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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, Whamcloud, Inc.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 #define DEBUG_SUBSYSTEM S_RPC
39 #include <liblustre.h>
41 #include <obd_support.h>
42 #include <obd_class.h>
43 #include <lustre_net.h>
44 #include <lu_object.h>
45 #include <lnet/types.h>
46 #include "ptlrpc_internal.h"
48 /* The following are visible and mutable through /sys/module/ptlrpc */
49 int test_req_buffer_pressure = 0;
50 CFS_MODULE_PARM(test_req_buffer_pressure, "i", int, 0444,
51 "set non-zero to put pressure on request buffer pools");
52 CFS_MODULE_PARM(at_min, "i", int, 0644,
53 "Adaptive timeout minimum (sec)");
54 CFS_MODULE_PARM(at_max, "i", int, 0644,
55 "Adaptive timeout maximum (sec)");
56 CFS_MODULE_PARM(at_history, "i", int, 0644,
57 "Adaptive timeouts remember the slowest event that took place "
58 "within this period (sec)");
59 CFS_MODULE_PARM(at_early_margin, "i", int, 0644,
60 "How soon before an RPC deadline to send an early reply");
61 CFS_MODULE_PARM(at_extra, "i", int, 0644,
62 "How much extra time to give with each early reply");
66 static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt);
67 static void ptlrpc_hpreq_fini(struct ptlrpc_request *req);
68 static void ptlrpc_at_remove_timed(struct ptlrpc_request *req);
70 static CFS_LIST_HEAD(ptlrpc_all_services);
71 cfs_spinlock_t ptlrpc_all_services_lock;
73 struct ptlrpc_request_buffer_desc *
74 ptlrpc_alloc_rqbd(struct ptlrpc_service_part *svcpt)
76 struct ptlrpc_service *svc = svcpt->scp_service;
77 struct ptlrpc_request_buffer_desc *rqbd;
79 OBD_CPT_ALLOC_PTR(rqbd, svc->srv_cptable, svcpt->scp_cpt);
83 rqbd->rqbd_svcpt = svcpt;
84 rqbd->rqbd_refcount = 0;
85 rqbd->rqbd_cbid.cbid_fn = request_in_callback;
86 rqbd->rqbd_cbid.cbid_arg = rqbd;
87 CFS_INIT_LIST_HEAD(&rqbd->rqbd_reqs);
88 OBD_CPT_ALLOC_LARGE(rqbd->rqbd_buffer, svc->srv_cptable,
89 svcpt->scp_cpt, svc->srv_buf_size);
90 if (rqbd->rqbd_buffer == NULL) {
95 cfs_spin_lock(&svcpt->scp_lock);
96 cfs_list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
97 svcpt->scp_nrqbds_total++;
98 cfs_spin_unlock(&svcpt->scp_lock);
104 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(cfs_list_empty(&rqbd->rqbd_reqs));
111 cfs_spin_lock(&svcpt->scp_lock);
112 cfs_list_del(&rqbd->rqbd_list);
113 svcpt->scp_nrqbds_total--;
114 cfs_spin_unlock(&svcpt->scp_lock);
116 OBD_FREE_LARGE(rqbd->rqbd_buffer, svcpt->scp_service->srv_buf_size);
121 ptlrpc_grow_req_bufs(struct ptlrpc_service_part *svcpt, int post)
123 struct ptlrpc_service *svc = svcpt->scp_service;
124 struct ptlrpc_request_buffer_desc *rqbd;
128 for (i = 0; i < svc->srv_nbuf_per_group; i++) {
129 /* NB: another thread might be doing this as well, we need to
130 * make sure that it wouldn't over-allocate, see LU-1212. */
131 if (svcpt->scp_nrqbds_posted >= svc->srv_nbuf_per_group)
134 rqbd = ptlrpc_alloc_rqbd(svcpt);
137 CERROR("%s: Can't allocate request buffer\n",
145 "%s: allocate %d new %d-byte reqbufs (%d/%d left), rc = %d\n",
146 svc->srv_name, i, svc->srv_buf_size, svcpt->scp_nrqbds_posted,
147 svcpt->scp_nrqbds_total, rc);
150 rc = ptlrpc_server_post_idle_rqbds(svcpt);
156 * Part of Rep-Ack logic.
157 * Puts a lock and its mode into reply state assotiated to request reply.
160 ptlrpc_save_lock(struct ptlrpc_request *req,
161 struct lustre_handle *lock, int mode, int no_ack)
163 struct ptlrpc_reply_state *rs = req->rq_reply_state;
167 LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
169 if (req->rq_export->exp_disconnected) {
170 ldlm_lock_decref(lock, mode);
172 idx = rs->rs_nlocks++;
173 rs->rs_locks[idx] = *lock;
174 rs->rs_modes[idx] = mode;
175 rs->rs_difficult = 1;
176 rs->rs_no_ack = !!no_ack;
179 EXPORT_SYMBOL(ptlrpc_save_lock);
183 struct ptlrpc_hr_partition;
185 struct ptlrpc_hr_thread {
186 int hrt_id; /* thread ID */
187 cfs_spinlock_t hrt_lock;
188 cfs_waitq_t hrt_waitq;
189 cfs_list_t hrt_queue; /* RS queue */
190 struct ptlrpc_hr_partition *hrt_partition;
193 struct ptlrpc_hr_partition {
194 /* # of started threads */
195 cfs_atomic_t hrp_nstarted;
196 /* # of stopped threads */
197 cfs_atomic_t hrp_nstopped;
198 /* cpu partition id */
200 /* round-robin rotor for choosing thread */
202 /* total number of threads on this partition */
205 struct ptlrpc_hr_thread *hrp_thrs;
208 #define HRT_RUNNING 0
209 #define HRT_STOPPING 1
211 struct ptlrpc_hr_service {
212 /* CPU partition table, it's just cfs_cpt_table for now */
213 struct cfs_cpt_table *hr_cpt_table;
214 /** controller sleep waitq */
215 cfs_waitq_t hr_waitq;
216 unsigned int hr_stopping;
217 /** roundrobin rotor for non-affinity service */
218 unsigned int hr_rotor;
220 struct ptlrpc_hr_partition **hr_partitions;
224 cfs_list_t rsb_replies;
225 unsigned int rsb_n_replies;
226 struct ptlrpc_service_part *rsb_svcpt;
229 /** reply handling service. */
230 static struct ptlrpc_hr_service ptlrpc_hr;
233 * maximum mumber of replies scheduled in one batch
235 #define MAX_SCHEDULED 256
238 * Initialize a reply batch.
242 static void rs_batch_init(struct rs_batch *b)
244 memset(b, 0, sizeof *b);
245 CFS_INIT_LIST_HEAD(&b->rsb_replies);
249 * Choose an hr thread to dispatch requests to.
251 static struct ptlrpc_hr_thread *
252 ptlrpc_hr_select(struct ptlrpc_service_part *svcpt)
254 struct ptlrpc_hr_partition *hrp;
257 if (svcpt->scp_cpt >= 0 &&
258 svcpt->scp_service->srv_cptable == ptlrpc_hr.hr_cpt_table) {
259 /* directly match partition */
260 hrp = ptlrpc_hr.hr_partitions[svcpt->scp_cpt];
263 rotor = ptlrpc_hr.hr_rotor++;
264 rotor %= cfs_cpt_number(ptlrpc_hr.hr_cpt_table);
266 hrp = ptlrpc_hr.hr_partitions[rotor];
269 rotor = hrp->hrp_rotor++;
270 return &hrp->hrp_thrs[rotor % hrp->hrp_nthrs];
274 * Dispatch all replies accumulated in the batch to one from
275 * dedicated reply handling threads.
279 static void rs_batch_dispatch(struct rs_batch *b)
281 if (b->rsb_n_replies != 0) {
282 struct ptlrpc_hr_thread *hrt;
284 hrt = ptlrpc_hr_select(b->rsb_svcpt);
286 cfs_spin_lock(&hrt->hrt_lock);
287 cfs_list_splice_init(&b->rsb_replies, &hrt->hrt_queue);
288 cfs_spin_unlock(&hrt->hrt_lock);
290 cfs_waitq_signal(&hrt->hrt_waitq);
291 b->rsb_n_replies = 0;
296 * Add a reply to a batch.
297 * Add one reply object to a batch, schedule batched replies if overload.
302 static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs)
304 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
306 if (svcpt != b->rsb_svcpt || b->rsb_n_replies >= MAX_SCHEDULED) {
307 if (b->rsb_svcpt != NULL) {
308 rs_batch_dispatch(b);
309 cfs_spin_unlock(&b->rsb_svcpt->scp_rep_lock);
311 cfs_spin_lock(&svcpt->scp_rep_lock);
312 b->rsb_svcpt = svcpt;
314 cfs_spin_lock(&rs->rs_lock);
315 rs->rs_scheduled_ever = 1;
316 if (rs->rs_scheduled == 0) {
317 cfs_list_move(&rs->rs_list, &b->rsb_replies);
318 rs->rs_scheduled = 1;
321 rs->rs_committed = 1;
322 cfs_spin_unlock(&rs->rs_lock);
326 * Reply batch finalization.
327 * Dispatch remaining replies from the batch
328 * and release remaining spinlock.
332 static void rs_batch_fini(struct rs_batch *b)
334 if (b->rsb_svcpt != NULL) {
335 rs_batch_dispatch(b);
336 cfs_spin_unlock(&b->rsb_svcpt->scp_rep_lock);
340 #define DECLARE_RS_BATCH(b) struct rs_batch b
342 #else /* __KERNEL__ */
344 #define rs_batch_init(b) do{}while(0)
345 #define rs_batch_fini(b) do{}while(0)
346 #define rs_batch_add(b, r) ptlrpc_schedule_difficult_reply(r)
347 #define DECLARE_RS_BATCH(b)
349 #endif /* __KERNEL__ */
352 * Put reply state into a queue for processing because we received
353 * ACK from the client
355 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
358 struct ptlrpc_hr_thread *hrt;
361 LASSERT(cfs_list_empty(&rs->rs_list));
363 hrt = ptlrpc_hr_select(rs->rs_svcpt);
365 cfs_spin_lock(&hrt->hrt_lock);
366 cfs_list_add_tail(&rs->rs_list, &hrt->hrt_queue);
367 cfs_spin_unlock(&hrt->hrt_lock);
369 cfs_waitq_signal(&hrt->hrt_waitq);
372 cfs_list_add_tail(&rs->rs_list, &rs->rs_svcpt->scp_rep_queue);
377 ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs)
381 LASSERT_SPIN_LOCKED(&rs->rs_svcpt->scp_rep_lock);
382 LASSERT_SPIN_LOCKED(&rs->rs_lock);
383 LASSERT (rs->rs_difficult);
384 rs->rs_scheduled_ever = 1; /* flag any notification attempt */
386 if (rs->rs_scheduled) { /* being set up or already notified */
391 rs->rs_scheduled = 1;
392 cfs_list_del_init(&rs->rs_list);
393 ptlrpc_dispatch_difficult_reply(rs);
396 EXPORT_SYMBOL(ptlrpc_schedule_difficult_reply);
398 void ptlrpc_commit_replies(struct obd_export *exp)
400 struct ptlrpc_reply_state *rs, *nxt;
401 DECLARE_RS_BATCH(batch);
404 rs_batch_init(&batch);
405 /* Find any replies that have been committed and get their service
406 * to attend to complete them. */
408 /* CAVEAT EMPTOR: spinlock ordering!!! */
409 cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
410 cfs_list_for_each_entry_safe(rs, nxt, &exp->exp_uncommitted_replies,
412 LASSERT (rs->rs_difficult);
413 /* VBR: per-export last_committed */
414 LASSERT(rs->rs_export);
415 if (rs->rs_transno <= exp->exp_last_committed) {
416 cfs_list_del_init(&rs->rs_obd_list);
417 rs_batch_add(&batch, rs);
420 cfs_spin_unlock(&exp->exp_uncommitted_replies_lock);
421 rs_batch_fini(&batch);
424 EXPORT_SYMBOL(ptlrpc_commit_replies);
427 ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt)
429 struct ptlrpc_request_buffer_desc *rqbd;
434 cfs_spin_lock(&svcpt->scp_lock);
436 if (cfs_list_empty(&svcpt->scp_rqbd_idle)) {
437 cfs_spin_unlock(&svcpt->scp_lock);
441 rqbd = cfs_list_entry(svcpt->scp_rqbd_idle.next,
442 struct ptlrpc_request_buffer_desc,
444 cfs_list_del(&rqbd->rqbd_list);
446 /* assume we will post successfully */
447 svcpt->scp_nrqbds_posted++;
448 cfs_list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_posted);
450 cfs_spin_unlock(&svcpt->scp_lock);
452 rc = ptlrpc_register_rqbd(rqbd);
459 cfs_spin_lock(&svcpt->scp_lock);
461 svcpt->scp_nrqbds_posted--;
462 cfs_list_del(&rqbd->rqbd_list);
463 cfs_list_add_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
465 /* Don't complain if no request buffers are posted right now; LNET
466 * won't drop requests because we set the portal lazy! */
468 cfs_spin_unlock(&svcpt->scp_lock);
473 static void ptlrpc_at_timer(unsigned long castmeharder)
475 struct ptlrpc_service_part *svcpt;
477 svcpt = (struct ptlrpc_service_part *)castmeharder;
479 svcpt->scp_at_check = 1;
480 svcpt->scp_at_checktime = cfs_time_current();
481 cfs_waitq_signal(&svcpt->scp_waitq);
485 ptlrpc_server_nthreads_check(struct ptlrpc_service *svc,
486 struct ptlrpc_service_conf *conf)
489 struct ptlrpc_service_thr_conf *tc = &conf->psc_thr;
496 * Common code for estimating & validating threads number.
