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, 2014, 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
38 #include <linux/kthread.h>
39 #include <obd_support.h>
40 #include <obd_class.h>
41 #include <lustre_net.h>
42 #include <lu_object.h>
43 #include <lnet/types.h>
44 #include "ptlrpc_internal.h"
46 /* The following are visible and mutable through /sys/module/ptlrpc */
47 int test_req_buffer_pressure = 0;
48 CFS_MODULE_PARM(test_req_buffer_pressure, "i", int, 0444,
49 "set non-zero to put pressure on request buffer pools");
50 CFS_MODULE_PARM(at_min, "i", int, 0644,
51 "Adaptive timeout minimum (sec)");
52 CFS_MODULE_PARM(at_max, "i", int, 0644,
53 "Adaptive timeout maximum (sec)");
54 CFS_MODULE_PARM(at_history, "i", int, 0644,
55 "Adaptive timeouts remember the slowest event that took place "
56 "within this period (sec)");
57 CFS_MODULE_PARM(at_early_margin, "i", int, 0644,
58 "How soon before an RPC deadline to send an early reply");
59 CFS_MODULE_PARM(at_extra, "i", int, 0644,
60 "How much extra time to give with each early reply");
64 static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt);
65 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req);
66 static void ptlrpc_at_remove_timed(struct ptlrpc_request *req);
68 /** Holds a list of all PTLRPC services */
69 struct list_head ptlrpc_all_services;
70 /** Used to protect the \e ptlrpc_all_services list */
71 struct mutex ptlrpc_all_services_mutex;
73 static struct ptlrpc_request_buffer_desc *
74 ptlrpc_alloc_rqbd(struct ptlrpc_service_part *svcpt)
76 struct ptlrpc_service *svc = svcpt->scp_service;
77 struct ptlrpc_request_buffer_desc *rqbd;
79 OBD_CPT_ALLOC_PTR(rqbd, svc->srv_cptable, svcpt->scp_cpt);
83 rqbd->rqbd_svcpt = svcpt;
84 rqbd->rqbd_refcount = 0;
85 rqbd->rqbd_cbid.cbid_fn = request_in_callback;
86 rqbd->rqbd_cbid.cbid_arg = rqbd;
87 INIT_LIST_HEAD(&rqbd->rqbd_reqs);
88 OBD_CPT_ALLOC_LARGE(rqbd->rqbd_buffer, svc->srv_cptable,
89 svcpt->scp_cpt, svc->srv_buf_size);
90 if (rqbd->rqbd_buffer == NULL) {
95 spin_lock(&svcpt->scp_lock);
96 list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
97 svcpt->scp_nrqbds_total++;
98 spin_unlock(&svcpt->scp_lock);
104 ptlrpc_free_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
106 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
108 LASSERT(rqbd->rqbd_refcount == 0);
109 LASSERT(list_empty(&rqbd->rqbd_reqs));
111 spin_lock(&svcpt->scp_lock);
112 list_del(&rqbd->rqbd_list);
113 svcpt->scp_nrqbds_total--;
114 spin_unlock(&svcpt->scp_lock);
116 OBD_FREE_LARGE(rqbd->rqbd_buffer, svcpt->scp_service->srv_buf_size);
121 ptlrpc_grow_req_bufs(struct ptlrpc_service_part *svcpt, int post)
123 struct ptlrpc_service *svc = svcpt->scp_service;
124 struct ptlrpc_request_buffer_desc *rqbd;
128 if (svcpt->scp_rqbd_allocating)
131 spin_lock(&svcpt->scp_lock);
132 /* check again with lock */
133 if (svcpt->scp_rqbd_allocating) {
134 /* NB: we might allow more than one thread in the future */
135 LASSERT(svcpt->scp_rqbd_allocating == 1);
136 spin_unlock(&svcpt->scp_lock);
140 svcpt->scp_rqbd_allocating++;
141 spin_unlock(&svcpt->scp_lock);
144 for (i = 0; i < svc->srv_nbuf_per_group; i++) {
145 /* NB: another thread might have recycled enough rqbds, we
146 * need to make sure it wouldn't over-allocate, see LU-1212. */
147 if (svcpt->scp_nrqbds_posted >= svc->srv_nbuf_per_group)
150 rqbd = ptlrpc_alloc_rqbd(svcpt);
153 CERROR("%s: Can't allocate request buffer\n",
160 spin_lock(&svcpt->scp_lock);
162 LASSERT(svcpt->scp_rqbd_allocating == 1);
163 svcpt->scp_rqbd_allocating--;
165 spin_unlock(&svcpt->scp_lock);
168 "%s: allocate %d new %d-byte reqbufs (%d/%d left), rc = %d\n",
169 svc->srv_name, i, svc->srv_buf_size, svcpt->scp_nrqbds_posted,
170 svcpt->scp_nrqbds_total, rc);
174 rc = ptlrpc_server_post_idle_rqbds(svcpt);
180 * Part of Rep-Ack logic.
181 * Puts a lock and its mode into reply state assotiated to request reply.
184 ptlrpc_save_lock(struct ptlrpc_request *req,
185 struct lustre_handle *lock, int mode, int no_ack)
187 struct ptlrpc_reply_state *rs = req->rq_reply_state;
191 LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
193 if (req->rq_export->exp_disconnected) {
194 ldlm_lock_decref(lock, mode);
196 idx = rs->rs_nlocks++;
197 rs->rs_locks[idx] = *lock;
198 rs->rs_modes[idx] = mode;
199 rs->rs_difficult = 1;
200 rs->rs_no_ack = !!no_ack;
203 EXPORT_SYMBOL(ptlrpc_save_lock);
206 struct ptlrpc_hr_partition;
208 struct ptlrpc_hr_thread {
209 int hrt_id; /* thread ID */
211 wait_queue_head_t hrt_waitq;
212 struct list_head hrt_queue; /* RS queue */
213 struct ptlrpc_hr_partition *hrt_partition;
216 struct ptlrpc_hr_partition {
217 /* # of started threads */
218 atomic_t hrp_nstarted;
219 /* # of stopped threads */
220 atomic_t hrp_nstopped;
221 /* cpu partition id */
223 /* round-robin rotor for choosing thread */
225 /* total number of threads on this partition */
228 struct ptlrpc_hr_thread *hrp_thrs;
231 #define HRT_RUNNING 0
232 #define HRT_STOPPING 1
234 struct ptlrpc_hr_service {
235 /* CPU partition table, it's just cfs_cpt_table for now */
236 struct cfs_cpt_table *hr_cpt_table;
237 /** controller sleep waitq */
238 wait_queue_head_t hr_waitq;
239 unsigned int hr_stopping;
240 /** roundrobin rotor for non-affinity service */
241 unsigned int hr_rotor;
243 struct ptlrpc_hr_partition **hr_partitions;
247 struct list_head rsb_replies;
248 unsigned int rsb_n_replies;
249 struct ptlrpc_service_part *rsb_svcpt;
252 /** reply handling service. */
253 static struct ptlrpc_hr_service ptlrpc_hr;
256 * maximum mumber of replies scheduled in one batch
258 #define MAX_SCHEDULED 256
261 * Initialize a reply batch.
265 static void rs_batch_init(struct rs_batch *b)
267 memset(b, 0, sizeof *b);
268 INIT_LIST_HEAD(&b->rsb_replies);
272 * Choose an hr thread to dispatch requests to.
274 static struct ptlrpc_hr_thread *
275 ptlrpc_hr_select(struct ptlrpc_service_part *svcpt)
277 struct ptlrpc_hr_partition *hrp;
280 if (svcpt->scp_cpt >= 0 &&
281 svcpt->scp_service->srv_cptable == ptlrpc_hr.hr_cpt_table) {
282 /* directly match partition */
283 hrp = ptlrpc_hr.hr_partitions[svcpt->scp_cpt];
286 rotor = ptlrpc_hr.hr_rotor++;
287 rotor %= cfs_cpt_number(ptlrpc_hr.hr_cpt_table);
289 hrp = ptlrpc_hr.hr_partitions[rotor];
292 rotor = hrp->hrp_rotor++;
293 return &hrp->hrp_thrs[rotor % hrp->hrp_nthrs];
297 * Dispatch all replies accumulated in the batch to one from
298 * dedicated reply handling threads.
302 static void rs_batch_dispatch(struct rs_batch *b)
304 if (b->rsb_n_replies != 0) {
305 struct ptlrpc_hr_thread *hrt;
307 hrt = ptlrpc_hr_select(b->rsb_svcpt);
309 spin_lock(&hrt->hrt_lock);
310 list_splice_init(&b->rsb_replies, &hrt->hrt_queue);
311 spin_unlock(&hrt->hrt_lock);
313 wake_up(&hrt->hrt_waitq);
314 b->rsb_n_replies = 0;
319 * Add a reply to a batch.
320 * Add one reply object to a batch, schedule batched replies if overload.
325 static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs)
327 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
329 if (svcpt != b->rsb_svcpt || b->rsb_n_replies >= MAX_SCHEDULED) {
330 if (b->rsb_svcpt != NULL) {
331 rs_batch_dispatch(b);
332 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
334 spin_lock(&svcpt->scp_rep_lock);
335 b->rsb_svcpt = svcpt;
337 spin_lock(&rs->rs_lock);
338 rs->rs_scheduled_ever = 1;
339 if (rs->rs_scheduled == 0) {
340 list_move(&rs->rs_list, &b->rsb_replies);
341 rs->rs_scheduled = 1;
344 rs->rs_committed = 1;
345 spin_unlock(&rs->rs_lock);
349 * Reply batch finalization.
350 * Dispatch remaining replies from the batch
351 * and release remaining spinlock.
