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_is_locked(&rs->rs_svcpt->scp_rep_lock));
408 LASSERT(spin_is_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 ptlrpc_request_cache_free(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 reqcopy = ptlrpc_request_cache_alloc(__GFP_IO);
1346 if (reqcopy == NULL)
1348 OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1350 GOTO(out_free, rc = -ENOMEM);
1353 reqcopy->rq_reply_state = NULL;
1354 reqcopy->rq_rep_swab_mask = 0;
1355 reqcopy->rq_pack_bulk = 0;
1356 reqcopy->rq_pack_udesc = 0;
1357 reqcopy->rq_packed_final = 0;
1358 sptlrpc_svc_ctx_addref(reqcopy);
1359 /* We only need the reqmsg for the magic */
1360 reqcopy->rq_reqmsg = reqmsg;
1361 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1363 LASSERT(cfs_atomic_read(&req->rq_refcount));
1364 /** if it is last refcount then early reply isn't needed */
1365 if (cfs_atomic_read(&req->rq_refcount) == 1) {
1366 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
1367 "abort sending early reply\n");
1368 GOTO(out, rc = -EINVAL);
1371 /* Connection ref */
1372 reqcopy->rq_export = class_conn2export(
1373 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1374 if (reqcopy->rq_export == NULL)
1375 GOTO(out, rc = -ENODEV);
1378 class_export_rpc_inc(reqcopy->rq_export);
1379 if (reqcopy->rq_export->exp_obd &&
1380 reqcopy->rq_export->exp_obd->obd_fail)
1381 GOTO(out_put, rc = -ENODEV);
1383 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1387 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1390 /* Adjust our own deadline to what we told the client */
1391 req->rq_deadline = newdl;
1392 req->rq_early_count++; /* number sent, server side */
1394 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1397 /* Free the (early) reply state from lustre_pack_reply.
1398 (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
1399 ptlrpc_req_drop_rs(reqcopy);
1402 class_export_rpc_dec(reqcopy->rq_export);
1403 class_export_put(reqcopy->rq_export);
1405 sptlrpc_svc_ctx_decref(reqcopy);
1406 OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1408 ptlrpc_request_cache_free(reqcopy);
1412 /* Send early replies to everybody expiring within at_early_margin
1413 asking for at_extra time */
1414 static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1416 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1417 struct ptlrpc_request *rq, *n;
1418 cfs_list_t work_list;
1421 time_t now = cfs_time_current_sec();
1422 cfs_duration_t delay;
1423 int first, counter = 0;
1426 spin_lock(&svcpt->scp_at_lock);
1427 if (svcpt->scp_at_check == 0) {
1428 spin_unlock(&svcpt->scp_at_lock);
1431 delay = cfs_time_sub(cfs_time_current(), svcpt->scp_at_checktime);
1432 svcpt->scp_at_check = 0;
1434 if (array->paa_count == 0) {
1435 spin_unlock(&svcpt->scp_at_lock);
1439 /* The timer went off, but maybe the nearest rpc already completed. */
1440 first = array->paa_deadline - now;
1441 if (first > at_early_margin) {
1442 /* We've still got plenty of time. Reset the timer. */
1443 ptlrpc_at_set_timer(svcpt);
1444 spin_unlock(&svcpt->scp_at_lock);
1448 /* We're close to a timeout, and we don't know how much longer the
1449 server will take. Send early replies to everyone expiring soon. */
1450 CFS_INIT_LIST_HEAD(&work_list);
1452 index = (unsigned long)array->paa_deadline % array->paa_size;
1453 count = array->paa_count;
1455 count -= array->paa_reqs_count[index];
1456 cfs_list_for_each_entry_safe(rq, n,
1457 &array->paa_reqs_array[index],
1459 if (rq->rq_deadline > now + at_early_margin) {
1460 /* update the earliest deadline */
1461 if (deadline == -1 ||
1462 rq->rq_deadline < deadline)
1463 deadline = rq->rq_deadline;
1467 ptlrpc_at_remove_timed(rq);
1469 * ptlrpc_server_drop_request() may drop
1470 * refcount to 0 already. Let's check this and
1471 * don't add entry to work_list
1473 if (likely(cfs_atomic_inc_not_zero(&rq->rq_refcount)))
1474 cfs_list_add(&rq->rq_timed_list, &work_list);
1478 if (++index >= array->paa_size)
1481 array->paa_deadline = deadline;
1482 /* we have a new earliest deadline, restart the timer */
1483 ptlrpc_at_set_timer(svcpt);
1485 spin_unlock(&svcpt->scp_at_lock);
1487 CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
1488 "replies\n", first, at_extra, counter);
1490 /* We're already past request deadlines before we even get a
1491 chance to send early replies */
1492 LCONSOLE_WARN("%s: This server is not able to keep up with "
1493 "request traffic (cpu-bound).\n",
1494 svcpt->scp_service->srv_name);
1495 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, "
1496 "delay="CFS_DURATION_T"(jiff)\n",
1497 counter, svcpt->scp_nreqs_incoming,
1498 svcpt->scp_nreqs_active,
1499 at_get(&svcpt->scp_at_estimate), delay);
1502 /* we took additional refcount so entries can't be deleted from list, no
1503 * locking is needed */
1504 while (!cfs_list_empty(&work_list)) {
1505 rq = cfs_list_entry(work_list.next, struct ptlrpc_request,
1507 cfs_list_del_init(&rq->rq_timed_list);
1509 if (ptlrpc_at_send_early_reply(rq) == 0)
1510 ptlrpc_at_add_timed(rq);
1512 ptlrpc_server_drop_request(rq);
1515 RETURN(1); /* return "did_something" for liblustre */
1519 * Put the request to the export list if the request may become
1520 * a high priority one.
