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 /* Increase max reply size to next power of two */
785 service->srv_max_reply_size = 1;
786 while (service->srv_max_reply_size <
787 conf->psc_buf.bc_rep_max_size + SPTLRPC_MAX_PAYLOAD)
788 service->srv_max_reply_size <<= 1;
790 service->srv_thread_name = conf->psc_thr.tc_thr_name;
791 service->srv_ctx_tags = conf->psc_thr.tc_ctx_tags;
792 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
793 service->srv_ops = conf->psc_ops;
795 for (i = 0; i < ncpts; i++) {
796 if (!cconf->cc_affinity)
799 cpt = cpts != NULL ? cpts[i] : i;
801 OBD_CPT_ALLOC(svcpt, cptable, cpt, sizeof(*svcpt));
803 GOTO(failed, rc = -ENOMEM);
805 service->srv_parts[i] = svcpt;
806 rc = ptlrpc_service_part_init(service, svcpt, cpt);
811 ptlrpc_server_nthreads_check(service, conf);
813 rc = LNetSetLazyPortal(service->srv_req_portal);
816 mutex_lock(&ptlrpc_all_services_mutex);
817 list_add(&service->srv_list, &ptlrpc_all_services);
818 mutex_unlock(&ptlrpc_all_services_mutex);
821 rc = ptlrpc_sysfs_register_service(parent, service);
826 if (debugfs_entry != NULL)
827 ptlrpc_ldebugfs_register_service(debugfs_entry, service);
829 rc = ptlrpc_service_nrs_setup(service);
833 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
834 service->srv_name, service->srv_req_portal);
836 rc = ptlrpc_start_threads(service);
838 CERROR("Failed to start threads for service %s: %d\n",
839 service->srv_name, rc);
845 ptlrpc_unregister_service(service);
848 EXPORT_SYMBOL(ptlrpc_register_service);
851 * to actually free the request, must be called without holding svc_lock.
852 * note it's caller's responsibility to unlink req->rq_list.
854 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
856 LASSERT(atomic_read(&req->rq_refcount) == 0);
857 LASSERT(list_empty(&req->rq_timed_list));
859 /* DEBUG_REQ() assumes the reply state of a request with a valid
860 * ref will not be destroyed until that reference is dropped. */
861 ptlrpc_req_drop_rs(req);
863 sptlrpc_svc_ctx_decref(req);
865 if (req != &req->rq_rqbd->rqbd_req) {
866 /* NB request buffers use an embedded
867 * req if the incoming req unlinked the
868 * MD; this isn't one of them! */
869 ptlrpc_request_cache_free(req);
874 * drop a reference count of the request. if it reaches 0, we either
875 * put it into history list, or free it immediately.
877 void ptlrpc_server_drop_request(struct ptlrpc_request *req)
879 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
880 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
881 struct ptlrpc_service *svc = svcpt->scp_service;
883 struct list_head *tmp;
884 struct list_head *nxt;
886 if (!atomic_dec_and_test(&req->rq_refcount))
889 if (req->rq_session.lc_state == LCS_ENTERED) {
890 lu_context_exit(&req->rq_session);
891 lu_context_fini(&req->rq_session);
894 if (req->rq_at_linked) {
895 spin_lock(&svcpt->scp_at_lock);
896 /* recheck with lock, in case it's unlinked by
897 * ptlrpc_at_check_timed() */
898 if (likely(req->rq_at_linked))
899 ptlrpc_at_remove_timed(req);
900 spin_unlock(&svcpt->scp_at_lock);
903 LASSERT(list_empty(&req->rq_timed_list));
905 /* finalize request */
906 if (req->rq_export) {
907 class_export_put(req->rq_export);
908 req->rq_export = NULL;
911 spin_lock(&svcpt->scp_lock);
913 list_add(&req->rq_list, &rqbd->rqbd_reqs);
915 refcount = --(rqbd->rqbd_refcount);
917 /* request buffer is now idle: add to history */
918 list_del(&rqbd->rqbd_list);
920 list_add_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds);
921 svcpt->scp_hist_nrqbds++;
923 /* cull some history?
924 * I expect only about 1 or 2 rqbds need to be recycled here */
925 while (svcpt->scp_hist_nrqbds > svc->srv_hist_nrqbds_cpt_max) {
926 rqbd = list_entry(svcpt->scp_hist_rqbds.next,
927 struct ptlrpc_request_buffer_desc,
930 list_del(&rqbd->rqbd_list);
931 svcpt->scp_hist_nrqbds--;
933 /* remove rqbd's reqs from svc's req history while
934 * I've got the service lock */
935 list_for_each(tmp, &rqbd->rqbd_reqs) {
936 req = list_entry(tmp, struct ptlrpc_request,
938 /* Track the highest culled req seq */
939 if (req->rq_history_seq >
940 svcpt->scp_hist_seq_culled) {
941 svcpt->scp_hist_seq_culled =
944 list_del(&req->rq_history_list);
947 spin_unlock(&svcpt->scp_lock);
949 list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
950 req = list_entry(rqbd->rqbd_reqs.next,
951 struct ptlrpc_request,
953 list_del(&req->rq_list);
954 ptlrpc_server_free_request(req);
957 spin_lock(&svcpt->scp_lock);
959 * now all reqs including the embedded req has been
960 * disposed, schedule request buffer for re-use
961 * or free it to drain some in excess.
963 LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) == 0);
964 if (svcpt->scp_nrqbds_posted >=
965 svc->srv_nbuf_per_group ||
966 (svc->srv_nrqbds_max != 0 &&
967 svcpt->scp_nrqbds_total > svc->srv_nrqbds_max) ||
968 test_req_buffer_pressure) {
969 /* like in ptlrpc_free_rqbd() */
970 svcpt->scp_nrqbds_total--;
971 OBD_FREE_LARGE(rqbd->rqbd_buffer,
975 list_add_tail(&rqbd->rqbd_list,
976 &svcpt->scp_rqbd_idle);
980 spin_unlock(&svcpt->scp_lock);
981 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
982 /* If we are low on memory, we are not interested in history */
983 list_del(&req->rq_list);
984 list_del_init(&req->rq_history_list);
986 /* Track the highest culled req seq */
987 if (req->rq_history_seq > svcpt->scp_hist_seq_culled)
988 svcpt->scp_hist_seq_culled = req->rq_history_seq;
990 spin_unlock(&svcpt->scp_lock);
992 ptlrpc_server_free_request(req);
994 spin_unlock(&svcpt->scp_lock);
998 /** Change request export and move hp request from old export to new */
999 void ptlrpc_request_change_export(struct ptlrpc_request *req,
1000 struct obd_export *export)
1002 if (req->rq_export != NULL) {
1003 LASSERT(!list_empty(&req->rq_exp_list));
1004 /* remove rq_exp_list from last export */
1005 spin_lock(&req->rq_export->exp_rpc_lock);
1006 list_del_init(&req->rq_exp_list);
1007 spin_unlock(&req->rq_export->exp_rpc_lock);
1008 /* export has one reference already, so it`s safe to
1009 * add req to export queue here and get another
1010 * reference for request later */
1011 spin_lock(&export->exp_rpc_lock);
1012 if (req->rq_ops != NULL) /* hp request */
1013 list_add(&req->rq_exp_list, &export->exp_hp_rpcs);
1015 list_add(&req->rq_exp_list, &export->exp_reg_rpcs);
1016 spin_unlock(&export->exp_rpc_lock);
1018 class_export_rpc_dec(req->rq_export);
1019 class_export_put(req->rq_export);
1022 /* request takes one export refcount */
1023 req->rq_export = class_export_get(export);
1024 class_export_rpc_inc(export);
1030 * to finish a request: stop sending more early replies, and release
1033 static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt,
1034 struct ptlrpc_request *req)
1036 ptlrpc_server_hpreq_fini(req);
1038 ptlrpc_server_drop_request(req);
1042 * to finish an active request: stop sending more early replies, and release
1043 * the request. should be called after we finished handling the request.
1045 static void ptlrpc_server_finish_active_request(
1046 struct ptlrpc_service_part *svcpt,
1047 struct ptlrpc_request *req)
1049 spin_lock(&svcpt->scp_req_lock);
1050 ptlrpc_nrs_req_stop_nolock(req);
1051 svcpt->scp_nreqs_active--;
1053 svcpt->scp_nhreqs_active--;
1054 spin_unlock(&svcpt->scp_req_lock);
1056 ptlrpc_nrs_req_finalize(req);
1058 if (req->rq_export != NULL)
1059 class_export_rpc_dec(req->rq_export);
1061 ptlrpc_server_finish_request(svcpt, req);
1065 * This function makes sure dead exports are evicted in a timely manner.
1066 * This function is only called when some export receives a message (i.e.,
1067 * the network is up.)
