4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
32 #define DEBUG_SUBSYSTEM S_RPC
34 #include <linux/fs_struct.h>
35 #include <linux/kthread.h>
36 #include <linux/ratelimit.h>
38 #include <obd_support.h>
39 #include <obd_class.h>
40 #include <lustre_net.h>
41 #include <lu_object.h>
42 #include <uapi/linux/lnet/lnet-types.h>
43 #include "ptlrpc_internal.h"
44 #include <linux/delay.h>
46 /* The following are visible and mutable through /sys/module/ptlrpc */
47 int test_req_buffer_pressure = 0;
48 module_param(test_req_buffer_pressure, int, 0444);
49 MODULE_PARM_DESC(test_req_buffer_pressure, "set non-zero to put pressure on request buffer pools");
50 module_param(at_min, int, 0644);
51 MODULE_PARM_DESC(at_min, "Adaptive timeout minimum (sec)");
52 module_param(at_max, int, 0644);
53 MODULE_PARM_DESC(at_max, "Adaptive timeout maximum (sec)");
54 module_param(at_history, int, 0644);
55 MODULE_PARM_DESC(at_history,
56 "Adaptive timeouts remember the slowest event that took place within this period (sec)");
57 module_param(at_early_margin, int, 0644);
58 MODULE_PARM_DESC(at_early_margin, "How soon before an RPC deadline to send an early reply");
59 module_param(at_extra, int, 0644);
60 MODULE_PARM_DESC(at_extra, "How much extra time to give with each early reply");
63 static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt);
64 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req);
65 static void ptlrpc_at_remove_timed(struct ptlrpc_request *req);
66 static int ptlrpc_start_threads(struct ptlrpc_service *svc);
67 static int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait);
69 /** Holds a list of all PTLRPC services */
70 LIST_HEAD(ptlrpc_all_services);
71 /** Used to protect the \e ptlrpc_all_services list */
72 struct mutex ptlrpc_all_services_mutex;
74 static struct ptlrpc_request_buffer_desc *
75 ptlrpc_alloc_rqbd(struct ptlrpc_service_part *svcpt)
77 struct ptlrpc_service *svc = svcpt->scp_service;
78 struct ptlrpc_request_buffer_desc *rqbd;
80 OBD_CPT_ALLOC_PTR(rqbd, svc->srv_cptable, svcpt->scp_cpt);
84 rqbd->rqbd_svcpt = svcpt;
85 rqbd->rqbd_refcount = 0;
86 rqbd->rqbd_cbid.cbid_fn = request_in_callback;
87 rqbd->rqbd_cbid.cbid_arg = rqbd;
88 INIT_LIST_HEAD(&rqbd->rqbd_reqs);
89 OBD_CPT_ALLOC_LARGE(rqbd->rqbd_buffer, svc->srv_cptable,
90 svcpt->scp_cpt, svc->srv_buf_size);
91 if (rqbd->rqbd_buffer == NULL) {
96 spin_lock(&svcpt->scp_lock);
97 list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
98 svcpt->scp_nrqbds_total++;
99 spin_unlock(&svcpt->scp_lock);
104 static void ptlrpc_free_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
106 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
108 LASSERT(rqbd->rqbd_refcount == 0);
109 LASSERT(list_empty(&rqbd->rqbd_reqs));
111 spin_lock(&svcpt->scp_lock);
112 list_del(&rqbd->rqbd_list);
113 svcpt->scp_nrqbds_total--;
114 spin_unlock(&svcpt->scp_lock);
116 OBD_FREE_LARGE(rqbd->rqbd_buffer, svcpt->scp_service->srv_buf_size);
120 static int ptlrpc_grow_req_bufs(struct ptlrpc_service_part *svcpt, int post)
122 struct ptlrpc_service *svc = svcpt->scp_service;
123 struct ptlrpc_request_buffer_desc *rqbd;
127 if (svcpt->scp_rqbd_allocating)
130 spin_lock(&svcpt->scp_lock);
131 /* check again with lock */
132 if (svcpt->scp_rqbd_allocating) {
133 /* NB: we might allow more than one thread in the future */
134 LASSERT(svcpt->scp_rqbd_allocating == 1);
135 spin_unlock(&svcpt->scp_lock);
139 svcpt->scp_rqbd_allocating++;
140 spin_unlock(&svcpt->scp_lock);
143 for (i = 0; i < svc->srv_nbuf_per_group; i++) {
145 * NB: another thread might have recycled enough rqbds, we
146 * need to make sure it wouldn't over-allocate, see LU-1212.
148 if (svcpt->scp_nrqbds_posted >= svc->srv_nbuf_per_group ||
149 (svc->srv_nrqbds_max != 0 &&
150 svcpt->scp_nrqbds_total > svc->srv_nrqbds_max))
153 rqbd = ptlrpc_alloc_rqbd(svcpt);
156 CERROR("%s: Can't allocate request buffer\n",
163 spin_lock(&svcpt->scp_lock);
165 LASSERT(svcpt->scp_rqbd_allocating == 1);
166 svcpt->scp_rqbd_allocating--;
168 spin_unlock(&svcpt->scp_lock);
171 "%s: allocate %d new %d-byte reqbufs (%d/%d left), rc = %d\n",
172 svc->srv_name, i, svc->srv_buf_size, svcpt->scp_nrqbds_posted,
173 svcpt->scp_nrqbds_total, rc);
177 rc = ptlrpc_server_post_idle_rqbds(svcpt);
183 * Part of Rep-Ack logic.
184 * Puts a lock and its mode into reply state assotiated to request reply.
186 void ptlrpc_save_lock(struct ptlrpc_request *req, struct lustre_handle *lock,
187 int mode, bool no_ack, bool convert_lock)
189 struct ptlrpc_reply_state *rs = req->rq_reply_state;
193 LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
195 idx = rs->rs_nlocks++;
196 rs->rs_locks[idx] = *lock;
197 rs->rs_modes[idx] = mode;
198 rs->rs_difficult = 1;
199 rs->rs_no_ack = no_ack;
200 rs->rs_convert_lock = convert_lock;
202 EXPORT_SYMBOL(ptlrpc_save_lock);
205 struct ptlrpc_hr_partition;
207 struct ptlrpc_hr_thread {
208 int hrt_id; /* thread ID */
210 wait_queue_head_t hrt_waitq;
211 struct list_head hrt_queue;
212 struct ptlrpc_hr_partition *hrt_partition;
215 struct ptlrpc_hr_partition {
216 /* # of started threads */
217 atomic_t hrp_nstarted;
218 /* # of stopped threads */
219 atomic_t hrp_nstopped;
220 /* cpu partition id */
222 /* round-robin rotor for choosing thread */
224 /* total number of threads on this partition */
227 struct ptlrpc_hr_thread *hrp_thrs;
230 #define HRT_RUNNING 0
231 #define HRT_STOPPING 1
233 struct ptlrpc_hr_service {
234 /* CPU partition table, it's just cfs_cpt_tab for now */
235 struct cfs_cpt_table *hr_cpt_table;
236 /** controller sleep waitq */
237 wait_queue_head_t hr_waitq;
238 unsigned int hr_stopping;
239 /** roundrobin rotor for non-affinity service */
240 unsigned int hr_rotor;
242 struct ptlrpc_hr_partition **hr_partitions;
246 struct list_head rsb_replies;
247 unsigned int rsb_n_replies;
248 struct ptlrpc_service_part *rsb_svcpt;
251 /** reply handling service. */
252 static struct ptlrpc_hr_service ptlrpc_hr;
255 * maximum mumber of replies scheduled in one batch
257 #define MAX_SCHEDULED 256
260 * Initialize a reply batch.
264 static void rs_batch_init(struct rs_batch *b)
266 memset(b, 0, sizeof(*b));
267 INIT_LIST_HEAD(&b->rsb_replies);
271 * Choose an hr thread to dispatch requests to.
274 struct ptlrpc_hr_thread *ptlrpc_hr_select(struct ptlrpc_service_part *svcpt)
276 struct ptlrpc_hr_partition *hrp;
279 if (svcpt->scp_cpt >= 0 &&
280 svcpt->scp_service->srv_cptable == ptlrpc_hr.hr_cpt_table) {
281 /* directly match partition */
282 hrp = ptlrpc_hr.hr_partitions[svcpt->scp_cpt];
285 rotor = ptlrpc_hr.hr_rotor++;
286 rotor %= cfs_cpt_number(ptlrpc_hr.hr_cpt_table);
288 hrp = ptlrpc_hr.hr_partitions[rotor];
291 rotor = hrp->hrp_rotor++;
292 return &hrp->hrp_thrs[rotor % hrp->hrp_nthrs];
296 * Dispatch all replies accumulated in the batch to one from
297 * dedicated reply handling threads.
301 static void rs_batch_dispatch(struct rs_batch *b)
303 if (b->rsb_n_replies != 0) {
304 struct ptlrpc_hr_thread *hrt;
306 hrt = ptlrpc_hr_select(b->rsb_svcpt);
308 spin_lock(&hrt->hrt_lock);
309 list_splice_init(&b->rsb_replies, &hrt->hrt_queue);
310 spin_unlock(&hrt->hrt_lock);
312 wake_up(&hrt->hrt_waitq);
313 b->rsb_n_replies = 0;
318 * Add a reply to a batch.
319 * Add one reply object to a batch, schedule batched replies if overload.
324 static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs)
326 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
328 if (svcpt != b->rsb_svcpt || b->rsb_n_replies >= MAX_SCHEDULED) {
329 if (b->rsb_svcpt != NULL) {
330 rs_batch_dispatch(b);
331 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
333 spin_lock(&svcpt->scp_rep_lock);
334 b->rsb_svcpt = svcpt;
336 spin_lock(&rs->rs_lock);
337 rs->rs_scheduled_ever = 1;
338 if (rs->rs_scheduled == 0) {
339 list_move(&rs->rs_list, &b->rsb_replies);
340 rs->rs_scheduled = 1;
343 rs->rs_committed = 1;
344 spin_unlock(&rs->rs_lock);
348 * Reply batch finalization.
349 * Dispatch remaining replies from the batch
350 * and release remaining spinlock.
354 static void rs_batch_fini(struct rs_batch *b)
356 if (b->rsb_svcpt != NULL) {
357 rs_batch_dispatch(b);
358 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
362 #define DECLARE_RS_BATCH(b) struct rs_batch b
366 * Put reply state into a queue for processing because we received
367 * ACK from the client
369 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
371 struct ptlrpc_hr_thread *hrt;
375 LASSERT(list_empty(&rs->rs_list));
377 hrt = ptlrpc_hr_select(rs->rs_svcpt);
379 spin_lock(&hrt->hrt_lock);
380 list_add_tail(&rs->rs_list, &hrt->hrt_queue);
381 spin_unlock(&hrt->hrt_lock);
383 wake_up(&hrt->hrt_waitq);
387 void ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs)
391 assert_spin_locked(&rs->rs_svcpt->scp_rep_lock);
392 assert_spin_locked(&rs->rs_lock);
393 LASSERT(rs->rs_difficult);
394 rs->rs_scheduled_ever = 1; /* flag any notification attempt */
396 if (rs->rs_scheduled) { /* being set up or already notified */
401 rs->rs_scheduled = 1;
402 list_del_init(&rs->rs_list);
403 ptlrpc_dispatch_difficult_reply(rs);
406 EXPORT_SYMBOL(ptlrpc_schedule_difficult_reply);
408 void ptlrpc_commit_replies(struct obd_export *exp)
410 struct ptlrpc_reply_state *rs, *nxt;
411 DECLARE_RS_BATCH(batch);
415 rs_batch_init(&batch);
417 * Find any replies that have been committed and get their service
418 * to attend to complete them.
421 /* CAVEAT EMPTOR: spinlock ordering!!! */
422 spin_lock(&exp->exp_uncommitted_replies_lock);
423 list_for_each_entry_safe(rs, nxt, &exp->exp_uncommitted_replies,
425 LASSERT(rs->rs_difficult);
426 /* VBR: per-export last_committed */
427 LASSERT(rs->rs_export);
428 if (rs->rs_transno <= exp->exp_last_committed) {
429 list_del_init(&rs->rs_obd_list);
430 rs_batch_add(&batch, rs);
433 spin_unlock(&exp->exp_uncommitted_replies_lock);
434 rs_batch_fini(&batch);
438 static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt)
440 struct ptlrpc_request_buffer_desc *rqbd;
445 spin_lock(&svcpt->scp_lock);
447 if (list_empty(&svcpt->scp_rqbd_idle)) {
448 spin_unlock(&svcpt->scp_lock);
452 rqbd = list_first_entry(&svcpt->scp_rqbd_idle,
453 struct ptlrpc_request_buffer_desc,
456 /* assume we will post successfully */
457 svcpt->scp_nrqbds_posted++;
458 list_move(&rqbd->rqbd_list, &svcpt->scp_rqbd_posted);
460 spin_unlock(&svcpt->scp_lock);
462 rc = ptlrpc_register_rqbd(rqbd);
469 spin_lock(&svcpt->scp_lock);
471 svcpt->scp_nrqbds_posted--;
472 list_move_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
475 * Don't complain if no request buffers are posted right now; LNET
476 * won't drop requests because we set the portal lazy!
