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"
44 #include <linux/delay.h>
46 /* The following are visible and mutable through /sys/module/ptlrpc */
47 int test_req_buffer_pressure = 0;
48 module_param(test_req_buffer_pressure, int, 0444);
49 MODULE_PARM_DESC(test_req_buffer_pressure, "set non-zero to put pressure on request buffer pools");
50 module_param(at_min, int, 0644);
51 MODULE_PARM_DESC(at_min, "Adaptive timeout minimum (sec)");
52 module_param(at_max, int, 0644);
53 MODULE_PARM_DESC(at_max, "Adaptive timeout maximum (sec)");
54 module_param(at_history, int, 0644);
55 MODULE_PARM_DESC(at_history,
56 "Adaptive timeouts remember the slowest event that took place within this period (sec)");
57 module_param(at_early_margin, int, 0644);
58 MODULE_PARM_DESC(at_early_margin, "How soon before an RPC deadline to send an early reply");
59 module_param(at_extra, int, 0644);
60 MODULE_PARM_DESC(at_extra, "How much extra time to give with each early reply");
63 static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt);
64 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req);
65 static void ptlrpc_at_remove_timed(struct ptlrpc_request *req);
66 static int ptlrpc_start_threads(struct ptlrpc_service *svc);
67 static int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait);
69 /** Holds a list of all PTLRPC services */
70 LIST_HEAD(ptlrpc_all_services);
71 /** Used to protect the \e ptlrpc_all_services list */
72 struct mutex ptlrpc_all_services_mutex;
74 static struct ptlrpc_request_buffer_desc *
75 ptlrpc_alloc_rqbd(struct ptlrpc_service_part *svcpt)
77 struct ptlrpc_service *svc = svcpt->scp_service;
78 struct ptlrpc_request_buffer_desc *rqbd;
80 OBD_CPT_ALLOC_PTR(rqbd, svc->srv_cptable, svcpt->scp_cpt);
84 rqbd->rqbd_svcpt = svcpt;
85 rqbd->rqbd_refcount = 0;
86 rqbd->rqbd_cbid.cbid_fn = request_in_callback;
87 rqbd->rqbd_cbid.cbid_arg = rqbd;
88 INIT_LIST_HEAD(&rqbd->rqbd_reqs);
89 OBD_CPT_ALLOC_LARGE(rqbd->rqbd_buffer, svc->srv_cptable,
90 svcpt->scp_cpt, svc->srv_buf_size);
91 if (rqbd->rqbd_buffer == NULL) {
96 spin_lock(&svcpt->scp_lock);
97 list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
98 svcpt->scp_nrqbds_total++;
99 spin_unlock(&svcpt->scp_lock);
104 static void ptlrpc_free_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
106 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
108 LASSERT(rqbd->rqbd_refcount == 0);
109 LASSERT(list_empty(&rqbd->rqbd_reqs));
111 spin_lock(&svcpt->scp_lock);
112 list_del(&rqbd->rqbd_list);
113 svcpt->scp_nrqbds_total--;
114 spin_unlock(&svcpt->scp_lock);
116 OBD_FREE_LARGE(rqbd->rqbd_buffer, svcpt->scp_service->srv_buf_size);
120 static int ptlrpc_grow_req_bufs(struct ptlrpc_service_part *svcpt, int post)
122 struct ptlrpc_service *svc = svcpt->scp_service;
123 struct ptlrpc_request_buffer_desc *rqbd;
127 if (svcpt->scp_rqbd_allocating)
130 spin_lock(&svcpt->scp_lock);
131 /* check again with lock */
132 if (svcpt->scp_rqbd_allocating) {
133 /* NB: we might allow more than one thread in the future */
134 LASSERT(svcpt->scp_rqbd_allocating == 1);
135 spin_unlock(&svcpt->scp_lock);
139 svcpt->scp_rqbd_allocating++;
140 spin_unlock(&svcpt->scp_lock);
143 for (i = 0; i < svc->srv_nbuf_per_group; i++) {
145 * NB: another thread might have recycled enough rqbds, we
146 * need to make sure it wouldn't over-allocate, see LU-1212.
148 if (svcpt->scp_nrqbds_posted >= svc->srv_nbuf_per_group ||
149 (svc->srv_nrqbds_max != 0 &&
150 svcpt->scp_nrqbds_total > svc->srv_nrqbds_max))
153 rqbd = ptlrpc_alloc_rqbd(svcpt);
156 CERROR("%s: Can't allocate request buffer\n",
163 spin_lock(&svcpt->scp_lock);
165 LASSERT(svcpt->scp_rqbd_allocating == 1);
166 svcpt->scp_rqbd_allocating--;
168 spin_unlock(&svcpt->scp_lock);
171 "%s: allocate %d new %d-byte reqbufs (%d/%d left), rc = %d\n",
172 svc->srv_name, i, svc->srv_buf_size, svcpt->scp_nrqbds_posted,
173 svcpt->scp_nrqbds_total, rc);
177 rc = ptlrpc_server_post_idle_rqbds(svcpt);
183 * Part of Rep-Ack logic.
184 * Puts a lock and its mode into reply state assotiated to request reply.
186 void ptlrpc_save_lock(struct ptlrpc_request *req, struct lustre_handle *lock,
187 int mode, bool no_ack, bool convert_lock)
189 struct ptlrpc_reply_state *rs = req->rq_reply_state;
193 LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
195 idx = rs->rs_nlocks++;
196 rs->rs_locks[idx] = *lock;
197 rs->rs_modes[idx] = mode;
198 rs->rs_difficult = 1;
199 rs->rs_no_ack = no_ack;
200 rs->rs_convert_lock = convert_lock;
202 EXPORT_SYMBOL(ptlrpc_save_lock);
205 struct ptlrpc_hr_partition;
207 struct ptlrpc_hr_thread {
208 int hrt_id; /* thread ID */
210 wait_queue_head_t hrt_waitq;
211 struct list_head hrt_queue;
212 struct ptlrpc_hr_partition *hrt_partition;
215 struct ptlrpc_hr_partition {
216 /* # of started threads */
217 atomic_t hrp_nstarted;
218 /* # of stopped threads */
219 atomic_t hrp_nstopped;
220 /* cpu partition id */
222 /* round-robin rotor for choosing thread */
224 /* total number of threads on this partition */
227 struct ptlrpc_hr_thread *hrp_thrs;
230 #define HRT_RUNNING 0
231 #define HRT_STOPPING 1
233 struct ptlrpc_hr_service {
234 /* CPU partition table, it's just cfs_cpt_tab for now */
235 struct cfs_cpt_table *hr_cpt_table;
236 /** controller sleep waitq */
237 wait_queue_head_t hr_waitq;
238 unsigned int hr_stopping;
239 /** roundrobin rotor for non-affinity service */
240 unsigned int hr_rotor;
242 struct ptlrpc_hr_partition **hr_partitions;
246 struct list_head rsb_replies;
247 unsigned int rsb_n_replies;
248 struct ptlrpc_service_part *rsb_svcpt;
251 /** reply handling service. */
252 static struct ptlrpc_hr_service ptlrpc_hr;
255 * maximum mumber of replies scheduled in one batch
257 #define MAX_SCHEDULED 256
260 * Initialize a reply batch.
264 static void rs_batch_init(struct rs_batch *b)
266 memset(b, 0, sizeof(*b));
267 INIT_LIST_HEAD(&b->rsb_replies);
271 * Choose an hr thread to dispatch requests to.
274 struct ptlrpc_hr_thread *ptlrpc_hr_select(struct ptlrpc_service_part *svcpt)
276 struct ptlrpc_hr_partition *hrp;
279 if (svcpt->scp_cpt >= 0 &&
280 svcpt->scp_service->srv_cptable == ptlrpc_hr.hr_cpt_table) {
281 /* directly match partition */
282 hrp = ptlrpc_hr.hr_partitions[svcpt->scp_cpt];
285 rotor = ptlrpc_hr.hr_rotor++;
286 rotor %= cfs_cpt_number(ptlrpc_hr.hr_cpt_table);
288 hrp = ptlrpc_hr.hr_partitions[rotor];
291 rotor = hrp->hrp_rotor++;
292 return &hrp->hrp_thrs[rotor % hrp->hrp_nthrs];
296 * Dispatch all replies accumulated in the batch to one from
297 * dedicated reply handling threads.
301 static void rs_batch_dispatch(struct rs_batch *b)
303 if (b->rsb_n_replies != 0) {
304 struct ptlrpc_hr_thread *hrt;
306 hrt = ptlrpc_hr_select(b->rsb_svcpt);
308 spin_lock(&hrt->hrt_lock);
309 list_splice_init(&b->rsb_replies, &hrt->hrt_queue);
310 spin_unlock(&hrt->hrt_lock);
312 wake_up(&hrt->hrt_waitq);
313 b->rsb_n_replies = 0;
318 * Add a reply to a batch.
319 * Add one reply object to a batch, schedule batched replies if overload.
324 static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs)
326 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
328 if (svcpt != b->rsb_svcpt || b->rsb_n_replies >= MAX_SCHEDULED) {
329 if (b->rsb_svcpt != NULL) {
330 rs_batch_dispatch(b);
331 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
333 spin_lock(&svcpt->scp_rep_lock);
334 b->rsb_svcpt = svcpt;
336 spin_lock(&rs->rs_lock);
337 rs->rs_scheduled_ever = 1;
338 if (rs->rs_scheduled == 0) {
339 list_move(&rs->rs_list, &b->rsb_replies);
340 rs->rs_scheduled = 1;
343 rs->rs_committed = 1;
344 spin_unlock(&rs->rs_lock);
348 * Reply batch finalization.
349 * Dispatch remaining replies from the batch
350 * and release remaining spinlock.
354 static void rs_batch_fini(struct rs_batch *b)
356 if (b->rsb_svcpt != NULL) {
357 rs_batch_dispatch(b);
358 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
362 #define DECLARE_RS_BATCH(b) struct rs_batch b
366 * Put reply state into a queue for processing because we received
367 * ACK from the client
369 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
371 struct ptlrpc_hr_thread *hrt;
375 LASSERT(list_empty(&rs->rs_list));
377 hrt = ptlrpc_hr_select(rs->rs_svcpt);
379 spin_lock(&hrt->hrt_lock);
380 list_add_tail(&rs->rs_list, &hrt->hrt_queue);
381 spin_unlock(&hrt->hrt_lock);
383 wake_up(&hrt->hrt_waitq);
387 void ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs)
391 assert_spin_locked(&rs->rs_svcpt->scp_rep_lock);
392 assert_spin_locked(&rs->rs_lock);
393 LASSERT(rs->rs_difficult);
394 rs->rs_scheduled_ever = 1; /* flag any notification attempt */
396 if (rs->rs_scheduled) { /* being set up or already notified */
401 rs->rs_scheduled = 1;
402 list_del_init(&rs->rs_list);
403 ptlrpc_dispatch_difficult_reply(rs);
406 EXPORT_SYMBOL(ptlrpc_schedule_difficult_reply);
408 void ptlrpc_commit_replies(struct obd_export *exp)
410 struct ptlrpc_reply_state *rs, *nxt;
411 DECLARE_RS_BATCH(batch);
415 rs_batch_init(&batch);
417 * Find any replies that have been committed and get their service
418 * to attend to complete them.
421 /* CAVEAT EMPTOR: spinlock ordering!!! */
422 spin_lock(&exp->exp_uncommitted_replies_lock);
423 list_for_each_entry_safe(rs, nxt, &exp->exp_uncommitted_replies,
425 LASSERT(rs->rs_difficult);
426 /* VBR: per-export last_committed */
427 LASSERT(rs->rs_export);
428 if (rs->rs_transno <= exp->exp_last_committed) {
429 list_del_init(&rs->rs_obd_list);
430 rs_batch_add(&batch, rs);
433 spin_unlock(&exp->exp_uncommitted_replies_lock);
434 rs_batch_fini(&batch);
438 static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt)
440 struct ptlrpc_request_buffer_desc *rqbd;
445 spin_lock(&svcpt->scp_lock);
447 if (list_empty(&svcpt->scp_rqbd_idle)) {
448 spin_unlock(&svcpt->scp_lock);
452 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
453 struct ptlrpc_request_buffer_desc,
456 /* assume we will post successfully */
457 svcpt->scp_nrqbds_posted++;
458 list_move(&rqbd->rqbd_list, &svcpt->scp_rqbd_posted);
460 spin_unlock(&svcpt->scp_lock);
462 rc = ptlrpc_register_rqbd(rqbd);
469 spin_lock(&svcpt->scp_lock);
471 svcpt->scp_nrqbds_posted--;
472 list_move_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
475 * Don't complain if no request buffers are posted right now; LNET
476 * won't drop requests because we set the portal lazy!
479 spin_unlock(&svcpt->scp_lock);
484 static void ptlrpc_at_timer(cfs_timer_cb_arg_t data)
486 struct ptlrpc_service_part *svcpt;
488 svcpt = cfs_from_timer(svcpt, data, scp_at_timer);
490 svcpt->scp_at_check = 1;
491 svcpt->scp_at_checktime = ktime_get();
492 wake_up(&svcpt->scp_waitq);
495 static void ptlrpc_server_nthreads_check(struct ptlrpc_service *svc,
496 struct ptlrpc_service_conf *conf)
498 struct ptlrpc_service_thr_conf *tc = &conf->psc_thr;
505 * Common code for estimating & validating threads number.
