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++) {
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 ||
144 (svc->srv_nrqbds_max != 0 &&
145 svcpt->scp_nrqbds_total > svc->srv_nrqbds_max))
148 rqbd = ptlrpc_alloc_rqbd(svcpt);
151 CERROR("%s: Can't allocate request buffer\n",
158 spin_lock(&svcpt->scp_lock);
160 LASSERT(svcpt->scp_rqbd_allocating == 1);
161 svcpt->scp_rqbd_allocating--;
163 spin_unlock(&svcpt->scp_lock);
166 "%s: allocate %d new %d-byte reqbufs (%d/%d left), rc = %d\n",
167 svc->srv_name, i, svc->srv_buf_size, svcpt->scp_nrqbds_posted,
168 svcpt->scp_nrqbds_total, rc);
172 rc = ptlrpc_server_post_idle_rqbds(svcpt);
178 * Part of Rep-Ack logic.
179 * Puts a lock and its mode into reply state assotiated to request reply.
181 void ptlrpc_save_lock(struct ptlrpc_request *req, struct lustre_handle *lock,
182 int mode, bool no_ack, bool convert_lock)
184 struct ptlrpc_reply_state *rs = req->rq_reply_state;
188 LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
190 idx = rs->rs_nlocks++;
191 rs->rs_locks[idx] = *lock;
192 rs->rs_modes[idx] = mode;
193 rs->rs_difficult = 1;
194 rs->rs_no_ack = no_ack;
195 rs->rs_convert_lock = convert_lock;
197 EXPORT_SYMBOL(ptlrpc_save_lock);
200 struct ptlrpc_hr_partition;
202 struct ptlrpc_hr_thread {
203 int hrt_id; /* thread ID */
205 wait_queue_head_t hrt_waitq;
206 struct list_head hrt_queue;
207 struct ptlrpc_hr_partition *hrt_partition;
210 struct ptlrpc_hr_partition {
211 /* # of started threads */
212 atomic_t hrp_nstarted;
213 /* # of stopped threads */
214 atomic_t hrp_nstopped;
215 /* cpu partition id */
217 /* round-robin rotor for choosing thread */
219 /* total number of threads on this partition */
222 struct ptlrpc_hr_thread *hrp_thrs;
225 #define HRT_RUNNING 0
226 #define HRT_STOPPING 1
228 struct ptlrpc_hr_service {
229 /* CPU partition table, it's just cfs_cpt_table for now */
230 struct cfs_cpt_table *hr_cpt_table;
231 /** controller sleep waitq */
232 wait_queue_head_t hr_waitq;
233 unsigned int hr_stopping;
234 /** roundrobin rotor for non-affinity service */
235 unsigned int hr_rotor;
237 struct ptlrpc_hr_partition **hr_partitions;
241 struct list_head rsb_replies;
242 unsigned int rsb_n_replies;
243 struct ptlrpc_service_part *rsb_svcpt;
246 /** reply handling service. */
247 static struct ptlrpc_hr_service ptlrpc_hr;
250 * maximum mumber of replies scheduled in one batch
252 #define MAX_SCHEDULED 256
255 * Initialize a reply batch.
259 static void rs_batch_init(struct rs_batch *b)
261 memset(b, 0, sizeof *b);
262 INIT_LIST_HEAD(&b->rsb_replies);
266 * Choose an hr thread to dispatch requests to.
269 struct ptlrpc_hr_thread *ptlrpc_hr_select(struct ptlrpc_service_part *svcpt)
271 struct ptlrpc_hr_partition *hrp;
274 if (svcpt->scp_cpt >= 0 &&
275 svcpt->scp_service->srv_cptable == ptlrpc_hr.hr_cpt_table) {
276 /* directly match partition */
277 hrp = ptlrpc_hr.hr_partitions[svcpt->scp_cpt];
280 rotor = ptlrpc_hr.hr_rotor++;
281 rotor %= cfs_cpt_number(ptlrpc_hr.hr_cpt_table);
283 hrp = ptlrpc_hr.hr_partitions[rotor];
286 rotor = hrp->hrp_rotor++;
287 return &hrp->hrp_thrs[rotor % hrp->hrp_nthrs];
291 * Dispatch all replies accumulated in the batch to one from
292 * dedicated reply handling threads.
296 static void rs_batch_dispatch(struct rs_batch *b)
298 if (b->rsb_n_replies != 0) {
299 struct ptlrpc_hr_thread *hrt;
301 hrt = ptlrpc_hr_select(b->rsb_svcpt);
303 spin_lock(&hrt->hrt_lock);
304 list_splice_init(&b->rsb_replies, &hrt->hrt_queue);
305 spin_unlock(&hrt->hrt_lock);
307 wake_up(&hrt->hrt_waitq);
308 b->rsb_n_replies = 0;
313 * Add a reply to a batch.
314 * Add one reply object to a batch, schedule batched replies if overload.
319 static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs)
321 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
323 if (svcpt != b->rsb_svcpt || b->rsb_n_replies >= MAX_SCHEDULED) {
324 if (b->rsb_svcpt != NULL) {
325 rs_batch_dispatch(b);
326 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
328 spin_lock(&svcpt->scp_rep_lock);
329 b->rsb_svcpt = svcpt;
331 spin_lock(&rs->rs_lock);
332 rs->rs_scheduled_ever = 1;
333 if (rs->rs_scheduled == 0) {
334 list_move(&rs->rs_list, &b->rsb_replies);
335 rs->rs_scheduled = 1;
338 rs->rs_committed = 1;
339 spin_unlock(&rs->rs_lock);
343 * Reply batch finalization.
344 * Dispatch remaining replies from the batch
345 * and release remaining spinlock.
349 static void rs_batch_fini(struct rs_batch *b)
351 if (b->rsb_svcpt != NULL) {
352 rs_batch_dispatch(b);
353 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
357 #define DECLARE_RS_BATCH(b) struct rs_batch b
361 * Put reply state into a queue for processing because we received
362 * ACK from the client
364 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
366 struct ptlrpc_hr_thread *hrt;
369 LASSERT(list_empty(&rs->rs_list));
371 hrt = ptlrpc_hr_select(rs->rs_svcpt);
373 spin_lock(&hrt->hrt_lock);
374 list_add_tail(&rs->rs_list, &hrt->hrt_queue);
375 spin_unlock(&hrt->hrt_lock);
377 wake_up(&hrt->hrt_waitq);
381 void 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);
429 static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt)
431 struct ptlrpc_request_buffer_desc *rqbd;
436 spin_lock(&svcpt->scp_lock);
438 if (list_empty(&svcpt->scp_rqbd_idle)) {
439 spin_unlock(&svcpt->scp_lock);
443 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
444 struct ptlrpc_request_buffer_desc,
446 list_del(&rqbd->rqbd_list);
448 /* assume we will post successfully */
449 svcpt->scp_nrqbds_posted++;
450 list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_posted);
452 spin_unlock(&svcpt->scp_lock);
454 rc = ptlrpc_register_rqbd(rqbd);
461 spin_lock(&svcpt->scp_lock);
463 svcpt->scp_nrqbds_posted--;
464 list_del(&rqbd->rqbd_list);
465 list_add_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
467 /* Don't complain if no request buffers are posted right now; LNET
468 * won't drop requests because we set the portal lazy! */
470 spin_unlock(&svcpt->scp_lock);
475 static void ptlrpc_at_timer(cfs_timer_cb_arg_t data)
477 struct ptlrpc_service_part *svcpt;
479 svcpt = cfs_from_timer(svcpt, data, scp_at_timer);
481 svcpt->scp_at_check = 1;
482 svcpt->scp_at_checktime = ktime_get();
483 wake_up(&svcpt->scp_waitq);
486 static void ptlrpc_server_nthreads_check(struct ptlrpc_service *svc,
487 struct ptlrpc_service_conf *conf)
489 struct ptlrpc_service_thr_conf *tc = &conf->psc_thr;
496 * Common code for estimating & validating threads number.
497 * CPT affinity service could have percpt thread-pool instead
498 * of a global thread-pool, which means user might not always
499 * get the threads number they give it in conf::tc_nthrs_user
500 * even they did set. It's because we need to validate threads
501 * number for each CPT to guarantee each pool will have enough
502 * threads to keep the service healthy.
504 init = PTLRPC_NTHRS_INIT + (svc->srv_ops.so_hpreq_handler != NULL);
505 init = max_t(int, init, tc->tc_nthrs_init);
507 /* NB: please see comments in lustre_lnet.h for definition
508 * details of these members */
509 LASSERT(tc->tc_nthrs_max != 0);
511 if (tc->tc_nthrs_user != 0) {
512 /* In case there is a reason to test a service with many
513 * threads, we give a less strict check here, it can
514 * be up to 8 * nthrs_max */
515 total = min(tc->tc_nthrs_max * 8, tc->tc_nthrs_user);
516 nthrs = total / svc->srv_ncpts;
517 init = max(init, nthrs);
521 total = tc->tc_nthrs_max;
522 if (tc->tc_nthrs_base == 0) {
523 /* don't care about base threads number per partition,
524 * this is most for non-affinity service */
525 nthrs = total / svc->srv_ncpts;
529 nthrs = tc->tc_nthrs_base;
530 if (svc->srv_ncpts == 1) {
533 /* NB: Increase the base number if it's single partition
534 * and total number of cores/HTs is larger or equal to 4.
535 * result will always < 2 * nthrs_base */
536 weight = cfs_cpt_weight(svc->srv_cptable, CFS_CPT_ANY);
537 for (i = 1; (weight >> (i + 1)) != 0 && /* >= 4 cores/HTs */
538 (tc->tc_nthrs_base >> i) != 0; i++)
539 nthrs += tc->tc_nthrs_base >> i;
542 if (tc->tc_thr_factor != 0) {
543 int factor = tc->tc_thr_factor;
547 * User wants to increase number of threads with for
548 * each CPU core/HT, most likely the factor is larger than
549 * one thread/core because service threads are supposed to
550 * be blocked by lock or wait for IO.