479 spin_unlock(&svcpt->scp_lock);
484 static void ptlrpc_at_timer(cfs_timer_cb_arg_t data)
486 struct ptlrpc_service_part *svcpt;
488 svcpt = cfs_from_timer(svcpt, data, scp_at_timer);
490 svcpt->scp_at_check = 1;
491 svcpt->scp_at_checktime = ktime_get();
492 wake_up(&svcpt->scp_waitq);
495 static void ptlrpc_server_nthreads_check(struct ptlrpc_service *svc,
496 struct ptlrpc_service_conf *conf)
498 struct ptlrpc_service_thr_conf *tc = &conf->psc_thr;
505 * Common code for estimating & validating threads number.
506 * CPT affinity service could have percpt thread-pool instead
507 * of a global thread-pool, which means user might not always
508 * get the threads number they give it in conf::tc_nthrs_user
509 * even they did set. It's because we need to validate threads
510 * number for each CPT to guarantee each pool will have enough
511 * threads to keep the service healthy.
513 init = PTLRPC_NTHRS_INIT + (svc->srv_ops.so_hpreq_handler != NULL);
514 init = max_t(int, init, tc->tc_nthrs_init);
517 * NB: please see comments in lustre_lnet.h for definition
518 * details of these members
520 LASSERT(tc->tc_nthrs_max != 0);
522 if (tc->tc_nthrs_user != 0) {
524 * In case there is a reason to test a service with many
525 * threads, we give a less strict check here, it can
526 * be up to 8 * nthrs_max
528 total = min(tc->tc_nthrs_max * 8, tc->tc_nthrs_user);
529 nthrs = total / svc->srv_ncpts;
530 init = max(init, nthrs);
534 total = tc->tc_nthrs_max;
535 if (tc->tc_nthrs_base == 0) {
537 * don't care about base threads number per partition,
538 * this is most for non-affinity service
540 nthrs = total / svc->srv_ncpts;
544 nthrs = tc->tc_nthrs_base;
545 if (svc->srv_ncpts == 1) {
549 * NB: Increase the base number if it's single partition
550 * and total number of cores/HTs is larger or equal to 4.
551 * result will always < 2 * nthrs_base
553 weight = cfs_cpt_weight(svc->srv_cptable, CFS_CPT_ANY);
554 for (i = 1; (weight >> (i + 1)) != 0 && /* >= 4 cores/HTs */
555 (tc->tc_nthrs_base >> i) != 0; i++)
556 nthrs += tc->tc_nthrs_base >> i;
559 if (tc->tc_thr_factor != 0) {
560 int factor = tc->tc_thr_factor;
564 * User wants to increase number of threads with for
565 * each CPU core/HT, most likely the factor is larger than
566 * one thread/core because service threads are supposed to
567 * be blocked by lock or wait for IO.
570 * Amdahl's law says that adding processors wouldn't give
571 * a linear increasing of parallelism, so it's nonsense to
572 * have too many threads no matter how many cores/HTs
577 (topology_sibling_cpumask(smp_processor_id())) > 1) {
578 /* weight is # of HTs */
579 /* depress thread factor for hyper-thread */
580 factor = factor - (factor >> 1) + (factor >> 3);
584 weight = cfs_cpt_weight(svc->srv_cptable, 0);
586 for (; factor > 0 && weight > 0; factor--, weight -= fade)
587 nthrs += min(weight, fade) * factor;
590 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
591 nthrs = max(tc->tc_nthrs_base,
592 tc->tc_nthrs_max / svc->srv_ncpts);
595 nthrs = max(nthrs, tc->tc_nthrs_init);
596 svc->srv_nthrs_cpt_limit = nthrs;
597 svc->srv_nthrs_cpt_init = init;
599 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
601 "%s: This service may have more threads (%d) than the given soft limit (%d)\n",
602 svc->srv_name, nthrs * svc->srv_ncpts,
608 * Initialize percpt data for a service
610 static int ptlrpc_service_part_init(struct ptlrpc_service *svc,
611 struct ptlrpc_service_part *svcpt, int cpt)
613 struct ptlrpc_at_array *array;
618 svcpt->scp_cpt = cpt;
619 INIT_LIST_HEAD(&svcpt->scp_threads);
621 /* rqbd and incoming request queue */
622 spin_lock_init(&svcpt->scp_lock);
623 mutex_init(&svcpt->scp_mutex);
624 INIT_LIST_HEAD(&svcpt->scp_rqbd_idle);
625 INIT_LIST_HEAD(&svcpt->scp_rqbd_posted);
626 INIT_LIST_HEAD(&svcpt->scp_req_incoming);
627 init_waitqueue_head(&svcpt->scp_waitq);
628 /* history request & rqbd list */
629 INIT_LIST_HEAD(&svcpt->scp_hist_reqs);
630 INIT_LIST_HEAD(&svcpt->scp_hist_rqbds);
632 /* acitve requests and hp requests */
633 spin_lock_init(&svcpt->scp_req_lock);
636 spin_lock_init(&svcpt->scp_rep_lock);
637 INIT_LIST_HEAD(&svcpt->scp_rep_active);
638 INIT_LIST_HEAD(&svcpt->scp_rep_idle);
639 init_waitqueue_head(&svcpt->scp_rep_waitq);
640 atomic_set(&svcpt->scp_nreps_difficult, 0);
642 /* adaptive timeout */
643 spin_lock_init(&svcpt->scp_at_lock);
644 array = &svcpt->scp_at_array;
646 size = at_est2timeout(at_max);
647 array->paa_size = size;
648 array->paa_count = 0;
649 array->paa_deadline = -1;
651 /* allocate memory for scp_at_array (ptlrpc_at_array) */
652 OBD_CPT_ALLOC(array->paa_reqs_array,
653 svc->srv_cptable, cpt, sizeof(struct list_head) * size);
654 if (array->paa_reqs_array == NULL)
657 for (index = 0; index < size; index++)
658 INIT_LIST_HEAD(&array->paa_reqs_array[index]);
660 OBD_CPT_ALLOC(array->paa_reqs_count,
661 svc->srv_cptable, cpt, sizeof(__u32) * size);
662 if (array->paa_reqs_count == NULL)
665 cfs_timer_setup(&svcpt->scp_at_timer, ptlrpc_at_timer,
666 (unsigned long)svcpt, 0);
669 * At SOW, service time should be quick; 10s seems generous. If client
670 * timeout is less than this, we'll be sending an early reply.
672 at_init(&svcpt->scp_at_estimate, 10, 0);
674 /* assign this before call ptlrpc_grow_req_bufs */
675 svcpt->scp_service = svc;
676 /* Now allocate the request buffers, but don't post them now */
677 rc = ptlrpc_grow_req_bufs(svcpt, 0);
679 * We shouldn't be under memory pressure at startup, so
680 * fail if we can't allocate all our buffers at this time.
688 if (array->paa_reqs_count != NULL) {
689 OBD_FREE_PTR_ARRAY(array->paa_reqs_count, size);
690 array->paa_reqs_count = NULL;
693 if (array->paa_reqs_array != NULL) {
694 OBD_FREE_PTR_ARRAY(array->paa_reqs_array, array->paa_size);
695 array->paa_reqs_array = NULL;
702 * Initialize service on a given portal.
703 * This includes starting serving threads , allocating and posting rqbds and
706 struct ptlrpc_service *ptlrpc_register_service(struct ptlrpc_service_conf *conf,
708 struct dentry *debugfs_entry)
710 struct ptlrpc_service_cpt_conf *cconf = &conf->psc_cpt;
711 struct ptlrpc_service *service;
712 struct ptlrpc_service_part *svcpt;
713 struct cfs_cpt_table *cptable;
722 LASSERT(conf->psc_buf.bc_nbufs > 0);
723 LASSERT(conf->psc_buf.bc_buf_size >=
724 conf->psc_buf.bc_req_max_size + SPTLRPC_MAX_PAYLOAD);
725 LASSERT(conf->psc_thr.tc_ctx_tags != 0);
727 cptable = cconf->cc_cptable;
729 cptable = cfs_cpt_tab;
731 if (conf->psc_thr.tc_cpu_bind > 1) {
732 CERROR("%s: Invalid cpu bind value %d, only 1 or 0 allowed\n",
733 conf->psc_name, conf->psc_thr.tc_cpu_bind);
734 RETURN(ERR_PTR(-EINVAL));
737 if (!cconf->cc_affinity) {
740 ncpts = cfs_cpt_number(cptable);
741 if (cconf->cc_pattern != NULL) {
742 struct cfs_expr_list *el;
744 rc = cfs_expr_list_parse(cconf->cc_pattern,
745 strlen(cconf->cc_pattern),
748 CERROR("%s: invalid CPT pattern string: %s\n",
749 conf->psc_name, cconf->cc_pattern);
750 RETURN(ERR_PTR(-EINVAL));
753 rc = cfs_expr_list_values(el, ncpts, &cpts);
754 cfs_expr_list_free(el);
756 CERROR("%s: failed to parse CPT array %s: %d\n",
757 conf->psc_name, cconf->cc_pattern, rc);
759 OBD_FREE_PTR_ARRAY(cpts, ncpts);
760 RETURN(ERR_PTR(rc < 0 ? rc : -EINVAL));
766 OBD_ALLOC(service, offsetof(struct ptlrpc_service, srv_parts[ncpts]));
767 if (service == NULL) {
769 OBD_FREE_PTR_ARRAY(cpts, ncpts);
770 RETURN(ERR_PTR(-ENOMEM));
773 service->srv_cptable = cptable;
774 service->srv_cpts = cpts;
775 service->srv_ncpts = ncpts;
776 service->srv_cpt_bind = conf->psc_thr.tc_cpu_bind;
778 service->srv_cpt_bits = 0; /* it's zero already, easy to read... */
779 while ((1 << service->srv_cpt_bits) < cfs_cpt_number(cptable))
780 service->srv_cpt_bits++;
783 spin_lock_init(&service->srv_lock);
784 service->srv_name = conf->psc_name;
785 service->srv_watchdog_factor = conf->psc_watchdog_factor;
786 INIT_LIST_HEAD(&service->srv_list); /* for safty of cleanup */
788 /* buffer configuration */
789 service->srv_nbuf_per_group = test_req_buffer_pressure ?
790 1 : conf->psc_buf.bc_nbufs;
791 /* do not limit max number of rqbds by default */
792 service->srv_nrqbds_max = 0;
794 service->srv_max_req_size = conf->psc_buf.bc_req_max_size +
796 service->srv_buf_size = conf->psc_buf.bc_buf_size;
797 service->srv_rep_portal = conf->psc_buf.bc_rep_portal;
798 service->srv_req_portal = conf->psc_buf.bc_req_portal;
800 /* With slab/alloc_pages buffer size will be rounded up to 2^n */
801 if (service->srv_buf_size & (service->srv_buf_size - 1)) {
802 int round = size_roundup_power2(service->srv_buf_size);
804 service->srv_buf_size = round;
807 /* Increase max reply size to next power of two */
808 service->srv_max_reply_size = 1;
809 while (service->srv_max_reply_size <
810 conf->psc_buf.bc_rep_max_size + SPTLRPC_MAX_PAYLOAD)
811 service->srv_max_reply_size <<= 1;
813 service->srv_thread_name = conf->psc_thr.tc_thr_name;
814 service->srv_ctx_tags = conf->psc_thr.tc_ctx_tags;
815 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
816 service->srv_ops = conf->psc_ops;
818 for (i = 0; i < ncpts; i++) {
819 if (!cconf->cc_affinity)
822 cpt = cpts != NULL ? cpts[i] : i;
824 OBD_CPT_ALLOC(svcpt, cptable, cpt, sizeof(*svcpt));
826 GOTO(failed, rc = -ENOMEM);
828 service->srv_parts[i] = svcpt;
829 rc = ptlrpc_service_part_init(service, svcpt, cpt);
834 ptlrpc_server_nthreads_check(service, conf);
836 rc = LNetSetLazyPortal(service->srv_req_portal);
839 mutex_lock(&ptlrpc_all_services_mutex);
840 list_add(&service->srv_list, &ptlrpc_all_services);
841 mutex_unlock(&ptlrpc_all_services_mutex);
844 rc = ptlrpc_sysfs_register_service(parent, service);
849 if (debugfs_entry != NULL)
850 ptlrpc_ldebugfs_register_service(debugfs_entry, service);
852 rc = ptlrpc_service_nrs_setup(service);
856 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
857 service->srv_name, service->srv_req_portal);
859 rc = ptlrpc_start_threads(service);
861 CERROR("Failed to start threads for service %s: %d\n",
862 service->srv_name, rc);
868 ptlrpc_unregister_service(service);
871 EXPORT_SYMBOL(ptlrpc_register_service);
874 * to actually free the request, must be called without holding svc_lock.
875 * note it's caller's responsibility to unlink req->rq_list.
877 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
879 LASSERT(atomic_read(&req->rq_refcount) == 0);
880 LASSERT(list_empty(&req->rq_timed_list));
883 * DEBUG_REQ() assumes the reply state of a request with a valid
884 * ref will not be destroyed until that reference is dropped.
886 ptlrpc_req_drop_rs(req);
888 sptlrpc_svc_ctx_decref(req);
890 if (req != &req->rq_rqbd->rqbd_req) {
892 * NB request buffers use an embedded
893 * req if the incoming req unlinked the
894 * MD; this isn't one of them!
