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;
178 EXPORT_SYMBOL(ptlrpc_save_lock);
182 struct ptlrpc_hr_partition;
184 struct ptlrpc_hr_thread {
185 int hrt_id; /* thread ID */
186 cfs_spinlock_t hrt_lock;
187 cfs_waitq_t hrt_waitq;
188 cfs_list_t hrt_queue; /* RS queue */
189 struct ptlrpc_hr_partition *hrt_partition;
192 struct ptlrpc_hr_partition {
193 /* # of started threads */
194 cfs_atomic_t hrp_nstarted;
195 /* # of stopped threads */
196 cfs_atomic_t hrp_nstopped;
197 /* cpu partition id */
199 /* round-robin rotor for choosing thread */
201 /* total number of threads on this partition */
204 struct ptlrpc_hr_thread *hrp_thrs;
207 #define HRT_RUNNING 0
208 #define HRT_STOPPING 1
210 struct ptlrpc_hr_service {
211 /* CPU partition table, it's just cfs_cpt_table for now */
212 struct cfs_cpt_table *hr_cpt_table;
213 /** controller sleep waitq */
214 cfs_waitq_t hr_waitq;
215 unsigned int hr_stopping;
216 /** roundrobin rotor for non-affinity service */
217 unsigned int hr_rotor;
219 struct ptlrpc_hr_partition **hr_partitions;
223 cfs_list_t rsb_replies;
224 unsigned int rsb_n_replies;
225 struct ptlrpc_service_part *rsb_svcpt;
228 /** reply handling service. */
229 static struct ptlrpc_hr_service ptlrpc_hr;
232 * maximum mumber of replies scheduled in one batch
234 #define MAX_SCHEDULED 256
237 * Initialize a reply batch.
241 static void rs_batch_init(struct rs_batch *b)
243 memset(b, 0, sizeof *b);
244 CFS_INIT_LIST_HEAD(&b->rsb_replies);
248 * Choose an hr thread to dispatch requests to.
250 static struct ptlrpc_hr_thread *
251 ptlrpc_hr_select(struct ptlrpc_service_part *svcpt)
253 struct ptlrpc_hr_partition *hrp;
256 if (svcpt->scp_cpt >= 0 &&
257 svcpt->scp_service->srv_cptable == ptlrpc_hr.hr_cpt_table) {
258 /* directly match partition */
259 hrp = ptlrpc_hr.hr_partitions[svcpt->scp_cpt];
262 rotor = ptlrpc_hr.hr_rotor++;
263 rotor %= cfs_cpt_number(ptlrpc_hr.hr_cpt_table);
265 hrp = ptlrpc_hr.hr_partitions[rotor];
268 rotor = hrp->hrp_rotor++;
269 return &hrp->hrp_thrs[rotor % hrp->hrp_nthrs];
273 * Dispatch all replies accumulated in the batch to one from
274 * dedicated reply handling threads.
278 static void rs_batch_dispatch(struct rs_batch *b)
280 if (b->rsb_n_replies != 0) {
281 struct ptlrpc_hr_thread *hrt;
283 hrt = ptlrpc_hr_select(b->rsb_svcpt);
285 cfs_spin_lock(&hrt->hrt_lock);
286 cfs_list_splice_init(&b->rsb_replies, &hrt->hrt_queue);
287 cfs_spin_unlock(&hrt->hrt_lock);
289 cfs_waitq_signal(&hrt->hrt_waitq);
290 b->rsb_n_replies = 0;
295 * Add a reply to a batch.
296 * Add one reply object to a batch, schedule batched replies if overload.
301 static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs)
303 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
305 if (svcpt != b->rsb_svcpt || b->rsb_n_replies >= MAX_SCHEDULED) {
306 if (b->rsb_svcpt != NULL) {
307 rs_batch_dispatch(b);
308 cfs_spin_unlock(&b->rsb_svcpt->scp_rep_lock);
310 cfs_spin_lock(&svcpt->scp_rep_lock);
311 b->rsb_svcpt = svcpt;
313 cfs_spin_lock(&rs->rs_lock);
314 rs->rs_scheduled_ever = 1;
315 if (rs->rs_scheduled == 0) {
316 cfs_list_move(&rs->rs_list, &b->rsb_replies);
317 rs->rs_scheduled = 1;
320 rs->rs_committed = 1;
321 cfs_spin_unlock(&rs->rs_lock);
325 * Reply batch finalization.
326 * Dispatch remaining replies from the batch
327 * and release remaining spinlock.
331 static void rs_batch_fini(struct rs_batch *b)
333 if (b->rsb_svcpt != NULL) {
334 rs_batch_dispatch(b);
335 cfs_spin_unlock(&b->rsb_svcpt->scp_rep_lock);
339 #define DECLARE_RS_BATCH(b) struct rs_batch b
341 #else /* __KERNEL__ */
343 #define rs_batch_init(b) do{}while(0)
344 #define rs_batch_fini(b) do{}while(0)
345 #define rs_batch_add(b, r) ptlrpc_schedule_difficult_reply(r)
346 #define DECLARE_RS_BATCH(b)
348 #endif /* __KERNEL__ */
351 * Put reply state into a queue for processing because we received
352 * ACK from the client
354 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
357 struct ptlrpc_hr_thread *hrt;
360 LASSERT(cfs_list_empty(&rs->rs_list));
362 hrt = ptlrpc_hr_select(rs->rs_svcpt);
364 cfs_spin_lock(&hrt->hrt_lock);
365 cfs_list_add_tail(&rs->rs_list, &hrt->hrt_queue);
366 cfs_spin_unlock(&hrt->hrt_lock);
368 cfs_waitq_signal(&hrt->hrt_waitq);
371 cfs_list_add_tail(&rs->rs_list, &rs->rs_svcpt->scp_rep_queue);
376 ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs)
380 LASSERT_SPIN_LOCKED(&rs->rs_svcpt->scp_rep_lock);
381 LASSERT_SPIN_LOCKED(&rs->rs_lock);
382 LASSERT (rs->rs_difficult);
383 rs->rs_scheduled_ever = 1; /* flag any notification attempt */
385 if (rs->rs_scheduled) { /* being set up or already notified */
390 rs->rs_scheduled = 1;
391 cfs_list_del_init(&rs->rs_list);
392 ptlrpc_dispatch_difficult_reply(rs);
395 EXPORT_SYMBOL(ptlrpc_schedule_difficult_reply);
397 void ptlrpc_commit_replies(struct obd_export *exp)
399 struct ptlrpc_reply_state *rs, *nxt;
400 DECLARE_RS_BATCH(batch);
403 rs_batch_init(&batch);
404 /* Find any replies that have been committed and get their service
405 * to attend to complete them. */
407 /* CAVEAT EMPTOR: spinlock ordering!!! */
408 cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
409 cfs_list_for_each_entry_safe(rs, nxt, &exp->exp_uncommitted_replies,
411 LASSERT (rs->rs_difficult);
412 /* VBR: per-export last_committed */
413 LASSERT(rs->rs_export);
414 if (rs->rs_transno <= exp->exp_last_committed) {
415 cfs_list_del_init(&rs->rs_obd_list);
416 rs_batch_add(&batch, rs);
419 cfs_spin_unlock(&exp->exp_uncommitted_replies_lock);
420 rs_batch_fini(&batch);
423 EXPORT_SYMBOL(ptlrpc_commit_replies);
426 ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt)
428 struct ptlrpc_request_buffer_desc *rqbd;
433 cfs_spin_lock(&svcpt->scp_lock);
435 if (cfs_list_empty(&svcpt->scp_rqbd_idle)) {
436 cfs_spin_unlock(&svcpt->scp_lock);
440 rqbd = cfs_list_entry(svcpt->scp_rqbd_idle.next,
441 struct ptlrpc_request_buffer_desc,
443 cfs_list_del(&rqbd->rqbd_list);
445 /* assume we will post successfully */
446 svcpt->scp_nrqbds_posted++;
447 cfs_list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_posted);
449 cfs_spin_unlock(&svcpt->scp_lock);
451 rc = ptlrpc_register_rqbd(rqbd);
458 cfs_spin_lock(&svcpt->scp_lock);
460 svcpt->scp_nrqbds_posted--;
461 cfs_list_del(&rqbd->rqbd_list);
462 cfs_list_add_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
464 /* Don't complain if no request buffers are posted right now; LNET
465 * won't drop requests because we set the portal lazy! */
467 cfs_spin_unlock(&svcpt->scp_lock);
472 static void ptlrpc_at_timer(unsigned long castmeharder)
474 struct ptlrpc_service_part *svcpt;
476 svcpt = (struct ptlrpc_service_part *)castmeharder;
478 svcpt->scp_at_check = 1;
479 svcpt->scp_at_checktime = cfs_time_current();
480 cfs_waitq_signal(&svcpt->scp_waitq);
484 ptlrpc_server_nthreads_check(struct ptlrpc_service *svc,
485 struct ptlrpc_service_conf *conf)
488 struct ptlrpc_service_thr_conf *tc = &conf->psc_thr;
495 * Common code for estimating & validating threads number.
496 * CPT affinity service could have percpt thread-pool instead
497 * of a global thread-pool, which means user might not always
498 * get the threads number they give it in conf::tc_nthrs_user
499 * even they did set. It's because we need to validate threads
500 * number for each CPT to guarantee each pool will have enough
501 * threads to keep the service healthy.
503 init = PTLRPC_NTHRS_INIT + (svc->srv_ops.so_hpreq_handler != NULL);
504 init = max_t(int, init, tc->tc_nthrs_init);
506 /* NB: please see comments in lustre_lnet.h for definition
507 * details of these members */
508 LASSERT(tc->tc_nthrs_max != 0);
510 if (tc->tc_nthrs_user != 0) {
511 /* In case there is a reason to test a service with many
512 * threads, we give a less strict check here, it can
513 * be up to 8 * nthrs_max */
514 total = min(tc->tc_nthrs_max * 8, tc->tc_nthrs_user);
515 nthrs = total / svc->srv_ncpts;
516 init = max(init, nthrs);
520 total = tc->tc_nthrs_max;
521 if (tc->tc_nthrs_base == 0) {
522 /* don't care about base threads number per partition,
523 * this is most for non-affinity service */
524 nthrs = total / svc->srv_ncpts;
528 nthrs = tc->tc_nthrs_base;
529 if (svc->srv_ncpts == 1) {
532 /* NB: Increase the base number if it's single partition
533 * and total number of cores/HTs is larger or equal to 4.
534 * result will always < 2 * nthrs_base */
535 weight = cfs_cpt_weight(svc->srv_cptable, CFS_CPT_ANY);
536 for (i = 1; (weight >> (i + 1)) != 0 && /* >= 4 cores/HTs */
537 (tc->tc_nthrs_base >> i) != 0; i++)
538 nthrs += tc->tc_nthrs_base >> i;
541 if (tc->tc_thr_factor != 0) {
542 int factor = tc->tc_thr_factor;
546 * User wants to increase number of threads with for
547 * each CPU core/HT, most likely the factor is larger then
548 * one thread/core because service threads are supposed to
549 * be blocked by lock or wait for IO.
