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_first_entry(&svcpt->scp_rqbd_idle,
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_first_entry(&svcpt->scp_hist_rqbds,
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 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_ENQ_RESEND)) {
1357 /* don't send early reply */
1362 * deadline is when the client expects us to reply, margin is the
1363 * difference between clients' and servers' expectations
1365 DEBUG_REQ(D_ADAPTTO, req,
1366 "%ssending early reply (deadline %+ds, margin %+ds) for %d+%d",
1367 AT_OFF ? "AT off - not " : "",
1368 olddl, olddl - at_get(&svcpt->scp_at_estimate),
1369 at_get(&svcpt->scp_at_estimate), at_extra);
1375 /* below message is checked in replay-ost-single.sh test_9 */
1376 DEBUG_REQ(D_WARNING, req,
1377 "Already past deadline (%+ds), not sending early reply. Consider increasing at_early_margin (%d)?",
1378 olddl, at_early_margin);
1380 /* Return an error so we're not re-added to the timed list. */
1384 if ((lustre_msghdr_get_flags(req->rq_reqmsg) &
1385 MSGHDR_AT_SUPPORT) == 0) {
1386 DEBUG_REQ(D_INFO, req,
1387 "Wanted to ask client for more time, but no AT support");
1391 if (req->rq_export &&
1392 lustre_msg_get_flags(req->rq_reqmsg) &
1393 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1394 struct obd_device *obd_exp = req->rq_export->exp_obd;
1397 * During recovery, we don't want to send too many early
1398 * replies, but on the other hand we want to make sure the
1399 * client has enough time to resend if the rpc is lost. So
1400 * during the recovery period send at least 4 early replies,
1401 * spacing them every at_extra if we can. at_estimate should
1402 * always equal this fixed value during recovery.
1406 * Don't account request processing time into AT history
1407 * during recovery, it is not service time we need but
1408 * includes also waiting time for recovering clients
1410 newdl = min_t(time64_t, at_extra,
1411 obd_exp->obd_recovery_timeout / 4) +
1412 ktime_get_real_seconds();
1415 * We want to extend the request deadline by at_extra seconds,
1416 * so we set our service estimate to reflect how much time has
1417 * passed since this request arrived plus an additional
1418 * at_extra seconds. The client will calculate the new deadline
1419 * based on this service estimate (plus some additional time to
1420 * account for network latency). See ptlrpc_at_recv_early_reply
1422 at_measured(&svcpt->scp_at_estimate, at_extra +
1423 ktime_get_real_seconds() -
1424 req->rq_arrival_time.tv_sec);
1425 newdl = req->rq_arrival_time.tv_sec +
1426 at_get(&svcpt->scp_at_estimate);
1430 * Check to see if we've actually increased the deadline -
1431 * we may be past adaptive_max
1433 if (req->rq_deadline >= newdl) {
1434 DEBUG_REQ(D_WARNING, req,
1435 "Could not add any time (%d/%lld), not sending early reply",
1436 olddl, newdl - ktime_get_real_seconds());
1440 reqcopy = ptlrpc_request_cache_alloc(GFP_NOFS);
1441 if (reqcopy == NULL)
1443 OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1445 GOTO(out_free, rc = -ENOMEM);
1448 reqcopy->rq_reply_state = NULL;
1449 reqcopy->rq_rep_swab_mask = 0;
1450 reqcopy->rq_pack_bulk = 0;
1451 reqcopy->rq_pack_udesc = 0;
1452 reqcopy->rq_packed_final = 0;
1453 sptlrpc_svc_ctx_addref(reqcopy);
1454 /* We only need the reqmsg for the magic */
1455 reqcopy->rq_reqmsg = reqmsg;
1456 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1459 * tgt_brw_read() and tgt_brw_write() may have decided not to reply.
1460 * Without this check, we would fail the rq_no_reply assertion in
1461 * ptlrpc_send_reply().
1463 if (reqcopy->rq_no_reply)
1464 GOTO(out, rc = -ETIMEDOUT);
1466 LASSERT(atomic_read(&req->rq_refcount));
1467 /* if it is last refcount then early reply isn't needed */
1468 if (atomic_read(&req->rq_refcount) == 1) {
1469 DEBUG_REQ(D_ADAPTTO, reqcopy,
1470 "Normal reply already sent, abort early reply");
1471 GOTO(out, rc = -EINVAL);
1474 /* Connection ref */
1475 reqcopy->rq_export = class_conn2export(
1476 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1477 if (reqcopy->rq_export == NULL)
1478 GOTO(out, rc = -ENODEV);
1481 class_export_rpc_inc(reqcopy->rq_export);
1482 if (reqcopy->rq_export->exp_obd &&
1483 reqcopy->rq_export->exp_obd->obd_fail)
1484 GOTO(out_put, rc = -ENODEV);
1486 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1490 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1493 /* Adjust our own deadline to what we told the client */
1494 req->rq_deadline = newdl;
1495 req->rq_early_count++; /* number sent, server side */
1497 DEBUG_REQ(D_ERROR, req, "Early reply send failed: rc = %d", rc);
1501 * Free the (early) reply state from lustre_pack_reply.
1502 * (ptlrpc_send_reply takes it's own rs ref, so this is safe here)
1504 ptlrpc_req_drop_rs(reqcopy);
1507 class_export_rpc_dec(reqcopy->rq_export);
1508 class_export_put(reqcopy->rq_export);
1510 sptlrpc_svc_ctx_decref(reqcopy);
1511 OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1513 ptlrpc_request_cache_free(reqcopy);
1518 * Send early replies to everybody expiring within at_early_margin
1519 * asking for at_extra time
1521 static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1523 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1524 struct ptlrpc_request *rq, *n;
1525 LIST_HEAD(work_list);
1528 time64_t now = ktime_get_real_seconds();
1530 int first, counter = 0;
1533 spin_lock(&svcpt->scp_at_lock);
1534 if (svcpt->scp_at_check == 0) {
1535 spin_unlock(&svcpt->scp_at_lock);
1538 delay_ms = ktime_ms_delta(ktime_get(), svcpt->scp_at_checktime);
1539 svcpt->scp_at_check = 0;
1541 if (array->paa_count == 0) {
1542 spin_unlock(&svcpt->scp_at_lock);
1546 /* The timer went off, but maybe the nearest rpc already completed. */
1547 first = array->paa_deadline - now;
1548 if (first > at_early_margin) {
1549 /* We've still got plenty of time. Reset the timer. */
1550 ptlrpc_at_set_timer(svcpt);
1551 spin_unlock(&svcpt->scp_at_lock);
1556 * We're close to a timeout, and we don't know how much longer the
1557 * server will take. Send early replies to everyone expiring soon.
1560 div_u64_rem(array->paa_deadline, array->paa_size, &index);
1561 count = array->paa_count;
1563 count -= array->paa_reqs_count[index];
1564 list_for_each_entry_safe(rq, n,
1565 &array->paa_reqs_array[index],
1567 if (rq->rq_deadline > now + at_early_margin) {
1568 /* update the earliest deadline */
1569 if (deadline == -1 ||
1570 rq->rq_deadline < deadline)
1571 deadline = rq->rq_deadline;
1576 * ptlrpc_server_drop_request() may drop
1577 * refcount to 0 already. Let's check this and
1578 * don't add entry to work_list
1580 if (likely(atomic_inc_not_zero(&rq->rq_refcount))) {
1581 ptlrpc_at_remove_timed(rq);
1582 list_add(&rq->rq_timed_list, &work_list);
1584 ptlrpc_at_remove_timed(rq);
1590 if (++index >= array->paa_size)
1593 array->paa_deadline = deadline;
1594 /* we have a new earliest deadline, restart the timer */
1595 ptlrpc_at_set_timer(svcpt);
1597 spin_unlock(&svcpt->scp_at_lock);
1600 "timeout in %+ds, asking for %d secs on %d early replies\n",
1601 first, at_extra, counter);
1604 * We're already past request deadlines before we even get a
1605 * chance to send early replies
1607 LCONSOLE_WARN("%s: This server is not able to keep up with request traffic (cpu-bound).\n",
1608 svcpt->scp_service->srv_name);
1609 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, delay=%lldms\n",
1610 counter, svcpt->scp_nreqs_incoming,
1611 svcpt->scp_nreqs_active,
1612 at_get(&svcpt->scp_at_estimate), delay_ms);
1616 * we took additional refcount so entries can't be deleted from list, no
1619 while ((rq = list_first_entry_or_null(&work_list,
1620 struct ptlrpc_request,
1621 rq_timed_list)) != NULL) {
1622 list_del_init(&rq->rq_timed_list);
1624 if (ptlrpc_at_send_early_reply(rq) == 0)
1625 ptlrpc_at_add_timed(rq);
1627 ptlrpc_server_drop_request(rq);
1630 RETURN(1); /* return "did_something" for liblustre */
1634 * Check if we are already handling earlier incarnation of this request.
