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 cfs_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 cfs_spin_lock(&svc->srv_lock);
115 cfs_list_add(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
117 cfs_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 (cfs_list_empty(&rqbd->rqbd_reqs));
130 cfs_spin_lock(&svc->srv_lock);
131 cfs_list_del(&rqbd->rqbd_list);
133 cfs_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 {
191 cfs_spinlock_t hrt_lock;
192 unsigned long hrt_flags;
193 cfs_waitq_t hrt_wait;
194 cfs_list_t hrt_queue;
195 cfs_completion_t hrt_completion;
198 struct ptlrpc_hr_service {
202 struct ptlrpc_hr_thread hr_threads[0];
206 cfs_list_t 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 * Choose an hr thread to dispatch requests to.
235 static unsigned int get_hr_thread_index(struct ptlrpc_hr_service *hr)
239 /* Concurrent modification of hr_index w/o any spinlock
240 protection is harmless as long as the result fits
241 [0..(hr_n_threads-1)] range and each thread gets near equal
244 hr->hr_index = (idx >= hr->hr_n_threads - 1) ? 0 : idx + 1;
249 * Dispatch all replies accumulated in the batch to one from
250 * dedicated reply handling threads.
254 static void rs_batch_dispatch(struct rs_batch *b)
256 if (b->rsb_n_replies != 0) {
257 struct ptlrpc_hr_service *hr = ptlrpc_hr;
260 idx = get_hr_thread_index(hr);
262 cfs_spin_lock(&hr->hr_threads[idx].hrt_lock);
263 cfs_list_splice_init(&b->rsb_replies,
264 &hr->hr_threads[idx].hrt_queue);
265 cfs_spin_unlock(&hr->hr_threads[idx].hrt_lock);
266 cfs_waitq_signal(&hr->hr_threads[idx].hrt_wait);
267 b->rsb_n_replies = 0;
272 * Add a reply to a batch.
273 * Add one reply object to a batch, schedule batched replies if overload.
278 static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs)
280 struct ptlrpc_service *svc = rs->rs_service;
282 if (svc != b->rsb_svc || b->rsb_n_replies >= MAX_SCHEDULED) {
283 if (b->rsb_svc != NULL) {
284 rs_batch_dispatch(b);
285 cfs_spin_unlock(&b->rsb_svc->srv_lock);
287 cfs_spin_lock(&svc->srv_lock);
290 cfs_spin_lock(&rs->rs_lock);
291 rs->rs_scheduled_ever = 1;
292 if (rs->rs_scheduled == 0) {
293 cfs_list_move(&rs->rs_list, &b->rsb_replies);
294 rs->rs_scheduled = 1;
297 rs->rs_committed = 1;
298 cfs_spin_unlock(&rs->rs_lock);
302 * Reply batch finalization.
303 * Dispatch remaining replies from the batch
304 * and release remaining spinlock.
308 static void rs_batch_fini(struct rs_batch *b)
310 if (b->rsb_svc != 0) {
311 rs_batch_dispatch(b);
312 cfs_spin_unlock(&b->rsb_svc->srv_lock);
316 #define DECLARE_RS_BATCH(b) struct rs_batch b
318 #else /* __KERNEL__ */
320 #define rs_batch_init(b) do{}while(0)
321 #define rs_batch_fini(b) do{}while(0)
322 #define rs_batch_add(b, r) ptlrpc_schedule_difficult_reply(r)
323 #define DECLARE_RS_BATCH(b)
325 #endif /* __KERNEL__ */
327 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
330 struct ptlrpc_hr_service *hr = ptlrpc_hr;
334 LASSERT(cfs_list_empty(&rs->rs_list));
336 idx = get_hr_thread_index(hr);
337 cfs_spin_lock(&hr->hr_threads[idx].hrt_lock);
338 cfs_list_add_tail(&rs->rs_list, &hr->hr_threads[idx].hrt_queue);
339 cfs_spin_unlock(&hr->hr_threads[idx].hrt_lock);
340 cfs_waitq_signal(&hr->hr_threads[idx].hrt_wait);
343 cfs_list_add_tail(&rs->rs_list, &rs->rs_service->srv_reply_queue);
348 ptlrpc_schedule_difficult_reply (struct ptlrpc_reply_state *rs)
352 LASSERT_SPIN_LOCKED(&rs->rs_service->srv_lock);
353 LASSERT_SPIN_LOCKED(&rs->rs_lock);
354 LASSERT (rs->rs_difficult);
355 rs->rs_scheduled_ever = 1; /* flag any notification attempt */
357 if (rs->rs_scheduled) { /* being set up or already notified */
362 rs->rs_scheduled = 1;
363 cfs_list_del_init(&rs->rs_list);
364 ptlrpc_dispatch_difficult_reply(rs);
368 void ptlrpc_commit_replies(struct obd_export *exp)
370 struct ptlrpc_reply_state *rs, *nxt;
371 DECLARE_RS_BATCH(batch);
374 rs_batch_init(&batch);
375 /* Find any replies that have been committed and get their service
376 * to attend to complete them. */
378 /* CAVEAT EMPTOR: spinlock ordering!!! */
379 cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
380 cfs_list_for_each_entry_safe(rs, nxt, &exp->exp_uncommitted_replies,
382 LASSERT (rs->rs_difficult);
383 /* VBR: per-export last_committed */
384 LASSERT(rs->rs_export);
385 if (rs->rs_transno <= exp->exp_last_committed) {
386 cfs_list_del_init(&rs->rs_obd_list);
387 rs_batch_add(&batch, rs);
390 cfs_spin_unlock(&exp->exp_uncommitted_replies_lock);
391 rs_batch_fini(&batch);
396 ptlrpc_server_post_idle_rqbds (struct ptlrpc_service *svc)
398 struct ptlrpc_request_buffer_desc *rqbd;
403 cfs_spin_lock(&svc->srv_lock);
405 if (cfs_list_empty (&svc->srv_idle_rqbds)) {
406 cfs_spin_unlock(&svc->srv_lock);
410 rqbd = cfs_list_entry(svc->srv_idle_rqbds.next,
411 struct ptlrpc_request_buffer_desc,
413 cfs_list_del (&rqbd->rqbd_list);
415 /* assume we will post successfully */
416 svc->srv_nrqbd_receiving++;
417 cfs_list_add (&rqbd->rqbd_list, &svc->srv_active_rqbds);
419 cfs_spin_unlock(&svc->srv_lock);
421 rc = ptlrpc_register_rqbd(rqbd);
428 cfs_spin_lock(&svc->srv_lock);
430 svc->srv_nrqbd_receiving--;
431 cfs_list_del(&rqbd->rqbd_list);
432 cfs_list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
434 /* Don't complain if no request buffers are posted right now; LNET
435 * won't drop requests because we set the portal lazy! */
437 cfs_spin_unlock(&svc->srv_lock);
442 struct ptlrpc_service *ptlrpc_init_svc_conf(struct ptlrpc_service_conf *c,
443 svc_handler_t h, char *name,
444 struct proc_dir_entry *proc_entry,
445 svcreq_printfn_t prntfn,
448 return ptlrpc_init_svc(c->psc_nbufs, c->psc_bufsize,
449 c->psc_max_req_size, c->psc_max_reply_size,
450 c->psc_req_portal, c->psc_rep_portal,
451 c->psc_watchdog_factor,
453 prntfn, c->psc_min_threads, c->psc_max_threads,
454 threadname, c->psc_ctx_tags, NULL);
456 EXPORT_SYMBOL(ptlrpc_init_svc_conf);
458 static void ptlrpc_at_timer(unsigned long castmeharder)
460 struct ptlrpc_service *svc = (struct ptlrpc_service *)castmeharder;
461 svc->srv_at_check = 1;
462 svc->srv_at_checktime = cfs_time_current();
463 cfs_waitq_signal(&svc->srv_waitq);
466 /* @threadname should be 11 characters or less - 3 will be added on */
467 struct ptlrpc_service *
468 ptlrpc_init_svc(int nbufs, int bufsize, int max_req_size, int max_reply_size,
469 int req_portal, int rep_portal, int watchdog_factor,
470 svc_handler_t handler, char *name,
471 cfs_proc_dir_entry_t *proc_entry,
472 svcreq_printfn_t svcreq_printfn,
473 int min_threads, int max_threads,
474 char *threadname, __u32 ctx_tags,
475 svc_hpreq_handler_t hp_handler)
478 struct ptlrpc_at_array *array;
479 struct ptlrpc_service *service;
480 unsigned int size, index;
484 LASSERT (bufsize >= max_req_size + SPTLRPC_MAX_PAYLOAD);
485 LASSERT (ctx_tags != 0);
487 OBD_ALLOC_PTR(service);
491 /* First initialise enough for early teardown */
493 service->srv_name = name;
494 cfs_spin_lock_init(&service->srv_lock);
495 CFS_INIT_LIST_HEAD(&service->srv_threads);
496 cfs_waitq_init(&service->srv_waitq);
498 service->srv_nbuf_per_group = test_req_buffer_pressure ? 1 : nbufs;
499 service->srv_max_req_size = max_req_size + SPTLRPC_MAX_PAYLOAD;
500 service->srv_buf_size = bufsize;
501 service->srv_rep_portal = rep_portal;
502 service->srv_req_portal = req_portal;
503 service->srv_watchdog_factor = watchdog_factor;
504 service->srv_handler = handler;
505 service->srv_request_history_print_fn = svcreq_printfn;
506 service->srv_request_seq = 1; /* valid seq #s start at 1 */
507 service->srv_request_max_cull_seq = 0;
508 service->srv_threads_min = min_threads;
509 service->srv_threads_max = max_threads;
510 service->srv_thread_name = threadname;
511 service->srv_ctx_tags = ctx_tags;
512 service->srv_hpreq_handler = hp_handler;
513 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
514 service->srv_hpreq_count = 0;
515 service->srv_n_hpreq = 0;
517 rc = LNetSetLazyPortal(service->srv_req_portal);
520 CFS_INIT_LIST_HEAD(&service->srv_request_queue);
521 CFS_INIT_LIST_HEAD(&service->srv_request_hpq);
522 CFS_INIT_LIST_HEAD(&service->srv_idle_rqbds);
523 CFS_INIT_LIST_HEAD(&service->srv_active_rqbds);
524 CFS_INIT_LIST_HEAD(&service->srv_history_rqbds);
525 CFS_INIT_LIST_HEAD(&service->srv_request_history);
526 CFS_INIT_LIST_HEAD(&service->srv_active_replies);
528 CFS_INIT_LIST_HEAD(&service->srv_reply_queue);
