1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
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);
68 static CFS_LIST_HEAD(ptlrpc_all_services);
69 spinlock_t ptlrpc_all_services_lock;
72 ptlrpc_alloc_request_buffer (int size)
76 if (size > SVC_BUF_VMALLOC_THRESHOLD)
77 OBD_VMALLOC(ptr, size);
85 ptlrpc_free_request_buffer (char *ptr, int size)
87 if (size > SVC_BUF_VMALLOC_THRESHOLD)
93 struct ptlrpc_request_buffer_desc *
94 ptlrpc_alloc_rqbd (struct ptlrpc_service *svc)
96 struct ptlrpc_request_buffer_desc *rqbd;
102 rqbd->rqbd_service = svc;
103 rqbd->rqbd_refcount = 0;
104 rqbd->rqbd_cbid.cbid_fn = request_in_callback;
105 rqbd->rqbd_cbid.cbid_arg = rqbd;
106 CFS_INIT_LIST_HEAD(&rqbd->rqbd_reqs);
107 rqbd->rqbd_buffer = ptlrpc_alloc_request_buffer(svc->srv_buf_size);
109 if (rqbd->rqbd_buffer == NULL) {
114 spin_lock(&svc->srv_lock);
115 list_add(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
117 spin_unlock(&svc->srv_lock);
123 ptlrpc_free_rqbd (struct ptlrpc_request_buffer_desc *rqbd)
125 struct ptlrpc_service *svc = rqbd->rqbd_service;
127 LASSERT (rqbd->rqbd_refcount == 0);
128 LASSERT (list_empty(&rqbd->rqbd_reqs));
130 spin_lock(&svc->srv_lock);
131 list_del(&rqbd->rqbd_list);
133 spin_unlock(&svc->srv_lock);
135 ptlrpc_free_request_buffer (rqbd->rqbd_buffer, svc->srv_buf_size);
140 ptlrpc_grow_req_bufs(struct ptlrpc_service *svc)
142 struct ptlrpc_request_buffer_desc *rqbd;
145 CDEBUG(D_RPCTRACE, "%s: allocate %d new %d-byte reqbufs (%d/%d left)\n",
146 svc->srv_name, svc->srv_nbuf_per_group, svc->srv_buf_size,
147 svc->srv_nrqbd_receiving, svc->srv_nbufs);
148 for (i = 0; i < svc->srv_nbuf_per_group; i++) {
149 rqbd = ptlrpc_alloc_rqbd(svc);
152 CERROR ("%s: Can't allocate request buffer\n",
157 if (ptlrpc_server_post_idle_rqbds(svc) < 0)
165 ptlrpc_save_lock (struct ptlrpc_request *req,
166 struct lustre_handle *lock, int mode, int no_ack)
168 struct ptlrpc_reply_state *rs = req->rq_reply_state;
172 LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
174 idx = rs->rs_nlocks++;
175 rs->rs_locks[idx] = *lock;
176 rs->rs_modes[idx] = mode;
177 rs->rs_difficult = 1;
178 rs->rs_no_ack = !!no_ack;
183 #define HRT_RUNNING 0
184 #define HRT_STOPPING 1
186 struct ptlrpc_hr_thread {
188 unsigned long hrt_flags;
189 cfs_waitq_t hrt_wait;
190 struct list_head hrt_queue;
191 struct completion hrt_completion;
194 struct ptlrpc_hr_service {
198 struct ptlrpc_hr_thread hr_threads[0];
202 struct list_head rsb_replies;
203 struct ptlrpc_service *rsb_svc;
204 unsigned int rsb_n_replies;
208 * A pointer to per-node reply handling service.
210 static struct ptlrpc_hr_service *ptlrpc_hr = NULL;
213 * maximum mumber of replies scheduled in one batch
215 #define MAX_SCHEDULED 256
218 * Initialize a reply batch.
222 static void rs_batch_init(struct rs_batch *b)
224 memset(b, 0, sizeof *b);
225 CFS_INIT_LIST_HEAD(&b->rsb_replies);
229 * Dispatch all replies accumulated in the batch to one from
230 * dedicated reply handing threads.
234 static void rs_batch_dispatch(struct rs_batch *b)
236 if (b->rsb_n_replies != 0) {
237 struct ptlrpc_hr_service *hr = ptlrpc_hr;
240 idx = hr->hr_index++;
241 if (hr->hr_index >= hr->hr_n_threads)
244 spin_lock(&hr->hr_threads[idx].hrt_lock);
245 list_splice_init(&b->rsb_replies,
246 &hr->hr_threads[idx].hrt_queue);
247 spin_unlock(&hr->hr_threads[idx].hrt_lock);
248 cfs_waitq_signal(&hr->hr_threads[idx].hrt_wait);
249 b->rsb_n_replies = 0;
254 * Add a reply to a batch.
255 * Add one reply object to a batch, schedule batched replies if overload.
260 static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs)
262 struct ptlrpc_service *svc = rs->rs_service;
264 if (svc != b->rsb_svc || b->rsb_n_replies >= MAX_SCHEDULED) {
265 if (b->rsb_svc != NULL) {
266 rs_batch_dispatch(b);
267 spin_unlock(&b->rsb_svc->srv_lock);
269 spin_lock(&svc->srv_lock);
272 spin_lock(&rs->rs_lock);
273 rs->rs_scheduled_ever = 1;
274 if (rs->rs_scheduled == 0) {
275 list_move(&rs->rs_list, &b->rsb_replies);
276 rs->rs_scheduled = 1;
279 spin_unlock(&rs->rs_lock);
283 * Reply batch finalization.
284 * Dispatch remaining replies from the batch
285 * and release remaining spinlock.
289 static void rs_batch_fini(struct rs_batch *b)
291 if (b->rsb_svc != 0) {
292 rs_batch_dispatch(b);
293 spin_unlock(&b->rsb_svc->srv_lock);
297 #define DECLARE_RS_BATCH(b) struct rs_batch b
299 #else /* __KERNEL__ */
301 #define rs_batch_init(b) do{}while(0)
302 #define rs_batch_fini(b) do{}while(0)
303 #define rs_batch_add(b, r) ptlrpc_schedule_difficult_reply(r)
304 #define DECLARE_RS_BATCH(b)
306 #endif /* __KERNEL__ */
308 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
311 struct ptlrpc_hr_service *hr = ptlrpc_hr;
315 LASSERT(list_empty(&rs->rs_list));
317 idx = hr->hr_index++;
318 if (hr->hr_index >= hr->hr_n_threads)
320 spin_lock(&hr->hr_threads[idx].hrt_lock);
321 list_add_tail(&rs->rs_list, &hr->hr_threads[idx].hrt_queue);
322 spin_unlock(&hr->hr_threads[idx].hrt_lock);
323 cfs_waitq_signal(&hr->hr_threads[idx].hrt_wait);
326 list_add_tail(&rs->rs_list, &rs->rs_service->srv_reply_queue);
331 ptlrpc_schedule_difficult_reply (struct ptlrpc_reply_state *rs)
335 LASSERT_SPIN_LOCKED(&rs->rs_service->srv_lock);
336 LASSERT_SPIN_LOCKED(&rs->rs_lock);
337 LASSERT (rs->rs_difficult);
338 rs->rs_scheduled_ever = 1; /* flag any notification attempt */
340 if (rs->rs_scheduled) { /* being set up or already notified */
345 rs->rs_scheduled = 1;
346 list_del_init(&rs->rs_list);
347 ptlrpc_dispatch_difficult_reply(rs);
351 void ptlrpc_commit_replies(struct obd_export *exp)
353 struct ptlrpc_reply_state *rs, *nxt;
354 DECLARE_RS_BATCH(batch);
357 rs_batch_init(&batch);
358 /* Find any replies that have been committed and get their service
359 * to attend to complete them. */
361 /* CAVEAT EMPTOR: spinlock ordering!!! */
362 spin_lock(&exp->exp_uncommitted_replies_lock);
363 list_for_each_entry_safe(rs, nxt, &exp->exp_uncommitted_replies,
365 LASSERT (rs->rs_difficult);
366 /* VBR: per-export last_committed */
367 LASSERT(rs->rs_export);
368 if (rs->rs_transno <= exp->exp_last_committed) {
369 list_del_init(&rs->rs_obd_list);
370 rs_batch_add(&batch, rs);
373 spin_unlock(&exp->exp_uncommitted_replies_lock);
374 rs_batch_fini(&batch);
379 ptlrpc_server_post_idle_rqbds (struct ptlrpc_service *svc)
381 struct ptlrpc_request_buffer_desc *rqbd;
386 spin_lock(&svc->srv_lock);
388 if (list_empty (&svc->srv_idle_rqbds)) {
389 spin_unlock(&svc->srv_lock);
393 rqbd = list_entry(svc->srv_idle_rqbds.next,
394 struct ptlrpc_request_buffer_desc,
396 list_del (&rqbd->rqbd_list);
398 /* assume we will post successfully */
399 svc->srv_nrqbd_receiving++;
400 list_add (&rqbd->rqbd_list, &svc->srv_active_rqbds);
402 spin_unlock(&svc->srv_lock);
404 rc = ptlrpc_register_rqbd(rqbd);
411 spin_lock(&svc->srv_lock);
413 svc->srv_nrqbd_receiving--;
414 list_del(&rqbd->rqbd_list);
415 list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
417 /* Don't complain if no request buffers are posted right now; LNET
418 * won't drop requests because we set the portal lazy! */
420 spin_unlock(&svc->srv_lock);
425 struct ptlrpc_service *ptlrpc_init_svc_conf(struct ptlrpc_service_conf *c,
426 svc_handler_t h, char *name,
427 struct proc_dir_entry *proc_entry,
428 svcreq_printfn_t prntfn,
431 return ptlrpc_init_svc(c->psc_nbufs, c->psc_bufsize,
432 c->psc_max_req_size, c->psc_max_reply_size,
433 c->psc_req_portal, c->psc_rep_portal,
434 c->psc_watchdog_factor,
436 prntfn, c->psc_min_threads, c->psc_max_threads,
