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,
453 /* assume we will post successfully */
454 svcpt->scp_nrqbds_posted++;
455 list_move(&rqbd->rqbd_list, &svcpt->scp_rqbd_posted);
457 spin_unlock(&svcpt->scp_lock);
459 rc = ptlrpc_register_rqbd(rqbd);
466 spin_lock(&svcpt->scp_lock);
468 svcpt->scp_nrqbds_posted--;
469 list_move_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
472 * Don't complain if no request buffers are posted right now; LNET
473 * won't drop requests because we set the portal lazy!
476 spin_unlock(&svcpt->scp_lock);
481 static void ptlrpc_at_timer(cfs_timer_cb_arg_t data)
483 struct ptlrpc_service_part *svcpt;
485 svcpt = cfs_from_timer(svcpt, data, scp_at_timer);
487 svcpt->scp_at_check = 1;
488 svcpt->scp_at_checktime = ktime_get();
489 wake_up(&svcpt->scp_waitq);
492 static void ptlrpc_server_nthreads_check(struct ptlrpc_service *svc,
493 struct ptlrpc_service_conf *conf)
495 struct ptlrpc_service_thr_conf *tc = &conf->psc_thr;
502 * Common code for estimating & validating threads number.
503 * CPT affinity service could have percpt thread-pool instead
504 * of a global thread-pool, which means user might not always
505 * get the threads number they give it in conf::tc_nthrs_user
506 * even they did set. It's because we need to validate threads
507 * number for each CPT to guarantee each pool will have enough
508 * threads to keep the service healthy.
510 init = PTLRPC_NTHRS_INIT + (svc->srv_ops.so_hpreq_handler != NULL);
511 init = max_t(int, init, tc->tc_nthrs_init);
514 * NB: please see comments in lustre_lnet.h for definition
515 * details of these members
517 LASSERT(tc->tc_nthrs_max != 0);
519 if (tc->tc_nthrs_user != 0) {
521 * In case there is a reason to test a service with many
522 * threads, we give a less strict check here, it can
523 * be up to 8 * nthrs_max
525 total = min(tc->tc_nthrs_max * 8, tc->tc_nthrs_user);
526 nthrs = total / svc->srv_ncpts;
527 init = max(init, nthrs);
531 total = tc->tc_nthrs_max;
532 if (tc->tc_nthrs_base == 0) {
534 * don't care about base threads number per partition,
535 * this is most for non-affinity service
537 nthrs = total / svc->srv_ncpts;
541 nthrs = tc->tc_nthrs_base;
542 if (svc->srv_ncpts == 1) {
546 * NB: Increase the base number if it's single partition
547 * and total number of cores/HTs is larger or equal to 4.
548 * result will always < 2 * nthrs_base
550 weight = cfs_cpt_weight(svc->srv_cptable, CFS_CPT_ANY);
551 for (i = 1; (weight >> (i + 1)) != 0 && /* >= 4 cores/HTs */
552 (tc->tc_nthrs_base >> i) != 0; i++)
553 nthrs += tc->tc_nthrs_base >> i;
556 if (tc->tc_thr_factor != 0) {
557 int factor = tc->tc_thr_factor;
561 * User wants to increase number of threads with for
562 * each CPU core/HT, most likely the factor is larger than
563 * one thread/core because service threads are supposed to
564 * be blocked by lock or wait for IO.
567 * Amdahl's law says that adding processors wouldn't give
568 * a linear increasing of parallelism, so it's nonsense to
569 * have too many threads no matter how many cores/HTs
574 (topology_sibling_cpumask(smp_processor_id())) > 1) {
575 /* weight is # of HTs */
576 /* depress thread factor for hyper-thread */
577 factor = factor - (factor >> 1) + (factor >> 3);
581 weight = cfs_cpt_weight(svc->srv_cptable, 0);
583 for (; factor > 0 && weight > 0; factor--, weight -= fade)
584 nthrs += min(weight, fade) * factor;
587 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
588 nthrs = max(tc->tc_nthrs_base,
589 tc->tc_nthrs_max / svc->srv_ncpts);
592 nthrs = max(nthrs, tc->tc_nthrs_init);
593 svc->srv_nthrs_cpt_limit = nthrs;
594 svc->srv_nthrs_cpt_init = init;
596 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
598 "%s: This service may have more threads (%d) than the given soft limit (%d)\n",
599 svc->srv_name, nthrs * svc->srv_ncpts,
605 * Initialize percpt data for a service
607 static int ptlrpc_service_part_init(struct ptlrpc_service *svc,
608 struct ptlrpc_service_part *svcpt, int cpt)
610 struct ptlrpc_at_array *array;
615 svcpt->scp_cpt = cpt;
616 INIT_LIST_HEAD(&svcpt->scp_threads);
618 /* rqbd and incoming request queue */
619 spin_lock_init(&svcpt->scp_lock);
620 mutex_init(&svcpt->scp_mutex);
621 INIT_LIST_HEAD(&svcpt->scp_rqbd_idle);
622 INIT_LIST_HEAD(&svcpt->scp_rqbd_posted);
623 INIT_LIST_HEAD(&svcpt->scp_req_incoming);
624 init_waitqueue_head(&svcpt->scp_waitq);
625 /* history request & rqbd list */
626 INIT_LIST_HEAD(&svcpt->scp_hist_reqs);
627 INIT_LIST_HEAD(&svcpt->scp_hist_rqbds);
629 /* acitve requests and hp requests */
630 spin_lock_init(&svcpt->scp_req_lock);
633 spin_lock_init(&svcpt->scp_rep_lock);
634 INIT_LIST_HEAD(&svcpt->scp_rep_active);
635 INIT_LIST_HEAD(&svcpt->scp_rep_idle);
636 init_waitqueue_head(&svcpt->scp_rep_waitq);
637 atomic_set(&svcpt->scp_nreps_difficult, 0);
639 /* adaptive timeout */
640 spin_lock_init(&svcpt->scp_at_lock);
641 array = &svcpt->scp_at_array;
643 size = at_est2timeout(at_max);
644 array->paa_size = size;
645 array->paa_count = 0;
646 array->paa_deadline = -1;
648 /* allocate memory for scp_at_array (ptlrpc_at_array) */
649 OBD_CPT_ALLOC(array->paa_reqs_array,
650 svc->srv_cptable, cpt, sizeof(struct list_head) * size);
651 if (array->paa_reqs_array == NULL)
654 for (index = 0; index < size; index++)
655 INIT_LIST_HEAD(&array->paa_reqs_array[index]);
657 OBD_CPT_ALLOC(array->paa_reqs_count,
658 svc->srv_cptable, cpt, sizeof(__u32) * size);
659 if (array->paa_reqs_count == NULL)
662 cfs_timer_setup(&svcpt->scp_at_timer, ptlrpc_at_timer,
663 (unsigned long)svcpt, 0);
666 * At SOW, service time should be quick; 10s seems generous. If client
667 * timeout is less than this, we'll be sending an early reply.
669 at_init(&svcpt->scp_at_estimate, 10, 0);
671 /* assign this before call ptlrpc_grow_req_bufs */
672 svcpt->scp_service = svc;
673 /* Now allocate the request buffers, but don't post them now */
674 rc = ptlrpc_grow_req_bufs(svcpt, 0);
676 * We shouldn't be under memory pressure at startup, so
677 * fail if we can't allocate all our buffers at this time.
685 if (array->paa_reqs_count != NULL) {
686 OBD_FREE(array->paa_reqs_count, sizeof(__u32) * size);
687 array->paa_reqs_count = NULL;
690 if (array->paa_reqs_array != NULL) {
691 OBD_FREE(array->paa_reqs_array,
692 sizeof(struct list_head) * array->paa_size);
693 array->paa_reqs_array = NULL;
700 * Initialize service on a given portal.
701 * This includes starting serving threads , allocating and posting rqbds and
704 struct ptlrpc_service *ptlrpc_register_service(struct ptlrpc_service_conf *conf,
706 struct dentry *debugfs_entry)
708 struct ptlrpc_service_cpt_conf *cconf = &conf->psc_cpt;
709 struct ptlrpc_service *service;
710 struct ptlrpc_service_part *svcpt;
711 struct cfs_cpt_table *cptable;
720 LASSERT(conf->psc_buf.bc_nbufs > 0);
721 LASSERT(conf->psc_buf.bc_buf_size >=
722 conf->psc_buf.bc_req_max_size + SPTLRPC_MAX_PAYLOAD);
723 LASSERT(conf->psc_thr.tc_ctx_tags != 0);
725 cptable = cconf->cc_cptable;
727 cptable = cfs_cpt_table;
729 if (conf->psc_thr.tc_cpu_bind > 1) {
730 CERROR("%s: Invalid cpu bind value %d, only 1 or 0 allowed\n",
731 conf->psc_name, conf->psc_thr.tc_cpu_bind);
732 RETURN(ERR_PTR(-EINVAL));
735 if (!cconf->cc_affinity) {
738 ncpts = cfs_cpt_number(cptable);
739 if (cconf->cc_pattern != NULL) {
740 struct cfs_expr_list *el;
742 rc = cfs_expr_list_parse(cconf->cc_pattern,
743 strlen(cconf->cc_pattern),
746 CERROR("%s: invalid CPT pattern string: %s",
747 conf->psc_name, cconf->cc_pattern);
748 RETURN(ERR_PTR(-EINVAL));
751 rc = cfs_expr_list_values(el, ncpts, &cpts);
752 cfs_expr_list_free(el);
754 CERROR("%s: failed to parse CPT array %s: %d\n",
755 conf->psc_name, cconf->cc_pattern, rc);
757 OBD_FREE(cpts, sizeof(*cpts) * ncpts);
758 RETURN(ERR_PTR(rc < 0 ? rc : -EINVAL));
764 OBD_ALLOC(service, offsetof(struct ptlrpc_service, srv_parts[ncpts]));
765 if (service == NULL) {
767 OBD_FREE(cpts, sizeof(*cpts) * ncpts);
768 RETURN(ERR_PTR(-ENOMEM));
771 service->srv_cptable = cptable;
772 service->srv_cpts = cpts;
773 service->srv_ncpts = ncpts;
774 service->srv_cpt_bind = conf->psc_thr.tc_cpu_bind;
776 service->srv_cpt_bits = 0; /* it's zero already, easy to read... */
777 while ((1 << service->srv_cpt_bits) < cfs_cpt_number(cptable))
778 service->srv_cpt_bits++;
781 spin_lock_init(&service->srv_lock);
782 service->srv_name = conf->psc_name;
783 service->srv_watchdog_factor = conf->psc_watchdog_factor;
784 INIT_LIST_HEAD(&service->srv_list); /* for safty of cleanup */
786 /* buffer configuration */
787 service->srv_nbuf_per_group = test_req_buffer_pressure ?
