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 (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
32 * Copyright (c) 2011 Whamcloud, Inc.
36 * This file is part of Lustre, http://www.lustre.org/
37 * Lustre is a trademark of Sun Microsystems, Inc.
40 #define DEBUG_SUBSYSTEM S_RPC
42 #include <liblustre.h>
44 #include <obd_support.h>
45 #include <obd_class.h>
46 #include <lustre_net.h>
47 #include <lu_object.h>
48 #include <lnet/types.h>
49 #include "ptlrpc_internal.h"
51 /* The following are visible and mutable through /sys/module/ptlrpc */
52 int test_req_buffer_pressure = 0;
53 CFS_MODULE_PARM(test_req_buffer_pressure, "i", int, 0444,
54 "set non-zero to put pressure on request buffer pools");
55 CFS_MODULE_PARM(at_min, "i", int, 0644,
56 "Adaptive timeout minimum (sec)");
57 CFS_MODULE_PARM(at_max, "i", int, 0644,
58 "Adaptive timeout maximum (sec)");
59 CFS_MODULE_PARM(at_history, "i", int, 0644,
60 "Adaptive timeouts remember the slowest event that took place "
61 "within this period (sec)");
62 CFS_MODULE_PARM(at_early_margin, "i", int, 0644,
63 "How soon before an RPC deadline to send an early reply");
64 CFS_MODULE_PARM(at_extra, "i", int, 0644,
65 "How much extra time to give with each early reply");
69 static int ptlrpc_server_post_idle_rqbds (struct ptlrpc_service *svc);
70 static void ptlrpc_hpreq_fini(struct ptlrpc_request *req);
72 static CFS_LIST_HEAD(ptlrpc_all_services);
73 cfs_spinlock_t ptlrpc_all_services_lock;
75 struct ptlrpc_request_buffer_desc *
76 ptlrpc_alloc_rqbd (struct ptlrpc_service *svc)
78 struct ptlrpc_request_buffer_desc *rqbd;
84 rqbd->rqbd_service = svc;
85 rqbd->rqbd_refcount = 0;
86 rqbd->rqbd_cbid.cbid_fn = request_in_callback;
87 rqbd->rqbd_cbid.cbid_arg = rqbd;
88 CFS_INIT_LIST_HEAD(&rqbd->rqbd_reqs);
89 OBD_ALLOC_LARGE(rqbd->rqbd_buffer, svc->srv_buf_size);
91 if (rqbd->rqbd_buffer == NULL) {
96 cfs_spin_lock(&svc->srv_lock);
97 cfs_list_add(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
99 cfs_spin_unlock(&svc->srv_lock);
105 ptlrpc_free_rqbd (struct ptlrpc_request_buffer_desc *rqbd)
107 struct ptlrpc_service *svc = rqbd->rqbd_service;
109 LASSERT (rqbd->rqbd_refcount == 0);
110 LASSERT (cfs_list_empty(&rqbd->rqbd_reqs));
112 cfs_spin_lock(&svc->srv_lock);
113 cfs_list_del(&rqbd->rqbd_list);
115 cfs_spin_unlock(&svc->srv_lock);
117 OBD_FREE_LARGE(rqbd->rqbd_buffer, svc->srv_buf_size);
122 ptlrpc_grow_req_bufs(struct ptlrpc_service *svc)
124 struct ptlrpc_request_buffer_desc *rqbd;
128 for (i = 0; i < svc->srv_nbuf_per_group; i++) {
129 /* NB: another thread might be doing this as well, we need to
130 * make sure that it wouldn't over-allocate, see LU-1212. */
131 if (svc->srv_nrqbd_receiving >= svc->srv_nbuf_per_group)
134 rqbd = ptlrpc_alloc_rqbd(svc);
137 CERROR("%s: Can't allocate request buffer\n",
143 if (ptlrpc_server_post_idle_rqbds(svc) < 0) {
150 "%s: allocate %d new %d-byte reqbufs (%d/%d left), rc = %d\n",
151 svc->srv_name, i, svc->srv_buf_size,
152 svc->srv_nrqbd_receiving, svc->srv_nbufs, rc);
158 * Part of Rep-Ack logic.
159 * Puts a lock and its mode into reply state assotiated to request reply.
162 ptlrpc_save_lock(struct ptlrpc_request *req,
163 struct lustre_handle *lock, int mode, int no_ack)
165 struct ptlrpc_reply_state *rs = req->rq_reply_state;
169 LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
171 if (req->rq_export->exp_disconnected) {
172 ldlm_lock_decref(lock, mode);
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;
184 #define HRT_RUNNING 0
185 #define HRT_STOPPING 1
187 struct ptlrpc_hr_thread {
188 cfs_spinlock_t hrt_lock;
189 unsigned long hrt_flags;
190 cfs_waitq_t hrt_wait;
191 cfs_list_t hrt_queue;
192 cfs_completion_t hrt_completion;
195 struct ptlrpc_hr_service {
199 struct ptlrpc_hr_thread hr_threads[0];
203 cfs_list_t rsb_replies;
204 struct ptlrpc_service *rsb_svc;
205 unsigned int rsb_n_replies;
209 * A pointer to per-node reply handling service.
211 static struct ptlrpc_hr_service *ptlrpc_hr = NULL;
214 * maximum mumber of replies scheduled in one batch
216 #define MAX_SCHEDULED 256
219 * Initialize a reply batch.
223 static void rs_batch_init(struct rs_batch *b)
225 memset(b, 0, sizeof *b);
226 CFS_INIT_LIST_HEAD(&b->rsb_replies);
230 * Choose an hr thread to dispatch requests to.
232 static unsigned int get_hr_thread_index(struct ptlrpc_hr_service *hr)
236 /* Concurrent modification of hr_index w/o any spinlock
237 protection is harmless as long as the result fits
238 [0..(hr_n_threads-1)] range and each thread gets near equal
241 hr->hr_index = (idx >= hr->hr_n_threads - 1) ? 0 : idx + 1;
246 * Dispatch all replies accumulated in the batch to one from
247 * dedicated reply handling threads.
251 static void rs_batch_dispatch(struct rs_batch *b)
253 if (b->rsb_n_replies != 0) {
254 struct ptlrpc_hr_service *hr = ptlrpc_hr;
257 idx = get_hr_thread_index(hr);
259 cfs_spin_lock(&hr->hr_threads[idx].hrt_lock);
260 cfs_list_splice_init(&b->rsb_replies,
261 &hr->hr_threads[idx].hrt_queue);
262 cfs_spin_unlock(&hr->hr_threads[idx].hrt_lock);
263 cfs_waitq_signal(&hr->hr_threads[idx].hrt_wait);
264 b->rsb_n_replies = 0;
269 * Add a reply to a batch.
270 * Add one reply object to a batch, schedule batched replies if overload.
275 static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs)
277 struct ptlrpc_service *svc = rs->rs_service;
279 if (svc != b->rsb_svc || b->rsb_n_replies >= MAX_SCHEDULED) {
280 if (b->rsb_svc != NULL) {
281 rs_batch_dispatch(b);
282 cfs_spin_unlock(&b->rsb_svc->srv_rs_lock);
284 cfs_spin_lock(&svc->srv_rs_lock);
287 cfs_spin_lock(&rs->rs_lock);
288 rs->rs_scheduled_ever = 1;
289 if (rs->rs_scheduled == 0) {
290 cfs_list_move(&rs->rs_list, &b->rsb_replies);
291 rs->rs_scheduled = 1;
294 rs->rs_committed = 1;
295 cfs_spin_unlock(&rs->rs_lock);
299 * Reply batch finalization.
300 * Dispatch remaining replies from the batch
301 * and release remaining spinlock.
305 static void rs_batch_fini(struct rs_batch *b)
307 if (b->rsb_svc != 0) {
308 rs_batch_dispatch(b);
309 cfs_spin_unlock(&b->rsb_svc->srv_rs_lock);
313 #define DECLARE_RS_BATCH(b) struct rs_batch b
315 #else /* __KERNEL__ */
317 #define rs_batch_init(b) do{}while(0)
318 #define rs_batch_fini(b) do{}while(0)
319 #define rs_batch_add(b, r) ptlrpc_schedule_difficult_reply(r)
320 #define DECLARE_RS_BATCH(b)
322 #endif /* __KERNEL__ */
325 * Put reply state into a queue for processing because we received
326 * ACK from the client
328 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
331 struct ptlrpc_hr_service *hr = ptlrpc_hr;
335 LASSERT(cfs_list_empty(&rs->rs_list));
337 idx = get_hr_thread_index(hr);
338 cfs_spin_lock(&hr->hr_threads[idx].hrt_lock);
339 cfs_list_add_tail(&rs->rs_list, &hr->hr_threads[idx].hrt_queue);
340 cfs_spin_unlock(&hr->hr_threads[idx].hrt_lock);
341 cfs_waitq_signal(&hr->hr_threads[idx].hrt_wait);
344 cfs_list_add_tail(&rs->rs_list, &rs->rs_service->srv_reply_queue);
349 ptlrpc_schedule_difficult_reply (struct ptlrpc_reply_state *rs)
353 LASSERT_SPIN_LOCKED(&rs->rs_service->srv_rs_lock);
354 LASSERT_SPIN_LOCKED(&rs->rs_lock);
355 LASSERT (rs->rs_difficult);
356 rs->rs_scheduled_ever = 1; /* flag any notification attempt */
358 if (rs->rs_scheduled) { /* being set up or already notified */
363 rs->rs_scheduled = 1;
364 cfs_list_del_init(&rs->rs_list);
365 ptlrpc_dispatch_difficult_reply(rs);
369 void ptlrpc_commit_replies(struct obd_export *exp)
371 struct ptlrpc_reply_state *rs, *nxt;
372 DECLARE_RS_BATCH(batch);
375 rs_batch_init(&batch);
376 /* Find any replies that have been committed and get their service
377 * to attend to complete them. */
379 /* CAVEAT EMPTOR: spinlock ordering!!! */
380 cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
381 cfs_list_for_each_entry_safe(rs, nxt, &exp->exp_uncommitted_replies,
383 LASSERT (rs->rs_difficult);
384 /* VBR: per-export last_committed */
385 LASSERT(rs->rs_export);
386 if (rs->rs_transno <= exp->exp_last_committed) {
387 cfs_list_del_init(&rs->rs_obd_list);
388 rs_batch_add(&batch, rs);
391 cfs_spin_unlock(&exp->exp_uncommitted_replies_lock);
392 rs_batch_fini(&batch);
397 ptlrpc_server_post_idle_rqbds (struct ptlrpc_service *svc)
399 struct ptlrpc_request_buffer_desc *rqbd;
404 cfs_spin_lock(&svc->srv_lock);
406 if (cfs_list_empty (&svc->srv_idle_rqbds)) {
407 cfs_spin_unlock(&svc->srv_lock);
411 rqbd = cfs_list_entry(svc->srv_idle_rqbds.next,
412 struct ptlrpc_request_buffer_desc,
414 cfs_list_del (&rqbd->rqbd_list);
416 /* assume we will post successfully */
417 svc->srv_nrqbd_receiving++;
418 cfs_list_add (&rqbd->rqbd_list, &svc->srv_active_rqbds);
420 cfs_spin_unlock(&svc->srv_lock);
422 rc = ptlrpc_register_rqbd(rqbd);
429 cfs_spin_lock(&svc->srv_lock);
431 svc->srv_nrqbd_receiving--;
432 cfs_list_del(&rqbd->rqbd_list);
433 cfs_list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
435 /* Don't complain if no request buffers are posted right now; LNET
436 * won't drop requests because we set the portal lazy! */
438 cfs_spin_unlock(&svc->srv_lock);
444 * Start a service with parameters from struct ptlrpc_service_conf \a c
445 * as opposed to directly calling ptlrpc_init_svc with tons of arguments.