437 threadname, c->psc_ctx_tags, NULL);
439 EXPORT_SYMBOL(ptlrpc_init_svc_conf);
441 static void ptlrpc_at_timer(unsigned long castmeharder)
443 struct ptlrpc_service *svc = (struct ptlrpc_service *)castmeharder;
444 svc->srv_at_check = 1;
445 svc->srv_at_checktime = cfs_time_current();
446 cfs_waitq_signal(&svc->srv_waitq);
449 /* @threadname should be 11 characters or less - 3 will be added on */
450 struct ptlrpc_service *
451 ptlrpc_init_svc(int nbufs, int bufsize, int max_req_size, int max_reply_size,
452 int req_portal, int rep_portal, int watchdog_factor,
453 svc_handler_t handler, char *name,
454 cfs_proc_dir_entry_t *proc_entry,
455 svcreq_printfn_t svcreq_printfn,
456 int min_threads, int max_threads,
457 char *threadname, __u32 ctx_tags,
458 svc_hpreq_handler_t hp_handler)
461 struct ptlrpc_at_array *array;
462 struct ptlrpc_service *service;
463 unsigned int size, index;
467 LASSERT (bufsize >= max_req_size + SPTLRPC_MAX_PAYLOAD);
468 LASSERT (ctx_tags != 0);
470 OBD_ALLOC_PTR(service);
474 /* First initialise enough for early teardown */
476 service->srv_name = name;
477 spin_lock_init(&service->srv_lock);
478 CFS_INIT_LIST_HEAD(&service->srv_threads);
479 cfs_waitq_init(&service->srv_waitq);
481 service->srv_nbuf_per_group = test_req_buffer_pressure ? 1 : nbufs;
482 service->srv_max_req_size = max_req_size + SPTLRPC_MAX_PAYLOAD;
483 service->srv_buf_size = bufsize;
484 service->srv_rep_portal = rep_portal;
485 service->srv_req_portal = req_portal;
486 service->srv_watchdog_factor = watchdog_factor;
487 service->srv_handler = handler;
488 service->srv_request_history_print_fn = svcreq_printfn;
489 service->srv_request_seq = 1; /* valid seq #s start at 1 */
490 service->srv_request_max_cull_seq = 0;
491 service->srv_threads_min = min_threads;
492 service->srv_threads_max = max_threads;
493 service->srv_thread_name = threadname;
494 service->srv_ctx_tags = ctx_tags;
495 service->srv_hpreq_handler = hp_handler;
496 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
497 service->srv_hpreq_count = 0;
498 service->srv_n_hpreq = 0;
500 rc = LNetSetLazyPortal(service->srv_req_portal);
503 CFS_INIT_LIST_HEAD(&service->srv_request_queue);
504 CFS_INIT_LIST_HEAD(&service->srv_request_hpq);
505 CFS_INIT_LIST_HEAD(&service->srv_idle_rqbds);
506 CFS_INIT_LIST_HEAD(&service->srv_active_rqbds);
507 CFS_INIT_LIST_HEAD(&service->srv_history_rqbds);
508 CFS_INIT_LIST_HEAD(&service->srv_request_history);
509 CFS_INIT_LIST_HEAD(&service->srv_active_replies);
511 CFS_INIT_LIST_HEAD(&service->srv_reply_queue);
513 CFS_INIT_LIST_HEAD(&service->srv_free_rs_list);
514 cfs_waitq_init(&service->srv_free_rs_waitq);
515 atomic_set(&service->srv_n_difficult_replies, 0);
517 spin_lock_init(&service->srv_at_lock);
518 CFS_INIT_LIST_HEAD(&service->srv_req_in_queue);
520 array = &service->srv_at_array;
521 size = at_est2timeout(at_max);
522 array->paa_size = size;
523 array->paa_count = 0;
524 array->paa_deadline = -1;
526 /* allocate memory for srv_at_array (ptlrpc_at_array) */
527 OBD_ALLOC(array->paa_reqs_array, sizeof(struct list_head) * size);
528 if (array->paa_reqs_array == NULL)
531 for (index = 0; index < size; index++)
532 CFS_INIT_LIST_HEAD(&array->paa_reqs_array[index]);
534 OBD_ALLOC(array->paa_reqs_count, sizeof(__u32) * size);
535 if (array->paa_reqs_count == NULL)
538 cfs_timer_init(&service->srv_at_timer, ptlrpc_at_timer, service);
539 /* At SOW, service time should be quick; 10s seems generous. If client
540 timeout is less than this, we'll be sending an early reply. */
541 at_init(&service->srv_at_estimate, 10, 0);
543 spin_lock (&ptlrpc_all_services_lock);
544 list_add (&service->srv_list, &ptlrpc_all_services);
545 spin_unlock (&ptlrpc_all_services_lock);
547 /* Now allocate the request buffers */
548 rc = ptlrpc_grow_req_bufs(service);
549 /* We shouldn't be under memory pressure at startup, so
550 * fail if we can't post all our buffers at this time. */
554 /* Now allocate pool of reply buffers */
555 /* Increase max reply size to next power of two */
556 service->srv_max_reply_size = 1;
557 while (service->srv_max_reply_size <
558 max_reply_size + SPTLRPC_MAX_PAYLOAD)
559 service->srv_max_reply_size <<= 1;
561 if (proc_entry != NULL)
562 ptlrpc_lprocfs_register_service(proc_entry, service);
564 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
565 service->srv_name, service->srv_req_portal);
569 ptlrpc_unregister_service(service);
574 * to actually free the request, must be called without holding svc_lock.
575 * note it's caller's responsibility to unlink req->rq_list.
577 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
579 LASSERT(atomic_read(&req->rq_refcount) == 0);
580 LASSERT(list_empty(&req->rq_timed_list));
582 /* DEBUG_REQ() assumes the reply state of a request with a valid
583 * ref will not be destroyed until that reference is dropped. */
584 ptlrpc_req_drop_rs(req);
586 sptlrpc_svc_ctx_decref(req);
588 if (req != &req->rq_rqbd->rqbd_req) {
589 /* NB request buffers use an embedded
590 * req if the incoming req unlinked the
591 * MD; this isn't one of them! */
592 OBD_FREE(req, sizeof(*req));
597 * drop a reference count of the request. if it reaches 0, we either
598 * put it into history list, or free it immediately.
600 static void ptlrpc_server_drop_request(struct ptlrpc_request *req)
602 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
603 struct ptlrpc_service *svc = rqbd->rqbd_service;
605 struct list_head *tmp;
606 struct list_head *nxt;
608 if (!atomic_dec_and_test(&req->rq_refcount))
611 spin_lock(&svc->srv_lock);
613 svc->srv_n_active_reqs--;
614 list_add(&req->rq_list, &rqbd->rqbd_reqs);
616 refcount = --(rqbd->rqbd_refcount);
618 /* request buffer is now idle: add to history */
619 list_del(&rqbd->rqbd_list);
620 list_add_tail(&rqbd->rqbd_list, &svc->srv_history_rqbds);
621 svc->srv_n_history_rqbds++;
623 /* cull some history?
624 * I expect only about 1 or 2 rqbds need to be recycled here */
625 while (svc->srv_n_history_rqbds > svc->srv_max_history_rqbds) {
626 rqbd = list_entry(svc->srv_history_rqbds.next,
627 struct ptlrpc_request_buffer_desc,
630 list_del(&rqbd->rqbd_list);
631 svc->srv_n_history_rqbds--;
633 /* remove rqbd's reqs from svc's req history while
634 * I've got the service lock */
635 list_for_each(tmp, &rqbd->rqbd_reqs) {
636 req = list_entry(tmp, struct ptlrpc_request,
638 /* Track the highest culled req seq */
639 if (req->rq_history_seq >
640 svc->srv_request_max_cull_seq)
641 svc->srv_request_max_cull_seq =
643 list_del(&req->rq_history_list);
646 spin_unlock(&svc->srv_lock);
648 list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
649 req = list_entry(rqbd->rqbd_reqs.next,
650 struct ptlrpc_request,
652 list_del(&req->rq_list);
653 ptlrpc_server_free_request(req);
656 spin_lock(&svc->srv_lock);
658 * now all reqs including the embedded req has been
659 * disposed, schedule request buffer for re-use.
661 LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) == 0);
662 list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
665 spin_unlock(&svc->srv_lock);
666 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
667 /* If we are low on memory, we are not interested in history */
668 list_del(&req->rq_list);
669 list_del_init(&req->rq_history_list);
670 spin_unlock(&svc->srv_lock);
672 ptlrpc_server_free_request(req);
674 spin_unlock(&svc->srv_lock);
679 * to finish a request: stop sending more early replies, and release
680 * the request. should be called after we finished handling the request.
