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 CFS_MODULE_PARM(test_req_buffer_pressure, "i", int, 0444,
48 "set non-zero to put pressure on request buffer pools");
49 CFS_MODULE_PARM(at_min, "i", int, 0644,
50 "Adaptive timeout minimum (sec)");
51 CFS_MODULE_PARM(at_max, "i", int, 0644,
52 "Adaptive timeout maximum (sec)");
53 CFS_MODULE_PARM(at_history, "i", int, 0644,
54 "Adaptive timeouts remember the slowest event that took place "
55 "within this period (sec)");
56 CFS_MODULE_PARM(at_early_margin, "i", int, 0644,
57 "How soon before an RPC deadline to send an early reply");
58 CFS_MODULE_PARM(at_extra, "i", int, 0644,
59 "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;
374 LASSERT(list_empty(&rs->rs_list));
376 hrt = ptlrpc_hr_select(rs->rs_svcpt);
378 spin_lock(&hrt->hrt_lock);
379 list_add_tail(&rs->rs_list, &hrt->hrt_queue);
380 spin_unlock(&hrt->hrt_lock);
382 wake_up(&hrt->hrt_waitq);
387 ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs)
391 LASSERT(spin_is_locked(&rs->rs_svcpt->scp_rep_lock));
392 LASSERT(spin_is_locked(&rs->rs_lock));
393 LASSERT (rs->rs_difficult);
394 rs->rs_scheduled_ever = 1; /* flag any notification attempt */
396 if (rs->rs_scheduled) { /* being set up or already notified */
401 rs->rs_scheduled = 1;
402 list_del_init(&rs->rs_list);
403 ptlrpc_dispatch_difficult_reply(rs);
406 EXPORT_SYMBOL(ptlrpc_schedule_difficult_reply);
408 void ptlrpc_commit_replies(struct obd_export *exp)
410 struct ptlrpc_reply_state *rs, *nxt;
411 DECLARE_RS_BATCH(batch);
414 rs_batch_init(&batch);
415 /* Find any replies that have been committed and get their service
416 * to attend to complete them. */
418 /* CAVEAT EMPTOR: spinlock ordering!!! */
419 spin_lock(&exp->exp_uncommitted_replies_lock);
420 list_for_each_entry_safe(rs, nxt, &exp->exp_uncommitted_replies,
422 LASSERT (rs->rs_difficult);
423 /* VBR: per-export last_committed */
424 LASSERT(rs->rs_export);
425 if (rs->rs_transno <= exp->exp_last_committed) {
426 list_del_init(&rs->rs_obd_list);
427 rs_batch_add(&batch, rs);
430 spin_unlock(&exp->exp_uncommitted_replies_lock);
431 rs_batch_fini(&batch);
434 EXPORT_SYMBOL(ptlrpc_commit_replies);
437 ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt)
439 struct ptlrpc_request_buffer_desc *rqbd;
444 spin_lock(&svcpt->scp_lock);
446 if (list_empty(&svcpt->scp_rqbd_idle)) {
447 spin_unlock(&svcpt->scp_lock);
451 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
452 struct ptlrpc_request_buffer_desc,
454 list_del(&rqbd->rqbd_list);
456 /* assume we will post successfully */
457 svcpt->scp_nrqbds_posted++;
458 list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_posted);
460 spin_unlock(&svcpt->scp_lock);
462 rc = ptlrpc_register_rqbd(rqbd);
469 spin_lock(&svcpt->scp_lock);
471 svcpt->scp_nrqbds_posted--;
472 list_del(&rqbd->rqbd_list);
473 list_add_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
475 /* Don't complain if no request buffers are posted right now; LNET
476 * won't drop requests because we set the portal lazy! */
478 spin_unlock(&svcpt->scp_lock);
483 static void ptlrpc_at_timer(unsigned long castmeharder)
485 struct ptlrpc_service_part *svcpt;
487 svcpt = (struct ptlrpc_service_part *)castmeharder;
489 svcpt->scp_at_check = 1;
490 svcpt->scp_at_checktime = cfs_time_current();
491 wake_up(&svcpt->scp_waitq);
495 ptlrpc_server_nthreads_check(struct ptlrpc_service *svc,
496 struct ptlrpc_service_conf *conf)
498 struct ptlrpc_service_thr_conf *tc = &conf->psc_thr;
505 * Common code for estimating & validating threads number.
506 * CPT affinity service could have percpt thread-pool instead
507 * of a global thread-pool, which means user might not always
508 * get the threads number they give it in conf::tc_nthrs_user
509 * even they did set. It's because we need to validate threads
510 * number for each CPT to guarantee each pool will have enough
511 * threads to keep the service healthy.
513 init = PTLRPC_NTHRS_INIT + (svc->srv_ops.so_hpreq_handler != NULL);
514 init = max_t(int, init, tc->tc_nthrs_init);
516 /* NB: please see comments in lustre_lnet.h for definition
517 * details of these members */
518 LASSERT(tc->tc_nthrs_max != 0);
520 if (tc->tc_nthrs_user != 0) {
521 /* In case there is a reason to test a service with many
522 * threads, we give a less strict check here, it can
523 * be up to 8 * nthrs_max */
524 total = min(tc->tc_nthrs_max * 8, tc->tc_nthrs_user);
525 nthrs = total / svc->srv_ncpts;
526 init = max(init, nthrs);
530 total = tc->tc_nthrs_max;
531 if (tc->tc_nthrs_base == 0) {
532 /* don't care about base threads number per partition,
533 * this is most for non-affinity service */
534 nthrs = total / svc->srv_ncpts;
538 nthrs = tc->tc_nthrs_base;
539 if (svc->srv_ncpts == 1) {
542 /* NB: Increase the base number if it's single partition
543 * and total number of cores/HTs is larger or equal to 4.
544 * result will always < 2 * nthrs_base */
545 weight = cfs_cpt_weight(svc->srv_cptable, CFS_CPT_ANY);
546 for (i = 1; (weight >> (i + 1)) != 0 && /* >= 4 cores/HTs */
547 (tc->tc_nthrs_base >> i) != 0; i++)
548 nthrs += tc->tc_nthrs_base >> i;
551 if (tc->tc_thr_factor != 0) {
552 int factor = tc->tc_thr_factor;
557 * User wants to increase number of threads with for
558 * each CPU core/HT, most likely the factor is larger then
559 * one thread/core because service threads are supposed to
560 * be blocked by lock or wait for IO.
563 * Amdahl's law says that adding processors wouldn't give
564 * a linear increasing of parallelism, so it's nonsense to
565 * have too many threads no matter how many cores/HTs
568 cpumask_copy(&mask, topology_thread_cpumask(0));
569 if (cpus_weight(mask) > 1) { /* weight is # of HTs */
570 /* depress thread factor for hyper-thread */
571 factor = factor - (factor >> 1) + (factor >> 3);
574 weight = cfs_cpt_weight(svc->srv_cptable, 0);
577 for (; factor > 0 && weight > 0; factor--, weight -= fade)
578 nthrs += min(weight, fade) * factor;
581 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
582 nthrs = max(tc->tc_nthrs_base,
583 tc->tc_nthrs_max / svc->srv_ncpts);
586 nthrs = max(nthrs, tc->tc_nthrs_init);
587 svc->srv_nthrs_cpt_limit = nthrs;
588 svc->srv_nthrs_cpt_init = init;
590 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
591 CDEBUG(D_OTHER, "%s: This service may have more threads (%d) "
592 "than the given soft limit (%d)\n",
593 svc->srv_name, nthrs * svc->srv_ncpts,
599 * Initialize percpt data for a service
602 ptlrpc_service_part_init(struct ptlrpc_service *svc,
603 struct ptlrpc_service_part *svcpt, int cpt)
605 struct ptlrpc_at_array *array;
610 svcpt->scp_cpt = cpt;
611 INIT_LIST_HEAD(&svcpt->scp_threads);
613 /* rqbd and incoming request queue */
614 spin_lock_init(&svcpt->scp_lock);
615 INIT_LIST_HEAD(&svcpt->scp_rqbd_idle);
616 INIT_LIST_HEAD(&svcpt->scp_rqbd_posted);
617 INIT_LIST_HEAD(&svcpt->scp_req_incoming);
618 init_waitqueue_head(&svcpt->scp_waitq);
619 /* history request & rqbd list */
620 INIT_LIST_HEAD(&svcpt->scp_hist_reqs);
621 INIT_LIST_HEAD(&svcpt->scp_hist_rqbds);
623 /* acitve requests and hp requests */
624 spin_lock_init(&svcpt->scp_req_lock);
627 spin_lock_init(&svcpt->scp_rep_lock);
628 INIT_LIST_HEAD(&svcpt->scp_rep_active);
629 INIT_LIST_HEAD(&svcpt->scp_rep_idle);
630 init_waitqueue_head(&svcpt->scp_rep_waitq);
631 atomic_set(&svcpt->scp_nreps_difficult, 0);
633 /* adaptive timeout */
634 spin_lock_init(&svcpt->scp_at_lock);
635 array = &svcpt->scp_at_array;
637 size = at_est2timeout(at_max);
638 array->paa_size = size;
639 array->paa_count = 0;
640 array->paa_deadline = -1;
642 /* allocate memory for scp_at_array (ptlrpc_at_array) */
643 OBD_CPT_ALLOC(array->paa_reqs_array,
644 svc->srv_cptable, cpt, sizeof(struct list_head) * size);
645 if (array->paa_reqs_array == NULL)
648 for (index = 0; index < size; index++)
649 INIT_LIST_HEAD(&array->paa_reqs_array[index]);
651 OBD_CPT_ALLOC(array->paa_reqs_count,
652 svc->srv_cptable, cpt, sizeof(__u32) * size);
653 if (array->paa_reqs_count == NULL)
656 cfs_timer_init(&svcpt->scp_at_timer, ptlrpc_at_timer, svcpt);
657 /* At SOW, service time should be quick; 10s seems generous. If client
658 * timeout is less than this, we'll be sending an early reply. */
659 at_init(&svcpt->scp_at_estimate, 10, 0);
661 /* assign this before call ptlrpc_grow_req_bufs */
662 svcpt->scp_service = svc;
663 /* Now allocate the request buffers, but don't post them now */
664 rc = ptlrpc_grow_req_bufs(svcpt, 0);
665 /* We shouldn't be under memory pressure at startup, so
666 * fail if we can't allocate all our buffers at this time. */
673 if (array->paa_reqs_count != NULL) {
674 OBD_FREE(array->paa_reqs_count, sizeof(__u32) * size);
675 array->paa_reqs_count = NULL;
678 if (array->paa_reqs_array != NULL) {
679 OBD_FREE(array->paa_reqs_array,
680 sizeof(struct list_head) * array->paa_size);
681 array->paa_reqs_array = NULL;
688 * Initialize service on a given portal.