530 CFS_INIT_LIST_HEAD(&service->srv_free_rs_list);
531 cfs_waitq_init(&service->srv_free_rs_waitq);
532 cfs_atomic_set(&service->srv_n_difficult_replies, 0);
534 cfs_spin_lock_init(&service->srv_at_lock);
535 CFS_INIT_LIST_HEAD(&service->srv_req_in_queue);
537 array = &service->srv_at_array;
538 size = at_est2timeout(at_max);
539 array->paa_size = size;
540 array->paa_count = 0;
541 array->paa_deadline = -1;
543 /* allocate memory for srv_at_array (ptlrpc_at_array) */
544 OBD_ALLOC(array->paa_reqs_array, sizeof(cfs_list_t) * size);
545 if (array->paa_reqs_array == NULL)
548 for (index = 0; index < size; index++)
549 CFS_INIT_LIST_HEAD(&array->paa_reqs_array[index]);
551 OBD_ALLOC(array->paa_reqs_count, sizeof(__u32) * size);
552 if (array->paa_reqs_count == NULL)
555 cfs_timer_init(&service->srv_at_timer, ptlrpc_at_timer, service);
556 /* At SOW, service time should be quick; 10s seems generous. If client
557 timeout is less than this, we'll be sending an early reply. */
558 at_init(&service->srv_at_estimate, 10, 0);
560 cfs_spin_lock (&ptlrpc_all_services_lock);
561 cfs_list_add (&service->srv_list, &ptlrpc_all_services);
562 cfs_spin_unlock (&ptlrpc_all_services_lock);
564 /* Now allocate the request buffers */
565 rc = ptlrpc_grow_req_bufs(service);
566 /* We shouldn't be under memory pressure at startup, so
567 * fail if we can't post all our buffers at this time. */
571 /* Now allocate pool of reply buffers */
572 /* Increase max reply size to next power of two */
573 service->srv_max_reply_size = 1;
574 while (service->srv_max_reply_size <
575 max_reply_size + SPTLRPC_MAX_PAYLOAD)
576 service->srv_max_reply_size <<= 1;
578 if (proc_entry != NULL)
579 ptlrpc_lprocfs_register_service(proc_entry, service);
581 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
582 service->srv_name, service->srv_req_portal);
586 ptlrpc_unregister_service(service);
591 * to actually free the request, must be called without holding svc_lock.
592 * note it's caller's responsibility to unlink req->rq_list.
594 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
596 LASSERT(cfs_atomic_read(&req->rq_refcount) == 0);
597 LASSERT(cfs_list_empty(&req->rq_timed_list));
599 /* DEBUG_REQ() assumes the reply state of a request with a valid
600 * ref will not be destroyed until that reference is dropped. */
601 ptlrpc_req_drop_rs(req);
603 sptlrpc_svc_ctx_decref(req);
605 if (req != &req->rq_rqbd->rqbd_req) {
606 /* NB request buffers use an embedded
607 * req if the incoming req unlinked the
608 * MD; this isn't one of them! */
609 OBD_FREE(req, sizeof(*req));
614 * increment the number of active requests consuming service threads.
616 void ptlrpc_server_active_request_inc(struct ptlrpc_request *req)
618 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
619 struct ptlrpc_service *svc = rqbd->rqbd_service;
621 cfs_spin_lock(&svc->srv_lock);
622 svc->srv_n_active_reqs++;
623 cfs_spin_unlock(&svc->srv_lock);
627 * decrement the number of active requests consuming service threads.
629 void ptlrpc_server_active_request_dec(struct ptlrpc_request *req)
631 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
632 struct ptlrpc_service *svc = rqbd->rqbd_service;
634 cfs_spin_lock(&svc->srv_lock);
635 svc->srv_n_active_reqs--;
636 cfs_spin_unlock(&svc->srv_lock);
640 * drop a reference count of the request. if it reaches 0, we either
641 * put it into history list, or free it immediately.
643 void ptlrpc_server_drop_request(struct ptlrpc_request *req)
645 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
646 struct ptlrpc_service *svc = rqbd->rqbd_service;
651 if (!cfs_atomic_dec_and_test(&req->rq_refcount))
654 cfs_spin_lock(&svc->srv_at_lock);
655 if (req->rq_at_linked) {
656 struct ptlrpc_at_array *array = &svc->srv_at_array;
657 __u32 index = req->rq_at_index;
659 LASSERT(!cfs_list_empty(&req->rq_timed_list));
660 cfs_list_del_init(&req->rq_timed_list);
661 req->rq_at_linked = 0;
662 array->paa_reqs_count[index]--;
665 LASSERT(cfs_list_empty(&req->rq_timed_list));
666 cfs_spin_unlock(&svc->srv_at_lock);
668 /* finalize request */
669 if (req->rq_export) {
670 class_export_put(req->rq_export);
671 req->rq_export = NULL;
674 cfs_spin_lock(&svc->srv_lock);
676 svc->srv_n_active_reqs--;
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_request *req)
749 ptlrpc_server_drop_request(req);
752 /* This function makes sure dead exports are evicted in a timely manner.
753 This function is only called when some export receives a message (i.e.,
754 the network is up.) */
755 static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
757 struct obd_export *oldest_exp;
758 time_t oldest_time, new_time;
764 /* Compensate for slow machines, etc, by faking our request time
765 into the future. Although this can break the strict time-ordering
766 of the list, we can be really lazy here - we don't have to evict
767 at the exact right moment. Eventually, all silent exports
768 will make it to the top of the list. */
770 /* Do not pay attention on 1sec or smaller renewals. */
771 new_time = cfs_time_current_sec() + extra_delay;
772 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
775 exp->exp_last_request_time = new_time;
776 CDEBUG(D_HA, "updating export %s at "CFS_TIME_T" exp %p\n",
777 exp->exp_client_uuid.uuid,
778 exp->exp_last_request_time, exp);
780 /* exports may get disconnected from the chain even though the
781 export has references, so we must keep the spin lock while
782 manipulating the lists */
783 cfs_spin_lock(&exp->exp_obd->obd_dev_lock);
785 if (cfs_list_empty(&exp->exp_obd_chain_timed)) {
786 /* this one is not timed */
787 cfs_spin_unlock(&exp->exp_obd->obd_dev_lock);
791 cfs_list_move_tail(&exp->exp_obd_chain_timed,
792 &exp->exp_obd->obd_exports_timed);
794 oldest_exp = cfs_list_entry(exp->exp_obd->obd_exports_timed.next,
795 struct obd_export, exp_obd_chain_timed);
796 oldest_time = oldest_exp->exp_last_request_time;
797 cfs_spin_unlock(&exp->exp_obd->obd_dev_lock);
799 if (exp->exp_obd->obd_recovering) {
800 /* be nice to everyone during recovery */
805 /* Note - racing to start/reset the obd_eviction timer is safe */
806 if (exp->exp_obd->obd_eviction_timer == 0) {
807 /* Check if the oldest entry is expired. */
808 if (cfs_time_current_sec() > (oldest_time + PING_EVICT_TIMEOUT +
810 /* We need a second timer, in case the net was down and
811 * it just came back. Since the pinger may skip every
812 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
813 * we better wait for 3. */
814 exp->exp_obd->obd_eviction_timer =
815 cfs_time_current_sec() + 3 * PING_INTERVAL;
816 CDEBUG(D_HA, "%s: Think about evicting %s from "CFS_TIME_T"\n",
817 exp->exp_obd->obd_name,
818 obd_export_nid2str(oldest_exp), oldest_time);
821 if (cfs_time_current_sec() >
822 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
823 /* The evictor won't evict anyone who we've heard from
824 * recently, so we don't have to check before we start
826 if (!ping_evictor_wake(exp))
827 exp->exp_obd->obd_eviction_timer = 0;
834 static int ptlrpc_check_req(struct ptlrpc_request *req)
836 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
837 req->rq_export->exp_conn_cnt)) {
838 DEBUG_REQ(D_ERROR, req,
839 "DROPPING req from old connection %d < %d",
840 lustre_msg_get_conn_cnt(req->rq_reqmsg),
841 req->rq_export->exp_conn_cnt);
844 if (unlikely(req->rq_export->exp_obd &&
845 req->rq_export->exp_obd->obd_fail)) {
846 /* Failing over, don't handle any more reqs, send
847 error response instead. */
848 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
849 req, req->rq_export->exp_obd->obd_name);
850 req->rq_status = -ENODEV;
858 static void ptlrpc_at_set_timer(struct ptlrpc_service *svc)
860 struct ptlrpc_at_array *array = &svc->srv_at_array;
863 cfs_spin_lock(&svc->srv_at_lock);
864 if (array->paa_count == 0) {
865 cfs_timer_disarm(&svc->srv_at_timer);
866 cfs_spin_unlock(&svc->srv_at_lock);
870 /* Set timer for closest deadline */
871 next = (__s32)(array->paa_deadline - cfs_time_current_sec() -
874 ptlrpc_at_timer((unsigned long)svc);
876 cfs_timer_arm(&svc->srv_at_timer, cfs_time_shift(next));
877 cfs_spin_unlock(&svc->srv_at_lock);
878 CDEBUG(D_INFO, "armed %s at %+ds\n", svc->srv_name, next);
881 /* Add rpc to early reply check list */
882 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
