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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
33 #define DEBUG_SUBSYSTEM S_RPC
34 #include <linux/kthread.h>
35 #include <obd_support.h>
36 #include <obd_class.h>
37 #include <lustre_net.h>
38 #include <lu_object.h>
39 #include <lnet/types.h>
40 #include "ptlrpc_internal.h"
42 /* The following are visible and mutable through /sys/module/ptlrpc */
43 int test_req_buffer_pressure = 0;
44 module_param(test_req_buffer_pressure, int, 0444);
45 MODULE_PARM_DESC(test_req_buffer_pressure, "set non-zero to put pressure on request buffer pools");
46 module_param(at_min, int, 0644);
47 MODULE_PARM_DESC(at_min, "Adaptive timeout minimum (sec)");
48 module_param(at_max, int, 0644);
49 MODULE_PARM_DESC(at_max, "Adaptive timeout maximum (sec)");
50 module_param(at_history, int, 0644);
51 MODULE_PARM_DESC(at_history,
52 "Adaptive timeouts remember the slowest event that took place within this period (sec)");
53 module_param(at_early_margin, int, 0644);
54 MODULE_PARM_DESC(at_early_margin, "How soon before an RPC deadline to send an early reply");
55 module_param(at_extra, int, 0644);
56 MODULE_PARM_DESC(at_extra, "How much extra time to give with each early reply");
59 static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt);
60 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req);
61 static void ptlrpc_at_remove_timed(struct ptlrpc_request *req);
63 /** Holds a list of all PTLRPC services */
64 struct list_head ptlrpc_all_services;
65 /** Used to protect the \e ptlrpc_all_services list */
66 struct mutex ptlrpc_all_services_mutex;
68 static struct ptlrpc_request_buffer_desc *
69 ptlrpc_alloc_rqbd(struct ptlrpc_service_part *svcpt)
71 struct ptlrpc_service *svc = svcpt->scp_service;
72 struct ptlrpc_request_buffer_desc *rqbd;
74 OBD_CPT_ALLOC_PTR(rqbd, svc->srv_cptable, svcpt->scp_cpt);
78 rqbd->rqbd_svcpt = svcpt;
79 rqbd->rqbd_refcount = 0;
80 rqbd->rqbd_cbid.cbid_fn = request_in_callback;
81 rqbd->rqbd_cbid.cbid_arg = rqbd;
82 INIT_LIST_HEAD(&rqbd->rqbd_reqs);
83 OBD_CPT_ALLOC_LARGE(rqbd->rqbd_buffer, svc->srv_cptable,
84 svcpt->scp_cpt, svc->srv_buf_size);
85 if (rqbd->rqbd_buffer == NULL) {
90 spin_lock(&svcpt->scp_lock);
91 list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
92 svcpt->scp_nrqbds_total++;
93 spin_unlock(&svcpt->scp_lock);
99 ptlrpc_free_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
101 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
103 LASSERT(rqbd->rqbd_refcount == 0);
104 LASSERT(list_empty(&rqbd->rqbd_reqs));
106 spin_lock(&svcpt->scp_lock);
107 list_del(&rqbd->rqbd_list);
108 svcpt->scp_nrqbds_total--;
109 spin_unlock(&svcpt->scp_lock);
111 OBD_FREE_LARGE(rqbd->rqbd_buffer, svcpt->scp_service->srv_buf_size);
116 ptlrpc_grow_req_bufs(struct ptlrpc_service_part *svcpt, int post)
118 struct ptlrpc_service *svc = svcpt->scp_service;
119 struct ptlrpc_request_buffer_desc *rqbd;
123 if (svcpt->scp_rqbd_allocating)
126 spin_lock(&svcpt->scp_lock);
127 /* check again with lock */
128 if (svcpt->scp_rqbd_allocating) {
129 /* NB: we might allow more than one thread in the future */
130 LASSERT(svcpt->scp_rqbd_allocating == 1);
131 spin_unlock(&svcpt->scp_lock);
135 svcpt->scp_rqbd_allocating++;
136 spin_unlock(&svcpt->scp_lock);
139 for (i = 0; i < svc->srv_nbuf_per_group; i++) {
140 /* NB: another thread might have recycled enough rqbds, we
141 * need to make sure it wouldn't over-allocate, see LU-1212. */
142 if (svcpt->scp_nrqbds_posted >= svc->srv_nbuf_per_group)
145 rqbd = ptlrpc_alloc_rqbd(svcpt);
148 CERROR("%s: Can't allocate request buffer\n",
155 spin_lock(&svcpt->scp_lock);
157 LASSERT(svcpt->scp_rqbd_allocating == 1);
158 svcpt->scp_rqbd_allocating--;
160 spin_unlock(&svcpt->scp_lock);
163 "%s: allocate %d new %d-byte reqbufs (%d/%d left), rc = %d\n",
164 svc->srv_name, i, svc->srv_buf_size, svcpt->scp_nrqbds_posted,
165 svcpt->scp_nrqbds_total, rc);
169 rc = ptlrpc_server_post_idle_rqbds(svcpt);
175 * Part of Rep-Ack logic.
176 * Puts a lock and its mode into reply state assotiated to request reply.
179 ptlrpc_save_lock(struct ptlrpc_request *req, struct lustre_handle *lock,
180 int mode, bool no_ack, bool convert_lock)
182 struct ptlrpc_reply_state *rs = req->rq_reply_state;
186 LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
188 if (req->rq_export->exp_disconnected) {
189 ldlm_lock_decref(lock, mode);
191 idx = rs->rs_nlocks++;
192 rs->rs_locks[idx] = *lock;
193 rs->rs_modes[idx] = mode;
194 rs->rs_difficult = 1;
195 rs->rs_no_ack = no_ack;
196 rs->rs_convert_lock = convert_lock;
199 EXPORT_SYMBOL(ptlrpc_save_lock);
202 struct ptlrpc_hr_partition;
204 struct ptlrpc_hr_thread {
205 int hrt_id; /* thread ID */
207 wait_queue_head_t hrt_waitq;
208 struct list_head hrt_queue;
209 struct ptlrpc_hr_partition *hrt_partition;
212 struct ptlrpc_hr_partition {
213 /* # of started threads */
214 atomic_t hrp_nstarted;
215 /* # of stopped threads */
216 atomic_t hrp_nstopped;
217 /* cpu partition id */
219 /* round-robin rotor for choosing thread */
221 /* total number of threads on this partition */
224 struct ptlrpc_hr_thread *hrp_thrs;
227 #define HRT_RUNNING 0
228 #define HRT_STOPPING 1
230 struct ptlrpc_hr_service {
231 /* CPU partition table, it's just cfs_cpt_table for now */
232 struct cfs_cpt_table *hr_cpt_table;
233 /** controller sleep waitq */
234 wait_queue_head_t hr_waitq;
235 unsigned int hr_stopping;
236 /** roundrobin rotor for non-affinity service */
237 unsigned int hr_rotor;
239 struct ptlrpc_hr_partition **hr_partitions;
243 struct list_head rsb_replies;
244 unsigned int rsb_n_replies;
245 struct ptlrpc_service_part *rsb_svcpt;
248 /** reply handling service. */
249 static struct ptlrpc_hr_service ptlrpc_hr;
252 * maximum mumber of replies scheduled in one batch
254 #define MAX_SCHEDULED 256
257 * Initialize a reply batch.
261 static void rs_batch_init(struct rs_batch *b)
263 memset(b, 0, sizeof *b);
264 INIT_LIST_HEAD(&b->rsb_replies);
268 * Choose an hr thread to dispatch requests to.
270 static struct ptlrpc_hr_thread *
271 ptlrpc_hr_select(struct ptlrpc_service_part *svcpt)
273 struct ptlrpc_hr_partition *hrp;
276 if (svcpt->scp_cpt >= 0 &&
277 svcpt->scp_service->srv_cptable == ptlrpc_hr.hr_cpt_table) {
278 /* directly match partition */
279 hrp = ptlrpc_hr.hr_partitions[svcpt->scp_cpt];
282 rotor = ptlrpc_hr.hr_rotor++;
283 rotor %= cfs_cpt_number(ptlrpc_hr.hr_cpt_table);
285 hrp = ptlrpc_hr.hr_partitions[rotor];
288 rotor = hrp->hrp_rotor++;
289 return &hrp->hrp_thrs[rotor % hrp->hrp_nthrs];
293 * Dispatch all replies accumulated in the batch to one from
294 * dedicated reply handling threads.
298 static void rs_batch_dispatch(struct rs_batch *b)
300 if (b->rsb_n_replies != 0) {
301 struct ptlrpc_hr_thread *hrt;
303 hrt = ptlrpc_hr_select(b->rsb_svcpt);
305 spin_lock(&hrt->hrt_lock);
306 list_splice_init(&b->rsb_replies, &hrt->hrt_queue);
307 spin_unlock(&hrt->hrt_lock);
309 wake_up(&hrt->hrt_waitq);
310 b->rsb_n_replies = 0;
315 * Add a reply to a batch.
316 * Add one reply object to a batch, schedule batched replies if overload.
321 static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs)
323 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
325 if (svcpt != b->rsb_svcpt || b->rsb_n_replies >= MAX_SCHEDULED) {
326 if (b->rsb_svcpt != NULL) {
327 rs_batch_dispatch(b);
328 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
330 spin_lock(&svcpt->scp_rep_lock);
331 b->rsb_svcpt = svcpt;
333 spin_lock(&rs->rs_lock);
334 rs->rs_scheduled_ever = 1;
335 if (rs->rs_scheduled == 0) {
336 list_move(&rs->rs_list, &b->rsb_replies);
337 rs->rs_scheduled = 1;
340 rs->rs_committed = 1;
341 spin_unlock(&rs->rs_lock);
345 * Reply batch finalization.
346 * Dispatch remaining replies from the batch
347 * and release remaining spinlock.
351 static void rs_batch_fini(struct rs_batch *b)
353 if (b->rsb_svcpt != NULL) {
354 rs_batch_dispatch(b);
355 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
359 #define DECLARE_RS_BATCH(b) struct rs_batch b
363 * Put reply state into a queue for processing because we received
364 * ACK from the client
366 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
368 struct ptlrpc_hr_thread *hrt;
371 LASSERT(list_empty(&rs->rs_list));
373 hrt = ptlrpc_hr_select(rs->rs_svcpt);
375 spin_lock(&hrt->hrt_lock);
376 list_add_tail(&rs->rs_list, &hrt->hrt_queue);
377 spin_unlock(&hrt->hrt_lock);
379 wake_up(&hrt->hrt_waitq);
384 ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs)
388 assert_spin_locked(&rs->rs_svcpt->scp_rep_lock);
389 assert_spin_locked(&rs->rs_lock);
390 LASSERT (rs->rs_difficult);
391 rs->rs_scheduled_ever = 1; /* flag any notification attempt */
393 if (rs->rs_scheduled) { /* being set up or already notified */
398 rs->rs_scheduled = 1;
399 list_del_init(&rs->rs_list);
400 ptlrpc_dispatch_difficult_reply(rs);
403 EXPORT_SYMBOL(ptlrpc_schedule_difficult_reply);
405 void ptlrpc_commit_replies(struct obd_export *exp)
407 struct ptlrpc_reply_state *rs, *nxt;
408 DECLARE_RS_BATCH(batch);
411 rs_batch_init(&batch);
412 /* Find any replies that have been committed and get their service
413 * to attend to complete them. */
415 /* CAVEAT EMPTOR: spinlock ordering!!! */
416 spin_lock(&exp->exp_uncommitted_replies_lock);
417 list_for_each_entry_safe(rs, nxt, &exp->exp_uncommitted_replies,
419 LASSERT (rs->rs_difficult);
420 /* VBR: per-export last_committed */
421 LASSERT(rs->rs_export);
422 if (rs->rs_transno <= exp->exp_last_committed) {
423 list_del_init(&rs->rs_obd_list);
424 rs_batch_add(&batch, rs);
427 spin_unlock(&exp->exp_uncommitted_replies_lock);
428 rs_batch_fini(&batch);
433 ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt)
435 struct ptlrpc_request_buffer_desc *rqbd;
440 spin_lock(&svcpt->scp_lock);
442 if (list_empty(&svcpt->scp_rqbd_idle)) {
443 spin_unlock(&svcpt->scp_lock);
447 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
448 struct ptlrpc_request_buffer_desc,
450 list_del(&rqbd->rqbd_list);
452 /* assume we will post successfully */
453 svcpt->scp_nrqbds_posted++;
454 list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_posted);
456 spin_unlock(&svcpt->scp_lock);
458 rc = ptlrpc_register_rqbd(rqbd);
465 spin_lock(&svcpt->scp_lock);
467 svcpt->scp_nrqbds_posted--;
468 list_del(&rqbd->rqbd_list);
469 list_add_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
471 /* Don't complain if no request buffers are posted right now; LNET
472 * won't drop requests because we set the portal lazy! */
474 spin_unlock(&svcpt->scp_lock);
479 static void ptlrpc_at_timer(unsigned long castmeharder)
481 struct ptlrpc_service_part *svcpt;
483 svcpt = (struct ptlrpc_service_part *)castmeharder;
485 svcpt->scp_at_check = 1;
486 svcpt->scp_at_checktime = cfs_time_current();
487 wake_up(&svcpt->scp_waitq);
491 ptlrpc_server_nthreads_check(struct ptlrpc_service *svc,
492 struct ptlrpc_service_conf *conf)
494 struct ptlrpc_service_thr_conf *tc = &conf->psc_thr;
501 * Common code for estimating & validating threads number.
