4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2010, 2012, Intel Corporation.
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 CFS_MODULE_PARM(at_min, "i", int, 0644,
53 "Adaptive timeout minimum (sec)");
54 CFS_MODULE_PARM(at_max, "i", int, 0644,
55 "Adaptive timeout maximum (sec)");
56 CFS_MODULE_PARM(at_history, "i", int, 0644,
57 "Adaptive timeouts remember the slowest event that took place "
58 "within this period (sec)");
59 CFS_MODULE_PARM(at_early_margin, "i", int, 0644,
60 "How soon before an RPC deadline to send an early reply");
61 CFS_MODULE_PARM(at_extra, "i", int, 0644,
62 "How much extra time to give with each early reply");
66 static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt);
67 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req);
68 static void ptlrpc_at_remove_timed(struct ptlrpc_request *req);
70 /** Holds a list of all PTLRPC services */
71 CFS_LIST_HEAD(ptlrpc_all_services);
72 /** Used to protect the \e ptlrpc_all_services list */
73 struct mutex ptlrpc_all_services_mutex;
75 struct ptlrpc_request_buffer_desc *
76 ptlrpc_alloc_rqbd(struct ptlrpc_service_part *svcpt)
78 struct ptlrpc_service *svc = svcpt->scp_service;
79 struct ptlrpc_request_buffer_desc *rqbd;
81 OBD_CPT_ALLOC_PTR(rqbd, svc->srv_cptable, svcpt->scp_cpt);
85 rqbd->rqbd_svcpt = svcpt;
86 rqbd->rqbd_refcount = 0;
87 rqbd->rqbd_cbid.cbid_fn = request_in_callback;
88 rqbd->rqbd_cbid.cbid_arg = rqbd;
89 CFS_INIT_LIST_HEAD(&rqbd->rqbd_reqs);
90 OBD_CPT_ALLOC_LARGE(rqbd->rqbd_buffer, svc->srv_cptable,
91 svcpt->scp_cpt, svc->srv_buf_size);
92 if (rqbd->rqbd_buffer == NULL) {
97 spin_lock(&svcpt->scp_lock);
98 cfs_list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
99 svcpt->scp_nrqbds_total++;
100 spin_unlock(&svcpt->scp_lock);
106 ptlrpc_free_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
108 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
110 LASSERT(rqbd->rqbd_refcount == 0);
111 LASSERT(cfs_list_empty(&rqbd->rqbd_reqs));
113 spin_lock(&svcpt->scp_lock);
114 cfs_list_del(&rqbd->rqbd_list);
115 svcpt->scp_nrqbds_total--;
116 spin_unlock(&svcpt->scp_lock);
118 OBD_FREE_LARGE(rqbd->rqbd_buffer, svcpt->scp_service->srv_buf_size);
123 ptlrpc_grow_req_bufs(struct ptlrpc_service_part *svcpt, int post)
125 struct ptlrpc_service *svc = svcpt->scp_service;
126 struct ptlrpc_request_buffer_desc *rqbd;
130 if (svcpt->scp_rqbd_allocating)
133 spin_lock(&svcpt->scp_lock);
134 /* check again with lock */
135 if (svcpt->scp_rqbd_allocating) {
136 /* NB: we might allow more than one thread in the future */
137 LASSERT(svcpt->scp_rqbd_allocating == 1);
138 spin_unlock(&svcpt->scp_lock);
142 svcpt->scp_rqbd_allocating++;
143 spin_unlock(&svcpt->scp_lock);
146 for (i = 0; i < svc->srv_nbuf_per_group; i++) {
147 /* NB: another thread might have recycled enough rqbds, we
148 * need to make sure it wouldn't over-allocate, see LU-1212. */
149 if (svcpt->scp_nrqbds_posted >= svc->srv_nbuf_per_group)
152 rqbd = ptlrpc_alloc_rqbd(svcpt);
155 CERROR("%s: Can't allocate request buffer\n",
162 spin_lock(&svcpt->scp_lock);
164 LASSERT(svcpt->scp_rqbd_allocating == 1);
165 svcpt->scp_rqbd_allocating--;
167 spin_unlock(&svcpt->scp_lock);
170 "%s: allocate %d new %d-byte reqbufs (%d/%d left), rc = %d\n",
171 svc->srv_name, i, svc->srv_buf_size, svcpt->scp_nrqbds_posted,
172 svcpt->scp_nrqbds_total, rc);
176 rc = ptlrpc_server_post_idle_rqbds(svcpt);
182 * Part of Rep-Ack logic.
183 * Puts a lock and its mode into reply state assotiated to request reply.
186 ptlrpc_save_lock(struct ptlrpc_request *req,
187 struct lustre_handle *lock, int mode, int no_ack)
189 struct ptlrpc_reply_state *rs = req->rq_reply_state;
193 LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
195 if (req->rq_export->exp_disconnected) {
196 ldlm_lock_decref(lock, mode);
198 idx = rs->rs_nlocks++;
199 rs->rs_locks[idx] = *lock;
200 rs->rs_modes[idx] = mode;
201 rs->rs_difficult = 1;
202 rs->rs_no_ack = !!no_ack;
205 EXPORT_SYMBOL(ptlrpc_save_lock);
209 struct ptlrpc_hr_partition;
211 struct ptlrpc_hr_thread {
212 int hrt_id; /* thread ID */
214 cfs_waitq_t hrt_waitq;
215 cfs_list_t hrt_queue; /* RS queue */
216 struct ptlrpc_hr_partition *hrt_partition;
219 struct ptlrpc_hr_partition {
220 /* # of started threads */
221 cfs_atomic_t hrp_nstarted;
222 /* # of stopped threads */
223 cfs_atomic_t hrp_nstopped;
224 /* cpu partition id */
226 /* round-robin rotor for choosing thread */
228 /* total number of threads on this partition */
231 struct ptlrpc_hr_thread *hrp_thrs;
234 #define HRT_RUNNING 0
235 #define HRT_STOPPING 1
237 struct ptlrpc_hr_service {
238 /* CPU partition table, it's just cfs_cpt_table for now */
239 struct cfs_cpt_table *hr_cpt_table;
240 /** controller sleep waitq */
241 cfs_waitq_t hr_waitq;
242 unsigned int hr_stopping;
243 /** roundrobin rotor for non-affinity service */
244 unsigned int hr_rotor;
246 struct ptlrpc_hr_partition **hr_partitions;
250 cfs_list_t rsb_replies;
251 unsigned int rsb_n_replies;
252 struct ptlrpc_service_part *rsb_svcpt;
255 /** reply handling service. */
256 static struct ptlrpc_hr_service ptlrpc_hr;
259 * maximum mumber of replies scheduled in one batch
261 #define MAX_SCHEDULED 256
264 * Initialize a reply batch.
268 static void rs_batch_init(struct rs_batch *b)
270 memset(b, 0, sizeof *b);
271 CFS_INIT_LIST_HEAD(&b->rsb_replies);
275 * Choose an hr thread to dispatch requests to.
277 static struct ptlrpc_hr_thread *
278 ptlrpc_hr_select(struct ptlrpc_service_part *svcpt)
280 struct ptlrpc_hr_partition *hrp;
283 if (svcpt->scp_cpt >= 0 &&
284 svcpt->scp_service->srv_cptable == ptlrpc_hr.hr_cpt_table) {
285 /* directly match partition */
286 hrp = ptlrpc_hr.hr_partitions[svcpt->scp_cpt];
289 rotor = ptlrpc_hr.hr_rotor++;
290 rotor %= cfs_cpt_number(ptlrpc_hr.hr_cpt_table);
292 hrp = ptlrpc_hr.hr_partitions[rotor];
295 rotor = hrp->hrp_rotor++;
296 return &hrp->hrp_thrs[rotor % hrp->hrp_nthrs];
300 * Dispatch all replies accumulated in the batch to one from
301 * dedicated reply handling threads.
305 static void rs_batch_dispatch(struct rs_batch *b)
307 if (b->rsb_n_replies != 0) {
308 struct ptlrpc_hr_thread *hrt;
310 hrt = ptlrpc_hr_select(b->rsb_svcpt);
312 spin_lock(&hrt->hrt_lock);
313 cfs_list_splice_init(&b->rsb_replies, &hrt->hrt_queue);
314 spin_unlock(&hrt->hrt_lock);
316 cfs_waitq_signal(&hrt->hrt_waitq);
317 b->rsb_n_replies = 0;
322 * Add a reply to a batch.
323 * Add one reply object to a batch, schedule batched replies if overload.
328 static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs)
330 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
332 if (svcpt != b->rsb_svcpt || b->rsb_n_replies >= MAX_SCHEDULED) {
333 if (b->rsb_svcpt != NULL) {
334 rs_batch_dispatch(b);
335 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
337 spin_lock(&svcpt->scp_rep_lock);
338 b->rsb_svcpt = svcpt;
340 spin_lock(&rs->rs_lock);
341 rs->rs_scheduled_ever = 1;
342 if (rs->rs_scheduled == 0) {
343 cfs_list_move(&rs->rs_list, &b->rsb_replies);
344 rs->rs_scheduled = 1;
347 rs->rs_committed = 1;
348 spin_unlock(&rs->rs_lock);
352 * Reply batch finalization.
353 * Dispatch remaining replies from the batch
354 * and release remaining spinlock.
358 static void rs_batch_fini(struct rs_batch *b)
360 if (b->rsb_svcpt != NULL) {
361 rs_batch_dispatch(b);
362 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
366 #define DECLARE_RS_BATCH(b) struct rs_batch b
368 #else /* __KERNEL__ */
370 #define rs_batch_init(b) do{}while(0)
371 #define rs_batch_fini(b) do{}while(0)
372 #define rs_batch_add(b, r) ptlrpc_schedule_difficult_reply(r)
373 #define DECLARE_RS_BATCH(b)
375 #endif /* __KERNEL__ */
378 * Put reply state into a queue for processing because we received
379 * ACK from the client
381 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
384 struct ptlrpc_hr_thread *hrt;
387 LASSERT(cfs_list_empty(&rs->rs_list));
389 hrt = ptlrpc_hr_select(rs->rs_svcpt);
391 spin_lock(&hrt->hrt_lock);
392 cfs_list_add_tail(&rs->rs_list, &hrt->hrt_queue);
393 spin_unlock(&hrt->hrt_lock);
395 cfs_waitq_signal(&hrt->hrt_waitq);
398 cfs_list_add_tail(&rs->rs_list, &rs->rs_svcpt->scp_rep_queue);
403 ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs)
407 LASSERT_SPIN_LOCKED(&rs->rs_svcpt->scp_rep_lock);
408 LASSERT_SPIN_LOCKED(&rs->rs_lock);
409 LASSERT (rs->rs_difficult);
410 rs->rs_scheduled_ever = 1; /* flag any notification attempt */
412 if (rs->rs_scheduled) { /* being set up or already notified */
417 rs->rs_scheduled = 1;
418 cfs_list_del_init(&rs->rs_list);
419 ptlrpc_dispatch_difficult_reply(rs);
422 EXPORT_SYMBOL(ptlrpc_schedule_difficult_reply);
424 void ptlrpc_commit_replies(struct obd_export *exp)
426 struct ptlrpc_reply_state *rs, *nxt;
427 DECLARE_RS_BATCH(batch);
430 rs_batch_init(&batch);
431 /* Find any replies that have been committed and get their service
432 * to attend to complete them. */
434 /* CAVEAT EMPTOR: spinlock ordering!!! */
435 spin_lock(&exp->exp_uncommitted_replies_lock);
436 cfs_list_for_each_entry_safe(rs, nxt, &exp->exp_uncommitted_replies,
438 LASSERT (rs->rs_difficult);
439 /* VBR: per-export last_committed */
440 LASSERT(rs->rs_export);
441 if (rs->rs_transno <= exp->exp_last_committed) {
442 cfs_list_del_init(&rs->rs_obd_list);
443 rs_batch_add(&batch, rs);
446 spin_unlock(&exp->exp_uncommitted_replies_lock);
447 rs_batch_fini(&batch);
450 EXPORT_SYMBOL(ptlrpc_commit_replies);
453 ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt)
455 struct ptlrpc_request_buffer_desc *rqbd;
460 spin_lock(&svcpt->scp_lock);
462 if (cfs_list_empty(&svcpt->scp_rqbd_idle)) {
463 spin_unlock(&svcpt->scp_lock);
467 rqbd = cfs_list_entry(svcpt->scp_rqbd_idle.next,
468 struct ptlrpc_request_buffer_desc,
470 cfs_list_del(&rqbd->rqbd_list);
472 /* assume we will post successfully */
473 svcpt->scp_nrqbds_posted++;
474 cfs_list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_posted);
476 spin_unlock(&svcpt->scp_lock);
478 rc = ptlrpc_register_rqbd(rqbd);
485 spin_lock(&svcpt->scp_lock);
487 svcpt->scp_nrqbds_posted--;
488 cfs_list_del(&rqbd->rqbd_list);
489 cfs_list_add_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
491 /* Don't complain if no request buffers are posted right now; LNET
492 * won't drop requests because we set the portal lazy! */
494 spin_unlock(&svcpt->scp_lock);
499 static void ptlrpc_at_timer(unsigned long castmeharder)
501 struct ptlrpc_service_part *svcpt;
503 svcpt = (struct ptlrpc_service_part *)castmeharder;
505 svcpt->scp_at_check = 1;
506 svcpt->scp_at_checktime = cfs_time_current();
507 cfs_waitq_signal(&svcpt->scp_waitq);
511 ptlrpc_server_nthreads_check(struct ptlrpc_service *svc,
512 struct ptlrpc_service_conf *conf)
515 struct ptlrpc_service_thr_conf *tc = &conf->psc_thr;
522 * Common code for estimating & validating threads number.