788 1 : conf->psc_buf.bc_nbufs;
789 /* do not limit max number of rqbds by default */
790 service->srv_nrqbds_max = 0;
792 service->srv_max_req_size = conf->psc_buf.bc_req_max_size +
794 service->srv_buf_size = conf->psc_buf.bc_buf_size;
795 service->srv_rep_portal = conf->psc_buf.bc_rep_portal;
796 service->srv_req_portal = conf->psc_buf.bc_req_portal;
798 /* With slab/alloc_pages buffer size will be rounded up to 2^n */
799 if (service->srv_buf_size & (service->srv_buf_size - 1)) {
800 int round = size_roundup_power2(service->srv_buf_size);
802 service->srv_buf_size = round;
805 /* Increase max reply size to next power of two */
806 service->srv_max_reply_size = 1;
807 while (service->srv_max_reply_size <
808 conf->psc_buf.bc_rep_max_size + SPTLRPC_MAX_PAYLOAD)
809 service->srv_max_reply_size <<= 1;
811 service->srv_thread_name = conf->psc_thr.tc_thr_name;
812 service->srv_ctx_tags = conf->psc_thr.tc_ctx_tags;
813 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
814 service->srv_ops = conf->psc_ops;
816 for (i = 0; i < ncpts; i++) {
817 if (!cconf->cc_affinity)
820 cpt = cpts != NULL ? cpts[i] : i;
822 OBD_CPT_ALLOC(svcpt, cptable, cpt, sizeof(*svcpt));
824 GOTO(failed, rc = -ENOMEM);
826 service->srv_parts[i] = svcpt;
827 rc = ptlrpc_service_part_init(service, svcpt, cpt);
832 ptlrpc_server_nthreads_check(service, conf);
834 rc = LNetSetLazyPortal(service->srv_req_portal);
837 mutex_lock(&ptlrpc_all_services_mutex);
838 list_add(&service->srv_list, &ptlrpc_all_services);
839 mutex_unlock(&ptlrpc_all_services_mutex);
842 rc = ptlrpc_sysfs_register_service(parent, service);
847 if (debugfs_entry != NULL)
848 ptlrpc_ldebugfs_register_service(debugfs_entry, service);
850 rc = ptlrpc_service_nrs_setup(service);
854 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
855 service->srv_name, service->srv_req_portal);
857 rc = ptlrpc_start_threads(service);
859 CERROR("Failed to start threads for service %s: %d\n",
860 service->srv_name, rc);
866 ptlrpc_unregister_service(service);
869 EXPORT_SYMBOL(ptlrpc_register_service);
872 * to actually free the request, must be called without holding svc_lock.
873 * note it's caller's responsibility to unlink req->rq_list.
875 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
877 LASSERT(atomic_read(&req->rq_refcount) == 0);
878 LASSERT(list_empty(&req->rq_timed_list));
881 * DEBUG_REQ() assumes the reply state of a request with a valid
882 * ref will not be destroyed until that reference is dropped.
884 ptlrpc_req_drop_rs(req);
886 sptlrpc_svc_ctx_decref(req);
888 if (req != &req->rq_rqbd->rqbd_req) {
890 * NB request buffers use an embedded
891 * req if the incoming req unlinked the
892 * MD; this isn't one of them!
894 ptlrpc_request_cache_free(req);
899 * drop a reference count of the request. if it reaches 0, we either
900 * put it into history list, or free it immediately.
902 void ptlrpc_server_drop_request(struct ptlrpc_request *req)
904 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
905 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
906 struct ptlrpc_service *svc = svcpt->scp_service;
908 struct list_head *tmp;
909 struct list_head *nxt;
911 if (!atomic_dec_and_test(&req->rq_refcount))
914 if (req->rq_session.lc_state == LCS_ENTERED) {
915 lu_context_exit(&req->rq_session);
916 lu_context_fini(&req->rq_session);
919 if (req->rq_at_linked) {
920 spin_lock(&svcpt->scp_at_lock);
922 * recheck with lock, in case it's unlinked by
923 * ptlrpc_at_check_timed()
925 if (likely(req->rq_at_linked))
926 ptlrpc_at_remove_timed(req);
927 spin_unlock(&svcpt->scp_at_lock);
930 LASSERT(list_empty(&req->rq_timed_list));
932 /* finalize request */
933 if (req->rq_export) {
934 class_export_put(req->rq_export);
935 req->rq_export = NULL;
938 spin_lock(&svcpt->scp_lock);
940 list_add(&req->rq_list, &rqbd->rqbd_reqs);
942 refcount = --(rqbd->rqbd_refcount);
944 /* request buffer is now idle: add to history */
945 list_move_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds);
946 svcpt->scp_hist_nrqbds++;
950 * I expect only about 1 or 2 rqbds need to be recycled here
952 while (svcpt->scp_hist_nrqbds > svc->srv_hist_nrqbds_cpt_max) {
953 rqbd = list_entry(svcpt->scp_hist_rqbds.next,
954 struct ptlrpc_request_buffer_desc,
957 list_del(&rqbd->rqbd_list);
958 svcpt->scp_hist_nrqbds--;
961 * remove rqbd's reqs from svc's req history while
962 * I've got the service lock
964 list_for_each(tmp, &rqbd->rqbd_reqs) {
965 req = list_entry(tmp, struct ptlrpc_request,
967 /* Track the highest culled req seq */
968 if (req->rq_history_seq >
969 svcpt->scp_hist_seq_culled) {
970 svcpt->scp_hist_seq_culled =
973 list_del(&req->rq_history_list);
976 spin_unlock(&svcpt->scp_lock);
978 list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
979 req = list_entry(rqbd->rqbd_reqs.next,
980 struct ptlrpc_request,
982 list_del(&req->rq_list);
983 ptlrpc_server_free_request(req);
986 spin_lock(&svcpt->scp_lock);
988 * now all reqs including the embedded req has been
989 * disposed, schedule request buffer for re-use
990 * or free it to drain some in excess.
992 LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) == 0);
993 if (svcpt->scp_nrqbds_posted >=
994 svc->srv_nbuf_per_group ||
995 (svc->srv_nrqbds_max != 0 &&
996 svcpt->scp_nrqbds_total > svc->srv_nrqbds_max) ||
997 test_req_buffer_pressure) {
998 /* like in ptlrpc_free_rqbd() */
999 svcpt->scp_nrqbds_total--;
1000 OBD_FREE_LARGE(rqbd->rqbd_buffer,
1004 list_add_tail(&rqbd->rqbd_list,
1005 &svcpt->scp_rqbd_idle);
1009 spin_unlock(&svcpt->scp_lock);
1010 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
1011 /* If we are low on memory, we are not interested in history */
1012 list_del(&req->rq_list);
1013 list_del_init(&req->rq_history_list);
1015 /* Track the highest culled req seq */
1016 if (req->rq_history_seq > svcpt->scp_hist_seq_culled)
1017 svcpt->scp_hist_seq_culled = req->rq_history_seq;
1019 spin_unlock(&svcpt->scp_lock);
1021 ptlrpc_server_free_request(req);
1023 spin_unlock(&svcpt->scp_lock);
1027 /** Change request export and move hp request from old export to new */
1028 void ptlrpc_request_change_export(struct ptlrpc_request *req,
1029 struct obd_export *export)
1031 if (req->rq_export != NULL) {
1032 LASSERT(!list_empty(&req->rq_exp_list));
1033 /* remove rq_exp_list from last export */
1034 spin_lock(&req->rq_export->exp_rpc_lock);
1035 list_del_init(&req->rq_exp_list);
1036 spin_unlock(&req->rq_export->exp_rpc_lock);
1038 * export has one reference already, so it`s safe to
1039 * add req to export queue here and get another
1040 * reference for request later
1042 spin_lock(&export->exp_rpc_lock);
1043 if (req->rq_ops != NULL) /* hp request */
1044 list_add(&req->rq_exp_list, &export->exp_hp_rpcs);
1046 list_add(&req->rq_exp_list, &export->exp_reg_rpcs);
1047 spin_unlock(&export->exp_rpc_lock);
1049 class_export_rpc_dec(req->rq_export);
1050 class_export_put(req->rq_export);
1053 /* request takes one export refcount */
1054 req->rq_export = class_export_get(export);
1055 class_export_rpc_inc(export);
1061 * to finish a request: stop sending more early replies, and release
1064 static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt,
1065 struct ptlrpc_request *req)
1067 ptlrpc_server_hpreq_fini(req);
1069 ptlrpc_server_drop_request(req);
1073 * to finish an active request: stop sending more early replies, and release
1074 * the request. should be called after we finished handling the request.
1076 static void ptlrpc_server_finish_active_request(
1077 struct ptlrpc_service_part *svcpt,
1078 struct ptlrpc_request *req)
1080 spin_lock(&svcpt->scp_req_lock);
1081 ptlrpc_nrs_req_stop_nolock(req);
1082 svcpt->scp_nreqs_active--;
1084 svcpt->scp_nhreqs_active--;
1085 spin_unlock(&svcpt->scp_req_lock);
1087 ptlrpc_nrs_req_finalize(req);
1089 if (req->rq_export != NULL)
1090 class_export_rpc_dec(req->rq_export);
1092 ptlrpc_server_finish_request(svcpt, req);
1096 * This function makes sure dead exports are evicted in a timely manner.
1097 * This function is only called when some export receives a message (i.e.,
1098 * the network is up.)
1100 void ptlrpc_update_export_timer(struct obd_export *exp, time64_t extra_delay)
1102 struct obd_export *oldest_exp;
1103 time64_t oldest_time, new_time;
1110 * Compensate for slow machines, etc, by faking our request time
1111 * into the future. Although this can break the strict time-ordering
1112 * of the list, we can be really lazy here - we don't have to evict
1113 * at the exact right moment. Eventually, all silent exports
1114 * will make it to the top of the list.
1117 /* Do not pay attention on 1sec or smaller renewals. */
1118 new_time = ktime_get_real_seconds() + extra_delay;
1119 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
1122 exp->exp_last_request_time = new_time;
1125 * exports may get disconnected from the chain even though the
1126 * export has references, so we must keep the spin lock while
1127 * manipulating the lists
1129 spin_lock(&exp->exp_obd->obd_dev_lock);
1131 if (list_empty(&exp->exp_obd_chain_timed)) {
1132 /* this one is not timed */
1133 spin_unlock(&exp->exp_obd->obd_dev_lock);
1137 list_move_tail(&exp->exp_obd_chain_timed,
1138 &exp->exp_obd->obd_exports_timed);
1140 oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
1141 struct obd_export, exp_obd_chain_timed);
1142 oldest_time = oldest_exp->exp_last_request_time;
1143 spin_unlock(&exp->exp_obd->obd_dev_lock);
1145 if (exp->exp_obd->obd_recovering) {
1146 /* be nice to everyone during recovery */
1151 /* Note - racing to start/reset the obd_eviction timer is safe */
1152 if (exp->exp_obd->obd_eviction_timer == 0) {
1153 /* Check if the oldest entry is expired. */
1154 if (ktime_get_real_seconds() >
1155 oldest_time + PING_EVICT_TIMEOUT + extra_delay) {
1157 * We need a second timer, in case the net was down and
1158 * it just came back. Since the pinger may skip every
1159 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
1160 * we better wait for 3.
1162 exp->exp_obd->obd_eviction_timer =
1163 ktime_get_real_seconds() + 3 * PING_INTERVAL;
1164 CDEBUG(D_HA, "%s: Think about evicting %s from %lld\n",
1165 exp->exp_obd->obd_name,
1166 obd_export_nid2str(oldest_exp), oldest_time);
1169 if (ktime_get_real_seconds() >
1170 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
1172 * The evictor won't evict anyone who we've heard from
1173 * recently, so we don't have to check before we start
1176 if (!ping_evictor_wake(exp))
1177 exp->exp_obd->obd_eviction_timer = 0;
1185 * Sanity check request \a req.