884 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
885 struct ptlrpc_request *rq = NULL;
886 struct ptlrpc_at_array *array = &svc->srv_at_array;
893 if (req->rq_no_reply)
896 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
899 cfs_spin_lock(&svc->srv_at_lock);
900 LASSERT(cfs_list_empty(&req->rq_timed_list));
902 index = (unsigned long)req->rq_deadline % array->paa_size;
903 if (array->paa_reqs_count[index] > 0) {
904 /* latest rpcs will have the latest deadlines in the list,
905 * so search backward. */
906 cfs_list_for_each_entry_reverse(rq,
907 &array->paa_reqs_array[index],
909 if (req->rq_deadline >= rq->rq_deadline) {
910 cfs_list_add(&req->rq_timed_list,
917 /* Add the request at the head of the list */
918 if (cfs_list_empty(&req->rq_timed_list))
919 cfs_list_add(&req->rq_timed_list,
920 &array->paa_reqs_array[index]);
922 req->rq_at_linked = 1;
923 req->rq_at_index = index;
924 array->paa_reqs_count[index]++;
926 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
927 array->paa_deadline = req->rq_deadline;
930 cfs_spin_unlock(&svc->srv_at_lock);
933 ptlrpc_at_set_timer(svc);
938 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
940 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
941 struct ptlrpc_request *reqcopy;
942 struct lustre_msg *reqmsg;
943 cfs_duration_t olddl = req->rq_deadline - cfs_time_current_sec();
948 /* deadline is when the client expects us to reply, margin is the
949 difference between clients' and servers' expectations */
950 DEBUG_REQ(D_ADAPTTO, req,
951 "%ssending early reply (deadline %+lds, margin %+lds) for "
952 "%d+%d", AT_OFF ? "AT off - not " : "",
953 olddl, olddl - at_get(&svc->srv_at_estimate),
954 at_get(&svc->srv_at_estimate), at_extra);
960 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), "
961 "not sending early reply. Consider increasing "
962 "at_early_margin (%d)?", olddl, at_early_margin);
964 /* Return an error so we're not re-added to the timed list. */
968 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
969 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
970 "but no AT support");
974 if (req->rq_export &&
975 lustre_msg_get_flags(req->rq_reqmsg) &
976 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
977 /* During recovery, we don't want to send too many early
978 * replies, but on the other hand we want to make sure the
979 * client has enough time to resend if the rpc is lost. So
980 * during the recovery period send at least 4 early replies,
981 * spacing them every at_extra if we can. at_estimate should
982 * always equal this fixed value during recovery. */
983 at_measured(&svc->srv_at_estimate, min(at_extra,
984 req->rq_export->exp_obd->obd_recovery_timeout / 4));
986 /* Fake our processing time into the future to ask the clients
987 * for some extra amount of time */
988 at_measured(&svc->srv_at_estimate, at_extra +
989 cfs_time_current_sec() -
990 req->rq_arrival_time.tv_sec);
992 /* Check to see if we've actually increased the deadline -
993 * we may be past adaptive_max */
994 if (req->rq_deadline >= req->rq_arrival_time.tv_sec +
995 at_get(&svc->srv_at_estimate)) {
996 DEBUG_REQ(D_WARNING, req, "Couldn't add any time "
997 "(%ld/%ld), not sending early reply\n",
998 olddl, req->rq_arrival_time.tv_sec +
999 at_get(&svc->srv_at_estimate) -
1000 cfs_time_current_sec());
1004 newdl = cfs_time_current_sec() + at_get(&svc->srv_at_estimate);
1006 OBD_ALLOC(reqcopy, sizeof *reqcopy);
1007 if (reqcopy == NULL)
1009 OBD_ALLOC(reqmsg, req->rq_reqlen);
1011 OBD_FREE(reqcopy, sizeof *reqcopy);
1016 reqcopy->rq_reply_state = NULL;
1017 reqcopy->rq_rep_swab_mask = 0;
1018 reqcopy->rq_pack_bulk = 0;
1019 reqcopy->rq_pack_udesc = 0;
1020 reqcopy->rq_packed_final = 0;
1021 sptlrpc_svc_ctx_addref(reqcopy);
1022 /* We only need the reqmsg for the magic */
1023 reqcopy->rq_reqmsg = reqmsg;
1024 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1026 LASSERT(cfs_atomic_read(&req->rq_refcount));
1027 /** if it is last refcount then early reply isn't needed */
1028 if (cfs_atomic_read(&req->rq_refcount) == 1) {
1029 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
1030 "abort sending early reply\n");
1031 GOTO(out, rc = -EINVAL);
1034 /* Connection ref */
1035 reqcopy->rq_export = class_conn2export(
1036 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1037 if (reqcopy->rq_export == NULL)
1038 GOTO(out, rc = -ENODEV);
1041 class_export_rpc_get(reqcopy->rq_export);
1042 if (reqcopy->rq_export->exp_obd &&
1043 reqcopy->rq_export->exp_obd->obd_fail)
1044 GOTO(out_put, rc = -ENODEV);
1046 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1050 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1053 /* Adjust our own deadline to what we told the client */
1054 req->rq_deadline = newdl;
1055 req->rq_early_count++; /* number sent, server side */
1057 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1060 /* Free the (early) reply state from lustre_pack_reply.
1061 (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
1062 ptlrpc_req_drop_rs(reqcopy);
1065 class_export_rpc_put(reqcopy->rq_export);
1066 class_export_put(reqcopy->rq_export);
1068 sptlrpc_svc_ctx_decref(reqcopy);
1069 OBD_FREE(reqmsg, req->rq_reqlen);
1070 OBD_FREE(reqcopy, sizeof *reqcopy);
1074 /* Send early replies to everybody expiring within at_early_margin
1075 asking for at_extra time */
1076 static int ptlrpc_at_check_timed(struct ptlrpc_service *svc)
1078 struct ptlrpc_request *rq, *n;
1079 cfs_list_t work_list;
1080 struct ptlrpc_at_array *array = &svc->srv_at_array;
1083 time_t now = cfs_time_current_sec();
1084 cfs_duration_t delay;
1085 int first, counter = 0;
1088 cfs_spin_lock(&svc->srv_at_lock);
1089 if (svc->srv_at_check == 0) {
1090 cfs_spin_unlock(&svc->srv_at_lock);
1093 delay = cfs_time_sub(cfs_time_current(), svc->srv_at_checktime);
1094 svc->srv_at_check = 0;
1096 if (array->paa_count == 0) {
1097 cfs_spin_unlock(&svc->srv_at_lock);
1101 /* The timer went off, but maybe the nearest rpc already completed. */
1102 first = array->paa_deadline - now;
1103 if (first > at_early_margin) {
1104 /* We've still got plenty of time. Reset the timer. */
1105 cfs_spin_unlock(&svc->srv_at_lock);
1106 ptlrpc_at_set_timer(svc);
1110 /* We're close to a timeout, and we don't know how much longer the
1111 server will take. Send early replies to everyone expiring soon. */
1112 CFS_INIT_LIST_HEAD(&work_list);
1114 index = (unsigned long)array->paa_deadline % array->paa_size;
1115 count = array->paa_count;
1117 count -= array->paa_reqs_count[index];
1118 cfs_list_for_each_entry_safe(rq, n,
1119 &array->paa_reqs_array[index],
1121 if (rq->rq_deadline <= now + at_early_margin) {
1122 cfs_list_del_init(&rq->rq_timed_list);
1124 * ptlrpc_server_drop_request() may drop
1125 * refcount to 0 already. Let's check this and
1126 * don't add entry to work_list
1128 if (likely(cfs_atomic_inc_not_zero(&rq->rq_refcount)))
1129 cfs_list_add(&rq->rq_timed_list, &work_list);
1131 array->paa_reqs_count[index]--;
1133 rq->rq_at_linked = 0;
1137 /* update the earliest deadline */
1138 if (deadline == -1 || rq->rq_deadline < deadline)
1139 deadline = rq->rq_deadline;
1144 if (++index >= array->paa_size)
1147 array->paa_deadline = deadline;
1148 cfs_spin_unlock(&svc->srv_at_lock);
1150 /* we have a new earliest deadline, restart the timer */
1151 ptlrpc_at_set_timer(svc);
1153 CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
1154 "replies\n", first, at_extra, counter);
1156 /* We're already past request deadlines before we even get a
1157 chance to send early replies */
1158 LCONSOLE_WARN("%s: This server is not able to keep up with "
1159 "request traffic (cpu-bound).\n", svc->srv_name);
1160 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, "
1161 "delay="CFS_DURATION_T"(jiff)\n",
1162 counter, svc->srv_n_queued_reqs, svc->srv_n_active_reqs,
1163 at_get(&svc->srv_at_estimate), delay);
1166 /* we took additional refcount so entries can't be deleted from list, no
1167 * locking is needed */
1168 while (!cfs_list_empty(&work_list)) {
1169 rq = cfs_list_entry(work_list.next, struct ptlrpc_request,
1171 cfs_list_del_init(&rq->rq_timed_list);
1173 if (ptlrpc_at_send_early_reply(rq) == 0)
1174 ptlrpc_at_add_timed(rq);
1176 ptlrpc_server_drop_request(rq);
1183 * Put the request to the export list if the request may become
1184 * a high priority one.
