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;
127 CDEBUG(D_RPCTRACE, "%s: allocate %d new %d-byte reqbufs (%d/%d left)\n",
128 svc->srv_name, svc->srv_nbuf_per_group, svc->srv_buf_size,
129 svc->srv_nrqbd_receiving, svc->srv_nbufs);
130 for (i = 0; i < svc->srv_nbuf_per_group; i++) {
131 rqbd = ptlrpc_alloc_rqbd(svc);
134 CERROR ("%s: Can't allocate request buffer\n",
139 if (ptlrpc_server_post_idle_rqbds(svc) < 0)
147 * Part of Rep-Ack logic.
148 * Puts a lock and its mode into reply state assotiated to request reply.
151 ptlrpc_save_lock(struct ptlrpc_request *req,
152 struct lustre_handle *lock, int mode, int no_ack)
154 struct ptlrpc_reply_state *rs = req->rq_reply_state;
158 LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
160 if (req->rq_export->exp_disconnected) {
161 ldlm_lock_decref(lock, mode);
163 idx = rs->rs_nlocks++;
164 rs->rs_locks[idx] = *lock;
165 rs->rs_modes[idx] = mode;
166 rs->rs_difficult = 1;
167 rs->rs_no_ack = !!no_ack;
173 #define HRT_RUNNING 0
174 #define HRT_STOPPING 1
176 struct ptlrpc_hr_thread {
177 cfs_spinlock_t hrt_lock;
178 unsigned long hrt_flags;
179 cfs_waitq_t hrt_wait;
180 cfs_list_t hrt_queue;
181 cfs_completion_t hrt_completion;
184 struct ptlrpc_hr_service {
188 struct ptlrpc_hr_thread hr_threads[0];
192 cfs_list_t rsb_replies;
193 struct ptlrpc_service *rsb_svc;
194 unsigned int rsb_n_replies;
198 * A pointer to per-node reply handling service.
200 static struct ptlrpc_hr_service *ptlrpc_hr = NULL;
203 * maximum mumber of replies scheduled in one batch
205 #define MAX_SCHEDULED 256
208 * Initialize a reply batch.
212 static void rs_batch_init(struct rs_batch *b)
214 memset(b, 0, sizeof *b);
215 CFS_INIT_LIST_HEAD(&b->rsb_replies);
219 * Choose an hr thread to dispatch requests to.
221 static unsigned int get_hr_thread_index(struct ptlrpc_hr_service *hr)
225 /* Concurrent modification of hr_index w/o any spinlock
226 protection is harmless as long as the result fits
227 [0..(hr_n_threads-1)] range and each thread gets near equal
230 hr->hr_index = (idx >= hr->hr_n_threads - 1) ? 0 : idx + 1;
235 * Dispatch all replies accumulated in the batch to one from
236 * dedicated reply handling threads.
240 static void rs_batch_dispatch(struct rs_batch *b)
242 if (b->rsb_n_replies != 0) {
243 struct ptlrpc_hr_service *hr = ptlrpc_hr;
246 idx = get_hr_thread_index(hr);
248 cfs_spin_lock(&hr->hr_threads[idx].hrt_lock);
249 cfs_list_splice_init(&b->rsb_replies,
250 &hr->hr_threads[idx].hrt_queue);
251 cfs_spin_unlock(&hr->hr_threads[idx].hrt_lock);
252 cfs_waitq_signal(&hr->hr_threads[idx].hrt_wait);
253 b->rsb_n_replies = 0;
258 * Add a reply to a batch.
259 * Add one reply object to a batch, schedule batched replies if overload.
264 static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs)
266 struct ptlrpc_service *svc = rs->rs_service;
268 if (svc != b->rsb_svc || b->rsb_n_replies >= MAX_SCHEDULED) {
269 if (b->rsb_svc != NULL) {
270 rs_batch_dispatch(b);
271 cfs_spin_unlock(&b->rsb_svc->srv_rs_lock);
273 cfs_spin_lock(&svc->srv_rs_lock);
276 cfs_spin_lock(&rs->rs_lock);
277 rs->rs_scheduled_ever = 1;
278 if (rs->rs_scheduled == 0) {
279 cfs_list_move(&rs->rs_list, &b->rsb_replies);
280 rs->rs_scheduled = 1;
283 rs->rs_committed = 1;
284 cfs_spin_unlock(&rs->rs_lock);
288 * Reply batch finalization.
289 * Dispatch remaining replies from the batch
290 * and release remaining spinlock.
294 static void rs_batch_fini(struct rs_batch *b)
296 if (b->rsb_svc != 0) {
297 rs_batch_dispatch(b);
298 cfs_spin_unlock(&b->rsb_svc->srv_rs_lock);
302 #define DECLARE_RS_BATCH(b) struct rs_batch b
304 #else /* __KERNEL__ */
306 #define rs_batch_init(b) do{}while(0)
307 #define rs_batch_fini(b) do{}while(0)
308 #define rs_batch_add(b, r) ptlrpc_schedule_difficult_reply(r)
309 #define DECLARE_RS_BATCH(b)
311 #endif /* __KERNEL__ */
314 * Put reply state into a queue for processing because we received
315 * ACK from the client
317 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
320 struct ptlrpc_hr_service *hr = ptlrpc_hr;
324 LASSERT(cfs_list_empty(&rs->rs_list));
326 idx = get_hr_thread_index(hr);
327 cfs_spin_lock(&hr->hr_threads[idx].hrt_lock);
328 cfs_list_add_tail(&rs->rs_list, &hr->hr_threads[idx].hrt_queue);
329 cfs_spin_unlock(&hr->hr_threads[idx].hrt_lock);
330 cfs_waitq_signal(&hr->hr_threads[idx].hrt_wait);
333 cfs_list_add_tail(&rs->rs_list, &rs->rs_service->srv_reply_queue);
338 ptlrpc_schedule_difficult_reply (struct ptlrpc_reply_state *rs)
342 LASSERT_SPIN_LOCKED(&rs->rs_service->srv_rs_lock);
343 LASSERT_SPIN_LOCKED(&rs->rs_lock);
344 LASSERT (rs->rs_difficult);
345 rs->rs_scheduled_ever = 1; /* flag any notification attempt */
347 if (rs->rs_scheduled) { /* being set up or already notified */
352 rs->rs_scheduled = 1;
353 cfs_list_del_init(&rs->rs_list);
354 ptlrpc_dispatch_difficult_reply(rs);
358 void ptlrpc_commit_replies(struct obd_export *exp)
360 struct ptlrpc_reply_state *rs, *nxt;
361 DECLARE_RS_BATCH(batch);
364 rs_batch_init(&batch);
365 /* Find any replies that have been committed and get their service
366 * to attend to complete them. */
368 /* CAVEAT EMPTOR: spinlock ordering!!! */
369 cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
370 cfs_list_for_each_entry_safe(rs, nxt, &exp->exp_uncommitted_replies,
372 LASSERT (rs->rs_difficult);
373 /* VBR: per-export last_committed */
374 LASSERT(rs->rs_export);
375 if (rs->rs_transno <= exp->exp_last_committed) {
376 cfs_list_del_init(&rs->rs_obd_list);
377 rs_batch_add(&batch, rs);
380 cfs_spin_unlock(&exp->exp_uncommitted_replies_lock);
381 rs_batch_fini(&batch);
386 ptlrpc_server_post_idle_rqbds (struct ptlrpc_service *svc)
388 struct ptlrpc_request_buffer_desc *rqbd;
393 cfs_spin_lock(&svc->srv_lock);
395 if (cfs_list_empty (&svc->srv_idle_rqbds)) {
396 cfs_spin_unlock(&svc->srv_lock);
400 rqbd = cfs_list_entry(svc->srv_idle_rqbds.next,
401 struct ptlrpc_request_buffer_desc,
403 cfs_list_del (&rqbd->rqbd_list);
405 /* assume we will post successfully */
406 svc->srv_nrqbd_receiving++;
407 cfs_list_add (&rqbd->rqbd_list, &svc->srv_active_rqbds);
409 cfs_spin_unlock(&svc->srv_lock);
411 rc = ptlrpc_register_rqbd(rqbd);
418 cfs_spin_lock(&svc->srv_lock);
420 svc->srv_nrqbd_receiving--;
421 cfs_list_del(&rqbd->rqbd_list);
422 cfs_list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
424 /* Don't complain if no request buffers are posted right now; LNET
425 * won't drop requests because we set the portal lazy! */
427 cfs_spin_unlock(&svc->srv_lock);
433 * Start a service with parameters from struct ptlrpc_service_conf \a c
434 * as opposed to directly calling ptlrpc_init_svc with tons of arguments.
436 struct ptlrpc_service *ptlrpc_init_svc_conf(struct ptlrpc_service_conf *c,
437 svc_handler_t h, char *name,
438 struct proc_dir_entry *proc_entry,
439 svc_req_printfn_t prntfn,
442 return ptlrpc_init_svc(c->psc_nbufs, c->psc_bufsize,
443 c->psc_max_req_size, c->psc_max_reply_size,
444 c->psc_req_portal, c->psc_rep_portal,
445 c->psc_watchdog_factor,
447 prntfn, c->psc_min_threads, c->psc_max_threads,
448 threadname, c->psc_ctx_tags, NULL);
450 EXPORT_SYMBOL(ptlrpc_init_svc_conf);
452 static void ptlrpc_at_timer(unsigned long castmeharder)
454 struct ptlrpc_service *svc = (struct ptlrpc_service *)castmeharder;
455 svc->srv_at_check = 1;
456 svc->srv_at_checktime = cfs_time_current();
457 cfs_waitq_signal(&svc->srv_waitq);
461 * Initialize service on a given portal.
462 * This includes starting serving threads , allocating and posting rqbds and
464 * \a nbufs is how many buffers to post
465 * \a bufsize is buffer size to post
466 * \a max_req_size - maximum request size to be accepted for this service
467 * \a max_reply_size maximum reply size this service can ever send
468 * \a req_portal - portal to listed for requests on
469 * \a rep_portal - portal of where to send replies to
470 * \a watchdog_factor soft watchdog timeout multiplifier to print stuck service traces.
