exp_conn_cnt)) {
805 DEBUG_REQ(D_ERROR, req,
806 "DROPPING req from old connection %d < %d",
807 lustre_msg_get_conn_cnt(req->rq_reqmsg),
808 req->rq_export->exp_conn_cnt);
811 if (unlikely(req->rq_export->exp_obd &&
812 req->rq_export->exp_obd->obd_fail)) {
813 /* Failing over, don't handle any more reqs, send
814 error response instead. */
815 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
816 req, req->rq_export->exp_obd->obd_name);
817 req->rq_status = -ENODEV;
825 static void ptlrpc_at_set_timer(struct ptlrpc_service *svc)
827 struct ptlrpc_at_array *array = &svc->srv_at_array;
830 spin_lock(&svc->srv_at_lock);
831 if (array->paa_count == 0) {
832 cfs_timer_disarm(&svc->srv_at_timer);
833 spin_unlock(&svc->srv_at_lock);
837 /* Set timer for closest deadline */
838 next = (__s32)(array->paa_deadline - cfs_time_current_sec() -
841 ptlrpc_at_timer((unsigned long)svc);
843 cfs_timer_arm(&svc->srv_at_timer, cfs_time_shift(next));
844 spin_unlock(&svc->srv_at_lock);
845 CDEBUG(D_INFO, "armed %s at %+ds\n", svc->srv_name, next);
848 /* Add rpc to early reply check list */
849 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
851 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
852 struct ptlrpc_request *rq = NULL;
853 struct ptlrpc_at_array *array = &svc->srv_at_array;
860 if (req->rq_no_reply)
863 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
866 spin_lock(&svc->srv_at_lock);
868 if (unlikely(req->rq_sent_final)) {
869 spin_unlock(&svc->srv_at_lock);
873 LASSERT(list_empty(&req->rq_timed_list));
875 wtimes = req->rq_deadline / array->paa_size;
876 index = req->rq_deadline % array->paa_size;
877 if (array->paa_reqs_count[index] > 0)
878 rq = list_entry(array->paa_reqs_array[index].next,
879 struct ptlrpc_request, rq_timed_list);
881 if (rq != NULL && (rq->rq_deadline / array->paa_size) < wtimes) {
882 /* latest rpcs will have the latest deadlines in the list,
883 * so search backward. */
884 list_for_each_entry_reverse(rq, &array->paa_reqs_array[index],
886 if (req->rq_deadline >= rq->rq_deadline) {
887 list_add(&req->rq_timed_list,
893 /* AT array is corrupted? */
894 LASSERT(!list_empty(&req->rq_timed_list));
896 /* Add the request at the head of the list */
897 list_add(&req->rq_timed_list, &array->paa_reqs_array[index]);
900 req->rq_at_linked = 1;
901 req->rq_at_index = index;
902 array->paa_reqs_count[index]++;
904 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
905 array->paa_deadline = req->rq_deadline;
908 spin_unlock(&svc->srv_at_lock);
911 ptlrpc_at_set_timer(svc);
916 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req,
919 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
920 struct ptlrpc_request *reqcopy;
921 struct lustre_msg *reqmsg;
922 cfs_duration_t olddl = req->rq_deadline - cfs_time_current_sec();
927 /* deadline is when the client expects us to reply, margin is the
928 difference between clients' and servers' expectations */
929 DEBUG_REQ(D_ADAPTTO, req,
930 "%ssending early reply (deadline %+lds, margin %+lds) for "
931 "%d+%d", AT_OFF ? "AT off - not " : "",
932 olddl, olddl - at_get(&svc->srv_at_estimate),
933 at_get(&svc->srv_at_estimate), extra_time);
939 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), "
940 "not sending early reply. Consider increasing "
941 "at_early_margin (%d)?", olddl, at_early_margin);
943 /* Return an error so we're not re-added to the timed list. */
947 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
948 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
949 "but no AT support");
953 if (req->rq_export && req->rq_export->exp_in_recovery) {
954 /* don't increase server estimates during recovery, and give
955 clients the full recovery time. */
956 newdl = cfs_time_current_sec() +
957 req->rq_export->exp_obd->obd_recovery_timeout;
960 /* Fake our processing time into the future to ask the
961 clients for some extra amount of time */
962 extra_time += cfs_time_current_sec() -
963 req->rq_arrival_time.tv_sec;
964 at_add(&svc->srv_at_estimate, extra_time);
966 newdl = req->rq_arrival_time.tv_sec +
967 at_get(&svc->srv_at_estimate);
969 if (req->rq_deadline >= newdl) {
970 /* We're not adding any time, no need to send an early reply
971 (e.g. maybe at adaptive_max) */
972 DEBUG_REQ(D_WARNING, req, "Couldn't add any time ("
973 CFS_DURATION_T"/"CFS_DURATION_T"), "
974 "not sending early reply\n", olddl,
975 cfs_time_sub(newdl, cfs_time_current_sec()));
979 OBD_ALLOC(reqcopy, sizeof *reqcopy);
982 OBD_ALLOC(reqmsg, req->rq_reqlen);
984 OBD_FREE(reqcopy, sizeof *reqcopy);
989 reqcopy->rq_reply_state = NULL;
990 reqcopy->rq_rep_swab_mask = 0;
991 reqcopy->rq_pack_bulk = 0;
992 reqcopy->rq_pack_udesc = 0;
993 reqcopy->rq_packed_final = 0;
994 sptlrpc_svc_ctx_addref(reqcopy);
995 /* We only need the reqmsg for the magic */
996 reqcopy->rq_reqmsg = reqmsg;
997 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
999 if (req->rq_sent_final) {
1000 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
1001 "abort sending early reply\n");
1005 /* Connection ref */
1006 reqcopy->rq_export = class_conn2export(
1007 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1008 if (reqcopy->rq_export == NULL)
1009 GOTO(out, rc = -ENODEV);
1012 class_export_rpc_get(reqcopy->rq_export);
1013 if (reqcopy->rq_export->exp_obd &&
1014 reqcopy->rq_export->exp_obd->obd_fail)
1015 GOTO(out_put, rc = -ENODEV);
1017 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1021 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1024 /* Adjust our own deadline to what we told the client */
1025 req->rq_deadline = newdl;
1026 req->rq_early_count++; /* number sent, server side */
1028 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1031 /* Free the (early) reply state from lustre_pack_reply.