355 static void rs_batch_fini(struct rs_batch *b)
357 if (b->rsb_svcpt != NULL) {
358 rs_batch_dispatch(b);
359 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
363 #define DECLARE_RS_BATCH(b) struct rs_batch b
367 * Put reply state into a queue for processing because we received
368 * ACK from the client
370 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
372 struct ptlrpc_hr_thread *hrt;
375 LASSERT(list_empty(&rs->rs_list));
377 hrt = ptlrpc_hr_select(rs->rs_svcpt);
379 spin_lock(&hrt->hrt_lock);
380 list_add_tail(&rs->rs_list, &hrt->hrt_queue);
381 spin_unlock(&hrt->hrt_lock);
383 wake_up(&hrt->hrt_waitq);
388 ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs)
392 assert_spin_locked(&rs->rs_svcpt->scp_rep_lock);
393 assert_spin_locked(&rs->rs_lock);
394 LASSERT (rs->rs_difficult);
395 rs->rs_scheduled_ever = 1; /* flag any notification attempt */
397 if (rs->rs_scheduled) { /* being set up or already notified */
402 rs->rs_scheduled = 1;
403 list_del_init(&rs->rs_list);
404 ptlrpc_dispatch_difficult_reply(rs);
407 EXPORT_SYMBOL(ptlrpc_schedule_difficult_reply);
409 void ptlrpc_commit_replies(struct obd_export *exp)
411 struct ptlrpc_reply_state *rs, *nxt;
412 DECLARE_RS_BATCH(batch);
415 rs_batch_init(&batch);
416 /* Find any replies that have been committed and get their service
417 * to attend to complete them. */
419 /* CAVEAT EMPTOR: spinlock ordering!!! */
420 spin_lock(&exp->exp_uncommitted_replies_lock);
421 list_for_each_entry_safe(rs, nxt, &exp->exp_uncommitted_replies,
423 LASSERT (rs->rs_difficult);
424 /* VBR: per-export last_committed */
425 LASSERT(rs->rs_export);
426 if (rs->rs_transno <= exp->exp_last_committed) {
427 list_del_init(&rs->rs_obd_list);
428 rs_batch_add(&batch, rs);
431 spin_unlock(&exp->exp_uncommitted_replies_lock);
432 rs_batch_fini(&batch);
437 ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt)
439 struct ptlrpc_request_buffer_desc *rqbd;
444 spin_lock(&svcpt->scp_lock);
446 if (list_empty(&svcpt->scp_rqbd_idle)) {
447 spin_unlock(&svcpt->scp_lock);
451 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
452 struct ptlrpc_request_buffer_desc,
454 list_del(&rqbd->rqbd_list);
456 /* assume we will post successfully */
457 svcpt->scp_nrqbds_posted++;
458 list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_posted);
460 spin_unlock(&svcpt->scp_lock);
462 rc = ptlrpc_register_rqbd(rqbd);
469 spin_lock(&svcpt->scp_lock);
471 svcpt->scp_nrqbds_posted--;
472 list_del(&rqbd->rqbd_list);
473 list_add_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
475 /* Don't complain if no request buffers are posted right now; LNET
476 * won't drop requests because we set the portal lazy! */
478 spin_unlock(&svcpt->scp_lock);
483 static void ptlrpc_at_timer(unsigned long castmeharder)
485 struct ptlrpc_service_part *svcpt;
487 svcpt = (struct ptlrpc_service_part *)castmeharder;
489 svcpt->scp_at_check = 1;
490 svcpt->scp_at_checktime = cfs_time_current();
491 wake_up(&svcpt->scp_waitq);
495 ptlrpc_server_nthreads_check(struct ptlrpc_service *svc,
496 struct ptlrpc_service_conf *conf)
498 struct ptlrpc_service_thr_conf *tc = &conf->psc_thr;
505 * Common code for estimating & validating threads number.
506 * CPT affinity service could have percpt thread-pool instead
507 * of a global thread-pool, which means user might not always
508 * get the threads number they give it in conf::tc_nthrs_user
509 * even they did set. It's because we need to validate threads
510 * number for each CPT to guarantee each pool will have enough
511 * threads to keep the service healthy.
513 init = PTLRPC_NTHRS_INIT + (svc->srv_ops.so_hpreq_handler != NULL);
514 init = max_t(int, init, tc->tc_nthrs_init);
516 /* NB: please see comments in lustre_lnet.h for definition
517 * details of these members */
518 LASSERT(tc->tc_nthrs_max != 0);
520 if (tc->tc_nthrs_user != 0) {
521 /* In case there is a reason to test a service with many
522 * threads, we give a less strict check here, it can
523 * be up to 8 * nthrs_max */
524 total = min(tc->tc_nthrs_max * 8, tc->tc_nthrs_user);
525 nthrs = total / svc->srv_ncpts;
526 init = max(init, nthrs);
530 total = tc->tc_nthrs_max;
531 if (tc->tc_nthrs_base == 0) {
532 /* don't care about base threads number per partition,
533 * this is most for non-affinity service */
534 nthrs = total / svc->srv_ncpts;
538 nthrs = tc->tc_nthrs_base;
539 if (svc->srv_ncpts == 1) {
542 /* NB: Increase the base number if it's single partition
543 * and total number of cores/HTs is larger or equal to 4.
544 * result will always < 2 * nthrs_base */
545 weight = cfs_cpt_weight(svc->srv_cptable, CFS_CPT_ANY);
546 for (i = 1; (weight >> (i + 1)) != 0 && /* >= 4 cores/HTs */
547 (tc->tc_nthrs_base >> i) != 0; i++)
548 nthrs += tc->tc_nthrs_base >> i;
551 if (tc->tc_thr_factor != 0) {
552 int factor = tc->tc_thr_factor;
556 * User wants to increase number of threads with for
557 * each CPU core/HT, most likely the factor is larger than
558 * one thread/core because service threads are supposed to
559 * be blocked by lock or wait for IO.
562 * Amdahl's law says that adding processors wouldn't give
563 * a linear increasing of parallelism, so it's nonsense to
564 * have too many threads no matter how many cores/HTs
567 if (cfs_cpu_ht_nsiblings(0) > 1) { /* weight is # of HTs */
568 /* depress thread factor for hyper-thread */
569 factor = factor - (factor >> 1) + (factor >> 3);
572 weight = cfs_cpt_weight(svc->srv_cptable, 0);
575 for (; factor > 0 && weight > 0; factor--, weight -= fade)
576 nthrs += min(weight, fade) * factor;
579 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
580 nthrs = max(tc->tc_nthrs_base,
581 tc->tc_nthrs_max / svc->srv_ncpts);
584 nthrs = max(nthrs, tc->tc_nthrs_init);
585 svc->srv_nthrs_cpt_limit = nthrs;
586 svc->srv_nthrs_cpt_init = init;
588 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
589 CDEBUG(D_OTHER, "%s: This service may have more threads (%d) "
590 "than the given soft limit (%d)\n",
591 svc->srv_name, nthrs * svc->srv_ncpts,
597 * Initialize percpt data for a service
600 ptlrpc_service_part_init(struct ptlrpc_service *svc,
601 struct ptlrpc_service_part *svcpt, int cpt)
603 struct ptlrpc_at_array *array;
608 svcpt->scp_cpt = cpt;
609 INIT_LIST_HEAD(&svcpt->scp_threads);
611 /* rqbd and incoming request queue */
612 spin_lock_init(&svcpt->scp_lock);
613 INIT_LIST_HEAD(&svcpt->scp_rqbd_idle);
614 INIT_LIST_HEAD(&svcpt->scp_rqbd_posted);
615 INIT_LIST_HEAD(&svcpt->scp_req_incoming);
616 init_waitqueue_head(&svcpt->scp_waitq);
617 /* history request & rqbd list */
618 INIT_LIST_HEAD(&svcpt->scp_hist_reqs);
619 INIT_LIST_HEAD(&svcpt->scp_hist_rqbds);
621 /* acitve requests and hp requests */
622 spin_lock_init(&svcpt->scp_req_lock);
625 spin_lock_init(&svcpt->scp_rep_lock);
626 INIT_LIST_HEAD(&svcpt->scp_rep_active);
627 INIT_LIST_HEAD(&svcpt->scp_rep_idle);
628 init_waitqueue_head(&svcpt->scp_rep_waitq);
629 atomic_set(&svcpt->scp_nreps_difficult, 0);
631 /* adaptive timeout */
632 spin_lock_init(&svcpt->scp_at_lock);
633 array = &svcpt->scp_at_array;
635 size = at_est2timeout(at_max);
636 array->paa_size = size;
637 array->paa_count = 0;
638 array->paa_deadline = -1;
640 /* allocate memory for scp_at_array (ptlrpc_at_array) */
641 OBD_CPT_ALLOC(array->paa_reqs_array,
642 svc->srv_cptable, cpt, sizeof(struct list_head) * size);
643 if (array->paa_reqs_array == NULL)
646 for (index = 0; index < size; index++)
647 INIT_LIST_HEAD(&array->paa_reqs_array[index]);
649 OBD_CPT_ALLOC(array->paa_reqs_count,
650 svc->srv_cptable, cpt, sizeof(__u32) * size);
651 if (array->paa_reqs_count == NULL)
654 cfs_timer_init(&svcpt->scp_at_timer, ptlrpc_at_timer, svcpt);
655 /* At SOW, service time should be quick; 10s seems generous. If client
656 * timeout is less than this, we'll be sending an early reply. */
657 at_init(&svcpt->scp_at_estimate, 10, 0);
659 /* assign this before call ptlrpc_grow_req_bufs */
660 svcpt->scp_service = svc;
661 /* Now allocate the request buffers, but don't post them now */
662 rc = ptlrpc_grow_req_bufs(svcpt, 0);
663 /* We shouldn't be under memory pressure at startup, so
664 * fail if we can't allocate all our buffers at this time. */
671 if (array->paa_reqs_count != NULL) {
672 OBD_FREE(array->paa_reqs_count, sizeof(__u32) * size);
673 array->paa_reqs_count = NULL;
676 if (array->paa_reqs_array != NULL) {
677 OBD_FREE(array->paa_reqs_array,
678 sizeof(struct list_head) * array->paa_size);
679 array->paa_reqs_array = NULL;
686 * Initialize service on a given portal.
687 * This includes starting serving threads , allocating and posting rqbds and
690 struct ptlrpc_service *
691 ptlrpc_register_service(struct ptlrpc_service_conf *conf,
692 struct proc_dir_entry *proc_entry)
694 struct ptlrpc_service_cpt_conf *cconf = &conf->psc_cpt;
695 struct ptlrpc_service *service;
696 struct ptlrpc_service_part *svcpt;
697 struct cfs_cpt_table *cptable;
705 LASSERT(conf->psc_buf.bc_nbufs > 0);
706 LASSERT(conf->psc_buf.bc_buf_size >=
707 conf->psc_buf.bc_req_max_size + SPTLRPC_MAX_PAYLOAD);
708 LASSERT(conf->psc_thr.tc_ctx_tags != 0);
710 cptable = cconf->cc_cptable;
712 cptable = cfs_cpt_table;
714 if (!conf->psc_thr.tc_cpu_affinity) {
717 ncpts = cfs_cpt_number(cptable);
718 if (cconf->cc_pattern != NULL) {
719 struct cfs_expr_list *el;
721 rc = cfs_expr_list_parse(cconf->cc_pattern,
722 strlen(cconf->cc_pattern),
725 CERROR("%s: invalid CPT pattern string: %s",
726 conf->psc_name, cconf->cc_pattern);
727 RETURN(ERR_PTR(-EINVAL));
730 rc = cfs_expr_list_values(el, ncpts, &cpts);
731 cfs_expr_list_free(el);
733 CERROR("%s: failed to parse CPT array %s: %d\n",
734 conf->psc_name, cconf->cc_pattern, rc);
736 OBD_FREE(cpts, sizeof(*cpts) * ncpts);
737 RETURN(ERR_PTR(rc < 0 ? rc : -EINVAL));
743 OBD_ALLOC(service, offsetof(struct ptlrpc_service, srv_parts[ncpts]));
744 if (service == NULL) {
746 OBD_FREE(cpts, sizeof(*cpts) * ncpts);
747 RETURN(ERR_PTR(-ENOMEM));
750 service->srv_cptable = cptable;
751 service->srv_cpts = cpts;
752 service->srv_ncpts = ncpts;
754 service->srv_cpt_bits = 0; /* it's zero already, easy to read... */
755 while ((1 << service->srv_cpt_bits) < cfs_cpt_number(cptable))
756 service->srv_cpt_bits++;
759 spin_lock_init(&service->srv_lock);
760 service->srv_name = conf->psc_name;
761 service->srv_watchdog_factor = conf->psc_watchdog_factor;
762 INIT_LIST_HEAD(&service->srv_list); /* for safty of cleanup */
764 /* buffer configuration */
765 service->srv_nbuf_per_group = test_req_buffer_pressure ?