1522 static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt,
1523 struct ptlrpc_request *req)
1528 if (svcpt->scp_service->srv_ops.so_hpreq_handler) {
1529 rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
1534 if (req->rq_export && req->rq_ops) {
1535 /* Perform request specific check. We should do this check
1536 * before the request is added into exp_hp_rpcs list otherwise
1537 * it may hit swab race at LU-1044. */
1538 if (req->rq_ops->hpreq_check) {
1539 rc = req->rq_ops->hpreq_check(req);
1541 * XXX: Out of all current
1542 * ptlrpc_hpreq_ops::hpreq_check(), only
1543 * ldlm_cancel_hpreq_check() can return an error code;
1544 * other functions assert in similar places, which seems
1545 * odd. What also does not seem right is that handlers
1546 * for those RPCs do not assert on the same checks, but
1547 * rather handle the error cases. e.g. see
1548 * ost_rw_hpreq_check(), and ost_brw_read(),
1553 LASSERT(rc == 0 || rc == 1);
1556 spin_lock_bh(&req->rq_export->exp_rpc_lock);
1557 cfs_list_add(&req->rq_exp_list,
1558 &req->rq_export->exp_hp_rpcs);
1559 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1562 ptlrpc_nrs_req_initialize(svcpt, req, rc);
1567 /** Remove the request from the export list. */
1568 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req)
1571 if (req->rq_export && req->rq_ops) {
1572 /* refresh lock timeout again so that client has more
1573 * room to send lock cancel RPC. */
1574 if (req->rq_ops->hpreq_fini)
1575 req->rq_ops->hpreq_fini(req);
1577 spin_lock_bh(&req->rq_export->exp_rpc_lock);
1578 cfs_list_del_init(&req->rq_exp_list);
1579 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1584 static int ptlrpc_hpreq_check(struct ptlrpc_request *req)
1589 static struct ptlrpc_hpreq_ops ptlrpc_hpreq_common = {
1590 .hpreq_check = ptlrpc_hpreq_check,
1593 /* Hi-Priority RPC check by RPC operation code. */
1594 int ptlrpc_hpreq_handler(struct ptlrpc_request *req)
1596 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1598 /* Check for export to let only reconnects for not yet evicted
1599 * export to become a HP rpc. */
1600 if ((req->rq_export != NULL) &&
1601 (opc == OBD_PING || opc == MDS_CONNECT || opc == OST_CONNECT))
1602 req->rq_ops = &ptlrpc_hpreq_common;
1606 EXPORT_SYMBOL(ptlrpc_hpreq_handler);
1608 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1609 struct ptlrpc_request *req)
1614 rc = ptlrpc_server_hpreq_init(svcpt, req);
1618 ptlrpc_nrs_req_add(svcpt, req, !!rc);
1624 * Allow to handle high priority request
1625 * User can call it w/o any lock but need to hold
1626 * ptlrpc_service_part::scp_req_lock to get reliable result
1628 static bool ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1631 int running = svcpt->scp_nthrs_running;
1633 if (!nrs_svcpt_has_hp(svcpt))
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)
1647 if (svcpt->scp_nreqs_active >= running - 1)
1650 if (svcpt->scp_nhreqs_active == 0)
1653 return !ptlrpc_nrs_req_pending_nolock(svcpt, false) ||
1654 svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1657 static bool ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1660 return ptlrpc_server_allow_high(svcpt, force) &&
1661 ptlrpc_nrs_req_pending_nolock(svcpt, true);
1665 * Only allow normal priority requests on a service that has a high-priority
1666 * queue if forced (i.e. cleanup), if there are other high priority requests
1667 * already being processed (i.e. those threads can service more high-priority
1668 * requests), or if there are enough idle threads that a later thread can do
1669 * a high priority request.
1670 * User can call it w/o any lock but need to hold
1671 * ptlrpc_service_part::scp_req_lock to get reliable result
1673 static bool ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1676 int running = svcpt->scp_nthrs_running;
1678 if (1) /* always allow to handle normal request for liblustre */
1681 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1682 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1683 /* leave just 1 thread for normal RPCs */
1684 running = PTLRPC_NTHRS_INIT;
1685 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1690 svcpt->scp_nreqs_active < running - 2)
1693 if (svcpt->scp_nreqs_active >= running - 1)
1696 return svcpt->scp_nhreqs_active > 0 || !nrs_svcpt_has_hp(svcpt);
1699 static bool ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1702 return ptlrpc_server_allow_normal(svcpt, force) &&
1703 ptlrpc_nrs_req_pending_nolock(svcpt, false);
1707 * Returns true if there are requests available in incoming
1708 * request queue for processing and it is allowed to fetch them.
1709 * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1710 * to get reliable result
1711 * \see ptlrpc_server_allow_normal
1712 * \see ptlrpc_server_allow high
1715 ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt, bool force)
1717 return ptlrpc_server_high_pending(svcpt, force) ||
1718 ptlrpc_server_normal_pending(svcpt, force);
1722 * Fetch a request for processing from queue of unprocessed requests.
1723 * Favors high-priority requests.
1724 * Returns a pointer to fetched request.
1726 static struct ptlrpc_request *
1727 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force)
1729 struct ptlrpc_request *req = NULL;
1732 spin_lock(&svcpt->scp_req_lock);
1734 /* !@%$# liblustre only has 1 thread */
1735 if (cfs_atomic_read(&svcpt->scp_nreps_difficult) != 0) {
1736 spin_unlock(&svcpt->scp_req_lock);
1741 if (ptlrpc_server_high_pending(svcpt, force)) {
1742 req = ptlrpc_nrs_req_get_nolock(svcpt, true, force);
1744 svcpt->scp_hreq_count++;
1749 if (ptlrpc_server_normal_pending(svcpt, force)) {
1750 req = ptlrpc_nrs_req_get_nolock(svcpt, false, force);
1752 svcpt->scp_hreq_count = 0;
1757 spin_unlock(&svcpt->scp_req_lock);
1761 svcpt->scp_nreqs_active++;
1763 svcpt->scp_nhreqs_active++;
1765 spin_unlock(&svcpt->scp_req_lock);
1767 if (likely(req->rq_export))
1768 class_export_rpc_inc(req->rq_export);
1774 * Handle freshly incoming reqs, add to timed early reply list,
1775 * pass on to regular request queue.
1776 * All incoming requests pass through here before getting into
1777 * ptlrpc_server_handle_req later on.
1780 ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt,
1781 struct ptlrpc_thread *thread)
1783 struct ptlrpc_service *svc = svcpt->scp_service;
1784 struct ptlrpc_request *req;
1789 spin_lock(&svcpt->scp_lock);
1790 if (cfs_list_empty(&svcpt->scp_req_incoming)) {
1791 spin_unlock(&svcpt->scp_lock);
1795 req = cfs_list_entry(svcpt->scp_req_incoming.next,
1796 struct ptlrpc_request, rq_list);
1797 cfs_list_del_init(&req->rq_list);
1798 svcpt->scp_nreqs_incoming--;
1799 /* Consider this still a "queued" request as far as stats are
1801 spin_unlock(&svcpt->scp_lock);
1803 /* go through security check/transform */
1804 rc = sptlrpc_svc_unwrap_request(req);
1808 case SECSVC_COMPLETE:
1809 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1818 * for null-flavored rpc, msg has been unpacked by sptlrpc, although
1819 * redo it wouldn't be harmful.