506 * CPT affinity service could have percpt thread-pool instead
507 * of a global thread-pool, which means user might not always
508 * get the threads number they give it in conf::tc_nthrs_user
509 * even they did set. It's because we need to validate threads
510 * number for each CPT to guarantee each pool will have enough
511 * threads to keep the service healthy.
513 init = PTLRPC_NTHRS_INIT + (svc->srv_ops.so_hpreq_handler != NULL);
514 init = max_t(int, init, tc->tc_nthrs_init);
517 * NB: please see comments in lustre_lnet.h for definition
518 * details of these members
520 LASSERT(tc->tc_nthrs_max != 0);
522 if (tc->tc_nthrs_user != 0) {
524 * In case there is a reason to test a service with many
525 * threads, we give a less strict check here, it can
526 * be up to 8 * nthrs_max
528 total = min(tc->tc_nthrs_max * 8, tc->tc_nthrs_user);
529 nthrs = total / svc->srv_ncpts;
530 init = max(init, nthrs);
534 total = tc->tc_nthrs_max;
535 if (tc->tc_nthrs_base == 0) {
537 * don't care about base threads number per partition,
538 * this is most for non-affinity service
540 nthrs = total / svc->srv_ncpts;
544 nthrs = tc->tc_nthrs_base;
545 if (svc->srv_ncpts == 1) {
549 * NB: Increase the base number if it's single partition
550 * and total number of cores/HTs is larger or equal to 4.
551 * result will always < 2 * nthrs_base
553 weight = cfs_cpt_weight(svc->srv_cptable, CFS_CPT_ANY);
554 for (i = 1; (weight >> (i + 1)) != 0 && /* >= 4 cores/HTs */
555 (tc->tc_nthrs_base >> i) != 0; i++)
556 nthrs += tc->tc_nthrs_base >> i;
559 if (tc->tc_thr_factor != 0) {
560 int factor = tc->tc_thr_factor;
564 * User wants to increase number of threads with for
565 * each CPU core/HT, most likely the factor is larger than
566 * one thread/core because service threads are supposed to
567 * be blocked by lock or wait for IO.
570 * Amdahl's law says that adding processors wouldn't give
571 * a linear increasing of parallelism, so it's nonsense to
572 * have too many threads no matter how many cores/HTs
577 (topology_sibling_cpumask(smp_processor_id())) > 1) {
578 /* weight is # of HTs */
579 /* depress thread factor for hyper-thread */
580 factor = factor - (factor >> 1) + (factor >> 3);
584 weight = cfs_cpt_weight(svc->srv_cptable, 0);
586 for (; factor > 0 && weight > 0; factor--, weight -= fade)
587 nthrs += min(weight, fade) * factor;
590 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
591 nthrs = max(tc->tc_nthrs_base,
592 tc->tc_nthrs_max / svc->srv_ncpts);
595 nthrs = max(nthrs, tc->tc_nthrs_init);
596 svc->srv_nthrs_cpt_limit = nthrs;
597 svc->srv_nthrs_cpt_init = init;
599 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
601 "%s: This service may have more threads (%d) than the given soft limit (%d)\n",
602 svc->srv_name, nthrs * svc->srv_ncpts,
608 * Initialize percpt data for a service
610 static int ptlrpc_service_part_init(struct ptlrpc_service *svc,
611 struct ptlrpc_service_part *svcpt, int cpt)
613 struct ptlrpc_at_array *array;
618 svcpt->scp_cpt = cpt;
619 INIT_LIST_HEAD(&svcpt->scp_threads);
621 /* rqbd and incoming request queue */
622 spin_lock_init(&svcpt->scp_lock);
623 mutex_init(&svcpt->scp_mutex);
624 INIT_LIST_HEAD(&svcpt->scp_rqbd_idle);
625 INIT_LIST_HEAD(&svcpt->scp_rqbd_posted);
626 INIT_LIST_HEAD(&svcpt->scp_req_incoming);
627 init_waitqueue_head(&svcpt->scp_waitq);
628 /* history request & rqbd list */
629 INIT_LIST_HEAD(&svcpt->scp_hist_reqs);
630 INIT_LIST_HEAD(&svcpt->scp_hist_rqbds);
632 /* acitve requests and hp requests */
633 spin_lock_init(&svcpt->scp_req_lock);
636 spin_lock_init(&svcpt->scp_rep_lock);
637 INIT_LIST_HEAD(&svcpt->scp_rep_active);
638 INIT_LIST_HEAD(&svcpt->scp_rep_idle);
639 init_waitqueue_head(&svcpt->scp_rep_waitq);
640 atomic_set(&svcpt->scp_nreps_difficult, 0);
642 /* adaptive timeout */
643 spin_lock_init(&svcpt->scp_at_lock);
644 array = &svcpt->scp_at_array;
646 size = at_est2timeout(at_max);
647 array->paa_size = size;
648 array->paa_count = 0;
649 array->paa_deadline = -1;
651 /* allocate memory for scp_at_array (ptlrpc_at_array) */
652 OBD_CPT_ALLOC(array->paa_reqs_array,
653 svc->srv_cptable, cpt, sizeof(struct list_head) * size);
654 if (array->paa_reqs_array == NULL)
657 for (index = 0; index < size; index++)
658 INIT_LIST_HEAD(&array->paa_reqs_array[index]);
660 OBD_CPT_ALLOC(array->paa_reqs_count,
661 svc->srv_cptable, cpt, sizeof(__u32) * size);
662 if (array->paa_reqs_count == NULL)
665 cfs_timer_setup(&svcpt->scp_at_timer, ptlrpc_at_timer,
666 (unsigned long)svcpt, 0);
669 * At SOW, service time should be quick; 10s seems generous. If client
670 * timeout is less than this, we'll be sending an early reply.
672 at_init(&svcpt->scp_at_estimate, 10, 0);
674 /* assign this before call ptlrpc_grow_req_bufs */
675 svcpt->scp_service = svc;
676 /* Now allocate the request buffers, but don't post them now */
677 rc = ptlrpc_grow_req_bufs(svcpt, 0);
679 * We shouldn't be under memory pressure at startup, so
680 * fail if we can't allocate all our buffers at this time.
688 if (array->paa_reqs_count != NULL) {
689 OBD_FREE_PTR_ARRAY(array->paa_reqs_count, size);
690 array->paa_reqs_count = NULL;
693 if (array->paa_reqs_array != NULL) {
694 OBD_FREE_PTR_ARRAY(array->paa_reqs_array, array->paa_size);
695 array->paa_reqs_array = NULL;
702 * Initialize service on a given portal.
703 * This includes starting serving threads , allocating and posting rqbds and
706 struct ptlrpc_service *ptlrpc_register_service(struct ptlrpc_service_conf *conf,
708 struct dentry *debugfs_entry)
710 struct ptlrpc_service_cpt_conf *cconf = &conf->psc_cpt;
711 struct ptlrpc_service *service;
712 struct ptlrpc_service_part *svcpt;
713 struct cfs_cpt_table *cptable;
722 LASSERT(conf->psc_buf.bc_nbufs > 0);
723 LASSERT(conf->psc_buf.bc_buf_size >=
724 conf->psc_buf.bc_req_max_size + SPTLRPC_MAX_PAYLOAD);
725 LASSERT(conf->psc_thr.tc_ctx_tags != 0);
727 cptable = cconf->cc_cptable;
729 cptable = cfs_cpt_tab;
731 if (conf->psc_thr.tc_cpu_bind > 1) {
732 CERROR("%s: Invalid cpu bind value %d, only 1 or 0 allowed\n",
733 conf->psc_name, conf->psc_thr.tc_cpu_bind);
734 RETURN(ERR_PTR(-EINVAL));
737 if (!cconf->cc_affinity) {
740 ncpts = cfs_cpt_number(cptable);
741 if (cconf->cc_pattern != NULL) {
742 struct cfs_expr_list *el;
744 rc = cfs_expr_list_parse(cconf->cc_pattern,
745 strlen(cconf->cc_pattern),
748 CERROR("%s: invalid CPT pattern string: %s\n",
749 conf->psc_name, cconf->cc_pattern);
750 RETURN(ERR_PTR(-EINVAL));
753 rc = cfs_expr_list_values(el, ncpts, &cpts);
754 cfs_expr_list_free(el);
756 CERROR("%s: failed to parse CPT array %s: %d\n",
757 conf->psc_name, cconf->cc_pattern, rc);
759 OBD_FREE_PTR_ARRAY(cpts, ncpts);
760 RETURN(ERR_PTR(rc < 0 ? rc : -EINVAL));
766 OBD_ALLOC(service, offsetof(struct ptlrpc_service, srv_parts[ncpts]));
767 if (service == NULL) {
769 OBD_FREE_PTR_ARRAY(cpts, ncpts);
770 RETURN(ERR_PTR(-ENOMEM));
773 service->srv_cptable = cptable;
774 service->srv_cpts = cpts;
775 service->srv_ncpts = ncpts;
776 service->srv_cpt_bind = conf->psc_thr.tc_cpu_bind;
778 service->srv_cpt_bits = 0; /* it's zero already, easy to read... */
779 while ((1 << service->srv_cpt_bits) < cfs_cpt_number(cptable))
780 service->srv_cpt_bits++;
783 spin_lock_init(&service->srv_lock);
784 service->srv_name = conf->psc_name;
785 service->srv_watchdog_factor = conf->psc_watchdog_factor;
786 INIT_LIST_HEAD(&service->srv_list); /* for safty of cleanup */
788 /* buffer configuration */
789 service->srv_nbuf_per_group = test_req_buffer_pressure ?
790 1 : conf->psc_buf.bc_nbufs;
791 /* do not limit max number of rqbds by default */
792 service->srv_nrqbds_max = 0;
794 service->srv_max_req_size = conf->psc_buf.bc_req_max_size +
796 service->srv_buf_size = conf->psc_buf.bc_buf_size;
797 service->srv_rep_portal = conf->psc_buf.bc_rep_portal;
798 service->srv_req_portal = conf->psc_buf.bc_req_portal;
800 /* With slab/alloc_pages buffer size will be rounded up to 2^n */
801 if (service->srv_buf_size & (service->srv_buf_size - 1)) {
802 int round = size_roundup_power2(service->srv_buf_size);
804 service->srv_buf_size = round;
807 /* Increase max reply size to next power of two */
808 service->srv_max_reply_size = 1;
809 while (service->srv_max_reply_size <
810 conf->psc_buf.bc_rep_max_size + SPTLRPC_MAX_PAYLOAD)
811 service->srv_max_reply_size <<= 1;
813 service->srv_thread_name = conf->psc_thr.tc_thr_name;
814 service->srv_ctx_tags = conf->psc_thr.tc_ctx_tags;
815 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
816 service->srv_ops = conf->psc_ops;
818 for (i = 0; i < ncpts; i++) {
819 if (!cconf->cc_affinity)
822 cpt = cpts != NULL ? cpts[i] : i;
824 OBD_CPT_ALLOC(svcpt, cptable, cpt, sizeof(*svcpt));
826 GOTO(failed, rc = -ENOMEM);
828 service->srv_parts[i] = svcpt;
829 rc = ptlrpc_service_part_init(service, svcpt, cpt);
834 ptlrpc_server_nthreads_check(service, conf);
836 rc = LNetSetLazyPortal(service->srv_req_portal);
839 mutex_lock(&ptlrpc_all_services_mutex);
840 list_add(&service->srv_list, &ptlrpc_all_services);
841 mutex_unlock(&ptlrpc_all_services_mutex);
844 rc = ptlrpc_sysfs_register_service(parent, service);
849 if (debugfs_entry != NULL)
850 ptlrpc_ldebugfs_register_service(debugfs_entry, service);
852 rc = ptlrpc_service_nrs_setup(service);
856 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
857 service->srv_name, service->srv_req_portal);
859 rc = ptlrpc_start_threads(service);
861 CERROR("Failed to start threads for service %s: %d\n",
862 service->srv_name, rc);
868 ptlrpc_unregister_service(service);
871 EXPORT_SYMBOL(ptlrpc_register_service);
874 * to actually free the request, must be called without holding svc_lock.
875 * note it's caller's responsibility to unlink req->rq_list.
877 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
879 LASSERT(atomic_read(&req->rq_refcount) == 0);
880 LASSERT(list_empty(&req->rq_timed_list));
883 * DEBUG_REQ() assumes the reply state of a request with a valid
884 * ref will not be destroyed until that reference is dropped.
886 ptlrpc_req_drop_rs(req);
888 sptlrpc_svc_ctx_decref(req);
890 if (req != &req->rq_rqbd->rqbd_req) {
892 * NB request buffers use an embedded
893 * req if the incoming req unlinked the
894 * MD; this isn't one of them!
896 ptlrpc_request_cache_free(req);
901 * drop a reference count of the request. if it reaches 0, we either
902 * put it into history list, or free it immediately.
