4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
33 #define DEBUG_SUBSYSTEM S_RPC
34 #include <linux/kthread.h>
35 #include <obd_support.h>
36 #include <obd_class.h>
37 #include <lustre_net.h>
38 #include <lu_object.h>
39 #include <lnet/types.h>
40 #include "ptlrpc_internal.h"
42 /* The following are visible and mutable through /sys/module/ptlrpc */
43 int test_req_buffer_pressure = 0;
44 module_param(test_req_buffer_pressure, int, 0444);
45 MODULE_PARM_DESC(test_req_buffer_pressure, "set non-zero to put pressure on request buffer pools");
46 module_param(at_min, int, 0644);
47 MODULE_PARM_DESC(at_min, "Adaptive timeout minimum (sec)");
48 module_param(at_max, int, 0644);
49 MODULE_PARM_DESC(at_max, "Adaptive timeout maximum (sec)");
50 module_param(at_history, int, 0644);
51 MODULE_PARM_DESC(at_history,
52 "Adaptive timeouts remember the slowest event that took place within this period (sec)");
53 module_param(at_early_margin, int, 0644);
54 MODULE_PARM_DESC(at_early_margin, "How soon before an RPC deadline to send an early reply");
55 module_param(at_extra, int, 0644);
56 MODULE_PARM_DESC(at_extra, "How much extra time to give with each early reply");
59 static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt);
60 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req);
61 static void ptlrpc_at_remove_timed(struct ptlrpc_request *req);
63 /** Holds a list of all PTLRPC services */
64 struct list_head ptlrpc_all_services;
65 /** Used to protect the \e ptlrpc_all_services list */
66 struct mutex ptlrpc_all_services_mutex;
68 static struct ptlrpc_request_buffer_desc *
69 ptlrpc_alloc_rqbd(struct ptlrpc_service_part *svcpt)
71 struct ptlrpc_service *svc = svcpt->scp_service;
72 struct ptlrpc_request_buffer_desc *rqbd;
74 OBD_CPT_ALLOC_PTR(rqbd, svc->srv_cptable, svcpt->scp_cpt);
78 rqbd->rqbd_svcpt = svcpt;
79 rqbd->rqbd_refcount = 0;
80 rqbd->rqbd_cbid.cbid_fn = request_in_callback;
81 rqbd->rqbd_cbid.cbid_arg = rqbd;
82 INIT_LIST_HEAD(&rqbd->rqbd_reqs);
83 OBD_CPT_ALLOC_LARGE(rqbd->rqbd_buffer, svc->srv_cptable,
84 svcpt->scp_cpt, svc->srv_buf_size);
85 if (rqbd->rqbd_buffer == NULL) {
90 spin_lock(&svcpt->scp_lock);
91 list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
92 svcpt->scp_nrqbds_total++;
93 spin_unlock(&svcpt->scp_lock);
99 ptlrpc_free_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
101 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
103 LASSERT(rqbd->rqbd_refcount == 0);
104 LASSERT(list_empty(&rqbd->rqbd_reqs));
106 spin_lock(&svcpt->scp_lock);
107 list_del(&rqbd->rqbd_list);
108 svcpt->scp_nrqbds_total--;
109 spin_unlock(&svcpt->scp_lock);
111 OBD_FREE_LARGE(rqbd->rqbd_buffer, svcpt->scp_service->srv_buf_size);
116 ptlrpc_grow_req_bufs(struct ptlrpc_service_part *svcpt, int post)
118 struct ptlrpc_service *svc = svcpt->scp_service;
119 struct ptlrpc_request_buffer_desc *rqbd;
123 if (svcpt->scp_rqbd_allocating)
126 spin_lock(&svcpt->scp_lock);
127 /* check again with lock */
128 if (svcpt->scp_rqbd_allocating) {
129 /* NB: we might allow more than one thread in the future */
130 LASSERT(svcpt->scp_rqbd_allocating == 1);
131 spin_unlock(&svcpt->scp_lock);
135 svcpt->scp_rqbd_allocating++;
136 spin_unlock(&svcpt->scp_lock);
139 for (i = 0; i < svc->srv_nbuf_per_group; i++) {
140 /* NB: another thread might have recycled enough rqbds, we
141 * need to make sure it wouldn't over-allocate, see LU-1212. */
142 if (svcpt->scp_nrqbds_posted >= svc->srv_nbuf_per_group)
145 rqbd = ptlrpc_alloc_rqbd(svcpt);
148 CERROR("%s: Can't allocate request buffer\n",
155 spin_lock(&svcpt->scp_lock);
157 LASSERT(svcpt->scp_rqbd_allocating == 1);
158 svcpt->scp_rqbd_allocating--;
160 spin_unlock(&svcpt->scp_lock);
163 "%s: allocate %d new %d-byte reqbufs (%d/%d left), rc = %d\n",
164 svc->srv_name, i, svc->srv_buf_size, svcpt->scp_nrqbds_posted,
165 svcpt->scp_nrqbds_total, rc);
169 rc = ptlrpc_server_post_idle_rqbds(svcpt);
175 * Part of Rep-Ack logic.
176 * Puts a lock and its mode into reply state assotiated to request reply.
179 ptlrpc_save_lock(struct ptlrpc_request *req, struct lustre_handle *lock,
180 int mode, bool no_ack, bool convert_lock)
182 struct ptlrpc_reply_state *rs = req->rq_reply_state;
186 LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
188 idx = rs->rs_nlocks++;
189 rs->rs_locks[idx] = *lock;
190 rs->rs_modes[idx] = mode;
191 rs->rs_difficult = 1;
192 rs->rs_no_ack = no_ack;
193 rs->rs_convert_lock = convert_lock;
195 EXPORT_SYMBOL(ptlrpc_save_lock);
198 struct ptlrpc_hr_partition;
200 struct ptlrpc_hr_thread {
201 int hrt_id; /* thread ID */
203 wait_queue_head_t hrt_waitq;
204 struct list_head hrt_queue;
205 struct ptlrpc_hr_partition *hrt_partition;
208 struct ptlrpc_hr_partition {
209 /* # of started threads */
210 atomic_t hrp_nstarted;
211 /* # of stopped threads */
212 atomic_t hrp_nstopped;
213 /* cpu partition id */
215 /* round-robin rotor for choosing thread */
217 /* total number of threads on this partition */
220 struct ptlrpc_hr_thread *hrp_thrs;
223 #define HRT_RUNNING 0
224 #define HRT_STOPPING 1
226 struct ptlrpc_hr_service {
227 /* CPU partition table, it's just cfs_cpt_table for now */
228 struct cfs_cpt_table *hr_cpt_table;
229 /** controller sleep waitq */
230 wait_queue_head_t hr_waitq;
231 unsigned int hr_stopping;
232 /** roundrobin rotor for non-affinity service */
233 unsigned int hr_rotor;
235 struct ptlrpc_hr_partition **hr_partitions;
239 struct list_head rsb_replies;
240 unsigned int rsb_n_replies;
241 struct ptlrpc_service_part *rsb_svcpt;
244 /** reply handling service. */
245 static struct ptlrpc_hr_service ptlrpc_hr;
248 * maximum mumber of replies scheduled in one batch
250 #define MAX_SCHEDULED 256
253 * Initialize a reply batch.
257 static void rs_batch_init(struct rs_batch *b)
259 memset(b, 0, sizeof *b);
260 INIT_LIST_HEAD(&b->rsb_replies);
264 * Choose an hr thread to dispatch requests to.
266 static struct ptlrpc_hr_thread *
267 ptlrpc_hr_select(struct ptlrpc_service_part *svcpt)
269 struct ptlrpc_hr_partition *hrp;
272 if (svcpt->scp_cpt >= 0 &&
273 svcpt->scp_service->srv_cptable == ptlrpc_hr.hr_cpt_table) {
274 /* directly match partition */
275 hrp = ptlrpc_hr.hr_partitions[svcpt->scp_cpt];
278 rotor = ptlrpc_hr.hr_rotor++;
279 rotor %= cfs_cpt_number(ptlrpc_hr.hr_cpt_table);
281 hrp = ptlrpc_hr.hr_partitions[rotor];
284 rotor = hrp->hrp_rotor++;
285 return &hrp->hrp_thrs[rotor % hrp->hrp_nthrs];
289 * Dispatch all replies accumulated in the batch to one from
290 * dedicated reply handling threads.
294 static void rs_batch_dispatch(struct rs_batch *b)
296 if (b->rsb_n_replies != 0) {
297 struct ptlrpc_hr_thread *hrt;
299 hrt = ptlrpc_hr_select(b->rsb_svcpt);
301 spin_lock(&hrt->hrt_lock);
302 list_splice_init(&b->rsb_replies, &hrt->hrt_queue);
303 spin_unlock(&hrt->hrt_lock);
305 wake_up(&hrt->hrt_waitq);
306 b->rsb_n_replies = 0;
311 * Add a reply to a batch.
312 * Add one reply object to a batch, schedule batched replies if overload.
317 static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs)
319 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
321 if (svcpt != b->rsb_svcpt || b->rsb_n_replies >= MAX_SCHEDULED) {
322 if (b->rsb_svcpt != NULL) {
323 rs_batch_dispatch(b);
324 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
326 spin_lock(&svcpt->scp_rep_lock);
327 b->rsb_svcpt = svcpt;
329 spin_lock(&rs->rs_lock);
330 rs->rs_scheduled_ever = 1;
331 if (rs->rs_scheduled == 0) {
332 list_move(&rs->rs_list, &b->rsb_replies);
333 rs->rs_scheduled = 1;
336 rs->rs_committed = 1;
337 spin_unlock(&rs->rs_lock);
341 * Reply batch finalization.
342 * Dispatch remaining replies from the batch
343 * and release remaining spinlock.
