4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2010, 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 #define DEBUG_SUBSYSTEM S_RPC
38 #include <obd_support.h>
39 #include <obd_class.h>
40 #include <lustre_net.h>
41 #include <lu_object.h>
42 #include <linux/lnet/types.h>
43 #include "ptlrpc_internal.h"
45 /* The following are visible and mutable through /sys/module/ptlrpc */
46 int test_req_buffer_pressure = 0;
47 module_param(test_req_buffer_pressure, int, 0444);
48 MODULE_PARM_DESC(test_req_buffer_pressure, "set non-zero to put pressure on request buffer pools");
49 module_param(at_min, int, 0644);
50 MODULE_PARM_DESC(at_min, "Adaptive timeout minimum (sec)");
51 module_param(at_max, int, 0644);
52 MODULE_PARM_DESC(at_max, "Adaptive timeout maximum (sec)");
53 module_param(at_history, int, 0644);
54 MODULE_PARM_DESC(at_history,
55 "Adaptive timeouts remember the slowest event that took place within this period (sec)");
56 module_param(at_early_margin, int, 0644);
57 MODULE_PARM_DESC(at_early_margin, "How soon before an RPC deadline to send an early reply");
58 module_param(at_extra, int, 0644);
59 MODULE_PARM_DESC(at_extra, "How much extra time to give with each early reply");
63 static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt);
64 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req);
65 static void ptlrpc_at_remove_timed(struct ptlrpc_request *req);
67 /** Holds a list of all PTLRPC services */
68 LIST_HEAD(ptlrpc_all_services);
69 /** Used to protect the \e ptlrpc_all_services list */
70 struct mutex ptlrpc_all_services_mutex;
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 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 spin_lock(&svcpt->scp_lock);
95 list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
96 svcpt->scp_nrqbds_total++;
97 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(list_empty(&rqbd->rqbd_reqs));
110 spin_lock(&svcpt->scp_lock);
111 list_del(&rqbd->rqbd_list);
112 svcpt->scp_nrqbds_total--;
113 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 if (svcpt->scp_rqbd_allocating)
130 spin_lock(&svcpt->scp_lock);
131 /* check again with lock */
132 if (svcpt->scp_rqbd_allocating) {
133 /* NB: we might allow more than one thread in the future */
134 LASSERT(svcpt->scp_rqbd_allocating == 1);
135 spin_unlock(&svcpt->scp_lock);
139 svcpt->scp_rqbd_allocating++;
140 spin_unlock(&svcpt->scp_lock);
143 for (i = 0; i < svc->srv_nbuf_per_group; i++) {
144 /* NB: another thread might have recycled enough rqbds, we
145 * need to make sure it wouldn't over-allocate, see LU-1212. */
146 if (svcpt->scp_nrqbds_posted >= svc->srv_nbuf_per_group)
149 rqbd = ptlrpc_alloc_rqbd(svcpt);
152 CERROR("%s: Can't allocate request buffer\n",
159 spin_lock(&svcpt->scp_lock);
161 LASSERT(svcpt->scp_rqbd_allocating == 1);
162 svcpt->scp_rqbd_allocating--;
164 spin_unlock(&svcpt->scp_lock);
167 "%s: allocate %d new %d-byte reqbufs (%d/%d left), rc = %d\n",
168 svc->srv_name, i, svc->srv_buf_size, svcpt->scp_nrqbds_posted,
169 svcpt->scp_nrqbds_total, rc);
173 rc = ptlrpc_server_post_idle_rqbds(svcpt);
179 * Part of Rep-Ack logic.
180 * Puts a lock and its mode into reply state assotiated to request reply.
183 ptlrpc_save_lock(struct ptlrpc_request *req,
184 struct lustre_handle *lock, int mode, int no_ack)
186 struct ptlrpc_reply_state *rs = req->rq_reply_state;
190 LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
192 if (req->rq_export->exp_disconnected) {
193 ldlm_lock_decref(lock, mode);
195 idx = rs->rs_nlocks++;
196 rs->rs_locks[idx] = *lock;
197 rs->rs_modes[idx] = mode;
198 rs->rs_difficult = 1;
199 rs->rs_no_ack = !!no_ack;
202 EXPORT_SYMBOL(ptlrpc_save_lock);
205 struct ptlrpc_hr_partition;
207 struct ptlrpc_hr_thread {
208 int hrt_id; /* thread ID */
210 wait_queue_head_t hrt_waitq;
211 struct list_head hrt_queue; /* RS queue */
212 struct ptlrpc_hr_partition *hrt_partition;
215 struct ptlrpc_hr_partition {
216 /* # of started threads */
217 atomic_t hrp_nstarted;
218 /* # of stopped threads */
219 atomic_t hrp_nstopped;
220 /* cpu partition id */
222 /* round-robin rotor for choosing thread */
224 /* total number of threads on this partition */
227 struct ptlrpc_hr_thread *hrp_thrs;
230 #define HRT_RUNNING 0
231 #define HRT_STOPPING 1
233 struct ptlrpc_hr_service {
234 /* CPU partition table, it's just cfs_cpt_table for now */
235 struct cfs_cpt_table *hr_cpt_table;
236 /** controller sleep waitq */
237 wait_queue_head_t hr_waitq;
238 unsigned int hr_stopping;
239 /** roundrobin rotor for non-affinity service */
240 unsigned int hr_rotor;
242 struct ptlrpc_hr_partition **hr_partitions;
246 struct list_head rsb_replies;
247 unsigned int rsb_n_replies;
248 struct ptlrpc_service_part *rsb_svcpt;
251 /** reply handling service. */
252 static struct ptlrpc_hr_service ptlrpc_hr;
255 * maximum mumber of replies scheduled in one batch
257 #define MAX_SCHEDULED 256
260 * Initialize a reply batch.
264 static void rs_batch_init(struct rs_batch *b)
266 memset(b, 0, sizeof(*b));
267 INIT_LIST_HEAD(&b->rsb_replies);
271 * Choose an hr thread to dispatch requests to.
273 static struct ptlrpc_hr_thread *
274 ptlrpc_hr_select(struct ptlrpc_service_part *svcpt)
276 struct ptlrpc_hr_partition *hrp;
279 if (svcpt->scp_cpt >= 0 &&
280 svcpt->scp_service->srv_cptable == ptlrpc_hr.hr_cpt_table) {
281 /* directly match partition */
282 hrp = ptlrpc_hr.hr_partitions[svcpt->scp_cpt];
285 rotor = ptlrpc_hr.hr_rotor++;
286 rotor %= cfs_cpt_number(ptlrpc_hr.hr_cpt_table);
288 hrp = ptlrpc_hr.hr_partitions[rotor];
291 rotor = hrp->hrp_rotor++;
292 return &hrp->hrp_thrs[rotor % hrp->hrp_nthrs];
296 * Dispatch all replies accumulated in the batch to one from
297 * dedicated reply handling threads.
301 static void rs_batch_dispatch(struct rs_batch *b)
303 if (b->rsb_n_replies != 0) {
304 struct ptlrpc_hr_thread *hrt;
306 hrt = ptlrpc_hr_select(b->rsb_svcpt);
308 spin_lock(&hrt->hrt_lock);
309 list_splice_init(&b->rsb_replies, &hrt->hrt_queue);
310 spin_unlock(&hrt->hrt_lock);
312 wake_up(&hrt->hrt_waitq);
313 b->rsb_n_replies = 0;
318 * Add a reply to a batch.
319 * Add one reply object to a batch, schedule batched replies if overload.
324 static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs)
326 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
328 if (svcpt != b->rsb_svcpt || b->rsb_n_replies >= MAX_SCHEDULED) {
329 if (b->rsb_svcpt != NULL) {
330 rs_batch_dispatch(b);
331 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
333 spin_lock(&svcpt->scp_rep_lock);
334 b->rsb_svcpt = svcpt;
336 spin_lock(&rs->rs_lock);
337 rs->rs_scheduled_ever = 1;
338 if (rs->rs_scheduled == 0) {
339 list_move(&rs->rs_list, &b->rsb_replies);
340 rs->rs_scheduled = 1;
343 rs->rs_committed = 1;
344 spin_unlock(&rs->rs_lock);
348 * Reply batch finalization.
349 * Dispatch remaining replies from the batch
350 * and release remaining spinlock.
354 static void rs_batch_fini(struct rs_batch *b)
356 if (b->rsb_svcpt != NULL) {
357 rs_batch_dispatch(b);
358 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
362 #define DECLARE_RS_BATCH(b) struct rs_batch b
366 * Put reply state into a queue for processing because we received
367 * ACK from the client
369 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
371 struct ptlrpc_hr_thread *hrt;
373 LASSERT(list_empty(&rs->rs_list));
375 hrt = ptlrpc_hr_select(rs->rs_svcpt);
377 spin_lock(&hrt->hrt_lock);
378 list_add_tail(&rs->rs_list, &hrt->hrt_queue);
379 spin_unlock(&hrt->hrt_lock);
381 wake_up(&hrt->hrt_waitq);
385 ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs)
387 LASSERT(spin_is_locked(&rs->rs_svcpt->scp_rep_lock));
388 LASSERT(spin_is_locked(&rs->rs_lock));
389 LASSERT(rs->rs_difficult);
390 rs->rs_scheduled_ever = 1; /* flag any notification attempt */
392 if (rs->rs_scheduled) { /* being set up or already notified */
396 rs->rs_scheduled = 1;
397 list_del_init(&rs->rs_list);
398 ptlrpc_dispatch_difficult_reply(rs);
400 EXPORT_SYMBOL(ptlrpc_schedule_difficult_reply);
402 void ptlrpc_commit_replies(struct obd_export *exp)
404 struct ptlrpc_reply_state *rs, *nxt;
405 DECLARE_RS_BATCH(batch);
407 rs_batch_init(&batch);
408 /* Find any replies that have been committed and get their service
409 * to attend to complete them. */
411 /* CAVEAT EMPTOR: spinlock ordering!!! */
412 spin_lock(&exp->exp_uncommitted_replies_lock);
413 list_for_each_entry_safe(rs, nxt, &exp->exp_uncommitted_replies,
415 LASSERT(rs->rs_difficult);
416 /* VBR: per-export last_committed */
417 LASSERT(rs->rs_export);
418 if (rs->rs_transno <= exp->exp_last_committed) {
419 list_del_init(&rs->rs_obd_list);
420 rs_batch_add(&batch, rs);
423 spin_unlock(&exp->exp_uncommitted_replies_lock);
424 rs_batch_fini(&batch);
426 EXPORT_SYMBOL(ptlrpc_commit_replies);
429 ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt)
431 struct ptlrpc_request_buffer_desc *rqbd;
436 spin_lock(&svcpt->scp_lock);
438 if (list_empty(&svcpt->scp_rqbd_idle)) {
439 spin_unlock(&svcpt->scp_lock);
443 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
444 struct ptlrpc_request_buffer_desc,
446 list_del(&rqbd->rqbd_list);
448 /* assume we will post successfully */
449 svcpt->scp_nrqbds_posted++;
450 list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_posted);
452 spin_unlock(&svcpt->scp_lock);
454 rc = ptlrpc_register_rqbd(rqbd);
461 spin_lock(&svcpt->scp_lock);
463 svcpt->scp_nrqbds_posted--;
464 list_del(&rqbd->rqbd_list);
465 list_add_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
467 /* Don't complain if no request buffers are posted right now; LNET
468 * won't drop requests because we set the portal lazy! */
470 spin_unlock(&svcpt->scp_lock);
475 static void ptlrpc_at_timer(unsigned long castmeharder)
477 struct ptlrpc_service_part *svcpt;
479 svcpt = (struct ptlrpc_service_part *)castmeharder;
481 svcpt->scp_at_check = 1;
482 svcpt->scp_at_checktime = cfs_time_current();
483 wake_up(&svcpt->scp_waitq);
487 ptlrpc_server_nthreads_check(struct ptlrpc_service *svc,
488 struct ptlrpc_service_conf *conf)
490 struct ptlrpc_service_thr_conf *tc = &conf->psc_thr;
497 * Common code for estimating & validating threads number.
