1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 #define DEBUG_SUBSYSTEM S_RPC
39 #include <liblustre.h>
40 #include <libcfs/kp30.h>
42 #include <obd_support.h>
43 #include <obd_class.h>
44 #include <lustre_net.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");
53 CFS_MODULE_PARM(at_min, "i", int, 0644,
54 "Adaptive timeout minimum (sec)");
55 CFS_MODULE_PARM(at_max, "i", int, 0644,
56 "Adaptive timeout maximum (sec)");
57 CFS_MODULE_PARM(at_history, "i", int, 0644,
58 "Adaptive timeouts remember the slowest event that took place "
59 "within this period (sec)");
60 CFS_MODULE_PARM(at_early_margin, "i", int, 0644,
61 "How soon before an RPC deadline to send an early reply");
62 CFS_MODULE_PARM(at_extra, "i", int, 0644,
63 "How much extra time to give with each early reply");
66 static int ptlrpc_server_post_idle_rqbds (struct ptlrpc_service *svc);
68 static CFS_LIST_HEAD (ptlrpc_all_services);
69 spinlock_t ptlrpc_all_services_lock;
72 ptlrpc_alloc_request_buffer (int size)
76 if (size > SVC_BUF_VMALLOC_THRESHOLD)
77 OBD_VMALLOC(ptr, size);
85 ptlrpc_free_request_buffer (char *ptr, int size)
87 if (size > SVC_BUF_VMALLOC_THRESHOLD)
93 struct ptlrpc_request_buffer_desc *
94 ptlrpc_alloc_rqbd (struct ptlrpc_service *svc)
96 struct ptlrpc_request_buffer_desc *rqbd;
98 OBD_ALLOC(rqbd, sizeof (*rqbd));
102 rqbd->rqbd_service = svc;
103 rqbd->rqbd_refcount = 0;
104 rqbd->rqbd_cbid.cbid_fn = request_in_callback;
105 rqbd->rqbd_cbid.cbid_arg = rqbd;
106 CFS_INIT_LIST_HEAD(&rqbd->rqbd_reqs);
107 rqbd->rqbd_buffer = ptlrpc_alloc_request_buffer(svc->srv_buf_size);
109 if (rqbd->rqbd_buffer == NULL) {
110 OBD_FREE(rqbd, sizeof (*rqbd));
114 spin_lock(&svc->srv_lock);
115 list_add(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
117 spin_unlock(&svc->srv_lock);
123 ptlrpc_free_rqbd (struct ptlrpc_request_buffer_desc *rqbd)
125 struct ptlrpc_service *svc = rqbd->rqbd_service;
127 LASSERT (rqbd->rqbd_refcount == 0);
128 LASSERT (list_empty(&rqbd->rqbd_reqs));
130 spin_lock(&svc->srv_lock);
131 list_del(&rqbd->rqbd_list);
133 spin_unlock(&svc->srv_lock);
135 ptlrpc_free_request_buffer (rqbd->rqbd_buffer, svc->srv_buf_size);
136 OBD_FREE (rqbd, sizeof (*rqbd));
140 ptlrpc_grow_req_bufs(struct ptlrpc_service *svc)
142 struct ptlrpc_request_buffer_desc *rqbd;
145 CDEBUG(D_RPCTRACE, "%s: allocate %d new %d-byte reqbufs (%d/%d left)\n",
146 svc->srv_name, svc->srv_nbuf_per_group, svc->srv_buf_size,
147 svc->srv_nrqbd_receiving, svc->srv_nbufs);
148 for (i = 0; i < svc->srv_nbuf_per_group; i++) {
149 rqbd = ptlrpc_alloc_rqbd(svc);
152 CERROR ("%s: Can't allocate request buffer\n",
157 if (ptlrpc_server_post_idle_rqbds(svc) < 0)
165 ptlrpc_save_lock(struct ptlrpc_request *req,
166 struct lustre_handle *lock, int mode)
168 struct ptlrpc_reply_state *rs = req->rq_reply_state;
172 LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
174 if (req->rq_export->exp_disconnected) {
175 ldlm_lock_decref(lock, mode);
177 idx = rs->rs_nlocks++;
178 rs->rs_locks[idx] = *lock;
179 rs->rs_modes[idx] = mode;
180 rs->rs_difficult = 1;
185 ptlrpc_schedule_difficult_reply (struct ptlrpc_reply_state *rs)
187 struct ptlrpc_service *svc = rs->rs_service;
190 LASSERT (spin_is_locked (&svc->srv_lock));
192 LASSERT (rs->rs_difficult);
193 rs->rs_scheduled_ever = 1; /* flag any notification attempt */
195 if (rs->rs_scheduled) /* being set up or already notified */
198 rs->rs_scheduled = 1;
199 list_del (&rs->rs_list);
200 list_add (&rs->rs_list, &svc->srv_reply_queue);
201 cfs_waitq_signal (&svc->srv_waitq);
205 ptlrpc_commit_replies (struct obd_export *exp)
207 struct list_head *tmp;
208 struct list_head *nxt;
210 /* Find any replies that have been committed and get their service
211 * to attend to complete them. */
213 /* CAVEAT EMPTOR: spinlock ordering!!! */
214 spin_lock(&exp->exp_uncommitted_replies_lock);
216 list_for_each_safe(tmp, nxt, &exp->exp_uncommitted_replies) {
217 struct ptlrpc_reply_state *rs =
218 list_entry(tmp, struct ptlrpc_reply_state, rs_obd_list);
220 LASSERT(rs->rs_difficult);
221 /* VBR: per-export last_committed */
222 LASSERT(rs->rs_export);
223 if (rs->rs_transno <= rs->rs_export->exp_last_committed) {
224 struct ptlrpc_service *svc = rs->rs_service;
226 spin_lock (&svc->srv_lock);
227 list_del_init (&rs->rs_obd_list);
228 ptlrpc_schedule_difficult_reply (rs);
229 spin_unlock (&svc->srv_lock);
233 spin_unlock(&exp->exp_uncommitted_replies_lock);
237 ptlrpc_server_post_idle_rqbds (struct ptlrpc_service *svc)
239 struct ptlrpc_request_buffer_desc *rqbd;
244 spin_lock(&svc->srv_lock);
246 if (list_empty (&svc->srv_idle_rqbds)) {
247 spin_unlock(&svc->srv_lock);
251 rqbd = list_entry(svc->srv_idle_rqbds.next,
252 struct ptlrpc_request_buffer_desc,
254 list_del (&rqbd->rqbd_list);
256 /* assume we will post successfully */
257 svc->srv_nrqbd_receiving++;
258 list_add (&rqbd->rqbd_list, &svc->srv_active_rqbds);
260 spin_unlock(&svc->srv_lock);
262 rc = ptlrpc_register_rqbd(rqbd);
269 spin_lock(&svc->srv_lock);
271 svc->srv_nrqbd_receiving--;
272 list_del(&rqbd->rqbd_list);
273 list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
275 /* Don't complain if no request buffers are posted right now; LNET
276 * won't drop requests because we set the portal lazy! */
278 spin_unlock(&svc->srv_lock);
283 static void ptlrpc_at_timer(unsigned long castmeharder)
285 struct ptlrpc_service *svc = (struct ptlrpc_service *)castmeharder;
286 svc->srv_at_check = 1;
287 svc->srv_at_checktime = cfs_time_current();
288 cfs_waitq_signal(&svc->srv_waitq);
291 /* @threadname should be 11 characters or less - 3 will be added on */
292 struct ptlrpc_service *
293 ptlrpc_init_svc(int nbufs, int bufsize, int max_req_size, int max_reply_size,
294 int req_portal, int rep_portal, int watchdog_factor,
295 svc_handler_t handler, char *name,
296 cfs_proc_dir_entry_t *proc_entry,
297 svcreq_printfn_t svcreq_printfn,
298 int min_threads, int max_threads, char *threadname,
299 svc_hpreq_handler_t hp_handler)
302 struct ptlrpc_at_array *array;
303 struct ptlrpc_service *service;
304 unsigned int size, index;
308 LASSERT (bufsize >= max_req_size);
310 OBD_ALLOC(service, sizeof(*service));
314 /* First initialise enough for early teardown */
316 service->srv_name = name;