689 * This includes starting serving threads , allocating and posting rqbds and
692 struct ptlrpc_service *
693 ptlrpc_register_service(struct ptlrpc_service_conf *conf,
694 proc_dir_entry_t *proc_entry)
696 struct ptlrpc_service_cpt_conf *cconf = &conf->psc_cpt;
697 struct ptlrpc_service *service;
698 struct ptlrpc_service_part *svcpt;
699 struct cfs_cpt_table *cptable;
707 LASSERT(conf->psc_buf.bc_nbufs > 0);
708 LASSERT(conf->psc_buf.bc_buf_size >=
709 conf->psc_buf.bc_req_max_size + SPTLRPC_MAX_PAYLOAD);
710 LASSERT(conf->psc_thr.tc_ctx_tags != 0);
712 cptable = cconf->cc_cptable;
714 cptable = cfs_cpt_table;
716 if (!conf->psc_thr.tc_cpu_affinity) {
719 ncpts = cfs_cpt_number(cptable);
720 if (cconf->cc_pattern != NULL) {
721 struct cfs_expr_list *el;
723 rc = cfs_expr_list_parse(cconf->cc_pattern,
724 strlen(cconf->cc_pattern),
727 CERROR("%s: invalid CPT pattern string: %s",
728 conf->psc_name, cconf->cc_pattern);
729 RETURN(ERR_PTR(-EINVAL));
732 rc = cfs_expr_list_values(el, ncpts, &cpts);
733 cfs_expr_list_free(el);
735 CERROR("%s: failed to parse CPT array %s: %d\n",
736 conf->psc_name, cconf->cc_pattern, rc);
738 OBD_FREE(cpts, sizeof(*cpts) * ncpts);
739 RETURN(ERR_PTR(rc < 0 ? rc : -EINVAL));
745 OBD_ALLOC(service, offsetof(struct ptlrpc_service, srv_parts[ncpts]));
746 if (service == NULL) {
748 OBD_FREE(cpts, sizeof(*cpts) * ncpts);
749 RETURN(ERR_PTR(-ENOMEM));
752 service->srv_cptable = cptable;
753 service->srv_cpts = cpts;
754 service->srv_ncpts = ncpts;
756 service->srv_cpt_bits = 0; /* it's zero already, easy to read... */
757 while ((1 << service->srv_cpt_bits) < cfs_cpt_number(cptable))
758 service->srv_cpt_bits++;
761 spin_lock_init(&service->srv_lock);
762 service->srv_name = conf->psc_name;
763 service->srv_watchdog_factor = conf->psc_watchdog_factor;
764 INIT_LIST_HEAD(&service->srv_list); /* for safty of cleanup */
766 /* buffer configuration */
767 service->srv_nbuf_per_group = test_req_buffer_pressure ?
768 1 : conf->psc_buf.bc_nbufs;
769 service->srv_max_req_size = conf->psc_buf.bc_req_max_size +
771 service->srv_buf_size = conf->psc_buf.bc_buf_size;
772 service->srv_rep_portal = conf->psc_buf.bc_rep_portal;
773 service->srv_req_portal = conf->psc_buf.bc_req_portal;
775 /* Increase max reply size to next power of two */
776 service->srv_max_reply_size = 1;
777 while (service->srv_max_reply_size <
778 conf->psc_buf.bc_rep_max_size + SPTLRPC_MAX_PAYLOAD)
779 service->srv_max_reply_size <<= 1;
781 service->srv_thread_name = conf->psc_thr.tc_thr_name;
782 service->srv_ctx_tags = conf->psc_thr.tc_ctx_tags;
783 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
784 service->srv_ops = conf->psc_ops;
786 for (i = 0; i < ncpts; i++) {
787 if (!conf->psc_thr.tc_cpu_affinity)
790 cpt = cpts != NULL ? cpts[i] : i;
792 OBD_CPT_ALLOC(svcpt, cptable, cpt, sizeof(*svcpt));
794 GOTO(failed, rc = -ENOMEM);
796 service->srv_parts[i] = svcpt;
797 rc = ptlrpc_service_part_init(service, svcpt, cpt);
802 ptlrpc_server_nthreads_check(service, conf);
804 rc = LNetSetLazyPortal(service->srv_req_portal);
807 mutex_lock(&ptlrpc_all_services_mutex);
808 list_add (&service->srv_list, &ptlrpc_all_services);
809 mutex_unlock(&ptlrpc_all_services_mutex);
811 if (proc_entry != NULL)
812 ptlrpc_lprocfs_register_service(proc_entry, service);
814 rc = ptlrpc_service_nrs_setup(service);
818 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
819 service->srv_name, service->srv_req_portal);
821 rc = ptlrpc_start_threads(service);
823 CERROR("Failed to start threads for service %s: %d\n",
824 service->srv_name, rc);
830 ptlrpc_unregister_service(service);
833 EXPORT_SYMBOL(ptlrpc_register_service);
836 * to actually free the request, must be called without holding svc_lock.
837 * note it's caller's responsibility to unlink req->rq_list.
839 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
841 LASSERT(atomic_read(&req->rq_refcount) == 0);
842 LASSERT(list_empty(&req->rq_timed_list));
844 /* DEBUG_REQ() assumes the reply state of a request with a valid
845 * ref will not be destroyed until that reference is dropped. */
846 ptlrpc_req_drop_rs(req);
848 sptlrpc_svc_ctx_decref(req);
850 if (req != &req->rq_rqbd->rqbd_req) {
851 /* NB request buffers use an embedded
852 * req if the incoming req unlinked the
853 * MD; this isn't one of them! */
854 OBD_FREE(req, sizeof(*req));
859 * drop a reference count of the request. if it reaches 0, we either
860 * put it into history list, or free it immediately.
862 void ptlrpc_server_drop_request(struct ptlrpc_request *req)
864 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
865 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
866 struct ptlrpc_service *svc = svcpt->scp_service;
868 struct list_head *tmp;
869 struct list_head *nxt;
871 if (!atomic_dec_and_test(&req->rq_refcount))
874 if (req->rq_at_linked) {
875 spin_lock(&svcpt->scp_at_lock);
876 /* recheck with lock, in case it's unlinked by
877 * ptlrpc_at_check_timed() */
878 if (likely(req->rq_at_linked))
879 ptlrpc_at_remove_timed(req);
880 spin_unlock(&svcpt->scp_at_lock);
883 LASSERT(list_empty(&req->rq_timed_list));
885 /* finalize request */
886 if (req->rq_export) {
887 class_export_put(req->rq_export);
888 req->rq_export = NULL;
891 spin_lock(&svcpt->scp_lock);
893 list_add(&req->rq_list, &rqbd->rqbd_reqs);
895 refcount = --(rqbd->rqbd_refcount);
897 /* request buffer is now idle: add to history */
898 list_del(&rqbd->rqbd_list);
900 list_add_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds);
901 svcpt->scp_hist_nrqbds++;
903 /* cull some history?
904 * I expect only about 1 or 2 rqbds need to be recycled here */
905 while (svcpt->scp_hist_nrqbds > svc->srv_hist_nrqbds_cpt_max) {
906 rqbd = list_entry(svcpt->scp_hist_rqbds.next,
907 struct ptlrpc_request_buffer_desc,
910 list_del(&rqbd->rqbd_list);
911 svcpt->scp_hist_nrqbds--;
913 /* remove rqbd's reqs from svc's req history while
914 * I've got the service lock */
915 list_for_each(tmp, &rqbd->rqbd_reqs) {
916 req = list_entry(tmp, struct ptlrpc_request,
918 /* Track the highest culled req seq */
919 if (req->rq_history_seq >
920 svcpt->scp_hist_seq_culled) {
921 svcpt->scp_hist_seq_culled =
924 list_del(&req->rq_history_list);
927 spin_unlock(&svcpt->scp_lock);
929 list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
930 req = list_entry(rqbd->rqbd_reqs.next,
931 struct ptlrpc_request,
933 list_del(&req->rq_list);
934 ptlrpc_server_free_request(req);
937 spin_lock(&svcpt->scp_lock);
939 * now all reqs including the embedded req has been
940 * disposed, schedule request buffer for re-use.
942 LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) ==
944 list_add_tail(&rqbd->rqbd_list,
945 &svcpt->scp_rqbd_idle);
948 spin_unlock(&svcpt->scp_lock);
949 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
950 /* If we are low on memory, we are not interested in history */
951 list_del(&req->rq_list);
952 list_del_init(&req->rq_history_list);
954 /* Track the highest culled req seq */
955 if (req->rq_history_seq > svcpt->scp_hist_seq_culled)
956 svcpt->scp_hist_seq_culled = req->rq_history_seq;
958 spin_unlock(&svcpt->scp_lock);
960 ptlrpc_server_free_request(req);
962 spin_unlock(&svcpt->scp_lock);
966 /** Change request export and move hp request from old export to new */
967 void ptlrpc_request_change_export(struct ptlrpc_request *req,
968 struct obd_export *export)
970 if (req->rq_export != NULL) {
971 if (!list_empty(&req->rq_exp_list)) {
972 /* remove rq_exp_list from last export */
973 spin_lock_bh(&req->rq_export->exp_rpc_lock);
974 list_del_init(&req->rq_exp_list);
975 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
977 /* export has one reference already, so it`s safe to
978 * add req to export queue here and get another
979 * reference for request later */
980 spin_lock_bh(&export->exp_rpc_lock);
981 list_add(&req->rq_exp_list, &export->exp_hp_rpcs);
982 spin_unlock_bh(&export->exp_rpc_lock);
984 class_export_rpc_dec(req->rq_export);
985 class_export_put(req->rq_export);
988 /* request takes one export refcount */
989 req->rq_export = class_export_get(export);
990 class_export_rpc_inc(export);
996 * to finish a request: stop sending more early replies, and release
999 static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt,
1000 struct ptlrpc_request *req)
1002 ptlrpc_server_hpreq_fini(req);
1004 ptlrpc_server_drop_request(req);
1008 * to finish a active request: stop sending more early replies, and release
1009 * the request. should be called after we finished handling the request.