1635 * Called under &req->rq_export->exp_rpc_lock locked
1637 static struct ptlrpc_request*
1638 ptlrpc_server_check_resend_in_progress(struct ptlrpc_request *req)
1640 struct ptlrpc_request *tmp = NULL;
1642 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ||
1643 (atomic_read(&req->rq_export->exp_rpc_count) == 0))
1647 * This list should not be longer than max_requests in
1648 * flights on the client, so it is not all that long.
1649 * Also we only hit this codepath in case of a resent
1650 * request which makes it even more rarely hit
1652 list_for_each_entry(tmp, &req->rq_export->exp_reg_rpcs,
1654 /* Found duplicate one */
1655 if (tmp->rq_xid == req->rq_xid)
1658 list_for_each_entry(tmp, &req->rq_export->exp_hp_rpcs,
1660 /* Found duplicate one */
1661 if (tmp->rq_xid == req->rq_xid)
1667 DEBUG_REQ(D_HA, req, "Found duplicate req in processing");
1668 DEBUG_REQ(D_HA, tmp, "Request being processed");
1672 #ifdef HAVE_SERVER_SUPPORT
1673 static void ptlrpc_server_mark_obsolete(struct ptlrpc_request *req)
1675 req->rq_obsolete = 1;
1679 ptlrpc_server_mark_in_progress_obsolete(struct ptlrpc_request *req)
1681 struct ptlrpc_request *tmp = NULL;
1684 if (!tgt_is_increasing_xid_client(req->rq_export) ||
1685 req->rq_export->exp_used_slots == NULL)
1688 tag = lustre_msg_get_tag(req->rq_reqmsg);
1692 if (!test_bit(tag - 1, req->rq_export->exp_used_slots))
1695 /* This list should not be longer than max_requests in
1696 * flights on the client, so it is not all that long.
1697 * Also we only hit this codepath in case of a resent
1698 * request which makes it even more rarely hit */
1699 list_for_each_entry(tmp, &req->rq_export->exp_reg_rpcs, rq_exp_list) {
1700 if (tag == lustre_msg_get_tag(tmp->rq_reqmsg) &&
1701 req->rq_xid > tmp->rq_xid)
1702 ptlrpc_server_mark_obsolete(tmp);
1705 list_for_each_entry(tmp, &req->rq_export->exp_hp_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);
1714 * Check if a request should be assigned with a high priority.
1716 * \retval < 0: error occurred
1717 * 0: normal RPC request
1718 * +1: high priority request
1720 static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt,
1721 struct ptlrpc_request *req)
1726 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL) {
1727 rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
1734 if (req->rq_export != NULL && req->rq_ops != NULL) {
1736 * Perform request specific check. We should do this
1737 * check before the request is added into exp_hp_rpcs
1738 * list otherwise it may hit swab race at LU-1044.
1740 if (req->rq_ops->hpreq_check != NULL) {
1741 rc = req->rq_ops->hpreq_check(req);
1742 if (rc == -ESTALE) {
1743 req->rq_status = rc;
1747 * can only return error,
1748 * 0 for normal request,
1749 * or 1 for high priority request
1758 /** Remove the request from the export list. */
1759 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req)
1762 if (req->rq_export) {
1764 * refresh lock timeout again so that client has more
1765 * room to send lock cancel RPC.
1767 if (req->rq_ops && req->rq_ops->hpreq_fini)
1768 req->rq_ops->hpreq_fini(req);
1770 ptlrpc_del_exp_list(req);
1775 static int ptlrpc_hpreq_check(struct ptlrpc_request *req)
1780 static struct ptlrpc_hpreq_ops ptlrpc_hpreq_common = {
1781 .hpreq_check = ptlrpc_hpreq_check,
1784 /* Hi-Priority RPC check by RPC operation code. */
1785 int ptlrpc_hpreq_handler(struct ptlrpc_request *req)
1787 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1790 * Check for export to let only reconnects for not yet evicted
1791 * export to become a HP rpc.
1793 if ((req->rq_export != NULL) &&
1794 (opc == OBD_PING || opc == MDS_CONNECT || opc == OST_CONNECT))
1795 req->rq_ops = &ptlrpc_hpreq_common;
1799 EXPORT_SYMBOL(ptlrpc_hpreq_handler);
1801 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1802 struct ptlrpc_request *req)
1806 struct ptlrpc_request *orig;
1810 rc = ptlrpc_server_hpreq_init(svcpt, req);
1815 ptlrpc_nrs_req_initialize(svcpt, req, hp);
1817 while (req->rq_export != NULL) {
1818 struct obd_export *exp = req->rq_export;
1821 * do search for duplicated xid and the adding to the list
1824 spin_lock_bh(&exp->exp_rpc_lock);
1825 #ifdef HAVE_SERVER_SUPPORT
1826 ptlrpc_server_mark_in_progress_obsolete(req);
1828 orig = ptlrpc_server_check_resend_in_progress(req);
1829 if (orig && OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_RESEND_RACE)) {
1830 spin_unlock_bh(&exp->exp_rpc_lock);
1832 OBD_RACE(OBD_FAIL_PTLRPC_RESEND_RACE);
1833 msleep(4 * MSEC_PER_SEC);
1837 if (orig && likely(atomic_inc_not_zero(&orig->rq_refcount))) {
1840 spin_unlock_bh(&exp->exp_rpc_lock);
1843 * When the client resend request and the server has
1844 * the previous copy of it, we need to update deadlines,
1845 * to be sure that the client and the server have equal
1846 * request deadlines.
1849 spin_lock(&orig->rq_rqbd->rqbd_svcpt->scp_at_lock);
1850 linked = orig->rq_at_linked;
1852 ptlrpc_at_remove_timed(orig);
1853 spin_unlock(&orig->rq_rqbd->rqbd_svcpt->scp_at_lock);
1854 orig->rq_deadline = req->rq_deadline;
1855 orig->rq_rep_mbits = req->rq_rep_mbits;
1857 ptlrpc_at_add_timed(orig);
1858 ptlrpc_server_drop_request(orig);
1859 ptlrpc_nrs_req_finalize(req);
1861 /* don't mark slot unused for resend in progress */
1862 req->rq_obsolete = 1;
1867 ptlrpc_add_exp_list_nolock(req, exp, hp || req->rq_ops != NULL);
1869 spin_unlock_bh(&exp->exp_rpc_lock);
1874 * the current thread is not the processing thread for this request
1875 * since that, but request is in exp_hp_list and can be find there.
1876 * Remove all relations between request and old thread.
1878 req->rq_svc_thread->t_env->le_ses = NULL;
1879 req->rq_svc_thread = NULL;
1880 req->rq_session.lc_thread = NULL;
1882 ptlrpc_nrs_req_add(svcpt, req, hp);
1888 * Allow to handle high priority request
1889 * User can call it w/o any lock but need to hold
1890 * ptlrpc_service_part::scp_req_lock to get reliable result
1892 static bool ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1895 int running = svcpt->scp_nthrs_running;
1897 if (!nrs_svcpt_has_hp(svcpt))
1903 if (ptlrpc_nrs_req_throttling_nolock(svcpt, true))
1906 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1907 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1908 /* leave just 1 thread for normal RPCs */
1909 running = PTLRPC_NTHRS_INIT;
1910 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1914 if (svcpt->scp_nreqs_active >= running - 1)
1917 if (svcpt->scp_nhreqs_active == 0)
1920 return !ptlrpc_nrs_req_pending_nolock(svcpt, false) ||
1921 svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1924 static bool ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1927 return ptlrpc_server_allow_high(svcpt, force) &&
1928 ptlrpc_nrs_req_pending_nolock(svcpt, true);
1932 * Only allow normal priority requests on a service that has a high-priority
1933 * queue if forced (i.e. cleanup), if there are other high priority requests
1934 * already being processed (i.e. those threads can service more high-priority
1935 * requests), or if there are enough idle threads that a later thread can do
1936 * a high priority request.
