4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
33 #define DEBUG_SUBSYSTEM S_RPC
35 #include <linux/kthread.h>
36 #include <obd_support.h>
37 #include <obd_class.h>
38 #include <lustre_net.h>
39 #include <lu_object.h>
40 #include <uapi/linux/lnet/lnet-types.h>
41 #include "ptlrpc_internal.h"
43 /* The following are visible and mutable through /sys/module/ptlrpc */
44 int test_req_buffer_pressure = 0;
45 module_param(test_req_buffer_pressure, int, 0444);
46 MODULE_PARM_DESC(test_req_buffer_pressure, "set non-zero to put pressure on request buffer pools");
47 module_param(at_min, int, 0644);
48 MODULE_PARM_DESC(at_min, "Adaptive timeout minimum (sec)");
49 module_param(at_max, int, 0644);
50 MODULE_PARM_DESC(at_max, "Adaptive timeout maximum (sec)");
51 module_param(at_history, int, 0644);
52 MODULE_PARM_DESC(at_history,
53 "Adaptive timeouts remember the slowest event that took place within this period (sec)");
54 module_param(at_early_margin, int, 0644);
55 MODULE_PARM_DESC(at_early_margin, "How soon before an RPC deadline to send an early reply");
56 module_param(at_extra, int, 0644);
57 MODULE_PARM_DESC(at_extra, "How much extra time to give with each early reply");
60 static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt);
61 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req);
62 static void ptlrpc_at_remove_timed(struct ptlrpc_request *req);
64 /** Holds a list of all PTLRPC services */
65 struct list_head ptlrpc_all_services;
66 /** Used to protect the \e ptlrpc_all_services list */
67 struct mutex ptlrpc_all_services_mutex;
69 static struct ptlrpc_request_buffer_desc *
70 ptlrpc_alloc_rqbd(struct ptlrpc_service_part *svcpt)
72 struct ptlrpc_service *svc = svcpt->scp_service;
73 struct ptlrpc_request_buffer_desc *rqbd;
75 OBD_CPT_ALLOC_PTR(rqbd, svc->srv_cptable, svcpt->scp_cpt);
79 rqbd->rqbd_svcpt = svcpt;
80 rqbd->rqbd_refcount = 0;
81 rqbd->rqbd_cbid.cbid_fn = request_in_callback;
82 rqbd->rqbd_cbid.cbid_arg = rqbd;
83 INIT_LIST_HEAD(&rqbd->rqbd_reqs);
84 OBD_CPT_ALLOC_LARGE(rqbd->rqbd_buffer, svc->srv_cptable,
85 svcpt->scp_cpt, svc->srv_buf_size);
86 if (rqbd->rqbd_buffer == NULL) {
91 spin_lock(&svcpt->scp_lock);
92 list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
93 svcpt->scp_nrqbds_total++;
94 spin_unlock(&svcpt->scp_lock);
100 ptlrpc_free_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
102 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
104 LASSERT(rqbd->rqbd_refcount == 0);
105 LASSERT(list_empty(&rqbd->rqbd_reqs));
107 spin_lock(&svcpt->scp_lock);
108 list_del(&rqbd->rqbd_list);
109 svcpt->scp_nrqbds_total--;
110 spin_unlock(&svcpt->scp_lock);
112 OBD_FREE_LARGE(rqbd->rqbd_buffer, svcpt->scp_service->srv_buf_size);
117 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++) {
141 /* NB: another thread might have recycled enough rqbds, we
142 * need to make sure it wouldn't over-allocate, see LU-1212. */
143 if (svcpt->scp_nrqbds_posted >= svc->srv_nbuf_per_group)
146 rqbd = ptlrpc_alloc_rqbd(svcpt);
149 CERROR("%s: Can't allocate request buffer\n",
156 spin_lock(&svcpt->scp_lock);
158 LASSERT(svcpt->scp_rqbd_allocating == 1);
159 svcpt->scp_rqbd_allocating--;
161 spin_unlock(&svcpt->scp_lock);
164 "%s: allocate %d new %d-byte reqbufs (%d/%d left), rc = %d\n",
165 svc->srv_name, i, svc->srv_buf_size, svcpt->scp_nrqbds_posted,
166 svcpt->scp_nrqbds_total, rc);
170 rc = ptlrpc_server_post_idle_rqbds(svcpt);
176 * Part of Rep-Ack logic.
177 * Puts a lock and its mode into reply state assotiated to request reply.
180 ptlrpc_save_lock(struct ptlrpc_request *req, struct lustre_handle *lock,
181 int mode, bool no_ack, bool convert_lock)
183 struct ptlrpc_reply_state *rs = req->rq_reply_state;
187 LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
189 idx = rs->rs_nlocks++;
190 rs->rs_locks[idx] = *lock;
191 rs->rs_modes[idx] = mode;
192 rs->rs_difficult = 1;
193 rs->rs_no_ack = no_ack;
194 rs->rs_convert_lock = convert_lock;
196 EXPORT_SYMBOL(ptlrpc_save_lock);
199 struct ptlrpc_hr_partition;
201 struct ptlrpc_hr_thread {
202 int hrt_id; /* thread ID */
204 wait_queue_head_t hrt_waitq;
205 struct list_head hrt_queue;
206 struct ptlrpc_hr_partition *hrt_partition;
209 struct ptlrpc_hr_partition {
210 /* # of started threads */
211 atomic_t hrp_nstarted;
212 /* # of stopped threads */
213 atomic_t hrp_nstopped;
214 /* cpu partition id */
216 /* round-robin rotor for choosing thread */
218 /* total number of threads on this partition */
221 struct ptlrpc_hr_thread *hrp_thrs;
224 #define HRT_RUNNING 0
225 #define HRT_STOPPING 1
227 struct ptlrpc_hr_service {
228 /* CPU partition table, it's just cfs_cpt_table for now */
229 struct cfs_cpt_table *hr_cpt_table;
230 /** controller sleep waitq */
231 wait_queue_head_t hr_waitq;
232 unsigned int hr_stopping;
233 /** roundrobin rotor for non-affinity service */
234 unsigned int hr_rotor;
236 struct ptlrpc_hr_partition **hr_partitions;
240 struct list_head rsb_replies;
241 unsigned int rsb_n_replies;
242 struct ptlrpc_service_part *rsb_svcpt;
245 /** reply handling service. */
246 static struct ptlrpc_hr_service ptlrpc_hr;
249 * maximum mumber of replies scheduled in one batch
251 #define MAX_SCHEDULED 256
254 * Initialize a reply batch.
258 static void rs_batch_init(struct rs_batch *b)
260 memset(b, 0, sizeof *b);
261 INIT_LIST_HEAD(&b->rsb_replies);
265 * Choose an hr thread to dispatch requests to.
267 static struct ptlrpc_hr_thread *
268 ptlrpc_hr_select(struct ptlrpc_service_part *svcpt)
270 struct ptlrpc_hr_partition *hrp;
273 if (svcpt->scp_cpt >= 0 &&
274 svcpt->scp_service->srv_cptable == ptlrpc_hr.hr_cpt_table) {
275 /* directly match partition */
276 hrp = ptlrpc_hr.hr_partitions[svcpt->scp_cpt];
279 rotor = ptlrpc_hr.hr_rotor++;
280 rotor %= cfs_cpt_number(ptlrpc_hr.hr_cpt_table);
282 hrp = ptlrpc_hr.hr_partitions[rotor];
285 rotor = hrp->hrp_rotor++;
286 return &hrp->hrp_thrs[rotor % hrp->hrp_nthrs];
290 * Dispatch all replies accumulated in the batch to one from
291 * dedicated reply handling threads.
295 static void rs_batch_dispatch(struct rs_batch *b)
297 if (b->rsb_n_replies != 0) {
298 struct ptlrpc_hr_thread *hrt;
300 hrt = ptlrpc_hr_select(b->rsb_svcpt);
302 spin_lock(&hrt->hrt_lock);
303 list_splice_init(&b->rsb_replies, &hrt->hrt_queue);
304 spin_unlock(&hrt->hrt_lock);
306 wake_up(&hrt->hrt_waitq);
307 b->rsb_n_replies = 0;
312 * Add a reply to a batch.
313 * Add one reply object to a batch, schedule batched replies if overload.
318 static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs)
320 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
322 if (svcpt != b->rsb_svcpt || b->rsb_n_replies >= MAX_SCHEDULED) {
323 if (b->rsb_svcpt != NULL) {
324 rs_batch_dispatch(b);
325 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
327 spin_lock(&svcpt->scp_rep_lock);
328 b->rsb_svcpt = svcpt;
330 spin_lock(&rs->rs_lock);
331 rs->rs_scheduled_ever = 1;
332 if (rs->rs_scheduled == 0) {
333 list_move(&rs->rs_list, &b->rsb_replies);
334 rs->rs_scheduled = 1;
337 rs->rs_committed = 1;
338 spin_unlock(&rs->rs_lock);
342 * Reply batch finalization.
343 * Dispatch remaining replies from the batch
344 * and release remaining spinlock.
348 static void rs_batch_fini(struct rs_batch *b)
350 if (b->rsb_svcpt != NULL) {
351 rs_batch_dispatch(b);
352 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
356 #define DECLARE_RS_BATCH(b) struct rs_batch b
360 * Put reply state into a queue for processing because we received
361 * ACK from the client
363 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
365 struct ptlrpc_hr_thread *hrt;
368 LASSERT(list_empty(&rs->rs_list));
370 hrt = ptlrpc_hr_select(rs->rs_svcpt);
372 spin_lock(&hrt->hrt_lock);
373 list_add_tail(&rs->rs_list, &hrt->hrt_queue);
374 spin_unlock(&hrt->hrt_lock);
376 wake_up(&hrt->hrt_waitq);
381 ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs)
385 assert_spin_locked(&rs->rs_svcpt->scp_rep_lock);
386 assert_spin_locked(&rs->rs_lock);
387 LASSERT (rs->rs_difficult);
388 rs->rs_scheduled_ever = 1; /* flag any notification attempt */
390 if (rs->rs_scheduled) { /* being set up or already notified */
395 rs->rs_scheduled = 1;
396 list_del_init(&rs->rs_list);
397 ptlrpc_dispatch_difficult_reply(rs);
400 EXPORT_SYMBOL(ptlrpc_schedule_difficult_reply);
402 void ptlrpc_commit_replies(struct obd_export *exp)
404 struct ptlrpc_reply_state *rs, *nxt;
405 DECLARE_RS_BATCH(batch);
408 rs_batch_init(&batch);
409 /* Find any replies that have been committed and get their service
410 * to attend to complete them. */
412 /* CAVEAT EMPTOR: spinlock ordering!!! */
413 spin_lock(&exp->exp_uncommitted_replies_lock);
414 list_for_each_entry_safe(rs, nxt, &exp->exp_uncommitted_replies,
416 LASSERT (rs->rs_difficult);
417 /* VBR: per-export last_committed */
418 LASSERT(rs->rs_export);
419 if (rs->rs_transno <= exp->exp_last_committed) {
420 list_del_init(&rs->rs_obd_list);
421 rs_batch_add(&batch, rs);
424 spin_unlock(&exp->exp_uncommitted_replies_lock);
425 rs_batch_fini(&batch);
430 ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt)
432 struct ptlrpc_request_buffer_desc *rqbd;
437 spin_lock(&svcpt->scp_lock);
439 if (list_empty(&svcpt->scp_rqbd_idle)) {
440 spin_unlock(&svcpt->scp_lock);
444 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
445 struct ptlrpc_request_buffer_desc,
447 list_del(&rqbd->rqbd_list);
449 /* assume we will post successfully */
450 svcpt->scp_nrqbds_posted++;
451 list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_posted);
453 spin_unlock(&svcpt->scp_lock);
455 rc = ptlrpc_register_rqbd(rqbd);
462 spin_lock(&svcpt->scp_lock);
464 svcpt->scp_nrqbds_posted--;
465 list_del(&rqbd->rqbd_list);
466 list_add_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
468 /* Don't complain if no request buffers are posted right now; LNET
469 * won't drop requests because we set the portal lazy! */
471 spin_unlock(&svcpt->scp_lock);
476 static void ptlrpc_at_timer(unsigned long castmeharder)
478 struct ptlrpc_service_part *svcpt;
480 svcpt = (struct ptlrpc_service_part *)castmeharder;
482 svcpt->scp_at_check = 1;
483 svcpt->scp_at_checktime = cfs_time_current();
484 wake_up(&svcpt->scp_waitq);
488 ptlrpc_server_nthreads_check(struct ptlrpc_service *svc,
489 struct ptlrpc_service_conf *conf)
491 struct ptlrpc_service_thr_conf *tc = &conf->psc_thr;
498 * Common code for estimating & validating threads number.
