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);
102 ptlrpc_free_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
104 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
106 LASSERT(rqbd->rqbd_refcount == 0);
107 LASSERT(list_empty(&rqbd->rqbd_reqs));
109 spin_lock(&svcpt->scp_lock);
110 list_del(&rqbd->rqbd_list);
111 svcpt->scp_nrqbds_total--;
112 spin_unlock(&svcpt->scp_lock);
114 OBD_FREE_LARGE(rqbd->rqbd_buffer, svcpt->scp_service->srv_buf_size);
119 ptlrpc_grow_req_bufs(struct ptlrpc_service_part *svcpt, int post)
121 struct ptlrpc_service *svc = svcpt->scp_service;
122 struct ptlrpc_request_buffer_desc *rqbd;
126 if (svcpt->scp_rqbd_allocating)
129 spin_lock(&svcpt->scp_lock);
130 /* check again with lock */
131 if (svcpt->scp_rqbd_allocating) {
132 /* NB: we might allow more than one thread in the future */
133 LASSERT(svcpt->scp_rqbd_allocating == 1);
134 spin_unlock(&svcpt->scp_lock);
138 svcpt->scp_rqbd_allocating++;
139 spin_unlock(&svcpt->scp_lock);
142 for (i = 0; i < svc->srv_nbuf_per_group; i++) {
143 /* NB: another thread might have recycled enough rqbds, we
144 * need to make sure it wouldn't over-allocate, see LU-1212. */
145 if (svcpt->scp_nrqbds_posted >= svc->srv_nbuf_per_group ||
146 (svc->srv_nrqbds_max != 0 &&
147 svcpt->scp_nrqbds_total > svc->srv_nrqbds_max))
150 rqbd = ptlrpc_alloc_rqbd(svcpt);
153 CERROR("%s: Can't allocate request buffer\n",
160 spin_lock(&svcpt->scp_lock);
162 LASSERT(svcpt->scp_rqbd_allocating == 1);
163 svcpt->scp_rqbd_allocating--;
165 spin_unlock(&svcpt->scp_lock);
168 "%s: allocate %d new %d-byte reqbufs (%d/%d left), rc = %d\n",
169 svc->srv_name, i, svc->srv_buf_size, svcpt->scp_nrqbds_posted,
170 svcpt->scp_nrqbds_total, rc);
174 rc = ptlrpc_server_post_idle_rqbds(svcpt);
180 * Part of Rep-Ack logic.
181 * Puts a lock and its mode into reply state assotiated to request reply.
184 ptlrpc_save_lock(struct ptlrpc_request *req, struct lustre_handle *lock,
185 int mode, bool no_ack, bool convert_lock)
187 struct ptlrpc_reply_state *rs = req->rq_reply_state;
191 LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
193 idx = rs->rs_nlocks++;
194 rs->rs_locks[idx] = *lock;
195 rs->rs_modes[idx] = mode;
196 rs->rs_difficult = 1;
197 rs->rs_no_ack = no_ack;
198 rs->rs_convert_lock = convert_lock;
200 EXPORT_SYMBOL(ptlrpc_save_lock);
203 struct ptlrpc_hr_partition;
205 struct ptlrpc_hr_thread {
206 int hrt_id; /* thread ID */
208 wait_queue_head_t hrt_waitq;
209 struct list_head hrt_queue;
210 struct ptlrpc_hr_partition *hrt_partition;
213 struct ptlrpc_hr_partition {
214 /* # of started threads */
215 atomic_t hrp_nstarted;
216 /* # of stopped threads */
217 atomic_t hrp_nstopped;
218 /* cpu partition id */
220 /* round-robin rotor for choosing thread */
222 /* total number of threads on this partition */
225 struct ptlrpc_hr_thread *hrp_thrs;
228 #define HRT_RUNNING 0
229 #define HRT_STOPPING 1
231 struct ptlrpc_hr_service {
232 /* CPU partition table, it's just cfs_cpt_table for now */
233 struct cfs_cpt_table *hr_cpt_table;
234 /** controller sleep waitq */
235 wait_queue_head_t hr_waitq;
236 unsigned int hr_stopping;
237 /** roundrobin rotor for non-affinity service */
238 unsigned int hr_rotor;
240 struct ptlrpc_hr_partition **hr_partitions;
244 struct list_head rsb_replies;
245 unsigned int rsb_n_replies;
246 struct ptlrpc_service_part *rsb_svcpt;
249 /** reply handling service. */
250 static struct ptlrpc_hr_service ptlrpc_hr;
253 * maximum mumber of replies scheduled in one batch
255 #define MAX_SCHEDULED 256
258 * Initialize a reply batch.
262 static void rs_batch_init(struct rs_batch *b)
264 memset(b, 0, sizeof *b);
265 INIT_LIST_HEAD(&b->rsb_replies);
269 * Choose an hr thread to dispatch requests to.
271 static struct ptlrpc_hr_thread *
272 ptlrpc_hr_select(struct ptlrpc_service_part *svcpt)
274 struct ptlrpc_hr_partition *hrp;
277 if (svcpt->scp_cpt >= 0 &&
278 svcpt->scp_service->srv_cptable == ptlrpc_hr.hr_cpt_table) {
279 /* directly match partition */
280 hrp = ptlrpc_hr.hr_partitions[svcpt->scp_cpt];
283 rotor = ptlrpc_hr.hr_rotor++;
284 rotor %= cfs_cpt_number(ptlrpc_hr.hr_cpt_table);
286 hrp = ptlrpc_hr.hr_partitions[rotor];
289 rotor = hrp->hrp_rotor++;
290 return &hrp->hrp_thrs[rotor % hrp->hrp_nthrs];
294 * Dispatch all replies accumulated in the batch to one from
295 * dedicated reply handling threads.
299 static void rs_batch_dispatch(struct rs_batch *b)
301 if (b->rsb_n_replies != 0) {
302 struct ptlrpc_hr_thread *hrt;
304 hrt = ptlrpc_hr_select(b->rsb_svcpt);
306 spin_lock(&hrt->hrt_lock);
307 list_splice_init(&b->rsb_replies, &hrt->hrt_queue);
308 spin_unlock(&hrt->hrt_lock);
310 wake_up(&hrt->hrt_waitq);
311 b->rsb_n_replies = 0;
316 * Add a reply to a batch.
317 * Add one reply object to a batch, schedule batched replies if overload.
322 static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs)
324 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
326 if (svcpt != b->rsb_svcpt || b->rsb_n_replies >= MAX_SCHEDULED) {
327 if (b->rsb_svcpt != NULL) {
328 rs_batch_dispatch(b);
329 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
331 spin_lock(&svcpt->scp_rep_lock);
332 b->rsb_svcpt = svcpt;
334 spin_lock(&rs->rs_lock);
335 rs->rs_scheduled_ever = 1;
336 if (rs->rs_scheduled == 0) {
337 list_move(&rs->rs_list, &b->rsb_replies);
338 rs->rs_scheduled = 1;
341 rs->rs_committed = 1;
342 spin_unlock(&rs->rs_lock);
346 * Reply batch finalization.
347 * Dispatch remaining replies from the batch
348 * and release remaining spinlock.
352 static void rs_batch_fini(struct rs_batch *b)
354 if (b->rsb_svcpt != NULL) {
355 rs_batch_dispatch(b);
356 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
360 #define DECLARE_RS_BATCH(b) struct rs_batch b
364 * Put reply state into a queue for processing because we received
365 * ACK from the client
367 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
369 struct ptlrpc_hr_thread *hrt;
372 LASSERT(list_empty(&rs->rs_list));
374 hrt = ptlrpc_hr_select(rs->rs_svcpt);
376 spin_lock(&hrt->hrt_lock);
377 list_add_tail(&rs->rs_list, &hrt->hrt_queue);
378 spin_unlock(&hrt->hrt_lock);
380 wake_up(&hrt->hrt_waitq);
385 ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs)
389 assert_spin_locked(&rs->rs_svcpt->scp_rep_lock);
390 assert_spin_locked(&rs->rs_lock);
391 LASSERT (rs->rs_difficult);
392 rs->rs_scheduled_ever = 1; /* flag any notification attempt */
394 if (rs->rs_scheduled) { /* being set up or already notified */
399 rs->rs_scheduled = 1;
400 list_del_init(&rs->rs_list);
401 ptlrpc_dispatch_difficult_reply(rs);
404 EXPORT_SYMBOL(ptlrpc_schedule_difficult_reply);
406 void ptlrpc_commit_replies(struct obd_export *exp)
408 struct ptlrpc_reply_state *rs, *nxt;
409 DECLARE_RS_BATCH(batch);
412 rs_batch_init(&batch);
413 /* Find any replies that have been committed and get their service
414 * to attend to complete them. */
416 /* CAVEAT EMPTOR: spinlock ordering!!! */
417 spin_lock(&exp->exp_uncommitted_replies_lock);
418 list_for_each_entry_safe(rs, nxt, &exp->exp_uncommitted_replies,
420 LASSERT (rs->rs_difficult);
421 /* VBR: per-export last_committed */
422 LASSERT(rs->rs_export);
423 if (rs->rs_transno <= exp->exp_last_committed) {
424 list_del_init(&rs->rs_obd_list);
425 rs_batch_add(&batch, rs);
428 spin_unlock(&exp->exp_uncommitted_replies_lock);
429 rs_batch_fini(&batch);
434 ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt)
436 struct ptlrpc_request_buffer_desc *rqbd;
441 spin_lock(&svcpt->scp_lock);
443 if (list_empty(&svcpt->scp_rqbd_idle)) {
444 spin_unlock(&svcpt->scp_lock);
448 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
449 struct ptlrpc_request_buffer_desc,
451 list_del(&rqbd->rqbd_list);
453 /* assume we will post successfully */
454 svcpt->scp_nrqbds_posted++;
455 list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_posted);
457 spin_unlock(&svcpt->scp_lock);
459 rc = ptlrpc_register_rqbd(rqbd);
466 spin_lock(&svcpt->scp_lock);
468 svcpt->scp_nrqbds_posted--;
469 list_del(&rqbd->rqbd_list);
470 list_add_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
472 /* Don't complain if no request buffers are posted right now; LNET
473 * won't drop requests because we set the portal lazy! */
475 spin_unlock(&svcpt->scp_lock);
480 static void ptlrpc_at_timer(cfs_timer_cb_arg_t data)
482 struct ptlrpc_service_part *svcpt;
484 svcpt = cfs_from_timer(svcpt, data, scp_at_timer);
486 svcpt->scp_at_check = 1;
487 svcpt->scp_at_checktime = ktime_get();
488 wake_up(&svcpt->scp_waitq);
492 ptlrpc_server_nthreads_check(struct ptlrpc_service *svc,
493 struct ptlrpc_service_conf *conf)
495 struct ptlrpc_service_thr_conf *tc = &conf->psc_thr;
502 * Common code for estimating & validating threads number.
503 * CPT affinity service could have percpt thread-pool instead
504 * of a global thread-pool, which means user might not always
505 * get the threads number they give it in conf::tc_nthrs_user
506 * even they did set. It's because we need to validate threads
507 * number for each CPT to guarantee each pool will have enough
508 * threads to keep the service healthy.
510 init = PTLRPC_NTHRS_INIT + (svc->srv_ops.so_hpreq_handler != NULL);
511 init = max_t(int, init, tc->tc_nthrs_init);
513 /* NB: please see comments in lustre_lnet.h for definition
514 * details of these members */
515 LASSERT(tc->tc_nthrs_max != 0);
517 if (tc->tc_nthrs_user != 0) {
518 /* In case there is a reason to test a service with many
519 * threads, we give a less strict check here, it can
520 * be up to 8 * nthrs_max */
521 total = min(tc->tc_nthrs_max * 8, tc->tc_nthrs_user);
522 nthrs = total / svc->srv_ncpts;
523 init = max(init, nthrs);
527 total = tc->tc_nthrs_max;
528 if (tc->tc_nthrs_base == 0) {
529 /* don't care about base threads number per partition,
530 * this is most for non-affinity service */
531 nthrs = total / svc->srv_ncpts;
535 nthrs = tc->tc_nthrs_base;
536 if (svc->srv_ncpts == 1) {
539 /* NB: Increase the base number if it's single partition
540 * and total number of cores/HTs is larger or equal to 4.