1069 void ptlrpc_update_export_timer(struct obd_export *exp, time64_t extra_delay)
1071 struct obd_export *oldest_exp;
1072 time64_t oldest_time, new_time;
1078 /* Compensate for slow machines, etc, by faking our request time
1079 into the future. Although this can break the strict time-ordering
1080 of the list, we can be really lazy here - we don't have to evict
1081 at the exact right moment. Eventually, all silent exports
1082 will make it to the top of the list. */
1084 /* Do not pay attention on 1sec or smaller renewals. */
1085 new_time = ktime_get_real_seconds() + extra_delay;
1086 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
1089 exp->exp_last_request_time = new_time;
1091 /* exports may get disconnected from the chain even though the
1092 export has references, so we must keep the spin lock while
1093 manipulating the lists */
1094 spin_lock(&exp->exp_obd->obd_dev_lock);
1096 if (list_empty(&exp->exp_obd_chain_timed)) {
1097 /* this one is not timed */
1098 spin_unlock(&exp->exp_obd->obd_dev_lock);
1102 list_move_tail(&exp->exp_obd_chain_timed,
1103 &exp->exp_obd->obd_exports_timed);
1105 oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
1106 struct obd_export, exp_obd_chain_timed);
1107 oldest_time = oldest_exp->exp_last_request_time;
1108 spin_unlock(&exp->exp_obd->obd_dev_lock);
1110 if (exp->exp_obd->obd_recovering) {
1111 /* be nice to everyone during recovery */
1116 /* Note - racing to start/reset the obd_eviction timer is safe */
1117 if (exp->exp_obd->obd_eviction_timer == 0) {
1118 /* Check if the oldest entry is expired. */
1119 if (ktime_get_real_seconds() >
1120 oldest_time + PING_EVICT_TIMEOUT + extra_delay) {
1121 /* We need a second timer, in case the net was down and
1122 * it just came back. Since the pinger may skip every
1123 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
1124 * we better wait for 3.
1126 exp->exp_obd->obd_eviction_timer =
1127 ktime_get_real_seconds() + 3 * PING_INTERVAL;
1128 CDEBUG(D_HA, "%s: Think about evicting %s from %lld\n",
1129 exp->exp_obd->obd_name,
1130 obd_export_nid2str(oldest_exp), oldest_time);
1133 if (ktime_get_real_seconds() >
1134 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
1135 /* The evictor won't evict anyone who we've heard from
1136 * recently, so we don't have to check before we start
1139 if (!ping_evictor_wake(exp))
1140 exp->exp_obd->obd_eviction_timer = 0;
1148 * Sanity check request \a req.
1149 * Return 0 if all is ok, error code otherwise.
1151 static int ptlrpc_check_req(struct ptlrpc_request *req)
1153 struct obd_device *obd = req->rq_export->exp_obd;
1156 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
1157 req->rq_export->exp_conn_cnt)) {
1158 DEBUG_REQ(D_RPCTRACE, req,
1159 "DROPPING req from old connection %d < %d",
1160 lustre_msg_get_conn_cnt(req->rq_reqmsg),
1161 req->rq_export->exp_conn_cnt);
1164 if (unlikely(obd == NULL || obd->obd_fail)) {
1165 /* Failing over, don't handle any more reqs,
1166 * send error response instead. */
1167 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
1168 req, (obd != NULL) ? obd->obd_name : "unknown");
1170 } else if (lustre_msg_get_flags(req->rq_reqmsg) &
1171 (MSG_REPLAY | MSG_REQ_REPLAY_DONE) &&
1172 !obd->obd_recovering) {
1173 DEBUG_REQ(D_ERROR, req,
1174 "Invalid replay without recovery");
1175 class_fail_export(req->rq_export);
1177 } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 &&
1178 !obd->obd_recovering) {
1179 DEBUG_REQ(D_ERROR, req, "Invalid req with transno "
1180 "%llu without recovery",
1181 lustre_msg_get_transno(req->rq_reqmsg));
1182 class_fail_export(req->rq_export);
1186 if (unlikely(rc < 0)) {
1187 req->rq_status = rc;
1193 static void ptlrpc_at_set_timer(struct ptlrpc_service_part *svcpt)
1195 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1198 if (array->paa_count == 0) {
1199 del_timer(&svcpt->scp_at_timer);
1203 /* Set timer for closest deadline */
1204 next = array->paa_deadline - ktime_get_real_seconds() -
1207 ptlrpc_at_timer(cfs_timer_cb_arg(svcpt, scp_at_timer));
1209 mod_timer(&svcpt->scp_at_timer,
1210 jiffies + nsecs_to_jiffies(next * NSEC_PER_SEC));
1211 CDEBUG(D_INFO, "armed %s at %+llds\n",
1212 svcpt->scp_service->srv_name, next);
1216 /* Add rpc to early reply check list */
1217 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
1219 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1220 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1221 struct ptlrpc_request *rq = NULL;
1227 if (req->rq_no_reply)
1230 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
1233 spin_lock(&svcpt->scp_at_lock);
1234 LASSERT(list_empty(&req->rq_timed_list));
1236 div_u64_rem(req->rq_deadline, array->paa_size, &index);
1237 if (array->paa_reqs_count[index] > 0) {
1238 /* latest rpcs will have the latest deadlines in the list,
1239 * so search backward. */
1240 list_for_each_entry_reverse(rq,
1241 &array->paa_reqs_array[index],
1243 if (req->rq_deadline >= rq->rq_deadline) {
1244 list_add(&req->rq_timed_list,
1245 &rq->rq_timed_list);
1251 /* Add the request at the head of the list */
1252 if (list_empty(&req->rq_timed_list))
1253 list_add(&req->rq_timed_list,
1254 &array->paa_reqs_array[index]);
1256 spin_lock(&req->rq_lock);
1257 req->rq_at_linked = 1;
1258 spin_unlock(&req->rq_lock);
1259 req->rq_at_index = index;
1260 array->paa_reqs_count[index]++;
1262 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
1263 array->paa_deadline = req->rq_deadline;
1264 ptlrpc_at_set_timer(svcpt);
1266 spin_unlock(&svcpt->scp_at_lock);
1272 ptlrpc_at_remove_timed(struct ptlrpc_request *req)
1274 struct ptlrpc_at_array *array;
1276 array = &req->rq_rqbd->rqbd_svcpt->scp_at_array;
1278 /* NB: must call with hold svcpt::scp_at_lock */
1279 LASSERT(!list_empty(&req->rq_timed_list));
1280 list_del_init(&req->rq_timed_list);
1282 spin_lock(&req->rq_lock);
1283 req->rq_at_linked = 0;
1284 spin_unlock(&req->rq_lock);
1286 array->paa_reqs_count[req->rq_at_index]--;
1291 * Attempt to extend the request deadline by sending an early reply to the
1294 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
1296 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1297 struct ptlrpc_request *reqcopy;
1298 struct lustre_msg *reqmsg;
1299 time64_t olddl = req->rq_deadline - ktime_get_real_seconds();
1305 if (CFS_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_RECONNECT)) {
1306 /* don't send early reply */
1310 /* deadline is when the client expects us to reply, margin is the
1311 difference between clients' and servers' expectations */
1312 DEBUG_REQ(D_ADAPTTO, req,
1313 "%ssending early reply (deadline %+llds, margin %+llds) for "
1314 "%d+%d", AT_OFF ? "AT off - not " : "",
1315 (s64)olddl, (s64)(olddl - at_get(&svcpt->scp_at_estimate)),
1316 at_get(&svcpt->scp_at_estimate), at_extra);
1322 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+llds), "
1323 "not sending early reply. Consider increasing "
1324 "at_early_margin (%d)?", (s64)olddl, at_early_margin);
1326 /* Return an error so we're not re-added to the timed list. */
1330 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
1331 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
1332 "but no AT support");
1336 if (req->rq_export &&
1337 lustre_msg_get_flags(req->rq_reqmsg) &
1338 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1339 struct obd_device *obd_exp = req->rq_export->exp_obd;
1341 /* During recovery, we don't want to send too many early
1342 * replies, but on the other hand we want to make sure the
1343 * client has enough time to resend if the rpc is lost. So
1344 * during the recovery period send at least 4 early replies,
1345 * spacing them every at_extra if we can. at_estimate should
1346 * always equal this fixed value during recovery.
1348 /* Don't account request processing time into AT history
1349 * during recovery, it is not service time we need but
1350 * includes also waiting time for recovering clients
1352 newdl = min_t(time64_t, at_extra,
1353 obd_exp->obd_recovery_timeout / 4) +
1354 ktime_get_real_seconds();
1356 /* We want to extend the request deadline by at_extra seconds,
1357 * so we set our service estimate to reflect how much time has
1358 * passed since this request arrived plus an additional
1359 * at_extra seconds. The client will calculate the new deadline
1360 * based on this service estimate (plus some additional time to
1361 * account for network latency). See ptlrpc_at_recv_early_reply
1363 at_measured(&svcpt->scp_at_estimate, at_extra +
1364 ktime_get_real_seconds() -
1365 req->rq_arrival_time.tv_sec);
1366 newdl = req->rq_arrival_time.tv_sec +
1367 at_get(&svcpt->scp_at_estimate);
1370 /* Check to see if we've actually increased the deadline -
1371 * we may be past adaptive_max */
1372 if (req->rq_deadline >= newdl) {
1373 DEBUG_REQ(D_WARNING, req, "Couldn't add any time (%lld/%lld), not sending early reply\n",
1374 (s64)olddl, (s64)(newdl - ktime_get_real_seconds()));
1378 reqcopy = ptlrpc_request_cache_alloc(GFP_NOFS);
1379 if (reqcopy == NULL)
1381 OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1383 GOTO(out_free, rc = -ENOMEM);
1386 reqcopy->rq_reply_state = NULL;
1387 reqcopy->rq_rep_swab_mask = 0;
1388 reqcopy->rq_pack_bulk = 0;
1389 reqcopy->rq_pack_udesc = 0;
1390 reqcopy->rq_packed_final = 0;
1391 sptlrpc_svc_ctx_addref(reqcopy);
1392 /* We only need the reqmsg for the magic */
1393 reqcopy->rq_reqmsg = reqmsg;
1394 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1397 * tgt_brw_read() and tgt_brw_write() may have decided not to reply.
1398 * Without this check, we would fail the rq_no_reply assertion in
1399 * ptlrpc_send_reply().