499 * CPT affinity service could have percpt thread-pool instead
500 * of a global thread-pool, which means user might not always
501 * get the threads number they give it in conf::tc_nthrs_user
502 * even they did set. It's because we need to validate threads
503 * number for each CPT to guarantee each pool will have enough
504 * threads to keep the service healthy.
506 init = PTLRPC_NTHRS_INIT + (svc->srv_ops.so_hpreq_handler != NULL);
507 init = max_t(int, init, tc->tc_nthrs_init);
509 /* NB: please see comments in lustre_lnet.h for definition
510 * details of these members */
511 LASSERT(tc->tc_nthrs_max != 0);
513 if (tc->tc_nthrs_user != 0) {
514 /* In case there is a reason to test a service with many
515 * threads, we give a less strict check here, it can
516 * be up to 8 * nthrs_max */
517 total = min(tc->tc_nthrs_max * 8, tc->tc_nthrs_user);
518 nthrs = total / svc->srv_ncpts;
519 init = max(init, nthrs);
523 total = tc->tc_nthrs_max;
524 if (tc->tc_nthrs_base == 0) {
525 /* don't care about base threads number per partition,
526 * this is most for non-affinity service */
527 nthrs = total / svc->srv_ncpts;
531 nthrs = tc->tc_nthrs_base;
532 if (svc->srv_ncpts == 1) {
535 /* NB: Increase the base number if it's single partition
536 * and total number of cores/HTs is larger or equal to 4.
537 * result will always < 2 * nthrs_base */
538 weight = cfs_cpt_weight(svc->srv_cptable, CFS_CPT_ANY);
539 for (i = 1; (weight >> (i + 1)) != 0 && /* >= 4 cores/HTs */
540 (tc->tc_nthrs_base >> i) != 0; i++)
541 nthrs += tc->tc_nthrs_base >> i;
544 if (tc->tc_thr_factor != 0) {
545 int factor = tc->tc_thr_factor;
549 * User wants to increase number of threads with for
550 * each CPU core/HT, most likely the factor is larger than
551 * one thread/core because service threads are supposed to
552 * be blocked by lock or wait for IO.
555 * Amdahl's law says that adding processors wouldn't give
556 * a linear increasing of parallelism, so it's nonsense to
557 * have too many threads no matter how many cores/HTs
560 if (cpumask_weight(topology_sibling_cpumask(smp_processor_id())) > 1) {
561 /* weight is # of HTs */
562 /* depress thread factor for hyper-thread */
563 factor = factor - (factor >> 1) + (factor >> 3);
566 weight = cfs_cpt_weight(svc->srv_cptable, 0);
568 for (; factor > 0 && weight > 0; factor--, weight -= fade)
569 nthrs += min(weight, fade) * factor;
572 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
573 nthrs = max(tc->tc_nthrs_base,
574 tc->tc_nthrs_max / svc->srv_ncpts);
577 nthrs = max(nthrs, tc->tc_nthrs_init);
578 svc->srv_nthrs_cpt_limit = nthrs;
579 svc->srv_nthrs_cpt_init = init;
581 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
582 CDEBUG(D_OTHER, "%s: This service may have more threads (%d) "
583 "than the given soft limit (%d)\n",
584 svc->srv_name, nthrs * svc->srv_ncpts,
590 * Initialize percpt data for a service
593 ptlrpc_service_part_init(struct ptlrpc_service *svc,
594 struct ptlrpc_service_part *svcpt, int cpt)
596 struct ptlrpc_at_array *array;
601 svcpt->scp_cpt = cpt;
602 INIT_LIST_HEAD(&svcpt->scp_threads);
604 /* rqbd and incoming request queue */
605 spin_lock_init(&svcpt->scp_lock);
606 INIT_LIST_HEAD(&svcpt->scp_rqbd_idle);
607 INIT_LIST_HEAD(&svcpt->scp_rqbd_posted);
608 INIT_LIST_HEAD(&svcpt->scp_req_incoming);
609 init_waitqueue_head(&svcpt->scp_waitq);
610 /* history request & rqbd list */
611 INIT_LIST_HEAD(&svcpt->scp_hist_reqs);
612 INIT_LIST_HEAD(&svcpt->scp_hist_rqbds);
614 /* acitve requests and hp requests */
615 spin_lock_init(&svcpt->scp_req_lock);
618 spin_lock_init(&svcpt->scp_rep_lock);
619 INIT_LIST_HEAD(&svcpt->scp_rep_active);
620 INIT_LIST_HEAD(&svcpt->scp_rep_idle);
621 init_waitqueue_head(&svcpt->scp_rep_waitq);
622 atomic_set(&svcpt->scp_nreps_difficult, 0);
624 /* adaptive timeout */
625 spin_lock_init(&svcpt->scp_at_lock);
626 array = &svcpt->scp_at_array;
628 size = at_est2timeout(at_max);
629 array->paa_size = size;
630 array->paa_count = 0;
631 array->paa_deadline = -1;
633 /* allocate memory for scp_at_array (ptlrpc_at_array) */
634 OBD_CPT_ALLOC(array->paa_reqs_array,
635 svc->srv_cptable, cpt, sizeof(struct list_head) * size);
636 if (array->paa_reqs_array == NULL)
639 for (index = 0; index < size; index++)
640 INIT_LIST_HEAD(&array->paa_reqs_array[index]);
642 OBD_CPT_ALLOC(array->paa_reqs_count,
643 svc->srv_cptable, cpt, sizeof(__u32) * size);
644 if (array->paa_reqs_count == NULL)
647 setup_timer(&svcpt->scp_at_timer, ptlrpc_at_timer,
648 (unsigned long)svcpt);
650 /* At SOW, service time should be quick; 10s seems generous. If client
651 * timeout is less than this, we'll be sending an early reply. */
652 at_init(&svcpt->scp_at_estimate, 10, 0);
654 /* assign this before call ptlrpc_grow_req_bufs */
655 svcpt->scp_service = svc;
656 /* Now allocate the request buffers, but don't post them now */
657 rc = ptlrpc_grow_req_bufs(svcpt, 0);
658 /* We shouldn't be under memory pressure at startup, so
659 * fail if we can't allocate all our buffers at this time. */
666 if (array->paa_reqs_count != NULL) {
667 OBD_FREE(array->paa_reqs_count, sizeof(__u32) * size);
668 array->paa_reqs_count = NULL;
671 if (array->paa_reqs_array != NULL) {
672 OBD_FREE(array->paa_reqs_array,
673 sizeof(struct list_head) * array->paa_size);
674 array->paa_reqs_array = NULL;
681 * Initialize service on a given portal.
682 * This includes starting serving threads , allocating and posting rqbds and
685 struct ptlrpc_service *
686 ptlrpc_register_service(struct ptlrpc_service_conf *conf,
688 struct proc_dir_entry *proc_entry)
690 struct ptlrpc_service_cpt_conf *cconf = &conf->psc_cpt;
691 struct ptlrpc_service *service;
692 struct ptlrpc_service_part *svcpt;
693 struct cfs_cpt_table *cptable;
701 LASSERT(conf->psc_buf.bc_nbufs > 0);
702 LASSERT(conf->psc_buf.bc_buf_size >=
703 conf->psc_buf.bc_req_max_size + SPTLRPC_MAX_PAYLOAD);
704 LASSERT(conf->psc_thr.tc_ctx_tags != 0);
706 cptable = cconf->cc_cptable;
708 cptable = cfs_cpt_table;
710 if (!conf->psc_thr.tc_cpu_affinity) {
713 ncpts = cfs_cpt_number(cptable);
714 if (cconf->cc_pattern != NULL) {
715 struct cfs_expr_list *el;
717 rc = cfs_expr_list_parse(cconf->cc_pattern,
718 strlen(cconf->cc_pattern),
721 CERROR("%s: invalid CPT pattern string: %s",
722 conf->psc_name, cconf->cc_pattern);
723 RETURN(ERR_PTR(-EINVAL));
726 rc = cfs_expr_list_values(el, ncpts, &cpts);
727 cfs_expr_list_free(el);
729 CERROR("%s: failed to parse CPT array %s: %d\n",
730 conf->psc_name, cconf->cc_pattern, rc);
732 OBD_FREE(cpts, sizeof(*cpts) * ncpts);
733 RETURN(ERR_PTR(rc < 0 ? rc : -EINVAL));
739 OBD_ALLOC(service, offsetof(struct ptlrpc_service, srv_parts[ncpts]));
740 if (service == NULL) {
742 OBD_FREE(cpts, sizeof(*cpts) * ncpts);
743 RETURN(ERR_PTR(-ENOMEM));
746 service->srv_cptable = cptable;
747 service->srv_cpts = cpts;
748 service->srv_ncpts = ncpts;
750 service->srv_cpt_bits = 0; /* it's zero already, easy to read... */
751 while ((1 << service->srv_cpt_bits) < cfs_cpt_number(cptable))
752 service->srv_cpt_bits++;
755 spin_lock_init(&service->srv_lock);
756 service->srv_name = conf->psc_name;
757 service->srv_watchdog_factor = conf->psc_watchdog_factor;
758 INIT_LIST_HEAD(&service->srv_list); /* for safty of cleanup */
760 /* buffer configuration */
761 service->srv_nbuf_per_group = test_req_buffer_pressure ?
762 1 : conf->psc_buf.bc_nbufs;
763 service->srv_max_req_size = conf->psc_buf.bc_req_max_size +
765 service->srv_buf_size = conf->psc_buf.bc_buf_size;
766 service->srv_rep_portal = conf->psc_buf.bc_rep_portal;
767 service->srv_req_portal = conf->psc_buf.bc_req_portal;
769 /* Increase max reply size to next power of two */
770 service->srv_max_reply_size = 1;
771 while (service->srv_max_reply_size <
772 conf->psc_buf.bc_rep_max_size + SPTLRPC_MAX_PAYLOAD)
773 service->srv_max_reply_size <<= 1;
775 service->srv_thread_name = conf->psc_thr.tc_thr_name;
776 service->srv_ctx_tags = conf->psc_thr.tc_ctx_tags;
777 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
778 service->srv_ops = conf->psc_ops;
780 for (i = 0; i < ncpts; i++) {
781 if (!conf->psc_thr.tc_cpu_affinity)
784 cpt = cpts != NULL ? cpts[i] : i;
786 OBD_CPT_ALLOC(svcpt, cptable, cpt, sizeof(*svcpt));
788 GOTO(failed, rc = -ENOMEM);
790 service->srv_parts[i] = svcpt;
791 rc = ptlrpc_service_part_init(service, svcpt, cpt);
796 ptlrpc_server_nthreads_check(service, conf);
798 rc = LNetSetLazyPortal(service->srv_req_portal);
801 mutex_lock(&ptlrpc_all_services_mutex);
802 list_add(&service->srv_list, &ptlrpc_all_services);
803 mutex_unlock(&ptlrpc_all_services_mutex);
806 rc = ptlrpc_sysfs_register_service(parent, service);
811 if (proc_entry != NULL)
812 ptlrpc_lprocfs_register_service(proc_entry, service);
814 rc = ptlrpc_service_nrs_setup(service);
818 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
819 service->srv_name, service->srv_req_portal);
821 rc = ptlrpc_start_threads(service);
823 CERROR("Failed to start threads for service %s: %d\n",
824 service->srv_name, rc);
830 ptlrpc_unregister_service(service);
833 EXPORT_SYMBOL(ptlrpc_register_service);
836 * to actually free the request, must be called without holding svc_lock.