447 struct ptlrpc_service *ptlrpc_init_svc_conf(struct ptlrpc_service_conf *c,
448 svc_handler_t h, char *name,
449 struct proc_dir_entry *proc_entry,
450 svc_req_printfn_t prntfn,
453 return ptlrpc_init_svc(c->psc_nbufs, c->psc_bufsize,
454 c->psc_max_req_size, c->psc_max_reply_size,
455 c->psc_req_portal, c->psc_rep_portal,
456 c->psc_watchdog_factor,
458 prntfn, c->psc_min_threads, c->psc_max_threads,
459 threadname, c->psc_ctx_tags, NULL);
461 EXPORT_SYMBOL(ptlrpc_init_svc_conf);
463 static void ptlrpc_at_timer(unsigned long castmeharder)
465 struct ptlrpc_service *svc = (struct ptlrpc_service *)castmeharder;
466 svc->srv_at_check = 1;
467 svc->srv_at_checktime = cfs_time_current();
468 cfs_waitq_signal(&svc->srv_waitq);
472 * Initialize service on a given portal.
473 * This includes starting serving threads , allocating and posting rqbds and
475 * \a nbufs is how many buffers to post
476 * \a bufsize is buffer size to post
477 * \a max_req_size - maximum request size to be accepted for this service
478 * \a max_reply_size maximum reply size this service can ever send
479 * \a req_portal - portal to listed for requests on
480 * \a rep_portal - portal of where to send replies to
481 * \a watchdog_factor soft watchdog timeout multiplifier to print stuck service traces.
482 * \a handler - function to process every new request
483 * \a name - service name
484 * \a proc_entry - entry in the /proc tree for sttistics reporting
485 * \a min_threads \a max_threads - min/max number of service threads to start.
486 * \a threadname should be 11 characters or less - 3 will be added on
487 * \a hp_handler - function to determine priority of the request, also called
488 * on every new request.
490 struct ptlrpc_service *
491 ptlrpc_init_svc(int nbufs, int bufsize, int max_req_size, int max_reply_size,
492 int req_portal, int rep_portal, int watchdog_factor,
493 svc_handler_t handler, char *name,
494 cfs_proc_dir_entry_t *proc_entry,
495 svc_req_printfn_t svcreq_printfn,
496 int min_threads, int max_threads,
497 char *threadname, __u32 ctx_tags,
498 svc_hpreq_handler_t hp_handler)
501 struct ptlrpc_at_array *array;
502 struct ptlrpc_service *service;
503 unsigned int size, index;
507 LASSERT (bufsize >= max_req_size + SPTLRPC_MAX_PAYLOAD);
508 LASSERT (ctx_tags != 0);
510 OBD_ALLOC_PTR(service);
514 /* First initialise enough for early teardown */
516 service->srv_name = name;
517 cfs_spin_lock_init(&service->srv_lock);
518 cfs_spin_lock_init(&service->srv_rq_lock);
519 cfs_spin_lock_init(&service->srv_rs_lock);
520 CFS_INIT_LIST_HEAD(&service->srv_threads);
521 cfs_waitq_init(&service->srv_waitq);
523 service->srv_nbuf_per_group = test_req_buffer_pressure ? 1 : nbufs;
524 service->srv_max_req_size = max_req_size + SPTLRPC_MAX_PAYLOAD;
525 service->srv_buf_size = bufsize;
526 service->srv_rep_portal = rep_portal;
527 service->srv_req_portal = req_portal;
528 service->srv_watchdog_factor = watchdog_factor;
529 service->srv_handler = handler;
530 service->srv_req_printfn = svcreq_printfn;
531 service->srv_request_seq = 1; /* valid seq #s start at 1 */
532 service->srv_request_max_cull_seq = 0;
533 service->srv_threads_min = min_threads;
534 service->srv_threads_max = max_threads;
535 service->srv_thread_name = threadname;
536 service->srv_ctx_tags = ctx_tags;
537 service->srv_hpreq_handler = hp_handler;
538 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
539 service->srv_hpreq_count = 0;
540 service->srv_n_active_hpreq = 0;
542 rc = LNetSetLazyPortal(service->srv_req_portal);
545 CFS_INIT_LIST_HEAD(&service->srv_request_queue);
546 CFS_INIT_LIST_HEAD(&service->srv_request_hpq);
547 CFS_INIT_LIST_HEAD(&service->srv_idle_rqbds);
548 CFS_INIT_LIST_HEAD(&service->srv_active_rqbds);
549 CFS_INIT_LIST_HEAD(&service->srv_history_rqbds);
550 CFS_INIT_LIST_HEAD(&service->srv_request_history);
551 CFS_INIT_LIST_HEAD(&service->srv_active_replies);
553 CFS_INIT_LIST_HEAD(&service->srv_reply_queue);
555 CFS_INIT_LIST_HEAD(&service->srv_free_rs_list);
556 cfs_waitq_init(&service->srv_free_rs_waitq);
557 cfs_atomic_set(&service->srv_n_difficult_replies, 0);
559 cfs_spin_lock_init(&service->srv_at_lock);
560 CFS_INIT_LIST_HEAD(&service->srv_req_in_queue);
562 array = &service->srv_at_array;
563 size = at_est2timeout(at_max);
564 array->paa_size = size;
565 array->paa_count = 0;
566 array->paa_deadline = -1;
568 /* allocate memory for srv_at_array (ptlrpc_at_array) */
569 OBD_ALLOC(array->paa_reqs_array, sizeof(cfs_list_t) * size);
570 if (array->paa_reqs_array == NULL)
573 for (index = 0; index < size; index++)
574 CFS_INIT_LIST_HEAD(&array->paa_reqs_array[index]);
576 OBD_ALLOC(array->paa_reqs_count, sizeof(__u32) * size);
577 if (array->paa_reqs_count == NULL)
580 cfs_timer_init(&service->srv_at_timer, ptlrpc_at_timer, service);
581 /* At SOW, service time should be quick; 10s seems generous. If client
582 timeout is less than this, we'll be sending an early reply. */
583 at_init(&service->srv_at_estimate, 10, 0);
585 cfs_spin_lock (&ptlrpc_all_services_lock);
586 cfs_list_add (&service->srv_list, &ptlrpc_all_services);
587 cfs_spin_unlock (&ptlrpc_all_services_lock);
589 /* Now allocate the request buffers */
590 rc = ptlrpc_grow_req_bufs(service);
591 /* We shouldn't be under memory pressure at startup, so
592 * fail if we can't post all our buffers at this time. */
596 /* Now allocate pool of reply buffers */
597 /* Increase max reply size to next power of two */
598 service->srv_max_reply_size = 1;
599 while (service->srv_max_reply_size <
600 max_reply_size + SPTLRPC_MAX_PAYLOAD)
601 service->srv_max_reply_size <<= 1;
603 if (proc_entry != NULL)
604 ptlrpc_lprocfs_register_service(proc_entry, service);
606 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
607 service->srv_name, service->srv_req_portal);
611 ptlrpc_unregister_service(service);
616 * to actually free the request, must be called without holding svc_lock.
617 * note it's caller's responsibility to unlink req->rq_list.
619 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
621 LASSERT(cfs_atomic_read(&req->rq_refcount) == 0);
622 LASSERT(cfs_list_empty(&req->rq_timed_list));
624 /* DEBUG_REQ() assumes the reply state of a request with a valid
625 * ref will not be destroyed until that reference is dropped. */
626 ptlrpc_req_drop_rs(req);
628 sptlrpc_svc_ctx_decref(req);
630 if (req != &req->rq_rqbd->rqbd_req) {
631 /* NB request buffers use an embedded
632 * req if the incoming req unlinked the
633 * MD; this isn't one of them! */
634 OBD_FREE(req, sizeof(*req));
639 * drop a reference count of the request. if it reaches 0, we either
640 * put it into history list, or free it immediately.
642 void ptlrpc_server_drop_request(struct ptlrpc_request *req)
644 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
645 struct ptlrpc_service *svc = rqbd->rqbd_service;
650 if (!cfs_atomic_dec_and_test(&req->rq_refcount))
653 cfs_spin_lock(&svc->srv_at_lock);
654 if (req->rq_at_linked) {
655 struct ptlrpc_at_array *array = &svc->srv_at_array;
656 __u32 index = req->rq_at_index;
658 LASSERT(!cfs_list_empty(&req->rq_timed_list));
659 cfs_list_del_init(&req->rq_timed_list);
660 cfs_spin_lock(&req->rq_lock);
661 req->rq_at_linked = 0;
662 cfs_spin_unlock(&req->rq_lock);
663 array->paa_reqs_count[index]--;
666 LASSERT(cfs_list_empty(&req->rq_timed_list));
667 cfs_spin_unlock(&svc->srv_at_lock);
669 /* finalize request */
670 if (req->rq_export) {
671 class_export_put(req->rq_export);
672 req->rq_export = NULL;
675 cfs_spin_lock(&svc->srv_lock);
677 cfs_list_add(&req->rq_list, &rqbd->rqbd_reqs);
679 refcount = --(rqbd->rqbd_refcount);
681 /* request buffer is now idle: add to history */
682 cfs_list_del(&rqbd->rqbd_list);
683 cfs_list_add_tail(&rqbd->rqbd_list, &svc->srv_history_rqbds);
684 svc->srv_n_history_rqbds++;
686 /* cull some history?