904 void ptlrpc_server_drop_request(struct ptlrpc_request *req)
906 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
907 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
908 struct ptlrpc_service *svc = svcpt->scp_service;
911 if (!atomic_dec_and_test(&req->rq_refcount))
914 if (req->rq_session.lc_state == LCS_ENTERED) {
915 lu_context_exit(&req->rq_session);
916 lu_context_fini(&req->rq_session);
919 if (req->rq_at_linked) {
920 spin_lock(&svcpt->scp_at_lock);
922 * recheck with lock, in case it's unlinked by
923 * ptlrpc_at_check_timed()
925 if (likely(req->rq_at_linked))
926 ptlrpc_at_remove_timed(req);
927 spin_unlock(&svcpt->scp_at_lock);
930 LASSERT(list_empty(&req->rq_timed_list));
932 /* finalize request */
933 if (req->rq_export) {
934 class_export_put(req->rq_export);
935 req->rq_export = NULL;
938 spin_lock(&svcpt->scp_lock);
940 list_add(&req->rq_list, &rqbd->rqbd_reqs);
942 refcount = --(rqbd->rqbd_refcount);
944 /* request buffer is now idle: add to history */
945 list_move_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds);
946 svcpt->scp_hist_nrqbds++;
950 * I expect only about 1 or 2 rqbds need to be recycled here
952 while (svcpt->scp_hist_nrqbds > svc->srv_hist_nrqbds_cpt_max) {
953 rqbd = list_entry(svcpt->scp_hist_rqbds.next,
954 struct ptlrpc_request_buffer_desc,
957 list_del(&rqbd->rqbd_list);
958 svcpt->scp_hist_nrqbds--;
961 * remove rqbd's reqs from svc's req history while
962 * I've got the service lock
964 list_for_each_entry(req, &rqbd->rqbd_reqs, rq_list) {
965 /* Track the highest culled req seq */
966 if (req->rq_history_seq >
967 svcpt->scp_hist_seq_culled) {
968 svcpt->scp_hist_seq_culled =
971 list_del(&req->rq_history_list);
974 spin_unlock(&svcpt->scp_lock);
976 while ((req = list_first_entry_or_null(
978 struct ptlrpc_request, rq_list))) {
979 list_del(&req->rq_list);
980 ptlrpc_server_free_request(req);
983 spin_lock(&svcpt->scp_lock);
985 * now all reqs including the embedded req has been
986 * disposed, schedule request buffer for re-use
987 * or free it to drain some in excess.
989 LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) == 0);
990 if (svcpt->scp_nrqbds_posted >=
991 svc->srv_nbuf_per_group ||
992 (svc->srv_nrqbds_max != 0 &&
993 svcpt->scp_nrqbds_total > svc->srv_nrqbds_max) ||
994 test_req_buffer_pressure) {
995 /* like in ptlrpc_free_rqbd() */
996 svcpt->scp_nrqbds_total--;
997 OBD_FREE_LARGE(rqbd->rqbd_buffer,
1001 list_add_tail(&rqbd->rqbd_list,
1002 &svcpt->scp_rqbd_idle);
1006 spin_unlock(&svcpt->scp_lock);
1007 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
1008 /* If we are low on memory, we are not interested in history */
1009 list_del(&req->rq_list);
1010 list_del_init(&req->rq_history_list);
1012 /* Track the highest culled req seq */
1013 if (req->rq_history_seq > svcpt->scp_hist_seq_culled)
1014 svcpt->scp_hist_seq_culled = req->rq_history_seq;
1016 spin_unlock(&svcpt->scp_lock);
1018 ptlrpc_server_free_request(req);
1020 spin_unlock(&svcpt->scp_lock);
1024 static void ptlrpc_add_exp_list_nolock(struct ptlrpc_request *req,
1025 struct obd_export *export, bool hp)
1027 __u16 tag = lustre_msg_get_tag(req->rq_reqmsg);
1030 list_add(&req->rq_exp_list, &export->exp_hp_rpcs);
1032 list_add(&req->rq_exp_list, &export->exp_reg_rpcs);
1033 if (tag && export->exp_used_slots)
1034 set_bit(tag - 1, export->exp_used_slots);
1037 static void ptlrpc_del_exp_list(struct ptlrpc_request *req)
1039 __u16 tag = lustre_msg_get_tag(req->rq_reqmsg);
1041 spin_lock(&req->rq_export->exp_rpc_lock);
1042 list_del_init(&req->rq_exp_list);
1043 if (tag && !req->rq_obsolete && req->rq_export->exp_used_slots)
1044 clear_bit(tag - 1, req->rq_export->exp_used_slots);
1045 spin_unlock(&req->rq_export->exp_rpc_lock);
1048 /** Change request export and move hp request from old export to new */
1049 void ptlrpc_request_change_export(struct ptlrpc_request *req,
1050 struct obd_export *export)
1052 if (req->rq_export != NULL) {
1053 LASSERT(!list_empty(&req->rq_exp_list));
1054 /* remove rq_exp_list from last export */
1055 ptlrpc_del_exp_list(req);
1056 /* export has one reference already, so it's safe to
1057 * add req to export queue here and get another
1058 * reference for request later
1060 spin_lock(&export->exp_rpc_lock);
1061 ptlrpc_add_exp_list_nolock(req, export, req->rq_ops != NULL);
1062 spin_unlock(&export->exp_rpc_lock);
1064 class_export_rpc_dec(req->rq_export);
1065 class_export_put(req->rq_export);
1068 /* request takes one export refcount */
1069 req->rq_export = class_export_get(export);
1070 class_export_rpc_inc(export);
1074 * to finish a request: stop sending more early replies, and release
1077 static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt,
1078 struct ptlrpc_request *req)
1080 ptlrpc_server_hpreq_fini(req);
1082 ptlrpc_server_drop_request(req);
1086 * to finish an active request: stop sending more early replies, and release
1087 * the request. should be called after we finished handling the request.
1089 static void ptlrpc_server_finish_active_request(
1090 struct ptlrpc_service_part *svcpt,
1091 struct ptlrpc_request *req)
1093 spin_lock(&svcpt->scp_req_lock);
1094 ptlrpc_nrs_req_stop_nolock(req);
1095 svcpt->scp_nreqs_active--;
1097 svcpt->scp_nhreqs_active--;
1098 spin_unlock(&svcpt->scp_req_lock);
1100 ptlrpc_nrs_req_finalize(req);
1102 if (req->rq_export != NULL)
1103 class_export_rpc_dec(req->rq_export);
1105 ptlrpc_server_finish_request(svcpt, req);
1109 * This function makes sure dead exports are evicted in a timely manner.
1110 * This function is only called when some export receives a message (i.e.,
1111 * the network is up.)
1113 void ptlrpc_update_export_timer(struct obd_export *exp, time64_t extra_delay)
1115 struct obd_export *oldest_exp;
1116 time64_t oldest_time, new_time;
1123 * Compensate for slow machines, etc, by faking our request time
1124 * into the future. Although this can break the strict time-ordering
1125 * of the list, we can be really lazy here - we don't have to evict
1126 * at the exact right moment. Eventually, all silent exports
1127 * will make it to the top of the list.
1130 /* Do not pay attention on 1sec or smaller renewals. */
1131 new_time = ktime_get_real_seconds() + extra_delay;
1132 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
1135 exp->exp_last_request_time = new_time;
1138 * exports may get disconnected from the chain even though the
1139 * export has references, so we must keep the spin lock while
1140 * manipulating the lists
1142 spin_lock(&exp->exp_obd->obd_dev_lock);
1144 if (list_empty(&exp->exp_obd_chain_timed)) {
1145 /* this one is not timed */
1146 spin_unlock(&exp->exp_obd->obd_dev_lock);
1150 list_move_tail(&exp->exp_obd_chain_timed,
1151 &exp->exp_obd->obd_exports_timed);
1153 oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
1154 struct obd_export, exp_obd_chain_timed);
1155 oldest_time = oldest_exp->exp_last_request_time;
1156 spin_unlock(&exp->exp_obd->obd_dev_lock);
1158 if (exp->exp_obd->obd_recovering) {
1159 /* be nice to everyone during recovery */
1164 /* Note - racing to start/reset the obd_eviction timer is safe */
1165 if (exp->exp_obd->obd_eviction_timer == 0) {
1166 /* Check if the oldest entry is expired. */
1167 if (ktime_get_real_seconds() >
1168 oldest_time + PING_EVICT_TIMEOUT + extra_delay) {
1170 * We need a second timer, in case the net was down and
1171 * it just came back. Since the pinger may skip every
1172 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
1173 * we better wait for 3.
1175 exp->exp_obd->obd_eviction_timer =
1176 ktime_get_real_seconds() + 3 * PING_INTERVAL;
1177 CDEBUG(D_HA, "%s: Think about evicting %s from %lld\n",
1178 exp->exp_obd->obd_name,
1179 obd_export_nid2str(oldest_exp), oldest_time);
1182 if (ktime_get_real_seconds() >
1183 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
1185 * The evictor won't evict anyone who we've heard from
1186 * recently, so we don't have to check before we start
1189 if (!ping_evictor_wake(exp))
1190 exp->exp_obd->obd_eviction_timer = 0;
1198 * Sanity check request \a req.
1199 * Return 0 if all is ok, error code otherwise.
1201 static int ptlrpc_check_req(struct ptlrpc_request *req)
1203 struct obd_device *obd = req->rq_export->exp_obd;
1206 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
1207 req->rq_export->exp_conn_cnt)) {
1208 DEBUG_REQ(D_RPCTRACE, req,
1209 "DROPPING req from old connection %d < %d",
1210 lustre_msg_get_conn_cnt(req->rq_reqmsg),
1211 req->rq_export->exp_conn_cnt);
1214 if (unlikely(obd == NULL || obd->obd_fail)) {
1216 * Failing over, don't handle any more reqs,
1217 * send error response instead.
1219 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
1220 req, (obd != NULL) ? obd->obd_name : "unknown");
1222 } else if (lustre_msg_get_flags(req->rq_reqmsg) &
1223 (MSG_REPLAY | MSG_REQ_REPLAY_DONE) &&
1224 !obd->obd_recovering) {
1225 DEBUG_REQ(D_ERROR, req,
1226 "Invalid replay without recovery");
1227 class_fail_export(req->rq_export);
1229 } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 &&
1230 !obd->obd_recovering) {
1231 DEBUG_REQ(D_ERROR, req,
1232 "Invalid req with transno %llu without recovery",
1233 lustre_msg_get_transno(req->rq_reqmsg));
1234 class_fail_export(req->rq_export);
1238 if (unlikely(rc < 0)) {
1239 req->rq_status = rc;
1245 static void ptlrpc_at_set_timer(struct ptlrpc_service_part *svcpt)
1247 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1250 if (array->paa_count == 0) {
1251 del_timer(&svcpt->scp_at_timer);
1255 /* Set timer for closest deadline */
1256 next = array->paa_deadline - ktime_get_real_seconds() -
1259 ptlrpc_at_timer(cfs_timer_cb_arg(svcpt, scp_at_timer));
1261 mod_timer(&svcpt->scp_at_timer,
1262 jiffies + nsecs_to_jiffies(next * NSEC_PER_SEC));
1263 CDEBUG(D_INFO, "armed %s at %+llds\n",
1264 svcpt->scp_service->srv_name, next);
1268 /* Add rpc to early reply check list */
1269 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
1271 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1272 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1273 struct ptlrpc_request *rq = NULL;
1279 if (req->rq_no_reply)
1282 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
1285 spin_lock(&svcpt->scp_at_lock);
1286 LASSERT(list_empty(&req->rq_timed_list));
1288 div_u64_rem(req->rq_deadline, array->paa_size, &index);
1289 if (array->paa_reqs_count[index] > 0) {
1291 * latest rpcs will have the latest deadlines in the list,
1292 * so search backward.
