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 <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 (test_req_buffer_pressure ||
144 svcpt->scp_nrqbds_posted >= svc->srv_nbuf_per_group ||
145 (svc->srv_nrqbds_max != 0 &&
146 svcpt->scp_nrqbds_total > svc->srv_nrqbds_max))
149 rqbd = ptlrpc_alloc_rqbd(svcpt);
152 CERROR("%s: Can't allocate request buffer\n",
159 spin_lock(&svcpt->scp_lock);
161 LASSERT(svcpt->scp_rqbd_allocating == 1);
162 svcpt->scp_rqbd_allocating--;
164 spin_unlock(&svcpt->scp_lock);
167 "%s: allocate %d new %d-byte reqbufs (%d/%d left), rc = %d\n",
168 svc->srv_name, i, svc->srv_buf_size, svcpt->scp_nrqbds_posted,
169 svcpt->scp_nrqbds_total, rc);
173 rc = ptlrpc_server_post_idle_rqbds(svcpt);
179 * Part of Rep-Ack logic.
180 * Puts a lock and its mode into reply state assotiated to request reply.
183 ptlrpc_save_lock(struct ptlrpc_request *req, struct lustre_handle *lock,
184 int mode, bool no_ack, bool convert_lock)
186 struct ptlrpc_reply_state *rs = req->rq_reply_state;
190 LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
192 idx = rs->rs_nlocks++;
193 rs->rs_locks[idx] = *lock;
194 rs->rs_modes[idx] = mode;
195 rs->rs_difficult = 1;
196 rs->rs_no_ack = no_ack;
197 rs->rs_convert_lock = convert_lock;
199 EXPORT_SYMBOL(ptlrpc_save_lock);
202 struct ptlrpc_hr_partition;
204 struct ptlrpc_hr_thread {
205 int hrt_id; /* thread ID */
207 wait_queue_head_t hrt_waitq;
208 struct list_head hrt_queue;
209 struct ptlrpc_hr_partition *hrt_partition;
212 struct ptlrpc_hr_partition {
213 /* # of started threads */
214 atomic_t hrp_nstarted;
215 /* # of stopped threads */
216 atomic_t hrp_nstopped;
217 /* cpu partition id */
219 /* round-robin rotor for choosing thread */
221 /* total number of threads on this partition */
224 struct ptlrpc_hr_thread *hrp_thrs;
227 #define HRT_RUNNING 0
228 #define HRT_STOPPING 1
230 struct ptlrpc_hr_service {
231 /* CPU partition table, it's just cfs_cpt_table for now */
232 struct cfs_cpt_table *hr_cpt_table;
233 /** controller sleep waitq */
234 wait_queue_head_t hr_waitq;
235 unsigned int hr_stopping;
236 /** roundrobin rotor for non-affinity service */
237 unsigned int hr_rotor;
239 struct ptlrpc_hr_partition **hr_partitions;
243 struct list_head rsb_replies;
244 unsigned int rsb_n_replies;
245 struct ptlrpc_service_part *rsb_svcpt;
248 /** reply handling service. */
249 static struct ptlrpc_hr_service ptlrpc_hr;
252 * maximum mumber of replies scheduled in one batch
254 #define MAX_SCHEDULED 256
257 * Initialize a reply batch.
261 static void rs_batch_init(struct rs_batch *b)
263 memset(b, 0, sizeof *b);
264 INIT_LIST_HEAD(&b->rsb_replies);
268 * Choose an hr thread to dispatch requests to.
270 static struct ptlrpc_hr_thread *
271 ptlrpc_hr_select(struct ptlrpc_service_part *svcpt)
273 struct ptlrpc_hr_partition *hrp;
276 if (svcpt->scp_cpt >= 0 &&
277 svcpt->scp_service->srv_cptable == ptlrpc_hr.hr_cpt_table) {
278 /* directly match partition */
279 hrp = ptlrpc_hr.hr_partitions[svcpt->scp_cpt];
282 rotor = ptlrpc_hr.hr_rotor++;
283 rotor %= cfs_cpt_number(ptlrpc_hr.hr_cpt_table);
285 hrp = ptlrpc_hr.hr_partitions[rotor];
288 rotor = hrp->hrp_rotor++;
289 return &hrp->hrp_thrs[rotor % hrp->hrp_nthrs];
293 * Dispatch all replies accumulated in the batch to one from
294 * dedicated reply handling threads.
298 static void rs_batch_dispatch(struct rs_batch *b)
300 if (b->rsb_n_replies != 0) {
301 struct ptlrpc_hr_thread *hrt;
303 hrt = ptlrpc_hr_select(b->rsb_svcpt);
305 spin_lock(&hrt->hrt_lock);
306 list_splice_init(&b->rsb_replies, &hrt->hrt_queue);
307 spin_unlock(&hrt->hrt_lock);
309 wake_up(&hrt->hrt_waitq);
310 b->rsb_n_replies = 0;
315 * Add a reply to a batch.
316 * Add one reply object to a batch, schedule batched replies if overload.
321 static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs)
323 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
325 if (svcpt != b->rsb_svcpt || b->rsb_n_replies >= MAX_SCHEDULED) {
326 if (b->rsb_svcpt != NULL) {
327 rs_batch_dispatch(b);
328 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
330 spin_lock(&svcpt->scp_rep_lock);
331 b->rsb_svcpt = svcpt;
333 spin_lock(&rs->rs_lock);
334 rs->rs_scheduled_ever = 1;
335 if (rs->rs_scheduled == 0) {
336 list_move(&rs->rs_list, &b->rsb_replies);
337 rs->rs_scheduled = 1;
340 rs->rs_committed = 1;
341 spin_unlock(&rs->rs_lock);
345 * Reply batch finalization.
346 * Dispatch remaining replies from the batch
347 * and release remaining spinlock.
351 static void rs_batch_fini(struct rs_batch *b)
353 if (b->rsb_svcpt != NULL) {
354 rs_batch_dispatch(b);
355 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
359 #define DECLARE_RS_BATCH(b) struct rs_batch b
363 * Put reply state into a queue for processing because we received
364 * ACK from the client
366 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
368 struct ptlrpc_hr_thread *hrt;
371 LASSERT(list_empty(&rs->rs_list));
373 hrt = ptlrpc_hr_select(rs->rs_svcpt);
375 spin_lock(&hrt->hrt_lock);
376 list_add_tail(&rs->rs_list, &hrt->hrt_queue);
377 spin_unlock(&hrt->hrt_lock);
379 wake_up(&hrt->hrt_waitq);
384 ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs)
388 assert_spin_locked(&rs->rs_svcpt->scp_rep_lock);
389 assert_spin_locked(&rs->rs_lock);
390 LASSERT (rs->rs_difficult);
391 rs->rs_scheduled_ever = 1; /* flag any notification attempt */
393 if (rs->rs_scheduled) { /* being set up or already notified */
398 rs->rs_scheduled = 1;
399 list_del_init(&rs->rs_list);
400 ptlrpc_dispatch_difficult_reply(rs);
403 EXPORT_SYMBOL(ptlrpc_schedule_difficult_reply);
405 void ptlrpc_commit_replies(struct obd_export *exp)
407 struct ptlrpc_reply_state *rs, *nxt;
408 DECLARE_RS_BATCH(batch);
411 rs_batch_init(&batch);
412 /* Find any replies that have been committed and get their service
413 * to attend to complete them. */
415 /* CAVEAT EMPTOR: spinlock ordering!!! */
416 spin_lock(&exp->exp_uncommitted_replies_lock);
417 list_for_each_entry_safe(rs, nxt, &exp->exp_uncommitted_replies,
419 LASSERT (rs->rs_difficult);
420 /* VBR: per-export last_committed */
421 LASSERT(rs->rs_export);
422 if (rs->rs_transno <= exp->exp_last_committed) {
423 list_del_init(&rs->rs_obd_list);
424 rs_batch_add(&batch, rs);
427 spin_unlock(&exp->exp_uncommitted_replies_lock);
428 rs_batch_fini(&batch);
433 ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt)
435 struct ptlrpc_request_buffer_desc *rqbd;
440 spin_lock(&svcpt->scp_lock);
442 if (list_empty(&svcpt->scp_rqbd_idle)) {
443 spin_unlock(&svcpt->scp_lock);
447 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
448 struct ptlrpc_request_buffer_desc,
450 list_del(&rqbd->rqbd_list);
452 /* assume we will post successfully */
453 svcpt->scp_nrqbds_posted++;
454 list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_posted);
456 spin_unlock(&svcpt->scp_lock);
458 rc = ptlrpc_register_rqbd(rqbd);
465 spin_lock(&svcpt->scp_lock);
467 svcpt->scp_nrqbds_posted--;
468 list_del(&rqbd->rqbd_list);
469 list_add_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
471 /* Don't complain if no request buffers are posted right now; LNET
472 * won't drop requests because we set the portal lazy! */
474 spin_unlock(&svcpt->scp_lock);
479 static void ptlrpc_at_timer(unsigned long castmeharder)
481 struct ptlrpc_service_part *svcpt;
483 svcpt = (struct ptlrpc_service_part *)castmeharder;
485 svcpt->scp_at_check = 1;
486 svcpt->scp_at_checktime = cfs_time_current();
487 wake_up(&svcpt->scp_waitq);
491 ptlrpc_server_nthreads_check(struct ptlrpc_service *svc,
492 struct ptlrpc_service_conf *conf)
494 struct ptlrpc_service_thr_conf *tc = &conf->psc_thr;
501 * Common code for estimating & validating threads number.
502 * CPT affinity service could have percpt thread-pool instead
503 * of a global thread-pool, which means user might not always
504 * get the threads number they give it in conf::tc_nthrs_user
505 * even they did set. It's because we need to validate threads
506 * number for each CPT to guarantee each pool will have enough
507 * threads to keep the service healthy.
509 init = PTLRPC_NTHRS_INIT + (svc->srv_ops.so_hpreq_handler != NULL);
510 init = max_t(int, init, tc->tc_nthrs_init);
512 /* NB: please see comments in lustre_lnet.h for definition
513 * details of these members */
514 LASSERT(tc->tc_nthrs_max != 0);
516 if (tc->tc_nthrs_user != 0) {
517 /* In case there is a reason to test a service with many
518 * threads, we give a less strict check here, it can
519 * be up to 8 * nthrs_max */
520 total = min(tc->tc_nthrs_max * 8, tc->tc_nthrs_user);
521 nthrs = total / svc->srv_ncpts;
522 init = max(init, nthrs);
526 total = tc->tc_nthrs_max;
527 if (tc->tc_nthrs_base == 0) {
528 /* don't care about base threads number per partition,
529 * this is most for non-affinity service */
530 nthrs = total / svc->srv_ncpts;
534 nthrs = tc->tc_nthrs_base;
535 if (svc->srv_ncpts == 1) {
538 /* NB: Increase the base number if it's single partition
539 * and total number of cores/HTs is larger or equal to 4.
540 * result will always < 2 * nthrs_base */
541 weight = cfs_cpt_weight(svc->srv_cptable, CFS_CPT_ANY);
542 for (i = 1; (weight >> (i + 1)) != 0 && /* >= 4 cores/HTs */
543 (tc->tc_nthrs_base >> i) != 0; i++)
544 nthrs += tc->tc_nthrs_base >> i;
547 if (tc->tc_thr_factor != 0) {
548 int factor = tc->tc_thr_factor;
552 * User wants to increase number of threads with for
553 * each CPU core/HT, most likely the factor is larger than
554 * one thread/core because service threads are supposed to
555 * be blocked by lock or wait for IO.