896 ptlrpc_request_cache_free(req);
901 * drop a reference count of the request. if it reaches 0, we either
902 * put it into history list, or free it immediately.
904 void ptlrpc_server_drop_request(struct ptlrpc_request *req)
906 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
907 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
908 struct ptlrpc_service *svc = svcpt->scp_service;
911 if (!atomic_dec_and_test(&req->rq_refcount))
914 if (req->rq_session.lc_state == LCS_ENTERED) {
915 lu_context_exit(&req->rq_session);
916 lu_context_fini(&req->rq_session);
919 if (req->rq_at_linked) {
920 spin_lock(&svcpt->scp_at_lock);
922 * recheck with lock, in case it's unlinked by
923 * ptlrpc_at_check_timed()
925 if (likely(req->rq_at_linked))
926 ptlrpc_at_remove_timed(req);
927 spin_unlock(&svcpt->scp_at_lock);
930 LASSERT(list_empty(&req->rq_timed_list));
932 /* finalize request */
933 if (req->rq_export) {
934 class_export_put(req->rq_export);
935 req->rq_export = NULL;
938 spin_lock(&svcpt->scp_lock);
940 list_add(&req->rq_list, &rqbd->rqbd_reqs);
942 refcount = --(rqbd->rqbd_refcount);
944 /* request buffer is now idle: add to history */
945 list_move_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds);
946 svcpt->scp_hist_nrqbds++;
950 * I expect only about 1 or 2 rqbds need to be recycled here
952 while (svcpt->scp_hist_nrqbds > svc->srv_hist_nrqbds_cpt_max) {
953 rqbd = list_first_entry(&svcpt->scp_hist_rqbds,
954 struct ptlrpc_request_buffer_desc,
957 list_del(&rqbd->rqbd_list);
958 svcpt->scp_hist_nrqbds--;
961 * remove rqbd's reqs from svc's req history while
962 * I've got the service lock
964 list_for_each_entry(req, &rqbd->rqbd_reqs, rq_list) {
965 /* Track the highest culled req seq */
966 if (req->rq_history_seq >
967 svcpt->scp_hist_seq_culled) {
968 svcpt->scp_hist_seq_culled =
971 list_del(&req->rq_history_list);
974 spin_unlock(&svcpt->scp_lock);
976 while ((req = list_first_entry_or_null(
978 struct ptlrpc_request, rq_list))) {
979 list_del(&req->rq_list);
980 ptlrpc_server_free_request(req);
983 spin_lock(&svcpt->scp_lock);
985 * now all reqs including the embedded req has been
986 * disposed, schedule request buffer for re-use
987 * or free it to drain some in excess.
989 LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) == 0);
990 if (svcpt->scp_nrqbds_posted >=
991 svc->srv_nbuf_per_group ||
992 (svc->srv_nrqbds_max != 0 &&
993 svcpt->scp_nrqbds_total > svc->srv_nrqbds_max) ||
994 test_req_buffer_pressure) {
995 /* like in ptlrpc_free_rqbd() */
996 svcpt->scp_nrqbds_total--;
997 OBD_FREE_LARGE(rqbd->rqbd_buffer,
1001 list_add_tail(&rqbd->rqbd_list,
1002 &svcpt->scp_rqbd_idle);
1006 spin_unlock(&svcpt->scp_lock);
1007 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
1008 /* If we are low on memory, we are not interested in history */
1009 list_del(&req->rq_list);
1010 list_del_init(&req->rq_history_list);
1012 /* Track the highest culled req seq */
1013 if (req->rq_history_seq > svcpt->scp_hist_seq_culled)
1014 svcpt->scp_hist_seq_culled = req->rq_history_seq;
1016 spin_unlock(&svcpt->scp_lock);
1018 ptlrpc_server_free_request(req);
1020 spin_unlock(&svcpt->scp_lock);
1024 static void ptlrpc_add_exp_list_nolock(struct ptlrpc_request *req,
1025 struct obd_export *export, bool hp)
1027 __u16 tag = lustre_msg_get_tag(req->rq_reqmsg);
1030 list_add(&req->rq_exp_list, &export->exp_hp_rpcs);
1032 list_add(&req->rq_exp_list, &export->exp_reg_rpcs);
1033 if (tag && export->exp_used_slots)
1034 set_bit(tag - 1, export->exp_used_slots);
1037 static void ptlrpc_del_exp_list(struct ptlrpc_request *req)
1039 __u16 tag = lustre_msg_get_tag(req->rq_reqmsg);
1041 spin_lock(&req->rq_export->exp_rpc_lock);
1042 list_del_init(&req->rq_exp_list);
1043 if (tag && !req->rq_obsolete && req->rq_export->exp_used_slots)
1044 clear_bit(tag - 1, req->rq_export->exp_used_slots);
1045 spin_unlock(&req->rq_export->exp_rpc_lock);
1048 /** Change request export and move hp request from old export to new */
1049 void ptlrpc_request_change_export(struct ptlrpc_request *req,
1050 struct obd_export *export)
1052 if (req->rq_export != NULL) {
1053 LASSERT(!list_empty(&req->rq_exp_list));
1054 /* remove rq_exp_list from last export */
1055 ptlrpc_del_exp_list(req);
1056 /* export has one reference already, so it's safe to
1057 * add req to export queue here and get another
1058 * reference for request later
1060 spin_lock(&export->exp_rpc_lock);
1061 ptlrpc_add_exp_list_nolock(req, export, req->rq_ops != NULL);
1062 spin_unlock(&export->exp_rpc_lock);
1064 class_export_rpc_dec(req->rq_export);
1065 class_export_put(req->rq_export);
1068 /* request takes one export refcount */
1069 req->rq_export = class_export_get(export);
1070 class_export_rpc_inc(export);
1074 * to finish a request: stop sending more early replies, and release
1077 static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt,
1078 struct ptlrpc_request *req)
1080 ptlrpc_server_hpreq_fini(req);
1082 ptlrpc_server_drop_request(req);
1086 * to finish an active request: stop sending more early replies, and release
1087 * the request. should be called after we finished handling the request.
1089 static void ptlrpc_server_finish_active_request(
1090 struct ptlrpc_service_part *svcpt,
1091 struct ptlrpc_request *req)
1093 spin_lock(&svcpt->scp_req_lock);
1094 ptlrpc_nrs_req_stop_nolock(req);
1095 svcpt->scp_nreqs_active--;
1097 svcpt->scp_nhreqs_active--;
1098 spin_unlock(&svcpt->scp_req_lock);
1100 ptlrpc_nrs_req_finalize(req);
1102 if (req->rq_export != NULL)
1103 class_export_rpc_dec(req->rq_export);
1105 ptlrpc_server_finish_request(svcpt, req);
1109 * This function makes sure dead exports are evicted in a timely manner.
1110 * This function is only called when some export receives a message (i.e.,
1111 * the network is up.)
1113 void ptlrpc_update_export_timer(struct obd_export *exp, time64_t extra_delay)
1115 struct obd_export *oldest_exp;
1116 time64_t oldest_time, new_time;
1123 * Compensate for slow machines, etc, by faking our request time
1124 * into the future. Although this can break the strict time-ordering
1125 * of the list, we can be really lazy here - we don't have to evict
1126 * at the exact right moment. Eventually, all silent exports
1127 * will make it to the top of the list.
1130 /* Do not pay attention on 1sec or smaller renewals. */
1131 new_time = ktime_get_real_seconds() + extra_delay;
1132 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
1135 exp->exp_last_request_time = new_time;
1138 * exports may get disconnected from the chain even though the
1139 * export has references, so we must keep the spin lock while
1140 * manipulating the lists
1142 spin_lock(&exp->exp_obd->obd_dev_lock);
1144 if (list_empty(&exp->exp_obd_chain_timed)) {
1145 /* this one is not timed */
1146 spin_unlock(&exp->exp_obd->obd_dev_lock);
1150 list_move_tail(&exp->exp_obd_chain_timed,
1151 &exp->exp_obd->obd_exports_timed);
1153 oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
1154 struct obd_export, exp_obd_chain_timed);
1155 oldest_time = oldest_exp->exp_last_request_time;
1156 spin_unlock(&exp->exp_obd->obd_dev_lock);
1158 if (exp->exp_obd->obd_recovering) {
1159 /* be nice to everyone during recovery */
1164 /* Note - racing to start/reset the obd_eviction timer is safe */
1165 if (exp->exp_obd->obd_eviction_timer == 0) {
1166 /* Check if the oldest entry is expired. */
1167 if (ktime_get_real_seconds() >
1168 oldest_time + PING_EVICT_TIMEOUT + extra_delay) {
1170 * We need a second timer, in case the net was down and
1171 * it just came back. Since the pinger may skip every
1172 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
1173 * we better wait for 3.
1175 exp->exp_obd->obd_eviction_timer =
1176 ktime_get_real_seconds() + 3 * PING_INTERVAL;
1177 CDEBUG(D_HA, "%s: Think about evicting %s from %lld\n",
1178 exp->exp_obd->obd_name,
1179 obd_export_nid2str(oldest_exp), oldest_time);
1182 if (ktime_get_real_seconds() >
1183 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
1185 * The evictor won't evict anyone who we've heard from
1186 * recently, so we don't have to check before we start
1189 if (!ping_evictor_wake(exp))
1190 exp->exp_obd->obd_eviction_timer = 0;
1198 * Sanity check request \a req.
1199 * Return 0 if all is ok, error code otherwise.
1201 static int ptlrpc_check_req(struct ptlrpc_request *req)
1203 struct obd_device *obd = req->rq_export->exp_obd;
1206 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
1207 req->rq_export->exp_conn_cnt)) {
1208 DEBUG_REQ(D_RPCTRACE, req,
1209 "DROPPING req from old connection %d < %d",
1210 lustre_msg_get_conn_cnt(req->rq_reqmsg),
1211 req->rq_export->exp_conn_cnt);
1214 if (unlikely(obd == NULL || obd->obd_fail)) {
1216 * Failing over, don't handle any more reqs,
1217 * send error response instead.
1219 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
1220 req, (obd != NULL) ? obd->obd_name : "unknown");
1222 } else if (lustre_msg_get_flags(req->rq_reqmsg) &
1223 (MSG_REPLAY | MSG_REQ_REPLAY_DONE) &&
1224 !obd->obd_recovering) {
1225 DEBUG_REQ(D_ERROR, req,
1226 "Invalid replay without recovery");
1227 class_fail_export(req->rq_export);
1229 } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 &&
1230 !obd->obd_recovering) {
1231 DEBUG_REQ(D_ERROR, req,
1232 "Invalid req with transno %llu without recovery",
1233 lustre_msg_get_transno(req->rq_reqmsg));
1234 class_fail_export(req->rq_export);
1238 if (unlikely(rc < 0)) {
1239 req->rq_status = rc;
1245 static void ptlrpc_at_set_timer(struct ptlrpc_service_part *svcpt)
1247 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1250 if (array->paa_count == 0) {
1251 del_timer(&svcpt->scp_at_timer);
1255 /* Set timer for closest deadline */
1256 next = array->paa_deadline - ktime_get_real_seconds() -
1259 ptlrpc_at_timer(cfs_timer_cb_arg(svcpt, scp_at_timer));
1261 mod_timer(&svcpt->scp_at_timer,
1262 jiffies + nsecs_to_jiffies(next * NSEC_PER_SEC));
1263 CDEBUG(D_INFO, "armed %s at %+llds\n",
1264 svcpt->scp_service->srv_name, next);
1268 /* Add rpc to early reply check list */
1269 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
1271 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1272 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1273 struct ptlrpc_request *rq = NULL;
1279 if (req->rq_no_reply)
1282 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
1285 spin_lock(&svcpt->scp_at_lock);
1286 LASSERT(list_empty(&req->rq_timed_list));
1288 div_u64_rem(req->rq_deadline, array->paa_size, &index);
1289 if (array->paa_reqs_count[index] > 0) {
1291 * latest rpcs will have the latest deadlines in the list,
1292 * so search backward.