687 * I expect only about 1 or 2 rqbds need to be recycled here */
688 while (svc->srv_n_history_rqbds > svc->srv_max_history_rqbds) {
689 rqbd = cfs_list_entry(svc->srv_history_rqbds.next,
690 struct ptlrpc_request_buffer_desc,
693 cfs_list_del(&rqbd->rqbd_list);
694 svc->srv_n_history_rqbds--;
696 /* remove rqbd's reqs from svc's req history while
697 * I've got the service lock */
698 cfs_list_for_each(tmp, &rqbd->rqbd_reqs) {
699 req = cfs_list_entry(tmp, struct ptlrpc_request,
701 /* Track the highest culled req seq */
702 if (req->rq_history_seq >
703 svc->srv_request_max_cull_seq)
704 svc->srv_request_max_cull_seq =
706 cfs_list_del(&req->rq_history_list);
709 cfs_spin_unlock(&svc->srv_lock);
711 cfs_list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
712 req = cfs_list_entry(rqbd->rqbd_reqs.next,
713 struct ptlrpc_request,
715 cfs_list_del(&req->rq_list);
716 ptlrpc_server_free_request(req);
719 cfs_spin_lock(&svc->srv_lock);
721 * now all reqs including the embedded req has been
722 * disposed, schedule request buffer for re-use.
724 LASSERT(cfs_atomic_read(&rqbd->rqbd_req.rq_refcount) ==
726 cfs_list_add_tail(&rqbd->rqbd_list,
727 &svc->srv_idle_rqbds);
730 cfs_spin_unlock(&svc->srv_lock);
731 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
732 /* If we are low on memory, we are not interested in history */
733 cfs_list_del(&req->rq_list);
734 cfs_list_del_init(&req->rq_history_list);
735 cfs_spin_unlock(&svc->srv_lock);
737 ptlrpc_server_free_request(req);
739 cfs_spin_unlock(&svc->srv_lock);
744 * to finish a request: stop sending more early replies, and release
745 * the request. should be called after we finished handling the request.
747 static void ptlrpc_server_finish_request(struct ptlrpc_service *svc,
748 struct ptlrpc_request *req)
750 ptlrpc_hpreq_fini(req);
752 cfs_spin_lock(&svc->srv_rq_lock);
753 svc->srv_n_active_reqs--;
755 svc->srv_n_active_hpreq--;
756 cfs_spin_unlock(&svc->srv_rq_lock);
758 ptlrpc_server_drop_request(req);
762 * This function makes sure dead exports are evicted in a timely manner.
763 * This function is only called when some export receives a message (i.e.,
764 * the network is up.)
766 static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
768 struct obd_export *oldest_exp;
769 time_t oldest_time, new_time;
775 /* Compensate for slow machines, etc, by faking our request time
776 into the future. Although this can break the strict time-ordering
777 of the list, we can be really lazy here - we don't have to evict
778 at the exact right moment. Eventually, all silent exports
779 will make it to the top of the list. */
781 /* Do not pay attention on 1sec or smaller renewals. */
782 new_time = cfs_time_current_sec() + extra_delay;
783 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
786 exp->exp_last_request_time = new_time;
787 CDEBUG(D_HA, "updating export %s at "CFS_TIME_T" exp %p\n",
788 exp->exp_client_uuid.uuid,
789 exp->exp_last_request_time, exp);
791 /* exports may get disconnected from the chain even though the
792 export has references, so we must keep the spin lock while
793 manipulating the lists */
794 cfs_spin_lock(&exp->exp_obd->obd_dev_lock);
796 if (cfs_list_empty(&exp->exp_obd_chain_timed)) {
797 /* this one is not timed */
798 cfs_spin_unlock(&exp->exp_obd->obd_dev_lock);
802 cfs_list_move_tail(&exp->exp_obd_chain_timed,
803 &exp->exp_obd->obd_exports_timed);
805 oldest_exp = cfs_list_entry(exp->exp_obd->obd_exports_timed.next,
806 struct obd_export, exp_obd_chain_timed);
807 oldest_time = oldest_exp->exp_last_request_time;
808 cfs_spin_unlock(&exp->exp_obd->obd_dev_lock);
810 if (exp->exp_obd->obd_recovering) {
811 /* be nice to everyone during recovery */
816 /* Note - racing to start/reset the obd_eviction timer is safe */
817 if (exp->exp_obd->obd_eviction_timer == 0) {
818 /* Check if the oldest entry is expired. */
819 if (cfs_time_current_sec() > (oldest_time + PING_EVICT_TIMEOUT +
821 /* We need a second timer, in case the net was down and
822 * it just came back. Since the pinger may skip every
823 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
824 * we better wait for 3. */
825 exp->exp_obd->obd_eviction_timer =
826 cfs_time_current_sec() + 3 * PING_INTERVAL;
827 CDEBUG(D_HA, "%s: Think about evicting %s from "CFS_TIME_T"\n",
828 exp->exp_obd->obd_name,
829 obd_export_nid2str(oldest_exp), oldest_time);
832 if (cfs_time_current_sec() >
833 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
834 /* The evictor won't evict anyone who we've heard from
835 * recently, so we don't have to check before we start
837 if (!ping_evictor_wake(exp))
838 exp->exp_obd->obd_eviction_timer = 0;
846 * Sanity check request \a req.
847 * Return 0 if all is ok, error code otherwise.
849 static int ptlrpc_check_req(struct ptlrpc_request *req)
853 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
854 req->rq_export->exp_conn_cnt)) {
855 DEBUG_REQ(D_ERROR, req,
856 "DROPPING req from old connection %d < %d",
857 lustre_msg_get_conn_cnt(req->rq_reqmsg),
858 req->rq_export->exp_conn_cnt);
861 if (unlikely(req->rq_export->exp_obd &&
862 req->rq_export->exp_obd->obd_fail)) {
863 /* Failing over, don't handle any more reqs, send
864 error response instead. */
865 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
866 req, req->rq_export->exp_obd->obd_name);
868 } else if (lustre_msg_get_flags(req->rq_reqmsg) &
869 (MSG_REPLAY | MSG_REQ_REPLAY_DONE) &&
870 !(req->rq_export->exp_obd->obd_recovering)) {
871 DEBUG_REQ(D_ERROR, req,
872 "Invalid replay without recovery");
873 class_fail_export(req->rq_export);
875 } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 &&
876 !(req->rq_export->exp_obd->obd_recovering)) {
877 DEBUG_REQ(D_ERROR, req, "Invalid req with transno "
878 LPU64" without recovery",
879 lustre_msg_get_transno(req->rq_reqmsg));
880 class_fail_export(req->rq_export);
884 if (unlikely(rc < 0)) {
891 static void ptlrpc_at_set_timer(struct ptlrpc_service *svc)
893 struct ptlrpc_at_array *array = &svc->srv_at_array;
896 cfs_spin_lock(&svc->srv_at_lock);
897 if (array->paa_count == 0) {
898 cfs_timer_disarm(&svc->srv_at_timer);
899 cfs_spin_unlock(&svc->srv_at_lock);
903 /* Set timer for closest deadline */
904 next = (__s32)(array->paa_deadline - cfs_time_current_sec() -
907 ptlrpc_at_timer((unsigned long)svc);
909 cfs_timer_arm(&svc->srv_at_timer, cfs_time_shift(next));
910 cfs_spin_unlock(&svc->srv_at_lock);
911 CDEBUG(D_INFO, "armed %s at %+ds\n", svc->srv_name, next);
914 /* Add rpc to early reply check list */
915 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
917 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
918 struct ptlrpc_request *rq = NULL;
919 struct ptlrpc_at_array *array = &svc->srv_at_array;
926 if (req->rq_no_reply)
929 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
932 cfs_spin_lock(&svc->srv_at_lock);
933 LASSERT(cfs_list_empty(&req->rq_timed_list));
935 index = (unsigned long)req->rq_deadline % array->paa_size;
936 if (array->paa_reqs_count[index] > 0) {
937 /* latest rpcs will have the latest deadlines in the list,
938 * so search backward. */
939 cfs_list_for_each_entry_reverse(rq,
940 &array->paa_reqs_array[index],
942 if (req->rq_deadline >= rq->rq_deadline) {
943 cfs_list_add(&req->rq_timed_list,
950 /* Add the request at the head of the list */
951 if (cfs_list_empty(&req->rq_timed_list))
952 cfs_list_add(&req->rq_timed_list,
953 &array->paa_reqs_array[index]);
955 cfs_spin_lock(&req->rq_lock);
956 req->rq_at_linked = 1;
957 cfs_spin_unlock(&req->rq_lock);
958 req->rq_at_index = index;
959 array->paa_reqs_count[index]++;
961 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
962 array->paa_deadline = req->rq_deadline;
965 cfs_spin_unlock(&svc->srv_at_lock);
968 ptlrpc_at_set_timer(svc);
973 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
975 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
976 struct ptlrpc_request *reqcopy;
977 struct lustre_msg *reqmsg;
978 cfs_duration_t olddl = req->rq_deadline - cfs_time_current_sec();
983 /* deadline is when the client expects us to reply, margin is the
984 difference between clients' and servers' expectations */
985 DEBUG_REQ(D_ADAPTTO, req,
986 "%ssending early reply (deadline %+lds, margin %+lds) for "
987 "%d+%d", AT_OFF ? "AT off - not " : "",
988 olddl, olddl - at_get(&svc->srv_at_estimate),
989 at_get(&svc->srv_at_estimate), at_extra);
995 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), "
996 "not sending early reply. Consider increasing "
997 "at_early_margin (%d)?", olddl, at_early_margin);
999 /* Return an error so we're not re-added to the timed list. */
1003 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
1004 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
1005 "but no AT support");
1009 if (req->rq_export &&
1010 lustre_msg_get_flags(req->rq_reqmsg) &
1011 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1012 /* During recovery, we don't want to send too many early
1013 * replies, but on the other hand we want to make sure the
1014 * client has enough time to resend if the rpc is lost. So
1015 * during the recovery period send at least 4 early replies,
1016 * spacing them every at_extra if we can. at_estimate should
1017 * always equal this fixed value during recovery. */
1018 at_measured(&svc->srv_at_estimate, min(at_extra,
1019 req->rq_export->exp_obd->obd_recovery_timeout / 4));
1021 /* Fake our processing time into the future to ask the clients
1022 * for some extra amount of time */
1023 at_measured(&svc->srv_at_estimate, at_extra +
1024 cfs_time_current_sec() -
1025 req->rq_arrival_time.tv_sec);
1027 /* Check to see if we've actually increased the deadline -
1028 * we may be past adaptive_max */
1029 if (req->rq_deadline >= req->rq_arrival_time.tv_sec +
1030 at_get(&svc->srv_at_estimate)) {
1031 DEBUG_REQ(D_WARNING, req, "Couldn't add any time "
1032 "(%ld/%ld), not sending early reply\n",
1033 olddl, req->rq_arrival_time.tv_sec +
1034 at_get(&svc->srv_at_estimate) -
1035 cfs_time_current_sec());
1039 newdl = cfs_time_current_sec() + at_get(&svc->srv_at_estimate);
1041 OBD_ALLOC(reqcopy, sizeof *reqcopy);
1042 if (reqcopy == NULL)
1044 OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1046 OBD_FREE(reqcopy, sizeof *reqcopy);
1051 reqcopy->rq_reply_state = NULL;
1052 reqcopy->rq_rep_swab_mask = 0;
1053 reqcopy->rq_pack_bulk = 0;
1054 reqcopy->rq_pack_udesc = 0;
1055 reqcopy->rq_packed_final = 0;
1056 sptlrpc_svc_ctx_addref(reqcopy);
1057 /* We only need the reqmsg for the magic */
1058 reqcopy->rq_reqmsg = reqmsg;
1059 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1061 LASSERT(cfs_atomic_read(&req->rq_refcount));
1062 /** if it is last refcount then early reply isn't needed */
1063 if (cfs_atomic_read(&req->rq_refcount) == 1) {
1064 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
1065 "abort sending early reply\n");
1066 GOTO(out, rc = -EINVAL);
1069 /* Connection ref */
1070 reqcopy->rq_export = class_conn2export(
1071 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1072 if (reqcopy->rq_export == NULL)
1073 GOTO(out, rc = -ENODEV);
1076 class_export_rpc_get(reqcopy->rq_export);
1077 if (reqcopy->rq_export->exp_obd &&
1078 reqcopy->rq_export->exp_obd->obd_fail)
1079 GOTO(out_put, rc = -ENODEV);
1081 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1085 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1088 /* Adjust our own deadline to what we told the client */
1089 req->rq_deadline = newdl;
1090 req->rq_early_count++; /* number sent, server side */
1092 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1095 /* Free the (early) reply state from lustre_pack_reply.
