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 struct 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_table for now */
233 struct cfs_cpt_table *hr_cpt_table;
234 /** controller sleep waitq */
235 wait_queue_head_t hr_waitq;
236 unsigned int hr_stopping;
237 /** roundrobin rotor for non-affinity service */
238 unsigned int hr_rotor;
240 struct ptlrpc_hr_partition **hr_partitions;
244 struct list_head rsb_replies;
245 unsigned int rsb_n_replies;
246 struct ptlrpc_service_part *rsb_svcpt;
249 /** reply handling service. */
250 static struct ptlrpc_hr_service ptlrpc_hr;
253 * maximum mumber of replies scheduled in one batch
255 #define MAX_SCHEDULED 256
258 * Initialize a reply batch.
262 static void rs_batch_init(struct rs_batch *b)
264 memset(b, 0, sizeof(*b));
265 INIT_LIST_HEAD(&b->rsb_replies);
269 * Choose an hr thread to dispatch requests to.
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(array->paa_reqs_count, sizeof(__u32) * size);
688 array->paa_reqs_count = NULL;
691 if (array->paa_reqs_array != NULL) {
692 OBD_FREE(array->paa_reqs_array,
693 sizeof(struct list_head) * array->paa_size);
694 array->paa_reqs_array = NULL;
701 * Initialize service on a given portal.
702 * This includes starting serving threads , allocating and posting rqbds and
705 struct ptlrpc_service *ptlrpc_register_service(struct ptlrpc_service_conf *conf,
707 struct dentry *debugfs_entry)
709 struct ptlrpc_service_cpt_conf *cconf = &conf->psc_cpt;
710 struct ptlrpc_service *service;
711 struct ptlrpc_service_part *svcpt;
712 struct cfs_cpt_table *cptable;
721 LASSERT(conf->psc_buf.bc_nbufs > 0);
722 LASSERT(conf->psc_buf.bc_buf_size >=
723 conf->psc_buf.bc_req_max_size + SPTLRPC_MAX_PAYLOAD);
724 LASSERT(conf->psc_thr.tc_ctx_tags != 0);
726 cptable = cconf->cc_cptable;
728 cptable = cfs_cpt_table;
730 if (conf->psc_thr.tc_cpu_bind > 1) {
731 CERROR("%s: Invalid cpu bind value %d, only 1 or 0 allowed\n",
732 conf->psc_name, conf->psc_thr.tc_cpu_bind);
733 RETURN(ERR_PTR(-EINVAL));
736 if (!cconf->cc_affinity) {
739 ncpts = cfs_cpt_number(cptable);
740 if (cconf->cc_pattern != NULL) {
741 struct cfs_expr_list *el;
743 rc = cfs_expr_list_parse(cconf->cc_pattern,
744 strlen(cconf->cc_pattern),
747 CERROR("%s: invalid CPT pattern string: %s",
748 conf->psc_name, cconf->cc_pattern);
749 RETURN(ERR_PTR(-EINVAL));
752 rc = cfs_expr_list_values(el, ncpts, &cpts);
753 cfs_expr_list_free(el);
755 CERROR("%s: failed to parse CPT array %s: %d\n",
756 conf->psc_name, cconf->cc_pattern, rc);
758 OBD_FREE(cpts, sizeof(*cpts) * ncpts);
759 RETURN(ERR_PTR(rc < 0 ? rc : -EINVAL));
765 OBD_ALLOC(service, offsetof(struct ptlrpc_service, srv_parts[ncpts]));
766 if (service == NULL) {
768 OBD_FREE(cpts, sizeof(*cpts) * ncpts);
769 RETURN(ERR_PTR(-ENOMEM));
772 service->srv_cptable = cptable;
773 service->srv_cpts = cpts;
774 service->srv_ncpts = ncpts;
775 service->srv_cpt_bind = conf->psc_thr.tc_cpu_bind;
777 service->srv_cpt_bits = 0; /* it's zero already, easy to read... */
778 while ((1 << service->srv_cpt_bits) < cfs_cpt_number(cptable))
779 service->srv_cpt_bits++;
782 spin_lock_init(&service->srv_lock);
783 service->srv_name = conf->psc_name;
784 service->srv_watchdog_factor = conf->psc_watchdog_factor;
785 INIT_LIST_HEAD(&service->srv_list); /* for safty of cleanup */
787 /* buffer configuration */
788 service->srv_nbuf_per_group = test_req_buffer_pressure ?
789 1 : conf->psc_buf.bc_nbufs;
790 /* do not limit max number of rqbds by default */
791 service->srv_nrqbds_max = 0;
793 service->srv_max_req_size = conf->psc_buf.bc_req_max_size +
795 service->srv_buf_size = conf->psc_buf.bc_buf_size;
796 service->srv_rep_portal = conf->psc_buf.bc_rep_portal;
797 service->srv_req_portal = conf->psc_buf.bc_req_portal;
799 /* With slab/alloc_pages buffer size will be rounded up to 2^n */
800 if (service->srv_buf_size & (service->srv_buf_size - 1)) {
801 int round = size_roundup_power2(service->srv_buf_size);
803 service->srv_buf_size = round;
806 /* Increase max reply size to next power of two */
807 service->srv_max_reply_size = 1;
808 while (service->srv_max_reply_size <
809 conf->psc_buf.bc_rep_max_size + SPTLRPC_MAX_PAYLOAD)
810 service->srv_max_reply_size <<= 1;
812 service->srv_thread_name = conf->psc_thr.tc_thr_name;
813 service->srv_ctx_tags = conf->psc_thr.tc_ctx_tags;
814 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
815 service->srv_ops = conf->psc_ops;
817 for (i = 0; i < ncpts; i++) {
818 if (!cconf->cc_affinity)
821 cpt = cpts != NULL ? cpts[i] : i;
823 OBD_CPT_ALLOC(svcpt, cptable, cpt, sizeof(*svcpt));
825 GOTO(failed, rc = -ENOMEM);
827 service->srv_parts[i] = svcpt;
828 rc = ptlrpc_service_part_init(service, svcpt, cpt);
833 ptlrpc_server_nthreads_check(service, conf);
835 rc = LNetSetLazyPortal(service->srv_req_portal);
838 mutex_lock(&ptlrpc_all_services_mutex);
839 list_add(&service->srv_list, &ptlrpc_all_services);
840 mutex_unlock(&ptlrpc_all_services_mutex);
843 rc = ptlrpc_sysfs_register_service(parent, service);
848 if (debugfs_entry != NULL)
849 ptlrpc_ldebugfs_register_service(debugfs_entry, service);
851 rc = ptlrpc_service_nrs_setup(service);
855 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
856 service->srv_name, service->srv_req_portal);
858 rc = ptlrpc_start_threads(service);
860 CERROR("Failed to start threads for service %s: %d\n",
861 service->srv_name, rc);
867 ptlrpc_unregister_service(service);
870 EXPORT_SYMBOL(ptlrpc_register_service);
873 * to actually free the request, must be called without holding svc_lock.
874 * note it's caller's responsibility to unlink req->rq_list.
876 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
878 LASSERT(atomic_read(&req->rq_refcount) == 0);
879 LASSERT(list_empty(&req->rq_timed_list));
882 * DEBUG_REQ() assumes the reply state of a request with a valid
883 * ref will not be destroyed until that reference is dropped.
885 ptlrpc_req_drop_rs(req);
887 sptlrpc_svc_ctx_decref(req);
889 if (req != &req->rq_rqbd->rqbd_req) {
891 * NB request buffers use an embedded
892 * req if the incoming req unlinked the
893 * MD; this isn't one of them!
895 ptlrpc_request_cache_free(req);
900 * drop a reference count of the request. if it reaches 0, we either
901 * put it into history list, or free it immediately.
903 void ptlrpc_server_drop_request(struct ptlrpc_request *req)
905 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
906 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
907 struct ptlrpc_service *svc = svcpt->scp_service;
909 struct list_head *tmp;
910 struct list_head *nxt;
912 if (!atomic_dec_and_test(&req->rq_refcount))
915 if (req->rq_session.lc_state == LCS_ENTERED) {
916 lu_context_exit(&req->rq_session);
917 lu_context_fini(&req->rq_session);
920 if (req->rq_at_linked) {
921 spin_lock(&svcpt->scp_at_lock);
923 * recheck with lock, in case it's unlinked by
924 * ptlrpc_at_check_timed()
926 if (likely(req->rq_at_linked))
927 ptlrpc_at_remove_timed(req);
928 spin_unlock(&svcpt->scp_at_lock);
931 LASSERT(list_empty(&req->rq_timed_list));
933 /* finalize request */
934 if (req->rq_export) {
935 class_export_put(req->rq_export);
936 req->rq_export = NULL;
939 spin_lock(&svcpt->scp_lock);
941 list_add(&req->rq_list, &rqbd->rqbd_reqs);
943 refcount = --(rqbd->rqbd_refcount);
945 /* request buffer is now idle: add to history */
946 list_move_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds);
947 svcpt->scp_hist_nrqbds++;
951 * I expect only about 1 or 2 rqbds need to be recycled here
953 while (svcpt->scp_hist_nrqbds > svc->srv_hist_nrqbds_cpt_max) {
954 rqbd = list_entry(svcpt->scp_hist_rqbds.next,
955 struct ptlrpc_request_buffer_desc,
958 list_del(&rqbd->rqbd_list);
959 svcpt->scp_hist_nrqbds--;
962 * remove rqbd's reqs from svc's req history while
963 * I've got the service lock
965 list_for_each(tmp, &rqbd->rqbd_reqs) {
966 req = list_entry(tmp, struct ptlrpc_request,
968 /* Track the highest culled req seq */
969 if (req->rq_history_seq >
970 svcpt->scp_hist_seq_culled) {
971 svcpt->scp_hist_seq_culled =
974 list_del(&req->rq_history_list);
977 spin_unlock(&svcpt->scp_lock);
979 list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
980 req = list_entry(rqbd->rqbd_reqs.next,
981 struct ptlrpc_request,
983 list_del(&req->rq_list);
984 ptlrpc_server_free_request(req);
987 spin_lock(&svcpt->scp_lock);
989 * now all reqs including the embedded req has been
990 * disposed, schedule request buffer for re-use
991 * or free it to drain some in excess.