1186 static int ptlrpc_hpreq_init(struct ptlrpc_service *svc,
1187 struct ptlrpc_request *req)
1192 if (svc->srv_hpreq_handler) {
1193 rc = svc->srv_hpreq_handler(req);
1197 if (req->rq_export && req->rq_ops) {
1198 cfs_spin_lock(&req->rq_export->exp_lock);
1199 cfs_list_add(&req->rq_exp_list,
1200 &req->rq_export->exp_queued_rpc);
1201 cfs_spin_unlock(&req->rq_export->exp_lock);
1207 /** Remove the request from the export list. */
1208 static void ptlrpc_hpreq_fini(struct ptlrpc_request *req)
1211 if (req->rq_export && req->rq_ops) {
1212 cfs_spin_lock(&req->rq_export->exp_lock);
1213 cfs_list_del_init(&req->rq_exp_list);
1214 cfs_spin_unlock(&req->rq_export->exp_lock);
1220 * Make the request a high priority one.
1222 * All the high priority requests are queued in a separate FIFO
1223 * ptlrpc_service::srv_request_hpq list which is parallel to
1224 * ptlrpc_service::srv_request_queue list but has a higher priority
1227 * \see ptlrpc_server_handle_request().
1229 static void ptlrpc_hpreq_reorder_nolock(struct ptlrpc_service *svc,
1230 struct ptlrpc_request *req)
1233 LASSERT(svc != NULL);
1234 cfs_spin_lock(&req->rq_lock);
1235 if (req->rq_hp == 0) {
1236 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1238 /* Add to the high priority queue. */
1239 cfs_list_move_tail(&req->rq_list, &svc->srv_request_hpq);
1241 if (opc != OBD_PING)
1242 DEBUG_REQ(D_NET, req, "high priority req");
1244 cfs_spin_unlock(&req->rq_lock);
1248 void ptlrpc_hpreq_reorder(struct ptlrpc_request *req)
1250 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
1253 cfs_spin_lock(&svc->srv_lock);
1254 /* It may happen that the request is already taken for the processing
1255 * but still in the export list, do not re-add it into the HP list. */
1256 if (req->rq_phase == RQ_PHASE_NEW)
1257 ptlrpc_hpreq_reorder_nolock(svc, req);
1258 cfs_spin_unlock(&svc->srv_lock);
1262 /** Check if the request is a high priority one. */
1263 static int ptlrpc_server_hpreq_check(struct ptlrpc_request *req)
1268 /* Check by request opc. */
1269 opc = lustre_msg_get_opc(req->rq_reqmsg);
1270 if (opc == OBD_PING)
1273 /* Perform request specific check. */
1274 if (req->rq_ops && req->rq_ops->hpreq_check)
1275 rc = req->rq_ops->hpreq_check(req);
1279 /** Check if a request is a high priority one. */
1280 static int ptlrpc_server_request_add(struct ptlrpc_service *svc,
1281 struct ptlrpc_request *req)
1286 rc = ptlrpc_server_hpreq_check(req);
1290 cfs_spin_lock(&svc->srv_lock);
1291 /* Before inserting the request into the queue, check if it is not
1292 * inserted yet, or even already handled -- it may happen due to
1293 * a racing ldlm_server_blocking_ast(). */
1294 if (req->rq_phase == RQ_PHASE_NEW && cfs_list_empty(&req->rq_list)) {
1296 ptlrpc_hpreq_reorder_nolock(svc, req);
1298 cfs_list_add_tail(&req->rq_list,
1299 &svc->srv_request_queue);
1301 cfs_spin_unlock(&svc->srv_lock);
1306 /* Only allow normal priority requests on a service that has a high-priority
1307 * queue if forced (i.e. cleanup), if there are other high priority requests
1308 * already being processed (i.e. those threads can service more high-priority
1309 * requests), or if there are enough idle threads that a later thread can do
1310 * a high priority request. */
1311 static int ptlrpc_server_allow_normal(struct ptlrpc_service *svc, int force)
1313 return force || !svc->srv_hpreq_handler || svc->srv_n_hpreq > 0 ||
1314 svc->srv_n_active_reqs < svc->srv_threads_running - 2;
1317 static struct ptlrpc_request *
1318 ptlrpc_server_request_get(struct ptlrpc_service *svc, int force)
1320 struct ptlrpc_request *req = NULL;
1323 if (ptlrpc_server_allow_normal(svc, force) &&
1324 !cfs_list_empty(&svc->srv_request_queue) &&
1325 (cfs_list_empty(&svc->srv_request_hpq) ||
1326 svc->srv_hpreq_count >= svc->srv_hpreq_ratio)) {
1327 req = cfs_list_entry(svc->srv_request_queue.next,
1328 struct ptlrpc_request, rq_list);
1329 svc->srv_hpreq_count = 0;
1330 } else if (!cfs_list_empty(&svc->srv_request_hpq)) {
1331 req = cfs_list_entry(svc->srv_request_hpq.next,
1332 struct ptlrpc_request, rq_list);
1333 svc->srv_hpreq_count++;
1338 static int ptlrpc_server_request_pending(struct ptlrpc_service *svc, int force)
1340 return ((ptlrpc_server_allow_normal(svc, force) &&
1341 !cfs_list_empty(&svc->srv_request_queue)) ||
1342 !cfs_list_empty(&svc->srv_request_hpq));
1345 /* Handle freshly incoming reqs, add to timed early reply list,
1346 pass on to regular request queue */
1348 ptlrpc_server_handle_req_in(struct ptlrpc_service *svc)
1350 struct ptlrpc_request *req;
1357 cfs_spin_lock(&svc->srv_lock);
1358 if (cfs_list_empty(&svc->srv_req_in_queue)) {
1359 cfs_spin_unlock(&svc->srv_lock);
1363 req = cfs_list_entry(svc->srv_req_in_queue.next,
1364 struct ptlrpc_request, rq_list);
1365 cfs_list_del_init (&req->rq_list);
1366 /* Consider this still a "queued" request as far as stats are
1368 cfs_spin_unlock(&svc->srv_lock);
1370 /* go through security check/transform */
1371 rc = sptlrpc_svc_unwrap_request(req);
1375 case SECSVC_COMPLETE:
1376 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1385 * for null-flavored rpc, msg has been unpacked by sptlrpc, although
1386 * redo it wouldn't be harmful.
1388 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
1389 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
1391 CERROR("error unpacking request: ptl %d from %s "
1392 "x"LPU64"\n", svc->srv_req_portal,
1393 libcfs_id2str(req->rq_peer), req->rq_xid);
1398 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1400 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
1401 LPU64"\n", svc->srv_req_portal,
1402 libcfs_id2str(req->rq_peer), req->rq_xid);
1406 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
1407 lustre_msg_get_opc(req->rq_reqmsg) == obd_fail_val) {
1408 CERROR("drop incoming rpc opc %u, x"LPU64"\n",
1409 obd_fail_val, req->rq_xid);
1414 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1415 CERROR("wrong packet type received (type=%u) from %s\n",
1416 lustre_msg_get_type(req->rq_reqmsg),
1417 libcfs_id2str(req->rq_peer));
1421 switch(lustre_msg_get_opc(req->rq_reqmsg)) {
1424 req->rq_bulk_write = 1;
1428 req->rq_bulk_read = 1;
1432 CDEBUG(D_NET, "got req "LPU64"\n", req->rq_xid);
1434 req->rq_export = class_conn2export(
1435 lustre_msg_get_handle(req->rq_reqmsg));
1436 if (req->rq_export) {
1437 rc = ptlrpc_check_req(req);
1439 rc = sptlrpc_target_export_check(req->rq_export, req);
1441 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
1442 "illegal security flavor,");
1447 ptlrpc_update_export_timer(req->rq_export, 0);
1450 /* req_in handling should/must be fast */
1451 if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5)
1452 DEBUG_REQ(D_WARNING, req, "Slow req_in handling "CFS_DURATION_T"s",
1453 cfs_time_sub(cfs_time_current_sec(),
1454 req->rq_arrival_time.tv_sec));
1456 /* Set rpc server deadline and add it to the timed list */
1457 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1458 MSGHDR_AT_SUPPORT) ?