1937 * User can call it w/o any lock but need to hold
1938 * ptlrpc_service_part::scp_req_lock to get reliable result
1940 static bool ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1943 int running = svcpt->scp_nthrs_running;
1945 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1946 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1947 /* leave just 1 thread for normal RPCs */
1948 running = PTLRPC_NTHRS_INIT;
1949 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1956 if (ptlrpc_nrs_req_throttling_nolock(svcpt, false))
1959 if (svcpt->scp_nreqs_active < running - 2)
1962 if (svcpt->scp_nreqs_active >= running - 1)
1965 return svcpt->scp_nhreqs_active > 0 || !nrs_svcpt_has_hp(svcpt);
1968 static bool ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1971 return ptlrpc_server_allow_normal(svcpt, force) &&
1972 ptlrpc_nrs_req_pending_nolock(svcpt, false);
1976 * Returns true if there are requests available in incoming
1977 * request queue for processing and it is allowed to fetch them.
1978 * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1979 * to get reliable result
1980 * \see ptlrpc_server_allow_normal
1981 * \see ptlrpc_server_allow high
1984 bool ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt,
1987 return ptlrpc_server_high_pending(svcpt, force) ||
1988 ptlrpc_server_normal_pending(svcpt, force);
1992 * Fetch a request for processing from queue of unprocessed requests.
1993 * Favors high-priority requests.
1994 * Returns a pointer to fetched request.
1996 static struct ptlrpc_request *
1997 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force)
1999 struct ptlrpc_request *req = NULL;
2003 spin_lock(&svcpt->scp_req_lock);
2005 if (ptlrpc_server_high_pending(svcpt, force)) {
2006 req = ptlrpc_nrs_req_get_nolock(svcpt, true, force);
2008 svcpt->scp_hreq_count++;
2013 if (ptlrpc_server_normal_pending(svcpt, force)) {
2014 req = ptlrpc_nrs_req_get_nolock(svcpt, false, force);
2016 svcpt->scp_hreq_count = 0;
2021 spin_unlock(&svcpt->scp_req_lock);
2025 svcpt->scp_nreqs_active++;
2027 svcpt->scp_nhreqs_active++;
2029 spin_unlock(&svcpt->scp_req_lock);
2031 if (likely(req->rq_export))
2032 class_export_rpc_inc(req->rq_export);
2038 * Handle freshly incoming reqs, add to timed early reply list,
2039 * pass on to regular request queue.
2040 * All incoming requests pass through here before getting into
2041 * ptlrpc_server_handle_req later on.
2043 static int ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt,
2044 struct ptlrpc_thread *thread)
2046 struct ptlrpc_service *svc = svcpt->scp_service;
2047 struct ptlrpc_request *req;
2054 spin_lock(&svcpt->scp_lock);
2055 if (list_empty(&svcpt->scp_req_incoming)) {
2056 spin_unlock(&svcpt->scp_lock);
2060 req = list_first_entry(&svcpt->scp_req_incoming,
2061 struct ptlrpc_request, rq_list);
2062 list_del_init(&req->rq_list);
2063 svcpt->scp_nreqs_incoming--;
2065 * Consider this still a "queued" request as far as stats are
2068 spin_unlock(&svcpt->scp_lock);
2070 /* go through security check/transform */
2071 rc = sptlrpc_svc_unwrap_request(req);
2075 case SECSVC_COMPLETE:
2076 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
2085 * for null-flavored rpc, msg has been unpacked by sptlrpc, although
2086 * redo it wouldn't be harmful.
2088 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
2089 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
2091 CERROR("error unpacking request: ptl %d from %s x%llu\n",
2092 svc->srv_req_portal, libcfs_id2str(req->rq_peer),
2098 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
2100 CERROR("error unpacking ptlrpc body: ptl %d from %s x %llu\n",
2101 svc->srv_req_portal, libcfs_id2str(req->rq_peer),
2106 opc = lustre_msg_get_opc(req->rq_reqmsg);
2107 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
2108 opc == cfs_fail_val) {
2109 CERROR("drop incoming rpc opc %u, x%llu\n",
2110 cfs_fail_val, req->rq_xid);
2115 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
2116 CERROR("wrong packet type received (type=%u) from %s\n",
2117 lustre_msg_get_type(req->rq_reqmsg),
2118 libcfs_id2str(req->rq_peer));
2126 req->rq_bulk_write = 1;
2130 case MGS_CONFIG_READ:
2131 req->rq_bulk_read = 1;
2135 CDEBUG(D_RPCTRACE, "got req x%llu\n", req->rq_xid);
2137 req->rq_export = class_conn2export(
2138 lustre_msg_get_handle(req->rq_reqmsg));
2139 if (req->rq_export) {
2140 rc = ptlrpc_check_req(req);
2142 rc = sptlrpc_target_export_check(req->rq_export, req);
2144 DEBUG_REQ(D_ERROR, req,
2145 "DROPPING req with illegal security flavor");
2150 ptlrpc_update_export_timer(req->rq_export, 0);
2153 /* req_in handling should/must be fast */
2154 if (ktime_get_real_seconds() - req->rq_arrival_time.tv_sec > 5)
2155 DEBUG_REQ(D_WARNING, req, "Slow req_in handling %llds",
2156 ktime_get_real_seconds() -
2157 req->rq_arrival_time.tv_sec);
2159 /* Set rpc server deadline and add it to the timed list */
2160 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
2161 MSGHDR_AT_SUPPORT) ?
2162 /* The max time the client expects us to take */
2163 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
2165 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
2166 if (unlikely(deadline == 0)) {
2167 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
2171 /* Skip early reply */
2172 if (OBD_FAIL_PRECHECK(OBD_FAIL_MDS_RESEND))
2173 req->rq_deadline += obd_timeout;
2175 req->rq_svc_thread = thread;
2176 if (thread != NULL) {
2178 * initialize request session, it is needed for request
2179 * processing by target
2181 rc = lu_context_init(&req->rq_session, LCT_SERVER_SESSION |
2184 CERROR("%s: failure to initialize session: rc = %d\n",
2185 thread->t_name, rc);
2188 req->rq_session.lc_thread = thread;
2189 lu_context_enter(&req->rq_session);
2190 thread->t_env->le_ses = &req->rq_session;
2194 if (unlikely(OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_ENQ_RESEND) &&
2195 (opc == LDLM_ENQUEUE) &&
2196 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT)))
2197 OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_ENQ_RESEND, 6);
2199 ptlrpc_at_add_timed(req);
2201 if (opc != OST_CONNECT && opc != MDS_CONNECT &&
2202 opc != MGS_CONNECT && req->rq_export != NULL) {
2203 if (exp_connect_flags2(req->rq_export) & OBD_CONNECT2_REP_MBITS)
2204 req->rq_rep_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
2207 /* Move it over to the request processing queue */
2208 rc = ptlrpc_server_request_add(svcpt, req);
2212 wake_up(&svcpt->scp_waitq);
2216 ptlrpc_server_finish_request(svcpt, req);
2222 * Main incoming request handling logic.
2223 * Calls handler function from service to do actual processing.
2225 static int ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
2226 struct ptlrpc_thread *thread)
2228 struct ptlrpc_service *svc = svcpt->scp_service;
2229 struct ptlrpc_request *request;
2239 request = ptlrpc_server_request_get(svcpt, false);
2240 if (request == NULL)
2243 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
2244 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
2245 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
2246 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
2248 if (unlikely(fail_opc)) {
2249 if (request->rq_export && request->rq_ops)
2250 OBD_FAIL_TIMEOUT(fail_opc, 4);
2253 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
2255 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
2256 libcfs_debug_dumplog();
2258 work_start = ktime_get_real();
2259 arrived = timespec64_to_ktime(request->rq_arrival_time);
2260 timediff_usecs = ktime_us_delta(work_start, arrived);
2261 if (likely(svc->srv_stats != NULL)) {
2262 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
2264 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
2265 svcpt->scp_nreqs_incoming);
2266 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
2267 svcpt->scp_nreqs_active);
2268 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
2269 at_get(&svcpt->scp_at_estimate));
2272 if (likely(request->rq_export)) {
2273 if (unlikely(ptlrpc_check_req(request)))
2275 ptlrpc_update_export_timer(request->rq_export,
2276 div_u64(timediff_usecs,
2281 * Discard requests queued for longer than the deadline.