993 LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) == 0);
994 if (svcpt->scp_nrqbds_posted >=
995 svc->srv_nbuf_per_group ||
996 (svc->srv_nrqbds_max != 0 &&
997 svcpt->scp_nrqbds_total > svc->srv_nrqbds_max) ||
998 test_req_buffer_pressure) {
999 /* like in ptlrpc_free_rqbd() */
1000 svcpt->scp_nrqbds_total--;
1001 OBD_FREE_LARGE(rqbd->rqbd_buffer,
1005 list_add_tail(&rqbd->rqbd_list,
1006 &svcpt->scp_rqbd_idle);
1010 spin_unlock(&svcpt->scp_lock);
1011 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
1012 /* If we are low on memory, we are not interested in history */
1013 list_del(&req->rq_list);
1014 list_del_init(&req->rq_history_list);
1016 /* Track the highest culled req seq */
1017 if (req->rq_history_seq > svcpt->scp_hist_seq_culled)
1018 svcpt->scp_hist_seq_culled = req->rq_history_seq;
1020 spin_unlock(&svcpt->scp_lock);
1022 ptlrpc_server_free_request(req);
1024 spin_unlock(&svcpt->scp_lock);
1028 static void ptlrpc_add_exp_list_nolock(struct ptlrpc_request *req,
1029 struct obd_export *export, bool hp)
1031 __u16 tag = lustre_msg_get_tag(req->rq_reqmsg);
1034 list_add(&req->rq_exp_list, &export->exp_hp_rpcs);
1036 list_add(&req->rq_exp_list, &export->exp_reg_rpcs);
1037 if (tag && export->exp_used_slots)
1038 set_bit(tag - 1, export->exp_used_slots);
1041 static void ptlrpc_del_exp_list(struct ptlrpc_request *req)
1043 __u16 tag = lustre_msg_get_tag(req->rq_reqmsg);
1045 spin_lock(&req->rq_export->exp_rpc_lock);
1046 list_del_init(&req->rq_exp_list);
1047 if (tag && !req->rq_obsolete && req->rq_export->exp_used_slots)
1048 clear_bit(tag - 1, req->rq_export->exp_used_slots);
1049 spin_unlock(&req->rq_export->exp_rpc_lock);
1052 /** Change request export and move hp request from old export to new */
1053 void ptlrpc_request_change_export(struct ptlrpc_request *req,
1054 struct obd_export *export)
1056 if (req->rq_export != NULL) {
1057 LASSERT(!list_empty(&req->rq_exp_list));
1058 /* remove rq_exp_list from last export */
1059 ptlrpc_del_exp_list(req);
1060 /* export has one reference already, so it's safe to
1061 * add req to export queue here and get another
1062 * reference for request later
1064 spin_lock(&export->exp_rpc_lock);
1065 ptlrpc_add_exp_list_nolock(req, export, req->rq_ops != NULL);
1066 spin_unlock(&export->exp_rpc_lock);
1068 class_export_rpc_dec(req->rq_export);
1069 class_export_put(req->rq_export);
1072 /* request takes one export refcount */
1073 req->rq_export = class_export_get(export);
1074 class_export_rpc_inc(export);
1080 * to finish a request: stop sending more early replies, and release
1083 static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt,
1084 struct ptlrpc_request *req)
1086 ptlrpc_server_hpreq_fini(req);
1088 ptlrpc_server_drop_request(req);
1092 * to finish an active request: stop sending more early replies, and release
1093 * the request. should be called after we finished handling the request.
1095 static void ptlrpc_server_finish_active_request(
1096 struct ptlrpc_service_part *svcpt,
1097 struct ptlrpc_request *req)
1099 spin_lock(&svcpt->scp_req_lock);
1100 ptlrpc_nrs_req_stop_nolock(req);
1101 svcpt->scp_nreqs_active--;
1103 svcpt->scp_nhreqs_active--;
1104 spin_unlock(&svcpt->scp_req_lock);
1106 ptlrpc_nrs_req_finalize(req);
1108 if (req->rq_export != NULL)
1109 class_export_rpc_dec(req->rq_export);
1111 ptlrpc_server_finish_request(svcpt, req);
1115 * This function makes sure dead exports are evicted in a timely manner.
1116 * This function is only called when some export receives a message (i.e.,
1117 * the network is up.)
1119 void ptlrpc_update_export_timer(struct obd_export *exp, time64_t extra_delay)
1121 struct obd_export *oldest_exp;
1122 time64_t oldest_time, new_time;
1129 * Compensate for slow machines, etc, by faking our request time
1130 * into the future. Although this can break the strict time-ordering
1131 * of the list, we can be really lazy here - we don't have to evict
1132 * at the exact right moment. Eventually, all silent exports
1133 * will make it to the top of the list.
1136 /* Do not pay attention on 1sec or smaller renewals. */
1137 new_time = ktime_get_real_seconds() + extra_delay;
1138 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
1141 exp->exp_last_request_time = new_time;
1144 * exports may get disconnected from the chain even though the
1145 * export has references, so we must keep the spin lock while
1146 * manipulating the lists
1148 spin_lock(&exp->exp_obd->obd_dev_lock);
1150 if (list_empty(&exp->exp_obd_chain_timed)) {
1151 /* this one is not timed */
1152 spin_unlock(&exp->exp_obd->obd_dev_lock);
1156 list_move_tail(&exp->exp_obd_chain_timed,
1157 &exp->exp_obd->obd_exports_timed);
1159 oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
1160 struct obd_export, exp_obd_chain_timed);
1161 oldest_time = oldest_exp->exp_last_request_time;
1162 spin_unlock(&exp->exp_obd->obd_dev_lock);
1164 if (exp->exp_obd->obd_recovering) {
1165 /* be nice to everyone during recovery */
1170 /* Note - racing to start/reset the obd_eviction timer is safe */
1171 if (exp->exp_obd->obd_eviction_timer == 0) {
1172 /* Check if the oldest entry is expired. */
1173 if (ktime_get_real_seconds() >
1174 oldest_time + PING_EVICT_TIMEOUT + extra_delay) {
1176 * We need a second timer, in case the net was down and
1177 * it just came back. Since the pinger may skip every
1178 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
1179 * we better wait for 3.
1181 exp->exp_obd->obd_eviction_timer =
1182 ktime_get_real_seconds() + 3 * PING_INTERVAL;
1183 CDEBUG(D_HA, "%s: Think about evicting %s from %lld\n",
1184 exp->exp_obd->obd_name,
1185 obd_export_nid2str(oldest_exp), oldest_time);
1188 if (ktime_get_real_seconds() >
1189 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
1191 * The evictor won't evict anyone who we've heard from
1192 * recently, so we don't have to check before we start
1195 if (!ping_evictor_wake(exp))
1196 exp->exp_obd->obd_eviction_timer = 0;
1204 * Sanity check request \a req.
1205 * Return 0 if all is ok, error code otherwise.
1207 static int ptlrpc_check_req(struct ptlrpc_request *req)
1209 struct obd_device *obd = req->rq_export->exp_obd;
1212 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
1213 req->rq_export->exp_conn_cnt)) {
1214 DEBUG_REQ(D_RPCTRACE, req,
1215 "DROPPING req from old connection %d < %d",
1216 lustre_msg_get_conn_cnt(req->rq_reqmsg),
1217 req->rq_export->exp_conn_cnt);
1220 if (unlikely(obd == NULL || obd->obd_fail)) {
1222 * Failing over, don't handle any more reqs,
1223 * send error response instead.
1225 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
1226 req, (obd != NULL) ? obd->obd_name : "unknown");
1228 } else if (lustre_msg_get_flags(req->rq_reqmsg) &
1229 (MSG_REPLAY | MSG_REQ_REPLAY_DONE) &&
1230 !obd->obd_recovering) {
1231 DEBUG_REQ(D_ERROR, req,
1232 "Invalid replay without recovery");
1233 class_fail_export(req->rq_export);
1235 } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 &&
1236 !obd->obd_recovering) {
1237 DEBUG_REQ(D_ERROR, req,
1238 "Invalid req with transno %llu without recovery",
1239 lustre_msg_get_transno(req->rq_reqmsg));
1240 class_fail_export(req->rq_export);
1244 if (unlikely(rc < 0)) {
1245 req->rq_status = rc;
1251 static void ptlrpc_at_set_timer(struct ptlrpc_service_part *svcpt)
1253 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1256 if (array->paa_count == 0) {
1257 del_timer(&svcpt->scp_at_timer);
1261 /* Set timer for closest deadline */
1262 next = array->paa_deadline - ktime_get_real_seconds() -
1265 ptlrpc_at_timer(cfs_timer_cb_arg(svcpt, scp_at_timer));
1267 mod_timer(&svcpt->scp_at_timer,
1268 jiffies + nsecs_to_jiffies(next * NSEC_PER_SEC));
1269 CDEBUG(D_INFO, "armed %s at %+llds\n",
1270 svcpt->scp_service->srv_name, next);
1274 /* Add rpc to early reply check list */
1275 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
1277 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1278 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1279 struct ptlrpc_request *rq = NULL;
1285 if (req->rq_no_reply)
1288 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
1291 spin_lock(&svcpt->scp_at_lock);
1292 LASSERT(list_empty(&req->rq_timed_list));
1294 div_u64_rem(req->rq_deadline, array->paa_size, &index);
1295 if (array->paa_reqs_count[index] > 0) {
1297 * latest rpcs will have the latest deadlines in the list,
1298 * so search backward.