502 * CPT affinity service could have percpt thread-pool instead
503 * of a global thread-pool, which means user might not always
504 * get the threads number they give it in conf::tc_nthrs_user
505 * even they did set. It's because we need to validate threads
506 * number for each CPT to guarantee each pool will have enough
507 * threads to keep the service healthy.
509 init = PTLRPC_NTHRS_INIT + (svc->srv_ops.so_hpreq_handler != NULL);
510 init = max_t(int, init, tc->tc_nthrs_init);
512 /* NB: please see comments in lustre_lnet.h for definition
513 * details of these members */
514 LASSERT(tc->tc_nthrs_max != 0);
516 if (tc->tc_nthrs_user != 0) {
517 /* In case there is a reason to test a service with many
518 * threads, we give a less strict check here, it can
519 * be up to 8 * nthrs_max */
520 total = min(tc->tc_nthrs_max * 8, tc->tc_nthrs_user);
521 nthrs = total / svc->srv_ncpts;
522 init = max(init, nthrs);
526 total = tc->tc_nthrs_max;
527 if (tc->tc_nthrs_base == 0) {
528 /* don't care about base threads number per partition,
529 * this is most for non-affinity service */
530 nthrs = total / svc->srv_ncpts;
534 nthrs = tc->tc_nthrs_base;
535 if (svc->srv_ncpts == 1) {
538 /* NB: Increase the base number if it's single partition
539 * and total number of cores/HTs is larger or equal to 4.
540 * result will always < 2 * nthrs_base */
541 weight = cfs_cpt_weight(svc->srv_cptable, CFS_CPT_ANY);
542 for (i = 1; (weight >> (i + 1)) != 0 && /* >= 4 cores/HTs */
543 (tc->tc_nthrs_base >> i) != 0; i++)
544 nthrs += tc->tc_nthrs_base >> i;
547 if (tc->tc_thr_factor != 0) {
548 int factor = tc->tc_thr_factor;
552 * User wants to increase number of threads with for
553 * each CPU core/HT, most likely the factor is larger than
554 * one thread/core because service threads are supposed to
555 * be blocked by lock or wait for IO.
558 * Amdahl's law says that adding processors wouldn't give
559 * a linear increasing of parallelism, so it's nonsense to
560 * have too many threads no matter how many cores/HTs
563 if (cpumask_weight(topology_sibling_cpumask(smp_processor_id())) > 1) {
564 /* weight is # of HTs */
565 /* depress thread factor for hyper-thread */
566 factor = factor - (factor >> 1) + (factor >> 3);
569 weight = cfs_cpt_weight(svc->srv_cptable, 0);
572 for (; factor > 0 && weight > 0; factor--, weight -= fade)
573 nthrs += min(weight, fade) * factor;
576 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
577 nthrs = max(tc->tc_nthrs_base,
578 tc->tc_nthrs_max / svc->srv_ncpts);
581 nthrs = max(nthrs, tc->tc_nthrs_init);
582 svc->srv_nthrs_cpt_limit = nthrs;
583 svc->srv_nthrs_cpt_init = init;
585 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
586 CDEBUG(D_OTHER, "%s: This service may have more threads (%d) "
587 "than the given soft limit (%d)\n",
588 svc->srv_name, nthrs * svc->srv_ncpts,
594 * Initialize percpt data for a service
597 ptlrpc_service_part_init(struct ptlrpc_service *svc,
598 struct ptlrpc_service_part *svcpt, int cpt)
600 struct ptlrpc_at_array *array;
605 svcpt->scp_cpt = cpt;
606 INIT_LIST_HEAD(&svcpt->scp_threads);
608 /* rqbd and incoming request queue */
609 spin_lock_init(&svcpt->scp_lock);
610 INIT_LIST_HEAD(&svcpt->scp_rqbd_idle);
611 INIT_LIST_HEAD(&svcpt->scp_rqbd_posted);
612 INIT_LIST_HEAD(&svcpt->scp_req_incoming);
613 init_waitqueue_head(&svcpt->scp_waitq);
614 /* history request & rqbd list */
615 INIT_LIST_HEAD(&svcpt->scp_hist_reqs);
616 INIT_LIST_HEAD(&svcpt->scp_hist_rqbds);
618 /* acitve requests and hp requests */
619 spin_lock_init(&svcpt->scp_req_lock);
622 spin_lock_init(&svcpt->scp_rep_lock);
623 INIT_LIST_HEAD(&svcpt->scp_rep_active);
624 INIT_LIST_HEAD(&svcpt->scp_rep_idle);
625 init_waitqueue_head(&svcpt->scp_rep_waitq);
626 atomic_set(&svcpt->scp_nreps_difficult, 0);
628 /* adaptive timeout */
629 spin_lock_init(&svcpt->scp_at_lock);
630 array = &svcpt->scp_at_array;
632 size = at_est2timeout(at_max);
633 array->paa_size = size;
634 array->paa_count = 0;
635 array->paa_deadline = -1;
637 /* allocate memory for scp_at_array (ptlrpc_at_array) */
638 OBD_CPT_ALLOC(array->paa_reqs_array,
639 svc->srv_cptable, cpt, sizeof(struct list_head) * size);
640 if (array->paa_reqs_array == NULL)
643 for (index = 0; index < size; index++)
644 INIT_LIST_HEAD(&array->paa_reqs_array[index]);
646 OBD_CPT_ALLOC(array->paa_reqs_count,
647 svc->srv_cptable, cpt, sizeof(__u32) * size);
648 if (array->paa_reqs_count == NULL)
651 setup_timer(&svcpt->scp_at_timer, ptlrpc_at_timer,
652 (unsigned long)svcpt);
654 /* At SOW, service time should be quick; 10s seems generous. If client
655 * timeout is less than this, we'll be sending an early reply. */
656 at_init(&svcpt->scp_at_estimate, 10, 0);
658 /* assign this before call ptlrpc_grow_req_bufs */
659 svcpt->scp_service = svc;
660 /* Now allocate the request buffers, but don't post them now */
661 rc = ptlrpc_grow_req_bufs(svcpt, 0);
662 /* We shouldn't be under memory pressure at startup, so
663 * fail if we can't allocate all our buffers at this time. */
670 if (array->paa_reqs_count != NULL) {
671 OBD_FREE(array->paa_reqs_count, sizeof(__u32) * size);
672 array->paa_reqs_count = NULL;
675 if (array->paa_reqs_array != NULL) {
676 OBD_FREE(array->paa_reqs_array,
677 sizeof(struct list_head) * array->paa_size);
678 array->paa_reqs_array = NULL;
685 * Initialize service on a given portal.
686 * This includes starting serving threads , allocating and posting rqbds and
689 struct ptlrpc_service *
690 ptlrpc_register_service(struct ptlrpc_service_conf *conf,
691 struct proc_dir_entry *proc_entry)
693 struct ptlrpc_service_cpt_conf *cconf = &conf->psc_cpt;
694 struct ptlrpc_service *service;
695 struct ptlrpc_service_part *svcpt;
696 struct cfs_cpt_table *cptable;
704 LASSERT(conf->psc_buf.bc_nbufs > 0);
705 LASSERT(conf->psc_buf.bc_buf_size >=
706 conf->psc_buf.bc_req_max_size + SPTLRPC_MAX_PAYLOAD);
707 LASSERT(conf->psc_thr.tc_ctx_tags != 0);
709 cptable = cconf->cc_cptable;
711 cptable = cfs_cpt_table;
713 if (!conf->psc_thr.tc_cpu_affinity) {
716 ncpts = cfs_cpt_number(cptable);
717 if (cconf->cc_pattern != NULL) {
718 struct cfs_expr_list *el;
720 rc = cfs_expr_list_parse(cconf->cc_pattern,
721 strlen(cconf->cc_pattern),
724 CERROR("%s: invalid CPT pattern string: %s",
725 conf->psc_name, cconf->cc_pattern);
726 RETURN(ERR_PTR(-EINVAL));
729 rc = cfs_expr_list_values(el, ncpts, &cpts);
730 cfs_expr_list_free(el);
732 CERROR("%s: failed to parse CPT array %s: %d\n",
733 conf->psc_name, cconf->cc_pattern, rc);
735 OBD_FREE(cpts, sizeof(*cpts) * ncpts);
736 RETURN(ERR_PTR(rc < 0 ? rc : -EINVAL));
742 OBD_ALLOC(service, offsetof(struct ptlrpc_service, srv_parts[ncpts]));
743 if (service == NULL) {
745 OBD_FREE(cpts, sizeof(*cpts) * ncpts);
746 RETURN(ERR_PTR(-ENOMEM));
749 service->srv_cptable = cptable;
750 service->srv_cpts = cpts;
751 service->srv_ncpts = ncpts;
753 service->srv_cpt_bits = 0; /* it's zero already, easy to read... */
754 while ((1 << service->srv_cpt_bits) < cfs_cpt_number(cptable))
755 service->srv_cpt_bits++;
758 spin_lock_init(&service->srv_lock);
759 service->srv_name = conf->psc_name;
760 service->srv_watchdog_factor = conf->psc_watchdog_factor;
761 INIT_LIST_HEAD(&service->srv_list); /* for safty of cleanup */
763 /* buffer configuration */
764 service->srv_nbuf_per_group = test_req_buffer_pressure ?
