4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, Whamcloud, Inc.
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);
67 static void ptlrpc_hpreq_fini(struct ptlrpc_request *req);
69 static CFS_LIST_HEAD(ptlrpc_all_services);
70 cfs_spinlock_t ptlrpc_all_services_lock;
72 struct ptlrpc_request_buffer_desc *
73 ptlrpc_alloc_rqbd (struct ptlrpc_service *svc)
75 struct ptlrpc_request_buffer_desc *rqbd;
81 rqbd->rqbd_service = svc;
82 rqbd->rqbd_refcount = 0;
83 rqbd->rqbd_cbid.cbid_fn = request_in_callback;
84 rqbd->rqbd_cbid.cbid_arg = rqbd;
85 CFS_INIT_LIST_HEAD(&rqbd->rqbd_reqs);
86 OBD_ALLOC_LARGE(rqbd->rqbd_buffer, svc->srv_buf_size);
88 if (rqbd->rqbd_buffer == NULL) {
93 cfs_spin_lock(&svc->srv_lock);
94 cfs_list_add(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
96 cfs_spin_unlock(&svc->srv_lock);
102 ptlrpc_free_rqbd (struct ptlrpc_request_buffer_desc *rqbd)
104 struct ptlrpc_service *svc = rqbd->rqbd_service;
106 LASSERT (rqbd->rqbd_refcount == 0);
107 LASSERT (cfs_list_empty(&rqbd->rqbd_reqs));
109 cfs_spin_lock(&svc->srv_lock);
110 cfs_list_del(&rqbd->rqbd_list);
112 cfs_spin_unlock(&svc->srv_lock);
114 OBD_FREE_LARGE(rqbd->rqbd_buffer, svc->srv_buf_size);
119 ptlrpc_grow_req_bufs(struct ptlrpc_service *svc)
121 struct ptlrpc_request_buffer_desc *rqbd;
125 for (i = 0; i < svc->srv_nbuf_per_group; i++) {
126 /* NB: another thread might be doing this as well, we need to
127 * make sure that it wouldn't over-allocate, see LU-1212. */
128 if (svc->srv_nrqbd_receiving >= svc->srv_nbuf_per_group)
131 rqbd = ptlrpc_alloc_rqbd(svc);
134 CERROR("%s: Can't allocate request buffer\n",
140 if (ptlrpc_server_post_idle_rqbds(svc) < 0) {
147 "%s: allocate %d new %d-byte reqbufs (%d/%d left), rc = %d\n",
148 svc->srv_name, i, svc->srv_buf_size,
149 svc->srv_nrqbd_receiving, svc->srv_nbufs, rc);
155 * Part of Rep-Ack logic.
156 * Puts a lock and its mode into reply state assotiated to request reply.
159 ptlrpc_save_lock(struct ptlrpc_request *req,
160 struct lustre_handle *lock, int mode, int no_ack)
162 struct ptlrpc_reply_state *rs = req->rq_reply_state;
166 LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
168 if (req->rq_export->exp_disconnected) {
169 ldlm_lock_decref(lock, mode);
171 idx = rs->rs_nlocks++;
172 rs->rs_locks[idx] = *lock;
173 rs->rs_modes[idx] = mode;
174 rs->rs_difficult = 1;
175 rs->rs_no_ack = !!no_ack;
181 #define HRT_RUNNING 0
182 #define HRT_STOPPING 1
184 struct ptlrpc_hr_thread {
185 cfs_spinlock_t hrt_lock;
186 unsigned long hrt_flags;
187 cfs_waitq_t hrt_wait;
188 cfs_list_t hrt_queue;
189 cfs_completion_t hrt_completion;
192 struct ptlrpc_hr_service {
196 struct ptlrpc_hr_thread hr_threads[0];
200 cfs_list_t rsb_replies;
201 struct ptlrpc_service *rsb_svc;
202 unsigned int rsb_n_replies;
206 * A pointer to per-node reply handling service.
208 static struct ptlrpc_hr_service *ptlrpc_hr = NULL;
211 * maximum mumber of replies scheduled in one batch
213 #define MAX_SCHEDULED 256
216 * Initialize a reply batch.
220 static void rs_batch_init(struct rs_batch *b)
222 memset(b, 0, sizeof *b);
223 CFS_INIT_LIST_HEAD(&b->rsb_replies);
227 * Choose an hr thread to dispatch requests to.
229 static unsigned int get_hr_thread_index(struct ptlrpc_hr_service *hr)
233 /* Concurrent modification of hr_index w/o any spinlock
234 protection is harmless as long as the result fits
235 [0..(hr_n_threads-1)] range and each thread gets near equal
238 hr->hr_index = (idx >= hr->hr_n_threads - 1) ? 0 : idx + 1;
243 * Dispatch all replies accumulated in the batch to one from
244 * dedicated reply handling threads.
248 static void rs_batch_dispatch(struct rs_batch *b)
250 if (b->rsb_n_replies != 0) {
251 struct ptlrpc_hr_service *hr = ptlrpc_hr;
254 idx = get_hr_thread_index(hr);
256 cfs_spin_lock(&hr->hr_threads[idx].hrt_lock);
257 cfs_list_splice_init(&b->rsb_replies,
258 &hr->hr_threads[idx].hrt_queue);
259 cfs_spin_unlock(&hr->hr_threads[idx].hrt_lock);
260 cfs_waitq_signal(&hr->hr_threads[idx].hrt_wait);
261 b->rsb_n_replies = 0;
266 * Add a reply to a batch.
267 * Add one reply object to a batch, schedule batched replies if overload.
272 static void rs_batch_add(struct rs_batch *b, struct ptlrpc_reply_state *rs)
274 struct ptlrpc_service *svc = rs->rs_service;
276 if (svc != b->rsb_svc || b->rsb_n_replies >= MAX_SCHEDULED) {
277 if (b->rsb_svc != NULL) {
278 rs_batch_dispatch(b);
279 cfs_spin_unlock(&b->rsb_svc->srv_rs_lock);
281 cfs_spin_lock(&svc->srv_rs_lock);
284 cfs_spin_lock(&rs->rs_lock);
285 rs->rs_scheduled_ever = 1;
286 if (rs->rs_scheduled == 0) {
287 cfs_list_move(&rs->rs_list, &b->rsb_replies);
288 rs->rs_scheduled = 1;
291 rs->rs_committed = 1;
292 cfs_spin_unlock(&rs->rs_lock);
296 * Reply batch finalization.
297 * Dispatch remaining replies from the batch
298 * and release remaining spinlock.
302 static void rs_batch_fini(struct rs_batch *b)
304 if (b->rsb_svc != 0) {
305 rs_batch_dispatch(b);
306 cfs_spin_unlock(&b->rsb_svc->srv_rs_lock);
310 #define DECLARE_RS_BATCH(b) struct rs_batch b
312 #else /* __KERNEL__ */
314 #define rs_batch_init(b) do{}while(0)
315 #define rs_batch_fini(b) do{}while(0)
316 #define rs_batch_add(b, r) ptlrpc_schedule_difficult_reply(r)
317 #define DECLARE_RS_BATCH(b)
319 #endif /* __KERNEL__ */
322 * Put reply state into a queue for processing because we received
323 * ACK from the client
325 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
328 struct ptlrpc_hr_service *hr = ptlrpc_hr;
332 LASSERT(cfs_list_empty(&rs->rs_list));
334 idx = get_hr_thread_index(hr);
335 cfs_spin_lock(&hr->hr_threads[idx].hrt_lock);
336 cfs_list_add_tail(&rs->rs_list, &hr->hr_threads[idx].hrt_queue);
337 cfs_spin_unlock(&hr->hr_threads[idx].hrt_lock);
338 cfs_waitq_signal(&hr->hr_threads[idx].hrt_wait);
341 cfs_list_add_tail(&rs->rs_list, &rs->rs_service->srv_reply_queue);
346 ptlrpc_schedule_difficult_reply (struct ptlrpc_reply_state *rs)
350 LASSERT_SPIN_LOCKED(&rs->rs_service->srv_rs_lock);
351 LASSERT_SPIN_LOCKED(&rs->rs_lock);
352 LASSERT (rs->rs_difficult);
353 rs->rs_scheduled_ever = 1; /* flag any notification attempt */
355 if (rs->rs_scheduled) { /* being set up or already notified */
360 rs->rs_scheduled = 1;
361 cfs_list_del_init(&rs->rs_list);
362 ptlrpc_dispatch_difficult_reply(rs);
366 void ptlrpc_commit_replies(struct obd_export *exp)
368 struct ptlrpc_reply_state *rs, *nxt;
369 DECLARE_RS_BATCH(batch);
372 rs_batch_init(&batch);
373 /* Find any replies that have been committed and get their service
374 * to attend to complete them. */
376 /* CAVEAT EMPTOR: spinlock ordering!!! */
377 cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
378 cfs_list_for_each_entry_safe(rs, nxt, &exp->exp_uncommitted_replies,
380 LASSERT (rs->rs_difficult);
381 /* VBR: per-export last_committed */
382 LASSERT(rs->rs_export);
383 if (rs->rs_transno <= exp->exp_last_committed) {
384 cfs_list_del_init(&rs->rs_obd_list);
385 rs_batch_add(&batch, rs);
388 cfs_spin_unlock(&exp->exp_uncommitted_replies_lock);
389 rs_batch_fini(&batch);
394 ptlrpc_server_post_idle_rqbds (struct ptlrpc_service *svc)
396 struct ptlrpc_request_buffer_desc *rqbd;
401 cfs_spin_lock(&svc->srv_lock);
403 if (cfs_list_empty (&svc->srv_idle_rqbds)) {
404 cfs_spin_unlock(&svc->srv_lock);
408 rqbd = cfs_list_entry(svc->srv_idle_rqbds.next,
409 struct ptlrpc_request_buffer_desc,
411 cfs_list_del (&rqbd->rqbd_list);
413 /* assume we will post successfully */
414 svc->srv_nrqbd_receiving++;
415 cfs_list_add (&rqbd->rqbd_list, &svc->srv_active_rqbds);
417 cfs_spin_unlock(&svc->srv_lock);
419 rc = ptlrpc_register_rqbd(rqbd);
426 cfs_spin_lock(&svc->srv_lock);
428 svc->srv_nrqbd_receiving--;
429 cfs_list_del(&rqbd->rqbd_list);
430 cfs_list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
432 /* Don't complain if no request buffers are posted right now; LNET
433 * won't drop requests because we set the portal lazy! */
435 cfs_spin_unlock(&svc->srv_lock);
441 * Start a service with parameters from struct ptlrpc_service_conf \a c
442 * as opposed to directly calling ptlrpc_init_svc with tons of arguments.
