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, 2013, 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 CDEBUG(D_OTHER, "%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);
987 /** Change request export and move hp request from old export to new */
988 void ptlrpc_request_change_export(struct ptlrpc_request *req,
989 struct obd_export *export)
991 if (req->rq_export != NULL) {
992 if (!cfs_list_empty(&req->rq_exp_list)) {
993 /* remove rq_exp_list from last export */
994 spin_lock_bh(&req->rq_export->exp_rpc_lock);
995 cfs_list_del_init(&req->rq_exp_list);
996 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
998 /* export has one reference already, so it`s safe to
999 * add req to export queue here and get another
1000 * reference for request later */
1001 spin_lock_bh(&export->exp_rpc_lock);
1002 cfs_list_add(&req->rq_exp_list, &export->exp_hp_rpcs);
1003 spin_unlock_bh(&export->exp_rpc_lock);
1005 class_export_rpc_dec(req->rq_export);
1006 class_export_put(req->rq_export);
1009 /* request takes one export refcount */
1010 req->rq_export = class_export_get(export);
1011 class_export_rpc_inc(export);
1017 * to finish a request: stop sending more early replies, and release
1020 static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt,
1021 struct ptlrpc_request *req)
1023 ptlrpc_server_hpreq_fini(req);
1025 ptlrpc_server_drop_request(req);
1029 * to finish a active request: stop sending more early replies, and release
1030 * the request. should be called after we finished handling the request.
1032 static void ptlrpc_server_finish_active_request(
1033 struct ptlrpc_service_part *svcpt,
1034 struct ptlrpc_request *req)
1036 spin_lock(&svcpt->scp_req_lock);
1037 ptlrpc_nrs_req_stop_nolock(req);
1038 svcpt->scp_nreqs_active--;
1040 svcpt->scp_nhreqs_active--;
1041 spin_unlock(&svcpt->scp_req_lock);
1043 ptlrpc_nrs_req_finalize(req);
1045 if (req->rq_export != NULL)
1046 class_export_rpc_dec(req->rq_export);
1048 ptlrpc_server_finish_request(svcpt, req);
1052 * This function makes sure dead exports are evicted in a timely manner.
1053 * This function is only called when some export receives a message (i.e.,
1054 * the network is up.)
1056 static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
1058 struct obd_export *oldest_exp;
1059 time_t oldest_time, new_time;
1065 /* Compensate for slow machines, etc, by faking our request time
1066 into the future. Although this can break the strict time-ordering
1067 of the list, we can be really lazy here - we don't have to evict
1068 at the exact right moment. Eventually, all silent exports
1069 will make it to the top of the list. */
1071 /* Do not pay attention on 1sec or smaller renewals. */
1072 new_time = cfs_time_current_sec() + extra_delay;
1073 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
1076 exp->exp_last_request_time = new_time;
1077 CDEBUG(D_HA, "updating export %s at "CFS_TIME_T" exp %p\n",
1078 exp->exp_client_uuid.uuid,
1079 exp->exp_last_request_time, exp);
1081 /* exports may get disconnected from the chain even though the
1082 export has references, so we must keep the spin lock while
1083 manipulating the lists */
1084 spin_lock(&exp->exp_obd->obd_dev_lock);
1086 if (cfs_list_empty(&exp->exp_obd_chain_timed)) {
1087 /* this one is not timed */
1088 spin_unlock(&exp->exp_obd->obd_dev_lock);
1092 cfs_list_move_tail(&exp->exp_obd_chain_timed,
1093 &exp->exp_obd->obd_exports_timed);
1095 oldest_exp = cfs_list_entry(exp->exp_obd->obd_exports_timed.next,
1096 struct obd_export, exp_obd_chain_timed);
1097 oldest_time = oldest_exp->exp_last_request_time;
1098 spin_unlock(&exp->exp_obd->obd_dev_lock);
1100 if (exp->exp_obd->obd_recovering) {
1101 /* be nice to everyone during recovery */
1106 /* Note - racing to start/reset the obd_eviction timer is safe */
1107 if (exp->exp_obd->obd_eviction_timer == 0) {
1108 /* Check if the oldest entry is expired. */
1109 if (cfs_time_current_sec() > (oldest_time + PING_EVICT_TIMEOUT +
1111 /* We need a second timer, in case the net was down and
1112 * it just came back. Since the pinger may skip every
1113 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
1114 * we better wait for 3. */
1115 exp->exp_obd->obd_eviction_timer =
1116 cfs_time_current_sec() + 3 * PING_INTERVAL;
1117 CDEBUG(D_HA, "%s: Think about evicting %s from "CFS_TIME_T"\n",
1118 exp->exp_obd->obd_name,
1119 obd_export_nid2str(oldest_exp), oldest_time);
1122 if (cfs_time_current_sec() >
1123 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
1124 /* The evictor won't evict anyone who we've heard from
1125 * recently, so we don't have to check before we start
1127 if (!ping_evictor_wake(exp))
1128 exp->exp_obd->obd_eviction_timer = 0;
1136 * Sanity check request \a req.
1137 * Return 0 if all is ok, error code otherwise.
1139 static int ptlrpc_check_req(struct ptlrpc_request *req)
1143 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
1144 req->rq_export->exp_conn_cnt)) {
1145 DEBUG_REQ(D_RPCTRACE, req,
1146 "DROPPING req from old connection %d < %d",
1147 lustre_msg_get_conn_cnt(req->rq_reqmsg),
1148 req->rq_export->exp_conn_cnt);
1151 if (unlikely(req->rq_export->exp_obd &&
1152 req->rq_export->exp_obd->obd_fail)) {
1153 /* Failing over, don't handle any more reqs, send
1154 error response instead. */
1155 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
1156 req, req->rq_export->exp_obd->obd_name);
1158 } else if (lustre_msg_get_flags(req->rq_reqmsg) &
1159 (MSG_REPLAY | MSG_REQ_REPLAY_DONE) &&
1160 !(req->rq_export->exp_obd->obd_recovering)) {
1161 DEBUG_REQ(D_ERROR, req,
1162 "Invalid replay without recovery");
1163 class_fail_export(req->rq_export);
1165 } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 &&
1166 !(req->rq_export->exp_obd->obd_recovering)) {
1167 DEBUG_REQ(D_ERROR, req, "Invalid req with transno "
1168 LPU64" without recovery",
1169 lustre_msg_get_transno(req->rq_reqmsg));
1170 class_fail_export(req->rq_export);
1174 if (unlikely(rc < 0)) {
1175 req->rq_status = rc;
1181 static void ptlrpc_at_set_timer(struct ptlrpc_service_part *svcpt)
1183 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1186 if (array->paa_count == 0) {
1187 cfs_timer_disarm(&svcpt->scp_at_timer);
1191 /* Set timer for closest deadline */
1192 next = (__s32)(array->paa_deadline - cfs_time_current_sec() -
1195 ptlrpc_at_timer((unsigned long)svcpt);
1197 cfs_timer_arm(&svcpt->scp_at_timer, cfs_time_shift(next));
1198 CDEBUG(D_INFO, "armed %s at %+ds\n",
1199 svcpt->scp_service->srv_name, next);
1203 /* Add rpc to early reply check list */
1204 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
1206 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1207 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1208 struct ptlrpc_request *rq = NULL;
1214 if (req->rq_no_reply)
1217 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
1220 spin_lock(&svcpt->scp_at_lock);
1221 LASSERT(cfs_list_empty(&req->rq_timed_list));
1223 index = (unsigned long)req->rq_deadline % array->paa_size;
1224 if (array->paa_reqs_count[index] > 0) {
1225 /* latest rpcs will have the latest deadlines in the list,
1226 * so search backward. */
1227 cfs_list_for_each_entry_reverse(rq,
1228 &array->paa_reqs_array[index],
1230 if (req->rq_deadline >= rq->rq_deadline) {
1231 cfs_list_add(&req->rq_timed_list,
1232 &rq->rq_timed_list);
1238 /* Add the request at the head of the list */
1239 if (cfs_list_empty(&req->rq_timed_list))
1240 cfs_list_add(&req->rq_timed_list,
1241 &array->paa_reqs_array[index]);
1243 spin_lock(&req->rq_lock);
1244 req->rq_at_linked = 1;
1245 spin_unlock(&req->rq_lock);
1246 req->rq_at_index = index;
1247 array->paa_reqs_count[index]++;
1249 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
1250 array->paa_deadline = req->rq_deadline;
1251 ptlrpc_at_set_timer(svcpt);