765 1 : conf->psc_buf.bc_nbufs;
766 service->srv_max_req_size = conf->psc_buf.bc_req_max_size +
768 service->srv_buf_size = conf->psc_buf.bc_buf_size;
769 service->srv_rep_portal = conf->psc_buf.bc_rep_portal;
770 service->srv_req_portal = conf->psc_buf.bc_req_portal;
772 /* Increase max reply size to next power of two */
773 service->srv_max_reply_size = 1;
774 while (service->srv_max_reply_size <
775 conf->psc_buf.bc_rep_max_size + SPTLRPC_MAX_PAYLOAD)
776 service->srv_max_reply_size <<= 1;
778 service->srv_thread_name = conf->psc_thr.tc_thr_name;
779 service->srv_ctx_tags = conf->psc_thr.tc_ctx_tags;
780 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
781 service->srv_ops = conf->psc_ops;
783 for (i = 0; i < ncpts; i++) {
784 if (!conf->psc_thr.tc_cpu_affinity)
787 cpt = cpts != NULL ? cpts[i] : i;
789 OBD_CPT_ALLOC(svcpt, cptable, cpt, sizeof(*svcpt));
791 GOTO(failed, rc = -ENOMEM);
793 service->srv_parts[i] = svcpt;
794 rc = ptlrpc_service_part_init(service, svcpt, cpt);
799 ptlrpc_server_nthreads_check(service, conf);
801 rc = LNetSetLazyPortal(service->srv_req_portal);
804 mutex_lock(&ptlrpc_all_services_mutex);
805 list_add(&service->srv_list, &ptlrpc_all_services);
806 mutex_unlock(&ptlrpc_all_services_mutex);
808 if (proc_entry != NULL)
809 ptlrpc_lprocfs_register_service(proc_entry, service);
811 rc = ptlrpc_service_nrs_setup(service);
815 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
816 service->srv_name, service->srv_req_portal);
818 rc = ptlrpc_start_threads(service);
820 CERROR("Failed to start threads for service %s: %d\n",
821 service->srv_name, rc);
827 ptlrpc_unregister_service(service);
830 EXPORT_SYMBOL(ptlrpc_register_service);
833 * to actually free the request, must be called without holding svc_lock.
834 * note it's caller's responsibility to unlink req->rq_list.
836 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
838 LASSERT(atomic_read(&req->rq_refcount) == 0);
839 LASSERT(list_empty(&req->rq_timed_list));
841 /* DEBUG_REQ() assumes the reply state of a request with a valid
842 * ref will not be destroyed until that reference is dropped. */
843 ptlrpc_req_drop_rs(req);
845 sptlrpc_svc_ctx_decref(req);
847 if (req != &req->rq_rqbd->rqbd_req) {
848 /* NB request buffers use an embedded
849 * req if the incoming req unlinked the
850 * MD; this isn't one of them! */
851 ptlrpc_request_cache_free(req);
856 * drop a reference count of the request. if it reaches 0, we either
857 * put it into history list, or free it immediately.
859 void ptlrpc_server_drop_request(struct ptlrpc_request *req)
861 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
862 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
863 struct ptlrpc_service *svc = svcpt->scp_service;
865 struct list_head *tmp;
866 struct list_head *nxt;
868 if (!atomic_dec_and_test(&req->rq_refcount))
871 if (req->rq_session.lc_state == LCS_ENTERED) {
872 lu_context_exit(&req->rq_session);
873 lu_context_fini(&req->rq_session);
876 if (req->rq_at_linked) {
877 spin_lock(&svcpt->scp_at_lock);
878 /* recheck with lock, in case it's unlinked by
879 * ptlrpc_at_check_timed() */
880 if (likely(req->rq_at_linked))
881 ptlrpc_at_remove_timed(req);
882 spin_unlock(&svcpt->scp_at_lock);
885 LASSERT(list_empty(&req->rq_timed_list));
887 /* finalize request */
888 if (req->rq_export) {
889 class_export_put(req->rq_export);
890 req->rq_export = NULL;
893 spin_lock(&svcpt->scp_lock);
895 list_add(&req->rq_list, &rqbd->rqbd_reqs);
897 refcount = --(rqbd->rqbd_refcount);
899 /* request buffer is now idle: add to history */
900 list_del(&rqbd->rqbd_list);
902 list_add_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds);
903 svcpt->scp_hist_nrqbds++;
905 /* cull some history?
906 * I expect only about 1 or 2 rqbds need to be recycled here */
907 while (svcpt->scp_hist_nrqbds > svc->srv_hist_nrqbds_cpt_max) {
908 rqbd = list_entry(svcpt->scp_hist_rqbds.next,
909 struct ptlrpc_request_buffer_desc,
912 list_del(&rqbd->rqbd_list);
913 svcpt->scp_hist_nrqbds--;
915 /* remove rqbd's reqs from svc's req history while
916 * I've got the service lock */
917 list_for_each(tmp, &rqbd->rqbd_reqs) {
918 req = list_entry(tmp, struct ptlrpc_request,
920 /* Track the highest culled req seq */
921 if (req->rq_history_seq >
922 svcpt->scp_hist_seq_culled) {
923 svcpt->scp_hist_seq_culled =
926 list_del(&req->rq_history_list);
929 spin_unlock(&svcpt->scp_lock);
931 list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
932 req = list_entry(rqbd->rqbd_reqs.next,
933 struct ptlrpc_request,
935 list_del(&req->rq_list);
936 ptlrpc_server_free_request(req);
939 spin_lock(&svcpt->scp_lock);
941 * now all reqs including the embedded req has been
942 * disposed, schedule request buffer for re-use.
944 LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) == 0);
945 list_add_tail(&rqbd->rqbd_list, &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 LASSERT(!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);
976 /* export has one reference already, so it`s safe to
977 * add req to export queue here and get another
978 * reference for request later */
979 spin_lock_bh(&export->exp_rpc_lock);
980 if (req->rq_ops != NULL) /* hp request */
981 list_add(&req->rq_exp_list, &export->exp_hp_rpcs);
983 list_add(&req->rq_exp_list, &export->exp_reg_rpcs);
984 spin_unlock_bh(&export->exp_rpc_lock);
986 class_export_rpc_dec(req->rq_export);
987 class_export_put(req->rq_export);
990 /* request takes one export refcount */
991 req->rq_export = class_export_get(export);
992 class_export_rpc_inc(export);
998 * to finish a request: stop sending more early replies, and release
1001 static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt,
1002 struct ptlrpc_request *req)
1004 ptlrpc_server_hpreq_fini(req);
1006 ptlrpc_server_drop_request(req);
1010 * to finish an active request: stop sending more early replies, and release
1011 * the request. should be called after we finished handling the request.
1013 static void ptlrpc_server_finish_active_request(
1014 struct ptlrpc_service_part *svcpt,
1015 struct ptlrpc_request *req)
1017 spin_lock(&svcpt->scp_req_lock);
1018 ptlrpc_nrs_req_stop_nolock(req);
1019 svcpt->scp_nreqs_active--;
1021 svcpt->scp_nhreqs_active--;
1022 spin_unlock(&svcpt->scp_req_lock);
1024 ptlrpc_nrs_req_finalize(req);
1026 if (req->rq_export != NULL)
1027 class_export_rpc_dec(req->rq_export);
1029 ptlrpc_server_finish_request(svcpt, req);
1033 * This function makes sure dead exports are evicted in a timely manner.
1034 * This function is only called when some export receives a message (i.e.,
1035 * the network is up.)
1037 void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
1039 struct obd_export *oldest_exp;
1040 time_t oldest_time, new_time;
1046 /* Compensate for slow machines, etc, by faking our request time
1047 into the future. Although this can break the strict time-ordering
1048 of the list, we can be really lazy here - we don't have to evict
1049 at the exact right moment. Eventually, all silent exports
1050 will make it to the top of the list. */
1052 /* Do not pay attention on 1sec or smaller renewals. */
1053 new_time = cfs_time_current_sec() + extra_delay;
1054 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
1057 exp->exp_last_request_time = new_time;
1059 /* exports may get disconnected from the chain even though the
1060 export has references, so we must keep the spin lock while
1061 manipulating the lists */
1062 spin_lock(&exp->exp_obd->obd_dev_lock);
1064 if (list_empty(&exp->exp_obd_chain_timed)) {
1065 /* this one is not timed */
1066 spin_unlock(&exp->exp_obd->obd_dev_lock);
1070 list_move_tail(&exp->exp_obd_chain_timed,
1071 &exp->exp_obd->obd_exports_timed);
1073 oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
1074 struct obd_export, exp_obd_chain_timed);
1075 oldest_time = oldest_exp->exp_last_request_time;
1076 spin_unlock(&exp->exp_obd->obd_dev_lock);
1078 if (exp->exp_obd->obd_recovering) {
1079 /* be nice to everyone during recovery */
1084 /* Note - racing to start/reset the obd_eviction timer is safe */
1085 if (exp->exp_obd->obd_eviction_timer == 0) {
1086 /* Check if the oldest entry is expired. */
1087 if (cfs_time_current_sec() > (oldest_time + PING_EVICT_TIMEOUT +
1089 /* We need a second timer, in case the net was down and
1090 * it just came back. Since the pinger may skip every
1091 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
1092 * we better wait for 3. */
1093 exp->exp_obd->obd_eviction_timer =
1094 cfs_time_current_sec() + 3 * PING_INTERVAL;
1095 CDEBUG(D_HA, "%s: Think about evicting %s from "CFS_TIME_T"\n",
1096 exp->exp_obd->obd_name,
1097 obd_export_nid2str(oldest_exp), oldest_time);
1100 if (cfs_time_current_sec() >
1101 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
1102 /* The evictor won't evict anyone who we've heard from
1103 * recently, so we don't have to check before we start
1105 if (!ping_evictor_wake(exp))
1106 exp->exp_obd->obd_eviction_timer = 0;
1114 * Sanity check request \a req.
1115 * Return 0 if all is ok, error code otherwise.