1011 static void ptlrpc_server_finish_active_request(
1012 struct ptlrpc_service_part *svcpt,
1013 struct ptlrpc_request *req)
1015 spin_lock(&svcpt->scp_req_lock);
1016 ptlrpc_nrs_req_stop_nolock(req);
1017 svcpt->scp_nreqs_active--;
1019 svcpt->scp_nhreqs_active--;
1020 spin_unlock(&svcpt->scp_req_lock);
1022 ptlrpc_nrs_req_finalize(req);
1024 if (req->rq_export != NULL)
1025 class_export_rpc_dec(req->rq_export);
1027 ptlrpc_server_finish_request(svcpt, req);
1031 * This function makes sure dead exports are evicted in a timely manner.
1032 * This function is only called when some export receives a message (i.e.,
1033 * the network is up.)
1035 static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
1037 struct obd_export *oldest_exp;
1038 time_t oldest_time, new_time;
1044 /* Compensate for slow machines, etc, by faking our request time
1045 into the future. Although this can break the strict time-ordering
1046 of the list, we can be really lazy here - we don't have to evict
1047 at the exact right moment. Eventually, all silent exports
1048 will make it to the top of the list. */
1050 /* Do not pay attention on 1sec or smaller renewals. */
1051 new_time = cfs_time_current_sec() + extra_delay;
1052 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
1055 exp->exp_last_request_time = new_time;
1056 CDEBUG(D_HA, "updating export %s at "CFS_TIME_T" exp %p\n",
1057 exp->exp_client_uuid.uuid,
1058 exp->exp_last_request_time, exp);
1060 /* exports may get disconnected from the chain even though the
1061 export has references, so we must keep the spin lock while
1062 manipulating the lists */
1063 spin_lock(&exp->exp_obd->obd_dev_lock);
1065 if (list_empty(&exp->exp_obd_chain_timed)) {
1066 /* this one is not timed */
1067 spin_unlock(&exp->exp_obd->obd_dev_lock);
1071 list_move_tail(&exp->exp_obd_chain_timed,
1072 &exp->exp_obd->obd_exports_timed);
1074 oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
1075 struct obd_export, exp_obd_chain_timed);
1076 oldest_time = oldest_exp->exp_last_request_time;
1077 spin_unlock(&exp->exp_obd->obd_dev_lock);
1079 if (exp->exp_obd->obd_recovering) {
1080 /* be nice to everyone during recovery */
1085 /* Note - racing to start/reset the obd_eviction timer is safe */
1086 if (exp->exp_obd->obd_eviction_timer == 0) {
1087 /* Check if the oldest entry is expired. */
1088 if (cfs_time_current_sec() > (oldest_time + PING_EVICT_TIMEOUT +
1090 /* We need a second timer, in case the net was down and
1091 * it just came back. Since the pinger may skip every
1092 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
1093 * we better wait for 3. */
1094 exp->exp_obd->obd_eviction_timer =
1095 cfs_time_current_sec() + 3 * PING_INTERVAL;
1096 CDEBUG(D_HA, "%s: Think about evicting %s from "CFS_TIME_T"\n",
1097 exp->exp_obd->obd_name,
1098 obd_export_nid2str(oldest_exp), oldest_time);
1101 if (cfs_time_current_sec() >
1102 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
1103 /* The evictor won't evict anyone who we've heard from
1104 * recently, so we don't have to check before we start
1106 if (!ping_evictor_wake(exp))
1107 exp->exp_obd->obd_eviction_timer = 0;
1115 * Sanity check request \a req.
1116 * Return 0 if all is ok, error code otherwise.
1118 static int ptlrpc_check_req(struct ptlrpc_request *req)
1122 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
1123 req->rq_export->exp_conn_cnt)) {
1124 DEBUG_REQ(D_RPCTRACE, req,
1125 "DROPPING req from old connection %d < %d",
1126 lustre_msg_get_conn_cnt(req->rq_reqmsg),
1127 req->rq_export->exp_conn_cnt);
1130 if (unlikely(req->rq_export->exp_obd &&
1131 req->rq_export->exp_obd->obd_fail)) {
1132 /* Failing over, don't handle any more reqs, send
1133 error response instead. */
1134 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
1135 req, req->rq_export->exp_obd->obd_name);
1137 } else if (lustre_msg_get_flags(req->rq_reqmsg) &
1138 (MSG_REPLAY | MSG_REQ_REPLAY_DONE) &&
1139 !(req->rq_export->exp_obd->obd_recovering)) {
1140 DEBUG_REQ(D_ERROR, req,
1141 "Invalid replay without recovery");
1142 class_fail_export(req->rq_export);
1144 } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 &&
1145 !(req->rq_export->exp_obd->obd_recovering)) {
1146 DEBUG_REQ(D_ERROR, req, "Invalid req with transno "
1147 LPU64" without recovery",
1148 lustre_msg_get_transno(req->rq_reqmsg));
1149 class_fail_export(req->rq_export);
1153 if (unlikely(rc < 0)) {
1154 req->rq_status = rc;
1160 static void ptlrpc_at_set_timer(struct ptlrpc_service_part *svcpt)
1162 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1165 if (array->paa_count == 0) {
1166 cfs_timer_disarm(&svcpt->scp_at_timer);
1170 /* Set timer for closest deadline */
1171 next = (__s32)(array->paa_deadline - cfs_time_current_sec() -
1174 ptlrpc_at_timer((unsigned long)svcpt);
1176 cfs_timer_arm(&svcpt->scp_at_timer, cfs_time_shift(next));
1177 CDEBUG(D_INFO, "armed %s at %+ds\n",
1178 svcpt->scp_service->srv_name, next);
1182 /* Add rpc to early reply check list */
1183 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
1185 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1186 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1187 struct ptlrpc_request *rq = NULL;
1193 if (req->rq_no_reply)
1196 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
1199 spin_lock(&svcpt->scp_at_lock);
1200 LASSERT(list_empty(&req->rq_timed_list));
1202 index = (unsigned long)req->rq_deadline % array->paa_size;
1203 if (array->paa_reqs_count[index] > 0) {
1204 /* latest rpcs will have the latest deadlines in the list,
1205 * so search backward. */
1206 list_for_each_entry_reverse(rq,
1207 &array->paa_reqs_array[index],
1209 if (req->rq_deadline >= rq->rq_deadline) {
1210 list_add(&req->rq_timed_list,
1211 &rq->rq_timed_list);
1217 /* Add the request at the head of the list */
1218 if (list_empty(&req->rq_timed_list))
1219 list_add(&req->rq_timed_list,
1220 &array->paa_reqs_array[index]);
1222 spin_lock(&req->rq_lock);
1223 req->rq_at_linked = 1;
1224 spin_unlock(&req->rq_lock);
1225 req->rq_at_index = index;
1226 array->paa_reqs_count[index]++;
1228 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
1229 array->paa_deadline = req->rq_deadline;
1230 ptlrpc_at_set_timer(svcpt);
1232 spin_unlock(&svcpt->scp_at_lock);
1238 ptlrpc_at_remove_timed(struct ptlrpc_request *req)
1240 struct ptlrpc_at_array *array;
1242 array = &req->rq_rqbd->rqbd_svcpt->scp_at_array;
1244 /* NB: must call with hold svcpt::scp_at_lock */
1245 LASSERT(!list_empty(&req->rq_timed_list));
1246 list_del_init(&req->rq_timed_list);
1248 spin_lock(&req->rq_lock);
1249 req->rq_at_linked = 0;
1250 spin_unlock(&req->rq_lock);
1252 array->paa_reqs_count[req->rq_at_index]--;
1256 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
1258 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1259 struct ptlrpc_request *reqcopy;
1260 struct lustre_msg *reqmsg;
1261 cfs_duration_t olddl = req->rq_deadline - cfs_time_current_sec();
1266 /* deadline is when the client expects us to reply, margin is the
1267 difference between clients' and servers' expectations */
1268 DEBUG_REQ(D_ADAPTTO, req,
1269 "%ssending early reply (deadline %+lds, margin %+lds) for "
1270 "%d+%d", AT_OFF ? "AT off - not " : "",
1271 olddl, olddl - at_get(&svcpt->scp_at_estimate),
1272 at_get(&svcpt->scp_at_estimate), at_extra);
1278 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), "
1279 "not sending early reply. Consider increasing "
1280 "at_early_margin (%d)?", olddl, at_early_margin);
1282 /* Return an error so we're not re-added to the timed list. */
1286 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
1287 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
1288 "but no AT support");
1292 if (req->rq_export &&
1293 lustre_msg_get_flags(req->rq_reqmsg) &
1294 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1295 /* During recovery, we don't want to send too many early
1296 * replies, but on the other hand we want to make sure the
1297 * client has enough time to resend if the rpc is lost. So
1298 * during the recovery period send at least 4 early replies,
1299 * spacing them every at_extra if we can. at_estimate should
1300 * always equal this fixed value during recovery. */
1301 at_measured(&svcpt->scp_at_estimate, min(at_extra,
1302 req->rq_export->exp_obd->obd_recovery_timeout / 4));
1304 /* Fake our processing time into the future to ask the clients
1305 * for some extra amount of time */
1306 at_measured(&svcpt->scp_at_estimate, at_extra +
1307 cfs_time_current_sec() -
1308 req->rq_arrival_time.tv_sec);
1310 /* Check to see if we've actually increased the deadline -
1311 * we may be past adaptive_max */
1312 if (req->rq_deadline >= req->rq_arrival_time.tv_sec +
1313 at_get(&svcpt->scp_at_estimate)) {
1314 DEBUG_REQ(D_WARNING, req, "Couldn't add any time "
1315 "(%ld/%ld), not sending early reply\n",
1316 olddl, req->rq_arrival_time.tv_sec +
1317 at_get(&svcpt->scp_at_estimate) -
1318 cfs_time_current_sec());
1322 newdl = cfs_time_current_sec() + at_get(&svcpt->scp_at_estimate);
1324 OBD_ALLOC(reqcopy, sizeof *reqcopy);
1325 if (reqcopy == NULL)
1327 OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1329 OBD_FREE(reqcopy, sizeof *reqcopy);
1334 reqcopy->rq_reply_state = NULL;
1335 reqcopy->rq_rep_swab_mask = 0;
1336 reqcopy->rq_pack_bulk = 0;
1337 reqcopy->rq_pack_udesc = 0;
1338 reqcopy->rq_packed_final = 0;
1339 sptlrpc_svc_ctx_addref(reqcopy);
1340 /* We only need the reqmsg for the magic */
1341 reqcopy->rq_reqmsg = reqmsg;
1342 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1344 LASSERT(atomic_read(&req->rq_refcount));
1345 /** if it is last refcount then early reply isn't needed */
1346 if (atomic_read(&req->rq_refcount) == 1) {
1347 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
1348 "abort sending early reply\n");
1349 GOTO(out, rc = -EINVAL);
1352 /* Connection ref */
1353 reqcopy->rq_export = class_conn2export(
1354 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1355 if (reqcopy->rq_export == NULL)
1356 GOTO(out, rc = -ENODEV);
1359 class_export_rpc_inc(reqcopy->rq_export);
1360 if (reqcopy->rq_export->exp_obd &&
1361 reqcopy->rq_export->exp_obd->obd_fail)
1362 GOTO(out_put, rc = -ENODEV);
1364 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1368 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1371 /* Adjust our own deadline to what we told the client */
1372 req->rq_deadline = newdl;
1373 req->rq_early_count++; /* number sent, server side */
1375 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1378 /* Free the (early) reply state from lustre_pack_reply.
