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 *svc);
67 static void ptlrpc_hpreq_fini(struct ptlrpc_request *req);
69 static CFS_LIST_HEAD(ptlrpc_all_services);
70 cfs_spinlock_t ptlrpc_all_services_lock;
72 struct ptlrpc_request_buffer_desc *
73 ptlrpc_alloc_rqbd (struct ptlrpc_service *svc)
75 struct ptlrpc_request_buffer_desc *rqbd;
81 rqbd->rqbd_service = svc;
82 rqbd->rqbd_refcount = 0;
83 rqbd->rqbd_cbid.cbid_fn = request_in_callback;
84 rqbd->rqbd_cbid.cbid_arg = rqbd;
85 CFS_INIT_LIST_HEAD(&rqbd->rqbd_reqs);
86 OBD_ALLOC_LARGE(rqbd->rqbd_buffer, svc->srv_buf_size);
88 if (rqbd->rqbd_buffer == NULL) {
93 cfs_spin_lock(&svc->srv_lock);
94 cfs_list_add(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
96 cfs_spin_unlock(&svc->srv_lock);
102 ptlrpc_free_rqbd (struct ptlrpc_request_buffer_desc *rqbd)
104 struct ptlrpc_service *svc = rqbd->rqbd_service;
106 LASSERT (rqbd->rqbd_refcount == 0);
107 LASSERT (cfs_list_empty(&rqbd->rqbd_reqs));
109 cfs_spin_lock(&svc->srv_lock);
110 cfs_list_del(&rqbd->rqbd_list);
112 cfs_spin_unlock(&svc->srv_lock);
114 OBD_FREE_LARGE(rqbd->rqbd_buffer, svc->srv_buf_size);
119 ptlrpc_grow_req_bufs(struct ptlrpc_service *svc)
121 struct ptlrpc_request_buffer_desc *rqbd;
125 for (i = 0; i < svc->srv_nbuf_per_group; i++) {
126 /* NB: another thread might be doing this as well, we need to
127 * make sure that it wouldn't over-allocate, see LU-1212. */
128 if (svc->srv_nrqbd_receiving >= svc->srv_nbuf_per_group)
131 rqbd = ptlrpc_alloc_rqbd(svc);
134 CERROR("%s: Can't allocate request buffer\n",
140 if (ptlrpc_server_post_idle_rqbds(svc) < 0) {
147 "%s: allocate %d new %d-byte reqbufs (%d/%d left), rc = %d\n",
148 svc->srv_name, i, svc->srv_buf_size,
149 svc->srv_nrqbd_receiving, svc->srv_nbufs, rc);
155 * Part of Rep-Ack logic.
156 * Puts a lock and its mode into reply state assotiated to request reply.
159 ptlrpc_save_lock(struct ptlrpc_request *req,
160 struct lustre_handle *lock, int mode, int no_ack)
162 struct ptlrpc_reply_state *rs = req->rq_reply_state;
166 LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
168 if (req->rq_export->exp_disconnected) {
169 ldlm_lock_decref(lock, mode);
171 idx = rs->rs_nlocks++;
172 rs->rs_locks[idx] = *lock;
173 rs->rs_modes[idx] = mode;
174 rs->rs_difficult = 1;
175 rs->rs_no_ack = !!no_ack;
181 #define HRT_RUNNING 0
182 #define HRT_STOPPING 1
184 struct ptlrpc_hr_thread {
185 cfs_spinlock_t hrt_lock;
186 unsigned long hrt_flags;
187 cfs_waitq_t hrt_wait;
188 cfs_list_t hrt_queue;
189 cfs_completion_t hrt_completion;
192 struct ptlrpc_hr_service {
196 struct ptlrpc_hr_thread hr_threads[0];
200 cfs_list_t rsb_replies;
201 struct ptlrpc_service *rsb_svc;
202 unsigned int rsb_n_replies;
206 * A pointer to per-node reply handling service.
208 static struct ptlrpc_hr_service *ptlrpc_hr = NULL;
211 * maximum mumber of replies scheduled in one batch
213 #define MAX_SCHEDULED 256
216 * Initialize a reply batch.
220 static void rs_batch_init(struct rs_batch *b)
222 memset(b, 0, sizeof *b);
223 CFS_INIT_LIST_HEAD(&b->rsb_replies);
227 * Choose an hr thread to dispatch requests to.
229 static unsigned int get_hr_thread_index(struct ptlrpc_hr_service *hr)
233 /* Concurrent modification of hr_index w/o any spinlock
234 protection is harmless as long as the result fits
235 [0..(hr_n_threads-1)] range and each thread gets near equal
238 hr->hr_index = (idx >= hr->hr_n_threads - 1) ? 0 : idx + 1;
243 * Dispatch all replies accumulated in the batch to one from
244 * dedicated reply handling threads.
248 static void rs_batch_dispatch(struct rs_batch *b)
250 if (b->rsb_n_replies != 0) {
251 struct ptlrpc_hr_service *hr = ptlrpc_hr;
254 idx = get_hr_thread_index(hr);
256 cfs_spin_lock(&hr->hr_threads[idx].hrt_lock);
257 cfs_list_splice_init(&b->rsb_replies,
258 &hr->hr_threads[idx].hrt_queue);
259 cfs_spin_unlock(&hr->hr_threads[idx].hrt_lock);
260 cfs_waitq_signal(&hr->hr_threads[idx].hrt_wait);
261 b->rsb_n_replies = 0;
266 * Add a reply to a batch.
267 * Add one reply object to a batch, schedule batched replies if overload.
272 static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs)
274 struct ptlrpc_service *svc = rs->rs_service;
276 if (svc != b->rsb_svc || b->rsb_n_replies >= MAX_SCHEDULED) {
277 if (b->rsb_svc != NULL) {
278 rs_batch_dispatch(b);
279 cfs_spin_unlock(&b->rsb_svc->srv_rs_lock);
281 cfs_spin_lock(&svc->srv_rs_lock);
284 cfs_spin_lock(&rs->rs_lock);
285 rs->rs_scheduled_ever = 1;
286 if (rs->rs_scheduled == 0) {
287 cfs_list_move(&rs->rs_list, &b->rsb_replies);
288 rs->rs_scheduled = 1;
291 rs->rs_committed = 1;
292 cfs_spin_unlock(&rs->rs_lock);
296 * Reply batch finalization.
297 * Dispatch remaining replies from the batch
298 * and release remaining spinlock.
302 static void rs_batch_fini(struct rs_batch *b)
304 if (b->rsb_svc != 0) {
305 rs_batch_dispatch(b);
306 cfs_spin_unlock(&b->rsb_svc->srv_rs_lock);
310 #define DECLARE_RS_BATCH(b) struct rs_batch b
312 #else /* __KERNEL__ */
314 #define rs_batch_init(b) do{}while(0)
315 #define rs_batch_fini(b) do{}while(0)
316 #define rs_batch_add(b, r) ptlrpc_schedule_difficult_reply(r)
317 #define DECLARE_RS_BATCH(b)
319 #endif /* __KERNEL__ */
322 * Put reply state into a queue for processing because we received
323 * ACK from the client
325 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
328 struct ptlrpc_hr_service *hr = ptlrpc_hr;
332 LASSERT(cfs_list_empty(&rs->rs_list));
334 idx = get_hr_thread_index(hr);
335 cfs_spin_lock(&hr->hr_threads[idx].hrt_lock);
336 cfs_list_add_tail(&rs->rs_list, &hr->hr_threads[idx].hrt_queue);
337 cfs_spin_unlock(&hr->hr_threads[idx].hrt_lock);
338 cfs_waitq_signal(&hr->hr_threads[idx].hrt_wait);
341 cfs_list_add_tail(&rs->rs_list, &rs->rs_service->srv_reply_queue);
346 ptlrpc_schedule_difficult_reply (struct ptlrpc_reply_state *rs)
350 LASSERT_SPIN_LOCKED(&rs->rs_service->srv_rs_lock);
351 LASSERT_SPIN_LOCKED(&rs->rs_lock);
352 LASSERT (rs->rs_difficult);
353 rs->rs_scheduled_ever = 1; /* flag any notification attempt */
355 if (rs->rs_scheduled) { /* being set up or already notified */
360 rs->rs_scheduled = 1;
361 cfs_list_del_init(&rs->rs_list);
362 ptlrpc_dispatch_difficult_reply(rs);
366 void ptlrpc_commit_replies(struct obd_export *exp)
368 struct ptlrpc_reply_state *rs, *nxt;
369 DECLARE_RS_BATCH(batch);
372 rs_batch_init(&batch);
373 /* Find any replies that have been committed and get their service
374 * to attend to complete them. */
376 /* CAVEAT EMPTOR: spinlock ordering!!! */
377 cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
378 cfs_list_for_each_entry_safe(rs, nxt, &exp->exp_uncommitted_replies,
380 LASSERT (rs->rs_difficult);
381 /* VBR: per-export last_committed */
382 LASSERT(rs->rs_export);
383 if (rs->rs_transno <= exp->exp_last_committed) {
384 cfs_list_del_init(&rs->rs_obd_list);
385 rs_batch_add(&batch, rs);
388 cfs_spin_unlock(&exp->exp_uncommitted_replies_lock);
389 rs_batch_fini(&batch);
394 ptlrpc_server_post_idle_rqbds (struct ptlrpc_service *svc)
396 struct ptlrpc_request_buffer_desc *rqbd;
401 cfs_spin_lock(&svc->srv_lock);
403 if (cfs_list_empty (&svc->srv_idle_rqbds)) {
404 cfs_spin_unlock(&svc->srv_lock);
408 rqbd = cfs_list_entry(svc->srv_idle_rqbds.next,
409 struct ptlrpc_request_buffer_desc,
411 cfs_list_del (&rqbd->rqbd_list);
413 /* assume we will post successfully */
414 svc->srv_nrqbd_receiving++;
415 cfs_list_add (&rqbd->rqbd_list, &svc->srv_active_rqbds);
417 cfs_spin_unlock(&svc->srv_lock);
419 rc = ptlrpc_register_rqbd(rqbd);
426 cfs_spin_lock(&svc->srv_lock);
428 svc->srv_nrqbd_receiving--;
429 cfs_list_del(&rqbd->rqbd_list);
430 cfs_list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
432 /* Don't complain if no request buffers are posted right now; LNET
433 * won't drop requests because we set the portal lazy! */
435 cfs_spin_unlock(&svc->srv_lock);
440 static void ptlrpc_at_timer(unsigned long castmeharder)
442 struct ptlrpc_service *svc = (struct ptlrpc_service *)castmeharder;
443 svc->srv_at_check = 1;
444 svc->srv_at_checktime = cfs_time_current();
445 cfs_waitq_signal(&svc->srv_waitq);
449 ptlrpc_server_nthreads_check(struct ptlrpc_service_conf *conf,
450 int *min_p, int *max_p)
453 struct ptlrpc_service_thr_conf *tc = &conf->psc_thr;
457 nthrs_min = PTLRPC_NTHRS_MIN + (conf->psc_ops.so_hpreq_handler != NULL);
458 nthrs_min = max_t(int, nthrs_min, tc->tc_nthrs_min);
460 nthrs = tc->tc_nthrs_user;
461 if (nthrs != 0) { /* validate it */
462 nthrs = min_t(int, nthrs, tc->tc_nthrs_max);
463 nthrs = max_t(int, nthrs, nthrs_min);
464 *min_p = *max_p = nthrs;
469 * NB: we will add some common at here for estimating, for example:
470 * add a new member ptlrpc_service_thr_conf::tc_factor, and estimate
471 * threads number based on:
472 * (online_cpus * conf::tc_factor) + conf::tc_nthrs_base.
