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, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
33 #define DEBUG_SUBSYSTEM S_RPC
34 #include <linux/kthread.h>
35 #include <obd_support.h>
36 #include <obd_class.h>
37 #include <lustre_net.h>
38 #include <lu_object.h>
39 #include <lnet/types.h>
40 #include "ptlrpc_internal.h"
42 /* The following are visible and mutable through /sys/module/ptlrpc */
43 int test_req_buffer_pressure = 0;
44 module_param(test_req_buffer_pressure, int, 0444);
45 MODULE_PARM_DESC(test_req_buffer_pressure, "set non-zero to put pressure on request buffer pools");
46 module_param(at_min, int, 0644);
47 MODULE_PARM_DESC(at_min, "Adaptive timeout minimum (sec)");
48 module_param(at_max, int, 0644);
49 MODULE_PARM_DESC(at_max, "Adaptive timeout maximum (sec)");
50 module_param(at_history, int, 0644);
51 MODULE_PARM_DESC(at_history,
52 "Adaptive timeouts remember the slowest event that took place within this period (sec)");
53 module_param(at_early_margin, int, 0644);
54 MODULE_PARM_DESC(at_early_margin, "How soon before an RPC deadline to send an early reply");
55 module_param(at_extra, int, 0644);
56 MODULE_PARM_DESC(at_extra, "How much extra time to give with each early reply");
59 static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt);
60 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req);
61 static void ptlrpc_at_remove_timed(struct ptlrpc_request *req);
63 /** Holds a list of all PTLRPC services */
64 struct list_head ptlrpc_all_services;
65 /** Used to protect the \e ptlrpc_all_services list */
66 struct mutex ptlrpc_all_services_mutex;
68 static struct ptlrpc_request_buffer_desc *
69 ptlrpc_alloc_rqbd(struct ptlrpc_service_part *svcpt)
71 struct ptlrpc_service *svc = svcpt->scp_service;
72 struct ptlrpc_request_buffer_desc *rqbd;
74 OBD_CPT_ALLOC_PTR(rqbd, svc->srv_cptable, svcpt->scp_cpt);
78 rqbd->rqbd_svcpt = svcpt;
79 rqbd->rqbd_refcount = 0;
80 rqbd->rqbd_cbid.cbid_fn = request_in_callback;
81 rqbd->rqbd_cbid.cbid_arg = rqbd;
82 INIT_LIST_HEAD(&rqbd->rqbd_reqs);
83 OBD_CPT_ALLOC_LARGE(rqbd->rqbd_buffer, svc->srv_cptable,
84 svcpt->scp_cpt, svc->srv_buf_size);
85 if (rqbd->rqbd_buffer == NULL) {
90 spin_lock(&svcpt->scp_lock);
91 list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
92 svcpt->scp_nrqbds_total++;
93 spin_unlock(&svcpt->scp_lock);
99 ptlrpc_free_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
101 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
103 LASSERT(rqbd->rqbd_refcount == 0);
104 LASSERT(list_empty(&rqbd->rqbd_reqs));
106 spin_lock(&svcpt->scp_lock);
107 list_del(&rqbd->rqbd_list);
108 svcpt->scp_nrqbds_total--;
109 spin_unlock(&svcpt->scp_lock);
111 OBD_FREE_LARGE(rqbd->rqbd_buffer, svcpt->scp_service->srv_buf_size);
116 ptlrpc_grow_req_bufs(struct ptlrpc_service_part *svcpt, int post)
118 struct ptlrpc_service *svc = svcpt->scp_service;
119 struct ptlrpc_request_buffer_desc *rqbd;
123 if (svcpt->scp_rqbd_allocating)
126 spin_lock(&svcpt->scp_lock);
127 /* check again with lock */
128 if (svcpt->scp_rqbd_allocating) {
129 /* NB: we might allow more than one thread in the future */
130 LASSERT(svcpt->scp_rqbd_allocating == 1);
131 spin_unlock(&svcpt->scp_lock);
135 svcpt->scp_rqbd_allocating++;
136 spin_unlock(&svcpt->scp_lock);
139 for (i = 0; i < svc->srv_nbuf_per_group; i++) {
140 /* NB: another thread might have recycled enough rqbds, we
141 * need to make sure it wouldn't over-allocate, see LU-1212. */
142 if (svcpt->scp_nrqbds_posted >= svc->srv_nbuf_per_group)
145 rqbd = ptlrpc_alloc_rqbd(svcpt);
148 CERROR("%s: Can't allocate request buffer\n",
155 spin_lock(&svcpt->scp_lock);
157 LASSERT(svcpt->scp_rqbd_allocating == 1);
158 svcpt->scp_rqbd_allocating--;
160 spin_unlock(&svcpt->scp_lock);
163 "%s: allocate %d new %d-byte reqbufs (%d/%d left), rc = %d\n",
164 svc->srv_name, i, svc->srv_buf_size, svcpt->scp_nrqbds_posted,
165 svcpt->scp_nrqbds_total, rc);
169 rc = ptlrpc_server_post_idle_rqbds(svcpt);
175 * Part of Rep-Ack logic.
176 * Puts a lock and its mode into reply state assotiated to request reply.
179 ptlrpc_save_lock(struct ptlrpc_request *req,
180 struct lustre_handle *lock, int mode, int no_ack)
182 struct ptlrpc_reply_state *rs = req->rq_reply_state;
186 LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
188 if (req->rq_export->exp_disconnected) {
189 ldlm_lock_decref(lock, mode);
191 idx = rs->rs_nlocks++;
192 rs->rs_locks[idx] = *lock;
193 rs->rs_modes[idx] = mode;
194 rs->rs_difficult = 1;
195 rs->rs_no_ack = !!no_ack;
198 EXPORT_SYMBOL(ptlrpc_save_lock);
201 struct ptlrpc_hr_partition;
203 struct ptlrpc_hr_thread {
204 int hrt_id; /* thread ID */
206 wait_queue_head_t hrt_waitq;
207 struct list_head hrt_queue; /* RS queue */
208 struct ptlrpc_hr_partition *hrt_partition;
211 struct ptlrpc_hr_partition {
212 /* # of started threads */
213 atomic_t hrp_nstarted;
214 /* # of stopped threads */
215 atomic_t hrp_nstopped;
216 /* cpu partition id */
218 /* round-robin rotor for choosing thread */
220 /* total number of threads on this partition */
223 struct ptlrpc_hr_thread *hrp_thrs;
226 #define HRT_RUNNING 0
227 #define HRT_STOPPING 1
229 struct ptlrpc_hr_service {
230 /* CPU partition table, it's just cfs_cpt_table for now */
231 struct cfs_cpt_table *hr_cpt_table;
232 /** controller sleep waitq */
233 wait_queue_head_t hr_waitq;
234 unsigned int hr_stopping;
235 /** roundrobin rotor for non-affinity service */
236 unsigned int hr_rotor;
238 struct ptlrpc_hr_partition **hr_partitions;
242 struct list_head rsb_replies;
243 unsigned int rsb_n_replies;
244 struct ptlrpc_service_part *rsb_svcpt;
247 /** reply handling service. */
248 static struct ptlrpc_hr_service ptlrpc_hr;
251 * maximum mumber of replies scheduled in one batch
253 #define MAX_SCHEDULED 256
256 * Initialize a reply batch.
260 static void rs_batch_init(struct rs_batch *b)
262 memset(b, 0, sizeof *b);
263 INIT_LIST_HEAD(&b->rsb_replies);
267 * Choose an hr thread to dispatch requests to.
269 static struct ptlrpc_hr_thread *
270 ptlrpc_hr_select(struct ptlrpc_service_part *svcpt)
272 struct ptlrpc_hr_partition *hrp;
275 if (svcpt->scp_cpt >= 0 &&
276 svcpt->scp_service->srv_cptable == ptlrpc_hr.hr_cpt_table) {
277 /* directly match partition */
278 hrp = ptlrpc_hr.hr_partitions[svcpt->scp_cpt];
281 rotor = ptlrpc_hr.hr_rotor++;
282 rotor %= cfs_cpt_number(ptlrpc_hr.hr_cpt_table);
284 hrp = ptlrpc_hr.hr_partitions[rotor];
287 rotor = hrp->hrp_rotor++;
288 return &hrp->hrp_thrs[rotor % hrp->hrp_nthrs];
292 * Dispatch all replies accumulated in the batch to one from
293 * dedicated reply handling threads.
297 static void rs_batch_dispatch(struct rs_batch *b)
299 if (b->rsb_n_replies != 0) {
300 struct ptlrpc_hr_thread *hrt;
302 hrt = ptlrpc_hr_select(b->rsb_svcpt);
304 spin_lock(&hrt->hrt_lock);
305 list_splice_init(&b->rsb_replies, &hrt->hrt_queue);
306 spin_unlock(&hrt->hrt_lock);
308 wake_up(&hrt->hrt_waitq);
309 b->rsb_n_replies = 0;
314 * Add a reply to a batch.
315 * Add one reply object to a batch, schedule batched replies if overload.
320 static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs)
322 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
324 if (svcpt != b->rsb_svcpt || b->rsb_n_replies >= MAX_SCHEDULED) {
325 if (b->rsb_svcpt != NULL) {
326 rs_batch_dispatch(b);
327 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
329 spin_lock(&svcpt->scp_rep_lock);
330 b->rsb_svcpt = svcpt;
332 spin_lock(&rs->rs_lock);
333 rs->rs_scheduled_ever = 1;
334 if (rs->rs_scheduled == 0) {
335 list_move(&rs->rs_list, &b->rsb_replies);
336 rs->rs_scheduled = 1;
339 rs->rs_committed = 1;
340 spin_unlock(&rs->rs_lock);
344 * Reply batch finalization.
345 * Dispatch remaining replies from the batch
346 * and release remaining spinlock.
350 static void rs_batch_fini(struct rs_batch *b)
352 if (b->rsb_svcpt != NULL) {
353 rs_batch_dispatch(b);
354 spin_unlock(&b->rsb_svcpt->scp_rep_lock);
358 #define DECLARE_RS_BATCH(b) struct rs_batch b
362 * Put reply state into a queue for processing because we received
363 * ACK from the client
365 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
367 struct ptlrpc_hr_thread *hrt;
370 LASSERT(list_empty(&rs->rs_list));
372 hrt = ptlrpc_hr_select(rs->rs_svcpt);
374 spin_lock(&hrt->hrt_lock);
375 list_add_tail(&rs->rs_list, &hrt->hrt_queue);
376 spin_unlock(&hrt->hrt_lock);
378 wake_up(&hrt->hrt_waitq);
383 ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs)
387 assert_spin_locked(&rs->rs_svcpt->scp_rep_lock);
388 assert_spin_locked(&rs->rs_lock);
389 LASSERT (rs->rs_difficult);
390 rs->rs_scheduled_ever = 1; /* flag any notification attempt */
392 if (rs->rs_scheduled) { /* being set up or already notified */
397 rs->rs_scheduled = 1;
398 list_del_init(&rs->rs_list);
399 ptlrpc_dispatch_difficult_reply(rs);
402 EXPORT_SYMBOL(ptlrpc_schedule_difficult_reply);
404 void ptlrpc_commit_replies(struct obd_export *exp)
406 struct ptlrpc_reply_state *rs, *nxt;
407 DECLARE_RS_BATCH(batch);
410 rs_batch_init(&batch);
411 /* Find any replies that have been committed and get their service
412 * to attend to complete them. */
414 /* CAVEAT EMPTOR: spinlock ordering!!! */
415 spin_lock(&exp->exp_uncommitted_replies_lock);
416 list_for_each_entry_safe(rs, nxt, &exp->exp_uncommitted_replies,
418 LASSERT (rs->rs_difficult);
419 /* VBR: per-export last_committed */
420 LASSERT(rs->rs_export);
421 if (rs->rs_transno <= exp->exp_last_committed) {
422 list_del_init(&rs->rs_obd_list);
423 rs_batch_add(&batch, rs);
426 spin_unlock(&exp->exp_uncommitted_replies_lock);
427 rs_batch_fini(&batch);
432 ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt)
434 struct ptlrpc_request_buffer_desc *rqbd;
439 spin_lock(&svcpt->scp_lock);
441 if (list_empty(&svcpt->scp_rqbd_idle)) {
442 spin_unlock(&svcpt->scp_lock);
446 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
447 struct ptlrpc_request_buffer_desc,
449 list_del(&rqbd->rqbd_list);
451 /* assume we will post successfully */
452 svcpt->scp_nrqbds_posted++;
453 list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_posted);
455 spin_unlock(&svcpt->scp_lock);
457 rc = ptlrpc_register_rqbd(rqbd);
464 spin_lock(&svcpt->scp_lock);
466 svcpt->scp_nrqbds_posted--;
467 list_del(&rqbd->rqbd_list);
468 list_add_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
470 /* Don't complain if no request buffers are posted right now; LNET
471 * won't drop requests because we set the portal lazy! */
473 spin_unlock(&svcpt->scp_lock);
478 static void ptlrpc_at_timer(unsigned long castmeharder)
480 struct ptlrpc_service_part *svcpt;
482 svcpt = (struct ptlrpc_service_part *)castmeharder;
484 svcpt->scp_at_check = 1;
485 svcpt->scp_at_checktime = cfs_time_current();
486 wake_up(&svcpt->scp_waitq);
490 ptlrpc_server_nthreads_check(struct ptlrpc_service *svc,
491 struct ptlrpc_service_conf *conf)
493 struct ptlrpc_service_thr_conf *tc = &conf->psc_thr;
500 * Common code for estimating & validating threads number.