552 * Amdahl's law says that adding processors wouldn't give
553 * a linear increasing of parallelism, so it's nonsense to
554 * have too many threads no matter how many cores/HTs
557 if (cfs_cpu_ht_nsiblings(0) > 1) { /* weight is # of HTs */
558 /* depress thread factor for hyper-thread */
559 factor = factor - (factor >> 1) + (factor >> 3);
562 weight = cfs_cpt_weight(svc->srv_cptable, 0);
565 for (; factor > 0 && weight > 0; factor--, weight -= fade)
566 nthrs += min(weight, fade) * factor;
569 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
570 nthrs = max(tc->tc_nthrs_base,
571 tc->tc_nthrs_max / svc->srv_ncpts);
574 nthrs = max(nthrs, tc->tc_nthrs_init);
575 svc->srv_nthrs_cpt_limit = nthrs;
576 svc->srv_nthrs_cpt_init = init;
578 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
579 LCONSOLE_WARN("%s: This service may have more threads (%d) "
580 "than the given soft limit (%d)\n",
581 svc->srv_name, nthrs * svc->srv_ncpts,
588 * Initialize percpt data for a service
591 ptlrpc_service_part_init(struct ptlrpc_service *svc,
592 struct ptlrpc_service_part *svcpt, int cpt)
594 struct ptlrpc_at_array *array;
599 svcpt->scp_cpt = cpt;
600 CFS_INIT_LIST_HEAD(&svcpt->scp_threads);
602 /* rqbd and incoming request queue */
603 cfs_spin_lock_init(&svcpt->scp_lock);
604 CFS_INIT_LIST_HEAD(&svcpt->scp_rqbd_idle);
605 CFS_INIT_LIST_HEAD(&svcpt->scp_rqbd_posted);
606 CFS_INIT_LIST_HEAD(&svcpt->scp_req_incoming);
607 cfs_waitq_init(&svcpt->scp_waitq);
608 /* history request & rqbd list */
609 CFS_INIT_LIST_HEAD(&svcpt->scp_hist_reqs);
610 CFS_INIT_LIST_HEAD(&svcpt->scp_hist_rqbds);
612 /* acitve requests and hp requests */
613 cfs_spin_lock_init(&svcpt->scp_req_lock);
614 CFS_INIT_LIST_HEAD(&svcpt->scp_req_pending);
615 CFS_INIT_LIST_HEAD(&svcpt->scp_hreq_pending);
618 cfs_spin_lock_init(&svcpt->scp_rep_lock);
619 CFS_INIT_LIST_HEAD(&svcpt->scp_rep_active);
621 CFS_INIT_LIST_HEAD(&svcpt->scp_rep_queue);
623 CFS_INIT_LIST_HEAD(&svcpt->scp_rep_idle);
624 cfs_waitq_init(&svcpt->scp_rep_waitq);
625 cfs_atomic_set(&svcpt->scp_nreps_difficult, 0);
627 /* adaptive timeout */
628 cfs_spin_lock_init(&svcpt->scp_at_lock);
629 array = &svcpt->scp_at_array;
631 size = at_est2timeout(at_max);
632 array->paa_size = size;
633 array->paa_count = 0;
634 array->paa_deadline = -1;
636 /* allocate memory for scp_at_array (ptlrpc_at_array) */
637 OBD_CPT_ALLOC(array->paa_reqs_array,
638 svc->srv_cptable, cpt, sizeof(cfs_list_t) * size);
639 if (array->paa_reqs_array == NULL)
642 for (index = 0; index < size; index++)
643 CFS_INIT_LIST_HEAD(&array->paa_reqs_array[index]);
645 OBD_CPT_ALLOC(array->paa_reqs_count,
646 svc->srv_cptable, cpt, sizeof(__u32) * size);
647 if (array->paa_reqs_count == NULL)
650 cfs_timer_init(&svcpt->scp_at_timer, ptlrpc_at_timer, svcpt);
651 /* At SOW, service time should be quick; 10s seems generous. If client
652 * timeout is less than this, we'll be sending an early reply. */
653 at_init(&svcpt->scp_at_estimate, 10, 0);
655 /* assign this before call ptlrpc_grow_req_bufs */
656 svcpt->scp_service = svc;
657 /* Now allocate the request buffers, but don't post them now */
658 rc = ptlrpc_grow_req_bufs(svcpt, 0);
659 /* We shouldn't be under memory pressure at startup, so
660 * fail if we can't allocate all our buffers at this time. */
667 if (array->paa_reqs_count != NULL) {
668 OBD_FREE(array->paa_reqs_count, sizeof(__u32) * size);
669 array->paa_reqs_count = NULL;
672 if (array->paa_reqs_array != NULL) {
673 OBD_FREE(array->paa_reqs_array,
674 sizeof(cfs_list_t) * array->paa_size);
675 array->paa_reqs_array = NULL;
682 * Initialize service on a given portal.
683 * This includes starting serving threads , allocating and posting rqbds and
686 struct ptlrpc_service *
687 ptlrpc_register_service(struct ptlrpc_service_conf *conf,
688 cfs_proc_dir_entry_t *proc_entry)
690 struct ptlrpc_service_cpt_conf *cconf = &conf->psc_cpt;
691 struct ptlrpc_service *service;
692 struct ptlrpc_service_part *svcpt;
693 struct cfs_cpt_table *cptable;
701 LASSERT(conf->psc_buf.bc_nbufs > 0);
702 LASSERT(conf->psc_buf.bc_buf_size >=
703 conf->psc_buf.bc_req_max_size + SPTLRPC_MAX_PAYLOAD);
704 LASSERT(conf->psc_thr.tc_ctx_tags != 0);
706 cptable = cconf->cc_cptable;
708 cptable = cfs_cpt_table;
710 if (!conf->psc_thr.tc_cpu_affinity) {
713 ncpts = cfs_cpt_number(cptable);
714 if (cconf->cc_pattern != NULL) {
715 struct cfs_expr_list *el;
717 rc = cfs_expr_list_parse(cconf->cc_pattern,
718 strlen(cconf->cc_pattern),
721 CERROR("%s: invalid CPT pattern string: %s",
722 conf->psc_name, cconf->cc_pattern);
723 RETURN(ERR_PTR(-EINVAL));
726 rc = cfs_expr_list_values(el, ncpts, &cpts);
727 cfs_expr_list_free(el);
729 CERROR("%s: failed to parse CPT array %s: %d\n",
730 conf->psc_name, cconf->cc_pattern, rc);
731 RETURN(ERR_PTR(rc < 0 ? rc : -EINVAL));
737 OBD_ALLOC(service, offsetof(struct ptlrpc_service, srv_parts[ncpts]));
738 if (service == NULL) {
740 OBD_FREE(cpts, sizeof(*cpts) * ncpts);
741 RETURN(ERR_PTR(-ENOMEM));
744 service->srv_cptable = cptable;
745 service->srv_cpts = cpts;
746 service->srv_ncpts = ncpts;
748 service->srv_cpt_bits = 0; /* it's zero already, easy to read... */
749 while ((1 << service->srv_cpt_bits) < cfs_cpt_number(cptable))
750 service->srv_cpt_bits++;
753 cfs_spin_lock_init(&service->srv_lock);
754 service->srv_name = conf->psc_name;
755 service->srv_watchdog_factor = conf->psc_watchdog_factor;
756 CFS_INIT_LIST_HEAD(&service->srv_list); /* for safty of cleanup */
758 /* buffer configuration */
759 service->srv_nbuf_per_group = test_req_buffer_pressure ? 1 :
760 max(conf->psc_buf.bc_nbufs /
761 service->srv_ncpts, 1U);
762 service->srv_max_req_size = conf->psc_buf.bc_req_max_size +
764 service->srv_buf_size = conf->psc_buf.bc_buf_size;
765 service->srv_rep_portal = conf->psc_buf.bc_rep_portal;
766 service->srv_req_portal = conf->psc_buf.bc_req_portal;
768 /* Increase max reply size to next power of two */
769 service->srv_max_reply_size = 1;
770 while (service->srv_max_reply_size <
771 conf->psc_buf.bc_rep_max_size + SPTLRPC_MAX_PAYLOAD)
772 service->srv_max_reply_size <<= 1;
774 service->srv_thread_name = conf->psc_thr.tc_thr_name;
775 service->srv_ctx_tags = conf->psc_thr.tc_ctx_tags;
776 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
777 service->srv_ops = conf->psc_ops;
779 for (i = 0; i < ncpts; i++) {
780 if (!conf->psc_thr.tc_cpu_affinity)
783 cpt = cpts != NULL ? cpts[i] : i;
785 OBD_CPT_ALLOC(svcpt, cptable, cpt, sizeof(*svcpt));
787 GOTO(failed, rc = -ENOMEM);
789 service->srv_parts[i] = svcpt;
790 rc = ptlrpc_service_part_init(service, svcpt, cpt);
795 ptlrpc_server_nthreads_check(service, conf);
797 rc = LNetSetLazyPortal(service->srv_req_portal);
800 cfs_spin_lock (&ptlrpc_all_services_lock);
801 cfs_list_add (&service->srv_list, &ptlrpc_all_services);
802 cfs_spin_unlock (&ptlrpc_all_services_lock);
804 if (proc_entry != NULL)
805 ptlrpc_lprocfs_register_service(proc_entry, service);
807 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
808 service->srv_name, service->srv_req_portal);
811 rc = ptlrpc_start_threads(service);
813 CERROR("Failed to start threads for service %s: %d\n",
814 service->srv_name, rc);
821 ptlrpc_unregister_service(service);
824 EXPORT_SYMBOL(ptlrpc_register_service);
827 * to actually free the request, must be called without holding svc_lock.
828 * note it's caller's responsibility to unlink req->rq_list.
830 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
832 LASSERT(cfs_atomic_read(&req->rq_refcount) == 0);
833 LASSERT(cfs_list_empty(&req->rq_timed_list));
835 /* DEBUG_REQ() assumes the reply state of a request with a valid
836 * ref will not be destroyed until that reference is dropped. */
837 ptlrpc_req_drop_rs(req);
839 sptlrpc_svc_ctx_decref(req);
841 if (req != &req->rq_rqbd->rqbd_req) {
842 /* NB request buffers use an embedded
843 * req if the incoming req unlinked the
844 * MD; this isn't one of them! */
845 OBD_FREE(req, sizeof(*req));
850 * drop a reference count of the request. if it reaches 0, we either
851 * put it into history list, or free it immediately.
853 void ptlrpc_server_drop_request(struct ptlrpc_request *req)
855 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
856 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
857 struct ptlrpc_service *svc = svcpt->scp_service;
862 if (!cfs_atomic_dec_and_test(&req->rq_refcount))
865 if (req->rq_at_linked) {
866 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
867 __u32 index = req->rq_at_index;
869 cfs_spin_lock(&svcpt->scp_at_lock);
871 LASSERT(!cfs_list_empty(&req->rq_timed_list));
872 cfs_list_del_init(&req->rq_timed_list);
873 cfs_spin_lock(&req->rq_lock);
874 req->rq_at_linked = 0;
875 cfs_spin_unlock(&req->rq_lock);
876 array->paa_reqs_count[index]--;
879 cfs_spin_unlock(&svcpt->scp_at_lock);
881 LASSERT(cfs_list_empty(&req->rq_timed_list));
884 /* finalize request */
885 if (req->rq_export) {
886 class_export_put(req->rq_export);
887 req->rq_export = NULL;
890 cfs_spin_lock(&svcpt->scp_lock);
892 cfs_list_add(&req->rq_list, &rqbd->rqbd_reqs);
894 refcount = --(rqbd->rqbd_refcount);
896 /* request buffer is now idle: add to history */
897 cfs_list_del(&rqbd->rqbd_list);
899 cfs_list_add_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds);
900 svcpt->scp_hist_nrqbds++;
902 /* cull some history?