1032 (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
1033 ptlrpc_req_drop_rs(reqcopy);
1036 class_export_rpc_put(reqcopy->rq_export);
1037 class_export_put(reqcopy->rq_export);
1039 sptlrpc_svc_ctx_decref(reqcopy);
1040 OBD_FREE(reqmsg, req->rq_reqlen);
1041 OBD_FREE(reqcopy, sizeof *reqcopy);
1045 /* Send early replies to everybody expiring within at_early_margin
1046 asking for at_extra time */
1047 static int ptlrpc_at_check_timed(struct ptlrpc_service *svc)
1049 struct ptlrpc_request *rq, *n;
1050 struct list_head work_list;
1051 struct ptlrpc_at_array *array = &svc->srv_at_array;
1054 time_t now = cfs_time_current_sec();
1055 cfs_duration_t delay;
1056 int first, counter = 0;
1059 spin_lock(&svc->srv_at_lock);
1060 if (svc->srv_at_check == 0) {
1061 spin_unlock(&svc->srv_at_lock);
1064 delay = cfs_time_sub(cfs_time_current(), svc->srv_at_checktime);
1065 svc->srv_at_check = 0;
1067 if (array->paa_count == 0) {
1068 spin_unlock(&svc->srv_at_lock);
1072 /* The timer went off, but maybe the nearest rpc already completed. */
1073 first = array->paa_deadline - now;
1074 if (first > at_early_margin) {
1075 /* We've still got plenty of time. Reset the timer. */
1076 spin_unlock(&svc->srv_at_lock);
1077 ptlrpc_at_set_timer(svc);
1081 /* We're close to a timeout, and we don't know how much longer the
1082 server will take. Send early replies to everyone expiring soon. */
1083 CFS_INIT_LIST_HEAD(&work_list);
1085 index = array->paa_deadline % array->paa_size;
1086 count = array->paa_count;
1088 count -= array->paa_reqs_count[index];
1089 list_for_each_entry_safe(rq, n, &array->paa_reqs_array[index],
1091 if (rq->rq_deadline <= now + at_early_margin) {
1092 list_move(&rq->rq_timed_list, &work_list);
1094 array->paa_reqs_count[index]--;
1096 rq->rq_at_linked = 0;
1100 /* update the earliest deadline */
1101 if (deadline == -1 || rq->rq_deadline < deadline)
1102 deadline = rq->rq_deadline;
1107 if (++index >= array->paa_size)
1110 array->paa_deadline = deadline;
1111 spin_unlock(&svc->srv_at_lock);
1113 /* we have a new earliest deadline, restart the timer */
1114 ptlrpc_at_set_timer(svc);
1116 CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
1117 "replies\n", first, at_extra, counter);
1119 /* We're already past request deadlines before we even get a
1120 chance to send early replies */
1121 LCONSOLE_WARN("%s: This server is not able to keep up with "
1122 "request traffic (cpu-bound).\n", svc->srv_name);
1123 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, "
1124 "delay="CFS_DURATION_T"(jiff)\n",
1125 counter, svc->srv_n_queued_reqs, svc->srv_n_active_reqs,
1126 at_get(&svc->srv_at_estimate), delay);
1129 /* ptlrpc_server_finish_request may delete an entry out of
1131 spin_lock(&svc->srv_at_lock);
1132 while (!list_empty(&work_list)) {
1133 rq = list_entry(work_list.next, struct ptlrpc_request,
1135 list_del_init(&rq->rq_timed_list);
1136 /* if the entry is still in the worklist, it hasn't been
1137 deleted, and is safe to take a ref to keep the req around */
1138 atomic_inc(&rq->rq_refcount);
1139 spin_unlock(&svc->srv_at_lock);
1141 if (ptlrpc_at_send_early_reply(rq, at_extra) == 0)
1142 ptlrpc_at_add_timed(rq);
1144 ptlrpc_server_drop_request(rq);
1145 spin_lock(&svc->srv_at_lock);
1147 spin_unlock(&svc->srv_at_lock);
1153 * Put the request to the export list if the request may become
1154 * a high priority one.
1156 static int ptlrpc_hpreq_init(struct ptlrpc_service *svc,
1157 struct ptlrpc_request *req)
1162 if (svc->srv_hpreq_handler) {
1163 rc = svc->srv_hpreq_handler(req);
1167 if (req->rq_export && req->rq_ops) {
1168 spin_lock(&req->rq_export->exp_lock);
1169 list_add(&req->rq_exp_list, &req->rq_export->exp_queued_rpc);
1170 spin_unlock(&req->rq_export->exp_lock);
1176 /** Remove the request from the export list. */
1177 static void ptlrpc_hpreq_fini(struct ptlrpc_request *req)
1180 if (req->rq_export && req->rq_ops) {
1181 spin_lock(&req->rq_export->exp_lock);
1182 list_del_init(&req->rq_exp_list);
1183 spin_unlock(&req->rq_export->exp_lock);
1189 * Make the request a high priority one.
1191 * All the high priority requests are queued in a separate FIFO
1192 * ptlrpc_service::srv_request_hpq list which is parallel to
1193 * ptlrpc_service::srv_request_queue list but has a higher priority
1196 * \see ptlrpc_server_handle_request().