474 * So we can remove code block like estimation in ost_setup, also,
475 * we might estimate MDS threads number as well instead of using
476 * absolute number, and have more threads on fat servers to improve
477 * availability of service.
479 * Also, we will need to validate threads number at here for
480 * CPT affinity service (CPU ParTiion) in the future.
481 * A service can have percpt thread-pool instead of a global thread
482 * pool for each service, which means user might not always get the
483 * threads number they want even they set it in conf::tc_nthrs_user,
484 * because we need to adjust threads number for each CPT, instead of
485 * just use (conf::tc_nthrs_user / NCPTS), to make sure each pool
488 *max_p = tc->tc_nthrs_max;
490 #else /* __KERNEL__ */
491 *max_p = *min_p = 1; /* whatever */
496 * Initialize service on a given portal.
497 * This includes starting serving threads , allocating and posting rqbds and
500 struct ptlrpc_service *
501 ptlrpc_register_service(struct ptlrpc_service_conf *conf,
502 cfs_proc_dir_entry_t *proc_entry)
504 struct ptlrpc_service *service;
505 struct ptlrpc_at_array *array;
511 LASSERT(conf->psc_buf.bc_nbufs > 0);
512 LASSERT(conf->psc_buf.bc_buf_size >=
513 conf->psc_buf.bc_req_max_size + SPTLRPC_MAX_PAYLOAD);
514 LASSERT(conf->psc_thr.tc_ctx_tags != 0);
516 OBD_ALLOC_PTR(service);
518 RETURN(ERR_PTR(-ENOMEM));
520 /* First initialise enough for early teardown */
522 cfs_spin_lock_init(&service->srv_lock);
523 cfs_spin_lock_init(&service->srv_rq_lock);
524 cfs_spin_lock_init(&service->srv_rs_lock);
525 CFS_INIT_LIST_HEAD(&service->srv_threads);
526 cfs_waitq_init(&service->srv_waitq);
528 service->srv_name = conf->psc_name;
529 service->srv_watchdog_factor = conf->psc_watchdog_factor;
530 service->srv_nbuf_per_group = test_req_buffer_pressure ?
531 1 : conf->psc_buf.bc_nbufs;
532 service->srv_max_req_size = conf->psc_buf.bc_req_max_size +
534 service->srv_buf_size = conf->psc_buf.bc_buf_size;
535 service->srv_rep_portal = conf->psc_buf.bc_rep_portal;
536 service->srv_req_portal = conf->psc_buf.bc_req_portal;
537 service->srv_request_seq = 1; /* valid seq #s start at 1 */
538 service->srv_request_max_cull_seq = 0;
540 ptlrpc_server_nthreads_check(conf, &service->srv_threads_min,
541 &service->srv_threads_max);
543 service->srv_thread_name = conf->psc_thr.tc_thr_name;
544 service->srv_ctx_tags = conf->psc_thr.tc_ctx_tags;
545 service->srv_cpu_affinity = !!conf->psc_thr.tc_cpu_affinity;
546 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
547 service->srv_hpreq_count = 0;
548 service->srv_n_active_hpreq = 0;
549 service->srv_ops = conf->psc_ops;
551 rc = LNetSetLazyPortal(service->srv_req_portal);
554 CFS_INIT_LIST_HEAD(&service->srv_request_queue);
555 CFS_INIT_LIST_HEAD(&service->srv_request_hpq);
556 CFS_INIT_LIST_HEAD(&service->srv_idle_rqbds);
557 CFS_INIT_LIST_HEAD(&service->srv_active_rqbds);
558 CFS_INIT_LIST_HEAD(&service->srv_history_rqbds);
559 CFS_INIT_LIST_HEAD(&service->srv_request_history);
560 CFS_INIT_LIST_HEAD(&service->srv_active_replies);
562 CFS_INIT_LIST_HEAD(&service->srv_reply_queue);
564 CFS_INIT_LIST_HEAD(&service->srv_free_rs_list);
565 cfs_waitq_init(&service->srv_free_rs_waitq);
566 cfs_atomic_set(&service->srv_n_difficult_replies, 0);
568 cfs_spin_lock_init(&service->srv_at_lock);
569 CFS_INIT_LIST_HEAD(&service->srv_req_in_queue);
571 array = &service->srv_at_array;
572 size = at_est2timeout(at_max);
573 array->paa_size = size;
574 array->paa_count = 0;
575 array->paa_deadline = -1;
577 /* allocate memory for srv_at_array (ptlrpc_at_array) */
578 OBD_ALLOC(array->paa_reqs_array, sizeof(cfs_list_t) * size);
579 if (array->paa_reqs_array == NULL)
580 GOTO(failed, rc = -ENOMEM);
582 for (index = 0; index < size; index++)
583 CFS_INIT_LIST_HEAD(&array->paa_reqs_array[index]);
585 OBD_ALLOC(array->paa_reqs_count, sizeof(__u32) * size);
586 if (array->paa_reqs_count == NULL)
587 GOTO(failed, rc = -ENOMEM);
589 cfs_timer_init(&service->srv_at_timer, ptlrpc_at_timer, service);
590 /* At SOW, service time should be quick; 10s seems generous. If client
591 timeout is less than this, we'll be sending an early reply. */
592 at_init(&service->srv_at_estimate, 10, 0);
594 cfs_spin_lock (&ptlrpc_all_services_lock);
595 cfs_list_add (&service->srv_list, &ptlrpc_all_services);
596 cfs_spin_unlock (&ptlrpc_all_services_lock);
598 /* Now allocate the request buffers */
599 rc = ptlrpc_grow_req_bufs(service);
600 /* We shouldn't be under memory pressure at startup, so
601 * fail if we can't post all our buffers at this time. */
603 GOTO(failed, rc = -ENOMEM);
605 /* Now allocate pool of reply buffers */
606 /* Increase max reply size to next power of two */
607 service->srv_max_reply_size = 1;
608 while (service->srv_max_reply_size <
609 conf->psc_buf.bc_rep_max_size + SPTLRPC_MAX_PAYLOAD)
610 service->srv_max_reply_size <<= 1;
612 if (proc_entry != NULL)
613 ptlrpc_lprocfs_register_service(proc_entry, service);
615 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
616 service->srv_name, service->srv_req_portal);
619 rc = ptlrpc_start_threads(service);
621 CERROR("Failed to start threads for service %s: %d\n",
622 service->srv_name, rc);
629 ptlrpc_unregister_service(service);
634 * to actually free the request, must be called without holding svc_lock.
635 * note it's caller's responsibility to unlink req->rq_list.
637 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
639 LASSERT(cfs_atomic_read(&req->rq_refcount) == 0);
640 LASSERT(cfs_list_empty(&req->rq_timed_list));
642 /* DEBUG_REQ() assumes the reply state of a request with a valid
643 * ref will not be destroyed until that reference is dropped. */
644 ptlrpc_req_drop_rs(req);
646 sptlrpc_svc_ctx_decref(req);
648 if (req != &req->rq_rqbd->rqbd_req) {
649 /* NB request buffers use an embedded
650 * req if the incoming req unlinked the
651 * MD; this isn't one of them! */
652 OBD_FREE(req, sizeof(*req));
657 * drop a reference count of the request. if it reaches 0, we either
658 * put it into history list, or free it immediately.
660 void ptlrpc_server_drop_request(struct ptlrpc_request *req)
662 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
663 struct ptlrpc_service *svc = rqbd->rqbd_service;
668 if (!cfs_atomic_dec_and_test(&req->rq_refcount))
671 cfs_spin_lock(&svc->srv_at_lock);
672 if (req->rq_at_linked) {
673 struct ptlrpc_at_array *array = &svc->srv_at_array;
674 __u32 index = req->rq_at_index;
676 LASSERT(!cfs_list_empty(&req->rq_timed_list));
677 cfs_list_del_init(&req->rq_timed_list);
678 cfs_spin_lock(&req->rq_lock);
679 req->rq_at_linked = 0;
680 cfs_spin_unlock(&req->rq_lock);
681 array->paa_reqs_count[index]--;
684 LASSERT(cfs_list_empty(&req->rq_timed_list));
685 cfs_spin_unlock(&svc->srv_at_lock);
687 /* finalize request */
688 if (req->rq_export) {
689 class_export_put(req->rq_export);
690 req->rq_export = NULL;
693 cfs_spin_lock(&svc->srv_lock);
695 cfs_list_add(&req->rq_list, &rqbd->rqbd_reqs);
697 refcount = --(rqbd->rqbd_refcount);
699 /* request buffer is now idle: add to history */
700 cfs_list_del(&rqbd->rqbd_list);
701 cfs_list_add_tail(&rqbd->rqbd_list, &svc->srv_history_rqbds);
702 svc->srv_n_history_rqbds++;
704 /* cull some history?
705 * I expect only about 1 or 2 rqbds need to be recycled here */
706 while (svc->srv_n_history_rqbds > svc->srv_max_history_rqbds) {
707 rqbd = cfs_list_entry(svc->srv_history_rqbds.next,
708 struct ptlrpc_request_buffer_desc,
711 cfs_list_del(&rqbd->rqbd_list);
712 svc->srv_n_history_rqbds--;
714 /* remove rqbd's reqs from svc's req history while
715 * I've got the service lock */
716 cfs_list_for_each(tmp, &rqbd->rqbd_reqs) {
717 req = cfs_list_entry(tmp, struct ptlrpc_request,
719 /* Track the highest culled req seq */
720 if (req->rq_history_seq >
721 svc->srv_request_max_cull_seq)
722 svc->srv_request_max_cull_seq =
724 cfs_list_del(&req->rq_history_list);
727 cfs_spin_unlock(&svc->srv_lock);
729 cfs_list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
730 req = cfs_list_entry(rqbd->rqbd_reqs.next,
731 struct ptlrpc_request,
733 cfs_list_del(&req->rq_list);
734 ptlrpc_server_free_request(req);
737 cfs_spin_lock(&svc->srv_lock);
739 * now all reqs including the embedded req has been
740 * disposed, schedule request buffer for re-use.
