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, 2014, 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 <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 struct list_head ptlrpc_all_services;
69 /** Used to protect the \e ptlrpc_all_services list */
70 struct mutex ptlrpc_all_services_mutex;
72 static 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 assert_spin_locked(&rs->rs_svcpt->scp_rep_lock);
392 assert_spin_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);
436 ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt)
438 struct ptlrpc_request_buffer_desc *rqbd;
443 spin_lock(&svcpt->scp_lock);
445 if (list_empty(&svcpt->scp_rqbd_idle)) {
446 spin_unlock(&svcpt->scp_lock);
450 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
451 struct ptlrpc_request_buffer_desc,
453 list_del(&rqbd->rqbd_list);
455 /* assume we will post successfully */
456 svcpt->scp_nrqbds_posted++;
457 list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_posted);
459 spin_unlock(&svcpt->scp_lock);
461 rc = ptlrpc_register_rqbd(rqbd);
468 spin_lock(&svcpt->scp_lock);
470 svcpt->scp_nrqbds_posted--;
471 list_del(&rqbd->rqbd_list);
472 list_add_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
474 /* Don't complain if no request buffers are posted right now; LNET
475 * won't drop requests because we set the portal lazy! */
477 spin_unlock(&svcpt->scp_lock);
482 static void ptlrpc_at_timer(unsigned long castmeharder)
484 struct ptlrpc_service_part *svcpt;
486 svcpt = (struct ptlrpc_service_part *)castmeharder;
488 svcpt->scp_at_check = 1;
489 svcpt->scp_at_checktime = cfs_time_current();
490 wake_up(&svcpt->scp_waitq);
494 ptlrpc_server_nthreads_check(struct ptlrpc_service *svc,
495 struct ptlrpc_service_conf *conf)
497 struct ptlrpc_service_thr_conf *tc = &conf->psc_thr;
504 * Common code for estimating & validating threads number.
505 * CPT affinity service could have percpt thread-pool instead
506 * of a global thread-pool, which means user might not always
507 * get the threads number they give it in conf::tc_nthrs_user
508 * even they did set. It's because we need to validate threads
509 * number for each CPT to guarantee each pool will have enough
510 * threads to keep the service healthy.
512 init = PTLRPC_NTHRS_INIT + (svc->srv_ops.so_hpreq_handler != NULL);
513 init = max_t(int, init, tc->tc_nthrs_init);
515 /* NB: please see comments in lustre_lnet.h for definition
516 * details of these members */
517 LASSERT(tc->tc_nthrs_max != 0);
519 if (tc->tc_nthrs_user != 0) {
520 /* In case there is a reason to test a service with many
521 * threads, we give a less strict check here, it can
522 * be up to 8 * nthrs_max */
523 total = min(tc->tc_nthrs_max * 8, tc->tc_nthrs_user);
524 nthrs = total / svc->srv_ncpts;
525 init = max(init, nthrs);
529 total = tc->tc_nthrs_max;
530 if (tc->tc_nthrs_base == 0) {
531 /* don't care about base threads number per partition,
532 * this is most for non-affinity service */
533 nthrs = total / svc->srv_ncpts;
537 nthrs = tc->tc_nthrs_base;
538 if (svc->srv_ncpts == 1) {
541 /* NB: Increase the base number if it's single partition
542 * and total number of cores/HTs is larger or equal to 4.
543 * result will always < 2 * nthrs_base */
544 weight = cfs_cpt_weight(svc->srv_cptable, CFS_CPT_ANY);
545 for (i = 1; (weight >> (i + 1)) != 0 && /* >= 4 cores/HTs */
546 (tc->tc_nthrs_base >> i) != 0; i++)
547 nthrs += tc->tc_nthrs_base >> i;
550 if (tc->tc_thr_factor != 0) {
551 int factor = tc->tc_thr_factor;
555 * User wants to increase number of threads with for
556 * each CPU core/HT, most likely the factor is larger then
557 * one thread/core because service threads are supposed to
558 * be blocked by lock or wait for IO.
561 * Amdahl's law says that adding processors wouldn't give
562 * a linear increasing of parallelism, so it's nonsense to
563 * have too many threads no matter how many cores/HTs
566 if (cfs_cpu_ht_nsiblings(0) > 1) { /* weight is # of HTs */
567 /* depress thread factor for hyper-thread */
568 factor = factor - (factor >> 1) + (factor >> 3);
571 weight = cfs_cpt_weight(svc->srv_cptable, 0);
574 for (; factor > 0 && weight > 0; factor--, weight -= fade)
575 nthrs += min(weight, fade) * factor;
578 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
579 nthrs = max(tc->tc_nthrs_base,
580 tc->tc_nthrs_max / svc->srv_ncpts);
583 nthrs = max(nthrs, tc->tc_nthrs_init);
584 svc->srv_nthrs_cpt_limit = nthrs;
585 svc->srv_nthrs_cpt_init = init;
587 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
588 CDEBUG(D_OTHER, "%s: This service may have more threads (%d) "
589 "than the given soft limit (%d)\n",
590 svc->srv_name, nthrs * svc->srv_ncpts,
596 * Initialize percpt data for a service
599 ptlrpc_service_part_init(struct ptlrpc_service *svc,
600 struct ptlrpc_service_part *svcpt, int cpt)
602 struct ptlrpc_at_array *array;
607 svcpt->scp_cpt = cpt;
608 INIT_LIST_HEAD(&svcpt->scp_threads);
610 /* rqbd and incoming request queue */
611 spin_lock_init(&svcpt->scp_lock);
612 INIT_LIST_HEAD(&svcpt->scp_rqbd_idle);
613 INIT_LIST_HEAD(&svcpt->scp_rqbd_posted);
614 INIT_LIST_HEAD(&svcpt->scp_req_incoming);
615 init_waitqueue_head(&svcpt->scp_waitq);
616 /* history request & rqbd list */
617 INIT_LIST_HEAD(&svcpt->scp_hist_reqs);
618 INIT_LIST_HEAD(&svcpt->scp_hist_rqbds);
620 /* acitve requests and hp requests */
621 spin_lock_init(&svcpt->scp_req_lock);
624 spin_lock_init(&svcpt->scp_rep_lock);
625 INIT_LIST_HEAD(&svcpt->scp_rep_active);
626 INIT_LIST_HEAD(&svcpt->scp_rep_idle);
627 init_waitqueue_head(&svcpt->scp_rep_waitq);
628 atomic_set(&svcpt->scp_nreps_difficult, 0);
630 /* adaptive timeout */
631 spin_lock_init(&svcpt->scp_at_lock);
632 array = &svcpt->scp_at_array;
634 size = at_est2timeout(at_max);
635 array->paa_size = size;
636 array->paa_count = 0;
637 array->paa_deadline = -1;
639 /* allocate memory for scp_at_array (ptlrpc_at_array) */
640 OBD_CPT_ALLOC(array->paa_reqs_array,
641 svc->srv_cptable, cpt, sizeof(struct list_head) * size);
642 if (array->paa_reqs_array == NULL)
645 for (index = 0; index < size; index++)
646 INIT_LIST_HEAD(&array->paa_reqs_array[index]);
648 OBD_CPT_ALLOC(array->paa_reqs_count,
649 svc->srv_cptable, cpt, sizeof(__u32) * size);
650 if (array->paa_reqs_count == NULL)
653 cfs_timer_init(&svcpt->scp_at_timer, ptlrpc_at_timer, 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 a 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 LPU64" 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 cfs_timer_disarm(&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 cfs_timer_arm(&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();
1268 if (CFS_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_RECONNECT)) {
1269 /* don't send early reply */
1273 /* deadline is when the client expects us to reply, margin is the
1274 difference between clients' and servers' expectations */
1275 DEBUG_REQ(D_ADAPTTO, req,
1276 "%ssending early reply (deadline %+lds, margin %+lds) for "
1277 "%d+%d", AT_OFF ? "AT off - not " : "",
1278 olddl, olddl - at_get(&svcpt->scp_at_estimate),
1279 at_get(&svcpt->scp_at_estimate), at_extra);
1285 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), "