837 * note it's caller's responsibility to unlink req->rq_list.
839 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
841 LASSERT(atomic_read(&req->rq_refcount) == 0);
842 LASSERT(list_empty(&req->rq_timed_list));
844 /* DEBUG_REQ() assumes the reply state of a request with a valid
845 * ref will not be destroyed until that reference is dropped. */
846 ptlrpc_req_drop_rs(req);
848 sptlrpc_svc_ctx_decref(req);
850 if (req != &req->rq_rqbd->rqbd_req) {
851 /* NB request buffers use an embedded
852 * req if the incoming req unlinked the
853 * MD; this isn't one of them! */
854 ptlrpc_request_cache_free(req);
859 * drop a reference count of the request. if it reaches 0, we either
860 * put it into history list, or free it immediately.
862 void ptlrpc_server_drop_request(struct ptlrpc_request *req)
864 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
865 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
866 struct ptlrpc_service *svc = svcpt->scp_service;
868 struct list_head *tmp;
869 struct list_head *nxt;
871 if (!atomic_dec_and_test(&req->rq_refcount))
874 if (req->rq_session.lc_state == LCS_ENTERED) {
875 lu_context_exit(&req->rq_session);
876 lu_context_fini(&req->rq_session);
879 if (req->rq_at_linked) {
880 spin_lock(&svcpt->scp_at_lock);
881 /* recheck with lock, in case it's unlinked by
882 * ptlrpc_at_check_timed() */
883 if (likely(req->rq_at_linked))
884 ptlrpc_at_remove_timed(req);
885 spin_unlock(&svcpt->scp_at_lock);
888 LASSERT(list_empty(&req->rq_timed_list));
890 /* finalize request */
891 if (req->rq_export) {
892 class_export_put(req->rq_export);
893 req->rq_export = NULL;
896 spin_lock(&svcpt->scp_lock);
898 list_add(&req->rq_list, &rqbd->rqbd_reqs);
900 refcount = --(rqbd->rqbd_refcount);
902 /* request buffer is now idle: add to history */
903 list_del(&rqbd->rqbd_list);
905 list_add_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds);
906 svcpt->scp_hist_nrqbds++;
908 /* cull some history?
909 * I expect only about 1 or 2 rqbds need to be recycled here */
910 while (svcpt->scp_hist_nrqbds > svc->srv_hist_nrqbds_cpt_max) {
911 rqbd = list_entry(svcpt->scp_hist_rqbds.next,
912 struct ptlrpc_request_buffer_desc,
915 list_del(&rqbd->rqbd_list);
916 svcpt->scp_hist_nrqbds--;
918 /* remove rqbd's reqs from svc's req history while
919 * I've got the service lock */
920 list_for_each(tmp, &rqbd->rqbd_reqs) {
921 req = list_entry(tmp, struct ptlrpc_request,
923 /* Track the highest culled req seq */
924 if (req->rq_history_seq >
925 svcpt->scp_hist_seq_culled) {
926 svcpt->scp_hist_seq_culled =
929 list_del(&req->rq_history_list);
932 spin_unlock(&svcpt->scp_lock);
934 list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
935 req = list_entry(rqbd->rqbd_reqs.next,
936 struct ptlrpc_request,
938 list_del(&req->rq_list);
939 ptlrpc_server_free_request(req);
942 spin_lock(&svcpt->scp_lock);
944 * now all reqs including the embedded req has been
945 * disposed, schedule request buffer for re-use
946 * or free it to drain some in excess.
948 LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) == 0);
949 if (svcpt->scp_nrqbds_posted >=
950 svc->srv_nbuf_per_group &&
951 !test_req_buffer_pressure) {
952 /* like in ptlrpc_free_rqbd() */
953 svcpt->scp_nrqbds_total--;
954 OBD_FREE_LARGE(rqbd->rqbd_buffer,
958 list_add_tail(&rqbd->rqbd_list,
959 &svcpt->scp_rqbd_idle);
963 spin_unlock(&svcpt->scp_lock);
964 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
965 /* If we are low on memory, we are not interested in history */
966 list_del(&req->rq_list);
967 list_del_init(&req->rq_history_list);
969 /* Track the highest culled req seq */
970 if (req->rq_history_seq > svcpt->scp_hist_seq_culled)
971 svcpt->scp_hist_seq_culled = req->rq_history_seq;
973 spin_unlock(&svcpt->scp_lock);
975 ptlrpc_server_free_request(req);
977 spin_unlock(&svcpt->scp_lock);
981 /** Change request export and move hp request from old export to new */
982 void ptlrpc_request_change_export(struct ptlrpc_request *req,
983 struct obd_export *export)
985 if (req->rq_export != NULL) {
986 LASSERT(!list_empty(&req->rq_exp_list));
987 /* remove rq_exp_list from last export */
988 spin_lock_bh(&req->rq_export->exp_rpc_lock);
989 list_del_init(&req->rq_exp_list);
990 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
991 /* export has one reference already, so it`s safe to
992 * add req to export queue here and get another
993 * reference for request later */
994 spin_lock_bh(&export->exp_rpc_lock);
995 if (req->rq_ops != NULL) /* hp request */
996 list_add(&req->rq_exp_list, &export->exp_hp_rpcs);
998 list_add(&req->rq_exp_list, &export->exp_reg_rpcs);
999 spin_unlock_bh(&export->exp_rpc_lock);
1001 class_export_rpc_dec(req->rq_export);
1002 class_export_put(req->rq_export);
1005 /* request takes one export refcount */
1006 req->rq_export = class_export_get(export);
1007 class_export_rpc_inc(export);
1013 * to finish a request: stop sending more early replies, and release
1016 static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt,
1017 struct ptlrpc_request *req)
1019 ptlrpc_server_hpreq_fini(req);
1021 ptlrpc_server_drop_request(req);
1025 * to finish an active request: stop sending more early replies, and release
1026 * the request. should be called after we finished handling the request.
1028 static void ptlrpc_server_finish_active_request(
1029 struct ptlrpc_service_part *svcpt,
1030 struct ptlrpc_request *req)
1032 spin_lock(&svcpt->scp_req_lock);
1033 ptlrpc_nrs_req_stop_nolock(req);
1034 svcpt->scp_nreqs_active--;
1036 svcpt->scp_nhreqs_active--;
1037 spin_unlock(&svcpt->scp_req_lock);
1039 ptlrpc_nrs_req_finalize(req);
1041 if (req->rq_export != NULL)
1042 class_export_rpc_dec(req->rq_export);
1044 ptlrpc_server_finish_request(svcpt, req);
1048 * This function makes sure dead exports are evicted in a timely manner.
1049 * This function is only called when some export receives a message (i.e.,
1050 * the network is up.)
1052 void ptlrpc_update_export_timer(struct obd_export *exp, time64_t extra_delay)
1054 struct obd_export *oldest_exp;
1055 time64_t oldest_time, new_time;
1061 /* Compensate for slow machines, etc, by faking our request time
1062 into the future. Although this can break the strict time-ordering
1063 of the list, we can be really lazy here - we don't have to evict
1064 at the exact right moment. Eventually, all silent exports
1065 will make it to the top of the list. */
1067 /* Do not pay attention on 1sec or smaller renewals. */
1068 new_time = ktime_get_real_seconds() + extra_delay;
1069 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
1072 exp->exp_last_request_time = new_time;
1074 /* exports may get disconnected from the chain even though the
1075 export has references, so we must keep the spin lock while
1076 manipulating the lists */
1077 spin_lock(&exp->exp_obd->obd_dev_lock);
1079 if (list_empty(&exp->exp_obd_chain_timed)) {
1080 /* this one is not timed */
1081 spin_unlock(&exp->exp_obd->obd_dev_lock);
1085 list_move_tail(&exp->exp_obd_chain_timed,
1086 &exp->exp_obd->obd_exports_timed);
1088 oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
1089 struct obd_export, exp_obd_chain_timed);
1090 oldest_time = oldest_exp->exp_last_request_time;
1091 spin_unlock(&exp->exp_obd->obd_dev_lock);
1093 if (exp->exp_obd->obd_recovering) {
1094 /* be nice to everyone during recovery */
1099 /* Note - racing to start/reset the obd_eviction timer is safe */
1100 if (exp->exp_obd->obd_eviction_timer == 0) {
1101 /* Check if the oldest entry is expired. */
1102 if (ktime_get_real_seconds() >
1103 oldest_time + PING_EVICT_TIMEOUT + extra_delay) {
1104 /* We need a second timer, in case the net was down and
1105 * it just came back. Since the pinger may skip every
1106 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
1107 * we better wait for 3.
1109 exp->exp_obd->obd_eviction_timer =
1110 ktime_get_real_seconds() + 3 * PING_INTERVAL;
1111 CDEBUG(D_HA, "%s: Think about evicting %s from %lld\n",
1112 exp->exp_obd->obd_name,
1113 obd_export_nid2str(oldest_exp), oldest_time);
1116 if (ktime_get_real_seconds() >
1117 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
1118 /* The evictor won't evict anyone who we've heard from
1119 * recently, so we don't have to check before we start
1122 if (!ping_evictor_wake(exp))
1123 exp->exp_obd->obd_eviction_timer = 0;
1131 * Sanity check request \a req.
1132 * Return 0 if all is ok, error code otherwise.