347 static void rs_batch_fini(struct rs_batch *b)
349 if (b->rsb_svcpt != NULL) {
350 rs_batch_dispatch(b);
351 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
355 #define DECLARE_RS_BATCH(b) struct rs_batch b
359 * Put reply state into a queue for processing because we received
360 * ACK from the client
362 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
364 struct ptlrpc_hr_thread *hrt;
367 LASSERT(list_empty(&rs->rs_list));
369 hrt = ptlrpc_hr_select(rs->rs_svcpt);
371 spin_lock(&hrt->hrt_lock);
372 list_add_tail(&rs->rs_list, &hrt->hrt_queue);
373 spin_unlock(&hrt->hrt_lock);
375 wake_up(&hrt->hrt_waitq);
380 ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs)
384 assert_spin_locked(&rs->rs_svcpt->scp_rep_lock);
385 assert_spin_locked(&rs->rs_lock);
386 LASSERT (rs->rs_difficult);
387 rs->rs_scheduled_ever = 1; /* flag any notification attempt */
389 if (rs->rs_scheduled) { /* being set up or already notified */
394 rs->rs_scheduled = 1;
395 list_del_init(&rs->rs_list);
396 ptlrpc_dispatch_difficult_reply(rs);
399 EXPORT_SYMBOL(ptlrpc_schedule_difficult_reply);
401 void ptlrpc_commit_replies(struct obd_export *exp)
403 struct ptlrpc_reply_state *rs, *nxt;
404 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);
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 than
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 if (cpumask_weight(topology_sibling_cpumask(smp_processor_id())) > 1) {
560 /* weight is # of HTs */
561 /* depress thread factor for hyper-thread */
562 factor = factor - (factor >> 1) + (factor >> 3);
565 weight = cfs_cpt_weight(svc->srv_cptable, 0);
567 for (; factor > 0 && weight > 0; factor--, weight -= fade)
568 nthrs += min(weight, fade) * factor;
571 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
572 nthrs = max(tc->tc_nthrs_base,
573 tc->tc_nthrs_max / svc->srv_ncpts);
576 nthrs = max(nthrs, tc->tc_nthrs_init);
577 svc->srv_nthrs_cpt_limit = nthrs;
578 svc->srv_nthrs_cpt_init = init;
580 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
581 CDEBUG(D_OTHER, "%s: This service may have more threads (%d) "
582 "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 /* acitve 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 setup_timer(&svcpt->scp_at_timer, ptlrpc_at_timer,
647 (unsigned long)svcpt);
649 /* At SOW, service time should be quick; 10s seems generous. If client
650 * timeout is less than this, we'll be sending an early reply. */
651 at_init(&svcpt->scp_at_estimate, 10, 0);
653 /* assign this before call ptlrpc_grow_req_bufs */
654 svcpt->scp_service = svc;
655 /* Now allocate the request buffers, but don't post them now */
656 rc = ptlrpc_grow_req_bufs(svcpt, 0);
657 /* We shouldn't be under memory pressure at startup, so
658 * fail if we can't allocate all our buffers at this time. */
665 if (array->paa_reqs_count != NULL) {
666 OBD_FREE(array->paa_reqs_count, sizeof(__u32) * size);
667 array->paa_reqs_count = NULL;
670 if (array->paa_reqs_array != NULL) {
671 OBD_FREE(array->paa_reqs_array,
672 sizeof(struct list_head) * array->paa_size);
673 array->paa_reqs_array = NULL;
680 * Initialize service on a given portal.
681 * This includes starting serving threads , allocating and posting rqbds and
684 struct ptlrpc_service *
685 ptlrpc_register_service(struct ptlrpc_service_conf *conf,
686 struct proc_dir_entry *proc_entry)
688 struct ptlrpc_service_cpt_conf *cconf = &conf->psc_cpt;
689 struct ptlrpc_service *service;
690 struct ptlrpc_service_part *svcpt;
691 struct cfs_cpt_table *cptable;
699 LASSERT(conf->psc_buf.bc_nbufs > 0);
700 LASSERT(conf->psc_buf.bc_buf_size >=
701 conf->psc_buf.bc_req_max_size + SPTLRPC_MAX_PAYLOAD);
702 LASSERT(conf->psc_thr.tc_ctx_tags != 0);
704 cptable = cconf->cc_cptable;
706 cptable = cfs_cpt_table;
708 if (!conf->psc_thr.tc_cpu_affinity) {
711 ncpts = cfs_cpt_number(cptable);
712 if (cconf->cc_pattern != NULL) {
713 struct cfs_expr_list *el;
715 rc = cfs_expr_list_parse(cconf->cc_pattern,
716 strlen(cconf->cc_pattern),
719 CERROR("%s: invalid CPT pattern string: %s",
720 conf->psc_name, cconf->cc_pattern);
721 RETURN(ERR_PTR(-EINVAL));
724 rc = cfs_expr_list_values(el, ncpts, &cpts);
725 cfs_expr_list_free(el);
727 CERROR("%s: failed to parse CPT array %s: %d\n",
728 conf->psc_name, cconf->cc_pattern, rc);
730 OBD_FREE(cpts, sizeof(*cpts) * ncpts);
731 RETURN(ERR_PTR(rc < 0 ? rc : -EINVAL));
737 OBD_ALLOC(service, offsetof(struct ptlrpc_service, srv_parts[ncpts]));
738 if (service == NULL) {
740 OBD_FREE(cpts, sizeof(*cpts) * ncpts);
741 RETURN(ERR_PTR(-ENOMEM));
744 service->srv_cptable = cptable;
745 service->srv_cpts = cpts;
746 service->srv_ncpts = ncpts;
748 service->srv_cpt_bits = 0; /* it's zero already, easy to read... */
749 while ((1 << service->srv_cpt_bits) < cfs_cpt_number(cptable))
750 service->srv_cpt_bits++;
753 spin_lock_init(&service->srv_lock);
754 service->srv_name = conf->psc_name;
755 service->srv_watchdog_factor = conf->psc_watchdog_factor;
756 INIT_LIST_HEAD(&service->srv_list); /* for safty of cleanup */
758 /* buffer configuration */
759 service->srv_nbuf_per_group = test_req_buffer_pressure ?
760 1 : conf->psc_buf.bc_nbufs;
761 service->srv_max_req_size = conf->psc_buf.bc_req_max_size +
763 service->srv_buf_size = conf->psc_buf.bc_buf_size;
764 service->srv_rep_portal = conf->psc_buf.bc_rep_portal;
765 service->srv_req_portal = conf->psc_buf.bc_req_portal;
767 /* Increase max reply size to next power of two */
768 service->srv_max_reply_size = 1;
769 while (service->srv_max_reply_size <
770 conf->psc_buf.bc_rep_max_size + SPTLRPC_MAX_PAYLOAD)
771 service->srv_max_reply_size <<= 1;
773 service->srv_thread_name = conf->psc_thr.tc_thr_name;
774 service->srv_ctx_tags = conf->psc_thr.tc_ctx_tags;
775 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
776 service->srv_ops = conf->psc_ops;
778 for (i = 0; i < ncpts; i++) {
779 if (!conf->psc_thr.tc_cpu_affinity)
782 cpt = cpts != NULL ? cpts[i] : i;
784 OBD_CPT_ALLOC(svcpt, cptable, cpt, sizeof(*svcpt));
786 GOTO(failed, rc = -ENOMEM);
788 service->srv_parts[i] = svcpt;
789 rc = ptlrpc_service_part_init(service, svcpt, cpt);
794 ptlrpc_server_nthreads_check(service, conf);
796 rc = LNetSetLazyPortal(service->srv_req_portal);
799 mutex_lock(&ptlrpc_all_services_mutex);
800 list_add(&service->srv_list, &ptlrpc_all_services);
801 mutex_unlock(&ptlrpc_all_services_mutex);
803 if (proc_entry != NULL)
804 ptlrpc_lprocfs_register_service(proc_entry, service);
806 rc = ptlrpc_service_nrs_setup(service);
810 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
811 service->srv_name, service->srv_req_portal);
813 rc = ptlrpc_start_threads(service);
815 CERROR("Failed to start threads for service %s: %d\n",
816 service->srv_name, rc);
822 ptlrpc_unregister_service(service);
825 EXPORT_SYMBOL(ptlrpc_register_service);
828 * to actually free the request, must be called without holding svc_lock.
829 * note it's caller's responsibility to unlink req->rq_list.
831 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
833 LASSERT(atomic_read(&req->rq_refcount) == 0);
834 LASSERT(list_empty(&req->rq_timed_list));
836 /* DEBUG_REQ() assumes the reply state of a request with a valid
837 * ref will not be destroyed until that reference is dropped. */
838 ptlrpc_req_drop_rs(req);
840 sptlrpc_svc_ctx_decref(req);
842 if (req != &req->rq_rqbd->rqbd_req) {
843 /* NB request buffers use an embedded
844 * req if the incoming req unlinked the
845 * MD; this isn't one of them! */
846 ptlrpc_request_cache_free(req);
851 * drop a reference count of the request. if it reaches 0, we either
852 * put it into history list, or free it immediately.
854 void ptlrpc_server_drop_request(struct ptlrpc_request *req)
856 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
857 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
858 struct ptlrpc_service *svc = svcpt->scp_service;
860 struct list_head *tmp;
861 struct list_head *nxt;
863 if (!atomic_dec_and_test(&req->rq_refcount))
866 if (req->rq_session.lc_state == LCS_ENTERED) {
867 lu_context_exit(&req->rq_session);
868 lu_context_fini(&req->rq_session);
871 if (req->rq_at_linked) {
872 spin_lock(&svcpt->scp_at_lock);
873 /* recheck with lock, in case it's unlinked by
874 * ptlrpc_at_check_timed() */
875 if (likely(req->rq_at_linked))
876 ptlrpc_at_remove_timed(req);
877 spin_unlock(&svcpt->scp_at_lock);
880 LASSERT(list_empty(&req->rq_timed_list));
882 /* finalize request */
883 if (req->rq_export) {
884 class_export_put(req->rq_export);
885 req->rq_export = NULL;
888 spin_lock(&svcpt->scp_lock);
890 list_add(&req->rq_list, &rqbd->rqbd_reqs);
892 refcount = --(rqbd->rqbd_refcount);
894 /* request buffer is now idle: add to history */
895 list_del(&rqbd->rqbd_list);
897 list_add_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds);
898 svcpt->scp_hist_nrqbds++;
900 /* cull some history?
901 * I expect only about 1 or 2 rqbds need to be recycled here */
902 while (svcpt->scp_hist_nrqbds > svc->srv_hist_nrqbds_cpt_max) {
903 rqbd = list_entry(svcpt->scp_hist_rqbds.next,
904 struct ptlrpc_request_buffer_desc,
907 list_del(&rqbd->rqbd_list);
908 svcpt->scp_hist_nrqbds--;
910 /* remove rqbd's reqs from svc's req history while
911 * I've got the service lock */
912 list_for_each(tmp, &rqbd->rqbd_reqs) {
913 req = list_entry(tmp, struct ptlrpc_request,
915 /* Track the highest culled req seq */
916 if (req->rq_history_seq >
917 svcpt->scp_hist_seq_culled) {
918 svcpt->scp_hist_seq_culled =
921 list_del(&req->rq_history_list);
924 spin_unlock(&svcpt->scp_lock);
926 list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
927 req = list_entry(rqbd->rqbd_reqs.next,
928 struct ptlrpc_request,
930 list_del(&req->rq_list);
931 ptlrpc_server_free_request(req);
934 spin_lock(&svcpt->scp_lock);
936 * now all reqs including the embedded req has been
937 * disposed, schedule request buffer for re-use
938 * or free it to drain some in excess.