903 * I expect only about 1 or 2 rqbds need to be recycled here */
904 while (svcpt->scp_hist_nrqbds > svc->srv_hist_nrqbds_cpt_max) {
905 rqbd = cfs_list_entry(svcpt->scp_hist_rqbds.next,
906 struct ptlrpc_request_buffer_desc,
909 cfs_list_del(&rqbd->rqbd_list);
910 svcpt->scp_hist_nrqbds--;
912 /* remove rqbd's reqs from svc's req history while
913 * I've got the service lock */
914 cfs_list_for_each(tmp, &rqbd->rqbd_reqs) {
915 req = cfs_list_entry(tmp, struct ptlrpc_request,
917 /* Track the highest culled req seq */
918 if (req->rq_history_seq >
919 svcpt->scp_hist_seq_culled) {
920 svcpt->scp_hist_seq_culled =
923 cfs_list_del(&req->rq_history_list);
926 cfs_spin_unlock(&svcpt->scp_lock);
928 cfs_list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
929 req = cfs_list_entry(rqbd->rqbd_reqs.next,
930 struct ptlrpc_request,
932 cfs_list_del(&req->rq_list);
933 ptlrpc_server_free_request(req);
936 cfs_spin_lock(&svcpt->scp_lock);
938 * now all reqs including the embedded req has been
939 * disposed, schedule request buffer for re-use.
941 LASSERT(cfs_atomic_read(&rqbd->rqbd_req.rq_refcount) ==
943 cfs_list_add_tail(&rqbd->rqbd_list,
944 &svcpt->scp_rqbd_idle);
947 cfs_spin_unlock(&svcpt->scp_lock);
948 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
949 /* If we are low on memory, we are not interested in history */
950 cfs_list_del(&req->rq_list);
951 cfs_list_del_init(&req->rq_history_list);
953 cfs_spin_unlock(&svcpt->scp_lock);
955 ptlrpc_server_free_request(req);
957 cfs_spin_unlock(&svcpt->scp_lock);
962 * to finish a request: stop sending more early replies, and release
963 * the request. should be called after we finished handling the request.
965 static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt,
966 struct ptlrpc_request *req)
968 ptlrpc_hpreq_fini(req);
970 cfs_spin_lock(&svcpt->scp_req_lock);
971 svcpt->scp_nreqs_active--;
973 svcpt->scp_nhreqs_active--;
974 cfs_spin_unlock(&svcpt->scp_req_lock);
976 ptlrpc_server_drop_request(req);
980 * This function makes sure dead exports are evicted in a timely manner.
981 * This function is only called when some export receives a message (i.e.,
982 * the network is up.)
984 static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
986 struct obd_export *oldest_exp;
987 time_t oldest_time, new_time;
993 /* Compensate for slow machines, etc, by faking our request time
994 into the future. Although this can break the strict time-ordering
995 of the list, we can be really lazy here - we don't have to evict
996 at the exact right moment. Eventually, all silent exports
997 will make it to the top of the list. */
999 /* Do not pay attention on 1sec or smaller renewals. */
1000 new_time = cfs_time_current_sec() + extra_delay;
1001 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
1004 exp->exp_last_request_time = new_time;
1005 CDEBUG(D_HA, "updating export %s at "CFS_TIME_T" exp %p\n",
1006 exp->exp_client_uuid.uuid,
1007 exp->exp_last_request_time, exp);
1009 /* exports may get disconnected from the chain even though the
1010 export has references, so we must keep the spin lock while
1011 manipulating the lists */
1012 cfs_spin_lock(&exp->exp_obd->obd_dev_lock);
1014 if (cfs_list_empty(&exp->exp_obd_chain_timed)) {
1015 /* this one is not timed */
1016 cfs_spin_unlock(&exp->exp_obd->obd_dev_lock);
1020 cfs_list_move_tail(&exp->exp_obd_chain_timed,
1021 &exp->exp_obd->obd_exports_timed);
1023 oldest_exp = cfs_list_entry(exp->exp_obd->obd_exports_timed.next,
1024 struct obd_export, exp_obd_chain_timed);
1025 oldest_time = oldest_exp->exp_last_request_time;
1026 cfs_spin_unlock(&exp->exp_obd->obd_dev_lock);
1028 if (exp->exp_obd->obd_recovering) {
1029 /* be nice to everyone during recovery */
1034 /* Note - racing to start/reset the obd_eviction timer is safe */
1035 if (exp->exp_obd->obd_eviction_timer == 0) {
1036 /* Check if the oldest entry is expired. */
1037 if (cfs_time_current_sec() > (oldest_time + PING_EVICT_TIMEOUT +
1039 /* We need a second timer, in case the net was down and
1040 * it just came back. Since the pinger may skip every
1041 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
1042 * we better wait for 3. */
1043 exp->exp_obd->obd_eviction_timer =
1044 cfs_time_current_sec() + 3 * PING_INTERVAL;
1045 CDEBUG(D_HA, "%s: Think about evicting %s from "CFS_TIME_T"\n",
1046 exp->exp_obd->obd_name,
1047 obd_export_nid2str(oldest_exp), oldest_time);
1050 if (cfs_time_current_sec() >
1051 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
1052 /* The evictor won't evict anyone who we've heard from
1053 * recently, so we don't have to check before we start
1055 if (!ping_evictor_wake(exp))
1056 exp->exp_obd->obd_eviction_timer = 0;
1064 * Sanity check request \a req.
1065 * Return 0 if all is ok, error code otherwise.
1067 static int ptlrpc_check_req(struct ptlrpc_request *req)
1071 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
1072 req->rq_export->exp_conn_cnt)) {
1073 DEBUG_REQ(D_ERROR, req,
1074 "DROPPING req from old connection %d < %d",
1075 lustre_msg_get_conn_cnt(req->rq_reqmsg),
1076 req->rq_export->exp_conn_cnt);
1079 if (unlikely(req->rq_export->exp_obd &&
1080 req->rq_export->exp_obd->obd_fail)) {
1081 /* Failing over, don't handle any more reqs, send
1082 error response instead. */
1083 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
1084 req, req->rq_export->exp_obd->obd_name);
1086 } else if (lustre_msg_get_flags(req->rq_reqmsg) &
1087 (MSG_REPLAY | MSG_REQ_REPLAY_DONE) &&
1088 !(req->rq_export->exp_obd->obd_recovering)) {
1089 DEBUG_REQ(D_ERROR, req,
1090 "Invalid replay without recovery");
1091 class_fail_export(req->rq_export);
1093 } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 &&
1094 !(req->rq_export->exp_obd->obd_recovering)) {
1095 DEBUG_REQ(D_ERROR, req, "Invalid req with transno "
1096 LPU64" without recovery",
1097 lustre_msg_get_transno(req->rq_reqmsg));
1098 class_fail_export(req->rq_export);
1102 if (unlikely(rc < 0)) {
1103 req->rq_status = rc;
1109 static void ptlrpc_at_set_timer(struct ptlrpc_service_part *svcpt)
1111 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1114 if (array->paa_count == 0) {
1115 cfs_timer_disarm(&svcpt->scp_at_timer);
1119 /* Set timer for closest deadline */
1120 next = (__s32)(array->paa_deadline - cfs_time_current_sec() -
1123 ptlrpc_at_timer((unsigned long)svcpt);
1125 cfs_timer_arm(&svcpt->scp_at_timer, cfs_time_shift(next));
1126 CDEBUG(D_INFO, "armed %s at %+ds\n",
1127 svcpt->scp_service->srv_name, next);
1131 /* Add rpc to early reply check list */
1132 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
1134 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1135 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1136 struct ptlrpc_request *rq = NULL;
1142 if (req->rq_no_reply)
1145 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
1148 cfs_spin_lock(&svcpt->scp_at_lock);
1149 LASSERT(cfs_list_empty(&req->rq_timed_list));
1151 index = (unsigned long)req->rq_deadline % array->paa_size;
1152 if (array->paa_reqs_count[index] > 0) {
1153 /* latest rpcs will have the latest deadlines in the list,
1154 * so search backward. */
1155 cfs_list_for_each_entry_reverse(rq,
1156 &array->paa_reqs_array[index],
1158 if (req->rq_deadline >= rq->rq_deadline) {
1159 cfs_list_add(&req->rq_timed_list,
1160 &rq->rq_timed_list);
1166 /* Add the request at the head of the list */
1167 if (cfs_list_empty(&req->rq_timed_list))
1168 cfs_list_add(&req->rq_timed_list,
1169 &array->paa_reqs_array[index]);
1171 cfs_spin_lock(&req->rq_lock);
1172 req->rq_at_linked = 1;
1173 cfs_spin_unlock(&req->rq_lock);
1174 req->rq_at_index = index;
1175 array->paa_reqs_count[index]++;
1177 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
1178 array->paa_deadline = req->rq_deadline;
1179 ptlrpc_at_set_timer(svcpt);
1181 cfs_spin_unlock(&svcpt->scp_at_lock);
1186 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
1188 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1189 struct ptlrpc_request *reqcopy;
1190 struct lustre_msg *reqmsg;
1191 cfs_duration_t olddl = req->rq_deadline - cfs_time_current_sec();
1196 /* deadline is when the client expects us to reply, margin is the
1197 difference between clients' and servers' expectations */
1198 DEBUG_REQ(D_ADAPTTO, req,
1199 "%ssending early reply (deadline %+lds, margin %+lds) for "
1200 "%d+%d", AT_OFF ? "AT off - not " : "",
1201 olddl, olddl - at_get(&svcpt->scp_at_estimate),
1202 at_get(&svcpt->scp_at_estimate), at_extra);
1208 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), "
1209 "not sending early reply. Consider increasing "
1210 "at_early_margin (%d)?", olddl, at_early_margin);
1212 /* Return an error so we're not re-added to the timed list. */
1216 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
1217 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
1218 "but no AT support");
1222 if (req->rq_export &&
1223 lustre_msg_get_flags(req->rq_reqmsg) &
1224 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1225 /* During recovery, we don't want to send too many early
1226 * replies, but on the other hand we want to make sure the
1227 * client has enough time to resend if the rpc is lost. So
1228 * during the recovery period send at least 4 early replies,
1229 * spacing them every at_extra if we can. at_estimate should
1230 * always equal this fixed value during recovery. */
1231 at_measured(&svcpt->scp_at_estimate, min(at_extra,
1232 req->rq_export->exp_obd->obd_recovery_timeout / 4));
1234 /* Fake our processing time into the future to ask the clients
1235 * for some extra amount of time */
1236 at_measured(&svcpt->scp_at_estimate, at_extra +
1237 cfs_time_current_sec() -
1238 req->rq_arrival_time.tv_sec);
1240 /* Check to see if we've actually increased the deadline -
1241 * we may be past adaptive_max */
1242 if (req->rq_deadline >= req->rq_arrival_time.tv_sec +
1243 at_get(&svcpt->scp_at_estimate)) {
1244 DEBUG_REQ(D_WARNING, req, "Couldn't add any time "
1245 "(%ld/%ld), not sending early reply\n",
1246 olddl, req->rq_arrival_time.tv_sec +
1247 at_get(&svcpt->scp_at_estimate) -
1248 cfs_time_current_sec());
1252 newdl = cfs_time_current_sec() + at_get(&svcpt->scp_at_estimate);
1254 OBD_ALLOC(reqcopy, sizeof *reqcopy);
1255 if (reqcopy == NULL)
1257 OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1259 OBD_FREE(reqcopy, sizeof *reqcopy);
1264 reqcopy->rq_reply_state = NULL;
1265 reqcopy->rq_rep_swab_mask = 0;
1266 reqcopy->rq_pack_bulk = 0;
1267 reqcopy->rq_pack_udesc = 0;
1268 reqcopy->rq_packed_final = 0;
1269 sptlrpc_svc_ctx_addref(reqcopy);
1270 /* We only need the reqmsg for the magic */
1271 reqcopy->rq_reqmsg = reqmsg;
1272 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1274 LASSERT(cfs_atomic_read(&req->rq_refcount));
1275 /** if it is last refcount then early reply isn't needed */
1276 if (cfs_atomic_read(&req->rq_refcount) == 1) {
1277 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
1278 "abort sending early reply\n");
1279 GOTO(out, rc = -EINVAL);
1282 /* Connection ref */
1283 reqcopy->rq_export = class_conn2export(
1284 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1285 if (reqcopy->rq_export == NULL)
1286 GOTO(out, rc = -ENODEV);
1289 class_export_rpc_get(reqcopy->rq_export);
1290 if (reqcopy->rq_export->exp_obd &&
1291 reqcopy->rq_export->exp_obd->obd_fail)
1292 GOTO(out_put, rc = -ENODEV);
1294 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1298 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1301 /* Adjust our own deadline to what we told the client */
1302 req->rq_deadline = newdl;
1303 req->rq_early_count++; /* number sent, server side */
1305 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1308 /* Free the (early) reply state from lustre_pack_reply.