1186 * Return 0 if all is ok, error code otherwise.
1188 static int ptlrpc_check_req(struct ptlrpc_request *req)
1190 struct obd_device *obd = req->rq_export->exp_obd;
1193 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
1194 req->rq_export->exp_conn_cnt)) {
1195 DEBUG_REQ(D_RPCTRACE, req,
1196 "DROPPING req from old connection %d < %d",
1197 lustre_msg_get_conn_cnt(req->rq_reqmsg),
1198 req->rq_export->exp_conn_cnt);
1201 if (unlikely(obd == NULL || obd->obd_fail)) {
1203 * Failing over, don't handle any more reqs,
1204 * send error response instead.
1206 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
1207 req, (obd != NULL) ? obd->obd_name : "unknown");
1209 } else if (lustre_msg_get_flags(req->rq_reqmsg) &
1210 (MSG_REPLAY | MSG_REQ_REPLAY_DONE) &&
1211 !obd->obd_recovering) {
1212 DEBUG_REQ(D_ERROR, req,
1213 "Invalid replay without recovery");
1214 class_fail_export(req->rq_export);
1216 } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 &&
1217 !obd->obd_recovering) {
1218 DEBUG_REQ(D_ERROR, req,
1219 "Invalid req with transno %llu without recovery",
1220 lustre_msg_get_transno(req->rq_reqmsg));
1221 class_fail_export(req->rq_export);
1225 if (unlikely(rc < 0)) {
1226 req->rq_status = rc;
1232 static void ptlrpc_at_set_timer(struct ptlrpc_service_part *svcpt)
1234 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1237 if (array->paa_count == 0) {
1238 del_timer(&svcpt->scp_at_timer);
1242 /* Set timer for closest deadline */
1243 next = array->paa_deadline - ktime_get_real_seconds() -
1246 ptlrpc_at_timer(cfs_timer_cb_arg(svcpt, scp_at_timer));
1248 mod_timer(&svcpt->scp_at_timer,
1249 jiffies + nsecs_to_jiffies(next * NSEC_PER_SEC));
1250 CDEBUG(D_INFO, "armed %s at %+llds\n",
1251 svcpt->scp_service->srv_name, next);
1255 /* Add rpc to early reply check list */
1256 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
1258 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1259 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1260 struct ptlrpc_request *rq = NULL;
1266 if (req->rq_no_reply)
1269 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
1272 spin_lock(&svcpt->scp_at_lock);
1273 LASSERT(list_empty(&req->rq_timed_list));
1275 div_u64_rem(req->rq_deadline, array->paa_size, &index);
1276 if (array->paa_reqs_count[index] > 0) {
1278 * latest rpcs will have the latest deadlines in the list,
1279 * so search backward.
1281 list_for_each_entry_reverse(rq, &array->paa_reqs_array[index],
1283 if (req->rq_deadline >= rq->rq_deadline) {
1284 list_add(&req->rq_timed_list,
1285 &rq->rq_timed_list);
1291 /* Add the request at the head of the list */
1292 if (list_empty(&req->rq_timed_list))
1293 list_add(&req->rq_timed_list, &array->paa_reqs_array[index]);
1295 spin_lock(&req->rq_lock);
1296 req->rq_at_linked = 1;
1297 spin_unlock(&req->rq_lock);
1298 req->rq_at_index = index;
1299 array->paa_reqs_count[index]++;
1301 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
1302 array->paa_deadline = req->rq_deadline;
1303 ptlrpc_at_set_timer(svcpt);
1305 spin_unlock(&svcpt->scp_at_lock);
1310 static void ptlrpc_at_remove_timed(struct ptlrpc_request *req)
1312 struct ptlrpc_at_array *array;
1314 array = &req->rq_rqbd->rqbd_svcpt->scp_at_array;
1316 /* NB: must call with hold svcpt::scp_at_lock */
1317 LASSERT(!list_empty(&req->rq_timed_list));
1318 list_del_init(&req->rq_timed_list);
1320 spin_lock(&req->rq_lock);
1321 req->rq_at_linked = 0;
1322 spin_unlock(&req->rq_lock);
1324 array->paa_reqs_count[req->rq_at_index]--;
1329 * Attempt to extend the request deadline by sending an early reply to the
1332 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
1334 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1335 struct ptlrpc_request *reqcopy;
1336 struct lustre_msg *reqmsg;
1337 time64_t olddl = req->rq_deadline - ktime_get_real_seconds();
1343 if (CFS_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_RECONNECT)) {
1344 /* don't send early reply */
1349 * deadline is when the client expects us to reply, margin is the
1350 * difference between clients' and servers' expectations
1352 DEBUG_REQ(D_ADAPTTO, req,
1353 "%ssending early reply (deadline %+llds, margin %+llds) for %d+%d",
1354 AT_OFF ? "AT off - not " : "",
1355 (s64)olddl, (s64)(olddl - at_get(&svcpt->scp_at_estimate)),
1356 at_get(&svcpt->scp_at_estimate), at_extra);
1362 DEBUG_REQ(D_WARNING, req,
1363 "Already past deadline (%+llds), not sending early reply. Consider increasing at_early_margin (%d)?",
1364 (s64)olddl, at_early_margin);
1366 /* Return an error so we're not re-added to the timed list. */
1370 if ((lustre_msghdr_get_flags(req->rq_reqmsg) &
1371 MSGHDR_AT_SUPPORT) == 0) {
1372 DEBUG_REQ(D_INFO, req,
1373 "Wanted to ask client for more time, but no AT support");
1377 if (req->rq_export &&
1378 lustre_msg_get_flags(req->rq_reqmsg) &
1379 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1380 struct obd_device *obd_exp = req->rq_export->exp_obd;
1383 * During recovery, we don't want to send too many early
1384 * replies, but on the other hand we want to make sure the
1385 * client has enough time to resend if the rpc is lost. So
1386 * during the recovery period send at least 4 early replies,
1387 * spacing them every at_extra if we can. at_estimate should
1388 * always equal this fixed value during recovery.
1392 * Don't account request processing time into AT history
1393 * during recovery, it is not service time we need but
1394 * includes also waiting time for recovering clients
1396 newdl = min_t(time64_t, at_extra,
1397 obd_exp->obd_recovery_timeout / 4) +
1398 ktime_get_real_seconds();
1401 * We want to extend the request deadline by at_extra seconds,
1402 * so we set our service estimate to reflect how much time has
1403 * passed since this request arrived plus an additional
1404 * at_extra seconds. The client will calculate the new deadline
1405 * based on this service estimate (plus some additional time to
1406 * account for network latency). See ptlrpc_at_recv_early_reply
1408 at_measured(&svcpt->scp_at_estimate, at_extra +
1409 ktime_get_real_seconds() -
1410 req->rq_arrival_time.tv_sec);
1411 newdl = req->rq_arrival_time.tv_sec +
1412 at_get(&svcpt->scp_at_estimate);
1416 * Check to see if we've actually increased the deadline -
1417 * we may be past adaptive_max
1419 if (req->rq_deadline >= newdl) {
1420 DEBUG_REQ(D_WARNING, req, "Couldn't add any time (%lld/%lld), not sending early reply\n",
1421 (s64)olddl, (s64)(newdl - ktime_get_real_seconds()));
1425 reqcopy = ptlrpc_request_cache_alloc(GFP_NOFS);
1426 if (reqcopy == NULL)
1428 OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1430 GOTO(out_free, rc = -ENOMEM);
1433 reqcopy->rq_reply_state = NULL;
1434 reqcopy->rq_rep_swab_mask = 0;
1435 reqcopy->rq_pack_bulk = 0;
1436 reqcopy->rq_pack_udesc = 0;
1437 reqcopy->rq_packed_final = 0;
1438 sptlrpc_svc_ctx_addref(reqcopy);
1439 /* We only need the reqmsg for the magic */
1440 reqcopy->rq_reqmsg = reqmsg;
1441 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1444 * tgt_brw_read() and tgt_brw_write() may have decided not to reply.
1445 * Without this check, we would fail the rq_no_reply assertion in
1446 * ptlrpc_send_reply().
1448 if (reqcopy->rq_no_reply)
1449 GOTO(out, rc = -ETIMEDOUT);
1451 LASSERT(atomic_read(&req->rq_refcount));
1452 /** if it is last refcount then early reply isn't needed */
1453 if (atomic_read(&req->rq_refcount) == 1) {
1454 DEBUG_REQ(D_ADAPTTO, reqcopy,
1455 "Normal reply already sent out, abort sending early reply\n");
1456 GOTO(out, rc = -EINVAL);
1459 /* Connection ref */
1460 reqcopy->rq_export = class_conn2export(
1461 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1462 if (reqcopy->rq_export == NULL)
1463 GOTO(out, rc = -ENODEV);
1466 class_export_rpc_inc(reqcopy->rq_export);
1467 if (reqcopy->rq_export->exp_obd &&
1468 reqcopy->rq_export->exp_obd->obd_fail)
1469 GOTO(out_put, rc = -ENODEV);
1471 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1475 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1478 /* Adjust our own deadline to what we told the client */
1479 req->rq_deadline = newdl;
1480 req->rq_early_count++; /* number sent, server side */
1482 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1486 * Free the (early) reply state from lustre_pack_reply.
1487 * (ptlrpc_send_reply takes it's own rs ref, so this is safe here)
1489 ptlrpc_req_drop_rs(reqcopy);
1492 class_export_rpc_dec(reqcopy->rq_export);
1493 class_export_put(reqcopy->rq_export);
1495 sptlrpc_svc_ctx_decref(reqcopy);
1496 OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1498 ptlrpc_request_cache_free(reqcopy);
1503 * Send early replies to everybody expiring within at_early_margin
1504 * asking for at_extra time
1506 static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1508 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1509 struct ptlrpc_request *rq, *n;
1510 struct list_head work_list;
1513 time64_t now = ktime_get_real_seconds();
1515 int first, counter = 0;
1518 spin_lock(&svcpt->scp_at_lock);
1519 if (svcpt->scp_at_check == 0) {
1520 spin_unlock(&svcpt->scp_at_lock);
1523 delay = ktime_ms_delta(ktime_get(), svcpt->scp_at_checktime);
1524 svcpt->scp_at_check = 0;
1526 if (array->paa_count == 0) {
1527 spin_unlock(&svcpt->scp_at_lock);
1531 /* The timer went off, but maybe the nearest rpc already completed. */
1532 first = array->paa_deadline - now;
1533 if (first > at_early_margin) {
1534 /* We've still got plenty of time. Reset the timer. */
1535 ptlrpc_at_set_timer(svcpt);
1536 spin_unlock(&svcpt->scp_at_lock);
1541 * We're close to a timeout, and we don't know how much longer the
1542 * server will take. Send early replies to everyone expiring soon.