1294 list_for_each_entry_reverse(rq, &array->paa_reqs_array[index],
1296 if (req->rq_deadline >= rq->rq_deadline) {
1297 list_add(&req->rq_timed_list,
1298 &rq->rq_timed_list);
1304 /* Add the request at the head of the list */
1305 if (list_empty(&req->rq_timed_list))
1306 list_add(&req->rq_timed_list, &array->paa_reqs_array[index]);
1308 spin_lock(&req->rq_lock);
1309 req->rq_at_linked = 1;
1310 spin_unlock(&req->rq_lock);
1311 req->rq_at_index = index;
1312 array->paa_reqs_count[index]++;
1314 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
1315 array->paa_deadline = req->rq_deadline;
1316 ptlrpc_at_set_timer(svcpt);
1318 spin_unlock(&svcpt->scp_at_lock);
1323 static void ptlrpc_at_remove_timed(struct ptlrpc_request *req)
1325 struct ptlrpc_at_array *array;
1327 array = &req->rq_rqbd->rqbd_svcpt->scp_at_array;
1329 /* NB: must call with hold svcpt::scp_at_lock */
1330 LASSERT(!list_empty(&req->rq_timed_list));
1331 list_del_init(&req->rq_timed_list);
1333 spin_lock(&req->rq_lock);
1334 req->rq_at_linked = 0;
1335 spin_unlock(&req->rq_lock);
1337 array->paa_reqs_count[req->rq_at_index]--;
1342 * Attempt to extend the request deadline by sending an early reply to the
1345 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
1347 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1348 struct ptlrpc_request *reqcopy;
1349 struct lustre_msg *reqmsg;
1350 timeout_t olddl = req->rq_deadline - ktime_get_real_seconds();
1356 if (CFS_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_RECONNECT) ||
1357 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_ENQ_RESEND)) {
1358 /* don't send early reply */
1363 * deadline is when the client expects us to reply, margin is the
1364 * difference between clients' and servers' expectations
1366 DEBUG_REQ(D_ADAPTTO, req,
1367 "%ssending early reply (deadline %+ds, margin %+ds) for %d+%d",
1368 AT_OFF ? "AT off - not " : "",
1369 olddl, olddl - at_get(&svcpt->scp_at_estimate),
1370 at_get(&svcpt->scp_at_estimate), at_extra);
1376 /* below message is checked in replay-ost-single.sh test_9 */
1377 DEBUG_REQ(D_WARNING, req,
1378 "Already past deadline (%+ds), not sending early reply. Consider increasing at_early_margin (%d)?",
1379 olddl, at_early_margin);
1381 /* Return an error so we're not re-added to the timed list. */
1385 if ((lustre_msghdr_get_flags(req->rq_reqmsg) &
1386 MSGHDR_AT_SUPPORT) == 0) {
1387 DEBUG_REQ(D_INFO, req,
1388 "Wanted to ask client for more time, but no AT support");
1392 if (req->rq_export &&
1393 lustre_msg_get_flags(req->rq_reqmsg) &
1394 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1395 struct obd_device *obd_exp = req->rq_export->exp_obd;
1398 * During recovery, we don't want to send too many early
1399 * replies, but on the other hand we want to make sure the
1400 * client has enough time to resend if the rpc is lost. So
1401 * during the recovery period send at least 4 early replies,
1402 * spacing them every at_extra if we can. at_estimate should
1403 * always equal this fixed value during recovery.
1407 * Don't account request processing time into AT history
1408 * during recovery, it is not service time we need but
1409 * includes also waiting time for recovering clients
1411 newdl = min_t(time64_t, at_extra,
1412 obd_exp->obd_recovery_timeout / 4) +
1413 ktime_get_real_seconds();
1416 * We want to extend the request deadline by at_extra seconds,
1417 * so we set our service estimate to reflect how much time has
1418 * passed since this request arrived plus an additional
1419 * at_extra seconds. The client will calculate the new deadline
1420 * based on this service estimate (plus some additional time to
1421 * account for network latency). See ptlrpc_at_recv_early_reply
1423 at_measured(&svcpt->scp_at_estimate, at_extra +
1424 ktime_get_real_seconds() -
1425 req->rq_arrival_time.tv_sec);
1426 newdl = req->rq_arrival_time.tv_sec +
1427 at_get(&svcpt->scp_at_estimate);
1431 * Check to see if we've actually increased the deadline -
1432 * we may be past adaptive_max
1434 if (req->rq_deadline >= newdl) {
1435 DEBUG_REQ(D_WARNING, req,
1436 "Could not add any time (%d/%lld), not sending early reply",
1437 olddl, newdl - ktime_get_real_seconds());
1441 reqcopy = ptlrpc_request_cache_alloc(GFP_NOFS);
1442 if (reqcopy == NULL)
1444 OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1446 GOTO(out_free, rc = -ENOMEM);
1449 reqcopy->rq_reply_state = NULL;
1450 reqcopy->rq_rep_swab_mask = 0;
1451 reqcopy->rq_pack_bulk = 0;
1452 reqcopy->rq_pack_udesc = 0;
1453 reqcopy->rq_packed_final = 0;
1454 sptlrpc_svc_ctx_addref(reqcopy);
1455 /* We only need the reqmsg for the magic */
1456 reqcopy->rq_reqmsg = reqmsg;
1457 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1460 * tgt_brw_read() and tgt_brw_write() may have decided not to reply.
1461 * Without this check, we would fail the rq_no_reply assertion in
1462 * ptlrpc_send_reply().
1464 if (reqcopy->rq_no_reply)
1465 GOTO(out, rc = -ETIMEDOUT);
1467 LASSERT(atomic_read(&req->rq_refcount));
1468 /* if it is last refcount then early reply isn't needed */
1469 if (atomic_read(&req->rq_refcount) == 1) {
1470 DEBUG_REQ(D_ADAPTTO, reqcopy,
1471 "Normal reply already sent, abort early reply");
1472 GOTO(out, rc = -EINVAL);
1475 /* Connection ref */
1476 reqcopy->rq_export = class_conn2export(
1477 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1478 if (reqcopy->rq_export == NULL)
1479 GOTO(out, rc = -ENODEV);
1482 class_export_rpc_inc(reqcopy->rq_export);
1483 if (reqcopy->rq_export->exp_obd &&
1484 reqcopy->rq_export->exp_obd->obd_fail)
1485 GOTO(out_put, rc = -ENODEV);
1487 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1491 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1494 /* Adjust our own deadline to what we told the client */
1495 req->rq_deadline = newdl;
1496 req->rq_early_count++; /* number sent, server side */
1498 DEBUG_REQ(D_ERROR, req, "Early reply send failed: rc = %d", rc);
1502 * Free the (early) reply state from lustre_pack_reply.
1503 * (ptlrpc_send_reply takes it's own rs ref, so this is safe here)
1505 ptlrpc_req_drop_rs(reqcopy);
1508 class_export_rpc_dec(reqcopy->rq_export);
1509 class_export_put(reqcopy->rq_export);
1511 sptlrpc_svc_ctx_decref(reqcopy);
1512 OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1514 ptlrpc_request_cache_free(reqcopy);
1519 * Send early replies to everybody expiring within at_early_margin
1520 * asking for at_extra time
1522 static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1524 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1525 struct ptlrpc_request *rq, *n;
1526 LIST_HEAD(work_list);
1529 time64_t now = ktime_get_real_seconds();
1531 int first, counter = 0;
1534 spin_lock(&svcpt->scp_at_lock);
1535 if (svcpt->scp_at_check == 0) {
1536 spin_unlock(&svcpt->scp_at_lock);
1539 delay_ms = ktime_ms_delta(ktime_get(), svcpt->scp_at_checktime);
1540 svcpt->scp_at_check = 0;
1542 if (array->paa_count == 0) {
1543 spin_unlock(&svcpt->scp_at_lock);
1547 /* The timer went off, but maybe the nearest rpc already completed. */
1548 first = array->paa_deadline - now;
1549 if (first > at_early_margin) {
1550 /* We've still got plenty of time. Reset the timer. */
1551 ptlrpc_at_set_timer(svcpt);
1552 spin_unlock(&svcpt->scp_at_lock);
1557 * We're close to a timeout, and we don't know how much longer the
1558 * server will take. Send early replies to everyone expiring soon.
1561 div_u64_rem(array->paa_deadline, array->paa_size, &index);
1562 count = array->paa_count;
1564 count -= array->paa_reqs_count[index];
1565 list_for_each_entry_safe(rq, n,
1566 &array->paa_reqs_array[index],
1568 if (rq->rq_deadline > now + at_early_margin) {
1569 /* update the earliest deadline */
1570 if (deadline == -1 ||
1571 rq->rq_deadline < deadline)
1572 deadline = rq->rq_deadline;
1577 * ptlrpc_server_drop_request() may drop
1578 * refcount to 0 already. Let's check this and
1579 * don't add entry to work_list
1581 if (likely(atomic_inc_not_zero(&rq->rq_refcount))) {
1582 ptlrpc_at_remove_timed(rq);
1583 list_add(&rq->rq_timed_list, &work_list);
1585 ptlrpc_at_remove_timed(rq);
1591 if (++index >= array->paa_size)
1594 array->paa_deadline = deadline;
1595 /* we have a new earliest deadline, restart the timer */
1596 ptlrpc_at_set_timer(svcpt);
1598 spin_unlock(&svcpt->scp_at_lock);
1601 "timeout in %+ds, asking for %d secs on %d early replies\n",
1602 first, at_extra, counter);
1605 * We're already past request deadlines before we even get a
1606 * chance to send early replies
1608 LCONSOLE_WARN("%s: This server is not able to keep up with request traffic (cpu-bound).\n",
1609 svcpt->scp_service->srv_name);
1610 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, delay=%lldms\n",
1611 counter, svcpt->scp_nreqs_incoming,
1612 svcpt->scp_nreqs_active,
1613 at_get(&svcpt->scp_at_estimate), delay_ms);
1617 * we took additional refcount so entries can't be deleted from list, no
1620 while ((rq = list_first_entry_or_null(&work_list,
1621 struct ptlrpc_request,
1622 rq_timed_list)) != NULL) {
1623 list_del_init(&rq->rq_timed_list);
1625 if (ptlrpc_at_send_early_reply(rq) == 0)
1626 ptlrpc_at_add_timed(rq);
1628 ptlrpc_server_drop_request(rq);
1631 RETURN(1); /* return "did_something" for liblustre */
1635 * Check if we are already handling earlier incarnation of this request.
1636 * Called under &req->rq_export->exp_rpc_lock locked
1638 static struct ptlrpc_request*
1639 ptlrpc_server_check_resend_in_progress(struct ptlrpc_request *req)
1641 struct ptlrpc_request *tmp = NULL;
1643 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
1647 * This list should not be longer than max_requests in
1648 * flights on the client, so it is not all that long.
1649 * Also we only hit this codepath in case of a resent
1650 * request which makes it even more rarely hit
1652 list_for_each_entry(tmp, &req->rq_export->exp_reg_rpcs,
1654 /* Found duplicate one */
1655 if (tmp->rq_xid == req->rq_xid)
1658 list_for_each_entry(tmp, &req->rq_export->exp_hp_rpcs,
1660 /* Found duplicate one */
1661 if (tmp->rq_xid == req->rq_xid)
1667 DEBUG_REQ(D_HA, req, "Found duplicate req in processing");
1668 DEBUG_REQ(D_HA, tmp, "Request being processed");
1672 #ifdef HAVE_SERVER_SUPPORT
1673 static void ptlrpc_server_mark_obsolete(struct ptlrpc_request *req)
1675 req->rq_obsolete = 1;
1679 ptlrpc_server_mark_in_progress_obsolete(struct ptlrpc_request *req)
1681 struct ptlrpc_request *tmp = NULL;
1684 if (!tgt_is_increasing_xid_client(req->rq_export) ||
1685 req->rq_export->exp_used_slots == NULL)
1688 tag = lustre_msg_get_tag(req->rq_reqmsg);
1692 if (!test_bit(tag - 1, req->rq_export->exp_used_slots))
1695 /* This list should not be longer than max_requests in
1696 * flights on the client, so it is not all that long.
1697 * Also we only hit this codepath in case of a resent
1698 * request which makes it even more rarely hit */
1699 list_for_each_entry(tmp, &req->rq_export->exp_reg_rpcs, rq_exp_list) {
1700 if (tag == lustre_msg_get_tag(tmp->rq_reqmsg) &&
1701 req->rq_xid > tmp->rq_xid)
1702 ptlrpc_server_mark_obsolete(tmp);
1705 list_for_each_entry(tmp, &req->rq_export->exp_hp_rpcs, rq_exp_list) {
1706 if (tag == lustre_msg_get_tag(tmp->rq_reqmsg) &&
1707 req->rq_xid > tmp->rq_xid)
1708 ptlrpc_server_mark_obsolete(tmp);
1714 * Check if a request should be assigned with a high priority.
1716 * \retval < 0: error occurred
1717 * 0: normal RPC request
1718 * +1: high priority request
1720 static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt,
1721 struct ptlrpc_request *req)
1726 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL) {
1727 rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
1734 if (req->rq_export != NULL && req->rq_ops != NULL) {
1736 * Perform request specific check. We should do this
1737 * check before the request is added into exp_hp_rpcs
1738 * list otherwise it may hit swab race at LU-1044.
1740 if (req->rq_ops->hpreq_check != NULL) {
1741 rc = req->rq_ops->hpreq_check(req);
1742 if (rc == -ESTALE) {
1743 req->rq_status = rc;
1747 * can only return error,
1748 * 0 for normal request,
1749 * or 1 for high priority request
1758 /** Remove the request from the export list. */
1759 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req)
1762 if (req->rq_export) {
1764 * refresh lock timeout again so that client has more
1765 * room to send lock cancel RPC.
1767 if (req->rq_ops && req->rq_ops->hpreq_fini)
1768 req->rq_ops->hpreq_fini(req);
1770 ptlrpc_del_exp_list(req);
1775 static int ptlrpc_hpreq_check(struct ptlrpc_request *req)
1780 static struct ptlrpc_hpreq_ops ptlrpc_hpreq_common = {
1781 .hpreq_check = ptlrpc_hpreq_check,
1784 /* Hi-Priority RPC check by RPC operation code. */
1785 int ptlrpc_hpreq_handler(struct ptlrpc_request *req)
1787 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1790 * Check for export to let only reconnects for not yet evicted
1791 * export to become a HP rpc.
