4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
33 #define DEBUG_SUBSYSTEM S_RPC
34 #include <linux/kthread.h>
35 #include <obd_support.h>
36 #include <obd_class.h>
37 #include <lustre_net.h>
38 #include <lu_object.h>
39 #include <lnet/types.h>
40 #include "ptlrpc_internal.h"
42 /* The following are visible and mutable through /sys/module/ptlrpc */
43 int test_req_buffer_pressure = 0;
44 module_param(test_req_buffer_pressure, int, 0444);
45 MODULE_PARM_DESC(test_req_buffer_pressure, "set non-zero to put pressure on request buffer pools");
46 module_param(at_min, int, 0644);
47 MODULE_PARM_DESC(at_min, "Adaptive timeout minimum (sec)");
48 module_param(at_max, int, 0644);
49 MODULE_PARM_DESC(at_max, "Adaptive timeout maximum (sec)");
50 module_param(at_history, int, 0644);
51 MODULE_PARM_DESC(at_history,
52 "Adaptive timeouts remember the slowest event that took place within this period (sec)");
53 module_param(at_early_margin, int, 0644);
54 MODULE_PARM_DESC(at_early_margin, "How soon before an RPC deadline to send an early reply");
55 module_param(at_extra, int, 0644);
56 MODULE_PARM_DESC(at_extra, "How much extra time to give with each early reply");
59 static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt);
60 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req);
61 static void ptlrpc_at_remove_timed(struct ptlrpc_request *req);
63 /** Holds a list of all PTLRPC services */
64 struct list_head ptlrpc_all_services;
65 /** Used to protect the \e ptlrpc_all_services list */
66 struct mutex ptlrpc_all_services_mutex;
68 static struct ptlrpc_request_buffer_desc *
69 ptlrpc_alloc_rqbd(struct ptlrpc_service_part *svcpt)
71 struct ptlrpc_service *svc = svcpt->scp_service;
72 struct ptlrpc_request_buffer_desc *rqbd;
74 OBD_CPT_ALLOC_PTR(rqbd, svc->srv_cptable, svcpt->scp_cpt);
78 rqbd->rqbd_svcpt = svcpt;
79 rqbd->rqbd_refcount = 0;
80 rqbd->rqbd_cbid.cbid_fn = request_in_callback;
81 rqbd->rqbd_cbid.cbid_arg = rqbd;
82 INIT_LIST_HEAD(&rqbd->rqbd_reqs);
83 OBD_CPT_ALLOC_LARGE(rqbd->rqbd_buffer, svc->srv_cptable,
84 svcpt->scp_cpt, svc->srv_buf_size);
85 if (rqbd->rqbd_buffer == NULL) {
90 spin_lock(&svcpt->scp_lock);
91 list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
92 svcpt->scp_nrqbds_total++;
93 spin_unlock(&svcpt->scp_lock);
99 ptlrpc_free_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
101 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
103 LASSERT(rqbd->rqbd_refcount == 0);
104 LASSERT(list_empty(&rqbd->rqbd_reqs));
106 spin_lock(&svcpt->scp_lock);
107 list_del(&rqbd->rqbd_list);
108 svcpt->scp_nrqbds_total--;
109 spin_unlock(&svcpt->scp_lock);
111 OBD_FREE_LARGE(rqbd->rqbd_buffer, svcpt->scp_service->srv_buf_size);
116 ptlrpc_grow_req_bufs(struct ptlrpc_service_part *svcpt, int post)
118 struct ptlrpc_service *svc = svcpt->scp_service;
119 struct ptlrpc_request_buffer_desc *rqbd;
123 if (svcpt->scp_rqbd_allocating)
126 spin_lock(&svcpt->scp_lock);
127 /* check again with lock */
128 if (svcpt->scp_rqbd_allocating) {
129 /* NB: we might allow more than one thread in the future */
130 LASSERT(svcpt->scp_rqbd_allocating == 1);
131 spin_unlock(&svcpt->scp_lock);
135 svcpt->scp_rqbd_allocating++;
136 spin_unlock(&svcpt->scp_lock);
139 for (i = 0; i < svc->srv_nbuf_per_group; i++) {
140 /* NB: another thread might have recycled enough rqbds, we
141 * need to make sure it wouldn't over-allocate, see LU-1212. */
142 if (svcpt->scp_nrqbds_posted >= svc->srv_nbuf_per_group)
145 rqbd = ptlrpc_alloc_rqbd(svcpt);
148 CERROR("%s: Can't allocate request buffer\n",
155 spin_lock(&svcpt->scp_lock);
157 LASSERT(svcpt->scp_rqbd_allocating == 1);
158 svcpt->scp_rqbd_allocating--;
160 spin_unlock(&svcpt->scp_lock);
163 "%s: allocate %d new %d-byte reqbufs (%d/%d left), rc = %d\n",
164 svc->srv_name, i, svc->srv_buf_size, svcpt->scp_nrqbds_posted,
165 svcpt->scp_nrqbds_total, rc);
169 rc = ptlrpc_server_post_idle_rqbds(svcpt);
175 * Part of Rep-Ack logic.
176 * Puts a lock and its mode into reply state assotiated to request reply.
179 ptlrpc_save_lock(struct ptlrpc_request *req,
180 struct lustre_handle *lock, int mode, int no_ack)
182 struct ptlrpc_reply_state *rs = req->rq_reply_state;
186 LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
188 if (req->rq_export->exp_disconnected) {
189 ldlm_lock_decref(lock, mode);
191 idx = rs->rs_nlocks++;
192 rs->rs_locks[idx] = *lock;
193 rs->rs_modes[idx] = mode;
194 rs->rs_difficult = 1;
195 rs->rs_no_ack = !!no_ack;
198 EXPORT_SYMBOL(ptlrpc_save_lock);
201 struct ptlrpc_hr_partition;
203 struct ptlrpc_hr_thread {
204 int hrt_id; /* thread ID */
206 wait_queue_head_t hrt_waitq;
207 struct list_head hrt_queue; /* RS queue */
208 struct ptlrpc_hr_partition *hrt_partition;
211 struct ptlrpc_hr_partition {
212 /* # of started threads */
213 atomic_t hrp_nstarted;
214 /* # of stopped threads */
215 atomic_t hrp_nstopped;
216 /* cpu partition id */
218 /* round-robin rotor for choosing thread */
220 /* total number of threads on this partition */
223 struct ptlrpc_hr_thread *hrp_thrs;
226 #define HRT_RUNNING 0
227 #define HRT_STOPPING 1
229 struct ptlrpc_hr_service {
230 /* CPU partition table, it's just cfs_cpt_table for now */
231 struct cfs_cpt_table *hr_cpt_table;
232 /** controller sleep waitq */
233 wait_queue_head_t hr_waitq;
234 unsigned int hr_stopping;
235 /** roundrobin rotor for non-affinity service */
236 unsigned int hr_rotor;
238 struct ptlrpc_hr_partition **hr_partitions;
242 struct list_head rsb_replies;
243 unsigned int rsb_n_replies;
244 struct ptlrpc_service_part *rsb_svcpt;
247 /** reply handling service. */
248 static struct ptlrpc_hr_service ptlrpc_hr;
251 * maximum mumber of replies scheduled in one batch
253 #define MAX_SCHEDULED 256
256 * Initialize a reply batch.
260 static void rs_batch_init(struct rs_batch *b)
262 memset(b, 0, sizeof *b);
263 INIT_LIST_HEAD(&b->rsb_replies);
267 * Choose an hr thread to dispatch requests to.
269 static struct ptlrpc_hr_thread *
270 ptlrpc_hr_select(struct ptlrpc_service_part *svcpt)
272 struct ptlrpc_hr_partition *hrp;
275 if (svcpt->scp_cpt >= 0 &&
276 svcpt->scp_service->srv_cptable == ptlrpc_hr.hr_cpt_table) {
277 /* directly match partition */
278 hrp = ptlrpc_hr.hr_partitions[svcpt->scp_cpt];
281 rotor = ptlrpc_hr.hr_rotor++;
282 rotor %= cfs_cpt_number(ptlrpc_hr.hr_cpt_table);
284 hrp = ptlrpc_hr.hr_partitions[rotor];
287 rotor = hrp->hrp_rotor++;
288 return &hrp->hrp_thrs[rotor % hrp->hrp_nthrs];
292 * Dispatch all replies accumulated in the batch to one from
293 * dedicated reply handling threads.
297 static void rs_batch_dispatch(struct rs_batch *b)
299 if (b->rsb_n_replies != 0) {
300 struct ptlrpc_hr_thread *hrt;
302 hrt = ptlrpc_hr_select(b->rsb_svcpt);
304 spin_lock(&hrt->hrt_lock);
305 list_splice_init(&b->rsb_replies, &hrt->hrt_queue);
306 spin_unlock(&hrt->hrt_lock);
308 wake_up(&hrt->hrt_waitq);
309 b->rsb_n_replies = 0;
314 * Add a reply to a batch.
315 * Add one reply object to a batch, schedule batched replies if overload.
320 static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs)
322 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
324 if (svcpt != b->rsb_svcpt || b->rsb_n_replies >= MAX_SCHEDULED) {
325 if (b->rsb_svcpt != NULL) {
326 rs_batch_dispatch(b);
327 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
329 spin_lock(&svcpt->scp_rep_lock);
330 b->rsb_svcpt = svcpt;
332 spin_lock(&rs->rs_lock);
333 rs->rs_scheduled_ever = 1;
334 if (rs->rs_scheduled == 0) {
335 list_move(&rs->rs_list, &b->rsb_replies);
336 rs->rs_scheduled = 1;
339 rs->rs_committed = 1;
340 spin_unlock(&rs->rs_lock);
344 * Reply batch finalization.
345 * Dispatch remaining replies from the batch
346 * and release remaining spinlock.
350 static void rs_batch_fini(struct rs_batch *b)
352 if (b->rsb_svcpt != NULL) {
353 rs_batch_dispatch(b);
354 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
358 #define DECLARE_RS_BATCH(b) struct rs_batch b
362 * Put reply state into a queue for processing because we received
363 * ACK from the client
365 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
367 struct ptlrpc_hr_thread *hrt;
370 LASSERT(list_empty(&rs->rs_list));
372 hrt = ptlrpc_hr_select(rs->rs_svcpt);
374 spin_lock(&hrt->hrt_lock);
375 list_add_tail(&rs->rs_list, &hrt->hrt_queue);
376 spin_unlock(&hrt->hrt_lock);
378 wake_up(&hrt->hrt_waitq);
383 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 */
397 rs->rs_scheduled = 1;
398 list_del_init(&rs->rs_list);
399 ptlrpc_dispatch_difficult_reply(rs);
402 EXPORT_SYMBOL(ptlrpc_schedule_difficult_reply);
404 void ptlrpc_commit_replies(struct obd_export *exp)
406 struct ptlrpc_reply_state *rs, *nxt;
407 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);
432 ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt)
434 struct ptlrpc_request_buffer_desc *rqbd;
439 spin_lock(&svcpt->scp_lock);
441 if (list_empty(&svcpt->scp_rqbd_idle)) {
442 spin_unlock(&svcpt->scp_lock);
446 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
447 struct ptlrpc_request_buffer_desc,
449 list_del(&rqbd->rqbd_list);
451 /* assume we will post successfully */
452 svcpt->scp_nrqbds_posted++;
453 list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_posted);
455 spin_unlock(&svcpt->scp_lock);
457 rc = ptlrpc_register_rqbd(rqbd);
464 spin_lock(&svcpt->scp_lock);
466 svcpt->scp_nrqbds_posted--;
467 list_del(&rqbd->rqbd_list);
468 list_add_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
470 /* Don't complain if no request buffers are posted right now; LNET
471 * won't drop requests because we set the portal lazy! */
473 spin_unlock(&svcpt->scp_lock);
478 static void ptlrpc_at_timer(unsigned long castmeharder)
480 struct ptlrpc_service_part *svcpt;
482 svcpt = (struct ptlrpc_service_part *)castmeharder;
484 svcpt->scp_at_check = 1;
485 svcpt->scp_at_checktime = cfs_time_current();
486 wake_up(&svcpt->scp_waitq);
490 ptlrpc_server_nthreads_check(struct ptlrpc_service *svc,
491 struct ptlrpc_service_conf *conf)
493 struct ptlrpc_service_thr_conf *tc = &conf->psc_thr;
500 * Common code for estimating & validating threads number.