1309 (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
1310 ptlrpc_req_drop_rs(reqcopy);
1313 class_export_rpc_put(reqcopy->rq_export);
1314 class_export_put(reqcopy->rq_export);
1316 sptlrpc_svc_ctx_decref(reqcopy);
1317 OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1318 OBD_FREE(reqcopy, sizeof *reqcopy);
1322 /* Send early replies to everybody expiring within at_early_margin
1323 asking for at_extra time */
1324 static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1326 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1327 struct ptlrpc_request *rq, *n;
1328 cfs_list_t work_list;
1331 time_t now = cfs_time_current_sec();
1332 cfs_duration_t delay;
1333 int first, counter = 0;
1336 cfs_spin_lock(&svcpt->scp_at_lock);
1337 if (svcpt->scp_at_check == 0) {
1338 cfs_spin_unlock(&svcpt->scp_at_lock);
1341 delay = cfs_time_sub(cfs_time_current(), svcpt->scp_at_checktime);
1342 svcpt->scp_at_check = 0;
1344 if (array->paa_count == 0) {
1345 cfs_spin_unlock(&svcpt->scp_at_lock);
1349 /* The timer went off, but maybe the nearest rpc already completed. */
1350 first = array->paa_deadline - now;
1351 if (first > at_early_margin) {
1352 /* We've still got plenty of time. Reset the timer. */
1353 ptlrpc_at_set_timer(svcpt);
1354 cfs_spin_unlock(&svcpt->scp_at_lock);
1358 /* We're close to a timeout, and we don't know how much longer the
1359 server will take. Send early replies to everyone expiring soon. */
1360 CFS_INIT_LIST_HEAD(&work_list);
1362 index = (unsigned long)array->paa_deadline % array->paa_size;
1363 count = array->paa_count;
1365 count -= array->paa_reqs_count[index];
1366 cfs_list_for_each_entry_safe(rq, n,
1367 &array->paa_reqs_array[index],
1369 if (rq->rq_deadline <= now + at_early_margin) {
1370 cfs_list_del_init(&rq->rq_timed_list);
1372 * ptlrpc_server_drop_request() may drop
1373 * refcount to 0 already. Let's check this and
1374 * don't add entry to work_list
1376 if (likely(cfs_atomic_inc_not_zero(&rq->rq_refcount)))
1377 cfs_list_add(&rq->rq_timed_list, &work_list);
1379 array->paa_reqs_count[index]--;
1381 cfs_spin_lock(&rq->rq_lock);
1382 rq->rq_at_linked = 0;
1383 cfs_spin_unlock(&rq->rq_lock);
1387 /* update the earliest deadline */
1388 if (deadline == -1 || rq->rq_deadline < deadline)
1389 deadline = rq->rq_deadline;
1394 if (++index >= array->paa_size)
1397 array->paa_deadline = deadline;
1398 /* we have a new earliest deadline, restart the timer */
1399 ptlrpc_at_set_timer(svcpt);
1401 cfs_spin_unlock(&svcpt->scp_at_lock);
1403 CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
1404 "replies\n", first, at_extra, counter);
1406 /* We're already past request deadlines before we even get a
1407 chance to send early replies */
1408 LCONSOLE_WARN("%s: This server is not able to keep up with "
1409 "request traffic (cpu-bound).\n",
1410 svcpt->scp_service->srv_name);
1411 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, "
1412 "delay="CFS_DURATION_T"(jiff)\n",
1413 counter, svcpt->scp_nreqs_incoming,
1414 svcpt->scp_nreqs_active,
1415 at_get(&svcpt->scp_at_estimate), delay);
1418 /* we took additional refcount so entries can't be deleted from list, no
1419 * locking is needed */
1420 while (!cfs_list_empty(&work_list)) {
1421 rq = cfs_list_entry(work_list.next, struct ptlrpc_request,
1423 cfs_list_del_init(&rq->rq_timed_list);
1425 if (ptlrpc_at_send_early_reply(rq) == 0)
1426 ptlrpc_at_add_timed(rq);
1428 ptlrpc_server_drop_request(rq);
1431 RETURN(1); /* return "did_something" for liblustre */
1435 * Put the request to the export list if the request may become
1436 * a high priority one.
1438 static int ptlrpc_hpreq_init(struct ptlrpc_service *svc,
1439 struct ptlrpc_request *req)
1444 if (svc->srv_ops.so_hpreq_handler) {
1445 rc = svc->srv_ops.so_hpreq_handler(req);
1449 if (req->rq_export && req->rq_ops) {
1450 /* Perform request specific check. We should do this check
1451 * before the request is added into exp_hp_rpcs list otherwise
1452 * it may hit swab race at LU-1044. */
1453 if (req->rq_ops->hpreq_check)
1454 rc = req->rq_ops->hpreq_check(req);
1456 cfs_spin_lock_bh(&req->rq_export->exp_rpc_lock);
1457 cfs_list_add(&req->rq_exp_list,
1458 &req->rq_export->exp_hp_rpcs);
1459 cfs_spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1465 /** Remove the request from the export list. */
1466 static void ptlrpc_hpreq_fini(struct ptlrpc_request *req)
1469 if (req->rq_export && req->rq_ops) {
1470 /* refresh lock timeout again so that client has more
1471 * room to send lock cancel RPC. */
1472 if (req->rq_ops->hpreq_fini)
1473 req->rq_ops->hpreq_fini(req);
1475 cfs_spin_lock_bh(&req->rq_export->exp_rpc_lock);
1476 cfs_list_del_init(&req->rq_exp_list);
1477 cfs_spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1482 static int ptlrpc_hpreq_check(struct ptlrpc_request *req)
1487 static struct ptlrpc_hpreq_ops ptlrpc_hpreq_common = {
1488 .hpreq_lock_match = NULL,
1489 .hpreq_check = ptlrpc_hpreq_check,
1493 /* Hi-Priority RPC check by RPC operation code. */
1494 int ptlrpc_hpreq_handler(struct ptlrpc_request *req)
1496 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1498 /* Check for export to let only reconnects for not yet evicted
1499 * export to become a HP rpc. */
1500 if ((req->rq_export != NULL) &&
1501 (opc == OBD_PING || opc == MDS_CONNECT || opc == OST_CONNECT))
1502 req->rq_ops = &ptlrpc_hpreq_common;
1506 EXPORT_SYMBOL(ptlrpc_hpreq_handler);
1509 * Make the request a high priority one.
1511 * All the high priority requests are queued in a separate FIFO
1512 * ptlrpc_service_part::scp_hpreq_pending list which is parallel to
1513 * ptlrpc_service_part::scp_req_pending list but has a higher priority
1516 * \see ptlrpc_server_handle_request().
1518 static void ptlrpc_hpreq_reorder_nolock(struct ptlrpc_service_part *svcpt,
1519 struct ptlrpc_request *req)
1523 cfs_spin_lock(&req->rq_lock);
1524 if (req->rq_hp == 0) {
1525 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1527 /* Add to the high priority queue. */
1528 cfs_list_move_tail(&req->rq_list, &svcpt->scp_hreq_pending);
1530 if (opc != OBD_PING)
1531 DEBUG_REQ(D_RPCTRACE, req, "high priority req");
1533 cfs_spin_unlock(&req->rq_lock);
1538 * \see ptlrpc_hpreq_reorder_nolock
1540 void ptlrpc_hpreq_reorder(struct ptlrpc_request *req)
1542 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1545 cfs_spin_lock(&svcpt->scp_req_lock);
1546 /* It may happen that the request is already taken for the processing
1547 * but still in the export list, or the request is not in the request
1548 * queue but in the export list already, do not add it into the
1550 if (!cfs_list_empty(&req->rq_list))
1551 ptlrpc_hpreq_reorder_nolock(svcpt, req);
1552 cfs_spin_unlock(&svcpt->scp_req_lock);
1555 EXPORT_SYMBOL(ptlrpc_hpreq_reorder);
1557 /** Check if the request is a high priority one. */
1558 static int ptlrpc_server_hpreq_check(struct ptlrpc_service *svc,
1559 struct ptlrpc_request *req)
1561 return ptlrpc_hpreq_init(svc, req);
1564 /** Check if a request is a high priority one. */
1565 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1566 struct ptlrpc_request *req)
1571 rc = ptlrpc_server_hpreq_check(svcpt->scp_service, req);
1575 cfs_spin_lock(&svcpt->scp_req_lock);
1578 ptlrpc_hpreq_reorder_nolock(svcpt, req);
1580 cfs_list_add_tail(&req->rq_list, &svcpt->scp_req_pending);
1582 cfs_spin_unlock(&svcpt->scp_req_lock);
1588 * Allow to handle high priority request
1589 * User can call it w/o any lock but need to hold
1590 * ptlrpc_service_part::scp_req_lock to get reliable result
1592 static int ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1598 if (svcpt->scp_nreqs_active >= svcpt->scp_nthrs_running - 1)
1601 return cfs_list_empty(&svcpt->scp_req_pending) ||
1602 svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1605 static int ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1608 return ptlrpc_server_allow_high(svcpt, force) &&
1609 !cfs_list_empty(&svcpt->scp_hreq_pending);
1613 * Only allow normal priority requests on a service that has a high-priority
1614 * queue if forced (i.e. cleanup), if there are other high priority requests
1615 * already being processed (i.e. those threads can service more high-priority
1616 * requests), or if there are enough idle threads that a later thread can do
1617 * a high priority request.
1618 * User can call it w/o any lock but need to hold
1619 * ptlrpc_service_part::scp_req_lock to get reliable result
1621 static int ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1625 if (1) /* always allow to handle normal request for liblustre */
1629 svcpt->scp_nreqs_active < svcpt->scp_nthrs_running - 2)
1632 if (svcpt->scp_nreqs_active >= svcpt->scp_nthrs_running - 1)
1635 return svcpt->scp_nhreqs_active > 0 ||
1636 svcpt->scp_service->srv_ops.so_hpreq_handler == NULL;
1639 static int ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1642 return ptlrpc_server_allow_normal(svcpt, force) &&
1643 !cfs_list_empty(&svcpt->scp_req_pending);
1647 * Returns true if there are requests available in incoming
1648 * request queue for processing and it is allowed to fetch them.