1544 INIT_LIST_HEAD(&work_list);
1546 div_u64_rem(array->paa_deadline, array->paa_size, &index);
1547 count = array->paa_count;
1549 count -= array->paa_reqs_count[index];
1550 list_for_each_entry_safe(rq, n,
1551 &array->paa_reqs_array[index],
1553 if (rq->rq_deadline > now + at_early_margin) {
1554 /* update the earliest deadline */
1555 if (deadline == -1 ||
1556 rq->rq_deadline < deadline)
1557 deadline = rq->rq_deadline;
1562 * ptlrpc_server_drop_request() may drop
1563 * refcount to 0 already. Let's check this and
1564 * don't add entry to work_list
1566 if (likely(atomic_inc_not_zero(&rq->rq_refcount))) {
1567 ptlrpc_at_remove_timed(rq);
1568 list_add(&rq->rq_timed_list, &work_list);
1570 ptlrpc_at_remove_timed(rq);
1576 if (++index >= array->paa_size)
1579 array->paa_deadline = deadline;
1580 /* we have a new earliest deadline, restart the timer */
1581 ptlrpc_at_set_timer(svcpt);
1583 spin_unlock(&svcpt->scp_at_lock);
1586 "timeout in %+ds, asking for %d secs on %d early replies\n",
1587 first, at_extra, counter);
1590 * We're already past request deadlines before we even get a
1591 * chance to send early replies
1593 LCONSOLE_WARN("%s: This server is not able to keep up with request traffic (cpu-bound).\n",
1594 svcpt->scp_service->srv_name);
1595 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, delay=%lld\n",
1596 counter, svcpt->scp_nreqs_incoming,
1597 svcpt->scp_nreqs_active,
1598 at_get(&svcpt->scp_at_estimate), delay);
1602 * we took additional refcount so entries can't be deleted from list, no
1605 while (!list_empty(&work_list)) {
1606 rq = list_entry(work_list.next, struct ptlrpc_request,
1608 list_del_init(&rq->rq_timed_list);
1610 if (ptlrpc_at_send_early_reply(rq) == 0)
1611 ptlrpc_at_add_timed(rq);
1613 ptlrpc_server_drop_request(rq);
1616 RETURN(1); /* return "did_something" for liblustre */
1620 * Check if we are already handling earlier incarnation of this request.
1621 * Called under &req->rq_export->exp_rpc_lock locked
1623 static struct ptlrpc_request*
1624 ptlrpc_server_check_resend_in_progress(struct ptlrpc_request *req)
1626 struct ptlrpc_request *tmp = NULL;
1628 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ||
1629 (atomic_read(&req->rq_export->exp_rpc_count) == 0))
1633 * bulk request are aborted upon reconnect, don't try to
1636 if (req->rq_bulk_write || req->rq_bulk_read)
1640 * This list should not be longer than max_requests in
1641 * flights on the client, so it is not all that long.
1642 * Also we only hit this codepath in case of a resent
1643 * request which makes it even more rarely hit
1645 list_for_each_entry(tmp, &req->rq_export->exp_reg_rpcs,
1647 /* Found duplicate one */
1648 if (tmp->rq_xid == req->rq_xid)
1651 list_for_each_entry(tmp, &req->rq_export->exp_hp_rpcs,
1653 /* Found duplicate one */
1654 if (tmp->rq_xid == req->rq_xid)
1660 DEBUG_REQ(D_HA, req, "Found duplicate req in processing");
1661 DEBUG_REQ(D_HA, tmp, "Request being processed");
1666 * Check if a request should be assigned with a high priority.
1668 * \retval < 0: error occurred
1669 * 0: normal RPC request
1670 * +1: high priority request
1672 static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt,
1673 struct ptlrpc_request *req)
1678 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL) {
1679 rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
1686 if (req->rq_export != NULL && req->rq_ops != NULL) {
1688 * Perform request specific check. We should do this
1689 * check before the request is added into exp_hp_rpcs
1690 * list otherwise it may hit swab race at LU-1044.
1692 if (req->rq_ops->hpreq_check != NULL) {
1693 rc = req->rq_ops->hpreq_check(req);
1694 if (rc == -ESTALE) {
1695 req->rq_status = rc;
1699 * can only return error,
1700 * 0 for normal request,
1701 * or 1 for high priority request
1710 /** Remove the request from the export list. */
1711 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req)
1714 if (req->rq_export) {
1716 * refresh lock timeout again so that client has more
1717 * room to send lock cancel RPC.
1719 if (req->rq_ops && req->rq_ops->hpreq_fini)
1720 req->rq_ops->hpreq_fini(req);
1722 spin_lock(&req->rq_export->exp_rpc_lock);
1723 list_del_init(&req->rq_exp_list);
1724 spin_unlock(&req->rq_export->exp_rpc_lock);
1729 static int ptlrpc_hpreq_check(struct ptlrpc_request *req)
1734 static struct ptlrpc_hpreq_ops ptlrpc_hpreq_common = {
1735 .hpreq_check = ptlrpc_hpreq_check,
1738 /* Hi-Priority RPC check by RPC operation code. */
1739 int ptlrpc_hpreq_handler(struct ptlrpc_request *req)
1741 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1744 * Check for export to let only reconnects for not yet evicted
1745 * export to become a HP rpc.
1747 if ((req->rq_export != NULL) &&
1748 (opc == OBD_PING || opc == MDS_CONNECT || opc == OST_CONNECT))
1749 req->rq_ops = &ptlrpc_hpreq_common;
1753 EXPORT_SYMBOL(ptlrpc_hpreq_handler);
1755 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1756 struct ptlrpc_request *req)
1760 struct ptlrpc_request *orig;
1764 rc = ptlrpc_server_hpreq_init(svcpt, req);
1769 ptlrpc_nrs_req_initialize(svcpt, req, hp);
1771 if (req->rq_export != NULL) {
1772 struct obd_export *exp = req->rq_export;
1775 * do search for duplicated xid and the adding to the list
1778 spin_lock_bh(&exp->exp_rpc_lock);
1779 orig = ptlrpc_server_check_resend_in_progress(req);
1780 if (orig && likely(atomic_inc_not_zero(&orig->rq_refcount))) {
1783 spin_unlock_bh(&exp->exp_rpc_lock);
1786 * When the client resend request and the server has
1787 * the previous copy of it, we need to update deadlines,
1788 * to be sure that the client and the server have equal
1789 * request deadlines.
1792 spin_lock(&orig->rq_rqbd->rqbd_svcpt->scp_at_lock);
1793 linked = orig->rq_at_linked;
1795 ptlrpc_at_remove_timed(orig);
1796 spin_unlock(&orig->rq_rqbd->rqbd_svcpt->scp_at_lock);
1797 orig->rq_deadline = req->rq_deadline;
1799 ptlrpc_at_add_timed(orig);
1800 ptlrpc_server_drop_request(orig);
1801 ptlrpc_nrs_req_finalize(req);
1805 if (hp || req->rq_ops != NULL)
1806 list_add(&req->rq_exp_list, &exp->exp_hp_rpcs);
1808 list_add(&req->rq_exp_list, &exp->exp_reg_rpcs);
1809 spin_unlock_bh(&exp->exp_rpc_lock);
1813 * the current thread is not the processing thread for this request
1814 * since that, but request is in exp_hp_list and can be find there.
1815 * Remove all relations between request and old thread.
1817 req->rq_svc_thread->t_env->le_ses = NULL;
1818 req->rq_svc_thread = NULL;
1819 req->rq_session.lc_thread = NULL;
1821 ptlrpc_nrs_req_add(svcpt, req, hp);
1827 * Allow to handle high priority request
1828 * User can call it w/o any lock but need to hold
1829 * ptlrpc_service_part::scp_req_lock to get reliable result
1831 static bool ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1834 int running = svcpt->scp_nthrs_running;
1836 if (!nrs_svcpt_has_hp(svcpt))
1842 if (ptlrpc_nrs_req_throttling_nolock(svcpt, true))
1845 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1846 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1847 /* leave just 1 thread for normal RPCs */
1848 running = PTLRPC_NTHRS_INIT;
1849 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1853 if (svcpt->scp_nreqs_active >= running - 1)
1856 if (svcpt->scp_nhreqs_active == 0)
1859 return !ptlrpc_nrs_req_pending_nolock(svcpt, false) ||
1860 svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1863 static bool ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1866 return ptlrpc_server_allow_high(svcpt, force) &&
1867 ptlrpc_nrs_req_pending_nolock(svcpt, true);
1871 * Only allow normal priority requests on a service that has a high-priority
1872 * queue if forced (i.e. cleanup), if there are other high priority requests
1873 * already being processed (i.e. those threads can service more high-priority
1874 * requests), or if there are enough idle threads that a later thread can do
1875 * a high priority request.
1876 * User can call it w/o any lock but need to hold
1877 * ptlrpc_service_part::scp_req_lock to get reliable result
1879 static bool ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1882 int running = svcpt->scp_nthrs_running;
1884 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1885 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1886 /* leave just 1 thread for normal RPCs */
1887 running = PTLRPC_NTHRS_INIT;
1888 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1895 if (ptlrpc_nrs_req_throttling_nolock(svcpt, false))
1898 if (svcpt->scp_nreqs_active < running - 2)
1901 if (svcpt->scp_nreqs_active >= running - 1)
1904 return svcpt->scp_nhreqs_active > 0 || !nrs_svcpt_has_hp(svcpt);
1907 static bool ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1910 return ptlrpc_server_allow_normal(svcpt, force) &&
1911 ptlrpc_nrs_req_pending_nolock(svcpt, false);
1915 * Returns true if there are requests available in incoming
1916 * request queue for processing and it is allowed to fetch them.
1917 * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1918 * to get reliable result
1919 * \see ptlrpc_server_allow_normal
1920 * \see ptlrpc_server_allow high
1923 bool ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt,
1926 return ptlrpc_server_high_pending(svcpt, force) ||
1927 ptlrpc_server_normal_pending(svcpt, force);
1931 * Fetch a request for processing from queue of unprocessed requests.
1932 * Favors high-priority requests.
1933 * Returns a pointer to fetched request.
1935 static struct ptlrpc_request *
1936 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force)
1938 struct ptlrpc_request *req = NULL;
1942 spin_lock(&svcpt->scp_req_lock);
1944 if (ptlrpc_server_high_pending(svcpt, force)) {
1945 req = ptlrpc_nrs_req_get_nolock(svcpt, true, force);
1947 svcpt->scp_hreq_count++;
1952 if (ptlrpc_server_normal_pending(svcpt, force)) {
1953 req = ptlrpc_nrs_req_get_nolock(svcpt, false, force);
1955 svcpt->scp_hreq_count = 0;
1960 spin_unlock(&svcpt->scp_req_lock);
1964 svcpt->scp_nreqs_active++;
1966 svcpt->scp_nhreqs_active++;
1968 spin_unlock(&svcpt->scp_req_lock);
1970 if (likely(req->rq_export))
1971 class_export_rpc_inc(req->rq_export);
1977 * Handle freshly incoming reqs, add to timed early reply list,
1978 * pass on to regular request queue.
1979 * All incoming requests pass through here before getting into
1980 * ptlrpc_server_handle_req later on.
1982 static int ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt,
1983 struct ptlrpc_thread *thread)
1985 struct ptlrpc_service *svc = svcpt->scp_service;
1986 struct ptlrpc_request *req;
1992 spin_lock(&svcpt->scp_lock);
1993 if (list_empty(&svcpt->scp_req_incoming)) {
1994 spin_unlock(&svcpt->scp_lock);
1998 req = list_entry(svcpt->scp_req_incoming.next,
1999 struct ptlrpc_request, rq_list);
2000 list_del_init(&req->rq_list);
2001 svcpt->scp_nreqs_incoming--;
2003 * Consider this still a "queued" request as far as stats are
2006 spin_unlock(&svcpt->scp_lock);
2008 /* go through security check/transform */
2009 rc = sptlrpc_svc_unwrap_request(req);
2013 case SECSVC_COMPLETE:
2014 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
2023 * for null-flavored rpc, msg has been unpacked by sptlrpc, although
2024 * redo it wouldn't be harmful.