1294 list_for_each_entry_reverse(rq, &array->paa_reqs_array[index],
1296 if (req->rq_deadline >= rq->rq_deadline) {
1297 list_add(&req->rq_timed_list,
1298 &rq->rq_timed_list);
1304 /* Add the request at the head of the list */
1305 if (list_empty(&req->rq_timed_list))
1306 list_add(&req->rq_timed_list, &array->paa_reqs_array[index]);
1308 spin_lock(&req->rq_lock);
1309 req->rq_at_linked = 1;
1310 spin_unlock(&req->rq_lock);
1311 req->rq_at_index = index;
1312 array->paa_reqs_count[index]++;
1314 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
1315 array->paa_deadline = req->rq_deadline;
1316 ptlrpc_at_set_timer(svcpt);
1318 spin_unlock(&svcpt->scp_at_lock);
1323 static void ptlrpc_at_remove_timed(struct ptlrpc_request *req)
1325 struct ptlrpc_at_array *array;
1327 array = &req->rq_rqbd->rqbd_svcpt->scp_at_array;
1329 /* NB: must call with hold svcpt::scp_at_lock */
1330 LASSERT(!list_empty(&req->rq_timed_list));
1331 list_del_init(&req->rq_timed_list);
1333 spin_lock(&req->rq_lock);
1334 req->rq_at_linked = 0;
1335 spin_unlock(&req->rq_lock);
1337 array->paa_reqs_count[req->rq_at_index]--;
1342 * Attempt to extend the request deadline by sending an early reply to the
1345 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
1347 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1348 struct ptlrpc_request *reqcopy;
1349 struct lustre_msg *reqmsg;
1350 timeout_t olddl = req->rq_deadline - ktime_get_real_seconds();
1356 if (CFS_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_RECONNECT)) {
1357 /* don't send early reply */
1362 * deadline is when the client expects us to reply, margin is the
1363 * difference between clients' and servers' expectations
1365 DEBUG_REQ(D_ADAPTTO, req,
1366 "%ssending early reply (deadline %+ds, margin %+ds) for %d+%d",
1367 AT_OFF ? "AT off - not " : "",
1368 olddl, olddl - at_get(&svcpt->scp_at_estimate),
1369 at_get(&svcpt->scp_at_estimate), at_extra);
1375 /* below message is checked in replay-ost-single.sh test_9 */
1376 DEBUG_REQ(D_WARNING, req,
1377 "Already past deadline (%+ds), not sending early reply. Consider increasing at_early_margin (%d)?",
1378 olddl, at_early_margin);
1380 /* Return an error so we're not re-added to the timed list. */
1384 if ((lustre_msghdr_get_flags(req->rq_reqmsg) &
1385 MSGHDR_AT_SUPPORT) == 0) {
1386 DEBUG_REQ(D_INFO, req,
1387 "Wanted to ask client for more time, but no AT support");
1391 if (req->rq_export &&
1392 lustre_msg_get_flags(req->rq_reqmsg) &
1393 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1394 struct obd_device *obd_exp = req->rq_export->exp_obd;
1397 * During recovery, we don't want to send too many early
1398 * replies, but on the other hand we want to make sure the
1399 * client has enough time to resend if the rpc is lost. So
1400 * during the recovery period send at least 4 early replies,
1401 * spacing them every at_extra if we can. at_estimate should
1402 * always equal this fixed value during recovery.
1406 * Don't account request processing time into AT history
1407 * during recovery, it is not service time we need but
1408 * includes also waiting time for recovering clients
1410 newdl = min_t(time64_t, at_extra,
1411 obd_exp->obd_recovery_timeout / 4) +
1412 ktime_get_real_seconds();
1415 * We want to extend the request deadline by at_extra seconds,
1416 * so we set our service estimate to reflect how much time has
1417 * passed since this request arrived plus an additional
1418 * at_extra seconds. The client will calculate the new deadline
1419 * based on this service estimate (plus some additional time to
1420 * account for network latency). See ptlrpc_at_recv_early_reply
1422 at_measured(&svcpt->scp_at_estimate, at_extra +
1423 ktime_get_real_seconds() -
1424 req->rq_arrival_time.tv_sec);
1425 newdl = req->rq_arrival_time.tv_sec +
1426 at_get(&svcpt->scp_at_estimate);
1430 * Check to see if we've actually increased the deadline -
1431 * we may be past adaptive_max
1433 if (req->rq_deadline >= newdl) {
1434 DEBUG_REQ(D_WARNING, req,
1435 "Could not add any time (%d/%lld), not sending early reply",
1436 olddl, newdl - ktime_get_real_seconds());
1440 reqcopy = ptlrpc_request_cache_alloc(GFP_NOFS);
1441 if (reqcopy == NULL)
1443 OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1445 GOTO(out_free, rc = -ENOMEM);
1448 reqcopy->rq_reply_state = NULL;
1449 reqcopy->rq_rep_swab_mask = 0;
1450 reqcopy->rq_pack_bulk = 0;
1451 reqcopy->rq_pack_udesc = 0;
1452 reqcopy->rq_packed_final = 0;
1453 sptlrpc_svc_ctx_addref(reqcopy);
1454 /* We only need the reqmsg for the magic */
1455 reqcopy->rq_reqmsg = reqmsg;
1456 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1459 * tgt_brw_read() and tgt_brw_write() may have decided not to reply.
1460 * Without this check, we would fail the rq_no_reply assertion in
1461 * ptlrpc_send_reply().
1463 if (reqcopy->rq_no_reply)
1464 GOTO(out, rc = -ETIMEDOUT);
1466 LASSERT(atomic_read(&req->rq_refcount));
1467 /* if it is last refcount then early reply isn't needed */
1468 if (atomic_read(&req->rq_refcount) == 1) {
1469 DEBUG_REQ(D_ADAPTTO, reqcopy,
1470 "Normal reply already sent, abort early reply");
1471 GOTO(out, rc = -EINVAL);
1474 /* Connection ref */
1475 reqcopy->rq_export = class_conn2export(
1476 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1477 if (reqcopy->rq_export == NULL)
1478 GOTO(out, rc = -ENODEV);
1481 class_export_rpc_inc(reqcopy->rq_export);
1482 if (reqcopy->rq_export->exp_obd &&
1483 reqcopy->rq_export->exp_obd->obd_fail)
1484 GOTO(out_put, rc = -ENODEV);
1486 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1490 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1493 /* Adjust our own deadline to what we told the client */
1494 req->rq_deadline = newdl;
1495 req->rq_early_count++; /* number sent, server side */
1497 DEBUG_REQ(D_ERROR, req, "Early reply send failed: rc = %d", rc);
1501 * Free the (early) reply state from lustre_pack_reply.
1502 * (ptlrpc_send_reply takes it's own rs ref, so this is safe here)
1504 ptlrpc_req_drop_rs(reqcopy);
1507 class_export_rpc_dec(reqcopy->rq_export);
1508 class_export_put(reqcopy->rq_export);
1510 sptlrpc_svc_ctx_decref(reqcopy);
1511 OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1513 ptlrpc_request_cache_free(reqcopy);
1518 * Send early replies to everybody expiring within at_early_margin
1519 * asking for at_extra time
1521 static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1523 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1524 struct ptlrpc_request *rq, *n;
1525 LIST_HEAD(work_list);
1528 time64_t now = ktime_get_real_seconds();
1530 int first, counter = 0;
1533 spin_lock(&svcpt->scp_at_lock);
1534 if (svcpt->scp_at_check == 0) {
1535 spin_unlock(&svcpt->scp_at_lock);
1538 delay_ms = ktime_ms_delta(ktime_get(), svcpt->scp_at_checktime);
1539 svcpt->scp_at_check = 0;
1541 if (array->paa_count == 0) {
1542 spin_unlock(&svcpt->scp_at_lock);
1546 /* The timer went off, but maybe the nearest rpc already completed. */
1547 first = array->paa_deadline - now;
1548 if (first > at_early_margin) {
1549 /* We've still got plenty of time. Reset the timer. */
1550 ptlrpc_at_set_timer(svcpt);
1551 spin_unlock(&svcpt->scp_at_lock);
1556 * We're close to a timeout, and we don't know how much longer the
1557 * server will take. Send early replies to everyone expiring soon.
1560 div_u64_rem(array->paa_deadline, array->paa_size, &index);
1561 count = array->paa_count;
1563 count -= array->paa_reqs_count[index];
1564 list_for_each_entry_safe(rq, n,
1565 &array->paa_reqs_array[index],
1567 if (rq->rq_deadline > now + at_early_margin) {
1568 /* update the earliest deadline */
1569 if (deadline == -1 ||
1570 rq->rq_deadline < deadline)
1571 deadline = rq->rq_deadline;
1576 * ptlrpc_server_drop_request() may drop
1577 * refcount to 0 already. Let's check this and
1578 * don't add entry to work_list
1580 if (likely(atomic_inc_not_zero(&rq->rq_refcount))) {
1581 ptlrpc_at_remove_timed(rq);
1582 list_add(&rq->rq_timed_list, &work_list);
1584 ptlrpc_at_remove_timed(rq);
1590 if (++index >= array->paa_size)
1593 array->paa_deadline = deadline;
1594 /* we have a new earliest deadline, restart the timer */
1595 ptlrpc_at_set_timer(svcpt);
1597 spin_unlock(&svcpt->scp_at_lock);
1600 "timeout in %+ds, asking for %d secs on %d early replies\n",
1601 first, at_extra, counter);
1604 * We're already past request deadlines before we even get a
1605 * chance to send early replies
1607 LCONSOLE_WARN("%s: This server is not able to keep up with request traffic (cpu-bound).\n",
1608 svcpt->scp_service->srv_name);
1609 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, delay=%lldms\n",
1610 counter, svcpt->scp_nreqs_incoming,
1611 svcpt->scp_nreqs_active,
1612 at_get(&svcpt->scp_at_estimate), delay_ms);
1616 * we took additional refcount so entries can't be deleted from list, no
1619 while (!list_empty(&work_list)) {
1620 rq = list_entry(work_list.next, struct ptlrpc_request,
1622 list_del_init(&rq->rq_timed_list);
1624 if (ptlrpc_at_send_early_reply(rq) == 0)
1625 ptlrpc_at_add_timed(rq);
1627 ptlrpc_server_drop_request(rq);
1630 RETURN(1); /* return "did_something" for liblustre */
1634 * Check if we are already handling earlier incarnation of this request.
1635 * Called under &req->rq_export->exp_rpc_lock locked
1637 static struct ptlrpc_request*
1638 ptlrpc_server_check_resend_in_progress(struct ptlrpc_request *req)
1640 struct ptlrpc_request *tmp = NULL;
1642 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ||
1643 (atomic_read(&req->rq_export->exp_rpc_count) == 0))
1647 * bulk request are aborted upon reconnect, don't try to
1650 if (req->rq_bulk_write || req->rq_bulk_read)
1654 * This list should not be longer than max_requests in
1655 * flights on the client, so it is not all that long.
1656 * Also we only hit this codepath in case of a resent
1657 * request which makes it even more rarely hit
1659 list_for_each_entry(tmp, &req->rq_export->exp_reg_rpcs,
1661 /* Found duplicate one */
1662 if (tmp->rq_xid == req->rq_xid)
1665 list_for_each_entry(tmp, &req->rq_export->exp_hp_rpcs,
1667 /* Found duplicate one */
1668 if (tmp->rq_xid == req->rq_xid)
1674 DEBUG_REQ(D_HA, req, "Found duplicate req in processing");
1675 DEBUG_REQ(D_HA, tmp, "Request being processed");
1679 #ifdef HAVE_SERVER_SUPPORT
1680 static void ptlrpc_server_mark_obsolete(struct ptlrpc_request *req)
1682 req->rq_obsolete = 1;
1686 ptlrpc_server_mark_in_progress_obsolete(struct ptlrpc_request *req)
1688 struct ptlrpc_request *tmp = NULL;
1691 if (!tgt_is_increasing_xid_client(req->rq_export) ||
1692 req->rq_export->exp_used_slots == NULL)
1695 tag = lustre_msg_get_tag(req->rq_reqmsg);
1699 if (!test_bit(tag - 1, req->rq_export->exp_used_slots))
1702 /* This list should not be longer than max_requests in
1703 * flights on the client, so it is not all that long.
1704 * Also we only hit this codepath in case of a resent
1705 * request which makes it even more rarely hit */
1706 list_for_each_entry(tmp, &req->rq_export->exp_reg_rpcs, rq_exp_list) {
1707 if (tag == lustre_msg_get_tag(tmp->rq_reqmsg) &&
1708 req->rq_xid > tmp->rq_xid)
1709 ptlrpc_server_mark_obsolete(tmp);
1712 list_for_each_entry(tmp, &req->rq_export->exp_hp_rpcs, rq_exp_list) {
1713 if (tag == lustre_msg_get_tag(tmp->rq_reqmsg) &&
1714 req->rq_xid > tmp->rq_xid)
1715 ptlrpc_server_mark_obsolete(tmp);
1721 * Check if a request should be assigned with a high priority.
1723 * \retval < 0: error occurred
1724 * 0: normal RPC request
1725 * +1: high priority request
1727 static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt,
1728 struct ptlrpc_request *req)
1733 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL) {
1734 rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
1741 if (req->rq_export != NULL && req->rq_ops != NULL) {
1743 * Perform request specific check. We should do this
1744 * check before the request is added into exp_hp_rpcs
1745 * list otherwise it may hit swab race at LU-1044.
1747 if (req->rq_ops->hpreq_check != NULL) {
1748 rc = req->rq_ops->hpreq_check(req);
1749 if (rc == -ESTALE) {
1750 req->rq_status = rc;
1754 * can only return error,
1755 * 0 for normal request,
1756 * or 1 for high priority request
1765 /** Remove the request from the export list. */
1766 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req)
1769 if (req->rq_export) {
1771 * refresh lock timeout again so that client has more
1772 * room to send lock cancel RPC.
1774 if (req->rq_ops && req->rq_ops->hpreq_fini)
1775 req->rq_ops->hpreq_fini(req);
1777 ptlrpc_del_exp_list(req);
1782 static int ptlrpc_hpreq_check(struct ptlrpc_request *req)
1787 static struct ptlrpc_hpreq_ops ptlrpc_hpreq_common = {
1788 .hpreq_check = ptlrpc_hpreq_check,
1791 /* Hi-Priority RPC check by RPC operation code. */
1792 int ptlrpc_hpreq_handler(struct ptlrpc_request *req)
1794 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1797 * Check for export to let only reconnects for not yet evicted
1798 * export to become a HP rpc.