553 * Amdahl's law says that adding processors wouldn't give
554 * a linear increasing of parallelism, so it's nonsense to
555 * have too many threads no matter how many cores/HTs
558 if (cpumask_weight(topology_sibling_cpumask(smp_processor_id())) > 1) {
559 /* weight is # of HTs */
560 /* depress thread factor for hyper-thread */
561 factor = factor - (factor >> 1) + (factor >> 3);
564 weight = cfs_cpt_weight(svc->srv_cptable, 0);
566 for (; factor > 0 && weight > 0; factor--, weight -= fade)
567 nthrs += min(weight, fade) * factor;
570 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
571 nthrs = max(tc->tc_nthrs_base,
572 tc->tc_nthrs_max / svc->srv_ncpts);
575 nthrs = max(nthrs, tc->tc_nthrs_init);
576 svc->srv_nthrs_cpt_limit = nthrs;
577 svc->srv_nthrs_cpt_init = init;
579 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
580 CDEBUG(D_OTHER, "%s: This service may have more threads (%d) "
581 "than the given soft limit (%d)\n",
582 svc->srv_name, nthrs * svc->srv_ncpts,
588 * Initialize percpt data for a service
590 static int ptlrpc_service_part_init(struct ptlrpc_service *svc,
591 struct ptlrpc_service_part *svcpt, int cpt)
593 struct ptlrpc_at_array *array;
598 svcpt->scp_cpt = cpt;
599 INIT_LIST_HEAD(&svcpt->scp_threads);
601 /* rqbd and incoming request queue */
602 spin_lock_init(&svcpt->scp_lock);
603 mutex_init(&svcpt->scp_mutex);
604 INIT_LIST_HEAD(&svcpt->scp_rqbd_idle);
605 INIT_LIST_HEAD(&svcpt->scp_rqbd_posted);
606 INIT_LIST_HEAD(&svcpt->scp_req_incoming);
607 init_waitqueue_head(&svcpt->scp_waitq);
608 /* history request & rqbd list */
609 INIT_LIST_HEAD(&svcpt->scp_hist_reqs);
610 INIT_LIST_HEAD(&svcpt->scp_hist_rqbds);
612 /* acitve requests and hp requests */
613 spin_lock_init(&svcpt->scp_req_lock);
616 spin_lock_init(&svcpt->scp_rep_lock);
617 INIT_LIST_HEAD(&svcpt->scp_rep_active);
618 INIT_LIST_HEAD(&svcpt->scp_rep_idle);
619 init_waitqueue_head(&svcpt->scp_rep_waitq);
620 atomic_set(&svcpt->scp_nreps_difficult, 0);
622 /* adaptive timeout */
623 spin_lock_init(&svcpt->scp_at_lock);
624 array = &svcpt->scp_at_array;
626 size = at_est2timeout(at_max);
627 array->paa_size = size;
628 array->paa_count = 0;
629 array->paa_deadline = -1;
631 /* allocate memory for scp_at_array (ptlrpc_at_array) */
632 OBD_CPT_ALLOC(array->paa_reqs_array,
633 svc->srv_cptable, cpt, sizeof(struct list_head) * size);
634 if (array->paa_reqs_array == NULL)
637 for (index = 0; index < size; index++)
638 INIT_LIST_HEAD(&array->paa_reqs_array[index]);
640 OBD_CPT_ALLOC(array->paa_reqs_count,
641 svc->srv_cptable, cpt, sizeof(__u32) * size);
642 if (array->paa_reqs_count == NULL)
645 cfs_timer_setup(&svcpt->scp_at_timer, ptlrpc_at_timer,
646 (unsigned long)svcpt, 0);
648 /* At SOW, service time should be quick; 10s seems generous. If client
649 * timeout is less than this, we'll be sending an early reply. */
650 at_init(&svcpt->scp_at_estimate, 10, 0);
652 /* assign this before call ptlrpc_grow_req_bufs */
653 svcpt->scp_service = svc;
654 /* Now allocate the request buffers, but don't post them now */
655 rc = ptlrpc_grow_req_bufs(svcpt, 0);
656 /* We shouldn't be under memory pressure at startup, so
657 * fail if we can't allocate all our buffers at this time. */
664 if (array->paa_reqs_count != NULL) {
665 OBD_FREE(array->paa_reqs_count, sizeof(__u32) * size);
666 array->paa_reqs_count = NULL;
669 if (array->paa_reqs_array != NULL) {
670 OBD_FREE(array->paa_reqs_array,
671 sizeof(struct list_head) * array->paa_size);
672 array->paa_reqs_array = NULL;
679 * Initialize service on a given portal.
680 * This includes starting serving threads , allocating and posting rqbds and
683 struct ptlrpc_service *ptlrpc_register_service(struct ptlrpc_service_conf *conf,
685 struct dentry *debugfs_entry)
687 struct ptlrpc_service_cpt_conf *cconf = &conf->psc_cpt;
688 struct ptlrpc_service *service;
689 struct ptlrpc_service_part *svcpt;
690 struct cfs_cpt_table *cptable;
698 LASSERT(conf->psc_buf.bc_nbufs > 0);
699 LASSERT(conf->psc_buf.bc_buf_size >=
700 conf->psc_buf.bc_req_max_size + SPTLRPC_MAX_PAYLOAD);
701 LASSERT(conf->psc_thr.tc_ctx_tags != 0);
703 cptable = cconf->cc_cptable;
705 cptable = cfs_cpt_table;
707 if (conf->psc_thr.tc_cpu_bind > 1) {
708 CERROR("%s: Invalid cpu bind value %d, only 1 or 0 allowed\n",
709 conf->psc_name, conf->psc_thr.tc_cpu_bind);
710 RETURN(ERR_PTR(-EINVAL));
713 if (!cconf->cc_affinity) {
716 ncpts = cfs_cpt_number(cptable);
717 if (cconf->cc_pattern != NULL) {
718 struct cfs_expr_list *el;
720 rc = cfs_expr_list_parse(cconf->cc_pattern,
721 strlen(cconf->cc_pattern),
724 CERROR("%s: invalid CPT pattern string: %s",
725 conf->psc_name, cconf->cc_pattern);
726 RETURN(ERR_PTR(-EINVAL));
729 rc = cfs_expr_list_values(el, ncpts, &cpts);
730 cfs_expr_list_free(el);
732 CERROR("%s: failed to parse CPT array %s: %d\n",
733 conf->psc_name, cconf->cc_pattern, rc);
735 OBD_FREE(cpts, sizeof(*cpts) * ncpts);
736 RETURN(ERR_PTR(rc < 0 ? rc : -EINVAL));
742 OBD_ALLOC(service, offsetof(struct ptlrpc_service, srv_parts[ncpts]));
743 if (service == NULL) {
745 OBD_FREE(cpts, sizeof(*cpts) * ncpts);
746 RETURN(ERR_PTR(-ENOMEM));
749 service->srv_cptable = cptable;
750 service->srv_cpts = cpts;
751 service->srv_ncpts = ncpts;
752 service->srv_cpt_bind = conf->psc_thr.tc_cpu_bind;
754 service->srv_cpt_bits = 0; /* it's zero already, easy to read... */
755 while ((1 << service->srv_cpt_bits) < cfs_cpt_number(cptable))
756 service->srv_cpt_bits++;
759 spin_lock_init(&service->srv_lock);
760 service->srv_name = conf->psc_name;
761 service->srv_watchdog_factor = conf->psc_watchdog_factor;
762 INIT_LIST_HEAD(&service->srv_list); /* for safty of cleanup */
764 /* buffer configuration */
765 service->srv_nbuf_per_group = test_req_buffer_pressure ?
766 1 : conf->psc_buf.bc_nbufs;
767 /* do not limit max number of rqbds by default */
768 service->srv_nrqbds_max = 0;
770 service->srv_max_req_size = conf->psc_buf.bc_req_max_size +
772 service->srv_buf_size = conf->psc_buf.bc_buf_size;
773 service->srv_rep_portal = conf->psc_buf.bc_rep_portal;
774 service->srv_req_portal = conf->psc_buf.bc_req_portal;
776 /* With slab/alloc_pages buffer size will be rounded up to 2^n */
777 if (service->srv_buf_size & (service->srv_buf_size - 1)) {
778 int round = size_roundup_power2(service->srv_buf_size);
779 service->srv_buf_size = round;
782 /* Increase max reply size to next power of two */
783 service->srv_max_reply_size = 1;
784 while (service->srv_max_reply_size <
785 conf->psc_buf.bc_rep_max_size + SPTLRPC_MAX_PAYLOAD)
786 service->srv_max_reply_size <<= 1;
788 service->srv_thread_name = conf->psc_thr.tc_thr_name;
789 service->srv_ctx_tags = conf->psc_thr.tc_ctx_tags;
790 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
791 service->srv_ops = conf->psc_ops;
793 for (i = 0; i < ncpts; i++) {
794 if (!cconf->cc_affinity)
797 cpt = cpts != NULL ? cpts[i] : i;
799 OBD_CPT_ALLOC(svcpt, cptable, cpt, sizeof(*svcpt));
801 GOTO(failed, rc = -ENOMEM);
803 service->srv_parts[i] = svcpt;
804 rc = ptlrpc_service_part_init(service, svcpt, cpt);
809 ptlrpc_server_nthreads_check(service, conf);
811 rc = LNetSetLazyPortal(service->srv_req_portal);
814 mutex_lock(&ptlrpc_all_services_mutex);
815 list_add(&service->srv_list, &ptlrpc_all_services);
816 mutex_unlock(&ptlrpc_all_services_mutex);
819 rc = ptlrpc_sysfs_register_service(parent, service);
824 if (debugfs_entry != NULL)
825 ptlrpc_ldebugfs_register_service(debugfs_entry, service);
827 rc = ptlrpc_service_nrs_setup(service);
831 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
832 service->srv_name, service->srv_req_portal);
834 rc = ptlrpc_start_threads(service);
836 CERROR("Failed to start threads for service %s: %d\n",
837 service->srv_name, rc);
843 ptlrpc_unregister_service(service);
846 EXPORT_SYMBOL(ptlrpc_register_service);
849 * to actually free the request, must be called without holding svc_lock.
850 * note it's caller's responsibility to unlink req->rq_list.
852 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
854 LASSERT(atomic_read(&req->rq_refcount) == 0);
855 LASSERT(list_empty(&req->rq_timed_list));
857 /* DEBUG_REQ() assumes the reply state of a request with a valid
858 * ref will not be destroyed until that reference is dropped. */
859 ptlrpc_req_drop_rs(req);
861 sptlrpc_svc_ctx_decref(req);
863 if (req != &req->rq_rqbd->rqbd_req) {
864 /* NB request buffers use an embedded
865 * req if the incoming req unlinked the
866 * MD; this isn't one of them! */
867 ptlrpc_request_cache_free(req);
872 * drop a reference count of the request. if it reaches 0, we either
873 * put it into history list, or free it immediately.
875 void ptlrpc_server_drop_request(struct ptlrpc_request *req)
877 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
878 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
879 struct ptlrpc_service *svc = svcpt->scp_service;
881 struct list_head *tmp;
882 struct list_head *nxt;
884 if (!atomic_dec_and_test(&req->rq_refcount))
887 if (req->rq_session.lc_state == LCS_ENTERED) {
888 lu_context_exit(&req->rq_session);
889 lu_context_fini(&req->rq_session);
892 if (req->rq_at_linked) {
893 spin_lock(&svcpt->scp_at_lock);
894 /* recheck with lock, in case it's unlinked by
895 * ptlrpc_at_check_timed() */
896 if (likely(req->rq_at_linked))
897 ptlrpc_at_remove_timed(req);
898 spin_unlock(&svcpt->scp_at_lock);
901 LASSERT(list_empty(&req->rq_timed_list));
903 /* finalize request */
904 if (req->rq_export) {
905 class_export_put(req->rq_export);
906 req->rq_export = NULL;
909 spin_lock(&svcpt->scp_lock);
911 list_add(&req->rq_list, &rqbd->rqbd_reqs);
913 refcount = --(rqbd->rqbd_refcount);
915 /* request buffer is now idle: add to history */
916 list_del(&rqbd->rqbd_list);
918 list_add_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds);
919 svcpt->scp_hist_nrqbds++;
921 /* cull some history?
922 * I expect only about 1 or 2 rqbds need to be recycled here */
923 while (svcpt->scp_hist_nrqbds > svc->srv_hist_nrqbds_cpt_max) {
924 rqbd = list_entry(svcpt->scp_hist_rqbds.next,
925 struct ptlrpc_request_buffer_desc,
928 list_del(&rqbd->rqbd_list);
929 svcpt->scp_hist_nrqbds--;
931 /* remove rqbd's reqs from svc's req history while
932 * I've got the service lock */
933 list_for_each(tmp, &rqbd->rqbd_reqs) {
934 req = list_entry(tmp, struct ptlrpc_request,
936 /* Track the highest culled req seq */
937 if (req->rq_history_seq >
938 svcpt->scp_hist_seq_culled) {
939 svcpt->scp_hist_seq_culled =
942 list_del(&req->rq_history_list);
945 spin_unlock(&svcpt->scp_lock);
947 list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
948 req = list_entry(rqbd->rqbd_reqs.next,
949 struct ptlrpc_request,
951 list_del(&req->rq_list);
952 ptlrpc_server_free_request(req);
955 spin_lock(&svcpt->scp_lock);
957 * now all reqs including the embedded req has been
958 * disposed, schedule request buffer for re-use
959 * or free it to drain some in excess.