1253 spin_unlock(&svcpt->scp_at_lock);
1259 ptlrpc_at_remove_timed(struct ptlrpc_request *req)
1261 struct ptlrpc_at_array *array;
1263 array = &req->rq_rqbd->rqbd_svcpt->scp_at_array;
1265 /* NB: must call with hold svcpt::scp_at_lock */
1266 LASSERT(!cfs_list_empty(&req->rq_timed_list));
1267 cfs_list_del_init(&req->rq_timed_list);
1269 spin_lock(&req->rq_lock);
1270 req->rq_at_linked = 0;
1271 spin_unlock(&req->rq_lock);
1273 array->paa_reqs_count[req->rq_at_index]--;
1277 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
1279 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1280 struct ptlrpc_request *reqcopy;
1281 struct lustre_msg *reqmsg;
1282 cfs_duration_t olddl = req->rq_deadline - cfs_time_current_sec();
1287 /* deadline is when the client expects us to reply, margin is the
1288 difference between clients' and servers' expectations */
1289 DEBUG_REQ(D_ADAPTTO, req,
1290 "%ssending early reply (deadline %+lds, margin %+lds) for "
1291 "%d+%d", AT_OFF ? "AT off - not " : "",
1292 olddl, olddl - at_get(&svcpt->scp_at_estimate),
1293 at_get(&svcpt->scp_at_estimate), at_extra);
1299 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), "
1300 "not sending early reply. Consider increasing "
1301 "at_early_margin (%d)?", olddl, at_early_margin);
1303 /* Return an error so we're not re-added to the timed list. */
1307 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
1308 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
1309 "but no AT support");
1313 if (req->rq_export &&
1314 lustre_msg_get_flags(req->rq_reqmsg) &
1315 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1316 /* During recovery, we don't want to send too many early
1317 * replies, but on the other hand we want to make sure the
1318 * client has enough time to resend if the rpc is lost. So
1319 * during the recovery period send at least 4 early replies,
1320 * spacing them every at_extra if we can. at_estimate should
1321 * always equal this fixed value during recovery. */
1322 at_measured(&svcpt->scp_at_estimate, min(at_extra,
1323 req->rq_export->exp_obd->obd_recovery_timeout / 4));
1325 /* Fake our processing time into the future to ask the clients
1326 * for some extra amount of time */
1327 at_measured(&svcpt->scp_at_estimate, at_extra +
1328 cfs_time_current_sec() -
1329 req->rq_arrival_time.tv_sec);
1331 /* Check to see if we've actually increased the deadline -
1332 * we may be past adaptive_max */
1333 if (req->rq_deadline >= req->rq_arrival_time.tv_sec +
1334 at_get(&svcpt->scp_at_estimate)) {
1335 DEBUG_REQ(D_WARNING, req, "Couldn't add any time "
1336 "(%ld/%ld), not sending early reply\n",
1337 olddl, req->rq_arrival_time.tv_sec +
1338 at_get(&svcpt->scp_at_estimate) -
1339 cfs_time_current_sec());
1343 newdl = cfs_time_current_sec() + at_get(&svcpt->scp_at_estimate);
1345 OBD_ALLOC(reqcopy, sizeof *reqcopy);
1346 if (reqcopy == NULL)
1348 OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1350 OBD_FREE(reqcopy, sizeof *reqcopy);
1355 reqcopy->rq_reply_state = NULL;
1356 reqcopy->rq_rep_swab_mask = 0;
1357 reqcopy->rq_pack_bulk = 0;
1358 reqcopy->rq_pack_udesc = 0;
1359 reqcopy->rq_packed_final = 0;
1360 sptlrpc_svc_ctx_addref(reqcopy);
1361 /* We only need the reqmsg for the magic */
1362 reqcopy->rq_reqmsg = reqmsg;
1363 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1365 LASSERT(cfs_atomic_read(&req->rq_refcount));
1366 /** if it is last refcount then early reply isn't needed */
1367 if (cfs_atomic_read(&req->rq_refcount) == 1) {
1368 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
1369 "abort sending early reply\n");
1370 GOTO(out, rc = -EINVAL);
1373 /* Connection ref */
1374 reqcopy->rq_export = class_conn2export(
1375 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1376 if (reqcopy->rq_export == NULL)
1377 GOTO(out, rc = -ENODEV);
1380 class_export_rpc_inc(reqcopy->rq_export);
1381 if (reqcopy->rq_export->exp_obd &&
1382 reqcopy->rq_export->exp_obd->obd_fail)
1383 GOTO(out_put, rc = -ENODEV);
1385 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1389 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1392 /* Adjust our own deadline to what we told the client */
1393 req->rq_deadline = newdl;
1394 req->rq_early_count++; /* number sent, server side */
1396 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1399 /* Free the (early) reply state from lustre_pack_reply.
1400 (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
1401 ptlrpc_req_drop_rs(reqcopy);
1404 class_export_rpc_dec(reqcopy->rq_export);
1405 class_export_put(reqcopy->rq_export);
1407 sptlrpc_svc_ctx_decref(reqcopy);
1408 OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1409 OBD_FREE(reqcopy, sizeof *reqcopy);
1413 /* Send early replies to everybody expiring within at_early_margin
1414 asking for at_extra time */
1415 static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1417 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1418 struct ptlrpc_request *rq, *n;
1419 cfs_list_t work_list;
1422 time_t now = cfs_time_current_sec();
1423 cfs_duration_t delay;
1424 int first, counter = 0;
1427 spin_lock(&svcpt->scp_at_lock);
1428 if (svcpt->scp_at_check == 0) {
1429 spin_unlock(&svcpt->scp_at_lock);
1432 delay = cfs_time_sub(cfs_time_current(), svcpt->scp_at_checktime);
1433 svcpt->scp_at_check = 0;
1435 if (array->paa_count == 0) {
1436 spin_unlock(&svcpt->scp_at_lock);
1440 /* The timer went off, but maybe the nearest rpc already completed. */
1441 first = array->paa_deadline - now;
1442 if (first > at_early_margin) {
1443 /* We've still got plenty of time. Reset the timer. */
1444 ptlrpc_at_set_timer(svcpt);
1445 spin_unlock(&svcpt->scp_at_lock);
1449 /* We're close to a timeout, and we don't know how much longer the
1450 server will take. Send early replies to everyone expiring soon. */
1451 CFS_INIT_LIST_HEAD(&work_list);
1453 index = (unsigned long)array->paa_deadline % array->paa_size;
1454 count = array->paa_count;
1456 count -= array->paa_reqs_count[index];
1457 cfs_list_for_each_entry_safe(rq, n,
1458 &array->paa_reqs_array[index],
1460 if (rq->rq_deadline > now + at_early_margin) {
1461 /* update the earliest deadline */
1462 if (deadline == -1 ||
1463 rq->rq_deadline < deadline)
1464 deadline = rq->rq_deadline;
1468 ptlrpc_at_remove_timed(rq);
1470 * ptlrpc_server_drop_request() may drop
1471 * refcount to 0 already. Let's check this and
1472 * don't add entry to work_list
1474 if (likely(cfs_atomic_inc_not_zero(&rq->rq_refcount)))
1475 cfs_list_add(&rq->rq_timed_list, &work_list);
1479 if (++index >= array->paa_size)
1482 array->paa_deadline = deadline;
1483 /* we have a new earliest deadline, restart the timer */
1484 ptlrpc_at_set_timer(svcpt);
1486 spin_unlock(&svcpt->scp_at_lock);
1488 CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
1489 "replies\n", first, at_extra, counter);
1491 /* We're already past request deadlines before we even get a
1492 chance to send early replies */
1493 LCONSOLE_WARN("%s: This server is not able to keep up with "
1494 "request traffic (cpu-bound).\n",
1495 svcpt->scp_service->srv_name);
1496 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, "
1497 "delay="CFS_DURATION_T"(jiff)\n",
1498 counter, svcpt->scp_nreqs_incoming,
1499 svcpt->scp_nreqs_active,
1500 at_get(&svcpt->scp_at_estimate), delay);
1503 /* we took additional refcount so entries can't be deleted from list, no
1504 * locking is needed */
1505 while (!cfs_list_empty(&work_list)) {
1506 rq = cfs_list_entry(work_list.next, struct ptlrpc_request,
1508 cfs_list_del_init(&rq->rq_timed_list);
1510 if (ptlrpc_at_send_early_reply(rq) == 0)
1511 ptlrpc_at_add_timed(rq);
1513 ptlrpc_server_drop_request(rq);
1516 RETURN(1); /* return "did_something" for liblustre */
1520 * Put the request to the export list if the request may become
1521 * a high priority one.
