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/
32 #define DEBUG_SUBSYSTEM S_RPC
34 #include <linux/kthread.h>
35 #include <linux/ratelimit.h>
37 #include <obd_support.h>
38 #include <obd_class.h>
39 #include <lustre_net.h>
40 #include <lu_object.h>
41 #include <uapi/linux/lnet/lnet-types.h>
42 #include "ptlrpc_internal.h"
43 #include <linux/delay.h>
45 /* The following are visible and mutable through /sys/module/ptlrpc */
46 int test_req_buffer_pressure = 0;
47 module_param(test_req_buffer_pressure, int, 0444);
48 MODULE_PARM_DESC(test_req_buffer_pressure, "set non-zero to put pressure on request buffer pools");
49 module_param(at_min, int, 0644);
50 MODULE_PARM_DESC(at_min, "Adaptive timeout minimum (sec)");
51 module_param(at_max, int, 0644);
52 MODULE_PARM_DESC(at_max, "Adaptive timeout maximum (sec)");
53 module_param(at_history, int, 0644);
54 MODULE_PARM_DESC(at_history,
55 "Adaptive timeouts remember the slowest event that took place within this period (sec)");
56 module_param(at_early_margin, int, 0644);
57 MODULE_PARM_DESC(at_early_margin, "How soon before an RPC deadline to send an early reply");
58 module_param(at_extra, int, 0644);
59 MODULE_PARM_DESC(at_extra, "How much extra time to give with each early reply");
62 static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt);
63 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req);
64 static void ptlrpc_at_remove_timed(struct ptlrpc_request *req);
65 static int ptlrpc_start_threads(struct ptlrpc_service *svc);
66 static int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait);
68 /** Holds a list of all PTLRPC services */
69 LIST_HEAD(ptlrpc_all_services);
70 /** Used to protect the \e ptlrpc_all_services list */
71 struct mutex ptlrpc_all_services_mutex;
73 static struct ptlrpc_request_buffer_desc *
74 ptlrpc_alloc_rqbd(struct ptlrpc_service_part *svcpt)
76 struct ptlrpc_service *svc = svcpt->scp_service;
77 struct ptlrpc_request_buffer_desc *rqbd;
79 OBD_CPT_ALLOC_PTR(rqbd, svc->srv_cptable, svcpt->scp_cpt);
83 rqbd->rqbd_svcpt = svcpt;
84 rqbd->rqbd_refcount = 0;
85 rqbd->rqbd_cbid.cbid_fn = request_in_callback;
86 rqbd->rqbd_cbid.cbid_arg = rqbd;
87 INIT_LIST_HEAD(&rqbd->rqbd_reqs);
88 OBD_CPT_ALLOC_LARGE(rqbd->rqbd_buffer, svc->srv_cptable,
89 svcpt->scp_cpt, svc->srv_buf_size);
90 if (rqbd->rqbd_buffer == NULL) {
95 spin_lock(&svcpt->scp_lock);
96 list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
97 svcpt->scp_nrqbds_total++;
98 spin_unlock(&svcpt->scp_lock);
103 static void ptlrpc_free_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
105 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
107 LASSERT(rqbd->rqbd_refcount == 0);
108 LASSERT(list_empty(&rqbd->rqbd_reqs));
110 spin_lock(&svcpt->scp_lock);
111 list_del(&rqbd->rqbd_list);
112 svcpt->scp_nrqbds_total--;
113 spin_unlock(&svcpt->scp_lock);
115 OBD_FREE_LARGE(rqbd->rqbd_buffer, svcpt->scp_service->srv_buf_size);
119 static int ptlrpc_grow_req_bufs(struct ptlrpc_service_part *svcpt, int post)
121 struct ptlrpc_service *svc = svcpt->scp_service;
122 struct ptlrpc_request_buffer_desc *rqbd;
126 if (svcpt->scp_rqbd_allocating)
129 spin_lock(&svcpt->scp_lock);
130 /* check again with lock */
131 if (svcpt->scp_rqbd_allocating) {
132 /* NB: we might allow more than one thread in the future */
133 LASSERT(svcpt->scp_rqbd_allocating == 1);
134 spin_unlock(&svcpt->scp_lock);
138 svcpt->scp_rqbd_allocating++;
139 spin_unlock(&svcpt->scp_lock);
142 for (i = 0; i < svc->srv_nbuf_per_group; i++) {
144 * NB: another thread might have recycled enough rqbds, we
145 * need to make sure it wouldn't over-allocate, see LU-1212.
147 if (svcpt->scp_nrqbds_posted >= svc->srv_nbuf_per_group ||
148 (svc->srv_nrqbds_max != 0 &&
149 svcpt->scp_nrqbds_total > svc->srv_nrqbds_max))
152 rqbd = ptlrpc_alloc_rqbd(svcpt);
155 CERROR("%s: Can't allocate request buffer\n",
162 spin_lock(&svcpt->scp_lock);
164 LASSERT(svcpt->scp_rqbd_allocating == 1);
165 svcpt->scp_rqbd_allocating--;
167 spin_unlock(&svcpt->scp_lock);
170 "%s: allocate %d new %d-byte reqbufs (%d/%d left), rc = %d\n",
171 svc->srv_name, i, svc->srv_buf_size, svcpt->scp_nrqbds_posted,
172 svcpt->scp_nrqbds_total, rc);
176 rc = ptlrpc_server_post_idle_rqbds(svcpt);
182 * Part of Rep-Ack logic.
183 * Puts a lock and its mode into reply state assotiated to request reply.
185 void ptlrpc_save_lock(struct ptlrpc_request *req, struct lustre_handle *lock,
186 int mode, bool no_ack, bool convert_lock)
188 struct ptlrpc_reply_state *rs = req->rq_reply_state;
192 LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
194 idx = rs->rs_nlocks++;
195 rs->rs_locks[idx] = *lock;
196 rs->rs_modes[idx] = mode;
197 rs->rs_difficult = 1;
198 rs->rs_no_ack = no_ack;
199 rs->rs_convert_lock = convert_lock;
201 EXPORT_SYMBOL(ptlrpc_save_lock);
204 struct ptlrpc_hr_partition;
206 struct ptlrpc_hr_thread {
207 int hrt_id; /* thread ID */
209 wait_queue_head_t hrt_waitq;
210 struct list_head hrt_queue;
211 struct ptlrpc_hr_partition *hrt_partition;
214 struct ptlrpc_hr_partition {
215 /* # of started threads */
216 atomic_t hrp_nstarted;
217 /* # of stopped threads */
218 atomic_t hrp_nstopped;
219 /* cpu partition id */
221 /* round-robin rotor for choosing thread */
223 /* total number of threads on this partition */
226 struct ptlrpc_hr_thread *hrp_thrs;
229 #define HRT_RUNNING 0
230 #define HRT_STOPPING 1
232 struct ptlrpc_hr_service {
233 /* CPU partition table, it's just cfs_cpt_tab for now */
234 struct cfs_cpt_table *hr_cpt_table;
235 /** controller sleep waitq */
236 wait_queue_head_t hr_waitq;
237 unsigned int hr_stopping;
238 /** roundrobin rotor for non-affinity service */
239 unsigned int hr_rotor;
241 struct ptlrpc_hr_partition **hr_partitions;
245 struct list_head rsb_replies;
246 unsigned int rsb_n_replies;
247 struct ptlrpc_service_part *rsb_svcpt;
250 /** reply handling service. */
251 static struct ptlrpc_hr_service ptlrpc_hr;
254 * maximum mumber of replies scheduled in one batch
256 #define MAX_SCHEDULED 256
259 * Initialize a reply batch.
263 static void rs_batch_init(struct rs_batch *b)
265 memset(b, 0, sizeof(*b));
266 INIT_LIST_HEAD(&b->rsb_replies);
270 * Choose an hr thread to dispatch requests to.
273 struct ptlrpc_hr_thread *ptlrpc_hr_select(struct ptlrpc_service_part *svcpt)
275 struct ptlrpc_hr_partition *hrp;
278 if (svcpt->scp_cpt >= 0 &&
279 svcpt->scp_service->srv_cptable == ptlrpc_hr.hr_cpt_table) {
280 /* directly match partition */
281 hrp = ptlrpc_hr.hr_partitions[svcpt->scp_cpt];
284 rotor = ptlrpc_hr.hr_rotor++;
285 rotor %= cfs_cpt_number(ptlrpc_hr.hr_cpt_table);
287 hrp = ptlrpc_hr.hr_partitions[rotor];
290 rotor = hrp->hrp_rotor++;
291 return &hrp->hrp_thrs[rotor % hrp->hrp_nthrs];
295 * Dispatch all replies accumulated in the batch to one from
296 * dedicated reply handling threads.
300 static void rs_batch_dispatch(struct rs_batch *b)
302 if (b->rsb_n_replies != 0) {
303 struct ptlrpc_hr_thread *hrt;
305 hrt = ptlrpc_hr_select(b->rsb_svcpt);
307 spin_lock(&hrt->hrt_lock);
308 list_splice_init(&b->rsb_replies, &hrt->hrt_queue);
309 spin_unlock(&hrt->hrt_lock);
311 wake_up(&hrt->hrt_waitq);
312 b->rsb_n_replies = 0;
317 * Add a reply to a batch.
318 * Add one reply object to a batch, schedule batched replies if overload.
323 static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs)
325 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
327 if (svcpt != b->rsb_svcpt || b->rsb_n_replies >= MAX_SCHEDULED) {
328 if (b->rsb_svcpt != NULL) {
329 rs_batch_dispatch(b);
330 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
332 spin_lock(&svcpt->scp_rep_lock);
333 b->rsb_svcpt = svcpt;
335 spin_lock(&rs->rs_lock);
336 rs->rs_scheduled_ever = 1;
337 if (rs->rs_scheduled == 0) {
338 list_move(&rs->rs_list, &b->rsb_replies);
339 rs->rs_scheduled = 1;
342 rs->rs_committed = 1;
343 spin_unlock(&rs->rs_lock);
347 * Reply batch finalization.
348 * Dispatch remaining replies from the batch
349 * and release remaining spinlock.
353 static void rs_batch_fini(struct rs_batch *b)
355 if (b->rsb_svcpt != NULL) {
356 rs_batch_dispatch(b);
357 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
361 #define DECLARE_RS_BATCH(b) struct rs_batch b
365 * Put reply state into a queue for processing because we received
366 * ACK from the client
368 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
370 struct ptlrpc_hr_thread *hrt;
374 LASSERT(list_empty(&rs->rs_list));
376 hrt = ptlrpc_hr_select(rs->rs_svcpt);
378 spin_lock(&hrt->hrt_lock);
379 list_add_tail(&rs->rs_list, &hrt->hrt_queue);
380 spin_unlock(&hrt->hrt_lock);
382 wake_up(&hrt->hrt_waitq);
386 void ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs)
390 assert_spin_locked(&rs->rs_svcpt->scp_rep_lock);
391 assert_spin_locked(&rs->rs_lock);
392 LASSERT(rs->rs_difficult);
393 rs->rs_scheduled_ever = 1; /* flag any notification attempt */
395 if (rs->rs_scheduled) { /* being set up or already notified */
400 rs->rs_scheduled = 1;
401 list_del_init(&rs->rs_list);
402 ptlrpc_dispatch_difficult_reply(rs);
405 EXPORT_SYMBOL(ptlrpc_schedule_difficult_reply);
407 void ptlrpc_commit_replies(struct obd_export *exp)
409 struct ptlrpc_reply_state *rs, *nxt;
410 DECLARE_RS_BATCH(batch);
414 rs_batch_init(&batch);
416 * Find any replies that have been committed and get their service
417 * to attend to complete them.
420 /* CAVEAT EMPTOR: spinlock ordering!!! */
421 spin_lock(&exp->exp_uncommitted_replies_lock);
422 list_for_each_entry_safe(rs, nxt, &exp->exp_uncommitted_replies,
424 LASSERT(rs->rs_difficult);
425 /* VBR: per-export last_committed */
426 LASSERT(rs->rs_export);
427 if (rs->rs_transno <= exp->exp_last_committed) {
428 list_del_init(&rs->rs_obd_list);
429 rs_batch_add(&batch, rs);
432 spin_unlock(&exp->exp_uncommitted_replies_lock);
433 rs_batch_fini(&batch);
437 static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt)
439 struct ptlrpc_request_buffer_desc *rqbd;
444 spin_lock(&svcpt->scp_lock);
446 if (list_empty(&svcpt->scp_rqbd_idle)) {
447 spin_unlock(&svcpt->scp_lock);
451 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
452 struct ptlrpc_request_buffer_desc,
455 /* assume we will post successfully */
456 svcpt->scp_nrqbds_posted++;
457 list_move(&rqbd->rqbd_list, &svcpt->scp_rqbd_posted);
459 spin_unlock(&svcpt->scp_lock);
461 rc = ptlrpc_register_rqbd(rqbd);
468 spin_lock(&svcpt->scp_lock);
470 svcpt->scp_nrqbds_posted--;
471 list_move_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
474 * Don't complain if no request buffers are posted right now; LNET
475 * won't drop requests because we set the portal lazy!