940 LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) == 0);
941 if (svcpt->scp_nrqbds_posted >=
942 svc->srv_nbuf_per_group &&
943 !test_req_buffer_pressure) {
944 /* like in ptlrpc_free_rqbd() */
945 svcpt->scp_nrqbds_total--;
946 OBD_FREE_LARGE(rqbd->rqbd_buffer,
950 list_add_tail(&rqbd->rqbd_list,
951 &svcpt->scp_rqbd_idle);
955 spin_unlock(&svcpt->scp_lock);
956 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
957 /* If we are low on memory, we are not interested in history */
958 list_del(&req->rq_list);
959 list_del_init(&req->rq_history_list);
961 /* Track the highest culled req seq */
962 if (req->rq_history_seq > svcpt->scp_hist_seq_culled)
963 svcpt->scp_hist_seq_culled = req->rq_history_seq;
965 spin_unlock(&svcpt->scp_lock);
967 ptlrpc_server_free_request(req);
969 spin_unlock(&svcpt->scp_lock);
973 /** Change request export and move hp request from old export to new */
974 void ptlrpc_request_change_export(struct ptlrpc_request *req,
975 struct obd_export *export)
977 if (req->rq_export != NULL) {
978 LASSERT(!list_empty(&req->rq_exp_list));
979 /* remove rq_exp_list from last export */
980 spin_lock_bh(&req->rq_export->exp_rpc_lock);
981 list_del_init(&req->rq_exp_list);
982 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
983 /* export has one reference already, so it`s safe to
984 * add req to export queue here and get another
985 * reference for request later */
986 spin_lock_bh(&export->exp_rpc_lock);
987 if (req->rq_ops != NULL) /* hp request */
988 list_add(&req->rq_exp_list, &export->exp_hp_rpcs);
990 list_add(&req->rq_exp_list, &export->exp_reg_rpcs);
991 spin_unlock_bh(&export->exp_rpc_lock);
993 class_export_rpc_dec(req->rq_export);
994 class_export_put(req->rq_export);
997 /* request takes one export refcount */
998 req->rq_export = class_export_get(export);
999 class_export_rpc_inc(export);
1005 * to finish a request: stop sending more early replies, and release
1008 static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt,
1009 struct ptlrpc_request *req)
1011 ptlrpc_server_hpreq_fini(req);
1013 ptlrpc_server_drop_request(req);
1017 * to finish an active request: stop sending more early replies, and release
1018 * the request. should be called after we finished handling the request.
1020 static void ptlrpc_server_finish_active_request(
1021 struct ptlrpc_service_part *svcpt,
1022 struct ptlrpc_request *req)
1024 spin_lock(&svcpt->scp_req_lock);
1025 ptlrpc_nrs_req_stop_nolock(req);
1026 svcpt->scp_nreqs_active--;
1028 svcpt->scp_nhreqs_active--;
1029 spin_unlock(&svcpt->scp_req_lock);
1031 ptlrpc_nrs_req_finalize(req);
1033 if (req->rq_export != NULL)
1034 class_export_rpc_dec(req->rq_export);
1036 ptlrpc_server_finish_request(svcpt, req);
1040 * This function makes sure dead exports are evicted in a timely manner.
1041 * This function is only called when some export receives a message (i.e.,
1042 * the network is up.)
1044 void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
1046 struct obd_export *oldest_exp;
1047 time_t oldest_time, new_time;
1053 /* Compensate for slow machines, etc, by faking our request time
1054 into the future. Although this can break the strict time-ordering
1055 of the list, we can be really lazy here - we don't have to evict
1056 at the exact right moment. Eventually, all silent exports
1057 will make it to the top of the list. */
1059 /* Do not pay attention on 1sec or smaller renewals. */
1060 new_time = cfs_time_current_sec() + extra_delay;
1061 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
1064 exp->exp_last_request_time = new_time;
1066 /* exports may get disconnected from the chain even though the
1067 export has references, so we must keep the spin lock while
1068 manipulating the lists */
1069 spin_lock(&exp->exp_obd->obd_dev_lock);
1071 if (list_empty(&exp->exp_obd_chain_timed)) {
1072 /* this one is not timed */
1073 spin_unlock(&exp->exp_obd->obd_dev_lock);
1077 list_move_tail(&exp->exp_obd_chain_timed,
1078 &exp->exp_obd->obd_exports_timed);
1080 oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
1081 struct obd_export, exp_obd_chain_timed);
1082 oldest_time = oldest_exp->exp_last_request_time;
1083 spin_unlock(&exp->exp_obd->obd_dev_lock);
1085 if (exp->exp_obd->obd_recovering) {
1086 /* be nice to everyone during recovery */
1091 /* Note - racing to start/reset the obd_eviction timer is safe */
1092 if (exp->exp_obd->obd_eviction_timer == 0) {
1093 /* Check if the oldest entry is expired. */
1094 if (cfs_time_current_sec() > (oldest_time + PING_EVICT_TIMEOUT +
1096 /* We need a second timer, in case the net was down and
1097 * it just came back. Since the pinger may skip every
1098 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
1099 * we better wait for 3. */
1100 exp->exp_obd->obd_eviction_timer =
1101 cfs_time_current_sec() + 3 * PING_INTERVAL;
1102 CDEBUG(D_HA, "%s: Think about evicting %s from %ld\n",
1103 exp->exp_obd->obd_name,
1104 obd_export_nid2str(oldest_exp), oldest_time);
1107 if (cfs_time_current_sec() >
1108 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
1109 /* The evictor won't evict anyone who we've heard from
1110 * recently, so we don't have to check before we start
1112 if (!ping_evictor_wake(exp))
1113 exp->exp_obd->obd_eviction_timer = 0;
1121 * Sanity check request \a req.
1122 * Return 0 if all is ok, error code otherwise.
1124 static int ptlrpc_check_req(struct ptlrpc_request *req)
1126 struct obd_device *obd = req->rq_export->exp_obd;
1129 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
1130 req->rq_export->exp_conn_cnt)) {
1131 DEBUG_REQ(D_RPCTRACE, req,
1132 "DROPPING req from old connection %d < %d",
1133 lustre_msg_get_conn_cnt(req->rq_reqmsg),
1134 req->rq_export->exp_conn_cnt);
1137 if (unlikely(obd == NULL || obd->obd_fail)) {
1138 /* Failing over, don't handle any more reqs,
1139 * send error response instead. */
1140 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
1141 req, (obd != NULL) ? obd->obd_name : "unknown");
1143 } else if (lustre_msg_get_flags(req->rq_reqmsg) &
1144 (MSG_REPLAY | MSG_REQ_REPLAY_DONE) &&
1145 !obd->obd_recovering) {
1146 DEBUG_REQ(D_ERROR, req,
1147 "Invalid replay without recovery");
1148 class_fail_export(req->rq_export);
1150 } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 &&
1151 !obd->obd_recovering) {
1152 DEBUG_REQ(D_ERROR, req, "Invalid req with transno "
1153 "%llu without recovery",
1154 lustre_msg_get_transno(req->rq_reqmsg));
1155 class_fail_export(req->rq_export);
1159 if (unlikely(rc < 0)) {
1160 req->rq_status = rc;
1166 static void ptlrpc_at_set_timer(struct ptlrpc_service_part *svcpt)
1168 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1171 if (array->paa_count == 0) {
1172 del_timer(&svcpt->scp_at_timer);
1176 /* Set timer for closest deadline */
1177 next = (__s32)(array->paa_deadline - ktime_get_real_seconds() -
1180 ptlrpc_at_timer((unsigned long)svcpt);
1182 mod_timer(&svcpt->scp_at_timer, cfs_time_shift(next));
1183 CDEBUG(D_INFO, "armed %s at %+ds\n",
1184 svcpt->scp_service->srv_name, next);
1188 /* Add rpc to early reply check list */
1189 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
1191 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1192 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1193 struct ptlrpc_request *rq = NULL;
1199 if (req->rq_no_reply)
1202 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
1205 spin_lock(&svcpt->scp_at_lock);
1206 LASSERT(list_empty(&req->rq_timed_list));
1208 div_u64_rem(req->rq_deadline, array->paa_size, &index);
1209 if (array->paa_reqs_count[index] > 0) {
1210 /* latest rpcs will have the latest deadlines in the list,
1211 * so search backward. */
1212 list_for_each_entry_reverse(rq,
1213 &array->paa_reqs_array[index],
1215 if (req->rq_deadline >= rq->rq_deadline) {
1216 list_add(&req->rq_timed_list,
1217 &rq->rq_timed_list);
1223 /* Add the request at the head of the list */
1224 if (list_empty(&req->rq_timed_list))
1225 list_add(&req->rq_timed_list,
1226 &array->paa_reqs_array[index]);
1228 spin_lock(&req->rq_lock);
1229 req->rq_at_linked = 1;
1230 spin_unlock(&req->rq_lock);
1231 req->rq_at_index = index;
1232 array->paa_reqs_count[index]++;
1234 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
1235 array->paa_deadline = req->rq_deadline;
1236 ptlrpc_at_set_timer(svcpt);
1238 spin_unlock(&svcpt->scp_at_lock);
1244 ptlrpc_at_remove_timed(struct ptlrpc_request *req)
1246 struct ptlrpc_at_array *array;
1248 array = &req->rq_rqbd->rqbd_svcpt->scp_at_array;
1250 /* NB: must call with hold svcpt::scp_at_lock */
1251 LASSERT(!list_empty(&req->rq_timed_list));
1252 list_del_init(&req->rq_timed_list);
1254 spin_lock(&req->rq_lock);
1255 req->rq_at_linked = 0;
1256 spin_unlock(&req->rq_lock);
1258 array->paa_reqs_count[req->rq_at_index]--;
1263 * Attempt to extend the request deadline by sending an early reply to the
1266 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
1268 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1269 struct ptlrpc_request *reqcopy;
1270 struct lustre_msg *reqmsg;
1271 time64_t olddl = req->rq_deadline - ktime_get_real_seconds();
1277 if (CFS_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_RECONNECT)) {
1278 /* don't send early reply */
1282 /* deadline is when the client expects us to reply, margin is the
1283 difference between clients' and servers' expectations */
1284 DEBUG_REQ(D_ADAPTTO, req,
1285 "%ssending early reply (deadline %+llds, margin %+llds) for "
1286 "%d+%d", AT_OFF ? "AT off - not " : "",
1287 (s64)olddl, (s64)(olddl - at_get(&svcpt->scp_at_estimate)),
1288 at_get(&svcpt->scp_at_estimate), at_extra);
1294 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+llds), "
1295 "not sending early reply. Consider increasing "
1296 "at_early_margin (%d)?", (s64)olddl, at_early_margin);
1298 /* Return an error so we're not re-added to the timed list. */
1302 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
1303 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
1304 "but no AT support");
1308 if (req->rq_export &&
1309 lustre_msg_get_flags(req->rq_reqmsg) &
1310 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1311 struct obd_device *obd_exp = req->rq_export->exp_obd;
1313 /* During recovery, we don't want to send too many early
1314 * replies, but on the other hand we want to make sure the
1315 * client has enough time to resend if the rpc is lost. So
1316 * during the recovery period send at least 4 early replies,
1317 * spacing them every at_extra if we can. at_estimate should
1318 * always equal this fixed value during recovery.
