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 if (req->rq_export->exp_disconnected) {
175 ldlm_lock_decref(lock, mode);
177 idx = rs->rs_nlocks++;
178 rs->rs_locks[idx] = *lock;
179 rs->rs_modes[idx] = mode;
180 rs->rs_difficult = 1;
181 rs->rs_no_ack = !!no_ack;
187 #define HRT_RUNNING 0
188 #define HRT_STOPPING 1
190 struct ptlrpc_hr_thread {
192 unsigned long hrt_flags;
193 cfs_waitq_t hrt_wait;
194 struct list_head hrt_queue;
195 struct completion hrt_completion;
198 struct ptlrpc_hr_service {
202 struct ptlrpc_hr_thread hr_threads[0];
206 struct list_head rsb_replies;
207 struct ptlrpc_service *rsb_svc;
208 unsigned int rsb_n_replies;
212 * A pointer to per-node reply handling service.
214 static struct ptlrpc_hr_service *ptlrpc_hr = NULL;
217 * maximum mumber of replies scheduled in one batch
219 #define MAX_SCHEDULED 256
222 * Initialize a reply batch.
226 static void rs_batch_init(struct rs_batch *b)
228 memset(b, 0, sizeof *b);
229 CFS_INIT_LIST_HEAD(&b->rsb_replies);
233 * Dispatch all replies accumulated in the batch to one from
234 * dedicated reply handling threads.
238 static void rs_batch_dispatch(struct rs_batch *b)
240 if (b->rsb_n_replies != 0) {
241 struct ptlrpc_hr_service *hr = ptlrpc_hr;
244 idx = hr->hr_index++;
245 if (hr->hr_index >= hr->hr_n_threads)
248 spin_lock(&hr->hr_threads[idx].hrt_lock);
249 list_splice_init(&b->rsb_replies,
250 &hr->hr_threads[idx].hrt_queue);
251 spin_unlock(&hr->hr_threads[idx].hrt_lock);
252 cfs_waitq_signal(&hr->hr_threads[idx].hrt_wait);
253 b->rsb_n_replies = 0;
258 * Add a reply to a batch.
259 * Add one reply object to a batch, schedule batched replies if overload.
264 static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs)
266 struct ptlrpc_service *svc = rs->rs_service;
268 if (svc != b->rsb_svc || b->rsb_n_replies >= MAX_SCHEDULED) {
269 if (b->rsb_svc != NULL) {
270 rs_batch_dispatch(b);
271 spin_unlock(&b->rsb_svc->srv_lock);
273 spin_lock(&svc->srv_lock);
276 spin_lock(&rs->rs_lock);
277 rs->rs_scheduled_ever = 1;
278 if (rs->rs_scheduled == 0) {
279 list_move(&rs->rs_list, &b->rsb_replies);
280 rs->rs_scheduled = 1;
283 spin_unlock(&rs->rs_lock);
287 * Reply batch finalization.
288 * Dispatch remaining replies from the batch
289 * and release remaining spinlock.
293 static void rs_batch_fini(struct rs_batch *b)
295 if (b->rsb_svc != 0) {
296 rs_batch_dispatch(b);
297 spin_unlock(&b->rsb_svc->srv_lock);
301 #define DECLARE_RS_BATCH(b) struct rs_batch b
303 #else /* __KERNEL__ */
305 #define rs_batch_init(b) do{}while(0)
306 #define rs_batch_fini(b) do{}while(0)
307 #define rs_batch_add(b, r) ptlrpc_schedule_difficult_reply(r)
308 #define DECLARE_RS_BATCH(b)
310 #endif /* __KERNEL__ */
312 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
315 struct ptlrpc_hr_service *hr = ptlrpc_hr;
319 LASSERT(list_empty(&rs->rs_list));
321 idx = hr->hr_index++;
322 if (hr->hr_index >= hr->hr_n_threads)
324 spin_lock(&hr->hr_threads[idx].hrt_lock);
325 list_add_tail(&rs->rs_list, &hr->hr_threads[idx].hrt_queue);
326 spin_unlock(&hr->hr_threads[idx].hrt_lock);
327 cfs_waitq_signal(&hr->hr_threads[idx].hrt_wait);
330 list_add_tail(&rs->rs_list, &rs->rs_service->srv_reply_queue);
335 ptlrpc_schedule_difficult_reply (struct ptlrpc_reply_state *rs)
339 LASSERT_SPIN_LOCKED(&rs->rs_service->srv_lock);
340 LASSERT_SPIN_LOCKED(&rs->rs_lock);
341 LASSERT (rs->rs_difficult);
342 rs->rs_scheduled_ever = 1; /* flag any notification attempt */
344 if (rs->rs_scheduled) { /* being set up or already notified */
349 rs->rs_scheduled = 1;
350 list_del_init(&rs->rs_list);
351 ptlrpc_dispatch_difficult_reply(rs);
355 void ptlrpc_commit_replies(struct obd_export *exp)
357 struct ptlrpc_reply_state *rs, *nxt;
358 DECLARE_RS_BATCH(batch);
361 rs_batch_init(&batch);
362 /* Find any replies that have been committed and get their service
363 * to attend to complete them. */
365 /* CAVEAT EMPTOR: spinlock ordering!!! */
366 spin_lock(&exp->exp_uncommitted_replies_lock);
367 list_for_each_entry_safe(rs, nxt, &exp->exp_uncommitted_replies,
369 LASSERT (rs->rs_difficult);
370 /* VBR: per-export last_committed */
371 LASSERT(rs->rs_export);
372 if (rs->rs_transno <= exp->exp_last_committed) {
373 list_del_init(&rs->rs_obd_list);
374 rs_batch_add(&batch, rs);
377 spin_unlock(&exp->exp_uncommitted_replies_lock);
378 rs_batch_fini(&batch);
383 ptlrpc_server_post_idle_rqbds (struct ptlrpc_service *svc)
385 struct ptlrpc_request_buffer_desc *rqbd;
390 spin_lock(&svc->srv_lock);
392 if (list_empty (&svc->srv_idle_rqbds)) {
393 spin_unlock(&svc->srv_lock);
397 rqbd = list_entry(svc->srv_idle_rqbds.next,
398 struct ptlrpc_request_buffer_desc,
400 list_del (&rqbd->rqbd_list);
402 /* assume we will post successfully */
403 svc->srv_nrqbd_receiving++;
404 list_add (&rqbd->rqbd_list, &svc->srv_active_rqbds);
406 spin_unlock(&svc->srv_lock);
408 rc = ptlrpc_register_rqbd(rqbd);
415 spin_lock(&svc->srv_lock);
417 svc->srv_nrqbd_receiving--;
418 list_del(&rqbd->rqbd_list);
419 list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
421 /* Don't complain if no request buffers are posted right now; LNET
422 * won't drop requests because we set the portal lazy! */
424 spin_unlock(&svc->srv_lock);
429 struct ptlrpc_service *ptlrpc_init_svc_conf(struct ptlrpc_service_conf *c,
430 svc_handler_t h, char *name,
431 struct proc_dir_entry *proc_entry,
432 svcreq_printfn_t prntfn,
435 return ptlrpc_init_svc(c->psc_nbufs, c->psc_bufsize,
436 c->psc_max_req_size, c->psc_max_reply_size,
437 c->psc_req_portal, c->psc_rep_portal,
438 c->psc_watchdog_factor,
440 prntfn, c->psc_min_threads, c->psc_max_threads,
441 threadname, c->psc_ctx_tags, NULL);
443 EXPORT_SYMBOL(ptlrpc_init_svc_conf);
445 static void ptlrpc_at_timer(unsigned long castmeharder)
447 struct ptlrpc_service *svc = (struct ptlrpc_service *)castmeharder;
448 svc->srv_at_check = 1;
449 svc->srv_at_checktime = cfs_time_current();
450 cfs_waitq_signal(&svc->srv_waitq);
453 /* @threadname should be 11 characters or less - 3 will be added on */
454 struct ptlrpc_service *
455 ptlrpc_init_svc(int nbufs, int bufsize, int max_req_size, int max_reply_size,
456 int req_portal, int rep_portal, int watchdog_factor,
457 svc_handler_t handler, char *name,
458 cfs_proc_dir_entry_t *proc_entry,
459 svcreq_printfn_t svcreq_printfn,
460 int min_threads, int max_threads,
461 char *threadname, __u32 ctx_tags,
462 svc_hpreq_handler_t hp_handler)
465 struct ptlrpc_at_array *array;
466 struct ptlrpc_service *service;
467 unsigned int size, index;
471 LASSERT (bufsize >= max_req_size + SPTLRPC_MAX_PAYLOAD);
472 LASSERT (ctx_tags != 0);
474 OBD_ALLOC_PTR(service);
478 /* First initialise enough for early teardown */
480 service->srv_name = name;
481 spin_lock_init(&service->srv_lock);
482 CFS_INIT_LIST_HEAD(&service->srv_threads);
483 cfs_waitq_init(&service->srv_waitq);
485 service->srv_nbuf_per_group = test_req_buffer_pressure ? 1 : nbufs;