742 LASSERT(cfs_atomic_read(&rqbd->rqbd_req.rq_refcount) ==
744 cfs_list_add_tail(&rqbd->rqbd_list,
745 &svc->srv_idle_rqbds);
748 cfs_spin_unlock(&svc->srv_lock);
749 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
750 /* If we are low on memory, we are not interested in history */
751 cfs_list_del(&req->rq_list);
752 cfs_list_del_init(&req->rq_history_list);
753 cfs_spin_unlock(&svc->srv_lock);
755 ptlrpc_server_free_request(req);
757 cfs_spin_unlock(&svc->srv_lock);
762 * to finish a request: stop sending more early replies, and release
763 * the request. should be called after we finished handling the request.
765 static void ptlrpc_server_finish_request(struct ptlrpc_service *svc,
766 struct ptlrpc_request *req)
768 ptlrpc_hpreq_fini(req);
770 cfs_spin_lock(&svc->srv_rq_lock);
771 svc->srv_n_active_reqs--;
773 svc->srv_n_active_hpreq--;
774 cfs_spin_unlock(&svc->srv_rq_lock);
776 ptlrpc_server_drop_request(req);
780 * This function makes sure dead exports are evicted in a timely manner.
781 * This function is only called when some export receives a message (i.e.,
782 * the network is up.)
784 static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
786 struct obd_export *oldest_exp;
787 time_t oldest_time, new_time;
793 /* Compensate for slow machines, etc, by faking our request time
794 into the future. Although this can break the strict time-ordering
795 of the list, we can be really lazy here - we don't have to evict
796 at the exact right moment. Eventually, all silent exports
797 will make it to the top of the list. */
799 /* Do not pay attention on 1sec or smaller renewals. */
800 new_time = cfs_time_current_sec() + extra_delay;
801 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
804 exp->exp_last_request_time = new_time;
805 CDEBUG(D_HA, "updating export %s at "CFS_TIME_T" exp %p\n",
806 exp->exp_client_uuid.uuid,
807 exp->exp_last_request_time, exp);
809 /* exports may get disconnected from the chain even though the
810 export has references, so we must keep the spin lock while
811 manipulating the lists */
812 cfs_spin_lock(&exp->exp_obd->obd_dev_lock);
814 if (cfs_list_empty(&exp->exp_obd_chain_timed)) {
815 /* this one is not timed */
816 cfs_spin_unlock(&exp->exp_obd->obd_dev_lock);
820 cfs_list_move_tail(&exp->exp_obd_chain_timed,
821 &exp->exp_obd->obd_exports_timed);
823 oldest_exp = cfs_list_entry(exp->exp_obd->obd_exports_timed.next,
824 struct obd_export, exp_obd_chain_timed);
825 oldest_time = oldest_exp->exp_last_request_time;
826 cfs_spin_unlock(&exp->exp_obd->obd_dev_lock);
828 if (exp->exp_obd->obd_recovering) {
829 /* be nice to everyone during recovery */
834 /* Note - racing to start/reset the obd_eviction timer is safe */
835 if (exp->exp_obd->obd_eviction_timer == 0) {
836 /* Check if the oldest entry is expired. */
837 if (cfs_time_current_sec() > (oldest_time + PING_EVICT_TIMEOUT +
839 /* We need a second timer, in case the net was down and
840 * it just came back. Since the pinger may skip every
841 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
842 * we better wait for 3. */
843 exp->exp_obd->obd_eviction_timer =
844 cfs_time_current_sec() + 3 * PING_INTERVAL;
845 CDEBUG(D_HA, "%s: Think about evicting %s from "CFS_TIME_T"\n",
846 exp->exp_obd->obd_name,
847 obd_export_nid2str(oldest_exp), oldest_time);
850 if (cfs_time_current_sec() >
851 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
852 /* The evictor won't evict anyone who we've heard from
853 * recently, so we don't have to check before we start
855 if (!ping_evictor_wake(exp))
856 exp->exp_obd->obd_eviction_timer = 0;
864 * Sanity check request \a req.
865 * Return 0 if all is ok, error code otherwise.
867 static int ptlrpc_check_req(struct ptlrpc_request *req)
871 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
872 req->rq_export->exp_conn_cnt)) {
873 DEBUG_REQ(D_ERROR, req,
874 "DROPPING req from old connection %d < %d",
875 lustre_msg_get_conn_cnt(req->rq_reqmsg),
876 req->rq_export->exp_conn_cnt);
879 if (unlikely(req->rq_export->exp_obd &&
880 req->rq_export->exp_obd->obd_fail)) {
881 /* Failing over, don't handle any more reqs, send
882 error response instead. */
883 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
884 req, req->rq_export->exp_obd->obd_name);
886 } else if (lustre_msg_get_flags(req->rq_reqmsg) &
887 (MSG_REPLAY | MSG_REQ_REPLAY_DONE) &&
888 !(req->rq_export->exp_obd->obd_recovering)) {
889 DEBUG_REQ(D_ERROR, req,
890 "Invalid replay without recovery");
891 class_fail_export(req->rq_export);
893 } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 &&
894 !(req->rq_export->exp_obd->obd_recovering)) {
895 DEBUG_REQ(D_ERROR, req, "Invalid req with transno "
896 LPU64" without recovery",
897 lustre_msg_get_transno(req->rq_reqmsg));
898 class_fail_export(req->rq_export);
902 if (unlikely(rc < 0)) {
909 static void ptlrpc_at_set_timer(struct ptlrpc_service *svc)
911 struct ptlrpc_at_array *array = &svc->srv_at_array;
914 cfs_spin_lock(&svc->srv_at_lock);
915 if (array->paa_count == 0) {
916 cfs_timer_disarm(&svc->srv_at_timer);
917 cfs_spin_unlock(&svc->srv_at_lock);
921 /* Set timer for closest deadline */
922 next = (__s32)(array->paa_deadline - cfs_time_current_sec() -
925 ptlrpc_at_timer((unsigned long)svc);
927 cfs_timer_arm(&svc->srv_at_timer, cfs_time_shift(next));
928 cfs_spin_unlock(&svc->srv_at_lock);
929 CDEBUG(D_INFO, "armed %s at %+ds\n", svc->srv_name, next);
932 /* Add rpc to early reply check list */
933 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
935 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
936 struct ptlrpc_request *rq = NULL;
937 struct ptlrpc_at_array *array = &svc->srv_at_array;
944 if (req->rq_no_reply)
947 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
950 cfs_spin_lock(&svc->srv_at_lock);
951 LASSERT(cfs_list_empty(&req->rq_timed_list));
953 index = (unsigned long)req->rq_deadline % array->paa_size;
954 if (array->paa_reqs_count[index] > 0) {
955 /* latest rpcs will have the latest deadlines in the list,
956 * so search backward. */
957 cfs_list_for_each_entry_reverse(rq,
958 &array->paa_reqs_array[index],
960 if (req->rq_deadline >= rq->rq_deadline) {
961 cfs_list_add(&req->rq_timed_list,
968 /* Add the request at the head of the list */
969 if (cfs_list_empty(&req->rq_timed_list))
970 cfs_list_add(&req->rq_timed_list,
971 &array->paa_reqs_array[index]);
973 cfs_spin_lock(&req->rq_lock);
974 req->rq_at_linked = 1;
975 cfs_spin_unlock(&req->rq_lock);
976 req->rq_at_index = index;
977 array->paa_reqs_count[index]++;
979 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
980 array->paa_deadline = req->rq_deadline;
983 cfs_spin_unlock(&svc->srv_at_lock);
986 ptlrpc_at_set_timer(svc);
991 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
993 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
994 struct ptlrpc_request *reqcopy;
995 struct lustre_msg *reqmsg;
996 cfs_duration_t olddl = req->rq_deadline - cfs_time_current_sec();
1001 /* deadline is when the client expects us to reply, margin is the
1002 difference between clients' and servers' expectations */
1003 DEBUG_REQ(D_ADAPTTO, req,
1004 "%ssending early reply (deadline %+lds, margin %+lds) for "
1005 "%d+%d", AT_OFF ? "AT off - not " : "",
1006 olddl, olddl - at_get(&svc->srv_at_estimate),
1007 at_get(&svc->srv_at_estimate), at_extra);
1013 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), "
1014 "not sending early reply. Consider increasing "
1015 "at_early_margin (%d)?", olddl, at_early_margin);
1017 /* Return an error so we're not re-added to the timed list. */
1021 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
1022 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
1023 "but no AT support");
1027 if (req->rq_export &&
1028 lustre_msg_get_flags(req->rq_reqmsg) &
1029 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1030 /* During recovery, we don't want to send too many early
1031 * replies, but on the other hand we want to make sure the
1032 * client has enough time to resend if the rpc is lost. So
1033 * during the recovery period send at least 4 early replies,
1034 * spacing them every at_extra if we can. at_estimate should
1035 * always equal this fixed value during recovery. */
1036 at_measured(&svc->srv_at_estimate, min(at_extra,
1037 req->rq_export->exp_obd->obd_recovery_timeout / 4));
1039 /* Fake our processing time into the future to ask the clients
1040 * for some extra amount of time */
1041 at_measured(&svc->srv_at_estimate, at_extra +
1042 cfs_time_current_sec() -
1043 req->rq_arrival_time.tv_sec);
1045 /* Check to see if we've actually increased the deadline -
1046 * we may be past adaptive_max */
1047 if (req->rq_deadline >= req->rq_arrival_time.tv_sec +
1048 at_get(&svc->srv_at_estimate)) {
1049 DEBUG_REQ(D_WARNING, req, "Couldn't add any time "
1050 "(%ld/%ld), not sending early reply\n",
1051 olddl, req->rq_arrival_time.tv_sec +
1052 at_get(&svc->srv_at_estimate) -
1053 cfs_time_current_sec());
1057 newdl = cfs_time_current_sec() + at_get(&svc->srv_at_estimate);
1059 OBD_ALLOC(reqcopy, sizeof *reqcopy);
1060 if (reqcopy == NULL)
1062 OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1064 OBD_FREE(reqcopy, sizeof *reqcopy);
1069 reqcopy->rq_reply_state = NULL;
1070 reqcopy->rq_rep_swab_mask = 0;
1071 reqcopy->rq_pack_bulk = 0;
1072 reqcopy->rq_pack_udesc = 0;
1073 reqcopy->rq_packed_final = 0;
1074 sptlrpc_svc_ctx_addref(reqcopy);
1075 /* We only need the reqmsg for the magic */
1076 reqcopy->rq_reqmsg = reqmsg;
1077 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1079 LASSERT(cfs_atomic_read(&req->rq_refcount));
1080 /** if it is last refcount then early reply isn't needed */
1081 if (cfs_atomic_read(&req->rq_refcount) == 1) {
1082 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
1083 "abort sending early reply\n");
1084 GOTO(out, rc = -EINVAL);
1087 /* Connection ref */
1088 reqcopy->rq_export = class_conn2export(
1089 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1090 if (reqcopy->rq_export == NULL)
1091 GOTO(out, rc = -ENODEV);
1094 class_export_rpc_get(reqcopy->rq_export);
1095 if (reqcopy->rq_export->exp_obd &&
1096 reqcopy->rq_export->exp_obd->obd_fail)
1097 GOTO(out_put, rc = -ENODEV);
1099 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1103 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1106 /* Adjust our own deadline to what we told the client */
1107 req->rq_deadline = newdl;
1108 req->rq_early_count++; /* number sent, server side */
1110 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1113 /* Free the (early) reply state from lustre_pack_reply.