2282 * The deadline is increased if we send an early reply.
2284 if (ktime_get_real_seconds() > request->rq_deadline) {
2285 DEBUG_REQ(D_ERROR, request,
2286 "Dropping timed-out request from %s: deadline %lld/%llds ago",
2287 libcfs_id2str(request->rq_peer),
2288 request->rq_deadline -
2289 request->rq_arrival_time.tv_sec,
2290 ktime_get_real_seconds() - request->rq_deadline);
2295 "Handling RPC req@%p pname:cluuid+ref:pid:xid:nid:opc:job %s:%s+%d:%d:x%llu:%s:%d:%s\n",
2296 request, current->comm,
2297 (request->rq_export ?
2298 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2299 (request->rq_export ?
2300 refcount_read(&request->rq_export->exp_handle.h_ref) : -99),
2301 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
2302 libcfs_id2str(request->rq_peer),
2303 lustre_msg_get_opc(request->rq_reqmsg),
2304 lustre_msg_get_jobid(request->rq_reqmsg) ?: "");
2306 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
2307 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
2309 CDEBUG(D_NET, "got req %llu\n", request->rq_xid);
2311 /* re-assign request and sesson thread to the current one */
2312 request->rq_svc_thread = thread;
2313 if (thread != NULL) {
2314 LASSERT(request->rq_session.lc_thread == NULL);
2315 request->rq_session.lc_thread = thread;
2316 thread->t_env->le_ses = &request->rq_session;
2318 svc->srv_ops.so_req_handler(request);
2320 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
2323 if (unlikely(ktime_get_real_seconds() > request->rq_deadline)) {
2324 DEBUG_REQ(D_WARNING, request,
2325 "Request took longer than estimated (%lld/%llds); client may timeout",
2326 request->rq_deadline -
2327 request->rq_arrival_time.tv_sec,
2328 ktime_get_real_seconds() - request->rq_deadline);
2331 work_end = ktime_get_real();
2332 timediff_usecs = ktime_us_delta(work_end, work_start);
2333 arrived_usecs = ktime_us_delta(work_end, arrived);
2335 "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",
2336 request, current->comm,
2337 (request->rq_export ?
2338 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2339 (request->rq_export ?
2340 refcount_read(&request->rq_export->exp_handle.h_ref) : -99),
2341 lustre_msg_get_status(request->rq_reqmsg),
2343 libcfs_id2str(request->rq_peer),
2344 lustre_msg_get_opc(request->rq_reqmsg),
2345 lustre_msg_get_jobid(request->rq_reqmsg) ?: "",
2348 (request->rq_repmsg ?
2349 lustre_msg_get_transno(request->rq_repmsg) :
2350 request->rq_transno),
2352 (request->rq_repmsg ?
2353 lustre_msg_get_status(request->rq_repmsg) : -999));
2354 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
2355 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
2356 int opc = opcode_offset(op);
2358 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
2359 LASSERT(opc < LUSTRE_MAX_OPCODES);
2360 lprocfs_counter_add(svc->srv_stats,
2361 opc + EXTRA_MAX_OPCODES,
2365 if (unlikely(request->rq_early_count)) {
2366 DEBUG_REQ(D_ADAPTTO, request,
2367 "sent %d early replies before finishing in %llds",
2368 request->rq_early_count,
2369 div_u64(arrived_usecs, USEC_PER_SEC));
2372 ptlrpc_server_finish_active_request(svcpt, request);
2378 * An internal function to process a single reply state object.
2380 static int ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
2382 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
2383 struct ptlrpc_service *svc = svcpt->scp_service;
2384 struct obd_export *exp;
2390 exp = rs->rs_export;
2392 LASSERT(rs->rs_difficult);
2393 LASSERT(rs->rs_scheduled);
2394 LASSERT(list_empty(&rs->rs_list));
2397 * The disk commit callback holds exp_uncommitted_replies_lock while it
2398 * iterates over newly committed replies, removing them from
2399 * exp_uncommitted_replies. It then drops this lock and schedules the
2400 * replies it found for handling here.
2402 * We can avoid contention for exp_uncommitted_replies_lock between the
2403 * HRT threads and further commit callbacks by checking rs_committed
2404 * which is set in the commit callback while it holds both
2405 * rs_lock and exp_uncommitted_reples.
2407 * If we see rs_committed clear, the commit callback _may_ not have
2408 * handled this reply yet and we race with it to grab
2409 * exp_uncommitted_replies_lock before removing the reply from
2410 * exp_uncommitted_replies. Note that if we lose the race and the
2411 * reply has already been removed, list_del_init() is a noop.
2413 * If we see rs_committed set, we know the commit callback is handling,
2414 * or has handled this reply since store reordering might allow us to
2415 * see rs_committed set out of sequence. But since this is done
2416 * holding rs_lock, we can be sure it has all completed once we hold
2417 * rs_lock, which we do right next.
2419 if (!rs->rs_committed) {
2421 * if rs was commited, no need to convert locks, don't check
2422 * rs_committed here because rs may never be added into
2423 * exp_uncommitted_replies and this flag never be set, see
2424 * target_send_reply()
2426 if (rs->rs_convert_lock &&
2427 rs->rs_transno > exp->exp_last_committed) {
2428 struct ldlm_lock *lock;
2429 struct ldlm_lock *ack_locks[RS_MAX_LOCKS] = { NULL };
2431 spin_lock(&rs->rs_lock);
2432 if (rs->rs_convert_lock &&
2433 rs->rs_transno > exp->exp_last_committed) {
2434 nlocks = rs->rs_nlocks;
2435 while (nlocks-- > 0) {
2437 * NB don't assume rs is always handled
2438 * by the same service thread (see
2439 * ptlrpc_hr_select, so REP-ACK hr may
2440 * race with trans commit, while the
2441 * latter will release locks, get locks
2442 * here early to convert to COS mode
2445 lock = ldlm_handle2lock(
2446 &rs->rs_locks[nlocks]);
2448 ack_locks[nlocks] = lock;
2449 rs->rs_modes[nlocks] = LCK_COS;
2451 nlocks = rs->rs_nlocks;
2452 rs->rs_convert_lock = 0;
2454 * clear rs_scheduled so that commit callback
2455 * can schedule again
2457 rs->rs_scheduled = 0;
2458 spin_unlock(&rs->rs_lock);
2460 while (nlocks-- > 0) {
2461 lock = ack_locks[nlocks];
2462 ldlm_lock_mode_downgrade(lock, LCK_COS);
2463 LDLM_LOCK_PUT(lock);
2467 spin_unlock(&rs->rs_lock);
2470 spin_lock(&exp->exp_uncommitted_replies_lock);
2471 list_del_init(&rs->rs_obd_list);
2472 spin_unlock(&exp->exp_uncommitted_replies_lock);
2475 spin_lock(&exp->exp_lock);
2476 /* Noop if removed already */
2477 list_del_init(&rs->rs_exp_list);
2478 spin_unlock(&exp->exp_lock);
2480 spin_lock(&rs->rs_lock);
2482 been_handled = rs->rs_handled;
2485 nlocks = rs->rs_nlocks; /* atomic "steal", but */
2486 rs->rs_nlocks = 0; /* locks still on rs_locks! */
2488 if (nlocks == 0 && !been_handled) {
2490 * If we see this, we should already have seen the warning
2491 * in mds_steal_ack_locks()
2494 "All locks stolen from rs %p x%lld.t%lld o%d NID %s\n",
2495 rs, rs->rs_xid, rs->rs_transno, rs->rs_opc,
2496 libcfs_nid2str(exp->exp_connection->c_peer.nid));
2499 if ((rs->rs_sent && !rs->rs_unlinked) || nlocks > 0) {
2500 spin_unlock(&rs->rs_lock);
2502 /* We can unlink if the LNET_EVENT_SEND has occurred.
2503 * If rs_unlinked is set then MD is already unlinked and no
2504 * need to do so here.