486 service->srv_max_req_size = max_req_size + SPTLRPC_MAX_PAYLOAD;
487 service->srv_buf_size = bufsize;
488 service->srv_rep_portal = rep_portal;
489 service->srv_req_portal = req_portal;
490 service->srv_watchdog_factor = watchdog_factor;
491 service->srv_handler = handler;
492 service->srv_request_history_print_fn = svcreq_printfn;
493 service->srv_request_seq = 1; /* valid seq #s start at 1 */
494 service->srv_request_max_cull_seq = 0;
495 service->srv_threads_min = min_threads;
496 service->srv_threads_max = max_threads;
497 service->srv_thread_name = threadname;
498 service->srv_ctx_tags = ctx_tags;
499 service->srv_hpreq_handler = hp_handler;
500 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
501 service->srv_hpreq_count = 0;
502 service->srv_n_hpreq = 0;
504 rc = LNetSetLazyPortal(service->srv_req_portal);
507 CFS_INIT_LIST_HEAD(&service->srv_request_queue);
508 CFS_INIT_LIST_HEAD(&service->srv_request_hpq);
509 CFS_INIT_LIST_HEAD(&service->srv_idle_rqbds);
510 CFS_INIT_LIST_HEAD(&service->srv_active_rqbds);
511 CFS_INIT_LIST_HEAD(&service->srv_history_rqbds);
512 CFS_INIT_LIST_HEAD(&service->srv_request_history);
513 CFS_INIT_LIST_HEAD(&service->srv_active_replies);
515 CFS_INIT_LIST_HEAD(&service->srv_reply_queue);
517 CFS_INIT_LIST_HEAD(&service->srv_free_rs_list);
518 cfs_waitq_init(&service->srv_free_rs_waitq);
519 atomic_set(&service->srv_n_difficult_replies, 0);
521 spin_lock_init(&service->srv_at_lock);
522 CFS_INIT_LIST_HEAD(&service->srv_req_in_queue);
524 array = &service->srv_at_array;
525 size = at_est2timeout(at_max);
526 array->paa_size = size;
527 array->paa_count = 0;
528 array->paa_deadline = -1;
530 /* allocate memory for srv_at_array (ptlrpc_at_array) */
531 OBD_ALLOC(array->paa_reqs_array, sizeof(struct list_head) * size);
532 if (array->paa_reqs_array == NULL)
535 for (index = 0; index < size; index++)
536 CFS_INIT_LIST_HEAD(&array->paa_reqs_array[index]);
538 OBD_ALLOC(array->paa_reqs_count, sizeof(__u32) * size);
539 if (array->paa_reqs_count == NULL)
542 cfs_timer_init(&service->srv_at_timer, ptlrpc_at_timer, service);
543 /* At SOW, service time should be quick; 10s seems generous. If client
544 timeout is less than this, we'll be sending an early reply. */
545 at_init(&service->srv_at_estimate, 10, 0);
547 spin_lock (&ptlrpc_all_services_lock);
548 list_add (&service->srv_list, &ptlrpc_all_services);
549 spin_unlock (&ptlrpc_all_services_lock);
551 /* Now allocate the request buffers */
552 rc = ptlrpc_grow_req_bufs(service);
553 /* We shouldn't be under memory pressure at startup, so
554 * fail if we can't post all our buffers at this time. */
558 /* Now allocate pool of reply buffers */
559 /* Increase max reply size to next power of two */
560 service->srv_max_reply_size = 1;
561 while (service->srv_max_reply_size <
562 max_reply_size + SPTLRPC_MAX_PAYLOAD)
563 service->srv_max_reply_size <<= 1;
565 if (proc_entry != NULL)
566 ptlrpc_lprocfs_register_service(proc_entry, service);
568 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
569 service->srv_name, service->srv_req_portal);
573 ptlrpc_unregister_service(service);
578 * to actually free the request, must be called without holding svc_lock.
579 * note it's caller's responsibility to unlink req->rq_list.
581 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
583 LASSERT(atomic_read(&req->rq_refcount) == 0);
584 LASSERT(list_empty(&req->rq_timed_list));
586 /* DEBUG_REQ() assumes the reply state of a request with a valid
587 * ref will not be destroyed until that reference is dropped. */
588 ptlrpc_req_drop_rs(req);
590 sptlrpc_svc_ctx_decref(req);
592 if (req != &req->rq_rqbd->rqbd_req) {
593 /* NB request buffers use an embedded
594 * req if the incoming req unlinked the
595 * MD; this isn't one of them! */
596 OBD_FREE(req, sizeof(*req));
601 * drop a reference count of the request. if it reaches 0, we either
602 * put it into history list, or free it immediately.
604 static void ptlrpc_server_drop_request(struct ptlrpc_request *req)
606 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
607 struct ptlrpc_service *svc = rqbd->rqbd_service;
609 struct list_head *tmp;
610 struct list_head *nxt;
612 if (!atomic_dec_and_test(&req->rq_refcount))
615 spin_lock(&svc->srv_lock);
617 svc->srv_n_active_reqs--;
618 list_add(&req->rq_list, &rqbd->rqbd_reqs);
620 refcount = --(rqbd->rqbd_refcount);
622 /* request buffer is now idle: add to history */
623 list_del(&rqbd->rqbd_list);
624 list_add_tail(&rqbd->rqbd_list, &svc->srv_history_rqbds);
625 svc->srv_n_history_rqbds++;
627 /* cull some history?
628 * I expect only about 1 or 2 rqbds need to be recycled here */
629 while (svc->srv_n_history_rqbds > svc->srv_max_history_rqbds) {
630 rqbd = list_entry(svc->srv_history_rqbds.next,
631 struct ptlrpc_request_buffer_desc,
634 list_del(&rqbd->rqbd_list);
635 svc->srv_n_history_rqbds--;
637 /* remove rqbd's reqs from svc's req history while
638 * I've got the service lock */
639 list_for_each(tmp, &rqbd->rqbd_reqs) {
640 req = list_entry(tmp, struct ptlrpc_request,
642 /* Track the highest culled req seq */
643 if (req->rq_history_seq >
644 svc->srv_request_max_cull_seq)
645 svc->srv_request_max_cull_seq =
647 list_del(&req->rq_history_list);
650 spin_unlock(&svc->srv_lock);
652 list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
653 req = list_entry(rqbd->rqbd_reqs.next,
654 struct ptlrpc_request,
656 list_del(&req->rq_list);
657 ptlrpc_server_free_request(req);
660 spin_lock(&svc->srv_lock);
662 * now all reqs including the embedded req has been
663 * disposed, schedule request buffer for re-use.
665 LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) == 0);
666 list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
669 spin_unlock(&svc->srv_lock);
670 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
671 /* If we are low on memory, we are not interested in history */
672 list_del(&req->rq_list);
673 list_del_init(&req->rq_history_list);
674 spin_unlock(&svc->srv_lock);
676 ptlrpc_server_free_request(req);
678 spin_unlock(&svc->srv_lock);
683 * to finish a request: stop sending more early replies, and release
684 * the request. should be called after we finished handling the request.
686 static void ptlrpc_server_finish_request(struct ptlrpc_request *req)
688 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
690 if (req->rq_export) {
691 class_export_put(req->rq_export);
692 req->rq_export = NULL;
695 if (req->rq_phase != RQ_PHASE_NEW) /* incorrect message magic */
696 DEBUG_REQ(D_INFO, req, "free req");
698 spin_lock(&svc->srv_at_lock);
699 req->rq_sent_final = 1;
700 list_del_init(&req->rq_timed_list);
701 if (req->rq_at_linked) {
702 struct ptlrpc_at_array *array = &svc->srv_at_array;
703 __u32 index = req->rq_at_index;
705 req->rq_at_linked = 0;
706 array->paa_reqs_count[index]--;
709 spin_unlock(&svc->srv_at_lock);
711 ptlrpc_server_drop_request(req);
714 /* This function makes sure dead exports are evicted in a timely manner.