1114 (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
1115 ptlrpc_req_drop_rs(reqcopy);
1118 class_export_rpc_put(reqcopy->rq_export);
1119 class_export_put(reqcopy->rq_export);
1121 sptlrpc_svc_ctx_decref(reqcopy);
1122 OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1123 OBD_FREE(reqcopy, sizeof *reqcopy);
1127 /* Send early replies to everybody expiring within at_early_margin
1128 asking for at_extra time */
1129 static int ptlrpc_at_check_timed(struct ptlrpc_service *svc)
1131 struct ptlrpc_request *rq, *n;
1132 cfs_list_t work_list;
1133 struct ptlrpc_at_array *array = &svc->srv_at_array;
1136 time_t now = cfs_time_current_sec();
1137 cfs_duration_t delay;
1138 int first, counter = 0;
1141 cfs_spin_lock(&svc->srv_at_lock);
1142 if (svc->srv_at_check == 0) {
1143 cfs_spin_unlock(&svc->srv_at_lock);
1146 delay = cfs_time_sub(cfs_time_current(), svc->srv_at_checktime);
1147 svc->srv_at_check = 0;
1149 if (array->paa_count == 0) {
1150 cfs_spin_unlock(&svc->srv_at_lock);
1154 /* The timer went off, but maybe the nearest rpc already completed. */
1155 first = array->paa_deadline - now;
1156 if (first > at_early_margin) {
1157 /* We've still got plenty of time. Reset the timer. */
1158 cfs_spin_unlock(&svc->srv_at_lock);
1159 ptlrpc_at_set_timer(svc);
1163 /* We're close to a timeout, and we don't know how much longer the
1164 server will take. Send early replies to everyone expiring soon. */
1165 CFS_INIT_LIST_HEAD(&work_list);
1167 index = (unsigned long)array->paa_deadline % array->paa_size;
1168 count = array->paa_count;
1170 count -= array->paa_reqs_count[index];
1171 cfs_list_for_each_entry_safe(rq, n,
1172 &array->paa_reqs_array[index],
1174 if (rq->rq_deadline <= now + at_early_margin) {
1175 cfs_list_del_init(&rq->rq_timed_list);
1177 * ptlrpc_server_drop_request() may drop
1178 * refcount to 0 already. Let's check this and
1179 * don't add entry to work_list
1181 if (likely(cfs_atomic_inc_not_zero(&rq->rq_refcount)))
1182 cfs_list_add(&rq->rq_timed_list, &work_list);
1184 array->paa_reqs_count[index]--;
1186 cfs_spin_lock(&rq->rq_lock);
1187 rq->rq_at_linked = 0;
1188 cfs_spin_unlock(&rq->rq_lock);
1192 /* update the earliest deadline */
1193 if (deadline == -1 || rq->rq_deadline < deadline)
1194 deadline = rq->rq_deadline;
1199 if (++index >= array->paa_size)
1202 array->paa_deadline = deadline;
1203 cfs_spin_unlock(&svc->srv_at_lock);
1205 /* we have a new earliest deadline, restart the timer */
1206 ptlrpc_at_set_timer(svc);
1208 CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
1209 "replies\n", first, at_extra, counter);
1211 /* We're already past request deadlines before we even get a
1212 chance to send early replies */
1213 LCONSOLE_WARN("%s: This server is not able to keep up with "
1214 "request traffic (cpu-bound).\n", svc->srv_name);
1215 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, "
1216 "delay="CFS_DURATION_T"(jiff)\n",
1217 counter, svc->srv_n_queued_reqs, svc->srv_n_active_reqs,
1218 at_get(&svc->srv_at_estimate), delay);
1221 /* we took additional refcount so entries can't be deleted from list, no
1222 * locking is needed */
1223 while (!cfs_list_empty(&work_list)) {
1224 rq = cfs_list_entry(work_list.next, struct ptlrpc_request,
1226 cfs_list_del_init(&rq->rq_timed_list);
1228 if (ptlrpc_at_send_early_reply(rq) == 0)
1229 ptlrpc_at_add_timed(rq);
1231 ptlrpc_server_drop_request(rq);
1238 * Put the request to the export list if the request may become
1239 * a high priority one.
1241 static int ptlrpc_hpreq_init(struct ptlrpc_service *svc,
1242 struct ptlrpc_request *req)
1247 if (svc->srv_ops.so_hpreq_handler) {
1248 rc = svc->srv_ops.so_hpreq_handler(req);
1252 if (req->rq_export && req->rq_ops) {
1253 /* Perform request specific check. We should do this check
1254 * before the request is added into exp_hp_rpcs list otherwise
1255 * it may hit swab race at LU-1044. */
1256 if (req->rq_ops->hpreq_check)
1257 rc = req->rq_ops->hpreq_check(req);
1259 cfs_spin_lock_bh(&req->rq_export->exp_rpc_lock);
1260 cfs_list_add(&req->rq_exp_list,
1261 &req->rq_export->exp_hp_rpcs);
1262 cfs_spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1268 /** Remove the request from the export list. */
1269 static void ptlrpc_hpreq_fini(struct ptlrpc_request *req)
1272 if (req->rq_export && req->rq_ops) {
1273 /* refresh lock timeout again so that client has more
1274 * room to send lock cancel RPC. */
1275 if (req->rq_ops->hpreq_fini)
1276 req->rq_ops->hpreq_fini(req);
1278 cfs_spin_lock_bh(&req->rq_export->exp_rpc_lock);
1279 cfs_list_del_init(&req->rq_exp_list);
1280 cfs_spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1286 * Make the request a high priority one.
1288 * All the high priority requests are queued in a separate FIFO
1289 * ptlrpc_service::srv_request_hpq list which is parallel to
1290 * ptlrpc_service::srv_request_queue list but has a higher priority
1293 * \see ptlrpc_server_handle_request().
1295 static void ptlrpc_hpreq_reorder_nolock(struct ptlrpc_service *svc,
1296 struct ptlrpc_request *req)
1299 LASSERT(svc != NULL);
1300 cfs_spin_lock(&req->rq_lock);
1301 if (req->rq_hp == 0) {
1302 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1304 /* Add to the high priority queue. */
1305 cfs_list_move_tail(&req->rq_list, &svc->srv_request_hpq);
1307 if (opc != OBD_PING)
1308 DEBUG_REQ(D_RPCTRACE, req, "high priority req");
1310 cfs_spin_unlock(&req->rq_lock);
1315 * \see ptlrpc_hpreq_reorder_nolock
1317 void ptlrpc_hpreq_reorder(struct ptlrpc_request *req)
1319 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
1322 cfs_spin_lock(&svc->srv_rq_lock);
1323 /* It may happen that the request is already taken for the processing
1324 * but still in the export list, or the request is not in the request
1325 * queue but in the export list already, do not add it into the
1327 if (!cfs_list_empty(&req->rq_list))
1328 ptlrpc_hpreq_reorder_nolock(svc, req);
1329 cfs_spin_unlock(&svc->srv_rq_lock);
1333 /** Check if the request is a high priority one. */
1334 static int ptlrpc_server_hpreq_check(struct ptlrpc_service *svc,
1335 struct ptlrpc_request *req)
1339 /* Check by request opc. */
1340 if (OBD_PING == lustre_msg_get_opc(req->rq_reqmsg))
1343 RETURN(ptlrpc_hpreq_init(svc, req));
1346 /** Check if a request is a high priority one. */
1347 static int ptlrpc_server_request_add(struct ptlrpc_service *svc,
1348 struct ptlrpc_request *req)
1353 rc = ptlrpc_server_hpreq_check(svc, req);
1357 cfs_spin_lock(&svc->srv_rq_lock);
1360 ptlrpc_hpreq_reorder_nolock(svc, req);
1362 cfs_list_add_tail(&req->rq_list,
1363 &svc->srv_request_queue);
1365 cfs_spin_unlock(&svc->srv_rq_lock);
1371 * Allow to handle high priority request
1372 * User can call it w/o any lock but need to hold ptlrpc_service::srv_rq_lock
1373 * to get reliable result
1375 static int ptlrpc_server_allow_high(struct ptlrpc_service *svc, int force)
1380 if (svc->srv_n_active_reqs >= svc->srv_threads_running - 1)
1383 return cfs_list_empty(&svc->srv_request_queue) ||
1384 svc->srv_hpreq_count < svc->srv_hpreq_ratio;
1387 static int ptlrpc_server_high_pending(struct ptlrpc_service *svc, int force)
1389 return ptlrpc_server_allow_high(svc, force) &&
1390 !cfs_list_empty(&svc->srv_request_hpq);
1394 * Only allow normal priority requests on a service that has a high-priority
1395 * queue if forced (i.e. cleanup), if there are other high priority requests
1396 * already being processed (i.e. those threads can service more high-priority
1397 * requests), or if there are enough idle threads that a later thread can do
1398 * a high priority request.
