4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
33 #define DEBUG_SUBSYSTEM S_RPC
35 #include <linux/kthread.h>
36 #include <linux/ratelimit.h>
38 #include <obd_support.h>
39 #include <obd_class.h>
40 #include <lustre_net.h>
41 #include <lu_object.h>
42 #include <uapi/linux/lnet/lnet-types.h>
43 #include "ptlrpc_internal.h"
44 #include <linux/delay.h>
46 /* The following are visible and mutable through /sys/module/ptlrpc */
47 int test_req_buffer_pressure = 0;
48 module_param(test_req_buffer_pressure, int, 0444);
49 MODULE_PARM_DESC(test_req_buffer_pressure, "set non-zero to put pressure on request buffer pools");
50 module_param(at_min, int, 0644);
51 MODULE_PARM_DESC(at_min, "Adaptive timeout minimum (sec)");
52 module_param(at_max, int, 0644);
53 MODULE_PARM_DESC(at_max, "Adaptive timeout maximum (sec)");
54 module_param(at_history, int, 0644);
55 MODULE_PARM_DESC(at_history,
56 "Adaptive timeouts remember the slowest event that took place within this period (sec)");
57 module_param(at_early_margin, int, 0644);
58 MODULE_PARM_DESC(at_early_margin, "How soon before an RPC deadline to send an early reply");
59 module_param(at_extra, int, 0644);
60 MODULE_PARM_DESC(at_extra, "How much extra time to give with each early reply");
63 static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt);
64 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req);
65 static void ptlrpc_at_remove_timed(struct ptlrpc_request *req);
67 /** Holds a list of all PTLRPC services */
68 LIST_HEAD(ptlrpc_all_services);
69 /** Used to protect the \e ptlrpc_all_services list */
70 struct mutex ptlrpc_all_services_mutex;
72 static struct ptlrpc_request_buffer_desc *
73 ptlrpc_alloc_rqbd(struct ptlrpc_service_part *svcpt)
75 struct ptlrpc_service *svc = svcpt->scp_service;
76 struct ptlrpc_request_buffer_desc *rqbd;
78 OBD_CPT_ALLOC_PTR(rqbd, svc->srv_cptable, svcpt->scp_cpt);
82 rqbd->rqbd_svcpt = svcpt;
83 rqbd->rqbd_refcount = 0;
84 rqbd->rqbd_cbid.cbid_fn = request_in_callback;
85 rqbd->rqbd_cbid.cbid_arg = rqbd;
86 INIT_LIST_HEAD(&rqbd->rqbd_reqs);
87 OBD_CPT_ALLOC_LARGE(rqbd->rqbd_buffer, svc->srv_cptable,
88 svcpt->scp_cpt, svc->srv_buf_size);
89 if (rqbd->rqbd_buffer == NULL) {
94 spin_lock(&svcpt->scp_lock);
95 list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
96 svcpt->scp_nrqbds_total++;
97 spin_unlock(&svcpt->scp_lock);
102 static void ptlrpc_free_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
104 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
106 LASSERT(rqbd->rqbd_refcount == 0);
107 LASSERT(list_empty(&rqbd->rqbd_reqs));
109 spin_lock(&svcpt->scp_lock);
110 list_del(&rqbd->rqbd_list);
111 svcpt->scp_nrqbds_total--;
112 spin_unlock(&svcpt->scp_lock);
114 OBD_FREE_LARGE(rqbd->rqbd_buffer, svcpt->scp_service->srv_buf_size);
118 static int ptlrpc_grow_req_bufs(struct ptlrpc_service_part *svcpt, int post)
120 struct ptlrpc_service *svc = svcpt->scp_service;
121 struct ptlrpc_request_buffer_desc *rqbd;
125 if (svcpt->scp_rqbd_allocating)
128 spin_lock(&svcpt->scp_lock);
129 /* check again with lock */
130 if (svcpt->scp_rqbd_allocating) {
131 /* NB: we might allow more than one thread in the future */
132 LASSERT(svcpt->scp_rqbd_allocating == 1);
133 spin_unlock(&svcpt->scp_lock);
137 svcpt->scp_rqbd_allocating++;
138 spin_unlock(&svcpt->scp_lock);
141 for (i = 0; i < svc->srv_nbuf_per_group; i++) {
143 * NB: another thread might have recycled enough rqbds, we
144 * need to make sure it wouldn't over-allocate, see LU-1212.
146 if (svcpt->scp_nrqbds_posted >= svc->srv_nbuf_per_group ||
147 (svc->srv_nrqbds_max != 0 &&
148 svcpt->scp_nrqbds_total > svc->srv_nrqbds_max))
151 rqbd = ptlrpc_alloc_rqbd(svcpt);
154 CERROR("%s: Can't allocate request buffer\n",
161 spin_lock(&svcpt->scp_lock);
163 LASSERT(svcpt->scp_rqbd_allocating == 1);
164 svcpt->scp_rqbd_allocating--;
166 spin_unlock(&svcpt->scp_lock);
169 "%s: allocate %d new %d-byte reqbufs (%d/%d left), rc = %d\n",
170 svc->srv_name, i, svc->srv_buf_size, svcpt->scp_nrqbds_posted,
171 svcpt->scp_nrqbds_total, rc);
175 rc = ptlrpc_server_post_idle_rqbds(svcpt);
181 * Part of Rep-Ack logic.
182 * Puts a lock and its mode into reply state assotiated to request reply.
184 void ptlrpc_save_lock(struct ptlrpc_request *req, struct lustre_handle *lock,
185 int mode, bool no_ack, bool convert_lock)
187 struct ptlrpc_reply_state *rs = req->rq_reply_state;
191 LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
193 idx = rs->rs_nlocks++;
194 rs->rs_locks[idx] = *lock;
195 rs->rs_modes[idx] = mode;
196 rs->rs_difficult = 1;
197 rs->rs_no_ack = no_ack;
198 rs->rs_convert_lock = convert_lock;
200 EXPORT_SYMBOL(ptlrpc_save_lock);
203 struct ptlrpc_hr_partition;
205 struct ptlrpc_hr_thread {
206 int hrt_id; /* thread ID */
208 wait_queue_head_t hrt_waitq;
209 struct list_head hrt_queue;
210 struct ptlrpc_hr_partition *hrt_partition;
213 struct ptlrpc_hr_partition {
214 /* # of started threads */
215 atomic_t hrp_nstarted;
216 /* # of stopped threads */
217 atomic_t hrp_nstopped;
218 /* cpu partition id */
220 /* round-robin rotor for choosing thread */
222 /* total number of threads on this partition */
225 struct ptlrpc_hr_thread *hrp_thrs;
228 #define HRT_RUNNING 0
229 #define HRT_STOPPING 1
231 struct ptlrpc_hr_service {
232 /* CPU partition table, it's just cfs_cpt_tab for now */
233 struct cfs_cpt_table *hr_cpt_table;
234 /** controller sleep waitq */
235 wait_queue_head_t hr_waitq;
236 unsigned int hr_stopping;
237 /** roundrobin rotor for non-affinity service */
238 unsigned int hr_rotor;
240 struct ptlrpc_hr_partition **hr_partitions;
244 struct list_head rsb_replies;
245 unsigned int rsb_n_replies;
246 struct ptlrpc_service_part *rsb_svcpt;
249 /** reply handling service. */
250 static struct ptlrpc_hr_service ptlrpc_hr;
253 * maximum mumber of replies scheduled in one batch
255 #define MAX_SCHEDULED 256
258 * Initialize a reply batch.
262 static void rs_batch_init(struct rs_batch *b)
264 memset(b, 0, sizeof(*b));
265 INIT_LIST_HEAD(&b->rsb_replies);
269 * Choose an hr thread to dispatch requests to.
272 struct ptlrpc_hr_thread *ptlrpc_hr_select(struct ptlrpc_service_part *svcpt)
274 struct ptlrpc_hr_partition *hrp;
277 if (svcpt->scp_cpt >= 0 &&
278 svcpt->scp_service->srv_cptable == ptlrpc_hr.hr_cpt_table) {
279 /* directly match partition */
280 hrp = ptlrpc_hr.hr_partitions[svcpt->scp_cpt];
283 rotor = ptlrpc_hr.hr_rotor++;
284 rotor %= cfs_cpt_number(ptlrpc_hr.hr_cpt_table);
286 hrp = ptlrpc_hr.hr_partitions[rotor];
289 rotor = hrp->hrp_rotor++;
290 return &hrp->hrp_thrs[rotor % hrp->hrp_nthrs];
294 * Dispatch all replies accumulated in the batch to one from
295 * dedicated reply handling threads.
299 static void rs_batch_dispatch(struct rs_batch *b)
301 if (b->rsb_n_replies != 0) {
302 struct ptlrpc_hr_thread *hrt;
304 hrt = ptlrpc_hr_select(b->rsb_svcpt);
306 spin_lock(&hrt->hrt_lock);
307 list_splice_init(&b->rsb_replies, &hrt->hrt_queue);
308 spin_unlock(&hrt->hrt_lock);
310 wake_up(&hrt->hrt_waitq);
311 b->rsb_n_replies = 0;
316 * Add a reply to a batch.
317 * Add one reply object to a batch, schedule batched replies if overload.
322 static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs)
324 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
326 if (svcpt != b->rsb_svcpt || b->rsb_n_replies >= MAX_SCHEDULED) {
327 if (b->rsb_svcpt != NULL) {
328 rs_batch_dispatch(b);
329 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
331 spin_lock(&svcpt->scp_rep_lock);
332 b->rsb_svcpt = svcpt;
334 spin_lock(&rs->rs_lock);
335 rs->rs_scheduled_ever = 1;
336 if (rs->rs_scheduled == 0) {
337 list_move(&rs->rs_list, &b->rsb_replies);
338 rs->rs_scheduled = 1;
341 rs->rs_committed = 1;
342 spin_unlock(&rs->rs_lock);
346 * Reply batch finalization.
347 * Dispatch remaining replies from the batch
348 * and release remaining spinlock.
352 static void rs_batch_fini(struct rs_batch *b)
354 if (b->rsb_svcpt != NULL) {
355 rs_batch_dispatch(b);
356 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
360 #define DECLARE_RS_BATCH(b) struct rs_batch b
364 * Put reply state into a queue for processing because we received
365 * ACK from the client
367 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
369 struct ptlrpc_hr_thread *hrt;
373 LASSERT(list_empty(&rs->rs_list));
375 hrt = ptlrpc_hr_select(rs->rs_svcpt);
377 spin_lock(&hrt->hrt_lock);
378 list_add_tail(&rs->rs_list, &hrt->hrt_queue);
379 spin_unlock(&hrt->hrt_lock);
381 wake_up(&hrt->hrt_waitq);
385 void ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs)
389 assert_spin_locked(&rs->rs_svcpt->scp_rep_lock);
390 assert_spin_locked(&rs->rs_lock);
391 LASSERT(rs->rs_difficult);
392 rs->rs_scheduled_ever = 1; /* flag any notification attempt */
394 if (rs->rs_scheduled) { /* being set up or already notified */
399 rs->rs_scheduled = 1;
400 list_del_init(&rs->rs_list);
401 ptlrpc_dispatch_difficult_reply(rs);
404 EXPORT_SYMBOL(ptlrpc_schedule_difficult_reply);
406 void ptlrpc_commit_replies(struct obd_export *exp)
408 struct ptlrpc_reply_state *rs, *nxt;
409 DECLARE_RS_BATCH(batch);
413 rs_batch_init(&batch);
415 * Find any replies that have been committed and get their service
416 * to attend to complete them.
419 /* CAVEAT EMPTOR: spinlock ordering!!! */
420 spin_lock(&exp->exp_uncommitted_replies_lock);
421 list_for_each_entry_safe(rs, nxt, &exp->exp_uncommitted_replies,
423 LASSERT(rs->rs_difficult);
424 /* VBR: per-export last_committed */
425 LASSERT(rs->rs_export);
426 if (rs->rs_transno <= exp->exp_last_committed) {
427 list_del_init(&rs->rs_obd_list);
428 rs_batch_add(&batch, rs);
431 spin_unlock(&exp->exp_uncommitted_replies_lock);
432 rs_batch_fini(&batch);
436 static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt)
438 struct ptlrpc_request_buffer_desc *rqbd;
443 spin_lock(&svcpt->scp_lock);
445 if (list_empty(&svcpt->scp_rqbd_idle)) {
446 spin_unlock(&svcpt->scp_lock);
450 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
451 struct ptlrpc_request_buffer_desc,
454 /* assume we will post successfully */
455 svcpt->scp_nrqbds_posted++;
456 list_move(&rqbd->rqbd_list, &svcpt->scp_rqbd_posted);
458 spin_unlock(&svcpt->scp_lock);
460 rc = ptlrpc_register_rqbd(rqbd);
467 spin_lock(&svcpt->scp_lock);
469 svcpt->scp_nrqbds_posted--;
470 list_move_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
473 * Don't complain if no request buffers are posted right now; LNET
474 * won't drop requests because we set the portal lazy!
