4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2010, 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 #define DEBUG_SUBSYSTEM S_RPC
38 #include <obd_support.h>
39 #include <obd_class.h>
40 #include <lustre_net.h>
41 #include <lu_object.h>
42 #include <linux/lnet/types.h>
43 #include "ptlrpc_internal.h"
45 /* The following are visible and mutable through /sys/module/ptlrpc */
46 int test_req_buffer_pressure = 0;
47 CFS_MODULE_PARM(test_req_buffer_pressure, "i", int, 0444,
48 "set non-zero to put pressure on request buffer pools");
49 CFS_MODULE_PARM(at_min, "i", int, 0644,
50 "Adaptive timeout minimum (sec)");
51 CFS_MODULE_PARM(at_max, "i", int, 0644,
52 "Adaptive timeout maximum (sec)");
53 CFS_MODULE_PARM(at_history, "i", int, 0644,
54 "Adaptive timeouts remember the slowest event that took place "
55 "within this period (sec)");
56 CFS_MODULE_PARM(at_early_margin, "i", int, 0644,
57 "How soon before an RPC deadline to send an early reply");
58 CFS_MODULE_PARM(at_extra, "i", int, 0644,
59 "How much extra time to give with each early reply");
63 static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt);
64 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req);
65 static void ptlrpc_at_remove_timed(struct ptlrpc_request *req);
67 /** Holds a list of all PTLRPC services */
68 LIST_HEAD(ptlrpc_all_services);
69 /** Used to protect the \e ptlrpc_all_services list */
70 struct mutex ptlrpc_all_services_mutex;
72 struct ptlrpc_request_buffer_desc *
73 ptlrpc_alloc_rqbd(struct ptlrpc_service_part *svcpt)
75 struct ptlrpc_service *svc = svcpt->scp_service;
76 struct ptlrpc_request_buffer_desc *rqbd;
78 OBD_CPT_ALLOC_PTR(rqbd, svc->srv_cptable, svcpt->scp_cpt);
82 rqbd->rqbd_svcpt = svcpt;
83 rqbd->rqbd_refcount = 0;
84 rqbd->rqbd_cbid.cbid_fn = request_in_callback;
85 rqbd->rqbd_cbid.cbid_arg = rqbd;
86 INIT_LIST_HEAD(&rqbd->rqbd_reqs);
87 OBD_CPT_ALLOC_LARGE(rqbd->rqbd_buffer, svc->srv_cptable,
88 svcpt->scp_cpt, svc->srv_buf_size);
89 if (rqbd->rqbd_buffer == NULL) {
94 spin_lock(&svcpt->scp_lock);
95 list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
96 svcpt->scp_nrqbds_total++;
97 spin_unlock(&svcpt->scp_lock);
103 ptlrpc_free_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
105 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
107 LASSERT(rqbd->rqbd_refcount == 0);
108 LASSERT(list_empty(&rqbd->rqbd_reqs));
110 spin_lock(&svcpt->scp_lock);
111 list_del(&rqbd->rqbd_list);
112 svcpt->scp_nrqbds_total--;
113 spin_unlock(&svcpt->scp_lock);
115 OBD_FREE_LARGE(rqbd->rqbd_buffer, svcpt->scp_service->srv_buf_size);
120 ptlrpc_grow_req_bufs(struct ptlrpc_service_part *svcpt, int post)
122 struct ptlrpc_service *svc = svcpt->scp_service;
123 struct ptlrpc_request_buffer_desc *rqbd;
127 if (svcpt->scp_rqbd_allocating)
130 spin_lock(&svcpt->scp_lock);
131 /* check again with lock */
132 if (svcpt->scp_rqbd_allocating) {
133 /* NB: we might allow more than one thread in the future */
134 LASSERT(svcpt->scp_rqbd_allocating == 1);
135 spin_unlock(&svcpt->scp_lock);
139 svcpt->scp_rqbd_allocating++;
140 spin_unlock(&svcpt->scp_lock);
143 for (i = 0; i < svc->srv_nbuf_per_group; i++) {
144 /* NB: another thread might have recycled enough rqbds, we
145 * need to make sure it wouldn't over-allocate, see LU-1212. */
146 if (svcpt->scp_nrqbds_posted >= svc->srv_nbuf_per_group)
149 rqbd = ptlrpc_alloc_rqbd(svcpt);
152 CERROR("%s: Can't allocate request buffer\n",
159 spin_lock(&svcpt->scp_lock);
161 LASSERT(svcpt->scp_rqbd_allocating == 1);
162 svcpt->scp_rqbd_allocating--;
164 spin_unlock(&svcpt->scp_lock);
167 "%s: allocate %d new %d-byte reqbufs (%d/%d left), rc = %d\n",
168 svc->srv_name, i, svc->srv_buf_size, svcpt->scp_nrqbds_posted,
169 svcpt->scp_nrqbds_total, rc);
173 rc = ptlrpc_server_post_idle_rqbds(svcpt);
179 * Part of Rep-Ack logic.
180 * Puts a lock and its mode into reply state assotiated to request reply.
183 ptlrpc_save_lock(struct ptlrpc_request *req,
184 struct lustre_handle *lock, int mode, int no_ack)
186 struct ptlrpc_reply_state *rs = req->rq_reply_state;
190 LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
192 if (req->rq_export->exp_disconnected) {
193 ldlm_lock_decref(lock, mode);
195 idx = rs->rs_nlocks++;
196 rs->rs_locks[idx] = *lock;
197 rs->rs_modes[idx] = mode;
198 rs->rs_difficult = 1;
199 rs->rs_no_ack = !!no_ack;
202 EXPORT_SYMBOL(ptlrpc_save_lock);
205 struct ptlrpc_hr_partition;
207 struct ptlrpc_hr_thread {
208 int hrt_id; /* thread ID */
210 wait_queue_head_t hrt_waitq;
211 struct list_head hrt_queue; /* RS queue */
212 struct ptlrpc_hr_partition *hrt_partition;
215 struct ptlrpc_hr_partition {
216 /* # of started threads */
217 atomic_t hrp_nstarted;
218 /* # of stopped threads */
219 atomic_t hrp_nstopped;
220 /* cpu partition id */
222 /* round-robin rotor for choosing thread */
224 /* total number of threads on this partition */
227 struct ptlrpc_hr_thread *hrp_thrs;
230 #define HRT_RUNNING 0
231 #define HRT_STOPPING 1
233 struct ptlrpc_hr_service {
234 /* CPU partition table, it's just cfs_cpt_table for now */
235 struct cfs_cpt_table *hr_cpt_table;
236 /** controller sleep waitq */
237 wait_queue_head_t hr_waitq;
238 unsigned int hr_stopping;
239 /** roundrobin rotor for non-affinity service */
240 unsigned int hr_rotor;
242 struct ptlrpc_hr_partition **hr_partitions;
246 struct list_head rsb_replies;
247 unsigned int rsb_n_replies;
248 struct ptlrpc_service_part *rsb_svcpt;
251 /** reply handling service. */
252 static struct ptlrpc_hr_service ptlrpc_hr;
255 * maximum mumber of replies scheduled in one batch
257 #define MAX_SCHEDULED 256
260 * Initialize a reply batch.
264 static void rs_batch_init(struct rs_batch *b)
266 memset(b, 0, sizeof *b);
267 INIT_LIST_HEAD(&b->rsb_replies);
271 * Choose an hr thread to dispatch requests to.
273 static struct ptlrpc_hr_thread *
274 ptlrpc_hr_select(struct ptlrpc_service_part *svcpt)
276 struct ptlrpc_hr_partition *hrp;
279 if (svcpt->scp_cpt >= 0 &&
280 svcpt->scp_service->srv_cptable == ptlrpc_hr.hr_cpt_table) {
281 /* directly match partition */
282 hrp = ptlrpc_hr.hr_partitions[svcpt->scp_cpt];
285 rotor = ptlrpc_hr.hr_rotor++;
286 rotor %= cfs_cpt_number(ptlrpc_hr.hr_cpt_table);
288 hrp = ptlrpc_hr.hr_partitions[rotor];
291 rotor = hrp->hrp_rotor++;
292 return &hrp->hrp_thrs[rotor % hrp->hrp_nthrs];
296 * Dispatch all replies accumulated in the batch to one from
297 * dedicated reply handling threads.
301 static void rs_batch_dispatch(struct rs_batch *b)
303 if (b->rsb_n_replies != 0) {
304 struct ptlrpc_hr_thread *hrt;
306 hrt = ptlrpc_hr_select(b->rsb_svcpt);
308 spin_lock(&hrt->hrt_lock);
309 list_splice_init(&b->rsb_replies, &hrt->hrt_queue);
310 spin_unlock(&hrt->hrt_lock);
312 wake_up(&hrt->hrt_waitq);
313 b->rsb_n_replies = 0;
318 * Add a reply to a batch.
319 * Add one reply object to a batch, schedule batched replies if overload.
324 static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs)
326 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
328 if (svcpt != b->rsb_svcpt || b->rsb_n_replies >= MAX_SCHEDULED) {
329 if (b->rsb_svcpt != NULL) {
330 rs_batch_dispatch(b);
331 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
333 spin_lock(&svcpt->scp_rep_lock);
334 b->rsb_svcpt = svcpt;
336 spin_lock(&rs->rs_lock);
337 rs->rs_scheduled_ever = 1;
338 if (rs->rs_scheduled == 0) {
339 list_move(&rs->rs_list, &b->rsb_replies);
340 rs->rs_scheduled = 1;
343 rs->rs_committed = 1;
344 spin_unlock(&rs->rs_lock);
348 * Reply batch finalization.
349 * Dispatch remaining replies from the batch
350 * and release remaining spinlock.
354 static void rs_batch_fini(struct rs_batch *b)
356 if (b->rsb_svcpt != NULL) {
357 rs_batch_dispatch(b);
358 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
362 #define DECLARE_RS_BATCH(b) struct rs_batch b
366 * Put reply state into a queue for processing because we received
367 * ACK from the client
369 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
371 struct ptlrpc_hr_thread *hrt;
373 LASSERT(list_empty(&rs->rs_list));
375 hrt = ptlrpc_hr_select(rs->rs_svcpt);
377 spin_lock(&hrt->hrt_lock);
378 list_add_tail(&rs->rs_list, &hrt->hrt_queue);
379 spin_unlock(&hrt->hrt_lock);
381 wake_up(&hrt->hrt_waitq);
386 ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs)
388 LASSERT(spin_is_locked(&rs->rs_svcpt->scp_rep_lock));
389 LASSERT(spin_is_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);
410 rs_batch_init(&batch);
411 /* Find any replies that have been committed and get their service
412 * to attend to complete them. */
414 /* CAVEAT EMPTOR: spinlock ordering!!! */
415 spin_lock(&exp->exp_uncommitted_replies_lock);
416 list_for_each_entry_safe(rs, nxt, &exp->exp_uncommitted_replies,
418 LASSERT (rs->rs_difficult);
419 /* VBR: per-export last_committed */
420 LASSERT(rs->rs_export);
421 if (rs->rs_transno <= exp->exp_last_committed) {
422 list_del_init(&rs->rs_obd_list);
423 rs_batch_add(&batch, rs);
426 spin_unlock(&exp->exp_uncommitted_replies_lock);
427 rs_batch_fini(&batch);
430 EXPORT_SYMBOL(ptlrpc_commit_replies);
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;
553 * User wants to increase number of threads with for
554 * each CPU core/HT, most likely the factor is larger then
555 * one thread/core because service threads are supposed to
556 * be blocked by lock or wait for IO.