1399 * User can call it w/o any lock but need to hold ptlrpc_service::srv_rq_lock
1400 * to get reliable result
1402 static int ptlrpc_server_allow_normal(struct ptlrpc_service *svc, int force)
1405 if (1) /* always allow to handle normal request for liblustre */
1409 svc->srv_n_active_reqs < svc->srv_threads_running - 2)
1412 if (svc->srv_n_active_reqs >= svc->srv_threads_running - 1)
1415 return svc->srv_n_active_hpreq > 0 ||
1416 svc->srv_ops.so_hpreq_handler == NULL;
1419 static int ptlrpc_server_normal_pending(struct ptlrpc_service *svc, int force)
1421 return ptlrpc_server_allow_normal(svc, force) &&
1422 !cfs_list_empty(&svc->srv_request_queue);
1426 * Returns true if there are requests available in incoming
1427 * request queue for processing and it is allowed to fetch them.
1428 * User can call it w/o any lock but need to hold ptlrpc_service::srv_rq_lock
1429 * to get reliable result
1430 * \see ptlrpc_server_allow_normal
1431 * \see ptlrpc_server_allow high
1434 ptlrpc_server_request_pending(struct ptlrpc_service *svc, int force)
1436 return ptlrpc_server_high_pending(svc, force) ||
1437 ptlrpc_server_normal_pending(svc, force);
1441 * Fetch a request for processing from queue of unprocessed requests.
1442 * Favors high-priority requests.
1443 * Returns a pointer to fetched request.
1445 static struct ptlrpc_request *
1446 ptlrpc_server_request_get(struct ptlrpc_service *svc, int force)
1448 struct ptlrpc_request *req;
1451 if (ptlrpc_server_high_pending(svc, force)) {
1452 req = cfs_list_entry(svc->srv_request_hpq.next,
1453 struct ptlrpc_request, rq_list);
1454 svc->srv_hpreq_count++;
1459 if (ptlrpc_server_normal_pending(svc, force)) {
1460 req = cfs_list_entry(svc->srv_request_queue.next,
1461 struct ptlrpc_request, rq_list);
1462 svc->srv_hpreq_count = 0;
1469 * Handle freshly incoming reqs, add to timed early reply list,
1470 * pass on to regular request queue.
1471 * All incoming requests pass through here before getting into
1472 * ptlrpc_server_handle_req later on.
1475 ptlrpc_server_handle_req_in(struct ptlrpc_service *svc)
1477 struct ptlrpc_request *req;
1484 cfs_spin_lock(&svc->srv_lock);
1485 if (cfs_list_empty(&svc->srv_req_in_queue)) {
1486 cfs_spin_unlock(&svc->srv_lock);
1490 req = cfs_list_entry(svc->srv_req_in_queue.next,
1491 struct ptlrpc_request, rq_list);
1492 cfs_list_del_init (&req->rq_list);
1493 svc->srv_n_queued_reqs--;
1494 /* Consider this still a "queued" request as far as stats are
1496 cfs_spin_unlock(&svc->srv_lock);
1498 /* go through security check/transform */
1499 rc = sptlrpc_svc_unwrap_request(req);
1503 case SECSVC_COMPLETE:
1504 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1513 * for null-flavored rpc, msg has been unpacked by sptlrpc, although
1514 * redo it wouldn't be harmful.
1516 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
1517 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
1519 CERROR("error unpacking request: ptl %d from %s "
1520 "x"LPU64"\n", svc->srv_req_portal,
1521 libcfs_id2str(req->rq_peer), req->rq_xid);
1526 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1528 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
1529 LPU64"\n", svc->srv_req_portal,
1530 libcfs_id2str(req->rq_peer), req->rq_xid);
1534 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
1535 lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
1536 CERROR("drop incoming rpc opc %u, x"LPU64"\n",
1537 cfs_fail_val, req->rq_xid);
1542 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1543 CERROR("wrong packet type received (type=%u) from %s\n",
1544 lustre_msg_get_type(req->rq_reqmsg),
1545 libcfs_id2str(req->rq_peer));
1549 switch(lustre_msg_get_opc(req->rq_reqmsg)) {
1552 req->rq_bulk_write = 1;
1556 case MGS_CONFIG_READ:
1557 req->rq_bulk_read = 1;
1561 CDEBUG(D_RPCTRACE, "got req x"LPU64"\n", req->rq_xid);
1563 req->rq_export = class_conn2export(
1564 lustre_msg_get_handle(req->rq_reqmsg));
1565 if (req->rq_export) {
1566 rc = ptlrpc_check_req(req);
1568 rc = sptlrpc_target_export_check(req->rq_export, req);
1570 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
1571 "illegal security flavor,");
1576 ptlrpc_update_export_timer(req->rq_export, 0);
1579 /* req_in handling should/must be fast */
1580 if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5)
1581 DEBUG_REQ(D_WARNING, req, "Slow req_in handling "CFS_DURATION_T"s",
1582 cfs_time_sub(cfs_time_current_sec(),
1583 req->rq_arrival_time.tv_sec));
1585 /* Set rpc server deadline and add it to the timed list */
1586 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1587 MSGHDR_AT_SUPPORT) ?
1588 /* The max time the client expects us to take */
1589 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1590 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1591 if (unlikely(deadline == 0)) {
1592 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1596 ptlrpc_at_add_timed(req);
1598 /* Move it over to the request processing queue */
1599 rc = ptlrpc_server_request_add(svc, req);
1601 ptlrpc_hpreq_fini(req);
1604 cfs_waitq_signal(&svc->srv_waitq);
1608 cfs_spin_lock(&svc->srv_rq_lock);
1609 svc->srv_n_active_reqs++;
1610 cfs_spin_unlock(&svc->srv_rq_lock);
1611 ptlrpc_server_finish_request(svc, req);
1617 * Main incoming request handling logic.
1618 * Calls handler function from service to do actual processing.
1621 ptlrpc_server_handle_request(struct ptlrpc_service *svc,
1622 struct ptlrpc_thread *thread)
1624 struct obd_export *export = NULL;
1625 struct ptlrpc_request *request;
1626 struct timeval work_start;
1627 struct timeval work_end;
1635 cfs_spin_lock(&svc->srv_rq_lock);
1637 /* !@%$# liblustre only has 1 thread */
1638 if (cfs_atomic_read(&svc->srv_n_difficult_replies) != 0) {
1639 cfs_spin_unlock(&svc->srv_rq_lock);
1643 request = ptlrpc_server_request_get(svc, 0);
1644 if (request == NULL) {
1645 cfs_spin_unlock(&svc->srv_rq_lock);
1649 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
1650 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
1651 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
1652 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
1654 if (unlikely(fail_opc)) {
1655 if (request->rq_export && request->rq_ops) {
1656 cfs_spin_unlock(&svc->srv_rq_lock);
1657 OBD_FAIL_TIMEOUT(fail_opc, 4);
1658 cfs_spin_lock(&svc->srv_rq_lock);
1659 request = ptlrpc_server_request_get(svc, 0);
1660 if (request == NULL) {
1661 cfs_spin_unlock(&svc->srv_rq_lock);
1667 cfs_list_del_init(&request->rq_list);
1668 svc->srv_n_active_reqs++;
1670 svc->srv_n_active_hpreq++;
1672 cfs_spin_unlock(&svc->srv_rq_lock);
1674 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
1676 if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
1677 libcfs_debug_dumplog();
1679 cfs_gettimeofday(&work_start);
1680 timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
1681 if (likely(svc->srv_stats != NULL)) {
1682 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
1684 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
1685 svc->srv_n_queued_reqs);
1686 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
1687 svc->srv_n_active_reqs);
1688 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
1689 at_get(&svc->srv_at_estimate));
1692 rc = lu_context_init(&request->rq_session,
1693 LCT_SESSION|LCT_REMEMBER|LCT_NOREF);
1695 CERROR("Failure to initialize session: %d\n", rc);
1698 request->rq_session.lc_thread = thread;
1699 request->rq_session.lc_cookie = 0x5;
1700 lu_context_enter(&request->rq_session);
1702 CDEBUG(D_NET, "got req "LPU64"\n", request->rq_xid);
1704 request->rq_svc_thread = thread;
1706 request->rq_svc_thread->t_env->le_ses = &request->rq_session;
1708 if (likely(request->rq_export)) {
1709 if (unlikely(ptlrpc_check_req(request)))
1711 ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
1712 export = class_export_rpc_get(request->rq_export);
1715 /* Discard requests queued for longer than the deadline.
1716 The deadline is increased if we send an early reply. */
1717 if (cfs_time_current_sec() > request->rq_deadline) {
1718 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s"
1719 ": deadline "CFS_DURATION_T":"CFS_DURATION_T"s ago\n",
1720 libcfs_id2str(request->rq_peer),
1721 cfs_time_sub(request->rq_deadline,
1722 request->rq_arrival_time.tv_sec),
1723 cfs_time_sub(cfs_time_current_sec(),
1724 request->rq_deadline));
1725 goto put_rpc_export;
1728 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
1729 "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1730 (request->rq_export ?
1731 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1732 (request->rq_export ?
1733 cfs_atomic_read(&request->rq_export->exp_refcount) : -99),
1734 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1735 libcfs_id2str(request->rq_peer),
1736 lustre_msg_get_opc(request->rq_reqmsg));
1738 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
1739 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
1741 rc = svc->srv_ops.so_req_handler(request);
1743 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
1747 class_export_rpc_put(export);
1749 lu_context_exit(&request->rq_session);
1750 lu_context_fini(&request->rq_session);
1752 if (unlikely(cfs_time_current_sec() > request->rq_deadline)) {
1753 DEBUG_REQ(D_WARNING, request, "Request x"LPU64" took longer "
1754 "than estimated ("CFS_DURATION_T":"CFS_DURATION_T"s);"
1755 " client may timeout.",
1756 request->rq_xid, cfs_time_sub(request->rq_deadline,
1757 request->rq_arrival_time.tv_sec),
1758 cfs_time_sub(cfs_time_current_sec(),
1759 request->rq_deadline));
1762 cfs_gettimeofday(&work_end);
1763 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1764 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
1765 "%s:%s+%d:%d:x"LPU64":%s:%d Request procesed in "
1766 "%ldus (%ldus total) trans "LPU64" rc %d/%d\n",
1768 (request->rq_export ?
1769 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1770 (request->rq_export ?
1771 cfs_atomic_read(&request->rq_export->exp_refcount) : -99),
1772 lustre_msg_get_status(request->rq_reqmsg),
1774 libcfs_id2str(request->rq_peer),
1775 lustre_msg_get_opc(request->rq_reqmsg),
1777 cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL),
1778 (request->rq_repmsg ?
1779 lustre_msg_get_transno(request->rq_repmsg) :
1780 request->rq_transno),
1782 (request->rq_repmsg ?