471 * \a handler - function to process every new request
472 * \a name - service name
473 * \a proc_entry - entry in the /proc tree for sttistics reporting
474 * \a min_threads \a max_threads - min/max number of service threads to start.
475 * \a threadname should be 11 characters or less - 3 will be added on
476 * \a hp_handler - function to determine priority of the request, also called
477 * on every new request.
479 struct ptlrpc_service *
480 ptlrpc_init_svc(int nbufs, int bufsize, int max_req_size, int max_reply_size,
481 int req_portal, int rep_portal, int watchdog_factor,
482 svc_handler_t handler, char *name,
483 cfs_proc_dir_entry_t *proc_entry,
484 svc_req_printfn_t svcreq_printfn,
485 int min_threads, int max_threads,
486 char *threadname, __u32 ctx_tags,
487 svc_hpreq_handler_t hp_handler)
490 struct ptlrpc_at_array *array;
491 struct ptlrpc_service *service;
492 unsigned int size, index;
496 LASSERT (bufsize >= max_req_size + SPTLRPC_MAX_PAYLOAD);
497 LASSERT (ctx_tags != 0);
499 OBD_ALLOC_PTR(service);
503 /* First initialise enough for early teardown */
505 service->srv_name = name;
506 cfs_spin_lock_init(&service->srv_lock);
507 cfs_spin_lock_init(&service->srv_rq_lock);
508 cfs_spin_lock_init(&service->srv_rs_lock);
509 CFS_INIT_LIST_HEAD(&service->srv_threads);
510 cfs_waitq_init(&service->srv_waitq);
512 service->srv_nbuf_per_group = test_req_buffer_pressure ? 1 : nbufs;
513 service->srv_max_req_size = max_req_size + SPTLRPC_MAX_PAYLOAD;
514 service->srv_buf_size = bufsize;
515 service->srv_rep_portal = rep_portal;
516 service->srv_req_portal = req_portal;
517 service->srv_watchdog_factor = watchdog_factor;
518 service->srv_handler = handler;
519 service->srv_req_printfn = svcreq_printfn;
520 service->srv_request_seq = 1; /* valid seq #s start at 1 */
521 service->srv_request_max_cull_seq = 0;
522 service->srv_threads_min = min_threads;
523 service->srv_threads_max = max_threads;
524 service->srv_thread_name = threadname;
525 service->srv_ctx_tags = ctx_tags;
526 service->srv_hpreq_handler = hp_handler;
527 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
528 service->srv_hpreq_count = 0;
529 service->srv_n_active_hpreq = 0;
531 rc = LNetSetLazyPortal(service->srv_req_portal);
534 CFS_INIT_LIST_HEAD(&service->srv_request_queue);
535 CFS_INIT_LIST_HEAD(&service->srv_request_hpq);
536 CFS_INIT_LIST_HEAD(&service->srv_idle_rqbds);
537 CFS_INIT_LIST_HEAD(&service->srv_active_rqbds);
538 CFS_INIT_LIST_HEAD(&service->srv_history_rqbds);
539 CFS_INIT_LIST_HEAD(&service->srv_request_history);
540 CFS_INIT_LIST_HEAD(&service->srv_active_replies);
542 CFS_INIT_LIST_HEAD(&service->srv_reply_queue);
544 CFS_INIT_LIST_HEAD(&service->srv_free_rs_list);
545 cfs_waitq_init(&service->srv_free_rs_waitq);
546 cfs_atomic_set(&service->srv_n_difficult_replies, 0);
548 cfs_spin_lock_init(&service->srv_at_lock);
549 CFS_INIT_LIST_HEAD(&service->srv_req_in_queue);
551 array = &service->srv_at_array;
552 size = at_est2timeout(at_max);
553 array->paa_size = size;
554 array->paa_count = 0;
555 array->paa_deadline = -1;
557 /* allocate memory for srv_at_array (ptlrpc_at_array) */
558 OBD_ALLOC(array->paa_reqs_array, sizeof(cfs_list_t) * size);
559 if (array->paa_reqs_array == NULL)
562 for (index = 0; index < size; index++)
563 CFS_INIT_LIST_HEAD(&array->paa_reqs_array[index]);
565 OBD_ALLOC(array->paa_reqs_count, sizeof(__u32) * size);
566 if (array->paa_reqs_count == NULL)
569 cfs_timer_init(&service->srv_at_timer, ptlrpc_at_timer, service);
570 /* At SOW, service time should be quick; 10s seems generous. If client
571 timeout is less than this, we'll be sending an early reply. */
572 at_init(&service->srv_at_estimate, 10, 0);
574 cfs_spin_lock (&ptlrpc_all_services_lock);
575 cfs_list_add (&service->srv_list, &ptlrpc_all_services);
576 cfs_spin_unlock (&ptlrpc_all_services_lock);
578 /* Now allocate the request buffers */
579 rc = ptlrpc_grow_req_bufs(service);
580 /* We shouldn't be under memory pressure at startup, so
581 * fail if we can't post all our buffers at this time. */
585 /* Now allocate pool of reply buffers */
586 /* Increase max reply size to next power of two */
587 service->srv_max_reply_size = 1;
588 while (service->srv_max_reply_size <
589 max_reply_size + SPTLRPC_MAX_PAYLOAD)
590 service->srv_max_reply_size <<= 1;
592 if (proc_entry != NULL)
593 ptlrpc_lprocfs_register_service(proc_entry, service);
595 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
596 service->srv_name, service->srv_req_portal);
600 ptlrpc_unregister_service(service);
605 * to actually free the request, must be called without holding svc_lock.
606 * note it's caller's responsibility to unlink req->rq_list.
608 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
610 LASSERT(cfs_atomic_read(&req->rq_refcount) == 0);
611 LASSERT(cfs_list_empty(&req->rq_timed_list));
613 /* DEBUG_REQ() assumes the reply state of a request with a valid
614 * ref will not be destroyed until that reference is dropped. */
615 ptlrpc_req_drop_rs(req);
617 sptlrpc_svc_ctx_decref(req);
619 if (req != &req->rq_rqbd->rqbd_req) {
620 /* NB request buffers use an embedded
621 * req if the incoming req unlinked the
622 * MD; this isn't one of them! */
623 OBD_FREE(req, sizeof(*req));
628 * drop a reference count of the request. if it reaches 0, we either
629 * put it into history list, or free it immediately.
631 void ptlrpc_server_drop_request(struct ptlrpc_request *req)
633 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
634 struct ptlrpc_service *svc = rqbd->rqbd_service;
639 if (!cfs_atomic_dec_and_test(&req->rq_refcount))
642 cfs_spin_lock(&svc->srv_at_lock);
643 if (req->rq_at_linked) {
644 struct ptlrpc_at_array *array = &svc->srv_at_array;
645 __u32 index = req->rq_at_index;
647 LASSERT(!cfs_list_empty(&req->rq_timed_list));
648 cfs_list_del_init(&req->rq_timed_list);
649 cfs_spin_lock(&req->rq_lock);
650 req->rq_at_linked = 0;
651 cfs_spin_unlock(&req->rq_lock);
652 array->paa_reqs_count[index]--;
655 LASSERT(cfs_list_empty(&req->rq_timed_list));
656 cfs_spin_unlock(&svc->srv_at_lock);
658 /* finalize request */
659 if (req->rq_export) {
660 class_export_put(req->rq_export);
661 req->rq_export = NULL;
664 cfs_spin_lock(&svc->srv_lock);
666 cfs_list_add(&req->rq_list, &rqbd->rqbd_reqs);
668 refcount = --(rqbd->rqbd_refcount);
670 /* request buffer is now idle: add to history */
671 cfs_list_del(&rqbd->rqbd_list);
672 cfs_list_add_tail(&rqbd->rqbd_list, &svc->srv_history_rqbds);
673 svc->srv_n_history_rqbds++;
675 /* cull some history?
676 * I expect only about 1 or 2 rqbds need to be recycled here */
677 while (svc->srv_n_history_rqbds > svc->srv_max_history_rqbds) {
678 rqbd = cfs_list_entry(svc->srv_history_rqbds.next,
679 struct ptlrpc_request_buffer_desc,
682 cfs_list_del(&rqbd->rqbd_list);
683 svc->srv_n_history_rqbds--;
685 /* remove rqbd's reqs from svc's req history while
686 * I've got the service lock */
687 cfs_list_for_each(tmp, &rqbd->rqbd_reqs) {
688 req = cfs_list_entry(tmp, struct ptlrpc_request,
690 /* Track the highest culled req seq */
691 if (req->rq_history_seq >
692 svc->srv_request_max_cull_seq)
693 svc->srv_request_max_cull_seq =
695 cfs_list_del(&req->rq_history_list);
698 cfs_spin_unlock(&svc->srv_lock);
700 cfs_list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
701 req = cfs_list_entry(rqbd->rqbd_reqs.next,
702 struct ptlrpc_request,
704 cfs_list_del(&req->rq_list);
705 ptlrpc_server_free_request(req);
708 cfs_spin_lock(&svc->srv_lock);
710 * now all reqs including the embedded req has been
711 * disposed, schedule request buffer for re-use.
713 LASSERT(cfs_atomic_read(&rqbd->rqbd_req.rq_refcount) ==
715 cfs_list_add_tail(&rqbd->rqbd_list,
716 &svc->srv_idle_rqbds);
719 cfs_spin_unlock(&svc->srv_lock);
720 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
721 /* If we are low on memory, we are not interested in history */
722 cfs_list_del(&req->rq_list);
723 cfs_list_del_init(&req->rq_history_list);
724 cfs_spin_unlock(&svc->srv_lock);
726 ptlrpc_server_free_request(req);
728 cfs_spin_unlock(&svc->srv_lock);
733 * to finish a request: stop sending more early replies, and release
734 * the request. should be called after we finished handling the request.