1793 if ((req->rq_export != NULL) &&
1794 (opc == OBD_PING || opc == MDS_CONNECT || opc == OST_CONNECT))
1795 req->rq_ops = &ptlrpc_hpreq_common;
1799 EXPORT_SYMBOL(ptlrpc_hpreq_handler);
1801 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1802 struct ptlrpc_request *req)
1806 struct ptlrpc_request *orig;
1810 rc = ptlrpc_server_hpreq_init(svcpt, req);
1815 ptlrpc_nrs_req_initialize(svcpt, req, hp);
1817 while (req->rq_export != NULL) {
1818 struct obd_export *exp = req->rq_export;
1821 * do search for duplicated xid and the adding to the list
1824 spin_lock_bh(&exp->exp_rpc_lock);
1825 #ifdef HAVE_SERVER_SUPPORT
1826 ptlrpc_server_mark_in_progress_obsolete(req);
1828 orig = ptlrpc_server_check_resend_in_progress(req);
1829 if (orig && OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_RESEND_RACE)) {
1830 spin_unlock_bh(&exp->exp_rpc_lock);
1832 OBD_RACE(OBD_FAIL_PTLRPC_RESEND_RACE);
1833 msleep(4 * MSEC_PER_SEC);
1837 if (orig && likely(atomic_inc_not_zero(&orig->rq_refcount))) {
1840 spin_unlock_bh(&exp->exp_rpc_lock);
1843 * When the client resend request and the server has
1844 * the previous copy of it, we need to update deadlines,
1845 * to be sure that the client and the server have equal
1846 * request deadlines.
1849 spin_lock(&orig->rq_rqbd->rqbd_svcpt->scp_at_lock);
1850 linked = orig->rq_at_linked;
1852 ptlrpc_at_remove_timed(orig);
1853 spin_unlock(&orig->rq_rqbd->rqbd_svcpt->scp_at_lock);
1854 orig->rq_deadline = req->rq_deadline;
1855 orig->rq_rep_mbits = req->rq_rep_mbits;
1857 ptlrpc_at_add_timed(orig);
1858 ptlrpc_server_drop_request(orig);
1859 ptlrpc_nrs_req_finalize(req);
1861 /* don't mark slot unused for resend in progress */
1862 req->rq_obsolete = 1;
1867 ptlrpc_add_exp_list_nolock(req, exp, hp || req->rq_ops != NULL);
1869 spin_unlock_bh(&exp->exp_rpc_lock);
1874 * the current thread is not the processing thread for this request
1875 * since that, but request is in exp_hp_list and can be find there.
1876 * Remove all relations between request and old thread.
1878 req->rq_svc_thread->t_env->le_ses = NULL;
1879 req->rq_svc_thread = NULL;
1880 req->rq_session.lc_thread = NULL;
1882 ptlrpc_nrs_req_add(svcpt, req, hp);
1888 * Allow to handle high priority request
1889 * User can call it w/o any lock but need to hold
1890 * ptlrpc_service_part::scp_req_lock to get reliable result
1892 static bool ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1895 int running = svcpt->scp_nthrs_running;
1897 if (!nrs_svcpt_has_hp(svcpt))
1903 if (ptlrpc_nrs_req_throttling_nolock(svcpt, true))
1906 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1907 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1908 /* leave just 1 thread for normal RPCs */
1909 running = PTLRPC_NTHRS_INIT;
1910 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1914 if (svcpt->scp_nreqs_active >= running - 1)
1917 if (svcpt->scp_nhreqs_active == 0)
1920 return !ptlrpc_nrs_req_pending_nolock(svcpt, false) ||
1921 svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1924 static bool ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1927 return ptlrpc_server_allow_high(svcpt, force) &&
1928 ptlrpc_nrs_req_pending_nolock(svcpt, true);
1932 * Only allow normal priority requests on a service that has a high-priority
1933 * queue if forced (i.e. cleanup), if there are other high priority requests
1934 * already being processed (i.e. those threads can service more high-priority
1935 * requests), or if there are enough idle threads that a later thread can do
1936 * a high priority request.
1937 * User can call it w/o any lock but need to hold
1938 * ptlrpc_service_part::scp_req_lock to get reliable result
1940 static bool ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1943 int running = svcpt->scp_nthrs_running;
1945 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1946 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1947 /* leave just 1 thread for normal RPCs */
1948 running = PTLRPC_NTHRS_INIT;
1949 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1956 if (ptlrpc_nrs_req_throttling_nolock(svcpt, false))
1959 if (svcpt->scp_nreqs_active < running - 2)
1962 if (svcpt->scp_nreqs_active >= running - 1)
1965 return svcpt->scp_nhreqs_active > 0 || !nrs_svcpt_has_hp(svcpt);
1968 static bool ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1971 return ptlrpc_server_allow_normal(svcpt, force) &&
1972 ptlrpc_nrs_req_pending_nolock(svcpt, false);
1976 * Returns true if there are requests available in incoming
1977 * request queue for processing and it is allowed to fetch them.
1978 * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1979 * to get reliable result
1980 * \see ptlrpc_server_allow_normal
1981 * \see ptlrpc_server_allow high
1984 bool ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt,
1987 return ptlrpc_server_high_pending(svcpt, force) ||
1988 ptlrpc_server_normal_pending(svcpt, force);
1992 * Fetch a request for processing from queue of unprocessed requests.
1993 * Favors high-priority requests.
1994 * Returns a pointer to fetched request.
1996 static struct ptlrpc_request *
1997 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force)
1999 struct ptlrpc_request *req = NULL;
2003 spin_lock(&svcpt->scp_req_lock);
2005 if (ptlrpc_server_high_pending(svcpt, force)) {
2006 req = ptlrpc_nrs_req_get_nolock(svcpt, true, force);
2008 svcpt->scp_hreq_count++;
2013 if (ptlrpc_server_normal_pending(svcpt, force)) {
2014 req = ptlrpc_nrs_req_get_nolock(svcpt, false, force);
2016 svcpt->scp_hreq_count = 0;
2021 spin_unlock(&svcpt->scp_req_lock);
2025 svcpt->scp_nreqs_active++;
2027 svcpt->scp_nhreqs_active++;
2029 spin_unlock(&svcpt->scp_req_lock);
2031 if (likely(req->rq_export))
2032 class_export_rpc_inc(req->rq_export);
2038 * Handle freshly incoming reqs, add to timed early reply list,
2039 * pass on to regular request queue.
2040 * All incoming requests pass through here before getting into
2041 * ptlrpc_server_handle_req later on.
2043 static int ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt,
2044 struct ptlrpc_thread *thread)
2046 struct ptlrpc_service *svc = svcpt->scp_service;
2047 struct ptlrpc_request *req;
2054 spin_lock(&svcpt->scp_lock);
2055 if (list_empty(&svcpt->scp_req_incoming)) {
2056 spin_unlock(&svcpt->scp_lock);
2060 req = list_first_entry(&svcpt->scp_req_incoming,
2061 struct ptlrpc_request, rq_list);
2062 list_del_init(&req->rq_list);
2063 svcpt->scp_nreqs_incoming--;
2065 * Consider this still a "queued" request as far as stats are
2068 spin_unlock(&svcpt->scp_lock);
2070 /* go through security check/transform */
2071 rc = sptlrpc_svc_unwrap_request(req);
2075 case SECSVC_COMPLETE:
2076 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
2085 * for null-flavored rpc, msg has been unpacked by sptlrpc, although
2086 * redo it wouldn't be harmful.
2088 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
2089 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
2091 CERROR("error unpacking request: ptl %d from %s x%llu\n",
2092 svc->srv_req_portal, libcfs_id2str(req->rq_peer),
2098 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
2100 CERROR("error unpacking ptlrpc body: ptl %d from %s x %llu\n",
2101 svc->srv_req_portal, libcfs_id2str(req->rq_peer),
2106 opc = lustre_msg_get_opc(req->rq_reqmsg);
2107 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
2108 opc == cfs_fail_val) {
2109 CERROR("drop incoming rpc opc %u, x%llu\n",
2110 cfs_fail_val, req->rq_xid);
2115 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
2116 CERROR("wrong packet type received (type=%u) from %s\n",
2117 lustre_msg_get_type(req->rq_reqmsg),
2118 libcfs_id2str(req->rq_peer));
2126 req->rq_bulk_write = 1;
2130 case MGS_CONFIG_READ:
2131 req->rq_bulk_read = 1;
2135 CDEBUG(D_RPCTRACE, "got req x%llu\n", req->rq_xid);
2137 req->rq_export = class_conn2export(
2138 lustre_msg_get_handle(req->rq_reqmsg));
2139 if (req->rq_export) {
2140 rc = ptlrpc_check_req(req);
2142 rc = sptlrpc_target_export_check(req->rq_export, req);
2144 DEBUG_REQ(D_ERROR, req,
2145 "DROPPING req with illegal security flavor");
2150 ptlrpc_update_export_timer(req->rq_export, 0);
2153 /* req_in handling should/must be fast */
2154 if (ktime_get_real_seconds() - req->rq_arrival_time.tv_sec > 5)
2155 DEBUG_REQ(D_WARNING, req, "Slow req_in handling %llds",
2156 ktime_get_real_seconds() -
2157 req->rq_arrival_time.tv_sec);
2159 /* Set rpc server deadline and add it to the timed list */
2160 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
2161 MSGHDR_AT_SUPPORT) ?
2162 /* The max time the client expects us to take */
2163 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
2165 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
2166 if (unlikely(deadline == 0)) {
2167 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
2171 /* Skip early reply */
2172 if (OBD_FAIL_PRECHECK(OBD_FAIL_MDS_RESEND))
2173 req->rq_deadline += obd_timeout;
2175 req->rq_svc_thread = thread;
2176 if (thread != NULL) {
2178 * initialize request session, it is needed for request
2179 * processing by target
2181 rc = lu_context_init(&req->rq_session, LCT_SERVER_SESSION |
2184 CERROR("%s: failure to initialize session: rc = %d\n",
2185 thread->t_name, rc);
2188 req->rq_session.lc_thread = thread;
2189 lu_context_enter(&req->rq_session);
2190 thread->t_env->le_ses = &req->rq_session;
2194 if (unlikely(OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_ENQ_RESEND) &&
2195 (opc == LDLM_ENQUEUE) &&
2196 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT)))
2197 OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_ENQ_RESEND, 6);
2199 ptlrpc_at_add_timed(req);
2201 if (opc != OST_CONNECT && opc != MDS_CONNECT &&
2202 opc != MGS_CONNECT && req->rq_export != NULL) {
2203 if (exp_connect_flags2(req->rq_export) & OBD_CONNECT2_REP_MBITS)
2204 req->rq_rep_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
2207 /* Move it over to the request processing queue */
2208 rc = ptlrpc_server_request_add(svcpt, req);
2212 wake_up(&svcpt->scp_waitq);
2216 ptlrpc_server_finish_request(svcpt, req);
2222 * Main incoming request handling logic.
2223 * Calls handler function from service to do actual processing.
2225 static int ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
2226 struct ptlrpc_thread *thread)
2228 struct ptlrpc_service *svc = svcpt->scp_service;
2229 struct ptlrpc_request *request;
2239 request = ptlrpc_server_request_get(svcpt, false);
2240 if (request == NULL)
2243 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
2244 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
2245 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
2246 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
2248 if (unlikely(fail_opc)) {
2249 if (request->rq_export && request->rq_ops)
2250 OBD_FAIL_TIMEOUT(fail_opc, 4);
2253 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
2255 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
2256 libcfs_debug_dumplog();
2258 work_start = ktime_get_real();
2259 arrived = timespec64_to_ktime(request->rq_arrival_time);
2260 timediff_usecs = ktime_us_delta(work_start, arrived);
2261 if (likely(svc->srv_stats != NULL)) {
2262 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
2264 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
2265 svcpt->scp_nreqs_incoming);
2266 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
2267 svcpt->scp_nreqs_active);
2268 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
2269 at_get(&svcpt->scp_at_estimate));
2272 if (likely(request->rq_export)) {
2273 if (unlikely(ptlrpc_check_req(request)))
2275 ptlrpc_update_export_timer(request->rq_export,
2276 div_u64(timediff_usecs,
2281 * Discard requests queued for longer than the deadline.
2282 * The deadline is increased if we send an early reply.
2284 if (ktime_get_real_seconds() > request->rq_deadline) {
2285 DEBUG_REQ(D_ERROR, request,
2286 "Dropping timed-out request from %s: deadline %lld/%llds ago",
2287 libcfs_id2str(request->rq_peer),
2288 request->rq_deadline -
2289 request->rq_arrival_time.tv_sec,
2290 ktime_get_real_seconds() - request->rq_deadline);
2295 "Handling RPC req@%p pname:cluuid+ref:pid:xid:nid:opc:job %s:%s+%d:%d:x%llu:%s:%d:%s\n",
2296 request, current->comm,
2297 (request->rq_export ?
2298 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2299 (request->rq_export ?