715 This function is only called when some export receives a message (i.e.,
716 the network is up.) */
717 static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
719 struct obd_export *oldest_exp;
720 time_t oldest_time, new_time;
726 /* Compensate for slow machines, etc, by faking our request time
727 into the future. Although this can break the strict time-ordering
728 of the list, we can be really lazy here - we don't have to evict
729 at the exact right moment. Eventually, all silent exports
730 will make it to the top of the list. */
732 /* Do not pay attention on 1sec or smaller renewals. */
733 new_time = cfs_time_current_sec() + extra_delay;
734 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
737 exp->exp_last_request_time = new_time;
738 CDEBUG(D_HA, "updating export %s at "CFS_TIME_T" exp %p\n",
739 exp->exp_client_uuid.uuid,
740 exp->exp_last_request_time, exp);
742 /* exports may get disconnected from the chain even though the
743 export has references, so we must keep the spin lock while
744 manipulating the lists */
745 spin_lock(&exp->exp_obd->obd_dev_lock);
747 if (list_empty(&exp->exp_obd_chain_timed)) {
748 /* this one is not timed */
749 spin_unlock(&exp->exp_obd->obd_dev_lock);
753 list_move_tail(&exp->exp_obd_chain_timed,
754 &exp->exp_obd->obd_exports_timed);
756 oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
757 struct obd_export, exp_obd_chain_timed);
758 oldest_time = oldest_exp->exp_last_request_time;
759 spin_unlock(&exp->exp_obd->obd_dev_lock);
761 if (exp->exp_obd->obd_recovering) {
762 /* be nice to everyone during recovery */
767 /* Note - racing to start/reset the obd_eviction timer is safe */
768 if (exp->exp_obd->obd_eviction_timer == 0) {
769 /* Check if the oldest entry is expired. */
770 if (cfs_time_current_sec() > (oldest_time + PING_EVICT_TIMEOUT +
772 /* We need a second timer, in case the net was down and
773 * it just came back. Since the pinger may skip every
774 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
775 * we better wait for 3. */
776 exp->exp_obd->obd_eviction_timer =
777 cfs_time_current_sec() + 3 * PING_INTERVAL;
778 CDEBUG(D_HA, "%s: Think about evicting %s from "CFS_TIME_T"\n",
779 exp->exp_obd->obd_name, obd_export_nid2str(exp),
783 if (cfs_time_current_sec() >
784 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
785 /* The evictor won't evict anyone who we've heard from
786 * recently, so we don't have to check before we start
788 if (!ping_evictor_wake(exp))
789 exp->exp_obd->obd_eviction_timer = 0;
796 static int ptlrpc_check_req(struct ptlrpc_request *req)
798 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
799 req->rq_export->exp_conn_cnt)) {
800 DEBUG_REQ(D_ERROR, req,
801 "DROPPING req from old connection %d < %d",
802 lustre_msg_get_conn_cnt(req->rq_reqmsg),
803 req->rq_export->exp_conn_cnt);
806 if (unlikely(req->rq_export->exp_obd &&
807 req->rq_export->exp_obd->obd_fail)) {
808 /* Failing over, don't handle any more reqs, send
809 error response instead. */
810 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
811 req, req->rq_export->exp_obd->obd_name);
812 req->rq_status = -ENODEV;
820 static void ptlrpc_at_set_timer(struct ptlrpc_service *svc)
822 struct ptlrpc_at_array *array = &svc->srv_at_array;
825 spin_lock(&svc->srv_at_lock);
826 if (array->paa_count == 0) {
827 cfs_timer_disarm(&svc->srv_at_timer);
828 spin_unlock(&svc->srv_at_lock);
832 /* Set timer for closest deadline */
833 next = (__s32)(array->paa_deadline - cfs_time_current_sec() -
836 ptlrpc_at_timer((unsigned long)svc);
838 cfs_timer_arm(&svc->srv_at_timer, cfs_time_shift(next));
839 spin_unlock(&svc->srv_at_lock);
840 CDEBUG(D_INFO, "armed %s at %+ds\n", svc->srv_name, next);
843 /* Add rpc to early reply check list */
844 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
846 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
847 struct ptlrpc_request *rq = NULL;
848 struct ptlrpc_at_array *array = &svc->srv_at_array;
855 if (req->rq_no_reply)
858 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
861 spin_lock(&svc->srv_at_lock);
863 if (unlikely(req->rq_sent_final)) {
864 spin_unlock(&svc->srv_at_lock);
868 LASSERT(list_empty(&req->rq_timed_list));
870 index = (unsigned long)req->rq_deadline % array->paa_size;
871 if (array->paa_reqs_count[index] > 0) {
872 /* latest rpcs will have the latest deadlines in the list,
873 * so search backward. */
874 list_for_each_entry_reverse(rq, &array->paa_reqs_array[index],
876 if (req->rq_deadline >= rq->rq_deadline) {
877 list_add(&req->rq_timed_list,
884 /* Add the request at the head of the list */
885 if (list_empty(&req->rq_timed_list))
886 list_add(&req->rq_timed_list, &array->paa_reqs_array[index]);
888 req->rq_at_linked = 1;
889 req->rq_at_index = index;
890 array->paa_reqs_count[index]++;
892 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
893 array->paa_deadline = req->rq_deadline;
896 spin_unlock(&svc->srv_at_lock);
899 ptlrpc_at_set_timer(svc);
904 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req,
907 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
908 struct ptlrpc_request *reqcopy;
909 struct lustre_msg *reqmsg;
910 cfs_duration_t olddl = req->rq_deadline - cfs_time_current_sec();
915 /* deadline is when the client expects us to reply, margin is the
916 difference between clients' and servers' expectations */
917 DEBUG_REQ(D_ADAPTTO, req,
918 "%ssending early reply (deadline %+lds, margin %+lds) for "
919 "%d+%d", AT_OFF ? "AT off - not " : "",
920 olddl, olddl - at_get(&svc->srv_at_estimate),
921 at_get(&svc->srv_at_estimate), extra_time);
927 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), "
928 "not sending early reply. Consider increasing "
929 "at_early_margin (%d)?", olddl, at_early_margin);
931 /* Return an error so we're not re-added to the timed list. */
935 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
936 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
937 "but no AT support");
941 if (req->rq_export && req->rq_export->exp_in_recovery) {
942 /* don't increase server estimates during recovery, and give
943 clients the full recovery time. */
944 newdl = cfs_time_current_sec() +
945 req->rq_export->exp_obd->obd_recovery_timeout;
948 /* Fake our processing time into the future to ask the
949 clients for some extra amount of time */
950 extra_time += cfs_time_current_sec() -
951 req->rq_arrival_time.tv_sec;
952 at_add(&svc->srv_at_estimate, extra_time);
954 newdl = req->rq_arrival_time.tv_sec +
955 at_get(&svc->srv_at_estimate);
957 if (req->rq_deadline >= newdl) {
958 /* We're not adding any time, no need to send an early reply
959 (e.g. maybe at adaptive_max) */
960 DEBUG_REQ(D_WARNING, req, "Couldn't add any time ("
961 CFS_DURATION_T"/"CFS_DURATION_T"), "
962 "not sending early reply\n", olddl,
963 cfs_time_sub(newdl, cfs_time_current_sec()));
967 OBD_ALLOC(reqcopy, sizeof *reqcopy);
970 OBD_ALLOC(reqmsg, req->rq_reqlen);
972 OBD_FREE(reqcopy, sizeof *reqcopy);
977 reqcopy->rq_reply_state = NULL;
978 reqcopy->rq_rep_swab_mask = 0;
979 reqcopy->rq_pack_bulk = 0;
980 reqcopy->rq_pack_udesc = 0;
981 reqcopy->rq_packed_final = 0;
982 sptlrpc_svc_ctx_addref(reqcopy);
983 /* We only need the reqmsg for the magic */
984 reqcopy->rq_reqmsg = reqmsg;
985 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
987 if (req->rq_sent_final) {
988 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
989 "abort sending early reply\n");
994 reqcopy->rq_export = class_conn2export(
995 lustre_msg_get_handle(reqcopy->rq_reqmsg));
996 if (reqcopy->rq_export == NULL)
997 GOTO(out, rc = -ENODEV);
1000 class_export_rpc_get(reqcopy->rq_export);
1001 if (reqcopy->rq_export->exp_obd &&
1002 reqcopy->rq_export->exp_obd->obd_fail)
1003 GOTO(out_put, rc = -ENODEV);
1005 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1009 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1012 /* Adjust our own deadline to what we told the client */
1013 req->rq_deadline = newdl;
1014 req->rq_early_count++; /* number sent, server side */
1016 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1019 /* Free the (early) reply state from lustre_pack_reply.