1134 static int ptlrpc_check_req(struct ptlrpc_request *req)
1136 struct obd_device *obd = req->rq_export->exp_obd;
1139 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
1140 req->rq_export->exp_conn_cnt)) {
1141 DEBUG_REQ(D_RPCTRACE, req,
1142 "DROPPING req from old connection %d < %d",
1143 lustre_msg_get_conn_cnt(req->rq_reqmsg),
1144 req->rq_export->exp_conn_cnt);
1147 if (unlikely(obd == NULL || obd->obd_fail)) {
1148 /* Failing over, don't handle any more reqs,
1149 * send error response instead. */
1150 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
1151 req, (obd != NULL) ? obd->obd_name : "unknown");
1153 } else if (lustre_msg_get_flags(req->rq_reqmsg) &
1154 (MSG_REPLAY | MSG_REQ_REPLAY_DONE) &&
1155 !obd->obd_recovering) {
1156 DEBUG_REQ(D_ERROR, req,
1157 "Invalid replay without recovery");
1158 class_fail_export(req->rq_export);
1160 } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 &&
1161 !obd->obd_recovering) {
1162 DEBUG_REQ(D_ERROR, req, "Invalid req with transno "
1163 "%llu without recovery",
1164 lustre_msg_get_transno(req->rq_reqmsg));
1165 class_fail_export(req->rq_export);
1169 if (unlikely(rc < 0)) {
1170 req->rq_status = rc;
1176 static void ptlrpc_at_set_timer(struct ptlrpc_service_part *svcpt)
1178 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1181 if (array->paa_count == 0) {
1182 del_timer(&svcpt->scp_at_timer);
1186 /* Set timer for closest deadline */
1187 next = (__s32)(array->paa_deadline - ktime_get_real_seconds() -
1190 ptlrpc_at_timer((unsigned long)svcpt);
1192 mod_timer(&svcpt->scp_at_timer, cfs_time_shift(next));
1193 CDEBUG(D_INFO, "armed %s at %+ds\n",
1194 svcpt->scp_service->srv_name, next);
1198 /* Add rpc to early reply check list */
1199 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
1201 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1202 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1203 struct ptlrpc_request *rq = NULL;
1209 if (req->rq_no_reply)
1212 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
1215 spin_lock(&svcpt->scp_at_lock);
1216 LASSERT(list_empty(&req->rq_timed_list));
1218 div_u64_rem(req->rq_deadline, array->paa_size, &index);
1219 if (array->paa_reqs_count[index] > 0) {
1220 /* latest rpcs will have the latest deadlines in the list,
1221 * so search backward. */
1222 list_for_each_entry_reverse(rq,
1223 &array->paa_reqs_array[index],
1225 if (req->rq_deadline >= rq->rq_deadline) {
1226 list_add(&req->rq_timed_list,
1227 &rq->rq_timed_list);
1233 /* Add the request at the head of the list */
1234 if (list_empty(&req->rq_timed_list))
1235 list_add(&req->rq_timed_list,
1236 &array->paa_reqs_array[index]);
1238 spin_lock(&req->rq_lock);
1239 req->rq_at_linked = 1;
1240 spin_unlock(&req->rq_lock);
1241 req->rq_at_index = index;
1242 array->paa_reqs_count[index]++;
1244 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
1245 array->paa_deadline = req->rq_deadline;
1246 ptlrpc_at_set_timer(svcpt);
1248 spin_unlock(&svcpt->scp_at_lock);
1254 ptlrpc_at_remove_timed(struct ptlrpc_request *req)
1256 struct ptlrpc_at_array *array;
1258 array = &req->rq_rqbd->rqbd_svcpt->scp_at_array;
1260 /* NB: must call with hold svcpt::scp_at_lock */
1261 LASSERT(!list_empty(&req->rq_timed_list));
1262 list_del_init(&req->rq_timed_list);
1264 spin_lock(&req->rq_lock);
1265 req->rq_at_linked = 0;
1266 spin_unlock(&req->rq_lock);
1268 array->paa_reqs_count[req->rq_at_index]--;
1273 * Attempt to extend the request deadline by sending an early reply to the
1276 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
1278 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1279 struct ptlrpc_request *reqcopy;
1280 struct lustre_msg *reqmsg;
1281 time64_t olddl = req->rq_deadline - ktime_get_real_seconds();
1287 if (CFS_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_RECONNECT)) {
1288 /* don't send early reply */
1292 /* deadline is when the client expects us to reply, margin is the
1293 difference between clients' and servers' expectations */
1294 DEBUG_REQ(D_ADAPTTO, req,
1295 "%ssending early reply (deadline %+llds, margin %+llds) for "
1296 "%d+%d", AT_OFF ? "AT off - not " : "",
1297 (s64)olddl, (s64)(olddl - at_get(&svcpt->scp_at_estimate)),
1298 at_get(&svcpt->scp_at_estimate), at_extra);
1304 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+llds), "
1305 "not sending early reply. Consider increasing "
1306 "at_early_margin (%d)?", (s64)olddl, at_early_margin);
1308 /* Return an error so we're not re-added to the timed list. */
1312 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
1313 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
1314 "but no AT support");
1318 if (req->rq_export &&
1319 lustre_msg_get_flags(req->rq_reqmsg) &
1320 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1321 struct obd_device *obd_exp = req->rq_export->exp_obd;
1323 /* During recovery, we don't want to send too many early
1324 * replies, but on the other hand we want to make sure the
1325 * client has enough time to resend if the rpc is lost. So
1326 * during the recovery period send at least 4 early replies,
1327 * spacing them every at_extra if we can. at_estimate should
1328 * always equal this fixed value during recovery.
1330 /* Don't account request processing time into AT history
1331 * during recovery, it is not service time we need but
1332 * includes also waiting time for recovering clients
1334 newdl = min_t(time64_t, at_extra,
1335 obd_exp->obd_recovery_timeout / 4) +
1336 ktime_get_real_seconds();
1338 /* We want to extend the request deadline by at_extra seconds,
1339 * so we set our service estimate to reflect how much time has
1340 * passed since this request arrived plus an additional
1341 * at_extra seconds. The client will calculate the new deadline
1342 * based on this service estimate (plus some additional time to
1343 * account for network latency). See ptlrpc_at_recv_early_reply
1345 at_measured(&svcpt->scp_at_estimate, at_extra +
1346 ktime_get_real_seconds() -
1347 req->rq_arrival_time.tv_sec);
1348 newdl = req->rq_arrival_time.tv_sec +
1349 at_get(&svcpt->scp_at_estimate);
1352 /* Check to see if we've actually increased the deadline -
1353 * we may be past adaptive_max */
1354 if (req->rq_deadline >= newdl) {
1355 DEBUG_REQ(D_WARNING, req, "Couldn't add any time (%lld/%lld), not sending early reply\n",
1356 (s64)olddl, (s64)(newdl - ktime_get_real_seconds()));
1360 reqcopy = ptlrpc_request_cache_alloc(GFP_NOFS);
1361 if (reqcopy == NULL)
1363 OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1365 GOTO(out_free, rc = -ENOMEM);
1368 reqcopy->rq_reply_state = NULL;
1369 reqcopy->rq_rep_swab_mask = 0;
1370 reqcopy->rq_pack_bulk = 0;
1371 reqcopy->rq_pack_udesc = 0;
1372 reqcopy->rq_packed_final = 0;
1373 sptlrpc_svc_ctx_addref(reqcopy);
1374 /* We only need the reqmsg for the magic */
1375 reqcopy->rq_reqmsg = reqmsg;
1376 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1379 * tgt_brw_read() and tgt_brw_write() may have decided not to reply.
1380 * Without this check, we would fail the rq_no_reply assertion in
1381 * ptlrpc_send_reply().
1383 if (reqcopy->rq_no_reply)
1384 GOTO(out, rc = -ETIMEDOUT);
1386 LASSERT(atomic_read(&req->rq_refcount));
1387 /** if it is last refcount then early reply isn't needed */
1388 if (atomic_read(&req->rq_refcount) == 1) {
1389 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
1390 "abort sending early reply\n");
1391 GOTO(out, rc = -EINVAL);
1394 /* Connection ref */
1395 reqcopy->rq_export = class_conn2export(
1396 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1397 if (reqcopy->rq_export == NULL)
1398 GOTO(out, rc = -ENODEV);
1401 class_export_rpc_inc(reqcopy->rq_export);
1402 if (reqcopy->rq_export->exp_obd &&
1403 reqcopy->rq_export->exp_obd->obd_fail)
1404 GOTO(out_put, rc = -ENODEV);
1406 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1410 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1413 /* Adjust our own deadline to what we told the client */
1414 req->rq_deadline = newdl;
1415 req->rq_early_count++; /* number sent, server side */
1417 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1420 /* Free the (early) reply state from lustre_pack_reply.
1421 (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
1422 ptlrpc_req_drop_rs(reqcopy);
1425 class_export_rpc_dec(reqcopy->rq_export);
1426 class_export_put(reqcopy->rq_export);
1428 sptlrpc_svc_ctx_decref(reqcopy);
1429 OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1431 ptlrpc_request_cache_free(reqcopy);
1435 /* Send early replies to everybody expiring within at_early_margin
1436 asking for at_extra time */
1437 static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1439 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1440 struct ptlrpc_request *rq, *n;
1441 struct list_head work_list;
1444 time64_t now = ktime_get_real_seconds();
1445 cfs_duration_t delay;
1446 int first, counter = 0;
1449 spin_lock(&svcpt->scp_at_lock);
1450 if (svcpt->scp_at_check == 0) {
1451 spin_unlock(&svcpt->scp_at_lock);
1454 delay = cfs_time_sub(cfs_time_current(), svcpt->scp_at_checktime);
1455 svcpt->scp_at_check = 0;
1457 if (array->paa_count == 0) {
1458 spin_unlock(&svcpt->scp_at_lock);
1462 /* The timer went off, but maybe the nearest rpc already completed. */
1463 first = array->paa_deadline - now;
1464 if (first > at_early_margin) {
1465 /* We've still got plenty of time. Reset the timer. */
1466 ptlrpc_at_set_timer(svcpt);
1467 spin_unlock(&svcpt->scp_at_lock);
1471 /* We're close to a timeout, and we don't know how much longer the
1472 server will take. Send early replies to everyone expiring soon. */
1473 INIT_LIST_HEAD(&work_list);
1475 div_u64_rem(array->paa_deadline, array->paa_size, &index);
1476 count = array->paa_count;
1478 count -= array->paa_reqs_count[index];
1479 list_for_each_entry_safe(rq, n,
1480 &array->paa_reqs_array[index],
1482 if (rq->rq_deadline > now + at_early_margin) {
1483 /* update the earliest deadline */
1484 if (deadline == -1 ||
1485 rq->rq_deadline < deadline)
1486 deadline = rq->rq_deadline;
1490 ptlrpc_at_remove_timed(rq);
1492 * ptlrpc_server_drop_request() may drop
1493 * refcount to 0 already. Let's check this and
1494 * don't add entry to work_list
1496 if (likely(atomic_inc_not_zero(&rq->rq_refcount)))
1497 list_add(&rq->rq_timed_list, &work_list);
1501 if (++index >= array->paa_size)
1504 array->paa_deadline = deadline;
1505 /* we have a new earliest deadline, restart the timer */
1506 ptlrpc_at_set_timer(svcpt);
1508 spin_unlock(&svcpt->scp_at_lock);
1510 CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
1511 "replies\n", first, at_extra, counter);
1513 /* We're already past request deadlines before we even get a
1514 chance to send early replies */
1515 LCONSOLE_WARN("%s: This server is not able to keep up with "
1516 "request traffic (cpu-bound).\n",
1517 svcpt->scp_service->srv_name);
1518 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, delay=%ld(jiff)\n",
1519 counter, svcpt->scp_nreqs_incoming,
1520 svcpt->scp_nreqs_active,
1521 at_get(&svcpt->scp_at_estimate), delay);
1524 /* we took additional refcount so entries can't be deleted from list, no
1525 * locking is needed */
1526 while (!list_empty(&work_list)) {
1527 rq = list_entry(work_list.next, struct ptlrpc_request,
1529 list_del_init(&rq->rq_timed_list);
1531 if (ptlrpc_at_send_early_reply(rq) == 0)
1532 ptlrpc_at_add_timed(rq);
1534 ptlrpc_server_drop_request(rq);
1537 RETURN(1); /* return "did_something" for liblustre */
1540 /* Check if we are already handling earlier incarnation of this request.
1541 * Called under &req->rq_export->exp_rpc_lock locked */
1542 static int ptlrpc_server_check_resend_in_progress(struct ptlrpc_request *req)
1544 struct ptlrpc_request *tmp = NULL;
1546 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ||
1547 (atomic_read(&req->rq_export->exp_rpc_count) == 0))
1550 /* bulk request are aborted upon reconnect, don't try to
1552 if (req->rq_bulk_write || req->rq_bulk_read)
1555 /* This list should not be longer than max_requests in
1556 * flights on the client, so it is not all that long.
1557 * Also we only hit this codepath in case of a resent
1558 * request which makes it even more rarely hit */
1559 list_for_each_entry(tmp, &req->rq_export->exp_reg_rpcs,
1561 /* Found duplicate one */
1562 if (tmp->rq_xid == req->rq_xid)
1565 list_for_each_entry(tmp, &req->rq_export->exp_hp_rpcs,
1567 /* Found duplicate one */
1568 if (tmp->rq_xid == req->rq_xid)
1574 DEBUG_REQ(D_HA, req, "Found duplicate req in processing");
1575 DEBUG_REQ(D_HA, tmp, "Request being processed");
1580 * Check if a request should be assigned with a high priority.
