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) 2011, 2012, Whamcloud, Inc.
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_hpreq_fini(struct ptlrpc_request *req);
69 static CFS_LIST_HEAD(ptlrpc_all_services);
70 cfs_spinlock_t ptlrpc_all_services_lock;
72 struct ptlrpc_request_buffer_desc *
73 ptlrpc_alloc_rqbd(struct ptlrpc_service_part *svcpt)
75 struct ptlrpc_service *svc = svcpt->scp_service;
76 struct ptlrpc_request_buffer_desc *rqbd;
78 OBD_CPT_ALLOC_PTR(rqbd, svc->srv_cptable, svcpt->scp_cpt);
82 rqbd->rqbd_svcpt = svcpt;
83 rqbd->rqbd_refcount = 0;
84 rqbd->rqbd_cbid.cbid_fn = request_in_callback;
85 rqbd->rqbd_cbid.cbid_arg = rqbd;
86 CFS_INIT_LIST_HEAD(&rqbd->rqbd_reqs);
87 OBD_CPT_ALLOC_LARGE(rqbd->rqbd_buffer, svc->srv_cptable,
88 svcpt->scp_cpt, svc->srv_buf_size);
89 if (rqbd->rqbd_buffer == NULL) {
94 cfs_spin_lock(&svcpt->scp_lock);
95 cfs_list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
96 svcpt->scp_nrqbds_total++;
97 cfs_spin_unlock(&svcpt->scp_lock);
103 ptlrpc_free_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
105 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
107 LASSERT(rqbd->rqbd_refcount == 0);
108 LASSERT(cfs_list_empty(&rqbd->rqbd_reqs));
110 cfs_spin_lock(&svcpt->scp_lock);
111 cfs_list_del(&rqbd->rqbd_list);
112 svcpt->scp_nrqbds_total--;
113 cfs_spin_unlock(&svcpt->scp_lock);
115 OBD_FREE_LARGE(rqbd->rqbd_buffer, svcpt->scp_service->srv_buf_size);
120 ptlrpc_grow_req_bufs(struct ptlrpc_service_part *svcpt, int post)
122 struct ptlrpc_service *svc = svcpt->scp_service;
123 struct ptlrpc_request_buffer_desc *rqbd;
127 for (i = 0; i < svc->srv_nbuf_per_group; i++) {
128 /* NB: another thread might be doing this as well, we need to
129 * make sure that it wouldn't over-allocate, see LU-1212. */
130 if (svcpt->scp_nrqbds_posted >= svc->srv_nbuf_per_group)
133 rqbd = ptlrpc_alloc_rqbd(svcpt);
136 CERROR("%s: Can't allocate request buffer\n",
144 "%s: allocate %d new %d-byte reqbufs (%d/%d left), rc = %d\n",
145 svc->srv_name, i, svc->srv_buf_size, svcpt->scp_nrqbds_posted,
146 svcpt->scp_nrqbds_total, rc);
149 rc = ptlrpc_server_post_idle_rqbds(svcpt);
155 * Part of Rep-Ack logic.
156 * Puts a lock and its mode into reply state assotiated to request reply.
159 ptlrpc_save_lock(struct ptlrpc_request *req,
160 struct lustre_handle *lock, int mode, int no_ack)
162 struct ptlrpc_reply_state *rs = req->rq_reply_state;
166 LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
168 if (req->rq_export->exp_disconnected) {
169 ldlm_lock_decref(lock, mode);
171 idx = rs->rs_nlocks++;
172 rs->rs_locks[idx] = *lock;
173 rs->rs_modes[idx] = mode;
174 rs->rs_difficult = 1;
175 rs->rs_no_ack = !!no_ack;
181 struct ptlrpc_hr_partition;
183 struct ptlrpc_hr_thread {
184 int hrt_id; /* thread ID */
185 cfs_spinlock_t hrt_lock;
186 cfs_waitq_t hrt_waitq;
187 cfs_list_t hrt_queue; /* RS queue */
188 struct ptlrpc_hr_partition *hrt_partition;
191 struct ptlrpc_hr_partition {
192 /* # of started threads */
193 cfs_atomic_t hrp_nstarted;
194 /* # of stopped threads */
195 cfs_atomic_t hrp_nstopped;
196 /* cpu partition id */
198 /* round-robin rotor for choosing thread */
200 /* total number of threads on this partition */
203 struct ptlrpc_hr_thread *hrp_thrs;
206 #define HRT_RUNNING 0
207 #define HRT_STOPPING 1
209 struct ptlrpc_hr_service {
210 /* CPU partition table, it's just cfs_cpt_table for now */
211 struct cfs_cpt_table *hr_cpt_table;
212 /** controller sleep waitq */
213 cfs_waitq_t hr_waitq;
214 unsigned int hr_stopping;
215 /** roundrobin rotor for non-affinity service */
216 unsigned int hr_rotor;
218 struct ptlrpc_hr_partition **hr_partitions;
222 cfs_list_t rsb_replies;
223 unsigned int rsb_n_replies;
224 struct ptlrpc_service_part *rsb_svcpt;
227 /** reply handling service. */
228 static struct ptlrpc_hr_service ptlrpc_hr;
231 * maximum mumber of replies scheduled in one batch
233 #define MAX_SCHEDULED 256
236 * Initialize a reply batch.
240 static void rs_batch_init(struct rs_batch *b)
242 memset(b, 0, sizeof *b);
243 CFS_INIT_LIST_HEAD(&b->rsb_replies);
247 * Choose an hr thread to dispatch requests to.
249 static struct ptlrpc_hr_thread *
250 ptlrpc_hr_select(struct ptlrpc_service_part *svcpt)
252 struct ptlrpc_hr_partition *hrp;
255 if (svcpt->scp_cpt >= 0 &&
256 svcpt->scp_service->srv_cptable == ptlrpc_hr.hr_cpt_table) {
257 /* directly match partition */
258 hrp = ptlrpc_hr.hr_partitions[svcpt->scp_cpt];
261 rotor = ptlrpc_hr.hr_rotor++;
262 rotor %= cfs_cpt_number(ptlrpc_hr.hr_cpt_table);
264 hrp = ptlrpc_hr.hr_partitions[rotor];
267 rotor = hrp->hrp_rotor++;
268 return &hrp->hrp_thrs[rotor % hrp->hrp_nthrs];
272 * Dispatch all replies accumulated in the batch to one from
273 * dedicated reply handling threads.
277 static void rs_batch_dispatch(struct rs_batch *b)
279 if (b->rsb_n_replies != 0) {
280 struct ptlrpc_hr_thread *hrt;
282 hrt = ptlrpc_hr_select(b->rsb_svcpt);
284 cfs_spin_lock(&hrt->hrt_lock);
285 cfs_list_splice_init(&b->rsb_replies, &hrt->hrt_queue);
286 cfs_spin_unlock(&hrt->hrt_lock);
288 cfs_waitq_signal(&hrt->hrt_waitq);
289 b->rsb_n_replies = 0;
294 * Add a reply to a batch.
295 * Add one reply object to a batch, schedule batched replies if overload.
300 static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs)
302 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
304 if (svcpt != b->rsb_svcpt || b->rsb_n_replies >= MAX_SCHEDULED) {
305 if (b->rsb_svcpt != NULL) {
306 rs_batch_dispatch(b);
307 cfs_spin_unlock(&b->rsb_svcpt->scp_rep_lock);
309 cfs_spin_lock(&svcpt->scp_rep_lock);
310 b->rsb_svcpt = svcpt;
312 cfs_spin_lock(&rs->rs_lock);
313 rs->rs_scheduled_ever = 1;
314 if (rs->rs_scheduled == 0) {
315 cfs_list_move(&rs->rs_list, &b->rsb_replies);
316 rs->rs_scheduled = 1;
319 rs->rs_committed = 1;
320 cfs_spin_unlock(&rs->rs_lock);
324 * Reply batch finalization.
325 * Dispatch remaining replies from the batch
326 * and release remaining spinlock.
330 static void rs_batch_fini(struct rs_batch *b)
332 if (b->rsb_svcpt != NULL) {
333 rs_batch_dispatch(b);
334 cfs_spin_unlock(&b->rsb_svcpt->scp_rep_lock);
338 #define DECLARE_RS_BATCH(b) struct rs_batch b
340 #else /* __KERNEL__ */
342 #define rs_batch_init(b) do{}while(0)
343 #define rs_batch_fini(b) do{}while(0)
344 #define rs_batch_add(b, r) ptlrpc_schedule_difficult_reply(r)
345 #define DECLARE_RS_BATCH(b)
347 #endif /* __KERNEL__ */
350 * Put reply state into a queue for processing because we received
351 * ACK from the client
353 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
356 struct ptlrpc_hr_thread *hrt;
359 LASSERT(cfs_list_empty(&rs->rs_list));
361 hrt = ptlrpc_hr_select(rs->rs_svcpt);
363 cfs_spin_lock(&hrt->hrt_lock);
364 cfs_list_add_tail(&rs->rs_list, &hrt->hrt_queue);
365 cfs_spin_unlock(&hrt->hrt_lock);
367 cfs_waitq_signal(&hrt->hrt_waitq);
370 cfs_list_add_tail(&rs->rs_list, &rs->rs_svcpt->scp_rep_queue);
375 ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs)
379 LASSERT_SPIN_LOCKED(&rs->rs_svcpt->scp_rep_lock);
380 LASSERT_SPIN_LOCKED(&rs->rs_lock);
381 LASSERT (rs->rs_difficult);
382 rs->rs_scheduled_ever = 1; /* flag any notification attempt */
384 if (rs->rs_scheduled) { /* being set up or already notified */
389 rs->rs_scheduled = 1;
390 cfs_list_del_init(&rs->rs_list);
391 ptlrpc_dispatch_difficult_reply(rs);
395 void ptlrpc_commit_replies(struct obd_export *exp)
397 struct ptlrpc_reply_state *rs, *nxt;
398 DECLARE_RS_BATCH(batch);
401 rs_batch_init(&batch);
402 /* Find any replies that have been committed and get their service
403 * to attend to complete them. */
405 /* CAVEAT EMPTOR: spinlock ordering!!! */
406 cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
407 cfs_list_for_each_entry_safe(rs, nxt, &exp->exp_uncommitted_replies,
409 LASSERT (rs->rs_difficult);
410 /* VBR: per-export last_committed */
411 LASSERT(rs->rs_export);
412 if (rs->rs_transno <= exp->exp_last_committed) {
413 cfs_list_del_init(&rs->rs_obd_list);
414 rs_batch_add(&batch, rs);
417 cfs_spin_unlock(&exp->exp_uncommitted_replies_lock);
418 rs_batch_fini(&batch);
423 ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt)
425 struct ptlrpc_request_buffer_desc *rqbd;
430 cfs_spin_lock(&svcpt->scp_lock);
432 if (cfs_list_empty(&svcpt->scp_rqbd_idle)) {
433 cfs_spin_unlock(&svcpt->scp_lock);
437 rqbd = cfs_list_entry(svcpt->scp_rqbd_idle.next,
438 struct ptlrpc_request_buffer_desc,
440 cfs_list_del(&rqbd->rqbd_list);
442 /* assume we will post successfully */
443 svcpt->scp_nrqbds_posted++;
444 cfs_list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_posted);
446 cfs_spin_unlock(&svcpt->scp_lock);
448 rc = ptlrpc_register_rqbd(rqbd);
455 cfs_spin_lock(&svcpt->scp_lock);
457 svcpt->scp_nrqbds_posted--;
458 cfs_list_del(&rqbd->rqbd_list);
459 cfs_list_add_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
461 /* Don't complain if no request buffers are posted right now; LNET
462 * won't drop requests because we set the portal lazy! */
464 cfs_spin_unlock(&svcpt->scp_lock);
469 static void ptlrpc_at_timer(unsigned long castmeharder)
471 struct ptlrpc_service_part *svcpt;
473 svcpt = (struct ptlrpc_service_part *)castmeharder;
475 svcpt->scp_at_check = 1;
476 svcpt->scp_at_checktime = cfs_time_current();
477 cfs_waitq_signal(&svcpt->scp_waitq);
481 ptlrpc_server_nthreads_check(struct ptlrpc_service *svc,
482 struct ptlrpc_service_conf *conf)
485 struct ptlrpc_service_thr_conf *tc = &conf->psc_thr;
492 * Common code for estimating & validating threads number.