558 * Amdahl's law says that adding processors wouldn't give
559 * a linear increasing of parallelism, so it's nonsense to
560 * have too many threads no matter how many cores/HTs
563 if (cpumask_weight(topology_sibling_cpumask(smp_processor_id())) > 1) {
564 /* weight is # of HTs */
565 /* depress thread factor for hyper-thread */
566 factor = factor - (factor >> 1) + (factor >> 3);
569 weight = cfs_cpt_weight(svc->srv_cptable, 0);
571 for (; factor > 0 && weight > 0; factor--, weight -= fade)
572 nthrs += min(weight, fade) * factor;
575 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
576 nthrs = max(tc->tc_nthrs_base,
577 tc->tc_nthrs_max / svc->srv_ncpts);
580 nthrs = max(nthrs, tc->tc_nthrs_init);
581 svc->srv_nthrs_cpt_limit = nthrs;
582 svc->srv_nthrs_cpt_init = init;
584 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
585 CDEBUG(D_OTHER, "%s: This service may have more threads (%d) "
586 "than the given soft limit (%d)\n",
587 svc->srv_name, nthrs * svc->srv_ncpts,
593 * Initialize percpt data for a service
596 ptlrpc_service_part_init(struct ptlrpc_service *svc,
597 struct ptlrpc_service_part *svcpt, int cpt)
599 struct ptlrpc_at_array *array;
604 svcpt->scp_cpt = cpt;
605 INIT_LIST_HEAD(&svcpt->scp_threads);
607 /* rqbd and incoming request queue */
608 spin_lock_init(&svcpt->scp_lock);
609 INIT_LIST_HEAD(&svcpt->scp_rqbd_idle);
610 INIT_LIST_HEAD(&svcpt->scp_rqbd_posted);
611 INIT_LIST_HEAD(&svcpt->scp_req_incoming);
612 init_waitqueue_head(&svcpt->scp_waitq);
613 /* history request & rqbd list */
614 INIT_LIST_HEAD(&svcpt->scp_hist_reqs);
615 INIT_LIST_HEAD(&svcpt->scp_hist_rqbds);
617 /* acitve requests and hp requests */
618 spin_lock_init(&svcpt->scp_req_lock);
621 spin_lock_init(&svcpt->scp_rep_lock);
622 INIT_LIST_HEAD(&svcpt->scp_rep_active);
623 INIT_LIST_HEAD(&svcpt->scp_rep_idle);
624 init_waitqueue_head(&svcpt->scp_rep_waitq);
625 atomic_set(&svcpt->scp_nreps_difficult, 0);
627 /* adaptive timeout */
628 spin_lock_init(&svcpt->scp_at_lock);
629 array = &svcpt->scp_at_array;
631 size = at_est2timeout(at_max);
632 array->paa_size = size;
633 array->paa_count = 0;
634 array->paa_deadline = -1;
636 /* allocate memory for scp_at_array (ptlrpc_at_array) */
637 OBD_CPT_ALLOC(array->paa_reqs_array,
638 svc->srv_cptable, cpt, sizeof(struct list_head) * size);
639 if (array->paa_reqs_array == NULL)
642 for (index = 0; index < size; index++)
643 INIT_LIST_HEAD(&array->paa_reqs_array[index]);
645 OBD_CPT_ALLOC(array->paa_reqs_count,
646 svc->srv_cptable, cpt, sizeof(__u32) * size);
647 if (array->paa_reqs_count == NULL)
650 setup_timer(&svcpt->scp_at_timer, ptlrpc_at_timer,
651 (unsigned long)svcpt);
653 /* At SOW, service time should be quick; 10s seems generous. If client
654 * timeout is less than this, we'll be sending an early reply. */
655 at_init(&svcpt->scp_at_estimate, 10, 0);
657 /* assign this before call ptlrpc_grow_req_bufs */
658 svcpt->scp_service = svc;
659 /* Now allocate the request buffers, but don't post them now */
660 rc = ptlrpc_grow_req_bufs(svcpt, 0);
661 /* We shouldn't be under memory pressure at startup, so
662 * fail if we can't allocate all our buffers at this time. */
669 if (array->paa_reqs_count != NULL) {
670 OBD_FREE(array->paa_reqs_count, sizeof(__u32) * size);
671 array->paa_reqs_count = NULL;
674 if (array->paa_reqs_array != NULL) {
675 OBD_FREE(array->paa_reqs_array,
676 sizeof(struct list_head) * array->paa_size);
677 array->paa_reqs_array = NULL;
684 * Initialize service on a given portal.
685 * This includes starting serving threads , allocating and posting rqbds and
688 struct ptlrpc_service *
689 ptlrpc_register_service(struct ptlrpc_service_conf *conf,
691 struct proc_dir_entry *proc_entry)
693 struct ptlrpc_service_cpt_conf *cconf = &conf->psc_cpt;
694 struct ptlrpc_service *service;
695 struct ptlrpc_service_part *svcpt;
696 struct cfs_cpt_table *cptable;
704 LASSERT(conf->psc_buf.bc_nbufs > 0);
705 LASSERT(conf->psc_buf.bc_buf_size >=
706 conf->psc_buf.bc_req_max_size + SPTLRPC_MAX_PAYLOAD);
707 LASSERT(conf->psc_thr.tc_ctx_tags != 0);
709 cptable = cconf->cc_cptable;
711 cptable = cfs_cpt_table;
713 if (!conf->psc_thr.tc_cpu_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;
753 service->srv_cpt_bits = 0; /* it's zero already, easy to read... */
754 while ((1 << service->srv_cpt_bits) < cfs_cpt_number(cptable))
755 service->srv_cpt_bits++;
758 spin_lock_init(&service->srv_lock);
759 service->srv_name = conf->psc_name;
760 service->srv_watchdog_factor = conf->psc_watchdog_factor;
761 INIT_LIST_HEAD(&service->srv_list); /* for safty of cleanup */
763 /* buffer configuration */
764 service->srv_nbuf_per_group = test_req_buffer_pressure ?
765 1 : conf->psc_buf.bc_nbufs;
766 /* do not limit max number of rqbds by default */
767 service->srv_nrqbds_max = 0;
769 service->srv_max_req_size = conf->psc_buf.bc_req_max_size +
771 service->srv_buf_size = conf->psc_buf.bc_buf_size;
772 service->srv_rep_portal = conf->psc_buf.bc_rep_portal;
773 service->srv_req_portal = conf->psc_buf.bc_req_portal;
775 /* Increase max reply size to next power of two */
776 service->srv_max_reply_size = 1;
777 while (service->srv_max_reply_size <
778 conf->psc_buf.bc_rep_max_size + SPTLRPC_MAX_PAYLOAD)
779 service->srv_max_reply_size <<= 1;
781 service->srv_thread_name = conf->psc_thr.tc_thr_name;
782 service->srv_ctx_tags = conf->psc_thr.tc_ctx_tags;
783 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
784 service->srv_ops = conf->psc_ops;
786 for (i = 0; i < ncpts; i++) {
787 if (!conf->psc_thr.tc_cpu_affinity)
790 cpt = cpts != NULL ? cpts[i] : i;
792 OBD_CPT_ALLOC(svcpt, cptable, cpt, sizeof(*svcpt));
794 GOTO(failed, rc = -ENOMEM);
796 service->srv_parts[i] = svcpt;
797 rc = ptlrpc_service_part_init(service, svcpt, cpt);
802 ptlrpc_server_nthreads_check(service, conf);
804 rc = LNetSetLazyPortal(service->srv_req_portal);
807 mutex_lock(&ptlrpc_all_services_mutex);
808 list_add(&service->srv_list, &ptlrpc_all_services);
809 mutex_unlock(&ptlrpc_all_services_mutex);
812 rc = ptlrpc_sysfs_register_service(parent, service);
817 if (proc_entry != NULL)
818 ptlrpc_lprocfs_register_service(proc_entry, service);
820 rc = ptlrpc_service_nrs_setup(service);
824 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
825 service->srv_name, service->srv_req_portal);
827 rc = ptlrpc_start_threads(service);
829 CERROR("Failed to start threads for service %s: %d\n",
830 service->srv_name, rc);
836 ptlrpc_unregister_service(service);
839 EXPORT_SYMBOL(ptlrpc_register_service);
842 * to actually free the request, must be called without holding svc_lock.
843 * note it's caller's responsibility to unlink req->rq_list.
845 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
847 LASSERT(atomic_read(&req->rq_refcount) == 0);
848 LASSERT(list_empty(&req->rq_timed_list));
850 /* DEBUG_REQ() assumes the reply state of a request with a valid
851 * ref will not be destroyed until that reference is dropped. */
852 ptlrpc_req_drop_rs(req);
854 sptlrpc_svc_ctx_decref(req);
856 if (req != &req->rq_rqbd->rqbd_req) {
857 /* NB request buffers use an embedded
858 * req if the incoming req unlinked the
859 * MD; this isn't one of them! */
860 ptlrpc_request_cache_free(req);
865 * drop a reference count of the request. if it reaches 0, we either
866 * put it into history list, or free it immediately.
868 void ptlrpc_server_drop_request(struct ptlrpc_request *req)
870 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
871 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
872 struct ptlrpc_service *svc = svcpt->scp_service;
874 struct list_head *tmp;
875 struct list_head *nxt;
877 if (!atomic_dec_and_test(&req->rq_refcount))
880 if (req->rq_session.lc_state == LCS_ENTERED) {
881 lu_context_exit(&req->rq_session);
882 lu_context_fini(&req->rq_session);
885 if (req->rq_at_linked) {
886 spin_lock(&svcpt->scp_at_lock);
887 /* recheck with lock, in case it's unlinked by
888 * ptlrpc_at_check_timed() */
889 if (likely(req->rq_at_linked))
890 ptlrpc_at_remove_timed(req);
891 spin_unlock(&svcpt->scp_at_lock);
894 LASSERT(list_empty(&req->rq_timed_list));
896 /* finalize request */
897 if (req->rq_export) {
898 class_export_put(req->rq_export);
899 req->rq_export = NULL;
902 spin_lock(&svcpt->scp_lock);
904 list_add(&req->rq_list, &rqbd->rqbd_reqs);
906 refcount = --(rqbd->rqbd_refcount);
908 /* request buffer is now idle: add to history */
909 list_del(&rqbd->rqbd_list);
911 list_add_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds);
912 svcpt->scp_hist_nrqbds++;
914 /* cull some history?
915 * I expect only about 1 or 2 rqbds need to be recycled here */
916 while (svcpt->scp_hist_nrqbds > svc->srv_hist_nrqbds_cpt_max) {
917 rqbd = list_entry(svcpt->scp_hist_rqbds.next,
918 struct ptlrpc_request_buffer_desc,
921 list_del(&rqbd->rqbd_list);
922 svcpt->scp_hist_nrqbds--;
924 /* remove rqbd's reqs from svc's req history while
925 * I've got the service lock */
926 list_for_each(tmp, &rqbd->rqbd_reqs) {
927 req = list_entry(tmp, struct ptlrpc_request,
929 /* Track the highest culled req seq */
930 if (req->rq_history_seq >
931 svcpt->scp_hist_seq_culled) {
932 svcpt->scp_hist_seq_culled =
935 list_del(&req->rq_history_list);
938 spin_unlock(&svcpt->scp_lock);
940 list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
941 req = list_entry(rqbd->rqbd_reqs.next,
942 struct ptlrpc_request,
944 list_del(&req->rq_list);
945 ptlrpc_server_free_request(req);
948 spin_lock(&svcpt->scp_lock);
950 * now all reqs including the embedded req has been
951 * disposed, schedule request buffer for re-use
952 * or free it to drain some in excess.