2026 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
2027 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
2029 CERROR("error unpacking request: ptl %d from %s x%llu\n",
2030 svc->srv_req_portal, libcfs_id2str(req->rq_peer),
2036 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
2038 CERROR("error unpacking ptlrpc body: ptl %d from %s x %llu\n",
2039 svc->srv_req_portal, libcfs_id2str(req->rq_peer),
2044 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
2045 lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
2046 CERROR("drop incoming rpc opc %u, x%llu\n",
2047 cfs_fail_val, req->rq_xid);
2052 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
2053 CERROR("wrong packet type received (type=%u) from %s\n",
2054 lustre_msg_get_type(req->rq_reqmsg),
2055 libcfs_id2str(req->rq_peer));
2059 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2063 req->rq_bulk_write = 1;
2067 case MGS_CONFIG_READ:
2068 req->rq_bulk_read = 1;
2072 CDEBUG(D_RPCTRACE, "got req x%llu\n", req->rq_xid);
2074 req->rq_export = class_conn2export(
2075 lustre_msg_get_handle(req->rq_reqmsg));
2076 if (req->rq_export) {
2077 rc = ptlrpc_check_req(req);
2079 rc = sptlrpc_target_export_check(req->rq_export, req);
2081 DEBUG_REQ(D_ERROR, req,
2082 "DROPPING req with illegal security flavor,");
2087 ptlrpc_update_export_timer(req->rq_export, 0);
2090 /* req_in handling should/must be fast */
2091 if (ktime_get_real_seconds() - req->rq_arrival_time.tv_sec > 5)
2092 DEBUG_REQ(D_WARNING, req, "Slow req_in handling %llds",
2093 (s64)(ktime_get_real_seconds() -
2094 req->rq_arrival_time.tv_sec));
2096 /* Set rpc server deadline and add it to the timed list */
2097 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
2098 MSGHDR_AT_SUPPORT) ?
2099 /* The max time the client expects us to take */
2100 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
2102 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
2103 if (unlikely(deadline == 0)) {
2104 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
2108 /* Skip early reply */
2109 if (OBD_FAIL_PRECHECK(OBD_FAIL_MDS_RESEND))
2110 req->rq_deadline += obd_timeout;
2112 req->rq_svc_thread = thread;
2113 if (thread != NULL) {
2115 * initialize request session, it is needed for request
2116 * processing by target
2118 rc = lu_context_init(&req->rq_session, LCT_SERVER_SESSION |
2121 CERROR("%s: failure to initialize session: rc = %d\n",
2122 thread->t_name, rc);
2125 req->rq_session.lc_thread = thread;
2126 lu_context_enter(&req->rq_session);
2127 thread->t_env->le_ses = &req->rq_session;
2130 ptlrpc_at_add_timed(req);
2132 /* Move it over to the request processing queue */
2133 rc = ptlrpc_server_request_add(svcpt, req);
2137 wake_up(&svcpt->scp_waitq);
2141 ptlrpc_server_finish_request(svcpt, req);
2147 * Main incoming request handling logic.
2148 * Calls handler function from service to do actual processing.
2150 static int ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
2151 struct ptlrpc_thread *thread)
2153 struct ptlrpc_service *svc = svcpt->scp_service;
2154 struct ptlrpc_request *request;
2164 request = ptlrpc_server_request_get(svcpt, false);
2165 if (request == NULL)
2168 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
2169 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
2170 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
2171 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
2173 if (unlikely(fail_opc)) {
2174 if (request->rq_export && request->rq_ops)
2175 OBD_FAIL_TIMEOUT(fail_opc, 4);
2178 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
2180 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
2181 libcfs_debug_dumplog();
2183 work_start = ktime_get_real();
2184 arrived = timespec64_to_ktime(request->rq_arrival_time);
2185 timediff_usecs = ktime_us_delta(work_start, arrived);
2186 if (likely(svc->srv_stats != NULL)) {
2187 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
2189 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
2190 svcpt->scp_nreqs_incoming);
2191 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
2192 svcpt->scp_nreqs_active);
2193 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
2194 at_get(&svcpt->scp_at_estimate));
2197 if (likely(request->rq_export)) {
2198 if (unlikely(ptlrpc_check_req(request)))
2200 ptlrpc_update_export_timer(request->rq_export,
2201 div_u64(timediff_usecs,
2206 * Discard requests queued for longer than the deadline.
2207 * The deadline is increased if we send an early reply.
2209 if (ktime_get_real_seconds() > request->rq_deadline) {
2210 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s: deadline %lld:%llds ago\n",
2211 libcfs_id2str(request->rq_peer),
2212 request->rq_deadline -
2213 request->rq_arrival_time.tv_sec,
2214 ktime_get_real_seconds() - request->rq_deadline);
2219 "Handling RPC req@%p pname:cluuid+ref:pid:xid:nid:opc:job %s:%s+%d:%d:x%llu:%s:%d:%s\n",
2220 request, current_comm(),
2221 (request->rq_export ?
2222 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2223 (request->rq_export ?
2224 atomic_read(&request->rq_export->exp_refcount) : -99),
2225 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
2226 libcfs_id2str(request->rq_peer),
2227 lustre_msg_get_opc(request->rq_reqmsg),
2228 lustre_msg_get_jobid(request->rq_reqmsg) ?: "");
2230 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
2231 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
2233 CDEBUG(D_NET, "got req %llu\n", request->rq_xid);
2235 /* re-assign request and sesson thread to the current one */
2236 request->rq_svc_thread = thread;
2237 if (thread != NULL) {
2238 LASSERT(request->rq_session.lc_thread == NULL);
2239 request->rq_session.lc_thread = thread;
2240 thread->t_env->le_ses = &request->rq_session;
2242 svc->srv_ops.so_req_handler(request);
2244 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
2247 if (unlikely(ktime_get_real_seconds() > request->rq_deadline)) {
2248 DEBUG_REQ(D_WARNING, request,
2249 "Request took longer than estimated (%lld:%llds); "
2250 "client may timeout.",
2251 request->rq_deadline -
2252 request->rq_arrival_time.tv_sec,
2253 ktime_get_real_seconds() - request->rq_deadline);
2256 work_end = ktime_get_real();
2257 timediff_usecs = ktime_us_delta(work_end, work_start);
2258 arrived_usecs = ktime_us_delta(work_end, arrived);
2260 "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",
2261 request, current_comm(),
2262 (request->rq_export ?
2263 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2264 (request->rq_export ?
2265 atomic_read(&request->rq_export->exp_refcount) : -99),
2266 lustre_msg_get_status(request->rq_reqmsg),
2268 libcfs_id2str(request->rq_peer),
2269 lustre_msg_get_opc(request->rq_reqmsg),
2270 lustre_msg_get_jobid(request->rq_reqmsg) ?: "",
2273 (request->rq_repmsg ?
2274 lustre_msg_get_transno(request->rq_repmsg) :
2275 request->rq_transno),
2277 (request->rq_repmsg ?
2278 lustre_msg_get_status(request->rq_repmsg) : -999));
2279 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
2280 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
2281 int opc = opcode_offset(op);
2283 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
2284 LASSERT(opc < LUSTRE_MAX_OPCODES);
2285 lprocfs_counter_add(svc->srv_stats,
2286 opc + EXTRA_MAX_OPCODES,
2290 if (unlikely(request->rq_early_count)) {
2291 DEBUG_REQ(D_ADAPTTO, request,
2292 "sent %d early replies before finishing in %llds",
2293 request->rq_early_count,
2294 div_u64(arrived_usecs, USEC_PER_SEC));
2297 ptlrpc_server_finish_active_request(svcpt, request);
2303 * An internal function to process a single reply state object.
2305 static int ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
2307 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
2308 struct ptlrpc_service *svc = svcpt->scp_service;
2309 struct obd_export *exp;
2315 exp = rs->rs_export;
2317 LASSERT(rs->rs_difficult);
2318 LASSERT(rs->rs_scheduled);
2319 LASSERT(list_empty(&rs->rs_list));
2322 * The disk commit callback holds exp_uncommitted_replies_lock while it
2323 * iterates over newly committed replies, removing them from
2324 * exp_uncommitted_replies. It then drops this lock and schedules the
2325 * replies it found for handling here.
2327 * We can avoid contention for exp_uncommitted_replies_lock between the
2328 * HRT threads and further commit callbacks by checking rs_committed
2329 * which is set in the commit callback while it holds both
2330 * rs_lock and exp_uncommitted_reples.
2332 * If we see rs_committed clear, the commit callback _may_ not have
2333 * handled this reply yet and we race with it to grab
2334 * exp_uncommitted_replies_lock before removing the reply from
2335 * exp_uncommitted_replies. Note that if we lose the race and the
2336 * reply has already been removed, list_del_init() is a noop.
2338 * If we see rs_committed set, we know the commit callback is handling,
2339 * or has handled this reply since store reordering might allow us to
2340 * see rs_committed set out of sequence. But since this is done
2341 * holding rs_lock, we can be sure it has all completed once we hold
2342 * rs_lock, which we do right next.