493 * CPT affinity service could have percpt thread-pool instead
494 * of a global thread-pool, which means user might not always
495 * get the threads number they give it in conf::tc_nthrs_user
496 * even they did set. It's because we need to validate threads
497 * number for each CPT to guarantee each pool will have enough
498 * threads to keep the service healthy.
500 init = PTLRPC_NTHRS_INIT + (svc->srv_ops.so_hpreq_handler != NULL);
501 init = max_t(int, init, tc->tc_nthrs_init);
503 /* NB: please see comments in lustre_lnet.h for definition
504 * details of these members */
505 LASSERT(tc->tc_nthrs_max != 0);
507 if (tc->tc_nthrs_user != 0) {
508 /* In case there is a reason to test a service with many
509 * threads, we give a less strict check here, it can
510 * be up to 8 * nthrs_max */
511 total = min(tc->tc_nthrs_max * 8, tc->tc_nthrs_user);
512 nthrs = total / svc->srv_ncpts;
513 init = max(init, nthrs);
517 total = tc->tc_nthrs_max;
518 if (tc->tc_nthrs_base == 0) {
519 /* don't care about base threads number per partition,
520 * this is most for non-affinity service */
521 nthrs = total / svc->srv_ncpts;
525 nthrs = tc->tc_nthrs_base;
526 if (svc->srv_ncpts == 1) {
529 /* NB: Increase the base number if it's single partition
530 * and total number of cores/HTs is larger or equal to 4.
531 * result will always < 2 * nthrs_base */
532 weight = cfs_cpt_weight(svc->srv_cptable, CFS_CPT_ANY);
533 for (i = 1; (weight >> (i + 1)) != 0 && /* >= 4 cores/HTs */
534 (tc->tc_nthrs_base >> i) != 0; i++)
535 nthrs += tc->tc_nthrs_base >> i;
538 if (tc->tc_thr_factor != 0) {
539 int factor = tc->tc_thr_factor;
543 * User wants to increase number of threads with for
544 * each CPU core/HT, most likely the factor is larger then
545 * one thread/core because service threads are supposed to
546 * be blocked by lock or wait for IO.
549 * Amdahl's law says that adding processors wouldn't give
550 * a linear increasing of parallelism, so it's nonsense to
551 * have too many threads no matter how many cores/HTs
554 if (cfs_cpu_ht_nsiblings(0) > 1) { /* weight is # of HTs */
555 /* depress thread factor for hyper-thread */
556 factor = factor - (factor >> 1) + (factor >> 3);
559 weight = cfs_cpt_weight(svc->srv_cptable, 0);
562 for (; factor > 0 && weight > 0; factor--, weight -= fade)
563 nthrs += min(weight, fade) * factor;
566 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
567 nthrs = max(tc->tc_nthrs_base,
568 tc->tc_nthrs_max / svc->srv_ncpts);
571 nthrs = max(nthrs, tc->tc_nthrs_init);
572 svc->srv_nthrs_cpt_limit = nthrs;
573 svc->srv_nthrs_cpt_init = init;
575 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
576 LCONSOLE_WARN("%s: This service may have more threads (%d) "
577 "than the given soft limit (%d)\n",
578 svc->srv_name, nthrs * svc->srv_ncpts,
585 * Initialize percpt data for a service
588 ptlrpc_service_part_init(struct ptlrpc_service *svc,
589 struct ptlrpc_service_part *svcpt, int cpt)
591 struct ptlrpc_at_array *array;
596 svcpt->scp_cpt = cpt;
597 CFS_INIT_LIST_HEAD(&svcpt->scp_threads);
599 /* rqbd and incoming request queue */
600 cfs_spin_lock_init(&svcpt->scp_lock);
601 CFS_INIT_LIST_HEAD(&svcpt->scp_rqbd_idle);
602 CFS_INIT_LIST_HEAD(&svcpt->scp_rqbd_posted);
603 CFS_INIT_LIST_HEAD(&svcpt->scp_req_incoming);
604 cfs_waitq_init(&svcpt->scp_waitq);
605 /* history request & rqbd list */
606 CFS_INIT_LIST_HEAD(&svcpt->scp_hist_reqs);
607 CFS_INIT_LIST_HEAD(&svcpt->scp_hist_rqbds);
609 /* acitve requests and hp requests */
610 cfs_spin_lock_init(&svcpt->scp_req_lock);
611 CFS_INIT_LIST_HEAD(&svcpt->scp_req_pending);
612 CFS_INIT_LIST_HEAD(&svcpt->scp_hreq_pending);
615 cfs_spin_lock_init(&svcpt->scp_rep_lock);
616 CFS_INIT_LIST_HEAD(&svcpt->scp_rep_active);
618 CFS_INIT_LIST_HEAD(&svcpt->scp_rep_queue);
620 CFS_INIT_LIST_HEAD(&svcpt->scp_rep_idle);
621 cfs_waitq_init(&svcpt->scp_rep_waitq);
622 cfs_atomic_set(&svcpt->scp_nreps_difficult, 0);
624 /* adaptive timeout */
625 cfs_spin_lock_init(&svcpt->scp_at_lock);
626 array = &svcpt->scp_at_array;
628 size = at_est2timeout(at_max);
629 array->paa_size = size;
630 array->paa_count = 0;
631 array->paa_deadline = -1;
633 /* allocate memory for scp_at_array (ptlrpc_at_array) */
634 OBD_CPT_ALLOC(array->paa_reqs_array,
635 svc->srv_cptable, cpt, sizeof(cfs_list_t) * size);
636 if (array->paa_reqs_array == NULL)
639 for (index = 0; index < size; index++)
640 CFS_INIT_LIST_HEAD(&array->paa_reqs_array[index]);
642 OBD_CPT_ALLOC(array->paa_reqs_count,
643 svc->srv_cptable, cpt, sizeof(__u32) * size);
644 if (array->paa_reqs_count == NULL)
647 cfs_timer_init(&svcpt->scp_at_timer, ptlrpc_at_timer, svcpt);
648 /* At SOW, service time should be quick; 10s seems generous. If client
649 * timeout is less than this, we'll be sending an early reply. */
650 at_init(&svcpt->scp_at_estimate, 10, 0);
652 /* assign this before call ptlrpc_grow_req_bufs */
653 svcpt->scp_service = svc;
654 /* Now allocate the request buffers, but don't post them now */
655 rc = ptlrpc_grow_req_bufs(svcpt, 0);
656 /* We shouldn't be under memory pressure at startup, so
657 * fail if we can't allocate all our buffers at this time. */
664 if (array->paa_reqs_count != NULL) {
665 OBD_FREE(array->paa_reqs_count, sizeof(__u32) * size);
666 array->paa_reqs_count = NULL;
669 if (array->paa_reqs_array != NULL) {
670 OBD_FREE(array->paa_reqs_array,
671 sizeof(cfs_list_t) * array->paa_size);
672 array->paa_reqs_array = NULL;
679 * Initialize service on a given portal.
680 * This includes starting serving threads , allocating and posting rqbds and
683 struct ptlrpc_service *
684 ptlrpc_register_service(struct ptlrpc_service_conf *conf,
685 cfs_proc_dir_entry_t *proc_entry)
687 struct ptlrpc_service_cpt_conf *cconf = &conf->psc_cpt;
688 struct ptlrpc_service *service;
689 struct ptlrpc_service_part *svcpt;
690 struct cfs_cpt_table *cptable;
698 LASSERT(conf->psc_buf.bc_nbufs > 0);
699 LASSERT(conf->psc_buf.bc_buf_size >=
700 conf->psc_buf.bc_req_max_size + SPTLRPC_MAX_PAYLOAD);
701 LASSERT(conf->psc_thr.tc_ctx_tags != 0);
703 cptable = cconf->cc_cptable;
705 cptable = cfs_cpt_table;
707 if (!conf->psc_thr.tc_cpu_affinity) {
710 ncpts = cfs_cpt_number(cptable);
711 if (cconf->cc_pattern != NULL) {
712 struct cfs_expr_list *el;
714 rc = cfs_expr_list_parse(cconf->cc_pattern,
715 strlen(cconf->cc_pattern),
718 CERROR("%s: invalid CPT pattern string: %s",
719 conf->psc_name, cconf->cc_pattern);
720 RETURN(ERR_PTR(-EINVAL));
723 rc = cfs_expr_list_values(el, ncpts, &cpts);
724 cfs_expr_list_free(el);
726 CERROR("%s: failed to parse CPT array %s: %d\n",
727 conf->psc_name, cconf->cc_pattern, rc);
728 RETURN(ERR_PTR(rc < 0 ? rc : -EINVAL));
734 OBD_ALLOC(service, offsetof(struct ptlrpc_service, srv_parts[ncpts]));
735 if (service == NULL) {
737 OBD_FREE(cpts, sizeof(*cpts) * ncpts);
738 RETURN(ERR_PTR(-ENOMEM));
741 service->srv_cptable = cptable;
742 service->srv_cpts = cpts;
743 service->srv_ncpts = ncpts;
745 service->srv_cpt_bits = 0; /* it's zero already, easy to read... */
746 while ((1 << service->srv_cpt_bits) < cfs_cpt_number(cptable))
747 service->srv_cpt_bits++;
750 cfs_spin_lock_init(&service->srv_lock);
751 service->srv_name = conf->psc_name;
752 service->srv_watchdog_factor = conf->psc_watchdog_factor;
753 CFS_INIT_LIST_HEAD(&service->srv_list); /* for safty of cleanup */
755 /* buffer configuration */
756 service->srv_nbuf_per_group = test_req_buffer_pressure ? 1 :
757 max(conf->psc_buf.bc_nbufs /
758 service->srv_ncpts, 1U);
759 service->srv_max_req_size = conf->psc_buf.bc_req_max_size +
761 service->srv_buf_size = conf->psc_buf.bc_buf_size;
762 service->srv_rep_portal = conf->psc_buf.bc_rep_portal;
763 service->srv_req_portal = conf->psc_buf.bc_req_portal;
765 /* Increase max reply size to next power of two */
766 service->srv_max_reply_size = 1;
767 while (service->srv_max_reply_size <
768 conf->psc_buf.bc_rep_max_size + SPTLRPC_MAX_PAYLOAD)
769 service->srv_max_reply_size <<= 1;
771 service->srv_thread_name = conf->psc_thr.tc_thr_name;
772 service->srv_ctx_tags = conf->psc_thr.tc_ctx_tags;
773 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
774 service->srv_ops = conf->psc_ops;
776 for (i = 0; i < ncpts; i++) {
777 if (!conf->psc_thr.tc_cpu_affinity)
780 cpt = cpts != NULL ? cpts[i] : i;
782 OBD_CPT_ALLOC(svcpt, cptable, cpt, sizeof(*svcpt));
784 GOTO(failed, rc = -ENOMEM);
786 service->srv_parts[i] = svcpt;
787 rc = ptlrpc_service_part_init(service, svcpt, cpt);
792 ptlrpc_server_nthreads_check(service, conf);
794 rc = LNetSetLazyPortal(service->srv_req_portal);
797 cfs_spin_lock (&ptlrpc_all_services_lock);
798 cfs_list_add (&service->srv_list, &ptlrpc_all_services);
799 cfs_spin_unlock (&ptlrpc_all_services_lock);
801 if (proc_entry != NULL)
802 ptlrpc_lprocfs_register_service(proc_entry, service);
804 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
805 service->srv_name, service->srv_req_portal);
808 rc = ptlrpc_start_threads(service);
810 CERROR("Failed to start threads for service %s: %d\n",
811 service->srv_name, rc);
818 ptlrpc_unregister_service(service);
823 * to actually free the request, must be called without holding svc_lock.