477 spin_unlock(&svcpt->scp_lock);
482 static void ptlrpc_at_timer(cfs_timer_cb_arg_t data)
484 struct ptlrpc_service_part *svcpt;
486 svcpt = cfs_from_timer(svcpt, data, scp_at_timer);
488 svcpt->scp_at_check = 1;
489 svcpt->scp_at_checktime = ktime_get();
490 wake_up(&svcpt->scp_waitq);
493 static void ptlrpc_server_nthreads_check(struct ptlrpc_service *svc,
494 struct ptlrpc_service_conf *conf)
496 struct ptlrpc_service_thr_conf *tc = &conf->psc_thr;
503 * Common code for estimating & validating threads number.
504 * CPT affinity service could have percpt thread-pool instead
505 * of a global thread-pool, which means user might not always
506 * get the threads number they give it in conf::tc_nthrs_user
507 * even they did set. It's because we need to validate threads
508 * number for each CPT to guarantee each pool will have enough
509 * threads to keep the service healthy.
511 init = PTLRPC_NTHRS_INIT + (svc->srv_ops.so_hpreq_handler != NULL);
512 init = max_t(int, init, tc->tc_nthrs_init);
515 * NB: please see comments in lustre_lnet.h for definition
516 * details of these members
518 LASSERT(tc->tc_nthrs_max != 0);
520 if (tc->tc_nthrs_user != 0) {
522 * In case there is a reason to test a service with many
523 * threads, we give a less strict check here, it can
524 * be up to 8 * nthrs_max
526 total = min(tc->tc_nthrs_max * 8, tc->tc_nthrs_user);
527 nthrs = total / svc->srv_ncpts;
528 init = max(init, nthrs);
532 total = tc->tc_nthrs_max;
533 if (tc->tc_nthrs_base == 0) {
535 * don't care about base threads number per partition,
536 * this is most for non-affinity service
538 nthrs = total / svc->srv_ncpts;
542 nthrs = tc->tc_nthrs_base;
543 if (svc->srv_ncpts == 1) {
547 * NB: Increase the base number if it's single partition
548 * and total number of cores/HTs is larger or equal to 4.
549 * result will always < 2 * nthrs_base
551 weight = cfs_cpt_weight(svc->srv_cptable, CFS_CPT_ANY);
552 for (i = 1; (weight >> (i + 1)) != 0 && /* >= 4 cores/HTs */
553 (tc->tc_nthrs_base >> i) != 0; i++)
554 nthrs += tc->tc_nthrs_base >> i;
557 if (tc->tc_thr_factor != 0) {
558 int factor = tc->tc_thr_factor;
562 * User wants to increase number of threads with for
563 * each CPU core/HT, most likely the factor is larger than
564 * one thread/core because service threads are supposed to
565 * be blocked by lock or wait for IO.
568 * Amdahl's law says that adding processors wouldn't give
569 * a linear increasing of parallelism, so it's nonsense to
570 * have too many threads no matter how many cores/HTs
575 (topology_sibling_cpumask(smp_processor_id())) > 1) {
576 /* weight is # of HTs */
577 /* depress thread factor for hyper-thread */
578 factor = factor - (factor >> 1) + (factor >> 3);
582 weight = cfs_cpt_weight(svc->srv_cptable, 0);
584 for (; factor > 0 && weight > 0; factor--, weight -= fade)
585 nthrs += min(weight, fade) * factor;
588 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
589 nthrs = max(tc->tc_nthrs_base,
590 tc->tc_nthrs_max / svc->srv_ncpts);
593 nthrs = max(nthrs, tc->tc_nthrs_init);
594 svc->srv_nthrs_cpt_limit = nthrs;
595 svc->srv_nthrs_cpt_init = init;
597 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
599 "%s: This service may have more threads (%d) than the given soft limit (%d)\n",
600 svc->srv_name, nthrs * svc->srv_ncpts,
606 * Initialize percpt data for a service
608 static int ptlrpc_service_part_init(struct ptlrpc_service *svc,
609 struct ptlrpc_service_part *svcpt, int cpt)
611 struct ptlrpc_at_array *array;
616 svcpt->scp_cpt = cpt;
617 INIT_LIST_HEAD(&svcpt->scp_threads);
619 /* rqbd and incoming request queue */
620 spin_lock_init(&svcpt->scp_lock);
621 mutex_init(&svcpt->scp_mutex);
622 INIT_LIST_HEAD(&svcpt->scp_rqbd_idle);
623 INIT_LIST_HEAD(&svcpt->scp_rqbd_posted);
624 INIT_LIST_HEAD(&svcpt->scp_req_incoming);
625 init_waitqueue_head(&svcpt->scp_waitq);
626 /* history request & rqbd list */
627 INIT_LIST_HEAD(&svcpt->scp_hist_reqs);
628 INIT_LIST_HEAD(&svcpt->scp_hist_rqbds);
630 /* acitve requests and hp requests */
631 spin_lock_init(&svcpt->scp_req_lock);
634 spin_lock_init(&svcpt->scp_rep_lock);
635 INIT_LIST_HEAD(&svcpt->scp_rep_active);
636 INIT_LIST_HEAD(&svcpt->scp_rep_idle);
637 init_waitqueue_head(&svcpt->scp_rep_waitq);
638 atomic_set(&svcpt->scp_nreps_difficult, 0);
640 /* adaptive timeout */
641 spin_lock_init(&svcpt->scp_at_lock);
642 array = &svcpt->scp_at_array;
644 size = at_est2timeout(at_max);
645 array->paa_size = size;
646 array->paa_count = 0;
647 array->paa_deadline = -1;
649 /* allocate memory for scp_at_array (ptlrpc_at_array) */
650 OBD_CPT_ALLOC(array->paa_reqs_array,
651 svc->srv_cptable, cpt, sizeof(struct list_head) * size);
652 if (array->paa_reqs_array == NULL)
655 for (index = 0; index < size; index++)
656 INIT_LIST_HEAD(&array->paa_reqs_array[index]);
658 OBD_CPT_ALLOC(array->paa_reqs_count,
659 svc->srv_cptable, cpt, sizeof(__u32) * size);
660 if (array->paa_reqs_count == NULL)
663 cfs_timer_setup(&svcpt->scp_at_timer, ptlrpc_at_timer,
664 (unsigned long)svcpt, 0);
667 * At SOW, service time should be quick; 10s seems generous. If client
668 * timeout is less than this, we'll be sending an early reply.
670 at_init(&svcpt->scp_at_estimate, 10, 0);
672 /* assign this before call ptlrpc_grow_req_bufs */
673 svcpt->scp_service = svc;
674 /* Now allocate the request buffers, but don't post them now */
675 rc = ptlrpc_grow_req_bufs(svcpt, 0);
677 * We shouldn't be under memory pressure at startup, so
678 * fail if we can't allocate all our buffers at this time.
686 if (array->paa_reqs_count != NULL) {
687 OBD_FREE_PTR_ARRAY(array->paa_reqs_count, size);
688 array->paa_reqs_count = NULL;
691 if (array->paa_reqs_array != NULL) {
692 OBD_FREE_PTR_ARRAY(array->paa_reqs_array, array->paa_size);
693 array->paa_reqs_array = NULL;
700 * Initialize service on a given portal.
701 * This includes starting serving threads , allocating and posting rqbds and
704 struct ptlrpc_service *ptlrpc_register_service(struct ptlrpc_service_conf *conf,
706 struct dentry *debugfs_entry)
708 struct ptlrpc_service_cpt_conf *cconf = &conf->psc_cpt;
709 struct ptlrpc_service *service;
710 struct ptlrpc_service_part *svcpt;
711 struct cfs_cpt_table *cptable;
720 LASSERT(conf->psc_buf.bc_nbufs > 0);
721 LASSERT(conf->psc_buf.bc_buf_size >=
722 conf->psc_buf.bc_req_max_size + SPTLRPC_MAX_PAYLOAD);
723 LASSERT(conf->psc_thr.tc_ctx_tags != 0);
725 cptable = cconf->cc_cptable;
727 cptable = cfs_cpt_tab;
729 if (conf->psc_thr.tc_cpu_bind > 1) {
730 CERROR("%s: Invalid cpu bind value %d, only 1 or 0 allowed\n",
731 conf->psc_name, conf->psc_thr.tc_cpu_bind);
732 RETURN(ERR_PTR(-EINVAL));
735 if (!cconf->cc_affinity) {
738 ncpts = cfs_cpt_number(cptable);
739 if (cconf->cc_pattern != NULL) {
740 struct cfs_expr_list *el;
742 rc = cfs_expr_list_parse(cconf->cc_pattern,
743 strlen(cconf->cc_pattern),
746 CERROR("%s: invalid CPT pattern string: %s\n",
747 conf->psc_name, cconf->cc_pattern);
748 RETURN(ERR_PTR(-EINVAL));
751 rc = cfs_expr_list_values(el, ncpts, &cpts);
752 cfs_expr_list_free(el);
754 CERROR("%s: failed to parse CPT array %s: %d\n",
755 conf->psc_name, cconf->cc_pattern, rc);
757 OBD_FREE_PTR_ARRAY(cpts, ncpts);
758 RETURN(ERR_PTR(rc < 0 ? rc : -EINVAL));
764 OBD_ALLOC(service, offsetof(struct ptlrpc_service, srv_parts[ncpts]));
765 if (service == NULL) {
767 OBD_FREE_PTR_ARRAY(cpts, ncpts);
768 RETURN(ERR_PTR(-ENOMEM));
771 service->srv_cptable = cptable;
772 service->srv_cpts = cpts;
773 service->srv_ncpts = ncpts;
774 service->srv_cpt_bind = conf->psc_thr.tc_cpu_bind;
776 service->srv_cpt_bits = 0; /* it's zero already, easy to read... */
777 while ((1 << service->srv_cpt_bits) < cfs_cpt_number(cptable))
778 service->srv_cpt_bits++;
781 spin_lock_init(&service->srv_lock);
782 service->srv_name = conf->psc_name;
783 service->srv_watchdog_factor = conf->psc_watchdog_factor;
784 INIT_LIST_HEAD(&service->srv_list); /* for safty of cleanup */
786 /* buffer configuration */
787 service->srv_nbuf_per_group = test_req_buffer_pressure ?
788 1 : conf->psc_buf.bc_nbufs;
789 /* do not limit max number of rqbds by default */
790 service->srv_nrqbds_max = 0;
792 service->srv_max_req_size = conf->psc_buf.bc_req_max_size +
794 service->srv_buf_size = conf->psc_buf.bc_buf_size;
795 service->srv_rep_portal = conf->psc_buf.bc_rep_portal;
796 service->srv_req_portal = conf->psc_buf.bc_req_portal;
798 /* With slab/alloc_pages buffer size will be rounded up to 2^n */
799 if (service->srv_buf_size & (service->srv_buf_size - 1)) {
800 int round = size_roundup_power2(service->srv_buf_size);
802 service->srv_buf_size = round;
805 /* Increase max reply size to next power of two */
806 service->srv_max_reply_size = 1;
807 while (service->srv_max_reply_size <
808 conf->psc_buf.bc_rep_max_size + SPTLRPC_MAX_PAYLOAD)
809 service->srv_max_reply_size <<= 1;
811 service->srv_thread_name = conf->psc_thr.tc_thr_name;
812 service->srv_ctx_tags = conf->psc_thr.tc_ctx_tags;
813 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
814 service->srv_ops = conf->psc_ops;
816 for (i = 0; i < ncpts; i++) {
817 if (!cconf->cc_affinity)
820 cpt = cpts != NULL ? cpts[i] : i;
822 OBD_CPT_ALLOC(svcpt, cptable, cpt, sizeof(*svcpt));
824 GOTO(failed, rc = -ENOMEM);
826 service->srv_parts[i] = svcpt;
827 rc = ptlrpc_service_part_init(service, svcpt, cpt);
832 ptlrpc_server_nthreads_check(service, conf);
834 rc = LNetSetLazyPortal(service->srv_req_portal);
837 mutex_lock(&ptlrpc_all_services_mutex);
838 list_add(&service->srv_list, &ptlrpc_all_services);
839 mutex_unlock(&ptlrpc_all_services_mutex);
842 rc = ptlrpc_sysfs_register_service(parent, service);
847 if (debugfs_entry != NULL)
848 ptlrpc_ldebugfs_register_service(debugfs_entry, service);
850 rc = ptlrpc_service_nrs_setup(service);
854 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
855 service->srv_name, service->srv_req_portal);
857 rc = ptlrpc_start_threads(service);
859 CERROR("Failed to start threads for service %s: %d\n",
860 service->srv_name, rc);
866 ptlrpc_unregister_service(service);
869 EXPORT_SYMBOL(ptlrpc_register_service);
872 * to actually free the request, must be called without holding svc_lock.
873 * note it's caller's responsibility to unlink req->rq_list.
875 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
877 LASSERT(atomic_read(&req->rq_refcount) == 0);
878 LASSERT(list_empty(&req->rq_timed_list));
881 * DEBUG_REQ() assumes the reply state of a request with a valid
882 * ref will not be destroyed until that reference is dropped.
884 ptlrpc_req_drop_rs(req);
886 sptlrpc_svc_ctx_decref(req);
888 if (req != &req->rq_rqbd->rqbd_req) {
890 * NB request buffers use an embedded
891 * req if the incoming req unlinked the
892 * MD; this isn't one of them!