444 struct ptlrpc_service *ptlrpc_init_svc_conf(struct ptlrpc_service_conf *c,
445 svc_handler_t h, char *name,
446 struct proc_dir_entry *proc_entry,
447 svc_req_printfn_t prntfn,
450 return ptlrpc_init_svc(c->psc_nbufs, c->psc_bufsize,
451 c->psc_max_req_size, c->psc_max_reply_size,
452 c->psc_req_portal, c->psc_rep_portal,
453 c->psc_watchdog_factor,
455 prntfn, c->psc_min_threads, c->psc_max_threads,
456 threadname, c->psc_ctx_tags, NULL);
458 EXPORT_SYMBOL(ptlrpc_init_svc_conf);
460 static void ptlrpc_at_timer(unsigned long castmeharder)
462 struct ptlrpc_service *svc = (struct ptlrpc_service *)castmeharder;
463 svc->srv_at_check = 1;
464 svc->srv_at_checktime = cfs_time_current();
465 cfs_waitq_signal(&svc->srv_waitq);
469 * Initialize service on a given portal.
470 * This includes starting serving threads , allocating and posting rqbds and
472 * \a nbufs is how many buffers to post
473 * \a bufsize is buffer size to post
474 * \a max_req_size - maximum request size to be accepted for this service
475 * \a max_reply_size maximum reply size this service can ever send
476 * \a req_portal - portal to listed for requests on
477 * \a rep_portal - portal of where to send replies to
478 * \a watchdog_factor soft watchdog timeout multiplifier to print stuck service traces.
479 * \a handler - function to process every new request
480 * \a name - service name
481 * \a proc_entry - entry in the /proc tree for sttistics reporting
482 * \a min_threads \a max_threads - min/max number of service threads to start.
483 * \a threadname should be 11 characters or less - 3 will be added on
484 * \a hp_handler - function to determine priority of the request, also called
485 * on every new request.
487 struct ptlrpc_service *
488 ptlrpc_init_svc(int nbufs, int bufsize, int max_req_size, int max_reply_size,
489 int req_portal, int rep_portal, int watchdog_factor,
490 svc_handler_t handler, char *name,
491 cfs_proc_dir_entry_t *proc_entry,
492 svc_req_printfn_t svcreq_printfn,
493 int min_threads, int max_threads,
494 char *threadname, __u32 ctx_tags,
495 svc_hpreq_handler_t hp_handler)
498 struct ptlrpc_at_array *array;
499 struct ptlrpc_service *service;
500 unsigned int size, index;
504 LASSERT (bufsize >= max_req_size + SPTLRPC_MAX_PAYLOAD);
505 LASSERT (ctx_tags != 0);
507 OBD_ALLOC_PTR(service);
511 /* First initialise enough for early teardown */
513 service->srv_name = name;
514 cfs_spin_lock_init(&service->srv_lock);
515 cfs_spin_lock_init(&service->srv_rq_lock);
516 cfs_spin_lock_init(&service->srv_rs_lock);
517 CFS_INIT_LIST_HEAD(&service->srv_threads);
518 cfs_waitq_init(&service->srv_waitq);
520 service->srv_nbuf_per_group = test_req_buffer_pressure ? 1 : nbufs;
521 service->srv_max_req_size = max_req_size + SPTLRPC_MAX_PAYLOAD;
522 service->srv_buf_size = bufsize;
523 service->srv_rep_portal = rep_portal;
524 service->srv_req_portal = req_portal;
525 service->srv_watchdog_factor = watchdog_factor;
526 service->srv_handler = handler;
527 service->srv_req_printfn = svcreq_printfn;
528 service->srv_request_seq = 1; /* valid seq #s start at 1 */
529 service->srv_request_max_cull_seq = 0;
530 service->srv_threads_min = min_threads;
531 service->srv_threads_max = max_threads;
532 service->srv_thread_name = threadname;
533 service->srv_ctx_tags = ctx_tags;
534 service->srv_hpreq_handler = hp_handler;
535 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
536 service->srv_hpreq_count = 0;
537 service->srv_n_active_hpreq = 0;
539 rc = LNetSetLazyPortal(service->srv_req_portal);
542 CFS_INIT_LIST_HEAD(&service->srv_request_queue);
543 CFS_INIT_LIST_HEAD(&service->srv_request_hpq);
544 CFS_INIT_LIST_HEAD(&service->srv_idle_rqbds);
545 CFS_INIT_LIST_HEAD(&service->srv_active_rqbds);
546 CFS_INIT_LIST_HEAD(&service->srv_history_rqbds);
547 CFS_INIT_LIST_HEAD(&service->srv_request_history);
548 CFS_INIT_LIST_HEAD(&service->srv_active_replies);
550 CFS_INIT_LIST_HEAD(&service->srv_reply_queue);
552 CFS_INIT_LIST_HEAD(&service->srv_free_rs_list);
553 cfs_waitq_init(&service->srv_free_rs_waitq);
554 cfs_atomic_set(&service->srv_n_difficult_replies, 0);
556 cfs_spin_lock_init(&service->srv_at_lock);
557 CFS_INIT_LIST_HEAD(&service->srv_req_in_queue);
559 array = &service->srv_at_array;
560 size = at_est2timeout(at_max);
561 array->paa_size = size;
562 array->paa_count = 0;
563 array->paa_deadline = -1;
565 /* allocate memory for srv_at_array (ptlrpc_at_array) */
566 OBD_ALLOC(array->paa_reqs_array, sizeof(cfs_list_t) * size);
567 if (array->paa_reqs_array == NULL)
570 for (index = 0; index < size; index++)
571 CFS_INIT_LIST_HEAD(&array->paa_reqs_array[index]);
573 OBD_ALLOC(array->paa_reqs_count, sizeof(__u32) * size);
574 if (array->paa_reqs_count == NULL)
577 cfs_timer_init(&service->srv_at_timer, ptlrpc_at_timer, service);
578 /* At SOW, service time should be quick; 10s seems generous. If client
579 timeout is less than this, we'll be sending an early reply. */
580 at_init(&service->srv_at_estimate, 10, 0);
582 cfs_spin_lock (&ptlrpc_all_services_lock);
583 cfs_list_add (&service->srv_list, &ptlrpc_all_services);
584 cfs_spin_unlock (&ptlrpc_all_services_lock);
586 /* Now allocate the request buffers */
587 rc = ptlrpc_grow_req_bufs(service);
588 /* We shouldn't be under memory pressure at startup, so
589 * fail if we can't post all our buffers at this time. */
593 /* Now allocate pool of reply buffers */
594 /* Increase max reply size to next power of two */
595 service->srv_max_reply_size = 1;
596 while (service->srv_max_reply_size <
597 max_reply_size + SPTLRPC_MAX_PAYLOAD)
598 service->srv_max_reply_size <<= 1;
600 if (proc_entry != NULL)
601 ptlrpc_lprocfs_register_service(proc_entry, service);
603 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
604 service->srv_name, service->srv_req_portal);
608 ptlrpc_unregister_service(service);
613 * to actually free the request, must be called without holding svc_lock.
614 * note it's caller's responsibility to unlink req->rq_list.
616 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
618 LASSERT(cfs_atomic_read(&req->rq_refcount) == 0);
619 LASSERT(cfs_list_empty(&req->rq_timed_list));
621 /* DEBUG_REQ() assumes the reply state of a request with a valid
622 * ref will not be destroyed until that reference is dropped. */
623 ptlrpc_req_drop_rs(req);
625 sptlrpc_svc_ctx_decref(req);
627 if (req != &req->rq_rqbd->rqbd_req) {
628 /* NB request buffers use an embedded
629 * req if the incoming req unlinked the
630 * MD; this isn't one of them! */
631 OBD_FREE(req, sizeof(*req));
636 * drop a reference count of the request. if it reaches 0, we either
637 * put it into history list, or free it immediately.
639 void ptlrpc_server_drop_request(struct ptlrpc_request *req)
641 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
642 struct ptlrpc_service *svc = rqbd->rqbd_service;
647 if (!cfs_atomic_dec_and_test(&req->rq_refcount))
650 cfs_spin_lock(&svc->srv_at_lock);
651 if (req->rq_at_linked) {
652 struct ptlrpc_at_array *array = &svc->srv_at_array;
653 __u32 index = req->rq_at_index;
655 LASSERT(!cfs_list_empty(&req->rq_timed_list));
656 cfs_list_del_init(&req->rq_timed_list);
657 cfs_spin_lock(&req->rq_lock);
658 req->rq_at_linked = 0;
659 cfs_spin_unlock(&req->rq_lock);
660 array->paa_reqs_count[index]--;
663 LASSERT(cfs_list_empty(&req->rq_timed_list));
664 cfs_spin_unlock(&svc->srv_at_lock);
666 /* finalize request */
667 if (req->rq_export) {
668 class_export_put(req->rq_export);
669 req->rq_export = NULL;
672 cfs_spin_lock(&svc->srv_lock);
674 cfs_list_add(&req->rq_list, &rqbd->rqbd_reqs);
676 refcount = --(rqbd->rqbd_refcount);
678 /* request buffer is now idle: add to history */
679 cfs_list_del(&rqbd->rqbd_list);
680 cfs_list_add_tail(&rqbd->rqbd_list, &svc->srv_history_rqbds);
681 svc->srv_n_history_rqbds++;
683 /* cull some history?
684 * I expect only about 1 or 2 rqbds need to be recycled here */
685 while (svc->srv_n_history_rqbds > svc->srv_max_history_rqbds) {
686 rqbd = cfs_list_entry(svc->srv_history_rqbds.next,
687 struct ptlrpc_request_buffer_desc,
690 cfs_list_del(&rqbd->rqbd_list);
691 svc->srv_n_history_rqbds--;
693 /* remove rqbd's reqs from svc's req history while
694 * I've got the service lock */
695 cfs_list_for_each(tmp, &rqbd->rqbd_reqs) {
696 req = cfs_list_entry(tmp, struct ptlrpc_request,
698 /* Track the highest culled req seq */
699 if (req->rq_history_seq >
700 svc->srv_request_max_cull_seq)
701 svc->srv_request_max_cull_seq =
703 cfs_list_del(&req->rq_history_list);
706 cfs_spin_unlock(&svc->srv_lock);
708 cfs_list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
709 req = cfs_list_entry(rqbd->rqbd_reqs.next,
710 struct ptlrpc_request,
712 cfs_list_del(&req->rq_list);
713 ptlrpc_server_free_request(req);
716 cfs_spin_lock(&svc->srv_lock);
718 * now all reqs including the embedded req has been
719 * disposed, schedule request buffer for re-use.