824 * note it's caller's responsibility to unlink req->rq_list.
826 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
828 LASSERT(cfs_atomic_read(&req->rq_refcount) == 0);
829 LASSERT(cfs_list_empty(&req->rq_timed_list));
831 /* DEBUG_REQ() assumes the reply state of a request with a valid
832 * ref will not be destroyed until that reference is dropped. */
833 ptlrpc_req_drop_rs(req);
835 sptlrpc_svc_ctx_decref(req);
837 if (req != &req->rq_rqbd->rqbd_req) {
838 /* NB request buffers use an embedded
839 * req if the incoming req unlinked the
840 * MD; this isn't one of them! */
841 OBD_FREE(req, sizeof(*req));
846 * drop a reference count of the request. if it reaches 0, we either
847 * put it into history list, or free it immediately.
849 void ptlrpc_server_drop_request(struct ptlrpc_request *req)
851 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
852 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
853 struct ptlrpc_service *svc = svcpt->scp_service;
858 if (!cfs_atomic_dec_and_test(&req->rq_refcount))
861 if (req->rq_at_linked) {
862 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
863 __u32 index = req->rq_at_index;
865 cfs_spin_lock(&svcpt->scp_at_lock);
867 LASSERT(!cfs_list_empty(&req->rq_timed_list));
868 cfs_list_del_init(&req->rq_timed_list);
869 cfs_spin_lock(&req->rq_lock);
870 req->rq_at_linked = 0;
871 cfs_spin_unlock(&req->rq_lock);
872 array->paa_reqs_count[index]--;
875 cfs_spin_unlock(&svcpt->scp_at_lock);
877 LASSERT(cfs_list_empty(&req->rq_timed_list));
880 /* finalize request */
881 if (req->rq_export) {
882 class_export_put(req->rq_export);
883 req->rq_export = NULL;
886 cfs_spin_lock(&svcpt->scp_lock);
888 cfs_list_add(&req->rq_list, &rqbd->rqbd_reqs);
890 refcount = --(rqbd->rqbd_refcount);
892 /* request buffer is now idle: add to history */
893 cfs_list_del(&rqbd->rqbd_list);
895 cfs_list_add_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds);
896 svcpt->scp_hist_nrqbds++;
898 /* cull some history?
899 * I expect only about 1 or 2 rqbds need to be recycled here */
900 while (svcpt->scp_hist_nrqbds > svc->srv_hist_nrqbds_cpt_max) {
901 rqbd = cfs_list_entry(svcpt->scp_hist_rqbds.next,
902 struct ptlrpc_request_buffer_desc,
905 cfs_list_del(&rqbd->rqbd_list);
906 svcpt->scp_hist_nrqbds--;
908 /* remove rqbd's reqs from svc's req history while
909 * I've got the service lock */
910 cfs_list_for_each(tmp, &rqbd->rqbd_reqs) {
911 req = cfs_list_entry(tmp, struct ptlrpc_request,
913 /* Track the highest culled req seq */
914 if (req->rq_history_seq >
915 svcpt->scp_hist_seq_culled) {
916 svcpt->scp_hist_seq_culled =
919 cfs_list_del(&req->rq_history_list);
922 cfs_spin_unlock(&svcpt->scp_lock);
924 cfs_list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
925 req = cfs_list_entry(rqbd->rqbd_reqs.next,
926 struct ptlrpc_request,
928 cfs_list_del(&req->rq_list);
929 ptlrpc_server_free_request(req);
932 cfs_spin_lock(&svcpt->scp_lock);
934 * now all reqs including the embedded req has been
935 * disposed, schedule request buffer for re-use.
937 LASSERT(cfs_atomic_read(&rqbd->rqbd_req.rq_refcount) ==
939 cfs_list_add_tail(&rqbd->rqbd_list,
940 &svcpt->scp_rqbd_idle);
943 cfs_spin_unlock(&svcpt->scp_lock);
944 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
945 /* If we are low on memory, we are not interested in history */
946 cfs_list_del(&req->rq_list);
947 cfs_list_del_init(&req->rq_history_list);
949 cfs_spin_unlock(&svcpt->scp_lock);
951 ptlrpc_server_free_request(req);
953 cfs_spin_unlock(&svcpt->scp_lock);
958 * to finish a request: stop sending more early replies, and release
959 * the request. should be called after we finished handling the request.
961 static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt,
962 struct ptlrpc_request *req)
964 ptlrpc_hpreq_fini(req);
966 cfs_spin_lock(&svcpt->scp_req_lock);
967 svcpt->scp_nreqs_active--;
969 svcpt->scp_nhreqs_active--;
970 cfs_spin_unlock(&svcpt->scp_req_lock);
972 ptlrpc_server_drop_request(req);
976 * This function makes sure dead exports are evicted in a timely manner.
977 * This function is only called when some export receives a message (i.e.,
978 * the network is up.)
980 static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
982 struct obd_export *oldest_exp;
983 time_t oldest_time, new_time;
989 /* Compensate for slow machines, etc, by faking our request time
990 into the future. Although this can break the strict time-ordering
991 of the list, we can be really lazy here - we don't have to evict
992 at the exact right moment. Eventually, all silent exports
993 will make it to the top of the list. */
995 /* Do not pay attention on 1sec or smaller renewals. */
996 new_time = cfs_time_current_sec() + extra_delay;
997 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
1000 exp->exp_last_request_time = new_time;
1001 CDEBUG(D_HA, "updating export %s at "CFS_TIME_T" exp %p\n",
1002 exp->exp_client_uuid.uuid,
1003 exp->exp_last_request_time, exp);
1005 /* exports may get disconnected from the chain even though the
1006 export has references, so we must keep the spin lock while
1007 manipulating the lists */
1008 cfs_spin_lock(&exp->exp_obd->obd_dev_lock);
1010 if (cfs_list_empty(&exp->exp_obd_chain_timed)) {
1011 /* this one is not timed */
1012 cfs_spin_unlock(&exp->exp_obd->obd_dev_lock);
1016 cfs_list_move_tail(&exp->exp_obd_chain_timed,
1017 &exp->exp_obd->obd_exports_timed);
1019 oldest_exp = cfs_list_entry(exp->exp_obd->obd_exports_timed.next,
1020 struct obd_export, exp_obd_chain_timed);
1021 oldest_time = oldest_exp->exp_last_request_time;
1022 cfs_spin_unlock(&exp->exp_obd->obd_dev_lock);
1024 if (exp->exp_obd->obd_recovering) {
1025 /* be nice to everyone during recovery */
1030 /* Note - racing to start/reset the obd_eviction timer is safe */
1031 if (exp->exp_obd->obd_eviction_timer == 0) {
1032 /* Check if the oldest entry is expired. */
1033 if (cfs_time_current_sec() > (oldest_time + PING_EVICT_TIMEOUT +
1035 /* We need a second timer, in case the net was down and
1036 * it just came back. Since the pinger may skip every
1037 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
1038 * we better wait for 3. */
1039 exp->exp_obd->obd_eviction_timer =
1040 cfs_time_current_sec() + 3 * PING_INTERVAL;
1041 CDEBUG(D_HA, "%s: Think about evicting %s from "CFS_TIME_T"\n",
1042 exp->exp_obd->obd_name,
1043 obd_export_nid2str(oldest_exp), oldest_time);
1046 if (cfs_time_current_sec() >
1047 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
1048 /* The evictor won't evict anyone who we've heard from
1049 * recently, so we don't have to check before we start
1051 if (!ping_evictor_wake(exp))
1052 exp->exp_obd->obd_eviction_timer = 0;
1060 * Sanity check request \a req.
1061 * Return 0 if all is ok, error code otherwise.