682 static void ptlrpc_server_finish_request(struct ptlrpc_request *req)
684 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
686 if (req->rq_export) {
687 class_export_put(req->rq_export);
688 req->rq_export = NULL;
691 if (req->rq_phase != RQ_PHASE_NEW) /* incorrect message magic */
692 DEBUG_REQ(D_INFO, req, "free req");
694 spin_lock(&svc->srv_at_lock);
695 req->rq_sent_final = 1;
696 list_del_init(&req->rq_timed_list);
697 if (req->rq_at_linked) {
698 struct ptlrpc_at_array *array = &svc->srv_at_array;
699 __u32 index = req->rq_at_index;
701 req->rq_at_linked = 0;
702 array->paa_reqs_count[index]--;
705 spin_unlock(&svc->srv_at_lock);
707 ptlrpc_server_drop_request(req);
710 /* This function makes sure dead exports are evicted in a timely manner.
711 This function is only called when some export receives a message (i.e.,
712 the network is up.) */
713 static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
715 struct obd_export *oldest_exp;
716 time_t oldest_time, new_time;
722 /* Compensate for slow machines, etc, by faking our request time
723 into the future. Although this can break the strict time-ordering
724 of the list, we can be really lazy here - we don't have to evict
725 at the exact right moment. Eventually, all silent exports
726 will make it to the top of the list. */
728 /* Do not pay attention on 1sec or smaller renewals. */
729 new_time = cfs_time_current_sec() + extra_delay;
730 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
733 exp->exp_last_request_time = new_time;
734 CDEBUG(D_HA, "updating export %s at "CFS_TIME_T" exp %p\n",
735 exp->exp_client_uuid.uuid,
736 exp->exp_last_request_time, exp);
738 /* exports may get disconnected from the chain even though the
739 export has references, so we must keep the spin lock while
740 manipulating the lists */
741 spin_lock(&exp->exp_obd->obd_dev_lock);
743 if (list_empty(&exp->exp_obd_chain_timed)) {
744 /* this one is not timed */
745 spin_unlock(&exp->exp_obd->obd_dev_lock);
749 list_move_tail(&exp->exp_obd_chain_timed,
750 &exp->exp_obd->obd_exports_timed);
752 oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
753 struct obd_export, exp_obd_chain_timed);
754 oldest_time = oldest_exp->exp_last_request_time;
755 spin_unlock(&exp->exp_obd->obd_dev_lock);
757 if (exp->exp_obd->obd_recovering) {
758 /* be nice to everyone during recovery */
763 /* Note - racing to start/reset the obd_eviction timer is safe */
764 if (exp->exp_obd->obd_eviction_timer == 0) {
765 /* Check if the oldest entry is expired. */
766 if (cfs_time_current_sec() > (oldest_time + PING_EVICT_TIMEOUT +
768 /* We need a second timer, in case the net was down and
769 * it just came back. Since the pinger may skip every
770 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
771 * we better wait for 3. */
772 exp->exp_obd->obd_eviction_timer =
773 cfs_time_current_sec() + 3 * PING_INTERVAL;
774 CDEBUG(D_HA, "%s: Think about evicting %s from "CFS_TIME_T"\n",
775 exp->exp_obd->obd_name, obd_export_nid2str(exp),
779 if (cfs_time_current_sec() >
780 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
781 /* The evictor won't evict anyone who we've heard from
782 * recently, so we don't have to check before we start
784 if (!ping_evictor_wake(exp))
785 exp->exp_obd->obd_eviction_timer = 0;
792 static int ptlrpc_check_req(struct ptlrpc_request *req)
794 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
795 req->rq_export->exp_conn_cnt)) {
796 DEBUG_REQ(D_ERROR, req,
797 "DROPPING req from old connection %d < %d",
798 lustre_msg_get_conn_cnt(req->rq_reqmsg),
799 req->rq_export->exp_conn_cnt);
802 if (unlikely(req->rq_export->exp_obd &&
803 req->rq_export->exp_obd->obd_fail)) {
804 /* Failing over, don't handle any more reqs, send
805 error response instead. */
806 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
807 req, req->rq_export->exp_obd->obd_name);
808 req->rq_status = -ENODEV;
816 static void ptlrpc_at_set_timer(struct ptlrpc_service *svc)
818 struct ptlrpc_at_array *array = &svc->srv_at_array;
821 spin_lock(&svc->srv_at_lock);
822 if (array->paa_count == 0) {
823 cfs_timer_disarm(&svc->srv_at_timer);
824 spin_unlock(&svc->srv_at_lock);
828 /* Set timer for closest deadline */
829 next = (__s32)(array->paa_deadline - cfs_time_current_sec() -
832 ptlrpc_at_timer((unsigned long)svc);
834 cfs_timer_arm(&svc->srv_at_timer, cfs_time_shift(next));
835 spin_unlock(&svc->srv_at_lock);
836 CDEBUG(D_INFO, "armed %s at %+ds\n", svc->srv_name, next);
839 /* Add rpc to early reply check list */
840 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
842 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
843 struct ptlrpc_request *rq = NULL;
844 struct ptlrpc_at_array *array = &svc->srv_at_array;
851 if (req->rq_no_reply)
854 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
857 spin_lock(&svc->srv_at_lock);
859 if (unlikely(req->rq_sent_final)) {
860 spin_unlock(&svc->srv_at_lock);
864 LASSERT(list_empty(&req->rq_timed_list));
866 index = req->rq_deadline % array->paa_size;
867 if (array->paa_reqs_count[index] > 0) {
868 /* latest rpcs will have the latest deadlines in the list,
869 * so search backward. */
870 list_for_each_entry_reverse(rq, &array->paa_reqs_array[index],
872 if (req->rq_deadline >= rq->rq_deadline) {
873 list_add(&req->rq_timed_list,
880 /* Add the request at the head of the list */
881 if (list_empty(&req->rq_timed_list))
882 list_add(&req->rq_timed_list, &array->paa_reqs_array[index]);
884 req->rq_at_linked = 1;
885 req->rq_at_index = index;
886 array->paa_reqs_count[index]++;
888 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
889 array->paa_deadline = req->rq_deadline;
892 spin_unlock(&svc->srv_at_lock);
895 ptlrpc_at_set_timer(svc);
900 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req,
903 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
904 struct ptlrpc_request *reqcopy;
905 struct lustre_msg *reqmsg;
906 cfs_duration_t olddl = req->rq_deadline - cfs_time_current_sec();
911 /* deadline is when the client expects us to reply, margin is the
912 difference between clients' and servers' expectations */
913 DEBUG_REQ(D_ADAPTTO, req,
914 "%ssending early reply (deadline %+lds, margin %+lds) for "
915 "%d+%d", AT_OFF ? "AT off - not " : "",
916 olddl, olddl - at_get(&svc->srv_at_estimate),
917 at_get(&svc->srv_at_estimate), extra_time);
923 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), "
924 "not sending early reply. Consider increasing "
925 "at_early_margin (%d)?", olddl, at_early_margin);
927 /* Return an error so we're not re-added to the timed list. */
931 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
932 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
933 "but no AT support");
937 if (req->rq_export && req->rq_export->exp_in_recovery) {
938 /* don't increase server estimates during recovery, and give
939 clients the full recovery time. */
940 newdl = cfs_time_current_sec() +
941 req->rq_export->exp_obd->obd_recovery_timeout;
944 /* Fake our processing time into the future to ask the
945 clients for some extra amount of time */
946 extra_time += cfs_time_current_sec() -
947 req->rq_arrival_time.tv_sec;
948 at_add(&svc->srv_at_estimate, extra_time);
950 newdl = req->rq_arrival_time.tv_sec +
951 at_get(&svc->srv_at_estimate);
953 if (req->rq_deadline >= newdl) {
954 /* We're not adding any time, no need to send an early reply
955 (e.g. maybe at adaptive_max) */
956 DEBUG_REQ(D_WARNING, req, "Couldn't add any time ("
957 CFS_DURATION_T"/"CFS_DURATION_T"), "
958 "not sending early reply\n", olddl,
959 cfs_time_sub(newdl, cfs_time_current_sec()));
963 OBD_ALLOC(reqcopy, sizeof *reqcopy);
966 OBD_ALLOC(reqmsg, req->rq_reqlen);
968 OBD_FREE(reqcopy, sizeof *reqcopy);
973 reqcopy->rq_reply_state = NULL;
974 reqcopy->rq_rep_swab_mask = 0;
975 reqcopy->rq_pack_bulk = 0;
976 reqcopy->rq_pack_udesc = 0;
977 reqcopy->rq_packed_final = 0;
978 sptlrpc_svc_ctx_addref(reqcopy);
979 /* We only need the reqmsg for the magic */
980 reqcopy->rq_reqmsg = reqmsg;
981 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
983 if (req->rq_sent_final) {
984 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
985 "abort sending early reply\n");
990 reqcopy->rq_export = class_conn2export(
991 lustre_msg_get_handle(reqcopy->rq_reqmsg));
992 if (reqcopy->rq_export == NULL)
993 GOTO(out, rc = -ENODEV);
996 class_export_rpc_get(reqcopy->rq_export);
997 if (reqcopy->rq_export->exp_obd &&
998 reqcopy->rq_export->exp_obd->obd_fail)
999 GOTO(out_put, rc = -ENODEV);
1001 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1005 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1008 /* Adjust our own deadline to what we told the client */
1009 req->rq_deadline = newdl;
1010 req->rq_early_count++; /* number sent, server side */
1012 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1015 /* Free the (early) reply state from lustre_pack_reply.