478 spin_unlock(&svcpt->scp_lock);
483 static void ptlrpc_at_timer(cfs_timer_cb_arg_t data)
485 struct ptlrpc_service_part *svcpt;
487 svcpt = cfs_from_timer(svcpt, data, scp_at_timer);
489 svcpt->scp_at_check = 1;
490 svcpt->scp_at_checktime = ktime_get();
491 wake_up(&svcpt->scp_waitq);
494 static void ptlrpc_server_nthreads_check(struct ptlrpc_service *svc,
495 struct ptlrpc_service_conf *conf)
497 struct ptlrpc_service_thr_conf *tc = &conf->psc_thr;
504 * Common code for estimating & validating threads number.
505 * CPT affinity service could have percpt thread-pool instead
506 * of a global thread-pool, which means user might not always
507 * get the threads number they give it in conf::tc_nthrs_user
508 * even they did set. It's because we need to validate threads
509 * number for each CPT to guarantee each pool will have enough
510 * threads to keep the service healthy.
512 init = PTLRPC_NTHRS_INIT + (svc->srv_ops.so_hpreq_handler != NULL);
513 init = max_t(int, init, tc->tc_nthrs_init);
516 * NB: please see comments in lustre_lnet.h for definition
517 * details of these members
519 LASSERT(tc->tc_nthrs_max != 0);
521 if (tc->tc_nthrs_user != 0) {
523 * In case there is a reason to test a service with many
524 * threads, we give a less strict check here, it can
525 * be up to 8 * nthrs_max
527 total = min(tc->tc_nthrs_max * 8, tc->tc_nthrs_user);
528 nthrs = total / svc->srv_ncpts;
529 init = max(init, nthrs);
533 total = tc->tc_nthrs_max;
534 if (tc->tc_nthrs_base == 0) {
536 * don't care about base threads number per partition,
537 * this is most for non-affinity service
539 nthrs = total / svc->srv_ncpts;
543 nthrs = tc->tc_nthrs_base;
544 if (svc->srv_ncpts == 1) {
548 * NB: Increase the base number if it's single partition
549 * and total number of cores/HTs is larger or equal to 4.
550 * result will always < 2 * nthrs_base
552 weight = cfs_cpt_weight(svc->srv_cptable, CFS_CPT_ANY);
553 for (i = 1; (weight >> (i + 1)) != 0 && /* >= 4 cores/HTs */
554 (tc->tc_nthrs_base >> i) != 0; i++)
555 nthrs += tc->tc_nthrs_base >> i;
558 if (tc->tc_thr_factor != 0) {
559 int factor = tc->tc_thr_factor;
563 * User wants to increase number of threads with for
564 * each CPU core/HT, most likely the factor is larger than
565 * one thread/core because service threads are supposed to
566 * be blocked by lock or wait for IO.
569 * Amdahl's law says that adding processors wouldn't give
570 * a linear increasing of parallelism, so it's nonsense to
571 * have too many threads no matter how many cores/HTs
576 (topology_sibling_cpumask(smp_processor_id())) > 1) {
577 /* weight is # of HTs */
578 /* depress thread factor for hyper-thread */
579 factor = factor - (factor >> 1) + (factor >> 3);
583 weight = cfs_cpt_weight(svc->srv_cptable, 0);
585 for (; factor > 0 && weight > 0; factor--, weight -= fade)
586 nthrs += min(weight, fade) * factor;
589 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
590 nthrs = max(tc->tc_nthrs_base,
591 tc->tc_nthrs_max / svc->srv_ncpts);
594 nthrs = max(nthrs, tc->tc_nthrs_init);
595 svc->srv_nthrs_cpt_limit = nthrs;
596 svc->srv_nthrs_cpt_init = init;
598 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
600 "%s: This service may have more threads (%d) than the given soft limit (%d)\n",
601 svc->srv_name, nthrs * svc->srv_ncpts,
607 * Initialize percpt data for a service
609 static int ptlrpc_service_part_init(struct ptlrpc_service *svc,
610 struct ptlrpc_service_part *svcpt, int cpt)
612 struct ptlrpc_at_array *array;
617 svcpt->scp_cpt = cpt;
618 INIT_LIST_HEAD(&svcpt->scp_threads);
620 /* rqbd and incoming request queue */
621 spin_lock_init(&svcpt->scp_lock);
622 mutex_init(&svcpt->scp_mutex);
623 INIT_LIST_HEAD(&svcpt->scp_rqbd_idle);
624 INIT_LIST_HEAD(&svcpt->scp_rqbd_posted);
625 INIT_LIST_HEAD(&svcpt->scp_req_incoming);
626 init_waitqueue_head(&svcpt->scp_waitq);
627 /* history request & rqbd list */
628 INIT_LIST_HEAD(&svcpt->scp_hist_reqs);
629 INIT_LIST_HEAD(&svcpt->scp_hist_rqbds);
631 /* acitve requests and hp requests */
632 spin_lock_init(&svcpt->scp_req_lock);
635 spin_lock_init(&svcpt->scp_rep_lock);
636 INIT_LIST_HEAD(&svcpt->scp_rep_active);
637 INIT_LIST_HEAD(&svcpt->scp_rep_idle);
638 init_waitqueue_head(&svcpt->scp_rep_waitq);
639 atomic_set(&svcpt->scp_nreps_difficult, 0);
641 /* adaptive timeout */
642 spin_lock_init(&svcpt->scp_at_lock);
643 array = &svcpt->scp_at_array;
645 size = at_est2timeout(at_max);
646 array->paa_size = size;
647 array->paa_count = 0;
648 array->paa_deadline = -1;
650 /* allocate memory for scp_at_array (ptlrpc_at_array) */
651 OBD_CPT_ALLOC(array->paa_reqs_array,
652 svc->srv_cptable, cpt, sizeof(struct list_head) * size);
653 if (array->paa_reqs_array == NULL)
656 for (index = 0; index < size; index++)
657 INIT_LIST_HEAD(&array->paa_reqs_array[index]);
659 OBD_CPT_ALLOC(array->paa_reqs_count,
660 svc->srv_cptable, cpt, sizeof(__u32) * size);
661 if (array->paa_reqs_count == NULL)
664 cfs_timer_setup(&svcpt->scp_at_timer, ptlrpc_at_timer,
665 (unsigned long)svcpt, 0);
668 * At SOW, service time should be quick; 10s seems generous. If client
669 * timeout is less than this, we'll be sending an early reply.
671 at_init(&svcpt->scp_at_estimate, 10, 0);
673 /* assign this before call ptlrpc_grow_req_bufs */
674 svcpt->scp_service = svc;
675 /* Now allocate the request buffers, but don't post them now */
676 rc = ptlrpc_grow_req_bufs(svcpt, 0);
678 * We shouldn't be under memory pressure at startup, so
679 * fail if we can't allocate all our buffers at this time.
687 if (array->paa_reqs_count != NULL) {
688 OBD_FREE_PTR_ARRAY(array->paa_reqs_count, size);
689 array->paa_reqs_count = NULL;
692 if (array->paa_reqs_array != NULL) {
693 OBD_FREE_PTR_ARRAY(array->paa_reqs_array, array->paa_size);
694 array->paa_reqs_array = NULL;
701 * Initialize service on a given portal.
702 * This includes starting serving threads , allocating and posting rqbds and
705 struct ptlrpc_service *ptlrpc_register_service(struct ptlrpc_service_conf *conf,
707 struct dentry *debugfs_entry)
709 struct ptlrpc_service_cpt_conf *cconf = &conf->psc_cpt;
710 struct ptlrpc_service *service;
711 struct ptlrpc_service_part *svcpt;
712 struct cfs_cpt_table *cptable;
721 LASSERT(conf->psc_buf.bc_nbufs > 0);
722 LASSERT(conf->psc_buf.bc_buf_size >=
723 conf->psc_buf.bc_req_max_size + SPTLRPC_MAX_PAYLOAD);
724 LASSERT(conf->psc_thr.tc_ctx_tags != 0);
726 cptable = cconf->cc_cptable;
728 cptable = cfs_cpt_tab;
730 if (conf->psc_thr.tc_cpu_bind > 1) {
731 CERROR("%s: Invalid cpu bind value %d, only 1 or 0 allowed\n",
732 conf->psc_name, conf->psc_thr.tc_cpu_bind);
733 RETURN(ERR_PTR(-EINVAL));
736 if (!cconf->cc_affinity) {
739 ncpts = cfs_cpt_number(cptable);
740 if (cconf->cc_pattern != NULL) {
741 struct cfs_expr_list *el;
743 rc = cfs_expr_list_parse(cconf->cc_pattern,
744 strlen(cconf->cc_pattern),
747 CERROR("%s: invalid CPT pattern string: %s\n",
748 conf->psc_name, cconf->cc_pattern);
749 RETURN(ERR_PTR(-EINVAL));
752 rc = cfs_expr_list_values(el, ncpts, &cpts);
753 cfs_expr_list_free(el);
755 CERROR("%s: failed to parse CPT array %s: %d\n",
756 conf->psc_name, cconf->cc_pattern, rc);
758 OBD_FREE_PTR_ARRAY(cpts, ncpts);
759 RETURN(ERR_PTR(rc < 0 ? rc : -EINVAL));
765 OBD_ALLOC(service, offsetof(struct ptlrpc_service, srv_parts[ncpts]));
766 if (service == NULL) {
768 OBD_FREE_PTR_ARRAY(cpts, ncpts);
769 RETURN(ERR_PTR(-ENOMEM));
772 service->srv_cptable = cptable;
773 service->srv_cpts = cpts;
774 service->srv_ncpts = ncpts;
775 service->srv_cpt_bind = conf->psc_thr.tc_cpu_bind;
777 service->srv_cpt_bits = 0; /* it's zero already, easy to read... */
778 while ((1 << service->srv_cpt_bits) < cfs_cpt_number(cptable))
779 service->srv_cpt_bits++;
782 spin_lock_init(&service->srv_lock);
783 service->srv_name = conf->psc_name;
784 service->srv_watchdog_factor = conf->psc_watchdog_factor;
785 INIT_LIST_HEAD(&service->srv_list); /* for safty of cleanup */
787 /* buffer configuration */
788 service->srv_nbuf_per_group = test_req_buffer_pressure ?
789 1 : conf->psc_buf.bc_nbufs;
790 /* do not limit max number of rqbds by default */
791 service->srv_nrqbds_max = 0;
793 service->srv_max_req_size = conf->psc_buf.bc_req_max_size +
795 service->srv_buf_size = conf->psc_buf.bc_buf_size;
796 service->srv_rep_portal = conf->psc_buf.bc_rep_portal;
797 service->srv_req_portal = conf->psc_buf.bc_req_portal;
799 /* With slab/alloc_pages buffer size will be rounded up to 2^n */
800 if (service->srv_buf_size & (service->srv_buf_size - 1)) {
801 int round = size_roundup_power2(service->srv_buf_size);
803 service->srv_buf_size = round;
806 /* Increase max reply size to next power of two */
807 service->srv_max_reply_size = 1;
808 while (service->srv_max_reply_size <
809 conf->psc_buf.bc_rep_max_size + SPTLRPC_MAX_PAYLOAD)
810 service->srv_max_reply_size <<= 1;
812 service->srv_thread_name = conf->psc_thr.tc_thr_name;
813 service->srv_ctx_tags = conf->psc_thr.tc_ctx_tags;
814 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
815 service->srv_ops = conf->psc_ops;
817 for (i = 0; i < ncpts; i++) {
818 if (!cconf->cc_affinity)
821 cpt = cpts != NULL ? cpts[i] : i;
823 OBD_CPT_ALLOC(svcpt, cptable, cpt, sizeof(*svcpt));
825 GOTO(failed, rc = -ENOMEM);
827 service->srv_parts[i] = svcpt;
828 rc = ptlrpc_service_part_init(service, svcpt, cpt);
833 ptlrpc_server_nthreads_check(service, conf);
835 rc = LNetSetLazyPortal(service->srv_req_portal);
838 mutex_lock(&ptlrpc_all_services_mutex);
839 list_add(&service->srv_list, &ptlrpc_all_services);
840 mutex_unlock(&ptlrpc_all_services_mutex);
843 rc = ptlrpc_sysfs_register_service(parent, service);
848 if (debugfs_entry != NULL)
849 ptlrpc_ldebugfs_register_service(debugfs_entry, service);
851 rc = ptlrpc_service_nrs_setup(service);
855 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
856 service->srv_name, service->srv_req_portal);
858 rc = ptlrpc_start_threads(service);
860 CERROR("Failed to start threads for service %s: %d\n",
861 service->srv_name, rc);
867 ptlrpc_unregister_service(service);
870 EXPORT_SYMBOL(ptlrpc_register_service);
873 * to actually free the request, must be called without holding svc_lock.
874 * note it's caller's responsibility to unlink req->rq_list.
876 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
878 LASSERT(atomic_read(&req->rq_refcount) == 0);
879 LASSERT(list_empty(&req->rq_timed_list));
882 * DEBUG_REQ() assumes the reply state of a request with a valid
883 * ref will not be destroyed until that reference is dropped.
885 ptlrpc_req_drop_rs(req);
887 sptlrpc_svc_ctx_decref(req);
889 if (req != &req->rq_rqbd->rqbd_req) {
891 * NB request buffers use an embedded
892 * req if the incoming req unlinked the
893 * MD; this isn't one of them!