1063 static int ptlrpc_check_req(struct ptlrpc_request *req)
1067 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
1068 req->rq_export->exp_conn_cnt)) {
1069 DEBUG_REQ(D_ERROR, req,
1070 "DROPPING req from old connection %d < %d",
1071 lustre_msg_get_conn_cnt(req->rq_reqmsg),
1072 req->rq_export->exp_conn_cnt);
1075 if (unlikely(req->rq_export->exp_obd &&
1076 req->rq_export->exp_obd->obd_fail)) {
1077 /* Failing over, don't handle any more reqs, send
1078 error response instead. */
1079 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
1080 req, req->rq_export->exp_obd->obd_name);
1082 } else if (lustre_msg_get_flags(req->rq_reqmsg) &
1083 (MSG_REPLAY | MSG_REQ_REPLAY_DONE) &&
1084 !(req->rq_export->exp_obd->obd_recovering)) {
1085 DEBUG_REQ(D_ERROR, req,
1086 "Invalid replay without recovery");
1087 class_fail_export(req->rq_export);
1089 } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 &&
1090 !(req->rq_export->exp_obd->obd_recovering)) {
1091 DEBUG_REQ(D_ERROR, req, "Invalid req with transno "
1092 LPU64" without recovery",
1093 lustre_msg_get_transno(req->rq_reqmsg));
1094 class_fail_export(req->rq_export);
1098 if (unlikely(rc < 0)) {
1099 req->rq_status = rc;
1105 static void ptlrpc_at_set_timer(struct ptlrpc_service_part *svcpt)
1107 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1110 if (array->paa_count == 0) {
1111 cfs_timer_disarm(&svcpt->scp_at_timer);
1115 /* Set timer for closest deadline */
1116 next = (__s32)(array->paa_deadline - cfs_time_current_sec() -
1119 ptlrpc_at_timer((unsigned long)svcpt);
1121 cfs_timer_arm(&svcpt->scp_at_timer, cfs_time_shift(next));
1122 CDEBUG(D_INFO, "armed %s at %+ds\n",
1123 svcpt->scp_service->srv_name, next);
1127 /* Add rpc to early reply check list */
1128 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
1130 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1131 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1132 struct ptlrpc_request *rq = NULL;
1138 if (req->rq_no_reply)
1141 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
1144 cfs_spin_lock(&svcpt->scp_at_lock);
1145 LASSERT(cfs_list_empty(&req->rq_timed_list));
1147 index = (unsigned long)req->rq_deadline % array->paa_size;
1148 if (array->paa_reqs_count[index] > 0) {
1149 /* latest rpcs will have the latest deadlines in the list,
1150 * so search backward. */
1151 cfs_list_for_each_entry_reverse(rq,
1152 &array->paa_reqs_array[index],
1154 if (req->rq_deadline >= rq->rq_deadline) {
1155 cfs_list_add(&req->rq_timed_list,
1156 &rq->rq_timed_list);
1162 /* Add the request at the head of the list */
1163 if (cfs_list_empty(&req->rq_timed_list))
1164 cfs_list_add(&req->rq_timed_list,
1165 &array->paa_reqs_array[index]);
1167 cfs_spin_lock(&req->rq_lock);
1168 req->rq_at_linked = 1;
1169 cfs_spin_unlock(&req->rq_lock);
1170 req->rq_at_index = index;
1171 array->paa_reqs_count[index]++;
1173 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
1174 array->paa_deadline = req->rq_deadline;
1175 ptlrpc_at_set_timer(svcpt);
1177 cfs_spin_unlock(&svcpt->scp_at_lock);
1182 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
1184 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1185 struct ptlrpc_request *reqcopy;
1186 struct lustre_msg *reqmsg;
1187 cfs_duration_t olddl = req->rq_deadline - cfs_time_current_sec();
1192 /* deadline is when the client expects us to reply, margin is the
1193 difference between clients' and servers' expectations */
1194 DEBUG_REQ(D_ADAPTTO, req,
1195 "%ssending early reply (deadline %+lds, margin %+lds) for "
1196 "%d+%d", AT_OFF ? "AT off - not " : "",
1197 olddl, olddl - at_get(&svcpt->scp_at_estimate),
1198 at_get(&svcpt->scp_at_estimate), at_extra);
1204 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), "
1205 "not sending early reply. Consider increasing "
1206 "at_early_margin (%d)?", olddl, at_early_margin);
1208 /* Return an error so we're not re-added to the timed list. */
1212 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
1213 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
1214 "but no AT support");
1218 if (req->rq_export &&
1219 lustre_msg_get_flags(req->rq_reqmsg) &
1220 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1221 /* During recovery, we don't want to send too many early
1222 * replies, but on the other hand we want to make sure the
1223 * client has enough time to resend if the rpc is lost. So
1224 * during the recovery period send at least 4 early replies,
1225 * spacing them every at_extra if we can. at_estimate should
1226 * always equal this fixed value during recovery. */
1227 at_measured(&svcpt->scp_at_estimate, min(at_extra,
1228 req->rq_export->exp_obd->obd_recovery_timeout / 4));
1230 /* Fake our processing time into the future to ask the clients
1231 * for some extra amount of time */
1232 at_measured(&svcpt->scp_at_estimate, at_extra +
1233 cfs_time_current_sec() -
1234 req->rq_arrival_time.tv_sec);
1236 /* Check to see if we've actually increased the deadline -
1237 * we may be past adaptive_max */
1238 if (req->rq_deadline >= req->rq_arrival_time.tv_sec +
1239 at_get(&svcpt->scp_at_estimate)) {
1240 DEBUG_REQ(D_WARNING, req, "Couldn't add any time "
1241 "(%ld/%ld), not sending early reply\n",
1242 olddl, req->rq_arrival_time.tv_sec +
1243 at_get(&svcpt->scp_at_estimate) -
1244 cfs_time_current_sec());
1248 newdl = cfs_time_current_sec() + at_get(&svcpt->scp_at_estimate);
1250 OBD_ALLOC(reqcopy, sizeof *reqcopy);
1251 if (reqcopy == NULL)
1253 OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1255 OBD_FREE(reqcopy, sizeof *reqcopy);
1260 reqcopy->rq_reply_state = NULL;
1261 reqcopy->rq_rep_swab_mask = 0;
1262 reqcopy->rq_pack_bulk = 0;
1263 reqcopy->rq_pack_udesc = 0;
1264 reqcopy->rq_packed_final = 0;
1265 sptlrpc_svc_ctx_addref(reqcopy);
1266 /* We only need the reqmsg for the magic */
1267 reqcopy->rq_reqmsg = reqmsg;
1268 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1270 LASSERT(cfs_atomic_read(&req->rq_refcount));
1271 /** if it is last refcount then early reply isn't needed */
1272 if (cfs_atomic_read(&req->rq_refcount) == 1) {
1273 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
1274 "abort sending early reply\n");
1275 GOTO(out, rc = -EINVAL);
1278 /* Connection ref */
1279 reqcopy->rq_export = class_conn2export(
1280 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1281 if (reqcopy->rq_export == NULL)
1282 GOTO(out, rc = -ENODEV);
1285 class_export_rpc_get(reqcopy->rq_export);
1286 if (reqcopy->rq_export->exp_obd &&
1287 reqcopy->rq_export->exp_obd->obd_fail)
1288 GOTO(out_put, rc = -ENODEV);
1290 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1294 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1297 /* Adjust our own deadline to what we told the client */
1298 req->rq_deadline = newdl;
1299 req->rq_early_count++; /* number sent, server side */
1301 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1304 /* Free the (early) reply state from lustre_pack_reply.
1305 (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
1306 ptlrpc_req_drop_rs(reqcopy);
1309 class_export_rpc_put(reqcopy->rq_export);
1310 class_export_put(reqcopy->rq_export);
1312 sptlrpc_svc_ctx_decref(reqcopy);
1313 OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1314 OBD_FREE(reqcopy, sizeof *reqcopy);
1318 /* Send early replies to everybody expiring within at_early_margin
1319 asking for at_extra time */
1320 static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1322 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1323 struct ptlrpc_request *rq, *n;
1324 cfs_list_t work_list;
1327 time_t now = cfs_time_current_sec();
1328 cfs_duration_t delay;
1329 int first, counter = 0;
1332 cfs_spin_lock(&svcpt->scp_at_lock);
1333 if (svcpt->scp_at_check == 0) {
1334 cfs_spin_unlock(&svcpt->scp_at_lock);
1337 delay = cfs_time_sub(cfs_time_current(), svcpt->scp_at_checktime);
1338 svcpt->scp_at_check = 0;
1340 if (array->paa_count == 0) {
1341 cfs_spin_unlock(&svcpt->scp_at_lock);
1345 /* The timer went off, but maybe the nearest rpc already completed. */
1346 first = array->paa_deadline - now;
1347 if (first > at_early_margin) {
1348 /* We've still got plenty of time. Reset the timer. */
1349 ptlrpc_at_set_timer(svcpt);
1350 cfs_spin_unlock(&svcpt->scp_at_lock);
1354 /* We're close to a timeout, and we don't know how much longer the
1355 server will take. Send early replies to everyone expiring soon. */
1356 CFS_INIT_LIST_HEAD(&work_list);
1358 index = (unsigned long)array->paa_deadline % array->paa_size;
1359 count = array->paa_count;
1361 count -= array->paa_reqs_count[index];
1362 cfs_list_for_each_entry_safe(rq, n,
1363 &array->paa_reqs_array[index],
1365 if (rq->rq_deadline <= now + at_early_margin) {
1366 cfs_list_del_init(&rq->rq_timed_list);
1368 * ptlrpc_server_drop_request() may drop
1369 * refcount to 0 already. Let's check this and
1370 * don't add entry to work_list
1372 if (likely(cfs_atomic_inc_not_zero(&rq->rq_refcount)))
1373 cfs_list_add(&rq->rq_timed_list, &work_list);
1375 array->paa_reqs_count[index]--;
1377 cfs_spin_lock(&rq->rq_lock);
1378 rq->rq_at_linked = 0;
1379 cfs_spin_unlock(&rq->rq_lock);
1383 /* update the earliest deadline */
1384 if (deadline == -1 || rq->rq_deadline < deadline)
1385 deadline = rq->rq_deadline;
1390 if (++index >= array->paa_size)
1393 array->paa_deadline = deadline;
1394 /* we have a new earliest deadline, restart the timer */
1395 ptlrpc_at_set_timer(svcpt);
1397 cfs_spin_unlock(&svcpt->scp_at_lock);
1399 CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
1400 "replies\n", first, at_extra, counter);
1402 /* We're already past request deadlines before we even get a
1403 chance to send early replies */
1404 LCONSOLE_WARN("%s: This server is not able to keep up with "
1405 "request traffic (cpu-bound).\n",
1406 svcpt->scp_service->srv_name);
1407 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, "
1408 "delay="CFS_DURATION_T"(jiff)\n",
1409 counter, svcpt->scp_nreqs_incoming,
1410 svcpt->scp_nreqs_active,
1411 at_get(&svcpt->scp_at_estimate), delay);
1414 /* we took additional refcount so entries can't be deleted from list, no
1415 * locking is needed */
1416 while (!cfs_list_empty(&work_list)) {
1417 rq = cfs_list_entry(work_list.next, struct ptlrpc_request,
1419 cfs_list_del_init(&rq->rq_timed_list);
1421 if (ptlrpc_at_send_early_reply(rq) == 0)
1422 ptlrpc_at_add_timed(rq);
1424 ptlrpc_server_drop_request(rq);
1427 RETURN(1); /* return "did_something" for liblustre */
1431 * Put the request to the export list if the request may become
1432 * a high priority one.
1434 static int ptlrpc_hpreq_init(struct ptlrpc_service *svc,
1435 struct ptlrpc_request *req)
1440 if (svc->srv_ops.so_hpreq_handler) {
1441 rc = svc->srv_ops.so_hpreq_handler(req);
1445 if (req->rq_export && req->rq_ops) {
1446 /* Perform request specific check. We should do this check
1447 * before the request is added into exp_hp_rpcs list otherwise
1448 * it may hit swab race at LU-1044. */
1449 if (req->rq_ops->hpreq_check)
1450 rc = req->rq_ops->hpreq_check(req);
1452 cfs_spin_lock_bh(&req->rq_export->exp_rpc_lock);
1453 cfs_list_add(&req->rq_exp_list,
1454 &req->rq_export->exp_hp_rpcs);
1455 cfs_spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1461 /** Remove the request from the export list. */
1462 static void ptlrpc_hpreq_fini(struct ptlrpc_request *req)
1465 if (req->rq_export && req->rq_ops) {
1466 /* refresh lock timeout again so that client has more
1467 * room to send lock cancel RPC. */
1468 if (req->rq_ops->hpreq_fini)
1469 req->rq_ops->hpreq_fini(req);
1471 cfs_spin_lock_bh(&req->rq_export->exp_rpc_lock);
1472 cfs_list_del_init(&req->rq_exp_list);
1473 cfs_spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1479 * Make the request a high priority one.
1481 * All the high priority requests are queued in a separate FIFO
1482 * ptlrpc_service_part::scp_hpreq_pending list which is parallel to
1483 * ptlrpc_service_part::scp_req_pending list but has a higher priority
1486 * \see ptlrpc_server_handle_request().
1488 static void ptlrpc_hpreq_reorder_nolock(struct ptlrpc_service_part *svcpt,
1489 struct ptlrpc_request *req)
1493 cfs_spin_lock(&req->rq_lock);
1494 if (req->rq_hp == 0) {
1495 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1497 /* Add to the high priority queue. */
1498 cfs_list_move_tail(&req->rq_list, &svcpt->scp_hreq_pending);
1500 if (opc != OBD_PING)
1501 DEBUG_REQ(D_RPCTRACE, req, "high priority req");
1503 cfs_spin_unlock(&req->rq_lock);
1508 * \see ptlrpc_hpreq_reorder_nolock
1510 void ptlrpc_hpreq_reorder(struct ptlrpc_request *req)
1512 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1515 cfs_spin_lock(&svcpt->scp_req_lock);
1516 /* It may happen that the request is already taken for the processing
1517 * but still in the export list, or the request is not in the request
1518 * queue but in the export list already, do not add it into the
1520 if (!cfs_list_empty(&req->rq_list))
1521 ptlrpc_hpreq_reorder_nolock(svcpt, req);
1522 cfs_spin_unlock(&svcpt->scp_req_lock);
1526 /** Check if the request is a high priority one. */
1527 static int ptlrpc_server_hpreq_check(struct ptlrpc_service *svc,
1528 struct ptlrpc_request *req)
1532 /* Check by request opc. */
1533 if (OBD_PING == lustre_msg_get_opc(req->rq_reqmsg))
1536 RETURN(ptlrpc_hpreq_init(svc, req));
1539 /** Check if a request is a high priority one. */
1540 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1541 struct ptlrpc_request *req)
1546 rc = ptlrpc_server_hpreq_check(svcpt->scp_service, req);
1550 cfs_spin_lock(&svcpt->scp_req_lock);
1553 ptlrpc_hpreq_reorder_nolock(svcpt, req);
1555 cfs_list_add_tail(&req->rq_list, &svcpt->scp_req_pending);
1557 cfs_spin_unlock(&svcpt->scp_req_lock);
1563 * Allow to handle high priority request
1564 * User can call it w/o any lock but need to hold
1565 * ptlrpc_service_part::scp_req_lock to get reliable result
1567 static int ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1573 if (svcpt->scp_nreqs_active >= svcpt->scp_nthrs_running - 1)
1576 return cfs_list_empty(&svcpt->scp_req_pending) ||
1577 svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1580 static int ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1583 return ptlrpc_server_allow_high(svcpt, force) &&
1584 !cfs_list_empty(&svcpt->scp_hreq_pending);
1588 * Only allow normal priority requests on a service that has a high-priority
1589 * queue if forced (i.e. cleanup), if there are other high priority requests
1590 * already being processed (i.e. those threads can service more high-priority
1591 * requests), or if there are enough idle threads that a later thread can do
1592 * a high priority request.