1198 static void ptlrpc_hpreq_reorder_nolock(struct ptlrpc_service *svc,
1199 struct ptlrpc_request *req)
1202 LASSERT(svc != NULL);
1203 spin_lock(&req->rq_lock);
1204 if (req->rq_hp == 0) {
1205 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1207 /* Add to the high priority queue. */
1208 list_move_tail(&req->rq_list, &svc->srv_request_hpq);
1210 if (opc != OBD_PING)
1211 DEBUG_REQ(D_NET, req, "high priority req");
1213 spin_unlock(&req->rq_lock);
1217 void ptlrpc_hpreq_reorder(struct ptlrpc_request *req)
1219 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
1222 spin_lock(&svc->srv_lock);
1223 /* It may happen that the request is already taken for the processing
1224 * but still in the export list, do not re-add it into the HP list. */
1225 if (req->rq_phase == RQ_PHASE_NEW)
1226 ptlrpc_hpreq_reorder_nolock(svc, req);
1227 spin_unlock(&svc->srv_lock);
1231 /** Check if the request if a high priority one. */
1232 static int ptlrpc_server_hpreq_check(struct ptlrpc_request *req)
1237 /* Check by request opc. */
1238 opc = lustre_msg_get_opc(req->rq_reqmsg);
1239 if (opc == OBD_PING)
1242 /* Perform request specific check. */
1243 if (req->rq_ops && req->rq_ops->hpreq_check)
1244 rc = req->rq_ops->hpreq_check(req);
1248 /** Check if a request is a high priority one. */
1249 static int ptlrpc_server_request_add(struct ptlrpc_service *svc,
1250 struct ptlrpc_request *req)
1255 rc = ptlrpc_server_hpreq_check(req);
1259 spin_lock(&svc->srv_lock);
1260 /* Before inserting the request into the queue, check if it is not
1261 * inserted yet, or even already handled -- it may happen due to
1262 * a racing ldlm_server_blocking_ast(). */
1263 if (req->rq_phase == RQ_PHASE_NEW && list_empty(&req->rq_list)) {
1265 ptlrpc_hpreq_reorder_nolock(svc, req);
1267 list_add_tail(&req->rq_list, &svc->srv_request_queue);
1269 spin_unlock(&svc->srv_lock);
1274 /* Only allow normal priority requests on a service that has a high-priority
1275 * queue if forced (i.e. cleanup), if there are other high priority requests
1276 * already being processed (i.e. those threads can service more high-priority
1277 * requests), or if there are enough idle threads that a later thread can do
1278 * a high priority request. */
1279 static int ptlrpc_server_allow_normal(struct ptlrpc_service *svc, int force)
1281 return force || !svc->srv_hpreq_handler || svc->srv_n_hpreq > 0 ||
1282 svc->srv_n_active_reqs < svc->srv_threads_running - 2;
1285 static struct ptlrpc_request *
1286 ptlrpc_server_request_get(struct ptlrpc_service *svc, int force)
1288 struct ptlrpc_request *req = NULL;
1291 if (ptlrpc_server_allow_normal(svc, force) &&
1292 !list_empty(&svc->srv_request_queue) &&
1293 (list_empty(&svc->srv_request_hpq) ||
1294 svc->srv_hpreq_count >= svc->srv_hpreq_ratio)) {
1295 req = list_entry(svc->srv_request_queue.next,
1296 struct ptlrpc_request, rq_list);
1297 svc->srv_hpreq_count = 0;
1298 } else if (!list_empty(&svc->srv_request_hpq)) {
1299 req = list_entry(svc->srv_request_hpq.next,
1300 struct ptlrpc_request, rq_list);
1301 svc->srv_hpreq_count++;
1306 static int ptlrpc_server_request_pending(struct ptlrpc_service *svc, int force)
1308 return ((ptlrpc_server_allow_normal(svc, force) &&
1309 !list_empty(&svc->srv_request_queue)) ||
1310 !list_empty(&svc->srv_request_hpq));
1313 /* Handle freshly incoming reqs, add to timed early reply list,
1314 pass on to regular request queue */
1316 ptlrpc_server_handle_req_in(struct ptlrpc_service *svc)
1318 struct ptlrpc_request *req;
1325 spin_lock(&svc->srv_lock);
1326 if (list_empty(&svc->srv_req_in_queue)) {
1327 spin_unlock(&svc->srv_lock);
1331 req = list_entry(svc->srv_req_in_queue.next,
1332 struct ptlrpc_request, rq_list);
1333 list_del_init (&req->rq_list);
1334 /* Consider this still a "queued" request as far as stats are
1336 spin_unlock(&svc->srv_lock);
1338 /* go through security check/transform */
1339 rc = sptlrpc_svc_unwrap_request(req);
1343 case SECSVC_COMPLETE:
1344 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1352 /* Clear request swab mask; this is a new request */
1353 req->rq_req_swab_mask = 0;
1355 rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
1357 CERROR("error unpacking request: ptl %d from %s x"LPU64"\n",
1358 svc->srv_req_portal, libcfs_id2str(req->rq_peer),
1363 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1365 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
1366 LPU64"\n", svc->srv_req_portal,
1367 libcfs_id2str(req->rq_peer), req->rq_xid);
1372 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1373 CERROR("wrong packet type received (type=%u) from %s\n",
1374 lustre_msg_get_type(req->rq_reqmsg),
1375 libcfs_id2str(req->rq_peer));
1379 switch(lustre_msg_get_opc(req->rq_reqmsg)) {
1382 req->rq_bulk_write = 1;
1386 req->rq_bulk_read = 1;
1390 CDEBUG(D_NET, "got req "LPD64"\n", req->rq_xid);
1392 req->rq_export = class_conn2export(
1393 lustre_msg_get_handle(req->rq_reqmsg));
1394 if (req->rq_export) {
1395 rc = ptlrpc_check_req(req);
1397 rc = sptlrpc_target_export_check(req->rq_export, req);
1399 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
1400 "illegal security flavor,");
1405 ptlrpc_update_export_timer(req->rq_export, 0);
1408 /* req_in handling should/must be fast */
1409 if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5)
1410 DEBUG_REQ(D_WARNING, req, "Slow req_in handling "CFS_DURATION_T"s",
1411 cfs_time_sub(cfs_time_current_sec(),
1412 req->rq_arrival_time.tv_sec));
1414 /* Set rpc server deadline and add it to the timed list */
1415 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1416 MSGHDR_AT_SUPPORT) ?
1417 /* The max time the client expects us to take */
1418 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1419 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1420 if (unlikely(deadline == 0)) {
1421 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1425 ptlrpc_at_add_timed(req);
1426 rc = ptlrpc_hpreq_init(svc, req);
1430 /* Move it over to the request processing queue */
1431 rc = ptlrpc_server_request_add(svc, req);
1434 cfs_waitq_signal(&svc->srv_waitq);
1438 spin_lock(&svc->srv_lock);
1439 svc->srv_n_queued_reqs--;
1440 svc->srv_n_active_reqs++;
1441 spin_unlock(&svc->srv_lock);
1442 ptlrpc_server_finish_request(req);
1448 ptlrpc_server_handle_request(struct ptlrpc_service *svc,
1449 struct ptlrpc_thread *thread)
1451 struct obd_export *export = NULL;
1452 struct ptlrpc_request *request;
1453 struct timeval work_start;
1454 struct timeval work_end;
1462 spin_lock(&svc->srv_lock);
1463 if (unlikely(!ptlrpc_server_request_pending(svc, 0) ||
1466 /* !@%$# liblustre only has 1 thread */
1467 atomic_read(&svc->srv_n_difficult_replies) != 0 &&
1469 svc->srv_n_active_reqs >= (svc->srv_threads_running - 1)))) {
1470 /* Don't handle regular requests in the last thread, in order * re
1471 * to handle difficult replies (which might block other threads)
1472 * as well as handle any incoming reqs, early replies, etc.