721 LASSERT(cfs_atomic_read(&rqbd->rqbd_req.rq_refcount) ==
723 cfs_list_add_tail(&rqbd->rqbd_list,
724 &svc->srv_idle_rqbds);
727 cfs_spin_unlock(&svc->srv_lock);
728 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
729 /* If we are low on memory, we are not interested in history */
730 cfs_list_del(&req->rq_list);
731 cfs_list_del_init(&req->rq_history_list);
732 cfs_spin_unlock(&svc->srv_lock);
734 ptlrpc_server_free_request(req);
736 cfs_spin_unlock(&svc->srv_lock);
741 * to finish a request: stop sending more early replies, and release
742 * the request. should be called after we finished handling the request.
744 static void ptlrpc_server_finish_request(struct ptlrpc_service *svc,
745 struct ptlrpc_request *req)
747 ptlrpc_hpreq_fini(req);
749 cfs_spin_lock(&svc->srv_rq_lock);
750 svc->srv_n_active_reqs--;
752 svc->srv_n_active_hpreq--;
753 cfs_spin_unlock(&svc->srv_rq_lock);
755 ptlrpc_server_drop_request(req);
759 * This function makes sure dead exports are evicted in a timely manner.
760 * This function is only called when some export receives a message (i.e.,
761 * the network is up.)
763 static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
765 struct obd_export *oldest_exp;
766 time_t oldest_time, new_time;
772 /* Compensate for slow machines, etc, by faking our request time
773 into the future. Although this can break the strict time-ordering
774 of the list, we can be really lazy here - we don't have to evict
775 at the exact right moment. Eventually, all silent exports
776 will make it to the top of the list. */
778 /* Do not pay attention on 1sec or smaller renewals. */
779 new_time = cfs_time_current_sec() + extra_delay;
780 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
783 exp->exp_last_request_time = new_time;
784 CDEBUG(D_HA, "updating export %s at "CFS_TIME_T" exp %p\n",
785 exp->exp_client_uuid.uuid,
786 exp->exp_last_request_time, exp);
788 /* exports may get disconnected from the chain even though the
789 export has references, so we must keep the spin lock while
790 manipulating the lists */
791 cfs_spin_lock(&exp->exp_obd->obd_dev_lock);
793 if (cfs_list_empty(&exp->exp_obd_chain_timed)) {
794 /* this one is not timed */
795 cfs_spin_unlock(&exp->exp_obd->obd_dev_lock);
799 cfs_list_move_tail(&exp->exp_obd_chain_timed,
800 &exp->exp_obd->obd_exports_timed);
802 oldest_exp = cfs_list_entry(exp->exp_obd->obd_exports_timed.next,
803 struct obd_export, exp_obd_chain_timed);
804 oldest_time = oldest_exp->exp_last_request_time;
805 cfs_spin_unlock(&exp->exp_obd->obd_dev_lock);
807 if (exp->exp_obd->obd_recovering) {
808 /* be nice to everyone during recovery */
813 /* Note - racing to start/reset the obd_eviction timer is safe */
814 if (exp->exp_obd->obd_eviction_timer == 0) {
815 /* Check if the oldest entry is expired. */
816 if (cfs_time_current_sec() > (oldest_time + PING_EVICT_TIMEOUT +
818 /* We need a second timer, in case the net was down and
819 * it just came back. Since the pinger may skip every
820 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
821 * we better wait for 3. */
822 exp->exp_obd->obd_eviction_timer =
823 cfs_time_current_sec() + 3 * PING_INTERVAL;
824 CDEBUG(D_HA, "%s: Think about evicting %s from "CFS_TIME_T"\n",
825 exp->exp_obd->obd_name,
826 obd_export_nid2str(oldest_exp), oldest_time);
829 if (cfs_time_current_sec() >
830 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
831 /* The evictor won't evict anyone who we've heard from
832 * recently, so we don't have to check before we start
834 if (!ping_evictor_wake(exp))
835 exp->exp_obd->obd_eviction_timer = 0;
843 * Sanity check request \a req.
844 * Return 0 if all is ok, error code otherwise.
846 static int ptlrpc_check_req(struct ptlrpc_request *req)
850 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
851 req->rq_export->exp_conn_cnt)) {
852 DEBUG_REQ(D_ERROR, req,
853 "DROPPING req from old connection %d < %d",
854 lustre_msg_get_conn_cnt(req->rq_reqmsg),
855 req->rq_export->exp_conn_cnt);
858 if (unlikely(req->rq_export->exp_obd &&
859 req->rq_export->exp_obd->obd_fail)) {
860 /* Failing over, don't handle any more reqs, send
861 error response instead. */
862 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
863 req, req->rq_export->exp_obd->obd_name);
865 } else if (lustre_msg_get_flags(req->rq_reqmsg) &
866 (MSG_REPLAY | MSG_REQ_REPLAY_DONE) &&
867 !(req->rq_export->exp_obd->obd_recovering)) {
868 DEBUG_REQ(D_ERROR, req,
869 "Invalid replay without recovery");
870 class_fail_export(req->rq_export);
872 } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 &&
873 !(req->rq_export->exp_obd->obd_recovering)) {
874 DEBUG_REQ(D_ERROR, req, "Invalid req with transno "
875 LPU64" without recovery",
876 lustre_msg_get_transno(req->rq_reqmsg));
877 class_fail_export(req->rq_export);
881 if (unlikely(rc < 0)) {
888 static void ptlrpc_at_set_timer(struct ptlrpc_service *svc)
890 struct ptlrpc_at_array *array = &svc->srv_at_array;
893 cfs_spin_lock(&svc->srv_at_lock);
894 if (array->paa_count == 0) {
895 cfs_timer_disarm(&svc->srv_at_timer);
896 cfs_spin_unlock(&svc->srv_at_lock);
900 /* Set timer for closest deadline */
901 next = (__s32)(array->paa_deadline - cfs_time_current_sec() -
904 ptlrpc_at_timer((unsigned long)svc);
906 cfs_timer_arm(&svc->srv_at_timer, cfs_time_shift(next));
907 cfs_spin_unlock(&svc->srv_at_lock);
908 CDEBUG(D_INFO, "armed %s at %+ds\n", svc->srv_name, next);
911 /* Add rpc to early reply check list */
912 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
914 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
915 struct ptlrpc_request *rq = NULL;
916 struct ptlrpc_at_array *array = &svc->srv_at_array;
923 if (req->rq_no_reply)
926 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
929 cfs_spin_lock(&svc->srv_at_lock);
930 LASSERT(cfs_list_empty(&req->rq_timed_list));
932 index = (unsigned long)req->rq_deadline % array->paa_size;
933 if (array->paa_reqs_count[index] > 0) {
934 /* latest rpcs will have the latest deadlines in the list,
935 * so search backward. */
936 cfs_list_for_each_entry_reverse(rq,
937 &array->paa_reqs_array[index],
939 if (req->rq_deadline >= rq->rq_deadline) {
940 cfs_list_add(&req->rq_timed_list,
947 /* Add the request at the head of the list */
948 if (cfs_list_empty(&req->rq_timed_list))
949 cfs_list_add(&req->rq_timed_list,
950 &array->paa_reqs_array[index]);
952 cfs_spin_lock(&req->rq_lock);
953 req->rq_at_linked = 1;
954 cfs_spin_unlock(&req->rq_lock);
955 req->rq_at_index = index;
956 array->paa_reqs_count[index]++;
958 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
959 array->paa_deadline = req->rq_deadline;
962 cfs_spin_unlock(&svc->srv_at_lock);
965 ptlrpc_at_set_timer(svc);
970 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
972 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
973 struct ptlrpc_request *reqcopy;
974 struct lustre_msg *reqmsg;
975 cfs_duration_t olddl = req->rq_deadline - cfs_time_current_sec();
980 /* deadline is when the client expects us to reply, margin is the
981 difference between clients' and servers' expectations */
982 DEBUG_REQ(D_ADAPTTO, req,
983 "%ssending early reply (deadline %+lds, margin %+lds) for "
984 "%d+%d", AT_OFF ? "AT off - not " : "",
985 olddl, olddl - at_get(&svc->srv_at_estimate),
986 at_get(&svc->srv_at_estimate), at_extra);
992 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), "
993 "not sending early reply. Consider increasing "
994 "at_early_margin (%d)?", olddl, at_early_margin);
996 /* Return an error so we're not re-added to the timed list. */
1000 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
1001 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
1002 "but no AT support");
1006 if (req->rq_export &&
1007 lustre_msg_get_flags(req->rq_reqmsg) &
1008 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
1009 /* During recovery, we don't want to send too many early
1010 * replies, but on the other hand we want to make sure the
1011 * client has enough time to resend if the rpc is lost. So
1012 * during the recovery period send at least 4 early replies,
1013 * spacing them every at_extra if we can. at_estimate should
1014 * always equal this fixed value during recovery. */
1015 at_measured(&svc->srv_at_estimate, min(at_extra,
1016 req->rq_export->exp_obd->obd_recovery_timeout / 4));
1018 /* Fake our processing time into the future to ask the clients
1019 * for some extra amount of time */
1020 at_measured(&svc->srv_at_estimate, at_extra +
1021 cfs_time_current_sec() -
1022 req->rq_arrival_time.tv_sec);
1024 /* Check to see if we've actually increased the deadline -
1025 * we may be past adaptive_max */
1026 if (req->rq_deadline >= req->rq_arrival_time.tv_sec +
1027 at_get(&svc->srv_at_estimate)) {
1028 DEBUG_REQ(D_WARNING, req, "Couldn't add any time "
1029 "(%ld/%ld), not sending early reply\n",
1030 olddl, req->rq_arrival_time.tv_sec +
1031 at_get(&svc->srv_at_estimate) -
1032 cfs_time_current_sec());
1036 newdl = cfs_time_current_sec() + at_get(&svc->srv_at_estimate);
1038 OBD_ALLOC(reqcopy, sizeof *reqcopy);
1039 if (reqcopy == NULL)
1041 OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
1043 OBD_FREE(reqcopy, sizeof *reqcopy);
1048 reqcopy->rq_reply_state = NULL;
1049 reqcopy->rq_rep_swab_mask = 0;
1050 reqcopy->rq_pack_bulk = 0;
1051 reqcopy->rq_pack_udesc = 0;
1052 reqcopy->rq_packed_final = 0;
1053 sptlrpc_svc_ctx_addref(reqcopy);
1054 /* We only need the reqmsg for the magic */
1055 reqcopy->rq_reqmsg = reqmsg;
1056 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1058 LASSERT(cfs_atomic_read(&req->rq_refcount));
1059 /** if it is last refcount then early reply isn't needed */
1060 if (cfs_atomic_read(&req->rq_refcount) == 1) {
1061 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
1062 "abort sending early reply\n");
1063 GOTO(out, rc = -EINVAL);
1066 /* Connection ref */
1067 reqcopy->rq_export = class_conn2export(
1068 lustre_msg_get_handle(reqcopy->rq_reqmsg));
1069 if (reqcopy->rq_export == NULL)
1070 GOTO(out, rc = -ENODEV);
1073 class_export_rpc_get(reqcopy->rq_export);
1074 if (reqcopy->rq_export->exp_obd &&
1075 reqcopy->rq_export->exp_obd->obd_fail)
1076 GOTO(out_put, rc = -ENODEV);
1078 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
1082 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
1085 /* Adjust our own deadline to what we told the client */
1086 req->rq_deadline = newdl;
1087 req->rq_early_count++; /* number sent, server side */
1089 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
1092 /* Free the (early) reply state from lustre_pack_reply.