501 * CPT affinity service could have percpt thread-pool instead
502 * of a global thread-pool, which means user might not always
503 * get the threads number they give it in conf::tc_nthrs_user
504 * even they did set. It's because we need to validate threads
505 * number for each CPT to guarantee each pool will have enough
506 * threads to keep the service healthy.
508 init = PTLRPC_NTHRS_INIT + (svc->srv_ops.so_hpreq_handler != NULL);
509 init = max_t(int, init, tc->tc_nthrs_init);
511 /* NB: please see comments in lustre_lnet.h for definition
512 * details of these members */
513 LASSERT(tc->tc_nthrs_max != 0);
515 if (tc->tc_nthrs_user != 0) {
516 /* In case there is a reason to test a service with many
517 * threads, we give a less strict check here, it can
518 * be up to 8 * nthrs_max */
519 total = min(tc->tc_nthrs_max * 8, tc->tc_nthrs_user);
520 nthrs = total / svc->srv_ncpts;
521 init = max(init, nthrs);
525 total = tc->tc_nthrs_max;
526 if (tc->tc_nthrs_base == 0) {
527 /* don't care about base threads number per partition,
528 * this is most for non-affinity service */
529 nthrs = total / svc->srv_ncpts;
533 nthrs = tc->tc_nthrs_base;
534 if (svc->srv_ncpts == 1) {
537 /* NB: Increase the base number if it's single partition
538 * and total number of cores/HTs is larger or equal to 4.
539 * result will always < 2 * nthrs_base */
540 weight = cfs_cpt_weight(svc->srv_cptable, CFS_CPT_ANY);
541 for (i = 1; (weight >> (i + 1)) != 0 && /* >= 4 cores/HTs */
542 (tc->tc_nthrs_base >> i) != 0; i++)
543 nthrs += tc->tc_nthrs_base >> i;
546 if (tc->tc_thr_factor != 0) {
547 int factor = tc->tc_thr_factor;
551 * User wants to increase number of threads with for
552 * each CPU core/HT, most likely the factor is larger than
553 * one thread/core because service threads are supposed to
554 * be blocked by lock or wait for IO.
557 * Amdahl's law says that adding processors wouldn't give
558 * a linear increasing of parallelism, so it's nonsense to
559 * have too many threads no matter how many cores/HTs
562 if (cfs_cpu_ht_nsiblings(0) > 1) { /* weight is # of HTs */
563 /* depress thread factor for hyper-thread */
564 factor = factor - (factor >> 1) + (factor >> 3);
567 weight = cfs_cpt_weight(svc->srv_cptable, 0);
570 for (; factor > 0 && weight > 0; factor--, weight -= fade)
571 nthrs += min(weight, fade) * factor;
574 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
575 nthrs = max(tc->tc_nthrs_base,
576 tc->tc_nthrs_max / svc->srv_ncpts);
579 nthrs = max(nthrs, tc->tc_nthrs_init);
580 svc->srv_nthrs_cpt_limit = nthrs;
581 svc->srv_nthrs_cpt_init = init;
583 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
584 CDEBUG(D_OTHER, "%s: This service may have more threads (%d) "
585 "than the given soft limit (%d)\n",
586 svc->srv_name, nthrs * svc->srv_ncpts,
592 * Initialize percpt data for a service
595 ptlrpc_service_part_init(struct ptlrpc_service *svc,
596 struct ptlrpc_service_part *svcpt, int cpt)
598 struct ptlrpc_at_array *array;
603 svcpt->scp_cpt = cpt;
604 INIT_LIST_HEAD(&svcpt->scp_threads);
606 /* rqbd and incoming request queue */
607 spin_lock_init(&svcpt->scp_lock);
608 INIT_LIST_HEAD(&svcpt->scp_rqbd_idle);
609 INIT_LIST_HEAD(&svcpt->scp_rqbd_posted);
610 INIT_LIST_HEAD(&svcpt->scp_req_incoming);
611 init_waitqueue_head(&svcpt->scp_waitq);
612 /* history request & rqbd list */
613 INIT_LIST_HEAD(&svcpt->scp_hist_reqs);
614 INIT_LIST_HEAD(&svcpt->scp_hist_rqbds);
616 /* acitve requests and hp requests */
617 spin_lock_init(&svcpt->scp_req_lock);
620 spin_lock_init(&svcpt->scp_rep_lock);
621 INIT_LIST_HEAD(&svcpt->scp_rep_active);
622 INIT_LIST_HEAD(&svcpt->scp_rep_idle);
623 init_waitqueue_head(&svcpt->scp_rep_waitq);
624 atomic_set(&svcpt->scp_nreps_difficult, 0);
626 /* adaptive timeout */
627 spin_lock_init(&svcpt->scp_at_lock);
628 array = &svcpt->scp_at_array;
630 size = at_est2timeout(at_max);
631 array->paa_size = size;
632 array->paa_count = 0;
633 array->paa_deadline = -1;
635 /* allocate memory for scp_at_array (ptlrpc_at_array) */
636 OBD_CPT_ALLOC(array->paa_reqs_array,
637 svc->srv_cptable, cpt, sizeof(struct list_head) * size);
638 if (array->paa_reqs_array == NULL)
641 for (index = 0; index < size; index++)
642 INIT_LIST_HEAD(&array->paa_reqs_array[index]);
644 OBD_CPT_ALLOC(array->paa_reqs_count,
645 svc->srv_cptable, cpt, sizeof(__u32) * size);
646 if (array->paa_reqs_count == NULL)
649 setup_timer(&svcpt->scp_at_timer, ptlrpc_at_timer,
650 (unsigned long)svcpt);
652 /* At SOW, service time should be quick; 10s seems generous. If client
653 * timeout is less than this, we'll be sending an early reply. */
654 at_init(&svcpt->scp_at_estimate, 10, 0);
656 /* assign this before call ptlrpc_grow_req_bufs */
657 svcpt->scp_service = svc;
658 /* Now allocate the request buffers, but don't post them now */
659 rc = ptlrpc_grow_req_bufs(svcpt, 0);
660 /* We shouldn't be under memory pressure at startup, so
661 * fail if we can't allocate all our buffers at this time. */
668 if (array->paa_reqs_count != NULL) {
669 OBD_FREE(array->paa_reqs_count, sizeof(__u32) * size);
670 array->paa_reqs_count = NULL;
673 if (array->paa_reqs_array != NULL) {
674 OBD_FREE(array->paa_reqs_array,
675 sizeof(struct list_head) * array->paa_size);
676 array->paa_reqs_array = NULL;
683 * Initialize service on a given portal.
684 * This includes starting serving threads , allocating and posting rqbds and
687 struct ptlrpc_service *
688 ptlrpc_register_service(struct ptlrpc_service_conf *conf,
689 struct proc_dir_entry *proc_entry)
691 struct ptlrpc_service_cpt_conf *cconf = &conf->psc_cpt;
692 struct ptlrpc_service *service;
693 struct ptlrpc_service_part *svcpt;
694 struct cfs_cpt_table *cptable;
702 LASSERT(conf->psc_buf.bc_nbufs > 0);
703 LASSERT(conf->psc_buf.bc_buf_size >=
704 conf->psc_buf.bc_req_max_size + SPTLRPC_MAX_PAYLOAD);
705 LASSERT(conf->psc_thr.tc_ctx_tags != 0);
707 cptable = cconf->cc_cptable;
709 cptable = cfs_cpt_table;
711 if (!conf->psc_thr.tc_cpu_affinity) {
714 ncpts = cfs_cpt_number(cptable);
715 if (cconf->cc_pattern != NULL) {
716 struct cfs_expr_list *el;
718 rc = cfs_expr_list_parse(cconf->cc_pattern,
719 strlen(cconf->cc_pattern),
722 CERROR("%s: invalid CPT pattern string: %s",
723 conf->psc_name, cconf->cc_pattern);
724 RETURN(ERR_PTR(-EINVAL));
727 rc = cfs_expr_list_values(el, ncpts, &cpts);
728 cfs_expr_list_free(el);
730 CERROR("%s: failed to parse CPT array %s: %d\n",
731 conf->psc_name, cconf->cc_pattern, rc);
733 OBD_FREE(cpts, sizeof(*cpts) * ncpts);
734 RETURN(ERR_PTR(rc < 0 ? rc : -EINVAL));
740 OBD_ALLOC(service, offsetof(struct ptlrpc_service, srv_parts[ncpts]));
741 if (service == NULL) {
743 OBD_FREE(cpts, sizeof(*cpts) * ncpts);
744 RETURN(ERR_PTR(-ENOMEM));
747 service->srv_cptable = cptable;
748 service->srv_cpts = cpts;
749 service->srv_ncpts = ncpts;
751 service->srv_cpt_bits = 0; /* it's zero already, easy to read... */
752 while ((1 << service->srv_cpt_bits) < cfs_cpt_number(cptable))
753 service->srv_cpt_bits++;
756 spin_lock_init(&service->srv_lock);
757 service->srv_name = conf->psc_name;
758 service->srv_watchdog_factor = conf->psc_watchdog_factor;
759 INIT_LIST_HEAD(&service->srv_list); /* for safty of cleanup */
761 /* buffer configuration */
762 service->srv_nbuf_per_group = test_req_buffer_pressure ?