1473 * That means we always need at least 2 service threads. */
1474 spin_unlock(&svc->srv_lock);
1478 request = ptlrpc_server_request_get(svc, 0);
1479 if (request == NULL) {
1480 spin_unlock(&svc->srv_lock);
1484 opc = lustre_msg_get_opc(request->rq_reqmsg);
1485 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
1486 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
1487 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
1488 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
1490 if (unlikely(fail_opc)) {
1491 if (request->rq_export && request->rq_ops) {
1492 spin_unlock(&svc->srv_lock);
1493 OBD_FAIL_TIMEOUT(fail_opc, 4);
1494 spin_lock(&svc->srv_lock);
1495 request = ptlrpc_server_request_get(svc, 0);
1496 if (request == NULL) {
1497 spin_unlock(&svc->srv_lock);
1500 LASSERT(ptlrpc_server_request_pending(svc, 0));
1504 list_del_init(&request->rq_list);
1505 svc->srv_n_queued_reqs--;
1506 svc->srv_n_active_reqs++;
1510 /* The phase is changed under the lock here because we need to know
1511 * the request is under processing (see ptlrpc_hpreq_reorder()). */
1512 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
1513 spin_unlock(&svc->srv_lock);
1515 ptlrpc_hpreq_fini(request);
1517 if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
1518 libcfs_debug_dumplog();
1520 do_gettimeofday(&work_start);
1521 timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
1522 if (likely(svc->srv_stats != NULL)) {
1523 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
1525 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
1526 svc->srv_n_queued_reqs);
1527 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
1528 svc->srv_n_active_reqs);
1529 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
1530 at_get(&svc->srv_at_estimate));
1533 rc = lu_context_init(&request->rq_session,
1534 LCT_SESSION|LCT_REMEMBER|LCT_NOREF);
1536 CERROR("Failure to initialize session: %d\n", rc);
1539 request->rq_session.lc_thread = thread;
1540 request->rq_session.lc_cookie = 0x5;
1541 lu_context_enter(&request->rq_session);
1543 CDEBUG(D_NET, "got req "LPU64"\n", request->rq_xid);
1545 request->rq_svc_thread = thread;
1547 request->rq_svc_thread->t_env->le_ses = &request->rq_session;
1549 if (likely(request->rq_export)) {
1550 if (unlikely(ptlrpc_check_req(request)))
1552 ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
1553 export = class_export_rpc_get(request->rq_export);
1556 /* Discard requests queued for longer than the deadline.
1557 The deadline is increased if we send an early reply. */
1558 if (cfs_time_current_sec() > request->rq_deadline) {
1559 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s"
1560 ": deadline "CFS_DURATION_T":"CFS_DURATION_T"s ago\n",
1561 libcfs_id2str(request->rq_peer),
1562 cfs_time_sub(request->rq_deadline,
1563 request->rq_arrival_time.tv_sec),
1564 cfs_time_sub(cfs_time_current_sec(),
1565 request->rq_deadline));
1566 goto put_rpc_export;
1569 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
1570 "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1571 (request->rq_export ?
1572 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1573 (request->rq_export ?
1574 atomic_read(&request->rq_export->exp_refcount) : -99),
1575 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1576 libcfs_id2str(request->rq_peer),
1577 lustre_msg_get_opc(request->rq_reqmsg));
1579 OBD_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, obd_fail_val);
1581 rc = svc->srv_handler(request);
1583 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
1585 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
1586 "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1587 (request->rq_export ?
1588 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1589 (request->rq_export ?
1590 atomic_read(&request->rq_export->exp_refcount) : -99),
1591 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1592 libcfs_id2str(request->rq_peer),
1593 lustre_msg_get_opc(request->rq_reqmsg));
1597 class_export_rpc_put(export);
1599 lu_context_exit(&request->rq_session);
1600 lu_context_fini(&request->rq_session);
1602 if (unlikely(cfs_time_current_sec() > request->rq_deadline)) {
1603 DEBUG_REQ(D_WARNING, request, "Request x"LPU64" took longer "
1604 "than estimated ("CFS_DURATION_T":"CFS_DURATION_T"s);"
1605 " client may timeout.",
1606 request->rq_xid, cfs_time_sub(request->rq_deadline,
1607 request->rq_arrival_time.tv_sec),
1608 cfs_time_sub(cfs_time_current_sec(),
1609 request->rq_deadline));
1612 do_gettimeofday(&work_end);
1613 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1614 CDEBUG(D_RPCTRACE, "request x"LPU64" opc %u from %s processed in "
1615 "%ldus (%ldus total) trans "LPU64" rc %d/%d\n",
1616 request->rq_xid, lustre_msg_get_opc(request->rq_reqmsg),
1617 libcfs_id2str(request->rq_peer), timediff,
1618 cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL),
1619 request->rq_repmsg ? lustre_msg_get_transno(request->rq_repmsg) :
1620 request->rq_transno, request->rq_status,
1621 request->rq_repmsg ? lustre_msg_get_status(request->rq_repmsg):
1623 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
1624 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
1625 int opc = opcode_offset(op);
1626 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
1627 LASSERT(opc < LUSTRE_MAX_OPCODES);
1628 lprocfs_counter_add(svc->srv_stats,
1629 opc + EXTRA_MAX_OPCODES,
1633 if (unlikely(request->rq_early_count)) {
1634 DEBUG_REQ(D_ADAPTTO, request,
1635 "sent %d early replies before finishing in "
1637 request->rq_early_count,
1638 cfs_time_sub(work_end.tv_sec,
1639 request->rq_arrival_time.tv_sec));
1643 spin_lock(&svc->srv_lock);
1646 spin_unlock(&svc->srv_lock);
1647 ptlrpc_server_finish_request(request);
1653 * An internal function to process a single reply state object.