895 ptlrpc_request_cache_free(req);
900 * drop a reference count of the request. if it reaches 0, we either
901 * put it into history list, or free it immediately.
903 void ptlrpc_server_drop_request(struct ptlrpc_request *req)
905 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
906 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
907 struct ptlrpc_service *svc = svcpt->scp_service;
910 if (!atomic_dec_and_test(&req->rq_refcount))
913 if (req->rq_session.lc_state == LCS_ENTERED) {
914 lu_context_exit(&req->rq_session);
915 lu_context_fini(&req->rq_session);
918 if (req->rq_at_linked) {
919 spin_lock(&svcpt->scp_at_lock);
921 * recheck with lock, in case it's unlinked by
922 * ptlrpc_at_check_timed()
924 if (likely(req->rq_at_linked))
925 ptlrpc_at_remove_timed(req);
926 spin_unlock(&svcpt->scp_at_lock);
929 LASSERT(list_empty(&req->rq_timed_list));
931 /* finalize request */
932 if (req->rq_export) {
933 class_export_put(req->rq_export);
934 req->rq_export = NULL;
937 spin_lock(&svcpt->scp_lock);
939 list_add(&req->rq_list, &rqbd->rqbd_reqs);
941 refcount = --(rqbd->rqbd_refcount);
943 /* request buffer is now idle: add to history */
944 list_move_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds);
945 svcpt->scp_hist_nrqbds++;
949 * I expect only about 1 or 2 rqbds need to be recycled here
951 while (svcpt->scp_hist_nrqbds > svc->srv_hist_nrqbds_cpt_max) {
952 rqbd = list_entry(svcpt->scp_hist_rqbds.next,
953 struct ptlrpc_request_buffer_desc,
956 list_del(&rqbd->rqbd_list);
957 svcpt->scp_hist_nrqbds--;
960 * remove rqbd's reqs from svc's req history while
961 * I've got the service lock
963 list_for_each_entry(req, &rqbd->rqbd_reqs, rq_list) {
964 /* Track the highest culled req seq */
965 if (req->rq_history_seq >
966 svcpt->scp_hist_seq_culled) {
967 svcpt->scp_hist_seq_culled =
970 list_del(&req->rq_history_list);
973 spin_unlock(&svcpt->scp_lock);
975 while ((req = list_first_entry_or_null(
977 struct ptlrpc_request, rq_list))) {
978 list_del(&req->rq_list);
979 ptlrpc_server_free_request(req);
982 spin_lock(&svcpt->scp_lock);
984 * now all reqs including the embedded req has been
985 * disposed, schedule request buffer for re-use
986 * or free it to drain some in excess.
988 LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) == 0);
989 if (svcpt->scp_nrqbds_posted >=
990 svc->srv_nbuf_per_group ||
991 (svc->srv_nrqbds_max != 0 &&
992 svcpt->scp_nrqbds_total > svc->srv_nrqbds_max) ||
993 test_req_buffer_pressure) {
994 /* like in ptlrpc_free_rqbd() */
995 svcpt->scp_nrqbds_total--;
996 OBD_FREE_LARGE(rqbd->rqbd_buffer,
1000 list_add_tail(&rqbd->rqbd_list,
1001 &svcpt->scp_rqbd_idle);
1005 spin_unlock(&svcpt->scp_lock);
1006 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
1007 /* If we are low on memory, we are not interested in history */
1008 list_del(&req->rq_list);
1009 list_del_init(&req->rq_history_list);
1011 /* Track the highest culled req seq */
1012 if (req->rq_history_seq > svcpt->scp_hist_seq_culled)
1013 svcpt->scp_hist_seq_culled = req->rq_history_seq;
1015 spin_unlock(&svcpt->scp_lock);
1017 ptlrpc_server_free_request(req);
1019 spin_unlock(&svcpt->scp_lock);
1023 static void ptlrpc_add_exp_list_nolock(struct ptlrpc_request *req,
1024 struct obd_export *export, bool hp)
1026 __u16 tag = lustre_msg_get_tag(req->rq_reqmsg);
1029 list_add(&req->rq_exp_list, &export->exp_hp_rpcs);
1031 list_add(&req->rq_exp_list, &export->exp_reg_rpcs);
1032 if (tag && export->exp_used_slots)
1033 set_bit(tag - 1, export->exp_used_slots);
1036 static void ptlrpc_del_exp_list(struct ptlrpc_request *req)
1038 __u16 tag = lustre_msg_get_tag(req->rq_reqmsg);
1040 spin_lock(&req->rq_export->exp_rpc_lock);
1041 list_del_init(&req->rq_exp_list);
1042 if (tag && !req->rq_obsolete && req->rq_export->exp_used_slots)
1043 clear_bit(tag - 1, req->rq_export->exp_used_slots);
1044 spin_unlock(&req->rq_export->exp_rpc_lock);
1047 /** Change request export and move hp request from old export to new */
1048 void ptlrpc_request_change_export(struct ptlrpc_request *req,
1049 struct obd_export *export)
1051 if (req->rq_export != NULL) {
1052 LASSERT(!list_empty(&req->rq_exp_list));
1053 /* remove rq_exp_list from last export */
1054 ptlrpc_del_exp_list(req);
1055 /* export has one reference already, so it's safe to
1056 * add req to export queue here and get another
1057 * reference for request later
1059 spin_lock(&export->exp_rpc_lock);
1060 ptlrpc_add_exp_list_nolock(req, export, req->rq_ops != NULL);
1061 spin_unlock(&export->exp_rpc_lock);
1063 class_export_rpc_dec(req->rq_export);
1064 class_export_put(req->rq_export);
1067 /* request takes one export refcount */
1068 req->rq_export = class_export_get(export);
1069 class_export_rpc_inc(export);
1073 * to finish a request: stop sending more early replies, and release
1076 static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt,
1077 struct ptlrpc_request *req)
1079 ptlrpc_server_hpreq_fini(req);
1081 ptlrpc_server_drop_request(req);
1085 * to finish an active request: stop sending more early replies, and release
1086 * the request. should be called after we finished handling the request.
1088 static void ptlrpc_server_finish_active_request(
1089 struct ptlrpc_service_part *svcpt,
1090 struct ptlrpc_request *req)
1092 spin_lock(&svcpt->scp_req_lock);
1093 ptlrpc_nrs_req_stop_nolock(req);
1094 svcpt->scp_nreqs_active--;
1096 svcpt->scp_nhreqs_active--;
1097 spin_unlock(&svcpt->scp_req_lock);
1099 ptlrpc_nrs_req_finalize(req);
1101 if (req->rq_export != NULL)
1102 class_export_rpc_dec(req->rq_export);
1104 ptlrpc_server_finish_request(svcpt, req);
1108 * This function makes sure dead exports are evicted in a timely manner.
1109 * This function is only called when some export receives a message (i.e.,
1110 * the network is up.)
1112 void ptlrpc_update_export_timer(struct obd_export *exp, time64_t extra_delay)
1114 struct obd_export *oldest_exp;
1115 time64_t oldest_time, new_time;
1122 * Compensate for slow machines, etc, by faking our request time
1123 * into the future. Although this can break the strict time-ordering
1124 * of the list, we can be really lazy here - we don't have to evict
1125 * at the exact right moment. Eventually, all silent exports
1126 * will make it to the top of the list.
1129 /* Do not pay attention on 1sec or smaller renewals. */
1130 new_time = ktime_get_real_seconds() + extra_delay;
1131 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
1134 exp->exp_last_request_time = new_time;
1137 * exports may get disconnected from the chain even though the
1138 * export has references, so we must keep the spin lock while
1139 * manipulating the lists
1141 spin_lock(&exp->exp_obd->obd_dev_lock);
1143 if (list_empty(&exp->exp_obd_chain_timed)) {
1144 /* this one is not timed */
1145 spin_unlock(&exp->exp_obd->obd_dev_lock);
1149 list_move_tail(&exp->exp_obd_chain_timed,
1150 &exp->exp_obd->obd_exports_timed);
1152 oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
1153 struct obd_export, exp_obd_chain_timed);
1154 oldest_time = oldest_exp->exp_last_request_time;
1155 spin_unlock(&exp->exp_obd->obd_dev_lock);
1157 if (exp->exp_obd->obd_recovering) {
1158 /* be nice to everyone during recovery */
1163 /* Note - racing to start/reset the obd_eviction timer is safe */
1164 if (exp->exp_obd->obd_eviction_timer == 0) {
1165 /* Check if the oldest entry is expired. */
1166 if (ktime_get_real_seconds() >
1167 oldest_time + PING_EVICT_TIMEOUT + extra_delay) {
1169 * We need a second timer, in case the net was down and
1170 * it just came back. Since the pinger may skip every
1171 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
1172 * we better wait for 3.
1174 exp->exp_obd->obd_eviction_timer =
1175 ktime_get_real_seconds() + 3 * PING_INTERVAL;
1176 CDEBUG(D_HA, "%s: Think about evicting %s from %lld\n",
1177 exp->exp_obd->obd_name,
1178 obd_export_nid2str(oldest_exp), oldest_time);
1181 if (ktime_get_real_seconds() >
1182 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
1184 * The evictor won't evict anyone who we've heard from
1185 * recently, so we don't have to check before we start
1188 if (!ping_evictor_wake(exp))
1189 exp->exp_obd->obd_eviction_timer = 0;
1197 * Sanity check request \a req.
1198 * Return 0 if all is ok, error code otherwise.
1200 static int ptlrpc_check_req(struct ptlrpc_request *req)
1202 struct obd_device *obd = req->rq_export->exp_obd;
1205 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
1206 req->rq_export->exp_conn_cnt)) {
1207 DEBUG_REQ(D_RPCTRACE, req,
1208 "DROPPING req from old connection %d < %d",
1209 lustre_msg_get_conn_cnt(req->rq_reqmsg),
1210 req->rq_export->exp_conn_cnt);
1213 if (unlikely(obd == NULL || obd->obd_fail)) {
1215 * Failing over, don't handle any more reqs,
1216 * send error response instead.
1218 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
1219 req, (obd != NULL) ? obd->obd_name : "unknown");
1221 } else if (lustre_msg_get_flags(req->rq_reqmsg) &
1222 (MSG_REPLAY | MSG_REQ_REPLAY_DONE) &&
1223 !obd->obd_recovering) {
1224 DEBUG_REQ(D_ERROR, req,
1225 "Invalid replay without recovery");
1226 class_fail_export(req->rq_export);
1228 } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 &&
1229 !obd->obd_recovering) {
1230 DEBUG_REQ(D_ERROR, req,
1231 "Invalid req with transno %llu without recovery",
1232 lustre_msg_get_transno(req->rq_reqmsg));
1233 class_fail_export(req->rq_export);
1237 if (unlikely(rc < 0)) {
1238 req->rq_status = rc;
1244 static void ptlrpc_at_set_timer(struct ptlrpc_service_part *svcpt)
1246 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1249 if (array->paa_count == 0) {
1250 del_timer(&svcpt->scp_at_timer);
1254 /* Set timer for closest deadline */
1255 next = array->paa_deadline - ktime_get_real_seconds() -
1258 ptlrpc_at_timer(cfs_timer_cb_arg(svcpt, scp_at_timer));
1260 mod_timer(&svcpt->scp_at_timer,
1261 jiffies + nsecs_to_jiffies(next * NSEC_PER_SEC));
1262 CDEBUG(D_INFO, "armed %s at %+llds\n",
1263 svcpt->scp_service->srv_name, next);
1267 /* Add rpc to early reply check list */
1268 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
1270 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1271 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1272 struct ptlrpc_request *rq = NULL;
1278 if (req->rq_no_reply)
1281 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
1284 spin_lock(&svcpt->scp_at_lock);
1285 LASSERT(list_empty(&req->rq_timed_list));
1287 div_u64_rem(req->rq_deadline, array->paa_size, &index);
1288 if (array->paa_reqs_count[index] > 0) {
1290 * latest rpcs will have the latest deadlines in the list,
1291 * so search backward.