498 * CPT affinity service could have percpt thread-pool instead
499 * of a global thread-pool, which means user might not always
500 * get the threads number they give it in conf::tc_nthrs_user
501 * even they did set. It's because we need to validate threads
502 * number for each CPT to guarantee each pool will have enough
503 * threads to keep the service healthy.
505 init = PTLRPC_NTHRS_INIT + (svc->srv_ops.so_hpreq_handler != NULL);
506 init = max_t(int, init, tc->tc_nthrs_init);
508 /* NB: please see comments in lustre_lnet.h for definition
509 * details of these members */
510 LASSERT(tc->tc_nthrs_max != 0);
512 if (tc->tc_nthrs_user != 0) {
513 /* In case there is a reason to test a service with many
514 * threads, we give a less strict check here, it can
515 * be up to 8 * nthrs_max */
516 total = min(tc->tc_nthrs_max * 8, tc->tc_nthrs_user);
517 nthrs = total / svc->srv_ncpts;
518 init = max(init, nthrs);
522 total = tc->tc_nthrs_max;
523 if (tc->tc_nthrs_base == 0) {
524 /* don't care about base threads number per partition,
525 * this is most for non-affinity service */
526 nthrs = total / svc->srv_ncpts;
530 nthrs = tc->tc_nthrs_base;
531 if (svc->srv_ncpts == 1) {
534 /* NB: Increase the base number if it's single partition
535 * and total number of cores/HTs is larger or equal to 4.
536 * result will always < 2 * nthrs_base */
537 weight = cfs_cpt_weight(svc->srv_cptable, CFS_CPT_ANY);
538 for (i = 1; (weight >> (i + 1)) != 0 && /* >= 4 cores/HTs */
539 (tc->tc_nthrs_base >> i) != 0; i++)
540 nthrs += tc->tc_nthrs_base >> i;
543 if (tc->tc_thr_factor != 0) {
544 int factor = tc->tc_thr_factor;
549 * User wants to increase number of threads with for
550 * each CPU core/HT, most likely the factor is larger then
551 * one thread/core because service threads are supposed to
552 * be blocked by lock or wait for IO.
555 * Amdahl's law says that adding processors wouldn't give
556 * a linear increasing of parallelism, so it's nonsense to
557 * have too many threads no matter how many cores/HTs
560 cpumask_copy(&mask, topology_thread_cpumask(0));
561 if (cpus_weight(mask) > 1) { /* weight is # of HTs */
562 /* depress thread factor for hyper-thread */
563 factor = factor - (factor >> 1) + (factor >> 3);
566 weight = cfs_cpt_weight(svc->srv_cptable, 0);
569 for (; factor > 0 && weight > 0; factor--, weight -= fade)
570 nthrs += min(weight, fade) * factor;
573 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
574 nthrs = max(tc->tc_nthrs_base,
575 tc->tc_nthrs_max / svc->srv_ncpts);
578 nthrs = max(nthrs, tc->tc_nthrs_init);
579 svc->srv_nthrs_cpt_limit = nthrs;
580 svc->srv_nthrs_cpt_init = init;
582 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
583 CDEBUG(D_OTHER, "%s: This service may have more threads (%d) "
584 "than the given soft limit (%d)\n",
585 svc->srv_name, nthrs * svc->srv_ncpts,
591 * Initialize percpt data for a service
594 ptlrpc_service_part_init(struct ptlrpc_service *svc,
595 struct ptlrpc_service_part *svcpt, int cpt)
597 struct ptlrpc_at_array *array;
602 svcpt->scp_cpt = cpt;
603 INIT_LIST_HEAD(&svcpt->scp_threads);
605 /* rqbd and incoming request queue */
606 spin_lock_init(&svcpt->scp_lock);
607 INIT_LIST_HEAD(&svcpt->scp_rqbd_idle);
608 INIT_LIST_HEAD(&svcpt->scp_rqbd_posted);
609 INIT_LIST_HEAD(&svcpt->scp_req_incoming);
610 init_waitqueue_head(&svcpt->scp_waitq);
611 /* history request & rqbd list */
612 INIT_LIST_HEAD(&svcpt->scp_hist_reqs);
613 INIT_LIST_HEAD(&svcpt->scp_hist_rqbds);
615 /* acitve requests and hp requests */
616 spin_lock_init(&svcpt->scp_req_lock);
619 spin_lock_init(&svcpt->scp_rep_lock);
620 INIT_LIST_HEAD(&svcpt->scp_rep_active);
621 INIT_LIST_HEAD(&svcpt->scp_rep_idle);
622 init_waitqueue_head(&svcpt->scp_rep_waitq);
623 atomic_set(&svcpt->scp_nreps_difficult, 0);
625 /* adaptive timeout */
626 spin_lock_init(&svcpt->scp_at_lock);
627 array = &svcpt->scp_at_array;
629 size = at_est2timeout(at_max);
630 array->paa_size = size;
631 array->paa_count = 0;
632 array->paa_deadline = -1;
634 /* allocate memory for scp_at_array (ptlrpc_at_array) */
635 OBD_CPT_ALLOC(array->paa_reqs_array,
636 svc->srv_cptable, cpt, sizeof(struct list_head) * size);
637 if (array->paa_reqs_array == NULL)
640 for (index = 0; index < size; index++)
641 INIT_LIST_HEAD(&array->paa_reqs_array[index]);
643 OBD_CPT_ALLOC(array->paa_reqs_count,
644 svc->srv_cptable, cpt, sizeof(__u32) * size);
645 if (array->paa_reqs_count == NULL)
648 cfs_timer_init(&svcpt->scp_at_timer, ptlrpc_at_timer, svcpt);
649 /* At SOW, service time should be quick; 10s seems generous. If client
650 * timeout is less than this, we'll be sending an early reply. */
651 at_init(&svcpt->scp_at_estimate, 10, 0);
653 /* assign this before call ptlrpc_grow_req_bufs */
654 svcpt->scp_service = svc;
655 /* Now allocate the request buffers, but don't post them now */
656 rc = ptlrpc_grow_req_bufs(svcpt, 0);
657 /* We shouldn't be under memory pressure at startup, so
658 * fail if we can't allocate all our buffers at this time. */
665 if (array->paa_reqs_count != NULL) {
666 OBD_FREE(array->paa_reqs_count, sizeof(__u32) * size);
667 array->paa_reqs_count = NULL;
670 if (array->paa_reqs_array != NULL) {
671 OBD_FREE(array->paa_reqs_array,
672 sizeof(struct list_head) * array->paa_size);
673 array->paa_reqs_array = NULL;
680 * Initialize service on a given portal.
681 * This includes starting serving threads , allocating and posting rqbds and
684 struct ptlrpc_service *
685 ptlrpc_register_service(struct ptlrpc_service_conf *conf,
686 struct proc_dir_entry *proc_entry)
688 struct ptlrpc_service_cpt_conf *cconf = &conf->psc_cpt;
689 struct ptlrpc_service *service;
690 struct ptlrpc_service_part *svcpt;
691 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);
729 OBD_FREE(cpts, sizeof(*cpts) * ncpts);
730 return ERR_PTR(rc < 0 ? rc : -EINVAL);
736 OBD_ALLOC(service, offsetof(struct ptlrpc_service, srv_parts[ncpts]));
737 if (service == NULL) {
739 OBD_FREE(cpts, sizeof(*cpts) * ncpts);
740 return ERR_PTR(-ENOMEM);
743 service->srv_cptable = cptable;
744 service->srv_cpts = cpts;
745 service->srv_ncpts = ncpts;
747 service->srv_cpt_bits = 0; /* it's zero already, easy to read... */
748 while ((1 << service->srv_cpt_bits) < cfs_cpt_number(cptable))
749 service->srv_cpt_bits++;
752 spin_lock_init(&service->srv_lock);
753 service->srv_name = conf->psc_name;
754 service->srv_watchdog_factor = conf->psc_watchdog_factor;
755 INIT_LIST_HEAD(&service->srv_list); /* for safty of cleanup */
757 /* buffer configuration */
758 service->srv_nbuf_per_group = test_req_buffer_pressure ?