2506 if ((rs->rs_sent && !rs->rs_unlinked)) {
2507 LNetMDUnlink(rs->rs_md_h);
2508 /* Ignore return code; we're racing with completion */
2511 while (nlocks-- > 0)
2512 ldlm_lock_decref(&rs->rs_locks[nlocks],
2513 rs->rs_modes[nlocks]);
2515 spin_lock(&rs->rs_lock);
2518 rs->rs_scheduled = 0;
2519 rs->rs_convert_lock = 0;
2521 if (rs->rs_unlinked) {
2523 spin_unlock(&rs->rs_lock);
2525 class_export_put(exp);
2526 rs->rs_export = NULL;
2527 ptlrpc_rs_decref(rs);
2528 if (atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
2529 svc->srv_is_stopping)
2530 wake_up_all(&svcpt->scp_waitq);
2534 /* still on the net; callback will schedule */
2535 spin_unlock(&rs->rs_lock);
2540 static void ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
2542 int avail = svcpt->scp_nrqbds_posted;
2543 int low_water = test_req_buffer_pressure ? 0 :
2544 svcpt->scp_service->srv_nbuf_per_group / 2;
2546 /* NB I'm not locking; just looking. */
2549 * CAVEAT EMPTOR: We might be allocating buffers here because we've
2550 * allowed the request history to grow out of control. We could put a
2551 * sanity check on that here and cull some history if we need the
2555 if (avail <= low_water)
2556 ptlrpc_grow_req_bufs(svcpt, 1);
2558 if (svcpt->scp_service->srv_stats) {
2559 lprocfs_counter_add(svcpt->scp_service->srv_stats,
2560 PTLRPC_REQBUF_AVAIL_CNTR, avail);
2564 static inline int ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
2566 return svcpt->scp_nreqs_active <
2567 svcpt->scp_nthrs_running - 1 -
2568 (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL);
2572 * allowed to create more threads
2573 * user can call it w/o any lock but need to hold
2574 * ptlrpc_service_part::scp_lock to get reliable result
2576 static inline int ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
2578 return svcpt->scp_nthrs_running +
2579 svcpt->scp_nthrs_starting <
2580 svcpt->scp_service->srv_nthrs_cpt_limit;
2584 * too many requests and allowed to create more threads
2586 static inline int ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
2588 return !ptlrpc_threads_enough(svcpt) &&
2589 ptlrpc_threads_increasable(svcpt);
2592 static inline int ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2594 return thread_is_stopping(thread) ||
2595 thread->t_svcpt->scp_service->srv_is_stopping;
2598 /* stop the highest numbered thread if there are too many threads running */
2599 static inline bool ptlrpc_thread_should_stop(struct ptlrpc_thread *thread)
2601 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2603 return thread->t_id >= svcpt->scp_service->srv_nthrs_cpt_limit &&
2604 thread->t_id == svcpt->scp_thr_nextid - 1;
2607 static void ptlrpc_stop_thread(struct ptlrpc_thread *thread)
2609 CDEBUG(D_INFO, "Stopping thread %s #%u\n",
2610 thread->t_svcpt->scp_service->srv_thread_name, thread->t_id);
2611 thread_add_flags(thread, SVC_STOPPING);
2614 static inline void ptlrpc_thread_stop(struct ptlrpc_thread *thread)
2616 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2618 spin_lock(&svcpt->scp_lock);
2619 if (ptlrpc_thread_should_stop(thread)) {
2620 ptlrpc_stop_thread(thread);
2621 svcpt->scp_thr_nextid--;
2623 spin_unlock(&svcpt->scp_lock);
2626 static inline int ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
2628 return !list_empty(&svcpt->scp_rqbd_idle) &&
2629 svcpt->scp_rqbd_timeout == 0;
2633 ptlrpc_at_check(struct ptlrpc_service_part *svcpt)
2635 return svcpt->scp_at_check;
2639 * If a thread runs too long or spends to much time on a single request,
2640 * we want to know about it, so we set up a delayed work item as a watchdog.
2641 * If it fires, we display a stack trace of the delayed thread,
2642 * providing we aren't rate-limited
2644 * Watchdog stack traces are limited to 3 per 'libcfs_watchdog_ratelimit'
2647 static struct ratelimit_state watchdog_limit;
2649 static void ptlrpc_watchdog_fire(struct work_struct *w)
2651 struct ptlrpc_thread *thread = container_of(w, struct ptlrpc_thread,
2653 u64 ms_lapse = ktime_ms_delta(ktime_get(), thread->t_touched);
2654 u32 ms_frac = do_div(ms_lapse, MSEC_PER_SEC);
2656 /* ___ratelimit() returns true if the action is NOT ratelimited */
2657 if (__ratelimit(&watchdog_limit)) {
2658 /* below message is checked in sanity-quota.sh test_6,18 */
2659 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",
2660 thread->t_task->comm, thread->t_task->pid,
2663 libcfs_debug_dumpstack(thread->t_task);
2665 /* below message is checked in sanity-quota.sh test_6,18 */
2666 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",
2667 thread->t_task->comm, thread->t_task->pid,
2668 ms_lapse, ms_frac, libcfs_watchdog_ratelimit);
2672 void ptlrpc_watchdog_init(struct delayed_work *work, timeout_t timeout)
2674 INIT_DELAYED_WORK(work, ptlrpc_watchdog_fire);
2675 schedule_delayed_work(work, cfs_time_seconds(timeout));
2678 void ptlrpc_watchdog_disable(struct delayed_work *work)
2680 cancel_delayed_work_sync(work);
2683 void ptlrpc_watchdog_touch(struct delayed_work *work, timeout_t timeout)
2685 struct ptlrpc_thread *thread = container_of(&work->work,
2686 struct ptlrpc_thread,
2688 thread->t_touched = ktime_get();
2689 mod_delayed_work(system_wq, work, cfs_time_seconds(timeout));
2693 * requests wait on preprocessing
2694 * user can call it w/o any lock but need to hold
2695 * ptlrpc_service_part::scp_lock to get reliable result
2698 ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
2700 return !list_empty(&svcpt->scp_req_incoming);
2703 static __attribute__((__noinline__)) int
2704 ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
2705 struct ptlrpc_thread *thread)
2707 ptlrpc_watchdog_disable(&thread->t_watchdog);
2711 if (svcpt->scp_rqbd_timeout == 0)
2712 /* Don't exit while there are replies to be handled */
2713 wait_event_idle_exclusive_lifo(
2715 ptlrpc_thread_stopping(thread) ||
2716 ptlrpc_server_request_incoming(svcpt) ||
2717 ptlrpc_server_request_pending(svcpt, false) ||
2718 ptlrpc_rqbd_pending(svcpt) ||
2719 ptlrpc_at_check(svcpt));
2720 else if (wait_event_idle_exclusive_lifo_timeout(
2722 ptlrpc_thread_stopping(thread) ||
2723 ptlrpc_server_request_incoming(svcpt) ||
2724 ptlrpc_server_request_pending(svcpt, false) ||
2725 ptlrpc_rqbd_pending(svcpt) ||
2726 ptlrpc_at_check(svcpt),
2727 svcpt->scp_rqbd_timeout) == 0)
2728 svcpt->scp_rqbd_timeout = 0;
2730 if (ptlrpc_thread_stopping(thread))
2733 ptlrpc_watchdog_touch(&thread->t_watchdog,
2734 ptlrpc_server_get_timeout(svcpt));
2739 * Main thread body for service threads.
2740 * Waits in a loop waiting for new requests to process to appear.
2741 * Every time an incoming requests is added to its queue, a waitq
2742 * is woken up and one of the threads will handle it.