954 LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) == 0);
955 if (svcpt->scp_nrqbds_posted >=
956 svc->srv_nbuf_per_group ||
957 (svc->srv_nrqbds_max != 0 &&
958 svcpt->scp_nrqbds_total > svc->srv_nrqbds_max) ||
959 test_req_buffer_pressure) {
960 /* like in ptlrpc_free_rqbd() */
961 svcpt->scp_nrqbds_total--;
962 OBD_FREE_LARGE(rqbd->rqbd_buffer,
966 list_add_tail(&rqbd->rqbd_list,
967 &svcpt->scp_rqbd_idle);
971 spin_unlock(&svcpt->scp_lock);
972 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
973 /* If we are low on memory, we are not interested in history */
974 list_del(&req->rq_list);
975 list_del_init(&req->rq_history_list);
977 /* Track the highest culled req seq */
978 if (req->rq_history_seq > svcpt->scp_hist_seq_culled)
979 svcpt->scp_hist_seq_culled = req->rq_history_seq;
981 spin_unlock(&svcpt->scp_lock);
983 ptlrpc_server_free_request(req);
985 spin_unlock(&svcpt->scp_lock);
989 /** Change request export and move hp request from old export to new */
990 void ptlrpc_request_change_export(struct ptlrpc_request *req,
991 struct obd_export *export)
993 if (req->rq_export != NULL) {
994 LASSERT(!list_empty(&req->rq_exp_list));
995 /* remove rq_exp_list from last export */
996 spin_lock_bh(&req->rq_export->exp_rpc_lock);
997 list_del_init(&req->rq_exp_list);
998 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
999 /* export has one reference already, so it`s safe to
1000 * add req to export queue here and get another
1001 * reference for request later */
1002 spin_lock_bh(&export->exp_rpc_lock);
1003 if (req->rq_ops != NULL) /* hp request */
1004 list_add(&req->rq_exp_list, &export->exp_hp_rpcs);
1006 list_add(&req->rq_exp_list, &export->exp_reg_rpcs);
1007 spin_unlock_bh(&export->exp_rpc_lock);
1009 class_export_rpc_dec(req->rq_export);
1010 class_export_put(req->rq_export);
1013 /* request takes one export refcount */
1014 req->rq_export = class_export_get(export);
1015 class_export_rpc_inc(export);
1021 * to finish a request: stop sending more early replies, and release
1024 static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt,
1025 struct ptlrpc_request *req)
1027 ptlrpc_server_hpreq_fini(req);
1029 ptlrpc_server_drop_request(req);
1033 * to finish an active request: stop sending more early replies, and release
1034 * the request. should be called after we finished handling the request.
1036 static void ptlrpc_server_finish_active_request(
1037 struct ptlrpc_service_part *svcpt,
1038 struct ptlrpc_request *req)
1040 spin_lock(&svcpt->scp_req_lock);
1041 ptlrpc_nrs_req_stop_nolock(req);
1042 svcpt->scp_nreqs_active--;
1044 svcpt->scp_nhreqs_active--;
1045 spin_unlock(&svcpt->scp_req_lock);
1047 ptlrpc_nrs_req_finalize(req);
1049 if (req->rq_export != NULL)
1050 class_export_rpc_dec(req->rq_export);
1052 ptlrpc_server_finish_request(svcpt, req);
1056 * This function makes sure dead exports are evicted in a timely manner.
1057 * This function is only called when some export receives a message (i.e.,
1058 * the network is up.)
1060 void ptlrpc_update_export_timer(struct obd_export *exp, time64_t extra_delay)
1062 struct obd_export *oldest_exp;
1063 time64_t oldest_time, new_time;
1069 /* Compensate for slow machines, etc, by faking our request time
1070 into the future. Although this can break the strict time-ordering
1071 of the list, we can be really lazy here - we don't have to evict
1072 at the exact right moment. Eventually, all silent exports
1073 will make it to the top of the list. */
1075 /* Do not pay attention on 1sec or smaller renewals. */
1076 new_time = ktime_get_real_seconds() + extra_delay;
1077 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
1080 exp->exp_last_request_time = new_time;
1082 /* exports may get disconnected from the chain even though the
1083 export has references, so we must keep the spin lock while
1084 manipulating the lists */
1085 spin_lock(&exp->exp_obd->obd_dev_lock);
1087 if (list_empty(&exp->exp_obd_chain_timed)) {
1088 /* this one is not timed */
1089 spin_unlock(&exp->exp_obd->obd_dev_lock);
1093 list_move_tail(&exp->exp_obd_chain_timed,
1094 &exp->exp_obd->obd_exports_timed);
1096 oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
1097 struct obd_export, exp_obd_chain_timed);
1098 oldest_time = oldest_exp->exp_last_request_time;
1099 spin_unlock(&exp->exp_obd->obd_dev_lock);
1101 if (exp->exp_obd->obd_recovering) {
1102 /* be nice to everyone during recovery */
1107 /* Note - racing to start/reset the obd_eviction timer is safe */
1108 if (exp->exp_obd->obd_eviction_timer == 0) {
1109 /* Check if the oldest entry is expired. */
1110 if (ktime_get_real_seconds() >
1111 oldest_time + PING_EVICT_TIMEOUT + extra_delay) {
1112 /* We need a second timer, in case the net was down and
1113 * it just came back. Since the pinger may skip every
1114 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
1115 * we better wait for 3.
1117 exp->exp_obd->obd_eviction_timer =
1118 ktime_get_real_seconds() + 3 * PING_INTERVAL;
1119 CDEBUG(D_HA, "%s: Think about evicting %s from %lld\n",
1120 exp->exp_obd->obd_name,
1121 obd_export_nid2str(oldest_exp), oldest_time);
1124 if (ktime_get_real_seconds() >
1125 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
1126 /* The evictor won't evict anyone who we've heard from
1127 * recently, so we don't have to check before we start
1130 if (!ping_evictor_wake(exp))
1131 exp->exp_obd->obd_eviction_timer = 0;
1139 * Sanity check request \a req.
1140 * Return 0 if all is ok, error code otherwise.
1142 static int ptlrpc_check_req(struct ptlrpc_request *req)
1144 struct obd_device *obd = req->rq_export->exp_obd;
1147 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
1148 req->rq_export->exp_conn_cnt)) {
1149 DEBUG_REQ(D_RPCTRACE, req,
1150 "DROPPING req from old connection %d < %d",
1151 lustre_msg_get_conn_cnt(req->rq_reqmsg),
1152 req->rq_export->exp_conn_cnt);
1155 if (unlikely(obd == NULL || obd->obd_fail)) {
1156 /* Failing over, don't handle any more reqs,
1157 * send error response instead. */
1158 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
1159 req, (obd != NULL) ? obd->obd_name : "unknown");
1161 } else if (lustre_msg_get_flags(req->rq_reqmsg) &
1162 (MSG_REPLAY | MSG_REQ_REPLAY_DONE) &&
1163 !obd->obd_recovering) {
1164 DEBUG_REQ(D_ERROR, req,
1165 "Invalid replay without recovery");
1166 class_fail_export(req->rq_export);
1168 } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 &&
1169 !obd->obd_recovering) {
1170 DEBUG_REQ(D_ERROR, req, "Invalid req with transno "
1171 "%llu without recovery",
1172 lustre_msg_get_transno(req->rq_reqmsg));
1173 class_fail_export(req->rq_export);
1177 if (unlikely(rc < 0)) {
1178 req->rq_status = rc;
1184 static void ptlrpc_at_set_timer(struct ptlrpc_service_part *svcpt)
1186 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1189 if (array->paa_count == 0) {
1190 del_timer(&svcpt->scp_at_timer);
1194 /* Set timer for closest deadline */
1195 next = (__s32)(array->paa_deadline - ktime_get_real_seconds() -
1198 ptlrpc_at_timer((unsigned long)svcpt);
1200 mod_timer(&svcpt->scp_at_timer, cfs_time_shift(next));
1201 CDEBUG(D_INFO, "armed %s at %+ds\n",
1202 svcpt->scp_service->srv_name, next);
1206 /* Add rpc to early reply check list */
1207 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
1209 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1210 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1211 struct ptlrpc_request *rq = NULL;
1217 if (req->rq_no_reply)
1220 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
1223 spin_lock(&svcpt->scp_at_lock);
1224 LASSERT(list_empty(&req->rq_timed_list));
1226 div_u64_rem(req->rq_deadline, array->paa_size, &index);
1227 if (array->paa_reqs_count[index] > 0) {
1228 /* latest rpcs will have the latest deadlines in the list,
1229 * so search backward. */
1230 list_for_each_entry_reverse(rq,
1231 &array->paa_reqs_array[index],
1233 if (req->rq_deadline >= rq->rq_deadline) {
1234 list_add(&req->rq_timed_list,
1235 &rq->rq_timed_list);
1241 /* Add the request at the head of the list */
1242 if (list_empty(&req->rq_timed_list))
1243 list_add(&req->rq_timed_list,
1244 &array->paa_reqs_array[index]);
1246 spin_lock(&req->rq_lock);
1247 req->rq_at_linked = 1;
1248 spin_unlock(&req->rq_lock);
1249 req->rq_at_index = index;
1250 array->paa_reqs_count[index]++;
1252 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
1253 array->paa_deadline = req->rq_deadline;
1254 ptlrpc_at_set_timer(svcpt);
1256 spin_unlock(&svcpt->scp_at_lock);
1262 ptlrpc_at_remove_timed(struct ptlrpc_request *req)
1264 struct ptlrpc_at_array *array;
1266 array = &req->rq_rqbd->rqbd_svcpt->scp_at_array;
1268 /* NB: must call with hold svcpt::scp_at_lock */
1269 LASSERT(!list_empty(&req->rq_timed_list));
1270 list_del_init(&req->rq_timed_list);
1272 spin_lock(&req->rq_lock);
1273 req->rq_at_linked = 0;
1274 spin_unlock(&req->rq_lock);
1276 array->paa_reqs_count[req->rq_at_index]--;
1281 * Attempt to extend the request deadline by sending an early reply to the
1284 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
1286 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1287 struct ptlrpc_request *reqcopy;
1288 struct lustre_msg *reqmsg;
1289 time64_t olddl = req->rq_deadline - ktime_get_real_seconds();
1295 if (CFS_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_RECONNECT)) {
1296 /* don't send early reply */
1300 /* deadline is when the client expects us to reply, margin is the
1301 difference between clients' and servers' expectations */
1302 DEBUG_REQ(D_ADAPTTO, req,
1303 "%ssending early reply (deadline %+llds, margin %+llds) for "
1304 "%d+%d", AT_OFF ? "AT off - not " : "",
1305 (s64)olddl, (s64)(olddl - at_get(&svcpt->scp_at_estimate)),
1306 at_get(&svcpt->scp_at_estimate), at_extra);
1312 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+llds), "
1313 "not sending early reply. Consider increasing "
1314 "at_early_margin (%d)?", (s64)olddl, at_early_margin);
1316 /* Return an error so we're not re-added to the timed list. */
1320 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
1321 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
1322 "but no AT support");
1326 if (req->rq_export &&
1327 lustre_msg_get_flags(req->rq_reqmsg) &
1328 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1329 struct obd_device *obd_exp = req->rq_export->exp_obd;
1331 /* During recovery, we don't want to send too many early
1332 * replies, but on the other hand we want to make sure the
1333 * client has enough time to resend if the rpc is lost. So
1334 * during the recovery period send at least 4 early replies,
1335 * spacing them every at_extra if we can. at_estimate should
1336 * always equal this fixed value during recovery.