541 * result will always < 2 * nthrs_base */
542 weight = cfs_cpt_weight(svc->srv_cptable, CFS_CPT_ANY);
543 for (i = 1; (weight >> (i + 1)) != 0 && /* >= 4 cores/HTs */
544 (tc->tc_nthrs_base >> i) != 0; i++)
545 nthrs += tc->tc_nthrs_base >> i;
548 if (tc->tc_thr_factor != 0) {
549 int factor = tc->tc_thr_factor;
553 * User wants to increase number of threads with for
554 * each CPU core/HT, most likely the factor is larger than
555 * one thread/core because service threads are supposed to
556 * be blocked by lock or wait for IO.
559 * Amdahl's law says that adding processors wouldn't give
560 * a linear increasing of parallelism, so it's nonsense to
561 * have too many threads no matter how many cores/HTs
564 if (cpumask_weight(topology_sibling_cpumask(smp_processor_id())) > 1) {
565 /* weight is # of HTs */
566 /* depress thread factor for hyper-thread */
567 factor = factor - (factor >> 1) + (factor >> 3);
570 weight = cfs_cpt_weight(svc->srv_cptable, 0);
572 for (; factor > 0 && weight > 0; factor--, weight -= fade)
573 nthrs += min(weight, fade) * factor;
576 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
577 nthrs = max(tc->tc_nthrs_base,
578 tc->tc_nthrs_max / svc->srv_ncpts);
581 nthrs = max(nthrs, tc->tc_nthrs_init);
582 svc->srv_nthrs_cpt_limit = nthrs;
583 svc->srv_nthrs_cpt_init = init;
585 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
586 CDEBUG(D_OTHER, "%s: This service may have more threads (%d) "
587 "than the given soft limit (%d)\n",
588 svc->srv_name, nthrs * svc->srv_ncpts,
594 * Initialize percpt data for a service
597 ptlrpc_service_part_init(struct ptlrpc_service *svc,
598 struct ptlrpc_service_part *svcpt, int cpt)
600 struct ptlrpc_at_array *array;
605 svcpt->scp_cpt = cpt;
606 INIT_LIST_HEAD(&svcpt->scp_threads);
608 /* rqbd and incoming request queue */
609 spin_lock_init(&svcpt->scp_lock);
610 mutex_init(&svcpt->scp_mutex);
611 INIT_LIST_HEAD(&svcpt->scp_rqbd_idle);
612 INIT_LIST_HEAD(&svcpt->scp_rqbd_posted);
613 INIT_LIST_HEAD(&svcpt->scp_req_incoming);
614 init_waitqueue_head(&svcpt->scp_waitq);
615 /* history request & rqbd list */
616 INIT_LIST_HEAD(&svcpt->scp_hist_reqs);
617 INIT_LIST_HEAD(&svcpt->scp_hist_rqbds);
619 /* acitve requests and hp requests */
620 spin_lock_init(&svcpt->scp_req_lock);
623 spin_lock_init(&svcpt->scp_rep_lock);
624 INIT_LIST_HEAD(&svcpt->scp_rep_active);
625 INIT_LIST_HEAD(&svcpt->scp_rep_idle);
626 init_waitqueue_head(&svcpt->scp_rep_waitq);
627 atomic_set(&svcpt->scp_nreps_difficult, 0);
629 /* adaptive timeout */
630 spin_lock_init(&svcpt->scp_at_lock);
631 array = &svcpt->scp_at_array;
633 size = at_est2timeout(at_max);
634 array->paa_size = size;
635 array->paa_count = 0;
636 array->paa_deadline = -1;
638 /* allocate memory for scp_at_array (ptlrpc_at_array) */
639 OBD_CPT_ALLOC(array->paa_reqs_array,
640 svc->srv_cptable, cpt, sizeof(struct list_head) * size);
641 if (array->paa_reqs_array == NULL)
644 for (index = 0; index < size; index++)
645 INIT_LIST_HEAD(&array->paa_reqs_array[index]);
647 OBD_CPT_ALLOC(array->paa_reqs_count,
648 svc->srv_cptable, cpt, sizeof(__u32) * size);
649 if (array->paa_reqs_count == NULL)
652 cfs_timer_setup(&svcpt->scp_at_timer, ptlrpc_at_timer,
653 (unsigned long)svcpt, 0);
655 /* At SOW, service time should be quick; 10s seems generous. If client
656 * timeout is less than this, we'll be sending an early reply. */
657 at_init(&svcpt->scp_at_estimate, 10, 0);
659 /* assign this before call ptlrpc_grow_req_bufs */
660 svcpt->scp_service = svc;
661 /* Now allocate the request buffers, but don't post them now */
662 rc = ptlrpc_grow_req_bufs(svcpt, 0);
663 /* We shouldn't be under memory pressure at startup, so
664 * fail if we can't allocate all our buffers at this time. */
671 if (array->paa_reqs_count != NULL) {
672 OBD_FREE(array->paa_reqs_count, sizeof(__u32) * size);
673 array->paa_reqs_count = NULL;
676 if (array->paa_reqs_array != NULL) {
677 OBD_FREE(array->paa_reqs_array,
678 sizeof(struct list_head) * array->paa_size);
679 array->paa_reqs_array = NULL;
686 * Initialize service on a given portal.
687 * This includes starting serving threads , allocating and posting rqbds and
690 struct ptlrpc_service *
691 ptlrpc_register_service(struct ptlrpc_service_conf *conf,
693 struct dentry *debugfs_entry)
695 struct ptlrpc_service_cpt_conf *cconf = &conf->psc_cpt;
696 struct ptlrpc_service *service;
697 struct ptlrpc_service_part *svcpt;
698 struct cfs_cpt_table *cptable;
706 LASSERT(conf->psc_buf.bc_nbufs > 0);
707 LASSERT(conf->psc_buf.bc_buf_size >=
708 conf->psc_buf.bc_req_max_size + SPTLRPC_MAX_PAYLOAD);
709 LASSERT(conf->psc_thr.tc_ctx_tags != 0);
711 cptable = cconf->cc_cptable;
713 cptable = cfs_cpt_table;
715 if (conf->psc_thr.tc_cpu_bind > 1) {
716 CERROR("%s: Invalid cpu bind value %d, only 1 or 0 allowed\n",
717 conf->psc_name, conf->psc_thr.tc_cpu_bind);
718 RETURN(ERR_PTR(-EINVAL));
721 if (!cconf->cc_affinity) {
724 ncpts = cfs_cpt_number(cptable);
725 if (cconf->cc_pattern != NULL) {
726 struct cfs_expr_list *el;
728 rc = cfs_expr_list_parse(cconf->cc_pattern,
729 strlen(cconf->cc_pattern),
732 CERROR("%s: invalid CPT pattern string: %s",
733 conf->psc_name, cconf->cc_pattern);
734 RETURN(ERR_PTR(-EINVAL));
737 rc = cfs_expr_list_values(el, ncpts, &cpts);
738 cfs_expr_list_free(el);
740 CERROR("%s: failed to parse CPT array %s: %d\n",
741 conf->psc_name, cconf->cc_pattern, rc);
743 OBD_FREE(cpts, sizeof(*cpts) * ncpts);
744 RETURN(ERR_PTR(rc < 0 ? rc : -EINVAL));
750 OBD_ALLOC(service, offsetof(struct ptlrpc_service, srv_parts[ncpts]));
751 if (service == NULL) {
753 OBD_FREE(cpts, sizeof(*cpts) * ncpts);
754 RETURN(ERR_PTR(-ENOMEM));
757 service->srv_cptable = cptable;
758 service->srv_cpts = cpts;
759 service->srv_ncpts = ncpts;
760 service->srv_cpt_bind = conf->psc_thr.tc_cpu_bind;
762 service->srv_cpt_bits = 0; /* it's zero already, easy to read... */
763 while ((1 << service->srv_cpt_bits) < cfs_cpt_number(cptable))
764 service->srv_cpt_bits++;
767 spin_lock_init(&service->srv_lock);
768 service->srv_name = conf->psc_name;
769 service->srv_watchdog_factor = conf->psc_watchdog_factor;
770 INIT_LIST_HEAD(&service->srv_list); /* for safty of cleanup */
772 /* buffer configuration */
773 service->srv_nbuf_per_group = test_req_buffer_pressure ?
774 1 : conf->psc_buf.bc_nbufs;
775 /* do not limit max number of rqbds by default */
776 service->srv_nrqbds_max = 0;
778 service->srv_max_req_size = conf->psc_buf.bc_req_max_size +
780 service->srv_buf_size = conf->psc_buf.bc_buf_size;
781 service->srv_rep_portal = conf->psc_buf.bc_rep_portal;
782 service->srv_req_portal = conf->psc_buf.bc_req_portal;
784 /* With slab/alloc_pages buffer size will be rounded up to 2^n */
785 if (service->srv_buf_size & (service->srv_buf_size - 1)) {
786 int round = size_roundup_power2(service->srv_buf_size);
787 service->srv_buf_size = round;
790 /* Increase max reply size to next power of two */
791 service->srv_max_reply_size = 1;
792 while (service->srv_max_reply_size <
793 conf->psc_buf.bc_rep_max_size + SPTLRPC_MAX_PAYLOAD)
794 service->srv_max_reply_size <<= 1;
796 service->srv_thread_name = conf->psc_thr.tc_thr_name;
797 service->srv_ctx_tags = conf->psc_thr.tc_ctx_tags;
798 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
799 service->srv_ops = conf->psc_ops;
801 for (i = 0; i < ncpts; i++) {
802 if (!cconf->cc_affinity)
805 cpt = cpts != NULL ? cpts[i] : i;
807 OBD_CPT_ALLOC(svcpt, cptable, cpt, sizeof(*svcpt));
809 GOTO(failed, rc = -ENOMEM);
811 service->srv_parts[i] = svcpt;
812 rc = ptlrpc_service_part_init(service, svcpt, cpt);
817 ptlrpc_server_nthreads_check(service, conf);
819 rc = LNetSetLazyPortal(service->srv_req_portal);
822 mutex_lock(&ptlrpc_all_services_mutex);
823 list_add(&service->srv_list, &ptlrpc_all_services);
824 mutex_unlock(&ptlrpc_all_services_mutex);
827 rc = ptlrpc_sysfs_register_service(parent, service);
832 if (debugfs_entry != NULL)
833 ptlrpc_ldebugfs_register_service(debugfs_entry, service);
835 rc = ptlrpc_service_nrs_setup(service);
839 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
840 service->srv_name, service->srv_req_portal);
842 rc = ptlrpc_start_threads(service);
844 CERROR("Failed to start threads for service %s: %d\n",
845 service->srv_name, rc);
851 ptlrpc_unregister_service(service);
854 EXPORT_SYMBOL(ptlrpc_register_service);
857 * to actually free the request, must be called without holding svc_lock.
858 * note it's caller's responsibility to unlink req->rq_list.
860 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
862 LASSERT(atomic_read(&req->rq_refcount) == 0);
863 LASSERT(list_empty(&req->rq_timed_list));
865 /* DEBUG_REQ() assumes the reply state of a request with a valid
866 * ref will not be destroyed until that reference is dropped. */
867 ptlrpc_req_drop_rs(req);
869 sptlrpc_svc_ctx_decref(req);
871 if (req != &req->rq_rqbd->rqbd_req) {
872 /* NB request buffers use an embedded
873 * req if the incoming req unlinked the
874 * MD; this isn't one of them! */
875 ptlrpc_request_cache_free(req);
880 * drop a reference count of the request. if it reaches 0, we either
881 * put it into history list, or free it immediately.