766 1 : conf->psc_buf.bc_nbufs;
767 service->srv_max_req_size = conf->psc_buf.bc_req_max_size +
769 service->srv_buf_size = conf->psc_buf.bc_buf_size;
770 service->srv_rep_portal = conf->psc_buf.bc_rep_portal;
771 service->srv_req_portal = conf->psc_buf.bc_req_portal;
773 /* Increase max reply size to next power of two */
774 service->srv_max_reply_size = 1;
775 while (service->srv_max_reply_size <
776 conf->psc_buf.bc_rep_max_size + SPTLRPC_MAX_PAYLOAD)
777 service->srv_max_reply_size <<= 1;
779 service->srv_thread_name = conf->psc_thr.tc_thr_name;
780 service->srv_ctx_tags = conf->psc_thr.tc_ctx_tags;
781 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
782 service->srv_ops = conf->psc_ops;
784 for (i = 0; i < ncpts; i++) {
785 if (!conf->psc_thr.tc_cpu_affinity)
788 cpt = cpts != NULL ? cpts[i] : i;
790 OBD_CPT_ALLOC(svcpt, cptable, cpt, sizeof(*svcpt));
792 GOTO(failed, rc = -ENOMEM);
794 service->srv_parts[i] = svcpt;
795 rc = ptlrpc_service_part_init(service, svcpt, cpt);
800 ptlrpc_server_nthreads_check(service, conf);
802 rc = LNetSetLazyPortal(service->srv_req_portal);
805 mutex_lock(&ptlrpc_all_services_mutex);
806 list_add(&service->srv_list, &ptlrpc_all_services);
807 mutex_unlock(&ptlrpc_all_services_mutex);
809 if (proc_entry != NULL)
810 ptlrpc_lprocfs_register_service(proc_entry, service);
812 rc = ptlrpc_service_nrs_setup(service);
816 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
817 service->srv_name, service->srv_req_portal);
819 rc = ptlrpc_start_threads(service);
821 CERROR("Failed to start threads for service %s: %d\n",
822 service->srv_name, rc);
828 ptlrpc_unregister_service(service);
831 EXPORT_SYMBOL(ptlrpc_register_service);
834 * to actually free the request, must be called without holding svc_lock.
835 * note it's caller's responsibility to unlink req->rq_list.
837 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
839 LASSERT(atomic_read(&req->rq_refcount) == 0);
840 LASSERT(list_empty(&req->rq_timed_list));
842 /* DEBUG_REQ() assumes the reply state of a request with a valid
843 * ref will not be destroyed until that reference is dropped. */
844 ptlrpc_req_drop_rs(req);
846 sptlrpc_svc_ctx_decref(req);
848 if (req != &req->rq_rqbd->rqbd_req) {
849 /* NB request buffers use an embedded
850 * req if the incoming req unlinked the
851 * MD; this isn't one of them! */
852 ptlrpc_request_cache_free(req);
857 * drop a reference count of the request. if it reaches 0, we either
858 * put it into history list, or free it immediately.
860 void ptlrpc_server_drop_request(struct ptlrpc_request *req)
862 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
863 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
864 struct ptlrpc_service *svc = svcpt->scp_service;
866 struct list_head *tmp;
867 struct list_head *nxt;
869 if (!atomic_dec_and_test(&req->rq_refcount))
872 if (req->rq_session.lc_state == LCS_ENTERED) {
873 lu_context_exit(&req->rq_session);
874 lu_context_fini(&req->rq_session);
877 if (req->rq_at_linked) {
878 spin_lock(&svcpt->scp_at_lock);
879 /* recheck with lock, in case it's unlinked by
880 * ptlrpc_at_check_timed() */
881 if (likely(req->rq_at_linked))
882 ptlrpc_at_remove_timed(req);
883 spin_unlock(&svcpt->scp_at_lock);
886 LASSERT(list_empty(&req->rq_timed_list));
888 /* finalize request */
889 if (req->rq_export) {
890 class_export_put(req->rq_export);
891 req->rq_export = NULL;
894 spin_lock(&svcpt->scp_lock);
896 list_add(&req->rq_list, &rqbd->rqbd_reqs);
898 refcount = --(rqbd->rqbd_refcount);
900 /* request buffer is now idle: add to history */
901 list_del(&rqbd->rqbd_list);
903 list_add_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds);
904 svcpt->scp_hist_nrqbds++;
906 /* cull some history?
907 * I expect only about 1 or 2 rqbds need to be recycled here */
908 while (svcpt->scp_hist_nrqbds > svc->srv_hist_nrqbds_cpt_max) {
909 rqbd = list_entry(svcpt->scp_hist_rqbds.next,
910 struct ptlrpc_request_buffer_desc,
913 list_del(&rqbd->rqbd_list);
914 svcpt->scp_hist_nrqbds--;
916 /* remove rqbd's reqs from svc's req history while
917 * I've got the service lock */
918 list_for_each(tmp, &rqbd->rqbd_reqs) {
919 req = list_entry(tmp, struct ptlrpc_request,
921 /* Track the highest culled req seq */
922 if (req->rq_history_seq >
923 svcpt->scp_hist_seq_culled) {
924 svcpt->scp_hist_seq_culled =
927 list_del(&req->rq_history_list);
930 spin_unlock(&svcpt->scp_lock);
932 list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
933 req = list_entry(rqbd->rqbd_reqs.next,
934 struct ptlrpc_request,
936 list_del(&req->rq_list);
937 ptlrpc_server_free_request(req);
940 spin_lock(&svcpt->scp_lock);
942 * now all reqs including the embedded req has been
943 * disposed, schedule request buffer for re-use.
945 LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) == 0);
946 list_add_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
949 spin_unlock(&svcpt->scp_lock);
950 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
951 /* If we are low on memory, we are not interested in history */
952 list_del(&req->rq_list);
953 list_del_init(&req->rq_history_list);
955 /* Track the highest culled req seq */
956 if (req->rq_history_seq > svcpt->scp_hist_seq_culled)
957 svcpt->scp_hist_seq_culled = req->rq_history_seq;
959 spin_unlock(&svcpt->scp_lock);
961 ptlrpc_server_free_request(req);
963 spin_unlock(&svcpt->scp_lock);
967 /** Change request export and move hp request from old export to new */
968 void ptlrpc_request_change_export(struct ptlrpc_request *req,
969 struct obd_export *export)
971 if (req->rq_export != NULL) {
972 LASSERT(!list_empty(&req->rq_exp_list));
973 /* remove rq_exp_list from last export */
974 spin_lock_bh(&req->rq_export->exp_rpc_lock);
975 list_del_init(&req->rq_exp_list);
976 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
977 /* export has one reference already, so it`s safe to
978 * add req to export queue here and get another
979 * reference for request later */
980 spin_lock_bh(&export->exp_rpc_lock);
981 if (req->rq_ops != NULL) /* hp request */
982 list_add(&req->rq_exp_list, &export->exp_hp_rpcs);
984 list_add(&req->rq_exp_list, &export->exp_reg_rpcs);
985 spin_unlock_bh(&export->exp_rpc_lock);
987 class_export_rpc_dec(req->rq_export);
988 class_export_put(req->rq_export);
991 /* request takes one export refcount */
992 req->rq_export = class_export_get(export);
993 class_export_rpc_inc(export);
999 * to finish a request: stop sending more early replies, and release
1002 static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt,
1003 struct ptlrpc_request *req)
1005 ptlrpc_server_hpreq_fini(req);
1007 ptlrpc_server_drop_request(req);
1011 * to finish an active request: stop sending more early replies, and release
1012 * the request. should be called after we finished handling the request.
1014 static void ptlrpc_server_finish_active_request(
1015 struct ptlrpc_service_part *svcpt,
1016 struct ptlrpc_request *req)
1018 spin_lock(&svcpt->scp_req_lock);
1019 ptlrpc_nrs_req_stop_nolock(req);
1020 svcpt->scp_nreqs_active--;
1022 svcpt->scp_nhreqs_active--;
1023 spin_unlock(&svcpt->scp_req_lock);
1025 ptlrpc_nrs_req_finalize(req);
1027 if (req->rq_export != NULL)
1028 class_export_rpc_dec(req->rq_export);
1030 ptlrpc_server_finish_request(svcpt, req);
1034 * This function makes sure dead exports are evicted in a timely manner.
1035 * This function is only called when some export receives a message (i.e.,
1036 * the network is up.)
1038 void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
1040 struct obd_export *oldest_exp;
1041 time_t oldest_time, new_time;
1047 /* Compensate for slow machines, etc, by faking our request time
1048 into the future. Although this can break the strict time-ordering
1049 of the list, we can be really lazy here - we don't have to evict
1050 at the exact right moment. Eventually, all silent exports
1051 will make it to the top of the list. */
1053 /* Do not pay attention on 1sec or smaller renewals. */
1054 new_time = cfs_time_current_sec() + extra_delay;
1055 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
1058 exp->exp_last_request_time = new_time;
1060 /* exports may get disconnected from the chain even though the
1061 export has references, so we must keep the spin lock while
1062 manipulating the lists */
1063 spin_lock(&exp->exp_obd->obd_dev_lock);
1065 if (list_empty(&exp->exp_obd_chain_timed)) {
1066 /* this one is not timed */
1067 spin_unlock(&exp->exp_obd->obd_dev_lock);
1071 list_move_tail(&exp->exp_obd_chain_timed,
1072 &exp->exp_obd->obd_exports_timed);
1074 oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
1075 struct obd_export, exp_obd_chain_timed);
1076 oldest_time = oldest_exp->exp_last_request_time;
1077 spin_unlock(&exp->exp_obd->obd_dev_lock);
1079 if (exp->exp_obd->obd_recovering) {
1080 /* be nice to everyone during recovery */
1085 /* Note - racing to start/reset the obd_eviction timer is safe */
1086 if (exp->exp_obd->obd_eviction_timer == 0) {
1087 /* Check if the oldest entry is expired. */
1088 if (cfs_time_current_sec() > (oldest_time + PING_EVICT_TIMEOUT +
1090 /* We need a second timer, in case the net was down and
1091 * it just came back. Since the pinger may skip every
1092 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
1093 * we better wait for 3. */
1094 exp->exp_obd->obd_eviction_timer =
1095 cfs_time_current_sec() + 3 * PING_INTERVAL;
1096 CDEBUG(D_HA, "%s: Think about evicting %s from "CFS_TIME_T"\n",
1097 exp->exp_obd->obd_name,
1098 obd_export_nid2str(oldest_exp), oldest_time);
1101 if (cfs_time_current_sec() >
1102 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
1103 /* The evictor won't evict anyone who we've heard from
1104 * recently, so we don't have to check before we start
1106 if (!ping_evictor_wake(exp))
1107 exp->exp_obd->obd_eviction_timer = 0;
1115 * Sanity check request \a req.
1116 * Return 0 if all is ok, error code otherwise.