1582 * \retval < 0: error occurred
1583 * 0: normal RPC request
1584 * +1: high priority request
1586 static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt,
1587 struct ptlrpc_request *req)
1592 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL) {
1593 rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
1600 if (req->rq_export != NULL && req->rq_ops != NULL) {
1601 /* Perform request specific check. We should do this
1602 * check before the request is added into exp_hp_rpcs
1603 * list otherwise it may hit swab race at LU-1044. */
1604 if (req->rq_ops->hpreq_check != NULL) {
1605 rc = req->rq_ops->hpreq_check(req);
1606 if (rc == -ESTALE) {
1607 req->rq_status = rc;
1610 /** can only return error,
1611 * 0 for normal request,
1612 * or 1 for high priority request */
1620 /** Remove the request from the export list. */
1621 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req)
1624 if (req->rq_export) {
1625 /* refresh lock timeout again so that client has more
1626 * room to send lock cancel RPC. */
1627 if (req->rq_ops && req->rq_ops->hpreq_fini)
1628 req->rq_ops->hpreq_fini(req);
1630 spin_lock_bh(&req->rq_export->exp_rpc_lock);
1631 list_del_init(&req->rq_exp_list);
1632 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1637 static int ptlrpc_hpreq_check(struct ptlrpc_request *req)
1642 static struct ptlrpc_hpreq_ops ptlrpc_hpreq_common = {
1643 .hpreq_check = ptlrpc_hpreq_check,
1646 /* Hi-Priority RPC check by RPC operation code. */
1647 int ptlrpc_hpreq_handler(struct ptlrpc_request *req)
1649 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1651 /* Check for export to let only reconnects for not yet evicted
1652 * export to become a HP rpc. */
1653 if ((req->rq_export != NULL) &&
1654 (opc == OBD_PING || opc == MDS_CONNECT || opc == OST_CONNECT))
1655 req->rq_ops = &ptlrpc_hpreq_common;
1659 EXPORT_SYMBOL(ptlrpc_hpreq_handler);
1661 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1662 struct ptlrpc_request *req)
1668 rc = ptlrpc_server_hpreq_init(svcpt, req);
1673 ptlrpc_nrs_req_initialize(svcpt, req, hp);
1675 if (req->rq_export != NULL) {
1676 struct obd_export *exp = req->rq_export;
1678 /* do search for duplicated xid and the adding to the list
1680 spin_lock_bh(&exp->exp_rpc_lock);
1681 rc = ptlrpc_server_check_resend_in_progress(req);
1683 spin_unlock_bh(&exp->exp_rpc_lock);
1685 ptlrpc_nrs_req_finalize(req);
1689 if (hp || req->rq_ops != NULL)
1690 list_add(&req->rq_exp_list, &exp->exp_hp_rpcs);
1692 list_add(&req->rq_exp_list, &exp->exp_reg_rpcs);
1693 spin_unlock_bh(&exp->exp_rpc_lock);
1696 /* the current thread is not the processing thread for this request
1697 * since that, but request is in exp_hp_list and can be find there.
1698 * Remove all relations between request and old thread. */
1699 req->rq_svc_thread->t_env->le_ses = NULL;
1700 req->rq_svc_thread = NULL;
1701 req->rq_session.lc_thread = NULL;
1703 ptlrpc_nrs_req_add(svcpt, req, hp);
1709 * Allow to handle high priority request
1710 * User can call it w/o any lock but need to hold
1711 * ptlrpc_service_part::scp_req_lock to get reliable result
1713 static bool ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1716 int running = svcpt->scp_nthrs_running;
1718 if (!nrs_svcpt_has_hp(svcpt))
1724 if (ptlrpc_nrs_req_throttling_nolock(svcpt, true))
1727 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1728 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1729 /* leave just 1 thread for normal RPCs */
1730 running = PTLRPC_NTHRS_INIT;
1731 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1735 if (svcpt->scp_nreqs_active >= running - 1)
1738 if (svcpt->scp_nhreqs_active == 0)
1741 return !ptlrpc_nrs_req_pending_nolock(svcpt, false) ||
1742 svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1745 static bool ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1748 return ptlrpc_server_allow_high(svcpt, force) &&
1749 ptlrpc_nrs_req_pending_nolock(svcpt, true);
1753 * Only allow normal priority requests on a service that has a high-priority
1754 * queue if forced (i.e. cleanup), if there are other high priority requests
1755 * already being processed (i.e. those threads can service more high-priority
1756 * requests), or if there are enough idle threads that a later thread can do
1757 * a high priority request.
1758 * User can call it w/o any lock but need to hold
1759 * ptlrpc_service_part::scp_req_lock to get reliable result
1761 static bool ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1764 int running = svcpt->scp_nthrs_running;
1765 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1766 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1767 /* leave just 1 thread for normal RPCs */
1768 running = PTLRPC_NTHRS_INIT;
1769 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1776 if (ptlrpc_nrs_req_throttling_nolock(svcpt, false))
1779 if (svcpt->scp_nreqs_active < running - 2)
1782 if (svcpt->scp_nreqs_active >= running - 1)
1785 return svcpt->scp_nhreqs_active > 0 || !nrs_svcpt_has_hp(svcpt);
1788 static bool ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1791 return ptlrpc_server_allow_normal(svcpt, force) &&
1792 ptlrpc_nrs_req_pending_nolock(svcpt, false);
1796 * Returns true if there are requests available in incoming
1797 * request queue for processing and it is allowed to fetch them.
1798 * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1799 * to get reliable result
1800 * \see ptlrpc_server_allow_normal
1801 * \see ptlrpc_server_allow high
1804 ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt, bool force)
1806 return ptlrpc_server_high_pending(svcpt, force) ||
1807 ptlrpc_server_normal_pending(svcpt, force);
1811 * Fetch a request for processing from queue of unprocessed requests.
1812 * Favors high-priority requests.
1813 * Returns a pointer to fetched request.
1815 static struct ptlrpc_request *
1816 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force)
1818 struct ptlrpc_request *req = NULL;
1821 spin_lock(&svcpt->scp_req_lock);
1823 if (ptlrpc_server_high_pending(svcpt, force)) {
1824 req = ptlrpc_nrs_req_get_nolock(svcpt, true, force);
1826 svcpt->scp_hreq_count++;
1831 if (ptlrpc_server_normal_pending(svcpt, force)) {
1832 req = ptlrpc_nrs_req_get_nolock(svcpt, false, force);
1834 svcpt->scp_hreq_count = 0;
1839 spin_unlock(&svcpt->scp_req_lock);
1843 svcpt->scp_nreqs_active++;
1845 svcpt->scp_nhreqs_active++;
1847 spin_unlock(&svcpt->scp_req_lock);
1849 if (likely(req->rq_export))
1850 class_export_rpc_inc(req->rq_export);
1856 * Handle freshly incoming reqs, add to timed early reply list,
1857 * pass on to regular request queue.
1858 * All incoming requests pass through here before getting into
1859 * ptlrpc_server_handle_req later on.
1862 ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt,
1863 struct ptlrpc_thread *thread)
1865 struct ptlrpc_service *svc = svcpt->scp_service;
1866 struct ptlrpc_request *req;
1871 spin_lock(&svcpt->scp_lock);
1872 if (list_empty(&svcpt->scp_req_incoming)) {
1873 spin_unlock(&svcpt->scp_lock);
1877 req = list_entry(svcpt->scp_req_incoming.next,
1878 struct ptlrpc_request, rq_list);
1879 list_del_init(&req->rq_list);
1880 svcpt->scp_nreqs_incoming--;
1881 /* Consider this still a "queued" request as far as stats are
1883 spin_unlock(&svcpt->scp_lock);
1885 /* go through security check/transform */
1886 rc = sptlrpc_svc_unwrap_request(req);
1890 case SECSVC_COMPLETE:
1891 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1900 * for null-flavored rpc, msg has been unpacked by sptlrpc, although
1901 * redo it wouldn't be harmful.
1903 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
1904 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
1906 CERROR("error unpacking request: ptl %d from %s "
1907 "x%llu\n", svc->srv_req_portal,
1908 libcfs_id2str(req->rq_peer), req->rq_xid);
1913 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1915 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
1916 "%llu\n", svc->srv_req_portal,
1917 libcfs_id2str(req->rq_peer), req->rq_xid);
1921 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
1922 lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
1923 CERROR("drop incoming rpc opc %u, x%llu\n",
1924 cfs_fail_val, req->rq_xid);
1929 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1930 CERROR("wrong packet type received (type=%u) from %s\n",
1931 lustre_msg_get_type(req->rq_reqmsg),
1932 libcfs_id2str(req->rq_peer));
1936 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1940 req->rq_bulk_write = 1;
1944 case MGS_CONFIG_READ:
1945 req->rq_bulk_read = 1;
1949 CDEBUG(D_RPCTRACE, "got req x%llu\n", req->rq_xid);
1951 req->rq_export = class_conn2export(
1952 lustre_msg_get_handle(req->rq_reqmsg));
1953 if (req->rq_export) {
1954 rc = ptlrpc_check_req(req);
1956 rc = sptlrpc_target_export_check(req->rq_export, req);
1958 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
1959 "illegal security flavor,");
1964 ptlrpc_update_export_timer(req->rq_export, 0);
1967 /* req_in handling should/must be fast */
1968 if (ktime_get_real_seconds() - req->rq_arrival_time.tv_sec > 5)
1969 DEBUG_REQ(D_WARNING, req, "Slow req_in handling %llds",
1970 (s64)(ktime_get_real_seconds() -
1971 req->rq_arrival_time.tv_sec));
1973 /* Set rpc server deadline and add it to the timed list */
1974 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1975 MSGHDR_AT_SUPPORT) ?
1976 /* The max time the client expects us to take */
1977 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1979 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1980 if (unlikely(deadline == 0)) {
1981 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1985 /* Skip early reply */
1986 if (OBD_FAIL_PRECHECK(OBD_FAIL_MDS_RESEND))
1987 req->rq_deadline += obd_timeout;
1989 req->rq_svc_thread = thread;
1990 if (thread != NULL) {
1991 /* initialize request session, it is needed for request
1992 * processing by target */
1993 rc = lu_context_init(&req->rq_session, LCT_SERVER_SESSION |
1996 CERROR("%s: failure to initialize session: rc = %d\n",
1997 thread->t_name, rc);
2000 req->rq_session.lc_thread = thread;
2001 lu_context_enter(&req->rq_session);
2002 thread->t_env->le_ses = &req->rq_session;
2005 ptlrpc_at_add_timed(req);
2007 /* Move it over to the request processing queue */
2008 rc = ptlrpc_server_request_add(svcpt, req);
2012 wake_up(&svcpt->scp_waitq);
2016 ptlrpc_server_finish_request(svcpt, req);
2022 * Main incoming request handling logic.
2023 * Calls handler function from service to do actual processing.
2026 ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
2027 struct ptlrpc_thread *thread)
2029 struct ptlrpc_service *svc = svcpt->scp_service;
2030 struct ptlrpc_request *request;
2040 request = ptlrpc_server_request_get(svcpt, false);
2041 if (request == NULL)
2044 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
2045 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
2046 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
2047 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
2049 if (unlikely(fail_opc)) {
2050 if (request->rq_export && request->rq_ops)
2051 OBD_FAIL_TIMEOUT(fail_opc, 4);
2054 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
2056 if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
2057 libcfs_debug_dumplog();
2059 work_start = ktime_get_real();
2060 arrived = timespec64_to_ktime(request->rq_arrival_time);
2061 timediff_usecs = ktime_us_delta(work_start, arrived);
2062 if (likely(svc->srv_stats != NULL)) {
2063 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
2065 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
2066 svcpt->scp_nreqs_incoming);
2067 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
2068 svcpt->scp_nreqs_active);
2069 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
2070 at_get(&svcpt->scp_at_estimate));
2073 if (likely(request->rq_export)) {
2074 if (unlikely(ptlrpc_check_req(request)))
2076 ptlrpc_update_export_timer(request->rq_export,
2077 timediff_usecs / (USEC_PER_SEC / 2));
2080 /* Discard requests queued for longer than the deadline.
2081 The deadline is increased if we send an early reply. */
2082 if (ktime_get_real_seconds() > request->rq_deadline) {
2083 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s: deadline %lld:%llds ago\n",
2084 libcfs_id2str(request->rq_peer),
2085 request->rq_deadline -
2086 request->rq_arrival_time.tv_sec,
2087 ktime_get_real_seconds() - request->rq_deadline);
2091 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
2092 "%s:%s+%d:%d:x%llu:%s:%d\n", current_comm(),
2093 (request->rq_export ?
2094 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2095 (request->rq_export ?