894 ptlrpc_request_cache_free(req);
899 * drop a reference count of the request. if it reaches 0, we either
900 * put it into history list, or free it immediately.
902 void ptlrpc_server_drop_request(struct ptlrpc_request *req)
904 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
905 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
906 struct ptlrpc_service *svc = svcpt->scp_service;
909 if (!atomic_dec_and_test(&req->rq_refcount))
912 if (req->rq_session.lc_state == LCS_ENTERED) {
913 lu_context_exit(&req->rq_session);
914 lu_context_fini(&req->rq_session);
917 if (req->rq_at_linked) {
918 spin_lock(&svcpt->scp_at_lock);
920 * recheck with lock, in case it's unlinked by
921 * ptlrpc_at_check_timed()
923 if (likely(req->rq_at_linked))
924 ptlrpc_at_remove_timed(req);
925 spin_unlock(&svcpt->scp_at_lock);
928 LASSERT(list_empty(&req->rq_timed_list));
930 /* finalize request */
931 if (req->rq_export) {
932 class_export_put(req->rq_export);
933 req->rq_export = NULL;
936 spin_lock(&svcpt->scp_lock);
938 list_add(&req->rq_list, &rqbd->rqbd_reqs);
940 refcount = --(rqbd->rqbd_refcount);
942 /* request buffer is now idle: add to history */
943 list_move_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds);
944 svcpt->scp_hist_nrqbds++;
948 * I expect only about 1 or 2 rqbds need to be recycled here
950 while (svcpt->scp_hist_nrqbds > svc->srv_hist_nrqbds_cpt_max) {
951 rqbd = list_entry(svcpt->scp_hist_rqbds.next,
952 struct ptlrpc_request_buffer_desc,
955 list_del(&rqbd->rqbd_list);
956 svcpt->scp_hist_nrqbds--;
959 * remove rqbd's reqs from svc's req history while
960 * I've got the service lock
962 list_for_each_entry(req, &rqbd->rqbd_reqs, rq_list) {
963 /* Track the highest culled req seq */
964 if (req->rq_history_seq >
965 svcpt->scp_hist_seq_culled) {
966 svcpt->scp_hist_seq_culled =
969 list_del(&req->rq_history_list);
972 spin_unlock(&svcpt->scp_lock);
974 while ((req = list_first_entry_or_null(
976 struct ptlrpc_request, rq_list))) {
977 list_del(&req->rq_list);
978 ptlrpc_server_free_request(req);
981 spin_lock(&svcpt->scp_lock);
983 * now all reqs including the embedded req has been
984 * disposed, schedule request buffer for re-use
985 * or free it to drain some in excess.
987 LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) == 0);
988 if (svcpt->scp_nrqbds_posted >=
989 svc->srv_nbuf_per_group ||
990 (svc->srv_nrqbds_max != 0 &&
991 svcpt->scp_nrqbds_total > svc->srv_nrqbds_max) ||
992 test_req_buffer_pressure) {
993 /* like in ptlrpc_free_rqbd() */
994 svcpt->scp_nrqbds_total--;
995 OBD_FREE_LARGE(rqbd->rqbd_buffer,
999 list_add_tail(&rqbd->rqbd_list,
1000 &svcpt->scp_rqbd_idle);
1004 spin_unlock(&svcpt->scp_lock);
1005 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
1006 /* If we are low on memory, we are not interested in history */
1007 list_del(&req->rq_list);
1008 list_del_init(&req->rq_history_list);
1010 /* Track the highest culled req seq */
1011 if (req->rq_history_seq > svcpt->scp_hist_seq_culled)
1012 svcpt->scp_hist_seq_culled = req->rq_history_seq;
1014 spin_unlock(&svcpt->scp_lock);
1016 ptlrpc_server_free_request(req);
1018 spin_unlock(&svcpt->scp_lock);
1022 static void ptlrpc_add_exp_list_nolock(struct ptlrpc_request *req,
1023 struct obd_export *export, bool hp)
1025 __u16 tag = lustre_msg_get_tag(req->rq_reqmsg);
1028 list_add(&req->rq_exp_list, &export->exp_hp_rpcs);
1030 list_add(&req->rq_exp_list, &export->exp_reg_rpcs);
1031 if (tag && export->exp_used_slots)
1032 set_bit(tag - 1, export->exp_used_slots);
1035 static void ptlrpc_del_exp_list(struct ptlrpc_request *req)
1037 __u16 tag = lustre_msg_get_tag(req->rq_reqmsg);
1039 spin_lock(&req->rq_export->exp_rpc_lock);
1040 list_del_init(&req->rq_exp_list);
1041 if (tag && !req->rq_obsolete && req->rq_export->exp_used_slots)
1042 clear_bit(tag - 1, req->rq_export->exp_used_slots);
1043 spin_unlock(&req->rq_export->exp_rpc_lock);
1046 /** Change request export and move hp request from old export to new */
1047 void ptlrpc_request_change_export(struct ptlrpc_request *req,
1048 struct obd_export *export)
1050 if (req->rq_export != NULL) {
1051 LASSERT(!list_empty(&req->rq_exp_list));
1052 /* remove rq_exp_list from last export */
1053 ptlrpc_del_exp_list(req);
1054 /* export has one reference already, so it's safe to
1055 * add req to export queue here and get another
1056 * reference for request later
1058 spin_lock(&export->exp_rpc_lock);
1059 ptlrpc_add_exp_list_nolock(req, export, req->rq_ops != NULL);
1060 spin_unlock(&export->exp_rpc_lock);
1062 class_export_rpc_dec(req->rq_export);
1063 class_export_put(req->rq_export);
1066 /* request takes one export refcount */
1067 req->rq_export = class_export_get(export);
1068 class_export_rpc_inc(export);
1072 * to finish a request: stop sending more early replies, and release
1075 static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt,
1076 struct ptlrpc_request *req)
1078 ptlrpc_server_hpreq_fini(req);
1080 ptlrpc_server_drop_request(req);
1084 * to finish an active request: stop sending more early replies, and release
1085 * the request. should be called after we finished handling the request.
1087 static void ptlrpc_server_finish_active_request(
1088 struct ptlrpc_service_part *svcpt,
1089 struct ptlrpc_request *req)
1091 spin_lock(&svcpt->scp_req_lock);
1092 ptlrpc_nrs_req_stop_nolock(req);
1093 svcpt->scp_nreqs_active--;
1095 svcpt->scp_nhreqs_active--;
1096 spin_unlock(&svcpt->scp_req_lock);
1098 ptlrpc_nrs_req_finalize(req);
1100 if (req->rq_export != NULL)
1101 class_export_rpc_dec(req->rq_export);
1103 ptlrpc_server_finish_request(svcpt, req);
1107 * This function makes sure dead exports are evicted in a timely manner.
1108 * This function is only called when some export receives a message (i.e.,
1109 * the network is up.)
1111 void ptlrpc_update_export_timer(struct obd_export *exp, time64_t extra_delay)
1113 struct obd_export *oldest_exp;
1114 time64_t oldest_time, new_time;
1121 * Compensate for slow machines, etc, by faking our request time
1122 * into the future. Although this can break the strict time-ordering
1123 * of the list, we can be really lazy here - we don't have to evict
1124 * at the exact right moment. Eventually, all silent exports
1125 * will make it to the top of the list.
1128 /* Do not pay attention on 1sec or smaller renewals. */
1129 new_time = ktime_get_real_seconds() + extra_delay;
1130 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
1133 exp->exp_last_request_time = new_time;
1136 * exports may get disconnected from the chain even though the
1137 * export has references, so we must keep the spin lock while
1138 * manipulating the lists
1140 spin_lock(&exp->exp_obd->obd_dev_lock);
1142 if (list_empty(&exp->exp_obd_chain_timed)) {
1143 /* this one is not timed */
1144 spin_unlock(&exp->exp_obd->obd_dev_lock);
1148 list_move_tail(&exp->exp_obd_chain_timed,
1149 &exp->exp_obd->obd_exports_timed);
1151 oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
1152 struct obd_export, exp_obd_chain_timed);
1153 oldest_time = oldest_exp->exp_last_request_time;
1154 spin_unlock(&exp->exp_obd->obd_dev_lock);
1156 if (exp->exp_obd->obd_recovering) {
1157 /* be nice to everyone during recovery */
1162 /* Note - racing to start/reset the obd_eviction timer is safe */
1163 if (exp->exp_obd->obd_eviction_timer == 0) {
1164 /* Check if the oldest entry is expired. */
1165 if (ktime_get_real_seconds() >
1166 oldest_time + PING_EVICT_TIMEOUT + extra_delay) {
1168 * We need a second timer, in case the net was down and
1169 * it just came back. Since the pinger may skip every
1170 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
1171 * we better wait for 3.
1173 exp->exp_obd->obd_eviction_timer =
1174 ktime_get_real_seconds() + 3 * PING_INTERVAL;
1175 CDEBUG(D_HA, "%s: Think about evicting %s from %lld\n",
1176 exp->exp_obd->obd_name,
1177 obd_export_nid2str(oldest_exp), oldest_time);
1180 if (ktime_get_real_seconds() >
1181 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
1183 * The evictor won't evict anyone who we've heard from
1184 * recently, so we don't have to check before we start
1187 if (!ping_evictor_wake(exp))
1188 exp->exp_obd->obd_eviction_timer = 0;
1196 * Sanity check request \a req.
1197 * Return 0 if all is ok, error code otherwise.
1199 static int ptlrpc_check_req(struct ptlrpc_request *req)
1201 struct obd_device *obd = req->rq_export->exp_obd;
1204 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
1205 req->rq_export->exp_conn_cnt)) {
1206 DEBUG_REQ(D_RPCTRACE, req,
1207 "DROPPING req from old connection %d < %d",
1208 lustre_msg_get_conn_cnt(req->rq_reqmsg),
1209 req->rq_export->exp_conn_cnt);
1212 if (unlikely(obd == NULL || obd->obd_fail)) {
1214 * Failing over, don't handle any more reqs,
1215 * send error response instead.
1217 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
1218 req, (obd != NULL) ? obd->obd_name : "unknown");
1220 } else if (lustre_msg_get_flags(req->rq_reqmsg) &
1221 (MSG_REPLAY | MSG_REQ_REPLAY_DONE) &&
1222 !obd->obd_recovering) {
1223 DEBUG_REQ(D_ERROR, req,
1224 "Invalid replay without recovery");
1225 class_fail_export(req->rq_export);
1227 } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 &&
1228 !obd->obd_recovering) {
1229 DEBUG_REQ(D_ERROR, req,
1230 "Invalid req with transno %llu without recovery",
1231 lustre_msg_get_transno(req->rq_reqmsg));
1232 class_fail_export(req->rq_export);
1236 if (unlikely(rc < 0)) {
1237 req->rq_status = rc;
1243 static void ptlrpc_at_set_timer(struct ptlrpc_service_part *svcpt)
1245 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1248 if (array->paa_count == 0) {
1249 del_timer(&svcpt->scp_at_timer);
1253 /* Set timer for closest deadline */
1254 next = array->paa_deadline - ktime_get_real_seconds() -
1257 ptlrpc_at_timer(cfs_timer_cb_arg(svcpt, scp_at_timer));
1259 mod_timer(&svcpt->scp_at_timer,
1260 jiffies + nsecs_to_jiffies(next * NSEC_PER_SEC));
1261 CDEBUG(D_INFO, "armed %s at %+llds\n",
1262 svcpt->scp_service->srv_name, next);
1266 /* Add rpc to early reply check list */
1267 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
1269 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1270 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1271 struct ptlrpc_request *rq = NULL;
1277 if (req->rq_no_reply)
1280 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
1283 spin_lock(&svcpt->scp_at_lock);
1284 LASSERT(list_empty(&req->rq_timed_list));
1286 div_u64_rem(req->rq_deadline, array->paa_size, &index);
1287 if (array->paa_reqs_count[index] > 0) {
1289 * latest rpcs will have the latest deadlines in the list,
1290 * so search backward.