1016 (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
1017 ptlrpc_req_drop_rs(reqcopy);
1020 class_export_rpc_put(reqcopy->rq_export);
1021 class_export_put(reqcopy->rq_export);
1023 sptlrpc_svc_ctx_decref(reqcopy);
1024 OBD_FREE(reqmsg, req->rq_reqlen);
1025 OBD_FREE(reqcopy, sizeof *reqcopy);
1029 /* Send early replies to everybody expiring within at_early_margin
1030 asking for at_extra time */
1031 static int ptlrpc_at_check_timed(struct ptlrpc_service *svc)
1033 struct ptlrpc_request *rq, *n;
1034 struct list_head work_list;
1035 struct ptlrpc_at_array *array = &svc->srv_at_array;
1038 time_t now = cfs_time_current_sec();
1039 cfs_duration_t delay;
1040 int first, counter = 0;
1043 spin_lock(&svc->srv_at_lock);
1044 if (svc->srv_at_check == 0) {
1045 spin_unlock(&svc->srv_at_lock);
1048 delay = cfs_time_sub(cfs_time_current(), svc->srv_at_checktime);
1049 svc->srv_at_check = 0;
1051 if (array->paa_count == 0) {
1052 spin_unlock(&svc->srv_at_lock);
1056 /* The timer went off, but maybe the nearest rpc already completed. */
1057 first = array->paa_deadline - now;
1058 if (first > at_early_margin) {
1059 /* We've still got plenty of time. Reset the timer. */
1060 spin_unlock(&svc->srv_at_lock);
1061 ptlrpc_at_set_timer(svc);
1065 /* We're close to a timeout, and we don't know how much longer the
1066 server will take. Send early replies to everyone expiring soon. */
1067 CFS_INIT_LIST_HEAD(&work_list);
1069 index = array->paa_deadline % array->paa_size;
1070 count = array->paa_count;
1072 count -= array->paa_reqs_count[index];
1073 list_for_each_entry_safe(rq, n, &array->paa_reqs_array[index],
1075 if (rq->rq_deadline <= now + at_early_margin) {
1076 list_move(&rq->rq_timed_list, &work_list);
1078 array->paa_reqs_count[index]--;
1080 rq->rq_at_linked = 0;
1084 /* update the earliest deadline */
1085 if (deadline == -1 || rq->rq_deadline < deadline)
1086 deadline = rq->rq_deadline;
1091 if (++index >= array->paa_size)
1094 array->paa_deadline = deadline;
1095 spin_unlock(&svc->srv_at_lock);
1097 /* we have a new earliest deadline, restart the timer */
1098 ptlrpc_at_set_timer(svc);
1100 CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
1101 "replies\n", first, at_extra, counter);
1103 /* We're already past request deadlines before we even get a
1104 chance to send early replies */
1105 LCONSOLE_WARN("%s: This server is not able to keep up with "
1106 "request traffic (cpu-bound).\n", svc->srv_name);
1107 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, "
1108 "delay="CFS_DURATION_T"(jiff)\n",
1109 counter, svc->srv_n_queued_reqs, svc->srv_n_active_reqs,
1110 at_get(&svc->srv_at_estimate), delay);
1113 /* ptlrpc_server_finish_request may delete an entry out of
1115 spin_lock(&svc->srv_at_lock);
1116 while (!list_empty(&work_list)) {
1117 rq = list_entry(work_list.next, struct ptlrpc_request,
1119 list_del_init(&rq->rq_timed_list);
1120 /* if the entry is still in the worklist, it hasn't been
1121 deleted, and is safe to take a ref to keep the req around */
1122 atomic_inc(&rq->rq_refcount);
1123 spin_unlock(&svc->srv_at_lock);
1125 if (ptlrpc_at_send_early_reply(rq, at_extra) == 0)
1126 ptlrpc_at_add_timed(rq);
1128 ptlrpc_server_drop_request(rq);
1129 spin_lock(&svc->srv_at_lock);
1131 spin_unlock(&svc->srv_at_lock);
1137 * Put the request to the export list if the request may become
1138 * a high priority one.
1140 static int ptlrpc_hpreq_init(struct ptlrpc_service *svc,
1141 struct ptlrpc_request *req)
1146 if (svc->srv_hpreq_handler) {
1147 rc = svc->srv_hpreq_handler(req);
1151 if (req->rq_export && req->rq_ops) {
1152 spin_lock(&req->rq_export->exp_lock);
1153 list_add(&req->rq_exp_list, &req->rq_export->exp_queued_rpc);
1154 spin_unlock(&req->rq_export->exp_lock);
1160 /** Remove the request from the export list. */
1161 static void ptlrpc_hpreq_fini(struct ptlrpc_request *req)
1164 if (req->rq_export && req->rq_ops) {
1165 spin_lock(&req->rq_export->exp_lock);
1166 list_del_init(&req->rq_exp_list);
1167 spin_unlock(&req->rq_export->exp_lock);
1173 * Make the request a high priority one.
1175 * All the high priority requests are queued in a separate FIFO
1176 * ptlrpc_service::srv_request_hpq list which is parallel to
1177 * ptlrpc_service::srv_request_queue list but has a higher priority
1180 * \see ptlrpc_server_handle_request().
1182 static void ptlrpc_hpreq_reorder_nolock(struct ptlrpc_service *svc,
1183 struct ptlrpc_request *req)
1186 LASSERT(svc != NULL);
1187 spin_lock(&req->rq_lock);
1188 if (req->rq_hp == 0) {
1189 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1191 /* Add to the high priority queue. */
1192 list_move_tail(&req->rq_list, &svc->srv_request_hpq);
1194 if (opc != OBD_PING)
1195 DEBUG_REQ(D_NET, req, "high priority req");
1197 spin_unlock(&req->rq_lock);
1201 void ptlrpc_hpreq_reorder(struct ptlrpc_request *req)
1203 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
1206 spin_lock(&svc->srv_lock);
1207 /* It may happen that the request is already taken for the processing
1208 * but still in the export list, do not re-add it into the HP list. */
1209 if (req->rq_phase == RQ_PHASE_NEW)
1210 ptlrpc_hpreq_reorder_nolock(svc, req);
1211 spin_unlock(&svc->srv_lock);
1215 /** Check if the request if a high priority one. */
1216 static int ptlrpc_server_hpreq_check(struct ptlrpc_request *req)
1221 /* Check by request opc. */
1222 opc = lustre_msg_get_opc(req->rq_reqmsg);
1223 if (opc == OBD_PING)
1226 /* Perform request specific check. */
1227 if (req->rq_ops && req->rq_ops->hpreq_check)
1228 rc = req->rq_ops->hpreq_check(req);
1232 /** Check if a request is a high priority one. */
1233 static int ptlrpc_server_request_add(struct ptlrpc_service *svc,
1234 struct ptlrpc_request *req)
1239 rc = ptlrpc_server_hpreq_check(req);
1243 spin_lock(&svc->srv_lock);
1244 /* Before inserting the request into the queue, check if it is not
1245 * inserted yet, or even already handled -- it may happen due to
1246 * a racing ldlm_server_blocking_ast(). */
1247 if (req->rq_phase == RQ_PHASE_NEW && list_empty(&req->rq_list)) {
1249 ptlrpc_hpreq_reorder_nolock(svc, req);
1251 list_add_tail(&req->rq_list, &svc->srv_request_queue);
1253 spin_unlock(&svc->srv_lock);
1258 /* Only allow normal priority requests on a service that has a high-priority
1259 * queue if forced (i.e. cleanup), if there are other high priority requests
1260 * already being processed (i.e. those threads can service more high-priority
1261 * requests), or if there are enough idle threads that a later thread can do
1262 * a high priority request. */
1263 static int ptlrpc_server_allow_normal(struct ptlrpc_service *svc, int force)
1265 return force || !svc->srv_hpreq_handler || svc->srv_n_hpreq > 0 ||
1266 svc->srv_n_active_reqs < svc->srv_threads_running - 2;
1269 static struct ptlrpc_request *
1270 ptlrpc_server_request_get(struct ptlrpc_service *svc, int force)
1272 struct ptlrpc_request *req = NULL;
1275 if (ptlrpc_server_allow_normal(svc, force) &&
1276 !list_empty(&svc->srv_request_queue) &&
1277 (list_empty(&svc->srv_request_hpq) ||
1278 svc->srv_hpreq_count >= svc->srv_hpreq_ratio)) {
1279 req = list_entry(svc->srv_request_queue.next,
1280 struct ptlrpc_request, rq_list);
1281 svc->srv_hpreq_count = 0;
1282 } else if (!list_empty(&svc->srv_request_hpq)) {
1283 req = list_entry(svc->srv_request_hpq.next,
1284 struct ptlrpc_request, rq_list);
1285 svc->srv_hpreq_count++;
1290 static int ptlrpc_server_request_pending(struct ptlrpc_service *svc, int force)
1292 return ((ptlrpc_server_allow_normal(svc, force) &&
1293 !list_empty(&svc->srv_request_queue)) ||
1294 !list_empty(&svc->srv_request_hpq));
1297 /* Handle freshly incoming reqs, add to timed early reply list,
1298 pass on to regular request queue */
1300 ptlrpc_server_handle_req_in(struct ptlrpc_service *svc)
1302 struct ptlrpc_request *req;
1309 spin_lock(&svc->srv_lock);
1310 if (list_empty(&svc->srv_req_in_queue)) {
1311 spin_unlock(&svc->srv_lock);
1315 req = list_entry(svc->srv_req_in_queue.next,
1316 struct ptlrpc_request, rq_list);
1317 list_del_init (&req->rq_list);
1318 /* Consider this still a "queued" request as far as stats are
1320 spin_unlock(&svc->srv_lock);
1322 /* go through security check/transform */
1323 rc = sptlrpc_svc_unwrap_request(req);
1327 case SECSVC_COMPLETE:
1328 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1336 /* Clear request swab mask; this is a new request */
1337 req->rq_req_swab_mask = 0;
1339 rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
1341 CERROR("error unpacking request: ptl %d from %s x"LPU64"\n",