1401 if (reqcopy->rq_no_reply)
1402 GOTO(out, rc = -ETIMEDOUT);
1404 LASSERT(atomic_read(&req->rq_refcount));
1405 /** if it is last refcount then early reply isn't needed */
1406 if (atomic_read(&req->rq_refcount) == 1) {
1407 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
1408 "abort sending early reply\n");
1409 GOTO(out, rc = -EINVAL);
1412 /* Connection ref */
1413 reqcopy->rq_export = class_conn2export(
1414 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1415 if (reqcopy->rq_export == NULL)
1416 GOTO(out, rc = -ENODEV);
1419 class_export_rpc_inc(reqcopy->rq_export);
1420 if (reqcopy->rq_export->exp_obd &&
1421 reqcopy->rq_export->exp_obd->obd_fail)
1422 GOTO(out_put, rc = -ENODEV);
1424 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1428 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1431 /* Adjust our own deadline to what we told the client */
1432 req->rq_deadline = newdl;
1433 req->rq_early_count++; /* number sent, server side */
1435 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1438 /* Free the (early) reply state from lustre_pack_reply.
1439 (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
1440 ptlrpc_req_drop_rs(reqcopy);
1443 class_export_rpc_dec(reqcopy->rq_export);
1444 class_export_put(reqcopy->rq_export);
1446 sptlrpc_svc_ctx_decref(reqcopy);
1447 OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1449 ptlrpc_request_cache_free(reqcopy);
1453 /* Send early replies to everybody expiring within at_early_margin
1454 asking for at_extra time */
1455 static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1457 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1458 struct ptlrpc_request *rq, *n;
1459 struct list_head work_list;
1462 time64_t now = ktime_get_real_seconds();
1464 int first, counter = 0;
1467 spin_lock(&svcpt->scp_at_lock);
1468 if (svcpt->scp_at_check == 0) {
1469 spin_unlock(&svcpt->scp_at_lock);
1472 delay = ktime_ms_delta(ktime_get(), svcpt->scp_at_checktime);
1473 svcpt->scp_at_check = 0;
1475 if (array->paa_count == 0) {
1476 spin_unlock(&svcpt->scp_at_lock);
1480 /* The timer went off, but maybe the nearest rpc already completed. */
1481 first = array->paa_deadline - now;
1482 if (first > at_early_margin) {
1483 /* We've still got plenty of time. Reset the timer. */
1484 ptlrpc_at_set_timer(svcpt);
1485 spin_unlock(&svcpt->scp_at_lock);
1489 /* We're close to a timeout, and we don't know how much longer the
1490 server will take. Send early replies to everyone expiring soon. */
1491 INIT_LIST_HEAD(&work_list);
1493 div_u64_rem(array->paa_deadline, array->paa_size, &index);
1494 count = array->paa_count;
1496 count -= array->paa_reqs_count[index];
1497 list_for_each_entry_safe(rq, n,
1498 &array->paa_reqs_array[index],
1500 if (rq->rq_deadline > now + at_early_margin) {
1501 /* update the earliest deadline */
1502 if (deadline == -1 ||
1503 rq->rq_deadline < deadline)
1504 deadline = rq->rq_deadline;
1509 * ptlrpc_server_drop_request() may drop
1510 * refcount to 0 already. Let's check this and
1511 * don't add entry to work_list
1513 if (likely(atomic_inc_not_zero(&rq->rq_refcount))) {
1514 ptlrpc_at_remove_timed(rq);
1515 list_add(&rq->rq_timed_list, &work_list);
1517 ptlrpc_at_remove_timed(rq);
1523 if (++index >= array->paa_size)
1526 array->paa_deadline = deadline;
1527 /* we have a new earliest deadline, restart the timer */
1528 ptlrpc_at_set_timer(svcpt);
1530 spin_unlock(&svcpt->scp_at_lock);
1532 CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
1533 "replies\n", first, at_extra, counter);
1535 /* We're already past request deadlines before we even get a
1536 chance to send early replies */
1537 LCONSOLE_WARN("%s: This server is not able to keep up with "
1538 "request traffic (cpu-bound).\n",
1539 svcpt->scp_service->srv_name);
1540 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, delay=%lld\n",
1541 counter, svcpt->scp_nreqs_incoming,
1542 svcpt->scp_nreqs_active,
1543 at_get(&svcpt->scp_at_estimate), delay);
1546 /* we took additional refcount so entries can't be deleted from list, no
1547 * locking is needed */
1548 while (!list_empty(&work_list)) {
1549 rq = list_entry(work_list.next, struct ptlrpc_request,
1551 list_del_init(&rq->rq_timed_list);
1553 if (ptlrpc_at_send_early_reply(rq) == 0)
1554 ptlrpc_at_add_timed(rq);
1556 ptlrpc_server_drop_request(rq);
1559 RETURN(1); /* return "did_something" for liblustre */
1562 /* Check if we are already handling earlier incarnation of this request.
1563 * Called under &req->rq_export->exp_rpc_lock locked */
1564 static struct ptlrpc_request*
1565 ptlrpc_server_check_resend_in_progress(struct ptlrpc_request *req)
1567 struct ptlrpc_request *tmp = NULL;
1569 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ||
1570 (atomic_read(&req->rq_export->exp_rpc_count) == 0))
1573 /* bulk request are aborted upon reconnect, don't try to
1575 if (req->rq_bulk_write || req->rq_bulk_read)
1578 /* This list should not be longer than max_requests in
1579 * flights on the client, so it is not all that long.
1580 * Also we only hit this codepath in case of a resent
1581 * request which makes it even more rarely hit */
1582 list_for_each_entry(tmp, &req->rq_export->exp_reg_rpcs,
1584 /* Found duplicate one */
1585 if (tmp->rq_xid == req->rq_xid)
1588 list_for_each_entry(tmp, &req->rq_export->exp_hp_rpcs,
1590 /* Found duplicate one */
1591 if (tmp->rq_xid == req->rq_xid)
1597 DEBUG_REQ(D_HA, req, "Found duplicate req in processing");
1598 DEBUG_REQ(D_HA, tmp, "Request being processed");
1603 * Check if a request should be assigned with a high priority.
1605 * \retval < 0: error occurred
1606 * 0: normal RPC request
1607 * +1: high priority request
1609 static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt,
1610 struct ptlrpc_request *req)
1615 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL) {
1616 rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
1623 if (req->rq_export != NULL && req->rq_ops != NULL) {
1624 /* Perform request specific check. We should do this
1625 * check before the request is added into exp_hp_rpcs
1626 * list otherwise it may hit swab race at LU-1044. */
1627 if (req->rq_ops->hpreq_check != NULL) {
1628 rc = req->rq_ops->hpreq_check(req);
1629 if (rc == -ESTALE) {
1630 req->rq_status = rc;
1633 /** can only return error,
1634 * 0 for normal request,
1635 * or 1 for high priority request */
1643 /** Remove the request from the export list. */
1644 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req)
1647 if (req->rq_export) {
1648 /* refresh lock timeout again so that client has more
1649 * room to send lock cancel RPC. */
1650 if (req->rq_ops && req->rq_ops->hpreq_fini)
1651 req->rq_ops->hpreq_fini(req);
1653 spin_lock(&req->rq_export->exp_rpc_lock);
1654 list_del_init(&req->rq_exp_list);
1655 spin_unlock(&req->rq_export->exp_rpc_lock);
1660 static int ptlrpc_hpreq_check(struct ptlrpc_request *req)
1665 static struct ptlrpc_hpreq_ops ptlrpc_hpreq_common = {
1666 .hpreq_check = ptlrpc_hpreq_check,
1669 /* Hi-Priority RPC check by RPC operation code. */
1670 int ptlrpc_hpreq_handler(struct ptlrpc_request *req)
1672 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1674 /* Check for export to let only reconnects for not yet evicted
1675 * export to become a HP rpc. */
1676 if ((req->rq_export != NULL) &&
1677 (opc == OBD_PING || opc == MDS_CONNECT || opc == OST_CONNECT))
1678 req->rq_ops = &ptlrpc_hpreq_common;
1682 EXPORT_SYMBOL(ptlrpc_hpreq_handler);
1684 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1685 struct ptlrpc_request *req)
1689 struct ptlrpc_request *orig;
1692 rc = ptlrpc_server_hpreq_init(svcpt, req);
1697 ptlrpc_nrs_req_initialize(svcpt, req, hp);
1699 if (req->rq_export != NULL) {
1700 struct obd_export *exp = req->rq_export;
1702 /* do search for duplicated xid and the adding to the list
1704 spin_lock_bh(&exp->exp_rpc_lock);
1705 orig = ptlrpc_server_check_resend_in_progress(req);
1706 if (orig && likely(atomic_inc_not_zero(&orig->rq_refcount))) {
1709 spin_unlock_bh(&exp->exp_rpc_lock);
1712 * When the client resend request and the server has
1713 * the previous copy of it, we need to update deadlines,
1714 * to be sure that the client and the server have equal
1715 * request deadlines.
1718 spin_lock(&orig->rq_rqbd->rqbd_svcpt->scp_at_lock);
1719 linked = orig->rq_at_linked;
1721 ptlrpc_at_remove_timed(orig);
1722 spin_unlock(&orig->rq_rqbd->rqbd_svcpt->scp_at_lock);
1723 orig->rq_deadline = req->rq_deadline;
1725 ptlrpc_at_add_timed(orig);
1726 ptlrpc_server_drop_request(orig);
1727 ptlrpc_nrs_req_finalize(req);
1731 if (hp || req->rq_ops != NULL)
1732 list_add(&req->rq_exp_list, &exp->exp_hp_rpcs);
1734 list_add(&req->rq_exp_list, &exp->exp_reg_rpcs);
1735 spin_unlock_bh(&exp->exp_rpc_lock);
1738 /* the current thread is not the processing thread for this request
1739 * since that, but request is in exp_hp_list and can be find there.