317 spin_lock_init(&service->srv_lock);
318 CFS_INIT_LIST_HEAD(&service->srv_threads);
319 cfs_waitq_init(&service->srv_waitq);
321 service->srv_nbuf_per_group = test_req_buffer_pressure ? 1 : nbufs;
322 service->srv_max_req_size = max_req_size;
323 service->srv_buf_size = bufsize;
324 service->srv_rep_portal = rep_portal;
325 service->srv_req_portal = req_portal;
326 service->srv_watchdog_factor = watchdog_factor;
327 service->srv_handler = handler;
328 service->srv_request_history_print_fn = svcreq_printfn;
329 service->srv_request_seq = 1; /* valid seq #s start at 1 */
330 service->srv_request_max_cull_seq = 0;
331 service->srv_threads_min = min_threads;
332 service->srv_threads_max = max_threads;
333 service->srv_thread_name = threadname;
334 service->srv_hpreq_handler = hp_handler;
335 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
336 service->srv_hpreq_count = 0;
337 service->srv_n_hpreq = 0;
339 rc = LNetSetLazyPortal(service->srv_req_portal);
342 CFS_INIT_LIST_HEAD(&service->srv_request_queue);
343 CFS_INIT_LIST_HEAD(&service->srv_request_hpq);
344 CFS_INIT_LIST_HEAD(&service->srv_idle_rqbds);
345 CFS_INIT_LIST_HEAD(&service->srv_active_rqbds);
346 CFS_INIT_LIST_HEAD(&service->srv_history_rqbds);
347 CFS_INIT_LIST_HEAD(&service->srv_request_history);
348 CFS_INIT_LIST_HEAD(&service->srv_active_replies);
349 CFS_INIT_LIST_HEAD(&service->srv_reply_queue);
350 CFS_INIT_LIST_HEAD(&service->srv_free_rs_list);
351 cfs_waitq_init(&service->srv_free_rs_waitq);
353 spin_lock_init(&service->srv_at_lock);
354 CFS_INIT_LIST_HEAD(&service->srv_req_in_queue);
356 array = &service->srv_at_array;
357 size = at_est2timeout(at_max);
358 array->paa_size = size;
359 array->paa_count = 0;
360 array->paa_deadline = -1;
362 /* allocate memory for srv_at_array (ptlrpc_at_array) */
363 OBD_ALLOC(array->paa_reqs_array, sizeof(struct list_head) * size);
364 if (array->paa_reqs_array == NULL)
367 for (index = 0; index < size; index++)
368 CFS_INIT_LIST_HEAD(&array->paa_reqs_array[index]);
370 OBD_ALLOC(array->paa_reqs_count, sizeof(__u32) * size);
371 if (array->paa_reqs_count == NULL)
374 cfs_timer_init(&service->srv_at_timer, ptlrpc_at_timer, service);
375 /* At SOW, service time should be quick; 10s seems generous. If client
376 timeout is less than this, we'll be sending an early reply. */
377 at_init(&service->srv_at_estimate, 10, 0);
379 spin_lock (&ptlrpc_all_services_lock);
380 list_add (&service->srv_list, &ptlrpc_all_services);
381 spin_unlock (&ptlrpc_all_services_lock);
383 /* Now allocate the request buffers */
384 rc = ptlrpc_grow_req_bufs(service);
385 /* We shouldn't be under memory pressure at startup, so
386 * fail if we can't post all our buffers at this time. */
390 /* Now allocate pool of reply buffers */
391 /* Increase max reply size to next power of two */
392 service->srv_max_reply_size = 1;
393 while (service->srv_max_reply_size < max_reply_size)
394 service->srv_max_reply_size <<= 1;
396 if (proc_entry != NULL)
397 ptlrpc_lprocfs_register_service(proc_entry, service);
399 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
400 service->srv_name, service->srv_req_portal);
404 ptlrpc_unregister_service(service);
409 * to actually free the request, must be called without holding svc_lock.
410 * note it's caller's responsibility to unlink req->rq_list.
412 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
414 LASSERT(atomic_read(&req->rq_refcount) == 0);
415 LASSERT(list_empty(&req->rq_timed_list));
417 /* DEBUG_REQ() assumes the reply state of a request with a valid
418 * ref will not be destroyed until that reference is dropped. */
419 ptlrpc_req_drop_rs(req);
421 if (req != &req->rq_rqbd->rqbd_req) {
422 /* NB request buffers use an embedded
423 * req if the incoming req unlinked the
424 * MD; this isn't one of them! */
425 OBD_FREE(req, sizeof(*req));
430 * drop a reference count of the request. if it reaches 0, we either
431 * put it into history list, or free it immediately.
433 static void ptlrpc_server_drop_request(struct ptlrpc_request *req)
435 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
436 struct ptlrpc_service *svc = rqbd->rqbd_service;
438 struct list_head *tmp;
439 struct list_head *nxt;
441 if (!atomic_dec_and_test(&req->rq_refcount))
444 spin_lock(&svc->srv_lock);
446 svc->srv_n_active_reqs--;
447 list_add(&req->rq_list, &rqbd->rqbd_reqs);
449 refcount = --(rqbd->rqbd_refcount);
451 /* request buffer is now idle: add to history */
452 list_del(&rqbd->rqbd_list);
453 list_add_tail(&rqbd->rqbd_list, &svc->srv_history_rqbds);
454 svc->srv_n_history_rqbds++;
456 /* cull some history?
457 * I expect only about 1 or 2 rqbds need to be recycled here */
458 while (svc->srv_n_history_rqbds > svc->srv_max_history_rqbds) {
459 rqbd = list_entry(svc->srv_history_rqbds.next,
460 struct ptlrpc_request_buffer_desc,
463 list_del(&rqbd->rqbd_list);
464 svc->srv_n_history_rqbds--;
466 /* remove rqbd's reqs from svc's req history while
467 * I've got the service lock */
468 list_for_each(tmp, &rqbd->rqbd_reqs) {
469 req = list_entry(tmp, struct ptlrpc_request,
471 /* Track the highest culled req seq */
472 if (req->rq_history_seq >
473 svc->srv_request_max_cull_seq)
474 svc->srv_request_max_cull_seq =
476 list_del(&req->rq_history_list);
479 spin_unlock(&svc->srv_lock);
481 list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
482 req = list_entry(rqbd->rqbd_reqs.next,
483 struct ptlrpc_request,
485 list_del(&req->rq_list);
486 ptlrpc_server_free_request(req);
489 spin_lock(&svc->srv_lock);
491 * now all reqs including the embedded req has been
492 * disposed, schedule request buffer for re-use.
494 LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) == 0);
495 list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
498 spin_unlock(&svc->srv_lock);
499 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
500 /* If we are low on memory, we are not interested in history */
501 list_del(&req->rq_list);
502 list_del_init(&req->rq_history_list);
503 spin_unlock(&svc->srv_lock);
505 ptlrpc_server_free_request(req);
507 spin_unlock(&svc->srv_lock);
512 * to finish a request: stop sending more early replies, and release
513 * the request. should be called after we finished handling the request.
