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 "../include/obd_support.h"
39 #include "../include/obd_class.h"
40 #include "../include/lustre_net.h"
41 #include "../include/lu_object.h"
42 #include "../../include/linux/lnet/types.h"
43 #include "ptlrpc_internal.h"
45 /* The following are visible and mutable through /sys/module/ptlrpc */
46 int test_req_buffer_pressure = 0;
47 module_param(test_req_buffer_pressure, int, 0444);
48 MODULE_PARM_DESC(test_req_buffer_pressure, "set non-zero to put pressure on request buffer pools");
49 module_param(at_min, int, 0644);
50 MODULE_PARM_DESC(at_min, "Adaptive timeout minimum (sec)");
51 module_param(at_max, int, 0644);
52 MODULE_PARM_DESC(at_max, "Adaptive timeout maximum (sec)");
53 module_param(at_history, int, 0644);
54 MODULE_PARM_DESC(at_history,
55 "Adaptive timeouts remember the slowest event that took place within this period (sec)");
56 module_param(at_early_margin, int, 0644);
57 MODULE_PARM_DESC(at_early_margin, "How soon before an RPC deadline to send an early reply");
58 module_param(at_extra, int, 0644);
59 MODULE_PARM_DESC(at_extra, "How much extra time to give with each early reply");
63 static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt);
64 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req);
65 static void ptlrpc_at_remove_timed(struct ptlrpc_request *req);
67 /** Holds a list of all PTLRPC services */
68 LIST_HEAD(ptlrpc_all_services);
69 /** Used to protect the \e ptlrpc_all_services list */
70 struct mutex ptlrpc_all_services_mutex;
72 struct ptlrpc_request_buffer_desc *
73 ptlrpc_alloc_rqbd(struct ptlrpc_service_part *svcpt)
75 struct ptlrpc_service *svc = svcpt->scp_service;
76 struct ptlrpc_request_buffer_desc *rqbd;
78 OBD_CPT_ALLOC_PTR(rqbd, svc->srv_cptable, svcpt->scp_cpt);
82 rqbd->rqbd_svcpt = svcpt;
83 rqbd->rqbd_refcount = 0;
84 rqbd->rqbd_cbid.cbid_fn = request_in_callback;
85 rqbd->rqbd_cbid.cbid_arg = rqbd;
86 INIT_LIST_HEAD(&rqbd->rqbd_reqs);
87 OBD_CPT_ALLOC_LARGE(rqbd->rqbd_buffer, svc->srv_cptable,
88 svcpt->scp_cpt, svc->srv_buf_size);
89 if (rqbd->rqbd_buffer == NULL) {
94 spin_lock(&svcpt->scp_lock);
95 list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
96 svcpt->scp_nrqbds_total++;
97 spin_unlock(&svcpt->scp_lock);
103 ptlrpc_free_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
105 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
107 LASSERT(rqbd->rqbd_refcount == 0);
108 LASSERT(list_empty(&rqbd->rqbd_reqs));
110 spin_lock(&svcpt->scp_lock);
111 list_del(&rqbd->rqbd_list);
112 svcpt->scp_nrqbds_total--;
113 spin_unlock(&svcpt->scp_lock);
115 OBD_FREE_LARGE(rqbd->rqbd_buffer, svcpt->scp_service->srv_buf_size);
120 ptlrpc_grow_req_bufs(struct ptlrpc_service_part *svcpt, int post)
122 struct ptlrpc_service *svc = svcpt->scp_service;
123 struct ptlrpc_request_buffer_desc *rqbd;
127 if (svcpt->scp_rqbd_allocating)
130 spin_lock(&svcpt->scp_lock);
131 /* check again with lock */
132 if (svcpt->scp_rqbd_allocating) {
133 /* NB: we might allow more than one thread in the future */
134 LASSERT(svcpt->scp_rqbd_allocating == 1);
135 spin_unlock(&svcpt->scp_lock);
139 svcpt->scp_rqbd_allocating++;
140 spin_unlock(&svcpt->scp_lock);
143 for (i = 0; i < svc->srv_nbuf_per_group; i++) {
144 /* NB: another thread might have recycled enough rqbds, we
145 * need to make sure it wouldn't over-allocate, see LU-1212. */
146 if (svcpt->scp_nrqbds_posted >= svc->srv_nbuf_per_group)
149 rqbd = ptlrpc_alloc_rqbd(svcpt);
152 CERROR("%s: Can't allocate request buffer\n",
159 spin_lock(&svcpt->scp_lock);
161 LASSERT(svcpt->scp_rqbd_allocating == 1);
162 svcpt->scp_rqbd_allocating--;
164 spin_unlock(&svcpt->scp_lock);
167 "%s: allocate %d new %d-byte reqbufs (%d/%d left), rc = %d\n",
168 svc->srv_name, i, svc->srv_buf_size, svcpt->scp_nrqbds_posted,
169 svcpt->scp_nrqbds_total, rc);
173 rc = ptlrpc_server_post_idle_rqbds(svcpt);
179 * Part of Rep-Ack logic.
180 * Puts a lock and its mode into reply state associated 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 number of replies scheduled in one batch
257 #define MAX_SCHEDULED 256
260 * Initialize a reply batch.
264 static void rs_batch_init(struct rs_batch *b)
266 memset(b, 0, sizeof(*b));
267 INIT_LIST_HEAD(&b->rsb_replies);
271 * Choose an hr thread to dispatch requests to.
273 static struct ptlrpc_hr_thread *
274 ptlrpc_hr_select(struct ptlrpc_service_part *svcpt)
276 struct ptlrpc_hr_partition *hrp;
279 if (svcpt->scp_cpt >= 0 &&
280 svcpt->scp_service->srv_cptable == ptlrpc_hr.hr_cpt_table) {
281 /* directly match partition */
282 hrp = ptlrpc_hr.hr_partitions[svcpt->scp_cpt];
285 rotor = ptlrpc_hr.hr_rotor++;
286 rotor %= cfs_cpt_number(ptlrpc_hr.hr_cpt_table);
288 hrp = ptlrpc_hr.hr_partitions[rotor];
291 rotor = hrp->hrp_rotor++;
292 return &hrp->hrp_thrs[rotor % hrp->hrp_nthrs];
296 * Dispatch all replies accumulated in the batch to one from
297 * dedicated reply handling threads.
301 static void rs_batch_dispatch(struct rs_batch *b)
303 if (b->rsb_n_replies != 0) {
304 struct ptlrpc_hr_thread *hrt;
306 hrt = ptlrpc_hr_select(b->rsb_svcpt);
308 spin_lock(&hrt->hrt_lock);
309 list_splice_init(&b->rsb_replies, &hrt->hrt_queue);
310 spin_unlock(&hrt->hrt_lock);
312 wake_up(&hrt->hrt_waitq);
313 b->rsb_n_replies = 0;
318 * Add a reply to a batch.
319 * Add one reply object to a batch, schedule batched replies if overload.
324 static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs)
326 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
328 if (svcpt != b->rsb_svcpt || b->rsb_n_replies >= MAX_SCHEDULED) {
329 if (b->rsb_svcpt != NULL) {
330 rs_batch_dispatch(b);
331 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
333 spin_lock(&svcpt->scp_rep_lock);
334 b->rsb_svcpt = svcpt;
336 spin_lock(&rs->rs_lock);
337 rs->rs_scheduled_ever = 1;
338 if (rs->rs_scheduled == 0) {
339 list_move(&rs->rs_list, &b->rsb_replies);
340 rs->rs_scheduled = 1;
343 rs->rs_committed = 1;
344 spin_unlock(&rs->rs_lock);
348 * Reply batch finalization.
349 * Dispatch remaining replies from the batch
350 * and release remaining spinlock.
354 static void rs_batch_fini(struct rs_batch *b)
356 if (b->rsb_svcpt != NULL) {
357 rs_batch_dispatch(b);
358 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
362 #define DECLARE_RS_BATCH(b) struct rs_batch b
366 * Put reply state into a queue for processing because we received
367 * ACK from the client
369 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
371 struct ptlrpc_hr_thread *hrt;
373 LASSERT(list_empty(&rs->rs_list));
375 hrt = ptlrpc_hr_select(rs->rs_svcpt);
377 spin_lock(&hrt->hrt_lock);
378 list_add_tail(&rs->rs_list, &hrt->hrt_queue);
379 spin_unlock(&hrt->hrt_lock);
381 wake_up(&hrt->hrt_waitq);
385 ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs)
387 assert_spin_locked(&rs->rs_svcpt->scp_rep_lock);
388 assert_spin_locked(&rs->rs_lock);
389 LASSERT(rs->rs_difficult);
390 rs->rs_scheduled_ever = 1; /* flag any notification attempt */
392 if (rs->rs_scheduled) { /* being set up or already notified */
396 rs->rs_scheduled = 1;
397 list_del_init(&rs->rs_list);
398 ptlrpc_dispatch_difficult_reply(rs);
400 EXPORT_SYMBOL(ptlrpc_schedule_difficult_reply);
402 void ptlrpc_commit_replies(struct obd_export *exp)
404 struct ptlrpc_reply_state *rs, *nxt;
405 DECLARE_RS_BATCH(batch);
407 rs_batch_init(&batch);
408 /* Find any replies that have been committed and get their service
409 * to attend to complete them. */
411 /* CAVEAT EMPTOR: spinlock ordering!!! */
412 spin_lock(&exp->exp_uncommitted_replies_lock);
413 list_for_each_entry_safe(rs, nxt, &exp->exp_uncommitted_replies,
415 LASSERT(rs->rs_difficult);
416 /* VBR: per-export last_committed */
417 LASSERT(rs->rs_export);
418 if (rs->rs_transno <= exp->exp_last_committed) {
419 list_del_init(&rs->rs_obd_list);
420 rs_batch_add(&batch, rs);
423 spin_unlock(&exp->exp_uncommitted_replies_lock);
424 rs_batch_fini(&batch);
426 EXPORT_SYMBOL(ptlrpc_commit_replies);
429 ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt)
431 struct ptlrpc_request_buffer_desc *rqbd;
436 spin_lock(&svcpt->scp_lock);
438 if (list_empty(&svcpt->scp_rqbd_idle)) {
439 spin_unlock(&svcpt->scp_lock);
443 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
444 struct ptlrpc_request_buffer_desc,
446 list_del(&rqbd->rqbd_list);
448 /* assume we will post successfully */
449 svcpt->scp_nrqbds_posted++;
450 list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_posted);
452 spin_unlock(&svcpt->scp_lock);
454 rc = ptlrpc_register_rqbd(rqbd);
461 spin_lock(&svcpt->scp_lock);
463 svcpt->scp_nrqbds_posted--;
464 list_del(&rqbd->rqbd_list);
465 list_add_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
467 /* Don't complain if no request buffers are posted right now; LNET
468 * won't drop requests because we set the portal lazy! */
470 spin_unlock(&svcpt->scp_lock);
475 static void ptlrpc_at_timer(unsigned long castmeharder)
477 struct ptlrpc_service_part *svcpt;
479 svcpt = (struct ptlrpc_service_part *)castmeharder;
481 svcpt->scp_at_check = 1;
482 svcpt->scp_at_checktime = cfs_time_current();
483 wake_up(&svcpt->scp_waitq);
487 ptlrpc_server_nthreads_check(struct ptlrpc_service *svc,
488 struct ptlrpc_service_conf *conf)
490 struct ptlrpc_service_thr_conf *tc = &conf->psc_thr;
497 * Common code for estimating & validating threads number.