1459 /* The max time the client expects us to take */
1460 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1461 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1462 if (unlikely(deadline == 0)) {
1463 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1467 ptlrpc_at_add_timed(req);
1468 rc = ptlrpc_hpreq_init(svc, req);
1472 /* Move it over to the request processing queue */
1473 rc = ptlrpc_server_request_add(svc, req);
1476 cfs_waitq_signal(&svc->srv_waitq);
1480 cfs_spin_lock(&svc->srv_lock);
1481 svc->srv_n_queued_reqs--;
1482 svc->srv_n_active_reqs++;
1483 cfs_spin_unlock(&svc->srv_lock);
1484 ptlrpc_server_finish_request(req);
1490 ptlrpc_server_handle_request(struct ptlrpc_service *svc,
1491 struct ptlrpc_thread *thread)
1493 struct obd_export *export = NULL;
1494 struct ptlrpc_request *request;
1495 struct timeval work_start;
1496 struct timeval work_end;
1504 cfs_spin_lock(&svc->srv_lock);
1505 if (unlikely(!ptlrpc_server_request_pending(svc, 0) ||
1508 /* !@%$# liblustre only has 1 thread */
1509 cfs_atomic_read(&svc->srv_n_difficult_replies) != 0 &&
1511 svc->srv_n_active_reqs >= (svc->srv_threads_running - 1)))) {
1512 /* Don't handle regular requests in the last thread, in order * re
1513 * to handle difficult replies (which might block other threads)
1514 * as well as handle any incoming reqs, early replies, etc.
1515 * That means we always need at least 2 service threads. */
1516 cfs_spin_unlock(&svc->srv_lock);
1520 request = ptlrpc_server_request_get(svc, 0);
1521 if (request == NULL) {
1522 cfs_spin_unlock(&svc->srv_lock);
1526 opc = lustre_msg_get_opc(request->rq_reqmsg);
1527 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
1528 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
1529 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
1530 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
1532 if (unlikely(fail_opc)) {
1533 if (request->rq_export && request->rq_ops) {
1534 cfs_spin_unlock(&svc->srv_lock);
1535 OBD_FAIL_TIMEOUT(fail_opc, 4);
1536 cfs_spin_lock(&svc->srv_lock);
1537 request = ptlrpc_server_request_get(svc, 0);
1538 if (request == NULL) {
1539 cfs_spin_unlock(&svc->srv_lock);
1542 LASSERT(ptlrpc_server_request_pending(svc, 0));
1546 cfs_list_del_init(&request->rq_list);
1547 svc->srv_n_queued_reqs--;
1548 svc->srv_n_active_reqs++;
1552 /* The phase is changed under the lock here because we need to know
1553 * the request is under processing (see ptlrpc_hpreq_reorder()). */
1554 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
1555 cfs_spin_unlock(&svc->srv_lock);
1557 ptlrpc_hpreq_fini(request);
1559 if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
1560 libcfs_debug_dumplog();
1562 cfs_gettimeofday(&work_start);
1563 timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
1564 if (likely(svc->srv_stats != NULL)) {
1565 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
1567 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
1568 svc->srv_n_queued_reqs);
1569 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
1570 svc->srv_n_active_reqs);
1571 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
1572 at_get(&svc->srv_at_estimate));
1575 rc = lu_context_init(&request->rq_session,
1576 LCT_SESSION|LCT_REMEMBER|LCT_NOREF);
1578 CERROR("Failure to initialize session: %d\n", rc);
1581 request->rq_session.lc_thread = thread;
1582 request->rq_session.lc_cookie = 0x5;
1583 lu_context_enter(&request->rq_session);
1585 CDEBUG(D_NET, "got req "LPU64"\n", request->rq_xid);
1587 request->rq_svc_thread = thread;
1589 request->rq_svc_thread->t_env->le_ses = &request->rq_session;
1591 if (likely(request->rq_export)) {
1592 if (unlikely(ptlrpc_check_req(request)))
1594 ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
1595 export = class_export_rpc_get(request->rq_export);
1598 /* Discard requests queued for longer than the deadline.
1599 The deadline is increased if we send an early reply. */
1600 if (cfs_time_current_sec() > request->rq_deadline) {
1601 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s"
1602 ": deadline "CFS_DURATION_T":"CFS_DURATION_T"s ago\n",
1603 libcfs_id2str(request->rq_peer),
1604 cfs_time_sub(request->rq_deadline,
1605 request->rq_arrival_time.tv_sec),
1606 cfs_time_sub(cfs_time_current_sec(),
1607 request->rq_deadline));
1608 goto put_rpc_export;
1611 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
1612 "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1613 (request->rq_export ?
1614 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1615 (request->rq_export ?
1616 cfs_atomic_read(&request->rq_export->exp_refcount) : -99),
1617 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1618 libcfs_id2str(request->rq_peer),
1619 lustre_msg_get_opc(request->rq_reqmsg));
1621 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
1622 OBD_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, obd_fail_val);
1624 rc = svc->srv_handler(request);
1626 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
1630 class_export_rpc_put(export);
1632 lu_context_exit(&request->rq_session);
1633 lu_context_fini(&request->rq_session);
1635 if (unlikely(cfs_time_current_sec() > request->rq_deadline)) {
1636 DEBUG_REQ(D_WARNING, request, "Request x"LPU64" took longer "
1637 "than estimated ("CFS_DURATION_T":"CFS_DURATION_T"s);"
1638 " client may timeout.",
1639 request->rq_xid, cfs_time_sub(request->rq_deadline,
1640 request->rq_arrival_time.tv_sec),
1641 cfs_time_sub(cfs_time_current_sec(),
1642 request->rq_deadline));
1645 cfs_gettimeofday(&work_end);
1646 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1647 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
1648 "%s:%s+%d:%d:x"LPU64":%s:%d Request procesed in "
1649 "%ldus (%ldus total) trans "LPU64" rc %d/%d\n",
1651 (request->rq_export ?
1652 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1653 (request->rq_export ?
1654 cfs_atomic_read(&request->rq_export->exp_refcount) : -99),
1655 lustre_msg_get_status(request->rq_reqmsg),
1657 libcfs_id2str(request->rq_peer),
1658 lustre_msg_get_opc(request->rq_reqmsg),
1660 cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL),
1661 (request->rq_repmsg ?
1662 lustre_msg_get_transno(request->rq_repmsg) :
1663 request->rq_transno),
1665 (request->rq_repmsg ?
1666 lustre_msg_get_status(request->rq_repmsg) : -999));
1667 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
1668 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
1669 int opc = opcode_offset(op);
1670 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
1671 LASSERT(opc < LUSTRE_MAX_OPCODES);
1672 lprocfs_counter_add(svc->srv_stats,
1673 opc + EXTRA_MAX_OPCODES,
1677 if (unlikely(request->rq_early_count)) {
1678 DEBUG_REQ(D_ADAPTTO, request,
1679 "sent %d early replies before finishing in "
1681 request->rq_early_count,
1682 cfs_time_sub(work_end.tv_sec,
1683 request->rq_arrival_time.tv_sec));
1687 cfs_spin_lock(&svc->srv_lock);
1690 cfs_spin_unlock(&svc->srv_lock);
1691 ptlrpc_server_finish_request(request);
1697 * An internal function to process a single reply state object.
1700 ptlrpc_handle_rs (struct ptlrpc_reply_state *rs)
1702 struct ptlrpc_service *svc = rs->rs_service;
1703 struct obd_export *exp;
1704 struct obd_device *obd;
1709 exp = rs->rs_export;
1712 LASSERT (rs->rs_difficult);
1713 LASSERT (rs->rs_scheduled);
1714 LASSERT (cfs_list_empty(&rs->rs_list));
1716 cfs_spin_lock (&exp->exp_lock);
1717 /* Noop if removed already */
1718 cfs_list_del_init (&rs->rs_exp_list);
1719 cfs_spin_unlock (&exp->exp_lock);
1721 /* The disk commit callback holds exp_uncommitted_replies_lock while it
1722 * iterates over newly committed replies, removing them from
1723 * exp_uncommitted_replies. It then drops this lock and schedules the
1724 * replies it found for handling here.
1726 * We can avoid contention for exp_uncommitted_replies_lock between the
1727 * HRT threads and further commit callbacks by checking rs_committed
1728 * which is set in the commit callback while it holds both
1729 * rs_lock and exp_uncommitted_reples.
1731 * If we see rs_committed clear, the commit callback _may_ not have
1732 * handled this reply yet and we race with it to grab
1733 * exp_uncommitted_replies_lock before removing the reply from
1734 * exp_uncommitted_replies. Note that if we lose the race and the
1735 * reply has already been removed, list_del_init() is a noop.
1737 * If we see rs_committed set, we know the commit callback is handling,
1738 * or has handled this reply since store reordering might allow us to
1739 * see rs_committed set out of sequence. But since this is done
1740 * holding rs_lock, we can be sure it has all completed once we hold
1741 * rs_lock, which we do right next.