1338 /* Don't account request processing time into AT history
1339 * during recovery, it is not service time we need but
1340 * includes also waiting time for recovering clients
1342 newdl = min_t(time64_t, at_extra,
1343 obd_exp->obd_recovery_timeout / 4) +
1344 ktime_get_real_seconds();
1346 /* We want to extend the request deadline by at_extra seconds,
1347 * so we set our service estimate to reflect how much time has
1348 * passed since this request arrived plus an additional
1349 * at_extra seconds. The client will calculate the new deadline
1350 * based on this service estimate (plus some additional time to
1351 * account for network latency). See ptlrpc_at_recv_early_reply
1353 at_measured(&svcpt->scp_at_estimate, at_extra +
1354 ktime_get_real_seconds() -
1355 req->rq_arrival_time.tv_sec);
1356 newdl = req->rq_arrival_time.tv_sec +
1357 at_get(&svcpt->scp_at_estimate);
1360 /* Check to see if we've actually increased the deadline -
1361 * we may be past adaptive_max */
1362 if (req->rq_deadline >= newdl) {
1363 DEBUG_REQ(D_WARNING, req, "Couldn't add any time (%lld/%lld), not sending early reply\n",
1364 (s64)olddl, (s64)(newdl - ktime_get_real_seconds()));
1368 reqcopy = ptlrpc_request_cache_alloc(GFP_NOFS);
1369 if (reqcopy == NULL)
1371 OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1373 GOTO(out_free, rc = -ENOMEM);
1376 reqcopy->rq_reply_state = NULL;
1377 reqcopy->rq_rep_swab_mask = 0;
1378 reqcopy->rq_pack_bulk = 0;
1379 reqcopy->rq_pack_udesc = 0;
1380 reqcopy->rq_packed_final = 0;
1381 sptlrpc_svc_ctx_addref(reqcopy);
1382 /* We only need the reqmsg for the magic */
1383 reqcopy->rq_reqmsg = reqmsg;
1384 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1387 * tgt_brw_read() and tgt_brw_write() may have decided not to reply.
1388 * Without this check, we would fail the rq_no_reply assertion in
1389 * ptlrpc_send_reply().
1391 if (reqcopy->rq_no_reply)
1392 GOTO(out, rc = -ETIMEDOUT);
1394 LASSERT(atomic_read(&req->rq_refcount));
1395 /** if it is last refcount then early reply isn't needed */
1396 if (atomic_read(&req->rq_refcount) == 1) {
1397 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
1398 "abort sending early reply\n");
1399 GOTO(out, rc = -EINVAL);
1402 /* Connection ref */
1403 reqcopy->rq_export = class_conn2export(
1404 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1405 if (reqcopy->rq_export == NULL)
1406 GOTO(out, rc = -ENODEV);
1409 class_export_rpc_inc(reqcopy->rq_export);
1410 if (reqcopy->rq_export->exp_obd &&
1411 reqcopy->rq_export->exp_obd->obd_fail)
1412 GOTO(out_put, rc = -ENODEV);
1414 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1418 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1421 /* Adjust our own deadline to what we told the client */
1422 req->rq_deadline = newdl;
1423 req->rq_early_count++; /* number sent, server side */
1425 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1428 /* Free the (early) reply state from lustre_pack_reply.
1429 (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
1430 ptlrpc_req_drop_rs(reqcopy);
1433 class_export_rpc_dec(reqcopy->rq_export);
1434 class_export_put(reqcopy->rq_export);
1436 sptlrpc_svc_ctx_decref(reqcopy);
1437 OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1439 ptlrpc_request_cache_free(reqcopy);
1443 /* Send early replies to everybody expiring within at_early_margin
1444 asking for at_extra time */
1445 static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1447 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1448 struct ptlrpc_request *rq, *n;
1449 struct list_head work_list;
1452 time64_t now = ktime_get_real_seconds();
1453 cfs_duration_t delay;
1454 int first, counter = 0;
1457 spin_lock(&svcpt->scp_at_lock);
1458 if (svcpt->scp_at_check == 0) {
1459 spin_unlock(&svcpt->scp_at_lock);
1462 delay = cfs_time_sub(cfs_time_current(), svcpt->scp_at_checktime);
1463 svcpt->scp_at_check = 0;
1465 if (array->paa_count == 0) {
1466 spin_unlock(&svcpt->scp_at_lock);
1470 /* The timer went off, but maybe the nearest rpc already completed. */
1471 first = array->paa_deadline - now;
1472 if (first > at_early_margin) {
1473 /* We've still got plenty of time. Reset the timer. */
1474 ptlrpc_at_set_timer(svcpt);
1475 spin_unlock(&svcpt->scp_at_lock);
1479 /* We're close to a timeout, and we don't know how much longer the
1480 server will take. Send early replies to everyone expiring soon. */
1481 INIT_LIST_HEAD(&work_list);
1483 div_u64_rem(array->paa_deadline, array->paa_size, &index);
1484 count = array->paa_count;
1486 count -= array->paa_reqs_count[index];
1487 list_for_each_entry_safe(rq, n,
1488 &array->paa_reqs_array[index],
1490 if (rq->rq_deadline > now + at_early_margin) {
1491 /* update the earliest deadline */
1492 if (deadline == -1 ||
1493 rq->rq_deadline < deadline)
1494 deadline = rq->rq_deadline;
1498 ptlrpc_at_remove_timed(rq);
1500 * ptlrpc_server_drop_request() may drop
1501 * refcount to 0 already. Let's check this and
1502 * don't add entry to work_list
1504 if (likely(atomic_inc_not_zero(&rq->rq_refcount)))
1505 list_add(&rq->rq_timed_list, &work_list);
1509 if (++index >= array->paa_size)
1512 array->paa_deadline = deadline;
1513 /* we have a new earliest deadline, restart the timer */
1514 ptlrpc_at_set_timer(svcpt);
1516 spin_unlock(&svcpt->scp_at_lock);
1518 CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
1519 "replies\n", first, at_extra, counter);
1521 /* We're already past request deadlines before we even get a
1522 chance to send early replies */
1523 LCONSOLE_WARN("%s: This server is not able to keep up with "
1524 "request traffic (cpu-bound).\n",
1525 svcpt->scp_service->srv_name);
1526 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, delay=%ld(jiff)\n",
1527 counter, svcpt->scp_nreqs_incoming,
1528 svcpt->scp_nreqs_active,
1529 at_get(&svcpt->scp_at_estimate), delay);
1532 /* we took additional refcount so entries can't be deleted from list, no
1533 * locking is needed */
1534 while (!list_empty(&work_list)) {
1535 rq = list_entry(work_list.next, struct ptlrpc_request,
1537 list_del_init(&rq->rq_timed_list);
1539 if (ptlrpc_at_send_early_reply(rq) == 0)
1540 ptlrpc_at_add_timed(rq);
1542 ptlrpc_server_drop_request(rq);
1545 RETURN(1); /* return "did_something" for liblustre */
1548 /* Check if we are already handling earlier incarnation of this request.
1549 * Called under &req->rq_export->exp_rpc_lock locked */
1550 static int ptlrpc_server_check_resend_in_progress(struct ptlrpc_request *req)
1552 struct ptlrpc_request *tmp = NULL;
1554 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ||
1555 (atomic_read(&req->rq_export->exp_rpc_count) == 0))
1558 /* bulk request are aborted upon reconnect, don't try to
1560 if (req->rq_bulk_write || req->rq_bulk_read)
1563 /* This list should not be longer than max_requests in
1564 * flights on the client, so it is not all that long.
1565 * Also we only hit this codepath in case of a resent
1566 * request which makes it even more rarely hit */
1567 list_for_each_entry(tmp, &req->rq_export->exp_reg_rpcs,
1569 /* Found duplicate one */
1570 if (tmp->rq_xid == req->rq_xid)
1573 list_for_each_entry(tmp, &req->rq_export->exp_hp_rpcs,
1575 /* Found duplicate one */
1576 if (tmp->rq_xid == req->rq_xid)
1582 DEBUG_REQ(D_HA, req, "Found duplicate req in processing");
1583 DEBUG_REQ(D_HA, tmp, "Request being processed");
1588 * Check if a request should be assigned with a high priority.
1590 * \retval < 0: error occurred
1591 * 0: normal RPC request
1592 * +1: high priority request
1594 static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt,
1595 struct ptlrpc_request *req)
1600 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL) {
1601 rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
1608 if (req->rq_export != NULL && req->rq_ops != NULL) {
1609 /* Perform request specific check. We should do this
1610 * check before the request is added into exp_hp_rpcs
1611 * list otherwise it may hit swab race at LU-1044. */
1612 if (req->rq_ops->hpreq_check != NULL) {
1613 rc = req->rq_ops->hpreq_check(req);
1614 if (rc == -ESTALE) {
1615 req->rq_status = rc;
1618 /** can only return error,
1619 * 0 for normal request,
1620 * or 1 for high priority request */
1628 /** Remove the request from the export list. */
1629 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req)
1632 if (req->rq_export) {
1633 /* refresh lock timeout again so that client has more
1634 * room to send lock cancel RPC. */
1635 if (req->rq_ops && req->rq_ops->hpreq_fini)
1636 req->rq_ops->hpreq_fini(req);
1638 spin_lock_bh(&req->rq_export->exp_rpc_lock);
1639 list_del_init(&req->rq_exp_list);
1640 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1645 static int ptlrpc_hpreq_check(struct ptlrpc_request *req)
1650 static struct ptlrpc_hpreq_ops ptlrpc_hpreq_common = {
1651 .hpreq_check = ptlrpc_hpreq_check,
1654 /* Hi-Priority RPC check by RPC operation code. */
1655 int ptlrpc_hpreq_handler(struct ptlrpc_request *req)
1657 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1659 /* Check for export to let only reconnects for not yet evicted
1660 * export to become a HP rpc. */
1661 if ((req->rq_export != NULL) &&
1662 (opc == OBD_PING || opc == MDS_CONNECT || opc == OST_CONNECT))
1663 req->rq_ops = &ptlrpc_hpreq_common;
1667 EXPORT_SYMBOL(ptlrpc_hpreq_handler);
1669 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1670 struct ptlrpc_request *req)
1676 rc = ptlrpc_server_hpreq_init(svcpt, req);
1681 ptlrpc_nrs_req_initialize(svcpt, req, hp);
1683 if (req->rq_export != NULL) {
1684 struct obd_export *exp = req->rq_export;
1686 /* do search for duplicated xid and the adding to the list
1688 spin_lock_bh(&exp->exp_rpc_lock);
1689 rc = ptlrpc_server_check_resend_in_progress(req);
1691 spin_unlock_bh(&exp->exp_rpc_lock);
1693 ptlrpc_nrs_req_finalize(req);
1697 if (hp || req->rq_ops != NULL)
1698 list_add(&req->rq_exp_list, &exp->exp_hp_rpcs);
1700 list_add(&req->rq_exp_list, &exp->exp_reg_rpcs);
1701 spin_unlock_bh(&exp->exp_rpc_lock);
1704 /* the current thread is not the processing thread for this request
1705 * since that, but request is in exp_hp_list and can be find there.
