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, 2017, 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
35 #include <linux/kthread.h>
36 #include <linux/ratelimit.h>
38 #include <obd_support.h>
39 #include <obd_class.h>
40 #include <lustre_net.h>
41 #include <lu_object.h>
42 #include <uapi/linux/lnet/lnet-types.h>
43 #include "ptlrpc_internal.h"
45 /* The following are visible and mutable through /sys/module/ptlrpc */
46 int test_req_buffer_pressure = 0;
47 module_param(test_req_buffer_pressure, int, 0444);
48 MODULE_PARM_DESC(test_req_buffer_pressure, "set non-zero to put pressure on request buffer pools");
49 module_param(at_min, int, 0644);
50 MODULE_PARM_DESC(at_min, "Adaptive timeout minimum (sec)");
51 module_param(at_max, int, 0644);
52 MODULE_PARM_DESC(at_max, "Adaptive timeout maximum (sec)");
53 module_param(at_history, int, 0644);
54 MODULE_PARM_DESC(at_history,
55 "Adaptive timeouts remember the slowest event that took place within this period (sec)");
56 module_param(at_early_margin, int, 0644);
57 MODULE_PARM_DESC(at_early_margin, "How soon before an RPC deadline to send an early reply");
58 module_param(at_extra, int, 0644);
59 MODULE_PARM_DESC(at_extra, "How much extra time to give with each early reply");
62 static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt);
63 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req);
64 static void ptlrpc_at_remove_timed(struct ptlrpc_request *req);
66 /** Holds a list of all PTLRPC services */
67 struct list_head ptlrpc_all_services;
68 /** Used to protect the \e ptlrpc_all_services list */
69 struct mutex ptlrpc_all_services_mutex;
71 static struct ptlrpc_request_buffer_desc *
72 ptlrpc_alloc_rqbd(struct ptlrpc_service_part *svcpt)
74 struct ptlrpc_service *svc = svcpt->scp_service;
75 struct ptlrpc_request_buffer_desc *rqbd;
77 OBD_CPT_ALLOC_PTR(rqbd, svc->srv_cptable, svcpt->scp_cpt);
81 rqbd->rqbd_svcpt = svcpt;
82 rqbd->rqbd_refcount = 0;
83 rqbd->rqbd_cbid.cbid_fn = request_in_callback;
84 rqbd->rqbd_cbid.cbid_arg = rqbd;
85 INIT_LIST_HEAD(&rqbd->rqbd_reqs);
86 OBD_CPT_ALLOC_LARGE(rqbd->rqbd_buffer, svc->srv_cptable,
87 svcpt->scp_cpt, svc->srv_buf_size);
88 if (rqbd->rqbd_buffer == NULL) {
93 spin_lock(&svcpt->scp_lock);
94 list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
95 svcpt->scp_nrqbds_total++;
96 spin_unlock(&svcpt->scp_lock);
101 static void ptlrpc_free_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
103 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
105 LASSERT(rqbd->rqbd_refcount == 0);
106 LASSERT(list_empty(&rqbd->rqbd_reqs));
108 spin_lock(&svcpt->scp_lock);
109 list_del(&rqbd->rqbd_list);
110 svcpt->scp_nrqbds_total--;
111 spin_unlock(&svcpt->scp_lock);
113 OBD_FREE_LARGE(rqbd->rqbd_buffer, svcpt->scp_service->srv_buf_size);
117 static int ptlrpc_grow_req_bufs(struct ptlrpc_service_part *svcpt, int post)
119 struct ptlrpc_service *svc = svcpt->scp_service;
120 struct ptlrpc_request_buffer_desc *rqbd;
124 if (svcpt->scp_rqbd_allocating)
127 spin_lock(&svcpt->scp_lock);
128 /* check again with lock */
129 if (svcpt->scp_rqbd_allocating) {
130 /* NB: we might allow more than one thread in the future */
131 LASSERT(svcpt->scp_rqbd_allocating == 1);
132 spin_unlock(&svcpt->scp_lock);
136 svcpt->scp_rqbd_allocating++;
137 spin_unlock(&svcpt->scp_lock);
140 for (i = 0; i < svc->srv_nbuf_per_group; i++) {
142 * NB: another thread might have recycled enough rqbds, we
143 * need to make sure it wouldn't over-allocate, see LU-1212.
145 if (svcpt->scp_nrqbds_posted >= svc->srv_nbuf_per_group ||
146 (svc->srv_nrqbds_max != 0 &&
147 svcpt->scp_nrqbds_total > svc->srv_nrqbds_max))
150 rqbd = ptlrpc_alloc_rqbd(svcpt);
153 CERROR("%s: Can't allocate request buffer\n",
160 spin_lock(&svcpt->scp_lock);
162 LASSERT(svcpt->scp_rqbd_allocating == 1);
163 svcpt->scp_rqbd_allocating--;
165 spin_unlock(&svcpt->scp_lock);
168 "%s: allocate %d new %d-byte reqbufs (%d/%d left), rc = %d\n",
169 svc->srv_name, i, svc->srv_buf_size, svcpt->scp_nrqbds_posted,
170 svcpt->scp_nrqbds_total, rc);
174 rc = ptlrpc_server_post_idle_rqbds(svcpt);
180 * Part of Rep-Ack logic.
181 * Puts a lock and its mode into reply state assotiated to request reply.
183 void ptlrpc_save_lock(struct ptlrpc_request *req, struct lustre_handle *lock,
184 int mode, bool no_ack, bool convert_lock)
186 struct ptlrpc_reply_state *rs = req->rq_reply_state;
190 LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
192 idx = rs->rs_nlocks++;
193 rs->rs_locks[idx] = *lock;
194 rs->rs_modes[idx] = mode;
195 rs->rs_difficult = 1;
196 rs->rs_no_ack = no_ack;
197 rs->rs_convert_lock = convert_lock;
199 EXPORT_SYMBOL(ptlrpc_save_lock);
202 struct ptlrpc_hr_partition;
204 struct ptlrpc_hr_thread {
205 int hrt_id; /* thread ID */
207 wait_queue_head_t hrt_waitq;
208 struct list_head hrt_queue;
209 struct ptlrpc_hr_partition *hrt_partition;
212 struct ptlrpc_hr_partition {
213 /* # of started threads */
214 atomic_t hrp_nstarted;
215 /* # of stopped threads */
216 atomic_t hrp_nstopped;
217 /* cpu partition id */
219 /* round-robin rotor for choosing thread */
221 /* total number of threads on this partition */
224 struct ptlrpc_hr_thread *hrp_thrs;
227 #define HRT_RUNNING 0
228 #define HRT_STOPPING 1
230 struct ptlrpc_hr_service {
231 /* CPU partition table, it's just cfs_cpt_table for now */
232 struct cfs_cpt_table *hr_cpt_table;
233 /** controller sleep waitq */
234 wait_queue_head_t hr_waitq;
235 unsigned int hr_stopping;
236 /** roundrobin rotor for non-affinity service */
237 unsigned int hr_rotor;
239 struct ptlrpc_hr_partition **hr_partitions;
243 struct list_head rsb_replies;
244 unsigned int rsb_n_replies;
245 struct ptlrpc_service_part *rsb_svcpt;
248 /** reply handling service. */
249 static struct ptlrpc_hr_service ptlrpc_hr;
252 * maximum mumber of replies scheduled in one batch
254 #define MAX_SCHEDULED 256
257 * Initialize a reply batch.
261 static void rs_batch_init(struct rs_batch *b)
263 memset(b, 0, sizeof(*b));
264 INIT_LIST_HEAD(&b->rsb_replies);
268 * Choose an hr thread to dispatch requests to.
271 struct ptlrpc_hr_thread *ptlrpc_hr_select(struct ptlrpc_service_part *svcpt)
273 struct ptlrpc_hr_partition *hrp;
276 if (svcpt->scp_cpt >= 0 &&
277 svcpt->scp_service->srv_cptable == ptlrpc_hr.hr_cpt_table) {
278 /* directly match partition */
279 hrp = ptlrpc_hr.hr_partitions[svcpt->scp_cpt];
282 rotor = ptlrpc_hr.hr_rotor++;
283 rotor %= cfs_cpt_number(ptlrpc_hr.hr_cpt_table);
285 hrp = ptlrpc_hr.hr_partitions[rotor];
288 rotor = hrp->hrp_rotor++;
289 return &hrp->hrp_thrs[rotor % hrp->hrp_nthrs];
293 * Dispatch all replies accumulated in the batch to one from
294 * dedicated reply handling threads.
298 static void rs_batch_dispatch(struct rs_batch *b)
300 if (b->rsb_n_replies != 0) {
301 struct ptlrpc_hr_thread *hrt;
303 hrt = ptlrpc_hr_select(b->rsb_svcpt);
305 spin_lock(&hrt->hrt_lock);
306 list_splice_init(&b->rsb_replies, &hrt->hrt_queue);
307 spin_unlock(&hrt->hrt_lock);
309 wake_up(&hrt->hrt_waitq);
310 b->rsb_n_replies = 0;
315 * Add a reply to a batch.
316 * Add one reply object to a batch, schedule batched replies if overload.
321 static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs)
323 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
325 if (svcpt != b->rsb_svcpt || b->rsb_n_replies >= MAX_SCHEDULED) {
326 if (b->rsb_svcpt != NULL) {
327 rs_batch_dispatch(b);
328 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
330 spin_lock(&svcpt->scp_rep_lock);
331 b->rsb_svcpt = svcpt;
333 spin_lock(&rs->rs_lock);
334 rs->rs_scheduled_ever = 1;
335 if (rs->rs_scheduled == 0) {
336 list_move(&rs->rs_list, &b->rsb_replies);
337 rs->rs_scheduled = 1;
340 rs->rs_committed = 1;
341 spin_unlock(&rs->rs_lock);
345 * Reply batch finalization.
346 * Dispatch remaining replies from the batch
347 * and release remaining spinlock.
351 static void rs_batch_fini(struct rs_batch *b)
353 if (b->rsb_svcpt != NULL) {
354 rs_batch_dispatch(b);
355 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
359 #define DECLARE_RS_BATCH(b) struct rs_batch b
363 * Put reply state into a queue for processing because we received
364 * ACK from the client
366 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
368 struct ptlrpc_hr_thread *hrt;
372 LASSERT(list_empty(&rs->rs_list));
374 hrt = ptlrpc_hr_select(rs->rs_svcpt);
376 spin_lock(&hrt->hrt_lock);
377 list_add_tail(&rs->rs_list, &hrt->hrt_queue);
378 spin_unlock(&hrt->hrt_lock);
380 wake_up(&hrt->hrt_waitq);
384 void ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs)
388 assert_spin_locked(&rs->rs_svcpt->scp_rep_lock);
389 assert_spin_locked(&rs->rs_lock);
390 LASSERT(rs->rs_difficult);
391 rs->rs_scheduled_ever = 1; /* flag any notification attempt */
393 if (rs->rs_scheduled) { /* being set up or already notified */
398 rs->rs_scheduled = 1;
399 list_del_init(&rs->rs_list);
400 ptlrpc_dispatch_difficult_reply(rs);
403 EXPORT_SYMBOL(ptlrpc_schedule_difficult_reply);
405 void ptlrpc_commit_replies(struct obd_export *exp)
407 struct ptlrpc_reply_state *rs, *nxt;
408 DECLARE_RS_BATCH(batch);
412 rs_batch_init(&batch);
414 * Find any replies that have been committed and get their service
415 * to attend to complete them.
418 /* CAVEAT EMPTOR: spinlock ordering!!! */
419 spin_lock(&exp->exp_uncommitted_replies_lock);
420 list_for_each_entry_safe(rs, nxt, &exp->exp_uncommitted_replies,
422 LASSERT(rs->rs_difficult);
423 /* VBR: per-export last_committed */
424 LASSERT(rs->rs_export);
425 if (rs->rs_transno <= exp->exp_last_committed) {
426 list_del_init(&rs->rs_obd_list);
427 rs_batch_add(&batch, rs);
430 spin_unlock(&exp->exp_uncommitted_replies_lock);
431 rs_batch_fini(&batch);
435 static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt)
437 struct ptlrpc_request_buffer_desc *rqbd;
442 spin_lock(&svcpt->scp_lock);
444 if (list_empty(&svcpt->scp_rqbd_idle)) {
445 spin_unlock(&svcpt->scp_lock);
449 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
450 struct ptlrpc_request_buffer_desc,
452 list_del(&rqbd->rqbd_list);
454 /* assume we will post successfully */
455 svcpt->scp_nrqbds_posted++;
456 list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_posted);
458 spin_unlock(&svcpt->scp_lock);
460 rc = ptlrpc_register_rqbd(rqbd);
467 spin_lock(&svcpt->scp_lock);
469 svcpt->scp_nrqbds_posted--;
470 list_del(&rqbd->rqbd_list);
471 list_add_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
474 * Don't complain if no request buffers are posted right now; LNET
475 * won't drop requests because we set the portal lazy!