501 * CPT affinity service could have percpt thread-pool instead
502 * of a global thread-pool, which means user might not always
503 * get the threads number they give it in conf::tc_nthrs_user
504 * even they did set. It's because we need to validate threads
505 * number for each CPT to guarantee each pool will have enough
506 * threads to keep the service healthy.
508 init = PTLRPC_NTHRS_INIT + (svc->srv_ops.so_hpreq_handler != NULL);
509 init = max_t(int, init, tc->tc_nthrs_init);
511 /* NB: please see comments in lustre_lnet.h for definition
512 * details of these members */
513 LASSERT(tc->tc_nthrs_max != 0);
515 if (tc->tc_nthrs_user != 0) {
516 /* In case there is a reason to test a service with many
517 * threads, we give a less strict check here, it can
518 * be up to 8 * nthrs_max */
519 total = min(tc->tc_nthrs_max * 8, tc->tc_nthrs_user);
520 nthrs = total / svc->srv_ncpts;
521 init = max(init, nthrs);
525 total = tc->tc_nthrs_max;
526 if (tc->tc_nthrs_base == 0) {
527 /* don't care about base threads number per partition,
528 * this is most for non-affinity service */
529 nthrs = total / svc->srv_ncpts;
533 nthrs = tc->tc_nthrs_base;
534 if (svc->srv_ncpts == 1) {
537 /* NB: Increase the base number if it's single partition
538 * and total number of cores/HTs is larger or equal to 4.
539 * result will always < 2 * nthrs_base */
540 weight = cfs_cpt_weight(svc->srv_cptable, CFS_CPT_ANY);
541 for (i = 1; (weight >> (i + 1)) != 0 && /* >= 4 cores/HTs */
542 (tc->tc_nthrs_base >> i) != 0; i++)
543 nthrs += tc->tc_nthrs_base >> i;
546 if (tc->tc_thr_factor != 0) {
547 int factor = tc->tc_thr_factor;
551 * User wants to increase number of threads with for
552 * each CPU core/HT, most likely the factor is larger than
553 * one thread/core because service threads are supposed to
554 * be blocked by lock or wait for IO.
557 * Amdahl's law says that adding processors wouldn't give
558 * a linear increasing of parallelism, so it's nonsense to
559 * have too many threads no matter how many cores/HTs
562 if (cpumask_weight(topology_sibling_cpumask(smp_processor_id())) > 1) {
563 /* weight is # of HTs */
564 /* depress thread factor for hyper-thread */
565 factor = factor - (factor >> 1) + (factor >> 3);
568 weight = cfs_cpt_weight(svc->srv_cptable, 0);
571 for (; factor > 0 && weight > 0; factor--, weight -= fade)
572 nthrs += min(weight, fade) * factor;
575 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
576 nthrs = max(tc->tc_nthrs_base,
577 tc->tc_nthrs_max / svc->srv_ncpts);
580 nthrs = max(nthrs, tc->tc_nthrs_init);
581 svc->srv_nthrs_cpt_limit = nthrs;
582 svc->srv_nthrs_cpt_init = init;
584 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
585 CDEBUG(D_OTHER, "%s: This service may have more threads (%d) "
586 "than the given soft limit (%d)\n",
587 svc->srv_name, nthrs * svc->srv_ncpts,
593 * Initialize percpt data for a service
596 ptlrpc_service_part_init(struct ptlrpc_service *svc,
597 struct ptlrpc_service_part *svcpt, int cpt)
599 struct ptlrpc_at_array *array;
604 svcpt->scp_cpt = cpt;
605 INIT_LIST_HEAD(&svcpt->scp_threads);
607 /* rqbd and incoming request queue */
608 spin_lock_init(&svcpt->scp_lock);
609 INIT_LIST_HEAD(&svcpt->scp_rqbd_idle);
610 INIT_LIST_HEAD(&svcpt->scp_rqbd_posted);
611 INIT_LIST_HEAD(&svcpt->scp_req_incoming);
612 init_waitqueue_head(&svcpt->scp_waitq);
613 /* history request & rqbd list */
614 INIT_LIST_HEAD(&svcpt->scp_hist_reqs);
615 INIT_LIST_HEAD(&svcpt->scp_hist_rqbds);
617 /* acitve requests and hp requests */
618 spin_lock_init(&svcpt->scp_req_lock);
621 spin_lock_init(&svcpt->scp_rep_lock);
622 INIT_LIST_HEAD(&svcpt->scp_rep_active);
623 INIT_LIST_HEAD(&svcpt->scp_rep_idle);
624 init_waitqueue_head(&svcpt->scp_rep_waitq);
625 atomic_set(&svcpt->scp_nreps_difficult, 0);
627 /* adaptive timeout */
628 spin_lock_init(&svcpt->scp_at_lock);
629 array = &svcpt->scp_at_array;
631 size = at_est2timeout(at_max);
632 array->paa_size = size;
633 array->paa_count = 0;
634 array->paa_deadline = -1;
636 /* allocate memory for scp_at_array (ptlrpc_at_array) */
637 OBD_CPT_ALLOC(array->paa_reqs_array,
638 svc->srv_cptable, cpt, sizeof(struct list_head) * size);
639 if (array->paa_reqs_array == NULL)
642 for (index = 0; index < size; index++)
643 INIT_LIST_HEAD(&array->paa_reqs_array[index]);
645 OBD_CPT_ALLOC(array->paa_reqs_count,
646 svc->srv_cptable, cpt, sizeof(__u32) * size);
647 if (array->paa_reqs_count == NULL)
650 setup_timer(&svcpt->scp_at_timer, ptlrpc_at_timer,
651 (unsigned long)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 struct proc_dir_entry *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;
703 LASSERT(conf->psc_buf.bc_nbufs > 0);
704 LASSERT(conf->psc_buf.bc_buf_size >=
705 conf->psc_buf.bc_req_max_size + SPTLRPC_MAX_PAYLOAD);
706 LASSERT(conf->psc_thr.tc_ctx_tags != 0);
708 cptable = cconf->cc_cptable;
710 cptable = cfs_cpt_table;
712 if (!conf->psc_thr.tc_cpu_affinity) {
715 ncpts = cfs_cpt_number(cptable);
716 if (cconf->cc_pattern != NULL) {
717 struct cfs_expr_list *el;
719 rc = cfs_expr_list_parse(cconf->cc_pattern,
720 strlen(cconf->cc_pattern),
723 CERROR("%s: invalid CPT pattern string: %s",
724 conf->psc_name, cconf->cc_pattern);
725 RETURN(ERR_PTR(-EINVAL));
728 rc = cfs_expr_list_values(el, ncpts, &cpts);
729 cfs_expr_list_free(el);
731 CERROR("%s: failed to parse CPT array %s: %d\n",
732 conf->psc_name, cconf->cc_pattern, rc);
734 OBD_FREE(cpts, sizeof(*cpts) * ncpts);
735 RETURN(ERR_PTR(rc < 0 ? rc : -EINVAL));
741 OBD_ALLOC(service, offsetof(struct ptlrpc_service, srv_parts[ncpts]));
742 if (service == NULL) {
744 OBD_FREE(cpts, sizeof(*cpts) * ncpts);
745 RETURN(ERR_PTR(-ENOMEM));
748 service->srv_cptable = cptable;
749 service->srv_cpts = cpts;
750 service->srv_ncpts = ncpts;
752 service->srv_cpt_bits = 0; /* it's zero already, easy to read... */
753 while ((1 << service->srv_cpt_bits) < cfs_cpt_number(cptable))
754 service->srv_cpt_bits++;
757 spin_lock_init(&service->srv_lock);
758 service->srv_name = conf->psc_name;
759 service->srv_watchdog_factor = conf->psc_watchdog_factor;
760 INIT_LIST_HEAD(&service->srv_list); /* for safty of cleanup */
762 /* buffer configuration */
763 service->srv_nbuf_per_group = test_req_buffer_pressure ?
764 1 : conf->psc_buf.bc_nbufs;
765 service->srv_max_req_size = conf->psc_buf.bc_req_max_size +
767 service->srv_buf_size = conf->psc_buf.bc_buf_size;
768 service->srv_rep_portal = conf->psc_buf.bc_rep_portal;
769 service->srv_req_portal = conf->psc_buf.bc_req_portal;
771 /* Increase max reply size to next power of two */
772 service->srv_max_reply_size = 1;
773 while (service->srv_max_reply_size <
774 conf->psc_buf.bc_rep_max_size + SPTLRPC_MAX_PAYLOAD)
775 service->srv_max_reply_size <<= 1;
777 service->srv_thread_name = conf->psc_thr.tc_thr_name;
778 service->srv_ctx_tags = conf->psc_thr.tc_ctx_tags;
779 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
780 service->srv_ops = conf->psc_ops;
782 for (i = 0; i < ncpts; i++) {
783 if (!conf->psc_thr.tc_cpu_affinity)
786 cpt = cpts != NULL ? cpts[i] : i;
788 OBD_CPT_ALLOC(svcpt, cptable, cpt, sizeof(*svcpt));
790 GOTO(failed, rc = -ENOMEM);
792 service->srv_parts[i] = svcpt;
793 rc = ptlrpc_service_part_init(service, svcpt, cpt);
798 ptlrpc_server_nthreads_check(service, conf);
800 rc = LNetSetLazyPortal(service->srv_req_portal);
803 mutex_lock(&ptlrpc_all_services_mutex);
804 list_add(&service->srv_list, &ptlrpc_all_services);
805 mutex_unlock(&ptlrpc_all_services_mutex);
807 if (proc_entry != NULL)
808 ptlrpc_lprocfs_register_service(proc_entry, service);
810 rc = ptlrpc_service_nrs_setup(service);
814 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
815 service->srv_name, service->srv_req_portal);
817 rc = ptlrpc_start_threads(service);
819 CERROR("Failed to start threads for service %s: %d\n",
820 service->srv_name, rc);
826 ptlrpc_unregister_service(service);
829 EXPORT_SYMBOL(ptlrpc_register_service);
832 * to actually free the request, must be called without holding svc_lock.
833 * note it's caller's responsibility to unlink req->rq_list.
835 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
837 LASSERT(atomic_read(&req->rq_refcount) == 0);
838 LASSERT(list_empty(&req->rq_timed_list));
840 /* DEBUG_REQ() assumes the reply state of a request with a valid
841 * ref will not be destroyed until that reference is dropped. */
842 ptlrpc_req_drop_rs(req);
844 sptlrpc_svc_ctx_decref(req);
846 if (req != &req->rq_rqbd->rqbd_req) {
847 /* NB request buffers use an embedded
848 * req if the incoming req unlinked the
849 * MD; this isn't one of them! */
850 ptlrpc_request_cache_free(req);
855 * drop a reference count of the request. if it reaches 0, we either
856 * put it into history list, or free it immediately.