498 * CPT affinity service could have percpt thread-pool instead
499 * of a global thread-pool, which means user might not always
500 * get the threads number they give it in conf::tc_nthrs_user
501 * even they did set. It's because we need to validate threads
502 * number for each CPT to guarantee each pool will have enough
503 * threads to keep the service healthy.
505 init = PTLRPC_NTHRS_INIT + (svc->srv_ops.so_hpreq_handler != NULL);
506 init = max_t(int, init, tc->tc_nthrs_init);
508 /* NB: please see comments in lustre_lnet.h for definition
509 * details of these members */
510 LASSERT(tc->tc_nthrs_max != 0);
512 if (tc->tc_nthrs_user != 0) {
513 /* In case there is a reason to test a service with many
514 * threads, we give a less strict check here, it can
515 * be up to 8 * nthrs_max */
516 total = min(tc->tc_nthrs_max * 8, tc->tc_nthrs_user);
517 nthrs = total / svc->srv_ncpts;
518 init = max(init, nthrs);
522 total = tc->tc_nthrs_max;
523 if (tc->tc_nthrs_base == 0) {
524 /* don't care about base threads number per partition,
525 * this is most for non-affinity service */
526 nthrs = total / svc->srv_ncpts;
530 nthrs = tc->tc_nthrs_base;
531 if (svc->srv_ncpts == 1) {
534 /* NB: Increase the base number if it's single partition
535 * and total number of cores/HTs is larger or equal to 4.
536 * result will always < 2 * nthrs_base */
537 weight = cfs_cpt_weight(svc->srv_cptable, CFS_CPT_ANY);
538 for (i = 1; (weight >> (i + 1)) != 0 && /* >= 4 cores/HTs */
539 (tc->tc_nthrs_base >> i) != 0; i++)
540 nthrs += tc->tc_nthrs_base >> i;
543 if (tc->tc_thr_factor != 0) {
544 int factor = tc->tc_thr_factor;
548 * User wants to increase number of threads with for
549 * each CPU core/HT, most likely the factor is larger then
550 * one thread/core because service threads are supposed to
551 * be blocked by lock or wait for IO.
554 * Amdahl's law says that adding processors wouldn't give
555 * a linear increasing of parallelism, so it's nonsense to
556 * have too many threads no matter how many cores/HTs
559 /* weight is # of HTs */
560 if (cpumask_weight(topology_thread_cpumask(0)) > 1) {
561 /* depress thread factor for hyper-thread */
562 factor = factor - (factor >> 1) + (factor >> 3);
565 weight = cfs_cpt_weight(svc->srv_cptable, 0);
568 for (; factor > 0 && weight > 0; factor--, weight -= fade)
569 nthrs += min(weight, fade) * factor;
572 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
573 nthrs = max(tc->tc_nthrs_base,
574 tc->tc_nthrs_max / svc->srv_ncpts);
577 nthrs = max(nthrs, tc->tc_nthrs_init);
578 svc->srv_nthrs_cpt_limit = nthrs;
579 svc->srv_nthrs_cpt_init = init;
581 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
582 CDEBUG(D_OTHER, "%s: This service may have more threads (%d) than the given soft limit (%d)\n",
583 svc->srv_name, nthrs * svc->srv_ncpts,
589 * Initialize percpt data for a service
592 ptlrpc_service_part_init(struct ptlrpc_service *svc,
593 struct ptlrpc_service_part *svcpt, int cpt)
595 struct ptlrpc_at_array *array;
600 svcpt->scp_cpt = cpt;
601 INIT_LIST_HEAD(&svcpt->scp_threads);
603 /* rqbd and incoming request queue */
604 spin_lock_init(&svcpt->scp_lock);
605 INIT_LIST_HEAD(&svcpt->scp_rqbd_idle);
606 INIT_LIST_HEAD(&svcpt->scp_rqbd_posted);
607 INIT_LIST_HEAD(&svcpt->scp_req_incoming);
608 init_waitqueue_head(&svcpt->scp_waitq);
609 /* history request & rqbd list */
610 INIT_LIST_HEAD(&svcpt->scp_hist_reqs);
611 INIT_LIST_HEAD(&svcpt->scp_hist_rqbds);
613 /* active requests and hp requests */
614 spin_lock_init(&svcpt->scp_req_lock);
617 spin_lock_init(&svcpt->scp_rep_lock);
618 INIT_LIST_HEAD(&svcpt->scp_rep_active);
619 INIT_LIST_HEAD(&svcpt->scp_rep_idle);
620 init_waitqueue_head(&svcpt->scp_rep_waitq);
621 atomic_set(&svcpt->scp_nreps_difficult, 0);
623 /* adaptive timeout */
624 spin_lock_init(&svcpt->scp_at_lock);
625 array = &svcpt->scp_at_array;
627 size = at_est2timeout(at_max);
628 array->paa_size = size;
629 array->paa_count = 0;
630 array->paa_deadline = -1;
632 /* allocate memory for scp_at_array (ptlrpc_at_array) */
633 OBD_CPT_ALLOC(array->paa_reqs_array,
634 svc->srv_cptable, cpt, sizeof(struct list_head) * size);
635 if (array->paa_reqs_array == NULL)
638 for (index = 0; index < size; index++)
639 INIT_LIST_HEAD(&array->paa_reqs_array[index]);
641 OBD_CPT_ALLOC(array->paa_reqs_count,
642 svc->srv_cptable, cpt, sizeof(__u32) * size);
643 if (array->paa_reqs_count == NULL)
646 cfs_timer_init(&svcpt->scp_at_timer, ptlrpc_at_timer, svcpt);
647 /* At SOW, service time should be quick; 10s seems generous. If client
648 * timeout is less than this, we'll be sending an early reply. */
649 at_init(&svcpt->scp_at_estimate, 10, 0);
651 /* assign this before call ptlrpc_grow_req_bufs */
652 svcpt->scp_service = svc;
653 /* Now allocate the request buffers, but don't post them now */
654 rc = ptlrpc_grow_req_bufs(svcpt, 0);
655 /* We shouldn't be under memory pressure at startup, so
656 * fail if we can't allocate all our buffers at this time. */
663 if (array->paa_reqs_count != NULL) {
664 OBD_FREE(array->paa_reqs_count, sizeof(__u32) * size);
665 array->paa_reqs_count = NULL;
668 if (array->paa_reqs_array != NULL) {
669 OBD_FREE(array->paa_reqs_array,
670 sizeof(struct list_head) * array->paa_size);
671 array->paa_reqs_array = NULL;
678 * Initialize service on a given portal.
679 * This includes starting serving threads , allocating and posting rqbds and
682 struct ptlrpc_service *
683 ptlrpc_register_service(struct ptlrpc_service_conf *conf,
684 struct proc_dir_entry *proc_entry)
686 struct ptlrpc_service_cpt_conf *cconf = &conf->psc_cpt;
687 struct ptlrpc_service *service;
688 struct ptlrpc_service_part *svcpt;
689 struct cfs_cpt_table *cptable;
696 LASSERT(conf->psc_buf.bc_nbufs > 0);
697 LASSERT(conf->psc_buf.bc_buf_size >=
698 conf->psc_buf.bc_req_max_size + SPTLRPC_MAX_PAYLOAD);
699 LASSERT(conf->psc_thr.tc_ctx_tags != 0);
701 cptable = cconf->cc_cptable;
703 cptable = cfs_cpt_table;
705 if (!conf->psc_thr.tc_cpu_affinity) {
708 ncpts = cfs_cpt_number(cptable);
709 if (cconf->cc_pattern != NULL) {
710 struct cfs_expr_list *el;
712 rc = cfs_expr_list_parse(cconf->cc_pattern,
713 strlen(cconf->cc_pattern),
716 CERROR("%s: invalid CPT pattern string: %s",
717 conf->psc_name, cconf->cc_pattern);
718 return ERR_PTR(-EINVAL);
721 rc = cfs_expr_list_values(el, ncpts, &cpts);
722 cfs_expr_list_free(el);
724 CERROR("%s: failed to parse CPT array %s: %d\n",
725 conf->psc_name, cconf->cc_pattern, rc);
727 OBD_FREE(cpts, sizeof(*cpts) * ncpts);
728 return ERR_PTR(rc < 0 ? rc : -EINVAL);
734 OBD_ALLOC(service, offsetof(struct ptlrpc_service, srv_parts[ncpts]));
735 if (service == NULL) {
737 OBD_FREE(cpts, sizeof(*cpts) * ncpts);
738 return ERR_PTR(-ENOMEM);
741 service->srv_cptable = cptable;
742 service->srv_cpts = cpts;
743 service->srv_ncpts = ncpts;
745 service->srv_cpt_bits = 0; /* it's zero already, easy to read... */
746 while ((1 << service->srv_cpt_bits) < cfs_cpt_number(cptable))
747 service->srv_cpt_bits++;
750 spin_lock_init(&service->srv_lock);
751 service->srv_name = conf->psc_name;
752 service->srv_watchdog_factor = conf->psc_watchdog_factor;
753 INIT_LIST_HEAD(&service->srv_list); /* for safety of cleanup */
755 /* buffer configuration */
756 service->srv_nbuf_per_group = test_req_buffer_pressure ?
757 1 : conf->psc_buf.bc_nbufs;
758 service->srv_max_req_size = conf->psc_buf.bc_req_max_size +
760 service->srv_buf_size = conf->psc_buf.bc_buf_size;
761 service->srv_rep_portal = conf->psc_buf.bc_rep_portal;
762 service->srv_req_portal = conf->psc_buf.bc_req_portal;
764 /* Increase max reply size to next power of two */
765 service->srv_max_reply_size = 1;
766 while (service->srv_max_reply_size <
767 conf->psc_buf.bc_rep_max_size + SPTLRPC_MAX_PAYLOAD)
768 service->srv_max_reply_size <<= 1;
770 service->srv_thread_name = conf->psc_thr.tc_thr_name;
771 service->srv_ctx_tags = conf->psc_thr.tc_ctx_tags;
772 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
773 service->srv_ops = conf->psc_ops;
775 for (i = 0; i < ncpts; i++) {
776 if (!conf->psc_thr.tc_cpu_affinity)
779 cpt = cpts != NULL ? cpts[i] : i;
781 OBD_CPT_ALLOC(svcpt, cptable, cpt, sizeof(*svcpt));
787 service->srv_parts[i] = svcpt;
788 rc = ptlrpc_service_part_init(service, svcpt, cpt);
793 ptlrpc_server_nthreads_check(service, conf);
795 rc = LNetSetLazyPortal(service->srv_req_portal);
798 mutex_lock(&ptlrpc_all_services_mutex);
799 list_add(&service->srv_list, &ptlrpc_all_services);
800 mutex_unlock(&ptlrpc_all_services_mutex);
802 if (proc_entry != NULL)
803 ptlrpc_lprocfs_register_service(proc_entry, service);
805 rc = ptlrpc_service_nrs_setup(service);
809 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
810 service->srv_name, service->srv_req_portal);
812 rc = ptlrpc_start_threads(service);
814 CERROR("Failed to start threads for service %s: %d\n",
815 service->srv_name, rc);
821 ptlrpc_unregister_service(service);
824 EXPORT_SYMBOL(ptlrpc_register_service);
827 * to actually free the request, must be called without holding svc_lock.