2096 atomic_read(&request->rq_export->exp_refcount) : -99),
2097 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
2098 libcfs_id2str(request->rq_peer),
2099 lustre_msg_get_opc(request->rq_reqmsg));
2101 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
2102 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
2104 CDEBUG(D_NET, "got req %llu\n", request->rq_xid);
2106 /* re-assign request and sesson thread to the current one */
2107 request->rq_svc_thread = thread;
2108 if (thread != NULL) {
2109 LASSERT(request->rq_session.lc_thread == NULL);
2110 request->rq_session.lc_thread = thread;
2111 thread->t_env->le_ses = &request->rq_session;
2113 svc->srv_ops.so_req_handler(request);
2115 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
2118 if (unlikely(ktime_get_real_seconds() > request->rq_deadline)) {
2119 DEBUG_REQ(D_WARNING, request, "Request took longer than estimated (%lld:%llds); client may timeout.",
2120 request->rq_deadline -
2121 request->rq_arrival_time.tv_sec,
2122 ktime_get_real_seconds() - request->rq_deadline);
2125 work_end = ktime_get_real();
2126 timediff_usecs = ktime_us_delta(work_end, work_start);
2127 arrived_usecs = ktime_us_delta(work_end, arrived);
2128 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc %s:%s+%d:%d:x%llu:%s:%d Request procesed in %lldus (%lldus total) trans %llu rc %d/%d\n",
2130 (request->rq_export ?
2131 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2132 (request->rq_export ?
2133 atomic_read(&request->rq_export->exp_refcount) : -99),
2134 lustre_msg_get_status(request->rq_reqmsg),
2136 libcfs_id2str(request->rq_peer),
2137 lustre_msg_get_opc(request->rq_reqmsg),
2140 (request->rq_repmsg ?
2141 lustre_msg_get_transno(request->rq_repmsg) :
2142 request->rq_transno),
2144 (request->rq_repmsg ?
2145 lustre_msg_get_status(request->rq_repmsg) : -999));
2146 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
2147 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
2148 int opc = opcode_offset(op);
2149 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
2150 LASSERT(opc < LUSTRE_MAX_OPCODES);
2151 lprocfs_counter_add(svc->srv_stats,
2152 opc + EXTRA_MAX_OPCODES,
2156 if (unlikely(request->rq_early_count)) {
2157 DEBUG_REQ(D_ADAPTTO, request,
2158 "sent %d early replies before finishing in %llds",
2159 request->rq_early_count,
2160 arrived_usecs / USEC_PER_SEC);
2163 ptlrpc_server_finish_active_request(svcpt, request);
2169 * An internal function to process a single reply state object.
2172 ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
2174 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
2175 struct ptlrpc_service *svc = svcpt->scp_service;
2176 struct obd_export *exp;
2181 exp = rs->rs_export;
2183 LASSERT(rs->rs_difficult);
2184 LASSERT(rs->rs_scheduled);
2185 LASSERT(list_empty(&rs->rs_list));
2187 /* The disk commit callback holds exp_uncommitted_replies_lock while it
2188 * iterates over newly committed replies, removing them from
2189 * exp_uncommitted_replies. It then drops this lock and schedules the
2190 * replies it found for handling here.
2192 * We can avoid contention for exp_uncommitted_replies_lock between the
2193 * HRT threads and further commit callbacks by checking rs_committed
2194 * which is set in the commit callback while it holds both
2195 * rs_lock and exp_uncommitted_reples.
2197 * If we see rs_committed clear, the commit callback _may_ not have
2198 * handled this reply yet and we race with it to grab
2199 * exp_uncommitted_replies_lock before removing the reply from
2200 * exp_uncommitted_replies. Note that if we lose the race and the
2201 * reply has already been removed, list_del_init() is a noop.
2203 * If we see rs_committed set, we know the commit callback is handling,
2204 * or has handled this reply since store reordering might allow us to
2205 * see rs_committed set out of sequence. But since this is done
2206 * holding rs_lock, we can be sure it has all completed once we hold
2207 * rs_lock, which we do right next.
2209 if (!rs->rs_committed) {
2210 /* if rs was commited, no need to convert locks, don't check
2211 * rs_committed here because rs may never be added into
2212 * exp_uncommitted_replies and this flag never be set, see
2213 * target_send_reply() */
2214 if (rs->rs_convert_lock &&
2215 rs->rs_transno > exp->exp_last_committed) {
2216 struct ldlm_lock *lock;
2217 struct ldlm_lock *ack_locks[RS_MAX_LOCKS] = { NULL };
2219 spin_lock(&rs->rs_lock);
2220 if (rs->rs_convert_lock &&
2221 rs->rs_transno > exp->exp_last_committed) {
2222 nlocks = rs->rs_nlocks;
2223 while (nlocks-- > 0) {
2225 * NB don't assume rs is always handled
2226 * by the same service thread (see
2227 * ptlrpc_hr_select, so REP-ACK hr may
2228 * race with trans commit, while the
2229 * latter will release locks, get locks
2230 * here early to convert to COS mode
2233 lock = ldlm_handle2lock(
2234 &rs->rs_locks[nlocks]);
2236 ack_locks[nlocks] = lock;
2237 rs->rs_modes[nlocks] = LCK_COS;
2239 nlocks = rs->rs_nlocks;
2240 rs->rs_convert_lock = 0;
2241 /* clear rs_scheduled so that commit callback
2242 * can schedule again */
2243 rs->rs_scheduled = 0;
2244 spin_unlock(&rs->rs_lock);
2246 while (nlocks-- > 0) {
2247 lock = ack_locks[nlocks];
2248 ldlm_lock_downgrade(lock, LCK_COS);
2249 LDLM_LOCK_PUT(lock);
2253 spin_unlock(&rs->rs_lock);
2256 spin_lock(&exp->exp_uncommitted_replies_lock);
2257 list_del_init(&rs->rs_obd_list);
2258 spin_unlock(&exp->exp_uncommitted_replies_lock);
2261 spin_lock(&exp->exp_lock);
2262 /* Noop if removed already */
2263 list_del_init(&rs->rs_exp_list);
2264 spin_unlock(&exp->exp_lock);
2266 spin_lock(&rs->rs_lock);
2268 been_handled = rs->rs_handled;
2271 nlocks = rs->rs_nlocks; /* atomic "steal", but */
2272 rs->rs_nlocks = 0; /* locks still on rs_locks! */
2274 if (nlocks == 0 && !been_handled) {
2275 /* If we see this, we should already have seen the warning
2276 * in mds_steal_ack_locks() */
2277 CDEBUG(D_HA, "All locks stolen from rs %p x%lld.t%lld"
2280 rs->rs_xid, rs->rs_transno, rs->rs_opc,
2281 libcfs_nid2str(exp->exp_connection->c_peer.nid));
2284 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
2285 spin_unlock(&rs->rs_lock);
2287 if (!been_handled && rs->rs_on_net) {
2288 LNetMDUnlink(rs->rs_md_h);
2289 /* Ignore return code; we're racing with completion */
2292 while (nlocks-- > 0)
2293 ldlm_lock_decref(&rs->rs_locks[nlocks],
2294 rs->rs_modes[nlocks]);
2296 spin_lock(&rs->rs_lock);
2299 rs->rs_scheduled = 0;
2300 rs->rs_convert_lock = 0;
2302 if (!rs->rs_on_net) {
2304 spin_unlock(&rs->rs_lock);
2306 class_export_put (exp);
2307 rs->rs_export = NULL;
2308 ptlrpc_rs_decref(rs);
2309 if (atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
2310 svc->srv_is_stopping)
2311 wake_up_all(&svcpt->scp_waitq);
2315 /* still on the net; callback will schedule */
2316 spin_unlock(&rs->rs_lock);
2322 ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
2324 int avail = svcpt->scp_nrqbds_posted;
2325 int low_water = test_req_buffer_pressure ? 0 :
2326 svcpt->scp_service->srv_nbuf_per_group / 2;
2328 /* NB I'm not locking; just looking. */
2330 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
2331 * allowed the request history to grow out of control. We could put a
2332 * sanity check on that here and cull some history if we need the
2335 if (avail <= low_water)
2336 ptlrpc_grow_req_bufs(svcpt, 1);
2338 if (svcpt->scp_service->srv_stats) {
2339 lprocfs_counter_add(svcpt->scp_service->srv_stats,
2340 PTLRPC_REQBUF_AVAIL_CNTR, avail);
2345 ptlrpc_retry_rqbds(void *arg)
2347 struct ptlrpc_service_part *svcpt = (struct ptlrpc_service_part *)arg;
2349 svcpt->scp_rqbd_timeout = 0;
2354 ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
2356 return svcpt->scp_nreqs_active <
2357 svcpt->scp_nthrs_running - 1 -
2358 (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL);
2362 * allowed to create more threads
2363 * user can call it w/o any lock but need to hold
2364 * ptlrpc_service_part::scp_lock to get reliable result
2367 ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
2369 return svcpt->scp_nthrs_running +
2370 svcpt->scp_nthrs_starting <
2371 svcpt->scp_service->srv_nthrs_cpt_limit;
2375 * too many requests and allowed to create more threads
2378 ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
2380 return !ptlrpc_threads_enough(svcpt) &&
2381 ptlrpc_threads_increasable(svcpt);
2385 ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2387 return thread_is_stopping(thread) ||
2388 thread->t_svcpt->scp_service->srv_is_stopping;
2392 ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
2394 return !list_empty(&svcpt->scp_rqbd_idle) &&
2395 svcpt->scp_rqbd_timeout == 0;
2399 ptlrpc_at_check(struct ptlrpc_service_part *svcpt)
2401 return svcpt->scp_at_check;
2405 * requests wait on preprocessing
2406 * user can call it w/o any lock but need to hold
2407 * ptlrpc_service_part::scp_lock to get reliable result
2410 ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
2412 return !list_empty(&svcpt->scp_req_incoming);
2415 static __attribute__((__noinline__)) int
2416 ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
2417 struct ptlrpc_thread *thread)
2419 /* Don't exit while there are replies to be handled */
2420 struct l_wait_info lwi = LWI_TIMEOUT(svcpt->scp_rqbd_timeout,
2421 ptlrpc_retry_rqbds, svcpt);
2423 lc_watchdog_disable(thread->t_watchdog);
2427 l_wait_event_exclusive_head(svcpt->scp_waitq,
2428 ptlrpc_thread_stopping(thread) ||
2429 ptlrpc_server_request_incoming(svcpt) ||
2430 ptlrpc_server_request_pending(svcpt, false) ||
2431 ptlrpc_rqbd_pending(svcpt) ||
2432 ptlrpc_at_check(svcpt), &lwi);
2434 if (ptlrpc_thread_stopping(thread))
2437 lc_watchdog_touch(thread->t_watchdog,
2438 ptlrpc_server_get_timeout(svcpt));
2443 * Main thread body for service threads.
2444 * Waits in a loop waiting for new requests to process to appear.
2445 * Every time an incoming requests is added to its queue, a waitq
2446 * is woken up and one of the threads will handle it.