1320 /* Don't account request processing time into AT history
1321 * during recovery, it is not service time we need but
1322 * includes also waiting time for recovering clients
1324 newdl = min_t(time64_t, at_extra,
1325 obd_exp->obd_recovery_timeout / 4) +
1326 ktime_get_real_seconds();
1328 /* We want to extend the request deadline by at_extra seconds,
1329 * so we set our service estimate to reflect how much time has
1330 * passed since this request arrived plus an additional
1331 * at_extra seconds. The client will calculate the new deadline
1332 * based on this service estimate (plus some additional time to
1333 * account for network latency). See ptlrpc_at_recv_early_reply
1335 at_measured(&svcpt->scp_at_estimate, at_extra +
1336 ktime_get_real_seconds() -
1337 req->rq_arrival_time.tv_sec);
1338 newdl = req->rq_arrival_time.tv_sec +
1339 at_get(&svcpt->scp_at_estimate);
1342 /* Check to see if we've actually increased the deadline -
1343 * we may be past adaptive_max */
1344 if (req->rq_deadline >= newdl) {
1345 DEBUG_REQ(D_WARNING, req, "Couldn't add any time (%lld/%lld), not sending early reply\n",
1346 (s64)olddl, (s64)(newdl - ktime_get_real_seconds()));
1350 reqcopy = ptlrpc_request_cache_alloc(GFP_NOFS);
1351 if (reqcopy == NULL)
1353 OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1355 GOTO(out_free, rc = -ENOMEM);
1358 reqcopy->rq_reply_state = NULL;
1359 reqcopy->rq_rep_swab_mask = 0;
1360 reqcopy->rq_pack_bulk = 0;
1361 reqcopy->rq_pack_udesc = 0;
1362 reqcopy->rq_packed_final = 0;
1363 sptlrpc_svc_ctx_addref(reqcopy);
1364 /* We only need the reqmsg for the magic */
1365 reqcopy->rq_reqmsg = reqmsg;
1366 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1369 * tgt_brw_read() and tgt_brw_write() may have decided not to reply.
1370 * Without this check, we would fail the rq_no_reply assertion in
1371 * ptlrpc_send_reply().
1373 if (reqcopy->rq_no_reply)
1374 GOTO(out, rc = -ETIMEDOUT);
1376 LASSERT(atomic_read(&req->rq_refcount));
1377 /** if it is last refcount then early reply isn't needed */
1378 if (atomic_read(&req->rq_refcount) == 1) {
1379 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
1380 "abort sending early reply\n");
1381 GOTO(out, rc = -EINVAL);
1384 /* Connection ref */
1385 reqcopy->rq_export = class_conn2export(
1386 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1387 if (reqcopy->rq_export == NULL)
1388 GOTO(out, rc = -ENODEV);
1391 class_export_rpc_inc(reqcopy->rq_export);
1392 if (reqcopy->rq_export->exp_obd &&
1393 reqcopy->rq_export->exp_obd->obd_fail)
1394 GOTO(out_put, rc = -ENODEV);
1396 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1400 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1403 /* Adjust our own deadline to what we told the client */
1404 req->rq_deadline = newdl;
1405 req->rq_early_count++; /* number sent, server side */
1407 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1410 /* Free the (early) reply state from lustre_pack_reply.
1411 (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
1412 ptlrpc_req_drop_rs(reqcopy);
1415 class_export_rpc_dec(reqcopy->rq_export);
1416 class_export_put(reqcopy->rq_export);
1418 sptlrpc_svc_ctx_decref(reqcopy);
1419 OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1421 ptlrpc_request_cache_free(reqcopy);
1425 /* Send early replies to everybody expiring within at_early_margin
1426 asking for at_extra time */
1427 static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1429 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1430 struct ptlrpc_request *rq, *n;
1431 struct list_head work_list;
1434 time64_t now = ktime_get_real_seconds();
1435 cfs_duration_t delay;
1436 int first, counter = 0;
1439 spin_lock(&svcpt->scp_at_lock);
1440 if (svcpt->scp_at_check == 0) {
1441 spin_unlock(&svcpt->scp_at_lock);
1444 delay = cfs_time_sub(cfs_time_current(), svcpt->scp_at_checktime);
1445 svcpt->scp_at_check = 0;
1447 if (array->paa_count == 0) {
1448 spin_unlock(&svcpt->scp_at_lock);
1452 /* The timer went off, but maybe the nearest rpc already completed. */
1453 first = array->paa_deadline - now;
1454 if (first > at_early_margin) {
1455 /* We've still got plenty of time. Reset the timer. */
1456 ptlrpc_at_set_timer(svcpt);
1457 spin_unlock(&svcpt->scp_at_lock);
1461 /* We're close to a timeout, and we don't know how much longer the
1462 server will take. Send early replies to everyone expiring soon. */
1463 INIT_LIST_HEAD(&work_list);
1465 div_u64_rem(array->paa_deadline, array->paa_size, &index);
1466 count = array->paa_count;
1468 count -= array->paa_reqs_count[index];
1469 list_for_each_entry_safe(rq, n,
1470 &array->paa_reqs_array[index],
1472 if (rq->rq_deadline > now + at_early_margin) {
1473 /* update the earliest deadline */
1474 if (deadline == -1 ||
1475 rq->rq_deadline < deadline)
1476 deadline = rq->rq_deadline;
1480 ptlrpc_at_remove_timed(rq);
1482 * ptlrpc_server_drop_request() may drop
1483 * refcount to 0 already. Let's check this and
1484 * don't add entry to work_list
1486 if (likely(atomic_inc_not_zero(&rq->rq_refcount)))
1487 list_add(&rq->rq_timed_list, &work_list);
1491 if (++index >= array->paa_size)
1494 array->paa_deadline = deadline;
1495 /* we have a new earliest deadline, restart the timer */
1496 ptlrpc_at_set_timer(svcpt);
1498 spin_unlock(&svcpt->scp_at_lock);
1500 CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
1501 "replies\n", first, at_extra, counter);
1503 /* We're already past request deadlines before we even get a
1504 chance to send early replies */
1505 LCONSOLE_WARN("%s: This server is not able to keep up with "
1506 "request traffic (cpu-bound).\n",
1507 svcpt->scp_service->srv_name);
1508 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, delay=%ld(jiff)\n",
1509 counter, svcpt->scp_nreqs_incoming,
1510 svcpt->scp_nreqs_active,
1511 at_get(&svcpt->scp_at_estimate), delay);
1514 /* we took additional refcount so entries can't be deleted from list, no
1515 * locking is needed */
1516 while (!list_empty(&work_list)) {
1517 rq = list_entry(work_list.next, struct ptlrpc_request,
1519 list_del_init(&rq->rq_timed_list);
1521 if (ptlrpc_at_send_early_reply(rq) == 0)
1522 ptlrpc_at_add_timed(rq);
1524 ptlrpc_server_drop_request(rq);
1527 RETURN(1); /* return "did_something" for liblustre */
1530 /* Check if we are already handling earlier incarnation of this request.
1531 * Called under &req->rq_export->exp_rpc_lock locked */
1532 static int ptlrpc_server_check_resend_in_progress(struct ptlrpc_request *req)
1534 struct ptlrpc_request *tmp = NULL;
1536 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ||
1537 (atomic_read(&req->rq_export->exp_rpc_count) == 0))
1540 /* bulk request are aborted upon reconnect, don't try to
1542 if (req->rq_bulk_write || req->rq_bulk_read)
1545 /* This list should not be longer than max_requests in
1546 * flights on the client, so it is not all that long.
1547 * Also we only hit this codepath in case of a resent
1548 * request which makes it even more rarely hit */
1549 list_for_each_entry(tmp, &req->rq_export->exp_reg_rpcs,
1551 /* Found duplicate one */
1552 if (tmp->rq_xid == req->rq_xid)
1555 list_for_each_entry(tmp, &req->rq_export->exp_hp_rpcs,
1557 /* Found duplicate one */
1558 if (tmp->rq_xid == req->rq_xid)
1564 DEBUG_REQ(D_HA, req, "Found duplicate req in processing");
1565 DEBUG_REQ(D_HA, tmp, "Request being processed");
1570 * Check if a request should be assigned with a high priority.
1572 * \retval < 0: error occurred
1573 * 0: normal RPC request
1574 * +1: high priority request
1576 static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt,
1577 struct ptlrpc_request *req)
1582 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL) {
1583 rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
1590 if (req->rq_export != NULL && req->rq_ops != NULL) {
1591 /* Perform request specific check. We should do this
1592 * check before the request is added into exp_hp_rpcs
1593 * list otherwise it may hit swab race at LU-1044. */
1594 if (req->rq_ops->hpreq_check != NULL) {
1595 rc = req->rq_ops->hpreq_check(req);
1596 if (rc == -ESTALE) {
1597 req->rq_status = rc;
1600 /** can only return error,
1601 * 0 for normal request,
1602 * or 1 for high priority request */
1610 /** Remove the request from the export list. */
1611 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req)
1614 if (req->rq_export) {
1615 /* refresh lock timeout again so that client has more
1616 * room to send lock cancel RPC. */
1617 if (req->rq_ops && req->rq_ops->hpreq_fini)
1618 req->rq_ops->hpreq_fini(req);
1620 spin_lock_bh(&req->rq_export->exp_rpc_lock);
1621 list_del_init(&req->rq_exp_list);
1622 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1627 static int ptlrpc_hpreq_check(struct ptlrpc_request *req)
1632 static struct ptlrpc_hpreq_ops ptlrpc_hpreq_common = {
1633 .hpreq_check = ptlrpc_hpreq_check,
1636 /* Hi-Priority RPC check by RPC operation code. */
1637 int ptlrpc_hpreq_handler(struct ptlrpc_request *req)
1639 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1641 /* Check for export to let only reconnects for not yet evicted
1642 * export to become a HP rpc. */
1643 if ((req->rq_export != NULL) &&
1644 (opc == OBD_PING || opc == MDS_CONNECT || opc == OST_CONNECT))
1645 req->rq_ops = &ptlrpc_hpreq_common;
1649 EXPORT_SYMBOL(ptlrpc_hpreq_handler);
1651 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1652 struct ptlrpc_request *req)
1658 rc = ptlrpc_server_hpreq_init(svcpt, req);
1663 ptlrpc_nrs_req_initialize(svcpt, req, hp);
1665 if (req->rq_export != NULL) {
1666 struct obd_export *exp = req->rq_export;
1668 /* do search for duplicated xid and the adding to the list
1670 spin_lock_bh(&exp->exp_rpc_lock);
1671 rc = ptlrpc_server_check_resend_in_progress(req);
1673 spin_unlock_bh(&exp->exp_rpc_lock);
1675 ptlrpc_nrs_req_finalize(req);
1679 if (hp || req->rq_ops != NULL)
1680 list_add(&req->rq_exp_list, &exp->exp_hp_rpcs);
1682 list_add(&req->rq_exp_list, &exp->exp_reg_rpcs);
1683 spin_unlock_bh(&exp->exp_rpc_lock);
1686 /* the current thread is not the processing thread for this request
1687 * since that, but request is in exp_hp_list and can be find there.
1688 * Remove all relations between request and old thread. */
1689 req->rq_svc_thread->t_env->le_ses = NULL;
1690 req->rq_svc_thread = NULL;
1691 req->rq_session.lc_thread = NULL;
1693 ptlrpc_nrs_req_add(svcpt, req, hp);
1699 * Allow to handle high priority request
1700 * User can call it w/o any lock but need to hold
1701 * ptlrpc_service_part::scp_req_lock to get reliable result
1703 static bool ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1706 int running = svcpt->scp_nthrs_running;
1708 if (!nrs_svcpt_has_hp(svcpt))
1714 if (ptlrpc_nrs_req_throttling_nolock(svcpt, true))
1717 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1718 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1719 /* leave just 1 thread for normal RPCs */
1720 running = PTLRPC_NTHRS_INIT;
1721 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1725 if (svcpt->scp_nreqs_active >= running - 1)
1728 if (svcpt->scp_nhreqs_active == 0)
1731 return !ptlrpc_nrs_req_pending_nolock(svcpt, false) ||
1732 svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1735 static bool ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1738 return ptlrpc_server_allow_high(svcpt, force) &&
1739 ptlrpc_nrs_req_pending_nolock(svcpt, true);
1743 * Only allow normal priority requests on a service that has a high-priority
1744 * queue if forced (i.e. cleanup), if there are other high priority requests
1745 * already being processed (i.e. those threads can service more high-priority
1746 * requests), or if there are enough idle threads that a later thread can do
1747 * a high priority request.