1286 "not sending early reply. Consider increasing "
1287 "at_early_margin (%d)?", olddl, at_early_margin);
1289 /* Return an error so we're not re-added to the timed list. */
1293 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
1294 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
1295 "but no AT support");
1299 if (req->rq_export &&
1300 lustre_msg_get_flags(req->rq_reqmsg) &
1301 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1302 /* During recovery, we don't want to send too many early
1303 * replies, but on the other hand we want to make sure the
1304 * client has enough time to resend if the rpc is lost. So
1305 * during the recovery period send at least 4 early replies,
1306 * spacing them every at_extra if we can. at_estimate should
1307 * always equal this fixed value during recovery. */
1308 at_measured(&svcpt->scp_at_estimate,
1309 cfs_time_current_sec() -
1310 req->rq_arrival_time.tv_sec + min(at_extra,
1311 req->rq_export->exp_obd->obd_recovery_timeout / 4));
1313 /* We want to extend the request deadline by at_extra seconds,
1314 * so we set our service estimate to reflect how much time has
1315 * passed since this request arrived plus an additional
1316 * at_extra seconds. The client will calculate the new deadline
1317 * based on this service estimate (plus some additional time to
1318 * account for network latency). See ptlrpc_at_recv_early_reply
1320 at_measured(&svcpt->scp_at_estimate, at_extra +
1321 cfs_time_current_sec() -
1322 req->rq_arrival_time.tv_sec);
1325 /* Check to see if we've actually increased the deadline -
1326 * we may be past adaptive_max */
1327 if (req->rq_deadline >= req->rq_arrival_time.tv_sec +
1328 at_get(&svcpt->scp_at_estimate)) {
1329 DEBUG_REQ(D_WARNING, req, "Couldn't add any time "
1330 "(%ld/%ld), not sending early reply\n",
1331 olddl, req->rq_arrival_time.tv_sec +
1332 at_get(&svcpt->scp_at_estimate) -
1333 cfs_time_current_sec());
1337 reqcopy = ptlrpc_request_cache_alloc(GFP_NOFS);
1338 if (reqcopy == NULL)
1340 OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1342 GOTO(out_free, rc = -ENOMEM);
1345 reqcopy->rq_reply_state = NULL;
1346 reqcopy->rq_rep_swab_mask = 0;
1347 reqcopy->rq_pack_bulk = 0;
1348 reqcopy->rq_pack_udesc = 0;
1349 reqcopy->rq_packed_final = 0;
1350 sptlrpc_svc_ctx_addref(reqcopy);
1351 /* We only need the reqmsg for the magic */
1352 reqcopy->rq_reqmsg = reqmsg;
1353 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1356 * tgt_brw_read() and tgt_brw_write() may have decided not to reply.
1357 * Without this check, we would fail the rq_no_reply assertion in
1358 * ptlrpc_send_reply().
1360 if (reqcopy->rq_no_reply)
1361 GOTO(out, rc = -ETIMEDOUT);
1363 LASSERT(atomic_read(&req->rq_refcount));
1364 /** if it is last refcount then early reply isn't needed */
1365 if (atomic_read(&req->rq_refcount) == 1) {
1366 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
1367 "abort sending early reply\n");
1368 GOTO(out, rc = -EINVAL);
1371 /* Connection ref */
1372 reqcopy->rq_export = class_conn2export(
1373 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1374 if (reqcopy->rq_export == NULL)
1375 GOTO(out, rc = -ENODEV);
1378 class_export_rpc_inc(reqcopy->rq_export);
1379 if (reqcopy->rq_export->exp_obd &&
1380 reqcopy->rq_export->exp_obd->obd_fail)
1381 GOTO(out_put, rc = -ENODEV);
1383 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1387 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1390 /* Adjust our own deadline to what we told the client */
1391 req->rq_deadline = req->rq_arrival_time.tv_sec +
1392 at_get(&svcpt->scp_at_estimate);
1393 req->rq_early_count++; /* number sent, server side */
1395 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1398 /* Free the (early) reply state from lustre_pack_reply.
1399 (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
1400 ptlrpc_req_drop_rs(reqcopy);
1403 class_export_rpc_dec(reqcopy->rq_export);
1404 class_export_put(reqcopy->rq_export);
1406 sptlrpc_svc_ctx_decref(reqcopy);
1407 OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1409 ptlrpc_request_cache_free(reqcopy);
1413 /* Send early replies to everybody expiring within at_early_margin
1414 asking for at_extra time */
1415 static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1417 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1418 struct ptlrpc_request *rq, *n;
1419 struct list_head work_list;
1422 time_t now = cfs_time_current_sec();
1423 cfs_duration_t delay;
1424 int first, counter = 0;
1427 spin_lock(&svcpt->scp_at_lock);
1428 if (svcpt->scp_at_check == 0) {
1429 spin_unlock(&svcpt->scp_at_lock);
1432 delay = cfs_time_sub(cfs_time_current(), svcpt->scp_at_checktime);
1433 svcpt->scp_at_check = 0;
1435 if (array->paa_count == 0) {
1436 spin_unlock(&svcpt->scp_at_lock);
1440 /* The timer went off, but maybe the nearest rpc already completed. */
1441 first = array->paa_deadline - now;
1442 if (first > at_early_margin) {
1443 /* We've still got plenty of time. Reset the timer. */
1444 ptlrpc_at_set_timer(svcpt);
1445 spin_unlock(&svcpt->scp_at_lock);
1449 /* We're close to a timeout, and we don't know how much longer the
1450 server will take. Send early replies to everyone expiring soon. */
1451 INIT_LIST_HEAD(&work_list);
1453 index = (unsigned long)array->paa_deadline % array->paa_size;
1454 count = array->paa_count;
1456 count -= array->paa_reqs_count[index];
1457 list_for_each_entry_safe(rq, n,
1458 &array->paa_reqs_array[index],
1460 if (rq->rq_deadline > now + at_early_margin) {
1461 /* update the earliest deadline */
1462 if (deadline == -1 ||
1463 rq->rq_deadline < deadline)
1464 deadline = rq->rq_deadline;
1468 ptlrpc_at_remove_timed(rq);
1470 * ptlrpc_server_drop_request() may drop
1471 * refcount to 0 already. Let's check this and
1472 * don't add entry to work_list
1474 if (likely(atomic_inc_not_zero(&rq->rq_refcount)))
1475 list_add(&rq->rq_timed_list, &work_list);
1479 if (++index >= array->paa_size)
1482 array->paa_deadline = deadline;
1483 /* we have a new earliest deadline, restart the timer */
1484 ptlrpc_at_set_timer(svcpt);
1486 spin_unlock(&svcpt->scp_at_lock);
1488 CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
1489 "replies\n", first, at_extra, counter);
1491 /* We're already past request deadlines before we even get a
1492 chance to send early replies */
1493 LCONSOLE_WARN("%s: This server is not able to keep up with "
1494 "request traffic (cpu-bound).\n",
1495 svcpt->scp_service->srv_name);
1496 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, "
1497 "delay="CFS_DURATION_T"(jiff)\n",
1498 counter, svcpt->scp_nreqs_incoming,
1499 svcpt->scp_nreqs_active,
1500 at_get(&svcpt->scp_at_estimate), delay);
1503 /* we took additional refcount so entries can't be deleted from list, no
1504 * locking is needed */
1505 while (!list_empty(&work_list)) {
1506 rq = list_entry(work_list.next, struct ptlrpc_request,
1508 list_del_init(&rq->rq_timed_list);
1510 if (ptlrpc_at_send_early_reply(rq) == 0)
1511 ptlrpc_at_add_timed(rq);
1513 ptlrpc_server_drop_request(rq);
1516 RETURN(1); /* return "did_something" for liblustre */
1519 /* Check if we are already handling earlier incarnation of this request.