828 * note it's caller's responsibility to unlink req->rq_list.
830 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
832 LASSERT(atomic_read(&req->rq_refcount) == 0);
833 LASSERT(list_empty(&req->rq_timed_list));
835 /* DEBUG_REQ() assumes the reply state of a request with a valid
836 * ref will not be destroyed until that reference is dropped. */
837 ptlrpc_req_drop_rs(req);
839 sptlrpc_svc_ctx_decref(req);
841 if (req != &req->rq_rqbd->rqbd_req) {
842 /* NB request buffers use an embedded
843 * req if the incoming req unlinked the
844 * MD; this isn't one of them! */
845 ptlrpc_request_cache_free(req);
850 * drop a reference count of the request. if it reaches 0, we either
851 * put it into history list, or free it immediately.
853 void ptlrpc_server_drop_request(struct ptlrpc_request *req)
855 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
856 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
857 struct ptlrpc_service *svc = svcpt->scp_service;
859 struct list_head *tmp;
860 struct list_head *nxt;
862 if (!atomic_dec_and_test(&req->rq_refcount))
865 if (req->rq_at_linked) {
866 spin_lock(&svcpt->scp_at_lock);
867 /* recheck with lock, in case it's unlinked by
868 * ptlrpc_at_check_timed() */
869 if (likely(req->rq_at_linked))
870 ptlrpc_at_remove_timed(req);
871 spin_unlock(&svcpt->scp_at_lock);
874 LASSERT(list_empty(&req->rq_timed_list));
876 /* finalize request */
877 if (req->rq_export) {
878 class_export_put(req->rq_export);
879 req->rq_export = NULL;
882 spin_lock(&svcpt->scp_lock);
884 list_add(&req->rq_list, &rqbd->rqbd_reqs);
886 refcount = --(rqbd->rqbd_refcount);
888 /* request buffer is now idle: add to history */
889 list_del(&rqbd->rqbd_list);
891 list_add_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds);
892 svcpt->scp_hist_nrqbds++;
894 /* cull some history?
895 * I expect only about 1 or 2 rqbds need to be recycled here */
896 while (svcpt->scp_hist_nrqbds > svc->srv_hist_nrqbds_cpt_max) {
897 rqbd = list_entry(svcpt->scp_hist_rqbds.next,
898 struct ptlrpc_request_buffer_desc,
901 list_del(&rqbd->rqbd_list);
902 svcpt->scp_hist_nrqbds--;
904 /* remove rqbd's reqs from svc's req history while
905 * I've got the service lock */
906 list_for_each(tmp, &rqbd->rqbd_reqs) {
907 req = list_entry(tmp, struct ptlrpc_request,
909 /* Track the highest culled req seq */
910 if (req->rq_history_seq >
911 svcpt->scp_hist_seq_culled) {
912 svcpt->scp_hist_seq_culled =
915 list_del(&req->rq_history_list);
918 spin_unlock(&svcpt->scp_lock);
920 list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
921 req = list_entry(rqbd->rqbd_reqs.next,
922 struct ptlrpc_request,
924 list_del(&req->rq_list);
925 ptlrpc_server_free_request(req);
928 spin_lock(&svcpt->scp_lock);
930 * now all reqs including the embedded req has been
931 * disposed, schedule request buffer for re-use.
933 LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) ==
935 list_add_tail(&rqbd->rqbd_list,
936 &svcpt->scp_rqbd_idle);
939 spin_unlock(&svcpt->scp_lock);
940 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
941 /* If we are low on memory, we are not interested in history */
942 list_del(&req->rq_list);
943 list_del_init(&req->rq_history_list);
945 /* Track the highest culled req seq */
946 if (req->rq_history_seq > svcpt->scp_hist_seq_culled)
947 svcpt->scp_hist_seq_culled = req->rq_history_seq;
949 spin_unlock(&svcpt->scp_lock);
951 ptlrpc_server_free_request(req);
953 spin_unlock(&svcpt->scp_lock);
957 /** Change request export and move hp request from old export to new */
958 void ptlrpc_request_change_export(struct ptlrpc_request *req,
959 struct obd_export *export)
961 if (req->rq_export != NULL) {
962 if (!list_empty(&req->rq_exp_list)) {
963 /* remove rq_exp_list from last export */
964 spin_lock_bh(&req->rq_export->exp_rpc_lock);
965 list_del_init(&req->rq_exp_list);
966 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
968 /* export has one reference already, so it`s safe to
969 * add req to export queue here and get another
970 * reference for request later */
971 spin_lock_bh(&export->exp_rpc_lock);
972 list_add(&req->rq_exp_list, &export->exp_hp_rpcs);
973 spin_unlock_bh(&export->exp_rpc_lock);
975 class_export_rpc_dec(req->rq_export);
976 class_export_put(req->rq_export);
979 /* request takes one export refcount */
980 req->rq_export = class_export_get(export);
981 class_export_rpc_inc(export);
987 * to finish a request: stop sending more early replies, and release
990 static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt,
991 struct ptlrpc_request *req)
993 ptlrpc_server_hpreq_fini(req);
995 ptlrpc_server_drop_request(req);
999 * to finish a active request: stop sending more early replies, and release
1000 * the request. should be called after we finished handling the request.
1002 static void ptlrpc_server_finish_active_request(
1003 struct ptlrpc_service_part *svcpt,
1004 struct ptlrpc_request *req)
1006 spin_lock(&svcpt->scp_req_lock);
1007 ptlrpc_nrs_req_stop_nolock(req);
1008 svcpt->scp_nreqs_active--;
1010 svcpt->scp_nhreqs_active--;
1011 spin_unlock(&svcpt->scp_req_lock);
1013 ptlrpc_nrs_req_finalize(req);
1015 if (req->rq_export != NULL)
1016 class_export_rpc_dec(req->rq_export);
1018 ptlrpc_server_finish_request(svcpt, req);
1022 * This function makes sure dead exports are evicted in a timely manner.
1023 * This function is only called when some export receives a message (i.e.,
1024 * the network is up.)
1026 static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
1028 struct obd_export *oldest_exp;
1029 time_t oldest_time, new_time;
1033 /* Compensate for slow machines, etc, by faking our request time
1034 into the future. Although this can break the strict time-ordering
1035 of the list, we can be really lazy here - we don't have to evict
1036 at the exact right moment. Eventually, all silent exports
1037 will make it to the top of the list. */
1039 /* Do not pay attention on 1sec or smaller renewals. */
1040 new_time = get_seconds() + extra_delay;
1041 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
1044 exp->exp_last_request_time = new_time;
1046 /* exports may get disconnected from the chain even though the
1047 export has references, so we must keep the spin lock while
1048 manipulating the lists */
1049 spin_lock(&exp->exp_obd->obd_dev_lock);
1051 if (list_empty(&exp->exp_obd_chain_timed)) {
1052 /* this one is not timed */
1053 spin_unlock(&exp->exp_obd->obd_dev_lock);
1057 list_move_tail(&exp->exp_obd_chain_timed,
1058 &exp->exp_obd->obd_exports_timed);
1060 oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
1061 struct obd_export, exp_obd_chain_timed);
1062 oldest_time = oldest_exp->exp_last_request_time;
1063 spin_unlock(&exp->exp_obd->obd_dev_lock);
1065 if (exp->exp_obd->obd_recovering) {
1066 /* be nice to everyone during recovery */
1070 /* Note - racing to start/reset the obd_eviction timer is safe */
1071 if (exp->exp_obd->obd_eviction_timer == 0) {
1072 /* Check if the oldest entry is expired. */
1073 if (get_seconds() > (oldest_time + PING_EVICT_TIMEOUT +
1075 /* We need a second timer, in case the net was down and
1076 * it just came back. Since the pinger may skip every
1077 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
1078 * we better wait for 3. */
1079 exp->exp_obd->obd_eviction_timer =
1080 get_seconds() + 3 * PING_INTERVAL;
1081 CDEBUG(D_HA, "%s: Think about evicting %s from "CFS_TIME_T"\n",
1082 exp->exp_obd->obd_name,
1083 obd_export_nid2str(oldest_exp), oldest_time);
1087 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
1088 /* The evictor won't evict anyone who we've heard from
1089 * recently, so we don't have to check before we start
1091 if (!ping_evictor_wake(exp))
1092 exp->exp_obd->obd_eviction_timer = 0;
1098 * Sanity check request \a req.
1099 * Return 0 if all is ok, error code otherwise.
1101 static int ptlrpc_check_req(struct ptlrpc_request *req)
1103 struct obd_device *obd = req->rq_export->exp_obd;
1106 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
1107 req->rq_export->exp_conn_cnt)) {
1108 DEBUG_REQ(D_RPCTRACE, req,
1109 "DROPPING req from old connection %d < %d",
1110 lustre_msg_get_conn_cnt(req->rq_reqmsg),
1111 req->rq_export->exp_conn_cnt);
1114 if (unlikely(obd == NULL || obd->obd_fail)) {
1116 * Failing over, don't handle any more reqs, send
1117 * error response instead.