478 spin_unlock(&svcpt->scp_lock);
483 static void ptlrpc_at_timer(cfs_timer_cb_arg_t data)
485 struct ptlrpc_service_part *svcpt;
487 svcpt = cfs_from_timer(svcpt, data, scp_at_timer);
489 svcpt->scp_at_check = 1;
490 svcpt->scp_at_checktime = ktime_get();
491 wake_up(&svcpt->scp_waitq);
494 static void ptlrpc_server_nthreads_check(struct ptlrpc_service *svc,
495 struct ptlrpc_service_conf *conf)
497 struct ptlrpc_service_thr_conf *tc = &conf->psc_thr;
504 * Common code for estimating & validating threads number.
505 * CPT affinity service could have percpt thread-pool instead
506 * of a global thread-pool, which means user might not always
507 * get the threads number they give it in conf::tc_nthrs_user
508 * even they did set. It's because we need to validate threads
509 * number for each CPT to guarantee each pool will have enough
510 * threads to keep the service healthy.
512 init = PTLRPC_NTHRS_INIT + (svc->srv_ops.so_hpreq_handler != NULL);
513 init = max_t(int, init, tc->tc_nthrs_init);
516 * NB: please see comments in lustre_lnet.h for definition
517 * details of these members
519 LASSERT(tc->tc_nthrs_max != 0);
521 if (tc->tc_nthrs_user != 0) {
523 * In case there is a reason to test a service with many
524 * threads, we give a less strict check here, it can
525 * be up to 8 * nthrs_max
527 total = min(tc->tc_nthrs_max * 8, tc->tc_nthrs_user);
528 nthrs = total / svc->srv_ncpts;
529 init = max(init, nthrs);
533 total = tc->tc_nthrs_max;
534 if (tc->tc_nthrs_base == 0) {
536 * don't care about base threads number per partition,
537 * this is most for non-affinity service
539 nthrs = total / svc->srv_ncpts;
543 nthrs = tc->tc_nthrs_base;
544 if (svc->srv_ncpts == 1) {
548 * NB: Increase the base number if it's single partition
549 * and total number of cores/HTs is larger or equal to 4.
550 * result will always < 2 * nthrs_base
552 weight = cfs_cpt_weight(svc->srv_cptable, CFS_CPT_ANY);
553 for (i = 1; (weight >> (i + 1)) != 0 && /* >= 4 cores/HTs */
554 (tc->tc_nthrs_base >> i) != 0; i++)
555 nthrs += tc->tc_nthrs_base >> i;
558 if (tc->tc_thr_factor != 0) {
559 int factor = tc->tc_thr_factor;
563 * User wants to increase number of threads with for
564 * each CPU core/HT, most likely the factor is larger than
565 * one thread/core because service threads are supposed to
566 * be blocked by lock or wait for IO.
569 * Amdahl's law says that adding processors wouldn't give
570 * a linear increasing of parallelism, so it's nonsense to
571 * have too many threads no matter how many cores/HTs
576 (topology_sibling_cpumask(smp_processor_id())) > 1) {
577 /* weight is # of HTs */
578 /* depress thread factor for hyper-thread */
579 factor = factor - (factor >> 1) + (factor >> 3);
583 weight = cfs_cpt_weight(svc->srv_cptable, 0);
585 for (; factor > 0 && weight > 0; factor--, weight -= fade)
586 nthrs += min(weight, fade) * factor;
589 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
590 nthrs = max(tc->tc_nthrs_base,
591 tc->tc_nthrs_max / svc->srv_ncpts);
594 nthrs = max(nthrs, tc->tc_nthrs_init);
595 svc->srv_nthrs_cpt_limit = nthrs;
596 svc->srv_nthrs_cpt_init = init;
598 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
600 "%s: This service may have more threads (%d) than the given soft limit (%d)\n",
601 svc->srv_name, nthrs * svc->srv_ncpts,
607 * Initialize percpt data for a service
609 static int ptlrpc_service_part_init(struct ptlrpc_service *svc,
610 struct ptlrpc_service_part *svcpt, int cpt)
612 struct ptlrpc_at_array *array;
617 svcpt->scp_cpt = cpt;
618 INIT_LIST_HEAD(&svcpt->scp_threads);
620 /* rqbd and incoming request queue */
621 spin_lock_init(&svcpt->scp_lock);
622 mutex_init(&svcpt->scp_mutex);
623 INIT_LIST_HEAD(&svcpt->scp_rqbd_idle);
624 INIT_LIST_HEAD(&svcpt->scp_rqbd_posted);
625 INIT_LIST_HEAD(&svcpt->scp_req_incoming);
626 init_waitqueue_head(&svcpt->scp_waitq);
627 /* history request & rqbd list */
628 INIT_LIST_HEAD(&svcpt->scp_hist_reqs);
629 INIT_LIST_HEAD(&svcpt->scp_hist_rqbds);
631 /* acitve requests and hp requests */
632 spin_lock_init(&svcpt->scp_req_lock);
635 spin_lock_init(&svcpt->scp_rep_lock);
636 INIT_LIST_HEAD(&svcpt->scp_rep_active);
637 INIT_LIST_HEAD(&svcpt->scp_rep_idle);
638 init_waitqueue_head(&svcpt->scp_rep_waitq);
639 atomic_set(&svcpt->scp_nreps_difficult, 0);
641 /* adaptive timeout */
642 spin_lock_init(&svcpt->scp_at_lock);
643 array = &svcpt->scp_at_array;
645 size = at_est2timeout(at_max);
646 array->paa_size = size;
647 array->paa_count = 0;
648 array->paa_deadline = -1;
650 /* allocate memory for scp_at_array (ptlrpc_at_array) */
651 OBD_CPT_ALLOC(array->paa_reqs_array,
652 svc->srv_cptable, cpt, sizeof(struct list_head) * size);
653 if (array->paa_reqs_array == NULL)
656 for (index = 0; index < size; index++)
657 INIT_LIST_HEAD(&array->paa_reqs_array[index]);
659 OBD_CPT_ALLOC(array->paa_reqs_count,
660 svc->srv_cptable, cpt, sizeof(__u32) * size);
661 if (array->paa_reqs_count == NULL)
664 cfs_timer_setup(&svcpt->scp_at_timer, ptlrpc_at_timer,
665 (unsigned long)svcpt, 0);
668 * At SOW, service time should be quick; 10s seems generous. If client
669 * timeout is less than this, we'll be sending an early reply.
671 at_init(&svcpt->scp_at_estimate, 10, 0);
673 /* assign this before call ptlrpc_grow_req_bufs */
674 svcpt->scp_service = svc;
675 /* Now allocate the request buffers, but don't post them now */
676 rc = ptlrpc_grow_req_bufs(svcpt, 0);
678 * We shouldn't be under memory pressure at startup, so
679 * fail if we can't allocate all our buffers at this time.
687 if (array->paa_reqs_count != NULL) {
688 OBD_FREE(array->paa_reqs_count, sizeof(__u32) * size);
689 array->paa_reqs_count = NULL;
692 if (array->paa_reqs_array != NULL) {
693 OBD_FREE(array->paa_reqs_array,
694 sizeof(struct list_head) * array->paa_size);
695 array->paa_reqs_array = NULL;
702 * Initialize service on a given portal.
703 * This includes starting serving threads , allocating and posting rqbds and
706 struct ptlrpc_service *ptlrpc_register_service(struct ptlrpc_service_conf *conf,
708 struct dentry *debugfs_entry)
710 struct ptlrpc_service_cpt_conf *cconf = &conf->psc_cpt;
711 struct ptlrpc_service *service;
712 struct ptlrpc_service_part *svcpt;
713 struct cfs_cpt_table *cptable;
722 LASSERT(conf->psc_buf.bc_nbufs > 0);
723 LASSERT(conf->psc_buf.bc_buf_size >=
724 conf->psc_buf.bc_req_max_size + SPTLRPC_MAX_PAYLOAD);
725 LASSERT(conf->psc_thr.tc_ctx_tags != 0);
727 cptable = cconf->cc_cptable;
729 cptable = cfs_cpt_table;
731 if (conf->psc_thr.tc_cpu_bind > 1) {
732 CERROR("%s: Invalid cpu bind value %d, only 1 or 0 allowed\n",
733 conf->psc_name, conf->psc_thr.tc_cpu_bind);
734 RETURN(ERR_PTR(-EINVAL));
737 if (!cconf->cc_affinity) {
740 ncpts = cfs_cpt_number(cptable);
741 if (cconf->cc_pattern != NULL) {
742 struct cfs_expr_list *el;
744 rc = cfs_expr_list_parse(cconf->cc_pattern,
745 strlen(cconf->cc_pattern),
748 CERROR("%s: invalid CPT pattern string: %s",
749 conf->psc_name, cconf->cc_pattern);
750 RETURN(ERR_PTR(-EINVAL));
753 rc = cfs_expr_list_values(el, ncpts, &cpts);
754 cfs_expr_list_free(el);
756 CERROR("%s: failed to parse CPT array %s: %d\n",
757 conf->psc_name, cconf->cc_pattern, rc);
759 OBD_FREE(cpts, sizeof(*cpts) * ncpts);
760 RETURN(ERR_PTR(rc < 0 ? rc : -EINVAL));
766 OBD_ALLOC(service, offsetof(struct ptlrpc_service, srv_parts[ncpts]));
767 if (service == NULL) {
769 OBD_FREE(cpts, sizeof(*cpts) * ncpts);
770 RETURN(ERR_PTR(-ENOMEM));
773 service->srv_cptable = cptable;
774 service->srv_cpts = cpts;
775 service->srv_ncpts = ncpts;
776 service->srv_cpt_bind = conf->psc_thr.tc_cpu_bind;
778 service->srv_cpt_bits = 0; /* it's zero already, easy to read... */
779 while ((1 << service->srv_cpt_bits) < cfs_cpt_number(cptable))
780 service->srv_cpt_bits++;
783 spin_lock_init(&service->srv_lock);
784 service->srv_name = conf->psc_name;
785 service->srv_watchdog_factor = conf->psc_watchdog_factor;
786 INIT_LIST_HEAD(&service->srv_list); /* for safty of cleanup */
788 /* buffer configuration */
789 service->srv_nbuf_per_group = test_req_buffer_pressure ?
790 1 : conf->psc_buf.bc_nbufs;
791 /* do not limit max number of rqbds by default */
792 service->srv_nrqbds_max = 0;
794 service->srv_max_req_size = conf->psc_buf.bc_req_max_size +
796 service->srv_buf_size = conf->psc_buf.bc_buf_size;
797 service->srv_rep_portal = conf->psc_buf.bc_rep_portal;
798 service->srv_req_portal = conf->psc_buf.bc_req_portal;
800 /* With slab/alloc_pages buffer size will be rounded up to 2^n */
801 if (service->srv_buf_size & (service->srv_buf_size - 1)) {
802 int round = size_roundup_power2(service->srv_buf_size);
804 service->srv_buf_size = round;
807 /* Increase max reply size to next power of two */
808 service->srv_max_reply_size = 1;
809 while (service->srv_max_reply_size <
810 conf->psc_buf.bc_rep_max_size + SPTLRPC_MAX_PAYLOAD)
811 service->srv_max_reply_size <<= 1;
813 service->srv_thread_name = conf->psc_thr.tc_thr_name;
814 service->srv_ctx_tags = conf->psc_thr.tc_ctx_tags;
815 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
816 service->srv_ops = conf->psc_ops;
818 for (i = 0; i < ncpts; i++) {
819 if (!cconf->cc_affinity)
822 cpt = cpts != NULL ? cpts[i] : i;
824 OBD_CPT_ALLOC(svcpt, cptable, cpt, sizeof(*svcpt));
826 GOTO(failed, rc = -ENOMEM);
828 service->srv_parts[i] = svcpt;
829 rc = ptlrpc_service_part_init(service, svcpt, cpt);
834 ptlrpc_server_nthreads_check(service, conf);
836 rc = LNetSetLazyPortal(service->srv_req_portal);
839 mutex_lock(&ptlrpc_all_services_mutex);
840 list_add(&service->srv_list, &ptlrpc_all_services);
841 mutex_unlock(&ptlrpc_all_services_mutex);
844 rc = ptlrpc_sysfs_register_service(parent, service);
849 if (debugfs_entry != NULL)
850 ptlrpc_ldebugfs_register_service(debugfs_entry, service);
852 rc = ptlrpc_service_nrs_setup(service);
856 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
857 service->srv_name, service->srv_req_portal);
859 rc = ptlrpc_start_threads(service);
861 CERROR("Failed to start threads for service %s: %d\n",
862 service->srv_name, rc);
868 ptlrpc_unregister_service(service);
871 EXPORT_SYMBOL(ptlrpc_register_service);
874 * to actually free the request, must be called without holding svc_lock.
875 * note it's caller's responsibility to unlink req->rq_list.