1520 * Called under &req->rq_export->exp_rpc_lock locked */
1521 static int ptlrpc_server_check_resend_in_progress(struct ptlrpc_request *req)
1523 struct ptlrpc_request *tmp = NULL;
1525 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ||
1526 (atomic_read(&req->rq_export->exp_rpc_count) == 0))
1529 /* bulk request are aborted upon reconnect, don't try to
1531 if (req->rq_bulk_write || req->rq_bulk_read)
1534 /* This list should not be longer than max_requests in
1535 * flights on the client, so it is not all that long.
1536 * Also we only hit this codepath in case of a resent
1537 * request which makes it even more rarely hit */
1538 list_for_each_entry(tmp, &req->rq_export->exp_reg_rpcs,
1540 /* Found duplicate one */
1541 if (tmp->rq_xid == req->rq_xid)
1544 list_for_each_entry(tmp, &req->rq_export->exp_hp_rpcs,
1546 /* Found duplicate one */
1547 if (tmp->rq_xid == req->rq_xid)
1553 DEBUG_REQ(D_HA, req, "Found duplicate req in processing");
1554 DEBUG_REQ(D_HA, tmp, "Request being processed");
1559 * Put the request to the export list if the request may become
1560 * a high priority one.
1562 static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt,
1563 struct ptlrpc_request *req)
1565 struct list_head *list;
1570 if (svcpt->scp_service->srv_ops.so_hpreq_handler) {
1571 rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
1576 if (req->rq_export) {
1578 /* Perform request specific check. We should do this
1579 * check before the request is added into exp_hp_rpcs
1580 * list otherwise it may hit swab race at LU-1044. */
1581 if (req->rq_ops->hpreq_check) {
1582 rc = req->rq_ops->hpreq_check(req);
1584 * XXX: Out of all current
1585 * ptlrpc_hpreq_ops::hpreq_check(), only
1586 * ldlm_cancel_hpreq_check() can return an
1587 * error code; other functions assert in
1588 * similar places, which seems odd.
1589 * What also does not seem right is that
1590 * handlers for those RPCs do not assert
1591 * on the same checks, but rather handle the
1592 * error cases. e.g. see ost_rw_hpreq_check(),
1593 * and ost_brw_read(), ost_brw_write().
1597 LASSERT(rc == 0 || rc == 1);
1600 list = &req->rq_export->exp_hp_rpcs;
1602 list = &req->rq_export->exp_reg_rpcs;
1605 /* do search for duplicated xid and the adding to the list
1607 spin_lock_bh(&req->rq_export->exp_rpc_lock);
1608 rc = ptlrpc_server_check_resend_in_progress(req);
1610 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1613 list_add(&req->rq_exp_list, list);
1614 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1617 ptlrpc_nrs_req_initialize(svcpt, req, !!hp);
1622 /** Remove the request from the export list. */
1623 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req)
1626 if (req->rq_export) {
1627 /* refresh lock timeout again so that client has more
1628 * room to send lock cancel RPC. */
1629 if (req->rq_ops && req->rq_ops->hpreq_fini)
1630 req->rq_ops->hpreq_fini(req);
1632 spin_lock_bh(&req->rq_export->exp_rpc_lock);
1633 list_del_init(&req->rq_exp_list);
1634 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1639 static int ptlrpc_hpreq_check(struct ptlrpc_request *req)
1644 static struct ptlrpc_hpreq_ops ptlrpc_hpreq_common = {
1645 .hpreq_check = ptlrpc_hpreq_check,
1648 /* Hi-Priority RPC check by RPC operation code. */
1649 int ptlrpc_hpreq_handler(struct ptlrpc_request *req)
1651 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1653 /* Check for export to let only reconnects for not yet evicted
1654 * export to become a HP rpc. */
1655 if ((req->rq_export != NULL) &&
1656 (opc == OBD_PING || opc == MDS_CONNECT || opc == OST_CONNECT))
1657 req->rq_ops = &ptlrpc_hpreq_common;
1661 EXPORT_SYMBOL(ptlrpc_hpreq_handler);
1663 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1664 struct ptlrpc_request *req)
1669 rc = ptlrpc_server_hpreq_init(svcpt, req);
1673 /* the current thread is not the processing thread for this request
1674 * since that, but request is in exp_hp_list and can be find there.
1675 * Remove all relations between request and old thread. */
1676 req->rq_svc_thread->t_env->le_ses = NULL;
1677 req->rq_svc_thread = NULL;
1678 req->rq_session.lc_thread = NULL;
1680 ptlrpc_nrs_req_add(svcpt, req, !!rc);
1686 * Allow to handle high priority request
1687 * User can call it w/o any lock but need to hold
1688 * ptlrpc_service_part::scp_req_lock to get reliable result
1690 static bool ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1693 int running = svcpt->scp_nthrs_running;
1695 if (!nrs_svcpt_has_hp(svcpt))
1701 if (ptlrpc_nrs_req_throttling_nolock(svcpt, true))
1704 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1705 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1706 /* leave just 1 thread for normal RPCs */
1707 running = PTLRPC_NTHRS_INIT;
1708 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1712 if (svcpt->scp_nreqs_active >= running - 1)
1715 if (svcpt->scp_nhreqs_active == 0)
1718 return !ptlrpc_nrs_req_pending_nolock(svcpt, false) ||
1719 svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1722 static bool ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1725 return ptlrpc_server_allow_high(svcpt, force) &&
1726 ptlrpc_nrs_req_pending_nolock(svcpt, true);
1730 * Only allow normal priority requests on a service that has a high-priority
1731 * queue if forced (i.e. cleanup), if there are other high priority requests
1732 * already being processed (i.e. those threads can service more high-priority
1733 * requests), or if there are enough idle threads that a later thread can do
1734 * a high priority request.
1735 * User can call it w/o any lock but need to hold
1736 * ptlrpc_service_part::scp_req_lock to get reliable result
1738 static bool ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1741 int running = svcpt->scp_nthrs_running;
1742 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1743 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1744 /* leave just 1 thread for normal RPCs */
1745 running = PTLRPC_NTHRS_INIT;
1746 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1753 if (ptlrpc_nrs_req_throttling_nolock(svcpt, false))
1756 if (svcpt->scp_nreqs_active < running - 2)
1759 if (svcpt->scp_nreqs_active >= running - 1)
1762 return svcpt->scp_nhreqs_active > 0 || !nrs_svcpt_has_hp(svcpt);
1765 static bool ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1768 return ptlrpc_server_allow_normal(svcpt, force) &&
1769 ptlrpc_nrs_req_pending_nolock(svcpt, false);
1773 * Returns true if there are requests available in incoming
1774 * request queue for processing and it is allowed to fetch them.
1775 * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1776 * to get reliable result
1777 * \see ptlrpc_server_allow_normal
1778 * \see ptlrpc_server_allow high
1781 ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt, bool force)
1783 return ptlrpc_server_high_pending(svcpt, force) ||
1784 ptlrpc_server_normal_pending(svcpt, force);
1788 * Fetch a request for processing from queue of unprocessed requests.
1789 * Favors high-priority requests.
1790 * Returns a pointer to fetched request.
1792 static struct ptlrpc_request *
1793 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force)
1795 struct ptlrpc_request *req = NULL;
1798 spin_lock(&svcpt->scp_req_lock);
1800 if (ptlrpc_server_high_pending(svcpt, force)) {
1801 req = ptlrpc_nrs_req_get_nolock(svcpt, true, force);
1803 svcpt->scp_hreq_count++;
1808 if (ptlrpc_server_normal_pending(svcpt, force)) {
1809 req = ptlrpc_nrs_req_get_nolock(svcpt, false, force);
1811 svcpt->scp_hreq_count = 0;
1816 spin_unlock(&svcpt->scp_req_lock);
1820 svcpt->scp_nreqs_active++;
1822 svcpt->scp_nhreqs_active++;
1824 spin_unlock(&svcpt->scp_req_lock);
1826 if (likely(req->rq_export))
1827 class_export_rpc_inc(req->rq_export);
1833 * Handle freshly incoming reqs, add to timed early reply list,
1834 * pass on to regular request queue.
1835 * All incoming requests pass through here before getting into
1836 * ptlrpc_server_handle_req later on.