559 * Amdahl's law says that adding processors wouldn't give
560 * a linear increasing of parallelism, so it's nonsense to
561 * have too many threads no matter how many cores/HTs
564 cpumask_copy(&mask, topology_thread_cpumask(0));
565 if (cpus_weight(mask) > 1) { /* weight is # of HTs */
566 /* depress thread factor for hyper-thread */
567 factor = factor - (factor >> 1) + (factor >> 3);
570 weight = cfs_cpt_weight(svc->srv_cptable, 0);
573 for (; factor > 0 && weight > 0; factor--, weight -= fade)
574 nthrs += min(weight, fade) * factor;
577 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
578 nthrs = max(tc->tc_nthrs_base,
579 tc->tc_nthrs_max / svc->srv_ncpts);
582 nthrs = max(nthrs, tc->tc_nthrs_init);
583 svc->srv_nthrs_cpt_limit = nthrs;
584 svc->srv_nthrs_cpt_init = init;
586 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
587 CDEBUG(D_OTHER, "%s: This service may have more threads (%d) "
588 "than the given soft limit (%d)\n",
589 svc->srv_name, nthrs * svc->srv_ncpts,
595 * Initialize percpt data for a service
598 ptlrpc_service_part_init(struct ptlrpc_service *svc,
599 struct ptlrpc_service_part *svcpt, int cpt)
601 struct ptlrpc_at_array *array;
606 svcpt->scp_cpt = cpt;
607 INIT_LIST_HEAD(&svcpt->scp_threads);
609 /* rqbd and incoming request queue */
610 spin_lock_init(&svcpt->scp_lock);
611 INIT_LIST_HEAD(&svcpt->scp_rqbd_idle);
612 INIT_LIST_HEAD(&svcpt->scp_rqbd_posted);
613 INIT_LIST_HEAD(&svcpt->scp_req_incoming);
614 init_waitqueue_head(&svcpt->scp_waitq);
615 /* history request & rqbd list */
616 INIT_LIST_HEAD(&svcpt->scp_hist_reqs);
617 INIT_LIST_HEAD(&svcpt->scp_hist_rqbds);
619 /* acitve requests and hp requests */
620 spin_lock_init(&svcpt->scp_req_lock);
623 spin_lock_init(&svcpt->scp_rep_lock);
624 INIT_LIST_HEAD(&svcpt->scp_rep_active);
625 INIT_LIST_HEAD(&svcpt->scp_rep_idle);
626 init_waitqueue_head(&svcpt->scp_rep_waitq);
627 atomic_set(&svcpt->scp_nreps_difficult, 0);
629 /* adaptive timeout */
630 spin_lock_init(&svcpt->scp_at_lock);
631 array = &svcpt->scp_at_array;
633 size = at_est2timeout(at_max);
634 array->paa_size = size;
635 array->paa_count = 0;
636 array->paa_deadline = -1;
638 /* allocate memory for scp_at_array (ptlrpc_at_array) */
639 OBD_CPT_ALLOC(array->paa_reqs_array,
640 svc->srv_cptable, cpt, sizeof(struct list_head) * size);
641 if (array->paa_reqs_array == NULL)
644 for (index = 0; index < size; index++)
645 INIT_LIST_HEAD(&array->paa_reqs_array[index]);
647 OBD_CPT_ALLOC(array->paa_reqs_count,
648 svc->srv_cptable, cpt, sizeof(__u32) * size);
649 if (array->paa_reqs_count == NULL)
652 cfs_timer_init(&svcpt->scp_at_timer, ptlrpc_at_timer, svcpt);
653 /* At SOW, service time should be quick; 10s seems generous. If client
654 * timeout is less than this, we'll be sending an early reply. */
655 at_init(&svcpt->scp_at_estimate, 10, 0);
657 /* assign this before call ptlrpc_grow_req_bufs */
658 svcpt->scp_service = svc;
659 /* Now allocate the request buffers, but don't post them now */
660 rc = ptlrpc_grow_req_bufs(svcpt, 0);
661 /* We shouldn't be under memory pressure at startup, so
662 * fail if we can't allocate all our buffers at this time. */
669 if (array->paa_reqs_count != NULL) {
670 OBD_FREE(array->paa_reqs_count, sizeof(__u32) * size);
671 array->paa_reqs_count = NULL;
674 if (array->paa_reqs_array != NULL) {
675 OBD_FREE(array->paa_reqs_array,
676 sizeof(struct list_head) * array->paa_size);
677 array->paa_reqs_array = NULL;
684 * Initialize service on a given portal.
685 * This includes starting serving threads , allocating and posting rqbds and
688 struct ptlrpc_service *
689 ptlrpc_register_service(struct ptlrpc_service_conf *conf,
690 proc_dir_entry_t *proc_entry)
692 struct ptlrpc_service_cpt_conf *cconf = &conf->psc_cpt;
693 struct ptlrpc_service *service;
694 struct ptlrpc_service_part *svcpt;
695 struct cfs_cpt_table *cptable;
702 LASSERT(conf->psc_buf.bc_nbufs > 0);
703 LASSERT(conf->psc_buf.bc_buf_size >=
704 conf->psc_buf.bc_req_max_size + SPTLRPC_MAX_PAYLOAD);
705 LASSERT(conf->psc_thr.tc_ctx_tags != 0);
707 cptable = cconf->cc_cptable;
709 cptable = cfs_cpt_table;
711 if (!conf->psc_thr.tc_cpu_affinity) {
714 ncpts = cfs_cpt_number(cptable);
715 if (cconf->cc_pattern != NULL) {
716 struct cfs_expr_list *el;
718 rc = cfs_expr_list_parse(cconf->cc_pattern,
719 strlen(cconf->cc_pattern),
722 CERROR("%s: invalid CPT pattern string: %s",
723 conf->psc_name, cconf->cc_pattern);
724 RETURN(ERR_PTR(-EINVAL));
727 rc = cfs_expr_list_values(el, ncpts, &cpts);
728 cfs_expr_list_free(el);
730 CERROR("%s: failed to parse CPT array %s: %d\n",
731 conf->psc_name, cconf->cc_pattern, rc);
733 OBD_FREE(cpts, sizeof(*cpts) * ncpts);
734 RETURN(ERR_PTR(rc < 0 ? rc : -EINVAL));
740 OBD_ALLOC(service, offsetof(struct ptlrpc_service, srv_parts[ncpts]));
741 if (service == NULL) {
743 OBD_FREE(cpts, sizeof(*cpts) * ncpts);
744 RETURN(ERR_PTR(-ENOMEM));
747 service->srv_cptable = cptable;
748 service->srv_cpts = cpts;
749 service->srv_ncpts = ncpts;
751 service->srv_cpt_bits = 0; /* it's zero already, easy to read... */
752 while ((1 << service->srv_cpt_bits) < cfs_cpt_number(cptable))
753 service->srv_cpt_bits++;
756 spin_lock_init(&service->srv_lock);
757 service->srv_name = conf->psc_name;
758 service->srv_watchdog_factor = conf->psc_watchdog_factor;
759 INIT_LIST_HEAD(&service->srv_list); /* for safty of cleanup */
761 /* buffer configuration */
762 service->srv_nbuf_per_group = test_req_buffer_pressure ?
763 1 : conf->psc_buf.bc_nbufs;
764 service->srv_max_req_size = conf->psc_buf.bc_req_max_size +
766 service->srv_buf_size = conf->psc_buf.bc_buf_size;
767 service->srv_rep_portal = conf->psc_buf.bc_rep_portal;
768 service->srv_req_portal = conf->psc_buf.bc_req_portal;
770 /* Increase max reply size to next power of two */
771 service->srv_max_reply_size = 1;
772 while (service->srv_max_reply_size <
773 conf->psc_buf.bc_rep_max_size + SPTLRPC_MAX_PAYLOAD)
774 service->srv_max_reply_size <<= 1;
776 service->srv_thread_name = conf->psc_thr.tc_thr_name;
777 service->srv_ctx_tags = conf->psc_thr.tc_ctx_tags;
778 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
779 service->srv_ops = conf->psc_ops;
781 for (i = 0; i < ncpts; i++) {
782 if (!conf->psc_thr.tc_cpu_affinity)
785 cpt = cpts != NULL ? cpts[i] : i;
787 OBD_CPT_ALLOC(svcpt, cptable, cpt, sizeof(*svcpt));
789 GOTO(failed, rc = -ENOMEM);
791 service->srv_parts[i] = svcpt;
792 rc = ptlrpc_service_part_init(service, svcpt, cpt);
797 ptlrpc_server_nthreads_check(service, conf);
799 rc = LNetSetLazyPortal(service->srv_req_portal);
802 mutex_lock(&ptlrpc_all_services_mutex);
803 list_add (&service->srv_list, &ptlrpc_all_services);
804 mutex_unlock(&ptlrpc_all_services_mutex);
806 if (proc_entry != NULL)
807 ptlrpc_lprocfs_register_service(proc_entry, service);
809 rc = ptlrpc_service_nrs_setup(service);
813 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
814 service->srv_name, service->srv_req_portal);
816 rc = ptlrpc_start_threads(service);
818 CERROR("Failed to start threads for service %s: %d\n",
819 service->srv_name, rc);
825 ptlrpc_unregister_service(service);
828 EXPORT_SYMBOL(ptlrpc_register_service);
831 * to actually free the request, must be called without holding svc_lock.
832 * note it's caller's responsibility to unlink req->rq_list.
834 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
836 LASSERT(atomic_read(&req->rq_refcount) == 0);
837 LASSERT(list_empty(&req->rq_timed_list));
839 /* DEBUG_REQ() assumes the reply state of a request with a valid
840 * ref will not be destroyed until that reference is dropped. */
841 ptlrpc_req_drop_rs(req);
843 sptlrpc_svc_ctx_decref(req);
845 if (req != &req->rq_rqbd->rqbd_req) {
846 /* NB request buffers use an embedded
847 * req if the incoming req unlinked the
848 * MD; this isn't one of them! */
849 OBD_FREE(req, sizeof(*req));
854 * drop a reference count of the request. if it reaches 0, we either
855 * put it into history list, or free it immediately.