497 * CPT affinity service could have percpt thread-pool instead
498 * of a global thread-pool, which means user might not always
499 * get the threads number they give it in conf::tc_nthrs_user
500 * even they did set. It's because we need to validate threads
501 * number for each CPT to guarantee each pool will have enough
502 * threads to keep the service healthy.
504 init = PTLRPC_NTHRS_INIT + (svc->srv_ops.so_hpreq_handler != NULL);
505 init = max_t(int, init, tc->tc_nthrs_init);
507 /* NB: please see comments in lustre_lnet.h for definition
508 * details of these members */
509 LASSERT(tc->tc_nthrs_max != 0);
511 if (tc->tc_nthrs_user != 0) {
512 /* In case there is a reason to test a service with many
513 * threads, we give a less strict check here, it can
514 * be up to 8 * nthrs_max */
515 total = min(tc->tc_nthrs_max * 8, tc->tc_nthrs_user);
516 nthrs = total / svc->srv_ncpts;
517 init = max(init, nthrs);
521 total = tc->tc_nthrs_max;
522 if (tc->tc_nthrs_base == 0) {
523 /* don't care about base threads number per partition,
524 * this is most for non-affinity service */
525 nthrs = total / svc->srv_ncpts;
529 nthrs = tc->tc_nthrs_base;
530 if (svc->srv_ncpts == 1) {
533 /* NB: Increase the base number if it's single partition
534 * and total number of cores/HTs is larger or equal to 4.
535 * result will always < 2 * nthrs_base */
536 weight = cfs_cpt_weight(svc->srv_cptable, CFS_CPT_ANY);
537 for (i = 1; (weight >> (i + 1)) != 0 && /* >= 4 cores/HTs */
538 (tc->tc_nthrs_base >> i) != 0; i++)
539 nthrs += tc->tc_nthrs_base >> i;
542 if (tc->tc_thr_factor != 0) {
543 int factor = tc->tc_thr_factor;
547 * User wants to increase number of threads with for
548 * each CPU core/HT, most likely the factor is larger then
549 * one thread/core because service threads are supposed to
550 * be blocked by lock or wait for IO.
553 * Amdahl's law says that adding processors wouldn't give
554 * a linear increasing of parallelism, so it's nonsense to
555 * have too many threads no matter how many cores/HTs
558 if (cfs_cpu_ht_nsiblings(0) > 1) { /* weight is # of HTs */
559 /* depress thread factor for hyper-thread */
560 factor = factor - (factor >> 1) + (factor >> 3);
563 weight = cfs_cpt_weight(svc->srv_cptable, 0);
566 for (; factor > 0 && weight > 0; factor--, weight -= fade)
567 nthrs += min(weight, fade) * factor;
570 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
571 nthrs = max(tc->tc_nthrs_base,
572 tc->tc_nthrs_max / svc->srv_ncpts);
575 nthrs = max(nthrs, tc->tc_nthrs_init);
576 svc->srv_nthrs_cpt_limit = nthrs;
577 svc->srv_nthrs_cpt_init = init;
579 if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
580 LCONSOLE_WARN("%s: This service may have more threads (%d) "
581 "than the given soft limit (%d)\n",
582 svc->srv_name, nthrs * svc->srv_ncpts,
589 * Initialize percpt data for a service
592 ptlrpc_service_part_init(struct ptlrpc_service *svc,
593 struct ptlrpc_service_part *svcpt, int cpt)
595 struct ptlrpc_at_array *array;
600 svcpt->scp_cpt = cpt;
601 CFS_INIT_LIST_HEAD(&svcpt->scp_threads);
603 /* rqbd and incoming request queue */
604 cfs_spin_lock_init(&svcpt->scp_lock);
605 CFS_INIT_LIST_HEAD(&svcpt->scp_rqbd_idle);
606 CFS_INIT_LIST_HEAD(&svcpt->scp_rqbd_posted);
607 CFS_INIT_LIST_HEAD(&svcpt->scp_req_incoming);
608 cfs_waitq_init(&svcpt->scp_waitq);
609 /* history request & rqbd list */
610 CFS_INIT_LIST_HEAD(&svcpt->scp_hist_reqs);
611 CFS_INIT_LIST_HEAD(&svcpt->scp_hist_rqbds);
613 /* acitve requests and hp requests */
614 cfs_spin_lock_init(&svcpt->scp_req_lock);
615 CFS_INIT_LIST_HEAD(&svcpt->scp_req_pending);
616 CFS_INIT_LIST_HEAD(&svcpt->scp_hreq_pending);
619 cfs_spin_lock_init(&svcpt->scp_rep_lock);
620 CFS_INIT_LIST_HEAD(&svcpt->scp_rep_active);
622 CFS_INIT_LIST_HEAD(&svcpt->scp_rep_queue);
624 CFS_INIT_LIST_HEAD(&svcpt->scp_rep_idle);
625 cfs_waitq_init(&svcpt->scp_rep_waitq);
626 cfs_atomic_set(&svcpt->scp_nreps_difficult, 0);
628 /* adaptive timeout */
629 cfs_spin_lock_init(&svcpt->scp_at_lock);
630 array = &svcpt->scp_at_array;
632 size = at_est2timeout(at_max);
633 array->paa_size = size;
634 array->paa_count = 0;
635 array->paa_deadline = -1;
637 /* allocate memory for scp_at_array (ptlrpc_at_array) */
638 OBD_CPT_ALLOC(array->paa_reqs_array,
639 svc->srv_cptable, cpt, sizeof(cfs_list_t) * size);
640 if (array->paa_reqs_array == NULL)
643 for (index = 0; index < size; index++)
644 CFS_INIT_LIST_HEAD(&array->paa_reqs_array[index]);
646 OBD_CPT_ALLOC(array->paa_reqs_count,
647 svc->srv_cptable, cpt, sizeof(__u32) * size);
648 if (array->paa_reqs_count == NULL)
651 cfs_timer_init(&svcpt->scp_at_timer, ptlrpc_at_timer, 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(cfs_list_t) * 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 cfs_proc_dir_entry_t *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);
732 RETURN(ERR_PTR(rc < 0 ? rc : -EINVAL));
738 OBD_ALLOC(service, offsetof(struct ptlrpc_service, srv_parts[ncpts]));
739 if (service == NULL) {
741 OBD_FREE(cpts, sizeof(*cpts) * ncpts);
742 RETURN(ERR_PTR(-ENOMEM));
745 service->srv_cptable = cptable;
746 service->srv_cpts = cpts;
747 service->srv_ncpts = ncpts;
749 service->srv_cpt_bits = 0; /* it's zero already, easy to read... */
750 while ((1 << service->srv_cpt_bits) < cfs_cpt_number(cptable))
751 service->srv_cpt_bits++;
754 cfs_spin_lock_init(&service->srv_lock);
755 service->srv_name = conf->psc_name;
756 service->srv_watchdog_factor = conf->psc_watchdog_factor;
757 CFS_INIT_LIST_HEAD(&service->srv_list); /* for safty of cleanup */
759 /* buffer configuration */
760 service->srv_nbuf_per_group = test_req_buffer_pressure ? 1 :
761 max(conf->psc_buf.bc_nbufs /
762 service->srv_ncpts, 1U);
763 service->srv_max_req_size = conf->psc_buf.bc_req_max_size +
765 service->srv_buf_size = conf->psc_buf.bc_buf_size;
766 service->srv_rep_portal = conf->psc_buf.bc_rep_portal;
767 service->srv_req_portal = conf->psc_buf.bc_req_portal;
769 /* Increase max reply size to next power of two */
770 service->srv_max_reply_size = 1;
771 while (service->srv_max_reply_size <
772 conf->psc_buf.bc_rep_max_size + SPTLRPC_MAX_PAYLOAD)
773 service->srv_max_reply_size <<= 1;
775 service->srv_thread_name = conf->psc_thr.tc_thr_name;
776 service->srv_ctx_tags = conf->psc_thr.tc_ctx_tags;
777 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
778 service->srv_ops = conf->psc_ops;
780 for (i = 0; i < ncpts; i++) {
781 if (!conf->psc_thr.tc_cpu_affinity)
784 cpt = cpts != NULL ? cpts[i] : i;
786 OBD_CPT_ALLOC(svcpt, cptable, cpt, sizeof(*svcpt));
788 GOTO(failed, rc = -ENOMEM);
790 service->srv_parts[i] = svcpt;
791 rc = ptlrpc_service_part_init(service, svcpt, cpt);
796 ptlrpc_server_nthreads_check(service, conf);
798 rc = LNetSetLazyPortal(service->srv_req_portal);
801 cfs_spin_lock (&ptlrpc_all_services_lock);
802 cfs_list_add (&service->srv_list, &ptlrpc_all_services);
803 cfs_spin_unlock (&ptlrpc_all_services_lock);
805 if (proc_entry != NULL)
806 ptlrpc_lprocfs_register_service(proc_entry, service);
808 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
809 service->srv_name, service->srv_req_portal);
812 rc = ptlrpc_start_threads(service);
814 CERROR("Failed to start threads for service %s: %d\n",
815 service->srv_name, rc);
822 ptlrpc_unregister_service(service);
825 EXPORT_SYMBOL(ptlrpc_register_service);
828 * to actually free the request, must be called without holding svc_lock.
829 * note it's caller's responsibility to unlink req->rq_list.
831 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
833 LASSERT(cfs_atomic_read(&req->rq_refcount) == 0);
834 LASSERT(cfs_list_empty(&req->rq_timed_list));
836 /* DEBUG_REQ() assumes the reply state of a request with a valid
837 * ref will not be destroyed until that reference is dropped. */
838 ptlrpc_req_drop_rs(req);
840 sptlrpc_svc_ctx_decref(req);
842 if (req != &req->rq_rqbd->rqbd_req) {
843 /* NB request buffers use an embedded
844 * req if the incoming req unlinked the
845 * MD; this isn't one of them! */
846 OBD_FREE(req, sizeof(*req));
851 * drop a reference count of the request. if it reaches 0, we either
852 * put it into history list, or free it immediately.