1740 * Remove all relations between request and old thread. */
1741 req->rq_svc_thread->t_env->le_ses = NULL;
1742 req->rq_svc_thread = NULL;
1743 req->rq_session.lc_thread = NULL;
1745 ptlrpc_nrs_req_add(svcpt, req, hp);
1751 * Allow to handle high priority request
1752 * User can call it w/o any lock but need to hold
1753 * ptlrpc_service_part::scp_req_lock to get reliable result
1755 static bool ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1758 int running = svcpt->scp_nthrs_running;
1760 if (!nrs_svcpt_has_hp(svcpt))
1766 if (ptlrpc_nrs_req_throttling_nolock(svcpt, true))
1769 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1770 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1771 /* leave just 1 thread for normal RPCs */
1772 running = PTLRPC_NTHRS_INIT;
1773 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1777 if (svcpt->scp_nreqs_active >= running - 1)
1780 if (svcpt->scp_nhreqs_active == 0)
1783 return !ptlrpc_nrs_req_pending_nolock(svcpt, false) ||
1784 svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1787 static bool ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1790 return ptlrpc_server_allow_high(svcpt, force) &&
1791 ptlrpc_nrs_req_pending_nolock(svcpt, true);
1795 * Only allow normal priority requests on a service that has a high-priority
1796 * queue if forced (i.e. cleanup), if there are other high priority requests
1797 * already being processed (i.e. those threads can service more high-priority
1798 * requests), or if there are enough idle threads that a later thread can do
1799 * a high priority request.
1800 * User can call it w/o any lock but need to hold
1801 * ptlrpc_service_part::scp_req_lock to get reliable result
1803 static bool ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1806 int running = svcpt->scp_nthrs_running;
1807 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1808 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1809 /* leave just 1 thread for normal RPCs */
1810 running = PTLRPC_NTHRS_INIT;
1811 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1818 if (ptlrpc_nrs_req_throttling_nolock(svcpt, false))
1821 if (svcpt->scp_nreqs_active < running - 2)
1824 if (svcpt->scp_nreqs_active >= running - 1)
1827 return svcpt->scp_nhreqs_active > 0 || !nrs_svcpt_has_hp(svcpt);
1830 static bool ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1833 return ptlrpc_server_allow_normal(svcpt, force) &&
1834 ptlrpc_nrs_req_pending_nolock(svcpt, false);
1838 * Returns true if there are requests available in incoming
1839 * request queue for processing and it is allowed to fetch them.
1840 * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1841 * to get reliable result
1842 * \see ptlrpc_server_allow_normal
1843 * \see ptlrpc_server_allow high
1846 ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt, bool force)
1848 return ptlrpc_server_high_pending(svcpt, force) ||
1849 ptlrpc_server_normal_pending(svcpt, force);
1853 * Fetch a request for processing from queue of unprocessed requests.
1854 * Favors high-priority requests.
1855 * Returns a pointer to fetched request.
1857 static struct ptlrpc_request *
1858 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force)
1860 struct ptlrpc_request *req = NULL;
1863 spin_lock(&svcpt->scp_req_lock);
1865 if (ptlrpc_server_high_pending(svcpt, force)) {
1866 req = ptlrpc_nrs_req_get_nolock(svcpt, true, force);
1868 svcpt->scp_hreq_count++;
1873 if (ptlrpc_server_normal_pending(svcpt, force)) {
1874 req = ptlrpc_nrs_req_get_nolock(svcpt, false, force);
1876 svcpt->scp_hreq_count = 0;
1881 spin_unlock(&svcpt->scp_req_lock);
1885 svcpt->scp_nreqs_active++;
1887 svcpt->scp_nhreqs_active++;
1889 spin_unlock(&svcpt->scp_req_lock);
1891 if (likely(req->rq_export))
1892 class_export_rpc_inc(req->rq_export);
1898 * Handle freshly incoming reqs, add to timed early reply list,
1899 * pass on to regular request queue.
1900 * All incoming requests pass through here before getting into
1901 * ptlrpc_server_handle_req later on.
1904 ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt,
1905 struct ptlrpc_thread *thread)
1907 struct ptlrpc_service *svc = svcpt->scp_service;
1908 struct ptlrpc_request *req;
1913 spin_lock(&svcpt->scp_lock);
1914 if (list_empty(&svcpt->scp_req_incoming)) {
1915 spin_unlock(&svcpt->scp_lock);
1919 req = list_entry(svcpt->scp_req_incoming.next,
1920 struct ptlrpc_request, rq_list);
1921 list_del_init(&req->rq_list);
1922 svcpt->scp_nreqs_incoming--;
1923 /* Consider this still a "queued" request as far as stats are
1925 spin_unlock(&svcpt->scp_lock);
1927 /* go through security check/transform */
1928 rc = sptlrpc_svc_unwrap_request(req);
1932 case SECSVC_COMPLETE:
1933 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1942 * for null-flavored rpc, msg has been unpacked by sptlrpc, although
1943 * redo it wouldn't be harmful.
1945 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
1946 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
1948 CERROR("error unpacking request: ptl %d from %s "
1949 "x%llu\n", svc->srv_req_portal,
1950 libcfs_id2str(req->rq_peer), req->rq_xid);
1955 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1957 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
1958 "%llu\n", svc->srv_req_portal,
1959 libcfs_id2str(req->rq_peer), req->rq_xid);
1963 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
1964 lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
1965 CERROR("drop incoming rpc opc %u, x%llu\n",
1966 cfs_fail_val, req->rq_xid);
1971 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1972 CERROR("wrong packet type received (type=%u) from %s\n",
1973 lustre_msg_get_type(req->rq_reqmsg),
1974 libcfs_id2str(req->rq_peer));
1978 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1982 req->rq_bulk_write = 1;
1986 case MGS_CONFIG_READ:
1987 req->rq_bulk_read = 1;
1991 CDEBUG(D_RPCTRACE, "got req x%llu\n", req->rq_xid);
1993 req->rq_export = class_conn2export(
1994 lustre_msg_get_handle(req->rq_reqmsg));
1995 if (req->rq_export) {
1996 rc = ptlrpc_check_req(req);
1998 rc = sptlrpc_target_export_check(req->rq_export, req);
2000 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
2001 "illegal security flavor,");
2006 ptlrpc_update_export_timer(req->rq_export, 0);
2009 /* req_in handling should/must be fast */
2010 if (ktime_get_real_seconds() - req->rq_arrival_time.tv_sec > 5)
2011 DEBUG_REQ(D_WARNING, req, "Slow req_in handling %llds",
2012 (s64)(ktime_get_real_seconds() -
2013 req->rq_arrival_time.tv_sec));
2015 /* Set rpc server deadline and add it to the timed list */
2016 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
2017 MSGHDR_AT_SUPPORT) ?
2018 /* The max time the client expects us to take */
2019 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
2021 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
2022 if (unlikely(deadline == 0)) {
2023 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
2027 /* Skip early reply */
2028 if (OBD_FAIL_PRECHECK(OBD_FAIL_MDS_RESEND))
2029 req->rq_deadline += obd_timeout;
2031 req->rq_svc_thread = thread;
2032 if (thread != NULL) {
2033 /* initialize request session, it is needed for request
2034 * processing by target */
2035 rc = lu_context_init(&req->rq_session, LCT_SERVER_SESSION |
2038 CERROR("%s: failure to initialize session: rc = %d\n",
2039 thread->t_name, rc);
2042 req->rq_session.lc_thread = thread;
2043 lu_context_enter(&req->rq_session);
2044 thread->t_env->le_ses = &req->rq_session;
2047 ptlrpc_at_add_timed(req);
2049 /* Move it over to the request processing queue */
2050 rc = ptlrpc_server_request_add(svcpt, req);
2054 wake_up(&svcpt->scp_waitq);
2058 ptlrpc_server_finish_request(svcpt, req);
2064 * Main incoming request handling logic.
2065 * Calls handler function from service to do actual processing.
2068 ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
2069 struct ptlrpc_thread *thread)
2071 struct ptlrpc_service *svc = svcpt->scp_service;
2072 struct ptlrpc_request *request;
2082 request = ptlrpc_server_request_get(svcpt, false);
2083 if (request == NULL)
2086 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
2087 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
2088 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
2089 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
2091 if (unlikely(fail_opc)) {
2092 if (request->rq_export && request->rq_ops)
2093 OBD_FAIL_TIMEOUT(fail_opc, 4);
2096 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
2098 if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
2099 libcfs_debug_dumplog();
2101 work_start = ktime_get_real();
2102 arrived = timespec64_to_ktime(request->rq_arrival_time);
2103 timediff_usecs = ktime_us_delta(work_start, arrived);
2104 if (likely(svc->srv_stats != NULL)) {
2105 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
2107 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
2108 svcpt->scp_nreqs_incoming);
2109 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
2110 svcpt->scp_nreqs_active);
2111 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
2112 at_get(&svcpt->scp_at_estimate));
2115 if (likely(request->rq_export)) {
2116 if (unlikely(ptlrpc_check_req(request)))
2118 ptlrpc_update_export_timer(request->rq_export,
2119 div_u64(timediff_usecs,
2123 /* Discard requests queued for longer than the deadline.
2124 The deadline is increased if we send an early reply. */
2125 if (ktime_get_real_seconds() > request->rq_deadline) {
2126 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s: deadline %lld:%llds ago\n",
2127 libcfs_id2str(request->rq_peer),
2128 request->rq_deadline -
2129 request->rq_arrival_time.tv_sec,
2130 ktime_get_real_seconds() - request->rq_deadline);
2134 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
2135 "%s:%s+%d:%d:x%llu:%s:%d\n", current_comm(),
2136 (request->rq_export ?