1379 (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
1380 ptlrpc_req_drop_rs(reqcopy);
1383 class_export_rpc_dec(reqcopy->rq_export);
1384 class_export_put(reqcopy->rq_export);
1386 sptlrpc_svc_ctx_decref(reqcopy);
1387 OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1388 OBD_FREE(reqcopy, sizeof *reqcopy);
1392 /* Send early replies to everybody expiring within at_early_margin
1393 asking for at_extra time */
1394 static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1396 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1397 struct ptlrpc_request *rq, *n;
1398 struct list_head work_list;
1401 time_t now = cfs_time_current_sec();
1402 cfs_duration_t delay;
1403 int first, counter = 0;
1406 spin_lock(&svcpt->scp_at_lock);
1407 if (svcpt->scp_at_check == 0) {
1408 spin_unlock(&svcpt->scp_at_lock);
1411 delay = cfs_time_sub(cfs_time_current(), svcpt->scp_at_checktime);
1412 svcpt->scp_at_check = 0;
1414 if (array->paa_count == 0) {
1415 spin_unlock(&svcpt->scp_at_lock);
1419 /* The timer went off, but maybe the nearest rpc already completed. */
1420 first = array->paa_deadline - now;
1421 if (first > at_early_margin) {
1422 /* We've still got plenty of time. Reset the timer. */
1423 ptlrpc_at_set_timer(svcpt);
1424 spin_unlock(&svcpt->scp_at_lock);
1428 /* We're close to a timeout, and we don't know how much longer the
1429 server will take. Send early replies to everyone expiring soon. */
1430 INIT_LIST_HEAD(&work_list);
1432 index = (unsigned long)array->paa_deadline % array->paa_size;
1433 count = array->paa_count;
1435 count -= array->paa_reqs_count[index];
1436 list_for_each_entry_safe(rq, n,
1437 &array->paa_reqs_array[index],
1439 if (rq->rq_deadline > now + at_early_margin) {
1440 /* update the earliest deadline */
1441 if (deadline == -1 ||
1442 rq->rq_deadline < deadline)
1443 deadline = rq->rq_deadline;
1447 ptlrpc_at_remove_timed(rq);
1449 * ptlrpc_server_drop_request() may drop
1450 * refcount to 0 already. Let's check this and
1451 * don't add entry to work_list
1453 if (likely(atomic_inc_not_zero(&rq->rq_refcount)))
1454 list_add(&rq->rq_timed_list, &work_list);
1458 if (++index >= array->paa_size)
1461 array->paa_deadline = deadline;
1462 /* we have a new earliest deadline, restart the timer */
1463 ptlrpc_at_set_timer(svcpt);
1465 spin_unlock(&svcpt->scp_at_lock);
1467 CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
1468 "replies\n", first, at_extra, counter);
1470 /* We're already past request deadlines before we even get a
1471 chance to send early replies */
1472 LCONSOLE_WARN("%s: This server is not able to keep up with "
1473 "request traffic (cpu-bound).\n",
1474 svcpt->scp_service->srv_name);
1475 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, "
1476 "delay="CFS_DURATION_T"(jiff)\n",
1477 counter, svcpt->scp_nreqs_incoming,
1478 svcpt->scp_nreqs_active,
1479 at_get(&svcpt->scp_at_estimate), delay);
1482 /* we took additional refcount so entries can't be deleted from list, no
1483 * locking is needed */
1484 while (!list_empty(&work_list)) {
1485 rq = list_entry(work_list.next, struct ptlrpc_request,
1487 list_del_init(&rq->rq_timed_list);
1489 if (ptlrpc_at_send_early_reply(rq) == 0)
1490 ptlrpc_at_add_timed(rq);
1492 ptlrpc_server_drop_request(rq);
1495 RETURN(1); /* return "did_something" for liblustre */
1499 * Put the request to the export list if the request may become
1500 * a high priority one.
1502 static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt,
1503 struct ptlrpc_request *req)
1508 if (svcpt->scp_service->srv_ops.so_hpreq_handler) {
1509 rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
1514 if (req->rq_export && req->rq_ops) {
1515 /* Perform request specific check. We should do this check
1516 * before the request is added into exp_hp_rpcs list otherwise
1517 * it may hit swab race at LU-1044. */
1518 if (req->rq_ops->hpreq_check) {
1519 rc = req->rq_ops->hpreq_check(req);
1521 * XXX: Out of all current
1522 * ptlrpc_hpreq_ops::hpreq_check(), only
1523 * ldlm_cancel_hpreq_check() can return an error code;
1524 * other functions assert in similar places, which seems
1525 * odd. What also does not seem right is that handlers
1526 * for those RPCs do not assert on the same checks, but
1527 * rather handle the error cases. e.g. see
1528 * ost_rw_hpreq_check(), and ost_brw_read(),
1533 LASSERT(rc == 0 || rc == 1);
1536 spin_lock_bh(&req->rq_export->exp_rpc_lock);
1537 list_add(&req->rq_exp_list,
1538 &req->rq_export->exp_hp_rpcs);
1539 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1542 ptlrpc_nrs_req_initialize(svcpt, req, rc);
1547 /** Remove the request from the export list. */
1548 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req)
1551 if (req->rq_export && req->rq_ops) {
1552 /* refresh lock timeout again so that client has more
1553 * room to send lock cancel RPC. */
1554 if (req->rq_ops->hpreq_fini)
1555 req->rq_ops->hpreq_fini(req);
1557 spin_lock_bh(&req->rq_export->exp_rpc_lock);
1558 list_del_init(&req->rq_exp_list);
1559 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1564 static int ptlrpc_hpreq_check(struct ptlrpc_request *req)
1569 static struct ptlrpc_hpreq_ops ptlrpc_hpreq_common = {
1570 .hpreq_check = ptlrpc_hpreq_check,
1573 /* Hi-Priority RPC check by RPC operation code. */
1574 int ptlrpc_hpreq_handler(struct ptlrpc_request *req)
1576 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1578 /* Check for export to let only reconnects for not yet evicted
1579 * export to become a HP rpc. */
1580 if ((req->rq_export != NULL) &&
1581 (opc == OBD_PING || opc == MDS_CONNECT || opc == OST_CONNECT))
1582 req->rq_ops = &ptlrpc_hpreq_common;
1586 EXPORT_SYMBOL(ptlrpc_hpreq_handler);
1588 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1589 struct ptlrpc_request *req)
1594 rc = ptlrpc_server_hpreq_init(svcpt, req);
1598 ptlrpc_nrs_req_add(svcpt, req, !!rc);
1604 * Allow to handle high priority request
1605 * User can call it w/o any lock but need to hold
1606 * ptlrpc_service_part::scp_req_lock to get reliable result
1608 static bool ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1611 int running = svcpt->scp_nthrs_running;
1613 if (!nrs_svcpt_has_hp(svcpt))
1619 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1620 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1621 /* leave just 1 thread for normal RPCs */
1622 running = PTLRPC_NTHRS_INIT;
1623 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1627 if (svcpt->scp_nreqs_active >= running - 1)
1630 if (svcpt->scp_nhreqs_active == 0)
1633 return !ptlrpc_nrs_req_pending_nolock(svcpt, false) ||
1634 svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1637 static bool ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1640 return ptlrpc_server_allow_high(svcpt, force) &&
1641 ptlrpc_nrs_req_pending_nolock(svcpt, true);
1645 * Only allow normal priority requests on a service that has a high-priority
1646 * queue if forced (i.e. cleanup), if there are other high priority requests
1647 * already being processed (i.e. those threads can service more high-priority
1648 * requests), or if there are enough idle threads that a later thread can do
1649 * a high priority request.
1650 * User can call it w/o any lock but need to hold
1651 * ptlrpc_service_part::scp_req_lock to get reliable result
1653 static bool ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1656 int running = svcpt->scp_nthrs_running;
1657 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1658 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1659 /* leave just 1 thread for normal RPCs */
1660 running = PTLRPC_NTHRS_INIT;
1661 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1666 svcpt->scp_nreqs_active < running - 2)
1669 if (svcpt->scp_nreqs_active >= running - 1)
1672 return svcpt->scp_nhreqs_active > 0 || !nrs_svcpt_has_hp(svcpt);
1675 static bool ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1678 return ptlrpc_server_allow_normal(svcpt, force) &&
1679 ptlrpc_nrs_req_pending_nolock(svcpt, false);
1683 * Returns true if there are requests available in incoming
1684 * request queue for processing and it is allowed to fetch them.