1293 list_for_each_entry_reverse(rq, &array->paa_reqs_array[index],
1295 if (req->rq_deadline >= rq->rq_deadline) {
1296 list_add(&req->rq_timed_list,
1297 &rq->rq_timed_list);
1303 /* Add the request at the head of the list */
1304 if (list_empty(&req->rq_timed_list))
1305 list_add(&req->rq_timed_list, &array->paa_reqs_array[index]);
1307 spin_lock(&req->rq_lock);
1308 req->rq_at_linked = 1;
1309 spin_unlock(&req->rq_lock);
1310 req->rq_at_index = index;
1311 array->paa_reqs_count[index]++;
1313 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
1314 array->paa_deadline = req->rq_deadline;
1315 ptlrpc_at_set_timer(svcpt);
1317 spin_unlock(&svcpt->scp_at_lock);
1322 static void ptlrpc_at_remove_timed(struct ptlrpc_request *req)
1324 struct ptlrpc_at_array *array;
1326 array = &req->rq_rqbd->rqbd_svcpt->scp_at_array;
1328 /* NB: must call with hold svcpt::scp_at_lock */
1329 LASSERT(!list_empty(&req->rq_timed_list));
1330 list_del_init(&req->rq_timed_list);
1332 spin_lock(&req->rq_lock);
1333 req->rq_at_linked = 0;
1334 spin_unlock(&req->rq_lock);
1336 array->paa_reqs_count[req->rq_at_index]--;
1341 * Attempt to extend the request deadline by sending an early reply to the
1344 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
1346 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1347 struct ptlrpc_request *reqcopy;
1348 struct lustre_msg *reqmsg;
1349 timeout_t olddl = req->rq_deadline - ktime_get_real_seconds();
1355 if (CFS_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_RECONNECT)) {
1356 /* don't send early reply */
1361 * deadline is when the client expects us to reply, margin is the
1362 * difference between clients' and servers' expectations
1364 DEBUG_REQ(D_ADAPTTO, req,
1365 "%ssending early reply (deadline %+ds, margin %+ds) for %d+%d",
1366 AT_OFF ? "AT off - not " : "",
1367 olddl, olddl - at_get(&svcpt->scp_at_estimate),
1368 at_get(&svcpt->scp_at_estimate), at_extra);
1374 /* below message is checked in replay-ost-single.sh test_9 */
1375 DEBUG_REQ(D_WARNING, req,
1376 "Already past deadline (%+ds), not sending early reply. Consider increasing at_early_margin (%d)?",
1377 olddl, at_early_margin);
1379 /* Return an error so we're not re-added to the timed list. */
1383 if ((lustre_msghdr_get_flags(req->rq_reqmsg) &
1384 MSGHDR_AT_SUPPORT) == 0) {
1385 DEBUG_REQ(D_INFO, req,
1386 "Wanted to ask client for more time, but no AT support");
1390 if (req->rq_export &&
1391 lustre_msg_get_flags(req->rq_reqmsg) &
1392 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1393 struct obd_device *obd_exp = req->rq_export->exp_obd;
1396 * During recovery, we don't want to send too many early
1397 * replies, but on the other hand we want to make sure the
1398 * client has enough time to resend if the rpc is lost. So
1399 * during the recovery period send at least 4 early replies,
1400 * spacing them every at_extra if we can. at_estimate should
1401 * always equal this fixed value during recovery.
1405 * Don't account request processing time into AT history
1406 * during recovery, it is not service time we need but
1407 * includes also waiting time for recovering clients
1409 newdl = min_t(time64_t, at_extra,
1410 obd_exp->obd_recovery_timeout / 4) +
1411 ktime_get_real_seconds();
1414 * We want to extend the request deadline by at_extra seconds,
1415 * so we set our service estimate to reflect how much time has
1416 * passed since this request arrived plus an additional
1417 * at_extra seconds. The client will calculate the new deadline
1418 * based on this service estimate (plus some additional time to
1419 * account for network latency). See ptlrpc_at_recv_early_reply
1421 at_measured(&svcpt->scp_at_estimate, at_extra +
1422 ktime_get_real_seconds() -
1423 req->rq_arrival_time.tv_sec);
1424 newdl = req->rq_arrival_time.tv_sec +
1425 at_get(&svcpt->scp_at_estimate);
1429 * Check to see if we've actually increased the deadline -
1430 * we may be past adaptive_max
1432 if (req->rq_deadline >= newdl) {
1433 DEBUG_REQ(D_WARNING, req,
1434 "Could not add any time (%d/%lld), not sending early reply",
1435 olddl, newdl - ktime_get_real_seconds());
1439 reqcopy = ptlrpc_request_cache_alloc(GFP_NOFS);
1440 if (reqcopy == NULL)
1442 OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1444 GOTO(out_free, rc = -ENOMEM);
1447 reqcopy->rq_reply_state = NULL;
1448 reqcopy->rq_rep_swab_mask = 0;
1449 reqcopy->rq_pack_bulk = 0;
1450 reqcopy->rq_pack_udesc = 0;
1451 reqcopy->rq_packed_final = 0;
1452 sptlrpc_svc_ctx_addref(reqcopy);
1453 /* We only need the reqmsg for the magic */
1454 reqcopy->rq_reqmsg = reqmsg;
1455 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1458 * tgt_brw_read() and tgt_brw_write() may have decided not to reply.
1459 * Without this check, we would fail the rq_no_reply assertion in
1460 * ptlrpc_send_reply().
1462 if (reqcopy->rq_no_reply)
1463 GOTO(out, rc = -ETIMEDOUT);
1465 LASSERT(atomic_read(&req->rq_refcount));
1466 /* if it is last refcount then early reply isn't needed */
1467 if (atomic_read(&req->rq_refcount) == 1) {
1468 DEBUG_REQ(D_ADAPTTO, reqcopy,
1469 "Normal reply already sent, abort early reply");
1470 GOTO(out, rc = -EINVAL);
1473 /* Connection ref */
1474 reqcopy->rq_export = class_conn2export(
1475 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1476 if (reqcopy->rq_export == NULL)
1477 GOTO(out, rc = -ENODEV);
1480 class_export_rpc_inc(reqcopy->rq_export);
1481 if (reqcopy->rq_export->exp_obd &&
1482 reqcopy->rq_export->exp_obd->obd_fail)
1483 GOTO(out_put, rc = -ENODEV);
1485 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1489 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1492 /* Adjust our own deadline to what we told the client */
1493 req->rq_deadline = newdl;
1494 req->rq_early_count++; /* number sent, server side */
1496 DEBUG_REQ(D_ERROR, req, "Early reply send failed: rc = %d", rc);
1500 * Free the (early) reply state from lustre_pack_reply.
1501 * (ptlrpc_send_reply takes it's own rs ref, so this is safe here)
1503 ptlrpc_req_drop_rs(reqcopy);
1506 class_export_rpc_dec(reqcopy->rq_export);
1507 class_export_put(reqcopy->rq_export);
1509 sptlrpc_svc_ctx_decref(reqcopy);
1510 OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1512 ptlrpc_request_cache_free(reqcopy);
1517 * Send early replies to everybody expiring within at_early_margin
1518 * asking for at_extra time
1520 static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1522 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1523 struct ptlrpc_request *rq, *n;
1524 LIST_HEAD(work_list);
1527 time64_t now = ktime_get_real_seconds();
1529 int first, counter = 0;
1532 spin_lock(&svcpt->scp_at_lock);
1533 if (svcpt->scp_at_check == 0) {
1534 spin_unlock(&svcpt->scp_at_lock);
1537 delay_ms = ktime_ms_delta(ktime_get(), svcpt->scp_at_checktime);
1538 svcpt->scp_at_check = 0;
1540 if (array->paa_count == 0) {
1541 spin_unlock(&svcpt->scp_at_lock);
1545 /* The timer went off, but maybe the nearest rpc already completed. */
1546 first = array->paa_deadline - now;
1547 if (first > at_early_margin) {
1548 /* We've still got plenty of time. Reset the timer. */
1549 ptlrpc_at_set_timer(svcpt);
1550 spin_unlock(&svcpt->scp_at_lock);
1555 * We're close to a timeout, and we don't know how much longer the
1556 * server will take. Send early replies to everyone expiring soon.
1559 div_u64_rem(array->paa_deadline, array->paa_size, &index);
1560 count = array->paa_count;
1562 count -= array->paa_reqs_count[index];
1563 list_for_each_entry_safe(rq, n,
1564 &array->paa_reqs_array[index],
1566 if (rq->rq_deadline > now + at_early_margin) {
1567 /* update the earliest deadline */
1568 if (deadline == -1 ||
1569 rq->rq_deadline < deadline)
1570 deadline = rq->rq_deadline;
1575 * ptlrpc_server_drop_request() may drop
1576 * refcount to 0 already. Let's check this and
1577 * don't add entry to work_list
1579 if (likely(atomic_inc_not_zero(&rq->rq_refcount))) {
1580 ptlrpc_at_remove_timed(rq);
1581 list_add(&rq->rq_timed_list, &work_list);
1583 ptlrpc_at_remove_timed(rq);
1589 if (++index >= array->paa_size)
1592 array->paa_deadline = deadline;
1593 /* we have a new earliest deadline, restart the timer */
1594 ptlrpc_at_set_timer(svcpt);
1596 spin_unlock(&svcpt->scp_at_lock);
1599 "timeout in %+ds, asking for %d secs on %d early replies\n",
1600 first, at_extra, counter);
1603 * We're already past request deadlines before we even get a
1604 * chance to send early replies
1606 LCONSOLE_WARN("%s: This server is not able to keep up with request traffic (cpu-bound).\n",
1607 svcpt->scp_service->srv_name);
1608 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, delay=%lldms\n",
1609 counter, svcpt->scp_nreqs_incoming,
1610 svcpt->scp_nreqs_active,
1611 at_get(&svcpt->scp_at_estimate), delay_ms);
1615 * we took additional refcount so entries can't be deleted from list, no
1618 while (!list_empty(&work_list)) {
1619 rq = list_entry(work_list.next, struct ptlrpc_request,
1621 list_del_init(&rq->rq_timed_list);
1623 if (ptlrpc_at_send_early_reply(rq) == 0)
1624 ptlrpc_at_add_timed(rq);
1626 ptlrpc_server_drop_request(rq);
1629 RETURN(1); /* return "did_something" for liblustre */
1633 * Check if we are already handling earlier incarnation of this request.
1634 * Called under &req->rq_export->exp_rpc_lock locked
1636 static struct ptlrpc_request*
1637 ptlrpc_server_check_resend_in_progress(struct ptlrpc_request *req)
1639 struct ptlrpc_request *tmp = NULL;
1641 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ||
1642 (atomic_read(&req->rq_export->exp_rpc_count) == 0))
1646 * bulk request are aborted upon reconnect, don't try to
1649 if (req->rq_bulk_write || req->rq_bulk_read)
1653 * This list should not be longer than max_requests in
1654 * flights on the client, so it is not all that long.
1655 * Also we only hit this codepath in case of a resent
1656 * request which makes it even more rarely hit
1658 list_for_each_entry(tmp, &req->rq_export->exp_reg_rpcs,
1660 /* Found duplicate one */
1661 if (tmp->rq_xid == req->rq_xid)
1664 list_for_each_entry(tmp, &req->rq_export->exp_hp_rpcs,
1666 /* Found duplicate one */
1667 if (tmp->rq_xid == req->rq_xid)
1673 DEBUG_REQ(D_HA, req, "Found duplicate req in processing");
1674 DEBUG_REQ(D_HA, tmp, "Request being processed");
1678 #ifdef HAVE_SERVER_SUPPORT
1679 static void ptlrpc_server_mark_obsolete(struct ptlrpc_request *req)
1681 req->rq_obsolete = 1;
1685 ptlrpc_server_mark_in_progress_obsolete(struct ptlrpc_request *req)
1687 struct ptlrpc_request *tmp = NULL;
1690 if (!tgt_is_increasing_xid_client(req->rq_export) ||
1691 req->rq_export->exp_used_slots == NULL)
1694 tag = lustre_msg_get_tag(req->rq_reqmsg);
1698 if (!test_bit(tag - 1, req->rq_export->exp_used_slots))
1701 /* This list should not be longer than max_requests in
1702 * flights on the client, so it is not all that long.
1703 * Also we only hit this codepath in case of a resent
1704 * request which makes it even more rarely hit */
1705 list_for_each_entry(tmp, &req->rq_export->exp_reg_rpcs, rq_exp_list) {
1706 if (tag == lustre_msg_get_tag(tmp->rq_reqmsg) &&
1707 req->rq_xid > tmp->rq_xid)
1708 ptlrpc_server_mark_obsolete(tmp);
1711 list_for_each_entry(tmp, &req->rq_export->exp_hp_rpcs, rq_exp_list) {
1712 if (tag == lustre_msg_get_tag(tmp->rq_reqmsg) &&
1713 req->rq_xid > tmp->rq_xid)
1714 ptlrpc_server_mark_obsolete(tmp);
1720 * Check if a request should be assigned with a high priority.
1722 * \retval < 0: error occurred
1723 * 0: normal RPC request
1724 * +1: high priority request
1726 static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt,
1727 struct ptlrpc_request *req)
1732 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL) {
1733 rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
1740 if (req->rq_export != NULL && req->rq_ops != NULL) {
1742 * Perform request specific check. We should do this
1743 * check before the request is added into exp_hp_rpcs
1744 * list otherwise it may hit swab race at LU-1044.
1746 if (req->rq_ops->hpreq_check != NULL) {
1747 rc = req->rq_ops->hpreq_check(req);
1748 if (rc == -ESTALE) {
1749 req->rq_status = rc;
1753 * can only return error,
1754 * 0 for normal request,
1755 * or 1 for high priority request
1764 /** Remove the request from the export list. */
1765 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req)
1768 if (req->rq_export) {
1770 * refresh lock timeout again so that client has more
1771 * room to send lock cancel RPC.
1773 if (req->rq_ops && req->rq_ops->hpreq_fini)
1774 req->rq_ops->hpreq_fini(req);
1776 ptlrpc_del_exp_list(req);
1781 static int ptlrpc_hpreq_check(struct ptlrpc_request *req)
1786 static struct ptlrpc_hpreq_ops ptlrpc_hpreq_common = {
1787 .hpreq_check = ptlrpc_hpreq_check,
1790 /* Hi-Priority RPC check by RPC operation code. */
1791 int ptlrpc_hpreq_handler(struct ptlrpc_request *req)
1793 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1796 * Check for export to let only reconnects for not yet evicted
1797 * export to become a HP rpc.