1292 list_for_each_entry_reverse(rq, &array->paa_reqs_array[index],
1294 if (req->rq_deadline >= rq->rq_deadline) {
1295 list_add(&req->rq_timed_list,
1296 &rq->rq_timed_list);
1302 /* Add the request at the head of the list */
1303 if (list_empty(&req->rq_timed_list))
1304 list_add(&req->rq_timed_list, &array->paa_reqs_array[index]);
1306 spin_lock(&req->rq_lock);
1307 req->rq_at_linked = 1;
1308 spin_unlock(&req->rq_lock);
1309 req->rq_at_index = index;
1310 array->paa_reqs_count[index]++;
1312 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
1313 array->paa_deadline = req->rq_deadline;
1314 ptlrpc_at_set_timer(svcpt);
1316 spin_unlock(&svcpt->scp_at_lock);
1321 static void ptlrpc_at_remove_timed(struct ptlrpc_request *req)
1323 struct ptlrpc_at_array *array;
1325 array = &req->rq_rqbd->rqbd_svcpt->scp_at_array;
1327 /* NB: must call with hold svcpt::scp_at_lock */
1328 LASSERT(!list_empty(&req->rq_timed_list));
1329 list_del_init(&req->rq_timed_list);
1331 spin_lock(&req->rq_lock);
1332 req->rq_at_linked = 0;
1333 spin_unlock(&req->rq_lock);
1335 array->paa_reqs_count[req->rq_at_index]--;
1340 * Attempt to extend the request deadline by sending an early reply to the
1343 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
1345 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1346 struct ptlrpc_request *reqcopy;
1347 struct lustre_msg *reqmsg;
1348 timeout_t olddl = req->rq_deadline - ktime_get_real_seconds();
1354 if (CFS_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_RECONNECT)) {
1355 /* don't send early reply */
1360 * deadline is when the client expects us to reply, margin is the
1361 * difference between clients' and servers' expectations
1363 DEBUG_REQ(D_ADAPTTO, req,
1364 "%ssending early reply (deadline %+ds, margin %+ds) for %d+%d",
1365 AT_OFF ? "AT off - not " : "",
1366 olddl, olddl - at_get(&svcpt->scp_at_estimate),
1367 at_get(&svcpt->scp_at_estimate), at_extra);
1373 /* below message is checked in replay-ost-single.sh test_9 */
1374 DEBUG_REQ(D_WARNING, req,
1375 "Already past deadline (%+ds), not sending early reply. Consider increasing at_early_margin (%d)?",
1376 olddl, at_early_margin);
1378 /* Return an error so we're not re-added to the timed list. */
1382 if ((lustre_msghdr_get_flags(req->rq_reqmsg) &
1383 MSGHDR_AT_SUPPORT) == 0) {
1384 DEBUG_REQ(D_INFO, req,
1385 "Wanted to ask client for more time, but no AT support");
1389 if (req->rq_export &&
1390 lustre_msg_get_flags(req->rq_reqmsg) &
1391 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1392 struct obd_device *obd_exp = req->rq_export->exp_obd;
1395 * During recovery, we don't want to send too many early
1396 * replies, but on the other hand we want to make sure the
1397 * client has enough time to resend if the rpc is lost. So
1398 * during the recovery period send at least 4 early replies,
1399 * spacing them every at_extra if we can. at_estimate should
1400 * always equal this fixed value during recovery.
1404 * Don't account request processing time into AT history
1405 * during recovery, it is not service time we need but
1406 * includes also waiting time for recovering clients
1408 newdl = min_t(time64_t, at_extra,
1409 obd_exp->obd_recovery_timeout / 4) +
1410 ktime_get_real_seconds();
1413 * We want to extend the request deadline by at_extra seconds,
1414 * so we set our service estimate to reflect how much time has
1415 * passed since this request arrived plus an additional
1416 * at_extra seconds. The client will calculate the new deadline
1417 * based on this service estimate (plus some additional time to
1418 * account for network latency). See ptlrpc_at_recv_early_reply
1420 at_measured(&svcpt->scp_at_estimate, at_extra +
1421 ktime_get_real_seconds() -
1422 req->rq_arrival_time.tv_sec);
1423 newdl = req->rq_arrival_time.tv_sec +
1424 at_get(&svcpt->scp_at_estimate);
1428 * Check to see if we've actually increased the deadline -
1429 * we may be past adaptive_max
1431 if (req->rq_deadline >= newdl) {
1432 DEBUG_REQ(D_WARNING, req,
1433 "Could not add any time (%d/%lld), not sending early reply",
1434 olddl, newdl - ktime_get_real_seconds());
1438 reqcopy = ptlrpc_request_cache_alloc(GFP_NOFS);
1439 if (reqcopy == NULL)
1441 OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1443 GOTO(out_free, rc = -ENOMEM);
1446 reqcopy->rq_reply_state = NULL;
1447 reqcopy->rq_rep_swab_mask = 0;
1448 reqcopy->rq_pack_bulk = 0;
1449 reqcopy->rq_pack_udesc = 0;
1450 reqcopy->rq_packed_final = 0;
1451 sptlrpc_svc_ctx_addref(reqcopy);
1452 /* We only need the reqmsg for the magic */
1453 reqcopy->rq_reqmsg = reqmsg;
1454 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1457 * tgt_brw_read() and tgt_brw_write() may have decided not to reply.
1458 * Without this check, we would fail the rq_no_reply assertion in
1459 * ptlrpc_send_reply().
1461 if (reqcopy->rq_no_reply)
1462 GOTO(out, rc = -ETIMEDOUT);
1464 LASSERT(atomic_read(&req->rq_refcount));
1465 /* if it is last refcount then early reply isn't needed */
1466 if (atomic_read(&req->rq_refcount) == 1) {
1467 DEBUG_REQ(D_ADAPTTO, reqcopy,
1468 "Normal reply already sent, abort early reply");
1469 GOTO(out, rc = -EINVAL);
1472 /* Connection ref */
1473 reqcopy->rq_export = class_conn2export(
1474 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1475 if (reqcopy->rq_export == NULL)
1476 GOTO(out, rc = -ENODEV);
1479 class_export_rpc_inc(reqcopy->rq_export);
1480 if (reqcopy->rq_export->exp_obd &&
1481 reqcopy->rq_export->exp_obd->obd_fail)
1482 GOTO(out_put, rc = -ENODEV);
1484 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1488 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1491 /* Adjust our own deadline to what we told the client */
1492 req->rq_deadline = newdl;
1493 req->rq_early_count++; /* number sent, server side */
1495 DEBUG_REQ(D_ERROR, req, "Early reply send failed: rc = %d", rc);
1499 * Free the (early) reply state from lustre_pack_reply.
1500 * (ptlrpc_send_reply takes it's own rs ref, so this is safe here)
1502 ptlrpc_req_drop_rs(reqcopy);
1505 class_export_rpc_dec(reqcopy->rq_export);
1506 class_export_put(reqcopy->rq_export);
1508 sptlrpc_svc_ctx_decref(reqcopy);
1509 OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1511 ptlrpc_request_cache_free(reqcopy);
1516 * Send early replies to everybody expiring within at_early_margin
1517 * asking for at_extra time
1519 static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1521 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1522 struct ptlrpc_request *rq, *n;
1523 LIST_HEAD(work_list);
1526 time64_t now = ktime_get_real_seconds();
1528 int first, counter = 0;
1531 spin_lock(&svcpt->scp_at_lock);
1532 if (svcpt->scp_at_check == 0) {
1533 spin_unlock(&svcpt->scp_at_lock);
1536 delay_ms = ktime_ms_delta(ktime_get(), svcpt->scp_at_checktime);
1537 svcpt->scp_at_check = 0;
1539 if (array->paa_count == 0) {
1540 spin_unlock(&svcpt->scp_at_lock);
1544 /* The timer went off, but maybe the nearest rpc already completed. */
1545 first = array->paa_deadline - now;
1546 if (first > at_early_margin) {
1547 /* We've still got plenty of time. Reset the timer. */
1548 ptlrpc_at_set_timer(svcpt);
1549 spin_unlock(&svcpt->scp_at_lock);
1554 * We're close to a timeout, and we don't know how much longer the
1555 * server will take. Send early replies to everyone expiring soon.
1558 div_u64_rem(array->paa_deadline, array->paa_size, &index);
1559 count = array->paa_count;
1561 count -= array->paa_reqs_count[index];
1562 list_for_each_entry_safe(rq, n,
1563 &array->paa_reqs_array[index],
1565 if (rq->rq_deadline > now + at_early_margin) {
1566 /* update the earliest deadline */
1567 if (deadline == -1 ||
1568 rq->rq_deadline < deadline)
1569 deadline = rq->rq_deadline;
1574 * ptlrpc_server_drop_request() may drop
1575 * refcount to 0 already. Let's check this and
1576 * don't add entry to work_list
1578 if (likely(atomic_inc_not_zero(&rq->rq_refcount))) {
1579 ptlrpc_at_remove_timed(rq);
1580 list_add(&rq->rq_timed_list, &work_list);
1582 ptlrpc_at_remove_timed(rq);
1588 if (++index >= array->paa_size)
1591 array->paa_deadline = deadline;
1592 /* we have a new earliest deadline, restart the timer */
1593 ptlrpc_at_set_timer(svcpt);
1595 spin_unlock(&svcpt->scp_at_lock);
1598 "timeout in %+ds, asking for %d secs on %d early replies\n",
1599 first, at_extra, counter);
1602 * We're already past request deadlines before we even get a
1603 * chance to send early replies
1605 LCONSOLE_WARN("%s: This server is not able to keep up with request traffic (cpu-bound).\n",
1606 svcpt->scp_service->srv_name);
1607 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, delay=%lldms\n",
1608 counter, svcpt->scp_nreqs_incoming,
1609 svcpt->scp_nreqs_active,
1610 at_get(&svcpt->scp_at_estimate), delay_ms);
1614 * we took additional refcount so entries can't be deleted from list, no
1617 while (!list_empty(&work_list)) {
1618 rq = list_entry(work_list.next, struct ptlrpc_request,
1620 list_del_init(&rq->rq_timed_list);
1622 if (ptlrpc_at_send_early_reply(rq) == 0)
1623 ptlrpc_at_add_timed(rq);
1625 ptlrpc_server_drop_request(rq);
1628 RETURN(1); /* return "did_something" for liblustre */
1632 * Check if we are already handling earlier incarnation of this request.
1633 * Called under &req->rq_export->exp_rpc_lock locked
1635 static struct ptlrpc_request*
1636 ptlrpc_server_check_resend_in_progress(struct ptlrpc_request *req)
1638 struct ptlrpc_request *tmp = NULL;
1640 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ||
1641 (atomic_read(&req->rq_export->exp_rpc_count) == 0))
1645 * bulk request are aborted upon reconnect, don't try to
1648 if (req->rq_bulk_write || req->rq_bulk_read)
1652 * This list should not be longer than max_requests in
1653 * flights on the client, so it is not all that long.
1654 * Also we only hit this codepath in case of a resent
1655 * request which makes it even more rarely hit
1657 list_for_each_entry(tmp, &req->rq_export->exp_reg_rpcs,
1659 /* Found duplicate one */
1660 if (tmp->rq_xid == req->rq_xid)
1663 list_for_each_entry(tmp, &req->rq_export->exp_hp_rpcs,
1665 /* Found duplicate one */
1666 if (tmp->rq_xid == req->rq_xid)
1672 DEBUG_REQ(D_HA, req, "Found duplicate req in processing");
1673 DEBUG_REQ(D_HA, tmp, "Request being processed");
1677 #ifdef HAVE_SERVER_SUPPORT
1678 static void ptlrpc_server_mark_obsolete(struct ptlrpc_request *req)
1680 req->rq_obsolete = 1;
1684 ptlrpc_server_mark_in_progress_obsolete(struct ptlrpc_request *req)
1686 struct ptlrpc_request *tmp = NULL;
1689 if (!tgt_is_increasing_xid_client(req->rq_export) ||
1690 req->rq_export->exp_used_slots == NULL)
1693 tag = lustre_msg_get_tag(req->rq_reqmsg);
1697 if (!test_bit(tag - 1, req->rq_export->exp_used_slots))
1700 /* This list should not be longer than max_requests in
1701 * flights on the client, so it is not all that long.
1702 * Also we only hit this codepath in case of a resent
1703 * request which makes it even more rarely hit */
1704 list_for_each_entry(tmp, &req->rq_export->exp_reg_rpcs, rq_exp_list) {
1705 if (tag == lustre_msg_get_tag(tmp->rq_reqmsg) &&
1706 req->rq_xid > tmp->rq_xid)
1707 ptlrpc_server_mark_obsolete(tmp);
1710 list_for_each_entry(tmp, &req->rq_export->exp_hp_rpcs, rq_exp_list) {
1711 if (tag == lustre_msg_get_tag(tmp->rq_reqmsg) &&
1712 req->rq_xid > tmp->rq_xid)
1713 ptlrpc_server_mark_obsolete(tmp);
1719 * Check if a request should be assigned with a high priority.
1721 * \retval < 0: error occurred
1722 * 0: normal RPC request
1723 * +1: high priority request
1725 static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt,
1726 struct ptlrpc_request *req)
1731 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL) {
1732 rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
1739 if (req->rq_export != NULL && req->rq_ops != NULL) {
1741 * Perform request specific check. We should do this
1742 * check before the request is added into exp_hp_rpcs
1743 * list otherwise it may hit swab race at LU-1044.
1745 if (req->rq_ops->hpreq_check != NULL) {
1746 rc = req->rq_ops->hpreq_check(req);
1747 if (rc == -ESTALE) {
1748 req->rq_status = rc;
1752 * can only return error,
1753 * 0 for normal request,
1754 * or 1 for high priority request
1763 /** Remove the request from the export list. */
1764 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req)
1767 if (req->rq_export) {
1769 * refresh lock timeout again so that client has more
1770 * room to send lock cancel RPC.
1772 if (req->rq_ops && req->rq_ops->hpreq_fini)
1773 req->rq_ops->hpreq_fini(req);
1775 ptlrpc_del_exp_list(req);
1780 static int ptlrpc_hpreq_check(struct ptlrpc_request *req)
1785 static struct ptlrpc_hpreq_ops ptlrpc_hpreq_common = {
1786 .hpreq_check = ptlrpc_hpreq_check,
1789 /* Hi-Priority RPC check by RPC operation code. */
1790 int ptlrpc_hpreq_handler(struct ptlrpc_request *req)
1792 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1795 * Check for export to let only reconnects for not yet evicted
1796 * export to become a HP rpc.