1593 * User can call it w/o any lock but need to hold
1594 * ptlrpc_service_part::scp_req_lock to get reliable result
1596 static int ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1600 if (1) /* always allow to handle normal request for liblustre */
1604 svcpt->scp_nreqs_active < svcpt->scp_nthrs_running - 2)
1607 if (svcpt->scp_nreqs_active >= svcpt->scp_nthrs_running - 1)
1610 return svcpt->scp_nhreqs_active > 0 ||
1611 svcpt->scp_service->srv_ops.so_hpreq_handler == NULL;
1614 static int ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1617 return ptlrpc_server_allow_normal(svcpt, force) &&
1618 !cfs_list_empty(&svcpt->scp_req_pending);
1622 * Returns true if there are requests available in incoming
1623 * request queue for processing and it is allowed to fetch them.
1624 * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1625 * to get reliable result
1626 * \see ptlrpc_server_allow_normal
1627 * \see ptlrpc_server_allow high
1630 ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt, int force)
1632 return ptlrpc_server_high_pending(svcpt, force) ||
1633 ptlrpc_server_normal_pending(svcpt, force);
1637 * Fetch a request for processing from queue of unprocessed requests.
1638 * Favors high-priority requests.
1639 * Returns a pointer to fetched request.
1641 static struct ptlrpc_request *
1642 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, int force)
1644 struct ptlrpc_request *req;
1647 if (ptlrpc_server_high_pending(svcpt, force)) {
1648 req = cfs_list_entry(svcpt->scp_hreq_pending.next,
1649 struct ptlrpc_request, rq_list);
1650 svcpt->scp_hreq_count++;
1654 if (ptlrpc_server_normal_pending(svcpt, force)) {
1655 req = cfs_list_entry(svcpt->scp_req_pending.next,
1656 struct ptlrpc_request, rq_list);
1657 svcpt->scp_hreq_count = 0;
1664 * Handle freshly incoming reqs, add to timed early reply list,
1665 * pass on to regular request queue.
1666 * All incoming requests pass through here before getting into
1667 * ptlrpc_server_handle_req later on.
1670 ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt)
1672 struct ptlrpc_service *svc = svcpt->scp_service;
1673 struct ptlrpc_request *req;
1678 cfs_spin_lock(&svcpt->scp_lock);
1679 if (cfs_list_empty(&svcpt->scp_req_incoming)) {
1680 cfs_spin_unlock(&svcpt->scp_lock);
1684 req = cfs_list_entry(svcpt->scp_req_incoming.next,
1685 struct ptlrpc_request, rq_list);
1686 cfs_list_del_init(&req->rq_list);
1687 svcpt->scp_nreqs_incoming--;
1688 /* Consider this still a "queued" request as far as stats are
1690 cfs_spin_unlock(&svcpt->scp_lock);
1692 /* go through security check/transform */
1693 rc = sptlrpc_svc_unwrap_request(req);
1697 case SECSVC_COMPLETE:
1698 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1707 * for null-flavored rpc, msg has been unpacked by sptlrpc, although
1708 * redo it wouldn't be harmful.
1710 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
1711 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
1713 CERROR("error unpacking request: ptl %d from %s "
1714 "x"LPU64"\n", svc->srv_req_portal,
1715 libcfs_id2str(req->rq_peer), req->rq_xid);
1720 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1722 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
1723 LPU64"\n", svc->srv_req_portal,
1724 libcfs_id2str(req->rq_peer), req->rq_xid);
1728 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
1729 lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
1730 CERROR("drop incoming rpc opc %u, x"LPU64"\n",
1731 cfs_fail_val, req->rq_xid);
1736 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1737 CERROR("wrong packet type received (type=%u) from %s\n",
1738 lustre_msg_get_type(req->rq_reqmsg),
1739 libcfs_id2str(req->rq_peer));
1743 switch(lustre_msg_get_opc(req->rq_reqmsg)) {
1746 req->rq_bulk_write = 1;
1750 case MGS_CONFIG_READ:
1751 req->rq_bulk_read = 1;
1755 CDEBUG(D_RPCTRACE, "got req x"LPU64"\n", req->rq_xid);
1757 req->rq_export = class_conn2export(
1758 lustre_msg_get_handle(req->rq_reqmsg));
1759 if (req->rq_export) {
1760 rc = ptlrpc_check_req(req);
1762 rc = sptlrpc_target_export_check(req->rq_export, req);
1764 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
1765 "illegal security flavor,");
1770 ptlrpc_update_export_timer(req->rq_export, 0);
1773 /* req_in handling should/must be fast */
1774 if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5)
1775 DEBUG_REQ(D_WARNING, req, "Slow req_in handling "CFS_DURATION_T"s",
1776 cfs_time_sub(cfs_time_current_sec(),
1777 req->rq_arrival_time.tv_sec));
1779 /* Set rpc server deadline and add it to the timed list */
1780 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1781 MSGHDR_AT_SUPPORT) ?
1782 /* The max time the client expects us to take */
1783 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1784 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1785 if (unlikely(deadline == 0)) {
1786 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1790 ptlrpc_at_add_timed(req);
1792 /* Move it over to the request processing queue */
1793 rc = ptlrpc_server_request_add(svcpt, req);
1795 ptlrpc_hpreq_fini(req);
1798 cfs_waitq_signal(&svcpt->scp_waitq);
1802 cfs_spin_lock(&svcpt->scp_req_lock);
1803 svcpt->scp_nreqs_active++;
1804 cfs_spin_unlock(&svcpt->scp_req_lock);
1805 ptlrpc_server_finish_request(svcpt, req);
1811 * Main incoming request handling logic.
1812 * Calls handler function from service to do actual processing.
1815 ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
1816 struct ptlrpc_thread *thread)
1818 struct ptlrpc_service *svc = svcpt->scp_service;
1819 struct obd_export *export = NULL;
1820 struct ptlrpc_request *request;
1821 struct timeval work_start;
1822 struct timeval work_end;
1828 cfs_spin_lock(&svcpt->scp_req_lock);
1830 /* !@%$# liblustre only has 1 thread */
1831 if (cfs_atomic_read(&svcpt->scp_nreps_difficult) != 0) {
1832 cfs_spin_unlock(&svcpt->scp_req_lock);
1836 request = ptlrpc_server_request_get(svcpt, 0);
1837 if (request == NULL) {
1838 cfs_spin_unlock(&svcpt->scp_req_lock);
1842 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
1843 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
1844 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
1845 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
1847 if (unlikely(fail_opc)) {
1848 if (request->rq_export && request->rq_ops) {
1849 cfs_spin_unlock(&svcpt->scp_req_lock);
1851 OBD_FAIL_TIMEOUT(fail_opc, 4);
1853 cfs_spin_lock(&svcpt->scp_req_lock);
1854 request = ptlrpc_server_request_get(svcpt, 0);
1855 if (request == NULL) {
1856 cfs_spin_unlock(&svcpt->scp_req_lock);
1862 cfs_list_del_init(&request->rq_list);
1863 svcpt->scp_nreqs_active++;
1865 svcpt->scp_nhreqs_active++;
1867 cfs_spin_unlock(&svcpt->scp_req_lock);
1869 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
1871 if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
1872 libcfs_debug_dumplog();
1874 cfs_gettimeofday(&work_start);
1875 timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
1876 if (likely(svc->srv_stats != NULL)) {
1877 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
1879 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
1880 svcpt->scp_nreqs_incoming);
1881 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
1882 svcpt->scp_nreqs_active);
1883 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
1884 at_get(&svcpt->scp_at_estimate));
1887 rc = lu_context_init(&request->rq_session, LCT_SESSION | LCT_NOREF);
1889 CERROR("Failure to initialize session: %d\n", rc);
1892 request->rq_session.lc_thread = thread;
1893 request->rq_session.lc_cookie = 0x5;
1894 lu_context_enter(&request->rq_session);
1896 CDEBUG(D_NET, "got req "LPU64"\n", request->rq_xid);
1898 request->rq_svc_thread = thread;
1900 request->rq_svc_thread->t_env->le_ses = &request->rq_session;
1902 if (likely(request->rq_export)) {
1903 if (unlikely(ptlrpc_check_req(request)))
1905 ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
1906 export = class_export_rpc_get(request->rq_export);
1909 /* Discard requests queued for longer than the deadline.
1910 The deadline is increased if we send an early reply. */
1911 if (cfs_time_current_sec() > request->rq_deadline) {
1912 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s"
1913 ": deadline "CFS_DURATION_T":"CFS_DURATION_T"s ago\n",
1914 libcfs_id2str(request->rq_peer),
1915 cfs_time_sub(request->rq_deadline,
1916 request->rq_arrival_time.tv_sec),
1917 cfs_time_sub(cfs_time_current_sec(),
1918 request->rq_deadline));
1919 goto put_rpc_export;
1922 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
1923 "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1924 (request->rq_export ?
1925 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1926 (request->rq_export ?
1927 cfs_atomic_read(&request->rq_export->exp_refcount) : -99),
1928 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1929 libcfs_id2str(request->rq_peer),
1930 lustre_msg_get_opc(request->rq_reqmsg));
1932 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
1933 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
1935 rc = svc->srv_ops.so_req_handler(request);
1937 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
1941 class_export_rpc_put(export);
1943 lu_context_exit(&request->rq_session);
1944 lu_context_fini(&request->rq_session);
1946 if (unlikely(cfs_time_current_sec() > request->rq_deadline)) {
1947 DEBUG_REQ(D_WARNING, request, "Request x"LPU64" took longer "
1948 "than estimated ("CFS_DURATION_T":"CFS_DURATION_T"s);"
1949 " client may timeout.",
1950 request->rq_xid, cfs_time_sub(request->rq_deadline,
1951 request->rq_arrival_time.tv_sec),
1952 cfs_time_sub(cfs_time_current_sec(),
1953 request->rq_deadline));
1956 cfs_gettimeofday(&work_end);
1957 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1958 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
1959 "%s:%s+%d:%d:x"LPU64":%s:%d Request procesed in "
1960 "%ldus (%ldus total) trans "LPU64" rc %d/%d\n",
1962 (request->rq_export ?
1963 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1964 (request->rq_export ?
1965 cfs_atomic_read(&request->rq_export->exp_refcount) : -99),
1966 lustre_msg_get_status(request->rq_reqmsg),
1968 libcfs_id2str(request->rq_peer),
1969 lustre_msg_get_opc(request->rq_reqmsg),
1971 cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL),
1972 (request->rq_repmsg ?
1973 lustre_msg_get_transno(request->rq_repmsg) :
1974 request->rq_transno),
1976 (request->rq_repmsg ?