1799 if ((req->rq_export != NULL) &&
1800 (opc == OBD_PING || opc == MDS_CONNECT || opc == OST_CONNECT))
1801 req->rq_ops = &ptlrpc_hpreq_common;
1805 EXPORT_SYMBOL(ptlrpc_hpreq_handler);
1807 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1808 struct ptlrpc_request *req)
1812 struct ptlrpc_request *orig;
1816 rc = ptlrpc_server_hpreq_init(svcpt, req);
1821 ptlrpc_nrs_req_initialize(svcpt, req, hp);
1823 while (req->rq_export != NULL) {
1824 struct obd_export *exp = req->rq_export;
1827 * do search for duplicated xid and the adding to the list
1830 spin_lock_bh(&exp->exp_rpc_lock);
1831 #ifdef HAVE_SERVER_SUPPORT
1832 ptlrpc_server_mark_in_progress_obsolete(req);
1834 orig = ptlrpc_server_check_resend_in_progress(req);
1835 if (orig && OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_RESEND_RACE)) {
1836 spin_unlock_bh(&exp->exp_rpc_lock);
1838 OBD_RACE(OBD_FAIL_PTLRPC_RESEND_RACE);
1839 msleep(4 * MSEC_PER_SEC);
1843 if (orig && likely(atomic_inc_not_zero(&orig->rq_refcount))) {
1846 spin_unlock_bh(&exp->exp_rpc_lock);
1849 * When the client resend request and the server has
1850 * the previous copy of it, we need to update deadlines,
1851 * to be sure that the client and the server have equal
1852 * request deadlines.
1855 spin_lock(&orig->rq_rqbd->rqbd_svcpt->scp_at_lock);
1856 linked = orig->rq_at_linked;
1858 ptlrpc_at_remove_timed(orig);
1859 spin_unlock(&orig->rq_rqbd->rqbd_svcpt->scp_at_lock);
1860 orig->rq_deadline = req->rq_deadline;
1862 ptlrpc_at_add_timed(orig);
1863 ptlrpc_server_drop_request(orig);
1864 ptlrpc_nrs_req_finalize(req);
1866 /* don't mark slot unused for resend in progress */
1867 req->rq_obsolete = 1;
1872 ptlrpc_add_exp_list_nolock(req, exp, hp || req->rq_ops != NULL);
1874 spin_unlock_bh(&exp->exp_rpc_lock);
1879 * the current thread is not the processing thread for this request
1880 * since that, but request is in exp_hp_list and can be find there.
1881 * Remove all relations between request and old thread.
1883 req->rq_svc_thread->t_env->le_ses = NULL;
1884 req->rq_svc_thread = NULL;
1885 req->rq_session.lc_thread = NULL;
1887 ptlrpc_nrs_req_add(svcpt, req, hp);
1893 * Allow to handle high priority request
1894 * User can call it w/o any lock but need to hold
1895 * ptlrpc_service_part::scp_req_lock to get reliable result
1897 static bool ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1900 int running = svcpt->scp_nthrs_running;
1902 if (!nrs_svcpt_has_hp(svcpt))
1908 if (ptlrpc_nrs_req_throttling_nolock(svcpt, true))
1911 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1912 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1913 /* leave just 1 thread for normal RPCs */
1914 running = PTLRPC_NTHRS_INIT;
1915 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1919 if (svcpt->scp_nreqs_active >= running - 1)
1922 if (svcpt->scp_nhreqs_active == 0)
1925 return !ptlrpc_nrs_req_pending_nolock(svcpt, false) ||
1926 svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1929 static bool ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1932 return ptlrpc_server_allow_high(svcpt, force) &&
1933 ptlrpc_nrs_req_pending_nolock(svcpt, true);
1937 * Only allow normal priority requests on a service that has a high-priority
1938 * queue if forced (i.e. cleanup), if there are other high priority requests
1939 * already being processed (i.e. those threads can service more high-priority
1940 * requests), or if there are enough idle threads that a later thread can do
1941 * a high priority request.
1942 * User can call it w/o any lock but need to hold
1943 * ptlrpc_service_part::scp_req_lock to get reliable result
1945 static bool ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1948 int running = svcpt->scp_nthrs_running;
1950 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1951 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1952 /* leave just 1 thread for normal RPCs */
1953 running = PTLRPC_NTHRS_INIT;
1954 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1961 if (ptlrpc_nrs_req_throttling_nolock(svcpt, false))
1964 if (svcpt->scp_nreqs_active < running - 2)
1967 if (svcpt->scp_nreqs_active >= running - 1)
1970 return svcpt->scp_nhreqs_active > 0 || !nrs_svcpt_has_hp(svcpt);
1973 static bool ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1976 return ptlrpc_server_allow_normal(svcpt, force) &&
1977 ptlrpc_nrs_req_pending_nolock(svcpt, false);
1981 * Returns true if there are requests available in incoming
1982 * request queue for processing and it is allowed to fetch them.
1983 * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1984 * to get reliable result
1985 * \see ptlrpc_server_allow_normal
1986 * \see ptlrpc_server_allow high
1989 bool ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt,
1992 return ptlrpc_server_high_pending(svcpt, force) ||
1993 ptlrpc_server_normal_pending(svcpt, force);
1997 * Fetch a request for processing from queue of unprocessed requests.
1998 * Favors high-priority requests.
1999 * Returns a pointer to fetched request.
2001 static struct ptlrpc_request *
2002 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force)
2004 struct ptlrpc_request *req = NULL;
2008 spin_lock(&svcpt->scp_req_lock);
2010 if (ptlrpc_server_high_pending(svcpt, force)) {
2011 req = ptlrpc_nrs_req_get_nolock(svcpt, true, force);
2013 svcpt->scp_hreq_count++;
2018 if (ptlrpc_server_normal_pending(svcpt, force)) {
2019 req = ptlrpc_nrs_req_get_nolock(svcpt, false, force);
2021 svcpt->scp_hreq_count = 0;
2026 spin_unlock(&svcpt->scp_req_lock);
2030 svcpt->scp_nreqs_active++;
2032 svcpt->scp_nhreqs_active++;
2034 spin_unlock(&svcpt->scp_req_lock);
2036 if (likely(req->rq_export))
2037 class_export_rpc_inc(req->rq_export);
2043 * Handle freshly incoming reqs, add to timed early reply list,
2044 * pass on to regular request queue.
2045 * All incoming requests pass through here before getting into
2046 * ptlrpc_server_handle_req later on.
2048 static int ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt,
2049 struct ptlrpc_thread *thread)
2051 struct ptlrpc_service *svc = svcpt->scp_service;
2052 struct ptlrpc_request *req;
2058 spin_lock(&svcpt->scp_lock);
2059 if (list_empty(&svcpt->scp_req_incoming)) {
2060 spin_unlock(&svcpt->scp_lock);
2064 req = list_entry(svcpt->scp_req_incoming.next,
2065 struct ptlrpc_request, rq_list);
2066 list_del_init(&req->rq_list);
2067 svcpt->scp_nreqs_incoming--;
2069 * Consider this still a "queued" request as far as stats are
2072 spin_unlock(&svcpt->scp_lock);
2074 /* go through security check/transform */
2075 rc = sptlrpc_svc_unwrap_request(req);
2079 case SECSVC_COMPLETE:
2080 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
2089 * for null-flavored rpc, msg has been unpacked by sptlrpc, although
2090 * redo it wouldn't be harmful.
2092 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
2093 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
2095 CERROR("error unpacking request: ptl %d from %s x%llu\n",
2096 svc->srv_req_portal, libcfs_id2str(req->rq_peer),
2102 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
2104 CERROR("error unpacking ptlrpc body: ptl %d from %s x %llu\n",
2105 svc->srv_req_portal, libcfs_id2str(req->rq_peer),
2110 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
2111 lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
2112 CERROR("drop incoming rpc opc %u, x%llu\n",
2113 cfs_fail_val, req->rq_xid);
2118 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
2119 CERROR("wrong packet type received (type=%u) from %s\n",
2120 lustre_msg_get_type(req->rq_reqmsg),
2121 libcfs_id2str(req->rq_peer));
2125 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2129 req->rq_bulk_write = 1;
2133 case MGS_CONFIG_READ:
2134 req->rq_bulk_read = 1;
2138 CDEBUG(D_RPCTRACE, "got req x%llu\n", req->rq_xid);
2140 req->rq_export = class_conn2export(
2141 lustre_msg_get_handle(req->rq_reqmsg));
2142 if (req->rq_export) {
2143 rc = ptlrpc_check_req(req);
2145 rc = sptlrpc_target_export_check(req->rq_export, req);
2147 DEBUG_REQ(D_ERROR, req,
2148 "DROPPING req with illegal security flavor");
2153 ptlrpc_update_export_timer(req->rq_export, 0);
2156 /* req_in handling should/must be fast */
2157 if (ktime_get_real_seconds() - req->rq_arrival_time.tv_sec > 5)
2158 DEBUG_REQ(D_WARNING, req, "Slow req_in handling %llds",
2159 ktime_get_real_seconds() -
2160 req->rq_arrival_time.tv_sec);
2162 /* Set rpc server deadline and add it to the timed list */
2163 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
2164 MSGHDR_AT_SUPPORT) ?
2165 /* The max time the client expects us to take */
2166 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
2168 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
2169 if (unlikely(deadline == 0)) {
2170 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
2174 /* Skip early reply */
2175 if (OBD_FAIL_PRECHECK(OBD_FAIL_MDS_RESEND))
2176 req->rq_deadline += obd_timeout;
2178 req->rq_svc_thread = thread;
2179 if (thread != NULL) {
2181 * initialize request session, it is needed for request
2182 * processing by target
2184 rc = lu_context_init(&req->rq_session, LCT_SERVER_SESSION |
2187 CERROR("%s: failure to initialize session: rc = %d\n",
2188 thread->t_name, rc);
2191 req->rq_session.lc_thread = thread;
2192 lu_context_enter(&req->rq_session);
2193 thread->t_env->le_ses = &req->rq_session;
2196 ptlrpc_at_add_timed(req);
2198 /* Move it over to the request processing queue */
2199 rc = ptlrpc_server_request_add(svcpt, req);
2203 wake_up(&svcpt->scp_waitq);
2207 ptlrpc_server_finish_request(svcpt, req);
2213 * Main incoming request handling logic.
2214 * Calls handler function from service to do actual processing.
2216 static int ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
2217 struct ptlrpc_thread *thread)
2219 struct ptlrpc_service *svc = svcpt->scp_service;
2220 struct ptlrpc_request *request;
2230 request = ptlrpc_server_request_get(svcpt, false);
2231 if (request == NULL)
2234 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
2235 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
2236 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
2237 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
2239 if (unlikely(fail_opc)) {
2240 if (request->rq_export && request->rq_ops)
2241 OBD_FAIL_TIMEOUT(fail_opc, 4);
2244 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
2246 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
2247 libcfs_debug_dumplog();
2249 work_start = ktime_get_real();
2250 arrived = timespec64_to_ktime(request->rq_arrival_time);
2251 timediff_usecs = ktime_us_delta(work_start, arrived);
2252 if (likely(svc->srv_stats != NULL)) {
2253 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
2255 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
2256 svcpt->scp_nreqs_incoming);
2257 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
2258 svcpt->scp_nreqs_active);
2259 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
2260 at_get(&svcpt->scp_at_estimate));
2263 if (likely(request->rq_export)) {
2264 if (unlikely(ptlrpc_check_req(request)))
2266 ptlrpc_update_export_timer(request->rq_export,
2267 div_u64(timediff_usecs,
2272 * Discard requests queued for longer than the deadline.
2273 * The deadline is increased if we send an early reply.
2275 if (ktime_get_real_seconds() > request->rq_deadline) {
2276 DEBUG_REQ(D_ERROR, request,
2277 "Dropping timed-out request from %s: deadline %lld/%llds ago",
2278 libcfs_id2str(request->rq_peer),
2279 request->rq_deadline -
2280 request->rq_arrival_time.tv_sec,
2281 ktime_get_real_seconds() - request->rq_deadline);
2286 "Handling RPC req@%p pname:cluuid+ref:pid:xid:nid:opc:job %s:%s+%d:%d:x%llu:%s:%d:%s\n",
2287 request, current->comm,
2288 (request->rq_export ?
2289 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2290 (request->rq_export ?