1093 (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
1094 ptlrpc_req_drop_rs(reqcopy);
1097 class_export_rpc_put(reqcopy->rq_export);
1098 class_export_put(reqcopy->rq_export);
1100 sptlrpc_svc_ctx_decref(reqcopy);
1101 OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
1102 OBD_FREE(reqcopy, sizeof *reqcopy);
1106 /* Send early replies to everybody expiring within at_early_margin
1107 asking for at_extra time */
1108 static int ptlrpc_at_check_timed(struct ptlrpc_service *svc)
1110 struct ptlrpc_request *rq, *n;
1111 cfs_list_t work_list;
1112 struct ptlrpc_at_array *array = &svc->srv_at_array;
1115 time_t now = cfs_time_current_sec();
1116 cfs_duration_t delay;
1117 int first, counter = 0;
1120 cfs_spin_lock(&svc->srv_at_lock);
1121 if (svc->srv_at_check == 0) {
1122 cfs_spin_unlock(&svc->srv_at_lock);
1125 delay = cfs_time_sub(cfs_time_current(), svc->srv_at_checktime);
1126 svc->srv_at_check = 0;
1128 if (array->paa_count == 0) {
1129 cfs_spin_unlock(&svc->srv_at_lock);
1133 /* The timer went off, but maybe the nearest rpc already completed. */
1134 first = array->paa_deadline - now;
1135 if (first > at_early_margin) {
1136 /* We've still got plenty of time. Reset the timer. */
1137 cfs_spin_unlock(&svc->srv_at_lock);
1138 ptlrpc_at_set_timer(svc);
1142 /* We're close to a timeout, and we don't know how much longer the
1143 server will take. Send early replies to everyone expiring soon. */
1144 CFS_INIT_LIST_HEAD(&work_list);
1146 index = (unsigned long)array->paa_deadline % array->paa_size;
1147 count = array->paa_count;
1149 count -= array->paa_reqs_count[index];
1150 cfs_list_for_each_entry_safe(rq, n,
1151 &array->paa_reqs_array[index],
1153 if (rq->rq_deadline <= now + at_early_margin) {
1154 cfs_list_del_init(&rq->rq_timed_list);
1156 * ptlrpc_server_drop_request() may drop
1157 * refcount to 0 already. Let's check this and
1158 * don't add entry to work_list
1160 if (likely(cfs_atomic_inc_not_zero(&rq->rq_refcount)))
1161 cfs_list_add(&rq->rq_timed_list, &work_list);
1163 array->paa_reqs_count[index]--;
1165 cfs_spin_lock(&rq->rq_lock);
1166 rq->rq_at_linked = 0;
1167 cfs_spin_unlock(&rq->rq_lock);
1171 /* update the earliest deadline */
1172 if (deadline == -1 || rq->rq_deadline < deadline)
1173 deadline = rq->rq_deadline;
1178 if (++index >= array->paa_size)
1181 array->paa_deadline = deadline;
1182 cfs_spin_unlock(&svc->srv_at_lock);
1184 /* we have a new earliest deadline, restart the timer */
1185 ptlrpc_at_set_timer(svc);
1187 CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
1188 "replies\n", first, at_extra, counter);
1190 /* We're already past request deadlines before we even get a
1191 chance to send early replies */
1192 LCONSOLE_WARN("%s: This server is not able to keep up with "
1193 "request traffic (cpu-bound).\n", svc->srv_name);
1194 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, "
1195 "delay="CFS_DURATION_T"(jiff)\n",
1196 counter, svc->srv_n_queued_reqs, svc->srv_n_active_reqs,
1197 at_get(&svc->srv_at_estimate), delay);
1200 /* we took additional refcount so entries can't be deleted from list, no
1201 * locking is needed */
1202 while (!cfs_list_empty(&work_list)) {
1203 rq = cfs_list_entry(work_list.next, struct ptlrpc_request,
1205 cfs_list_del_init(&rq->rq_timed_list);
1207 if (ptlrpc_at_send_early_reply(rq) == 0)
1208 ptlrpc_at_add_timed(rq);
1210 ptlrpc_server_drop_request(rq);
1217 * Put the request to the export list if the request may become
1218 * a high priority one.
1220 static int ptlrpc_hpreq_init(struct ptlrpc_service *svc,
1221 struct ptlrpc_request *req)
1226 if (svc->srv_hpreq_handler) {
1227 rc = svc->srv_hpreq_handler(req);
1231 if (req->rq_export && req->rq_ops) {
1232 /* Perform request specific check. We should do this check
1233 * before the request is added into exp_hp_rpcs list otherwise
1234 * it may hit swab race at LU-1044. */
1235 if (req->rq_ops->hpreq_check)
1236 rc = req->rq_ops->hpreq_check(req);
1238 cfs_spin_lock_bh(&req->rq_export->exp_rpc_lock);
1239 cfs_list_add(&req->rq_exp_list,
1240 &req->rq_export->exp_hp_rpcs);
1241 cfs_spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1247 /** Remove the request from the export list. */
1248 static void ptlrpc_hpreq_fini(struct ptlrpc_request *req)
1251 if (req->rq_export && req->rq_ops) {
1252 /* refresh lock timeout again so that client has more
1253 * room to send lock cancel RPC. */
1254 if (req->rq_ops->hpreq_fini)
1255 req->rq_ops->hpreq_fini(req);
1257 cfs_spin_lock_bh(&req->rq_export->exp_rpc_lock);
1258 cfs_list_del_init(&req->rq_exp_list);
1259 cfs_spin_unlock_bh(&req->rq_export->exp_rpc_lock);
1265 * Make the request a high priority one.
1267 * All the high priority requests are queued in a separate FIFO
1268 * ptlrpc_service::srv_request_hpq list which is parallel to
1269 * ptlrpc_service::srv_request_queue list but has a higher priority
1272 * \see ptlrpc_server_handle_request().
1274 static void ptlrpc_hpreq_reorder_nolock(struct ptlrpc_service *svc,
1275 struct ptlrpc_request *req)
1278 LASSERT(svc != NULL);
1279 cfs_spin_lock(&req->rq_lock);
1280 if (req->rq_hp == 0) {
1281 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1283 /* Add to the high priority queue. */
1284 cfs_list_move_tail(&req->rq_list, &svc->srv_request_hpq);
1286 if (opc != OBD_PING)
1287 DEBUG_REQ(D_RPCTRACE, req, "high priority req");
1289 cfs_spin_unlock(&req->rq_lock);
1294 * \see ptlrpc_hpreq_reorder_nolock
1296 void ptlrpc_hpreq_reorder(struct ptlrpc_request *req)
1298 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
1301 cfs_spin_lock(&svc->srv_rq_lock);
1302 /* It may happen that the request is already taken for the processing
1303 * but still in the export list, or the request is not in the request
1304 * queue but in the export list already, do not add it into the
1306 if (!cfs_list_empty(&req->rq_list))
1307 ptlrpc_hpreq_reorder_nolock(svc, req);
1308 cfs_spin_unlock(&svc->srv_rq_lock);
1312 /** Check if the request is a high priority one. */
1313 static int ptlrpc_server_hpreq_check(struct ptlrpc_service *svc,
1314 struct ptlrpc_request *req)
1318 /* Check by request opc. */
1319 if (OBD_PING == lustre_msg_get_opc(req->rq_reqmsg))
1322 RETURN(ptlrpc_hpreq_init(svc, req));
1325 /** Check if a request is a high priority one. */
1326 static int ptlrpc_server_request_add(struct ptlrpc_service *svc,
1327 struct ptlrpc_request *req)
1332 rc = ptlrpc_server_hpreq_check(svc, req);
1336 cfs_spin_lock(&svc->srv_rq_lock);
1339 ptlrpc_hpreq_reorder_nolock(svc, req);
1341 cfs_list_add_tail(&req->rq_list,
1342 &svc->srv_request_queue);
1344 cfs_spin_unlock(&svc->srv_rq_lock);
1350 * Allow to handle high priority request
1351 * User can call it w/o any lock but need to hold ptlrpc_service::srv_rq_lock
1352 * to get reliable result
1354 static int ptlrpc_server_allow_high(struct ptlrpc_service *svc, int force)
1359 if (svc->srv_n_active_reqs >= svc->srv_threads_running - 1)
1362 return cfs_list_empty(&svc->srv_request_queue) ||
1363 svc->srv_hpreq_count < svc->srv_hpreq_ratio;
1366 static int ptlrpc_server_high_pending(struct ptlrpc_service *svc, int force)
1368 return ptlrpc_server_allow_high(svc, force) &&
1369 !cfs_list_empty(&svc->srv_request_hpq);
1373 * Only allow normal priority requests on a service that has a high-priority
1374 * queue if forced (i.e. cleanup), if there are other high priority requests
1375 * already being processed (i.e. those threads can service more high-priority
1376 * requests), or if there are enough idle threads that a later thread can do
1377 * a high priority request.