961 LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) == 0);
962 if (svcpt->scp_nrqbds_posted >=
963 svc->srv_nbuf_per_group ||
964 (svc->srv_nrqbds_max != 0 &&
965 svcpt->scp_nrqbds_total > svc->srv_nrqbds_max) ||
966 test_req_buffer_pressure) {
967 /* like in ptlrpc_free_rqbd() */
968 svcpt->scp_nrqbds_total--;
969 OBD_FREE_LARGE(rqbd->rqbd_buffer,
973 list_add_tail(&rqbd->rqbd_list,
974 &svcpt->scp_rqbd_idle);
978 spin_unlock(&svcpt->scp_lock);
979 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
980 /* If we are low on memory, we are not interested in history */
981 list_del(&req->rq_list);
982 list_del_init(&req->rq_history_list);
984 /* Track the highest culled req seq */
985 if (req->rq_history_seq > svcpt->scp_hist_seq_culled)
986 svcpt->scp_hist_seq_culled = req->rq_history_seq;
988 spin_unlock(&svcpt->scp_lock);
990 ptlrpc_server_free_request(req);
992 spin_unlock(&svcpt->scp_lock);
996 /** Change request export and move hp request from old export to new */
997 void ptlrpc_request_change_export(struct ptlrpc_request *req,
998 struct obd_export *export)
1000 if (req->rq_export != NULL) {
1001 LASSERT(!list_empty(&req->rq_exp_list));
1002 /* remove rq_exp_list from last export */
1003 spin_lock(&req->rq_export->exp_rpc_lock);
1004 list_del_init(&req->rq_exp_list);
1005 spin_unlock(&req->rq_export->exp_rpc_lock);
1006 /* export has one reference already, so it`s safe to
1007 * add req to export queue here and get another
1008 * reference for request later */
1009 spin_lock(&export->exp_rpc_lock);
1010 if (req->rq_ops != NULL) /* hp request */
1011 list_add(&req->rq_exp_list, &export->exp_hp_rpcs);
1013 list_add(&req->rq_exp_list, &export->exp_reg_rpcs);
1014 spin_unlock(&export->exp_rpc_lock);
1016 class_export_rpc_dec(req->rq_export);
1017 class_export_put(req->rq_export);
1020 /* request takes one export refcount */
1021 req->rq_export = class_export_get(export);
1022 class_export_rpc_inc(export);
1028 * to finish a request: stop sending more early replies, and release
1031 static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt,
1032 struct ptlrpc_request *req)
1034 ptlrpc_server_hpreq_fini(req);
1036 ptlrpc_server_drop_request(req);
1040 * to finish an active request: stop sending more early replies, and release
1041 * the request. should be called after we finished handling the request.
1043 static void ptlrpc_server_finish_active_request(
1044 struct ptlrpc_service_part *svcpt,
1045 struct ptlrpc_request *req)
1047 spin_lock(&svcpt->scp_req_lock);
1048 ptlrpc_nrs_req_stop_nolock(req);
1049 svcpt->scp_nreqs_active--;
1051 svcpt->scp_nhreqs_active--;
1052 spin_unlock(&svcpt->scp_req_lock);
1054 ptlrpc_nrs_req_finalize(req);
1056 if (req->rq_export != NULL)
1057 class_export_rpc_dec(req->rq_export);
1059 ptlrpc_server_finish_request(svcpt, req);
1063 * This function makes sure dead exports are evicted in a timely manner.
1064 * This function is only called when some export receives a message (i.e.,
1065 * the network is up.)
1067 void ptlrpc_update_export_timer(struct obd_export *exp, time64_t extra_delay)
1069 struct obd_export *oldest_exp;
1070 time64_t oldest_time, new_time;
1076 /* Compensate for slow machines, etc, by faking our request time
1077 into the future. Although this can break the strict time-ordering
1078 of the list, we can be really lazy here - we don't have to evict
1079 at the exact right moment. Eventually, all silent exports
1080 will make it to the top of the list. */
1082 /* Do not pay attention on 1sec or smaller renewals. */
1083 new_time = ktime_get_real_seconds() + extra_delay;
1084 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
1087 exp->exp_last_request_time = new_time;
1089 /* exports may get disconnected from the chain even though the
1090 export has references, so we must keep the spin lock while
1091 manipulating the lists */
1092 spin_lock(&exp->exp_obd->obd_dev_lock);
1094 if (list_empty(&exp->exp_obd_chain_timed)) {
1095 /* this one is not timed */
1096 spin_unlock(&exp->exp_obd->obd_dev_lock);
1100 list_move_tail(&exp->exp_obd_chain_timed,
1101 &exp->exp_obd->obd_exports_timed);
1103 oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
1104 struct obd_export, exp_obd_chain_timed);
1105 oldest_time = oldest_exp->exp_last_request_time;
1106 spin_unlock(&exp->exp_obd->obd_dev_lock);
1108 if (exp->exp_obd->obd_recovering) {
1109 /* be nice to everyone during recovery */
1114 /* Note - racing to start/reset the obd_eviction timer is safe */
1115 if (exp->exp_obd->obd_eviction_timer == 0) {
1116 /* Check if the oldest entry is expired. */
1117 if (ktime_get_real_seconds() >
1118 oldest_time + PING_EVICT_TIMEOUT + extra_delay) {
1119 /* We need a second timer, in case the net was down and
1120 * it just came back. Since the pinger may skip every
1121 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
1122 * we better wait for 3.
1124 exp->exp_obd->obd_eviction_timer =
1125 ktime_get_real_seconds() + 3 * PING_INTERVAL;
1126 CDEBUG(D_HA, "%s: Think about evicting %s from %lld\n",
1127 exp->exp_obd->obd_name,
1128 obd_export_nid2str(oldest_exp), oldest_time);
1131 if (ktime_get_real_seconds() >
1132 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
1133 /* The evictor won't evict anyone who we've heard from
1134 * recently, so we don't have to check before we start
1137 if (!ping_evictor_wake(exp))
1138 exp->exp_obd->obd_eviction_timer = 0;
1146 * Sanity check request \a req.
1147 * Return 0 if all is ok, error code otherwise.
1149 static int ptlrpc_check_req(struct ptlrpc_request *req)
1151 struct obd_device *obd = req->rq_export->exp_obd;
1154 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
1155 req->rq_export->exp_conn_cnt)) {
1156 DEBUG_REQ(D_RPCTRACE, req,
1157 "DROPPING req from old connection %d < %d",
1158 lustre_msg_get_conn_cnt(req->rq_reqmsg),
1159 req->rq_export->exp_conn_cnt);
1162 if (unlikely(obd == NULL || obd->obd_fail)) {
1163 /* Failing over, don't handle any more reqs,
1164 * send error response instead. */
1165 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
1166 req, (obd != NULL) ? obd->obd_name : "unknown");
1168 } else if (lustre_msg_get_flags(req->rq_reqmsg) &
1169 (MSG_REPLAY | MSG_REQ_REPLAY_DONE) &&
1170 !obd->obd_recovering) {
1171 DEBUG_REQ(D_ERROR, req,
1172 "Invalid replay without recovery");
1173 class_fail_export(req->rq_export);
1175 } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 &&
1176 !obd->obd_recovering) {
1177 DEBUG_REQ(D_ERROR, req, "Invalid req with transno "
1178 "%llu without recovery",
1179 lustre_msg_get_transno(req->rq_reqmsg));
1180 class_fail_export(req->rq_export);
1184 if (unlikely(rc < 0)) {
1185 req->rq_status = rc;
1191 static void ptlrpc_at_set_timer(struct ptlrpc_service_part *svcpt)
1193 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1196 if (array->paa_count == 0) {
1197 del_timer(&svcpt->scp_at_timer);
1201 /* Set timer for closest deadline */
1202 next = array->paa_deadline - ktime_get_real_seconds() -
1205 ptlrpc_at_timer(cfs_timer_cb_arg(svcpt, scp_at_timer));
1207 mod_timer(&svcpt->scp_at_timer,
1208 jiffies + nsecs_to_jiffies(next * NSEC_PER_SEC));
1209 CDEBUG(D_INFO, "armed %s at %+llds\n",
1210 svcpt->scp_service->srv_name, next);
1214 /* Add rpc to early reply check list */
1215 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
1217 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1218 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1219 struct ptlrpc_request *rq = NULL;
1225 if (req->rq_no_reply)
1228 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
1231 spin_lock(&svcpt->scp_at_lock);
1232 LASSERT(list_empty(&req->rq_timed_list));
1234 div_u64_rem(req->rq_deadline, array->paa_size, &index);
1235 if (array->paa_reqs_count[index] > 0) {
1236 /* latest rpcs will have the latest deadlines in the list,
1237 * so search backward. */
1238 list_for_each_entry_reverse(rq,
1239 &array->paa_reqs_array[index],
1241 if (req->rq_deadline >= rq->rq_deadline) {
1242 list_add(&req->rq_timed_list,
1243 &rq->rq_timed_list);
1249 /* Add the request at the head of the list */
1250 if (list_empty(&req->rq_timed_list))
1251 list_add(&req->rq_timed_list,
1252 &array->paa_reqs_array[index]);
1254 spin_lock(&req->rq_lock);
1255 req->rq_at_linked = 1;
1256 spin_unlock(&req->rq_lock);
1257 req->rq_at_index = index;
1258 array->paa_reqs_count[index]++;
1260 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
1261 array->paa_deadline = req->rq_deadline;
1262 ptlrpc_at_set_timer(svcpt);
1264 spin_unlock(&svcpt->scp_at_lock);
1269 static void ptlrpc_at_remove_timed(struct ptlrpc_request *req)
1271 struct ptlrpc_at_array *array;
1273 array = &req->rq_rqbd->rqbd_svcpt->scp_at_array;
1275 /* NB: must call with hold svcpt::scp_at_lock */
1276 LASSERT(!list_empty(&req->rq_timed_list));
1277 list_del_init(&req->rq_timed_list);
1279 spin_lock(&req->rq_lock);
1280 req->rq_at_linked = 0;
1281 spin_unlock(&req->rq_lock);
1283 array->paa_reqs_count[req->rq_at_index]--;
1288 * Attempt to extend the request deadline by sending an early reply to the
1291 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
1293 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1294 struct ptlrpc_request *reqcopy;
1295 struct lustre_msg *reqmsg;
1296 time64_t olddl = req->rq_deadline - ktime_get_real_seconds();
1302 if (CFS_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_RECONNECT)) {
1303 /* don't send early reply */
1307 /* deadline is when the client expects us to reply, margin is the
1308 difference between clients' and servers' expectations */
1309 DEBUG_REQ(D_ADAPTTO, req,
1310 "%ssending early reply (deadline %+llds, margin %+llds) for "
1311 "%d+%d", AT_OFF ? "AT off - not " : "",
1312 (s64)olddl, (s64)(olddl - at_get(&svcpt->scp_at_estimate)),
1313 at_get(&svcpt->scp_at_estimate), at_extra);
1319 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+llds), "
1320 "not sending early reply. Consider increasing "
1321 "at_early_margin (%d)?", (s64)olddl, at_early_margin);
1323 /* Return an error so we're not re-added to the timed list. */
1327 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
1328 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
1329 "but no AT support");
1333 if (req->rq_export &&
1334 lustre_msg_get_flags(req->rq_reqmsg) &
1335 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1336 struct obd_device *obd_exp = req->rq_export->exp_obd;
1338 /* During recovery, we don't want to send too many early
1339 * replies, but on the other hand we want to make sure the
1340 * client has enough time to resend if the rpc is lost. So
1341 * during the recovery period send at least 4 early replies,
1342 * spacing them every at_extra if we can. at_estimate should
1343 * always equal this fixed value during recovery.