515 static void ptlrpc_server_finish_request(struct ptlrpc_request *req)
517 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
519 if (req->rq_export) {
520 class_export_put(req->rq_export);
521 req->rq_export = NULL;
524 if (req->rq_phase != RQ_PHASE_NEW) /* incorrect message magic */
525 DEBUG_REQ(D_INFO, req, "free req");
527 spin_lock(&svc->srv_at_lock);
528 req->rq_sent_final = 1;
529 list_del_init(&req->rq_timed_list);
530 if (req->rq_at_linked) {
531 struct ptlrpc_at_array *array = &svc->srv_at_array;
532 __u32 index = req->rq_at_index;
534 req->rq_at_linked = 0;
535 array->paa_reqs_count[index]--;
538 spin_unlock(&svc->srv_at_lock);
540 ptlrpc_server_drop_request(req);
543 /* This function makes sure dead exports are evicted in a timely manner.
544 This function is only called when some export receives a message (i.e.,
545 the network is up.) */
546 static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
548 struct obd_export *oldest_exp;
549 time_t oldest_time, new_time;
555 /* Compensate for slow machines, etc, by faking our request time
556 into the future. Although this can break the strict time-ordering
557 of the list, we can be really lazy here - we don't have to evict
558 at the exact right moment. Eventually, all silent exports
559 will make it to the top of the list. */
561 /* Do not pay attention on 1sec or smaller renewals. */
562 new_time = cfs_time_current_sec() + extra_delay;
563 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
566 exp->exp_last_request_time = new_time;
567 CDEBUG(D_INFO, "updating export %s at %ld\n",
568 exp->exp_client_uuid.uuid,
569 exp->exp_last_request_time);
571 /* exports may get disconnected from the chain even though the
572 export has references, so we must keep the spin lock while
573 manipulating the lists */
574 spin_lock(&exp->exp_obd->obd_dev_lock);
576 if (list_empty(&exp->exp_obd_chain_timed)) {
577 /* this one is not timed */
578 spin_unlock(&exp->exp_obd->obd_dev_lock);
582 list_move_tail(&exp->exp_obd_chain_timed,
583 &exp->exp_obd->obd_exports_timed);
585 oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
586 struct obd_export, exp_obd_chain_timed);
587 oldest_time = oldest_exp->exp_last_request_time;
588 spin_unlock(&exp->exp_obd->obd_dev_lock);
590 if (exp->exp_obd->obd_recovering) {
591 /* be nice to everyone during recovery */
596 /* Note - racing to start/reset the obd_eviction timer is safe */
597 if (exp->exp_obd->obd_eviction_timer == 0) {
598 /* Check if the oldest entry is expired. */
599 if (cfs_time_current_sec() > (oldest_time + PING_EVICT_TIMEOUT +
601 /* We need a second timer, in case the net was down and
602 * it just came back. Since the pinger may skip every
603 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
604 * we better wait for 3. */
605 exp->exp_obd->obd_eviction_timer =
606 cfs_time_current_sec() + 3 * PING_INTERVAL;
607 CDEBUG(D_HA, "%s: Think about evicting %s from %ld\n",
608 exp->exp_obd->obd_name, obd_export_nid2str(exp),
612 if (cfs_time_current_sec() >
613 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
614 /* The evictor won't evict anyone who we've heard from
615 * recently, so we don't have to check before we start
617 if (!ping_evictor_wake(exp))
618 exp->exp_obd->obd_eviction_timer = 0;
625 static int ptlrpc_check_req(struct ptlrpc_request *req)
627 if (lustre_msg_get_conn_cnt(req->rq_reqmsg) <
628 req->rq_export->exp_conn_cnt) {
629 DEBUG_REQ(D_ERROR, req,
630 "DROPPING req from old connection %d < %d",
631 lustre_msg_get_conn_cnt(req->rq_reqmsg),
632 req->rq_export->exp_conn_cnt);
635 if (req->rq_export->exp_obd && req->rq_export->exp_obd->obd_fail) {
636 /* Failing over, don't handle any more reqs, send
637 error response instead. */
638 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
639 req, req->rq_export->exp_obd->obd_name);
640 req->rq_status = -ENODEV;
647 static void ptlrpc_at_set_timer(struct ptlrpc_service *svc)
649 struct ptlrpc_at_array *array = &svc->srv_at_array;
652 spin_lock(&svc->srv_at_lock);
653 if (array->paa_count == 0) {
654 cfs_timer_disarm(&svc->srv_at_timer);
655 spin_unlock(&svc->srv_at_lock);
659 /* Set timer for closest deadline */
660 next = (__s32)(array->paa_deadline - cfs_time_current_sec() -
663 ptlrpc_at_timer((unsigned long)svc);
665 cfs_timer_arm(&svc->srv_at_timer, cfs_time_shift(next));
666 spin_unlock(&svc->srv_at_lock);
667 CDEBUG(D_INFO, "armed %s at %+ds\n", svc->srv_name, next);
670 /* Add rpc to early reply check list */
671 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
673 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
674 struct ptlrpc_request *rq = NULL;
675 struct ptlrpc_at_array *array = &svc->srv_at_array;
682 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
685 spin_lock(&svc->srv_at_lock);
687 if (unlikely(req->rq_sent_final)) {
688 spin_unlock(&svc->srv_at_lock);
692 LASSERT(list_empty(&req->rq_timed_list));
694 index = (unsigned long)req->rq_deadline % array->paa_size;
695 if (array->paa_reqs_count[index] > 0) {
696 /* latest rpcs will have the latest deadlines in the list,
697 * so search backward. */
698 list_for_each_entry_reverse(rq, &array->paa_reqs_array[index],
700 if (req->rq_deadline >= rq->rq_deadline) {
701 list_add(&req->rq_timed_list,
708 /* Add the request at the head of the list */
709 if (list_empty(&req->rq_timed_list))
710 list_add(&req->rq_timed_list, &array->paa_reqs_array[index]);
712 req->rq_at_linked = 1;
713 req->rq_at_index = index;
714 array->paa_reqs_count[index]++;
716 if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
717 array->paa_deadline = req->rq_deadline;
720 spin_unlock(&svc->srv_at_lock);
723 ptlrpc_at_set_timer(svc);
728 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req,
731 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
732 struct ptlrpc_request *reqcopy;
733 struct lustre_msg *reqmsg;
734 long olddl = req->rq_deadline - cfs_time_current_sec();
739 /* deadline is when the client expects us to reply, margin is the
740 difference between clients' and servers' expectations */
741 DEBUG_REQ(D_ADAPTTO, req,
742 "%ssending early reply (deadline %+lds, margin %+lds) for "
743 "%d+%d", AT_OFF ? "AT off - not " : "",
744 olddl, olddl - at_get(&svc->srv_at_estimate),
745 at_get(&svc->srv_at_estimate), extra_time);
751 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), "
752 "not sending early reply. Consider increasing "
753 "at_early_margin (%d)?", olddl, at_early_margin);
755 /* Return an error so we're not re-added to the timed list. */
759 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
760 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
761 "but no AT support");
765 if (req->rq_export && req->rq_export->exp_in_recovery) {
766 /* don't increase server estimates during recovery, and give
767 clients the full recovery time. */
768 newdl = cfs_time_current_sec() +
769 req->rq_export->exp_obd->obd_recovery_timeout;
772 /* Fake our processing time into the future to ask the
773 clients for some extra amount of time */
774 extra_time += cfs_time_current_sec() -
775 req->rq_arrival_time.tv_sec;
776 at_add(&svc->srv_at_estimate, extra_time);
778 newdl = req->rq_arrival_time.tv_sec +
779 at_get(&svc->srv_at_estimate);
781 if (req->rq_deadline >= newdl) {
782 /* We're not adding any time, no need to send an early reply
783 (e.g. maybe at adaptive_max) */
784 DEBUG_REQ(D_WARNING, req, "Couldn't add any time "
785 "(%ld/%ld), not sending early reply\n",
786 olddl, newdl - cfs_time_current_sec());
790 OBD_ALLOC(reqcopy, sizeof *reqcopy);
793 OBD_ALLOC(reqmsg, req->rq_reqlen);
795 OBD_FREE(reqcopy, sizeof *reqcopy);
800 reqcopy->rq_reply_state = NULL;
801 reqcopy->rq_rep_swab_mask = 0;
802 /* We only need the reqmsg for the magic */
803 reqcopy->rq_reqmsg = reqmsg;
804 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
806 if (req->rq_sent_final) {
807 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
808 "abort sending early reply\n");
813 reqcopy->rq_export = class_conn2export(
814 lustre_msg_get_handle(reqcopy->rq_reqmsg));
815 if (reqcopy->rq_export == NULL)
816 GOTO(out, rc = -ENODEV);
819 class_export_rpc_get(reqcopy->rq_export);
820 if (reqcopy->rq_export->exp_obd &&
821 reqcopy->rq_export->exp_obd->obd_fail)
822 GOTO(out_put, rc = -ENODEV);
824 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
828 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
831 /* Adjust our own deadline to what we told the client */
832 req->rq_deadline = newdl;
833 req->rq_early_count++; /* number sent, server side */
835 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
838 /* Free the (early) reply state from lustre_pack_reply.