1743 if (!rs->rs_committed) {
1744 cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
1745 cfs_list_del_init(&rs->rs_obd_list);
1746 cfs_spin_unlock(&exp->exp_uncommitted_replies_lock);
1749 cfs_spin_lock(&rs->rs_lock);
1751 been_handled = rs->rs_handled;
1754 nlocks = rs->rs_nlocks; /* atomic "steal", but */
1755 rs->rs_nlocks = 0; /* locks still on rs_locks! */
1757 if (nlocks == 0 && !been_handled) {
1758 /* If we see this, we should already have seen the warning
1759 * in mds_steal_ack_locks() */
1760 CWARN("All locks stolen from rs %p x"LPD64".t"LPD64
1763 rs->rs_xid, rs->rs_transno, rs->rs_opc,
1764 libcfs_nid2str(exp->exp_connection->c_peer.nid));
1767 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
1768 cfs_spin_unlock(&rs->rs_lock);
1770 if (!been_handled && rs->rs_on_net) {
1771 LNetMDUnlink(rs->rs_md_h);
1772 /* Ignore return code; we're racing with
1776 while (nlocks-- > 0)
1777 ldlm_lock_decref(&rs->rs_locks[nlocks],
1778 rs->rs_modes[nlocks]);
1780 cfs_spin_lock(&rs->rs_lock);
1783 rs->rs_scheduled = 0;
1785 if (!rs->rs_on_net) {
1787 cfs_spin_unlock(&rs->rs_lock);
1789 class_export_put (exp);
1790 rs->rs_export = NULL;
1791 ptlrpc_rs_decref (rs);
1792 cfs_atomic_dec (&svc->srv_outstanding_replies);
1793 if (cfs_atomic_dec_and_test(&svc->srv_n_difficult_replies) &&
1794 svc->srv_is_stopping)
1795 cfs_waitq_broadcast(&svc->srv_waitq);
1799 /* still on the net; callback will schedule */
1800 cfs_spin_unlock(&rs->rs_lock);
1807 * Check whether given service has a reply available for processing
1810 * \param svc a ptlrpc service
1811 * \retval 0 no replies processed
1812 * \retval 1 one reply processed
1815 ptlrpc_server_handle_reply(struct ptlrpc_service *svc)
1817 struct ptlrpc_reply_state *rs = NULL;
1820 cfs_spin_lock(&svc->srv_lock);
1821 if (!cfs_list_empty(&svc->srv_reply_queue)) {
1822 rs = cfs_list_entry(svc->srv_reply_queue.prev,
1823 struct ptlrpc_reply_state,
1825 cfs_list_del_init(&rs->rs_list);
1827 cfs_spin_unlock(&svc->srv_lock);
1829 ptlrpc_handle_rs(rs);
1833 /* FIXME make use of timeout later */
1835 liblustre_check_services (void *arg)
1837 int did_something = 0;
1839 cfs_list_t *tmp, *nxt;
1842 /* I'm relying on being single threaded, not to have to lock
1843 * ptlrpc_all_services etc */
1844 cfs_list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
1845 struct ptlrpc_service *svc =
1846 cfs_list_entry (tmp, struct ptlrpc_service, srv_list);
1848 if (svc->srv_threads_running != 0) /* I've recursed */
1851 /* service threads can block for bulk, so this limits us
1852 * (arbitrarily) to recursing 1 stack frame per service.
1853 * Note that the problem with recursion is that we have to
1854 * unwind completely before our caller can resume. */
1856 svc->srv_threads_running++;
1859 rc = ptlrpc_server_handle_req_in(svc);
1860 rc |= ptlrpc_server_handle_reply(svc);
1861 rc |= ptlrpc_at_check_timed(svc);
1862 rc |= ptlrpc_server_handle_request(svc, NULL);
1863 rc |= (ptlrpc_server_post_idle_rqbds(svc) > 0);
1864 did_something |= rc;
1867 svc->srv_threads_running--;
1870 RETURN(did_something);
1872 #define ptlrpc_stop_all_threads(s) do {} while (0)
1874 #else /* __KERNEL__ */
1877 ptlrpc_check_rqbd_pool(struct ptlrpc_service *svc)
1879 int avail = svc->srv_nrqbd_receiving;
1880 int low_water = test_req_buffer_pressure ? 0 :
1881 svc->srv_nbuf_per_group/2;
1883 /* NB I'm not locking; just looking. */
1885 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
1886 * allowed the request history to grow out of control. We could put a
1887 * sanity check on that here and cull some history if we need the
1890 if (avail <= low_water)
1891 ptlrpc_grow_req_bufs(svc);
1894 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQBUF_AVAIL_CNTR,
1899 ptlrpc_retry_rqbds(void *arg)
1901 struct ptlrpc_service *svc = (struct ptlrpc_service *)arg;
1903 svc->srv_rqbd_timeout = 0;
1904 return (-ETIMEDOUT);
1907 static int ptlrpc_main(void *arg)
1909 struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
1910 struct ptlrpc_service *svc = data->svc;
1911 struct ptlrpc_thread *thread = data->thread;
1912 struct obd_device *dev = data->dev;
1913 struct ptlrpc_reply_state *rs;
1914 #ifdef WITH_GROUP_INFO
1915 cfs_group_info_t *ginfo = NULL;
1918 int counter = 0, rc = 0;
1921 thread->t_pid = cfs_curproc_pid();
1922 cfs_daemonize_ctxt(data->name);
1924 #if defined(HAVE_NODE_TO_CPUMASK) && defined(CONFIG_NUMA)
1925 /* we need to do this before any per-thread allocation is done so that
1926 * we get the per-thread allocations on local node. bug 7342 */
1927 if (svc->srv_cpu_affinity) {
1930 for (cpu = 0, num_cpu = 0; cpu < cfs_num_possible_cpus();
1932 if (!cfs_cpu_online(cpu))
1934 if (num_cpu == thread->t_id % cfs_num_online_cpus())
1938 cfs_set_cpus_allowed(cfs_current(),
1939 node_to_cpumask(cpu_to_node(cpu)));
1943 #ifdef WITH_GROUP_INFO
1944 ginfo = cfs_groups_alloc(0);
1950 cfs_set_current_groups(ginfo);
1951 cfs_put_group_info(ginfo);
1954 if (svc->srv_init != NULL) {
1955 rc = svc->srv_init(thread);
1960 rc = lu_context_init(&env.le_ctx,
1961 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
1965 thread->t_env = &env;
1966 env.le_ctx.lc_thread = thread;
1967 env.le_ctx.lc_cookie = 0x6;
1969 /* Alloc reply state structure for this one */
1970 OBD_ALLOC_GFP(rs, svc->srv_max_reply_size, CFS_ALLOC_STD);
1976 cfs_spin_lock(&svc->srv_lock);
1977 /* SVC_STOPPING may already be set here if someone else is trying
1978 * to stop the service while this new thread has been dynamically
1979 * forked. We still set SVC_RUNNING to let our creator know that
1980 * we are now running, however we will exit as soon as possible */
1981 thread->t_flags |= SVC_RUNNING;
1982 cfs_spin_unlock(&svc->srv_lock);
1985 * wake up our creator. Note: @data is invalid after this point,
1986 * because it's allocated on ptlrpc_start_thread() stack.