1345 /* Don't account request processing time into AT history
1346 * during recovery, it is not service time we need but
1347 * includes also waiting time for recovering clients
1349 newdl = min_t(time64_t, at_extra,
1350 obd_exp->obd_recovery_timeout / 4) +
1351 ktime_get_real_seconds();
1353 /* We want to extend the request deadline by at_extra seconds,
1354 * so we set our service estimate to reflect how much time has
1355 * passed since this request arrived plus an additional
1356 * at_extra seconds. The client will calculate the new deadline
1357 * based on this service estimate (plus some additional time to
1358 * account for network latency). See ptlrpc_at_recv_early_reply
1360 at_measured(&svcpt->scp_at_estimate, at_extra +
1361 ktime_get_real_seconds() -
1362 req->rq_arrival_time.tv_sec);
1363 newdl = req->rq_arrival_time.tv_sec +
1364 at_get(&svcpt->scp_at_estimate);
1367 /* Check to see if we've actually increased the deadline -
1368 * we may be past adaptive_max */
1369 if (req->rq_deadline >= newdl) {
1370 DEBUG_REQ(D_WARNING, req, "Couldn't add any time (%lld/%lld), not sending early reply\n",
1371 (s64)olddl, (s64)(newdl - ktime_get_real_seconds()));
1375 reqcopy = ptlrpc_request_cache_alloc(GFP_NOFS);
1376 if (reqcopy == NULL)
1378 OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1380 GOTO(out_free, rc = -ENOMEM);
1383 reqcopy->rq_reply_state = NULL;
1384 reqcopy->rq_rep_swab_mask = 0;
1385 reqcopy->rq_pack_bulk = 0;
1386 reqcopy->rq_pack_udesc = 0;
1387 reqcopy->rq_packed_final = 0;
1388 sptlrpc_svc_ctx_addref(reqcopy);
1389 /* We only need the reqmsg for the magic */
1390 reqcopy->rq_reqmsg = reqmsg;
1391 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1394 * tgt_brw_read() and tgt_brw_write() may have decided not to reply.
1395 * Without this check, we would fail the rq_no_reply assertion in
1396 * ptlrpc_send_reply().
1398 if (reqcopy->rq_no_reply)
1399 GOTO(out, rc = -ETIMEDOUT);
1401 LASSERT(atomic_read(&req->rq_refcount));
1402 /** if it is last refcount then early reply isn't needed */
1403 if (atomic_read(&req->rq_refcount) == 1) {
1404 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
1405 "abort sending early reply\n");
1406 GOTO(out, rc = -EINVAL);
1409 /* Connection ref */
1410 reqcopy->rq_export = class_conn2export(
1411 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1412 if (reqcopy->rq_export == NULL)
1413 GOTO(out, rc = -ENODEV);
1416 class_export_rpc_inc(reqcopy->rq_export);
1417 if (reqcopy->rq_export->exp_obd &&
1418 reqcopy->rq_export->exp_obd->obd_fail)
1419 GOTO(out_put, rc = -ENODEV);
1421 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1425 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1428 /* Adjust our own deadline to what we told the client */
1429 req->rq_deadline = newdl;
1430 req->rq_early_count++; /* number sent, server side */
1432 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1435 /* Free the (early) reply state from lustre_pack_reply.
1436 (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
1437 ptlrpc_req_drop_rs(reqcopy);
1440 class_export_rpc_dec(reqcopy->rq_export);
1441 class_export_put(reqcopy->rq_export);
1443 sptlrpc_svc_ctx_decref(reqcopy);
1444 OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1446 ptlrpc_request_cache_free(reqcopy);
1450 /* Send early replies to everybody expiring within at_early_margin
1451 asking for at_extra time */
1452 static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1454 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1455 struct ptlrpc_request *rq, *n;
1456 struct list_head work_list;
1459 time64_t now = ktime_get_real_seconds();
1461 int first, counter = 0;
1464 spin_lock(&svcpt->scp_at_lock);
1465 if (svcpt->scp_at_check == 0) {
1466 spin_unlock(&svcpt->scp_at_lock);
1469 delay = ktime_ms_delta(ktime_get(), svcpt->scp_at_checktime);
1470 svcpt->scp_at_check = 0;
1472 if (array->paa_count == 0) {
1473 spin_unlock(&svcpt->scp_at_lock);
1477 /* The timer went off, but maybe the nearest rpc already completed. */
1478 first = array->paa_deadline - now;
1479 if (first > at_early_margin) {
1480 /* We've still got plenty of time. Reset the timer. */
1481 ptlrpc_at_set_timer(svcpt);
1482 spin_unlock(&svcpt->scp_at_lock);
1486 /* We're close to a timeout, and we don't know how much longer the
1487 server will take. Send early replies to everyone expiring soon. */
1488 INIT_LIST_HEAD(&work_list);
1490 div_u64_rem(array->paa_deadline, array->paa_size, &index);
1491 count = array->paa_count;
1493 count -= array->paa_reqs_count[index];
1494 list_for_each_entry_safe(rq, n,
1495 &array->paa_reqs_array[index],
1497 if (rq->rq_deadline > now + at_early_margin) {
1498 /* update the earliest deadline */
1499 if (deadline == -1 ||
1500 rq->rq_deadline < deadline)
1501 deadline = rq->rq_deadline;
1506 * ptlrpc_server_drop_request() may drop
1507 * refcount to 0 already. Let's check this and
1508 * don't add entry to work_list
1510 if (likely(atomic_inc_not_zero(&rq->rq_refcount))) {
1511 ptlrpc_at_remove_timed(rq);
1512 list_add(&rq->rq_timed_list, &work_list);
1514 ptlrpc_at_remove_timed(rq);
1520 if (++index >= array->paa_size)
1523 array->paa_deadline = deadline;
1524 /* we have a new earliest deadline, restart the timer */
1525 ptlrpc_at_set_timer(svcpt);
1527 spin_unlock(&svcpt->scp_at_lock);
1529 CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
1530 "replies\n", first, at_extra, counter);
1532 /* We're already past request deadlines before we even get a
1533 chance to send early replies */
1534 LCONSOLE_WARN("%s: This server is not able to keep up with "
1535 "request traffic (cpu-bound).\n",
1536 svcpt->scp_service->srv_name);
1537 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, delay=%lld\n",
1538 counter, svcpt->scp_nreqs_incoming,
1539 svcpt->scp_nreqs_active,
1540 at_get(&svcpt->scp_at_estimate), delay);
1543 /* we took additional refcount so entries can't be deleted from list, no
1544 * locking is needed */
1545 while (!list_empty(&work_list)) {
1546 rq = list_entry(work_list.next, struct ptlrpc_request,
1548 list_del_init(&rq->rq_timed_list);
1550 if (ptlrpc_at_send_early_reply(rq) == 0)
1551 ptlrpc_at_add_timed(rq);
1553 ptlrpc_server_drop_request(rq);
1556 RETURN(1); /* return "did_something" for liblustre */
1559 /* Check if we are already handling earlier incarnation of this request.
1560 * Called under &req->rq_export->exp_rpc_lock locked */
1561 static struct ptlrpc_request *
1562 ptlrpc_server_check_resend_in_progress(struct ptlrpc_request *req)
1564 struct ptlrpc_request *tmp = NULL;
1566 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ||
1567 (atomic_read(&req->rq_export->exp_rpc_count) == 0))
1570 /* bulk request are aborted upon reconnect, don't try to
1572 if (req->rq_bulk_write || req->rq_bulk_read)
1575 /* This list should not be longer than max_requests in
1576 * flights on the client, so it is not all that long.
1577 * Also we only hit this codepath in case of a resent
1578 * request which makes it even more rarely hit */
1579 list_for_each_entry(tmp, &req->rq_export->exp_reg_rpcs,
1581 /* Found duplicate one */
1582 if (tmp->rq_xid == req->rq_xid)
1585 list_for_each_entry(tmp, &req->rq_export->exp_hp_rpcs,
1587 /* Found duplicate one */
1588 if (tmp->rq_xid == req->rq_xid)
1594 DEBUG_REQ(D_HA, req, "Found duplicate req in processing");
1595 DEBUG_REQ(D_HA, tmp, "Request being processed");
1600 * Check if a request should be assigned with a high priority.
1602 * \retval < 0: error occurred
1603 * 0: normal RPC request
1604 * +1: high priority request
1606 static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt,
1607 struct ptlrpc_request *req)
1612 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL) {
1613 rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
1620 if (req->rq_export != NULL && req->rq_ops != NULL) {
1621 /* Perform request specific check. We should do this
1622 * check before the request is added into exp_hp_rpcs
1623 * list otherwise it may hit swab race at LU-1044. */
1624 if (req->rq_ops->hpreq_check != NULL) {
1625 rc = req->rq_ops->hpreq_check(req);
1626 if (rc == -ESTALE) {
1627 req->rq_status = rc;
1630 /** can only return error,
1631 * 0 for normal request,
1632 * or 1 for high priority request */
1640 /** Remove the request from the export list. */
1641 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req)
1644 if (req->rq_export) {
1645 /* refresh lock timeout again so that client has more
1646 * room to send lock cancel RPC. */
1647 if (req->rq_ops && req->rq_ops->hpreq_fini)
1648 req->rq_ops->hpreq_fini(req);
1650 spin_lock(&req->rq_export->exp_rpc_lock);
1651 list_del_init(&req->rq_exp_list);
1652 spin_unlock(&req->rq_export->exp_rpc_lock);
1657 static int ptlrpc_hpreq_check(struct ptlrpc_request *req)
1662 static struct ptlrpc_hpreq_ops ptlrpc_hpreq_common = {
1663 .hpreq_check = ptlrpc_hpreq_check,
1666 /* Hi-Priority RPC check by RPC operation code. */
1667 int ptlrpc_hpreq_handler(struct ptlrpc_request *req)
1669 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1671 /* Check for export to let only reconnects for not yet evicted
1672 * export to become a HP rpc. */
1673 if ((req->rq_export != NULL) &&
1674 (opc == OBD_PING || opc == MDS_CONNECT || opc == OST_CONNECT))
1675 req->rq_ops = &ptlrpc_hpreq_common;
1679 EXPORT_SYMBOL(ptlrpc_hpreq_handler);
1681 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1682 struct ptlrpc_request *req)
1686 struct ptlrpc_request *orig;
1689 rc = ptlrpc_server_hpreq_init(svcpt, req);
1694 ptlrpc_nrs_req_initialize(svcpt, req, hp);
1696 if (req->rq_export != NULL) {
1697 struct obd_export *exp = req->rq_export;
1699 /* do search for duplicated xid and the adding to the list
1701 spin_lock_bh(&exp->exp_rpc_lock);
1702 orig = ptlrpc_server_check_resend_in_progress(req);
1703 if (orig && likely(atomic_inc_not_zero(&orig->rq_refcount))) {
1706 spin_unlock_bh(&exp->exp_rpc_lock);
1709 * When the client resend request and the server has
1710 * the previous copy of it, we need to update deadlines,
1711 * to be sure that the client and the server have equal
1712 * request deadlines.
1715 spin_lock(&orig->rq_rqbd->rqbd_svcpt->scp_at_lock);
1716 linked = orig->rq_at_linked;
1718 ptlrpc_at_remove_timed(orig);
1719 spin_unlock(&orig->rq_rqbd->rqbd_svcpt->scp_at_lock);
1720 orig->rq_deadline = req->rq_deadline;
1722 ptlrpc_at_add_timed(orig);
1723 ptlrpc_server_drop_request(orig);
1724 ptlrpc_nrs_req_finalize(req);
1728 if (hp || req->rq_ops != NULL)
1729 list_add(&req->rq_exp_list, &exp->exp_hp_rpcs);
1731 list_add(&req->rq_exp_list, &exp->exp_reg_rpcs);
1732 spin_unlock_bh(&exp->exp_rpc_lock);
1735 /* the current thread is not the processing thread for this request
1736 * since that, but request is in exp_hp_list and can be find there.
1737 * Remove all relations between request and old thread. */
1738 req->rq_svc_thread->t_env->le_ses = NULL;
1739 req->rq_svc_thread = NULL;
1740 req->rq_session.lc_thread = NULL;
1742 ptlrpc_nrs_req_add(svcpt, req, hp);
1748 * Allow to handle high priority request
1749 * User can call it w/o any lock but need to hold
1750 * ptlrpc_service_part::scp_req_lock to get reliable result
1752 static bool ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1755 int running = svcpt->scp_nthrs_running;
1757 if (!nrs_svcpt_has_hp(svcpt))
1763 if (ptlrpc_nrs_req_throttling_nolock(svcpt, true))
1766 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1767 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1768 /* leave just 1 thread for normal RPCs */
1769 running = PTLRPC_NTHRS_INIT;
1770 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1774 if (svcpt->scp_nreqs_active >= running - 1)
1777 if (svcpt->scp_nhreqs_active == 0)
1780 return !ptlrpc_nrs_req_pending_nolock(svcpt, false) ||
1781 svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1784 static bool ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1787 return ptlrpc_server_allow_high(svcpt, force) &&
1788 ptlrpc_nrs_req_pending_nolock(svcpt, true);
1792 * Only allow normal priority requests on a service that has a high-priority
1793 * queue if forced (i.e. cleanup), if there are other high priority requests
1794 * already being processed (i.e. those threads can service more high-priority
1795 * requests), or if there are enough idle threads that a later thread can do
1796 * a high priority request.