2744 static int ptlrpc_main(void *arg)
2746 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
2747 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2748 struct ptlrpc_service *svc = svcpt->scp_service;
2749 struct ptlrpc_reply_state *rs;
2750 struct group_info *ginfo = NULL;
2752 int counter = 0, rc = 0;
2756 thread->t_task = current;
2757 thread->t_pid = current->pid;
2759 if (svc->srv_cpt_bind) {
2760 rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
2762 CWARN("%s: failed to bind %s on CPT %d\n",
2763 svc->srv_name, thread->t_name, svcpt->scp_cpt);
2767 ginfo = groups_alloc(0);
2769 GOTO(out, rc = -ENOMEM);
2771 set_current_groups(ginfo);
2772 put_group_info(ginfo);
2774 if (svc->srv_ops.so_thr_init != NULL) {
2775 rc = svc->srv_ops.so_thr_init(thread);
2782 GOTO(out_srv_fini, rc = -ENOMEM);
2783 rc = lu_env_add(env);
2787 rc = lu_context_init(&env->le_ctx,
2788 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2790 GOTO(out_env_remove, rc);
2792 thread->t_env = env;
2793 env->le_ctx.lc_thread = thread;
2794 env->le_ctx.lc_cookie = 0x6;
2796 while (!list_empty(&svcpt->scp_rqbd_idle)) {
2797 rc = ptlrpc_server_post_idle_rqbds(svcpt);
2801 CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
2802 svc->srv_name, svcpt->scp_cpt, rc);
2803 GOTO(out_ctx_fini, rc);
2806 /* Alloc reply state structure for this one */
2807 OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2809 GOTO(out_ctx_fini, rc = -ENOMEM);
2811 spin_lock(&svcpt->scp_lock);
2813 LASSERT(thread_is_starting(thread));
2814 thread_clear_flags(thread, SVC_STARTING);
2816 LASSERT(svcpt->scp_nthrs_starting == 1);
2817 svcpt->scp_nthrs_starting--;
2820 * SVC_STOPPING may already be set here if someone else is trying
2821 * to stop the service while this new thread has been dynamically
2822 * forked. We still set SVC_RUNNING to let our creator know that
2823 * we are now running, however we will exit as soon as possible
2825 thread_add_flags(thread, SVC_RUNNING);
2826 svcpt->scp_nthrs_running++;
2827 spin_unlock(&svcpt->scp_lock);
2829 /* wake up our creator in case he's still waiting. */
2830 wake_up(&thread->t_ctl_waitq);
2832 thread->t_touched = ktime_get();
2833 ptlrpc_watchdog_init(&thread->t_watchdog,
2834 ptlrpc_server_get_timeout(svcpt));
2836 spin_lock(&svcpt->scp_rep_lock);
2837 list_add(&rs->rs_list, &svcpt->scp_rep_idle);
2838 wake_up(&svcpt->scp_rep_waitq);
2839 spin_unlock(&svcpt->scp_rep_lock);
2841 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2842 svcpt->scp_nthrs_running);
2844 /* XXX maintain a list of all managed devices: insert here */
2845 while (!ptlrpc_thread_stopping(thread)) {
2846 if (ptlrpc_wait_event(svcpt, thread))
2849 ptlrpc_check_rqbd_pool(svcpt);
2851 if (ptlrpc_threads_need_create(svcpt)) {
2852 /* Ignore return code - we tried... */
2853 ptlrpc_start_thread(svcpt, 0);
2856 /* reset le_ses to initial state */
2858 /* Refill the context before execution to make sure
2859 * all thread keys are allocated */
2861 /* Process all incoming reqs before handling any */
2862 if (ptlrpc_server_request_incoming(svcpt)) {
2863 lu_context_enter(&env->le_ctx);
2864 ptlrpc_server_handle_req_in(svcpt, thread);
2865 lu_context_exit(&env->le_ctx);
2867 /* but limit ourselves in case of flood */
2868 if (counter++ < 100)
2873 if (ptlrpc_at_check(svcpt))
2874 ptlrpc_at_check_timed(svcpt);
2876 if (ptlrpc_server_request_pending(svcpt, false)) {
2877 lu_context_enter(&env->le_ctx);
2878 ptlrpc_server_handle_request(svcpt, thread);
2879 lu_context_exit(&env->le_ctx);
2882 if (ptlrpc_rqbd_pending(svcpt) &&
2883 ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
2885 * I just failed to repost request buffers.
2886 * Wait for a timeout (unless something else
2887 * happens) before I try again
2889 svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
2890 CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
2891 svcpt->scp_nrqbds_posted);
2894 * If the number of threads has been tuned downward and this
2895 * thread should be stopped, then stop in reverse order so the
2896 * the threads always have contiguous thread index values.
2898 if (unlikely(ptlrpc_thread_should_stop(thread)))
2899 ptlrpc_thread_stop(thread);
2902 ptlrpc_watchdog_disable(&thread->t_watchdog);
2905 lu_context_fini(&env->le_ctx);
2911 /* deconstruct service thread state created by ptlrpc_start_thread() */
2912 if (svc->srv_ops.so_thr_done != NULL)
2913 svc->srv_ops.so_thr_done(thread);
2915 CDEBUG(D_RPCTRACE, "%s: service thread [%p:%u] %d exiting: rc = %d\n",
2916 thread->t_name, thread, thread->t_pid, thread->t_id, rc);
2917 spin_lock(&svcpt->scp_lock);
2918 if (thread_test_and_clear_flags(thread, SVC_STARTING))
2919 svcpt->scp_nthrs_starting--;
2921 if (thread_test_and_clear_flags(thread, SVC_RUNNING)) {
2922 /* must know immediately */
2923 svcpt->scp_nthrs_running--;
2927 thread_add_flags(thread, SVC_STOPPED);
2929 wake_up(&thread->t_ctl_waitq);
2930 spin_unlock(&svcpt->scp_lock);
2935 static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
2936 struct list_head *replies)
2940 spin_lock(&hrt->hrt_lock);
2942 list_splice_init(&hrt->hrt_queue, replies);
2943 result = ptlrpc_hr.hr_stopping || !list_empty(replies);
2945 spin_unlock(&hrt->hrt_lock);
2950 * Main body of "handle reply" function.
2951 * It processes acked reply states
2953 static int ptlrpc_hr_main(void *arg)
2955 struct ptlrpc_hr_thread *hrt = (struct ptlrpc_hr_thread *)arg;
2956 struct ptlrpc_hr_partition *hrp = hrt->hrt_partition;
2965 rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
2967 char threadname[20];
2969 snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
2970 hrp->hrp_cpt, hrt->hrt_id);
2971 CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
2972 threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
2975 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD |
2976 LCT_REMEMBER | LCT_NOREF);
2980 rc = lu_env_add(env);
2982 GOTO(out_ctx_fini, rc);
2984 atomic_inc(&hrp->hrp_nstarted);
2985 wake_up(&ptlrpc_hr.hr_waitq);
2987 while (!ptlrpc_hr.hr_stopping) {
2988 wait_event_idle(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
2990 while (!list_empty(&replies)) {
2991 struct ptlrpc_reply_state *rs;
2993 rs = list_entry(replies.prev,
2994 struct ptlrpc_reply_state,
2996 list_del_init(&rs->rs_list);
2997 /* refill keys if needed */
2999 lu_context_enter(&env->le_ctx);
3000 ptlrpc_handle_rs(rs);
3001 lu_context_exit(&env->le_ctx);
3005 atomic_inc(&hrp->hrp_nstopped);
3006 wake_up(&ptlrpc_hr.hr_waitq);
3010 lu_context_fini(&env->le_ctx);
3016 static void ptlrpc_stop_hr_threads(void)
3018 struct ptlrpc_hr_partition *hrp;
3022 ptlrpc_hr.hr_stopping = 1;
3024 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
3025 if (hrp->hrp_thrs == NULL)
3026 continue; /* uninitialized */
3027 for (j = 0; j < hrp->hrp_nthrs; j++)
3028 wake_up(&hrp->hrp_thrs[j].hrt_waitq);
3031 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