1649 * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1650 * to get reliable result
1651 * \see ptlrpc_server_allow_normal
1652 * \see ptlrpc_server_allow high
1655 ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt, int force)
1657 return ptlrpc_server_high_pending(svcpt, force) ||
1658 ptlrpc_server_normal_pending(svcpt, force);
1662 * Fetch a request for processing from queue of unprocessed requests.
1663 * Favors high-priority requests.
1664 * Returns a pointer to fetched request.
1666 static struct ptlrpc_request *
1667 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, int force)
1669 struct ptlrpc_request *req;
1672 if (ptlrpc_server_high_pending(svcpt, force)) {
1673 req = cfs_list_entry(svcpt->scp_hreq_pending.next,
1674 struct ptlrpc_request, rq_list);
1675 svcpt->scp_hreq_count++;
1679 if (ptlrpc_server_normal_pending(svcpt, force)) {
1680 req = cfs_list_entry(svcpt->scp_req_pending.next,
1681 struct ptlrpc_request, rq_list);
1682 svcpt->scp_hreq_count = 0;
1689 * Handle freshly incoming reqs, add to timed early reply list,
1690 * pass on to regular request queue.
1691 * All incoming requests pass through here before getting into
1692 * ptlrpc_server_handle_req later on.
1695 ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt)
1697 struct ptlrpc_service *svc = svcpt->scp_service;
1698 struct ptlrpc_request *req;
1703 cfs_spin_lock(&svcpt->scp_lock);
1704 if (cfs_list_empty(&svcpt->scp_req_incoming)) {
1705 cfs_spin_unlock(&svcpt->scp_lock);
1709 req = cfs_list_entry(svcpt->scp_req_incoming.next,
1710 struct ptlrpc_request, rq_list);
1711 cfs_list_del_init(&req->rq_list);
1712 svcpt->scp_nreqs_incoming--;
1713 /* Consider this still a "queued" request as far as stats are
1715 cfs_spin_unlock(&svcpt->scp_lock);
1717 /* go through security check/transform */
1718 rc = sptlrpc_svc_unwrap_request(req);
1722 case SECSVC_COMPLETE:
1723 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1732 * for null-flavored rpc, msg has been unpacked by sptlrpc, although
1733 * redo it wouldn't be harmful.
1735 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
1736 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
1738 CERROR("error unpacking request: ptl %d from %s "
1739 "x"LPU64"\n", svc->srv_req_portal,
1740 libcfs_id2str(req->rq_peer), req->rq_xid);
1745 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1747 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
1748 LPU64"\n", svc->srv_req_portal,
1749 libcfs_id2str(req->rq_peer), req->rq_xid);
1753 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
1754 lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
1755 CERROR("drop incoming rpc opc %u, x"LPU64"\n",
1756 cfs_fail_val, req->rq_xid);
1761 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1762 CERROR("wrong packet type received (type=%u) from %s\n",
1763 lustre_msg_get_type(req->rq_reqmsg),
1764 libcfs_id2str(req->rq_peer));
1768 switch(lustre_msg_get_opc(req->rq_reqmsg)) {
1771 req->rq_bulk_write = 1;
1775 case MGS_CONFIG_READ:
1776 req->rq_bulk_read = 1;
1780 CDEBUG(D_RPCTRACE, "got req x"LPU64"\n", req->rq_xid);
1782 req->rq_export = class_conn2export(
1783 lustre_msg_get_handle(req->rq_reqmsg));
1784 if (req->rq_export) {
1785 rc = ptlrpc_check_req(req);
1787 rc = sptlrpc_target_export_check(req->rq_export, req);
1789 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
1790 "illegal security flavor,");
1795 ptlrpc_update_export_timer(req->rq_export, 0);
1798 /* req_in handling should/must be fast */
1799 if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5)
1800 DEBUG_REQ(D_WARNING, req, "Slow req_in handling "CFS_DURATION_T"s",
1801 cfs_time_sub(cfs_time_current_sec(),
1802 req->rq_arrival_time.tv_sec));
1804 /* Set rpc server deadline and add it to the timed list */
1805 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1806 MSGHDR_AT_SUPPORT) ?
1807 /* The max time the client expects us to take */
1808 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1809 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1810 if (unlikely(deadline == 0)) {
1811 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1815 ptlrpc_at_add_timed(req);
1817 /* Move it over to the request processing queue */
1818 rc = ptlrpc_server_request_add(svcpt, req);
1820 ptlrpc_hpreq_fini(req);
1823 cfs_waitq_signal(&svcpt->scp_waitq);
1827 cfs_spin_lock(&svcpt->scp_req_lock);
1828 svcpt->scp_nreqs_active++;
1829 cfs_spin_unlock(&svcpt->scp_req_lock);
1830 ptlrpc_server_finish_request(svcpt, req);
1836 * Main incoming request handling logic.
1837 * Calls handler function from service to do actual processing.
1840 ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
1841 struct ptlrpc_thread *thread)
1843 struct ptlrpc_service *svc = svcpt->scp_service;
1844 struct obd_export *export = NULL;
1845 struct ptlrpc_request *request;
1846 struct timeval work_start;
1847 struct timeval work_end;
1853 cfs_spin_lock(&svcpt->scp_req_lock);
1855 /* !@%$# liblustre only has 1 thread */
1856 if (cfs_atomic_read(&svcpt->scp_nreps_difficult) != 0) {
1857 cfs_spin_unlock(&svcpt->scp_req_lock);
1861 request = ptlrpc_server_request_get(svcpt, 0);
1862 if (request == NULL) {
1863 cfs_spin_unlock(&svcpt->scp_req_lock);
1867 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
1868 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
1869 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
1870 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
1872 if (unlikely(fail_opc)) {
1873 if (request->rq_export && request->rq_ops) {
1874 cfs_spin_unlock(&svcpt->scp_req_lock);
1876 OBD_FAIL_TIMEOUT(fail_opc, 4);
1878 cfs_spin_lock(&svcpt->scp_req_lock);
1879 request = ptlrpc_server_request_get(svcpt, 0);
1880 if (request == NULL) {
1881 cfs_spin_unlock(&svcpt->scp_req_lock);
1887 cfs_list_del_init(&request->rq_list);
1888 svcpt->scp_nreqs_active++;
1890 svcpt->scp_nhreqs_active++;
1892 cfs_spin_unlock(&svcpt->scp_req_lock);
1894 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
1896 if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
1897 libcfs_debug_dumplog();
1899 cfs_gettimeofday(&work_start);
1900 timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
1901 if (likely(svc->srv_stats != NULL)) {
1902 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
1904 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
1905 svcpt->scp_nreqs_incoming);
1906 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
1907 svcpt->scp_nreqs_active);
1908 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
1909 at_get(&svcpt->scp_at_estimate));
1912 rc = lu_context_init(&request->rq_session, LCT_SESSION | LCT_NOREF);
1914 CERROR("Failure to initialize session: %d\n", rc);
1917 request->rq_session.lc_thread = thread;
1918 request->rq_session.lc_cookie = 0x5;
1919 lu_context_enter(&request->rq_session);
1921 CDEBUG(D_NET, "got req "LPU64"\n", request->rq_xid);
1923 request->rq_svc_thread = thread;
1925 request->rq_svc_thread->t_env->le_ses = &request->rq_session;
1927 if (likely(request->rq_export)) {
1928 if (unlikely(ptlrpc_check_req(request)))
1930 ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
1931 export = class_export_rpc_get(request->rq_export);
1934 /* Discard requests queued for longer than the deadline.
1935 The deadline is increased if we send an early reply. */
1936 if (cfs_time_current_sec() > request->rq_deadline) {
1937 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s"
1938 ": deadline "CFS_DURATION_T":"CFS_DURATION_T"s ago\n",
1939 libcfs_id2str(request->rq_peer),
1940 cfs_time_sub(request->rq_deadline,
1941 request->rq_arrival_time.tv_sec),
1942 cfs_time_sub(cfs_time_current_sec(),
1943 request->rq_deadline));
1944 goto put_rpc_export;
1947 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
1948 "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1949 (request->rq_export ?
1950 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1951 (request->rq_export ?
1952 cfs_atomic_read(&request->rq_export->exp_refcount) : -99),
1953 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1954 libcfs_id2str(request->rq_peer),
1955 lustre_msg_get_opc(request->rq_reqmsg));
1957 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
1958 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
1960 rc = svc->srv_ops.so_req_handler(request);
1962 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
1966 class_export_rpc_put(export);
1968 lu_context_exit(&request->rq_session);
1969 lu_context_fini(&request->rq_session);
1971 if (unlikely(cfs_time_current_sec() > request->rq_deadline)) {
1972 DEBUG_REQ(D_WARNING, request, "Request x"LPU64" took longer "
1973 "than estimated ("CFS_DURATION_T":"CFS_DURATION_T"s);"
1974 " client may timeout.",
1975 request->rq_xid, cfs_time_sub(request->rq_deadline,
1976 request->rq_arrival_time.tv_sec),
1977 cfs_time_sub(cfs_time_current_sec(),
1978 request->rq_deadline));
1981 cfs_gettimeofday(&work_end);
1982 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1983 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
1984 "%s:%s+%d:%d:x"LPU64":%s:%d Request procesed in "
1985 "%ldus (%ldus total) trans "LPU64" rc %d/%d\n",
1987 (request->rq_export ?
1988 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1989 (request->rq_export ?
1990 cfs_atomic_read(&request->rq_export->exp_refcount) : -99),
1991 lustre_msg_get_status(request->rq_reqmsg),
1993 libcfs_id2str(request->rq_peer),
1994 lustre_msg_get_opc(request->rq_reqmsg),
1996 cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL),
1997 (request->rq_repmsg ?
1998 lustre_msg_get_transno(request->rq_repmsg) :
1999 request->rq_transno),
2001 (request->rq_repmsg ?
2002 lustre_msg_get_status(request->rq_repmsg) : -999));
2003 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
2004 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
2005 int opc = opcode_offset(op);
2006 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
2007 LASSERT(opc < LUSTRE_MAX_OPCODES);
2008 lprocfs_counter_add(svc->srv_stats,
2009 opc + EXTRA_MAX_OPCODES,
2013 if (unlikely(request->rq_early_count)) {
2014 DEBUG_REQ(D_ADAPTTO, request,
2015 "sent %d early replies before finishing in "
2017 request->rq_early_count,
2018 cfs_time_sub(work_end.tv_sec,
2019 request->rq_arrival_time.tv_sec));
2023 ptlrpc_server_finish_request(svcpt, request);
2029 * An internal function to process a single reply state object.
2032 ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
2034 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
2035 struct ptlrpc_service *svc = svcpt->scp_service;
2036 struct obd_export *exp;
2041 exp = rs->rs_export;
2043 LASSERT (rs->rs_difficult);
2044 LASSERT (rs->rs_scheduled);
2045 LASSERT (cfs_list_empty(&rs->rs_list));
2047 cfs_spin_lock (&exp->exp_lock);
2048 /* Noop if removed already */
2049 cfs_list_del_init (&rs->rs_exp_list);
2050 cfs_spin_unlock (&exp->exp_lock);
2052 /* The disk commit callback holds exp_uncommitted_replies_lock while it
2053 * iterates over newly committed replies, removing them from
2054 * exp_uncommitted_replies. It then drops this lock and schedules the
2055 * replies it found for handling here.