1797 * User can call it w/o any lock but need to hold
1798 * ptlrpc_service_part::scp_req_lock to get reliable result
1800 static bool ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1803 int running = svcpt->scp_nthrs_running;
1804 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1805 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1806 /* leave just 1 thread for normal RPCs */
1807 running = PTLRPC_NTHRS_INIT;
1808 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1815 if (ptlrpc_nrs_req_throttling_nolock(svcpt, false))
1818 if (svcpt->scp_nreqs_active < running - 2)
1821 if (svcpt->scp_nreqs_active >= running - 1)
1824 return svcpt->scp_nhreqs_active > 0 || !nrs_svcpt_has_hp(svcpt);
1827 static bool ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1830 return ptlrpc_server_allow_normal(svcpt, force) &&
1831 ptlrpc_nrs_req_pending_nolock(svcpt, false);
1835 * Returns true if there are requests available in incoming
1836 * request queue for processing and it is allowed to fetch them.
1837 * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1838 * to get reliable result
1839 * \see ptlrpc_server_allow_normal
1840 * \see ptlrpc_server_allow high
1843 bool ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt,
1846 return ptlrpc_server_high_pending(svcpt, force) ||
1847 ptlrpc_server_normal_pending(svcpt, force);
1851 * Fetch a request for processing from queue of unprocessed requests.
1852 * Favors high-priority requests.
1853 * Returns a pointer to fetched request.
1855 static struct ptlrpc_request *
1856 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force)
1858 struct ptlrpc_request *req = NULL;
1861 spin_lock(&svcpt->scp_req_lock);
1863 if (ptlrpc_server_high_pending(svcpt, force)) {
1864 req = ptlrpc_nrs_req_get_nolock(svcpt, true, force);
1866 svcpt->scp_hreq_count++;
1871 if (ptlrpc_server_normal_pending(svcpt, force)) {
1872 req = ptlrpc_nrs_req_get_nolock(svcpt, false, force);
1874 svcpt->scp_hreq_count = 0;
1879 spin_unlock(&svcpt->scp_req_lock);
1883 svcpt->scp_nreqs_active++;
1885 svcpt->scp_nhreqs_active++;
1887 spin_unlock(&svcpt->scp_req_lock);
1889 if (likely(req->rq_export))
1890 class_export_rpc_inc(req->rq_export);
1896 * Handle freshly incoming reqs, add to timed early reply list,
1897 * pass on to regular request queue.
1898 * All incoming requests pass through here before getting into
1899 * ptlrpc_server_handle_req later on.
1901 static int ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt,
1902 struct ptlrpc_thread *thread)
1904 struct ptlrpc_service *svc = svcpt->scp_service;
1905 struct ptlrpc_request *req;
1910 spin_lock(&svcpt->scp_lock);
1911 if (list_empty(&svcpt->scp_req_incoming)) {
1912 spin_unlock(&svcpt->scp_lock);
1916 req = list_entry(svcpt->scp_req_incoming.next,
1917 struct ptlrpc_request, rq_list);
1918 list_del_init(&req->rq_list);
1919 svcpt->scp_nreqs_incoming--;
1920 /* Consider this still a "queued" request as far as stats are
1922 spin_unlock(&svcpt->scp_lock);
1924 /* go through security check/transform */
1925 rc = sptlrpc_svc_unwrap_request(req);
1929 case SECSVC_COMPLETE:
1930 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1939 * for null-flavored rpc, msg has been unpacked by sptlrpc, although
1940 * redo it wouldn't be harmful.
1942 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
1943 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
1945 CERROR("error unpacking request: ptl %d from %s "
1946 "x%llu\n", svc->srv_req_portal,
1947 libcfs_id2str(req->rq_peer), req->rq_xid);
1952 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1954 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
1955 "%llu\n", svc->srv_req_portal,
1956 libcfs_id2str(req->rq_peer), req->rq_xid);
1960 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
1961 lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
1962 CERROR("drop incoming rpc opc %u, x%llu\n",
1963 cfs_fail_val, req->rq_xid);
1968 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1969 CERROR("wrong packet type received (type=%u) from %s\n",
1970 lustre_msg_get_type(req->rq_reqmsg),
1971 libcfs_id2str(req->rq_peer));
1975 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1979 req->rq_bulk_write = 1;
1983 case MGS_CONFIG_READ:
1984 req->rq_bulk_read = 1;
1988 CDEBUG(D_RPCTRACE, "got req x%llu\n", req->rq_xid);
1990 req->rq_export = class_conn2export(
1991 lustre_msg_get_handle(req->rq_reqmsg));
1992 if (req->rq_export) {
1993 rc = ptlrpc_check_req(req);
1995 rc = sptlrpc_target_export_check(req->rq_export, req);
1997 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
1998 "illegal security flavor,");
2003 ptlrpc_update_export_timer(req->rq_export, 0);
2006 /* req_in handling should/must be fast */
2007 if (ktime_get_real_seconds() - req->rq_arrival_time.tv_sec > 5)
2008 DEBUG_REQ(D_WARNING, req, "Slow req_in handling %llds",
2009 (s64)(ktime_get_real_seconds() -
2010 req->rq_arrival_time.tv_sec));
2012 /* Set rpc server deadline and add it to the timed list */
2013 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
2014 MSGHDR_AT_SUPPORT) ?
2015 /* The max time the client expects us to take */
2016 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
2018 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
2019 if (unlikely(deadline == 0)) {
2020 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
2024 /* Skip early reply */
2025 if (OBD_FAIL_PRECHECK(OBD_FAIL_MDS_RESEND))
2026 req->rq_deadline += obd_timeout;
2028 req->rq_svc_thread = thread;
2029 if (thread != NULL) {
2030 /* initialize request session, it is needed for request
2031 * processing by target */
2032 rc = lu_context_init(&req->rq_session, LCT_SERVER_SESSION |
2035 CERROR("%s: failure to initialize session: rc = %d\n",
2036 thread->t_name, rc);
2039 req->rq_session.lc_thread = thread;
2040 lu_context_enter(&req->rq_session);
2041 thread->t_env->le_ses = &req->rq_session;
2044 ptlrpc_at_add_timed(req);
2046 /* Move it over to the request processing queue */
2047 rc = ptlrpc_server_request_add(svcpt, req);
2051 wake_up(&svcpt->scp_waitq);
2055 ptlrpc_server_finish_request(svcpt, req);
2061 * Main incoming request handling logic.
2062 * Calls handler function from service to do actual processing.
2064 static int ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
2065 struct ptlrpc_thread *thread)
2067 struct ptlrpc_service *svc = svcpt->scp_service;
2068 struct ptlrpc_request *request;
2078 request = ptlrpc_server_request_get(svcpt, false);
2079 if (request == NULL)
2082 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
2083 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
2084 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
2085 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
2087 if (unlikely(fail_opc)) {
2088 if (request->rq_export && request->rq_ops)
2089 OBD_FAIL_TIMEOUT(fail_opc, 4);
2092 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
2094 if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
2095 libcfs_debug_dumplog();
2097 work_start = ktime_get_real();
2098 arrived = timespec64_to_ktime(request->rq_arrival_time);
2099 timediff_usecs = ktime_us_delta(work_start, arrived);
2100 if (likely(svc->srv_stats != NULL)) {
2101 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
2103 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
2104 svcpt->scp_nreqs_incoming);
2105 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
2106 svcpt->scp_nreqs_active);
2107 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
2108 at_get(&svcpt->scp_at_estimate));
2111 if (likely(request->rq_export)) {
2112 if (unlikely(ptlrpc_check_req(request)))
2114 ptlrpc_update_export_timer(request->rq_export,
2115 div_u64(timediff_usecs,
2119 /* Discard requests queued for longer than the deadline.
2120 The deadline is increased if we send an early reply. */
2121 if (ktime_get_real_seconds() > request->rq_deadline) {
2122 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s: deadline %lld:%llds ago\n",
2123 libcfs_id2str(request->rq_peer),
2124 request->rq_deadline -
2125 request->rq_arrival_time.tv_sec,
2126 ktime_get_real_seconds() - request->rq_deadline);
2130 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
2131 "%s:%s+%d:%d:x%llu:%s:%d\n", current_comm(),
2132 (request->rq_export ?
2133 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2134 (request->rq_export ?
2135 atomic_read(&request->rq_export->exp_refcount) : -99),
2136 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
2137 libcfs_id2str(request->rq_peer),
2138 lustre_msg_get_opc(request->rq_reqmsg));
2140 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
2141 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
2143 CDEBUG(D_NET, "got req %llu\n", request->rq_xid);
2145 /* re-assign request and sesson thread to the current one */
2146 request->rq_svc_thread = thread;
2147 if (thread != NULL) {
2148 LASSERT(request->rq_session.lc_thread == NULL);
2149 request->rq_session.lc_thread = thread;
2150 thread->t_env->le_ses = &request->rq_session;
2152 svc->srv_ops.so_req_handler(request);
2154 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
2157 if (unlikely(ktime_get_real_seconds() > request->rq_deadline)) {
2158 DEBUG_REQ(D_WARNING, request,
2159 "Request took longer than estimated (%lld:%llds); "
2160 "client may timeout.",
2161 request->rq_deadline -
2162 request->rq_arrival_time.tv_sec,
2163 ktime_get_real_seconds() - request->rq_deadline);
2166 work_end = ktime_get_real();
2167 timediff_usecs = ktime_us_delta(work_end, work_start);
2168 arrived_usecs = ktime_us_delta(work_end, arrived);
2169 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
2170 "%s:%s+%d:%d:x%llu:%s:%d Request processed in %lldus "
2171 "(%lldus total) trans %llu rc %d/%d\n",
2173 (request->rq_export ?
2174 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2175 (request->rq_export ?
2176 atomic_read(&request->rq_export->exp_refcount) : -99),
2177 lustre_msg_get_status(request->rq_reqmsg),
2179 libcfs_id2str(request->rq_peer),
2180 lustre_msg_get_opc(request->rq_reqmsg),
2183 (request->rq_repmsg ?
2184 lustre_msg_get_transno(request->rq_repmsg) :
2185 request->rq_transno),
2187 (request->rq_repmsg ?
2188 lustre_msg_get_status(request->rq_repmsg) : -999));
2189 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
2190 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
2191 int opc = opcode_offset(op);
2192 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
2193 LASSERT(opc < LUSTRE_MAX_OPCODES);
2194 lprocfs_counter_add(svc->srv_stats,
2195 opc + EXTRA_MAX_OPCODES,
2199 if (unlikely(request->rq_early_count)) {
2200 DEBUG_REQ(D_ADAPTTO, request,
2201 "sent %d early replies before finishing in %llds",
2202 request->rq_early_count,
2203 div_u64(arrived_usecs, USEC_PER_SEC));
2206 ptlrpc_server_finish_active_request(svcpt, request);
2212 * An internal function to process a single reply state object.
2214 static int ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
2216 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
2217 struct ptlrpc_service *svc = svcpt->scp_service;
2218 struct obd_export *exp;
2223 exp = rs->rs_export;
2225 LASSERT(rs->rs_difficult);
2226 LASSERT(rs->rs_scheduled);
2227 LASSERT(list_empty(&rs->rs_list));
2229 /* The disk commit callback holds exp_uncommitted_replies_lock while it
2230 * iterates over newly committed replies, removing them from
2231 * exp_uncommitted_replies. It then drops this lock and schedules the
2232 * replies it found for handling here.