523 * CPT affinity service could have percpt thread-pool instead
524 * of a global thread-pool, which means user might not always
525 * get the threads number they give it in conf::tc_nthrs_user
526 * even they did set. It's because we need to validate threads
527 * number for each CPT to guarantee each pool will have enough
528 * threads to keep the service healthy.
530 init = PTLRPC_NTHRS_INIT + (svc->srv_ops.so_hpreq_handler != NULL);
531 init = max_t(int, init, tc->tc_nthrs_init);
533 /* NB: please see comments in lustre_lnet.h for definition
534 * details of these members */
535 LASSERT(tc->tc_nthrs_max != 0);
537 if (tc->tc_nthrs_user != 0) {
538 /* In case there is a reason to test a service with many
539 * threads, we give a less strict check here, it can
540 * be up to 8 * nthrs_max */
541 total = min(tc->tc_nthrs_max * 8, tc->tc_nthrs_user);
542 nthrs = total / svc->srv_ncpts;
543 init = max(init, nthrs);
547 total = tc->tc_nthrs_max;
548 if (tc->tc_nthrs_base == 0) {
549 /* don't care about base threads number per partition,
550 * this is most for non-affinity service */
551 nthrs = total / svc->srv_ncpts;
555 nthrs = tc->tc_nthrs_base;
556 if (svc->srv_ncpts == 1) {
559 /* NB: Increase the base number if it's single partition
560 * and total number of cores/HTs is larger or equal to 4.
561 * result will always < 2 * nthrs_base */
562 weight = cfs_cpt_weight(svc->srv_cptable, CFS_CPT_ANY);
563 for (i = 1; (weight >> (i + 1)) != 0 && /* >= 4 cores/HTs */
564 (tc->tc_nthrs_base >> i) != 0; i++)
565 nthrs += tc->tc_nthrs_base >> i;
568 if (tc->tc_thr_factor != 0) {
569 int factor = tc->tc_thr_factor;
573 * User wants to increase number of threads with for
574 * each CPU core/HT, most likely the factor is larger then
575 * one thread/core because service threads are supposed to
576 * be blocked by lock or wait for IO.
579 * Amdahl's law says that adding processors wouldn't give
580 * a linear increasing of parallelism, so it's nonsense to
581 * have too many threads no matter how many cores/HTs
584 if (cfs_cpu_ht_nsiblings(0) > 1) { /* weight is # of HTs */
585 /* depress thread factor for hyper-thread */
586 factor = factor - (factor >> 1) + (factor >> 3);
589 weight = cfs_cpt_weight(svc->srv_cptable, 0);
592 for (; factor > 0 && weight > 0; factor--, weight -= fade)
593 nthrs += min(weight, fade) * factor;
596 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
597 nthrs = max(tc->tc_nthrs_base,
598 tc->tc_nthrs_max / svc->srv_ncpts);
601 nthrs = max(nthrs, tc->tc_nthrs_init);
602 svc->srv_nthrs_cpt_limit = nthrs;
603 svc->srv_nthrs_cpt_init = init;
605 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
606 LCONSOLE_WARN("%s: This service may have more threads (%d) "
607 "than the given soft limit (%d)\n",
608 svc->srv_name, nthrs * svc->srv_ncpts,
615 * Initialize percpt data for a service
618 ptlrpc_service_part_init(struct ptlrpc_service *svc,
619 struct ptlrpc_service_part *svcpt, int cpt)
621 struct ptlrpc_at_array *array;
626 svcpt->scp_cpt = cpt;
627 CFS_INIT_LIST_HEAD(&svcpt->scp_threads);
629 /* rqbd and incoming request queue */
630 spin_lock_init(&svcpt->scp_lock);
631 CFS_INIT_LIST_HEAD(&svcpt->scp_rqbd_idle);
632 CFS_INIT_LIST_HEAD(&svcpt->scp_rqbd_posted);
633 CFS_INIT_LIST_HEAD(&svcpt->scp_req_incoming);
634 cfs_waitq_init(&svcpt->scp_waitq);
635 /* history request & rqbd list */
636 CFS_INIT_LIST_HEAD(&svcpt->scp_hist_reqs);
637 CFS_INIT_LIST_HEAD(&svcpt->scp_hist_rqbds);
639 /* acitve requests and hp requests */
640 spin_lock_init(&svcpt->scp_req_lock);
643 spin_lock_init(&svcpt->scp_rep_lock);
644 CFS_INIT_LIST_HEAD(&svcpt->scp_rep_active);
646 CFS_INIT_LIST_HEAD(&svcpt->scp_rep_queue);
648 CFS_INIT_LIST_HEAD(&svcpt->scp_rep_idle);
649 cfs_waitq_init(&svcpt->scp_rep_waitq);
650 cfs_atomic_set(&svcpt->scp_nreps_difficult, 0);
652 /* adaptive timeout */
653 spin_lock_init(&svcpt->scp_at_lock);
654 array = &svcpt->scp_at_array;
656 size = at_est2timeout(at_max);
657 array->paa_size = size;
658 array->paa_count = 0;
659 array->paa_deadline = -1;
661 /* allocate memory for scp_at_array (ptlrpc_at_array) */
662 OBD_CPT_ALLOC(array->paa_reqs_array,
663 svc->srv_cptable, cpt, sizeof(cfs_list_t) * size);
664 if (array->paa_reqs_array == NULL)
667 for (index = 0; index < size; index++)
668 CFS_INIT_LIST_HEAD(&array->paa_reqs_array[index]);
670 OBD_CPT_ALLOC(array->paa_reqs_count,
671 svc->srv_cptable, cpt, sizeof(__u32) * size);
672 if (array->paa_reqs_count == NULL)
675 cfs_timer_init(&svcpt->scp_at_timer, ptlrpc_at_timer, svcpt);
676 /* At SOW, service time should be quick; 10s seems generous. If client
677 * timeout is less than this, we'll be sending an early reply. */
678 at_init(&svcpt->scp_at_estimate, 10, 0);
680 /* assign this before call ptlrpc_grow_req_bufs */
681 svcpt->scp_service = svc;
682 /* Now allocate the request buffers, but don't post them now */
683 rc = ptlrpc_grow_req_bufs(svcpt, 0);
684 /* We shouldn't be under memory pressure at startup, so
685 * fail if we can't allocate all our buffers at this time. */
692 if (array->paa_reqs_count != NULL) {
693 OBD_FREE(array->paa_reqs_count, sizeof(__u32) * size);
694 array->paa_reqs_count = NULL;
697 if (array->paa_reqs_array != NULL) {
698 OBD_FREE(array->paa_reqs_array,
699 sizeof(cfs_list_t) * array->paa_size);
700 array->paa_reqs_array = NULL;
707 * Initialize service on a given portal.
708 * This includes starting serving threads , allocating and posting rqbds and
711 struct ptlrpc_service *
712 ptlrpc_register_service(struct ptlrpc_service_conf *conf,
713 cfs_proc_dir_entry_t *proc_entry)
715 struct ptlrpc_service_cpt_conf *cconf = &conf->psc_cpt;
716 struct ptlrpc_service *service;
717 struct ptlrpc_service_part *svcpt;
718 struct cfs_cpt_table *cptable;
726 LASSERT(conf->psc_buf.bc_nbufs > 0);
727 LASSERT(conf->psc_buf.bc_buf_size >=
728 conf->psc_buf.bc_req_max_size + SPTLRPC_MAX_PAYLOAD);
729 LASSERT(conf->psc_thr.tc_ctx_tags != 0);
731 cptable = cconf->cc_cptable;
733 cptable = cfs_cpt_table;
735 if (!conf->psc_thr.tc_cpu_affinity) {
738 ncpts = cfs_cpt_number(cptable);
739 if (cconf->cc_pattern != NULL) {
740 struct cfs_expr_list *el;
742 rc = cfs_expr_list_parse(cconf->cc_pattern,
743 strlen(cconf->cc_pattern),
746 CERROR("%s: invalid CPT pattern string: %s",
747 conf->psc_name, cconf->cc_pattern);
748 RETURN(ERR_PTR(-EINVAL));
751 rc = cfs_expr_list_values(el, ncpts, &cpts);
752 cfs_expr_list_free(el);
754 CERROR("%s: failed to parse CPT array %s: %d\n",
755 conf->psc_name, cconf->cc_pattern, rc);
757 OBD_FREE(cpts, sizeof(*cpts) * ncpts);
758 RETURN(ERR_PTR(rc < 0 ? rc : -EINVAL));
764 OBD_ALLOC(service, offsetof(struct ptlrpc_service, srv_parts[ncpts]));
765 if (service == NULL) {
767 OBD_FREE(cpts, sizeof(*cpts) * ncpts);
768 RETURN(ERR_PTR(-ENOMEM));
771 service->srv_cptable = cptable;
772 service->srv_cpts = cpts;
773 service->srv_ncpts = ncpts;
775 service->srv_cpt_bits = 0; /* it's zero already, easy to read... */
776 while ((1 << service->srv_cpt_bits) < cfs_cpt_number(cptable))
777 service->srv_cpt_bits++;
780 spin_lock_init(&service->srv_lock);
781 service->srv_name = conf->psc_name;
782 service->srv_watchdog_factor = conf->psc_watchdog_factor;
783 CFS_INIT_LIST_HEAD(&service->srv_list); /* for safty of cleanup */
785 /* buffer configuration */
786 service->srv_nbuf_per_group = test_req_buffer_pressure ?