883 void ptlrpc_server_drop_request(struct ptlrpc_request *req)
885 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
886 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
887 struct ptlrpc_service *svc = svcpt->scp_service;
889 struct list_head *tmp;
890 struct list_head *nxt;
892 if (!atomic_dec_and_test(&req->rq_refcount))
895 if (req->rq_session.lc_state == LCS_ENTERED) {
896 lu_context_exit(&req->rq_session);
897 lu_context_fini(&req->rq_session);
900 if (req->rq_at_linked) {
901 spin_lock(&svcpt->scp_at_lock);
902 /* recheck with lock, in case it's unlinked by
903 * ptlrpc_at_check_timed() */
904 if (likely(req->rq_at_linked))
905 ptlrpc_at_remove_timed(req);
906 spin_unlock(&svcpt->scp_at_lock);
909 LASSERT(list_empty(&req->rq_timed_list));
911 /* finalize request */
912 if (req->rq_export) {
913 class_export_put(req->rq_export);
914 req->rq_export = NULL;
917 spin_lock(&svcpt->scp_lock);
919 list_add(&req->rq_list, &rqbd->rqbd_reqs);
921 refcount = --(rqbd->rqbd_refcount);
923 /* request buffer is now idle: add to history */
924 list_del(&rqbd->rqbd_list);
926 list_add_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds);
927 svcpt->scp_hist_nrqbds++;
929 /* cull some history?
930 * I expect only about 1 or 2 rqbds need to be recycled here */
931 while (svcpt->scp_hist_nrqbds > svc->srv_hist_nrqbds_cpt_max) {
932 rqbd = list_entry(svcpt->scp_hist_rqbds.next,
933 struct ptlrpc_request_buffer_desc,
936 list_del(&rqbd->rqbd_list);
937 svcpt->scp_hist_nrqbds--;
939 /* remove rqbd's reqs from svc's req history while
940 * I've got the service lock */
941 list_for_each(tmp, &rqbd->rqbd_reqs) {
942 req = list_entry(tmp, struct ptlrpc_request,
944 /* Track the highest culled req seq */
945 if (req->rq_history_seq >
946 svcpt->scp_hist_seq_culled) {
947 svcpt->scp_hist_seq_culled =
950 list_del(&req->rq_history_list);
953 spin_unlock(&svcpt->scp_lock);
955 list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
956 req = list_entry(rqbd->rqbd_reqs.next,
957 struct ptlrpc_request,
959 list_del(&req->rq_list);
960 ptlrpc_server_free_request(req);
963 spin_lock(&svcpt->scp_lock);
965 * now all reqs including the embedded req has been
966 * disposed, schedule request buffer for re-use
967 * or free it to drain some in excess.
969 LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) == 0);
970 if (svcpt->scp_nrqbds_posted >=
971 svc->srv_nbuf_per_group ||
972 (svc->srv_nrqbds_max != 0 &&
973 svcpt->scp_nrqbds_total > svc->srv_nrqbds_max) ||
974 test_req_buffer_pressure) {
975 /* like in ptlrpc_free_rqbd() */
976 svcpt->scp_nrqbds_total--;
977 OBD_FREE_LARGE(rqbd->rqbd_buffer,
981 list_add_tail(&rqbd->rqbd_list,
982 &svcpt->scp_rqbd_idle);
986 spin_unlock(&svcpt->scp_lock);
987 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
988 /* If we are low on memory, we are not interested in history */
989 list_del(&req->rq_list);
990 list_del_init(&req->rq_history_list);
992 /* Track the highest culled req seq */
993 if (req->rq_history_seq > svcpt->scp_hist_seq_culled)
994 svcpt->scp_hist_seq_culled = req->rq_history_seq;
996 spin_unlock(&svcpt->scp_lock);
998 ptlrpc_server_free_request(req);
1000 spin_unlock(&svcpt->scp_lock);
1004 /** Change request export and move hp request from old export to new */
1005 void ptlrpc_request_change_export(struct ptlrpc_request *req,
1006 struct obd_export *export)
1008 if (req->rq_export != NULL) {
1009 LASSERT(!list_empty(&req->rq_exp_list));
1010 /* remove rq_exp_list from last export */
1011 spin_lock(&req->rq_export->exp_rpc_lock);
1012 list_del_init(&req->rq_exp_list);
1013 spin_unlock(&req->rq_export->exp_rpc_lock);
1014 /* export has one reference already, so it`s safe to
1015 * add req to export queue here and get another
1016 * reference for request later */
1017 spin_lock(&export->exp_rpc_lock);
1018 if (req->rq_ops != NULL) /* hp request */
1019 list_add(&req->rq_exp_list, &export->exp_hp_rpcs);
1021 list_add(&req->rq_exp_list, &export->exp_reg_rpcs);
1022 spin_unlock(&export->exp_rpc_lock);
1024 class_export_rpc_dec(req->rq_export);
1025 class_export_put(req->rq_export);
1028 /* request takes one export refcount */
1029 req->rq_export = class_export_get(export);
1030 class_export_rpc_inc(export);
1036 * to finish a request: stop sending more early replies, and release
1039 static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt,
1040 struct ptlrpc_request *req)
1042 ptlrpc_server_hpreq_fini(req);
1044 ptlrpc_server_drop_request(req);
1048 * to finish an active request: stop sending more early replies, and release
1049 * the request. should be called after we finished handling the request.
1051 static void ptlrpc_server_finish_active_request(
1052 struct ptlrpc_service_part *svcpt,
1053 struct ptlrpc_request *req)
1055 spin_lock(&svcpt->scp_req_lock);
1056 ptlrpc_nrs_req_stop_nolock(req);
1057 svcpt->scp_nreqs_active--;
1059 svcpt->scp_nhreqs_active--;
1060 spin_unlock(&svcpt->scp_req_lock);
1062 ptlrpc_nrs_req_finalize(req);
1064 if (req->rq_export != NULL)
1065 class_export_rpc_dec(req->rq_export);
1067 ptlrpc_server_finish_request(svcpt, req);
1071 * This function makes sure dead exports are evicted in a timely manner.
1072 * This function is only called when some export receives a message (i.e.,
1073 * the network is up.)
1075 void ptlrpc_update_export_timer(struct obd_export *exp, time64_t extra_delay)
1077 struct obd_export *oldest_exp;
1078 time64_t oldest_time, new_time;
1084 /* Compensate for slow machines, etc, by faking our request time
1085 into the future. Although this can break the strict time-ordering
1086 of the list, we can be really lazy here - we don't have to evict
1087 at the exact right moment. Eventually, all silent exports
1088 will make it to the top of the list. */
1090 /* Do not pay attention on 1sec or smaller renewals. */
1091 new_time = ktime_get_real_seconds() + extra_delay;
1092 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
1095 exp->exp_last_request_time = new_time;
1097 /* exports may get disconnected from the chain even though the
1098 export has references, so we must keep the spin lock while
1099 manipulating the lists */
1100 spin_lock(&exp->exp_obd->obd_dev_lock);
1102 if (list_empty(&exp->exp_obd_chain_timed)) {
1103 /* this one is not timed */
1104 spin_unlock(&exp->exp_obd->obd_dev_lock);
1108 list_move_tail(&exp->exp_obd_chain_timed,
1109 &exp->exp_obd->obd_exports_timed);
1111 oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
1112 struct obd_export, exp_obd_chain_timed);
1113 oldest_time = oldest_exp->exp_last_request_time;
1114 spin_unlock(&exp->exp_obd->obd_dev_lock);
1116 if (exp->exp_obd->obd_recovering) {
1117 /* be nice to everyone during recovery */
1122 /* Note - racing to start/reset the obd_eviction timer is safe */
1123 if (exp->exp_obd->obd_eviction_timer == 0) {
1124 /* Check if the oldest entry is expired. */
1125 if (ktime_get_real_seconds() >
1126 oldest_time + PING_EVICT_TIMEOUT + extra_delay) {
1127 /* We need a second timer, in case the net was down and
1128 * it just came back. Since the pinger may skip every
1129 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
1130 * we better wait for 3.
1132 exp->exp_obd->obd_eviction_timer =
1133 ktime_get_real_seconds() + 3 * PING_INTERVAL;
1134 CDEBUG(D_HA, "%s: Think about evicting %s from %lld\n",
1135 exp->exp_obd->obd_name,
1136 obd_export_nid2str(oldest_exp), oldest_time);
1139 if (ktime_get_real_seconds() >
1140 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
1141 /* The evictor won't evict anyone who we've heard from
1142 * recently, so we don't have to check before we start
1145 if (!ping_evictor_wake(exp))
1146 exp->exp_obd->obd_eviction_timer = 0;
1154 * Sanity check request \a req.
1155 * Return 0 if all is ok, error code otherwise.
1157 static int ptlrpc_check_req(struct ptlrpc_request *req)
1159 struct obd_device *obd = req->rq_export->exp_obd;
1162 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
1163 req->rq_export->exp_conn_cnt)) {
1164 DEBUG_REQ(D_RPCTRACE, req,
1165 "DROPPING req from old connection %d < %d",
1166 lustre_msg_get_conn_cnt(req->rq_reqmsg),
1167 req->rq_export->exp_conn_cnt);
1170 if (unlikely(obd == NULL || obd->obd_fail)) {
1171 /* Failing over, don't handle any more reqs,
1172 * send error response instead. */
1173 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
1174 req, (obd != NULL) ? obd->obd_name : "unknown");
1176 } else if (lustre_msg_get_flags(req->rq_reqmsg) &
1177 (MSG_REPLAY | MSG_REQ_REPLAY_DONE) &&
1178 !obd->obd_recovering) {
1179 DEBUG_REQ(D_ERROR, req,
1180 "Invalid replay without recovery");
1181 class_fail_export(req->rq_export);
1183 } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 &&
1184 !obd->obd_recovering) {
1185 DEBUG_REQ(D_ERROR, req, "Invalid req with transno "
1186 "%llu without recovery",
1187 lustre_msg_get_transno(req->rq_reqmsg));
1188 class_fail_export(req->rq_export);
1192 if (unlikely(rc < 0)) {
1193 req->rq_status = rc;
1199 static void ptlrpc_at_set_timer(struct ptlrpc_service_part *svcpt)
1201 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1204 if (array->paa_count == 0) {
1205 del_timer(&svcpt->scp_at_timer);
1209 /* Set timer for closest deadline */
1210 next = array->paa_deadline - ktime_get_real_seconds() -
1213 ptlrpc_at_timer(cfs_timer_cb_arg(svcpt, scp_at_timer));
1215 mod_timer(&svcpt->scp_at_timer,
1216 jiffies + nsecs_to_jiffies(next * NSEC_PER_SEC));
1217 CDEBUG(D_INFO, "armed %s at %+llds\n",
1218 svcpt->scp_service->srv_name, next);
1222 /* Add rpc to early reply check list */
1223 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
1225 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1226 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1227 struct ptlrpc_request *rq = NULL;
1233 if (req->rq_no_reply)
1236 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
1239 spin_lock(&svcpt->scp_at_lock);
1240 LASSERT(list_empty(&req->rq_timed_list));
1242 div_u64_rem(req->rq_deadline, array->paa_size, &index);
1243 if (array->paa_reqs_count[index] > 0) {
1244 /* latest rpcs will have the latest deadlines in the list,
1245 * so search backward. */
1246 list_for_each_entry_reverse(rq,
1247 &array->paa_reqs_array[index],
1249 if (req->rq_deadline >= rq->rq_deadline) {
1250 list_add(&req->rq_timed_list,
1251 &rq->rq_timed_list);
1257 /* Add the request at the head of the list */
1258 if (list_empty(&req->rq_timed_list))
1259 list_add(&req->rq_timed_list,
1260 &array->paa_reqs_array[index]);
1262 spin_lock(&req->rq_lock);
1263 req->rq_at_linked = 1;
1264 spin_unlock(&req->rq_lock);
1265 req->rq_at_index = index;
1266 array->paa_reqs_count[index]++;
1268 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
1269 array->paa_deadline = req->rq_deadline;
1270 ptlrpc_at_set_timer(svcpt);
1272 spin_unlock(&svcpt->scp_at_lock);
1278 ptlrpc_at_remove_timed(struct ptlrpc_request *req)
1280 struct ptlrpc_at_array *array;
1282 array = &req->rq_rqbd->rqbd_svcpt->scp_at_array;
1284 /* NB: must call with hold svcpt::scp_at_lock */
1285 LASSERT(!list_empty(&req->rq_timed_list));
1286 list_del_init(&req->rq_timed_list);
1288 spin_lock(&req->rq_lock);
1289 req->rq_at_linked = 0;
1290 spin_unlock(&req->rq_lock);
1292 array->paa_reqs_count[req->rq_at_index]--;
1297 * Attempt to extend the request deadline by sending an early reply to the
1300 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
1302 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1303 struct ptlrpc_request *reqcopy;
1304 struct lustre_msg *reqmsg;
1305 time64_t olddl = req->rq_deadline - ktime_get_real_seconds();
1311 if (CFS_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_RECONNECT)) {
1312 /* don't send early reply */
1316 /* deadline is when the client expects us to reply, margin is the
1317 difference between clients' and servers' expectations */
1318 DEBUG_REQ(D_ADAPTTO, req,
1319 "%ssending early reply (deadline %+llds, margin %+llds) for "
1320 "%d+%d", AT_OFF ? "AT off - not " : "",
1321 (s64)olddl, (s64)(olddl - at_get(&svcpt->scp_at_estimate)),
1322 at_get(&svcpt->scp_at_estimate), at_extra);
1328 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+llds), "
1329 "not sending early reply. Consider increasing "
1330 "at_early_margin (%d)?", (s64)olddl, at_early_margin);
1332 /* Return an error so we're not re-added to the timed list. */
1336 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
1337 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
1338 "but no AT support");
1342 if (req->rq_export &&
1343 lustre_msg_get_flags(req->rq_reqmsg) &
1344 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1345 struct obd_device *obd_exp = req->rq_export->exp_obd;
1347 /* During recovery, we don't want to send too many early
1348 * replies, but on the other hand we want to make sure the
1349 * client has enough time to resend if the rpc is lost. So
1350 * during the recovery period send at least 4 early replies,
1351 * spacing them every at_extra if we can. at_estimate should
1352 * always equal this fixed value during recovery.