2291 refcount_read(&request->rq_export->exp_handle.h_ref) : -99),
2292 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
2293 libcfs_id2str(request->rq_peer),
2294 lustre_msg_get_opc(request->rq_reqmsg),
2295 lustre_msg_get_jobid(request->rq_reqmsg) ?: "");
2297 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
2298 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
2300 CDEBUG(D_NET, "got req %llu\n", request->rq_xid);
2302 /* re-assign request and sesson thread to the current one */
2303 request->rq_svc_thread = thread;
2304 if (thread != NULL) {
2305 LASSERT(request->rq_session.lc_thread == NULL);
2306 request->rq_session.lc_thread = thread;
2307 thread->t_env->le_ses = &request->rq_session;
2309 svc->srv_ops.so_req_handler(request);
2311 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
2314 if (unlikely(ktime_get_real_seconds() > request->rq_deadline)) {
2315 DEBUG_REQ(D_WARNING, request,
2316 "Request took longer than estimated (%lld/%llds); client may timeout",
2317 request->rq_deadline -
2318 request->rq_arrival_time.tv_sec,
2319 ktime_get_real_seconds() - request->rq_deadline);
2322 work_end = ktime_get_real();
2323 timediff_usecs = ktime_us_delta(work_end, work_start);
2324 arrived_usecs = ktime_us_delta(work_end, arrived);
2326 "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",
2327 request, current->comm,
2328 (request->rq_export ?
2329 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2330 (request->rq_export ?
2331 refcount_read(&request->rq_export->exp_handle.h_ref) : -99),
2332 lustre_msg_get_status(request->rq_reqmsg),
2334 libcfs_id2str(request->rq_peer),
2335 lustre_msg_get_opc(request->rq_reqmsg),
2336 lustre_msg_get_jobid(request->rq_reqmsg) ?: "",
2339 (request->rq_repmsg ?
2340 lustre_msg_get_transno(request->rq_repmsg) :
2341 request->rq_transno),
2343 (request->rq_repmsg ?
2344 lustre_msg_get_status(request->rq_repmsg) : -999));
2345 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
2346 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
2347 int opc = opcode_offset(op);
2349 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
2350 LASSERT(opc < LUSTRE_MAX_OPCODES);
2351 lprocfs_counter_add(svc->srv_stats,
2352 opc + EXTRA_MAX_OPCODES,
2356 if (unlikely(request->rq_early_count)) {
2357 DEBUG_REQ(D_ADAPTTO, request,
2358 "sent %d early replies before finishing in %llds",
2359 request->rq_early_count,
2360 div_u64(arrived_usecs, USEC_PER_SEC));
2363 ptlrpc_server_finish_active_request(svcpt, request);
2369 * An internal function to process a single reply state object.
2371 static int ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
2373 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
2374 struct ptlrpc_service *svc = svcpt->scp_service;
2375 struct obd_export *exp;
2381 exp = rs->rs_export;
2383 LASSERT(rs->rs_difficult);
2384 LASSERT(rs->rs_scheduled);
2385 LASSERT(list_empty(&rs->rs_list));
2388 * The disk commit callback holds exp_uncommitted_replies_lock while it
2389 * iterates over newly committed replies, removing them from
2390 * exp_uncommitted_replies. It then drops this lock and schedules the
2391 * replies it found for handling here.
2393 * We can avoid contention for exp_uncommitted_replies_lock between the
2394 * HRT threads and further commit callbacks by checking rs_committed
2395 * which is set in the commit callback while it holds both
2396 * rs_lock and exp_uncommitted_reples.
2398 * If we see rs_committed clear, the commit callback _may_ not have
2399 * handled this reply yet and we race with it to grab
2400 * exp_uncommitted_replies_lock before removing the reply from
2401 * exp_uncommitted_replies. Note that if we lose the race and the
2402 * reply has already been removed, list_del_init() is a noop.
2404 * If we see rs_committed set, we know the commit callback is handling,
2405 * or has handled this reply since store reordering might allow us to
2406 * see rs_committed set out of sequence. But since this is done
2407 * holding rs_lock, we can be sure it has all completed once we hold
2408 * rs_lock, which we do right next.
2410 if (!rs->rs_committed) {
2412 * if rs was commited, no need to convert locks, don't check
2413 * rs_committed here because rs may never be added into
2414 * exp_uncommitted_replies and this flag never be set, see
2415 * target_send_reply()
2417 if (rs->rs_convert_lock &&
2418 rs->rs_transno > exp->exp_last_committed) {
2419 struct ldlm_lock *lock;
2420 struct ldlm_lock *ack_locks[RS_MAX_LOCKS] = { NULL };
2422 spin_lock(&rs->rs_lock);
2423 if (rs->rs_convert_lock &&
2424 rs->rs_transno > exp->exp_last_committed) {
2425 nlocks = rs->rs_nlocks;
2426 while (nlocks-- > 0) {
2428 * NB don't assume rs is always handled
2429 * by the same service thread (see
2430 * ptlrpc_hr_select, so REP-ACK hr may
2431 * race with trans commit, while the
2432 * latter will release locks, get locks
2433 * here early to convert to COS mode
2436 lock = ldlm_handle2lock(
2437 &rs->rs_locks[nlocks]);
2439 ack_locks[nlocks] = lock;
2440 rs->rs_modes[nlocks] = LCK_COS;
2442 nlocks = rs->rs_nlocks;
2443 rs->rs_convert_lock = 0;
2445 * clear rs_scheduled so that commit callback
2446 * can schedule again
2448 rs->rs_scheduled = 0;
2449 spin_unlock(&rs->rs_lock);
2451 while (nlocks-- > 0) {
2452 lock = ack_locks[nlocks];
2453 ldlm_lock_mode_downgrade(lock, LCK_COS);
2454 LDLM_LOCK_PUT(lock);
2458 spin_unlock(&rs->rs_lock);
2461 spin_lock(&exp->exp_uncommitted_replies_lock);
2462 list_del_init(&rs->rs_obd_list);
2463 spin_unlock(&exp->exp_uncommitted_replies_lock);
2466 spin_lock(&exp->exp_lock);
2467 /* Noop if removed already */
2468 list_del_init(&rs->rs_exp_list);
2469 spin_unlock(&exp->exp_lock);
2471 spin_lock(&rs->rs_lock);
2473 been_handled = rs->rs_handled;
2476 nlocks = rs->rs_nlocks; /* atomic "steal", but */
2477 rs->rs_nlocks = 0; /* locks still on rs_locks! */
2479 if (nlocks == 0 && !been_handled) {
2481 * If we see this, we should already have seen the warning
2482 * in mds_steal_ack_locks()
2485 "All locks stolen from rs %p x%lld.t%lld o%d NID %s\n",
2486 rs, rs->rs_xid, rs->rs_transno, rs->rs_opc,
2487 libcfs_nid2str(exp->exp_connection->c_peer.nid));
2490 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
2491 spin_unlock(&rs->rs_lock);
2493 if (!been_handled && rs->rs_on_net) {
2494 LNetMDUnlink(rs->rs_md_h);
2495 /* Ignore return code; we're racing with completion */
2498 while (nlocks-- > 0)
2499 ldlm_lock_decref(&rs->rs_locks[nlocks],
2500 rs->rs_modes[nlocks]);
2502 spin_lock(&rs->rs_lock);
2505 rs->rs_scheduled = 0;
2506 rs->rs_convert_lock = 0;
2508 if (!rs->rs_on_net) {
2510 spin_unlock(&rs->rs_lock);
2512 class_export_put(exp);
2513 rs->rs_export = NULL;
2514 ptlrpc_rs_decref(rs);
2515 if (atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
2516 svc->srv_is_stopping)
2517 wake_up_all(&svcpt->scp_waitq);
2521 /* still on the net; callback will schedule */
2522 spin_unlock(&rs->rs_lock);
2527 static void ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
2529 int avail = svcpt->scp_nrqbds_posted;
2530 int low_water = test_req_buffer_pressure ? 0 :
2531 svcpt->scp_service->srv_nbuf_per_group / 2;
2533 /* NB I'm not locking; just looking. */
2536 * CAVEAT EMPTOR: We might be allocating buffers here because we've
2537 * allowed the request history to grow out of control. We could put a
2538 * sanity check on that here and cull some history if we need the
2542 if (avail <= low_water)
2543 ptlrpc_grow_req_bufs(svcpt, 1);
2545 if (svcpt->scp_service->srv_stats) {
2546 lprocfs_counter_add(svcpt->scp_service->srv_stats,
2547 PTLRPC_REQBUF_AVAIL_CNTR, avail);
2551 static inline int ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
2553 return svcpt->scp_nreqs_active <
2554 svcpt->scp_nthrs_running - 1 -
2555 (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL);
2559 * allowed to create more threads
2560 * user can call it w/o any lock but need to hold
2561 * ptlrpc_service_part::scp_lock to get reliable result
2563 static inline int ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
2565 return svcpt->scp_nthrs_running +
2566 svcpt->scp_nthrs_starting <
2567 svcpt->scp_service->srv_nthrs_cpt_limit;
2571 * too many requests and allowed to create more threads
2573 static inline int ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
2575 return !ptlrpc_threads_enough(svcpt) &&
2576 ptlrpc_threads_increasable(svcpt);
2579 static inline int ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2581 return thread_is_stopping(thread) ||
2582 thread->t_svcpt->scp_service->srv_is_stopping;
2585 /* stop the highest numbered thread if there are too many threads running */
2586 static inline bool ptlrpc_thread_should_stop(struct ptlrpc_thread *thread)
2588 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2590 return thread->t_id >= svcpt->scp_service->srv_nthrs_cpt_limit &&
2591 thread->t_id == svcpt->scp_thr_nextid - 1;
2594 static void ptlrpc_stop_thread(struct ptlrpc_thread *thread)
2596 CDEBUG(D_INFO, "Stopping thread %s #%u\n",
2597 thread->t_svcpt->scp_service->srv_thread_name, thread->t_id);
2598 thread_add_flags(thread, SVC_STOPPING);
2601 static inline void ptlrpc_thread_stop(struct ptlrpc_thread *thread)
2603 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2605 spin_lock(&svcpt->scp_lock);
2606 if (ptlrpc_thread_should_stop(thread)) {
2607 ptlrpc_stop_thread(thread);
2608 svcpt->scp_thr_nextid--;
2610 spin_unlock(&svcpt->scp_lock);
2613 static inline int ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
2615 return !list_empty(&svcpt->scp_rqbd_idle) &&
2616 svcpt->scp_rqbd_timeout == 0;
2620 ptlrpc_at_check(struct ptlrpc_service_part *svcpt)
2622 return svcpt->scp_at_check;
2626 * If a thread runs too long or spends to much time on a single request,
2627 * we want to know about it, so we set up a delayed work item as a watchdog.
2628 * If it fires, we display a stack trace of the delayed thread,
2629 * providing we aren't rate-limited
2631 * Watchdog stack traces are limited to 3 per 'libcfs_watchdog_ratelimit'
2634 static struct ratelimit_state watchdog_limit;
2636 static void ptlrpc_watchdog_fire(struct work_struct *w)
2638 struct ptlrpc_thread *thread = container_of(w, struct ptlrpc_thread,
2640 u64 ms_lapse = ktime_ms_delta(ktime_get(), thread->t_touched);
2641 u32 ms_frac = do_div(ms_lapse, MSEC_PER_SEC);
2643 /* ___ratelimit() returns true if the action is NOT ratelimited */
2644 if (__ratelimit(&watchdog_limit)) {
2645 /* below message is checked in sanity-quota.sh test_6,18 */
2646 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",
2647 thread->t_task->comm, thread->t_task->pid,
2650 libcfs_debug_dumpstack(thread->t_task);
2652 /* below message is checked in sanity-quota.sh test_6,18 */
2653 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",
2654 thread->t_task->comm, thread->t_task->pid,
2655 ms_lapse, ms_frac, libcfs_watchdog_ratelimit);
2659 void ptlrpc_watchdog_init(struct delayed_work *work, timeout_t timeout)
2661 INIT_DELAYED_WORK(work, ptlrpc_watchdog_fire);
2662 schedule_delayed_work(work, cfs_time_seconds(timeout));
2665 void ptlrpc_watchdog_disable(struct delayed_work *work)
2667 cancel_delayed_work_sync(work);
2670 void ptlrpc_watchdog_touch(struct delayed_work *work, timeout_t timeout)
2672 struct ptlrpc_thread *thread = container_of(&work->work,
2673 struct ptlrpc_thread,
2675 thread->t_touched = ktime_get();
2676 mod_delayed_work(system_wq, work, cfs_time_seconds(timeout));
2680 * requests wait on preprocessing
2681 * user can call it w/o any lock but need to hold
2682 * ptlrpc_service_part::scp_lock to get reliable result
2685 ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
2687 return !list_empty(&svcpt->scp_req_incoming);
2690 static __attribute__((__noinline__)) int
2691 ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
2692 struct ptlrpc_thread *thread)
2694 ptlrpc_watchdog_disable(&thread->t_watchdog);
2698 if (svcpt->scp_rqbd_timeout == 0)
2699 /* Don't exit while there are replies to be handled */
2700 wait_event_idle_exclusive_lifo(
2702 ptlrpc_thread_stopping(thread) ||
2703 ptlrpc_server_request_incoming(svcpt) ||
2704 ptlrpc_server_request_pending(svcpt, false) ||
2705 ptlrpc_rqbd_pending(svcpt) ||
2706 ptlrpc_at_check(svcpt));
2707 else if (wait_event_idle_exclusive_lifo_timeout(
2709 ptlrpc_thread_stopping(thread) ||
2710 ptlrpc_server_request_incoming(svcpt) ||
2711 ptlrpc_server_request_pending(svcpt, false) ||
2712 ptlrpc_rqbd_pending(svcpt) ||
2713 ptlrpc_at_check(svcpt),
2714 svcpt->scp_rqbd_timeout) == 0)
2715 svcpt->scp_rqbd_timeout = 0;
2717 if (ptlrpc_thread_stopping(thread))
2720 ptlrpc_watchdog_touch(&thread->t_watchdog,
2721 ptlrpc_server_get_timeout(svcpt));
2726 * Main thread body for service threads.