1656 ptlrpc_handle_rs (struct ptlrpc_reply_state *rs)
1658 struct ptlrpc_service *svc = rs->rs_service;
1659 struct obd_export *exp;
1660 struct obd_device *obd;
1665 exp = rs->rs_export;
1668 LASSERT (rs->rs_difficult);
1669 LASSERT (rs->rs_scheduled);
1670 LASSERT (list_empty(&rs->rs_list));
1672 spin_lock (&exp->exp_lock);
1673 /* Noop if removed already */
1674 list_del_init (&rs->rs_exp_list);
1675 spin_unlock (&exp->exp_lock);
1677 /* Avoid obd_uncommitted_replies_lock contention if we 100% sure that
1678 * rs has been removed from the list already */
1679 if (!list_empty_careful(&rs->rs_obd_list)) {
1680 spin_lock(&obd->obd_uncommitted_replies_lock);
1681 list_del_init(&rs->rs_obd_list);
1682 spin_unlock(&obd->obd_uncommitted_replies_lock);
1685 spin_lock(&rs->rs_lock);
1687 been_handled = rs->rs_handled;
1690 nlocks = rs->rs_nlocks; /* atomic "steal", but */
1691 rs->rs_nlocks = 0; /* locks still on rs_locks! */
1693 if (nlocks == 0 && !been_handled) {
1694 /* If we see this, we should already have seen the warning
1695 * in mds_steal_ack_locks() */
1696 CWARN("All locks stolen from rs %p x"LPD64".t"LPD64
1699 rs->rs_xid, rs->rs_transno,
1700 lustre_msg_get_opc(rs->rs_msg),
1701 libcfs_nid2str(exp->exp_connection->c_peer.nid));
1704 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
1705 spin_unlock(&rs->rs_lock);
1707 if (!been_handled && rs->rs_on_net) {
1708 LNetMDUnlink(rs->rs_md_h);
1709 /* Ignore return code; we're racing with
1713 while (nlocks-- > 0)
1714 ldlm_lock_decref(&rs->rs_locks[nlocks],
1715 rs->rs_modes[nlocks]);
1717 spin_lock(&rs->rs_lock);
1720 rs->rs_scheduled = 0;
1722 if (!rs->rs_on_net) {
1724 spin_unlock(&rs->rs_lock);
1726 class_export_put (exp);
1727 rs->rs_export = NULL;
1728 ptlrpc_rs_decref (rs);
1729 atomic_dec (&svc->srv_outstanding_replies);
1730 if (atomic_dec_and_test(&svc->srv_n_difficult_replies) &&
1731 svc->srv_is_stopping)
1732 cfs_waitq_broadcast(&svc->srv_waitq);
1736 /* still on the net; callback will schedule */
1737 spin_unlock(&rs->rs_lock);
1744 * Check whether given service has a reply available for processing
1747 * \param svc a ptlrpc service
1748 * \retval 0 no replies processes
1749 * \retval 1 one reply processed
1752 ptlrpc_server_handle_reply(struct ptlrpc_service *svc)
1754 struct ptlrpc_reply_state *rs = NULL;
1757 spin_lock(&svc->srv_lock);
1758 if (!list_empty(&svc->srv_reply_queue)) {
1759 rs = list_entry(svc->srv_reply_queue.prev,
1760 struct ptlrpc_reply_state,
1762 list_del_init(&rs->rs_list);
1764 spin_unlock(&svc->srv_lock);
1766 ptlrpc_handle_rs(rs);
1770 /* FIXME make use of timeout later */
1772 liblustre_check_services (void *arg)
1774 int did_something = 0;
1776 struct list_head *tmp, *nxt;
1779 /* I'm relying on being single threaded, not to have to lock
1780 * ptlrpc_all_services etc */
1781 list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
1782 struct ptlrpc_service *svc =
1783 list_entry (tmp, struct ptlrpc_service, srv_list);
1785 if (svc->srv_threads_running != 0) /* I've recursed */
1788 /* service threads can block for bulk, so this limits us
1789 * (arbitrarily) to recursing 1 stack frame per service.
1790 * Note that the problem with recursion is that we have to
1791 * unwind completely before our caller can resume. */
1793 svc->srv_threads_running++;
1796 rc = ptlrpc_server_handle_req_in(svc);
1797 rc |= ptlrpc_server_handle_reply(svc);
1798 rc |= ptlrpc_at_check_timed(svc);
1799 rc |= ptlrpc_server_handle_request(svc, NULL);
1800 rc |= (ptlrpc_server_post_idle_rqbds(svc) > 0);
1801 did_something |= rc;
1804 svc->srv_threads_running--;
1807 RETURN(did_something);
1809 #define ptlrpc_stop_all_threads(s) do {} while (0)
1811 #else /* __KERNEL__ */
1813 /* Don't use daemonize, it removes fs struct from new thread (bug 418) */
1814 void ptlrpc_daemonize(char *name)
1816 struct fs_struct *fs = current->fs;
1818 atomic_inc(&fs->count);
1819 cfs_daemonize(name);
1820 exit_fs(cfs_current());
1822 ll_set_fs_pwd(current->fs, init_task.fs->pwdmnt, init_task.fs->pwd);
1826 ptlrpc_check_rqbd_pool(struct ptlrpc_service *svc)
1828 int avail = svc->srv_nrqbd_receiving;
1829 int low_water = test_req_buffer_pressure ? 0 :
1830 svc->srv_nbuf_per_group/2;
1832 /* NB I'm not locking; just looking. */
1834 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
1835 * allowed the request history to grow out of control. We could put a
1836 * sanity check on that here and cull some history if we need the
1839 if (avail <= low_water)
1840 ptlrpc_grow_req_bufs(svc);
1843 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQBUF_AVAIL_CNTR,
1848 ptlrpc_retry_rqbds(void *arg)
1850 struct ptlrpc_service *svc = (struct ptlrpc_service *)arg;
1852 svc->srv_rqbd_timeout = 0;
1853 return (-ETIMEDOUT);
1856 static int ptlrpc_main(void *arg)
1858 struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
1859 struct ptlrpc_service *svc = data->svc;
1860 struct ptlrpc_thread *thread = data->thread;
1861 struct obd_device *dev = data->dev;
1862 struct ptlrpc_reply_state *rs;
1863 #ifdef WITH_GROUP_INFO
1864 struct group_info *ginfo = NULL;
1867 int counter = 0, rc = 0;
1870 ptlrpc_daemonize(data->name);
1872 #if defined(HAVE_NODE_TO_CPUMASK) && defined(CONFIG_NUMA)
1873 /* we need to do this before any per-thread allocation is done so that
1874 * we get the per-thread allocations on local node. bug 7342 */
1875 if (svc->srv_cpu_affinity) {
1878 for (cpu = 0, num_cpu = 0; cpu < num_possible_cpus(); cpu++) {
1879 if (!cpu_online(cpu))
1881 if (num_cpu == thread->t_id % num_online_cpus())
1885 set_cpus_allowed(cfs_current(), node_to_cpumask(cpu_to_node(cpu)));
1889 #ifdef WITH_GROUP_INFO
1890 ginfo = groups_alloc(0);
1896 set_current_groups(ginfo);
1897 put_group_info(ginfo);
1900 if (svc->srv_init != NULL) {
1901 rc = svc->srv_init(thread);
1906 rc = lu_context_init(&env.le_ctx,
1907 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
1911 thread->t_env = &env;
1912 env.le_ctx.lc_thread = thread;
1913 env.le_ctx.lc_cookie = 0x6;
1915 /* Alloc reply state structure for this one */
1916 OBD_ALLOC_GFP(rs, svc->srv_max_reply_size, CFS_ALLOC_STD);
1922 /* Record that the thread is running */
1923 thread->t_flags = SVC_RUNNING;
1925 * wake up our creator. Note: @data is invalid after this point,
1926 * because it's allocated on ptlrpc_start_thread() stack.