858 void ptlrpc_server_drop_request(struct ptlrpc_request *req)
860 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
861 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
862 struct ptlrpc_service *svc = svcpt->scp_service;
864 struct list_head *tmp;
865 struct list_head *nxt;
867 if (!atomic_dec_and_test(&req->rq_refcount))
870 if (req->rq_session.lc_state == LCS_ENTERED) {
871 lu_context_exit(&req->rq_session);
872 lu_context_fini(&req->rq_session);
875 if (req->rq_at_linked) {
876 spin_lock(&svcpt->scp_at_lock);
877 /* recheck with lock, in case it's unlinked by
878 * ptlrpc_at_check_timed() */
879 if (likely(req->rq_at_linked))
880 ptlrpc_at_remove_timed(req);
881 spin_unlock(&svcpt->scp_at_lock);
884 LASSERT(list_empty(&req->rq_timed_list));
886 /* finalize request */
887 if (req->rq_export) {
888 class_export_put(req->rq_export);
889 req->rq_export = NULL;
892 spin_lock(&svcpt->scp_lock);
894 list_add(&req->rq_list, &rqbd->rqbd_reqs);
896 refcount = --(rqbd->rqbd_refcount);
898 /* request buffer is now idle: add to history */
899 list_del(&rqbd->rqbd_list);
901 list_add_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds);
902 svcpt->scp_hist_nrqbds++;
904 /* cull some history?
905 * I expect only about 1 or 2 rqbds need to be recycled here */
906 while (svcpt->scp_hist_nrqbds > svc->srv_hist_nrqbds_cpt_max) {
907 rqbd = list_entry(svcpt->scp_hist_rqbds.next,
908 struct ptlrpc_request_buffer_desc,
911 list_del(&rqbd->rqbd_list);
912 svcpt->scp_hist_nrqbds--;
914 /* remove rqbd's reqs from svc's req history while
915 * I've got the service lock */
916 list_for_each(tmp, &rqbd->rqbd_reqs) {
917 req = list_entry(tmp, struct ptlrpc_request,
919 /* Track the highest culled req seq */
920 if (req->rq_history_seq >
921 svcpt->scp_hist_seq_culled) {
922 svcpt->scp_hist_seq_culled =
925 list_del(&req->rq_history_list);
928 spin_unlock(&svcpt->scp_lock);
930 list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
931 req = list_entry(rqbd->rqbd_reqs.next,
932 struct ptlrpc_request,
934 list_del(&req->rq_list);
935 ptlrpc_server_free_request(req);
938 spin_lock(&svcpt->scp_lock);
940 * now all reqs including the embedded req has been
941 * disposed, schedule request buffer for re-use.
943 LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) == 0);
944 list_add_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
947 spin_unlock(&svcpt->scp_lock);
948 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
949 /* If we are low on memory, we are not interested in history */
950 list_del(&req->rq_list);
951 list_del_init(&req->rq_history_list);
953 /* Track the highest culled req seq */
954 if (req->rq_history_seq > svcpt->scp_hist_seq_culled)
955 svcpt->scp_hist_seq_culled = req->rq_history_seq;
957 spin_unlock(&svcpt->scp_lock);
959 ptlrpc_server_free_request(req);
961 spin_unlock(&svcpt->scp_lock);
965 /** Change request export and move hp request from old export to new */
966 void ptlrpc_request_change_export(struct ptlrpc_request *req,
967 struct obd_export *export)
969 if (req->rq_export != NULL) {
970 LASSERT(!list_empty(&req->rq_exp_list));
971 /* remove rq_exp_list from last export */
972 spin_lock_bh(&req->rq_export->exp_rpc_lock);
973 list_del_init(&req->rq_exp_list);
974 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
975 /* export has one reference already, so it`s safe to
976 * add req to export queue here and get another
977 * reference for request later */
978 spin_lock_bh(&export->exp_rpc_lock);
979 if (req->rq_ops != NULL) /* hp request */
980 list_add(&req->rq_exp_list, &export->exp_hp_rpcs);
982 list_add(&req->rq_exp_list, &export->exp_reg_rpcs);
983 spin_unlock_bh(&export->exp_rpc_lock);
985 class_export_rpc_dec(req->rq_export);
986 class_export_put(req->rq_export);
989 /* request takes one export refcount */
990 req->rq_export = class_export_get(export);
991 class_export_rpc_inc(export);
997 * to finish a request: stop sending more early replies, and release
1000 static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt,
1001 struct ptlrpc_request *req)
1003 ptlrpc_server_hpreq_fini(req);
1005 ptlrpc_server_drop_request(req);
1009 * to finish an active request: stop sending more early replies, and release
1010 * the request. should be called after we finished handling the request.
1012 static void ptlrpc_server_finish_active_request(
1013 struct ptlrpc_service_part *svcpt,
1014 struct ptlrpc_request *req)
1016 spin_lock(&svcpt->scp_req_lock);
1017 ptlrpc_nrs_req_stop_nolock(req);
1018 svcpt->scp_nreqs_active--;
1020 svcpt->scp_nhreqs_active--;
1021 spin_unlock(&svcpt->scp_req_lock);
1023 ptlrpc_nrs_req_finalize(req);
1025 if (req->rq_export != NULL)
1026 class_export_rpc_dec(req->rq_export);
1028 ptlrpc_server_finish_request(svcpt, req);
1032 * This function makes sure dead exports are evicted in a timely manner.
1033 * This function is only called when some export receives a message (i.e.,
1034 * the network is up.)
1036 void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
1038 struct obd_export *oldest_exp;
1039 time_t oldest_time, new_time;
1045 /* Compensate for slow machines, etc, by faking our request time
1046 into the future. Although this can break the strict time-ordering
1047 of the list, we can be really lazy here - we don't have to evict
1048 at the exact right moment. Eventually, all silent exports
1049 will make it to the top of the list. */
1051 /* Do not pay attention on 1sec or smaller renewals. */
1052 new_time = cfs_time_current_sec() + extra_delay;
1053 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
1056 exp->exp_last_request_time = new_time;
1058 /* exports may get disconnected from the chain even though the
1059 export has references, so we must keep the spin lock while
1060 manipulating the lists */
1061 spin_lock(&exp->exp_obd->obd_dev_lock);
1063 if (list_empty(&exp->exp_obd_chain_timed)) {
1064 /* this one is not timed */
1065 spin_unlock(&exp->exp_obd->obd_dev_lock);
1069 list_move_tail(&exp->exp_obd_chain_timed,
1070 &exp->exp_obd->obd_exports_timed);
1072 oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
1073 struct obd_export, exp_obd_chain_timed);
1074 oldest_time = oldest_exp->exp_last_request_time;
1075 spin_unlock(&exp->exp_obd->obd_dev_lock);
1077 if (exp->exp_obd->obd_recovering) {
1078 /* be nice to everyone during recovery */
1083 /* Note - racing to start/reset the obd_eviction timer is safe */
1084 if (exp->exp_obd->obd_eviction_timer == 0) {
1085 /* Check if the oldest entry is expired. */
1086 if (cfs_time_current_sec() > (oldest_time + PING_EVICT_TIMEOUT +
1088 /* We need a second timer, in case the net was down and
1089 * it just came back. Since the pinger may skip every
1090 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
1091 * we better wait for 3. */
1092 exp->exp_obd->obd_eviction_timer =
1093 cfs_time_current_sec() + 3 * PING_INTERVAL;
1094 CDEBUG(D_HA, "%s: Think about evicting %s from "CFS_TIME_T"\n",
1095 exp->exp_obd->obd_name,
1096 obd_export_nid2str(oldest_exp), oldest_time);
1099 if (cfs_time_current_sec() >
1100 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
1101 /* The evictor won't evict anyone who we've heard from
1102 * recently, so we don't have to check before we start
1104 if (!ping_evictor_wake(exp))
1105 exp->exp_obd->obd_eviction_timer = 0;
1113 * Sanity check request \a req.
1114 * Return 0 if all is ok, error code otherwise.
1116 static int ptlrpc_check_req(struct ptlrpc_request *req)
1118 struct obd_device *obd = req->rq_export->exp_obd;
1121 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
1122 req->rq_export->exp_conn_cnt)) {
1123 DEBUG_REQ(D_RPCTRACE, req,
1124 "DROPPING req from old connection %d < %d",
1125 lustre_msg_get_conn_cnt(req->rq_reqmsg),
1126 req->rq_export->exp_conn_cnt);
1129 if (unlikely(obd == NULL || obd->obd_fail)) {
1130 /* Failing over, don't handle any more reqs,
1131 * send error response instead. */
1132 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
1133 req, (obd != NULL) ? obd->obd_name : "unknown");
1135 } else if (lustre_msg_get_flags(req->rq_reqmsg) &
1136 (MSG_REPLAY | MSG_REQ_REPLAY_DONE) &&
1137 !obd->obd_recovering) {
1138 DEBUG_REQ(D_ERROR, req,
1139 "Invalid replay without recovery");
1140 class_fail_export(req->rq_export);
1142 } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 &&
1143 !obd->obd_recovering) {
1144 DEBUG_REQ(D_ERROR, req, "Invalid req with transno "
1145 "%llu without recovery",
1146 lustre_msg_get_transno(req->rq_reqmsg));
1147 class_fail_export(req->rq_export);
1151 if (unlikely(rc < 0)) {
1152 req->rq_status = rc;
1158 static void ptlrpc_at_set_timer(struct ptlrpc_service_part *svcpt)
1160 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1163 if (array->paa_count == 0) {
1164 del_timer(&svcpt->scp_at_timer);
1168 /* Set timer for closest deadline */
1169 next = (__s32)(array->paa_deadline - cfs_time_current_sec() -
1172 ptlrpc_at_timer((unsigned long)svcpt);
1174 mod_timer(&svcpt->scp_at_timer, cfs_time_shift(next));
1175 CDEBUG(D_INFO, "armed %s at %+ds\n",
1176 svcpt->scp_service->srv_name, next);
1180 /* Add rpc to early reply check list */
1181 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
1183 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1184 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1185 struct ptlrpc_request *rq = NULL;
1191 if (req->rq_no_reply)
1194 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
1197 spin_lock(&svcpt->scp_at_lock);
1198 LASSERT(list_empty(&req->rq_timed_list));
1200 index = (unsigned long)req->rq_deadline % array->paa_size;
1201 if (array->paa_reqs_count[index] > 0) {
1202 /* latest rpcs will have the latest deadlines in the list,
1203 * so search backward. */
1204 list_for_each_entry_reverse(rq,
1205 &array->paa_reqs_array[index],
1207 if (req->rq_deadline >= rq->rq_deadline) {
1208 list_add(&req->rq_timed_list,
1209 &rq->rq_timed_list);
1215 /* Add the request at the head of the list */
1216 if (list_empty(&req->rq_timed_list))
1217 list_add(&req->rq_timed_list,
1218 &array->paa_reqs_array[index]);
1220 spin_lock(&req->rq_lock);
1221 req->rq_at_linked = 1;
1222 spin_unlock(&req->rq_lock);
1223 req->rq_at_index = index;
1224 array->paa_reqs_count[index]++;
1226 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
1227 array->paa_deadline = req->rq_deadline;
1228 ptlrpc_at_set_timer(svcpt);
1230 spin_unlock(&svcpt->scp_at_lock);
1236 ptlrpc_at_remove_timed(struct ptlrpc_request *req)
1238 struct ptlrpc_at_array *array;
1240 array = &req->rq_rqbd->rqbd_svcpt->scp_at_array;
1242 /* NB: must call with hold svcpt::scp_at_lock */
1243 LASSERT(!list_empty(&req->rq_timed_list));
1244 list_del_init(&req->rq_timed_list);
1246 spin_lock(&req->rq_lock);
1247 req->rq_at_linked = 0;
1248 spin_unlock(&req->rq_lock);
1250 array->paa_reqs_count[req->rq_at_index]--;
1255 * Attempt to extend the request deadline by sending an early reply to the
1258 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
1260 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1261 struct ptlrpc_request *reqcopy;
1262 struct lustre_msg *reqmsg;
1263 cfs_duration_t olddl = req->rq_deadline - cfs_time_current_sec();
1269 if (CFS_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_RECONNECT)) {
1270 /* don't send early reply */
1274 /* deadline is when the client expects us to reply, margin is the
1275 difference between clients' and servers' expectations */
1276 DEBUG_REQ(D_ADAPTTO, req,
1277 "%ssending early reply (deadline %+lds, margin %+lds) for "
1278 "%d+%d", AT_OFF ? "AT off - not " : "",
1279 olddl, olddl - at_get(&svcpt->scp_at_estimate),
1280 at_get(&svcpt->scp_at_estimate), at_extra);
1286 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), "
1287 "not sending early reply. Consider increasing "
1288 "at_early_margin (%d)?", olddl, at_early_margin);
1290 /* Return an error so we're not re-added to the timed list. */
1294 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
1295 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
1296 "but no AT support");
1300 if (req->rq_export &&
1301 lustre_msg_get_flags(req->rq_reqmsg) &
1302 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1303 struct obd_device *obd_exp = req->rq_export->exp_obd;
1305 /* During recovery, we don't want to send too many early
1306 * replies, but on the other hand we want to make sure the
1307 * client has enough time to resend if the rpc is lost. So
1308 * during the recovery period send at least 4 early replies,
1309 * spacing them every at_extra if we can. at_estimate should
1310 * always equal this fixed value during recovery.