1117 static int ptlrpc_check_req(struct ptlrpc_request *req)
1119 struct obd_device *obd = req->rq_export->exp_obd;
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(obd == NULL || obd->obd_fail)) {
1131 /* Failing over, don't handle any more reqs,
1132 * send error response instead. */
1133 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
1134 req, (obd != NULL) ? obd->obd_name : "unknown");
1136 } else if (lustre_msg_get_flags(req->rq_reqmsg) &
1137 (MSG_REPLAY | MSG_REQ_REPLAY_DONE) &&
1138 !obd->obd_recovering) {
1139 DEBUG_REQ(D_ERROR, req,
1140 "Invalid replay without recovery");
1141 class_fail_export(req->rq_export);
1143 } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 &&
1144 !obd->obd_recovering) {
1145 DEBUG_REQ(D_ERROR, req, "Invalid req with transno "
1146 "%llu without recovery",
1147 lustre_msg_get_transno(req->rq_reqmsg));
1148 class_fail_export(req->rq_export);
1152 if (unlikely(rc < 0)) {
1153 req->rq_status = rc;
1159 static void ptlrpc_at_set_timer(struct ptlrpc_service_part *svcpt)
1161 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1164 if (array->paa_count == 0) {
1165 del_timer(&svcpt->scp_at_timer);
1169 /* Set timer for closest deadline */
1170 next = (__s32)(array->paa_deadline - cfs_time_current_sec() -
1173 ptlrpc_at_timer((unsigned long)svcpt);
1175 mod_timer(&svcpt->scp_at_timer, cfs_time_shift(next));
1176 CDEBUG(D_INFO, "armed %s at %+ds\n",
1177 svcpt->scp_service->srv_name, next);
1181 /* Add rpc to early reply check list */
1182 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
1184 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1185 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1186 struct ptlrpc_request *rq = NULL;
1192 if (req->rq_no_reply)
1195 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
1198 spin_lock(&svcpt->scp_at_lock);
1199 LASSERT(list_empty(&req->rq_timed_list));
1201 index = (unsigned long)req->rq_deadline % array->paa_size;
1202 if (array->paa_reqs_count[index] > 0) {
1203 /* latest rpcs will have the latest deadlines in the list,
1204 * so search backward. */
1205 list_for_each_entry_reverse(rq,
1206 &array->paa_reqs_array[index],
1208 if (req->rq_deadline >= rq->rq_deadline) {
1209 list_add(&req->rq_timed_list,
1210 &rq->rq_timed_list);
1216 /* Add the request at the head of the list */
1217 if (list_empty(&req->rq_timed_list))
1218 list_add(&req->rq_timed_list,
1219 &array->paa_reqs_array[index]);
1221 spin_lock(&req->rq_lock);
1222 req->rq_at_linked = 1;
1223 spin_unlock(&req->rq_lock);
1224 req->rq_at_index = index;
1225 array->paa_reqs_count[index]++;
1227 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
1228 array->paa_deadline = req->rq_deadline;
1229 ptlrpc_at_set_timer(svcpt);
1231 spin_unlock(&svcpt->scp_at_lock);
1237 ptlrpc_at_remove_timed(struct ptlrpc_request *req)
1239 struct ptlrpc_at_array *array;
1241 array = &req->rq_rqbd->rqbd_svcpt->scp_at_array;
1243 /* NB: must call with hold svcpt::scp_at_lock */
1244 LASSERT(!list_empty(&req->rq_timed_list));
1245 list_del_init(&req->rq_timed_list);
1247 spin_lock(&req->rq_lock);
1248 req->rq_at_linked = 0;
1249 spin_unlock(&req->rq_lock);
1251 array->paa_reqs_count[req->rq_at_index]--;
1256 * Attempt to extend the request deadline by sending an early reply to the
1259 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
1261 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1262 struct ptlrpc_request *reqcopy;
1263 struct lustre_msg *reqmsg;
1264 cfs_duration_t olddl = req->rq_deadline - cfs_time_current_sec();
1270 if (CFS_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_RECONNECT)) {
1271 /* don't send early reply */
1275 /* deadline is when the client expects us to reply, margin is the
1276 difference between clients' and servers' expectations */
1277 DEBUG_REQ(D_ADAPTTO, req,
1278 "%ssending early reply (deadline %+lds, margin %+lds) for "
1279 "%d+%d", AT_OFF ? "AT off - not " : "",
1280 olddl, olddl - at_get(&svcpt->scp_at_estimate),
1281 at_get(&svcpt->scp_at_estimate), at_extra);
1287 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), "
1288 "not sending early reply. Consider increasing "
1289 "at_early_margin (%d)?", olddl, at_early_margin);
1291 /* Return an error so we're not re-added to the timed list. */
1295 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
1296 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
1297 "but no AT support");
1301 if (req->rq_export &&
1302 lustre_msg_get_flags(req->rq_reqmsg) &
1303 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1304 struct obd_device *obd_exp = req->rq_export->exp_obd;
1306 /* During recovery, we don't want to send too many early
1307 * replies, but on the other hand we want to make sure the
1308 * client has enough time to resend if the rpc is lost. So
1309 * during the recovery period send at least 4 early replies,
1310 * spacing them every at_extra if we can. at_estimate should
1311 * always equal this fixed value during recovery.
1313 /* Don't account request processing time into AT history
1314 * during recovery, it is not service time we need but
1315 * includes also waiting time for recovering clients
1317 newdl = min_t(time64_t, at_extra,
1318 obd_exp->obd_recovery_timeout / 4) +
1319 ktime_get_real_seconds();
1321 /* We want to extend the request deadline by at_extra seconds,
1322 * so we set our service estimate to reflect how much time has
1323 * passed since this request arrived plus an additional
1324 * at_extra seconds. The client will calculate the new deadline
1325 * based on this service estimate (plus some additional time to
1326 * account for network latency). See ptlrpc_at_recv_early_reply
1328 at_measured(&svcpt->scp_at_estimate, at_extra +
1329 cfs_time_current_sec() -
1330 req->rq_arrival_time.tv_sec);
1331 newdl = req->rq_arrival_time.tv_sec +
1332 at_get(&svcpt->scp_at_estimate);
1335 /* Check to see if we've actually increased the deadline -
1336 * we may be past adaptive_max */
1337 if (req->rq_deadline >= newdl) {
1338 DEBUG_REQ(D_WARNING, req, "Couldn't add any time "
1339 "(%ld/%ld), not sending early reply\n",
1340 olddl, newdl - cfs_time_current_sec());
1344 reqcopy = ptlrpc_request_cache_alloc(GFP_NOFS);
1345 if (reqcopy == NULL)
1347 OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1349 GOTO(out_free, rc = -ENOMEM);
1352 reqcopy->rq_reply_state = NULL;
1353 reqcopy->rq_rep_swab_mask = 0;
1354 reqcopy->rq_pack_bulk = 0;
1355 reqcopy->rq_pack_udesc = 0;
1356 reqcopy->rq_packed_final = 0;
1357 sptlrpc_svc_ctx_addref(reqcopy);
1358 /* We only need the reqmsg for the magic */
1359 reqcopy->rq_reqmsg = reqmsg;
1360 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1363 * tgt_brw_read() and tgt_brw_write() may have decided not to reply.
1364 * Without this check, we would fail the rq_no_reply assertion in
1365 * ptlrpc_send_reply().
1367 if (reqcopy->rq_no_reply)
1368 GOTO(out, rc = -ETIMEDOUT);
1370 LASSERT(atomic_read(&req->rq_refcount));
1371 /** if it is last refcount then early reply isn't needed */
1372 if (atomic_read(&req->rq_refcount) == 1) {
1373 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
1374 "abort sending early reply\n");
1375 GOTO(out, rc = -EINVAL);
1378 /* Connection ref */
1379 reqcopy->rq_export = class_conn2export(
1380 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1381 if (reqcopy->rq_export == NULL)
1382 GOTO(out, rc = -ENODEV);
1385 class_export_rpc_inc(reqcopy->rq_export);
1386 if (reqcopy->rq_export->exp_obd &&
1387 reqcopy->rq_export->exp_obd->obd_fail)
1388 GOTO(out_put, rc = -ENODEV);
1390 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1394 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1397 /* Adjust our own deadline to what we told the client */
1398 req->rq_deadline = newdl;
1399 req->rq_early_count++; /* number sent, server side */
1401 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1404 /* Free the (early) reply state from lustre_pack_reply.
1405 (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
1406 ptlrpc_req_drop_rs(reqcopy);
1409 class_export_rpc_dec(reqcopy->rq_export);
1410 class_export_put(reqcopy->rq_export);
1412 sptlrpc_svc_ctx_decref(reqcopy);
1413 OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1415 ptlrpc_request_cache_free(reqcopy);
1419 /* Send early replies to everybody expiring within at_early_margin
1420 asking for at_extra time */
1421 static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1423 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1424 struct ptlrpc_request *rq, *n;
1425 struct list_head work_list;
1428 time_t now = cfs_time_current_sec();
1429 cfs_duration_t delay;
1430 int first, counter = 0;
1433 spin_lock(&svcpt->scp_at_lock);
1434 if (svcpt->scp_at_check == 0) {
1435 spin_unlock(&svcpt->scp_at_lock);
1438 delay = cfs_time_sub(cfs_time_current(), svcpt->scp_at_checktime);
1439 svcpt->scp_at_check = 0;
1441 if (array->paa_count == 0) {
1442 spin_unlock(&svcpt->scp_at_lock);
1446 /* The timer went off, but maybe the nearest rpc already completed. */
1447 first = array->paa_deadline - now;
1448 if (first > at_early_margin) {
1449 /* We've still got plenty of time. Reset the timer. */
1450 ptlrpc_at_set_timer(svcpt);
1451 spin_unlock(&svcpt->scp_at_lock);
1455 /* We're close to a timeout, and we don't know how much longer the
1456 server will take. Send early replies to everyone expiring soon. */
1457 INIT_LIST_HEAD(&work_list);
1459 index = (unsigned long)array->paa_deadline % array->paa_size;
1460 count = array->paa_count;
1462 count -= array->paa_reqs_count[index];
1463 list_for_each_entry_safe(rq, n,
1464 &array->paa_reqs_array[index],
1466 if (rq->rq_deadline > now + at_early_margin) {
1467 /* update the earliest deadline */
1468 if (deadline == -1 ||
1469 rq->rq_deadline < deadline)
1470 deadline = rq->rq_deadline;
1474 ptlrpc_at_remove_timed(rq);
1476 * ptlrpc_server_drop_request() may drop
1477 * refcount to 0 already. Let's check this and
1478 * don't add entry to work_list
1480 if (likely(atomic_inc_not_zero(&rq->rq_refcount)))
1481 list_add(&rq->rq_timed_list, &work_list);
1485 if (++index >= array->paa_size)
1488 array->paa_deadline = deadline;
1489 /* we have a new earliest deadline, restart the timer */
1490 ptlrpc_at_set_timer(svcpt);
1492 spin_unlock(&svcpt->scp_at_lock);
1494 CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
1495 "replies\n", first, at_extra, counter);
1497 /* We're already past request deadlines before we even get a
1498 chance to send early replies */
1499 LCONSOLE_WARN("%s: This server is not able to keep up with "
1500 "request traffic (cpu-bound).\n",
1501 svcpt->scp_service->srv_name);
1502 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, "
1503 "delay="CFS_DURATION_T"(jiff)\n",
1504 counter, svcpt->scp_nreqs_incoming,
1505 svcpt->scp_nreqs_active,
1506 at_get(&svcpt->scp_at_estimate), delay);
1509 /* we took additional refcount so entries can't be deleted from list, no
1510 * locking is needed */
1511 while (!list_empty(&work_list)) {
1512 rq = list_entry(work_list.next, struct ptlrpc_request,
1514 list_del_init(&rq->rq_timed_list);
1516 if (ptlrpc_at_send_early_reply(rq) == 0)
1517 ptlrpc_at_add_timed(rq);
1519 ptlrpc_server_drop_request(rq);
1522 RETURN(1); /* return "did_something" for liblustre */
1525 /* Check if we are already handling earlier incarnation of this request.
1526 * Called under &req->rq_export->exp_rpc_lock locked */
1527 static int ptlrpc_server_check_resend_in_progress(struct ptlrpc_request *req)
1529 struct ptlrpc_request *tmp = NULL;
1531 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ||
1532 (atomic_read(&req->rq_export->exp_rpc_count) == 0))
1535 /* bulk request are aborted upon reconnect, don't try to
1537 if (req->rq_bulk_write || req->rq_bulk_read)
1540 /* This list should not be longer than max_requests in
1541 * flights on the client, so it is not all that long.
1542 * Also we only hit this codepath in case of a resent
1543 * request which makes it even more rarely hit */
1544 list_for_each_entry(tmp, &req->rq_export->exp_reg_rpcs,
1546 /* Found duplicate one */
1547 if (tmp->rq_xid == req->rq_xid)
1550 list_for_each_entry(tmp, &req->rq_export->exp_hp_rpcs,
1552 /* Found duplicate one */
1553 if (tmp->rq_xid == req->rq_xid)
1559 DEBUG_REQ(D_HA, req, "Found duplicate req in processing");
1560 DEBUG_REQ(D_HA, tmp, "Request being processed");
1565 * Check if a request should be assigned with a high priority.