1685 * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1686 * to get reliable result
1687 * \see ptlrpc_server_allow_normal
1688 * \see ptlrpc_server_allow high
1691 ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt, bool force)
1693 return ptlrpc_server_high_pending(svcpt, force) ||
1694 ptlrpc_server_normal_pending(svcpt, force);
1698 * Fetch a request for processing from queue of unprocessed requests.
1699 * Favors high-priority requests.
1700 * Returns a pointer to fetched request.
1702 static struct ptlrpc_request *
1703 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force)
1705 struct ptlrpc_request *req = NULL;
1708 spin_lock(&svcpt->scp_req_lock);
1710 if (ptlrpc_server_high_pending(svcpt, force)) {
1711 req = ptlrpc_nrs_req_get_nolock(svcpt, true, force);
1713 svcpt->scp_hreq_count++;
1718 if (ptlrpc_server_normal_pending(svcpt, force)) {
1719 req = ptlrpc_nrs_req_get_nolock(svcpt, false, force);
1721 svcpt->scp_hreq_count = 0;
1726 spin_unlock(&svcpt->scp_req_lock);
1730 svcpt->scp_nreqs_active++;
1732 svcpt->scp_nhreqs_active++;
1734 spin_unlock(&svcpt->scp_req_lock);
1736 if (likely(req->rq_export))
1737 class_export_rpc_inc(req->rq_export);
1743 * Handle freshly incoming reqs, add to timed early reply list,
1744 * pass on to regular request queue.
1745 * All incoming requests pass through here before getting into
1746 * ptlrpc_server_handle_req later on.
1749 ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt,
1750 struct ptlrpc_thread *thread)
1752 struct ptlrpc_service *svc = svcpt->scp_service;
1753 struct ptlrpc_request *req;
1758 spin_lock(&svcpt->scp_lock);
1759 if (list_empty(&svcpt->scp_req_incoming)) {
1760 spin_unlock(&svcpt->scp_lock);
1764 req = list_entry(svcpt->scp_req_incoming.next,
1765 struct ptlrpc_request, rq_list);
1766 list_del_init(&req->rq_list);
1767 svcpt->scp_nreqs_incoming--;
1768 /* Consider this still a "queued" request as far as stats are
1770 spin_unlock(&svcpt->scp_lock);
1772 /* go through security check/transform */
1773 rc = sptlrpc_svc_unwrap_request(req);
1777 case SECSVC_COMPLETE:
1778 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1787 * for null-flavored rpc, msg has been unpacked by sptlrpc, although
1788 * redo it wouldn't be harmful.
1790 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
1791 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
1793 CERROR("error unpacking request: ptl %d from %s "
1794 "x"LPU64"\n", svc->srv_req_portal,
1795 libcfs_id2str(req->rq_peer), req->rq_xid);
1800 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1802 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
1803 LPU64"\n", svc->srv_req_portal,
1804 libcfs_id2str(req->rq_peer), req->rq_xid);
1808 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
1809 lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
1810 CERROR("drop incoming rpc opc %u, x"LPU64"\n",
1811 cfs_fail_val, req->rq_xid);
1816 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1817 CERROR("wrong packet type received (type=%u) from %s\n",
1818 lustre_msg_get_type(req->rq_reqmsg),
1819 libcfs_id2str(req->rq_peer));
1823 switch(lustre_msg_get_opc(req->rq_reqmsg)) {
1826 req->rq_bulk_write = 1;
1830 case MGS_CONFIG_READ:
1831 req->rq_bulk_read = 1;
1835 CDEBUG(D_RPCTRACE, "got req x"LPU64"\n", req->rq_xid);
1837 req->rq_export = class_conn2export(
1838 lustre_msg_get_handle(req->rq_reqmsg));
1839 if (req->rq_export) {
1840 rc = ptlrpc_check_req(req);
1842 rc = sptlrpc_target_export_check(req->rq_export, req);
1844 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
1845 "illegal security flavor,");
1850 ptlrpc_update_export_timer(req->rq_export, 0);
1853 /* req_in handling should/must be fast */
1854 if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5)
1855 DEBUG_REQ(D_WARNING, req, "Slow req_in handling "CFS_DURATION_T"s",
1856 cfs_time_sub(cfs_time_current_sec(),
1857 req->rq_arrival_time.tv_sec));
1859 /* Set rpc server deadline and add it to the timed list */
1860 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1861 MSGHDR_AT_SUPPORT) ?
1862 /* The max time the client expects us to take */
1863 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1864 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1865 if (unlikely(deadline == 0)) {
1866 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1870 req->rq_svc_thread = thread;
1872 ptlrpc_at_add_timed(req);
1874 /* Move it over to the request processing queue */
1875 rc = ptlrpc_server_request_add(svcpt, req);
1879 wake_up(&svcpt->scp_waitq);
1883 ptlrpc_server_finish_request(svcpt, req);
1889 * Main incoming request handling logic.
1890 * Calls handler function from service to do actual processing.
1893 ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
1894 struct ptlrpc_thread *thread)
1896 struct ptlrpc_service *svc = svcpt->scp_service;
1897 struct ptlrpc_request *request;
1898 struct timeval work_start;
1899 struct timeval work_end;
1905 request = ptlrpc_server_request_get(svcpt, false);
1906 if (request == NULL)
1909 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
1910 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
1911 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
1912 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
1914 if (unlikely(fail_opc)) {
1915 if (request->rq_export && request->rq_ops)
1916 OBD_FAIL_TIMEOUT(fail_opc, 4);
1919 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
1921 if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
1922 libcfs_debug_dumplog();
1924 do_gettimeofday(&work_start);
1925 timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
1926 if (likely(svc->srv_stats != NULL)) {
1927 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
1929 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
1930 svcpt->scp_nreqs_incoming);
1931 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
1932 svcpt->scp_nreqs_active);
1933 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
1934 at_get(&svcpt->scp_at_estimate));
1937 rc = lu_context_init(&request->rq_session, LCT_SESSION | LCT_NOREF);
1939 CERROR("Failure to initialize session: %d\n", rc);
1942 request->rq_session.lc_thread = thread;
1943 request->rq_session.lc_cookie = 0x5;
1944 lu_context_enter(&request->rq_session);
1946 CDEBUG(D_NET, "got req "LPU64"\n", request->rq_xid);
1948 request->rq_svc_thread = thread;
1950 request->rq_svc_thread->t_env->le_ses = &request->rq_session;
1952 if (likely(request->rq_export)) {
1953 if (unlikely(ptlrpc_check_req(request)))
1955 ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
1958 /* Discard requests queued for longer than the deadline.
1959 The deadline is increased if we send an early reply. */
1960 if (cfs_time_current_sec() > request->rq_deadline) {
1961 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s"
1962 ": deadline "CFS_DURATION_T":"CFS_DURATION_T"s ago\n",
1963 libcfs_id2str(request->rq_peer),
1964 cfs_time_sub(request->rq_deadline,
1965 request->rq_arrival_time.tv_sec),
1966 cfs_time_sub(cfs_time_current_sec(),
1967 request->rq_deadline));
1971 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
1972 "%s:%s+%d:%d:x"LPU64":%s:%d\n", current_comm(),
1973 (request->rq_export ?
1974 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1975 (request->rq_export ?
1976 atomic_read(&request->rq_export->exp_refcount) : -99),
1977 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1978 libcfs_id2str(request->rq_peer),
1979 lustre_msg_get_opc(request->rq_reqmsg));
1981 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
1982 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
1984 rc = svc->srv_ops.so_req_handler(request);
1986 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
1989 lu_context_exit(&request->rq_session);
1990 lu_context_fini(&request->rq_session);
1992 if (unlikely(cfs_time_current_sec() > request->rq_deadline)) {
1993 DEBUG_REQ(D_WARNING, request, "Request took longer "
1994 "than estimated ("CFS_DURATION_T":"CFS_DURATION_T"s);"
1995 " client may timeout.",
1996 cfs_time_sub(request->rq_deadline,
1997 request->rq_arrival_time.tv_sec),
1998 cfs_time_sub(cfs_time_current_sec(),
1999 request->rq_deadline));
2002 do_gettimeofday(&work_end);
2003 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
2004 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
2005 "%s:%s+%d:%d:x"LPU64":%s:%d Request procesed in "
2006 "%ldus (%ldus total) trans "LPU64" rc %d/%d\n",
2008 (request->rq_export ?
2009 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2010 (request->rq_export ?
2011 atomic_read(&request->rq_export->exp_refcount) : -99),
2012 lustre_msg_get_status(request->rq_reqmsg),
2014 libcfs_id2str(request->rq_peer),
2015 lustre_msg_get_opc(request->rq_reqmsg),
2017 cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL),
2018 (request->rq_repmsg ?
2019 lustre_msg_get_transno(request->rq_repmsg) :
2020 request->rq_transno),
2022 (request->rq_repmsg ?
2023 lustre_msg_get_status(request->rq_repmsg) : -999));
2024 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
2025 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
2026 int opc = opcode_offset(op);
2027 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
2028 LASSERT(opc < LUSTRE_MAX_OPCODES);
2029 lprocfs_counter_add(svc->srv_stats,
2030 opc + EXTRA_MAX_OPCODES,
2034 if (unlikely(request->rq_early_count)) {
2035 DEBUG_REQ(D_ADAPTTO, request,
2036 "sent %d early replies before finishing in "
2038 request->rq_early_count,
2039 cfs_time_sub(work_end.tv_sec,
2040 request->rq_arrival_time.tv_sec));
2044 ptlrpc_server_finish_active_request(svcpt, request);
2050 * An internal function to process a single reply state object.
2053 ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
2055 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
2056 struct ptlrpc_service *svc = svcpt->scp_service;
2057 struct obd_export *exp;
2062 exp = rs->rs_export;
2064 LASSERT (rs->rs_difficult);
2065 LASSERT (rs->rs_scheduled);
2066 LASSERT (list_empty(&rs->rs_list));
2068 spin_lock(&exp->exp_lock);
2069 /* Noop if removed already */
2070 list_del_init (&rs->rs_exp_list);
2071 spin_unlock(&exp->exp_lock);
2073 /* The disk commit callback holds exp_uncommitted_replies_lock while it
2074 * iterates over newly committed replies, removing them from
2075 * exp_uncommitted_replies. It then drops this lock and schedules the
2076 * replies it found for handling here.