854 void ptlrpc_server_drop_request(struct ptlrpc_request *req)
856 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
857 struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
858 struct ptlrpc_service *svc = svcpt->scp_service;
863 if (!cfs_atomic_dec_and_test(&req->rq_refcount))
866 if (req->rq_at_linked) {
867 cfs_spin_lock(&svcpt->scp_at_lock);
868 /* recheck with lock, in case it's unlinked by
869 * ptlrpc_at_check_timed() */
870 if (likely(req->rq_at_linked))
871 ptlrpc_at_remove_timed(req);
872 cfs_spin_unlock(&svcpt->scp_at_lock);
875 LASSERT(cfs_list_empty(&req->rq_timed_list));
877 /* finalize request */
878 if (req->rq_export) {
879 class_export_put(req->rq_export);
880 req->rq_export = NULL;
883 cfs_spin_lock(&svcpt->scp_lock);
885 cfs_list_add(&req->rq_list, &rqbd->rqbd_reqs);
887 refcount = --(rqbd->rqbd_refcount);
889 /* request buffer is now idle: add to history */
890 cfs_list_del(&rqbd->rqbd_list);
892 cfs_list_add_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds);
893 svcpt->scp_hist_nrqbds++;
895 /* cull some history?
896 * I expect only about 1 or 2 rqbds need to be recycled here */
897 while (svcpt->scp_hist_nrqbds > svc->srv_hist_nrqbds_cpt_max) {
898 rqbd = cfs_list_entry(svcpt->scp_hist_rqbds.next,
899 struct ptlrpc_request_buffer_desc,
902 cfs_list_del(&rqbd->rqbd_list);
903 svcpt->scp_hist_nrqbds--;
905 /* remove rqbd's reqs from svc's req history while
906 * I've got the service lock */
907 cfs_list_for_each(tmp, &rqbd->rqbd_reqs) {
908 req = cfs_list_entry(tmp, struct ptlrpc_request,
910 /* Track the highest culled req seq */
911 if (req->rq_history_seq >
912 svcpt->scp_hist_seq_culled) {
913 svcpt->scp_hist_seq_culled =
916 cfs_list_del(&req->rq_history_list);
919 cfs_spin_unlock(&svcpt->scp_lock);
921 cfs_list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
922 req = cfs_list_entry(rqbd->rqbd_reqs.next,
923 struct ptlrpc_request,
925 cfs_list_del(&req->rq_list);
926 ptlrpc_server_free_request(req);
929 cfs_spin_lock(&svcpt->scp_lock);
931 * now all reqs including the embedded req has been
932 * disposed, schedule request buffer for re-use.
934 LASSERT(cfs_atomic_read(&rqbd->rqbd_req.rq_refcount) ==
936 cfs_list_add_tail(&rqbd->rqbd_list,
937 &svcpt->scp_rqbd_idle);
940 cfs_spin_unlock(&svcpt->scp_lock);
941 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
942 /* If we are low on memory, we are not interested in history */
943 cfs_list_del(&req->rq_list);
944 cfs_list_del_init(&req->rq_history_list);
946 /* Track the highest culled req seq */
947 if (req->rq_history_seq > svcpt->scp_hist_seq_culled)
948 svcpt->scp_hist_seq_culled = req->rq_history_seq;
950 cfs_spin_unlock(&svcpt->scp_lock);
952 ptlrpc_server_free_request(req);
954 cfs_spin_unlock(&svcpt->scp_lock);
959 * to finish a request: stop sending more early replies, and release
960 * the request. should be called after we finished handling the request.
962 static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt,
963 struct ptlrpc_request *req)
965 ptlrpc_hpreq_fini(req);
967 cfs_spin_lock(&svcpt->scp_req_lock);
968 svcpt->scp_nreqs_active--;
970 svcpt->scp_nhreqs_active--;
971 cfs_spin_unlock(&svcpt->scp_req_lock);
973 ptlrpc_server_drop_request(req);
977 * This function makes sure dead exports are evicted in a timely manner.
978 * This function is only called when some export receives a message (i.e.,
979 * the network is up.)
981 static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
983 struct obd_export *oldest_exp;
984 time_t oldest_time, new_time;
990 /* Compensate for slow machines, etc, by faking our request time
991 into the future. Although this can break the strict time-ordering
992 of the list, we can be really lazy here - we don't have to evict
993 at the exact right moment. Eventually, all silent exports
994 will make it to the top of the list. */
996 /* Do not pay attention on 1sec or smaller renewals. */
997 new_time = cfs_time_current_sec() + extra_delay;
998 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
1001 exp->exp_last_request_time = new_time;
1002 CDEBUG(D_HA, "updating export %s at "CFS_TIME_T" exp %p\n",
1003 exp->exp_client_uuid.uuid,
1004 exp->exp_last_request_time, exp);
1006 /* exports may get disconnected from the chain even though the
1007 export has references, so we must keep the spin lock while
1008 manipulating the lists */
1009 cfs_spin_lock(&exp->exp_obd->obd_dev_lock);
1011 if (cfs_list_empty(&exp->exp_obd_chain_timed)) {
1012 /* this one is not timed */
1013 cfs_spin_unlock(&exp->exp_obd->obd_dev_lock);
1017 cfs_list_move_tail(&exp->exp_obd_chain_timed,
1018 &exp->exp_obd->obd_exports_timed);
1020 oldest_exp = cfs_list_entry(exp->exp_obd->obd_exports_timed.next,
1021 struct obd_export, exp_obd_chain_timed);
1022 oldest_time = oldest_exp->exp_last_request_time;
1023 cfs_spin_unlock(&exp->exp_obd->obd_dev_lock);
1025 if (exp->exp_obd->obd_recovering) {
1026 /* be nice to everyone during recovery */
1031 /* Note - racing to start/reset the obd_eviction timer is safe */
1032 if (exp->exp_obd->obd_eviction_timer == 0) {
1033 /* Check if the oldest entry is expired. */
1034 if (cfs_time_current_sec() > (oldest_time + PING_EVICT_TIMEOUT +
1036 /* We need a second timer, in case the net was down and
1037 * it just came back. Since the pinger may skip every
1038 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
1039 * we better wait for 3. */
1040 exp->exp_obd->obd_eviction_timer =
1041 cfs_time_current_sec() + 3 * PING_INTERVAL;
1042 CDEBUG(D_HA, "%s: Think about evicting %s from "CFS_TIME_T"\n",
1043 exp->exp_obd->obd_name,
1044 obd_export_nid2str(oldest_exp), oldest_time);
1047 if (cfs_time_current_sec() >
1048 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
1049 /* The evictor won't evict anyone who we've heard from
1050 * recently, so we don't have to check before we start
1052 if (!ping_evictor_wake(exp))
1053 exp->exp_obd->obd_eviction_timer = 0;
1061 * Sanity check request \a req.
1062 * Return 0 if all is ok, error code otherwise.
1064 static int ptlrpc_check_req(struct ptlrpc_request *req)
1068 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
1069 req->rq_export->exp_conn_cnt)) {
1070 DEBUG_REQ(D_ERROR, req,
1071 "DROPPING req from old connection %d < %d",
1072 lustre_msg_get_conn_cnt(req->rq_reqmsg),
1073 req->rq_export->exp_conn_cnt);
1076 if (unlikely(req->rq_export->exp_obd &&
1077 req->rq_export->exp_obd->obd_fail)) {
1078 /* Failing over, don't handle any more reqs, send
1079 error response instead. */
1080 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
1081 req, req->rq_export->exp_obd->obd_name);
1083 } else if (lustre_msg_get_flags(req->rq_reqmsg) &
1084 (MSG_REPLAY | MSG_REQ_REPLAY_DONE) &&
1085 !(req->rq_export->exp_obd->obd_recovering)) {
1086 DEBUG_REQ(D_ERROR, req,
1087 "Invalid replay without recovery");
1088 class_fail_export(req->rq_export);
1090 } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 &&
1091 !(req->rq_export->exp_obd->obd_recovering)) {
1092 DEBUG_REQ(D_ERROR, req, "Invalid req with transno "
1093 LPU64" without recovery",
1094 lustre_msg_get_transno(req->rq_reqmsg));
1095 class_fail_export(req->rq_export);
1099 if (unlikely(rc < 0)) {
1100 req->rq_status = rc;
1106 static void ptlrpc_at_set_timer(struct ptlrpc_service_part *svcpt)
1108 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1111 if (array->paa_count == 0) {
1112 cfs_timer_disarm(&svcpt->scp_at_timer);
1116 /* Set timer for closest deadline */
1117 next = (__s32)(array->paa_deadline - cfs_time_current_sec() -
1120 ptlrpc_at_timer((unsigned long)svcpt);
1122 cfs_timer_arm(&svcpt->scp_at_timer, cfs_time_shift(next));
1123 CDEBUG(D_INFO, "armed %s at %+ds\n",
1124 svcpt->scp_service->srv_name, next);
1128 /* Add rpc to early reply check list */
1129 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
1131 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1132 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1133 struct ptlrpc_request *rq = NULL;
1139 if (req->rq_no_reply)
1142 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
1145 cfs_spin_lock(&svcpt->scp_at_lock);
1146 LASSERT(cfs_list_empty(&req->rq_timed_list));
1148 index = (unsigned long)req->rq_deadline % array->paa_size;
1149 if (array->paa_reqs_count[index] > 0) {
1150 /* latest rpcs will have the latest deadlines in the list,
1151 * so search backward. */
1152 cfs_list_for_each_entry_reverse(rq,
1153 &array->paa_reqs_array[index],
1155 if (req->rq_deadline >= rq->rq_deadline) {
1156 cfs_list_add(&req->rq_timed_list,
1157 &rq->rq_timed_list);
1163 /* Add the request at the head of the list */
1164 if (cfs_list_empty(&req->rq_timed_list))
1165 cfs_list_add(&req->rq_timed_list,
1166 &array->paa_reqs_array[index]);
1168 cfs_spin_lock(&req->rq_lock);
1169 req->rq_at_linked = 1;
1170 cfs_spin_unlock(&req->rq_lock);
1171 req->rq_at_index = index;
1172 array->paa_reqs_count[index]++;
1174 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
1175 array->paa_deadline = req->rq_deadline;
1176 ptlrpc_at_set_timer(svcpt);
1178 cfs_spin_unlock(&svcpt->scp_at_lock);
1184 ptlrpc_at_remove_timed(struct ptlrpc_request *req)
1186 struct ptlrpc_at_array *array;
1188 array = &req->rq_rqbd->rqbd_svcpt->scp_at_array;
1190 /* NB: must call with hold svcpt::scp_at_lock */
1191 LASSERT(!cfs_list_empty(&req->rq_timed_list));
1192 cfs_list_del_init(&req->rq_timed_list);
1194 cfs_spin_lock(&req->rq_lock);
1195 req->rq_at_linked = 0;
1196 cfs_spin_unlock(&req->rq_lock);
1198 array->paa_reqs_count[req->rq_at_index]--;
1202 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
1204 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1205 struct ptlrpc_request *reqcopy;
1206 struct lustre_msg *reqmsg;
1207 cfs_duration_t olddl = req->rq_deadline - cfs_time_current_sec();
1212 /* deadline is when the client expects us to reply, margin is the
1213 difference between clients' and servers' expectations */
1214 DEBUG_REQ(D_ADAPTTO, req,
1215 "%ssending early reply (deadline %+lds, margin %+lds) for "
1216 "%d+%d", AT_OFF ? "AT off - not " : "",
1217 olddl, olddl - at_get(&svcpt->scp_at_estimate),
1218 at_get(&svcpt->scp_at_estimate), at_extra);
1224 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), "
1225 "not sending early reply. Consider increasing "
1226 "at_early_margin (%d)?", olddl, at_early_margin);
1228 /* Return an error so we're not re-added to the timed list. */
1232 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
1233 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
1234 "but no AT support");
1238 if (req->rq_export &&
1239 lustre_msg_get_flags(req->rq_reqmsg) &
1240 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1241 /* During recovery, we don't want to send too many early
1242 * replies, but on the other hand we want to make sure the
1243 * client has enough time to resend if the rpc is lost. So
1244 * during the recovery period send at least 4 early replies,
1245 * spacing them every at_extra if we can. at_estimate should
1246 * always equal this fixed value during recovery. */
1247 at_measured(&svcpt->scp_at_estimate, min(at_extra,
1248 req->rq_export->exp_obd->obd_recovery_timeout / 4));
1250 /* Fake our processing time into the future to ask the clients
1251 * for some extra amount of time */
1252 at_measured(&svcpt->scp_at_estimate, at_extra +
1253 cfs_time_current_sec() -
1254 req->rq_arrival_time.tv_sec);
1256 /* Check to see if we've actually increased the deadline -
1257 * we may be past adaptive_max */
1258 if (req->rq_deadline >= req->rq_arrival_time.tv_sec +
1259 at_get(&svcpt->scp_at_estimate)) {
1260 DEBUG_REQ(D_WARNING, req, "Couldn't add any time "
1261 "(%ld/%ld), not sending early reply\n",
1262 olddl, req->rq_arrival_time.tv_sec +
1263 at_get(&svcpt->scp_at_estimate) -
1264 cfs_time_current_sec());
1268 newdl = cfs_time_current_sec() + at_get(&svcpt->scp_at_estimate);
1270 OBD_ALLOC(reqcopy, sizeof *reqcopy);
1271 if (reqcopy == NULL)
1273 OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1275 OBD_FREE(reqcopy, sizeof *reqcopy);
1280 reqcopy->rq_reply_state = NULL;
1281 reqcopy->rq_rep_swab_mask = 0;
1282 reqcopy->rq_pack_bulk = 0;
1283 reqcopy->rq_pack_udesc = 0;
1284 reqcopy->rq_packed_final = 0;
1285 sptlrpc_svc_ctx_addref(reqcopy);
1286 /* We only need the reqmsg for the magic */
1287 reqcopy->rq_reqmsg = reqmsg;
1288 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1290 LASSERT(cfs_atomic_read(&req->rq_refcount));
1291 /** if it is last refcount then early reply isn't needed */
1292 if (cfs_atomic_read(&req->rq_refcount) == 1) {
1293 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
1294 "abort sending early reply\n");
1295 GOTO(out, rc = -EINVAL);
1298 /* Connection ref */
1299 reqcopy->rq_export = class_conn2export(
1300 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1301 if (reqcopy->rq_export == NULL)
1302 GOTO(out, rc = -ENODEV);
1305 class_export_rpc_get(reqcopy->rq_export);
1306 if (reqcopy->rq_export->exp_obd &&
1307 reqcopy->rq_export->exp_obd->obd_fail)
1308 GOTO(out_put, rc = -ENODEV);
1310 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1314 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1317 /* Adjust our own deadline to what we told the client */
1318 req->rq_deadline = newdl;
1319 req->rq_early_count++; /* number sent, server side */
1321 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1324 /* Free the (early) reply state from lustre_pack_reply.