763 1 : conf->psc_buf.bc_nbufs;
764 service->srv_max_req_size = conf->psc_buf.bc_req_max_size +
766 service->srv_buf_size = conf->psc_buf.bc_buf_size;
767 service->srv_rep_portal = conf->psc_buf.bc_rep_portal;
768 service->srv_req_portal = conf->psc_buf.bc_req_portal;
770 /* Increase max reply size to next power of two */
771 service->srv_max_reply_size = 1;
772 while (service->srv_max_reply_size <
773 conf->psc_buf.bc_rep_max_size + SPTLRPC_MAX_PAYLOAD)
774 service->srv_max_reply_size <<= 1;
776 service->srv_thread_name = conf->psc_thr.tc_thr_name;
777 service->srv_ctx_tags = conf->psc_thr.tc_ctx_tags;
778 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
779 service->srv_ops = conf->psc_ops;
781 for (i = 0; i < ncpts; i++) {
782 if (!conf->psc_thr.tc_cpu_affinity)
785 cpt = cpts != NULL ? cpts[i] : i;
787 OBD_CPT_ALLOC(svcpt, cptable, cpt, sizeof(*svcpt));
789 GOTO(failed, rc = -ENOMEM);
791 service->srv_parts[i] = svcpt;
792 rc = ptlrpc_service_part_init(service, svcpt, cpt);
797 ptlrpc_server_nthreads_check(service, conf);
799 rc = LNetSetLazyPortal(service->srv_req_portal);
802 mutex_lock(&ptlrpc_all_services_mutex);
803 list_add(&service->srv_list, &ptlrpc_all_services);
804 mutex_unlock(&ptlrpc_all_services_mutex);
806 if (proc_entry != NULL)
807 ptlrpc_lprocfs_register_service(proc_entry, service);
809 rc = ptlrpc_service_nrs_setup(service);
813 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
814 service->srv_name, service->srv_req_portal);
816 rc = ptlrpc_start_threads(service);
818 CERROR("Failed to start threads for service %s: %d\n",
819 service->srv_name, rc);
825 ptlrpc_unregister_service(service);
828 EXPORT_SYMBOL(ptlrpc_register_service);
831 * to actually free the request, must be called without holding svc_lock.
832 * note it's caller's responsibility to unlink req->rq_list.
834 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
836 LASSERT(atomic_read(&req->rq_refcount) == 0);
837 LASSERT(list_empty(&req->rq_timed_list));
839 /* DEBUG_REQ() assumes the reply state of a request with a valid
840 * ref will not be destroyed until that reference is dropped. */
841 ptlrpc_req_drop_rs(req);
843 sptlrpc_svc_ctx_decref(req);
845 if (req != &req->rq_rqbd->rqbd_req) {
846 /* NB request buffers use an embedded
847 * req if the incoming req unlinked the
848 * MD; this isn't one of them! */
849 ptlrpc_request_cache_free(req);
854 * drop a reference count of the request. if it reaches 0, we either
855 * put it into history list, or free it immediately.
857 void ptlrpc_server_drop_request(struct ptlrpc_request *req)
859 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
860 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
861 struct ptlrpc_service *svc = svcpt->scp_service;
863 struct list_head *tmp;
864 struct list_head *nxt;
866 if (!atomic_dec_and_test(&req->rq_refcount))
869 if (req->rq_session.lc_state == LCS_ENTERED) {
870 lu_context_exit(&req->rq_session);
871 lu_context_fini(&req->rq_session);
874 if (req->rq_at_linked) {
875 spin_lock(&svcpt->scp_at_lock);
876 /* recheck with lock, in case it's unlinked by
877 * ptlrpc_at_check_timed() */
878 if (likely(req->rq_at_linked))
879 ptlrpc_at_remove_timed(req);
880 spin_unlock(&svcpt->scp_at_lock);
883 LASSERT(list_empty(&req->rq_timed_list));
885 /* finalize request */
886 if (req->rq_export) {
887 class_export_put(req->rq_export);
888 req->rq_export = NULL;
891 spin_lock(&svcpt->scp_lock);
893 list_add(&req->rq_list, &rqbd->rqbd_reqs);
895 refcount = --(rqbd->rqbd_refcount);
897 /* request buffer is now idle: add to history */
898 list_del(&rqbd->rqbd_list);
900 list_add_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds);
901 svcpt->scp_hist_nrqbds++;
903 /* cull some history?
904 * I expect only about 1 or 2 rqbds need to be recycled here */
905 while (svcpt->scp_hist_nrqbds > svc->srv_hist_nrqbds_cpt_max) {
906 rqbd = list_entry(svcpt->scp_hist_rqbds.next,
907 struct ptlrpc_request_buffer_desc,
910 list_del(&rqbd->rqbd_list);
911 svcpt->scp_hist_nrqbds--;
913 /* remove rqbd's reqs from svc's req history while
914 * I've got the service lock */
915 list_for_each(tmp, &rqbd->rqbd_reqs) {
916 req = list_entry(tmp, struct ptlrpc_request,
918 /* Track the highest culled req seq */
919 if (req->rq_history_seq >
920 svcpt->scp_hist_seq_culled) {
921 svcpt->scp_hist_seq_culled =
924 list_del(&req->rq_history_list);
927 spin_unlock(&svcpt->scp_lock);
929 list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
930 req = list_entry(rqbd->rqbd_reqs.next,
931 struct ptlrpc_request,
933 list_del(&req->rq_list);
934 ptlrpc_server_free_request(req);
937 spin_lock(&svcpt->scp_lock);
939 * now all reqs including the embedded req has been
940 * disposed, schedule request buffer for re-use.
942 LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) == 0);
943 list_add_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
946 spin_unlock(&svcpt->scp_lock);
947 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
948 /* If we are low on memory, we are not interested in history */
949 list_del(&req->rq_list);
950 list_del_init(&req->rq_history_list);
952 /* Track the highest culled req seq */
953 if (req->rq_history_seq > svcpt->scp_hist_seq_culled)
954 svcpt->scp_hist_seq_culled = req->rq_history_seq;
956 spin_unlock(&svcpt->scp_lock);
958 ptlrpc_server_free_request(req);
960 spin_unlock(&svcpt->scp_lock);
964 /** Change request export and move hp request from old export to new */
965 void ptlrpc_request_change_export(struct ptlrpc_request *req,
966 struct obd_export *export)
968 if (req->rq_export != NULL) {
969 LASSERT(!list_empty(&req->rq_exp_list));
970 /* remove rq_exp_list from last export */
971 spin_lock_bh(&req->rq_export->exp_rpc_lock);
972 list_del_init(&req->rq_exp_list);
973 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
974 /* export has one reference already, so it`s safe to
975 * add req to export queue here and get another
976 * reference for request later */
977 spin_lock_bh(&export->exp_rpc_lock);
978 if (req->rq_ops != NULL) /* hp request */
979 list_add(&req->rq_exp_list, &export->exp_hp_rpcs);
981 list_add(&req->rq_exp_list, &export->exp_reg_rpcs);
982 spin_unlock_bh(&export->exp_rpc_lock);
984 class_export_rpc_dec(req->rq_export);
985 class_export_put(req->rq_export);
988 /* request takes one export refcount */
989 req->rq_export = class_export_get(export);
990 class_export_rpc_inc(export);
996 * to finish a request: stop sending more early replies, and release
999 static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt,
1000 struct ptlrpc_request *req)
1002 ptlrpc_server_hpreq_fini(req);
1004 ptlrpc_server_drop_request(req);
1008 * to finish an active request: stop sending more early replies, and release
1009 * the request. should be called after we finished handling the request.
1011 static void ptlrpc_server_finish_active_request(
1012 struct ptlrpc_service_part *svcpt,
1013 struct ptlrpc_request *req)
1015 spin_lock(&svcpt->scp_req_lock);
1016 ptlrpc_nrs_req_stop_nolock(req);
1017 svcpt->scp_nreqs_active--;
1019 svcpt->scp_nhreqs_active--;
1020 spin_unlock(&svcpt->scp_req_lock);
1022 ptlrpc_nrs_req_finalize(req);
1024 if (req->rq_export != NULL)
1025 class_export_rpc_dec(req->rq_export);
1027 ptlrpc_server_finish_request(svcpt, req);
1031 * This function makes sure dead exports are evicted in a timely manner.
1032 * This function is only called when some export receives a message (i.e.,
1033 * the network is up.)
1035 void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
1037 struct obd_export *oldest_exp;
1038 time_t oldest_time, new_time;
1044 /* Compensate for slow machines, etc, by faking our request time
1045 into the future. Although this can break the strict time-ordering
1046 of the list, we can be really lazy here - we don't have to evict
1047 at the exact right moment. Eventually, all silent exports
1048 will make it to the top of the list. */
1050 /* Do not pay attention on 1sec or smaller renewals. */
1051 new_time = cfs_time_current_sec() + extra_delay;
1052 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
1055 exp->exp_last_request_time = new_time;
1057 /* exports may get disconnected from the chain even though the
1058 export has references, so we must keep the spin lock while
1059 manipulating the lists */
1060 spin_lock(&exp->exp_obd->obd_dev_lock);
1062 if (list_empty(&exp->exp_obd_chain_timed)) {
1063 /* this one is not timed */
1064 spin_unlock(&exp->exp_obd->obd_dev_lock);
1068 list_move_tail(&exp->exp_obd_chain_timed,
1069 &exp->exp_obd->obd_exports_timed);
1071 oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
1072 struct obd_export, exp_obd_chain_timed);
1073 oldest_time = oldest_exp->exp_last_request_time;
1074 spin_unlock(&exp->exp_obd->obd_dev_lock);
1076 if (exp->exp_obd->obd_recovering) {
1077 /* be nice to everyone during recovery */
1082 /* Note - racing to start/reset the obd_eviction timer is safe */
1083 if (exp->exp_obd->obd_eviction_timer == 0) {
1084 /* Check if the oldest entry is expired. */
1085 if (cfs_time_current_sec() > (oldest_time + PING_EVICT_TIMEOUT +
1087 /* We need a second timer, in case the net was down and
1088 * it just came back. Since the pinger may skip every
1089 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
1090 * we better wait for 3. */
1091 exp->exp_obd->obd_eviction_timer =
1092 cfs_time_current_sec() + 3 * PING_INTERVAL;
1093 CDEBUG(D_HA, "%s: Think about evicting %s from "CFS_TIME_T"\n",
1094 exp->exp_obd->obd_name,
1095 obd_export_nid2str(oldest_exp), oldest_time);
1098 if (cfs_time_current_sec() >
1099 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
1100 /* The evictor won't evict anyone who we've heard from
1101 * recently, so we don't have to check before we start
1103 if (!ping_evictor_wake(exp))
1104 exp->exp_obd->obd_eviction_timer = 0;
1112 * Sanity check request \a req.
1113 * Return 0 if all is ok, error code otherwise.