1312 /* Don't account request processing time into AT history
1313 * during recovery, it is not service time we need but
1314 * includes also waiting time for recovering clients
1316 newdl = min_t(time64_t, at_extra,
1317 obd_exp->obd_recovery_timeout / 4) +
1318 ktime_get_real_seconds();
1320 /* We want to extend the request deadline by at_extra seconds,
1321 * so we set our service estimate to reflect how much time has
1322 * passed since this request arrived plus an additional
1323 * at_extra seconds. The client will calculate the new deadline
1324 * based on this service estimate (plus some additional time to
1325 * account for network latency). See ptlrpc_at_recv_early_reply
1327 at_measured(&svcpt->scp_at_estimate, at_extra +
1328 cfs_time_current_sec() -
1329 req->rq_arrival_time.tv_sec);
1330 newdl = req->rq_arrival_time.tv_sec +
1331 at_get(&svcpt->scp_at_estimate);
1334 /* Check to see if we've actually increased the deadline -
1335 * we may be past adaptive_max */
1336 if (req->rq_deadline >= newdl) {
1337 DEBUG_REQ(D_WARNING, req, "Couldn't add any time "
1338 "(%ld/%ld), not sending early reply\n",
1339 olddl, newdl - cfs_time_current_sec());
1343 reqcopy = ptlrpc_request_cache_alloc(GFP_NOFS);
1344 if (reqcopy == NULL)
1346 OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1348 GOTO(out_free, rc = -ENOMEM);
1351 reqcopy->rq_reply_state = NULL;
1352 reqcopy->rq_rep_swab_mask = 0;
1353 reqcopy->rq_pack_bulk = 0;
1354 reqcopy->rq_pack_udesc = 0;
1355 reqcopy->rq_packed_final = 0;
1356 sptlrpc_svc_ctx_addref(reqcopy);
1357 /* We only need the reqmsg for the magic */
1358 reqcopy->rq_reqmsg = reqmsg;
1359 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1362 * tgt_brw_read() and tgt_brw_write() may have decided not to reply.
1363 * Without this check, we would fail the rq_no_reply assertion in
1364 * ptlrpc_send_reply().
1366 if (reqcopy->rq_no_reply)
1367 GOTO(out, rc = -ETIMEDOUT);
1369 LASSERT(atomic_read(&req->rq_refcount));
1370 /** if it is last refcount then early reply isn't needed */
1371 if (atomic_read(&req->rq_refcount) == 1) {
1372 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
1373 "abort sending early reply\n");
1374 GOTO(out, rc = -EINVAL);
1377 /* Connection ref */
1378 reqcopy->rq_export = class_conn2export(
1379 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1380 if (reqcopy->rq_export == NULL)
1381 GOTO(out, rc = -ENODEV);
1384 class_export_rpc_inc(reqcopy->rq_export);
1385 if (reqcopy->rq_export->exp_obd &&
1386 reqcopy->rq_export->exp_obd->obd_fail)
1387 GOTO(out_put, rc = -ENODEV);
1389 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1393 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1396 /* Adjust our own deadline to what we told the client */
1397 req->rq_deadline = newdl;
1398 req->rq_early_count++; /* number sent, server side */
1400 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1403 /* Free the (early) reply state from lustre_pack_reply.
1404 (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
1405 ptlrpc_req_drop_rs(reqcopy);
1408 class_export_rpc_dec(reqcopy->rq_export);
1409 class_export_put(reqcopy->rq_export);
1411 sptlrpc_svc_ctx_decref(reqcopy);
1412 OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1414 ptlrpc_request_cache_free(reqcopy);
1418 /* Send early replies to everybody expiring within at_early_margin
1419 asking for at_extra time */
1420 static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1422 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1423 struct ptlrpc_request *rq, *n;
1424 struct list_head work_list;
1427 time_t now = cfs_time_current_sec();
1428 cfs_duration_t delay;
1429 int first, counter = 0;
1432 spin_lock(&svcpt->scp_at_lock);
1433 if (svcpt->scp_at_check == 0) {
1434 spin_unlock(&svcpt->scp_at_lock);
1437 delay = cfs_time_sub(cfs_time_current(), svcpt->scp_at_checktime);
1438 svcpt->scp_at_check = 0;
1440 if (array->paa_count == 0) {
1441 spin_unlock(&svcpt->scp_at_lock);
1445 /* The timer went off, but maybe the nearest rpc already completed. */
1446 first = array->paa_deadline - now;
1447 if (first > at_early_margin) {
1448 /* We've still got plenty of time. Reset the timer. */
1449 ptlrpc_at_set_timer(svcpt);
1450 spin_unlock(&svcpt->scp_at_lock);
1454 /* We're close to a timeout, and we don't know how much longer the
1455 server will take. Send early replies to everyone expiring soon. */
1456 INIT_LIST_HEAD(&work_list);
1458 index = (unsigned long)array->paa_deadline % array->paa_size;
1459 count = array->paa_count;
1461 count -= array->paa_reqs_count[index];
1462 list_for_each_entry_safe(rq, n,
1463 &array->paa_reqs_array[index],
1465 if (rq->rq_deadline > now + at_early_margin) {
1466 /* update the earliest deadline */
1467 if (deadline == -1 ||
1468 rq->rq_deadline < deadline)
1469 deadline = rq->rq_deadline;
1473 ptlrpc_at_remove_timed(rq);
1475 * ptlrpc_server_drop_request() may drop
1476 * refcount to 0 already. Let's check this and
1477 * don't add entry to work_list
1479 if (likely(atomic_inc_not_zero(&rq->rq_refcount)))
1480 list_add(&rq->rq_timed_list, &work_list);
1484 if (++index >= array->paa_size)
1487 array->paa_deadline = deadline;
1488 /* we have a new earliest deadline, restart the timer */
1489 ptlrpc_at_set_timer(svcpt);
1491 spin_unlock(&svcpt->scp_at_lock);
1493 CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
1494 "replies\n", first, at_extra, counter);
1496 /* We're already past request deadlines before we even get a
1497 chance to send early replies */
1498 LCONSOLE_WARN("%s: This server is not able to keep up with "
1499 "request traffic (cpu-bound).\n",
1500 svcpt->scp_service->srv_name);
1501 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, "
1502 "delay="CFS_DURATION_T"(jiff)\n",
1503 counter, svcpt->scp_nreqs_incoming,
1504 svcpt->scp_nreqs_active,
1505 at_get(&svcpt->scp_at_estimate), delay);
1508 /* we took additional refcount so entries can't be deleted from list, no
1509 * locking is needed */
1510 while (!list_empty(&work_list)) {
1511 rq = list_entry(work_list.next, struct ptlrpc_request,
1513 list_del_init(&rq->rq_timed_list);
1515 if (ptlrpc_at_send_early_reply(rq) == 0)
1516 ptlrpc_at_add_timed(rq);
1518 ptlrpc_server_drop_request(rq);
1521 RETURN(1); /* return "did_something" for liblustre */
1524 /* Check if we are already handling earlier incarnation of this request.
1525 * Called under &req->rq_export->exp_rpc_lock locked */
1526 static int ptlrpc_server_check_resend_in_progress(struct ptlrpc_request *req)
1528 struct ptlrpc_request *tmp = NULL;
1530 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ||
1531 (atomic_read(&req->rq_export->exp_rpc_count) == 0))
1534 /* bulk request are aborted upon reconnect, don't try to
1536 if (req->rq_bulk_write || req->rq_bulk_read)
1539 /* This list should not be longer than max_requests in
1540 * flights on the client, so it is not all that long.
1541 * Also we only hit this codepath in case of a resent
1542 * request which makes it even more rarely hit */
1543 list_for_each_entry(tmp, &req->rq_export->exp_reg_rpcs,
1545 /* Found duplicate one */
1546 if (tmp->rq_xid == req->rq_xid)
1549 list_for_each_entry(tmp, &req->rq_export->exp_hp_rpcs,
1551 /* Found duplicate one */
1552 if (tmp->rq_xid == req->rq_xid)
1558 DEBUG_REQ(D_HA, req, "Found duplicate req in processing");
1559 DEBUG_REQ(D_HA, tmp, "Request being processed");
1564 * Check if a request should be assigned with a high priority.