1325 (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
1326 ptlrpc_req_drop_rs(reqcopy);
1329 class_export_rpc_put(reqcopy->rq_export);
1330 class_export_put(reqcopy->rq_export);
1332 sptlrpc_svc_ctx_decref(reqcopy);
1333 OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1334 OBD_FREE(reqcopy, sizeof *reqcopy);
1338 /* Send early replies to everybody expiring within at_early_margin
1339 asking for at_extra time */
1340 static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
1342 struct ptlrpc_at_array *array = &svcpt->scp_at_array;
1343 struct ptlrpc_request *rq, *n;
1344 cfs_list_t work_list;
1347 time_t now = cfs_time_current_sec();
1348 cfs_duration_t delay;
1349 int first, counter = 0;
1352 cfs_spin_lock(&svcpt->scp_at_lock);
1353 if (svcpt->scp_at_check == 0) {
1354 cfs_spin_unlock(&svcpt->scp_at_lock);
1357 delay = cfs_time_sub(cfs_time_current(), svcpt->scp_at_checktime);
1358 svcpt->scp_at_check = 0;
1360 if (array->paa_count == 0) {
1361 cfs_spin_unlock(&svcpt->scp_at_lock);
1365 /* The timer went off, but maybe the nearest rpc already completed. */
1366 first = array->paa_deadline - now;
1367 if (first > at_early_margin) {
1368 /* We've still got plenty of time. Reset the timer. */
1369 ptlrpc_at_set_timer(svcpt);
1370 cfs_spin_unlock(&svcpt->scp_at_lock);
1374 /* We're close to a timeout, and we don't know how much longer the
1375 server will take. Send early replies to everyone expiring soon. */
1376 CFS_INIT_LIST_HEAD(&work_list);
1378 index = (unsigned long)array->paa_deadline % array->paa_size;
1379 count = array->paa_count;
1381 count -= array->paa_reqs_count[index];
1382 cfs_list_for_each_entry_safe(rq, n,
1383 &array->paa_reqs_array[index],
1385 if (rq->rq_deadline > now + at_early_margin) {
1386 /* update the earliest deadline */
1387 if (deadline == -1 ||
1388 rq->rq_deadline < deadline)
1389 deadline = rq->rq_deadline;
1393 ptlrpc_at_remove_timed(rq);
1395 * ptlrpc_server_drop_request() may drop
1396 * refcount to 0 already. Let's check this and
1397 * don't add entry to work_list
1399 if (likely(cfs_atomic_inc_not_zero(&rq->rq_refcount)))
1400 cfs_list_add(&rq->rq_timed_list, &work_list);
1404 if (++index >= array->paa_size)
1407 array->paa_deadline = deadline;
1408 /* we have a new earliest deadline, restart the timer */
1409 ptlrpc_at_set_timer(svcpt);
1411 cfs_spin_unlock(&svcpt->scp_at_lock);
1413 CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
1414 "replies\n", first, at_extra, counter);
1416 /* We're already past request deadlines before we even get a
1417 chance to send early replies */
1418 LCONSOLE_WARN("%s: This server is not able to keep up with "
1419 "request traffic (cpu-bound).\n",
1420 svcpt->scp_service->srv_name);
1421 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, "
1422 "delay="CFS_DURATION_T"(jiff)\n",
1423 counter, svcpt->scp_nreqs_incoming,
1424 svcpt->scp_nreqs_active,
1425 at_get(&svcpt->scp_at_estimate), delay);
1428 /* we took additional refcount so entries can't be deleted from list, no
1429 * locking is needed */
1430 while (!cfs_list_empty(&work_list)) {
1431 rq = cfs_list_entry(work_list.next, struct ptlrpc_request,
1433 cfs_list_del_init(&rq->rq_timed_list);
1435 if (ptlrpc_at_send_early_reply(rq) == 0)
1436 ptlrpc_at_add_timed(rq);
1438 ptlrpc_server_drop_request(rq);
1441 RETURN(1); /* return "did_something" for liblustre */
1445 * Put the request to the export list if the request may become
1446 * a high priority one.
1448 static int ptlrpc_hpreq_init(struct ptlrpc_service *svc,
1449 struct ptlrpc_request *req)
1454 if (svc->srv_ops.so_hpreq_handler) {
1455 rc = svc->srv_ops.so_hpreq_handler(req);
1459 if (req->rq_export && req->rq_ops) {
1460 /* Perform request specific check. We should do this check
1461 * before the request is added into exp_hp_rpcs list otherwise
1462 * it may hit swab race at LU-1044. */
1463 if (req->rq_ops->hpreq_check)
1464 rc = req->rq_ops->hpreq_check(req);
1466 cfs_spin_lock_bh(&req->rq_export->exp_rpc_lock);
1467 cfs_list_add(&req->rq_exp_list,
1468 &req->rq_export->exp_hp_rpcs);
1469 cfs_spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1475 /** Remove the request from the export list. */
1476 static void ptlrpc_hpreq_fini(struct ptlrpc_request *req)
1479 if (req->rq_export && req->rq_ops) {
1480 /* refresh lock timeout again so that client has more
1481 * room to send lock cancel RPC. */
1482 if (req->rq_ops->hpreq_fini)
1483 req->rq_ops->hpreq_fini(req);
1485 cfs_spin_lock_bh(&req->rq_export->exp_rpc_lock);
1486 cfs_list_del_init(&req->rq_exp_list);
1487 cfs_spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1492 static int ptlrpc_hpreq_check(struct ptlrpc_request *req)
1497 static struct ptlrpc_hpreq_ops ptlrpc_hpreq_common = {
1498 .hpreq_lock_match = NULL,
1499 .hpreq_check = ptlrpc_hpreq_check,
1503 /* Hi-Priority RPC check by RPC operation code. */
1504 int ptlrpc_hpreq_handler(struct ptlrpc_request *req)
1506 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1508 /* Check for export to let only reconnects for not yet evicted
1509 * export to become a HP rpc. */
1510 if ((req->rq_export != NULL) &&
1511 (opc == OBD_PING || opc == MDS_CONNECT || opc == OST_CONNECT))
1512 req->rq_ops = &ptlrpc_hpreq_common;
1516 EXPORT_SYMBOL(ptlrpc_hpreq_handler);
1519 * Make the request a high priority one.
1521 * All the high priority requests are queued in a separate FIFO
1522 * ptlrpc_service_part::scp_hpreq_pending list which is parallel to
1523 * ptlrpc_service_part::scp_req_pending list but has a higher priority
1526 * \see ptlrpc_server_handle_request().
1528 static void ptlrpc_hpreq_reorder_nolock(struct ptlrpc_service_part *svcpt,
1529 struct ptlrpc_request *req)
1533 cfs_spin_lock(&req->rq_lock);
1534 if (req->rq_hp == 0) {
1535 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1537 /* Add to the high priority queue. */
1538 cfs_list_move_tail(&req->rq_list, &svcpt->scp_hreq_pending);
1540 if (opc != OBD_PING)
1541 DEBUG_REQ(D_RPCTRACE, req, "high priority req");
1543 cfs_spin_unlock(&req->rq_lock);
1548 * \see ptlrpc_hpreq_reorder_nolock
1550 void ptlrpc_hpreq_reorder(struct ptlrpc_request *req)
1552 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
1555 cfs_spin_lock(&svcpt->scp_req_lock);
1556 /* It may happen that the request is already taken for the processing
1557 * but still in the export list, or the request is not in the request
1558 * queue but in the export list already, do not add it into the
1560 if (!cfs_list_empty(&req->rq_list))
1561 ptlrpc_hpreq_reorder_nolock(svcpt, req);
1562 cfs_spin_unlock(&svcpt->scp_req_lock);
1565 EXPORT_SYMBOL(ptlrpc_hpreq_reorder);
1567 /** Check if the request is a high priority one. */
1568 static int ptlrpc_server_hpreq_check(struct ptlrpc_service *svc,
1569 struct ptlrpc_request *req)
1571 return ptlrpc_hpreq_init(svc, req);
1574 /** Check if a request is a high priority one. */
1575 static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
1576 struct ptlrpc_request *req)
1581 rc = ptlrpc_server_hpreq_check(svcpt->scp_service, req);
1585 cfs_spin_lock(&svcpt->scp_req_lock);
1588 ptlrpc_hpreq_reorder_nolock(svcpt, req);
1590 cfs_list_add_tail(&req->rq_list, &svcpt->scp_req_pending);
1592 cfs_spin_unlock(&svcpt->scp_req_lock);
1598 * Allow to handle high priority request
1599 * User can call it w/o any lock but need to hold
1600 * ptlrpc_service_part::scp_req_lock to get reliable result
1602 static int ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
1608 if (svcpt->scp_nreqs_active >= svcpt->scp_nthrs_running - 1)
1611 return cfs_list_empty(&svcpt->scp_req_pending) ||
1612 svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
1615 static int ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
1618 return ptlrpc_server_allow_high(svcpt, force) &&
1619 !cfs_list_empty(&svcpt->scp_hreq_pending);
1623 * Only allow normal priority requests on a service that has a high-priority
1624 * queue if forced (i.e. cleanup), if there are other high priority requests
1625 * already being processed (i.e. those threads can service more high-priority
1626 * requests), or if there are enough idle threads that a later thread can do
1627 * a high priority request.