2137 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2138 (request->rq_export ?
2139 atomic_read(&request->rq_export->exp_refcount) : -99),
2140 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
2141 libcfs_id2str(request->rq_peer),
2142 lustre_msg_get_opc(request->rq_reqmsg));
2144 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
2145 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
2147 CDEBUG(D_NET, "got req %llu\n", request->rq_xid);
2149 /* re-assign request and sesson thread to the current one */
2150 request->rq_svc_thread = thread;
2151 if (thread != NULL) {
2152 LASSERT(request->rq_session.lc_thread == NULL);
2153 request->rq_session.lc_thread = thread;
2154 thread->t_env->le_ses = &request->rq_session;
2156 svc->srv_ops.so_req_handler(request);
2158 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
2161 if (unlikely(ktime_get_real_seconds() > request->rq_deadline)) {
2162 DEBUG_REQ(D_WARNING, request,
2163 "Request took longer than estimated (%lld:%llds); "
2164 "client may timeout.",
2165 request->rq_deadline -
2166 request->rq_arrival_time.tv_sec,
2167 ktime_get_real_seconds() - request->rq_deadline);
2170 work_end = ktime_get_real();
2171 timediff_usecs = ktime_us_delta(work_end, work_start);
2172 arrived_usecs = ktime_us_delta(work_end, arrived);
2173 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
2174 "%s:%s+%d:%d:x%llu:%s:%d Request processed in %lldus "
2175 "(%lldus total) trans %llu rc %d/%d\n",
2177 (request->rq_export ?
2178 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2179 (request->rq_export ?
2180 atomic_read(&request->rq_export->exp_refcount) : -99),
2181 lustre_msg_get_status(request->rq_reqmsg),
2183 libcfs_id2str(request->rq_peer),
2184 lustre_msg_get_opc(request->rq_reqmsg),
2187 (request->rq_repmsg ?
2188 lustre_msg_get_transno(request->rq_repmsg) :
2189 request->rq_transno),
2191 (request->rq_repmsg ?
2192 lustre_msg_get_status(request->rq_repmsg) : -999));
2193 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
2194 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
2195 int opc = opcode_offset(op);
2196 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
2197 LASSERT(opc < LUSTRE_MAX_OPCODES);
2198 lprocfs_counter_add(svc->srv_stats,
2199 opc + EXTRA_MAX_OPCODES,
2203 if (unlikely(request->rq_early_count)) {
2204 DEBUG_REQ(D_ADAPTTO, request,
2205 "sent %d early replies before finishing in %llds",
2206 request->rq_early_count,
2207 div_u64(arrived_usecs, USEC_PER_SEC));
2210 ptlrpc_server_finish_active_request(svcpt, request);
2216 * An internal function to process a single reply state object.
2219 ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
2221 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
2222 struct ptlrpc_service *svc = svcpt->scp_service;
2223 struct obd_export *exp;
2228 exp = rs->rs_export;
2230 LASSERT(rs->rs_difficult);
2231 LASSERT(rs->rs_scheduled);
2232 LASSERT(list_empty(&rs->rs_list));
2234 /* The disk commit callback holds exp_uncommitted_replies_lock while it
2235 * iterates over newly committed replies, removing them from
2236 * exp_uncommitted_replies. It then drops this lock and schedules the
2237 * replies it found for handling here.
2239 * We can avoid contention for exp_uncommitted_replies_lock between the
2240 * HRT threads and further commit callbacks by checking rs_committed
2241 * which is set in the commit callback while it holds both
2242 * rs_lock and exp_uncommitted_reples.
2244 * If we see rs_committed clear, the commit callback _may_ not have
2245 * handled this reply yet and we race with it to grab
2246 * exp_uncommitted_replies_lock before removing the reply from
2247 * exp_uncommitted_replies. Note that if we lose the race and the
2248 * reply has already been removed, list_del_init() is a noop.
2250 * If we see rs_committed set, we know the commit callback is handling,
2251 * or has handled this reply since store reordering might allow us to
2252 * see rs_committed set out of sequence. But since this is done
2253 * holding rs_lock, we can be sure it has all completed once we hold
2254 * rs_lock, which we do right next.
2256 if (!rs->rs_committed) {
2257 /* if rs was commited, no need to convert locks, don't check
2258 * rs_committed here because rs may never be added into
2259 * exp_uncommitted_replies and this flag never be set, see
2260 * target_send_reply() */
2261 if (rs->rs_convert_lock &&
2262 rs->rs_transno > exp->exp_last_committed) {
2263 struct ldlm_lock *lock;
2264 struct ldlm_lock *ack_locks[RS_MAX_LOCKS] = { NULL };
2266 spin_lock(&rs->rs_lock);
2267 if (rs->rs_convert_lock &&
2268 rs->rs_transno > exp->exp_last_committed) {
2269 nlocks = rs->rs_nlocks;
2270 while (nlocks-- > 0) {
2272 * NB don't assume rs is always handled
2273 * by the same service thread (see
2274 * ptlrpc_hr_select, so REP-ACK hr may
2275 * race with trans commit, while the
2276 * latter will release locks, get locks
2277 * here early to convert to COS mode
2280 lock = ldlm_handle2lock(
2281 &rs->rs_locks[nlocks]);
2283 ack_locks[nlocks] = lock;
2284 rs->rs_modes[nlocks] = LCK_COS;
2286 nlocks = rs->rs_nlocks;
2287 rs->rs_convert_lock = 0;
2288 /* clear rs_scheduled so that commit callback
2289 * can schedule again */
2290 rs->rs_scheduled = 0;
2291 spin_unlock(&rs->rs_lock);
2293 while (nlocks-- > 0) {
2294 lock = ack_locks[nlocks];
2295 ldlm_lock_mode_downgrade(lock, LCK_COS);
2296 LDLM_LOCK_PUT(lock);
2300 spin_unlock(&rs->rs_lock);
2303 spin_lock(&exp->exp_uncommitted_replies_lock);
2304 list_del_init(&rs->rs_obd_list);
2305 spin_unlock(&exp->exp_uncommitted_replies_lock);
2308 spin_lock(&exp->exp_lock);
2309 /* Noop if removed already */
2310 list_del_init(&rs->rs_exp_list);
2311 spin_unlock(&exp->exp_lock);
2313 spin_lock(&rs->rs_lock);
2315 been_handled = rs->rs_handled;
2318 nlocks = rs->rs_nlocks; /* atomic "steal", but */
2319 rs->rs_nlocks = 0; /* locks still on rs_locks! */
2321 if (nlocks == 0 && !been_handled) {
2322 /* If we see this, we should already have seen the warning
2323 * in mds_steal_ack_locks() */
2324 CDEBUG(D_HA, "All locks stolen from rs %p x%lld.t%lld"
2327 rs->rs_xid, rs->rs_transno, rs->rs_opc,
2328 libcfs_nid2str(exp->exp_connection->c_peer.nid));
2331 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
2332 spin_unlock(&rs->rs_lock);
2334 if (!been_handled && rs->rs_on_net) {
2335 LNetMDUnlink(rs->rs_md_h);
2336 /* Ignore return code; we're racing with completion */
2339 while (nlocks-- > 0)
2340 ldlm_lock_decref(&rs->rs_locks[nlocks],
2341 rs->rs_modes[nlocks]);
2343 spin_lock(&rs->rs_lock);
2346 rs->rs_scheduled = 0;
2347 rs->rs_convert_lock = 0;
2349 if (!rs->rs_on_net) {
2351 spin_unlock(&rs->rs_lock);
2353 class_export_put (exp);
2354 rs->rs_export = NULL;
2355 ptlrpc_rs_decref(rs);
2356 if (atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
2357 svc->srv_is_stopping)
2358 wake_up_all(&svcpt->scp_waitq);
2362 /* still on the net; callback will schedule */
2363 spin_unlock(&rs->rs_lock);
2369 ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
2371 int avail = svcpt->scp_nrqbds_posted;
2372 int low_water = test_req_buffer_pressure ? 0 :
2373 svcpt->scp_service->srv_nbuf_per_group / 2;
2375 /* NB I'm not locking; just looking. */
2377 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
2378 * allowed the request history to grow out of control. We could put a
2379 * sanity check on that here and cull some history if we need the
2382 if (avail <= low_water)
2383 ptlrpc_grow_req_bufs(svcpt, 1);
2385 if (svcpt->scp_service->srv_stats) {
2386 lprocfs_counter_add(svcpt->scp_service->srv_stats,
2387 PTLRPC_REQBUF_AVAIL_CNTR, avail);
2392 ptlrpc_retry_rqbds(void *arg)
2394 struct ptlrpc_service_part *svcpt = (struct ptlrpc_service_part *)arg;
2396 svcpt->scp_rqbd_timeout = 0;
2401 ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
2403 return svcpt->scp_nreqs_active <
2404 svcpt->scp_nthrs_running - 1 -
2405 (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL);
2409 * allowed to create more threads
2410 * user can call it w/o any lock but need to hold
2411 * ptlrpc_service_part::scp_lock to get reliable result
2414 ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
2416 return svcpt->scp_nthrs_running +
2417 svcpt->scp_nthrs_starting <
2418 svcpt->scp_service->srv_nthrs_cpt_limit;
2422 * too many requests and allowed to create more threads
2425 ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
2427 return !ptlrpc_threads_enough(svcpt) &&
2428 ptlrpc_threads_increasable(svcpt);
2432 ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2434 return thread_is_stopping(thread) ||
2435 thread->t_svcpt->scp_service->srv_is_stopping;
2439 ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
2441 return !list_empty(&svcpt->scp_rqbd_idle) &&
2442 svcpt->scp_rqbd_timeout == 0;
2446 ptlrpc_at_check(struct ptlrpc_service_part *svcpt)
2448 return svcpt->scp_at_check;
2452 * If a thread runs too long or spends to much time on a single request,
2453 * we want to know about it, so we set up a delayed work item as a watchdog.