1300 list_for_each_entry_reverse(rq, &array->paa_reqs_array[index],
1302 if (req->rq_deadline >= rq->rq_deadline) {
1303 list_add(&req->rq_timed_list,
1304 &rq->rq_timed_list);
1310 /* Add the request at the head of the list */
1311 if (list_empty(&req->rq_timed_list))
1312 list_add(&req->rq_timed_list, &array->paa_reqs_array[index]);
1314 spin_lock(&req->rq_lock);
1315 req->rq_at_linked = 1;
1316 spin_unlock(&req->rq_lock);
1317 req->rq_at_index = index;
1318 array->paa_reqs_count[index]++;
1320 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
1321 array->paa_deadline = req->rq_deadline;
1322 ptlrpc_at_set_timer(svcpt);
1324 spin_unlock(&svcpt->scp_at_lock);
1329 static void ptlrpc_at_remove_timed(struct ptlrpc_request *req)
1331 struct ptlrpc_at_array *array;
1333 array = &req->rq_rqbd->rqbd_svcpt->scp_at_array;
1335 /* NB: must call with hold svcpt::scp_at_lock */
1336 LASSERT(!list_empty(&req->rq_timed_list));
1337 list_del_init(&req->rq_timed_list);
1339 spin_lock(&req->rq_lock);
1340 req->rq_at_linked = 0;
1341 spin_unlock(&req->rq_lock);
1343 array->paa_reqs_count[req->rq_at_index]--;
1348 * Attempt to extend the request deadline by sending an early reply to the
1351 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
1353 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1354 struct ptlrpc_request *reqcopy;
1355 struct lustre_msg *reqmsg;
1356 time64_t olddl = req->rq_deadline - ktime_get_real_seconds();
1362 if (CFS_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_RECONNECT)) {
1363 /* don't send early reply */
1368 * deadline is when the client expects us to reply, margin is the
1369 * difference between clients' and servers' expectations
1371 DEBUG_REQ(D_ADAPTTO, req,
1372 "%ssending early reply (deadline %+llds, margin %+llds) for %d+%d",
1373 AT_OFF ? "AT off - not " : "",
1374 (s64)olddl, (s64)(olddl - at_get(&svcpt->scp_at_estimate)),
1375 at_get(&svcpt->scp_at_estimate), at_extra);
1381 /* below message is checked in replay-ost-single.sh test_9 */
1382 DEBUG_REQ(D_WARNING, req,
1383 "Already past deadline (%+llds), not sending early reply. Consider increasing at_early_margin (%d)?",
1384 (s64)olddl, at_early_margin);
1386 /* Return an error so we're not re-added to the timed list. */
1390 if ((lustre_msghdr_get_flags(req->rq_reqmsg) &
1391 MSGHDR_AT_SUPPORT) == 0) {
1392 DEBUG_REQ(D_INFO, req,
1393 "Wanted to ask client for more time, but no AT support");
1397 if (req->rq_export &&
1398 lustre_msg_get_flags(req->rq_reqmsg) &
1399 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1400 struct obd_device *obd_exp = req->rq_export->exp_obd;
1403 * During recovery, we don't want to send too many early
1404 * replies, but on the other hand we want to make sure the
1405 * client has enough time to resend if the rpc is lost. So
1406 * during the recovery period send at least 4 early replies,
1407 * spacing them every at_extra if we can. at_estimate should
1408 * always equal this fixed value during recovery.
1412 * Don't account request processing time into AT history
1413 * during recovery, it is not service time we need but
1414 * includes also waiting time for recovering clients
1416 newdl = min_t(time64_t, at_extra,
1417 obd_exp->obd_recovery_timeout / 4) +
1418 ktime_get_real_seconds();
1421 * We want to extend the request deadline by at_extra seconds,
1422 * so we set our service estimate to reflect how much time has
1423 * passed since this request arrived plus an additional
1424 * at_extra seconds. The client will calculate the new deadline
1425 * based on this service estimate (plus some additional time to
1426 * account for network latency). See ptlrpc_at_recv_early_reply
1428 at_measured(&svcpt->scp_at_estimate, at_extra +
1429 ktime_get_real_seconds() -
1430 req->rq_arrival_time.tv_sec);
1431 newdl = req->rq_arrival_time.tv_sec +
1432 at_get(&svcpt->scp_at_estimate);
1436 * Check to see if we've actually increased the deadline -
1437 * we may be past adaptive_max
1439 if (req->rq_deadline >= newdl) {
1440 DEBUG_REQ(D_WARNING, req,
1441 "Could not add any time (%lld/%lld), not sending early reply",
1442 (s64)olddl, (s64)(newdl - ktime_get_real_seconds()));
1446 reqcopy = ptlrpc_request_cache_alloc(GFP_NOFS);
1447 if (reqcopy == NULL)
1449 OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1451 GOTO(out_free, rc = -ENOMEM);
1454 reqcopy->rq_reply_state = NULL;
1455 reqcopy->rq_rep_swab_mask = 0;
1456 reqcopy->rq_pack_bulk = 0;
1457 reqcopy->rq_pack_udesc = 0;
1458 reqcopy->rq_packed_final = 0;
1459 sptlrpc_svc_ctx_addref(reqcopy);
1460 /* We only need the reqmsg for the magic */
1461 reqcopy->rq_reqmsg = reqmsg;
1462 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1465 * tgt_brw_read() and tgt_brw_write() may have decided not to reply.
1466 * Without this check, we would fail the rq_no_reply assertion in
1467 * ptlrpc_send_reply().
1469 if (reqcopy->rq_no_reply)
1470 GOTO(out, rc = -ETIMEDOUT);
1472 LASSERT(atomic_read(&req->rq_refcount));
1473 /* if it is last refcount then early reply isn't needed */
1474 if (atomic_read(&req->rq_refcount) == 1) {
1475 DEBUG_REQ(D_ADAPTTO, reqcopy,
1476 "Normal reply already sent, abort early reply");
1477 GOTO(out, rc = -EINVAL);
1480 /* Connection ref */
1481 reqcopy->rq_export = class_conn2export(
1482 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1483 if (reqcopy->rq_export == NULL)
1484 GOTO(out, rc = -ENODEV);
1487 class_export_rpc_inc(reqcopy->rq_export);
1488 if (reqcopy->rq_export->exp_obd &&
1489 reqcopy->rq_export->exp_obd->obd_fail)
1490 GOTO(out_put, rc = -ENODEV);
1492 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1496 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1499 /* Adjust our own deadline to what we told the client */
1500 req->rq_deadline = newdl;
1501 req->rq_early_count++; /* number sent, server side */
1503 DEBUG_REQ(D_ERROR, req, "Early reply send failed: rc = %d", rc);
1507 * Free the (early) reply state from lustre_pack_reply.
1508 * (ptlrpc_send_reply takes it's own rs ref, so this is safe here)
1510 ptlrpc_req_drop_rs(reqcopy);
1513 class_export_rpc_dec(reqcopy->rq_export);
1514 class_export_put(reqcopy->rq_export);
1516 sptlrpc_svc_ctx_decref(reqcopy);
1517 OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1519 ptlrpc_request_cache_free(reqcopy);
1524 * Send early replies to everybody expiring within at_early_margin
1525 * asking for at_extra time
1527 static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1529 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1530 struct ptlrpc_request *rq, *n;
1531 struct list_head work_list;
1534 time64_t now = ktime_get_real_seconds();
1536 int first, counter = 0;
1539 spin_lock(&svcpt->scp_at_lock);
1540 if (svcpt->scp_at_check == 0) {
1541 spin_unlock(&svcpt->scp_at_lock);
1544 delay = ktime_ms_delta(ktime_get(), svcpt->scp_at_checktime);
1545 svcpt->scp_at_check = 0;
1547 if (array->paa_count == 0) {
1548 spin_unlock(&svcpt->scp_at_lock);
1552 /* The timer went off, but maybe the nearest rpc already completed. */
1553 first = array->paa_deadline - now;
1554 if (first > at_early_margin) {
1555 /* We've still got plenty of time. Reset the timer. */
1556 ptlrpc_at_set_timer(svcpt);
1557 spin_unlock(&svcpt->scp_at_lock);
1562 * We're close to a timeout, and we don't know how much longer the
1563 * server will take. Send early replies to everyone expiring soon.
1565 INIT_LIST_HEAD(&work_list);
1567 div_u64_rem(array->paa_deadline, array->paa_size, &index);
1568 count = array->paa_count;
1570 count -= array->paa_reqs_count[index];
1571 list_for_each_entry_safe(rq, n,
1572 &array->paa_reqs_array[index],
1574 if (rq->rq_deadline > now + at_early_margin) {
1575 /* update the earliest deadline */
1576 if (deadline == -1 ||
1577 rq->rq_deadline < deadline)
1578 deadline = rq->rq_deadline;
1583 * ptlrpc_server_drop_request() may drop
1584 * refcount to 0 already. Let's check this and
1585 * don't add entry to work_list
1587 if (likely(atomic_inc_not_zero(&rq->rq_refcount))) {
1588 ptlrpc_at_remove_timed(rq);
1589 list_add(&rq->rq_timed_list, &work_list);
1591 ptlrpc_at_remove_timed(rq);
1597 if (++index >= array->paa_size)
1600 array->paa_deadline = deadline;
1601 /* we have a new earliest deadline, restart the timer */
1602 ptlrpc_at_set_timer(svcpt);
1604 spin_unlock(&svcpt->scp_at_lock);
1607 "timeout in %+ds, asking for %d secs on %d early replies\n",
1608 first, at_extra, counter);
1611 * We're already past request deadlines before we even get a
1612 * chance to send early replies
1614 LCONSOLE_WARN("%s: This server is not able to keep up with request traffic (cpu-bound).\n",
1615 svcpt->scp_service->srv_name);
1616 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, delay=%lld\n",
1617 counter, svcpt->scp_nreqs_incoming,
1618 svcpt->scp_nreqs_active,
1619 at_get(&svcpt->scp_at_estimate), delay);
1623 * we took additional refcount so entries can't be deleted from list, no
1626 while (!list_empty(&work_list)) {
1627 rq = list_entry(work_list.next, struct ptlrpc_request,
1629 list_del_init(&rq->rq_timed_list);
1631 if (ptlrpc_at_send_early_reply(rq) == 0)
1632 ptlrpc_at_add_timed(rq);
1634 ptlrpc_server_drop_request(rq);
1637 RETURN(1); /* return "did_something" for liblustre */
1641 * Check if we are already handling earlier incarnation of this request.
1642 * Called under &req->rq_export->exp_rpc_lock locked
1644 static struct ptlrpc_request*
1645 ptlrpc_server_check_resend_in_progress(struct ptlrpc_request *req)
1647 struct ptlrpc_request *tmp = NULL;
1649 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ||
1650 (atomic_read(&req->rq_export->exp_rpc_count) == 0))
1654 * bulk request are aborted upon reconnect, don't try to
1657 if (req->rq_bulk_write || req->rq_bulk_read)
1661 * This list should not be longer than max_requests in
1662 * flights on the client, so it is not all that long.