2078 * We can avoid contention for exp_uncommitted_replies_lock between the
2079 * HRT threads and further commit callbacks by checking rs_committed
2080 * which is set in the commit callback while it holds both
2081 * rs_lock and exp_uncommitted_reples.
2083 * If we see rs_committed clear, the commit callback _may_ not have
2084 * handled this reply yet and we race with it to grab
2085 * exp_uncommitted_replies_lock before removing the reply from
2086 * exp_uncommitted_replies. Note that if we lose the race and the
2087 * reply has already been removed, list_del_init() is a noop.
2089 * If we see rs_committed set, we know the commit callback is handling,
2090 * or has handled this reply since store reordering might allow us to
2091 * see rs_committed set out of sequence. But since this is done
2092 * holding rs_lock, we can be sure it has all completed once we hold
2093 * rs_lock, which we do right next.
2095 if (!rs->rs_committed) {
2096 spin_lock(&exp->exp_uncommitted_replies_lock);
2097 list_del_init(&rs->rs_obd_list);
2098 spin_unlock(&exp->exp_uncommitted_replies_lock);
2101 spin_lock(&rs->rs_lock);
2103 been_handled = rs->rs_handled;
2106 nlocks = rs->rs_nlocks; /* atomic "steal", but */
2107 rs->rs_nlocks = 0; /* locks still on rs_locks! */
2109 if (nlocks == 0 && !been_handled) {
2110 /* If we see this, we should already have seen the warning
2111 * in mds_steal_ack_locks() */
2112 CDEBUG(D_HA, "All locks stolen from rs %p x"LPD64".t"LPD64
2115 rs->rs_xid, rs->rs_transno, rs->rs_opc,
2116 libcfs_nid2str(exp->exp_connection->c_peer.nid));
2119 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
2120 spin_unlock(&rs->rs_lock);
2122 if (!been_handled && rs->rs_on_net) {
2123 LNetMDUnlink(rs->rs_md_h);
2124 /* Ignore return code; we're racing with completion */
2127 while (nlocks-- > 0)
2128 ldlm_lock_decref(&rs->rs_locks[nlocks],
2129 rs->rs_modes[nlocks]);
2131 spin_lock(&rs->rs_lock);
2134 rs->rs_scheduled = 0;
2136 if (!rs->rs_on_net) {
2138 spin_unlock(&rs->rs_lock);
2140 class_export_put (exp);
2141 rs->rs_export = NULL;
2142 ptlrpc_rs_decref (rs);
2143 if (atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
2144 svc->srv_is_stopping)
2145 wake_up_all(&svcpt->scp_waitq);
2149 /* still on the net; callback will schedule */
2150 spin_unlock(&rs->rs_lock);
2156 ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
2158 int avail = svcpt->scp_nrqbds_posted;
2159 int low_water = test_req_buffer_pressure ? 0 :
2160 svcpt->scp_service->srv_nbuf_per_group / 2;
2162 /* NB I'm not locking; just looking. */
2164 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
2165 * allowed the request history to grow out of control. We could put a
2166 * sanity check on that here and cull some history if we need the
2169 if (avail <= low_water)
2170 ptlrpc_grow_req_bufs(svcpt, 1);
2172 if (svcpt->scp_service->srv_stats) {
2173 lprocfs_counter_add(svcpt->scp_service->srv_stats,
2174 PTLRPC_REQBUF_AVAIL_CNTR, avail);
2179 ptlrpc_retry_rqbds(void *arg)
2181 struct ptlrpc_service_part *svcpt = (struct ptlrpc_service_part *)arg;
2183 svcpt->scp_rqbd_timeout = 0;
2188 ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
2190 return svcpt->scp_nreqs_active <
2191 svcpt->scp_nthrs_running - 1 -
2192 (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL);
2196 * allowed to create more threads
2197 * user can call it w/o any lock but need to hold
2198 * ptlrpc_service_part::scp_lock to get reliable result
2201 ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
2203 return svcpt->scp_nthrs_running +
2204 svcpt->scp_nthrs_starting <
2205 svcpt->scp_service->srv_nthrs_cpt_limit;
2209 * too many requests and allowed to create more threads
2212 ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
2214 return !ptlrpc_threads_enough(svcpt) &&
2215 ptlrpc_threads_increasable(svcpt);
2219 ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2221 return thread_is_stopping(thread) ||
2222 thread->t_svcpt->scp_service->srv_is_stopping;
2226 ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
2228 return !list_empty(&svcpt->scp_rqbd_idle) &&
2229 svcpt->scp_rqbd_timeout == 0;
2233 ptlrpc_at_check(struct ptlrpc_service_part *svcpt)
2235 return svcpt->scp_at_check;
2239 * requests wait on preprocessing
2240 * user can call it w/o any lock but need to hold
2241 * ptlrpc_service_part::scp_lock to get reliable result
2244 ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
2246 return !list_empty(&svcpt->scp_req_incoming);
2249 static __attribute__((__noinline__)) int
2250 ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
2251 struct ptlrpc_thread *thread)
2253 /* Don't exit while there are replies to be handled */
2254 struct l_wait_info lwi = LWI_TIMEOUT(svcpt->scp_rqbd_timeout,
2255 ptlrpc_retry_rqbds, svcpt);
2257 /* XXX: Add this back when libcfs watchdog is merged upstream
2258 lc_watchdog_disable(thread->t_watchdog);
2263 l_wait_event_exclusive_head(svcpt->scp_waitq,
2264 ptlrpc_thread_stopping(thread) ||
2265 ptlrpc_server_request_incoming(svcpt) ||
2266 ptlrpc_server_request_pending(svcpt, false) ||
2267 ptlrpc_rqbd_pending(svcpt) ||
2268 ptlrpc_at_check(svcpt), &lwi);
2270 if (ptlrpc_thread_stopping(thread))
2274 lc_watchdog_touch(thread->t_watchdog,
2275 ptlrpc_server_get_timeout(svcpt));
2281 * Main thread body for service threads.
2282 * Waits in a loop waiting for new requests to process to appear.
2283 * Every time an incoming requests is added to its queue, a waitq
2284 * is woken up and one of the threads will handle it.
2286 static int ptlrpc_main(void *arg)
2288 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
2289 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2290 struct ptlrpc_service *svc = svcpt->scp_service;
2291 struct ptlrpc_reply_state *rs;
2292 #ifdef WITH_GROUP_INFO
2293 group_info_t *ginfo = NULL;
2296 int counter = 0, rc = 0;
2299 thread->t_pid = current_pid();
2300 unshare_fs_struct();
2302 /* NB: we will call cfs_cpt_bind() for all threads, because we
2303 * might want to run lustre server only on a subset of system CPUs,
2304 * in that case ->scp_cpt is CFS_CPT_ANY */
2305 rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
2307 CWARN("%s: failed to bind %s on CPT %d\n",
2308 svc->srv_name, thread->t_name, svcpt->scp_cpt);
2311 #ifdef WITH_GROUP_INFO
2312 ginfo = groups_alloc(0);
2318 set_current_groups(ginfo);
2319 put_group_info(ginfo);
2322 if (svc->srv_ops.so_thr_init != NULL) {
2323 rc = svc->srv_ops.so_thr_init(thread);
2334 rc = lu_context_init(&env->le_ctx,
2335 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2339 thread->t_env = env;
2340 env->le_ctx.lc_thread = thread;
2341 env->le_ctx.lc_cookie = 0x6;
2343 while (!list_empty(&svcpt->scp_rqbd_idle)) {
2344 rc = ptlrpc_server_post_idle_rqbds(svcpt);
2348 CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
2349 svc->srv_name, svcpt->scp_cpt, rc);
2353 /* Alloc reply state structure for this one */
2354 OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2360 spin_lock(&svcpt->scp_lock);
2362 LASSERT(thread_is_starting(thread));
2363 thread_clear_flags(thread, SVC_STARTING);
2365 LASSERT(svcpt->scp_nthrs_starting == 1);
2366 svcpt->scp_nthrs_starting--;
2368 /* SVC_STOPPING may already be set here if someone else is trying
2369 * to stop the service while this new thread has been dynamically
2370 * forked. We still set SVC_RUNNING to let our creator know that
2371 * we are now running, however we will exit as soon as possible */
2372 thread_add_flags(thread, SVC_RUNNING);
2373 svcpt->scp_nthrs_running++;
2374 spin_unlock(&svcpt->scp_lock);
2376 /* wake up our creator in case he's still waiting. */
2377 wake_up(&thread->t_ctl_waitq);
2380 thread->t_watchdog = lc_watchdog_add(ptlrpc_server_get_timeout(svcpt),
2384 spin_lock(&svcpt->scp_rep_lock);
2385 list_add(&rs->rs_list, &svcpt->scp_rep_idle);
2386 wake_up(&svcpt->scp_rep_waitq);
2387 spin_unlock(&svcpt->scp_rep_lock);
2389 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2390 svcpt->scp_nthrs_running);
2392 /* XXX maintain a list of all managed devices: insert here */
2393 while (!ptlrpc_thread_stopping(thread)) {
2394 if (ptlrpc_wait_event(svcpt, thread))
2397 ptlrpc_check_rqbd_pool(svcpt);
2399 if (ptlrpc_threads_need_create(svcpt)) {
2400 /* Ignore return code - we tried... */
2401 ptlrpc_start_thread(svcpt, 0);
2404 /* Process all incoming reqs before handling any */
2405 if (ptlrpc_server_request_incoming(svcpt)) {
2406 lu_context_enter(&env->le_ctx);
2408 ptlrpc_server_handle_req_in(svcpt, thread);
2409 lu_context_exit(&env->le_ctx);
2411 /* but limit ourselves in case of flood */
2412 if (counter++ < 100)
2417 if (ptlrpc_at_check(svcpt))
2418 ptlrpc_at_check_timed(svcpt);
2420 if (ptlrpc_server_request_pending(svcpt, false)) {
2421 lu_context_enter(&env->le_ctx);
2422 ptlrpc_server_handle_request(svcpt, thread);
2423 lu_context_exit(&env->le_ctx);
2426 if (ptlrpc_rqbd_pending(svcpt) &&
2427 ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
2428 /* I just failed to repost request buffers.