1119 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
1120 req, (obd != NULL) ? obd->obd_name : "unknown");
1122 } else if (lustre_msg_get_flags(req->rq_reqmsg) &
1123 (MSG_REPLAY | MSG_REQ_REPLAY_DONE) &&
1124 !obd->obd_recovering) {
1125 DEBUG_REQ(D_ERROR, req,
1126 "Invalid replay without recovery");
1127 class_fail_export(req->rq_export);
1129 } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 &&
1130 !obd->obd_recovering) {
1131 DEBUG_REQ(D_ERROR, req, "Invalid req with transno %llu without recovery",
1132 lustre_msg_get_transno(req->rq_reqmsg));
1133 class_fail_export(req->rq_export);
1137 if (unlikely(rc < 0)) {
1138 req->rq_status = rc;
1144 static void ptlrpc_at_set_timer(struct ptlrpc_service_part *svcpt)
1146 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1149 if (array->paa_count == 0) {
1150 cfs_timer_disarm(&svcpt->scp_at_timer);
1154 /* Set timer for closest deadline */
1155 next = (__s32)(array->paa_deadline - get_seconds() -
1158 ptlrpc_at_timer((unsigned long)svcpt);
1160 cfs_timer_arm(&svcpt->scp_at_timer, cfs_time_shift(next));
1161 CDEBUG(D_INFO, "armed %s at %+ds\n",
1162 svcpt->scp_service->srv_name, next);
1166 /* Add rpc to early reply check list */
1167 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
1169 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1170 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1171 struct ptlrpc_request *rq = NULL;
1177 if (req->rq_no_reply)
1180 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
1183 spin_lock(&svcpt->scp_at_lock);
1184 LASSERT(list_empty(&req->rq_timed_list));
1186 index = (unsigned long)req->rq_deadline % array->paa_size;
1187 if (array->paa_reqs_count[index] > 0) {
1188 /* latest rpcs will have the latest deadlines in the list,
1189 * so search backward. */
1190 list_for_each_entry_reverse(rq,
1191 &array->paa_reqs_array[index],
1193 if (req->rq_deadline >= rq->rq_deadline) {
1194 list_add(&req->rq_timed_list,
1195 &rq->rq_timed_list);
1201 /* Add the request at the head of the list */
1202 if (list_empty(&req->rq_timed_list))
1203 list_add(&req->rq_timed_list,
1204 &array->paa_reqs_array[index]);
1206 spin_lock(&req->rq_lock);
1207 req->rq_at_linked = 1;
1208 spin_unlock(&req->rq_lock);
1209 req->rq_at_index = index;
1210 array->paa_reqs_count[index]++;
1212 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
1213 array->paa_deadline = req->rq_deadline;
1214 ptlrpc_at_set_timer(svcpt);
1216 spin_unlock(&svcpt->scp_at_lock);
1222 ptlrpc_at_remove_timed(struct ptlrpc_request *req)
1224 struct ptlrpc_at_array *array;
1226 array = &req->rq_rqbd->rqbd_svcpt->scp_at_array;
1228 /* NB: must call with hold svcpt::scp_at_lock */
1229 LASSERT(!list_empty(&req->rq_timed_list));
1230 list_del_init(&req->rq_timed_list);
1232 spin_lock(&req->rq_lock);
1233 req->rq_at_linked = 0;
1234 spin_unlock(&req->rq_lock);
1236 array->paa_reqs_count[req->rq_at_index]--;
1240 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
1242 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1243 struct ptlrpc_request *reqcopy;
1244 struct lustre_msg *reqmsg;
1245 long olddl = req->rq_deadline - get_seconds();
1249 /* deadline is when the client expects us to reply, margin is the
1250 difference between clients' and servers' expectations */
1251 DEBUG_REQ(D_ADAPTTO, req,
1252 "%ssending early reply (deadline %+lds, margin %+lds) for %d+%d",
1253 AT_OFF ? "AT off - not " : "",
1254 olddl, olddl - at_get(&svcpt->scp_at_estimate),
1255 at_get(&svcpt->scp_at_estimate), at_extra);
1261 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), not sending early reply. Consider increasing at_early_margin (%d)?",
1262 olddl, at_early_margin);
1264 /* Return an error so we're not re-added to the timed list. */
1268 if (!(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
1269 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, but no AT support");
1273 if (req->rq_export &&
1274 lustre_msg_get_flags(req->rq_reqmsg) &
1275 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1276 /* During recovery, we don't want to send too many early
1277 * replies, but on the other hand we want to make sure the
1278 * client has enough time to resend if the rpc is lost. So
1279 * during the recovery period send at least 4 early replies,
1280 * spacing them every at_extra if we can. at_estimate should
1281 * always equal this fixed value during recovery. */
1282 at_measured(&svcpt->scp_at_estimate, min(at_extra,
1283 req->rq_export->exp_obd->obd_recovery_timeout / 4));
1285 /* Fake our processing time into the future to ask the clients
1286 * for some extra amount of time */
1287 at_measured(&svcpt->scp_at_estimate, at_extra +
1289 req->rq_arrival_time.tv_sec);
1291 /* Check to see if we've actually increased the deadline -
1292 * we may be past adaptive_max */
1293 if (req->rq_deadline >= req->rq_arrival_time.tv_sec +
1294 at_get(&svcpt->scp_at_estimate)) {
1295 DEBUG_REQ(D_WARNING, req, "Couldn't add any time (%ld/%ld), not sending early reply\n",
1296 olddl, req->rq_arrival_time.tv_sec +
1297 at_get(&svcpt->scp_at_estimate) -
1302 newdl = get_seconds() + at_get(&svcpt->scp_at_estimate);
1304 reqcopy = ptlrpc_request_cache_alloc(GFP_NOFS);
1305 if (reqcopy == NULL)
1307 OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1314 reqcopy->rq_reply_state = NULL;
1315 reqcopy->rq_rep_swab_mask = 0;
1316 reqcopy->rq_pack_bulk = 0;
1317 reqcopy->rq_pack_udesc = 0;
1318 reqcopy->rq_packed_final = 0;
1319 sptlrpc_svc_ctx_addref(reqcopy);
1320 /* We only need the reqmsg for the magic */
1321 reqcopy->rq_reqmsg = reqmsg;
1322 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1324 LASSERT(atomic_read(&req->rq_refcount));
1325 /** if it is last refcount then early reply isn't needed */
1326 if (atomic_read(&req->rq_refcount) == 1) {
1327 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, abort sending early reply\n");
1332 /* Connection ref */
1333 reqcopy->rq_export = class_conn2export(
1334 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1335 if (reqcopy->rq_export == NULL) {
1341 class_export_rpc_inc(reqcopy->rq_export);
1342 if (reqcopy->rq_export->exp_obd &&
1343 reqcopy->rq_export->exp_obd->obd_fail) {
1348 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1352 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1355 /* Adjust our own deadline to what we told the client */
1356 req->rq_deadline = newdl;
1357 req->rq_early_count++; /* number sent, server side */
1359 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1362 /* Free the (early) reply state from lustre_pack_reply.
1363 (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
1364 ptlrpc_req_drop_rs(reqcopy);
1367 class_export_rpc_dec(reqcopy->rq_export);
1368 class_export_put(reqcopy->rq_export);
1370 sptlrpc_svc_ctx_decref(reqcopy);
1371 OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1373 ptlrpc_request_cache_free(reqcopy);
1377 /* Send early replies to everybody expiring within at_early_margin
1378 asking for at_extra time */
1379 static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1381 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1382 struct ptlrpc_request *rq, *n;
1383 struct list_head work_list;
1386 time_t now = get_seconds();
1388 int first, counter = 0;
1390 spin_lock(&svcpt->scp_at_lock);
1391 if (svcpt->scp_at_check == 0) {
1392 spin_unlock(&svcpt->scp_at_lock);
1395 delay = cfs_time_sub(cfs_time_current(), svcpt->scp_at_checktime);
1396 svcpt->scp_at_check = 0;
1398 if (array->paa_count == 0) {
1399 spin_unlock(&svcpt->scp_at_lock);
1403 /* The timer went off, but maybe the nearest rpc already completed. */
1404 first = array->paa_deadline - now;
1405 if (first > at_early_margin) {
1406 /* We've still got plenty of time. Reset the timer. */
1407 ptlrpc_at_set_timer(svcpt);
1408 spin_unlock(&svcpt->scp_at_lock);
1412 /* We're close to a timeout, and we don't know how much longer the
1413 server will take. Send early replies to everyone expiring soon. */
1414 INIT_LIST_HEAD(&work_list);
1416 index = (unsigned long)array->paa_deadline % array->paa_size;
1417 count = array->paa_count;
1419 count -= array->paa_reqs_count[index];
1420 list_for_each_entry_safe(rq, n,
1421 &array->paa_reqs_array[index],
1423 if (rq->rq_deadline > now + at_early_margin) {
1424 /* update the earliest deadline */
1425 if (deadline == -1 ||
1426 rq->rq_deadline < deadline)
1427 deadline = rq->rq_deadline;
1431 ptlrpc_at_remove_timed(rq);
1433 * ptlrpc_server_drop_request() may drop
1434 * refcount to 0 already. Let's check this and
1435 * don't add entry to work_list
1437 if (likely(atomic_inc_not_zero(&rq->rq_refcount)))
1438 list_add(&rq->rq_timed_list, &work_list);
1442 if (++index >= array->paa_size)
1445 array->paa_deadline = deadline;
1446 /* we have a new earliest deadline, restart the timer */
1447 ptlrpc_at_set_timer(svcpt);
1449 spin_unlock(&svcpt->scp_at_lock);
1451 CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early replies\n",
1452 first, at_extra, counter);
1454 /* We're already past request deadlines before we even get a
1455 chance to send early replies */
1456 LCONSOLE_WARN("%s: This server is not able to keep up with request traffic (cpu-bound).\n",
1457 svcpt->scp_service->srv_name);
1458 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, delay=" CFS_DURATION_T "(jiff)\n",
1459 counter, svcpt->scp_nreqs_incoming,
1460 svcpt->scp_nreqs_active,
1461 at_get(&svcpt->scp_at_estimate), delay);
1464 /* we took additional refcount so entries can't be deleted from list, no
1465 * locking is needed */
1466 while (!list_empty(&work_list)) {
1467 rq = list_entry(work_list.next, struct ptlrpc_request,
1469 list_del_init(&rq->rq_timed_list);
1471 if (ptlrpc_at_send_early_reply(rq) == 0)
1472 ptlrpc_at_add_timed(rq);
1474 ptlrpc_server_drop_request(rq);
1477 return 1; /* return "did_something" for liblustre */
1481 * Put the request to the export list if the request may become
1482 * a high priority one.