1839 ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt,
1840 struct ptlrpc_thread *thread)
1842 struct ptlrpc_service *svc = svcpt->scp_service;
1843 struct ptlrpc_request *req;
1848 spin_lock(&svcpt->scp_lock);
1849 if (list_empty(&svcpt->scp_req_incoming)) {
1850 spin_unlock(&svcpt->scp_lock);
1854 req = list_entry(svcpt->scp_req_incoming.next,
1855 struct ptlrpc_request, rq_list);
1856 list_del_init(&req->rq_list);
1857 svcpt->scp_nreqs_incoming--;
1858 /* Consider this still a "queued" request as far as stats are
1860 spin_unlock(&svcpt->scp_lock);
1862 /* go through security check/transform */
1863 rc = sptlrpc_svc_unwrap_request(req);
1867 case SECSVC_COMPLETE:
1868 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1877 * for null-flavored rpc, msg has been unpacked by sptlrpc, although
1878 * redo it wouldn't be harmful.
1880 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
1881 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
1883 CERROR("error unpacking request: ptl %d from %s "
1884 "x"LPU64"\n", svc->srv_req_portal,
1885 libcfs_id2str(req->rq_peer), req->rq_xid);
1890 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1892 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
1893 LPU64"\n", svc->srv_req_portal,
1894 libcfs_id2str(req->rq_peer), req->rq_xid);
1898 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
1899 lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
1900 CERROR("drop incoming rpc opc %u, x"LPU64"\n",
1901 cfs_fail_val, req->rq_xid);
1906 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1907 CERROR("wrong packet type received (type=%u) from %s\n",
1908 lustre_msg_get_type(req->rq_reqmsg),
1909 libcfs_id2str(req->rq_peer));
1913 switch(lustre_msg_get_opc(req->rq_reqmsg)) {
1916 req->rq_bulk_write = 1;
1920 case MGS_CONFIG_READ:
1921 req->rq_bulk_read = 1;
1925 CDEBUG(D_RPCTRACE, "got req x"LPU64"\n", req->rq_xid);
1927 req->rq_export = class_conn2export(
1928 lustre_msg_get_handle(req->rq_reqmsg));
1929 if (req->rq_export) {
1930 rc = ptlrpc_check_req(req);
1932 rc = sptlrpc_target_export_check(req->rq_export, req);
1934 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
1935 "illegal security flavor,");
1940 ptlrpc_update_export_timer(req->rq_export, 0);
1943 /* req_in handling should/must be fast */
1944 if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5)
1945 DEBUG_REQ(D_WARNING, req, "Slow req_in handling "CFS_DURATION_T"s",
1946 cfs_time_sub(cfs_time_current_sec(),
1947 req->rq_arrival_time.tv_sec));
1949 /* Set rpc server deadline and add it to the timed list */
1950 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1951 MSGHDR_AT_SUPPORT) ?
1952 /* The max time the client expects us to take */
1953 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1954 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1955 if (unlikely(deadline == 0)) {
1956 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1960 /* Skip early reply */
1961 if (OBD_FAIL_PRECHECK(OBD_FAIL_MDS_RESEND))
1962 req->rq_deadline += obd_timeout;
1964 req->rq_svc_thread = thread;
1965 if (thread != NULL) {
1966 /* initialize request session, it is needed for request
1967 * processing by target */
1968 rc = lu_context_init(&req->rq_session, LCT_SERVER_SESSION |
1971 CERROR("%s: failure to initialize session: rc = %d\n",
1972 thread->t_name, rc);
1975 req->rq_session.lc_thread = thread;
1976 lu_context_enter(&req->rq_session);
1977 thread->t_env->le_ses = &req->rq_session;
1980 ptlrpc_at_add_timed(req);
1982 /* Move it over to the request processing queue */
1983 rc = ptlrpc_server_request_add(svcpt, req);
1987 wake_up(&svcpt->scp_waitq);
1991 ptlrpc_server_finish_request(svcpt, req);
1997 * Main incoming request handling logic.
1998 * Calls handler function from service to do actual processing.
2001 ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
2002 struct ptlrpc_thread *thread)
2004 struct ptlrpc_service *svc = svcpt->scp_service;
2005 struct ptlrpc_request *request;
2006 struct timeval work_start;
2007 struct timeval work_end;
2013 request = ptlrpc_server_request_get(svcpt, false);
2014 if (request == NULL)
2017 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
2018 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
2019 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
2020 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
2022 if (unlikely(fail_opc)) {
2023 if (request->rq_export && request->rq_ops)
2024 OBD_FAIL_TIMEOUT(fail_opc, 4);
2027 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
2029 if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
2030 libcfs_debug_dumplog();
2032 do_gettimeofday(&work_start);
2033 timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
2034 if (likely(svc->srv_stats != NULL)) {
2035 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
2037 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
2038 svcpt->scp_nreqs_incoming);
2039 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
2040 svcpt->scp_nreqs_active);
2041 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
2042 at_get(&svcpt->scp_at_estimate));
2045 if (likely(request->rq_export)) {
2046 if (unlikely(ptlrpc_check_req(request)))
2048 ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
2051 /* Discard requests queued for longer than the deadline.
2052 The deadline is increased if we send an early reply. */
2053 if (cfs_time_current_sec() > request->rq_deadline) {
2054 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s"
2055 ": deadline "CFS_DURATION_T":"CFS_DURATION_T"s ago\n",
2056 libcfs_id2str(request->rq_peer),
2057 cfs_time_sub(request->rq_deadline,
2058 request->rq_arrival_time.tv_sec),
2059 cfs_time_sub(cfs_time_current_sec(),
2060 request->rq_deadline));
2064 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
2065 "%s:%s+%d:%d:x"LPU64":%s:%d\n", current_comm(),
2066 (request->rq_export ?
2067 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2068 (request->rq_export ?
2069 atomic_read(&request->rq_export->exp_refcount) : -99),
2070 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
2071 libcfs_id2str(request->rq_peer),
2072 lustre_msg_get_opc(request->rq_reqmsg));
2074 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
2075 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
2077 CDEBUG(D_NET, "got req "LPU64"\n", request->rq_xid);
2079 /* re-assign request and sesson thread to the current one */
2080 request->rq_svc_thread = thread;
2081 if (thread != NULL) {
2082 LASSERT(request->rq_session.lc_thread == NULL);
2083 request->rq_session.lc_thread = thread;
2084 thread->t_env->le_ses = &request->rq_session;
2086 svc->srv_ops.so_req_handler(request);
2088 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
2091 if (unlikely(cfs_time_current_sec() > request->rq_deadline)) {
2092 DEBUG_REQ(D_WARNING, request, "Request took longer "
2093 "than estimated ("CFS_DURATION_T":"CFS_DURATION_T"s);"
2094 " client may timeout.",
2095 cfs_time_sub(request->rq_deadline,
2096 request->rq_arrival_time.tv_sec),
2097 cfs_time_sub(cfs_time_current_sec(),
2098 request->rq_deadline));
2101 do_gettimeofday(&work_end);
2102 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
2103 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
2104 "%s:%s+%d:%d:x"LPU64":%s:%d Request procesed in "
2105 "%ldus (%ldus total) trans "LPU64" rc %d/%d\n",
2107 (request->rq_export ?
2108 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2109 (request->rq_export ?
2110 atomic_read(&request->rq_export->exp_refcount) : -99),
2111 lustre_msg_get_status(request->rq_reqmsg),
2113 libcfs_id2str(request->rq_peer),
2114 lustre_msg_get_opc(request->rq_reqmsg),
2116 cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL),
2117 (request->rq_repmsg ?
2118 lustre_msg_get_transno(request->rq_repmsg) :
2119 request->rq_transno),
2121 (request->rq_repmsg ?