1798 if ((req->rq_export != NULL) &&
1799 (opc == OBD_PING || opc == MDS_CONNECT || opc == OST_CONNECT))
1800 req->rq_ops = &ptlrpc_hpreq_common;
1804 EXPORT_SYMBOL(ptlrpc_hpreq_handler);
1806 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1807 struct ptlrpc_request *req)
1811 struct ptlrpc_request *orig;
1815 rc = ptlrpc_server_hpreq_init(svcpt, req);
1820 ptlrpc_nrs_req_initialize(svcpt, req, hp);
1822 while (req->rq_export != NULL) {
1823 struct obd_export *exp = req->rq_export;
1826 * do search for duplicated xid and the adding to the list
1829 spin_lock_bh(&exp->exp_rpc_lock);
1830 #ifdef HAVE_SERVER_SUPPORT
1831 ptlrpc_server_mark_in_progress_obsolete(req);
1833 orig = ptlrpc_server_check_resend_in_progress(req);
1834 if (orig && OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_RESEND_RACE)) {
1835 spin_unlock_bh(&exp->exp_rpc_lock);
1837 OBD_RACE(OBD_FAIL_PTLRPC_RESEND_RACE);
1838 msleep(4 * MSEC_PER_SEC);
1842 if (orig && likely(atomic_inc_not_zero(&orig->rq_refcount))) {
1845 spin_unlock_bh(&exp->exp_rpc_lock);
1848 * When the client resend request and the server has
1849 * the previous copy of it, we need to update deadlines,
1850 * to be sure that the client and the server have equal
1851 * request deadlines.
1854 spin_lock(&orig->rq_rqbd->rqbd_svcpt->scp_at_lock);
1855 linked = orig->rq_at_linked;
1857 ptlrpc_at_remove_timed(orig);
1858 spin_unlock(&orig->rq_rqbd->rqbd_svcpt->scp_at_lock);
1859 orig->rq_deadline = req->rq_deadline;
1861 ptlrpc_at_add_timed(orig);
1862 ptlrpc_server_drop_request(orig);
1863 ptlrpc_nrs_req_finalize(req);
1865 /* don't mark slot unused for resend in progress */
1866 req->rq_obsolete = 1;
1871 ptlrpc_add_exp_list_nolock(req, exp, hp || req->rq_ops != NULL);
1873 spin_unlock_bh(&exp->exp_rpc_lock);
1878 * the current thread is not the processing thread for this request
1879 * since that, but request is in exp_hp_list and can be find there.
1880 * Remove all relations between request and old thread.
1882 req->rq_svc_thread->t_env->le_ses = NULL;
1883 req->rq_svc_thread = NULL;
1884 req->rq_session.lc_thread = NULL;
1886 ptlrpc_nrs_req_add(svcpt, req, hp);
1892 * Allow to handle high priority request
1893 * User can call it w/o any lock but need to hold
1894 * ptlrpc_service_part::scp_req_lock to get reliable result
1896 static bool ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1899 int running = svcpt->scp_nthrs_running;
1901 if (!nrs_svcpt_has_hp(svcpt))
1907 if (ptlrpc_nrs_req_throttling_nolock(svcpt, true))
1910 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1911 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1912 /* leave just 1 thread for normal RPCs */
1913 running = PTLRPC_NTHRS_INIT;
1914 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1918 if (svcpt->scp_nreqs_active >= running - 1)
1921 if (svcpt->scp_nhreqs_active == 0)
1924 return !ptlrpc_nrs_req_pending_nolock(svcpt, false) ||
1925 svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1928 static bool ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1931 return ptlrpc_server_allow_high(svcpt, force) &&
1932 ptlrpc_nrs_req_pending_nolock(svcpt, true);
1936 * Only allow normal priority requests on a service that has a high-priority
1937 * queue if forced (i.e. cleanup), if there are other high priority requests
1938 * already being processed (i.e. those threads can service more high-priority
1939 * requests), or if there are enough idle threads that a later thread can do
1940 * a high priority request.
1941 * User can call it w/o any lock but need to hold
1942 * ptlrpc_service_part::scp_req_lock to get reliable result
1944 static bool ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1947 int running = svcpt->scp_nthrs_running;
1949 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1950 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1951 /* leave just 1 thread for normal RPCs */
1952 running = PTLRPC_NTHRS_INIT;
1953 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1960 if (ptlrpc_nrs_req_throttling_nolock(svcpt, false))
1963 if (svcpt->scp_nreqs_active < running - 2)
1966 if (svcpt->scp_nreqs_active >= running - 1)
1969 return svcpt->scp_nhreqs_active > 0 || !nrs_svcpt_has_hp(svcpt);
1972 static bool ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1975 return ptlrpc_server_allow_normal(svcpt, force) &&
1976 ptlrpc_nrs_req_pending_nolock(svcpt, false);
1980 * Returns true if there are requests available in incoming
1981 * request queue for processing and it is allowed to fetch them.
1982 * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1983 * to get reliable result
1984 * \see ptlrpc_server_allow_normal
1985 * \see ptlrpc_server_allow high
1988 bool ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt,
1991 return ptlrpc_server_high_pending(svcpt, force) ||
1992 ptlrpc_server_normal_pending(svcpt, force);
1996 * Fetch a request for processing from queue of unprocessed requests.
1997 * Favors high-priority requests.
1998 * Returns a pointer to fetched request.
2000 static struct ptlrpc_request *
2001 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force)
2003 struct ptlrpc_request *req = NULL;
2007 spin_lock(&svcpt->scp_req_lock);
2009 if (ptlrpc_server_high_pending(svcpt, force)) {
2010 req = ptlrpc_nrs_req_get_nolock(svcpt, true, force);
2012 svcpt->scp_hreq_count++;
2017 if (ptlrpc_server_normal_pending(svcpt, force)) {
2018 req = ptlrpc_nrs_req_get_nolock(svcpt, false, force);
2020 svcpt->scp_hreq_count = 0;
2025 spin_unlock(&svcpt->scp_req_lock);
2029 svcpt->scp_nreqs_active++;
2031 svcpt->scp_nhreqs_active++;
2033 spin_unlock(&svcpt->scp_req_lock);
2035 if (likely(req->rq_export))
2036 class_export_rpc_inc(req->rq_export);
2042 * Handle freshly incoming reqs, add to timed early reply list,
2043 * pass on to regular request queue.
2044 * All incoming requests pass through here before getting into
2045 * ptlrpc_server_handle_req later on.
2047 static int ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt,
2048 struct ptlrpc_thread *thread)
2050 struct ptlrpc_service *svc = svcpt->scp_service;
2051 struct ptlrpc_request *req;
2057 spin_lock(&svcpt->scp_lock);
2058 if (list_empty(&svcpt->scp_req_incoming)) {
2059 spin_unlock(&svcpt->scp_lock);
2063 req = list_entry(svcpt->scp_req_incoming.next,
2064 struct ptlrpc_request, rq_list);
2065 list_del_init(&req->rq_list);
2066 svcpt->scp_nreqs_incoming--;
2068 * Consider this still a "queued" request as far as stats are
2071 spin_unlock(&svcpt->scp_lock);
2073 /* go through security check/transform */
2074 rc = sptlrpc_svc_unwrap_request(req);
2078 case SECSVC_COMPLETE:
2079 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
2088 * for null-flavored rpc, msg has been unpacked by sptlrpc, although
2089 * redo it wouldn't be harmful.
2091 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
2092 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
2094 CERROR("error unpacking request: ptl %d from %s x%llu\n",
2095 svc->srv_req_portal, libcfs_id2str(req->rq_peer),
2101 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
2103 CERROR("error unpacking ptlrpc body: ptl %d from %s x %llu\n",
2104 svc->srv_req_portal, libcfs_id2str(req->rq_peer),
2109 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
2110 lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
2111 CERROR("drop incoming rpc opc %u, x%llu\n",
2112 cfs_fail_val, req->rq_xid);
2117 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
2118 CERROR("wrong packet type received (type=%u) from %s\n",
2119 lustre_msg_get_type(req->rq_reqmsg),
2120 libcfs_id2str(req->rq_peer));
2124 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2128 req->rq_bulk_write = 1;
2132 case MGS_CONFIG_READ:
2133 req->rq_bulk_read = 1;
2137 CDEBUG(D_RPCTRACE, "got req x%llu\n", req->rq_xid);
2139 req->rq_export = class_conn2export(
2140 lustre_msg_get_handle(req->rq_reqmsg));
2141 if (req->rq_export) {
2142 rc = ptlrpc_check_req(req);
2144 rc = sptlrpc_target_export_check(req->rq_export, req);
2146 DEBUG_REQ(D_ERROR, req,
2147 "DROPPING req with illegal security flavor");
2152 ptlrpc_update_export_timer(req->rq_export, 0);
2155 /* req_in handling should/must be fast */
2156 if (ktime_get_real_seconds() - req->rq_arrival_time.tv_sec > 5)
2157 DEBUG_REQ(D_WARNING, req, "Slow req_in handling %llds",
2158 ktime_get_real_seconds() -
2159 req->rq_arrival_time.tv_sec);
2161 /* Set rpc server deadline and add it to the timed list */
2162 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
2163 MSGHDR_AT_SUPPORT) ?
2164 /* The max time the client expects us to take */
2165 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
2167 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
2168 if (unlikely(deadline == 0)) {
2169 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
2173 /* Skip early reply */
2174 if (OBD_FAIL_PRECHECK(OBD_FAIL_MDS_RESEND))
2175 req->rq_deadline += obd_timeout;
2177 req->rq_svc_thread = thread;
2178 if (thread != NULL) {
2180 * initialize request session, it is needed for request
2181 * processing by target
2183 rc = lu_context_init(&req->rq_session, LCT_SERVER_SESSION |
2186 CERROR("%s: failure to initialize session: rc = %d\n",
2187 thread->t_name, rc);
2190 req->rq_session.lc_thread = thread;
2191 lu_context_enter(&req->rq_session);
2192 thread->t_env->le_ses = &req->rq_session;
2195 ptlrpc_at_add_timed(req);
2197 /* Move it over to the request processing queue */
2198 rc = ptlrpc_server_request_add(svcpt, req);
2202 wake_up(&svcpt->scp_waitq);
2206 ptlrpc_server_finish_request(svcpt, req);
2212 * Main incoming request handling logic.
2213 * Calls handler function from service to do actual processing.
2215 static int ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
2216 struct ptlrpc_thread *thread)
2218 struct ptlrpc_service *svc = svcpt->scp_service;
2219 struct ptlrpc_request *request;
2229 request = ptlrpc_server_request_get(svcpt, false);
2230 if (request == NULL)
2233 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
2234 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
2235 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
2236 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
2238 if (unlikely(fail_opc)) {
2239 if (request->rq_export && request->rq_ops)
2240 OBD_FAIL_TIMEOUT(fail_opc, 4);
2243 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
2245 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
2246 libcfs_debug_dumplog();
2248 work_start = ktime_get_real();
2249 arrived = timespec64_to_ktime(request->rq_arrival_time);
2250 timediff_usecs = ktime_us_delta(work_start, arrived);
2251 if (likely(svc->srv_stats != NULL)) {
2252 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
2254 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
2255 svcpt->scp_nreqs_incoming);
2256 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
2257 svcpt->scp_nreqs_active);
2258 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
2259 at_get(&svcpt->scp_at_estimate));
2262 if (likely(request->rq_export)) {
2263 if (unlikely(ptlrpc_check_req(request)))
2265 ptlrpc_update_export_timer(request->rq_export,
2266 div_u64(timediff_usecs,
2271 * Discard requests queued for longer than the deadline.
2272 * The deadline is increased if we send an early reply.
2274 if (ktime_get_real_seconds() > request->rq_deadline) {
2275 DEBUG_REQ(D_ERROR, request,
2276 "Dropping timed-out request from %s: deadline %lld/%llds ago",
2277 libcfs_id2str(request->rq_peer),
2278 request->rq_deadline -
2279 request->rq_arrival_time.tv_sec,
2280 ktime_get_real_seconds() - request->rq_deadline);
2285 "Handling RPC req@%p pname:cluuid+ref:pid:xid:nid:opc:job %s:%s+%d:%d:x%llu:%s:%d:%s\n",
2286 request, current->comm,
2287 (request->rq_export ?
2288 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2289 (request->rq_export ?
2290 refcount_read(&request->rq_export->exp_handle.h_ref) : -99),
2291 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
2292 libcfs_id2str(request->rq_peer),
2293 lustre_msg_get_opc(request->rq_reqmsg),
2294 lustre_msg_get_jobid(request->rq_reqmsg) ?: "");
2296 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
2297 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
2299 CDEBUG(D_NET, "got req %llu\n", request->rq_xid);
2301 /* re-assign request and sesson thread to the current one */
2302 request->rq_svc_thread = thread;
2303 if (thread != NULL) {
2304 LASSERT(request->rq_session.lc_thread == NULL);
2305 request->rq_session.lc_thread = thread;
2306 thread->t_env->le_ses = &request->rq_session;
2308 svc->srv_ops.so_req_handler(request);
2310 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
2313 if (unlikely(ktime_get_real_seconds() > request->rq_deadline)) {
2314 DEBUG_REQ(D_WARNING, request,
2315 "Request took longer than estimated (%lld/%llds); client may timeout",
2316 request->rq_deadline -
2317 request->rq_arrival_time.tv_sec,
2318 ktime_get_real_seconds() - request->rq_deadline);
2321 work_end = ktime_get_real();
2322 timediff_usecs = ktime_us_delta(work_end, work_start);
2323 arrived_usecs = ktime_us_delta(work_end, arrived);
2325 "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",
2326 request, current->comm,
2327 (request->rq_export ?