1484 static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt,
1485 struct ptlrpc_request *req)
1489 if (svcpt->scp_service->srv_ops.so_hpreq_handler) {
1490 rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
1495 if (req->rq_export && req->rq_ops) {
1496 /* Perform request specific check. We should do this check
1497 * before the request is added into exp_hp_rpcs list otherwise
1498 * it may hit swab race at LU-1044. */
1499 if (req->rq_ops->hpreq_check) {
1500 rc = req->rq_ops->hpreq_check(req);
1502 * XXX: Out of all current
1503 * ptlrpc_hpreq_ops::hpreq_check(), only
1504 * ldlm_cancel_hpreq_check() can return an error code;
1505 * other functions assert in similar places, which seems
1506 * odd. What also does not seem right is that handlers
1507 * for those RPCs do not assert on the same checks, but
1508 * rather handle the error cases. e.g. see
1509 * ost_rw_hpreq_check(), and ost_brw_read(),
1514 LASSERT(rc == 0 || rc == 1);
1517 spin_lock_bh(&req->rq_export->exp_rpc_lock);
1518 list_add(&req->rq_exp_list,
1519 &req->rq_export->exp_hp_rpcs);
1520 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1523 ptlrpc_nrs_req_initialize(svcpt, req, rc);
1528 /** Remove the request from the export list. */
1529 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req)
1531 if (req->rq_export && req->rq_ops) {
1532 /* refresh lock timeout again so that client has more
1533 * room to send lock cancel RPC. */
1534 if (req->rq_ops->hpreq_fini)
1535 req->rq_ops->hpreq_fini(req);
1537 spin_lock_bh(&req->rq_export->exp_rpc_lock);
1538 list_del_init(&req->rq_exp_list);
1539 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1543 static int ptlrpc_hpreq_check(struct ptlrpc_request *req)
1548 static struct ptlrpc_hpreq_ops ptlrpc_hpreq_common = {
1549 .hpreq_check = ptlrpc_hpreq_check,
1552 /* Hi-Priority RPC check by RPC operation code. */
1553 int ptlrpc_hpreq_handler(struct ptlrpc_request *req)
1555 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1557 /* Check for export to let only reconnects for not yet evicted
1558 * export to become a HP rpc. */
1559 if ((req->rq_export != NULL) &&
1560 (opc == OBD_PING || opc == MDS_CONNECT || opc == OST_CONNECT))
1561 req->rq_ops = &ptlrpc_hpreq_common;
1565 EXPORT_SYMBOL(ptlrpc_hpreq_handler);
1567 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1568 struct ptlrpc_request *req)
1572 rc = ptlrpc_server_hpreq_init(svcpt, req);
1576 ptlrpc_nrs_req_add(svcpt, req, !!rc);
1582 * Allow to handle high priority request
1583 * User can call it w/o any lock but need to hold
1584 * ptlrpc_service_part::scp_req_lock to get reliable result
1586 static bool ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1589 int running = svcpt->scp_nthrs_running;
1591 if (!nrs_svcpt_has_hp(svcpt))
1597 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1598 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1599 /* leave just 1 thread for normal RPCs */
1600 running = PTLRPC_NTHRS_INIT;
1601 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1605 if (svcpt->scp_nreqs_active >= running - 1)
1608 if (svcpt->scp_nhreqs_active == 0)
1611 return !ptlrpc_nrs_req_pending_nolock(svcpt, false) ||
1612 svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1615 static bool ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1618 return ptlrpc_server_allow_high(svcpt, force) &&
1619 ptlrpc_nrs_req_pending_nolock(svcpt, true);
1623 * Only allow normal priority requests on a service that has a high-priority
1624 * queue if forced (i.e. cleanup), if there are other high priority requests
1625 * already being processed (i.e. those threads can service more high-priority
1626 * requests), or if there are enough idle threads that a later thread can do
1627 * a high priority request.
1628 * User can call it w/o any lock but need to hold
1629 * ptlrpc_service_part::scp_req_lock to get reliable result
1631 static bool ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1634 int running = svcpt->scp_nthrs_running;
1635 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1636 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1637 /* leave just 1 thread for normal RPCs */
1638 running = PTLRPC_NTHRS_INIT;
1639 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1644 svcpt->scp_nreqs_active < running - 2)
1647 if (svcpt->scp_nreqs_active >= running - 1)
1650 return svcpt->scp_nhreqs_active > 0 || !nrs_svcpt_has_hp(svcpt);
1653 static bool ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1656 return ptlrpc_server_allow_normal(svcpt, force) &&
1657 ptlrpc_nrs_req_pending_nolock(svcpt, false);
1661 * Returns true if there are requests available in incoming
1662 * request queue for processing and it is allowed to fetch them.
1663 * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1664 * to get reliable result
1665 * \see ptlrpc_server_allow_normal
1666 * \see ptlrpc_server_allow high
1669 ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt, bool force)
1671 return ptlrpc_server_high_pending(svcpt, force) ||
1672 ptlrpc_server_normal_pending(svcpt, force);
1676 * Fetch a request for processing from queue of unprocessed requests.
1677 * Favors high-priority requests.
1678 * Returns a pointer to fetched request.
1680 static struct ptlrpc_request *
1681 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force)
1683 struct ptlrpc_request *req = NULL;
1685 spin_lock(&svcpt->scp_req_lock);
1687 if (ptlrpc_server_high_pending(svcpt, force)) {
1688 req = ptlrpc_nrs_req_get_nolock(svcpt, true, force);
1690 svcpt->scp_hreq_count++;
1695 if (ptlrpc_server_normal_pending(svcpt, force)) {
1696 req = ptlrpc_nrs_req_get_nolock(svcpt, false, force);
1698 svcpt->scp_hreq_count = 0;
1703 spin_unlock(&svcpt->scp_req_lock);
1707 svcpt->scp_nreqs_active++;
1709 svcpt->scp_nhreqs_active++;
1711 spin_unlock(&svcpt->scp_req_lock);
1713 if (likely(req->rq_export))
1714 class_export_rpc_inc(req->rq_export);
1720 * Handle freshly incoming reqs, add to timed early reply list,
1721 * pass on to regular request queue.
1722 * All incoming requests pass through here before getting into
1723 * ptlrpc_server_handle_req later on.
1726 ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt,
1727 struct ptlrpc_thread *thread)
1729 struct ptlrpc_service *svc = svcpt->scp_service;
1730 struct ptlrpc_request *req;
1734 spin_lock(&svcpt->scp_lock);
1735 if (list_empty(&svcpt->scp_req_incoming)) {
1736 spin_unlock(&svcpt->scp_lock);
1740 req = list_entry(svcpt->scp_req_incoming.next,
1741 struct ptlrpc_request, rq_list);
1742 list_del_init(&req->rq_list);
1743 svcpt->scp_nreqs_incoming--;
1744 /* Consider this still a "queued" request as far as stats are
1746 spin_unlock(&svcpt->scp_lock);
1748 /* go through security check/transform */
1749 rc = sptlrpc_svc_unwrap_request(req);
1753 case SECSVC_COMPLETE:
1754 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1763 * for null-flavored rpc, msg has been unpacked by sptlrpc, although
1764 * redo it wouldn't be harmful.
1766 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
1767 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
1769 CERROR("error unpacking request: ptl %d from %s x%llu\n",
1770 svc->srv_req_portal, libcfs_id2str(req->rq_peer),
1776 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1778 CERROR("error unpacking ptlrpc body: ptl %d from %s x%llu\n",
1779 svc->srv_req_portal, libcfs_id2str(req->rq_peer),
1784 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
1785 lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
1786 CERROR("drop incoming rpc opc %u, x%llu\n",
1787 cfs_fail_val, req->rq_xid);
1792 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1793 CERROR("wrong packet type received (type=%u) from %s\n",
1794 lustre_msg_get_type(req->rq_reqmsg),
1795 libcfs_id2str(req->rq_peer));
1799 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1802 req->rq_bulk_write = 1;
1806 case MGS_CONFIG_READ:
1807 req->rq_bulk_read = 1;
1811 CDEBUG(D_RPCTRACE, "got req x%llu\n", req->rq_xid);
1813 req->rq_export = class_conn2export(
1814 lustre_msg_get_handle(req->rq_reqmsg));
1815 if (req->rq_export) {
1816 rc = ptlrpc_check_req(req);
1818 rc = sptlrpc_target_export_check(req->rq_export, req);
1820 DEBUG_REQ(D_ERROR, req, "DROPPING req with illegal security flavor,");
1825 ptlrpc_update_export_timer(req->rq_export, 0);
1828 /* req_in handling should/must be fast */
1829 if (get_seconds() - req->rq_arrival_time.tv_sec > 5)
1830 DEBUG_REQ(D_WARNING, req, "Slow req_in handling "CFS_DURATION_T"s",
1831 cfs_time_sub(get_seconds(),
1832 req->rq_arrival_time.tv_sec));
1834 /* Set rpc server deadline and add it to the timed list */
1835 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1836 MSGHDR_AT_SUPPORT) ?
1837 /* The max time the client expects us to take */
1838 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1839 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1840 if (unlikely(deadline == 0)) {
1841 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1845 req->rq_svc_thread = thread;
1847 ptlrpc_at_add_timed(req);
1849 /* Move it over to the request processing queue */
1850 rc = ptlrpc_server_request_add(svcpt, req);
1854 wake_up(&svcpt->scp_waitq);
1858 ptlrpc_server_finish_request(svcpt, req);
1864 * Main incoming request handling logic.
1865 * Calls handler function from service to do actual processing.
1868 ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
1869 struct ptlrpc_thread *thread)
1871 struct ptlrpc_service *svc = svcpt->scp_service;
1872 struct ptlrpc_request *request;
1873 struct timeval work_start;
1874 struct timeval work_end;
1879 request = ptlrpc_server_request_get(svcpt, false);
1880 if (request == NULL)
1883 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
1884 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
1885 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
1886 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
1888 if (unlikely(fail_opc)) {
1889 if (request->rq_export && request->rq_ops)
1890 OBD_FAIL_TIMEOUT(fail_opc, 4);
1893 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
1895 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
1896 libcfs_debug_dumplog();
1898 do_gettimeofday(&work_start);
1899 timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,
1901 if (likely(svc->srv_stats != NULL)) {
1902 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
1904 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
1905 svcpt->scp_nreqs_incoming);
1906 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
1907 svcpt->scp_nreqs_active);
1908 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
1909 at_get(&svcpt->scp_at_estimate));
1912 rc = lu_context_init(&request->rq_session, LCT_SESSION | LCT_NOREF);
1914 CERROR("Failure to initialize session: %d\n", rc);
1917 request->rq_session.lc_thread = thread;
1918 request->rq_session.lc_cookie = 0x5;
1919 lu_context_enter(&request->rq_session);
1921 CDEBUG(D_NET, "got req %llu\n", request->rq_xid);
1923 request->rq_svc_thread = thread;
1925 request->rq_svc_thread->t_env->le_ses = &request->rq_session;
1927 if (likely(request->rq_export)) {
1928 if (unlikely(ptlrpc_check_req(request)))
1930 ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
1933 /* Discard requests queued for longer than the deadline.
1934 The deadline is increased if we send an early reply. */
1935 if (get_seconds() > request->rq_deadline) {
1936 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s: deadline " CFS_DURATION_T ":" CFS_DURATION_T "s ago\n",
1937 libcfs_id2str(request->rq_peer),
1938 cfs_time_sub(request->rq_deadline,
1939 request->rq_arrival_time.tv_sec),
1940 cfs_time_sub(get_seconds(),
1941 request->rq_deadline));
1945 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc %s:%s+%d:%d:x%llu:%s:%d\n",
1947 (request->rq_export ?
1948 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1949 (request->rq_export ?
1950 atomic_read(&request->rq_export->exp_refcount) : -99),
1951 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1952 libcfs_id2str(request->rq_peer),
1953 lustre_msg_get_opc(request->rq_reqmsg));
1955 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
1956 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
1958 rc = svc->srv_ops.so_req_handler(request);
1960 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
1963 lu_context_exit(&request->rq_session);
1964 lu_context_fini(&request->rq_session);
1966 if (unlikely(get_seconds() > request->rq_deadline)) {
1967 DEBUG_REQ(D_WARNING, request,
1968 "Request took longer than estimated ("
1969 CFS_DURATION_T":"CFS_DURATION_T
1970 "s); client may timeout.",
1971 cfs_time_sub(request->rq_deadline,
1972 request->rq_arrival_time.tv_sec),
1973 cfs_time_sub(get_seconds(),
1974 request->rq_deadline));
1977 do_gettimeofday(&work_end);
1978 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1979 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc %s:%s+%d:%d:x%llu:%s:%d Request processed in %ldus (%ldus total) trans %llu rc %d/%d\n",
1981 (request->rq_export ?