1342 svc->srv_req_portal, libcfs_id2str(req->rq_peer),
1347 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1349 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
1350 LPU64"\n", svc->srv_req_portal,
1351 libcfs_id2str(req->rq_peer), req->rq_xid);
1356 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1357 CERROR("wrong packet type received (type=%u) from %s\n",
1358 lustre_msg_get_type(req->rq_reqmsg),
1359 libcfs_id2str(req->rq_peer));
1363 switch(lustre_msg_get_opc(req->rq_reqmsg)) {
1366 req->rq_bulk_write = 1;
1370 req->rq_bulk_read = 1;
1374 CDEBUG(D_NET, "got req "LPU64"\n", req->rq_xid);
1376 req->rq_export = class_conn2export(
1377 lustre_msg_get_handle(req->rq_reqmsg));
1378 if (req->rq_export) {
1379 rc = ptlrpc_check_req(req);
1381 rc = sptlrpc_target_export_check(req->rq_export, req);
1383 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
1384 "illegal security flavor,");
1389 ptlrpc_update_export_timer(req->rq_export, 0);
1392 /* req_in handling should/must be fast */
1393 if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5)
1394 DEBUG_REQ(D_WARNING, req, "Slow req_in handling "CFS_DURATION_T"s",
1395 cfs_time_sub(cfs_time_current_sec(),
1396 req->rq_arrival_time.tv_sec));
1398 /* Set rpc server deadline and add it to the timed list */
1399 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1400 MSGHDR_AT_SUPPORT) ?
1401 /* The max time the client expects us to take */
1402 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1403 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1404 if (unlikely(deadline == 0)) {
1405 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1409 ptlrpc_at_add_timed(req);
1410 rc = ptlrpc_hpreq_init(svc, req);
1414 /* Move it over to the request processing queue */
1415 rc = ptlrpc_server_request_add(svc, req);
1418 cfs_waitq_signal(&svc->srv_waitq);
1422 spin_lock(&svc->srv_lock);
1423 svc->srv_n_queued_reqs--;
1424 svc->srv_n_active_reqs++;
1425 spin_unlock(&svc->srv_lock);
1426 ptlrpc_server_finish_request(req);
1432 ptlrpc_server_handle_request(struct ptlrpc_service *svc,
1433 struct ptlrpc_thread *thread)
1435 struct obd_export *export = NULL;
1436 struct ptlrpc_request *request;
1437 struct timeval work_start;
1438 struct timeval work_end;
1446 spin_lock(&svc->srv_lock);
1447 if (unlikely(!ptlrpc_server_request_pending(svc, 0) ||
1450 /* !@%$# liblustre only has 1 thread */
1451 atomic_read(&svc->srv_n_difficult_replies) != 0 &&
1453 svc->srv_n_active_reqs >= (svc->srv_threads_running - 1)))) {
1454 /* Don't handle regular requests in the last thread, in order * re
1455 * to handle difficult replies (which might block other threads)
1456 * as well as handle any incoming reqs, early replies, etc.
1457 * That means we always need at least 2 service threads. */
1458 spin_unlock(&svc->srv_lock);
1462 request = ptlrpc_server_request_get(svc, 0);
1463 if (request == NULL) {
1464 spin_unlock(&svc->srv_lock);
1468 opc = lustre_msg_get_opc(request->rq_reqmsg);
1469 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
1470 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
1471 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
1472 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
1474 if (unlikely(fail_opc)) {
1475 if (request->rq_export && request->rq_ops) {
1476 spin_unlock(&svc->srv_lock);
1477 OBD_FAIL_TIMEOUT(fail_opc, 4);
1478 spin_lock(&svc->srv_lock);
1479 request = ptlrpc_server_request_get(svc, 0);
1480 if (request == NULL) {
1481 spin_unlock(&svc->srv_lock);
1484 LASSERT(ptlrpc_server_request_pending(svc, 0));
1488 list_del_init(&request->rq_list);
1489 svc->srv_n_queued_reqs--;
1490 svc->srv_n_active_reqs++;
1494 /* The phase is changed under the lock here because we need to know
1495 * the request is under processing (see ptlrpc_hpreq_reorder()). */
1496 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
1497 spin_unlock(&svc->srv_lock);
1499 ptlrpc_hpreq_fini(request);
1501 if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
1502 libcfs_debug_dumplog();
1504 do_gettimeofday(&work_start);
1505 timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
1506 if (likely(svc->srv_stats != NULL)) {
1507 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
1509 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
1510 svc->srv_n_queued_reqs);
1511 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
1512 svc->srv_n_active_reqs);
1513 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
1514 at_get(&svc->srv_at_estimate));
1517 rc = lu_context_init(&request->rq_session,
1518 LCT_SESSION|LCT_REMEMBER|LCT_NOREF);
1520 CERROR("Failure to initialize session: %d\n", rc);
1523 request->rq_session.lc_thread = thread;
1524 request->rq_session.lc_cookie = 0x5;
1525 lu_context_enter(&request->rq_session);
1527 CDEBUG(D_NET, "got req "LPU64"\n", request->rq_xid);
1529 request->rq_svc_thread = thread;
1531 request->rq_svc_thread->t_env->le_ses = &request->rq_session;
1533 if (likely(request->rq_export)) {
1534 if (unlikely(ptlrpc_check_req(request)))
1536 ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
1537 export = class_export_rpc_get(request->rq_export);
1540 /* Discard requests queued for longer than the deadline.
1541 The deadline is increased if we send an early reply. */
1542 if (cfs_time_current_sec() > request->rq_deadline) {
1543 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s"
1544 ": deadline "CFS_DURATION_T":"CFS_DURATION_T"s ago\n",
1545 libcfs_id2str(request->rq_peer),
1546 cfs_time_sub(request->rq_deadline,
1547 request->rq_arrival_time.tv_sec),
1548 cfs_time_sub(cfs_time_current_sec(),
1549 request->rq_deadline));
1550 goto put_rpc_export;
1553 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
1554 "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1555 (request->rq_export ?
1556 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1557 (request->rq_export ?
1558 atomic_read(&request->rq_export->exp_refcount) : -99),
1559 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1560 libcfs_id2str(request->rq_peer),
1561 lustre_msg_get_opc(request->rq_reqmsg));
1563 OBD_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, obd_fail_val);
1565 rc = svc->srv_handler(request);
1567 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
1569 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
1570 "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1571 (request->rq_export ?
1572 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1573 (request->rq_export ?
1574 atomic_read(&request->rq_export->exp_refcount) : -99),
1575 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1576 libcfs_id2str(request->rq_peer),
1577 lustre_msg_get_opc(request->rq_reqmsg));
1581 class_export_rpc_put(export);
1583 lu_context_exit(&request->rq_session);
1584 lu_context_fini(&request->rq_session);
1586 if (unlikely(cfs_time_current_sec() > request->rq_deadline)) {
1587 DEBUG_REQ(D_WARNING, request, "Request x"LPU64" took longer "
1588 "than estimated ("CFS_DURATION_T":"CFS_DURATION_T"s);"
1589 " client may timeout.",
1590 request->rq_xid, cfs_time_sub(request->rq_deadline,
1591 request->rq_arrival_time.tv_sec),
1592 cfs_time_sub(cfs_time_current_sec(),
1593 request->rq_deadline));
1596 do_gettimeofday(&work_end);
1597 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1598 CDEBUG(D_RPCTRACE, "request x"LPU64" opc %u from %s processed in "
1599 "%ldus (%ldus total) trans "LPU64" rc %d/%d\n",
1600 request->rq_xid, lustre_msg_get_opc(request->rq_reqmsg),
1601 libcfs_id2str(request->rq_peer), timediff,
1602 cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL),
1603 request->rq_repmsg ? lustre_msg_get_transno(request->rq_repmsg) :
1604 request->rq_transno, request->rq_status,
1605 request->rq_repmsg ? lustre_msg_get_status(request->rq_repmsg):
1607 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
1608 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
1609 int opc = opcode_offset(op);
1610 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
1611 LASSERT(opc < LUSTRE_MAX_OPCODES);
1612 lprocfs_counter_add(svc->srv_stats,
1613 opc + EXTRA_MAX_OPCODES,
1617 if (unlikely(request->rq_early_count)) {
1618 DEBUG_REQ(D_ADAPTTO, request,
1619 "sent %d early replies before finishing in "
1621 request->rq_early_count,
1622 cfs_time_sub(work_end.tv_sec,
1623 request->rq_arrival_time.tv_sec));
1627 spin_lock(&svc->srv_lock);
1630 spin_unlock(&svc->srv_lock);
1631 ptlrpc_server_finish_request(request);
1637 * An internal function to process a single reply state object.