1354 /* Don't account request processing time into AT history
1355 * during recovery, it is not service time we need but
1356 * includes also waiting time for recovering clients
1358 newdl = min_t(time64_t, at_extra,
1359 obd_exp->obd_recovery_timeout / 4) +
1360 ktime_get_real_seconds();
1362 /* We want to extend the request deadline by at_extra seconds,
1363 * so we set our service estimate to reflect how much time has
1364 * passed since this request arrived plus an additional
1365 * at_extra seconds. The client will calculate the new deadline
1366 * based on this service estimate (plus some additional time to
1367 * account for network latency). See ptlrpc_at_recv_early_reply
1369 at_measured(&svcpt->scp_at_estimate, at_extra +
1370 ktime_get_real_seconds() -
1371 req->rq_arrival_time.tv_sec);
1372 newdl = req->rq_arrival_time.tv_sec +
1373 at_get(&svcpt->scp_at_estimate);
1376 /* Check to see if we've actually increased the deadline -
1377 * we may be past adaptive_max */
1378 if (req->rq_deadline >= newdl) {
1379 DEBUG_REQ(D_WARNING, req, "Couldn't add any time (%lld/%lld), not sending early reply\n",
1380 (s64)olddl, (s64)(newdl - ktime_get_real_seconds()));
1384 reqcopy = ptlrpc_request_cache_alloc(GFP_NOFS);
1385 if (reqcopy == NULL)
1387 OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1389 GOTO(out_free, rc = -ENOMEM);
1392 reqcopy->rq_reply_state = NULL;
1393 reqcopy->rq_rep_swab_mask = 0;
1394 reqcopy->rq_pack_bulk = 0;
1395 reqcopy->rq_pack_udesc = 0;
1396 reqcopy->rq_packed_final = 0;
1397 sptlrpc_svc_ctx_addref(reqcopy);
1398 /* We only need the reqmsg for the magic */
1399 reqcopy->rq_reqmsg = reqmsg;
1400 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1403 * tgt_brw_read() and tgt_brw_write() may have decided not to reply.
1404 * Without this check, we would fail the rq_no_reply assertion in
1405 * ptlrpc_send_reply().
1407 if (reqcopy->rq_no_reply)
1408 GOTO(out, rc = -ETIMEDOUT);
1410 LASSERT(atomic_read(&req->rq_refcount));
1411 /** if it is last refcount then early reply isn't needed */
1412 if (atomic_read(&req->rq_refcount) == 1) {
1413 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
1414 "abort sending early reply\n");
1415 GOTO(out, rc = -EINVAL);
1418 /* Connection ref */
1419 reqcopy->rq_export = class_conn2export(
1420 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1421 if (reqcopy->rq_export == NULL)
1422 GOTO(out, rc = -ENODEV);
1425 class_export_rpc_inc(reqcopy->rq_export);
1426 if (reqcopy->rq_export->exp_obd &&
1427 reqcopy->rq_export->exp_obd->obd_fail)
1428 GOTO(out_put, rc = -ENODEV);
1430 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1434 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1437 /* Adjust our own deadline to what we told the client */
1438 req->rq_deadline = newdl;
1439 req->rq_early_count++; /* number sent, server side */
1441 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1444 /* Free the (early) reply state from lustre_pack_reply.
1445 (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
1446 ptlrpc_req_drop_rs(reqcopy);
1449 class_export_rpc_dec(reqcopy->rq_export);
1450 class_export_put(reqcopy->rq_export);
1452 sptlrpc_svc_ctx_decref(reqcopy);
1453 OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1455 ptlrpc_request_cache_free(reqcopy);
1459 /* Send early replies to everybody expiring within at_early_margin
1460 asking for at_extra time */
1461 static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1463 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1464 struct ptlrpc_request *rq, *n;
1465 struct list_head work_list;
1468 time64_t now = ktime_get_real_seconds();
1470 int first, counter = 0;
1473 spin_lock(&svcpt->scp_at_lock);
1474 if (svcpt->scp_at_check == 0) {
1475 spin_unlock(&svcpt->scp_at_lock);
1478 delay = ktime_ms_delta(ktime_get(), svcpt->scp_at_checktime);
1479 svcpt->scp_at_check = 0;
1481 if (array->paa_count == 0) {
1482 spin_unlock(&svcpt->scp_at_lock);
1486 /* The timer went off, but maybe the nearest rpc already completed. */
1487 first = array->paa_deadline - now;
1488 if (first > at_early_margin) {
1489 /* We've still got plenty of time. Reset the timer. */
1490 ptlrpc_at_set_timer(svcpt);
1491 spin_unlock(&svcpt->scp_at_lock);
1495 /* We're close to a timeout, and we don't know how much longer the
1496 server will take. Send early replies to everyone expiring soon. */
1497 INIT_LIST_HEAD(&work_list);
1499 div_u64_rem(array->paa_deadline, array->paa_size, &index);
1500 count = array->paa_count;
1502 count -= array->paa_reqs_count[index];
1503 list_for_each_entry_safe(rq, n,
1504 &array->paa_reqs_array[index],
1506 if (rq->rq_deadline > now + at_early_margin) {
1507 /* update the earliest deadline */
1508 if (deadline == -1 ||
1509 rq->rq_deadline < deadline)
1510 deadline = rq->rq_deadline;
1515 * ptlrpc_server_drop_request() may drop
1516 * refcount to 0 already. Let's check this and
1517 * don't add entry to work_list
1519 if (likely(atomic_inc_not_zero(&rq->rq_refcount))) {
1520 ptlrpc_at_remove_timed(rq);
1521 list_add(&rq->rq_timed_list, &work_list);
1523 ptlrpc_at_remove_timed(rq);
1529 if (++index >= array->paa_size)
1532 array->paa_deadline = deadline;
1533 /* we have a new earliest deadline, restart the timer */
1534 ptlrpc_at_set_timer(svcpt);
1536 spin_unlock(&svcpt->scp_at_lock);
1538 CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
1539 "replies\n", first, at_extra, counter);
1541 /* We're already past request deadlines before we even get a
1542 chance to send early replies */
1543 LCONSOLE_WARN("%s: This server is not able to keep up with "
1544 "request traffic (cpu-bound).\n",
1545 svcpt->scp_service->srv_name);
1546 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, delay=%lld\n",
1547 counter, svcpt->scp_nreqs_incoming,
1548 svcpt->scp_nreqs_active,
1549 at_get(&svcpt->scp_at_estimate), delay);
1552 /* we took additional refcount so entries can't be deleted from list, no
1553 * locking is needed */
1554 while (!list_empty(&work_list)) {
1555 rq = list_entry(work_list.next, struct ptlrpc_request,
1557 list_del_init(&rq->rq_timed_list);
1559 if (ptlrpc_at_send_early_reply(rq) == 0)
1560 ptlrpc_at_add_timed(rq);
1562 ptlrpc_server_drop_request(rq);
1565 RETURN(1); /* return "did_something" for liblustre */
1568 /* Check if we are already handling earlier incarnation of this request.
1569 * Called under &req->rq_export->exp_rpc_lock locked */
1570 static struct ptlrpc_request*
1571 ptlrpc_server_check_resend_in_progress(struct ptlrpc_request *req)
1573 struct ptlrpc_request *tmp = NULL;
1575 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ||
1576 (atomic_read(&req->rq_export->exp_rpc_count) == 0))
1579 /* bulk request are aborted upon reconnect, don't try to
1581 if (req->rq_bulk_write || req->rq_bulk_read)
1584 /* This list should not be longer than max_requests in
1585 * flights on the client, so it is not all that long.
1586 * Also we only hit this codepath in case of a resent
1587 * request which makes it even more rarely hit */
1588 list_for_each_entry(tmp, &req->rq_export->exp_reg_rpcs,
1590 /* Found duplicate one */
1591 if (tmp->rq_xid == req->rq_xid)
1594 list_for_each_entry(tmp, &req->rq_export->exp_hp_rpcs,
1596 /* Found duplicate one */
1597 if (tmp->rq_xid == req->rq_xid)
1603 DEBUG_REQ(D_HA, req, "Found duplicate req in processing");
1604 DEBUG_REQ(D_HA, tmp, "Request being processed");
1609 * Check if a request should be assigned with a high priority.
1611 * \retval < 0: error occurred
1612 * 0: normal RPC request
1613 * +1: high priority request
1615 static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt,
1616 struct ptlrpc_request *req)
1621 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL) {
1622 rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
1629 if (req->rq_export != NULL && req->rq_ops != NULL) {
1630 /* Perform request specific check. We should do this
1631 * check before the request is added into exp_hp_rpcs
1632 * list otherwise it may hit swab race at LU-1044. */
1633 if (req->rq_ops->hpreq_check != NULL) {
1634 rc = req->rq_ops->hpreq_check(req);
1635 if (rc == -ESTALE) {
1636 req->rq_status = rc;
1639 /** can only return error,
1640 * 0 for normal request,
1641 * or 1 for high priority request */
1649 /** Remove the request from the export list. */
1650 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req)
1653 if (req->rq_export) {
1654 /* refresh lock timeout again so that client has more
1655 * room to send lock cancel RPC. */
1656 if (req->rq_ops && req->rq_ops->hpreq_fini)
1657 req->rq_ops->hpreq_fini(req);
1659 spin_lock(&req->rq_export->exp_rpc_lock);
1660 list_del_init(&req->rq_exp_list);
1661 spin_unlock(&req->rq_export->exp_rpc_lock);
1666 static int ptlrpc_hpreq_check(struct ptlrpc_request *req)
1671 static struct ptlrpc_hpreq_ops ptlrpc_hpreq_common = {
1672 .hpreq_check = ptlrpc_hpreq_check,
1675 /* Hi-Priority RPC check by RPC operation code. */
1676 int ptlrpc_hpreq_handler(struct ptlrpc_request *req)
1678 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1680 /* Check for export to let only reconnects for not yet evicted
1681 * export to become a HP rpc. */
1682 if ((req->rq_export != NULL) &&
1683 (opc == OBD_PING || opc == MDS_CONNECT || opc == OST_CONNECT))
1684 req->rq_ops = &ptlrpc_hpreq_common;
1688 EXPORT_SYMBOL(ptlrpc_hpreq_handler);
1690 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1691 struct ptlrpc_request *req)
1695 struct ptlrpc_request *orig;
1698 rc = ptlrpc_server_hpreq_init(svcpt, req);
1703 ptlrpc_nrs_req_initialize(svcpt, req, hp);
1705 if (req->rq_export != NULL) {
1706 struct obd_export *exp = req->rq_export;
1708 /* do search for duplicated xid and the adding to the list
1710 spin_lock_bh(&exp->exp_rpc_lock);
1711 orig = ptlrpc_server_check_resend_in_progress(req);
1712 if (orig && likely(atomic_inc_not_zero(&orig->rq_refcount))) {
1715 spin_unlock_bh(&exp->exp_rpc_lock);
1718 * When the client resend request and the server has
1719 * the previous copy of it, we need to update deadlines,
1720 * to be sure that the client and the server have equal
1721 * request deadlines.