1800 if ((req->rq_export != NULL) &&
1801 (opc == OBD_PING || opc == MDS_CONNECT || opc == OST_CONNECT))
1802 req->rq_ops = &ptlrpc_hpreq_common;
1806 EXPORT_SYMBOL(ptlrpc_hpreq_handler);
1808 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1809 struct ptlrpc_request *req)
1813 struct ptlrpc_request *orig;
1817 rc = ptlrpc_server_hpreq_init(svcpt, req);
1822 ptlrpc_nrs_req_initialize(svcpt, req, hp);
1824 while (req->rq_export != NULL) {
1825 struct obd_export *exp = req->rq_export;
1828 * do search for duplicated xid and the adding to the list
1831 spin_lock_bh(&exp->exp_rpc_lock);
1832 #ifdef HAVE_SERVER_SUPPORT
1833 ptlrpc_server_mark_in_progress_obsolete(req);
1835 orig = ptlrpc_server_check_resend_in_progress(req);
1836 if (orig && OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_RESEND_RACE)) {
1837 spin_unlock_bh(&exp->exp_rpc_lock);
1839 OBD_RACE(OBD_FAIL_PTLRPC_RESEND_RACE);
1840 msleep(4 * MSEC_PER_SEC);
1844 if (orig && likely(atomic_inc_not_zero(&orig->rq_refcount))) {
1847 spin_unlock_bh(&exp->exp_rpc_lock);
1850 * When the client resend request and the server has
1851 * the previous copy of it, we need to update deadlines,
1852 * to be sure that the client and the server have equal
1853 * request deadlines.
1856 spin_lock(&orig->rq_rqbd->rqbd_svcpt->scp_at_lock);
1857 linked = orig->rq_at_linked;
1859 ptlrpc_at_remove_timed(orig);
1860 spin_unlock(&orig->rq_rqbd->rqbd_svcpt->scp_at_lock);
1861 orig->rq_deadline = req->rq_deadline;
1863 ptlrpc_at_add_timed(orig);
1864 ptlrpc_server_drop_request(orig);
1865 ptlrpc_nrs_req_finalize(req);
1867 /* don't mark slot unused for resend in progress */
1868 req->rq_obsolete = 1;
1873 ptlrpc_add_exp_list_nolock(req, exp, hp || req->rq_ops != NULL);
1875 spin_unlock_bh(&exp->exp_rpc_lock);
1880 * the current thread is not the processing thread for this request
1881 * since that, but request is in exp_hp_list and can be find there.
1882 * Remove all relations between request and old thread.
1884 req->rq_svc_thread->t_env->le_ses = NULL;
1885 req->rq_svc_thread = NULL;
1886 req->rq_session.lc_thread = NULL;
1888 ptlrpc_nrs_req_add(svcpt, req, hp);
1894 * Allow to handle high priority request
1895 * User can call it w/o any lock but need to hold
1896 * ptlrpc_service_part::scp_req_lock to get reliable result
1898 static bool ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1901 int running = svcpt->scp_nthrs_running;
1903 if (!nrs_svcpt_has_hp(svcpt))
1909 if (ptlrpc_nrs_req_throttling_nolock(svcpt, true))
1912 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1913 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1914 /* leave just 1 thread for normal RPCs */
1915 running = PTLRPC_NTHRS_INIT;
1916 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1920 if (svcpt->scp_nreqs_active >= running - 1)
1923 if (svcpt->scp_nhreqs_active == 0)
1926 return !ptlrpc_nrs_req_pending_nolock(svcpt, false) ||
1927 svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1930 static bool ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1933 return ptlrpc_server_allow_high(svcpt, force) &&
1934 ptlrpc_nrs_req_pending_nolock(svcpt, true);
1938 * Only allow normal priority requests on a service that has a high-priority
1939 * queue if forced (i.e. cleanup), if there are other high priority requests
1940 * already being processed (i.e. those threads can service more high-priority
1941 * requests), or if there are enough idle threads that a later thread can do
1942 * a high priority request.
1943 * User can call it w/o any lock but need to hold
1944 * ptlrpc_service_part::scp_req_lock to get reliable result
1946 static bool ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1949 int running = svcpt->scp_nthrs_running;
1951 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1952 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1953 /* leave just 1 thread for normal RPCs */
1954 running = PTLRPC_NTHRS_INIT;
1955 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1962 if (ptlrpc_nrs_req_throttling_nolock(svcpt, false))
1965 if (svcpt->scp_nreqs_active < running - 2)
1968 if (svcpt->scp_nreqs_active >= running - 1)
1971 return svcpt->scp_nhreqs_active > 0 || !nrs_svcpt_has_hp(svcpt);
1974 static bool ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1977 return ptlrpc_server_allow_normal(svcpt, force) &&
1978 ptlrpc_nrs_req_pending_nolock(svcpt, false);
1982 * Returns true if there are requests available in incoming
1983 * request queue for processing and it is allowed to fetch them.
1984 * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1985 * to get reliable result
1986 * \see ptlrpc_server_allow_normal
1987 * \see ptlrpc_server_allow high
1990 bool ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt,
1993 return ptlrpc_server_high_pending(svcpt, force) ||
1994 ptlrpc_server_normal_pending(svcpt, force);
1998 * Fetch a request for processing from queue of unprocessed requests.
1999 * Favors high-priority requests.
2000 * Returns a pointer to fetched request.
2002 static struct ptlrpc_request *
2003 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force)
2005 struct ptlrpc_request *req = NULL;
2009 spin_lock(&svcpt->scp_req_lock);
2011 if (ptlrpc_server_high_pending(svcpt, force)) {
2012 req = ptlrpc_nrs_req_get_nolock(svcpt, true, force);
2014 svcpt->scp_hreq_count++;
2019 if (ptlrpc_server_normal_pending(svcpt, force)) {
2020 req = ptlrpc_nrs_req_get_nolock(svcpt, false, force);
2022 svcpt->scp_hreq_count = 0;
2027 spin_unlock(&svcpt->scp_req_lock);
2031 svcpt->scp_nreqs_active++;
2033 svcpt->scp_nhreqs_active++;
2035 spin_unlock(&svcpt->scp_req_lock);
2037 if (likely(req->rq_export))
2038 class_export_rpc_inc(req->rq_export);
2044 * Handle freshly incoming reqs, add to timed early reply list,
2045 * pass on to regular request queue.
2046 * All incoming requests pass through here before getting into
2047 * ptlrpc_server_handle_req later on.
2049 static int ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt,
2050 struct ptlrpc_thread *thread)
2052 struct ptlrpc_service *svc = svcpt->scp_service;
2053 struct ptlrpc_request *req;
2059 spin_lock(&svcpt->scp_lock);
2060 if (list_empty(&svcpt->scp_req_incoming)) {
2061 spin_unlock(&svcpt->scp_lock);
2065 req = list_entry(svcpt->scp_req_incoming.next,
2066 struct ptlrpc_request, rq_list);
2067 list_del_init(&req->rq_list);
2068 svcpt->scp_nreqs_incoming--;
2070 * Consider this still a "queued" request as far as stats are
2073 spin_unlock(&svcpt->scp_lock);
2075 /* go through security check/transform */
2076 rc = sptlrpc_svc_unwrap_request(req);
2080 case SECSVC_COMPLETE:
2081 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
2090 * for null-flavored rpc, msg has been unpacked by sptlrpc, although
2091 * redo it wouldn't be harmful.
2093 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
2094 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
2096 CERROR("error unpacking request: ptl %d from %s x%llu\n",
2097 svc->srv_req_portal, libcfs_id2str(req->rq_peer),
2103 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
2105 CERROR("error unpacking ptlrpc body: ptl %d from %s x %llu\n",
2106 svc->srv_req_portal, libcfs_id2str(req->rq_peer),
2111 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
2112 lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
2113 CERROR("drop incoming rpc opc %u, x%llu\n",
2114 cfs_fail_val, req->rq_xid);
2119 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
2120 CERROR("wrong packet type received (type=%u) from %s\n",
2121 lustre_msg_get_type(req->rq_reqmsg),
2122 libcfs_id2str(req->rq_peer));
2126 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2130 req->rq_bulk_write = 1;
2134 case MGS_CONFIG_READ:
2135 req->rq_bulk_read = 1;
2139 CDEBUG(D_RPCTRACE, "got req x%llu\n", req->rq_xid);
2141 req->rq_export = class_conn2export(
2142 lustre_msg_get_handle(req->rq_reqmsg));
2143 if (req->rq_export) {
2144 rc = ptlrpc_check_req(req);
2146 rc = sptlrpc_target_export_check(req->rq_export, req);
2148 DEBUG_REQ(D_ERROR, req,
2149 "DROPPING req with illegal security flavor");
2154 ptlrpc_update_export_timer(req->rq_export, 0);
2157 /* req_in handling should/must be fast */
2158 if (ktime_get_real_seconds() - req->rq_arrival_time.tv_sec > 5)
2159 DEBUG_REQ(D_WARNING, req, "Slow req_in handling %llds",
2160 ktime_get_real_seconds() -
2161 req->rq_arrival_time.tv_sec);
2163 /* Set rpc server deadline and add it to the timed list */
2164 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
2165 MSGHDR_AT_SUPPORT) ?
2166 /* The max time the client expects us to take */
2167 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
2169 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
2170 if (unlikely(deadline == 0)) {
2171 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
2175 /* Skip early reply */
2176 if (OBD_FAIL_PRECHECK(OBD_FAIL_MDS_RESEND))
2177 req->rq_deadline += obd_timeout;
2179 req->rq_svc_thread = thread;
2180 if (thread != NULL) {
2182 * initialize request session, it is needed for request
2183 * processing by target
2185 rc = lu_context_init(&req->rq_session, LCT_SERVER_SESSION |
2188 CERROR("%s: failure to initialize session: rc = %d\n",
2189 thread->t_name, rc);
2192 req->rq_session.lc_thread = thread;
2193 lu_context_enter(&req->rq_session);
2194 thread->t_env->le_ses = &req->rq_session;
2197 ptlrpc_at_add_timed(req);
2199 /* Move it over to the request processing queue */
2200 rc = ptlrpc_server_request_add(svcpt, req);
2204 wake_up(&svcpt->scp_waitq);
2208 ptlrpc_server_finish_request(svcpt, req);
2214 * Main incoming request handling logic.
2215 * Calls handler function from service to do actual processing.
2217 static int ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
2218 struct ptlrpc_thread *thread)
2220 struct ptlrpc_service *svc = svcpt->scp_service;
2221 struct ptlrpc_request *request;
2231 request = ptlrpc_server_request_get(svcpt, false);
2232 if (request == NULL)
2235 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
2236 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
2237 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
2238 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
2240 if (unlikely(fail_opc)) {
2241 if (request->rq_export && request->rq_ops)
2242 OBD_FAIL_TIMEOUT(fail_opc, 4);
2245 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
2247 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
2248 libcfs_debug_dumplog();
2250 work_start = ktime_get_real();
2251 arrived = timespec64_to_ktime(request->rq_arrival_time);
2252 timediff_usecs = ktime_us_delta(work_start, arrived);
2253 if (likely(svc->srv_stats != NULL)) {
2254 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
2256 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
2257 svcpt->scp_nreqs_incoming);
2258 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
2259 svcpt->scp_nreqs_active);
2260 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
2261 at_get(&svcpt->scp_at_estimate));
2264 if (likely(request->rq_export)) {
2265 if (unlikely(ptlrpc_check_req(request)))
2267 ptlrpc_update_export_timer(request->rq_export,
2268 div_u64(timediff_usecs,
2273 * Discard requests queued for longer than the deadline.
2274 * The deadline is increased if we send an early reply.
2276 if (ktime_get_real_seconds() > request->rq_deadline) {
2277 DEBUG_REQ(D_ERROR, request,
2278 "Dropping timed-out request from %s: deadline %lld/%llds ago",
2279 libcfs_id2str(request->rq_peer),
2280 request->rq_deadline -
2281 request->rq_arrival_time.tv_sec,
2282 ktime_get_real_seconds() - request->rq_deadline);
2287 "Handling RPC req@%p pname:cluuid+ref:pid:xid:nid:opc:job %s:%s+%d:%d:x%llu:%s:%d:%s\n",
2288 request, current->comm,
2289 (request->rq_export ?
2290 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2291 (request->rq_export ?
2292 refcount_read(&request->rq_export->exp_handle.h_ref) : -99),
2293 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
2294 libcfs_id2str(request->rq_peer),
2295 lustre_msg_get_opc(request->rq_reqmsg),
2296 lustre_msg_get_jobid(request->rq_reqmsg) ?: "");
2298 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
2299 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
2301 CDEBUG(D_NET, "got req %llu\n", request->rq_xid);
2303 /* re-assign request and sesson thread to the current one */
2304 request->rq_svc_thread = thread;
2305 if (thread != NULL) {
2306 LASSERT(request->rq_session.lc_thread == NULL);
2307 request->rq_session.lc_thread = thread;
2308 thread->t_env->le_ses = &request->rq_session;
2310 svc->srv_ops.so_req_handler(request);
2312 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
2315 if (unlikely(ktime_get_real_seconds() > request->rq_deadline)) {
2316 DEBUG_REQ(D_WARNING, request,
2317 "Request took longer than estimated (%lld/%llds); client may timeout",
2318 request->rq_deadline -
2319 request->rq_arrival_time.tv_sec,
2320 ktime_get_real_seconds() - request->rq_deadline);
2323 work_end = ktime_get_real();
2324 timediff_usecs = ktime_us_delta(work_end, work_start);
2325 arrived_usecs = ktime_us_delta(work_end, arrived);
2327 "Handled RPC req@%p pname:cluuid+ref:pid:xid:nid:opc:job %s:%s+%d:%d:x%llu:%s:%d:%s Request processed in %lldus (%lldus total) trans %llu rc %d/%d\n",
2328 request, current->comm,
2329 (request->rq_export ?