857 void ptlrpc_server_drop_request(struct ptlrpc_request *req)
859 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
860 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
861 struct ptlrpc_service *svc = svcpt->scp_service;
863 struct list_head *tmp;
864 struct list_head *nxt;
866 if (!atomic_dec_and_test(&req->rq_refcount))
869 if (req->rq_at_linked) {
870 spin_lock(&svcpt->scp_at_lock);
871 /* recheck with lock, in case it's unlinked by
872 * ptlrpc_at_check_timed() */
873 if (likely(req->rq_at_linked))
874 ptlrpc_at_remove_timed(req);
875 spin_unlock(&svcpt->scp_at_lock);
878 LASSERT(list_empty(&req->rq_timed_list));
880 /* finalize request */
881 if (req->rq_export) {
882 class_export_put(req->rq_export);
883 req->rq_export = NULL;
886 spin_lock(&svcpt->scp_lock);
888 list_add(&req->rq_list, &rqbd->rqbd_reqs);
890 refcount = --(rqbd->rqbd_refcount);
892 /* request buffer is now idle: add to history */
893 list_del(&rqbd->rqbd_list);
895 list_add_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds);
896 svcpt->scp_hist_nrqbds++;
898 /* cull some history?
899 * I expect only about 1 or 2 rqbds need to be recycled here */
900 while (svcpt->scp_hist_nrqbds > svc->srv_hist_nrqbds_cpt_max) {
901 rqbd = list_entry(svcpt->scp_hist_rqbds.next,
902 struct ptlrpc_request_buffer_desc,
905 list_del(&rqbd->rqbd_list);
906 svcpt->scp_hist_nrqbds--;
908 /* remove rqbd's reqs from svc's req history while
909 * I've got the service lock */
910 list_for_each(tmp, &rqbd->rqbd_reqs) {
911 req = list_entry(tmp, struct ptlrpc_request,
913 /* Track the highest culled req seq */
914 if (req->rq_history_seq >
915 svcpt->scp_hist_seq_culled) {
916 svcpt->scp_hist_seq_culled =
919 list_del(&req->rq_history_list);
922 spin_unlock(&svcpt->scp_lock);
924 list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
925 req = list_entry(rqbd->rqbd_reqs.next,
926 struct ptlrpc_request,
928 list_del(&req->rq_list);
929 ptlrpc_server_free_request(req);
932 spin_lock(&svcpt->scp_lock);
934 * now all reqs including the embedded req has been
935 * disposed, schedule request buffer for re-use.
937 LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) ==
939 list_add_tail(&rqbd->rqbd_list,
940 &svcpt->scp_rqbd_idle);
943 spin_unlock(&svcpt->scp_lock);
944 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
945 /* If we are low on memory, we are not interested in history */
946 list_del(&req->rq_list);
947 list_del_init(&req->rq_history_list);
949 /* Track the highest culled req seq */
950 if (req->rq_history_seq > svcpt->scp_hist_seq_culled)
951 svcpt->scp_hist_seq_culled = req->rq_history_seq;
953 spin_unlock(&svcpt->scp_lock);
955 ptlrpc_server_free_request(req);
957 spin_unlock(&svcpt->scp_lock);
961 /** Change request export and move hp request from old export to new */
962 void ptlrpc_request_change_export(struct ptlrpc_request *req,
963 struct obd_export *export)
965 if (req->rq_export != NULL) {
966 if (!list_empty(&req->rq_exp_list)) {
967 /* remove rq_exp_list from last export */
968 spin_lock_bh(&req->rq_export->exp_rpc_lock);
969 list_del_init(&req->rq_exp_list);
970 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
972 /* export has one reference already, so it`s safe to
973 * add req to export queue here and get another
974 * reference for request later */
975 spin_lock_bh(&export->exp_rpc_lock);
976 list_add(&req->rq_exp_list, &export->exp_hp_rpcs);
977 spin_unlock_bh(&export->exp_rpc_lock);
979 class_export_rpc_dec(req->rq_export);
980 class_export_put(req->rq_export);
983 /* request takes one export refcount */
984 req->rq_export = class_export_get(export);
985 class_export_rpc_inc(export);
991 * to finish a request: stop sending more early replies, and release
994 static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt,
995 struct ptlrpc_request *req)
997 ptlrpc_server_hpreq_fini(req);
999 ptlrpc_server_drop_request(req);
1003 * to finish a active request: stop sending more early replies, and release
1004 * the request. should be called after we finished handling the request.
1006 static void ptlrpc_server_finish_active_request(
1007 struct ptlrpc_service_part *svcpt,
1008 struct ptlrpc_request *req)
1010 spin_lock(&svcpt->scp_req_lock);
1011 ptlrpc_nrs_req_stop_nolock(req);
1012 svcpt->scp_nreqs_active--;
1014 svcpt->scp_nhreqs_active--;
1015 spin_unlock(&svcpt->scp_req_lock);
1017 ptlrpc_nrs_req_finalize(req);
1019 if (req->rq_export != NULL)
1020 class_export_rpc_dec(req->rq_export);
1022 ptlrpc_server_finish_request(svcpt, req);
1026 * This function makes sure dead exports are evicted in a timely manner.
1027 * This function is only called when some export receives a message (i.e.,
1028 * the network is up.)
1030 static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
1032 struct obd_export *oldest_exp;
1033 time_t oldest_time, new_time;
1037 /* Compensate for slow machines, etc, by faking our request time
1038 into the future. Although this can break the strict time-ordering
1039 of the list, we can be really lazy here - we don't have to evict
1040 at the exact right moment. Eventually, all silent exports
1041 will make it to the top of the list. */
1043 /* Do not pay attention on 1sec or smaller renewals. */
1044 new_time = cfs_time_current_sec() + extra_delay;
1045 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
1048 exp->exp_last_request_time = new_time;
1049 CDEBUG(D_HA, "updating export %s at "CFS_TIME_T" exp %p\n",
1050 exp->exp_client_uuid.uuid,
1051 exp->exp_last_request_time, exp);
1053 /* exports may get disconnected from the chain even though the
1054 export has references, so we must keep the spin lock while
1055 manipulating the lists */
1056 spin_lock(&exp->exp_obd->obd_dev_lock);
1058 if (list_empty(&exp->exp_obd_chain_timed)) {
1059 /* this one is not timed */
1060 spin_unlock(&exp->exp_obd->obd_dev_lock);
1064 list_move_tail(&exp->exp_obd_chain_timed,
1065 &exp->exp_obd->obd_exports_timed);
1067 oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
1068 struct obd_export, exp_obd_chain_timed);
1069 oldest_time = oldest_exp->exp_last_request_time;
1070 spin_unlock(&exp->exp_obd->obd_dev_lock);
1072 if (exp->exp_obd->obd_recovering) {
1073 /* be nice to everyone during recovery */
1078 /* Note - racing to start/reset the obd_eviction timer is safe */
1079 if (exp->exp_obd->obd_eviction_timer == 0) {
1080 /* Check if the oldest entry is expired. */
1081 if (cfs_time_current_sec() > (oldest_time + PING_EVICT_TIMEOUT +
1083 /* We need a second timer, in case the net was down and
1084 * it just came back. Since the pinger may skip every
1085 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
1086 * we better wait for 3. */
1087 exp->exp_obd->obd_eviction_timer =
1088 cfs_time_current_sec() + 3 * PING_INTERVAL;
1089 CDEBUG(D_HA, "%s: Think about evicting %s from "CFS_TIME_T"\n",
1090 exp->exp_obd->obd_name,
1091 obd_export_nid2str(oldest_exp), oldest_time);
1094 if (cfs_time_current_sec() >
1095 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
1096 /* The evictor won't evict anyone who we've heard from
1097 * recently, so we don't have to check before we start
1099 if (!ping_evictor_wake(exp))
1100 exp->exp_obd->obd_eviction_timer = 0;
1108 * Sanity check request \a req.
1109 * Return 0 if all is ok, error code otherwise.