1977 lustre_msg_get_status(request->rq_repmsg) : -999));
1978 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
1979 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
1980 int opc = opcode_offset(op);
1981 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
1982 LASSERT(opc < LUSTRE_MAX_OPCODES);
1983 lprocfs_counter_add(svc->srv_stats,
1984 opc + EXTRA_MAX_OPCODES,
1988 if (unlikely(request->rq_early_count)) {
1989 DEBUG_REQ(D_ADAPTTO, request,
1990 "sent %d early replies before finishing in "
1992 request->rq_early_count,
1993 cfs_time_sub(work_end.tv_sec,
1994 request->rq_arrival_time.tv_sec));
1998 ptlrpc_server_finish_request(svcpt, request);
2004 * An internal function to process a single reply state object.
2007 ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
2009 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
2010 struct ptlrpc_service *svc = svcpt->scp_service;
2011 struct obd_export *exp;
2016 exp = rs->rs_export;
2018 LASSERT (rs->rs_difficult);
2019 LASSERT (rs->rs_scheduled);
2020 LASSERT (cfs_list_empty(&rs->rs_list));
2022 cfs_spin_lock (&exp->exp_lock);
2023 /* Noop if removed already */
2024 cfs_list_del_init (&rs->rs_exp_list);
2025 cfs_spin_unlock (&exp->exp_lock);
2027 /* The disk commit callback holds exp_uncommitted_replies_lock while it
2028 * iterates over newly committed replies, removing them from
2029 * exp_uncommitted_replies. It then drops this lock and schedules the
2030 * replies it found for handling here.
2032 * We can avoid contention for exp_uncommitted_replies_lock between the
2033 * HRT threads and further commit callbacks by checking rs_committed
2034 * which is set in the commit callback while it holds both
2035 * rs_lock and exp_uncommitted_reples.
2037 * If we see rs_committed clear, the commit callback _may_ not have
2038 * handled this reply yet and we race with it to grab
2039 * exp_uncommitted_replies_lock before removing the reply from
2040 * exp_uncommitted_replies. Note that if we lose the race and the
2041 * reply has already been removed, list_del_init() is a noop.
2043 * If we see rs_committed set, we know the commit callback is handling,
2044 * or has handled this reply since store reordering might allow us to
2045 * see rs_committed set out of sequence. But since this is done
2046 * holding rs_lock, we can be sure it has all completed once we hold
2047 * rs_lock, which we do right next.
2049 if (!rs->rs_committed) {
2050 cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
2051 cfs_list_del_init(&rs->rs_obd_list);
2052 cfs_spin_unlock(&exp->exp_uncommitted_replies_lock);
2055 cfs_spin_lock(&rs->rs_lock);
2057 been_handled = rs->rs_handled;
2060 nlocks = rs->rs_nlocks; /* atomic "steal", but */
2061 rs->rs_nlocks = 0; /* locks still on rs_locks! */
2063 if (nlocks == 0 && !been_handled) {
2064 /* If we see this, we should already have seen the warning
2065 * in mds_steal_ack_locks() */
2066 CWARN("All locks stolen from rs %p x"LPD64".t"LPD64
2069 rs->rs_xid, rs->rs_transno, rs->rs_opc,
2070 libcfs_nid2str(exp->exp_connection->c_peer.nid));
2073 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
2074 cfs_spin_unlock(&rs->rs_lock);
2076 if (!been_handled && rs->rs_on_net) {
2077 LNetMDUnlink(rs->rs_md_h);
2078 /* Ignore return code; we're racing with
2082 while (nlocks-- > 0)
2083 ldlm_lock_decref(&rs->rs_locks[nlocks],
2084 rs->rs_modes[nlocks]);
2086 cfs_spin_lock(&rs->rs_lock);
2089 rs->rs_scheduled = 0;
2091 if (!rs->rs_on_net) {
2093 cfs_spin_unlock(&rs->rs_lock);
2095 class_export_put (exp);
2096 rs->rs_export = NULL;
2097 ptlrpc_rs_decref (rs);
2098 if (cfs_atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
2099 svc->srv_is_stopping)
2100 cfs_waitq_broadcast(&svcpt->scp_waitq);
2104 /* still on the net; callback will schedule */
2105 cfs_spin_unlock(&rs->rs_lock);
2112 * Check whether given service has a reply available for processing
2115 * \param svc a ptlrpc service
2116 * \retval 0 no replies processed
2117 * \retval 1 one reply processed
2120 ptlrpc_server_handle_reply(struct ptlrpc_service_part *svcpt)
2122 struct ptlrpc_reply_state *rs = NULL;
2125 cfs_spin_lock(&svcpt->scp_rep_lock);
2126 if (!cfs_list_empty(&svcpt->scp_rep_queue)) {
2127 rs = cfs_list_entry(svcpt->scp_rep_queue.prev,
2128 struct ptlrpc_reply_state,
2130 cfs_list_del_init(&rs->rs_list);
2132 cfs_spin_unlock(&svcpt->scp_rep_lock);
2134 ptlrpc_handle_rs(rs);
2138 /* FIXME make use of timeout later */
2140 liblustre_check_services (void *arg)
2142 int did_something = 0;
2144 cfs_list_t *tmp, *nxt;
2147 /* I'm relying on being single threaded, not to have to lock
2148 * ptlrpc_all_services etc */
2149 cfs_list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
2150 struct ptlrpc_service *svc =
2151 cfs_list_entry (tmp, struct ptlrpc_service, srv_list);
2152 struct ptlrpc_service_part *svcpt;
2154 LASSERT(svc->srv_ncpts == 1);
2155 svcpt = svc->srv_parts[0];
2157 if (svcpt->scp_nthrs_running != 0) /* I've recursed */
2160 /* service threads can block for bulk, so this limits us
2161 * (arbitrarily) to recursing 1 stack frame per service.
2162 * Note that the problem with recursion is that we have to
2163 * unwind completely before our caller can resume. */
2165 svcpt->scp_nthrs_running++;
2168 rc = ptlrpc_server_handle_req_in(svcpt);
2169 rc |= ptlrpc_server_handle_reply(svcpt);
2170 rc |= ptlrpc_at_check_timed(svcpt);
2171 rc |= ptlrpc_server_handle_request(svcpt, NULL);
2172 rc |= (ptlrpc_server_post_idle_rqbds(svcpt) > 0);
2173 did_something |= rc;
2176 svcpt->scp_nthrs_running--;
2179 RETURN(did_something);
2181 #define ptlrpc_stop_all_threads(s) do {} while (0)
2183 #else /* __KERNEL__ */
2186 ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
2188 int avail = svcpt->scp_nrqbds_posted;
2189 int low_water = test_req_buffer_pressure ? 0 :
2190 svcpt->scp_service->srv_nbuf_per_group / 2;
2192 /* NB I'm not locking; just looking. */
2194 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
2195 * allowed the request history to grow out of control. We could put a
2196 * sanity check on that here and cull some history if we need the
2199 if (avail <= low_water)
2200 ptlrpc_grow_req_bufs(svcpt, 1);
2202 if (svcpt->scp_service->srv_stats) {
2203 lprocfs_counter_add(svcpt->scp_service->srv_stats,
2204 PTLRPC_REQBUF_AVAIL_CNTR, avail);
2209 ptlrpc_retry_rqbds(void *arg)
2211 struct ptlrpc_service_part *svcpt = (struct ptlrpc_service_part *)arg;
2213 svcpt->scp_rqbd_timeout = 0;
2218 ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
2220 return svcpt->scp_nreqs_active <
2221 svcpt->scp_nthrs_running - 1 -
2222 (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL);
2226 * allowed to create more threads
2227 * user can call it w/o any lock but need to hold
2228 * ptlrpc_service_part::scp_lock to get reliable result
2231 ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
2233 return svcpt->scp_nthrs_running +
2234 svcpt->scp_nthrs_starting <
2235 svcpt->scp_service->srv_nthrs_cpt_limit;
2239 * too many requests and allowed to create more threads
2242 ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
2244 return !ptlrpc_threads_enough(svcpt) &&
2245 ptlrpc_threads_increasable(svcpt);
2249 ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2251 return thread_is_stopping(thread) ||
2252 thread->t_svcpt->scp_service->srv_is_stopping;
2256 ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
2258 return !cfs_list_empty(&svcpt->scp_rqbd_idle) &&
2259 svcpt->scp_rqbd_timeout == 0;
2263 ptlrpc_at_check(struct ptlrpc_service_part *svcpt)
2265 return svcpt->scp_at_check;
2269 * requests wait on preprocessing
2270 * user can call it w/o any lock but need to hold
2271 * ptlrpc_service_part::scp_lock to get reliable result
2274 ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
2276 return !cfs_list_empty(&svcpt->scp_req_incoming);
2279 static __attribute__((__noinline__)) int
2280 ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
2281 struct ptlrpc_thread *thread)
2283 /* Don't exit while there are replies to be handled */
2284 struct l_wait_info lwi = LWI_TIMEOUT(svcpt->scp_rqbd_timeout,
2285 ptlrpc_retry_rqbds, svcpt);
2287 lc_watchdog_disable(thread->t_watchdog);
2291 l_wait_event_exclusive_head(svcpt->scp_waitq,
2292 ptlrpc_thread_stopping(thread) ||
2293 ptlrpc_server_request_incoming(svcpt) ||
2294 ptlrpc_server_request_pending(svcpt, 0) ||
2295 ptlrpc_rqbd_pending(svcpt) ||
2296 ptlrpc_at_check(svcpt), &lwi);
2298 if (ptlrpc_thread_stopping(thread))
2301 lc_watchdog_touch(thread->t_watchdog,
2302 ptlrpc_server_get_timeout(svcpt));
2307 * Main thread body for service threads.
2308 * Waits in a loop waiting for new requests to process to appear.
2309 * Every time an incoming requests is added to its queue, a waitq
2310 * is woken up and one of the threads will handle it.