1118 static int ptlrpc_check_req(struct ptlrpc_request *req)
1120 struct obd_device *obd = req->rq_export->exp_obd;
1123 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
1124 req->rq_export->exp_conn_cnt)) {
1125 DEBUG_REQ(D_RPCTRACE, req,
1126 "DROPPING req from old connection %d < %d",
1127 lustre_msg_get_conn_cnt(req->rq_reqmsg),
1128 req->rq_export->exp_conn_cnt);
1131 if (unlikely(obd == NULL || obd->obd_fail)) {
1132 /* Failing over, don't handle any more reqs,
1133 * send error response instead. */
1134 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
1135 req, (obd != NULL) ? obd->obd_name : "unknown");
1137 } else if (lustre_msg_get_flags(req->rq_reqmsg) &
1138 (MSG_REPLAY | MSG_REQ_REPLAY_DONE) &&
1139 !obd->obd_recovering) {
1140 DEBUG_REQ(D_ERROR, req,
1141 "Invalid replay without recovery");
1142 class_fail_export(req->rq_export);
1144 } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 &&
1145 !obd->obd_recovering) {
1146 DEBUG_REQ(D_ERROR, req, "Invalid req with transno "
1147 LPU64" without recovery",
1148 lustre_msg_get_transno(req->rq_reqmsg));
1149 class_fail_export(req->rq_export);
1153 if (unlikely(rc < 0)) {
1154 req->rq_status = rc;
1160 static void ptlrpc_at_set_timer(struct ptlrpc_service_part *svcpt)
1162 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1165 if (array->paa_count == 0) {
1166 cfs_timer_disarm(&svcpt->scp_at_timer);
1170 /* Set timer for closest deadline */
1171 next = (__s32)(array->paa_deadline - cfs_time_current_sec() -
1174 ptlrpc_at_timer((unsigned long)svcpt);
1176 cfs_timer_arm(&svcpt->scp_at_timer, cfs_time_shift(next));
1177 CDEBUG(D_INFO, "armed %s at %+ds\n",
1178 svcpt->scp_service->srv_name, next);
1182 /* Add rpc to early reply check list */
1183 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
1185 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1186 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1187 struct ptlrpc_request *rq = NULL;
1193 if (req->rq_no_reply)
1196 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
1199 spin_lock(&svcpt->scp_at_lock);
1200 LASSERT(list_empty(&req->rq_timed_list));
1202 index = (unsigned long)req->rq_deadline % array->paa_size;
1203 if (array->paa_reqs_count[index] > 0) {
1204 /* latest rpcs will have the latest deadlines in the list,
1205 * so search backward. */
1206 list_for_each_entry_reverse(rq,
1207 &array->paa_reqs_array[index],
1209 if (req->rq_deadline >= rq->rq_deadline) {
1210 list_add(&req->rq_timed_list,
1211 &rq->rq_timed_list);
1217 /* Add the request at the head of the list */
1218 if (list_empty(&req->rq_timed_list))
1219 list_add(&req->rq_timed_list,
1220 &array->paa_reqs_array[index]);
1222 spin_lock(&req->rq_lock);
1223 req->rq_at_linked = 1;
1224 spin_unlock(&req->rq_lock);
1225 req->rq_at_index = index;
1226 array->paa_reqs_count[index]++;
1228 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
1229 array->paa_deadline = req->rq_deadline;
1230 ptlrpc_at_set_timer(svcpt);
1232 spin_unlock(&svcpt->scp_at_lock);
1238 ptlrpc_at_remove_timed(struct ptlrpc_request *req)
1240 struct ptlrpc_at_array *array;
1242 array = &req->rq_rqbd->rqbd_svcpt->scp_at_array;
1244 /* NB: must call with hold svcpt::scp_at_lock */
1245 LASSERT(!list_empty(&req->rq_timed_list));
1246 list_del_init(&req->rq_timed_list);
1248 spin_lock(&req->rq_lock);
1249 req->rq_at_linked = 0;
1250 spin_unlock(&req->rq_lock);
1252 array->paa_reqs_count[req->rq_at_index]--;
1257 * Attempt to extend the request deadline by sending an early reply to the
1260 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
1262 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1263 struct ptlrpc_request *reqcopy;
1264 struct lustre_msg *reqmsg;
1265 cfs_duration_t olddl = req->rq_deadline - cfs_time_current_sec();
1271 if (CFS_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_RECONNECT)) {
1272 /* don't send early reply */
1276 /* deadline is when the client expects us to reply, margin is the
1277 difference between clients' and servers' expectations */
1278 DEBUG_REQ(D_ADAPTTO, req,
1279 "%ssending early reply (deadline %+lds, margin %+lds) for "
1280 "%d+%d", AT_OFF ? "AT off - not " : "",
1281 olddl, olddl - at_get(&svcpt->scp_at_estimate),
1282 at_get(&svcpt->scp_at_estimate), at_extra);
1288 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), "
1289 "not sending early reply. Consider increasing "
1290 "at_early_margin (%d)?", olddl, at_early_margin);
1292 /* Return an error so we're not re-added to the timed list. */
1296 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
1297 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
1298 "but no AT support");
1302 if (req->rq_export &&
1303 lustre_msg_get_flags(req->rq_reqmsg) &
1304 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1305 /* During recovery, we don't want to send too many early
1306 * replies, but on the other hand we want to make sure the
1307 * client has enough time to resend if the rpc is lost. So
1308 * during the recovery period send at least 4 early replies,
1309 * spacing them every at_extra if we can. at_estimate should
1310 * always equal this fixed value during recovery. */
1311 /* Don't account request processing time into AT history
1312 * during recovery, it is not service time we need but
1313 * includes also waiting time for recovering clients */
1314 newdl = cfs_time_current_sec() + min(at_extra,
1315 req->rq_export->exp_obd->obd_recovery_timeout / 4);
1317 /* We want to extend the request deadline by at_extra seconds,
1318 * so we set our service estimate to reflect how much time has
1319 * passed since this request arrived plus an additional
1320 * at_extra seconds. The client will calculate the new deadline
1321 * based on this service estimate (plus some additional time to
1322 * account for network latency). See ptlrpc_at_recv_early_reply
1324 at_measured(&svcpt->scp_at_estimate, at_extra +
1325 cfs_time_current_sec() -
1326 req->rq_arrival_time.tv_sec);
1327 newdl = req->rq_arrival_time.tv_sec +
1328 at_get(&svcpt->scp_at_estimate);
1331 /* Check to see if we've actually increased the deadline -
1332 * we may be past adaptive_max */
1333 if (req->rq_deadline >= newdl) {
1334 DEBUG_REQ(D_WARNING, req, "Couldn't add any time "
1335 "(%ld/%ld), not sending early reply\n",
1336 olddl, newdl - cfs_time_current_sec());
1340 reqcopy = ptlrpc_request_cache_alloc(GFP_NOFS);
1341 if (reqcopy == NULL)
1343 OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1345 GOTO(out_free, rc = -ENOMEM);
1348 reqcopy->rq_reply_state = NULL;
1349 reqcopy->rq_rep_swab_mask = 0;
1350 reqcopy->rq_pack_bulk = 0;
1351 reqcopy->rq_pack_udesc = 0;
1352 reqcopy->rq_packed_final = 0;
1353 sptlrpc_svc_ctx_addref(reqcopy);
1354 /* We only need the reqmsg for the magic */
1355 reqcopy->rq_reqmsg = reqmsg;
1356 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1359 * tgt_brw_read() and tgt_brw_write() may have decided not to reply.
1360 * Without this check, we would fail the rq_no_reply assertion in
1361 * ptlrpc_send_reply().
1363 if (reqcopy->rq_no_reply)
1364 GOTO(out, rc = -ETIMEDOUT);
1366 LASSERT(atomic_read(&req->rq_refcount));
1367 /** if it is last refcount then early reply isn't needed */
1368 if (atomic_read(&req->rq_refcount) == 1) {
1369 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
1370 "abort sending early reply\n");
1371 GOTO(out, rc = -EINVAL);
1374 /* Connection ref */
1375 reqcopy->rq_export = class_conn2export(
1376 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1377 if (reqcopy->rq_export == NULL)
1378 GOTO(out, rc = -ENODEV);
1381 class_export_rpc_inc(reqcopy->rq_export);
1382 if (reqcopy->rq_export->exp_obd &&
1383 reqcopy->rq_export->exp_obd->obd_fail)
1384 GOTO(out_put, rc = -ENODEV);
1386 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1390 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1393 /* Adjust our own deadline to what we told the client */
1394 req->rq_deadline = newdl;
1395 req->rq_early_count++; /* number sent, server side */
1397 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1400 /* Free the (early) reply state from lustre_pack_reply.
1401 (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
1402 ptlrpc_req_drop_rs(reqcopy);
1405 class_export_rpc_dec(reqcopy->rq_export);
1406 class_export_put(reqcopy->rq_export);
1408 sptlrpc_svc_ctx_decref(reqcopy);
1409 OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1411 ptlrpc_request_cache_free(reqcopy);
1415 /* Send early replies to everybody expiring within at_early_margin
1416 asking for at_extra time */
1417 static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1419 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1420 struct ptlrpc_request *rq, *n;
1421 struct list_head work_list;
1424 time_t now = cfs_time_current_sec();
1425 cfs_duration_t delay;
1426 int first, counter = 0;
1429 spin_lock(&svcpt->scp_at_lock);
1430 if (svcpt->scp_at_check == 0) {
1431 spin_unlock(&svcpt->scp_at_lock);
1434 delay = cfs_time_sub(cfs_time_current(), svcpt->scp_at_checktime);
1435 svcpt->scp_at_check = 0;
1437 if (array->paa_count == 0) {
1438 spin_unlock(&svcpt->scp_at_lock);
1442 /* The timer went off, but maybe the nearest rpc already completed. */
1443 first = array->paa_deadline - now;
1444 if (first > at_early_margin) {
1445 /* We've still got plenty of time. Reset the timer. */
1446 ptlrpc_at_set_timer(svcpt);
1447 spin_unlock(&svcpt->scp_at_lock);
1451 /* We're close to a timeout, and we don't know how much longer the
1452 server will take. Send early replies to everyone expiring soon. */
1453 INIT_LIST_HEAD(&work_list);
1455 index = (unsigned long)array->paa_deadline % array->paa_size;
1456 count = array->paa_count;
1458 count -= array->paa_reqs_count[index];
1459 list_for_each_entry_safe(rq, n,
1460 &array->paa_reqs_array[index],
1462 if (rq->rq_deadline > now + at_early_margin) {
1463 /* update the earliest deadline */
1464 if (deadline == -1 ||
1465 rq->rq_deadline < deadline)
1466 deadline = rq->rq_deadline;
1470 ptlrpc_at_remove_timed(rq);
1472 * ptlrpc_server_drop_request() may drop
1473 * refcount to 0 already. Let's check this and
1474 * don't add entry to work_list
1476 if (likely(atomic_inc_not_zero(&rq->rq_refcount)))
1477 list_add(&rq->rq_timed_list, &work_list);
1481 if (++index >= array->paa_size)
1484 array->paa_deadline = deadline;
1485 /* we have a new earliest deadline, restart the timer */
1486 ptlrpc_at_set_timer(svcpt);
1488 spin_unlock(&svcpt->scp_at_lock);
1490 CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
1491 "replies\n", first, at_extra, counter);
1493 /* We're already past request deadlines before we even get a
1494 chance to send early replies */
1495 LCONSOLE_WARN("%s: This server is not able to keep up with "
1496 "request traffic (cpu-bound).\n",
1497 svcpt->scp_service->srv_name);
1498 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, "
1499 "delay="CFS_DURATION_T"(jiff)\n",
1500 counter, svcpt->scp_nreqs_incoming,
1501 svcpt->scp_nreqs_active,
1502 at_get(&svcpt->scp_at_estimate), delay);
1505 /* we took additional refcount so entries can't be deleted from list, no
1506 * locking is needed */
1507 while (!list_empty(&work_list)) {
1508 rq = list_entry(work_list.next, struct ptlrpc_request,
1510 list_del_init(&rq->rq_timed_list);
1512 if (ptlrpc_at_send_early_reply(rq) == 0)
1513 ptlrpc_at_add_timed(rq);
1515 ptlrpc_server_drop_request(rq);
1518 RETURN(1); /* return "did_something" for liblustre */
1521 /* Check if we are already handling earlier incarnation of this request.