1628 * User can call it w/o any lock but need to hold
1629 * ptlrpc_service_part::scp_req_lock to get reliable result
1631 static int ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
1635 if (1) /* always allow to handle normal request for liblustre */
1639 svcpt->scp_nreqs_active < svcpt->scp_nthrs_running - 2)
1642 if (svcpt->scp_nreqs_active >= svcpt->scp_nthrs_running - 1)
1645 return svcpt->scp_nhreqs_active > 0 ||
1646 svcpt->scp_service->srv_ops.so_hpreq_handler == NULL;
1649 static int ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
1652 return ptlrpc_server_allow_normal(svcpt, force) &&
1653 !cfs_list_empty(&svcpt->scp_req_pending);
1657 * Returns true if there are requests available in incoming
1658 * request queue for processing and it is allowed to fetch them.
1659 * User can call it w/o any lock but need to hold ptlrpc_service::scp_req_lock
1660 * to get reliable result
1661 * \see ptlrpc_server_allow_normal
1662 * \see ptlrpc_server_allow high
1665 ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt, int force)
1667 return ptlrpc_server_high_pending(svcpt, force) ||
1668 ptlrpc_server_normal_pending(svcpt, force);
1672 * Fetch a request for processing from queue of unprocessed requests.
1673 * Favors high-priority requests.
1674 * Returns a pointer to fetched request.
1676 static struct ptlrpc_request *
1677 ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, int force)
1679 struct ptlrpc_request *req;
1682 if (ptlrpc_server_high_pending(svcpt, force)) {
1683 req = cfs_list_entry(svcpt->scp_hreq_pending.next,
1684 struct ptlrpc_request, rq_list);
1685 svcpt->scp_hreq_count++;
1689 if (ptlrpc_server_normal_pending(svcpt, force)) {
1690 req = cfs_list_entry(svcpt->scp_req_pending.next,
1691 struct ptlrpc_request, rq_list);
1692 svcpt->scp_hreq_count = 0;
1699 * Handle freshly incoming reqs, add to timed early reply list,
1700 * pass on to regular request queue.
1701 * All incoming requests pass through here before getting into
1702 * ptlrpc_server_handle_req later on.
1705 ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt)
1707 struct ptlrpc_service *svc = svcpt->scp_service;
1708 struct ptlrpc_request *req;
1713 cfs_spin_lock(&svcpt->scp_lock);
1714 if (cfs_list_empty(&svcpt->scp_req_incoming)) {
1715 cfs_spin_unlock(&svcpt->scp_lock);
1719 req = cfs_list_entry(svcpt->scp_req_incoming.next,
1720 struct ptlrpc_request, rq_list);
1721 cfs_list_del_init(&req->rq_list);
1722 svcpt->scp_nreqs_incoming--;
1723 /* Consider this still a "queued" request as far as stats are
1725 cfs_spin_unlock(&svcpt->scp_lock);
1727 /* go through security check/transform */
1728 rc = sptlrpc_svc_unwrap_request(req);
1732 case SECSVC_COMPLETE:
1733 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1742 * for null-flavored rpc, msg has been unpacked by sptlrpc, although
1743 * redo it wouldn't be harmful.
1745 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
1746 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
1748 CERROR("error unpacking request: ptl %d from %s "
1749 "x"LPU64"\n", svc->srv_req_portal,
1750 libcfs_id2str(req->rq_peer), req->rq_xid);
1755 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1757 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
1758 LPU64"\n", svc->srv_req_portal,
1759 libcfs_id2str(req->rq_peer), req->rq_xid);
1763 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
1764 lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
1765 CERROR("drop incoming rpc opc %u, x"LPU64"\n",
1766 cfs_fail_val, req->rq_xid);
1771 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1772 CERROR("wrong packet type received (type=%u) from %s\n",
1773 lustre_msg_get_type(req->rq_reqmsg),
1774 libcfs_id2str(req->rq_peer));
1778 switch(lustre_msg_get_opc(req->rq_reqmsg)) {
1781 req->rq_bulk_write = 1;
1785 case MGS_CONFIG_READ:
1786 req->rq_bulk_read = 1;
1790 CDEBUG(D_RPCTRACE, "got req x"LPU64"\n", req->rq_xid);
1792 req->rq_export = class_conn2export(
1793 lustre_msg_get_handle(req->rq_reqmsg));
1794 if (req->rq_export) {
1795 rc = ptlrpc_check_req(req);
1797 rc = sptlrpc_target_export_check(req->rq_export, req);
1799 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
1800 "illegal security flavor,");
1805 ptlrpc_update_export_timer(req->rq_export, 0);
1808 /* req_in handling should/must be fast */
1809 if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5)
1810 DEBUG_REQ(D_WARNING, req, "Slow req_in handling "CFS_DURATION_T"s",
1811 cfs_time_sub(cfs_time_current_sec(),
1812 req->rq_arrival_time.tv_sec));
1814 /* Set rpc server deadline and add it to the timed list */
1815 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1816 MSGHDR_AT_SUPPORT) ?
1817 /* The max time the client expects us to take */
1818 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1819 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1820 if (unlikely(deadline == 0)) {
1821 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1825 ptlrpc_at_add_timed(req);
1827 /* Move it over to the request processing queue */
1828 rc = ptlrpc_server_request_add(svcpt, req);
1830 ptlrpc_hpreq_fini(req);
1833 cfs_waitq_signal(&svcpt->scp_waitq);
1837 cfs_spin_lock(&svcpt->scp_req_lock);
1838 svcpt->scp_nreqs_active++;
1839 cfs_spin_unlock(&svcpt->scp_req_lock);
1840 ptlrpc_server_finish_request(svcpt, req);
1846 * Main incoming request handling logic.
1847 * Calls handler function from service to do actual processing.
1850 ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
1851 struct ptlrpc_thread *thread)
1853 struct ptlrpc_service *svc = svcpt->scp_service;
1854 struct obd_export *export = NULL;
1855 struct ptlrpc_request *request;
1856 struct timeval work_start;
1857 struct timeval work_end;
1863 cfs_spin_lock(&svcpt->scp_req_lock);
1865 /* !@%$# liblustre only has 1 thread */
1866 if (cfs_atomic_read(&svcpt->scp_nreps_difficult) != 0) {
1867 cfs_spin_unlock(&svcpt->scp_req_lock);
1871 request = ptlrpc_server_request_get(svcpt, 0);
1872 if (request == NULL) {
1873 cfs_spin_unlock(&svcpt->scp_req_lock);
1877 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
1878 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
1879 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
1880 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
1882 if (unlikely(fail_opc)) {
1883 if (request->rq_export && request->rq_ops) {
1884 cfs_spin_unlock(&svcpt->scp_req_lock);
1886 OBD_FAIL_TIMEOUT(fail_opc, 4);
1888 cfs_spin_lock(&svcpt->scp_req_lock);
1889 request = ptlrpc_server_request_get(svcpt, 0);
1890 if (request == NULL) {
1891 cfs_spin_unlock(&svcpt->scp_req_lock);
1897 cfs_list_del_init(&request->rq_list);
1898 svcpt->scp_nreqs_active++;
1900 svcpt->scp_nhreqs_active++;
1902 cfs_spin_unlock(&svcpt->scp_req_lock);
1904 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
1906 if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
1907 libcfs_debug_dumplog();
1909 cfs_gettimeofday(&work_start);
1910 timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
1911 if (likely(svc->srv_stats != NULL)) {
1912 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
1914 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
1915 svcpt->scp_nreqs_incoming);
1916 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
1917 svcpt->scp_nreqs_active);
1918 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
1919 at_get(&svcpt->scp_at_estimate));
1922 rc = lu_context_init(&request->rq_session, LCT_SESSION | LCT_NOREF);
1924 CERROR("Failure to initialize session: %d\n", rc);
1927 request->rq_session.lc_thread = thread;
1928 request->rq_session.lc_cookie = 0x5;
1929 lu_context_enter(&request->rq_session);
1931 CDEBUG(D_NET, "got req "LPU64"\n", request->rq_xid);
1933 request->rq_svc_thread = thread;
1935 request->rq_svc_thread->t_env->le_ses = &request->rq_session;
1937 if (likely(request->rq_export)) {
1938 if (unlikely(ptlrpc_check_req(request)))
1940 ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
1941 export = class_export_rpc_get(request->rq_export);
1944 /* Discard requests queued for longer than the deadline.
1945 The deadline is increased if we send an early reply. */
1946 if (cfs_time_current_sec() > request->rq_deadline) {
1947 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s"
1948 ": deadline "CFS_DURATION_T":"CFS_DURATION_T"s ago\n",
1949 libcfs_id2str(request->rq_peer),
1950 cfs_time_sub(request->rq_deadline,
1951 request->rq_arrival_time.tv_sec),
1952 cfs_time_sub(cfs_time_current_sec(),
1953 request->rq_deadline));
1954 goto put_rpc_export;
1957 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
1958 "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1959 (request->rq_export ?
1960 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1961 (request->rq_export ?
1962 cfs_atomic_read(&request->rq_export->exp_refcount) : -99),
1963 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1964 libcfs_id2str(request->rq_peer),
1965 lustre_msg_get_opc(request->rq_reqmsg));
1967 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
1968 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
1970 rc = svc->srv_ops.so_req_handler(request);
1972 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
1976 class_export_rpc_put(export);
1978 lu_context_exit(&request->rq_session);
1979 lu_context_fini(&request->rq_session);
1981 if (unlikely(cfs_time_current_sec() > request->rq_deadline)) {
1982 DEBUG_REQ(D_WARNING, request, "Request x"LPU64" took longer "
1983 "than estimated ("CFS_DURATION_T":"CFS_DURATION_T"s);"
1984 " client may timeout.",
1985 request->rq_xid, cfs_time_sub(request->rq_deadline,
1986 request->rq_arrival_time.tv_sec),
1987 cfs_time_sub(cfs_time_current_sec(),
1988 request->rq_deadline));
1991 cfs_gettimeofday(&work_end);
1992 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1993 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
1994 "%s:%s+%d:%d:x"LPU64":%s:%d Request procesed in "
1995 "%ldus (%ldus total) trans "LPU64" rc %d/%d\n",
1997 (request->rq_export ?
1998 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1999 (request->rq_export ?
2000 cfs_atomic_read(&request->rq_export->exp_refcount) : -99),
2001 lustre_msg_get_status(request->rq_reqmsg),
2003 libcfs_id2str(request->rq_peer),
2004 lustre_msg_get_opc(request->rq_reqmsg),
2006 cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL),
2007 (request->rq_repmsg ?
2008 lustre_msg_get_transno(request->rq_repmsg) :
2009 request->rq_transno),
2011 (request->rq_repmsg ?
2012 lustre_msg_get_status(request->rq_repmsg) : -999));
2013 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
2014 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
2015 int opc = opcode_offset(op);
2016 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
2017 LASSERT(opc < LUSTRE_MAX_OPCODES);
2018 lprocfs_counter_add(svc->srv_stats,
2019 opc + EXTRA_MAX_OPCODES,
2023 if (unlikely(request->rq_early_count)) {
2024 DEBUG_REQ(D_ADAPTTO, request,
2025 "sent %d early replies before finishing in "
2027 request->rq_early_count,
2028 cfs_time_sub(work_end.tv_sec,
2029 request->rq_arrival_time.tv_sec));
2033 ptlrpc_server_finish_request(svcpt, request);
2039 * An internal function to process a single reply state object.