2312 static int ptlrpc_main(void *arg)
2314 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
2315 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2316 struct ptlrpc_service *svc = svcpt->scp_service;
2317 struct ptlrpc_reply_state *rs;
2318 #ifdef WITH_GROUP_INFO
2319 cfs_group_info_t *ginfo = NULL;
2322 int counter = 0, rc = 0;
2325 thread->t_pid = cfs_curproc_pid();
2326 cfs_daemonize_ctxt(thread->t_name);
2328 /* NB: we will call cfs_cpt_bind() for all threads, because we
2329 * might want to run lustre server only on a subset of system CPUs,
2330 * in that case ->scp_cpt is CFS_CPT_ANY */
2331 rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
2333 CWARN("%s: failed to bind %s on CPT %d\n",
2334 svc->srv_name, thread->t_name, svcpt->scp_cpt);
2337 #ifdef WITH_GROUP_INFO
2338 ginfo = cfs_groups_alloc(0);
2344 cfs_set_current_groups(ginfo);
2345 cfs_put_group_info(ginfo);
2348 if (svc->srv_ops.so_thr_init != NULL) {
2349 rc = svc->srv_ops.so_thr_init(thread);
2360 rc = lu_context_init(&env->le_ctx,
2361 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2365 thread->t_env = env;
2366 env->le_ctx.lc_thread = thread;
2367 env->le_ctx.lc_cookie = 0x6;
2369 while (!cfs_list_empty(&svcpt->scp_rqbd_idle)) {
2370 rc = ptlrpc_server_post_idle_rqbds(svcpt);
2374 CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
2375 svc->srv_name, svcpt->scp_cpt, rc);
2379 /* Alloc reply state structure for this one */
2380 OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2386 cfs_spin_lock(&svcpt->scp_lock);
2388 LASSERT(thread_is_starting(thread));
2389 thread_clear_flags(thread, SVC_STARTING);
2391 LASSERT(svcpt->scp_nthrs_starting == 1);
2392 svcpt->scp_nthrs_starting--;
2394 /* SVC_STOPPING may already be set here if someone else is trying
2395 * to stop the service while this new thread has been dynamically
2396 * forked. We still set SVC_RUNNING to let our creator know that
2397 * we are now running, however we will exit as soon as possible */
2398 thread_add_flags(thread, SVC_RUNNING);
2399 svcpt->scp_nthrs_running++;
2400 cfs_spin_unlock(&svcpt->scp_lock);
2402 /* wake up our creator in case he's still waiting. */
2403 cfs_waitq_signal(&thread->t_ctl_waitq);
2405 thread->t_watchdog = lc_watchdog_add(ptlrpc_server_get_timeout(svcpt),
2408 cfs_spin_lock(&svcpt->scp_rep_lock);
2409 cfs_list_add(&rs->rs_list, &svcpt->scp_rep_idle);
2410 cfs_waitq_signal(&svcpt->scp_rep_waitq);
2411 cfs_spin_unlock(&svcpt->scp_rep_lock);
2413 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2414 svcpt->scp_nthrs_running);
2416 /* XXX maintain a list of all managed devices: insert here */
2417 while (!ptlrpc_thread_stopping(thread)) {
2418 if (ptlrpc_wait_event(svcpt, thread))
2421 ptlrpc_check_rqbd_pool(svcpt);
2423 if (ptlrpc_threads_need_create(svcpt)) {
2424 /* Ignore return code - we tried... */
2425 ptlrpc_start_thread(svcpt, 0);
2428 /* Process all incoming reqs before handling any */
2429 if (ptlrpc_server_request_incoming(svcpt)) {
2430 ptlrpc_server_handle_req_in(svcpt);
2431 /* but limit ourselves in case of flood */
2432 if (counter++ < 100)
2437 if (ptlrpc_at_check(svcpt))
2438 ptlrpc_at_check_timed(svcpt);
2440 if (ptlrpc_server_request_pending(svcpt, 0)) {
2441 lu_context_enter(&env->le_ctx);
2442 ptlrpc_server_handle_request(svcpt, thread);
2443 lu_context_exit(&env->le_ctx);
2446 if (ptlrpc_rqbd_pending(svcpt) &&
2447 ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
2448 /* I just failed to repost request buffers.
2449 * Wait for a timeout (unless something else
2450 * happens) before I try again */
2451 svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
2452 CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
2453 svcpt->scp_nrqbds_posted);
2457 lc_watchdog_delete(thread->t_watchdog);
2458 thread->t_watchdog = NULL;
2462 * deconstruct service specific state created by ptlrpc_start_thread()
2464 if (svc->srv_ops.so_thr_done != NULL)
2465 svc->srv_ops.so_thr_done(thread);
2468 lu_context_fini(&env->le_ctx);
2472 CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n",
2473 thread, thread->t_pid, thread->t_id, rc);
2475 cfs_spin_lock(&svcpt->scp_lock);
2476 if (thread_test_and_clear_flags(thread, SVC_STARTING))
2477 svcpt->scp_nthrs_starting--;
2479 if (thread_test_and_clear_flags(thread, SVC_RUNNING)) {
2480 /* must know immediately */
2481 svcpt->scp_nthrs_running--;
2485 thread_add_flags(thread, SVC_STOPPED);
2487 cfs_waitq_signal(&thread->t_ctl_waitq);
2488 cfs_spin_unlock(&svcpt->scp_lock);
2493 static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
2494 cfs_list_t *replies)
2498 cfs_spin_lock(&hrt->hrt_lock);
2500 cfs_list_splice_init(&hrt->hrt_queue, replies);
2501 result = ptlrpc_hr.hr_stopping || !cfs_list_empty(replies);
2503 cfs_spin_unlock(&hrt->hrt_lock);
2508 * Main body of "handle reply" function.
2509 * It processes acked reply states
2511 static int ptlrpc_hr_main(void *arg)
2513 struct ptlrpc_hr_thread *hrt = (struct ptlrpc_hr_thread *)arg;
2514 struct ptlrpc_hr_partition *hrp = hrt->hrt_partition;
2515 CFS_LIST_HEAD (replies);
2516 char threadname[20];
2519 snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
2520 hrp->hrp_cpt, hrt->hrt_id);
2521 cfs_daemonize_ctxt(threadname);
2523 rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
2525 CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
2526 threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
2529 cfs_atomic_inc(&hrp->hrp_nstarted);
2530 cfs_waitq_signal(&ptlrpc_hr.hr_waitq);
2532 while (!ptlrpc_hr.hr_stopping) {
2533 l_wait_condition(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
2535 while (!cfs_list_empty(&replies)) {
2536 struct ptlrpc_reply_state *rs;
2538 rs = cfs_list_entry(replies.prev,
2539 struct ptlrpc_reply_state,
2541 cfs_list_del_init(&rs->rs_list);
2542 ptlrpc_handle_rs(rs);
2546 cfs_atomic_inc(&hrp->hrp_nstopped);
2547 cfs_waitq_signal(&ptlrpc_hr.hr_waitq);
2552 static void ptlrpc_stop_hr_threads(void)
2554 struct ptlrpc_hr_partition *hrp;
2558 ptlrpc_hr.hr_stopping = 1;
2560 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2561 if (hrp->hrp_thrs == NULL)
2562 continue; /* uninitialized */
2563 for (j = 0; j < hrp->hrp_nthrs; j++)
2564 cfs_waitq_broadcast(&hrp->hrp_thrs[j].hrt_waitq);
2567 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2568 if (hrp->hrp_thrs == NULL)
2569 continue; /* uninitialized */
2570 cfs_wait_event(ptlrpc_hr.hr_waitq,
2571 cfs_atomic_read(&hrp->hrp_nstopped) ==
2572 cfs_atomic_read(&hrp->hrp_nstarted));
2576 static int ptlrpc_start_hr_threads(void)
2578 struct ptlrpc_hr_partition *hrp;
2583 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2586 for (j = 0; j < hrp->hrp_nthrs; j++) {
2587 rc = cfs_create_thread(ptlrpc_hr_main,
2589 CLONE_VM | CLONE_FILES);
2593 cfs_wait_event(ptlrpc_hr.hr_waitq,
2594 cfs_atomic_read(&hrp->hrp_nstarted) == j);
2598 CERROR("Reply handling thread %d:%d Failed on starting: "
2599 "rc = %d\n", i, j, rc);
2600 ptlrpc_stop_hr_threads();
2606 static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
2608 struct l_wait_info lwi = { 0 };
2609 struct ptlrpc_thread *thread;
2610 CFS_LIST_HEAD (zombie);
2614 CDEBUG(D_INFO, "Stopping threads for service %s\n",
2615 svcpt->scp_service->srv_name);
2617 cfs_spin_lock(&svcpt->scp_lock);
2618 /* let the thread know that we would like it to stop asap */
2619 list_for_each_entry(thread, &svcpt->scp_threads, t_link) {
2620 CDEBUG(D_INFO, "Stopping thread %s #%u\n",
2621 svcpt->scp_service->srv_thread_name, thread->t_id);
2622 thread_add_flags(thread, SVC_STOPPING);
2625 cfs_waitq_broadcast(&svcpt->scp_waitq);
2627 while (!cfs_list_empty(&svcpt->scp_threads)) {
2628 thread = cfs_list_entry(svcpt->scp_threads.next,
2629 struct ptlrpc_thread, t_link);
2630 if (thread_is_stopped(thread)) {
2631 cfs_list_del(&thread->t_link);
2632 cfs_list_add(&thread->t_link, &zombie);
2635 cfs_spin_unlock(&svcpt->scp_lock);
2637 CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
2638 svcpt->scp_service->srv_thread_name, thread->t_id);
2639 l_wait_event(thread->t_ctl_waitq,
2640 thread_is_stopped(thread), &lwi);
2642 cfs_spin_lock(&svcpt->scp_lock);
2645 cfs_spin_unlock(&svcpt->scp_lock);
2647 while (!cfs_list_empty(&zombie)) {
2648 thread = cfs_list_entry(zombie.next,
2649 struct ptlrpc_thread, t_link);
2650 cfs_list_del(&thread->t_link);
2651 OBD_FREE_PTR(thread);
2657 * Stops all threads of a particular service \a svc
2659 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2661 struct ptlrpc_service_part *svcpt;
2665 ptlrpc_service_for_each_part(svcpt, i, svc) {
2666 if (svcpt->scp_service != NULL)
2667 ptlrpc_svcpt_stop_threads(svcpt);
2673 int ptlrpc_start_threads(struct ptlrpc_service *svc)
2680 /* We require 2 threads min, see note in ptlrpc_server_handle_request */
2681 LASSERT(svc->srv_nthrs_cpt_init >= PTLRPC_NTHRS_INIT);
2683 for (i = 0; i < svc->srv_ncpts; i++) {
2684 for (j = 0; j < svc->srv_nthrs_cpt_init; j++) {
2685 rc = ptlrpc_start_thread(svc->srv_parts[i], 1);
2691 /* We have enough threads, don't start more. b=15759 */
2698 CERROR("cannot start %s thread #%d_%d: rc %d\n",
2699 svc->srv_thread_name, i, j, rc);
2700 ptlrpc_stop_all_threads(svc);
2704 int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
2706 struct l_wait_info lwi = { 0 };
2707 struct ptlrpc_thread *thread;
2708 struct ptlrpc_service *svc = svcpt->scp_service;
2712 LASSERT(svcpt != NULL);
2714 CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n",
2715 svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running,
2716 svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit);
2719 if (unlikely(svc->srv_is_stopping))
2722 if (!ptlrpc_threads_increasable(svcpt) ||
2723 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
2724 svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1))
2727 OBD_CPT_ALLOC_PTR(thread, svc->srv_cptable, svcpt->scp_cpt);
2730 cfs_waitq_init(&thread->t_ctl_waitq);
2732 cfs_spin_lock(&svcpt->scp_lock);
2733 if (!ptlrpc_threads_increasable(svcpt)) {
2734 cfs_spin_unlock(&svcpt->scp_lock);
2735 OBD_FREE_PTR(thread);
2739 if (svcpt->scp_nthrs_starting != 0) {
2740 /* serialize starting because some modules (obdfilter)
2741 * might require unique and contiguous t_id */
2742 LASSERT(svcpt->scp_nthrs_starting == 1);
2743 cfs_spin_unlock(&svcpt->scp_lock);
2744 OBD_FREE_PTR(thread);
2746 CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n",
2747 svc->srv_thread_name, svcpt->scp_thr_nextid);
2752 CDEBUG(D_INFO, "Creating thread %s #%d race, retry later\n",
2753 svc->srv_thread_name, svcpt->scp_thr_nextid);
2757 svcpt->scp_nthrs_starting++;
2758 thread->t_id = svcpt->scp_thr_nextid++;
2759 thread_add_flags(thread, SVC_STARTING);
2760 thread->t_svcpt = svcpt;
2762 cfs_list_add(&thread->t_link, &svcpt->scp_threads);
2763 cfs_spin_unlock(&svcpt->scp_lock);
2765 if (svcpt->scp_cpt >= 0) {
2766 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s%02d_%03d",
2767 svc->srv_thread_name, svcpt->scp_cpt, thread->t_id);
2769 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s_%04d",
2770 svc->srv_thread_name, thread->t_id);
2773 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name);
2775 * CLONE_VM and CLONE_FILES just avoid a needless copy, because we
2776 * just drop the VM and FILES in cfs_daemonize_ctxt() right away.