1928 cfs_waitq_signal(&thread->t_ctl_waitq);
1930 thread->t_watchdog = lc_watchdog_add(max_t(int, obd_timeout, AT_OFF ? 0 :
1931 at_get(&svc->srv_at_estimate))
1932 * svc->srv_watchdog_factor,
1935 spin_lock(&svc->srv_lock);
1936 svc->srv_threads_running++;
1937 list_add(&rs->rs_list, &svc->srv_free_rs_list);
1938 spin_unlock(&svc->srv_lock);
1939 cfs_waitq_signal(&svc->srv_free_rs_waitq);
1941 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
1942 svc->srv_threads_running);
1944 /* XXX maintain a list of all managed devices: insert here */
1946 while ((thread->t_flags & SVC_STOPPING) == 0) {
1947 /* Don't exit while there are replies to be handled */
1948 struct l_wait_info lwi = LWI_TIMEOUT(svc->srv_rqbd_timeout,
1949 ptlrpc_retry_rqbds, svc);
1951 lc_watchdog_disable(thread->t_watchdog);
1955 l_wait_event_exclusive (svc->srv_waitq,
1956 ((thread->t_flags & SVC_STOPPING) != 0) ||
1957 (!list_empty(&svc->srv_idle_rqbds) &&
1958 svc->srv_rqbd_timeout == 0) ||
1959 !list_empty(&svc->srv_req_in_queue) ||
1960 (ptlrpc_server_request_pending(svc, 0) &&
1961 (svc->srv_n_active_reqs <
1962 (svc->srv_threads_running - 1))) ||
1966 lc_watchdog_touch_ms(thread->t_watchdog, max_t(int, obd_timeout,
1968 at_get(&svc->srv_at_estimate)) *
1969 svc->srv_watchdog_factor);
1971 ptlrpc_check_rqbd_pool(svc);
1973 if ((svc->srv_threads_started < svc->srv_threads_max) &&
1974 (svc->srv_n_active_reqs >= (svc->srv_threads_started - 1))){
1975 /* Ignore return code - we tried... */
1976 ptlrpc_start_thread(dev, svc);
1979 if (!list_empty(&svc->srv_req_in_queue)) {
1980 /* Process all incoming reqs before handling any */
1981 ptlrpc_server_handle_req_in(svc);
1982 /* but limit ourselves in case of flood */
1983 if (counter++ < 1000)
1988 if (svc->srv_at_check)
1989 ptlrpc_at_check_timed(svc);
1991 /* don't handle requests in the last thread */
1992 if (ptlrpc_server_request_pending(svc, 0) &&
1993 (svc->srv_n_active_reqs < (svc->srv_threads_running - 1))) {
1994 lu_context_enter(&env.le_ctx);
1995 ptlrpc_server_handle_request(svc, thread);
1996 lu_context_exit(&env.le_ctx);
1999 if (!list_empty(&svc->srv_idle_rqbds) &&
2000 ptlrpc_server_post_idle_rqbds(svc) < 0) {
2001 /* I just failed to repost request buffers. Wait
2002 * for a timeout (unless something else happens)
2003 * before I try again */
2004 svc->srv_rqbd_timeout = cfs_time_seconds(1)/10;
2005 CDEBUG(D_RPCTRACE,"Posted buffers: %d\n",
2006 svc->srv_nrqbd_receiving);
2010 lc_watchdog_delete(thread->t_watchdog);
2011 thread->t_watchdog = NULL;
2015 * deconstruct service specific state created by ptlrpc_start_thread()
2017 if (svc->srv_done != NULL)
2018 svc->srv_done(thread);
2020 lu_context_fini(&env.le_ctx);
2022 CDEBUG(D_NET, "service thread %d exiting: rc %d\n", thread->t_id, rc);
2024 spin_lock(&svc->srv_lock);
2025 svc->srv_threads_running--; /* must know immediately */
2027 thread->t_flags = SVC_STOPPED;
2029 cfs_waitq_signal(&thread->t_ctl_waitq);
2030 spin_unlock(&svc->srv_lock);
2035 struct ptlrpc_hr_args {
2038 struct ptlrpc_hr_service *hrs;
2041 static int hrt_dont_sleep(struct ptlrpc_hr_thread *t,
2042 struct list_head *replies)
2046 spin_lock(&t->hrt_lock);
2047 list_splice_init(&t->hrt_queue, replies);
2048 result = test_bit(HRT_STOPPING, &t->hrt_flags) ||
2049 !list_empty(replies);
2050 spin_unlock(&t->hrt_lock);
2054 static int ptlrpc_hr_main(void *arg)
2056 struct ptlrpc_hr_args * hr_args = arg;
2057 struct ptlrpc_hr_service *hr = hr_args->hrs;
2058 struct ptlrpc_hr_thread *t = &hr->hr_threads[hr_args->thread_index];
2059 char threadname[20];
2060 CFS_LIST_HEAD(replies);
2062 snprintf(threadname, sizeof(threadname),
2063 "ptlrpc_hr_%d", hr_args->thread_index);
2065 ptlrpc_daemonize(threadname);
2066 #if defined(HAVE_NODE_TO_CPUMASK)
2067 set_cpus_allowed(cfs_current(),
2068 node_to_cpumask(cpu_to_node(hr_args->cpu_index)));
2070 set_bit(HRT_RUNNING, &t->hrt_flags);
2071 cfs_waitq_signal(&t->hrt_wait);
2073 while (!test_bit(HRT_STOPPING, &t->hrt_flags)) {
2075 cfs_wait_event(t->hrt_wait, hrt_dont_sleep(t, &replies));
2076 while (!list_empty(&replies)) {
2077 struct ptlrpc_reply_state *rs;
2079 rs = list_entry(replies.prev,
2080 struct ptlrpc_reply_state,
2082 list_del_init(&rs->rs_list);
2083 ptlrpc_handle_rs(rs);
2087 clear_bit(HRT_RUNNING, &t->hrt_flags);
2088 complete(&t->hrt_completion);
2093 static int ptlrpc_start_hr_thread(struct ptlrpc_hr_service *hr, int n, int cpu)
2095 struct ptlrpc_hr_thread *t = &hr->hr_threads[n];
2096 struct ptlrpc_hr_args args;
2100 args.thread_index = n;
2101 args.cpu_index = cpu;
2104 rc = cfs_kernel_thread(ptlrpc_hr_main, (void*)&args,
2105 CLONE_VM|CLONE_FILES);
2107 complete(&t->hrt_completion);
2110 cfs_wait_event(t->hrt_wait, test_bit(HRT_RUNNING, &t->hrt_flags));