2344 if (!rs->rs_committed) {
2346 * if rs was commited, no need to convert locks, don't check
2347 * rs_committed here because rs may never be added into
2348 * exp_uncommitted_replies and this flag never be set, see
2349 * target_send_reply()
2351 if (rs->rs_convert_lock &&
2352 rs->rs_transno > exp->exp_last_committed) {
2353 struct ldlm_lock *lock;
2354 struct ldlm_lock *ack_locks[RS_MAX_LOCKS] = { NULL };
2356 spin_lock(&rs->rs_lock);
2357 if (rs->rs_convert_lock &&
2358 rs->rs_transno > exp->exp_last_committed) {
2359 nlocks = rs->rs_nlocks;
2360 while (nlocks-- > 0) {
2362 * NB don't assume rs is always handled
2363 * by the same service thread (see
2364 * ptlrpc_hr_select, so REP-ACK hr may
2365 * race with trans commit, while the
2366 * latter will release locks, get locks
2367 * here early to convert to COS mode
2370 lock = ldlm_handle2lock(
2371 &rs->rs_locks[nlocks]);
2373 ack_locks[nlocks] = lock;
2374 rs->rs_modes[nlocks] = LCK_COS;
2376 nlocks = rs->rs_nlocks;
2377 rs->rs_convert_lock = 0;
2379 * clear rs_scheduled so that commit callback
2380 * can schedule again
2382 rs->rs_scheduled = 0;
2383 spin_unlock(&rs->rs_lock);
2385 while (nlocks-- > 0) {
2386 lock = ack_locks[nlocks];
2387 ldlm_lock_mode_downgrade(lock, LCK_COS);
2388 LDLM_LOCK_PUT(lock);
2392 spin_unlock(&rs->rs_lock);
2395 spin_lock(&exp->exp_uncommitted_replies_lock);
2396 list_del_init(&rs->rs_obd_list);
2397 spin_unlock(&exp->exp_uncommitted_replies_lock);
2400 spin_lock(&exp->exp_lock);
2401 /* Noop if removed already */
2402 list_del_init(&rs->rs_exp_list);
2403 spin_unlock(&exp->exp_lock);
2405 spin_lock(&rs->rs_lock);
2407 been_handled = rs->rs_handled;
2410 nlocks = rs->rs_nlocks; /* atomic "steal", but */
2411 rs->rs_nlocks = 0; /* locks still on rs_locks! */
2413 if (nlocks == 0 && !been_handled) {
2415 * If we see this, we should already have seen the warning
2416 * in mds_steal_ack_locks()
2419 "All locks stolen from rs %p x%lld.t%lld o%d NID %s\n",
2420 rs, rs->rs_xid, rs->rs_transno, rs->rs_opc,
2421 libcfs_nid2str(exp->exp_connection->c_peer.nid));
2424 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
2425 spin_unlock(&rs->rs_lock);
2427 if (!been_handled && rs->rs_on_net) {
2428 LNetMDUnlink(rs->rs_md_h);
2429 /* Ignore return code; we're racing with completion */
2432 while (nlocks-- > 0)
2433 ldlm_lock_decref(&rs->rs_locks[nlocks],
2434 rs->rs_modes[nlocks]);
2436 spin_lock(&rs->rs_lock);
2439 rs->rs_scheduled = 0;
2440 rs->rs_convert_lock = 0;
2442 if (!rs->rs_on_net) {
2444 spin_unlock(&rs->rs_lock);
2446 class_export_put(exp);
2447 rs->rs_export = NULL;
2448 ptlrpc_rs_decref(rs);
2449 if (atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
2450 svc->srv_is_stopping)
2451 wake_up_all(&svcpt->scp_waitq);
2455 /* still on the net; callback will schedule */
2456 spin_unlock(&rs->rs_lock);
2461 static void ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
2463 int avail = svcpt->scp_nrqbds_posted;
2464 int low_water = test_req_buffer_pressure ? 0 :
2465 svcpt->scp_service->srv_nbuf_per_group / 2;
2467 /* NB I'm not locking; just looking. */
2470 * CAVEAT EMPTOR: We might be allocating buffers here because we've
2471 * allowed the request history to grow out of control. We could put a
2472 * sanity check on that here and cull some history if we need the
2476 if (avail <= low_water)
2477 ptlrpc_grow_req_bufs(svcpt, 1);
2479 if (svcpt->scp_service->srv_stats) {
2480 lprocfs_counter_add(svcpt->scp_service->srv_stats,
2481 PTLRPC_REQBUF_AVAIL_CNTR, avail);
2485 static int ptlrpc_retry_rqbds(void *arg)
2487 struct ptlrpc_service_part *svcpt = (struct ptlrpc_service_part *)arg;
2489 svcpt->scp_rqbd_timeout = 0;
2493 static inline int ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
2495 return svcpt->scp_nreqs_active <
2496 svcpt->scp_nthrs_running - 1 -
2497 (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL);
2501 * allowed to create more threads
2502 * user can call it w/o any lock but need to hold
2503 * ptlrpc_service_part::scp_lock to get reliable result
2505 static inline int ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
2507 return svcpt->scp_nthrs_running +
2508 svcpt->scp_nthrs_starting <
2509 svcpt->scp_service->srv_nthrs_cpt_limit;
2513 * too many requests and allowed to create more threads
2515 static inline int ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
2517 return !ptlrpc_threads_enough(svcpt) &&
2518 ptlrpc_threads_increasable(svcpt);
2521 static inline int ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2523 return thread_is_stopping(thread) ||
2524 thread->t_svcpt->scp_service->srv_is_stopping;
2527 /* stop the highest numbered thread if there are too many threads running */
2528 static inline bool ptlrpc_thread_should_stop(struct ptlrpc_thread *thread)
2530 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2532 return thread->t_id >= svcpt->scp_service->srv_nthrs_cpt_limit &&
2533 thread->t_id == svcpt->scp_thr_nextid - 1;
2536 static void ptlrpc_stop_thread(struct ptlrpc_thread *thread)
2538 CDEBUG(D_INFO, "Stopping thread %s #%u\n",
2539 thread->t_svcpt->scp_service->srv_thread_name, thread->t_id);
2540 thread_add_flags(thread, SVC_STOPPING);
2543 static inline void ptlrpc_thread_stop(struct ptlrpc_thread *thread)
2545 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2547 spin_lock(&svcpt->scp_lock);
2548 if (ptlrpc_thread_should_stop(thread)) {
2549 ptlrpc_stop_thread(thread);
2550 svcpt->scp_thr_nextid--;
2552 spin_unlock(&svcpt->scp_lock);
2555 static inline int ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
2557 return !list_empty(&svcpt->scp_rqbd_idle) &&
2558 svcpt->scp_rqbd_timeout == 0;
2562 ptlrpc_at_check(struct ptlrpc_service_part *svcpt)
2564 return svcpt->scp_at_check;
2568 * If a thread runs too long or spends to much time on a single request,
2569 * we want to know about it, so we set up a delayed work item as a watchdog.
2570 * If it fires, we display a stack trace of the delayed thread,
2571 * providing we aren't rate-limited
2573 * Watchdog stack traces are limited to 3 per 'libcfs_watchdog_ratelimit'
2576 static struct ratelimit_state watchdog_limit;
2578 static void ptlrpc_watchdog_fire(struct work_struct *w)
2580 struct ptlrpc_thread *thread = container_of(w, struct ptlrpc_thread,
2582 u64 ms_lapse = ktime_ms_delta(ktime_get(), thread->t_touched);
2583 u32 ms_frac = do_div(ms_lapse, MSEC_PER_SEC);
2585 if (!__ratelimit(&watchdog_limit)) {
2586 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",
2587 thread->t_task->comm, thread->t_task->pid,
2590 libcfs_debug_dumpstack(thread->t_task);
2592 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",
2593 thread->t_task->comm, thread->t_task->pid,
2594 ms_lapse, ms_frac, libcfs_watchdog_ratelimit);
2598 static void ptlrpc_watchdog_init(struct delayed_work *work, time_t time)
2600 INIT_DELAYED_WORK(work, ptlrpc_watchdog_fire);
2601 schedule_delayed_work(work, cfs_time_seconds(time));
2604 static void ptlrpc_watchdog_disable(struct delayed_work *work)
2606 cancel_delayed_work_sync(work);
2609 static void ptlrpc_watchdog_touch(struct delayed_work *work, time_t time)
2611 struct ptlrpc_thread *thread = container_of(&work->work,
2612 struct ptlrpc_thread,
2614 thread->t_touched = ktime_get();
2615 mod_delayed_work(system_wq, work, cfs_time_seconds(time));
2619 * requests wait on preprocessing
2620 * user can call it w/o any lock but need to hold
2621 * ptlrpc_service_part::scp_lock to get reliable result
2624 ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
2626 return !list_empty(&svcpt->scp_req_incoming);
2629 static __attribute__((__noinline__)) int
2630 ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
2631 struct ptlrpc_thread *thread)
2633 /* Don't exit while there are replies to be handled */
2634 struct l_wait_info lwi = LWI_TIMEOUT(svcpt->scp_rqbd_timeout,
2635 ptlrpc_retry_rqbds, svcpt);
2637 ptlrpc_watchdog_disable(&thread->t_watchdog);
2641 l_wait_event_exclusive_head(svcpt->scp_waitq,
2642 ptlrpc_thread_stopping(thread) ||
2643 ptlrpc_server_request_incoming(svcpt) ||
2644 ptlrpc_server_request_pending(svcpt, false) ||
2645 ptlrpc_rqbd_pending(svcpt) ||
2646 ptlrpc_at_check(svcpt), &lwi);
2648 if (ptlrpc_thread_stopping(thread))
2651 ptlrpc_watchdog_touch(&thread->t_watchdog,
2652 ptlrpc_server_get_timeout(svcpt));
2657 * Main thread body for service threads.
2658 * Waits in a loop waiting for new requests to process to appear.
2659 * Every time an incoming requests is added to its queue, a waitq
2660 * is woken up and one of the threads will handle it.
2662 static int ptlrpc_main(void *arg)
2664 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
2665 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2666 struct ptlrpc_service *svc = svcpt->scp_service;
2667 struct ptlrpc_reply_state *rs;
2668 struct group_info *ginfo = NULL;
2670 int counter = 0, rc = 0;
2674 thread->t_task = current;
2675 thread->t_pid = current_pid();
2676 unshare_fs_struct();
2678 if (svc->srv_cpt_bind) {
2679 rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
2681 CWARN("%s: failed to bind %s on CPT %d\n",
2682 svc->srv_name, thread->t_name, svcpt->scp_cpt);
2686 ginfo = groups_alloc(0);
2688 GOTO(out, rc = -ENOMEM);
2690 set_current_groups(ginfo);
2691 put_group_info(ginfo);
2693 if (svc->srv_ops.so_thr_init != NULL) {
2694 rc = svc->srv_ops.so_thr_init(thread);
2701 GOTO(out_srv_fini, rc = -ENOMEM);
2702 rc = lu_env_add(env);
2706 rc = lu_context_init(&env->le_ctx,
2707 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2709 GOTO(out_env_remove, rc);
2711 thread->t_env = env;
2712 env->le_ctx.lc_thread = thread;
2713 env->le_ctx.lc_cookie = 0x6;
2715 while (!list_empty(&svcpt->scp_rqbd_idle)) {
2716 rc = ptlrpc_server_post_idle_rqbds(svcpt);
2720 CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
2721 svc->srv_name, svcpt->scp_cpt, rc);
2722 GOTO(out_ctx_fini, rc);
2725 /* Alloc reply state structure for this one */
2726 OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2728 GOTO(out_ctx_fini, rc = -ENOMEM);
2730 spin_lock(&svcpt->scp_lock);
2732 LASSERT(thread_is_starting(thread));
2733 thread_clear_flags(thread, SVC_STARTING);
2735 LASSERT(svcpt->scp_nthrs_starting == 1);
2736 svcpt->scp_nthrs_starting--;
2739 * SVC_STOPPING may already be set here if someone else is trying
2740 * to stop the service while this new thread has been dynamically
2741 * forked. We still set SVC_RUNNING to let our creator know that
2742 * we are now running, however we will exit as soon as possible
2744 thread_add_flags(thread, SVC_RUNNING);
2745 svcpt->scp_nthrs_running++;
2746 spin_unlock(&svcpt->scp_lock);
2748 /* wake up our creator in case he's still waiting. */
2749 wake_up(&thread->t_ctl_waitq);
2751 thread->t_touched = ktime_get();
2752 ptlrpc_watchdog_init(&thread->t_watchdog,
2753 ptlrpc_server_get_timeout(svcpt));
2755 spin_lock(&svcpt->scp_rep_lock);
2756 list_add(&rs->rs_list, &svcpt->scp_rep_idle);
2757 wake_up(&svcpt->scp_rep_waitq);
2758 spin_unlock(&svcpt->scp_rep_lock);
2760 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2761 svcpt->scp_nthrs_running);
2763 /* XXX maintain a list of all managed devices: insert here */
2764 while (!ptlrpc_thread_stopping(thread)) {
2765 if (ptlrpc_wait_event(svcpt, thread))
2768 ptlrpc_check_rqbd_pool(svcpt);
2770 if (ptlrpc_threads_need_create(svcpt)) {
2771 /* Ignore return code - we tried... */
2772 ptlrpc_start_thread(svcpt, 0);
2775 /* reset le_ses to initial state */
2777 /* Process all incoming reqs before handling any */
2778 if (ptlrpc_server_request_incoming(svcpt)) {
2779 lu_context_enter(&env->le_ctx);
2780 ptlrpc_server_handle_req_in(svcpt, thread);
2781 lu_context_exit(&env->le_ctx);
2783 /* but limit ourselves in case of flood */
2784 if (counter++ < 100)
2789 if (ptlrpc_at_check(svcpt))
2790 ptlrpc_at_check_timed(svcpt);
2792 if (ptlrpc_server_request_pending(svcpt, false)) {
2793 lu_context_enter(&env->le_ctx);
2794 ptlrpc_server_handle_request(svcpt, thread);
2795 lu_context_exit(&env->le_ctx);
2798 if (ptlrpc_rqbd_pending(svcpt) &&
2799 ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
2801 * I just failed to repost request buffers.