2234 * We can avoid contention for exp_uncommitted_replies_lock between the
2235 * HRT threads and further commit callbacks by checking rs_committed
2236 * which is set in the commit callback while it holds both
2237 * rs_lock and exp_uncommitted_reples.
2239 * If we see rs_committed clear, the commit callback _may_ not have
2240 * handled this reply yet and we race with it to grab
2241 * exp_uncommitted_replies_lock before removing the reply from
2242 * exp_uncommitted_replies. Note that if we lose the race and the
2243 * reply has already been removed, list_del_init() is a noop.
2245 * If we see rs_committed set, we know the commit callback is handling,
2246 * or has handled this reply since store reordering might allow us to
2247 * see rs_committed set out of sequence. But since this is done
2248 * holding rs_lock, we can be sure it has all completed once we hold
2249 * rs_lock, which we do right next.
2251 if (!rs->rs_committed) {
2252 /* if rs was commited, no need to convert locks, don't check
2253 * rs_committed here because rs may never be added into
2254 * exp_uncommitted_replies and this flag never be set, see
2255 * target_send_reply() */
2256 if (rs->rs_convert_lock &&
2257 rs->rs_transno > exp->exp_last_committed) {
2258 struct ldlm_lock *lock;
2259 struct ldlm_lock *ack_locks[RS_MAX_LOCKS] = { NULL };
2261 spin_lock(&rs->rs_lock);
2262 if (rs->rs_convert_lock &&
2263 rs->rs_transno > exp->exp_last_committed) {
2264 nlocks = rs->rs_nlocks;
2265 while (nlocks-- > 0) {
2267 * NB don't assume rs is always handled
2268 * by the same service thread (see
2269 * ptlrpc_hr_select, so REP-ACK hr may
2270 * race with trans commit, while the
2271 * latter will release locks, get locks
2272 * here early to convert to COS mode
2275 lock = ldlm_handle2lock(
2276 &rs->rs_locks[nlocks]);
2278 ack_locks[nlocks] = lock;
2279 rs->rs_modes[nlocks] = LCK_COS;
2281 nlocks = rs->rs_nlocks;
2282 rs->rs_convert_lock = 0;
2283 /* clear rs_scheduled so that commit callback
2284 * can schedule again */
2285 rs->rs_scheduled = 0;
2286 spin_unlock(&rs->rs_lock);
2288 while (nlocks-- > 0) {
2289 lock = ack_locks[nlocks];
2290 ldlm_lock_mode_downgrade(lock, LCK_COS);
2291 LDLM_LOCK_PUT(lock);
2295 spin_unlock(&rs->rs_lock);
2298 spin_lock(&exp->exp_uncommitted_replies_lock);
2299 list_del_init(&rs->rs_obd_list);
2300 spin_unlock(&exp->exp_uncommitted_replies_lock);
2303 spin_lock(&exp->exp_lock);
2304 /* Noop if removed already */
2305 list_del_init(&rs->rs_exp_list);
2306 spin_unlock(&exp->exp_lock);
2308 spin_lock(&rs->rs_lock);
2310 been_handled = rs->rs_handled;
2313 nlocks = rs->rs_nlocks; /* atomic "steal", but */
2314 rs->rs_nlocks = 0; /* locks still on rs_locks! */
2316 if (nlocks == 0 && !been_handled) {
2317 /* If we see this, we should already have seen the warning
2318 * in mds_steal_ack_locks() */
2319 CDEBUG(D_HA, "All locks stolen from rs %p x%lld.t%lld"
2322 rs->rs_xid, rs->rs_transno, rs->rs_opc,
2323 libcfs_nid2str(exp->exp_connection->c_peer.nid));
2326 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
2327 spin_unlock(&rs->rs_lock);
2329 if (!been_handled && rs->rs_on_net) {
2330 LNetMDUnlink(rs->rs_md_h);
2331 /* Ignore return code; we're racing with completion */
2334 while (nlocks-- > 0)
2335 ldlm_lock_decref(&rs->rs_locks[nlocks],
2336 rs->rs_modes[nlocks]);
2338 spin_lock(&rs->rs_lock);
2341 rs->rs_scheduled = 0;
2342 rs->rs_convert_lock = 0;
2344 if (!rs->rs_on_net) {
2346 spin_unlock(&rs->rs_lock);
2348 class_export_put (exp);
2349 rs->rs_export = NULL;
2350 ptlrpc_rs_decref(rs);
2351 if (atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
2352 svc->srv_is_stopping)
2353 wake_up_all(&svcpt->scp_waitq);
2357 /* still on the net; callback will schedule */
2358 spin_unlock(&rs->rs_lock);
2363 static void ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
2365 int avail = svcpt->scp_nrqbds_posted;
2366 int low_water = test_req_buffer_pressure ? 0 :
2367 svcpt->scp_service->srv_nbuf_per_group / 2;
2369 /* NB I'm not locking; just looking. */
2371 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
2372 * allowed the request history to grow out of control. We could put a
2373 * sanity check on that here and cull some history if we need the
2376 if (avail <= low_water)
2377 ptlrpc_grow_req_bufs(svcpt, 1);
2379 if (svcpt->scp_service->srv_stats) {
2380 lprocfs_counter_add(svcpt->scp_service->srv_stats,
2381 PTLRPC_REQBUF_AVAIL_CNTR, avail);
2385 static int ptlrpc_retry_rqbds(void *arg)
2387 struct ptlrpc_service_part *svcpt = (struct ptlrpc_service_part *)arg;
2389 svcpt->scp_rqbd_timeout = 0;
2393 static inline int ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
2395 return svcpt->scp_nreqs_active <
2396 svcpt->scp_nthrs_running - 1 -
2397 (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL);
2401 * allowed to create more threads
2402 * user can call it w/o any lock but need to hold
2403 * ptlrpc_service_part::scp_lock to get reliable result
2405 static inline int ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
2407 return svcpt->scp_nthrs_running +
2408 svcpt->scp_nthrs_starting <
2409 svcpt->scp_service->srv_nthrs_cpt_limit;
2413 * too many requests and allowed to create more threads
2415 static inline int ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
2417 return !ptlrpc_threads_enough(svcpt) &&
2418 ptlrpc_threads_increasable(svcpt);
2421 static inline int ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2423 return thread_is_stopping(thread) ||
2424 thread->t_svcpt->scp_service->srv_is_stopping;
2427 /* stop the highest numbered thread if there are too many threads running */
2428 static inline bool ptlrpc_thread_should_stop(struct ptlrpc_thread *thread)
2430 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2432 return thread->t_id >= svcpt->scp_service->srv_nthrs_cpt_limit &&
2433 thread->t_id == svcpt->scp_thr_nextid - 1;
2436 static void ptlrpc_stop_thread(struct ptlrpc_thread *thread)
2438 CDEBUG(D_INFO, "Stopping thread %s #%u\n",
2439 thread->t_svcpt->scp_service->srv_thread_name, thread->t_id);
2440 thread_add_flags(thread, SVC_STOPPING);
2443 static inline void ptlrpc_thread_stop(struct ptlrpc_thread *thread)
2445 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2447 spin_lock(&svcpt->scp_lock);
2448 if (ptlrpc_thread_should_stop(thread)) {
2449 ptlrpc_stop_thread(thread);
2450 svcpt->scp_thr_nextid--;
2452 spin_unlock(&svcpt->scp_lock);
2455 static inline int ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
2457 return !list_empty(&svcpt->scp_rqbd_idle) &&
2458 svcpt->scp_rqbd_timeout == 0;
2462 ptlrpc_at_check(struct ptlrpc_service_part *svcpt)
2464 return svcpt->scp_at_check;
2468 * If a thread runs too long or spends to much time on a single request,
2469 * we want to know about it, so we set up a delayed work item as a watchdog.
2470 * If it fires, we display a stack trace of the delayed thread,
2471 * providing we aren't rate-limited
2473 * Watchdog stack traces are limited to 3 per 'libcfs_watchdog_ratelimit'
2476 static struct ratelimit_state watchdog_limit;
2478 static void ptlrpc_watchdog_fire(struct work_struct *w)
2480 struct ptlrpc_thread *thread = container_of(w, struct ptlrpc_thread,
2482 u64 ms_lapse = ktime_ms_delta(ktime_get(), thread->t_touched);
2483 u32 ms_frac = do_div(ms_lapse, MSEC_PER_SEC);
2485 if (!__ratelimit(&watchdog_limit)) {
2486 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",
2487 thread->t_task->comm, thread->t_task->pid,
2490 libcfs_debug_dumpstack(thread->t_task);
2492 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",
2493 thread->t_task->comm, thread->t_task->pid,
2494 ms_lapse, ms_frac, libcfs_watchdog_ratelimit);
2498 static void ptlrpc_watchdog_init(struct delayed_work *work, time_t time)
2500 INIT_DELAYED_WORK(work, ptlrpc_watchdog_fire);
2501 schedule_delayed_work(work, cfs_time_seconds(time));
2504 static void ptlrpc_watchdog_disable(struct delayed_work *work)
2506 cancel_delayed_work_sync(work);
2509 static void ptlrpc_watchdog_touch(struct delayed_work *work, time_t time)
2511 struct ptlrpc_thread *thread = container_of(&work->work,
2512 struct ptlrpc_thread,
2514 thread->t_touched = ktime_get();
2515 mod_delayed_work(system_wq, work, cfs_time_seconds(time));
2519 * requests wait on preprocessing
2520 * user can call it w/o any lock but need to hold
2521 * ptlrpc_service_part::scp_lock to get reliable result
2524 ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
2526 return !list_empty(&svcpt->scp_req_incoming);
2529 static __attribute__((__noinline__)) int
2530 ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
2531 struct ptlrpc_thread *thread)
2533 /* Don't exit while there are replies to be handled */
2534 struct l_wait_info lwi = LWI_TIMEOUT(svcpt->scp_rqbd_timeout,
2535 ptlrpc_retry_rqbds, svcpt);
2537 ptlrpc_watchdog_disable(&thread->t_watchdog);
2541 l_wait_event_exclusive_head(svcpt->scp_waitq,
2542 ptlrpc_thread_stopping(thread) ||
2543 ptlrpc_server_request_incoming(svcpt) ||
2544 ptlrpc_server_request_pending(svcpt, false) ||
2545 ptlrpc_rqbd_pending(svcpt) ||
2546 ptlrpc_at_check(svcpt), &lwi);
2548 if (ptlrpc_thread_stopping(thread))
2551 ptlrpc_watchdog_touch(&thread->t_watchdog,
2552 ptlrpc_server_get_timeout(svcpt));
2557 * Main thread body for service threads.
2558 * Waits in a loop waiting for new requests to process to appear.
2559 * Every time an incoming requests is added to its queue, a waitq
2560 * is woken up and one of the threads will handle it.