1821 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
1822 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
1824 CERROR("error unpacking request: ptl %d from %s "
1825 "x"LPU64"\n", svc->srv_req_portal,
1826 libcfs_id2str(req->rq_peer), req->rq_xid);
1831 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1833 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
1834 LPU64"\n", svc->srv_req_portal,
1835 libcfs_id2str(req->rq_peer), req->rq_xid);
1839 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
1840 lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
1841 CERROR("drop incoming rpc opc %u, x"LPU64"\n",
1842 cfs_fail_val, req->rq_xid);
1847 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1848 CERROR("wrong packet type received (type=%u) from %s\n",
1849 lustre_msg_get_type(req->rq_reqmsg),
1850 libcfs_id2str(req->rq_peer));
1854 switch(lustre_msg_get_opc(req->rq_reqmsg)) {
1857 req->rq_bulk_write = 1;
1861 case MGS_CONFIG_READ:
1862 req->rq_bulk_read = 1;
1866 CDEBUG(D_RPCTRACE, "got req x"LPU64"\n", req->rq_xid);
1868 req->rq_export = class_conn2export(
1869 lustre_msg_get_handle(req->rq_reqmsg));
1870 if (req->rq_export) {
1871 rc = ptlrpc_check_req(req);
1873 rc = sptlrpc_target_export_check(req->rq_export, req);
1875 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
1876 "illegal security flavor,");
1881 ptlrpc_update_export_timer(req->rq_export, 0);
1884 /* req_in handling should/must be fast */
1885 if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5)
1886 DEBUG_REQ(D_WARNING, req, "Slow req_in handling "CFS_DURATION_T"s",
1887 cfs_time_sub(cfs_time_current_sec(),
1888 req->rq_arrival_time.tv_sec));
1890 /* Set rpc server deadline and add it to the timed list */
1891 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1892 MSGHDR_AT_SUPPORT) ?
1893 /* The max time the client expects us to take */
1894 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1895 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1896 if (unlikely(deadline == 0)) {
1897 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1901 req->rq_svc_thread = thread;
1903 ptlrpc_at_add_timed(req);
1905 /* Move it over to the request processing queue */
1906 rc = ptlrpc_server_request_add(svcpt, req);
1910 cfs_waitq_signal(&svcpt->scp_waitq);
1914 ptlrpc_server_finish_request(svcpt, req);
1920 * Main incoming request handling logic.
1921 * Calls handler function from service to do actual processing.
1924 ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
1925 struct ptlrpc_thread *thread)
1927 struct ptlrpc_service *svc = svcpt->scp_service;
1928 struct ptlrpc_request *request;
1929 struct timeval work_start;
1930 struct timeval work_end;
1936 request = ptlrpc_server_request_get(svcpt, false);
1937 if (request == NULL)
1940 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
1941 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
1942 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
1943 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
1945 if (unlikely(fail_opc)) {
1946 if (request->rq_export && request->rq_ops)
1947 OBD_FAIL_TIMEOUT(fail_opc, 4);
1950 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
1952 if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
1953 libcfs_debug_dumplog();
1955 do_gettimeofday(&work_start);
1956 timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
1957 if (likely(svc->srv_stats != NULL)) {
1958 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
1960 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
1961 svcpt->scp_nreqs_incoming);
1962 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
1963 svcpt->scp_nreqs_active);
1964 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
1965 at_get(&svcpt->scp_at_estimate));
1968 rc = lu_context_init(&request->rq_session, LCT_SESSION | LCT_NOREF);
1970 CERROR("Failure to initialize session: %d\n", rc);
1973 request->rq_session.lc_thread = thread;
1974 request->rq_session.lc_cookie = 0x5;
1975 lu_context_enter(&request->rq_session);
1977 CDEBUG(D_NET, "got req "LPU64"\n", request->rq_xid);
1979 request->rq_svc_thread = thread;
1981 request->rq_svc_thread->t_env->le_ses = &request->rq_session;
1983 if (likely(request->rq_export)) {
1984 if (unlikely(ptlrpc_check_req(request)))
1986 ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
1989 /* Discard requests queued for longer than the deadline.
1990 The deadline is increased if we send an early reply. */
1991 if (cfs_time_current_sec() > request->rq_deadline) {
1992 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s"
1993 ": deadline "CFS_DURATION_T":"CFS_DURATION_T"s ago\n",
1994 libcfs_id2str(request->rq_peer),
1995 cfs_time_sub(request->rq_deadline,
1996 request->rq_arrival_time.tv_sec),
1997 cfs_time_sub(cfs_time_current_sec(),
1998 request->rq_deadline));
2002 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
2003 "%s:%s+%d:%d:x"LPU64":%s:%d\n", current_comm(),
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), request->rq_xid,
2009 libcfs_id2str(request->rq_peer),
2010 lustre_msg_get_opc(request->rq_reqmsg));
2012 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
2013 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
2015 rc = svc->srv_ops.so_req_handler(request);
2017 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
2020 lu_context_exit(&request->rq_session);
2021 lu_context_fini(&request->rq_session);
2023 if (unlikely(cfs_time_current_sec() > request->rq_deadline)) {
2024 DEBUG_REQ(D_WARNING, request, "Request took longer "
2025 "than estimated ("CFS_DURATION_T":"CFS_DURATION_T"s);"
2026 " client may timeout.",
2027 cfs_time_sub(request->rq_deadline,
2028 request->rq_arrival_time.tv_sec),
2029 cfs_time_sub(cfs_time_current_sec(),
2030 request->rq_deadline));
2033 do_gettimeofday(&work_end);
2034 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
2035 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
2036 "%s:%s+%d:%d:x"LPU64":%s:%d Request procesed in "
2037 "%ldus (%ldus total) trans "LPU64" rc %d/%d\n",
2039 (request->rq_export ?
2040 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2041 (request->rq_export ?
2042 cfs_atomic_read(&request->rq_export->exp_refcount) : -99),
2043 lustre_msg_get_status(request->rq_reqmsg),
2045 libcfs_id2str(request->rq_peer),
2046 lustre_msg_get_opc(request->rq_reqmsg),
2048 cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL),
2049 (request->rq_repmsg ?
2050 lustre_msg_get_transno(request->rq_repmsg) :
2051 request->rq_transno),
2053 (request->rq_repmsg ?
2054 lustre_msg_get_status(request->rq_repmsg) : -999));
2055 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
2056 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
2057 int opc = opcode_offset(op);
2058 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
2059 LASSERT(opc < LUSTRE_MAX_OPCODES);
2060 lprocfs_counter_add(svc->srv_stats,
2061 opc + EXTRA_MAX_OPCODES,
2065 if (unlikely(request->rq_early_count)) {
2066 DEBUG_REQ(D_ADAPTTO, request,
2067 "sent %d early replies before finishing in "
2069 request->rq_early_count,
2070 cfs_time_sub(work_end.tv_sec,
2071 request->rq_arrival_time.tv_sec));
2075 ptlrpc_server_finish_active_request(svcpt, request);
2081 * An internal function to process a single reply state object.
2084 ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
2086 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
2087 struct ptlrpc_service *svc = svcpt->scp_service;
2088 struct obd_export *exp;
2093 exp = rs->rs_export;
2095 LASSERT (rs->rs_difficult);
2096 LASSERT (rs->rs_scheduled);
2097 LASSERT (cfs_list_empty(&rs->rs_list));
2099 spin_lock(&exp->exp_lock);
2100 /* Noop if removed already */
2101 cfs_list_del_init (&rs->rs_exp_list);
2102 spin_unlock(&exp->exp_lock);
2104 /* The disk commit callback holds exp_uncommitted_replies_lock while it
2105 * iterates over newly committed replies, removing them from
2106 * exp_uncommitted_replies. It then drops this lock and schedules the
2107 * replies it found for handling here.