2042 ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
2044 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
2045 struct ptlrpc_service *svc = svcpt->scp_service;
2046 struct obd_export *exp;
2051 exp = rs->rs_export;
2053 LASSERT (rs->rs_difficult);
2054 LASSERT (rs->rs_scheduled);
2055 LASSERT (cfs_list_empty(&rs->rs_list));
2057 cfs_spin_lock (&exp->exp_lock);
2058 /* Noop if removed already */
2059 cfs_list_del_init (&rs->rs_exp_list);
2060 cfs_spin_unlock (&exp->exp_lock);
2062 /* The disk commit callback holds exp_uncommitted_replies_lock while it
2063 * iterates over newly committed replies, removing them from
2064 * exp_uncommitted_replies. It then drops this lock and schedules the
2065 * replies it found for handling here.
2067 * We can avoid contention for exp_uncommitted_replies_lock between the
2068 * HRT threads and further commit callbacks by checking rs_committed
2069 * which is set in the commit callback while it holds both
2070 * rs_lock and exp_uncommitted_reples.
2072 * If we see rs_committed clear, the commit callback _may_ not have
2073 * handled this reply yet and we race with it to grab
2074 * exp_uncommitted_replies_lock before removing the reply from
2075 * exp_uncommitted_replies. Note that if we lose the race and the
2076 * reply has already been removed, list_del_init() is a noop.
2078 * If we see rs_committed set, we know the commit callback is handling,
2079 * or has handled this reply since store reordering might allow us to
2080 * see rs_committed set out of sequence. But since this is done
2081 * holding rs_lock, we can be sure it has all completed once we hold
2082 * rs_lock, which we do right next.
2084 if (!rs->rs_committed) {
2085 cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
2086 cfs_list_del_init(&rs->rs_obd_list);
2087 cfs_spin_unlock(&exp->exp_uncommitted_replies_lock);
2090 cfs_spin_lock(&rs->rs_lock);
2092 been_handled = rs->rs_handled;
2095 nlocks = rs->rs_nlocks; /* atomic "steal", but */
2096 rs->rs_nlocks = 0; /* locks still on rs_locks! */
2098 if (nlocks == 0 && !been_handled) {
2099 /* If we see this, we should already have seen the warning
2100 * in mds_steal_ack_locks() */
2101 CWARN("All locks stolen from rs %p x"LPD64".t"LPD64
2104 rs->rs_xid, rs->rs_transno, rs->rs_opc,
2105 libcfs_nid2str(exp->exp_connection->c_peer.nid));
2108 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
2109 cfs_spin_unlock(&rs->rs_lock);
2111 if (!been_handled && rs->rs_on_net) {
2112 LNetMDUnlink(rs->rs_md_h);
2113 /* Ignore return code; we're racing with
2117 while (nlocks-- > 0)
2118 ldlm_lock_decref(&rs->rs_locks[nlocks],
2119 rs->rs_modes[nlocks]);
2121 cfs_spin_lock(&rs->rs_lock);
2124 rs->rs_scheduled = 0;
2126 if (!rs->rs_on_net) {
2128 cfs_spin_unlock(&rs->rs_lock);
2130 class_export_put (exp);
2131 rs->rs_export = NULL;
2132 ptlrpc_rs_decref (rs);
2133 if (cfs_atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
2134 svc->srv_is_stopping)
2135 cfs_waitq_broadcast(&svcpt->scp_waitq);
2139 /* still on the net; callback will schedule */
2140 cfs_spin_unlock(&rs->rs_lock);
2147 * Check whether given service has a reply available for processing
2150 * \param svc a ptlrpc service
2151 * \retval 0 no replies processed
2152 * \retval 1 one reply processed
2155 ptlrpc_server_handle_reply(struct ptlrpc_service_part *svcpt)
2157 struct ptlrpc_reply_state *rs = NULL;
2160 cfs_spin_lock(&svcpt->scp_rep_lock);
2161 if (!cfs_list_empty(&svcpt->scp_rep_queue)) {
2162 rs = cfs_list_entry(svcpt->scp_rep_queue.prev,
2163 struct ptlrpc_reply_state,
2165 cfs_list_del_init(&rs->rs_list);
2167 cfs_spin_unlock(&svcpt->scp_rep_lock);
2169 ptlrpc_handle_rs(rs);
2173 /* FIXME make use of timeout later */
2175 liblustre_check_services (void *arg)
2177 int did_something = 0;
2179 cfs_list_t *tmp, *nxt;
2182 /* I'm relying on being single threaded, not to have to lock
2183 * ptlrpc_all_services etc */
2184 cfs_list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
2185 struct ptlrpc_service *svc =
2186 cfs_list_entry (tmp, struct ptlrpc_service, srv_list);
2187 struct ptlrpc_service_part *svcpt;
2189 LASSERT(svc->srv_ncpts == 1);
2190 svcpt = svc->srv_parts[0];
2192 if (svcpt->scp_nthrs_running != 0) /* I've recursed */
2195 /* service threads can block for bulk, so this limits us
2196 * (arbitrarily) to recursing 1 stack frame per service.
2197 * Note that the problem with recursion is that we have to
2198 * unwind completely before our caller can resume. */
2200 svcpt->scp_nthrs_running++;
2203 rc = ptlrpc_server_handle_req_in(svcpt);
2204 rc |= ptlrpc_server_handle_reply(svcpt);
2205 rc |= ptlrpc_at_check_timed(svcpt);
2206 rc |= ptlrpc_server_handle_request(svcpt, NULL);
2207 rc |= (ptlrpc_server_post_idle_rqbds(svcpt) > 0);
2208 did_something |= rc;
2211 svcpt->scp_nthrs_running--;
2214 RETURN(did_something);
2216 #define ptlrpc_stop_all_threads(s) do {} while (0)
2218 #else /* __KERNEL__ */
2221 ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
2223 int avail = svcpt->scp_nrqbds_posted;
2224 int low_water = test_req_buffer_pressure ? 0 :
2225 svcpt->scp_service->srv_nbuf_per_group / 2;
2227 /* NB I'm not locking; just looking. */
2229 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
2230 * allowed the request history to grow out of control. We could put a
2231 * sanity check on that here and cull some history if we need the
2234 if (avail <= low_water)
2235 ptlrpc_grow_req_bufs(svcpt, 1);
2237 if (svcpt->scp_service->srv_stats) {
2238 lprocfs_counter_add(svcpt->scp_service->srv_stats,
2239 PTLRPC_REQBUF_AVAIL_CNTR, avail);
2244 ptlrpc_retry_rqbds(void *arg)
2246 struct ptlrpc_service_part *svcpt = (struct ptlrpc_service_part *)arg;
2248 svcpt->scp_rqbd_timeout = 0;
2253 ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
2255 return svcpt->scp_nreqs_active <
2256 svcpt->scp_nthrs_running - 1 -
2257 (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL);
2261 * allowed to create more threads
2262 * user can call it w/o any lock but need to hold
2263 * ptlrpc_service_part::scp_lock to get reliable result
2266 ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
2268 return svcpt->scp_nthrs_running +
2269 svcpt->scp_nthrs_starting <
2270 svcpt->scp_service->srv_nthrs_cpt_limit;
2274 * too many requests and allowed to create more threads
2277 ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
2279 return !ptlrpc_threads_enough(svcpt) &&
2280 ptlrpc_threads_increasable(svcpt);
2284 ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2286 return thread_is_stopping(thread) ||
2287 thread->t_svcpt->scp_service->srv_is_stopping;
2291 ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
2293 return !cfs_list_empty(&svcpt->scp_rqbd_idle) &&
2294 svcpt->scp_rqbd_timeout == 0;
2298 ptlrpc_at_check(struct ptlrpc_service_part *svcpt)
2300 return svcpt->scp_at_check;
2304 * requests wait on preprocessing
2305 * user can call it w/o any lock but need to hold
2306 * ptlrpc_service_part::scp_lock to get reliable result
2309 ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
2311 return !cfs_list_empty(&svcpt->scp_req_incoming);
2314 static __attribute__((__noinline__)) int
2315 ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
2316 struct ptlrpc_thread *thread)
2318 /* Don't exit while there are replies to be handled */
2319 struct l_wait_info lwi = LWI_TIMEOUT(svcpt->scp_rqbd_timeout,
2320 ptlrpc_retry_rqbds, svcpt);
2322 lc_watchdog_disable(thread->t_watchdog);
2326 l_wait_event_exclusive_head(svcpt->scp_waitq,
2327 ptlrpc_thread_stopping(thread) ||
2328 ptlrpc_server_request_incoming(svcpt) ||
2329 ptlrpc_server_request_pending(svcpt, 0) ||
2330 ptlrpc_rqbd_pending(svcpt) ||
2331 ptlrpc_at_check(svcpt), &lwi);
2333 if (ptlrpc_thread_stopping(thread))
2336 lc_watchdog_touch(thread->t_watchdog,
2337 ptlrpc_server_get_timeout(svcpt));
2342 * Main thread body for service threads.
2343 * Waits in a loop waiting for new requests to process to appear.
2344 * Every time an incoming requests is added to its queue, a waitq
2345 * is woken up and one of the threads will handle it.
2347 static int ptlrpc_main(void *arg)
2349 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
2350 struct ptlrpc_service_part *svcpt = thread->t_svcpt;
2351 struct ptlrpc_service *svc = svcpt->scp_service;
2352 struct ptlrpc_reply_state *rs;
2353 #ifdef WITH_GROUP_INFO
2354 cfs_group_info_t *ginfo = NULL;
2357 int counter = 0, rc = 0;
2360 thread->t_pid = cfs_curproc_pid();
2361 cfs_daemonize_ctxt(thread->t_name);
2363 /* NB: we will call cfs_cpt_bind() for all threads, because we
2364 * might want to run lustre server only on a subset of system CPUs,
2365 * in that case ->scp_cpt is CFS_CPT_ANY */
2366 rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
2368 CWARN("%s: failed to bind %s on CPT %d\n",
2369 svc->srv_name, thread->t_name, svcpt->scp_cpt);
2372 #ifdef WITH_GROUP_INFO
2373 ginfo = cfs_groups_alloc(0);
2379 cfs_set_current_groups(ginfo);
2380 cfs_put_group_info(ginfo);
2383 if (svc->srv_ops.so_thr_init != NULL) {
2384 rc = svc->srv_ops.so_thr_init(thread);
2395 rc = lu_context_init(&env->le_ctx,
2396 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2400 thread->t_env = env;
2401 env->le_ctx.lc_thread = thread;
2402 env->le_ctx.lc_cookie = 0x6;
2404 while (!cfs_list_empty(&svcpt->scp_rqbd_idle)) {
2405 rc = ptlrpc_server_post_idle_rqbds(svcpt);
2409 CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
2410 svc->srv_name, svcpt->scp_cpt, rc);
2414 /* Alloc reply state structure for this one */
2415 OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2421 cfs_spin_lock(&svcpt->scp_lock);
2423 LASSERT(thread_is_starting(thread));
2424 thread_clear_flags(thread, SVC_STARTING);
2426 LASSERT(svcpt->scp_nthrs_starting == 1);
2427 svcpt->scp_nthrs_starting--;
2429 /* SVC_STOPPING may already be set here if someone else is trying
2430 * to stop the service while this new thread has been dynamically
2431 * forked. We still set SVC_RUNNING to let our creator know that
2432 * we are now running, however we will exit as soon as possible */
2433 thread_add_flags(thread, SVC_RUNNING);
2434 svcpt->scp_nthrs_running++;
2435 cfs_spin_unlock(&svcpt->scp_lock);
2437 /* wake up our creator in case he's still waiting. */
2438 cfs_waitq_signal(&thread->t_ctl_waitq);
2440 thread->t_watchdog = lc_watchdog_add(ptlrpc_server_get_timeout(svcpt),
2443 cfs_spin_lock(&svcpt->scp_rep_lock);
2444 cfs_list_add(&rs->rs_list, &svcpt->scp_rep_idle);
2445 cfs_waitq_signal(&svcpt->scp_rep_waitq);
2446 cfs_spin_unlock(&svcpt->scp_rep_lock);
2448 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2449 svcpt->scp_nthrs_running);
2451 /* XXX maintain a list of all managed devices: insert here */
2452 while (!ptlrpc_thread_stopping(thread)) {
2453 if (ptlrpc_wait_event(svcpt, thread))
2456 ptlrpc_check_rqbd_pool(svcpt);
2458 if (ptlrpc_threads_need_create(svcpt)) {
2459 /* Ignore return code - we tried... */
2460 ptlrpc_start_thread(svcpt, 0);
2463 /* Process all incoming reqs before handling any */
2464 if (ptlrpc_server_request_incoming(svcpt)) {
2465 ptlrpc_server_handle_req_in(svcpt);
2466 /* but limit ourselves in case of flood */
2467 if (counter++ < 100)
2472 if (ptlrpc_at_check(svcpt))
2473 ptlrpc_at_check_timed(svcpt);
2475 if (ptlrpc_server_request_pending(svcpt, 0)) {
2476 lu_context_enter(&env->le_ctx);
2477 ptlrpc_server_handle_request(svcpt, thread);
2478 lu_context_exit(&env->le_ctx);
2481 if (ptlrpc_rqbd_pending(svcpt) &&
2482 ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
2483 /* I just failed to repost request buffers.