1724 spin_lock(&orig->rq_rqbd->rqbd_svcpt->scp_at_lock);
1725 linked = orig->rq_at_linked;
1727 ptlrpc_at_remove_timed(orig);
1728 spin_unlock(&orig->rq_rqbd->rqbd_svcpt->scp_at_lock);
1729 orig->rq_deadline = req->rq_deadline;
1731 ptlrpc_at_add_timed(orig);
1732 ptlrpc_server_drop_request(orig);
1733 ptlrpc_nrs_req_finalize(req);
1737 if (hp || req->rq_ops != NULL)
1738 list_add(&req->rq_exp_list, &exp->exp_hp_rpcs);
1740 list_add(&req->rq_exp_list, &exp->exp_reg_rpcs);
1741 spin_unlock_bh(&exp->exp_rpc_lock);
1744 /* the current thread is not the processing thread for this request
1745 * since that, but request is in exp_hp_list and can be find there.
1746 * Remove all relations between request and old thread. */
1747 req->rq_svc_thread->t_env->le_ses = NULL;
1748 req->rq_svc_thread = NULL;
1749 req->rq_session.lc_thread = NULL;
1751 ptlrpc_nrs_req_add(svcpt, req, hp);
1757 * Allow to handle high priority request
1758 * User can call it w/o any lock but need to hold
1759 * ptlrpc_service_part::scp_req_lock to get reliable result
1761 static bool ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1764 int running = svcpt->scp_nthrs_running;
1766 if (!nrs_svcpt_has_hp(svcpt))
1772 if (ptlrpc_nrs_req_throttling_nolock(svcpt, true))
1775 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1776 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1777 /* leave just 1 thread for normal RPCs */
1778 running = PTLRPC_NTHRS_INIT;
1779 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1783 if (svcpt->scp_nreqs_active >= running - 1)
1786 if (svcpt->scp_nhreqs_active == 0)
1789 return !ptlrpc_nrs_req_pending_nolock(svcpt, false) ||
1790 svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1793 static bool ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1796 return ptlrpc_server_allow_high(svcpt, force) &&
1797 ptlrpc_nrs_req_pending_nolock(svcpt, true);
1801 * Only allow normal priority requests on a service that has a high-priority
1802 * queue if forced (i.e. cleanup), if there are other high priority requests
1803 * already being processed (i.e. those threads can service more high-priority
1804 * requests), or if there are enough idle threads that a later thread can do
1805 * a high priority request.
1806 * User can call it w/o any lock but need to hold
1807 * ptlrpc_service_part::scp_req_lock to get reliable result
1809 static bool ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1812 int running = svcpt->scp_nthrs_running;
1813 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1814 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1815 /* leave just 1 thread for normal RPCs */
1816 running = PTLRPC_NTHRS_INIT;
1817 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1824 if (ptlrpc_nrs_req_throttling_nolock(svcpt, false))
1827 if (svcpt->scp_nreqs_active < running - 2)
1830 if (svcpt->scp_nreqs_active >= running - 1)
1833 return svcpt->scp_nhreqs_active > 0 || !nrs_svcpt_has_hp(svcpt);
1836 static bool ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1839 return ptlrpc_server_allow_normal(svcpt, force) &&
1840 ptlrpc_nrs_req_pending_nolock(svcpt, false);
1844 * Returns true if there are requests available in incoming
1845 * request queue for processing and it is allowed to fetch them.
1846 * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1847 * to get reliable result
1848 * \see ptlrpc_server_allow_normal
1849 * \see ptlrpc_server_allow high
1852 ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt, bool force)
1854 return ptlrpc_server_high_pending(svcpt, force) ||
1855 ptlrpc_server_normal_pending(svcpt, force);
1859 * Fetch a request for processing from queue of unprocessed requests.
1860 * Favors high-priority requests.
1861 * Returns a pointer to fetched request.
1863 static struct ptlrpc_request *
1864 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force)
1866 struct ptlrpc_request *req = NULL;
1869 spin_lock(&svcpt->scp_req_lock);
1871 if (ptlrpc_server_high_pending(svcpt, force)) {
1872 req = ptlrpc_nrs_req_get_nolock(svcpt, true, force);
1874 svcpt->scp_hreq_count++;
1879 if (ptlrpc_server_normal_pending(svcpt, force)) {
1880 req = ptlrpc_nrs_req_get_nolock(svcpt, false, force);
1882 svcpt->scp_hreq_count = 0;
1887 spin_unlock(&svcpt->scp_req_lock);
1891 svcpt->scp_nreqs_active++;
1893 svcpt->scp_nhreqs_active++;
1895 spin_unlock(&svcpt->scp_req_lock);
1897 if (likely(req->rq_export))
1898 class_export_rpc_inc(req->rq_export);
1904 * Handle freshly incoming reqs, add to timed early reply list,
1905 * pass on to regular request queue.
1906 * All incoming requests pass through here before getting into
1907 * ptlrpc_server_handle_req later on.
1910 ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt,
1911 struct ptlrpc_thread *thread)
1913 struct ptlrpc_service *svc = svcpt->scp_service;
1914 struct ptlrpc_request *req;
1919 spin_lock(&svcpt->scp_lock);
1920 if (list_empty(&svcpt->scp_req_incoming)) {
1921 spin_unlock(&svcpt->scp_lock);
1925 req = list_entry(svcpt->scp_req_incoming.next,
1926 struct ptlrpc_request, rq_list);
1927 list_del_init(&req->rq_list);
1928 svcpt->scp_nreqs_incoming--;
1929 /* Consider this still a "queued" request as far as stats are
1931 spin_unlock(&svcpt->scp_lock);
1933 /* go through security check/transform */
1934 rc = sptlrpc_svc_unwrap_request(req);
1938 case SECSVC_COMPLETE:
1939 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1948 * for null-flavored rpc, msg has been unpacked by sptlrpc, although
1949 * redo it wouldn't be harmful.
1951 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
1952 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
1954 CERROR("error unpacking request: ptl %d from %s "
1955 "x%llu\n", svc->srv_req_portal,
1956 libcfs_id2str(req->rq_peer), req->rq_xid);
1961 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1963 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
1964 "%llu\n", svc->srv_req_portal,
1965 libcfs_id2str(req->rq_peer), req->rq_xid);
1969 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
1970 lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
1971 CERROR("drop incoming rpc opc %u, x%llu\n",
1972 cfs_fail_val, req->rq_xid);
1977 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1978 CERROR("wrong packet type received (type=%u) from %s\n",
1979 lustre_msg_get_type(req->rq_reqmsg),
1980 libcfs_id2str(req->rq_peer));
1984 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1988 req->rq_bulk_write = 1;
1992 case MGS_CONFIG_READ:
1993 req->rq_bulk_read = 1;
1997 CDEBUG(D_RPCTRACE, "got req x%llu\n", req->rq_xid);
1999 req->rq_export = class_conn2export(
2000 lustre_msg_get_handle(req->rq_reqmsg));
2001 if (req->rq_export) {
2002 rc = ptlrpc_check_req(req);
2004 rc = sptlrpc_target_export_check(req->rq_export, req);
2006 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
2007 "illegal security flavor,");
2012 ptlrpc_update_export_timer(req->rq_export, 0);
2015 /* req_in handling should/must be fast */
2016 if (ktime_get_real_seconds() - req->rq_arrival_time.tv_sec > 5)
2017 DEBUG_REQ(D_WARNING, req, "Slow req_in handling %llds",
2018 (s64)(ktime_get_real_seconds() -
2019 req->rq_arrival_time.tv_sec));
2021 /* Set rpc server deadline and add it to the timed list */
2022 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
2023 MSGHDR_AT_SUPPORT) ?
2024 /* The max time the client expects us to take */
2025 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
2027 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
2028 if (unlikely(deadline == 0)) {
2029 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
2033 /* Skip early reply */
2034 if (OBD_FAIL_PRECHECK(OBD_FAIL_MDS_RESEND))
2035 req->rq_deadline += obd_timeout;
2037 req->rq_svc_thread = thread;
2038 if (thread != NULL) {
2039 /* initialize request session, it is needed for request
2040 * processing by target */
2041 rc = lu_context_init(&req->rq_session, LCT_SERVER_SESSION |
2044 CERROR("%s: failure to initialize session: rc = %d\n",
2045 thread->t_name, rc);
2048 req->rq_session.lc_thread = thread;
2049 lu_context_enter(&req->rq_session);
2050 thread->t_env->le_ses = &req->rq_session;
2053 ptlrpc_at_add_timed(req);
2055 /* Move it over to the request processing queue */
2056 rc = ptlrpc_server_request_add(svcpt, req);
2060 wake_up(&svcpt->scp_waitq);
2064 ptlrpc_server_finish_request(svcpt, req);
2070 * Main incoming request handling logic.
2071 * Calls handler function from service to do actual processing.
2074 ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
2075 struct ptlrpc_thread *thread)
2077 struct ptlrpc_service *svc = svcpt->scp_service;
2078 struct ptlrpc_request *request;
2088 request = ptlrpc_server_request_get(svcpt, false);
2089 if (request == NULL)
2092 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
2093 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
2094 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
2095 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
2097 if (unlikely(fail_opc)) {
2098 if (request->rq_export && request->rq_ops)
2099 OBD_FAIL_TIMEOUT(fail_opc, 4);
2102 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
2104 if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
2105 libcfs_debug_dumplog();
2107 work_start = ktime_get_real();
2108 arrived = timespec64_to_ktime(request->rq_arrival_time);
2109 timediff_usecs = ktime_us_delta(work_start, arrived);
2110 if (likely(svc->srv_stats != NULL)) {
2111 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
2113 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
2114 svcpt->scp_nreqs_incoming);
2115 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
2116 svcpt->scp_nreqs_active);
2117 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
2118 at_get(&svcpt->scp_at_estimate));
2121 if (likely(request->rq_export)) {
2122 if (unlikely(ptlrpc_check_req(request)))
2124 ptlrpc_update_export_timer(request->rq_export,
2125 div_u64(timediff_usecs,
2129 /* Discard requests queued for longer than the deadline.
2130 The deadline is increased if we send an early reply. */
2131 if (ktime_get_real_seconds() > request->rq_deadline) {
2132 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s: deadline %lld:%llds ago\n",
2133 libcfs_id2str(request->rq_peer),
2134 request->rq_deadline -
2135 request->rq_arrival_time.tv_sec,
2136 ktime_get_real_seconds() - request->rq_deadline);
2140 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
2141 "%s:%s+%d:%d:x%llu:%s:%d\n", current_comm(),
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), request->rq_xid,
2147 libcfs_id2str(request->rq_peer),
2148 lustre_msg_get_opc(request->rq_reqmsg));
2150 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
2151 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
2153 CDEBUG(D_NET, "got req %llu\n", request->rq_xid);
2155 /* re-assign request and sesson thread to the current one */
2156 request->rq_svc_thread = thread;
2157 if (thread != NULL) {
2158 LASSERT(request->rq_session.lc_thread == NULL);
2159 request->rq_session.lc_thread = thread;
2160 thread->t_env->le_ses = &request->rq_session;
2162 svc->srv_ops.so_req_handler(request);
2164 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
2167 if (unlikely(ktime_get_real_seconds() > request->rq_deadline)) {
2168 DEBUG_REQ(D_WARNING, request,
2169 "Request took longer than estimated (%lld:%llds); "
2170 "client may timeout.",
2171 request->rq_deadline -
2172 request->rq_arrival_time.tv_sec,
2173 ktime_get_real_seconds() - request->rq_deadline);
2176 work_end = ktime_get_real();
2177 timediff_usecs = ktime_us_delta(work_end, work_start);
2178 arrived_usecs = ktime_us_delta(work_end, arrived);
2179 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
2180 "%s:%s+%d:%d:x%llu:%s:%d Request processed in %lldus "
2181 "(%lldus total) trans %llu rc %d/%d\n",
2183 (request->rq_export ?