2330 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2331 (request->rq_export ?
2332 refcount_read(&request->rq_export->exp_handle.h_ref) : -99),
2333 lustre_msg_get_status(request->rq_reqmsg),
2335 libcfs_id2str(request->rq_peer),
2336 lustre_msg_get_opc(request->rq_reqmsg),
2337 lustre_msg_get_jobid(request->rq_reqmsg) ?: "",
2340 (request->rq_repmsg ?
2341 lustre_msg_get_transno(request->rq_repmsg) :
2342 request->rq_transno),
2344 (request->rq_repmsg ?
2345 lustre_msg_get_status(request->rq_repmsg) : -999));
2346 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
2347 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
2348 int opc = opcode_offset(op);
2350 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
2351 LASSERT(opc < LUSTRE_MAX_OPCODES);
2352 lprocfs_counter_add(svc->srv_stats,
2353 opc + EXTRA_MAX_OPCODES,
2357 if (unlikely(request->rq_early_count)) {
2358 DEBUG_REQ(D_ADAPTTO, request,
2359 "sent %d early replies before finishing in %llds",
2360 request->rq_early_count,
2361 div_u64(arrived_usecs, USEC_PER_SEC));
2364 ptlrpc_server_finish_active_request(svcpt, request);
2370 * An internal function to process a single reply state object.
2372 static int ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
2374 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
2375 struct ptlrpc_service *svc = svcpt->scp_service;
2376 struct obd_export *exp;
2382 exp = rs->rs_export;
2384 LASSERT(rs->rs_difficult);
2385 LASSERT(rs->rs_scheduled);
2386 LASSERT(list_empty(&rs->rs_list));
2389 * The disk commit callback holds exp_uncommitted_replies_lock while it
2390 * iterates over newly committed replies, removing them from
2391 * exp_uncommitted_replies. It then drops this lock and schedules the
2392 * replies it found for handling here.
2394 * We can avoid contention for exp_uncommitted_replies_lock between the
2395 * HRT threads and further commit callbacks by checking rs_committed
2396 * which is set in the commit callback while it holds both
2397 * rs_lock and exp_uncommitted_reples.
2399 * If we see rs_committed clear, the commit callback _may_ not have
2400 * handled this reply yet and we race with it to grab
2401 * exp_uncommitted_replies_lock before removing the reply from
2402 * exp_uncommitted_replies. Note that if we lose the race and the
2403 * reply has already been removed, list_del_init() is a noop.
2405 * If we see rs_committed set, we know the commit callback is handling,
2406 * or has handled this reply since store reordering might allow us to
2407 * see rs_committed set out of sequence. But since this is done
2408 * holding rs_lock, we can be sure it has all completed once we hold
2409 * rs_lock, which we do right next.
2411 if (!rs->rs_committed) {
2413 * if rs was commited, no need to convert locks, don't check
2414 * rs_committed here because rs may never be added into
2415 * exp_uncommitted_replies and this flag never be set, see
2416 * target_send_reply()
2418 if (rs->rs_convert_lock &&
2419 rs->rs_transno > exp->exp_last_committed) {
2420 struct ldlm_lock *lock;
2421 struct ldlm_lock *ack_locks[RS_MAX_LOCKS] = { NULL };
2423 spin_lock(&rs->rs_lock);
2424 if (rs->rs_convert_lock &&
2425 rs->rs_transno > exp->exp_last_committed) {
2426 nlocks = rs->rs_nlocks;
2427 while (nlocks-- > 0) {
2429 * NB don't assume rs is always handled
2430 * by the same service thread (see
2431 * ptlrpc_hr_select, so REP-ACK hr may
2432 * race with trans commit, while the
2433 * latter will release locks, get locks
2434 * here early to convert to COS mode
2437 lock = ldlm_handle2lock(
2438 &rs->rs_locks[nlocks]);
2440 ack_locks[nlocks] = lock;
2441 rs->rs_modes[nlocks] = LCK_COS;
2443 nlocks = rs->rs_nlocks;
2444 rs->rs_convert_lock = 0;
2446 * clear rs_scheduled so that commit callback
2447 * can schedule again
2449 rs->rs_scheduled = 0;
2450 spin_unlock(&rs->rs_lock);
2452 while (nlocks-- > 0) {
2453 lock = ack_locks[nlocks];
2454 ldlm_lock_mode_downgrade(lock, LCK_COS);
2455 LDLM_LOCK_PUT(lock);
2459 spin_unlock(&rs->rs_lock);
2462 spin_lock(&exp->exp_uncommitted_replies_lock);
2463 list_del_init(&rs->rs_obd_list);
2464 spin_unlock(&exp->exp_uncommitted_replies_lock);
2467 spin_lock(&exp->exp_lock);
2468 /* Noop if removed already */
2469 list_del_init(&rs->rs_exp_list);
2470 spin_unlock(&exp->exp_lock);
2472 spin_lock(&rs->rs_lock);
2474 been_handled = rs->rs_handled;
2477 nlocks = rs->rs_nlocks; /* atomic "steal", but */
2478 rs->rs_nlocks = 0; /* locks still on rs_locks! */
2480 if (nlocks == 0 && !been_handled) {
2482 * If we see this, we should already have seen the warning
2483 * in mds_steal_ack_locks()
2486 "All locks stolen from rs %p x%lld.t%lld o%d NID %s\n",
2487 rs, rs->rs_xid, rs->rs_transno, rs->rs_opc,
2488 libcfs_nid2str(exp->exp_connection->c_peer.nid));
2491 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
2492 spin_unlock(&rs->rs_lock);
2494 if (!been_handled && rs->rs_on_net) {
2495 LNetMDUnlink(rs->rs_md_h);
2496 /* Ignore return code; we're racing with completion */
2499 while (nlocks-- > 0)
2500 ldlm_lock_decref(&rs->rs_locks[nlocks],
2501 rs->rs_modes[nlocks]);
2503 spin_lock(&rs->rs_lock);
2506 rs->rs_scheduled = 0;
2507 rs->rs_convert_lock = 0;
2509 if (!rs->rs_on_net) {
2511 spin_unlock(&rs->rs_lock);
2513 class_export_put(exp);
2514 rs->rs_export = NULL;
2515 ptlrpc_rs_decref(rs);
2516 if (atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
2517 svc->srv_is_stopping)
2518 wake_up_all(&svcpt->scp_waitq);
2522 /* still on the net; callback will schedule */
2523 spin_unlock(&rs->rs_lock);
2528 static void ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
2530 int avail = svcpt->scp_nrqbds_posted;
2531 int low_water = test_req_buffer_pressure ? 0 :
2532 svcpt->scp_service->srv_nbuf_per_group / 2;
2534 /* NB I'm not locking; just looking. */
2537 * CAVEAT EMPTOR: We might be allocating buffers here because we've
2538 * allowed the request history to grow out of control. We could put a
2539 * sanity check on that here and cull some history if we need the
2543 if (avail <= low_water)
2544 ptlrpc_grow_req_bufs(svcpt, 1);
2546 if (svcpt->scp_service->srv_stats) {
2547 lprocfs_counter_add(svcpt->scp_service->srv_stats,
2548 PTLRPC_REQBUF_AVAIL_CNTR, avail);
2552 static inline int ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
2554 return svcpt->scp_nreqs_active <
2555 svcpt->scp_nthrs_running - 1 -
2556 (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL);
2560 * allowed to create more threads
2561 * user can call it w/o any lock but need to hold
2562 * ptlrpc_service_part::scp_lock to get reliable result
2564 static inline int ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
2566 return svcpt->scp_nthrs_running +
2567 svcpt->scp_nthrs_starting <
2568 svcpt->scp_service->srv_nthrs_cpt_limit;
2572 * too many requests and allowed to create more threads
2574 static inline int ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
2576 return !ptlrpc_threads_enough(svcpt) &&
2577 ptlrpc_threads_increasable(svcpt);
2580 static inline int ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2582 return thread_is_stopping(thread) ||
2583 thread->t_svcpt->scp_service->srv_is_stopping;
2586 /* stop the highest numbered thread if there are too many threads running */
2587 static inline bool ptlrpc_thread_should_stop(struct ptlrpc_thread *thread)
2589 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2591 return thread->t_id >= svcpt->scp_service->srv_nthrs_cpt_limit &&
2592 thread->t_id == svcpt->scp_thr_nextid - 1;
2595 static void ptlrpc_stop_thread(struct ptlrpc_thread *thread)
2597 CDEBUG(D_INFO, "Stopping thread %s #%u\n",
2598 thread->t_svcpt->scp_service->srv_thread_name, thread->t_id);
2599 thread_add_flags(thread, SVC_STOPPING);
2602 static inline void ptlrpc_thread_stop(struct ptlrpc_thread *thread)
2604 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2606 spin_lock(&svcpt->scp_lock);
2607 if (ptlrpc_thread_should_stop(thread)) {
2608 ptlrpc_stop_thread(thread);
2609 svcpt->scp_thr_nextid--;
2611 spin_unlock(&svcpt->scp_lock);
2614 static inline int ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
2616 return !list_empty(&svcpt->scp_rqbd_idle) &&
2617 svcpt->scp_rqbd_timeout == 0;
2621 ptlrpc_at_check(struct ptlrpc_service_part *svcpt)
2623 return svcpt->scp_at_check;
2627 * If a thread runs too long or spends to much time on a single request,
2628 * we want to know about it, so we set up a delayed work item as a watchdog.
2629 * If it fires, we display a stack trace of the delayed thread,
2630 * providing we aren't rate-limited
2632 * Watchdog stack traces are limited to 3 per 'libcfs_watchdog_ratelimit'
2635 static struct ratelimit_state watchdog_limit;
2637 static void ptlrpc_watchdog_fire(struct work_struct *w)
2639 struct ptlrpc_thread *thread = container_of(w, struct ptlrpc_thread,
2641 u64 ms_lapse = ktime_ms_delta(ktime_get(), thread->t_touched);
2642 u32 ms_frac = do_div(ms_lapse, MSEC_PER_SEC);
2644 /* ___ratelimit() returns true if the action is NOT ratelimited */
2645 if (__ratelimit(&watchdog_limit)) {
2646 /* below message is checked in sanity-quota.sh test_6,18 */
2647 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",
2648 thread->t_task->comm, thread->t_task->pid,
2651 libcfs_debug_dumpstack(thread->t_task);
2653 /* below message is checked in sanity-quota.sh test_6,18 */
2654 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",
2655 thread->t_task->comm, thread->t_task->pid,
2656 ms_lapse, ms_frac, libcfs_watchdog_ratelimit);
2660 void ptlrpc_watchdog_init(struct delayed_work *work, timeout_t timeout)
2662 INIT_DELAYED_WORK(work, ptlrpc_watchdog_fire);
2663 schedule_delayed_work(work, cfs_time_seconds(timeout));
2666 void ptlrpc_watchdog_disable(struct delayed_work *work)
2668 cancel_delayed_work_sync(work);
2671 void ptlrpc_watchdog_touch(struct delayed_work *work, timeout_t timeout)
2673 struct ptlrpc_thread *thread = container_of(&work->work,
2674 struct ptlrpc_thread,
2676 thread->t_touched = ktime_get();
2677 mod_delayed_work(system_wq, work, cfs_time_seconds(timeout));
2681 * requests wait on preprocessing
2682 * user can call it w/o any lock but need to hold
2683 * ptlrpc_service_part::scp_lock to get reliable result
2686 ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
2688 return !list_empty(&svcpt->scp_req_incoming);
2691 static __attribute__((__noinline__)) int
2692 ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
2693 struct ptlrpc_thread *thread)
2695 ptlrpc_watchdog_disable(&thread->t_watchdog);
2699 if (svcpt->scp_rqbd_timeout == 0)
2700 /* Don't exit while there are replies to be handled */
2701 wait_event_idle_exclusive_lifo(
2703 ptlrpc_thread_stopping(thread) ||
2704 ptlrpc_server_request_incoming(svcpt) ||
2705 ptlrpc_server_request_pending(svcpt, false) ||
2706 ptlrpc_rqbd_pending(svcpt) ||
2707 ptlrpc_at_check(svcpt));
2708 else if (wait_event_idle_exclusive_lifo_timeout(
2710 ptlrpc_thread_stopping(thread) ||
2711 ptlrpc_server_request_incoming(svcpt) ||
2712 ptlrpc_server_request_pending(svcpt, false) ||
2713 ptlrpc_rqbd_pending(svcpt) ||
2714 ptlrpc_at_check(svcpt),
2715 svcpt->scp_rqbd_timeout) == 0)
2716 svcpt->scp_rqbd_timeout = 0;
2718 if (ptlrpc_thread_stopping(thread))
2721 ptlrpc_watchdog_touch(&thread->t_watchdog,
2722 ptlrpc_server_get_timeout(svcpt));
2727 * Main thread body for service threads.