877 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
879 LASSERT(atomic_read(&req->rq_refcount) == 0);
880 LASSERT(list_empty(&req->rq_timed_list));
883 * DEBUG_REQ() assumes the reply state of a request with a valid
884 * ref will not be destroyed until that reference is dropped.
886 ptlrpc_req_drop_rs(req);
888 sptlrpc_svc_ctx_decref(req);
890 if (req != &req->rq_rqbd->rqbd_req) {
892 * NB request buffers use an embedded
893 * req if the incoming req unlinked the
894 * MD; this isn't one of them!
896 ptlrpc_request_cache_free(req);
901 * drop a reference count of the request. if it reaches 0, we either
902 * put it into history list, or free it immediately.
904 void ptlrpc_server_drop_request(struct ptlrpc_request *req)
906 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
907 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
908 struct ptlrpc_service *svc = svcpt->scp_service;
910 struct list_head *tmp;
911 struct list_head *nxt;
913 if (!atomic_dec_and_test(&req->rq_refcount))
916 if (req->rq_session.lc_state == LCS_ENTERED) {
917 lu_context_exit(&req->rq_session);
918 lu_context_fini(&req->rq_session);
921 if (req->rq_at_linked) {
922 spin_lock(&svcpt->scp_at_lock);
924 * recheck with lock, in case it's unlinked by
925 * ptlrpc_at_check_timed()
927 if (likely(req->rq_at_linked))
928 ptlrpc_at_remove_timed(req);
929 spin_unlock(&svcpt->scp_at_lock);
932 LASSERT(list_empty(&req->rq_timed_list));
934 /* finalize request */
935 if (req->rq_export) {
936 class_export_put(req->rq_export);
937 req->rq_export = NULL;
940 spin_lock(&svcpt->scp_lock);
942 list_add(&req->rq_list, &rqbd->rqbd_reqs);
944 refcount = --(rqbd->rqbd_refcount);
946 /* request buffer is now idle: add to history */
947 list_del(&rqbd->rqbd_list);
949 list_add_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds);
950 svcpt->scp_hist_nrqbds++;
954 * I expect only about 1 or 2 rqbds need to be recycled here
956 while (svcpt->scp_hist_nrqbds > svc->srv_hist_nrqbds_cpt_max) {
957 rqbd = list_entry(svcpt->scp_hist_rqbds.next,
958 struct ptlrpc_request_buffer_desc,
961 list_del(&rqbd->rqbd_list);
962 svcpt->scp_hist_nrqbds--;
965 * remove rqbd's reqs from svc's req history while
966 * I've got the service lock
968 list_for_each(tmp, &rqbd->rqbd_reqs) {
969 req = list_entry(tmp, struct ptlrpc_request,
971 /* Track the highest culled req seq */
972 if (req->rq_history_seq >
973 svcpt->scp_hist_seq_culled) {
974 svcpt->scp_hist_seq_culled =
977 list_del(&req->rq_history_list);
980 spin_unlock(&svcpt->scp_lock);
982 list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
983 req = list_entry(rqbd->rqbd_reqs.next,
984 struct ptlrpc_request,
986 list_del(&req->rq_list);
987 ptlrpc_server_free_request(req);
990 spin_lock(&svcpt->scp_lock);
992 * now all reqs including the embedded req has been
993 * disposed, schedule request buffer for re-use
994 * or free it to drain some in excess.
996 LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) == 0);
997 if (svcpt->scp_nrqbds_posted >=
998 svc->srv_nbuf_per_group ||
999 (svc->srv_nrqbds_max != 0 &&
1000 svcpt->scp_nrqbds_total > svc->srv_nrqbds_max) ||
1001 test_req_buffer_pressure) {
1002 /* like in ptlrpc_free_rqbd() */
1003 svcpt->scp_nrqbds_total--;
1004 OBD_FREE_LARGE(rqbd->rqbd_buffer,
1008 list_add_tail(&rqbd->rqbd_list,
1009 &svcpt->scp_rqbd_idle);
1013 spin_unlock(&svcpt->scp_lock);
1014 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
1015 /* If we are low on memory, we are not interested in history */
1016 list_del(&req->rq_list);
1017 list_del_init(&req->rq_history_list);
1019 /* Track the highest culled req seq */
1020 if (req->rq_history_seq > svcpt->scp_hist_seq_culled)
1021 svcpt->scp_hist_seq_culled = req->rq_history_seq;
1023 spin_unlock(&svcpt->scp_lock);
1025 ptlrpc_server_free_request(req);
1027 spin_unlock(&svcpt->scp_lock);
1031 /** Change request export and move hp request from old export to new */
1032 void ptlrpc_request_change_export(struct ptlrpc_request *req,
1033 struct obd_export *export)
1035 if (req->rq_export != NULL) {
1036 LASSERT(!list_empty(&req->rq_exp_list));
1037 /* remove rq_exp_list from last export */
1038 spin_lock(&req->rq_export->exp_rpc_lock);
1039 list_del_init(&req->rq_exp_list);
1040 spin_unlock(&req->rq_export->exp_rpc_lock);
1042 * export has one reference already, so it`s safe to
1043 * add req to export queue here and get another
1044 * reference for request later
1046 spin_lock(&export->exp_rpc_lock);
1047 if (req->rq_ops != NULL) /* hp request */
1048 list_add(&req->rq_exp_list, &export->exp_hp_rpcs);
1050 list_add(&req->rq_exp_list, &export->exp_reg_rpcs);
1051 spin_unlock(&export->exp_rpc_lock);
1053 class_export_rpc_dec(req->rq_export);
1054 class_export_put(req->rq_export);
1057 /* request takes one export refcount */
1058 req->rq_export = class_export_get(export);
1059 class_export_rpc_inc(export);
1065 * to finish a request: stop sending more early replies, and release
1068 static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt,
1069 struct ptlrpc_request *req)
1071 ptlrpc_server_hpreq_fini(req);
1073 ptlrpc_server_drop_request(req);
1077 * to finish an active request: stop sending more early replies, and release
1078 * the request. should be called after we finished handling the request.
1080 static void ptlrpc_server_finish_active_request(
1081 struct ptlrpc_service_part *svcpt,
1082 struct ptlrpc_request *req)
1084 spin_lock(&svcpt->scp_req_lock);
1085 ptlrpc_nrs_req_stop_nolock(req);
1086 svcpt->scp_nreqs_active--;
1088 svcpt->scp_nhreqs_active--;
1089 spin_unlock(&svcpt->scp_req_lock);
1091 ptlrpc_nrs_req_finalize(req);
1093 if (req->rq_export != NULL)
1094 class_export_rpc_dec(req->rq_export);
1096 ptlrpc_server_finish_request(svcpt, req);
1100 * This function makes sure dead exports are evicted in a timely manner.
1101 * This function is only called when some export receives a message (i.e.,
1102 * the network is up.)
1104 void ptlrpc_update_export_timer(struct obd_export *exp, time64_t extra_delay)
1106 struct obd_export *oldest_exp;
1107 time64_t oldest_time, new_time;
1114 * Compensate for slow machines, etc, by faking our request time
1115 * into the future. Although this can break the strict time-ordering
1116 * of the list, we can be really lazy here - we don't have to evict
1117 * at the exact right moment. Eventually, all silent exports
1118 * will make it to the top of the list.
1121 /* Do not pay attention on 1sec or smaller renewals. */
1122 new_time = ktime_get_real_seconds() + extra_delay;
1123 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
1126 exp->exp_last_request_time = new_time;
1129 * exports may get disconnected from the chain even though the
1130 * export has references, so we must keep the spin lock while
1131 * manipulating the lists
1133 spin_lock(&exp->exp_obd->obd_dev_lock);
1135 if (list_empty(&exp->exp_obd_chain_timed)) {
1136 /* this one is not timed */
1137 spin_unlock(&exp->exp_obd->obd_dev_lock);
1141 list_move_tail(&exp->exp_obd_chain_timed,
1142 &exp->exp_obd->obd_exports_timed);
1144 oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
1145 struct obd_export, exp_obd_chain_timed);
1146 oldest_time = oldest_exp->exp_last_request_time;
1147 spin_unlock(&exp->exp_obd->obd_dev_lock);
1149 if (exp->exp_obd->obd_recovering) {
1150 /* be nice to everyone during recovery */
1155 /* Note - racing to start/reset the obd_eviction timer is safe */
1156 if (exp->exp_obd->obd_eviction_timer == 0) {
1157 /* Check if the oldest entry is expired. */
1158 if (ktime_get_real_seconds() >
1159 oldest_time + PING_EVICT_TIMEOUT + extra_delay) {
1161 * We need a second timer, in case the net was down and
1162 * it just came back. Since the pinger may skip every
1163 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
1164 * we better wait for 3.
1166 exp->exp_obd->obd_eviction_timer =
1167 ktime_get_real_seconds() + 3 * PING_INTERVAL;
1168 CDEBUG(D_HA, "%s: Think about evicting %s from %lld\n",
1169 exp->exp_obd->obd_name,
1170 obd_export_nid2str(oldest_exp), oldest_time);
1173 if (ktime_get_real_seconds() >
1174 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
1176 * The evictor won't evict anyone who we've heard from
1177 * recently, so we don't have to check before we start
1180 if (!ping_evictor_wake(exp))
1181 exp->exp_obd->obd_eviction_timer = 0;
1189 * Sanity check request \a req.
1190 * Return 0 if all is ok, error code otherwise.
1192 static int ptlrpc_check_req(struct ptlrpc_request *req)
1194 struct obd_device *obd = req->rq_export->exp_obd;
1197 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
1198 req->rq_export->exp_conn_cnt)) {
1199 DEBUG_REQ(D_RPCTRACE, req,
1200 "DROPPING req from old connection %d < %d",
1201 lustre_msg_get_conn_cnt(req->rq_reqmsg),
1202 req->rq_export->exp_conn_cnt);
1205 if (unlikely(obd == NULL || obd->obd_fail)) {
1207 * Failing over, don't handle any more reqs,
1208 * send error response instead.
1210 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
1211 req, (obd != NULL) ? obd->obd_name : "unknown");
1213 } else if (lustre_msg_get_flags(req->rq_reqmsg) &
1214 (MSG_REPLAY | MSG_REQ_REPLAY_DONE) &&
1215 !obd->obd_recovering) {
1216 DEBUG_REQ(D_ERROR, req,
1217 "Invalid replay without recovery");
1218 class_fail_export(req->rq_export);
1220 } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 &&
1221 !obd->obd_recovering) {
1222 DEBUG_REQ(D_ERROR, req,
1223 "Invalid req with transno %llu without recovery",
1224 lustre_msg_get_transno(req->rq_reqmsg));
1225 class_fail_export(req->rq_export);
1229 if (unlikely(rc < 0)) {
1230 req->rq_status = rc;
1236 static void ptlrpc_at_set_timer(struct ptlrpc_service_part *svcpt)
1238 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1241 if (array->paa_count == 0) {
1242 del_timer(&svcpt->scp_at_timer);
1246 /* Set timer for closest deadline */
1247 next = array->paa_deadline - ktime_get_real_seconds() -
1250 ptlrpc_at_timer(cfs_timer_cb_arg(svcpt, scp_at_timer));
1252 mod_timer(&svcpt->scp_at_timer,
1253 jiffies + nsecs_to_jiffies(next * NSEC_PER_SEC));
1254 CDEBUG(D_INFO, "armed %s at %+llds\n",
1255 svcpt->scp_service->srv_name, next);
1259 /* Add rpc to early reply check list */
1260 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
1262 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1263 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1264 struct ptlrpc_request *rq = NULL;
1270 if (req->rq_no_reply)
1273 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
1276 spin_lock(&svcpt->scp_at_lock);
1277 LASSERT(list_empty(&req->rq_timed_list));
1279 div_u64_rem(req->rq_deadline, array->paa_size, &index);
1280 if (array->paa_reqs_count[index] > 0) {
1282 * latest rpcs will have the latest deadlines in the list,
1283 * so search backward.