2727 * Waits in a loop waiting for new requests to process to appear.
2728 * Every time an incoming requests is added to its queue, a waitq
2729 * is woken up and one of the threads will handle it.
2731 static int ptlrpc_main(void *arg)
2733 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
2734 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2735 struct ptlrpc_service *svc = svcpt->scp_service;
2736 struct ptlrpc_reply_state *rs;
2737 struct group_info *ginfo = NULL;
2739 int counter = 0, rc = 0;
2743 thread->t_task = current;
2744 thread->t_pid = current->pid;
2746 if (svc->srv_cpt_bind) {
2747 rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
2749 CWARN("%s: failed to bind %s on CPT %d\n",
2750 svc->srv_name, thread->t_name, svcpt->scp_cpt);
2754 ginfo = groups_alloc(0);
2756 GOTO(out, rc = -ENOMEM);
2758 set_current_groups(ginfo);
2759 put_group_info(ginfo);
2761 if (svc->srv_ops.so_thr_init != NULL) {
2762 rc = svc->srv_ops.so_thr_init(thread);
2769 GOTO(out_srv_fini, rc = -ENOMEM);
2770 rc = lu_env_add(env);
2774 rc = lu_context_init(&env->le_ctx,
2775 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2777 GOTO(out_env_remove, rc);
2779 thread->t_env = env;
2780 env->le_ctx.lc_thread = thread;
2781 env->le_ctx.lc_cookie = 0x6;
2783 while (!list_empty(&svcpt->scp_rqbd_idle)) {
2784 rc = ptlrpc_server_post_idle_rqbds(svcpt);
2788 CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
2789 svc->srv_name, svcpt->scp_cpt, rc);
2790 GOTO(out_ctx_fini, rc);
2793 /* Alloc reply state structure for this one */
2794 OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2796 GOTO(out_ctx_fini, rc = -ENOMEM);
2798 spin_lock(&svcpt->scp_lock);
2800 LASSERT(thread_is_starting(thread));
2801 thread_clear_flags(thread, SVC_STARTING);
2803 LASSERT(svcpt->scp_nthrs_starting == 1);
2804 svcpt->scp_nthrs_starting--;
2807 * SVC_STOPPING may already be set here if someone else is trying
2808 * to stop the service while this new thread has been dynamically
2809 * forked. We still set SVC_RUNNING to let our creator know that
2810 * we are now running, however we will exit as soon as possible
2812 thread_add_flags(thread, SVC_RUNNING);
2813 svcpt->scp_nthrs_running++;
2814 spin_unlock(&svcpt->scp_lock);
2816 /* wake up our creator in case he's still waiting. */
2817 wake_up(&thread->t_ctl_waitq);
2819 thread->t_touched = ktime_get();
2820 ptlrpc_watchdog_init(&thread->t_watchdog,
2821 ptlrpc_server_get_timeout(svcpt));
2823 spin_lock(&svcpt->scp_rep_lock);
2824 list_add(&rs->rs_list, &svcpt->scp_rep_idle);
2825 wake_up(&svcpt->scp_rep_waitq);
2826 spin_unlock(&svcpt->scp_rep_lock);
2828 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2829 svcpt->scp_nthrs_running);
2831 /* XXX maintain a list of all managed devices: insert here */
2832 while (!ptlrpc_thread_stopping(thread)) {
2833 if (ptlrpc_wait_event(svcpt, thread))
2836 ptlrpc_check_rqbd_pool(svcpt);
2838 if (ptlrpc_threads_need_create(svcpt)) {
2839 /* Ignore return code - we tried... */
2840 ptlrpc_start_thread(svcpt, 0);
2843 /* reset le_ses to initial state */
2845 /* Refill the context before execution to make sure
2846 * all thread keys are allocated */
2848 /* Process all incoming reqs before handling any */
2849 if (ptlrpc_server_request_incoming(svcpt)) {
2850 lu_context_enter(&env->le_ctx);
2851 ptlrpc_server_handle_req_in(svcpt, thread);
2852 lu_context_exit(&env->le_ctx);
2854 /* but limit ourselves in case of flood */
2855 if (counter++ < 100)
2860 if (ptlrpc_at_check(svcpt))
2861 ptlrpc_at_check_timed(svcpt);
2863 if (ptlrpc_server_request_pending(svcpt, false)) {
2864 lu_context_enter(&env->le_ctx);
2865 ptlrpc_server_handle_request(svcpt, thread);
2866 lu_context_exit(&env->le_ctx);
2869 if (ptlrpc_rqbd_pending(svcpt) &&
2870 ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
2872 * I just failed to repost request buffers.
2873 * Wait for a timeout (unless something else
2874 * happens) before I try again
2876 svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
2877 CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
2878 svcpt->scp_nrqbds_posted);
2881 * If the number of threads has been tuned downward and this
2882 * thread should be stopped, then stop in reverse order so the
2883 * the threads always have contiguous thread index values.
2885 if (unlikely(ptlrpc_thread_should_stop(thread)))
2886 ptlrpc_thread_stop(thread);
2889 ptlrpc_watchdog_disable(&thread->t_watchdog);
2892 lu_context_fini(&env->le_ctx);
2898 /* deconstruct service thread state created by ptlrpc_start_thread() */
2899 if (svc->srv_ops.so_thr_done != NULL)
2900 svc->srv_ops.so_thr_done(thread);
2902 CDEBUG(D_RPCTRACE, "%s: service thread [%p:%u] %d exiting: rc = %d\n",
2903 thread->t_name, thread, thread->t_pid, thread->t_id, rc);
2904 spin_lock(&svcpt->scp_lock);
2905 if (thread_test_and_clear_flags(thread, SVC_STARTING))
2906 svcpt->scp_nthrs_starting--;
2908 if (thread_test_and_clear_flags(thread, SVC_RUNNING)) {
2909 /* must know immediately */
2910 svcpt->scp_nthrs_running--;
2914 thread_add_flags(thread, SVC_STOPPED);
2916 wake_up(&thread->t_ctl_waitq);
2917 spin_unlock(&svcpt->scp_lock);
2922 static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
2923 struct list_head *replies)
2927 spin_lock(&hrt->hrt_lock);
2929 list_splice_init(&hrt->hrt_queue, replies);
2930 result = ptlrpc_hr.hr_stopping || !list_empty(replies);
2932 spin_unlock(&hrt->hrt_lock);
2937 * Main body of "handle reply" function.
2938 * It processes acked reply states
2940 static int ptlrpc_hr_main(void *arg)
2942 struct ptlrpc_hr_thread *hrt = (struct ptlrpc_hr_thread *)arg;
2943 struct ptlrpc_hr_partition *hrp = hrt->hrt_partition;
2952 rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
2954 char threadname[20];
2956 snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
2957 hrp->hrp_cpt, hrt->hrt_id);
2958 CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
2959 threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
2962 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD |
2963 LCT_REMEMBER | LCT_NOREF);
2967 rc = lu_env_add(env);
2969 GOTO(out_ctx_fini, rc);
2971 atomic_inc(&hrp->hrp_nstarted);
2972 wake_up(&ptlrpc_hr.hr_waitq);
2974 while (!ptlrpc_hr.hr_stopping) {
2975 wait_event_idle(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
2977 while (!list_empty(&replies)) {
2978 struct ptlrpc_reply_state *rs;
2980 rs = list_entry(replies.prev,
2981 struct ptlrpc_reply_state,
2983 list_del_init(&rs->rs_list);
2984 /* refill keys if needed */
2986 lu_context_enter(&env->le_ctx);
2987 ptlrpc_handle_rs(rs);
2988 lu_context_exit(&env->le_ctx);
2992 atomic_inc(&hrp->hrp_nstopped);
2993 wake_up(&ptlrpc_hr.hr_waitq);
2997 lu_context_fini(&env->le_ctx);
3003 static void ptlrpc_stop_hr_threads(void)
3005 struct ptlrpc_hr_partition *hrp;
3009 ptlrpc_hr.hr_stopping = 1;
3011 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
3012 if (hrp->hrp_thrs == NULL)
3013 continue; /* uninitialized */
3014 for (j = 0; j < hrp->hrp_nthrs; j++)
3015 wake_up(&hrp->hrp_thrs[j].hrt_waitq);
3018 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
3019 if (hrp->hrp_thrs == NULL)
3020 continue; /* uninitialized */
3021 wait_event(ptlrpc_hr.hr_waitq,
3022 atomic_read(&hrp->hrp_nstopped) ==
3023 atomic_read(&hrp->hrp_nstarted));
3027 static int ptlrpc_start_hr_threads(void)
3029 struct ptlrpc_hr_partition *hrp;
3035 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
3038 for (j = 0; j < hrp->hrp_nthrs; j++) {
3039 struct ptlrpc_hr_thread *hrt = &hrp->hrp_thrs[j];
3040 struct task_struct *task;
3042 task = kthread_run(ptlrpc_hr_main,
3044 "ptlrpc_hr%02d_%03d",
3053 wait_event(ptlrpc_hr.hr_waitq,
3054 atomic_read(&hrp->hrp_nstarted) == j);
3057 CERROR("cannot start reply handler thread %d:%d: rc = %d\n",
3059 ptlrpc_stop_hr_threads();
3067 static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
3069 struct ptlrpc_thread *thread;
3074 CDEBUG(D_INFO, "Stopping threads for service %s\n",
3075 svcpt->scp_service->srv_name);
3077 spin_lock(&svcpt->scp_lock);
3078 /* let the thread know that we would like it to stop asap */
3079 list_for_each_entry(thread, &svcpt->scp_threads, t_link)
3080 ptlrpc_stop_thread(thread);
3082 wake_up_all(&svcpt->scp_waitq);
3084 while (!list_empty(&svcpt->scp_threads)) {
3085 thread = list_entry(svcpt->scp_threads.next,
3086 struct ptlrpc_thread, t_link);
3087 if (thread_is_stopped(thread)) {
3088 list_move(&thread->t_link, &zombie);
3091 spin_unlock(&svcpt->scp_lock);
3093 CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
3094 svcpt->scp_service->srv_thread_name, thread->t_id);
3095 wait_event_idle(thread->t_ctl_waitq,
3096 thread_is_stopped(thread));
3098 spin_lock(&svcpt->scp_lock);
3101 spin_unlock(&svcpt->scp_lock);
3103 while (!list_empty(&zombie)) {
3104 thread = list_entry(zombie.next,
3105 struct ptlrpc_thread, t_link);
3106 list_del(&thread->t_link);
3107 OBD_FREE_PTR(thread);
3113 * Stops all threads of a particular service \a svc
3115 static void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
3117 struct ptlrpc_service_part *svcpt;
3122 ptlrpc_service_for_each_part(svcpt, i, svc) {
3123 if (svcpt->scp_service != NULL)
3124 ptlrpc_svcpt_stop_threads(svcpt);
3130 static int ptlrpc_start_threads(struct ptlrpc_service *svc)
3138 /* We require 2 threads min, see note in ptlrpc_server_handle_request */
3139 LASSERT(svc->srv_nthrs_cpt_init >= PTLRPC_NTHRS_INIT);
3141 for (i = 0; i < svc->srv_ncpts; i++) {
3142 for (j = 0; j < svc->srv_nthrs_cpt_init; j++) {
3143 rc = ptlrpc_start_thread(svc->srv_parts[i], 1);
3149 /* We have enough threads, don't start more. b=15759 */
3156 CERROR("cannot start %s thread #%d_%d: rc %d\n",
3157 svc->srv_thread_name, i, j, rc);
3158 ptlrpc_stop_all_threads(svc);
3162 static int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
3164 struct ptlrpc_thread *thread;
3165 struct ptlrpc_service *svc;
3166 struct task_struct *task;
3171 LASSERT(svcpt != NULL);
3173 svc = svcpt->scp_service;
3175 CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n",
3176 svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running,
3177 svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit);
3180 if (unlikely(svc->srv_is_stopping))
3183 if (!ptlrpc_threads_increasable(svcpt) ||
3184 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
3185 svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1))
3188 OBD_CPT_ALLOC_PTR(thread, svc->srv_cptable, svcpt->scp_cpt);
3191 init_waitqueue_head(&thread->t_ctl_waitq);
3193 spin_lock(&svcpt->scp_lock);
3194 if (!ptlrpc_threads_increasable(svcpt)) {
3195 spin_unlock(&svcpt->scp_lock);
3196 OBD_FREE_PTR(thread);
3200 if (svcpt->scp_nthrs_starting != 0) {
3202 * serialize starting because some modules (obdfilter)
3203 * might require unique and contiguous t_id
3205 LASSERT(svcpt->scp_nthrs_starting == 1);
3206 spin_unlock(&svcpt->scp_lock);
3207 OBD_FREE_PTR(thread);
3209 CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n",
3210 svc->srv_thread_name, svcpt->scp_thr_nextid);
3215 CDEBUG(D_INFO, "Creating thread %s #%d race, retry later\n",
3216 svc->srv_thread_name, svcpt->scp_thr_nextid);
3220 svcpt->scp_nthrs_starting++;
3221 thread->t_id = svcpt->scp_thr_nextid++;
3222 thread_add_flags(thread, SVC_STARTING);
3223 thread->t_svcpt = svcpt;
3225 list_add(&thread->t_link, &svcpt->scp_threads);
3226 spin_unlock(&svcpt->scp_lock);
3228 if (svcpt->scp_cpt >= 0) {
3229 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s%02d_%03d",
3230 svc->srv_thread_name, svcpt->scp_cpt, thread->t_id);
3232 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s_%04d",
3233 svc->srv_thread_name, thread->t_id);
3236 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name);
3237 task = kthread_run(ptlrpc_main, thread, "%s", thread->t_name);
3240 CERROR("cannot start thread '%s': rc = %d\n",
3241 thread->t_name, rc);
3242 spin_lock(&svcpt->scp_lock);
3243 --svcpt->scp_nthrs_starting;
3244 if (thread_is_stopping(thread)) {
3246 * this ptlrpc_thread is being hanled
3247 * by ptlrpc_svcpt_stop_threads now
3249 thread_add_flags(thread, SVC_STOPPED);
3250 wake_up(&thread->t_ctl_waitq);
3251 spin_unlock(&svcpt->scp_lock);
3253 list_del(&thread->t_link);
3254 spin_unlock(&svcpt->scp_lock);
3255 OBD_FREE_PTR(thread);
3263 wait_event_idle(thread->t_ctl_waitq,
3264 thread_is_running(thread) || thread_is_stopped(thread));
3266 rc = thread_is_stopped(thread) ? thread->t_id : 0;
3270 int ptlrpc_hr_init(void)
3272 struct ptlrpc_hr_partition *hrp;
3273 struct ptlrpc_hr_thread *hrt;
3281 memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
3282 ptlrpc_hr.hr_cpt_table = cfs_cpt_tab;
3284 ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
3286 if (ptlrpc_hr.hr_partitions == NULL)
3289 ratelimit_state_init(&watchdog_limit,
3290 cfs_time_seconds(libcfs_watchdog_ratelimit), 3);
3292 init_waitqueue_head(&ptlrpc_hr.hr_waitq);
3295 weight = cpumask_weight(topology_sibling_cpumask(smp_processor_id()));
3298 cfs_percpt_for_each(hrp, cpt, ptlrpc_hr.hr_partitions) {
3301 atomic_set(&hrp->hrp_nstarted, 0);
3302 atomic_set(&hrp->hrp_nstopped, 0);
3304 hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, cpt);
3305 hrp->hrp_nthrs /= weight;
3306 if (hrp->hrp_nthrs == 0)
3309 OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, cpt,
3310 hrp->hrp_nthrs * sizeof(*hrt));
3311 if (hrp->hrp_thrs == NULL)
3312 GOTO(out, rc = -ENOMEM);
3314 for (i = 0; i < hrp->hrp_nthrs; i++) {
3315 hrt = &hrp->hrp_thrs[i];
3318 hrt->hrt_partition = hrp;
3319 init_waitqueue_head(&hrt->hrt_waitq);
3320 spin_lock_init(&hrt->hrt_lock);
3321 INIT_LIST_HEAD(&hrt->hrt_queue);
3325 rc = ptlrpc_start_hr_threads();
3332 void ptlrpc_hr_fini(void)
3334 struct ptlrpc_hr_partition *hrp;
3337 if (ptlrpc_hr.hr_partitions == NULL)
3340 ptlrpc_stop_hr_threads();
3342 cfs_percpt_for_each(hrp, cpt, ptlrpc_hr.hr_partitions) {
3344 OBD_FREE_PTR_ARRAY(hrp->hrp_thrs, hrp->hrp_nthrs);
3347 cfs_percpt_free(ptlrpc_hr.hr_partitions);
3348 ptlrpc_hr.hr_partitions = NULL;
3353 * Wait until all already scheduled replies are processed.