2057 * We can avoid contention for exp_uncommitted_replies_lock between the
2058 * HRT threads and further commit callbacks by checking rs_committed
2059 * which is set in the commit callback while it holds both
2060 * rs_lock and exp_uncommitted_reples.
2062 * If we see rs_committed clear, the commit callback _may_ not have
2063 * handled this reply yet and we race with it to grab
2064 * exp_uncommitted_replies_lock before removing the reply from
2065 * exp_uncommitted_replies. Note that if we lose the race and the
2066 * reply has already been removed, list_del_init() is a noop.
2068 * If we see rs_committed set, we know the commit callback is handling,
2069 * or has handled this reply since store reordering might allow us to
2070 * see rs_committed set out of sequence. But since this is done
2071 * holding rs_lock, we can be sure it has all completed once we hold
2072 * rs_lock, which we do right next.
2074 if (!rs->rs_committed) {
2075 cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
2076 cfs_list_del_init(&rs->rs_obd_list);
2077 cfs_spin_unlock(&exp->exp_uncommitted_replies_lock);
2080 cfs_spin_lock(&rs->rs_lock);
2082 been_handled = rs->rs_handled;
2085 nlocks = rs->rs_nlocks; /* atomic "steal", but */
2086 rs->rs_nlocks = 0; /* locks still on rs_locks! */
2088 if (nlocks == 0 && !been_handled) {
2089 /* If we see this, we should already have seen the warning
2090 * in mds_steal_ack_locks() */
2091 CWARN("All locks stolen from rs %p x"LPD64".t"LPD64
2094 rs->rs_xid, rs->rs_transno, rs->rs_opc,
2095 libcfs_nid2str(exp->exp_connection->c_peer.nid));
2098 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
2099 cfs_spin_unlock(&rs->rs_lock);
2101 if (!been_handled && rs->rs_on_net) {
2102 LNetMDUnlink(rs->rs_md_h);
2103 /* Ignore return code; we're racing with
2107 while (nlocks-- > 0)
2108 ldlm_lock_decref(&rs->rs_locks[nlocks],
2109 rs->rs_modes[nlocks]);
2111 cfs_spin_lock(&rs->rs_lock);
2114 rs->rs_scheduled = 0;
2116 if (!rs->rs_on_net) {
2118 cfs_spin_unlock(&rs->rs_lock);
2120 class_export_put (exp);
2121 rs->rs_export = NULL;
2122 ptlrpc_rs_decref (rs);
2123 if (cfs_atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
2124 svc->srv_is_stopping)
2125 cfs_waitq_broadcast(&svcpt->scp_waitq);
2129 /* still on the net; callback will schedule */
2130 cfs_spin_unlock(&rs->rs_lock);
2137 * Check whether given service has a reply available for processing
2140 * \param svc a ptlrpc service
2141 * \retval 0 no replies processed
2142 * \retval 1 one reply processed
2145 ptlrpc_server_handle_reply(struct ptlrpc_service_part *svcpt)
2147 struct ptlrpc_reply_state *rs = NULL;
2150 cfs_spin_lock(&svcpt->scp_rep_lock);
2151 if (!cfs_list_empty(&svcpt->scp_rep_queue)) {
2152 rs = cfs_list_entry(svcpt->scp_rep_queue.prev,
2153 struct ptlrpc_reply_state,
2155 cfs_list_del_init(&rs->rs_list);
2157 cfs_spin_unlock(&svcpt->scp_rep_lock);
2159 ptlrpc_handle_rs(rs);
2163 /* FIXME make use of timeout later */
2165 liblustre_check_services (void *arg)
2167 int did_something = 0;
2169 cfs_list_t *tmp, *nxt;
2172 /* I'm relying on being single threaded, not to have to lock
2173 * ptlrpc_all_services etc */
2174 cfs_list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
2175 struct ptlrpc_service *svc =
2176 cfs_list_entry (tmp, struct ptlrpc_service, srv_list);
2177 struct ptlrpc_service_part *svcpt;
2179 LASSERT(svc->srv_ncpts == 1);
2180 svcpt = svc->srv_parts[0];
2182 if (svcpt->scp_nthrs_running != 0) /* I've recursed */
2185 /* service threads can block for bulk, so this limits us
2186 * (arbitrarily) to recursing 1 stack frame per service.
2187 * Note that the problem with recursion is that we have to
2188 * unwind completely before our caller can resume. */
2190 svcpt->scp_nthrs_running++;
2193 rc = ptlrpc_server_handle_req_in(svcpt);
2194 rc |= ptlrpc_server_handle_reply(svcpt);
2195 rc |= ptlrpc_at_check_timed(svcpt);
2196 rc |= ptlrpc_server_handle_request(svcpt, NULL);
2197 rc |= (ptlrpc_server_post_idle_rqbds(svcpt) > 0);
2198 did_something |= rc;
2201 svcpt->scp_nthrs_running--;
2204 RETURN(did_something);
2206 #define ptlrpc_stop_all_threads(s) do {} while (0)
2208 #else /* __KERNEL__ */
2211 ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
2213 int avail = svcpt->scp_nrqbds_posted;
2214 int low_water = test_req_buffer_pressure ? 0 :
2215 svcpt->scp_service->srv_nbuf_per_group / 2;
2217 /* NB I'm not locking; just looking. */
2219 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
2220 * allowed the request history to grow out of control. We could put a
2221 * sanity check on that here and cull some history if we need the
2224 if (avail <= low_water)
2225 ptlrpc_grow_req_bufs(svcpt, 1);
2227 if (svcpt->scp_service->srv_stats) {
2228 lprocfs_counter_add(svcpt->scp_service->srv_stats,
2229 PTLRPC_REQBUF_AVAIL_CNTR, avail);
2234 ptlrpc_retry_rqbds(void *arg)
2236 struct ptlrpc_service_part *svcpt = (struct ptlrpc_service_part *)arg;
2238 svcpt->scp_rqbd_timeout = 0;
2243 ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
2245 return svcpt->scp_nreqs_active <
2246 svcpt->scp_nthrs_running - 1 -
2247 (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL);
2251 * allowed to create more threads
2252 * user can call it w/o any lock but need to hold
2253 * ptlrpc_service_part::scp_lock to get reliable result
2256 ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
2258 return svcpt->scp_nthrs_running +
2259 svcpt->scp_nthrs_starting <
2260 svcpt->scp_service->srv_nthrs_cpt_limit;
2264 * too many requests and allowed to create more threads
2267 ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
2269 return !ptlrpc_threads_enough(svcpt) &&
2270 ptlrpc_threads_increasable(svcpt);
2274 ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2276 return thread_is_stopping(thread) ||
2277 thread->t_svcpt->scp_service->srv_is_stopping;
2281 ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
2283 return !cfs_list_empty(&svcpt->scp_rqbd_idle) &&
2284 svcpt->scp_rqbd_timeout == 0;
2288 ptlrpc_at_check(struct ptlrpc_service_part *svcpt)
2290 return svcpt->scp_at_check;
2294 * requests wait on preprocessing
2295 * user can call it w/o any lock but need to hold
2296 * ptlrpc_service_part::scp_lock to get reliable result
2299 ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
2301 return !cfs_list_empty(&svcpt->scp_req_incoming);
2304 static __attribute__((__noinline__)) int
2305 ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
2306 struct ptlrpc_thread *thread)
2308 /* Don't exit while there are replies to be handled */
2309 struct l_wait_info lwi = LWI_TIMEOUT(svcpt->scp_rqbd_timeout,
2310 ptlrpc_retry_rqbds, svcpt);
2312 lc_watchdog_disable(thread->t_watchdog);
2316 l_wait_event_exclusive_head(svcpt->scp_waitq,
2317 ptlrpc_thread_stopping(thread) ||
2318 ptlrpc_server_request_incoming(svcpt) ||
2319 ptlrpc_server_request_pending(svcpt, 0) ||
2320 ptlrpc_rqbd_pending(svcpt) ||
2321 ptlrpc_at_check(svcpt), &lwi);
2323 if (ptlrpc_thread_stopping(thread))
2326 lc_watchdog_touch(thread->t_watchdog,
2327 ptlrpc_server_get_timeout(svcpt));
2332 * Main thread body for service threads.
2333 * Waits in a loop waiting for new requests to process to appear.
2334 * Every time an incoming requests is added to its queue, a waitq
2335 * is woken up and one of the threads will handle it.
2337 static int ptlrpc_main(void *arg)
2339 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
2340 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2341 struct ptlrpc_service *svc = svcpt->scp_service;
2342 struct ptlrpc_reply_state *rs;
2343 #ifdef WITH_GROUP_INFO
2344 cfs_group_info_t *ginfo = NULL;
2347 int counter = 0, rc = 0;
2350 thread->t_pid = cfs_curproc_pid();
2351 cfs_daemonize_ctxt(thread->t_name);
2353 /* NB: we will call cfs_cpt_bind() for all threads, because we
2354 * might want to run lustre server only on a subset of system CPUs,
2355 * in that case ->scp_cpt is CFS_CPT_ANY */
2356 rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
2358 CWARN("%s: failed to bind %s on CPT %d\n",
2359 svc->srv_name, thread->t_name, svcpt->scp_cpt);
2362 #ifdef WITH_GROUP_INFO
2363 ginfo = cfs_groups_alloc(0);
2369 cfs_set_current_groups(ginfo);
2370 cfs_put_group_info(ginfo);
2373 if (svc->srv_ops.so_thr_init != NULL) {
2374 rc = svc->srv_ops.so_thr_init(thread);
2385 rc = lu_context_init(&env->le_ctx,
2386 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2390 thread->t_env = env;
2391 env->le_ctx.lc_thread = thread;
2392 env->le_ctx.lc_cookie = 0x6;
2394 while (!cfs_list_empty(&svcpt->scp_rqbd_idle)) {
2395 rc = ptlrpc_server_post_idle_rqbds(svcpt);
2399 CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
2400 svc->srv_name, svcpt->scp_cpt, rc);
2404 /* Alloc reply state structure for this one */
2405 OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2411 cfs_spin_lock(&svcpt->scp_lock);
2413 LASSERT(thread_is_starting(thread));
2414 thread_clear_flags(thread, SVC_STARTING);
2416 LASSERT(svcpt->scp_nthrs_starting == 1);
2417 svcpt->scp_nthrs_starting--;
2419 /* SVC_STOPPING may already be set here if someone else is trying
2420 * to stop the service while this new thread has been dynamically
2421 * forked. We still set SVC_RUNNING to let our creator know that
2422 * we are now running, however we will exit as soon as possible */
2423 thread_add_flags(thread, SVC_RUNNING);
2424 svcpt->scp_nthrs_running++;
2425 cfs_spin_unlock(&svcpt->scp_lock);
2427 /* wake up our creator in case he's still waiting. */
2428 cfs_waitq_signal(&thread->t_ctl_waitq);
2430 thread->t_watchdog = lc_watchdog_add(ptlrpc_server_get_timeout(svcpt),
2433 cfs_spin_lock(&svcpt->scp_rep_lock);
2434 cfs_list_add(&rs->rs_list, &svcpt->scp_rep_idle);
2435 cfs_waitq_signal(&svcpt->scp_rep_waitq);
2436 cfs_spin_unlock(&svcpt->scp_rep_lock);
2438 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2439 svcpt->scp_nthrs_running);
2441 /* XXX maintain a list of all managed devices: insert here */
2442 while (!ptlrpc_thread_stopping(thread)) {
2443 if (ptlrpc_wait_event(svcpt, thread))
2446 ptlrpc_check_rqbd_pool(svcpt);
2448 if (ptlrpc_threads_need_create(svcpt)) {
2449 /* Ignore return code - we tried... */
2450 ptlrpc_start_thread(svcpt, 0);
2453 /* Process all incoming reqs before handling any */
2454 if (ptlrpc_server_request_incoming(svcpt)) {
2455 ptlrpc_server_handle_req_in(svcpt);
2456 /* but limit ourselves in case of flood */
2457 if (counter++ < 100)
2462 if (ptlrpc_at_check(svcpt))
2463 ptlrpc_at_check_timed(svcpt);
2465 if (ptlrpc_server_request_pending(svcpt, 0)) {
2466 lu_context_enter(&env->le_ctx);
2467 ptlrpc_server_handle_request(svcpt, thread);
2468 lu_context_exit(&env->le_ctx);
2471 if (ptlrpc_rqbd_pending(svcpt) &&
2472 ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
2473 /* I just failed to repost request buffers.