2802 * Wait for a timeout (unless something else
2803 * happens) before I try again
2805 svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
2806 CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
2807 svcpt->scp_nrqbds_posted);
2810 * If the number of threads has been tuned downward and this
2811 * thread should be stopped, then stop in reverse order so the
2812 * the threads always have contiguous thread index values.
2814 if (unlikely(ptlrpc_thread_should_stop(thread)))
2815 ptlrpc_thread_stop(thread);
2818 ptlrpc_watchdog_disable(&thread->t_watchdog);
2821 lu_context_fini(&env->le_ctx);
2827 /* deconstruct service thread state created by ptlrpc_start_thread() */
2828 if (svc->srv_ops.so_thr_done != NULL)
2829 svc->srv_ops.so_thr_done(thread);
2831 CDEBUG(D_RPCTRACE, "%s: service thread [%p:%u] %d exiting: rc = %d\n",
2832 thread->t_name, thread, thread->t_pid, thread->t_id, rc);
2833 spin_lock(&svcpt->scp_lock);
2834 if (thread_test_and_clear_flags(thread, SVC_STARTING))
2835 svcpt->scp_nthrs_starting--;
2837 if (thread_test_and_clear_flags(thread, SVC_RUNNING)) {
2838 /* must know immediately */
2839 svcpt->scp_nthrs_running--;
2843 thread_add_flags(thread, SVC_STOPPED);
2845 wake_up(&thread->t_ctl_waitq);
2846 spin_unlock(&svcpt->scp_lock);
2851 static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
2852 struct list_head *replies)
2856 spin_lock(&hrt->hrt_lock);
2858 list_splice_init(&hrt->hrt_queue, replies);
2859 result = ptlrpc_hr.hr_stopping || !list_empty(replies);
2861 spin_unlock(&hrt->hrt_lock);
2866 * Main body of "handle reply" function.
2867 * It processes acked reply states
2869 static int ptlrpc_hr_main(void *arg)
2871 struct ptlrpc_hr_thread *hrt = (struct ptlrpc_hr_thread *)arg;
2872 struct ptlrpc_hr_partition *hrp = hrt->hrt_partition;
2873 struct list_head replies;
2881 INIT_LIST_HEAD(&replies);
2882 unshare_fs_struct();
2884 rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
2886 char threadname[20];
2888 snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
2889 hrp->hrp_cpt, hrt->hrt_id);
2890 CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
2891 threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
2894 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD |
2895 LCT_REMEMBER | LCT_NOREF);
2899 rc = lu_env_add(env);
2901 GOTO(out_ctx_fini, rc);
2903 atomic_inc(&hrp->hrp_nstarted);
2904 wake_up(&ptlrpc_hr.hr_waitq);
2906 while (!ptlrpc_hr.hr_stopping) {
2907 l_wait_condition(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
2909 while (!list_empty(&replies)) {
2910 struct ptlrpc_reply_state *rs;
2912 rs = list_entry(replies.prev,
2913 struct ptlrpc_reply_state,
2915 list_del_init(&rs->rs_list);
2916 /* refill keys if needed */
2918 lu_context_enter(&env->le_ctx);
2919 ptlrpc_handle_rs(rs);
2920 lu_context_exit(&env->le_ctx);
2924 atomic_inc(&hrp->hrp_nstopped);
2925 wake_up(&ptlrpc_hr.hr_waitq);
2929 lu_context_fini(&env->le_ctx);
2935 static void ptlrpc_stop_hr_threads(void)
2937 struct ptlrpc_hr_partition *hrp;
2941 ptlrpc_hr.hr_stopping = 1;
2943 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2944 if (hrp->hrp_thrs == NULL)
2945 continue; /* uninitialized */
2946 for (j = 0; j < hrp->hrp_nthrs; j++)
2947 wake_up_all(&hrp->hrp_thrs[j].hrt_waitq);
2950 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2951 if (hrp->hrp_thrs == NULL)
2952 continue; /* uninitialized */
2953 wait_event(ptlrpc_hr.hr_waitq,
2954 atomic_read(&hrp->hrp_nstopped) ==
2955 atomic_read(&hrp->hrp_nstarted));
2959 static int ptlrpc_start_hr_threads(void)
2961 struct ptlrpc_hr_partition *hrp;
2967 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2970 for (j = 0; j < hrp->hrp_nthrs; j++) {
2971 struct ptlrpc_hr_thread *hrt = &hrp->hrp_thrs[j];
2972 struct task_struct *task;
2974 task = kthread_run(ptlrpc_hr_main,
2976 "ptlrpc_hr%02d_%03d",
2985 wait_event(ptlrpc_hr.hr_waitq,
2986 atomic_read(&hrp->hrp_nstarted) == j);
2989 CERROR("cannot start reply handler thread %d:%d: rc = %d\n",
2991 ptlrpc_stop_hr_threads();
2999 static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
3001 struct l_wait_info lwi = { 0 };
3002 struct ptlrpc_thread *thread;
3003 struct list_head zombie;
3007 CDEBUG(D_INFO, "Stopping threads for service %s\n",
3008 svcpt->scp_service->srv_name);
3010 INIT_LIST_HEAD(&zombie);
3011 spin_lock(&svcpt->scp_lock);
3012 /* let the thread know that we would like it to stop asap */
3013 list_for_each_entry(thread, &svcpt->scp_threads, t_link)
3014 ptlrpc_stop_thread(thread);
3016 wake_up_all(&svcpt->scp_waitq);
3018 while (!list_empty(&svcpt->scp_threads)) {
3019 thread = list_entry(svcpt->scp_threads.next,
3020 struct ptlrpc_thread, t_link);
3021 if (thread_is_stopped(thread)) {
3022 list_move(&thread->t_link, &zombie);
3025 spin_unlock(&svcpt->scp_lock);
3027 CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
3028 svcpt->scp_service->srv_thread_name, thread->t_id);
3029 l_wait_event(thread->t_ctl_waitq,
3030 thread_is_stopped(thread), &lwi);
3032 spin_lock(&svcpt->scp_lock);
3035 spin_unlock(&svcpt->scp_lock);
3037 while (!list_empty(&zombie)) {
3038 thread = list_entry(zombie.next,
3039 struct ptlrpc_thread, t_link);
3040 list_del(&thread->t_link);
3041 OBD_FREE_PTR(thread);
3047 * Stops all threads of a particular service \a svc
3049 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
3051 struct ptlrpc_service_part *svcpt;
3056 ptlrpc_service_for_each_part(svcpt, i, svc) {
3057 if (svcpt->scp_service != NULL)
3058 ptlrpc_svcpt_stop_threads(svcpt);
3064 int ptlrpc_start_threads(struct ptlrpc_service *svc)
3072 /* We require 2 threads min, see note in ptlrpc_server_handle_request */
3073 LASSERT(svc->srv_nthrs_cpt_init >= PTLRPC_NTHRS_INIT);
3075 for (i = 0; i < svc->srv_ncpts; i++) {
3076 for (j = 0; j < svc->srv_nthrs_cpt_init; j++) {
3077 rc = ptlrpc_start_thread(svc->srv_parts[i], 1);
3083 /* We have enough threads, don't start more. b=15759 */
3090 CERROR("cannot start %s thread #%d_%d: rc %d\n",
3091 svc->srv_thread_name, i, j, rc);
3092 ptlrpc_stop_all_threads(svc);
3096 int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
3098 struct l_wait_info lwi = { 0 };
3099 struct ptlrpc_thread *thread;
3100 struct ptlrpc_service *svc;
3101 struct task_struct *task;
3106 LASSERT(svcpt != NULL);
3108 svc = svcpt->scp_service;
3110 CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n",
3111 svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running,
3112 svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit);
3115 if (unlikely(svc->srv_is_stopping))
3118 if (!ptlrpc_threads_increasable(svcpt) ||
3119 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
3120 svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1))
3123 OBD_CPT_ALLOC_PTR(thread, svc->srv_cptable, svcpt->scp_cpt);
3126 init_waitqueue_head(&thread->t_ctl_waitq);
3128 spin_lock(&svcpt->scp_lock);
3129 if (!ptlrpc_threads_increasable(svcpt)) {
3130 spin_unlock(&svcpt->scp_lock);
3131 OBD_FREE_PTR(thread);
3135 if (svcpt->scp_nthrs_starting != 0) {
3137 * serialize starting because some modules (obdfilter)
3138 * might require unique and contiguous t_id
3140 LASSERT(svcpt->scp_nthrs_starting == 1);
3141 spin_unlock(&svcpt->scp_lock);
3142 OBD_FREE_PTR(thread);
3144 CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n",
3145 svc->srv_thread_name, svcpt->scp_thr_nextid);
3150 CDEBUG(D_INFO, "Creating thread %s #%d race, retry later\n",
3151 svc->srv_thread_name, svcpt->scp_thr_nextid);
3155 svcpt->scp_nthrs_starting++;
3156 thread->t_id = svcpt->scp_thr_nextid++;
3157 thread_add_flags(thread, SVC_STARTING);
3158 thread->t_svcpt = svcpt;
3160 list_add(&thread->t_link, &svcpt->scp_threads);
3161 spin_unlock(&svcpt->scp_lock);
3163 if (svcpt->scp_cpt >= 0) {
3164 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s%02d_%03d",
3165 svc->srv_thread_name, svcpt->scp_cpt, thread->t_id);
3167 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s_%04d",
3168 svc->srv_thread_name, thread->t_id);
3171 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name);
3172 task = kthread_run(ptlrpc_main, thread, "%s", thread->t_name);
3175 CERROR("cannot start thread '%s': rc = %d\n",
3176 thread->t_name, rc);
3177 spin_lock(&svcpt->scp_lock);
3178 --svcpt->scp_nthrs_starting;
3179 if (thread_is_stopping(thread)) {
3181 * this ptlrpc_thread is being hanled
3182 * by ptlrpc_svcpt_stop_threads now
3184 thread_add_flags(thread, SVC_STOPPED);
3185 wake_up(&thread->t_ctl_waitq);
3186 spin_unlock(&svcpt->scp_lock);
3188 list_del(&thread->t_link);
3189 spin_unlock(&svcpt->scp_lock);
3190 OBD_FREE_PTR(thread);
3198 l_wait_event(thread->t_ctl_waitq,
3199 thread_is_running(thread) || thread_is_stopped(thread),
3202 rc = thread_is_stopped(thread) ? thread->t_id : 0;
3206 int ptlrpc_hr_init(void)
3208 struct ptlrpc_hr_partition *hrp;
3209 struct ptlrpc_hr_thread *hrt;
3217 memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
3218 ptlrpc_hr.hr_cpt_table = cfs_cpt_table;
3220 ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
3222 if (ptlrpc_hr.hr_partitions == NULL)
3225 ratelimit_state_init(&watchdog_limit,
3226 cfs_time_seconds(libcfs_watchdog_ratelimit), 3);
3228 init_waitqueue_head(&ptlrpc_hr.hr_waitq);
3231 weight = cpumask_weight(topology_sibling_cpumask(smp_processor_id()));
3234 cfs_percpt_for_each(hrp, cpt, ptlrpc_hr.hr_partitions) {
3237 atomic_set(&hrp->hrp_nstarted, 0);
3238 atomic_set(&hrp->hrp_nstopped, 0);
3240 hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, cpt);
3241 hrp->hrp_nthrs /= weight;
3242 if (hrp->hrp_nthrs == 0)
3245 OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, cpt,
3246 hrp->hrp_nthrs * sizeof(*hrt));
3247 if (hrp->hrp_thrs == NULL)
3248 GOTO(out, rc = -ENOMEM);
3250 for (i = 0; i < hrp->hrp_nthrs; i++) {
3251 hrt = &hrp->hrp_thrs[i];
3254 hrt->hrt_partition = hrp;
3255 init_waitqueue_head(&hrt->hrt_waitq);
3256 spin_lock_init(&hrt->hrt_lock);
3257 INIT_LIST_HEAD(&hrt->hrt_queue);
3261 rc = ptlrpc_start_hr_threads();
3268 void ptlrpc_hr_fini(void)
3270 struct ptlrpc_hr_partition *hrp;
3273 if (ptlrpc_hr.hr_partitions == NULL)
3276 ptlrpc_stop_hr_threads();
3278 cfs_percpt_for_each(hrp, cpt, ptlrpc_hr.hr_partitions) {
3279 if (hrp->hrp_thrs != NULL) {
3280 OBD_FREE(hrp->hrp_thrs,
3281 hrp->hrp_nthrs * sizeof(hrp->hrp_thrs[0]));
3285 cfs_percpt_free(ptlrpc_hr.hr_partitions);
3286 ptlrpc_hr.hr_partitions = NULL;
3291 * Wait until all already scheduled replies are processed.