787 1 : conf->psc_buf.bc_nbufs;
788 service->srv_max_req_size = conf->psc_buf.bc_req_max_size +
790 service->srv_buf_size = conf->psc_buf.bc_buf_size;
791 service->srv_rep_portal = conf->psc_buf.bc_rep_portal;
792 service->srv_req_portal = conf->psc_buf.bc_req_portal;
794 /* Increase max reply size to next power of two */
795 service->srv_max_reply_size = 1;
796 while (service->srv_max_reply_size <
797 conf->psc_buf.bc_rep_max_size + SPTLRPC_MAX_PAYLOAD)
798 service->srv_max_reply_size <<= 1;
800 service->srv_thread_name = conf->psc_thr.tc_thr_name;
801 service->srv_ctx_tags = conf->psc_thr.tc_ctx_tags;
802 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
803 service->srv_ops = conf->psc_ops;
805 for (i = 0; i < ncpts; i++) {
806 if (!conf->psc_thr.tc_cpu_affinity)
809 cpt = cpts != NULL ? cpts[i] : i;
811 OBD_CPT_ALLOC(svcpt, cptable, cpt, sizeof(*svcpt));
813 GOTO(failed, rc = -ENOMEM);
815 service->srv_parts[i] = svcpt;
816 rc = ptlrpc_service_part_init(service, svcpt, cpt);
821 ptlrpc_server_nthreads_check(service, conf);
823 rc = LNetSetLazyPortal(service->srv_req_portal);
826 mutex_lock(&ptlrpc_all_services_mutex);
827 cfs_list_add (&service->srv_list, &ptlrpc_all_services);
828 mutex_unlock(&ptlrpc_all_services_mutex);
830 if (proc_entry != NULL)
831 ptlrpc_lprocfs_register_service(proc_entry, service);
833 rc = ptlrpc_service_nrs_setup(service);
837 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
838 service->srv_name, service->srv_req_portal);
841 rc = ptlrpc_start_threads(service);
843 CERROR("Failed to start threads for service %s: %d\n",
844 service->srv_name, rc);
851 ptlrpc_unregister_service(service);
854 EXPORT_SYMBOL(ptlrpc_register_service);
857 * to actually free the request, must be called without holding svc_lock.
858 * note it's caller's responsibility to unlink req->rq_list.
860 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
862 LASSERT(cfs_atomic_read(&req->rq_refcount) == 0);
863 LASSERT(cfs_list_empty(&req->rq_timed_list));
865 /* DEBUG_REQ() assumes the reply state of a request with a valid
866 * ref will not be destroyed until that reference is dropped. */
867 ptlrpc_req_drop_rs(req);
869 sptlrpc_svc_ctx_decref(req);
871 if (req != &req->rq_rqbd->rqbd_req) {
872 /* NB request buffers use an embedded
873 * req if the incoming req unlinked the
874 * MD; this isn't one of them! */
875 OBD_FREE(req, sizeof(*req));
880 * drop a reference count of the request. if it reaches 0, we either
881 * put it into history list, or free it immediately.
883 void ptlrpc_server_drop_request(struct ptlrpc_request *req)
885 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
886 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
887 struct ptlrpc_service *svc = svcpt->scp_service;
892 if (!cfs_atomic_dec_and_test(&req->rq_refcount))
895 if (req->rq_at_linked) {
896 spin_lock(&svcpt->scp_at_lock);
897 /* recheck with lock, in case it's unlinked by
898 * ptlrpc_at_check_timed() */
899 if (likely(req->rq_at_linked))
900 ptlrpc_at_remove_timed(req);
901 spin_unlock(&svcpt->scp_at_lock);
904 LASSERT(cfs_list_empty(&req->rq_timed_list));
906 /* finalize request */
907 if (req->rq_export) {
908 class_export_put(req->rq_export);
909 req->rq_export = NULL;
912 spin_lock(&svcpt->scp_lock);
914 cfs_list_add(&req->rq_list, &rqbd->rqbd_reqs);
916 refcount = --(rqbd->rqbd_refcount);
918 /* request buffer is now idle: add to history */
919 cfs_list_del(&rqbd->rqbd_list);
921 cfs_list_add_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds);
922 svcpt->scp_hist_nrqbds++;
924 /* cull some history?
925 * I expect only about 1 or 2 rqbds need to be recycled here */
926 while (svcpt->scp_hist_nrqbds > svc->srv_hist_nrqbds_cpt_max) {
927 rqbd = cfs_list_entry(svcpt->scp_hist_rqbds.next,
928 struct ptlrpc_request_buffer_desc,
931 cfs_list_del(&rqbd->rqbd_list);
932 svcpt->scp_hist_nrqbds--;
934 /* remove rqbd's reqs from svc's req history while
935 * I've got the service lock */
936 cfs_list_for_each(tmp, &rqbd->rqbd_reqs) {
937 req = cfs_list_entry(tmp, struct ptlrpc_request,
939 /* Track the highest culled req seq */
940 if (req->rq_history_seq >
941 svcpt->scp_hist_seq_culled) {
942 svcpt->scp_hist_seq_culled =
945 cfs_list_del(&req->rq_history_list);
948 spin_unlock(&svcpt->scp_lock);
950 cfs_list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
951 req = cfs_list_entry(rqbd->rqbd_reqs.next,
952 struct ptlrpc_request,
954 cfs_list_del(&req->rq_list);
955 ptlrpc_server_free_request(req);
958 spin_lock(&svcpt->scp_lock);
960 * now all reqs including the embedded req has been
961 * disposed, schedule request buffer for re-use.
963 LASSERT(cfs_atomic_read(&rqbd->rqbd_req.rq_refcount) ==
965 cfs_list_add_tail(&rqbd->rqbd_list,
966 &svcpt->scp_rqbd_idle);
969 spin_unlock(&svcpt->scp_lock);
970 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
971 /* If we are low on memory, we are not interested in history */
972 cfs_list_del(&req->rq_list);
973 cfs_list_del_init(&req->rq_history_list);
975 /* Track the highest culled req seq */
976 if (req->rq_history_seq > svcpt->scp_hist_seq_culled)
977 svcpt->scp_hist_seq_culled = req->rq_history_seq;
979 spin_unlock(&svcpt->scp_lock);
981 ptlrpc_server_free_request(req);
983 spin_unlock(&svcpt->scp_lock);
988 * to finish a request: stop sending more early replies, and release
989 * the request. should be called after we finished handling the request.
991 static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt,
992 struct ptlrpc_request *req)
994 ptlrpc_server_hpreq_fini(req);
996 spin_lock(&svcpt->scp_req_lock);
997 ptlrpc_nrs_req_stop_nolock(req);
998 svcpt->scp_nreqs_active--;
1000 svcpt->scp_nhreqs_active--;
1001 spin_unlock(&svcpt->scp_req_lock);
1003 ptlrpc_nrs_req_finalize(req);
1005 ptlrpc_server_drop_request(req);
1009 * This function makes sure dead exports are evicted in a timely manner.
1010 * This function is only called when some export receives a message (i.e.,
1011 * the network is up.)
1013 static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
1015 struct obd_export *oldest_exp;
1016 time_t oldest_time, new_time;
1022 /* Compensate for slow machines, etc, by faking our request time
1023 into the future. Although this can break the strict time-ordering
1024 of the list, we can be really lazy here - we don't have to evict
1025 at the exact right moment. Eventually, all silent exports
1026 will make it to the top of the list. */
1028 /* Do not pay attention on 1sec or smaller renewals. */
1029 new_time = cfs_time_current_sec() + extra_delay;
1030 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
1033 exp->exp_last_request_time = new_time;
1034 CDEBUG(D_HA, "updating export %s at "CFS_TIME_T" exp %p\n",
1035 exp->exp_client_uuid.uuid,
1036 exp->exp_last_request_time, exp);
1038 /* exports may get disconnected from the chain even though the
1039 export has references, so we must keep the spin lock while
1040 manipulating the lists */
1041 spin_lock(&exp->exp_obd->obd_dev_lock);
1043 if (cfs_list_empty(&exp->exp_obd_chain_timed)) {
1044 /* this one is not timed */
1045 spin_unlock(&exp->exp_obd->obd_dev_lock);
1049 cfs_list_move_tail(&exp->exp_obd_chain_timed,
1050 &exp->exp_obd->obd_exports_timed);
1052 oldest_exp = cfs_list_entry(exp->exp_obd->obd_exports_timed.next,
1053 struct obd_export, exp_obd_chain_timed);
1054 oldest_time = oldest_exp->exp_last_request_time;
1055 spin_unlock(&exp->exp_obd->obd_dev_lock);
1057 if (exp->exp_obd->obd_recovering) {
1058 /* be nice to everyone during recovery */
1063 /* Note - racing to start/reset the obd_eviction timer is safe */
1064 if (exp->exp_obd->obd_eviction_timer == 0) {
1065 /* Check if the oldest entry is expired. */
1066 if (cfs_time_current_sec() > (oldest_time + PING_EVICT_TIMEOUT +
1068 /* We need a second timer, in case the net was down and
1069 * it just came back. Since the pinger may skip every
1070 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
1071 * we better wait for 3. */
1072 exp->exp_obd->obd_eviction_timer =
1073 cfs_time_current_sec() + 3 * PING_INTERVAL;
1074 CDEBUG(D_HA, "%s: Think about evicting %s from "CFS_TIME_T"\n",
1075 exp->exp_obd->obd_name,
1076 obd_export_nid2str(oldest_exp), oldest_time);
1079 if (cfs_time_current_sec() >
1080 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
1081 /* The evictor won't evict anyone who we've heard from
1082 * recently, so we don't have to check before we start
1084 if (!ping_evictor_wake(exp))
1085 exp->exp_obd->obd_eviction_timer = 0;
1093 * Sanity check request \a req.
1094 * Return 0 if all is ok, error code otherwise.