1706 * Remove all relations between request and old thread. */
1707 req->rq_svc_thread->t_env->le_ses = NULL;
1708 req->rq_svc_thread = NULL;
1709 req->rq_session.lc_thread = NULL;
1711 ptlrpc_nrs_req_add(svcpt, req, hp);
1717 * Allow to handle high priority request
1718 * User can call it w/o any lock but need to hold
1719 * ptlrpc_service_part::scp_req_lock to get reliable result
1721 static bool ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1724 int running = svcpt->scp_nthrs_running;
1726 if (!nrs_svcpt_has_hp(svcpt))
1732 if (ptlrpc_nrs_req_throttling_nolock(svcpt, true))
1735 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1736 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1737 /* leave just 1 thread for normal RPCs */
1738 running = PTLRPC_NTHRS_INIT;
1739 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1743 if (svcpt->scp_nreqs_active >= running - 1)
1746 if (svcpt->scp_nhreqs_active == 0)
1749 return !ptlrpc_nrs_req_pending_nolock(svcpt, false) ||
1750 svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1753 static bool ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1756 return ptlrpc_server_allow_high(svcpt, force) &&
1757 ptlrpc_nrs_req_pending_nolock(svcpt, true);
1761 * Only allow normal priority requests on a service that has a high-priority
1762 * queue if forced (i.e. cleanup), if there are other high priority requests
1763 * already being processed (i.e. those threads can service more high-priority
1764 * requests), or if there are enough idle threads that a later thread can do
1765 * a high priority request.
1766 * User can call it w/o any lock but need to hold
1767 * ptlrpc_service_part::scp_req_lock to get reliable result
1769 static bool ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1772 int running = svcpt->scp_nthrs_running;
1773 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1774 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1775 /* leave just 1 thread for normal RPCs */
1776 running = PTLRPC_NTHRS_INIT;
1777 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1784 if (ptlrpc_nrs_req_throttling_nolock(svcpt, false))
1787 if (svcpt->scp_nreqs_active < running - 2)
1790 if (svcpt->scp_nreqs_active >= running - 1)
1793 return svcpt->scp_nhreqs_active > 0 || !nrs_svcpt_has_hp(svcpt);
1796 static bool ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1799 return ptlrpc_server_allow_normal(svcpt, force) &&
1800 ptlrpc_nrs_req_pending_nolock(svcpt, false);
1804 * Returns true if there are requests available in incoming
1805 * request queue for processing and it is allowed to fetch them.
1806 * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1807 * to get reliable result
1808 * \see ptlrpc_server_allow_normal
1809 * \see ptlrpc_server_allow high
1812 ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt, bool force)
1814 return ptlrpc_server_high_pending(svcpt, force) ||
1815 ptlrpc_server_normal_pending(svcpt, force);
1819 * Fetch a request for processing from queue of unprocessed requests.
1820 * Favors high-priority requests.
1821 * Returns a pointer to fetched request.
1823 static struct ptlrpc_request *
1824 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force)
1826 struct ptlrpc_request *req = NULL;
1829 spin_lock(&svcpt->scp_req_lock);
1831 if (ptlrpc_server_high_pending(svcpt, force)) {
1832 req = ptlrpc_nrs_req_get_nolock(svcpt, true, force);
1834 svcpt->scp_hreq_count++;
1839 if (ptlrpc_server_normal_pending(svcpt, force)) {
1840 req = ptlrpc_nrs_req_get_nolock(svcpt, false, force);
1842 svcpt->scp_hreq_count = 0;
1847 spin_unlock(&svcpt->scp_req_lock);
1851 svcpt->scp_nreqs_active++;
1853 svcpt->scp_nhreqs_active++;
1855 spin_unlock(&svcpt->scp_req_lock);
1857 if (likely(req->rq_export))
1858 class_export_rpc_inc(req->rq_export);
1864 * Handle freshly incoming reqs, add to timed early reply list,
1865 * pass on to regular request queue.
1866 * All incoming requests pass through here before getting into
1867 * ptlrpc_server_handle_req later on.
1870 ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt,
1871 struct ptlrpc_thread *thread)
1873 struct ptlrpc_service *svc = svcpt->scp_service;
1874 struct ptlrpc_request *req;
1879 spin_lock(&svcpt->scp_lock);
1880 if (list_empty(&svcpt->scp_req_incoming)) {
1881 spin_unlock(&svcpt->scp_lock);
1885 req = list_entry(svcpt->scp_req_incoming.next,
1886 struct ptlrpc_request, rq_list);
1887 list_del_init(&req->rq_list);
1888 svcpt->scp_nreqs_incoming--;
1889 /* Consider this still a "queued" request as far as stats are
1891 spin_unlock(&svcpt->scp_lock);
1893 /* go through security check/transform */
1894 rc = sptlrpc_svc_unwrap_request(req);
1898 case SECSVC_COMPLETE:
1899 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1908 * for null-flavored rpc, msg has been unpacked by sptlrpc, although
1909 * redo it wouldn't be harmful.
1911 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
1912 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
1914 CERROR("error unpacking request: ptl %d from %s "
1915 "x%llu\n", svc->srv_req_portal,
1916 libcfs_id2str(req->rq_peer), req->rq_xid);
1921 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1923 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
1924 "%llu\n", svc->srv_req_portal,
1925 libcfs_id2str(req->rq_peer), req->rq_xid);
1929 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
1930 lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
1931 CERROR("drop incoming rpc opc %u, x%llu\n",
1932 cfs_fail_val, req->rq_xid);
1937 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1938 CERROR("wrong packet type received (type=%u) from %s\n",
1939 lustre_msg_get_type(req->rq_reqmsg),
1940 libcfs_id2str(req->rq_peer));
1944 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1948 req->rq_bulk_write = 1;
1952 case MGS_CONFIG_READ:
1953 req->rq_bulk_read = 1;
1957 CDEBUG(D_RPCTRACE, "got req x%llu\n", req->rq_xid);
1959 req->rq_export = class_conn2export(
1960 lustre_msg_get_handle(req->rq_reqmsg));
1961 if (req->rq_export) {
1962 rc = ptlrpc_check_req(req);
1964 rc = sptlrpc_target_export_check(req->rq_export, req);
1966 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
1967 "illegal security flavor,");
1972 ptlrpc_update_export_timer(req->rq_export, 0);
1975 /* req_in handling should/must be fast */
1976 if (ktime_get_real_seconds() - req->rq_arrival_time.tv_sec > 5)
1977 DEBUG_REQ(D_WARNING, req, "Slow req_in handling %llds",
1978 (s64)(ktime_get_real_seconds() -
1979 req->rq_arrival_time.tv_sec));
1981 /* Set rpc server deadline and add it to the timed list */
1982 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1983 MSGHDR_AT_SUPPORT) ?
1984 /* The max time the client expects us to take */
1985 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1987 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1988 if (unlikely(deadline == 0)) {
1989 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1993 /* Skip early reply */
1994 if (OBD_FAIL_PRECHECK(OBD_FAIL_MDS_RESEND))
1995 req->rq_deadline += obd_timeout;
1997 req->rq_svc_thread = thread;
1998 if (thread != NULL) {
1999 /* initialize request session, it is needed for request
2000 * processing by target */
2001 rc = lu_context_init(&req->rq_session, LCT_SERVER_SESSION |
2004 CERROR("%s: failure to initialize session: rc = %d\n",
2005 thread->t_name, rc);
2008 req->rq_session.lc_thread = thread;
2009 lu_context_enter(&req->rq_session);
2010 thread->t_env->le_ses = &req->rq_session;
2013 ptlrpc_at_add_timed(req);
2015 /* Move it over to the request processing queue */
2016 rc = ptlrpc_server_request_add(svcpt, req);
2020 wake_up(&svcpt->scp_waitq);
2024 ptlrpc_server_finish_request(svcpt, req);
2030 * Main incoming request handling logic.
2031 * Calls handler function from service to do actual processing.
2034 ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
2035 struct ptlrpc_thread *thread)
2037 struct ptlrpc_service *svc = svcpt->scp_service;
2038 struct ptlrpc_request *request;
2048 request = ptlrpc_server_request_get(svcpt, false);
2049 if (request == NULL)
2052 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
2053 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
2054 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
2055 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
2057 if (unlikely(fail_opc)) {
2058 if (request->rq_export && request->rq_ops)
2059 OBD_FAIL_TIMEOUT(fail_opc, 4);
2062 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
2064 if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
2065 libcfs_debug_dumplog();
2067 work_start = ktime_get_real();
2068 arrived = timespec64_to_ktime(request->rq_arrival_time);
2069 timediff_usecs = ktime_us_delta(work_start, arrived);
2070 if (likely(svc->srv_stats != NULL)) {
2071 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
2073 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
2074 svcpt->scp_nreqs_incoming);
2075 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
2076 svcpt->scp_nreqs_active);
2077 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
2078 at_get(&svcpt->scp_at_estimate));
2081 if (likely(request->rq_export)) {
2082 if (unlikely(ptlrpc_check_req(request)))
2084 ptlrpc_update_export_timer(request->rq_export,
2085 timediff_usecs / (USEC_PER_SEC / 2));
2088 /* Discard requests queued for longer than the deadline.
2089 The deadline is increased if we send an early reply. */
2090 if (ktime_get_real_seconds() > request->rq_deadline) {
2091 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s: deadline %lld:%llds ago\n",
2092 libcfs_id2str(request->rq_peer),
2093 request->rq_deadline -
2094 request->rq_arrival_time.tv_sec,
2095 ktime_get_real_seconds() - request->rq_deadline);
2099 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
2100 "%s:%s+%d:%d:x%llu:%s:%d\n", current_comm(),
2101 (request->rq_export ?
2102 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2103 (request->rq_export ?
2104 atomic_read(&request->rq_export->exp_refcount) : -99),
2105 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
2106 libcfs_id2str(request->rq_peer),
2107 lustre_msg_get_opc(request->rq_reqmsg));
2109 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
2110 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
2112 CDEBUG(D_NET, "got req %llu\n", request->rq_xid);
2114 /* re-assign request and sesson thread to the current one */
2115 request->rq_svc_thread = thread;
2116 if (thread != NULL) {
2117 LASSERT(request->rq_session.lc_thread == NULL);
2118 request->rq_session.lc_thread = thread;
2119 thread->t_env->le_ses = &request->rq_session;
2121 svc->srv_ops.so_req_handler(request);
2123 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
2126 if (unlikely(ktime_get_real_seconds() > request->rq_deadline)) {
2127 DEBUG_REQ(D_WARNING, request,
2128 "Request took longer than estimated (%lld:%llds); "
2129 "client may timeout.",
2130 request->rq_deadline -
2131 request->rq_arrival_time.tv_sec,
2132 ktime_get_real_seconds() - request->rq_deadline);
2135 work_end = ktime_get_real();
2136 timediff_usecs = ktime_us_delta(work_end, work_start);
2137 arrived_usecs = ktime_us_delta(work_end, arrived);
2138 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
2139 "%s:%s+%d:%d:x%llu:%s:%d Request processed in %lldus "
2140 "(%lldus total) trans %llu rc %d/%d\n",
2142 (request->rq_export ?
2143 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2144 (request->rq_export ?
2145 atomic_read(&request->rq_export->exp_refcount) : -99),
2146 lustre_msg_get_status(request->rq_reqmsg),
2148 libcfs_id2str(request->rq_peer),
2149 lustre_msg_get_opc(request->rq_reqmsg),
2152 (request->rq_repmsg ?