2116 static void ptlrpc_stop_hr_thread(struct ptlrpc_hr_thread *t)
2120 set_bit(HRT_STOPPING, &t->hrt_flags);
2121 cfs_waitq_signal(&t->hrt_wait);
2122 wait_for_completion(&t->hrt_completion);
2127 static void ptlrpc_stop_hr_threads(struct ptlrpc_hr_service *hrs)
2132 for (n = 0; n < hrs->hr_n_threads; n++)
2133 ptlrpc_stop_hr_thread(&hrs->hr_threads[n]);
2138 static int ptlrpc_start_hr_threads(struct ptlrpc_hr_service *hr)
2141 int n, cpu, threads_started = 0;
2144 LASSERT(hr != NULL);
2145 LASSERT(hr->hr_n_threads > 0);
2147 for (n = 0, cpu = 0; n < hr->hr_n_threads; n++) {
2148 #if defined(HAVE_NODE_TO_CPUMASK)
2149 while(!cpu_online(cpu)) {
2151 if (cpu >= num_possible_cpus())
2155 rc = ptlrpc_start_hr_thread(hr, n, cpu);
2161 if (threads_started == 0) {
2162 CERROR("No reply handling threads started\n");
2165 if (threads_started < hr->hr_n_threads) {
2166 CWARN("Started only %d reply handling threads from %d\n",
2167 threads_started, hr->hr_n_threads);
2168 hr->hr_n_threads = threads_started;
2173 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
2174 struct ptlrpc_thread *thread)
2176 struct l_wait_info lwi = { 0 };
2179 CDEBUG(D_RPCTRACE, "Stopping thread %p\n", thread);
2180 spin_lock(&svc->srv_lock);
2181 thread->t_flags = SVC_STOPPING;
2182 spin_unlock(&svc->srv_lock);
2184 cfs_waitq_broadcast(&svc->srv_waitq);
2185 l_wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED),
2188 spin_lock(&svc->srv_lock);
2189 list_del(&thread->t_link);
2190 spin_unlock(&svc->srv_lock);
2192 OBD_FREE_PTR(thread);
2196 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2198 struct ptlrpc_thread *thread;
2201 spin_lock(&svc->srv_lock);
2202 while (!list_empty(&svc->srv_threads)) {
2203 thread = list_entry(svc->srv_threads.next,
2204 struct ptlrpc_thread, t_link);
2206 spin_unlock(&svc->srv_lock);
2207 ptlrpc_stop_thread(svc, thread);
2208 spin_lock(&svc->srv_lock);
2211 spin_unlock(&svc->srv_lock);
2215 int ptlrpc_start_threads(struct obd_device *dev, struct ptlrpc_service *svc)
2220 /* We require 2 threads min - see note in
2221 ptlrpc_server_handle_request */
2222 LASSERT(svc->srv_threads_min >= 2);
2223 for (i = 0; i < svc->srv_threads_min; i++) {
2224 rc = ptlrpc_start_thread(dev, svc);
2225 /* We have enough threads, don't start more. b=15759 */
2229 CERROR("cannot start %s thread #%d: rc %d\n",
2230 svc->srv_thread_name, i, rc);
2231 ptlrpc_stop_all_threads(svc);
2237 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc)
2239 struct l_wait_info lwi = { 0 };
2240 struct ptlrpc_svc_data d;
2241 struct ptlrpc_thread *thread;
2246 CDEBUG(D_RPCTRACE, "%s started %d min %d max %d running %d\n",
2247 svc->srv_name, svc->srv_threads_started, svc->srv_threads_min,
2248 svc->srv_threads_max, svc->srv_threads_running);
2249 if (unlikely(svc->srv_threads_started >= svc->srv_threads_max) ||
2250 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
2251 svc->srv_threads_started == svc->srv_threads_min - 1))
2254 OBD_ALLOC_PTR(thread);
2257 cfs_waitq_init(&thread->t_ctl_waitq);
2259 spin_lock(&svc->srv_lock);
2260 if (svc->srv_threads_started >= svc->srv_threads_max) {
2261 spin_unlock(&svc->srv_lock);
2262 OBD_FREE_PTR(thread);
2265 list_add(&thread->t_link, &svc->srv_threads);
2266 id = svc->srv_threads_started++;
2267 spin_unlock(&svc->srv_lock);
2270 sprintf(name, "%s_%02d", svc->srv_thread_name, id);
2276 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", name);
2278 /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
2279 * just drop the VM and FILES in ptlrpc_daemonize() right away.
2281 rc = cfs_kernel_thread(ptlrpc_main, &d, CLONE_VM | CLONE_FILES);
2283 CERROR("cannot start thread '%s': rc %d\n", name, rc);
2285 spin_lock(&svc->srv_lock);
2286 list_del(&thread->t_link);
2287 --svc->srv_threads_started;
2288 spin_unlock(&svc->srv_lock);
2290 OBD_FREE(thread, sizeof(*thread));
2293 l_wait_event(thread->t_ctl_waitq,
2294 thread->t_flags & (SVC_RUNNING | SVC_STOPPED), &lwi);
2296 rc = (thread->t_flags & SVC_STOPPED) ? thread->t_id : 0;
2301 int ptlrpc_hr_init(void)
2304 int n_cpus = num_online_cpus();
2305 struct ptlrpc_hr_service *hr;
2309 LASSERT(ptlrpc_hr == NULL);
2311 size = offsetof(struct ptlrpc_hr_service, hr_threads[n_cpus]);
2312 OBD_ALLOC(hr, size);
2315 for (i = 0; i < n_cpus; i++) {
2316 struct ptlrpc_hr_thread *t = &hr->hr_threads[i];
2318 spin_lock_init(&t->hrt_lock);
2319 cfs_waitq_init(&t->hrt_wait);
2320 CFS_INIT_LIST_HEAD(&t->hrt_queue);
2321 init_completion(&t->hrt_completion);
2323 hr->hr_n_threads = n_cpus;
2327 RETURN(ptlrpc_start_hr_threads(hr));
2330 void ptlrpc_hr_fini(void)
2332 if (ptlrpc_hr != NULL) {
2333 ptlrpc_stop_hr_threads(ptlrpc_hr);
2334 OBD_FREE(ptlrpc_hr, ptlrpc_hr->hr_size);
2339 #endif /* __KERNEL__ */
2342 * Wait until all already scheduled replies are processed.