2728 * Waits in a loop waiting for new requests to process to appear.
2729 * Every time an incoming requests is added to its queue, a waitq
2730 * is woken up and one of the threads will handle it.
2732 static int ptlrpc_main(void *arg)
2734 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
2735 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2736 struct ptlrpc_service *svc = svcpt->scp_service;
2737 struct ptlrpc_reply_state *rs;
2738 struct group_info *ginfo = NULL;
2740 int counter = 0, rc = 0;
2744 thread->t_task = current;
2745 thread->t_pid = current->pid;
2747 if (svc->srv_cpt_bind) {
2748 rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
2750 CWARN("%s: failed to bind %s on CPT %d\n",
2751 svc->srv_name, thread->t_name, svcpt->scp_cpt);
2755 ginfo = groups_alloc(0);
2757 GOTO(out, rc = -ENOMEM);
2759 set_current_groups(ginfo);
2760 put_group_info(ginfo);
2762 if (svc->srv_ops.so_thr_init != NULL) {
2763 rc = svc->srv_ops.so_thr_init(thread);
2770 GOTO(out_srv_fini, rc = -ENOMEM);
2771 rc = lu_env_add(env);
2775 rc = lu_context_init(&env->le_ctx,
2776 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2778 GOTO(out_env_remove, rc);
2780 thread->t_env = env;
2781 env->le_ctx.lc_thread = thread;
2782 env->le_ctx.lc_cookie = 0x6;
2784 while (!list_empty(&svcpt->scp_rqbd_idle)) {
2785 rc = ptlrpc_server_post_idle_rqbds(svcpt);
2789 CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
2790 svc->srv_name, svcpt->scp_cpt, rc);
2791 GOTO(out_ctx_fini, rc);
2794 /* Alloc reply state structure for this one */
2795 OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2797 GOTO(out_ctx_fini, rc = -ENOMEM);
2799 spin_lock(&svcpt->scp_lock);
2801 LASSERT(thread_is_starting(thread));
2802 thread_clear_flags(thread, SVC_STARTING);
2804 LASSERT(svcpt->scp_nthrs_starting == 1);
2805 svcpt->scp_nthrs_starting--;
2808 * SVC_STOPPING may already be set here if someone else is trying
2809 * to stop the service while this new thread has been dynamically
2810 * forked. We still set SVC_RUNNING to let our creator know that
2811 * we are now running, however we will exit as soon as possible
2813 thread_add_flags(thread, SVC_RUNNING);
2814 svcpt->scp_nthrs_running++;
2815 spin_unlock(&svcpt->scp_lock);
2817 /* wake up our creator in case he's still waiting. */
2818 wake_up(&thread->t_ctl_waitq);
2820 thread->t_touched = ktime_get();
2821 ptlrpc_watchdog_init(&thread->t_watchdog,
2822 ptlrpc_server_get_timeout(svcpt));
2824 spin_lock(&svcpt->scp_rep_lock);
2825 list_add(&rs->rs_list, &svcpt->scp_rep_idle);
2826 wake_up(&svcpt->scp_rep_waitq);
2827 spin_unlock(&svcpt->scp_rep_lock);
2829 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2830 svcpt->scp_nthrs_running);
2832 /* XXX maintain a list of all managed devices: insert here */
2833 while (!ptlrpc_thread_stopping(thread)) {
2834 if (ptlrpc_wait_event(svcpt, thread))
2837 ptlrpc_check_rqbd_pool(svcpt);
2839 if (ptlrpc_threads_need_create(svcpt)) {
2840 /* Ignore return code - we tried... */
2841 ptlrpc_start_thread(svcpt, 0);
2844 /* reset le_ses to initial state */
2846 /* Refill the context before execution to make sure
2847 * all thread keys are allocated */
2849 /* Process all incoming reqs before handling any */
2850 if (ptlrpc_server_request_incoming(svcpt)) {
2851 lu_context_enter(&env->le_ctx);
2852 ptlrpc_server_handle_req_in(svcpt, thread);
2853 lu_context_exit(&env->le_ctx);
2855 /* but limit ourselves in case of flood */
2856 if (counter++ < 100)
2861 if (ptlrpc_at_check(svcpt))
2862 ptlrpc_at_check_timed(svcpt);
2864 if (ptlrpc_server_request_pending(svcpt, false)) {
2865 lu_context_enter(&env->le_ctx);
2866 ptlrpc_server_handle_request(svcpt, thread);
2867 lu_context_exit(&env->le_ctx);
2870 if (ptlrpc_rqbd_pending(svcpt) &&
2871 ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
2873 * I just failed to repost request buffers.
2874 * Wait for a timeout (unless something else
2875 * happens) before I try again
2877 svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
2878 CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
2879 svcpt->scp_nrqbds_posted);
2882 * If the number of threads has been tuned downward and this
2883 * thread should be stopped, then stop in reverse order so the
2884 * the threads always have contiguous thread index values.
2886 if (unlikely(ptlrpc_thread_should_stop(thread)))
2887 ptlrpc_thread_stop(thread);
2890 ptlrpc_watchdog_disable(&thread->t_watchdog);
2893 lu_context_fini(&env->le_ctx);
2899 /* deconstruct service thread state created by ptlrpc_start_thread() */
2900 if (svc->srv_ops.so_thr_done != NULL)
2901 svc->srv_ops.so_thr_done(thread);
2903 CDEBUG(D_RPCTRACE, "%s: service thread [%p:%u] %d exiting: rc = %d\n",
2904 thread->t_name, thread, thread->t_pid, thread->t_id, rc);
2905 spin_lock(&svcpt->scp_lock);
2906 if (thread_test_and_clear_flags(thread, SVC_STARTING))
2907 svcpt->scp_nthrs_starting--;
2909 if (thread_test_and_clear_flags(thread, SVC_RUNNING)) {
2910 /* must know immediately */
2911 svcpt->scp_nthrs_running--;
2915 thread_add_flags(thread, SVC_STOPPED);
2917 wake_up(&thread->t_ctl_waitq);
2918 spin_unlock(&svcpt->scp_lock);
2923 static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
2924 struct list_head *replies)
2928 spin_lock(&hrt->hrt_lock);
2930 list_splice_init(&hrt->hrt_queue, replies);
2931 result = ptlrpc_hr.hr_stopping || !list_empty(replies);
2933 spin_unlock(&hrt->hrt_lock);
2938 * Main body of "handle reply" function.
2939 * It processes acked reply states
2941 static int ptlrpc_hr_main(void *arg)
2943 struct ptlrpc_hr_thread *hrt = (struct ptlrpc_hr_thread *)arg;
2944 struct ptlrpc_hr_partition *hrp = hrt->hrt_partition;
2953 rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
2955 char threadname[20];
2957 snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
2958 hrp->hrp_cpt, hrt->hrt_id);
2959 CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
2960 threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
2963 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD |
2964 LCT_REMEMBER | LCT_NOREF);
2968 rc = lu_env_add(env);
2970 GOTO(out_ctx_fini, rc);
2972 atomic_inc(&hrp->hrp_nstarted);
2973 wake_up(&ptlrpc_hr.hr_waitq);
2975 while (!ptlrpc_hr.hr_stopping) {
2976 wait_event_idle(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
2978 while (!list_empty(&replies)) {
2979 struct ptlrpc_reply_state *rs;
2981 rs = list_entry(replies.prev,
2982 struct ptlrpc_reply_state,
2984 list_del_init(&rs->rs_list);
2985 /* refill keys if needed */
2987 lu_context_enter(&env->le_ctx);
2988 ptlrpc_handle_rs(rs);
2989 lu_context_exit(&env->le_ctx);
2993 atomic_inc(&hrp->hrp_nstopped);
2994 wake_up(&ptlrpc_hr.hr_waitq);
2998 lu_context_fini(&env->le_ctx);
3004 static void ptlrpc_stop_hr_threads(void)
3006 struct ptlrpc_hr_partition *hrp;
3010 ptlrpc_hr.hr_stopping = 1;
3012 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
3013 if (hrp->hrp_thrs == NULL)
3014 continue; /* uninitialized */
3015 for (j = 0; j < hrp->hrp_nthrs; j++)
3016 wake_up_all(&hrp->hrp_thrs[j].hrt_waitq);
3019 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
3020 if (hrp->hrp_thrs == NULL)
3021 continue; /* uninitialized */
3022 wait_event(ptlrpc_hr.hr_waitq,
3023 atomic_read(&hrp->hrp_nstopped) ==
3024 atomic_read(&hrp->hrp_nstarted));
3028 static int ptlrpc_start_hr_threads(void)
3030 struct ptlrpc_hr_partition *hrp;
3036 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
3039 for (j = 0; j < hrp->hrp_nthrs; j++) {
3040 struct ptlrpc_hr_thread *hrt = &hrp->hrp_thrs[j];
3041 struct task_struct *task;
3043 task = kthread_run(ptlrpc_hr_main,
3045 "ptlrpc_hr%02d_%03d",
3054 wait_event(ptlrpc_hr.hr_waitq,
3055 atomic_read(&hrp->hrp_nstarted) == j);
3058 CERROR("cannot start reply handler thread %d:%d: rc = %d\n",
3060 ptlrpc_stop_hr_threads();
3068 static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
3070 struct ptlrpc_thread *thread;
3075 CDEBUG(D_INFO, "Stopping threads for service %s\n",
3076 svcpt->scp_service->srv_name);
3078 spin_lock(&svcpt->scp_lock);
3079 /* let the thread know that we would like it to stop asap */
3080 list_for_each_entry(thread, &svcpt->scp_threads, t_link)
3081 ptlrpc_stop_thread(thread);
3083 wake_up_all(&svcpt->scp_waitq);
3085 while (!list_empty(&svcpt->scp_threads)) {
3086 thread = list_entry(svcpt->scp_threads.next,
3087 struct ptlrpc_thread, t_link);
3088 if (thread_is_stopped(thread)) {
3089 list_move(&thread->t_link, &zombie);
3092 spin_unlock(&svcpt->scp_lock);
3094 CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
3095 svcpt->scp_service->srv_thread_name, thread->t_id);
3096 wait_event_idle(thread->t_ctl_waitq,
3097 thread_is_stopped(thread));
3099 spin_lock(&svcpt->scp_lock);
3102 spin_unlock(&svcpt->scp_lock);
3104 while (!list_empty(&zombie)) {
3105 thread = list_entry(zombie.next,
3106 struct ptlrpc_thread, t_link);
3107 list_del(&thread->t_link);
3108 OBD_FREE_PTR(thread);
3114 * Stops all threads of a particular service \a svc
3116 static void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
3118 struct ptlrpc_service_part *svcpt;
3123 ptlrpc_service_for_each_part(svcpt, i, svc) {
3124 if (svcpt->scp_service != NULL)
3125 ptlrpc_svcpt_stop_threads(svcpt);
3131 static int ptlrpc_start_threads(struct ptlrpc_service *svc)
3139 /* We require 2 threads min, see note in ptlrpc_server_handle_request */
3140 LASSERT(svc->srv_nthrs_cpt_init >= PTLRPC_NTHRS_INIT);
3142 for (i = 0; i < svc->srv_ncpts; i++) {
3143 for (j = 0; j < svc->srv_nthrs_cpt_init; j++) {
3144 rc = ptlrpc_start_thread(svc->srv_parts[i], 1);
3150 /* We have enough threads, don't start more. b=15759 */
3157 CERROR("cannot start %s thread #%d_%d: rc %d\n",
3158 svc->srv_thread_name, i, j, rc);
3159 ptlrpc_stop_all_threads(svc);
3163 static int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
3165 struct ptlrpc_thread *thread;
3166 struct ptlrpc_service *svc;
3167 struct task_struct *task;
3172 LASSERT(svcpt != NULL);
3174 svc = svcpt->scp_service;
3176 CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n",
3177 svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running,
3178 svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit);
3181 if (unlikely(svc->srv_is_stopping))
3184 if (!ptlrpc_threads_increasable(svcpt) ||
3185 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
3186 svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1))
3189 OBD_CPT_ALLOC_PTR(thread, svc->srv_cptable, svcpt->scp_cpt);
3192 init_waitqueue_head(&thread->t_ctl_waitq);
3194 spin_lock(&svcpt->scp_lock);
3195 if (!ptlrpc_threads_increasable(svcpt)) {
3196 spin_unlock(&svcpt->scp_lock);
3197 OBD_FREE_PTR(thread);
3201 if (svcpt->scp_nthrs_starting != 0) {
3203 * serialize starting because some modules (obdfilter)
3204 * might require unique and contiguous t_id
3206 LASSERT(svcpt->scp_nthrs_starting == 1);
3207 spin_unlock(&svcpt->scp_lock);
3208 OBD_FREE_PTR(thread);
3210 CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n",
3211 svc->srv_thread_name, svcpt->scp_thr_nextid);
3216 CDEBUG(D_INFO, "Creating thread %s #%d race, retry later\n",
3217 svc->srv_thread_name, svcpt->scp_thr_nextid);
3221 svcpt->scp_nthrs_starting++;
3222 thread->t_id = svcpt->scp_thr_nextid++;
3223 thread_add_flags(thread, SVC_STARTING);
3224 thread->t_svcpt = svcpt;
3226 list_add(&thread->t_link, &svcpt->scp_threads);
3227 spin_unlock(&svcpt->scp_lock);
3229 if (svcpt->scp_cpt >= 0) {
3230 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s%02d_%03d",
3231 svc->srv_thread_name, svcpt->scp_cpt, thread->t_id);
3233 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s_%04d",
3234 svc->srv_thread_name, thread->t_id);
3237 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name);
3238 task = kthread_run(ptlrpc_main, thread, "%s", thread->t_name);
3241 CERROR("cannot start thread '%s': rc = %d\n",
3242 thread->t_name, rc);
3243 spin_lock(&svcpt->scp_lock);
3244 --svcpt->scp_nthrs_starting;
3245 if (thread_is_stopping(thread)) {
3247 * this ptlrpc_thread is being hanled
3248 * by ptlrpc_svcpt_stop_threads now
3250 thread_add_flags(thread, SVC_STOPPED);
3251 wake_up(&thread->t_ctl_waitq);
3252 spin_unlock(&svcpt->scp_lock);
3254 list_del(&thread->t_link);
3255 spin_unlock(&svcpt->scp_lock);
3256 OBD_FREE_PTR(thread);
3264 wait_event_idle(thread->t_ctl_waitq,
3265 thread_is_running(thread) || thread_is_stopped(thread));
3267 rc = thread_is_stopped(thread) ? thread->t_id : 0;
3271 int ptlrpc_hr_init(void)
3273 struct ptlrpc_hr_partition *hrp;
3274 struct ptlrpc_hr_thread *hrt;
3282 memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
3283 ptlrpc_hr.hr_cpt_table = cfs_cpt_tab;
3285 ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
3287 if (ptlrpc_hr.hr_partitions == NULL)
3290 ratelimit_state_init(&watchdog_limit,
3291 cfs_time_seconds(libcfs_watchdog_ratelimit), 3);
3293 init_waitqueue_head(&ptlrpc_hr.hr_waitq);
3296 weight = cpumask_weight(topology_sibling_cpumask(smp_processor_id()));
3299 cfs_percpt_for_each(hrp, cpt, ptlrpc_hr.hr_partitions) {
3302 atomic_set(&hrp->hrp_nstarted, 0);
3303 atomic_set(&hrp->hrp_nstopped, 0);
3305 hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, cpt);
3306 hrp->hrp_nthrs /= weight;
3307 if (hrp->hrp_nthrs == 0)
3310 OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, cpt,
3311 hrp->hrp_nthrs * sizeof(*hrt));
3312 if (hrp->hrp_thrs == NULL)
3313 GOTO(out, rc = -ENOMEM);
3315 for (i = 0; i < hrp->hrp_nthrs; i++) {
3316 hrt = &hrp->hrp_thrs[i];
3319 hrt->hrt_partition = hrp;
3320 init_waitqueue_head(&hrt->hrt_waitq);
3321 spin_lock_init(&hrt->hrt_lock);
3322 INIT_LIST_HEAD(&hrt->hrt_queue);
3326 rc = ptlrpc_start_hr_threads();
3333 void ptlrpc_hr_fini(void)
3335 struct ptlrpc_hr_partition *hrp;
3338 if (ptlrpc_hr.hr_partitions == NULL)
3341 ptlrpc_stop_hr_threads();
3343 cfs_percpt_for_each(hrp, cpt, ptlrpc_hr.hr_partitions) {
3345 OBD_FREE_PTR_ARRAY(hrp->hrp_thrs, hrp->hrp_nthrs);
3348 cfs_percpt_free(ptlrpc_hr.hr_partitions);
3349 ptlrpc_hr.hr_partitions = NULL;
3354 * Wait until all already scheduled replies are processed.