736 static void ptlrpc_server_finish_request(struct ptlrpc_service *svc,
737 struct ptlrpc_request *req)
739 ptlrpc_hpreq_fini(req);
741 cfs_spin_lock(&svc->srv_rq_lock);
742 svc->srv_n_active_reqs--;
744 svc->srv_n_active_hpreq--;
745 cfs_spin_unlock(&svc->srv_rq_lock);
747 ptlrpc_server_drop_request(req);
751 * This function makes sure dead exports are evicted in a timely manner.
752 * This function is only called when some export receives a message (i.e.,
753 * 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;
835 * Sanity check request \a req.
836 * Return 0 if all is ok, error code otherwise.
838 static int ptlrpc_check_req(struct ptlrpc_request *req)
842 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
843 req->rq_export->exp_conn_cnt)) {
844 DEBUG_REQ(D_ERROR, req,
845 "DROPPING req from old connection %d < %d",
846 lustre_msg_get_conn_cnt(req->rq_reqmsg),
847 req->rq_export->exp_conn_cnt);
850 if (unlikely(req->rq_export->exp_obd &&
851 req->rq_export->exp_obd->obd_fail)) {
852 /* Failing over, don't handle any more reqs, send
853 error response instead. */
854 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
855 req, req->rq_export->exp_obd->obd_name);
857 } else if (lustre_msg_get_flags(req->rq_reqmsg) &
858 (MSG_REPLAY | MSG_REQ_REPLAY_DONE) &&
859 !(req->rq_export->exp_obd->obd_recovering)) {
860 DEBUG_REQ(D_ERROR, req,
861 "Invalid replay without recovery");
862 class_fail_export(req->rq_export);
864 } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 &&
865 !(req->rq_export->exp_obd->obd_recovering)) {
866 DEBUG_REQ(D_ERROR, req, "Invalid req with transno "
867 LPU64" without recovery",
868 lustre_msg_get_transno(req->rq_reqmsg));
869 class_fail_export(req->rq_export);
873 if (unlikely(rc < 0)) {
880 static void ptlrpc_at_set_timer(struct ptlrpc_service *svc)
882 struct ptlrpc_at_array *array = &svc->srv_at_array;
885 cfs_spin_lock(&svc->srv_at_lock);
886 if (array->paa_count == 0) {
887 cfs_timer_disarm(&svc->srv_at_timer);
888 cfs_spin_unlock(&svc->srv_at_lock);
892 /* Set timer for closest deadline */
893 next = (__s32)(array->paa_deadline - cfs_time_current_sec() -
896 ptlrpc_at_timer((unsigned long)svc);
898 cfs_timer_arm(&svc->srv_at_timer, cfs_time_shift(next));
899 cfs_spin_unlock(&svc->srv_at_lock);
900 CDEBUG(D_INFO, "armed %s at %+ds\n", svc->srv_name, next);
903 /* Add rpc to early reply check list */
904 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
906 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
907 struct ptlrpc_request *rq = NULL;
908 struct ptlrpc_at_array *array = &svc->srv_at_array;
915 if (req->rq_no_reply)
918 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
921 cfs_spin_lock(&svc->srv_at_lock);
922 LASSERT(cfs_list_empty(&req->rq_timed_list));
924 index = (unsigned long)req->rq_deadline % array->paa_size;
925 if (array->paa_reqs_count[index] > 0) {
926 /* latest rpcs will have the latest deadlines in the list,
927 * so search backward. */
928 cfs_list_for_each_entry_reverse(rq,
929 &array->paa_reqs_array[index],
931 if (req->rq_deadline >= rq->rq_deadline) {
932 cfs_list_add(&req->rq_timed_list,
939 /* Add the request at the head of the list */
940 if (cfs_list_empty(&req->rq_timed_list))
941 cfs_list_add(&req->rq_timed_list,
942 &array->paa_reqs_array[index]);
944 cfs_spin_lock(&req->rq_lock);
945 req->rq_at_linked = 1;
946 cfs_spin_unlock(&req->rq_lock);
947 req->rq_at_index = index;
948 array->paa_reqs_count[index]++;
950 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
951 array->paa_deadline = req->rq_deadline;
954 cfs_spin_unlock(&svc->srv_at_lock);
957 ptlrpc_at_set_timer(svc);
962 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
964 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
965 struct ptlrpc_request *reqcopy;
966 struct lustre_msg *reqmsg;
967 cfs_duration_t olddl = req->rq_deadline - cfs_time_current_sec();
972 /* deadline is when the client expects us to reply, margin is the
973 difference between clients' and servers' expectations */
974 DEBUG_REQ(D_ADAPTTO, req,
975 "%ssending early reply (deadline %+lds, margin %+lds) for "
976 "%d+%d", AT_OFF ? "AT off - not " : "",
977 olddl, olddl - at_get(&svc->srv_at_estimate),
978 at_get(&svc->srv_at_estimate), at_extra);
984 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), "
985 "not sending early reply. Consider increasing "
986 "at_early_margin (%d)?", olddl, at_early_margin);
988 /* Return an error so we're not re-added to the timed list. */
992 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
993 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
994 "but no AT support");
998 if (req->rq_export &&
999 lustre_msg_get_flags(req->rq_reqmsg) &
1000 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1001 /* During recovery, we don't want to send too many early
1002 * replies, but on the other hand we want to make sure the
1003 * client has enough time to resend if the rpc is lost. So
1004 * during the recovery period send at least 4 early replies,
1005 * spacing them every at_extra if we can. at_estimate should
1006 * always equal this fixed value during recovery. */
1007 at_measured(&svc->srv_at_estimate, min(at_extra,
1008 req->rq_export->exp_obd->obd_recovery_timeout / 4));
1010 /* Fake our processing time into the future to ask the clients
1011 * for some extra amount of time */
1012 at_measured(&svc->srv_at_estimate, at_extra +
1013 cfs_time_current_sec() -
1014 req->rq_arrival_time.tv_sec);
1016 /* Check to see if we've actually increased the deadline -
1017 * we may be past adaptive_max */
1018 if (req->rq_deadline >= req->rq_arrival_time.tv_sec +
1019 at_get(&svc->srv_at_estimate)) {
1020 DEBUG_REQ(D_WARNING, req, "Couldn't add any time "
1021 "(%ld/%ld), not sending early reply\n",
1022 olddl, req->rq_arrival_time.tv_sec +
1023 at_get(&svc->srv_at_estimate) -
1024 cfs_time_current_sec());
1028 newdl = cfs_time_current_sec() + at_get(&svc->srv_at_estimate);
1030 OBD_ALLOC(reqcopy, sizeof *reqcopy);
1031 if (reqcopy == NULL)
1033 OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1035 OBD_FREE(reqcopy, sizeof *reqcopy);
1040 reqcopy->rq_reply_state = NULL;
1041 reqcopy->rq_rep_swab_mask = 0;
1042 reqcopy->rq_pack_bulk = 0;
1043 reqcopy->rq_pack_udesc = 0;
1044 reqcopy->rq_packed_final = 0;
1045 sptlrpc_svc_ctx_addref(reqcopy);
1046 /* We only need the reqmsg for the magic */
1047 reqcopy->rq_reqmsg = reqmsg;
1048 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1050 LASSERT(cfs_atomic_read(&req->rq_refcount));
1051 /** if it is last refcount then early reply isn't needed */
1052 if (cfs_atomic_read(&req->rq_refcount) == 1) {
1053 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
1054 "abort sending early reply\n");
1055 GOTO(out, rc = -EINVAL);
1058 /* Connection ref */
1059 reqcopy->rq_export = class_conn2export(
1060 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1061 if (reqcopy->rq_export == NULL)
1062 GOTO(out, rc = -ENODEV);
1065 class_export_rpc_get(reqcopy->rq_export);
1066 if (reqcopy->rq_export->exp_obd &&
1067 reqcopy->rq_export->exp_obd->obd_fail)
1068 GOTO(out_put, rc = -ENODEV);
1070 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1074 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1077 /* Adjust our own deadline to what we told the client */
1078 req->rq_deadline = newdl;
1079 req->rq_early_count++; /* number sent, server side */
1081 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1084 /* Free the (early) reply state from lustre_pack_reply.