1748 * User can call it w/o any lock but need to hold
1749 * ptlrpc_service_part::scp_req_lock to get reliable result
1751 static bool ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1754 int running = svcpt->scp_nthrs_running;
1755 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1756 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1757 /* leave just 1 thread for normal RPCs */
1758 running = PTLRPC_NTHRS_INIT;
1759 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1766 if (ptlrpc_nrs_req_throttling_nolock(svcpt, false))
1769 if (svcpt->scp_nreqs_active < running - 2)
1772 if (svcpt->scp_nreqs_active >= running - 1)
1775 return svcpt->scp_nhreqs_active > 0 || !nrs_svcpt_has_hp(svcpt);
1778 static bool ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1781 return ptlrpc_server_allow_normal(svcpt, force) &&
1782 ptlrpc_nrs_req_pending_nolock(svcpt, false);
1786 * Returns true if there are requests available in incoming
1787 * request queue for processing and it is allowed to fetch them.
1788 * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1789 * to get reliable result
1790 * \see ptlrpc_server_allow_normal
1791 * \see ptlrpc_server_allow high
1794 ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt, bool force)
1796 return ptlrpc_server_high_pending(svcpt, force) ||
1797 ptlrpc_server_normal_pending(svcpt, force);
1801 * Fetch a request for processing from queue of unprocessed requests.
1802 * Favors high-priority requests.
1803 * Returns a pointer to fetched request.
1805 static struct ptlrpc_request *
1806 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force)
1808 struct ptlrpc_request *req = NULL;
1811 spin_lock(&svcpt->scp_req_lock);
1813 if (ptlrpc_server_high_pending(svcpt, force)) {
1814 req = ptlrpc_nrs_req_get_nolock(svcpt, true, force);
1816 svcpt->scp_hreq_count++;
1821 if (ptlrpc_server_normal_pending(svcpt, force)) {
1822 req = ptlrpc_nrs_req_get_nolock(svcpt, false, force);
1824 svcpt->scp_hreq_count = 0;
1829 spin_unlock(&svcpt->scp_req_lock);
1833 svcpt->scp_nreqs_active++;
1835 svcpt->scp_nhreqs_active++;
1837 spin_unlock(&svcpt->scp_req_lock);
1839 if (likely(req->rq_export))
1840 class_export_rpc_inc(req->rq_export);
1846 * Handle freshly incoming reqs, add to timed early reply list,
1847 * pass on to regular request queue.
1848 * All incoming requests pass through here before getting into
1849 * ptlrpc_server_handle_req later on.
1852 ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt,
1853 struct ptlrpc_thread *thread)
1855 struct ptlrpc_service *svc = svcpt->scp_service;
1856 struct ptlrpc_request *req;
1861 spin_lock(&svcpt->scp_lock);
1862 if (list_empty(&svcpt->scp_req_incoming)) {
1863 spin_unlock(&svcpt->scp_lock);
1867 req = list_entry(svcpt->scp_req_incoming.next,
1868 struct ptlrpc_request, rq_list);
1869 list_del_init(&req->rq_list);
1870 svcpt->scp_nreqs_incoming--;
1871 /* Consider this still a "queued" request as far as stats are
1873 spin_unlock(&svcpt->scp_lock);
1875 /* go through security check/transform */
1876 rc = sptlrpc_svc_unwrap_request(req);
1880 case SECSVC_COMPLETE:
1881 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1890 * for null-flavored rpc, msg has been unpacked by sptlrpc, although
1891 * redo it wouldn't be harmful.
1893 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
1894 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
1896 CERROR("error unpacking request: ptl %d from %s "
1897 "x%llu\n", svc->srv_req_portal,
1898 libcfs_id2str(req->rq_peer), req->rq_xid);
1903 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1905 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
1906 "%llu\n", svc->srv_req_portal,
1907 libcfs_id2str(req->rq_peer), req->rq_xid);
1911 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
1912 lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
1913 CERROR("drop incoming rpc opc %u, x%llu\n",
1914 cfs_fail_val, req->rq_xid);
1919 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1920 CERROR("wrong packet type received (type=%u) from %s\n",
1921 lustre_msg_get_type(req->rq_reqmsg),
1922 libcfs_id2str(req->rq_peer));
1926 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1930 req->rq_bulk_write = 1;
1934 case MGS_CONFIG_READ:
1935 req->rq_bulk_read = 1;
1939 CDEBUG(D_RPCTRACE, "got req x%llu\n", req->rq_xid);
1941 req->rq_export = class_conn2export(
1942 lustre_msg_get_handle(req->rq_reqmsg));
1943 if (req->rq_export) {
1944 rc = ptlrpc_check_req(req);
1946 rc = sptlrpc_target_export_check(req->rq_export, req);
1948 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
1949 "illegal security flavor,");
1954 ptlrpc_update_export_timer(req->rq_export, 0);
1957 /* req_in handling should/must be fast */
1958 if (ktime_get_real_seconds() - req->rq_arrival_time.tv_sec > 5)
1959 DEBUG_REQ(D_WARNING, req, "Slow req_in handling %llds",
1960 (s64)(ktime_get_real_seconds() -
1961 req->rq_arrival_time.tv_sec));
1963 /* Set rpc server deadline and add it to the timed list */
1964 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1965 MSGHDR_AT_SUPPORT) ?
1966 /* The max time the client expects us to take */
1967 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1969 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1970 if (unlikely(deadline == 0)) {
1971 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1975 /* Skip early reply */
1976 if (OBD_FAIL_PRECHECK(OBD_FAIL_MDS_RESEND))
1977 req->rq_deadline += obd_timeout;
1979 req->rq_svc_thread = thread;
1980 if (thread != NULL) {
1981 /* initialize request session, it is needed for request
1982 * processing by target */
1983 rc = lu_context_init(&req->rq_session, LCT_SERVER_SESSION |
1986 CERROR("%s: failure to initialize session: rc = %d\n",
1987 thread->t_name, rc);
1990 req->rq_session.lc_thread = thread;
1991 lu_context_enter(&req->rq_session);
1992 thread->t_env->le_ses = &req->rq_session;
1995 ptlrpc_at_add_timed(req);
1997 /* Move it over to the request processing queue */
1998 rc = ptlrpc_server_request_add(svcpt, req);
2002 wake_up(&svcpt->scp_waitq);
2006 ptlrpc_server_finish_request(svcpt, req);
2012 * Main incoming request handling logic.
2013 * Calls handler function from service to do actual processing.
2016 ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
2017 struct ptlrpc_thread *thread)
2019 struct ptlrpc_service *svc = svcpt->scp_service;
2020 struct ptlrpc_request *request;
2030 request = ptlrpc_server_request_get(svcpt, false);
2031 if (request == NULL)
2034 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
2035 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
2036 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
2037 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
2039 if (unlikely(fail_opc)) {
2040 if (request->rq_export && request->rq_ops)
2041 OBD_FAIL_TIMEOUT(fail_opc, 4);
2044 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
2046 if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
2047 libcfs_debug_dumplog();
2049 work_start = ktime_get_real();
2050 arrived = timespec64_to_ktime(request->rq_arrival_time);
2051 timediff_usecs = ktime_us_delta(arrived, work_start);
2052 if (likely(svc->srv_stats != NULL)) {
2053 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
2055 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
2056 svcpt->scp_nreqs_incoming);
2057 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
2058 svcpt->scp_nreqs_active);
2059 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
2060 at_get(&svcpt->scp_at_estimate));
2063 if (likely(request->rq_export)) {
2064 if (unlikely(ptlrpc_check_req(request)))
2066 ptlrpc_update_export_timer(request->rq_export,
2067 timediff_usecs >> 19);
2070 /* Discard requests queued for longer than the deadline.
2071 The deadline is increased if we send an early reply. */
2072 if (ktime_get_real_seconds() > request->rq_deadline) {
2073 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s: deadline %lld:%llds ago\n",
2074 libcfs_id2str(request->rq_peer),
2075 request->rq_deadline -
2076 request->rq_arrival_time.tv_sec,
2077 ktime_get_real_seconds() - request->rq_deadline);
2081 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
2082 "%s:%s+%d:%d:x%llu:%s:%d\n", current_comm(),
2083 (request->rq_export ?
2084 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2085 (request->rq_export ?
2086 atomic_read(&request->rq_export->exp_refcount) : -99),
2087 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
2088 libcfs_id2str(request->rq_peer),
2089 lustre_msg_get_opc(request->rq_reqmsg));
2091 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
2092 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
2094 CDEBUG(D_NET, "got req %llu\n", request->rq_xid);
2096 /* re-assign request and sesson thread to the current one */
2097 request->rq_svc_thread = thread;
2098 if (thread != NULL) {
2099 LASSERT(request->rq_session.lc_thread == NULL);
2100 request->rq_session.lc_thread = thread;
2101 thread->t_env->le_ses = &request->rq_session;
2103 svc->srv_ops.so_req_handler(request);
2105 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
2108 if (unlikely(ktime_get_real_seconds() > request->rq_deadline)) {
2109 DEBUG_REQ(D_WARNING, request, "Request took longer than estimated (%lld:%llds); client may timeout.",
2110 request->rq_deadline -
2111 request->rq_arrival_time.tv_sec,
2112 ktime_get_real_seconds() - request->rq_deadline);
2115 work_end = ktime_get_real();
2116 timediff_usecs = ktime_us_delta(work_end, work_start);
2117 arrived_usecs = ktime_us_delta(work_end, arrived);
2118 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc %s:%s+%d:%d:x%llu:%s:%d Request procesed in %lldus (%lldus total) trans %llu rc %d/%d\n",
2120 (request->rq_export ?
2121 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2122 (request->rq_export ?
2123 atomic_read(&request->rq_export->exp_refcount) : -99),
2124 lustre_msg_get_status(request->rq_reqmsg),
2126 libcfs_id2str(request->rq_peer),
2127 lustre_msg_get_opc(request->rq_reqmsg),
2130 (request->rq_repmsg ?
2131 lustre_msg_get_transno(request->rq_repmsg) :
2132 request->rq_transno),
2134 (request->rq_repmsg ?
2135 lustre_msg_get_status(request->rq_repmsg) : -999));
2136 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
2137 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
2138 int opc = opcode_offset(op);
2139 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
2140 LASSERT(opc < LUSTRE_MAX_OPCODES);
2141 lprocfs_counter_add(svc->srv_stats,
2142 opc + EXTRA_MAX_OPCODES,
2146 if (unlikely(request->rq_early_count)) {
2147 DEBUG_REQ(D_ADAPTTO, request,
2148 "sent %d early replies before finishing in %llds",
2149 request->rq_early_count,
2150 arrived_usecs / USEC_PER_SEC);
2153 ptlrpc_server_finish_active_request(svcpt, request);
2159 * An internal function to process a single reply state object.