1096 static int ptlrpc_check_req(struct ptlrpc_request *req)
1100 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
1101 req->rq_export->exp_conn_cnt)) {
1102 DEBUG_REQ(D_RPCTRACE, req,
1103 "DROPPING req from old connection %d < %d",
1104 lustre_msg_get_conn_cnt(req->rq_reqmsg),
1105 req->rq_export->exp_conn_cnt);
1108 if (unlikely(req->rq_export->exp_obd &&
1109 req->rq_export->exp_obd->obd_fail)) {
1110 /* Failing over, don't handle any more reqs, send
1111 error response instead. */
1112 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
1113 req, req->rq_export->exp_obd->obd_name);
1115 } else if (lustre_msg_get_flags(req->rq_reqmsg) &
1116 (MSG_REPLAY | MSG_REQ_REPLAY_DONE) &&
1117 !(req->rq_export->exp_obd->obd_recovering)) {
1118 DEBUG_REQ(D_ERROR, req,
1119 "Invalid replay without recovery");
1120 class_fail_export(req->rq_export);
1122 } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 &&
1123 !(req->rq_export->exp_obd->obd_recovering)) {
1124 DEBUG_REQ(D_ERROR, req, "Invalid req with transno "
1125 LPU64" without recovery",
1126 lustre_msg_get_transno(req->rq_reqmsg));
1127 class_fail_export(req->rq_export);
1131 if (unlikely(rc < 0)) {
1132 req->rq_status = rc;
1138 static void ptlrpc_at_set_timer(struct ptlrpc_service_part *svcpt)
1140 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1143 if (array->paa_count == 0) {
1144 cfs_timer_disarm(&svcpt->scp_at_timer);
1148 /* Set timer for closest deadline */
1149 next = (__s32)(array->paa_deadline - cfs_time_current_sec() -
1152 ptlrpc_at_timer((unsigned long)svcpt);
1154 cfs_timer_arm(&svcpt->scp_at_timer, cfs_time_shift(next));
1155 CDEBUG(D_INFO, "armed %s at %+ds\n",
1156 svcpt->scp_service->srv_name, next);
1160 /* Add rpc to early reply check list */
1161 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
1163 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1164 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1165 struct ptlrpc_request *rq = NULL;
1171 if (req->rq_no_reply)
1174 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
1177 spin_lock(&svcpt->scp_at_lock);
1178 LASSERT(cfs_list_empty(&req->rq_timed_list));
1180 index = (unsigned long)req->rq_deadline % array->paa_size;
1181 if (array->paa_reqs_count[index] > 0) {
1182 /* latest rpcs will have the latest deadlines in the list,
1183 * so search backward. */
1184 cfs_list_for_each_entry_reverse(rq,
1185 &array->paa_reqs_array[index],
1187 if (req->rq_deadline >= rq->rq_deadline) {
1188 cfs_list_add(&req->rq_timed_list,
1189 &rq->rq_timed_list);
1195 /* Add the request at the head of the list */
1196 if (cfs_list_empty(&req->rq_timed_list))
1197 cfs_list_add(&req->rq_timed_list,
1198 &array->paa_reqs_array[index]);
1200 spin_lock(&req->rq_lock);
1201 req->rq_at_linked = 1;
1202 spin_unlock(&req->rq_lock);
1203 req->rq_at_index = index;
1204 array->paa_reqs_count[index]++;
1206 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
1207 array->paa_deadline = req->rq_deadline;
1208 ptlrpc_at_set_timer(svcpt);
1210 spin_unlock(&svcpt->scp_at_lock);
1216 ptlrpc_at_remove_timed(struct ptlrpc_request *req)
1218 struct ptlrpc_at_array *array;
1220 array = &req->rq_rqbd->rqbd_svcpt->scp_at_array;
1222 /* NB: must call with hold svcpt::scp_at_lock */
1223 LASSERT(!cfs_list_empty(&req->rq_timed_list));
1224 cfs_list_del_init(&req->rq_timed_list);
1226 spin_lock(&req->rq_lock);
1227 req->rq_at_linked = 0;
1228 spin_unlock(&req->rq_lock);
1230 array->paa_reqs_count[req->rq_at_index]--;
1234 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
1236 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1237 struct ptlrpc_request *reqcopy;
1238 struct lustre_msg *reqmsg;
1239 cfs_duration_t olddl = req->rq_deadline - cfs_time_current_sec();
1244 /* deadline is when the client expects us to reply, margin is the
1245 difference between clients' and servers' expectations */
1246 DEBUG_REQ(D_ADAPTTO, req,
1247 "%ssending early reply (deadline %+lds, margin %+lds) for "
1248 "%d+%d", AT_OFF ? "AT off - not " : "",
1249 olddl, olddl - at_get(&svcpt->scp_at_estimate),
1250 at_get(&svcpt->scp_at_estimate), at_extra);
1256 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), "
1257 "not sending early reply. Consider increasing "
1258 "at_early_margin (%d)?", olddl, at_early_margin);
1260 /* Return an error so we're not re-added to the timed list. */
1264 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
1265 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
1266 "but no AT support");
1270 if (req->rq_export &&
1271 lustre_msg_get_flags(req->rq_reqmsg) &
1272 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1273 /* During recovery, we don't want to send too many early
1274 * replies, but on the other hand we want to make sure the
1275 * client has enough time to resend if the rpc is lost. So
1276 * during the recovery period send at least 4 early replies,
1277 * spacing them every at_extra if we can. at_estimate should
1278 * always equal this fixed value during recovery. */
1279 at_measured(&svcpt->scp_at_estimate, min(at_extra,
1280 req->rq_export->exp_obd->obd_recovery_timeout / 4));
1282 /* Fake our processing time into the future to ask the clients
1283 * for some extra amount of time */
1284 at_measured(&svcpt->scp_at_estimate, at_extra +
1285 cfs_time_current_sec() -
1286 req->rq_arrival_time.tv_sec);
1288 /* Check to see if we've actually increased the deadline -
1289 * we may be past adaptive_max */
1290 if (req->rq_deadline >= req->rq_arrival_time.tv_sec +
1291 at_get(&svcpt->scp_at_estimate)) {
1292 DEBUG_REQ(D_WARNING, req, "Couldn't add any time "
1293 "(%ld/%ld), not sending early reply\n",
1294 olddl, req->rq_arrival_time.tv_sec +
1295 at_get(&svcpt->scp_at_estimate) -
1296 cfs_time_current_sec());
1300 newdl = cfs_time_current_sec() + at_get(&svcpt->scp_at_estimate);
1302 OBD_ALLOC(reqcopy, sizeof *reqcopy);
1303 if (reqcopy == NULL)
1305 OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1307 OBD_FREE(reqcopy, sizeof *reqcopy);
1312 reqcopy->rq_reply_state = NULL;
1313 reqcopy->rq_rep_swab_mask = 0;
1314 reqcopy->rq_pack_bulk = 0;
1315 reqcopy->rq_pack_udesc = 0;
1316 reqcopy->rq_packed_final = 0;
1317 sptlrpc_svc_ctx_addref(reqcopy);
1318 /* We only need the reqmsg for the magic */
1319 reqcopy->rq_reqmsg = reqmsg;
1320 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1322 LASSERT(cfs_atomic_read(&req->rq_refcount));
1323 /** if it is last refcount then early reply isn't needed */
1324 if (cfs_atomic_read(&req->rq_refcount) == 1) {
1325 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
1326 "abort sending early reply\n");
1327 GOTO(out, rc = -EINVAL);
1330 /* Connection ref */
1331 reqcopy->rq_export = class_conn2export(
1332 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1333 if (reqcopy->rq_export == NULL)
1334 GOTO(out, rc = -ENODEV);
1337 class_export_rpc_get(reqcopy->rq_export);
1338 if (reqcopy->rq_export->exp_obd &&
1339 reqcopy->rq_export->exp_obd->obd_fail)
1340 GOTO(out_put, rc = -ENODEV);
1342 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1346 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1349 /* Adjust our own deadline to what we told the client */
1350 req->rq_deadline = newdl;
1351 req->rq_early_count++; /* number sent, server side */
1353 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1356 /* Free the (early) reply state from lustre_pack_reply.
1357 (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
1358 ptlrpc_req_drop_rs(reqcopy);
1361 class_export_rpc_put(reqcopy->rq_export);
1362 class_export_put(reqcopy->rq_export);
1364 sptlrpc_svc_ctx_decref(reqcopy);
1365 OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1366 OBD_FREE(reqcopy, sizeof *reqcopy);
1370 /* Send early replies to everybody expiring within at_early_margin
1371 asking for at_extra time */
1372 static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1374 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1375 struct ptlrpc_request *rq, *n;
1376 cfs_list_t work_list;
1379 time_t now = cfs_time_current_sec();
1380 cfs_duration_t delay;
1381 int first, counter = 0;
1384 spin_lock(&svcpt->scp_at_lock);
1385 if (svcpt->scp_at_check == 0) {
1386 spin_unlock(&svcpt->scp_at_lock);
1389 delay = cfs_time_sub(cfs_time_current(), svcpt->scp_at_checktime);
1390 svcpt->scp_at_check = 0;
1392 if (array->paa_count == 0) {
1393 spin_unlock(&svcpt->scp_at_lock);
1397 /* The timer went off, but maybe the nearest rpc already completed. */
1398 first = array->paa_deadline - now;
1399 if (first > at_early_margin) {
1400 /* We've still got plenty of time. Reset the timer. */
1401 ptlrpc_at_set_timer(svcpt);
1402 spin_unlock(&svcpt->scp_at_lock);
1406 /* We're close to a timeout, and we don't know how much longer the
1407 server will take. Send early replies to everyone expiring soon. */
1408 CFS_INIT_LIST_HEAD(&work_list);
1410 index = (unsigned long)array->paa_deadline % array->paa_size;
1411 count = array->paa_count;
1413 count -= array->paa_reqs_count[index];
1414 cfs_list_for_each_entry_safe(rq, n,
1415 &array->paa_reqs_array[index],
1417 if (rq->rq_deadline > now + at_early_margin) {
1418 /* update the earliest deadline */
1419 if (deadline == -1 ||
1420 rq->rq_deadline < deadline)
1421 deadline = rq->rq_deadline;
1425 ptlrpc_at_remove_timed(rq);
1427 * ptlrpc_server_drop_request() may drop
1428 * refcount to 0 already. Let's check this and
1429 * don't add entry to work_list
1431 if (likely(cfs_atomic_inc_not_zero(&rq->rq_refcount)))
1432 cfs_list_add(&rq->rq_timed_list, &work_list);
1436 if (++index >= array->paa_size)
1439 array->paa_deadline = deadline;
1440 /* we have a new earliest deadline, restart the timer */
1441 ptlrpc_at_set_timer(svcpt);
1443 spin_unlock(&svcpt->scp_at_lock);
1445 CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
1446 "replies\n", first, at_extra, counter);
1448 /* We're already past request deadlines before we even get a
1449 chance to send early replies */
1450 LCONSOLE_WARN("%s: This server is not able to keep up with "
1451 "request traffic (cpu-bound).\n",
1452 svcpt->scp_service->srv_name);
1453 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, "
1454 "delay="CFS_DURATION_T"(jiff)\n",
1455 counter, svcpt->scp_nreqs_incoming,
1456 svcpt->scp_nreqs_active,
1457 at_get(&svcpt->scp_at_estimate), delay);
1460 /* we took additional refcount so entries can't be deleted from list, no
1461 * locking is needed */
1462 while (!cfs_list_empty(&work_list)) {
1463 rq = cfs_list_entry(work_list.next, struct ptlrpc_request,
1465 cfs_list_del_init(&rq->rq_timed_list);
1467 if (ptlrpc_at_send_early_reply(rq) == 0)
1468 ptlrpc_at_add_timed(rq);
1470 ptlrpc_server_drop_request(rq);
1473 RETURN(1); /* return "did_something" for liblustre */
1477 * Put the request to the export list if the request may become
1478 * a high priority one.
1480 static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt,
1481 struct ptlrpc_request *req)
1486 if (svcpt->scp_service->srv_ops.so_hpreq_handler) {
1487 rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
1492 if (req->rq_export && req->rq_ops) {
1493 /* Perform request specific check. We should do this check
1494 * before the request is added into exp_hp_rpcs list otherwise
1495 * it may hit swab race at LU-1044. */
1496 if (req->rq_ops->hpreq_check) {
1497 rc = req->rq_ops->hpreq_check(req);
1499 * XXX: Out of all current
1500 * ptlrpc_hpreq_ops::hpreq_check(), only
1501 * ldlm_cancel_hpreq_check() can return an error code;
1502 * other functions assert in similar places, which seems
1503 * odd. What also does not seem right is that handlers
1504 * for those RPCs do not assert on the same checks, but
1505 * rather handle the error cases. e.g. see
1506 * ost_rw_hpreq_check(), and ost_brw_read(),
1511 LASSERT(rc == 0 || rc == 1);
1514 spin_lock_bh(&req->rq_export->exp_rpc_lock);
1515 cfs_list_add(&req->rq_exp_list,
1516 &req->rq_export->exp_hp_rpcs);
1517 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1520 ptlrpc_nrs_req_initialize(svcpt, req, rc);
1525 /** Remove the request from the export list. */
1526 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req)
1529 if (req->rq_export && req->rq_ops) {
1530 /* refresh lock timeout again so that client has more
1531 * room to send lock cancel RPC. */
1532 if (req->rq_ops->hpreq_fini)
1533 req->rq_ops->hpreq_fini(req);
1535 spin_lock_bh(&req->rq_export->exp_rpc_lock);
1536 cfs_list_del_init(&req->rq_exp_list);
1537 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1542 static int ptlrpc_hpreq_check(struct ptlrpc_request *req)
1547 static struct ptlrpc_hpreq_ops ptlrpc_hpreq_common = {
1548 .hpreq_check = ptlrpc_hpreq_check,
1551 /* Hi-Priority RPC check by RPC operation code. */
1552 int ptlrpc_hpreq_handler(struct ptlrpc_request *req)
1554 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1556 /* Check for export to let only reconnects for not yet evicted
1557 * export to become a HP rpc. */
1558 if ((req->rq_export != NULL) &&
1559 (opc == OBD_PING || opc == MDS_CONNECT || opc == OST_CONNECT))
1560 req->rq_ops = &ptlrpc_hpreq_common;
1564 EXPORT_SYMBOL(ptlrpc_hpreq_handler);
1566 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1567 struct ptlrpc_request *req)
1572 rc = ptlrpc_server_hpreq_init(svcpt, req);
1576 ptlrpc_nrs_req_add(svcpt, req, !!rc);
1582 * Allow to handle high priority request
1583 * User can call it w/o any lock but need to hold
1584 * ptlrpc_service_part::scp_req_lock to get reliable result
1586 static int ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1589 int running = svcpt->scp_nthrs_running;
1591 if (!nrs_svcpt_has_hp(svcpt))
1597 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1598 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1599 /* leave just 1 thread for normal RPCs */
1600 running = PTLRPC_NTHRS_INIT;
1601 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1605 if (svcpt->scp_nreqs_active >= running - 1)
1608 if (svcpt->scp_nhreqs_active == 0)
1611 return !ptlrpc_nrs_req_pending_nolock(svcpt, false) ||
1612 svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1615 static int ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1618 return ptlrpc_server_allow_high(svcpt, force) &&
1619 ptlrpc_nrs_req_pending_nolock(svcpt, true);
1623 * Only allow normal priority requests on a service that has a high-priority
1624 * queue if forced (i.e. cleanup), if there are other high priority requests
1625 * already being processed (i.e. those threads can service more high-priority
1626 * requests), or if there are enough idle threads that a later thread can do
1627 * a high priority request.