1115 static int ptlrpc_check_req(struct ptlrpc_request *req)
1117 struct obd_device *obd = req->rq_export->exp_obd;
1120 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
1121 req->rq_export->exp_conn_cnt)) {
1122 DEBUG_REQ(D_RPCTRACE, req,
1123 "DROPPING req from old connection %d < %d",
1124 lustre_msg_get_conn_cnt(req->rq_reqmsg),
1125 req->rq_export->exp_conn_cnt);
1128 if (unlikely(obd == NULL || obd->obd_fail)) {
1129 /* Failing over, don't handle any more reqs,
1130 * send error response instead. */
1131 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
1132 req, (obd != NULL) ? obd->obd_name : "unknown");
1134 } else if (lustre_msg_get_flags(req->rq_reqmsg) &
1135 (MSG_REPLAY | MSG_REQ_REPLAY_DONE) &&
1136 !obd->obd_recovering) {
1137 DEBUG_REQ(D_ERROR, req,
1138 "Invalid replay without recovery");
1139 class_fail_export(req->rq_export);
1141 } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 &&
1142 !obd->obd_recovering) {
1143 DEBUG_REQ(D_ERROR, req, "Invalid req with transno "
1144 "%llu without recovery",
1145 lustre_msg_get_transno(req->rq_reqmsg));
1146 class_fail_export(req->rq_export);
1150 if (unlikely(rc < 0)) {
1151 req->rq_status = rc;
1157 static void ptlrpc_at_set_timer(struct ptlrpc_service_part *svcpt)
1159 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1162 if (array->paa_count == 0) {
1163 del_timer(&svcpt->scp_at_timer);
1167 /* Set timer for closest deadline */
1168 next = (__s32)(array->paa_deadline - cfs_time_current_sec() -
1171 ptlrpc_at_timer((unsigned long)svcpt);
1173 mod_timer(&svcpt->scp_at_timer, cfs_time_shift(next));
1174 CDEBUG(D_INFO, "armed %s at %+ds\n",
1175 svcpt->scp_service->srv_name, next);
1179 /* Add rpc to early reply check list */
1180 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
1182 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1183 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1184 struct ptlrpc_request *rq = NULL;
1190 if (req->rq_no_reply)
1193 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
1196 spin_lock(&svcpt->scp_at_lock);
1197 LASSERT(list_empty(&req->rq_timed_list));
1199 index = (unsigned long)req->rq_deadline % array->paa_size;
1200 if (array->paa_reqs_count[index] > 0) {
1201 /* latest rpcs will have the latest deadlines in the list,
1202 * so search backward. */
1203 list_for_each_entry_reverse(rq,
1204 &array->paa_reqs_array[index],
1206 if (req->rq_deadline >= rq->rq_deadline) {
1207 list_add(&req->rq_timed_list,
1208 &rq->rq_timed_list);
1214 /* Add the request at the head of the list */
1215 if (list_empty(&req->rq_timed_list))
1216 list_add(&req->rq_timed_list,
1217 &array->paa_reqs_array[index]);
1219 spin_lock(&req->rq_lock);
1220 req->rq_at_linked = 1;
1221 spin_unlock(&req->rq_lock);
1222 req->rq_at_index = index;
1223 array->paa_reqs_count[index]++;
1225 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
1226 array->paa_deadline = req->rq_deadline;
1227 ptlrpc_at_set_timer(svcpt);
1229 spin_unlock(&svcpt->scp_at_lock);
1235 ptlrpc_at_remove_timed(struct ptlrpc_request *req)
1237 struct ptlrpc_at_array *array;
1239 array = &req->rq_rqbd->rqbd_svcpt->scp_at_array;
1241 /* NB: must call with hold svcpt::scp_at_lock */
1242 LASSERT(!list_empty(&req->rq_timed_list));
1243 list_del_init(&req->rq_timed_list);
1245 spin_lock(&req->rq_lock);
1246 req->rq_at_linked = 0;
1247 spin_unlock(&req->rq_lock);
1249 array->paa_reqs_count[req->rq_at_index]--;
1254 * Attempt to extend the request deadline by sending an early reply to the
1257 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
1259 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1260 struct ptlrpc_request *reqcopy;
1261 struct lustre_msg *reqmsg;
1262 cfs_duration_t olddl = req->rq_deadline - cfs_time_current_sec();
1268 if (CFS_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_RECONNECT)) {
1269 /* don't send early reply */
1273 /* deadline is when the client expects us to reply, margin is the
1274 difference between clients' and servers' expectations */
1275 DEBUG_REQ(D_ADAPTTO, req,
1276 "%ssending early reply (deadline %+lds, margin %+lds) for "
1277 "%d+%d", AT_OFF ? "AT off - not " : "",
1278 olddl, olddl - at_get(&svcpt->scp_at_estimate),
1279 at_get(&svcpt->scp_at_estimate), at_extra);
1285 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), "
1286 "not sending early reply. Consider increasing "
1287 "at_early_margin (%d)?", olddl, at_early_margin);
1289 /* Return an error so we're not re-added to the timed list. */
1293 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
1294 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
1295 "but no AT support");
1299 if (req->rq_export &&
1300 lustre_msg_get_flags(req->rq_reqmsg) &
1301 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1302 /* During recovery, we don't want to send too many early
1303 * replies, but on the other hand we want to make sure the
1304 * client has enough time to resend if the rpc is lost. So
1305 * during the recovery period send at least 4 early replies,
1306 * spacing them every at_extra if we can. at_estimate should
1307 * always equal this fixed value during recovery. */
1308 /* Don't account request processing time into AT history
1309 * during recovery, it is not service time we need but
1310 * includes also waiting time for recovering clients */
1311 newdl = cfs_time_current_sec() + min(at_extra,
1312 req->rq_export->exp_obd->obd_recovery_timeout / 4);
1314 /* We want to extend the request deadline by at_extra seconds,
1315 * so we set our service estimate to reflect how much time has
1316 * passed since this request arrived plus an additional
1317 * at_extra seconds. The client will calculate the new deadline
1318 * based on this service estimate (plus some additional time to
1319 * account for network latency). See ptlrpc_at_recv_early_reply
1321 at_measured(&svcpt->scp_at_estimate, at_extra +
1322 cfs_time_current_sec() -
1323 req->rq_arrival_time.tv_sec);
1324 newdl = req->rq_arrival_time.tv_sec +
1325 at_get(&svcpt->scp_at_estimate);
1328 /* Check to see if we've actually increased the deadline -
1329 * we may be past adaptive_max */
1330 if (req->rq_deadline >= newdl) {
1331 DEBUG_REQ(D_WARNING, req, "Couldn't add any time "
1332 "(%ld/%ld), not sending early reply\n",
1333 olddl, newdl - cfs_time_current_sec());
1337 reqcopy = ptlrpc_request_cache_alloc(GFP_NOFS);
1338 if (reqcopy == NULL)
1340 OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1342 GOTO(out_free, rc = -ENOMEM);
1345 reqcopy->rq_reply_state = NULL;
1346 reqcopy->rq_rep_swab_mask = 0;
1347 reqcopy->rq_pack_bulk = 0;
1348 reqcopy->rq_pack_udesc = 0;
1349 reqcopy->rq_packed_final = 0;
1350 sptlrpc_svc_ctx_addref(reqcopy);
1351 /* We only need the reqmsg for the magic */
1352 reqcopy->rq_reqmsg = reqmsg;
1353 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1356 * tgt_brw_read() and tgt_brw_write() may have decided not to reply.
1357 * Without this check, we would fail the rq_no_reply assertion in
1358 * ptlrpc_send_reply().
1360 if (reqcopy->rq_no_reply)
1361 GOTO(out, rc = -ETIMEDOUT);
1363 LASSERT(atomic_read(&req->rq_refcount));
1364 /** if it is last refcount then early reply isn't needed */
1365 if (atomic_read(&req->rq_refcount) == 1) {
1366 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
1367 "abort sending early reply\n");
1368 GOTO(out, rc = -EINVAL);
1371 /* Connection ref */
1372 reqcopy->rq_export = class_conn2export(
1373 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1374 if (reqcopy->rq_export == NULL)
1375 GOTO(out, rc = -ENODEV);
1378 class_export_rpc_inc(reqcopy->rq_export);
1379 if (reqcopy->rq_export->exp_obd &&
1380 reqcopy->rq_export->exp_obd->obd_fail)
1381 GOTO(out_put, rc = -ENODEV);
1383 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1387 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1390 /* Adjust our own deadline to what we told the client */
1391 req->rq_deadline = newdl;
1392 req->rq_early_count++; /* number sent, server side */
1394 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1397 /* Free the (early) reply state from lustre_pack_reply.
1398 (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
1399 ptlrpc_req_drop_rs(reqcopy);
1402 class_export_rpc_dec(reqcopy->rq_export);
1403 class_export_put(reqcopy->rq_export);
1405 sptlrpc_svc_ctx_decref(reqcopy);
1406 OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1408 ptlrpc_request_cache_free(reqcopy);
1412 /* Send early replies to everybody expiring within at_early_margin
1413 asking for at_extra time */
1414 static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1416 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1417 struct ptlrpc_request *rq, *n;
1418 struct list_head work_list;
1421 time_t now = cfs_time_current_sec();
1422 cfs_duration_t delay;
1423 int first, counter = 0;
1426 spin_lock(&svcpt->scp_at_lock);
1427 if (svcpt->scp_at_check == 0) {
1428 spin_unlock(&svcpt->scp_at_lock);
1431 delay = cfs_time_sub(cfs_time_current(), svcpt->scp_at_checktime);
1432 svcpt->scp_at_check = 0;
1434 if (array->paa_count == 0) {
1435 spin_unlock(&svcpt->scp_at_lock);
1439 /* The timer went off, but maybe the nearest rpc already completed. */
1440 first = array->paa_deadline - now;
1441 if (first > at_early_margin) {
1442 /* We've still got plenty of time. Reset the timer. */
1443 ptlrpc_at_set_timer(svcpt);
1444 spin_unlock(&svcpt->scp_at_lock);
1448 /* We're close to a timeout, and we don't know how much longer the
1449 server will take. Send early replies to everyone expiring soon. */
1450 INIT_LIST_HEAD(&work_list);
1452 index = (unsigned long)array->paa_deadline % array->paa_size;
1453 count = array->paa_count;
1455 count -= array->paa_reqs_count[index];
1456 list_for_each_entry_safe(rq, n,
1457 &array->paa_reqs_array[index],
1459 if (rq->rq_deadline > now + at_early_margin) {
1460 /* update the earliest deadline */
1461 if (deadline == -1 ||
1462 rq->rq_deadline < deadline)
1463 deadline = rq->rq_deadline;
1467 ptlrpc_at_remove_timed(rq);
1469 * ptlrpc_server_drop_request() may drop
1470 * refcount to 0 already. Let's check this and
1471 * don't add entry to work_list
1473 if (likely(atomic_inc_not_zero(&rq->rq_refcount)))
1474 list_add(&rq->rq_timed_list, &work_list);
1478 if (++index >= array->paa_size)
1481 array->paa_deadline = deadline;
1482 /* we have a new earliest deadline, restart the timer */
1483 ptlrpc_at_set_timer(svcpt);
1485 spin_unlock(&svcpt->scp_at_lock);
1487 CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
1488 "replies\n", first, at_extra, counter);
1490 /* We're already past request deadlines before we even get a
1491 chance to send early replies */
1492 LCONSOLE_WARN("%s: This server is not able to keep up with "
1493 "request traffic (cpu-bound).\n",
1494 svcpt->scp_service->srv_name);
1495 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, "
1496 "delay="CFS_DURATION_T"(jiff)\n",
1497 counter, svcpt->scp_nreqs_incoming,
1498 svcpt->scp_nreqs_active,
1499 at_get(&svcpt->scp_at_estimate), delay);
1502 /* we took additional refcount so entries can't be deleted from list, no
1503 * locking is needed */
1504 while (!list_empty(&work_list)) {
1505 rq = list_entry(work_list.next, struct ptlrpc_request,
1507 list_del_init(&rq->rq_timed_list);
1509 if (ptlrpc_at_send_early_reply(rq) == 0)
1510 ptlrpc_at_add_timed(rq);
1512 ptlrpc_server_drop_request(rq);
1515 RETURN(1); /* return "did_something" for liblustre */
1518 /* Check if we are already handling earlier incarnation of this request.
1519 * Called under &req->rq_export->exp_rpc_lock locked */
1520 static int ptlrpc_server_check_resend_in_progress(struct ptlrpc_request *req)
1522 struct ptlrpc_request *tmp = NULL;
1524 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ||
1525 (atomic_read(&req->rq_export->exp_rpc_count) == 0))
1528 /* bulk request are aborted upon reconnect, don't try to
1530 if (req->rq_bulk_write || req->rq_bulk_read)
1533 /* This list should not be longer than max_requests in
1534 * flights on the client, so it is not all that long.
1535 * Also we only hit this codepath in case of a resent
1536 * request which makes it even more rarely hit */
1537 list_for_each_entry(tmp, &req->rq_export->exp_reg_rpcs,
1539 /* Found duplicate one */
1540 if (tmp->rq_xid == req->rq_xid)
1543 list_for_each_entry(tmp, &req->rq_export->exp_hp_rpcs,
1545 /* Found duplicate one */
1546 if (tmp->rq_xid == req->rq_xid)
1552 DEBUG_REQ(D_HA, req, "Found duplicate req in processing");
1553 DEBUG_REQ(D_HA, tmp, "Request being processed");
1558 * Check if a request should be assigned with a high priority.