1522 * Called under &req->rq_export->exp_rpc_lock locked */
1523 static int ptlrpc_server_check_resend_in_progress(struct ptlrpc_request *req)
1525 struct ptlrpc_request *tmp = NULL;
1527 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ||
1528 (atomic_read(&req->rq_export->exp_rpc_count) == 0))
1531 /* bulk request are aborted upon reconnect, don't try to
1533 if (req->rq_bulk_write || req->rq_bulk_read)
1536 /* This list should not be longer than max_requests in
1537 * flights on the client, so it is not all that long.
1538 * Also we only hit this codepath in case of a resent
1539 * request which makes it even more rarely hit */
1540 list_for_each_entry(tmp, &req->rq_export->exp_reg_rpcs,
1542 /* Found duplicate one */
1543 if (tmp->rq_xid == req->rq_xid)
1546 list_for_each_entry(tmp, &req->rq_export->exp_hp_rpcs,
1548 /* Found duplicate one */
1549 if (tmp->rq_xid == req->rq_xid)
1555 DEBUG_REQ(D_HA, req, "Found duplicate req in processing");
1556 DEBUG_REQ(D_HA, tmp, "Request being processed");
1561 * Put the request to the export list if the request may become
1562 * a high priority one.
1564 static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt,
1565 struct ptlrpc_request *req)
1567 struct list_head *list;
1572 if (svcpt->scp_service->srv_ops.so_hpreq_handler) {
1573 rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
1578 if (req->rq_export) {
1580 /* Perform request specific check. We should do this
1581 * check before the request is added into exp_hp_rpcs
1582 * list otherwise it may hit swab race at LU-1044. */
1583 if (req->rq_ops->hpreq_check) {
1584 rc = req->rq_ops->hpreq_check(req);
1586 * XXX: Out of all current
1587 * ptlrpc_hpreq_ops::hpreq_check(), only
1588 * ldlm_cancel_hpreq_check() can return an
1589 * error code; other functions assert in
1590 * similar places, which seems odd.
1591 * What also does not seem right is that
1592 * handlers for those RPCs do not assert
1593 * on the same checks, but rather handle the
1594 * error cases. e.g. see ost_rw_hpreq_check(),
1595 * and ost_brw_read(), ost_brw_write().
1599 LASSERT(rc == 0 || rc == 1);
1602 list = &req->rq_export->exp_hp_rpcs;
1604 list = &req->rq_export->exp_reg_rpcs;
1607 /* do search for duplicated xid and the adding to the list
1609 spin_lock_bh(&req->rq_export->exp_rpc_lock);
1610 rc = ptlrpc_server_check_resend_in_progress(req);
1612 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1615 list_add(&req->rq_exp_list, list);
1616 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1619 ptlrpc_nrs_req_initialize(svcpt, req, !!hp);
1624 /** Remove the request from the export list. */
1625 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req)
1628 if (req->rq_export) {
1629 /* refresh lock timeout again so that client has more
1630 * room to send lock cancel RPC. */
1631 if (req->rq_ops && req->rq_ops->hpreq_fini)
1632 req->rq_ops->hpreq_fini(req);
1634 spin_lock_bh(&req->rq_export->exp_rpc_lock);
1635 list_del_init(&req->rq_exp_list);
1636 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1641 static int ptlrpc_hpreq_check(struct ptlrpc_request *req)
1646 static struct ptlrpc_hpreq_ops ptlrpc_hpreq_common = {
1647 .hpreq_check = ptlrpc_hpreq_check,
1650 /* Hi-Priority RPC check by RPC operation code. */
1651 int ptlrpc_hpreq_handler(struct ptlrpc_request *req)
1653 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1655 /* Check for export to let only reconnects for not yet evicted
1656 * export to become a HP rpc. */
1657 if ((req->rq_export != NULL) &&
1658 (opc == OBD_PING || opc == MDS_CONNECT || opc == OST_CONNECT))
1659 req->rq_ops = &ptlrpc_hpreq_common;
1663 EXPORT_SYMBOL(ptlrpc_hpreq_handler);
1665 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1666 struct ptlrpc_request *req)
1671 rc = ptlrpc_server_hpreq_init(svcpt, req);
1675 /* the current thread is not the processing thread for this request
1676 * since that, but request is in exp_hp_list and can be find there.
1677 * Remove all relations between request and old thread. */
1678 req->rq_svc_thread->t_env->le_ses = NULL;
1679 req->rq_svc_thread = NULL;
1680 req->rq_session.lc_thread = NULL;
1682 ptlrpc_nrs_req_add(svcpt, req, !!rc);
1688 * Allow to handle high priority request
1689 * User can call it w/o any lock but need to hold
1690 * ptlrpc_service_part::scp_req_lock to get reliable result
1692 static bool ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1695 int running = svcpt->scp_nthrs_running;
1697 if (!nrs_svcpt_has_hp(svcpt))
1703 if (ptlrpc_nrs_req_throttling_nolock(svcpt, true))
1706 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1707 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1708 /* leave just 1 thread for normal RPCs */
1709 running = PTLRPC_NTHRS_INIT;
1710 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1714 if (svcpt->scp_nreqs_active >= running - 1)
1717 if (svcpt->scp_nhreqs_active == 0)
1720 return !ptlrpc_nrs_req_pending_nolock(svcpt, false) ||
1721 svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1724 static bool ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1727 return ptlrpc_server_allow_high(svcpt, force) &&
1728 ptlrpc_nrs_req_pending_nolock(svcpt, true);
1732 * Only allow normal priority requests on a service that has a high-priority
1733 * queue if forced (i.e. cleanup), if there are other high priority requests
1734 * already being processed (i.e. those threads can service more high-priority
1735 * requests), or if there are enough idle threads that a later thread can do
1736 * a high priority request.
1737 * User can call it w/o any lock but need to hold
1738 * ptlrpc_service_part::scp_req_lock to get reliable result
1740 static bool ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1743 int running = svcpt->scp_nthrs_running;
1744 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1745 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1746 /* leave just 1 thread for normal RPCs */
1747 running = PTLRPC_NTHRS_INIT;
1748 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1755 if (ptlrpc_nrs_req_throttling_nolock(svcpt, false))
1758 if (svcpt->scp_nreqs_active < running - 2)
1761 if (svcpt->scp_nreqs_active >= running - 1)
1764 return svcpt->scp_nhreqs_active > 0 || !nrs_svcpt_has_hp(svcpt);
1767 static bool ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1770 return ptlrpc_server_allow_normal(svcpt, force) &&
1771 ptlrpc_nrs_req_pending_nolock(svcpt, false);
1775 * Returns true if there are requests available in incoming
1776 * request queue for processing and it is allowed to fetch them.
1777 * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1778 * to get reliable result
1779 * \see ptlrpc_server_allow_normal
1780 * \see ptlrpc_server_allow high
1783 ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt, bool force)
1785 return ptlrpc_server_high_pending(svcpt, force) ||
1786 ptlrpc_server_normal_pending(svcpt, force);
1790 * Fetch a request for processing from queue of unprocessed requests.
1791 * Favors high-priority requests.
1792 * Returns a pointer to fetched request.
1794 static struct ptlrpc_request *
1795 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force)
1797 struct ptlrpc_request *req = NULL;
1800 spin_lock(&svcpt->scp_req_lock);
1802 if (ptlrpc_server_high_pending(svcpt, force)) {
1803 req = ptlrpc_nrs_req_get_nolock(svcpt, true, force);
1805 svcpt->scp_hreq_count++;
1810 if (ptlrpc_server_normal_pending(svcpt, force)) {
1811 req = ptlrpc_nrs_req_get_nolock(svcpt, false, force);
1813 svcpt->scp_hreq_count = 0;
1818 spin_unlock(&svcpt->scp_req_lock);
1822 svcpt->scp_nreqs_active++;
1824 svcpt->scp_nhreqs_active++;
1826 spin_unlock(&svcpt->scp_req_lock);
1828 if (likely(req->rq_export))
1829 class_export_rpc_inc(req->rq_export);
1835 * Handle freshly incoming reqs, add to timed early reply list,
1836 * pass on to regular request queue.
1837 * All incoming requests pass through here before getting into
1838 * ptlrpc_server_handle_req later on.
1841 ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt,
1842 struct ptlrpc_thread *thread)
1844 struct ptlrpc_service *svc = svcpt->scp_service;
1845 struct ptlrpc_request *req;
1850 spin_lock(&svcpt->scp_lock);
1851 if (list_empty(&svcpt->scp_req_incoming)) {
1852 spin_unlock(&svcpt->scp_lock);
1856 req = list_entry(svcpt->scp_req_incoming.next,
1857 struct ptlrpc_request, rq_list);
1858 list_del_init(&req->rq_list);
1859 svcpt->scp_nreqs_incoming--;
1860 /* Consider this still a "queued" request as far as stats are
1862 spin_unlock(&svcpt->scp_lock);
1864 /* go through security check/transform */
1865 rc = sptlrpc_svc_unwrap_request(req);
1869 case SECSVC_COMPLETE:
1870 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1879 * for null-flavored rpc, msg has been unpacked by sptlrpc, although
1880 * redo it wouldn't be harmful.
1882 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
1883 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
1885 CERROR("error unpacking request: ptl %d from %s "
1886 "x"LPU64"\n", svc->srv_req_portal,
1887 libcfs_id2str(req->rq_peer), req->rq_xid);
1892 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1894 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
1895 LPU64"\n", svc->srv_req_portal,
1896 libcfs_id2str(req->rq_peer), req->rq_xid);
1900 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
1901 lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
1902 CERROR("drop incoming rpc opc %u, x"LPU64"\n",
1903 cfs_fail_val, req->rq_xid);
1908 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1909 CERROR("wrong packet type received (type=%u) from %s\n",
1910 lustre_msg_get_type(req->rq_reqmsg),
1911 libcfs_id2str(req->rq_peer));
1915 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1919 req->rq_bulk_write = 1;
1923 case MGS_CONFIG_READ:
1924 req->rq_bulk_read = 1;
1928 CDEBUG(D_RPCTRACE, "got req x"LPU64"\n", req->rq_xid);
1930 req->rq_export = class_conn2export(
1931 lustre_msg_get_handle(req->rq_reqmsg));
1932 if (req->rq_export) {
1933 rc = ptlrpc_check_req(req);
1935 rc = sptlrpc_target_export_check(req->rq_export, req);
1937 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
1938 "illegal security flavor,");
1943 ptlrpc_update_export_timer(req->rq_export, 0);
1946 /* req_in handling should/must be fast */
1947 if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5)
1948 DEBUG_REQ(D_WARNING, req, "Slow req_in handling "CFS_DURATION_T"s",
1949 cfs_time_sub(cfs_time_current_sec(),
1950 req->rq_arrival_time.tv_sec));
1952 /* Set rpc server deadline and add it to the timed list */
1953 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1954 MSGHDR_AT_SUPPORT) ?