2474 * Wait for a timeout (unless something else
2475 * happens) before I try again */
2476 svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
2477 CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
2478 svcpt->scp_nrqbds_posted);
2482 lc_watchdog_delete(thread->t_watchdog);
2483 thread->t_watchdog = NULL;
2487 * deconstruct service specific state created by ptlrpc_start_thread()
2489 if (svc->srv_ops.so_thr_done != NULL)
2490 svc->srv_ops.so_thr_done(thread);
2493 lu_context_fini(&env->le_ctx);
2497 CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n",
2498 thread, thread->t_pid, thread->t_id, rc);
2500 cfs_spin_lock(&svcpt->scp_lock);
2501 if (thread_test_and_clear_flags(thread, SVC_STARTING))
2502 svcpt->scp_nthrs_starting--;
2504 if (thread_test_and_clear_flags(thread, SVC_RUNNING)) {
2505 /* must know immediately */
2506 svcpt->scp_nthrs_running--;
2510 thread_add_flags(thread, SVC_STOPPED);
2512 cfs_waitq_signal(&thread->t_ctl_waitq);
2513 cfs_spin_unlock(&svcpt->scp_lock);
2518 static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
2519 cfs_list_t *replies)
2523 cfs_spin_lock(&hrt->hrt_lock);
2525 cfs_list_splice_init(&hrt->hrt_queue, replies);
2526 result = ptlrpc_hr.hr_stopping || !cfs_list_empty(replies);
2528 cfs_spin_unlock(&hrt->hrt_lock);
2533 * Main body of "handle reply" function.
2534 * It processes acked reply states
2536 static int ptlrpc_hr_main(void *arg)
2538 struct ptlrpc_hr_thread *hrt = (struct ptlrpc_hr_thread *)arg;
2539 struct ptlrpc_hr_partition *hrp = hrt->hrt_partition;
2540 CFS_LIST_HEAD (replies);
2541 char threadname[20];
2544 snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
2545 hrp->hrp_cpt, hrt->hrt_id);
2546 cfs_daemonize_ctxt(threadname);
2548 rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
2550 CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
2551 threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
2554 cfs_atomic_inc(&hrp->hrp_nstarted);
2555 cfs_waitq_signal(&ptlrpc_hr.hr_waitq);
2557 while (!ptlrpc_hr.hr_stopping) {
2558 l_wait_condition(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
2560 while (!cfs_list_empty(&replies)) {
2561 struct ptlrpc_reply_state *rs;
2563 rs = cfs_list_entry(replies.prev,
2564 struct ptlrpc_reply_state,
2566 cfs_list_del_init(&rs->rs_list);
2567 ptlrpc_handle_rs(rs);
2571 cfs_atomic_inc(&hrp->hrp_nstopped);
2572 cfs_waitq_signal(&ptlrpc_hr.hr_waitq);
2577 static void ptlrpc_stop_hr_threads(void)
2579 struct ptlrpc_hr_partition *hrp;
2583 ptlrpc_hr.hr_stopping = 1;
2585 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2586 if (hrp->hrp_thrs == NULL)
2587 continue; /* uninitialized */
2588 for (j = 0; j < hrp->hrp_nthrs; j++)
2589 cfs_waitq_broadcast(&hrp->hrp_thrs[j].hrt_waitq);
2592 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2593 if (hrp->hrp_thrs == NULL)
2594 continue; /* uninitialized */
2595 cfs_wait_event(ptlrpc_hr.hr_waitq,
2596 cfs_atomic_read(&hrp->hrp_nstopped) ==
2597 cfs_atomic_read(&hrp->hrp_nstarted));
2601 static int ptlrpc_start_hr_threads(void)
2603 struct ptlrpc_hr_partition *hrp;
2608 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2611 for (j = 0; j < hrp->hrp_nthrs; j++) {
2612 rc = cfs_create_thread(ptlrpc_hr_main,
2614 CLONE_VM | CLONE_FILES);
2618 cfs_wait_event(ptlrpc_hr.hr_waitq,
2619 cfs_atomic_read(&hrp->hrp_nstarted) == j);
2623 CERROR("Reply handling thread %d:%d Failed on starting: "
2624 "rc = %d\n", i, j, rc);
2625 ptlrpc_stop_hr_threads();
2631 static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
2633 struct l_wait_info lwi = { 0 };
2634 struct ptlrpc_thread *thread;
2635 CFS_LIST_HEAD (zombie);
2639 CDEBUG(D_INFO, "Stopping threads for service %s\n",
2640 svcpt->scp_service->srv_name);
2642 cfs_spin_lock(&svcpt->scp_lock);
2643 /* let the thread know that we would like it to stop asap */
2644 list_for_each_entry(thread, &svcpt->scp_threads, t_link) {
2645 CDEBUG(D_INFO, "Stopping thread %s #%u\n",
2646 svcpt->scp_service->srv_thread_name, thread->t_id);
2647 thread_add_flags(thread, SVC_STOPPING);
2650 cfs_waitq_broadcast(&svcpt->scp_waitq);
2652 while (!cfs_list_empty(&svcpt->scp_threads)) {
2653 thread = cfs_list_entry(svcpt->scp_threads.next,
2654 struct ptlrpc_thread, t_link);
2655 if (thread_is_stopped(thread)) {
2656 cfs_list_del(&thread->t_link);
2657 cfs_list_add(&thread->t_link, &zombie);
2660 cfs_spin_unlock(&svcpt->scp_lock);
2662 CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
2663 svcpt->scp_service->srv_thread_name, thread->t_id);
2664 l_wait_event(thread->t_ctl_waitq,
2665 thread_is_stopped(thread), &lwi);
2667 cfs_spin_lock(&svcpt->scp_lock);
2670 cfs_spin_unlock(&svcpt->scp_lock);
2672 while (!cfs_list_empty(&zombie)) {
2673 thread = cfs_list_entry(zombie.next,
2674 struct ptlrpc_thread, t_link);
2675 cfs_list_del(&thread->t_link);
2676 OBD_FREE_PTR(thread);
2682 * Stops all threads of a particular service \a svc
2684 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2686 struct ptlrpc_service_part *svcpt;
2690 ptlrpc_service_for_each_part(svcpt, i, svc) {
2691 if (svcpt->scp_service != NULL)
2692 ptlrpc_svcpt_stop_threads(svcpt);
2697 EXPORT_SYMBOL(ptlrpc_stop_all_threads);
2699 int ptlrpc_start_threads(struct ptlrpc_service *svc)
2706 /* We require 2 threads min, see note in ptlrpc_server_handle_request */
2707 LASSERT(svc->srv_nthrs_cpt_init >= PTLRPC_NTHRS_INIT);
2709 for (i = 0; i < svc->srv_ncpts; i++) {
2710 for (j = 0; j < svc->srv_nthrs_cpt_init; j++) {
2711 rc = ptlrpc_start_thread(svc->srv_parts[i], 1);
2717 /* We have enough threads, don't start more. b=15759 */
2724 CERROR("cannot start %s thread #%d_%d: rc %d\n",
2725 svc->srv_thread_name, i, j, rc);
2726 ptlrpc_stop_all_threads(svc);
2729 EXPORT_SYMBOL(ptlrpc_start_threads);
2731 int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
2733 struct l_wait_info lwi = { 0 };
2734 struct ptlrpc_thread *thread;
2735 struct ptlrpc_service *svc = svcpt->scp_service;
2739 LASSERT(svcpt != NULL);
2741 CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n",
2742 svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running,
2743 svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit);
2746 if (unlikely(svc->srv_is_stopping))
2749 if (!ptlrpc_threads_increasable(svcpt) ||
2750 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
2751 svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1))
2754 OBD_CPT_ALLOC_PTR(thread, svc->srv_cptable, svcpt->scp_cpt);
2757 cfs_waitq_init(&thread->t_ctl_waitq);
2759 cfs_spin_lock(&svcpt->scp_lock);
2760 if (!ptlrpc_threads_increasable(svcpt)) {
2761 cfs_spin_unlock(&svcpt->scp_lock);
2762 OBD_FREE_PTR(thread);
2766 if (svcpt->scp_nthrs_starting != 0) {
2767 /* serialize starting because some modules (obdfilter)
2768 * might require unique and contiguous t_id */
2769 LASSERT(svcpt->scp_nthrs_starting == 1);
2770 cfs_spin_unlock(&svcpt->scp_lock);
2771 OBD_FREE_PTR(thread);
2773 CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n",
2774 svc->srv_thread_name, svcpt->scp_thr_nextid);
2779 CDEBUG(D_INFO, "Creating thread %s #%d race, retry later\n",
2780 svc->srv_thread_name, svcpt->scp_thr_nextid);
2784 svcpt->scp_nthrs_starting++;
2785 thread->t_id = svcpt->scp_thr_nextid++;
2786 thread_add_flags(thread, SVC_STARTING);
2787 thread->t_svcpt = svcpt;
2789 cfs_list_add(&thread->t_link, &svcpt->scp_threads);
2790 cfs_spin_unlock(&svcpt->scp_lock);
2792 if (svcpt->scp_cpt >= 0) {
2793 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s%02d_%03d",
2794 svc->srv_thread_name, svcpt->scp_cpt, thread->t_id);
2796 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s_%04d",
2797 svc->srv_thread_name, thread->t_id);
2800 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name);
2802 * CLONE_VM and CLONE_FILES just avoid a needless copy, because we
2803 * just drop the VM and FILES in cfs_daemonize_ctxt() right away.