1560 * \retval < 0: error occurred
1561 * 0: normal RPC request
1562 * +1: high priority request
1564 static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt,
1565 struct ptlrpc_request *req)
1570 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL) {
1571 rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
1578 if (req->rq_export != NULL && req->rq_ops != NULL) {
1579 /* Perform request specific check. We should do this
1580 * check before the request is added into exp_hp_rpcs
1581 * list otherwise it may hit swab race at LU-1044. */
1582 if (req->rq_ops->hpreq_check != NULL) {
1583 rc = req->rq_ops->hpreq_check(req);
1584 if (rc == -ESTALE) {
1585 req->rq_status = rc;
1588 /** can only return error,
1589 * 0 for normal request,
1590 * or 1 for high priority request */
1598 /** Remove the request from the export list. */
1599 static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req)
1602 if (req->rq_export) {
1603 /* refresh lock timeout again so that client has more
1604 * room to send lock cancel RPC. */
1605 if (req->rq_ops && req->rq_ops->hpreq_fini)
1606 req->rq_ops->hpreq_fini(req);
1608 spin_lock_bh(&req->rq_export->exp_rpc_lock);
1609 list_del_init(&req->rq_exp_list);
1610 spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1615 static int ptlrpc_hpreq_check(struct ptlrpc_request *req)
1620 static struct ptlrpc_hpreq_ops ptlrpc_hpreq_common = {
1621 .hpreq_check = ptlrpc_hpreq_check,
1624 /* Hi-Priority RPC check by RPC operation code. */
1625 int ptlrpc_hpreq_handler(struct ptlrpc_request *req)
1627 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1629 /* Check for export to let only reconnects for not yet evicted
1630 * export to become a HP rpc. */
1631 if ((req->rq_export != NULL) &&
1632 (opc == OBD_PING || opc == MDS_CONNECT || opc == OST_CONNECT))
1633 req->rq_ops = &ptlrpc_hpreq_common;
1637 EXPORT_SYMBOL(ptlrpc_hpreq_handler);
1639 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1640 struct ptlrpc_request *req)
1646 rc = ptlrpc_server_hpreq_init(svcpt, req);
1651 ptlrpc_nrs_req_initialize(svcpt, req, hp);
1653 if (req->rq_export != NULL) {
1654 struct obd_export *exp = req->rq_export;
1656 /* do search for duplicated xid and the adding to the list
1658 spin_lock_bh(&exp->exp_rpc_lock);
1659 rc = ptlrpc_server_check_resend_in_progress(req);
1661 spin_unlock_bh(&exp->exp_rpc_lock);
1663 ptlrpc_nrs_req_finalize(req);
1667 if (hp || req->rq_ops != NULL)
1668 list_add(&req->rq_exp_list, &exp->exp_hp_rpcs);
1670 list_add(&req->rq_exp_list, &exp->exp_reg_rpcs);
1671 spin_unlock_bh(&exp->exp_rpc_lock);
1674 /* the current thread is not the processing thread for this request
1675 * since that, but request is in exp_hp_list and can be find there.
1676 * Remove all relations between request and old thread. */
1677 req->rq_svc_thread->t_env->le_ses = NULL;
1678 req->rq_svc_thread = NULL;
1679 req->rq_session.lc_thread = NULL;
1681 ptlrpc_nrs_req_add(svcpt, req, hp);
1687 * Allow to handle high priority request
1688 * User can call it w/o any lock but need to hold
1689 * ptlrpc_service_part::scp_req_lock to get reliable result
1691 static bool ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1694 int running = svcpt->scp_nthrs_running;
1696 if (!nrs_svcpt_has_hp(svcpt))
1702 if (ptlrpc_nrs_req_throttling_nolock(svcpt, true))
1705 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1706 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1707 /* leave just 1 thread for normal RPCs */
1708 running = PTLRPC_NTHRS_INIT;
1709 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1713 if (svcpt->scp_nreqs_active >= running - 1)
1716 if (svcpt->scp_nhreqs_active == 0)
1719 return !ptlrpc_nrs_req_pending_nolock(svcpt, false) ||
1720 svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1723 static bool ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1726 return ptlrpc_server_allow_high(svcpt, force) &&
1727 ptlrpc_nrs_req_pending_nolock(svcpt, true);
1731 * Only allow normal priority requests on a service that has a high-priority
1732 * queue if forced (i.e. cleanup), if there are other high priority requests
1733 * already being processed (i.e. those threads can service more high-priority
1734 * requests), or if there are enough idle threads that a later thread can do
1735 * a high priority request.
1736 * User can call it w/o any lock but need to hold
1737 * ptlrpc_service_part::scp_req_lock to get reliable result
1739 static bool ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1742 int running = svcpt->scp_nthrs_running;
1743 if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
1744 CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
1745 /* leave just 1 thread for normal RPCs */
1746 running = PTLRPC_NTHRS_INIT;
1747 if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
1754 if (ptlrpc_nrs_req_throttling_nolock(svcpt, false))
1757 if (svcpt->scp_nreqs_active < running - 2)
1760 if (svcpt->scp_nreqs_active >= running - 1)
1763 return svcpt->scp_nhreqs_active > 0 || !nrs_svcpt_has_hp(svcpt);
1766 static bool ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1769 return ptlrpc_server_allow_normal(svcpt, force) &&
1770 ptlrpc_nrs_req_pending_nolock(svcpt, false);
1774 * Returns true if there are requests available in incoming
1775 * request queue for processing and it is allowed to fetch them.
1776 * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1777 * to get reliable result
1778 * \see ptlrpc_server_allow_normal
1779 * \see ptlrpc_server_allow high
1782 ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt, bool force)
1784 return ptlrpc_server_high_pending(svcpt, force) ||
1785 ptlrpc_server_normal_pending(svcpt, force);
1789 * Fetch a request for processing from queue of unprocessed requests.
1790 * Favors high-priority requests.
1791 * Returns a pointer to fetched request.
1793 static struct ptlrpc_request *
1794 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force)
1796 struct ptlrpc_request *req = NULL;
1799 spin_lock(&svcpt->scp_req_lock);
1801 if (ptlrpc_server_high_pending(svcpt, force)) {
1802 req = ptlrpc_nrs_req_get_nolock(svcpt, true, force);
1804 svcpt->scp_hreq_count++;
1809 if (ptlrpc_server_normal_pending(svcpt, force)) {
1810 req = ptlrpc_nrs_req_get_nolock(svcpt, false, force);
1812 svcpt->scp_hreq_count = 0;
1817 spin_unlock(&svcpt->scp_req_lock);
1821 svcpt->scp_nreqs_active++;
1823 svcpt->scp_nhreqs_active++;
1825 spin_unlock(&svcpt->scp_req_lock);
1827 if (likely(req->rq_export))
1828 class_export_rpc_inc(req->rq_export);
1834 * Handle freshly incoming reqs, add to timed early reply list,
1835 * pass on to regular request queue.
1836 * All incoming requests pass through here before getting into
1837 * ptlrpc_server_handle_req later on.
1840 ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt,
1841 struct ptlrpc_thread *thread)
1843 struct ptlrpc_service *svc = svcpt->scp_service;
1844 struct ptlrpc_request *req;
1849 spin_lock(&svcpt->scp_lock);
1850 if (list_empty(&svcpt->scp_req_incoming)) {
1851 spin_unlock(&svcpt->scp_lock);
1855 req = list_entry(svcpt->scp_req_incoming.next,
1856 struct ptlrpc_request, rq_list);
1857 list_del_init(&req->rq_list);
1858 svcpt->scp_nreqs_incoming--;
1859 /* Consider this still a "queued" request as far as stats are
1861 spin_unlock(&svcpt->scp_lock);
1863 /* go through security check/transform */
1864 rc = sptlrpc_svc_unwrap_request(req);
1868 case SECSVC_COMPLETE:
1869 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1878 * for null-flavored rpc, msg has been unpacked by sptlrpc, although
1879 * redo it wouldn't be harmful.
1881 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
1882 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
1884 CERROR("error unpacking request: ptl %d from %s "
1885 "x%llu\n", svc->srv_req_portal,
1886 libcfs_id2str(req->rq_peer), req->rq_xid);
1891 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1893 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
1894 "%llu\n", svc->srv_req_portal,
1895 libcfs_id2str(req->rq_peer), req->rq_xid);
1899 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
1900 lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
1901 CERROR("drop incoming rpc opc %u, x%llu\n",
1902 cfs_fail_val, req->rq_xid);
1907 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1908 CERROR("wrong packet type received (type=%u) from %s\n",
1909 lustre_msg_get_type(req->rq_reqmsg),
1910 libcfs_id2str(req->rq_peer));
1914 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1918 req->rq_bulk_write = 1;
1922 case MGS_CONFIG_READ:
1923 req->rq_bulk_read = 1;
1927 CDEBUG(D_RPCTRACE, "got req x%llu\n", req->rq_xid);
1929 req->rq_export = class_conn2export(
1930 lustre_msg_get_handle(req->rq_reqmsg));
1931 if (req->rq_export) {
1932 rc = ptlrpc_check_req(req);
1934 rc = sptlrpc_target_export_check(req->rq_export, req);
1936 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
1937 "illegal security flavor,");
1942 ptlrpc_update_export_timer(req->rq_export, 0);
1945 /* req_in handling should/must be fast */
1946 if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5)
1947 DEBUG_REQ(D_WARNING, req, "Slow req_in handling "CFS_DURATION_T"s",
1948 cfs_time_sub(cfs_time_current_sec(),
1949 req->rq_arrival_time.tv_sec));
1951 /* Set rpc server deadline and add it to the timed list */
1952 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1953 MSGHDR_AT_SUPPORT) ?
1954 /* The max time the client expects us to take */
1955 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1957 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1958 if (unlikely(deadline == 0)) {
1959 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1963 /* Skip early reply */
1964 if (OBD_FAIL_PRECHECK(OBD_FAIL_MDS_RESEND))
1965 req->rq_deadline += obd_timeout;
1967 req->rq_svc_thread = thread;
1968 if (thread != NULL) {
1969 /* initialize request session, it is needed for request
1970 * processing by target */
1971 rc = lu_context_init(&req->rq_session, LCT_SERVER_SESSION |
1974 CERROR("%s: failure to initialize session: rc = %d\n",
1975 thread->t_name, rc);
1978 req->rq_session.lc_thread = thread;
1979 lu_context_enter(&req->rq_session);
1980 thread->t_env->le_ses = &req->rq_session;
1983 ptlrpc_at_add_timed(req);
1985 /* Move it over to the request processing queue */
1986 rc = ptlrpc_server_request_add(svcpt, req);
1990 wake_up(&svcpt->scp_waitq);
1994 ptlrpc_server_finish_request(svcpt, req);
2000 * Main incoming request handling logic.
2001 * Calls handler function from service to do actual processing.
2004 ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
2005 struct ptlrpc_thread *thread)
2007 struct ptlrpc_service *svc = svcpt->scp_service;
2008 struct ptlrpc_request *request;
2009 struct timeval work_start;
2010 struct timeval work_end;
2016 request = ptlrpc_server_request_get(svcpt, false);
2017 if (request == NULL)
2020 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
2021 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
2022 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
2023 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
2025 if (unlikely(fail_opc)) {
2026 if (request->rq_export && request->rq_ops)
2027 OBD_FAIL_TIMEOUT(fail_opc, 4);
2030 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
2032 if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
2033 libcfs_debug_dumplog();
2035 do_gettimeofday(&work_start);
2036 timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
2037 if (likely(svc->srv_stats != NULL)) {
2038 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
2040 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
2041 svcpt->scp_nreqs_incoming);
2042 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
2043 svcpt->scp_nreqs_active);
2044 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
2045 at_get(&svcpt->scp_at_estimate));
2048 if (likely(request->rq_export)) {
2049 if (unlikely(ptlrpc_check_req(request)))
2051 ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
2054 /* Discard requests queued for longer than the deadline.