2184 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2185 (request->rq_export ?
2186 atomic_read(&request->rq_export->exp_refcount) : -99),
2187 lustre_msg_get_status(request->rq_reqmsg),
2189 libcfs_id2str(request->rq_peer),
2190 lustre_msg_get_opc(request->rq_reqmsg),
2193 (request->rq_repmsg ?
2194 lustre_msg_get_transno(request->rq_repmsg) :
2195 request->rq_transno),
2197 (request->rq_repmsg ?
2198 lustre_msg_get_status(request->rq_repmsg) : -999));
2199 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
2200 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
2201 int opc = opcode_offset(op);
2202 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
2203 LASSERT(opc < LUSTRE_MAX_OPCODES);
2204 lprocfs_counter_add(svc->srv_stats,
2205 opc + EXTRA_MAX_OPCODES,
2209 if (unlikely(request->rq_early_count)) {
2210 DEBUG_REQ(D_ADAPTTO, request,
2211 "sent %d early replies before finishing in %llds",
2212 request->rq_early_count,
2213 div_u64(arrived_usecs, USEC_PER_SEC));
2216 ptlrpc_server_finish_active_request(svcpt, request);
2222 * An internal function to process a single reply state object.
2225 ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
2227 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
2228 struct ptlrpc_service *svc = svcpt->scp_service;
2229 struct obd_export *exp;
2234 exp = rs->rs_export;
2236 LASSERT(rs->rs_difficult);
2237 LASSERT(rs->rs_scheduled);
2238 LASSERT(list_empty(&rs->rs_list));
2240 /* The disk commit callback holds exp_uncommitted_replies_lock while it
2241 * iterates over newly committed replies, removing them from
2242 * exp_uncommitted_replies. It then drops this lock and schedules the
2243 * replies it found for handling here.
2245 * We can avoid contention for exp_uncommitted_replies_lock between the
2246 * HRT threads and further commit callbacks by checking rs_committed
2247 * which is set in the commit callback while it holds both
2248 * rs_lock and exp_uncommitted_reples.
2250 * If we see rs_committed clear, the commit callback _may_ not have
2251 * handled this reply yet and we race with it to grab
2252 * exp_uncommitted_replies_lock before removing the reply from
2253 * exp_uncommitted_replies. Note that if we lose the race and the
2254 * reply has already been removed, list_del_init() is a noop.
2256 * If we see rs_committed set, we know the commit callback is handling,
2257 * or has handled this reply since store reordering might allow us to
2258 * see rs_committed set out of sequence. But since this is done
2259 * holding rs_lock, we can be sure it has all completed once we hold
2260 * rs_lock, which we do right next.
2262 if (!rs->rs_committed) {
2263 /* if rs was commited, no need to convert locks, don't check
2264 * rs_committed here because rs may never be added into
2265 * exp_uncommitted_replies and this flag never be set, see
2266 * target_send_reply() */
2267 if (rs->rs_convert_lock &&
2268 rs->rs_transno > exp->exp_last_committed) {
2269 struct ldlm_lock *lock;
2270 struct ldlm_lock *ack_locks[RS_MAX_LOCKS] = { NULL };
2272 spin_lock(&rs->rs_lock);
2273 if (rs->rs_convert_lock &&
2274 rs->rs_transno > exp->exp_last_committed) {
2275 nlocks = rs->rs_nlocks;
2276 while (nlocks-- > 0) {
2278 * NB don't assume rs is always handled
2279 * by the same service thread (see
2280 * ptlrpc_hr_select, so REP-ACK hr may
2281 * race with trans commit, while the
2282 * latter will release locks, get locks
2283 * here early to convert to COS mode
2286 lock = ldlm_handle2lock(
2287 &rs->rs_locks[nlocks]);
2289 ack_locks[nlocks] = lock;
2290 rs->rs_modes[nlocks] = LCK_COS;
2292 nlocks = rs->rs_nlocks;
2293 rs->rs_convert_lock = 0;
2294 /* clear rs_scheduled so that commit callback
2295 * can schedule again */
2296 rs->rs_scheduled = 0;
2297 spin_unlock(&rs->rs_lock);
2299 while (nlocks-- > 0) {
2300 lock = ack_locks[nlocks];
2301 ldlm_lock_mode_downgrade(lock, LCK_COS);
2302 LDLM_LOCK_PUT(lock);
2306 spin_unlock(&rs->rs_lock);
2309 spin_lock(&exp->exp_uncommitted_replies_lock);
2310 list_del_init(&rs->rs_obd_list);
2311 spin_unlock(&exp->exp_uncommitted_replies_lock);
2314 spin_lock(&exp->exp_lock);
2315 /* Noop if removed already */
2316 list_del_init(&rs->rs_exp_list);
2317 spin_unlock(&exp->exp_lock);
2319 spin_lock(&rs->rs_lock);
2321 been_handled = rs->rs_handled;
2324 nlocks = rs->rs_nlocks; /* atomic "steal", but */
2325 rs->rs_nlocks = 0; /* locks still on rs_locks! */
2327 if (nlocks == 0 && !been_handled) {
2328 /* If we see this, we should already have seen the warning
2329 * in mds_steal_ack_locks() */
2330 CDEBUG(D_HA, "All locks stolen from rs %p x%lld.t%lld"
2333 rs->rs_xid, rs->rs_transno, rs->rs_opc,
2334 libcfs_nid2str(exp->exp_connection->c_peer.nid));
2337 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
2338 spin_unlock(&rs->rs_lock);
2340 if (!been_handled && rs->rs_on_net) {
2341 LNetMDUnlink(rs->rs_md_h);
2342 /* Ignore return code; we're racing with completion */
2345 while (nlocks-- > 0)
2346 ldlm_lock_decref(&rs->rs_locks[nlocks],
2347 rs->rs_modes[nlocks]);
2349 spin_lock(&rs->rs_lock);
2352 rs->rs_scheduled = 0;
2353 rs->rs_convert_lock = 0;
2355 if (!rs->rs_on_net) {
2357 spin_unlock(&rs->rs_lock);
2359 class_export_put (exp);
2360 rs->rs_export = NULL;
2361 ptlrpc_rs_decref(rs);
2362 if (atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
2363 svc->srv_is_stopping)
2364 wake_up_all(&svcpt->scp_waitq);
2368 /* still on the net; callback will schedule */
2369 spin_unlock(&rs->rs_lock);
2375 ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
2377 int avail = svcpt->scp_nrqbds_posted;
2378 int low_water = test_req_buffer_pressure ? 0 :
2379 svcpt->scp_service->srv_nbuf_per_group / 2;
2381 /* NB I'm not locking; just looking. */
2383 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
2384 * allowed the request history to grow out of control. We could put a
2385 * sanity check on that here and cull some history if we need the
2388 if (avail <= low_water)
2389 ptlrpc_grow_req_bufs(svcpt, 1);
2391 if (svcpt->scp_service->srv_stats) {
2392 lprocfs_counter_add(svcpt->scp_service->srv_stats,
2393 PTLRPC_REQBUF_AVAIL_CNTR, avail);
2398 ptlrpc_retry_rqbds(void *arg)
2400 struct ptlrpc_service_part *svcpt = (struct ptlrpc_service_part *)arg;
2402 svcpt->scp_rqbd_timeout = 0;
2407 ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
2409 return svcpt->scp_nreqs_active <
2410 svcpt->scp_nthrs_running - 1 -
2411 (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL);
2415 * allowed to create more threads
2416 * user can call it w/o any lock but need to hold
2417 * ptlrpc_service_part::scp_lock to get reliable result
2420 ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
2422 return svcpt->scp_nthrs_running +
2423 svcpt->scp_nthrs_starting <
2424 svcpt->scp_service->srv_nthrs_cpt_limit;
2428 * too many requests and allowed to create more threads
2431 ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
2433 return !ptlrpc_threads_enough(svcpt) &&
2434 ptlrpc_threads_increasable(svcpt);
2438 ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2440 return thread_is_stopping(thread) ||
2441 thread->t_svcpt->scp_service->srv_is_stopping;
2445 ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
2447 return !list_empty(&svcpt->scp_rqbd_idle) &&
2448 svcpt->scp_rqbd_timeout == 0;
2452 ptlrpc_at_check(struct ptlrpc_service_part *svcpt)
2454 return svcpt->scp_at_check;
2458 * If a thread runs too long or spends to much time on a single request,
2459 * we want to know about it, so we set up a delayed work item as a watchdog.
2460 * If it fires, we display a stack trace of the delayed thread,
2461 * providing we aren't rate-limited
2463 * Watchdog stack traces are limited to 3 per 'libcfs_watchdog_ratelimit'
2466 static struct ratelimit_state watchdog_limit;
2468 static void ptlrpc_watchdog_fire(struct work_struct *w)
2470 struct ptlrpc_thread *thread = container_of(w, struct ptlrpc_thread,
2472 u64 ms_lapse = ktime_ms_delta(ktime_get(), thread->t_touched);
2473 u32 ms_frac = do_div(ms_lapse, MSEC_PER_SEC);
2475 if (!__ratelimit(&watchdog_limit)) {
2476 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",
2477 thread->t_task->comm, thread->t_task->pid,
2480 libcfs_debug_dumpstack(thread->t_task);
2482 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",
2483 thread->t_task->comm, thread->t_task->pid,
2484 ms_lapse, ms_frac, libcfs_watchdog_ratelimit);
2488 static void ptlrpc_watchdog_init(struct delayed_work *work, time_t time)
2490 INIT_DELAYED_WORK(work, ptlrpc_watchdog_fire);
2491 schedule_delayed_work(work, cfs_time_seconds(time));
2494 static void ptlrpc_watchdog_disable(struct delayed_work *work)
2496 cancel_delayed_work_sync(work);
2499 static void ptlrpc_watchdog_touch(struct delayed_work *work, time_t time)
2501 struct ptlrpc_thread *thread = container_of(&work->work,
2502 struct ptlrpc_thread,
2504 thread->t_touched = ktime_get();
2505 mod_delayed_work(system_wq, work, cfs_time_seconds(time));
2509 * requests wait on preprocessing
2510 * user can call it w/o any lock but need to hold
2511 * ptlrpc_service_part::scp_lock to get reliable result
2514 ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
2516 return !list_empty(&svcpt->scp_req_incoming);
2519 static __attribute__((__noinline__)) int
2520 ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
2521 struct ptlrpc_thread *thread)
2523 /* Don't exit while there are replies to be handled */
2524 struct l_wait_info lwi = LWI_TIMEOUT(svcpt->scp_rqbd_timeout,
2525 ptlrpc_retry_rqbds, svcpt);
2527 ptlrpc_watchdog_disable(&thread->t_watchdog);
2531 l_wait_event_exclusive_head(svcpt->scp_waitq,
2532 ptlrpc_thread_stopping(thread) ||
2533 ptlrpc_server_request_incoming(svcpt) ||
2534 ptlrpc_server_request_pending(svcpt, false) ||
2535 ptlrpc_rqbd_pending(svcpt) ||
2536 ptlrpc_at_check(svcpt), &lwi);
2538 if (ptlrpc_thread_stopping(thread))
2541 ptlrpc_watchdog_touch(&thread->t_watchdog,
2542 ptlrpc_server_get_timeout(svcpt));
2547 * Main thread body for service threads.
2548 * Waits in a loop waiting for new requests to process to appear.
2549 * Every time an incoming requests is added to its queue, a waitq
2550 * is woken up and one of the threads will handle it.