1628 * User can call it w/o any lock but need to hold
1629 * ptlrpc_service_part::scp_req_lock to get reliable result
1631 static int ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1634 int running = svcpt->scp_nthrs_running;
1636 if (1) /* always allow to handle normal request for liblustre */
1639 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1640 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1641 /* leave just 1 thread for normal RPCs */
1642 running = PTLRPC_NTHRS_INIT;
1643 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1648 svcpt->scp_nreqs_active < running - 2)
1651 if (svcpt->scp_nreqs_active >= running - 1)
1654 return svcpt->scp_nhreqs_active > 0 ||
1655 !nrs_svcpt_has_hp(svcpt);
1658 static int ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1661 return ptlrpc_server_allow_normal(svcpt, force) &&
1662 ptlrpc_nrs_req_pending_nolock(svcpt, false);
1666 * Returns true if there are requests available in incoming
1667 * request queue for processing and it is allowed to fetch them.
1668 * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1669 * to get reliable result
1670 * \see ptlrpc_server_allow_normal
1671 * \see ptlrpc_server_allow high
1674 ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt, int force)
1676 return ptlrpc_server_high_pending(svcpt, force) ||
1677 ptlrpc_server_normal_pending(svcpt, force);
1681 * Fetch a request for processing from queue of unprocessed requests.
1682 * Favors high-priority requests.
1683 * Returns a pointer to fetched request.
1685 static struct ptlrpc_request *
1686 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, int force)
1688 struct ptlrpc_request *req;
1691 if (ptlrpc_server_high_pending(svcpt, force)) {
1692 req = ptlrpc_nrs_req_poll_nolock(svcpt, true);
1693 svcpt->scp_hreq_count++;
1697 if (ptlrpc_server_normal_pending(svcpt, force)) {
1698 req = ptlrpc_nrs_req_poll_nolock(svcpt, false);
1699 svcpt->scp_hreq_count = 0;
1706 * Handle freshly incoming reqs, add to timed early reply list,
1707 * pass on to regular request queue.
1708 * All incoming requests pass through here before getting into
1709 * ptlrpc_server_handle_req later on.
1712 ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt)
1714 struct ptlrpc_service *svc = svcpt->scp_service;
1715 struct ptlrpc_request *req;
1720 spin_lock(&svcpt->scp_lock);
1721 if (cfs_list_empty(&svcpt->scp_req_incoming)) {
1722 spin_unlock(&svcpt->scp_lock);
1726 req = cfs_list_entry(svcpt->scp_req_incoming.next,
1727 struct ptlrpc_request, rq_list);
1728 cfs_list_del_init(&req->rq_list);
1729 svcpt->scp_nreqs_incoming--;
1730 /* Consider this still a "queued" request as far as stats are
1732 spin_unlock(&svcpt->scp_lock);
1734 /* go through security check/transform */
1735 rc = sptlrpc_svc_unwrap_request(req);
1739 case SECSVC_COMPLETE:
1740 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1749 * for null-flavored rpc, msg has been unpacked by sptlrpc, although
1750 * redo it wouldn't be harmful.
1752 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
1753 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
1755 CERROR("error unpacking request: ptl %d from %s "
1756 "x"LPU64"\n", svc->srv_req_portal,
1757 libcfs_id2str(req->rq_peer), req->rq_xid);
1762 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1764 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
1765 LPU64"\n", svc->srv_req_portal,
1766 libcfs_id2str(req->rq_peer), req->rq_xid);
1770 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
1771 lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
1772 CERROR("drop incoming rpc opc %u, x"LPU64"\n",
1773 cfs_fail_val, req->rq_xid);
1778 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1779 CERROR("wrong packet type received (type=%u) from %s\n",
1780 lustre_msg_get_type(req->rq_reqmsg),
1781 libcfs_id2str(req->rq_peer));
1785 switch(lustre_msg_get_opc(req->rq_reqmsg)) {
1788 req->rq_bulk_write = 1;
1792 case MGS_CONFIG_READ:
1793 req->rq_bulk_read = 1;
1797 CDEBUG(D_RPCTRACE, "got req x"LPU64"\n", req->rq_xid);
1799 req->rq_export = class_conn2export(
1800 lustre_msg_get_handle(req->rq_reqmsg));
1801 if (req->rq_export) {
1802 class_export_rpc_get(req->rq_export);
1803 rc = ptlrpc_check_req(req);
1805 rc = sptlrpc_target_export_check(req->rq_export, req);
1807 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
1808 "illegal security flavor,");
1813 ptlrpc_update_export_timer(req->rq_export, 0);
1816 /* req_in handling should/must be fast */
1817 if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5)
1818 DEBUG_REQ(D_WARNING, req, "Slow req_in handling "CFS_DURATION_T"s",
1819 cfs_time_sub(cfs_time_current_sec(),
1820 req->rq_arrival_time.tv_sec));
1822 /* Set rpc server deadline and add it to the timed list */
1823 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1824 MSGHDR_AT_SUPPORT) ?
1825 /* The max time the client expects us to take */
1826 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1827 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1828 if (unlikely(deadline == 0)) {
1829 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1833 ptlrpc_at_add_timed(req);
1835 /* Move it over to the request processing queue */
1836 rc = ptlrpc_server_request_add(svcpt, req);
1838 ptlrpc_server_hpreq_fini(req);
1841 cfs_waitq_signal(&svcpt->scp_waitq);
1846 class_export_rpc_put(req->rq_export);
1847 spin_lock(&svcpt->scp_req_lock);
1848 svcpt->scp_nreqs_active++;
1849 spin_unlock(&svcpt->scp_req_lock);
1850 ptlrpc_server_finish_request(svcpt, req);
1856 * Main incoming request handling logic.
1857 * Calls handler function from service to do actual processing.
1860 ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
1861 struct ptlrpc_thread *thread)
1863 struct ptlrpc_service *svc = svcpt->scp_service;
1864 struct obd_export *export = NULL;
1865 struct ptlrpc_request *request;
1866 struct timeval work_start;
1867 struct timeval work_end;
1873 spin_lock(&svcpt->scp_req_lock);
1875 /* !@%$# liblustre only has 1 thread */
1876 if (cfs_atomic_read(&svcpt->scp_nreps_difficult) != 0) {
1877 spin_unlock(&svcpt->scp_req_lock);
1881 request = ptlrpc_server_request_get(svcpt, 0);
1882 if (request == NULL) {
1883 spin_unlock(&svcpt->scp_req_lock);
1887 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
1888 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
1889 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
1890 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
1892 if (unlikely(fail_opc)) {
1893 if (request->rq_export && request->rq_ops) {
1894 spin_unlock(&svcpt->scp_req_lock);
1896 OBD_FAIL_TIMEOUT(fail_opc, 4);
1898 spin_lock(&svcpt->scp_req_lock);
1899 request = ptlrpc_server_request_get(svcpt, 0);
1900 if (request == NULL) {
1901 spin_unlock(&svcpt->scp_req_lock);
1906 ptlrpc_nrs_req_del_nolock(request);
1907 svcpt->scp_nreqs_active++;
1909 svcpt->scp_nhreqs_active++;
1911 ptlrpc_nrs_req_start_nolock(request);
1912 spin_unlock(&svcpt->scp_req_lock);
1914 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
1916 if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
1917 libcfs_debug_dumplog();
1919 cfs_gettimeofday(&work_start);
1920 timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
1921 if (likely(svc->srv_stats != NULL)) {
1922 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
1924 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
1925 svcpt->scp_nreqs_incoming);
1926 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
1927 svcpt->scp_nreqs_active);
1928 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
1929 at_get(&svcpt->scp_at_estimate));
1932 export = request->rq_export;
1933 rc = lu_context_init(&request->rq_session, LCT_SESSION | LCT_NOREF);
1935 CERROR("Failure to initialize session: %d\n", rc);
1938 request->rq_session.lc_thread = thread;
1939 request->rq_session.lc_cookie = 0x5;
1940 lu_context_enter(&request->rq_session);
1942 CDEBUG(D_NET, "got req "LPU64"\n", request->rq_xid);
1944 request->rq_svc_thread = thread;
1946 request->rq_svc_thread->t_env->le_ses = &request->rq_session;
1948 if (likely(request->rq_export)) {
1949 if (unlikely(ptlrpc_check_req(request)))
1951 ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
1954 /* Discard requests queued for longer than the deadline.
1955 The deadline is increased if we send an early reply. */
1956 if (cfs_time_current_sec() > request->rq_deadline) {
1957 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s"
1958 ": deadline "CFS_DURATION_T":"CFS_DURATION_T"s ago\n",
1959 libcfs_id2str(request->rq_peer),
1960 cfs_time_sub(request->rq_deadline,
1961 request->rq_arrival_time.tv_sec),
1962 cfs_time_sub(cfs_time_current_sec(),
1963 request->rq_deadline));
1967 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
1968 "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1969 (request->rq_export ?
1970 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1971 (request->rq_export ?
1972 cfs_atomic_read(&request->rq_export->exp_refcount) : -99),
1973 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1974 libcfs_id2str(request->rq_peer),
1975 lustre_msg_get_opc(request->rq_reqmsg));
1977 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
1978 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
1980 rc = svc->srv_ops.so_req_handler(request);
1982 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
1985 lu_context_exit(&request->rq_session);
1986 lu_context_fini(&request->rq_session);
1988 if (unlikely(cfs_time_current_sec() > request->rq_deadline)) {
1989 DEBUG_REQ(D_WARNING, request, "Request took longer "
1990 "than estimated ("CFS_DURATION_T":"CFS_DURATION_T"s);"
1991 " client may timeout.",
1992 cfs_time_sub(request->rq_deadline,
1993 request->rq_arrival_time.tv_sec),
1994 cfs_time_sub(cfs_time_current_sec(),
1995 request->rq_deadline));
1998 cfs_gettimeofday(&work_end);
1999 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
2000 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
2001 "%s:%s+%d:%d:x"LPU64":%s:%d Request procesed in "
2002 "%ldus (%ldus total) trans "LPU64" rc %d/%d\n",
2004 (request->rq_export ?
2005 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2006 (request->rq_export ?
2007 cfs_atomic_read(&request->rq_export->exp_refcount) : -99),
2008 lustre_msg_get_status(request->rq_reqmsg),
2010 libcfs_id2str(request->rq_peer),
2011 lustre_msg_get_opc(request->rq_reqmsg),
2013 cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL),
2014 (request->rq_repmsg ?