2162 ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
2164 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
2165 struct ptlrpc_service *svc = svcpt->scp_service;
2166 struct obd_export *exp;
2171 exp = rs->rs_export;
2173 LASSERT(rs->rs_difficult);
2174 LASSERT(rs->rs_scheduled);
2175 LASSERT(list_empty(&rs->rs_list));
2177 /* The disk commit callback holds exp_uncommitted_replies_lock while it
2178 * iterates over newly committed replies, removing them from
2179 * exp_uncommitted_replies. It then drops this lock and schedules the
2180 * replies it found for handling here.
2182 * We can avoid contention for exp_uncommitted_replies_lock between the
2183 * HRT threads and further commit callbacks by checking rs_committed
2184 * which is set in the commit callback while it holds both
2185 * rs_lock and exp_uncommitted_reples.
2187 * If we see rs_committed clear, the commit callback _may_ not have
2188 * handled this reply yet and we race with it to grab
2189 * exp_uncommitted_replies_lock before removing the reply from
2190 * exp_uncommitted_replies. Note that if we lose the race and the
2191 * reply has already been removed, list_del_init() is a noop.
2193 * If we see rs_committed set, we know the commit callback is handling,
2194 * or has handled this reply since store reordering might allow us to
2195 * see rs_committed set out of sequence. But since this is done
2196 * holding rs_lock, we can be sure it has all completed once we hold
2197 * rs_lock, which we do right next.
2199 if (!rs->rs_committed) {
2200 /* if rs was commited, no need to convert locks, don't check
2201 * rs_committed here because rs may never be added into
2202 * exp_uncommitted_replies and this flag never be set, see
2203 * target_send_reply() */
2204 if (rs->rs_convert_lock &&
2205 rs->rs_transno > exp->exp_last_committed) {
2206 struct ldlm_lock *lock;
2207 struct ldlm_lock *ack_locks[RS_MAX_LOCKS] = { NULL };
2209 spin_lock(&rs->rs_lock);
2210 if (rs->rs_convert_lock &&
2211 rs->rs_transno > exp->exp_last_committed) {
2212 nlocks = rs->rs_nlocks;
2213 while (nlocks-- > 0) {
2215 * NB don't assume rs is always handled
2216 * by the same service thread (see
2217 * ptlrpc_hr_select, so REP-ACK hr may
2218 * race with trans commit, while the
2219 * latter will release locks, get locks
2220 * here early to convert to COS mode
2223 lock = ldlm_handle2lock(
2224 &rs->rs_locks[nlocks]);
2226 ack_locks[nlocks] = lock;
2227 rs->rs_modes[nlocks] = LCK_COS;
2229 nlocks = rs->rs_nlocks;
2230 rs->rs_convert_lock = 0;
2231 /* clear rs_scheduled so that commit callback
2232 * can schedule again */
2233 rs->rs_scheduled = 0;
2234 spin_unlock(&rs->rs_lock);
2236 while (nlocks-- > 0) {
2237 lock = ack_locks[nlocks];
2238 ldlm_lock_downgrade(lock, LCK_COS);
2239 LDLM_LOCK_PUT(lock);
2243 spin_unlock(&rs->rs_lock);
2246 spin_lock(&exp->exp_uncommitted_replies_lock);
2247 list_del_init(&rs->rs_obd_list);
2248 spin_unlock(&exp->exp_uncommitted_replies_lock);
2251 spin_lock(&exp->exp_lock);
2252 /* Noop if removed already */
2253 list_del_init(&rs->rs_exp_list);
2254 spin_unlock(&exp->exp_lock);
2256 spin_lock(&rs->rs_lock);
2258 been_handled = rs->rs_handled;
2261 nlocks = rs->rs_nlocks; /* atomic "steal", but */
2262 rs->rs_nlocks = 0; /* locks still on rs_locks! */
2264 if (nlocks == 0 && !been_handled) {
2265 /* If we see this, we should already have seen the warning
2266 * in mds_steal_ack_locks() */
2267 CDEBUG(D_HA, "All locks stolen from rs %p x%lld.t%lld"
2270 rs->rs_xid, rs->rs_transno, rs->rs_opc,
2271 libcfs_nid2str(exp->exp_connection->c_peer.nid));
2274 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
2275 spin_unlock(&rs->rs_lock);
2277 if (!been_handled && rs->rs_on_net) {
2278 LNetMDUnlink(rs->rs_md_h);
2279 /* Ignore return code; we're racing with completion */
2282 while (nlocks-- > 0)
2283 ldlm_lock_decref(&rs->rs_locks[nlocks],
2284 rs->rs_modes[nlocks]);
2286 spin_lock(&rs->rs_lock);
2289 rs->rs_scheduled = 0;
2290 rs->rs_convert_lock = 0;
2292 if (!rs->rs_on_net) {
2294 spin_unlock(&rs->rs_lock);
2296 class_export_put (exp);
2297 rs->rs_export = NULL;
2298 ptlrpc_rs_decref(rs);
2299 if (atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
2300 svc->srv_is_stopping)
2301 wake_up_all(&svcpt->scp_waitq);
2305 /* still on the net; callback will schedule */
2306 spin_unlock(&rs->rs_lock);
2312 ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
2314 int avail = svcpt->scp_nrqbds_posted;
2315 int low_water = test_req_buffer_pressure ? 0 :
2316 svcpt->scp_service->srv_nbuf_per_group / 2;
2318 /* NB I'm not locking; just looking. */
2320 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
2321 * allowed the request history to grow out of control. We could put a
2322 * sanity check on that here and cull some history if we need the
2325 if (avail <= low_water)
2326 ptlrpc_grow_req_bufs(svcpt, 1);
2328 if (svcpt->scp_service->srv_stats) {
2329 lprocfs_counter_add(svcpt->scp_service->srv_stats,
2330 PTLRPC_REQBUF_AVAIL_CNTR, avail);
2335 ptlrpc_retry_rqbds(void *arg)
2337 struct ptlrpc_service_part *svcpt = (struct ptlrpc_service_part *)arg;
2339 svcpt->scp_rqbd_timeout = 0;
2344 ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
2346 return svcpt->scp_nreqs_active <
2347 svcpt->scp_nthrs_running - 1 -
2348 (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL);
2352 * allowed to create more threads
2353 * user can call it w/o any lock but need to hold
2354 * ptlrpc_service_part::scp_lock to get reliable result
2357 ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
2359 return svcpt->scp_nthrs_running +
2360 svcpt->scp_nthrs_starting <
2361 svcpt->scp_service->srv_nthrs_cpt_limit;
2365 * too many requests and allowed to create more threads
2368 ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
2370 return !ptlrpc_threads_enough(svcpt) &&
2371 ptlrpc_threads_increasable(svcpt);
2375 ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2377 return thread_is_stopping(thread) ||
2378 thread->t_svcpt->scp_service->srv_is_stopping;
2382 ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
2384 return !list_empty(&svcpt->scp_rqbd_idle) &&
2385 svcpt->scp_rqbd_timeout == 0;
2389 ptlrpc_at_check(struct ptlrpc_service_part *svcpt)
2391 return svcpt->scp_at_check;
2395 * requests wait on preprocessing
2396 * user can call it w/o any lock but need to hold
2397 * ptlrpc_service_part::scp_lock to get reliable result
2400 ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
2402 return !list_empty(&svcpt->scp_req_incoming);
2405 static __attribute__((__noinline__)) int
2406 ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
2407 struct ptlrpc_thread *thread)
2409 /* Don't exit while there are replies to be handled */
2410 struct l_wait_info lwi = LWI_TIMEOUT(svcpt->scp_rqbd_timeout,
2411 ptlrpc_retry_rqbds, svcpt);
2413 lc_watchdog_disable(thread->t_watchdog);
2417 l_wait_event_exclusive_head(svcpt->scp_waitq,
2418 ptlrpc_thread_stopping(thread) ||
2419 ptlrpc_server_request_incoming(svcpt) ||
2420 ptlrpc_server_request_pending(svcpt, false) ||
2421 ptlrpc_rqbd_pending(svcpt) ||
2422 ptlrpc_at_check(svcpt), &lwi);
2424 if (ptlrpc_thread_stopping(thread))
2427 lc_watchdog_touch(thread->t_watchdog,
2428 ptlrpc_server_get_timeout(svcpt));
2433 * Main thread body for service threads.
2434 * Waits in a loop waiting for new requests to process to appear.
2435 * Every time an incoming requests is added to its queue, a waitq
2436 * is woken up and one of the threads will handle it.
2438 static int ptlrpc_main(void *arg)
2440 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
2441 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2442 struct ptlrpc_service *svc = svcpt->scp_service;
2443 struct ptlrpc_reply_state *rs;
2444 struct group_info *ginfo = NULL;
2446 int counter = 0, rc = 0;
2449 thread->t_pid = current_pid();
2450 unshare_fs_struct();
2452 /* NB: we will call cfs_cpt_bind() for all threads, because we
2453 * might want to run lustre server only on a subset of system CPUs,
2454 * in that case ->scp_cpt is CFS_CPT_ANY */
2455 rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
2457 CWARN("%s: failed to bind %s on CPT %d\n",
2458 svc->srv_name, thread->t_name, svcpt->scp_cpt);
2461 ginfo = groups_alloc(0);
2467 set_current_groups(ginfo);
2468 put_group_info(ginfo);
2470 if (svc->srv_ops.so_thr_init != NULL) {
2471 rc = svc->srv_ops.so_thr_init(thread);
2482 rc = lu_context_init(&env->le_ctx,
2483 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2487 thread->t_env = env;
2488 env->le_ctx.lc_thread = thread;
2489 env->le_ctx.lc_cookie = 0x6;
2491 while (!list_empty(&svcpt->scp_rqbd_idle)) {
2492 rc = ptlrpc_server_post_idle_rqbds(svcpt);
2496 CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
2497 svc->srv_name, svcpt->scp_cpt, rc);
2501 /* Alloc reply state structure for this one */
2502 OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2508 spin_lock(&svcpt->scp_lock);
2510 LASSERT(thread_is_starting(thread));
2511 thread_clear_flags(thread, SVC_STARTING);
2513 LASSERT(svcpt->scp_nthrs_starting == 1);
2514 svcpt->scp_nthrs_starting--;
2516 /* SVC_STOPPING may already be set here if someone else is trying
2517 * to stop the service while this new thread has been dynamically
2518 * forked. We still set SVC_RUNNING to let our creator know that
2519 * we are now running, however we will exit as soon as possible */
2520 thread_add_flags(thread, SVC_RUNNING);
2521 svcpt->scp_nthrs_running++;
2522 spin_unlock(&svcpt->scp_lock);
2524 /* wake up our creator in case he's still waiting. */
2525 wake_up(&thread->t_ctl_waitq);
2527 thread->t_watchdog = lc_watchdog_add(ptlrpc_server_get_timeout(svcpt),
2530 spin_lock(&svcpt->scp_rep_lock);
2531 list_add(&rs->rs_list, &svcpt->scp_rep_idle);
2532 wake_up(&svcpt->scp_rep_waitq);
2533 spin_unlock(&svcpt->scp_rep_lock);
2535 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2536 svcpt->scp_nthrs_running);
2538 /* XXX maintain a list of all managed devices: insert here */
2539 while (!ptlrpc_thread_stopping(thread)) {
2540 if (ptlrpc_wait_event(svcpt, thread))
2543 ptlrpc_check_rqbd_pool(svcpt);
2545 if (ptlrpc_threads_need_create(svcpt)) {
2546 /* Ignore return code - we tried... */
2547 ptlrpc_start_thread(svcpt, 0);
2550 /* reset le_ses to initial state */
2552 /* Process all incoming reqs before handling any */
2553 if (ptlrpc_server_request_incoming(svcpt)) {
2554 lu_context_enter(&env->le_ctx);
2555 ptlrpc_server_handle_req_in(svcpt, thread);
2556 lu_context_exit(&env->le_ctx);
2558 /* but limit ourselves in case of flood */
2559 if (counter++ < 100)
2564 if (ptlrpc_at_check(svcpt))
2565 ptlrpc_at_check_timed(svcpt);
2567 if (ptlrpc_server_request_pending(svcpt, false)) {
2568 lu_context_enter(&env->le_ctx);
2569 ptlrpc_server_handle_request(svcpt, thread);
2570 lu_context_exit(&env->le_ctx);
2573 if (ptlrpc_rqbd_pending(svcpt) &&
2574 ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
2575 /* I just failed to repost request buffers.