3355 static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
3358 if (wait_event_idle_timeout(
3360 atomic_read(&svcpt->scp_nreps_difficult) == 0,
3361 cfs_time_seconds(10)) > 0)
3363 CWARN("Unexpectedly long timeout %s %p\n",
3364 svcpt->scp_service->srv_name, svcpt->scp_service);
3369 ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
3371 struct ptlrpc_service_part *svcpt;
3374 /* early disarm AT timer... */
3375 ptlrpc_service_for_each_part(svcpt, i, svc) {
3376 if (svcpt->scp_service != NULL)
3377 del_timer(&svcpt->scp_at_timer);
3382 ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
3384 struct ptlrpc_service_part *svcpt;
3385 struct ptlrpc_request_buffer_desc *rqbd;
3390 * All history will be culled when the next request buffer is
3391 * freed in ptlrpc_service_purge_all()
3393 svc->srv_hist_nrqbds_cpt_max = 0;
3395 rc = LNetClearLazyPortal(svc->srv_req_portal);
3398 ptlrpc_service_for_each_part(svcpt, i, svc) {
3399 if (svcpt->scp_service == NULL)
3403 * Unlink all the request buffers. This forces a 'final'
3404 * event with its 'unlink' flag set for each posted rqbd
3406 list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
3408 rc = LNetMDUnlink(rqbd->rqbd_md_h);
3409 LASSERT(rc == 0 || rc == -ENOENT);
3413 ptlrpc_service_for_each_part(svcpt, i, svc) {
3414 if (svcpt->scp_service == NULL)
3418 * Wait for the network to release any buffers
3419 * it's currently filling
3421 spin_lock(&svcpt->scp_lock);
3422 while (svcpt->scp_nrqbds_posted != 0) {
3423 int seconds = PTLRPC_REQ_LONG_UNLINK;
3425 spin_unlock(&svcpt->scp_lock);
3427 * Network access will complete in finite time but
3428 * the HUGE timeout lets us CWARN for visibility
3431 while (seconds > 0 &&
3432 wait_event_idle_timeout(
3434 svcpt->scp_nrqbds_posted == 0,
3435 cfs_time_seconds(1)) == 0)
3438 CWARN("Service %s waiting for request buffers\n",
3439 svcpt->scp_service->srv_name);
3441 spin_lock(&svcpt->scp_lock);
3443 spin_unlock(&svcpt->scp_lock);
3448 ptlrpc_service_purge_all(struct ptlrpc_service *svc)
3450 struct ptlrpc_service_part *svcpt;
3451 struct ptlrpc_request_buffer_desc *rqbd;
3452 struct ptlrpc_request *req;
3453 struct ptlrpc_reply_state *rs;
3456 ptlrpc_service_for_each_part(svcpt, i, svc) {
3457 if (svcpt->scp_service == NULL)
3460 spin_lock(&svcpt->scp_rep_lock);
3461 while (!list_empty(&svcpt->scp_rep_active)) {
3462 rs = list_entry(svcpt->scp_rep_active.next,
3463 struct ptlrpc_reply_state, rs_list);
3464 spin_lock(&rs->rs_lock);
3465 ptlrpc_schedule_difficult_reply(rs);
3466 spin_unlock(&rs->rs_lock);
3468 spin_unlock(&svcpt->scp_rep_lock);
3471 * purge the request queue. NB No new replies (rqbds
3472 * all unlinked) and no service threads, so I'm the only
3473 * thread noodling the request queue now
3475 while (!list_empty(&svcpt->scp_req_incoming)) {
3476 req = list_entry(svcpt->scp_req_incoming.next,
3477 struct ptlrpc_request, rq_list);
3479 list_del(&req->rq_list);
3480 svcpt->scp_nreqs_incoming--;
3481 ptlrpc_server_finish_request(svcpt, req);
3484 while (ptlrpc_server_request_pending(svcpt, true)) {
3485 req = ptlrpc_server_request_get(svcpt, true);
3486 ptlrpc_server_finish_active_request(svcpt, req);
3490 * The portal may be shared by several services (eg:OUT_PORTAL).
3491 * So the request could be referenced by other target. So we
3492 * have to wait the ptlrpc_server_drop_request invoked.
3494 * TODO: move the req_buffer as global rather than per service.
3496 spin_lock(&svcpt->scp_lock);
3497 while (!list_empty(&svcpt->scp_rqbd_posted)) {
3498 spin_unlock(&svcpt->scp_lock);
3499 wait_event_idle_timeout(svcpt->scp_waitq,
3500 list_empty(&svcpt->scp_rqbd_posted),
3501 cfs_time_seconds(1));
3502 spin_lock(&svcpt->scp_lock);
3504 spin_unlock(&svcpt->scp_lock);
3506 LASSERT(svcpt->scp_nreqs_incoming == 0);
3507 LASSERT(svcpt->scp_nreqs_active == 0);
3509 * history should have been culled by
3510 * ptlrpc_server_finish_request
3512 LASSERT(svcpt->scp_hist_nrqbds == 0);
3515 * Now free all the request buffers since nothing
3516 * references them any more...
3519 while (!list_empty(&svcpt->scp_rqbd_idle)) {
3520 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
3521 struct ptlrpc_request_buffer_desc,
3523 ptlrpc_free_rqbd(rqbd);
3525 ptlrpc_wait_replies(svcpt);
3527 while (!list_empty(&svcpt->scp_rep_idle)) {
3528 rs = list_entry(svcpt->scp_rep_idle.next,
3529 struct ptlrpc_reply_state,
3531 list_del(&rs->rs_list);
3532 OBD_FREE_LARGE(rs, svc->srv_max_reply_size);
3538 ptlrpc_service_free(struct ptlrpc_service *svc)
3540 struct ptlrpc_service_part *svcpt;
3541 struct ptlrpc_at_array *array;
3544 ptlrpc_service_for_each_part(svcpt, i, svc) {
3545 if (svcpt->scp_service == NULL)
3548 /* In case somebody rearmed this in the meantime */
3549 del_timer(&svcpt->scp_at_timer);
3550 array = &svcpt->scp_at_array;
3552 if (array->paa_reqs_array != NULL) {
3553 OBD_FREE_PTR_ARRAY(array->paa_reqs_array,
3555 array->paa_reqs_array = NULL;
3558 if (array->paa_reqs_count != NULL) {
3559 OBD_FREE_PTR_ARRAY(array->paa_reqs_count,
3561 array->paa_reqs_count = NULL;
3565 ptlrpc_service_for_each_part(svcpt, i, svc)
3566 OBD_FREE_PTR(svcpt);
3568 if (svc->srv_cpts != NULL)
3569 cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts);
3571 OBD_FREE(svc, offsetof(struct ptlrpc_service,
3572 srv_parts[svc->srv_ncpts]));
3575 int ptlrpc_unregister_service(struct ptlrpc_service *service)
3579 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
3581 service->srv_is_stopping = 1;
3583 mutex_lock(&ptlrpc_all_services_mutex);
3584 list_del_init(&service->srv_list);
3585 mutex_unlock(&ptlrpc_all_services_mutex);
3587 ptlrpc_service_del_atimer(service);
3588 ptlrpc_stop_all_threads(service);
3590 ptlrpc_service_unlink_rqbd(service);
3591 ptlrpc_service_purge_all(service);
3592 ptlrpc_service_nrs_cleanup(service);
3594 ptlrpc_lprocfs_unregister_service(service);
3595 ptlrpc_sysfs_unregister_service(service);
3597 ptlrpc_service_free(service);
3601 EXPORT_SYMBOL(ptlrpc_unregister_service);
3604 * Returns 0 if the service is healthy.
3606 * Right now, it just checks to make sure that requests aren't languishing
3607 * in the queue. We'll use this health check to govern whether a node needs
3608 * to be shot, so it's intentionally non-aggressive.
3610 static int ptlrpc_svcpt_health_check(struct ptlrpc_service_part *svcpt)
3612 struct ptlrpc_request *request = NULL;
3613 struct timespec64 right_now;
3614 struct timespec64 timediff;
3616 ktime_get_real_ts64(&right_now);
3618 spin_lock(&svcpt->scp_req_lock);
3619 /* How long has the next entry been waiting? */
3620 if (ptlrpc_server_high_pending(svcpt, true))
3621 request = ptlrpc_nrs_req_peek_nolock(svcpt, true);
3622 else if (ptlrpc_server_normal_pending(svcpt, true))
3623 request = ptlrpc_nrs_req_peek_nolock(svcpt, false);
3625 if (request == NULL) {
3626 spin_unlock(&svcpt->scp_req_lock);
3630 timediff = timespec64_sub(right_now, request->rq_arrival_time);
3631 spin_unlock(&svcpt->scp_req_lock);
3633 if ((timediff.tv_sec) >
3634 (AT_OFF ? obd_timeout * 3 / 2 : at_max)) {
3635 CERROR("%s: unhealthy - request has been waiting %llds\n",
3636 svcpt->scp_service->srv_name, (s64)timediff.tv_sec);
3644 ptlrpc_service_health_check(struct ptlrpc_service *svc)
3646 struct ptlrpc_service_part *svcpt;
3652 ptlrpc_service_for_each_part(svcpt, i, svc) {
3653 int rc = ptlrpc_svcpt_health_check(svcpt);
3660 EXPORT_SYMBOL(ptlrpc_service_health_check);