2109 * We can avoid contention for exp_uncommitted_replies_lock between the
2110 * HRT threads and further commit callbacks by checking rs_committed
2111 * which is set in the commit callback while it holds both
2112 * rs_lock and exp_uncommitted_reples.
2114 * If we see rs_committed clear, the commit callback _may_ not have
2115 * handled this reply yet and we race with it to grab
2116 * exp_uncommitted_replies_lock before removing the reply from
2117 * exp_uncommitted_replies. Note that if we lose the race and the
2118 * reply has already been removed, list_del_init() is a noop.
2120 * If we see rs_committed set, we know the commit callback is handling,
2121 * or has handled this reply since store reordering might allow us to
2122 * see rs_committed set out of sequence. But since this is done
2123 * holding rs_lock, we can be sure it has all completed once we hold
2124 * rs_lock, which we do right next.
2126 if (!rs->rs_committed) {
2127 spin_lock(&exp->exp_uncommitted_replies_lock);
2128 cfs_list_del_init(&rs->rs_obd_list);
2129 spin_unlock(&exp->exp_uncommitted_replies_lock);
2132 spin_lock(&rs->rs_lock);
2134 been_handled = rs->rs_handled;
2137 nlocks = rs->rs_nlocks; /* atomic "steal", but */
2138 rs->rs_nlocks = 0; /* locks still on rs_locks! */
2140 if (nlocks == 0 && !been_handled) {
2141 /* If we see this, we should already have seen the warning
2142 * in mds_steal_ack_locks() */
2143 CDEBUG(D_HA, "All locks stolen from rs %p x"LPD64".t"LPD64
2146 rs->rs_xid, rs->rs_transno, rs->rs_opc,
2147 libcfs_nid2str(exp->exp_connection->c_peer.nid));
2150 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
2151 spin_unlock(&rs->rs_lock);
2153 if (!been_handled && rs->rs_on_net) {
2154 LNetMDUnlink(rs->rs_md_h);
2155 /* Ignore return code; we're racing with completion */
2158 while (nlocks-- > 0)
2159 ldlm_lock_decref(&rs->rs_locks[nlocks],
2160 rs->rs_modes[nlocks]);
2162 spin_lock(&rs->rs_lock);
2165 rs->rs_scheduled = 0;
2167 if (!rs->rs_on_net) {
2169 spin_unlock(&rs->rs_lock);
2171 class_export_put (exp);
2172 rs->rs_export = NULL;
2173 ptlrpc_rs_decref (rs);
2174 if (cfs_atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
2175 svc->srv_is_stopping)
2176 cfs_waitq_broadcast(&svcpt->scp_waitq);
2180 /* still on the net; callback will schedule */
2181 spin_unlock(&rs->rs_lock);
2188 * Check whether given service has a reply available for processing
2191 * \param svc a ptlrpc service
2192 * \retval 0 no replies processed
2193 * \retval 1 one reply processed
2196 ptlrpc_server_handle_reply(struct ptlrpc_service_part *svcpt)
2198 struct ptlrpc_reply_state *rs = NULL;
2201 spin_lock(&svcpt->scp_rep_lock);
2202 if (!cfs_list_empty(&svcpt->scp_rep_queue)) {
2203 rs = cfs_list_entry(svcpt->scp_rep_queue.prev,
2204 struct ptlrpc_reply_state,
2206 cfs_list_del_init(&rs->rs_list);
2208 spin_unlock(&svcpt->scp_rep_lock);
2210 ptlrpc_handle_rs(rs);
2214 /* FIXME make use of timeout later */
2216 liblustre_check_services (void *arg)
2218 int did_something = 0;
2220 cfs_list_t *tmp, *nxt;
2223 /* I'm relying on being single threaded, not to have to lock
2224 * ptlrpc_all_services etc */
2225 cfs_list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
2226 struct ptlrpc_service *svc =
2227 cfs_list_entry (tmp, struct ptlrpc_service, srv_list);
2228 struct ptlrpc_service_part *svcpt;
2230 LASSERT(svc->srv_ncpts == 1);
2231 svcpt = svc->srv_parts[0];
2233 if (svcpt->scp_nthrs_running != 0) /* I've recursed */
2236 /* service threads can block for bulk, so this limits us
2237 * (arbitrarily) to recursing 1 stack frame per service.
2238 * Note that the problem with recursion is that we have to
2239 * unwind completely before our caller can resume. */
2241 svcpt->scp_nthrs_running++;
2244 rc = ptlrpc_server_handle_req_in(svcpt, NULL);
2245 rc |= ptlrpc_server_handle_reply(svcpt);
2246 rc |= ptlrpc_at_check_timed(svcpt);
2247 rc |= ptlrpc_server_handle_request(svcpt, NULL);
2248 rc |= (ptlrpc_server_post_idle_rqbds(svcpt) > 0);
2249 did_something |= rc;
2252 svcpt->scp_nthrs_running--;
2255 RETURN(did_something);
2257 #define ptlrpc_stop_all_threads(s) do {} while (0)
2259 #else /* __KERNEL__ */
2262 ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
2264 int avail = svcpt->scp_nrqbds_posted;
2265 int low_water = test_req_buffer_pressure ? 0 :
2266 svcpt->scp_service->srv_nbuf_per_group / 2;
2268 /* NB I'm not locking; just looking. */
2270 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
2271 * allowed the request history to grow out of control. We could put a
2272 * sanity check on that here and cull some history if we need the
2275 if (avail <= low_water)
2276 ptlrpc_grow_req_bufs(svcpt, 1);
2278 if (svcpt->scp_service->srv_stats) {
2279 lprocfs_counter_add(svcpt->scp_service->srv_stats,
2280 PTLRPC_REQBUF_AVAIL_CNTR, avail);
2285 ptlrpc_retry_rqbds(void *arg)
2287 struct ptlrpc_service_part *svcpt = (struct ptlrpc_service_part *)arg;
2289 svcpt->scp_rqbd_timeout = 0;
2294 ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
2296 return svcpt->scp_nreqs_active <
2297 svcpt->scp_nthrs_running - 1 -
2298 (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL);
2302 * allowed to create more threads
2303 * user can call it w/o any lock but need to hold
2304 * ptlrpc_service_part::scp_lock to get reliable result
2307 ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
2309 return svcpt->scp_nthrs_running +
2310 svcpt->scp_nthrs_starting <
2311 svcpt->scp_service->srv_nthrs_cpt_limit;
2315 * too many requests and allowed to create more threads
2318 ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
2320 return !ptlrpc_threads_enough(svcpt) &&
2321 ptlrpc_threads_increasable(svcpt);
2325 ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2327 return thread_is_stopping(thread) ||
2328 thread->t_svcpt->scp_service->srv_is_stopping;
2332 ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
2334 return !cfs_list_empty(&svcpt->scp_rqbd_idle) &&
2335 svcpt->scp_rqbd_timeout == 0;
2339 ptlrpc_at_check(struct ptlrpc_service_part *svcpt)
2341 return svcpt->scp_at_check;
2345 * requests wait on preprocessing
2346 * user can call it w/o any lock but need to hold
2347 * ptlrpc_service_part::scp_lock to get reliable result
2350 ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
2352 return !cfs_list_empty(&svcpt->scp_req_incoming);
2355 static __attribute__((__noinline__)) int
2356 ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
2357 struct ptlrpc_thread *thread)
2359 /* Don't exit while there are replies to be handled */
2360 struct l_wait_info lwi = LWI_TIMEOUT(svcpt->scp_rqbd_timeout,
2361 ptlrpc_retry_rqbds, svcpt);
2363 lc_watchdog_disable(thread->t_watchdog);
2367 l_wait_event_exclusive_head(svcpt->scp_waitq,
2368 ptlrpc_thread_stopping(thread) ||
2369 ptlrpc_server_request_incoming(svcpt) ||
2370 ptlrpc_server_request_pending(svcpt, false) ||
2371 ptlrpc_rqbd_pending(svcpt) ||
2372 ptlrpc_at_check(svcpt), &lwi);
2374 if (ptlrpc_thread_stopping(thread))
2377 lc_watchdog_touch(thread->t_watchdog,
2378 ptlrpc_server_get_timeout(svcpt));
2383 * Main thread body for service threads.