2122 lustre_msg_get_status(request->rq_repmsg) : -999));
2123 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
2124 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
2125 int opc = opcode_offset(op);
2126 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
2127 LASSERT(opc < LUSTRE_MAX_OPCODES);
2128 lprocfs_counter_add(svc->srv_stats,
2129 opc + EXTRA_MAX_OPCODES,
2133 if (unlikely(request->rq_early_count)) {
2134 DEBUG_REQ(D_ADAPTTO, request,
2135 "sent %d early replies before finishing in "
2137 request->rq_early_count,
2138 cfs_time_sub(work_end.tv_sec,
2139 request->rq_arrival_time.tv_sec));
2142 ptlrpc_server_finish_active_request(svcpt, request);
2148 * An internal function to process a single reply state object.
2151 ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
2153 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
2154 struct ptlrpc_service *svc = svcpt->scp_service;
2155 struct obd_export *exp;
2160 exp = rs->rs_export;
2162 LASSERT(rs->rs_difficult);
2163 LASSERT(rs->rs_scheduled);
2164 LASSERT(list_empty(&rs->rs_list));
2166 spin_lock(&exp->exp_lock);
2167 /* Noop if removed already */
2168 list_del_init(&rs->rs_exp_list);
2169 spin_unlock(&exp->exp_lock);
2171 /* The disk commit callback holds exp_uncommitted_replies_lock while it
2172 * iterates over newly committed replies, removing them from
2173 * exp_uncommitted_replies. It then drops this lock and schedules the
2174 * replies it found for handling here.
2176 * We can avoid contention for exp_uncommitted_replies_lock between the
2177 * HRT threads and further commit callbacks by checking rs_committed
2178 * which is set in the commit callback while it holds both
2179 * rs_lock and exp_uncommitted_reples.
2181 * If we see rs_committed clear, the commit callback _may_ not have
2182 * handled this reply yet and we race with it to grab
2183 * exp_uncommitted_replies_lock before removing the reply from
2184 * exp_uncommitted_replies. Note that if we lose the race and the
2185 * reply has already been removed, list_del_init() is a noop.
2187 * If we see rs_committed set, we know the commit callback is handling,
2188 * or has handled this reply since store reordering might allow us to
2189 * see rs_committed set out of sequence. But since this is done
2190 * holding rs_lock, we can be sure it has all completed once we hold
2191 * rs_lock, which we do right next.
2193 if (!rs->rs_committed) {
2194 spin_lock(&exp->exp_uncommitted_replies_lock);
2195 list_del_init(&rs->rs_obd_list);
2196 spin_unlock(&exp->exp_uncommitted_replies_lock);
2199 spin_lock(&rs->rs_lock);
2201 been_handled = rs->rs_handled;
2204 nlocks = rs->rs_nlocks; /* atomic "steal", but */
2205 rs->rs_nlocks = 0; /* locks still on rs_locks! */
2207 if (nlocks == 0 && !been_handled) {
2208 /* If we see this, we should already have seen the warning
2209 * in mds_steal_ack_locks() */
2210 CDEBUG(D_HA, "All locks stolen from rs %p x"LPD64".t"LPD64
2213 rs->rs_xid, rs->rs_transno, rs->rs_opc,
2214 libcfs_nid2str(exp->exp_connection->c_peer.nid));
2217 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
2218 spin_unlock(&rs->rs_lock);
2220 if (!been_handled && rs->rs_on_net) {
2221 LNetMDUnlink(rs->rs_md_h);
2222 /* Ignore return code; we're racing with completion */
2225 while (nlocks-- > 0)
2226 ldlm_lock_decref(&rs->rs_locks[nlocks],
2227 rs->rs_modes[nlocks]);
2229 spin_lock(&rs->rs_lock);
2232 rs->rs_scheduled = 0;
2234 if (!rs->rs_on_net) {
2236 spin_unlock(&rs->rs_lock);
2238 class_export_put (exp);
2239 rs->rs_export = NULL;
2240 ptlrpc_rs_decref(rs);
2241 if (atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
2242 svc->srv_is_stopping)
2243 wake_up_all(&svcpt->scp_waitq);
2247 /* still on the net; callback will schedule */
2248 spin_unlock(&rs->rs_lock);
2254 ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
2256 int avail = svcpt->scp_nrqbds_posted;
2257 int low_water = test_req_buffer_pressure ? 0 :
2258 svcpt->scp_service->srv_nbuf_per_group / 2;
2260 /* NB I'm not locking; just looking. */
2262 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
2263 * allowed the request history to grow out of control. We could put a
2264 * sanity check on that here and cull some history if we need the
2267 if (avail <= low_water)
2268 ptlrpc_grow_req_bufs(svcpt, 1);
2270 if (svcpt->scp_service->srv_stats) {
2271 lprocfs_counter_add(svcpt->scp_service->srv_stats,
2272 PTLRPC_REQBUF_AVAIL_CNTR, avail);
2277 ptlrpc_retry_rqbds(void *arg)
2279 struct ptlrpc_service_part *svcpt = (struct ptlrpc_service_part *)arg;
2281 svcpt->scp_rqbd_timeout = 0;
2286 ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
2288 return svcpt->scp_nreqs_active <
2289 svcpt->scp_nthrs_running - 1 -
2290 (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL);
2294 * allowed to create more threads
2295 * user can call it w/o any lock but need to hold
2296 * ptlrpc_service_part::scp_lock to get reliable result
2299 ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
2301 return svcpt->scp_nthrs_running +
2302 svcpt->scp_nthrs_starting <
2303 svcpt->scp_service->srv_nthrs_cpt_limit;
2307 * too many requests and allowed to create more threads
2310 ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
2312 return !ptlrpc_threads_enough(svcpt) &&
2313 ptlrpc_threads_increasable(svcpt);
2317 ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2319 return thread_is_stopping(thread) ||
2320 thread->t_svcpt->scp_service->srv_is_stopping;
2324 ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
2326 return !list_empty(&svcpt->scp_rqbd_idle) &&
2327 svcpt->scp_rqbd_timeout == 0;
2331 ptlrpc_at_check(struct ptlrpc_service_part *svcpt)
2333 return svcpt->scp_at_check;
2337 * requests wait on preprocessing
2338 * user can call it w/o any lock but need to hold
2339 * ptlrpc_service_part::scp_lock to get reliable result
2342 ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
2344 return !list_empty(&svcpt->scp_req_incoming);
2347 static __attribute__((__noinline__)) int
2348 ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
2349 struct ptlrpc_thread *thread)
2351 /* Don't exit while there are replies to be handled */
2352 struct l_wait_info lwi = LWI_TIMEOUT(svcpt->scp_rqbd_timeout,
2353 ptlrpc_retry_rqbds, svcpt);
2355 lc_watchdog_disable(thread->t_watchdog);
2359 l_wait_event_exclusive_head(svcpt->scp_waitq,
2360 ptlrpc_thread_stopping(thread) ||
2361 ptlrpc_server_request_incoming(svcpt) ||
2362 ptlrpc_server_request_pending(svcpt, false) ||
2363 ptlrpc_rqbd_pending(svcpt) ||
2364 ptlrpc_at_check(svcpt), &lwi);
2366 if (ptlrpc_thread_stopping(thread))
2369 lc_watchdog_touch(thread->t_watchdog,
2370 ptlrpc_server_get_timeout(svcpt));
2375 * Main thread body for service threads.
2376 * Waits in a loop waiting for new requests to process to appear.
2377 * Every time an incoming requests is added to its queue, a waitq
2378 * is woken up and one of the threads will handle it.