1663 * Also we only hit this codepath in case of a resent
1664 * request which makes it even more rarely hit
1666 list_for_each_entry(tmp, &req->rq_export->exp_reg_rpcs,
1668 /* Found duplicate one */
1669 if (tmp->rq_xid == req->rq_xid)
1672 list_for_each_entry(tmp, &req->rq_export->exp_hp_rpcs,
1674 /* Found duplicate one */
1675 if (tmp->rq_xid == req->rq_xid)
1681 DEBUG_REQ(D_HA, req, "Found duplicate req in processing");
1682 DEBUG_REQ(D_HA, tmp, "Request being processed");
1686 #ifdef HAVE_SERVER_SUPPORT
1687 static void ptlrpc_server_mark_obsolete(struct ptlrpc_request *req)
1689 req->rq_obsolete = 1;
1693 ptlrpc_server_mark_in_progress_obsolete(struct ptlrpc_request *req)
1695 struct ptlrpc_request *tmp = NULL;
1698 if (!tgt_is_increasing_xid_client(req->rq_export) ||
1699 req->rq_export->exp_used_slots == NULL)
1702 tag = lustre_msg_get_tag(req->rq_reqmsg);
1706 if (!test_bit(tag - 1, req->rq_export->exp_used_slots))
1709 /* This list should not be longer than max_requests in
1710 * flights on the client, so it is not all that long.
1711 * Also we only hit this codepath in case of a resent
1712 * request which makes it even more rarely hit */
1713 list_for_each_entry(tmp, &req->rq_export->exp_reg_rpcs, rq_exp_list) {
1714 if (tag == lustre_msg_get_tag(tmp->rq_reqmsg) &&
1715 req->rq_xid > tmp->rq_xid)
1716 ptlrpc_server_mark_obsolete(tmp);
1719 list_for_each_entry(tmp, &req->rq_export->exp_hp_rpcs, rq_exp_list) {
1720 if (tag == lustre_msg_get_tag(tmp->rq_reqmsg) &&
1721 req->rq_xid > tmp->rq_xid)
1722 ptlrpc_server_mark_obsolete(tmp);
1728 * Check if a request should be assigned with a high priority.
1730 * \retval < 0: error occurred
1731 * 0: normal RPC request
1732 * +1: high priority request
1734 static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt,
1735 struct ptlrpc_request *req)
1740 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL) {
1741 rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
1748 if (req->rq_export != NULL && req->rq_ops != NULL) {
1750 * Perform request specific check. We should do this
1751 * check before the request is added into exp_hp_rpcs
1752 * list otherwise it may hit swab race at LU-1044.
1754 if (req->rq_ops->hpreq_check != NULL) {
1755 rc = req->rq_ops->hpreq_check(req);
1756 if (rc == -ESTALE) {
1757 req->rq_status = rc;
1761 * can only return error,
1762 * 0 for normal request,
1763 * or 1 for high priority request
1772 /** Remove the request from the export list. */
1773 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req)
1776 if (req->rq_export) {
1778 * refresh lock timeout again so that client has more
1779 * room to send lock cancel RPC.
1781 if (req->rq_ops && req->rq_ops->hpreq_fini)
1782 req->rq_ops->hpreq_fini(req);
1784 ptlrpc_del_exp_list(req);
1789 static int ptlrpc_hpreq_check(struct ptlrpc_request *req)
1794 static struct ptlrpc_hpreq_ops ptlrpc_hpreq_common = {
1795 .hpreq_check = ptlrpc_hpreq_check,
1798 /* Hi-Priority RPC check by RPC operation code. */
1799 int ptlrpc_hpreq_handler(struct ptlrpc_request *req)
1801 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1804 * Check for export to let only reconnects for not yet evicted
1805 * export to become a HP rpc.
1807 if ((req->rq_export != NULL) &&
1808 (opc == OBD_PING || opc == MDS_CONNECT || opc == OST_CONNECT))
1809 req->rq_ops = &ptlrpc_hpreq_common;
1813 EXPORT_SYMBOL(ptlrpc_hpreq_handler);
1815 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1816 struct ptlrpc_request *req)
1820 struct ptlrpc_request *orig;
1824 rc = ptlrpc_server_hpreq_init(svcpt, req);
1829 ptlrpc_nrs_req_initialize(svcpt, req, hp);
1831 while (req->rq_export != NULL) {
1832 struct obd_export *exp = req->rq_export;
1835 * do search for duplicated xid and the adding to the list
1838 spin_lock_bh(&exp->exp_rpc_lock);
1839 #ifdef HAVE_SERVER_SUPPORT
1840 ptlrpc_server_mark_in_progress_obsolete(req);
1842 orig = ptlrpc_server_check_resend_in_progress(req);
1843 if (orig && OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_RESEND_RACE)) {
1844 spin_unlock_bh(&exp->exp_rpc_lock);
1846 OBD_RACE(OBD_FAIL_PTLRPC_RESEND_RACE);
1847 msleep(4 * MSEC_PER_SEC);
1851 if (orig && likely(atomic_inc_not_zero(&orig->rq_refcount))) {
1854 spin_unlock_bh(&exp->exp_rpc_lock);
1857 * When the client resend request and the server has
1858 * the previous copy of it, we need to update deadlines,
1859 * to be sure that the client and the server have equal
1860 * request deadlines.
1863 spin_lock(&orig->rq_rqbd->rqbd_svcpt->scp_at_lock);
1864 linked = orig->rq_at_linked;
1866 ptlrpc_at_remove_timed(orig);
1867 spin_unlock(&orig->rq_rqbd->rqbd_svcpt->scp_at_lock);
1868 orig->rq_deadline = req->rq_deadline;
1870 ptlrpc_at_add_timed(orig);
1871 ptlrpc_server_drop_request(orig);
1872 ptlrpc_nrs_req_finalize(req);
1874 /* don't mark slot unused for resend in progress */
1875 req->rq_obsolete = 1;
1880 ptlrpc_add_exp_list_nolock(req, exp, hp || req->rq_ops != NULL);
1882 spin_unlock_bh(&exp->exp_rpc_lock);
1887 * the current thread is not the processing thread for this request
1888 * since that, but request is in exp_hp_list and can be find there.
1889 * Remove all relations between request and old thread.
1891 req->rq_svc_thread->t_env->le_ses = NULL;
1892 req->rq_svc_thread = NULL;
1893 req->rq_session.lc_thread = NULL;
1895 ptlrpc_nrs_req_add(svcpt, req, hp);
1901 * Allow to handle high priority request
1902 * User can call it w/o any lock but need to hold
1903 * ptlrpc_service_part::scp_req_lock to get reliable result
1905 static bool ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1908 int running = svcpt->scp_nthrs_running;
1910 if (!nrs_svcpt_has_hp(svcpt))
1916 if (ptlrpc_nrs_req_throttling_nolock(svcpt, true))
1919 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1920 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1921 /* leave just 1 thread for normal RPCs */
1922 running = PTLRPC_NTHRS_INIT;
1923 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1927 if (svcpt->scp_nreqs_active >= running - 1)
1930 if (svcpt->scp_nhreqs_active == 0)
1933 return !ptlrpc_nrs_req_pending_nolock(svcpt, false) ||
1934 svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1937 static bool ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1940 return ptlrpc_server_allow_high(svcpt, force) &&
1941 ptlrpc_nrs_req_pending_nolock(svcpt, true);
1945 * Only allow normal priority requests on a service that has a high-priority
1946 * queue if forced (i.e. cleanup), if there are other high priority requests
1947 * already being processed (i.e. those threads can service more high-priority
1948 * requests), or if there are enough idle threads that a later thread can do
1949 * a high priority request.
1950 * User can call it w/o any lock but need to hold
1951 * ptlrpc_service_part::scp_req_lock to get reliable result
1953 static bool ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1956 int running = svcpt->scp_nthrs_running;
1958 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1959 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1960 /* leave just 1 thread for normal RPCs */
1961 running = PTLRPC_NTHRS_INIT;
1962 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1969 if (ptlrpc_nrs_req_throttling_nolock(svcpt, false))
1972 if (svcpt->scp_nreqs_active < running - 2)
1975 if (svcpt->scp_nreqs_active >= running - 1)
1978 return svcpt->scp_nhreqs_active > 0 || !nrs_svcpt_has_hp(svcpt);
1981 static bool ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1984 return ptlrpc_server_allow_normal(svcpt, force) &&
1985 ptlrpc_nrs_req_pending_nolock(svcpt, false);
1989 * Returns true if there are requests available in incoming
1990 * request queue for processing and it is allowed to fetch them.
1991 * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1992 * to get reliable result
1993 * \see ptlrpc_server_allow_normal
1994 * \see ptlrpc_server_allow high
1997 bool ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt,
2000 return ptlrpc_server_high_pending(svcpt, force) ||
2001 ptlrpc_server_normal_pending(svcpt, force);
2005 * Fetch a request for processing from queue of unprocessed requests.
2006 * Favors high-priority requests.
2007 * Returns a pointer to fetched request.
2009 static struct ptlrpc_request *
2010 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force)
2012 struct ptlrpc_request *req = NULL;
2016 spin_lock(&svcpt->scp_req_lock);
2018 if (ptlrpc_server_high_pending(svcpt, force)) {
2019 req = ptlrpc_nrs_req_get_nolock(svcpt, true, force);
2021 svcpt->scp_hreq_count++;
2026 if (ptlrpc_server_normal_pending(svcpt, force)) {
2027 req = ptlrpc_nrs_req_get_nolock(svcpt, false, force);
2029 svcpt->scp_hreq_count = 0;
2034 spin_unlock(&svcpt->scp_req_lock);
2038 svcpt->scp_nreqs_active++;
2040 svcpt->scp_nhreqs_active++;
2042 spin_unlock(&svcpt->scp_req_lock);
2044 if (likely(req->rq_export))
2045 class_export_rpc_inc(req->rq_export);
2051 * Handle freshly incoming reqs, add to timed early reply list,
2052 * pass on to regular request queue.
2053 * All incoming requests pass through here before getting into
2054 * ptlrpc_server_handle_req later on.