2300 refcount_read(&request->rq_export->exp_handle.h_ref) : -99),
2301 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
2302 libcfs_id2str(request->rq_peer),
2303 lustre_msg_get_opc(request->rq_reqmsg),
2304 lustre_msg_get_jobid(request->rq_reqmsg) ?: "");
2306 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
2307 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
2309 CDEBUG(D_NET, "got req %llu\n", request->rq_xid);
2311 /* re-assign request and sesson thread to the current one */
2312 request->rq_svc_thread = thread;
2313 if (thread != NULL) {
2314 LASSERT(request->rq_session.lc_thread == NULL);
2315 request->rq_session.lc_thread = thread;
2316 thread->t_env->le_ses = &request->rq_session;
2318 svc->srv_ops.so_req_handler(request);
2320 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
2323 if (unlikely(ktime_get_real_seconds() > request->rq_deadline)) {
2324 DEBUG_REQ(D_WARNING, request,
2325 "Request took longer than estimated (%lld/%llds); client may timeout",
2326 request->rq_deadline -
2327 request->rq_arrival_time.tv_sec,
2328 ktime_get_real_seconds() - request->rq_deadline);
2331 work_end = ktime_get_real();
2332 timediff_usecs = ktime_us_delta(work_end, work_start);
2333 arrived_usecs = ktime_us_delta(work_end, arrived);
2335 "Handled RPC req@%p pname:cluuid+ref:pid:xid:nid:opc:job %s:%s+%d:%d:x%llu:%s:%d:%s Request processed in %lldus (%lldus total) trans %llu rc %d/%d\n",
2336 request, current->comm,
2337 (request->rq_export ?
2338 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2339 (request->rq_export ?
2340 refcount_read(&request->rq_export->exp_handle.h_ref) : -99),
2341 lustre_msg_get_status(request->rq_reqmsg),
2343 libcfs_id2str(request->rq_peer),
2344 lustre_msg_get_opc(request->rq_reqmsg),
2345 lustre_msg_get_jobid(request->rq_reqmsg) ?: "",
2348 (request->rq_repmsg ?
2349 lustre_msg_get_transno(request->rq_repmsg) :
2350 request->rq_transno),
2352 (request->rq_repmsg ?
2353 lustre_msg_get_status(request->rq_repmsg) : -999));
2354 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
2355 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
2356 int opc = opcode_offset(op);
2358 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
2359 LASSERT(opc < LUSTRE_MAX_OPCODES);
2360 lprocfs_counter_add(svc->srv_stats,
2361 opc + EXTRA_MAX_OPCODES,
2365 if (unlikely(request->rq_early_count)) {
2366 DEBUG_REQ(D_ADAPTTO, request,
2367 "sent %d early replies before finishing in %llds",
2368 request->rq_early_count,
2369 div_u64(arrived_usecs, USEC_PER_SEC));
2372 ptlrpc_server_finish_active_request(svcpt, request);
2378 * An internal function to process a single reply state object.
2380 static int ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
2382 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
2383 struct ptlrpc_service *svc = svcpt->scp_service;
2384 struct obd_export *exp;
2390 exp = rs->rs_export;
2392 LASSERT(rs->rs_difficult);
2393 LASSERT(rs->rs_scheduled);
2394 LASSERT(list_empty(&rs->rs_list));
2397 * The disk commit callback holds exp_uncommitted_replies_lock while it
2398 * iterates over newly committed replies, removing them from
2399 * exp_uncommitted_replies. It then drops this lock and schedules the
2400 * replies it found for handling here.
2402 * We can avoid contention for exp_uncommitted_replies_lock between the
2403 * HRT threads and further commit callbacks by checking rs_committed
2404 * which is set in the commit callback while it holds both
2405 * rs_lock and exp_uncommitted_reples.
2407 * If we see rs_committed clear, the commit callback _may_ not have
2408 * handled this reply yet and we race with it to grab
2409 * exp_uncommitted_replies_lock before removing the reply from
2410 * exp_uncommitted_replies. Note that if we lose the race and the
2411 * reply has already been removed, list_del_init() is a noop.
2413 * If we see rs_committed set, we know the commit callback is handling,
2414 * or has handled this reply since store reordering might allow us to
2415 * see rs_committed set out of sequence. But since this is done
2416 * holding rs_lock, we can be sure it has all completed once we hold
2417 * rs_lock, which we do right next.
2419 if (!rs->rs_committed) {
2421 * if rs was commited, no need to convert locks, don't check
2422 * rs_committed here because rs may never be added into
2423 * exp_uncommitted_replies and this flag never be set, see
2424 * target_send_reply()
2426 if (rs->rs_convert_lock &&
2427 rs->rs_transno > exp->exp_last_committed) {
2428 struct ldlm_lock *lock;
2429 struct ldlm_lock *ack_locks[RS_MAX_LOCKS] = { NULL };
2431 spin_lock(&rs->rs_lock);
2432 if (rs->rs_convert_lock &&
2433 rs->rs_transno > exp->exp_last_committed) {
2434 nlocks = rs->rs_nlocks;
2435 while (nlocks-- > 0) {
2437 * NB don't assume rs is always handled
2438 * by the same service thread (see
2439 * ptlrpc_hr_select, so REP-ACK hr may
2440 * race with trans commit, while the
2441 * latter will release locks, get locks
2442 * here early to convert to COS mode
2445 lock = ldlm_handle2lock(
2446 &rs->rs_locks[nlocks]);
2448 ack_locks[nlocks] = lock;
2449 rs->rs_modes[nlocks] = LCK_COS;
2451 nlocks = rs->rs_nlocks;
2452 rs->rs_convert_lock = 0;
2454 * clear rs_scheduled so that commit callback
2455 * can schedule again
2457 rs->rs_scheduled = 0;
2458 spin_unlock(&rs->rs_lock);
2460 while (nlocks-- > 0) {
2461 lock = ack_locks[nlocks];
2462 ldlm_lock_mode_downgrade(lock, LCK_COS);
2463 LDLM_LOCK_PUT(lock);
2467 spin_unlock(&rs->rs_lock);
2470 spin_lock(&exp->exp_uncommitted_replies_lock);
2471 list_del_init(&rs->rs_obd_list);
2472 spin_unlock(&exp->exp_uncommitted_replies_lock);
2475 spin_lock(&exp->exp_lock);
2476 /* Noop if removed already */
2477 list_del_init(&rs->rs_exp_list);
2478 spin_unlock(&exp->exp_lock);
2480 spin_lock(&rs->rs_lock);
2482 been_handled = rs->rs_handled;
2485 nlocks = rs->rs_nlocks; /* atomic "steal", but */
2486 rs->rs_nlocks = 0; /* locks still on rs_locks! */
2488 if (nlocks == 0 && !been_handled) {
2490 * If we see this, we should already have seen the warning
2491 * in mds_steal_ack_locks()
2494 "All locks stolen from rs %p x%lld.t%lld o%d NID %s\n",
2495 rs, rs->rs_xid, rs->rs_transno, rs->rs_opc,
2496 libcfs_nidstr(&exp->exp_connection->c_peer.nid));
2499 if ((rs->rs_sent && !rs->rs_unlinked) || nlocks > 0) {
2500 spin_unlock(&rs->rs_lock);
2502 /* We can unlink if the LNET_EVENT_SEND has occurred.
2503 * If rs_unlinked is set then MD is already unlinked and no
2504 * need to do so here.
2506 if ((rs->rs_sent && !rs->rs_unlinked)) {
2507 LNetMDUnlink(rs->rs_md_h);
2508 /* Ignore return code; we're racing with completion */
2511 while (nlocks-- > 0)
2512 ldlm_lock_decref(&rs->rs_locks[nlocks],
2513 rs->rs_modes[nlocks]);
2515 spin_lock(&rs->rs_lock);
2518 rs->rs_scheduled = 0;
2519 rs->rs_convert_lock = 0;
2521 if (rs->rs_unlinked) {
2523 spin_unlock(&rs->rs_lock);
2525 class_export_put(exp);
2526 rs->rs_export = NULL;
2527 ptlrpc_rs_decref(rs);
2528 if (atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
2529 svc->srv_is_stopping)
2530 wake_up_all(&svcpt->scp_waitq);
2534 /* still on the net; callback will schedule */
2535 spin_unlock(&rs->rs_lock);
2540 static void ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
2542 int avail = svcpt->scp_nrqbds_posted;
2543 int low_water = test_req_buffer_pressure ? 0 :
2544 svcpt->scp_service->srv_nbuf_per_group / 2;
2546 /* NB I'm not locking; just looking. */
2549 * CAVEAT EMPTOR: We might be allocating buffers here because we've
2550 * allowed the request history to grow out of control. We could put a
2551 * sanity check on that here and cull some history if we need the
2555 if (avail <= low_water)
2556 ptlrpc_grow_req_bufs(svcpt, 1);
2558 if (svcpt->scp_service->srv_stats) {
2559 lprocfs_counter_add(svcpt->scp_service->srv_stats,
2560 PTLRPC_REQBUF_AVAIL_CNTR, avail);
2564 static inline int ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
2566 return svcpt->scp_nreqs_active <
2567 svcpt->scp_nthrs_running - 1 -
2568 (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL);
2572 * allowed to create more threads
2573 * user can call it w/o any lock but need to hold
2574 * ptlrpc_service_part::scp_lock to get reliable result
2576 static inline int ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
2578 return svcpt->scp_nthrs_running +
2579 svcpt->scp_nthrs_starting <
2580 svcpt->scp_service->srv_nthrs_cpt_limit;
2584 * too many requests and allowed to create more threads
2586 static inline int ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
2588 return !ptlrpc_threads_enough(svcpt) &&
2589 ptlrpc_threads_increasable(svcpt);
2592 static inline int ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2594 return thread_is_stopping(thread) ||
2595 thread->t_svcpt->scp_service->srv_is_stopping;
2598 /* stop the highest numbered thread if there are too many threads running */
2599 static inline bool ptlrpc_thread_should_stop(struct ptlrpc_thread *thread)
2601 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2603 return thread->t_id >= svcpt->scp_service->srv_nthrs_cpt_limit &&
2604 thread->t_id == svcpt->scp_thr_nextid - 1;
2607 static void ptlrpc_stop_thread(struct ptlrpc_thread *thread)
2609 CDEBUG(D_INFO, "Stopping thread %s #%u\n",
2610 thread->t_svcpt->scp_service->srv_thread_name, thread->t_id);
2611 thread_add_flags(thread, SVC_STOPPING);
2614 static inline void ptlrpc_thread_stop(struct ptlrpc_thread *thread)
2616 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2618 spin_lock(&svcpt->scp_lock);
2619 if (ptlrpc_thread_should_stop(thread)) {
2620 ptlrpc_stop_thread(thread);
2621 svcpt->scp_thr_nextid--;
2623 spin_unlock(&svcpt->scp_lock);
2626 static inline int ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
2628 return !list_empty(&svcpt->scp_rqbd_idle) &&
2629 svcpt->scp_rqbd_timeout == 0;
2633 ptlrpc_at_check(struct ptlrpc_service_part *svcpt)
2635 return svcpt->scp_at_check;
2639 * If a thread runs too long or spends to much time on a single request,
2640 * we want to know about it, so we set up a delayed work item as a watchdog.
2641 * If it fires, we display a stack trace of the delayed thread,
2642 * providing we aren't rate-limited
2644 * Watchdog stack traces are limited to 3 per 'libcfs_watchdog_ratelimit'
2647 static struct ratelimit_state watchdog_limit;
2649 static void ptlrpc_watchdog_fire(struct work_struct *w)
2651 struct ptlrpc_thread *thread = container_of(w, struct ptlrpc_thread,
2653 u64 ms_lapse = ktime_ms_delta(ktime_get(), thread->t_touched);
2654 u32 ms_frac = do_div(ms_lapse, MSEC_PER_SEC);
2656 /* ___ratelimit() returns true if the action is NOT ratelimited */
2657 if (__ratelimit(&watchdog_limit)) {
2658 /* below message is checked in sanity-quota.sh test_6,18 */
2659 LCONSOLE_WARN("%s: service thread pid %u was inactive for %llu.%03u seconds. The thread might be hung, or it might only be slow and will resume later. Dumping the stack trace for debugging purposes:\n",
2660 thread->t_task->comm, thread->t_task->pid,
2663 libcfs_debug_dumpstack(thread->t_task);
2665 /* below message is checked in sanity-quota.sh test_6,18 */
2666 LCONSOLE_WARN("%s: service thread pid %u was inactive for %llu.%03u seconds. Watchdog stack traces are limited to 3 per %u seconds, skipping this one.\n",
2667 thread->t_task->comm, thread->t_task->pid,
2668 ms_lapse, ms_frac, libcfs_watchdog_ratelimit);
2672 void ptlrpc_watchdog_init(struct delayed_work *work, timeout_t timeout)
2674 INIT_DELAYED_WORK(work, ptlrpc_watchdog_fire);
2675 schedule_delayed_work(work, cfs_time_seconds(timeout));
2678 void ptlrpc_watchdog_disable(struct delayed_work *work)
2680 cancel_delayed_work_sync(work);
2683 void ptlrpc_watchdog_touch(struct delayed_work *work, timeout_t timeout)
2685 struct ptlrpc_thread *thread = container_of(&work->work,
2686 struct ptlrpc_thread,
2688 thread->t_touched = ktime_get();
2689 mod_delayed_work(system_wq, work, cfs_time_seconds(timeout));
2693 * requests wait on preprocessing
2694 * user can call it w/o any lock but need to hold
2695 * ptlrpc_service_part::scp_lock to get reliable result
2698 ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
2700 return !list_empty(&svcpt->scp_req_incoming);
2703 static __attribute__((__noinline__)) int
2704 ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
2705 struct ptlrpc_thread *thread)
2707 ptlrpc_watchdog_disable(&thread->t_watchdog);
2711 if (svcpt->scp_rqbd_timeout == 0)
2712 /* Don't exit while there are replies to be handled */
2713 wait_event_idle_exclusive_lifo(
2715 ptlrpc_thread_stopping(thread) ||
2716 ptlrpc_server_request_incoming(svcpt) ||
2717 ptlrpc_server_request_pending(svcpt, false) ||
2718 ptlrpc_rqbd_pending(svcpt) ||
2719 ptlrpc_at_check(svcpt));
2720 else if (wait_event_idle_exclusive_lifo_timeout(
2722 ptlrpc_thread_stopping(thread) ||
2723 ptlrpc_server_request_incoming(svcpt) ||
2724 ptlrpc_server_request_pending(svcpt, false) ||
2725 ptlrpc_rqbd_pending(svcpt) ||
2726 ptlrpc_at_check(svcpt),
2727 svcpt->scp_rqbd_timeout) == 0)
2728 svcpt->scp_rqbd_timeout = 0;
2730 if (ptlrpc_thread_stopping(thread))
2733 ptlrpc_watchdog_touch(&thread->t_watchdog,
2734 ptlrpc_server_get_timeout(svcpt));
2739 * Main thread body for service threads.