2344 static void ptlrpc_wait_replies(struct ptlrpc_service *svc)
2348 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
2350 rc = l_wait_event(svc->srv_waitq,
2351 atomic_read(&svc->srv_n_difficult_replies) == 0,
2355 CWARN("Unexpectedly long timeout %p\n", svc);
2359 int ptlrpc_unregister_service(struct ptlrpc_service *service)
2362 struct l_wait_info lwi;
2363 struct list_head *tmp;
2364 struct ptlrpc_reply_state *rs, *t;
2365 struct ptlrpc_at_array *array = &service->srv_at_array;
2368 service->srv_is_stopping = 1;
2369 cfs_timer_disarm(&service->srv_at_timer);
2371 ptlrpc_stop_all_threads(service);
2372 LASSERT(list_empty(&service->srv_threads));
2374 spin_lock (&ptlrpc_all_services_lock);
2375 list_del_init (&service->srv_list);
2376 spin_unlock (&ptlrpc_all_services_lock);
2378 ptlrpc_lprocfs_unregister_service(service);
2380 /* All history will be culled when the next request buffer is
2382 service->srv_max_history_rqbds = 0;
2384 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
2386 rc = LNetClearLazyPortal(service->srv_req_portal);
2389 /* Unlink all the request buffers. This forces a 'final' event with
2390 * its 'unlink' flag set for each posted rqbd */
2391 list_for_each(tmp, &service->srv_active_rqbds) {
2392 struct ptlrpc_request_buffer_desc *rqbd =
2393 list_entry(tmp, struct ptlrpc_request_buffer_desc,
2396 rc = LNetMDUnlink(rqbd->rqbd_md_h);
2397 LASSERT (rc == 0 || rc == -ENOENT);
2400 /* Wait for the network to release any buffers it's currently
2403 spin_lock(&service->srv_lock);
2404 rc = service->srv_nrqbd_receiving;
2405 spin_unlock(&service->srv_lock);
2410 /* Network access will complete in finite time but the HUGE
2411 * timeout lets us CWARN for visibility of sluggish NALs */
2412 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2413 cfs_time_seconds(1), NULL, NULL);
2414 rc = l_wait_event(service->srv_waitq,
2415 service->srv_nrqbd_receiving == 0,
2417 if (rc == -ETIMEDOUT)
2418 CWARN("Service %s waiting for request buffers\n",
2422 /* schedule all outstanding replies to terminate them */
2423 spin_lock(&service->srv_lock);
2424 while (!list_empty(&service->srv_active_replies)) {
2425 struct ptlrpc_reply_state *rs =
2426 list_entry(service->srv_active_replies.next,
2427 struct ptlrpc_reply_state, rs_list);
2428 spin_lock(&rs->rs_lock);
2429 ptlrpc_schedule_difficult_reply(rs);
2430 spin_unlock(&rs->rs_lock);
2432 spin_unlock(&service->srv_lock);
2434 /* purge the request queue. NB No new replies (rqbds all unlinked)
2435 * and no service threads, so I'm the only thread noodling the
2436 * request queue now */
2437 while (!list_empty(&service->srv_req_in_queue)) {
2438 struct ptlrpc_request *req =
2439 list_entry(service->srv_req_in_queue.next,
2440 struct ptlrpc_request,
2443 list_del(&req->rq_list);
2444 service->srv_n_queued_reqs--;
2445 service->srv_n_active_reqs++;
2446 ptlrpc_server_finish_request(req);
2448 while (ptlrpc_server_request_pending(service, 1)) {
2449 struct ptlrpc_request *req;
2451 req = ptlrpc_server_request_get(service, 1);
2452 list_del(&req->rq_list);
2453 service->srv_n_queued_reqs--;
2454 service->srv_n_active_reqs++;
2455 ptlrpc_hpreq_fini(req);
2456 ptlrpc_server_finish_request(req);
2458 LASSERT(service->srv_n_queued_reqs == 0);
2459 LASSERT(service->srv_n_active_reqs == 0);
2460 LASSERT(service->srv_n_history_rqbds == 0);
2461 LASSERT(list_empty(&service->srv_active_rqbds));
2463 /* Now free all the request buffers since nothing references them
2465 while (!list_empty(&service->srv_idle_rqbds)) {
2466 struct ptlrpc_request_buffer_desc *rqbd =
2467 list_entry(service->srv_idle_rqbds.next,
2468 struct ptlrpc_request_buffer_desc,
2471 ptlrpc_free_rqbd(rqbd);
2474 ptlrpc_wait_replies(service);
2476 list_for_each_entry_safe(rs, t, &service->srv_free_rs_list, rs_list) {
2477 list_del(&rs->rs_list);
2478 OBD_FREE(rs, service->srv_max_reply_size);
2481 /* In case somebody rearmed this in the meantime */
2482 cfs_timer_disarm(&service->srv_at_timer);
2484 if (array->paa_reqs_array != NULL) {
2485 OBD_FREE(array->paa_reqs_array,
2486 sizeof(struct list_head) * array->paa_size);
2487 array->paa_reqs_array = NULL;
2490 if (array->paa_reqs_count != NULL) {
2491 OBD_FREE(array->paa_reqs_count,
2492 sizeof(__u32) * array->paa_size);
2493 array->paa_reqs_count= NULL;
2496 OBD_FREE_PTR(service);
2500 /* Returns 0 if the service is healthy.
2502 * Right now, it just checks to make sure that requests aren't languishing
2503 * in the queue. We'll use this health check to govern whether a node needs
2504 * to be shot, so it's intentionally non-aggressive. */
2505 int ptlrpc_service_health_check(struct ptlrpc_service *svc)
2507 struct ptlrpc_request *request;
2508 struct timeval right_now;
2514 do_gettimeofday(&right_now);
2516 spin_lock(&svc->srv_lock);
2517 if (!ptlrpc_server_request_pending(svc, 1)) {
2518 spin_unlock(&svc->srv_lock);
2522 /* How long has the next entry been waiting? */
2523 if (list_empty(&svc->srv_request_queue))
2524 request = list_entry(svc->srv_request_hpq.next,
2525 struct ptlrpc_request, rq_list);
2527 request = list_entry(svc->srv_request_queue.next,
2528 struct ptlrpc_request, rq_list);
2529 timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
2530 spin_unlock(&svc->srv_lock);
2532 if ((timediff / ONE_MILLION) > (AT_OFF ? obd_timeout * 3/2 :
2534 CERROR("%s: unhealthy - request has been waiting %lds\n",
2535 svc->srv_name, timediff / ONE_MILLION);