1783 lustre_msg_get_status(request->rq_repmsg) : -999));
1784 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
1785 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
1786 int opc = opcode_offset(op);
1787 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
1788 LASSERT(opc < LUSTRE_MAX_OPCODES);
1789 lprocfs_counter_add(svc->srv_stats,
1790 opc + EXTRA_MAX_OPCODES,
1794 if (unlikely(request->rq_early_count)) {
1795 DEBUG_REQ(D_ADAPTTO, request,
1796 "sent %d early replies before finishing in "
1798 request->rq_early_count,
1799 cfs_time_sub(work_end.tv_sec,
1800 request->rq_arrival_time.tv_sec));
1804 ptlrpc_server_finish_request(svc, request);
1810 * An internal function to process a single reply state object.
1813 ptlrpc_handle_rs (struct ptlrpc_reply_state *rs)
1815 struct ptlrpc_service *svc = rs->rs_service;
1816 struct obd_export *exp;
1821 exp = rs->rs_export;
1823 LASSERT (rs->rs_difficult);
1824 LASSERT (rs->rs_scheduled);
1825 LASSERT (cfs_list_empty(&rs->rs_list));
1827 cfs_spin_lock (&exp->exp_lock);
1828 /* Noop if removed already */
1829 cfs_list_del_init (&rs->rs_exp_list);
1830 cfs_spin_unlock (&exp->exp_lock);
1832 /* The disk commit callback holds exp_uncommitted_replies_lock while it
1833 * iterates over newly committed replies, removing them from
1834 * exp_uncommitted_replies. It then drops this lock and schedules the
1835 * replies it found for handling here.
1837 * We can avoid contention for exp_uncommitted_replies_lock between the
1838 * HRT threads and further commit callbacks by checking rs_committed
1839 * which is set in the commit callback while it holds both
1840 * rs_lock and exp_uncommitted_reples.
1842 * If we see rs_committed clear, the commit callback _may_ not have
1843 * handled this reply yet and we race with it to grab
1844 * exp_uncommitted_replies_lock before removing the reply from
1845 * exp_uncommitted_replies. Note that if we lose the race and the
1846 * reply has already been removed, list_del_init() is a noop.
1848 * If we see rs_committed set, we know the commit callback is handling,
1849 * or has handled this reply since store reordering might allow us to
1850 * see rs_committed set out of sequence. But since this is done
1851 * holding rs_lock, we can be sure it has all completed once we hold
1852 * rs_lock, which we do right next.
1854 if (!rs->rs_committed) {
1855 cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
1856 cfs_list_del_init(&rs->rs_obd_list);
1857 cfs_spin_unlock(&exp->exp_uncommitted_replies_lock);
1860 cfs_spin_lock(&rs->rs_lock);
1862 been_handled = rs->rs_handled;
1865 nlocks = rs->rs_nlocks; /* atomic "steal", but */
1866 rs->rs_nlocks = 0; /* locks still on rs_locks! */
1868 if (nlocks == 0 && !been_handled) {
1869 /* If we see this, we should already have seen the warning
1870 * in mds_steal_ack_locks() */
1871 CWARN("All locks stolen from rs %p x"LPD64".t"LPD64
1874 rs->rs_xid, rs->rs_transno, rs->rs_opc,
1875 libcfs_nid2str(exp->exp_connection->c_peer.nid));
1878 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
1879 cfs_spin_unlock(&rs->rs_lock);
1881 if (!been_handled && rs->rs_on_net) {
1882 LNetMDUnlink(rs->rs_md_h);
1883 /* Ignore return code; we're racing with
1887 while (nlocks-- > 0)
1888 ldlm_lock_decref(&rs->rs_locks[nlocks],
1889 rs->rs_modes[nlocks]);
1891 cfs_spin_lock(&rs->rs_lock);
1894 rs->rs_scheduled = 0;
1896 if (!rs->rs_on_net) {
1898 cfs_spin_unlock(&rs->rs_lock);
1900 class_export_put (exp);
1901 rs->rs_export = NULL;
1902 ptlrpc_rs_decref (rs);
1903 if (cfs_atomic_dec_and_test(&svc->srv_n_difficult_replies) &&
1904 svc->srv_is_stopping)
1905 cfs_waitq_broadcast(&svc->srv_waitq);
1909 /* still on the net; callback will schedule */
1910 cfs_spin_unlock(&rs->rs_lock);
1917 * Check whether given service has a reply available for processing
1920 * \param svc a ptlrpc service
1921 * \retval 0 no replies processed
1922 * \retval 1 one reply processed
1925 ptlrpc_server_handle_reply(struct ptlrpc_service *svc)
1927 struct ptlrpc_reply_state *rs = NULL;
1930 cfs_spin_lock(&svc->srv_rs_lock);
1931 if (!cfs_list_empty(&svc->srv_reply_queue)) {
1932 rs = cfs_list_entry(svc->srv_reply_queue.prev,
1933 struct ptlrpc_reply_state,
1935 cfs_list_del_init(&rs->rs_list);
1937 cfs_spin_unlock(&svc->srv_rs_lock);
1939 ptlrpc_handle_rs(rs);
1943 /* FIXME make use of timeout later */
1945 liblustre_check_services (void *arg)
1947 int did_something = 0;
1949 cfs_list_t *tmp, *nxt;
1952 /* I'm relying on being single threaded, not to have to lock
1953 * ptlrpc_all_services etc */
1954 cfs_list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
1955 struct ptlrpc_service *svc =
1956 cfs_list_entry (tmp, struct ptlrpc_service, srv_list);
1958 if (svc->srv_threads_running != 0) /* I've recursed */
1961 /* service threads can block for bulk, so this limits us
1962 * (arbitrarily) to recursing 1 stack frame per service.
1963 * Note that the problem with recursion is that we have to
1964 * unwind completely before our caller can resume. */
1966 svc->srv_threads_running++;
1969 rc = ptlrpc_server_handle_req_in(svc);
1970 rc |= ptlrpc_server_handle_reply(svc);
1971 rc |= ptlrpc_at_check_timed(svc);
1972 rc |= ptlrpc_server_handle_request(svc, NULL);
1973 rc |= (ptlrpc_server_post_idle_rqbds(svc) > 0);
1974 did_something |= rc;
1977 svc->srv_threads_running--;
1980 RETURN(did_something);
1982 #define ptlrpc_stop_all_threads(s) do {} while (0)
1984 #else /* __KERNEL__ */
1987 ptlrpc_check_rqbd_pool(struct ptlrpc_service *svc)
1989 int avail = svc->srv_nrqbd_receiving;
1990 int low_water = test_req_buffer_pressure ? 0 :
1991 svc->srv_nbuf_per_group / 2;
1993 /* NB I'm not locking; just looking. */
1995 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
1996 * allowed the request history to grow out of control. We could put a
1997 * sanity check on that here and cull some history if we need the
2000 if (avail <= low_water)
2001 ptlrpc_grow_req_bufs(svc);
2004 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQBUF_AVAIL_CNTR,
2009 ptlrpc_retry_rqbds(void *arg)
2011 struct ptlrpc_service *svc = (struct ptlrpc_service *)arg;
2013 svc->srv_rqbd_timeout = 0;
2014 return (-ETIMEDOUT);
2018 ptlrpc_threads_enough(struct ptlrpc_service *svc)
2020 return svc->srv_n_active_reqs <
2021 svc->srv_threads_running - 1 -
2022 (svc->srv_ops.so_hpreq_handler != NULL);
2026 * allowed to create more threads
2027 * user can call it w/o any lock but need to hold ptlrpc_service::srv_lock to
2028 * get reliable result
2031 ptlrpc_threads_increasable(struct ptlrpc_service *svc)
2033 return svc->srv_threads_running +
2034 svc->srv_threads_starting < svc->srv_threads_max;
2038 * too many requests and allowed to create more threads
2041 ptlrpc_threads_need_create(struct ptlrpc_service *svc)
2043 return !ptlrpc_threads_enough(svc) && ptlrpc_threads_increasable(svc);
2047 ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2049 return thread_is_stopping(thread) ||
2050 thread->t_svc->srv_is_stopping;
2054 ptlrpc_rqbd_pending(struct ptlrpc_service *svc)
2056 return !cfs_list_empty(&svc->srv_idle_rqbds) &&
2057 svc->srv_rqbd_timeout == 0;
2061 ptlrpc_at_check(struct ptlrpc_service *svc)
2063 return svc->srv_at_check;
2067 * requests wait on preprocessing
2068 * user can call it w/o any lock but need to hold ptlrpc_service::srv_lock to
2069 * get reliable result
2072 ptlrpc_server_request_waiting(struct ptlrpc_service *svc)
2074 return !cfs_list_empty(&svc->srv_req_in_queue);
2077 static __attribute__((__noinline__)) int
2078 ptlrpc_wait_event(struct ptlrpc_service *svc,
2079 struct ptlrpc_thread *thread)
2081 /* Don't exit while there are replies to be handled */
2082 struct l_wait_info lwi = LWI_TIMEOUT(svc->srv_rqbd_timeout,
2083 ptlrpc_retry_rqbds, svc);
2085 lc_watchdog_disable(thread->t_watchdog);
2089 l_wait_event_exclusive_head(svc->srv_waitq,
2090 ptlrpc_thread_stopping(thread) ||
2091 ptlrpc_server_request_waiting(svc) ||
2092 ptlrpc_server_request_pending(svc, 0) ||
2093 ptlrpc_rqbd_pending(svc) ||
2094 ptlrpc_at_check(svc), &lwi);
2096 if (ptlrpc_thread_stopping(thread))
2099 lc_watchdog_touch(thread->t_watchdog, CFS_GET_TIMEOUT(svc));
2105 * Main thread body for service threads.
2106 * Waits in a loop waiting for new requests to process to appear.
2107 * Every time an incoming requests is added to its queue, a waitq
2108 * is woken up and one of the threads will handle it.