1955 /* The max time the client expects us to take */
1956 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1958 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1959 if (unlikely(deadline == 0)) {
1960 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1964 /* Skip early reply */
1965 if (OBD_FAIL_PRECHECK(OBD_FAIL_MDS_RESEND))
1966 req->rq_deadline += obd_timeout;
1968 req->rq_svc_thread = thread;
1969 if (thread != NULL) {
1970 /* initialize request session, it is needed for request
1971 * processing by target */
1972 rc = lu_context_init(&req->rq_session, LCT_SERVER_SESSION |
1975 CERROR("%s: failure to initialize session: rc = %d\n",
1976 thread->t_name, rc);
1979 req->rq_session.lc_thread = thread;
1980 lu_context_enter(&req->rq_session);
1981 thread->t_env->le_ses = &req->rq_session;
1984 ptlrpc_at_add_timed(req);
1986 /* Move it over to the request processing queue */
1987 rc = ptlrpc_server_request_add(svcpt, req);
1991 wake_up(&svcpt->scp_waitq);
1995 ptlrpc_server_finish_request(svcpt, req);
2001 * Main incoming request handling logic.
2002 * Calls handler function from service to do actual processing.
2005 ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
2006 struct ptlrpc_thread *thread)
2008 struct ptlrpc_service *svc = svcpt->scp_service;
2009 struct ptlrpc_request *request;
2010 struct timeval work_start;
2011 struct timeval work_end;
2017 request = ptlrpc_server_request_get(svcpt, false);
2018 if (request == NULL)
2021 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
2022 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
2023 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
2024 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
2026 if (unlikely(fail_opc)) {
2027 if (request->rq_export && request->rq_ops)
2028 OBD_FAIL_TIMEOUT(fail_opc, 4);
2031 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
2033 if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
2034 libcfs_debug_dumplog();
2036 do_gettimeofday(&work_start);
2037 timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
2038 if (likely(svc->srv_stats != NULL)) {
2039 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
2041 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
2042 svcpt->scp_nreqs_incoming);
2043 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
2044 svcpt->scp_nreqs_active);
2045 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
2046 at_get(&svcpt->scp_at_estimate));
2049 if (likely(request->rq_export)) {
2050 if (unlikely(ptlrpc_check_req(request)))
2052 ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
2055 /* Discard requests queued for longer than the deadline.
2056 The deadline is increased if we send an early reply. */
2057 if (cfs_time_current_sec() > request->rq_deadline) {
2058 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s"
2059 ": deadline "CFS_DURATION_T":"CFS_DURATION_T"s ago\n",
2060 libcfs_id2str(request->rq_peer),
2061 cfs_time_sub(request->rq_deadline,
2062 request->rq_arrival_time.tv_sec),
2063 cfs_time_sub(cfs_time_current_sec(),
2064 request->rq_deadline));
2068 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
2069 "%s:%s+%d:%d:x"LPU64":%s:%d\n", current_comm(),
2070 (request->rq_export ?
2071 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2072 (request->rq_export ?
2073 atomic_read(&request->rq_export->exp_refcount) : -99),
2074 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
2075 libcfs_id2str(request->rq_peer),
2076 lustre_msg_get_opc(request->rq_reqmsg));
2078 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
2079 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
2081 CDEBUG(D_NET, "got req "LPU64"\n", request->rq_xid);
2083 /* re-assign request and sesson thread to the current one */
2084 request->rq_svc_thread = thread;
2085 if (thread != NULL) {
2086 LASSERT(request->rq_session.lc_thread == NULL);
2087 request->rq_session.lc_thread = thread;
2088 thread->t_env->le_ses = &request->rq_session;
2090 svc->srv_ops.so_req_handler(request);
2092 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
2095 if (unlikely(cfs_time_current_sec() > request->rq_deadline)) {
2096 DEBUG_REQ(D_WARNING, request, "Request took longer "
2097 "than estimated ("CFS_DURATION_T":"CFS_DURATION_T"s);"
2098 " client may timeout.",
2099 cfs_time_sub(request->rq_deadline,
2100 request->rq_arrival_time.tv_sec),
2101 cfs_time_sub(cfs_time_current_sec(),
2102 request->rq_deadline));
2105 do_gettimeofday(&work_end);
2106 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
2107 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
2108 "%s:%s+%d:%d:x"LPU64":%s:%d Request procesed in "
2109 "%ldus (%ldus total) trans "LPU64" rc %d/%d\n",
2111 (request->rq_export ?
2112 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2113 (request->rq_export ?
2114 atomic_read(&request->rq_export->exp_refcount) : -99),
2115 lustre_msg_get_status(request->rq_reqmsg),
2117 libcfs_id2str(request->rq_peer),
2118 lustre_msg_get_opc(request->rq_reqmsg),
2120 cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL),
2121 (request->rq_repmsg ?
2122 lustre_msg_get_transno(request->rq_repmsg) :
2123 request->rq_transno),
2125 (request->rq_repmsg ?
2126 lustre_msg_get_status(request->rq_repmsg) : -999));
2127 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
2128 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
2129 int opc = opcode_offset(op);
2130 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
2131 LASSERT(opc < LUSTRE_MAX_OPCODES);
2132 lprocfs_counter_add(svc->srv_stats,
2133 opc + EXTRA_MAX_OPCODES,
2137 if (unlikely(request->rq_early_count)) {
2138 DEBUG_REQ(D_ADAPTTO, request,
2139 "sent %d early replies before finishing in "
2141 request->rq_early_count,
2142 cfs_time_sub(work_end.tv_sec,
2143 request->rq_arrival_time.tv_sec));
2146 ptlrpc_server_finish_active_request(svcpt, request);
2152 * An internal function to process a single reply state object.
2155 ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
2157 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
2158 struct ptlrpc_service *svc = svcpt->scp_service;
2159 struct obd_export *exp;
2164 exp = rs->rs_export;
2166 LASSERT(rs->rs_difficult);
2167 LASSERT(rs->rs_scheduled);
2168 LASSERT(list_empty(&rs->rs_list));
2170 spin_lock(&exp->exp_lock);
2171 /* Noop if removed already */
2172 list_del_init(&rs->rs_exp_list);
2173 spin_unlock(&exp->exp_lock);
2175 /* The disk commit callback holds exp_uncommitted_replies_lock while it
2176 * iterates over newly committed replies, removing them from
2177 * exp_uncommitted_replies. It then drops this lock and schedules the
2178 * replies it found for handling here.
2180 * We can avoid contention for exp_uncommitted_replies_lock between the
2181 * HRT threads and further commit callbacks by checking rs_committed
2182 * which is set in the commit callback while it holds both
2183 * rs_lock and exp_uncommitted_reples.
2185 * If we see rs_committed clear, the commit callback _may_ not have
2186 * handled this reply yet and we race with it to grab
2187 * exp_uncommitted_replies_lock before removing the reply from
2188 * exp_uncommitted_replies. Note that if we lose the race and the
2189 * reply has already been removed, list_del_init() is a noop.
2191 * If we see rs_committed set, we know the commit callback is handling,
2192 * or has handled this reply since store reordering might allow us to
2193 * see rs_committed set out of sequence. But since this is done
2194 * holding rs_lock, we can be sure it has all completed once we hold
2195 * rs_lock, which we do right next.
2197 if (!rs->rs_committed) {
2198 spin_lock(&exp->exp_uncommitted_replies_lock);
2199 list_del_init(&rs->rs_obd_list);
2200 spin_unlock(&exp->exp_uncommitted_replies_lock);
2203 spin_lock(&rs->rs_lock);
2205 been_handled = rs->rs_handled;
2208 nlocks = rs->rs_nlocks; /* atomic "steal", but */
2209 rs->rs_nlocks = 0; /* locks still on rs_locks! */
2211 if (nlocks == 0 && !been_handled) {
2212 /* If we see this, we should already have seen the warning
2213 * in mds_steal_ack_locks() */
2214 CDEBUG(D_HA, "All locks stolen from rs %p x"LPD64".t"LPD64
2217 rs->rs_xid, rs->rs_transno, rs->rs_opc,
2218 libcfs_nid2str(exp->exp_connection->c_peer.nid));
2221 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
2222 spin_unlock(&rs->rs_lock);
2224 if (!been_handled && rs->rs_on_net) {
2225 LNetMDUnlink(rs->rs_md_h);
2226 /* Ignore return code; we're racing with completion */
2229 while (nlocks-- > 0)
2230 ldlm_lock_decref(&rs->rs_locks[nlocks],
2231 rs->rs_modes[nlocks]);
2233 spin_lock(&rs->rs_lock);
2236 rs->rs_scheduled = 0;
2238 if (!rs->rs_on_net) {
2240 spin_unlock(&rs->rs_lock);
2242 class_export_put (exp);
2243 rs->rs_export = NULL;
2244 ptlrpc_rs_decref(rs);
2245 if (atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
2246 svc->srv_is_stopping)
2247 wake_up_all(&svcpt->scp_waitq);
2251 /* still on the net; callback will schedule */
2252 spin_unlock(&rs->rs_lock);
2258 ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
2260 int avail = svcpt->scp_nrqbds_posted;
2261 int low_water = test_req_buffer_pressure ? 0 :
2262 svcpt->scp_service->srv_nbuf_per_group / 2;
2264 /* NB I'm not locking; just looking. */
2266 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
2267 * allowed the request history to grow out of control. We could put a
2268 * sanity check on that here and cull some history if we need the
2271 if (avail <= low_water)
2272 ptlrpc_grow_req_bufs(svcpt, 1);
2274 if (svcpt->scp_service->srv_stats) {
2275 lprocfs_counter_add(svcpt->scp_service->srv_stats,
2276 PTLRPC_REQBUF_AVAIL_CNTR, avail);
2281 ptlrpc_retry_rqbds(void *arg)
2283 struct ptlrpc_service_part *svcpt = (struct ptlrpc_service_part *)arg;
2285 svcpt->scp_rqbd_timeout = 0;
2290 ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
2292 return svcpt->scp_nreqs_active <
2293 svcpt->scp_nthrs_running - 1 -
2294 (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL);
2298 * allowed to create more threads
2299 * user can call it w/o any lock but need to hold
2300 * ptlrpc_service_part::scp_lock to get reliable result
2303 ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
2305 return svcpt->scp_nthrs_running +
2306 svcpt->scp_nthrs_starting <
2307 svcpt->scp_service->srv_nthrs_cpt_limit;
2311 * too many requests and allowed to create more threads
2314 ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
2316 return !ptlrpc_threads_enough(svcpt) &&
2317 ptlrpc_threads_increasable(svcpt);
2321 ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2323 return thread_is_stopping(thread) ||
2324 thread->t_svcpt->scp_service->srv_is_stopping;
2328 ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
2330 return !list_empty(&svcpt->scp_rqbd_idle) &&
2331 svcpt->scp_rqbd_timeout == 0;
2335 ptlrpc_at_check(struct ptlrpc_service_part *svcpt)
2337 return svcpt->scp_at_check;
2341 * requests wait on preprocessing
2342 * user can call it w/o any lock but need to hold
2343 * ptlrpc_service_part::scp_lock to get reliable result
2346 ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
2348 return !list_empty(&svcpt->scp_req_incoming);
2351 static __attribute__((__noinline__)) int
2352 ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
2353 struct ptlrpc_thread *thread)
2355 /* Don't exit while there are replies to be handled */
2356 struct l_wait_info lwi = LWI_TIMEOUT(svcpt->scp_rqbd_timeout,
2357 ptlrpc_retry_rqbds, svcpt);
2359 lc_watchdog_disable(thread->t_watchdog);
2363 l_wait_event_exclusive_head(svcpt->scp_waitq,
2364 ptlrpc_thread_stopping(thread) ||
2365 ptlrpc_server_request_incoming(svcpt) ||
2366 ptlrpc_server_request_pending(svcpt, false) ||
2367 ptlrpc_rqbd_pending(svcpt) ||
2368 ptlrpc_at_check(svcpt), &lwi);
2370 if (ptlrpc_thread_stopping(thread))
2373 lc_watchdog_touch(thread->t_watchdog,
2374 ptlrpc_server_get_timeout(svcpt));
2379 * Main thread body for service threads.