2380 static int ptlrpc_main(void *arg)
2382 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
2383 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2384 struct ptlrpc_service *svc = svcpt->scp_service;
2385 struct ptlrpc_reply_state *rs;
2386 struct group_info *ginfo = NULL;
2388 int counter = 0, rc = 0;
2391 thread->t_pid = current_pid();
2392 unshare_fs_struct();
2394 /* NB: we will call cfs_cpt_bind() for all threads, because we
2395 * might want to run lustre server only on a subset of system CPUs,
2396 * in that case ->scp_cpt is CFS_CPT_ANY */
2397 rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
2399 CWARN("%s: failed to bind %s on CPT %d\n",
2400 svc->srv_name, thread->t_name, svcpt->scp_cpt);
2403 ginfo = groups_alloc(0);
2409 set_current_groups(ginfo);
2410 put_group_info(ginfo);
2412 if (svc->srv_ops.so_thr_init != NULL) {
2413 rc = svc->srv_ops.so_thr_init(thread);
2424 rc = lu_context_init(&env->le_ctx,
2425 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2429 thread->t_env = env;
2430 env->le_ctx.lc_thread = thread;
2431 env->le_ctx.lc_cookie = 0x6;
2433 while (!list_empty(&svcpt->scp_rqbd_idle)) {
2434 rc = ptlrpc_server_post_idle_rqbds(svcpt);
2438 CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
2439 svc->srv_name, svcpt->scp_cpt, rc);
2443 /* Alloc reply state structure for this one */
2444 OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2450 spin_lock(&svcpt->scp_lock);
2452 LASSERT(thread_is_starting(thread));
2453 thread_clear_flags(thread, SVC_STARTING);
2455 LASSERT(svcpt->scp_nthrs_starting == 1);
2456 svcpt->scp_nthrs_starting--;
2458 /* SVC_STOPPING may already be set here if someone else is trying
2459 * to stop the service while this new thread has been dynamically
2460 * forked. We still set SVC_RUNNING to let our creator know that
2461 * we are now running, however we will exit as soon as possible */
2462 thread_add_flags(thread, SVC_RUNNING);
2463 svcpt->scp_nthrs_running++;
2464 spin_unlock(&svcpt->scp_lock);
2466 /* wake up our creator in case he's still waiting. */
2467 wake_up(&thread->t_ctl_waitq);
2469 thread->t_watchdog = lc_watchdog_add(ptlrpc_server_get_timeout(svcpt),
2472 spin_lock(&svcpt->scp_rep_lock);
2473 list_add(&rs->rs_list, &svcpt->scp_rep_idle);
2474 wake_up(&svcpt->scp_rep_waitq);
2475 spin_unlock(&svcpt->scp_rep_lock);
2477 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2478 svcpt->scp_nthrs_running);
2480 /* XXX maintain a list of all managed devices: insert here */
2481 while (!ptlrpc_thread_stopping(thread)) {
2482 if (ptlrpc_wait_event(svcpt, thread))
2485 ptlrpc_check_rqbd_pool(svcpt);
2487 if (ptlrpc_threads_need_create(svcpt)) {
2488 /* Ignore return code - we tried... */
2489 ptlrpc_start_thread(svcpt, 0);
2492 /* reset le_ses to initial state */
2494 /* Process all incoming reqs before handling any */
2495 if (ptlrpc_server_request_incoming(svcpt)) {
2496 lu_context_enter(&env->le_ctx);
2497 ptlrpc_server_handle_req_in(svcpt, thread);
2498 lu_context_exit(&env->le_ctx);
2500 /* but limit ourselves in case of flood */
2501 if (counter++ < 100)
2506 if (ptlrpc_at_check(svcpt))
2507 ptlrpc_at_check_timed(svcpt);
2509 if (ptlrpc_server_request_pending(svcpt, false)) {
2510 lu_context_enter(&env->le_ctx);
2511 ptlrpc_server_handle_request(svcpt, thread);
2512 lu_context_exit(&env->le_ctx);
2515 if (ptlrpc_rqbd_pending(svcpt) &&
2516 ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
2517 /* I just failed to repost request buffers.
2518 * Wait for a timeout (unless something else
2519 * happens) before I try again */
2520 svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
2521 CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
2522 svcpt->scp_nrqbds_posted);
2526 lc_watchdog_delete(thread->t_watchdog);
2527 thread->t_watchdog = NULL;
2531 * deconstruct service specific state created by ptlrpc_start_thread()
2533 if (svc->srv_ops.so_thr_done != NULL)
2534 svc->srv_ops.so_thr_done(thread);
2537 lu_context_fini(&env->le_ctx);
2541 CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n",
2542 thread, thread->t_pid, thread->t_id, rc);
2544 spin_lock(&svcpt->scp_lock);
2545 if (thread_test_and_clear_flags(thread, SVC_STARTING))
2546 svcpt->scp_nthrs_starting--;
2548 if (thread_test_and_clear_flags(thread, SVC_RUNNING)) {
2549 /* must know immediately */
2550 svcpt->scp_nthrs_running--;
2554 thread_add_flags(thread, SVC_STOPPED);
2556 wake_up(&thread->t_ctl_waitq);
2557 spin_unlock(&svcpt->scp_lock);
2562 static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
2563 struct list_head *replies)
2567 spin_lock(&hrt->hrt_lock);
2569 list_splice_init(&hrt->hrt_queue, replies);
2570 result = ptlrpc_hr.hr_stopping || !list_empty(replies);
2572 spin_unlock(&hrt->hrt_lock);
2577 * Main body of "handle reply" function.
2578 * It processes acked reply states
2580 static int ptlrpc_hr_main(void *arg)
2582 struct ptlrpc_hr_thread *hrt = (struct ptlrpc_hr_thread *)arg;
2583 struct ptlrpc_hr_partition *hrp = hrt->hrt_partition;
2584 struct list_head replies;
2585 char threadname[20];
2588 INIT_LIST_HEAD(&replies);
2589 snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
2590 hrp->hrp_cpt, hrt->hrt_id);
2591 unshare_fs_struct();
2593 rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
2595 CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
2596 threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
2599 atomic_inc(&hrp->hrp_nstarted);
2600 wake_up(&ptlrpc_hr.hr_waitq);
2602 while (!ptlrpc_hr.hr_stopping) {
2603 l_wait_condition(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
2605 while (!list_empty(&replies)) {
2606 struct ptlrpc_reply_state *rs;
2608 rs = list_entry(replies.prev,
2609 struct ptlrpc_reply_state,
2611 list_del_init(&rs->rs_list);
2612 ptlrpc_handle_rs(rs);
2616 atomic_inc(&hrp->hrp_nstopped);
2617 wake_up(&ptlrpc_hr.hr_waitq);
2622 static void ptlrpc_stop_hr_threads(void)
2624 struct ptlrpc_hr_partition *hrp;
2628 ptlrpc_hr.hr_stopping = 1;
2630 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2631 if (hrp->hrp_thrs == NULL)
2632 continue; /* uninitialized */
2633 for (j = 0; j < hrp->hrp_nthrs; j++)
2634 wake_up_all(&hrp->hrp_thrs[j].hrt_waitq);
2637 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2638 if (hrp->hrp_thrs == NULL)
2639 continue; /* uninitialized */
2640 wait_event(ptlrpc_hr.hr_waitq,
2641 atomic_read(&hrp->hrp_nstopped) ==
2642 atomic_read(&hrp->hrp_nstarted));
2646 static int ptlrpc_start_hr_threads(void)
2648 struct ptlrpc_hr_partition *hrp;
2653 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2656 for (j = 0; j < hrp->hrp_nthrs; j++) {
2657 struct ptlrpc_hr_thread *hrt = &hrp->hrp_thrs[j];