759 1 : conf->psc_buf.bc_nbufs;
760 service->srv_max_req_size = conf->psc_buf.bc_req_max_size +
762 service->srv_buf_size = conf->psc_buf.bc_buf_size;
763 service->srv_rep_portal = conf->psc_buf.bc_rep_portal;
764 service->srv_req_portal = conf->psc_buf.bc_req_portal;
766 /* Increase max reply size to next power of two */
767 service->srv_max_reply_size = 1;
768 while (service->srv_max_reply_size <
769 conf->psc_buf.bc_rep_max_size + SPTLRPC_MAX_PAYLOAD)
770 service->srv_max_reply_size <<= 1;
772 service->srv_thread_name = conf->psc_thr.tc_thr_name;
773 service->srv_ctx_tags = conf->psc_thr.tc_ctx_tags;
774 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
775 service->srv_ops = conf->psc_ops;
777 for (i = 0; i < ncpts; i++) {
778 if (!conf->psc_thr.tc_cpu_affinity)
781 cpt = cpts != NULL ? cpts[i] : i;
783 OBD_CPT_ALLOC(svcpt, cptable, cpt, sizeof(*svcpt));
785 GOTO(failed, rc = -ENOMEM);
787 service->srv_parts[i] = svcpt;
788 rc = ptlrpc_service_part_init(service, svcpt, cpt);
793 ptlrpc_server_nthreads_check(service, conf);
795 rc = LNetSetLazyPortal(service->srv_req_portal);
798 mutex_lock(&ptlrpc_all_services_mutex);
799 list_add(&service->srv_list, &ptlrpc_all_services);
800 mutex_unlock(&ptlrpc_all_services_mutex);
802 if (proc_entry != NULL)
803 ptlrpc_lprocfs_register_service(proc_entry, service);
805 rc = ptlrpc_service_nrs_setup(service);
809 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
810 service->srv_name, service->srv_req_portal);
812 rc = ptlrpc_start_threads(service);
814 CERROR("Failed to start threads for service %s: %d\n",
815 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(atomic_read(&req->rq_refcount) == 0);
833 LASSERT(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;
859 struct list_head *tmp;
860 struct list_head *nxt;
862 if (!atomic_dec_and_test(&req->rq_refcount))
865 if (req->rq_at_linked) {
866 spin_lock(&svcpt->scp_at_lock);
867 /* recheck with lock, in case it's unlinked by
868 * ptlrpc_at_check_timed() */
869 if (likely(req->rq_at_linked))
870 ptlrpc_at_remove_timed(req);
871 spin_unlock(&svcpt->scp_at_lock);
874 LASSERT(list_empty(&req->rq_timed_list));
876 /* finalize request */
877 if (req->rq_export) {
878 class_export_put(req->rq_export);
879 req->rq_export = NULL;
882 spin_lock(&svcpt->scp_lock);
884 list_add(&req->rq_list, &rqbd->rqbd_reqs);
886 refcount = --(rqbd->rqbd_refcount);
888 /* request buffer is now idle: add to history */
889 list_del(&rqbd->rqbd_list);
891 list_add_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds);
892 svcpt->scp_hist_nrqbds++;
894 /* cull some history?
895 * I expect only about 1 or 2 rqbds need to be recycled here */
896 while (svcpt->scp_hist_nrqbds > svc->srv_hist_nrqbds_cpt_max) {
897 rqbd = list_entry(svcpt->scp_hist_rqbds.next,
898 struct ptlrpc_request_buffer_desc,
901 list_del(&rqbd->rqbd_list);
902 svcpt->scp_hist_nrqbds--;
904 /* remove rqbd's reqs from svc's req history while
905 * I've got the service lock */
906 list_for_each(tmp, &rqbd->rqbd_reqs) {
907 req = list_entry(tmp, struct ptlrpc_request,
909 /* Track the highest culled req seq */
910 if (req->rq_history_seq >
911 svcpt->scp_hist_seq_culled) {
912 svcpt->scp_hist_seq_culled =
915 list_del(&req->rq_history_list);
918 spin_unlock(&svcpt->scp_lock);
920 list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
921 req = list_entry(rqbd->rqbd_reqs.next,
922 struct ptlrpc_request,
924 list_del(&req->rq_list);
925 ptlrpc_server_free_request(req);
928 spin_lock(&svcpt->scp_lock);
930 * now all reqs including the embedded req has been
931 * disposed, schedule request buffer for re-use.
933 LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) ==
935 list_add_tail(&rqbd->rqbd_list,
936 &svcpt->scp_rqbd_idle);
939 spin_unlock(&svcpt->scp_lock);
940 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
941 /* If we are low on memory, we are not interested in history */
942 list_del(&req->rq_list);
943 list_del_init(&req->rq_history_list);
945 /* Track the highest culled req seq */
946 if (req->rq_history_seq > svcpt->scp_hist_seq_culled)
947 svcpt->scp_hist_seq_culled = req->rq_history_seq;
949 spin_unlock(&svcpt->scp_lock);
951 ptlrpc_server_free_request(req);
953 spin_unlock(&svcpt->scp_lock);
957 /** Change request export and move hp request from old export to new */
958 void ptlrpc_request_change_export(struct ptlrpc_request *req,
959 struct obd_export *export)
961 if (req->rq_export != NULL) {
962 if (!list_empty(&req->rq_exp_list)) {
963 /* remove rq_exp_list from last export */
964 spin_lock_bh(&req->rq_export->exp_rpc_lock);
965 list_del_init(&req->rq_exp_list);
966 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
968 /* export has one reference already, so it`s safe to
969 * add req to export queue here and get another
970 * reference for request later */
971 spin_lock_bh(&export->exp_rpc_lock);
972 list_add(&req->rq_exp_list, &export->exp_hp_rpcs);
973 spin_unlock_bh(&export->exp_rpc_lock);
975 class_export_rpc_dec(req->rq_export);
976 class_export_put(req->rq_export);
979 /* request takes one export refcount */
980 req->rq_export = class_export_get(export);
981 class_export_rpc_inc(export);
987 * to finish a request: stop sending more early replies, and release
990 static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt,
991 struct ptlrpc_request *req)
993 ptlrpc_server_hpreq_fini(req);
995 ptlrpc_server_drop_request(req);
999 * to finish a active request: stop sending more early replies, and release
1000 * the request. should be called after we finished handling the request.
1002 static void ptlrpc_server_finish_active_request(
1003 struct ptlrpc_service_part *svcpt,
1004 struct ptlrpc_request *req)
1006 spin_lock(&svcpt->scp_req_lock);
1007 ptlrpc_nrs_req_stop_nolock(req);
1008 svcpt->scp_nreqs_active--;
1010 svcpt->scp_nhreqs_active--;
1011 spin_unlock(&svcpt->scp_req_lock);
1013 ptlrpc_nrs_req_finalize(req);
1015 if (req->rq_export != NULL)
1016 class_export_rpc_dec(req->rq_export);
1018 ptlrpc_server_finish_request(svcpt, req);
1022 * This function makes sure dead exports are evicted in a timely manner.
1023 * This function is only called when some export receives a message (i.e.,
1024 * the network is up.)
1026 static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
1028 struct obd_export *oldest_exp;
1029 time_t oldest_time, new_time;
1033 /* Compensate for slow machines, etc, by faking our request time
1034 into the future. Although this can break the strict time-ordering
1035 of the list, we can be really lazy here - we don't have to evict
1036 at the exact right moment. Eventually, all silent exports
1037 will make it to the top of the list. */
1039 /* Do not pay attention on 1sec or smaller renewals. */
1040 new_time = cfs_time_current_sec() + extra_delay;
1041 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
1044 exp->exp_last_request_time = new_time;
1045 CDEBUG(D_HA, "updating export %s at "CFS_TIME_T" exp %p\n",
1046 exp->exp_client_uuid.uuid,
1047 exp->exp_last_request_time, exp);
1049 /* exports may get disconnected from the chain even though the
1050 export has references, so we must keep the spin lock while
1051 manipulating the lists */
1052 spin_lock(&exp->exp_obd->obd_dev_lock);
1054 if (list_empty(&exp->exp_obd_chain_timed)) {
1055 /* this one is not timed */
1056 spin_unlock(&exp->exp_obd->obd_dev_lock);
1060 list_move_tail(&exp->exp_obd_chain_timed,
1061 &exp->exp_obd->obd_exports_timed);
1063 oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
1064 struct obd_export, exp_obd_chain_timed);
1065 oldest_time = oldest_exp->exp_last_request_time;
1066 spin_unlock(&exp->exp_obd->obd_dev_lock);
1068 if (exp->exp_obd->obd_recovering) {
1069 /* be nice to everyone during recovery */
1073 /* Note - racing to start/reset the obd_eviction timer is safe */
1074 if (exp->exp_obd->obd_eviction_timer == 0) {
1075 /* Check if the oldest entry is expired. */
1076 if (cfs_time_current_sec() > (oldest_time + PING_EVICT_TIMEOUT +
1078 /* We need a second timer, in case the net was down and
1079 * it just came back. Since the pinger may skip every
1080 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
1081 * we better wait for 3. */
1082 exp->exp_obd->obd_eviction_timer =
1083 cfs_time_current_sec() + 3 * PING_INTERVAL;
1084 CDEBUG(D_HA, "%s: Think about evicting %s from "CFS_TIME_T"\n",
1085 exp->exp_obd->obd_name,
1086 obd_export_nid2str(oldest_exp), oldest_time);
1089 if (cfs_time_current_sec() >
1090 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
1091 /* The evictor won't evict anyone who we've heard from
1092 * recently, so we don't have to check before we start
1094 if (!ping_evictor_wake(exp))
1095 exp->exp_obd->obd_eviction_timer = 0;
1101 * Sanity check request \a req.
1102 * Return 0 if all is ok, error code otherwise.
1104 static int ptlrpc_check_req(struct ptlrpc_request *req)
1108 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
1109 req->rq_export->exp_conn_cnt)) {
1110 DEBUG_REQ(D_RPCTRACE, req,
1111 "DROPPING req from old connection %d < %d",
1112 lustre_msg_get_conn_cnt(req->rq_reqmsg),
1113 req->rq_export->exp_conn_cnt);
1116 if (unlikely(req->rq_export->exp_obd &&
1117 req->rq_export->exp_obd->obd_fail)) {
1119 * Failing over, don't handle any more reqs, send
1120 * error response instead.