2056 static int ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt,
2057 struct ptlrpc_thread *thread)
2059 struct ptlrpc_service *svc = svcpt->scp_service;
2060 struct ptlrpc_request *req;
2066 spin_lock(&svcpt->scp_lock);
2067 if (list_empty(&svcpt->scp_req_incoming)) {
2068 spin_unlock(&svcpt->scp_lock);
2072 req = list_entry(svcpt->scp_req_incoming.next,
2073 struct ptlrpc_request, rq_list);
2074 list_del_init(&req->rq_list);
2075 svcpt->scp_nreqs_incoming--;
2077 * Consider this still a "queued" request as far as stats are
2080 spin_unlock(&svcpt->scp_lock);
2082 /* go through security check/transform */
2083 rc = sptlrpc_svc_unwrap_request(req);
2087 case SECSVC_COMPLETE:
2088 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
2097 * for null-flavored rpc, msg has been unpacked by sptlrpc, although
2098 * redo it wouldn't be harmful.
2100 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
2101 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
2103 CERROR("error unpacking request: ptl %d from %s x%llu\n",
2104 svc->srv_req_portal, libcfs_id2str(req->rq_peer),
2110 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
2112 CERROR("error unpacking ptlrpc body: ptl %d from %s x %llu\n",
2113 svc->srv_req_portal, libcfs_id2str(req->rq_peer),
2118 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
2119 lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
2120 CERROR("drop incoming rpc opc %u, x%llu\n",
2121 cfs_fail_val, req->rq_xid);
2126 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
2127 CERROR("wrong packet type received (type=%u) from %s\n",
2128 lustre_msg_get_type(req->rq_reqmsg),
2129 libcfs_id2str(req->rq_peer));
2133 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2137 req->rq_bulk_write = 1;
2141 case MGS_CONFIG_READ:
2142 req->rq_bulk_read = 1;
2146 CDEBUG(D_RPCTRACE, "got req x%llu\n", req->rq_xid);
2148 req->rq_export = class_conn2export(
2149 lustre_msg_get_handle(req->rq_reqmsg));
2150 if (req->rq_export) {
2151 rc = ptlrpc_check_req(req);
2153 rc = sptlrpc_target_export_check(req->rq_export, req);
2155 DEBUG_REQ(D_ERROR, req,
2156 "DROPPING req with illegal security flavor");
2161 ptlrpc_update_export_timer(req->rq_export, 0);
2164 /* req_in handling should/must be fast */
2165 if (ktime_get_real_seconds() - req->rq_arrival_time.tv_sec > 5)
2166 DEBUG_REQ(D_WARNING, req, "Slow req_in handling %llds",
2167 (s64)(ktime_get_real_seconds() -
2168 req->rq_arrival_time.tv_sec));
2170 /* Set rpc server deadline and add it to the timed list */
2171 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
2172 MSGHDR_AT_SUPPORT) ?
2173 /* The max time the client expects us to take */
2174 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
2176 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
2177 if (unlikely(deadline == 0)) {
2178 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
2182 /* Skip early reply */
2183 if (OBD_FAIL_PRECHECK(OBD_FAIL_MDS_RESEND))
2184 req->rq_deadline += obd_timeout;
2186 req->rq_svc_thread = thread;
2187 if (thread != NULL) {
2189 * initialize request session, it is needed for request
2190 * processing by target
2192 rc = lu_context_init(&req->rq_session, LCT_SERVER_SESSION |
2195 CERROR("%s: failure to initialize session: rc = %d\n",
2196 thread->t_name, rc);
2199 req->rq_session.lc_thread = thread;
2200 lu_context_enter(&req->rq_session);
2201 thread->t_env->le_ses = &req->rq_session;
2204 ptlrpc_at_add_timed(req);
2206 /* Move it over to the request processing queue */
2207 rc = ptlrpc_server_request_add(svcpt, req);
2211 wake_up(&svcpt->scp_waitq);
2215 ptlrpc_server_finish_request(svcpt, req);
2221 * Main incoming request handling logic.
2222 * Calls handler function from service to do actual processing.
2224 static int ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
2225 struct ptlrpc_thread *thread)
2227 struct ptlrpc_service *svc = svcpt->scp_service;
2228 struct ptlrpc_request *request;
2238 request = ptlrpc_server_request_get(svcpt, false);
2239 if (request == NULL)
2242 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
2243 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
2244 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
2245 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
2247 if (unlikely(fail_opc)) {
2248 if (request->rq_export && request->rq_ops)
2249 OBD_FAIL_TIMEOUT(fail_opc, 4);
2252 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
2254 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
2255 libcfs_debug_dumplog();
2257 work_start = ktime_get_real();
2258 arrived = timespec64_to_ktime(request->rq_arrival_time);
2259 timediff_usecs = ktime_us_delta(work_start, arrived);
2260 if (likely(svc->srv_stats != NULL)) {
2261 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
2263 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
2264 svcpt->scp_nreqs_incoming);
2265 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
2266 svcpt->scp_nreqs_active);
2267 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
2268 at_get(&svcpt->scp_at_estimate));
2271 if (likely(request->rq_export)) {
2272 if (unlikely(ptlrpc_check_req(request)))
2274 ptlrpc_update_export_timer(request->rq_export,
2275 div_u64(timediff_usecs,
2280 * Discard requests queued for longer than the deadline.
2281 * The deadline is increased if we send an early reply.
2283 if (ktime_get_real_seconds() > request->rq_deadline) {
2284 DEBUG_REQ(D_ERROR, request,
2285 "Dropping timed-out request from %s: deadline %lld/%llds ago",
2286 libcfs_id2str(request->rq_peer),
2287 request->rq_deadline -
2288 request->rq_arrival_time.tv_sec,
2289 ktime_get_real_seconds() - request->rq_deadline);
2294 "Handling RPC req@%p pname:cluuid+ref:pid:xid:nid:opc:job %s:%s+%d:%d:x%llu:%s:%d:%s\n",
2295 request, current_comm(),
2296 (request->rq_export ?
2297 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2298 (request->rq_export ?
2299 atomic_read(&request->rq_export->exp_refcount) : -99),
2300 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
2301 libcfs_id2str(request->rq_peer),
2302 lustre_msg_get_opc(request->rq_reqmsg),
2303 lustre_msg_get_jobid(request->rq_reqmsg) ?: "");
2305 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
2306 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
2308 CDEBUG(D_NET, "got req %llu\n", request->rq_xid);
2310 /* re-assign request and sesson thread to the current one */
2311 request->rq_svc_thread = thread;
2312 if (thread != NULL) {
2313 LASSERT(request->rq_session.lc_thread == NULL);
2314 request->rq_session.lc_thread = thread;
2315 thread->t_env->le_ses = &request->rq_session;
2317 svc->srv_ops.so_req_handler(request);
2319 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
2322 if (unlikely(ktime_get_real_seconds() > request->rq_deadline)) {
2323 DEBUG_REQ(D_WARNING, request,
2324 "Request took longer than estimated (%lld/%llds); client may timeout",
2325 request->rq_deadline -
2326 request->rq_arrival_time.tv_sec,
2327 ktime_get_real_seconds() - request->rq_deadline);
2330 work_end = ktime_get_real();
2331 timediff_usecs = ktime_us_delta(work_end, work_start);
2332 arrived_usecs = ktime_us_delta(work_end, arrived);
2334 "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",
2335 request, current_comm(),
2336 (request->rq_export ?
2337 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2338 (request->rq_export ?
2339 atomic_read(&request->rq_export->exp_refcount) : -99),
2340 lustre_msg_get_status(request->rq_reqmsg),
2342 libcfs_id2str(request->rq_peer),
2343 lustre_msg_get_opc(request->rq_reqmsg),
2344 lustre_msg_get_jobid(request->rq_reqmsg) ?: "",
2347 (request->rq_repmsg ?
2348 lustre_msg_get_transno(request->rq_repmsg) :
2349 request->rq_transno),
2351 (request->rq_repmsg ?
2352 lustre_msg_get_status(request->rq_repmsg) : -999));
2353 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
2354 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
2355 int opc = opcode_offset(op);
2357 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
2358 LASSERT(opc < LUSTRE_MAX_OPCODES);
2359 lprocfs_counter_add(svc->srv_stats,
2360 opc + EXTRA_MAX_OPCODES,
2364 if (unlikely(request->rq_early_count)) {
2365 DEBUG_REQ(D_ADAPTTO, request,
2366 "sent %d early replies before finishing in %llds",
2367 request->rq_early_count,
2368 div_u64(arrived_usecs, USEC_PER_SEC));
2371 ptlrpc_server_finish_active_request(svcpt, request);
2377 * An internal function to process a single reply state object.
2379 static int ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
2381 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
2382 struct ptlrpc_service *svc = svcpt->scp_service;
2383 struct obd_export *exp;
2389 exp = rs->rs_export;
2391 LASSERT(rs->rs_difficult);
2392 LASSERT(rs->rs_scheduled);
2393 LASSERT(list_empty(&rs->rs_list));
2396 * The disk commit callback holds exp_uncommitted_replies_lock while it
2397 * iterates over newly committed replies, removing them from
2398 * exp_uncommitted_replies. It then drops this lock and schedules the
2399 * replies it found for handling here.
2401 * We can avoid contention for exp_uncommitted_replies_lock between the
2402 * HRT threads and further commit callbacks by checking rs_committed
2403 * which is set in the commit callback while it holds both
2404 * rs_lock and exp_uncommitted_reples.
2406 * If we see rs_committed clear, the commit callback _may_ not have
2407 * handled this reply yet and we race with it to grab
2408 * exp_uncommitted_replies_lock before removing the reply from
2409 * exp_uncommitted_replies. Note that if we lose the race and the
2410 * reply has already been removed, list_del_init() is a noop.
2412 * If we see rs_committed set, we know the commit callback is handling,
2413 * or has handled this reply since store reordering might allow us to
2414 * see rs_committed set out of sequence. But since this is done
2415 * holding rs_lock, we can be sure it has all completed once we hold
2416 * rs_lock, which we do right next.