839 (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
840 ptlrpc_req_drop_rs(reqcopy);
843 class_export_rpc_put(reqcopy->rq_export);
844 class_export_put(reqcopy->rq_export);
846 OBD_FREE(reqmsg, req->rq_reqlen);
847 OBD_FREE(reqcopy, sizeof *reqcopy);
851 /* Send early replies to everybody expiring within at_early_margin
852 asking for at_extra time */
853 static int ptlrpc_at_check_timed(struct ptlrpc_service *svc)
855 struct ptlrpc_request *rq, *n;
856 struct list_head work_list;
857 struct ptlrpc_at_array *array = &svc->srv_at_array;
860 time_t now = cfs_time_current_sec();
861 cfs_duration_t delay;
862 int first, counter = 0;
865 spin_lock(&svc->srv_at_lock);
866 if (svc->srv_at_check == 0) {
867 spin_unlock(&svc->srv_at_lock);
870 delay = cfs_time_sub(cfs_time_current(), svc->srv_at_checktime);
871 svc->srv_at_check = 0;
873 if (array->paa_count == 0) {
874 spin_unlock(&svc->srv_at_lock);
878 /* The timer went off, but maybe the nearest rpc already completed. */
879 first = array->paa_deadline - now;
880 if (first > at_early_margin) {
881 /* We've still got plenty of time. Reset the timer. */
882 spin_unlock(&svc->srv_at_lock);
883 ptlrpc_at_set_timer(svc);
887 /* We're close to a timeout, and we don't know how much longer the
888 server will take. Send early replies to everyone expiring soon. */
889 CFS_INIT_LIST_HEAD(&work_list);
891 index = (unsigned long)array->paa_deadline % array->paa_size;
892 count = array->paa_count;
894 count -= array->paa_reqs_count[index];
895 list_for_each_entry_safe(rq, n, &array->paa_reqs_array[index],
897 if (rq->rq_deadline <= now + at_early_margin) {
898 list_move(&rq->rq_timed_list, &work_list);
900 array->paa_reqs_count[index]--;
902 rq->rq_at_linked = 0;
906 /* update the earliest deadline */
907 if (deadline == -1 || rq->rq_deadline < deadline)
908 deadline = rq->rq_deadline;
913 if (++index >= array->paa_size)
916 array->paa_deadline = deadline;
917 spin_unlock(&svc->srv_at_lock);
919 /* we have a new earliest deadline, restart the timer */
920 ptlrpc_at_set_timer(svc);
922 CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
923 "replies\n", first, at_extra, counter);
926 /* We're already past request deadlines before we even get a
927 chance to send early replies */
928 LCONSOLE_WARN("%s: This server is not able to keep up with "
929 "request traffic (cpu-bound).\n", svc->srv_name);
930 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, "
931 "delay="CFS_DURATION_T"(jiff)\n",
932 counter, svc->srv_n_queued_reqs, svc->srv_n_active_reqs,
933 at_get(&svc->srv_at_estimate), delay);
936 /* ptlrpc_server_finish_request may delete an entry out of
938 spin_lock(&svc->srv_at_lock);
939 while (!list_empty(&work_list)) {
940 rq = list_entry(work_list.next, struct ptlrpc_request,
942 list_del_init(&rq->rq_timed_list);
943 /* if the entry is still in the worklist, it hasn't been
944 deleted, and is safe to take a ref to keep the req around */
945 atomic_inc(&rq->rq_refcount);
946 spin_unlock(&svc->srv_at_lock);
948 if (ptlrpc_at_send_early_reply(rq, at_extra) == 0)
949 ptlrpc_at_add_timed(rq);
951 ptlrpc_server_drop_request(rq);
952 spin_lock(&svc->srv_at_lock);
954 spin_unlock(&svc->srv_at_lock);
960 * Put the request to the export list if the request may become
961 * a high priority one.
963 static int ptlrpc_hpreq_init(struct ptlrpc_service *svc,
964 struct ptlrpc_request *req)
969 if (svc->srv_hpreq_handler) {
970 rc = svc->srv_hpreq_handler(req);
974 if (req->rq_export && req->rq_ops) {
975 spin_lock(&req->rq_export->exp_lock);
976 list_add(&req->rq_exp_list, &req->rq_export->exp_queued_rpc);
977 spin_unlock(&req->rq_export->exp_lock);
983 /** Remove the request from the export list. */
984 static void ptlrpc_hpreq_fini(struct ptlrpc_request *req)
987 if (req->rq_export && req->rq_ops) {
988 spin_lock(&req->rq_export->exp_lock);
989 list_del_init(&req->rq_exp_list);
990 spin_unlock(&req->rq_export->exp_lock);
996 * Make the request a high priority one.
998 * All the high priority requests are queued in a separate FIFO
999 * ptlrpc_service::srv_request_hpq list which is parallel to
1000 * ptlrpc_service::srv_request_queue list but has a higher priority
1003 * \see ptlrpc_server_handle_request().