2740 * Waits in a loop waiting for new requests to process to appear.
2741 * Every time an incoming requests is added to its queue, a waitq
2742 * is woken up and one of the threads will handle it.
2744 static int ptlrpc_main(void *arg)
2746 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
2747 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2748 struct ptlrpc_service *svc = svcpt->scp_service;
2749 struct ptlrpc_reply_state *rs;
2750 struct group_info *ginfo = NULL;
2752 int counter = 0, rc = 0;
2755 unshare_fs_struct();
2757 thread->t_task = current;
2758 thread->t_pid = current->pid;
2760 if (svc->srv_cpt_bind) {
2761 rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
2763 CWARN("%s: failed to bind %s on CPT %d\n",
2764 svc->srv_name, thread->t_name, svcpt->scp_cpt);
2768 ginfo = groups_alloc(0);
2770 GOTO(out, rc = -ENOMEM);
2772 set_current_groups(ginfo);
2773 put_group_info(ginfo);
2775 if (svc->srv_ops.so_thr_init != NULL) {
2776 rc = svc->srv_ops.so_thr_init(thread);
2783 GOTO(out_srv_fini, rc = -ENOMEM);
2784 rc = lu_env_add(env);
2788 rc = lu_context_init(&env->le_ctx,
2789 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2791 GOTO(out_env_remove, rc);
2793 thread->t_env = env;
2794 env->le_ctx.lc_thread = thread;
2795 env->le_ctx.lc_cookie = 0x6;
2797 while (!list_empty(&svcpt->scp_rqbd_idle)) {
2798 rc = ptlrpc_server_post_idle_rqbds(svcpt);
2802 CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
2803 svc->srv_name, svcpt->scp_cpt, rc);
2804 GOTO(out_ctx_fini, rc);
2807 /* Alloc reply state structure for this one */
2808 OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2810 GOTO(out_ctx_fini, rc = -ENOMEM);
2812 spin_lock(&svcpt->scp_lock);
2814 LASSERT(thread_is_starting(thread));
2815 thread_clear_flags(thread, SVC_STARTING);
2817 LASSERT(svcpt->scp_nthrs_starting == 1);
2818 svcpt->scp_nthrs_starting--;
2821 * SVC_STOPPING may already be set here if someone else is trying
2822 * to stop the service while this new thread has been dynamically
2823 * forked. We still set SVC_RUNNING to let our creator know that
2824 * we are now running, however we will exit as soon as possible
2826 thread_add_flags(thread, SVC_RUNNING);
2827 svcpt->scp_nthrs_running++;
2828 spin_unlock(&svcpt->scp_lock);
2830 /* wake up our creator in case he's still waiting. */
2831 wake_up(&thread->t_ctl_waitq);
2833 thread->t_touched = ktime_get();
2834 ptlrpc_watchdog_init(&thread->t_watchdog,
2835 ptlrpc_server_get_timeout(svcpt));
2837 spin_lock(&svcpt->scp_rep_lock);
2838 list_add(&rs->rs_list, &svcpt->scp_rep_idle);
2839 wake_up(&svcpt->scp_rep_waitq);
2840 spin_unlock(&svcpt->scp_rep_lock);
2842 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2843 svcpt->scp_nthrs_running);
2845 /* XXX maintain a list of all managed devices: insert here */
2846 while (!ptlrpc_thread_stopping(thread)) {
2847 if (ptlrpc_wait_event(svcpt, thread))
2850 ptlrpc_check_rqbd_pool(svcpt);
2852 if (ptlrpc_threads_need_create(svcpt)) {
2853 /* Ignore return code - we tried... */
2854 ptlrpc_start_thread(svcpt, 0);
2857 /* reset le_ses to initial state */
2859 /* Refill the context before execution to make sure
2860 * all thread keys are allocated */
2862 /* Process all incoming reqs before handling any */
2863 if (ptlrpc_server_request_incoming(svcpt)) {
2864 lu_context_enter(&env->le_ctx);
2865 ptlrpc_server_handle_req_in(svcpt, thread);
2866 lu_context_exit(&env->le_ctx);
2868 /* but limit ourselves in case of flood */
2869 if (counter++ < 100)
2874 if (ptlrpc_at_check(svcpt))
2875 ptlrpc_at_check_timed(svcpt);
2877 if (ptlrpc_server_request_pending(svcpt, false)) {
2878 lu_context_enter(&env->le_ctx);
2879 ptlrpc_server_handle_request(svcpt, thread);
2880 lu_context_exit(&env->le_ctx);
2883 if (ptlrpc_rqbd_pending(svcpt) &&
2884 ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
2886 * I just failed to repost request buffers.
2887 * Wait for a timeout (unless something else
2888 * happens) before I try again
2890 svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
2891 CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
2892 svcpt->scp_nrqbds_posted);
2895 * If the number of threads has been tuned downward and this
2896 * thread should be stopped, then stop in reverse order so the
2897 * the threads always have contiguous thread index values.
2899 if (unlikely(ptlrpc_thread_should_stop(thread)))
2900 ptlrpc_thread_stop(thread);
2903 ptlrpc_watchdog_disable(&thread->t_watchdog);
2906 lu_context_fini(&env->le_ctx);
2912 /* deconstruct service thread state created by ptlrpc_start_thread() */
2913 if (svc->srv_ops.so_thr_done != NULL)
2914 svc->srv_ops.so_thr_done(thread);
2916 CDEBUG(D_RPCTRACE, "%s: service thread [%p:%u] %d exiting: rc = %d\n",
2917 thread->t_name, thread, thread->t_pid, thread->t_id, rc);
2918 spin_lock(&svcpt->scp_lock);
2919 if (thread_test_and_clear_flags(thread, SVC_STARTING))
2920 svcpt->scp_nthrs_starting--;
2922 if (thread_test_and_clear_flags(thread, SVC_RUNNING)) {
2923 /* must know immediately */
2924 svcpt->scp_nthrs_running--;
2928 thread_add_flags(thread, SVC_STOPPED);
2930 wake_up(&thread->t_ctl_waitq);
2931 spin_unlock(&svcpt->scp_lock);
2936 static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
2937 struct list_head *replies)
2941 spin_lock(&hrt->hrt_lock);
2943 list_splice_init(&hrt->hrt_queue, replies);
2944 result = ptlrpc_hr.hr_stopping || !list_empty(replies);
2946 spin_unlock(&hrt->hrt_lock);
2951 * Main body of "handle reply" function.
2952 * It processes acked reply states
2954 static int ptlrpc_hr_main(void *arg)
2956 struct ptlrpc_hr_thread *hrt = (struct ptlrpc_hr_thread *)arg;
2957 struct ptlrpc_hr_partition *hrp = hrt->hrt_partition;
2962 unshare_fs_struct();
2967 rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
2969 char threadname[20];
2971 snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
2972 hrp->hrp_cpt, hrt->hrt_id);
2973 CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
2974 threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
2977 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD |
2978 LCT_REMEMBER | LCT_NOREF);
2982 rc = lu_env_add(env);
2984 GOTO(out_ctx_fini, rc);
2986 atomic_inc(&hrp->hrp_nstarted);
2987 wake_up(&ptlrpc_hr.hr_waitq);
2989 while (!ptlrpc_hr.hr_stopping) {
2990 wait_event_idle(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
2992 while (!list_empty(&replies)) {
2993 struct ptlrpc_reply_state *rs;
2995 rs = list_entry(replies.prev,
2996 struct ptlrpc_reply_state,
2998 list_del_init(&rs->rs_list);
2999 /* refill keys if needed */
3001 lu_context_enter(&env->le_ctx);
3002 ptlrpc_handle_rs(rs);
3003 lu_context_exit(&env->le_ctx);
3007 atomic_inc(&hrp->hrp_nstopped);
3008 wake_up(&ptlrpc_hr.hr_waitq);
3012 lu_context_fini(&env->le_ctx);
3018 static void ptlrpc_stop_hr_threads(void)
3020 struct ptlrpc_hr_partition *hrp;
3024 ptlrpc_hr.hr_stopping = 1;
3026 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
3027 if (hrp->hrp_thrs == NULL)
3028 continue; /* uninitialized */
3029 for (j = 0; j < hrp->hrp_nthrs; j++)
3030 wake_up(&hrp->hrp_thrs[j].hrt_waitq);
3033 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
3034 if (hrp->hrp_thrs == NULL)
3035 continue; /* uninitialized */
3036 wait_event(ptlrpc_hr.hr_waitq,
3037 atomic_read(&hrp->hrp_nstopped) ==
3038 atomic_read(&hrp->hrp_nstarted));
3042 static int ptlrpc_start_hr_threads(void)
3044 struct ptlrpc_hr_partition *hrp;
3050 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
3053 for (j = 0; j < hrp->hrp_nthrs; j++) {
3054 struct ptlrpc_hr_thread *hrt = &hrp->hrp_thrs[j];
3055 struct task_struct *task;
3057 task = kthread_run(ptlrpc_hr_main,
3059 "ptlrpc_hr%02d_%03d",
3068 wait_event(ptlrpc_hr.hr_waitq,
3069 atomic_read(&hrp->hrp_nstarted) == j);
3072 CERROR("cannot start reply handler thread %d:%d: rc = %d\n",
3074 ptlrpc_stop_hr_threads();
3082 static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
3084 struct ptlrpc_thread *thread;
3089 CDEBUG(D_INFO, "Stopping threads for service %s\n",
3090 svcpt->scp_service->srv_name);
3092 spin_lock(&svcpt->scp_lock);
3093 /* let the thread know that we would like it to stop asap */
3094 list_for_each_entry(thread, &svcpt->scp_threads, t_link)
3095 ptlrpc_stop_thread(thread);
3097 wake_up_all(&svcpt->scp_waitq);
3099 while ((thread = list_first_entry_or_null(&svcpt->scp_threads,
3100 struct ptlrpc_thread,
3102 if (thread_is_stopped(thread)) {
3103 list_move(&thread->t_link, &zombie);
3106 spin_unlock(&svcpt->scp_lock);
3108 CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
3109 svcpt->scp_service->srv_thread_name, thread->t_id);
3110 wait_event_idle(thread->t_ctl_waitq,
3111 thread_is_stopped(thread));
3113 spin_lock(&svcpt->scp_lock);
3116 spin_unlock(&svcpt->scp_lock);
3118 while ((thread = list_first_entry_or_null(&zombie,
3119 struct ptlrpc_thread,
3121 list_del(&thread->t_link);
3122 OBD_FREE_PTR(thread);
3128 * Stops all threads of a particular service \a svc
3130 static void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
3132 struct ptlrpc_service_part *svcpt;
3137 ptlrpc_service_for_each_part(svcpt, i, svc) {
3138 if (svcpt->scp_service != NULL)
3139 ptlrpc_svcpt_stop_threads(svcpt);
3145 static int ptlrpc_start_threads(struct ptlrpc_service *svc)
3153 /* We require 2 threads min, see note in ptlrpc_server_handle_request */
3154 LASSERT(svc->srv_nthrs_cpt_init >= PTLRPC_NTHRS_INIT);
3156 for (i = 0; i < svc->srv_ncpts; i++) {
3157 for (j = 0; j < svc->srv_nthrs_cpt_init; j++) {
3158 rc = ptlrpc_start_thread(svc->srv_parts[i], 1);
3164 /* We have enough threads, don't start more. b=15759 */
3171 CERROR("cannot start %s thread #%d_%d: rc %d\n",
3172 svc->srv_thread_name, i, j, rc);
3173 ptlrpc_stop_all_threads(svc);
3177 static int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
3179 struct ptlrpc_thread *thread;
3180 struct ptlrpc_service *svc;
3181 struct task_struct *task;
3186 LASSERT(svcpt != NULL);
3188 svc = svcpt->scp_service;
3190 CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n",
3191 svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running,
3192 svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit);
3195 if (unlikely(svc->srv_is_stopping))
3198 if (!ptlrpc_threads_increasable(svcpt) ||
3199 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