1111 static int ptlrpc_check_req(struct ptlrpc_request *req)
1115 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
1116 req->rq_export->exp_conn_cnt)) {
1117 DEBUG_REQ(D_RPCTRACE, req,
1118 "DROPPING req from old connection %d < %d",
1119 lustre_msg_get_conn_cnt(req->rq_reqmsg),
1120 req->rq_export->exp_conn_cnt);
1123 if (unlikely(req->rq_export->exp_obd &&
1124 req->rq_export->exp_obd->obd_fail)) {
1125 /* Failing over, don't handle any more reqs, send
1126 error response instead. */
1127 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
1128 req, req->rq_export->exp_obd->obd_name);
1130 } else if (lustre_msg_get_flags(req->rq_reqmsg) &
1131 (MSG_REPLAY | MSG_REQ_REPLAY_DONE) &&
1132 !(req->rq_export->exp_obd->obd_recovering)) {
1133 DEBUG_REQ(D_ERROR, req,
1134 "Invalid replay without recovery");
1135 class_fail_export(req->rq_export);
1137 } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 &&
1138 !(req->rq_export->exp_obd->obd_recovering)) {
1139 DEBUG_REQ(D_ERROR, req, "Invalid req with transno "
1140 LPU64" without recovery",
1141 lustre_msg_get_transno(req->rq_reqmsg));
1142 class_fail_export(req->rq_export);
1146 if (unlikely(rc < 0)) {
1147 req->rq_status = rc;
1153 static void ptlrpc_at_set_timer(struct ptlrpc_service_part *svcpt)
1155 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1158 if (array->paa_count == 0) {
1159 cfs_timer_disarm(&svcpt->scp_at_timer);
1163 /* Set timer for closest deadline */
1164 next = (__s32)(array->paa_deadline - cfs_time_current_sec() -
1167 ptlrpc_at_timer((unsigned long)svcpt);
1169 cfs_timer_arm(&svcpt->scp_at_timer, cfs_time_shift(next));
1170 CDEBUG(D_INFO, "armed %s at %+ds\n",
1171 svcpt->scp_service->srv_name, next);
1175 /* Add rpc to early reply check list */
1176 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
1178 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1179 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1180 struct ptlrpc_request *rq = NULL;
1186 if (req->rq_no_reply)
1189 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
1192 spin_lock(&svcpt->scp_at_lock);
1193 LASSERT(list_empty(&req->rq_timed_list));
1195 index = (unsigned long)req->rq_deadline % array->paa_size;
1196 if (array->paa_reqs_count[index] > 0) {
1197 /* latest rpcs will have the latest deadlines in the list,
1198 * so search backward. */
1199 list_for_each_entry_reverse(rq,
1200 &array->paa_reqs_array[index],
1202 if (req->rq_deadline >= rq->rq_deadline) {
1203 list_add(&req->rq_timed_list,
1204 &rq->rq_timed_list);
1210 /* Add the request at the head of the list */
1211 if (list_empty(&req->rq_timed_list))
1212 list_add(&req->rq_timed_list,
1213 &array->paa_reqs_array[index]);
1215 spin_lock(&req->rq_lock);
1216 req->rq_at_linked = 1;
1217 spin_unlock(&req->rq_lock);
1218 req->rq_at_index = index;
1219 array->paa_reqs_count[index]++;
1221 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
1222 array->paa_deadline = req->rq_deadline;
1223 ptlrpc_at_set_timer(svcpt);
1225 spin_unlock(&svcpt->scp_at_lock);
1231 ptlrpc_at_remove_timed(struct ptlrpc_request *req)
1233 struct ptlrpc_at_array *array;
1235 array = &req->rq_rqbd->rqbd_svcpt->scp_at_array;
1237 /* NB: must call with hold svcpt::scp_at_lock */
1238 LASSERT(!list_empty(&req->rq_timed_list));
1239 list_del_init(&req->rq_timed_list);
1241 spin_lock(&req->rq_lock);
1242 req->rq_at_linked = 0;
1243 spin_unlock(&req->rq_lock);
1245 array->paa_reqs_count[req->rq_at_index]--;
1249 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
1251 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1252 struct ptlrpc_request *reqcopy;
1253 struct lustre_msg *reqmsg;
1254 cfs_duration_t olddl = req->rq_deadline - cfs_time_current_sec();
1258 /* deadline is when the client expects us to reply, margin is the
1259 difference between clients' and servers' expectations */
1260 DEBUG_REQ(D_ADAPTTO, req,
1261 "%ssending early reply (deadline %+lds, margin %+lds) for "
1262 "%d+%d", AT_OFF ? "AT off - not " : "",
1263 olddl, olddl - at_get(&svcpt->scp_at_estimate),
1264 at_get(&svcpt->scp_at_estimate), at_extra);
1270 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), "
1271 "not sending early reply. Consider increasing "
1272 "at_early_margin (%d)?", olddl, at_early_margin);
1274 /* Return an error so we're not re-added to the timed list. */
1278 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
1279 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
1280 "but no AT support");
1284 if (req->rq_export &&
1285 lustre_msg_get_flags(req->rq_reqmsg) &
1286 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1287 /* During recovery, we don't want to send too many early
1288 * replies, but on the other hand we want to make sure the
1289 * client has enough time to resend if the rpc is lost. So
1290 * during the recovery period send at least 4 early replies,
1291 * spacing them every at_extra if we can. at_estimate should
1292 * always equal this fixed value during recovery. */
1293 at_measured(&svcpt->scp_at_estimate, min(at_extra,
1294 req->rq_export->exp_obd->obd_recovery_timeout / 4));
1296 /* Fake our processing time into the future to ask the clients
1297 * for some extra amount of time */
1298 at_measured(&svcpt->scp_at_estimate, at_extra +
1299 cfs_time_current_sec() -
1300 req->rq_arrival_time.tv_sec);
1302 /* Check to see if we've actually increased the deadline -
1303 * we may be past adaptive_max */
1304 if (req->rq_deadline >= req->rq_arrival_time.tv_sec +
1305 at_get(&svcpt->scp_at_estimate)) {
1306 DEBUG_REQ(D_WARNING, req, "Couldn't add any time "
1307 "(%ld/%ld), not sending early reply\n",
1308 olddl, req->rq_arrival_time.tv_sec +
1309 at_get(&svcpt->scp_at_estimate) -
1310 cfs_time_current_sec());
1314 newdl = cfs_time_current_sec() + at_get(&svcpt->scp_at_estimate);
1316 OBD_ALLOC(reqcopy, sizeof *reqcopy);
1317 if (reqcopy == NULL)
1319 OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1321 OBD_FREE(reqcopy, sizeof *reqcopy);
1326 reqcopy->rq_reply_state = NULL;
1327 reqcopy->rq_rep_swab_mask = 0;
1328 reqcopy->rq_pack_bulk = 0;
1329 reqcopy->rq_pack_udesc = 0;
1330 reqcopy->rq_packed_final = 0;
1331 sptlrpc_svc_ctx_addref(reqcopy);
1332 /* We only need the reqmsg for the magic */
1333 reqcopy->rq_reqmsg = reqmsg;
1334 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1336 LASSERT(atomic_read(&req->rq_refcount));
1337 /** if it is last refcount then early reply isn't needed */
1338 if (atomic_read(&req->rq_refcount) == 1) {
1339 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
1340 "abort sending early reply\n");
1341 GOTO(out, rc = -EINVAL);
1344 /* Connection ref */
1345 reqcopy->rq_export = class_conn2export(
1346 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1347 if (reqcopy->rq_export == NULL)
1348 GOTO(out, rc = -ENODEV);
1351 class_export_rpc_inc(reqcopy->rq_export);
1352 if (reqcopy->rq_export->exp_obd &&
1353 reqcopy->rq_export->exp_obd->obd_fail)
1354 GOTO(out_put, rc = -ENODEV);
1356 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1360 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1363 /* Adjust our own deadline to what we told the client */
1364 req->rq_deadline = newdl;
1365 req->rq_early_count++; /* number sent, server side */
1367 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1370 /* Free the (early) reply state from lustre_pack_reply.
1371 (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
1372 ptlrpc_req_drop_rs(reqcopy);
1375 class_export_rpc_dec(reqcopy->rq_export);
1376 class_export_put(reqcopy->rq_export);
1378 sptlrpc_svc_ctx_decref(reqcopy);
1379 OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1380 OBD_FREE(reqcopy, sizeof *reqcopy);
1384 /* Send early replies to everybody expiring within at_early_margin
1385 asking for at_extra time */
1386 static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1388 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1389 struct ptlrpc_request *rq, *n;
1390 struct list_head work_list;
1393 time_t now = cfs_time_current_sec();
1394 cfs_duration_t delay;
1395 int first, counter = 0;
1397 spin_lock(&svcpt->scp_at_lock);
1398 if (svcpt->scp_at_check == 0) {
1399 spin_unlock(&svcpt->scp_at_lock);
1402 delay = cfs_time_sub(cfs_time_current(), svcpt->scp_at_checktime);
1403 svcpt->scp_at_check = 0;
1405 if (array->paa_count == 0) {
1406 spin_unlock(&svcpt->scp_at_lock);
1410 /* The timer went off, but maybe the nearest rpc already completed. */
1411 first = array->paa_deadline - now;
1412 if (first > at_early_margin) {
1413 /* We've still got plenty of time. Reset the timer. */
1414 ptlrpc_at_set_timer(svcpt);
1415 spin_unlock(&svcpt->scp_at_lock);
1419 /* We're close to a timeout, and we don't know how much longer the
1420 server will take. Send early replies to everyone expiring soon. */
1421 INIT_LIST_HEAD(&work_list);
1423 index = (unsigned long)array->paa_deadline % array->paa_size;
1424 count = array->paa_count;
1426 count -= array->paa_reqs_count[index];
1427 list_for_each_entry_safe(rq, n,
1428 &array->paa_reqs_array[index],
1430 if (rq->rq_deadline > now + at_early_margin) {
1431 /* update the earliest deadline */
1432 if (deadline == -1 ||
1433 rq->rq_deadline < deadline)
1434 deadline = rq->rq_deadline;
1438 ptlrpc_at_remove_timed(rq);
1440 * ptlrpc_server_drop_request() may drop
1441 * refcount to 0 already. Let's check this and
1442 * don't add entry to work_list
1444 if (likely(atomic_inc_not_zero(&rq->rq_refcount)))
1445 list_add(&rq->rq_timed_list, &work_list);
1449 if (++index >= array->paa_size)
1452 array->paa_deadline = deadline;
1453 /* we have a new earliest deadline, restart the timer */
1454 ptlrpc_at_set_timer(svcpt);
1456 spin_unlock(&svcpt->scp_at_lock);
1458 CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
1459 "replies\n", first, at_extra, counter);
1461 /* We're already past request deadlines before we even get a
1462 chance to send early replies */
1463 LCONSOLE_WARN("%s: This server is not able to keep up with "
1464 "request traffic (cpu-bound).\n",
1465 svcpt->scp_service->srv_name);
1466 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, "
1467 "delay="CFS_DURATION_T"(jiff)\n",
1468 counter, svcpt->scp_nreqs_incoming,
1469 svcpt->scp_nreqs_active,
1470 at_get(&svcpt->scp_at_estimate), delay);
1473 /* we took additional refcount so entries can't be deleted from list, no
1474 * locking is needed */
1475 while (!list_empty(&work_list)) {
1476 rq = list_entry(work_list.next, struct ptlrpc_request,
1478 list_del_init(&rq->rq_timed_list);
1480 if (ptlrpc_at_send_early_reply(rq) == 0)
1481 ptlrpc_at_add_timed(rq);
1483 ptlrpc_server_drop_request(rq);
1486 RETURN(1); /* return "did_something" for liblustre */
1490 * Put the request to the export list if the request may become
1491 * a high priority one.
1493 static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt,
1494 struct ptlrpc_request *req)
1498 if (svcpt->scp_service->srv_ops.so_hpreq_handler) {
1499 rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
1504 if (req->rq_export && req->rq_ops) {
1505 /* Perform request specific check. We should do this check
1506 * before the request is added into exp_hp_rpcs list otherwise
1507 * it may hit swab race at LU-1044. */
1508 if (req->rq_ops->hpreq_check) {
1509 rc = req->rq_ops->hpreq_check(req);
1511 * XXX: Out of all current
1512 * ptlrpc_hpreq_ops::hpreq_check(), only
1513 * ldlm_cancel_hpreq_check() can return an error code;
1514 * other functions assert in similar places, which seems
1515 * odd. What also does not seem right is that handlers
1516 * for those RPCs do not assert on the same checks, but
1517 * rather handle the error cases. e.g. see
1518 * ost_rw_hpreq_check(), and ost_brw_read(),
1523 LASSERT(rc == 0 || rc == 1);
1526 spin_lock_bh(&req->rq_export->exp_rpc_lock);
1527 list_add(&req->rq_exp_list,
1528 &req->rq_export->exp_hp_rpcs);
1529 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1532 ptlrpc_nrs_req_initialize(svcpt, req, rc);
1537 /** Remove the request from the export list. */
1538 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req)
1540 if (req->rq_export && req->rq_ops) {
1541 /* refresh lock timeout again so that client has more
1542 * room to send lock cancel RPC. */
1543 if (req->rq_ops->hpreq_fini)
1544 req->rq_ops->hpreq_fini(req);
1546 spin_lock_bh(&req->rq_export->exp_rpc_lock);
1547 list_del_init(&req->rq_exp_list);
1548 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1553 static int ptlrpc_hpreq_check(struct ptlrpc_request *req)
1558 static struct ptlrpc_hpreq_ops ptlrpc_hpreq_common = {
1559 .hpreq_check = ptlrpc_hpreq_check,
1562 /* Hi-Priority RPC check by RPC operation code. */
1563 int ptlrpc_hpreq_handler(struct ptlrpc_request *req)
1565 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1567 /* Check for export to let only reconnects for not yet evicted
1568 * export to become a HP rpc. */
1569 if ((req->rq_export != NULL) &&
1570 (opc == OBD_PING || opc == MDS_CONNECT || opc == OST_CONNECT))
1571 req->rq_ops = &ptlrpc_hpreq_common;
1575 EXPORT_SYMBOL(ptlrpc_hpreq_handler);
1577 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1578 struct ptlrpc_request *req)
1582 rc = ptlrpc_server_hpreq_init(svcpt, req);
1586 ptlrpc_nrs_req_add(svcpt, req, !!rc);
1592 * Allow to handle high priority request
1593 * User can call it w/o any lock but need to hold
1594 * ptlrpc_service_part::scp_req_lock to get reliable result
1596 static bool ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1599 int running = svcpt->scp_nthrs_running;
1601 if (!nrs_svcpt_has_hp(svcpt))
1607 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1608 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1609 /* leave just 1 thread for normal RPCs */
1610 running = PTLRPC_NTHRS_INIT;
1611 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1615 if (svcpt->scp_nreqs_active >= running - 1)
1618 if (svcpt->scp_nhreqs_active == 0)
1621 return !ptlrpc_nrs_req_pending_nolock(svcpt, false) ||
1622 svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1625 static bool ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1628 return ptlrpc_server_allow_high(svcpt, force) &&
1629 ptlrpc_nrs_req_pending_nolock(svcpt, true);
1633 * Only allow normal priority requests on a service that has a high-priority
1634 * queue if forced (i.e. cleanup), if there are other high priority requests
1635 * already being processed (i.e. those threads can service more high-priority
1636 * requests), or if there are enough idle threads that a later thread can do
1637 * a high priority request.