1005 static void ptlrpc_hpreq_reorder_nolock(struct ptlrpc_service *svc,
1006 struct ptlrpc_request *req)
1009 LASSERT(svc != NULL);
1010 spin_lock(&req->rq_lock);
1011 if (req->rq_hp == 0) {
1012 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1014 /* Add to the high priority queue. */
1015 list_move_tail(&req->rq_list, &svc->srv_request_hpq);
1017 if (opc != OBD_PING)
1018 DEBUG_REQ(D_NET, req, "high priority req");
1020 spin_unlock(&req->rq_lock);
1024 void ptlrpc_hpreq_reorder(struct ptlrpc_request *req)
1026 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
1029 spin_lock(&svc->srv_lock);
1030 /* It may happen that the request is already taken for the processing
1031 * but still in the export list, do not re-add it into the HP list. */
1032 if (req->rq_phase == RQ_PHASE_NEW)
1033 ptlrpc_hpreq_reorder_nolock(svc, req);
1034 spin_unlock(&svc->srv_lock);
1038 /** Check if the request if a high priority one. */
1039 static int ptlrpc_server_hpreq_check(struct ptlrpc_request *req)
1044 /* Check by request opc. */
1045 opc = lustre_msg_get_opc(req->rq_reqmsg);
1046 if (opc == OBD_PING)
1049 /* Perform request specific check. */
1050 if (req->rq_ops && req->rq_ops->hpreq_check)
1051 rc = req->rq_ops->hpreq_check(req);
1055 /** Check if a request is a high priority one. */
1056 static int ptlrpc_server_request_add(struct ptlrpc_service *svc,
1057 struct ptlrpc_request *req)
1062 rc = ptlrpc_server_hpreq_check(req);
1066 spin_lock(&svc->srv_lock);
1067 /* Before inserting the request into the queue, check if it is not
1068 * inserted yet, or even already handled -- it may happen due to
1069 * a racing ldlm_server_blocking_ast(). */
1070 if (req->rq_phase == RQ_PHASE_NEW && list_empty(&req->rq_list)) {
1072 ptlrpc_hpreq_reorder_nolock(svc, req);
1074 list_add_tail(&req->rq_list, &svc->srv_request_queue);
1076 spin_unlock(&svc->srv_lock);
1081 /* Only allow normal priority requests on a service that has a high-priority
1082 * queue if forced (i.e. cleanup), if there are other high priority requests
1083 * already being processed (i.e. those threads can service more high-priority
1084 * requests), or if there are enough idle threads that a later thread can do
1085 * a high priority request. */
1086 static int ptlrpc_server_allow_normal(struct ptlrpc_service *svc, int force)
1088 return force || !svc->srv_hpreq_handler || svc->srv_n_hpreq > 0 ||
1089 svc->srv_n_active_reqs < svc->srv_threads_running - 2;
1092 static struct ptlrpc_request *
1093 ptlrpc_server_request_get(struct ptlrpc_service *svc, int force)
1095 struct ptlrpc_request *req = NULL;
1098 if (ptlrpc_server_allow_normal(svc, force) &&
1099 !list_empty(&svc->srv_request_queue) &&
1100 (list_empty(&svc->srv_request_hpq) ||
1101 svc->srv_hpreq_count >= svc->srv_hpreq_ratio)) {
1102 req = list_entry(svc->srv_request_queue.next,
1103 struct ptlrpc_request, rq_list);
1104 svc->srv_hpreq_count = 0;
1105 } else if (!list_empty(&svc->srv_request_hpq)) {
1106 req = list_entry(svc->srv_request_hpq.next,
1107 struct ptlrpc_request, rq_list);
1108 svc->srv_hpreq_count++;
1113 static int ptlrpc_server_request_pending(struct ptlrpc_service *svc, int force)
1115 return ((ptlrpc_server_allow_normal(svc, force) &&
1116 !list_empty(&svc->srv_request_queue)) ||
1117 !list_empty(&svc->srv_request_hpq));
1120 /* Handle freshly incoming reqs, add to timed early reply list,
1121 pass on to regular request queue */
1123 ptlrpc_server_handle_req_in(struct ptlrpc_service *svc)
1125 struct ptlrpc_request *req;
1132 spin_lock(&svc->srv_lock);
1133 if (list_empty(&svc->srv_req_in_queue)) {
1134 spin_unlock(&svc->srv_lock);
1138 req = list_entry(svc->srv_req_in_queue.next,
1139 struct ptlrpc_request, rq_list);
1140 list_del_init (&req->rq_list);
1141 /* Consider this still a "queued" request as far as stats are
1143 spin_unlock(&svc->srv_lock);
1145 /* Clear request swab mask; this is a new request */
1146 req->rq_req_swab_mask = 0;
1148 rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
1150 CERROR ("error unpacking request: ptl %d from %s"
1151 " xid "LPU64"\n", svc->srv_req_portal,
1152 libcfs_id2str(req->rq_peer), req->rq_xid);
1157 lustre_set_req_swabbed(req, MSG_PTLRPC_HEADER_OFF);
1159 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1161 CERROR ("error unpacking ptlrpc body: ptl %d from %s"
1162 " xid "LPU64"\n", svc->srv_req_portal,
1163 libcfs_id2str(req->rq_peer), req->rq_xid);
1168 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1169 CERROR("wrong packet type received (type=%u) from %s\n",
1170 lustre_msg_get_type(req->rq_reqmsg),
1171 libcfs_id2str(req->rq_peer));
1175 CDEBUG(D_NET, "got req "LPU64"\n", req->rq_xid);
1177 req->rq_export = class_conn2export(
1178 lustre_msg_get_handle(req->rq_reqmsg));
1179 if (req->rq_export) {
1180 rc = ptlrpc_check_req(req);
1183 ptlrpc_update_export_timer(req->rq_export, 0);
1186 /* req_in handling should/must be fast */
1187 if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5)
1188 DEBUG_REQ(D_WARNING, req, "Slow req_in handling %lus",
1189 cfs_time_current_sec() - req->rq_arrival_time.tv_sec);
1191 /* Set rpc server deadline and add it to the timed list */
1192 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1193 MSGHDR_AT_SUPPORT) ?
1194 /* The max time the client expects us to take */
1195 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1196 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1197 if (unlikely(deadline == 0)) {
1198 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1202 ptlrpc_at_add_timed(req);
1203 rc = ptlrpc_hpreq_init(svc, req);
1207 /* Move it over to the request processing queue */
1208 rc = ptlrpc_server_request_add(svc, req);
1211 cfs_waitq_signal(&svc->srv_waitq);
1215 spin_lock(&svc->srv_lock);
1216 svc->srv_n_queued_reqs--;
1217 svc->srv_n_active_reqs++;
1218 spin_unlock(&svc->srv_lock);
1219 ptlrpc_server_finish_request(req);
1225 ptlrpc_server_handle_request(struct ptlrpc_service *svc,
1226 struct ptlrpc_thread *thread)
1228 struct obd_export *export = NULL;
1229 struct ptlrpc_request *request;
1230 struct timeval work_start;
1231 struct timeval work_end;
1239 spin_lock(&svc->srv_lock);
1240 if (!ptlrpc_server_request_pending(svc, 0) ||
1243 /* !@%$# liblustre only has 1 thread */
1244 svc->srv_n_difficult_replies != 0 &&
1246 svc->srv_n_active_reqs >= (svc->srv_threads_running - 1))) {
1247 /* Don't handle regular requests in the last thread, in order
1248 * to handle difficult replies (which might block other threads)
1249 * as well as handle any incoming reqs, early replies, etc.
1250 * That means we always need at least 2 service threads. */
1251 spin_unlock(&svc->srv_lock);
1255 request = ptlrpc_server_request_get(svc, 0);
1256 if (request == NULL) {
1257 spin_unlock(&svc->srv_lock);
1261 opc = lustre_msg_get_opc(request->rq_reqmsg);
1262 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
1263 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
1264 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
1265 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
1267 if (unlikely(fail_opc)) {
1268 if (request->rq_export && request->rq_ops) {
1269 spin_unlock(&svc->srv_lock);
1270 OBD_FAIL_TIMEOUT(fail_opc, 4);
1271 spin_lock(&svc->srv_lock);
1272 request = ptlrpc_server_request_get(svc, 0);
1273 if (request == NULL) {
1274 spin_unlock(&svc->srv_lock);
1277 LASSERT(ptlrpc_server_request_pending(svc, 0));
1281 list_del_init(&request->rq_list);
1282 svc->srv_n_queued_reqs--;
1283 svc->srv_n_active_reqs++;
1288 /* The phase is changed under the lock here because we need to know
1289 * the request is under processing (see ptlrpc_hpreq_reorder()). */
1290 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
1291 spin_unlock(&svc->srv_lock);
1293 ptlrpc_hpreq_fini(request);
1295 if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
1296 libcfs_debug_dumplog();
1298 do_gettimeofday(&work_start);
1299 timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
1300 if (svc->srv_stats != NULL) {
1301 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
1303 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
1304 svc->srv_n_queued_reqs);
1305 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
1306 svc->srv_n_active_reqs);
1307 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
1308 at_get(&svc->srv_at_estimate));
1311 CDEBUG(D_NET, "got req "LPU64"\n", request->rq_xid);
1313 request->rq_svc_thread = thread;
1314 if (request->rq_export) {
1315 if (ptlrpc_check_req(request))
1317 ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
1318 export = class_export_rpc_get(request->rq_export);
1321 /* Discard requests queued for longer than the deadline.