2429 * Wait for a timeout (unless something else
2430 * happens) before I try again */
2431 svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
2432 CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
2433 svcpt->scp_nrqbds_posted);
2438 lc_watchdog_delete(thread->t_watchdog);
2439 thread->t_watchdog = NULL;
2444 * deconstruct service specific state created by ptlrpc_start_thread()
2446 if (svc->srv_ops.so_thr_done != NULL)
2447 svc->srv_ops.so_thr_done(thread);
2450 lu_context_fini(&env->le_ctx);
2454 CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n",
2455 thread, thread->t_pid, thread->t_id, rc);
2457 spin_lock(&svcpt->scp_lock);
2458 if (thread_test_and_clear_flags(thread, SVC_STARTING))
2459 svcpt->scp_nthrs_starting--;
2461 if (thread_test_and_clear_flags(thread, SVC_RUNNING)) {
2462 /* must know immediately */
2463 svcpt->scp_nthrs_running--;
2467 thread_add_flags(thread, SVC_STOPPED);
2469 wake_up(&thread->t_ctl_waitq);
2470 spin_unlock(&svcpt->scp_lock);
2475 static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
2476 struct list_head *replies)
2480 spin_lock(&hrt->hrt_lock);
2482 list_splice_init(&hrt->hrt_queue, replies);
2483 result = ptlrpc_hr.hr_stopping || !list_empty(replies);
2485 spin_unlock(&hrt->hrt_lock);
2490 * Main body of "handle reply" function.
2491 * It processes acked reply states
2493 static int ptlrpc_hr_main(void *arg)
2495 struct ptlrpc_hr_thread *hrt = (struct ptlrpc_hr_thread *)arg;
2496 struct ptlrpc_hr_partition *hrp = hrt->hrt_partition;
2497 LIST_HEAD (replies);
2498 char threadname[20];
2501 snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
2502 hrp->hrp_cpt, hrt->hrt_id);
2503 unshare_fs_struct();
2505 rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
2507 CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
2508 threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
2511 atomic_inc(&hrp->hrp_nstarted);
2512 wake_up(&ptlrpc_hr.hr_waitq);
2514 while (!ptlrpc_hr.hr_stopping) {
2515 l_wait_condition(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
2517 while (!list_empty(&replies)) {
2518 struct ptlrpc_reply_state *rs;
2520 rs = list_entry(replies.prev,
2521 struct ptlrpc_reply_state,
2523 list_del_init(&rs->rs_list);
2524 ptlrpc_handle_rs(rs);
2528 atomic_inc(&hrp->hrp_nstopped);
2529 wake_up(&ptlrpc_hr.hr_waitq);
2534 static void ptlrpc_stop_hr_threads(void)
2536 struct ptlrpc_hr_partition *hrp;
2540 ptlrpc_hr.hr_stopping = 1;
2542 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2543 if (hrp->hrp_thrs == NULL)
2544 continue; /* uninitialized */
2545 for (j = 0; j < hrp->hrp_nthrs; j++)
2546 wake_up_all(&hrp->hrp_thrs[j].hrt_waitq);
2549 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2550 if (hrp->hrp_thrs == NULL)
2551 continue; /* uninitialized */
2552 wait_event(ptlrpc_hr.hr_waitq,
2553 atomic_read(&hrp->hrp_nstopped) ==
2554 atomic_read(&hrp->hrp_nstarted));
2558 static int ptlrpc_start_hr_threads(void)
2560 struct ptlrpc_hr_partition *hrp;
2565 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2568 for (j = 0; j < hrp->hrp_nthrs; j++) {
2569 struct ptlrpc_hr_thread *hrt = &hrp->hrp_thrs[j];
2570 rc = PTR_ERR(kthread_run(ptlrpc_hr_main,
2572 "ptlrpc_hr%02d_%03d",
2575 if (IS_ERR_VALUE(rc))
2578 wait_event(ptlrpc_hr.hr_waitq,
2579 atomic_read(&hrp->hrp_nstarted) == j);
2580 if (!IS_ERR_VALUE(rc))
2583 CERROR("Reply handling thread %d:%d Failed on starting: "
2584 "rc = %d\n", i, j, rc);
2585 ptlrpc_stop_hr_threads();
2591 static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
2593 struct l_wait_info lwi = { 0 };
2594 struct ptlrpc_thread *thread;
2599 CDEBUG(D_INFO, "Stopping threads for service %s\n",
2600 svcpt->scp_service->srv_name);
2602 spin_lock(&svcpt->scp_lock);
2603 /* let the thread know that we would like it to stop asap */
2604 list_for_each_entry(thread, &svcpt->scp_threads, t_link) {
2605 CDEBUG(D_INFO, "Stopping thread %s #%u\n",
2606 svcpt->scp_service->srv_thread_name, thread->t_id);
2607 thread_add_flags(thread, SVC_STOPPING);
2610 wake_up_all(&svcpt->scp_waitq);
2612 while (!list_empty(&svcpt->scp_threads)) {
2613 thread = list_entry(svcpt->scp_threads.next,
2614 struct ptlrpc_thread, t_link);
2615 if (thread_is_stopped(thread)) {
2616 list_del(&thread->t_link);
2617 list_add(&thread->t_link, &zombie);
2620 spin_unlock(&svcpt->scp_lock);
2622 CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
2623 svcpt->scp_service->srv_thread_name, thread->t_id);
2624 l_wait_event(thread->t_ctl_waitq,
2625 thread_is_stopped(thread), &lwi);
2627 spin_lock(&svcpt->scp_lock);
2630 spin_unlock(&svcpt->scp_lock);
2632 while (!list_empty(&zombie)) {
2633 thread = list_entry(zombie.next,
2634 struct ptlrpc_thread, t_link);
2635 list_del(&thread->t_link);
2636 OBD_FREE_PTR(thread);
2642 * Stops all threads of a particular service \a svc
2644 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2646 struct ptlrpc_service_part *svcpt;
2650 ptlrpc_service_for_each_part(svcpt, i, svc) {
2651 if (svcpt->scp_service != NULL)
2652 ptlrpc_svcpt_stop_threads(svcpt);
2657 EXPORT_SYMBOL(ptlrpc_stop_all_threads);
2659 int ptlrpc_start_threads(struct ptlrpc_service *svc)
2666 /* We require 2 threads min, see note in ptlrpc_server_handle_request */
2667 LASSERT(svc->srv_nthrs_cpt_init >= PTLRPC_NTHRS_INIT);
2669 for (i = 0; i < svc->srv_ncpts; i++) {
2670 for (j = 0; j < svc->srv_nthrs_cpt_init; j++) {
2671 rc = ptlrpc_start_thread(svc->srv_parts[i], 1);
2677 /* We have enough threads, don't start more. b=15759 */
2684 CERROR("cannot start %s thread #%d_%d: rc %d\n",
2685 svc->srv_thread_name, i, j, rc);
2686 ptlrpc_stop_all_threads(svc);
2689 EXPORT_SYMBOL(ptlrpc_start_threads);
2691 int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
2693 struct l_wait_info lwi = { 0 };
2694 struct ptlrpc_thread *thread;
2695 struct ptlrpc_service *svc;
2699 LASSERT(svcpt != NULL);
2701 svc = svcpt->scp_service;
2703 CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n",
2704 svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running,
2705 svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit);
2708 if (unlikely(svc->srv_is_stopping))
2711 if (!ptlrpc_threads_increasable(svcpt) ||
2712 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
2713 svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1))
2716 OBD_CPT_ALLOC_PTR(thread, svc->srv_cptable, svcpt->scp_cpt);
2719 init_waitqueue_head(&thread->t_ctl_waitq);
2721 spin_lock(&svcpt->scp_lock);
2722 if (!ptlrpc_threads_increasable(svcpt)) {
2723 spin_unlock(&svcpt->scp_lock);
2724 OBD_FREE_PTR(thread);
2728 if (svcpt->scp_nthrs_starting != 0) {
2729 /* serialize starting because some modules (obdfilter)
2730 * might require unique and contiguous t_id */
2731 LASSERT(svcpt->scp_nthrs_starting == 1);
2732 spin_unlock(&svcpt->scp_lock);
2733 OBD_FREE_PTR(thread);
2735 CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n",
2736 svc->srv_thread_name, svcpt->scp_thr_nextid);
2741 CDEBUG(D_INFO, "Creating thread %s #%d race, retry later\n",
2742 svc->srv_thread_name, svcpt->scp_thr_nextid);
2746 svcpt->scp_nthrs_starting++;
2747 thread->t_id = svcpt->scp_thr_nextid++;
2748 thread_add_flags(thread, SVC_STARTING);
2749 thread->t_svcpt = svcpt;
2751 list_add(&thread->t_link, &svcpt->scp_threads);
2752 spin_unlock(&svcpt->scp_lock);
2754 if (svcpt->scp_cpt >= 0) {
2755 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s%02d_%03d",
2756 svc->srv_thread_name, svcpt->scp_cpt, thread->t_id);
2758 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s_%04d",
2759 svc->srv_thread_name, thread->t_id);
2762 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name);
2763 rc = PTR_ERR(kthread_run(ptlrpc_main, thread, thread->t_name));
2764 if (IS_ERR_VALUE(rc)) {
2765 CERROR("cannot start thread '%s': rc %d\n",
2766 thread->t_name, rc);
2767 spin_lock(&svcpt->scp_lock);
2768 list_del(&thread->t_link);
2769 --svcpt->scp_nthrs_starting;
2770 spin_unlock(&svcpt->scp_lock);
2772 OBD_FREE(thread, sizeof(*thread));
2779 l_wait_event(thread->t_ctl_waitq,
2780 thread_is_running(thread) || thread_is_stopped(thread),
2783 rc = thread_is_stopped(thread) ? thread->t_id : 0;
2787 int ptlrpc_hr_init(void)
2790 struct ptlrpc_hr_partition *hrp;
2791 struct ptlrpc_hr_thread *hrt;
2798 memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
2799 ptlrpc_hr.hr_cpt_table = cfs_cpt_table;
2801 ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
2803 if (ptlrpc_hr.hr_partitions == NULL)
2806 init_waitqueue_head(&ptlrpc_hr.hr_waitq);
2808 cpumask_copy(&mask, topology_thread_cpumask(0));
2809 weight = cpus_weight(mask);
2811 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2814 atomic_set(&hrp->hrp_nstarted, 0);
2815 atomic_set(&hrp->hrp_nstopped, 0);
2817 hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, i);
2818 hrp->hrp_nthrs /= weight;
2820 LASSERT(hrp->hrp_nthrs > 0);
2821 OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, i,
2822 hrp->hrp_nthrs * sizeof(*hrt));
2823 if (hrp->hrp_thrs == NULL)
2824 GOTO(out, rc = -ENOMEM);
2826 for (j = 0; j < hrp->hrp_nthrs; j++) {
2827 hrt = &hrp->hrp_thrs[j];
2830 hrt->hrt_partition = hrp;
2831 init_waitqueue_head(&hrt->hrt_waitq);
2832 spin_lock_init(&hrt->hrt_lock);
2833 INIT_LIST_HEAD(&hrt->hrt_queue);
2837 rc = ptlrpc_start_hr_threads();
2844 void ptlrpc_hr_fini(void)
2846 struct ptlrpc_hr_partition *hrp;
2849 if (ptlrpc_hr.hr_partitions == NULL)
2852 ptlrpc_stop_hr_threads();
2854 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2855 if (hrp->hrp_thrs != NULL) {
2856 OBD_FREE(hrp->hrp_thrs,
2857 hrp->hrp_nthrs * sizeof(hrp->hrp_thrs[0]));
2861 cfs_percpt_free(ptlrpc_hr.hr_partitions);
2862 ptlrpc_hr.hr_partitions = NULL;
2867 * Wait until all already scheduled replies are processed.