2384 * Waits in a loop waiting for new requests to process to appear.
2385 * Every time an incoming requests is added to its queue, a waitq
2386 * is woken up and one of the threads will handle it.
2388 static int ptlrpc_main(void *arg)
2390 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
2391 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2392 struct ptlrpc_service *svc = svcpt->scp_service;
2393 struct ptlrpc_reply_state *rs;
2394 #ifdef WITH_GROUP_INFO
2395 cfs_group_info_t *ginfo = NULL;
2398 int counter = 0, rc = 0;
2401 thread->t_pid = current_pid();
2402 unshare_fs_struct();
2404 /* NB: we will call cfs_cpt_bind() for all threads, because we
2405 * might want to run lustre server only on a subset of system CPUs,
2406 * in that case ->scp_cpt is CFS_CPT_ANY */
2407 rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
2409 CWARN("%s: failed to bind %s on CPT %d\n",
2410 svc->srv_name, thread->t_name, svcpt->scp_cpt);
2413 #ifdef WITH_GROUP_INFO
2414 ginfo = cfs_groups_alloc(0);
2420 cfs_set_current_groups(ginfo);
2421 cfs_put_group_info(ginfo);
2424 if (svc->srv_ops.so_thr_init != NULL) {
2425 rc = svc->srv_ops.so_thr_init(thread);
2436 rc = lu_context_init(&env->le_ctx,
2437 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2441 thread->t_env = env;
2442 env->le_ctx.lc_thread = thread;
2443 env->le_ctx.lc_cookie = 0x6;
2445 while (!cfs_list_empty(&svcpt->scp_rqbd_idle)) {
2446 rc = ptlrpc_server_post_idle_rqbds(svcpt);
2450 CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
2451 svc->srv_name, svcpt->scp_cpt, rc);
2455 /* Alloc reply state structure for this one */
2456 OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2462 spin_lock(&svcpt->scp_lock);
2464 LASSERT(thread_is_starting(thread));
2465 thread_clear_flags(thread, SVC_STARTING);
2467 LASSERT(svcpt->scp_nthrs_starting == 1);
2468 svcpt->scp_nthrs_starting--;
2470 /* SVC_STOPPING may already be set here if someone else is trying
2471 * to stop the service while this new thread has been dynamically
2472 * forked. We still set SVC_RUNNING to let our creator know that
2473 * we are now running, however we will exit as soon as possible */
2474 thread_add_flags(thread, SVC_RUNNING);
2475 svcpt->scp_nthrs_running++;
2476 spin_unlock(&svcpt->scp_lock);
2478 /* wake up our creator in case he's still waiting. */
2479 cfs_waitq_signal(&thread->t_ctl_waitq);
2481 thread->t_watchdog = lc_watchdog_add(ptlrpc_server_get_timeout(svcpt),
2484 spin_lock(&svcpt->scp_rep_lock);
2485 cfs_list_add(&rs->rs_list, &svcpt->scp_rep_idle);
2486 cfs_waitq_signal(&svcpt->scp_rep_waitq);
2487 spin_unlock(&svcpt->scp_rep_lock);
2489 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2490 svcpt->scp_nthrs_running);
2492 /* XXX maintain a list of all managed devices: insert here */
2493 while (!ptlrpc_thread_stopping(thread)) {
2494 if (ptlrpc_wait_event(svcpt, thread))
2497 ptlrpc_check_rqbd_pool(svcpt);
2499 if (ptlrpc_threads_need_create(svcpt)) {
2500 /* Ignore return code - we tried... */
2501 ptlrpc_start_thread(svcpt, 0);
2504 /* Process all incoming reqs before handling any */
2505 if (ptlrpc_server_request_incoming(svcpt)) {
2506 lu_context_enter(&env->le_ctx);
2508 ptlrpc_server_handle_req_in(svcpt, thread);
2509 lu_context_exit(&env->le_ctx);
2511 /* but limit ourselves in case of flood */
2512 if (counter++ < 100)
2517 if (ptlrpc_at_check(svcpt))
2518 ptlrpc_at_check_timed(svcpt);
2520 if (ptlrpc_server_request_pending(svcpt, false)) {
2521 lu_context_enter(&env->le_ctx);
2522 ptlrpc_server_handle_request(svcpt, thread);
2523 lu_context_exit(&env->le_ctx);
2526 if (ptlrpc_rqbd_pending(svcpt) &&
2527 ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
2528 /* I just failed to repost request buffers.
2529 * Wait for a timeout (unless something else
2530 * happens) before I try again */
2531 svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
2532 CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
2533 svcpt->scp_nrqbds_posted);
2537 lc_watchdog_delete(thread->t_watchdog);
2538 thread->t_watchdog = NULL;
2542 * deconstruct service specific state created by ptlrpc_start_thread()
2544 if (svc->srv_ops.so_thr_done != NULL)
2545 svc->srv_ops.so_thr_done(thread);
2548 lu_context_fini(&env->le_ctx);
2552 CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n",
2553 thread, thread->t_pid, thread->t_id, rc);
2555 spin_lock(&svcpt->scp_lock);
2556 if (thread_test_and_clear_flags(thread, SVC_STARTING))
2557 svcpt->scp_nthrs_starting--;
2559 if (thread_test_and_clear_flags(thread, SVC_RUNNING)) {
2560 /* must know immediately */
2561 svcpt->scp_nthrs_running--;
2565 thread_add_flags(thread, SVC_STOPPED);
2567 cfs_waitq_signal(&thread->t_ctl_waitq);
2568 spin_unlock(&svcpt->scp_lock);
2573 static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
2574 cfs_list_t *replies)
2578 spin_lock(&hrt->hrt_lock);
2580 cfs_list_splice_init(&hrt->hrt_queue, replies);
2581 result = ptlrpc_hr.hr_stopping || !cfs_list_empty(replies);
2583 spin_unlock(&hrt->hrt_lock);
2588 * Main body of "handle reply" function.