1322 The deadline is increased if we send an early reply. */
1323 if (cfs_time_current_sec() > request->rq_deadline) {
1324 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s"
1325 ": deadline %ld%+lds ago\n",
1326 libcfs_id2str(request->rq_peer),
1327 request->rq_deadline -
1328 request->rq_arrival_time.tv_sec,
1329 cfs_time_current_sec() - request->rq_deadline);
1330 goto put_rpc_export;
1333 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
1334 "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1335 (request->rq_export ?
1336 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1337 (request->rq_export ?
1338 atomic_read(&request->rq_export->exp_refcount) : -99),
1339 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1340 libcfs_id2str(request->rq_peer),
1341 lustre_msg_get_opc(request->rq_reqmsg));
1343 if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
1344 OBD_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, obd_fail_val);
1346 rc = svc->srv_handler(request);
1348 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
1350 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
1351 "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1352 (request->rq_export ?
1353 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1354 (request->rq_export ?
1355 atomic_read(&request->rq_export->exp_refcount) : -99),
1356 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1357 libcfs_id2str(request->rq_peer),
1358 lustre_msg_get_opc(request->rq_reqmsg));
1361 if (export != NULL && !request->rq_copy_queued)
1362 class_export_rpc_put(export);
1365 if (cfs_time_current_sec() > request->rq_deadline) {
1366 DEBUG_REQ(D_WARNING, request, "Request x"LPU64" took longer "
1367 "than estimated (%ld%+lds); client may timeout.",
1368 request->rq_xid, request->rq_deadline -
1369 request->rq_arrival_time.tv_sec,
1370 cfs_time_current_sec() - request->rq_deadline);
1373 do_gettimeofday(&work_end);
1374 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1375 CDEBUG(D_RPCTRACE, "request x"LPU64" opc %u from %s processed in "
1376 "%ldus (%ldus total) trans "LPU64" rc %d/%d\n",
1377 request->rq_xid, lustre_msg_get_opc(request->rq_reqmsg),
1378 libcfs_id2str(request->rq_peer), timediff,
1379 cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL),
1380 request->rq_repmsg ? lustre_msg_get_transno(request->rq_repmsg) :
1381 request->rq_transno, request->rq_status,
1382 request->rq_repmsg ? lustre_msg_get_status(request->rq_repmsg):
1384 if (svc->srv_stats != NULL) {
1385 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
1386 int opc = opcode_offset(op);
1387 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
1388 LASSERT(opc < LUSTRE_MAX_OPCODES);
1389 lprocfs_counter_add(svc->srv_stats,
1390 opc + EXTRA_MAX_OPCODES,
1394 if (request->rq_early_count) {
1395 DEBUG_REQ(D_ADAPTTO, request,
1396 "sent %d early replies before finishing in %lds",
1397 request->rq_early_count,
1398 work_end.tv_sec - request->rq_arrival_time.tv_sec);
1401 spin_lock(&svc->srv_lock);
1404 spin_unlock(&svc->srv_lock);
1405 ptlrpc_server_finish_request(request);
1411 ptlrpc_server_handle_reply (struct ptlrpc_service *svc)
1413 struct ptlrpc_reply_state *rs;
1414 struct obd_export *exp;
1415 struct obd_device *obd;
1420 spin_lock(&svc->srv_lock);
1421 if (list_empty (&svc->srv_reply_queue)) {
1422 spin_unlock(&svc->srv_lock);
1426 rs = list_entry (svc->srv_reply_queue.next,
1427 struct ptlrpc_reply_state, rs_list);
1429 exp = rs->rs_export;
1432 LASSERT (rs->rs_difficult);
1433 LASSERT (rs->rs_scheduled);
1435 list_del_init (&rs->rs_list);
1437 /* Disengage from notifiers carefully (lock order - irqrestore below!)*/
1438 spin_unlock(&svc->srv_lock);
1440 spin_lock (&exp->exp_uncommitted_replies_lock);
1441 /* Noop if removed already */
1442 list_del_init (&rs->rs_obd_list);
1443 spin_unlock (&exp->exp_uncommitted_replies_lock);
1445 spin_lock (&exp->exp_lock);
1446 /* Noop if removed already */
1447 list_del_init (&rs->rs_exp_list);
1448 spin_unlock (&exp->exp_lock);
1450 spin_lock(&svc->srv_lock);
1452 been_handled = rs->rs_handled;
1455 nlocks = rs->rs_nlocks; /* atomic "steal", but */
1456 rs->rs_nlocks = 0; /* locks still on rs_locks! */
1458 if (nlocks == 0 && !been_handled) {
1459 /* If we see this, we should already have seen the warning
1460 * in mds_steal_ack_locks() */
1461 CWARN("All locks stolen from rs %p x"LPD64".t"LPD64
1462 " o%d NID %s\n", rs, rs->rs_xid, rs->rs_transno,
1463 lustre_msg_get_opc(rs->rs_msg),
1464 libcfs_nid2str(exp->exp_connection->c_peer.nid));
1467 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
1468 spin_unlock(&svc->srv_lock);
1470 if (!been_handled && rs->rs_on_net) {
1471 LNetMDUnlink(rs->rs_md_h);
1472 /* Ignore return code; we're racing with
1476 while (nlocks-- > 0)
1477 ldlm_lock_decref(&rs->rs_locks[nlocks],
1478 rs->rs_modes[nlocks]);
1480 spin_lock(&svc->srv_lock);
1483 rs->rs_scheduled = 0;
1485 if (!rs->rs_on_net) {
1487 svc->srv_n_difficult_replies--;
1488 spin_unlock(&svc->srv_lock);
1490 class_export_put (exp);
1491 rs->rs_export = NULL;
1492 ptlrpc_rs_decref (rs);
1493 atomic_dec (&svc->srv_outstanding_replies);
1497 /* still on the net; callback will schedule */
1498 spin_unlock(&svc->srv_lock);
1503 /* FIXME make use of timeout later */
1505 liblustre_check_services (void *arg)
1507 int did_something = 0;
1509 struct list_head *tmp, *nxt;
1512 /* I'm relying on being single threaded, not to have to lock
1513 * ptlrpc_all_services etc */
1514 list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
1515 struct ptlrpc_service *svc =
1516 list_entry (tmp, struct ptlrpc_service, srv_list);
1518 if (svc->srv_threads_running != 0) /* I've recursed */
1521 /* service threads can block for bulk, so this limits us
1522 * (arbitrarily) to recursing 1 stack frame per service.