1638 * User can call it w/o any lock but need to hold
1639 * ptlrpc_service_part::scp_req_lock to get reliable result
1641 static bool ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1644 int running = svcpt->scp_nthrs_running;
1645 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1646 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1647 /* leave just 1 thread for normal RPCs */
1648 running = PTLRPC_NTHRS_INIT;
1649 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1654 svcpt->scp_nreqs_active < running - 2)
1657 if (svcpt->scp_nreqs_active >= running - 1)
1660 return svcpt->scp_nhreqs_active > 0 || !nrs_svcpt_has_hp(svcpt);
1663 static bool ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1666 return ptlrpc_server_allow_normal(svcpt, force) &&
1667 ptlrpc_nrs_req_pending_nolock(svcpt, false);
1671 * Returns true if there are requests available in incoming
1672 * request queue for processing and it is allowed to fetch them.
1673 * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1674 * to get reliable result
1675 * \see ptlrpc_server_allow_normal
1676 * \see ptlrpc_server_allow high
1679 ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt, bool force)
1681 return ptlrpc_server_high_pending(svcpt, force) ||
1682 ptlrpc_server_normal_pending(svcpt, force);
1686 * Fetch a request for processing from queue of unprocessed requests.
1687 * Favors high-priority requests.
1688 * Returns a pointer to fetched request.
1690 static struct ptlrpc_request *
1691 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force)
1693 struct ptlrpc_request *req = NULL;
1695 spin_lock(&svcpt->scp_req_lock);
1697 if (ptlrpc_server_high_pending(svcpt, force)) {
1698 req = ptlrpc_nrs_req_get_nolock(svcpt, true, force);
1700 svcpt->scp_hreq_count++;
1705 if (ptlrpc_server_normal_pending(svcpt, force)) {
1706 req = ptlrpc_nrs_req_get_nolock(svcpt, false, force);
1708 svcpt->scp_hreq_count = 0;
1713 spin_unlock(&svcpt->scp_req_lock);
1717 svcpt->scp_nreqs_active++;
1719 svcpt->scp_nhreqs_active++;
1721 spin_unlock(&svcpt->scp_req_lock);
1723 if (likely(req->rq_export))
1724 class_export_rpc_inc(req->rq_export);
1730 * Handle freshly incoming reqs, add to timed early reply list,
1731 * pass on to regular request queue.
1732 * All incoming requests pass through here before getting into
1733 * ptlrpc_server_handle_req later on.
1736 ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt,
1737 struct ptlrpc_thread *thread)
1739 struct ptlrpc_service *svc = svcpt->scp_service;
1740 struct ptlrpc_request *req;
1744 spin_lock(&svcpt->scp_lock);
1745 if (list_empty(&svcpt->scp_req_incoming)) {
1746 spin_unlock(&svcpt->scp_lock);
1750 req = list_entry(svcpt->scp_req_incoming.next,
1751 struct ptlrpc_request, rq_list);
1752 list_del_init(&req->rq_list);
1753 svcpt->scp_nreqs_incoming--;
1754 /* Consider this still a "queued" request as far as stats are
1756 spin_unlock(&svcpt->scp_lock);
1758 /* go through security check/transform */
1759 rc = sptlrpc_svc_unwrap_request(req);
1763 case SECSVC_COMPLETE:
1764 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1773 * for null-flavored rpc, msg has been unpacked by sptlrpc, although
1774 * redo it wouldn't be harmful.
1776 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
1777 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
1779 CERROR("error unpacking request: ptl %d from %s "
1780 "x"LPU64"\n", svc->srv_req_portal,
1781 libcfs_id2str(req->rq_peer), req->rq_xid);
1786 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1788 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
1789 LPU64"\n", svc->srv_req_portal,
1790 libcfs_id2str(req->rq_peer), req->rq_xid);
1794 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
1795 lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
1796 CERROR("drop incoming rpc opc %u, x"LPU64"\n",
1797 cfs_fail_val, req->rq_xid);
1802 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1803 CERROR("wrong packet type received (type=%u) from %s\n",
1804 lustre_msg_get_type(req->rq_reqmsg),
1805 libcfs_id2str(req->rq_peer));
1809 switch(lustre_msg_get_opc(req->rq_reqmsg)) {
1812 req->rq_bulk_write = 1;
1816 case MGS_CONFIG_READ:
1817 req->rq_bulk_read = 1;
1821 CDEBUG(D_RPCTRACE, "got req x"LPU64"\n", req->rq_xid);
1823 req->rq_export = class_conn2export(
1824 lustre_msg_get_handle(req->rq_reqmsg));
1825 if (req->rq_export) {
1826 rc = ptlrpc_check_req(req);
1828 rc = sptlrpc_target_export_check(req->rq_export, req);
1830 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
1831 "illegal security flavor,");
1836 ptlrpc_update_export_timer(req->rq_export, 0);
1839 /* req_in handling should/must be fast */
1840 if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5)
1841 DEBUG_REQ(D_WARNING, req, "Slow req_in handling "CFS_DURATION_T"s",
1842 cfs_time_sub(cfs_time_current_sec(),
1843 req->rq_arrival_time.tv_sec));
1845 /* Set rpc server deadline and add it to the timed list */
1846 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1847 MSGHDR_AT_SUPPORT) ?
1848 /* The max time the client expects us to take */
1849 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1850 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1851 if (unlikely(deadline == 0)) {
1852 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1856 req->rq_svc_thread = thread;
1858 ptlrpc_at_add_timed(req);
1860 /* Move it over to the request processing queue */
1861 rc = ptlrpc_server_request_add(svcpt, req);
1865 wake_up(&svcpt->scp_waitq);
1869 ptlrpc_server_finish_request(svcpt, req);
1875 * Main incoming request handling logic.
1876 * Calls handler function from service to do actual processing.
1879 ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
1880 struct ptlrpc_thread *thread)
1882 struct ptlrpc_service *svc = svcpt->scp_service;
1883 struct ptlrpc_request *request;
1884 struct timeval work_start;
1885 struct timeval work_end;
1890 request = ptlrpc_server_request_get(svcpt, false);
1891 if (request == NULL)
1894 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
1895 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
1896 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
1897 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
1899 if (unlikely(fail_opc)) {
1900 if (request->rq_export && request->rq_ops)
1901 OBD_FAIL_TIMEOUT(fail_opc, 4);
1904 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
1906 if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
1907 libcfs_debug_dumplog();
1909 do_gettimeofday(&work_start);
1910 timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
1911 if (likely(svc->srv_stats != NULL)) {
1912 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
1914 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
1915 svcpt->scp_nreqs_incoming);
1916 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
1917 svcpt->scp_nreqs_active);
1918 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
1919 at_get(&svcpt->scp_at_estimate));
1922 rc = lu_context_init(&request->rq_session, LCT_SESSION | LCT_NOREF);
1924 CERROR("Failure to initialize session: %d\n", rc);
1927 request->rq_session.lc_thread = thread;
1928 request->rq_session.lc_cookie = 0x5;
1929 lu_context_enter(&request->rq_session);
1931 CDEBUG(D_NET, "got req "LPU64"\n", request->rq_xid);
1933 request->rq_svc_thread = thread;
1935 request->rq_svc_thread->t_env->le_ses = &request->rq_session;
1937 if (likely(request->rq_export)) {
1938 if (unlikely(ptlrpc_check_req(request)))
1940 ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
1943 /* Discard requests queued for longer than the deadline.
1944 The deadline is increased if we send an early reply. */
1945 if (cfs_time_current_sec() > request->rq_deadline) {
1946 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s"
1947 ": deadline "CFS_DURATION_T":"CFS_DURATION_T"s ago\n",
1948 libcfs_id2str(request->rq_peer),
1949 cfs_time_sub(request->rq_deadline,
1950 request->rq_arrival_time.tv_sec),
1951 cfs_time_sub(cfs_time_current_sec(),
1952 request->rq_deadline));
1956 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
1957 "%s:%s+%d:%d:x"LPU64":%s:%d\n", current_comm(),
1958 (request->rq_export ?
1959 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1960 (request->rq_export ?
1961 atomic_read(&request->rq_export->exp_refcount) : -99),
1962 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1963 libcfs_id2str(request->rq_peer),
1964 lustre_msg_get_opc(request->rq_reqmsg));
1966 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
1967 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
1969 rc = svc->srv_ops.so_req_handler(request);
1971 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
1974 lu_context_exit(&request->rq_session);
1975 lu_context_fini(&request->rq_session);
1977 if (unlikely(cfs_time_current_sec() > request->rq_deadline)) {
1978 DEBUG_REQ(D_WARNING, request, "Request took longer "
1979 "than estimated ("CFS_DURATION_T":"CFS_DURATION_T"s);"
1980 " client may timeout.",
1981 cfs_time_sub(request->rq_deadline,
1982 request->rq_arrival_time.tv_sec),
1983 cfs_time_sub(cfs_time_current_sec(),
1984 request->rq_deadline));
1987 do_gettimeofday(&work_end);
1988 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1989 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
1990 "%s:%s+%d:%d:x"LPU64":%s:%d Request procesed in "
1991 "%ldus (%ldus total) trans "LPU64" rc %d/%d\n",
1993 (request->rq_export ?
1994 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1995 (request->rq_export ?
1996 atomic_read(&request->rq_export->exp_refcount) : -99),
1997 lustre_msg_get_status(request->rq_reqmsg),
1999 libcfs_id2str(request->rq_peer),
2000 lustre_msg_get_opc(request->rq_reqmsg),
2002 cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL),
2003 (request->rq_repmsg ?