1096 (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
1097 ptlrpc_req_drop_rs(reqcopy);
1100 class_export_rpc_put(reqcopy->rq_export);
1101 class_export_put(reqcopy->rq_export);
1103 sptlrpc_svc_ctx_decref(reqcopy);
1104 OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1105 OBD_FREE(reqcopy, sizeof *reqcopy);
1109 /* Send early replies to everybody expiring within at_early_margin
1110 asking for at_extra time */
1111 static int ptlrpc_at_check_timed(struct ptlrpc_service *svc)
1113 struct ptlrpc_request *rq, *n;
1114 cfs_list_t work_list;
1115 struct ptlrpc_at_array *array = &svc->srv_at_array;
1118 time_t now = cfs_time_current_sec();
1119 cfs_duration_t delay;
1120 int first, counter = 0;
1123 cfs_spin_lock(&svc->srv_at_lock);
1124 if (svc->srv_at_check == 0) {
1125 cfs_spin_unlock(&svc->srv_at_lock);
1128 delay = cfs_time_sub(cfs_time_current(), svc->srv_at_checktime);
1129 svc->srv_at_check = 0;
1131 if (array->paa_count == 0) {
1132 cfs_spin_unlock(&svc->srv_at_lock);
1136 /* The timer went off, but maybe the nearest rpc already completed. */
1137 first = array->paa_deadline - now;
1138 if (first > at_early_margin) {
1139 /* We've still got plenty of time. Reset the timer. */
1140 cfs_spin_unlock(&svc->srv_at_lock);
1141 ptlrpc_at_set_timer(svc);
1145 /* We're close to a timeout, and we don't know how much longer the
1146 server will take. Send early replies to everyone expiring soon. */
1147 CFS_INIT_LIST_HEAD(&work_list);
1149 index = (unsigned long)array->paa_deadline % array->paa_size;
1150 count = array->paa_count;
1152 count -= array->paa_reqs_count[index];
1153 cfs_list_for_each_entry_safe(rq, n,
1154 &array->paa_reqs_array[index],
1156 if (rq->rq_deadline <= now + at_early_margin) {
1157 cfs_list_del_init(&rq->rq_timed_list);
1159 * ptlrpc_server_drop_request() may drop
1160 * refcount to 0 already. Let's check this and
1161 * don't add entry to work_list
1163 if (likely(cfs_atomic_inc_not_zero(&rq->rq_refcount)))
1164 cfs_list_add(&rq->rq_timed_list, &work_list);
1166 array->paa_reqs_count[index]--;
1168 cfs_spin_lock(&rq->rq_lock);
1169 rq->rq_at_linked = 0;
1170 cfs_spin_unlock(&rq->rq_lock);
1174 /* update the earliest deadline */
1175 if (deadline == -1 || rq->rq_deadline < deadline)
1176 deadline = rq->rq_deadline;
1181 if (++index >= array->paa_size)
1184 array->paa_deadline = deadline;
1185 cfs_spin_unlock(&svc->srv_at_lock);
1187 /* we have a new earliest deadline, restart the timer */
1188 ptlrpc_at_set_timer(svc);
1190 CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
1191 "replies\n", first, at_extra, counter);
1193 /* We're already past request deadlines before we even get a
1194 chance to send early replies */
1195 LCONSOLE_WARN("%s: This server is not able to keep up with "
1196 "request traffic (cpu-bound).\n", svc->srv_name);
1197 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, "
1198 "delay="CFS_DURATION_T"(jiff)\n",
1199 counter, svc->srv_n_queued_reqs, svc->srv_n_active_reqs,
1200 at_get(&svc->srv_at_estimate), delay);
1203 /* we took additional refcount so entries can't be deleted from list, no
1204 * locking is needed */
1205 while (!cfs_list_empty(&work_list)) {
1206 rq = cfs_list_entry(work_list.next, struct ptlrpc_request,
1208 cfs_list_del_init(&rq->rq_timed_list);
1210 if (ptlrpc_at_send_early_reply(rq) == 0)
1211 ptlrpc_at_add_timed(rq);
1213 ptlrpc_server_drop_request(rq);
1220 * Put the request to the export list if the request may become
1221 * a high priority one.
1223 static int ptlrpc_hpreq_init(struct ptlrpc_service *svc,
1224 struct ptlrpc_request *req)
1229 if (svc->srv_hpreq_handler) {
1230 rc = svc->srv_hpreq_handler(req);
1234 if (req->rq_export && req->rq_ops) {
1235 /* Perform request specific check. We should do this check
1236 * before the request is added into exp_hp_rpcs list otherwise
1237 * it may hit swab race at LU-1044. */
1238 if (req->rq_ops->hpreq_check)
1239 rc = req->rq_ops->hpreq_check(req);
1241 cfs_spin_lock_bh(&req->rq_export->exp_rpc_lock);
1242 cfs_list_add(&req->rq_exp_list,
1243 &req->rq_export->exp_hp_rpcs);
1244 cfs_spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1250 /** Remove the request from the export list. */
1251 static void ptlrpc_hpreq_fini(struct ptlrpc_request *req)
1254 if (req->rq_export && req->rq_ops) {
1255 /* refresh lock timeout again so that client has more
1256 * room to send lock cancel RPC. */
1257 if (req->rq_ops->hpreq_fini)
1258 req->rq_ops->hpreq_fini(req);
1260 cfs_spin_lock_bh(&req->rq_export->exp_rpc_lock);
1261 cfs_list_del_init(&req->rq_exp_list);
1262 cfs_spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1268 * Make the request a high priority one.
1270 * All the high priority requests are queued in a separate FIFO
1271 * ptlrpc_service::srv_request_hpq list which is parallel to
1272 * ptlrpc_service::srv_request_queue list but has a higher priority
1275 * \see ptlrpc_server_handle_request().
1277 static void ptlrpc_hpreq_reorder_nolock(struct ptlrpc_service *svc,
1278 struct ptlrpc_request *req)
1281 LASSERT(svc != NULL);
1282 cfs_spin_lock(&req->rq_lock);
1283 if (req->rq_hp == 0) {
1284 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1286 /* Add to the high priority queue. */
1287 cfs_list_move_tail(&req->rq_list, &svc->srv_request_hpq);
1289 if (opc != OBD_PING)
1290 DEBUG_REQ(D_RPCTRACE, req, "high priority req");
1292 cfs_spin_unlock(&req->rq_lock);
1297 * \see ptlrpc_hpreq_reorder_nolock
1299 void ptlrpc_hpreq_reorder(struct ptlrpc_request *req)
1301 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
1304 cfs_spin_lock(&svc->srv_rq_lock);
1305 /* It may happen that the request is already taken for the processing
1306 * but still in the export list, do not re-add it into the HP list. */
1307 if (req->rq_phase == RQ_PHASE_NEW)
1308 ptlrpc_hpreq_reorder_nolock(svc, req);
1309 cfs_spin_unlock(&svc->srv_rq_lock);
1313 /** Check if the request is a high priority one. */
1314 static int ptlrpc_server_hpreq_check(struct ptlrpc_service *svc,
1315 struct ptlrpc_request *req)
1319 /* Check by request opc. */
1320 if (OBD_PING == lustre_msg_get_opc(req->rq_reqmsg))
1323 RETURN(ptlrpc_hpreq_init(svc, req));
1326 /** Check if a request is a high priority one. */
1327 static int ptlrpc_server_request_add(struct ptlrpc_service *svc,
1328 struct ptlrpc_request *req)
1333 rc = ptlrpc_server_hpreq_check(svc, req);
1337 cfs_spin_lock(&svc->srv_rq_lock);
1338 /* Before inserting the request into the queue, check if it is not
1339 * inserted yet, or even already handled -- it may happen due to
1340 * a racing ldlm_server_blocking_ast(). */
1341 if (req->rq_phase == RQ_PHASE_NEW && cfs_list_empty(&req->rq_list)) {
1343 ptlrpc_hpreq_reorder_nolock(svc, req);
1345 cfs_list_add_tail(&req->rq_list,
1346 &svc->srv_request_queue);
1348 cfs_spin_unlock(&svc->srv_rq_lock);
1354 * Allow to handle high priority request
1355 * User can call it w/o any lock but need to hold ptlrpc_service::srv_rq_lock
1356 * to get reliable result
1358 static int ptlrpc_server_allow_high(struct ptlrpc_service *svc, int force)
1363 if (svc->srv_n_active_reqs >= svc->srv_threads_running - 1)
1366 return cfs_list_empty(&svc->srv_request_queue) ||
1367 svc->srv_hpreq_count < svc->srv_hpreq_ratio;
1370 static int ptlrpc_server_high_pending(struct ptlrpc_service *svc, int force)
1372 return ptlrpc_server_allow_high(svc, force) &&
1373 !cfs_list_empty(&svc->srv_request_hpq);
1377 * Only allow normal priority requests on a service that has a high-priority
1378 * queue if forced (i.e. cleanup), if there are other high priority requests
1379 * already being processed (i.e. those threads can service more high-priority
1380 * requests), or if there are enough idle threads that a later thread can do
1381 * a high priority request.