2589 * It processes acked reply states
2591 static int ptlrpc_hr_main(void *arg)
2593 struct ptlrpc_hr_thread *hrt = (struct ptlrpc_hr_thread *)arg;
2594 struct ptlrpc_hr_partition *hrp = hrt->hrt_partition;
2595 CFS_LIST_HEAD (replies);
2596 char threadname[20];
2599 snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
2600 hrp->hrp_cpt, hrt->hrt_id);
2601 unshare_fs_struct();
2603 rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
2605 CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
2606 threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
2609 cfs_atomic_inc(&hrp->hrp_nstarted);
2610 cfs_waitq_signal(&ptlrpc_hr.hr_waitq);
2612 while (!ptlrpc_hr.hr_stopping) {
2613 l_wait_condition(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
2615 while (!cfs_list_empty(&replies)) {
2616 struct ptlrpc_reply_state *rs;
2618 rs = cfs_list_entry(replies.prev,
2619 struct ptlrpc_reply_state,
2621 cfs_list_del_init(&rs->rs_list);
2622 ptlrpc_handle_rs(rs);
2626 cfs_atomic_inc(&hrp->hrp_nstopped);
2627 cfs_waitq_signal(&ptlrpc_hr.hr_waitq);
2632 static void ptlrpc_stop_hr_threads(void)
2634 struct ptlrpc_hr_partition *hrp;
2638 ptlrpc_hr.hr_stopping = 1;
2640 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2641 if (hrp->hrp_thrs == NULL)
2642 continue; /* uninitialized */
2643 for (j = 0; j < hrp->hrp_nthrs; j++)
2644 cfs_waitq_broadcast(&hrp->hrp_thrs[j].hrt_waitq);
2647 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2648 if (hrp->hrp_thrs == NULL)
2649 continue; /* uninitialized */
2650 cfs_wait_event(ptlrpc_hr.hr_waitq,
2651 cfs_atomic_read(&hrp->hrp_nstopped) ==
2652 cfs_atomic_read(&hrp->hrp_nstarted));
2656 static int ptlrpc_start_hr_threads(void)
2658 struct ptlrpc_hr_partition *hrp;
2663 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2666 for (j = 0; j < hrp->hrp_nthrs; j++) {
2667 struct ptlrpc_hr_thread *hrt = &hrp->hrp_thrs[j];
2668 rc = PTR_ERR(kthread_run(ptlrpc_hr_main,
2670 "ptlrpc_hr%02d_%03d",
2673 if (IS_ERR_VALUE(rc))
2676 cfs_wait_event(ptlrpc_hr.hr_waitq,
2677 cfs_atomic_read(&hrp->hrp_nstarted) == j);
2678 if (!IS_ERR_VALUE(rc))
2681 CERROR("Reply handling thread %d:%d Failed on starting: "
2682 "rc = %d\n", i, j, rc);
2683 ptlrpc_stop_hr_threads();
2689 static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
2691 struct l_wait_info lwi = { 0 };
2692 struct ptlrpc_thread *thread;
2693 CFS_LIST_HEAD (zombie);
2697 CDEBUG(D_INFO, "Stopping threads for service %s\n",
2698 svcpt->scp_service->srv_name);
2700 spin_lock(&svcpt->scp_lock);
2701 /* let the thread know that we would like it to stop asap */
2702 list_for_each_entry(thread, &svcpt->scp_threads, t_link) {
2703 CDEBUG(D_INFO, "Stopping thread %s #%u\n",
2704 svcpt->scp_service->srv_thread_name, thread->t_id);
2705 thread_add_flags(thread, SVC_STOPPING);
2708 cfs_waitq_broadcast(&svcpt->scp_waitq);
2710 while (!cfs_list_empty(&svcpt->scp_threads)) {
2711 thread = cfs_list_entry(svcpt->scp_threads.next,
2712 struct ptlrpc_thread, t_link);
2713 if (thread_is_stopped(thread)) {
2714 cfs_list_del(&thread->t_link);
2715 cfs_list_add(&thread->t_link, &zombie);
2718 spin_unlock(&svcpt->scp_lock);
2720 CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
2721 svcpt->scp_service->srv_thread_name, thread->t_id);
2722 l_wait_event(thread->t_ctl_waitq,
2723 thread_is_stopped(thread), &lwi);
2725 spin_lock(&svcpt->scp_lock);
2728 spin_unlock(&svcpt->scp_lock);
2730 while (!cfs_list_empty(&zombie)) {
2731 thread = cfs_list_entry(zombie.next,
2732 struct ptlrpc_thread, t_link);
2733 cfs_list_del(&thread->t_link);
2734 OBD_FREE_PTR(thread);
2740 * Stops all threads of a particular service \a svc
2742 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2744 struct ptlrpc_service_part *svcpt;
2748 ptlrpc_service_for_each_part(svcpt, i, svc) {
2749 if (svcpt->scp_service != NULL)
2750 ptlrpc_svcpt_stop_threads(svcpt);
2755 EXPORT_SYMBOL(ptlrpc_stop_all_threads);
2757 int ptlrpc_start_threads(struct ptlrpc_service *svc)
2764 /* We require 2 threads min, see note in ptlrpc_server_handle_request */
2765 LASSERT(svc->srv_nthrs_cpt_init >= PTLRPC_NTHRS_INIT);
2767 for (i = 0; i < svc->srv_ncpts; i++) {
2768 for (j = 0; j < svc->srv_nthrs_cpt_init; j++) {
2769 rc = ptlrpc_start_thread(svc->srv_parts[i], 1);
2775 /* We have enough threads, don't start more. b=15759 */
2782 CERROR("cannot start %s thread #%d_%d: rc %d\n",
2783 svc->srv_thread_name, i, j, rc);
2784 ptlrpc_stop_all_threads(svc);
2787 EXPORT_SYMBOL(ptlrpc_start_threads);
2789 int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
2791 struct l_wait_info lwi = { 0 };
2792 struct ptlrpc_thread *thread;
2793 struct ptlrpc_service *svc;
2797 LASSERT(svcpt != NULL);
2799 svc = svcpt->scp_service;
2801 CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n",
2802 svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running,
2803 svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit);
2806 if (unlikely(svc->srv_is_stopping))
2809 if (!ptlrpc_threads_increasable(svcpt) ||
2810 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
2811 svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1))
2814 OBD_CPT_ALLOC_PTR(thread, svc->srv_cptable, svcpt->scp_cpt);
2817 cfs_waitq_init(&thread->t_ctl_waitq);
2819 spin_lock(&svcpt->scp_lock);