1523 static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt,
1524 struct ptlrpc_request *req)
1529 if (svcpt->scp_service->srv_ops.so_hpreq_handler) {
1530 rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
1535 if (req->rq_export && req->rq_ops) {
1536 /* Perform request specific check. We should do this check
1537 * before the request is added into exp_hp_rpcs list otherwise
1538 * it may hit swab race at LU-1044. */
1539 if (req->rq_ops->hpreq_check) {
1540 rc = req->rq_ops->hpreq_check(req);
1542 * XXX: Out of all current
1543 * ptlrpc_hpreq_ops::hpreq_check(), only
1544 * ldlm_cancel_hpreq_check() can return an error code;
1545 * other functions assert in similar places, which seems
1546 * odd. What also does not seem right is that handlers
1547 * for those RPCs do not assert on the same checks, but
1548 * rather handle the error cases. e.g. see
1549 * ost_rw_hpreq_check(), and ost_brw_read(),
1554 LASSERT(rc == 0 || rc == 1);
1557 spin_lock_bh(&req->rq_export->exp_rpc_lock);
1558 cfs_list_add(&req->rq_exp_list,
1559 &req->rq_export->exp_hp_rpcs);
1560 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1563 ptlrpc_nrs_req_initialize(svcpt, req, rc);
1568 /** Remove the request from the export list. */
1569 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req)
1572 if (req->rq_export && req->rq_ops) {
1573 /* refresh lock timeout again so that client has more
1574 * room to send lock cancel RPC. */
1575 if (req->rq_ops->hpreq_fini)
1576 req->rq_ops->hpreq_fini(req);
1578 spin_lock_bh(&req->rq_export->exp_rpc_lock);
1579 cfs_list_del_init(&req->rq_exp_list);
1580 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1585 static int ptlrpc_hpreq_check(struct ptlrpc_request *req)
1590 static struct ptlrpc_hpreq_ops ptlrpc_hpreq_common = {
1591 .hpreq_check = ptlrpc_hpreq_check,
1594 /* Hi-Priority RPC check by RPC operation code. */
1595 int ptlrpc_hpreq_handler(struct ptlrpc_request *req)
1597 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1599 /* Check for export to let only reconnects for not yet evicted
1600 * export to become a HP rpc. */
1601 if ((req->rq_export != NULL) &&
1602 (opc == OBD_PING || opc == MDS_CONNECT || opc == OST_CONNECT))
1603 req->rq_ops = &ptlrpc_hpreq_common;
1607 EXPORT_SYMBOL(ptlrpc_hpreq_handler);
1609 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1610 struct ptlrpc_request *req)
1615 rc = ptlrpc_server_hpreq_init(svcpt, req);
1619 ptlrpc_nrs_req_add(svcpt, req, !!rc);
1625 * Allow to handle high priority request
1626 * User can call it w/o any lock but need to hold
1627 * ptlrpc_service_part::scp_req_lock to get reliable result
1629 static bool ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1632 int running = svcpt->scp_nthrs_running;
1634 if (!nrs_svcpt_has_hp(svcpt))
1640 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1641 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1642 /* leave just 1 thread for normal RPCs */
1643 running = PTLRPC_NTHRS_INIT;
1644 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1648 if (svcpt->scp_nreqs_active >= running - 1)
1651 if (svcpt->scp_nhreqs_active == 0)
1654 return !ptlrpc_nrs_req_pending_nolock(svcpt, false) ||
1655 svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1658 static bool ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1661 return ptlrpc_server_allow_high(svcpt, force) &&
1662 ptlrpc_nrs_req_pending_nolock(svcpt, true);
1666 * Only allow normal priority requests on a service that has a high-priority
1667 * queue if forced (i.e. cleanup), if there are other high priority requests
1668 * already being processed (i.e. those threads can service more high-priority
1669 * requests), or if there are enough idle threads that a later thread can do
1670 * a high priority request.
1671 * User can call it w/o any lock but need to hold
1672 * ptlrpc_service_part::scp_req_lock to get reliable result
1674 static bool ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1677 int running = svcpt->scp_nthrs_running;
1679 if (1) /* always allow to handle normal request for liblustre */
1682 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1683 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1684 /* leave just 1 thread for normal RPCs */
1685 running = PTLRPC_NTHRS_INIT;
1686 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1691 svcpt->scp_nreqs_active < running - 2)
1694 if (svcpt->scp_nreqs_active >= running - 1)
1697 return svcpt->scp_nhreqs_active > 0 || !nrs_svcpt_has_hp(svcpt);
1700 static bool ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1703 return ptlrpc_server_allow_normal(svcpt, force) &&
1704 ptlrpc_nrs_req_pending_nolock(svcpt, false);
1708 * Returns true if there are requests available in incoming
1709 * request queue for processing and it is allowed to fetch them.
1710 * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1711 * to get reliable result
1712 * \see ptlrpc_server_allow_normal
1713 * \see ptlrpc_server_allow high
1716 ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt, bool force)
1718 return ptlrpc_server_high_pending(svcpt, force) ||
1719 ptlrpc_server_normal_pending(svcpt, force);
1723 * Fetch a request for processing from queue of unprocessed requests.
1724 * Favors high-priority requests.
1725 * Returns a pointer to fetched request.
1727 static struct ptlrpc_request *
1728 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force)
1730 struct ptlrpc_request *req = NULL;
1733 spin_lock(&svcpt->scp_req_lock);
1735 /* !@%$# liblustre only has 1 thread */
1736 if (cfs_atomic_read(&svcpt->scp_nreps_difficult) != 0) {
1737 spin_unlock(&svcpt->scp_req_lock);
1742 if (ptlrpc_server_high_pending(svcpt, force)) {
1743 req = ptlrpc_nrs_req_get_nolock(svcpt, true, force);
1745 svcpt->scp_hreq_count++;
1750 if (ptlrpc_server_normal_pending(svcpt, force)) {
1751 req = ptlrpc_nrs_req_get_nolock(svcpt, false, force);
1753 svcpt->scp_hreq_count = 0;
1758 spin_unlock(&svcpt->scp_req_lock);
1762 svcpt->scp_nreqs_active++;
1764 svcpt->scp_nhreqs_active++;
1766 spin_unlock(&svcpt->scp_req_lock);
1768 if (likely(req->rq_export))
1769 class_export_rpc_inc(req->rq_export);
1775 * Handle freshly incoming reqs, add to timed early reply list,
1776 * pass on to regular request queue.
1777 * All incoming requests pass through here before getting into
1778 * ptlrpc_server_handle_req later on.
1781 ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt,
1782 struct ptlrpc_thread *thread)
1784 struct ptlrpc_service *svc = svcpt->scp_service;
1785 struct ptlrpc_request *req;
1790 spin_lock(&svcpt->scp_lock);
1791 if (cfs_list_empty(&svcpt->scp_req_incoming)) {
1792 spin_unlock(&svcpt->scp_lock);
1796 req = cfs_list_entry(svcpt->scp_req_incoming.next,
1797 struct ptlrpc_request, rq_list);
1798 cfs_list_del_init(&req->rq_list);
1799 svcpt->scp_nreqs_incoming--;
1800 /* Consider this still a "queued" request as far as stats are
1802 spin_unlock(&svcpt->scp_lock);
1804 /* go through security check/transform */
1805 rc = sptlrpc_svc_unwrap_request(req);
1809 case SECSVC_COMPLETE:
1810 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1819 * for null-flavored rpc, msg has been unpacked by sptlrpc, although
1820 * redo it wouldn't be harmful.