1566 * \retval < 0: error occurred
1567 * 0: normal RPC request
1568 * +1: high priority request
1570 static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt,
1571 struct ptlrpc_request *req)
1576 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL) {
1577 rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
1584 if (req->rq_export != NULL && req->rq_ops != NULL) {
1585 /* Perform request specific check. We should do this
1586 * check before the request is added into exp_hp_rpcs
1587 * list otherwise it may hit swab race at LU-1044. */
1588 if (req->rq_ops->hpreq_check != NULL) {
1589 rc = req->rq_ops->hpreq_check(req);
1590 if (rc == -ESTALE) {
1591 req->rq_status = rc;
1594 /** can only return error,
1595 * 0 for normal request,
1596 * or 1 for high priority request */
1604 /** Remove the request from the export list. */
1605 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req)
1608 if (req->rq_export) {
1609 /* refresh lock timeout again so that client has more
1610 * room to send lock cancel RPC. */
1611 if (req->rq_ops && req->rq_ops->hpreq_fini)
1612 req->rq_ops->hpreq_fini(req);
1614 spin_lock_bh(&req->rq_export->exp_rpc_lock);
1615 list_del_init(&req->rq_exp_list);
1616 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1621 static int ptlrpc_hpreq_check(struct ptlrpc_request *req)
1626 static struct ptlrpc_hpreq_ops ptlrpc_hpreq_common = {
1627 .hpreq_check = ptlrpc_hpreq_check,
1630 /* Hi-Priority RPC check by RPC operation code. */
1631 int ptlrpc_hpreq_handler(struct ptlrpc_request *req)
1633 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1635 /* Check for export to let only reconnects for not yet evicted
1636 * export to become a HP rpc. */
1637 if ((req->rq_export != NULL) &&
1638 (opc == OBD_PING || opc == MDS_CONNECT || opc == OST_CONNECT))
1639 req->rq_ops = &ptlrpc_hpreq_common;
1643 EXPORT_SYMBOL(ptlrpc_hpreq_handler);
1645 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1646 struct ptlrpc_request *req)
1652 rc = ptlrpc_server_hpreq_init(svcpt, req);
1657 ptlrpc_nrs_req_initialize(svcpt, req, hp);
1659 if (req->rq_export != NULL) {
1660 struct obd_export *exp = req->rq_export;
1662 /* do search for duplicated xid and the adding to the list
1664 spin_lock_bh(&exp->exp_rpc_lock);
1665 rc = ptlrpc_server_check_resend_in_progress(req);
1667 spin_unlock_bh(&exp->exp_rpc_lock);
1669 ptlrpc_nrs_req_finalize(req);
1673 if (hp || req->rq_ops != NULL)
1674 list_add(&req->rq_exp_list, &exp->exp_hp_rpcs);
1676 list_add(&req->rq_exp_list, &exp->exp_reg_rpcs);
1677 spin_unlock_bh(&exp->exp_rpc_lock);
1680 /* the current thread is not the processing thread for this request
1681 * since that, but request is in exp_hp_list and can be find there.
1682 * Remove all relations between request and old thread. */
1683 req->rq_svc_thread->t_env->le_ses = NULL;
1684 req->rq_svc_thread = NULL;
1685 req->rq_session.lc_thread = NULL;
1687 ptlrpc_nrs_req_add(svcpt, req, hp);
1693 * Allow to handle high priority request
1694 * User can call it w/o any lock but need to hold
1695 * ptlrpc_service_part::scp_req_lock to get reliable result
1697 static bool ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1700 int running = svcpt->scp_nthrs_running;
1702 if (!nrs_svcpt_has_hp(svcpt))
1708 if (ptlrpc_nrs_req_throttling_nolock(svcpt, true))
1711 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1712 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1713 /* leave just 1 thread for normal RPCs */
1714 running = PTLRPC_NTHRS_INIT;
1715 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1719 if (svcpt->scp_nreqs_active >= running - 1)
1722 if (svcpt->scp_nhreqs_active == 0)
1725 return !ptlrpc_nrs_req_pending_nolock(svcpt, false) ||
1726 svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1729 static bool ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1732 return ptlrpc_server_allow_high(svcpt, force) &&
1733 ptlrpc_nrs_req_pending_nolock(svcpt, true);
1737 * Only allow normal priority requests on a service that has a high-priority
1738 * queue if forced (i.e. cleanup), if there are other high priority requests
1739 * already being processed (i.e. those threads can service more high-priority
1740 * requests), or if there are enough idle threads that a later thread can do
1741 * a high priority request.
1742 * User can call it w/o any lock but need to hold
1743 * ptlrpc_service_part::scp_req_lock to get reliable result
1745 static bool ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1748 int running = svcpt->scp_nthrs_running;
1749 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1750 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1751 /* leave just 1 thread for normal RPCs */
1752 running = PTLRPC_NTHRS_INIT;
1753 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1760 if (ptlrpc_nrs_req_throttling_nolock(svcpt, false))
1763 if (svcpt->scp_nreqs_active < running - 2)
1766 if (svcpt->scp_nreqs_active >= running - 1)
1769 return svcpt->scp_nhreqs_active > 0 || !nrs_svcpt_has_hp(svcpt);
1772 static bool ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1775 return ptlrpc_server_allow_normal(svcpt, force) &&
1776 ptlrpc_nrs_req_pending_nolock(svcpt, false);
1780 * Returns true if there are requests available in incoming
1781 * request queue for processing and it is allowed to fetch them.
1782 * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1783 * to get reliable result
1784 * \see ptlrpc_server_allow_normal
1785 * \see ptlrpc_server_allow high
1788 ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt, bool force)
1790 return ptlrpc_server_high_pending(svcpt, force) ||
1791 ptlrpc_server_normal_pending(svcpt, force);
1795 * Fetch a request for processing from queue of unprocessed requests.
1796 * Favors high-priority requests.
1797 * Returns a pointer to fetched request.
1799 static struct ptlrpc_request *
1800 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force)
1802 struct ptlrpc_request *req = NULL;
1805 spin_lock(&svcpt->scp_req_lock);
1807 if (ptlrpc_server_high_pending(svcpt, force)) {
1808 req = ptlrpc_nrs_req_get_nolock(svcpt, true, force);
1810 svcpt->scp_hreq_count++;
1815 if (ptlrpc_server_normal_pending(svcpt, force)) {
1816 req = ptlrpc_nrs_req_get_nolock(svcpt, false, force);
1818 svcpt->scp_hreq_count = 0;
1823 spin_unlock(&svcpt->scp_req_lock);
1827 svcpt->scp_nreqs_active++;
1829 svcpt->scp_nhreqs_active++;
1831 spin_unlock(&svcpt->scp_req_lock);
1833 if (likely(req->rq_export))
1834 class_export_rpc_inc(req->rq_export);
1840 * Handle freshly incoming reqs, add to timed early reply list,
1841 * pass on to regular request queue.
1842 * All incoming requests pass through here before getting into
1843 * ptlrpc_server_handle_req later on.
1846 ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt,
1847 struct ptlrpc_thread *thread)
1849 struct ptlrpc_service *svc = svcpt->scp_service;
1850 struct ptlrpc_request *req;
1855 spin_lock(&svcpt->scp_lock);
1856 if (list_empty(&svcpt->scp_req_incoming)) {
1857 spin_unlock(&svcpt->scp_lock);
1861 req = list_entry(svcpt->scp_req_incoming.next,
1862 struct ptlrpc_request, rq_list);
1863 list_del_init(&req->rq_list);
1864 svcpt->scp_nreqs_incoming--;
1865 /* Consider this still a "queued" request as far as stats are
1867 spin_unlock(&svcpt->scp_lock);
1869 /* go through security check/transform */
1870 rc = sptlrpc_svc_unwrap_request(req);
1874 case SECSVC_COMPLETE:
1875 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1884 * for null-flavored rpc, msg has been unpacked by sptlrpc, although
1885 * redo it wouldn't be harmful.
1887 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
1888 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
1890 CERROR("error unpacking request: ptl %d from %s "
1891 "x%llu\n", svc->srv_req_portal,
1892 libcfs_id2str(req->rq_peer), req->rq_xid);
1897 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1899 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
1900 "%llu\n", svc->srv_req_portal,
1901 libcfs_id2str(req->rq_peer), req->rq_xid);
1905 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
1906 lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
1907 CERROR("drop incoming rpc opc %u, x%llu\n",
1908 cfs_fail_val, req->rq_xid);
1913 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1914 CERROR("wrong packet type received (type=%u) from %s\n",
1915 lustre_msg_get_type(req->rq_reqmsg),
1916 libcfs_id2str(req->rq_peer));
1920 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1924 req->rq_bulk_write = 1;
1928 case MGS_CONFIG_READ:
1929 req->rq_bulk_read = 1;
1933 CDEBUG(D_RPCTRACE, "got req x%llu\n", req->rq_xid);
1935 req->rq_export = class_conn2export(
1936 lustre_msg_get_handle(req->rq_reqmsg));
1937 if (req->rq_export) {
1938 rc = ptlrpc_check_req(req);
1940 rc = sptlrpc_target_export_check(req->rq_export, req);
1942 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
1943 "illegal security flavor,");
1948 ptlrpc_update_export_timer(req->rq_export, 0);
1951 /* req_in handling should/must be fast */
1952 if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5)
1953 DEBUG_REQ(D_WARNING, req, "Slow req_in handling "CFS_DURATION_T"s",
1954 cfs_time_sub(cfs_time_current_sec(),
1955 req->rq_arrival_time.tv_sec));
1957 /* Set rpc server deadline and add it to the timed list */
1958 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1959 MSGHDR_AT_SUPPORT) ?
1960 /* The max time the client expects us to take */
1961 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1963 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1964 if (unlikely(deadline == 0)) {
1965 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1969 /* Skip early reply */
1970 if (OBD_FAIL_PRECHECK(OBD_FAIL_MDS_RESEND))
1971 req->rq_deadline += obd_timeout;
1973 req->rq_svc_thread = thread;
1974 if (thread != NULL) {
1975 /* initialize request session, it is needed for request
1976 * processing by target */
1977 rc = lu_context_init(&req->rq_session, LCT_SERVER_SESSION |
1980 CERROR("%s: failure to initialize session: rc = %d\n",
1981 thread->t_name, rc);
1984 req->rq_session.lc_thread = thread;
1985 lu_context_enter(&req->rq_session);
1986 thread->t_env->le_ses = &req->rq_session;
1989 ptlrpc_at_add_timed(req);
1991 /* Move it over to the request processing queue */
1992 rc = ptlrpc_server_request_add(svcpt, req);
1996 wake_up(&svcpt->scp_waitq);
2000 ptlrpc_server_finish_request(svcpt, req);
2006 * Main incoming request handling logic.
2007 * Calls handler function from service to do actual processing.
2010 ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
2011 struct ptlrpc_thread *thread)
2013 struct ptlrpc_service *svc = svcpt->scp_service;
2014 struct ptlrpc_request *request;
2015 struct timeval work_start;
2016 struct timeval work_end;
2022 request = ptlrpc_server_request_get(svcpt, false);
2023 if (request == NULL)
2026 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
2027 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
2028 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
2029 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
2031 if (unlikely(fail_opc)) {
2032 if (request->rq_export && request->rq_ops)
2033 OBD_FAIL_TIMEOUT(fail_opc, 4);
2036 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
2038 if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
2039 libcfs_debug_dumplog();
2041 do_gettimeofday(&work_start);
2042 timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
2043 if (likely(svc->srv_stats != NULL)) {
2044 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
2046 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
2047 svcpt->scp_nreqs_incoming);
2048 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
2049 svcpt->scp_nreqs_active);
2050 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
2051 at_get(&svcpt->scp_at_estimate));
2054 if (likely(request->rq_export)) {
2055 if (unlikely(ptlrpc_check_req(request)))
2057 ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
2060 /* Discard requests queued for longer than the deadline.
2061 The deadline is increased if we send an early reply. */
2062 if (cfs_time_current_sec() > request->rq_deadline) {
2063 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s"
2064 ": deadline "CFS_DURATION_T":"CFS_DURATION_T"s ago\n",
2065 libcfs_id2str(request->rq_peer),
2066 cfs_time_sub(request->rq_deadline,
2067 request->rq_arrival_time.tv_sec),
2068 cfs_time_sub(cfs_time_current_sec(),
2069 request->rq_deadline));
2073 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
2074 "%s:%s+%d:%d:x%llu:%s:%d\n", current_comm(),
2075 (request->rq_export ?