1020 (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
1021 ptlrpc_req_drop_rs(reqcopy);
1024 class_export_rpc_put(reqcopy->rq_export);
1025 class_export_put(reqcopy->rq_export);
1027 sptlrpc_svc_ctx_decref(reqcopy);
1028 OBD_FREE(reqmsg, req->rq_reqlen);
1029 OBD_FREE(reqcopy, sizeof *reqcopy);
1033 /* Send early replies to everybody expiring within at_early_margin
1034 asking for at_extra time */
1035 static int ptlrpc_at_check_timed(struct ptlrpc_service *svc)
1037 struct ptlrpc_request *rq, *n;
1038 struct list_head work_list;
1039 struct ptlrpc_at_array *array = &svc->srv_at_array;
1042 time_t now = cfs_time_current_sec();
1043 cfs_duration_t delay;
1044 int first, counter = 0;
1047 spin_lock(&svc->srv_at_lock);
1048 if (svc->srv_at_check == 0) {
1049 spin_unlock(&svc->srv_at_lock);
1052 delay = cfs_time_sub(cfs_time_current(), svc->srv_at_checktime);
1053 svc->srv_at_check = 0;
1055 if (array->paa_count == 0) {
1056 spin_unlock(&svc->srv_at_lock);
1060 /* The timer went off, but maybe the nearest rpc already completed. */
1061 first = array->paa_deadline - now;
1062 if (first > at_early_margin) {
1063 /* We've still got plenty of time. Reset the timer. */
1064 spin_unlock(&svc->srv_at_lock);
1065 ptlrpc_at_set_timer(svc);
1069 /* We're close to a timeout, and we don't know how much longer the
1070 server will take. Send early replies to everyone expiring soon. */
1071 CFS_INIT_LIST_HEAD(&work_list);
1073 index = (unsigned long)array->paa_deadline % array->paa_size;
1074 count = array->paa_count;
1076 count -= array->paa_reqs_count[index];
1077 list_for_each_entry_safe(rq, n, &array->paa_reqs_array[index],
1079 if (rq->rq_deadline <= now + at_early_margin) {
1080 list_move(&rq->rq_timed_list, &work_list);
1082 array->paa_reqs_count[index]--;
1084 rq->rq_at_linked = 0;
1088 /* update the earliest deadline */
1089 if (deadline == -1 || rq->rq_deadline < deadline)
1090 deadline = rq->rq_deadline;
1095 if (++index >= array->paa_size)
1098 array->paa_deadline = deadline;
1099 spin_unlock(&svc->srv_at_lock);
1101 /* we have a new earliest deadline, restart the timer */
1102 ptlrpc_at_set_timer(svc);
1104 CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
1105 "replies\n", first, at_extra, counter);
1107 /* We're already past request deadlines before we even get a
1108 chance to send early replies */
1109 LCONSOLE_WARN("%s: This server is not able to keep up with "
1110 "request traffic (cpu-bound).\n", svc->srv_name);
1111 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, "
1112 "delay="CFS_DURATION_T"(jiff)\n",
1113 counter, svc->srv_n_queued_reqs, svc->srv_n_active_reqs,
1114 at_get(&svc->srv_at_estimate), delay);
1117 /* ptlrpc_server_finish_request may delete an entry out of
1119 spin_lock(&svc->srv_at_lock);
1120 while (!list_empty(&work_list)) {
1121 rq = list_entry(work_list.next, struct ptlrpc_request,
1123 list_del_init(&rq->rq_timed_list);
1124 /* if the entry is still in the worklist, it hasn't been
1125 deleted, and is safe to take a ref to keep the req around */
1126 atomic_inc(&rq->rq_refcount);
1127 spin_unlock(&svc->srv_at_lock);
1129 if (ptlrpc_at_send_early_reply(rq, at_extra) == 0)
1130 ptlrpc_at_add_timed(rq);
1132 ptlrpc_server_drop_request(rq);
1133 spin_lock(&svc->srv_at_lock);
1135 spin_unlock(&svc->srv_at_lock);
1141 * Put the request to the export list if the request may become
1142 * a high priority one.
1144 static int ptlrpc_hpreq_init(struct ptlrpc_service *svc,
1145 struct ptlrpc_request *req)
1150 if (svc->srv_hpreq_handler) {
1151 rc = svc->srv_hpreq_handler(req);
1155 if (req->rq_export && req->rq_ops) {
1156 spin_lock(&req->rq_export->exp_lock);
1157 list_add(&req->rq_exp_list, &req->rq_export->exp_queued_rpc);
1158 spin_unlock(&req->rq_export->exp_lock);
1164 /** Remove the request from the export list. */
1165 static void ptlrpc_hpreq_fini(struct ptlrpc_request *req)
1168 if (req->rq_export && req->rq_ops) {
1169 spin_lock(&req->rq_export->exp_lock);
1170 list_del_init(&req->rq_exp_list);
1171 spin_unlock(&req->rq_export->exp_lock);
1177 * Make the request a high priority one.
1179 * All the high priority requests are queued in a separate FIFO
1180 * ptlrpc_service::srv_request_hpq list which is parallel to
1181 * ptlrpc_service::srv_request_queue list but has a higher priority
1184 * \see ptlrpc_server_handle_request().
1186 static void ptlrpc_hpreq_reorder_nolock(struct ptlrpc_service *svc,
1187 struct ptlrpc_request *req)
1190 LASSERT(svc != NULL);
1191 spin_lock(&req->rq_lock);
1192 if (req->rq_hp == 0) {
1193 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1195 /* Add to the high priority queue. */
1196 list_move_tail(&req->rq_list, &svc->srv_request_hpq);
1198 if (opc != OBD_PING)
1199 DEBUG_REQ(D_NET, req, "high priority req");
1201 spin_unlock(&req->rq_lock);
1205 void ptlrpc_hpreq_reorder(struct ptlrpc_request *req)
1207 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
1210 spin_lock(&svc->srv_lock);
1211 /* It may happen that the request is already taken for the processing
1212 * but still in the export list, do not re-add it into the HP list. */
1213 if (req->rq_phase == RQ_PHASE_NEW)
1214 ptlrpc_hpreq_reorder_nolock(svc, req);
1215 spin_unlock(&svc->srv_lock);
1219 /** Check if the request if a high priority one. */
1220 static int ptlrpc_server_hpreq_check(struct ptlrpc_request *req)
1225 /* Check by request opc. */
1226 opc = lustre_msg_get_opc(req->rq_reqmsg);
1227 if (opc == OBD_PING)
1230 /* Perform request specific check. */
1231 if (req->rq_ops && req->rq_ops->hpreq_check)
1232 rc = req->rq_ops->hpreq_check(req);
1236 /** Check if a request is a high priority one. */
1237 static int ptlrpc_server_request_add(struct ptlrpc_service *svc,
1238 struct ptlrpc_request *req)
1243 rc = ptlrpc_server_hpreq_check(req);
1247 spin_lock(&svc->srv_lock);
1248 /* Before inserting the request into the queue, check if it is not
1249 * inserted yet, or even already handled -- it may happen due to
1250 * a racing ldlm_server_blocking_ast(). */
1251 if (req->rq_phase == RQ_PHASE_NEW && list_empty(&req->rq_list)) {
1253 ptlrpc_hpreq_reorder_nolock(svc, req);
1255 list_add_tail(&req->rq_list, &svc->srv_request_queue);
1257 spin_unlock(&svc->srv_lock);
1262 /* Only allow normal priority requests on a service that has a high-priority
1263 * queue if forced (i.e. cleanup), if there are other high priority requests
1264 * already being processed (i.e. those threads can service more high-priority
1265 * requests), or if there are enough idle threads that a later thread can do
1266 * a high priority request. */
1267 static int ptlrpc_server_allow_normal(struct ptlrpc_service *svc, int force)
1269 return force || !svc->srv_hpreq_handler || svc->srv_n_hpreq > 0 ||
1270 svc->srv_n_active_reqs < svc->srv_threads_running - 2;
1273 static struct ptlrpc_request *
1274 ptlrpc_server_request_get(struct ptlrpc_service *svc, int force)
1276 struct ptlrpc_request *req = NULL;
1279 if (ptlrpc_server_allow_normal(svc, force) &&
1280 !list_empty(&svc->srv_request_queue) &&
1281 (list_empty(&svc->srv_request_hpq) ||
1282 svc->srv_hpreq_count >= svc->srv_hpreq_ratio)) {
1283 req = list_entry(svc->srv_request_queue.next,
1284 struct ptlrpc_request, rq_list);
1285 svc->srv_hpreq_count = 0;
1286 } else if (!list_empty(&svc->srv_request_hpq)) {
1287 req = list_entry(svc->srv_request_hpq.next,
1288 struct ptlrpc_request, rq_list);
1289 svc->srv_hpreq_count++;
1294 static int ptlrpc_server_request_pending(struct ptlrpc_service *svc, int force)
1296 return ((ptlrpc_server_allow_normal(svc, force) &&
1297 !list_empty(&svc->srv_request_queue)) ||
1298 !list_empty(&svc->srv_request_hpq));
1301 /* Handle freshly incoming reqs, add to timed early reply list,
1302 pass on to regular request queue */
1304 ptlrpc_server_handle_req_in(struct ptlrpc_service *svc)
1306 struct ptlrpc_request *req;
1313 spin_lock(&svc->srv_lock);
1314 if (list_empty(&svc->srv_req_in_queue)) {
1315 spin_unlock(&svc->srv_lock);
1319 req = list_entry(svc->srv_req_in_queue.next,
1320 struct ptlrpc_request, rq_list);
1321 list_del_init (&req->rq_list);
1322 /* Consider this still a "queued" request as far as stats are
1324 spin_unlock(&svc->srv_lock);
1326 /* go through security check/transform */
1327 rc = sptlrpc_svc_unwrap_request(req);
1331 case SECSVC_COMPLETE:
1332 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1340 /* Clear request swab mask; this is a new request */
1341 req->rq_req_swab_mask = 0;
1343 rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
1345 CERROR("error unpacking request: ptl %d from %s x"LPU64"\n",
1346 svc->srv_req_portal, libcfs_id2str(req->rq_peer),
1351 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1353 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
1354 LPU64"\n", svc->srv_req_portal,
1355 libcfs_id2str(req->rq_peer), req->rq_xid);
1359 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
1360 lustre_msg_get_opc(req->rq_reqmsg) == obd_fail_val) {
1361 CERROR("drop incoming rpc opc %u, x"LPU64"\n",
1362 obd_fail_val, req->rq_xid);
1367 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1368 CERROR("wrong packet type received (type=%u) from %s\n",
1369 lustre_msg_get_type(req->rq_reqmsg),
1370 libcfs_id2str(req->rq_peer));
1374 switch(lustre_msg_get_opc(req->rq_reqmsg)) {
1377 req->rq_bulk_write = 1;
1381 req->rq_bulk_read = 1;
1385 CDEBUG(D_NET, "got req "LPU64"\n", req->rq_xid);
1387 req->rq_export = class_conn2export(
1388 lustre_msg_get_handle(req->rq_reqmsg));
1389 if (req->rq_export) {
1390 rc = ptlrpc_check_req(req);
1392 rc = sptlrpc_target_export_check(req->rq_export, req);
1394 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
1395 "illegal security flavor,");
1400 ptlrpc_update_export_timer(req->rq_export, 0);
1403 /* req_in handling should/must be fast */
1404 if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5)
1405 DEBUG_REQ(D_WARNING, req, "Slow req_in handling "CFS_DURATION_T"s",
1406 cfs_time_sub(cfs_time_current_sec(),
1407 req->rq_arrival_time.tv_sec));
1409 /* Set rpc server deadline and add it to the timed list */
1410 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1411 MSGHDR_AT_SUPPORT) ?