1382 * User can call it w/o any lock but need to hold ptlrpc_service::srv_rq_lock
1383 * to get reliable result
1385 static int ptlrpc_server_allow_normal(struct ptlrpc_service *svc, int force)
1388 if (1) /* always allow to handle normal request for liblustre */
1392 svc->srv_n_active_reqs < svc->srv_threads_running - 2)
1395 if (svc->srv_n_active_reqs >= svc->srv_threads_running - 1)
1398 return svc->srv_n_active_hpreq > 0 || svc->srv_hpreq_handler == NULL;
1401 static int ptlrpc_server_normal_pending(struct ptlrpc_service *svc, int force)
1403 return ptlrpc_server_allow_normal(svc, force) &&
1404 !cfs_list_empty(&svc->srv_request_queue);
1408 * Returns true if there are requests available in incoming
1409 * request queue for processing and it is allowed to fetch them.
1410 * User can call it w/o any lock but need to hold ptlrpc_service::srv_rq_lock
1411 * to get reliable result
1412 * \see ptlrpc_server_allow_normal
1413 * \see ptlrpc_server_allow high
1416 ptlrpc_server_request_pending(struct ptlrpc_service *svc, int force)
1418 return ptlrpc_server_high_pending(svc, force) ||
1419 ptlrpc_server_normal_pending(svc, force);
1423 * Fetch a request for processing from queue of unprocessed requests.
1424 * Favors high-priority requests.
1425 * Returns a pointer to fetched request.
1427 static struct ptlrpc_request *
1428 ptlrpc_server_request_get(struct ptlrpc_service *svc, int force)
1430 struct ptlrpc_request *req;
1433 if (ptlrpc_server_high_pending(svc, force)) {
1434 req = cfs_list_entry(svc->srv_request_hpq.next,
1435 struct ptlrpc_request, rq_list);
1436 svc->srv_hpreq_count++;
1441 if (ptlrpc_server_normal_pending(svc, force)) {
1442 req = cfs_list_entry(svc->srv_request_queue.next,
1443 struct ptlrpc_request, rq_list);
1444 svc->srv_hpreq_count = 0;
1451 * Handle freshly incoming reqs, add to timed early reply list,
1452 * pass on to regular request queue.
1453 * All incoming requests pass through here before getting into
1454 * ptlrpc_server_handle_req later on.
1457 ptlrpc_server_handle_req_in(struct ptlrpc_service *svc)
1459 struct ptlrpc_request *req;
1466 cfs_spin_lock(&svc->srv_lock);
1467 if (cfs_list_empty(&svc->srv_req_in_queue)) {
1468 cfs_spin_unlock(&svc->srv_lock);
1472 req = cfs_list_entry(svc->srv_req_in_queue.next,
1473 struct ptlrpc_request, rq_list);
1474 cfs_list_del_init (&req->rq_list);
1475 svc->srv_n_queued_reqs--;
1476 /* Consider this still a "queued" request as far as stats are
1478 /* ptlrpc_hpreq_init() inserts it to the export list and by the time
1479 * of ptlrpc_server_request_add() it could be already handled and
1480 * released. To not lose request in between, take an extra reference
1481 * on the request. */
1482 ptlrpc_request_addref(req);
1483 cfs_spin_unlock(&svc->srv_lock);
1485 /* go through security check/transform */
1486 rc = sptlrpc_svc_unwrap_request(req);
1490 case SECSVC_COMPLETE:
1491 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1500 * for null-flavored rpc, msg has been unpacked by sptlrpc, although
1501 * redo it wouldn't be harmful.
1503 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
1504 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
1506 CERROR("error unpacking request: ptl %d from %s "
1507 "x"LPU64"\n", svc->srv_req_portal,
1508 libcfs_id2str(req->rq_peer), req->rq_xid);
1513 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1515 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
1516 LPU64"\n", svc->srv_req_portal,
1517 libcfs_id2str(req->rq_peer), req->rq_xid);
1521 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
1522 lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
1523 CERROR("drop incoming rpc opc %u, x"LPU64"\n",
1524 cfs_fail_val, req->rq_xid);
1529 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1530 CERROR("wrong packet type received (type=%u) from %s\n",
1531 lustre_msg_get_type(req->rq_reqmsg),
1532 libcfs_id2str(req->rq_peer));
1536 switch(lustre_msg_get_opc(req->rq_reqmsg)) {
1539 req->rq_bulk_write = 1;
1543 req->rq_bulk_read = 1;
1547 CDEBUG(D_RPCTRACE, "got req x"LPU64"\n", req->rq_xid);
1549 req->rq_export = class_conn2export(
1550 lustre_msg_get_handle(req->rq_reqmsg));
1551 if (req->rq_export) {
1552 rc = ptlrpc_check_req(req);
1554 rc = sptlrpc_target_export_check(req->rq_export, req);
1556 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
1557 "illegal security flavor,");
1562 ptlrpc_update_export_timer(req->rq_export, 0);
1565 /* req_in handling should/must be fast */
1566 if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5)
1567 DEBUG_REQ(D_WARNING, req, "Slow req_in handling "CFS_DURATION_T"s",
1568 cfs_time_sub(cfs_time_current_sec(),
1569 req->rq_arrival_time.tv_sec));
1571 /* Set rpc server deadline and add it to the timed list */
1572 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1573 MSGHDR_AT_SUPPORT) ?
1574 /* The max time the client expects us to take */
1575 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1576 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1577 if (unlikely(deadline == 0)) {
1578 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1582 ptlrpc_at_add_timed(req);
1584 /* Move it over to the request processing queue */
1585 rc = ptlrpc_server_request_add(svc, req);
1588 cfs_waitq_signal(&svc->srv_waitq);
1589 ptlrpc_server_drop_request(req);
1593 ptlrpc_server_drop_request(req);
1594 cfs_spin_lock(&svc->srv_rq_lock);
1595 svc->srv_n_active_reqs++;
1596 cfs_spin_unlock(&svc->srv_rq_lock);
1597 ptlrpc_server_finish_request(svc, req);
1603 * Main incoming request handling logic.
1604 * Calls handler function from service to do actual processing.
1607 ptlrpc_server_handle_request(struct ptlrpc_service *svc,
1608 struct ptlrpc_thread *thread)
1610 struct obd_export *export = NULL;
1611 struct ptlrpc_request *request;
1612 struct timeval work_start;
1613 struct timeval work_end;
1621 cfs_spin_lock(&svc->srv_rq_lock);
1623 /* !@%$# liblustre only has 1 thread */
1624 if (cfs_atomic_read(&svc->srv_n_difficult_replies) != 0) {
1625 cfs_spin_unlock(&svc->srv_rq_lock);
1629 request = ptlrpc_server_request_get(svc, 0);
1630 if (request == NULL) {
1631 cfs_spin_unlock(&svc->srv_rq_lock);
1635 opc = lustre_msg_get_opc(request->rq_reqmsg);
1636 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
1637 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
1638 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
1639 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
1641 if (unlikely(fail_opc)) {
1642 if (request->rq_export && request->rq_ops) {
1643 cfs_spin_unlock(&svc->srv_rq_lock);
1644 OBD_FAIL_TIMEOUT(fail_opc, 4);
1645 cfs_spin_lock(&svc->srv_rq_lock);
1646 request = ptlrpc_server_request_get(svc, 0);
1647 if (request == NULL) {
1648 cfs_spin_unlock(&svc->srv_rq_lock);
1654 cfs_list_del_init(&request->rq_list);
1655 svc->srv_n_active_reqs++;
1657 svc->srv_n_active_hpreq++;
1659 /* The phase is changed under the lock here because we need to know
1660 * the request is under processing (see ptlrpc_hpreq_reorder()). */
1661 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
1662 cfs_spin_unlock(&svc->srv_rq_lock);
1664 if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
1665 libcfs_debug_dumplog();
1667 cfs_gettimeofday(&work_start);
1668 timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
1669 if (likely(svc->srv_stats != NULL)) {
1670 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
1672 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
1673 svc->srv_n_queued_reqs);
1674 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
1675 svc->srv_n_active_reqs);
1676 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
1677 at_get(&svc->srv_at_estimate));
1680 rc = lu_context_init(&request->rq_session,
1681 LCT_SESSION|LCT_REMEMBER|LCT_NOREF);
1683 CERROR("Failure to initialize session: %d\n", rc);
1686 request->rq_session.lc_thread = thread;
1687 request->rq_session.lc_cookie = 0x5;
1688 lu_context_enter(&request->rq_session);
1690 CDEBUG(D_NET, "got req "LPU64"\n", request->rq_xid);
1692 request->rq_svc_thread = thread;
1694 request->rq_svc_thread->t_env->le_ses = &request->rq_session;
1696 if (likely(request->rq_export)) {
1697 if (unlikely(ptlrpc_check_req(request)))
1699 ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
1700 export = class_export_rpc_get(request->rq_export);
1703 /* Discard requests queued for longer than the deadline.
1704 The deadline is increased if we send an early reply. */
1705 if (cfs_time_current_sec() > request->rq_deadline) {
1706 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s"
1707 ": deadline "CFS_DURATION_T":"CFS_DURATION_T"s ago\n",
1708 libcfs_id2str(request->rq_peer),
1709 cfs_time_sub(request->rq_deadline,
1710 request->rq_arrival_time.tv_sec),
1711 cfs_time_sub(cfs_time_current_sec(),
1712 request->rq_deadline));
1713 goto put_rpc_export;
1716 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
1717 "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1718 (request->rq_export ?
1719 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1720 (request->rq_export ?
1721 cfs_atomic_read(&request->rq_export->exp_refcount) : -99),
1722 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1723 libcfs_id2str(request->rq_peer),
1724 lustre_msg_get_opc(request->rq_reqmsg));
1726 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
1727 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
1729 rc = svc->srv_handler(request);
1731 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
1735 class_export_rpc_put(export);
1737 lu_context_exit(&request->rq_session);
1738 lu_context_fini(&request->rq_session);
1740 if (unlikely(cfs_time_current_sec() > request->rq_deadline)) {
1741 DEBUG_REQ(D_WARNING, request, "Request x"LPU64" took longer "
1742 "than estimated ("CFS_DURATION_T":"CFS_DURATION_T"s);"
1743 " client may timeout.",
1744 request->rq_xid, cfs_time_sub(request->rq_deadline,
1745 request->rq_arrival_time.tv_sec),
1746 cfs_time_sub(cfs_time_current_sec(),
1747 request->rq_deadline));
1750 cfs_gettimeofday(&work_end);
1751 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1752 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
1753 "%s:%s+%d:%d:x"LPU64":%s:%d Request procesed in "
1754 "%ldus (%ldus total) trans "LPU64" rc %d/%d\n",
1756 (request->rq_export ?
1757 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1758 (request->rq_export ?