2805 rc = cfs_create_thread(ptlrpc_main, thread, CFS_DAEMON_FLAGS);
2807 CERROR("cannot start thread '%s': rc %d\n",
2808 thread->t_name, rc);
2809 cfs_spin_lock(&svcpt->scp_lock);
2810 cfs_list_del(&thread->t_link);
2811 --svcpt->scp_nthrs_starting;
2812 cfs_spin_unlock(&svcpt->scp_lock);
2814 OBD_FREE(thread, sizeof(*thread));
2821 l_wait_event(thread->t_ctl_waitq,
2822 thread_is_running(thread) || thread_is_stopped(thread),
2825 rc = thread_is_stopped(thread) ? thread->t_id : 0;
2829 int ptlrpc_hr_init(void)
2831 struct ptlrpc_hr_partition *hrp;
2832 struct ptlrpc_hr_thread *hrt;
2838 memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
2839 ptlrpc_hr.hr_cpt_table = cfs_cpt_table;
2841 ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
2843 if (ptlrpc_hr.hr_partitions == NULL)
2846 cfs_waitq_init(&ptlrpc_hr.hr_waitq);
2848 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2851 cfs_atomic_set(&hrp->hrp_nstarted, 0);
2852 cfs_atomic_set(&hrp->hrp_nstopped, 0);
2854 hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, i);
2855 hrp->hrp_nthrs /= cfs_cpu_ht_nsiblings(0);
2857 LASSERT(hrp->hrp_nthrs > 0);
2858 OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, i,
2859 hrp->hrp_nthrs * sizeof(*hrt));
2860 if (hrp->hrp_thrs == NULL)
2861 GOTO(out, rc = -ENOMEM);
2863 for (j = 0; j < hrp->hrp_nthrs; j++) {
2864 hrt = &hrp->hrp_thrs[j];
2867 hrt->hrt_partition = hrp;
2868 cfs_waitq_init(&hrt->hrt_waitq);
2869 cfs_spin_lock_init(&hrt->hrt_lock);
2870 CFS_INIT_LIST_HEAD(&hrt->hrt_queue);
2874 rc = ptlrpc_start_hr_threads();
2881 void ptlrpc_hr_fini(void)
2883 struct ptlrpc_hr_partition *hrp;
2886 if (ptlrpc_hr.hr_partitions == NULL)
2889 ptlrpc_stop_hr_threads();
2891 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2892 if (hrp->hrp_thrs != NULL) {
2893 OBD_FREE(hrp->hrp_thrs,
2894 hrp->hrp_nthrs * sizeof(hrp->hrp_thrs[0]));
2898 cfs_percpt_free(ptlrpc_hr.hr_partitions);
2899 ptlrpc_hr.hr_partitions = NULL;
2902 #endif /* __KERNEL__ */
2905 * Wait until all already scheduled replies are processed.
2907 static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
2911 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
2914 rc = l_wait_event(svcpt->scp_waitq,
2915 cfs_atomic_read(&svcpt->scp_nreps_difficult) == 0, &lwi);
2918 CWARN("Unexpectedly long timeout %s %p\n",
2919 svcpt->scp_service->srv_name, svcpt->scp_service);
2924 ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
2926 struct ptlrpc_service_part *svcpt;
2929 /* early disarm AT timer... */
2930 ptlrpc_service_for_each_part(svcpt, i, svc) {
2931 if (svcpt->scp_service != NULL)
2932 cfs_timer_disarm(&svcpt->scp_at_timer);
2937 ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
2939 struct ptlrpc_service_part *svcpt;
2940 struct ptlrpc_request_buffer_desc *rqbd;
2941 struct l_wait_info lwi;
2945 /* All history will be culled when the next request buffer is
2946 * freed in ptlrpc_service_purge_all() */
2947 svc->srv_hist_nrqbds_cpt_max = 0;
2949 rc = LNetClearLazyPortal(svc->srv_req_portal);
2952 ptlrpc_service_for_each_part(svcpt, i, svc) {
2953 if (svcpt->scp_service == NULL)
2956 /* Unlink all the request buffers. This forces a 'final'
2957 * event with its 'unlink' flag set for each posted rqbd */
2958 cfs_list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
2960 rc = LNetMDUnlink(rqbd->rqbd_md_h);
2961 LASSERT(rc == 0 || rc == -ENOENT);
2965 ptlrpc_service_for_each_part(svcpt, i, svc) {
2966 if (svcpt->scp_service == NULL)
2969 /* Wait for the network to release any buffers
2970 * it's currently filling */
2971 cfs_spin_lock(&svcpt->scp_lock);
2972 while (svcpt->scp_nrqbds_posted != 0) {
2973 cfs_spin_unlock(&svcpt->scp_lock);
2974 /* Network access will complete in finite time but
2975 * the HUGE timeout lets us CWARN for visibility
2976 * of sluggish NALs */
2977 lwi = LWI_TIMEOUT_INTERVAL(
2978 cfs_time_seconds(LONG_UNLINK),
2979 cfs_time_seconds(1), NULL, NULL);
2980 rc = l_wait_event(svcpt->scp_waitq,
2981 svcpt->scp_nrqbds_posted == 0, &lwi);
2982 if (rc == -ETIMEDOUT) {
2983 CWARN("Service %s waiting for "
2984 "request buffers\n",
2985 svcpt->scp_service->srv_name);
2987 cfs_spin_lock(&svcpt->scp_lock);
2989 cfs_spin_unlock(&svcpt->scp_lock);
2994 ptlrpc_service_purge_all(struct ptlrpc_service *svc)
2996 struct ptlrpc_service_part *svcpt;
2997 struct ptlrpc_request_buffer_desc *rqbd;
2998 struct ptlrpc_request *req;
2999 struct ptlrpc_reply_state *rs;
3002 ptlrpc_service_for_each_part(svcpt, i, svc) {
3003 if (svcpt->scp_service == NULL)
3006 cfs_spin_lock(&svcpt->scp_rep_lock);
3007 while (!cfs_list_empty(&svcpt->scp_rep_active)) {
3008 rs = cfs_list_entry(svcpt->scp_rep_active.next,
3009 struct ptlrpc_reply_state, rs_list);
3010 cfs_spin_lock(&rs->rs_lock);
3011 ptlrpc_schedule_difficult_reply(rs);
3012 cfs_spin_unlock(&rs->rs_lock);
3014 cfs_spin_unlock(&svcpt->scp_rep_lock);
3016 /* purge the request queue. NB No new replies (rqbds
3017 * all unlinked) and no service threads, so I'm the only
3018 * thread noodling the request queue now */
3019 while (!cfs_list_empty(&svcpt->scp_req_incoming)) {
3020 req = cfs_list_entry(svcpt->scp_req_incoming.next,
3021 struct ptlrpc_request, rq_list);
3023 cfs_list_del(&req->rq_list);
3024 svcpt->scp_nreqs_incoming--;
3025 svcpt->scp_nreqs_active++;
3026 ptlrpc_server_finish_request(svcpt, req);
3029 while (ptlrpc_server_request_pending(svcpt, 1)) {
3030 req = ptlrpc_server_request_get(svcpt, 1);
3031 cfs_list_del(&req->rq_list);
3032 svcpt->scp_nreqs_active++;
3033 ptlrpc_hpreq_fini(req);
3034 ptlrpc_server_finish_request(svcpt, req);
3037 LASSERT(cfs_list_empty(&svcpt->scp_rqbd_posted));
3038 LASSERT(svcpt->scp_nreqs_incoming == 0);
3039 LASSERT(svcpt->scp_nreqs_active == 0);
3040 /* history should have been culled by
3041 * ptlrpc_server_finish_request */
3042 LASSERT(svcpt->scp_hist_nrqbds == 0);
3044 /* Now free all the request buffers since nothing
3045 * references them any more... */
3047 while (!cfs_list_empty(&svcpt->scp_rqbd_idle)) {
3048 rqbd = cfs_list_entry(svcpt->scp_rqbd_idle.next,
3049 struct ptlrpc_request_buffer_desc,
3051 ptlrpc_free_rqbd(rqbd);
3053 ptlrpc_wait_replies(svcpt);
3055 while (!cfs_list_empty(&svcpt->scp_rep_idle)) {
3056 rs = cfs_list_entry(svcpt->scp_rep_idle.next,
3057 struct ptlrpc_reply_state,
3059 cfs_list_del(&rs->rs_list);
3060 OBD_FREE_LARGE(rs, svc->srv_max_reply_size);
3066 ptlrpc_service_free(struct ptlrpc_service *svc)
3068 struct ptlrpc_service_part *svcpt;
3069 struct ptlrpc_at_array *array;
3072 ptlrpc_service_for_each_part(svcpt, i, svc) {
3073 if (svcpt->scp_service == NULL)
3076 /* In case somebody rearmed this in the meantime */
3077 cfs_timer_disarm(&svcpt->scp_at_timer);
3078 array = &svcpt->scp_at_array;
3080 if (array->paa_reqs_array != NULL) {
3081 OBD_FREE(array->paa_reqs_array,
3082 sizeof(cfs_list_t) * array->paa_size);
3083 array->paa_reqs_array = NULL;
3086 if (array->paa_reqs_count != NULL) {
3087 OBD_FREE(array->paa_reqs_count,
3088 sizeof(__u32) * array->paa_size);
3089 array->paa_reqs_count = NULL;
3093 ptlrpc_service_for_each_part(svcpt, i, svc)
3094 OBD_FREE_PTR(svcpt);
3096 if (svc->srv_cpts != NULL)
3097 cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts);
3099 OBD_FREE(svc, offsetof(struct ptlrpc_service,
3100 srv_parts[svc->srv_ncpts]));
3103 int ptlrpc_unregister_service(struct ptlrpc_service *service)
3107 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
3109 service->srv_is_stopping = 1;
3111 cfs_spin_lock(&ptlrpc_all_services_lock);
3112 cfs_list_del_init(&service->srv_list);
3113 cfs_spin_unlock(&ptlrpc_all_services_lock);
3115 ptlrpc_lprocfs_unregister_service(service);
3117 ptlrpc_service_del_atimer(service);
3118 ptlrpc_stop_all_threads(service);
3120 ptlrpc_service_unlink_rqbd(service);
3121 ptlrpc_service_purge_all(service);
3122 ptlrpc_service_free(service);
3126 EXPORT_SYMBOL(ptlrpc_unregister_service);
3129 * Returns 0 if the service is healthy.
3131 * Right now, it just checks to make sure that requests aren't languishing
3132 * in the queue. We'll use this health check to govern whether a node needs
3133 * to be shot, so it's intentionally non-aggressive. */
3134 int ptlrpc_svcpt_health_check(struct ptlrpc_service_part *svcpt)
3136 struct ptlrpc_request *request;
3137 struct timeval right_now;
3140 cfs_gettimeofday(&right_now);
3142 cfs_spin_lock(&svcpt->scp_req_lock);
3143 if (!ptlrpc_server_request_pending(svcpt, 1)) {
3144 cfs_spin_unlock(&svcpt->scp_req_lock);
3148 /* How long has the next entry been waiting? */
3149 if (cfs_list_empty(&svcpt->scp_req_pending)) {
3150 request = cfs_list_entry(svcpt->scp_hreq_pending.next,
3151 struct ptlrpc_request, rq_list);
3153 request = cfs_list_entry(svcpt->scp_req_pending.next,
3154 struct ptlrpc_request, rq_list);
3157 timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
3158 cfs_spin_unlock(&svcpt->scp_req_lock);
3160 if ((timediff / ONE_MILLION) >
3161 (AT_OFF ? obd_timeout * 3 / 2 : at_max)) {
3162 CERROR("%s: unhealthy - request has been waiting %lds\n",
3163 svcpt->scp_service->srv_name, timediff / ONE_MILLION);
3171 ptlrpc_service_health_check(struct ptlrpc_service *svc)
3173 struct ptlrpc_service_part *svcpt;
3176 if (svc == NULL || svc->srv_parts == NULL)
3179 ptlrpc_service_for_each_part(svcpt, i, svc) {
3180 int rc = ptlrpc_svcpt_health_check(svcpt);
3187 EXPORT_SYMBOL(ptlrpc_service_health_check);