2055 The deadline is increased if we send an early reply. */
2056 if (cfs_time_current_sec() > request->rq_deadline) {
2057 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s"
2058 ": deadline "CFS_DURATION_T":"CFS_DURATION_T"s ago\n",
2059 libcfs_id2str(request->rq_peer),
2060 cfs_time_sub(request->rq_deadline,
2061 request->rq_arrival_time.tv_sec),
2062 cfs_time_sub(cfs_time_current_sec(),
2063 request->rq_deadline));
2067 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
2068 "%s:%s+%d:%d:x%llu:%s:%d\n", current_comm(),
2069 (request->rq_export ?
2070 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2071 (request->rq_export ?
2072 atomic_read(&request->rq_export->exp_refcount) : -99),
2073 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
2074 libcfs_id2str(request->rq_peer),
2075 lustre_msg_get_opc(request->rq_reqmsg));
2077 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
2078 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
2080 CDEBUG(D_NET, "got req %llu\n", request->rq_xid);
2082 /* re-assign request and sesson thread to the current one */
2083 request->rq_svc_thread = thread;
2084 if (thread != NULL) {
2085 LASSERT(request->rq_session.lc_thread == NULL);
2086 request->rq_session.lc_thread = thread;
2087 thread->t_env->le_ses = &request->rq_session;
2089 svc->srv_ops.so_req_handler(request);
2091 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
2094 if (unlikely(cfs_time_current_sec() > request->rq_deadline)) {
2095 DEBUG_REQ(D_WARNING, request, "Request took longer "
2096 "than estimated ("CFS_DURATION_T":"CFS_DURATION_T"s);"
2097 " client may timeout.",
2098 cfs_time_sub(request->rq_deadline,
2099 request->rq_arrival_time.tv_sec),
2100 cfs_time_sub(cfs_time_current_sec(),
2101 request->rq_deadline));
2104 do_gettimeofday(&work_end);
2105 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
2106 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
2107 "%s:%s+%d:%d:x%llu:%s:%d Request procesed in "
2108 "%ldus (%ldus total) trans %llu rc %d/%d\n",
2110 (request->rq_export ?
2111 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
2112 (request->rq_export ?
2113 atomic_read(&request->rq_export->exp_refcount) : -99),
2114 lustre_msg_get_status(request->rq_reqmsg),
2116 libcfs_id2str(request->rq_peer),
2117 lustre_msg_get_opc(request->rq_reqmsg),
2119 cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL),
2120 (request->rq_repmsg ?
2121 lustre_msg_get_transno(request->rq_repmsg) :
2122 request->rq_transno),
2124 (request->rq_repmsg ?
2125 lustre_msg_get_status(request->rq_repmsg) : -999));
2126 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
2127 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
2128 int opc = opcode_offset(op);
2129 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
2130 LASSERT(opc < LUSTRE_MAX_OPCODES);
2131 lprocfs_counter_add(svc->srv_stats,
2132 opc + EXTRA_MAX_OPCODES,
2136 if (unlikely(request->rq_early_count)) {
2137 DEBUG_REQ(D_ADAPTTO, request,
2138 "sent %d early replies before finishing in "
2140 request->rq_early_count,
2141 cfs_time_sub(work_end.tv_sec,
2142 request->rq_arrival_time.tv_sec));
2145 ptlrpc_server_finish_active_request(svcpt, request);
2151 * An internal function to process a single reply state object.
2154 ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
2156 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
2157 struct ptlrpc_service *svc = svcpt->scp_service;
2158 struct obd_export *exp;
2163 exp = rs->rs_export;
2165 LASSERT(rs->rs_difficult);
2166 LASSERT(rs->rs_scheduled);
2167 LASSERT(list_empty(&rs->rs_list));
2169 spin_lock(&exp->exp_lock);
2170 /* Noop if removed already */
2171 list_del_init(&rs->rs_exp_list);
2172 spin_unlock(&exp->exp_lock);
2174 /* The disk commit callback holds exp_uncommitted_replies_lock while it
2175 * iterates over newly committed replies, removing them from
2176 * exp_uncommitted_replies. It then drops this lock and schedules the
2177 * replies it found for handling here.
2179 * We can avoid contention for exp_uncommitted_replies_lock between the
2180 * HRT threads and further commit callbacks by checking rs_committed
2181 * which is set in the commit callback while it holds both
2182 * rs_lock and exp_uncommitted_reples.
2184 * If we see rs_committed clear, the commit callback _may_ not have
2185 * handled this reply yet and we race with it to grab
2186 * exp_uncommitted_replies_lock before removing the reply from
2187 * exp_uncommitted_replies. Note that if we lose the race and the
2188 * reply has already been removed, list_del_init() is a noop.
2190 * If we see rs_committed set, we know the commit callback is handling,
2191 * or has handled this reply since store reordering might allow us to
2192 * see rs_committed set out of sequence. But since this is done
2193 * holding rs_lock, we can be sure it has all completed once we hold
2194 * rs_lock, which we do right next.
2196 if (!rs->rs_committed) {
2197 spin_lock(&exp->exp_uncommitted_replies_lock);
2198 list_del_init(&rs->rs_obd_list);
2199 spin_unlock(&exp->exp_uncommitted_replies_lock);
2202 spin_lock(&rs->rs_lock);
2204 been_handled = rs->rs_handled;
2207 nlocks = rs->rs_nlocks; /* atomic "steal", but */
2208 rs->rs_nlocks = 0; /* locks still on rs_locks! */
2210 if (nlocks == 0 && !been_handled) {
2211 /* If we see this, we should already have seen the warning
2212 * in mds_steal_ack_locks() */
2213 CDEBUG(D_HA, "All locks stolen from rs %p x%lld.t%lld"
2216 rs->rs_xid, rs->rs_transno, rs->rs_opc,
2217 libcfs_nid2str(exp->exp_connection->c_peer.nid));
2220 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
2221 spin_unlock(&rs->rs_lock);
2223 if (!been_handled && rs->rs_on_net) {
2224 LNetMDUnlink(rs->rs_md_h);
2225 /* Ignore return code; we're racing with completion */
2228 while (nlocks-- > 0)
2229 ldlm_lock_decref(&rs->rs_locks[nlocks],
2230 rs->rs_modes[nlocks]);
2232 spin_lock(&rs->rs_lock);
2235 rs->rs_scheduled = 0;
2237 if (!rs->rs_on_net) {
2239 spin_unlock(&rs->rs_lock);
2241 class_export_put (exp);
2242 rs->rs_export = NULL;
2243 ptlrpc_rs_decref(rs);
2244 if (atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
2245 svc->srv_is_stopping)
2246 wake_up_all(&svcpt->scp_waitq);
2250 /* still on the net; callback will schedule */
2251 spin_unlock(&rs->rs_lock);
2257 ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
2259 int avail = svcpt->scp_nrqbds_posted;
2260 int low_water = test_req_buffer_pressure ? 0 :
2261 svcpt->scp_service->srv_nbuf_per_group / 2;
2263 /* NB I'm not locking; just looking. */
2265 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
2266 * allowed the request history to grow out of control. We could put a
2267 * sanity check on that here and cull some history if we need the
2270 if (avail <= low_water)
2271 ptlrpc_grow_req_bufs(svcpt, 1);
2273 if (svcpt->scp_service->srv_stats) {
2274 lprocfs_counter_add(svcpt->scp_service->srv_stats,
2275 PTLRPC_REQBUF_AVAIL_CNTR, avail);
2280 ptlrpc_retry_rqbds(void *arg)
2282 struct ptlrpc_service_part *svcpt = (struct ptlrpc_service_part *)arg;
2284 svcpt->scp_rqbd_timeout = 0;
2289 ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
2291 return svcpt->scp_nreqs_active <
2292 svcpt->scp_nthrs_running - 1 -
2293 (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL);
2297 * allowed to create more threads
2298 * user can call it w/o any lock but need to hold
2299 * ptlrpc_service_part::scp_lock to get reliable result
2302 ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
2304 return svcpt->scp_nthrs_running +
2305 svcpt->scp_nthrs_starting <
2306 svcpt->scp_service->srv_nthrs_cpt_limit;
2310 * too many requests and allowed to create more threads
2313 ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
2315 return !ptlrpc_threads_enough(svcpt) &&
2316 ptlrpc_threads_increasable(svcpt);
2320 ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2322 return thread_is_stopping(thread) ||
2323 thread->t_svcpt->scp_service->srv_is_stopping;
2327 ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
2329 return !list_empty(&svcpt->scp_rqbd_idle) &&
2330 svcpt->scp_rqbd_timeout == 0;
2334 ptlrpc_at_check(struct ptlrpc_service_part *svcpt)
2336 return svcpt->scp_at_check;
2340 * requests wait on preprocessing
2341 * user can call it w/o any lock but need to hold
2342 * ptlrpc_service_part::scp_lock to get reliable result
2345 ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
2347 return !list_empty(&svcpt->scp_req_incoming);
2350 static __attribute__((__noinline__)) int
2351 ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
2352 struct ptlrpc_thread *thread)
2354 /* Don't exit while there are replies to be handled */
2355 struct l_wait_info lwi = LWI_TIMEOUT(svcpt->scp_rqbd_timeout,
2356 ptlrpc_retry_rqbds, svcpt);
2358 lc_watchdog_disable(thread->t_watchdog);
2362 l_wait_event_exclusive_head(svcpt->scp_waitq,
2363 ptlrpc_thread_stopping(thread) ||
2364 ptlrpc_server_request_incoming(svcpt) ||
2365 ptlrpc_server_request_pending(svcpt, false) ||
2366 ptlrpc_rqbd_pending(svcpt) ||
2367 ptlrpc_at_check(svcpt), &lwi);
2369 if (ptlrpc_thread_stopping(thread))
2372 lc_watchdog_touch(thread->t_watchdog,
2373 ptlrpc_server_get_timeout(svcpt));
2378 * Main thread body for service threads.
2379 * Waits in a loop waiting for new requests to process to appear.
2380 * Every time an incoming requests is added to its queue, a waitq
2381 * is woken up and one of the threads will handle it.