1640 ptlrpc_handle_rs (struct ptlrpc_reply_state *rs)
1642 struct ptlrpc_service *svc = rs->rs_service;
1643 struct obd_export *exp;
1644 struct obd_device *obd;
1649 exp = rs->rs_export;
1652 LASSERT (rs->rs_difficult);
1653 LASSERT (rs->rs_scheduled);
1654 LASSERT (list_empty(&rs->rs_list));
1656 spin_lock (&exp->exp_lock);
1657 /* Noop if removed already */
1658 list_del_init (&rs->rs_exp_list);
1659 spin_unlock (&exp->exp_lock);
1661 /* Avoid exp_uncommitted_replies_lock contention if we 100% sure that
1662 * rs has been removed from the list already */
1663 if (!list_empty_careful(&rs->rs_obd_list)) {
1664 spin_lock(&exp->exp_uncommitted_replies_lock);
1665 list_del_init(&rs->rs_obd_list);
1666 spin_unlock(&exp->exp_uncommitted_replies_lock);
1669 spin_lock(&rs->rs_lock);
1671 been_handled = rs->rs_handled;
1674 nlocks = rs->rs_nlocks; /* atomic "steal", but */
1675 rs->rs_nlocks = 0; /* locks still on rs_locks! */
1677 if (nlocks == 0 && !been_handled) {
1678 /* If we see this, we should already have seen the warning
1679 * in mds_steal_ack_locks() */
1680 CWARN("All locks stolen from rs %p x"LPD64".t"LPD64
1683 rs->rs_xid, rs->rs_transno,
1684 lustre_msg_get_opc(rs->rs_msg),
1685 libcfs_nid2str(exp->exp_connection->c_peer.nid));
1688 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
1689 spin_unlock(&rs->rs_lock);
1691 if (!been_handled && rs->rs_on_net) {
1692 LNetMDUnlink(rs->rs_md_h);
1693 /* Ignore return code; we're racing with
1697 while (nlocks-- > 0)
1698 ldlm_lock_decref(&rs->rs_locks[nlocks],
1699 rs->rs_modes[nlocks]);
1701 spin_lock(&rs->rs_lock);
1704 rs->rs_scheduled = 0;
1706 if (!rs->rs_on_net) {
1708 spin_unlock(&rs->rs_lock);
1710 class_export_put (exp);
1711 rs->rs_export = NULL;
1712 ptlrpc_rs_decref (rs);
1713 atomic_dec (&svc->srv_outstanding_replies);
1714 if (atomic_dec_and_test(&svc->srv_n_difficult_replies) &&
1715 svc->srv_is_stopping)
1716 cfs_waitq_broadcast(&svc->srv_waitq);
1720 /* still on the net; callback will schedule */
1721 spin_unlock(&rs->rs_lock);
1728 * Check whether given service has a reply available for processing
1731 * \param svc a ptlrpc service
1732 * \retval 0 no replies processes
1733 * \retval 1 one reply processed
1736 ptlrpc_server_handle_reply(struct ptlrpc_service *svc)
1738 struct ptlrpc_reply_state *rs = NULL;
1741 spin_lock(&svc->srv_lock);
1742 if (!list_empty(&svc->srv_reply_queue)) {
1743 rs = list_entry(svc->srv_reply_queue.prev,
1744 struct ptlrpc_reply_state,
1746 list_del_init(&rs->rs_list);
1748 spin_unlock(&svc->srv_lock);
1750 ptlrpc_handle_rs(rs);
1754 /* FIXME make use of timeout later */
1756 liblustre_check_services (void *arg)
1758 int did_something = 0;
1760 struct list_head *tmp, *nxt;
1763 /* I'm relying on being single threaded, not to have to lock
1764 * ptlrpc_all_services etc */
1765 list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
1766 struct ptlrpc_service *svc =
1767 list_entry (tmp, struct ptlrpc_service, srv_list);
1769 if (svc->srv_threads_running != 0) /* I've recursed */
1772 /* service threads can block for bulk, so this limits us
1773 * (arbitrarily) to recursing 1 stack frame per service.
1774 * Note that the problem with recursion is that we have to
1775 * unwind completely before our caller can resume. */
1777 svc->srv_threads_running++;
1780 rc = ptlrpc_server_handle_req_in(svc);
1781 rc |= ptlrpc_server_handle_reply(svc);
1782 rc |= ptlrpc_at_check_timed(svc);
1783 rc |= ptlrpc_server_handle_request(svc, NULL);
1784 rc |= (ptlrpc_server_post_idle_rqbds(svc) > 0);
1785 did_something |= rc;
1788 svc->srv_threads_running--;
1791 RETURN(did_something);
1793 #define ptlrpc_stop_all_threads(s) do {} while (0)
1795 #else /* __KERNEL__ */
1797 /* Don't use daemonize, it removes fs struct from new thread (bug 418) */
1798 void ptlrpc_daemonize(char *name)
1800 struct fs_struct *fs = current->fs;
1802 atomic_inc(&fs->count);
1803 cfs_daemonize(name);
1804 exit_fs(cfs_current());
1806 ll_set_fs_pwd(current->fs, init_task.fs->pwdmnt, init_task.fs->pwd);
1810 ptlrpc_check_rqbd_pool(struct ptlrpc_service *svc)
1812 int avail = svc->srv_nrqbd_receiving;
1813 int low_water = test_req_buffer_pressure ? 0 :
1814 svc->srv_nbuf_per_group/2;
1816 /* NB I'm not locking; just looking. */
1818 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
1819 * allowed the request history to grow out of control. We could put a
1820 * sanity check on that here and cull some history if we need the
1823 if (avail <= low_water)
1824 ptlrpc_grow_req_bufs(svc);
1827 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQBUF_AVAIL_CNTR,
1832 ptlrpc_retry_rqbds(void *arg)
1834 struct ptlrpc_service *svc = (struct ptlrpc_service *)arg;
1836 svc->srv_rqbd_timeout = 0;
1837 return (-ETIMEDOUT);
1840 static int ptlrpc_main(void *arg)
1842 struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
1843 struct ptlrpc_service *svc = data->svc;
1844 struct ptlrpc_thread *thread = data->thread;
1845 struct obd_device *dev = data->dev;
1846 struct ptlrpc_reply_state *rs;
1847 #ifdef WITH_GROUP_INFO
1848 struct group_info *ginfo = NULL;
1851 int counter = 0, rc = 0;
1854 ptlrpc_daemonize(data->name);
1856 #if defined(HAVE_NODE_TO_CPUMASK) && defined(CONFIG_NUMA)
1857 /* we need to do this before any per-thread allocation is done so that
1858 * we get the per-thread allocations on local node. bug 7342 */
1859 if (svc->srv_cpu_affinity) {
1862 for (cpu = 0, num_cpu = 0; cpu < num_possible_cpus(); cpu++) {
1863 if (!cpu_online(cpu))
1865 if (num_cpu == thread->t_id % num_online_cpus())
1869 set_cpus_allowed(cfs_current(), node_to_cpumask(cpu_to_node(cpu)));
1873 #ifdef WITH_GROUP_INFO
1874 ginfo = groups_alloc(0);
1880 set_current_groups(ginfo);
1881 put_group_info(ginfo);
1884 if (svc->srv_init != NULL) {
1885 rc = svc->srv_init(thread);
1890 rc = lu_context_init(&env.le_ctx,
1891 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
1895 thread->t_env = &env;
1896 env.le_ctx.lc_thread = thread;
1897 env.le_ctx.lc_cookie = 0x6;
1899 /* Alloc reply state structure for this one */
1900 OBD_ALLOC_GFP(rs, svc->srv_max_reply_size, CFS_ALLOC_STD);
1906 /* Record that the thread is running */
1907 thread->t_flags = SVC_RUNNING;
1909 * wake up our creator. Note: @data is invalid after this point,
1910 * because it's allocated on ptlrpc_start_thread() stack.