1759 cfs_atomic_read(&request->rq_export->exp_refcount) : -99),
1760 lustre_msg_get_status(request->rq_reqmsg),
1762 libcfs_id2str(request->rq_peer),
1763 lustre_msg_get_opc(request->rq_reqmsg),
1765 cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL),
1766 (request->rq_repmsg ?
1767 lustre_msg_get_transno(request->rq_repmsg) :
1768 request->rq_transno),
1770 (request->rq_repmsg ?
1771 lustre_msg_get_status(request->rq_repmsg) : -999));
1772 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
1773 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
1774 int opc = opcode_offset(op);
1775 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
1776 LASSERT(opc < LUSTRE_MAX_OPCODES);
1777 lprocfs_counter_add(svc->srv_stats,
1778 opc + EXTRA_MAX_OPCODES,
1782 if (unlikely(request->rq_early_count)) {
1783 DEBUG_REQ(D_ADAPTTO, request,
1784 "sent %d early replies before finishing in "
1786 request->rq_early_count,
1787 cfs_time_sub(work_end.tv_sec,
1788 request->rq_arrival_time.tv_sec));
1792 ptlrpc_server_finish_request(svc, request);
1798 * An internal function to process a single reply state object.
1801 ptlrpc_handle_rs (struct ptlrpc_reply_state *rs)
1803 struct ptlrpc_service *svc = rs->rs_service;
1804 struct obd_export *exp;
1805 struct obd_device *obd;
1810 exp = rs->rs_export;
1813 LASSERT (rs->rs_difficult);
1814 LASSERT (rs->rs_scheduled);
1815 LASSERT (cfs_list_empty(&rs->rs_list));
1817 cfs_spin_lock (&exp->exp_lock);
1818 /* Noop if removed already */
1819 cfs_list_del_init (&rs->rs_exp_list);
1820 cfs_spin_unlock (&exp->exp_lock);
1822 /* The disk commit callback holds exp_uncommitted_replies_lock while it
1823 * iterates over newly committed replies, removing them from
1824 * exp_uncommitted_replies. It then drops this lock and schedules the
1825 * replies it found for handling here.
1827 * We can avoid contention for exp_uncommitted_replies_lock between the
1828 * HRT threads and further commit callbacks by checking rs_committed
1829 * which is set in the commit callback while it holds both
1830 * rs_lock and exp_uncommitted_reples.
1832 * If we see rs_committed clear, the commit callback _may_ not have
1833 * handled this reply yet and we race with it to grab
1834 * exp_uncommitted_replies_lock before removing the reply from
1835 * exp_uncommitted_replies. Note that if we lose the race and the
1836 * reply has already been removed, list_del_init() is a noop.
1838 * If we see rs_committed set, we know the commit callback is handling,
1839 * or has handled this reply since store reordering might allow us to
1840 * see rs_committed set out of sequence. But since this is done
1841 * holding rs_lock, we can be sure it has all completed once we hold
1842 * rs_lock, which we do right next.
1844 if (!rs->rs_committed) {
1845 cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
1846 cfs_list_del_init(&rs->rs_obd_list);
1847 cfs_spin_unlock(&exp->exp_uncommitted_replies_lock);
1850 cfs_spin_lock(&rs->rs_lock);
1852 been_handled = rs->rs_handled;
1855 nlocks = rs->rs_nlocks; /* atomic "steal", but */
1856 rs->rs_nlocks = 0; /* locks still on rs_locks! */
1858 if (nlocks == 0 && !been_handled) {
1859 /* If we see this, we should already have seen the warning
1860 * in mds_steal_ack_locks() */
1861 CWARN("All locks stolen from rs %p x"LPD64".t"LPD64
1864 rs->rs_xid, rs->rs_transno, rs->rs_opc,
1865 libcfs_nid2str(exp->exp_connection->c_peer.nid));
1868 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
1869 cfs_spin_unlock(&rs->rs_lock);
1871 if (!been_handled && rs->rs_on_net) {
1872 LNetMDUnlink(rs->rs_md_h);
1873 /* Ignore return code; we're racing with
1877 while (nlocks-- > 0)
1878 ldlm_lock_decref(&rs->rs_locks[nlocks],
1879 rs->rs_modes[nlocks]);
1881 cfs_spin_lock(&rs->rs_lock);
1884 rs->rs_scheduled = 0;
1886 if (!rs->rs_on_net) {
1888 cfs_spin_unlock(&rs->rs_lock);
1890 class_export_put (exp);
1891 rs->rs_export = NULL;
1892 ptlrpc_rs_decref (rs);
1893 if (cfs_atomic_dec_and_test(&svc->srv_n_difficult_replies) &&
1894 svc->srv_is_stopping)
1895 cfs_waitq_broadcast(&svc->srv_waitq);
1899 /* still on the net; callback will schedule */
1900 cfs_spin_unlock(&rs->rs_lock);
1907 * Check whether given service has a reply available for processing
1910 * \param svc a ptlrpc service
1911 * \retval 0 no replies processed
1912 * \retval 1 one reply processed
1915 ptlrpc_server_handle_reply(struct ptlrpc_service *svc)
1917 struct ptlrpc_reply_state *rs = NULL;
1920 cfs_spin_lock(&svc->srv_rs_lock);
1921 if (!cfs_list_empty(&svc->srv_reply_queue)) {
1922 rs = cfs_list_entry(svc->srv_reply_queue.prev,
1923 struct ptlrpc_reply_state,
1925 cfs_list_del_init(&rs->rs_list);
1927 cfs_spin_unlock(&svc->srv_rs_lock);
1929 ptlrpc_handle_rs(rs);
1933 /* FIXME make use of timeout later */
1935 liblustre_check_services (void *arg)
1937 int did_something = 0;
1939 cfs_list_t *tmp, *nxt;
1942 /* I'm relying on being single threaded, not to have to lock
1943 * ptlrpc_all_services etc */
1944 cfs_list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
1945 struct ptlrpc_service *svc =
1946 cfs_list_entry (tmp, struct ptlrpc_service, srv_list);
1948 if (svc->srv_threads_running != 0) /* I've recursed */
1951 /* service threads can block for bulk, so this limits us
1952 * (arbitrarily) to recursing 1 stack frame per service.
1953 * Note that the problem with recursion is that we have to
1954 * unwind completely before our caller can resume. */
1956 svc->srv_threads_running++;
1959 rc = ptlrpc_server_handle_req_in(svc);
1960 rc |= ptlrpc_server_handle_reply(svc);
1961 rc |= ptlrpc_at_check_timed(svc);
1962 rc |= ptlrpc_server_handle_request(svc, NULL);
1963 rc |= (ptlrpc_server_post_idle_rqbds(svc) > 0);
1964 did_something |= rc;
1967 svc->srv_threads_running--;
1970 RETURN(did_something);
1972 #define ptlrpc_stop_all_threads(s) do {} while (0)
1974 #else /* __KERNEL__ */
1977 ptlrpc_check_rqbd_pool(struct ptlrpc_service *svc)
1979 int avail = svc->srv_nrqbd_receiving;
1980 int low_water = test_req_buffer_pressure ? 0 :
1981 svc->srv_nbuf_per_group / 2;
1983 /* NB I'm not locking; just looking. */
1985 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
1986 * allowed the request history to grow out of control. We could put a
1987 * sanity check on that here and cull some history if we need the
1990 if (avail <= low_water)
1991 ptlrpc_grow_req_bufs(svc);
1994 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQBUF_AVAIL_CNTR,
1999 ptlrpc_retry_rqbds(void *arg)
2001 struct ptlrpc_service *svc = (struct ptlrpc_service *)arg;
2003 svc->srv_rqbd_timeout = 0;
2004 return (-ETIMEDOUT);
2008 ptlrpc_threads_enough(struct ptlrpc_service *svc)
2010 return svc->srv_n_active_reqs <
2011 svc->srv_threads_running - 1 - (svc->srv_hpreq_handler != NULL);
2015 * allowed to create more threads
2016 * user can call it w/o any lock but need to hold ptlrpc_service::srv_lock to
2017 * get reliable result
2020 ptlrpc_threads_increasable(struct ptlrpc_service *svc)
2022 return svc->srv_threads_running +
2023 svc->srv_threads_starting < svc->srv_threads_max;
2027 * too many requests and allowed to create more threads
2030 ptlrpc_threads_need_create(struct ptlrpc_service *svc)
2032 return !ptlrpc_threads_enough(svc) && ptlrpc_threads_increasable(svc);
2036 ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2038 return (thread->t_flags & SVC_STOPPING) != 0 ||
2039 thread->t_svc->srv_is_stopping;
2043 ptlrpc_rqbd_pending(struct ptlrpc_service *svc)
2045 return !cfs_list_empty(&svc->srv_idle_rqbds) &&
2046 svc->srv_rqbd_timeout == 0;
2050 ptlrpc_at_check(struct ptlrpc_service *svc)
2052 return svc->srv_at_check;
2056 * requests wait on preprocessing
2057 * user can call it w/o any lock but need to hold ptlrpc_service::srv_lock to
2058 * get reliable result
2061 ptlrpc_server_request_waiting(struct ptlrpc_service *svc)
2063 return !cfs_list_empty(&svc->srv_req_in_queue);
2066 static __attribute__((__noinline__)) int
2067 ptlrpc_wait_event(struct ptlrpc_service *svc,
2068 struct ptlrpc_thread *thread)
2070 /* Don't exit while there are replies to be handled */
2071 struct l_wait_info lwi = LWI_TIMEOUT(svc->srv_rqbd_timeout,
2072 ptlrpc_retry_rqbds, svc);
2074 lc_watchdog_disable(thread->t_watchdog);
2078 l_wait_event_exclusive_head(svc->srv_waitq,
2079 ptlrpc_thread_stopping(thread) ||
2080 ptlrpc_server_request_waiting(svc) ||
2081 ptlrpc_server_request_pending(svc, 0) ||
2082 ptlrpc_rqbd_pending(svc) ||
2083 ptlrpc_at_check(svc), &lwi);
2085 if (ptlrpc_thread_stopping(thread))
2088 lc_watchdog_touch(thread->t_watchdog, CFS_GET_TIMEOUT(svc));
2094 * Main thread body for service threads.
2095 * Waits in a loop waiting for new requests to process to appear.
2096 * Every time an incoming requests is added to its queue, a waitq
2097 * is woken up and one of the threads will handle it.