1085 (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
1086 ptlrpc_req_drop_rs(reqcopy);
1089 class_export_rpc_put(reqcopy->rq_export);
1090 class_export_put(reqcopy->rq_export);
1092 sptlrpc_svc_ctx_decref(reqcopy);
1093 OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1094 OBD_FREE(reqcopy, sizeof *reqcopy);
1098 /* Send early replies to everybody expiring within at_early_margin
1099 asking for at_extra time */
1100 static int ptlrpc_at_check_timed(struct ptlrpc_service *svc)
1102 struct ptlrpc_request *rq, *n;
1103 cfs_list_t work_list;
1104 struct ptlrpc_at_array *array = &svc->srv_at_array;
1107 time_t now = cfs_time_current_sec();
1108 cfs_duration_t delay;
1109 int first, counter = 0;
1112 cfs_spin_lock(&svc->srv_at_lock);
1113 if (svc->srv_at_check == 0) {
1114 cfs_spin_unlock(&svc->srv_at_lock);
1117 delay = cfs_time_sub(cfs_time_current(), svc->srv_at_checktime);
1118 svc->srv_at_check = 0;
1120 if (array->paa_count == 0) {
1121 cfs_spin_unlock(&svc->srv_at_lock);
1125 /* The timer went off, but maybe the nearest rpc already completed. */
1126 first = array->paa_deadline - now;
1127 if (first > at_early_margin) {
1128 /* We've still got plenty of time. Reset the timer. */
1129 cfs_spin_unlock(&svc->srv_at_lock);
1130 ptlrpc_at_set_timer(svc);
1134 /* We're close to a timeout, and we don't know how much longer the
1135 server will take. Send early replies to everyone expiring soon. */
1136 CFS_INIT_LIST_HEAD(&work_list);
1138 index = (unsigned long)array->paa_deadline % array->paa_size;
1139 count = array->paa_count;
1141 count -= array->paa_reqs_count[index];
1142 cfs_list_for_each_entry_safe(rq, n,
1143 &array->paa_reqs_array[index],
1145 if (rq->rq_deadline <= now + at_early_margin) {
1146 cfs_list_del_init(&rq->rq_timed_list);
1148 * ptlrpc_server_drop_request() may drop
1149 * refcount to 0 already. Let's check this and
1150 * don't add entry to work_list
1152 if (likely(cfs_atomic_inc_not_zero(&rq->rq_refcount)))
1153 cfs_list_add(&rq->rq_timed_list, &work_list);
1155 array->paa_reqs_count[index]--;
1157 cfs_spin_lock(&rq->rq_lock);
1158 rq->rq_at_linked = 0;
1159 cfs_spin_unlock(&rq->rq_lock);
1163 /* update the earliest deadline */
1164 if (deadline == -1 || rq->rq_deadline < deadline)
1165 deadline = rq->rq_deadline;
1170 if (++index >= array->paa_size)
1173 array->paa_deadline = deadline;
1174 cfs_spin_unlock(&svc->srv_at_lock);
1176 /* we have a new earliest deadline, restart the timer */
1177 ptlrpc_at_set_timer(svc);
1179 CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
1180 "replies\n", first, at_extra, counter);
1182 /* We're already past request deadlines before we even get a
1183 chance to send early replies */
1184 LCONSOLE_WARN("%s: This server is not able to keep up with "
1185 "request traffic (cpu-bound).\n", svc->srv_name);
1186 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, "
1187 "delay="CFS_DURATION_T"(jiff)\n",
1188 counter, svc->srv_n_queued_reqs, svc->srv_n_active_reqs,
1189 at_get(&svc->srv_at_estimate), delay);
1192 /* we took additional refcount so entries can't be deleted from list, no
1193 * locking is needed */
1194 while (!cfs_list_empty(&work_list)) {
1195 rq = cfs_list_entry(work_list.next, struct ptlrpc_request,
1197 cfs_list_del_init(&rq->rq_timed_list);
1199 if (ptlrpc_at_send_early_reply(rq) == 0)
1200 ptlrpc_at_add_timed(rq);
1202 ptlrpc_server_drop_request(rq);
1209 * Put the request to the export list if the request may become
1210 * a high priority one.
1212 static int ptlrpc_hpreq_init(struct ptlrpc_service *svc,
1213 struct ptlrpc_request *req)
1218 if (svc->srv_hpreq_handler) {
1219 rc = svc->srv_hpreq_handler(req);
1223 if (req->rq_export && req->rq_ops) {
1224 /* Perform request specific check. We should do this check
1225 * before the request is added into exp_hp_rpcs list otherwise
1226 * it may hit swab race at LU-1044. */
1227 if (req->rq_ops->hpreq_check)
1228 rc = req->rq_ops->hpreq_check(req);
1230 cfs_spin_lock_bh(&req->rq_export->exp_rpc_lock);
1231 cfs_list_add(&req->rq_exp_list,
1232 &req->rq_export->exp_hp_rpcs);
1233 cfs_spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1239 /** Remove the request from the export list. */
1240 static void ptlrpc_hpreq_fini(struct ptlrpc_request *req)
1243 if (req->rq_export && req->rq_ops) {
1244 /* refresh lock timeout again so that client has more
1245 * room to send lock cancel RPC. */
1246 if (req->rq_ops->hpreq_fini)
1247 req->rq_ops->hpreq_fini(req);
1249 cfs_spin_lock_bh(&req->rq_export->exp_rpc_lock);
1250 cfs_list_del_init(&req->rq_exp_list);
1251 cfs_spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1257 * Make the request a high priority one.
1259 * All the high priority requests are queued in a separate FIFO
1260 * ptlrpc_service::srv_request_hpq list which is parallel to
1261 * ptlrpc_service::srv_request_queue list but has a higher priority
1264 * \see ptlrpc_server_handle_request().
1266 static void ptlrpc_hpreq_reorder_nolock(struct ptlrpc_service *svc,
1267 struct ptlrpc_request *req)
1270 LASSERT(svc != NULL);
1271 cfs_spin_lock(&req->rq_lock);
1272 if (req->rq_hp == 0) {
1273 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1275 /* Add to the high priority queue. */
1276 cfs_list_move_tail(&req->rq_list, &svc->srv_request_hpq);
1278 if (opc != OBD_PING)
1279 DEBUG_REQ(D_RPCTRACE, req, "high priority req");
1281 cfs_spin_unlock(&req->rq_lock);
1286 * \see ptlrpc_hpreq_reorder_nolock
1288 void ptlrpc_hpreq_reorder(struct ptlrpc_request *req)
1290 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
1293 cfs_spin_lock(&svc->srv_rq_lock);
1294 /* It may happen that the request is already taken for the processing
1295 * but still in the export list, or the request is not in the request
1296 * queue but in the export list already, do not add it into the
1298 if (!cfs_list_empty(&req->rq_list))
1299 ptlrpc_hpreq_reorder_nolock(svc, req);
1300 cfs_spin_unlock(&svc->srv_rq_lock);
1304 /** Check if the request is a high priority one. */
1305 static int ptlrpc_server_hpreq_check(struct ptlrpc_service *svc,
1306 struct ptlrpc_request *req)
1310 /* Check by request opc. */
1311 if (OBD_PING == lustre_msg_get_opc(req->rq_reqmsg))
1314 RETURN(ptlrpc_hpreq_init(svc, req));
1317 /** Check if a request is a high priority one. */
1318 static int ptlrpc_server_request_add(struct ptlrpc_service *svc,
1319 struct ptlrpc_request *req)
1324 rc = ptlrpc_server_hpreq_check(svc, req);
1328 cfs_spin_lock(&svc->srv_rq_lock);
1331 ptlrpc_hpreq_reorder_nolock(svc, req);
1333 cfs_list_add_tail(&req->rq_list,
1334 &svc->srv_request_queue);
1336 cfs_spin_unlock(&svc->srv_rq_lock);
1342 * Allow to handle high priority request
1343 * User can call it w/o any lock but need to hold ptlrpc_service::srv_rq_lock
1344 * to get reliable result
1346 static int ptlrpc_server_allow_high(struct ptlrpc_service *svc, int force)
1351 if (svc->srv_n_active_reqs >= svc->srv_threads_running - 1)
1354 return cfs_list_empty(&svc->srv_request_queue) ||
1355 svc->srv_hpreq_count < svc->srv_hpreq_ratio;
1358 static int ptlrpc_server_high_pending(struct ptlrpc_service *svc, int force)
1360 return ptlrpc_server_allow_high(svc, force) &&
1361 !cfs_list_empty(&svc->srv_request_hpq);
1365 * Only allow normal priority requests on a service that has a high-priority
1366 * queue if forced (i.e. cleanup), if there are other high priority requests
1367 * already being processed (i.e. those threads can service more high-priority
1368 * requests), or if there are enough idle threads that a later thread can do
1369 * a high priority request.
1370 * User can call it w/o any lock but need to hold ptlrpc_service::srv_rq_lock
1371 * to get reliable result
1373 static int ptlrpc_server_allow_normal(struct ptlrpc_service *svc, int force)
1376 if (1) /* always allow to handle normal request for liblustre */
1380 svc->srv_n_active_reqs < svc->srv_threads_running - 2)
1383 if (svc->srv_n_active_reqs >= svc->srv_threads_running - 1)
1386 return svc->srv_n_active_hpreq > 0 || svc->srv_hpreq_handler == NULL;
1389 static int ptlrpc_server_normal_pending(struct ptlrpc_service *svc, int force)
1391 return ptlrpc_server_allow_normal(svc, force) &&
1392 !cfs_list_empty(&svc->srv_request_queue);
1396 * Returns true if there are requests available in incoming
1397 * request queue for processing and it is allowed to fetch them.
1398 * User can call it w/o any lock but need to hold ptlrpc_service::srv_rq_lock
1399 * to get reliable result
1400 * \see ptlrpc_server_allow_normal
1401 * \see ptlrpc_server_allow high
1404 ptlrpc_server_request_pending(struct ptlrpc_service *svc, int force)
1406 return ptlrpc_server_high_pending(svc, force) ||
1407 ptlrpc_server_normal_pending(svc, force);
1411 * Fetch a request for processing from queue of unprocessed requests.
1412 * Favors high-priority requests.
1413 * Returns a pointer to fetched request.
1415 static struct ptlrpc_request *
1416 ptlrpc_server_request_get(struct ptlrpc_service *svc, int force)
1418 struct ptlrpc_request *req;
1421 if (ptlrpc_server_high_pending(svc, force)) {
1422 req = cfs_list_entry(svc->srv_request_hpq.next,
1423 struct ptlrpc_request, rq_list);
1424 svc->srv_hpreq_count++;
1429 if (ptlrpc_server_normal_pending(svc, force)) {
1430 req = cfs_list_entry(svc->srv_request_queue.next,
1431 struct ptlrpc_request, rq_list);
1432 svc->srv_hpreq_count = 0;
1439 * Handle freshly incoming reqs, add to timed early reply list,
1440 * pass on to regular request queue.
1441 * All incoming requests pass through here before getting into
1442 * ptlrpc_server_handle_req later on.
1445 ptlrpc_server_handle_req_in(struct ptlrpc_service *svc)
1447 struct ptlrpc_request *req;
1454 cfs_spin_lock(&svc->srv_lock);
1455 if (cfs_list_empty(&svc->srv_req_in_queue)) {
1456 cfs_spin_unlock(&svc->srv_lock);
1460 req = cfs_list_entry(svc->srv_req_in_queue.next,
1461 struct ptlrpc_request, rq_list);
1462 cfs_list_del_init (&req->rq_list);
1463 svc->srv_n_queued_reqs--;
1464 /* Consider this still a "queued" request as far as stats are
1466 cfs_spin_unlock(&svc->srv_lock);
1468 /* go through security check/transform */
1469 rc = sptlrpc_svc_unwrap_request(req);
1473 case SECSVC_COMPLETE:
1474 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1483 * for null-flavored rpc, msg has been unpacked by sptlrpc, although
1484 * redo it wouldn't be harmful.