2328 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2329 (request->rq_export ?
2330 refcount_read(&request->rq_export->exp_handle.h_ref) : -99),
2331 lustre_msg_get_status(request->rq_reqmsg),
2333 libcfs_id2str(request->rq_peer),
2334 lustre_msg_get_opc(request->rq_reqmsg),
2335 lustre_msg_get_jobid(request->rq_reqmsg) ?: "",
2338 (request->rq_repmsg ?
2339 lustre_msg_get_transno(request->rq_repmsg) :
2340 request->rq_transno),
2342 (request->rq_repmsg ?
2343 lustre_msg_get_status(request->rq_repmsg) : -999));
2344 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
2345 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
2346 int opc = opcode_offset(op);
2348 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
2349 LASSERT(opc < LUSTRE_MAX_OPCODES);
2350 lprocfs_counter_add(svc->srv_stats,
2351 opc + EXTRA_MAX_OPCODES,
2355 if (unlikely(request->rq_early_count)) {
2356 DEBUG_REQ(D_ADAPTTO, request,
2357 "sent %d early replies before finishing in %llds",
2358 request->rq_early_count,
2359 div_u64(arrived_usecs, USEC_PER_SEC));
2362 ptlrpc_server_finish_active_request(svcpt, request);
2368 * An internal function to process a single reply state object.
2370 static int ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
2372 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
2373 struct ptlrpc_service *svc = svcpt->scp_service;
2374 struct obd_export *exp;
2380 exp = rs->rs_export;
2382 LASSERT(rs->rs_difficult);
2383 LASSERT(rs->rs_scheduled);
2384 LASSERT(list_empty(&rs->rs_list));
2387 * The disk commit callback holds exp_uncommitted_replies_lock while it
2388 * iterates over newly committed replies, removing them from
2389 * exp_uncommitted_replies. It then drops this lock and schedules the
2390 * replies it found for handling here.
2392 * We can avoid contention for exp_uncommitted_replies_lock between the
2393 * HRT threads and further commit callbacks by checking rs_committed
2394 * which is set in the commit callback while it holds both
2395 * rs_lock and exp_uncommitted_reples.
2397 * If we see rs_committed clear, the commit callback _may_ not have
2398 * handled this reply yet and we race with it to grab
2399 * exp_uncommitted_replies_lock before removing the reply from
2400 * exp_uncommitted_replies. Note that if we lose the race and the
2401 * reply has already been removed, list_del_init() is a noop.
2403 * If we see rs_committed set, we know the commit callback is handling,
2404 * or has handled this reply since store reordering might allow us to
2405 * see rs_committed set out of sequence. But since this is done
2406 * holding rs_lock, we can be sure it has all completed once we hold
2407 * rs_lock, which we do right next.
2409 if (!rs->rs_committed) {
2411 * if rs was commited, no need to convert locks, don't check
2412 * rs_committed here because rs may never be added into
2413 * exp_uncommitted_replies and this flag never be set, see
2414 * target_send_reply()
2416 if (rs->rs_convert_lock &&
2417 rs->rs_transno > exp->exp_last_committed) {
2418 struct ldlm_lock *lock;
2419 struct ldlm_lock *ack_locks[RS_MAX_LOCKS] = { NULL };
2421 spin_lock(&rs->rs_lock);
2422 if (rs->rs_convert_lock &&
2423 rs->rs_transno > exp->exp_last_committed) {
2424 nlocks = rs->rs_nlocks;
2425 while (nlocks-- > 0) {
2427 * NB don't assume rs is always handled
2428 * by the same service thread (see
2429 * ptlrpc_hr_select, so REP-ACK hr may
2430 * race with trans commit, while the
2431 * latter will release locks, get locks
2432 * here early to convert to COS mode
2435 lock = ldlm_handle2lock(
2436 &rs->rs_locks[nlocks]);
2438 ack_locks[nlocks] = lock;
2439 rs->rs_modes[nlocks] = LCK_COS;
2441 nlocks = rs->rs_nlocks;
2442 rs->rs_convert_lock = 0;
2444 * clear rs_scheduled so that commit callback
2445 * can schedule again
2447 rs->rs_scheduled = 0;
2448 spin_unlock(&rs->rs_lock);
2450 while (nlocks-- > 0) {
2451 lock = ack_locks[nlocks];
2452 ldlm_lock_mode_downgrade(lock, LCK_COS);
2453 LDLM_LOCK_PUT(lock);
2457 spin_unlock(&rs->rs_lock);
2460 spin_lock(&exp->exp_uncommitted_replies_lock);
2461 list_del_init(&rs->rs_obd_list);
2462 spin_unlock(&exp->exp_uncommitted_replies_lock);
2465 spin_lock(&exp->exp_lock);
2466 /* Noop if removed already */
2467 list_del_init(&rs->rs_exp_list);
2468 spin_unlock(&exp->exp_lock);
2470 spin_lock(&rs->rs_lock);
2472 been_handled = rs->rs_handled;
2475 nlocks = rs->rs_nlocks; /* atomic "steal", but */
2476 rs->rs_nlocks = 0; /* locks still on rs_locks! */
2478 if (nlocks == 0 && !been_handled) {
2480 * If we see this, we should already have seen the warning
2481 * in mds_steal_ack_locks()
2484 "All locks stolen from rs %p x%lld.t%lld o%d NID %s\n",
2485 rs, rs->rs_xid, rs->rs_transno, rs->rs_opc,
2486 libcfs_nid2str(exp->exp_connection->c_peer.nid));
2489 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
2490 spin_unlock(&rs->rs_lock);
2492 if (!been_handled && rs->rs_on_net) {
2493 LNetMDUnlink(rs->rs_md_h);
2494 /* Ignore return code; we're racing with completion */
2497 while (nlocks-- > 0)
2498 ldlm_lock_decref(&rs->rs_locks[nlocks],
2499 rs->rs_modes[nlocks]);
2501 spin_lock(&rs->rs_lock);
2504 rs->rs_scheduled = 0;
2505 rs->rs_convert_lock = 0;
2507 if (!rs->rs_on_net) {
2509 spin_unlock(&rs->rs_lock);
2511 class_export_put(exp);
2512 rs->rs_export = NULL;
2513 ptlrpc_rs_decref(rs);
2514 if (atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
2515 svc->srv_is_stopping)
2516 wake_up_all(&svcpt->scp_waitq);
2520 /* still on the net; callback will schedule */
2521 spin_unlock(&rs->rs_lock);
2526 static void ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
2528 int avail = svcpt->scp_nrqbds_posted;
2529 int low_water = test_req_buffer_pressure ? 0 :
2530 svcpt->scp_service->srv_nbuf_per_group / 2;
2532 /* NB I'm not locking; just looking. */
2535 * CAVEAT EMPTOR: We might be allocating buffers here because we've
2536 * allowed the request history to grow out of control. We could put a
2537 * sanity check on that here and cull some history if we need the
2541 if (avail <= low_water)
2542 ptlrpc_grow_req_bufs(svcpt, 1);
2544 if (svcpt->scp_service->srv_stats) {
2545 lprocfs_counter_add(svcpt->scp_service->srv_stats,
2546 PTLRPC_REQBUF_AVAIL_CNTR, avail);
2550 static inline int ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
2552 return svcpt->scp_nreqs_active <
2553 svcpt->scp_nthrs_running - 1 -
2554 (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL);
2558 * allowed to create more threads
2559 * user can call it w/o any lock but need to hold
2560 * ptlrpc_service_part::scp_lock to get reliable result
2562 static inline int ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
2564 return svcpt->scp_nthrs_running +
2565 svcpt->scp_nthrs_starting <
2566 svcpt->scp_service->srv_nthrs_cpt_limit;
2570 * too many requests and allowed to create more threads
2572 static inline int ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
2574 return !ptlrpc_threads_enough(svcpt) &&
2575 ptlrpc_threads_increasable(svcpt);
2578 static inline int ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2580 return thread_is_stopping(thread) ||
2581 thread->t_svcpt->scp_service->srv_is_stopping;
2584 /* stop the highest numbered thread if there are too many threads running */
2585 static inline bool ptlrpc_thread_should_stop(struct ptlrpc_thread *thread)
2587 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2589 return thread->t_id >= svcpt->scp_service->srv_nthrs_cpt_limit &&
2590 thread->t_id == svcpt->scp_thr_nextid - 1;
2593 static void ptlrpc_stop_thread(struct ptlrpc_thread *thread)
2595 CDEBUG(D_INFO, "Stopping thread %s #%u\n",
2596 thread->t_svcpt->scp_service->srv_thread_name, thread->t_id);
2597 thread_add_flags(thread, SVC_STOPPING);
2600 static inline void ptlrpc_thread_stop(struct ptlrpc_thread *thread)
2602 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2604 spin_lock(&svcpt->scp_lock);
2605 if (ptlrpc_thread_should_stop(thread)) {
2606 ptlrpc_stop_thread(thread);
2607 svcpt->scp_thr_nextid--;
2609 spin_unlock(&svcpt->scp_lock);
2612 static inline int ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
2614 return !list_empty(&svcpt->scp_rqbd_idle) &&
2615 svcpt->scp_rqbd_timeout == 0;
2619 ptlrpc_at_check(struct ptlrpc_service_part *svcpt)
2621 return svcpt->scp_at_check;
2625 * If a thread runs too long or spends to much time on a single request,
2626 * we want to know about it, so we set up a delayed work item as a watchdog.
2627 * If it fires, we display a stack trace of the delayed thread,
2628 * providing we aren't rate-limited
2630 * Watchdog stack traces are limited to 3 per 'libcfs_watchdog_ratelimit'
2633 static struct ratelimit_state watchdog_limit;
2635 static void ptlrpc_watchdog_fire(struct work_struct *w)
2637 struct ptlrpc_thread *thread = container_of(w, struct ptlrpc_thread,
2639 u64 ms_lapse = ktime_ms_delta(ktime_get(), thread->t_touched);
2640 u32 ms_frac = do_div(ms_lapse, MSEC_PER_SEC);
2642 /* ___ratelimit() returns true if the action is NOT ratelimited */
2643 if (__ratelimit(&watchdog_limit)) {
2644 /* below message is checked in sanity-quota.sh test_6,18 */
2645 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",
2646 thread->t_task->comm, thread->t_task->pid,
2649 libcfs_debug_dumpstack(thread->t_task);
2651 /* below message is checked in sanity-quota.sh test_6,18 */
2652 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",
2653 thread->t_task->comm, thread->t_task->pid,
2654 ms_lapse, ms_frac, libcfs_watchdog_ratelimit);
2658 void ptlrpc_watchdog_init(struct delayed_work *work, timeout_t timeout)
2660 INIT_DELAYED_WORK(work, ptlrpc_watchdog_fire);
2661 schedule_delayed_work(work, cfs_time_seconds(timeout));
2664 void ptlrpc_watchdog_disable(struct delayed_work *work)
2666 cancel_delayed_work_sync(work);
2669 void ptlrpc_watchdog_touch(struct delayed_work *work, timeout_t timeout)
2671 struct ptlrpc_thread *thread = container_of(&work->work,
2672 struct ptlrpc_thread,
2674 thread->t_touched = ktime_get();
2675 mod_delayed_work(system_wq, work, cfs_time_seconds(timeout));
2679 * requests wait on preprocessing
2680 * user can call it w/o any lock but need to hold
2681 * ptlrpc_service_part::scp_lock to get reliable result
2684 ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
2686 return !list_empty(&svcpt->scp_req_incoming);
2689 static __attribute__((__noinline__)) int
2690 ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
2691 struct ptlrpc_thread *thread)
2693 ptlrpc_watchdog_disable(&thread->t_watchdog);
2697 if (svcpt->scp_rqbd_timeout == 0)
2698 /* Don't exit while there are replies to be handled */
2699 wait_event_idle_exclusive_lifo(
2701 ptlrpc_thread_stopping(thread) ||
2702 ptlrpc_server_request_incoming(svcpt) ||
2703 ptlrpc_server_request_pending(svcpt, false) ||
2704 ptlrpc_rqbd_pending(svcpt) ||
2705 ptlrpc_at_check(svcpt));
2706 else if (wait_event_idle_exclusive_lifo_timeout(
2708 ptlrpc_thread_stopping(thread) ||
2709 ptlrpc_server_request_incoming(svcpt) ||
2710 ptlrpc_server_request_pending(svcpt, false) ||
2711 ptlrpc_rqbd_pending(svcpt) ||
2712 ptlrpc_at_check(svcpt),
2713 svcpt->scp_rqbd_timeout) == 0)
2714 svcpt->scp_rqbd_timeout = 0;
2716 if (ptlrpc_thread_stopping(thread))
2719 ptlrpc_watchdog_touch(&thread->t_watchdog,
2720 ptlrpc_server_get_timeout(svcpt));
2725 * Main thread body for service threads.