1982 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1983 (request->rq_export ?
1984 atomic_read(&request->rq_export->exp_refcount) : -99),
1985 lustre_msg_get_status(request->rq_reqmsg),
1987 libcfs_id2str(request->rq_peer),
1988 lustre_msg_get_opc(request->rq_reqmsg),
1990 cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL),
1991 (request->rq_repmsg ?
1992 lustre_msg_get_transno(request->rq_repmsg) :
1993 request->rq_transno),
1995 (request->rq_repmsg ?
1996 lustre_msg_get_status(request->rq_repmsg) : -999));
1997 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
1998 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
1999 int opc = opcode_offset(op);
2000 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
2001 LASSERT(opc < LUSTRE_MAX_OPCODES);
2002 lprocfs_counter_add(svc->srv_stats,
2003 opc + EXTRA_MAX_OPCODES,
2007 if (unlikely(request->rq_early_count)) {
2008 DEBUG_REQ(D_ADAPTTO, request,
2009 "sent %d early replies before finishing in "
2011 request->rq_early_count,
2012 cfs_time_sub(work_end.tv_sec,
2013 request->rq_arrival_time.tv_sec));
2017 ptlrpc_server_finish_active_request(svcpt, request);
2023 * An internal function to process a single reply state object.
2026 ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
2028 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
2029 struct ptlrpc_service *svc = svcpt->scp_service;
2030 struct obd_export *exp;
2034 exp = rs->rs_export;
2036 LASSERT(rs->rs_difficult);
2037 LASSERT(rs->rs_scheduled);
2038 LASSERT(list_empty(&rs->rs_list));
2040 spin_lock(&exp->exp_lock);
2041 /* Noop if removed already */
2042 list_del_init(&rs->rs_exp_list);
2043 spin_unlock(&exp->exp_lock);
2045 /* The disk commit callback holds exp_uncommitted_replies_lock while it
2046 * iterates over newly committed replies, removing them from
2047 * exp_uncommitted_replies. It then drops this lock and schedules the
2048 * replies it found for handling here.
2050 * We can avoid contention for exp_uncommitted_replies_lock between the
2051 * HRT threads and further commit callbacks by checking rs_committed
2052 * which is set in the commit callback while it holds both
2053 * rs_lock and exp_uncommitted_reples.
2055 * If we see rs_committed clear, the commit callback _may_ not have
2056 * handled this reply yet and we race with it to grab
2057 * exp_uncommitted_replies_lock before removing the reply from
2058 * exp_uncommitted_replies. Note that if we lose the race and the
2059 * reply has already been removed, list_del_init() is a noop.
2061 * If we see rs_committed set, we know the commit callback is handling,
2062 * or has handled this reply since store reordering might allow us to
2063 * see rs_committed set out of sequence. But since this is done
2064 * holding rs_lock, we can be sure it has all completed once we hold
2065 * rs_lock, which we do right next.
2067 if (!rs->rs_committed) {
2068 spin_lock(&exp->exp_uncommitted_replies_lock);
2069 list_del_init(&rs->rs_obd_list);
2070 spin_unlock(&exp->exp_uncommitted_replies_lock);
2073 spin_lock(&rs->rs_lock);
2075 been_handled = rs->rs_handled;
2078 nlocks = rs->rs_nlocks; /* atomic "steal", but */
2079 rs->rs_nlocks = 0; /* locks still on rs_locks! */
2081 if (nlocks == 0 && !been_handled) {
2082 /* If we see this, we should already have seen the warning
2083 * in mds_steal_ack_locks() */
2084 CDEBUG(D_HA, "All locks stolen from rs %p x%lld.t%lld o%d NID %s\n",
2086 rs->rs_xid, rs->rs_transno, rs->rs_opc,
2087 libcfs_nid2str(exp->exp_connection->c_peer.nid));
2090 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
2091 spin_unlock(&rs->rs_lock);
2093 if (!been_handled && rs->rs_on_net) {
2094 LNetMDUnlink(rs->rs_md_h);
2095 /* Ignore return code; we're racing with completion */
2098 while (nlocks-- > 0)
2099 ldlm_lock_decref(&rs->rs_locks[nlocks],
2100 rs->rs_modes[nlocks]);
2102 spin_lock(&rs->rs_lock);
2105 rs->rs_scheduled = 0;
2107 if (!rs->rs_on_net) {
2109 spin_unlock(&rs->rs_lock);
2111 class_export_put(exp);
2112 rs->rs_export = NULL;
2113 ptlrpc_rs_decref(rs);
2114 if (atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
2115 svc->srv_is_stopping)
2116 wake_up_all(&svcpt->scp_waitq);
2120 /* still on the net; callback will schedule */
2121 spin_unlock(&rs->rs_lock);
2127 ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
2129 int avail = svcpt->scp_nrqbds_posted;
2130 int low_water = test_req_buffer_pressure ? 0 :
2131 svcpt->scp_service->srv_nbuf_per_group / 2;
2133 /* NB I'm not locking; just looking. */
2135 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
2136 * allowed the request history to grow out of control. We could put a
2137 * sanity check on that here and cull some history if we need the
2140 if (avail <= low_water)
2141 ptlrpc_grow_req_bufs(svcpt, 1);
2143 if (svcpt->scp_service->srv_stats) {
2144 lprocfs_counter_add(svcpt->scp_service->srv_stats,
2145 PTLRPC_REQBUF_AVAIL_CNTR, avail);
2150 ptlrpc_retry_rqbds(void *arg)
2152 struct ptlrpc_service_part *svcpt = (struct ptlrpc_service_part *)arg;
2154 svcpt->scp_rqbd_timeout = 0;
2159 ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
2161 return svcpt->scp_nreqs_active <
2162 svcpt->scp_nthrs_running - 1 -
2163 (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL);
2167 * allowed to create more threads
2168 * user can call it w/o any lock but need to hold
2169 * ptlrpc_service_part::scp_lock to get reliable result
2172 ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
2174 return svcpt->scp_nthrs_running +
2175 svcpt->scp_nthrs_starting <
2176 svcpt->scp_service->srv_nthrs_cpt_limit;
2180 * too many requests and allowed to create more threads
2183 ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
2185 return !ptlrpc_threads_enough(svcpt) &&
2186 ptlrpc_threads_increasable(svcpt);
2190 ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2192 return thread_is_stopping(thread) ||
2193 thread->t_svcpt->scp_service->srv_is_stopping;
2197 ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
2199 return !list_empty(&svcpt->scp_rqbd_idle) &&
2200 svcpt->scp_rqbd_timeout == 0;
2204 ptlrpc_at_check(struct ptlrpc_service_part *svcpt)
2206 return svcpt->scp_at_check;
2210 * requests wait on preprocessing
2211 * user can call it w/o any lock but need to hold
2212 * ptlrpc_service_part::scp_lock to get reliable result
2215 ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
2217 return !list_empty(&svcpt->scp_req_incoming);
2220 static __attribute__((__noinline__)) int
2221 ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
2222 struct ptlrpc_thread *thread)
2224 /* Don't exit while there are replies to be handled */
2225 struct l_wait_info lwi = LWI_TIMEOUT(svcpt->scp_rqbd_timeout,
2226 ptlrpc_retry_rqbds, svcpt);
2228 /* XXX: Add this back when libcfs watchdog is merged upstream
2229 lc_watchdog_disable(thread->t_watchdog);
2234 l_wait_event_exclusive_head(svcpt->scp_waitq,
2235 ptlrpc_thread_stopping(thread) ||
2236 ptlrpc_server_request_incoming(svcpt) ||
2237 ptlrpc_server_request_pending(svcpt, false) ||
2238 ptlrpc_rqbd_pending(svcpt) ||
2239 ptlrpc_at_check(svcpt), &lwi);
2241 if (ptlrpc_thread_stopping(thread))
2245 lc_watchdog_touch(thread->t_watchdog,
2246 ptlrpc_server_get_timeout(svcpt));
2252 * Main thread body for service threads.
2253 * Waits in a loop waiting for new requests to process to appear.
2254 * Every time an incoming requests is added to its queue, a waitq
2255 * is woken up and one of the threads will handle it.
2257 static int ptlrpc_main(void *arg)
2259 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
2260 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2261 struct ptlrpc_service *svc = svcpt->scp_service;
2262 struct ptlrpc_reply_state *rs;
2263 struct group_info *ginfo = NULL;
2265 int counter = 0, rc = 0;
2267 thread->t_pid = current_pid();
2268 unshare_fs_struct();
2270 /* NB: we will call cfs_cpt_bind() for all threads, because we
2271 * might want to run lustre server only on a subset of system CPUs,
2272 * in that case ->scp_cpt is CFS_CPT_ANY */
2273 rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
2275 CWARN("%s: failed to bind %s on CPT %d\n",
2276 svc->srv_name, thread->t_name, svcpt->scp_cpt);
2279 ginfo = groups_alloc(0);
2285 set_current_groups(ginfo);
2286 put_group_info(ginfo);
2288 if (svc->srv_ops.so_thr_init != NULL) {
2289 rc = svc->srv_ops.so_thr_init(thread);
2300 rc = lu_context_init(&env->le_ctx,
2301 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2305 thread->t_env = env;
2306 env->le_ctx.lc_thread = thread;
2307 env->le_ctx.lc_cookie = 0x6;
2309 while (!list_empty(&svcpt->scp_rqbd_idle)) {
2310 rc = ptlrpc_server_post_idle_rqbds(svcpt);
2314 CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
2315 svc->srv_name, svcpt->scp_cpt, rc);
2319 /* Alloc reply state structure for this one */
2320 OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2326 spin_lock(&svcpt->scp_lock);
2328 LASSERT(thread_is_starting(thread));
2329 thread_clear_flags(thread, SVC_STARTING);
2331 LASSERT(svcpt->scp_nthrs_starting == 1);
2332 svcpt->scp_nthrs_starting--;
2334 /* SVC_STOPPING may already be set here if someone else is trying
2335 * to stop the service while this new thread has been dynamically
2336 * forked. We still set SVC_RUNNING to let our creator know that
2337 * we are now running, however we will exit as soon as possible */
2338 thread_add_flags(thread, SVC_RUNNING);
2339 svcpt->scp_nthrs_running++;
2340 spin_unlock(&svcpt->scp_lock);
2342 /* wake up our creator in case he's still waiting. */
2343 wake_up(&thread->t_ctl_waitq);
2346 thread->t_watchdog = lc_watchdog_add(ptlrpc_server_get_timeout(svcpt),
2350 spin_lock(&svcpt->scp_rep_lock);
2351 list_add(&rs->rs_list, &svcpt->scp_rep_idle);
2352 wake_up(&svcpt->scp_rep_waitq);
2353 spin_unlock(&svcpt->scp_rep_lock);
2355 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2356 svcpt->scp_nthrs_running);
2358 /* XXX maintain a list of all managed devices: insert here */
2359 while (!ptlrpc_thread_stopping(thread)) {
2360 if (ptlrpc_wait_event(svcpt, thread))
2363 ptlrpc_check_rqbd_pool(svcpt);
2365 if (ptlrpc_threads_need_create(svcpt)) {
2366 /* Ignore return code - we tried... */
2367 ptlrpc_start_thread(svcpt, 0);
2370 /* Process all incoming reqs before handling any */
2371 if (ptlrpc_server_request_incoming(svcpt)) {
2372 lu_context_enter(&env->le_ctx);
2374 ptlrpc_server_handle_req_in(svcpt, thread);
2375 lu_context_exit(&env->le_ctx);
2377 /* but limit ourselves in case of flood */
2378 if (counter++ < 100)
2383 if (ptlrpc_at_check(svcpt))
2384 ptlrpc_at_check_timed(svcpt);
2386 if (ptlrpc_server_request_pending(svcpt, false)) {
2387 lu_context_enter(&env->le_ctx);
2388 ptlrpc_server_handle_request(svcpt, thread);
2389 lu_context_exit(&env->le_ctx);
2392 if (ptlrpc_rqbd_pending(svcpt) &&
2393 ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
2394 /* I just failed to repost request buffers.