1486 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
1487 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
1489 CERROR("error unpacking request: ptl %d from %s "
1490 "x"LPU64"\n", svc->srv_req_portal,
1491 libcfs_id2str(req->rq_peer), req->rq_xid);
1496 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1498 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
1499 LPU64"\n", svc->srv_req_portal,
1500 libcfs_id2str(req->rq_peer), req->rq_xid);
1504 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
1505 lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
1506 CERROR("drop incoming rpc opc %u, x"LPU64"\n",
1507 cfs_fail_val, req->rq_xid);
1512 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1513 CERROR("wrong packet type received (type=%u) from %s\n",
1514 lustre_msg_get_type(req->rq_reqmsg),
1515 libcfs_id2str(req->rq_peer));
1519 switch(lustre_msg_get_opc(req->rq_reqmsg)) {
1522 req->rq_bulk_write = 1;
1526 case MGS_CONFIG_READ:
1527 req->rq_bulk_read = 1;
1531 CDEBUG(D_RPCTRACE, "got req x"LPU64"\n", req->rq_xid);
1533 req->rq_export = class_conn2export(
1534 lustre_msg_get_handle(req->rq_reqmsg));
1535 if (req->rq_export) {
1536 rc = ptlrpc_check_req(req);
1538 rc = sptlrpc_target_export_check(req->rq_export, req);
1540 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
1541 "illegal security flavor,");
1546 ptlrpc_update_export_timer(req->rq_export, 0);
1549 /* req_in handling should/must be fast */
1550 if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5)
1551 DEBUG_REQ(D_WARNING, req, "Slow req_in handling "CFS_DURATION_T"s",
1552 cfs_time_sub(cfs_time_current_sec(),
1553 req->rq_arrival_time.tv_sec));
1555 /* Set rpc server deadline and add it to the timed list */
1556 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1557 MSGHDR_AT_SUPPORT) ?
1558 /* The max time the client expects us to take */
1559 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1560 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1561 if (unlikely(deadline == 0)) {
1562 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1566 ptlrpc_at_add_timed(req);
1568 /* Move it over to the request processing queue */
1569 rc = ptlrpc_server_request_add(svc, req);
1571 ptlrpc_hpreq_fini(req);
1574 cfs_waitq_signal(&svc->srv_waitq);
1578 cfs_spin_lock(&svc->srv_rq_lock);
1579 svc->srv_n_active_reqs++;
1580 cfs_spin_unlock(&svc->srv_rq_lock);
1581 ptlrpc_server_finish_request(svc, req);
1587 * Main incoming request handling logic.
1588 * Calls handler function from service to do actual processing.
1591 ptlrpc_server_handle_request(struct ptlrpc_service *svc,
1592 struct ptlrpc_thread *thread)
1594 struct obd_export *export = NULL;
1595 struct ptlrpc_request *request;
1596 struct timeval work_start;
1597 struct timeval work_end;
1605 cfs_spin_lock(&svc->srv_rq_lock);
1607 /* !@%$# liblustre only has 1 thread */
1608 if (cfs_atomic_read(&svc->srv_n_difficult_replies) != 0) {
1609 cfs_spin_unlock(&svc->srv_rq_lock);
1613 request = ptlrpc_server_request_get(svc, 0);
1614 if (request == NULL) {
1615 cfs_spin_unlock(&svc->srv_rq_lock);
1619 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
1620 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
1621 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
1622 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
1624 if (unlikely(fail_opc)) {
1625 if (request->rq_export && request->rq_ops) {
1626 cfs_spin_unlock(&svc->srv_rq_lock);
1627 OBD_FAIL_TIMEOUT(fail_opc, 4);
1628 cfs_spin_lock(&svc->srv_rq_lock);
1629 request = ptlrpc_server_request_get(svc, 0);
1630 if (request == NULL) {
1631 cfs_spin_unlock(&svc->srv_rq_lock);
1637 cfs_list_del_init(&request->rq_list);
1638 svc->srv_n_active_reqs++;
1640 svc->srv_n_active_hpreq++;
1642 cfs_spin_unlock(&svc->srv_rq_lock);
1644 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
1646 if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
1647 libcfs_debug_dumplog();
1649 cfs_gettimeofday(&work_start);
1650 timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
1651 if (likely(svc->srv_stats != NULL)) {
1652 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
1654 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
1655 svc->srv_n_queued_reqs);
1656 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
1657 svc->srv_n_active_reqs);
1658 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
1659 at_get(&svc->srv_at_estimate));
1662 rc = lu_context_init(&request->rq_session,
1663 LCT_SESSION|LCT_REMEMBER|LCT_NOREF);
1665 CERROR("Failure to initialize session: %d\n", rc);
1668 request->rq_session.lc_thread = thread;
1669 request->rq_session.lc_cookie = 0x5;
1670 lu_context_enter(&request->rq_session);
1672 CDEBUG(D_NET, "got req "LPU64"\n", request->rq_xid);
1674 request->rq_svc_thread = thread;
1676 request->rq_svc_thread->t_env->le_ses = &request->rq_session;
1678 if (likely(request->rq_export)) {
1679 if (unlikely(ptlrpc_check_req(request)))
1681 ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
1682 export = class_export_rpc_get(request->rq_export);
1685 /* Discard requests queued for longer than the deadline.
1686 The deadline is increased if we send an early reply. */
1687 if (cfs_time_current_sec() > request->rq_deadline) {
1688 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s"
1689 ": deadline "CFS_DURATION_T":"CFS_DURATION_T"s ago\n",
1690 libcfs_id2str(request->rq_peer),
1691 cfs_time_sub(request->rq_deadline,
1692 request->rq_arrival_time.tv_sec),
1693 cfs_time_sub(cfs_time_current_sec(),
1694 request->rq_deadline));
1695 goto put_rpc_export;
1698 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
1699 "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1700 (request->rq_export ?
1701 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1702 (request->rq_export ?
1703 cfs_atomic_read(&request->rq_export->exp_refcount) : -99),
1704 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1705 libcfs_id2str(request->rq_peer),
1706 lustre_msg_get_opc(request->rq_reqmsg));
1708 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
1709 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
1711 rc = svc->srv_handler(request);
1713 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
1717 class_export_rpc_put(export);
1719 lu_context_exit(&request->rq_session);
1720 lu_context_fini(&request->rq_session);
1722 if (unlikely(cfs_time_current_sec() > request->rq_deadline)) {
1723 DEBUG_REQ(D_WARNING, request, "Request x"LPU64" took longer "
1724 "than estimated ("CFS_DURATION_T":"CFS_DURATION_T"s);"
1725 " client may timeout.",
1726 request->rq_xid, cfs_time_sub(request->rq_deadline,
1727 request->rq_arrival_time.tv_sec),
1728 cfs_time_sub(cfs_time_current_sec(),
1729 request->rq_deadline));
1732 cfs_gettimeofday(&work_end);
1733 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1734 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
1735 "%s:%s+%d:%d:x"LPU64":%s:%d Request procesed in "
1736 "%ldus (%ldus total) trans "LPU64" rc %d/%d\n",
1738 (request->rq_export ?
1739 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1740 (request->rq_export ?
1741 cfs_atomic_read(&request->rq_export->exp_refcount) : -99),
1742 lustre_msg_get_status(request->rq_reqmsg),
1744 libcfs_id2str(request->rq_peer),
1745 lustre_msg_get_opc(request->rq_reqmsg),
1747 cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL),
1748 (request->rq_repmsg ?
1749 lustre_msg_get_transno(request->rq_repmsg) :
1750 request->rq_transno),
1752 (request->rq_repmsg ?
1753 lustre_msg_get_status(request->rq_repmsg) : -999));
1754 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
1755 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
1756 int opc = opcode_offset(op);
1757 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
1758 LASSERT(opc < LUSTRE_MAX_OPCODES);
1759 lprocfs_counter_add(svc->srv_stats,
1760 opc + EXTRA_MAX_OPCODES,
1764 if (unlikely(request->rq_early_count)) {
1765 DEBUG_REQ(D_ADAPTTO, request,
1766 "sent %d early replies before finishing in "
1768 request->rq_early_count,
1769 cfs_time_sub(work_end.tv_sec,
1770 request->rq_arrival_time.tv_sec));
1774 ptlrpc_server_finish_request(svc, request);
1780 * An internal function to process a single reply state object.
1783 ptlrpc_handle_rs (struct ptlrpc_reply_state *rs)
1785 struct ptlrpc_service *svc = rs->rs_service;
1786 struct obd_export *exp;
1791 exp = rs->rs_export;
1793 LASSERT (rs->rs_difficult);
1794 LASSERT (rs->rs_scheduled);
1795 LASSERT (cfs_list_empty(&rs->rs_list));
1797 cfs_spin_lock (&exp->exp_lock);
1798 /* Noop if removed already */
1799 cfs_list_del_init (&rs->rs_exp_list);
1800 cfs_spin_unlock (&exp->exp_lock);
1802 /* The disk commit callback holds exp_uncommitted_replies_lock while it
1803 * iterates over newly committed replies, removing them from
1804 * exp_uncommitted_replies. It then drops this lock and schedules the
1805 * replies it found for handling here.
1807 * We can avoid contention for exp_uncommitted_replies_lock between the
1808 * HRT threads and further commit callbacks by checking rs_committed
1809 * which is set in the commit callback while it holds both
1810 * rs_lock and exp_uncommitted_reples.
1812 * If we see rs_committed clear, the commit callback _may_ not have
1813 * handled this reply yet and we race with it to grab
1814 * exp_uncommitted_replies_lock before removing the reply from
1815 * exp_uncommitted_replies. Note that if we lose the race and the
1816 * reply has already been removed, list_del_init() is a noop.