2448 static int ptlrpc_main(void *arg)
2450 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
2451 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2452 struct ptlrpc_service *svc = svcpt->scp_service;
2453 struct ptlrpc_reply_state *rs;
2454 struct group_info *ginfo = NULL;
2456 int counter = 0, rc = 0;
2459 thread->t_pid = current_pid();
2460 unshare_fs_struct();
2462 /* NB: we will call cfs_cpt_bind() for all threads, because we
2463 * might want to run lustre server only on a subset of system CPUs,
2464 * in that case ->scp_cpt is CFS_CPT_ANY */
2465 rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
2467 CWARN("%s: failed to bind %s on CPT %d\n",
2468 svc->srv_name, thread->t_name, svcpt->scp_cpt);
2471 ginfo = groups_alloc(0);
2477 set_current_groups(ginfo);
2478 put_group_info(ginfo);
2480 if (svc->srv_ops.so_thr_init != NULL) {
2481 rc = svc->srv_ops.so_thr_init(thread);
2492 rc = lu_context_init(&env->le_ctx,
2493 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2497 thread->t_env = env;
2498 env->le_ctx.lc_thread = thread;
2499 env->le_ctx.lc_cookie = 0x6;
2501 while (!list_empty(&svcpt->scp_rqbd_idle)) {
2502 rc = ptlrpc_server_post_idle_rqbds(svcpt);
2506 CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
2507 svc->srv_name, svcpt->scp_cpt, rc);
2511 /* Alloc reply state structure for this one */
2512 OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2518 spin_lock(&svcpt->scp_lock);
2520 LASSERT(thread_is_starting(thread));
2521 thread_clear_flags(thread, SVC_STARTING);
2523 LASSERT(svcpt->scp_nthrs_starting == 1);
2524 svcpt->scp_nthrs_starting--;
2526 /* SVC_STOPPING may already be set here if someone else is trying
2527 * to stop the service while this new thread has been dynamically
2528 * forked. We still set SVC_RUNNING to let our creator know that
2529 * we are now running, however we will exit as soon as possible */
2530 thread_add_flags(thread, SVC_RUNNING);
2531 svcpt->scp_nthrs_running++;
2532 spin_unlock(&svcpt->scp_lock);
2534 /* wake up our creator in case he's still waiting. */
2535 wake_up(&thread->t_ctl_waitq);
2537 thread->t_watchdog = lc_watchdog_add(ptlrpc_server_get_timeout(svcpt),
2540 spin_lock(&svcpt->scp_rep_lock);
2541 list_add(&rs->rs_list, &svcpt->scp_rep_idle);
2542 wake_up(&svcpt->scp_rep_waitq);
2543 spin_unlock(&svcpt->scp_rep_lock);
2545 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2546 svcpt->scp_nthrs_running);
2548 /* XXX maintain a list of all managed devices: insert here */
2549 while (!ptlrpc_thread_stopping(thread)) {
2550 if (ptlrpc_wait_event(svcpt, thread))
2553 ptlrpc_check_rqbd_pool(svcpt);
2555 if (ptlrpc_threads_need_create(svcpt)) {
2556 /* Ignore return code - we tried... */
2557 ptlrpc_start_thread(svcpt, 0);
2560 /* reset le_ses to initial state */
2562 /* Process all incoming reqs before handling any */
2563 if (ptlrpc_server_request_incoming(svcpt)) {
2564 lu_context_enter(&env->le_ctx);
2565 ptlrpc_server_handle_req_in(svcpt, thread);
2566 lu_context_exit(&env->le_ctx);
2568 /* but limit ourselves in case of flood */
2569 if (counter++ < 100)
2574 if (ptlrpc_at_check(svcpt))
2575 ptlrpc_at_check_timed(svcpt);
2577 if (ptlrpc_server_request_pending(svcpt, false)) {
2578 lu_context_enter(&env->le_ctx);
2579 ptlrpc_server_handle_request(svcpt, thread);
2580 lu_context_exit(&env->le_ctx);
2583 if (ptlrpc_rqbd_pending(svcpt) &&
2584 ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
2585 /* I just failed to repost request buffers.
2586 * Wait for a timeout (unless something else
2587 * happens) before I try again */
2588 svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
2589 CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
2590 svcpt->scp_nrqbds_posted);
2594 lc_watchdog_delete(thread->t_watchdog);
2595 thread->t_watchdog = NULL;
2599 * deconstruct service specific state created by ptlrpc_start_thread()
2601 if (svc->srv_ops.so_thr_done != NULL)
2602 svc->srv_ops.so_thr_done(thread);
2605 lu_context_fini(&env->le_ctx);
2609 CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n",
2610 thread, thread->t_pid, thread->t_id, rc);
2612 spin_lock(&svcpt->scp_lock);
2613 if (thread_test_and_clear_flags(thread, SVC_STARTING))
2614 svcpt->scp_nthrs_starting--;
2616 if (thread_test_and_clear_flags(thread, SVC_RUNNING)) {
2617 /* must know immediately */
2618 svcpt->scp_nthrs_running--;
2622 thread_add_flags(thread, SVC_STOPPED);
2624 wake_up(&thread->t_ctl_waitq);
2625 spin_unlock(&svcpt->scp_lock);
2630 static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
2631 struct list_head *replies)
2635 spin_lock(&hrt->hrt_lock);
2637 list_splice_init(&hrt->hrt_queue, replies);
2638 result = ptlrpc_hr.hr_stopping || !list_empty(replies);
2640 spin_unlock(&hrt->hrt_lock);
2645 * Main body of "handle reply" function.
2646 * It processes acked reply states
2648 static int ptlrpc_hr_main(void *arg)
2650 struct ptlrpc_hr_thread *hrt = (struct ptlrpc_hr_thread *)arg;
2651 struct ptlrpc_hr_partition *hrp = hrt->hrt_partition;
2652 struct list_head replies;
2655 INIT_LIST_HEAD(&replies);
2656 unshare_fs_struct();
2658 rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
2660 char threadname[20];
2662 snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
2663 hrp->hrp_cpt, hrt->hrt_id);
2664 CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
2665 threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
2668 atomic_inc(&hrp->hrp_nstarted);
2669 wake_up(&ptlrpc_hr.hr_waitq);
2671 while (!ptlrpc_hr.hr_stopping) {
2672 l_wait_condition(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
2674 while (!list_empty(&replies)) {
2675 struct ptlrpc_reply_state *rs;
2677 rs = list_entry(replies.prev,
2678 struct ptlrpc_reply_state,
2680 list_del_init(&rs->rs_list);
2681 ptlrpc_handle_rs(rs);
2685 atomic_inc(&hrp->hrp_nstopped);
2686 wake_up(&ptlrpc_hr.hr_waitq);
2691 static void ptlrpc_stop_hr_threads(void)
2693 struct ptlrpc_hr_partition *hrp;
2697 ptlrpc_hr.hr_stopping = 1;
2699 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2700 if (hrp->hrp_thrs == NULL)
2701 continue; /* uninitialized */
2702 for (j = 0; j < hrp->hrp_nthrs; j++)
2703 wake_up_all(&hrp->hrp_thrs[j].hrt_waitq);
2706 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2707 if (hrp->hrp_thrs == NULL)
2708 continue; /* uninitialized */
2709 wait_event(ptlrpc_hr.hr_waitq,
2710 atomic_read(&hrp->hrp_nstopped) ==
2711 atomic_read(&hrp->hrp_nstarted));
2715 static int ptlrpc_start_hr_threads(void)
2717 struct ptlrpc_hr_partition *hrp;
2722 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2725 for (j = 0; j < hrp->hrp_nthrs; j++) {
2726 struct ptlrpc_hr_thread *hrt = &hrp->hrp_thrs[j];
2727 struct task_struct *task;
2729 task = kthread_run(ptlrpc_hr_main,
2731 "ptlrpc_hr%02d_%03d",
2740 wait_event(ptlrpc_hr.hr_waitq,
2741 atomic_read(&hrp->hrp_nstarted) == j);
2744 CERROR("cannot start reply handler thread %d:%d: "
2745 "rc = %d\n", i, j, rc);
2746 ptlrpc_stop_hr_threads();
2754 static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
2756 struct l_wait_info lwi = { 0 };
2757 struct ptlrpc_thread *thread;
2758 struct list_head zombie;
2762 CDEBUG(D_INFO, "Stopping threads for service %s\n",
2763 svcpt->scp_service->srv_name);
2765 INIT_LIST_HEAD(&zombie);
2766 spin_lock(&svcpt->scp_lock);
2767 /* let the thread know that we would like it to stop asap */
2768 list_for_each_entry(thread, &svcpt->scp_threads, t_link) {
2769 CDEBUG(D_INFO, "Stopping thread %s #%u\n",
2770 svcpt->scp_service->srv_thread_name, thread->t_id);
2771 thread_add_flags(thread, SVC_STOPPING);
2774 wake_up_all(&svcpt->scp_waitq);
2776 while (!list_empty(&svcpt->scp_threads)) {
2777 thread = list_entry(svcpt->scp_threads.next,
2778 struct ptlrpc_thread, t_link);
2779 if (thread_is_stopped(thread)) {
2780 list_del(&thread->t_link);
2781 list_add(&thread->t_link, &zombie);
2784 spin_unlock(&svcpt->scp_lock);
2786 CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
2787 svcpt->scp_service->srv_thread_name, thread->t_id);
2788 l_wait_event(thread->t_ctl_waitq,
2789 thread_is_stopped(thread), &lwi);
2791 spin_lock(&svcpt->scp_lock);
2794 spin_unlock(&svcpt->scp_lock);
2796 while (!list_empty(&zombie)) {
2797 thread = list_entry(zombie.next,
2798 struct ptlrpc_thread, t_link);
2799 list_del(&thread->t_link);
2800 OBD_FREE_PTR(thread);
2806 * Stops all threads of a particular service \a svc
2808 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2810 struct ptlrpc_service_part *svcpt;
2814 ptlrpc_service_for_each_part(svcpt, i, svc) {
2815 if (svcpt->scp_service != NULL)
2816 ptlrpc_svcpt_stop_threads(svcpt);
2822 int ptlrpc_start_threads(struct ptlrpc_service *svc)
2829 /* We require 2 threads min, see note in ptlrpc_server_handle_request */
2830 LASSERT(svc->srv_nthrs_cpt_init >= PTLRPC_NTHRS_INIT);
2832 for (i = 0; i < svc->srv_ncpts; i++) {
2833 for (j = 0; j < svc->srv_nthrs_cpt_init; j++) {
2834 rc = ptlrpc_start_thread(svc->srv_parts[i], 1);
2840 /* We have enough threads, don't start more. b=15759 */
2847 CERROR("cannot start %s thread #%d_%d: rc %d\n",
2848 svc->srv_thread_name, i, j, rc);
2849 ptlrpc_stop_all_threads(svc);
2853 int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
2855 struct l_wait_info lwi = { 0 };
2856 struct ptlrpc_thread *thread;
2857 struct ptlrpc_service *svc;
2858 struct task_struct *task;
2862 LASSERT(svcpt != NULL);
2864 svc = svcpt->scp_service;
2866 CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n",
2867 svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running,
2868 svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit);
2871 if (unlikely(svc->srv_is_stopping))
2874 if (!ptlrpc_threads_increasable(svcpt) ||
2875 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
2876 svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1))
2879 OBD_CPT_ALLOC_PTR(thread, svc->srv_cptable, svcpt->scp_cpt);
2882 init_waitqueue_head(&thread->t_ctl_waitq);
2884 spin_lock(&svcpt->scp_lock);
2885 if (!ptlrpc_threads_increasable(svcpt)) {
2886 spin_unlock(&svcpt->scp_lock);
2887 OBD_FREE_PTR(thread);
2891 if (svcpt->scp_nthrs_starting != 0) {
2892 /* serialize starting because some modules (obdfilter)
2893 * might require unique and contiguous t_id */
2894 LASSERT(svcpt->scp_nthrs_starting == 1);
2895 spin_unlock(&svcpt->scp_lock);
2896 OBD_FREE_PTR(thread);
2898 CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n",
2899 svc->srv_thread_name, svcpt->scp_thr_nextid);
2904 CDEBUG(D_INFO, "Creating thread %s #%d race, retry later\n",
2905 svc->srv_thread_name, svcpt->scp_thr_nextid);
2909 svcpt->scp_nthrs_starting++;
2910 thread->t_id = svcpt->scp_thr_nextid++;
2911 thread_add_flags(thread, SVC_STARTING);
2912 thread->t_svcpt = svcpt;
2914 list_add(&thread->t_link, &svcpt->scp_threads);
2915 spin_unlock(&svcpt->scp_lock);
2917 if (svcpt->scp_cpt >= 0) {
2918 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s%02d_%03d",
2919 svc->srv_thread_name, svcpt->scp_cpt, thread->t_id);
2921 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s_%04d",
2922 svc->srv_thread_name, thread->t_id);
2925 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name);
2926 task = kthread_run(ptlrpc_main, thread, "%s", thread->t_name);
2929 CERROR("cannot start thread '%s': rc = %d\n",
2930 thread->t_name, rc);
2931 spin_lock(&svcpt->scp_lock);
2932 --svcpt->scp_nthrs_starting;
2933 if (thread_is_stopping(thread)) {
2934 /* this ptlrpc_thread is being hanled
2935 * by ptlrpc_svcpt_stop_threads now
2937 thread_add_flags(thread, SVC_STOPPED);
2938 wake_up(&thread->t_ctl_waitq);
2939 spin_unlock(&svcpt->scp_lock);
2941 list_del(&thread->t_link);
2942 spin_unlock(&svcpt->scp_lock);
2943 OBD_FREE_PTR(thread);
2951 l_wait_event(thread->t_ctl_waitq,
2952 thread_is_running(thread) || thread_is_stopped(thread),
2955 rc = thread_is_stopped(thread) ? thread->t_id : 0;
2959 int ptlrpc_hr_init(void)
2961 struct ptlrpc_hr_partition *hrp;
2962 struct ptlrpc_hr_thread *hrt;
2969 memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
2970 ptlrpc_hr.hr_cpt_table = cfs_cpt_table;
2972 ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
2974 if (ptlrpc_hr.hr_partitions == NULL)
2977 init_waitqueue_head(&ptlrpc_hr.hr_waitq);
2979 weight = cpumask_weight(topology_sibling_cpumask(smp_processor_id()));
2981 cfs_percpt_for_each(hrp, cpt, ptlrpc_hr.hr_partitions) {
2984 atomic_set(&hrp->hrp_nstarted, 0);
2985 atomic_set(&hrp->hrp_nstopped, 0);
2987 hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, cpt);
2988 hrp->hrp_nthrs /= weight;
2989 if (hrp->hrp_nthrs == 0)
2992 OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, cpt,
2993 hrp->hrp_nthrs * sizeof(*hrt));
2994 if (hrp->hrp_thrs == NULL)
2995 GOTO(out, rc = -ENOMEM);
2997 for (i = 0; i < hrp->hrp_nthrs; i++) {
2998 hrt = &hrp->hrp_thrs[i];
3001 hrt->hrt_partition = hrp;
3002 init_waitqueue_head(&hrt->hrt_waitq);
3003 spin_lock_init(&hrt->hrt_lock);
3004 INIT_LIST_HEAD(&hrt->hrt_queue);
3008 rc = ptlrpc_start_hr_threads();
3015 void ptlrpc_hr_fini(void)
3017 struct ptlrpc_hr_partition *hrp;
3020 if (ptlrpc_hr.hr_partitions == NULL)
3023 ptlrpc_stop_hr_threads();
3025 cfs_percpt_for_each(hrp, cpt, ptlrpc_hr.hr_partitions) {
3026 if (hrp->hrp_thrs != NULL) {
3027 OBD_FREE(hrp->hrp_thrs,
3028 hrp->hrp_nthrs * sizeof(hrp->hrp_thrs[0]));
3032 cfs_percpt_free(ptlrpc_hr.hr_partitions);
3033 ptlrpc_hr.hr_partitions = NULL;
3038 * Wait until all already scheduled replies are processed.