1523 * Note that the problem with recursion is that we have to
1524 * unwind completely before our caller can resume. */
1526 svc->srv_threads_running++;
1529 rc = ptlrpc_server_handle_req_in(svc);
1530 rc |= ptlrpc_server_handle_reply(svc);
1531 rc |= ptlrpc_at_check_timed(svc);
1532 rc |= ptlrpc_server_handle_request(svc, NULL);
1533 rc |= (ptlrpc_server_post_idle_rqbds(svc) > 0);
1534 did_something |= rc;
1537 svc->srv_threads_running--;
1540 RETURN(did_something);
1542 #define ptlrpc_stop_all_threads(s) do {} while (0)
1544 #else /* __KERNEL__ */
1547 ptlrpc_check_rqbd_pool(struct ptlrpc_service *svc)
1549 int avail = svc->srv_nrqbd_receiving;
1550 int low_water = test_req_buffer_pressure ? 0 :
1551 svc->srv_nbuf_per_group/2;
1553 /* NB I'm not locking; just looking. */
1555 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
1556 * allowed the request history to grow out of control. We could put a
1557 * sanity check on that here and cull some history if we need the
1560 if (avail <= low_water)
1561 ptlrpc_grow_req_bufs(svc);
1564 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQBUF_AVAIL_CNTR,
1569 ptlrpc_retry_rqbds(void *arg)
1571 struct ptlrpc_service *svc = (struct ptlrpc_service *)arg;
1573 svc->srv_rqbd_timeout = 0;
1574 return (-ETIMEDOUT);
1577 static int ptlrpc_main(void *arg)
1579 struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
1580 struct ptlrpc_service *svc = data->svc;
1581 struct ptlrpc_thread *thread = data->thread;
1582 struct obd_device *dev = data->dev;
1583 struct ptlrpc_reply_state *rs;
1584 #ifdef WITH_GROUP_INFO
1585 struct group_info *ginfo = NULL;
1587 int counter = 0, rc = 0;
1590 cfs_daemonize_ctxt(data->name);
1592 #if defined(HAVE_NODE_TO_CPUMASK) && defined(CONFIG_NUMA)
1593 /* we need to do this before any per-thread allocation is done so that
1594 * we get the per-thread allocations on local node. bug 7342 */
1595 if (svc->srv_cpu_affinity) {
1598 for (cpu = 0, num_cpu = 0; cpu < num_possible_cpus(); cpu++) {
1599 if (!cpu_online(cpu))
1601 if (num_cpu == thread->t_id % num_online_cpus())
1605 set_cpus_allowed(cfs_current(), node_to_cpumask(cpu_to_node(cpu)));
1609 #ifdef WITH_GROUP_INFO
1610 ginfo = groups_alloc(0);
1616 set_current_groups(ginfo);
1617 put_group_info(ginfo);
1620 if (svc->srv_init != NULL) {
1621 rc = svc->srv_init(thread);
1626 /* Alloc reply state structure for this one */
1627 OBD_ALLOC_GFP(rs, svc->srv_max_reply_size, CFS_ALLOC_STD);
1633 /* Record that the thread is running */
1634 thread->t_flags = SVC_RUNNING;
1636 * wake up our creator. Note: @data is invalid after this point,
1637 * because it's allocated on ptlrpc_start_thread() stack.
1639 cfs_waitq_signal(&thread->t_ctl_waitq);
1641 thread->t_watchdog = lc_watchdog_add(GET_TIMEOUT(svc), NULL, NULL);
1643 spin_lock(&svc->srv_lock);
1644 svc->srv_threads_running++;
1645 list_add(&rs->rs_list, &svc->srv_free_rs_list);
1646 spin_unlock(&svc->srv_lock);
1647 cfs_waitq_signal(&svc->srv_free_rs_waitq);
1649 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
1650 svc->srv_threads_running);
1652 /* XXX maintain a list of all managed devices: insert here */
1654 while ((thread->t_flags & SVC_STOPPING) == 0 ||
1655 svc->srv_n_difficult_replies != 0) {
1656 /* Don't exit while there are replies to be handled */
1657 struct l_wait_info lwi = LWI_TIMEOUT(svc->srv_rqbd_timeout,
1658 ptlrpc_retry_rqbds, svc);
1660 lc_watchdog_disable(thread->t_watchdog);
1664 l_wait_event_exclusive (svc->srv_waitq,
1665 ((thread->t_flags & SVC_STOPPING) != 0 &&
1666 svc->srv_n_difficult_replies == 0) ||
1667 (!list_empty(&svc->srv_idle_rqbds) &&
1668 svc->srv_rqbd_timeout == 0) ||
1669 !list_empty(&svc->srv_req_in_queue) ||
1670 !list_empty(&svc->srv_reply_queue) ||
1671 (ptlrpc_server_request_pending(svc, 0) &&
1672 (svc->srv_n_active_reqs <
1673 (svc->srv_threads_running - 1))) ||
1677 lc_watchdog_touch(thread->t_watchdog, GET_TIMEOUT(svc));
1679 ptlrpc_check_rqbd_pool(svc);
1681 if ((svc->srv_threads_started < svc->srv_threads_max) &&
1682 (svc->srv_n_active_reqs >= (svc->srv_threads_started - 1))){
1683 /* Ignore return code - we tried... */
1684 ptlrpc_start_thread(dev, svc);
1687 if (!list_empty(&svc->srv_reply_queue))
1688 ptlrpc_server_handle_reply (svc);
1690 if (!list_empty(&svc->srv_req_in_queue)) {
1691 /* Process all incoming reqs before handling any */
1692 ptlrpc_server_handle_req_in(svc);
1693 /* but limit ourselves in case of flood */
1694 if (counter++ < 1000)
1699 if (svc->srv_at_check)
1700 ptlrpc_at_check_timed(svc);
1702 /* don't handle requests in the last thread */
1703 if (ptlrpc_server_request_pending(svc, 0) &&
1704 (svc->srv_n_active_reqs < (svc->srv_threads_running - 1)))
1705 ptlrpc_server_handle_request(svc, thread);
1707 if (!list_empty(&svc->srv_idle_rqbds) &&
1708 ptlrpc_server_post_idle_rqbds(svc) < 0) {
1709 /* I just failed to repost request buffers. Wait
1710 * for a timeout (unless something else happens)
1711 * before I try again */
1712 svc->srv_rqbd_timeout = cfs_time_seconds(1)/10;
1713 CDEBUG(D_RPCTRACE,"Posted buffers: %d\n",
1714 svc->srv_nrqbd_receiving);
1718 lc_watchdog_delete(thread->t_watchdog);
1719 thread->t_watchdog = NULL;
1723 * deconstruct service specific state created by ptlrpc_start_thread()
1725 if (svc->srv_done != NULL)
1726 svc->srv_done(thread);
1729 CDEBUG(D_NET, "service thread %d exiting: rc %d\n", thread->t_id, rc);
1731 spin_lock(&svc->srv_lock);
1732 svc->srv_threads_running--; /* must know immediately */
1734 thread->t_flags = SVC_STOPPED;
1736 cfs_waitq_signal(&thread->t_ctl_waitq);
1737 spin_unlock(&svc->srv_lock);
1742 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
1743 struct ptlrpc_thread *thread)
1745 struct l_wait_info lwi = { 0 };
1747 spin_lock(&svc->srv_lock);
1748 thread->t_flags = SVC_STOPPING;
1749 spin_unlock(&svc->srv_lock);
1751 cfs_waitq_broadcast(&svc->srv_waitq);
1752 l_wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED),
1755 spin_lock(&svc->srv_lock);
1756 list_del(&thread->t_link);
1757 spin_unlock(&svc->srv_lock);
1759 OBD_FREE(thread, sizeof(*thread));
1762 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
1764 struct ptlrpc_thread *thread;
1766 spin_lock(&svc->srv_lock);
1767 while (!list_empty(&svc->srv_threads)) {
1768 thread = list_entry(svc->srv_threads.next,
1769 struct ptlrpc_thread, t_link);
1771 spin_unlock(&svc->srv_lock);
1772 ptlrpc_stop_thread(svc, thread);
1773 spin_lock(&svc->srv_lock);
1776 spin_unlock(&svc->srv_lock);
1779 int ptlrpc_start_threads(struct obd_device *dev, struct ptlrpc_service *svc)
1784 /* We require 2 threads min - see note in
1785 * ptlrpc_server_handle_request() */
1787 LASSERT(svc->srv_threads_min >= 2);
1788 for (i = 0; i < svc->srv_threads_min; i++) {
1789 rc = ptlrpc_start_thread(dev, svc);
1790 /* We have enough threads, don't start more. b=15759 */
1794 CERROR("cannot start %s thread #%d: rc %d\n",
1795 svc->srv_thread_name, i, rc);
1796 ptlrpc_stop_all_threads(svc);
1802 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc)
1804 struct l_wait_info lwi = { 0 };
1805 struct ptlrpc_svc_data d;
1806 struct ptlrpc_thread *thread;
1811 CDEBUG(D_RPCTRACE, "%s started %d min %d max %d running %d\n",
1812 svc->srv_name, svc->srv_threads_started, svc->srv_threads_min,
1813 svc->srv_threads_max, svc->srv_threads_running);
1814 if (unlikely(svc->srv_threads_started >= svc->srv_threads_max) ||
1815 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
1816 svc->srv_threads_started == svc->srv_threads_min - 1))
1819 OBD_ALLOC(thread, sizeof(*thread));
1822 cfs_waitq_init(&thread->t_ctl_waitq);
1824 spin_lock(&svc->srv_lock);
1825 if (svc->srv_threads_started >= svc->srv_threads_max) {
1826 spin_unlock(&svc->srv_lock);
1827 OBD_FREE(thread, sizeof(*thread));
1830 list_add(&thread->t_link, &svc->srv_threads);
1831 id = svc->srv_threads_started++;
1832 spin_unlock(&svc->srv_lock);
1834 thread->t_svc = svc;
1836 sprintf(name, "%s_%02d", svc->srv_thread_name, id);
1842 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", name);
1844 /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
1845 * just drop the VM and FILES in cfs_daemonize_ctxt() right away.