1818 * If we see rs_committed set, we know the commit callback is handling,
1819 * or has handled this reply since store reordering might allow us to
1820 * see rs_committed set out of sequence. But since this is done
1821 * holding rs_lock, we can be sure it has all completed once we hold
1822 * rs_lock, which we do right next.
1824 if (!rs->rs_committed) {
1825 cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
1826 cfs_list_del_init(&rs->rs_obd_list);
1827 cfs_spin_unlock(&exp->exp_uncommitted_replies_lock);
1830 cfs_spin_lock(&rs->rs_lock);
1832 been_handled = rs->rs_handled;
1835 nlocks = rs->rs_nlocks; /* atomic "steal", but */
1836 rs->rs_nlocks = 0; /* locks still on rs_locks! */
1838 if (nlocks == 0 && !been_handled) {
1839 /* If we see this, we should already have seen the warning
1840 * in mds_steal_ack_locks() */
1841 CWARN("All locks stolen from rs %p x"LPD64".t"LPD64
1844 rs->rs_xid, rs->rs_transno, rs->rs_opc,
1845 libcfs_nid2str(exp->exp_connection->c_peer.nid));
1848 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
1849 cfs_spin_unlock(&rs->rs_lock);
1851 if (!been_handled && rs->rs_on_net) {
1852 LNetMDUnlink(rs->rs_md_h);
1853 /* Ignore return code; we're racing with
1857 while (nlocks-- > 0)
1858 ldlm_lock_decref(&rs->rs_locks[nlocks],
1859 rs->rs_modes[nlocks]);
1861 cfs_spin_lock(&rs->rs_lock);
1864 rs->rs_scheduled = 0;
1866 if (!rs->rs_on_net) {
1868 cfs_spin_unlock(&rs->rs_lock);
1870 class_export_put (exp);
1871 rs->rs_export = NULL;
1872 ptlrpc_rs_decref (rs);
1873 if (cfs_atomic_dec_and_test(&svc->srv_n_difficult_replies) &&
1874 svc->srv_is_stopping)
1875 cfs_waitq_broadcast(&svc->srv_waitq);
1879 /* still on the net; callback will schedule */
1880 cfs_spin_unlock(&rs->rs_lock);
1887 * Check whether given service has a reply available for processing
1890 * \param svc a ptlrpc service
1891 * \retval 0 no replies processed
1892 * \retval 1 one reply processed
1895 ptlrpc_server_handle_reply(struct ptlrpc_service *svc)
1897 struct ptlrpc_reply_state *rs = NULL;
1900 cfs_spin_lock(&svc->srv_rs_lock);
1901 if (!cfs_list_empty(&svc->srv_reply_queue)) {
1902 rs = cfs_list_entry(svc->srv_reply_queue.prev,
1903 struct ptlrpc_reply_state,
1905 cfs_list_del_init(&rs->rs_list);
1907 cfs_spin_unlock(&svc->srv_rs_lock);
1909 ptlrpc_handle_rs(rs);
1913 /* FIXME make use of timeout later */
1915 liblustre_check_services (void *arg)
1917 int did_something = 0;
1919 cfs_list_t *tmp, *nxt;
1922 /* I'm relying on being single threaded, not to have to lock
1923 * ptlrpc_all_services etc */
1924 cfs_list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
1925 struct ptlrpc_service *svc =
1926 cfs_list_entry (tmp, struct ptlrpc_service, srv_list);
1928 if (svc->srv_threads_running != 0) /* I've recursed */
1931 /* service threads can block for bulk, so this limits us
1932 * (arbitrarily) to recursing 1 stack frame per service.
1933 * Note that the problem with recursion is that we have to
1934 * unwind completely before our caller can resume. */
1936 svc->srv_threads_running++;
1939 rc = ptlrpc_server_handle_req_in(svc);
1940 rc |= ptlrpc_server_handle_reply(svc);
1941 rc |= ptlrpc_at_check_timed(svc);
1942 rc |= ptlrpc_server_handle_request(svc, NULL);
1943 rc |= (ptlrpc_server_post_idle_rqbds(svc) > 0);
1944 did_something |= rc;
1947 svc->srv_threads_running--;
1950 RETURN(did_something);
1952 #define ptlrpc_stop_all_threads(s) do {} while (0)
1954 #else /* __KERNEL__ */
1957 ptlrpc_check_rqbd_pool(struct ptlrpc_service *svc)
1959 int avail = svc->srv_nrqbd_receiving;
1960 int low_water = test_req_buffer_pressure ? 0 :
1961 svc->srv_nbuf_per_group/2;
1963 /* NB I'm not locking; just looking. */
1965 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
1966 * allowed the request history to grow out of control. We could put a
1967 * sanity check on that here and cull some history if we need the
1970 if (avail <= low_water)
1971 ptlrpc_grow_req_bufs(svc);
1974 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQBUF_AVAIL_CNTR,
1979 ptlrpc_retry_rqbds(void *arg)
1981 struct ptlrpc_service *svc = (struct ptlrpc_service *)arg;
1983 svc->srv_rqbd_timeout = 0;
1984 return (-ETIMEDOUT);
1988 ptlrpc_threads_enough(struct ptlrpc_service *svc)
1990 return svc->srv_n_active_reqs <
1991 svc->srv_threads_running - 1 - (svc->srv_hpreq_handler != NULL);
1995 * allowed to create more threads
1996 * user can call it w/o any lock but need to hold ptlrpc_service::srv_lock to
1997 * get reliable result
2000 ptlrpc_threads_increasable(struct ptlrpc_service *svc)
2002 return svc->srv_threads_running +
2003 svc->srv_threads_starting < svc->srv_threads_max;
2007 * too many requests and allowed to create more threads
2010 ptlrpc_threads_need_create(struct ptlrpc_service *svc)
2012 return !ptlrpc_threads_enough(svc) && ptlrpc_threads_increasable(svc);
2016 ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2018 return thread_is_stopping(thread) ||
2019 thread->t_svc->srv_is_stopping;
2023 ptlrpc_rqbd_pending(struct ptlrpc_service *svc)
2025 return !cfs_list_empty(&svc->srv_idle_rqbds) &&
2026 svc->srv_rqbd_timeout == 0;
2030 ptlrpc_at_check(struct ptlrpc_service *svc)
2032 return svc->srv_at_check;
2036 * requests wait on preprocessing
2037 * user can call it w/o any lock but need to hold ptlrpc_service::srv_lock to
2038 * get reliable result
2041 ptlrpc_server_request_waiting(struct ptlrpc_service *svc)
2043 return !cfs_list_empty(&svc->srv_req_in_queue);
2046 static __attribute__((__noinline__)) int
2047 ptlrpc_wait_event(struct ptlrpc_service *svc,
2048 struct ptlrpc_thread *thread)
2050 /* Don't exit while there are replies to be handled */
2051 struct l_wait_info lwi = LWI_TIMEOUT(svc->srv_rqbd_timeout,
2052 ptlrpc_retry_rqbds, svc);
2054 lc_watchdog_disable(thread->t_watchdog);
2058 l_wait_event_exclusive_head(svc->srv_waitq,
2059 ptlrpc_thread_stopping(thread) ||
2060 ptlrpc_server_request_waiting(svc) ||
2061 ptlrpc_server_request_pending(svc, 0) ||
2062 ptlrpc_rqbd_pending(svc) ||
2063 ptlrpc_at_check(svc), &lwi);
2065 if (ptlrpc_thread_stopping(thread))
2068 lc_watchdog_touch(thread->t_watchdog, CFS_GET_TIMEOUT(svc));
2074 * Main thread body for service threads.
2075 * Waits in a loop waiting for new requests to process to appear.
2076 * Every time an incoming requests is added to its queue, a waitq
2077 * is woken up and one of the threads will handle it.
2079 static int ptlrpc_main(void *arg)
2081 struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
2082 struct ptlrpc_service *svc = data->svc;
2083 struct ptlrpc_thread *thread = data->thread;
2084 struct ptlrpc_reply_state *rs;
2085 #ifdef WITH_GROUP_INFO
2086 cfs_group_info_t *ginfo = NULL;
2089 int counter = 0, rc = 0;
2092 thread->t_pid = cfs_curproc_pid();
2093 cfs_daemonize_ctxt(data->name);
2095 #if defined(HAVE_NODE_TO_CPUMASK) && defined(CONFIG_NUMA)
2096 /* we need to do this before any per-thread allocation is done so that
2097 * we get the per-thread allocations on local node. bug 7342 */
2098 if (svc->srv_cpu_affinity) {
2101 for (cpu = 0, num_cpu = 0; cpu < cfs_num_possible_cpus();
2103 if (!cfs_cpu_online(cpu))
2105 if (num_cpu == thread->t_id % cfs_num_online_cpus())
2109 cfs_set_cpus_allowed(cfs_current(),
2110 node_to_cpumask(cpu_to_node(cpu)));
2114 #ifdef WITH_GROUP_INFO
2115 ginfo = cfs_groups_alloc(0);
2121 cfs_set_current_groups(ginfo);
2122 cfs_put_group_info(ginfo);
2125 if (svc->srv_init != NULL) {
2126 rc = svc->srv_init(thread);
2131 rc = lu_context_init(&env.le_ctx,
2132 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2136 thread->t_env = &env;
2137 env.le_ctx.lc_thread = thread;
2138 env.le_ctx.lc_cookie = 0x6;
2140 /* Alloc reply state structure for this one */
2141 OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2147 cfs_spin_lock(&svc->srv_lock);
2149 LASSERT(thread_is_starting(thread));
2150 thread_clear_flags(thread, SVC_STARTING);
2151 svc->srv_threads_starting--;
2153 /* SVC_STOPPING may already be set here if someone else is trying
2154 * to stop the service while this new thread has been dynamically
2155 * forked. We still set SVC_RUNNING to let our creator know that
2156 * we are now running, however we will exit as soon as possible */
2157 thread_add_flags(thread, SVC_RUNNING);
2158 svc->srv_threads_running++;
2159 cfs_spin_unlock(&svc->srv_lock);
2162 * wake up our creator. Note: @data is invalid after this point,
2163 * because it's allocated on ptlrpc_start_thread() stack.