2454 * If it fires, we display a stack trace of the delayed thread,
2455 * providing we aren't rate-limited
2457 * Watchdog stack traces are limited to 3 per 'libcfs_watchdog_ratelimit'
2460 static struct ratelimit_state watchdog_limit;
2462 static void ptlrpc_watchdog_fire(struct work_struct *w)
2464 struct ptlrpc_thread *thread = container_of(w, struct ptlrpc_thread,
2466 u64 ms_lapse = ktime_ms_delta(ktime_get(), thread->t_touched);
2467 u32 ms_frac = do_div(ms_lapse, MSEC_PER_SEC);
2469 if (!__ratelimit(&watchdog_limit)) {
2470 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",
2471 thread->t_task->comm, thread->t_task->pid,
2474 libcfs_debug_dumpstack(thread->t_task);
2476 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",
2477 thread->t_task->comm, thread->t_task->pid,
2478 ms_lapse, ms_frac, libcfs_watchdog_ratelimit);
2482 static void ptlrpc_watchdog_init(struct delayed_work *work, time_t time)
2484 INIT_DELAYED_WORK(work, ptlrpc_watchdog_fire);
2485 schedule_delayed_work(work, cfs_time_seconds(time));
2488 static void ptlrpc_watchdog_disable(struct delayed_work *work)
2490 cancel_delayed_work_sync(work);
2493 static void ptlrpc_watchdog_touch(struct delayed_work *work, time_t time)
2495 struct ptlrpc_thread *thread = container_of(&work->work,
2496 struct ptlrpc_thread,
2498 thread->t_touched = ktime_get();
2499 mod_delayed_work(system_wq, work, cfs_time_seconds(time));
2503 * requests wait on preprocessing
2504 * user can call it w/o any lock but need to hold
2505 * ptlrpc_service_part::scp_lock to get reliable result
2508 ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
2510 return !list_empty(&svcpt->scp_req_incoming);
2513 static __attribute__((__noinline__)) int
2514 ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
2515 struct ptlrpc_thread *thread)
2517 /* Don't exit while there are replies to be handled */
2518 struct l_wait_info lwi = LWI_TIMEOUT(svcpt->scp_rqbd_timeout,
2519 ptlrpc_retry_rqbds, svcpt);
2521 ptlrpc_watchdog_disable(&thread->t_watchdog);
2525 l_wait_event_exclusive_head(svcpt->scp_waitq,
2526 ptlrpc_thread_stopping(thread) ||
2527 ptlrpc_server_request_incoming(svcpt) ||
2528 ptlrpc_server_request_pending(svcpt, false) ||
2529 ptlrpc_rqbd_pending(svcpt) ||
2530 ptlrpc_at_check(svcpt), &lwi);
2532 if (ptlrpc_thread_stopping(thread))
2535 ptlrpc_watchdog_touch(&thread->t_watchdog,
2536 ptlrpc_server_get_timeout(svcpt));
2541 * Main thread body for service threads.
2542 * Waits in a loop waiting for new requests to process to appear.
2543 * Every time an incoming requests is added to its queue, a waitq
2544 * is woken up and one of the threads will handle it.
2546 static int ptlrpc_main(void *arg)
2548 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
2549 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2550 struct ptlrpc_service *svc = svcpt->scp_service;
2551 struct ptlrpc_reply_state *rs;
2552 struct group_info *ginfo = NULL;
2554 int counter = 0, rc = 0;
2557 thread->t_task = current;
2558 thread->t_pid = current_pid();
2559 unshare_fs_struct();
2561 if (svc->srv_cpt_bind) {
2562 rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
2564 CWARN("%s: failed to bind %s on CPT %d\n",
2565 svc->srv_name, thread->t_name, svcpt->scp_cpt);
2569 ginfo = groups_alloc(0);
2575 set_current_groups(ginfo);
2576 put_group_info(ginfo);
2578 if (svc->srv_ops.so_thr_init != NULL) {
2579 rc = svc->srv_ops.so_thr_init(thread);
2590 rc = lu_context_init(&env->le_ctx,
2591 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2595 thread->t_env = env;
2596 env->le_ctx.lc_thread = thread;
2597 env->le_ctx.lc_cookie = 0x6;
2599 while (!list_empty(&svcpt->scp_rqbd_idle)) {
2600 rc = ptlrpc_server_post_idle_rqbds(svcpt);
2604 CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
2605 svc->srv_name, svcpt->scp_cpt, rc);
2609 /* Alloc reply state structure for this one */
2610 OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2616 spin_lock(&svcpt->scp_lock);
2618 LASSERT(thread_is_starting(thread));
2619 thread_clear_flags(thread, SVC_STARTING);
2621 LASSERT(svcpt->scp_nthrs_starting == 1);
2622 svcpt->scp_nthrs_starting--;
2624 /* SVC_STOPPING may already be set here if someone else is trying
2625 * to stop the service while this new thread has been dynamically
2626 * forked. We still set SVC_RUNNING to let our creator know that
2627 * we are now running, however we will exit as soon as possible */
2628 thread_add_flags(thread, SVC_RUNNING);
2629 svcpt->scp_nthrs_running++;
2630 spin_unlock(&svcpt->scp_lock);
2632 /* wake up our creator in case he's still waiting. */
2633 wake_up(&thread->t_ctl_waitq);
2635 thread->t_touched = ktime_get();
2636 ptlrpc_watchdog_init(&thread->t_watchdog,
2637 ptlrpc_server_get_timeout(svcpt));
2639 spin_lock(&svcpt->scp_rep_lock);
2640 list_add(&rs->rs_list, &svcpt->scp_rep_idle);
2641 wake_up(&svcpt->scp_rep_waitq);
2642 spin_unlock(&svcpt->scp_rep_lock);
2644 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2645 svcpt->scp_nthrs_running);
2647 /* XXX maintain a list of all managed devices: insert here */
2648 while (!ptlrpc_thread_stopping(thread)) {
2649 if (ptlrpc_wait_event(svcpt, thread))
2652 ptlrpc_check_rqbd_pool(svcpt);
2654 if (ptlrpc_threads_need_create(svcpt)) {
2655 /* Ignore return code - we tried... */
2656 ptlrpc_start_thread(svcpt, 0);
2659 /* reset le_ses to initial state */
2661 /* Process all incoming reqs before handling any */
2662 if (ptlrpc_server_request_incoming(svcpt)) {
2663 lu_context_enter(&env->le_ctx);
2664 ptlrpc_server_handle_req_in(svcpt, thread);
2665 lu_context_exit(&env->le_ctx);
2667 /* but limit ourselves in case of flood */
2668 if (counter++ < 100)
2673 if (ptlrpc_at_check(svcpt))
2674 ptlrpc_at_check_timed(svcpt);
2676 if (ptlrpc_server_request_pending(svcpt, false)) {
2677 lu_context_enter(&env->le_ctx);
2678 ptlrpc_server_handle_request(svcpt, thread);
2679 lu_context_exit(&env->le_ctx);
2682 if (ptlrpc_rqbd_pending(svcpt) &&
2683 ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
2684 /* I just failed to repost request buffers.
2685 * Wait for a timeout (unless something else
2686 * happens) before I try again */
2687 svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
2688 CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
2689 svcpt->scp_nrqbds_posted);
2693 ptlrpc_watchdog_disable(&thread->t_watchdog);
2697 * deconstruct service specific state created by ptlrpc_start_thread()
2699 if (svc->srv_ops.so_thr_done != NULL)
2700 svc->srv_ops.so_thr_done(thread);
2703 lu_context_fini(&env->le_ctx);
2707 CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n",
2708 thread, thread->t_pid, thread->t_id, rc);
2710 spin_lock(&svcpt->scp_lock);
2711 if (thread_test_and_clear_flags(thread, SVC_STARTING))
2712 svcpt->scp_nthrs_starting--;
2714 if (thread_test_and_clear_flags(thread, SVC_RUNNING)) {
2715 /* must know immediately */
2716 svcpt->scp_nthrs_running--;
2720 thread_add_flags(thread, SVC_STOPPED);
2722 wake_up(&thread->t_ctl_waitq);
2723 spin_unlock(&svcpt->scp_lock);
2728 static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
2729 struct list_head *replies)
2733 spin_lock(&hrt->hrt_lock);
2735 list_splice_init(&hrt->hrt_queue, replies);
2736 result = ptlrpc_hr.hr_stopping || !list_empty(replies);
2738 spin_unlock(&hrt->hrt_lock);
2743 * Main body of "handle reply" function.