2552 static int ptlrpc_main(void *arg)
2554 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
2555 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2556 struct ptlrpc_service *svc = svcpt->scp_service;
2557 struct ptlrpc_reply_state *rs;
2558 struct group_info *ginfo = NULL;
2560 int counter = 0, rc = 0;
2563 thread->t_task = current;
2564 thread->t_pid = current_pid();
2565 unshare_fs_struct();
2567 if (svc->srv_cpt_bind) {
2568 rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
2570 CWARN("%s: failed to bind %s on CPT %d\n",
2571 svc->srv_name, thread->t_name, svcpt->scp_cpt);
2575 ginfo = groups_alloc(0);
2581 set_current_groups(ginfo);
2582 put_group_info(ginfo);
2584 if (svc->srv_ops.so_thr_init != NULL) {
2585 rc = svc->srv_ops.so_thr_init(thread);
2596 rc = lu_context_init(&env->le_ctx,
2597 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2601 thread->t_env = env;
2602 env->le_ctx.lc_thread = thread;
2603 env->le_ctx.lc_cookie = 0x6;
2605 while (!list_empty(&svcpt->scp_rqbd_idle)) {
2606 rc = ptlrpc_server_post_idle_rqbds(svcpt);
2610 CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
2611 svc->srv_name, svcpt->scp_cpt, rc);
2615 /* Alloc reply state structure for this one */
2616 OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2622 spin_lock(&svcpt->scp_lock);
2624 LASSERT(thread_is_starting(thread));
2625 thread_clear_flags(thread, SVC_STARTING);
2627 LASSERT(svcpt->scp_nthrs_starting == 1);
2628 svcpt->scp_nthrs_starting--;
2630 /* SVC_STOPPING may already be set here if someone else is trying
2631 * to stop the service while this new thread has been dynamically
2632 * forked. We still set SVC_RUNNING to let our creator know that
2633 * we are now running, however we will exit as soon as possible */
2634 thread_add_flags(thread, SVC_RUNNING);
2635 svcpt->scp_nthrs_running++;
2636 spin_unlock(&svcpt->scp_lock);
2638 /* wake up our creator in case he's still waiting. */
2639 wake_up(&thread->t_ctl_waitq);
2641 thread->t_touched = ktime_get();
2642 ptlrpc_watchdog_init(&thread->t_watchdog,
2643 ptlrpc_server_get_timeout(svcpt));
2645 spin_lock(&svcpt->scp_rep_lock);
2646 list_add(&rs->rs_list, &svcpt->scp_rep_idle);
2647 wake_up(&svcpt->scp_rep_waitq);
2648 spin_unlock(&svcpt->scp_rep_lock);
2650 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2651 svcpt->scp_nthrs_running);
2653 /* XXX maintain a list of all managed devices: insert here */
2654 while (!ptlrpc_thread_stopping(thread)) {
2655 if (ptlrpc_wait_event(svcpt, thread))
2658 ptlrpc_check_rqbd_pool(svcpt);
2660 if (ptlrpc_threads_need_create(svcpt)) {
2661 /* Ignore return code - we tried... */
2662 ptlrpc_start_thread(svcpt, 0);
2665 /* reset le_ses to initial state */
2667 /* Process all incoming reqs before handling any */
2668 if (ptlrpc_server_request_incoming(svcpt)) {
2669 lu_context_enter(&env->le_ctx);
2670 ptlrpc_server_handle_req_in(svcpt, thread);
2671 lu_context_exit(&env->le_ctx);
2673 /* but limit ourselves in case of flood */
2674 if (counter++ < 100)
2679 if (ptlrpc_at_check(svcpt))
2680 ptlrpc_at_check_timed(svcpt);
2682 if (ptlrpc_server_request_pending(svcpt, false)) {
2683 lu_context_enter(&env->le_ctx);
2684 ptlrpc_server_handle_request(svcpt, thread);
2685 lu_context_exit(&env->le_ctx);
2688 if (ptlrpc_rqbd_pending(svcpt) &&
2689 ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
2690 /* I just failed to repost request buffers.
2691 * Wait for a timeout (unless something else
2692 * happens) before I try again */
2693 svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
2694 CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
2695 svcpt->scp_nrqbds_posted);
2699 ptlrpc_watchdog_disable(&thread->t_watchdog);
2703 * deconstruct service specific state created by ptlrpc_start_thread()
2705 if (svc->srv_ops.so_thr_done != NULL)
2706 svc->srv_ops.so_thr_done(thread);
2709 lu_context_fini(&env->le_ctx);
2713 CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n",
2714 thread, thread->t_pid, thread->t_id, rc);
2716 spin_lock(&svcpt->scp_lock);
2717 if (thread_test_and_clear_flags(thread, SVC_STARTING))
2718 svcpt->scp_nthrs_starting--;
2720 if (thread_test_and_clear_flags(thread, SVC_RUNNING)) {
2721 /* must know immediately */
2722 svcpt->scp_nthrs_running--;
2726 thread_add_flags(thread, SVC_STOPPED);
2728 wake_up(&thread->t_ctl_waitq);
2729 spin_unlock(&svcpt->scp_lock);
2734 static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
2735 struct list_head *replies)
2739 spin_lock(&hrt->hrt_lock);
2741 list_splice_init(&hrt->hrt_queue, replies);
2742 result = ptlrpc_hr.hr_stopping || !list_empty(replies);
2744 spin_unlock(&hrt->hrt_lock);
2749 * Main body of "handle reply" function.
2750 * It processes acked reply states
2752 static int ptlrpc_hr_main(void *arg)
2754 struct ptlrpc_hr_thread *hrt = (struct ptlrpc_hr_thread *)arg;
2755 struct ptlrpc_hr_partition *hrp = hrt->hrt_partition;
2756 struct list_head replies;
2759 INIT_LIST_HEAD(&replies);
2760 unshare_fs_struct();
2762 rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
2764 char threadname[20];
2766 snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
2767 hrp->hrp_cpt, hrt->hrt_id);
2768 CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
2769 threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
2772 atomic_inc(&hrp->hrp_nstarted);
2773 wake_up(&ptlrpc_hr.hr_waitq);
2775 while (!ptlrpc_hr.hr_stopping) {
2776 l_wait_condition(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
2778 while (!list_empty(&replies)) {
2779 struct ptlrpc_reply_state *rs;
2781 rs = list_entry(replies.prev,
2782 struct ptlrpc_reply_state,
2784 list_del_init(&rs->rs_list);
2785 ptlrpc_handle_rs(rs);
2789 atomic_inc(&hrp->hrp_nstopped);
2790 wake_up(&ptlrpc_hr.hr_waitq);
2795 static void ptlrpc_stop_hr_threads(void)
2797 struct ptlrpc_hr_partition *hrp;
2801 ptlrpc_hr.hr_stopping = 1;
2803 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2804 if (hrp->hrp_thrs == NULL)
2805 continue; /* uninitialized */
2806 for (j = 0; j < hrp->hrp_nthrs; j++)
2807 wake_up_all(&hrp->hrp_thrs[j].hrt_waitq);
2810 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2811 if (hrp->hrp_thrs == NULL)
2812 continue; /* uninitialized */
2813 wait_event(ptlrpc_hr.hr_waitq,
2814 atomic_read(&hrp->hrp_nstopped) ==
2815 atomic_read(&hrp->hrp_nstarted));
2819 static int ptlrpc_start_hr_threads(void)
2821 struct ptlrpc_hr_partition *hrp;
2826 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2829 for (j = 0; j < hrp->hrp_nthrs; j++) {
2830 struct ptlrpc_hr_thread *hrt = &hrp->hrp_thrs[j];
2831 struct task_struct *task;
2833 task = kthread_run(ptlrpc_hr_main,
2835 "ptlrpc_hr%02d_%03d",
2844 wait_event(ptlrpc_hr.hr_waitq,
2845 atomic_read(&hrp->hrp_nstarted) == j);
2848 CERROR("cannot start reply handler thread %d:%d: "
2849 "rc = %d\n", i, j, rc);
2850 ptlrpc_stop_hr_threads();
2858 static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
2860 struct l_wait_info lwi = { 0 };
2861 struct ptlrpc_thread *thread;
2862 struct list_head zombie;
2866 CDEBUG(D_INFO, "Stopping threads for service %s\n",
2867 svcpt->scp_service->srv_name);
2869 INIT_LIST_HEAD(&zombie);
2870 spin_lock(&svcpt->scp_lock);
2871 /* let the thread know that we would like it to stop asap */
2872 list_for_each_entry(thread, &svcpt->scp_threads, t_link) {
2873 CDEBUG(D_INFO, "Stopping thread %s #%u\n",
2874 svcpt->scp_service->srv_thread_name, thread->t_id);
2875 thread_add_flags(thread, SVC_STOPPING);
2878 wake_up_all(&svcpt->scp_waitq);
2880 while (!list_empty(&svcpt->scp_threads)) {
2881 thread = list_entry(svcpt->scp_threads.next,
2882 struct ptlrpc_thread, t_link);
2883 if (thread_is_stopped(thread)) {
2884 list_del(&thread->t_link);
2885 list_add(&thread->t_link, &zombie);
2888 spin_unlock(&svcpt->scp_lock);
2890 CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
2891 svcpt->scp_service->srv_thread_name, thread->t_id);
2892 l_wait_event(thread->t_ctl_waitq,
2893 thread_is_stopped(thread), &lwi);
2895 spin_lock(&svcpt->scp_lock);
2898 spin_unlock(&svcpt->scp_lock);
2900 while (!list_empty(&zombie)) {
2901 thread = list_entry(zombie.next,
2902 struct ptlrpc_thread, t_link);
2903 list_del(&thread->t_link);
2904 OBD_FREE_PTR(thread);
2910 * Stops all threads of a particular service \a svc
2912 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2914 struct ptlrpc_service_part *svcpt;
2918 ptlrpc_service_for_each_part(svcpt, i, svc) {
2919 if (svcpt->scp_service != NULL)
2920 ptlrpc_svcpt_stop_threads(svcpt);
2926 int ptlrpc_start_threads(struct ptlrpc_service *svc)
2933 /* We require 2 threads min, see note in ptlrpc_server_handle_request */
2934 LASSERT(svc->srv_nthrs_cpt_init >= PTLRPC_NTHRS_INIT);
2936 for (i = 0; i < svc->srv_ncpts; i++) {
2937 for (j = 0; j < svc->srv_nthrs_cpt_init; j++) {
2938 rc = ptlrpc_start_thread(svc->srv_parts[i], 1);
2944 /* We have enough threads, don't start more. b=15759 */
2951 CERROR("cannot start %s thread #%d_%d: rc %d\n",
2952 svc->srv_thread_name, i, j, rc);
2953 ptlrpc_stop_all_threads(svc);
2957 int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
2959 struct l_wait_info lwi = { 0 };
2960 struct ptlrpc_thread *thread;
2961 struct ptlrpc_service *svc;
2962 struct task_struct *task;
2966 LASSERT(svcpt != NULL);
2968 svc = svcpt->scp_service;
2970 CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n",
2971 svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running,
2972 svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit);
2975 if (unlikely(svc->srv_is_stopping))
2978 if (!ptlrpc_threads_increasable(svcpt) ||
2979 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
2980 svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1))
2983 OBD_CPT_ALLOC_PTR(thread, svc->srv_cptable, svcpt->scp_cpt);
2986 init_waitqueue_head(&thread->t_ctl_waitq);
2988 spin_lock(&svcpt->scp_lock);
2989 if (!ptlrpc_threads_increasable(svcpt)) {
2990 spin_unlock(&svcpt->scp_lock);
2991 OBD_FREE_PTR(thread);
2995 if (svcpt->scp_nthrs_starting != 0) {
2996 /* serialize starting because some modules (obdfilter)
2997 * might require unique and contiguous t_id */
2998 LASSERT(svcpt->scp_nthrs_starting == 1);
2999 spin_unlock(&svcpt->scp_lock);
3000 OBD_FREE_PTR(thread);
3002 CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n",
3003 svc->srv_thread_name, svcpt->scp_thr_nextid);
3008 CDEBUG(D_INFO, "Creating thread %s #%d race, retry later\n",
3009 svc->srv_thread_name, svcpt->scp_thr_nextid);
3013 svcpt->scp_nthrs_starting++;
3014 thread->t_id = svcpt->scp_thr_nextid++;
3015 thread_add_flags(thread, SVC_STARTING);
3016 thread->t_svcpt = svcpt;
3018 list_add(&thread->t_link, &svcpt->scp_threads);
3019 spin_unlock(&svcpt->scp_lock);
3021 if (svcpt->scp_cpt >= 0) {
3022 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s%02d_%03d",
3023 svc->srv_thread_name, svcpt->scp_cpt, thread->t_id);
3025 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s_%04d",
3026 svc->srv_thread_name, thread->t_id);
3029 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name);
3030 task = kthread_run(ptlrpc_main, thread, "%s", thread->t_name);
3033 CERROR("cannot start thread '%s': rc = %d\n",
3034 thread->t_name, rc);
3035 spin_lock(&svcpt->scp_lock);
3036 --svcpt->scp_nthrs_starting;
3037 if (thread_is_stopping(thread)) {
3038 /* this ptlrpc_thread is being hanled
3039 * by ptlrpc_svcpt_stop_threads now
3041 thread_add_flags(thread, SVC_STOPPED);
3042 wake_up(&thread->t_ctl_waitq);
3043 spin_unlock(&svcpt->scp_lock);
3045 list_del(&thread->t_link);
3046 spin_unlock(&svcpt->scp_lock);
3047 OBD_FREE_PTR(thread);
3055 l_wait_event(thread->t_ctl_waitq,
3056 thread_is_running(thread) || thread_is_stopped(thread),
3059 rc = thread_is_stopped(thread) ? thread->t_id : 0;
3063 int ptlrpc_hr_init(void)
3065 struct ptlrpc_hr_partition *hrp;
3066 struct ptlrpc_hr_thread *hrt;
3073 memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
3074 ptlrpc_hr.hr_cpt_table = cfs_cpt_table;
3076 ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
3078 if (ptlrpc_hr.hr_partitions == NULL)
3081 ratelimit_state_init(&watchdog_limit,
3082 cfs_time_seconds(libcfs_watchdog_ratelimit), 3);
3084 init_waitqueue_head(&ptlrpc_hr.hr_waitq);
3086 weight = cpumask_weight(topology_sibling_cpumask(smp_processor_id()));
3088 cfs_percpt_for_each(hrp, cpt, ptlrpc_hr.hr_partitions) {
3091 atomic_set(&hrp->hrp_nstarted, 0);
3092 atomic_set(&hrp->hrp_nstopped, 0);
3094 hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, cpt);
3095 hrp->hrp_nthrs /= weight;
3096 if (hrp->hrp_nthrs == 0)
3099 OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, cpt,
3100 hrp->hrp_nthrs * sizeof(*hrt));
3101 if (hrp->hrp_thrs == NULL)
3102 GOTO(out, rc = -ENOMEM);
3104 for (i = 0; i < hrp->hrp_nthrs; i++) {
3105 hrt = &hrp->hrp_thrs[i];
3108 hrt->hrt_partition = hrp;
3109 init_waitqueue_head(&hrt->hrt_waitq);
3110 spin_lock_init(&hrt->hrt_lock);
3111 INIT_LIST_HEAD(&hrt->hrt_queue);
3115 rc = ptlrpc_start_hr_threads();
3122 void ptlrpc_hr_fini(void)
3124 struct ptlrpc_hr_partition *hrp;
3127 if (ptlrpc_hr.hr_partitions == NULL)
3130 ptlrpc_stop_hr_threads();
3132 cfs_percpt_for_each(hrp, cpt, ptlrpc_hr.hr_partitions) {
3133 if (hrp->hrp_thrs != NULL) {
3134 OBD_FREE(hrp->hrp_thrs,
3135 hrp->hrp_nthrs * sizeof(hrp->hrp_thrs[0]));
3139 cfs_percpt_free(ptlrpc_hr.hr_partitions);
3140 ptlrpc_hr.hr_partitions = NULL;
3145 * Wait until all already scheduled replies are processed.