2562 static int ptlrpc_main(void *arg)
2564 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
2565 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2566 struct ptlrpc_service *svc = svcpt->scp_service;
2567 struct ptlrpc_reply_state *rs;
2568 struct group_info *ginfo = NULL;
2570 int counter = 0, rc = 0;
2573 thread->t_task = current;
2574 thread->t_pid = current_pid();
2575 unshare_fs_struct();
2577 if (svc->srv_cpt_bind) {
2578 rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
2580 CWARN("%s: failed to bind %s on CPT %d\n",
2581 svc->srv_name, thread->t_name, svcpt->scp_cpt);
2585 ginfo = groups_alloc(0);
2591 set_current_groups(ginfo);
2592 put_group_info(ginfo);
2594 if (svc->srv_ops.so_thr_init != NULL) {
2595 rc = svc->srv_ops.so_thr_init(thread);
2606 rc = lu_context_init(&env->le_ctx,
2607 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2611 thread->t_env = env;
2612 env->le_ctx.lc_thread = thread;
2613 env->le_ctx.lc_cookie = 0x6;
2615 while (!list_empty(&svcpt->scp_rqbd_idle)) {
2616 rc = ptlrpc_server_post_idle_rqbds(svcpt);
2620 CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
2621 svc->srv_name, svcpt->scp_cpt, rc);
2625 /* Alloc reply state structure for this one */
2626 OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2632 spin_lock(&svcpt->scp_lock);
2634 LASSERT(thread_is_starting(thread));
2635 thread_clear_flags(thread, SVC_STARTING);
2637 LASSERT(svcpt->scp_nthrs_starting == 1);
2638 svcpt->scp_nthrs_starting--;
2640 /* SVC_STOPPING may already be set here if someone else is trying
2641 * to stop the service while this new thread has been dynamically
2642 * forked. We still set SVC_RUNNING to let our creator know that
2643 * we are now running, however we will exit as soon as possible */
2644 thread_add_flags(thread, SVC_RUNNING);
2645 svcpt->scp_nthrs_running++;
2646 spin_unlock(&svcpt->scp_lock);
2648 /* wake up our creator in case he's still waiting. */
2649 wake_up(&thread->t_ctl_waitq);
2651 thread->t_touched = ktime_get();
2652 ptlrpc_watchdog_init(&thread->t_watchdog,
2653 ptlrpc_server_get_timeout(svcpt));
2655 spin_lock(&svcpt->scp_rep_lock);
2656 list_add(&rs->rs_list, &svcpt->scp_rep_idle);
2657 wake_up(&svcpt->scp_rep_waitq);
2658 spin_unlock(&svcpt->scp_rep_lock);
2660 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2661 svcpt->scp_nthrs_running);
2663 /* XXX maintain a list of all managed devices: insert here */
2664 while (!ptlrpc_thread_stopping(thread)) {
2665 if (ptlrpc_wait_event(svcpt, thread))
2668 ptlrpc_check_rqbd_pool(svcpt);
2670 if (ptlrpc_threads_need_create(svcpt)) {
2671 /* Ignore return code - we tried... */
2672 ptlrpc_start_thread(svcpt, 0);
2675 /* reset le_ses to initial state */
2677 /* Process all incoming reqs before handling any */
2678 if (ptlrpc_server_request_incoming(svcpt)) {
2679 lu_context_enter(&env->le_ctx);
2680 ptlrpc_server_handle_req_in(svcpt, thread);
2681 lu_context_exit(&env->le_ctx);
2683 /* but limit ourselves in case of flood */
2684 if (counter++ < 100)
2689 if (ptlrpc_at_check(svcpt))
2690 ptlrpc_at_check_timed(svcpt);
2692 if (ptlrpc_server_request_pending(svcpt, false)) {
2693 lu_context_enter(&env->le_ctx);
2694 ptlrpc_server_handle_request(svcpt, thread);
2695 lu_context_exit(&env->le_ctx);
2698 if (ptlrpc_rqbd_pending(svcpt) &&
2699 ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
2700 /* I just failed to repost request buffers.
2701 * Wait for a timeout (unless something else
2702 * happens) before I try again */
2703 svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
2704 CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
2705 svcpt->scp_nrqbds_posted);
2708 /* If the number of threads has been tuned downward and this
2709 * thread should be stopped, then stop in reverse order so the
2710 * the threads always have contiguous thread index values.
2712 if (unlikely(ptlrpc_thread_should_stop(thread)))
2713 ptlrpc_thread_stop(thread);
2716 ptlrpc_watchdog_disable(&thread->t_watchdog);
2719 /* deconstruct service thread state created by ptlrpc_start_thread() */
2720 if (svc->srv_ops.so_thr_done != NULL)
2721 svc->srv_ops.so_thr_done(thread);
2724 lu_context_fini(&env->le_ctx);
2728 CDEBUG(D_RPCTRACE, "%s: service thread [%p:%u] %d exiting: rc = %d\n",
2729 thread->t_name, thread, thread->t_pid, thread->t_id, rc);
2731 spin_lock(&svcpt->scp_lock);
2732 if (thread_test_and_clear_flags(thread, SVC_STARTING))
2733 svcpt->scp_nthrs_starting--;
2735 if (thread_test_and_clear_flags(thread, SVC_RUNNING)) {
2736 /* must know immediately */
2737 svcpt->scp_nthrs_running--;
2741 thread_add_flags(thread, SVC_STOPPED);
2743 wake_up(&thread->t_ctl_waitq);
2744 spin_unlock(&svcpt->scp_lock);
2749 static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
2750 struct list_head *replies)
2754 spin_lock(&hrt->hrt_lock);
2756 list_splice_init(&hrt->hrt_queue, replies);
2757 result = ptlrpc_hr.hr_stopping || !list_empty(replies);
2759 spin_unlock(&hrt->hrt_lock);
2764 * Main body of "handle reply" function.
2765 * It processes acked reply states
2767 static int ptlrpc_hr_main(void *arg)
2769 struct ptlrpc_hr_thread *hrt = (struct ptlrpc_hr_thread *)arg;
2770 struct ptlrpc_hr_partition *hrp = hrt->hrt_partition;
2771 struct list_head replies;
2774 INIT_LIST_HEAD(&replies);
2775 unshare_fs_struct();
2777 rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
2779 char threadname[20];
2781 snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
2782 hrp->hrp_cpt, hrt->hrt_id);
2783 CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
2784 threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
2787 atomic_inc(&hrp->hrp_nstarted);
2788 wake_up(&ptlrpc_hr.hr_waitq);
2790 while (!ptlrpc_hr.hr_stopping) {
2791 l_wait_condition(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
2793 while (!list_empty(&replies)) {
2794 struct ptlrpc_reply_state *rs;
2796 rs = list_entry(replies.prev,
2797 struct ptlrpc_reply_state,
2799 list_del_init(&rs->rs_list);
2800 ptlrpc_handle_rs(rs);
2804 atomic_inc(&hrp->hrp_nstopped);
2805 wake_up(&ptlrpc_hr.hr_waitq);
2810 static void ptlrpc_stop_hr_threads(void)
2812 struct ptlrpc_hr_partition *hrp;
2816 ptlrpc_hr.hr_stopping = 1;
2818 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2819 if (hrp->hrp_thrs == NULL)
2820 continue; /* uninitialized */
2821 for (j = 0; j < hrp->hrp_nthrs; j++)
2822 wake_up_all(&hrp->hrp_thrs[j].hrt_waitq);
2825 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2826 if (hrp->hrp_thrs == NULL)
2827 continue; /* uninitialized */
2828 wait_event(ptlrpc_hr.hr_waitq,
2829 atomic_read(&hrp->hrp_nstopped) ==
2830 atomic_read(&hrp->hrp_nstarted));
2834 static int ptlrpc_start_hr_threads(void)
2836 struct ptlrpc_hr_partition *hrp;
2841 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2844 for (j = 0; j < hrp->hrp_nthrs; j++) {
2845 struct ptlrpc_hr_thread *hrt = &hrp->hrp_thrs[j];
2846 struct task_struct *task;
2848 task = kthread_run(ptlrpc_hr_main,
2850 "ptlrpc_hr%02d_%03d",
2859 wait_event(ptlrpc_hr.hr_waitq,
2860 atomic_read(&hrp->hrp_nstarted) == j);
2863 CERROR("cannot start reply handler thread %d:%d: "
2864 "rc = %d\n", i, j, rc);
2865 ptlrpc_stop_hr_threads();
2873 static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
2875 struct l_wait_info lwi = { 0 };
2876 struct ptlrpc_thread *thread;
2877 struct list_head zombie;
2881 CDEBUG(D_INFO, "Stopping threads for service %s\n",
2882 svcpt->scp_service->srv_name);
2884 INIT_LIST_HEAD(&zombie);
2885 spin_lock(&svcpt->scp_lock);
2886 /* let the thread know that we would like it to stop asap */
2887 list_for_each_entry(thread, &svcpt->scp_threads, t_link)
2888 ptlrpc_stop_thread(thread);
2890 wake_up_all(&svcpt->scp_waitq);
2892 while (!list_empty(&svcpt->scp_threads)) {
2893 thread = list_entry(svcpt->scp_threads.next,
2894 struct ptlrpc_thread, t_link);
2895 if (thread_is_stopped(thread)) {
2896 list_del(&thread->t_link);
2897 list_add(&thread->t_link, &zombie);
2900 spin_unlock(&svcpt->scp_lock);
2902 CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
2903 svcpt->scp_service->srv_thread_name, thread->t_id);
2904 l_wait_event(thread->t_ctl_waitq,
2905 thread_is_stopped(thread), &lwi);
2907 spin_lock(&svcpt->scp_lock);
2910 spin_unlock(&svcpt->scp_lock);
2912 while (!list_empty(&zombie)) {
2913 thread = list_entry(zombie.next,
2914 struct ptlrpc_thread, t_link);
2915 list_del(&thread->t_link);
2916 OBD_FREE_PTR(thread);
2922 * Stops all threads of a particular service \a svc
2924 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2926 struct ptlrpc_service_part *svcpt;
2930 ptlrpc_service_for_each_part(svcpt, i, svc) {
2931 if (svcpt->scp_service != NULL)
2932 ptlrpc_svcpt_stop_threads(svcpt);
2938 int ptlrpc_start_threads(struct ptlrpc_service *svc)
2945 /* We require 2 threads min, see note in ptlrpc_server_handle_request */
2946 LASSERT(svc->srv_nthrs_cpt_init >= PTLRPC_NTHRS_INIT);
2948 for (i = 0; i < svc->srv_ncpts; i++) {
2949 for (j = 0; j < svc->srv_nthrs_cpt_init; j++) {
2950 rc = ptlrpc_start_thread(svc->srv_parts[i], 1);
2956 /* We have enough threads, don't start more. b=15759 */
2963 CERROR("cannot start %s thread #%d_%d: rc %d\n",
2964 svc->srv_thread_name, i, j, rc);
2965 ptlrpc_stop_all_threads(svc);
2969 int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
2971 struct l_wait_info lwi = { 0 };
2972 struct ptlrpc_thread *thread;
2973 struct ptlrpc_service *svc;
2974 struct task_struct *task;
2978 LASSERT(svcpt != NULL);
2980 svc = svcpt->scp_service;
2982 CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n",
2983 svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running,
2984 svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit);
2987 if (unlikely(svc->srv_is_stopping))
2990 if (!ptlrpc_threads_increasable(svcpt) ||
2991 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
2992 svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1))
2995 OBD_CPT_ALLOC_PTR(thread, svc->srv_cptable, svcpt->scp_cpt);
2998 init_waitqueue_head(&thread->t_ctl_waitq);
3000 spin_lock(&svcpt->scp_lock);
3001 if (!ptlrpc_threads_increasable(svcpt)) {
3002 spin_unlock(&svcpt->scp_lock);
3003 OBD_FREE_PTR(thread);
3007 if (svcpt->scp_nthrs_starting != 0) {
3008 /* serialize starting because some modules (obdfilter)
3009 * might require unique and contiguous t_id */
3010 LASSERT(svcpt->scp_nthrs_starting == 1);
3011 spin_unlock(&svcpt->scp_lock);
3012 OBD_FREE_PTR(thread);
3014 CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n",
3015 svc->srv_thread_name, svcpt->scp_thr_nextid);
3020 CDEBUG(D_INFO, "Creating thread %s #%d race, retry later\n",
3021 svc->srv_thread_name, svcpt->scp_thr_nextid);
3025 svcpt->scp_nthrs_starting++;
3026 thread->t_id = svcpt->scp_thr_nextid++;
3027 thread_add_flags(thread, SVC_STARTING);
3028 thread->t_svcpt = svcpt;
3030 list_add(&thread->t_link, &svcpt->scp_threads);
3031 spin_unlock(&svcpt->scp_lock);
3033 if (svcpt->scp_cpt >= 0) {
3034 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s%02d_%03d",
3035 svc->srv_thread_name, svcpt->scp_cpt, thread->t_id);
3037 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s_%04d",
3038 svc->srv_thread_name, thread->t_id);
3041 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name);
3042 task = kthread_run(ptlrpc_main, thread, "%s", thread->t_name);
3045 CERROR("cannot start thread '%s': rc = %d\n",
3046 thread->t_name, rc);
3047 spin_lock(&svcpt->scp_lock);
3048 --svcpt->scp_nthrs_starting;
3049 if (thread_is_stopping(thread)) {
3050 /* this ptlrpc_thread is being hanled
3051 * by ptlrpc_svcpt_stop_threads now
3053 thread_add_flags(thread, SVC_STOPPED);
3054 wake_up(&thread->t_ctl_waitq);
3055 spin_unlock(&svcpt->scp_lock);
3057 list_del(&thread->t_link);
3058 spin_unlock(&svcpt->scp_lock);
3059 OBD_FREE_PTR(thread);
3067 l_wait_event(thread->t_ctl_waitq,
3068 thread_is_running(thread) || thread_is_stopped(thread),
3071 rc = thread_is_stopped(thread) ? thread->t_id : 0;
3075 int ptlrpc_hr_init(void)
3077 struct ptlrpc_hr_partition *hrp;
3078 struct ptlrpc_hr_thread *hrt;
3085 memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
3086 ptlrpc_hr.hr_cpt_table = cfs_cpt_table;
3088 ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
3090 if (ptlrpc_hr.hr_partitions == NULL)
3093 ratelimit_state_init(&watchdog_limit,
3094 cfs_time_seconds(libcfs_watchdog_ratelimit), 3);
3096 init_waitqueue_head(&ptlrpc_hr.hr_waitq);
3098 weight = cpumask_weight(topology_sibling_cpumask(smp_processor_id()));
3100 cfs_percpt_for_each(hrp, cpt, ptlrpc_hr.hr_partitions) {
3103 atomic_set(&hrp->hrp_nstarted, 0);
3104 atomic_set(&hrp->hrp_nstopped, 0);
3106 hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, cpt);
3107 hrp->hrp_nthrs /= weight;
3108 if (hrp->hrp_nthrs == 0)
3111 OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, cpt,
3112 hrp->hrp_nthrs * sizeof(*hrt));
3113 if (hrp->hrp_thrs == NULL)
3114 GOTO(out, rc = -ENOMEM);
3116 for (i = 0; i < hrp->hrp_nthrs; i++) {
3117 hrt = &hrp->hrp_thrs[i];
3120 hrt->hrt_partition = hrp;
3121 init_waitqueue_head(&hrt->hrt_waitq);
3122 spin_lock_init(&hrt->hrt_lock);
3123 INIT_LIST_HEAD(&hrt->hrt_queue);
3127 rc = ptlrpc_start_hr_threads();
3134 void ptlrpc_hr_fini(void)
3136 struct ptlrpc_hr_partition *hrp;
3139 if (ptlrpc_hr.hr_partitions == NULL)
3142 ptlrpc_stop_hr_threads();
3144 cfs_percpt_for_each(hrp, cpt, ptlrpc_hr.hr_partitions) {
3145 if (hrp->hrp_thrs != NULL) {
3146 OBD_FREE(hrp->hrp_thrs,
3147 hrp->hrp_nthrs * sizeof(hrp->hrp_thrs[0]));
3151 cfs_percpt_free(ptlrpc_hr.hr_partitions);
3152 ptlrpc_hr.hr_partitions = NULL;
3157 * Wait until all already scheduled replies are processed.