1378 * User can call it w/o any lock but need to hold ptlrpc_service::srv_rq_lock
1379 * to get reliable result
1381 static int ptlrpc_server_allow_normal(struct ptlrpc_service *svc, int force)
1384 if (1) /* always allow to handle normal request for liblustre */
1388 svc->srv_n_active_reqs < svc->srv_threads_running - 2)
1391 if (svc->srv_n_active_reqs >= svc->srv_threads_running - 1)
1394 return svc->srv_n_active_hpreq > 0 || svc->srv_hpreq_handler == NULL;
1397 static int ptlrpc_server_normal_pending(struct ptlrpc_service *svc, int force)
1399 return ptlrpc_server_allow_normal(svc, force) &&
1400 !cfs_list_empty(&svc->srv_request_queue);
1404 * Returns true if there are requests available in incoming
1405 * request queue for processing and it is allowed to fetch them.
1406 * User can call it w/o any lock but need to hold ptlrpc_service::srv_rq_lock
1407 * to get reliable result
1408 * \see ptlrpc_server_allow_normal
1409 * \see ptlrpc_server_allow high
1412 ptlrpc_server_request_pending(struct ptlrpc_service *svc, int force)
1414 return ptlrpc_server_high_pending(svc, force) ||
1415 ptlrpc_server_normal_pending(svc, force);
1419 * Fetch a request for processing from queue of unprocessed requests.
1420 * Favors high-priority requests.
1421 * Returns a pointer to fetched request.
1423 static struct ptlrpc_request *
1424 ptlrpc_server_request_get(struct ptlrpc_service *svc, int force)
1426 struct ptlrpc_request *req;
1429 if (ptlrpc_server_high_pending(svc, force)) {
1430 req = cfs_list_entry(svc->srv_request_hpq.next,
1431 struct ptlrpc_request, rq_list);
1432 svc->srv_hpreq_count++;
1437 if (ptlrpc_server_normal_pending(svc, force)) {
1438 req = cfs_list_entry(svc->srv_request_queue.next,
1439 struct ptlrpc_request, rq_list);
1440 svc->srv_hpreq_count = 0;
1447 * Handle freshly incoming reqs, add to timed early reply list,
1448 * pass on to regular request queue.
1449 * All incoming requests pass through here before getting into
1450 * ptlrpc_server_handle_req later on.
1453 ptlrpc_server_handle_req_in(struct ptlrpc_service *svc)
1455 struct ptlrpc_request *req;
1462 cfs_spin_lock(&svc->srv_lock);
1463 if (cfs_list_empty(&svc->srv_req_in_queue)) {
1464 cfs_spin_unlock(&svc->srv_lock);
1468 req = cfs_list_entry(svc->srv_req_in_queue.next,
1469 struct ptlrpc_request, rq_list);
1470 cfs_list_del_init (&req->rq_list);
1471 svc->srv_n_queued_reqs--;
1472 /* Consider this still a "queued" request as far as stats are
1474 cfs_spin_unlock(&svc->srv_lock);
1476 /* go through security check/transform */
1477 rc = sptlrpc_svc_unwrap_request(req);
1481 case SECSVC_COMPLETE:
1482 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1491 * for null-flavored rpc, msg has been unpacked by sptlrpc, although
1492 * redo it wouldn't be harmful.
1494 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
1495 rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
1497 CERROR("error unpacking request: ptl %d from %s "
1498 "x"LPU64"\n", svc->srv_req_portal,
1499 libcfs_id2str(req->rq_peer), req->rq_xid);
1504 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1506 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
1507 LPU64"\n", svc->srv_req_portal,
1508 libcfs_id2str(req->rq_peer), req->rq_xid);
1512 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
1513 lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
1514 CERROR("drop incoming rpc opc %u, x"LPU64"\n",
1515 cfs_fail_val, req->rq_xid);
1520 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1521 CERROR("wrong packet type received (type=%u) from %s\n",
1522 lustre_msg_get_type(req->rq_reqmsg),
1523 libcfs_id2str(req->rq_peer));
1527 switch(lustre_msg_get_opc(req->rq_reqmsg)) {
1530 req->rq_bulk_write = 1;
1534 case MGS_CONFIG_READ:
1535 req->rq_bulk_read = 1;
1539 CDEBUG(D_RPCTRACE, "got req x"LPU64"\n", req->rq_xid);
1541 req->rq_export = class_conn2export(
1542 lustre_msg_get_handle(req->rq_reqmsg));
1543 if (req->rq_export) {
1544 rc = ptlrpc_check_req(req);
1546 rc = sptlrpc_target_export_check(req->rq_export, req);
1548 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
1549 "illegal security flavor,");
1554 ptlrpc_update_export_timer(req->rq_export, 0);
1557 /* req_in handling should/must be fast */
1558 if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5)
1559 DEBUG_REQ(D_WARNING, req, "Slow req_in handling "CFS_DURATION_T"s",
1560 cfs_time_sub(cfs_time_current_sec(),
1561 req->rq_arrival_time.tv_sec));
1563 /* Set rpc server deadline and add it to the timed list */
1564 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1565 MSGHDR_AT_SUPPORT) ?
1566 /* The max time the client expects us to take */
1567 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1568 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1569 if (unlikely(deadline == 0)) {
1570 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1574 ptlrpc_at_add_timed(req);
1576 /* Move it over to the request processing queue */
1577 rc = ptlrpc_server_request_add(svc, req);
1579 ptlrpc_hpreq_fini(req);
1582 cfs_waitq_signal(&svc->srv_waitq);
1586 cfs_spin_lock(&svc->srv_rq_lock);
1587 svc->srv_n_active_reqs++;
1588 cfs_spin_unlock(&svc->srv_rq_lock);
1589 ptlrpc_server_finish_request(svc, req);
1595 * Main incoming request handling logic.
1596 * Calls handler function from service to do actual processing.
1599 ptlrpc_server_handle_request(struct ptlrpc_service *svc,
1600 struct ptlrpc_thread *thread)
1602 struct obd_export *export = NULL;
1603 struct ptlrpc_request *request;
1604 struct timeval work_start;
1605 struct timeval work_end;
1613 cfs_spin_lock(&svc->srv_rq_lock);
1615 /* !@%$# liblustre only has 1 thread */
1616 if (cfs_atomic_read(&svc->srv_n_difficult_replies) != 0) {
1617 cfs_spin_unlock(&svc->srv_rq_lock);
1621 request = ptlrpc_server_request_get(svc, 0);
1622 if (request == NULL) {
1623 cfs_spin_unlock(&svc->srv_rq_lock);
1627 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
1628 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
1629 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
1630 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
1632 if (unlikely(fail_opc)) {
1633 if (request->rq_export && request->rq_ops) {
1634 cfs_spin_unlock(&svc->srv_rq_lock);
1635 OBD_FAIL_TIMEOUT(fail_opc, 4);
1636 cfs_spin_lock(&svc->srv_rq_lock);
1637 request = ptlrpc_server_request_get(svc, 0);
1638 if (request == NULL) {
1639 cfs_spin_unlock(&svc->srv_rq_lock);
1645 cfs_list_del_init(&request->rq_list);
1646 svc->srv_n_active_reqs++;
1648 svc->srv_n_active_hpreq++;
1650 cfs_spin_unlock(&svc->srv_rq_lock);
1652 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
1654 if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
1655 libcfs_debug_dumplog();
1657 cfs_gettimeofday(&work_start);
1658 timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
1659 if (likely(svc->srv_stats != NULL)) {
1660 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
1662 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
1663 svc->srv_n_queued_reqs);
1664 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
1665 svc->srv_n_active_reqs);
1666 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
1667 at_get(&svc->srv_at_estimate));
1670 rc = lu_context_init(&request->rq_session,
1671 LCT_SESSION|LCT_REMEMBER|LCT_NOREF);
1673 CERROR("Failure to initialize session: %d\n", rc);
1676 request->rq_session.lc_thread = thread;
1677 request->rq_session.lc_cookie = 0x5;
1678 lu_context_enter(&request->rq_session);
1680 CDEBUG(D_NET, "got req "LPU64"\n", request->rq_xid);
1682 request->rq_svc_thread = thread;
1684 request->rq_svc_thread->t_env->le_ses = &request->rq_session;
1686 if (likely(request->rq_export)) {
1687 if (unlikely(ptlrpc_check_req(request)))
1689 ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
1690 export = class_export_rpc_get(request->rq_export);
1693 /* Discard requests queued for longer than the deadline.
1694 The deadline is increased if we send an early reply. */
1695 if (cfs_time_current_sec() > request->rq_deadline) {
1696 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s"
1697 ": deadline "CFS_DURATION_T":"CFS_DURATION_T"s ago\n",
1698 libcfs_id2str(request->rq_peer),
1699 cfs_time_sub(request->rq_deadline,
1700 request->rq_arrival_time.tv_sec),
1701 cfs_time_sub(cfs_time_current_sec(),
1702 request->rq_deadline));
1703 goto put_rpc_export;
1706 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
1707 "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1708 (request->rq_export ?
1709 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1710 (request->rq_export ?
1711 cfs_atomic_read(&request->rq_export->exp_refcount) : -99),
1712 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1713 libcfs_id2str(request->rq_peer),
1714 lustre_msg_get_opc(request->rq_reqmsg));
1716 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
1717 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
1719 rc = svc->srv_handler(request);
1721 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
1725 class_export_rpc_put(export);
1727 lu_context_exit(&request->rq_session);
1728 lu_context_fini(&request->rq_session);
1730 if (unlikely(cfs_time_current_sec() > request->rq_deadline)) {
1731 DEBUG_REQ(D_WARNING, request, "Request x"LPU64" took longer "
1732 "than estimated ("CFS_DURATION_T":"CFS_DURATION_T"s);"
1733 " client may timeout.",
1734 request->rq_xid, cfs_time_sub(request->rq_deadline,
1735 request->rq_arrival_time.tv_sec),
1736 cfs_time_sub(cfs_time_current_sec(),
1737 request->rq_deadline));
1740 cfs_gettimeofday(&work_end);
1741 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1742 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
1743 "%s:%s+%d:%d:x"LPU64":%s:%d Request procesed in "
1744 "%ldus (%ldus total) trans "LPU64" rc %d/%d\n",
1746 (request->rq_export ?
1747 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1748 (request->rq_export ?
1749 cfs_atomic_read(&request->rq_export->exp_refcount) : -99),
1750 lustre_msg_get_status(request->rq_reqmsg),
1752 libcfs_id2str(request->rq_peer),
1753 lustre_msg_get_opc(request->rq_reqmsg),
1755 cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL),
1756 (request->rq_repmsg ?
1757 lustre_msg_get_transno(request->rq_repmsg) :
1758 request->rq_transno),
1760 (request->rq_repmsg ?