2110 static int ptlrpc_main(void *arg)
2112 struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
2113 struct ptlrpc_service *svc = data->svc;
2114 struct ptlrpc_thread *thread = data->thread;
2115 struct ptlrpc_reply_state *rs;
2116 #ifdef WITH_GROUP_INFO
2117 cfs_group_info_t *ginfo = NULL;
2120 int counter = 0, rc = 0;
2123 thread->t_pid = cfs_curproc_pid();
2124 cfs_daemonize_ctxt(data->name);
2126 #if defined(HAVE_NODE_TO_CPUMASK) && defined(CONFIG_NUMA)
2127 /* we need to do this before any per-thread allocation is done so that
2128 * we get the per-thread allocations on local node. bug 7342 */
2129 if (svc->srv_cpu_affinity) {
2132 for (cpu = 0, num_cpu = 0; cpu < cfs_num_possible_cpus();
2134 if (!cpu_online(cpu))
2136 if (num_cpu == thread->t_id % cfs_num_online_cpus())
2140 cfs_set_cpus_allowed(cfs_current(),
2141 node_to_cpumask(cpu_to_node(cpu)));
2145 #ifdef WITH_GROUP_INFO
2146 ginfo = cfs_groups_alloc(0);
2152 cfs_set_current_groups(ginfo);
2153 cfs_put_group_info(ginfo);
2156 if (svc->srv_ops.so_thr_init != NULL) {
2157 rc = svc->srv_ops.so_thr_init(thread);
2168 rc = lu_context_init(&env->le_ctx,
2169 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2173 thread->t_env = env;
2174 env->le_ctx.lc_thread = thread;
2175 env->le_ctx.lc_cookie = 0x6;
2177 /* Alloc reply state structure for this one */
2178 OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2184 cfs_spin_lock(&svc->srv_lock);
2186 LASSERT(thread_is_starting(thread));
2187 thread_clear_flags(thread, SVC_STARTING);
2188 svc->srv_threads_starting--;
2190 /* SVC_STOPPING may already be set here if someone else is trying
2191 * to stop the service while this new thread has been dynamically
2192 * forked. We still set SVC_RUNNING to let our creator know that
2193 * we are now running, however we will exit as soon as possible */
2194 thread_add_flags(thread, SVC_RUNNING);
2195 svc->srv_threads_running++;
2196 cfs_spin_unlock(&svc->srv_lock);
2199 * wake up our creator. Note: @data is invalid after this point,
2200 * because it's allocated on ptlrpc_start_thread() stack.
2202 cfs_waitq_signal(&thread->t_ctl_waitq);
2204 thread->t_watchdog = lc_watchdog_add(CFS_GET_TIMEOUT(svc), NULL, NULL);
2206 cfs_spin_lock(&svc->srv_rs_lock);
2207 cfs_list_add(&rs->rs_list, &svc->srv_free_rs_list);
2208 cfs_waitq_signal(&svc->srv_free_rs_waitq);
2209 cfs_spin_unlock(&svc->srv_rs_lock);
2211 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2212 svc->srv_threads_running);
2214 /* XXX maintain a list of all managed devices: insert here */
2215 while (!ptlrpc_thread_stopping(thread)) {
2216 if (ptlrpc_wait_event(svc, thread))
2219 ptlrpc_check_rqbd_pool(svc);
2221 if (ptlrpc_threads_need_create(svc)) {
2222 /* Ignore return code - we tried... */
2223 ptlrpc_start_thread(svc);
2226 /* Process all incoming reqs before handling any */
2227 if (ptlrpc_server_request_waiting(svc)) {
2228 ptlrpc_server_handle_req_in(svc);
2229 /* but limit ourselves in case of flood */
2230 if (counter++ < 100)
2235 if (ptlrpc_at_check(svc))
2236 ptlrpc_at_check_timed(svc);
2238 if (ptlrpc_server_request_pending(svc, 0)) {
2239 lu_context_enter(&env->le_ctx);
2240 ptlrpc_server_handle_request(svc, thread);
2241 lu_context_exit(&env->le_ctx);
2244 if (ptlrpc_rqbd_pending(svc) &&
2245 ptlrpc_server_post_idle_rqbds(svc) < 0) {
2246 /* I just failed to repost request buffers.
2247 * Wait for a timeout (unless something else
2248 * happens) before I try again */
2249 svc->srv_rqbd_timeout = cfs_time_seconds(1)/10;
2250 CDEBUG(D_RPCTRACE,"Posted buffers: %d\n",
2251 svc->srv_nrqbd_receiving);
2255 lc_watchdog_delete(thread->t_watchdog);
2256 thread->t_watchdog = NULL;
2260 * deconstruct service specific state created by ptlrpc_start_thread()
2262 if (svc->srv_ops.so_thr_done != NULL)
2263 svc->srv_ops.so_thr_done(thread);
2266 lu_context_fini(&env->le_ctx);
2270 CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n",
2271 thread, thread->t_pid, thread->t_id, rc);
2273 cfs_spin_lock(&svc->srv_lock);
2274 if (thread_test_and_clear_flags(thread, SVC_STARTING))
2275 svc->srv_threads_starting--;
2277 if (thread_test_and_clear_flags(thread, SVC_RUNNING))
2278 /* must know immediately */
2279 svc->srv_threads_running--;
2282 thread_add_flags(thread, SVC_STOPPED);
2284 cfs_waitq_signal(&thread->t_ctl_waitq);
2285 cfs_spin_unlock(&svc->srv_lock);
2290 struct ptlrpc_hr_args {
2293 struct ptlrpc_hr_service *hrs;
2296 static int hrt_dont_sleep(struct ptlrpc_hr_thread *t,
2297 cfs_list_t *replies)
2301 cfs_spin_lock(&t->hrt_lock);
2302 cfs_list_splice_init(&t->hrt_queue, replies);
2303 result = cfs_test_bit(HRT_STOPPING, &t->hrt_flags) ||
2304 !cfs_list_empty(replies);
2305 cfs_spin_unlock(&t->hrt_lock);
2310 * Main body of "handle reply" function.
2311 * It processes acked reply states
2313 static int ptlrpc_hr_main(void *arg)
2315 struct ptlrpc_hr_args * hr_args = arg;
2316 struct ptlrpc_hr_service *hr = hr_args->hrs;
2317 struct ptlrpc_hr_thread *t = &hr->hr_threads[hr_args->thread_index];
2318 char threadname[20];
2319 CFS_LIST_HEAD(replies);
2321 snprintf(threadname, sizeof(threadname),
2322 "ptlrpc_hr_%d", hr_args->thread_index);
2324 cfs_daemonize_ctxt(threadname);
2325 #if defined(CONFIG_NUMA) && defined(HAVE_NODE_TO_CPUMASK)
2326 cfs_set_cpus_allowed(cfs_current(),
2327 node_to_cpumask(cpu_to_node(hr_args->cpu_index)));
2329 cfs_set_bit(HRT_RUNNING, &t->hrt_flags);
2330 cfs_waitq_signal(&t->hrt_wait);
2332 while (!cfs_test_bit(HRT_STOPPING, &t->hrt_flags)) {
2334 l_wait_condition(t->hrt_wait, hrt_dont_sleep(t, &replies));
2335 while (!cfs_list_empty(&replies)) {
2336 struct ptlrpc_reply_state *rs;
2338 rs = cfs_list_entry(replies.prev,
2339 struct ptlrpc_reply_state,
2341 cfs_list_del_init(&rs->rs_list);
2342 ptlrpc_handle_rs(rs);
2346 cfs_clear_bit(HRT_RUNNING, &t->hrt_flags);
2347 cfs_complete(&t->hrt_completion);
2352 static int ptlrpc_start_hr_thread(struct ptlrpc_hr_service *hr, int n, int cpu)
2354 struct ptlrpc_hr_thread *t = &hr->hr_threads[n];
2355 struct ptlrpc_hr_args args;
2359 args.thread_index = n;
2360 args.cpu_index = cpu;
2363 rc = cfs_create_thread(ptlrpc_hr_main, (void*)&args, CFS_DAEMON_FLAGS);
2365 cfs_complete(&t->hrt_completion);
2368 l_wait_condition(t->hrt_wait, cfs_test_bit(HRT_RUNNING, &t->hrt_flags));
2374 static void ptlrpc_stop_hr_thread(struct ptlrpc_hr_thread *t)
2378 cfs_set_bit(HRT_STOPPING, &t->hrt_flags);
2379 cfs_waitq_signal(&t->hrt_wait);
2380 cfs_wait_for_completion(&t->hrt_completion);
2385 static void ptlrpc_stop_hr_threads(struct ptlrpc_hr_service *hrs)
2390 for (n = 0; n < hrs->hr_n_threads; n++)
2391 ptlrpc_stop_hr_thread(&hrs->hr_threads[n]);
2396 static int ptlrpc_start_hr_threads(struct ptlrpc_hr_service *hr)
2399 int n, cpu, threads_started = 0;
2402 LASSERT(hr != NULL);
2403 LASSERT(hr->hr_n_threads > 0);
2405 for (n = 0, cpu = 0; n < hr->hr_n_threads; n++) {
2406 #if defined(CONFIG_SMP) && defined(HAVE_NODE_TO_CPUMASK)
2407 while (!cpu_online(cpu)) {
2409 if (cpu >= cfs_num_possible_cpus())
2413 rc = ptlrpc_start_hr_thread(hr, n, cpu);
2419 if (threads_started == 0) {
2420 CERROR("No reply handling threads started\n");
2423 if (threads_started < hr->hr_n_threads) {
2424 CWARN("Started only %d reply handling threads from %d\n",
2425 threads_started, hr->hr_n_threads);
2426 hr->hr_n_threads = threads_started;
2431 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
2432 struct ptlrpc_thread *thread)
2434 struct l_wait_info lwi = { 0 };
2437 CDEBUG(D_RPCTRACE, "Stopping thread [ %p : %u ]\n",
2438 thread, thread->t_pid);
2440 cfs_spin_lock(&svc->srv_lock);
2441 /* let the thread know that we would like it to stop asap */
2442 thread_add_flags(thread, SVC_STOPPING);
2443 cfs_spin_unlock(&svc->srv_lock);
2445 cfs_waitq_broadcast(&svc->srv_waitq);
2446 l_wait_event(thread->t_ctl_waitq,
2447 thread_is_stopped(thread), &lwi);
2449 cfs_spin_lock(&svc->srv_lock);
2450 cfs_list_del(&thread->t_link);
2451 cfs_spin_unlock(&svc->srv_lock);
2453 OBD_FREE_PTR(thread);
2458 * Stops all threads of a particular service \a svc
2460 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2462 struct ptlrpc_thread *thread;
2465 cfs_spin_lock(&svc->srv_lock);
2466 while (!cfs_list_empty(&svc->srv_threads)) {
2467 thread = cfs_list_entry(svc->srv_threads.next,
2468 struct ptlrpc_thread, t_link);
2470 cfs_spin_unlock(&svc->srv_lock);
2471 ptlrpc_stop_thread(svc, thread);
2472 cfs_spin_lock(&svc->srv_lock);
2475 cfs_spin_unlock(&svc->srv_lock);
2479 int ptlrpc_start_threads(struct ptlrpc_service *svc)
2484 /* We require 2 threads min - see note in
2485 ptlrpc_server_handle_request */
2486 LASSERT(svc->srv_threads_min >= 2);
2487 for (i = 0; i < svc->srv_threads_min; i++) {
2488 rc = ptlrpc_start_thread(svc);
2489 /* We have enough threads, don't start more. b=15759 */
2490 if (rc == -EMFILE) {
2495 CERROR("cannot start %s thread #%d: rc %d\n",
2496 svc->srv_thread_name, i, rc);
2497 ptlrpc_stop_all_threads(svc);
2504 int ptlrpc_start_thread(struct ptlrpc_service *svc)
2506 struct l_wait_info lwi = { 0 };
2507 struct ptlrpc_svc_data d;
2508 struct ptlrpc_thread *thread;
2513 CDEBUG(D_RPCTRACE, "%s started %d min %d max %d running %d\n",
2514 svc->srv_name, svc->srv_threads_running, svc->srv_threads_min,
2515 svc->srv_threads_max, svc->srv_threads_running);
2517 if (unlikely(svc->srv_is_stopping))
2520 if (!ptlrpc_threads_increasable(svc) ||
2521 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
2522 svc->srv_threads_running == svc->srv_threads_min - 1))
2525 OBD_ALLOC_PTR(thread);
2528 cfs_waitq_init(&thread->t_ctl_waitq);
2530 cfs_spin_lock(&svc->srv_lock);
2531 if (!ptlrpc_threads_increasable(svc)) {
2532 cfs_spin_unlock(&svc->srv_lock);
2533 OBD_FREE_PTR(thread);
2537 svc->srv_threads_starting++;
2538 thread->t_id = svc->srv_threads_next_id++;
2539 thread_add_flags(thread, SVC_STARTING);
2540 thread->t_svc = svc;
2542 cfs_list_add(&thread->t_link, &svc->srv_threads);
2543 cfs_spin_unlock(&svc->srv_lock);
2545 sprintf(name, "%s_%02d", svc->srv_thread_name, thread->t_id);
2550 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", name);
2552 /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
2553 * just drop the VM and FILES in cfs_daemonize_ctxt() right away.