2576 * Wait for a timeout (unless something else
2577 * happens) before I try again */
2578 svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
2579 CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
2580 svcpt->scp_nrqbds_posted);
2584 lc_watchdog_delete(thread->t_watchdog);
2585 thread->t_watchdog = NULL;
2589 * deconstruct service specific state created by ptlrpc_start_thread()
2591 if (svc->srv_ops.so_thr_done != NULL)
2592 svc->srv_ops.so_thr_done(thread);
2595 lu_context_fini(&env->le_ctx);
2599 CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n",
2600 thread, thread->t_pid, thread->t_id, rc);
2602 spin_lock(&svcpt->scp_lock);
2603 if (thread_test_and_clear_flags(thread, SVC_STARTING))
2604 svcpt->scp_nthrs_starting--;
2606 if (thread_test_and_clear_flags(thread, SVC_RUNNING)) {
2607 /* must know immediately */
2608 svcpt->scp_nthrs_running--;
2612 thread_add_flags(thread, SVC_STOPPED);
2614 wake_up(&thread->t_ctl_waitq);
2615 spin_unlock(&svcpt->scp_lock);
2620 static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
2621 struct list_head *replies)
2625 spin_lock(&hrt->hrt_lock);
2627 list_splice_init(&hrt->hrt_queue, replies);
2628 result = ptlrpc_hr.hr_stopping || !list_empty(replies);
2630 spin_unlock(&hrt->hrt_lock);
2635 * Main body of "handle reply" function.
2636 * It processes acked reply states
2638 static int ptlrpc_hr_main(void *arg)
2640 struct ptlrpc_hr_thread *hrt = (struct ptlrpc_hr_thread *)arg;
2641 struct ptlrpc_hr_partition *hrp = hrt->hrt_partition;
2642 struct list_head replies;
2645 INIT_LIST_HEAD(&replies);
2646 unshare_fs_struct();
2648 rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
2650 char threadname[20];
2652 snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
2653 hrp->hrp_cpt, hrt->hrt_id);
2654 CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
2655 threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
2658 atomic_inc(&hrp->hrp_nstarted);
2659 wake_up(&ptlrpc_hr.hr_waitq);
2661 while (!ptlrpc_hr.hr_stopping) {
2662 l_wait_condition(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
2664 while (!list_empty(&replies)) {
2665 struct ptlrpc_reply_state *rs;
2667 rs = list_entry(replies.prev,
2668 struct ptlrpc_reply_state,
2670 list_del_init(&rs->rs_list);
2671 ptlrpc_handle_rs(rs);
2675 atomic_inc(&hrp->hrp_nstopped);
2676 wake_up(&ptlrpc_hr.hr_waitq);
2681 static void ptlrpc_stop_hr_threads(void)
2683 struct ptlrpc_hr_partition *hrp;
2687 ptlrpc_hr.hr_stopping = 1;
2689 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2690 if (hrp->hrp_thrs == NULL)
2691 continue; /* uninitialized */
2692 for (j = 0; j < hrp->hrp_nthrs; j++)
2693 wake_up_all(&hrp->hrp_thrs[j].hrt_waitq);
2696 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2697 if (hrp->hrp_thrs == NULL)
2698 continue; /* uninitialized */
2699 wait_event(ptlrpc_hr.hr_waitq,
2700 atomic_read(&hrp->hrp_nstopped) ==
2701 atomic_read(&hrp->hrp_nstarted));
2705 static int ptlrpc_start_hr_threads(void)
2707 struct ptlrpc_hr_partition *hrp;
2712 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2715 for (j = 0; j < hrp->hrp_nthrs; j++) {
2716 struct ptlrpc_hr_thread *hrt = &hrp->hrp_thrs[j];
2717 struct task_struct *task;
2719 task = kthread_run(ptlrpc_hr_main,
2721 "ptlrpc_hr%02d_%03d",
2730 wait_event(ptlrpc_hr.hr_waitq,
2731 atomic_read(&hrp->hrp_nstarted) == j);
2734 CERROR("cannot start reply handler thread %d:%d: "
2735 "rc = %d\n", i, j, rc);
2736 ptlrpc_stop_hr_threads();
2744 static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
2746 struct l_wait_info lwi = { 0 };
2747 struct ptlrpc_thread *thread;
2748 struct list_head zombie;
2752 CDEBUG(D_INFO, "Stopping threads for service %s\n",
2753 svcpt->scp_service->srv_name);
2755 INIT_LIST_HEAD(&zombie);
2756 spin_lock(&svcpt->scp_lock);
2757 /* let the thread know that we would like it to stop asap */
2758 list_for_each_entry(thread, &svcpt->scp_threads, t_link) {
2759 CDEBUG(D_INFO, "Stopping thread %s #%u\n",
2760 svcpt->scp_service->srv_thread_name, thread->t_id);
2761 thread_add_flags(thread, SVC_STOPPING);
2764 wake_up_all(&svcpt->scp_waitq);
2766 while (!list_empty(&svcpt->scp_threads)) {
2767 thread = list_entry(svcpt->scp_threads.next,
2768 struct ptlrpc_thread, t_link);
2769 if (thread_is_stopped(thread)) {
2770 list_del(&thread->t_link);
2771 list_add(&thread->t_link, &zombie);
2774 spin_unlock(&svcpt->scp_lock);
2776 CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
2777 svcpt->scp_service->srv_thread_name, thread->t_id);
2778 l_wait_event(thread->t_ctl_waitq,
2779 thread_is_stopped(thread), &lwi);
2781 spin_lock(&svcpt->scp_lock);
2784 spin_unlock(&svcpt->scp_lock);
2786 while (!list_empty(&zombie)) {
2787 thread = list_entry(zombie.next,
2788 struct ptlrpc_thread, t_link);
2789 list_del(&thread->t_link);
2790 OBD_FREE_PTR(thread);
2796 * Stops all threads of a particular service \a svc
2798 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2800 struct ptlrpc_service_part *svcpt;
2804 ptlrpc_service_for_each_part(svcpt, i, svc) {
2805 if (svcpt->scp_service != NULL)
2806 ptlrpc_svcpt_stop_threads(svcpt);
2812 int ptlrpc_start_threads(struct ptlrpc_service *svc)
2819 /* We require 2 threads min, see note in ptlrpc_server_handle_request */
2820 LASSERT(svc->srv_nthrs_cpt_init >= PTLRPC_NTHRS_INIT);
2822 for (i = 0; i < svc->srv_ncpts; i++) {
2823 for (j = 0; j < svc->srv_nthrs_cpt_init; j++) {
2824 rc = ptlrpc_start_thread(svc->srv_parts[i], 1);
2830 /* We have enough threads, don't start more. b=15759 */
2837 CERROR("cannot start %s thread #%d_%d: rc %d\n",
2838 svc->srv_thread_name, i, j, rc);
2839 ptlrpc_stop_all_threads(svc);
2843 int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
2845 struct l_wait_info lwi = { 0 };
2846 struct ptlrpc_thread *thread;
2847 struct ptlrpc_service *svc;
2848 struct task_struct *task;
2852 LASSERT(svcpt != NULL);
2854 svc = svcpt->scp_service;
2856 CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n",
2857 svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running,
2858 svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit);
2861 if (unlikely(svc->srv_is_stopping))
2864 if (!ptlrpc_threads_increasable(svcpt) ||
2865 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
2866 svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1))
2869 OBD_CPT_ALLOC_PTR(thread, svc->srv_cptable, svcpt->scp_cpt);
2872 init_waitqueue_head(&thread->t_ctl_waitq);
2874 spin_lock(&svcpt->scp_lock);
2875 if (!ptlrpc_threads_increasable(svcpt)) {
2876 spin_unlock(&svcpt->scp_lock);
2877 OBD_FREE_PTR(thread);
2881 if (svcpt->scp_nthrs_starting != 0) {
2882 /* serialize starting because some modules (obdfilter)
2883 * might require unique and contiguous t_id */
2884 LASSERT(svcpt->scp_nthrs_starting == 1);
2885 spin_unlock(&svcpt->scp_lock);
2886 OBD_FREE_PTR(thread);
2888 CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n",
2889 svc->srv_thread_name, svcpt->scp_thr_nextid);
2894 CDEBUG(D_INFO, "Creating thread %s #%d race, retry later\n",
2895 svc->srv_thread_name, svcpt->scp_thr_nextid);
2899 svcpt->scp_nthrs_starting++;
2900 thread->t_id = svcpt->scp_thr_nextid++;
2901 thread_add_flags(thread, SVC_STARTING);
2902 thread->t_svcpt = svcpt;
2904 list_add(&thread->t_link, &svcpt->scp_threads);
2905 spin_unlock(&svcpt->scp_lock);
2907 if (svcpt->scp_cpt >= 0) {
2908 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s%02d_%03d",
2909 svc->srv_thread_name, svcpt->scp_cpt, thread->t_id);
2911 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s_%04d",
2912 svc->srv_thread_name, thread->t_id);
2915 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name);
2916 task = kthread_run(ptlrpc_main, thread, "%s", thread->t_name);
2919 CERROR("cannot start thread '%s': rc = %d\n",
2920 thread->t_name, rc);
2921 spin_lock(&svcpt->scp_lock);
2922 --svcpt->scp_nthrs_starting;
2923 if (thread_is_stopping(thread)) {
2924 /* this ptlrpc_thread is being hanled
2925 * by ptlrpc_svcpt_stop_threads now
2927 thread_add_flags(thread, SVC_STOPPED);
2928 wake_up(&thread->t_ctl_waitq);
2929 spin_unlock(&svcpt->scp_lock);
2931 list_del(&thread->t_link);
2932 spin_unlock(&svcpt->scp_lock);
2933 OBD_FREE_PTR(thread);
2941 l_wait_event(thread->t_ctl_waitq,
2942 thread_is_running(thread) || thread_is_stopped(thread),
2945 rc = thread_is_stopped(thread) ? thread->t_id : 0;
2949 int ptlrpc_hr_init(void)
2951 struct ptlrpc_hr_partition *hrp;
2952 struct ptlrpc_hr_thread *hrt;
2959 memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
2960 ptlrpc_hr.hr_cpt_table = cfs_cpt_table;
2962 ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
2964 if (ptlrpc_hr.hr_partitions == NULL)
2967 init_waitqueue_head(&ptlrpc_hr.hr_waitq);
2969 weight = cpumask_weight(topology_sibling_cpumask(smp_processor_id()));
2971 cfs_percpt_for_each(hrp, cpt, ptlrpc_hr.hr_partitions) {
2974 atomic_set(&hrp->hrp_nstarted, 0);
2975 atomic_set(&hrp->hrp_nstopped, 0);
2977 hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, cpt);
2978 hrp->hrp_nthrs /= weight;
2979 if (hrp->hrp_nthrs == 0)
2982 OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, cpt,
2983 hrp->hrp_nthrs * sizeof(*hrt));
2984 if (hrp->hrp_thrs == NULL)
2985 GOTO(out, rc = -ENOMEM);
2987 for (i = 0; i < hrp->hrp_nthrs; i++) {
2988 hrt = &hrp->hrp_thrs[i];
2991 hrt->hrt_partition = hrp;
2992 init_waitqueue_head(&hrt->hrt_waitq);
2993 spin_lock_init(&hrt->hrt_lock);
2994 INIT_LIST_HEAD(&hrt->hrt_queue);
2998 rc = ptlrpc_start_hr_threads();
3005 void ptlrpc_hr_fini(void)
3007 struct ptlrpc_hr_partition *hrp;
3010 if (ptlrpc_hr.hr_partitions == NULL)
3013 ptlrpc_stop_hr_threads();
3015 cfs_percpt_for_each(hrp, cpt, ptlrpc_hr.hr_partitions) {
3016 if (hrp->hrp_thrs != NULL) {
3017 OBD_FREE(hrp->hrp_thrs,
3018 hrp->hrp_nthrs * sizeof(hrp->hrp_thrs[0]));
3022 cfs_percpt_free(ptlrpc_hr.hr_partitions);
3023 ptlrpc_hr.hr_partitions = NULL;
3028 * Wait until all already scheduled replies are processed.