3032 if (hrp->hrp_thrs == NULL)
3033 continue; /* uninitialized */
3034 wait_event(ptlrpc_hr.hr_waitq,
3035 atomic_read(&hrp->hrp_nstopped) ==
3036 atomic_read(&hrp->hrp_nstarted));
3040 static int ptlrpc_start_hr_threads(void)
3042 struct ptlrpc_hr_partition *hrp;
3048 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
3051 for (j = 0; j < hrp->hrp_nthrs; j++) {
3052 struct ptlrpc_hr_thread *hrt = &hrp->hrp_thrs[j];
3053 struct task_struct *task;
3055 task = kthread_run(ptlrpc_hr_main,
3057 "ptlrpc_hr%02d_%03d",
3066 wait_event(ptlrpc_hr.hr_waitq,
3067 atomic_read(&hrp->hrp_nstarted) == j);
3070 CERROR("cannot start reply handler thread %d:%d: rc = %d\n",
3072 ptlrpc_stop_hr_threads();
3080 static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
3082 struct ptlrpc_thread *thread;
3087 CDEBUG(D_INFO, "Stopping threads for service %s\n",
3088 svcpt->scp_service->srv_name);
3090 spin_lock(&svcpt->scp_lock);
3091 /* let the thread know that we would like it to stop asap */
3092 list_for_each_entry(thread, &svcpt->scp_threads, t_link)
3093 ptlrpc_stop_thread(thread);
3095 wake_up_all(&svcpt->scp_waitq);
3097 while ((thread = list_first_entry_or_null(&svcpt->scp_threads,
3098 struct ptlrpc_thread,
3100 if (thread_is_stopped(thread)) {
3101 list_move(&thread->t_link, &zombie);
3104 spin_unlock(&svcpt->scp_lock);
3106 CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
3107 svcpt->scp_service->srv_thread_name, thread->t_id);
3108 wait_event_idle(thread->t_ctl_waitq,
3109 thread_is_stopped(thread));
3111 spin_lock(&svcpt->scp_lock);
3114 spin_unlock(&svcpt->scp_lock);
3116 while ((thread = list_first_entry_or_null(&zombie,
3117 struct ptlrpc_thread,
3119 list_del(&thread->t_link);
3120 OBD_FREE_PTR(thread);
3126 * Stops all threads of a particular service \a svc
3128 static void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
3130 struct ptlrpc_service_part *svcpt;
3135 ptlrpc_service_for_each_part(svcpt, i, svc) {
3136 if (svcpt->scp_service != NULL)
3137 ptlrpc_svcpt_stop_threads(svcpt);
3143 static int ptlrpc_start_threads(struct ptlrpc_service *svc)
3151 /* We require 2 threads min, see note in ptlrpc_server_handle_request */
3152 LASSERT(svc->srv_nthrs_cpt_init >= PTLRPC_NTHRS_INIT);
3154 for (i = 0; i < svc->srv_ncpts; i++) {
3155 for (j = 0; j < svc->srv_nthrs_cpt_init; j++) {
3156 rc = ptlrpc_start_thread(svc->srv_parts[i], 1);
3162 /* We have enough threads, don't start more. b=15759 */
3169 CERROR("cannot start %s thread #%d_%d: rc %d\n",
3170 svc->srv_thread_name, i, j, rc);
3171 ptlrpc_stop_all_threads(svc);
3175 static int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
3177 struct ptlrpc_thread *thread;
3178 struct ptlrpc_service *svc;
3179 struct task_struct *task;
3184 LASSERT(svcpt != NULL);
3186 svc = svcpt->scp_service;
3188 CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n",
3189 svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running,
3190 svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit);
3193 if (unlikely(svc->srv_is_stopping))
3196 if (!ptlrpc_threads_increasable(svcpt) ||
3197 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
3198 svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1))
3201 OBD_CPT_ALLOC_PTR(thread, svc->srv_cptable, svcpt->scp_cpt);
3204 init_waitqueue_head(&thread->t_ctl_waitq);
3206 spin_lock(&svcpt->scp_lock);
3207 if (!ptlrpc_threads_increasable(svcpt)) {
3208 spin_unlock(&svcpt->scp_lock);
3209 OBD_FREE_PTR(thread);
3213 if (svcpt->scp_nthrs_starting != 0) {
3215 * serialize starting because some modules (obdfilter)
3216 * might require unique and contiguous t_id
3218 LASSERT(svcpt->scp_nthrs_starting == 1);
3219 spin_unlock(&svcpt->scp_lock);
3220 OBD_FREE_PTR(thread);
3222 CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n",
3223 svc->srv_thread_name, svcpt->scp_thr_nextid);
3228 CDEBUG(D_INFO, "Creating thread %s #%d race, retry later\n",
3229 svc->srv_thread_name, svcpt->scp_thr_nextid);
3233 svcpt->scp_nthrs_starting++;
3234 thread->t_id = svcpt->scp_thr_nextid++;
3235 thread_add_flags(thread, SVC_STARTING);
3236 thread->t_svcpt = svcpt;
3238 list_add(&thread->t_link, &svcpt->scp_threads);
3239 spin_unlock(&svcpt->scp_lock);
3241 if (svcpt->scp_cpt >= 0) {
3242 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s%02d_%03d",
3243 svc->srv_thread_name, svcpt->scp_cpt, thread->t_id);
3245 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s_%04d",
3246 svc->srv_thread_name, thread->t_id);
3249 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name);
3250 task = kthread_run(ptlrpc_main, thread, "%s", thread->t_name);
3253 CERROR("cannot start thread '%s': rc = %d\n",
3254 thread->t_name, rc);
3255 spin_lock(&svcpt->scp_lock);
3256 --svcpt->scp_nthrs_starting;
3257 if (thread_is_stopping(thread)) {
3259 * this ptlrpc_thread is being hanled
3260 * by ptlrpc_svcpt_stop_threads now
3262 thread_add_flags(thread, SVC_STOPPED);
3263 wake_up(&thread->t_ctl_waitq);
3264 spin_unlock(&svcpt->scp_lock);
3266 list_del(&thread->t_link);
3267 spin_unlock(&svcpt->scp_lock);
3268 OBD_FREE_PTR(thread);
3276 wait_event_idle(thread->t_ctl_waitq,
3277 thread_is_running(thread) || thread_is_stopped(thread));
3279 rc = thread_is_stopped(thread) ? thread->t_id : 0;
3283 int ptlrpc_hr_init(void)
3285 struct ptlrpc_hr_partition *hrp;
3286 struct ptlrpc_hr_thread *hrt;
3294 memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
3295 ptlrpc_hr.hr_cpt_table = cfs_cpt_tab;
3297 ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
3299 if (ptlrpc_hr.hr_partitions == NULL)
3302 ratelimit_state_init(&watchdog_limit,
3303 cfs_time_seconds(libcfs_watchdog_ratelimit), 3);
3305 init_waitqueue_head(&ptlrpc_hr.hr_waitq);
3308 weight = cpumask_weight(topology_sibling_cpumask(smp_processor_id()));
3311 cfs_percpt_for_each(hrp, cpt, ptlrpc_hr.hr_partitions) {
3314 atomic_set(&hrp->hrp_nstarted, 0);
3315 atomic_set(&hrp->hrp_nstopped, 0);
3317 hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, cpt);
3318 hrp->hrp_nthrs /= weight;
3319 if (hrp->hrp_nthrs == 0)
3322 OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, cpt,
3323 hrp->hrp_nthrs * sizeof(*hrt));
3324 if (hrp->hrp_thrs == NULL)
3325 GOTO(out, rc = -ENOMEM);
3327 for (i = 0; i < hrp->hrp_nthrs; i++) {
3328 hrt = &hrp->hrp_thrs[i];
3331 hrt->hrt_partition = hrp;
3332 init_waitqueue_head(&hrt->hrt_waitq);
3333 spin_lock_init(&hrt->hrt_lock);
3334 INIT_LIST_HEAD(&hrt->hrt_queue);
3338 rc = ptlrpc_start_hr_threads();
3345 void ptlrpc_hr_fini(void)
3347 struct ptlrpc_hr_partition *hrp;
3350 if (ptlrpc_hr.hr_partitions == NULL)
3353 ptlrpc_stop_hr_threads();
3355 cfs_percpt_for_each(hrp, cpt, ptlrpc_hr.hr_partitions) {
3357 OBD_FREE_PTR_ARRAY(hrp->hrp_thrs, hrp->hrp_nthrs);
3360 cfs_percpt_free(ptlrpc_hr.hr_partitions);
3361 ptlrpc_hr.hr_partitions = NULL;
3366 * Wait until all already scheduled replies are processed.