1412 /* The max time the client expects us to take */
1413 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1414 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1415 if (unlikely(deadline == 0)) {
1416 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1420 ptlrpc_at_add_timed(req);
1421 rc = ptlrpc_hpreq_init(svc, req);
1425 /* Move it over to the request processing queue */
1426 rc = ptlrpc_server_request_add(svc, req);
1429 cfs_waitq_signal(&svc->srv_waitq);
1433 spin_lock(&svc->srv_lock);
1434 svc->srv_n_queued_reqs--;
1435 svc->srv_n_active_reqs++;
1436 spin_unlock(&svc->srv_lock);
1437 ptlrpc_server_finish_request(req);
1443 ptlrpc_server_handle_request(struct ptlrpc_service *svc,
1444 struct ptlrpc_thread *thread)
1446 struct obd_export *export = NULL;
1447 struct ptlrpc_request *request;
1448 struct timeval work_start;
1449 struct timeval work_end;
1457 spin_lock(&svc->srv_lock);
1458 if (unlikely(!ptlrpc_server_request_pending(svc, 0) ||
1461 /* !@%$# liblustre only has 1 thread */
1462 atomic_read(&svc->srv_n_difficult_replies) != 0 &&
1464 svc->srv_n_active_reqs >= (svc->srv_threads_running - 1)))) {
1465 /* Don't handle regular requests in the last thread, in order * re
1466 * to handle difficult replies (which might block other threads)
1467 * as well as handle any incoming reqs, early replies, etc.
1468 * That means we always need at least 2 service threads. */
1469 spin_unlock(&svc->srv_lock);
1473 request = ptlrpc_server_request_get(svc, 0);
1474 if (request == NULL) {
1475 spin_unlock(&svc->srv_lock);
1479 opc = lustre_msg_get_opc(request->rq_reqmsg);
1480 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
1481 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
1482 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
1483 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
1485 if (unlikely(fail_opc)) {
1486 if (request->rq_export && request->rq_ops) {
1487 spin_unlock(&svc->srv_lock);
1488 OBD_FAIL_TIMEOUT(fail_opc, 4);
1489 spin_lock(&svc->srv_lock);
1490 request = ptlrpc_server_request_get(svc, 0);
1491 if (request == NULL) {
1492 spin_unlock(&svc->srv_lock);
1495 LASSERT(ptlrpc_server_request_pending(svc, 0));
1499 list_del_init(&request->rq_list);
1500 svc->srv_n_queued_reqs--;
1501 svc->srv_n_active_reqs++;
1505 /* The phase is changed under the lock here because we need to know
1506 * the request is under processing (see ptlrpc_hpreq_reorder()). */
1507 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
1508 spin_unlock(&svc->srv_lock);
1510 ptlrpc_hpreq_fini(request);
1512 if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
1513 libcfs_debug_dumplog();
1515 do_gettimeofday(&work_start);
1516 timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
1517 if (likely(svc->srv_stats != NULL)) {
1518 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
1520 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
1521 svc->srv_n_queued_reqs);
1522 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
1523 svc->srv_n_active_reqs);
1524 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
1525 at_get(&svc->srv_at_estimate));
1528 rc = lu_context_init(&request->rq_session,
1529 LCT_SESSION|LCT_REMEMBER|LCT_NOREF);
1531 CERROR("Failure to initialize session: %d\n", rc);
1534 request->rq_session.lc_thread = thread;
1535 request->rq_session.lc_cookie = 0x5;
1536 lu_context_enter(&request->rq_session);
1538 CDEBUG(D_NET, "got req "LPU64"\n", request->rq_xid);
1540 request->rq_svc_thread = thread;
1542 request->rq_svc_thread->t_env->le_ses = &request->rq_session;
1544 if (likely(request->rq_export)) {
1545 if (unlikely(ptlrpc_check_req(request)))
1547 ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
1548 export = class_export_rpc_get(request->rq_export);
1551 /* Discard requests queued for longer than the deadline.
1552 The deadline is increased if we send an early reply. */
1553 if (cfs_time_current_sec() > request->rq_deadline) {
1554 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s"
1555 ": deadline "CFS_DURATION_T":"CFS_DURATION_T"s ago\n",
1556 libcfs_id2str(request->rq_peer),
1557 cfs_time_sub(request->rq_deadline,
1558 request->rq_arrival_time.tv_sec),
1559 cfs_time_sub(cfs_time_current_sec(),
1560 request->rq_deadline));
1561 goto put_rpc_export;
1564 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
1565 "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1566 (request->rq_export ?
1567 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1568 (request->rq_export ?
1569 atomic_read(&request->rq_export->exp_refcount) : -99),
1570 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1571 libcfs_id2str(request->rq_peer),
1572 lustre_msg_get_opc(request->rq_reqmsg));
1574 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
1575 OBD_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, obd_fail_val);
1577 rc = svc->srv_handler(request);
1579 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
1581 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
1582 "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1583 (request->rq_export ?
1584 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1585 (request->rq_export ?
1586 atomic_read(&request->rq_export->exp_refcount) : -99),
1587 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1588 libcfs_id2str(request->rq_peer),
1589 lustre_msg_get_opc(request->rq_reqmsg));
1593 class_export_rpc_put(export);
1595 lu_context_exit(&request->rq_session);
1596 lu_context_fini(&request->rq_session);
1598 if (unlikely(cfs_time_current_sec() > request->rq_deadline)) {
1599 DEBUG_REQ(D_WARNING, request, "Request x"LPU64" took longer "
1600 "than estimated ("CFS_DURATION_T":"CFS_DURATION_T"s);"
1601 " client may timeout.",
1602 request->rq_xid, cfs_time_sub(request->rq_deadline,
1603 request->rq_arrival_time.tv_sec),
1604 cfs_time_sub(cfs_time_current_sec(),
1605 request->rq_deadline));
1608 do_gettimeofday(&work_end);
1609 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1610 CDEBUG(D_RPCTRACE, "request x"LPU64" opc %u from %s processed in "
1611 "%ldus (%ldus total) trans "LPU64" rc %d/%d\n",
1612 request->rq_xid, lustre_msg_get_opc(request->rq_reqmsg),
1613 libcfs_id2str(request->rq_peer), timediff,
1614 cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL),
1615 request->rq_repmsg ? lustre_msg_get_transno(request->rq_repmsg) :
1616 request->rq_transno, request->rq_status,
1617 request->rq_repmsg ? lustre_msg_get_status(request->rq_repmsg):
1619 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
1620 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
1621 int opc = opcode_offset(op);
1622 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
1623 LASSERT(opc < LUSTRE_MAX_OPCODES);
1624 lprocfs_counter_add(svc->srv_stats,
1625 opc + EXTRA_MAX_OPCODES,
1629 if (unlikely(request->rq_early_count)) {
1630 DEBUG_REQ(D_ADAPTTO, request,
1631 "sent %d early replies before finishing in "
1633 request->rq_early_count,
1634 cfs_time_sub(work_end.tv_sec,
1635 request->rq_arrival_time.tv_sec));
1639 spin_lock(&svc->srv_lock);
1642 spin_unlock(&svc->srv_lock);
1643 ptlrpc_server_finish_request(request);
1649 * An internal function to process a single reply state object.