3030 static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
3034 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
3037 rc = l_wait_event(svcpt->scp_waitq,
3038 atomic_read(&svcpt->scp_nreps_difficult) == 0, &lwi);
3041 CWARN("Unexpectedly long timeout %s %p\n",
3042 svcpt->scp_service->srv_name, svcpt->scp_service);
3047 ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
3049 struct ptlrpc_service_part *svcpt;
3052 /* early disarm AT timer... */
3053 ptlrpc_service_for_each_part(svcpt, i, svc) {
3054 if (svcpt->scp_service != NULL)
3055 del_timer(&svcpt->scp_at_timer);
3060 ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
3062 struct ptlrpc_service_part *svcpt;
3063 struct ptlrpc_request_buffer_desc *rqbd;
3064 struct l_wait_info lwi;
3068 /* All history will be culled when the next request buffer is
3069 * freed in ptlrpc_service_purge_all() */
3070 svc->srv_hist_nrqbds_cpt_max = 0;
3072 rc = LNetClearLazyPortal(svc->srv_req_portal);
3075 ptlrpc_service_for_each_part(svcpt, i, svc) {
3076 if (svcpt->scp_service == NULL)
3079 /* Unlink all the request buffers. This forces a 'final'
3080 * event with its 'unlink' flag set for each posted rqbd */
3081 list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
3083 rc = LNetMDUnlink(rqbd->rqbd_md_h);
3084 LASSERT(rc == 0 || rc == -ENOENT);
3088 ptlrpc_service_for_each_part(svcpt, i, svc) {
3089 if (svcpt->scp_service == NULL)
3092 /* Wait for the network to release any buffers
3093 * it's currently filling */
3094 spin_lock(&svcpt->scp_lock);
3095 while (svcpt->scp_nrqbds_posted != 0) {
3096 spin_unlock(&svcpt->scp_lock);
3097 /* Network access will complete in finite time but
3098 * the HUGE timeout lets us CWARN for visibility
3099 * of sluggish NALs */
3100 lwi = LWI_TIMEOUT_INTERVAL(
3101 cfs_time_seconds(LONG_UNLINK),
3102 cfs_time_seconds(1), NULL, NULL);
3103 rc = l_wait_event(svcpt->scp_waitq,
3104 svcpt->scp_nrqbds_posted == 0, &lwi);
3105 if (rc == -ETIMEDOUT) {
3106 CWARN("Service %s waiting for "
3107 "request buffers\n",
3108 svcpt->scp_service->srv_name);
3110 spin_lock(&svcpt->scp_lock);
3112 spin_unlock(&svcpt->scp_lock);
3117 ptlrpc_service_purge_all(struct ptlrpc_service *svc)
3119 struct ptlrpc_service_part *svcpt;
3120 struct ptlrpc_request_buffer_desc *rqbd;
3121 struct ptlrpc_request *req;
3122 struct ptlrpc_reply_state *rs;
3125 ptlrpc_service_for_each_part(svcpt, i, svc) {
3126 if (svcpt->scp_service == NULL)
3129 spin_lock(&svcpt->scp_rep_lock);
3130 while (!list_empty(&svcpt->scp_rep_active)) {
3131 rs = list_entry(svcpt->scp_rep_active.next,
3132 struct ptlrpc_reply_state, rs_list);
3133 spin_lock(&rs->rs_lock);
3134 ptlrpc_schedule_difficult_reply(rs);
3135 spin_unlock(&rs->rs_lock);
3137 spin_unlock(&svcpt->scp_rep_lock);
3139 /* purge the request queue. NB No new replies (rqbds
3140 * all unlinked) and no service threads, so I'm the only
3141 * thread noodling the request queue now */
3142 while (!list_empty(&svcpt->scp_req_incoming)) {
3143 req = list_entry(svcpt->scp_req_incoming.next,
3144 struct ptlrpc_request, rq_list);
3146 list_del(&req->rq_list);
3147 svcpt->scp_nreqs_incoming--;
3148 ptlrpc_server_finish_request(svcpt, req);
3151 while (ptlrpc_server_request_pending(svcpt, true)) {
3152 req = ptlrpc_server_request_get(svcpt, true);
3153 ptlrpc_server_finish_active_request(svcpt, req);
3156 LASSERT(list_empty(&svcpt->scp_rqbd_posted));
3157 LASSERT(svcpt->scp_nreqs_incoming == 0);
3158 LASSERT(svcpt->scp_nreqs_active == 0);
3159 /* history should have been culled by
3160 * ptlrpc_server_finish_request */
3161 LASSERT(svcpt->scp_hist_nrqbds == 0);
3163 /* Now free all the request buffers since nothing
3164 * references them any more... */
3166 while (!list_empty(&svcpt->scp_rqbd_idle)) {
3167 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
3168 struct ptlrpc_request_buffer_desc,
3170 ptlrpc_free_rqbd(rqbd);
3172 ptlrpc_wait_replies(svcpt);
3174 while (!list_empty(&svcpt->scp_rep_idle)) {
3175 rs = list_entry(svcpt->scp_rep_idle.next,
3176 struct ptlrpc_reply_state,
3178 list_del(&rs->rs_list);
3179 OBD_FREE_LARGE(rs, svc->srv_max_reply_size);
3185 ptlrpc_service_free(struct ptlrpc_service *svc)
3187 struct ptlrpc_service_part *svcpt;
3188 struct ptlrpc_at_array *array;
3191 ptlrpc_service_for_each_part(svcpt, i, svc) {
3192 if (svcpt->scp_service == NULL)
3195 /* In case somebody rearmed this in the meantime */
3196 del_timer(&svcpt->scp_at_timer);
3197 array = &svcpt->scp_at_array;
3199 if (array->paa_reqs_array != NULL) {
3200 OBD_FREE(array->paa_reqs_array,
3201 sizeof(struct list_head) * array->paa_size);
3202 array->paa_reqs_array = NULL;
3205 if (array->paa_reqs_count != NULL) {
3206 OBD_FREE(array->paa_reqs_count,
3207 sizeof(__u32) * array->paa_size);
3208 array->paa_reqs_count = NULL;
3212 ptlrpc_service_for_each_part(svcpt, i, svc)
3213 OBD_FREE_PTR(svcpt);
3215 if (svc->srv_cpts != NULL)
3216 cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts);
3218 OBD_FREE(svc, offsetof(struct ptlrpc_service,
3219 srv_parts[svc->srv_ncpts]));
3222 int ptlrpc_unregister_service(struct ptlrpc_service *service)
3226 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
3228 service->srv_is_stopping = 1;
3230 mutex_lock(&ptlrpc_all_services_mutex);
3231 list_del_init(&service->srv_list);
3232 mutex_unlock(&ptlrpc_all_services_mutex);
3234 ptlrpc_service_del_atimer(service);
3235 ptlrpc_stop_all_threads(service);
3237 ptlrpc_service_unlink_rqbd(service);
3238 ptlrpc_service_purge_all(service);
3239 ptlrpc_service_nrs_cleanup(service);
3241 ptlrpc_lprocfs_unregister_service(service);
3243 ptlrpc_service_free(service);
3247 EXPORT_SYMBOL(ptlrpc_unregister_service);
3250 * Returns 0 if the service is healthy.
3252 * Right now, it just checks to make sure that requests aren't languishing
3253 * in the queue. We'll use this health check to govern whether a node needs
3254 * to be shot, so it's intentionally non-aggressive. */
3255 static int ptlrpc_svcpt_health_check(struct ptlrpc_service_part *svcpt)
3257 struct ptlrpc_request *request = NULL;
3258 struct timespec64 right_now;
3259 struct timespec64 timediff;
3261 ktime_get_real_ts64(&right_now);
3263 spin_lock(&svcpt->scp_req_lock);
3264 /* How long has the next entry been waiting? */
3265 if (ptlrpc_server_high_pending(svcpt, true))
3266 request = ptlrpc_nrs_req_peek_nolock(svcpt, true);
3267 else if (ptlrpc_server_normal_pending(svcpt, true))
3268 request = ptlrpc_nrs_req_peek_nolock(svcpt, false);
3270 if (request == NULL) {
3271 spin_unlock(&svcpt->scp_req_lock);
3275 timediff = timespec64_sub(right_now, request->rq_arrival_time);
3276 spin_unlock(&svcpt->scp_req_lock);
3278 if ((timediff.tv_sec) >
3279 (AT_OFF ? obd_timeout * 3 / 2 : at_max)) {
3280 CERROR("%s: unhealthy - request has been waiting %llds\n",
3281 svcpt->scp_service->srv_name, (s64)timediff.tv_sec);
3289 ptlrpc_service_health_check(struct ptlrpc_service *svc)
3291 struct ptlrpc_service_part *svcpt;
3297 ptlrpc_service_for_each_part(svcpt, i, svc) {
3298 int rc = ptlrpc_svcpt_health_check(svcpt);
3305 EXPORT_SYMBOL(ptlrpc_service_health_check);