2004 lustre_msg_get_transno(request->rq_repmsg) :
2005 request->rq_transno),
2007 (request->rq_repmsg ?
2008 lustre_msg_get_status(request->rq_repmsg) : -999));
2009 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
2010 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
2011 int opc = opcode_offset(op);
2012 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
2013 LASSERT(opc < LUSTRE_MAX_OPCODES);
2014 lprocfs_counter_add(svc->srv_stats,
2015 opc + EXTRA_MAX_OPCODES,
2019 if (unlikely(request->rq_early_count)) {
2020 DEBUG_REQ(D_ADAPTTO, request,
2021 "sent %d early replies before finishing in "
2023 request->rq_early_count,
2024 cfs_time_sub(work_end.tv_sec,
2025 request->rq_arrival_time.tv_sec));
2029 ptlrpc_server_finish_active_request(svcpt, request);
2035 * An internal function to process a single reply state object.
2038 ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
2040 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
2041 struct ptlrpc_service *svc = svcpt->scp_service;
2042 struct obd_export *exp;
2046 exp = rs->rs_export;
2048 LASSERT (rs->rs_difficult);
2049 LASSERT (rs->rs_scheduled);
2050 LASSERT (list_empty(&rs->rs_list));
2052 spin_lock(&exp->exp_lock);
2053 /* Noop if removed already */
2054 list_del_init (&rs->rs_exp_list);
2055 spin_unlock(&exp->exp_lock);
2057 /* The disk commit callback holds exp_uncommitted_replies_lock while it
2058 * iterates over newly committed replies, removing them from
2059 * exp_uncommitted_replies. It then drops this lock and schedules the
2060 * replies it found for handling here.
2062 * We can avoid contention for exp_uncommitted_replies_lock between the
2063 * HRT threads and further commit callbacks by checking rs_committed
2064 * which is set in the commit callback while it holds both
2065 * rs_lock and exp_uncommitted_reples.
2067 * If we see rs_committed clear, the commit callback _may_ not have
2068 * handled this reply yet and we race with it to grab
2069 * exp_uncommitted_replies_lock before removing the reply from
2070 * exp_uncommitted_replies. Note that if we lose the race and the
2071 * reply has already been removed, list_del_init() is a noop.
2073 * If we see rs_committed set, we know the commit callback is handling,
2074 * or has handled this reply since store reordering might allow us to
2075 * see rs_committed set out of sequence. But since this is done
2076 * holding rs_lock, we can be sure it has all completed once we hold
2077 * rs_lock, which we do right next.
2079 if (!rs->rs_committed) {
2080 spin_lock(&exp->exp_uncommitted_replies_lock);
2081 list_del_init(&rs->rs_obd_list);
2082 spin_unlock(&exp->exp_uncommitted_replies_lock);
2085 spin_lock(&rs->rs_lock);
2087 been_handled = rs->rs_handled;
2090 nlocks = rs->rs_nlocks; /* atomic "steal", but */
2091 rs->rs_nlocks = 0; /* locks still on rs_locks! */
2093 if (nlocks == 0 && !been_handled) {
2094 /* If we see this, we should already have seen the warning
2095 * in mds_steal_ack_locks() */
2096 CDEBUG(D_HA, "All locks stolen from rs %p x"LPD64".t"LPD64
2099 rs->rs_xid, rs->rs_transno, rs->rs_opc,
2100 libcfs_nid2str(exp->exp_connection->c_peer.nid));
2103 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
2104 spin_unlock(&rs->rs_lock);
2106 if (!been_handled && rs->rs_on_net) {
2107 LNetMDUnlink(rs->rs_md_h);
2108 /* Ignore return code; we're racing with completion */
2111 while (nlocks-- > 0)
2112 ldlm_lock_decref(&rs->rs_locks[nlocks],
2113 rs->rs_modes[nlocks]);
2115 spin_lock(&rs->rs_lock);
2118 rs->rs_scheduled = 0;
2120 if (!rs->rs_on_net) {
2122 spin_unlock(&rs->rs_lock);
2124 class_export_put (exp);
2125 rs->rs_export = NULL;
2126 ptlrpc_rs_decref (rs);
2127 if (atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
2128 svc->srv_is_stopping)
2129 wake_up_all(&svcpt->scp_waitq);
2133 /* still on the net; callback will schedule */
2134 spin_unlock(&rs->rs_lock);
2140 ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
2142 int avail = svcpt->scp_nrqbds_posted;
2143 int low_water = test_req_buffer_pressure ? 0 :
2144 svcpt->scp_service->srv_nbuf_per_group / 2;
2146 /* NB I'm not locking; just looking. */
2148 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
2149 * allowed the request history to grow out of control. We could put a
2150 * sanity check on that here and cull some history if we need the
2153 if (avail <= low_water)
2154 ptlrpc_grow_req_bufs(svcpt, 1);
2156 if (svcpt->scp_service->srv_stats) {
2157 lprocfs_counter_add(svcpt->scp_service->srv_stats,
2158 PTLRPC_REQBUF_AVAIL_CNTR, avail);
2163 ptlrpc_retry_rqbds(void *arg)
2165 struct ptlrpc_service_part *svcpt = (struct ptlrpc_service_part *)arg;
2167 svcpt->scp_rqbd_timeout = 0;
2172 ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
2174 return svcpt->scp_nreqs_active <
2175 svcpt->scp_nthrs_running - 1 -
2176 (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL);
2180 * allowed to create more threads
2181 * user can call it w/o any lock but need to hold
2182 * ptlrpc_service_part::scp_lock to get reliable result
2185 ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
2187 return svcpt->scp_nthrs_running +
2188 svcpt->scp_nthrs_starting <
2189 svcpt->scp_service->srv_nthrs_cpt_limit;
2193 * too many requests and allowed to create more threads
2196 ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
2198 return !ptlrpc_threads_enough(svcpt) &&
2199 ptlrpc_threads_increasable(svcpt);
2203 ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2205 return thread_is_stopping(thread) ||
2206 thread->t_svcpt->scp_service->srv_is_stopping;
2210 ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
2212 return !list_empty(&svcpt->scp_rqbd_idle) &&
2213 svcpt->scp_rqbd_timeout == 0;
2217 ptlrpc_at_check(struct ptlrpc_service_part *svcpt)
2219 return svcpt->scp_at_check;
2223 * requests wait on preprocessing
2224 * user can call it w/o any lock but need to hold
2225 * ptlrpc_service_part::scp_lock to get reliable result
2228 ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
2230 return !list_empty(&svcpt->scp_req_incoming);
2233 static __attribute__((__noinline__)) int
2234 ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
2235 struct ptlrpc_thread *thread)
2237 /* Don't exit while there are replies to be handled */
2238 struct l_wait_info lwi = LWI_TIMEOUT(svcpt->scp_rqbd_timeout,
2239 ptlrpc_retry_rqbds, svcpt);
2241 /* XXX: Add this back when libcfs watchdog is merged upstream
2242 lc_watchdog_disable(thread->t_watchdog);
2247 l_wait_event_exclusive_head(svcpt->scp_waitq,
2248 ptlrpc_thread_stopping(thread) ||
2249 ptlrpc_server_request_incoming(svcpt) ||
2250 ptlrpc_server_request_pending(svcpt, false) ||
2251 ptlrpc_rqbd_pending(svcpt) ||
2252 ptlrpc_at_check(svcpt), &lwi);
2254 if (ptlrpc_thread_stopping(thread))
2258 lc_watchdog_touch(thread->t_watchdog,
2259 ptlrpc_server_get_timeout(svcpt));
2265 * Main thread body for service threads.
2266 * Waits in a loop waiting for new requests to process to appear.
2267 * Every time an incoming requests is added to its queue, a waitq
2268 * is woken up and one of the threads will handle it.
2270 static int ptlrpc_main(void *arg)
2272 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
2273 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2274 struct ptlrpc_service *svc = svcpt->scp_service;
2275 struct ptlrpc_reply_state *rs;
2276 #ifdef WITH_GROUP_INFO
2277 group_info_t *ginfo = NULL;
2280 int counter = 0, rc = 0;
2282 thread->t_pid = current_pid();
2283 unshare_fs_struct();
2285 /* NB: we will call cfs_cpt_bind() for all threads, because we
2286 * might want to run lustre server only on a subset of system CPUs,
2287 * in that case ->scp_cpt is CFS_CPT_ANY */
2288 rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
2290 CWARN("%s: failed to bind %s on CPT %d\n",
2291 svc->srv_name, thread->t_name, svcpt->scp_cpt);
2294 #ifdef WITH_GROUP_INFO
2295 ginfo = groups_alloc(0);
2301 set_current_groups(ginfo);
2302 put_group_info(ginfo);
2305 if (svc->srv_ops.so_thr_init != NULL) {
2306 rc = svc->srv_ops.so_thr_init(thread);
2317 rc = lu_context_init(&env->le_ctx,
2318 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2322 thread->t_env = env;
2323 env->le_ctx.lc_thread = thread;
2324 env->le_ctx.lc_cookie = 0x6;
2326 while (!list_empty(&svcpt->scp_rqbd_idle)) {
2327 rc = ptlrpc_server_post_idle_rqbds(svcpt);
2331 CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
2332 svc->srv_name, svcpt->scp_cpt, rc);
2336 /* Alloc reply state structure for this one */
2337 OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2343 spin_lock(&svcpt->scp_lock);
2345 LASSERT(thread_is_starting(thread));
2346 thread_clear_flags(thread, SVC_STARTING);
2348 LASSERT(svcpt->scp_nthrs_starting == 1);
2349 svcpt->scp_nthrs_starting--;
2351 /* SVC_STOPPING may already be set here if someone else is trying
2352 * to stop the service while this new thread has been dynamically
2353 * forked. We still set SVC_RUNNING to let our creator know that
2354 * we are now running, however we will exit as soon as possible */
2355 thread_add_flags(thread, SVC_RUNNING);
2356 svcpt->scp_nthrs_running++;
2357 spin_unlock(&svcpt->scp_lock);
2359 /* wake up our creator in case he's still waiting. */
2360 wake_up(&thread->t_ctl_waitq);
2363 thread->t_watchdog = lc_watchdog_add(ptlrpc_server_get_timeout(svcpt),
2367 spin_lock(&svcpt->scp_rep_lock);
2368 list_add(&rs->rs_list, &svcpt->scp_rep_idle);
2369 wake_up(&svcpt->scp_rep_waitq);
2370 spin_unlock(&svcpt->scp_rep_lock);
2372 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2373 svcpt->scp_nthrs_running);
2375 /* XXX maintain a list of all managed devices: insert here */
2376 while (!ptlrpc_thread_stopping(thread)) {
2377 if (ptlrpc_wait_event(svcpt, thread))
2380 ptlrpc_check_rqbd_pool(svcpt);
2382 if (ptlrpc_threads_need_create(svcpt)) {
2383 /* Ignore return code - we tried... */
2384 ptlrpc_start_thread(svcpt, 0);
2387 /* Process all incoming reqs before handling any */
2388 if (ptlrpc_server_request_incoming(svcpt)) {
2389 lu_context_enter(&env->le_ctx);
2391 ptlrpc_server_handle_req_in(svcpt, thread);
2392 lu_context_exit(&env->le_ctx);
2394 /* but limit ourselves in case of flood */
2395 if (counter++ < 100)
2400 if (ptlrpc_at_check(svcpt))
2401 ptlrpc_at_check_timed(svcpt);
2403 if (ptlrpc_server_request_pending(svcpt, false)) {
2404 lu_context_enter(&env->le_ctx);
2405 ptlrpc_server_handle_request(svcpt, thread);
2406 lu_context_exit(&env->le_ctx);
2409 if (ptlrpc_rqbd_pending(svcpt) &&
2410 ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
2411 /* I just failed to repost request buffers.