1912 cfs_waitq_signal(&thread->t_ctl_waitq);
1914 thread->t_watchdog = lc_watchdog_add(GET_TIMEOUT(svc), NULL, NULL);
1916 spin_lock(&svc->srv_lock);
1917 svc->srv_threads_running++;
1918 list_add(&rs->rs_list, &svc->srv_free_rs_list);
1919 spin_unlock(&svc->srv_lock);
1920 cfs_waitq_signal(&svc->srv_free_rs_waitq);
1922 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
1923 svc->srv_threads_running);
1925 /* XXX maintain a list of all managed devices: insert here */
1927 while ((thread->t_flags & SVC_STOPPING) == 0) {
1928 /* Don't exit while there are replies to be handled */
1929 struct l_wait_info lwi = LWI_TIMEOUT(svc->srv_rqbd_timeout,
1930 ptlrpc_retry_rqbds, svc);
1932 lc_watchdog_disable(thread->t_watchdog);
1936 l_wait_event_exclusive (svc->srv_waitq,
1937 ((thread->t_flags & SVC_STOPPING) != 0) ||
1938 (!list_empty(&svc->srv_idle_rqbds) &&
1939 svc->srv_rqbd_timeout == 0) ||
1940 !list_empty(&svc->srv_req_in_queue) ||
1941 (ptlrpc_server_request_pending(svc, 0) &&
1942 (svc->srv_n_active_reqs <
1943 (svc->srv_threads_running - 1))) ||
1947 lc_watchdog_touch(thread->t_watchdog, GET_TIMEOUT(svc));
1949 ptlrpc_check_rqbd_pool(svc);
1951 if ((svc->srv_threads_started < svc->srv_threads_max) &&
1952 (svc->srv_n_active_reqs >= (svc->srv_threads_started - 1))){
1953 /* Ignore return code - we tried... */
1954 ptlrpc_start_thread(dev, svc);
1957 if (!list_empty(&svc->srv_req_in_queue)) {
1958 /* Process all incoming reqs before handling any */
1959 ptlrpc_server_handle_req_in(svc);
1960 /* but limit ourselves in case of flood */
1961 if (counter++ < 1000)
1966 if (svc->srv_at_check)
1967 ptlrpc_at_check_timed(svc);
1969 /* don't handle requests in the last thread */
1970 if (ptlrpc_server_request_pending(svc, 0) &&
1971 (svc->srv_n_active_reqs < (svc->srv_threads_running - 1))) {
1972 lu_context_enter(&env.le_ctx);
1973 ptlrpc_server_handle_request(svc, thread);
1974 lu_context_exit(&env.le_ctx);
1977 if (!list_empty(&svc->srv_idle_rqbds) &&
1978 ptlrpc_server_post_idle_rqbds(svc) < 0) {
1979 /* I just failed to repost request buffers. Wait
1980 * for a timeout (unless something else happens)
1981 * before I try again */
1982 svc->srv_rqbd_timeout = cfs_time_seconds(1)/10;
1983 CDEBUG(D_RPCTRACE,"Posted buffers: %d\n",
1984 svc->srv_nrqbd_receiving);
1988 lc_watchdog_delete(thread->t_watchdog);
1989 thread->t_watchdog = NULL;
1993 * deconstruct service specific state created by ptlrpc_start_thread()
1995 if (svc->srv_done != NULL)
1996 svc->srv_done(thread);
1998 lu_context_fini(&env.le_ctx);
2000 CDEBUG(D_NET, "service thread %d exiting: rc %d\n", thread->t_id, rc);
2002 spin_lock(&svc->srv_lock);
2003 svc->srv_threads_running--; /* must know immediately */
2005 thread->t_flags = SVC_STOPPED;
2007 cfs_waitq_signal(&thread->t_ctl_waitq);
2008 spin_unlock(&svc->srv_lock);
2013 struct ptlrpc_hr_args {
2016 struct ptlrpc_hr_service *hrs;
2019 static int hrt_dont_sleep(struct ptlrpc_hr_thread *t,
2020 struct list_head *replies)
2024 spin_lock(&t->hrt_lock);
2025 list_splice_init(&t->hrt_queue, replies);
2026 result = test_bit(HRT_STOPPING, &t->hrt_flags) ||
2027 !list_empty(replies);
2028 spin_unlock(&t->hrt_lock);
2032 static int ptlrpc_hr_main(void *arg)
2034 struct ptlrpc_hr_args * hr_args = arg;
2035 struct ptlrpc_hr_service *hr = hr_args->hrs;
2036 struct ptlrpc_hr_thread *t = &hr->hr_threads[hr_args->thread_index];
2037 char threadname[20];
2038 CFS_LIST_HEAD(replies);
2040 snprintf(threadname, sizeof(threadname),
2041 "ptlrpc_hr_%d", hr_args->thread_index);
2043 ptlrpc_daemonize(threadname);
2044 #if defined(HAVE_NODE_TO_CPUMASK)
2045 set_cpus_allowed(cfs_current(),
2046 node_to_cpumask(cpu_to_node(hr_args->cpu_index)));
2048 set_bit(HRT_RUNNING, &t->hrt_flags);
2049 cfs_waitq_signal(&t->hrt_wait);
2051 while (!test_bit(HRT_STOPPING, &t->hrt_flags)) {
2053 cfs_wait_event(t->hrt_wait, hrt_dont_sleep(t, &replies));
2054 while (!list_empty(&replies)) {
2055 struct ptlrpc_reply_state *rs;
2057 rs = list_entry(replies.prev,
2058 struct ptlrpc_reply_state,
2060 list_del_init(&rs->rs_list);
2061 ptlrpc_handle_rs(rs);
2065 clear_bit(HRT_RUNNING, &t->hrt_flags);
2066 complete(&t->hrt_completion);
2071 static int ptlrpc_start_hr_thread(struct ptlrpc_hr_service *hr, int n, int cpu)
2073 struct ptlrpc_hr_thread *t = &hr->hr_threads[n];
2074 struct ptlrpc_hr_args args;
2078 args.thread_index = n;
2079 args.cpu_index = cpu;
2082 rc = cfs_kernel_thread(ptlrpc_hr_main, (void*)&args,
2083 CLONE_VM|CLONE_FILES);
2085 complete(&t->hrt_completion);
2088 cfs_wait_event(t->hrt_wait, test_bit(HRT_RUNNING, &t->hrt_flags));
2094 static void ptlrpc_stop_hr_thread(struct ptlrpc_hr_thread *t)
2098 set_bit(HRT_STOPPING, &t->hrt_flags);
2099 cfs_waitq_signal(&t->hrt_wait);
2100 wait_for_completion(&t->hrt_completion);
2105 static void ptlrpc_stop_hr_threads(struct ptlrpc_hr_service *hrs)
2110 for (n = 0; n < hrs->hr_n_threads; n++)
2111 ptlrpc_stop_hr_thread(&hrs->hr_threads[n]);
2116 static int ptlrpc_start_hr_threads(struct ptlrpc_hr_service *hr)
2119 int n, cpu, threads_started = 0;
2122 LASSERT(hr != NULL);
2123 LASSERT(hr->hr_n_threads > 0);
2125 for (n = 0, cpu = 0; n < hr->hr_n_threads; n++) {
2126 #if defined(HAVE_NODE_TO_CPUMASK)
2127 while(!cpu_online(cpu)) {
2129 if (cpu >= num_possible_cpus())
2133 rc = ptlrpc_start_hr_thread(hr, n, cpu);
2139 if (threads_started == 0) {
2140 CERROR("No reply handling threads started\n");
2143 if (threads_started < hr->hr_n_threads) {
2144 CWARN("Started only %d reply handling threads from %d\n",
2145 threads_started, hr->hr_n_threads);
2146 hr->hr_n_threads = threads_started;
2151 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
2152 struct ptlrpc_thread *thread)
2154 struct l_wait_info lwi = { 0 };
2157 CDEBUG(D_RPCTRACE, "Stopping thread %p\n", thread);
2158 spin_lock(&svc->srv_lock);
2159 thread->t_flags = SVC_STOPPING;
2160 spin_unlock(&svc->srv_lock);
2162 cfs_waitq_broadcast(&svc->srv_waitq);
2163 l_wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED),
2166 spin_lock(&svc->srv_lock);
2167 list_del(&thread->t_link);
2168 spin_unlock(&svc->srv_lock);
2170 OBD_FREE_PTR(thread);
2174 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2176 struct ptlrpc_thread *thread;
2179 spin_lock(&svc->srv_lock);
2180 while (!list_empty(&svc->srv_threads)) {
2181 thread = list_entry(svc->srv_threads.next,
2182 struct ptlrpc_thread, t_link);
2184 spin_unlock(&svc->srv_lock);
2185 ptlrpc_stop_thread(svc, thread);
2186 spin_lock(&svc->srv_lock);
2189 spin_unlock(&svc->srv_lock);
2193 int ptlrpc_start_threads(struct obd_device *dev, struct ptlrpc_service *svc)
2198 /* We require 2 threads min - see note in
2199 ptlrpc_server_handle_request */
2200 LASSERT(svc->srv_threads_min >= 2);
2201 for (i = 0; i < svc->srv_threads_min; i++) {
2202 rc = ptlrpc_start_thread(dev, svc);
2203 /* We have enough threads, don't start more. b=15759 */
2207 CERROR("cannot start %s thread #%d: rc %d\n",
2208 svc->srv_thread_name, i, rc);
2209 ptlrpc_stop_all_threads(svc);
2215 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc)
2217 struct l_wait_info lwi = { 0 };
2218 struct ptlrpc_svc_data d;
2219 struct ptlrpc_thread *thread;
2224 CDEBUG(D_RPCTRACE, "%s started %d min %d max %d running %d\n",
2225 svc->srv_name, svc->srv_threads_started, svc->srv_threads_min,
2226 svc->srv_threads_max, svc->srv_threads_running);
2227 if (unlikely(svc->srv_threads_started >= svc->srv_threads_max) ||
2228 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
2229 svc->srv_threads_started == svc->srv_threads_min - 1))
2232 OBD_ALLOC_PTR(thread);
2235 cfs_waitq_init(&thread->t_ctl_waitq);
2237 spin_lock(&svc->srv_lock);
2238 if (svc->srv_threads_started >= svc->srv_threads_max) {
2239 spin_unlock(&svc->srv_lock);
2240 OBD_FREE_PTR(thread);
2243 list_add(&thread->t_link, &svc->srv_threads);
2244 id = svc->srv_threads_started++;
2245 spin_unlock(&svc->srv_lock);
2247 thread->t_svc = svc;
2249 sprintf(name, "%s_%02d", svc->srv_thread_name, id);
2255 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", name);
2257 /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
2258 * just drop the VM and FILES in ptlrpc_daemonize() right away.