1122 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
1123 req, req->rq_export->exp_obd->obd_name);
1125 } else if (lustre_msg_get_flags(req->rq_reqmsg) &
1126 (MSG_REPLAY | MSG_REQ_REPLAY_DONE) &&
1127 !(req->rq_export->exp_obd->obd_recovering)) {
1128 DEBUG_REQ(D_ERROR, req,
1129 "Invalid replay without recovery");
1130 class_fail_export(req->rq_export);
1132 } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 &&
1133 !(req->rq_export->exp_obd->obd_recovering)) {
1134 DEBUG_REQ(D_ERROR, req, "Invalid req with transno "
1135 LPU64" without recovery",
1136 lustre_msg_get_transno(req->rq_reqmsg));
1137 class_fail_export(req->rq_export);
1141 if (unlikely(rc < 0)) {
1142 req->rq_status = rc;
1148 static void ptlrpc_at_set_timer(struct ptlrpc_service_part *svcpt)
1150 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1153 if (array->paa_count == 0) {
1154 cfs_timer_disarm(&svcpt->scp_at_timer);
1158 /* Set timer for closest deadline */
1159 next = (__s32)(array->paa_deadline - cfs_time_current_sec() -
1162 ptlrpc_at_timer((unsigned long)svcpt);
1164 cfs_timer_arm(&svcpt->scp_at_timer, cfs_time_shift(next));
1165 CDEBUG(D_INFO, "armed %s at %+ds\n",
1166 svcpt->scp_service->srv_name, next);
1170 /* Add rpc to early reply check list */
1171 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
1173 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1174 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1175 struct ptlrpc_request *rq = NULL;
1181 if (req->rq_no_reply)
1184 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
1187 spin_lock(&svcpt->scp_at_lock);
1188 LASSERT(list_empty(&req->rq_timed_list));
1190 index = (unsigned long)req->rq_deadline % array->paa_size;
1191 if (array->paa_reqs_count[index] > 0) {
1192 /* latest rpcs will have the latest deadlines in the list,
1193 * so search backward. */
1194 list_for_each_entry_reverse(rq,
1195 &array->paa_reqs_array[index],
1197 if (req->rq_deadline >= rq->rq_deadline) {
1198 list_add(&req->rq_timed_list,
1199 &rq->rq_timed_list);
1205 /* Add the request at the head of the list */
1206 if (list_empty(&req->rq_timed_list))
1207 list_add(&req->rq_timed_list,
1208 &array->paa_reqs_array[index]);
1210 spin_lock(&req->rq_lock);
1211 req->rq_at_linked = 1;
1212 spin_unlock(&req->rq_lock);
1213 req->rq_at_index = index;
1214 array->paa_reqs_count[index]++;
1216 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
1217 array->paa_deadline = req->rq_deadline;
1218 ptlrpc_at_set_timer(svcpt);
1220 spin_unlock(&svcpt->scp_at_lock);
1226 ptlrpc_at_remove_timed(struct ptlrpc_request *req)
1228 struct ptlrpc_at_array *array;
1230 array = &req->rq_rqbd->rqbd_svcpt->scp_at_array;
1232 /* NB: must call with hold svcpt::scp_at_lock */
1233 LASSERT(!list_empty(&req->rq_timed_list));
1234 list_del_init(&req->rq_timed_list);
1236 spin_lock(&req->rq_lock);
1237 req->rq_at_linked = 0;
1238 spin_unlock(&req->rq_lock);
1240 array->paa_reqs_count[req->rq_at_index]--;
1244 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
1246 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1247 struct ptlrpc_request *reqcopy;
1248 struct lustre_msg *reqmsg;
1249 cfs_duration_t olddl = req->rq_deadline - cfs_time_current_sec();
1253 /* deadline is when the client expects us to reply, margin is the
1254 difference between clients' and servers' expectations */
1255 DEBUG_REQ(D_ADAPTTO, req,
1256 "%ssending early reply (deadline %+lds, margin %+lds) for "
1257 "%d+%d", AT_OFF ? "AT off - not " : "",
1258 olddl, olddl - at_get(&svcpt->scp_at_estimate),
1259 at_get(&svcpt->scp_at_estimate), at_extra);
1265 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), "
1266 "not sending early reply. Consider increasing "
1267 "at_early_margin (%d)?", olddl, at_early_margin);
1269 /* Return an error so we're not re-added to the timed list. */
1273 if (!(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
1274 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
1275 "but no AT support");
1279 if (req->rq_export &&
1280 lustre_msg_get_flags(req->rq_reqmsg) &
1281 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1282 /* During recovery, we don't want to send too many early
1283 * replies, but on the other hand we want to make sure the
1284 * client has enough time to resend if the rpc is lost. So
1285 * during the recovery period send at least 4 early replies,
1286 * spacing them every at_extra if we can. at_estimate should
1287 * always equal this fixed value during recovery. */
1288 at_measured(&svcpt->scp_at_estimate, min(at_extra,
1289 req->rq_export->exp_obd->obd_recovery_timeout / 4));
1291 /* Fake our processing time into the future to ask the clients
1292 * for some extra amount of time */
1293 at_measured(&svcpt->scp_at_estimate, at_extra +
1294 cfs_time_current_sec() -
1295 req->rq_arrival_time.tv_sec);
1297 /* Check to see if we've actually increased the deadline -
1298 * we may be past adaptive_max */
1299 if (req->rq_deadline >= req->rq_arrival_time.tv_sec +
1300 at_get(&svcpt->scp_at_estimate)) {
1301 DEBUG_REQ(D_WARNING, req, "Couldn't add any time "
1302 "(%ld/%ld), not sending early reply\n",
1303 olddl, req->rq_arrival_time.tv_sec +
1304 at_get(&svcpt->scp_at_estimate) -
1305 cfs_time_current_sec());
1309 newdl = cfs_time_current_sec() + at_get(&svcpt->scp_at_estimate);
1311 OBD_ALLOC(reqcopy, sizeof(*reqcopy));
1312 if (reqcopy == NULL)
1314 OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1316 OBD_FREE(reqcopy, sizeof(*reqcopy));
1321 reqcopy->rq_reply_state = NULL;
1322 reqcopy->rq_rep_swab_mask = 0;
1323 reqcopy->rq_pack_bulk = 0;
1324 reqcopy->rq_pack_udesc = 0;
1325 reqcopy->rq_packed_final = 0;
1326 sptlrpc_svc_ctx_addref(reqcopy);
1327 /* We only need the reqmsg for the magic */
1328 reqcopy->rq_reqmsg = reqmsg;
1329 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1331 LASSERT(atomic_read(&req->rq_refcount));
1332 /** if it is last refcount then early reply isn't needed */
1333 if (atomic_read(&req->rq_refcount) == 1) {
1334 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
1335 "abort sending early reply\n");
1336 GOTO(out, rc = -EINVAL);
1339 /* Connection ref */
1340 reqcopy->rq_export = class_conn2export(
1341 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1342 if (reqcopy->rq_export == NULL)
1343 GOTO(out, rc = -ENODEV);
1346 class_export_rpc_inc(reqcopy->rq_export);
1347 if (reqcopy->rq_export->exp_obd &&
1348 reqcopy->rq_export->exp_obd->obd_fail)
1349 GOTO(out_put, rc = -ENODEV);
1351 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1355 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1358 /* Adjust our own deadline to what we told the client */
1359 req->rq_deadline = newdl;
1360 req->rq_early_count++; /* number sent, server side */
1362 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1365 /* Free the (early) reply state from lustre_pack_reply.
1366 (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
1367 ptlrpc_req_drop_rs(reqcopy);
1370 class_export_rpc_dec(reqcopy->rq_export);
1371 class_export_put(reqcopy->rq_export);
1373 sptlrpc_svc_ctx_decref(reqcopy);
1374 OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1375 OBD_FREE(reqcopy, sizeof(*reqcopy));
1379 /* Send early replies to everybody expiring within at_early_margin
1380 asking for at_extra time */
1381 static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1383 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1384 struct ptlrpc_request *rq, *n;
1385 struct list_head work_list;
1388 time_t now = cfs_time_current_sec();
1389 cfs_duration_t delay;
1390 int first, counter = 0;
1392 spin_lock(&svcpt->scp_at_lock);
1393 if (svcpt->scp_at_check == 0) {
1394 spin_unlock(&svcpt->scp_at_lock);
1397 delay = cfs_time_sub(cfs_time_current(), svcpt->scp_at_checktime);
1398 svcpt->scp_at_check = 0;
1400 if (array->paa_count == 0) {
1401 spin_unlock(&svcpt->scp_at_lock);
1405 /* The timer went off, but maybe the nearest rpc already completed. */
1406 first = array->paa_deadline - now;
1407 if (first > at_early_margin) {
1408 /* We've still got plenty of time. Reset the timer. */
1409 ptlrpc_at_set_timer(svcpt);
1410 spin_unlock(&svcpt->scp_at_lock);
1414 /* We're close to a timeout, and we don't know how much longer the
1415 server will take. Send early replies to everyone expiring soon. */
1416 INIT_LIST_HEAD(&work_list);
1418 index = (unsigned long)array->paa_deadline % array->paa_size;
1419 count = array->paa_count;
1421 count -= array->paa_reqs_count[index];
1422 list_for_each_entry_safe(rq, n,
1423 &array->paa_reqs_array[index],
1425 if (rq->rq_deadline > now + at_early_margin) {
1426 /* update the earliest deadline */
1427 if (deadline == -1 ||
1428 rq->rq_deadline < deadline)
1429 deadline = rq->rq_deadline;
1433 ptlrpc_at_remove_timed(rq);
1435 * ptlrpc_server_drop_request() may drop
1436 * refcount to 0 already. Let's check this and
1437 * don't add entry to work_list
1439 if (likely(atomic_inc_not_zero(&rq->rq_refcount)))
1440 list_add(&rq->rq_timed_list, &work_list);
1444 if (++index >= array->paa_size)
1447 array->paa_deadline = deadline;
1448 /* we have a new earliest deadline, restart the timer */
1449 ptlrpc_at_set_timer(svcpt);
1451 spin_unlock(&svcpt->scp_at_lock);
1453 CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
1454 "replies\n", first, at_extra, counter);
1456 /* We're already past request deadlines before we even get a
1457 chance to send early replies */
1458 LCONSOLE_WARN("%s: This server is not able to keep up with "
1459 "request traffic (cpu-bound).\n",
1460 svcpt->scp_service->srv_name);
1461 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, "
1462 "delay="CFS_DURATION_T"(jiff)\n",
1463 counter, svcpt->scp_nreqs_incoming,
1464 svcpt->scp_nreqs_active,
1465 at_get(&svcpt->scp_at_estimate), delay);
1468 /* we took additional refcount so entries can't be deleted from list, no
1469 * locking is needed */
1470 while (!list_empty(&work_list)) {
1471 rq = list_entry(work_list.next, struct ptlrpc_request,
1473 list_del_init(&rq->rq_timed_list);
1475 if (ptlrpc_at_send_early_reply(rq) == 0)
1476 ptlrpc_at_add_timed(rq);
1478 ptlrpc_server_drop_request(rq);
1481 return 1; /* return "did_something" for liblustre */
1485 * Put the request to the export list if the request may become
1486 * a high priority one.