1761 lustre_msg_get_status(request->rq_repmsg) : -999));
1762 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
1763 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
1764 int opc = opcode_offset(op);
1765 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
1766 LASSERT(opc < LUSTRE_MAX_OPCODES);
1767 lprocfs_counter_add(svc->srv_stats,
1768 opc + EXTRA_MAX_OPCODES,
1772 if (unlikely(request->rq_early_count)) {
1773 DEBUG_REQ(D_ADAPTTO, request,
1774 "sent %d early replies before finishing in "
1776 request->rq_early_count,
1777 cfs_time_sub(work_end.tv_sec,
1778 request->rq_arrival_time.tv_sec));
1782 ptlrpc_server_finish_request(svc, request);
1788 * An internal function to process a single reply state object.
1791 ptlrpc_handle_rs (struct ptlrpc_reply_state *rs)
1793 struct ptlrpc_service *svc = rs->rs_service;
1794 struct obd_export *exp;
1799 exp = rs->rs_export;
1801 LASSERT (rs->rs_difficult);
1802 LASSERT (rs->rs_scheduled);
1803 LASSERT (cfs_list_empty(&rs->rs_list));
1805 cfs_spin_lock (&exp->exp_lock);
1806 /* Noop if removed already */
1807 cfs_list_del_init (&rs->rs_exp_list);
1808 cfs_spin_unlock (&exp->exp_lock);
1810 /* The disk commit callback holds exp_uncommitted_replies_lock while it
1811 * iterates over newly committed replies, removing them from
1812 * exp_uncommitted_replies. It then drops this lock and schedules the
1813 * replies it found for handling here.
1815 * We can avoid contention for exp_uncommitted_replies_lock between the
1816 * HRT threads and further commit callbacks by checking rs_committed
1817 * which is set in the commit callback while it holds both
1818 * rs_lock and exp_uncommitted_reples.
1820 * If we see rs_committed clear, the commit callback _may_ not have
1821 * handled this reply yet and we race with it to grab
1822 * exp_uncommitted_replies_lock before removing the reply from
1823 * exp_uncommitted_replies. Note that if we lose the race and the
1824 * reply has already been removed, list_del_init() is a noop.
1826 * If we see rs_committed set, we know the commit callback is handling,
1827 * or has handled this reply since store reordering might allow us to
1828 * see rs_committed set out of sequence. But since this is done
1829 * holding rs_lock, we can be sure it has all completed once we hold
1830 * rs_lock, which we do right next.
1832 if (!rs->rs_committed) {
1833 cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
1834 cfs_list_del_init(&rs->rs_obd_list);
1835 cfs_spin_unlock(&exp->exp_uncommitted_replies_lock);
1838 cfs_spin_lock(&rs->rs_lock);
1840 been_handled = rs->rs_handled;
1843 nlocks = rs->rs_nlocks; /* atomic "steal", but */
1844 rs->rs_nlocks = 0; /* locks still on rs_locks! */
1846 if (nlocks == 0 && !been_handled) {
1847 /* If we see this, we should already have seen the warning
1848 * in mds_steal_ack_locks() */
1849 CWARN("All locks stolen from rs %p x"LPD64".t"LPD64
1852 rs->rs_xid, rs->rs_transno, rs->rs_opc,
1853 libcfs_nid2str(exp->exp_connection->c_peer.nid));
1856 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
1857 cfs_spin_unlock(&rs->rs_lock);
1859 if (!been_handled && rs->rs_on_net) {
1860 LNetMDUnlink(rs->rs_md_h);
1861 /* Ignore return code; we're racing with
1865 while (nlocks-- > 0)
1866 ldlm_lock_decref(&rs->rs_locks[nlocks],
1867 rs->rs_modes[nlocks]);
1869 cfs_spin_lock(&rs->rs_lock);
1872 rs->rs_scheduled = 0;
1874 if (!rs->rs_on_net) {
1876 cfs_spin_unlock(&rs->rs_lock);
1878 class_export_put (exp);
1879 rs->rs_export = NULL;
1880 ptlrpc_rs_decref (rs);
1881 if (cfs_atomic_dec_and_test(&svc->srv_n_difficult_replies) &&
1882 svc->srv_is_stopping)
1883 cfs_waitq_broadcast(&svc->srv_waitq);
1887 /* still on the net; callback will schedule */
1888 cfs_spin_unlock(&rs->rs_lock);
1895 * Check whether given service has a reply available for processing
1898 * \param svc a ptlrpc service
1899 * \retval 0 no replies processed
1900 * \retval 1 one reply processed
1903 ptlrpc_server_handle_reply(struct ptlrpc_service *svc)
1905 struct ptlrpc_reply_state *rs = NULL;
1908 cfs_spin_lock(&svc->srv_rs_lock);
1909 if (!cfs_list_empty(&svc->srv_reply_queue)) {
1910 rs = cfs_list_entry(svc->srv_reply_queue.prev,
1911 struct ptlrpc_reply_state,
1913 cfs_list_del_init(&rs->rs_list);
1915 cfs_spin_unlock(&svc->srv_rs_lock);
1917 ptlrpc_handle_rs(rs);
1921 /* FIXME make use of timeout later */
1923 liblustre_check_services (void *arg)
1925 int did_something = 0;
1927 cfs_list_t *tmp, *nxt;
1930 /* I'm relying on being single threaded, not to have to lock
1931 * ptlrpc_all_services etc */
1932 cfs_list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
1933 struct ptlrpc_service *svc =
1934 cfs_list_entry (tmp, struct ptlrpc_service, srv_list);
1936 if (svc->srv_threads_running != 0) /* I've recursed */
1939 /* service threads can block for bulk, so this limits us
1940 * (arbitrarily) to recursing 1 stack frame per service.
1941 * Note that the problem with recursion is that we have to
1942 * unwind completely before our caller can resume. */
1944 svc->srv_threads_running++;
1947 rc = ptlrpc_server_handle_req_in(svc);
1948 rc |= ptlrpc_server_handle_reply(svc);
1949 rc |= ptlrpc_at_check_timed(svc);
1950 rc |= ptlrpc_server_handle_request(svc, NULL);
1951 rc |= (ptlrpc_server_post_idle_rqbds(svc) > 0);
1952 did_something |= rc;
1955 svc->srv_threads_running--;
1958 RETURN(did_something);
1960 #define ptlrpc_stop_all_threads(s) do {} while (0)
1962 #else /* __KERNEL__ */
1965 ptlrpc_check_rqbd_pool(struct ptlrpc_service *svc)
1967 int avail = svc->srv_nrqbd_receiving;
1968 int low_water = test_req_buffer_pressure ? 0 :
1969 svc->srv_nbuf_per_group / 2;
1971 /* NB I'm not locking; just looking. */
1973 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
1974 * allowed the request history to grow out of control. We could put a
1975 * sanity check on that here and cull some history if we need the
1978 if (avail <= low_water)
1979 ptlrpc_grow_req_bufs(svc);
1982 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQBUF_AVAIL_CNTR,
1987 ptlrpc_retry_rqbds(void *arg)
1989 struct ptlrpc_service *svc = (struct ptlrpc_service *)arg;
1991 svc->srv_rqbd_timeout = 0;
1992 return (-ETIMEDOUT);
1996 ptlrpc_threads_enough(struct ptlrpc_service *svc)
1998 return svc->srv_n_active_reqs <
1999 svc->srv_threads_running - 1 - (svc->srv_hpreq_handler != NULL);
2003 * allowed to create more threads
2004 * user can call it w/o any lock but need to hold ptlrpc_service::srv_lock to
2005 * get reliable result
2008 ptlrpc_threads_increasable(struct ptlrpc_service *svc)
2010 return svc->srv_threads_running +
2011 svc->srv_threads_starting < svc->srv_threads_max;
2015 * too many requests and allowed to create more threads
2018 ptlrpc_threads_need_create(struct ptlrpc_service *svc)
2020 return !ptlrpc_threads_enough(svc) && ptlrpc_threads_increasable(svc);
2024 ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
2026 return thread_is_stopping(thread) ||
2027 thread->t_svc->srv_is_stopping;
2031 ptlrpc_rqbd_pending(struct ptlrpc_service *svc)
2033 return !cfs_list_empty(&svc->srv_idle_rqbds) &&
2034 svc->srv_rqbd_timeout == 0;
2038 ptlrpc_at_check(struct ptlrpc_service *svc)
2040 return svc->srv_at_check;
2044 * requests wait on preprocessing
2045 * user can call it w/o any lock but need to hold ptlrpc_service::srv_lock to
2046 * get reliable result
2049 ptlrpc_server_request_waiting(struct ptlrpc_service *svc)
2051 return !cfs_list_empty(&svc->srv_req_in_queue);
2054 static __attribute__((__noinline__)) int
2055 ptlrpc_wait_event(struct ptlrpc_service *svc,
2056 struct ptlrpc_thread *thread)
2058 /* Don't exit while there are replies to be handled */
2059 struct l_wait_info lwi = LWI_TIMEOUT(svc->srv_rqbd_timeout,
2060 ptlrpc_retry_rqbds, svc);
2062 lc_watchdog_disable(thread->t_watchdog);
2066 l_wait_event_exclusive_head(svc->srv_waitq,
2067 ptlrpc_thread_stopping(thread) ||
2068 ptlrpc_server_request_waiting(svc) ||
2069 ptlrpc_server_request_pending(svc, 0) ||
2070 ptlrpc_rqbd_pending(svc) ||
2071 ptlrpc_at_check(svc), &lwi);
2073 if (ptlrpc_thread_stopping(thread))
2076 lc_watchdog_touch(thread->t_watchdog, CFS_GET_TIMEOUT(svc));
2082 * Main thread body for service threads.
2083 * Waits in a loop waiting for new requests to process to appear.
2084 * Every time an incoming requests is added to its queue, a waitq
2085 * is woken up and one of the threads will handle it.