1822 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
1823 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
1825 CERROR("error unpacking request: ptl %d from %s "
1826 "x"LPU64"\n", svc->srv_req_portal,
1827 libcfs_id2str(req->rq_peer), req->rq_xid);
1832 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1834 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
1835 LPU64"\n", svc->srv_req_portal,
1836 libcfs_id2str(req->rq_peer), req->rq_xid);
1840 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
1841 lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
1842 CERROR("drop incoming rpc opc %u, x"LPU64"\n",
1843 cfs_fail_val, req->rq_xid);
1848 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1849 CERROR("wrong packet type received (type=%u) from %s\n",
1850 lustre_msg_get_type(req->rq_reqmsg),
1851 libcfs_id2str(req->rq_peer));
1855 switch(lustre_msg_get_opc(req->rq_reqmsg)) {
1858 req->rq_bulk_write = 1;
1862 case MGS_CONFIG_READ:
1863 req->rq_bulk_read = 1;
1867 CDEBUG(D_RPCTRACE, "got req x"LPU64"\n", req->rq_xid);
1869 req->rq_export = class_conn2export(
1870 lustre_msg_get_handle(req->rq_reqmsg));
1871 if (req->rq_export) {
1872 rc = ptlrpc_check_req(req);
1874 rc = sptlrpc_target_export_check(req->rq_export, req);
1876 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
1877 "illegal security flavor,");
1882 ptlrpc_update_export_timer(req->rq_export, 0);
1885 /* req_in handling should/must be fast */
1886 if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5)
1887 DEBUG_REQ(D_WARNING, req, "Slow req_in handling "CFS_DURATION_T"s",
1888 cfs_time_sub(cfs_time_current_sec(),
1889 req->rq_arrival_time.tv_sec));
1891 /* Set rpc server deadline and add it to the timed list */
1892 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1893 MSGHDR_AT_SUPPORT) ?
1894 /* The max time the client expects us to take */
1895 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1896 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1897 if (unlikely(deadline == 0)) {
1898 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1902 req->rq_svc_thread = thread;
1904 ptlrpc_at_add_timed(req);
1906 /* Move it over to the request processing queue */
1907 rc = ptlrpc_server_request_add(svcpt, req);
1911 cfs_waitq_signal(&svcpt->scp_waitq);
1915 ptlrpc_server_finish_request(svcpt, req);
1921 * Main incoming request handling logic.
1922 * Calls handler function from service to do actual processing.
1925 ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
1926 struct ptlrpc_thread *thread)
1928 struct ptlrpc_service *svc = svcpt->scp_service;
1929 struct ptlrpc_request *request;
1930 struct timeval work_start;
1931 struct timeval work_end;
1937 request = ptlrpc_server_request_get(svcpt, false);
1938 if (request == NULL)
1941 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
1942 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
1943 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
1944 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
1946 if (unlikely(fail_opc)) {
1947 if (request->rq_export && request->rq_ops)
1948 OBD_FAIL_TIMEOUT(fail_opc, 4);
1951 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
1953 if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
1954 libcfs_debug_dumplog();
1956 cfs_gettimeofday(&work_start);
1957 timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
1958 if (likely(svc->srv_stats != NULL)) {
1959 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
1961 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
1962 svcpt->scp_nreqs_incoming);
1963 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
1964 svcpt->scp_nreqs_active);
1965 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
1966 at_get(&svcpt->scp_at_estimate));
1969 rc = lu_context_init(&request->rq_session, LCT_SESSION | LCT_NOREF);
1971 CERROR("Failure to initialize session: %d\n", rc);
1974 request->rq_session.lc_thread = thread;
1975 request->rq_session.lc_cookie = 0x5;
1976 lu_context_enter(&request->rq_session);
1978 CDEBUG(D_NET, "got req "LPU64"\n", request->rq_xid);
1980 request->rq_svc_thread = thread;
1982 request->rq_svc_thread->t_env->le_ses = &request->rq_session;
1984 if (likely(request->rq_export)) {
1985 if (unlikely(ptlrpc_check_req(request)))
1987 ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
1990 /* Discard requests queued for longer than the deadline.
1991 The deadline is increased if we send an early reply. */
1992 if (cfs_time_current_sec() > request->rq_deadline) {
1993 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s"
1994 ": deadline "CFS_DURATION_T":"CFS_DURATION_T"s ago\n",
1995 libcfs_id2str(request->rq_peer),
1996 cfs_time_sub(request->rq_deadline,
1997 request->rq_arrival_time.tv_sec),
1998 cfs_time_sub(cfs_time_current_sec(),
1999 request->rq_deadline));
2003 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
2004 "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
2005 (request->rq_export ?
2006 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2007 (request->rq_export ?
2008 cfs_atomic_read(&request->rq_export->exp_refcount) : -99),
2009 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
2010 libcfs_id2str(request->rq_peer),
2011 lustre_msg_get_opc(request->rq_reqmsg));
2013 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
2014 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
2016 rc = svc->srv_ops.so_req_handler(request);
2018 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
2021 lu_context_exit(&request->rq_session);
2022 lu_context_fini(&request->rq_session);
2024 if (unlikely(cfs_time_current_sec() > request->rq_deadline)) {
2025 DEBUG_REQ(D_WARNING, request, "Request took longer "
2026 "than estimated ("CFS_DURATION_T":"CFS_DURATION_T"s);"
2027 " client may timeout.",
2028 cfs_time_sub(request->rq_deadline,
2029 request->rq_arrival_time.tv_sec),
2030 cfs_time_sub(cfs_time_current_sec(),
2031 request->rq_deadline));
2034 cfs_gettimeofday(&work_end);
2035 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
2036 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
2037 "%s:%s+%d:%d:x"LPU64":%s:%d Request procesed in "
2038 "%ldus (%ldus total) trans "LPU64" rc %d/%d\n",
2040 (request->rq_export ?
2041 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2042 (request->rq_export ?
2043 cfs_atomic_read(&request->rq_export->exp_refcount) : -99),
2044 lustre_msg_get_status(request->rq_reqmsg),
2046 libcfs_id2str(request->rq_peer),
2047 lustre_msg_get_opc(request->rq_reqmsg),
2049 cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL),
2050 (request->rq_repmsg ?
2051 lustre_msg_get_transno(request->rq_repmsg) :
2052 request->rq_transno),
2054 (request->rq_repmsg ?
2055 lustre_msg_get_status(request->rq_repmsg) : -999));
2056 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
2057 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
2058 int opc = opcode_offset(op);
2059 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
2060 LASSERT(opc < LUSTRE_MAX_OPCODES);
2061 lprocfs_counter_add(svc->srv_stats,
2062 opc + EXTRA_MAX_OPCODES,
2066 if (unlikely(request->rq_early_count)) {
2067 DEBUG_REQ(D_ADAPTTO, request,
2068 "sent %d early replies before finishing in "
2070 request->rq_early_count,
2071 cfs_time_sub(work_end.tv_sec,
2072 request->rq_arrival_time.tv_sec));
2076 ptlrpc_server_finish_active_request(svcpt, request);
2082 * An internal function to process a single reply state object.
2085 ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
2087 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
2088 struct ptlrpc_service *svc = svcpt->scp_service;
2089 struct obd_export *exp;
2094 exp = rs->rs_export;
2096 LASSERT (rs->rs_difficult);
2097 LASSERT (rs->rs_scheduled);
2098 LASSERT (cfs_list_empty(&rs->rs_list));
2100 spin_lock(&exp->exp_lock);
2101 /* Noop if removed already */
2102 cfs_list_del_init (&rs->rs_exp_list);
2103 spin_unlock(&exp->exp_lock);
2105 /* The disk commit callback holds exp_uncommitted_replies_lock while it
2106 * iterates over newly committed replies, removing them from
2107 * exp_uncommitted_replies. It then drops this lock and schedules the
2108 * replies it found for handling here.
2110 * We can avoid contention for exp_uncommitted_replies_lock between the
2111 * HRT threads and further commit callbacks by checking rs_committed
2112 * which is set in the commit callback while it holds both
2113 * rs_lock and exp_uncommitted_reples.
2115 * If we see rs_committed clear, the commit callback _may_ not have
2116 * handled this reply yet and we race with it to grab
2117 * exp_uncommitted_replies_lock before removing the reply from
2118 * exp_uncommitted_replies. Note that if we lose the race and the
2119 * reply has already been removed, list_del_init() is a noop.
2121 * If we see rs_committed set, we know the commit callback is handling,
2122 * or has handled this reply since store reordering might allow us to
2123 * see rs_committed set out of sequence. But since this is done
2124 * holding rs_lock, we can be sure it has all completed once we hold
2125 * rs_lock, which we do right next.