3159 static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
3163 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
3166 rc = l_wait_event(svcpt->scp_waitq,
3167 atomic_read(&svcpt->scp_nreps_difficult) == 0, &lwi);
3170 CWARN("Unexpectedly long timeout %s %p\n",
3171 svcpt->scp_service->srv_name, svcpt->scp_service);
3176 ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
3178 struct ptlrpc_service_part *svcpt;
3181 /* early disarm AT timer... */
3182 ptlrpc_service_for_each_part(svcpt, i, svc) {
3183 if (svcpt->scp_service != NULL)
3184 del_timer(&svcpt->scp_at_timer);
3189 ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
3191 struct ptlrpc_service_part *svcpt;
3192 struct ptlrpc_request_buffer_desc *rqbd;
3193 struct l_wait_info lwi;
3197 /* All history will be culled when the next request buffer is
3198 * freed in ptlrpc_service_purge_all() */
3199 svc->srv_hist_nrqbds_cpt_max = 0;
3201 rc = LNetClearLazyPortal(svc->srv_req_portal);
3204 ptlrpc_service_for_each_part(svcpt, i, svc) {
3205 if (svcpt->scp_service == NULL)
3208 /* Unlink all the request buffers. This forces a 'final'
3209 * event with its 'unlink' flag set for each posted rqbd */
3210 list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
3212 rc = LNetMDUnlink(rqbd->rqbd_md_h);
3213 LASSERT(rc == 0 || rc == -ENOENT);
3217 ptlrpc_service_for_each_part(svcpt, i, svc) {
3218 if (svcpt->scp_service == NULL)
3221 /* Wait for the network to release any buffers
3222 * it's currently filling */
3223 spin_lock(&svcpt->scp_lock);
3224 while (svcpt->scp_nrqbds_posted != 0) {
3225 spin_unlock(&svcpt->scp_lock);
3226 /* Network access will complete in finite time but
3227 * the HUGE timeout lets us CWARN for visibility
3228 * of sluggish NALs */
3229 lwi = LWI_TIMEOUT_INTERVAL(
3230 cfs_time_seconds(LONG_UNLINK),
3231 cfs_time_seconds(1), NULL, NULL);
3232 rc = l_wait_event(svcpt->scp_waitq,
3233 svcpt->scp_nrqbds_posted == 0, &lwi);
3234 if (rc == -ETIMEDOUT) {
3235 CWARN("Service %s waiting for "
3236 "request buffers\n",
3237 svcpt->scp_service->srv_name);
3239 spin_lock(&svcpt->scp_lock);
3241 spin_unlock(&svcpt->scp_lock);
3246 ptlrpc_service_purge_all(struct ptlrpc_service *svc)
3248 struct ptlrpc_service_part *svcpt;
3249 struct ptlrpc_request_buffer_desc *rqbd;
3250 struct ptlrpc_request *req;
3251 struct ptlrpc_reply_state *rs;
3254 ptlrpc_service_for_each_part(svcpt, i, svc) {
3255 if (svcpt->scp_service == NULL)
3258 spin_lock(&svcpt->scp_rep_lock);
3259 while (!list_empty(&svcpt->scp_rep_active)) {
3260 rs = list_entry(svcpt->scp_rep_active.next,
3261 struct ptlrpc_reply_state, rs_list);
3262 spin_lock(&rs->rs_lock);
3263 ptlrpc_schedule_difficult_reply(rs);
3264 spin_unlock(&rs->rs_lock);
3266 spin_unlock(&svcpt->scp_rep_lock);
3268 /* purge the request queue. NB No new replies (rqbds
3269 * all unlinked) and no service threads, so I'm the only
3270 * thread noodling the request queue now */
3271 while (!list_empty(&svcpt->scp_req_incoming)) {
3272 req = list_entry(svcpt->scp_req_incoming.next,
3273 struct ptlrpc_request, rq_list);
3275 list_del(&req->rq_list);
3276 svcpt->scp_nreqs_incoming--;
3277 ptlrpc_server_finish_request(svcpt, req);
3280 while (ptlrpc_server_request_pending(svcpt, true)) {
3281 req = ptlrpc_server_request_get(svcpt, true);
3282 ptlrpc_server_finish_active_request(svcpt, req);
3285 LASSERT(list_empty(&svcpt->scp_rqbd_posted));
3286 LASSERT(svcpt->scp_nreqs_incoming == 0);
3287 LASSERT(svcpt->scp_nreqs_active == 0);
3288 /* history should have been culled by
3289 * ptlrpc_server_finish_request */
3290 LASSERT(svcpt->scp_hist_nrqbds == 0);
3292 /* Now free all the request buffers since nothing
3293 * references them any more... */
3295 while (!list_empty(&svcpt->scp_rqbd_idle)) {
3296 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
3297 struct ptlrpc_request_buffer_desc,
3299 ptlrpc_free_rqbd(rqbd);
3301 ptlrpc_wait_replies(svcpt);
3303 while (!list_empty(&svcpt->scp_rep_idle)) {
3304 rs = list_entry(svcpt->scp_rep_idle.next,
3305 struct ptlrpc_reply_state,
3307 list_del(&rs->rs_list);
3308 OBD_FREE_LARGE(rs, svc->srv_max_reply_size);
3314 ptlrpc_service_free(struct ptlrpc_service *svc)
3316 struct ptlrpc_service_part *svcpt;
3317 struct ptlrpc_at_array *array;
3320 ptlrpc_service_for_each_part(svcpt, i, svc) {
3321 if (svcpt->scp_service == NULL)
3324 /* In case somebody rearmed this in the meantime */
3325 del_timer(&svcpt->scp_at_timer);
3326 array = &svcpt->scp_at_array;
3328 if (array->paa_reqs_array != NULL) {
3329 OBD_FREE(array->paa_reqs_array,
3330 sizeof(struct list_head) * array->paa_size);
3331 array->paa_reqs_array = NULL;
3334 if (array->paa_reqs_count != NULL) {
3335 OBD_FREE(array->paa_reqs_count,
3336 sizeof(__u32) * array->paa_size);
3337 array->paa_reqs_count = NULL;
3341 ptlrpc_service_for_each_part(svcpt, i, svc)
3342 OBD_FREE_PTR(svcpt);
3344 if (svc->srv_cpts != NULL)
3345 cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts);
3347 OBD_FREE(svc, offsetof(struct ptlrpc_service,
3348 srv_parts[svc->srv_ncpts]));
3351 int ptlrpc_unregister_service(struct ptlrpc_service *service)
3355 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
3357 service->srv_is_stopping = 1;
3359 mutex_lock(&ptlrpc_all_services_mutex);
3360 list_del_init(&service->srv_list);
3361 mutex_unlock(&ptlrpc_all_services_mutex);
3363 ptlrpc_service_del_atimer(service);
3364 ptlrpc_stop_all_threads(service);
3366 ptlrpc_service_unlink_rqbd(service);
3367 ptlrpc_service_purge_all(service);
3368 ptlrpc_service_nrs_cleanup(service);
3370 ptlrpc_lprocfs_unregister_service(service);
3371 ptlrpc_sysfs_unregister_service(service);
3373 ptlrpc_service_free(service);
3377 EXPORT_SYMBOL(ptlrpc_unregister_service);
3380 * Returns 0 if the service is healthy.
3382 * Right now, it just checks to make sure that requests aren't languishing
3383 * in the queue. We'll use this health check to govern whether a node needs
3384 * to be shot, so it's intentionally non-aggressive. */
3385 static int ptlrpc_svcpt_health_check(struct ptlrpc_service_part *svcpt)
3387 struct ptlrpc_request *request = NULL;
3388 struct timespec64 right_now;
3389 struct timespec64 timediff;
3391 ktime_get_real_ts64(&right_now);
3393 spin_lock(&svcpt->scp_req_lock);
3394 /* How long has the next entry been waiting? */
3395 if (ptlrpc_server_high_pending(svcpt, true))
3396 request = ptlrpc_nrs_req_peek_nolock(svcpt, true);
3397 else if (ptlrpc_server_normal_pending(svcpt, true))
3398 request = ptlrpc_nrs_req_peek_nolock(svcpt, false);
3400 if (request == NULL) {
3401 spin_unlock(&svcpt->scp_req_lock);
3405 timediff = timespec64_sub(right_now, request->rq_arrival_time);
3406 spin_unlock(&svcpt->scp_req_lock);
3408 if ((timediff.tv_sec) >
3409 (AT_OFF ? obd_timeout * 3 / 2 : at_max)) {
3410 CERROR("%s: unhealthy - request has been waiting %llds\n",
3411 svcpt->scp_service->srv_name, (s64)timediff.tv_sec);
3419 ptlrpc_service_health_check(struct ptlrpc_service *svc)
3421 struct ptlrpc_service_part *svcpt;
3427 ptlrpc_service_for_each_part(svcpt, i, svc) {
3428 int rc = ptlrpc_svcpt_health_check(svcpt);
3435 EXPORT_SYMBOL(ptlrpc_service_health_check);