2260 rc = cfs_kernel_thread(ptlrpc_main, &d, CLONE_VM | CLONE_FILES);
2262 CERROR("cannot start thread '%s': rc %d\n", name, rc);
2264 spin_lock(&svc->srv_lock);
2265 list_del(&thread->t_link);
2266 --svc->srv_threads_started;
2267 spin_unlock(&svc->srv_lock);
2269 OBD_FREE(thread, sizeof(*thread));
2272 l_wait_event(thread->t_ctl_waitq,
2273 thread->t_flags & (SVC_RUNNING | SVC_STOPPED), &lwi);
2275 rc = (thread->t_flags & SVC_STOPPED) ? thread->t_id : 0;
2280 int ptlrpc_hr_init(void)
2283 int n_cpus = num_online_cpus();
2284 struct ptlrpc_hr_service *hr;
2288 LASSERT(ptlrpc_hr == NULL);
2290 size = offsetof(struct ptlrpc_hr_service, hr_threads[n_cpus]);
2291 OBD_ALLOC(hr, size);
2294 for (i = 0; i < n_cpus; i++) {
2295 struct ptlrpc_hr_thread *t = &hr->hr_threads[i];
2297 spin_lock_init(&t->hrt_lock);
2298 cfs_waitq_init(&t->hrt_wait);
2299 CFS_INIT_LIST_HEAD(&t->hrt_queue);
2300 init_completion(&t->hrt_completion);
2302 hr->hr_n_threads = n_cpus;
2306 RETURN(ptlrpc_start_hr_threads(hr));
2309 void ptlrpc_hr_fini(void)
2311 if (ptlrpc_hr != NULL) {
2312 ptlrpc_stop_hr_threads(ptlrpc_hr);
2313 OBD_FREE(ptlrpc_hr, ptlrpc_hr->hr_size);
2318 #endif /* __KERNEL__ */
2321 * Wait until all already scheduled replies are processed.
2323 static void ptlrpc_wait_replies(struct ptlrpc_service *svc)
2327 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
2329 rc = l_wait_event(svc->srv_waitq,
2330 atomic_read(&svc->srv_n_difficult_replies) == 0,
2334 CWARN("Unexpectedly long timeout %p\n", svc);
2338 int ptlrpc_unregister_service(struct ptlrpc_service *service)
2341 struct l_wait_info lwi;
2342 struct list_head *tmp;
2343 struct ptlrpc_reply_state *rs, *t;
2344 struct ptlrpc_at_array *array = &service->srv_at_array;
2347 service->srv_is_stopping = 1;
2348 cfs_timer_disarm(&service->srv_at_timer);
2350 ptlrpc_stop_all_threads(service);
2351 LASSERT(list_empty(&service->srv_threads));
2353 spin_lock (&ptlrpc_all_services_lock);
2354 list_del_init (&service->srv_list);
2355 spin_unlock (&ptlrpc_all_services_lock);
2357 ptlrpc_lprocfs_unregister_service(service);
2359 /* All history will be culled when the next request buffer is
2361 service->srv_max_history_rqbds = 0;
2363 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
2365 rc = LNetClearLazyPortal(service->srv_req_portal);
2368 /* Unlink all the request buffers. This forces a 'final' event with
2369 * its 'unlink' flag set for each posted rqbd */
2370 list_for_each(tmp, &service->srv_active_rqbds) {
2371 struct ptlrpc_request_buffer_desc *rqbd =
2372 list_entry(tmp, struct ptlrpc_request_buffer_desc,
2375 rc = LNetMDUnlink(rqbd->rqbd_md_h);
2376 LASSERT (rc == 0 || rc == -ENOENT);
2379 /* Wait for the network to release any buffers it's currently
2382 spin_lock(&service->srv_lock);
2383 rc = service->srv_nrqbd_receiving;
2384 spin_unlock(&service->srv_lock);
2389 /* Network access will complete in finite time but the HUGE
2390 * timeout lets us CWARN for visibility of sluggish NALs */
2391 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2392 cfs_time_seconds(1), NULL, NULL);
2393 rc = l_wait_event(service->srv_waitq,
2394 service->srv_nrqbd_receiving == 0,
2396 if (rc == -ETIMEDOUT)
2397 CWARN("Service %s waiting for request buffers\n",
2401 /* schedule all outstanding replies to terminate them */
2402 spin_lock(&service->srv_lock);
2403 while (!list_empty(&service->srv_active_replies)) {
2404 struct ptlrpc_reply_state *rs =
2405 list_entry(service->srv_active_replies.next,
2406 struct ptlrpc_reply_state, rs_list);
2407 spin_lock(&rs->rs_lock);
2408 ptlrpc_schedule_difficult_reply(rs);
2409 spin_unlock(&rs->rs_lock);
2411 spin_unlock(&service->srv_lock);
2413 /* purge the request queue. NB No new replies (rqbds all unlinked)
2414 * and no service threads, so I'm the only thread noodling the
2415 * request queue now */
2416 while (!list_empty(&service->srv_req_in_queue)) {
2417 struct ptlrpc_request *req =
2418 list_entry(service->srv_req_in_queue.next,
2419 struct ptlrpc_request,
2422 list_del(&req->rq_list);
2423 service->srv_n_queued_reqs--;
2424 service->srv_n_active_reqs++;
2425 ptlrpc_server_finish_request(req);
2427 while (ptlrpc_server_request_pending(service, 1)) {
2428 struct ptlrpc_request *req;
2430 req = ptlrpc_server_request_get(service, 1);
2431 list_del(&req->rq_list);
2432 service->srv_n_queued_reqs--;
2433 service->srv_n_active_reqs++;
2434 ptlrpc_hpreq_fini(req);
2435 ptlrpc_server_finish_request(req);
2437 LASSERT(service->srv_n_queued_reqs == 0);
2438 LASSERT(service->srv_n_active_reqs == 0);
2439 LASSERT(service->srv_n_history_rqbds == 0);
2440 LASSERT(list_empty(&service->srv_active_rqbds));
2442 /* Now free all the request buffers since nothing references them
2444 while (!list_empty(&service->srv_idle_rqbds)) {
2445 struct ptlrpc_request_buffer_desc *rqbd =
2446 list_entry(service->srv_idle_rqbds.next,
2447 struct ptlrpc_request_buffer_desc,
2450 ptlrpc_free_rqbd(rqbd);
2453 ptlrpc_wait_replies(service);
2455 list_for_each_entry_safe(rs, t, &service->srv_free_rs_list, rs_list) {
2456 list_del(&rs->rs_list);
2457 OBD_FREE(rs, service->srv_max_reply_size);
2460 /* In case somebody rearmed this in the meantime */
2461 cfs_timer_disarm(&service->srv_at_timer);
2463 if (array->paa_reqs_array != NULL) {
2464 OBD_FREE(array->paa_reqs_array,
2465 sizeof(struct list_head) * array->paa_size);
2466 array->paa_reqs_array = NULL;
2469 if (array->paa_reqs_count != NULL) {
2470 OBD_FREE(array->paa_reqs_count,
2471 sizeof(__u32) * array->paa_size);
2472 array->paa_reqs_count= NULL;
2475 OBD_FREE_PTR(service);
2479 /* Returns 0 if the service is healthy.
2481 * Right now, it just checks to make sure that requests aren't languishing
2482 * in the queue. We'll use this health check to govern whether a node needs
2483 * to be shot, so it's intentionally non-aggressive. */
2484 int ptlrpc_service_health_check(struct ptlrpc_service *svc)
2486 struct ptlrpc_request *request;
2487 struct timeval right_now;
2493 do_gettimeofday(&right_now);
2495 spin_lock(&svc->srv_lock);
2496 if (!ptlrpc_server_request_pending(svc, 1)) {
2497 spin_unlock(&svc->srv_lock);
2501 /* How long has the next entry been waiting? */
2502 if (list_empty(&svc->srv_request_queue))
2503 request = list_entry(svc->srv_request_hpq.next,
2504 struct ptlrpc_request, rq_list);
2506 request = list_entry(svc->srv_request_queue.next,
2507 struct ptlrpc_request, rq_list);
2508 timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
2509 spin_unlock(&svc->srv_lock);
2511 if ((timediff / ONE_MILLION) > (AT_OFF ? obd_timeout * 3/2 :
2513 CERROR("%s: unhealthy - request has been waiting %lds\n",
2514 svc->srv_name, timediff / ONE_MILLION);