2555 rc = cfs_create_thread(ptlrpc_main, &d, CFS_DAEMON_FLAGS);
2557 CERROR("cannot start thread '%s': rc %d\n", name, rc);
2559 cfs_spin_lock(&svc->srv_lock);
2560 cfs_list_del(&thread->t_link);
2561 --svc->srv_threads_starting;
2562 cfs_spin_unlock(&svc->srv_lock);
2564 OBD_FREE(thread, sizeof(*thread));
2567 l_wait_event(thread->t_ctl_waitq,
2568 thread_is_running(thread) || thread_is_stopped(thread),
2571 rc = thread_is_stopped(thread) ? thread->t_id : 0;
2576 int ptlrpc_hr_init(void)
2579 int n_cpus = cfs_num_online_cpus();
2580 struct ptlrpc_hr_service *hr;
2585 LASSERT(ptlrpc_hr == NULL);
2587 size = offsetof(struct ptlrpc_hr_service, hr_threads[n_cpus]);
2588 OBD_ALLOC(hr, size);
2591 for (i = 0; i < n_cpus; i++) {
2592 struct ptlrpc_hr_thread *t = &hr->hr_threads[i];
2594 cfs_spin_lock_init(&t->hrt_lock);
2595 cfs_waitq_init(&t->hrt_wait);
2596 CFS_INIT_LIST_HEAD(&t->hrt_queue);
2597 cfs_init_completion(&t->hrt_completion);
2599 hr->hr_n_threads = n_cpus;
2603 rc = ptlrpc_start_hr_threads(hr);
2605 OBD_FREE(hr, hr->hr_size);
2611 void ptlrpc_hr_fini(void)
2613 if (ptlrpc_hr != NULL) {
2614 ptlrpc_stop_hr_threads(ptlrpc_hr);
2615 OBD_FREE(ptlrpc_hr, ptlrpc_hr->hr_size);
2620 #endif /* __KERNEL__ */
2623 * Wait until all already scheduled replies are processed.
2625 static void ptlrpc_wait_replies(struct ptlrpc_service *svc)
2629 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
2631 rc = l_wait_event(svc->srv_waitq, cfs_atomic_read(&svc-> \
2632 srv_n_difficult_replies) == 0,
2636 CWARN("Unexpectedly long timeout %p\n", svc);
2640 int ptlrpc_unregister_service(struct ptlrpc_service *service)
2643 struct l_wait_info lwi;
2645 struct ptlrpc_reply_state *rs, *t;
2646 struct ptlrpc_at_array *array = &service->srv_at_array;
2649 service->srv_is_stopping = 1;
2650 cfs_timer_disarm(&service->srv_at_timer);
2652 ptlrpc_stop_all_threads(service);
2653 LASSERT(cfs_list_empty(&service->srv_threads));
2655 cfs_spin_lock (&ptlrpc_all_services_lock);
2656 cfs_list_del_init (&service->srv_list);
2657 cfs_spin_unlock (&ptlrpc_all_services_lock);
2659 ptlrpc_lprocfs_unregister_service(service);
2661 /* All history will be culled when the next request buffer is
2663 service->srv_max_history_rqbds = 0;
2665 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
2667 rc = LNetClearLazyPortal(service->srv_req_portal);
2670 /* Unlink all the request buffers. This forces a 'final' event with
2671 * its 'unlink' flag set for each posted rqbd */
2672 cfs_list_for_each(tmp, &service->srv_active_rqbds) {
2673 struct ptlrpc_request_buffer_desc *rqbd =
2674 cfs_list_entry(tmp, struct ptlrpc_request_buffer_desc,
2677 rc = LNetMDUnlink(rqbd->rqbd_md_h);
2678 LASSERT (rc == 0 || rc == -ENOENT);
2681 /* Wait for the network to release any buffers it's currently
2684 cfs_spin_lock(&service->srv_lock);
2685 rc = service->srv_nrqbd_receiving;
2686 cfs_spin_unlock(&service->srv_lock);
2691 /* Network access will complete in finite time but the HUGE
2692 * timeout lets us CWARN for visibility of sluggish NALs */
2693 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2694 cfs_time_seconds(1), NULL, NULL);
2695 rc = l_wait_event(service->srv_waitq,
2696 service->srv_nrqbd_receiving == 0,
2698 if (rc == -ETIMEDOUT)
2699 CWARN("Service %s waiting for request buffers\n",
2703 /* schedule all outstanding replies to terminate them */
2704 cfs_spin_lock(&service->srv_rs_lock);
2705 while (!cfs_list_empty(&service->srv_active_replies)) {
2706 struct ptlrpc_reply_state *rs =
2707 cfs_list_entry(service->srv_active_replies.next,
2708 struct ptlrpc_reply_state, rs_list);
2709 cfs_spin_lock(&rs->rs_lock);
2710 ptlrpc_schedule_difficult_reply(rs);
2711 cfs_spin_unlock(&rs->rs_lock);
2713 cfs_spin_unlock(&service->srv_rs_lock);
2715 /* purge the request queue. NB No new replies (rqbds all unlinked)
2716 * and no service threads, so I'm the only thread noodling the
2717 * request queue now */
2718 while (!cfs_list_empty(&service->srv_req_in_queue)) {
2719 struct ptlrpc_request *req =
2720 cfs_list_entry(service->srv_req_in_queue.next,
2721 struct ptlrpc_request,
2724 cfs_list_del(&req->rq_list);
2725 service->srv_n_queued_reqs--;
2726 service->srv_n_active_reqs++;
2727 ptlrpc_server_finish_request(service, req);
2729 while (ptlrpc_server_request_pending(service, 1)) {
2730 struct ptlrpc_request *req;
2732 req = ptlrpc_server_request_get(service, 1);
2733 cfs_list_del(&req->rq_list);
2734 service->srv_n_active_reqs++;
2735 ptlrpc_server_finish_request(service, req);
2737 LASSERT(service->srv_n_queued_reqs == 0);
2738 LASSERT(service->srv_n_active_reqs == 0);
2739 LASSERT(service->srv_n_history_rqbds == 0);
2740 LASSERT(cfs_list_empty(&service->srv_active_rqbds));
2742 /* Now free all the request buffers since nothing references them
2744 while (!cfs_list_empty(&service->srv_idle_rqbds)) {
2745 struct ptlrpc_request_buffer_desc *rqbd =
2746 cfs_list_entry(service->srv_idle_rqbds.next,
2747 struct ptlrpc_request_buffer_desc,
2750 ptlrpc_free_rqbd(rqbd);
2753 ptlrpc_wait_replies(service);
2755 cfs_list_for_each_entry_safe(rs, t, &service->srv_free_rs_list,
2757 cfs_list_del(&rs->rs_list);
2758 OBD_FREE_LARGE(rs, service->srv_max_reply_size);
2761 /* In case somebody rearmed this in the meantime */
2762 cfs_timer_disarm(&service->srv_at_timer);
2764 if (array->paa_reqs_array != NULL) {
2765 OBD_FREE(array->paa_reqs_array,
2766 sizeof(cfs_list_t) * array->paa_size);
2767 array->paa_reqs_array = NULL;
2770 if (array->paa_reqs_count != NULL) {
2771 OBD_FREE(array->paa_reqs_count,
2772 sizeof(__u32) * array->paa_size);
2773 array->paa_reqs_count= NULL;
2776 OBD_FREE_PTR(service);
2781 * Returns 0 if the service is healthy.
2783 * Right now, it just checks to make sure that requests aren't languishing
2784 * in the queue. We'll use this health check to govern whether a node needs
2785 * to be shot, so it's intentionally non-aggressive. */
2786 int ptlrpc_service_health_check(struct ptlrpc_service *svc)
2788 struct ptlrpc_request *request;
2789 struct timeval right_now;
2795 cfs_gettimeofday(&right_now);
2797 cfs_spin_lock(&svc->srv_rq_lock);
2798 if (!ptlrpc_server_request_pending(svc, 1)) {
2799 cfs_spin_unlock(&svc->srv_rq_lock);
2803 /* How long has the next entry been waiting? */
2804 if (cfs_list_empty(&svc->srv_request_queue))
2805 request = cfs_list_entry(svc->srv_request_hpq.next,
2806 struct ptlrpc_request, rq_list);
2808 request = cfs_list_entry(svc->srv_request_queue.next,
2809 struct ptlrpc_request, rq_list);
2810 timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
2811 cfs_spin_unlock(&svc->srv_rq_lock);
2813 if ((timediff / ONE_MILLION) > (AT_OFF ? obd_timeout * 3/2 :
2815 CERROR("%s: unhealthy - request has been waiting %lds\n",
2816 svc->srv_name, timediff / ONE_MILLION);