1567 * \retval < 0: error occurred
1568 * 0: normal RPC request
1569 * +1: high priority request
1571 static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt,
1572 struct ptlrpc_request *req)
1577 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL) {
1578 rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
1585 if (req->rq_export != NULL && req->rq_ops != NULL) {
1586 /* Perform request specific check. We should do this
1587 * check before the request is added into exp_hp_rpcs
1588 * list otherwise it may hit swab race at LU-1044. */
1589 if (req->rq_ops->hpreq_check != NULL) {
1590 rc = req->rq_ops->hpreq_check(req);
1591 if (rc == -ESTALE) {
1592 req->rq_status = rc;
1595 /** can only return error,
1596 * 0 for normal request,
1597 * or 1 for high priority request */
1605 /** Remove the request from the export list. */
1606 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req)
1609 if (req->rq_export) {
1610 /* refresh lock timeout again so that client has more
1611 * room to send lock cancel RPC. */
1612 if (req->rq_ops && req->rq_ops->hpreq_fini)
1613 req->rq_ops->hpreq_fini(req);
1615 spin_lock_bh(&req->rq_export->exp_rpc_lock);
1616 list_del_init(&req->rq_exp_list);
1617 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1622 static int ptlrpc_hpreq_check(struct ptlrpc_request *req)
1627 static struct ptlrpc_hpreq_ops ptlrpc_hpreq_common = {
1628 .hpreq_check = ptlrpc_hpreq_check,
1631 /* Hi-Priority RPC check by RPC operation code. */
1632 int ptlrpc_hpreq_handler(struct ptlrpc_request *req)
1634 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1636 /* Check for export to let only reconnects for not yet evicted
1637 * export to become a HP rpc. */
1638 if ((req->rq_export != NULL) &&
1639 (opc == OBD_PING || opc == MDS_CONNECT || opc == OST_CONNECT))
1640 req->rq_ops = &ptlrpc_hpreq_common;
1644 EXPORT_SYMBOL(ptlrpc_hpreq_handler);
1646 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1647 struct ptlrpc_request *req)
1653 rc = ptlrpc_server_hpreq_init(svcpt, req);
1658 ptlrpc_nrs_req_initialize(svcpt, req, hp);
1660 if (req->rq_export != NULL) {
1661 struct obd_export *exp = req->rq_export;
1663 /* do search for duplicated xid and the adding to the list
1665 spin_lock_bh(&exp->exp_rpc_lock);
1666 rc = ptlrpc_server_check_resend_in_progress(req);
1668 spin_unlock_bh(&exp->exp_rpc_lock);
1670 ptlrpc_nrs_req_finalize(req);
1674 if (hp || req->rq_ops != NULL)
1675 list_add(&req->rq_exp_list, &exp->exp_hp_rpcs);
1677 list_add(&req->rq_exp_list, &exp->exp_reg_rpcs);
1678 spin_unlock_bh(&exp->exp_rpc_lock);
1681 /* the current thread is not the processing thread for this request
1682 * since that, but request is in exp_hp_list and can be find there.
1683 * Remove all relations between request and old thread. */
1684 req->rq_svc_thread->t_env->le_ses = NULL;
1685 req->rq_svc_thread = NULL;
1686 req->rq_session.lc_thread = NULL;
1688 ptlrpc_nrs_req_add(svcpt, req, hp);
1694 * Allow to handle high priority request
1695 * User can call it w/o any lock but need to hold
1696 * ptlrpc_service_part::scp_req_lock to get reliable result
1698 static bool ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1701 int running = svcpt->scp_nthrs_running;
1703 if (!nrs_svcpt_has_hp(svcpt))
1709 if (ptlrpc_nrs_req_throttling_nolock(svcpt, true))
1712 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1713 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1714 /* leave just 1 thread for normal RPCs */
1715 running = PTLRPC_NTHRS_INIT;
1716 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1720 if (svcpt->scp_nreqs_active >= running - 1)
1723 if (svcpt->scp_nhreqs_active == 0)
1726 return !ptlrpc_nrs_req_pending_nolock(svcpt, false) ||
1727 svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1730 static bool ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1733 return ptlrpc_server_allow_high(svcpt, force) &&
1734 ptlrpc_nrs_req_pending_nolock(svcpt, true);
1738 * Only allow normal priority requests on a service that has a high-priority
1739 * queue if forced (i.e. cleanup), if there are other high priority requests
1740 * already being processed (i.e. those threads can service more high-priority
1741 * requests), or if there are enough idle threads that a later thread can do
1742 * a high priority request.
1743 * User can call it w/o any lock but need to hold
1744 * ptlrpc_service_part::scp_req_lock to get reliable result
1746 static bool ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1749 int running = svcpt->scp_nthrs_running;
1750 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1751 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1752 /* leave just 1 thread for normal RPCs */
1753 running = PTLRPC_NTHRS_INIT;
1754 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1761 if (ptlrpc_nrs_req_throttling_nolock(svcpt, false))
1764 if (svcpt->scp_nreqs_active < running - 2)
1767 if (svcpt->scp_nreqs_active >= running - 1)
1770 return svcpt->scp_nhreqs_active > 0 || !nrs_svcpt_has_hp(svcpt);
1773 static bool ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1776 return ptlrpc_server_allow_normal(svcpt, force) &&
1777 ptlrpc_nrs_req_pending_nolock(svcpt, false);
1781 * Returns true if there are requests available in incoming
1782 * request queue for processing and it is allowed to fetch them.
1783 * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1784 * to get reliable result
1785 * \see ptlrpc_server_allow_normal
1786 * \see ptlrpc_server_allow high
1789 ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt, bool force)
1791 return ptlrpc_server_high_pending(svcpt, force) ||
1792 ptlrpc_server_normal_pending(svcpt, force);
1796 * Fetch a request for processing from queue of unprocessed requests.
1797 * Favors high-priority requests.
1798 * Returns a pointer to fetched request.
1800 static struct ptlrpc_request *
1801 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force)
1803 struct ptlrpc_request *req = NULL;
1806 spin_lock(&svcpt->scp_req_lock);
1808 if (ptlrpc_server_high_pending(svcpt, force)) {
1809 req = ptlrpc_nrs_req_get_nolock(svcpt, true, force);
1811 svcpt->scp_hreq_count++;
1816 if (ptlrpc_server_normal_pending(svcpt, force)) {
1817 req = ptlrpc_nrs_req_get_nolock(svcpt, false, force);
1819 svcpt->scp_hreq_count = 0;
1824 spin_unlock(&svcpt->scp_req_lock);
1828 svcpt->scp_nreqs_active++;
1830 svcpt->scp_nhreqs_active++;
1832 spin_unlock(&svcpt->scp_req_lock);
1834 if (likely(req->rq_export))
1835 class_export_rpc_inc(req->rq_export);
1841 * Handle freshly incoming reqs, add to timed early reply list,
1842 * pass on to regular request queue.
1843 * All incoming requests pass through here before getting into
1844 * ptlrpc_server_handle_req later on.
1847 ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt,
1848 struct ptlrpc_thread *thread)
1850 struct ptlrpc_service *svc = svcpt->scp_service;
1851 struct ptlrpc_request *req;
1856 spin_lock(&svcpt->scp_lock);
1857 if (list_empty(&svcpt->scp_req_incoming)) {
1858 spin_unlock(&svcpt->scp_lock);
1862 req = list_entry(svcpt->scp_req_incoming.next,
1863 struct ptlrpc_request, rq_list);
1864 list_del_init(&req->rq_list);
1865 svcpt->scp_nreqs_incoming--;
1866 /* Consider this still a "queued" request as far as stats are
1868 spin_unlock(&svcpt->scp_lock);
1870 /* go through security check/transform */
1871 rc = sptlrpc_svc_unwrap_request(req);
1875 case SECSVC_COMPLETE:
1876 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1885 * for null-flavored rpc, msg has been unpacked by sptlrpc, although
1886 * redo it wouldn't be harmful.
1888 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
1889 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
1891 CERROR("error unpacking request: ptl %d from %s "
1892 "x%llu\n", svc->srv_req_portal,
1893 libcfs_id2str(req->rq_peer), req->rq_xid);
1898 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1900 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
1901 "%llu\n", svc->srv_req_portal,
1902 libcfs_id2str(req->rq_peer), req->rq_xid);
1906 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
1907 lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
1908 CERROR("drop incoming rpc opc %u, x%llu\n",
1909 cfs_fail_val, req->rq_xid);
1914 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1915 CERROR("wrong packet type received (type=%u) from %s\n",
1916 lustre_msg_get_type(req->rq_reqmsg),
1917 libcfs_id2str(req->rq_peer));
1921 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1925 req->rq_bulk_write = 1;
1929 case MGS_CONFIG_READ:
1930 req->rq_bulk_read = 1;
1934 CDEBUG(D_RPCTRACE, "got req x%llu\n", req->rq_xid);
1936 req->rq_export = class_conn2export(
1937 lustre_msg_get_handle(req->rq_reqmsg));
1938 if (req->rq_export) {
1939 rc = ptlrpc_check_req(req);
1941 rc = sptlrpc_target_export_check(req->rq_export, req);
1943 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
1944 "illegal security flavor,");
1949 ptlrpc_update_export_timer(req->rq_export, 0);
1952 /* req_in handling should/must be fast */
1953 if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5)
1954 DEBUG_REQ(D_WARNING, req, "Slow req_in handling "CFS_DURATION_T"s",
1955 cfs_time_sub(cfs_time_current_sec(),
1956 req->rq_arrival_time.tv_sec));
1958 /* Set rpc server deadline and add it to the timed list */
1959 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1960 MSGHDR_AT_SUPPORT) ?
1961 /* The max time the client expects us to take */
1962 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1964 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1965 if (unlikely(deadline == 0)) {
1966 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1970 /* Skip early reply */
1971 if (OBD_FAIL_PRECHECK(OBD_FAIL_MDS_RESEND))
1972 req->rq_deadline += obd_timeout;
1974 req->rq_svc_thread = thread;
1975 if (thread != NULL) {
1976 /* initialize request session, it is needed for request
1977 * processing by target */
1978 rc = lu_context_init(&req->rq_session, LCT_SERVER_SESSION |
1981 CERROR("%s: failure to initialize session: rc = %d\n",
1982 thread->t_name, rc);
1985 req->rq_session.lc_thread = thread;
1986 lu_context_enter(&req->rq_session);
1987 thread->t_env->le_ses = &req->rq_session;
1990 ptlrpc_at_add_timed(req);
1992 /* Move it over to the request processing queue */
1993 rc = ptlrpc_server_request_add(svcpt, req);
1997 wake_up(&svcpt->scp_waitq);
2001 ptlrpc_server_finish_request(svcpt, req);
2007 * Main incoming request handling logic.
2008 * Calls handler function from service to do actual processing.
2011 ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
2012 struct ptlrpc_thread *thread)
2014 struct ptlrpc_service *svc = svcpt->scp_service;
2015 struct ptlrpc_request *request;
2016 struct timeval work_start;
2017 struct timeval work_end;
2023 request = ptlrpc_server_request_get(svcpt, false);
2024 if (request == NULL)
2027 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
2028 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
2029 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
2030 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
2032 if (unlikely(fail_opc)) {
2033 if (request->rq_export && request->rq_ops)
2034 OBD_FAIL_TIMEOUT(fail_opc, 4);
2037 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
2039 if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
2040 libcfs_debug_dumplog();
2042 do_gettimeofday(&work_start);
2043 timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
2044 if (likely(svc->srv_stats != NULL)) {
2045 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
2047 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
2048 svcpt->scp_nreqs_incoming);
2049 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
2050 svcpt->scp_nreqs_active);
2051 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
2052 at_get(&svcpt->scp_at_estimate));
2055 if (likely(request->rq_export)) {
2056 if (unlikely(ptlrpc_check_req(request)))
2058 ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
2061 /* Discard requests queued for longer than the deadline.
2062 The deadline is increased if we send an early reply. */
2063 if (cfs_time_current_sec() > request->rq_deadline) {
2064 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s"
2065 ": deadline "CFS_DURATION_T":"CFS_DURATION_T"s ago\n",
2066 libcfs_id2str(request->rq_peer),
2067 cfs_time_sub(request->rq_deadline,
2068 request->rq_arrival_time.tv_sec),
2069 cfs_time_sub(cfs_time_current_sec(),
2070 request->rq_deadline));
2074 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
2075 "%s:%s+%d:%d:x%llu:%s:%d\n", current_comm(),
2076 (request->rq_export ?
2077 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2078 (request->rq_export ?