2015 lustre_msg_get_transno(request->rq_repmsg) :
2016 request->rq_transno),
2018 (request->rq_repmsg ?
2019 lustre_msg_get_status(request->rq_repmsg) : -999));
2020 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
2021 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
2022 int opc = opcode_offset(op);
2023 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
2024 LASSERT(opc < LUSTRE_MAX_OPCODES);
2025 lprocfs_counter_add(svc->srv_stats,
2026 opc + EXTRA_MAX_OPCODES,
2030 if (unlikely(request->rq_early_count)) {
2031 DEBUG_REQ(D_ADAPTTO, request,
2032 "sent %d early replies before finishing in "
2034 request->rq_early_count,
2035 cfs_time_sub(work_end.tv_sec,
2036 request->rq_arrival_time.tv_sec));
2041 class_export_rpc_put(export);
2042 ptlrpc_server_finish_request(svcpt, request);
2048 * An internal function to process a single reply state object.
2051 ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
2053 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
2054 struct ptlrpc_service *svc = svcpt->scp_service;
2055 struct obd_export *exp;
2060 exp = rs->rs_export;
2062 LASSERT (rs->rs_difficult);
2063 LASSERT (rs->rs_scheduled);
2064 LASSERT (cfs_list_empty(&rs->rs_list));
2066 spin_lock(&exp->exp_lock);
2067 /* Noop if removed already */
2068 cfs_list_del_init (&rs->rs_exp_list);
2069 spin_unlock(&exp->exp_lock);
2071 /* The disk commit callback holds exp_uncommitted_replies_lock while it
2072 * iterates over newly committed replies, removing them from
2073 * exp_uncommitted_replies. It then drops this lock and schedules the
2074 * replies it found for handling here.
2076 * We can avoid contention for exp_uncommitted_replies_lock between the
2077 * HRT threads and further commit callbacks by checking rs_committed
2078 * which is set in the commit callback while it holds both
2079 * rs_lock and exp_uncommitted_reples.
2081 * If we see rs_committed clear, the commit callback _may_ not have
2082 * handled this reply yet and we race with it to grab
2083 * exp_uncommitted_replies_lock before removing the reply from
2084 * exp_uncommitted_replies. Note that if we lose the race and the
2085 * reply has already been removed, list_del_init() is a noop.
2087 * If we see rs_committed set, we know the commit callback is handling,
2088 * or has handled this reply since store reordering might allow us to
2089 * see rs_committed set out of sequence. But since this is done
2090 * holding rs_lock, we can be sure it has all completed once we hold
2091 * rs_lock, which we do right next.
2093 if (!rs->rs_committed) {
2094 spin_lock(&exp->exp_uncommitted_replies_lock);
2095 cfs_list_del_init(&rs->rs_obd_list);
2096 spin_unlock(&exp->exp_uncommitted_replies_lock);
2099 spin_lock(&rs->rs_lock);
2101 been_handled = rs->rs_handled;
2104 nlocks = rs->rs_nlocks; /* atomic "steal", but */
2105 rs->rs_nlocks = 0; /* locks still on rs_locks! */
2107 if (nlocks == 0 && !been_handled) {
2108 /* If we see this, we should already have seen the warning
2109 * in mds_steal_ack_locks() */
2110 CDEBUG(D_HA, "All locks stolen from rs %p x"LPD64".t"LPD64
2113 rs->rs_xid, rs->rs_transno, rs->rs_opc,
2114 libcfs_nid2str(exp->exp_connection->c_peer.nid));
2117 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
2118 spin_unlock(&rs->rs_lock);
2120 if (!been_handled && rs->rs_on_net) {
2121 LNetMDUnlink(rs->rs_md_h);
2122 /* Ignore return code; we're racing with completion */
2125 while (nlocks-- > 0)
2126 ldlm_lock_decref(&rs->rs_locks[nlocks],
2127 rs->rs_modes[nlocks]);
2129 spin_lock(&rs->rs_lock);
2132 rs->rs_scheduled = 0;
2134 if (!rs->rs_on_net) {
2136 spin_unlock(&rs->rs_lock);
2138 class_export_put (exp);
2139 rs->rs_export = NULL;
2140 ptlrpc_rs_decref (rs);
2141 if (cfs_atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
2142 svc->srv_is_stopping)
2143 cfs_waitq_broadcast(&svcpt->scp_waitq);
2147 /* still on the net; callback will schedule */
2148 spin_unlock(&rs->rs_lock);
2155 * Check whether given service has a reply available for processing
2158 * \param svc a ptlrpc service
2159 * \retval 0 no replies processed
2160 * \retval 1 one reply processed
2163 ptlrpc_server_handle_reply(struct ptlrpc_service_part *svcpt)
2165 struct ptlrpc_reply_state *rs = NULL;
2168 spin_lock(&svcpt->scp_rep_lock);
2169 if (!cfs_list_empty(&svcpt->scp_rep_queue)) {
2170 rs = cfs_list_entry(svcpt->scp_rep_queue.prev,
2171 struct ptlrpc_reply_state,
2173 cfs_list_del_init(&rs->rs_list);
2175 spin_unlock(&svcpt->scp_rep_lock);
2177 ptlrpc_handle_rs(rs);
2181 /* FIXME make use of timeout later */
2183 liblustre_check_services (void *arg)
2185 int did_something = 0;
2187 cfs_list_t *tmp, *nxt;
2190 /* I'm relying on being single threaded, not to have to lock
2191 * ptlrpc_all_services etc */
2192 cfs_list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
2193 struct ptlrpc_service *svc =
2194 cfs_list_entry (tmp, struct ptlrpc_service, srv_list);
2195 struct ptlrpc_service_part *svcpt;
2197 LASSERT(svc->srv_ncpts == 1);
2198 svcpt = svc->srv_parts[0];
2200 if (svcpt->scp_nthrs_running != 0) /* I've recursed */
2203 /* service threads can block for bulk, so this limits us
2204 * (arbitrarily) to recursing 1 stack frame per service.
2205 * Note that the problem with recursion is that we have to
2206 * unwind completely before our caller can resume. */
2208 svcpt->scp_nthrs_running++;
2211 rc = ptlrpc_server_handle_req_in(svcpt);
2212 rc |= ptlrpc_server_handle_reply(svcpt);
2213 rc |= ptlrpc_at_check_timed(svcpt);
2214 rc |= ptlrpc_server_handle_request(svcpt, NULL);
2215 rc |= (ptlrpc_server_post_idle_rqbds(svcpt) > 0);
2216 did_something |= rc;
2219 svcpt->scp_nthrs_running--;
2222 RETURN(did_something);
2224 #define ptlrpc_stop_all_threads(s) do {} while (0)
2226 #else /* __KERNEL__ */
2229 ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
2231 int avail = svcpt->scp_nrqbds_posted;
2232 int low_water = test_req_buffer_pressure ? 0 :
2233 svcpt->scp_service->srv_nbuf_per_group / 2;
2235 /* NB I'm not locking; just looking. */
2237 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
2238 * allowed the request history to grow out of control. We could put a
2239 * sanity check on that here and cull some history if we need the
2242 if (avail <= low_water)
2243 ptlrpc_grow_req_bufs(svcpt, 1);
2245 if (svcpt->scp_service->srv_stats) {
2246 lprocfs_counter_add(svcpt->scp_service->srv_stats,
2247 PTLRPC_REQBUF_AVAIL_CNTR, avail);
2252 ptlrpc_retry_rqbds(void *arg)
2254 struct ptlrpc_service_part *svcpt = (struct ptlrpc_service_part *)arg;
2256 svcpt->scp_rqbd_timeout = 0;
2261 ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
2263 return svcpt->scp_nreqs_active <
2264 svcpt->scp_nthrs_running - 1 -
2265 (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL);
2269 * allowed to create more threads
2270 * user can call it w/o any lock but need to hold
2271 * ptlrpc_service_part::scp_lock to get reliable result
2274 ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
2276 return svcpt->scp_nthrs_running +
2277 svcpt->scp_nthrs_starting <
2278 svcpt->scp_service->srv_nthrs_cpt_limit;
2282 * too many requests and allowed to create more threads
2285 ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
2287 return !ptlrpc_threads_enough(svcpt) &&
2288 ptlrpc_threads_increasable(svcpt);
2292 ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2294 return thread_is_stopping(thread) ||
2295 thread->t_svcpt->scp_service->srv_is_stopping;
2299 ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
2301 return !cfs_list_empty(&svcpt->scp_rqbd_idle) &&
2302 svcpt->scp_rqbd_timeout == 0;
2306 ptlrpc_at_check(struct ptlrpc_service_part *svcpt)
2308 return svcpt->scp_at_check;
2312 * requests wait on preprocessing
2313 * user can call it w/o any lock but need to hold
2314 * ptlrpc_service_part::scp_lock to get reliable result
2317 ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
2319 return !cfs_list_empty(&svcpt->scp_req_incoming);
2322 static __attribute__((__noinline__)) int
2323 ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
2324 struct ptlrpc_thread *thread)
2326 /* Don't exit while there are replies to be handled */
2327 struct l_wait_info lwi = LWI_TIMEOUT(svcpt->scp_rqbd_timeout,
2328 ptlrpc_retry_rqbds, svcpt);
2330 lc_watchdog_disable(thread->t_watchdog);
2334 l_wait_event_exclusive_head(svcpt->scp_waitq,
2335 ptlrpc_thread_stopping(thread) ||
2336 ptlrpc_server_request_incoming(svcpt) ||
2337 ptlrpc_server_request_pending(svcpt, 0) ||
2338 ptlrpc_rqbd_pending(svcpt) ||
2339 ptlrpc_at_check(svcpt), &lwi);
2341 if (ptlrpc_thread_stopping(thread))
2344 lc_watchdog_touch(thread->t_watchdog,
2345 ptlrpc_server_get_timeout(svcpt));
2350 * Main thread body for service threads.
2351 * Waits in a loop waiting for new requests to process to appear.
2352 * Every time an incoming requests is added to its queue, a waitq
2353 * is woken up and one of the threads will handle it.