3293 static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
3297 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
3300 rc = l_wait_event(svcpt->scp_waitq,
3301 atomic_read(&svcpt->scp_nreps_difficult) == 0, &lwi);
3304 CWARN("Unexpectedly long timeout %s %p\n",
3305 svcpt->scp_service->srv_name, svcpt->scp_service);
3310 ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
3312 struct ptlrpc_service_part *svcpt;
3315 /* early disarm AT timer... */
3316 ptlrpc_service_for_each_part(svcpt, i, svc) {
3317 if (svcpt->scp_service != NULL)
3318 del_timer(&svcpt->scp_at_timer);
3323 ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
3325 struct ptlrpc_service_part *svcpt;
3326 struct ptlrpc_request_buffer_desc *rqbd;
3327 struct l_wait_info lwi;
3332 * All history will be culled when the next request buffer is
3333 * freed in ptlrpc_service_purge_all()
3335 svc->srv_hist_nrqbds_cpt_max = 0;
3337 rc = LNetClearLazyPortal(svc->srv_req_portal);
3340 ptlrpc_service_for_each_part(svcpt, i, svc) {
3341 if (svcpt->scp_service == NULL)
3345 * Unlink all the request buffers. This forces a 'final'
3346 * event with its 'unlink' flag set for each posted rqbd
3348 list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
3350 rc = LNetMDUnlink(rqbd->rqbd_md_h);
3351 LASSERT(rc == 0 || rc == -ENOENT);
3355 ptlrpc_service_for_each_part(svcpt, i, svc) {
3356 if (svcpt->scp_service == NULL)
3360 * Wait for the network to release any buffers
3361 * it's currently filling
3363 spin_lock(&svcpt->scp_lock);
3364 while (svcpt->scp_nrqbds_posted != 0) {
3365 spin_unlock(&svcpt->scp_lock);
3367 * Network access will complete in finite time but
3368 * the HUGE timeout lets us CWARN for visibility
3371 lwi = LWI_TIMEOUT_INTERVAL(
3372 cfs_time_seconds(LONG_UNLINK),
3373 cfs_time_seconds(1), NULL, NULL);
3374 rc = l_wait_event(svcpt->scp_waitq,
3375 svcpt->scp_nrqbds_posted == 0, &lwi);
3376 if (rc == -ETIMEDOUT) {
3377 CWARN("Service %s waiting for request buffers\n",
3378 svcpt->scp_service->srv_name);
3380 spin_lock(&svcpt->scp_lock);
3382 spin_unlock(&svcpt->scp_lock);
3387 ptlrpc_service_purge_all(struct ptlrpc_service *svc)
3389 struct ptlrpc_service_part *svcpt;
3390 struct ptlrpc_request_buffer_desc *rqbd;
3391 struct ptlrpc_request *req;
3392 struct ptlrpc_reply_state *rs;
3395 ptlrpc_service_for_each_part(svcpt, i, svc) {
3396 if (svcpt->scp_service == NULL)
3399 spin_lock(&svcpt->scp_rep_lock);
3400 while (!list_empty(&svcpt->scp_rep_active)) {
3401 rs = list_entry(svcpt->scp_rep_active.next,
3402 struct ptlrpc_reply_state, rs_list);
3403 spin_lock(&rs->rs_lock);
3404 ptlrpc_schedule_difficult_reply(rs);
3405 spin_unlock(&rs->rs_lock);
3407 spin_unlock(&svcpt->scp_rep_lock);
3410 * purge the request queue. NB No new replies (rqbds
3411 * all unlinked) and no service threads, so I'm the only
3412 * thread noodling the request queue now
3414 while (!list_empty(&svcpt->scp_req_incoming)) {
3415 req = list_entry(svcpt->scp_req_incoming.next,
3416 struct ptlrpc_request, rq_list);
3418 list_del(&req->rq_list);
3419 svcpt->scp_nreqs_incoming--;
3420 ptlrpc_server_finish_request(svcpt, req);
3423 while (ptlrpc_server_request_pending(svcpt, true)) {
3424 req = ptlrpc_server_request_get(svcpt, true);
3425 ptlrpc_server_finish_active_request(svcpt, req);
3428 LASSERT(list_empty(&svcpt->scp_rqbd_posted));
3429 LASSERT(svcpt->scp_nreqs_incoming == 0);
3430 LASSERT(svcpt->scp_nreqs_active == 0);
3432 * history should have been culled by
3433 * ptlrpc_server_finish_request
3435 LASSERT(svcpt->scp_hist_nrqbds == 0);
3438 * Now free all the request buffers since nothing
3439 * references them any more...
3442 while (!list_empty(&svcpt->scp_rqbd_idle)) {
3443 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
3444 struct ptlrpc_request_buffer_desc,
3446 ptlrpc_free_rqbd(rqbd);
3448 ptlrpc_wait_replies(svcpt);
3450 while (!list_empty(&svcpt->scp_rep_idle)) {
3451 rs = list_entry(svcpt->scp_rep_idle.next,
3452 struct ptlrpc_reply_state,
3454 list_del(&rs->rs_list);
3455 OBD_FREE_LARGE(rs, svc->srv_max_reply_size);
3461 ptlrpc_service_free(struct ptlrpc_service *svc)
3463 struct ptlrpc_service_part *svcpt;
3464 struct ptlrpc_at_array *array;
3467 ptlrpc_service_for_each_part(svcpt, i, svc) {
3468 if (svcpt->scp_service == NULL)
3471 /* In case somebody rearmed this in the meantime */
3472 del_timer(&svcpt->scp_at_timer);
3473 array = &svcpt->scp_at_array;
3475 if (array->paa_reqs_array != NULL) {
3476 OBD_FREE(array->paa_reqs_array,
3477 sizeof(struct list_head) * array->paa_size);
3478 array->paa_reqs_array = NULL;
3481 if (array->paa_reqs_count != NULL) {
3482 OBD_FREE(array->paa_reqs_count,
3483 sizeof(__u32) * array->paa_size);
3484 array->paa_reqs_count = NULL;
3488 ptlrpc_service_for_each_part(svcpt, i, svc)
3489 OBD_FREE_PTR(svcpt);
3491 if (svc->srv_cpts != NULL)
3492 cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts);
3494 OBD_FREE(svc, offsetof(struct ptlrpc_service,
3495 srv_parts[svc->srv_ncpts]));
3498 int ptlrpc_unregister_service(struct ptlrpc_service *service)
3502 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
3504 service->srv_is_stopping = 1;
3506 mutex_lock(&ptlrpc_all_services_mutex);
3507 list_del_init(&service->srv_list);
3508 mutex_unlock(&ptlrpc_all_services_mutex);
3510 ptlrpc_service_del_atimer(service);
3511 ptlrpc_stop_all_threads(service);
3513 ptlrpc_service_unlink_rqbd(service);
3514 ptlrpc_service_purge_all(service);
3515 ptlrpc_service_nrs_cleanup(service);
3517 ptlrpc_lprocfs_unregister_service(service);
3518 ptlrpc_sysfs_unregister_service(service);
3520 ptlrpc_service_free(service);
3524 EXPORT_SYMBOL(ptlrpc_unregister_service);
3527 * Returns 0 if the service is healthy.
3529 * Right now, it just checks to make sure that requests aren't languishing
3530 * in the queue. We'll use this health check to govern whether a node needs
3531 * to be shot, so it's intentionally non-aggressive.
3533 static int ptlrpc_svcpt_health_check(struct ptlrpc_service_part *svcpt)
3535 struct ptlrpc_request *request = NULL;
3536 struct timespec64 right_now;
3537 struct timespec64 timediff;
3539 ktime_get_real_ts64(&right_now);
3541 spin_lock(&svcpt->scp_req_lock);
3542 /* How long has the next entry been waiting? */
3543 if (ptlrpc_server_high_pending(svcpt, true))
3544 request = ptlrpc_nrs_req_peek_nolock(svcpt, true);
3545 else if (ptlrpc_server_normal_pending(svcpt, true))
3546 request = ptlrpc_nrs_req_peek_nolock(svcpt, false);
3548 if (request == NULL) {
3549 spin_unlock(&svcpt->scp_req_lock);
3553 timediff = timespec64_sub(right_now, request->rq_arrival_time);
3554 spin_unlock(&svcpt->scp_req_lock);
3556 if ((timediff.tv_sec) >
3557 (AT_OFF ? obd_timeout * 3 / 2 : at_max)) {
3558 CERROR("%s: unhealthy - request has been waiting %llds\n",
3559 svcpt->scp_service->srv_name, (s64)timediff.tv_sec);
3567 ptlrpc_service_health_check(struct ptlrpc_service *svc)
3569 struct ptlrpc_service_part *svcpt;
3575 ptlrpc_service_for_each_part(svcpt, i, svc) {
3576 int rc = ptlrpc_svcpt_health_check(svcpt);
3583 EXPORT_SYMBOL(ptlrpc_service_health_check);