1652 ptlrpc_handle_rs (struct ptlrpc_reply_state *rs)
1654 struct ptlrpc_service *svc = rs->rs_service;
1655 struct obd_export *exp;
1656 struct obd_device *obd;
1661 exp = rs->rs_export;
1664 LASSERT (rs->rs_difficult);
1665 LASSERT (rs->rs_scheduled);
1666 LASSERT (list_empty(&rs->rs_list));
1668 spin_lock (&exp->exp_lock);
1669 /* Noop if removed already */
1670 list_del_init (&rs->rs_exp_list);
1671 spin_unlock (&exp->exp_lock);
1673 /* Avoid exp_uncommitted_replies_lock contention if we 100% sure that
1674 * rs has been removed from the list already */
1675 if (!list_empty_careful(&rs->rs_obd_list)) {
1676 spin_lock(&exp->exp_uncommitted_replies_lock);
1677 list_del_init(&rs->rs_obd_list);
1678 spin_unlock(&exp->exp_uncommitted_replies_lock);
1681 spin_lock(&rs->rs_lock);
1683 been_handled = rs->rs_handled;
1686 nlocks = rs->rs_nlocks; /* atomic "steal", but */
1687 rs->rs_nlocks = 0; /* locks still on rs_locks! */
1689 if (nlocks == 0 && !been_handled) {
1690 /* If we see this, we should already have seen the warning
1691 * in mds_steal_ack_locks() */
1692 CWARN("All locks stolen from rs %p x"LPD64".t"LPD64
1695 rs->rs_xid, rs->rs_transno, rs->rs_opc,
1696 libcfs_nid2str(exp->exp_connection->c_peer.nid));
1699 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
1700 spin_unlock(&rs->rs_lock);
1702 if (!been_handled && rs->rs_on_net) {
1703 LNetMDUnlink(rs->rs_md_h);
1704 /* Ignore return code; we're racing with
1708 while (nlocks-- > 0)
1709 ldlm_lock_decref(&rs->rs_locks[nlocks],
1710 rs->rs_modes[nlocks]);
1712 spin_lock(&rs->rs_lock);
1715 rs->rs_scheduled = 0;
1717 if (!rs->rs_on_net) {
1719 spin_unlock(&rs->rs_lock);
1721 class_export_put (exp);
1722 rs->rs_export = NULL;
1723 ptlrpc_rs_decref (rs);
1724 atomic_dec (&svc->srv_outstanding_replies);
1725 if (atomic_dec_and_test(&svc->srv_n_difficult_replies) &&
1726 svc->srv_is_stopping)
1727 cfs_waitq_broadcast(&svc->srv_waitq);
1731 /* still on the net; callback will schedule */
1732 spin_unlock(&rs->rs_lock);
1739 * Check whether given service has a reply available for processing
1742 * \param svc a ptlrpc service
1743 * \retval 0 no replies processed
1744 * \retval 1 one reply processed
1747 ptlrpc_server_handle_reply(struct ptlrpc_service *svc)
1749 struct ptlrpc_reply_state *rs = NULL;
1752 spin_lock(&svc->srv_lock);
1753 if (!list_empty(&svc->srv_reply_queue)) {
1754 rs = list_entry(svc->srv_reply_queue.prev,
1755 struct ptlrpc_reply_state,
1757 list_del_init(&rs->rs_list);
1759 spin_unlock(&svc->srv_lock);
1761 ptlrpc_handle_rs(rs);
1765 /* FIXME make use of timeout later */
1767 liblustre_check_services (void *arg)
1769 int did_something = 0;
1771 struct list_head *tmp, *nxt;
1774 /* I'm relying on being single threaded, not to have to lock
1775 * ptlrpc_all_services etc */
1776 list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
1777 struct ptlrpc_service *svc =
1778 list_entry (tmp, struct ptlrpc_service, srv_list);
1780 if (svc->srv_threads_running != 0) /* I've recursed */
1783 /* service threads can block for bulk, so this limits us
1784 * (arbitrarily) to recursing 1 stack frame per service.
1785 * Note that the problem with recursion is that we have to
1786 * unwind completely before our caller can resume. */
1788 svc->srv_threads_running++;
1791 rc = ptlrpc_server_handle_req_in(svc);
1792 rc |= ptlrpc_server_handle_reply(svc);
1793 rc |= ptlrpc_at_check_timed(svc);
1794 rc |= ptlrpc_server_handle_request(svc, NULL);
1795 rc |= (ptlrpc_server_post_idle_rqbds(svc) > 0);
1796 did_something |= rc;
1799 svc->srv_threads_running--;
1802 RETURN(did_something);
1804 #define ptlrpc_stop_all_threads(s) do {} while (0)
1806 #else /* __KERNEL__ */
1809 ptlrpc_check_rqbd_pool(struct ptlrpc_service *svc)
1811 int avail = svc->srv_nrqbd_receiving;
1812 int low_water = test_req_buffer_pressure ? 0 :
1813 svc->srv_nbuf_per_group/2;
1815 /* NB I'm not locking; just looking. */
1817 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
1818 * allowed the request history to grow out of control. We could put a
1819 * sanity check on that here and cull some history if we need the
1822 if (avail <= low_water)
1823 ptlrpc_grow_req_bufs(svc);
1826 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQBUF_AVAIL_CNTR,
1831 ptlrpc_retry_rqbds(void *arg)
1833 struct ptlrpc_service *svc = (struct ptlrpc_service *)arg;
1835 svc->srv_rqbd_timeout = 0;
1836 return (-ETIMEDOUT);
1839 static int ptlrpc_main(void *arg)
1841 struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
1842 struct ptlrpc_service *svc = data->svc;
1843 struct ptlrpc_thread *thread = data->thread;
1844 struct obd_device *dev = data->dev;
1845 struct ptlrpc_reply_state *rs;
1846 #ifdef WITH_GROUP_INFO
1847 struct group_info *ginfo = NULL;
1850 int counter = 0, rc = 0;
1853 cfs_daemonize_ctxt(data->name);
1855 #if defined(HAVE_NODE_TO_CPUMASK) && defined(CONFIG_NUMA)
1856 /* we need to do this before any per-thread allocation is done so that
1857 * we get the per-thread allocations on local node. bug 7342 */
1858 if (svc->srv_cpu_affinity) {
1861 for (cpu = 0, num_cpu = 0; cpu < num_possible_cpus(); cpu++) {
1862 if (!cpu_online(cpu))
1864 if (num_cpu == thread->t_id % num_online_cpus())
1868 set_cpus_allowed(cfs_current(), node_to_cpumask(cpu_to_node(cpu)));
1872 #ifdef WITH_GROUP_INFO
1873 ginfo = groups_alloc(0);
1879 set_current_groups(ginfo);
1880 put_group_info(ginfo);
1883 if (svc->srv_init != NULL) {
1884 rc = svc->srv_init(thread);
1889 rc = lu_context_init(&env.le_ctx,
1890 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
1894 thread->t_env = &env;
1895 env.le_ctx.lc_thread = thread;
1896 env.le_ctx.lc_cookie = 0x6;
1898 /* Alloc reply state structure for this one */
1899 OBD_ALLOC_GFP(rs, svc->srv_max_reply_size, CFS_ALLOC_STD);
1905 /* Record that the thread is running */
1906 thread->t_flags = SVC_RUNNING;
1908 * wake up our creator. Note: @data is invalid after this point,
1909 * because it's allocated on ptlrpc_start_thread() stack.