1488 static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt,
1489 struct ptlrpc_request *req)
1493 if (svcpt->scp_service->srv_ops.so_hpreq_handler) {
1494 rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
1499 if (req->rq_export && req->rq_ops) {
1500 /* Perform request specific check. We should do this check
1501 * before the request is added into exp_hp_rpcs list otherwise
1502 * it may hit swab race at LU-1044. */
1503 if (req->rq_ops->hpreq_check) {
1504 rc = req->rq_ops->hpreq_check(req);
1506 * XXX: Out of all current
1507 * ptlrpc_hpreq_ops::hpreq_check(), only
1508 * ldlm_cancel_hpreq_check() can return an error code;
1509 * other functions assert in similar places, which seems
1510 * odd. What also does not seem right is that handlers
1511 * for those RPCs do not assert on the same checks, but
1512 * rather handle the error cases. e.g. see
1513 * ost_rw_hpreq_check(), and ost_brw_read(),
1518 LASSERT(rc == 0 || rc == 1);
1521 spin_lock_bh(&req->rq_export->exp_rpc_lock);
1522 list_add(&req->rq_exp_list,
1523 &req->rq_export->exp_hp_rpcs);
1524 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1527 ptlrpc_nrs_req_initialize(svcpt, req, rc);
1532 /** Remove the request from the export list. */
1533 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req)
1535 if (req->rq_export && req->rq_ops) {
1536 /* refresh lock timeout again so that client has more
1537 * room to send lock cancel RPC. */
1538 if (req->rq_ops->hpreq_fini)
1539 req->rq_ops->hpreq_fini(req);
1541 spin_lock_bh(&req->rq_export->exp_rpc_lock);
1542 list_del_init(&req->rq_exp_list);
1543 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1547 static int ptlrpc_hpreq_check(struct ptlrpc_request *req)
1552 static struct ptlrpc_hpreq_ops ptlrpc_hpreq_common = {
1553 .hpreq_check = ptlrpc_hpreq_check,
1556 /* Hi-Priority RPC check by RPC operation code. */
1557 int ptlrpc_hpreq_handler(struct ptlrpc_request *req)
1559 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1561 /* Check for export to let only reconnects for not yet evicted
1562 * export to become a HP rpc. */
1563 if ((req->rq_export != NULL) &&
1564 (opc == OBD_PING || opc == MDS_CONNECT || opc == OST_CONNECT))
1565 req->rq_ops = &ptlrpc_hpreq_common;
1569 EXPORT_SYMBOL(ptlrpc_hpreq_handler);
1571 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1572 struct ptlrpc_request *req)
1576 rc = ptlrpc_server_hpreq_init(svcpt, req);
1580 ptlrpc_nrs_req_add(svcpt, req, !!rc);
1586 * Allow to handle high priority request
1587 * User can call it w/o any lock but need to hold
1588 * ptlrpc_service_part::scp_req_lock to get reliable result
1590 static bool ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1593 int running = svcpt->scp_nthrs_running;
1595 if (!nrs_svcpt_has_hp(svcpt))
1601 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1602 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1603 /* leave just 1 thread for normal RPCs */
1604 running = PTLRPC_NTHRS_INIT;
1605 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1609 if (svcpt->scp_nreqs_active >= running - 1)
1612 if (svcpt->scp_nhreqs_active == 0)
1615 return !ptlrpc_nrs_req_pending_nolock(svcpt, false) ||
1616 svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1619 static bool ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1622 return ptlrpc_server_allow_high(svcpt, force) &&
1623 ptlrpc_nrs_req_pending_nolock(svcpt, true);
1627 * Only allow normal priority requests on a service that has a high-priority
1628 * queue if forced (i.e. cleanup), if there are other high priority requests
1629 * already being processed (i.e. those threads can service more high-priority
1630 * requests), or if there are enough idle threads that a later thread can do
1631 * a high priority request.
1632 * User can call it w/o any lock but need to hold
1633 * ptlrpc_service_part::scp_req_lock to get reliable result
1635 static bool ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1638 int running = svcpt->scp_nthrs_running;
1639 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1640 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1641 /* leave just 1 thread for normal RPCs */
1642 running = PTLRPC_NTHRS_INIT;
1643 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1648 svcpt->scp_nreqs_active < running - 2)
1651 if (svcpt->scp_nreqs_active >= running - 1)
1654 return svcpt->scp_nhreqs_active > 0 || !nrs_svcpt_has_hp(svcpt);
1657 static bool ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1660 return ptlrpc_server_allow_normal(svcpt, force) &&
1661 ptlrpc_nrs_req_pending_nolock(svcpt, false);
1665 * Returns true if there are requests available in incoming
1666 * request queue for processing and it is allowed to fetch them.
1667 * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1668 * to get reliable result
1669 * \see ptlrpc_server_allow_normal
1670 * \see ptlrpc_server_allow high
1673 ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt, bool force)
1675 return ptlrpc_server_high_pending(svcpt, force) ||
1676 ptlrpc_server_normal_pending(svcpt, force);
1680 * Fetch a request for processing from queue of unprocessed requests.
1681 * Favors high-priority requests.
1682 * Returns a pointer to fetched request.
1684 static struct ptlrpc_request *
1685 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force)
1687 struct ptlrpc_request *req = NULL;
1689 spin_lock(&svcpt->scp_req_lock);
1691 if (ptlrpc_server_high_pending(svcpt, force)) {
1692 req = ptlrpc_nrs_req_get_nolock(svcpt, true, force);
1694 svcpt->scp_hreq_count++;
1699 if (ptlrpc_server_normal_pending(svcpt, force)) {
1700 req = ptlrpc_nrs_req_get_nolock(svcpt, false, force);
1702 svcpt->scp_hreq_count = 0;
1707 spin_unlock(&svcpt->scp_req_lock);
1711 svcpt->scp_nreqs_active++;
1713 svcpt->scp_nhreqs_active++;
1715 spin_unlock(&svcpt->scp_req_lock);
1717 if (likely(req->rq_export))
1718 class_export_rpc_inc(req->rq_export);
1724 * Handle freshly incoming reqs, add to timed early reply list,
1725 * pass on to regular request queue.
1726 * All incoming requests pass through here before getting into
1727 * ptlrpc_server_handle_req later on.
1730 ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt,
1731 struct ptlrpc_thread *thread)
1733 struct ptlrpc_service *svc = svcpt->scp_service;
1734 struct ptlrpc_request *req;
1738 spin_lock(&svcpt->scp_lock);
1739 if (list_empty(&svcpt->scp_req_incoming)) {
1740 spin_unlock(&svcpt->scp_lock);
1744 req = list_entry(svcpt->scp_req_incoming.next,
1745 struct ptlrpc_request, rq_list);
1746 list_del_init(&req->rq_list);
1747 svcpt->scp_nreqs_incoming--;
1748 /* Consider this still a "queued" request as far as stats are
1750 spin_unlock(&svcpt->scp_lock);
1752 /* go through security check/transform */
1753 rc = sptlrpc_svc_unwrap_request(req);
1757 case SECSVC_COMPLETE:
1758 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1767 * for null-flavored rpc, msg has been unpacked by sptlrpc, although
1768 * redo it wouldn't be harmful.
1770 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
1771 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
1773 CERROR("error unpacking request: ptl %d from %s "
1774 "x"LPU64"\n", svc->srv_req_portal,
1775 libcfs_id2str(req->rq_peer), req->rq_xid);
1780 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1782 CERROR("error unpacking ptlrpc body: ptl %d from %s x"
1783 LPU64"\n", svc->srv_req_portal,
1784 libcfs_id2str(req->rq_peer), req->rq_xid);
1788 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
1789 lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
1790 CERROR("drop incoming rpc opc %u, x"LPU64"\n",
1791 cfs_fail_val, req->rq_xid);
1796 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1797 CERROR("wrong packet type received (type=%u) from %s\n",
1798 lustre_msg_get_type(req->rq_reqmsg),
1799 libcfs_id2str(req->rq_peer));
1803 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1806 req->rq_bulk_write = 1;
1810 case MGS_CONFIG_READ:
1811 req->rq_bulk_read = 1;
1815 CDEBUG(D_RPCTRACE, "got req x"LPU64"\n", req->rq_xid);
1817 req->rq_export = class_conn2export(
1818 lustre_msg_get_handle(req->rq_reqmsg));
1819 if (req->rq_export) {
1820 rc = ptlrpc_check_req(req);
1822 rc = sptlrpc_target_export_check(req->rq_export, req);
1824 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
1825 "illegal security flavor,");
1830 ptlrpc_update_export_timer(req->rq_export, 0);
1833 /* req_in handling should/must be fast */
1834 if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5)
1835 DEBUG_REQ(D_WARNING, req, "Slow req_in handling "CFS_DURATION_T"s",
1836 cfs_time_sub(cfs_time_current_sec(),
1837 req->rq_arrival_time.tv_sec));
1839 /* Set rpc server deadline and add it to the timed list */
1840 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1841 MSGHDR_AT_SUPPORT) ?
1842 /* The max time the client expects us to take */
1843 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1844 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1845 if (unlikely(deadline == 0)) {
1846 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1850 req->rq_svc_thread = thread;
1852 ptlrpc_at_add_timed(req);
1854 /* Move it over to the request processing queue */
1855 rc = ptlrpc_server_request_add(svcpt, req);
1859 wake_up(&svcpt->scp_waitq);
1863 ptlrpc_server_finish_request(svcpt, req);
1869 * Main incoming request handling logic.
1870 * Calls handler function from service to do actual processing.
1873 ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
1874 struct ptlrpc_thread *thread)
1876 struct ptlrpc_service *svc = svcpt->scp_service;
1877 struct ptlrpc_request *request;
1878 struct timeval work_start;
1879 struct timeval work_end;
1884 request = ptlrpc_server_request_get(svcpt, false);
1885 if (request == NULL)
1888 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
1889 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
1890 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
1891 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
1893 if (unlikely(fail_opc)) {
1894 if (request->rq_export && request->rq_ops)
1895 OBD_FAIL_TIMEOUT(fail_opc, 4);
1898 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
1900 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
1901 libcfs_debug_dumplog();
1903 do_gettimeofday(&work_start);
1904 timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
1905 if (likely(svc->srv_stats != NULL)) {
1906 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
1908 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
1909 svcpt->scp_nreqs_incoming);
1910 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
1911 svcpt->scp_nreqs_active);
1912 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
1913 at_get(&svcpt->scp_at_estimate));
1916 rc = lu_context_init(&request->rq_session, LCT_SESSION | LCT_NOREF);
1918 CERROR("Failure to initialize session: %d\n", rc);
1921 request->rq_session.lc_thread = thread;
1922 request->rq_session.lc_cookie = 0x5;
1923 lu_context_enter(&request->rq_session);
1925 CDEBUG(D_NET, "got req "LPU64"\n", request->rq_xid);
1927 request->rq_svc_thread = thread;
1929 request->rq_svc_thread->t_env->le_ses = &request->rq_session;
1931 if (likely(request->rq_export)) {
1932 if (unlikely(ptlrpc_check_req(request)))
1934 ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
1937 /* Discard requests queued for longer than the deadline.
1938 The deadline is increased if we send an early reply. */
1939 if (cfs_time_current_sec() > request->rq_deadline) {
1940 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s"
1941 ": deadline "CFS_DURATION_T":"CFS_DURATION_T"s ago\n",
1942 libcfs_id2str(request->rq_peer),
1943 cfs_time_sub(request->rq_deadline,
1944 request->rq_arrival_time.tv_sec),
1945 cfs_time_sub(cfs_time_current_sec(),
1946 request->rq_deadline));
1950 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
1951 "%s:%s+%d:%d:x"LPU64":%s:%d\n", current_comm(),
1952 (request->rq_export ?
1953 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1954 (request->rq_export ?