2869 static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
2873 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
2876 rc = l_wait_event(svcpt->scp_waitq,
2877 atomic_read(&svcpt->scp_nreps_difficult) == 0, &lwi);
2880 CWARN("Unexpectedly long timeout %s %p\n",
2881 svcpt->scp_service->srv_name, svcpt->scp_service);
2886 ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
2888 struct ptlrpc_service_part *svcpt;
2891 /* early disarm AT timer... */
2892 ptlrpc_service_for_each_part(svcpt, i, svc) {
2893 if (svcpt->scp_service != NULL)
2894 cfs_timer_disarm(&svcpt->scp_at_timer);
2899 ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
2901 struct ptlrpc_service_part *svcpt;
2902 struct ptlrpc_request_buffer_desc *rqbd;
2903 struct l_wait_info lwi;
2907 /* All history will be culled when the next request buffer is
2908 * freed in ptlrpc_service_purge_all() */
2909 svc->srv_hist_nrqbds_cpt_max = 0;
2911 rc = LNetClearLazyPortal(svc->srv_req_portal);
2914 ptlrpc_service_for_each_part(svcpt, i, svc) {
2915 if (svcpt->scp_service == NULL)
2918 /* Unlink all the request buffers. This forces a 'final'
2919 * event with its 'unlink' flag set for each posted rqbd */
2920 list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
2922 rc = LNetMDUnlink(rqbd->rqbd_md_h);
2923 LASSERT(rc == 0 || rc == -ENOENT);
2927 ptlrpc_service_for_each_part(svcpt, i, svc) {
2928 if (svcpt->scp_service == NULL)
2931 /* Wait for the network to release any buffers
2932 * it's currently filling */
2933 spin_lock(&svcpt->scp_lock);
2934 while (svcpt->scp_nrqbds_posted != 0) {
2935 spin_unlock(&svcpt->scp_lock);
2936 /* Network access will complete in finite time but
2937 * the HUGE timeout lets us CWARN for visibility
2938 * of sluggish NALs */
2939 lwi = LWI_TIMEOUT_INTERVAL(
2940 cfs_time_seconds(LONG_UNLINK),
2941 cfs_time_seconds(1), NULL, NULL);
2942 rc = l_wait_event(svcpt->scp_waitq,
2943 svcpt->scp_nrqbds_posted == 0, &lwi);
2944 if (rc == -ETIMEDOUT) {
2945 CWARN("Service %s waiting for "
2946 "request buffers\n",
2947 svcpt->scp_service->srv_name);
2949 spin_lock(&svcpt->scp_lock);
2951 spin_unlock(&svcpt->scp_lock);
2956 ptlrpc_service_purge_all(struct ptlrpc_service *svc)
2958 struct ptlrpc_service_part *svcpt;
2959 struct ptlrpc_request_buffer_desc *rqbd;
2960 struct ptlrpc_request *req;
2961 struct ptlrpc_reply_state *rs;
2964 ptlrpc_service_for_each_part(svcpt, i, svc) {
2965 if (svcpt->scp_service == NULL)
2968 spin_lock(&svcpt->scp_rep_lock);
2969 while (!list_empty(&svcpt->scp_rep_active)) {
2970 rs = list_entry(svcpt->scp_rep_active.next,
2971 struct ptlrpc_reply_state, rs_list);
2972 spin_lock(&rs->rs_lock);
2973 ptlrpc_schedule_difficult_reply(rs);
2974 spin_unlock(&rs->rs_lock);
2976 spin_unlock(&svcpt->scp_rep_lock);
2978 /* purge the request queue. NB No new replies (rqbds
2979 * all unlinked) and no service threads, so I'm the only
2980 * thread noodling the request queue now */
2981 while (!list_empty(&svcpt->scp_req_incoming)) {
2982 req = list_entry(svcpt->scp_req_incoming.next,
2983 struct ptlrpc_request, rq_list);
2985 list_del(&req->rq_list);
2986 svcpt->scp_nreqs_incoming--;
2987 ptlrpc_server_finish_request(svcpt, req);
2990 while (ptlrpc_server_request_pending(svcpt, true)) {
2991 req = ptlrpc_server_request_get(svcpt, true);
2992 ptlrpc_server_finish_active_request(svcpt, req);
2995 LASSERT(list_empty(&svcpt->scp_rqbd_posted));
2996 LASSERT(svcpt->scp_nreqs_incoming == 0);
2997 LASSERT(svcpt->scp_nreqs_active == 0);
2998 /* history should have been culled by
2999 * ptlrpc_server_finish_request */
3000 LASSERT(svcpt->scp_hist_nrqbds == 0);
3002 /* Now free all the request buffers since nothing
3003 * references them any more... */
3005 while (!list_empty(&svcpt->scp_rqbd_idle)) {
3006 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
3007 struct ptlrpc_request_buffer_desc,
3009 ptlrpc_free_rqbd(rqbd);
3011 ptlrpc_wait_replies(svcpt);
3013 while (!list_empty(&svcpt->scp_rep_idle)) {
3014 rs = list_entry(svcpt->scp_rep_idle.next,
3015 struct ptlrpc_reply_state,
3017 list_del(&rs->rs_list);
3018 OBD_FREE_LARGE(rs, svc->srv_max_reply_size);
3024 ptlrpc_service_free(struct ptlrpc_service *svc)
3026 struct ptlrpc_service_part *svcpt;
3027 struct ptlrpc_at_array *array;
3030 ptlrpc_service_for_each_part(svcpt, i, svc) {
3031 if (svcpt->scp_service == NULL)
3034 /* In case somebody rearmed this in the meantime */
3035 cfs_timer_disarm(&svcpt->scp_at_timer);
3036 array = &svcpt->scp_at_array;
3038 if (array->paa_reqs_array != NULL) {
3039 OBD_FREE(array->paa_reqs_array,
3040 sizeof(struct list_head) * array->paa_size);
3041 array->paa_reqs_array = NULL;
3044 if (array->paa_reqs_count != NULL) {
3045 OBD_FREE(array->paa_reqs_count,
3046 sizeof(__u32) * array->paa_size);
3047 array->paa_reqs_count = NULL;
3051 ptlrpc_service_for_each_part(svcpt, i, svc)
3052 OBD_FREE_PTR(svcpt);
3054 if (svc->srv_cpts != NULL)
3055 cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts);
3057 OBD_FREE(svc, offsetof(struct ptlrpc_service,
3058 srv_parts[svc->srv_ncpts]));
3061 int ptlrpc_unregister_service(struct ptlrpc_service *service)
3065 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
3067 service->srv_is_stopping = 1;
3069 mutex_lock(&ptlrpc_all_services_mutex);
3070 list_del_init(&service->srv_list);
3071 mutex_unlock(&ptlrpc_all_services_mutex);
3073 ptlrpc_service_del_atimer(service);
3074 ptlrpc_stop_all_threads(service);
3076 ptlrpc_service_unlink_rqbd(service);
3077 ptlrpc_service_purge_all(service);
3078 ptlrpc_service_nrs_cleanup(service);
3080 ptlrpc_lprocfs_unregister_service(service);
3082 ptlrpc_service_free(service);
3086 EXPORT_SYMBOL(ptlrpc_unregister_service);
3089 * Returns 0 if the service is healthy.
3091 * Right now, it just checks to make sure that requests aren't languishing
3092 * in the queue. We'll use this health check to govern whether a node needs
3093 * to be shot, so it's intentionally non-aggressive. */
3094 int ptlrpc_svcpt_health_check(struct ptlrpc_service_part *svcpt)
3096 struct ptlrpc_request *request = NULL;
3097 struct timeval right_now;
3100 do_gettimeofday(&right_now);
3102 spin_lock(&svcpt->scp_req_lock);
3103 /* How long has the next entry been waiting? */
3104 if (ptlrpc_server_high_pending(svcpt, true))
3105 request = ptlrpc_nrs_req_peek_nolock(svcpt, true);
3106 else if (ptlrpc_server_normal_pending(svcpt, true))
3107 request = ptlrpc_nrs_req_peek_nolock(svcpt, false);
3109 if (request == NULL) {
3110 spin_unlock(&svcpt->scp_req_lock);
3114 timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
3115 spin_unlock(&svcpt->scp_req_lock);
3117 if ((timediff / ONE_MILLION) >
3118 (AT_OFF ? obd_timeout * 3 / 2 : at_max)) {
3119 CERROR("%s: unhealthy - request has been waiting %lds\n",
3120 svcpt->scp_service->srv_name, timediff / ONE_MILLION);
3128 ptlrpc_service_health_check(struct ptlrpc_service *svc)
3130 struct ptlrpc_service_part *svcpt;
3136 ptlrpc_service_for_each_part(svcpt, i, svc) {
3137 int rc = ptlrpc_svcpt_health_check(svcpt);
3144 EXPORT_SYMBOL(ptlrpc_service_health_check);