1911 cfs_waitq_signal(&thread->t_ctl_waitq);
1913 thread->t_watchdog = lc_watchdog_add(GET_TIMEOUT(svc), NULL, NULL);
1915 spin_lock(&svc->srv_lock);
1916 svc->srv_threads_running++;
1917 list_add(&rs->rs_list, &svc->srv_free_rs_list);
1918 spin_unlock(&svc->srv_lock);
1919 cfs_waitq_signal(&svc->srv_free_rs_waitq);
1921 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
1922 svc->srv_threads_running);
1924 /* XXX maintain a list of all managed devices: insert here */
1926 while ((thread->t_flags & SVC_STOPPING) == 0) {
1927 /* Don't exit while there are replies to be handled */
1928 struct l_wait_info lwi = LWI_TIMEOUT(svc->srv_rqbd_timeout,
1929 ptlrpc_retry_rqbds, svc);
1931 lc_watchdog_disable(thread->t_watchdog);
1935 l_wait_event_exclusive (svc->srv_waitq,
1936 ((thread->t_flags & SVC_STOPPING) != 0) ||
1937 (!list_empty(&svc->srv_idle_rqbds) &&
1938 svc->srv_rqbd_timeout == 0) ||
1939 !list_empty(&svc->srv_req_in_queue) ||
1940 (ptlrpc_server_request_pending(svc, 0) &&
1941 (svc->srv_n_active_reqs <
1942 (svc->srv_threads_running - 1))) ||
1946 lc_watchdog_touch(thread->t_watchdog, GET_TIMEOUT(svc));
1948 ptlrpc_check_rqbd_pool(svc);
1950 if ((svc->srv_threads_started < svc->srv_threads_max) &&
1951 (svc->srv_n_active_reqs >= (svc->srv_threads_started - 1))){
1952 /* Ignore return code - we tried... */
1953 ptlrpc_start_thread(dev, svc);
1956 if (!list_empty(&svc->srv_req_in_queue)) {
1957 /* Process all incoming reqs before handling any */
1958 ptlrpc_server_handle_req_in(svc);
1959 /* but limit ourselves in case of flood */
1960 if (counter++ < 1000)
1965 if (svc->srv_at_check)
1966 ptlrpc_at_check_timed(svc);
1968 /* don't handle requests in the last thread */
1969 if (ptlrpc_server_request_pending(svc, 0) &&
1970 (svc->srv_n_active_reqs < (svc->srv_threads_running - 1))) {
1971 lu_context_enter(&env.le_ctx);
1972 ptlrpc_server_handle_request(svc, thread);
1973 lu_context_exit(&env.le_ctx);
1976 if (!list_empty(&svc->srv_idle_rqbds) &&
1977 ptlrpc_server_post_idle_rqbds(svc) < 0) {
1978 /* I just failed to repost request buffers. Wait
1979 * for a timeout (unless something else happens)
1980 * before I try again */
1981 svc->srv_rqbd_timeout = cfs_time_seconds(1)/10;
1982 CDEBUG(D_RPCTRACE,"Posted buffers: %d\n",
1983 svc->srv_nrqbd_receiving);
1987 lc_watchdog_delete(thread->t_watchdog);
1988 thread->t_watchdog = NULL;
1992 * deconstruct service specific state created by ptlrpc_start_thread()
1994 if (svc->srv_done != NULL)
1995 svc->srv_done(thread);
1997 lu_context_fini(&env.le_ctx);
1999 CDEBUG(D_NET, "service thread %d exiting: rc %d\n", thread->t_id, rc);
2001 spin_lock(&svc->srv_lock);
2002 svc->srv_threads_running--; /* must know immediately */
2004 thread->t_flags = SVC_STOPPED;
2006 cfs_waitq_signal(&thread->t_ctl_waitq);
2007 spin_unlock(&svc->srv_lock);
2012 struct ptlrpc_hr_args {
2015 struct ptlrpc_hr_service *hrs;
2018 static int hrt_dont_sleep(struct ptlrpc_hr_thread *t,
2019 struct list_head *replies)
2023 spin_lock(&t->hrt_lock);
2024 list_splice_init(&t->hrt_queue, replies);
2025 result = test_bit(HRT_STOPPING, &t->hrt_flags) ||
2026 !list_empty(replies);
2027 spin_unlock(&t->hrt_lock);
2031 static int ptlrpc_hr_main(void *arg)
2033 struct ptlrpc_hr_args * hr_args = arg;
2034 struct ptlrpc_hr_service *hr = hr_args->hrs;
2035 struct ptlrpc_hr_thread *t = &hr->hr_threads[hr_args->thread_index];
2036 char threadname[20];
2037 CFS_LIST_HEAD(replies);
2039 snprintf(threadname, sizeof(threadname),
2040 "ptlrpc_hr_%d", hr_args->thread_index);
2042 cfs_daemonize_ctxt(threadname);
2043 #if defined(HAVE_NODE_TO_CPUMASK)
2044 set_cpus_allowed(cfs_current(),
2045 node_to_cpumask(cpu_to_node(hr_args->cpu_index)));
2047 set_bit(HRT_RUNNING, &t->hrt_flags);
2048 cfs_waitq_signal(&t->hrt_wait);
2050 while (!test_bit(HRT_STOPPING, &t->hrt_flags)) {
2052 cfs_wait_event(t->hrt_wait, hrt_dont_sleep(t, &replies));
2053 while (!list_empty(&replies)) {
2054 struct ptlrpc_reply_state *rs;
2056 rs = list_entry(replies.prev,
2057 struct ptlrpc_reply_state,
2059 list_del_init(&rs->rs_list);
2060 ptlrpc_handle_rs(rs);
2064 clear_bit(HRT_RUNNING, &t->hrt_flags);
2065 complete(&t->hrt_completion);
2070 static int ptlrpc_start_hr_thread(struct ptlrpc_hr_service *hr, int n, int cpu)
2072 struct ptlrpc_hr_thread *t = &hr->hr_threads[n];
2073 struct ptlrpc_hr_args args;
2077 args.thread_index = n;
2078 args.cpu_index = cpu;
2081 rc = cfs_kernel_thread(ptlrpc_hr_main, (void*)&args,
2082 CLONE_VM|CLONE_FILES);
2084 complete(&t->hrt_completion);
2087 cfs_wait_event(t->hrt_wait, test_bit(HRT_RUNNING, &t->hrt_flags));
2093 static void ptlrpc_stop_hr_thread(struct ptlrpc_hr_thread *t)
2097 set_bit(HRT_STOPPING, &t->hrt_flags);
2098 cfs_waitq_signal(&t->hrt_wait);
2099 wait_for_completion(&t->hrt_completion);
2104 static void ptlrpc_stop_hr_threads(struct ptlrpc_hr_service *hrs)
2109 for (n = 0; n < hrs->hr_n_threads; n++)
2110 ptlrpc_stop_hr_thread(&hrs->hr_threads[n]);
2115 static int ptlrpc_start_hr_threads(struct ptlrpc_hr_service *hr)
2118 int n, cpu, threads_started = 0;
2121 LASSERT(hr != NULL);
2122 LASSERT(hr->hr_n_threads > 0);
2124 for (n = 0, cpu = 0; n < hr->hr_n_threads; n++) {
2125 #if defined(HAVE_NODE_TO_CPUMASK)
2126 while(!cpu_online(cpu)) {
2128 if (cpu >= num_possible_cpus())
2132 rc = ptlrpc_start_hr_thread(hr, n, cpu);
2138 if (threads_started == 0) {
2139 CERROR("No reply handling threads started\n");
2142 if (threads_started < hr->hr_n_threads) {
2143 CWARN("Started only %d reply handling threads from %d\n",
2144 threads_started, hr->hr_n_threads);
2145 hr->hr_n_threads = threads_started;
2150 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
2151 struct ptlrpc_thread *thread)
2153 struct l_wait_info lwi = { 0 };
2156 CDEBUG(D_RPCTRACE, "Stopping thread %p\n", thread);
2157 spin_lock(&svc->srv_lock);
2158 thread->t_flags = SVC_STOPPING;
2159 spin_unlock(&svc->srv_lock);
2161 cfs_waitq_broadcast(&svc->srv_waitq);
2162 l_wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED),
2165 spin_lock(&svc->srv_lock);
2166 list_del(&thread->t_link);
2167 spin_unlock(&svc->srv_lock);
2169 OBD_FREE_PTR(thread);
2173 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2175 struct ptlrpc_thread *thread;
2178 spin_lock(&svc->srv_lock);
2179 while (!list_empty(&svc->srv_threads)) {
2180 thread = list_entry(svc->srv_threads.next,
2181 struct ptlrpc_thread, t_link);
2183 spin_unlock(&svc->srv_lock);
2184 ptlrpc_stop_thread(svc, thread);
2185 spin_lock(&svc->srv_lock);
2188 spin_unlock(&svc->srv_lock);
2192 int ptlrpc_start_threads(struct obd_device *dev, struct ptlrpc_service *svc)
2197 /* We require 2 threads min - see note in
2198 ptlrpc_server_handle_request */
2199 LASSERT(svc->srv_threads_min >= 2);
2200 for (i = 0; i < svc->srv_threads_min; i++) {
2201 rc = ptlrpc_start_thread(dev, svc);
2202 /* We have enough threads, don't start more. b=15759 */
2206 CERROR("cannot start %s thread #%d: rc %d\n",
2207 svc->srv_thread_name, i, rc);
2208 ptlrpc_stop_all_threads(svc);
2214 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc)
2216 struct l_wait_info lwi = { 0 };
2217 struct ptlrpc_svc_data d;
2218 struct ptlrpc_thread *thread;
2223 CDEBUG(D_RPCTRACE, "%s started %d min %d max %d running %d\n",
2224 svc->srv_name, svc->srv_threads_started, svc->srv_threads_min,
2225 svc->srv_threads_max, svc->srv_threads_running);
2226 if (unlikely(svc->srv_threads_started >= svc->srv_threads_max) ||
2227 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
2228 svc->srv_threads_started == svc->srv_threads_min - 1))
2231 OBD_ALLOC_PTR(thread);
2234 cfs_waitq_init(&thread->t_ctl_waitq);
2236 spin_lock(&svc->srv_lock);
2237 if (svc->srv_threads_started >= svc->srv_threads_max) {
2238 spin_unlock(&svc->srv_lock);
2239 OBD_FREE_PTR(thread);
2242 list_add(&thread->t_link, &svc->srv_threads);
2243 id = svc->srv_threads_started++;
2244 spin_unlock(&svc->srv_lock);
2246 thread->t_svc = svc;
2248 sprintf(name, "%s_%02d", svc->srv_thread_name, id);
2254 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", name);
2256 /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
2257 * just drop the VM and FILES in cfs_daemonize_ctxt() right away.