2820 if (!ptlrpc_threads_increasable(svcpt)) {
2821 spin_unlock(&svcpt->scp_lock);
2822 OBD_FREE_PTR(thread);
2826 if (svcpt->scp_nthrs_starting != 0) {
2827 /* serialize starting because some modules (obdfilter)
2828 * might require unique and contiguous t_id */
2829 LASSERT(svcpt->scp_nthrs_starting == 1);
2830 spin_unlock(&svcpt->scp_lock);
2831 OBD_FREE_PTR(thread);
2833 CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n",
2834 svc->srv_thread_name, svcpt->scp_thr_nextid);
2839 CDEBUG(D_INFO, "Creating thread %s #%d race, retry later\n",
2840 svc->srv_thread_name, svcpt->scp_thr_nextid);
2844 svcpt->scp_nthrs_starting++;
2845 thread->t_id = svcpt->scp_thr_nextid++;
2846 thread_add_flags(thread, SVC_STARTING);
2847 thread->t_svcpt = svcpt;
2849 cfs_list_add(&thread->t_link, &svcpt->scp_threads);
2850 spin_unlock(&svcpt->scp_lock);
2852 if (svcpt->scp_cpt >= 0) {
2853 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s%02d_%03d",
2854 svc->srv_thread_name, svcpt->scp_cpt, thread->t_id);
2856 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s_%04d",
2857 svc->srv_thread_name, thread->t_id);
2860 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name);
2861 rc = PTR_ERR(kthread_run(ptlrpc_main, thread, thread->t_name));
2862 if (IS_ERR_VALUE(rc)) {
2863 CERROR("cannot start thread '%s': rc %d\n",
2864 thread->t_name, rc);
2865 spin_lock(&svcpt->scp_lock);
2866 --svcpt->scp_nthrs_starting;
2867 if (thread_is_stopping(thread)) {
2868 /* this ptlrpc_thread is being hanled
2869 * by ptlrpc_svcpt_stop_threads now
2871 thread_add_flags(thread, SVC_STOPPED);
2872 cfs_waitq_signal(&thread->t_ctl_waitq);
2873 spin_unlock(&svcpt->scp_lock);
2875 cfs_list_del(&thread->t_link);
2876 spin_unlock(&svcpt->scp_lock);
2877 OBD_FREE_PTR(thread);
2885 l_wait_event(thread->t_ctl_waitq,
2886 thread_is_running(thread) || thread_is_stopped(thread),
2889 rc = thread_is_stopped(thread) ? thread->t_id : 0;
2893 int ptlrpc_hr_init(void)
2895 struct ptlrpc_hr_partition *hrp;
2896 struct ptlrpc_hr_thread *hrt;
2902 memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
2903 ptlrpc_hr.hr_cpt_table = cfs_cpt_table;
2905 ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
2907 if (ptlrpc_hr.hr_partitions == NULL)
2910 cfs_waitq_init(&ptlrpc_hr.hr_waitq);
2912 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2915 cfs_atomic_set(&hrp->hrp_nstarted, 0);
2916 cfs_atomic_set(&hrp->hrp_nstopped, 0);
2918 hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, i);
2919 hrp->hrp_nthrs /= cfs_cpu_ht_nsiblings(0);
2921 LASSERT(hrp->hrp_nthrs > 0);
2922 OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, i,
2923 hrp->hrp_nthrs * sizeof(*hrt));
2924 if (hrp->hrp_thrs == NULL)
2925 GOTO(out, rc = -ENOMEM);
2927 for (j = 0; j < hrp->hrp_nthrs; j++) {
2928 hrt = &hrp->hrp_thrs[j];
2931 hrt->hrt_partition = hrp;
2932 cfs_waitq_init(&hrt->hrt_waitq);
2933 spin_lock_init(&hrt->hrt_lock);
2934 CFS_INIT_LIST_HEAD(&hrt->hrt_queue);
2938 rc = ptlrpc_start_hr_threads();
2945 void ptlrpc_hr_fini(void)
2947 struct ptlrpc_hr_partition *hrp;
2950 if (ptlrpc_hr.hr_partitions == NULL)
2953 ptlrpc_stop_hr_threads();
2955 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2956 if (hrp->hrp_thrs != NULL) {
2957 OBD_FREE(hrp->hrp_thrs,
2958 hrp->hrp_nthrs * sizeof(hrp->hrp_thrs[0]));
2962 cfs_percpt_free(ptlrpc_hr.hr_partitions);
2963 ptlrpc_hr.hr_partitions = NULL;
2966 #endif /* __KERNEL__ */
2969 * Wait until all already scheduled replies are processed.
2971 static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
2975 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
2978 rc = l_wait_event(svcpt->scp_waitq,
2979 cfs_atomic_read(&svcpt->scp_nreps_difficult) == 0, &lwi);
2982 CWARN("Unexpectedly long timeout %s %p\n",
2983 svcpt->scp_service->srv_name, svcpt->scp_service);
2988 ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
2990 struct ptlrpc_service_part *svcpt;
2993 /* early disarm AT timer... */
2994 ptlrpc_service_for_each_part(svcpt, i, svc) {
2995 if (svcpt->scp_service != NULL)
2996 cfs_timer_disarm(&svcpt->scp_at_timer);
3001 ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
3003 struct ptlrpc_service_part *svcpt;
3004 struct ptlrpc_request_buffer_desc *rqbd;
3005 struct l_wait_info lwi;
3009 /* All history will be culled when the next request buffer is
3010 * freed in ptlrpc_service_purge_all() */
3011 svc->srv_hist_nrqbds_cpt_max = 0;
3013 rc = LNetClearLazyPortal(svc->srv_req_portal);
3016 ptlrpc_service_for_each_part(svcpt, i, svc) {
3017 if (svcpt->scp_service == NULL)
3020 /* Unlink all the request buffers. This forces a 'final'
3021 * event with its 'unlink' flag set for each posted rqbd */
3022 cfs_list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
3024 rc = LNetMDUnlink(rqbd->rqbd_md_h);
3025 LASSERT(rc == 0 || rc == -ENOENT);
3029 ptlrpc_service_for_each_part(svcpt, i, svc) {
3030 if (svcpt->scp_service == NULL)
3033 /* Wait for the network to release any buffers
3034 * it's currently filling */
3035 spin_lock(&svcpt->scp_lock);
3036 while (svcpt->scp_nrqbds_posted != 0) {
3037 spin_unlock(&svcpt->scp_lock);
3038 /* Network access will complete in finite time but
3039 * the HUGE timeout lets us CWARN for visibility
3040 * of sluggish NALs */
3041 lwi = LWI_TIMEOUT_INTERVAL(
3042 cfs_time_seconds(LONG_UNLINK),