2726 * Waits in a loop waiting for new requests to process to appear.
2727 * Every time an incoming requests is added to its queue, a waitq
2728 * is woken up and one of the threads will handle it.
2730 static int ptlrpc_main(void *arg)
2732 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
2733 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2734 struct ptlrpc_service *svc = svcpt->scp_service;
2735 struct ptlrpc_reply_state *rs;
2736 struct group_info *ginfo = NULL;
2738 int counter = 0, rc = 0;
2742 thread->t_task = current;
2743 thread->t_pid = current->pid;
2745 if (svc->srv_cpt_bind) {
2746 rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
2748 CWARN("%s: failed to bind %s on CPT %d\n",
2749 svc->srv_name, thread->t_name, svcpt->scp_cpt);
2753 ginfo = groups_alloc(0);
2755 GOTO(out, rc = -ENOMEM);
2757 set_current_groups(ginfo);
2758 put_group_info(ginfo);
2760 if (svc->srv_ops.so_thr_init != NULL) {
2761 rc = svc->srv_ops.so_thr_init(thread);
2768 GOTO(out_srv_fini, rc = -ENOMEM);
2769 rc = lu_env_add(env);
2773 rc = lu_context_init(&env->le_ctx,
2774 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2776 GOTO(out_env_remove, rc);
2778 thread->t_env = env;
2779 env->le_ctx.lc_thread = thread;
2780 env->le_ctx.lc_cookie = 0x6;
2782 while (!list_empty(&svcpt->scp_rqbd_idle)) {
2783 rc = ptlrpc_server_post_idle_rqbds(svcpt);
2787 CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
2788 svc->srv_name, svcpt->scp_cpt, rc);
2789 GOTO(out_ctx_fini, rc);
2792 /* Alloc reply state structure for this one */
2793 OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2795 GOTO(out_ctx_fini, rc = -ENOMEM);
2797 spin_lock(&svcpt->scp_lock);
2799 LASSERT(thread_is_starting(thread));
2800 thread_clear_flags(thread, SVC_STARTING);
2802 LASSERT(svcpt->scp_nthrs_starting == 1);
2803 svcpt->scp_nthrs_starting--;
2806 * SVC_STOPPING may already be set here if someone else is trying
2807 * to stop the service while this new thread has been dynamically
2808 * forked. We still set SVC_RUNNING to let our creator know that
2809 * we are now running, however we will exit as soon as possible
2811 thread_add_flags(thread, SVC_RUNNING);
2812 svcpt->scp_nthrs_running++;
2813 spin_unlock(&svcpt->scp_lock);
2815 /* wake up our creator in case he's still waiting. */
2816 wake_up(&thread->t_ctl_waitq);
2818 thread->t_touched = ktime_get();
2819 ptlrpc_watchdog_init(&thread->t_watchdog,
2820 ptlrpc_server_get_timeout(svcpt));
2822 spin_lock(&svcpt->scp_rep_lock);
2823 list_add(&rs->rs_list, &svcpt->scp_rep_idle);
2824 wake_up(&svcpt->scp_rep_waitq);
2825 spin_unlock(&svcpt->scp_rep_lock);
2827 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2828 svcpt->scp_nthrs_running);
2830 /* XXX maintain a list of all managed devices: insert here */
2831 while (!ptlrpc_thread_stopping(thread)) {
2832 if (ptlrpc_wait_event(svcpt, thread))
2835 ptlrpc_check_rqbd_pool(svcpt);
2837 if (ptlrpc_threads_need_create(svcpt)) {
2838 /* Ignore return code - we tried... */
2839 ptlrpc_start_thread(svcpt, 0);
2842 /* reset le_ses to initial state */
2844 /* Refill the context before execution to make sure
2845 * all thread keys are allocated */
2847 /* Process all incoming reqs before handling any */
2848 if (ptlrpc_server_request_incoming(svcpt)) {
2849 lu_context_enter(&env->le_ctx);
2850 ptlrpc_server_handle_req_in(svcpt, thread);
2851 lu_context_exit(&env->le_ctx);
2853 /* but limit ourselves in case of flood */
2854 if (counter++ < 100)
2859 if (ptlrpc_at_check(svcpt))
2860 ptlrpc_at_check_timed(svcpt);
2862 if (ptlrpc_server_request_pending(svcpt, false)) {
2863 lu_context_enter(&env->le_ctx);
2864 ptlrpc_server_handle_request(svcpt, thread);
2865 lu_context_exit(&env->le_ctx);
2868 if (ptlrpc_rqbd_pending(svcpt) &&
2869 ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
2871 * I just failed to repost request buffers.
2872 * Wait for a timeout (unless something else
2873 * happens) before I try again
2875 svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
2876 CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
2877 svcpt->scp_nrqbds_posted);
2880 * If the number of threads has been tuned downward and this
2881 * thread should be stopped, then stop in reverse order so the
2882 * the threads always have contiguous thread index values.
2884 if (unlikely(ptlrpc_thread_should_stop(thread)))
2885 ptlrpc_thread_stop(thread);
2888 ptlrpc_watchdog_disable(&thread->t_watchdog);
2891 lu_context_fini(&env->le_ctx);
2897 /* deconstruct service thread state created by ptlrpc_start_thread() */
2898 if (svc->srv_ops.so_thr_done != NULL)
2899 svc->srv_ops.so_thr_done(thread);
2901 CDEBUG(D_RPCTRACE, "%s: service thread [%p:%u] %d exiting: rc = %d\n",
2902 thread->t_name, thread, thread->t_pid, thread->t_id, rc);
2903 spin_lock(&svcpt->scp_lock);
2904 if (thread_test_and_clear_flags(thread, SVC_STARTING))
2905 svcpt->scp_nthrs_starting--;
2907 if (thread_test_and_clear_flags(thread, SVC_RUNNING)) {
2908 /* must know immediately */
2909 svcpt->scp_nthrs_running--;
2913 thread_add_flags(thread, SVC_STOPPED);
2915 wake_up(&thread->t_ctl_waitq);
2916 spin_unlock(&svcpt->scp_lock);
2921 static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
2922 struct list_head *replies)
2926 spin_lock(&hrt->hrt_lock);
2928 list_splice_init(&hrt->hrt_queue, replies);
2929 result = ptlrpc_hr.hr_stopping || !list_empty(replies);
2931 spin_unlock(&hrt->hrt_lock);
2936 * Main body of "handle reply" function.
2937 * It processes acked reply states
2939 static int ptlrpc_hr_main(void *arg)
2941 struct ptlrpc_hr_thread *hrt = (struct ptlrpc_hr_thread *)arg;
2942 struct ptlrpc_hr_partition *hrp = hrt->hrt_partition;
2951 rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
2953 char threadname[20];
2955 snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
2956 hrp->hrp_cpt, hrt->hrt_id);
2957 CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
2958 threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
2961 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD |
2962 LCT_REMEMBER | LCT_NOREF);
2966 rc = lu_env_add(env);
2968 GOTO(out_ctx_fini, rc);
2970 atomic_inc(&hrp->hrp_nstarted);
2971 wake_up(&ptlrpc_hr.hr_waitq);
2973 while (!ptlrpc_hr.hr_stopping) {
2974 wait_event_idle(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
2976 while (!list_empty(&replies)) {
2977 struct ptlrpc_reply_state *rs;
2979 rs = list_entry(replies.prev,
2980 struct ptlrpc_reply_state,
2982 list_del_init(&rs->rs_list);
2983 /* refill keys if needed */
2985 lu_context_enter(&env->le_ctx);
2986 ptlrpc_handle_rs(rs);
2987 lu_context_exit(&env->le_ctx);
2991 atomic_inc(&hrp->hrp_nstopped);
2992 wake_up(&ptlrpc_hr.hr_waitq);
2996 lu_context_fini(&env->le_ctx);
3002 static void ptlrpc_stop_hr_threads(void)
3004 struct ptlrpc_hr_partition *hrp;
3008 ptlrpc_hr.hr_stopping = 1;
3010 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
3011 if (hrp->hrp_thrs == NULL)
3012 continue; /* uninitialized */
3013 for (j = 0; j < hrp->hrp_nthrs; j++)
3014 wake_up_all(&hrp->hrp_thrs[j].hrt_waitq);
3017 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
3018 if (hrp->hrp_thrs == NULL)
3019 continue; /* uninitialized */
3020 wait_event(ptlrpc_hr.hr_waitq,
3021 atomic_read(&hrp->hrp_nstopped) ==
3022 atomic_read(&hrp->hrp_nstarted));
3026 static int ptlrpc_start_hr_threads(void)
3028 struct ptlrpc_hr_partition *hrp;
3034 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
3037 for (j = 0; j < hrp->hrp_nthrs; j++) {
3038 struct ptlrpc_hr_thread *hrt = &hrp->hrp_thrs[j];
3039 struct task_struct *task;
3041 task = kthread_run(ptlrpc_hr_main,
3043 "ptlrpc_hr%02d_%03d",
3052 wait_event(ptlrpc_hr.hr_waitq,
3053 atomic_read(&hrp->hrp_nstarted) == j);
3056 CERROR("cannot start reply handler thread %d:%d: rc = %d\n",
3058 ptlrpc_stop_hr_threads();
3066 static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
3068 struct ptlrpc_thread *thread;
3073 CDEBUG(D_INFO, "Stopping threads for service %s\n",
3074 svcpt->scp_service->srv_name);
3076 spin_lock(&svcpt->scp_lock);
3077 /* let the thread know that we would like it to stop asap */
3078 list_for_each_entry(thread, &svcpt->scp_threads, t_link)
3079 ptlrpc_stop_thread(thread);
3081 wake_up_all(&svcpt->scp_waitq);
3083 while (!list_empty(&svcpt->scp_threads)) {
3084 thread = list_entry(svcpt->scp_threads.next,
3085 struct ptlrpc_thread, t_link);
3086 if (thread_is_stopped(thread)) {
3087 list_move(&thread->t_link, &zombie);
3090 spin_unlock(&svcpt->scp_lock);
3092 CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
3093 svcpt->scp_service->srv_thread_name, thread->t_id);
3094 wait_event_idle(thread->t_ctl_waitq,
3095 thread_is_stopped(thread));
3097 spin_lock(&svcpt->scp_lock);
3100 spin_unlock(&svcpt->scp_lock);
3102 while (!list_empty(&zombie)) {
3103 thread = list_entry(zombie.next,
3104 struct ptlrpc_thread, t_link);
3105 list_del(&thread->t_link);
3106 OBD_FREE_PTR(thread);
3112 * Stops all threads of a particular service \a svc
3114 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
3116 struct ptlrpc_service_part *svcpt;
3121 ptlrpc_service_for_each_part(svcpt, i, svc) {
3122 if (svcpt->scp_service != NULL)
3123 ptlrpc_svcpt_stop_threads(svcpt);
3129 int ptlrpc_start_threads(struct ptlrpc_service *svc)
3137 /* We require 2 threads min, see note in ptlrpc_server_handle_request */
3138 LASSERT(svc->srv_nthrs_cpt_init >= PTLRPC_NTHRS_INIT);
3140 for (i = 0; i < svc->srv_ncpts; i++) {
3141 for (j = 0; j < svc->srv_nthrs_cpt_init; j++) {
3142 rc = ptlrpc_start_thread(svc->srv_parts[i], 1);
3148 /* We have enough threads, don't start more. b=15759 */
3155 CERROR("cannot start %s thread #%d_%d: rc %d\n",
3156 svc->srv_thread_name, i, j, rc);
3157 ptlrpc_stop_all_threads(svc);
3161 int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
3163 struct ptlrpc_thread *thread;
3164 struct ptlrpc_service *svc;
3165 struct task_struct *task;
3170 LASSERT(svcpt != NULL);
3172 svc = svcpt->scp_service;
3174 CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n",
3175 svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running,
3176 svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit);
3179 if (unlikely(svc->srv_is_stopping))
3182 if (!ptlrpc_threads_increasable(svcpt) ||
3183 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
3184 svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1))
3187 OBD_CPT_ALLOC_PTR(thread, svc->srv_cptable, svcpt->scp_cpt);
3190 init_waitqueue_head(&thread->t_ctl_waitq);
3192 spin_lock(&svcpt->scp_lock);
3193 if (!ptlrpc_threads_increasable(svcpt)) {
3194 spin_unlock(&svcpt->scp_lock);
3195 OBD_FREE_PTR(thread);
3199 if (svcpt->scp_nthrs_starting != 0) {
3201 * serialize starting because some modules (obdfilter)
3202 * might require unique and contiguous t_id
3204 LASSERT(svcpt->scp_nthrs_starting == 1);
3205 spin_unlock(&svcpt->scp_lock);
3206 OBD_FREE_PTR(thread);
3208 CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n",
3209 svc->srv_thread_name, svcpt->scp_thr_nextid);
3214 CDEBUG(D_INFO, "Creating thread %s #%d race, retry later\n",
3215 svc->srv_thread_name, svcpt->scp_thr_nextid);
3219 svcpt->scp_nthrs_starting++;
3220 thread->t_id = svcpt->scp_thr_nextid++;
3221 thread_add_flags(thread, SVC_STARTING);
3222 thread->t_svcpt = svcpt;
3224 list_add(&thread->t_link, &svcpt->scp_threads);
3225 spin_unlock(&svcpt->scp_lock);
3227 if (svcpt->scp_cpt >= 0) {
3228 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s%02d_%03d",
3229 svc->srv_thread_name, svcpt->scp_cpt, thread->t_id);
3231 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s_%04d",
3232 svc->srv_thread_name, thread->t_id);
3235 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name);
3236 task = kthread_run(ptlrpc_main, thread, "%s", thread->t_name);
3239 CERROR("cannot start thread '%s': rc = %d\n",
3240 thread->t_name, rc);
3241 spin_lock(&svcpt->scp_lock);
3242 --svcpt->scp_nthrs_starting;
3243 if (thread_is_stopping(thread)) {
3245 * this ptlrpc_thread is being hanled
3246 * by ptlrpc_svcpt_stop_threads now
3248 thread_add_flags(thread, SVC_STOPPED);
3249 wake_up(&thread->t_ctl_waitq);
3250 spin_unlock(&svcpt->scp_lock);
3252 list_del(&thread->t_link);
3253 spin_unlock(&svcpt->scp_lock);
3254 OBD_FREE_PTR(thread);
3262 wait_event_idle(thread->t_ctl_waitq,
3263 thread_is_running(thread) || thread_is_stopped(thread));
3265 rc = thread_is_stopped(thread) ? thread->t_id : 0;
3269 int ptlrpc_hr_init(void)
3271 struct ptlrpc_hr_partition *hrp;
3272 struct ptlrpc_hr_thread *hrt;
3280 memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
3281 ptlrpc_hr.hr_cpt_table = cfs_cpt_tab;
3283 ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
3285 if (ptlrpc_hr.hr_partitions == NULL)
3288 ratelimit_state_init(&watchdog_limit,
3289 cfs_time_seconds(libcfs_watchdog_ratelimit), 3);
3291 init_waitqueue_head(&ptlrpc_hr.hr_waitq);
3294 weight = cpumask_weight(topology_sibling_cpumask(smp_processor_id()));
3297 cfs_percpt_for_each(hrp, cpt, ptlrpc_hr.hr_partitions) {
3300 atomic_set(&hrp->hrp_nstarted, 0);
3301 atomic_set(&hrp->hrp_nstopped, 0);
3303 hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, cpt);
3304 hrp->hrp_nthrs /= weight;
3305 if (hrp->hrp_nthrs == 0)
3308 OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, cpt,
3309 hrp->hrp_nthrs * sizeof(*hrt));
3310 if (hrp->hrp_thrs == NULL)
3311 GOTO(out, rc = -ENOMEM);
3313 for (i = 0; i < hrp->hrp_nthrs; i++) {
3314 hrt = &hrp->hrp_thrs[i];
3317 hrt->hrt_partition = hrp;
3318 init_waitqueue_head(&hrt->hrt_waitq);
3319 spin_lock_init(&hrt->hrt_lock);
3320 INIT_LIST_HEAD(&hrt->hrt_queue);
3324 rc = ptlrpc_start_hr_threads();
3331 void ptlrpc_hr_fini(void)
3333 struct ptlrpc_hr_partition *hrp;
3336 if (ptlrpc_hr.hr_partitions == NULL)
3339 ptlrpc_stop_hr_threads();
3341 cfs_percpt_for_each(hrp, cpt, ptlrpc_hr.hr_partitions) {
3343 OBD_FREE_PTR_ARRAY(hrp->hrp_thrs, hrp->hrp_nthrs);
3346 cfs_percpt_free(ptlrpc_hr.hr_partitions);
3347 ptlrpc_hr.hr_partitions = NULL;
3352 * Wait until all already scheduled replies are processed.