3368 static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
3371 if (wait_event_idle_timeout(
3373 atomic_read(&svcpt->scp_nreps_difficult) == 0,
3374 cfs_time_seconds(10)) > 0)
3376 CWARN("Unexpectedly long timeout %s %p\n",
3377 svcpt->scp_service->srv_name, svcpt->scp_service);
3382 ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
3384 struct ptlrpc_service_part *svcpt;
3387 /* early disarm AT timer... */
3388 ptlrpc_service_for_each_part(svcpt, i, svc) {
3389 if (svcpt->scp_service != NULL)
3390 del_timer(&svcpt->scp_at_timer);
3395 ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
3397 struct ptlrpc_service_part *svcpt;
3398 struct ptlrpc_request_buffer_desc *rqbd;
3403 * All history will be culled when the next request buffer is
3404 * freed in ptlrpc_service_purge_all()
3406 svc->srv_hist_nrqbds_cpt_max = 0;
3408 rc = LNetClearLazyPortal(svc->srv_req_portal);
3411 ptlrpc_service_for_each_part(svcpt, i, svc) {
3412 if (svcpt->scp_service == NULL)
3416 * Unlink all the request buffers. This forces a 'final'
3417 * event with its 'unlink' flag set for each posted rqbd
3419 list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
3421 rc = LNetMDUnlink(rqbd->rqbd_md_h);
3422 LASSERT(rc == 0 || rc == -ENOENT);
3426 ptlrpc_service_for_each_part(svcpt, i, svc) {
3427 if (svcpt->scp_service == NULL)
3431 * Wait for the network to release any buffers
3432 * it's currently filling
3434 spin_lock(&svcpt->scp_lock);
3435 while (svcpt->scp_nrqbds_posted != 0) {
3436 int seconds = PTLRPC_REQ_LONG_UNLINK;
3438 spin_unlock(&svcpt->scp_lock);
3440 * Network access will complete in finite time but
3441 * the HUGE timeout lets us CWARN for visibility
3444 while (seconds > 0 &&
3445 wait_event_idle_timeout(
3447 svcpt->scp_nrqbds_posted == 0,
3448 cfs_time_seconds(1)) == 0)
3451 CWARN("Service %s waiting for request buffers\n",
3452 svcpt->scp_service->srv_name);
3454 spin_lock(&svcpt->scp_lock);
3456 spin_unlock(&svcpt->scp_lock);
3461 ptlrpc_service_purge_all(struct ptlrpc_service *svc)
3463 struct ptlrpc_service_part *svcpt;
3464 struct ptlrpc_request_buffer_desc *rqbd;
3465 struct ptlrpc_request *req;
3466 struct ptlrpc_reply_state *rs;
3469 ptlrpc_service_for_each_part(svcpt, i, svc) {
3470 if (svcpt->scp_service == NULL)
3473 spin_lock(&svcpt->scp_rep_lock);
3474 while ((rs = list_first_entry_or_null(&svcpt->scp_rep_active,
3475 struct ptlrpc_reply_state,
3476 rs_list)) != NULL) {
3477 spin_lock(&rs->rs_lock);
3478 ptlrpc_schedule_difficult_reply(rs);
3479 spin_unlock(&rs->rs_lock);
3481 spin_unlock(&svcpt->scp_rep_lock);
3484 * purge the request queue. NB No new replies (rqbds
3485 * all unlinked) and no service threads, so I'm the only
3486 * thread noodling the request queue now
3488 while ((req = list_first_entry_or_null(&svcpt->scp_req_incoming,
3489 struct ptlrpc_request,
3490 rq_list)) != NULL) {
3491 list_del(&req->rq_list);
3492 svcpt->scp_nreqs_incoming--;
3493 ptlrpc_server_finish_request(svcpt, req);
3496 while (ptlrpc_server_request_pending(svcpt, true)) {
3497 req = ptlrpc_server_request_get(svcpt, true);
3498 ptlrpc_server_finish_active_request(svcpt, req);
3502 * The portal may be shared by several services (eg:OUT_PORTAL).
3503 * So the request could be referenced by other target. So we
3504 * have to wait the ptlrpc_server_drop_request invoked.
3506 * TODO: move the req_buffer as global rather than per service.
3508 spin_lock(&svcpt->scp_lock);
3509 while (!list_empty(&svcpt->scp_rqbd_posted)) {
3510 spin_unlock(&svcpt->scp_lock);
3511 wait_event_idle_timeout(svcpt->scp_waitq,
3512 list_empty(&svcpt->scp_rqbd_posted),
3513 cfs_time_seconds(1));
3514 spin_lock(&svcpt->scp_lock);
3516 spin_unlock(&svcpt->scp_lock);
3518 LASSERT(svcpt->scp_nreqs_incoming == 0);
3519 LASSERT(svcpt->scp_nreqs_active == 0);
3521 * history should have been culled by
3522 * ptlrpc_server_finish_request
3524 LASSERT(svcpt->scp_hist_nrqbds == 0);
3527 * Now free all the request buffers since nothing
3528 * references them any more...
3530 while ((rqbd = list_first_entry_or_null(&svcpt->scp_rqbd_idle,
3531 struct ptlrpc_request_buffer_desc,
3532 rqbd_list)) != NULL)
3533 ptlrpc_free_rqbd(rqbd);
3535 ptlrpc_wait_replies(svcpt);
3537 while ((rs = list_first_entry_or_null(&svcpt->scp_rep_idle,
3538 struct ptlrpc_reply_state,
3539 rs_list)) != NULL) {
3540 list_del(&rs->rs_list);
3541 OBD_FREE_LARGE(rs, svc->srv_max_reply_size);
3547 ptlrpc_service_free(struct ptlrpc_service *svc)
3549 struct ptlrpc_service_part *svcpt;
3550 struct ptlrpc_at_array *array;
3553 ptlrpc_service_for_each_part(svcpt, i, svc) {
3554 if (svcpt->scp_service == NULL)
3557 /* In case somebody rearmed this in the meantime */
3558 del_timer(&svcpt->scp_at_timer);
3559 array = &svcpt->scp_at_array;
3561 if (array->paa_reqs_array != NULL) {
3562 OBD_FREE_PTR_ARRAY(array->paa_reqs_array,
3564 array->paa_reqs_array = NULL;
3567 if (array->paa_reqs_count != NULL) {
3568 OBD_FREE_PTR_ARRAY(array->paa_reqs_count,
3570 array->paa_reqs_count = NULL;
3574 ptlrpc_service_for_each_part(svcpt, i, svc)
3575 OBD_FREE_PTR(svcpt);
3577 if (svc->srv_cpts != NULL)
3578 cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts);
3580 OBD_FREE(svc, offsetof(struct ptlrpc_service,
3581 srv_parts[svc->srv_ncpts]));
3584 int ptlrpc_unregister_service(struct ptlrpc_service *service)
3588 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
3590 service->srv_is_stopping = 1;
3592 mutex_lock(&ptlrpc_all_services_mutex);
3593 list_del_init(&service->srv_list);
3594 mutex_unlock(&ptlrpc_all_services_mutex);
3596 ptlrpc_service_del_atimer(service);
3597 ptlrpc_stop_all_threads(service);
3599 ptlrpc_service_unlink_rqbd(service);
3600 ptlrpc_service_purge_all(service);
3601 ptlrpc_service_nrs_cleanup(service);
3603 ptlrpc_lprocfs_unregister_service(service);
3604 ptlrpc_sysfs_unregister_service(service);
3606 ptlrpc_service_free(service);
3610 EXPORT_SYMBOL(ptlrpc_unregister_service);
3613 * Returns 0 if the service is healthy.
3615 * Right now, it just checks to make sure that requests aren't languishing
3616 * in the queue. We'll use this health check to govern whether a node needs
3617 * to be shot, so it's intentionally non-aggressive.
3619 static int ptlrpc_svcpt_health_check(struct ptlrpc_service_part *svcpt)
3621 struct ptlrpc_request *request = NULL;
3622 struct timespec64 right_now;
3623 struct timespec64 timediff;
3625 ktime_get_real_ts64(&right_now);
3627 spin_lock(&svcpt->scp_req_lock);
3628 /* How long has the next entry been waiting? */
3629 if (ptlrpc_server_high_pending(svcpt, true))
3630 request = ptlrpc_nrs_req_peek_nolock(svcpt, true);
3631 else if (ptlrpc_server_normal_pending(svcpt, true))
3632 request = ptlrpc_nrs_req_peek_nolock(svcpt, false);
3634 if (request == NULL) {
3635 spin_unlock(&svcpt->scp_req_lock);
3639 timediff = timespec64_sub(right_now, request->rq_arrival_time);
3640 spin_unlock(&svcpt->scp_req_lock);
3642 if ((timediff.tv_sec) >
3643 (AT_OFF ? obd_timeout * 3 / 2 : at_max)) {
3644 CERROR("%s: unhealthy - request has been waiting %llds\n",
3645 svcpt->scp_service->srv_name, (s64)timediff.tv_sec);
3653 ptlrpc_service_health_check(struct ptlrpc_service *svc)
3655 struct ptlrpc_service_part *svcpt;
3661 ptlrpc_service_for_each_part(svcpt, i, svc) {
3662 int rc = ptlrpc_svcpt_health_check(svcpt);
3669 EXPORT_SYMBOL(ptlrpc_service_health_check);