2099 static int ptlrpc_main(void *arg)
2101 struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
2102 struct ptlrpc_service *svc = data->svc;
2103 struct ptlrpc_thread *thread = data->thread;
2104 struct ptlrpc_reply_state *rs;
2105 #ifdef WITH_GROUP_INFO
2106 cfs_group_info_t *ginfo = NULL;
2109 int counter = 0, rc = 0;
2112 thread->t_pid = cfs_curproc_pid();
2113 cfs_daemonize_ctxt(data->name);
2115 #if defined(HAVE_NODE_TO_CPUMASK) && defined(CONFIG_NUMA)
2116 /* we need to do this before any per-thread allocation is done so that
2117 * we get the per-thread allocations on local node. bug 7342 */
2118 if (svc->srv_cpu_affinity) {
2121 for (cpu = 0, num_cpu = 0; cpu < cfs_num_possible_cpus();
2123 if (!cfs_cpu_online(cpu))
2125 if (num_cpu == thread->t_id % cfs_num_online_cpus())
2129 cfs_set_cpus_allowed(cfs_current(),
2130 node_to_cpumask(cpu_to_node(cpu)));
2134 #ifdef WITH_GROUP_INFO
2135 ginfo = cfs_groups_alloc(0);
2141 cfs_set_current_groups(ginfo);
2142 cfs_put_group_info(ginfo);
2145 if (svc->srv_init != NULL) {
2146 rc = svc->srv_init(thread);
2151 rc = lu_context_init(&env.le_ctx,
2152 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2156 thread->t_env = &env;
2157 env.le_ctx.lc_thread = thread;
2158 env.le_ctx.lc_cookie = 0x6;
2160 /* Alloc reply state structure for this one */
2161 OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2167 cfs_spin_lock(&svc->srv_lock);
2169 LASSERT((thread->t_flags & SVC_STARTING) != 0);
2170 thread->t_flags &= ~SVC_STARTING;
2171 svc->srv_threads_starting--;
2173 /* SVC_STOPPING may already be set here if someone else is trying
2174 * to stop the service while this new thread has been dynamically
2175 * forked. We still set SVC_RUNNING to let our creator know that
2176 * we are now running, however we will exit as soon as possible */
2177 thread->t_flags |= SVC_RUNNING;
2178 svc->srv_threads_running++;
2179 cfs_spin_unlock(&svc->srv_lock);
2182 * wake up our creator. Note: @data is invalid after this point,
2183 * because it's allocated on ptlrpc_start_thread() stack.
2185 cfs_waitq_signal(&thread->t_ctl_waitq);
2187 thread->t_watchdog = lc_watchdog_add(CFS_GET_TIMEOUT(svc), NULL, NULL);
2189 cfs_spin_lock(&svc->srv_rs_lock);
2190 cfs_list_add(&rs->rs_list, &svc->srv_free_rs_list);
2191 cfs_waitq_signal(&svc->srv_free_rs_waitq);
2192 cfs_spin_unlock(&svc->srv_rs_lock);
2194 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2195 svc->srv_threads_running);
2197 /* XXX maintain a list of all managed devices: insert here */
2198 while (!ptlrpc_thread_stopping(thread)) {
2199 if (ptlrpc_wait_event(svc, thread))
2202 ptlrpc_check_rqbd_pool(svc);
2204 if (ptlrpc_threads_need_create(svc)) {
2205 /* Ignore return code - we tried... */
2206 ptlrpc_start_thread(svc);
2209 /* Process all incoming reqs before handling any */
2210 if (ptlrpc_server_request_waiting(svc)) {
2211 ptlrpc_server_handle_req_in(svc);
2212 /* but limit ourselves in case of flood */
2213 if (counter++ < 100)
2218 if (ptlrpc_at_check(svc))
2219 ptlrpc_at_check_timed(svc);
2221 if (ptlrpc_server_request_pending(svc, 0)) {
2222 lu_context_enter(&env.le_ctx);
2223 ptlrpc_server_handle_request(svc, thread);
2224 lu_context_exit(&env.le_ctx);
2227 if (ptlrpc_rqbd_pending(svc) &&
2228 ptlrpc_server_post_idle_rqbds(svc) < 0) {
2229 /* I just failed to repost request buffers.
2230 * Wait for a timeout (unless something else
2231 * happens) before I try again */
2232 svc->srv_rqbd_timeout = cfs_time_seconds(1)/10;
2233 CDEBUG(D_RPCTRACE,"Posted buffers: %d\n",
2234 svc->srv_nrqbd_receiving);
2238 lc_watchdog_delete(thread->t_watchdog);
2239 thread->t_watchdog = NULL;
2243 * deconstruct service specific state created by ptlrpc_start_thread()
2245 if (svc->srv_done != NULL)
2246 svc->srv_done(thread);
2248 lu_context_fini(&env.le_ctx);
2250 CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n",
2251 thread, thread->t_pid, thread->t_id, rc);
2253 cfs_spin_lock(&svc->srv_lock);
2254 if ((thread->t_flags & SVC_STARTING) != 0) {
2255 svc->srv_threads_starting--;
2256 thread->t_flags &= ~SVC_STARTING;
2259 if ((thread->t_flags & SVC_RUNNING) != 0) {
2260 /* must know immediately */
2261 svc->srv_threads_running--;
2262 thread->t_flags &= ~SVC_RUNNING;
2266 thread->t_flags |= SVC_STOPPED;
2268 cfs_waitq_signal(&thread->t_ctl_waitq);
2269 cfs_spin_unlock(&svc->srv_lock);
2274 struct ptlrpc_hr_args {
2277 struct ptlrpc_hr_service *hrs;
2280 static int hrt_dont_sleep(struct ptlrpc_hr_thread *t,
2281 cfs_list_t *replies)
2285 cfs_spin_lock(&t->hrt_lock);
2286 cfs_list_splice_init(&t->hrt_queue, replies);
2287 result = cfs_test_bit(HRT_STOPPING, &t->hrt_flags) ||
2288 !cfs_list_empty(replies);
2289 cfs_spin_unlock(&t->hrt_lock);
2294 * Main body of "handle reply" function.
2295 * It processes acked reply states
2297 static int ptlrpc_hr_main(void *arg)
2299 struct ptlrpc_hr_args * hr_args = arg;
2300 struct ptlrpc_hr_service *hr = hr_args->hrs;
2301 struct ptlrpc_hr_thread *t = &hr->hr_threads[hr_args->thread_index];
2302 char threadname[20];
2303 CFS_LIST_HEAD(replies);
2305 snprintf(threadname, sizeof(threadname),
2306 "ptlrpc_hr_%d", hr_args->thread_index);
2308 cfs_daemonize_ctxt(threadname);
2309 #if defined(CONFIG_NUMA) && defined(HAVE_NODE_TO_CPUMASK)
2310 cfs_set_cpus_allowed(cfs_current(),
2311 node_to_cpumask(cpu_to_node(hr_args->cpu_index)));
2313 cfs_set_bit(HRT_RUNNING, &t->hrt_flags);
2314 cfs_waitq_signal(&t->hrt_wait);
2316 while (!cfs_test_bit(HRT_STOPPING, &t->hrt_flags)) {
2318 l_wait_condition(t->hrt_wait, hrt_dont_sleep(t, &replies));
2319 while (!cfs_list_empty(&replies)) {
2320 struct ptlrpc_reply_state *rs;
2322 rs = cfs_list_entry(replies.prev,
2323 struct ptlrpc_reply_state,
2325 cfs_list_del_init(&rs->rs_list);
2326 ptlrpc_handle_rs(rs);
2330 cfs_clear_bit(HRT_RUNNING, &t->hrt_flags);
2331 cfs_complete(&t->hrt_completion);
2336 static int ptlrpc_start_hr_thread(struct ptlrpc_hr_service *hr, int n, int cpu)
2338 struct ptlrpc_hr_thread *t = &hr->hr_threads[n];
2339 struct ptlrpc_hr_args args;
2343 args.thread_index = n;
2344 args.cpu_index = cpu;
2347 rc = cfs_create_thread(ptlrpc_hr_main, (void*)&args, CFS_DAEMON_FLAGS);
2349 cfs_complete(&t->hrt_completion);
2352 l_wait_condition(t->hrt_wait, cfs_test_bit(HRT_RUNNING, &t->hrt_flags));
2358 static void ptlrpc_stop_hr_thread(struct ptlrpc_hr_thread *t)
2362 cfs_set_bit(HRT_STOPPING, &t->hrt_flags);
2363 cfs_waitq_signal(&t->hrt_wait);
2364 cfs_wait_for_completion(&t->hrt_completion);
2369 static void ptlrpc_stop_hr_threads(struct ptlrpc_hr_service *hrs)
2374 for (n = 0; n < hrs->hr_n_threads; n++)
2375 ptlrpc_stop_hr_thread(&hrs->hr_threads[n]);
2380 static int ptlrpc_start_hr_threads(struct ptlrpc_hr_service *hr)
2383 int n, cpu, threads_started = 0;
2386 LASSERT(hr != NULL);
2387 LASSERT(hr->hr_n_threads > 0);
2389 for (n = 0, cpu = 0; n < hr->hr_n_threads; n++) {
2390 #if defined(CONFIG_SMP) && defined(HAVE_NODE_TO_CPUMASK)
2391 while(!cfs_cpu_online(cpu)) {
2393 if (cpu >= cfs_num_possible_cpus())
2397 rc = ptlrpc_start_hr_thread(hr, n, cpu);
2403 if (threads_started == 0) {
2404 CERROR("No reply handling threads started\n");
2407 if (threads_started < hr->hr_n_threads) {
2408 CWARN("Started only %d reply handling threads from %d\n",
2409 threads_started, hr->hr_n_threads);
2410 hr->hr_n_threads = threads_started;
2415 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
2416 struct ptlrpc_thread *thread)
2418 struct l_wait_info lwi = { 0 };
2421 CDEBUG(D_RPCTRACE, "Stopping thread [ %p : %u ]\n",
2422 thread, thread->t_pid);
2424 cfs_spin_lock(&svc->srv_lock);
2425 /* let the thread know that we would like it to stop asap */
2426 thread->t_flags |= SVC_STOPPING;
2427 cfs_spin_unlock(&svc->srv_lock);
2429 cfs_waitq_broadcast(&svc->srv_waitq);
2430 l_wait_event(thread->t_ctl_waitq,
2431 (thread->t_flags & SVC_STOPPED), &lwi);
2433 cfs_spin_lock(&svc->srv_lock);
2434 cfs_list_del(&thread->t_link);
2435 cfs_spin_unlock(&svc->srv_lock);
2437 OBD_FREE_PTR(thread);
2442 * Stops all threads of a particular service \a svc
2444 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2446 struct ptlrpc_thread *thread;
2449 cfs_spin_lock(&svc->srv_lock);
2450 while (!cfs_list_empty(&svc->srv_threads)) {
2451 thread = cfs_list_entry(svc->srv_threads.next,
2452 struct ptlrpc_thread, t_link);
2454 cfs_spin_unlock(&svc->srv_lock);
2455 ptlrpc_stop_thread(svc, thread);
2456 cfs_spin_lock(&svc->srv_lock);
2459 cfs_spin_unlock(&svc->srv_lock);
2463 int ptlrpc_start_threads(struct ptlrpc_service *svc)
2468 /* We require 2 threads min - see note in
2469 ptlrpc_server_handle_request */
2470 LASSERT(svc->srv_threads_min >= 2);
2471 for (i = 0; i < svc->srv_threads_min; i++) {
2472 rc = ptlrpc_start_thread(svc);
2473 /* We have enough threads, don't start more. b=15759 */
2474 if (rc == -EMFILE) {
2479 CERROR("cannot start %s thread #%d: rc %d\n",
2480 svc->srv_thread_name, i, rc);
2481 ptlrpc_stop_all_threads(svc);
2488 int ptlrpc_start_thread(struct ptlrpc_service *svc)
2490 struct l_wait_info lwi = { 0 };
2491 struct ptlrpc_svc_data d;
2492 struct ptlrpc_thread *thread;
2497 CDEBUG(D_RPCTRACE, "%s started %d min %d max %d running %d\n",
2498 svc->srv_name, svc->srv_threads_running, svc->srv_threads_min,
2499 svc->srv_threads_max, svc->srv_threads_running);
2501 if (unlikely(svc->srv_is_stopping))
2504 if (!ptlrpc_threads_increasable(svc) ||
2505 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
2506 svc->srv_threads_running == svc->srv_threads_min - 1))
2509 OBD_ALLOC_PTR(thread);
2512 cfs_waitq_init(&thread->t_ctl_waitq);
2514 cfs_spin_lock(&svc->srv_lock);
2515 if (!ptlrpc_threads_increasable(svc)) {
2516 cfs_spin_unlock(&svc->srv_lock);
2517 OBD_FREE_PTR(thread);
2521 svc->srv_threads_starting++;
2522 thread->t_id = svc->srv_threads_next_id++;
2523 thread->t_flags |= SVC_STARTING;
2524 thread->t_svc = svc;
2526 cfs_list_add(&thread->t_link, &svc->srv_threads);
2527 cfs_spin_unlock(&svc->srv_lock);
2529 sprintf(name, "%s_%02d", svc->srv_thread_name, thread->t_id);
2534 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", name);
2536 /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
2537 * just drop the VM and FILES in cfs_daemonize_ctxt() right away.