2127 if (!rs->rs_committed) {
2128 spin_lock(&exp->exp_uncommitted_replies_lock);
2129 cfs_list_del_init(&rs->rs_obd_list);
2130 spin_unlock(&exp->exp_uncommitted_replies_lock);
2133 spin_lock(&rs->rs_lock);
2135 been_handled = rs->rs_handled;
2138 nlocks = rs->rs_nlocks; /* atomic "steal", but */
2139 rs->rs_nlocks = 0; /* locks still on rs_locks! */
2141 if (nlocks == 0 && !been_handled) {
2142 /* If we see this, we should already have seen the warning
2143 * in mds_steal_ack_locks() */
2144 CDEBUG(D_HA, "All locks stolen from rs %p x"LPD64".t"LPD64
2147 rs->rs_xid, rs->rs_transno, rs->rs_opc,
2148 libcfs_nid2str(exp->exp_connection->c_peer.nid));
2151 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
2152 spin_unlock(&rs->rs_lock);
2154 if (!been_handled && rs->rs_on_net) {
2155 LNetMDUnlink(rs->rs_md_h);
2156 /* Ignore return code; we're racing with completion */
2159 while (nlocks-- > 0)
2160 ldlm_lock_decref(&rs->rs_locks[nlocks],
2161 rs->rs_modes[nlocks]);
2163 spin_lock(&rs->rs_lock);
2166 rs->rs_scheduled = 0;
2168 if (!rs->rs_on_net) {
2170 spin_unlock(&rs->rs_lock);
2172 class_export_put (exp);
2173 rs->rs_export = NULL;
2174 ptlrpc_rs_decref (rs);
2175 if (cfs_atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
2176 svc->srv_is_stopping)
2177 cfs_waitq_broadcast(&svcpt->scp_waitq);
2181 /* still on the net; callback will schedule */
2182 spin_unlock(&rs->rs_lock);
2189 * Check whether given service has a reply available for processing
2192 * \param svc a ptlrpc service
2193 * \retval 0 no replies processed
2194 * \retval 1 one reply processed
2197 ptlrpc_server_handle_reply(struct ptlrpc_service_part *svcpt)
2199 struct ptlrpc_reply_state *rs = NULL;
2202 spin_lock(&svcpt->scp_rep_lock);
2203 if (!cfs_list_empty(&svcpt->scp_rep_queue)) {
2204 rs = cfs_list_entry(svcpt->scp_rep_queue.prev,
2205 struct ptlrpc_reply_state,
2207 cfs_list_del_init(&rs->rs_list);
2209 spin_unlock(&svcpt->scp_rep_lock);
2211 ptlrpc_handle_rs(rs);
2215 /* FIXME make use of timeout later */
2217 liblustre_check_services (void *arg)
2219 int did_something = 0;
2221 cfs_list_t *tmp, *nxt;
2224 /* I'm relying on being single threaded, not to have to lock
2225 * ptlrpc_all_services etc */
2226 cfs_list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
2227 struct ptlrpc_service *svc =
2228 cfs_list_entry (tmp, struct ptlrpc_service, srv_list);
2229 struct ptlrpc_service_part *svcpt;
2231 LASSERT(svc->srv_ncpts == 1);
2232 svcpt = svc->srv_parts[0];
2234 if (svcpt->scp_nthrs_running != 0) /* I've recursed */
2237 /* service threads can block for bulk, so this limits us
2238 * (arbitrarily) to recursing 1 stack frame per service.
2239 * Note that the problem with recursion is that we have to
2240 * unwind completely before our caller can resume. */
2242 svcpt->scp_nthrs_running++;
2245 rc = ptlrpc_server_handle_req_in(svcpt, NULL);
2246 rc |= ptlrpc_server_handle_reply(svcpt);
2247 rc |= ptlrpc_at_check_timed(svcpt);
2248 rc |= ptlrpc_server_handle_request(svcpt, NULL);
2249 rc |= (ptlrpc_server_post_idle_rqbds(svcpt) > 0);
2250 did_something |= rc;
2253 svcpt->scp_nthrs_running--;
2256 RETURN(did_something);
2258 #define ptlrpc_stop_all_threads(s) do {} while (0)
2260 #else /* __KERNEL__ */
2263 ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
2265 int avail = svcpt->scp_nrqbds_posted;
2266 int low_water = test_req_buffer_pressure ? 0 :
2267 svcpt->scp_service->srv_nbuf_per_group / 2;
2269 /* NB I'm not locking; just looking. */
2271 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
2272 * allowed the request history to grow out of control. We could put a
2273 * sanity check on that here and cull some history if we need the
2276 if (avail <= low_water)
2277 ptlrpc_grow_req_bufs(svcpt, 1);
2279 if (svcpt->scp_service->srv_stats) {
2280 lprocfs_counter_add(svcpt->scp_service->srv_stats,
2281 PTLRPC_REQBUF_AVAIL_CNTR, avail);
2286 ptlrpc_retry_rqbds(void *arg)
2288 struct ptlrpc_service_part *svcpt = (struct ptlrpc_service_part *)arg;
2290 svcpt->scp_rqbd_timeout = 0;
2295 ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
2297 return svcpt->scp_nreqs_active <
2298 svcpt->scp_nthrs_running - 1 -
2299 (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL);
2303 * allowed to create more threads
2304 * user can call it w/o any lock but need to hold
2305 * ptlrpc_service_part::scp_lock to get reliable result
2308 ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
2310 return svcpt->scp_nthrs_running +
2311 svcpt->scp_nthrs_starting <
2312 svcpt->scp_service->srv_nthrs_cpt_limit;
2316 * too many requests and allowed to create more threads
2319 ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
2321 return !ptlrpc_threads_enough(svcpt) &&
2322 ptlrpc_threads_increasable(svcpt);
2326 ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2328 return thread_is_stopping(thread) ||
2329 thread->t_svcpt->scp_service->srv_is_stopping;
2333 ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
2335 return !cfs_list_empty(&svcpt->scp_rqbd_idle) &&
2336 svcpt->scp_rqbd_timeout == 0;
2340 ptlrpc_at_check(struct ptlrpc_service_part *svcpt)
2342 return svcpt->scp_at_check;
2346 * requests wait on preprocessing
2347 * user can call it w/o any lock but need to hold
2348 * ptlrpc_service_part::scp_lock to get reliable result
2351 ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
2353 return !cfs_list_empty(&svcpt->scp_req_incoming);
2356 static __attribute__((__noinline__)) int
2357 ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
2358 struct ptlrpc_thread *thread)
2360 /* Don't exit while there are replies to be handled */
2361 struct l_wait_info lwi = LWI_TIMEOUT(svcpt->scp_rqbd_timeout,
2362 ptlrpc_retry_rqbds, svcpt);
2364 lc_watchdog_disable(thread->t_watchdog);
2368 l_wait_event_exclusive_head(svcpt->scp_waitq,
2369 ptlrpc_thread_stopping(thread) ||
2370 ptlrpc_server_request_incoming(svcpt) ||
2371 ptlrpc_server_request_pending(svcpt, false) ||
2372 ptlrpc_rqbd_pending(svcpt) ||
2373 ptlrpc_at_check(svcpt), &lwi);
2375 if (ptlrpc_thread_stopping(thread))
2378 lc_watchdog_touch(thread->t_watchdog,
2379 ptlrpc_server_get_timeout(svcpt));
2384 * Main thread body for service threads.
2385 * Waits in a loop waiting for new requests to process to appear.
2386 * Every time an incoming requests is added to its queue, a waitq
2387 * is woken up and one of the threads will handle it.