2079 atomic_read(&request->rq_export->exp_refcount) : -99),
2080 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
2081 libcfs_id2str(request->rq_peer),
2082 lustre_msg_get_opc(request->rq_reqmsg));
2084 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
2085 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
2087 CDEBUG(D_NET, "got req %llu\n", request->rq_xid);
2089 /* re-assign request and sesson thread to the current one */
2090 request->rq_svc_thread = thread;
2091 if (thread != NULL) {
2092 LASSERT(request->rq_session.lc_thread == NULL);
2093 request->rq_session.lc_thread = thread;
2094 thread->t_env->le_ses = &request->rq_session;
2096 svc->srv_ops.so_req_handler(request);
2098 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
2101 if (unlikely(cfs_time_current_sec() > request->rq_deadline)) {
2102 DEBUG_REQ(D_WARNING, request, "Request took longer "
2103 "than estimated ("CFS_DURATION_T":"CFS_DURATION_T"s);"
2104 " client may timeout.",
2105 cfs_time_sub(request->rq_deadline,
2106 request->rq_arrival_time.tv_sec),
2107 cfs_time_sub(cfs_time_current_sec(),
2108 request->rq_deadline));
2111 do_gettimeofday(&work_end);
2112 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
2113 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
2114 "%s:%s+%d:%d:x%llu:%s:%d Request procesed in "
2115 "%ldus (%ldus total) trans %llu rc %d/%d\n",
2117 (request->rq_export ?
2118 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2119 (request->rq_export ?
2120 atomic_read(&request->rq_export->exp_refcount) : -99),
2121 lustre_msg_get_status(request->rq_reqmsg),
2123 libcfs_id2str(request->rq_peer),
2124 lustre_msg_get_opc(request->rq_reqmsg),
2126 cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL),
2127 (request->rq_repmsg ?
2128 lustre_msg_get_transno(request->rq_repmsg) :
2129 request->rq_transno),
2131 (request->rq_repmsg ?
2132 lustre_msg_get_status(request->rq_repmsg) : -999));
2133 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
2134 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
2135 int opc = opcode_offset(op);
2136 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
2137 LASSERT(opc < LUSTRE_MAX_OPCODES);
2138 lprocfs_counter_add(svc->srv_stats,
2139 opc + EXTRA_MAX_OPCODES,
2143 if (unlikely(request->rq_early_count)) {
2144 DEBUG_REQ(D_ADAPTTO, request,
2145 "sent %d early replies before finishing in "
2147 request->rq_early_count,
2148 cfs_time_sub(work_end.tv_sec,
2149 request->rq_arrival_time.tv_sec));
2152 ptlrpc_server_finish_active_request(svcpt, request);
2158 * An internal function to process a single reply state object.
2161 ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
2163 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
2164 struct ptlrpc_service *svc = svcpt->scp_service;
2165 struct obd_export *exp;
2170 exp = rs->rs_export;
2172 LASSERT(rs->rs_difficult);
2173 LASSERT(rs->rs_scheduled);
2174 LASSERT(list_empty(&rs->rs_list));
2176 /* The disk commit callback holds exp_uncommitted_replies_lock while it
2177 * iterates over newly committed replies, removing them from
2178 * exp_uncommitted_replies. It then drops this lock and schedules the
2179 * replies it found for handling here.
2181 * We can avoid contention for exp_uncommitted_replies_lock between the
2182 * HRT threads and further commit callbacks by checking rs_committed
2183 * which is set in the commit callback while it holds both
2184 * rs_lock and exp_uncommitted_reples.
2186 * If we see rs_committed clear, the commit callback _may_ not have
2187 * handled this reply yet and we race with it to grab
2188 * exp_uncommitted_replies_lock before removing the reply from
2189 * exp_uncommitted_replies. Note that if we lose the race and the
2190 * reply has already been removed, list_del_init() is a noop.
2192 * If we see rs_committed set, we know the commit callback is handling,
2193 * or has handled this reply since store reordering might allow us to
2194 * see rs_committed set out of sequence. But since this is done
2195 * holding rs_lock, we can be sure it has all completed once we hold
2196 * rs_lock, which we do right next.
2198 if (!rs->rs_committed) {
2199 /* if rs was commited, no need to convert locks, don't check
2200 * rs_committed here because rs may never be added into
2201 * exp_uncommitted_replies and this flag never be set, see
2202 * target_send_reply() */
2203 if (rs->rs_convert_lock &&
2204 rs->rs_transno > exp->exp_last_committed) {
2205 struct ldlm_lock *lock;
2207 spin_lock(&rs->rs_lock);
2208 if (rs->rs_convert_lock &&
2209 rs->rs_transno > exp->exp_last_committed) {
2210 nlocks = rs->rs_nlocks;
2211 while (nlocks-- > 0)
2212 rs->rs_modes[nlocks] = LCK_COS;
2213 nlocks = rs->rs_nlocks;
2214 rs->rs_convert_lock = 0;
2215 /* clear rs_scheduled so that commit callback
2216 * can schedule again */
2217 rs->rs_scheduled = 0;
2218 spin_unlock(&rs->rs_lock);
2220 while (nlocks-- > 0) {
2221 lock = ldlm_handle2lock(
2222 &rs->rs_locks[nlocks]);
2223 LASSERT(lock != NULL);
2224 ldlm_lock_downgrade(lock, LCK_COS);
2225 LDLM_LOCK_PUT(lock);
2229 spin_unlock(&rs->rs_lock);
2232 spin_lock(&exp->exp_uncommitted_replies_lock);
2233 list_del_init(&rs->rs_obd_list);
2234 spin_unlock(&exp->exp_uncommitted_replies_lock);
2237 spin_lock(&exp->exp_lock);
2238 /* Noop if removed already */
2239 list_del_init(&rs->rs_exp_list);
2240 spin_unlock(&exp->exp_lock);
2242 spin_lock(&rs->rs_lock);
2244 been_handled = rs->rs_handled;
2247 nlocks = rs->rs_nlocks; /* atomic "steal", but */
2248 rs->rs_nlocks = 0; /* locks still on rs_locks! */
2250 if (nlocks == 0 && !been_handled) {
2251 /* If we see this, we should already have seen the warning
2252 * in mds_steal_ack_locks() */
2253 CDEBUG(D_HA, "All locks stolen from rs %p x%lld.t%lld"
2256 rs->rs_xid, rs->rs_transno, rs->rs_opc,
2257 libcfs_nid2str(exp->exp_connection->c_peer.nid));
2260 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
2261 spin_unlock(&rs->rs_lock);
2263 if (!been_handled && rs->rs_on_net) {
2264 LNetMDUnlink(rs->rs_md_h);
2265 /* Ignore return code; we're racing with completion */
2268 while (nlocks-- > 0)
2269 ldlm_lock_decref(&rs->rs_locks[nlocks],
2270 rs->rs_modes[nlocks]);
2272 spin_lock(&rs->rs_lock);
2275 rs->rs_scheduled = 0;
2276 rs->rs_convert_lock = 0;
2278 if (!rs->rs_on_net) {
2280 spin_unlock(&rs->rs_lock);
2282 class_export_put (exp);
2283 rs->rs_export = NULL;
2284 ptlrpc_rs_decref(rs);
2285 if (atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
2286 svc->srv_is_stopping)
2287 wake_up_all(&svcpt->scp_waitq);
2291 /* still on the net; callback will schedule */
2292 spin_unlock(&rs->rs_lock);
2298 ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
2300 int avail = svcpt->scp_nrqbds_posted;
2301 int low_water = test_req_buffer_pressure ? 0 :
2302 svcpt->scp_service->srv_nbuf_per_group / 2;
2304 /* NB I'm not locking; just looking. */
2306 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
2307 * allowed the request history to grow out of control. We could put a
2308 * sanity check on that here and cull some history if we need the
2311 if (avail <= low_water)
2312 ptlrpc_grow_req_bufs(svcpt, 1);
2314 if (svcpt->scp_service->srv_stats) {
2315 lprocfs_counter_add(svcpt->scp_service->srv_stats,
2316 PTLRPC_REQBUF_AVAIL_CNTR, avail);
2321 ptlrpc_retry_rqbds(void *arg)
2323 struct ptlrpc_service_part *svcpt = (struct ptlrpc_service_part *)arg;
2325 svcpt->scp_rqbd_timeout = 0;
2330 ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
2332 return svcpt->scp_nreqs_active <
2333 svcpt->scp_nthrs_running - 1 -
2334 (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL);
2338 * allowed to create more threads
2339 * user can call it w/o any lock but need to hold
2340 * ptlrpc_service_part::scp_lock to get reliable result
2343 ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
2345 return svcpt->scp_nthrs_running +
2346 svcpt->scp_nthrs_starting <
2347 svcpt->scp_service->srv_nthrs_cpt_limit;
2351 * too many requests and allowed to create more threads
2354 ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
2356 return !ptlrpc_threads_enough(svcpt) &&
2357 ptlrpc_threads_increasable(svcpt);
2361 ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2363 return thread_is_stopping(thread) ||
2364 thread->t_svcpt->scp_service->srv_is_stopping;
2368 ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
2370 return !list_empty(&svcpt->scp_rqbd_idle) &&
2371 svcpt->scp_rqbd_timeout == 0;
2375 ptlrpc_at_check(struct ptlrpc_service_part *svcpt)
2377 return svcpt->scp_at_check;
2381 * requests wait on preprocessing
2382 * user can call it w/o any lock but need to hold
2383 * ptlrpc_service_part::scp_lock to get reliable result
2386 ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
2388 return !list_empty(&svcpt->scp_req_incoming);
2391 static __attribute__((__noinline__)) int
2392 ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
2393 struct ptlrpc_thread *thread)
2395 /* Don't exit while there are replies to be handled */
2396 struct l_wait_info lwi = LWI_TIMEOUT(svcpt->scp_rqbd_timeout,
2397 ptlrpc_retry_rqbds, svcpt);
2399 lc_watchdog_disable(thread->t_watchdog);
2403 l_wait_event_exclusive_head(svcpt->scp_waitq,
2404 ptlrpc_thread_stopping(thread) ||
2405 ptlrpc_server_request_incoming(svcpt) ||
2406 ptlrpc_server_request_pending(svcpt, false) ||
2407 ptlrpc_rqbd_pending(svcpt) ||
2408 ptlrpc_at_check(svcpt), &lwi);
2410 if (ptlrpc_thread_stopping(thread))
2413 lc_watchdog_touch(thread->t_watchdog,
2414 ptlrpc_server_get_timeout(svcpt));
2419 * Main thread body for service threads.
2420 * Waits in a loop waiting for new requests to process to appear.
2421 * Every time an incoming requests is added to its queue, a waitq
2422 * is woken up and one of the threads will handle it.