2412 * Wait for a timeout (unless something else
2413 * happens) before I try again */
2414 svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
2415 CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
2416 svcpt->scp_nrqbds_posted);
2421 lc_watchdog_delete(thread->t_watchdog);
2422 thread->t_watchdog = NULL;
2427 * deconstruct service specific state created by ptlrpc_start_thread()
2429 if (svc->srv_ops.so_thr_done != NULL)
2430 svc->srv_ops.so_thr_done(thread);
2433 lu_context_fini(&env->le_ctx);
2437 CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n",
2438 thread, thread->t_pid, thread->t_id, rc);
2440 spin_lock(&svcpt->scp_lock);
2441 if (thread_test_and_clear_flags(thread, SVC_STARTING))
2442 svcpt->scp_nthrs_starting--;
2444 if (thread_test_and_clear_flags(thread, SVC_RUNNING)) {
2445 /* must know immediately */
2446 svcpt->scp_nthrs_running--;
2450 thread_add_flags(thread, SVC_STOPPED);
2452 wake_up(&thread->t_ctl_waitq);
2453 spin_unlock(&svcpt->scp_lock);
2458 static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
2459 struct list_head *replies)
2463 spin_lock(&hrt->hrt_lock);
2465 list_splice_init(&hrt->hrt_queue, replies);
2466 result = ptlrpc_hr.hr_stopping || !list_empty(replies);
2468 spin_unlock(&hrt->hrt_lock);
2473 * Main body of "handle reply" function.
2474 * It processes acked reply states
2476 static int ptlrpc_hr_main(void *arg)
2478 struct ptlrpc_hr_thread *hrt = (struct ptlrpc_hr_thread *)arg;
2479 struct ptlrpc_hr_partition *hrp = hrt->hrt_partition;
2480 LIST_HEAD (replies);
2481 char threadname[20];
2484 snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
2485 hrp->hrp_cpt, hrt->hrt_id);
2486 unshare_fs_struct();
2488 rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
2490 CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
2491 threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
2494 atomic_inc(&hrp->hrp_nstarted);
2495 wake_up(&ptlrpc_hr.hr_waitq);
2497 while (!ptlrpc_hr.hr_stopping) {
2498 l_wait_condition(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
2500 while (!list_empty(&replies)) {
2501 struct ptlrpc_reply_state *rs;
2503 rs = list_entry(replies.prev,
2504 struct ptlrpc_reply_state,
2506 list_del_init(&rs->rs_list);
2507 ptlrpc_handle_rs(rs);
2511 atomic_inc(&hrp->hrp_nstopped);
2512 wake_up(&ptlrpc_hr.hr_waitq);
2517 static void ptlrpc_stop_hr_threads(void)
2519 struct ptlrpc_hr_partition *hrp;
2523 ptlrpc_hr.hr_stopping = 1;
2525 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2526 if (hrp->hrp_thrs == NULL)
2527 continue; /* uninitialized */
2528 for (j = 0; j < hrp->hrp_nthrs; j++)
2529 wake_up_all(&hrp->hrp_thrs[j].hrt_waitq);
2532 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2533 if (hrp->hrp_thrs == NULL)
2534 continue; /* uninitialized */
2535 wait_event(ptlrpc_hr.hr_waitq,
2536 atomic_read(&hrp->hrp_nstopped) ==
2537 atomic_read(&hrp->hrp_nstarted));
2541 static int ptlrpc_start_hr_threads(void)
2543 struct ptlrpc_hr_partition *hrp;
2547 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2550 for (j = 0; j < hrp->hrp_nthrs; j++) {
2551 struct ptlrpc_hr_thread *hrt = &hrp->hrp_thrs[j];
2552 rc = PTR_ERR(kthread_run(ptlrpc_hr_main,
2554 "ptlrpc_hr%02d_%03d",
2557 if (IS_ERR_VALUE(rc))
2560 wait_event(ptlrpc_hr.hr_waitq,
2561 atomic_read(&hrp->hrp_nstarted) == j);
2562 if (!IS_ERR_VALUE(rc))
2565 CERROR("Reply handling thread %d:%d Failed on starting: "
2566 "rc = %d\n", i, j, rc);
2567 ptlrpc_stop_hr_threads();
2573 static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
2575 struct l_wait_info lwi = { 0 };
2576 struct ptlrpc_thread *thread;
2579 CDEBUG(D_INFO, "Stopping threads for service %s\n",
2580 svcpt->scp_service->srv_name);
2582 spin_lock(&svcpt->scp_lock);
2583 /* let the thread know that we would like it to stop asap */
2584 list_for_each_entry(thread, &svcpt->scp_threads, t_link) {
2585 CDEBUG(D_INFO, "Stopping thread %s #%u\n",
2586 svcpt->scp_service->srv_thread_name, thread->t_id);
2587 thread_add_flags(thread, SVC_STOPPING);
2590 wake_up_all(&svcpt->scp_waitq);
2592 while (!list_empty(&svcpt->scp_threads)) {
2593 thread = list_entry(svcpt->scp_threads.next,
2594 struct ptlrpc_thread, t_link);
2595 if (thread_is_stopped(thread)) {
2596 list_del(&thread->t_link);
2597 list_add(&thread->t_link, &zombie);
2600 spin_unlock(&svcpt->scp_lock);
2602 CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
2603 svcpt->scp_service->srv_thread_name, thread->t_id);
2604 l_wait_event(thread->t_ctl_waitq,
2605 thread_is_stopped(thread), &lwi);
2607 spin_lock(&svcpt->scp_lock);
2610 spin_unlock(&svcpt->scp_lock);
2612 while (!list_empty(&zombie)) {
2613 thread = list_entry(zombie.next,
2614 struct ptlrpc_thread, t_link);
2615 list_del(&thread->t_link);
2616 OBD_FREE_PTR(thread);
2622 * Stops all threads of a particular service \a svc
2624 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2626 struct ptlrpc_service_part *svcpt;
2629 ptlrpc_service_for_each_part(svcpt, i, svc) {
2630 if (svcpt->scp_service != NULL)
2631 ptlrpc_svcpt_stop_threads(svcpt);
2636 EXPORT_SYMBOL(ptlrpc_stop_all_threads);
2638 int ptlrpc_start_threads(struct ptlrpc_service *svc)
2644 /* We require 2 threads min, see note in ptlrpc_server_handle_request */
2645 LASSERT(svc->srv_nthrs_cpt_init >= PTLRPC_NTHRS_INIT);
2647 for (i = 0; i < svc->srv_ncpts; i++) {
2648 for (j = 0; j < svc->srv_nthrs_cpt_init; j++) {
2649 rc = ptlrpc_start_thread(svc->srv_parts[i], 1);
2655 /* We have enough threads, don't start more. b=15759 */
2662 CERROR("cannot start %s thread #%d_%d: rc %d\n",
2663 svc->srv_thread_name, i, j, rc);
2664 ptlrpc_stop_all_threads(svc);
2667 EXPORT_SYMBOL(ptlrpc_start_threads);
2669 int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
2671 struct l_wait_info lwi = { 0 };
2672 struct ptlrpc_thread *thread;
2673 struct ptlrpc_service *svc;
2676 LASSERT(svcpt != NULL);
2678 svc = svcpt->scp_service;
2680 CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n",
2681 svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running,
2682 svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit);
2685 if (unlikely(svc->srv_is_stopping))
2688 if (!ptlrpc_threads_increasable(svcpt) ||
2689 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
2690 svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1))
2693 OBD_CPT_ALLOC_PTR(thread, svc->srv_cptable, svcpt->scp_cpt);
2696 init_waitqueue_head(&thread->t_ctl_waitq);
2698 spin_lock(&svcpt->scp_lock);
2699 if (!ptlrpc_threads_increasable(svcpt)) {
2700 spin_unlock(&svcpt->scp_lock);
2701 OBD_FREE_PTR(thread);
2705 if (svcpt->scp_nthrs_starting != 0) {
2706 /* serialize starting because some modules (obdfilter)
2707 * might require unique and contiguous t_id */
2708 LASSERT(svcpt->scp_nthrs_starting == 1);
2709 spin_unlock(&svcpt->scp_lock);
2710 OBD_FREE_PTR(thread);
2712 CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n",
2713 svc->srv_thread_name, svcpt->scp_thr_nextid);
2718 CDEBUG(D_INFO, "Creating thread %s #%d race, retry later\n",
2719 svc->srv_thread_name, svcpt->scp_thr_nextid);
2723 svcpt->scp_nthrs_starting++;
2724 thread->t_id = svcpt->scp_thr_nextid++;
2725 thread_add_flags(thread, SVC_STARTING);
2726 thread->t_svcpt = svcpt;
2728 list_add(&thread->t_link, &svcpt->scp_threads);
2729 spin_unlock(&svcpt->scp_lock);
2731 if (svcpt->scp_cpt >= 0) {
2732 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s%02d_%03d",
2733 svc->srv_thread_name, svcpt->scp_cpt, thread->t_id);
2735 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s_%04d",
2736 svc->srv_thread_name, thread->t_id);
2739 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name);
2740 rc = PTR_ERR(kthread_run(ptlrpc_main, thread, thread->t_name));
2741 if (IS_ERR_VALUE(rc)) {
2742 CERROR("cannot start thread '%s': rc %d\n",
2743 thread->t_name, rc);
2744 spin_lock(&svcpt->scp_lock);
2745 --svcpt->scp_nthrs_starting;
2746 if (thread_is_stopping(thread)) {
2747 /* this ptlrpc_thread is being hanled
2748 * by ptlrpc_svcpt_stop_threads now
2750 thread_add_flags(thread, SVC_STOPPED);
2751 wake_up(&thread->t_ctl_waitq);
2752 spin_unlock(&svcpt->scp_lock);
2754 list_del(&thread->t_link);
2755 spin_unlock(&svcpt->scp_lock);
2756 OBD_FREE_PTR(thread);
2764 l_wait_event(thread->t_ctl_waitq,
2765 thread_is_running(thread) || thread_is_stopped(thread),
2768 rc = thread_is_stopped(thread) ? thread->t_id : 0;
2772 int ptlrpc_hr_init(void)
2775 struct ptlrpc_hr_partition *hrp;
2776 struct ptlrpc_hr_thread *hrt;
2782 memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
2783 ptlrpc_hr.hr_cpt_table = cfs_cpt_table;
2785 ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
2787 if (ptlrpc_hr.hr_partitions == NULL)
2790 init_waitqueue_head(&ptlrpc_hr.hr_waitq);
2792 cpumask_copy(&mask, topology_thread_cpumask(0));
2793 weight = cpus_weight(mask);
2795 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2798 atomic_set(&hrp->hrp_nstarted, 0);
2799 atomic_set(&hrp->hrp_nstopped, 0);
2801 hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, i);
2802 hrp->hrp_nthrs /= weight;
2804 LASSERT(hrp->hrp_nthrs > 0);
2805 OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, i,
2806 hrp->hrp_nthrs * sizeof(*hrt));
2807 if (hrp->hrp_thrs == NULL)
2808 GOTO(out, rc = -ENOMEM);
2810 for (j = 0; j < hrp->hrp_nthrs; j++) {
2811 hrt = &hrp->hrp_thrs[j];
2814 hrt->hrt_partition = hrp;
2815 init_waitqueue_head(&hrt->hrt_waitq);
2816 spin_lock_init(&hrt->hrt_lock);
2817 INIT_LIST_HEAD(&hrt->hrt_queue);
2821 rc = ptlrpc_start_hr_threads();
2828 void ptlrpc_hr_fini(void)
2830 struct ptlrpc_hr_partition *hrp;
2833 if (ptlrpc_hr.hr_partitions == NULL)
2836 ptlrpc_stop_hr_threads();
2838 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2839 if (hrp->hrp_thrs != NULL) {
2840 OBD_FREE(hrp->hrp_thrs,
2841 hrp->hrp_nthrs * sizeof(hrp->hrp_thrs[0]));
2845 cfs_percpt_free(ptlrpc_hr.hr_partitions);
2846 ptlrpc_hr.hr_partitions = NULL;
2851 * Wait until all already scheduled replies are processed.