3147 static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
3151 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
3154 rc = l_wait_event(svcpt->scp_waitq,
3155 atomic_read(&svcpt->scp_nreps_difficult) == 0, &lwi);
3158 CWARN("Unexpectedly long timeout %s %p\n",
3159 svcpt->scp_service->srv_name, svcpt->scp_service);
3164 ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
3166 struct ptlrpc_service_part *svcpt;
3169 /* early disarm AT timer... */
3170 ptlrpc_service_for_each_part(svcpt, i, svc) {
3171 if (svcpt->scp_service != NULL)
3172 del_timer(&svcpt->scp_at_timer);
3177 ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
3179 struct ptlrpc_service_part *svcpt;
3180 struct ptlrpc_request_buffer_desc *rqbd;
3181 struct l_wait_info lwi;
3185 /* All history will be culled when the next request buffer is
3186 * freed in ptlrpc_service_purge_all() */
3187 svc->srv_hist_nrqbds_cpt_max = 0;
3189 rc = LNetClearLazyPortal(svc->srv_req_portal);
3192 ptlrpc_service_for_each_part(svcpt, i, svc) {
3193 if (svcpt->scp_service == NULL)
3196 /* Unlink all the request buffers. This forces a 'final'
3197 * event with its 'unlink' flag set for each posted rqbd */
3198 list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
3200 rc = LNetMDUnlink(rqbd->rqbd_md_h);
3201 LASSERT(rc == 0 || rc == -ENOENT);
3205 ptlrpc_service_for_each_part(svcpt, i, svc) {
3206 if (svcpt->scp_service == NULL)
3209 /* Wait for the network to release any buffers
3210 * it's currently filling */
3211 spin_lock(&svcpt->scp_lock);
3212 while (svcpt->scp_nrqbds_posted != 0) {
3213 spin_unlock(&svcpt->scp_lock);
3214 /* Network access will complete in finite time but
3215 * the HUGE timeout lets us CWARN for visibility
3216 * of sluggish NALs */
3217 lwi = LWI_TIMEOUT_INTERVAL(
3218 cfs_time_seconds(LONG_UNLINK),
3219 cfs_time_seconds(1), NULL, NULL);
3220 rc = l_wait_event(svcpt->scp_waitq,
3221 svcpt->scp_nrqbds_posted == 0, &lwi);
3222 if (rc == -ETIMEDOUT) {
3223 CWARN("Service %s waiting for "
3224 "request buffers\n",
3225 svcpt->scp_service->srv_name);
3227 spin_lock(&svcpt->scp_lock);
3229 spin_unlock(&svcpt->scp_lock);
3234 ptlrpc_service_purge_all(struct ptlrpc_service *svc)
3236 struct ptlrpc_service_part *svcpt;
3237 struct ptlrpc_request_buffer_desc *rqbd;
3238 struct ptlrpc_request *req;
3239 struct ptlrpc_reply_state *rs;
3242 ptlrpc_service_for_each_part(svcpt, i, svc) {
3243 if (svcpt->scp_service == NULL)
3246 spin_lock(&svcpt->scp_rep_lock);
3247 while (!list_empty(&svcpt->scp_rep_active)) {
3248 rs = list_entry(svcpt->scp_rep_active.next,
3249 struct ptlrpc_reply_state, rs_list);
3250 spin_lock(&rs->rs_lock);
3251 ptlrpc_schedule_difficult_reply(rs);
3252 spin_unlock(&rs->rs_lock);
3254 spin_unlock(&svcpt->scp_rep_lock);
3256 /* purge the request queue. NB No new replies (rqbds
3257 * all unlinked) and no service threads, so I'm the only
3258 * thread noodling the request queue now */
3259 while (!list_empty(&svcpt->scp_req_incoming)) {
3260 req = list_entry(svcpt->scp_req_incoming.next,
3261 struct ptlrpc_request, rq_list);
3263 list_del(&req->rq_list);
3264 svcpt->scp_nreqs_incoming--;
3265 ptlrpc_server_finish_request(svcpt, req);
3268 while (ptlrpc_server_request_pending(svcpt, true)) {
3269 req = ptlrpc_server_request_get(svcpt, true);
3270 ptlrpc_server_finish_active_request(svcpt, req);
3273 LASSERT(list_empty(&svcpt->scp_rqbd_posted));
3274 LASSERT(svcpt->scp_nreqs_incoming == 0);
3275 LASSERT(svcpt->scp_nreqs_active == 0);
3276 /* history should have been culled by
3277 * ptlrpc_server_finish_request */
3278 LASSERT(svcpt->scp_hist_nrqbds == 0);
3280 /* Now free all the request buffers since nothing
3281 * references them any more... */
3283 while (!list_empty(&svcpt->scp_rqbd_idle)) {
3284 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
3285 struct ptlrpc_request_buffer_desc,
3287 ptlrpc_free_rqbd(rqbd);
3289 ptlrpc_wait_replies(svcpt);
3291 while (!list_empty(&svcpt->scp_rep_idle)) {
3292 rs = list_entry(svcpt->scp_rep_idle.next,
3293 struct ptlrpc_reply_state,
3295 list_del(&rs->rs_list);
3296 OBD_FREE_LARGE(rs, svc->srv_max_reply_size);
3302 ptlrpc_service_free(struct ptlrpc_service *svc)
3304 struct ptlrpc_service_part *svcpt;
3305 struct ptlrpc_at_array *array;
3308 ptlrpc_service_for_each_part(svcpt, i, svc) {
3309 if (svcpt->scp_service == NULL)
3312 /* In case somebody rearmed this in the meantime */
3313 del_timer(&svcpt->scp_at_timer);
3314 array = &svcpt->scp_at_array;
3316 if (array->paa_reqs_array != NULL) {
3317 OBD_FREE(array->paa_reqs_array,
3318 sizeof(struct list_head) * array->paa_size);
3319 array->paa_reqs_array = NULL;
3322 if (array->paa_reqs_count != NULL) {
3323 OBD_FREE(array->paa_reqs_count,
3324 sizeof(__u32) * array->paa_size);
3325 array->paa_reqs_count = NULL;
3329 ptlrpc_service_for_each_part(svcpt, i, svc)
3330 OBD_FREE_PTR(svcpt);
3332 if (svc->srv_cpts != NULL)
3333 cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts);
3335 OBD_FREE(svc, offsetof(struct ptlrpc_service,
3336 srv_parts[svc->srv_ncpts]));
3339 int ptlrpc_unregister_service(struct ptlrpc_service *service)
3343 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
3345 service->srv_is_stopping = 1;
3347 mutex_lock(&ptlrpc_all_services_mutex);
3348 list_del_init(&service->srv_list);
3349 mutex_unlock(&ptlrpc_all_services_mutex);
3351 ptlrpc_service_del_atimer(service);
3352 ptlrpc_stop_all_threads(service);
3354 ptlrpc_service_unlink_rqbd(service);
3355 ptlrpc_service_purge_all(service);
3356 ptlrpc_service_nrs_cleanup(service);
3358 ptlrpc_lprocfs_unregister_service(service);
3359 ptlrpc_sysfs_unregister_service(service);
3361 ptlrpc_service_free(service);
3365 EXPORT_SYMBOL(ptlrpc_unregister_service);
3368 * Returns 0 if the service is healthy.
3370 * Right now, it just checks to make sure that requests aren't languishing
3371 * in the queue. We'll use this health check to govern whether a node needs
3372 * to be shot, so it's intentionally non-aggressive. */
3373 static int ptlrpc_svcpt_health_check(struct ptlrpc_service_part *svcpt)
3375 struct ptlrpc_request *request = NULL;
3376 struct timespec64 right_now;
3377 struct timespec64 timediff;
3379 ktime_get_real_ts64(&right_now);
3381 spin_lock(&svcpt->scp_req_lock);
3382 /* How long has the next entry been waiting? */
3383 if (ptlrpc_server_high_pending(svcpt, true))
3384 request = ptlrpc_nrs_req_peek_nolock(svcpt, true);
3385 else if (ptlrpc_server_normal_pending(svcpt, true))
3386 request = ptlrpc_nrs_req_peek_nolock(svcpt, false);
3388 if (request == NULL) {
3389 spin_unlock(&svcpt->scp_req_lock);
3393 timediff = timespec64_sub(right_now, request->rq_arrival_time);
3394 spin_unlock(&svcpt->scp_req_lock);
3396 if ((timediff.tv_sec) >
3397 (AT_OFF ? obd_timeout * 3 / 2 : at_max)) {
3398 CERROR("%s: unhealthy - request has been waiting %llds\n",
3399 svcpt->scp_service->srv_name, (s64)timediff.tv_sec);
3407 ptlrpc_service_health_check(struct ptlrpc_service *svc)
3409 struct ptlrpc_service_part *svcpt;
3415 ptlrpc_service_for_each_part(svcpt, i, svc) {
3416 int rc = ptlrpc_svcpt_health_check(svcpt);
3423 EXPORT_SYMBOL(ptlrpc_service_health_check);