2418 if (!rs->rs_committed) {
2420 * if rs was commited, no need to convert locks, don't check
2421 * rs_committed here because rs may never be added into
2422 * exp_uncommitted_replies and this flag never be set, see
2423 * target_send_reply()
2425 if (rs->rs_convert_lock &&
2426 rs->rs_transno > exp->exp_last_committed) {
2427 struct ldlm_lock *lock;
2428 struct ldlm_lock *ack_locks[RS_MAX_LOCKS] = { NULL };
2430 spin_lock(&rs->rs_lock);
2431 if (rs->rs_convert_lock &&
2432 rs->rs_transno > exp->exp_last_committed) {
2433 nlocks = rs->rs_nlocks;
2434 while (nlocks-- > 0) {
2436 * NB don't assume rs is always handled
2437 * by the same service thread (see
2438 * ptlrpc_hr_select, so REP-ACK hr may
2439 * race with trans commit, while the
2440 * latter will release locks, get locks
2441 * here early to convert to COS mode
2444 lock = ldlm_handle2lock(
2445 &rs->rs_locks[nlocks]);
2447 ack_locks[nlocks] = lock;
2448 rs->rs_modes[nlocks] = LCK_COS;
2450 nlocks = rs->rs_nlocks;
2451 rs->rs_convert_lock = 0;
2453 * clear rs_scheduled so that commit callback
2454 * can schedule again
2456 rs->rs_scheduled = 0;
2457 spin_unlock(&rs->rs_lock);
2459 while (nlocks-- > 0) {
2460 lock = ack_locks[nlocks];
2461 ldlm_lock_mode_downgrade(lock, LCK_COS);
2462 LDLM_LOCK_PUT(lock);
2466 spin_unlock(&rs->rs_lock);
2469 spin_lock(&exp->exp_uncommitted_replies_lock);
2470 list_del_init(&rs->rs_obd_list);
2471 spin_unlock(&exp->exp_uncommitted_replies_lock);
2474 spin_lock(&exp->exp_lock);
2475 /* Noop if removed already */
2476 list_del_init(&rs->rs_exp_list);
2477 spin_unlock(&exp->exp_lock);
2479 spin_lock(&rs->rs_lock);
2481 been_handled = rs->rs_handled;
2484 nlocks = rs->rs_nlocks; /* atomic "steal", but */
2485 rs->rs_nlocks = 0; /* locks still on rs_locks! */
2487 if (nlocks == 0 && !been_handled) {
2489 * If we see this, we should already have seen the warning
2490 * in mds_steal_ack_locks()
2493 "All locks stolen from rs %p x%lld.t%lld o%d NID %s\n",
2494 rs, rs->rs_xid, rs->rs_transno, rs->rs_opc,
2495 libcfs_nid2str(exp->exp_connection->c_peer.nid));
2498 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
2499 spin_unlock(&rs->rs_lock);
2501 if (!been_handled && rs->rs_on_net) {
2502 LNetMDUnlink(rs->rs_md_h);
2503 /* Ignore return code; we're racing with completion */
2506 while (nlocks-- > 0)
2507 ldlm_lock_decref(&rs->rs_locks[nlocks],
2508 rs->rs_modes[nlocks]);
2510 spin_lock(&rs->rs_lock);
2513 rs->rs_scheduled = 0;
2514 rs->rs_convert_lock = 0;
2516 if (!rs->rs_on_net) {
2518 spin_unlock(&rs->rs_lock);
2520 class_export_put(exp);
2521 rs->rs_export = NULL;
2522 ptlrpc_rs_decref(rs);
2523 if (atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
2524 svc->srv_is_stopping)
2525 wake_up_all(&svcpt->scp_waitq);
2529 /* still on the net; callback will schedule */
2530 spin_unlock(&rs->rs_lock);
2535 static void ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
2537 int avail = svcpt->scp_nrqbds_posted;
2538 int low_water = test_req_buffer_pressure ? 0 :
2539 svcpt->scp_service->srv_nbuf_per_group / 2;
2541 /* NB I'm not locking; just looking. */
2544 * CAVEAT EMPTOR: We might be allocating buffers here because we've
2545 * allowed the request history to grow out of control. We could put a
2546 * sanity check on that here and cull some history if we need the
2550 if (avail <= low_water)
2551 ptlrpc_grow_req_bufs(svcpt, 1);
2553 if (svcpt->scp_service->srv_stats) {
2554 lprocfs_counter_add(svcpt->scp_service->srv_stats,
2555 PTLRPC_REQBUF_AVAIL_CNTR, avail);
2559 static int ptlrpc_retry_rqbds(void *arg)
2561 struct ptlrpc_service_part *svcpt = (struct ptlrpc_service_part *)arg;
2563 svcpt->scp_rqbd_timeout = 0;
2567 static inline int ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
2569 return svcpt->scp_nreqs_active <
2570 svcpt->scp_nthrs_running - 1 -
2571 (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL);
2575 * allowed to create more threads
2576 * user can call it w/o any lock but need to hold
2577 * ptlrpc_service_part::scp_lock to get reliable result
2579 static inline int ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
2581 return svcpt->scp_nthrs_running +
2582 svcpt->scp_nthrs_starting <
2583 svcpt->scp_service->srv_nthrs_cpt_limit;
2587 * too many requests and allowed to create more threads
2589 static inline int ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
2591 return !ptlrpc_threads_enough(svcpt) &&
2592 ptlrpc_threads_increasable(svcpt);
2595 static inline int ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2597 return thread_is_stopping(thread) ||
2598 thread->t_svcpt->scp_service->srv_is_stopping;
2601 /* stop the highest numbered thread if there are too many threads running */
2602 static inline bool ptlrpc_thread_should_stop(struct ptlrpc_thread *thread)
2604 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2606 return thread->t_id >= svcpt->scp_service->srv_nthrs_cpt_limit &&
2607 thread->t_id == svcpt->scp_thr_nextid - 1;
2610 static void ptlrpc_stop_thread(struct ptlrpc_thread *thread)
2612 CDEBUG(D_INFO, "Stopping thread %s #%u\n",
2613 thread->t_svcpt->scp_service->srv_thread_name, thread->t_id);
2614 thread_add_flags(thread, SVC_STOPPING);
2617 static inline void ptlrpc_thread_stop(struct ptlrpc_thread *thread)
2619 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2621 spin_lock(&svcpt->scp_lock);
2622 if (ptlrpc_thread_should_stop(thread)) {
2623 ptlrpc_stop_thread(thread);
2624 svcpt->scp_thr_nextid--;
2626 spin_unlock(&svcpt->scp_lock);
2629 static inline int ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
2631 return !list_empty(&svcpt->scp_rqbd_idle) &&
2632 svcpt->scp_rqbd_timeout == 0;
2636 ptlrpc_at_check(struct ptlrpc_service_part *svcpt)
2638 return svcpt->scp_at_check;
2642 * If a thread runs too long or spends to much time on a single request,
2643 * we want to know about it, so we set up a delayed work item as a watchdog.
2644 * If it fires, we display a stack trace of the delayed thread,
2645 * providing we aren't rate-limited
2647 * Watchdog stack traces are limited to 3 per 'libcfs_watchdog_ratelimit'
2650 static struct ratelimit_state watchdog_limit;
2652 static void ptlrpc_watchdog_fire(struct work_struct *w)
2654 struct ptlrpc_thread *thread = container_of(w, struct ptlrpc_thread,
2656 u64 ms_lapse = ktime_ms_delta(ktime_get(), thread->t_touched);
2657 u32 ms_frac = do_div(ms_lapse, MSEC_PER_SEC);
2659 if (!__ratelimit(&watchdog_limit)) {
2660 /* below message is checked in sanity-quota.sh test_6,18 */
2661 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",
2662 thread->t_task->comm, thread->t_task->pid,
2665 libcfs_debug_dumpstack(thread->t_task);
2667 /* below message is checked in sanity-quota.sh test_6,18 */
2668 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",
2669 thread->t_task->comm, thread->t_task->pid,
2670 ms_lapse, ms_frac, libcfs_watchdog_ratelimit);
2674 static void ptlrpc_watchdog_init(struct delayed_work *work, time_t time)
2676 INIT_DELAYED_WORK(work, ptlrpc_watchdog_fire);
2677 schedule_delayed_work(work, cfs_time_seconds(time));
2680 static void ptlrpc_watchdog_disable(struct delayed_work *work)
2682 cancel_delayed_work_sync(work);
2685 static void ptlrpc_watchdog_touch(struct delayed_work *work, time_t time)
2687 struct ptlrpc_thread *thread = container_of(&work->work,
2688 struct ptlrpc_thread,
2690 thread->t_touched = ktime_get();
2691 mod_delayed_work(system_wq, work, cfs_time_seconds(time));
2695 * requests wait on preprocessing
2696 * user can call it w/o any lock but need to hold
2697 * ptlrpc_service_part::scp_lock to get reliable result
2700 ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
2702 return !list_empty(&svcpt->scp_req_incoming);
2705 static __attribute__((__noinline__)) int
2706 ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
2707 struct ptlrpc_thread *thread)
2709 /* Don't exit while there are replies to be handled */
2710 struct l_wait_info lwi = LWI_TIMEOUT(svcpt->scp_rqbd_timeout,
2711 ptlrpc_retry_rqbds, svcpt);
2713 ptlrpc_watchdog_disable(&thread->t_watchdog);
2717 l_wait_event_exclusive_head(svcpt->scp_waitq,
2718 ptlrpc_thread_stopping(thread) ||
2719 ptlrpc_server_request_incoming(svcpt) ||
2720 ptlrpc_server_request_pending(svcpt, false) ||
2721 ptlrpc_rqbd_pending(svcpt) ||
2722 ptlrpc_at_check(svcpt), &lwi);
2724 if (ptlrpc_thread_stopping(thread))
2727 ptlrpc_watchdog_touch(&thread->t_watchdog,
2728 ptlrpc_server_get_timeout(svcpt));
2733 * Main thread body for service threads.
2734 * Waits in a loop waiting for new requests to process to appear.
2735 * Every time an incoming requests is added to its queue, a waitq
2736 * is woken up and one of the threads will handle it.