2778 rc = cfs_create_thread(ptlrpc_main, thread, CFS_DAEMON_FLAGS);
2780 CERROR("cannot start thread '%s': rc %d\n",
2781 thread->t_name, rc);
2782 cfs_spin_lock(&svcpt->scp_lock);
2783 cfs_list_del(&thread->t_link);
2784 --svcpt->scp_nthrs_starting;
2785 cfs_spin_unlock(&svcpt->scp_lock);
2787 OBD_FREE(thread, sizeof(*thread));
2794 l_wait_event(thread->t_ctl_waitq,
2795 thread_is_running(thread) || thread_is_stopped(thread),
2798 rc = thread_is_stopped(thread) ? thread->t_id : 0;
2802 int ptlrpc_hr_init(void)
2804 struct ptlrpc_hr_partition *hrp;
2805 struct ptlrpc_hr_thread *hrt;
2811 memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
2812 ptlrpc_hr.hr_cpt_table = cfs_cpt_table;
2814 ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
2816 if (ptlrpc_hr.hr_partitions == NULL)
2819 cfs_waitq_init(&ptlrpc_hr.hr_waitq);
2821 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2824 cfs_atomic_set(&hrp->hrp_nstarted, 0);
2825 cfs_atomic_set(&hrp->hrp_nstopped, 0);
2827 hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, i);
2828 hrp->hrp_nthrs /= cfs_cpu_ht_nsiblings(0);
2830 LASSERT(hrp->hrp_nthrs > 0);
2831 OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, i,
2832 hrp->hrp_nthrs * sizeof(*hrt));
2833 if (hrp->hrp_thrs == NULL)
2834 GOTO(out, rc = -ENOMEM);
2836 for (j = 0; j < hrp->hrp_nthrs; j++) {
2837 hrt = &hrp->hrp_thrs[j];
2840 hrt->hrt_partition = hrp;
2841 cfs_waitq_init(&hrt->hrt_waitq);
2842 cfs_spin_lock_init(&hrt->hrt_lock);
2843 CFS_INIT_LIST_HEAD(&hrt->hrt_queue);
2847 rc = ptlrpc_start_hr_threads();
2854 void ptlrpc_hr_fini(void)
2856 struct ptlrpc_hr_partition *hrp;
2859 if (ptlrpc_hr.hr_partitions == NULL)
2862 ptlrpc_stop_hr_threads();
2864 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2865 if (hrp->hrp_thrs != NULL) {
2866 OBD_FREE(hrp->hrp_thrs,
2867 hrp->hrp_nthrs * sizeof(hrp->hrp_thrs[0]));
2871 cfs_percpt_free(ptlrpc_hr.hr_partitions);
2872 ptlrpc_hr.hr_partitions = NULL;
2875 #endif /* __KERNEL__ */
2878 * Wait until all already scheduled replies are processed.
2880 static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
2884 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
2887 rc = l_wait_event(svcpt->scp_waitq,
2888 cfs_atomic_read(&svcpt->scp_nreps_difficult) == 0, &lwi);
2891 CWARN("Unexpectedly long timeout %s %p\n",
2892 svcpt->scp_service->srv_name, svcpt->scp_service);
2897 ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
2899 struct ptlrpc_service_part *svcpt;
2902 /* early disarm AT timer... */
2903 ptlrpc_service_for_each_part(svcpt, i, svc) {
2904 if (svcpt->scp_service != NULL)
2905 cfs_timer_disarm(&svcpt->scp_at_timer);
2910 ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
2912 struct ptlrpc_service_part *svcpt;
2913 struct ptlrpc_request_buffer_desc *rqbd;
2914 struct l_wait_info lwi;
2918 /* All history will be culled when the next request buffer is
2919 * freed in ptlrpc_service_purge_all() */
2920 svc->srv_hist_nrqbds_cpt_max = 0;
2922 rc = LNetClearLazyPortal(svc->srv_req_portal);
2925 ptlrpc_service_for_each_part(svcpt, i, svc) {
2926 if (svcpt->scp_service == NULL)
2929 /* Unlink all the request buffers. This forces a 'final'
2930 * event with its 'unlink' flag set for each posted rqbd */
2931 cfs_list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
2933 rc = LNetMDUnlink(rqbd->rqbd_md_h);
2934 LASSERT(rc == 0 || rc == -ENOENT);
2938 ptlrpc_service_for_each_part(svcpt, i, svc) {
2939 if (svcpt->scp_service == NULL)
2942 /* Wait for the network to release any buffers
2943 * it's currently filling */
2944 cfs_spin_lock(&svcpt->scp_lock);
2945 while (svcpt->scp_nrqbds_posted != 0) {
2946 cfs_spin_unlock(&svcpt->scp_lock);
2947 /* Network access will complete in finite time but
2948 * the HUGE timeout lets us CWARN for visibility
2949 * of sluggish NALs */
2950 lwi = LWI_TIMEOUT_INTERVAL(
2951 cfs_time_seconds(LONG_UNLINK),
2952 cfs_time_seconds(1), NULL, NULL);
2953 rc = l_wait_event(svcpt->scp_waitq,
2954 svcpt->scp_nrqbds_posted == 0, &lwi);
2955 if (rc == -ETIMEDOUT) {
2956 CWARN("Service %s waiting for "
2957 "request buffers\n",
2958 svcpt->scp_service->srv_name);
2960 cfs_spin_lock(&svcpt->scp_lock);
2962 cfs_spin_unlock(&svcpt->scp_lock);
2967 ptlrpc_service_purge_all(struct ptlrpc_service *svc)
2969 struct ptlrpc_service_part *svcpt;
2970 struct ptlrpc_request_buffer_desc *rqbd;
2971 struct ptlrpc_request *req;
2972 struct ptlrpc_reply_state *rs;
2975 ptlrpc_service_for_each_part(svcpt, i, svc) {
2976 if (svcpt->scp_service == NULL)
2979 cfs_spin_lock(&svcpt->scp_rep_lock);
2980 while (!cfs_list_empty(&svcpt->scp_rep_active)) {
2981 rs = cfs_list_entry(svcpt->scp_rep_active.next,
2982 struct ptlrpc_reply_state, rs_list);
2983 cfs_spin_lock(&rs->rs_lock);
2984 ptlrpc_schedule_difficult_reply(rs);
2985 cfs_spin_unlock(&rs->rs_lock);
2987 cfs_spin_unlock(&svcpt->scp_rep_lock);
2989 /* purge the request queue. NB No new replies (rqbds
2990 * all unlinked) and no service threads, so I'm the only
2991 * thread noodling the request queue now */
2992 while (!cfs_list_empty(&svcpt->scp_req_incoming)) {
2993 req = cfs_list_entry(svcpt->scp_req_incoming.next,
2994 struct ptlrpc_request, rq_list);
2996 cfs_list_del(&req->rq_list);
2997 svcpt->scp_nreqs_incoming--;
2998 svcpt->scp_nreqs_active++;
2999 ptlrpc_server_finish_request(svcpt, req);
3002 while (ptlrpc_server_request_pending(svcpt, 1)) {
3003 req = ptlrpc_server_request_get(svcpt, 1);
3004 cfs_list_del(&req->rq_list);
3005 svcpt->scp_nreqs_active++;
3006 ptlrpc_hpreq_fini(req);
3007 ptlrpc_server_finish_request(svcpt, req);
3010 LASSERT(cfs_list_empty(&svcpt->scp_rqbd_posted));
3011 LASSERT(svcpt->scp_nreqs_incoming == 0);
3012 LASSERT(svcpt->scp_nreqs_active == 0);
3013 /* history should have been culled by
3014 * ptlrpc_server_finish_request */
3015 LASSERT(svcpt->scp_hist_nrqbds == 0);
3017 /* Now free all the request buffers since nothing
3018 * references them any more... */
3020 while (!cfs_list_empty(&svcpt->scp_rqbd_idle)) {
3021 rqbd = cfs_list_entry(svcpt->scp_rqbd_idle.next,
3022 struct ptlrpc_request_buffer_desc,
3024 ptlrpc_free_rqbd(rqbd);
3026 ptlrpc_wait_replies(svcpt);
3028 while (!cfs_list_empty(&svcpt->scp_rep_idle)) {
3029 rs = cfs_list_entry(svcpt->scp_rep_idle.next,
3030 struct ptlrpc_reply_state,
3032 cfs_list_del(&rs->rs_list);
3033 OBD_FREE_LARGE(rs, svc->srv_max_reply_size);
3039 ptlrpc_service_free(struct ptlrpc_service *svc)
3041 struct ptlrpc_service_part *svcpt;
3042 struct ptlrpc_at_array *array;
3045 ptlrpc_service_for_each_part(svcpt, i, svc) {
3046 if (svcpt->scp_service == NULL)
3049 /* In case somebody rearmed this in the meantime */
3050 cfs_timer_disarm(&svcpt->scp_at_timer);
3051 array = &svcpt->scp_at_array;
3053 if (array->paa_reqs_array != NULL) {
3054 OBD_FREE(array->paa_reqs_array,
3055 sizeof(cfs_list_t) * array->paa_size);
3056 array->paa_reqs_array = NULL;
3059 if (array->paa_reqs_count != NULL) {
3060 OBD_FREE(array->paa_reqs_count,
3061 sizeof(__u32) * array->paa_size);
3062 array->paa_reqs_count = NULL;
3066 ptlrpc_service_for_each_part(svcpt, i, svc)
3067 OBD_FREE_PTR(svcpt);
3069 if (svc->srv_cpts != NULL)
3070 cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts);
3072 OBD_FREE(svc, offsetof(struct ptlrpc_service,
3073 srv_parts[svc->srv_ncpts]));
3076 int ptlrpc_unregister_service(struct ptlrpc_service *service)
3080 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
3082 service->srv_is_stopping = 1;
3084 cfs_spin_lock(&ptlrpc_all_services_lock);
3085 cfs_list_del_init(&service->srv_list);
3086 cfs_spin_unlock(&ptlrpc_all_services_lock);
3088 ptlrpc_lprocfs_unregister_service(service);
3090 ptlrpc_service_del_atimer(service);
3091 ptlrpc_stop_all_threads(service);
3093 ptlrpc_service_unlink_rqbd(service);
3094 ptlrpc_service_purge_all(service);
3095 ptlrpc_service_free(service);
3101 * Returns 0 if the service is healthy.
3103 * Right now, it just checks to make sure that requests aren't languishing
3104 * in the queue. We'll use this health check to govern whether a node needs
3105 * to be shot, so it's intentionally non-aggressive. */
3106 int ptlrpc_svcpt_health_check(struct ptlrpc_service_part *svcpt)
3108 struct ptlrpc_request *request;
3109 struct timeval right_now;
3112 cfs_gettimeofday(&right_now);
3114 cfs_spin_lock(&svcpt->scp_req_lock);
3115 if (!ptlrpc_server_request_pending(svcpt, 1)) {
3116 cfs_spin_unlock(&svcpt->scp_req_lock);
3120 /* How long has the next entry been waiting? */
3121 if (cfs_list_empty(&svcpt->scp_req_pending)) {
3122 request = cfs_list_entry(svcpt->scp_hreq_pending.next,
3123 struct ptlrpc_request, rq_list);
3125 request = cfs_list_entry(svcpt->scp_req_pending.next,
3126 struct ptlrpc_request, rq_list);
3129 timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
3130 cfs_spin_unlock(&svcpt->scp_req_lock);
3132 if ((timediff / ONE_MILLION) >
3133 (AT_OFF ? obd_timeout * 3 / 2 : at_max)) {
3134 CERROR("%s: unhealthy - request has been waiting %lds\n",
3135 svcpt->scp_service->srv_name, timediff / ONE_MILLION);
3143 ptlrpc_service_health_check(struct ptlrpc_service *svc)
3145 struct ptlrpc_service_part *svcpt;
3148 if (svc == NULL || svc->srv_parts == NULL)
3151 ptlrpc_service_for_each_part(svcpt, i, svc) {
3152 int rc = ptlrpc_svcpt_health_check(svcpt);