1285 list_for_each_entry_reverse(rq, &array->paa_reqs_array[index],
1287 if (req->rq_deadline >= rq->rq_deadline) {
1288 list_add(&req->rq_timed_list,
1289 &rq->rq_timed_list);
1295 /* Add the request at the head of the list */
1296 if (list_empty(&req->rq_timed_list))
1297 list_add(&req->rq_timed_list, &array->paa_reqs_array[index]);
1299 spin_lock(&req->rq_lock);
1300 req->rq_at_linked = 1;
1301 spin_unlock(&req->rq_lock);
1302 req->rq_at_index = index;
1303 array->paa_reqs_count[index]++;
1305 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
1306 array->paa_deadline = req->rq_deadline;
1307 ptlrpc_at_set_timer(svcpt);
1309 spin_unlock(&svcpt->scp_at_lock);
1314 static void ptlrpc_at_remove_timed(struct ptlrpc_request *req)
1316 struct ptlrpc_at_array *array;
1318 array = &req->rq_rqbd->rqbd_svcpt->scp_at_array;
1320 /* NB: must call with hold svcpt::scp_at_lock */
1321 LASSERT(!list_empty(&req->rq_timed_list));
1322 list_del_init(&req->rq_timed_list);
1324 spin_lock(&req->rq_lock);
1325 req->rq_at_linked = 0;
1326 spin_unlock(&req->rq_lock);
1328 array->paa_reqs_count[req->rq_at_index]--;
1333 * Attempt to extend the request deadline by sending an early reply to the
1336 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
1338 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1339 struct ptlrpc_request *reqcopy;
1340 struct lustre_msg *reqmsg;
1341 time64_t olddl = req->rq_deadline - ktime_get_real_seconds();
1347 if (CFS_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_RECONNECT)) {
1348 /* don't send early reply */
1353 * deadline is when the client expects us to reply, margin is the
1354 * difference between clients' and servers' expectations
1356 DEBUG_REQ(D_ADAPTTO, req,
1357 "%ssending early reply (deadline %+llds, margin %+llds) for %d+%d",
1358 AT_OFF ? "AT off - not " : "",
1359 (s64)olddl, (s64)(olddl - at_get(&svcpt->scp_at_estimate)),
1360 at_get(&svcpt->scp_at_estimate), at_extra);
1366 DEBUG_REQ(D_WARNING, req,
1367 "Already past deadline (%+llds), not sending early reply. Consider increasing at_early_margin (%d)?",
1368 (s64)olddl, at_early_margin);
1370 /* Return an error so we're not re-added to the timed list. */
1374 if ((lustre_msghdr_get_flags(req->rq_reqmsg) &
1375 MSGHDR_AT_SUPPORT) == 0) {
1376 DEBUG_REQ(D_INFO, req,
1377 "Wanted to ask client for more time, but no AT support");
1381 if (req->rq_export &&
1382 lustre_msg_get_flags(req->rq_reqmsg) &
1383 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1384 struct obd_device *obd_exp = req->rq_export->exp_obd;
1387 * During recovery, we don't want to send too many early
1388 * replies, but on the other hand we want to make sure the
1389 * client has enough time to resend if the rpc is lost. So
1390 * during the recovery period send at least 4 early replies,
1391 * spacing them every at_extra if we can. at_estimate should
1392 * always equal this fixed value during recovery.
1396 * Don't account request processing time into AT history
1397 * during recovery, it is not service time we need but
1398 * includes also waiting time for recovering clients
1400 newdl = min_t(time64_t, at_extra,
1401 obd_exp->obd_recovery_timeout / 4) +
1402 ktime_get_real_seconds();
1405 * We want to extend the request deadline by at_extra seconds,
1406 * so we set our service estimate to reflect how much time has
1407 * passed since this request arrived plus an additional
1408 * at_extra seconds. The client will calculate the new deadline
1409 * based on this service estimate (plus some additional time to
1410 * account for network latency). See ptlrpc_at_recv_early_reply
1412 at_measured(&svcpt->scp_at_estimate, at_extra +
1413 ktime_get_real_seconds() -
1414 req->rq_arrival_time.tv_sec);
1415 newdl = req->rq_arrival_time.tv_sec +
1416 at_get(&svcpt->scp_at_estimate);
1420 * Check to see if we've actually increased the deadline -
1421 * we may be past adaptive_max
1423 if (req->rq_deadline >= newdl) {
1424 DEBUG_REQ(D_WARNING, req, "Couldn't add any time (%lld/%lld), not sending early reply\n",
1425 (s64)olddl, (s64)(newdl - ktime_get_real_seconds()));
1429 reqcopy = ptlrpc_request_cache_alloc(GFP_NOFS);
1430 if (reqcopy == NULL)
1432 OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1434 GOTO(out_free, rc = -ENOMEM);
1437 reqcopy->rq_reply_state = NULL;
1438 reqcopy->rq_rep_swab_mask = 0;
1439 reqcopy->rq_pack_bulk = 0;
1440 reqcopy->rq_pack_udesc = 0;
1441 reqcopy->rq_packed_final = 0;
1442 sptlrpc_svc_ctx_addref(reqcopy);
1443 /* We only need the reqmsg for the magic */
1444 reqcopy->rq_reqmsg = reqmsg;
1445 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1448 * tgt_brw_read() and tgt_brw_write() may have decided not to reply.
1449 * Without this check, we would fail the rq_no_reply assertion in
1450 * ptlrpc_send_reply().
1452 if (reqcopy->rq_no_reply)
1453 GOTO(out, rc = -ETIMEDOUT);
1455 LASSERT(atomic_read(&req->rq_refcount));
1456 /** if it is last refcount then early reply isn't needed */
1457 if (atomic_read(&req->rq_refcount) == 1) {
1458 DEBUG_REQ(D_ADAPTTO, reqcopy,
1459 "Normal reply already sent out, abort sending early reply\n");
1460 GOTO(out, rc = -EINVAL);
1463 /* Connection ref */
1464 reqcopy->rq_export = class_conn2export(
1465 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1466 if (reqcopy->rq_export == NULL)
1467 GOTO(out, rc = -ENODEV);
1470 class_export_rpc_inc(reqcopy->rq_export);
1471 if (reqcopy->rq_export->exp_obd &&
1472 reqcopy->rq_export->exp_obd->obd_fail)
1473 GOTO(out_put, rc = -ENODEV);
1475 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1479 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1482 /* Adjust our own deadline to what we told the client */
1483 req->rq_deadline = newdl;
1484 req->rq_early_count++; /* number sent, server side */
1486 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1490 * Free the (early) reply state from lustre_pack_reply.
1491 * (ptlrpc_send_reply takes it's own rs ref, so this is safe here)
1493 ptlrpc_req_drop_rs(reqcopy);
1496 class_export_rpc_dec(reqcopy->rq_export);
1497 class_export_put(reqcopy->rq_export);
1499 sptlrpc_svc_ctx_decref(reqcopy);
1500 OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1502 ptlrpc_request_cache_free(reqcopy);
1507 * Send early replies to everybody expiring within at_early_margin
1508 * asking for at_extra time
1510 static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1512 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1513 struct ptlrpc_request *rq, *n;
1514 struct list_head work_list;
1517 time64_t now = ktime_get_real_seconds();
1519 int first, counter = 0;
1522 spin_lock(&svcpt->scp_at_lock);
1523 if (svcpt->scp_at_check == 0) {
1524 spin_unlock(&svcpt->scp_at_lock);
1527 delay = ktime_ms_delta(ktime_get(), svcpt->scp_at_checktime);
1528 svcpt->scp_at_check = 0;
1530 if (array->paa_count == 0) {
1531 spin_unlock(&svcpt->scp_at_lock);
1535 /* The timer went off, but maybe the nearest rpc already completed. */
1536 first = array->paa_deadline - now;
1537 if (first > at_early_margin) {
1538 /* We've still got plenty of time. Reset the timer. */
1539 ptlrpc_at_set_timer(svcpt);
1540 spin_unlock(&svcpt->scp_at_lock);
1545 * We're close to a timeout, and we don't know how much longer the
1546 * server will take. Send early replies to everyone expiring soon.
1548 INIT_LIST_HEAD(&work_list);
1550 div_u64_rem(array->paa_deadline, array->paa_size, &index);
1551 count = array->paa_count;
1553 count -= array->paa_reqs_count[index];
1554 list_for_each_entry_safe(rq, n,
1555 &array->paa_reqs_array[index],
1557 if (rq->rq_deadline > now + at_early_margin) {
1558 /* update the earliest deadline */
1559 if (deadline == -1 ||
1560 rq->rq_deadline < deadline)
1561 deadline = rq->rq_deadline;
1566 * ptlrpc_server_drop_request() may drop
1567 * refcount to 0 already. Let's check this and
1568 * don't add entry to work_list
1570 if (likely(atomic_inc_not_zero(&rq->rq_refcount))) {
1571 ptlrpc_at_remove_timed(rq);
1572 list_add(&rq->rq_timed_list, &work_list);
1574 ptlrpc_at_remove_timed(rq);
1580 if (++index >= array->paa_size)
1583 array->paa_deadline = deadline;
1584 /* we have a new earliest deadline, restart the timer */
1585 ptlrpc_at_set_timer(svcpt);
1587 spin_unlock(&svcpt->scp_at_lock);
1590 "timeout in %+ds, asking for %d secs on %d early replies\n",
1591 first, at_extra, counter);
1594 * We're already past request deadlines before we even get a
1595 * chance to send early replies
1597 LCONSOLE_WARN("%s: This server is not able to keep up with request traffic (cpu-bound).\n",
1598 svcpt->scp_service->srv_name);
1599 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, delay=%lld\n",
1600 counter, svcpt->scp_nreqs_incoming,
1601 svcpt->scp_nreqs_active,
1602 at_get(&svcpt->scp_at_estimate), delay);
1606 * we took additional refcount so entries can't be deleted from list, no
1609 while (!list_empty(&work_list)) {
1610 rq = list_entry(work_list.next, struct ptlrpc_request,
1612 list_del_init(&rq->rq_timed_list);
1614 if (ptlrpc_at_send_early_reply(rq) == 0)
1615 ptlrpc_at_add_timed(rq);
1617 ptlrpc_server_drop_request(rq);
1620 RETURN(1); /* return "did_something" for liblustre */
1624 * Check if we are already handling earlier incarnation of this request.
1625 * Called under &req->rq_export->exp_rpc_lock locked
1627 static struct ptlrpc_request*
1628 ptlrpc_server_check_resend_in_progress(struct ptlrpc_request *req)
1630 struct ptlrpc_request *tmp = NULL;
1632 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ||
1633 (atomic_read(&req->rq_export->exp_rpc_count) == 0))
1637 * bulk request are aborted upon reconnect, don't try to
1640 if (req->rq_bulk_write || req->rq_bulk_read)
1644 * This list should not be longer than max_requests in
1645 * flights on the client, so it is not all that long.
1646 * Also we only hit this codepath in case of a resent
1647 * request which makes it even more rarely hit
1649 list_for_each_entry(tmp, &req->rq_export->exp_reg_rpcs,
1651 /* Found duplicate one */
1652 if (tmp->rq_xid == req->rq_xid)
1655 list_for_each_entry(tmp, &req->rq_export->exp_hp_rpcs,
1657 /* Found duplicate one */
1658 if (tmp->rq_xid == req->rq_xid)
1664 DEBUG_REQ(D_HA, req, "Found duplicate req in processing");
1665 DEBUG_REQ(D_HA, tmp, "Request being processed");
1670 * Check if a request should be assigned with a high priority.
1672 * \retval < 0: error occurred
1673 * 0: normal RPC request
1674 * +1: high priority request
1676 static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt,
1677 struct ptlrpc_request *req)
1682 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL) {
1683 rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
1690 if (req->rq_export != NULL && req->rq_ops != NULL) {
1692 * Perform request specific check. We should do this
1693 * check before the request is added into exp_hp_rpcs
1694 * list otherwise it may hit swab race at LU-1044.
1696 if (req->rq_ops->hpreq_check != NULL) {
1697 rc = req->rq_ops->hpreq_check(req);
1698 if (rc == -ESTALE) {
1699 req->rq_status = rc;
1703 * can only return error,
1704 * 0 for normal request,
1705 * or 1 for high priority request
1714 /** Remove the request from the export list. */
1715 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req)
1718 if (req->rq_export) {
1720 * refresh lock timeout again so that client has more
1721 * room to send lock cancel RPC.
1723 if (req->rq_ops && req->rq_ops->hpreq_fini)
1724 req->rq_ops->hpreq_fini(req);
1726 spin_lock(&req->rq_export->exp_rpc_lock);
1727 list_del_init(&req->rq_exp_list);
1728 spin_unlock(&req->rq_export->exp_rpc_lock);
1733 static int ptlrpc_hpreq_check(struct ptlrpc_request *req)
1738 static struct ptlrpc_hpreq_ops ptlrpc_hpreq_common = {
1739 .hpreq_check = ptlrpc_hpreq_check,
1742 /* Hi-Priority RPC check by RPC operation code. */
1743 int ptlrpc_hpreq_handler(struct ptlrpc_request *req)
1745 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1748 * Check for export to let only reconnects for not yet evicted
1749 * export to become a HP rpc.
1751 if ((req->rq_export != NULL) &&
1752 (opc == OBD_PING || opc == MDS_CONNECT || opc == OST_CONNECT))
1753 req->rq_ops = &ptlrpc_hpreq_common;
1757 EXPORT_SYMBOL(ptlrpc_hpreq_handler);
1759 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1760 struct ptlrpc_request *req)
1764 struct ptlrpc_request *orig;
1768 rc = ptlrpc_server_hpreq_init(svcpt, req);
1773 ptlrpc_nrs_req_initialize(svcpt, req, hp);
1775 if (req->rq_export != NULL) {
1776 struct obd_export *exp = req->rq_export;
1779 * do search for duplicated xid and the adding to the list
1782 spin_lock_bh(&exp->exp_rpc_lock);
1783 orig = ptlrpc_server_check_resend_in_progress(req);
1784 if (orig && likely(atomic_inc_not_zero(&orig->rq_refcount))) {
1787 spin_unlock_bh(&exp->exp_rpc_lock);
1790 * When the client resend request and the server has
1791 * the previous copy of it, we need to update deadlines,
1792 * to be sure that the client and the server have equal
1793 * request deadlines.