2153 lustre_msg_get_transno(request->rq_repmsg) :
2154 request->rq_transno),
2156 (request->rq_repmsg ?
2157 lustre_msg_get_status(request->rq_repmsg) : -999));
2158 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
2159 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
2160 int opc = opcode_offset(op);
2161 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
2162 LASSERT(opc < LUSTRE_MAX_OPCODES);
2163 lprocfs_counter_add(svc->srv_stats,
2164 opc + EXTRA_MAX_OPCODES,
2168 if (unlikely(request->rq_early_count)) {
2169 DEBUG_REQ(D_ADAPTTO, request,
2170 "sent %d early replies before finishing in %llds",
2171 request->rq_early_count,
2172 arrived_usecs / USEC_PER_SEC);
2175 ptlrpc_server_finish_active_request(svcpt, request);
2181 * An internal function to process a single reply state object.
2184 ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
2186 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
2187 struct ptlrpc_service *svc = svcpt->scp_service;
2188 struct obd_export *exp;
2193 exp = rs->rs_export;
2195 LASSERT(rs->rs_difficult);
2196 LASSERT(rs->rs_scheduled);
2197 LASSERT(list_empty(&rs->rs_list));
2199 /* The disk commit callback holds exp_uncommitted_replies_lock while it
2200 * iterates over newly committed replies, removing them from
2201 * exp_uncommitted_replies. It then drops this lock and schedules the
2202 * replies it found for handling here.
2204 * We can avoid contention for exp_uncommitted_replies_lock between the
2205 * HRT threads and further commit callbacks by checking rs_committed
2206 * which is set in the commit callback while it holds both
2207 * rs_lock and exp_uncommitted_reples.
2209 * If we see rs_committed clear, the commit callback _may_ not have
2210 * handled this reply yet and we race with it to grab
2211 * exp_uncommitted_replies_lock before removing the reply from
2212 * exp_uncommitted_replies. Note that if we lose the race and the
2213 * reply has already been removed, list_del_init() is a noop.
2215 * If we see rs_committed set, we know the commit callback is handling,
2216 * or has handled this reply since store reordering might allow us to
2217 * see rs_committed set out of sequence. But since this is done
2218 * holding rs_lock, we can be sure it has all completed once we hold
2219 * rs_lock, which we do right next.
2221 if (!rs->rs_committed) {
2222 /* if rs was commited, no need to convert locks, don't check
2223 * rs_committed here because rs may never be added into
2224 * exp_uncommitted_replies and this flag never be set, see
2225 * target_send_reply() */
2226 if (rs->rs_convert_lock &&
2227 rs->rs_transno > exp->exp_last_committed) {
2228 struct ldlm_lock *lock;
2229 struct ldlm_lock *ack_locks[RS_MAX_LOCKS] = { NULL };
2231 spin_lock(&rs->rs_lock);
2232 if (rs->rs_convert_lock &&
2233 rs->rs_transno > exp->exp_last_committed) {
2234 nlocks = rs->rs_nlocks;
2235 while (nlocks-- > 0) {
2237 * NB don't assume rs is always handled
2238 * by the same service thread (see
2239 * ptlrpc_hr_select, so REP-ACK hr may
2240 * race with trans commit, while the
2241 * latter will release locks, get locks
2242 * here early to convert to COS mode
2245 lock = ldlm_handle2lock(
2246 &rs->rs_locks[nlocks]);
2248 ack_locks[nlocks] = lock;
2249 rs->rs_modes[nlocks] = LCK_COS;
2251 nlocks = rs->rs_nlocks;
2252 rs->rs_convert_lock = 0;
2253 /* clear rs_scheduled so that commit callback
2254 * can schedule again */
2255 rs->rs_scheduled = 0;
2256 spin_unlock(&rs->rs_lock);
2258 while (nlocks-- > 0) {
2259 lock = ack_locks[nlocks];
2260 ldlm_lock_mode_downgrade(lock, LCK_COS);
2261 LDLM_LOCK_PUT(lock);
2265 spin_unlock(&rs->rs_lock);
2268 spin_lock(&exp->exp_uncommitted_replies_lock);
2269 list_del_init(&rs->rs_obd_list);
2270 spin_unlock(&exp->exp_uncommitted_replies_lock);
2273 spin_lock(&exp->exp_lock);
2274 /* Noop if removed already */
2275 list_del_init(&rs->rs_exp_list);
2276 spin_unlock(&exp->exp_lock);
2278 spin_lock(&rs->rs_lock);
2280 been_handled = rs->rs_handled;
2283 nlocks = rs->rs_nlocks; /* atomic "steal", but */
2284 rs->rs_nlocks = 0; /* locks still on rs_locks! */
2286 if (nlocks == 0 && !been_handled) {
2287 /* If we see this, we should already have seen the warning
2288 * in mds_steal_ack_locks() */
2289 CDEBUG(D_HA, "All locks stolen from rs %p x%lld.t%lld"
2292 rs->rs_xid, rs->rs_transno, rs->rs_opc,
2293 libcfs_nid2str(exp->exp_connection->c_peer.nid));
2296 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
2297 spin_unlock(&rs->rs_lock);
2299 if (!been_handled && rs->rs_on_net) {
2300 LNetMDUnlink(rs->rs_md_h);
2301 /* Ignore return code; we're racing with completion */
2304 while (nlocks-- > 0)
2305 ldlm_lock_decref(&rs->rs_locks[nlocks],
2306 rs->rs_modes[nlocks]);
2308 spin_lock(&rs->rs_lock);
2311 rs->rs_scheduled = 0;
2312 rs->rs_convert_lock = 0;
2314 if (!rs->rs_on_net) {
2316 spin_unlock(&rs->rs_lock);
2318 class_export_put (exp);
2319 rs->rs_export = NULL;
2320 ptlrpc_rs_decref(rs);
2321 if (atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
2322 svc->srv_is_stopping)
2323 wake_up_all(&svcpt->scp_waitq);
2327 /* still on the net; callback will schedule */
2328 spin_unlock(&rs->rs_lock);
2334 ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
2336 int avail = svcpt->scp_nrqbds_posted;
2337 int low_water = test_req_buffer_pressure ? 0 :
2338 svcpt->scp_service->srv_nbuf_per_group / 2;
2340 /* NB I'm not locking; just looking. */
2342 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
2343 * allowed the request history to grow out of control. We could put a
2344 * sanity check on that here and cull some history if we need the
2347 if (avail <= low_water)
2348 ptlrpc_grow_req_bufs(svcpt, 1);
2350 if (svcpt->scp_service->srv_stats) {
2351 lprocfs_counter_add(svcpt->scp_service->srv_stats,
2352 PTLRPC_REQBUF_AVAIL_CNTR, avail);
2357 ptlrpc_retry_rqbds(void *arg)
2359 struct ptlrpc_service_part *svcpt = (struct ptlrpc_service_part *)arg;
2361 svcpt->scp_rqbd_timeout = 0;
2366 ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
2368 return svcpt->scp_nreqs_active <
2369 svcpt->scp_nthrs_running - 1 -
2370 (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL);
2374 * allowed to create more threads
2375 * user can call it w/o any lock but need to hold
2376 * ptlrpc_service_part::scp_lock to get reliable result
2379 ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
2381 return svcpt->scp_nthrs_running +
2382 svcpt->scp_nthrs_starting <
2383 svcpt->scp_service->srv_nthrs_cpt_limit;
2387 * too many requests and allowed to create more threads
2390 ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
2392 return !ptlrpc_threads_enough(svcpt) &&
2393 ptlrpc_threads_increasable(svcpt);
2397 ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2399 return thread_is_stopping(thread) ||
2400 thread->t_svcpt->scp_service->srv_is_stopping;
2404 ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
2406 return !list_empty(&svcpt->scp_rqbd_idle) &&
2407 svcpt->scp_rqbd_timeout == 0;
2411 ptlrpc_at_check(struct ptlrpc_service_part *svcpt)
2413 return svcpt->scp_at_check;
2417 * requests wait on preprocessing
2418 * user can call it w/o any lock but need to hold
2419 * ptlrpc_service_part::scp_lock to get reliable result
2422 ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
2424 return !list_empty(&svcpt->scp_req_incoming);
2427 static __attribute__((__noinline__)) int
2428 ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
2429 struct ptlrpc_thread *thread)
2431 /* Don't exit while there are replies to be handled */
2432 struct l_wait_info lwi = LWI_TIMEOUT(svcpt->scp_rqbd_timeout,
2433 ptlrpc_retry_rqbds, svcpt);
2435 lc_watchdog_disable(thread->t_watchdog);
2439 l_wait_event_exclusive_head(svcpt->scp_waitq,
2440 ptlrpc_thread_stopping(thread) ||
2441 ptlrpc_server_request_incoming(svcpt) ||
2442 ptlrpc_server_request_pending(svcpt, false) ||
2443 ptlrpc_rqbd_pending(svcpt) ||
2444 ptlrpc_at_check(svcpt), &lwi);
2446 if (ptlrpc_thread_stopping(thread))
2449 lc_watchdog_touch(thread->t_watchdog,
2450 ptlrpc_server_get_timeout(svcpt));
2455 * Main thread body for service threads.
2456 * Waits in a loop waiting for new requests to process to appear.
2457 * Every time an incoming requests is added to its queue, a waitq
2458 * is woken up and one of the threads will handle it.
2460 static int ptlrpc_main(void *arg)
2462 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
2463 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2464 struct ptlrpc_service *svc = svcpt->scp_service;
2465 struct ptlrpc_reply_state *rs;
2466 struct group_info *ginfo = NULL;
2468 int counter = 0, rc = 0;
2471 thread->t_pid = current_pid();
2472 unshare_fs_struct();
2474 /* NB: we will call cfs_cpt_bind() for all threads, because we
2475 * might want to run lustre server only on a subset of system CPUs,
2476 * in that case ->scp_cpt is CFS_CPT_ANY */
2477 rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
2479 CWARN("%s: failed to bind %s on CPT %d\n",
2480 svc->srv_name, thread->t_name, svcpt->scp_cpt);
2483 ginfo = groups_alloc(0);
2489 set_current_groups(ginfo);
2490 put_group_info(ginfo);
2492 if (svc->srv_ops.so_thr_init != NULL) {
2493 rc = svc->srv_ops.so_thr_init(thread);
2504 rc = lu_context_init(&env->le_ctx,
2505 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2509 thread->t_env = env;
2510 env->le_ctx.lc_thread = thread;
2511 env->le_ctx.lc_cookie = 0x6;
2513 while (!list_empty(&svcpt->scp_rqbd_idle)) {
2514 rc = ptlrpc_server_post_idle_rqbds(svcpt);
2518 CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
2519 svc->srv_name, svcpt->scp_cpt, rc);
2523 /* Alloc reply state structure for this one */
2524 OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2530 spin_lock(&svcpt->scp_lock);
2532 LASSERT(thread_is_starting(thread));
2533 thread_clear_flags(thread, SVC_STARTING);
2535 LASSERT(svcpt->scp_nthrs_starting == 1);
2536 svcpt->scp_nthrs_starting--;
2538 /* SVC_STOPPING may already be set here if someone else is trying
2539 * to stop the service while this new thread has been dynamically
2540 * forked. We still set SVC_RUNNING to let our creator know that
2541 * we are now running, however we will exit as soon as possible */
2542 thread_add_flags(thread, SVC_RUNNING);
2543 svcpt->scp_nthrs_running++;
2544 spin_unlock(&svcpt->scp_lock);
2546 /* wake up our creator in case he's still waiting. */
2547 wake_up(&thread->t_ctl_waitq);
2549 thread->t_watchdog = lc_watchdog_add(ptlrpc_server_get_timeout(svcpt),
2552 spin_lock(&svcpt->scp_rep_lock);
2553 list_add(&rs->rs_list, &svcpt->scp_rep_idle);
2554 wake_up(&svcpt->scp_rep_waitq);
2555 spin_unlock(&svcpt->scp_rep_lock);
2557 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2558 svcpt->scp_nthrs_running);
2560 /* XXX maintain a list of all managed devices: insert here */
2561 while (!ptlrpc_thread_stopping(thread)) {
2562 if (ptlrpc_wait_event(svcpt, thread))
2565 ptlrpc_check_rqbd_pool(svcpt);
2567 if (ptlrpc_threads_need_create(svcpt)) {
2568 /* Ignore return code - we tried... */
2569 ptlrpc_start_thread(svcpt, 0);
2572 /* reset le_ses to initial state */
2574 /* Process all incoming reqs before handling any */
2575 if (ptlrpc_server_request_incoming(svcpt)) {
2576 lu_context_enter(&env->le_ctx);
2577 ptlrpc_server_handle_req_in(svcpt, thread);
2578 lu_context_exit(&env->le_ctx);
2580 /* but limit ourselves in case of flood */
2581 if (counter++ < 100)
2586 if (ptlrpc_at_check(svcpt))
2587 ptlrpc_at_check_timed(svcpt);
2589 if (ptlrpc_server_request_pending(svcpt, false)) {
2590 lu_context_enter(&env->le_ctx);
2591 ptlrpc_server_handle_request(svcpt, thread);
2592 lu_context_exit(&env->le_ctx);
2595 if (ptlrpc_rqbd_pending(svcpt) &&
2596 ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
2597 /* I just failed to repost request buffers.