2165 cfs_waitq_signal(&thread->t_ctl_waitq);
2167 thread->t_watchdog = lc_watchdog_add(CFS_GET_TIMEOUT(svc), NULL, NULL);
2169 cfs_spin_lock(&svc->srv_rs_lock);
2170 cfs_list_add(&rs->rs_list, &svc->srv_free_rs_list);
2171 cfs_waitq_signal(&svc->srv_free_rs_waitq);
2172 cfs_spin_unlock(&svc->srv_rs_lock);
2174 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2175 svc->srv_threads_running);
2177 /* XXX maintain a list of all managed devices: insert here */
2178 while (!ptlrpc_thread_stopping(thread)) {
2179 if (ptlrpc_wait_event(svc, thread))
2182 ptlrpc_check_rqbd_pool(svc);
2184 if (ptlrpc_threads_need_create(svc)) {
2185 /* Ignore return code - we tried... */
2186 ptlrpc_start_thread(svc);
2189 /* Process all incoming reqs before handling any */
2190 if (ptlrpc_server_request_waiting(svc)) {
2191 ptlrpc_server_handle_req_in(svc);
2192 /* but limit ourselves in case of flood */
2193 if (counter++ < 100)
2198 if (ptlrpc_at_check(svc))
2199 ptlrpc_at_check_timed(svc);
2201 if (ptlrpc_server_request_pending(svc, 0)) {
2202 lu_context_enter(&env.le_ctx);
2203 ptlrpc_server_handle_request(svc, thread);
2204 lu_context_exit(&env.le_ctx);
2207 if (ptlrpc_rqbd_pending(svc) &&
2208 ptlrpc_server_post_idle_rqbds(svc) < 0) {
2209 /* I just failed to repost request buffers.
2210 * Wait for a timeout (unless something else
2211 * happens) before I try again */
2212 svc->srv_rqbd_timeout = cfs_time_seconds(1)/10;
2213 CDEBUG(D_RPCTRACE,"Posted buffers: %d\n",
2214 svc->srv_nrqbd_receiving);
2218 lc_watchdog_delete(thread->t_watchdog);
2219 thread->t_watchdog = NULL;
2223 * deconstruct service specific state created by ptlrpc_start_thread()
2225 if (svc->srv_done != NULL)
2226 svc->srv_done(thread);
2228 lu_context_fini(&env.le_ctx);
2230 CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n",
2231 thread, thread->t_pid, thread->t_id, rc);
2233 cfs_spin_lock(&svc->srv_lock);
2234 if (thread_test_and_clear_flags(thread, SVC_STARTING))
2235 svc->srv_threads_starting--;
2237 if (thread_test_and_clear_flags(thread, SVC_RUNNING))
2238 /* must know immediately */
2239 svc->srv_threads_running--;
2242 thread_add_flags(thread, SVC_STOPPED);
2244 cfs_waitq_signal(&thread->t_ctl_waitq);
2245 cfs_spin_unlock(&svc->srv_lock);
2250 struct ptlrpc_hr_args {
2253 struct ptlrpc_hr_service *hrs;
2256 static int hrt_dont_sleep(struct ptlrpc_hr_thread *t,
2257 cfs_list_t *replies)
2261 cfs_spin_lock(&t->hrt_lock);
2262 cfs_list_splice_init(&t->hrt_queue, replies);
2263 result = cfs_test_bit(HRT_STOPPING, &t->hrt_flags) ||
2264 !cfs_list_empty(replies);
2265 cfs_spin_unlock(&t->hrt_lock);
2270 * Main body of "handle reply" function.
2271 * It processes acked reply states
2273 static int ptlrpc_hr_main(void *arg)
2275 struct ptlrpc_hr_args * hr_args = arg;
2276 struct ptlrpc_hr_service *hr = hr_args->hrs;
2277 struct ptlrpc_hr_thread *t = &hr->hr_threads[hr_args->thread_index];
2278 char threadname[20];
2279 CFS_LIST_HEAD(replies);
2281 snprintf(threadname, sizeof(threadname),
2282 "ptlrpc_hr_%d", hr_args->thread_index);
2284 cfs_daemonize_ctxt(threadname);
2285 #if defined(CONFIG_NUMA) && defined(HAVE_NODE_TO_CPUMASK)
2286 cfs_set_cpus_allowed(cfs_current(),
2287 node_to_cpumask(cpu_to_node(hr_args->cpu_index)));
2289 cfs_set_bit(HRT_RUNNING, &t->hrt_flags);
2290 cfs_waitq_signal(&t->hrt_wait);
2292 while (!cfs_test_bit(HRT_STOPPING, &t->hrt_flags)) {
2294 l_wait_condition(t->hrt_wait, hrt_dont_sleep(t, &replies));
2295 while (!cfs_list_empty(&replies)) {
2296 struct ptlrpc_reply_state *rs;
2298 rs = cfs_list_entry(replies.prev,
2299 struct ptlrpc_reply_state,
2301 cfs_list_del_init(&rs->rs_list);
2302 ptlrpc_handle_rs(rs);
2306 cfs_clear_bit(HRT_RUNNING, &t->hrt_flags);
2307 cfs_complete(&t->hrt_completion);
2312 static int ptlrpc_start_hr_thread(struct ptlrpc_hr_service *hr, int n, int cpu)
2314 struct ptlrpc_hr_thread *t = &hr->hr_threads[n];
2315 struct ptlrpc_hr_args args;
2319 args.thread_index = n;
2320 args.cpu_index = cpu;
2323 rc = cfs_create_thread(ptlrpc_hr_main, (void*)&args, CFS_DAEMON_FLAGS);
2325 cfs_complete(&t->hrt_completion);
2328 l_wait_condition(t->hrt_wait, cfs_test_bit(HRT_RUNNING, &t->hrt_flags));
2334 static void ptlrpc_stop_hr_thread(struct ptlrpc_hr_thread *t)
2338 cfs_set_bit(HRT_STOPPING, &t->hrt_flags);
2339 cfs_waitq_signal(&t->hrt_wait);
2340 cfs_wait_for_completion(&t->hrt_completion);
2345 static void ptlrpc_stop_hr_threads(struct ptlrpc_hr_service *hrs)
2350 for (n = 0; n < hrs->hr_n_threads; n++)
2351 ptlrpc_stop_hr_thread(&hrs->hr_threads[n]);
2356 static int ptlrpc_start_hr_threads(struct ptlrpc_hr_service *hr)
2359 int n, cpu, threads_started = 0;
2362 LASSERT(hr != NULL);
2363 LASSERT(hr->hr_n_threads > 0);
2365 for (n = 0, cpu = 0; n < hr->hr_n_threads; n++) {
2366 #if defined(CONFIG_SMP) && defined(HAVE_NODE_TO_CPUMASK)
2367 while(!cfs_cpu_online(cpu)) {
2369 if (cpu >= cfs_num_possible_cpus())
2373 rc = ptlrpc_start_hr_thread(hr, n, cpu);
2379 if (threads_started == 0) {
2380 CERROR("No reply handling threads started\n");
2383 if (threads_started < hr->hr_n_threads) {
2384 CWARN("Started only %d reply handling threads from %d\n",
2385 threads_started, hr->hr_n_threads);
2386 hr->hr_n_threads = threads_started;
2391 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
2392 struct ptlrpc_thread *thread)
2394 struct l_wait_info lwi = { 0 };
2397 CDEBUG(D_RPCTRACE, "Stopping thread [ %p : %u ]\n",
2398 thread, thread->t_pid);
2400 cfs_spin_lock(&svc->srv_lock);
2401 /* let the thread know that we would like it to stop asap */
2402 thread_add_flags(thread, SVC_STOPPING);
2403 cfs_spin_unlock(&svc->srv_lock);
2405 cfs_waitq_broadcast(&svc->srv_waitq);
2406 l_wait_event(thread->t_ctl_waitq,
2407 thread_is_stopped(thread), &lwi);
2409 cfs_spin_lock(&svc->srv_lock);
2410 cfs_list_del(&thread->t_link);
2411 cfs_spin_unlock(&svc->srv_lock);
2413 OBD_FREE_PTR(thread);
2418 * Stops all threads of a particular service \a svc
2420 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2422 struct ptlrpc_thread *thread;
2425 cfs_spin_lock(&svc->srv_lock);
2426 while (!cfs_list_empty(&svc->srv_threads)) {
2427 thread = cfs_list_entry(svc->srv_threads.next,
2428 struct ptlrpc_thread, t_link);
2430 cfs_spin_unlock(&svc->srv_lock);
2431 ptlrpc_stop_thread(svc, thread);
2432 cfs_spin_lock(&svc->srv_lock);
2435 cfs_spin_unlock(&svc->srv_lock);
2439 int ptlrpc_start_threads(struct ptlrpc_service *svc)
2444 /* We require 2 threads min - see note in
2445 ptlrpc_server_handle_request */
2446 LASSERT(svc->srv_threads_min >= 2);
2447 for (i = 0; i < svc->srv_threads_min; i++) {
2448 rc = ptlrpc_start_thread(svc);
2449 /* We have enough threads, don't start more. b=15759 */
2450 if (rc == -EMFILE) {
2455 CERROR("cannot start %s thread #%d: rc %d\n",
2456 svc->srv_thread_name, i, rc);
2457 ptlrpc_stop_all_threads(svc);
2464 int ptlrpc_start_thread(struct ptlrpc_service *svc)
2466 struct l_wait_info lwi = { 0 };
2467 struct ptlrpc_svc_data d;
2468 struct ptlrpc_thread *thread;
2473 CDEBUG(D_RPCTRACE, "%s started %d min %d max %d running %d\n",
2474 svc->srv_name, svc->srv_threads_running, svc->srv_threads_min,
2475 svc->srv_threads_max, svc->srv_threads_running);
2477 if (unlikely(svc->srv_is_stopping))
2480 if (!ptlrpc_threads_increasable(svc) ||
2481 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
2482 svc->srv_threads_running == svc->srv_threads_min - 1))
2485 OBD_ALLOC_PTR(thread);
2488 cfs_waitq_init(&thread->t_ctl_waitq);
2490 cfs_spin_lock(&svc->srv_lock);
2491 if (!ptlrpc_threads_increasable(svc)) {
2492 cfs_spin_unlock(&svc->srv_lock);
2493 OBD_FREE_PTR(thread);
2497 svc->srv_threads_starting++;
2498 thread->t_id = svc->srv_threads_next_id++;
2499 thread_add_flags(thread, SVC_STARTING);
2500 thread->t_svc = svc;
2502 cfs_list_add(&thread->t_link, &svc->srv_threads);
2503 cfs_spin_unlock(&svc->srv_lock);
2505 sprintf(name, "%s_%02d", svc->srv_thread_name, thread->t_id);
2510 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", name);
2512 /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
2513 * just drop the VM and FILES in cfs_daemonize_ctxt() right away.