1796 spin_lock(&orig->rq_rqbd->rqbd_svcpt->scp_at_lock);
1797 linked = orig->rq_at_linked;
1799 ptlrpc_at_remove_timed(orig);
1800 spin_unlock(&orig->rq_rqbd->rqbd_svcpt->scp_at_lock);
1801 orig->rq_deadline = req->rq_deadline;
1803 ptlrpc_at_add_timed(orig);
1804 ptlrpc_server_drop_request(orig);
1805 ptlrpc_nrs_req_finalize(req);
1809 if (hp || req->rq_ops != NULL)
1810 list_add(&req->rq_exp_list, &exp->exp_hp_rpcs);
1812 list_add(&req->rq_exp_list, &exp->exp_reg_rpcs);
1813 spin_unlock_bh(&exp->exp_rpc_lock);
1817 * the current thread is not the processing thread for this request
1818 * since that, but request is in exp_hp_list and can be find there.
1819 * Remove all relations between request and old thread.
1821 req->rq_svc_thread->t_env->le_ses = NULL;
1822 req->rq_svc_thread = NULL;
1823 req->rq_session.lc_thread = NULL;
1825 ptlrpc_nrs_req_add(svcpt, req, hp);
1831 * Allow to handle high priority request
1832 * User can call it w/o any lock but need to hold
1833 * ptlrpc_service_part::scp_req_lock to get reliable result
1835 static bool ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1838 int running = svcpt->scp_nthrs_running;
1840 if (!nrs_svcpt_has_hp(svcpt))
1846 if (ptlrpc_nrs_req_throttling_nolock(svcpt, true))
1849 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1850 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1851 /* leave just 1 thread for normal RPCs */
1852 running = PTLRPC_NTHRS_INIT;
1853 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1857 if (svcpt->scp_nreqs_active >= running - 1)
1860 if (svcpt->scp_nhreqs_active == 0)
1863 return !ptlrpc_nrs_req_pending_nolock(svcpt, false) ||
1864 svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1867 static bool ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1870 return ptlrpc_server_allow_high(svcpt, force) &&
1871 ptlrpc_nrs_req_pending_nolock(svcpt, true);
1875 * Only allow normal priority requests on a service that has a high-priority
1876 * queue if forced (i.e. cleanup), if there are other high priority requests
1877 * already being processed (i.e. those threads can service more high-priority
1878 * requests), or if there are enough idle threads that a later thread can do
1879 * a high priority request.
1880 * User can call it w/o any lock but need to hold
1881 * ptlrpc_service_part::scp_req_lock to get reliable result
1883 static bool ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1886 int running = svcpt->scp_nthrs_running;
1888 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1889 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1890 /* leave just 1 thread for normal RPCs */
1891 running = PTLRPC_NTHRS_INIT;
1892 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1899 if (ptlrpc_nrs_req_throttling_nolock(svcpt, false))
1902 if (svcpt->scp_nreqs_active < running - 2)
1905 if (svcpt->scp_nreqs_active >= running - 1)
1908 return svcpt->scp_nhreqs_active > 0 || !nrs_svcpt_has_hp(svcpt);
1911 static bool ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1914 return ptlrpc_server_allow_normal(svcpt, force) &&
1915 ptlrpc_nrs_req_pending_nolock(svcpt, false);
1919 * Returns true if there are requests available in incoming
1920 * request queue for processing and it is allowed to fetch them.
1921 * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1922 * to get reliable result
1923 * \see ptlrpc_server_allow_normal
1924 * \see ptlrpc_server_allow high
1927 bool ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt,
1930 return ptlrpc_server_high_pending(svcpt, force) ||
1931 ptlrpc_server_normal_pending(svcpt, force);
1935 * Fetch a request for processing from queue of unprocessed requests.
1936 * Favors high-priority requests.
1937 * Returns a pointer to fetched request.
1939 static struct ptlrpc_request *
1940 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force)
1942 struct ptlrpc_request *req = NULL;
1946 spin_lock(&svcpt->scp_req_lock);
1948 if (ptlrpc_server_high_pending(svcpt, force)) {
1949 req = ptlrpc_nrs_req_get_nolock(svcpt, true, force);
1951 svcpt->scp_hreq_count++;
1956 if (ptlrpc_server_normal_pending(svcpt, force)) {
1957 req = ptlrpc_nrs_req_get_nolock(svcpt, false, force);
1959 svcpt->scp_hreq_count = 0;
1964 spin_unlock(&svcpt->scp_req_lock);
1968 svcpt->scp_nreqs_active++;
1970 svcpt->scp_nhreqs_active++;
1972 spin_unlock(&svcpt->scp_req_lock);
1974 if (likely(req->rq_export))
1975 class_export_rpc_inc(req->rq_export);
1981 * Handle freshly incoming reqs, add to timed early reply list,
1982 * pass on to regular request queue.
1983 * All incoming requests pass through here before getting into
1984 * ptlrpc_server_handle_req later on.
1986 static int ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt,
1987 struct ptlrpc_thread *thread)
1989 struct ptlrpc_service *svc = svcpt->scp_service;
1990 struct ptlrpc_request *req;
1996 spin_lock(&svcpt->scp_lock);
1997 if (list_empty(&svcpt->scp_req_incoming)) {
1998 spin_unlock(&svcpt->scp_lock);
2002 req = list_entry(svcpt->scp_req_incoming.next,
2003 struct ptlrpc_request, rq_list);
2004 list_del_init(&req->rq_list);
2005 svcpt->scp_nreqs_incoming--;
2007 * Consider this still a "queued" request as far as stats are
2010 spin_unlock(&svcpt->scp_lock);
2012 /* go through security check/transform */
2013 rc = sptlrpc_svc_unwrap_request(req);
2017 case SECSVC_COMPLETE:
2018 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
2027 * for null-flavored rpc, msg has been unpacked by sptlrpc, although
2028 * redo it wouldn't be harmful.
2030 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
2031 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
2033 CERROR("error unpacking request: ptl %d from %s x%llu\n",
2034 svc->srv_req_portal, libcfs_id2str(req->rq_peer),
2040 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
2042 CERROR("error unpacking ptlrpc body: ptl %d from %s x %llu\n",
2043 svc->srv_req_portal, libcfs_id2str(req->rq_peer),
2048 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
2049 lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
2050 CERROR("drop incoming rpc opc %u, x%llu\n",
2051 cfs_fail_val, req->rq_xid);
2056 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
2057 CERROR("wrong packet type received (type=%u) from %s\n",
2058 lustre_msg_get_type(req->rq_reqmsg),
2059 libcfs_id2str(req->rq_peer));
2063 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2067 req->rq_bulk_write = 1;
2071 case MGS_CONFIG_READ:
2072 req->rq_bulk_read = 1;
2076 CDEBUG(D_RPCTRACE, "got req x%llu\n", req->rq_xid);
2078 req->rq_export = class_conn2export(
2079 lustre_msg_get_handle(req->rq_reqmsg));
2080 if (req->rq_export) {
2081 rc = ptlrpc_check_req(req);
2083 rc = sptlrpc_target_export_check(req->rq_export, req);
2085 DEBUG_REQ(D_ERROR, req,
2086 "DROPPING req with illegal security flavor,");
2091 ptlrpc_update_export_timer(req->rq_export, 0);
2094 /* req_in handling should/must be fast */
2095 if (ktime_get_real_seconds() - req->rq_arrival_time.tv_sec > 5)
2096 DEBUG_REQ(D_WARNING, req, "Slow req_in handling %llds",
2097 (s64)(ktime_get_real_seconds() -
2098 req->rq_arrival_time.tv_sec));
2100 /* Set rpc server deadline and add it to the timed list */
2101 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
2102 MSGHDR_AT_SUPPORT) ?
2103 /* The max time the client expects us to take */
2104 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
2106 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
2107 if (unlikely(deadline == 0)) {
2108 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
2112 /* Skip early reply */
2113 if (OBD_FAIL_PRECHECK(OBD_FAIL_MDS_RESEND))
2114 req->rq_deadline += obd_timeout;
2116 req->rq_svc_thread = thread;
2117 if (thread != NULL) {
2119 * initialize request session, it is needed for request
2120 * processing by target
2122 rc = lu_context_init(&req->rq_session, LCT_SERVER_SESSION |
2125 CERROR("%s: failure to initialize session: rc = %d\n",
2126 thread->t_name, rc);
2129 req->rq_session.lc_thread = thread;
2130 lu_context_enter(&req->rq_session);
2131 thread->t_env->le_ses = &req->rq_session;
2134 ptlrpc_at_add_timed(req);
2136 /* Move it over to the request processing queue */
2137 rc = ptlrpc_server_request_add(svcpt, req);
2141 wake_up(&svcpt->scp_waitq);
2145 ptlrpc_server_finish_request(svcpt, req);
2151 * Main incoming request handling logic.
2152 * Calls handler function from service to do actual processing.
2154 static int ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
2155 struct ptlrpc_thread *thread)
2157 struct ptlrpc_service *svc = svcpt->scp_service;
2158 struct ptlrpc_request *request;
2168 request = ptlrpc_server_request_get(svcpt, false);
2169 if (request == NULL)
2172 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
2173 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
2174 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
2175 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
2177 if (unlikely(fail_opc)) {
2178 if (request->rq_export && request->rq_ops)
2179 OBD_FAIL_TIMEOUT(fail_opc, 4);
2182 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
2184 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
2185 libcfs_debug_dumplog();
2187 work_start = ktime_get_real();
2188 arrived = timespec64_to_ktime(request->rq_arrival_time);
2189 timediff_usecs = ktime_us_delta(work_start, arrived);
2190 if (likely(svc->srv_stats != NULL)) {
2191 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
2193 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
2194 svcpt->scp_nreqs_incoming);
2195 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
2196 svcpt->scp_nreqs_active);
2197 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
2198 at_get(&svcpt->scp_at_estimate));
2201 if (likely(request->rq_export)) {
2202 if (unlikely(ptlrpc_check_req(request)))
2204 ptlrpc_update_export_timer(request->rq_export,
2205 div_u64(timediff_usecs,
2210 * Discard requests queued for longer than the deadline.
2211 * The deadline is increased if we send an early reply.
2213 if (ktime_get_real_seconds() > request->rq_deadline) {
2214 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s: deadline %lld:%llds ago\n",
2215 libcfs_id2str(request->rq_peer),
2216 request->rq_deadline -
2217 request->rq_arrival_time.tv_sec,
2218 ktime_get_real_seconds() - request->rq_deadline);
2223 "Handling RPC req@%p pname:cluuid+ref:pid:xid:nid:opc:job %s:%s+%d:%d:x%llu:%s:%d:%s\n",
2224 request, current_comm(),
2225 (request->rq_export ?
2226 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2227 (request->rq_export ?
2228 atomic_read(&request->rq_export->exp_refcount) : -99),
2229 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
2230 libcfs_id2str(request->rq_peer),
2231 lustre_msg_get_opc(request->rq_reqmsg),
2232 lustre_msg_get_jobid(request->rq_reqmsg));
2234 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
2235 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
2237 CDEBUG(D_NET, "got req %llu\n", request->rq_xid);
2239 /* re-assign request and sesson thread to the current one */
2240 request->rq_svc_thread = thread;
2241 if (thread != NULL) {
2242 LASSERT(request->rq_session.lc_thread == NULL);
2243 request->rq_session.lc_thread = thread;
2244 thread->t_env->le_ses = &request->rq_session;
2246 svc->srv_ops.so_req_handler(request);
2248 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
2251 if (unlikely(ktime_get_real_seconds() > request->rq_deadline)) {
2252 DEBUG_REQ(D_WARNING, request,
2253 "Request took longer than estimated (%lld:%llds); "
2254 "client may timeout.",
2255 request->rq_deadline -
2256 request->rq_arrival_time.tv_sec,
2257 ktime_get_real_seconds() - request->rq_deadline);
2260 work_end = ktime_get_real();
2261 timediff_usecs = ktime_us_delta(work_end, work_start);
2262 arrived_usecs = ktime_us_delta(work_end, arrived);
2264 "Handled RPC req@%p pname:cluuid+ref:pid:xid:nid:opc:job %s:%s+%d:%d:x%llu:%s:%d:%s Request processed in %lldus (%lldus total) trans %llu rc %d/%d\n",
2265 request, current_comm(),
2266 (request->rq_export ?
2267 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2268 (request->rq_export ?