2598 * Wait for a timeout (unless something else
2599 * happens) before I try again */
2600 svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
2601 CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
2602 svcpt->scp_nrqbds_posted);
2606 lc_watchdog_delete(thread->t_watchdog);
2607 thread->t_watchdog = NULL;
2611 * deconstruct service specific state created by ptlrpc_start_thread()
2613 if (svc->srv_ops.so_thr_done != NULL)
2614 svc->srv_ops.so_thr_done(thread);
2617 lu_context_fini(&env->le_ctx);
2621 CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n",
2622 thread, thread->t_pid, thread->t_id, rc);
2624 spin_lock(&svcpt->scp_lock);
2625 if (thread_test_and_clear_flags(thread, SVC_STARTING))
2626 svcpt->scp_nthrs_starting--;
2628 if (thread_test_and_clear_flags(thread, SVC_RUNNING)) {
2629 /* must know immediately */
2630 svcpt->scp_nthrs_running--;
2634 thread_add_flags(thread, SVC_STOPPED);
2636 wake_up(&thread->t_ctl_waitq);
2637 spin_unlock(&svcpt->scp_lock);
2642 static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
2643 struct list_head *replies)
2647 spin_lock(&hrt->hrt_lock);
2649 list_splice_init(&hrt->hrt_queue, replies);
2650 result = ptlrpc_hr.hr_stopping || !list_empty(replies);
2652 spin_unlock(&hrt->hrt_lock);
2657 * Main body of "handle reply" function.
2658 * It processes acked reply states
2660 static int ptlrpc_hr_main(void *arg)
2662 struct ptlrpc_hr_thread *hrt = (struct ptlrpc_hr_thread *)arg;
2663 struct ptlrpc_hr_partition *hrp = hrt->hrt_partition;
2664 struct list_head replies;
2667 INIT_LIST_HEAD(&replies);
2668 unshare_fs_struct();
2670 rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
2672 char threadname[20];
2674 snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
2675 hrp->hrp_cpt, hrt->hrt_id);
2676 CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
2677 threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
2680 atomic_inc(&hrp->hrp_nstarted);
2681 wake_up(&ptlrpc_hr.hr_waitq);
2683 while (!ptlrpc_hr.hr_stopping) {
2684 l_wait_condition(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
2686 while (!list_empty(&replies)) {
2687 struct ptlrpc_reply_state *rs;
2689 rs = list_entry(replies.prev,
2690 struct ptlrpc_reply_state,
2692 list_del_init(&rs->rs_list);
2693 ptlrpc_handle_rs(rs);
2697 atomic_inc(&hrp->hrp_nstopped);
2698 wake_up(&ptlrpc_hr.hr_waitq);
2703 static void ptlrpc_stop_hr_threads(void)
2705 struct ptlrpc_hr_partition *hrp;
2709 ptlrpc_hr.hr_stopping = 1;
2711 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2712 if (hrp->hrp_thrs == NULL)
2713 continue; /* uninitialized */
2714 for (j = 0; j < hrp->hrp_nthrs; j++)
2715 wake_up_all(&hrp->hrp_thrs[j].hrt_waitq);
2718 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2719 if (hrp->hrp_thrs == NULL)
2720 continue; /* uninitialized */
2721 wait_event(ptlrpc_hr.hr_waitq,
2722 atomic_read(&hrp->hrp_nstopped) ==
2723 atomic_read(&hrp->hrp_nstarted));
2727 static int ptlrpc_start_hr_threads(void)
2729 struct ptlrpc_hr_partition *hrp;
2734 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2737 for (j = 0; j < hrp->hrp_nthrs; j++) {
2738 struct ptlrpc_hr_thread *hrt = &hrp->hrp_thrs[j];
2739 struct task_struct *task;
2741 task = kthread_run(ptlrpc_hr_main,
2743 "ptlrpc_hr%02d_%03d",
2752 wait_event(ptlrpc_hr.hr_waitq,
2753 atomic_read(&hrp->hrp_nstarted) == j);
2756 CERROR("cannot start reply handler thread %d:%d: "
2757 "rc = %d\n", i, j, rc);
2758 ptlrpc_stop_hr_threads();
2766 static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
2768 struct l_wait_info lwi = { 0 };
2769 struct ptlrpc_thread *thread;
2770 struct list_head zombie;
2774 CDEBUG(D_INFO, "Stopping threads for service %s\n",
2775 svcpt->scp_service->srv_name);
2777 INIT_LIST_HEAD(&zombie);
2778 spin_lock(&svcpt->scp_lock);
2779 /* let the thread know that we would like it to stop asap */
2780 list_for_each_entry(thread, &svcpt->scp_threads, t_link) {
2781 CDEBUG(D_INFO, "Stopping thread %s #%u\n",
2782 svcpt->scp_service->srv_thread_name, thread->t_id);
2783 thread_add_flags(thread, SVC_STOPPING);
2786 wake_up_all(&svcpt->scp_waitq);
2788 while (!list_empty(&svcpt->scp_threads)) {
2789 thread = list_entry(svcpt->scp_threads.next,
2790 struct ptlrpc_thread, t_link);
2791 if (thread_is_stopped(thread)) {
2792 list_del(&thread->t_link);
2793 list_add(&thread->t_link, &zombie);
2796 spin_unlock(&svcpt->scp_lock);
2798 CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
2799 svcpt->scp_service->srv_thread_name, thread->t_id);
2800 l_wait_event(thread->t_ctl_waitq,
2801 thread_is_stopped(thread), &lwi);
2803 spin_lock(&svcpt->scp_lock);
2806 spin_unlock(&svcpt->scp_lock);
2808 while (!list_empty(&zombie)) {
2809 thread = list_entry(zombie.next,
2810 struct ptlrpc_thread, t_link);
2811 list_del(&thread->t_link);
2812 OBD_FREE_PTR(thread);
2818 * Stops all threads of a particular service \a svc
2820 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2822 struct ptlrpc_service_part *svcpt;
2826 ptlrpc_service_for_each_part(svcpt, i, svc) {
2827 if (svcpt->scp_service != NULL)
2828 ptlrpc_svcpt_stop_threads(svcpt);
2834 int ptlrpc_start_threads(struct ptlrpc_service *svc)
2841 /* We require 2 threads min, see note in ptlrpc_server_handle_request */
2842 LASSERT(svc->srv_nthrs_cpt_init >= PTLRPC_NTHRS_INIT);
2844 for (i = 0; i < svc->srv_ncpts; i++) {
2845 for (j = 0; j < svc->srv_nthrs_cpt_init; j++) {
2846 rc = ptlrpc_start_thread(svc->srv_parts[i], 1);
2852 /* We have enough threads, don't start more. b=15759 */
2859 CERROR("cannot start %s thread #%d_%d: rc %d\n",
2860 svc->srv_thread_name, i, j, rc);
2861 ptlrpc_stop_all_threads(svc);
2865 int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
2867 struct l_wait_info lwi = { 0 };
2868 struct ptlrpc_thread *thread;
2869 struct ptlrpc_service *svc;
2870 struct task_struct *task;
2874 LASSERT(svcpt != NULL);
2876 svc = svcpt->scp_service;
2878 CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n",
2879 svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running,
2880 svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit);
2883 if (unlikely(svc->srv_is_stopping))
2886 if (!ptlrpc_threads_increasable(svcpt) ||
2887 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
2888 svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1))
2891 OBD_CPT_ALLOC_PTR(thread, svc->srv_cptable, svcpt->scp_cpt);
2894 init_waitqueue_head(&thread->t_ctl_waitq);
2896 spin_lock(&svcpt->scp_lock);
2897 if (!ptlrpc_threads_increasable(svcpt)) {
2898 spin_unlock(&svcpt->scp_lock);
2899 OBD_FREE_PTR(thread);
2903 if (svcpt->scp_nthrs_starting != 0) {
2904 /* serialize starting because some modules (obdfilter)
2905 * might require unique and contiguous t_id */
2906 LASSERT(svcpt->scp_nthrs_starting == 1);
2907 spin_unlock(&svcpt->scp_lock);
2908 OBD_FREE_PTR(thread);
2910 CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n",
2911 svc->srv_thread_name, svcpt->scp_thr_nextid);
2916 CDEBUG(D_INFO, "Creating thread %s #%d race, retry later\n",
2917 svc->srv_thread_name, svcpt->scp_thr_nextid);
2921 svcpt->scp_nthrs_starting++;
2922 thread->t_id = svcpt->scp_thr_nextid++;
2923 thread_add_flags(thread, SVC_STARTING);
2924 thread->t_svcpt = svcpt;
2926 list_add(&thread->t_link, &svcpt->scp_threads);
2927 spin_unlock(&svcpt->scp_lock);
2929 if (svcpt->scp_cpt >= 0) {
2930 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s%02d_%03d",
2931 svc->srv_thread_name, svcpt->scp_cpt, thread->t_id);
2933 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s_%04d",
2934 svc->srv_thread_name, thread->t_id);
2937 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name);
2938 task = kthread_run(ptlrpc_main, thread, "%s", thread->t_name);
2941 CERROR("cannot start thread '%s': rc = %d\n",
2942 thread->t_name, rc);
2943 spin_lock(&svcpt->scp_lock);
2944 --svcpt->scp_nthrs_starting;
2945 if (thread_is_stopping(thread)) {
2946 /* this ptlrpc_thread is being hanled
2947 * by ptlrpc_svcpt_stop_threads now
2949 thread_add_flags(thread, SVC_STOPPED);
2950 wake_up(&thread->t_ctl_waitq);
2951 spin_unlock(&svcpt->scp_lock);
2953 list_del(&thread->t_link);
2954 spin_unlock(&svcpt->scp_lock);
2955 OBD_FREE_PTR(thread);
2963 l_wait_event(thread->t_ctl_waitq,
2964 thread_is_running(thread) || thread_is_stopped(thread),
2967 rc = thread_is_stopped(thread) ? thread->t_id : 0;
2971 int ptlrpc_hr_init(void)
2973 struct ptlrpc_hr_partition *hrp;
2974 struct ptlrpc_hr_thread *hrt;
2981 memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
2982 ptlrpc_hr.hr_cpt_table = cfs_cpt_table;
2984 ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
2986 if (ptlrpc_hr.hr_partitions == NULL)
2989 init_waitqueue_head(&ptlrpc_hr.hr_waitq);
2991 weight = cpumask_weight(topology_sibling_cpumask(smp_processor_id()));
2993 cfs_percpt_for_each(hrp, cpt, ptlrpc_hr.hr_partitions) {
2996 atomic_set(&hrp->hrp_nstarted, 0);
2997 atomic_set(&hrp->hrp_nstopped, 0);
2999 hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, cpt);
3000 hrp->hrp_nthrs /= weight;
3001 if (hrp->hrp_nthrs == 0)
3004 OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, cpt,
3005 hrp->hrp_nthrs * sizeof(*hrt));
3006 if (hrp->hrp_thrs == NULL)
3007 GOTO(out, rc = -ENOMEM);
3009 for (i = 0; i < hrp->hrp_nthrs; i++) {
3010 hrt = &hrp->hrp_thrs[i];
3013 hrt->hrt_partition = hrp;
3014 init_waitqueue_head(&hrt->hrt_waitq);
3015 spin_lock_init(&hrt->hrt_lock);
3016 INIT_LIST_HEAD(&hrt->hrt_queue);
3020 rc = ptlrpc_start_hr_threads();
3027 void ptlrpc_hr_fini(void)
3029 struct ptlrpc_hr_partition *hrp;
3032 if (ptlrpc_hr.hr_partitions == NULL)
3035 ptlrpc_stop_hr_threads();
3037 cfs_percpt_for_each(hrp, cpt, ptlrpc_hr.hr_partitions) {
3038 if (hrp->hrp_thrs != NULL) {
3039 OBD_FREE(hrp->hrp_thrs,
3040 hrp->hrp_nthrs * sizeof(hrp->hrp_thrs[0]));
3044 cfs_percpt_free(ptlrpc_hr.hr_partitions);
3045 ptlrpc_hr.hr_partitions = NULL;
3050 * Wait until all already scheduled replies are processed.