2395 * Wait for a timeout (unless something else
2396 * happens) before I try again */
2397 svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
2398 CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
2399 svcpt->scp_nrqbds_posted);
2404 lc_watchdog_delete(thread->t_watchdog);
2405 thread->t_watchdog = NULL;
2410 * deconstruct service specific state created by ptlrpc_start_thread()
2412 if (svc->srv_ops.so_thr_done != NULL)
2413 svc->srv_ops.so_thr_done(thread);
2416 lu_context_fini(&env->le_ctx);
2420 CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n",
2421 thread, thread->t_pid, thread->t_id, rc);
2423 spin_lock(&svcpt->scp_lock);
2424 if (thread_test_and_clear_flags(thread, SVC_STARTING))
2425 svcpt->scp_nthrs_starting--;
2427 if (thread_test_and_clear_flags(thread, SVC_RUNNING)) {
2428 /* must know immediately */
2429 svcpt->scp_nthrs_running--;
2433 thread_add_flags(thread, SVC_STOPPED);
2435 wake_up(&thread->t_ctl_waitq);
2436 spin_unlock(&svcpt->scp_lock);
2441 static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
2442 struct list_head *replies)
2446 spin_lock(&hrt->hrt_lock);
2448 list_splice_init(&hrt->hrt_queue, replies);
2449 result = ptlrpc_hr.hr_stopping || !list_empty(replies);
2451 spin_unlock(&hrt->hrt_lock);
2456 * Main body of "handle reply" function.
2457 * It processes acked reply states
2459 static int ptlrpc_hr_main(void *arg)
2461 struct ptlrpc_hr_thread *hrt = (struct ptlrpc_hr_thread *)arg;
2462 struct ptlrpc_hr_partition *hrp = hrt->hrt_partition;
2463 LIST_HEAD (replies);
2464 char threadname[20];
2467 snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
2468 hrp->hrp_cpt, hrt->hrt_id);
2469 unshare_fs_struct();
2471 rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
2473 CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
2474 threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
2477 atomic_inc(&hrp->hrp_nstarted);
2478 wake_up(&ptlrpc_hr.hr_waitq);
2480 while (!ptlrpc_hr.hr_stopping) {
2481 l_wait_condition(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
2483 while (!list_empty(&replies)) {
2484 struct ptlrpc_reply_state *rs;
2486 rs = list_entry(replies.prev,
2487 struct ptlrpc_reply_state,
2489 list_del_init(&rs->rs_list);
2490 ptlrpc_handle_rs(rs);
2494 atomic_inc(&hrp->hrp_nstopped);
2495 wake_up(&ptlrpc_hr.hr_waitq);
2500 static void ptlrpc_stop_hr_threads(void)
2502 struct ptlrpc_hr_partition *hrp;
2506 ptlrpc_hr.hr_stopping = 1;
2508 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2509 if (hrp->hrp_thrs == NULL)
2510 continue; /* uninitialized */
2511 for (j = 0; j < hrp->hrp_nthrs; j++)
2512 wake_up_all(&hrp->hrp_thrs[j].hrt_waitq);
2515 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2516 if (hrp->hrp_thrs == NULL)
2517 continue; /* uninitialized */
2518 wait_event(ptlrpc_hr.hr_waitq,
2519 atomic_read(&hrp->hrp_nstopped) ==
2520 atomic_read(&hrp->hrp_nstarted));
2524 static int ptlrpc_start_hr_threads(void)
2526 struct ptlrpc_hr_partition *hrp;
2530 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2533 for (j = 0; j < hrp->hrp_nthrs; j++) {
2534 struct ptlrpc_hr_thread *hrt = &hrp->hrp_thrs[j];
2535 rc = PTR_ERR(kthread_run(ptlrpc_hr_main,
2537 "ptlrpc_hr%02d_%03d",
2540 if (IS_ERR_VALUE(rc))
2543 wait_event(ptlrpc_hr.hr_waitq,
2544 atomic_read(&hrp->hrp_nstarted) == j);
2545 if (!IS_ERR_VALUE(rc))
2548 CERROR("Reply handling thread %d:%d Failed on starting: rc = %d\n",
2550 ptlrpc_stop_hr_threads();
2556 static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
2558 struct l_wait_info lwi = { 0 };
2559 struct ptlrpc_thread *thread;
2562 CDEBUG(D_INFO, "Stopping threads for service %s\n",
2563 svcpt->scp_service->srv_name);
2565 spin_lock(&svcpt->scp_lock);
2566 /* let the thread know that we would like it to stop asap */
2567 list_for_each_entry(thread, &svcpt->scp_threads, t_link) {
2568 CDEBUG(D_INFO, "Stopping thread %s #%u\n",
2569 svcpt->scp_service->srv_thread_name, thread->t_id);
2570 thread_add_flags(thread, SVC_STOPPING);
2573 wake_up_all(&svcpt->scp_waitq);
2575 while (!list_empty(&svcpt->scp_threads)) {
2576 thread = list_entry(svcpt->scp_threads.next,
2577 struct ptlrpc_thread, t_link);
2578 if (thread_is_stopped(thread)) {
2579 list_del(&thread->t_link);
2580 list_add(&thread->t_link, &zombie);
2583 spin_unlock(&svcpt->scp_lock);
2585 CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
2586 svcpt->scp_service->srv_thread_name, thread->t_id);
2587 l_wait_event(thread->t_ctl_waitq,
2588 thread_is_stopped(thread), &lwi);
2590 spin_lock(&svcpt->scp_lock);
2593 spin_unlock(&svcpt->scp_lock);
2595 while (!list_empty(&zombie)) {
2596 thread = list_entry(zombie.next,
2597 struct ptlrpc_thread, t_link);
2598 list_del(&thread->t_link);
2599 OBD_FREE_PTR(thread);
2604 * Stops all threads of a particular service \a svc
2606 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2608 struct ptlrpc_service_part *svcpt;
2611 ptlrpc_service_for_each_part(svcpt, i, svc) {
2612 if (svcpt->scp_service != NULL)
2613 ptlrpc_svcpt_stop_threads(svcpt);
2616 EXPORT_SYMBOL(ptlrpc_stop_all_threads);
2618 int ptlrpc_start_threads(struct ptlrpc_service *svc)
2624 /* We require 2 threads min, see note in ptlrpc_server_handle_request */
2625 LASSERT(svc->srv_nthrs_cpt_init >= PTLRPC_NTHRS_INIT);
2627 for (i = 0; i < svc->srv_ncpts; i++) {
2628 for (j = 0; j < svc->srv_nthrs_cpt_init; j++) {
2629 rc = ptlrpc_start_thread(svc->srv_parts[i], 1);
2635 /* We have enough threads, don't start more. b=15759 */
2642 CERROR("cannot start %s thread #%d_%d: rc %d\n",
2643 svc->srv_thread_name, i, j, rc);
2644 ptlrpc_stop_all_threads(svc);
2647 EXPORT_SYMBOL(ptlrpc_start_threads);
2649 int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
2651 struct l_wait_info lwi = { 0 };
2652 struct ptlrpc_thread *thread;
2653 struct ptlrpc_service *svc;
2656 LASSERT(svcpt != NULL);
2658 svc = svcpt->scp_service;
2660 CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n",
2661 svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running,
2662 svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit);
2665 if (unlikely(svc->srv_is_stopping))
2668 if (!ptlrpc_threads_increasable(svcpt) ||
2669 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
2670 svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1))
2673 OBD_CPT_ALLOC_PTR(thread, svc->srv_cptable, svcpt->scp_cpt);
2676 init_waitqueue_head(&thread->t_ctl_waitq);
2678 spin_lock(&svcpt->scp_lock);
2679 if (!ptlrpc_threads_increasable(svcpt)) {
2680 spin_unlock(&svcpt->scp_lock);
2681 OBD_FREE_PTR(thread);
2685 if (svcpt->scp_nthrs_starting != 0) {
2686 /* serialize starting because some modules (obdfilter)
2687 * might require unique and contiguous t_id */
2688 LASSERT(svcpt->scp_nthrs_starting == 1);
2689 spin_unlock(&svcpt->scp_lock);
2690 OBD_FREE_PTR(thread);
2692 CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n",
2693 svc->srv_thread_name, svcpt->scp_thr_nextid);
2698 CDEBUG(D_INFO, "Creating thread %s #%d race, retry later\n",
2699 svc->srv_thread_name, svcpt->scp_thr_nextid);
2703 svcpt->scp_nthrs_starting++;
2704 thread->t_id = svcpt->scp_thr_nextid++;
2705 thread_add_flags(thread, SVC_STARTING);
2706 thread->t_svcpt = svcpt;
2708 list_add(&thread->t_link, &svcpt->scp_threads);
2709 spin_unlock(&svcpt->scp_lock);
2711 if (svcpt->scp_cpt >= 0) {
2712 snprintf(thread->t_name, sizeof(thread->t_name), "%s%02d_%03d",
2713 svc->srv_thread_name, svcpt->scp_cpt, thread->t_id);
2715 snprintf(thread->t_name, sizeof(thread->t_name), "%s_%04d",
2716 svc->srv_thread_name, thread->t_id);
2719 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name);
2720 rc = PTR_ERR(kthread_run(ptlrpc_main, thread, "%s", thread->t_name));
2721 if (IS_ERR_VALUE(rc)) {
2722 CERROR("cannot start thread '%s': rc %d\n",
2723 thread->t_name, rc);
2724 spin_lock(&svcpt->scp_lock);
2725 --svcpt->scp_nthrs_starting;
2726 if (thread_is_stopping(thread)) {
2727 /* this ptlrpc_thread is being handled
2728 * by ptlrpc_svcpt_stop_threads now
2730 thread_add_flags(thread, SVC_STOPPED);
2731 wake_up(&thread->t_ctl_waitq);
2732 spin_unlock(&svcpt->scp_lock);
2734 list_del(&thread->t_link);
2735 spin_unlock(&svcpt->scp_lock);
2736 OBD_FREE_PTR(thread);
2744 l_wait_event(thread->t_ctl_waitq,
2745 thread_is_running(thread) || thread_is_stopped(thread),
2748 rc = thread_is_stopped(thread) ? thread->t_id : 0;
2752 int ptlrpc_hr_init(void)
2754 struct ptlrpc_hr_partition *hrp;
2755 struct ptlrpc_hr_thread *hrt;
2761 memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
2762 ptlrpc_hr.hr_cpt_table = cfs_cpt_table;
2764 ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
2766 if (ptlrpc_hr.hr_partitions == NULL)
2769 init_waitqueue_head(&ptlrpc_hr.hr_waitq);
2771 weight = cpumask_weight(topology_thread_cpumask(0));
2773 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2776 atomic_set(&hrp->hrp_nstarted, 0);
2777 atomic_set(&hrp->hrp_nstopped, 0);
2779 hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, i);
2780 hrp->hrp_nthrs /= weight;
2782 LASSERT(hrp->hrp_nthrs > 0);
2783 OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, i,
2784 hrp->hrp_nthrs * sizeof(*hrt));
2785 if (hrp->hrp_thrs == NULL) {
2790 for (j = 0; j < hrp->hrp_nthrs; j++) {
2791 hrt = &hrp->hrp_thrs[j];
2794 hrt->hrt_partition = hrp;
2795 init_waitqueue_head(&hrt->hrt_waitq);
2796 spin_lock_init(&hrt->hrt_lock);
2797 INIT_LIST_HEAD(&hrt->hrt_queue);
2801 rc = ptlrpc_start_hr_threads();
2808 void ptlrpc_hr_fini(void)
2810 struct ptlrpc_hr_partition *hrp;
2813 if (ptlrpc_hr.hr_partitions == NULL)
2816 ptlrpc_stop_hr_threads();
2818 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2819 if (hrp->hrp_thrs != NULL) {
2820 OBD_FREE(hrp->hrp_thrs,
2821 hrp->hrp_nthrs * sizeof(hrp->hrp_thrs[0]));
2825 cfs_percpt_free(ptlrpc_hr.hr_partitions);
2826 ptlrpc_hr.hr_partitions = NULL;
2831 * Wait until all already scheduled replies are processed.