2087 static int ptlrpc_main(void *arg)
2089 struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
2090 struct ptlrpc_service *svc = data->svc;
2091 struct ptlrpc_thread *thread = data->thread;
2092 struct ptlrpc_reply_state *rs;
2093 #ifdef WITH_GROUP_INFO
2094 cfs_group_info_t *ginfo = NULL;
2097 int counter = 0, rc = 0;
2100 thread->t_pid = cfs_curproc_pid();
2101 cfs_daemonize_ctxt(data->name);
2103 #if defined(HAVE_NODE_TO_CPUMASK) && defined(CONFIG_NUMA)
2104 /* we need to do this before any per-thread allocation is done so that
2105 * we get the per-thread allocations on local node. bug 7342 */
2106 if (svc->srv_cpu_affinity) {
2109 for (cpu = 0, num_cpu = 0; cpu < cfs_num_possible_cpus();
2111 if (!cpu_online(cpu))
2113 if (num_cpu == thread->t_id % cfs_num_online_cpus())
2117 cfs_set_cpus_allowed(cfs_current(),
2118 node_to_cpumask(cpu_to_node(cpu)));
2122 #ifdef WITH_GROUP_INFO
2123 ginfo = cfs_groups_alloc(0);
2129 cfs_set_current_groups(ginfo);
2130 cfs_put_group_info(ginfo);
2133 if (svc->srv_init != NULL) {
2134 rc = svc->srv_init(thread);
2145 rc = lu_context_init(&env->le_ctx,
2146 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
2150 thread->t_env = env;
2151 env->le_ctx.lc_thread = thread;
2152 env->le_ctx.lc_cookie = 0x6;
2154 /* Alloc reply state structure for this one */
2155 OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
2161 cfs_spin_lock(&svc->srv_lock);
2163 LASSERT(thread_is_starting(thread));
2164 thread_clear_flags(thread, SVC_STARTING);
2165 svc->srv_threads_starting--;
2167 /* SVC_STOPPING may already be set here if someone else is trying
2168 * to stop the service while this new thread has been dynamically
2169 * forked. We still set SVC_RUNNING to let our creator know that
2170 * we are now running, however we will exit as soon as possible */
2171 thread_add_flags(thread, SVC_RUNNING);
2172 svc->srv_threads_running++;
2173 cfs_spin_unlock(&svc->srv_lock);
2176 * wake up our creator. Note: @data is invalid after this point,
2177 * because it's allocated on ptlrpc_start_thread() stack.
2179 cfs_waitq_signal(&thread->t_ctl_waitq);
2181 thread->t_watchdog = lc_watchdog_add(CFS_GET_TIMEOUT(svc), NULL, NULL);
2183 cfs_spin_lock(&svc->srv_rs_lock);
2184 cfs_list_add(&rs->rs_list, &svc->srv_free_rs_list);
2185 cfs_waitq_signal(&svc->srv_free_rs_waitq);
2186 cfs_spin_unlock(&svc->srv_rs_lock);
2188 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
2189 svc->srv_threads_running);
2191 /* XXX maintain a list of all managed devices: insert here */
2192 while (!ptlrpc_thread_stopping(thread)) {
2193 if (ptlrpc_wait_event(svc, thread))
2196 ptlrpc_check_rqbd_pool(svc);
2198 if (ptlrpc_threads_need_create(svc)) {
2199 /* Ignore return code - we tried... */
2200 ptlrpc_start_thread(svc);
2203 /* Process all incoming reqs before handling any */
2204 if (ptlrpc_server_request_waiting(svc)) {
2205 ptlrpc_server_handle_req_in(svc);
2206 /* but limit ourselves in case of flood */
2207 if (counter++ < 100)
2212 if (ptlrpc_at_check(svc))
2213 ptlrpc_at_check_timed(svc);
2215 if (ptlrpc_server_request_pending(svc, 0)) {
2216 lu_context_enter(&env->le_ctx);
2217 ptlrpc_server_handle_request(svc, thread);
2218 lu_context_exit(&env->le_ctx);
2221 if (ptlrpc_rqbd_pending(svc) &&
2222 ptlrpc_server_post_idle_rqbds(svc) < 0) {
2223 /* I just failed to repost request buffers.
2224 * Wait for a timeout (unless something else
2225 * happens) before I try again */
2226 svc->srv_rqbd_timeout = cfs_time_seconds(1)/10;
2227 CDEBUG(D_RPCTRACE,"Posted buffers: %d\n",
2228 svc->srv_nrqbd_receiving);
2232 lc_watchdog_delete(thread->t_watchdog);
2233 thread->t_watchdog = NULL;
2237 * deconstruct service specific state created by ptlrpc_start_thread()
2239 if (svc->srv_done != NULL)
2240 svc->srv_done(thread);
2243 lu_context_fini(&env->le_ctx);
2247 CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n",
2248 thread, thread->t_pid, thread->t_id, rc);
2250 cfs_spin_lock(&svc->srv_lock);
2251 if (thread_test_and_clear_flags(thread, SVC_STARTING))
2252 svc->srv_threads_starting--;
2254 if (thread_test_and_clear_flags(thread, SVC_RUNNING))
2255 /* must know immediately */
2256 svc->srv_threads_running--;
2259 thread_add_flags(thread, SVC_STOPPED);
2261 cfs_waitq_signal(&thread->t_ctl_waitq);
2262 cfs_spin_unlock(&svc->srv_lock);
2267 struct ptlrpc_hr_args {
2270 struct ptlrpc_hr_service *hrs;
2273 static int hrt_dont_sleep(struct ptlrpc_hr_thread *t,
2274 cfs_list_t *replies)
2278 cfs_spin_lock(&t->hrt_lock);
2279 cfs_list_splice_init(&t->hrt_queue, replies);
2280 result = cfs_test_bit(HRT_STOPPING, &t->hrt_flags) ||
2281 !cfs_list_empty(replies);
2282 cfs_spin_unlock(&t->hrt_lock);
2287 * Main body of "handle reply" function.
2288 * It processes acked reply states
2290 static int ptlrpc_hr_main(void *arg)
2292 struct ptlrpc_hr_args * hr_args = arg;
2293 struct ptlrpc_hr_service *hr = hr_args->hrs;
2294 struct ptlrpc_hr_thread *t = &hr->hr_threads[hr_args->thread_index];
2295 char threadname[20];
2296 CFS_LIST_HEAD(replies);
2298 snprintf(threadname, sizeof(threadname),
2299 "ptlrpc_hr_%d", hr_args->thread_index);
2301 cfs_daemonize_ctxt(threadname);
2302 #if defined(CONFIG_NUMA) && defined(HAVE_NODE_TO_CPUMASK)
2303 cfs_set_cpus_allowed(cfs_current(),
2304 node_to_cpumask(cpu_to_node(hr_args->cpu_index)));
2306 cfs_set_bit(HRT_RUNNING, &t->hrt_flags);
2307 cfs_waitq_signal(&t->hrt_wait);
2309 while (!cfs_test_bit(HRT_STOPPING, &t->hrt_flags)) {
2311 l_wait_condition(t->hrt_wait, hrt_dont_sleep(t, &replies));
2312 while (!cfs_list_empty(&replies)) {
2313 struct ptlrpc_reply_state *rs;
2315 rs = cfs_list_entry(replies.prev,
2316 struct ptlrpc_reply_state,
2318 cfs_list_del_init(&rs->rs_list);
2319 ptlrpc_handle_rs(rs);
2323 cfs_clear_bit(HRT_RUNNING, &t->hrt_flags);
2324 cfs_complete(&t->hrt_completion);
2329 static int ptlrpc_start_hr_thread(struct ptlrpc_hr_service *hr, int n, int cpu)
2331 struct ptlrpc_hr_thread *t = &hr->hr_threads[n];
2332 struct ptlrpc_hr_args args;
2336 args.thread_index = n;
2337 args.cpu_index = cpu;
2340 rc = cfs_create_thread(ptlrpc_hr_main, (void*)&args, CFS_DAEMON_FLAGS);
2342 cfs_complete(&t->hrt_completion);
2345 l_wait_condition(t->hrt_wait, cfs_test_bit(HRT_RUNNING, &t->hrt_flags));
2351 static void ptlrpc_stop_hr_thread(struct ptlrpc_hr_thread *t)
2355 cfs_set_bit(HRT_STOPPING, &t->hrt_flags);
2356 cfs_waitq_signal(&t->hrt_wait);
2357 cfs_wait_for_completion(&t->hrt_completion);
2362 static void ptlrpc_stop_hr_threads(struct ptlrpc_hr_service *hrs)
2367 for (n = 0; n < hrs->hr_n_threads; n++)
2368 ptlrpc_stop_hr_thread(&hrs->hr_threads[n]);
2373 static int ptlrpc_start_hr_threads(struct ptlrpc_hr_service *hr)
2376 int n, cpu, threads_started = 0;
2379 LASSERT(hr != NULL);
2380 LASSERT(hr->hr_n_threads > 0);
2382 for (n = 0, cpu = 0; n < hr->hr_n_threads; n++) {
2383 #if defined(CONFIG_SMP) && defined(HAVE_NODE_TO_CPUMASK)
2384 while (!cpu_online(cpu)) {
2386 if (cpu >= cfs_num_possible_cpus())
2390 rc = ptlrpc_start_hr_thread(hr, n, cpu);
2396 if (threads_started == 0) {
2397 CERROR("No reply handling threads started\n");
2400 if (threads_started < hr->hr_n_threads) {
2401 CWARN("Started only %d reply handling threads from %d\n",
2402 threads_started, hr->hr_n_threads);
2403 hr->hr_n_threads = threads_started;
2408 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
2409 struct ptlrpc_thread *thread)
2411 struct l_wait_info lwi = { 0 };
2414 CDEBUG(D_RPCTRACE, "Stopping thread [ %p : %u ]\n",
2415 thread, thread->t_pid);
2417 cfs_spin_lock(&svc->srv_lock);
2418 /* let the thread know that we would like it to stop asap */
2419 thread_add_flags(thread, SVC_STOPPING);
2420 cfs_spin_unlock(&svc->srv_lock);
2422 cfs_waitq_broadcast(&svc->srv_waitq);
2423 l_wait_event(thread->t_ctl_waitq,
2424 thread_is_stopped(thread), &lwi);
2426 cfs_spin_lock(&svc->srv_lock);
2427 cfs_list_del(&thread->t_link);
2428 cfs_spin_unlock(&svc->srv_lock);
2430 OBD_FREE_PTR(thread);
2435 * Stops all threads of a particular service \a svc
2437 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
2439 struct ptlrpc_thread *thread;
2442 cfs_spin_lock(&svc->srv_lock);
2443 while (!cfs_list_empty(&svc->srv_threads)) {
2444 thread = cfs_list_entry(svc->srv_threads.next,
2445 struct ptlrpc_thread, t_link);
2447 cfs_spin_unlock(&svc->srv_lock);
2448 ptlrpc_stop_thread(svc, thread);
2449 cfs_spin_lock(&svc->srv_lock);
2452 cfs_spin_unlock(&svc->srv_lock);
2456 int ptlrpc_start_threads(struct ptlrpc_service *svc)
2461 /* We require 2 threads min - see note in
2462 ptlrpc_server_handle_request */
2463 LASSERT(svc->srv_threads_min >= 2);
2464 for (i = 0; i < svc->srv_threads_min; i++) {
2465 rc = ptlrpc_start_thread(svc);
2466 /* We have enough threads, don't start more. b=15759 */
2467 if (rc == -EMFILE) {
2472 CERROR("cannot start %s thread #%d: rc %d\n",
2473 svc->srv_thread_name, i, rc);
2474 ptlrpc_stop_all_threads(svc);
2481 int ptlrpc_start_thread(struct ptlrpc_service *svc)
2483 struct l_wait_info lwi = { 0 };
2484 struct ptlrpc_svc_data d;
2485 struct ptlrpc_thread *thread;
2490 CDEBUG(D_RPCTRACE, "%s started %d min %d max %d running %d\n",
2491 svc->srv_name, svc->srv_threads_running, svc->srv_threads_min,
2492 svc->srv_threads_max, svc->srv_threads_running);
2494 if (unlikely(svc->srv_is_stopping))
2497 if (!ptlrpc_threads_increasable(svc) ||
2498 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
2499 svc->srv_threads_running == svc->srv_threads_min - 1))
2502 OBD_ALLOC_PTR(thread);
2505 cfs_waitq_init(&thread->t_ctl_waitq);
2507 cfs_spin_lock(&svc->srv_lock);
2508 if (!ptlrpc_threads_increasable(svc)) {
2509 cfs_spin_unlock(&svc->srv_lock);
2510 OBD_FREE_PTR(thread);
2514 svc->srv_threads_starting++;
2515 thread->t_id = svc->srv_threads_next_id++;
2516 thread_add_flags(thread, SVC_STARTING);
2517 thread->t_svc = svc;
2519 cfs_list_add(&thread->t_link, &svc->srv_threads);
2520 cfs_spin_unlock(&svc->srv_lock);
2522 sprintf(name, "%s_%02d", svc->srv_thread_name, thread->t_id);
2527 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", name);
2529 /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
2530 * just drop the VM and FILES in cfs_daemonize_ctxt() right away.