2853 static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
2857 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
2860 rc = l_wait_event(svcpt->scp_waitq,
2861 atomic_read(&svcpt->scp_nreps_difficult) == 0, &lwi);
2864 CWARN("Unexpectedly long timeout %s %p\n",
2865 svcpt->scp_service->srv_name, svcpt->scp_service);
2870 ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
2872 struct ptlrpc_service_part *svcpt;
2875 /* early disarm AT timer... */
2876 ptlrpc_service_for_each_part(svcpt, i, svc) {
2877 if (svcpt->scp_service != NULL)
2878 cfs_timer_disarm(&svcpt->scp_at_timer);
2883 ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
2885 struct ptlrpc_service_part *svcpt;
2886 struct ptlrpc_request_buffer_desc *rqbd;
2887 struct l_wait_info lwi;
2891 /* All history will be culled when the next request buffer is
2892 * freed in ptlrpc_service_purge_all() */
2893 svc->srv_hist_nrqbds_cpt_max = 0;
2895 rc = LNetClearLazyPortal(svc->srv_req_portal);
2898 ptlrpc_service_for_each_part(svcpt, i, svc) {
2899 if (svcpt->scp_service == NULL)
2902 /* Unlink all the request buffers. This forces a 'final'
2903 * event with its 'unlink' flag set for each posted rqbd */
2904 list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
2906 rc = LNetMDUnlink(rqbd->rqbd_md_h);
2907 LASSERT(rc == 0 || rc == -ENOENT);
2911 ptlrpc_service_for_each_part(svcpt, i, svc) {
2912 if (svcpt->scp_service == NULL)
2915 /* Wait for the network to release any buffers
2916 * it's currently filling */
2917 spin_lock(&svcpt->scp_lock);
2918 while (svcpt->scp_nrqbds_posted != 0) {
2919 spin_unlock(&svcpt->scp_lock);
2920 /* Network access will complete in finite time but
2921 * the HUGE timeout lets us CWARN for visibility
2922 * of sluggish NALs */
2923 lwi = LWI_TIMEOUT_INTERVAL(
2924 cfs_time_seconds(LONG_UNLINK),
2925 cfs_time_seconds(1), NULL, NULL);
2926 rc = l_wait_event(svcpt->scp_waitq,
2927 svcpt->scp_nrqbds_posted == 0, &lwi);
2928 if (rc == -ETIMEDOUT) {
2929 CWARN("Service %s waiting for "
2930 "request buffers\n",
2931 svcpt->scp_service->srv_name);
2933 spin_lock(&svcpt->scp_lock);
2935 spin_unlock(&svcpt->scp_lock);
2940 ptlrpc_service_purge_all(struct ptlrpc_service *svc)
2942 struct ptlrpc_service_part *svcpt;
2943 struct ptlrpc_request_buffer_desc *rqbd;
2944 struct ptlrpc_request *req;
2945 struct ptlrpc_reply_state *rs;
2948 ptlrpc_service_for_each_part(svcpt, i, svc) {
2949 if (svcpt->scp_service == NULL)
2952 spin_lock(&svcpt->scp_rep_lock);
2953 while (!list_empty(&svcpt->scp_rep_active)) {
2954 rs = list_entry(svcpt->scp_rep_active.next,
2955 struct ptlrpc_reply_state, rs_list);
2956 spin_lock(&rs->rs_lock);
2957 ptlrpc_schedule_difficult_reply(rs);
2958 spin_unlock(&rs->rs_lock);
2960 spin_unlock(&svcpt->scp_rep_lock);
2962 /* purge the request queue. NB No new replies (rqbds
2963 * all unlinked) and no service threads, so I'm the only
2964 * thread noodling the request queue now */
2965 while (!list_empty(&svcpt->scp_req_incoming)) {
2966 req = list_entry(svcpt->scp_req_incoming.next,
2967 struct ptlrpc_request, rq_list);
2969 list_del(&req->rq_list);
2970 svcpt->scp_nreqs_incoming--;
2971 ptlrpc_server_finish_request(svcpt, req);
2974 while (ptlrpc_server_request_pending(svcpt, true)) {
2975 req = ptlrpc_server_request_get(svcpt, true);
2976 ptlrpc_server_finish_active_request(svcpt, req);
2979 LASSERT(list_empty(&svcpt->scp_rqbd_posted));
2980 LASSERT(svcpt->scp_nreqs_incoming == 0);
2981 LASSERT(svcpt->scp_nreqs_active == 0);
2982 /* history should have been culled by
2983 * ptlrpc_server_finish_request */
2984 LASSERT(svcpt->scp_hist_nrqbds == 0);
2986 /* Now free all the request buffers since nothing
2987 * references them any more... */
2989 while (!list_empty(&svcpt->scp_rqbd_idle)) {
2990 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
2991 struct ptlrpc_request_buffer_desc,
2993 ptlrpc_free_rqbd(rqbd);
2995 ptlrpc_wait_replies(svcpt);
2997 while (!list_empty(&svcpt->scp_rep_idle)) {
2998 rs = list_entry(svcpt->scp_rep_idle.next,
2999 struct ptlrpc_reply_state,
3001 list_del(&rs->rs_list);
3002 OBD_FREE_LARGE(rs, svc->srv_max_reply_size);
3008 ptlrpc_service_free(struct ptlrpc_service *svc)
3010 struct ptlrpc_service_part *svcpt;
3011 struct ptlrpc_at_array *array;
3014 ptlrpc_service_for_each_part(svcpt, i, svc) {
3015 if (svcpt->scp_service == NULL)
3018 /* In case somebody rearmed this in the meantime */
3019 cfs_timer_disarm(&svcpt->scp_at_timer);
3020 array = &svcpt->scp_at_array;
3022 if (array->paa_reqs_array != NULL) {
3023 OBD_FREE(array->paa_reqs_array,
3024 sizeof(struct list_head) * array->paa_size);
3025 array->paa_reqs_array = NULL;
3028 if (array->paa_reqs_count != NULL) {
3029 OBD_FREE(array->paa_reqs_count,
3030 sizeof(__u32) * array->paa_size);
3031 array->paa_reqs_count = NULL;
3035 ptlrpc_service_for_each_part(svcpt, i, svc)
3036 OBD_FREE_PTR(svcpt);
3038 if (svc->srv_cpts != NULL)
3039 cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts);
3041 OBD_FREE(svc, offsetof(struct ptlrpc_service,
3042 srv_parts[svc->srv_ncpts]));
3045 int ptlrpc_unregister_service(struct ptlrpc_service *service)
3047 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
3049 service->srv_is_stopping = 1;
3051 mutex_lock(&ptlrpc_all_services_mutex);
3052 list_del_init(&service->srv_list);
3053 mutex_unlock(&ptlrpc_all_services_mutex);
3055 ptlrpc_service_del_atimer(service);
3056 ptlrpc_stop_all_threads(service);
3058 ptlrpc_service_unlink_rqbd(service);
3059 ptlrpc_service_purge_all(service);
3060 ptlrpc_service_nrs_cleanup(service);
3062 ptlrpc_lprocfs_unregister_service(service);
3064 ptlrpc_service_free(service);
3068 EXPORT_SYMBOL(ptlrpc_unregister_service);
3071 * Returns 0 if the service is healthy.
3073 * Right now, it just checks to make sure that requests aren't languishing
3074 * in the queue. We'll use this health check to govern whether a node needs
3075 * to be shot, so it's intentionally non-aggressive. */
3076 int ptlrpc_svcpt_health_check(struct ptlrpc_service_part *svcpt)
3078 struct ptlrpc_request *request = NULL;
3079 struct timeval right_now;
3082 do_gettimeofday(&right_now);
3084 spin_lock(&svcpt->scp_req_lock);
3085 /* How long has the next entry been waiting? */
3086 if (ptlrpc_server_high_pending(svcpt, true))
3087 request = ptlrpc_nrs_req_peek_nolock(svcpt, true);
3088 else if (ptlrpc_server_normal_pending(svcpt, true))
3089 request = ptlrpc_nrs_req_peek_nolock(svcpt, false);
3091 if (request == NULL) {
3092 spin_unlock(&svcpt->scp_req_lock);
3096 timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
3097 spin_unlock(&svcpt->scp_req_lock);
3099 if ((timediff / ONE_MILLION) >
3100 (AT_OFF ? obd_timeout * 3 / 2 : at_max)) {
3101 CERROR("%s: unhealthy - request has been waiting %lds\n",
3102 svcpt->scp_service->srv_name, timediff / ONE_MILLION);
3110 ptlrpc_service_health_check(struct ptlrpc_service *svc)
3112 struct ptlrpc_service_part *svcpt;
3118 ptlrpc_service_for_each_part(svcpt, i, svc) {
3119 int rc = ptlrpc_svcpt_health_check(svcpt);
3126 EXPORT_SYMBOL(ptlrpc_service_health_check);