3052 static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
3056 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
3059 rc = l_wait_event(svcpt->scp_waitq,
3060 atomic_read(&svcpt->scp_nreps_difficult) == 0, &lwi);
3063 CWARN("Unexpectedly long timeout %s %p\n",
3064 svcpt->scp_service->srv_name, svcpt->scp_service);
3069 ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
3071 struct ptlrpc_service_part *svcpt;
3074 /* early disarm AT timer... */
3075 ptlrpc_service_for_each_part(svcpt, i, svc) {
3076 if (svcpt->scp_service != NULL)
3077 del_timer(&svcpt->scp_at_timer);
3082 ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
3084 struct ptlrpc_service_part *svcpt;
3085 struct ptlrpc_request_buffer_desc *rqbd;
3086 struct l_wait_info lwi;
3090 /* All history will be culled when the next request buffer is
3091 * freed in ptlrpc_service_purge_all() */
3092 svc->srv_hist_nrqbds_cpt_max = 0;
3094 rc = LNetClearLazyPortal(svc->srv_req_portal);
3097 ptlrpc_service_for_each_part(svcpt, i, svc) {
3098 if (svcpt->scp_service == NULL)
3101 /* Unlink all the request buffers. This forces a 'final'
3102 * event with its 'unlink' flag set for each posted rqbd */
3103 list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
3105 rc = LNetMDUnlink(rqbd->rqbd_md_h);
3106 LASSERT(rc == 0 || rc == -ENOENT);
3110 ptlrpc_service_for_each_part(svcpt, i, svc) {
3111 if (svcpt->scp_service == NULL)
3114 /* Wait for the network to release any buffers
3115 * it's currently filling */
3116 spin_lock(&svcpt->scp_lock);
3117 while (svcpt->scp_nrqbds_posted != 0) {
3118 spin_unlock(&svcpt->scp_lock);
3119 /* Network access will complete in finite time but
3120 * the HUGE timeout lets us CWARN for visibility
3121 * of sluggish NALs */
3122 lwi = LWI_TIMEOUT_INTERVAL(
3123 cfs_time_seconds(LONG_UNLINK),
3124 cfs_time_seconds(1), NULL, NULL);
3125 rc = l_wait_event(svcpt->scp_waitq,
3126 svcpt->scp_nrqbds_posted == 0, &lwi);
3127 if (rc == -ETIMEDOUT) {
3128 CWARN("Service %s waiting for "
3129 "request buffers\n",
3130 svcpt->scp_service->srv_name);
3132 spin_lock(&svcpt->scp_lock);
3134 spin_unlock(&svcpt->scp_lock);
3139 ptlrpc_service_purge_all(struct ptlrpc_service *svc)
3141 struct ptlrpc_service_part *svcpt;
3142 struct ptlrpc_request_buffer_desc *rqbd;
3143 struct ptlrpc_request *req;
3144 struct ptlrpc_reply_state *rs;
3147 ptlrpc_service_for_each_part(svcpt, i, svc) {
3148 if (svcpt->scp_service == NULL)
3151 spin_lock(&svcpt->scp_rep_lock);
3152 while (!list_empty(&svcpt->scp_rep_active)) {
3153 rs = list_entry(svcpt->scp_rep_active.next,
3154 struct ptlrpc_reply_state, rs_list);
3155 spin_lock(&rs->rs_lock);
3156 ptlrpc_schedule_difficult_reply(rs);
3157 spin_unlock(&rs->rs_lock);
3159 spin_unlock(&svcpt->scp_rep_lock);
3161 /* purge the request queue. NB No new replies (rqbds
3162 * all unlinked) and no service threads, so I'm the only
3163 * thread noodling the request queue now */
3164 while (!list_empty(&svcpt->scp_req_incoming)) {
3165 req = list_entry(svcpt->scp_req_incoming.next,
3166 struct ptlrpc_request, rq_list);
3168 list_del(&req->rq_list);
3169 svcpt->scp_nreqs_incoming--;
3170 ptlrpc_server_finish_request(svcpt, req);
3173 while (ptlrpc_server_request_pending(svcpt, true)) {
3174 req = ptlrpc_server_request_get(svcpt, true);
3175 ptlrpc_server_finish_active_request(svcpt, req);
3178 LASSERT(list_empty(&svcpt->scp_rqbd_posted));
3179 LASSERT(svcpt->scp_nreqs_incoming == 0);
3180 LASSERT(svcpt->scp_nreqs_active == 0);
3181 /* history should have been culled by
3182 * ptlrpc_server_finish_request */
3183 LASSERT(svcpt->scp_hist_nrqbds == 0);
3185 /* Now free all the request buffers since nothing
3186 * references them any more... */
3188 while (!list_empty(&svcpt->scp_rqbd_idle)) {
3189 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
3190 struct ptlrpc_request_buffer_desc,
3192 ptlrpc_free_rqbd(rqbd);
3194 ptlrpc_wait_replies(svcpt);
3196 while (!list_empty(&svcpt->scp_rep_idle)) {
3197 rs = list_entry(svcpt->scp_rep_idle.next,
3198 struct ptlrpc_reply_state,
3200 list_del(&rs->rs_list);
3201 OBD_FREE_LARGE(rs, svc->srv_max_reply_size);
3207 ptlrpc_service_free(struct ptlrpc_service *svc)
3209 struct ptlrpc_service_part *svcpt;
3210 struct ptlrpc_at_array *array;
3213 ptlrpc_service_for_each_part(svcpt, i, svc) {
3214 if (svcpt->scp_service == NULL)
3217 /* In case somebody rearmed this in the meantime */
3218 del_timer(&svcpt->scp_at_timer);
3219 array = &svcpt->scp_at_array;
3221 if (array->paa_reqs_array != NULL) {
3222 OBD_FREE(array->paa_reqs_array,
3223 sizeof(struct list_head) * array->paa_size);
3224 array->paa_reqs_array = NULL;
3227 if (array->paa_reqs_count != NULL) {
3228 OBD_FREE(array->paa_reqs_count,
3229 sizeof(__u32) * array->paa_size);
3230 array->paa_reqs_count = NULL;
3234 ptlrpc_service_for_each_part(svcpt, i, svc)
3235 OBD_FREE_PTR(svcpt);
3237 if (svc->srv_cpts != NULL)
3238 cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts);
3240 OBD_FREE(svc, offsetof(struct ptlrpc_service,
3241 srv_parts[svc->srv_ncpts]));
3244 int ptlrpc_unregister_service(struct ptlrpc_service *service)
3248 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
3250 service->srv_is_stopping = 1;
3252 mutex_lock(&ptlrpc_all_services_mutex);
3253 list_del_init(&service->srv_list);
3254 mutex_unlock(&ptlrpc_all_services_mutex);
3256 ptlrpc_service_del_atimer(service);
3257 ptlrpc_stop_all_threads(service);
3259 ptlrpc_service_unlink_rqbd(service);
3260 ptlrpc_service_purge_all(service);
3261 ptlrpc_service_nrs_cleanup(service);
3263 ptlrpc_lprocfs_unregister_service(service);
3264 ptlrpc_sysfs_unregister_service(service);
3266 ptlrpc_service_free(service);
3270 EXPORT_SYMBOL(ptlrpc_unregister_service);
3273 * Returns 0 if the service is healthy.
3275 * Right now, it just checks to make sure that requests aren't languishing
3276 * in the queue. We'll use this health check to govern whether a node needs
3277 * to be shot, so it's intentionally non-aggressive. */
3278 static int ptlrpc_svcpt_health_check(struct ptlrpc_service_part *svcpt)
3280 struct ptlrpc_request *request = NULL;
3281 struct timespec64 right_now;
3282 struct timespec64 timediff;
3284 ktime_get_real_ts64(&right_now);
3286 spin_lock(&svcpt->scp_req_lock);
3287 /* How long has the next entry been waiting? */
3288 if (ptlrpc_server_high_pending(svcpt, true))
3289 request = ptlrpc_nrs_req_peek_nolock(svcpt, true);
3290 else if (ptlrpc_server_normal_pending(svcpt, true))
3291 request = ptlrpc_nrs_req_peek_nolock(svcpt, false);
3293 if (request == NULL) {
3294 spin_unlock(&svcpt->scp_req_lock);
3298 timediff = timespec64_sub(right_now, request->rq_arrival_time);
3299 spin_unlock(&svcpt->scp_req_lock);
3301 if ((timediff.tv_sec) >
3302 (AT_OFF ? obd_timeout * 3 / 2 : at_max)) {
3303 CERROR("%s: unhealthy - request has been waiting %llds\n",
3304 svcpt->scp_service->srv_name, (s64)timediff.tv_sec);
3312 ptlrpc_service_health_check(struct ptlrpc_service *svc)
3314 struct ptlrpc_service_part *svcpt;
3320 ptlrpc_service_for_each_part(svcpt, i, svc) {
3321 int rc = ptlrpc_svcpt_health_check(svcpt);
3328 EXPORT_SYMBOL(ptlrpc_service_health_check);