3043 cfs_time_seconds(1), NULL, NULL);
3044 rc = l_wait_event(svcpt->scp_waitq,
3045 svcpt->scp_nrqbds_posted == 0, &lwi);
3046 if (rc == -ETIMEDOUT) {
3047 CWARN("Service %s waiting for "
3048 "request buffers\n",
3049 svcpt->scp_service->srv_name);
3051 spin_lock(&svcpt->scp_lock);
3053 spin_unlock(&svcpt->scp_lock);
3058 ptlrpc_service_purge_all(struct ptlrpc_service *svc)
3060 struct ptlrpc_service_part *svcpt;
3061 struct ptlrpc_request_buffer_desc *rqbd;
3062 struct ptlrpc_request *req;
3063 struct ptlrpc_reply_state *rs;
3066 ptlrpc_service_for_each_part(svcpt, i, svc) {
3067 if (svcpt->scp_service == NULL)
3070 spin_lock(&svcpt->scp_rep_lock);
3071 while (!cfs_list_empty(&svcpt->scp_rep_active)) {
3072 rs = cfs_list_entry(svcpt->scp_rep_active.next,
3073 struct ptlrpc_reply_state, rs_list);
3074 spin_lock(&rs->rs_lock);
3075 ptlrpc_schedule_difficult_reply(rs);
3076 spin_unlock(&rs->rs_lock);
3078 spin_unlock(&svcpt->scp_rep_lock);
3080 /* purge the request queue. NB No new replies (rqbds
3081 * all unlinked) and no service threads, so I'm the only
3082 * thread noodling the request queue now */
3083 while (!cfs_list_empty(&svcpt->scp_req_incoming)) {
3084 req = cfs_list_entry(svcpt->scp_req_incoming.next,
3085 struct ptlrpc_request, rq_list);
3087 cfs_list_del(&req->rq_list);
3088 svcpt->scp_nreqs_incoming--;
3089 ptlrpc_server_finish_request(svcpt, req);
3092 while (ptlrpc_server_request_pending(svcpt, true)) {
3093 req = ptlrpc_server_request_get(svcpt, true);
3094 ptlrpc_server_finish_active_request(svcpt, req);
3097 LASSERT(cfs_list_empty(&svcpt->scp_rqbd_posted));
3098 LASSERT(svcpt->scp_nreqs_incoming == 0);
3099 LASSERT(svcpt->scp_nreqs_active == 0);
3100 /* history should have been culled by
3101 * ptlrpc_server_finish_request */
3102 LASSERT(svcpt->scp_hist_nrqbds == 0);
3104 /* Now free all the request buffers since nothing
3105 * references them any more... */
3107 while (!cfs_list_empty(&svcpt->scp_rqbd_idle)) {
3108 rqbd = cfs_list_entry(svcpt->scp_rqbd_idle.next,
3109 struct ptlrpc_request_buffer_desc,
3111 ptlrpc_free_rqbd(rqbd);
3113 ptlrpc_wait_replies(svcpt);
3115 while (!cfs_list_empty(&svcpt->scp_rep_idle)) {
3116 rs = cfs_list_entry(svcpt->scp_rep_idle.next,
3117 struct ptlrpc_reply_state,
3119 cfs_list_del(&rs->rs_list);
3120 OBD_FREE_LARGE(rs, svc->srv_max_reply_size);
3126 ptlrpc_service_free(struct ptlrpc_service *svc)
3128 struct ptlrpc_service_part *svcpt;
3129 struct ptlrpc_at_array *array;
3132 ptlrpc_service_for_each_part(svcpt, i, svc) {
3133 if (svcpt->scp_service == NULL)
3136 /* In case somebody rearmed this in the meantime */
3137 cfs_timer_disarm(&svcpt->scp_at_timer);
3138 array = &svcpt->scp_at_array;
3140 if (array->paa_reqs_array != NULL) {
3141 OBD_FREE(array->paa_reqs_array,
3142 sizeof(cfs_list_t) * array->paa_size);
3143 array->paa_reqs_array = NULL;
3146 if (array->paa_reqs_count != NULL) {
3147 OBD_FREE(array->paa_reqs_count,
3148 sizeof(__u32) * array->paa_size);
3149 array->paa_reqs_count = NULL;
3153 ptlrpc_service_for_each_part(svcpt, i, svc)
3154 OBD_FREE_PTR(svcpt);
3156 if (svc->srv_cpts != NULL)
3157 cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts);
3159 OBD_FREE(svc, offsetof(struct ptlrpc_service,
3160 srv_parts[svc->srv_ncpts]));
3163 int ptlrpc_unregister_service(struct ptlrpc_service *service)
3167 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
3169 service->srv_is_stopping = 1;
3171 mutex_lock(&ptlrpc_all_services_mutex);
3172 cfs_list_del_init(&service->srv_list);
3173 mutex_unlock(&ptlrpc_all_services_mutex);
3175 ptlrpc_service_del_atimer(service);
3176 ptlrpc_stop_all_threads(service);
3178 ptlrpc_service_unlink_rqbd(service);
3179 ptlrpc_service_purge_all(service);
3180 ptlrpc_service_nrs_cleanup(service);
3182 ptlrpc_lprocfs_unregister_service(service);
3184 ptlrpc_service_free(service);
3188 EXPORT_SYMBOL(ptlrpc_unregister_service);
3191 * Returns 0 if the service is healthy.
3193 * Right now, it just checks to make sure that requests aren't languishing
3194 * in the queue. We'll use this health check to govern whether a node needs
3195 * to be shot, so it's intentionally non-aggressive. */
3196 int ptlrpc_svcpt_health_check(struct ptlrpc_service_part *svcpt)
3198 struct ptlrpc_request *request = NULL;
3199 struct timeval right_now;
3202 do_gettimeofday(&right_now);
3204 spin_lock(&svcpt->scp_req_lock);
3205 /* How long has the next entry been waiting? */
3206 if (ptlrpc_server_high_pending(svcpt, true))
3207 request = ptlrpc_nrs_req_peek_nolock(svcpt, true);
3208 else if (ptlrpc_server_normal_pending(svcpt, true))
3209 request = ptlrpc_nrs_req_peek_nolock(svcpt, false);
3211 if (request == NULL) {
3212 spin_unlock(&svcpt->scp_req_lock);
3216 timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
3217 spin_unlock(&svcpt->scp_req_lock);
3219 if ((timediff / ONE_MILLION) >
3220 (AT_OFF ? obd_timeout * 3 / 2 : at_max)) {
3221 CERROR("%s: unhealthy - request has been waiting %lds\n",
3222 svcpt->scp_service->srv_name, timediff / ONE_MILLION);
3230 ptlrpc_service_health_check(struct ptlrpc_service *svc)
3232 struct ptlrpc_service_part *svcpt;
3238 ptlrpc_service_for_each_part(svcpt, i, svc) {
3239 int rc = ptlrpc_svcpt_health_check(svcpt);
3246 EXPORT_SYMBOL(ptlrpc_service_health_check);