2424 static int ptlrpc_main(void *arg)
2426 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
2427 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2428 struct ptlrpc_service *svc = svcpt->scp_service;
2429 struct ptlrpc_reply_state *rs;
2430 struct group_info *ginfo = NULL;
2432 int counter = 0, rc = 0;
2435 thread->t_pid = current_pid();
2436 unshare_fs_struct();
2438 /* NB: we will call cfs_cpt_bind() for all threads, because we
2439 * might want to run lustre server only on a subset of system CPUs,
2440 * in that case ->scp_cpt is CFS_CPT_ANY */
2441 rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
2443 CWARN("%s: failed to bind %s on CPT %d\n",
2444 svc->srv_name, thread->t_name, svcpt->scp_cpt);
2447 ginfo = groups_alloc(0);
2453 set_current_groups(ginfo);
2454 put_group_info(ginfo);
2456 if (svc->srv_ops.so_thr_init != NULL) {
2457 rc = svc->srv_ops.so_thr_init(thread);
2468 rc = lu_context_init(&env->le_ctx,
2469 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2473 thread->t_env = env;
2474 env->le_ctx.lc_thread = thread;
2475 env->le_ctx.lc_cookie = 0x6;
2477 while (!list_empty(&svcpt->scp_rqbd_idle)) {
2478 rc = ptlrpc_server_post_idle_rqbds(svcpt);
2482 CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
2483 svc->srv_name, svcpt->scp_cpt, rc);
2487 /* Alloc reply state structure for this one */
2488 OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2494 spin_lock(&svcpt->scp_lock);
2496 LASSERT(thread_is_starting(thread));
2497 thread_clear_flags(thread, SVC_STARTING);
2499 LASSERT(svcpt->scp_nthrs_starting == 1);
2500 svcpt->scp_nthrs_starting--;
2502 /* SVC_STOPPING may already be set here if someone else is trying
2503 * to stop the service while this new thread has been dynamically
2504 * forked. We still set SVC_RUNNING to let our creator know that
2505 * we are now running, however we will exit as soon as possible */
2506 thread_add_flags(thread, SVC_RUNNING);
2507 svcpt->scp_nthrs_running++;
2508 spin_unlock(&svcpt->scp_lock);
2510 /* wake up our creator in case he's still waiting. */
2511 wake_up(&thread->t_ctl_waitq);
2513 thread->t_watchdog = lc_watchdog_add(ptlrpc_server_get_timeout(svcpt),
2516 spin_lock(&svcpt->scp_rep_lock);
2517 list_add(&rs->rs_list, &svcpt->scp_rep_idle);
2518 wake_up(&svcpt->scp_rep_waitq);
2519 spin_unlock(&svcpt->scp_rep_lock);
2521 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2522 svcpt->scp_nthrs_running);
2524 /* XXX maintain a list of all managed devices: insert here */
2525 while (!ptlrpc_thread_stopping(thread)) {
2526 if (ptlrpc_wait_event(svcpt, thread))
2529 ptlrpc_check_rqbd_pool(svcpt);
2531 if (ptlrpc_threads_need_create(svcpt)) {
2532 /* Ignore return code - we tried... */
2533 ptlrpc_start_thread(svcpt, 0);
2536 /* reset le_ses to initial state */
2538 /* Process all incoming reqs before handling any */
2539 if (ptlrpc_server_request_incoming(svcpt)) {
2540 lu_context_enter(&env->le_ctx);
2541 ptlrpc_server_handle_req_in(svcpt, thread);
2542 lu_context_exit(&env->le_ctx);
2544 /* but limit ourselves in case of flood */
2545 if (counter++ < 100)
2550 if (ptlrpc_at_check(svcpt))
2551 ptlrpc_at_check_timed(svcpt);
2553 if (ptlrpc_server_request_pending(svcpt, false)) {
2554 lu_context_enter(&env->le_ctx);
2555 ptlrpc_server_handle_request(svcpt, thread);
2556 lu_context_exit(&env->le_ctx);
2559 if (ptlrpc_rqbd_pending(svcpt) &&
2560 ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
2561 /* I just failed to repost request buffers.
2562 * Wait for a timeout (unless something else
2563 * happens) before I try again */
2564 svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
2565 CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
2566 svcpt->scp_nrqbds_posted);
2570 lc_watchdog_delete(thread->t_watchdog);
2571 thread->t_watchdog = NULL;
2575 * deconstruct service specific state created by ptlrpc_start_thread()
2577 if (svc->srv_ops.so_thr_done != NULL)
2578 svc->srv_ops.so_thr_done(thread);
2581 lu_context_fini(&env->le_ctx);
2585 CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n",
2586 thread, thread->t_pid, thread->t_id, rc);
2588 spin_lock(&svcpt->scp_lock);
2589 if (thread_test_and_clear_flags(thread, SVC_STARTING))
2590 svcpt->scp_nthrs_starting--;
2592 if (thread_test_and_clear_flags(thread, SVC_RUNNING)) {
2593 /* must know immediately */
2594 svcpt->scp_nthrs_running--;
2598 thread_add_flags(thread, SVC_STOPPED);
2600 wake_up(&thread->t_ctl_waitq);
2601 spin_unlock(&svcpt->scp_lock);
2606 static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
2607 struct list_head *replies)
2611 spin_lock(&hrt->hrt_lock);
2613 list_splice_init(&hrt->hrt_queue, replies);
2614 result = ptlrpc_hr.hr_stopping || !list_empty(replies);
2616 spin_unlock(&hrt->hrt_lock);
2621 * Main body of "handle reply" function.
2622 * It processes acked reply states
2624 static int ptlrpc_hr_main(void *arg)
2626 struct ptlrpc_hr_thread *hrt = (struct ptlrpc_hr_thread *)arg;
2627 struct ptlrpc_hr_partition *hrp = hrt->hrt_partition;
2628 struct list_head replies;
2631 INIT_LIST_HEAD(&replies);
2632 unshare_fs_struct();
2634 rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
2636 char threadname[20];
2638 snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
2639 hrp->hrp_cpt, hrt->hrt_id);
2640 CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
2641 threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
2644 atomic_inc(&hrp->hrp_nstarted);
2645 wake_up(&ptlrpc_hr.hr_waitq);
2647 while (!ptlrpc_hr.hr_stopping) {
2648 l_wait_condition(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
2650 while (!list_empty(&replies)) {
2651 struct ptlrpc_reply_state *rs;
2653 rs = list_entry(replies.prev,
2654 struct ptlrpc_reply_state,
2656 list_del_init(&rs->rs_list);
2657 ptlrpc_handle_rs(rs);
2661 atomic_inc(&hrp->hrp_nstopped);
2662 wake_up(&ptlrpc_hr.hr_waitq);
2667 static void ptlrpc_stop_hr_threads(void)
2669 struct ptlrpc_hr_partition *hrp;
2673 ptlrpc_hr.hr_stopping = 1;
2675 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2676 if (hrp->hrp_thrs == NULL)
2677 continue; /* uninitialized */
2678 for (j = 0; j < hrp->hrp_nthrs; j++)
2679 wake_up_all(&hrp->hrp_thrs[j].hrt_waitq);
2682 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2683 if (hrp->hrp_thrs == NULL)
2684 continue; /* uninitialized */
2685 wait_event(ptlrpc_hr.hr_waitq,
2686 atomic_read(&hrp->hrp_nstopped) ==
2687 atomic_read(&hrp->hrp_nstarted));
2691 static int ptlrpc_start_hr_threads(void)
2693 struct ptlrpc_hr_partition *hrp;
2698 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2701 for (j = 0; j < hrp->hrp_nthrs; j++) {
2702 struct ptlrpc_hr_thread *hrt = &hrp->hrp_thrs[j];
2703 struct task_struct *task;
2705 task = kthread_run(ptlrpc_hr_main,
2707 "ptlrpc_hr%02d_%03d",
2716 wait_event(ptlrpc_hr.hr_waitq,
2717 atomic_read(&hrp->hrp_nstarted) == j);
2720 CERROR("cannot start reply handler thread %d:%d: "
2721 "rc = %d\n", i, j, rc);
2722 ptlrpc_stop_hr_threads();
2730 static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
2732 struct l_wait_info lwi = { 0 };
2733 struct ptlrpc_thread *thread;
2734 struct list_head zombie;
2738 CDEBUG(D_INFO, "Stopping threads for service %s\n",
2739 svcpt->scp_service->srv_name);
2741 INIT_LIST_HEAD(&zombie);
2742 spin_lock(&svcpt->scp_lock);
2743 /* let the thread know that we would like it to stop asap */
2744 list_for_each_entry(thread, &svcpt->scp_threads, t_link) {
2745 CDEBUG(D_INFO, "Stopping thread %s #%u\n",
2746 svcpt->scp_service->srv_thread_name, thread->t_id);
2747 thread_add_flags(thread, SVC_STOPPING);
2750 wake_up_all(&svcpt->scp_waitq);
2752 while (!list_empty(&svcpt->scp_threads)) {
2753 thread = list_entry(svcpt->scp_threads.next,
2754 struct ptlrpc_thread, t_link);
2755 if (thread_is_stopped(thread)) {
2756 list_del(&thread->t_link);
2757 list_add(&thread->t_link, &zombie);
2760 spin_unlock(&svcpt->scp_lock);
2762 CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
2763 svcpt->scp_service->srv_thread_name, thread->t_id);
2764 l_wait_event(thread->t_ctl_waitq,
2765 thread_is_stopped(thread), &lwi);
2767 spin_lock(&svcpt->scp_lock);
2770 spin_unlock(&svcpt->scp_lock);
2772 while (!list_empty(&zombie)) {
2773 thread = list_entry(zombie.next,
2774 struct ptlrpc_thread, t_link);
2775 list_del(&thread->t_link);
2776 OBD_FREE_PTR(thread);
2782 * Stops all threads of a particular service \a svc
2784 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2786 struct ptlrpc_service_part *svcpt;
2790 ptlrpc_service_for_each_part(svcpt, i, svc) {
2791 if (svcpt->scp_service != NULL)
2792 ptlrpc_svcpt_stop_threads(svcpt);
2798 int ptlrpc_start_threads(struct ptlrpc_service *svc)
2805 /* We require 2 threads min, see note in ptlrpc_server_handle_request */
2806 LASSERT(svc->srv_nthrs_cpt_init >= PTLRPC_NTHRS_INIT);
2808 for (i = 0; i < svc->srv_ncpts; i++) {
2809 for (j = 0; j < svc->srv_nthrs_cpt_init; j++) {
2810 rc = ptlrpc_start_thread(svc->srv_parts[i], 1);
2816 /* We have enough threads, don't start more. b=15759 */
2823 CERROR("cannot start %s thread #%d_%d: rc %d\n",
2824 svc->srv_thread_name, i, j, rc);
2825 ptlrpc_stop_all_threads(svc);
2829 int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
2831 struct l_wait_info lwi = { 0 };
2832 struct ptlrpc_thread *thread;
2833 struct ptlrpc_service *svc;
2834 struct task_struct *task;
2838 LASSERT(svcpt != NULL);
2840 svc = svcpt->scp_service;
2842 CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n",
2843 svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running,
2844 svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit);
2847 if (unlikely(svc->srv_is_stopping))
2850 if (!ptlrpc_threads_increasable(svcpt) ||
2851 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
2852 svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1))
2855 OBD_CPT_ALLOC_PTR(thread, svc->srv_cptable, svcpt->scp_cpt);
2858 init_waitqueue_head(&thread->t_ctl_waitq);
2860 spin_lock(&svcpt->scp_lock);
2861 if (!ptlrpc_threads_increasable(svcpt)) {
2862 spin_unlock(&svcpt->scp_lock);
2863 OBD_FREE_PTR(thread);
2867 if (svcpt->scp_nthrs_starting != 0) {
2868 /* serialize starting because some modules (obdfilter)
2869 * might require unique and contiguous t_id */
2870 LASSERT(svcpt->scp_nthrs_starting == 1);
2871 spin_unlock(&svcpt->scp_lock);
2872 OBD_FREE_PTR(thread);
2874 CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n",
2875 svc->srv_thread_name, svcpt->scp_thr_nextid);
2880 CDEBUG(D_INFO, "Creating thread %s #%d race, retry later\n",
2881 svc->srv_thread_name, svcpt->scp_thr_nextid);
2885 svcpt->scp_nthrs_starting++;
2886 thread->t_id = svcpt->scp_thr_nextid++;
2887 thread_add_flags(thread, SVC_STARTING);
2888 thread->t_svcpt = svcpt;
2890 list_add(&thread->t_link, &svcpt->scp_threads);
2891 spin_unlock(&svcpt->scp_lock);
2893 if (svcpt->scp_cpt >= 0) {
2894 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s%02d_%03d",
2895 svc->srv_thread_name, svcpt->scp_cpt, thread->t_id);
2897 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s_%04d",
2898 svc->srv_thread_name, thread->t_id);
2901 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name);
2902 task = kthread_run(ptlrpc_main, thread, "%s", thread->t_name);
2905 CERROR("cannot start thread '%s': rc = %d\n",
2906 thread->t_name, rc);
2907 spin_lock(&svcpt->scp_lock);
2908 --svcpt->scp_nthrs_starting;
2909 if (thread_is_stopping(thread)) {
2910 /* this ptlrpc_thread is being hanled
2911 * by ptlrpc_svcpt_stop_threads now
2913 thread_add_flags(thread, SVC_STOPPED);
2914 wake_up(&thread->t_ctl_waitq);
2915 spin_unlock(&svcpt->scp_lock);
2917 list_del(&thread->t_link);
2918 spin_unlock(&svcpt->scp_lock);
2919 OBD_FREE_PTR(thread);
2927 l_wait_event(thread->t_ctl_waitq,
2928 thread_is_running(thread) || thread_is_stopped(thread),
2931 rc = thread_is_stopped(thread) ? thread->t_id : 0;
2935 int ptlrpc_hr_init(void)
2937 struct ptlrpc_hr_partition *hrp;
2938 struct ptlrpc_hr_thread *hrt;
2945 memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
2946 ptlrpc_hr.hr_cpt_table = cfs_cpt_table;
2948 ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
2950 if (ptlrpc_hr.hr_partitions == NULL)
2953 init_waitqueue_head(&ptlrpc_hr.hr_waitq);
2955 weight = cpumask_weight(topology_sibling_cpumask(smp_processor_id()));
2957 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2960 atomic_set(&hrp->hrp_nstarted, 0);
2961 atomic_set(&hrp->hrp_nstopped, 0);
2963 hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, i);
2965 hrp->hrp_nthrs /= weight;
2966 if (hrp->hrp_nthrs == 0)
2969 OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, i,
2970 hrp->hrp_nthrs * sizeof(*hrt));
2971 if (hrp->hrp_thrs == NULL)
2972 GOTO(out, rc = -ENOMEM);
2974 for (j = 0; j < hrp->hrp_nthrs; j++) {
2975 hrt = &hrp->hrp_thrs[j];
2978 hrt->hrt_partition = hrp;
2979 init_waitqueue_head(&hrt->hrt_waitq);
2980 spin_lock_init(&hrt->hrt_lock);
2981 INIT_LIST_HEAD(&hrt->hrt_queue);
2985 rc = ptlrpc_start_hr_threads();
2992 void ptlrpc_hr_fini(void)
2994 struct ptlrpc_hr_partition *hrp;
2997 if (ptlrpc_hr.hr_partitions == NULL)
3000 ptlrpc_stop_hr_threads();
3002 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
3003 if (hrp->hrp_thrs != NULL) {
3004 OBD_FREE(hrp->hrp_thrs,
3005 hrp->hrp_nthrs * sizeof(hrp->hrp_thrs[0]));
3009 cfs_percpt_free(ptlrpc_hr.hr_partitions);
3010 ptlrpc_hr.hr_partitions = NULL;
3015 * Wait until all already scheduled replies are processed.