2380 * Waits in a loop waiting for new requests to process to appear.
2381 * Every time an incoming requests is added to its queue, a waitq
2382 * is woken up and one of the threads will handle it.
2384 static int ptlrpc_main(void *arg)
2386 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
2387 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2388 struct ptlrpc_service *svc = svcpt->scp_service;
2389 struct ptlrpc_reply_state *rs;
2390 struct group_info *ginfo = NULL;
2392 int counter = 0, rc = 0;
2395 thread->t_pid = current_pid();
2396 unshare_fs_struct();
2398 /* NB: we will call cfs_cpt_bind() for all threads, because we
2399 * might want to run lustre server only on a subset of system CPUs,
2400 * in that case ->scp_cpt is CFS_CPT_ANY */
2401 rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
2403 CWARN("%s: failed to bind %s on CPT %d\n",
2404 svc->srv_name, thread->t_name, svcpt->scp_cpt);
2407 ginfo = groups_alloc(0);
2413 set_current_groups(ginfo);
2414 put_group_info(ginfo);
2416 if (svc->srv_ops.so_thr_init != NULL) {
2417 rc = svc->srv_ops.so_thr_init(thread);
2428 rc = lu_context_init(&env->le_ctx,
2429 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2433 thread->t_env = env;
2434 env->le_ctx.lc_thread = thread;
2435 env->le_ctx.lc_cookie = 0x6;
2437 while (!list_empty(&svcpt->scp_rqbd_idle)) {
2438 rc = ptlrpc_server_post_idle_rqbds(svcpt);
2442 CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
2443 svc->srv_name, svcpt->scp_cpt, rc);
2447 /* Alloc reply state structure for this one */
2448 OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2454 spin_lock(&svcpt->scp_lock);
2456 LASSERT(thread_is_starting(thread));
2457 thread_clear_flags(thread, SVC_STARTING);
2459 LASSERT(svcpt->scp_nthrs_starting == 1);
2460 svcpt->scp_nthrs_starting--;
2462 /* SVC_STOPPING may already be set here if someone else is trying
2463 * to stop the service while this new thread has been dynamically
2464 * forked. We still set SVC_RUNNING to let our creator know that
2465 * we are now running, however we will exit as soon as possible */
2466 thread_add_flags(thread, SVC_RUNNING);
2467 svcpt->scp_nthrs_running++;
2468 spin_unlock(&svcpt->scp_lock);
2470 /* wake up our creator in case he's still waiting. */
2471 wake_up(&thread->t_ctl_waitq);
2473 thread->t_watchdog = lc_watchdog_add(ptlrpc_server_get_timeout(svcpt),
2476 spin_lock(&svcpt->scp_rep_lock);
2477 list_add(&rs->rs_list, &svcpt->scp_rep_idle);
2478 wake_up(&svcpt->scp_rep_waitq);
2479 spin_unlock(&svcpt->scp_rep_lock);
2481 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2482 svcpt->scp_nthrs_running);
2484 /* XXX maintain a list of all managed devices: insert here */
2485 while (!ptlrpc_thread_stopping(thread)) {
2486 if (ptlrpc_wait_event(svcpt, thread))
2489 ptlrpc_check_rqbd_pool(svcpt);
2491 if (ptlrpc_threads_need_create(svcpt)) {
2492 /* Ignore return code - we tried... */
2493 ptlrpc_start_thread(svcpt, 0);
2496 /* reset le_ses to initial state */
2498 /* Process all incoming reqs before handling any */
2499 if (ptlrpc_server_request_incoming(svcpt)) {
2500 lu_context_enter(&env->le_ctx);
2501 ptlrpc_server_handle_req_in(svcpt, thread);
2502 lu_context_exit(&env->le_ctx);
2504 /* but limit ourselves in case of flood */
2505 if (counter++ < 100)
2510 if (ptlrpc_at_check(svcpt))
2511 ptlrpc_at_check_timed(svcpt);
2513 if (ptlrpc_server_request_pending(svcpt, false)) {
2514 lu_context_enter(&env->le_ctx);
2515 ptlrpc_server_handle_request(svcpt, thread);
2516 lu_context_exit(&env->le_ctx);
2519 if (ptlrpc_rqbd_pending(svcpt) &&
2520 ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
2521 /* I just failed to repost request buffers.
2522 * Wait for a timeout (unless something else
2523 * happens) before I try again */
2524 svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
2525 CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
2526 svcpt->scp_nrqbds_posted);
2530 lc_watchdog_delete(thread->t_watchdog);
2531 thread->t_watchdog = NULL;
2535 * deconstruct service specific state created by ptlrpc_start_thread()
2537 if (svc->srv_ops.so_thr_done != NULL)
2538 svc->srv_ops.so_thr_done(thread);
2541 lu_context_fini(&env->le_ctx);
2545 CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n",
2546 thread, thread->t_pid, thread->t_id, rc);
2548 spin_lock(&svcpt->scp_lock);
2549 if (thread_test_and_clear_flags(thread, SVC_STARTING))
2550 svcpt->scp_nthrs_starting--;
2552 if (thread_test_and_clear_flags(thread, SVC_RUNNING)) {
2553 /* must know immediately */
2554 svcpt->scp_nthrs_running--;
2558 thread_add_flags(thread, SVC_STOPPED);
2560 wake_up(&thread->t_ctl_waitq);
2561 spin_unlock(&svcpt->scp_lock);
2566 static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
2567 struct list_head *replies)
2571 spin_lock(&hrt->hrt_lock);
2573 list_splice_init(&hrt->hrt_queue, replies);
2574 result = ptlrpc_hr.hr_stopping || !list_empty(replies);
2576 spin_unlock(&hrt->hrt_lock);
2581 * Main body of "handle reply" function.
2582 * It processes acked reply states
2584 static int ptlrpc_hr_main(void *arg)
2586 struct ptlrpc_hr_thread *hrt = (struct ptlrpc_hr_thread *)arg;
2587 struct ptlrpc_hr_partition *hrp = hrt->hrt_partition;
2588 struct list_head replies;
2589 char threadname[20];
2592 INIT_LIST_HEAD(&replies);
2593 snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
2594 hrp->hrp_cpt, hrt->hrt_id);
2595 unshare_fs_struct();
2597 rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
2599 CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
2600 threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
2603 atomic_inc(&hrp->hrp_nstarted);
2604 wake_up(&ptlrpc_hr.hr_waitq);
2606 while (!ptlrpc_hr.hr_stopping) {
2607 l_wait_condition(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
2609 while (!list_empty(&replies)) {
2610 struct ptlrpc_reply_state *rs;
2612 rs = list_entry(replies.prev,
2613 struct ptlrpc_reply_state,
2615 list_del_init(&rs->rs_list);
2616 ptlrpc_handle_rs(rs);
2620 atomic_inc(&hrp->hrp_nstopped);
2621 wake_up(&ptlrpc_hr.hr_waitq);
2626 static void ptlrpc_stop_hr_threads(void)
2628 struct ptlrpc_hr_partition *hrp;
2632 ptlrpc_hr.hr_stopping = 1;
2634 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2635 if (hrp->hrp_thrs == NULL)
2636 continue; /* uninitialized */
2637 for (j = 0; j < hrp->hrp_nthrs; j++)
2638 wake_up_all(&hrp->hrp_thrs[j].hrt_waitq);
2641 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2642 if (hrp->hrp_thrs == NULL)
2643 continue; /* uninitialized */
2644 wait_event(ptlrpc_hr.hr_waitq,
2645 atomic_read(&hrp->hrp_nstopped) ==
2646 atomic_read(&hrp->hrp_nstarted));
2650 static int ptlrpc_start_hr_threads(void)
2652 struct ptlrpc_hr_partition *hrp;
2657 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2660 for (j = 0; j < hrp->hrp_nthrs; j++) {
2661 struct ptlrpc_hr_thread *hrt = &hrp->hrp_thrs[j];
2662 struct task_struct *task;
2664 task = kthread_run(ptlrpc_hr_main,
2666 "ptlrpc_hr%02d_%03d",
2675 wait_event(ptlrpc_hr.hr_waitq,
2676 atomic_read(&hrp->hrp_nstarted) == j);
2679 CERROR("cannot start reply handler thread %d:%d: "
2680 "rc = %d\n", i, j, rc);
2681 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 struct list_head zombie;
2697 CDEBUG(D_INFO, "Stopping threads for service %s\n",
2698 svcpt->scp_service->srv_name);
2700 INIT_LIST_HEAD(&zombie);
2701 spin_lock(&svcpt->scp_lock);
2702 /* let the thread know that we would like it to stop asap */
2703 list_for_each_entry(thread, &svcpt->scp_threads, t_link) {
2704 CDEBUG(D_INFO, "Stopping thread %s #%u\n",
2705 svcpt->scp_service->srv_thread_name, thread->t_id);
2706 thread_add_flags(thread, SVC_STOPPING);
2709 wake_up_all(&svcpt->scp_waitq);
2711 while (!list_empty(&svcpt->scp_threads)) {
2712 thread = list_entry(svcpt->scp_threads.next,
2713 struct ptlrpc_thread, t_link);
2714 if (thread_is_stopped(thread)) {
2715 list_del(&thread->t_link);
2716 list_add(&thread->t_link, &zombie);
2719 spin_unlock(&svcpt->scp_lock);
2721 CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
2722 svcpt->scp_service->srv_thread_name, thread->t_id);
2723 l_wait_event(thread->t_ctl_waitq,
2724 thread_is_stopped(thread), &lwi);
2726 spin_lock(&svcpt->scp_lock);
2729 spin_unlock(&svcpt->scp_lock);
2731 while (!list_empty(&zombie)) {
2732 thread = list_entry(zombie.next,
2733 struct ptlrpc_thread, t_link);
2734 list_del(&thread->t_link);
2735 OBD_FREE_PTR(thread);
2741 * Stops all threads of a particular service \a svc
2743 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2745 struct ptlrpc_service_part *svcpt;
2749 ptlrpc_service_for_each_part(svcpt, i, svc) {
2750 if (svcpt->scp_service != NULL)
2751 ptlrpc_svcpt_stop_threads(svcpt);
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);
2788 int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
2790 struct l_wait_info lwi = { 0 };
2791 struct ptlrpc_thread *thread;
2792 struct ptlrpc_service *svc;
2793 struct task_struct *task;
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 init_waitqueue_head(&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 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 task = kthread_run(ptlrpc_main, thread, "%s", thread->t_name);
2864 CERROR("cannot start thread '%s': rc = %d\n",
2865 thread->t_name, rc);
2866 spin_lock(&svcpt->scp_lock);
2867 --svcpt->scp_nthrs_starting;
2868 if (thread_is_stopping(thread)) {
2869 /* this ptlrpc_thread is being hanled
2870 * by ptlrpc_svcpt_stop_threads now
2872 thread_add_flags(thread, SVC_STOPPED);
2873 wake_up(&thread->t_ctl_waitq);
2874 spin_unlock(&svcpt->scp_lock);
2876 list_del(&thread->t_link);
2877 spin_unlock(&svcpt->scp_lock);
2878 OBD_FREE_PTR(thread);
2886 l_wait_event(thread->t_ctl_waitq,
2887 thread_is_running(thread) || thread_is_stopped(thread),
2890 rc = thread_is_stopped(thread) ? thread->t_id : 0;
2894 int ptlrpc_hr_init(void)
2896 struct ptlrpc_hr_partition *hrp;
2897 struct ptlrpc_hr_thread *hrt;
2904 memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
2905 ptlrpc_hr.hr_cpt_table = cfs_cpt_table;
2907 ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
2909 if (ptlrpc_hr.hr_partitions == NULL)
2912 init_waitqueue_head(&ptlrpc_hr.hr_waitq);
2914 weight = cfs_cpu_ht_nsiblings(0);
2916 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2919 atomic_set(&hrp->hrp_nstarted, 0);
2920 atomic_set(&hrp->hrp_nstopped, 0);
2922 hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, i);
2923 hrp->hrp_nthrs /= weight;
2925 LASSERT(hrp->hrp_nthrs > 0);
2926 OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, i,
2927 hrp->hrp_nthrs * sizeof(*hrt));
2928 if (hrp->hrp_thrs == NULL)
2929 GOTO(out, rc = -ENOMEM);
2931 for (j = 0; j < hrp->hrp_nthrs; j++) {
2932 hrt = &hrp->hrp_thrs[j];
2935 hrt->hrt_partition = hrp;
2936 init_waitqueue_head(&hrt->hrt_waitq);
2937 spin_lock_init(&hrt->hrt_lock);
2938 INIT_LIST_HEAD(&hrt->hrt_queue);
2942 rc = ptlrpc_start_hr_threads();
2949 void ptlrpc_hr_fini(void)
2951 struct ptlrpc_hr_partition *hrp;
2954 if (ptlrpc_hr.hr_partitions == NULL)
2957 ptlrpc_stop_hr_threads();
2959 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2960 if (hrp->hrp_thrs != NULL) {
2961 OBD_FREE(hrp->hrp_thrs,
2962 hrp->hrp_nthrs * sizeof(hrp->hrp_thrs[0]));
2966 cfs_percpt_free(ptlrpc_hr.hr_partitions);
2967 ptlrpc_hr.hr_partitions = NULL;
2972 * Wait until all already scheduled replies are processed.