3354 static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
3357 if (wait_event_idle_timeout(
3359 atomic_read(&svcpt->scp_nreps_difficult) == 0,
3360 cfs_time_seconds(10)) > 0)
3362 CWARN("Unexpectedly long timeout %s %p\n",
3363 svcpt->scp_service->srv_name, svcpt->scp_service);
3368 ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
3370 struct ptlrpc_service_part *svcpt;
3373 /* early disarm AT timer... */
3374 ptlrpc_service_for_each_part(svcpt, i, svc) {
3375 if (svcpt->scp_service != NULL)
3376 del_timer(&svcpt->scp_at_timer);
3381 ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
3383 struct ptlrpc_service_part *svcpt;
3384 struct ptlrpc_request_buffer_desc *rqbd;
3389 * All history will be culled when the next request buffer is
3390 * freed in ptlrpc_service_purge_all()
3392 svc->srv_hist_nrqbds_cpt_max = 0;
3394 rc = LNetClearLazyPortal(svc->srv_req_portal);
3397 ptlrpc_service_for_each_part(svcpt, i, svc) {
3398 if (svcpt->scp_service == NULL)
3402 * Unlink all the request buffers. This forces a 'final'
3403 * event with its 'unlink' flag set for each posted rqbd
3405 list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
3407 rc = LNetMDUnlink(rqbd->rqbd_md_h);
3408 LASSERT(rc == 0 || rc == -ENOENT);
3412 ptlrpc_service_for_each_part(svcpt, i, svc) {
3413 if (svcpt->scp_service == NULL)
3417 * Wait for the network to release any buffers
3418 * it's currently filling
3420 spin_lock(&svcpt->scp_lock);
3421 while (svcpt->scp_nrqbds_posted != 0) {
3422 int seconds = PTLRPC_REQ_LONG_UNLINK;
3424 spin_unlock(&svcpt->scp_lock);
3426 * Network access will complete in finite time but
3427 * the HUGE timeout lets us CWARN for visibility
3430 while (seconds > 0 &&
3431 wait_event_idle_timeout(
3433 svcpt->scp_nrqbds_posted == 0,
3434 cfs_time_seconds(1)) == 0)
3437 CWARN("Service %s waiting for request buffers\n",
3438 svcpt->scp_service->srv_name);
3440 spin_lock(&svcpt->scp_lock);
3442 spin_unlock(&svcpt->scp_lock);
3447 ptlrpc_service_purge_all(struct ptlrpc_service *svc)
3449 struct ptlrpc_service_part *svcpt;
3450 struct ptlrpc_request_buffer_desc *rqbd;
3451 struct ptlrpc_request *req;
3452 struct ptlrpc_reply_state *rs;
3455 ptlrpc_service_for_each_part(svcpt, i, svc) {
3456 if (svcpt->scp_service == NULL)
3459 spin_lock(&svcpt->scp_rep_lock);
3460 while (!list_empty(&svcpt->scp_rep_active)) {
3461 rs = list_entry(svcpt->scp_rep_active.next,
3462 struct ptlrpc_reply_state, rs_list);
3463 spin_lock(&rs->rs_lock);
3464 ptlrpc_schedule_difficult_reply(rs);
3465 spin_unlock(&rs->rs_lock);
3467 spin_unlock(&svcpt->scp_rep_lock);
3470 * purge the request queue. NB No new replies (rqbds
3471 * all unlinked) and no service threads, so I'm the only
3472 * thread noodling the request queue now
3474 while (!list_empty(&svcpt->scp_req_incoming)) {
3475 req = list_entry(svcpt->scp_req_incoming.next,
3476 struct ptlrpc_request, rq_list);
3478 list_del(&req->rq_list);
3479 svcpt->scp_nreqs_incoming--;
3480 ptlrpc_server_finish_request(svcpt, req);
3483 while (ptlrpc_server_request_pending(svcpt, true)) {
3484 req = ptlrpc_server_request_get(svcpt, true);
3485 ptlrpc_server_finish_active_request(svcpt, req);
3489 * The portal may be shared by several services (eg:OUT_PORTAL).
3490 * So the request could be referenced by other target. So we
3491 * have to wait the ptlrpc_server_drop_request invoked.
3493 * TODO: move the req_buffer as global rather than per service.
3495 spin_lock(&svcpt->scp_lock);
3496 while (!list_empty(&svcpt->scp_rqbd_posted)) {
3497 spin_unlock(&svcpt->scp_lock);
3498 wait_event_idle_timeout(svcpt->scp_waitq,
3499 list_empty(&svcpt->scp_rqbd_posted),
3500 cfs_time_seconds(1));
3501 spin_lock(&svcpt->scp_lock);
3503 spin_unlock(&svcpt->scp_lock);
3505 LASSERT(svcpt->scp_nreqs_incoming == 0);
3506 LASSERT(svcpt->scp_nreqs_active == 0);
3508 * history should have been culled by
3509 * ptlrpc_server_finish_request
3511 LASSERT(svcpt->scp_hist_nrqbds == 0);
3514 * Now free all the request buffers since nothing
3515 * references them any more...
3518 while (!list_empty(&svcpt->scp_rqbd_idle)) {
3519 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
3520 struct ptlrpc_request_buffer_desc,
3522 ptlrpc_free_rqbd(rqbd);
3524 ptlrpc_wait_replies(svcpt);
3526 while (!list_empty(&svcpt->scp_rep_idle)) {
3527 rs = list_entry(svcpt->scp_rep_idle.next,
3528 struct ptlrpc_reply_state,
3530 list_del(&rs->rs_list);
3531 OBD_FREE_LARGE(rs, svc->srv_max_reply_size);
3537 ptlrpc_service_free(struct ptlrpc_service *svc)
3539 struct ptlrpc_service_part *svcpt;
3540 struct ptlrpc_at_array *array;
3543 ptlrpc_service_for_each_part(svcpt, i, svc) {
3544 if (svcpt->scp_service == NULL)
3547 /* In case somebody rearmed this in the meantime */
3548 del_timer(&svcpt->scp_at_timer);
3549 array = &svcpt->scp_at_array;
3551 if (array->paa_reqs_array != NULL) {
3552 OBD_FREE_PTR_ARRAY(array->paa_reqs_array,
3554 array->paa_reqs_array = NULL;
3557 if (array->paa_reqs_count != NULL) {
3558 OBD_FREE_PTR_ARRAY(array->paa_reqs_count,
3560 array->paa_reqs_count = NULL;
3564 ptlrpc_service_for_each_part(svcpt, i, svc)
3565 OBD_FREE_PTR(svcpt);
3567 if (svc->srv_cpts != NULL)
3568 cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts);
3570 OBD_FREE(svc, offsetof(struct ptlrpc_service,
3571 srv_parts[svc->srv_ncpts]));
3574 int ptlrpc_unregister_service(struct ptlrpc_service *service)
3578 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
3580 service->srv_is_stopping = 1;
3582 mutex_lock(&ptlrpc_all_services_mutex);
3583 list_del_init(&service->srv_list);
3584 mutex_unlock(&ptlrpc_all_services_mutex);
3586 ptlrpc_service_del_atimer(service);
3587 ptlrpc_stop_all_threads(service);
3589 ptlrpc_service_unlink_rqbd(service);
3590 ptlrpc_service_purge_all(service);
3591 ptlrpc_service_nrs_cleanup(service);
3593 ptlrpc_lprocfs_unregister_service(service);
3594 ptlrpc_sysfs_unregister_service(service);
3596 ptlrpc_service_free(service);
3600 EXPORT_SYMBOL(ptlrpc_unregister_service);
3603 * Returns 0 if the service is healthy.
3605 * Right now, it just checks to make sure that requests aren't languishing
3606 * in the queue. We'll use this health check to govern whether a node needs
3607 * to be shot, so it's intentionally non-aggressive.
3609 static int ptlrpc_svcpt_health_check(struct ptlrpc_service_part *svcpt)
3611 struct ptlrpc_request *request = NULL;
3612 struct timespec64 right_now;
3613 struct timespec64 timediff;
3615 ktime_get_real_ts64(&right_now);
3617 spin_lock(&svcpt->scp_req_lock);
3618 /* How long has the next entry been waiting? */
3619 if (ptlrpc_server_high_pending(svcpt, true))
3620 request = ptlrpc_nrs_req_peek_nolock(svcpt, true);
3621 else if (ptlrpc_server_normal_pending(svcpt, true))
3622 request = ptlrpc_nrs_req_peek_nolock(svcpt, false);
3624 if (request == NULL) {
3625 spin_unlock(&svcpt->scp_req_lock);
3629 timediff = timespec64_sub(right_now, request->rq_arrival_time);
3630 spin_unlock(&svcpt->scp_req_lock);
3632 if ((timediff.tv_sec) >
3633 (AT_OFF ? obd_timeout * 3 / 2 : at_max)) {
3634 CERROR("%s: unhealthy - request has been waiting %llds\n",
3635 svcpt->scp_service->srv_name, (s64)timediff.tv_sec);
3643 ptlrpc_service_health_check(struct ptlrpc_service *svc)
3645 struct ptlrpc_service_part *svcpt;
3651 ptlrpc_service_for_each_part(svcpt, i, svc) {
3652 int rc = ptlrpc_svcpt_health_check(svcpt);
3659 EXPORT_SYMBOL(ptlrpc_service_health_check);