3040 static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
3044 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
3047 rc = l_wait_event(svcpt->scp_waitq,
3048 atomic_read(&svcpt->scp_nreps_difficult) == 0, &lwi);
3051 CWARN("Unexpectedly long timeout %s %p\n",
3052 svcpt->scp_service->srv_name, svcpt->scp_service);
3057 ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
3059 struct ptlrpc_service_part *svcpt;
3062 /* early disarm AT timer... */
3063 ptlrpc_service_for_each_part(svcpt, i, svc) {
3064 if (svcpt->scp_service != NULL)
3065 del_timer(&svcpt->scp_at_timer);
3070 ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
3072 struct ptlrpc_service_part *svcpt;
3073 struct ptlrpc_request_buffer_desc *rqbd;
3074 struct l_wait_info lwi;
3078 /* All history will be culled when the next request buffer is
3079 * freed in ptlrpc_service_purge_all() */
3080 svc->srv_hist_nrqbds_cpt_max = 0;
3082 rc = LNetClearLazyPortal(svc->srv_req_portal);
3085 ptlrpc_service_for_each_part(svcpt, i, svc) {
3086 if (svcpt->scp_service == NULL)
3089 /* Unlink all the request buffers. This forces a 'final'
3090 * event with its 'unlink' flag set for each posted rqbd */
3091 list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
3093 rc = LNetMDUnlink(rqbd->rqbd_md_h);
3094 LASSERT(rc == 0 || rc == -ENOENT);
3098 ptlrpc_service_for_each_part(svcpt, i, svc) {
3099 if (svcpt->scp_service == NULL)
3102 /* Wait for the network to release any buffers
3103 * it's currently filling */
3104 spin_lock(&svcpt->scp_lock);
3105 while (svcpt->scp_nrqbds_posted != 0) {
3106 spin_unlock(&svcpt->scp_lock);
3107 /* Network access will complete in finite time but
3108 * the HUGE timeout lets us CWARN for visibility
3109 * of sluggish NALs */
3110 lwi = LWI_TIMEOUT_INTERVAL(
3111 cfs_time_seconds(LONG_UNLINK),
3112 cfs_time_seconds(1), NULL, NULL);
3113 rc = l_wait_event(svcpt->scp_waitq,
3114 svcpt->scp_nrqbds_posted == 0, &lwi);
3115 if (rc == -ETIMEDOUT) {
3116 CWARN("Service %s waiting for "
3117 "request buffers\n",
3118 svcpt->scp_service->srv_name);
3120 spin_lock(&svcpt->scp_lock);
3122 spin_unlock(&svcpt->scp_lock);
3127 ptlrpc_service_purge_all(struct ptlrpc_service *svc)
3129 struct ptlrpc_service_part *svcpt;
3130 struct ptlrpc_request_buffer_desc *rqbd;
3131 struct ptlrpc_request *req;
3132 struct ptlrpc_reply_state *rs;
3135 ptlrpc_service_for_each_part(svcpt, i, svc) {
3136 if (svcpt->scp_service == NULL)
3139 spin_lock(&svcpt->scp_rep_lock);
3140 while (!list_empty(&svcpt->scp_rep_active)) {
3141 rs = list_entry(svcpt->scp_rep_active.next,
3142 struct ptlrpc_reply_state, rs_list);
3143 spin_lock(&rs->rs_lock);
3144 ptlrpc_schedule_difficult_reply(rs);
3145 spin_unlock(&rs->rs_lock);
3147 spin_unlock(&svcpt->scp_rep_lock);
3149 /* purge the request queue. NB No new replies (rqbds
3150 * all unlinked) and no service threads, so I'm the only
3151 * thread noodling the request queue now */
3152 while (!list_empty(&svcpt->scp_req_incoming)) {
3153 req = list_entry(svcpt->scp_req_incoming.next,
3154 struct ptlrpc_request, rq_list);
3156 list_del(&req->rq_list);
3157 svcpt->scp_nreqs_incoming--;
3158 ptlrpc_server_finish_request(svcpt, req);
3161 while (ptlrpc_server_request_pending(svcpt, true)) {
3162 req = ptlrpc_server_request_get(svcpt, true);
3163 ptlrpc_server_finish_active_request(svcpt, req);
3166 LASSERT(list_empty(&svcpt->scp_rqbd_posted));
3167 LASSERT(svcpt->scp_nreqs_incoming == 0);
3168 LASSERT(svcpt->scp_nreqs_active == 0);
3169 /* history should have been culled by
3170 * ptlrpc_server_finish_request */
3171 LASSERT(svcpt->scp_hist_nrqbds == 0);
3173 /* Now free all the request buffers since nothing
3174 * references them any more... */
3176 while (!list_empty(&svcpt->scp_rqbd_idle)) {
3177 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
3178 struct ptlrpc_request_buffer_desc,
3180 ptlrpc_free_rqbd(rqbd);
3182 ptlrpc_wait_replies(svcpt);
3184 while (!list_empty(&svcpt->scp_rep_idle)) {
3185 rs = list_entry(svcpt->scp_rep_idle.next,
3186 struct ptlrpc_reply_state,
3188 list_del(&rs->rs_list);
3189 OBD_FREE_LARGE(rs, svc->srv_max_reply_size);
3195 ptlrpc_service_free(struct ptlrpc_service *svc)
3197 struct ptlrpc_service_part *svcpt;
3198 struct ptlrpc_at_array *array;
3201 ptlrpc_service_for_each_part(svcpt, i, svc) {
3202 if (svcpt->scp_service == NULL)
3205 /* In case somebody rearmed this in the meantime */
3206 del_timer(&svcpt->scp_at_timer);
3207 array = &svcpt->scp_at_array;
3209 if (array->paa_reqs_array != NULL) {
3210 OBD_FREE(array->paa_reqs_array,
3211 sizeof(struct list_head) * array->paa_size);
3212 array->paa_reqs_array = NULL;
3215 if (array->paa_reqs_count != NULL) {
3216 OBD_FREE(array->paa_reqs_count,
3217 sizeof(__u32) * array->paa_size);
3218 array->paa_reqs_count = NULL;
3222 ptlrpc_service_for_each_part(svcpt, i, svc)
3223 OBD_FREE_PTR(svcpt);
3225 if (svc->srv_cpts != NULL)
3226 cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts);
3228 OBD_FREE(svc, offsetof(struct ptlrpc_service,
3229 srv_parts[svc->srv_ncpts]));
3232 int ptlrpc_unregister_service(struct ptlrpc_service *service)
3236 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
3238 service->srv_is_stopping = 1;
3240 mutex_lock(&ptlrpc_all_services_mutex);
3241 list_del_init(&service->srv_list);
3242 mutex_unlock(&ptlrpc_all_services_mutex);
3244 ptlrpc_service_del_atimer(service);
3245 ptlrpc_stop_all_threads(service);
3247 ptlrpc_service_unlink_rqbd(service);
3248 ptlrpc_service_purge_all(service);
3249 ptlrpc_service_nrs_cleanup(service);
3251 ptlrpc_lprocfs_unregister_service(service);
3252 ptlrpc_sysfs_unregister_service(service);
3254 ptlrpc_service_free(service);
3258 EXPORT_SYMBOL(ptlrpc_unregister_service);
3261 * Returns 0 if the service is healthy.
3263 * Right now, it just checks to make sure that requests aren't languishing
3264 * in the queue. We'll use this health check to govern whether a node needs
3265 * to be shot, so it's intentionally non-aggressive. */
3266 static int ptlrpc_svcpt_health_check(struct ptlrpc_service_part *svcpt)
3268 struct ptlrpc_request *request = NULL;
3269 struct timespec64 right_now;
3270 struct timespec64 timediff;
3272 ktime_get_real_ts64(&right_now);
3274 spin_lock(&svcpt->scp_req_lock);
3275 /* How long has the next entry been waiting? */
3276 if (ptlrpc_server_high_pending(svcpt, true))
3277 request = ptlrpc_nrs_req_peek_nolock(svcpt, true);
3278 else if (ptlrpc_server_normal_pending(svcpt, true))
3279 request = ptlrpc_nrs_req_peek_nolock(svcpt, false);
3281 if (request == NULL) {
3282 spin_unlock(&svcpt->scp_req_lock);
3286 timediff = timespec64_sub(right_now, request->rq_arrival_time);
3287 spin_unlock(&svcpt->scp_req_lock);
3289 if ((timediff.tv_sec) >
3290 (AT_OFF ? obd_timeout * 3 / 2 : at_max)) {
3291 CERROR("%s: unhealthy - request has been waiting %llds\n",
3292 svcpt->scp_service->srv_name, (s64)timediff.tv_sec);
3300 ptlrpc_service_health_check(struct ptlrpc_service *svc)
3302 struct ptlrpc_service_part *svcpt;
3308 ptlrpc_service_for_each_part(svcpt, i, svc) {
3309 int rc = ptlrpc_svcpt_health_check(svcpt);
3316 EXPORT_SYMBOL(ptlrpc_service_health_check);