2974 static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
2978 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
2981 rc = l_wait_event(svcpt->scp_waitq,
2982 atomic_read(&svcpt->scp_nreps_difficult) == 0, &lwi);
2985 CWARN("Unexpectedly long timeout %s %p\n",
2986 svcpt->scp_service->srv_name, svcpt->scp_service);
2991 ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
2993 struct ptlrpc_service_part *svcpt;
2996 /* early disarm AT timer... */
2997 ptlrpc_service_for_each_part(svcpt, i, svc) {
2998 if (svcpt->scp_service != NULL)
2999 cfs_timer_disarm(&svcpt->scp_at_timer);
3004 ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
3006 struct ptlrpc_service_part *svcpt;
3007 struct ptlrpc_request_buffer_desc *rqbd;
3008 struct l_wait_info lwi;
3012 /* All history will be culled when the next request buffer is
3013 * freed in ptlrpc_service_purge_all() */
3014 svc->srv_hist_nrqbds_cpt_max = 0;
3016 rc = LNetClearLazyPortal(svc->srv_req_portal);
3019 ptlrpc_service_for_each_part(svcpt, i, svc) {
3020 if (svcpt->scp_service == NULL)
3023 /* Unlink all the request buffers. This forces a 'final'
3024 * event with its 'unlink' flag set for each posted rqbd */
3025 list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
3027 rc = LNetMDUnlink(rqbd->rqbd_md_h);
3028 LASSERT(rc == 0 || rc == -ENOENT);
3032 ptlrpc_service_for_each_part(svcpt, i, svc) {
3033 if (svcpt->scp_service == NULL)
3036 /* Wait for the network to release any buffers
3037 * it's currently filling */
3038 spin_lock(&svcpt->scp_lock);
3039 while (svcpt->scp_nrqbds_posted != 0) {
3040 spin_unlock(&svcpt->scp_lock);
3041 /* Network access will complete in finite time but
3042 * the HUGE timeout lets us CWARN for visibility
3043 * of sluggish NALs */
3044 lwi = LWI_TIMEOUT_INTERVAL(
3045 cfs_time_seconds(LONG_UNLINK),
3046 cfs_time_seconds(1), NULL, NULL);
3047 rc = l_wait_event(svcpt->scp_waitq,
3048 svcpt->scp_nrqbds_posted == 0, &lwi);
3049 if (rc == -ETIMEDOUT) {
3050 CWARN("Service %s waiting for "
3051 "request buffers\n",
3052 svcpt->scp_service->srv_name);
3054 spin_lock(&svcpt->scp_lock);
3056 spin_unlock(&svcpt->scp_lock);
3061 ptlrpc_service_purge_all(struct ptlrpc_service *svc)
3063 struct ptlrpc_service_part *svcpt;
3064 struct ptlrpc_request_buffer_desc *rqbd;
3065 struct ptlrpc_request *req;
3066 struct ptlrpc_reply_state *rs;
3069 ptlrpc_service_for_each_part(svcpt, i, svc) {
3070 if (svcpt->scp_service == NULL)
3073 spin_lock(&svcpt->scp_rep_lock);
3074 while (!list_empty(&svcpt->scp_rep_active)) {
3075 rs = list_entry(svcpt->scp_rep_active.next,
3076 struct ptlrpc_reply_state, rs_list);
3077 spin_lock(&rs->rs_lock);
3078 ptlrpc_schedule_difficult_reply(rs);
3079 spin_unlock(&rs->rs_lock);
3081 spin_unlock(&svcpt->scp_rep_lock);
3083 /* purge the request queue. NB No new replies (rqbds
3084 * all unlinked) and no service threads, so I'm the only
3085 * thread noodling the request queue now */
3086 while (!list_empty(&svcpt->scp_req_incoming)) {
3087 req = list_entry(svcpt->scp_req_incoming.next,
3088 struct ptlrpc_request, rq_list);
3090 list_del(&req->rq_list);
3091 svcpt->scp_nreqs_incoming--;
3092 ptlrpc_server_finish_request(svcpt, req);
3095 while (ptlrpc_server_request_pending(svcpt, true)) {
3096 req = ptlrpc_server_request_get(svcpt, true);
3097 ptlrpc_server_finish_active_request(svcpt, req);
3100 LASSERT(list_empty(&svcpt->scp_rqbd_posted));
3101 LASSERT(svcpt->scp_nreqs_incoming == 0);
3102 LASSERT(svcpt->scp_nreqs_active == 0);
3103 /* history should have been culled by
3104 * ptlrpc_server_finish_request */
3105 LASSERT(svcpt->scp_hist_nrqbds == 0);
3107 /* Now free all the request buffers since nothing
3108 * references them any more... */
3110 while (!list_empty(&svcpt->scp_rqbd_idle)) {
3111 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
3112 struct ptlrpc_request_buffer_desc,
3114 ptlrpc_free_rqbd(rqbd);
3116 ptlrpc_wait_replies(svcpt);
3118 while (!list_empty(&svcpt->scp_rep_idle)) {
3119 rs = list_entry(svcpt->scp_rep_idle.next,
3120 struct ptlrpc_reply_state,
3122 list_del(&rs->rs_list);
3123 OBD_FREE_LARGE(rs, svc->srv_max_reply_size);
3129 ptlrpc_service_free(struct ptlrpc_service *svc)
3131 struct ptlrpc_service_part *svcpt;
3132 struct ptlrpc_at_array *array;
3135 ptlrpc_service_for_each_part(svcpt, i, svc) {
3136 if (svcpt->scp_service == NULL)
3139 /* In case somebody rearmed this in the meantime */
3140 cfs_timer_disarm(&svcpt->scp_at_timer);
3141 array = &svcpt->scp_at_array;
3143 if (array->paa_reqs_array != NULL) {
3144 OBD_FREE(array->paa_reqs_array,
3145 sizeof(struct list_head) * array->paa_size);
3146 array->paa_reqs_array = NULL;
3149 if (array->paa_reqs_count != NULL) {
3150 OBD_FREE(array->paa_reqs_count,
3151 sizeof(__u32) * array->paa_size);
3152 array->paa_reqs_count = NULL;
3156 ptlrpc_service_for_each_part(svcpt, i, svc)
3157 OBD_FREE_PTR(svcpt);
3159 if (svc->srv_cpts != NULL)
3160 cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts);
3162 OBD_FREE(svc, offsetof(struct ptlrpc_service,
3163 srv_parts[svc->srv_ncpts]));
3166 int ptlrpc_unregister_service(struct ptlrpc_service *service)
3170 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
3172 service->srv_is_stopping = 1;
3174 mutex_lock(&ptlrpc_all_services_mutex);
3175 list_del_init(&service->srv_list);
3176 mutex_unlock(&ptlrpc_all_services_mutex);
3178 ptlrpc_service_del_atimer(service);
3179 ptlrpc_stop_all_threads(service);
3181 ptlrpc_service_unlink_rqbd(service);
3182 ptlrpc_service_purge_all(service);
3183 ptlrpc_service_nrs_cleanup(service);
3185 ptlrpc_lprocfs_unregister_service(service);
3187 ptlrpc_service_free(service);
3191 EXPORT_SYMBOL(ptlrpc_unregister_service);
3194 * Returns 0 if the service is healthy.
3196 * Right now, it just checks to make sure that requests aren't languishing
3197 * in the queue. We'll use this health check to govern whether a node needs
3198 * to be shot, so it's intentionally non-aggressive. */
3199 static int ptlrpc_svcpt_health_check(struct ptlrpc_service_part *svcpt)
3201 struct ptlrpc_request *request = NULL;
3202 struct timeval right_now;
3205 do_gettimeofday(&right_now);
3207 spin_lock(&svcpt->scp_req_lock);
3208 /* How long has the next entry been waiting? */
3209 if (ptlrpc_server_high_pending(svcpt, true))
3210 request = ptlrpc_nrs_req_peek_nolock(svcpt, true);
3211 else if (ptlrpc_server_normal_pending(svcpt, true))
3212 request = ptlrpc_nrs_req_peek_nolock(svcpt, false);
3214 if (request == NULL) {
3215 spin_unlock(&svcpt->scp_req_lock);
3219 timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
3220 spin_unlock(&svcpt->scp_req_lock);
3222 if ((timediff / ONE_MILLION) >
3223 (AT_OFF ? obd_timeout * 3 / 2 : at_max)) {
3224 CERROR("%s: unhealthy - request has been waiting %lds\n",
3225 svcpt->scp_service->srv_name, timediff / ONE_MILLION);
3233 ptlrpc_service_health_check(struct ptlrpc_service *svc)
3235 struct ptlrpc_service_part *svcpt;
3241 ptlrpc_service_for_each_part(svcpt, i, svc) {
3242 int rc = ptlrpc_svcpt_health_check(svcpt);
3249 EXPORT_SYMBOL(ptlrpc_service_health_check);