1955 atomic_read(&request->rq_export->exp_refcount) : -99),
1956 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1957 libcfs_id2str(request->rq_peer),
1958 lustre_msg_get_opc(request->rq_reqmsg));
1960 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
1961 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
1963 rc = svc->srv_ops.so_req_handler(request);
1965 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
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,
1973 "Request took longer than estimated ("
1974 CFS_DURATION_T":"CFS_DURATION_T
1975 "s); client may timeout.",
1976 cfs_time_sub(request->rq_deadline,
1977 request->rq_arrival_time.tv_sec),
1978 cfs_time_sub(cfs_time_current_sec(),
1979 request->rq_deadline));
1982 do_gettimeofday(&work_end);
1983 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1984 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
1985 "%s:%s+%d:%d:x"LPU64":%s:%d Request procesed in "
1986 "%ldus (%ldus total) trans "LPU64" rc %d/%d\n",
1988 (request->rq_export ?
1989 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1990 (request->rq_export ?
1991 atomic_read(&request->rq_export->exp_refcount) : -99),
1992 lustre_msg_get_status(request->rq_reqmsg),
1994 libcfs_id2str(request->rq_peer),
1995 lustre_msg_get_opc(request->rq_reqmsg),
1997 cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL),
1998 (request->rq_repmsg ?
1999 lustre_msg_get_transno(request->rq_repmsg) :
2000 request->rq_transno),
2002 (request->rq_repmsg ?
2003 lustre_msg_get_status(request->rq_repmsg) : -999));
2004 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
2005 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
2006 int opc = opcode_offset(op);
2007 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
2008 LASSERT(opc < LUSTRE_MAX_OPCODES);
2009 lprocfs_counter_add(svc->srv_stats,
2010 opc + EXTRA_MAX_OPCODES,
2014 if (unlikely(request->rq_early_count)) {
2015 DEBUG_REQ(D_ADAPTTO, request,
2016 "sent %d early replies before finishing in "
2018 request->rq_early_count,
2019 cfs_time_sub(work_end.tv_sec,
2020 request->rq_arrival_time.tv_sec));
2024 ptlrpc_server_finish_active_request(svcpt, request);
2030 * An internal function to process a single reply state object.
2033 ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
2035 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
2036 struct ptlrpc_service *svc = svcpt->scp_service;
2037 struct obd_export *exp;
2041 exp = rs->rs_export;
2043 LASSERT(rs->rs_difficult);
2044 LASSERT(rs->rs_scheduled);
2045 LASSERT(list_empty(&rs->rs_list));
2047 spin_lock(&exp->exp_lock);
2048 /* Noop if removed already */
2049 list_del_init(&rs->rs_exp_list);
2050 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 spin_lock(&exp->exp_uncommitted_replies_lock);
2076 list_del_init(&rs->rs_obd_list);
2077 spin_unlock(&exp->exp_uncommitted_replies_lock);
2080 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 CDEBUG(D_HA, "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 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 completion */
2106 while (nlocks-- > 0)
2107 ldlm_lock_decref(&rs->rs_locks[nlocks],
2108 rs->rs_modes[nlocks]);
2110 spin_lock(&rs->rs_lock);
2113 rs->rs_scheduled = 0;
2115 if (!rs->rs_on_net) {
2117 spin_unlock(&rs->rs_lock);
2119 class_export_put(exp);
2120 rs->rs_export = NULL;
2121 ptlrpc_rs_decref(rs);
2122 if (atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
2123 svc->srv_is_stopping)
2124 wake_up_all(&svcpt->scp_waitq);
2128 /* still on the net; callback will schedule */
2129 spin_unlock(&rs->rs_lock);
2135 ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
2137 int avail = svcpt->scp_nrqbds_posted;
2138 int low_water = test_req_buffer_pressure ? 0 :
2139 svcpt->scp_service->srv_nbuf_per_group / 2;
2141 /* NB I'm not locking; just looking. */
2143 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
2144 * allowed the request history to grow out of control. We could put a
2145 * sanity check on that here and cull some history if we need the
2148 if (avail <= low_water)
2149 ptlrpc_grow_req_bufs(svcpt, 1);
2151 if (svcpt->scp_service->srv_stats) {
2152 lprocfs_counter_add(svcpt->scp_service->srv_stats,
2153 PTLRPC_REQBUF_AVAIL_CNTR, avail);
2158 ptlrpc_retry_rqbds(void *arg)
2160 struct ptlrpc_service_part *svcpt = (struct ptlrpc_service_part *)arg;
2162 svcpt->scp_rqbd_timeout = 0;
2167 ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
2169 return svcpt->scp_nreqs_active <
2170 svcpt->scp_nthrs_running - 1 -
2171 (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL);
2175 * allowed to create more threads
2176 * user can call it w/o any lock but need to hold
2177 * ptlrpc_service_part::scp_lock to get reliable result
2180 ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
2182 return svcpt->scp_nthrs_running +
2183 svcpt->scp_nthrs_starting <
2184 svcpt->scp_service->srv_nthrs_cpt_limit;
2188 * too many requests and allowed to create more threads
2191 ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
2193 return !ptlrpc_threads_enough(svcpt) &&
2194 ptlrpc_threads_increasable(svcpt);
2198 ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2200 return thread_is_stopping(thread) ||
2201 thread->t_svcpt->scp_service->srv_is_stopping;
2205 ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
2207 return !list_empty(&svcpt->scp_rqbd_idle) &&
2208 svcpt->scp_rqbd_timeout == 0;
2212 ptlrpc_at_check(struct ptlrpc_service_part *svcpt)
2214 return svcpt->scp_at_check;
2218 * requests wait on preprocessing
2219 * user can call it w/o any lock but need to hold
2220 * ptlrpc_service_part::scp_lock to get reliable result
2223 ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
2225 return !list_empty(&svcpt->scp_req_incoming);
2228 static __attribute__((__noinline__)) int
2229 ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
2230 struct ptlrpc_thread *thread)
2232 /* Don't exit while there are replies to be handled */
2233 struct l_wait_info lwi = LWI_TIMEOUT(svcpt->scp_rqbd_timeout,
2234 ptlrpc_retry_rqbds, svcpt);
2236 /* XXX: Add this back when libcfs watchdog is merged upstream
2237 lc_watchdog_disable(thread->t_watchdog);
2242 l_wait_event_exclusive_head(svcpt->scp_waitq,
2243 ptlrpc_thread_stopping(thread) ||
2244 ptlrpc_server_request_incoming(svcpt) ||
2245 ptlrpc_server_request_pending(svcpt, false) ||
2246 ptlrpc_rqbd_pending(svcpt) ||
2247 ptlrpc_at_check(svcpt), &lwi);
2249 if (ptlrpc_thread_stopping(thread))
2253 lc_watchdog_touch(thread->t_watchdog,
2254 ptlrpc_server_get_timeout(svcpt));
2260 * Main thread body for service threads.
2261 * Waits in a loop waiting for new requests to process to appear.
2262 * Every time an incoming requests is added to its queue, a waitq
2263 * is woken up and one of the threads will handle it.
2265 static int ptlrpc_main(void *arg)
2267 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
2268 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2269 struct ptlrpc_service *svc = svcpt->scp_service;
2270 struct ptlrpc_reply_state *rs;
2271 #ifdef WITH_GROUP_INFO
2272 struct group_info *ginfo = NULL;
2275 int counter = 0, rc = 0;
2277 thread->t_pid = current_pid();
2278 unshare_fs_struct();
2280 /* NB: we will call cfs_cpt_bind() for all threads, because we
2281 * might want to run lustre server only on a subset of system CPUs,
2282 * in that case ->scp_cpt is CFS_CPT_ANY */
2283 rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
2285 CWARN("%s: failed to bind %s on CPT %d\n",
2286 svc->srv_name, thread->t_name, svcpt->scp_cpt);
2289 #ifdef WITH_GROUP_INFO
2290 ginfo = groups_alloc(0);
2296 set_current_groups(ginfo);
2297 put_group_info(ginfo);
2300 if (svc->srv_ops.so_thr_init != NULL) {
2301 rc = svc->srv_ops.so_thr_init(thread);
2312 rc = lu_context_init(&env->le_ctx,
2313 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2317 thread->t_env = env;
2318 env->le_ctx.lc_thread = thread;
2319 env->le_ctx.lc_cookie = 0x6;
2321 while (!list_empty(&svcpt->scp_rqbd_idle)) {
2322 rc = ptlrpc_server_post_idle_rqbds(svcpt);
2326 CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
2327 svc->srv_name, svcpt->scp_cpt, rc);
2331 /* Alloc reply state structure for this one */
2332 OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2338 spin_lock(&svcpt->scp_lock);
2340 LASSERT(thread_is_starting(thread));
2341 thread_clear_flags(thread, SVC_STARTING);
2343 LASSERT(svcpt->scp_nthrs_starting == 1);
2344 svcpt->scp_nthrs_starting--;
2346 /* SVC_STOPPING may already be set here if someone else is trying
2347 * to stop the service while this new thread has been dynamically
2348 * forked. We still set SVC_RUNNING to let our creator know that
2349 * we are now running, however we will exit as soon as possible */
2350 thread_add_flags(thread, SVC_RUNNING);
2351 svcpt->scp_nthrs_running++;
2352 spin_unlock(&svcpt->scp_lock);
2354 /* wake up our creator in case he's still waiting. */
2355 wake_up(&thread->t_ctl_waitq);
2358 thread->t_watchdog = lc_watchdog_add(ptlrpc_server_get_timeout(svcpt),
2362 spin_lock(&svcpt->scp_rep_lock);
2363 list_add(&rs->rs_list, &svcpt->scp_rep_idle);
2364 wake_up(&svcpt->scp_rep_waitq);
2365 spin_unlock(&svcpt->scp_rep_lock);
2367 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2368 svcpt->scp_nthrs_running);
2370 /* XXX maintain a list of all managed devices: insert here */
2371 while (!ptlrpc_thread_stopping(thread)) {
2372 if (ptlrpc_wait_event(svcpt, thread))
2375 ptlrpc_check_rqbd_pool(svcpt);
2377 if (ptlrpc_threads_need_create(svcpt)) {
2378 /* Ignore return code - we tried... */
2379 ptlrpc_start_thread(svcpt, 0);
2382 /* Process all incoming reqs before handling any */
2383 if (ptlrpc_server_request_incoming(svcpt)) {
2384 lu_context_enter(&env->le_ctx);
2386 ptlrpc_server_handle_req_in(svcpt, thread);
2387 lu_context_exit(&env->le_ctx);
2389 /* but limit ourselves in case of flood */
2390 if (counter++ < 100)
2395 if (ptlrpc_at_check(svcpt))
2396 ptlrpc_at_check_timed(svcpt);
2398 if (ptlrpc_server_request_pending(svcpt, false)) {
2399 lu_context_enter(&env->le_ctx);
2400 ptlrpc_server_handle_request(svcpt, thread);
2401 lu_context_exit(&env->le_ctx);
2404 if (ptlrpc_rqbd_pending(svcpt) &&
2405 ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
2406 /* I just failed to repost request buffers.