2383 static int ptlrpc_main(void *arg)
2385 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
2386 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2387 struct ptlrpc_service *svc = svcpt->scp_service;
2388 struct ptlrpc_reply_state *rs;
2389 struct group_info *ginfo = NULL;
2391 int counter = 0, rc = 0;
2394 thread->t_pid = current_pid();
2395 unshare_fs_struct();
2397 /* NB: we will call cfs_cpt_bind() for all threads, because we
2398 * might want to run lustre server only on a subset of system CPUs,
2399 * in that case ->scp_cpt is CFS_CPT_ANY */
2400 rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
2402 CWARN("%s: failed to bind %s on CPT %d\n",
2403 svc->srv_name, thread->t_name, svcpt->scp_cpt);
2406 ginfo = groups_alloc(0);
2412 set_current_groups(ginfo);
2413 put_group_info(ginfo);
2415 if (svc->srv_ops.so_thr_init != NULL) {
2416 rc = svc->srv_ops.so_thr_init(thread);
2427 rc = lu_context_init(&env->le_ctx,
2428 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2432 thread->t_env = env;
2433 env->le_ctx.lc_thread = thread;
2434 env->le_ctx.lc_cookie = 0x6;
2436 while (!list_empty(&svcpt->scp_rqbd_idle)) {
2437 rc = ptlrpc_server_post_idle_rqbds(svcpt);
2441 CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
2442 svc->srv_name, svcpt->scp_cpt, rc);
2446 /* Alloc reply state structure for this one */
2447 OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2453 spin_lock(&svcpt->scp_lock);
2455 LASSERT(thread_is_starting(thread));
2456 thread_clear_flags(thread, SVC_STARTING);
2458 LASSERT(svcpt->scp_nthrs_starting == 1);
2459 svcpt->scp_nthrs_starting--;
2461 /* SVC_STOPPING may already be set here if someone else is trying
2462 * to stop the service while this new thread has been dynamically
2463 * forked. We still set SVC_RUNNING to let our creator know that
2464 * we are now running, however we will exit as soon as possible */
2465 thread_add_flags(thread, SVC_RUNNING);
2466 svcpt->scp_nthrs_running++;
2467 spin_unlock(&svcpt->scp_lock);
2469 /* wake up our creator in case he's still waiting. */
2470 wake_up(&thread->t_ctl_waitq);
2472 thread->t_watchdog = lc_watchdog_add(ptlrpc_server_get_timeout(svcpt),
2475 spin_lock(&svcpt->scp_rep_lock);
2476 list_add(&rs->rs_list, &svcpt->scp_rep_idle);
2477 wake_up(&svcpt->scp_rep_waitq);
2478 spin_unlock(&svcpt->scp_rep_lock);
2480 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2481 svcpt->scp_nthrs_running);
2483 /* XXX maintain a list of all managed devices: insert here */
2484 while (!ptlrpc_thread_stopping(thread)) {
2485 if (ptlrpc_wait_event(svcpt, thread))
2488 ptlrpc_check_rqbd_pool(svcpt);
2490 if (ptlrpc_threads_need_create(svcpt)) {
2491 /* Ignore return code - we tried... */
2492 ptlrpc_start_thread(svcpt, 0);
2495 /* reset le_ses to initial state */
2497 /* Process all incoming reqs before handling any */
2498 if (ptlrpc_server_request_incoming(svcpt)) {
2499 lu_context_enter(&env->le_ctx);
2500 ptlrpc_server_handle_req_in(svcpt, thread);
2501 lu_context_exit(&env->le_ctx);
2503 /* but limit ourselves in case of flood */
2504 if (counter++ < 100)
2509 if (ptlrpc_at_check(svcpt))
2510 ptlrpc_at_check_timed(svcpt);
2512 if (ptlrpc_server_request_pending(svcpt, false)) {
2513 lu_context_enter(&env->le_ctx);
2514 ptlrpc_server_handle_request(svcpt, thread);
2515 lu_context_exit(&env->le_ctx);
2518 if (ptlrpc_rqbd_pending(svcpt) &&
2519 ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
2520 /* I just failed to repost request buffers.
2521 * Wait for a timeout (unless something else
2522 * happens) before I try again */
2523 svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
2524 CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
2525 svcpt->scp_nrqbds_posted);
2529 lc_watchdog_delete(thread->t_watchdog);
2530 thread->t_watchdog = NULL;
2534 * deconstruct service specific state created by ptlrpc_start_thread()
2536 if (svc->srv_ops.so_thr_done != NULL)
2537 svc->srv_ops.so_thr_done(thread);
2540 lu_context_fini(&env->le_ctx);
2544 CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n",
2545 thread, thread->t_pid, thread->t_id, rc);
2547 spin_lock(&svcpt->scp_lock);
2548 if (thread_test_and_clear_flags(thread, SVC_STARTING))
2549 svcpt->scp_nthrs_starting--;
2551 if (thread_test_and_clear_flags(thread, SVC_RUNNING)) {
2552 /* must know immediately */
2553 svcpt->scp_nthrs_running--;
2557 thread_add_flags(thread, SVC_STOPPED);
2559 wake_up(&thread->t_ctl_waitq);
2560 spin_unlock(&svcpt->scp_lock);
2565 static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
2566 struct list_head *replies)
2570 spin_lock(&hrt->hrt_lock);
2572 list_splice_init(&hrt->hrt_queue, replies);
2573 result = ptlrpc_hr.hr_stopping || !list_empty(replies);
2575 spin_unlock(&hrt->hrt_lock);
2580 * Main body of "handle reply" function.
2581 * It processes acked reply states
2583 static int ptlrpc_hr_main(void *arg)
2585 struct ptlrpc_hr_thread *hrt = (struct ptlrpc_hr_thread *)arg;
2586 struct ptlrpc_hr_partition *hrp = hrt->hrt_partition;
2587 struct list_head replies;
2588 char threadname[20];
2591 INIT_LIST_HEAD(&replies);
2592 snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
2593 hrp->hrp_cpt, hrt->hrt_id);
2594 unshare_fs_struct();
2596 rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
2598 CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
2599 threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
2602 atomic_inc(&hrp->hrp_nstarted);
2603 wake_up(&ptlrpc_hr.hr_waitq);
2605 while (!ptlrpc_hr.hr_stopping) {
2606 l_wait_condition(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
2608 while (!list_empty(&replies)) {
2609 struct ptlrpc_reply_state *rs;
2611 rs = list_entry(replies.prev,
2612 struct ptlrpc_reply_state,
2614 list_del_init(&rs->rs_list);
2615 ptlrpc_handle_rs(rs);
2619 atomic_inc(&hrp->hrp_nstopped);
2620 wake_up(&ptlrpc_hr.hr_waitq);
2625 static void ptlrpc_stop_hr_threads(void)
2627 struct ptlrpc_hr_partition *hrp;
2631 ptlrpc_hr.hr_stopping = 1;
2633 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2634 if (hrp->hrp_thrs == NULL)
2635 continue; /* uninitialized */
2636 for (j = 0; j < hrp->hrp_nthrs; j++)
2637 wake_up_all(&hrp->hrp_thrs[j].hrt_waitq);
2640 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2641 if (hrp->hrp_thrs == NULL)
2642 continue; /* uninitialized */
2643 wait_event(ptlrpc_hr.hr_waitq,
2644 atomic_read(&hrp->hrp_nstopped) ==
2645 atomic_read(&hrp->hrp_nstarted));
2649 static int ptlrpc_start_hr_threads(void)
2651 struct ptlrpc_hr_partition *hrp;
2656 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2659 for (j = 0; j < hrp->hrp_nthrs; j++) {
2660 struct ptlrpc_hr_thread *hrt = &hrp->hrp_thrs[j];
2661 struct task_struct *task;
2663 task = kthread_run(ptlrpc_hr_main,
2665 "ptlrpc_hr%02d_%03d",
2674 wait_event(ptlrpc_hr.hr_waitq,
2675 atomic_read(&hrp->hrp_nstarted) == j);
2678 CERROR("cannot start reply handler thread %d:%d: "
2679 "rc = %d\n", i, j, rc);
2680 ptlrpc_stop_hr_threads();
2688 static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
2690 struct l_wait_info lwi = { 0 };
2691 struct ptlrpc_thread *thread;
2692 struct list_head zombie;
2696 CDEBUG(D_INFO, "Stopping threads for service %s\n",
2697 svcpt->scp_service->srv_name);
2699 INIT_LIST_HEAD(&zombie);
2700 spin_lock(&svcpt->scp_lock);
2701 /* let the thread know that we would like it to stop asap */
2702 list_for_each_entry(thread, &svcpt->scp_threads, t_link) {
2703 CDEBUG(D_INFO, "Stopping thread %s #%u\n",
2704 svcpt->scp_service->srv_thread_name, thread->t_id);
2705 thread_add_flags(thread, SVC_STOPPING);
2708 wake_up_all(&svcpt->scp_waitq);
2710 while (!list_empty(&svcpt->scp_threads)) {
2711 thread = list_entry(svcpt->scp_threads.next,
2712 struct ptlrpc_thread, t_link);
2713 if (thread_is_stopped(thread)) {
2714 list_del(&thread->t_link);
2715 list_add(&thread->t_link, &zombie);
2718 spin_unlock(&svcpt->scp_lock);
2720 CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
2721 svcpt->scp_service->srv_thread_name, thread->t_id);
2722 l_wait_event(thread->t_ctl_waitq,
2723 thread_is_stopped(thread), &lwi);
2725 spin_lock(&svcpt->scp_lock);
2728 spin_unlock(&svcpt->scp_lock);
2730 while (!list_empty(&zombie)) {
2731 thread = list_entry(zombie.next,
2732 struct ptlrpc_thread, t_link);
2733 list_del(&thread->t_link);
2734 OBD_FREE_PTR(thread);
2740 * Stops all threads of a particular service \a svc
2742 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2744 struct ptlrpc_service_part *svcpt;
2748 ptlrpc_service_for_each_part(svcpt, i, svc) {
2749 if (svcpt->scp_service != NULL)
2750 ptlrpc_svcpt_stop_threads(svcpt);
2756 int ptlrpc_start_threads(struct ptlrpc_service *svc)
2763 /* We require 2 threads min, see note in ptlrpc_server_handle_request */
2764 LASSERT(svc->srv_nthrs_cpt_init >= PTLRPC_NTHRS_INIT);
2766 for (i = 0; i < svc->srv_ncpts; i++) {
2767 for (j = 0; j < svc->srv_nthrs_cpt_init; j++) {
2768 rc = ptlrpc_start_thread(svc->srv_parts[i], 1);
2774 /* We have enough threads, don't start more. b=15759 */
2781 CERROR("cannot start %s thread #%d_%d: rc %d\n",
2782 svc->srv_thread_name, i, j, rc);
2783 ptlrpc_stop_all_threads(svc);
2787 int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
2789 struct l_wait_info lwi = { 0 };
2790 struct ptlrpc_thread *thread;
2791 struct ptlrpc_service *svc;
2792 struct task_struct *task;
2796 LASSERT(svcpt != NULL);
2798 svc = svcpt->scp_service;
2800 CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n",
2801 svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running,
2802 svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit);
2805 if (unlikely(svc->srv_is_stopping))
2808 if (!ptlrpc_threads_increasable(svcpt) ||
2809 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
2810 svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1))
2813 OBD_CPT_ALLOC_PTR(thread, svc->srv_cptable, svcpt->scp_cpt);
2816 init_waitqueue_head(&thread->t_ctl_waitq);
2818 spin_lock(&svcpt->scp_lock);
2819 if (!ptlrpc_threads_increasable(svcpt)) {
2820 spin_unlock(&svcpt->scp_lock);
2821 OBD_FREE_PTR(thread);
2825 if (svcpt->scp_nthrs_starting != 0) {
2826 /* serialize starting because some modules (obdfilter)
2827 * might require unique and contiguous t_id */
2828 LASSERT(svcpt->scp_nthrs_starting == 1);
2829 spin_unlock(&svcpt->scp_lock);
2830 OBD_FREE_PTR(thread);
2832 CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n",
2833 svc->srv_thread_name, svcpt->scp_thr_nextid);
2838 CDEBUG(D_INFO, "Creating thread %s #%d race, retry later\n",
2839 svc->srv_thread_name, svcpt->scp_thr_nextid);
2843 svcpt->scp_nthrs_starting++;
2844 thread->t_id = svcpt->scp_thr_nextid++;
2845 thread_add_flags(thread, SVC_STARTING);
2846 thread->t_svcpt = svcpt;
2848 list_add(&thread->t_link, &svcpt->scp_threads);
2849 spin_unlock(&svcpt->scp_lock);
2851 if (svcpt->scp_cpt >= 0) {
2852 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s%02d_%03d",
2853 svc->srv_thread_name, svcpt->scp_cpt, thread->t_id);
2855 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s_%04d",
2856 svc->srv_thread_name, thread->t_id);
2859 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name);
2860 task = kthread_run(ptlrpc_main, thread, "%s", thread->t_name);
2863 CERROR("cannot start thread '%s': rc = %d\n",
2864 thread->t_name, rc);
2865 spin_lock(&svcpt->scp_lock);
2866 --svcpt->scp_nthrs_starting;
2867 if (thread_is_stopping(thread)) {
2868 /* this ptlrpc_thread is being hanled
2869 * by ptlrpc_svcpt_stop_threads now
2871 thread_add_flags(thread, SVC_STOPPED);
2872 wake_up(&thread->t_ctl_waitq);
2873 spin_unlock(&svcpt->scp_lock);
2875 list_del(&thread->t_link);
2876 spin_unlock(&svcpt->scp_lock);
2877 OBD_FREE_PTR(thread);
2885 l_wait_event(thread->t_ctl_waitq,
2886 thread_is_running(thread) || thread_is_stopped(thread),
2889 rc = thread_is_stopped(thread) ? thread->t_id : 0;
2893 int ptlrpc_hr_init(void)
2895 struct ptlrpc_hr_partition *hrp;
2896 struct ptlrpc_hr_thread *hrt;
2903 memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
2904 ptlrpc_hr.hr_cpt_table = cfs_cpt_table;
2906 ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
2908 if (ptlrpc_hr.hr_partitions == NULL)
2911 init_waitqueue_head(&ptlrpc_hr.hr_waitq);
2913 weight = cfs_cpu_ht_nsiblings(0);
2915 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2918 atomic_set(&hrp->hrp_nstarted, 0);
2919 atomic_set(&hrp->hrp_nstopped, 0);
2921 hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, i);
2923 hrp->hrp_nthrs /= weight;
2924 if (hrp->hrp_nthrs == 0)
2927 OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, i,
2928 hrp->hrp_nthrs * sizeof(*hrt));
2929 if (hrp->hrp_thrs == NULL)
2930 GOTO(out, rc = -ENOMEM);
2932 for (j = 0; j < hrp->hrp_nthrs; j++) {
2933 hrt = &hrp->hrp_thrs[j];
2936 hrt->hrt_partition = hrp;
2937 init_waitqueue_head(&hrt->hrt_waitq);
2938 spin_lock_init(&hrt->hrt_lock);
2939 INIT_LIST_HEAD(&hrt->hrt_queue);
2943 rc = ptlrpc_start_hr_threads();
2950 void ptlrpc_hr_fini(void)
2952 struct ptlrpc_hr_partition *hrp;
2955 if (ptlrpc_hr.hr_partitions == NULL)
2958 ptlrpc_stop_hr_threads();
2960 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2961 if (hrp->hrp_thrs != NULL) {
2962 OBD_FREE(hrp->hrp_thrs,
2963 hrp->hrp_nthrs * sizeof(hrp->hrp_thrs[0]));
2967 cfs_percpt_free(ptlrpc_hr.hr_partitions);
2968 ptlrpc_hr.hr_partitions = NULL;
2973 * Wait until all already scheduled replies are processed.