2389 static int ptlrpc_main(void *arg)
2391 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
2392 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2393 struct ptlrpc_service *svc = svcpt->scp_service;
2394 struct ptlrpc_reply_state *rs;
2395 #ifdef WITH_GROUP_INFO
2396 cfs_group_info_t *ginfo = NULL;
2399 int counter = 0, rc = 0;
2402 thread->t_pid = cfs_curproc_pid();
2403 cfs_daemonize_ctxt(thread->t_name);
2405 /* NB: we will call cfs_cpt_bind() for all threads, because we
2406 * might want to run lustre server only on a subset of system CPUs,
2407 * in that case ->scp_cpt is CFS_CPT_ANY */
2408 rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
2410 CWARN("%s: failed to bind %s on CPT %d\n",
2411 svc->srv_name, thread->t_name, svcpt->scp_cpt);
2414 #ifdef WITH_GROUP_INFO
2415 ginfo = cfs_groups_alloc(0);
2421 cfs_set_current_groups(ginfo);
2422 cfs_put_group_info(ginfo);
2425 if (svc->srv_ops.so_thr_init != NULL) {
2426 rc = svc->srv_ops.so_thr_init(thread);
2437 rc = lu_context_init(&env->le_ctx,
2438 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2442 thread->t_env = env;
2443 env->le_ctx.lc_thread = thread;
2444 env->le_ctx.lc_cookie = 0x6;
2446 while (!cfs_list_empty(&svcpt->scp_rqbd_idle)) {
2447 rc = ptlrpc_server_post_idle_rqbds(svcpt);
2451 CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
2452 svc->srv_name, svcpt->scp_cpt, rc);
2456 /* Alloc reply state structure for this one */
2457 OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2463 spin_lock(&svcpt->scp_lock);
2465 LASSERT(thread_is_starting(thread));
2466 thread_clear_flags(thread, SVC_STARTING);
2468 LASSERT(svcpt->scp_nthrs_starting == 1);
2469 svcpt->scp_nthrs_starting--;
2471 /* SVC_STOPPING may already be set here if someone else is trying
2472 * to stop the service while this new thread has been dynamically
2473 * forked. We still set SVC_RUNNING to let our creator know that
2474 * we are now running, however we will exit as soon as possible */
2475 thread_add_flags(thread, SVC_RUNNING);
2476 svcpt->scp_nthrs_running++;
2477 spin_unlock(&svcpt->scp_lock);
2479 /* wake up our creator in case he's still waiting. */
2480 cfs_waitq_signal(&thread->t_ctl_waitq);
2482 thread->t_watchdog = lc_watchdog_add(ptlrpc_server_get_timeout(svcpt),
2485 spin_lock(&svcpt->scp_rep_lock);
2486 cfs_list_add(&rs->rs_list, &svcpt->scp_rep_idle);
2487 cfs_waitq_signal(&svcpt->scp_rep_waitq);
2488 spin_unlock(&svcpt->scp_rep_lock);
2490 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2491 svcpt->scp_nthrs_running);
2493 /* XXX maintain a list of all managed devices: insert here */
2494 while (!ptlrpc_thread_stopping(thread)) {
2495 if (ptlrpc_wait_event(svcpt, thread))
2498 ptlrpc_check_rqbd_pool(svcpt);
2500 if (ptlrpc_threads_need_create(svcpt)) {
2501 /* Ignore return code - we tried... */
2502 ptlrpc_start_thread(svcpt, 0);
2505 /* Process all incoming reqs before handling any */
2506 if (ptlrpc_server_request_incoming(svcpt)) {
2507 lu_context_enter(&env->le_ctx);
2509 ptlrpc_server_handle_req_in(svcpt, thread);
2510 lu_context_exit(&env->le_ctx);
2512 /* but limit ourselves in case of flood */
2513 if (counter++ < 100)
2518 if (ptlrpc_at_check(svcpt))
2519 ptlrpc_at_check_timed(svcpt);
2521 if (ptlrpc_server_request_pending(svcpt, false)) {
2522 lu_context_enter(&env->le_ctx);
2523 ptlrpc_server_handle_request(svcpt, thread);
2524 lu_context_exit(&env->le_ctx);
2527 if (ptlrpc_rqbd_pending(svcpt) &&
2528 ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
2529 /* I just failed to repost request buffers.
2530 * Wait for a timeout (unless something else
2531 * happens) before I try again */
2532 svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
2533 CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
2534 svcpt->scp_nrqbds_posted);
2538 lc_watchdog_delete(thread->t_watchdog);
2539 thread->t_watchdog = NULL;
2543 * deconstruct service specific state created by ptlrpc_start_thread()
2545 if (svc->srv_ops.so_thr_done != NULL)
2546 svc->srv_ops.so_thr_done(thread);
2549 lu_context_fini(&env->le_ctx);
2553 CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n",
2554 thread, thread->t_pid, thread->t_id, rc);
2556 spin_lock(&svcpt->scp_lock);
2557 if (thread_test_and_clear_flags(thread, SVC_STARTING))
2558 svcpt->scp_nthrs_starting--;
2560 if (thread_test_and_clear_flags(thread, SVC_RUNNING)) {
2561 /* must know immediately */
2562 svcpt->scp_nthrs_running--;
2566 thread_add_flags(thread, SVC_STOPPED);
2568 cfs_waitq_signal(&thread->t_ctl_waitq);
2569 spin_unlock(&svcpt->scp_lock);
2574 static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
2575 cfs_list_t *replies)
2579 spin_lock(&hrt->hrt_lock);
2581 cfs_list_splice_init(&hrt->hrt_queue, replies);
2582 result = ptlrpc_hr.hr_stopping || !cfs_list_empty(replies);
2584 spin_unlock(&hrt->hrt_lock);
2589 * Main body of "handle reply" function.
2590 * It processes acked reply states
2592 static int ptlrpc_hr_main(void *arg)
2594 struct ptlrpc_hr_thread *hrt = (struct ptlrpc_hr_thread *)arg;
2595 struct ptlrpc_hr_partition *hrp = hrt->hrt_partition;
2596 CFS_LIST_HEAD (replies);
2597 char threadname[20];
2600 snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
2601 hrp->hrp_cpt, hrt->hrt_id);
2602 cfs_daemonize_ctxt(threadname);
2604 rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
2606 CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
2607 threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
2610 cfs_atomic_inc(&hrp->hrp_nstarted);
2611 cfs_waitq_signal(&ptlrpc_hr.hr_waitq);
2613 while (!ptlrpc_hr.hr_stopping) {
2614 l_wait_condition(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
2616 while (!cfs_list_empty(&replies)) {
2617 struct ptlrpc_reply_state *rs;
2619 rs = cfs_list_entry(replies.prev,
2620 struct ptlrpc_reply_state,
2622 cfs_list_del_init(&rs->rs_list);
2623 ptlrpc_handle_rs(rs);
2627 cfs_atomic_inc(&hrp->hrp_nstopped);
2628 cfs_waitq_signal(&ptlrpc_hr.hr_waitq);
2633 static void ptlrpc_stop_hr_threads(void)
2635 struct ptlrpc_hr_partition *hrp;
2639 ptlrpc_hr.hr_stopping = 1;
2641 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2642 if (hrp->hrp_thrs == NULL)
2643 continue; /* uninitialized */
2644 for (j = 0; j < hrp->hrp_nthrs; j++)
2645 cfs_waitq_broadcast(&hrp->hrp_thrs[j].hrt_waitq);
2648 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2649 if (hrp->hrp_thrs == NULL)
2650 continue; /* uninitialized */
2651 cfs_wait_event(ptlrpc_hr.hr_waitq,
2652 cfs_atomic_read(&hrp->hrp_nstopped) ==
2653 cfs_atomic_read(&hrp->hrp_nstarted));
2657 static int ptlrpc_start_hr_threads(void)
2659 struct ptlrpc_hr_partition *hrp;
2664 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2667 for (j = 0; j < hrp->hrp_nthrs; j++) {