2515 rc = cfs_create_thread(ptlrpc_main, &d, CFS_DAEMON_FLAGS);
2517 CERROR("cannot start thread '%s': rc %d\n", name, rc);
2519 cfs_spin_lock(&svc->srv_lock);
2520 cfs_list_del(&thread->t_link);
2521 --svc->srv_threads_starting;
2522 cfs_spin_unlock(&svc->srv_lock);
2524 OBD_FREE(thread, sizeof(*thread));
2527 l_wait_event(thread->t_ctl_waitq,
2528 thread_is_running(thread) || thread_is_stopped(thread),
2531 rc = thread_is_stopped(thread) ? thread->t_id : 0;
2536 int ptlrpc_hr_init(void)
2539 int n_cpus = cfs_num_online_cpus();
2540 struct ptlrpc_hr_service *hr;
2545 LASSERT(ptlrpc_hr == NULL);
2547 size = offsetof(struct ptlrpc_hr_service, hr_threads[n_cpus]);
2548 OBD_ALLOC(hr, size);
2551 for (i = 0; i < n_cpus; i++) {
2552 struct ptlrpc_hr_thread *t = &hr->hr_threads[i];
2554 cfs_spin_lock_init(&t->hrt_lock);
2555 cfs_waitq_init(&t->hrt_wait);
2556 CFS_INIT_LIST_HEAD(&t->hrt_queue);
2557 cfs_init_completion(&t->hrt_completion);
2559 hr->hr_n_threads = n_cpus;
2563 rc = ptlrpc_start_hr_threads(hr);
2565 OBD_FREE(hr, hr->hr_size);
2571 void ptlrpc_hr_fini(void)
2573 if (ptlrpc_hr != NULL) {
2574 ptlrpc_stop_hr_threads(ptlrpc_hr);
2575 OBD_FREE(ptlrpc_hr, ptlrpc_hr->hr_size);
2580 #endif /* __KERNEL__ */
2583 * Wait until all already scheduled replies are processed.
2585 static void ptlrpc_wait_replies(struct ptlrpc_service *svc)
2589 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
2591 rc = l_wait_event(svc->srv_waitq, cfs_atomic_read(&svc-> \
2592 srv_n_difficult_replies) == 0,
2596 CWARN("Unexpectedly long timeout %p\n", svc);
2600 int ptlrpc_unregister_service(struct ptlrpc_service *service)
2603 struct l_wait_info lwi;
2605 struct ptlrpc_reply_state *rs, *t;
2606 struct ptlrpc_at_array *array = &service->srv_at_array;
2609 service->srv_is_stopping = 1;
2610 cfs_timer_disarm(&service->srv_at_timer);
2612 ptlrpc_stop_all_threads(service);
2613 LASSERT(cfs_list_empty(&service->srv_threads));
2615 cfs_spin_lock (&ptlrpc_all_services_lock);
2616 cfs_list_del_init (&service->srv_list);
2617 cfs_spin_unlock (&ptlrpc_all_services_lock);
2619 ptlrpc_lprocfs_unregister_service(service);
2621 /* All history will be culled when the next request buffer is
2623 service->srv_max_history_rqbds = 0;
2625 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
2627 rc = LNetClearLazyPortal(service->srv_req_portal);
2630 /* Unlink all the request buffers. This forces a 'final' event with
2631 * its 'unlink' flag set for each posted rqbd */
2632 cfs_list_for_each(tmp, &service->srv_active_rqbds) {
2633 struct ptlrpc_request_buffer_desc *rqbd =
2634 cfs_list_entry(tmp, struct ptlrpc_request_buffer_desc,
2637 rc = LNetMDUnlink(rqbd->rqbd_md_h);
2638 LASSERT (rc == 0 || rc == -ENOENT);
2641 /* Wait for the network to release any buffers it's currently
2644 cfs_spin_lock(&service->srv_lock);
2645 rc = service->srv_nrqbd_receiving;
2646 cfs_spin_unlock(&service->srv_lock);
2651 /* Network access will complete in finite time but the HUGE
2652 * timeout lets us CWARN for visibility of sluggish NALs */
2653 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2654 cfs_time_seconds(1), NULL, NULL);
2655 rc = l_wait_event(service->srv_waitq,
2656 service->srv_nrqbd_receiving == 0,
2658 if (rc == -ETIMEDOUT)
2659 CWARN("Service %s waiting for request buffers\n",
2663 /* schedule all outstanding replies to terminate them */
2664 cfs_spin_lock(&service->srv_rs_lock);
2665 while (!cfs_list_empty(&service->srv_active_replies)) {
2666 struct ptlrpc_reply_state *rs =
2667 cfs_list_entry(service->srv_active_replies.next,
2668 struct ptlrpc_reply_state, rs_list);
2669 cfs_spin_lock(&rs->rs_lock);
2670 ptlrpc_schedule_difficult_reply(rs);
2671 cfs_spin_unlock(&rs->rs_lock);
2673 cfs_spin_unlock(&service->srv_rs_lock);
2675 /* purge the request queue. NB No new replies (rqbds all unlinked)
2676 * and no service threads, so I'm the only thread noodling the
2677 * request queue now */
2678 while (!cfs_list_empty(&service->srv_req_in_queue)) {
2679 struct ptlrpc_request *req =
2680 cfs_list_entry(service->srv_req_in_queue.next,
2681 struct ptlrpc_request,
2684 cfs_list_del(&req->rq_list);
2685 service->srv_n_queued_reqs--;
2686 service->srv_n_active_reqs++;
2687 ptlrpc_server_finish_request(service, req);
2689 while (ptlrpc_server_request_pending(service, 1)) {
2690 struct ptlrpc_request *req;
2692 req = ptlrpc_server_request_get(service, 1);
2693 cfs_list_del(&req->rq_list);
2694 service->srv_n_active_reqs++;
2695 ptlrpc_server_finish_request(service, req);
2697 LASSERT(service->srv_n_queued_reqs == 0);
2698 LASSERT(service->srv_n_active_reqs == 0);
2699 LASSERT(service->srv_n_history_rqbds == 0);
2700 LASSERT(cfs_list_empty(&service->srv_active_rqbds));
2702 /* Now free all the request buffers since nothing references them
2704 while (!cfs_list_empty(&service->srv_idle_rqbds)) {
2705 struct ptlrpc_request_buffer_desc *rqbd =
2706 cfs_list_entry(service->srv_idle_rqbds.next,
2707 struct ptlrpc_request_buffer_desc,
2710 ptlrpc_free_rqbd(rqbd);
2713 ptlrpc_wait_replies(service);
2715 cfs_list_for_each_entry_safe(rs, t, &service->srv_free_rs_list,
2717 cfs_list_del(&rs->rs_list);
2718 OBD_FREE_LARGE(rs, service->srv_max_reply_size);
2721 /* In case somebody rearmed this in the meantime */
2722 cfs_timer_disarm(&service->srv_at_timer);
2724 if (array->paa_reqs_array != NULL) {
2725 OBD_FREE(array->paa_reqs_array,
2726 sizeof(cfs_list_t) * array->paa_size);
2727 array->paa_reqs_array = NULL;
2730 if (array->paa_reqs_count != NULL) {
2731 OBD_FREE(array->paa_reqs_count,
2732 sizeof(__u32) * array->paa_size);
2733 array->paa_reqs_count= NULL;
2736 OBD_FREE_PTR(service);
2741 * Returns 0 if the service is healthy.
2743 * Right now, it just checks to make sure that requests aren't languishing
2744 * in the queue. We'll use this health check to govern whether a node needs
2745 * to be shot, so it's intentionally non-aggressive. */
2746 int ptlrpc_service_health_check(struct ptlrpc_service *svc)
2748 struct ptlrpc_request *request;
2749 struct timeval right_now;
2755 cfs_gettimeofday(&right_now);
2757 cfs_spin_lock(&svc->srv_rq_lock);
2758 if (!ptlrpc_server_request_pending(svc, 1)) {
2759 cfs_spin_unlock(&svc->srv_rq_lock);
2763 /* How long has the next entry been waiting? */
2764 if (cfs_list_empty(&svc->srv_request_queue))
2765 request = cfs_list_entry(svc->srv_request_hpq.next,
2766 struct ptlrpc_request, rq_list);
2768 request = cfs_list_entry(svc->srv_request_queue.next,
2769 struct ptlrpc_request, rq_list);
2770 timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
2771 cfs_spin_unlock(&svc->srv_rq_lock);
2773 if ((timediff / ONE_MILLION) > (AT_OFF ? obd_timeout * 3/2 :
2775 CERROR("%s: unhealthy - request has been waiting %lds\n",
2776 svc->srv_name, timediff / ONE_MILLION);