2076 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2077 (request->rq_export ?
2078 atomic_read(&request->rq_export->exp_refcount) : -99),
2079 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
2080 libcfs_id2str(request->rq_peer),
2081 lustre_msg_get_opc(request->rq_reqmsg));
2083 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
2084 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
2086 CDEBUG(D_NET, "got req %llu\n", request->rq_xid);
2088 /* re-assign request and sesson thread to the current one */
2089 request->rq_svc_thread = thread;
2090 if (thread != NULL) {
2091 LASSERT(request->rq_session.lc_thread == NULL);
2092 request->rq_session.lc_thread = thread;
2093 thread->t_env->le_ses = &request->rq_session;
2095 svc->srv_ops.so_req_handler(request);
2097 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
2100 if (unlikely(cfs_time_current_sec() > request->rq_deadline)) {
2101 DEBUG_REQ(D_WARNING, request, "Request took longer "
2102 "than estimated ("CFS_DURATION_T":"CFS_DURATION_T"s);"
2103 " client may timeout.",
2104 cfs_time_sub(request->rq_deadline,
2105 request->rq_arrival_time.tv_sec),
2106 cfs_time_sub(cfs_time_current_sec(),
2107 request->rq_deadline));
2110 do_gettimeofday(&work_end);
2111 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
2112 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
2113 "%s:%s+%d:%d:x%llu:%s:%d Request procesed in "
2114 "%ldus (%ldus total) trans %llu rc %d/%d\n",
2116 (request->rq_export ?
2117 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2118 (request->rq_export ?
2119 atomic_read(&request->rq_export->exp_refcount) : -99),
2120 lustre_msg_get_status(request->rq_reqmsg),
2122 libcfs_id2str(request->rq_peer),
2123 lustre_msg_get_opc(request->rq_reqmsg),
2125 cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL),
2126 (request->rq_repmsg ?
2127 lustre_msg_get_transno(request->rq_repmsg) :
2128 request->rq_transno),
2130 (request->rq_repmsg ?
2131 lustre_msg_get_status(request->rq_repmsg) : -999));
2132 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
2133 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
2134 int opc = opcode_offset(op);
2135 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
2136 LASSERT(opc < LUSTRE_MAX_OPCODES);
2137 lprocfs_counter_add(svc->srv_stats,
2138 opc + EXTRA_MAX_OPCODES,
2142 if (unlikely(request->rq_early_count)) {
2143 DEBUG_REQ(D_ADAPTTO, request,
2144 "sent %d early replies before finishing in "
2146 request->rq_early_count,
2147 cfs_time_sub(work_end.tv_sec,
2148 request->rq_arrival_time.tv_sec));
2151 ptlrpc_server_finish_active_request(svcpt, request);
2157 * An internal function to process a single reply state object.
2160 ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
2162 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
2163 struct ptlrpc_service *svc = svcpt->scp_service;
2164 struct obd_export *exp;
2169 exp = rs->rs_export;
2171 LASSERT(rs->rs_difficult);
2172 LASSERT(rs->rs_scheduled);
2173 LASSERT(list_empty(&rs->rs_list));
2175 spin_lock(&exp->exp_lock);
2176 /* Noop if removed already */
2177 list_del_init(&rs->rs_exp_list);
2178 spin_unlock(&exp->exp_lock);
2180 /* The disk commit callback holds exp_uncommitted_replies_lock while it
2181 * iterates over newly committed replies, removing them from
2182 * exp_uncommitted_replies. It then drops this lock and schedules the
2183 * replies it found for handling here.
2185 * We can avoid contention for exp_uncommitted_replies_lock between the
2186 * HRT threads and further commit callbacks by checking rs_committed
2187 * which is set in the commit callback while it holds both
2188 * rs_lock and exp_uncommitted_reples.
2190 * If we see rs_committed clear, the commit callback _may_ not have
2191 * handled this reply yet and we race with it to grab
2192 * exp_uncommitted_replies_lock before removing the reply from
2193 * exp_uncommitted_replies. Note that if we lose the race and the
2194 * reply has already been removed, list_del_init() is a noop.
2196 * If we see rs_committed set, we know the commit callback is handling,
2197 * or has handled this reply since store reordering might allow us to
2198 * see rs_committed set out of sequence. But since this is done
2199 * holding rs_lock, we can be sure it has all completed once we hold
2200 * rs_lock, which we do right next.
2202 if (!rs->rs_committed) {
2203 spin_lock(&exp->exp_uncommitted_replies_lock);
2204 list_del_init(&rs->rs_obd_list);
2205 spin_unlock(&exp->exp_uncommitted_replies_lock);
2208 spin_lock(&rs->rs_lock);
2210 been_handled = rs->rs_handled;
2213 nlocks = rs->rs_nlocks; /* atomic "steal", but */
2214 rs->rs_nlocks = 0; /* locks still on rs_locks! */
2216 if (nlocks == 0 && !been_handled) {
2217 /* If we see this, we should already have seen the warning
2218 * in mds_steal_ack_locks() */
2219 CDEBUG(D_HA, "All locks stolen from rs %p x%lld.t%lld"
2222 rs->rs_xid, rs->rs_transno, rs->rs_opc,
2223 libcfs_nid2str(exp->exp_connection->c_peer.nid));
2226 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
2227 spin_unlock(&rs->rs_lock);
2229 if (!been_handled && rs->rs_on_net) {
2230 LNetMDUnlink(rs->rs_md_h);
2231 /* Ignore return code; we're racing with completion */
2234 while (nlocks-- > 0)
2235 ldlm_lock_decref(&rs->rs_locks[nlocks],
2236 rs->rs_modes[nlocks]);
2238 spin_lock(&rs->rs_lock);
2241 rs->rs_scheduled = 0;
2243 if (!rs->rs_on_net) {
2245 spin_unlock(&rs->rs_lock);
2247 class_export_put (exp);
2248 rs->rs_export = NULL;
2249 ptlrpc_rs_decref(rs);
2250 if (atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
2251 svc->srv_is_stopping)
2252 wake_up_all(&svcpt->scp_waitq);
2256 /* still on the net; callback will schedule */
2257 spin_unlock(&rs->rs_lock);
2263 ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
2265 int avail = svcpt->scp_nrqbds_posted;
2266 int low_water = test_req_buffer_pressure ? 0 :
2267 svcpt->scp_service->srv_nbuf_per_group / 2;
2269 /* NB I'm not locking; just looking. */
2271 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
2272 * allowed the request history to grow out of control. We could put a
2273 * sanity check on that here and cull some history if we need the
2276 if (avail <= low_water)
2277 ptlrpc_grow_req_bufs(svcpt, 1);
2279 if (svcpt->scp_service->srv_stats) {
2280 lprocfs_counter_add(svcpt->scp_service->srv_stats,
2281 PTLRPC_REQBUF_AVAIL_CNTR, avail);
2286 ptlrpc_retry_rqbds(void *arg)
2288 struct ptlrpc_service_part *svcpt = (struct ptlrpc_service_part *)arg;
2290 svcpt->scp_rqbd_timeout = 0;
2295 ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
2297 return svcpt->scp_nreqs_active <
2298 svcpt->scp_nthrs_running - 1 -
2299 (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL);
2303 * allowed to create more threads
2304 * user can call it w/o any lock but need to hold
2305 * ptlrpc_service_part::scp_lock to get reliable result
2308 ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
2310 return svcpt->scp_nthrs_running +
2311 svcpt->scp_nthrs_starting <
2312 svcpt->scp_service->srv_nthrs_cpt_limit;
2316 * too many requests and allowed to create more threads
2319 ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
2321 return !ptlrpc_threads_enough(svcpt) &&
2322 ptlrpc_threads_increasable(svcpt);
2326 ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2328 return thread_is_stopping(thread) ||
2329 thread->t_svcpt->scp_service->srv_is_stopping;
2333 ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
2335 return !list_empty(&svcpt->scp_rqbd_idle) &&
2336 svcpt->scp_rqbd_timeout == 0;
2340 ptlrpc_at_check(struct ptlrpc_service_part *svcpt)
2342 return svcpt->scp_at_check;
2346 * requests wait on preprocessing
2347 * user can call it w/o any lock but need to hold
2348 * ptlrpc_service_part::scp_lock to get reliable result
2351 ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
2353 return !list_empty(&svcpt->scp_req_incoming);
2356 static __attribute__((__noinline__)) int
2357 ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
2358 struct ptlrpc_thread *thread)
2360 /* Don't exit while there are replies to be handled */
2361 struct l_wait_info lwi = LWI_TIMEOUT(svcpt->scp_rqbd_timeout,
2362 ptlrpc_retry_rqbds, svcpt);
2364 lc_watchdog_disable(thread->t_watchdog);
2368 l_wait_event_exclusive_head(svcpt->scp_waitq,
2369 ptlrpc_thread_stopping(thread) ||
2370 ptlrpc_server_request_incoming(svcpt) ||
2371 ptlrpc_server_request_pending(svcpt, false) ||
2372 ptlrpc_rqbd_pending(svcpt) ||
2373 ptlrpc_at_check(svcpt), &lwi);
2375 if (ptlrpc_thread_stopping(thread))
2378 lc_watchdog_touch(thread->t_watchdog,
2379 ptlrpc_server_get_timeout(svcpt));
2384 * Main thread body for service threads.
2385 * Waits in a loop waiting for new requests to process to appear.
2386 * Every time an incoming requests is added to its queue, a waitq
2387 * is woken up and one of the threads will handle it.