2668 rc = cfs_create_thread(ptlrpc_hr_main,
2670 CLONE_VM | CLONE_FILES);
2674 cfs_wait_event(ptlrpc_hr.hr_waitq,
2675 cfs_atomic_read(&hrp->hrp_nstarted) == j);
2679 CERROR("Reply handling thread %d:%d Failed on starting: "
2680 "rc = %d\n", i, j, rc);
2681 ptlrpc_stop_hr_threads();
2687 static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
2689 struct l_wait_info lwi = { 0 };
2690 struct ptlrpc_thread *thread;
2691 CFS_LIST_HEAD (zombie);
2695 CDEBUG(D_INFO, "Stopping threads for service %s\n",
2696 svcpt->scp_service->srv_name);
2698 spin_lock(&svcpt->scp_lock);
2699 /* let the thread know that we would like it to stop asap */
2700 list_for_each_entry(thread, &svcpt->scp_threads, t_link) {
2701 CDEBUG(D_INFO, "Stopping thread %s #%u\n",
2702 svcpt->scp_service->srv_thread_name, thread->t_id);
2703 thread_add_flags(thread, SVC_STOPPING);
2706 cfs_waitq_broadcast(&svcpt->scp_waitq);
2708 while (!cfs_list_empty(&svcpt->scp_threads)) {
2709 thread = cfs_list_entry(svcpt->scp_threads.next,
2710 struct ptlrpc_thread, t_link);
2711 if (thread_is_stopped(thread)) {
2712 cfs_list_del(&thread->t_link);
2713 cfs_list_add(&thread->t_link, &zombie);
2716 spin_unlock(&svcpt->scp_lock);
2718 CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
2719 svcpt->scp_service->srv_thread_name, thread->t_id);
2720 l_wait_event(thread->t_ctl_waitq,
2721 thread_is_stopped(thread), &lwi);
2723 spin_lock(&svcpt->scp_lock);
2726 spin_unlock(&svcpt->scp_lock);
2728 while (!cfs_list_empty(&zombie)) {
2729 thread = cfs_list_entry(zombie.next,
2730 struct ptlrpc_thread, t_link);
2731 cfs_list_del(&thread->t_link);
2732 OBD_FREE_PTR(thread);
2738 * Stops all threads of a particular service \a svc
2740 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2742 struct ptlrpc_service_part *svcpt;
2746 ptlrpc_service_for_each_part(svcpt, i, svc) {
2747 if (svcpt->scp_service != NULL)
2748 ptlrpc_svcpt_stop_threads(svcpt);
2753 EXPORT_SYMBOL(ptlrpc_stop_all_threads);
2755 int ptlrpc_start_threads(struct ptlrpc_service *svc)
2762 /* We require 2 threads min, see note in ptlrpc_server_handle_request */
2763 LASSERT(svc->srv_nthrs_cpt_init >= PTLRPC_NTHRS_INIT);
2765 for (i = 0; i < svc->srv_ncpts; i++) {
2766 for (j = 0; j < svc->srv_nthrs_cpt_init; j++) {
2767 rc = ptlrpc_start_thread(svc->srv_parts[i], 1);
2773 /* We have enough threads, don't start more. b=15759 */
2780 CERROR("cannot start %s thread #%d_%d: rc %d\n",
2781 svc->srv_thread_name, i, j, rc);
2782 ptlrpc_stop_all_threads(svc);
2785 EXPORT_SYMBOL(ptlrpc_start_threads);
2787 int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
2789 struct l_wait_info lwi = { 0 };
2790 struct ptlrpc_thread *thread;
2791 struct ptlrpc_service *svc;
2795 LASSERT(svcpt != NULL);
2797 svc = svcpt->scp_service;
2799 CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n",
2800 svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running,
2801 svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit);
2804 if (unlikely(svc->srv_is_stopping))
2807 if (!ptlrpc_threads_increasable(svcpt) ||
2808 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
2809 svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1))
2812 OBD_CPT_ALLOC_PTR(thread, svc->srv_cptable, svcpt->scp_cpt);
2815 cfs_waitq_init(&thread->t_ctl_waitq);
2817 spin_lock(&svcpt->scp_lock);
2818 if (!ptlrpc_threads_increasable(svcpt)) {
2819 spin_unlock(&svcpt->scp_lock);
2820 OBD_FREE_PTR(thread);
2824 if (svcpt->scp_nthrs_starting != 0) {
2825 /* serialize starting because some modules (obdfilter)
2826 * might require unique and contiguous t_id */
2827 LASSERT(svcpt->scp_nthrs_starting == 1);
2828 spin_unlock(&svcpt->scp_lock);
2829 OBD_FREE_PTR(thread);
2831 CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n",
2832 svc->srv_thread_name, svcpt->scp_thr_nextid);
2837 CDEBUG(D_INFO, "Creating thread %s #%d race, retry later\n",
2838 svc->srv_thread_name, svcpt->scp_thr_nextid);
2842 svcpt->scp_nthrs_starting++;
2843 thread->t_id = svcpt->scp_thr_nextid++;
2844 thread_add_flags(thread, SVC_STARTING);
2845 thread->t_svcpt = svcpt;
2847 cfs_list_add(&thread->t_link, &svcpt->scp_threads);
2848 spin_unlock(&svcpt->scp_lock);
2850 if (svcpt->scp_cpt >= 0) {
2851 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s%02d_%03d",
2852 svc->srv_thread_name, svcpt->scp_cpt, thread->t_id);
2854 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s_%04d",
2855 svc->srv_thread_name, thread->t_id);
2858 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name);
2860 * CLONE_VM and CLONE_FILES just avoid a needless copy, because we
2861 * just drop the VM and FILES in cfs_daemonize_ctxt() right away.
2863 rc = cfs_create_thread(ptlrpc_main, thread, CFS_DAEMON_FLAGS);
2865 CERROR("cannot start thread '%s': rc %d\n",
2866 thread->t_name, rc);
2867 spin_lock(&svcpt->scp_lock);
2868 cfs_list_del(&thread->t_link);
2869 --svcpt->scp_nthrs_starting;
2870 spin_unlock(&svcpt->scp_lock);
2872 OBD_FREE(thread, sizeof(*thread));
2879 l_wait_event(thread->t_ctl_waitq,
2880 thread_is_running(thread) || thread_is_stopped(thread),
2883 rc = thread_is_stopped(thread) ? thread->t_id : 0;
2887 int ptlrpc_hr_init(void)
2889 struct ptlrpc_hr_partition *hrp;
2890 struct ptlrpc_hr_thread *hrt;
2896 memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
2897 ptlrpc_hr.hr_cpt_table = cfs_cpt_table;
2899 ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
2901 if (ptlrpc_hr.hr_partitions == NULL)
2904 cfs_waitq_init(&ptlrpc_hr.hr_waitq);
2906 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2909 cfs_atomic_set(&hrp->hrp_nstarted, 0);
2910 cfs_atomic_set(&hrp->hrp_nstopped, 0);
2912 hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, i);
2913 hrp->hrp_nthrs /= cfs_cpu_ht_nsiblings(0);
2915 LASSERT(hrp->hrp_nthrs > 0);
2916 OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, i,
2917 hrp->hrp_nthrs * sizeof(*hrt));
2918 if (hrp->hrp_thrs == NULL)
2919 GOTO(out, rc = -ENOMEM);
2921 for (j = 0; j < hrp->hrp_nthrs; j++) {
2922 hrt = &hrp->hrp_thrs[j];
2925 hrt->hrt_partition = hrp;
2926 cfs_waitq_init(&hrt->hrt_waitq);
2927 spin_lock_init(&hrt->hrt_lock);
2928 CFS_INIT_LIST_HEAD(&hrt->hrt_queue);
2932 rc = ptlrpc_start_hr_threads();
2939 void ptlrpc_hr_fini(void)
2941 struct ptlrpc_hr_partition *hrp;
2944 if (ptlrpc_hr.hr_partitions == NULL)
2947 ptlrpc_stop_hr_threads();
2949 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2950 if (hrp->hrp_thrs != NULL) {
2951 OBD_FREE(hrp->hrp_thrs,
2952 hrp->hrp_nthrs * sizeof(hrp->hrp_thrs[0]));
2956 cfs_percpt_free(ptlrpc_hr.hr_partitions);
2957 ptlrpc_hr.hr_partitions = NULL;
2960 #endif /* __KERNEL__ */
2963 * Wait until all already scheduled replies are processed.