2658 struct task_struct *task;
2660 task = kthread_run(ptlrpc_hr_main,
2662 "ptlrpc_hr%02d_%03d",
2671 wait_event(ptlrpc_hr.hr_waitq,
2672 atomic_read(&hrp->hrp_nstarted) == j);
2675 CERROR("cannot start reply handler thread %d:%d: "
2676 "rc = %d\n", i, j, rc);
2677 ptlrpc_stop_hr_threads();
2685 static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
2687 struct l_wait_info lwi = { 0 };
2688 struct ptlrpc_thread *thread;
2689 struct list_head zombie;
2693 CDEBUG(D_INFO, "Stopping threads for service %s\n",
2694 svcpt->scp_service->srv_name);
2696 INIT_LIST_HEAD(&zombie);
2697 spin_lock(&svcpt->scp_lock);
2698 /* let the thread know that we would like it to stop asap */
2699 list_for_each_entry(thread, &svcpt->scp_threads, t_link) {
2700 CDEBUG(D_INFO, "Stopping thread %s #%u\n",
2701 svcpt->scp_service->srv_thread_name, thread->t_id);
2702 thread_add_flags(thread, SVC_STOPPING);
2705 wake_up_all(&svcpt->scp_waitq);
2707 while (!list_empty(&svcpt->scp_threads)) {
2708 thread = list_entry(svcpt->scp_threads.next,
2709 struct ptlrpc_thread, t_link);
2710 if (thread_is_stopped(thread)) {
2711 list_del(&thread->t_link);
2712 list_add(&thread->t_link, &zombie);
2715 spin_unlock(&svcpt->scp_lock);
2717 CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
2718 svcpt->scp_service->srv_thread_name, thread->t_id);
2719 l_wait_event(thread->t_ctl_waitq,
2720 thread_is_stopped(thread), &lwi);
2722 spin_lock(&svcpt->scp_lock);
2725 spin_unlock(&svcpt->scp_lock);
2727 while (!list_empty(&zombie)) {
2728 thread = list_entry(zombie.next,
2729 struct ptlrpc_thread, t_link);
2730 list_del(&thread->t_link);
2731 OBD_FREE_PTR(thread);
2737 * Stops all threads of a particular service \a svc
2739 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2741 struct ptlrpc_service_part *svcpt;
2745 ptlrpc_service_for_each_part(svcpt, i, svc) {
2746 if (svcpt->scp_service != NULL)
2747 ptlrpc_svcpt_stop_threads(svcpt);
2753 int ptlrpc_start_threads(struct ptlrpc_service *svc)
2760 /* We require 2 threads min, see note in ptlrpc_server_handle_request */
2761 LASSERT(svc->srv_nthrs_cpt_init >= PTLRPC_NTHRS_INIT);
2763 for (i = 0; i < svc->srv_ncpts; i++) {
2764 for (j = 0; j < svc->srv_nthrs_cpt_init; j++) {
2765 rc = ptlrpc_start_thread(svc->srv_parts[i], 1);
2771 /* We have enough threads, don't start more. b=15759 */
2778 CERROR("cannot start %s thread #%d_%d: rc %d\n",
2779 svc->srv_thread_name, i, j, rc);
2780 ptlrpc_stop_all_threads(svc);
2784 int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
2786 struct l_wait_info lwi = { 0 };
2787 struct ptlrpc_thread *thread;
2788 struct ptlrpc_service *svc;
2789 struct task_struct *task;
2793 LASSERT(svcpt != NULL);
2795 svc = svcpt->scp_service;
2797 CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n",
2798 svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running,
2799 svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit);
2802 if (unlikely(svc->srv_is_stopping))
2805 if (!ptlrpc_threads_increasable(svcpt) ||
2806 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
2807 svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1))
2810 OBD_CPT_ALLOC_PTR(thread, svc->srv_cptable, svcpt->scp_cpt);
2813 init_waitqueue_head(&thread->t_ctl_waitq);
2815 spin_lock(&svcpt->scp_lock);
2816 if (!ptlrpc_threads_increasable(svcpt)) {
2817 spin_unlock(&svcpt->scp_lock);
2818 OBD_FREE_PTR(thread);
2822 if (svcpt->scp_nthrs_starting != 0) {
2823 /* serialize starting because some modules (obdfilter)
2824 * might require unique and contiguous t_id */
2825 LASSERT(svcpt->scp_nthrs_starting == 1);
2826 spin_unlock(&svcpt->scp_lock);
2827 OBD_FREE_PTR(thread);
2829 CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n",
2830 svc->srv_thread_name, svcpt->scp_thr_nextid);
2835 CDEBUG(D_INFO, "Creating thread %s #%d race, retry later\n",
2836 svc->srv_thread_name, svcpt->scp_thr_nextid);
2840 svcpt->scp_nthrs_starting++;
2841 thread->t_id = svcpt->scp_thr_nextid++;
2842 thread_add_flags(thread, SVC_STARTING);
2843 thread->t_svcpt = svcpt;
2845 list_add(&thread->t_link, &svcpt->scp_threads);
2846 spin_unlock(&svcpt->scp_lock);
2848 if (svcpt->scp_cpt >= 0) {
2849 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s%02d_%03d",
2850 svc->srv_thread_name, svcpt->scp_cpt, thread->t_id);
2852 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s_%04d",
2853 svc->srv_thread_name, thread->t_id);
2856 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name);
2857 task = kthread_run(ptlrpc_main, thread, "%s", thread->t_name);
2860 CERROR("cannot start thread '%s': rc = %d\n",
2861 thread->t_name, rc);
2862 spin_lock(&svcpt->scp_lock);
2863 --svcpt->scp_nthrs_starting;
2864 if (thread_is_stopping(thread)) {
2865 /* this ptlrpc_thread is being hanled
2866 * by ptlrpc_svcpt_stop_threads now
2868 thread_add_flags(thread, SVC_STOPPED);
2869 wake_up(&thread->t_ctl_waitq);
2870 spin_unlock(&svcpt->scp_lock);
2872 list_del(&thread->t_link);
2873 spin_unlock(&svcpt->scp_lock);
2874 OBD_FREE_PTR(thread);
2882 l_wait_event(thread->t_ctl_waitq,
2883 thread_is_running(thread) || thread_is_stopped(thread),
2886 rc = thread_is_stopped(thread) ? thread->t_id : 0;
2890 int ptlrpc_hr_init(void)
2892 struct ptlrpc_hr_partition *hrp;
2893 struct ptlrpc_hr_thread *hrt;
2899 memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
2900 ptlrpc_hr.hr_cpt_table = cfs_cpt_table;
2902 ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
2904 if (ptlrpc_hr.hr_partitions == NULL)
2907 init_waitqueue_head(&ptlrpc_hr.hr_waitq);
2909 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2912 atomic_set(&hrp->hrp_nstarted, 0);
2913 atomic_set(&hrp->hrp_nstopped, 0);
2915 hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, i);
2916 hrp->hrp_nthrs /= cfs_cpu_ht_nsiblings(0);
2918 LASSERT(hrp->hrp_nthrs > 0);
2919 OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, i,
2920 hrp->hrp_nthrs * sizeof(*hrt));
2921 if (hrp->hrp_thrs == NULL)
2922 GOTO(out, rc = -ENOMEM);
2924 for (j = 0; j < hrp->hrp_nthrs; j++) {
2925 hrt = &hrp->hrp_thrs[j];
2928 hrt->hrt_partition = hrp;
2929 init_waitqueue_head(&hrt->hrt_waitq);
2930 spin_lock_init(&hrt->hrt_lock);
2931 INIT_LIST_HEAD(&hrt->hrt_queue);
2935 rc = ptlrpc_start_hr_threads();
2942 void ptlrpc_hr_fini(void)
2944 struct ptlrpc_hr_partition *hrp;
2947 if (ptlrpc_hr.hr_partitions == NULL)
2950 ptlrpc_stop_hr_threads();
2952 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2953 if (hrp->hrp_thrs != NULL) {
2954 OBD_FREE(hrp->hrp_thrs,
2955 hrp->hrp_nthrs * sizeof(hrp->hrp_thrs[0]));
2959 cfs_percpt_free(ptlrpc_hr.hr_partitions);
2960 ptlrpc_hr.hr_partitions = NULL;
2965 * Wait until all already scheduled replies are processed.