2269 atomic_read(&request->rq_export->exp_refcount) : -99),
2270 lustre_msg_get_status(request->rq_reqmsg),
2272 libcfs_id2str(request->rq_peer),
2273 lustre_msg_get_opc(request->rq_reqmsg),
2274 lustre_msg_get_jobid(request->rq_reqmsg),
2277 (request->rq_repmsg ?
2278 lustre_msg_get_transno(request->rq_repmsg) :
2279 request->rq_transno),
2281 (request->rq_repmsg ?
2282 lustre_msg_get_status(request->rq_repmsg) : -999));
2283 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
2284 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
2285 int opc = opcode_offset(op);
2287 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
2288 LASSERT(opc < LUSTRE_MAX_OPCODES);
2289 lprocfs_counter_add(svc->srv_stats,
2290 opc + EXTRA_MAX_OPCODES,
2294 if (unlikely(request->rq_early_count)) {
2295 DEBUG_REQ(D_ADAPTTO, request,
2296 "sent %d early replies before finishing in %llds",
2297 request->rq_early_count,
2298 div_u64(arrived_usecs, USEC_PER_SEC));
2301 ptlrpc_server_finish_active_request(svcpt, request);
2307 * An internal function to process a single reply state object.
2309 static int ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
2311 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
2312 struct ptlrpc_service *svc = svcpt->scp_service;
2313 struct obd_export *exp;
2319 exp = rs->rs_export;
2321 LASSERT(rs->rs_difficult);
2322 LASSERT(rs->rs_scheduled);
2323 LASSERT(list_empty(&rs->rs_list));
2326 * The disk commit callback holds exp_uncommitted_replies_lock while it
2327 * iterates over newly committed replies, removing them from
2328 * exp_uncommitted_replies. It then drops this lock and schedules the
2329 * replies it found for handling here.
2331 * We can avoid contention for exp_uncommitted_replies_lock between the
2332 * HRT threads and further commit callbacks by checking rs_committed
2333 * which is set in the commit callback while it holds both
2334 * rs_lock and exp_uncommitted_reples.
2336 * If we see rs_committed clear, the commit callback _may_ not have
2337 * handled this reply yet and we race with it to grab
2338 * exp_uncommitted_replies_lock before removing the reply from
2339 * exp_uncommitted_replies. Note that if we lose the race and the
2340 * reply has already been removed, list_del_init() is a noop.
2342 * If we see rs_committed set, we know the commit callback is handling,
2343 * or has handled this reply since store reordering might allow us to
2344 * see rs_committed set out of sequence. But since this is done
2345 * holding rs_lock, we can be sure it has all completed once we hold
2346 * rs_lock, which we do right next.
2348 if (!rs->rs_committed) {
2350 * if rs was commited, no need to convert locks, don't check
2351 * rs_committed here because rs may never be added into
2352 * exp_uncommitted_replies and this flag never be set, see
2353 * target_send_reply()
2355 if (rs->rs_convert_lock &&
2356 rs->rs_transno > exp->exp_last_committed) {
2357 struct ldlm_lock *lock;
2358 struct ldlm_lock *ack_locks[RS_MAX_LOCKS] = { NULL };
2360 spin_lock(&rs->rs_lock);
2361 if (rs->rs_convert_lock &&
2362 rs->rs_transno > exp->exp_last_committed) {
2363 nlocks = rs->rs_nlocks;
2364 while (nlocks-- > 0) {
2366 * NB don't assume rs is always handled
2367 * by the same service thread (see
2368 * ptlrpc_hr_select, so REP-ACK hr may
2369 * race with trans commit, while the
2370 * latter will release locks, get locks
2371 * here early to convert to COS mode
2374 lock = ldlm_handle2lock(
2375 &rs->rs_locks[nlocks]);
2377 ack_locks[nlocks] = lock;
2378 rs->rs_modes[nlocks] = LCK_COS;
2380 nlocks = rs->rs_nlocks;
2381 rs->rs_convert_lock = 0;
2383 * clear rs_scheduled so that commit callback
2384 * can schedule again
2386 rs->rs_scheduled = 0;
2387 spin_unlock(&rs->rs_lock);
2389 while (nlocks-- > 0) {
2390 lock = ack_locks[nlocks];
2391 ldlm_lock_mode_downgrade(lock, LCK_COS);
2392 LDLM_LOCK_PUT(lock);
2396 spin_unlock(&rs->rs_lock);
2399 spin_lock(&exp->exp_uncommitted_replies_lock);
2400 list_del_init(&rs->rs_obd_list);
2401 spin_unlock(&exp->exp_uncommitted_replies_lock);
2404 spin_lock(&exp->exp_lock);
2405 /* Noop if removed already */
2406 list_del_init(&rs->rs_exp_list);
2407 spin_unlock(&exp->exp_lock);
2409 spin_lock(&rs->rs_lock);
2411 been_handled = rs->rs_handled;
2414 nlocks = rs->rs_nlocks; /* atomic "steal", but */
2415 rs->rs_nlocks = 0; /* locks still on rs_locks! */
2417 if (nlocks == 0 && !been_handled) {
2419 * If we see this, we should already have seen the warning
2420 * in mds_steal_ack_locks()
2423 "All locks stolen from rs %p x%lld.t%lld o%d NID %s\n",
2424 rs, rs->rs_xid, rs->rs_transno, rs->rs_opc,
2425 libcfs_nid2str(exp->exp_connection->c_peer.nid));
2428 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
2429 spin_unlock(&rs->rs_lock);
2431 if (!been_handled && rs->rs_on_net) {
2432 LNetMDUnlink(rs->rs_md_h);
2433 /* Ignore return code; we're racing with completion */
2436 while (nlocks-- > 0)
2437 ldlm_lock_decref(&rs->rs_locks[nlocks],
2438 rs->rs_modes[nlocks]);
2440 spin_lock(&rs->rs_lock);
2443 rs->rs_scheduled = 0;
2444 rs->rs_convert_lock = 0;
2446 if (!rs->rs_on_net) {
2448 spin_unlock(&rs->rs_lock);
2450 class_export_put(exp);
2451 rs->rs_export = NULL;
2452 ptlrpc_rs_decref(rs);
2453 if (atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
2454 svc->srv_is_stopping)
2455 wake_up_all(&svcpt->scp_waitq);
2459 /* still on the net; callback will schedule */
2460 spin_unlock(&rs->rs_lock);
2465 static void ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
2467 int avail = svcpt->scp_nrqbds_posted;
2468 int low_water = test_req_buffer_pressure ? 0 :
2469 svcpt->scp_service->srv_nbuf_per_group / 2;
2471 /* NB I'm not locking; just looking. */
2474 * CAVEAT EMPTOR: We might be allocating buffers here because we've
2475 * allowed the request history to grow out of control. We could put a
2476 * sanity check on that here and cull some history if we need the
2480 if (avail <= low_water)
2481 ptlrpc_grow_req_bufs(svcpt, 1);
2483 if (svcpt->scp_service->srv_stats) {
2484 lprocfs_counter_add(svcpt->scp_service->srv_stats,
2485 PTLRPC_REQBUF_AVAIL_CNTR, avail);
2489 static int ptlrpc_retry_rqbds(void *arg)
2491 struct ptlrpc_service_part *svcpt = (struct ptlrpc_service_part *)arg;
2493 svcpt->scp_rqbd_timeout = 0;
2497 static inline int ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
2499 return svcpt->scp_nreqs_active <
2500 svcpt->scp_nthrs_running - 1 -
2501 (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL);
2505 * allowed to create more threads
2506 * user can call it w/o any lock but need to hold
2507 * ptlrpc_service_part::scp_lock to get reliable result
2509 static inline int ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
2511 return svcpt->scp_nthrs_running +
2512 svcpt->scp_nthrs_starting <
2513 svcpt->scp_service->srv_nthrs_cpt_limit;
2517 * too many requests and allowed to create more threads
2519 static inline int ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
2521 return !ptlrpc_threads_enough(svcpt) &&
2522 ptlrpc_threads_increasable(svcpt);
2525 static inline int ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2527 return thread_is_stopping(thread) ||
2528 thread->t_svcpt->scp_service->srv_is_stopping;
2531 /* stop the highest numbered thread if there are too many threads running */
2532 static inline bool ptlrpc_thread_should_stop(struct ptlrpc_thread *thread)
2534 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2536 return thread->t_id >= svcpt->scp_service->srv_nthrs_cpt_limit &&
2537 thread->t_id == svcpt->scp_thr_nextid - 1;
2540 static void ptlrpc_stop_thread(struct ptlrpc_thread *thread)
2542 CDEBUG(D_INFO, "Stopping thread %s #%u\n",
2543 thread->t_svcpt->scp_service->srv_thread_name, thread->t_id);
2544 thread_add_flags(thread, SVC_STOPPING);
2547 static inline void ptlrpc_thread_stop(struct ptlrpc_thread *thread)
2549 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2551 spin_lock(&svcpt->scp_lock);
2552 if (ptlrpc_thread_should_stop(thread)) {
2553 ptlrpc_stop_thread(thread);
2554 svcpt->scp_thr_nextid--;
2556 spin_unlock(&svcpt->scp_lock);
2559 static inline int ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
2561 return !list_empty(&svcpt->scp_rqbd_idle) &&
2562 svcpt->scp_rqbd_timeout == 0;
2566 ptlrpc_at_check(struct ptlrpc_service_part *svcpt)
2568 return svcpt->scp_at_check;
2572 * If a thread runs too long or spends to much time on a single request,
2573 * we want to know about it, so we set up a delayed work item as a watchdog.
2574 * If it fires, we display a stack trace of the delayed thread,
2575 * providing we aren't rate-limited
2577 * Watchdog stack traces are limited to 3 per 'libcfs_watchdog_ratelimit'
2580 static struct ratelimit_state watchdog_limit;
2582 static void ptlrpc_watchdog_fire(struct work_struct *w)
2584 struct ptlrpc_thread *thread = container_of(w, struct ptlrpc_thread,
2586 u64 ms_lapse = ktime_ms_delta(ktime_get(), thread->t_touched);
2587 u32 ms_frac = do_div(ms_lapse, MSEC_PER_SEC);
2589 if (!__ratelimit(&watchdog_limit)) {
2590 LCONSOLE_WARN("%s: service thread pid %u was inactive for %llu.%03u seconds. The thread might be hung, or it might only be slow and will resume later. Dumping the stack trace for debugging purposes:\n",
2591 thread->t_task->comm, thread->t_task->pid,
2594 libcfs_debug_dumpstack(thread->t_task);
2596 LCONSOLE_WARN("%s: service thread pid %u was inactive for %llu.%03u seconds. Watchdog stack traces are limited to 3 per %u seconds, skipping this one.\n",
2597 thread->t_task->comm, thread->t_task->pid,
2598 ms_lapse, ms_frac, libcfs_watchdog_ratelimit);
2602 static void ptlrpc_watchdog_init(struct delayed_work *work, time_t time)
2604 INIT_DELAYED_WORK(work, ptlrpc_watchdog_fire);
2605 schedule_delayed_work(work, cfs_time_seconds(time));
2608 static void ptlrpc_watchdog_disable(struct delayed_work *work)
2610 cancel_delayed_work_sync(work);
2613 static void ptlrpc_watchdog_touch(struct delayed_work *work, time_t time)
2615 struct ptlrpc_thread *thread = container_of(&work->work,
2616 struct ptlrpc_thread,
2618 thread->t_touched = ktime_get();
2619 mod_delayed_work(system_wq, work, cfs_time_seconds(time));
2623 * requests wait on preprocessing
2624 * user can call it w/o any lock but need to hold
2625 * ptlrpc_service_part::scp_lock to get reliable result
2628 ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
2630 return !list_empty(&svcpt->scp_req_incoming);
2633 static __attribute__((__noinline__)) int
2634 ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
2635 struct ptlrpc_thread *thread)
2637 /* Don't exit while there are replies to be handled */
2638 struct l_wait_info lwi = LWI_TIMEOUT(svcpt->scp_rqbd_timeout,
2639 ptlrpc_retry_rqbds, svcpt);
2641 ptlrpc_watchdog_disable(&thread->t_watchdog);
2645 l_wait_event_exclusive_head(svcpt->scp_waitq,
2646 ptlrpc_thread_stopping(thread) ||
2647 ptlrpc_server_request_incoming(svcpt) ||
2648 ptlrpc_server_request_pending(svcpt, false) ||
2649 ptlrpc_rqbd_pending(svcpt) ||
2650 ptlrpc_at_check(svcpt), &lwi);
2652 if (ptlrpc_thread_stopping(thread))
2655 ptlrpc_watchdog_touch(&thread->t_watchdog,
2656 ptlrpc_server_get_timeout(svcpt));
2661 * Main thread body for service threads.
2662 * Waits in a loop waiting for new requests to process to appear.
2663 * Every time an incoming requests is added to its queue, a waitq
2664 * is woken up and one of the threads will handle it.