2965 static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
2969 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
2972 rc = l_wait_event(svcpt->scp_waitq,
2973 cfs_atomic_read(&svcpt->scp_nreps_difficult) == 0, &lwi);
2976 CWARN("Unexpectedly long timeout %s %p\n",
2977 svcpt->scp_service->srv_name, svcpt->scp_service);
2982 ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
2984 struct ptlrpc_service_part *svcpt;
2987 /* early disarm AT timer... */
2988 ptlrpc_service_for_each_part(svcpt, i, svc) {
2989 if (svcpt->scp_service != NULL)
2990 cfs_timer_disarm(&svcpt->scp_at_timer);
2995 ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
2997 struct ptlrpc_service_part *svcpt;
2998 struct ptlrpc_request_buffer_desc *rqbd;
2999 struct l_wait_info lwi;
3003 /* All history will be culled when the next request buffer is
3004 * freed in ptlrpc_service_purge_all() */
3005 svc->srv_hist_nrqbds_cpt_max = 0;
3007 rc = LNetClearLazyPortal(svc->srv_req_portal);
3010 ptlrpc_service_for_each_part(svcpt, i, svc) {
3011 if (svcpt->scp_service == NULL)
3014 /* Unlink all the request buffers. This forces a 'final'
3015 * event with its 'unlink' flag set for each posted rqbd */
3016 cfs_list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
3018 rc = LNetMDUnlink(rqbd->rqbd_md_h);
3019 LASSERT(rc == 0 || rc == -ENOENT);
3023 ptlrpc_service_for_each_part(svcpt, i, svc) {
3024 if (svcpt->scp_service == NULL)
3027 /* Wait for the network to release any buffers
3028 * it's currently filling */
3029 spin_lock(&svcpt->scp_lock);
3030 while (svcpt->scp_nrqbds_posted != 0) {
3031 spin_unlock(&svcpt->scp_lock);
3032 /* Network access will complete in finite time but
3033 * the HUGE timeout lets us CWARN for visibility
3034 * of sluggish NALs */
3035 lwi = LWI_TIMEOUT_INTERVAL(
3036 cfs_time_seconds(LONG_UNLINK),
3037 cfs_time_seconds(1), NULL, NULL);
3038 rc = l_wait_event(svcpt->scp_waitq,
3039 svcpt->scp_nrqbds_posted == 0, &lwi);
3040 if (rc == -ETIMEDOUT) {
3041 CWARN("Service %s waiting for "
3042 "request buffers\n",
3043 svcpt->scp_service->srv_name);
3045 spin_lock(&svcpt->scp_lock);
3047 spin_unlock(&svcpt->scp_lock);
3052 ptlrpc_service_purge_all(struct ptlrpc_service *svc)
3054 struct ptlrpc_service_part *svcpt;
3055 struct ptlrpc_request_buffer_desc *rqbd;
3056 struct ptlrpc_request *req;
3057 struct ptlrpc_reply_state *rs;
3060 ptlrpc_service_for_each_part(svcpt, i, svc) {
3061 if (svcpt->scp_service == NULL)
3064 spin_lock(&svcpt->scp_rep_lock);
3065 while (!cfs_list_empty(&svcpt->scp_rep_active)) {
3066 rs = cfs_list_entry(svcpt->scp_rep_active.next,
3067 struct ptlrpc_reply_state, rs_list);
3068 spin_lock(&rs->rs_lock);
3069 ptlrpc_schedule_difficult_reply(rs);
3070 spin_unlock(&rs->rs_lock);
3072 spin_unlock(&svcpt->scp_rep_lock);
3074 /* purge the request queue. NB No new replies (rqbds
3075 * all unlinked) and no service threads, so I'm the only
3076 * thread noodling the request queue now */
3077 while (!cfs_list_empty(&svcpt->scp_req_incoming)) {
3078 req = cfs_list_entry(svcpt->scp_req_incoming.next,
3079 struct ptlrpc_request, rq_list);
3081 cfs_list_del(&req->rq_list);
3082 svcpt->scp_nreqs_incoming--;
3083 ptlrpc_server_finish_request(svcpt, req);
3086 while (ptlrpc_server_request_pending(svcpt, true)) {
3087 req = ptlrpc_server_request_get(svcpt, true);
3088 ptlrpc_server_finish_active_request(svcpt, req);
3091 LASSERT(cfs_list_empty(&svcpt->scp_rqbd_posted));
3092 LASSERT(svcpt->scp_nreqs_incoming == 0);
3093 LASSERT(svcpt->scp_nreqs_active == 0);
3094 /* history should have been culled by
3095 * ptlrpc_server_finish_request */
3096 LASSERT(svcpt->scp_hist_nrqbds == 0);
3098 /* Now free all the request buffers since nothing
3099 * references them any more... */
3101 while (!cfs_list_empty(&svcpt->scp_rqbd_idle)) {
3102 rqbd = cfs_list_entry(svcpt->scp_rqbd_idle.next,
3103 struct ptlrpc_request_buffer_desc,
3105 ptlrpc_free_rqbd(rqbd);
3107 ptlrpc_wait_replies(svcpt);
3109 while (!cfs_list_empty(&svcpt->scp_rep_idle)) {
3110 rs = cfs_list_entry(svcpt->scp_rep_idle.next,
3111 struct ptlrpc_reply_state,
3113 cfs_list_del(&rs->rs_list);
3114 OBD_FREE_LARGE(rs, svc->srv_max_reply_size);
3120 ptlrpc_service_free(struct ptlrpc_service *svc)
3122 struct ptlrpc_service_part *svcpt;
3123 struct ptlrpc_at_array *array;
3126 ptlrpc_service_for_each_part(svcpt, i, svc) {
3127 if (svcpt->scp_service == NULL)
3130 /* In case somebody rearmed this in the meantime */
3131 cfs_timer_disarm(&svcpt->scp_at_timer);
3132 array = &svcpt->scp_at_array;
3134 if (array->paa_reqs_array != NULL) {
3135 OBD_FREE(array->paa_reqs_array,
3136 sizeof(cfs_list_t) * array->paa_size);
3137 array->paa_reqs_array = NULL;
3140 if (array->paa_reqs_count != NULL) {
3141 OBD_FREE(array->paa_reqs_count,
3142 sizeof(__u32) * array->paa_size);
3143 array->paa_reqs_count = NULL;
3147 ptlrpc_service_for_each_part(svcpt, i, svc)
3148 OBD_FREE_PTR(svcpt);
3150 if (svc->srv_cpts != NULL)
3151 cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts);
3153 OBD_FREE(svc, offsetof(struct ptlrpc_service,
3154 srv_parts[svc->srv_ncpts]));
3157 int ptlrpc_unregister_service(struct ptlrpc_service *service)
3161 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
3163 service->srv_is_stopping = 1;
3165 mutex_lock(&ptlrpc_all_services_mutex);
3166 cfs_list_del_init(&service->srv_list);
3167 mutex_unlock(&ptlrpc_all_services_mutex);
3169 ptlrpc_service_del_atimer(service);
3170 ptlrpc_stop_all_threads(service);
3172 ptlrpc_service_unlink_rqbd(service);
3173 ptlrpc_service_purge_all(service);
3174 ptlrpc_service_nrs_cleanup(service);
3176 ptlrpc_lprocfs_unregister_service(service);
3178 ptlrpc_service_free(service);
3182 EXPORT_SYMBOL(ptlrpc_unregister_service);
3185 * Returns 0 if the service is healthy.
3187 * Right now, it just checks to make sure that requests aren't languishing
3188 * in the queue. We'll use this health check to govern whether a node needs
3189 * to be shot, so it's intentionally non-aggressive. */
3190 int ptlrpc_svcpt_health_check(struct ptlrpc_service_part *svcpt)
3192 struct ptlrpc_request *request = NULL;
3193 struct timeval right_now;
3196 cfs_gettimeofday(&right_now);
3198 spin_lock(&svcpt->scp_req_lock);
3199 /* How long has the next entry been waiting? */
3200 if (ptlrpc_server_high_pending(svcpt, true))
3201 request = ptlrpc_nrs_req_peek_nolock(svcpt, true);
3202 else if (ptlrpc_server_normal_pending(svcpt, true))
3203 request = ptlrpc_nrs_req_peek_nolock(svcpt, false);
3205 if (request == NULL) {
3206 spin_unlock(&svcpt->scp_req_lock);
3210 timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
3211 spin_unlock(&svcpt->scp_req_lock);
3213 if ((timediff / ONE_MILLION) >
3214 (AT_OFF ? obd_timeout * 3 / 2 : at_max)) {
3215 CERROR("%s: unhealthy - request has been waiting %lds\n",
3216 svcpt->scp_service->srv_name, timediff / ONE_MILLION);
3224 ptlrpc_service_health_check(struct ptlrpc_service *svc)
3226 struct ptlrpc_service_part *svcpt;
3232 ptlrpc_service_for_each_part(svcpt, i, svc) {
3233 int rc = ptlrpc_svcpt_health_check(svcpt);
3240 EXPORT_SYMBOL(ptlrpc_service_health_check);