1988 cfs_waitq_signal(&thread->t_ctl_waitq);
1990 thread->t_watchdog = lc_watchdog_add(CFS_GET_TIMEOUT(svc), NULL, NULL);
1992 cfs_spin_lock(&svc->srv_lock);
1993 svc->srv_threads_running++;
1994 cfs_list_add(&rs->rs_list, &svc->srv_free_rs_list);
1995 cfs_spin_unlock(&svc->srv_lock);
1996 cfs_waitq_signal(&svc->srv_free_rs_waitq);
1998 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
1999 svc->srv_threads_running);
2001 /* XXX maintain a list of all managed devices: insert here */
2003 while (!(thread->t_flags & SVC_STOPPING) && !svc->srv_is_stopping) {
2004 /* Don't exit while there are replies to be handled */
2005 struct l_wait_info lwi = LWI_TIMEOUT(svc->srv_rqbd_timeout,
2006 ptlrpc_retry_rqbds, svc);
2008 lc_watchdog_disable(thread->t_watchdog);
2012 l_wait_event_exclusive (svc->srv_waitq,
2013 thread->t_flags & SVC_STOPPING ||
2014 svc->srv_is_stopping ||
2015 (!cfs_list_empty(&svc->srv_idle_rqbds) &&
2016 svc->srv_rqbd_timeout == 0) ||
2017 !cfs_list_empty(&svc->srv_req_in_queue) ||
2018 (ptlrpc_server_request_pending(svc, 0) &&
2019 (svc->srv_n_active_reqs <
2020 (svc->srv_threads_running - 1))) ||
2024 if (thread->t_flags & SVC_STOPPING || svc->srv_is_stopping)
2027 lc_watchdog_touch(thread->t_watchdog, CFS_GET_TIMEOUT(svc));
2029 ptlrpc_check_rqbd_pool(svc);
2031 if (svc->srv_threads_started < svc->srv_threads_max &&
2032 svc->srv_n_active_reqs >= (svc->srv_threads_started - 1))
2033 /* Ignore return code - we tried... */
2034 ptlrpc_start_thread(dev, svc);
2036 if (!cfs_list_empty(&svc->srv_req_in_queue)) {
2037 /* Process all incoming reqs before handling any */
2038 ptlrpc_server_handle_req_in(svc);
2039 /* but limit ourselves in case of flood */
2040 if (counter++ < 1000)
2045 if (svc->srv_at_check)
2046 ptlrpc_at_check_timed(svc);
2048 /* don't handle requests in the last thread */
2049 if (ptlrpc_server_request_pending(svc, 0) &&
2050 (svc->srv_n_active_reqs < (svc->srv_threads_running - 1))) {
2051 lu_context_enter(&env.le_ctx);
2052 ptlrpc_server_handle_request(svc, thread);
2053 lu_context_exit(&env.le_ctx);
2056 if (!cfs_list_empty(&svc->srv_idle_rqbds) &&
2057 ptlrpc_server_post_idle_rqbds(svc) < 0) {
2058 /* I just failed to repost request buffers. Wait
2059 * for a timeout (unless something else happens)
2060 * before I try again */
2061 svc->srv_rqbd_timeout = cfs_time_seconds(1)/10;
2062 CDEBUG(D_RPCTRACE,"Posted buffers: %d\n",
2063 svc->srv_nrqbd_receiving);
2067 lc_watchdog_delete(thread->t_watchdog);
2068 thread->t_watchdog = NULL;
2072 * deconstruct service specific state created by ptlrpc_start_thread()
2074 if (svc->srv_done != NULL)
2075 svc->srv_done(thread);
2077 lu_context_fini(&env.le_ctx);
2079 CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n",
2080 thread, thread->t_pid, thread->t_id, rc);
2082 cfs_spin_lock(&svc->srv_lock);
2083 svc->srv_threads_running--; /* must know immediately */
2085 thread->t_flags = SVC_STOPPED;
2087 cfs_waitq_signal(&thread->t_ctl_waitq);
2088 cfs_spin_unlock(&svc->srv_lock);
2093 struct ptlrpc_hr_args {
2096 struct ptlrpc_hr_service *hrs;
2099 static int hrt_dont_sleep(struct ptlrpc_hr_thread *t,
2100 cfs_list_t *replies)
2104 cfs_spin_lock(&t->hrt_lock);
2105 cfs_list_splice_init(&t->hrt_queue, replies);
2106 result = cfs_test_bit(HRT_STOPPING, &t->hrt_flags) ||
2107 !cfs_list_empty(replies);
2108 cfs_spin_unlock(&t->hrt_lock);
2112 static int ptlrpc_hr_main(void *arg)
2114 struct ptlrpc_hr_args * hr_args = arg;
2115 struct ptlrpc_hr_service *hr = hr_args->hrs;
2116 struct ptlrpc_hr_thread *t = &hr->hr_threads[hr_args->thread_index];
2117 char threadname[20];
2118 CFS_LIST_HEAD(replies);
2120 snprintf(threadname, sizeof(threadname),
2121 "ptlrpc_hr_%d", hr_args->thread_index);
2123 cfs_daemonize_ctxt(threadname);
2124 #if defined(CONFIG_SMP) && defined(HAVE_NODE_TO_CPUMASK)
2125 cfs_set_cpus_allowed(cfs_current(),
2126 node_to_cpumask(cpu_to_node(hr_args->cpu_index)));
2128 cfs_set_bit(HRT_RUNNING, &t->hrt_flags);
2129 cfs_waitq_signal(&t->hrt_wait);
2131 while (!cfs_test_bit(HRT_STOPPING, &t->hrt_flags)) {
2133 l_cfs_wait_event(t->hrt_wait, hrt_dont_sleep(t, &replies));
2134 while (!cfs_list_empty(&replies)) {
2135 struct ptlrpc_reply_state *rs;
2137 rs = cfs_list_entry(replies.prev,
2138 struct ptlrpc_reply_state,
2140 cfs_list_del_init(&rs->rs_list);
2141 ptlrpc_handle_rs(rs);
2145 cfs_clear_bit(HRT_RUNNING, &t->hrt_flags);
2146 cfs_complete(&t->hrt_completion);
2151 static int ptlrpc_start_hr_thread(struct ptlrpc_hr_service *hr, int n, int cpu)
2153 struct ptlrpc_hr_thread *t = &hr->hr_threads[n];
2154 struct ptlrpc_hr_args args;
2158 args.thread_index = n;
2159 args.cpu_index = cpu;
2162 rc = cfs_kernel_thread(ptlrpc_hr_main, (void*)&args,
2163 CLONE_VM|CLONE_FILES);
2165 cfs_complete(&t->hrt_completion);
2168 l_cfs_wait_event(t->hrt_wait, cfs_test_bit(HRT_RUNNING, &t->hrt_flags));
2174 static void ptlrpc_stop_hr_thread(struct ptlrpc_hr_thread *t)
2178 cfs_set_bit(HRT_STOPPING, &t->hrt_flags);
2179 cfs_waitq_signal(&t->hrt_wait);
2180 cfs_wait_for_completion(&t->hrt_completion);
2185 static void ptlrpc_stop_hr_threads(struct ptlrpc_hr_service *hrs)
2190 for (n = 0; n < hrs->hr_n_threads; n++)
2191 ptlrpc_stop_hr_thread(&hrs->hr_threads[n]);
2196 static int ptlrpc_start_hr_threads(struct ptlrpc_hr_service *hr)
2199 int n, cpu, threads_started = 0;
2202 LASSERT(hr != NULL);
2203 LASSERT(hr->hr_n_threads > 0);
2205 for (n = 0, cpu = 0; n < hr->hr_n_threads; n++) {
2206 #if defined(CONFIG_SMP) && defined(HAVE_NODE_TO_CPUMASK)
2207 while(!cfs_cpu_online(cpu)) {
2209 if (cpu >= cfs_num_possible_cpus())
2213 rc = ptlrpc_start_hr_thread(hr, n, cpu);
2219 if (threads_started == 0) {
2220 CERROR("No reply handling threads started\n");
2223 if (threads_started < hr->hr_n_threads) {
2224 CWARN("Started only %d reply handling threads from %d\n",
2225 threads_started, hr->hr_n_threads);
2226 hr->hr_n_threads = threads_started;
2231 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
2232 struct ptlrpc_thread *thread)
2234 struct l_wait_info lwi = { 0 };
2237 CDEBUG(D_RPCTRACE, "Stopping thread [ %p : %u ]\n",
2238 thread, thread->t_pid);
2240 cfs_spin_lock(&svc->srv_lock);
2241 /* let the thread know that we would like it to stop asap */
2242 thread->t_flags |= SVC_STOPPING;
2243 cfs_spin_unlock(&svc->srv_lock);
2245 cfs_waitq_broadcast(&svc->srv_waitq);
2246 l_wait_event(thread->t_ctl_waitq,
2247 (thread->t_flags & SVC_STOPPED), &lwi);
2249 cfs_spin_lock(&svc->srv_lock);
2250 cfs_list_del(&thread->t_link);
2251 cfs_spin_unlock(&svc->srv_lock);
2253 OBD_FREE_PTR(thread);
2257 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2259 struct ptlrpc_thread *thread;
2262 cfs_spin_lock(&svc->srv_lock);
2263 while (!cfs_list_empty(&svc->srv_threads)) {
2264 thread = cfs_list_entry(svc->srv_threads.next,
2265 struct ptlrpc_thread, t_link);
2267 cfs_spin_unlock(&svc->srv_lock);
2268 ptlrpc_stop_thread(svc, thread);
2269 cfs_spin_lock(&svc->srv_lock);
2272 cfs_spin_unlock(&svc->srv_lock);
2276 int ptlrpc_start_threads(struct obd_device *dev, struct ptlrpc_service *svc)
2281 /* We require 2 threads min - see note in
2282 ptlrpc_server_handle_request */
2283 LASSERT(svc->srv_threads_min >= 2);
2284 for (i = 0; i < svc->srv_threads_min; i++) {
2285 rc = ptlrpc_start_thread(dev, svc);
2286 /* We have enough threads, don't start more. b=15759 */
2290 CERROR("cannot start %s thread #%d: rc %d\n",
2291 svc->srv_thread_name, i, rc);
2292 ptlrpc_stop_all_threads(svc);
2299 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc)
2301 struct l_wait_info lwi = { 0 };
2302 struct ptlrpc_svc_data d;
2303 struct ptlrpc_thread *thread;
2308 CDEBUG(D_RPCTRACE, "%s started %d min %d max %d running %d\n",
2309 svc->srv_name, svc->srv_threads_started, svc->srv_threads_min,
2310 svc->srv_threads_max, svc->srv_threads_running);
2312 if (unlikely(svc->srv_is_stopping))
2315 if (unlikely(svc->srv_threads_started >= svc->srv_threads_max) ||
2316 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
2317 svc->srv_threads_started == svc->srv_threads_min - 1))
2320 OBD_ALLOC_PTR(thread);
2323 cfs_waitq_init(&thread->t_ctl_waitq);
2325 cfs_spin_lock(&svc->srv_lock);
2326 if (svc->srv_threads_started >= svc->srv_threads_max) {
2327 cfs_spin_unlock(&svc->srv_lock);
2328 OBD_FREE_PTR(thread);
2331 cfs_list_add(&thread->t_link, &svc->srv_threads);
2332 id = svc->srv_threads_started++;
2333 cfs_spin_unlock(&svc->srv_lock);
2335 thread->t_svc = svc;
2337 sprintf(name, "%s_%02d", svc->srv_thread_name, id);
2343 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", name);
2345 /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
2346 * just drop the VM and FILES in cfs_daemonize_ctxt() right away.