2407 * Wait for a timeout (unless something else
2408 * happens) before I try again */
2409 svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
2410 CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
2411 svcpt->scp_nrqbds_posted);
2416 lc_watchdog_delete(thread->t_watchdog);
2417 thread->t_watchdog = NULL;
2422 * deconstruct service specific state created by ptlrpc_start_thread()
2424 if (svc->srv_ops.so_thr_done != NULL)
2425 svc->srv_ops.so_thr_done(thread);
2428 lu_context_fini(&env->le_ctx);
2432 CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n",
2433 thread, thread->t_pid, thread->t_id, rc);
2435 spin_lock(&svcpt->scp_lock);
2436 if (thread_test_and_clear_flags(thread, SVC_STARTING))
2437 svcpt->scp_nthrs_starting--;
2439 if (thread_test_and_clear_flags(thread, SVC_RUNNING)) {
2440 /* must know immediately */
2441 svcpt->scp_nthrs_running--;
2445 thread_add_flags(thread, SVC_STOPPED);
2447 wake_up(&thread->t_ctl_waitq);
2448 spin_unlock(&svcpt->scp_lock);
2453 static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
2454 struct list_head *replies)
2458 spin_lock(&hrt->hrt_lock);
2460 list_splice_init(&hrt->hrt_queue, replies);
2461 result = ptlrpc_hr.hr_stopping || !list_empty(replies);
2463 spin_unlock(&hrt->hrt_lock);
2468 * Main body of "handle reply" function.
2469 * It processes acked reply states
2471 static int ptlrpc_hr_main(void *arg)
2473 struct ptlrpc_hr_thread *hrt = (struct ptlrpc_hr_thread *)arg;
2474 struct ptlrpc_hr_partition *hrp = hrt->hrt_partition;
2475 LIST_HEAD (replies);
2476 char threadname[20];
2479 snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
2480 hrp->hrp_cpt, hrt->hrt_id);
2481 unshare_fs_struct();
2483 rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
2485 CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
2486 threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
2489 atomic_inc(&hrp->hrp_nstarted);
2490 wake_up(&ptlrpc_hr.hr_waitq);
2492 while (!ptlrpc_hr.hr_stopping) {
2493 l_wait_condition(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
2495 while (!list_empty(&replies)) {
2496 struct ptlrpc_reply_state *rs;
2498 rs = list_entry(replies.prev,
2499 struct ptlrpc_reply_state,
2501 list_del_init(&rs->rs_list);
2502 ptlrpc_handle_rs(rs);
2506 atomic_inc(&hrp->hrp_nstopped);
2507 wake_up(&ptlrpc_hr.hr_waitq);
2512 static void ptlrpc_stop_hr_threads(void)
2514 struct ptlrpc_hr_partition *hrp;
2518 ptlrpc_hr.hr_stopping = 1;
2520 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2521 if (hrp->hrp_thrs == NULL)
2522 continue; /* uninitialized */
2523 for (j = 0; j < hrp->hrp_nthrs; j++)
2524 wake_up_all(&hrp->hrp_thrs[j].hrt_waitq);
2527 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2528 if (hrp->hrp_thrs == NULL)
2529 continue; /* uninitialized */
2530 wait_event(ptlrpc_hr.hr_waitq,
2531 atomic_read(&hrp->hrp_nstopped) ==
2532 atomic_read(&hrp->hrp_nstarted));
2536 static int ptlrpc_start_hr_threads(void)
2538 struct ptlrpc_hr_partition *hrp;
2542 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2545 for (j = 0; j < hrp->hrp_nthrs; j++) {
2546 struct ptlrpc_hr_thread *hrt = &hrp->hrp_thrs[j];
2547 rc = PTR_ERR(kthread_run(ptlrpc_hr_main,
2549 "ptlrpc_hr%02d_%03d",
2552 if (IS_ERR_VALUE(rc))
2555 wait_event(ptlrpc_hr.hr_waitq,
2556 atomic_read(&hrp->hrp_nstarted) == j);
2557 if (!IS_ERR_VALUE(rc))
2560 CERROR("Reply handling thread %d:%d Failed on starting: "
2561 "rc = %d\n", i, j, rc);
2562 ptlrpc_stop_hr_threads();
2568 static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
2570 struct l_wait_info lwi = { 0 };
2571 struct ptlrpc_thread *thread;
2574 CDEBUG(D_INFO, "Stopping threads for service %s\n",
2575 svcpt->scp_service->srv_name);
2577 spin_lock(&svcpt->scp_lock);
2578 /* let the thread know that we would like it to stop asap */
2579 list_for_each_entry(thread, &svcpt->scp_threads, t_link) {
2580 CDEBUG(D_INFO, "Stopping thread %s #%u\n",
2581 svcpt->scp_service->srv_thread_name, thread->t_id);
2582 thread_add_flags(thread, SVC_STOPPING);
2585 wake_up_all(&svcpt->scp_waitq);
2587 while (!list_empty(&svcpt->scp_threads)) {
2588 thread = list_entry(svcpt->scp_threads.next,
2589 struct ptlrpc_thread, t_link);
2590 if (thread_is_stopped(thread)) {
2591 list_del(&thread->t_link);
2592 list_add(&thread->t_link, &zombie);
2595 spin_unlock(&svcpt->scp_lock);
2597 CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
2598 svcpt->scp_service->srv_thread_name, thread->t_id);
2599 l_wait_event(thread->t_ctl_waitq,
2600 thread_is_stopped(thread), &lwi);
2602 spin_lock(&svcpt->scp_lock);
2605 spin_unlock(&svcpt->scp_lock);
2607 while (!list_empty(&zombie)) {
2608 thread = list_entry(zombie.next,
2609 struct ptlrpc_thread, t_link);
2610 list_del(&thread->t_link);
2611 OBD_FREE_PTR(thread);
2616 * Stops all threads of a particular service \a svc
2618 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2620 struct ptlrpc_service_part *svcpt;
2623 ptlrpc_service_for_each_part(svcpt, i, svc) {
2624 if (svcpt->scp_service != NULL)
2625 ptlrpc_svcpt_stop_threads(svcpt);
2628 EXPORT_SYMBOL(ptlrpc_stop_all_threads);
2630 int ptlrpc_start_threads(struct ptlrpc_service *svc)
2636 /* We require 2 threads min, see note in ptlrpc_server_handle_request */
2637 LASSERT(svc->srv_nthrs_cpt_init >= PTLRPC_NTHRS_INIT);
2639 for (i = 0; i < svc->srv_ncpts; i++) {
2640 for (j = 0; j < svc->srv_nthrs_cpt_init; j++) {
2641 rc = ptlrpc_start_thread(svc->srv_parts[i], 1);
2647 /* We have enough threads, don't start more. b=15759 */
2654 CERROR("cannot start %s thread #%d_%d: rc %d\n",
2655 svc->srv_thread_name, i, j, rc);
2656 ptlrpc_stop_all_threads(svc);
2659 EXPORT_SYMBOL(ptlrpc_start_threads);
2661 int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
2663 struct l_wait_info lwi = { 0 };
2664 struct ptlrpc_thread *thread;
2665 struct ptlrpc_service *svc;
2668 LASSERT(svcpt != NULL);
2670 svc = svcpt->scp_service;
2672 CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n",
2673 svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running,
2674 svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit);
2677 if (unlikely(svc->srv_is_stopping))
2680 if (!ptlrpc_threads_increasable(svcpt) ||
2681 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
2682 svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1))
2685 OBD_CPT_ALLOC_PTR(thread, svc->srv_cptable, svcpt->scp_cpt);
2688 init_waitqueue_head(&thread->t_ctl_waitq);
2690 spin_lock(&svcpt->scp_lock);
2691 if (!ptlrpc_threads_increasable(svcpt)) {
2692 spin_unlock(&svcpt->scp_lock);
2693 OBD_FREE_PTR(thread);
2697 if (svcpt->scp_nthrs_starting != 0) {
2698 /* serialize starting because some modules (obdfilter)
2699 * might require unique and contiguous t_id */
2700 LASSERT(svcpt->scp_nthrs_starting == 1);
2701 spin_unlock(&svcpt->scp_lock);
2702 OBD_FREE_PTR(thread);
2704 CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n",
2705 svc->srv_thread_name, svcpt->scp_thr_nextid);
2710 CDEBUG(D_INFO, "Creating thread %s #%d race, retry later\n",
2711 svc->srv_thread_name, svcpt->scp_thr_nextid);
2715 svcpt->scp_nthrs_starting++;
2716 thread->t_id = svcpt->scp_thr_nextid++;
2717 thread_add_flags(thread, SVC_STARTING);
2718 thread->t_svcpt = svcpt;
2720 list_add(&thread->t_link, &svcpt->scp_threads);
2721 spin_unlock(&svcpt->scp_lock);
2723 if (svcpt->scp_cpt >= 0) {
2724 snprintf(thread->t_name, sizeof(thread->t_name), "%s%02d_%03d",
2725 svc->srv_thread_name, svcpt->scp_cpt, thread->t_id);
2727 snprintf(thread->t_name, sizeof(thread->t_name), "%s_%04d",
2728 svc->srv_thread_name, thread->t_id);
2731 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name);
2732 rc = PTR_ERR(kthread_run(ptlrpc_main, thread, "%s", thread->t_name));
2733 if (IS_ERR_VALUE(rc)) {
2734 CERROR("cannot start thread '%s': rc %d\n",
2735 thread->t_name, rc);
2736 spin_lock(&svcpt->scp_lock);
2737 --svcpt->scp_nthrs_starting;
2738 if (thread_is_stopping(thread)) {
2739 /* this ptlrpc_thread is being hanled
2740 * by ptlrpc_svcpt_stop_threads now
2742 thread_add_flags(thread, SVC_STOPPED);
2743 wake_up(&thread->t_ctl_waitq);
2744 spin_unlock(&svcpt->scp_lock);
2746 list_del(&thread->t_link);
2747 spin_unlock(&svcpt->scp_lock);
2748 OBD_FREE_PTR(thread);
2756 l_wait_event(thread->t_ctl_waitq,
2757 thread_is_running(thread) || thread_is_stopped(thread),
2760 rc = thread_is_stopped(thread) ? thread->t_id : 0;
2764 int ptlrpc_hr_init(void)
2767 struct ptlrpc_hr_partition *hrp;
2768 struct ptlrpc_hr_thread *hrt;
2774 memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
2775 ptlrpc_hr.hr_cpt_table = cfs_cpt_table;
2777 ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
2779 if (ptlrpc_hr.hr_partitions == NULL)
2782 init_waitqueue_head(&ptlrpc_hr.hr_waitq);
2784 cpumask_copy(&mask, topology_thread_cpumask(0));
2785 weight = cpus_weight(mask);
2787 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2790 atomic_set(&hrp->hrp_nstarted, 0);
2791 atomic_set(&hrp->hrp_nstopped, 0);
2793 hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, i);
2794 hrp->hrp_nthrs /= weight;
2796 LASSERT(hrp->hrp_nthrs > 0);
2797 OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, i,
2798 hrp->hrp_nthrs * sizeof(*hrt));
2799 if (hrp->hrp_thrs == NULL)
2800 GOTO(out, rc = -ENOMEM);
2802 for (j = 0; j < hrp->hrp_nthrs; j++) {
2803 hrt = &hrp->hrp_thrs[j];
2806 hrt->hrt_partition = hrp;
2807 init_waitqueue_head(&hrt->hrt_waitq);
2808 spin_lock_init(&hrt->hrt_lock);
2809 INIT_LIST_HEAD(&hrt->hrt_queue);
2813 rc = ptlrpc_start_hr_threads();
2820 void ptlrpc_hr_fini(void)
2822 struct ptlrpc_hr_partition *hrp;
2825 if (ptlrpc_hr.hr_partitions == NULL)
2828 ptlrpc_stop_hr_threads();
2830 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2831 if (hrp->hrp_thrs != NULL) {
2832 OBD_FREE(hrp->hrp_thrs,
2833 hrp->hrp_nthrs * sizeof(hrp->hrp_thrs[0]));
2837 cfs_percpt_free(ptlrpc_hr.hr_partitions);
2838 ptlrpc_hr.hr_partitions = NULL;
2843 * Wait until all already scheduled replies are processed.