2389 static int ptlrpc_main(void *arg)
2391 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
2392 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2393 struct ptlrpc_service *svc = svcpt->scp_service;
2394 struct ptlrpc_reply_state *rs;
2395 struct group_info *ginfo = NULL;
2397 int counter = 0, rc = 0;
2400 thread->t_pid = current_pid();
2401 unshare_fs_struct();
2403 /* NB: we will call cfs_cpt_bind() for all threads, because we
2404 * might want to run lustre server only on a subset of system CPUs,
2405 * in that case ->scp_cpt is CFS_CPT_ANY */
2406 rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
2408 CWARN("%s: failed to bind %s on CPT %d\n",
2409 svc->srv_name, thread->t_name, svcpt->scp_cpt);
2412 ginfo = groups_alloc(0);
2418 set_current_groups(ginfo);
2419 put_group_info(ginfo);
2421 if (svc->srv_ops.so_thr_init != NULL) {
2422 rc = svc->srv_ops.so_thr_init(thread);
2433 rc = lu_context_init(&env->le_ctx,
2434 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2438 thread->t_env = env;
2439 env->le_ctx.lc_thread = thread;
2440 env->le_ctx.lc_cookie = 0x6;
2442 while (!list_empty(&svcpt->scp_rqbd_idle)) {
2443 rc = ptlrpc_server_post_idle_rqbds(svcpt);
2447 CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
2448 svc->srv_name, svcpt->scp_cpt, rc);
2452 /* Alloc reply state structure for this one */
2453 OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2459 spin_lock(&svcpt->scp_lock);
2461 LASSERT(thread_is_starting(thread));
2462 thread_clear_flags(thread, SVC_STARTING);
2464 LASSERT(svcpt->scp_nthrs_starting == 1);
2465 svcpt->scp_nthrs_starting--;
2467 /* SVC_STOPPING may already be set here if someone else is trying
2468 * to stop the service while this new thread has been dynamically
2469 * forked. We still set SVC_RUNNING to let our creator know that
2470 * we are now running, however we will exit as soon as possible */
2471 thread_add_flags(thread, SVC_RUNNING);
2472 svcpt->scp_nthrs_running++;
2473 spin_unlock(&svcpt->scp_lock);
2475 /* wake up our creator in case he's still waiting. */
2476 wake_up(&thread->t_ctl_waitq);
2478 thread->t_watchdog = lc_watchdog_add(ptlrpc_server_get_timeout(svcpt),
2481 spin_lock(&svcpt->scp_rep_lock);
2482 list_add(&rs->rs_list, &svcpt->scp_rep_idle);
2483 wake_up(&svcpt->scp_rep_waitq);
2484 spin_unlock(&svcpt->scp_rep_lock);
2486 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2487 svcpt->scp_nthrs_running);
2489 /* XXX maintain a list of all managed devices: insert here */
2490 while (!ptlrpc_thread_stopping(thread)) {
2491 if (ptlrpc_wait_event(svcpt, thread))
2494 ptlrpc_check_rqbd_pool(svcpt);
2496 if (ptlrpc_threads_need_create(svcpt)) {
2497 /* Ignore return code - we tried... */
2498 ptlrpc_start_thread(svcpt, 0);
2501 /* reset le_ses to initial state */
2503 /* Process all incoming reqs before handling any */
2504 if (ptlrpc_server_request_incoming(svcpt)) {
2505 lu_context_enter(&env->le_ctx);
2506 ptlrpc_server_handle_req_in(svcpt, thread);
2507 lu_context_exit(&env->le_ctx);
2509 /* but limit ourselves in case of flood */
2510 if (counter++ < 100)
2515 if (ptlrpc_at_check(svcpt))
2516 ptlrpc_at_check_timed(svcpt);
2518 if (ptlrpc_server_request_pending(svcpt, false)) {
2519 lu_context_enter(&env->le_ctx);
2520 ptlrpc_server_handle_request(svcpt, thread);
2521 lu_context_exit(&env->le_ctx);
2524 if (ptlrpc_rqbd_pending(svcpt) &&
2525 ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
2526 /* I just failed to repost request buffers.
2527 * Wait for a timeout (unless something else
2528 * happens) before I try again */
2529 svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
2530 CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
2531 svcpt->scp_nrqbds_posted);
2535 lc_watchdog_delete(thread->t_watchdog);
2536 thread->t_watchdog = NULL;
2540 * deconstruct service specific state created by ptlrpc_start_thread()
2542 if (svc->srv_ops.so_thr_done != NULL)
2543 svc->srv_ops.so_thr_done(thread);
2546 lu_context_fini(&env->le_ctx);
2550 CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n",
2551 thread, thread->t_pid, thread->t_id, rc);
2553 spin_lock(&svcpt->scp_lock);
2554 if (thread_test_and_clear_flags(thread, SVC_STARTING))
2555 svcpt->scp_nthrs_starting--;
2557 if (thread_test_and_clear_flags(thread, SVC_RUNNING)) {
2558 /* must know immediately */
2559 svcpt->scp_nthrs_running--;
2563 thread_add_flags(thread, SVC_STOPPED);
2565 wake_up(&thread->t_ctl_waitq);
2566 spin_unlock(&svcpt->scp_lock);
2571 static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
2572 struct list_head *replies)
2576 spin_lock(&hrt->hrt_lock);
2578 list_splice_init(&hrt->hrt_queue, replies);
2579 result = ptlrpc_hr.hr_stopping || !list_empty(replies);
2581 spin_unlock(&hrt->hrt_lock);
2586 * Main body of "handle reply" function.
2587 * It processes acked reply states
2589 static int ptlrpc_hr_main(void *arg)
2591 struct ptlrpc_hr_thread *hrt = (struct ptlrpc_hr_thread *)arg;
2592 struct ptlrpc_hr_partition *hrp = hrt->hrt_partition;
2593 struct list_head replies;
2596 INIT_LIST_HEAD(&replies);
2597 unshare_fs_struct();
2599 rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
2601 char threadname[20];
2603 snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
2604 hrp->hrp_cpt, hrt->hrt_id);
2605 CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
2606 threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
2609 atomic_inc(&hrp->hrp_nstarted);
2610 wake_up(&ptlrpc_hr.hr_waitq);
2612 while (!ptlrpc_hr.hr_stopping) {
2613 l_wait_condition(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
2615 while (!list_empty(&replies)) {
2616 struct ptlrpc_reply_state *rs;
2618 rs = list_entry(replies.prev,
2619 struct ptlrpc_reply_state,
2621 list_del_init(&rs->rs_list);
2622 ptlrpc_handle_rs(rs);
2626 atomic_inc(&hrp->hrp_nstopped);
2627 wake_up(&ptlrpc_hr.hr_waitq);
2632 static void ptlrpc_stop_hr_threads(void)
2634 struct ptlrpc_hr_partition *hrp;
2638 ptlrpc_hr.hr_stopping = 1;
2640 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2641 if (hrp->hrp_thrs == NULL)
2642 continue; /* uninitialized */
2643 for (j = 0; j < hrp->hrp_nthrs; j++)
2644 wake_up_all(&hrp->hrp_thrs[j].hrt_waitq);
2647 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2648 if (hrp->hrp_thrs == NULL)
2649 continue; /* uninitialized */
2650 wait_event(ptlrpc_hr.hr_waitq,
2651 atomic_read(&hrp->hrp_nstopped) ==
2652 atomic_read(&hrp->hrp_nstarted));
2656 static int ptlrpc_start_hr_threads(void)
2658 struct ptlrpc_hr_partition *hrp;
2663 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2666 for (j = 0; j < hrp->hrp_nthrs; j++) {
2667 struct ptlrpc_hr_thread *hrt = &hrp->hrp_thrs[j];
2668 struct task_struct *task;
2670 task = kthread_run(ptlrpc_hr_main,
2672 "ptlrpc_hr%02d_%03d",
2681 wait_event(ptlrpc_hr.hr_waitq,
2682 atomic_read(&hrp->hrp_nstarted) == j);
2685 CERROR("cannot start reply handler thread %d:%d: "
2686 "rc = %d\n", i, j, rc);
2687 ptlrpc_stop_hr_threads();
2695 static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
2697 struct l_wait_info lwi = { 0 };
2698 struct ptlrpc_thread *thread;
2699 struct list_head zombie;
2703 CDEBUG(D_INFO, "Stopping threads for service %s\n",
2704 svcpt->scp_service->srv_name);
2706 INIT_LIST_HEAD(&zombie);
2707 spin_lock(&svcpt->scp_lock);
2708 /* let the thread know that we would like it to stop asap */
2709 list_for_each_entry(thread, &svcpt->scp_threads, t_link) {
2710 CDEBUG(D_INFO, "Stopping thread %s #%u\n",
2711 svcpt->scp_service->srv_thread_name, thread->t_id);
2712 thread_add_flags(thread, SVC_STOPPING);
2715 wake_up_all(&svcpt->scp_waitq);
2717 while (!list_empty(&svcpt->scp_threads)) {
2718 thread = list_entry(svcpt->scp_threads.next,
2719 struct ptlrpc_thread, t_link);
2720 if (thread_is_stopped(thread)) {
2721 list_del(&thread->t_link);
2722 list_add(&thread->t_link, &zombie);
2725 spin_unlock(&svcpt->scp_lock);
2727 CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
2728 svcpt->scp_service->srv_thread_name, thread->t_id);
2729 l_wait_event(thread->t_ctl_waitq,
2730 thread_is_stopped(thread), &lwi);
2732 spin_lock(&svcpt->scp_lock);
2735 spin_unlock(&svcpt->scp_lock);
2737 while (!list_empty(&zombie)) {
2738 thread = list_entry(zombie.next,
2739 struct ptlrpc_thread, t_link);
2740 list_del(&thread->t_link);
2741 OBD_FREE_PTR(thread);
2747 * Stops all threads of a particular service \a svc
2749 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2751 struct ptlrpc_service_part *svcpt;
2755 ptlrpc_service_for_each_part(svcpt, i, svc) {
2756 if (svcpt->scp_service != NULL)
2757 ptlrpc_svcpt_stop_threads(svcpt);
2763 int ptlrpc_start_threads(struct ptlrpc_service *svc)
2770 /* We require 2 threads min, see note in ptlrpc_server_handle_request */
2771 LASSERT(svc->srv_nthrs_cpt_init >= PTLRPC_NTHRS_INIT);
2773 for (i = 0; i < svc->srv_ncpts; i++) {
2774 for (j = 0; j < svc->srv_nthrs_cpt_init; j++) {
2775 rc = ptlrpc_start_thread(svc->srv_parts[i], 1);
2781 /* We have enough threads, don't start more. b=15759 */
2788 CERROR("cannot start %s thread #%d_%d: rc %d\n",
2789 svc->srv_thread_name, i, j, rc);
2790 ptlrpc_stop_all_threads(svc);
2794 int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
2796 struct l_wait_info lwi = { 0 };
2797 struct ptlrpc_thread *thread;
2798 struct ptlrpc_service *svc;
2799 struct task_struct *task;
2803 LASSERT(svcpt != NULL);
2805 svc = svcpt->scp_service;
2807 CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n",
2808 svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running,
2809 svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit);
2812 if (unlikely(svc->srv_is_stopping))
2815 if (!ptlrpc_threads_increasable(svcpt) ||
2816 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
2817 svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1))
2820 OBD_CPT_ALLOC_PTR(thread, svc->srv_cptable, svcpt->scp_cpt);
2823 init_waitqueue_head(&thread->t_ctl_waitq);
2825 spin_lock(&svcpt->scp_lock);
2826 if (!ptlrpc_threads_increasable(svcpt)) {
2827 spin_unlock(&svcpt->scp_lock);
2828 OBD_FREE_PTR(thread);
2832 if (svcpt->scp_nthrs_starting != 0) {
2833 /* serialize starting because some modules (obdfilter)
2834 * might require unique and contiguous t_id */
2835 LASSERT(svcpt->scp_nthrs_starting == 1);
2836 spin_unlock(&svcpt->scp_lock);
2837 OBD_FREE_PTR(thread);
2839 CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n",
2840 svc->srv_thread_name, svcpt->scp_thr_nextid);
2845 CDEBUG(D_INFO, "Creating thread %s #%d race, retry later\n",
2846 svc->srv_thread_name, svcpt->scp_thr_nextid);
2850 svcpt->scp_nthrs_starting++;
2851 thread->t_id = svcpt->scp_thr_nextid++;
2852 thread_add_flags(thread, SVC_STARTING);
2853 thread->t_svcpt = svcpt;
2855 list_add(&thread->t_link, &svcpt->scp_threads);
2856 spin_unlock(&svcpt->scp_lock);
2858 if (svcpt->scp_cpt >= 0) {
2859 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s%02d_%03d",
2860 svc->srv_thread_name, svcpt->scp_cpt, thread->t_id);
2862 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s_%04d",
2863 svc->srv_thread_name, thread->t_id);
2866 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name);
2867 task = kthread_run(ptlrpc_main, thread, "%s", thread->t_name);
2870 CERROR("cannot start thread '%s': rc = %d\n",
2871 thread->t_name, rc);
2872 spin_lock(&svcpt->scp_lock);
2873 --svcpt->scp_nthrs_starting;
2874 if (thread_is_stopping(thread)) {
2875 /* this ptlrpc_thread is being hanled
2876 * by ptlrpc_svcpt_stop_threads now
2878 thread_add_flags(thread, SVC_STOPPED);
2879 wake_up(&thread->t_ctl_waitq);
2880 spin_unlock(&svcpt->scp_lock);
2882 list_del(&thread->t_link);
2883 spin_unlock(&svcpt->scp_lock);
2884 OBD_FREE_PTR(thread);
2892 l_wait_event(thread->t_ctl_waitq,
2893 thread_is_running(thread) || thread_is_stopped(thread),
2896 rc = thread_is_stopped(thread) ? thread->t_id : 0;
2900 int ptlrpc_hr_init(void)
2902 struct ptlrpc_hr_partition *hrp;
2903 struct ptlrpc_hr_thread *hrt;
2910 memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
2911 ptlrpc_hr.hr_cpt_table = cfs_cpt_table;
2913 ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
2915 if (ptlrpc_hr.hr_partitions == NULL)
2918 init_waitqueue_head(&ptlrpc_hr.hr_waitq);
2920 weight = cpumask_weight(topology_sibling_cpumask(smp_processor_id()));
2922 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2925 atomic_set(&hrp->hrp_nstarted, 0);
2926 atomic_set(&hrp->hrp_nstopped, 0);
2928 hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, i);
2930 hrp->hrp_nthrs /= weight;
2931 if (hrp->hrp_nthrs == 0)
2934 OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, i,
2935 hrp->hrp_nthrs * sizeof(*hrt));
2936 if (hrp->hrp_thrs == NULL)
2937 GOTO(out, rc = -ENOMEM);
2939 for (j = 0; j < hrp->hrp_nthrs; j++) {
2940 hrt = &hrp->hrp_thrs[j];
2943 hrt->hrt_partition = hrp;
2944 init_waitqueue_head(&hrt->hrt_waitq);
2945 spin_lock_init(&hrt->hrt_lock);
2946 INIT_LIST_HEAD(&hrt->hrt_queue);
2950 rc = ptlrpc_start_hr_threads();
2957 void ptlrpc_hr_fini(void)
2959 struct ptlrpc_hr_partition *hrp;
2962 if (ptlrpc_hr.hr_partitions == NULL)
2965 ptlrpc_stop_hr_threads();
2967 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2968 if (hrp->hrp_thrs != NULL) {
2969 OBD_FREE(hrp->hrp_thrs,
2970 hrp->hrp_nthrs * sizeof(hrp->hrp_thrs[0]));
2974 cfs_percpt_free(ptlrpc_hr.hr_partitions);
2975 ptlrpc_hr.hr_partitions = NULL;
2980 * Wait until all already scheduled replies are processed.