2744 * It processes acked reply states
2746 static int ptlrpc_hr_main(void *arg)
2748 struct ptlrpc_hr_thread *hrt = (struct ptlrpc_hr_thread *)arg;
2749 struct ptlrpc_hr_partition *hrp = hrt->hrt_partition;
2750 struct list_head replies;
2753 INIT_LIST_HEAD(&replies);
2754 unshare_fs_struct();
2756 rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
2758 char threadname[20];
2760 snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
2761 hrp->hrp_cpt, hrt->hrt_id);
2762 CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
2763 threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
2766 atomic_inc(&hrp->hrp_nstarted);
2767 wake_up(&ptlrpc_hr.hr_waitq);
2769 while (!ptlrpc_hr.hr_stopping) {
2770 l_wait_condition(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
2772 while (!list_empty(&replies)) {
2773 struct ptlrpc_reply_state *rs;
2775 rs = list_entry(replies.prev,
2776 struct ptlrpc_reply_state,
2778 list_del_init(&rs->rs_list);
2779 ptlrpc_handle_rs(rs);
2783 atomic_inc(&hrp->hrp_nstopped);
2784 wake_up(&ptlrpc_hr.hr_waitq);
2789 static void ptlrpc_stop_hr_threads(void)
2791 struct ptlrpc_hr_partition *hrp;
2795 ptlrpc_hr.hr_stopping = 1;
2797 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2798 if (hrp->hrp_thrs == NULL)
2799 continue; /* uninitialized */
2800 for (j = 0; j < hrp->hrp_nthrs; j++)
2801 wake_up_all(&hrp->hrp_thrs[j].hrt_waitq);
2804 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2805 if (hrp->hrp_thrs == NULL)
2806 continue; /* uninitialized */
2807 wait_event(ptlrpc_hr.hr_waitq,
2808 atomic_read(&hrp->hrp_nstopped) ==
2809 atomic_read(&hrp->hrp_nstarted));
2813 static int ptlrpc_start_hr_threads(void)
2815 struct ptlrpc_hr_partition *hrp;
2820 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2823 for (j = 0; j < hrp->hrp_nthrs; j++) {
2824 struct ptlrpc_hr_thread *hrt = &hrp->hrp_thrs[j];
2825 struct task_struct *task;
2827 task = kthread_run(ptlrpc_hr_main,
2829 "ptlrpc_hr%02d_%03d",
2838 wait_event(ptlrpc_hr.hr_waitq,
2839 atomic_read(&hrp->hrp_nstarted) == j);
2842 CERROR("cannot start reply handler thread %d:%d: "
2843 "rc = %d\n", i, j, rc);
2844 ptlrpc_stop_hr_threads();
2852 static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
2854 struct l_wait_info lwi = { 0 };
2855 struct ptlrpc_thread *thread;
2856 struct list_head zombie;
2860 CDEBUG(D_INFO, "Stopping threads for service %s\n",
2861 svcpt->scp_service->srv_name);
2863 INIT_LIST_HEAD(&zombie);
2864 spin_lock(&svcpt->scp_lock);
2865 /* let the thread know that we would like it to stop asap */
2866 list_for_each_entry(thread, &svcpt->scp_threads, t_link) {
2867 CDEBUG(D_INFO, "Stopping thread %s #%u\n",
2868 svcpt->scp_service->srv_thread_name, thread->t_id);
2869 thread_add_flags(thread, SVC_STOPPING);
2872 wake_up_all(&svcpt->scp_waitq);
2874 while (!list_empty(&svcpt->scp_threads)) {
2875 thread = list_entry(svcpt->scp_threads.next,
2876 struct ptlrpc_thread, t_link);
2877 if (thread_is_stopped(thread)) {
2878 list_del(&thread->t_link);
2879 list_add(&thread->t_link, &zombie);
2882 spin_unlock(&svcpt->scp_lock);
2884 CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
2885 svcpt->scp_service->srv_thread_name, thread->t_id);
2886 l_wait_event(thread->t_ctl_waitq,
2887 thread_is_stopped(thread), &lwi);
2889 spin_lock(&svcpt->scp_lock);
2892 spin_unlock(&svcpt->scp_lock);
2894 while (!list_empty(&zombie)) {
2895 thread = list_entry(zombie.next,
2896 struct ptlrpc_thread, t_link);
2897 list_del(&thread->t_link);
2898 OBD_FREE_PTR(thread);
2904 * Stops all threads of a particular service \a svc
2906 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2908 struct ptlrpc_service_part *svcpt;
2912 ptlrpc_service_for_each_part(svcpt, i, svc) {
2913 if (svcpt->scp_service != NULL)
2914 ptlrpc_svcpt_stop_threads(svcpt);
2920 int ptlrpc_start_threads(struct ptlrpc_service *svc)
2927 /* We require 2 threads min, see note in ptlrpc_server_handle_request */
2928 LASSERT(svc->srv_nthrs_cpt_init >= PTLRPC_NTHRS_INIT);
2930 for (i = 0; i < svc->srv_ncpts; i++) {
2931 for (j = 0; j < svc->srv_nthrs_cpt_init; j++) {
2932 rc = ptlrpc_start_thread(svc->srv_parts[i], 1);
2938 /* We have enough threads, don't start more. b=15759 */
2945 CERROR("cannot start %s thread #%d_%d: rc %d\n",
2946 svc->srv_thread_name, i, j, rc);
2947 ptlrpc_stop_all_threads(svc);
2951 int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
2953 struct l_wait_info lwi = { 0 };
2954 struct ptlrpc_thread *thread;
2955 struct ptlrpc_service *svc;
2956 struct task_struct *task;
2960 LASSERT(svcpt != NULL);
2962 svc = svcpt->scp_service;
2964 CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n",
2965 svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running,
2966 svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit);
2969 if (unlikely(svc->srv_is_stopping))
2972 if (!ptlrpc_threads_increasable(svcpt) ||
2973 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
2974 svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1))
2977 OBD_CPT_ALLOC_PTR(thread, svc->srv_cptable, svcpt->scp_cpt);
2980 init_waitqueue_head(&thread->t_ctl_waitq);
2982 spin_lock(&svcpt->scp_lock);
2983 if (!ptlrpc_threads_increasable(svcpt)) {
2984 spin_unlock(&svcpt->scp_lock);
2985 OBD_FREE_PTR(thread);
2989 if (svcpt->scp_nthrs_starting != 0) {
2990 /* serialize starting because some modules (obdfilter)
2991 * might require unique and contiguous t_id */
2992 LASSERT(svcpt->scp_nthrs_starting == 1);
2993 spin_unlock(&svcpt->scp_lock);
2994 OBD_FREE_PTR(thread);
2996 CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n",
2997 svc->srv_thread_name, svcpt->scp_thr_nextid);
3002 CDEBUG(D_INFO, "Creating thread %s #%d race, retry later\n",
3003 svc->srv_thread_name, svcpt->scp_thr_nextid);
3007 svcpt->scp_nthrs_starting++;
3008 thread->t_id = svcpt->scp_thr_nextid++;
3009 thread_add_flags(thread, SVC_STARTING);
3010 thread->t_svcpt = svcpt;
3012 list_add(&thread->t_link, &svcpt->scp_threads);
3013 spin_unlock(&svcpt->scp_lock);
3015 if (svcpt->scp_cpt >= 0) {
3016 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s%02d_%03d",
3017 svc->srv_thread_name, svcpt->scp_cpt, thread->t_id);
3019 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s_%04d",
3020 svc->srv_thread_name, thread->t_id);
3023 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name);
3024 task = kthread_run(ptlrpc_main, thread, "%s", thread->t_name);
3027 CERROR("cannot start thread '%s': rc = %d\n",
3028 thread->t_name, rc);
3029 spin_lock(&svcpt->scp_lock);
3030 --svcpt->scp_nthrs_starting;
3031 if (thread_is_stopping(thread)) {
3032 /* this ptlrpc_thread is being hanled
3033 * by ptlrpc_svcpt_stop_threads now
3035 thread_add_flags(thread, SVC_STOPPED);
3036 wake_up(&thread->t_ctl_waitq);
3037 spin_unlock(&svcpt->scp_lock);
3039 list_del(&thread->t_link);
3040 spin_unlock(&svcpt->scp_lock);
3041 OBD_FREE_PTR(thread);
3049 l_wait_event(thread->t_ctl_waitq,
3050 thread_is_running(thread) || thread_is_stopped(thread),
3053 rc = thread_is_stopped(thread) ? thread->t_id : 0;
3057 int ptlrpc_hr_init(void)
3059 struct ptlrpc_hr_partition *hrp;
3060 struct ptlrpc_hr_thread *hrt;
3067 memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
3068 ptlrpc_hr.hr_cpt_table = cfs_cpt_table;
3070 ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
3072 if (ptlrpc_hr.hr_partitions == NULL)
3075 ratelimit_state_init(&watchdog_limit,
3076 cfs_time_seconds(libcfs_watchdog_ratelimit), 3);
3078 init_waitqueue_head(&ptlrpc_hr.hr_waitq);
3080 weight = cpumask_weight(topology_sibling_cpumask(smp_processor_id()));
3082 cfs_percpt_for_each(hrp, cpt, ptlrpc_hr.hr_partitions) {
3085 atomic_set(&hrp->hrp_nstarted, 0);
3086 atomic_set(&hrp->hrp_nstopped, 0);
3088 hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, cpt);
3089 hrp->hrp_nthrs /= weight;
3090 if (hrp->hrp_nthrs == 0)
3093 OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, cpt,
3094 hrp->hrp_nthrs * sizeof(*hrt));
3095 if (hrp->hrp_thrs == NULL)
3096 GOTO(out, rc = -ENOMEM);
3098 for (i = 0; i < hrp->hrp_nthrs; i++) {
3099 hrt = &hrp->hrp_thrs[i];
3102 hrt->hrt_partition = hrp;
3103 init_waitqueue_head(&hrt->hrt_waitq);
3104 spin_lock_init(&hrt->hrt_lock);
3105 INIT_LIST_HEAD(&hrt->hrt_queue);
3109 rc = ptlrpc_start_hr_threads();
3116 void ptlrpc_hr_fini(void)
3118 struct ptlrpc_hr_partition *hrp;
3121 if (ptlrpc_hr.hr_partitions == NULL)
3124 ptlrpc_stop_hr_threads();
3126 cfs_percpt_for_each(hrp, cpt, ptlrpc_hr.hr_partitions) {
3127 if (hrp->hrp_thrs != NULL) {
3128 OBD_FREE(hrp->hrp_thrs,
3129 hrp->hrp_nthrs * sizeof(hrp->hrp_thrs[0]));
3133 cfs_percpt_free(ptlrpc_hr.hr_partitions);
3134 ptlrpc_hr.hr_partitions = NULL;
3139 * Wait until all already scheduled replies are processed.