2666 static int ptlrpc_main(void *arg)
2668 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
2669 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2670 struct ptlrpc_service *svc = svcpt->scp_service;
2671 struct ptlrpc_reply_state *rs;
2672 struct group_info *ginfo = NULL;
2674 int counter = 0, rc = 0;
2678 thread->t_task = current;
2679 thread->t_pid = current_pid();
2680 unshare_fs_struct();
2682 if (svc->srv_cpt_bind) {
2683 rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
2685 CWARN("%s: failed to bind %s on CPT %d\n",
2686 svc->srv_name, thread->t_name, svcpt->scp_cpt);
2690 ginfo = groups_alloc(0);
2692 GOTO(out, rc = -ENOMEM);
2694 set_current_groups(ginfo);
2695 put_group_info(ginfo);
2697 if (svc->srv_ops.so_thr_init != NULL) {
2698 rc = svc->srv_ops.so_thr_init(thread);
2705 GOTO(out_srv_fini, rc = -ENOMEM);
2706 rc = lu_env_add(env);
2710 rc = lu_context_init(&env->le_ctx,
2711 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2713 GOTO(out_env_remove, rc);
2715 thread->t_env = env;
2716 env->le_ctx.lc_thread = thread;
2717 env->le_ctx.lc_cookie = 0x6;
2719 while (!list_empty(&svcpt->scp_rqbd_idle)) {
2720 rc = ptlrpc_server_post_idle_rqbds(svcpt);
2724 CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
2725 svc->srv_name, svcpt->scp_cpt, rc);
2726 GOTO(out_ctx_fini, rc);
2729 /* Alloc reply state structure for this one */
2730 OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2732 GOTO(out_ctx_fini, rc = -ENOMEM);
2734 spin_lock(&svcpt->scp_lock);
2736 LASSERT(thread_is_starting(thread));
2737 thread_clear_flags(thread, SVC_STARTING);
2739 LASSERT(svcpt->scp_nthrs_starting == 1);
2740 svcpt->scp_nthrs_starting--;
2743 * SVC_STOPPING may already be set here if someone else is trying
2744 * to stop the service while this new thread has been dynamically
2745 * forked. We still set SVC_RUNNING to let our creator know that
2746 * we are now running, however we will exit as soon as possible
2748 thread_add_flags(thread, SVC_RUNNING);
2749 svcpt->scp_nthrs_running++;
2750 spin_unlock(&svcpt->scp_lock);
2752 /* wake up our creator in case he's still waiting. */
2753 wake_up(&thread->t_ctl_waitq);
2755 thread->t_touched = ktime_get();
2756 ptlrpc_watchdog_init(&thread->t_watchdog,
2757 ptlrpc_server_get_timeout(svcpt));
2759 spin_lock(&svcpt->scp_rep_lock);
2760 list_add(&rs->rs_list, &svcpt->scp_rep_idle);
2761 wake_up(&svcpt->scp_rep_waitq);
2762 spin_unlock(&svcpt->scp_rep_lock);
2764 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2765 svcpt->scp_nthrs_running);
2767 /* XXX maintain a list of all managed devices: insert here */
2768 while (!ptlrpc_thread_stopping(thread)) {
2769 if (ptlrpc_wait_event(svcpt, thread))
2772 ptlrpc_check_rqbd_pool(svcpt);
2774 if (ptlrpc_threads_need_create(svcpt)) {
2775 /* Ignore return code - we tried... */
2776 ptlrpc_start_thread(svcpt, 0);
2779 /* reset le_ses to initial state */
2781 /* Process all incoming reqs before handling any */
2782 if (ptlrpc_server_request_incoming(svcpt)) {
2783 lu_context_enter(&env->le_ctx);
2784 ptlrpc_server_handle_req_in(svcpt, thread);
2785 lu_context_exit(&env->le_ctx);
2787 /* but limit ourselves in case of flood */
2788 if (counter++ < 100)
2793 if (ptlrpc_at_check(svcpt))
2794 ptlrpc_at_check_timed(svcpt);
2796 if (ptlrpc_server_request_pending(svcpt, false)) {
2797 lu_context_enter(&env->le_ctx);
2798 ptlrpc_server_handle_request(svcpt, thread);
2799 lu_context_exit(&env->le_ctx);
2802 if (ptlrpc_rqbd_pending(svcpt) &&
2803 ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
2805 * I just failed to repost request buffers.
2806 * Wait for a timeout (unless something else
2807 * happens) before I try again
2809 svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
2810 CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
2811 svcpt->scp_nrqbds_posted);
2814 * If the number of threads has been tuned downward and this
2815 * thread should be stopped, then stop in reverse order so the
2816 * the threads always have contiguous thread index values.
2818 if (unlikely(ptlrpc_thread_should_stop(thread)))
2819 ptlrpc_thread_stop(thread);
2822 ptlrpc_watchdog_disable(&thread->t_watchdog);
2825 lu_context_fini(&env->le_ctx);
2831 /* deconstruct service thread state created by ptlrpc_start_thread() */
2832 if (svc->srv_ops.so_thr_done != NULL)
2833 svc->srv_ops.so_thr_done(thread);
2835 CDEBUG(D_RPCTRACE, "%s: service thread [%p:%u] %d exiting: rc = %d\n",
2836 thread->t_name, thread, thread->t_pid, thread->t_id, rc);
2837 spin_lock(&svcpt->scp_lock);
2838 if (thread_test_and_clear_flags(thread, SVC_STARTING))
2839 svcpt->scp_nthrs_starting--;
2841 if (thread_test_and_clear_flags(thread, SVC_RUNNING)) {
2842 /* must know immediately */
2843 svcpt->scp_nthrs_running--;
2847 thread_add_flags(thread, SVC_STOPPED);
2849 wake_up(&thread->t_ctl_waitq);
2850 spin_unlock(&svcpt->scp_lock);
2855 static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
2856 struct list_head *replies)
2860 spin_lock(&hrt->hrt_lock);
2862 list_splice_init(&hrt->hrt_queue, replies);
2863 result = ptlrpc_hr.hr_stopping || !list_empty(replies);
2865 spin_unlock(&hrt->hrt_lock);
2870 * Main body of "handle reply" function.
2871 * It processes acked reply states
2873 static int ptlrpc_hr_main(void *arg)
2875 struct ptlrpc_hr_thread *hrt = (struct ptlrpc_hr_thread *)arg;
2876 struct ptlrpc_hr_partition *hrp = hrt->hrt_partition;
2877 struct list_head replies;
2885 INIT_LIST_HEAD(&replies);
2886 unshare_fs_struct();
2888 rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
2890 char threadname[20];
2892 snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
2893 hrp->hrp_cpt, hrt->hrt_id);
2894 CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
2895 threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
2898 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD |
2899 LCT_REMEMBER | LCT_NOREF);
2903 rc = lu_env_add(env);
2905 GOTO(out_ctx_fini, rc);
2907 atomic_inc(&hrp->hrp_nstarted);
2908 wake_up(&ptlrpc_hr.hr_waitq);
2910 while (!ptlrpc_hr.hr_stopping) {
2911 l_wait_condition(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
2913 while (!list_empty(&replies)) {
2914 struct ptlrpc_reply_state *rs;
2916 rs = list_entry(replies.prev,
2917 struct ptlrpc_reply_state,
2919 list_del_init(&rs->rs_list);
2920 /* refill keys if needed */
2922 lu_context_enter(&env->le_ctx);
2923 ptlrpc_handle_rs(rs);
2924 lu_context_exit(&env->le_ctx);
2928 atomic_inc(&hrp->hrp_nstopped);
2929 wake_up(&ptlrpc_hr.hr_waitq);
2933 lu_context_fini(&env->le_ctx);
2939 static void ptlrpc_stop_hr_threads(void)
2941 struct ptlrpc_hr_partition *hrp;
2945 ptlrpc_hr.hr_stopping = 1;
2947 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2948 if (hrp->hrp_thrs == NULL)
2949 continue; /* uninitialized */
2950 for (j = 0; j < hrp->hrp_nthrs; j++)
2951 wake_up_all(&hrp->hrp_thrs[j].hrt_waitq);
2954 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2955 if (hrp->hrp_thrs == NULL)
2956 continue; /* uninitialized */
2957 wait_event(ptlrpc_hr.hr_waitq,
2958 atomic_read(&hrp->hrp_nstopped) ==
2959 atomic_read(&hrp->hrp_nstarted));
2963 static int ptlrpc_start_hr_threads(void)
2965 struct ptlrpc_hr_partition *hrp;
2971 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2974 for (j = 0; j < hrp->hrp_nthrs; j++) {
2975 struct ptlrpc_hr_thread *hrt = &hrp->hrp_thrs[j];
2976 struct task_struct *task;
2978 task = kthread_run(ptlrpc_hr_main,
2980 "ptlrpc_hr%02d_%03d",
2989 wait_event(ptlrpc_hr.hr_waitq,
2990 atomic_read(&hrp->hrp_nstarted) == j);
2993 CERROR("cannot start reply handler thread %d:%d: rc = %d\n",
2995 ptlrpc_stop_hr_threads();
3003 static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
3005 struct l_wait_info lwi = { 0 };
3006 struct ptlrpc_thread *thread;
3007 struct list_head zombie;
3011 CDEBUG(D_INFO, "Stopping threads for service %s\n",
3012 svcpt->scp_service->srv_name);
3014 INIT_LIST_HEAD(&zombie);
3015 spin_lock(&svcpt->scp_lock);
3016 /* let the thread know that we would like it to stop asap */
3017 list_for_each_entry(thread, &svcpt->scp_threads, t_link)
3018 ptlrpc_stop_thread(thread);
3020 wake_up_all(&svcpt->scp_waitq);
3022 while (!list_empty(&svcpt->scp_threads)) {
3023 thread = list_entry(svcpt->scp_threads.next,
3024 struct ptlrpc_thread, t_link);
3025 if (thread_is_stopped(thread)) {
3026 list_del(&thread->t_link);
3027 list_add(&thread->t_link, &zombie);
3030 spin_unlock(&svcpt->scp_lock);
3032 CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
3033 svcpt->scp_service->srv_thread_name, thread->t_id);
3034 l_wait_event(thread->t_ctl_waitq,
3035 thread_is_stopped(thread), &lwi);
3037 spin_lock(&svcpt->scp_lock);
3040 spin_unlock(&svcpt->scp_lock);
3042 while (!list_empty(&zombie)) {
3043 thread = list_entry(zombie.next,
3044 struct ptlrpc_thread, t_link);
3045 list_del(&thread->t_link);
3046 OBD_FREE_PTR(thread);
3052 * Stops all threads of a particular service \a svc
3054 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
3056 struct ptlrpc_service_part *svcpt;
3061 ptlrpc_service_for_each_part(svcpt, i, svc) {
3062 if (svcpt->scp_service != NULL)
3063 ptlrpc_svcpt_stop_threads(svcpt);
3069 int ptlrpc_start_threads(struct ptlrpc_service *svc)
3077 /* We require 2 threads min, see note in ptlrpc_server_handle_request */
3078 LASSERT(svc->srv_nthrs_cpt_init >= PTLRPC_NTHRS_INIT);
3080 for (i = 0; i < svc->srv_ncpts; i++) {
3081 for (j = 0; j < svc->srv_nthrs_cpt_init; j++) {
3082 rc = ptlrpc_start_thread(svc->srv_parts[i], 1);
3088 /* We have enough threads, don't start more. b=15759 */
3095 CERROR("cannot start %s thread #%d_%d: rc %d\n",
3096 svc->srv_thread_name, i, j, rc);
3097 ptlrpc_stop_all_threads(svc);
3101 int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
3103 struct l_wait_info lwi = { 0 };
3104 struct ptlrpc_thread *thread;
3105 struct ptlrpc_service *svc;
3106 struct task_struct *task;
3111 LASSERT(svcpt != NULL);
3113 svc = svcpt->scp_service;
3115 CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n",
3116 svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running,
3117 svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit);
3120 if (unlikely(svc->srv_is_stopping))
3123 if (!ptlrpc_threads_increasable(svcpt) ||
3124 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
3125 svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1))
3128 OBD_CPT_ALLOC_PTR(thread, svc->srv_cptable, svcpt->scp_cpt);
3131 init_waitqueue_head(&thread->t_ctl_waitq);
3133 spin_lock(&svcpt->scp_lock);
3134 if (!ptlrpc_threads_increasable(svcpt)) {
3135 spin_unlock(&svcpt->scp_lock);