2738 static int ptlrpc_main(void *arg)
2740 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
2741 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2742 struct ptlrpc_service *svc = svcpt->scp_service;
2743 struct ptlrpc_reply_state *rs;
2744 struct group_info *ginfo = NULL;
2746 int counter = 0, rc = 0;
2750 thread->t_task = current;
2751 thread->t_pid = current_pid();
2752 unshare_fs_struct();
2754 if (svc->srv_cpt_bind) {
2755 rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
2757 CWARN("%s: failed to bind %s on CPT %d\n",
2758 svc->srv_name, thread->t_name, svcpt->scp_cpt);
2762 ginfo = groups_alloc(0);
2764 GOTO(out, rc = -ENOMEM);
2766 set_current_groups(ginfo);
2767 put_group_info(ginfo);
2769 if (svc->srv_ops.so_thr_init != NULL) {
2770 rc = svc->srv_ops.so_thr_init(thread);
2777 GOTO(out_srv_fini, rc = -ENOMEM);
2778 rc = lu_env_add(env);
2782 rc = lu_context_init(&env->le_ctx,
2783 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2785 GOTO(out_env_remove, rc);
2787 thread->t_env = env;
2788 env->le_ctx.lc_thread = thread;
2789 env->le_ctx.lc_cookie = 0x6;
2791 while (!list_empty(&svcpt->scp_rqbd_idle)) {
2792 rc = ptlrpc_server_post_idle_rqbds(svcpt);
2796 CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
2797 svc->srv_name, svcpt->scp_cpt, rc);
2798 GOTO(out_ctx_fini, rc);
2801 /* Alloc reply state structure for this one */
2802 OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2804 GOTO(out_ctx_fini, rc = -ENOMEM);
2806 spin_lock(&svcpt->scp_lock);
2808 LASSERT(thread_is_starting(thread));
2809 thread_clear_flags(thread, SVC_STARTING);
2811 LASSERT(svcpt->scp_nthrs_starting == 1);
2812 svcpt->scp_nthrs_starting--;
2815 * SVC_STOPPING may already be set here if someone else is trying
2816 * to stop the service while this new thread has been dynamically
2817 * forked. We still set SVC_RUNNING to let our creator know that
2818 * we are now running, however we will exit as soon as possible
2820 thread_add_flags(thread, SVC_RUNNING);
2821 svcpt->scp_nthrs_running++;
2822 spin_unlock(&svcpt->scp_lock);
2824 /* wake up our creator in case he's still waiting. */
2825 wake_up(&thread->t_ctl_waitq);
2827 thread->t_touched = ktime_get();
2828 ptlrpc_watchdog_init(&thread->t_watchdog,
2829 ptlrpc_server_get_timeout(svcpt));
2831 spin_lock(&svcpt->scp_rep_lock);
2832 list_add(&rs->rs_list, &svcpt->scp_rep_idle);
2833 wake_up(&svcpt->scp_rep_waitq);
2834 spin_unlock(&svcpt->scp_rep_lock);
2836 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2837 svcpt->scp_nthrs_running);
2839 /* XXX maintain a list of all managed devices: insert here */
2840 while (!ptlrpc_thread_stopping(thread)) {
2841 if (ptlrpc_wait_event(svcpt, thread))
2844 ptlrpc_check_rqbd_pool(svcpt);
2846 if (ptlrpc_threads_need_create(svcpt)) {
2847 /* Ignore return code - we tried... */
2848 ptlrpc_start_thread(svcpt, 0);
2851 /* reset le_ses to initial state */
2853 /* Process all incoming reqs before handling any */
2854 if (ptlrpc_server_request_incoming(svcpt)) {
2855 lu_context_enter(&env->le_ctx);
2856 ptlrpc_server_handle_req_in(svcpt, thread);
2857 lu_context_exit(&env->le_ctx);
2859 /* but limit ourselves in case of flood */
2860 if (counter++ < 100)
2865 if (ptlrpc_at_check(svcpt))
2866 ptlrpc_at_check_timed(svcpt);
2868 if (ptlrpc_server_request_pending(svcpt, false)) {
2869 lu_context_enter(&env->le_ctx);
2870 ptlrpc_server_handle_request(svcpt, thread);
2871 lu_context_exit(&env->le_ctx);
2874 if (ptlrpc_rqbd_pending(svcpt) &&
2875 ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
2877 * I just failed to repost request buffers.
2878 * Wait for a timeout (unless something else
2879 * happens) before I try again
2881 svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
2882 CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
2883 svcpt->scp_nrqbds_posted);
2886 * If the number of threads has been tuned downward and this
2887 * thread should be stopped, then stop in reverse order so the
2888 * the threads always have contiguous thread index values.
2890 if (unlikely(ptlrpc_thread_should_stop(thread)))
2891 ptlrpc_thread_stop(thread);
2894 ptlrpc_watchdog_disable(&thread->t_watchdog);
2897 lu_context_fini(&env->le_ctx);
2903 /* deconstruct service thread state created by ptlrpc_start_thread() */
2904 if (svc->srv_ops.so_thr_done != NULL)
2905 svc->srv_ops.so_thr_done(thread);
2907 CDEBUG(D_RPCTRACE, "%s: service thread [%p:%u] %d exiting: rc = %d\n",
2908 thread->t_name, thread, thread->t_pid, thread->t_id, rc);
2909 spin_lock(&svcpt->scp_lock);
2910 if (thread_test_and_clear_flags(thread, SVC_STARTING))
2911 svcpt->scp_nthrs_starting--;
2913 if (thread_test_and_clear_flags(thread, SVC_RUNNING)) {
2914 /* must know immediately */
2915 svcpt->scp_nthrs_running--;
2919 thread_add_flags(thread, SVC_STOPPED);
2921 wake_up(&thread->t_ctl_waitq);
2922 spin_unlock(&svcpt->scp_lock);
2927 static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
2928 struct list_head *replies)
2932 spin_lock(&hrt->hrt_lock);
2934 list_splice_init(&hrt->hrt_queue, replies);
2935 result = ptlrpc_hr.hr_stopping || !list_empty(replies);
2937 spin_unlock(&hrt->hrt_lock);
2942 * Main body of "handle reply" function.
2943 * It processes acked reply states
2945 static int ptlrpc_hr_main(void *arg)
2947 struct ptlrpc_hr_thread *hrt = (struct ptlrpc_hr_thread *)arg;
2948 struct ptlrpc_hr_partition *hrp = hrt->hrt_partition;
2949 struct list_head replies;
2957 INIT_LIST_HEAD(&replies);
2958 unshare_fs_struct();
2960 rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
2962 char threadname[20];
2964 snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
2965 hrp->hrp_cpt, hrt->hrt_id);
2966 CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
2967 threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
2970 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD |
2971 LCT_REMEMBER | LCT_NOREF);
2975 rc = lu_env_add(env);
2977 GOTO(out_ctx_fini, rc);
2979 atomic_inc(&hrp->hrp_nstarted);
2980 wake_up(&ptlrpc_hr.hr_waitq);
2982 while (!ptlrpc_hr.hr_stopping) {
2983 l_wait_condition(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
2985 while (!list_empty(&replies)) {
2986 struct ptlrpc_reply_state *rs;
2988 rs = list_entry(replies.prev,
2989 struct ptlrpc_reply_state,
2991 list_del_init(&rs->rs_list);
2992 /* refill keys if needed */
2994 lu_context_enter(&env->le_ctx);
2995 ptlrpc_handle_rs(rs);
2996 lu_context_exit(&env->le_ctx);
3000 atomic_inc(&hrp->hrp_nstopped);
3001 wake_up(&ptlrpc_hr.hr_waitq);
3005 lu_context_fini(&env->le_ctx);
3011 static void ptlrpc_stop_hr_threads(void)
3013 struct ptlrpc_hr_partition *hrp;
3017 ptlrpc_hr.hr_stopping = 1;
3019 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
3020 if (hrp->hrp_thrs == NULL)
3021 continue; /* uninitialized */
3022 for (j = 0; j < hrp->hrp_nthrs; j++)
3023 wake_up_all(&hrp->hrp_thrs[j].hrt_waitq);
3026 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
3027 if (hrp->hrp_thrs == NULL)
3028 continue; /* uninitialized */
3029 wait_event(ptlrpc_hr.hr_waitq,
3030 atomic_read(&hrp->hrp_nstopped) ==
3031 atomic_read(&hrp->hrp_nstarted));
3035 static int ptlrpc_start_hr_threads(void)
3037 struct ptlrpc_hr_partition *hrp;
3043 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
3046 for (j = 0; j < hrp->hrp_nthrs; j++) {
3047 struct ptlrpc_hr_thread *hrt = &hrp->hrp_thrs[j];
3048 struct task_struct *task;
3050 task = kthread_run(ptlrpc_hr_main,
3052 "ptlrpc_hr%02d_%03d",
3061 wait_event(ptlrpc_hr.hr_waitq,
3062 atomic_read(&hrp->hrp_nstarted) == j);
3065 CERROR("cannot start reply handler thread %d:%d: rc = %d\n",
3067 ptlrpc_stop_hr_threads();
3075 static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
3077 struct l_wait_info lwi = { 0 };
3078 struct ptlrpc_thread *thread;
3079 struct list_head zombie;
3083 CDEBUG(D_INFO, "Stopping threads for service %s\n",
3084 svcpt->scp_service->srv_name);
3086 INIT_LIST_HEAD(&zombie);
3087 spin_lock(&svcpt->scp_lock);
3088 /* let the thread know that we would like it to stop asap */
3089 list_for_each_entry(thread, &svcpt->scp_threads, t_link)
3090 ptlrpc_stop_thread(thread);
3092 wake_up_all(&svcpt->scp_waitq);
3094 while (!list_empty(&svcpt->scp_threads)) {
3095 thread = list_entry(svcpt->scp_threads.next,
3096 struct ptlrpc_thread, t_link);
3097 if (thread_is_stopped(thread)) {
3098 list_move(&thread->t_link, &zombie);
3101 spin_unlock(&svcpt->scp_lock);
3103 CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
3104 svcpt->scp_service->srv_thread_name, thread->t_id);
3105 l_wait_event(thread->t_ctl_waitq,
3106 thread_is_stopped(thread), &lwi);
3108 spin_lock(&svcpt->scp_lock);
3111 spin_unlock(&svcpt->scp_lock);
3113 while (!list_empty(&zombie)) {
3114 thread = list_entry(zombie.next,
3115 struct ptlrpc_thread, t_link);
3116 list_del(&thread->t_link);
3117 OBD_FREE_PTR(thread);