2975 static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
2979 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
2982 rc = l_wait_event(svcpt->scp_waitq,
2983 atomic_read(&svcpt->scp_nreps_difficult) == 0, &lwi);
2986 CWARN("Unexpectedly long timeout %s %p\n",
2987 svcpt->scp_service->srv_name, svcpt->scp_service);
2992 ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
2994 struct ptlrpc_service_part *svcpt;
2997 /* early disarm AT timer... */
2998 ptlrpc_service_for_each_part(svcpt, i, svc) {
2999 if (svcpt->scp_service != NULL)
3000 del_timer(&svcpt->scp_at_timer);
3005 ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
3007 struct ptlrpc_service_part *svcpt;
3008 struct ptlrpc_request_buffer_desc *rqbd;
3009 struct l_wait_info lwi;
3013 /* All history will be culled when the next request buffer is
3014 * freed in ptlrpc_service_purge_all() */
3015 svc->srv_hist_nrqbds_cpt_max = 0;
3017 rc = LNetClearLazyPortal(svc->srv_req_portal);
3020 ptlrpc_service_for_each_part(svcpt, i, svc) {
3021 if (svcpt->scp_service == NULL)
3024 /* Unlink all the request buffers. This forces a 'final'
3025 * event with its 'unlink' flag set for each posted rqbd */
3026 list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
3028 rc = LNetMDUnlink(rqbd->rqbd_md_h);
3029 LASSERT(rc == 0 || rc == -ENOENT);
3033 ptlrpc_service_for_each_part(svcpt, i, svc) {
3034 if (svcpt->scp_service == NULL)
3037 /* Wait for the network to release any buffers
3038 * it's currently filling */
3039 spin_lock(&svcpt->scp_lock);
3040 while (svcpt->scp_nrqbds_posted != 0) {
3041 spin_unlock(&svcpt->scp_lock);
3042 /* Network access will complete in finite time but
3043 * the HUGE timeout lets us CWARN for visibility
3044 * of sluggish NALs */
3045 lwi = LWI_TIMEOUT_INTERVAL(
3046 cfs_time_seconds(LONG_UNLINK),
3047 cfs_time_seconds(1), NULL, NULL);
3048 rc = l_wait_event(svcpt->scp_waitq,
3049 svcpt->scp_nrqbds_posted == 0, &lwi);
3050 if (rc == -ETIMEDOUT) {
3051 CWARN("Service %s waiting for "
3052 "request buffers\n",
3053 svcpt->scp_service->srv_name);
3055 spin_lock(&svcpt->scp_lock);
3057 spin_unlock(&svcpt->scp_lock);
3062 ptlrpc_service_purge_all(struct ptlrpc_service *svc)
3064 struct ptlrpc_service_part *svcpt;
3065 struct ptlrpc_request_buffer_desc *rqbd;
3066 struct ptlrpc_request *req;
3067 struct ptlrpc_reply_state *rs;
3070 ptlrpc_service_for_each_part(svcpt, i, svc) {
3071 if (svcpt->scp_service == NULL)
3074 spin_lock(&svcpt->scp_rep_lock);
3075 while (!list_empty(&svcpt->scp_rep_active)) {
3076 rs = list_entry(svcpt->scp_rep_active.next,
3077 struct ptlrpc_reply_state, rs_list);
3078 spin_lock(&rs->rs_lock);
3079 ptlrpc_schedule_difficult_reply(rs);
3080 spin_unlock(&rs->rs_lock);
3082 spin_unlock(&svcpt->scp_rep_lock);
3084 /* purge the request queue. NB No new replies (rqbds
3085 * all unlinked) and no service threads, so I'm the only
3086 * thread noodling the request queue now */
3087 while (!list_empty(&svcpt->scp_req_incoming)) {
3088 req = list_entry(svcpt->scp_req_incoming.next,
3089 struct ptlrpc_request, rq_list);
3091 list_del(&req->rq_list);
3092 svcpt->scp_nreqs_incoming--;
3093 ptlrpc_server_finish_request(svcpt, req);
3096 while (ptlrpc_server_request_pending(svcpt, true)) {
3097 req = ptlrpc_server_request_get(svcpt, true);
3098 ptlrpc_server_finish_active_request(svcpt, req);
3101 LASSERT(list_empty(&svcpt->scp_rqbd_posted));
3102 LASSERT(svcpt->scp_nreqs_incoming == 0);
3103 LASSERT(svcpt->scp_nreqs_active == 0);
3104 /* history should have been culled by
3105 * ptlrpc_server_finish_request */
3106 LASSERT(svcpt->scp_hist_nrqbds == 0);
3108 /* Now free all the request buffers since nothing
3109 * references them any more... */
3111 while (!list_empty(&svcpt->scp_rqbd_idle)) {
3112 rqbd = list_entry(svcpt->scp_rqbd_idle.next,
3113 struct ptlrpc_request_buffer_desc,
3115 ptlrpc_free_rqbd(rqbd);
3117 ptlrpc_wait_replies(svcpt);
3119 while (!list_empty(&svcpt->scp_rep_idle)) {
3120 rs = list_entry(svcpt->scp_rep_idle.next,
3121 struct ptlrpc_reply_state,
3123 list_del(&rs->rs_list);
3124 OBD_FREE_LARGE(rs, svc->srv_max_reply_size);
3130 ptlrpc_service_free(struct ptlrpc_service *svc)
3132 struct ptlrpc_service_part *svcpt;
3133 struct ptlrpc_at_array *array;
3136 ptlrpc_service_for_each_part(svcpt, i, svc) {
3137 if (svcpt->scp_service == NULL)
3140 /* In case somebody rearmed this in the meantime */
3141 del_timer(&svcpt->scp_at_timer);
3142 array = &svcpt->scp_at_array;
3144 if (array->paa_reqs_array != NULL) {
3145 OBD_FREE(array->paa_reqs_array,
3146 sizeof(struct list_head) * array->paa_size);
3147 array->paa_reqs_array = NULL;
3150 if (array->paa_reqs_count != NULL) {
3151 OBD_FREE(array->paa_reqs_count,
3152 sizeof(__u32) * array->paa_size);
3153 array->paa_reqs_count = NULL;
3157 ptlrpc_service_for_each_part(svcpt, i, svc)
3158 OBD_FREE_PTR(svcpt);
3160 if (svc->srv_cpts != NULL)
3161 cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts);
3163 OBD_FREE(svc, offsetof(struct ptlrpc_service,
3164 srv_parts[svc->srv_ncpts]));
3167 int ptlrpc_unregister_service(struct ptlrpc_service *service)
3171 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
3173 service->srv_is_stopping = 1;
3175 mutex_lock(&ptlrpc_all_services_mutex);
3176 list_del_init(&service->srv_list);
3177 mutex_unlock(&ptlrpc_all_services_mutex);
3179 ptlrpc_service_del_atimer(service);
3180 ptlrpc_stop_all_threads(service);
3182 ptlrpc_service_unlink_rqbd(service);
3183 ptlrpc_service_purge_all(service);
3184 ptlrpc_service_nrs_cleanup(service);
3186 ptlrpc_lprocfs_unregister_service(service);
3188 ptlrpc_service_free(service);
3192 EXPORT_SYMBOL(ptlrpc_unregister_service);
3195 * Returns 0 if the service is healthy.
3197 * Right now, it just checks to make sure that requests aren't languishing
3198 * in the queue. We'll use this health check to govern whether a node needs
3199 * to be shot, so it's intentionally non-aggressive. */
3200 static int ptlrpc_svcpt_health_check(struct ptlrpc_service_part *svcpt)
3202 struct ptlrpc_request *request = NULL;
3203 struct timeval right_now;
3206 do_gettimeofday(&right_now);
3208 spin_lock(&svcpt->scp_req_lock);
3209 /* How long has the next entry been waiting? */
3210 if (ptlrpc_server_high_pending(svcpt, true))
3211 request = ptlrpc_nrs_req_peek_nolock(svcpt, true);
3212 else if (ptlrpc_server_normal_pending(svcpt, true))
3213 request = ptlrpc_nrs_req_peek_nolock(svcpt, false);
3215 if (request == NULL) {
3216 spin_unlock(&svcpt->scp_req_lock);
3220 timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
3221 spin_unlock(&svcpt->scp_req_lock);
3223 if ((timediff / ONE_MILLION) >
3224 (AT_OFF ? obd_timeout * 3 / 2 : at_max)) {
3225 CERROR("%s: unhealthy - request has been waiting %lds\n",
3226 svcpt->scp_service->srv_name, timediff / ONE_MILLION);
3234 ptlrpc_service_health_check(struct ptlrpc_service *svc)
3236 struct ptlrpc_service_part *svcpt;
3242 ptlrpc_service_for_each_part(svcpt, i, svc) {
3243 int rc = ptlrpc_svcpt_health_check(svcpt);
3250 EXPORT_SYMBOL(ptlrpc_service_health_check);