3356 static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
3359 if (wait_event_idle_timeout(
3361 atomic_read(&svcpt->scp_nreps_difficult) == 0,
3362 cfs_time_seconds(10)) > 0)
3364 CWARN("Unexpectedly long timeout %s %p\n",
3365 svcpt->scp_service->srv_name, svcpt->scp_service);
3370 ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
3372 struct ptlrpc_service_part *svcpt;
3375 /* early disarm AT timer... */
3376 ptlrpc_service_for_each_part(svcpt, i, svc) {
3377 if (svcpt->scp_service != NULL)
3378 del_timer(&svcpt->scp_at_timer);
3383 ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
3385 struct ptlrpc_service_part *svcpt;
3386 struct ptlrpc_request_buffer_desc *rqbd;
3391 * All history will be culled when the next request buffer is
3392 * freed in ptlrpc_service_purge_all()
3394 svc->srv_hist_nrqbds_cpt_max = 0;
3396 rc = LNetClearLazyPortal(svc->srv_req_portal);
3399 ptlrpc_service_for_each_part(svcpt, i, svc) {
3400 if (svcpt->scp_service == NULL)
3404 * Unlink all the request buffers. This forces a 'final'
3405 * event with its 'unlink' flag set for each posted rqbd
3407 list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
3409 rc = LNetMDUnlink(rqbd->rqbd_md_h);
3410 LASSERT(rc == 0 || rc == -ENOENT);
3414 ptlrpc_service_for_each_part(svcpt, i, svc) {
3415 if (svcpt->scp_service == NULL)
3419 * Wait for the network to release any buffers
3420 * it's currently filling
3422 spin_lock(&svcpt->scp_lock);
3423 while (svcpt->scp_nrqbds_posted != 0) {
3424 int seconds = PTLRPC_REQ_LONG_UNLINK;
3426 spin_unlock(&svcpt->scp_lock);
3428 * Network access will complete in finite time but
3429 * the HUGE timeout lets us CWARN for visibility
3432 while (seconds > 0 &&
3433 wait_event_idle_timeout(
3435 svcpt->scp_nrqbds_posted == 0,
3436 cfs_time_seconds(1)) == 0)
3439 CWARN("Service %s waiting for request buffers\n",
3440 svcpt->scp_service->srv_name);
3442 spin_lock(&svcpt->scp_lock);
3444 spin_unlock(&svcpt->scp_lock);
3449 ptlrpc_service_purge_all(struct ptlrpc_service *svc)
3451 struct ptlrpc_service_part *svcpt;
3452 struct ptlrpc_request_buffer_desc *rqbd;
3453 struct ptlrpc_request *req;
3454 struct ptlrpc_reply_state *rs;
3457 ptlrpc_service_for_each_part(svcpt, i, svc) {
3458 if (svcpt->scp_service == NULL)
3461 spin_lock(&svcpt->scp_rep_lock);
3462 while (!list_empty(&svcpt->scp_rep_active)) {
3463 rs = list_entry(svcpt->scp_rep_active.next,
3464 struct ptlrpc_reply_state, rs_list);
3465 spin_lock(&rs->rs_lock);
3466 ptlrpc_schedule_difficult_reply(rs);
3467 spin_unlock(&rs->rs_lock);
3469 spin_unlock(&svcpt->scp_rep_lock);
3472 * purge the request queue. NB No new replies (rqbds
3473 * all unlinked) and no service threads, so I'm the only
3474 * thread noodling the request queue now
3476 while (!list_empty(&svcpt->scp_req_incoming)) {
3477 req = list_entry(svcpt->scp_req_incoming.next,
3478 struct ptlrpc_request, rq_list);
3480 list_del(&req->rq_list);
3481 svcpt->scp_nreqs_incoming--;
3482 ptlrpc_server_finish_request(svcpt, req);
3485 while (ptlrpc_server_request_pending(svcpt, true)) {
3486 req = ptlrpc_server_request_get(svcpt, true);
3487 ptlrpc_server_finish_active_request(svcpt, req);
3491 * The portal may be shared by several services (eg:OUT_PORTAL).
3492 * So the request could be referenced by other target. So we
3493 * have to wait the ptlrpc_server_drop_request invoked.
3495 * TODO: move the req_buffer as global rather than per service.
3497 spin_lock(&svcpt->scp_lock);
3498 while (!list_empty(&svcpt->scp_rqbd_posted)) {
3499 spin_unlock(&svcpt->scp_lock);
3500 wait_event_idle_timeout(svcpt->scp_waitq,
3501 list_empty(&svcpt->scp_rqbd_posted),
3502 cfs_time_seconds(1));
3503 spin_lock(&svcpt->scp_lock);
3505 spin_unlock(&svcpt->scp_lock);
3507 LASSERT(svcpt->scp_nreqs_incoming == 0);
3508 LASSERT(svcpt->scp_nreqs_active == 0);
3510 * history should have been culled by
3511 * ptlrpc_server_finish_request
3513 LASSERT(svcpt->scp_hist_nrqbds == 0);
3516 * Now free all the request buffers since nothing
3517 * references them any more...
3520 while (!list_empty(&svcpt->scp_rqbd_idle)) {
3521 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
3522 struct ptlrpc_request_buffer_desc,
3524 ptlrpc_free_rqbd(rqbd);
3526 ptlrpc_wait_replies(svcpt);
3528 while (!list_empty(&svcpt->scp_rep_idle)) {
3529 rs = list_entry(svcpt->scp_rep_idle.next,
3530 struct ptlrpc_reply_state,
3532 list_del(&rs->rs_list);
3533 OBD_FREE_LARGE(rs, svc->srv_max_reply_size);
3539 ptlrpc_service_free(struct ptlrpc_service *svc)
3541 struct ptlrpc_service_part *svcpt;
3542 struct ptlrpc_at_array *array;
3545 ptlrpc_service_for_each_part(svcpt, i, svc) {
3546 if (svcpt->scp_service == NULL)
3549 /* In case somebody rearmed this in the meantime */
3550 del_timer(&svcpt->scp_at_timer);
3551 array = &svcpt->scp_at_array;
3553 if (array->paa_reqs_array != NULL) {
3554 OBD_FREE_PTR_ARRAY(array->paa_reqs_array,
3556 array->paa_reqs_array = NULL;
3559 if (array->paa_reqs_count != NULL) {
3560 OBD_FREE_PTR_ARRAY(array->paa_reqs_count,
3562 array->paa_reqs_count = NULL;
3566 ptlrpc_service_for_each_part(svcpt, i, svc)
3567 OBD_FREE_PTR(svcpt);
3569 if (svc->srv_cpts != NULL)
3570 cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts);
3572 OBD_FREE(svc, offsetof(struct ptlrpc_service,
3573 srv_parts[svc->srv_ncpts]));
3576 int ptlrpc_unregister_service(struct ptlrpc_service *service)
3580 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
3582 service->srv_is_stopping = 1;
3584 mutex_lock(&ptlrpc_all_services_mutex);
3585 list_del_init(&service->srv_list);
3586 mutex_unlock(&ptlrpc_all_services_mutex);
3588 ptlrpc_service_del_atimer(service);
3589 ptlrpc_stop_all_threads(service);
3591 ptlrpc_service_unlink_rqbd(service);
3592 ptlrpc_service_purge_all(service);
3593 ptlrpc_service_nrs_cleanup(service);
3595 ptlrpc_lprocfs_unregister_service(service);
3596 ptlrpc_sysfs_unregister_service(service);
3598 ptlrpc_service_free(service);
3602 EXPORT_SYMBOL(ptlrpc_unregister_service);
3605 * Returns 0 if the service is healthy.
3607 * Right now, it just checks to make sure that requests aren't languishing
3608 * in the queue. We'll use this health check to govern whether a node needs
3609 * to be shot, so it's intentionally non-aggressive.
3611 static int ptlrpc_svcpt_health_check(struct ptlrpc_service_part *svcpt)
3613 struct ptlrpc_request *request = NULL;
3614 struct timespec64 right_now;
3615 struct timespec64 timediff;
3617 ktime_get_real_ts64(&right_now);
3619 spin_lock(&svcpt->scp_req_lock);
3620 /* How long has the next entry been waiting? */
3621 if (ptlrpc_server_high_pending(svcpt, true))
3622 request = ptlrpc_nrs_req_peek_nolock(svcpt, true);
3623 else if (ptlrpc_server_normal_pending(svcpt, true))
3624 request = ptlrpc_nrs_req_peek_nolock(svcpt, false);
3626 if (request == NULL) {
3627 spin_unlock(&svcpt->scp_req_lock);
3631 timediff = timespec64_sub(right_now, request->rq_arrival_time);
3632 spin_unlock(&svcpt->scp_req_lock);
3634 if ((timediff.tv_sec) >
3635 (AT_OFF ? obd_timeout * 3 / 2 : at_max)) {
3636 CERROR("%s: unhealthy - request has been waiting %llds\n",
3637 svcpt->scp_service->srv_name, (s64)timediff.tv_sec);
3645 ptlrpc_service_health_check(struct ptlrpc_service *svc)
3647 struct ptlrpc_service_part *svcpt;
3653 ptlrpc_service_for_each_part(svcpt, i, svc) {
3654 int rc = ptlrpc_svcpt_health_check(svcpt);
3661 EXPORT_SYMBOL(ptlrpc_service_health_check);