3123 * Stops all threads of a particular service \a svc
3125 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
3127 struct ptlrpc_service_part *svcpt;
3132 ptlrpc_service_for_each_part(svcpt, i, svc) {
3133 if (svcpt->scp_service != NULL)
3134 ptlrpc_svcpt_stop_threads(svcpt);
3140 int ptlrpc_start_threads(struct ptlrpc_service *svc)
3148 /* We require 2 threads min, see note in ptlrpc_server_handle_request */
3149 LASSERT(svc->srv_nthrs_cpt_init >= PTLRPC_NTHRS_INIT);
3151 for (i = 0; i < svc->srv_ncpts; i++) {
3152 for (j = 0; j < svc->srv_nthrs_cpt_init; j++) {
3153 rc = ptlrpc_start_thread(svc->srv_parts[i], 1);
3159 /* We have enough threads, don't start more. b=15759 */
3166 CERROR("cannot start %s thread #%d_%d: rc %d\n",
3167 svc->srv_thread_name, i, j, rc);
3168 ptlrpc_stop_all_threads(svc);
3172 int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
3174 struct l_wait_info lwi = { 0 };
3175 struct ptlrpc_thread *thread;
3176 struct ptlrpc_service *svc;
3177 struct task_struct *task;
3182 LASSERT(svcpt != NULL);
3184 svc = svcpt->scp_service;
3186 CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n",
3187 svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running,
3188 svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit);
3191 if (unlikely(svc->srv_is_stopping))
3194 if (!ptlrpc_threads_increasable(svcpt) ||
3195 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
3196 svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1))
3199 OBD_CPT_ALLOC_PTR(thread, svc->srv_cptable, svcpt->scp_cpt);
3202 init_waitqueue_head(&thread->t_ctl_waitq);
3204 spin_lock(&svcpt->scp_lock);
3205 if (!ptlrpc_threads_increasable(svcpt)) {
3206 spin_unlock(&svcpt->scp_lock);
3207 OBD_FREE_PTR(thread);
3211 if (svcpt->scp_nthrs_starting != 0) {
3213 * serialize starting because some modules (obdfilter)
3214 * might require unique and contiguous t_id
3216 LASSERT(svcpt->scp_nthrs_starting == 1);
3217 spin_unlock(&svcpt->scp_lock);
3218 OBD_FREE_PTR(thread);
3220 CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n",
3221 svc->srv_thread_name, svcpt->scp_thr_nextid);
3226 CDEBUG(D_INFO, "Creating thread %s #%d race, retry later\n",
3227 svc->srv_thread_name, svcpt->scp_thr_nextid);
3231 svcpt->scp_nthrs_starting++;
3232 thread->t_id = svcpt->scp_thr_nextid++;
3233 thread_add_flags(thread, SVC_STARTING);
3234 thread->t_svcpt = svcpt;
3236 list_add(&thread->t_link, &svcpt->scp_threads);
3237 spin_unlock(&svcpt->scp_lock);
3239 if (svcpt->scp_cpt >= 0) {
3240 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s%02d_%03d",
3241 svc->srv_thread_name, svcpt->scp_cpt, thread->t_id);
3243 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s_%04d",
3244 svc->srv_thread_name, thread->t_id);
3247 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name);
3248 task = kthread_run(ptlrpc_main, thread, "%s", thread->t_name);
3251 CERROR("cannot start thread '%s': rc = %d\n",
3252 thread->t_name, rc);
3253 spin_lock(&svcpt->scp_lock);
3254 --svcpt->scp_nthrs_starting;
3255 if (thread_is_stopping(thread)) {
3257 * this ptlrpc_thread is being hanled
3258 * by ptlrpc_svcpt_stop_threads now
3260 thread_add_flags(thread, SVC_STOPPED);
3261 wake_up(&thread->t_ctl_waitq);
3262 spin_unlock(&svcpt->scp_lock);
3264 list_del(&thread->t_link);
3265 spin_unlock(&svcpt->scp_lock);
3266 OBD_FREE_PTR(thread);
3274 l_wait_event(thread->t_ctl_waitq,
3275 thread_is_running(thread) || thread_is_stopped(thread),
3278 rc = thread_is_stopped(thread) ? thread->t_id : 0;
3282 int ptlrpc_hr_init(void)
3284 struct ptlrpc_hr_partition *hrp;
3285 struct ptlrpc_hr_thread *hrt;
3293 memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
3294 ptlrpc_hr.hr_cpt_table = cfs_cpt_table;
3296 ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
3298 if (ptlrpc_hr.hr_partitions == NULL)
3301 ratelimit_state_init(&watchdog_limit,
3302 cfs_time_seconds(libcfs_watchdog_ratelimit), 3);
3304 init_waitqueue_head(&ptlrpc_hr.hr_waitq);
3307 weight = cpumask_weight(topology_sibling_cpumask(smp_processor_id()));
3310 cfs_percpt_for_each(hrp, cpt, ptlrpc_hr.hr_partitions) {
3313 atomic_set(&hrp->hrp_nstarted, 0);
3314 atomic_set(&hrp->hrp_nstopped, 0);
3316 hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, cpt);
3317 hrp->hrp_nthrs /= weight;
3318 if (hrp->hrp_nthrs == 0)
3321 OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, cpt,
3322 hrp->hrp_nthrs * sizeof(*hrt));
3323 if (hrp->hrp_thrs == NULL)
3324 GOTO(out, rc = -ENOMEM);
3326 for (i = 0; i < hrp->hrp_nthrs; i++) {
3327 hrt = &hrp->hrp_thrs[i];
3330 hrt->hrt_partition = hrp;
3331 init_waitqueue_head(&hrt->hrt_waitq);
3332 spin_lock_init(&hrt->hrt_lock);
3333 INIT_LIST_HEAD(&hrt->hrt_queue);
3337 rc = ptlrpc_start_hr_threads();
3344 void ptlrpc_hr_fini(void)
3346 struct ptlrpc_hr_partition *hrp;
3349 if (ptlrpc_hr.hr_partitions == NULL)
3352 ptlrpc_stop_hr_threads();
3354 cfs_percpt_for_each(hrp, cpt, ptlrpc_hr.hr_partitions) {
3355 if (hrp->hrp_thrs != NULL) {
3356 OBD_FREE(hrp->hrp_thrs,
3357 hrp->hrp_nthrs * sizeof(hrp->hrp_thrs[0]));
3361 cfs_percpt_free(ptlrpc_hr.hr_partitions);
3362 ptlrpc_hr.hr_partitions = NULL;
3367 * Wait until all already scheduled replies are processed.
3369 static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
3373 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
3376 rc = l_wait_event(svcpt->scp_waitq,
3377 atomic_read(&svcpt->scp_nreps_difficult) == 0, &lwi);
3380 CWARN("Unexpectedly long timeout %s %p\n",
3381 svcpt->scp_service->srv_name, svcpt->scp_service);
3386 ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
3388 struct ptlrpc_service_part *svcpt;
3391 /* early disarm AT timer... */
3392 ptlrpc_service_for_each_part(svcpt, i, svc) {
3393 if (svcpt->scp_service != NULL)
3394 del_timer(&svcpt->scp_at_timer);
3399 ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
3401 struct ptlrpc_service_part *svcpt;
3402 struct ptlrpc_request_buffer_desc *rqbd;
3403 struct l_wait_info lwi;
3408 * All history will be culled when the next request buffer is
3409 * freed in ptlrpc_service_purge_all()
3411 svc->srv_hist_nrqbds_cpt_max = 0;
3413 rc = LNetClearLazyPortal(svc->srv_req_portal);
3416 ptlrpc_service_for_each_part(svcpt, i, svc) {
3417 if (svcpt->scp_service == NULL)
3421 * Unlink all the request buffers. This forces a 'final'
3422 * event with its 'unlink' flag set for each posted rqbd
3424 list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
3426 rc = LNetMDUnlink(rqbd->rqbd_md_h);
3427 LASSERT(rc == 0 || rc == -ENOENT);
3431 ptlrpc_service_for_each_part(svcpt, i, svc) {
3432 if (svcpt->scp_service == NULL)
3436 * Wait for the network to release any buffers
3437 * it's currently filling
3439 spin_lock(&svcpt->scp_lock);
3440 while (svcpt->scp_nrqbds_posted != 0) {
3441 spin_unlock(&svcpt->scp_lock);
3443 * Network access will complete in finite time but
3444 * the HUGE timeout lets us CWARN for visibility
3447 lwi = LWI_TIMEOUT_INTERVAL(
3448 cfs_time_seconds(LONG_UNLINK),
3449 cfs_time_seconds(1), NULL, NULL);
3450 rc = l_wait_event(svcpt->scp_waitq,
3451 svcpt->scp_nrqbds_posted == 0, &lwi);
3452 if (rc == -ETIMEDOUT) {
3453 CWARN("Service %s waiting for request buffers\n",
3454 svcpt->scp_service->srv_name);
3456 spin_lock(&svcpt->scp_lock);
3458 spin_unlock(&svcpt->scp_lock);
3463 ptlrpc_service_purge_all(struct ptlrpc_service *svc)
3465 struct ptlrpc_service_part *svcpt;
3466 struct ptlrpc_request_buffer_desc *rqbd;
3467 struct ptlrpc_request *req;
3468 struct ptlrpc_reply_state *rs;
3471 ptlrpc_service_for_each_part(svcpt, i, svc) {
3472 if (svcpt->scp_service == NULL)
3475 spin_lock(&svcpt->scp_rep_lock);
3476 while (!list_empty(&svcpt->scp_rep_active)) {
3477 rs = list_entry(svcpt->scp_rep_active.next,
3478 struct ptlrpc_reply_state, rs_list);
3479 spin_lock(&rs->rs_lock);
3480 ptlrpc_schedule_difficult_reply(rs);
3481 spin_unlock(&rs->rs_lock);
3483 spin_unlock(&svcpt->scp_rep_lock);
3486 * purge the request queue. NB No new replies (rqbds
3487 * all unlinked) and no service threads, so I'm the only
3488 * thread noodling the request queue now
3490 while (!list_empty(&svcpt->scp_req_incoming)) {
3491 req = list_entry(svcpt->scp_req_incoming.next,
3492 struct ptlrpc_request, rq_list);
3494 list_del(&req->rq_list);
3495 svcpt->scp_nreqs_incoming--;
3496 ptlrpc_server_finish_request(svcpt, req);
3499 while (ptlrpc_server_request_pending(svcpt, true)) {
3500 req = ptlrpc_server_request_get(svcpt, true);
3501 ptlrpc_server_finish_active_request(svcpt, req);
3504 LASSERT(list_empty(&svcpt->scp_rqbd_posted));
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(array->paa_reqs_array,
3553 sizeof(struct list_head) * array->paa_size);
3554 array->paa_reqs_array = NULL;
3557 if (array->paa_reqs_count != NULL) {
3558 OBD_FREE(array->paa_reqs_count,
3559 sizeof(__u32) * array->paa_size);
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);