3136 OBD_FREE_PTR(thread);
3140 if (svcpt->scp_nthrs_starting != 0) {
3142 * serialize starting because some modules (obdfilter)
3143 * might require unique and contiguous t_id
3145 LASSERT(svcpt->scp_nthrs_starting == 1);
3146 spin_unlock(&svcpt->scp_lock);
3147 OBD_FREE_PTR(thread);
3149 CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n",
3150 svc->srv_thread_name, svcpt->scp_thr_nextid);
3155 CDEBUG(D_INFO, "Creating thread %s #%d race, retry later\n",
3156 svc->srv_thread_name, svcpt->scp_thr_nextid);
3160 svcpt->scp_nthrs_starting++;
3161 thread->t_id = svcpt->scp_thr_nextid++;
3162 thread_add_flags(thread, SVC_STARTING);
3163 thread->t_svcpt = svcpt;
3165 list_add(&thread->t_link, &svcpt->scp_threads);
3166 spin_unlock(&svcpt->scp_lock);
3168 if (svcpt->scp_cpt >= 0) {
3169 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s%02d_%03d",
3170 svc->srv_thread_name, svcpt->scp_cpt, thread->t_id);
3172 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s_%04d",
3173 svc->srv_thread_name, thread->t_id);
3176 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name);
3177 task = kthread_run(ptlrpc_main, thread, "%s", thread->t_name);
3180 CERROR("cannot start thread '%s': rc = %d\n",
3181 thread->t_name, rc);
3182 spin_lock(&svcpt->scp_lock);
3183 --svcpt->scp_nthrs_starting;
3184 if (thread_is_stopping(thread)) {
3186 * this ptlrpc_thread is being hanled
3187 * by ptlrpc_svcpt_stop_threads now
3189 thread_add_flags(thread, SVC_STOPPED);
3190 wake_up(&thread->t_ctl_waitq);
3191 spin_unlock(&svcpt->scp_lock);
3193 list_del(&thread->t_link);
3194 spin_unlock(&svcpt->scp_lock);
3195 OBD_FREE_PTR(thread);
3203 l_wait_event(thread->t_ctl_waitq,
3204 thread_is_running(thread) || thread_is_stopped(thread),
3207 rc = thread_is_stopped(thread) ? thread->t_id : 0;
3211 int ptlrpc_hr_init(void)
3213 struct ptlrpc_hr_partition *hrp;
3214 struct ptlrpc_hr_thread *hrt;
3222 memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
3223 ptlrpc_hr.hr_cpt_table = cfs_cpt_table;
3225 ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
3227 if (ptlrpc_hr.hr_partitions == NULL)
3230 ratelimit_state_init(&watchdog_limit,
3231 cfs_time_seconds(libcfs_watchdog_ratelimit), 3);
3233 init_waitqueue_head(&ptlrpc_hr.hr_waitq);
3236 weight = cpumask_weight(topology_sibling_cpumask(smp_processor_id()));
3239 cfs_percpt_for_each(hrp, cpt, ptlrpc_hr.hr_partitions) {
3242 atomic_set(&hrp->hrp_nstarted, 0);
3243 atomic_set(&hrp->hrp_nstopped, 0);
3245 hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, cpt);
3246 hrp->hrp_nthrs /= weight;
3247 if (hrp->hrp_nthrs == 0)
3250 OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, cpt,
3251 hrp->hrp_nthrs * sizeof(*hrt));
3252 if (hrp->hrp_thrs == NULL)
3253 GOTO(out, rc = -ENOMEM);
3255 for (i = 0; i < hrp->hrp_nthrs; i++) {
3256 hrt = &hrp->hrp_thrs[i];
3259 hrt->hrt_partition = hrp;
3260 init_waitqueue_head(&hrt->hrt_waitq);
3261 spin_lock_init(&hrt->hrt_lock);
3262 INIT_LIST_HEAD(&hrt->hrt_queue);
3266 rc = ptlrpc_start_hr_threads();
3273 void ptlrpc_hr_fini(void)
3275 struct ptlrpc_hr_partition *hrp;
3278 if (ptlrpc_hr.hr_partitions == NULL)
3281 ptlrpc_stop_hr_threads();
3283 cfs_percpt_for_each(hrp, cpt, ptlrpc_hr.hr_partitions) {
3284 if (hrp->hrp_thrs != NULL) {
3285 OBD_FREE(hrp->hrp_thrs,
3286 hrp->hrp_nthrs * sizeof(hrp->hrp_thrs[0]));
3290 cfs_percpt_free(ptlrpc_hr.hr_partitions);
3291 ptlrpc_hr.hr_partitions = NULL;
3296 * Wait until all already scheduled replies are processed.
3298 static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
3302 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
3305 rc = l_wait_event(svcpt->scp_waitq,
3306 atomic_read(&svcpt->scp_nreps_difficult) == 0, &lwi);
3309 CWARN("Unexpectedly long timeout %s %p\n",
3310 svcpt->scp_service->srv_name, svcpt->scp_service);
3315 ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
3317 struct ptlrpc_service_part *svcpt;
3320 /* early disarm AT timer... */
3321 ptlrpc_service_for_each_part(svcpt, i, svc) {
3322 if (svcpt->scp_service != NULL)
3323 del_timer(&svcpt->scp_at_timer);
3328 ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
3330 struct ptlrpc_service_part *svcpt;
3331 struct ptlrpc_request_buffer_desc *rqbd;
3332 struct l_wait_info lwi;
3337 * All history will be culled when the next request buffer is
3338 * freed in ptlrpc_service_purge_all()
3340 svc->srv_hist_nrqbds_cpt_max = 0;
3342 rc = LNetClearLazyPortal(svc->srv_req_portal);
3345 ptlrpc_service_for_each_part(svcpt, i, svc) {
3346 if (svcpt->scp_service == NULL)
3350 * Unlink all the request buffers. This forces a 'final'
3351 * event with its 'unlink' flag set for each posted rqbd
3353 list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
3355 rc = LNetMDUnlink(rqbd->rqbd_md_h);
3356 LASSERT(rc == 0 || rc == -ENOENT);
3360 ptlrpc_service_for_each_part(svcpt, i, svc) {
3361 if (svcpt->scp_service == NULL)
3365 * Wait for the network to release any buffers
3366 * it's currently filling
3368 spin_lock(&svcpt->scp_lock);
3369 while (svcpt->scp_nrqbds_posted != 0) {
3370 spin_unlock(&svcpt->scp_lock);
3372 * Network access will complete in finite time but
3373 * the HUGE timeout lets us CWARN for visibility
3376 lwi = LWI_TIMEOUT_INTERVAL(
3377 cfs_time_seconds(LONG_UNLINK),
3378 cfs_time_seconds(1), NULL, NULL);
3379 rc = l_wait_event(svcpt->scp_waitq,
3380 svcpt->scp_nrqbds_posted == 0, &lwi);
3381 if (rc == -ETIMEDOUT) {
3382 CWARN("Service %s waiting for request buffers\n",
3383 svcpt->scp_service->srv_name);
3385 spin_lock(&svcpt->scp_lock);
3387 spin_unlock(&svcpt->scp_lock);
3392 ptlrpc_service_purge_all(struct ptlrpc_service *svc)
3394 struct ptlrpc_service_part *svcpt;
3395 struct ptlrpc_request_buffer_desc *rqbd;
3396 struct ptlrpc_request *req;
3397 struct ptlrpc_reply_state *rs;
3400 ptlrpc_service_for_each_part(svcpt, i, svc) {
3401 if (svcpt->scp_service == NULL)
3404 spin_lock(&svcpt->scp_rep_lock);
3405 while (!list_empty(&svcpt->scp_rep_active)) {
3406 rs = list_entry(svcpt->scp_rep_active.next,
3407 struct ptlrpc_reply_state, rs_list);
3408 spin_lock(&rs->rs_lock);
3409 ptlrpc_schedule_difficult_reply(rs);
3410 spin_unlock(&rs->rs_lock);
3412 spin_unlock(&svcpt->scp_rep_lock);
3415 * purge the request queue. NB No new replies (rqbds
3416 * all unlinked) and no service threads, so I'm the only
3417 * thread noodling the request queue now
3419 while (!list_empty(&svcpt->scp_req_incoming)) {
3420 req = list_entry(svcpt->scp_req_incoming.next,
3421 struct ptlrpc_request, rq_list);
3423 list_del(&req->rq_list);
3424 svcpt->scp_nreqs_incoming--;
3425 ptlrpc_server_finish_request(svcpt, req);
3428 while (ptlrpc_server_request_pending(svcpt, true)) {
3429 req = ptlrpc_server_request_get(svcpt, true);
3430 ptlrpc_server_finish_active_request(svcpt, req);
3433 LASSERT(list_empty(&svcpt->scp_rqbd_posted));
3434 LASSERT(svcpt->scp_nreqs_incoming == 0);
3435 LASSERT(svcpt->scp_nreqs_active == 0);
3437 * history should have been culled by
3438 * ptlrpc_server_finish_request
3440 LASSERT(svcpt->scp_hist_nrqbds == 0);
3443 * Now free all the request buffers since nothing
3444 * references them any more...
3447 while (!list_empty(&svcpt->scp_rqbd_idle)) {
3448 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
3449 struct ptlrpc_request_buffer_desc,
3451 ptlrpc_free_rqbd(rqbd);
3453 ptlrpc_wait_replies(svcpt);
3455 while (!list_empty(&svcpt->scp_rep_idle)) {
3456 rs = list_entry(svcpt->scp_rep_idle.next,
3457 struct ptlrpc_reply_state,
3459 list_del(&rs->rs_list);
3460 OBD_FREE_LARGE(rs, svc->srv_max_reply_size);
3466 ptlrpc_service_free(struct ptlrpc_service *svc)
3468 struct ptlrpc_service_part *svcpt;
3469 struct ptlrpc_at_array *array;
3472 ptlrpc_service_for_each_part(svcpt, i, svc) {
3473 if (svcpt->scp_service == NULL)
3476 /* In case somebody rearmed this in the meantime */
3477 del_timer(&svcpt->scp_at_timer);
3478 array = &svcpt->scp_at_array;
3480 if (array->paa_reqs_array != NULL) {
3481 OBD_FREE(array->paa_reqs_array,
3482 sizeof(struct list_head) * array->paa_size);
3483 array->paa_reqs_array = NULL;
3486 if (array->paa_reqs_count != NULL) {
3487 OBD_FREE(array->paa_reqs_count,
3488 sizeof(__u32) * array->paa_size);
3489 array->paa_reqs_count = NULL;
3493 ptlrpc_service_for_each_part(svcpt, i, svc)
3494 OBD_FREE_PTR(svcpt);
3496 if (svc->srv_cpts != NULL)
3497 cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts);
3499 OBD_FREE(svc, offsetof(struct ptlrpc_service,
3500 srv_parts[svc->srv_ncpts]));
3503 int ptlrpc_unregister_service(struct ptlrpc_service *service)
3507 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
3509 service->srv_is_stopping = 1;
3511 mutex_lock(&ptlrpc_all_services_mutex);
3512 list_del_init(&service->srv_list);
3513 mutex_unlock(&ptlrpc_all_services_mutex);
3515 ptlrpc_service_del_atimer(service);
3516 ptlrpc_stop_all_threads(service);
3518 ptlrpc_service_unlink_rqbd(service);
3519 ptlrpc_service_purge_all(service);
3520 ptlrpc_service_nrs_cleanup(service);
3522 ptlrpc_lprocfs_unregister_service(service);
3523 ptlrpc_sysfs_unregister_service(service);
3525 ptlrpc_service_free(service);
3529 EXPORT_SYMBOL(ptlrpc_unregister_service);
3532 * Returns 0 if the service is healthy.
3534 * Right now, it just checks to make sure that requests aren't languishing
3535 * in the queue. We'll use this health check to govern whether a node needs
3536 * to be shot, so it's intentionally non-aggressive.
3538 static int ptlrpc_svcpt_health_check(struct ptlrpc_service_part *svcpt)
3540 struct ptlrpc_request *request = NULL;
3541 struct timespec64 right_now;
3542 struct timespec64 timediff;
3544 ktime_get_real_ts64(&right_now);
3546 spin_lock(&svcpt->scp_req_lock);
3547 /* How long has the next entry been waiting? */
3548 if (ptlrpc_server_high_pending(svcpt, true))
3549 request = ptlrpc_nrs_req_peek_nolock(svcpt, true);
3550 else if (ptlrpc_server_normal_pending(svcpt, true))
3551 request = ptlrpc_nrs_req_peek_nolock(svcpt, false);
3553 if (request == NULL) {
3554 spin_unlock(&svcpt->scp_req_lock);
3558 timediff = timespec64_sub(right_now, request->rq_arrival_time);
3559 spin_unlock(&svcpt->scp_req_lock);
3561 if ((timediff.tv_sec) >
3562 (AT_OFF ? obd_timeout * 3 / 2 : at_max)) {
3563 CERROR("%s: unhealthy - request has been waiting %llds\n",
3564 svcpt->scp_service->srv_name, (s64)timediff.tv_sec);
3572 ptlrpc_service_health_check(struct ptlrpc_service *svc)
3574 struct ptlrpc_service_part *svcpt;
3580 ptlrpc_service_for_each_part(svcpt, i, svc) {
3581 int rc = ptlrpc_svcpt_health_check(svcpt);
3588 EXPORT_SYMBOL(ptlrpc_service_health_check);