2484 * Wait for a timeout (unless something else
2485 * happens) before I try again */
2486 svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
2487 CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
2488 svcpt->scp_nrqbds_posted);
2492 lc_watchdog_delete(thread->t_watchdog);
2493 thread->t_watchdog = NULL;
2497 * deconstruct service specific state created by ptlrpc_start_thread()
2499 if (svc->srv_ops.so_thr_done != NULL)
2500 svc->srv_ops.so_thr_done(thread);
2503 lu_context_fini(&env->le_ctx);
2507 CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n",
2508 thread, thread->t_pid, thread->t_id, rc);
2510 cfs_spin_lock(&svcpt->scp_lock);
2511 if (thread_test_and_clear_flags(thread, SVC_STARTING))
2512 svcpt->scp_nthrs_starting--;
2514 if (thread_test_and_clear_flags(thread, SVC_RUNNING)) {
2515 /* must know immediately */
2516 svcpt->scp_nthrs_running--;
2520 thread_add_flags(thread, SVC_STOPPED);
2522 cfs_waitq_signal(&thread->t_ctl_waitq);
2523 cfs_spin_unlock(&svcpt->scp_lock);
2528 static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
2529 cfs_list_t *replies)
2533 cfs_spin_lock(&hrt->hrt_lock);
2535 cfs_list_splice_init(&hrt->hrt_queue, replies);
2536 result = ptlrpc_hr.hr_stopping || !cfs_list_empty(replies);
2538 cfs_spin_unlock(&hrt->hrt_lock);
2543 * Main body of "handle reply" function.
2544 * It processes acked reply states
2546 static int ptlrpc_hr_main(void *arg)
2548 struct ptlrpc_hr_thread *hrt = (struct ptlrpc_hr_thread *)arg;
2549 struct ptlrpc_hr_partition *hrp = hrt->hrt_partition;
2550 CFS_LIST_HEAD (replies);
2551 char threadname[20];
2554 snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
2555 hrp->hrp_cpt, hrt->hrt_id);
2556 cfs_daemonize_ctxt(threadname);
2558 rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
2560 CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
2561 threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
2564 cfs_atomic_inc(&hrp->hrp_nstarted);
2565 cfs_waitq_signal(&ptlrpc_hr.hr_waitq);
2567 while (!ptlrpc_hr.hr_stopping) {
2568 l_wait_condition(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
2570 while (!cfs_list_empty(&replies)) {
2571 struct ptlrpc_reply_state *rs;
2573 rs = cfs_list_entry(replies.prev,
2574 struct ptlrpc_reply_state,
2576 cfs_list_del_init(&rs->rs_list);
2577 ptlrpc_handle_rs(rs);
2581 cfs_atomic_inc(&hrp->hrp_nstopped);
2582 cfs_waitq_signal(&ptlrpc_hr.hr_waitq);
2587 static void ptlrpc_stop_hr_threads(void)
2589 struct ptlrpc_hr_partition *hrp;
2593 ptlrpc_hr.hr_stopping = 1;
2595 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2596 if (hrp->hrp_thrs == NULL)
2597 continue; /* uninitialized */
2598 for (j = 0; j < hrp->hrp_nthrs; j++)
2599 cfs_waitq_broadcast(&hrp->hrp_thrs[j].hrt_waitq);
2602 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2603 if (hrp->hrp_thrs == NULL)
2604 continue; /* uninitialized */
2605 cfs_wait_event(ptlrpc_hr.hr_waitq,
2606 cfs_atomic_read(&hrp->hrp_nstopped) ==
2607 cfs_atomic_read(&hrp->hrp_nstarted));
2611 static int ptlrpc_start_hr_threads(void)
2613 struct ptlrpc_hr_partition *hrp;
2618 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2621 for (j = 0; j < hrp->hrp_nthrs; j++) {
2622 rc = cfs_create_thread(ptlrpc_hr_main,
2624 CLONE_VM | CLONE_FILES);
2628 cfs_wait_event(ptlrpc_hr.hr_waitq,
2629 cfs_atomic_read(&hrp->hrp_nstarted) == j);
2633 CERROR("Reply handling thread %d:%d Failed on starting: "
2634 "rc = %d\n", i, j, rc);
2635 ptlrpc_stop_hr_threads();
2641 static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
2643 struct l_wait_info lwi = { 0 };
2644 struct ptlrpc_thread *thread;
2645 CFS_LIST_HEAD (zombie);
2649 CDEBUG(D_INFO, "Stopping threads for service %s\n",
2650 svcpt->scp_service->srv_name);
2652 cfs_spin_lock(&svcpt->scp_lock);
2653 /* let the thread know that we would like it to stop asap */
2654 list_for_each_entry(thread, &svcpt->scp_threads, t_link) {
2655 CDEBUG(D_INFO, "Stopping thread %s #%u\n",
2656 svcpt->scp_service->srv_thread_name, thread->t_id);
2657 thread_add_flags(thread, SVC_STOPPING);
2660 cfs_waitq_broadcast(&svcpt->scp_waitq);
2662 while (!cfs_list_empty(&svcpt->scp_threads)) {
2663 thread = cfs_list_entry(svcpt->scp_threads.next,
2664 struct ptlrpc_thread, t_link);
2665 if (thread_is_stopped(thread)) {
2666 cfs_list_del(&thread->t_link);
2667 cfs_list_add(&thread->t_link, &zombie);
2670 cfs_spin_unlock(&svcpt->scp_lock);
2672 CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
2673 svcpt->scp_service->srv_thread_name, thread->t_id);
2674 l_wait_event(thread->t_ctl_waitq,
2675 thread_is_stopped(thread), &lwi);
2677 cfs_spin_lock(&svcpt->scp_lock);
2680 cfs_spin_unlock(&svcpt->scp_lock);
2682 while (!cfs_list_empty(&zombie)) {
2683 thread = cfs_list_entry(zombie.next,
2684 struct ptlrpc_thread, t_link);
2685 cfs_list_del(&thread->t_link);
2686 OBD_FREE_PTR(thread);
2692 * Stops all threads of a particular service \a svc
2694 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2696 struct ptlrpc_service_part *svcpt;
2700 ptlrpc_service_for_each_part(svcpt, i, svc) {
2701 if (svcpt->scp_service != NULL)
2702 ptlrpc_svcpt_stop_threads(svcpt);
2707 EXPORT_SYMBOL(ptlrpc_stop_all_threads);
2709 int ptlrpc_start_threads(struct ptlrpc_service *svc)
2716 /* We require 2 threads min, see note in ptlrpc_server_handle_request */
2717 LASSERT(svc->srv_nthrs_cpt_init >= PTLRPC_NTHRS_INIT);
2719 for (i = 0; i < svc->srv_ncpts; i++) {
2720 for (j = 0; j < svc->srv_nthrs_cpt_init; j++) {
2721 rc = ptlrpc_start_thread(svc->srv_parts[i], 1);
2727 /* We have enough threads, don't start more. b=15759 */
2734 CERROR("cannot start %s thread #%d_%d: rc %d\n",
2735 svc->srv_thread_name, i, j, rc);
2736 ptlrpc_stop_all_threads(svc);
2739 EXPORT_SYMBOL(ptlrpc_start_threads);
2741 int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
2743 struct l_wait_info lwi = { 0 };
2744 struct ptlrpc_thread *thread;
2745 struct ptlrpc_service *svc = svcpt->scp_service;
2749 LASSERT(svcpt != NULL);
2751 CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n",
2752 svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running,
2753 svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit);
2756 if (unlikely(svc->srv_is_stopping))
2759 if (!ptlrpc_threads_increasable(svcpt) ||
2760 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
2761 svcpt->scp_nthrs_running == svc->srv_nthrs_cpt_init - 1))
2764 OBD_CPT_ALLOC_PTR(thread, svc->srv_cptable, svcpt->scp_cpt);
2767 cfs_waitq_init(&thread->t_ctl_waitq);
2769 cfs_spin_lock(&svcpt->scp_lock);
2770 if (!ptlrpc_threads_increasable(svcpt)) {
2771 cfs_spin_unlock(&svcpt->scp_lock);
2772 OBD_FREE_PTR(thread);
2776 if (svcpt->scp_nthrs_starting != 0) {
2777 /* serialize starting because some modules (obdfilter)
2778 * might require unique and contiguous t_id */
2779 LASSERT(svcpt->scp_nthrs_starting == 1);
2780 cfs_spin_unlock(&svcpt->scp_lock);
2781 OBD_FREE_PTR(thread);
2783 CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n",
2784 svc->srv_thread_name, svcpt->scp_thr_nextid);
2789 CDEBUG(D_INFO, "Creating thread %s #%d race, retry later\n",
2790 svc->srv_thread_name, svcpt->scp_thr_nextid);
2794 svcpt->scp_nthrs_starting++;
2795 thread->t_id = svcpt->scp_thr_nextid++;
2796 thread_add_flags(thread, SVC_STARTING);
2797 thread->t_svcpt = svcpt;
2799 cfs_list_add(&thread->t_link, &svcpt->scp_threads);
2800 cfs_spin_unlock(&svcpt->scp_lock);
2802 if (svcpt->scp_cpt >= 0) {
2803 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s%02d_%03d",
2804 svc->srv_thread_name, svcpt->scp_cpt, thread->t_id);
2806 snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s_%04d",
2807 svc->srv_thread_name, thread->t_id);
2810 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name);
2812 * CLONE_VM and CLONE_FILES just avoid a needless copy, because we
2813 * just drop the VM and FILES in cfs_daemonize_ctxt() right away.