2355 static int ptlrpc_main(void *arg)
2357 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
2358 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2359 struct ptlrpc_service *svc = svcpt->scp_service;
2360 struct ptlrpc_reply_state *rs;
2361 #ifdef WITH_GROUP_INFO
2362 cfs_group_info_t *ginfo = NULL;
2365 int counter = 0, rc = 0;
2368 thread->t_pid = cfs_curproc_pid();
2369 cfs_daemonize_ctxt(thread->t_name);
2371 /* NB: we will call cfs_cpt_bind() for all threads, because we
2372 * might want to run lustre server only on a subset of system CPUs,
2373 * in that case ->scp_cpt is CFS_CPT_ANY */
2374 rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
2376 CWARN("%s: failed to bind %s on CPT %d\n",
2377 svc->srv_name, thread->t_name, svcpt->scp_cpt);
2380 #ifdef WITH_GROUP_INFO
2381 ginfo = cfs_groups_alloc(0);
2387 cfs_set_current_groups(ginfo);
2388 cfs_put_group_info(ginfo);
2391 if (svc->srv_ops.so_thr_init != NULL) {
2392 rc = svc->srv_ops.so_thr_init(thread);
2403 rc = lu_context_init(&env->le_ctx,
2404 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2408 thread->t_env = env;
2409 env->le_ctx.lc_thread = thread;
2410 env->le_ctx.lc_cookie = 0x6;
2412 while (!cfs_list_empty(&svcpt->scp_rqbd_idle)) {
2413 rc = ptlrpc_server_post_idle_rqbds(svcpt);
2417 CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
2418 svc->srv_name, svcpt->scp_cpt, rc);
2422 /* Alloc reply state structure for this one */
2423 OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2429 spin_lock(&svcpt->scp_lock);
2431 LASSERT(thread_is_starting(thread));
2432 thread_clear_flags(thread, SVC_STARTING);
2434 LASSERT(svcpt->scp_nthrs_starting == 1);
2435 svcpt->scp_nthrs_starting--;
2437 /* SVC_STOPPING may already be set here if someone else is trying
2438 * to stop the service while this new thread has been dynamically
2439 * forked. We still set SVC_RUNNING to let our creator know that
2440 * we are now running, however we will exit as soon as possible */
2441 thread_add_flags(thread, SVC_RUNNING);
2442 svcpt->scp_nthrs_running++;
2443 spin_unlock(&svcpt->scp_lock);
2445 /* wake up our creator in case he's still waiting. */
2446 cfs_waitq_signal(&thread->t_ctl_waitq);
2448 thread->t_watchdog = lc_watchdog_add(ptlrpc_server_get_timeout(svcpt),
2451 spin_lock(&svcpt->scp_rep_lock);
2452 cfs_list_add(&rs->rs_list, &svcpt->scp_rep_idle);
2453 cfs_waitq_signal(&svcpt->scp_rep_waitq);
2454 spin_unlock(&svcpt->scp_rep_lock);
2456 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2457 svcpt->scp_nthrs_running);
2459 /* XXX maintain a list of all managed devices: insert here */
2460 while (!ptlrpc_thread_stopping(thread)) {
2461 if (ptlrpc_wait_event(svcpt, thread))
2464 ptlrpc_check_rqbd_pool(svcpt);
2466 if (ptlrpc_threads_need_create(svcpt)) {
2467 /* Ignore return code - we tried... */
2468 ptlrpc_start_thread(svcpt, 0);
2471 /* Process all incoming reqs before handling any */
2472 if (ptlrpc_server_request_incoming(svcpt)) {
2473 ptlrpc_server_handle_req_in(svcpt);
2474 /* but limit ourselves in case of flood */
2475 if (counter++ < 100)
2480 if (ptlrpc_at_check(svcpt))
2481 ptlrpc_at_check_timed(svcpt);
2483 if (ptlrpc_server_request_pending(svcpt, 0)) {
2484 lu_context_enter(&env->le_ctx);
2485 ptlrpc_server_handle_request(svcpt, thread);
2486 lu_context_exit(&env->le_ctx);
2489 if (ptlrpc_rqbd_pending(svcpt) &&
2490 ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
2491 /* I just failed to repost request buffers.
2492 * Wait for a timeout (unless something else
2493 * happens) before I try again */
2494 svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
2495 CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
2496 svcpt->scp_nrqbds_posted);
2500 lc_watchdog_delete(thread->t_watchdog);
2501 thread->t_watchdog = NULL;
2505 * deconstruct service specific state created by ptlrpc_start_thread()
2507 if (svc->srv_ops.so_thr_done != NULL)
2508 svc->srv_ops.so_thr_done(thread);
2511 lu_context_fini(&env->le_ctx);
2515 CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n",
2516 thread, thread->t_pid, thread->t_id, rc);
2518 spin_lock(&svcpt->scp_lock);
2519 if (thread_test_and_clear_flags(thread, SVC_STARTING))
2520 svcpt->scp_nthrs_starting--;
2522 if (thread_test_and_clear_flags(thread, SVC_RUNNING)) {
2523 /* must know immediately */
2524 svcpt->scp_nthrs_running--;
2528 thread_add_flags(thread, SVC_STOPPED);
2530 cfs_waitq_signal(&thread->t_ctl_waitq);
2531 spin_unlock(&svcpt->scp_lock);
2536 static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
2537 cfs_list_t *replies)
2541 spin_lock(&hrt->hrt_lock);
2543 cfs_list_splice_init(&hrt->hrt_queue, replies);
2544 result = ptlrpc_hr.hr_stopping || !cfs_list_empty(replies);
2546 spin_unlock(&hrt->hrt_lock);
2551 * Main body of "handle reply" function.
2552 * It processes acked reply states
2554 static int ptlrpc_hr_main(void *arg)
2556 struct ptlrpc_hr_thread *hrt = (struct ptlrpc_hr_thread *)arg;
2557 struct ptlrpc_hr_partition *hrp = hrt->hrt_partition;
2558 CFS_LIST_HEAD (replies);
2559 char threadname[20];
2562 snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
2563 hrp->hrp_cpt, hrt->hrt_id);
2564 cfs_daemonize_ctxt(threadname);
2566 rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
2568 CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
2569 threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
2572 cfs_atomic_inc(&hrp->hrp_nstarted);
2573 cfs_waitq_signal(&ptlrpc_hr.hr_waitq);
2575 while (!ptlrpc_hr.hr_stopping) {
2576 l_wait_condition(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
2578 while (!cfs_list_empty(&replies)) {
2579 struct ptlrpc_reply_state *rs;
2581 rs = cfs_list_entry(replies.prev,
2582 struct ptlrpc_reply_state,
2584 cfs_list_del_init(&rs->rs_list);
2585 ptlrpc_handle_rs(rs);
2589 cfs_atomic_inc(&hrp->hrp_nstopped);
2590 cfs_waitq_signal(&ptlrpc_hr.hr_waitq);
2595 static void ptlrpc_stop_hr_threads(void)
2597 struct ptlrpc_hr_partition *hrp;
2601 ptlrpc_hr.hr_stopping = 1;
2603 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2604 if (hrp->hrp_thrs == NULL)
2605 continue; /* uninitialized */
2606 for (j = 0; j < hrp->hrp_nthrs; j++)
2607 cfs_waitq_broadcast(&hrp->hrp_thrs[j].hrt_waitq);
2610 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2611 if (hrp->hrp_thrs == NULL)
2612 continue; /* uninitialized */
2613 cfs_wait_event(ptlrpc_hr.hr_waitq,
2614 cfs_atomic_read(&hrp->hrp_nstopped) ==
2615 cfs_atomic_read(&hrp->hrp_nstarted));
2619 static int ptlrpc_start_hr_threads(void)
2621 struct ptlrpc_hr_partition *hrp;
2626 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2629 for (j = 0; j < hrp->hrp_nthrs; j++) {
2630 rc = cfs_create_thread(ptlrpc_hr_main,
2632 CLONE_VM | CLONE_FILES);
2636 cfs_wait_event(ptlrpc_hr.hr_waitq,
2637 cfs_atomic_read(&hrp->hrp_nstarted) == j);
2641 CERROR("Reply handling thread %d:%d Failed on starting: "
2642 "rc = %d\n", i, j, rc);
2643 ptlrpc_stop_hr_threads();
2649 static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
2651 struct l_wait_info lwi = { 0 };
2652 struct ptlrpc_thread *thread;
2653 CFS_LIST_HEAD (zombie);
2657 CDEBUG(D_INFO, "Stopping threads for service %s\n",
2658 svcpt->scp_service->srv_name);
2660 spin_lock(&svcpt->scp_lock);
2661 /* let the thread know that we would like it to stop asap */
2662 list_for_each_entry(thread, &svcpt->scp_threads, t_link) {
2663 CDEBUG(D_INFO, "Stopping thread %s #%u\n",
2664 svcpt->scp_service->srv_thread_name, thread->t_id);
2665 thread_add_flags(thread, SVC_STOPPING);
2668 cfs_waitq_broadcast(&svcpt->scp_waitq);
2670 while (!cfs_list_empty(&svcpt->scp_threads)) {
2671 thread = cfs_list_entry(svcpt->scp_threads.next,
2672 struct ptlrpc_thread, t_link);
2673 if (thread_is_stopped(thread)) {
2674 cfs_list_del(&thread->t_link);
2675 cfs_list_add(&thread->t_link, &zombie);
2678 spin_unlock(&svcpt->scp_lock);
2680 CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
2681 svcpt->scp_service->srv_thread_name, thread->t_id);
2682 l_wait_event(thread->t_ctl_waitq,
2683 thread_is_stopped(thread), &lwi);
2685 spin_lock(&svcpt->scp_lock);
2688 spin_unlock(&svcpt->scp_lock);
2690 while (!cfs_list_empty(&zombie)) {
2691 thread = cfs_list_entry(zombie.next,
2692 struct ptlrpc_thread, t_link);
2693 cfs_list_del(&thread->t_link);
2694 OBD_FREE_PTR(thread);
2700 * Stops all threads of a particular service \a svc
2702 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2704 struct ptlrpc_service_part *svcpt;
2708 ptlrpc_service_for_each_part(svcpt, i, svc) {
2709 if (svcpt->scp_service != NULL)
2710 ptlrpc_svcpt_stop_threads(svcpt);
2715 EXPORT_SYMBOL(ptlrpc_stop_all_threads);
2717 int ptlrpc_start_threads(struct ptlrpc_service *svc)
2724 /* We require 2 threads min, see note in ptlrpc_server_handle_request */
2725 LASSERT(svc->srv_nthrs_cpt_init >= PTLRPC_NTHRS_INIT);
2727 for (i = 0; i < svc->srv_ncpts; i++) {
2728 for (j = 0; j < svc->srv_nthrs_cpt_init; j++) {
2729 rc = ptlrpc_start_thread(svc->srv_parts[i], 1);
2735 /* We have enough threads, don't start more. b=15759 */
2742 CERROR("cannot start %s thread #%d_%d: rc %d\n",
2743 svc->srv_thread_name, i, j, rc);
2744 ptlrpc_stop_all_threads(svc);
2747 EXPORT_SYMBOL(ptlrpc_start_threads);
2749 int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
2751 struct l_wait_info lwi = { 0 };
2752 struct ptlrpc_thread *thread;
2753 struct ptlrpc_service *svc;
2757 LASSERT(svcpt != NULL);
2759 svc = svcpt->scp_service;
2761 CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n",
2762 svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running,
2763 svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit);
2766 if (unlikely(svc->srv_is_stopping))
2769 if (!ptlrpc_threads_increasable(svcpt) ||
2770 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
2771 svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1))
2774 OBD_CPT_ALLOC_PTR(thread, svc->srv_cptable, svcpt->scp_cpt);
2777 cfs_waitq_init(&thread->t_ctl_waitq);
2779 spin_lock(&svcpt->scp_lock);
2780 if (!ptlrpc_threads_increasable(svcpt)) {
2781 spin_unlock(&svcpt->scp_lock);
2782 OBD_FREE_PTR(thread);
2786 if (svcpt->scp_nthrs_starting != 0) {
2787 /* serialize starting because some modules (obdfilter)
2788 * might require unique and contiguous t_id */
2789 LASSERT(svcpt->scp_nthrs_starting == 1);
2790 spin_unlock(&svcpt->scp_lock);
2791 OBD_FREE_PTR(thread);
2793 CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n",
2794 svc->srv_thread_name, svcpt->scp_thr_nextid);
2799 CDEBUG(D_INFO, "Creating thread %s #%d race, retry later\n",
2800 svc->srv_thread_name, svcpt->scp_thr_nextid);
2804 svcpt->scp_nthrs_starting++;
2805 thread->t_id = svcpt->scp_thr_nextid++;
2806 thread_add_flags(thread, SVC_STARTING);
2807 thread->t_svcpt = svcpt;
2809 cfs_list_add(&thread->t_link, &svcpt->scp_threads);
2810 spin_unlock(&svcpt->scp_lock);
2812 if (svcpt->scp_cpt >= 0) {
2813 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s%02d_%03d",
2814 svc->srv_thread_name, svcpt->scp_cpt, thread->t_id);
2816 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s_%04d",
2817 svc->srv_thread_name, thread->t_id);
2820 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name);
2822 * CLONE_VM and CLONE_FILES just avoid a needless copy, because we
2823 * just drop the VM and FILES in cfs_daemonize_ctxt() right away.