2833 static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
2837 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
2840 rc = l_wait_event(svcpt->scp_waitq,
2841 atomic_read(&svcpt->scp_nreps_difficult) == 0, &lwi);
2844 CWARN("Unexpectedly long timeout %s %p\n",
2845 svcpt->scp_service->srv_name, svcpt->scp_service);
2850 ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
2852 struct ptlrpc_service_part *svcpt;
2855 /* early disarm AT timer... */
2856 ptlrpc_service_for_each_part(svcpt, i, svc) {
2857 if (svcpt->scp_service != NULL)
2858 cfs_timer_disarm(&svcpt->scp_at_timer);
2863 ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
2865 struct ptlrpc_service_part *svcpt;
2866 struct ptlrpc_request_buffer_desc *rqbd;
2867 struct l_wait_info lwi;
2871 /* All history will be culled when the next request buffer is
2872 * freed in ptlrpc_service_purge_all() */
2873 svc->srv_hist_nrqbds_cpt_max = 0;
2875 rc = LNetClearLazyPortal(svc->srv_req_portal);
2878 ptlrpc_service_for_each_part(svcpt, i, svc) {
2879 if (svcpt->scp_service == NULL)
2882 /* Unlink all the request buffers. This forces a 'final'
2883 * event with its 'unlink' flag set for each posted rqbd */
2884 list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
2886 rc = LNetMDUnlink(rqbd->rqbd_md_h);
2887 LASSERT(rc == 0 || rc == -ENOENT);
2891 ptlrpc_service_for_each_part(svcpt, i, svc) {
2892 if (svcpt->scp_service == NULL)
2895 /* Wait for the network to release any buffers
2896 * it's currently filling */
2897 spin_lock(&svcpt->scp_lock);
2898 while (svcpt->scp_nrqbds_posted != 0) {
2899 spin_unlock(&svcpt->scp_lock);
2900 /* Network access will complete in finite time but
2901 * the HUGE timeout lets us CWARN for visibility
2902 * of sluggish NALs */
2903 lwi = LWI_TIMEOUT_INTERVAL(
2904 cfs_time_seconds(LONG_UNLINK),
2905 cfs_time_seconds(1), NULL, NULL);
2906 rc = l_wait_event(svcpt->scp_waitq,
2907 svcpt->scp_nrqbds_posted == 0, &lwi);
2908 if (rc == -ETIMEDOUT) {
2909 CWARN("Service %s waiting for request buffers\n",
2910 svcpt->scp_service->srv_name);
2912 spin_lock(&svcpt->scp_lock);
2914 spin_unlock(&svcpt->scp_lock);
2919 ptlrpc_service_purge_all(struct ptlrpc_service *svc)
2921 struct ptlrpc_service_part *svcpt;
2922 struct ptlrpc_request_buffer_desc *rqbd;
2923 struct ptlrpc_request *req;
2924 struct ptlrpc_reply_state *rs;
2927 ptlrpc_service_for_each_part(svcpt, i, svc) {
2928 if (svcpt->scp_service == NULL)
2931 spin_lock(&svcpt->scp_rep_lock);
2932 while (!list_empty(&svcpt->scp_rep_active)) {
2933 rs = list_entry(svcpt->scp_rep_active.next,
2934 struct ptlrpc_reply_state, rs_list);
2935 spin_lock(&rs->rs_lock);
2936 ptlrpc_schedule_difficult_reply(rs);
2937 spin_unlock(&rs->rs_lock);
2939 spin_unlock(&svcpt->scp_rep_lock);
2941 /* purge the request queue. NB No new replies (rqbds
2942 * all unlinked) and no service threads, so I'm the only
2943 * thread noodling the request queue now */
2944 while (!list_empty(&svcpt->scp_req_incoming)) {
2945 req = list_entry(svcpt->scp_req_incoming.next,
2946 struct ptlrpc_request, rq_list);
2948 list_del(&req->rq_list);
2949 svcpt->scp_nreqs_incoming--;
2950 ptlrpc_server_finish_request(svcpt, req);
2953 while (ptlrpc_server_request_pending(svcpt, true)) {
2954 req = ptlrpc_server_request_get(svcpt, true);
2955 ptlrpc_server_finish_active_request(svcpt, req);
2958 LASSERT(list_empty(&svcpt->scp_rqbd_posted));
2959 LASSERT(svcpt->scp_nreqs_incoming == 0);
2960 LASSERT(svcpt->scp_nreqs_active == 0);
2961 /* history should have been culled by
2962 * ptlrpc_server_finish_request */
2963 LASSERT(svcpt->scp_hist_nrqbds == 0);
2965 /* Now free all the request buffers since nothing
2966 * references them any more... */
2968 while (!list_empty(&svcpt->scp_rqbd_idle)) {
2969 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
2970 struct ptlrpc_request_buffer_desc,
2972 ptlrpc_free_rqbd(rqbd);
2974 ptlrpc_wait_replies(svcpt);
2976 while (!list_empty(&svcpt->scp_rep_idle)) {
2977 rs = list_entry(svcpt->scp_rep_idle.next,
2978 struct ptlrpc_reply_state,
2980 list_del(&rs->rs_list);
2981 OBD_FREE_LARGE(rs, svc->srv_max_reply_size);
2987 ptlrpc_service_free(struct ptlrpc_service *svc)
2989 struct ptlrpc_service_part *svcpt;
2990 struct ptlrpc_at_array *array;
2993 ptlrpc_service_for_each_part(svcpt, i, svc) {
2994 if (svcpt->scp_service == NULL)
2997 /* In case somebody rearmed this in the meantime */
2998 cfs_timer_disarm(&svcpt->scp_at_timer);
2999 array = &svcpt->scp_at_array;
3001 if (array->paa_reqs_array != NULL) {
3002 OBD_FREE(array->paa_reqs_array,
3003 sizeof(struct list_head) * array->paa_size);
3004 array->paa_reqs_array = NULL;
3007 if (array->paa_reqs_count != NULL) {
3008 OBD_FREE(array->paa_reqs_count,
3009 sizeof(__u32) * array->paa_size);
3010 array->paa_reqs_count = NULL;
3014 ptlrpc_service_for_each_part(svcpt, i, svc)
3015 OBD_FREE_PTR(svcpt);
3017 if (svc->srv_cpts != NULL)
3018 cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts);
3020 OBD_FREE(svc, offsetof(struct ptlrpc_service,
3021 srv_parts[svc->srv_ncpts]));
3024 int ptlrpc_unregister_service(struct ptlrpc_service *service)
3026 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
3028 service->srv_is_stopping = 1;
3030 mutex_lock(&ptlrpc_all_services_mutex);
3031 list_del_init(&service->srv_list);
3032 mutex_unlock(&ptlrpc_all_services_mutex);
3034 ptlrpc_service_del_atimer(service);
3035 ptlrpc_stop_all_threads(service);
3037 ptlrpc_service_unlink_rqbd(service);
3038 ptlrpc_service_purge_all(service);
3039 ptlrpc_service_nrs_cleanup(service);
3041 ptlrpc_lprocfs_unregister_service(service);
3043 ptlrpc_service_free(service);
3047 EXPORT_SYMBOL(ptlrpc_unregister_service);
3050 * Returns 0 if the service is healthy.
3052 * Right now, it just checks to make sure that requests aren't languishing
3053 * in the queue. We'll use this health check to govern whether a node needs
3054 * to be shot, so it's intentionally non-aggressive. */
3055 int ptlrpc_svcpt_health_check(struct ptlrpc_service_part *svcpt)
3057 struct ptlrpc_request *request = NULL;
3058 struct timeval right_now;
3061 do_gettimeofday(&right_now);
3063 spin_lock(&svcpt->scp_req_lock);
3064 /* How long has the next entry been waiting? */
3065 if (ptlrpc_server_high_pending(svcpt, true))
3066 request = ptlrpc_nrs_req_peek_nolock(svcpt, true);
3067 else if (ptlrpc_server_normal_pending(svcpt, true))
3068 request = ptlrpc_nrs_req_peek_nolock(svcpt, false);
3070 if (request == NULL) {
3071 spin_unlock(&svcpt->scp_req_lock);
3075 timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
3076 spin_unlock(&svcpt->scp_req_lock);
3078 if ((timediff / ONE_MILLION) >
3079 (AT_OFF ? obd_timeout * 3 / 2 : at_max)) {
3080 CERROR("%s: unhealthy - request has been waiting %lds\n",
3081 svcpt->scp_service->srv_name, timediff / ONE_MILLION);
3089 ptlrpc_service_health_check(struct ptlrpc_service *svc)
3091 struct ptlrpc_service_part *svcpt;
3097 ptlrpc_service_for_each_part(svcpt, i, svc) {
3098 int rc = ptlrpc_svcpt_health_check(svcpt);
3105 EXPORT_SYMBOL(ptlrpc_service_health_check);