2982 static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
2986 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
2989 rc = l_wait_event(svcpt->scp_waitq,
2990 atomic_read(&svcpt->scp_nreps_difficult) == 0, &lwi);
2993 CWARN("Unexpectedly long timeout %s %p\n",
2994 svcpt->scp_service->srv_name, svcpt->scp_service);
2999 ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
3001 struct ptlrpc_service_part *svcpt;
3004 /* early disarm AT timer... */
3005 ptlrpc_service_for_each_part(svcpt, i, svc) {
3006 if (svcpt->scp_service != NULL)
3007 del_timer(&svcpt->scp_at_timer);
3012 ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
3014 struct ptlrpc_service_part *svcpt;
3015 struct ptlrpc_request_buffer_desc *rqbd;
3016 struct l_wait_info lwi;
3020 /* All history will be culled when the next request buffer is
3021 * freed in ptlrpc_service_purge_all() */
3022 svc->srv_hist_nrqbds_cpt_max = 0;
3024 rc = LNetClearLazyPortal(svc->srv_req_portal);
3027 ptlrpc_service_for_each_part(svcpt, i, svc) {
3028 if (svcpt->scp_service == NULL)
3031 /* Unlink all the request buffers. This forces a 'final'
3032 * event with its 'unlink' flag set for each posted rqbd */
3033 list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
3035 rc = LNetMDUnlink(rqbd->rqbd_md_h);
3036 LASSERT(rc == 0 || rc == -ENOENT);
3040 ptlrpc_service_for_each_part(svcpt, i, svc) {
3041 if (svcpt->scp_service == NULL)
3044 /* Wait for the network to release any buffers
3045 * it's currently filling */
3046 spin_lock(&svcpt->scp_lock);
3047 while (svcpt->scp_nrqbds_posted != 0) {
3048 spin_unlock(&svcpt->scp_lock);
3049 /* Network access will complete in finite time but
3050 * the HUGE timeout lets us CWARN for visibility
3051 * of sluggish NALs */
3052 lwi = LWI_TIMEOUT_INTERVAL(
3053 cfs_time_seconds(LONG_UNLINK),
3054 cfs_time_seconds(1), NULL, NULL);
3055 rc = l_wait_event(svcpt->scp_waitq,
3056 svcpt->scp_nrqbds_posted == 0, &lwi);
3057 if (rc == -ETIMEDOUT) {
3058 CWARN("Service %s waiting for "
3059 "request buffers\n",
3060 svcpt->scp_service->srv_name);
3062 spin_lock(&svcpt->scp_lock);
3064 spin_unlock(&svcpt->scp_lock);
3069 ptlrpc_service_purge_all(struct ptlrpc_service *svc)
3071 struct ptlrpc_service_part *svcpt;
3072 struct ptlrpc_request_buffer_desc *rqbd;
3073 struct ptlrpc_request *req;
3074 struct ptlrpc_reply_state *rs;
3077 ptlrpc_service_for_each_part(svcpt, i, svc) {
3078 if (svcpt->scp_service == NULL)
3081 spin_lock(&svcpt->scp_rep_lock);
3082 while (!list_empty(&svcpt->scp_rep_active)) {
3083 rs = list_entry(svcpt->scp_rep_active.next,
3084 struct ptlrpc_reply_state, rs_list);
3085 spin_lock(&rs->rs_lock);
3086 ptlrpc_schedule_difficult_reply(rs);
3087 spin_unlock(&rs->rs_lock);
3089 spin_unlock(&svcpt->scp_rep_lock);
3091 /* purge the request queue. NB No new replies (rqbds
3092 * all unlinked) and no service threads, so I'm the only
3093 * thread noodling the request queue now */
3094 while (!list_empty(&svcpt->scp_req_incoming)) {
3095 req = list_entry(svcpt->scp_req_incoming.next,
3096 struct ptlrpc_request, rq_list);
3098 list_del(&req->rq_list);
3099 svcpt->scp_nreqs_incoming--;
3100 ptlrpc_server_finish_request(svcpt, req);
3103 while (ptlrpc_server_request_pending(svcpt, true)) {
3104 req = ptlrpc_server_request_get(svcpt, true);
3105 ptlrpc_server_finish_active_request(svcpt, req);
3108 LASSERT(list_empty(&svcpt->scp_rqbd_posted));
3109 LASSERT(svcpt->scp_nreqs_incoming == 0);
3110 LASSERT(svcpt->scp_nreqs_active == 0);
3111 /* history should have been culled by
3112 * ptlrpc_server_finish_request */
3113 LASSERT(svcpt->scp_hist_nrqbds == 0);
3115 /* Now free all the request buffers since nothing
3116 * references them any more... */
3118 while (!list_empty(&svcpt->scp_rqbd_idle)) {
3119 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
3120 struct ptlrpc_request_buffer_desc,
3122 ptlrpc_free_rqbd(rqbd);
3124 ptlrpc_wait_replies(svcpt);
3126 while (!list_empty(&svcpt->scp_rep_idle)) {
3127 rs = list_entry(svcpt->scp_rep_idle.next,
3128 struct ptlrpc_reply_state,
3130 list_del(&rs->rs_list);
3131 OBD_FREE_LARGE(rs, svc->srv_max_reply_size);
3137 ptlrpc_service_free(struct ptlrpc_service *svc)
3139 struct ptlrpc_service_part *svcpt;
3140 struct ptlrpc_at_array *array;
3143 ptlrpc_service_for_each_part(svcpt, i, svc) {
3144 if (svcpt->scp_service == NULL)
3147 /* In case somebody rearmed this in the meantime */
3148 del_timer(&svcpt->scp_at_timer);
3149 array = &svcpt->scp_at_array;
3151 if (array->paa_reqs_array != NULL) {
3152 OBD_FREE(array->paa_reqs_array,
3153 sizeof(struct list_head) * array->paa_size);
3154 array->paa_reqs_array = NULL;
3157 if (array->paa_reqs_count != NULL) {
3158 OBD_FREE(array->paa_reqs_count,
3159 sizeof(__u32) * array->paa_size);
3160 array->paa_reqs_count = NULL;
3164 ptlrpc_service_for_each_part(svcpt, i, svc)
3165 OBD_FREE_PTR(svcpt);
3167 if (svc->srv_cpts != NULL)
3168 cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts);
3170 OBD_FREE(svc, offsetof(struct ptlrpc_service,
3171 srv_parts[svc->srv_ncpts]));
3174 int ptlrpc_unregister_service(struct ptlrpc_service *service)
3178 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
3180 service->srv_is_stopping = 1;
3182 mutex_lock(&ptlrpc_all_services_mutex);
3183 list_del_init(&service->srv_list);
3184 mutex_unlock(&ptlrpc_all_services_mutex);
3186 ptlrpc_service_del_atimer(service);
3187 ptlrpc_stop_all_threads(service);
3189 ptlrpc_service_unlink_rqbd(service);
3190 ptlrpc_service_purge_all(service);
3191 ptlrpc_service_nrs_cleanup(service);
3193 ptlrpc_lprocfs_unregister_service(service);
3195 ptlrpc_service_free(service);
3199 EXPORT_SYMBOL(ptlrpc_unregister_service);
3202 * Returns 0 if the service is healthy.
3204 * Right now, it just checks to make sure that requests aren't languishing
3205 * in the queue. We'll use this health check to govern whether a node needs
3206 * to be shot, so it's intentionally non-aggressive. */
3207 static int ptlrpc_svcpt_health_check(struct ptlrpc_service_part *svcpt)
3209 struct ptlrpc_request *request = NULL;
3210 struct timeval right_now;
3213 do_gettimeofday(&right_now);
3215 spin_lock(&svcpt->scp_req_lock);
3216 /* How long has the next entry been waiting? */
3217 if (ptlrpc_server_high_pending(svcpt, true))
3218 request = ptlrpc_nrs_req_peek_nolock(svcpt, true);
3219 else if (ptlrpc_server_normal_pending(svcpt, true))
3220 request = ptlrpc_nrs_req_peek_nolock(svcpt, false);
3222 if (request == NULL) {
3223 spin_unlock(&svcpt->scp_req_lock);
3227 timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
3228 spin_unlock(&svcpt->scp_req_lock);
3230 if ((timediff / ONE_MILLION) >
3231 (AT_OFF ? obd_timeout * 3 / 2 : at_max)) {
3232 CERROR("%s: unhealthy - request has been waiting %lds\n",
3233 svcpt->scp_service->srv_name, timediff / ONE_MILLION);
3241 ptlrpc_service_health_check(struct ptlrpc_service *svc)
3243 struct ptlrpc_service_part *svcpt;
3249 ptlrpc_service_for_each_part(svcpt, i, svc) {
3250 int rc = ptlrpc_svcpt_health_check(svcpt);
3257 EXPORT_SYMBOL(ptlrpc_service_health_check);