2825 rc = cfs_create_thread(ptlrpc_main, thread, CFS_DAEMON_FLAGS);
2827 CERROR("cannot start thread '%s': rc %d\n",
2828 thread->t_name, rc);
2829 spin_lock(&svcpt->scp_lock);
2830 cfs_list_del(&thread->t_link);
2831 --svcpt->scp_nthrs_starting;
2832 spin_unlock(&svcpt->scp_lock);
2834 OBD_FREE(thread, sizeof(*thread));
2841 l_wait_event(thread->t_ctl_waitq,
2842 thread_is_running(thread) || thread_is_stopped(thread),
2845 rc = thread_is_stopped(thread) ? thread->t_id : 0;
2849 int ptlrpc_hr_init(void)
2851 struct ptlrpc_hr_partition *hrp;
2852 struct ptlrpc_hr_thread *hrt;
2858 memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
2859 ptlrpc_hr.hr_cpt_table = cfs_cpt_table;
2861 ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
2863 if (ptlrpc_hr.hr_partitions == NULL)
2866 cfs_waitq_init(&ptlrpc_hr.hr_waitq);
2868 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2871 cfs_atomic_set(&hrp->hrp_nstarted, 0);
2872 cfs_atomic_set(&hrp->hrp_nstopped, 0);
2874 hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, i);
2875 hrp->hrp_nthrs /= cfs_cpu_ht_nsiblings(0);
2877 LASSERT(hrp->hrp_nthrs > 0);
2878 OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, i,
2879 hrp->hrp_nthrs * sizeof(*hrt));
2880 if (hrp->hrp_thrs == NULL)
2881 GOTO(out, rc = -ENOMEM);
2883 for (j = 0; j < hrp->hrp_nthrs; j++) {
2884 hrt = &hrp->hrp_thrs[j];
2887 hrt->hrt_partition = hrp;
2888 cfs_waitq_init(&hrt->hrt_waitq);
2889 spin_lock_init(&hrt->hrt_lock);
2890 CFS_INIT_LIST_HEAD(&hrt->hrt_queue);
2894 rc = ptlrpc_start_hr_threads();
2901 void ptlrpc_hr_fini(void)
2903 struct ptlrpc_hr_partition *hrp;
2906 if (ptlrpc_hr.hr_partitions == NULL)
2909 ptlrpc_stop_hr_threads();
2911 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2912 if (hrp->hrp_thrs != NULL) {
2913 OBD_FREE(hrp->hrp_thrs,
2914 hrp->hrp_nthrs * sizeof(hrp->hrp_thrs[0]));
2918 cfs_percpt_free(ptlrpc_hr.hr_partitions);
2919 ptlrpc_hr.hr_partitions = NULL;
2922 #endif /* __KERNEL__ */
2925 * Wait until all already scheduled replies are processed.
2927 static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
2931 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
2934 rc = l_wait_event(svcpt->scp_waitq,
2935 cfs_atomic_read(&svcpt->scp_nreps_difficult) == 0, &lwi);
2938 CWARN("Unexpectedly long timeout %s %p\n",
2939 svcpt->scp_service->srv_name, svcpt->scp_service);
2944 ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
2946 struct ptlrpc_service_part *svcpt;
2949 /* early disarm AT timer... */
2950 ptlrpc_service_for_each_part(svcpt, i, svc) {
2951 if (svcpt->scp_service != NULL)
2952 cfs_timer_disarm(&svcpt->scp_at_timer);
2957 ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
2959 struct ptlrpc_service_part *svcpt;
2960 struct ptlrpc_request_buffer_desc *rqbd;
2961 struct l_wait_info lwi;
2965 /* All history will be culled when the next request buffer is
2966 * freed in ptlrpc_service_purge_all() */
2967 svc->srv_hist_nrqbds_cpt_max = 0;
2969 rc = LNetClearLazyPortal(svc->srv_req_portal);
2972 ptlrpc_service_for_each_part(svcpt, i, svc) {
2973 if (svcpt->scp_service == NULL)
2976 /* Unlink all the request buffers. This forces a 'final'
2977 * event with its 'unlink' flag set for each posted rqbd */
2978 cfs_list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
2980 rc = LNetMDUnlink(rqbd->rqbd_md_h);
2981 LASSERT(rc == 0 || rc == -ENOENT);
2985 ptlrpc_service_for_each_part(svcpt, i, svc) {
2986 if (svcpt->scp_service == NULL)
2989 /* Wait for the network to release any buffers
2990 * it's currently filling */
2991 spin_lock(&svcpt->scp_lock);
2992 while (svcpt->scp_nrqbds_posted != 0) {
2993 spin_unlock(&svcpt->scp_lock);
2994 /* Network access will complete in finite time but
2995 * the HUGE timeout lets us CWARN for visibility
2996 * of sluggish NALs */
2997 lwi = LWI_TIMEOUT_INTERVAL(
2998 cfs_time_seconds(LONG_UNLINK),
2999 cfs_time_seconds(1), NULL, NULL);
3000 rc = l_wait_event(svcpt->scp_waitq,
3001 svcpt->scp_nrqbds_posted == 0, &lwi);
3002 if (rc == -ETIMEDOUT) {
3003 CWARN("Service %s waiting for "
3004 "request buffers\n",
3005 svcpt->scp_service->srv_name);
3007 spin_lock(&svcpt->scp_lock);
3009 spin_unlock(&svcpt->scp_lock);
3014 ptlrpc_service_purge_all(struct ptlrpc_service *svc)
3016 struct ptlrpc_service_part *svcpt;
3017 struct ptlrpc_request_buffer_desc *rqbd;
3018 struct ptlrpc_request *req;
3019 struct ptlrpc_reply_state *rs;
3022 ptlrpc_service_for_each_part(svcpt, i, svc) {
3023 if (svcpt->scp_service == NULL)
3026 spin_lock(&svcpt->scp_rep_lock);
3027 while (!cfs_list_empty(&svcpt->scp_rep_active)) {
3028 rs = cfs_list_entry(svcpt->scp_rep_active.next,
3029 struct ptlrpc_reply_state, rs_list);
3030 spin_lock(&rs->rs_lock);
3031 ptlrpc_schedule_difficult_reply(rs);
3032 spin_unlock(&rs->rs_lock);
3034 spin_unlock(&svcpt->scp_rep_lock);
3036 /* purge the request queue. NB No new replies (rqbds
3037 * all unlinked) and no service threads, so I'm the only
3038 * thread noodling the request queue now */
3039 while (!cfs_list_empty(&svcpt->scp_req_incoming)) {
3040 req = cfs_list_entry(svcpt->scp_req_incoming.next,
3041 struct ptlrpc_request, rq_list);
3043 cfs_list_del(&req->rq_list);
3044 svcpt->scp_nreqs_incoming--;
3045 svcpt->scp_nreqs_active++;
3046 ptlrpc_server_finish_request(svcpt, req);
3049 while (ptlrpc_server_request_pending(svcpt, 1)) {
3050 req = ptlrpc_server_request_get(svcpt, 1);
3051 ptlrpc_nrs_req_del_nolock(req);
3052 svcpt->scp_nreqs_active++;
3053 ptlrpc_server_hpreq_fini(req);
3055 if (req->rq_export != NULL)
3056 class_export_rpc_put(req->rq_export);
3057 ptlrpc_server_finish_request(svcpt, req);
3060 LASSERT(cfs_list_empty(&svcpt->scp_rqbd_posted));
3061 LASSERT(svcpt->scp_nreqs_incoming == 0);
3062 LASSERT(svcpt->scp_nreqs_active == 0);
3063 /* history should have been culled by
3064 * ptlrpc_server_finish_request */
3065 LASSERT(svcpt->scp_hist_nrqbds == 0);
3067 /* Now free all the request buffers since nothing
3068 * references them any more... */
3070 while (!cfs_list_empty(&svcpt->scp_rqbd_idle)) {
3071 rqbd = cfs_list_entry(svcpt->scp_rqbd_idle.next,
3072 struct ptlrpc_request_buffer_desc,
3074 ptlrpc_free_rqbd(rqbd);
3076 ptlrpc_wait_replies(svcpt);
3078 while (!cfs_list_empty(&svcpt->scp_rep_idle)) {
3079 rs = cfs_list_entry(svcpt->scp_rep_idle.next,
3080 struct ptlrpc_reply_state,
3082 cfs_list_del(&rs->rs_list);
3083 OBD_FREE_LARGE(rs, svc->srv_max_reply_size);
3089 ptlrpc_service_free(struct ptlrpc_service *svc)
3091 struct ptlrpc_service_part *svcpt;
3092 struct ptlrpc_at_array *array;
3095 ptlrpc_service_for_each_part(svcpt, i, svc) {
3096 if (svcpt->scp_service == NULL)
3099 /* In case somebody rearmed this in the meantime */
3100 cfs_timer_disarm(&svcpt->scp_at_timer);
3101 array = &svcpt->scp_at_array;
3103 if (array->paa_reqs_array != NULL) {
3104 OBD_FREE(array->paa_reqs_array,
3105 sizeof(cfs_list_t) * array->paa_size);
3106 array->paa_reqs_array = NULL;
3109 if (array->paa_reqs_count != NULL) {
3110 OBD_FREE(array->paa_reqs_count,
3111 sizeof(__u32) * array->paa_size);
3112 array->paa_reqs_count = NULL;
3116 ptlrpc_service_for_each_part(svcpt, i, svc)
3117 OBD_FREE_PTR(svcpt);
3119 if (svc->srv_cpts != NULL)
3120 cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts);
3122 OBD_FREE(svc, offsetof(struct ptlrpc_service,
3123 srv_parts[svc->srv_ncpts]));
3126 int ptlrpc_unregister_service(struct ptlrpc_service *service)
3130 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
3132 service->srv_is_stopping = 1;
3134 mutex_lock(&ptlrpc_all_services_mutex);
3135 cfs_list_del_init(&service->srv_list);
3136 mutex_unlock(&ptlrpc_all_services_mutex);
3138 ptlrpc_service_del_atimer(service);
3139 ptlrpc_stop_all_threads(service);
3141 ptlrpc_service_unlink_rqbd(service);
3142 ptlrpc_service_purge_all(service);
3143 ptlrpc_service_nrs_cleanup(service);
3145 ptlrpc_lprocfs_unregister_service(service);
3147 ptlrpc_service_free(service);
3151 EXPORT_SYMBOL(ptlrpc_unregister_service);
3154 * Returns 0 if the service is healthy.
3156 * Right now, it just checks to make sure that requests aren't languishing
3157 * in the queue. We'll use this health check to govern whether a node needs
3158 * to be shot, so it's intentionally non-aggressive. */
3159 int ptlrpc_svcpt_health_check(struct ptlrpc_service_part *svcpt)
3161 struct ptlrpc_request *request = NULL;
3162 struct timeval right_now;
3165 cfs_gettimeofday(&right_now);
3167 spin_lock(&svcpt->scp_req_lock);
3168 /* How long has the next entry been waiting? */
3169 if (ptlrpc_server_high_pending(svcpt, 1))
3170 request = ptlrpc_nrs_req_poll_nolock(svcpt, true);
3171 else if (ptlrpc_server_normal_pending(svcpt, 1))
3172 request = ptlrpc_nrs_req_poll_nolock(svcpt, false);
3174 if (request == NULL) {
3175 spin_unlock(&svcpt->scp_req_lock);
3179 timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
3180 spin_unlock(&svcpt->scp_req_lock);
3182 if ((timediff / ONE_MILLION) >
3183 (AT_OFF ? obd_timeout * 3 / 2 : at_max)) {
3184 CERROR("%s: unhealthy - request has been waiting %lds\n",
3185 svcpt->scp_service->srv_name, timediff / ONE_MILLION);
3193 ptlrpc_service_health_check(struct ptlrpc_service *svc)
3195 struct ptlrpc_service_part *svcpt;
3198 if (svc == NULL || svc->srv_parts == NULL)
3201 ptlrpc_service_for_each_part(svcpt, i, svc) {
3202 int rc = ptlrpc_svcpt_health_check(svcpt);
3209 EXPORT_SYMBOL(ptlrpc_service_health_check);