2348 rc = cfs_kernel_thread(ptlrpc_main, &d, CLONE_VM | CLONE_FILES);
2350 CERROR("cannot start thread '%s': rc %d\n", name, rc);
2352 cfs_spin_lock(&svc->srv_lock);
2353 cfs_list_del(&thread->t_link);
2354 --svc->srv_threads_started;
2355 cfs_spin_unlock(&svc->srv_lock);
2357 OBD_FREE(thread, sizeof(*thread));
2360 l_wait_event(thread->t_ctl_waitq,
2361 thread->t_flags & (SVC_RUNNING | SVC_STOPPED), &lwi);
2363 rc = (thread->t_flags & SVC_STOPPED) ? thread->t_id : 0;
2368 int ptlrpc_hr_init(void)
2371 int n_cpus = cfs_num_online_cpus();
2372 struct ptlrpc_hr_service *hr;
2377 LASSERT(ptlrpc_hr == NULL);
2379 size = offsetof(struct ptlrpc_hr_service, hr_threads[n_cpus]);
2380 OBD_ALLOC(hr, size);
2383 for (i = 0; i < n_cpus; i++) {
2384 struct ptlrpc_hr_thread *t = &hr->hr_threads[i];
2386 cfs_spin_lock_init(&t->hrt_lock);
2387 cfs_waitq_init(&t->hrt_wait);
2388 CFS_INIT_LIST_HEAD(&t->hrt_queue);
2389 cfs_init_completion(&t->hrt_completion);
2391 hr->hr_n_threads = n_cpus;
2395 rc = ptlrpc_start_hr_threads(hr);
2397 OBD_FREE(hr, hr->hr_size);
2403 void ptlrpc_hr_fini(void)
2405 if (ptlrpc_hr != NULL) {
2406 ptlrpc_stop_hr_threads(ptlrpc_hr);
2407 OBD_FREE(ptlrpc_hr, ptlrpc_hr->hr_size);
2412 #endif /* __KERNEL__ */
2415 * Wait until all already scheduled replies are processed.
2417 static void ptlrpc_wait_replies(struct ptlrpc_service *svc)
2421 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
2423 rc = l_wait_event(svc->srv_waitq, cfs_atomic_read(&svc-> \
2424 srv_n_difficult_replies) == 0,
2428 CWARN("Unexpectedly long timeout %p\n", svc);
2432 int ptlrpc_unregister_service(struct ptlrpc_service *service)
2435 struct l_wait_info lwi;
2437 struct ptlrpc_reply_state *rs, *t;
2438 struct ptlrpc_at_array *array = &service->srv_at_array;
2441 service->srv_is_stopping = 1;
2442 cfs_timer_disarm(&service->srv_at_timer);
2444 ptlrpc_stop_all_threads(service);
2445 LASSERT(cfs_list_empty(&service->srv_threads));
2447 cfs_spin_lock (&ptlrpc_all_services_lock);
2448 cfs_list_del_init (&service->srv_list);
2449 cfs_spin_unlock (&ptlrpc_all_services_lock);
2451 ptlrpc_lprocfs_unregister_service(service);
2453 /* All history will be culled when the next request buffer is
2455 service->srv_max_history_rqbds = 0;
2457 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
2459 rc = LNetClearLazyPortal(service->srv_req_portal);
2462 /* Unlink all the request buffers. This forces a 'final' event with
2463 * its 'unlink' flag set for each posted rqbd */
2464 cfs_list_for_each(tmp, &service->srv_active_rqbds) {
2465 struct ptlrpc_request_buffer_desc *rqbd =
2466 cfs_list_entry(tmp, struct ptlrpc_request_buffer_desc,
2469 rc = LNetMDUnlink(rqbd->rqbd_md_h);
2470 LASSERT (rc == 0 || rc == -ENOENT);
2473 /* Wait for the network to release any buffers it's currently
2476 cfs_spin_lock(&service->srv_lock);
2477 rc = service->srv_nrqbd_receiving;
2478 cfs_spin_unlock(&service->srv_lock);
2483 /* Network access will complete in finite time but the HUGE
2484 * timeout lets us CWARN for visibility of sluggish NALs */
2485 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2486 cfs_time_seconds(1), NULL, NULL);
2487 rc = l_wait_event(service->srv_waitq,
2488 service->srv_nrqbd_receiving == 0,
2490 if (rc == -ETIMEDOUT)
2491 CWARN("Service %s waiting for request buffers\n",
2495 /* schedule all outstanding replies to terminate them */
2496 cfs_spin_lock(&service->srv_lock);
2497 while (!cfs_list_empty(&service->srv_active_replies)) {
2498 struct ptlrpc_reply_state *rs =
2499 cfs_list_entry(service->srv_active_replies.next,
2500 struct ptlrpc_reply_state, rs_list);
2501 cfs_spin_lock(&rs->rs_lock);
2502 ptlrpc_schedule_difficult_reply(rs);
2503 cfs_spin_unlock(&rs->rs_lock);
2505 cfs_spin_unlock(&service->srv_lock);
2507 /* purge the request queue. NB No new replies (rqbds all unlinked)
2508 * and no service threads, so I'm the only thread noodling the
2509 * request queue now */
2510 while (!cfs_list_empty(&service->srv_req_in_queue)) {
2511 struct ptlrpc_request *req =
2512 cfs_list_entry(service->srv_req_in_queue.next,
2513 struct ptlrpc_request,
2516 cfs_list_del(&req->rq_list);
2517 service->srv_n_queued_reqs--;
2518 service->srv_n_active_reqs++;
2519 ptlrpc_server_finish_request(req);
2521 while (ptlrpc_server_request_pending(service, 1)) {
2522 struct ptlrpc_request *req;
2524 req = ptlrpc_server_request_get(service, 1);
2525 cfs_list_del(&req->rq_list);
2526 service->srv_n_queued_reqs--;
2527 service->srv_n_active_reqs++;
2528 ptlrpc_hpreq_fini(req);
2529 ptlrpc_server_finish_request(req);
2531 LASSERT(service->srv_n_queued_reqs == 0);
2532 LASSERT(service->srv_n_active_reqs == 0);
2533 LASSERT(service->srv_n_history_rqbds == 0);
2534 LASSERT(cfs_list_empty(&service->srv_active_rqbds));
2536 /* Now free all the request buffers since nothing references them
2538 while (!cfs_list_empty(&service->srv_idle_rqbds)) {
2539 struct ptlrpc_request_buffer_desc *rqbd =
2540 cfs_list_entry(service->srv_idle_rqbds.next,
2541 struct ptlrpc_request_buffer_desc,
2544 ptlrpc_free_rqbd(rqbd);
2547 ptlrpc_wait_replies(service);
2549 cfs_list_for_each_entry_safe(rs, t, &service->srv_free_rs_list,
2551 cfs_list_del(&rs->rs_list);
2552 OBD_FREE(rs, service->srv_max_reply_size);
2555 /* In case somebody rearmed this in the meantime */
2556 cfs_timer_disarm(&service->srv_at_timer);
2558 if (array->paa_reqs_array != NULL) {
2559 OBD_FREE(array->paa_reqs_array,
2560 sizeof(cfs_list_t) * array->paa_size);
2561 array->paa_reqs_array = NULL;
2564 if (array->paa_reqs_count != NULL) {
2565 OBD_FREE(array->paa_reqs_count,
2566 sizeof(__u32) * array->paa_size);
2567 array->paa_reqs_count= NULL;
2570 OBD_FREE_PTR(service);
2574 /* Returns 0 if the service is healthy.
2576 * Right now, it just checks to make sure that requests aren't languishing
2577 * in the queue. We'll use this health check to govern whether a node needs
2578 * to be shot, so it's intentionally non-aggressive. */
2579 int ptlrpc_service_health_check(struct ptlrpc_service *svc)
2581 struct ptlrpc_request *request;
2582 struct timeval right_now;
2588 cfs_gettimeofday(&right_now);
2590 cfs_spin_lock(&svc->srv_lock);
2591 if (!ptlrpc_server_request_pending(svc, 1)) {
2592 cfs_spin_unlock(&svc->srv_lock);
2596 /* How long has the next entry been waiting? */
2597 if (cfs_list_empty(&svc->srv_request_queue))
2598 request = cfs_list_entry(svc->srv_request_hpq.next,
2599 struct ptlrpc_request, rq_list);
2601 request = cfs_list_entry(svc->srv_request_queue.next,
2602 struct ptlrpc_request, rq_list);
2603 timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
2604 cfs_spin_unlock(&svc->srv_lock);
2606 if ((timediff / ONE_MILLION) > (AT_OFF ? obd_timeout * 3/2 :
2608 CERROR("%s: unhealthy - request has been waiting %lds\n",
2609 svc->srv_name, timediff / ONE_MILLION);