2539 rc = cfs_create_thread(ptlrpc_main, &d, CFS_DAEMON_FLAGS);
2541 CERROR("cannot start thread '%s': rc %d\n", name, rc);
2543 cfs_spin_lock(&svc->srv_lock);
2544 cfs_list_del(&thread->t_link);
2545 --svc->srv_threads_starting;
2546 cfs_spin_unlock(&svc->srv_lock);
2548 OBD_FREE(thread, sizeof(*thread));
2551 l_wait_event(thread->t_ctl_waitq,
2552 thread->t_flags & (SVC_RUNNING | SVC_STOPPED), &lwi);
2554 rc = (thread->t_flags & SVC_STOPPED) ? thread->t_id : 0;
2559 int ptlrpc_hr_init(void)
2562 int n_cpus = cfs_num_online_cpus();
2563 struct ptlrpc_hr_service *hr;
2568 LASSERT(ptlrpc_hr == NULL);
2570 size = offsetof(struct ptlrpc_hr_service, hr_threads[n_cpus]);
2571 OBD_ALLOC(hr, size);
2574 for (i = 0; i < n_cpus; i++) {
2575 struct ptlrpc_hr_thread *t = &hr->hr_threads[i];
2577 cfs_spin_lock_init(&t->hrt_lock);
2578 cfs_waitq_init(&t->hrt_wait);
2579 CFS_INIT_LIST_HEAD(&t->hrt_queue);
2580 cfs_init_completion(&t->hrt_completion);
2582 hr->hr_n_threads = n_cpus;
2586 rc = ptlrpc_start_hr_threads(hr);
2588 OBD_FREE(hr, hr->hr_size);
2594 void ptlrpc_hr_fini(void)
2596 if (ptlrpc_hr != NULL) {
2597 ptlrpc_stop_hr_threads(ptlrpc_hr);
2598 OBD_FREE(ptlrpc_hr, ptlrpc_hr->hr_size);
2603 #endif /* __KERNEL__ */
2606 * Wait until all already scheduled replies are processed.
2608 static void ptlrpc_wait_replies(struct ptlrpc_service *svc)
2612 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
2614 rc = l_wait_event(svc->srv_waitq, cfs_atomic_read(&svc-> \
2615 srv_n_difficult_replies) == 0,
2619 CWARN("Unexpectedly long timeout %p\n", svc);
2623 int ptlrpc_unregister_service(struct ptlrpc_service *service)
2626 struct l_wait_info lwi;
2628 struct ptlrpc_reply_state *rs, *t;
2629 struct ptlrpc_at_array *array = &service->srv_at_array;
2632 service->srv_is_stopping = 1;
2633 cfs_timer_disarm(&service->srv_at_timer);
2635 ptlrpc_stop_all_threads(service);
2636 LASSERT(cfs_list_empty(&service->srv_threads));
2638 cfs_spin_lock (&ptlrpc_all_services_lock);
2639 cfs_list_del_init (&service->srv_list);
2640 cfs_spin_unlock (&ptlrpc_all_services_lock);
2642 ptlrpc_lprocfs_unregister_service(service);
2644 /* All history will be culled when the next request buffer is
2646 service->srv_max_history_rqbds = 0;
2648 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
2650 rc = LNetClearLazyPortal(service->srv_req_portal);
2653 /* Unlink all the request buffers. This forces a 'final' event with
2654 * its 'unlink' flag set for each posted rqbd */
2655 cfs_list_for_each(tmp, &service->srv_active_rqbds) {
2656 struct ptlrpc_request_buffer_desc *rqbd =
2657 cfs_list_entry(tmp, struct ptlrpc_request_buffer_desc,
2660 rc = LNetMDUnlink(rqbd->rqbd_md_h);
2661 LASSERT (rc == 0 || rc == -ENOENT);
2664 /* Wait for the network to release any buffers it's currently
2667 cfs_spin_lock(&service->srv_lock);
2668 rc = service->srv_nrqbd_receiving;
2669 cfs_spin_unlock(&service->srv_lock);
2674 /* Network access will complete in finite time but the HUGE
2675 * timeout lets us CWARN for visibility of sluggish NALs */
2676 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2677 cfs_time_seconds(1), NULL, NULL);
2678 rc = l_wait_event(service->srv_waitq,
2679 service->srv_nrqbd_receiving == 0,
2681 if (rc == -ETIMEDOUT)
2682 CWARN("Service %s waiting for request buffers\n",
2686 /* schedule all outstanding replies to terminate them */
2687 cfs_spin_lock(&service->srv_rs_lock);
2688 while (!cfs_list_empty(&service->srv_active_replies)) {
2689 struct ptlrpc_reply_state *rs =
2690 cfs_list_entry(service->srv_active_replies.next,
2691 struct ptlrpc_reply_state, rs_list);
2692 cfs_spin_lock(&rs->rs_lock);
2693 ptlrpc_schedule_difficult_reply(rs);
2694 cfs_spin_unlock(&rs->rs_lock);
2696 cfs_spin_unlock(&service->srv_rs_lock);
2698 /* purge the request queue. NB No new replies (rqbds all unlinked)
2699 * and no service threads, so I'm the only thread noodling the
2700 * request queue now */
2701 while (!cfs_list_empty(&service->srv_req_in_queue)) {
2702 struct ptlrpc_request *req =
2703 cfs_list_entry(service->srv_req_in_queue.next,
2704 struct ptlrpc_request,
2707 cfs_list_del(&req->rq_list);
2708 service->srv_n_queued_reqs--;
2709 service->srv_n_active_reqs++;
2710 ptlrpc_server_finish_request(service, req);
2712 while (ptlrpc_server_request_pending(service, 1)) {
2713 struct ptlrpc_request *req;
2715 req = ptlrpc_server_request_get(service, 1);
2716 cfs_list_del(&req->rq_list);
2717 service->srv_n_active_reqs++;
2718 ptlrpc_server_finish_request(service, req);
2720 LASSERT(service->srv_n_queued_reqs == 0);
2721 LASSERT(service->srv_n_active_reqs == 0);
2722 LASSERT(service->srv_n_history_rqbds == 0);
2723 LASSERT(cfs_list_empty(&service->srv_active_rqbds));
2725 /* Now free all the request buffers since nothing references them
2727 while (!cfs_list_empty(&service->srv_idle_rqbds)) {
2728 struct ptlrpc_request_buffer_desc *rqbd =
2729 cfs_list_entry(service->srv_idle_rqbds.next,
2730 struct ptlrpc_request_buffer_desc,
2733 ptlrpc_free_rqbd(rqbd);
2736 ptlrpc_wait_replies(service);
2738 cfs_list_for_each_entry_safe(rs, t, &service->srv_free_rs_list,
2740 cfs_list_del(&rs->rs_list);
2741 OBD_FREE_LARGE(rs, service->srv_max_reply_size);
2744 /* In case somebody rearmed this in the meantime */
2745 cfs_timer_disarm(&service->srv_at_timer);
2747 if (array->paa_reqs_array != NULL) {
2748 OBD_FREE(array->paa_reqs_array,
2749 sizeof(cfs_list_t) * array->paa_size);
2750 array->paa_reqs_array = NULL;
2753 if (array->paa_reqs_count != NULL) {
2754 OBD_FREE(array->paa_reqs_count,
2755 sizeof(__u32) * array->paa_size);
2756 array->paa_reqs_count= NULL;
2759 OBD_FREE_PTR(service);
2764 * Returns 0 if the service is healthy.
2766 * Right now, it just checks to make sure that requests aren't languishing
2767 * in the queue. We'll use this health check to govern whether a node needs
2768 * to be shot, so it's intentionally non-aggressive. */
2769 int ptlrpc_service_health_check(struct ptlrpc_service *svc)
2771 struct ptlrpc_request *request;
2772 struct timeval right_now;
2778 cfs_gettimeofday(&right_now);
2780 cfs_spin_lock(&svc->srv_rq_lock);
2781 if (!ptlrpc_server_request_pending(svc, 1)) {
2782 cfs_spin_unlock(&svc->srv_rq_lock);
2786 /* How long has the next entry been waiting? */
2787 if (cfs_list_empty(&svc->srv_request_queue))
2788 request = cfs_list_entry(svc->srv_request_hpq.next,
2789 struct ptlrpc_request, rq_list);
2791 request = cfs_list_entry(svc->srv_request_queue.next,
2792 struct ptlrpc_request, rq_list);
2793 timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
2794 cfs_spin_unlock(&svc->srv_rq_lock);
2796 if ((timediff / ONE_MILLION) > (AT_OFF ? obd_timeout * 3/2 :
2798 CERROR("%s: unhealthy - request has been waiting %lds\n",
2799 svc->srv_name, timediff / ONE_MILLION);