2815 rc = cfs_create_thread(ptlrpc_main, thread, CFS_DAEMON_FLAGS);
2817 CERROR("cannot start thread '%s': rc %d\n",
2818 thread->t_name, rc);
2819 cfs_spin_lock(&svcpt->scp_lock);
2820 cfs_list_del(&thread->t_link);
2821 --svcpt->scp_nthrs_starting;
2822 cfs_spin_unlock(&svcpt->scp_lock);
2824 OBD_FREE(thread, sizeof(*thread));
2831 l_wait_event(thread->t_ctl_waitq,
2832 thread_is_running(thread) || thread_is_stopped(thread),
2835 rc = thread_is_stopped(thread) ? thread->t_id : 0;
2839 int ptlrpc_hr_init(void)
2841 struct ptlrpc_hr_partition *hrp;
2842 struct ptlrpc_hr_thread *hrt;
2848 memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
2849 ptlrpc_hr.hr_cpt_table = cfs_cpt_table;
2851 ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
2853 if (ptlrpc_hr.hr_partitions == NULL)
2856 cfs_waitq_init(&ptlrpc_hr.hr_waitq);
2858 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2861 cfs_atomic_set(&hrp->hrp_nstarted, 0);
2862 cfs_atomic_set(&hrp->hrp_nstopped, 0);
2864 hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, i);
2865 hrp->hrp_nthrs /= cfs_cpu_ht_nsiblings(0);
2867 LASSERT(hrp->hrp_nthrs > 0);
2868 OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, i,
2869 hrp->hrp_nthrs * sizeof(*hrt));
2870 if (hrp->hrp_thrs == NULL)
2871 GOTO(out, rc = -ENOMEM);
2873 for (j = 0; j < hrp->hrp_nthrs; j++) {
2874 hrt = &hrp->hrp_thrs[j];
2877 hrt->hrt_partition = hrp;
2878 cfs_waitq_init(&hrt->hrt_waitq);
2879 cfs_spin_lock_init(&hrt->hrt_lock);
2880 CFS_INIT_LIST_HEAD(&hrt->hrt_queue);
2884 rc = ptlrpc_start_hr_threads();
2891 void ptlrpc_hr_fini(void)
2893 struct ptlrpc_hr_partition *hrp;
2896 if (ptlrpc_hr.hr_partitions == NULL)
2899 ptlrpc_stop_hr_threads();
2901 cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
2902 if (hrp->hrp_thrs != NULL) {
2903 OBD_FREE(hrp->hrp_thrs,
2904 hrp->hrp_nthrs * sizeof(hrp->hrp_thrs[0]));
2908 cfs_percpt_free(ptlrpc_hr.hr_partitions);
2909 ptlrpc_hr.hr_partitions = NULL;
2912 #endif /* __KERNEL__ */
2915 * Wait until all already scheduled replies are processed.
2917 static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
2921 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
2924 rc = l_wait_event(svcpt->scp_waitq,
2925 cfs_atomic_read(&svcpt->scp_nreps_difficult) == 0, &lwi);
2928 CWARN("Unexpectedly long timeout %s %p\n",
2929 svcpt->scp_service->srv_name, svcpt->scp_service);
2934 ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
2936 struct ptlrpc_service_part *svcpt;
2939 /* early disarm AT timer... */
2940 ptlrpc_service_for_each_part(svcpt, i, svc) {
2941 if (svcpt->scp_service != NULL)
2942 cfs_timer_disarm(&svcpt->scp_at_timer);
2947 ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
2949 struct ptlrpc_service_part *svcpt;
2950 struct ptlrpc_request_buffer_desc *rqbd;
2951 struct l_wait_info lwi;
2955 /* All history will be culled when the next request buffer is
2956 * freed in ptlrpc_service_purge_all() */
2957 svc->srv_hist_nrqbds_cpt_max = 0;
2959 rc = LNetClearLazyPortal(svc->srv_req_portal);
2962 ptlrpc_service_for_each_part(svcpt, i, svc) {
2963 if (svcpt->scp_service == NULL)
2966 /* Unlink all the request buffers. This forces a 'final'
2967 * event with its 'unlink' flag set for each posted rqbd */
2968 cfs_list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
2970 rc = LNetMDUnlink(rqbd->rqbd_md_h);
2971 LASSERT(rc == 0 || rc == -ENOENT);
2975 ptlrpc_service_for_each_part(svcpt, i, svc) {
2976 if (svcpt->scp_service == NULL)
2979 /* Wait for the network to release any buffers
2980 * it's currently filling */
2981 cfs_spin_lock(&svcpt->scp_lock);
2982 while (svcpt->scp_nrqbds_posted != 0) {
2983 cfs_spin_unlock(&svcpt->scp_lock);
2984 /* Network access will complete in finite time but
2985 * the HUGE timeout lets us CWARN for visibility
2986 * of sluggish NALs */
2987 lwi = LWI_TIMEOUT_INTERVAL(
2988 cfs_time_seconds(LONG_UNLINK),
2989 cfs_time_seconds(1), NULL, NULL);
2990 rc = l_wait_event(svcpt->scp_waitq,
2991 svcpt->scp_nrqbds_posted == 0, &lwi);
2992 if (rc == -ETIMEDOUT) {
2993 CWARN("Service %s waiting for "
2994 "request buffers\n",
2995 svcpt->scp_service->srv_name);
2997 cfs_spin_lock(&svcpt->scp_lock);
2999 cfs_spin_unlock(&svcpt->scp_lock);
3004 ptlrpc_service_purge_all(struct ptlrpc_service *svc)
3006 struct ptlrpc_service_part *svcpt;
3007 struct ptlrpc_request_buffer_desc *rqbd;
3008 struct ptlrpc_request *req;
3009 struct ptlrpc_reply_state *rs;
3012 ptlrpc_service_for_each_part(svcpt, i, svc) {
3013 if (svcpt->scp_service == NULL)
3016 cfs_spin_lock(&svcpt->scp_rep_lock);
3017 while (!cfs_list_empty(&svcpt->scp_rep_active)) {
3018 rs = cfs_list_entry(svcpt->scp_rep_active.next,
3019 struct ptlrpc_reply_state, rs_list);
3020 cfs_spin_lock(&rs->rs_lock);
3021 ptlrpc_schedule_difficult_reply(rs);
3022 cfs_spin_unlock(&rs->rs_lock);
3024 cfs_spin_unlock(&svcpt->scp_rep_lock);
3026 /* purge the request queue. NB No new replies (rqbds
3027 * all unlinked) and no service threads, so I'm the only
3028 * thread noodling the request queue now */
3029 while (!cfs_list_empty(&svcpt->scp_req_incoming)) {
3030 req = cfs_list_entry(svcpt->scp_req_incoming.next,
3031 struct ptlrpc_request, rq_list);
3033 cfs_list_del(&req->rq_list);
3034 svcpt->scp_nreqs_incoming--;
3035 svcpt->scp_nreqs_active++;
3036 ptlrpc_server_finish_request(svcpt, req);
3039 while (ptlrpc_server_request_pending(svcpt, 1)) {
3040 req = ptlrpc_server_request_get(svcpt, 1);
3041 cfs_list_del(&req->rq_list);
3042 svcpt->scp_nreqs_active++;
3043 ptlrpc_hpreq_fini(req);
3044 ptlrpc_server_finish_request(svcpt, req);
3047 LASSERT(cfs_list_empty(&svcpt->scp_rqbd_posted));
3048 LASSERT(svcpt->scp_nreqs_incoming == 0);
3049 LASSERT(svcpt->scp_nreqs_active == 0);
3050 /* history should have been culled by
3051 * ptlrpc_server_finish_request */
3052 LASSERT(svcpt->scp_hist_nrqbds == 0);
3054 /* Now free all the request buffers since nothing
3055 * references them any more... */
3057 while (!cfs_list_empty(&svcpt->scp_rqbd_idle)) {
3058 rqbd = cfs_list_entry(svcpt->scp_rqbd_idle.next,
3059 struct ptlrpc_request_buffer_desc,
3061 ptlrpc_free_rqbd(rqbd);
3063 ptlrpc_wait_replies(svcpt);
3065 while (!cfs_list_empty(&svcpt->scp_rep_idle)) {
3066 rs = cfs_list_entry(svcpt->scp_rep_idle.next,
3067 struct ptlrpc_reply_state,
3069 cfs_list_del(&rs->rs_list);
3070 OBD_FREE_LARGE(rs, svc->srv_max_reply_size);
3076 ptlrpc_service_free(struct ptlrpc_service *svc)
3078 struct ptlrpc_service_part *svcpt;
3079 struct ptlrpc_at_array *array;
3082 ptlrpc_service_for_each_part(svcpt, i, svc) {
3083 if (svcpt->scp_service == NULL)
3086 /* In case somebody rearmed this in the meantime */
3087 cfs_timer_disarm(&svcpt->scp_at_timer);
3088 array = &svcpt->scp_at_array;
3090 if (array->paa_reqs_array != NULL) {
3091 OBD_FREE(array->paa_reqs_array,
3092 sizeof(cfs_list_t) * array->paa_size);
3093 array->paa_reqs_array = NULL;
3096 if (array->paa_reqs_count != NULL) {
3097 OBD_FREE(array->paa_reqs_count,
3098 sizeof(__u32) * array->paa_size);
3099 array->paa_reqs_count = NULL;
3103 ptlrpc_service_for_each_part(svcpt, i, svc)
3104 OBD_FREE_PTR(svcpt);
3106 if (svc->srv_cpts != NULL)
3107 cfs_expr_list_values_free(svc->srv_cpts, svc->srv_ncpts);
3109 OBD_FREE(svc, offsetof(struct ptlrpc_service,
3110 srv_parts[svc->srv_ncpts]));
3113 int ptlrpc_unregister_service(struct ptlrpc_service *service)
3117 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
3119 service->srv_is_stopping = 1;
3121 cfs_spin_lock(&ptlrpc_all_services_lock);
3122 cfs_list_del_init(&service->srv_list);
3123 cfs_spin_unlock(&ptlrpc_all_services_lock);
3125 ptlrpc_lprocfs_unregister_service(service);
3127 ptlrpc_service_del_atimer(service);
3128 ptlrpc_stop_all_threads(service);
3130 ptlrpc_service_unlink_rqbd(service);
3131 ptlrpc_service_purge_all(service);
3132 ptlrpc_service_free(service);
3136 EXPORT_SYMBOL(ptlrpc_unregister_service);
3139 * Returns 0 if the service is healthy.
3141 * Right now, it just checks to make sure that requests aren't languishing
3142 * in the queue. We'll use this health check to govern whether a node needs
3143 * to be shot, so it's intentionally non-aggressive. */
3144 int ptlrpc_svcpt_health_check(struct ptlrpc_service_part *svcpt)
3146 struct ptlrpc_request *request;
3147 struct timeval right_now;
3150 cfs_gettimeofday(&right_now);
3152 cfs_spin_lock(&svcpt->scp_req_lock);
3153 if (!ptlrpc_server_request_pending(svcpt, 1)) {
3154 cfs_spin_unlock(&svcpt->scp_req_lock);
3158 /* How long has the next entry been waiting? */
3159 if (cfs_list_empty(&svcpt->scp_req_pending)) {
3160 request = cfs_list_entry(svcpt->scp_hreq_pending.next,
3161 struct ptlrpc_request, rq_list);
3163 request = cfs_list_entry(svcpt->scp_req_pending.next,
3164 struct ptlrpc_request, rq_list);
3167 timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
3168 cfs_spin_unlock(&svcpt->scp_req_lock);
3170 if ((timediff / ONE_MILLION) >
3171 (AT_OFF ? obd_timeout * 3 / 2 : at_max)) {
3172 CERROR("%s: unhealthy - request has been waiting %lds\n",
3173 svcpt->scp_service->srv_name, timediff / ONE_MILLION);
3181 ptlrpc_service_health_check(struct ptlrpc_service *svc)
3183 struct ptlrpc_service_part *svcpt;
3186 if (svc == NULL || svc->srv_parts == NULL)
3189 ptlrpc_service_for_each_part(svcpt, i, svc) {
3190 int rc = ptlrpc_svcpt_health_check(svcpt);
3197 EXPORT_SYMBOL(ptlrpc_service_health_check);