2532 rc = cfs_create_thread(ptlrpc_main, &d, CFS_DAEMON_FLAGS);
2534 CERROR("cannot start thread '%s': rc %d\n", name, rc);
2536 cfs_spin_lock(&svc->srv_lock);
2537 cfs_list_del(&thread->t_link);
2538 --svc->srv_threads_starting;
2539 cfs_spin_unlock(&svc->srv_lock);
2541 OBD_FREE(thread, sizeof(*thread));
2544 l_wait_event(thread->t_ctl_waitq,
2545 thread_is_running(thread) || thread_is_stopped(thread),
2548 rc = thread_is_stopped(thread) ? thread->t_id : 0;
2553 int ptlrpc_hr_init(void)
2556 int n_cpus = cfs_num_online_cpus();
2557 struct ptlrpc_hr_service *hr;
2562 LASSERT(ptlrpc_hr == NULL);
2564 size = offsetof(struct ptlrpc_hr_service, hr_threads[n_cpus]);
2565 OBD_ALLOC(hr, size);
2568 for (i = 0; i < n_cpus; i++) {
2569 struct ptlrpc_hr_thread *t = &hr->hr_threads[i];
2571 cfs_spin_lock_init(&t->hrt_lock);
2572 cfs_waitq_init(&t->hrt_wait);
2573 CFS_INIT_LIST_HEAD(&t->hrt_queue);
2574 cfs_init_completion(&t->hrt_completion);
2576 hr->hr_n_threads = n_cpus;
2580 rc = ptlrpc_start_hr_threads(hr);
2582 OBD_FREE(hr, hr->hr_size);
2588 void ptlrpc_hr_fini(void)
2590 if (ptlrpc_hr != NULL) {
2591 ptlrpc_stop_hr_threads(ptlrpc_hr);
2592 OBD_FREE(ptlrpc_hr, ptlrpc_hr->hr_size);
2597 #endif /* __KERNEL__ */
2600 * Wait until all already scheduled replies are processed.
2602 static void ptlrpc_wait_replies(struct ptlrpc_service *svc)
2606 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
2608 rc = l_wait_event(svc->srv_waitq, cfs_atomic_read(&svc-> \
2609 srv_n_difficult_replies) == 0,
2613 CWARN("Unexpectedly long timeout %p\n", svc);
2617 int ptlrpc_unregister_service(struct ptlrpc_service *service)
2620 struct l_wait_info lwi;
2622 struct ptlrpc_reply_state *rs, *t;
2623 struct ptlrpc_at_array *array = &service->srv_at_array;
2626 service->srv_is_stopping = 1;
2627 cfs_timer_disarm(&service->srv_at_timer);
2629 ptlrpc_stop_all_threads(service);
2630 LASSERT(cfs_list_empty(&service->srv_threads));
2632 cfs_spin_lock (&ptlrpc_all_services_lock);
2633 cfs_list_del_init (&service->srv_list);
2634 cfs_spin_unlock (&ptlrpc_all_services_lock);
2636 ptlrpc_lprocfs_unregister_service(service);
2638 /* All history will be culled when the next request buffer is
2640 service->srv_max_history_rqbds = 0;
2642 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
2644 rc = LNetClearLazyPortal(service->srv_req_portal);
2647 /* Unlink all the request buffers. This forces a 'final' event with
2648 * its 'unlink' flag set for each posted rqbd */
2649 cfs_list_for_each(tmp, &service->srv_active_rqbds) {
2650 struct ptlrpc_request_buffer_desc *rqbd =
2651 cfs_list_entry(tmp, struct ptlrpc_request_buffer_desc,
2654 rc = LNetMDUnlink(rqbd->rqbd_md_h);
2655 LASSERT (rc == 0 || rc == -ENOENT);
2658 /* Wait for the network to release any buffers it's currently
2661 cfs_spin_lock(&service->srv_lock);
2662 rc = service->srv_nrqbd_receiving;
2663 cfs_spin_unlock(&service->srv_lock);
2668 /* Network access will complete in finite time but the HUGE
2669 * timeout lets us CWARN for visibility of sluggish NALs */
2670 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2671 cfs_time_seconds(1), NULL, NULL);
2672 rc = l_wait_event(service->srv_waitq,
2673 service->srv_nrqbd_receiving == 0,
2675 if (rc == -ETIMEDOUT)
2676 CWARN("Service %s waiting for request buffers\n",
2680 /* schedule all outstanding replies to terminate them */
2681 cfs_spin_lock(&service->srv_rs_lock);
2682 while (!cfs_list_empty(&service->srv_active_replies)) {
2683 struct ptlrpc_reply_state *rs =
2684 cfs_list_entry(service->srv_active_replies.next,
2685 struct ptlrpc_reply_state, rs_list);
2686 cfs_spin_lock(&rs->rs_lock);
2687 ptlrpc_schedule_difficult_reply(rs);
2688 cfs_spin_unlock(&rs->rs_lock);
2690 cfs_spin_unlock(&service->srv_rs_lock);
2692 /* purge the request queue. NB No new replies (rqbds all unlinked)
2693 * and no service threads, so I'm the only thread noodling the
2694 * request queue now */
2695 while (!cfs_list_empty(&service->srv_req_in_queue)) {
2696 struct ptlrpc_request *req =
2697 cfs_list_entry(service->srv_req_in_queue.next,
2698 struct ptlrpc_request,
2701 cfs_list_del(&req->rq_list);
2702 service->srv_n_queued_reqs--;
2703 service->srv_n_active_reqs++;
2704 ptlrpc_server_finish_request(service, req);
2706 while (ptlrpc_server_request_pending(service, 1)) {
2707 struct ptlrpc_request *req;
2709 req = ptlrpc_server_request_get(service, 1);
2710 cfs_list_del(&req->rq_list);
2711 service->srv_n_active_reqs++;
2712 ptlrpc_server_finish_request(service, req);
2714 LASSERT(service->srv_n_queued_reqs == 0);
2715 LASSERT(service->srv_n_active_reqs == 0);
2716 LASSERT(service->srv_n_history_rqbds == 0);
2717 LASSERT(cfs_list_empty(&service->srv_active_rqbds));
2719 /* Now free all the request buffers since nothing references them
2721 while (!cfs_list_empty(&service->srv_idle_rqbds)) {
2722 struct ptlrpc_request_buffer_desc *rqbd =
2723 cfs_list_entry(service->srv_idle_rqbds.next,
2724 struct ptlrpc_request_buffer_desc,
2727 ptlrpc_free_rqbd(rqbd);
2730 ptlrpc_wait_replies(service);
2732 cfs_list_for_each_entry_safe(rs, t, &service->srv_free_rs_list,
2734 cfs_list_del(&rs->rs_list);
2735 OBD_FREE_LARGE(rs, service->srv_max_reply_size);
2738 /* In case somebody rearmed this in the meantime */
2739 cfs_timer_disarm(&service->srv_at_timer);
2741 if (array->paa_reqs_array != NULL) {
2742 OBD_FREE(array->paa_reqs_array,
2743 sizeof(cfs_list_t) * array->paa_size);
2744 array->paa_reqs_array = NULL;
2747 if (array->paa_reqs_count != NULL) {
2748 OBD_FREE(array->paa_reqs_count,
2749 sizeof(__u32) * array->paa_size);
2750 array->paa_reqs_count= NULL;
2753 OBD_FREE_PTR(service);
2758 * Returns 0 if the service is healthy.
2760 * Right now, it just checks to make sure that requests aren't languishing
2761 * in the queue. We'll use this health check to govern whether a node needs
2762 * to be shot, so it's intentionally non-aggressive. */
2763 int ptlrpc_service_health_check(struct ptlrpc_service *svc)
2765 struct ptlrpc_request *request;
2766 struct timeval right_now;
2772 cfs_gettimeofday(&right_now);
2774 cfs_spin_lock(&svc->srv_rq_lock);
2775 if (!ptlrpc_server_request_pending(svc, 1)) {
2776 cfs_spin_unlock(&svc->srv_rq_lock);
2780 /* How long has the next entry been waiting? */
2781 if (cfs_list_empty(&svc->srv_request_queue))
2782 request = cfs_list_entry(svc->srv_request_hpq.next,
2783 struct ptlrpc_request, rq_list);
2785 request = cfs_list_entry(svc->srv_request_queue.next,
2786 struct ptlrpc_request, rq_list);
2787 timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
2788 cfs_spin_unlock(&svc->srv_rq_lock);
2790 if ((timediff / ONE_MILLION) > (AT_OFF ? obd_timeout * 3/2 :
2792 CERROR("%s: unhealthy - request has been waiting %lds\n",
2793 svc->srv_name, timediff / ONE_MILLION);