3141 static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
3145 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
3148 rc = l_wait_event(svcpt->scp_waitq,
3149 atomic_read(&svcpt->scp_nreps_difficult) == 0, &lwi);
3152 CWARN("Unexpectedly long timeout %s %p\n",
3153 svcpt->scp_service->srv_name, svcpt->scp_service);
3158 ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
3160 struct ptlrpc_service_part *svcpt;
3163 /* early disarm AT timer... */
3164 ptlrpc_service_for_each_part(svcpt, i, svc) {
3165 if (svcpt->scp_service != NULL)
3166 del_timer(&svcpt->scp_at_timer);
3171 ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
3173 struct ptlrpc_service_part *svcpt;
3174 struct ptlrpc_request_buffer_desc *rqbd;
3175 struct l_wait_info lwi;
3179 /* All history will be culled when the next request buffer is
3180 * freed in ptlrpc_service_purge_all() */
3181 svc->srv_hist_nrqbds_cpt_max = 0;
3183 rc = LNetClearLazyPortal(svc->srv_req_portal);
3186 ptlrpc_service_for_each_part(svcpt, i, svc) {
3187 if (svcpt->scp_service == NULL)
3190 /* Unlink all the request buffers. This forces a 'final'
3191 * event with its 'unlink' flag set for each posted rqbd */
3192 list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
3194 rc = LNetMDUnlink(rqbd->rqbd_md_h);
3195 LASSERT(rc == 0 || rc == -ENOENT);
3199 ptlrpc_service_for_each_part(svcpt, i, svc) {
3200 if (svcpt->scp_service == NULL)
3203 /* Wait for the network to release any buffers
3204 * it's currently filling */
3205 spin_lock(&svcpt->scp_lock);
3206 while (svcpt->scp_nrqbds_posted != 0) {
3207 spin_unlock(&svcpt->scp_lock);
3208 /* Network access will complete in finite time but
3209 * the HUGE timeout lets us CWARN for visibility
3210 * of sluggish NALs */
3211 lwi = LWI_TIMEOUT_INTERVAL(
3212 cfs_time_seconds(LONG_UNLINK),
3213 cfs_time_seconds(1), NULL, NULL);
3214 rc = l_wait_event(svcpt->scp_waitq,
3215 svcpt->scp_nrqbds_posted == 0, &lwi);
3216 if (rc == -ETIMEDOUT) {
3217 CWARN("Service %s waiting for "
3218 "request buffers\n",
3219 svcpt->scp_service->srv_name);
3221 spin_lock(&svcpt->scp_lock);
3223 spin_unlock(&svcpt->scp_lock);
3228 ptlrpc_service_purge_all(struct ptlrpc_service *svc)
3230 struct ptlrpc_service_part *svcpt;
3231 struct ptlrpc_request_buffer_desc *rqbd;
3232 struct ptlrpc_request *req;
3233 struct ptlrpc_reply_state *rs;
3236 ptlrpc_service_for_each_part(svcpt, i, svc) {
3237 if (svcpt->scp_service == NULL)
3240 spin_lock(&svcpt->scp_rep_lock);
3241 while (!list_empty(&svcpt->scp_rep_active)) {
3242 rs = list_entry(svcpt->scp_rep_active.next,
3243 struct ptlrpc_reply_state, rs_list);
3244 spin_lock(&rs->rs_lock);
3245 ptlrpc_schedule_difficult_reply(rs);
3246 spin_unlock(&rs->rs_lock);
3248 spin_unlock(&svcpt->scp_rep_lock);
3250 /* purge the request queue. NB No new replies (rqbds
3251 * all unlinked) and no service threads, so I'm the only
3252 * thread noodling the request queue now */
3253 while (!list_empty(&svcpt->scp_req_incoming)) {
3254 req = list_entry(svcpt->scp_req_incoming.next,
3255 struct ptlrpc_request, rq_list);
3257 list_del(&req->rq_list);
3258 svcpt->scp_nreqs_incoming--;
3259 ptlrpc_server_finish_request(svcpt, req);
3262 while (ptlrpc_server_request_pending(svcpt, true)) {
3263 req = ptlrpc_server_request_get(svcpt, true);
3264 ptlrpc_server_finish_active_request(svcpt, req);
3267 LASSERT(list_empty(&svcpt->scp_rqbd_posted));
3268 LASSERT(svcpt->scp_nreqs_incoming == 0);
3269 LASSERT(svcpt->scp_nreqs_active == 0);
3270 /* history should have been culled by
3271 * ptlrpc_server_finish_request */
3272 LASSERT(svcpt->scp_hist_nrqbds == 0);
3274 /* Now free all the request buffers since nothing
3275 * references them any more... */
3277 while (!list_empty(&svcpt->scp_rqbd_idle)) {
3278 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
3279 struct ptlrpc_request_buffer_desc,
3281 ptlrpc_free_rqbd(rqbd);
3283 ptlrpc_wait_replies(svcpt);
3285 while (!list_empty(&svcpt->scp_rep_idle)) {
3286 rs = list_entry(svcpt->scp_rep_idle.next,
3287 struct ptlrpc_reply_state,
3289 list_del(&rs->rs_list);
3290 OBD_FREE_LARGE(rs, svc->srv_max_reply_size);
3296 ptlrpc_service_free(struct ptlrpc_service *svc)
3298 struct ptlrpc_service_part *svcpt;
3299 struct ptlrpc_at_array *array;
3302 ptlrpc_service_for_each_part(svcpt, i, svc) {
3303 if (svcpt->scp_service == NULL)
3306 /* In case somebody rearmed this in the meantime */
3307 del_timer(&svcpt->scp_at_timer);
3308 array = &svcpt->scp_at_array;
3310 if (array->paa_reqs_array != NULL) {
3311 OBD_FREE(array->paa_reqs_array,
3312 sizeof(struct list_head) * array->paa_size);
3313 array->paa_reqs_array = NULL;
3316 if (array->paa_reqs_count != NULL) {
3317 OBD_FREE(array->paa_reqs_count,
3318 sizeof(__u32) * array->paa_size);
3319 array->paa_reqs_count = NULL;
3323 ptlrpc_service_for_each_part(svcpt, i, svc)
3324 OBD_FREE_PTR(svcpt);
3326 if (svc->srv_cpts != NULL)
3327 cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts);
3329 OBD_FREE(svc, offsetof(struct ptlrpc_service,
3330 srv_parts[svc->srv_ncpts]));
3333 int ptlrpc_unregister_service(struct ptlrpc_service *service)
3337 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
3339 service->srv_is_stopping = 1;
3341 mutex_lock(&ptlrpc_all_services_mutex);
3342 list_del_init(&service->srv_list);
3343 mutex_unlock(&ptlrpc_all_services_mutex);
3345 ptlrpc_service_del_atimer(service);
3346 ptlrpc_stop_all_threads(service);
3348 ptlrpc_service_unlink_rqbd(service);
3349 ptlrpc_service_purge_all(service);
3350 ptlrpc_service_nrs_cleanup(service);
3352 ptlrpc_lprocfs_unregister_service(service);
3353 ptlrpc_sysfs_unregister_service(service);
3355 ptlrpc_service_free(service);
3359 EXPORT_SYMBOL(ptlrpc_unregister_service);
3362 * Returns 0 if the service is healthy.
3364 * Right now, it just checks to make sure that requests aren't languishing
3365 * in the queue. We'll use this health check to govern whether a node needs
3366 * to be shot, so it's intentionally non-aggressive. */
3367 static int ptlrpc_svcpt_health_check(struct ptlrpc_service_part *svcpt)
3369 struct ptlrpc_request *request = NULL;
3370 struct timespec64 right_now;
3371 struct timespec64 timediff;
3373 ktime_get_real_ts64(&right_now);
3375 spin_lock(&svcpt->scp_req_lock);
3376 /* How long has the next entry been waiting? */
3377 if (ptlrpc_server_high_pending(svcpt, true))
3378 request = ptlrpc_nrs_req_peek_nolock(svcpt, true);
3379 else if (ptlrpc_server_normal_pending(svcpt, true))
3380 request = ptlrpc_nrs_req_peek_nolock(svcpt, false);
3382 if (request == NULL) {
3383 spin_unlock(&svcpt->scp_req_lock);
3387 timediff = timespec64_sub(right_now, request->rq_arrival_time);
3388 spin_unlock(&svcpt->scp_req_lock);
3390 if ((timediff.tv_sec) >
3391 (AT_OFF ? obd_timeout * 3 / 2 : at_max)) {
3392 CERROR("%s: unhealthy - request has been waiting %llds\n",
3393 svcpt->scp_service->srv_name, (s64)timediff.tv_sec);
3401 ptlrpc_service_health_check(struct ptlrpc_service *svc)
3403 struct ptlrpc_service_part *svcpt;
3409 ptlrpc_service_for_each_part(svcpt, i, svc) {
3410 int rc = ptlrpc_svcpt_health_check(svcpt);
3417 EXPORT_SYMBOL(ptlrpc_service_health_check);