2845 static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
2849 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
2852 rc = l_wait_event(svcpt->scp_waitq,
2853 atomic_read(&svcpt->scp_nreps_difficult) == 0, &lwi);
2856 CWARN("Unexpectedly long timeout %s %p\n",
2857 svcpt->scp_service->srv_name, svcpt->scp_service);
2862 ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
2864 struct ptlrpc_service_part *svcpt;
2867 /* early disarm AT timer... */
2868 ptlrpc_service_for_each_part(svcpt, i, svc) {
2869 if (svcpt->scp_service != NULL)
2870 cfs_timer_disarm(&svcpt->scp_at_timer);
2875 ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
2877 struct ptlrpc_service_part *svcpt;
2878 struct ptlrpc_request_buffer_desc *rqbd;
2879 struct l_wait_info lwi;
2883 /* All history will be culled when the next request buffer is
2884 * freed in ptlrpc_service_purge_all() */
2885 svc->srv_hist_nrqbds_cpt_max = 0;
2887 rc = LNetClearLazyPortal(svc->srv_req_portal);
2890 ptlrpc_service_for_each_part(svcpt, i, svc) {
2891 if (svcpt->scp_service == NULL)
2894 /* Unlink all the request buffers. This forces a 'final'
2895 * event with its 'unlink' flag set for each posted rqbd */
2896 list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
2898 rc = LNetMDUnlink(rqbd->rqbd_md_h);
2899 LASSERT(rc == 0 || rc == -ENOENT);
2903 ptlrpc_service_for_each_part(svcpt, i, svc) {
2904 if (svcpt->scp_service == NULL)
2907 /* Wait for the network to release any buffers
2908 * it's currently filling */
2909 spin_lock(&svcpt->scp_lock);
2910 while (svcpt->scp_nrqbds_posted != 0) {
2911 spin_unlock(&svcpt->scp_lock);
2912 /* Network access will complete in finite time but
2913 * the HUGE timeout lets us CWARN for visibility
2914 * of sluggish NALs */
2915 lwi = LWI_TIMEOUT_INTERVAL(
2916 cfs_time_seconds(LONG_UNLINK),
2917 cfs_time_seconds(1), NULL, NULL);
2918 rc = l_wait_event(svcpt->scp_waitq,
2919 svcpt->scp_nrqbds_posted == 0, &lwi);
2920 if (rc == -ETIMEDOUT) {
2921 CWARN("Service %s waiting for "
2922 "request buffers\n",
2923 svcpt->scp_service->srv_name);
2925 spin_lock(&svcpt->scp_lock);
2927 spin_unlock(&svcpt->scp_lock);
2932 ptlrpc_service_purge_all(struct ptlrpc_service *svc)
2934 struct ptlrpc_service_part *svcpt;
2935 struct ptlrpc_request_buffer_desc *rqbd;
2936 struct ptlrpc_request *req;
2937 struct ptlrpc_reply_state *rs;
2940 ptlrpc_service_for_each_part(svcpt, i, svc) {
2941 if (svcpt->scp_service == NULL)
2944 spin_lock(&svcpt->scp_rep_lock);
2945 while (!list_empty(&svcpt->scp_rep_active)) {
2946 rs = list_entry(svcpt->scp_rep_active.next,
2947 struct ptlrpc_reply_state, rs_list);
2948 spin_lock(&rs->rs_lock);
2949 ptlrpc_schedule_difficult_reply(rs);
2950 spin_unlock(&rs->rs_lock);
2952 spin_unlock(&svcpt->scp_rep_lock);
2954 /* purge the request queue. NB No new replies (rqbds
2955 * all unlinked) and no service threads, so I'm the only
2956 * thread noodling the request queue now */
2957 while (!list_empty(&svcpt->scp_req_incoming)) {
2958 req = list_entry(svcpt->scp_req_incoming.next,
2959 struct ptlrpc_request, rq_list);
2961 list_del(&req->rq_list);
2962 svcpt->scp_nreqs_incoming--;
2963 ptlrpc_server_finish_request(svcpt, req);
2966 while (ptlrpc_server_request_pending(svcpt, true)) {
2967 req = ptlrpc_server_request_get(svcpt, true);
2968 ptlrpc_server_finish_active_request(svcpt, req);
2971 LASSERT(list_empty(&svcpt->scp_rqbd_posted));
2972 LASSERT(svcpt->scp_nreqs_incoming == 0);
2973 LASSERT(svcpt->scp_nreqs_active == 0);
2974 /* history should have been culled by
2975 * ptlrpc_server_finish_request */
2976 LASSERT(svcpt->scp_hist_nrqbds == 0);
2978 /* Now free all the request buffers since nothing
2979 * references them any more... */
2981 while (!list_empty(&svcpt->scp_rqbd_idle)) {
2982 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
2983 struct ptlrpc_request_buffer_desc,
2985 ptlrpc_free_rqbd(rqbd);
2987 ptlrpc_wait_replies(svcpt);
2989 while (!list_empty(&svcpt->scp_rep_idle)) {
2990 rs = list_entry(svcpt->scp_rep_idle.next,
2991 struct ptlrpc_reply_state,
2993 list_del(&rs->rs_list);
2994 OBD_FREE_LARGE(rs, svc->srv_max_reply_size);
3000 ptlrpc_service_free(struct ptlrpc_service *svc)
3002 struct ptlrpc_service_part *svcpt;
3003 struct ptlrpc_at_array *array;
3006 ptlrpc_service_for_each_part(svcpt, i, svc) {
3007 if (svcpt->scp_service == NULL)
3010 /* In case somebody rearmed this in the meantime */
3011 cfs_timer_disarm(&svcpt->scp_at_timer);
3012 array = &svcpt->scp_at_array;
3014 if (array->paa_reqs_array != NULL) {
3015 OBD_FREE(array->paa_reqs_array,
3016 sizeof(struct list_head) * array->paa_size);
3017 array->paa_reqs_array = NULL;
3020 if (array->paa_reqs_count != NULL) {
3021 OBD_FREE(array->paa_reqs_count,
3022 sizeof(__u32) * array->paa_size);
3023 array->paa_reqs_count = NULL;
3027 ptlrpc_service_for_each_part(svcpt, i, svc)
3028 OBD_FREE_PTR(svcpt);
3030 if (svc->srv_cpts != NULL)
3031 cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts);
3033 OBD_FREE(svc, offsetof(struct ptlrpc_service,
3034 srv_parts[svc->srv_ncpts]));
3037 int ptlrpc_unregister_service(struct ptlrpc_service *service)
3039 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
3041 service->srv_is_stopping = 1;
3043 mutex_lock(&ptlrpc_all_services_mutex);
3044 list_del_init(&service->srv_list);
3045 mutex_unlock(&ptlrpc_all_services_mutex);
3047 ptlrpc_service_del_atimer(service);
3048 ptlrpc_stop_all_threads(service);
3050 ptlrpc_service_unlink_rqbd(service);
3051 ptlrpc_service_purge_all(service);
3052 ptlrpc_service_nrs_cleanup(service);
3054 ptlrpc_lprocfs_unregister_service(service);
3056 ptlrpc_service_free(service);
3060 EXPORT_SYMBOL(ptlrpc_unregister_service);
3063 * Returns 0 if the service is healthy.
3065 * Right now, it just checks to make sure that requests aren't languishing
3066 * in the queue. We'll use this health check to govern whether a node needs
3067 * to be shot, so it's intentionally non-aggressive. */
3068 int ptlrpc_svcpt_health_check(struct ptlrpc_service_part *svcpt)
3070 struct ptlrpc_request *request = NULL;
3071 struct timeval right_now;
3074 do_gettimeofday(&right_now);
3076 spin_lock(&svcpt->scp_req_lock);
3077 /* How long has the next entry been waiting? */
3078 if (ptlrpc_server_high_pending(svcpt, true))
3079 request = ptlrpc_nrs_req_peek_nolock(svcpt, true);
3080 else if (ptlrpc_server_normal_pending(svcpt, true))
3081 request = ptlrpc_nrs_req_peek_nolock(svcpt, false);
3083 if (request == NULL) {
3084 spin_unlock(&svcpt->scp_req_lock);
3088 timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
3089 spin_unlock(&svcpt->scp_req_lock);
3091 if ((timediff / ONE_MILLION) >
3092 (AT_OFF ? obd_timeout * 3 / 2 : at_max)) {
3093 CERROR("%s: unhealthy - request has been waiting %lds\n",
3094 svcpt->scp_service->srv_name, timediff / ONE_MILLION);
3102 ptlrpc_service_health_check(struct ptlrpc_service *svc)
3104 struct ptlrpc_service_part *svcpt;
3110 ptlrpc_service_for_each_part(svcpt, i, svc) {
3111 int rc = ptlrpc_svcpt_health_check(svcpt);
3118 EXPORT_SYMBOL(ptlrpc_service_health_check);