2967 static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
2971 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
2974 rc = l_wait_event(svcpt->scp_waitq,
2975 atomic_read(&svcpt->scp_nreps_difficult) == 0, &lwi);
2978 CWARN("Unexpectedly long timeout %s %p\n",
2979 svcpt->scp_service->srv_name, svcpt->scp_service);
2984 ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
2986 struct ptlrpc_service_part *svcpt;
2989 /* early disarm AT timer... */
2990 ptlrpc_service_for_each_part(svcpt, i, svc) {
2991 if (svcpt->scp_service != NULL)
2992 cfs_timer_disarm(&svcpt->scp_at_timer);
2997 ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
2999 struct ptlrpc_service_part *svcpt;
3000 struct ptlrpc_request_buffer_desc *rqbd;
3001 struct l_wait_info lwi;
3005 /* All history will be culled when the next request buffer is
3006 * freed in ptlrpc_service_purge_all() */
3007 svc->srv_hist_nrqbds_cpt_max = 0;
3009 rc = LNetClearLazyPortal(svc->srv_req_portal);
3012 ptlrpc_service_for_each_part(svcpt, i, svc) {
3013 if (svcpt->scp_service == NULL)
3016 /* Unlink all the request buffers. This forces a 'final'
3017 * event with its 'unlink' flag set for each posted rqbd */
3018 list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
3020 rc = LNetMDUnlink(rqbd->rqbd_md_h);
3021 LASSERT(rc == 0 || rc == -ENOENT);
3025 ptlrpc_service_for_each_part(svcpt, i, svc) {
3026 if (svcpt->scp_service == NULL)
3029 /* Wait for the network to release any buffers
3030 * it's currently filling */
3031 spin_lock(&svcpt->scp_lock);
3032 while (svcpt->scp_nrqbds_posted != 0) {
3033 spin_unlock(&svcpt->scp_lock);
3034 /* Network access will complete in finite time but
3035 * the HUGE timeout lets us CWARN for visibility
3036 * of sluggish NALs */
3037 lwi = LWI_TIMEOUT_INTERVAL(
3038 cfs_time_seconds(LONG_UNLINK),
3039 cfs_time_seconds(1), NULL, NULL);
3040 rc = l_wait_event(svcpt->scp_waitq,
3041 svcpt->scp_nrqbds_posted == 0, &lwi);
3042 if (rc == -ETIMEDOUT) {
3043 CWARN("Service %s waiting for "
3044 "request buffers\n",
3045 svcpt->scp_service->srv_name);
3047 spin_lock(&svcpt->scp_lock);
3049 spin_unlock(&svcpt->scp_lock);
3054 ptlrpc_service_purge_all(struct ptlrpc_service *svc)
3056 struct ptlrpc_service_part *svcpt;
3057 struct ptlrpc_request_buffer_desc *rqbd;
3058 struct ptlrpc_request *req;
3059 struct ptlrpc_reply_state *rs;
3062 ptlrpc_service_for_each_part(svcpt, i, svc) {
3063 if (svcpt->scp_service == NULL)
3066 spin_lock(&svcpt->scp_rep_lock);
3067 while (!list_empty(&svcpt->scp_rep_active)) {
3068 rs = list_entry(svcpt->scp_rep_active.next,
3069 struct ptlrpc_reply_state, rs_list);
3070 spin_lock(&rs->rs_lock);
3071 ptlrpc_schedule_difficult_reply(rs);
3072 spin_unlock(&rs->rs_lock);
3074 spin_unlock(&svcpt->scp_rep_lock);
3076 /* purge the request queue. NB No new replies (rqbds
3077 * all unlinked) and no service threads, so I'm the only
3078 * thread noodling the request queue now */
3079 while (!list_empty(&svcpt->scp_req_incoming)) {
3080 req = list_entry(svcpt->scp_req_incoming.next,
3081 struct ptlrpc_request, rq_list);
3083 list_del(&req->rq_list);
3084 svcpt->scp_nreqs_incoming--;
3085 ptlrpc_server_finish_request(svcpt, req);
3088 while (ptlrpc_server_request_pending(svcpt, true)) {
3089 req = ptlrpc_server_request_get(svcpt, true);
3090 ptlrpc_server_finish_active_request(svcpt, req);
3093 LASSERT(list_empty(&svcpt->scp_rqbd_posted));
3094 LASSERT(svcpt->scp_nreqs_incoming == 0);
3095 LASSERT(svcpt->scp_nreqs_active == 0);
3096 /* history should have been culled by
3097 * ptlrpc_server_finish_request */
3098 LASSERT(svcpt->scp_hist_nrqbds == 0);
3100 /* Now free all the request buffers since nothing
3101 * references them any more... */
3103 while (!list_empty(&svcpt->scp_rqbd_idle)) {
3104 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
3105 struct ptlrpc_request_buffer_desc,
3107 ptlrpc_free_rqbd(rqbd);
3109 ptlrpc_wait_replies(svcpt);
3111 while (!list_empty(&svcpt->scp_rep_idle)) {
3112 rs = list_entry(svcpt->scp_rep_idle.next,
3113 struct ptlrpc_reply_state,
3115 list_del(&rs->rs_list);
3116 OBD_FREE_LARGE(rs, svc->srv_max_reply_size);
3122 ptlrpc_service_free(struct ptlrpc_service *svc)
3124 struct ptlrpc_service_part *svcpt;
3125 struct ptlrpc_at_array *array;
3128 ptlrpc_service_for_each_part(svcpt, i, svc) {
3129 if (svcpt->scp_service == NULL)
3132 /* In case somebody rearmed this in the meantime */
3133 cfs_timer_disarm(&svcpt->scp_at_timer);
3134 array = &svcpt->scp_at_array;
3136 if (array->paa_reqs_array != NULL) {
3137 OBD_FREE(array->paa_reqs_array,
3138 sizeof(struct list_head) * array->paa_size);
3139 array->paa_reqs_array = NULL;
3142 if (array->paa_reqs_count != NULL) {
3143 OBD_FREE(array->paa_reqs_count,
3144 sizeof(__u32) * array->paa_size);
3145 array->paa_reqs_count = NULL;
3149 ptlrpc_service_for_each_part(svcpt, i, svc)
3150 OBD_FREE_PTR(svcpt);
3152 if (svc->srv_cpts != NULL)
3153 cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts);
3155 OBD_FREE(svc, offsetof(struct ptlrpc_service,
3156 srv_parts[svc->srv_ncpts]));
3159 int ptlrpc_unregister_service(struct ptlrpc_service *service)
3163 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
3165 service->srv_is_stopping = 1;
3167 mutex_lock(&ptlrpc_all_services_mutex);
3168 list_del_init(&service->srv_list);
3169 mutex_unlock(&ptlrpc_all_services_mutex);
3171 ptlrpc_service_del_atimer(service);
3172 ptlrpc_stop_all_threads(service);
3174 ptlrpc_service_unlink_rqbd(service);
3175 ptlrpc_service_purge_all(service);
3176 ptlrpc_service_nrs_cleanup(service);
3178 ptlrpc_lprocfs_unregister_service(service);
3180 ptlrpc_service_free(service);
3184 EXPORT_SYMBOL(ptlrpc_unregister_service);
3187 * Returns 0 if the service is healthy.
3189 * Right now, it just checks to make sure that requests aren't languishing
3190 * in the queue. We'll use this health check to govern whether a node needs
3191 * to be shot, so it's intentionally non-aggressive. */
3192 static int ptlrpc_svcpt_health_check(struct ptlrpc_service_part *svcpt)
3194 struct ptlrpc_request *request = NULL;
3195 struct timeval right_now;
3198 do_gettimeofday(&right_now);
3200 spin_lock(&svcpt->scp_req_lock);
3201 /* How long has the next entry been waiting? */
3202 if (ptlrpc_server_high_pending(svcpt, true))
3203 request = ptlrpc_nrs_req_peek_nolock(svcpt, true);
3204 else if (ptlrpc_server_normal_pending(svcpt, true))
3205 request = ptlrpc_nrs_req_peek_nolock(svcpt, false);
3207 if (request == NULL) {
3208 spin_unlock(&svcpt->scp_req_lock);
3212 timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
3213 spin_unlock(&svcpt->scp_req_lock);
3215 if ((timediff / ONE_MILLION) >
3216 (AT_OFF ? obd_timeout * 3 / 2 : at_max)) {
3217 CERROR("%s: unhealthy - request has been waiting %lds\n",
3218 svcpt->scp_service->srv_name, timediff / ONE_MILLION);
3226 ptlrpc_service_health_check(struct ptlrpc_service *svc)
3228 struct ptlrpc_service_part *svcpt;
3234 ptlrpc_service_for_each_part(svcpt, i, svc) {
3235 int rc = ptlrpc_svcpt_health_check(svcpt);
3242 EXPORT_SYMBOL(ptlrpc_service_health_check);