2259 rc = cfs_kernel_thread(ptlrpc_main, &d, CLONE_VM | CLONE_FILES);
2261 CERROR("cannot start thread '%s': rc %d\n", name, rc);
2263 spin_lock(&svc->srv_lock);
2264 list_del(&thread->t_link);
2265 --svc->srv_threads_started;
2266 spin_unlock(&svc->srv_lock);
2268 OBD_FREE(thread, sizeof(*thread));
2271 l_wait_event(thread->t_ctl_waitq,
2272 thread->t_flags & (SVC_RUNNING | SVC_STOPPED), &lwi);
2274 rc = (thread->t_flags & SVC_STOPPED) ? thread->t_id : 0;
2279 int ptlrpc_hr_init(void)
2282 int n_cpus = num_online_cpus();
2283 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 rc = ptlrpc_start_hr_threads(hr);
2308 OBD_FREE(hr, hr->hr_size);
2314 void ptlrpc_hr_fini(void)
2316 if (ptlrpc_hr != NULL) {
2317 ptlrpc_stop_hr_threads(ptlrpc_hr);
2318 OBD_FREE(ptlrpc_hr, ptlrpc_hr->hr_size);
2323 #endif /* __KERNEL__ */
2326 * Wait until all already scheduled replies are processed.
2328 static void ptlrpc_wait_replies(struct ptlrpc_service *svc)
2332 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
2334 rc = l_wait_event(svc->srv_waitq,
2335 atomic_read(&svc->srv_n_difficult_replies) == 0,
2339 CWARN("Unexpectedly long timeout %p\n", svc);
2343 int ptlrpc_unregister_service(struct ptlrpc_service *service)
2346 struct l_wait_info lwi;
2347 struct list_head *tmp;
2348 struct ptlrpc_reply_state *rs, *t;
2349 struct ptlrpc_at_array *array = &service->srv_at_array;
2352 service->srv_is_stopping = 1;
2353 cfs_timer_disarm(&service->srv_at_timer);
2355 ptlrpc_stop_all_threads(service);
2356 LASSERT(list_empty(&service->srv_threads));
2358 spin_lock (&ptlrpc_all_services_lock);
2359 list_del_init (&service->srv_list);
2360 spin_unlock (&ptlrpc_all_services_lock);
2362 ptlrpc_lprocfs_unregister_service(service);
2364 /* All history will be culled when the next request buffer is
2366 service->srv_max_history_rqbds = 0;
2368 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
2370 rc = LNetClearLazyPortal(service->srv_req_portal);
2373 /* Unlink all the request buffers. This forces a 'final' event with
2374 * its 'unlink' flag set for each posted rqbd */
2375 list_for_each(tmp, &service->srv_active_rqbds) {
2376 struct ptlrpc_request_buffer_desc *rqbd =
2377 list_entry(tmp, struct ptlrpc_request_buffer_desc,
2380 rc = LNetMDUnlink(rqbd->rqbd_md_h);
2381 LASSERT (rc == 0 || rc == -ENOENT);
2384 /* Wait for the network to release any buffers it's currently
2387 spin_lock(&service->srv_lock);
2388 rc = service->srv_nrqbd_receiving;
2389 spin_unlock(&service->srv_lock);
2394 /* Network access will complete in finite time but the HUGE
2395 * timeout lets us CWARN for visibility of sluggish NALs */
2396 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2397 cfs_time_seconds(1), NULL, NULL);
2398 rc = l_wait_event(service->srv_waitq,
2399 service->srv_nrqbd_receiving == 0,
2401 if (rc == -ETIMEDOUT)
2402 CWARN("Service %s waiting for request buffers\n",
2406 /* schedule all outstanding replies to terminate them */
2407 spin_lock(&service->srv_lock);
2408 while (!list_empty(&service->srv_active_replies)) {
2409 struct ptlrpc_reply_state *rs =
2410 list_entry(service->srv_active_replies.next,
2411 struct ptlrpc_reply_state, rs_list);
2412 spin_lock(&rs->rs_lock);
2413 ptlrpc_schedule_difficult_reply(rs);
2414 spin_unlock(&rs->rs_lock);
2416 spin_unlock(&service->srv_lock);
2418 /* purge the request queue. NB No new replies (rqbds all unlinked)
2419 * and no service threads, so I'm the only thread noodling the
2420 * request queue now */
2421 while (!list_empty(&service->srv_req_in_queue)) {
2422 struct ptlrpc_request *req =
2423 list_entry(service->srv_req_in_queue.next,
2424 struct ptlrpc_request,
2427 list_del(&req->rq_list);
2428 service->srv_n_queued_reqs--;
2429 service->srv_n_active_reqs++;
2430 ptlrpc_server_finish_request(req);
2432 while (ptlrpc_server_request_pending(service, 1)) {
2433 struct ptlrpc_request *req;
2435 req = ptlrpc_server_request_get(service, 1);
2436 list_del(&req->rq_list);
2437 service->srv_n_queued_reqs--;
2438 service->srv_n_active_reqs++;
2439 ptlrpc_hpreq_fini(req);
2440 ptlrpc_server_finish_request(req);
2442 LASSERT(service->srv_n_queued_reqs == 0);
2443 LASSERT(service->srv_n_active_reqs == 0);
2444 LASSERT(service->srv_n_history_rqbds == 0);
2445 LASSERT(list_empty(&service->srv_active_rqbds));
2447 /* Now free all the request buffers since nothing references them
2449 while (!list_empty(&service->srv_idle_rqbds)) {
2450 struct ptlrpc_request_buffer_desc *rqbd =
2451 list_entry(service->srv_idle_rqbds.next,
2452 struct ptlrpc_request_buffer_desc,
2455 ptlrpc_free_rqbd(rqbd);
2458 ptlrpc_wait_replies(service);
2460 list_for_each_entry_safe(rs, t, &service->srv_free_rs_list, rs_list) {
2461 list_del(&rs->rs_list);
2462 OBD_FREE(rs, service->srv_max_reply_size);
2465 /* In case somebody rearmed this in the meantime */
2466 cfs_timer_disarm(&service->srv_at_timer);
2468 if (array->paa_reqs_array != NULL) {
2469 OBD_FREE(array->paa_reqs_array,
2470 sizeof(struct list_head) * array->paa_size);
2471 array->paa_reqs_array = NULL;
2474 if (array->paa_reqs_count != NULL) {
2475 OBD_FREE(array->paa_reqs_count,
2476 sizeof(__u32) * array->paa_size);
2477 array->paa_reqs_count= NULL;
2480 OBD_FREE_PTR(service);
2484 /* Returns 0 if the service is healthy.
2486 * Right now, it just checks to make sure that requests aren't languishing
2487 * in the queue. We'll use this health check to govern whether a node needs
2488 * to be shot, so it's intentionally non-aggressive. */
2489 int ptlrpc_service_health_check(struct ptlrpc_service *svc)
2491 struct ptlrpc_request *request;
2492 struct timeval right_now;
2498 do_gettimeofday(&right_now);
2500 spin_lock(&svc->srv_lock);
2501 if (!ptlrpc_server_request_pending(svc, 1)) {
2502 spin_unlock(&svc->srv_lock);
2506 /* How long has the next entry been waiting? */
2507 if (list_empty(&svc->srv_request_queue))
2508 request = list_entry(svc->srv_request_hpq.next,
2509 struct ptlrpc_request, rq_list);
2511 request = list_entry(svc->srv_request_queue.next,
2512 struct ptlrpc_request, rq_list);
2513 timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
2514 spin_unlock(&svc->srv_lock);
2516 if ((timediff / ONE_MILLION) > (AT_OFF ? obd_timeout * 3/2 :
2518 CERROR("%s: unhealthy - request has been waiting %lds\n",
2519 svc->srv_name, timediff / ONE_MILLION);