3017 static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
3021 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
3024 rc = l_wait_event(svcpt->scp_waitq,
3025 atomic_read(&svcpt->scp_nreps_difficult) == 0, &lwi);
3028 CWARN("Unexpectedly long timeout %s %p\n",
3029 svcpt->scp_service->srv_name, svcpt->scp_service);
3034 ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
3036 struct ptlrpc_service_part *svcpt;
3039 /* early disarm AT timer... */
3040 ptlrpc_service_for_each_part(svcpt, i, svc) {
3041 if (svcpt->scp_service != NULL)
3042 del_timer(&svcpt->scp_at_timer);
3047 ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
3049 struct ptlrpc_service_part *svcpt;
3050 struct ptlrpc_request_buffer_desc *rqbd;
3051 struct l_wait_info lwi;
3055 /* All history will be culled when the next request buffer is
3056 * freed in ptlrpc_service_purge_all() */
3057 svc->srv_hist_nrqbds_cpt_max = 0;
3059 rc = LNetClearLazyPortal(svc->srv_req_portal);
3062 ptlrpc_service_for_each_part(svcpt, i, svc) {
3063 if (svcpt->scp_service == NULL)
3066 /* Unlink all the request buffers. This forces a 'final'
3067 * event with its 'unlink' flag set for each posted rqbd */
3068 list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
3070 rc = LNetMDUnlink(rqbd->rqbd_md_h);
3071 LASSERT(rc == 0 || rc == -ENOENT);
3075 ptlrpc_service_for_each_part(svcpt, i, svc) {
3076 if (svcpt->scp_service == NULL)
3079 /* Wait for the network to release any buffers
3080 * it's currently filling */
3081 spin_lock(&svcpt->scp_lock);
3082 while (svcpt->scp_nrqbds_posted != 0) {
3083 spin_unlock(&svcpt->scp_lock);
3084 /* Network access will complete in finite time but
3085 * the HUGE timeout lets us CWARN for visibility
3086 * of sluggish NALs */
3087 lwi = LWI_TIMEOUT_INTERVAL(
3088 cfs_time_seconds(LONG_UNLINK),
3089 cfs_time_seconds(1), NULL, NULL);
3090 rc = l_wait_event(svcpt->scp_waitq,
3091 svcpt->scp_nrqbds_posted == 0, &lwi);
3092 if (rc == -ETIMEDOUT) {
3093 CWARN("Service %s waiting for "
3094 "request buffers\n",
3095 svcpt->scp_service->srv_name);
3097 spin_lock(&svcpt->scp_lock);
3099 spin_unlock(&svcpt->scp_lock);
3104 ptlrpc_service_purge_all(struct ptlrpc_service *svc)
3106 struct ptlrpc_service_part *svcpt;
3107 struct ptlrpc_request_buffer_desc *rqbd;
3108 struct ptlrpc_request *req;
3109 struct ptlrpc_reply_state *rs;
3112 ptlrpc_service_for_each_part(svcpt, i, svc) {
3113 if (svcpt->scp_service == NULL)
3116 spin_lock(&svcpt->scp_rep_lock);
3117 while (!list_empty(&svcpt->scp_rep_active)) {
3118 rs = list_entry(svcpt->scp_rep_active.next,
3119 struct ptlrpc_reply_state, rs_list);
3120 spin_lock(&rs->rs_lock);
3121 ptlrpc_schedule_difficult_reply(rs);
3122 spin_unlock(&rs->rs_lock);
3124 spin_unlock(&svcpt->scp_rep_lock);
3126 /* purge the request queue. NB No new replies (rqbds
3127 * all unlinked) and no service threads, so I'm the only
3128 * thread noodling the request queue now */
3129 while (!list_empty(&svcpt->scp_req_incoming)) {
3130 req = list_entry(svcpt->scp_req_incoming.next,
3131 struct ptlrpc_request, rq_list);
3133 list_del(&req->rq_list);
3134 svcpt->scp_nreqs_incoming--;
3135 ptlrpc_server_finish_request(svcpt, req);
3138 while (ptlrpc_server_request_pending(svcpt, true)) {
3139 req = ptlrpc_server_request_get(svcpt, true);
3140 ptlrpc_server_finish_active_request(svcpt, req);
3143 LASSERT(list_empty(&svcpt->scp_rqbd_posted));
3144 LASSERT(svcpt->scp_nreqs_incoming == 0);
3145 LASSERT(svcpt->scp_nreqs_active == 0);
3146 /* history should have been culled by
3147 * ptlrpc_server_finish_request */
3148 LASSERT(svcpt->scp_hist_nrqbds == 0);
3150 /* Now free all the request buffers since nothing
3151 * references them any more... */
3153 while (!list_empty(&svcpt->scp_rqbd_idle)) {
3154 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
3155 struct ptlrpc_request_buffer_desc,
3157 ptlrpc_free_rqbd(rqbd);
3159 ptlrpc_wait_replies(svcpt);
3161 while (!list_empty(&svcpt->scp_rep_idle)) {
3162 rs = list_entry(svcpt->scp_rep_idle.next,
3163 struct ptlrpc_reply_state,
3165 list_del(&rs->rs_list);
3166 OBD_FREE_LARGE(rs, svc->srv_max_reply_size);
3172 ptlrpc_service_free(struct ptlrpc_service *svc)
3174 struct ptlrpc_service_part *svcpt;
3175 struct ptlrpc_at_array *array;
3178 ptlrpc_service_for_each_part(svcpt, i, svc) {
3179 if (svcpt->scp_service == NULL)
3182 /* In case somebody rearmed this in the meantime */
3183 del_timer(&svcpt->scp_at_timer);
3184 array = &svcpt->scp_at_array;
3186 if (array->paa_reqs_array != NULL) {
3187 OBD_FREE(array->paa_reqs_array,
3188 sizeof(struct list_head) * array->paa_size);
3189 array->paa_reqs_array = NULL;
3192 if (array->paa_reqs_count != NULL) {
3193 OBD_FREE(array->paa_reqs_count,
3194 sizeof(__u32) * array->paa_size);
3195 array->paa_reqs_count = NULL;
3199 ptlrpc_service_for_each_part(svcpt, i, svc)
3200 OBD_FREE_PTR(svcpt);
3202 if (svc->srv_cpts != NULL)
3203 cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts);
3205 OBD_FREE(svc, offsetof(struct ptlrpc_service,
3206 srv_parts[svc->srv_ncpts]));
3209 int ptlrpc_unregister_service(struct ptlrpc_service *service)
3213 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
3215 service->srv_is_stopping = 1;
3217 mutex_lock(&ptlrpc_all_services_mutex);
3218 list_del_init(&service->srv_list);
3219 mutex_unlock(&ptlrpc_all_services_mutex);
3221 ptlrpc_service_del_atimer(service);
3222 ptlrpc_stop_all_threads(service);
3224 ptlrpc_service_unlink_rqbd(service);
3225 ptlrpc_service_purge_all(service);
3226 ptlrpc_service_nrs_cleanup(service);
3228 ptlrpc_lprocfs_unregister_service(service);
3230 ptlrpc_service_free(service);
3234 EXPORT_SYMBOL(ptlrpc_unregister_service);
3237 * Returns 0 if the service is healthy.
3239 * Right now, it just checks to make sure that requests aren't languishing
3240 * in the queue. We'll use this health check to govern whether a node needs
3241 * to be shot, so it's intentionally non-aggressive. */
3242 static int ptlrpc_svcpt_health_check(struct ptlrpc_service_part *svcpt)
3244 struct ptlrpc_request *request = NULL;
3245 struct timeval right_now;
3248 do_gettimeofday(&right_now);
3250 spin_lock(&svcpt->scp_req_lock);
3251 /* How long has the next entry been waiting? */
3252 if (ptlrpc_server_high_pending(svcpt, true))
3253 request = ptlrpc_nrs_req_peek_nolock(svcpt, true);
3254 else if (ptlrpc_server_normal_pending(svcpt, true))
3255 request = ptlrpc_nrs_req_peek_nolock(svcpt, false);
3257 if (request == NULL) {
3258 spin_unlock(&svcpt->scp_req_lock);
3262 timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
3263 spin_unlock(&svcpt->scp_req_lock);
3265 if ((timediff / ONE_MILLION) >
3266 (AT_OFF ? obd_timeout * 3 / 2 : at_max)) {
3267 CERROR("%s: unhealthy - request has been waiting %lds\n",
3268 svcpt->scp_service->srv_name, timediff / ONE_MILLION);
3276 ptlrpc_service_health_check(struct ptlrpc_service *svc)
3278 struct ptlrpc_service_part *svcpt;
3284 ptlrpc_service_for_each_part(svcpt, i, svc) {
3285 int rc = ptlrpc_svcpt_health_check(svcpt);
3292 EXPORT_SYMBOL(ptlrpc_service_health_check);