3200 svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1))
3203 OBD_CPT_ALLOC_PTR(thread, svc->srv_cptable, svcpt->scp_cpt);
3206 init_waitqueue_head(&thread->t_ctl_waitq);
3208 spin_lock(&svcpt->scp_lock);
3209 if (!ptlrpc_threads_increasable(svcpt)) {
3210 spin_unlock(&svcpt->scp_lock);
3211 OBD_FREE_PTR(thread);
3215 if (svcpt->scp_nthrs_starting != 0) {
3217 * serialize starting because some modules (obdfilter)
3218 * might require unique and contiguous t_id
3220 LASSERT(svcpt->scp_nthrs_starting == 1);
3221 spin_unlock(&svcpt->scp_lock);
3222 OBD_FREE_PTR(thread);
3224 CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n",
3225 svc->srv_thread_name, svcpt->scp_thr_nextid);
3230 CDEBUG(D_INFO, "Creating thread %s #%d race, retry later\n",
3231 svc->srv_thread_name, svcpt->scp_thr_nextid);
3235 svcpt->scp_nthrs_starting++;
3236 thread->t_id = svcpt->scp_thr_nextid++;
3237 thread_add_flags(thread, SVC_STARTING);
3238 thread->t_svcpt = svcpt;
3240 list_add(&thread->t_link, &svcpt->scp_threads);
3241 spin_unlock(&svcpt->scp_lock);
3243 if (svcpt->scp_cpt >= 0) {
3244 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s%02d_%03d",
3245 svc->srv_thread_name, svcpt->scp_cpt, thread->t_id);
3247 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s_%04d",
3248 svc->srv_thread_name, thread->t_id);
3251 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name);
3252 task = kthread_run(ptlrpc_main, thread, "%s", thread->t_name);
3255 CERROR("cannot start thread '%s': rc = %d\n",
3256 thread->t_name, rc);
3257 spin_lock(&svcpt->scp_lock);
3258 --svcpt->scp_nthrs_starting;
3259 if (thread_is_stopping(thread)) {
3261 * this ptlrpc_thread is being hanled
3262 * by ptlrpc_svcpt_stop_threads now
3264 thread_add_flags(thread, SVC_STOPPED);
3265 wake_up(&thread->t_ctl_waitq);
3266 spin_unlock(&svcpt->scp_lock);
3268 list_del(&thread->t_link);
3269 spin_unlock(&svcpt->scp_lock);
3270 OBD_FREE_PTR(thread);
3278 wait_event_idle(thread->t_ctl_waitq,
3279 thread_is_running(thread) || thread_is_stopped(thread));
3281 rc = thread_is_stopped(thread) ? thread->t_id : 0;
3285 int ptlrpc_hr_init(void)
3287 struct ptlrpc_hr_partition *hrp;
3288 struct ptlrpc_hr_thread *hrt;
3296 memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
3297 ptlrpc_hr.hr_cpt_table = cfs_cpt_tab;
3299 ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
3301 if (ptlrpc_hr.hr_partitions == NULL)
3304 ratelimit_state_init(&watchdog_limit,
3305 cfs_time_seconds(libcfs_watchdog_ratelimit), 3);
3307 init_waitqueue_head(&ptlrpc_hr.hr_waitq);
3310 weight = cpumask_weight(topology_sibling_cpumask(smp_processor_id()));
3313 cfs_percpt_for_each(hrp, cpt, ptlrpc_hr.hr_partitions) {
3316 atomic_set(&hrp->hrp_nstarted, 0);
3317 atomic_set(&hrp->hrp_nstopped, 0);
3319 hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, cpt);
3320 hrp->hrp_nthrs /= weight;
3321 if (hrp->hrp_nthrs == 0)
3324 OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, cpt,
3325 hrp->hrp_nthrs * sizeof(*hrt));
3326 if (hrp->hrp_thrs == NULL)
3327 GOTO(out, rc = -ENOMEM);
3329 for (i = 0; i < hrp->hrp_nthrs; i++) {
3330 hrt = &hrp->hrp_thrs[i];
3333 hrt->hrt_partition = hrp;
3334 init_waitqueue_head(&hrt->hrt_waitq);
3335 spin_lock_init(&hrt->hrt_lock);
3336 INIT_LIST_HEAD(&hrt->hrt_queue);
3340 rc = ptlrpc_start_hr_threads();
3347 void ptlrpc_hr_fini(void)
3349 struct ptlrpc_hr_partition *hrp;
3352 if (ptlrpc_hr.hr_partitions == NULL)
3355 ptlrpc_stop_hr_threads();
3357 cfs_percpt_for_each(hrp, cpt, ptlrpc_hr.hr_partitions) {
3359 OBD_FREE_PTR_ARRAY(hrp->hrp_thrs, hrp->hrp_nthrs);
3362 cfs_percpt_free(ptlrpc_hr.hr_partitions);
3363 ptlrpc_hr.hr_partitions = NULL;
3368 * Wait until all already scheduled replies are processed.
3370 static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
3373 if (wait_event_idle_timeout(
3375 atomic_read(&svcpt->scp_nreps_difficult) == 0,
3376 cfs_time_seconds(10)) > 0)
3378 CWARN("Unexpectedly long timeout %s %p\n",
3379 svcpt->scp_service->srv_name, svcpt->scp_service);
3384 ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
3386 struct ptlrpc_service_part *svcpt;
3389 /* early disarm AT timer... */
3390 ptlrpc_service_for_each_part(svcpt, i, svc) {
3391 if (svcpt->scp_service != NULL)
3392 del_timer(&svcpt->scp_at_timer);
3397 ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
3399 struct ptlrpc_service_part *svcpt;
3400 struct ptlrpc_request_buffer_desc *rqbd;
3405 * All history will be culled when the next request buffer is
3406 * freed in ptlrpc_service_purge_all()
3408 svc->srv_hist_nrqbds_cpt_max = 0;
3410 rc = LNetClearLazyPortal(svc->srv_req_portal);
3413 ptlrpc_service_for_each_part(svcpt, i, svc) {
3414 if (svcpt->scp_service == NULL)
3418 * Unlink all the request buffers. This forces a 'final'
3419 * event with its 'unlink' flag set for each posted rqbd
3421 list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
3423 rc = LNetMDUnlink(rqbd->rqbd_md_h);
3424 LASSERT(rc == 0 || rc == -ENOENT);
3428 ptlrpc_service_for_each_part(svcpt, i, svc) {
3429 if (svcpt->scp_service == NULL)
3433 * Wait for the network to release any buffers
3434 * it's currently filling
3436 spin_lock(&svcpt->scp_lock);
3437 while (svcpt->scp_nrqbds_posted != 0) {
3438 int seconds = PTLRPC_REQ_LONG_UNLINK;
3440 spin_unlock(&svcpt->scp_lock);
3442 * Network access will complete in finite time but
3443 * the HUGE timeout lets us CWARN for visibility
3446 while (seconds > 0 &&
3447 wait_event_idle_timeout(
3449 svcpt->scp_nrqbds_posted == 0,
3450 cfs_time_seconds(1)) == 0)
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 ((rs = list_first_entry_or_null(&svcpt->scp_rep_active,
3477 struct ptlrpc_reply_state,
3478 rs_list)) != NULL) {
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 ((req = list_first_entry_or_null(&svcpt->scp_req_incoming,
3491 struct ptlrpc_request,
3492 rq_list)) != NULL) {
3493 list_del(&req->rq_list);
3494 svcpt->scp_nreqs_incoming--;
3495 ptlrpc_server_finish_request(svcpt, req);
3498 while (ptlrpc_server_request_pending(svcpt, true)) {
3499 req = ptlrpc_server_request_get(svcpt, true);
3500 ptlrpc_server_finish_active_request(svcpt, req);
3504 * The portal may be shared by several services (eg:OUT_PORTAL).
3505 * So the request could be referenced by other target. So we
3506 * have to wait the ptlrpc_server_drop_request invoked.
3508 * TODO: move the req_buffer as global rather than per service.
3510 spin_lock(&svcpt->scp_lock);
3511 while (!list_empty(&svcpt->scp_rqbd_posted)) {
3512 spin_unlock(&svcpt->scp_lock);
3513 wait_event_idle_timeout(svcpt->scp_waitq,
3514 list_empty(&svcpt->scp_rqbd_posted),
3515 cfs_time_seconds(1));
3516 spin_lock(&svcpt->scp_lock);
3518 spin_unlock(&svcpt->scp_lock);
3520 LASSERT(svcpt->scp_nreqs_incoming == 0);
3521 LASSERT(svcpt->scp_nreqs_active == 0);
3523 * history should have been culled by
3524 * ptlrpc_server_finish_request
3526 LASSERT(svcpt->scp_hist_nrqbds == 0);
3529 * Now free all the request buffers since nothing
3530 * references them any more...
3532 while ((rqbd = list_first_entry_or_null(&svcpt->scp_rqbd_idle,
3533 struct ptlrpc_request_buffer_desc,
3534 rqbd_list)) != NULL)
3535 ptlrpc_free_rqbd(rqbd);
3537 ptlrpc_wait_replies(svcpt);
3539 while ((rs = list_first_entry_or_null(&svcpt->scp_rep_idle,
3540 struct ptlrpc_reply_state,
3541 rs_list)) != NULL) {
3542 list_del(&rs->rs_list);
3543 OBD_FREE_LARGE(rs, svc->srv_max_reply_size);
3549 ptlrpc_service_free(struct ptlrpc_service *svc)
3551 struct ptlrpc_service_part *svcpt;
3552 struct ptlrpc_at_array *array;
3555 ptlrpc_service_for_each_part(svcpt, i, svc) {
3556 if (svcpt->scp_service == NULL)
3559 /* In case somebody rearmed this in the meantime */
3560 del_timer(&svcpt->scp_at_timer);
3561 array = &svcpt->scp_at_array;
3563 if (array->paa_reqs_array != NULL) {
3564 OBD_FREE_PTR_ARRAY(array->paa_reqs_array,
3566 array->paa_reqs_array = NULL;
3569 if (array->paa_reqs_count != NULL) {
3570 OBD_FREE_PTR_ARRAY(array->paa_reqs_count,
3572 array->paa_reqs_count = NULL;
3576 ptlrpc_service_for_each_part(svcpt, i, svc)
3577 OBD_FREE_PTR(svcpt);
3579 if (svc->srv_cpts != NULL)
3580 cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts);
3582 OBD_FREE(svc, offsetof(struct ptlrpc_service,
3583 srv_parts[svc->srv_ncpts]));
3586 int ptlrpc_unregister_service(struct ptlrpc_service *service)
3590 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
3592 service->srv_is_stopping = 1;
3594 mutex_lock(&ptlrpc_all_services_mutex);
3595 list_del_init(&service->srv_list);
3596 mutex_unlock(&ptlrpc_all_services_mutex);
3598 ptlrpc_service_del_atimer(service);
3599 ptlrpc_stop_all_threads(service);
3601 ptlrpc_service_unlink_rqbd(service);
3602 ptlrpc_service_purge_all(service);
3603 ptlrpc_service_nrs_cleanup(service);
3605 ptlrpc_lprocfs_unregister_service(service);
3606 ptlrpc_sysfs_unregister_service(service);
3608 ptlrpc_service_free(service);
3612 EXPORT_SYMBOL(ptlrpc_unregister_service);
3615 * Returns 0 if the service is healthy.
3617 * Right now, it just checks to make sure that requests aren't languishing
3618 * in the queue. We'll use this health check to govern whether a node needs
3619 * to be shot, so it's intentionally non-aggressive.
3621 static int ptlrpc_svcpt_health_check(struct ptlrpc_service_part *svcpt)
3623 struct ptlrpc_request *request = NULL;
3624 struct timespec64 right_now;
3625 struct timespec64 timediff;
3627 ktime_get_real_ts64(&right_now);
3629 spin_lock(&svcpt->scp_req_lock);
3630 /* How long has the next entry been waiting? */
3631 if (ptlrpc_server_high_pending(svcpt, true))
3632 request = ptlrpc_nrs_req_peek_nolock(svcpt, true);
3633 else if (ptlrpc_server_normal_pending(svcpt, true))
3634 request = ptlrpc_nrs_req_peek_nolock(svcpt, false);
3636 if (request == NULL) {
3637 spin_unlock(&svcpt->scp_req_lock);
3641 timediff = timespec64_sub(right_now, request->rq_arrival_time);
3642 spin_unlock(&svcpt->scp_req_lock);
3644 if ((timediff.tv_sec) >
3645 (AT_OFF ? obd_timeout * 3 / 2 : at_max)) {
3646 CERROR("%s: unhealthy - request has been waiting %llds\n",
3647 svcpt->scp_service->srv_name, (s64)timediff.tv_sec);
3655 ptlrpc_service_health_check(struct ptlrpc_service *svc)
3657 struct ptlrpc_service_part *svcpt;
3663 ptlrpc_service_for_each_part(svcpt, i, svc) {
3664 int rc = ptlrpc_svcpt_health_check(svcpt);
3671 EXPORT_SYMBOL(ptlrpc_service_health_check);