1847 rc = cfs_kernel_thread(ptlrpc_main, &d, CLONE_VM | CLONE_FILES);
1849 CERROR("cannot start thread '%s': rc %d\n", name, rc);
1851 spin_lock(&svc->srv_lock);
1852 list_del(&thread->t_link);
1853 --svc->srv_threads_started;
1854 spin_unlock(&svc->srv_lock);
1856 OBD_FREE(thread, sizeof(*thread));
1859 l_wait_event(thread->t_ctl_waitq,
1860 thread->t_flags & (SVC_RUNNING | SVC_STOPPED), &lwi);
1862 rc = (thread->t_flags & SVC_STOPPED) ? thread->t_id : 0;
1867 int ptlrpc_unregister_service(struct ptlrpc_service *service)
1870 struct l_wait_info lwi;
1871 struct list_head *tmp;
1872 struct ptlrpc_reply_state *rs, *t;
1873 struct ptlrpc_at_array *array = &service->srv_at_array;
1875 cfs_timer_disarm(&service->srv_at_timer);
1877 ptlrpc_stop_all_threads(service);
1878 LASSERT(list_empty(&service->srv_threads));
1880 spin_lock (&ptlrpc_all_services_lock);
1881 list_del_init (&service->srv_list);
1882 spin_unlock (&ptlrpc_all_services_lock);
1884 ptlrpc_lprocfs_unregister_service(service);
1886 /* All history will be culled when the next request buffer is
1888 service->srv_max_history_rqbds = 0;
1890 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
1892 rc = LNetClearLazyPortal(service->srv_req_portal);
1895 /* Unlink all the request buffers. This forces a 'final' event with
1896 * its 'unlink' flag set for each posted rqbd */
1897 list_for_each(tmp, &service->srv_active_rqbds) {
1898 struct ptlrpc_request_buffer_desc *rqbd =
1899 list_entry(tmp, struct ptlrpc_request_buffer_desc,
1902 rc = LNetMDUnlink(rqbd->rqbd_md_h);
1903 LASSERT (rc == 0 || rc == -ENOENT);
1906 /* Wait for the network to release any buffers it's currently
1909 spin_lock(&service->srv_lock);
1910 rc = service->srv_nrqbd_receiving;
1911 spin_unlock(&service->srv_lock);
1916 /* Network access will complete in finite time but the HUGE
1917 * timeout lets us CWARN for visibility of sluggish NALs */
1918 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
1919 cfs_time_seconds(1), NULL, NULL);
1920 rc = l_wait_event(service->srv_waitq,
1921 service->srv_nrqbd_receiving == 0,
1923 if (rc == -ETIMEDOUT)
1924 CWARN("Service %s waiting for request buffers\n",
1928 /* schedule all outstanding replies to terminate them */
1929 spin_lock(&service->srv_lock);
1930 while (!list_empty(&service->srv_active_replies)) {
1931 struct ptlrpc_reply_state *rs =
1932 list_entry(service->srv_active_replies.next,
1933 struct ptlrpc_reply_state, rs_list);
1934 ptlrpc_schedule_difficult_reply(rs);
1936 spin_unlock(&service->srv_lock);
1938 /* purge the request queue. NB No new replies (rqbds all unlinked)
1939 * and no service threads, so I'm the only thread noodling the
1940 * request queue now */
1941 while (!list_empty(&service->srv_req_in_queue)) {
1942 struct ptlrpc_request *req =
1943 list_entry(service->srv_req_in_queue.next,
1944 struct ptlrpc_request,
1947 list_del(&req->rq_list);
1948 service->srv_n_queued_reqs--;
1949 service->srv_n_active_reqs++;
1950 ptlrpc_server_finish_request(req);
1952 while (ptlrpc_server_request_pending(service, 1)) {
1953 struct ptlrpc_request *req;
1955 req = ptlrpc_server_request_get(service, 1);
1956 list_del(&req->rq_list);
1957 service->srv_n_queued_reqs--;
1958 service->srv_n_active_reqs++;
1959 ptlrpc_hpreq_fini(req);
1960 ptlrpc_server_finish_request(req);
1962 LASSERT(service->srv_n_queued_reqs == 0);
1963 LASSERT(service->srv_n_active_reqs == 0);
1964 LASSERT(service->srv_n_history_rqbds == 0);
1965 LASSERT(list_empty(&service->srv_active_rqbds));
1967 /* Now free all the request buffers since nothing references them
1969 while (!list_empty(&service->srv_idle_rqbds)) {
1970 struct ptlrpc_request_buffer_desc *rqbd =
1971 list_entry(service->srv_idle_rqbds.next,
1972 struct ptlrpc_request_buffer_desc,
1975 ptlrpc_free_rqbd(rqbd);
1978 /* wait for all outstanding replies to complete (they were
1979 * scheduled having been flagged to abort above) */
1980 while (atomic_read(&service->srv_outstanding_replies) != 0) {
1981 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10), NULL, NULL);
1983 rc = l_wait_event(service->srv_waitq,
1984 !list_empty(&service->srv_reply_queue), &lwi);
1985 LASSERT(rc == 0 || rc == -ETIMEDOUT);
1988 ptlrpc_server_handle_reply(service);
1991 CWARN("Unexpectedly long timeout %p\n", service);
1994 list_for_each_entry_safe(rs, t, &service->srv_free_rs_list, rs_list) {
1995 list_del(&rs->rs_list);
1996 OBD_FREE(rs, service->srv_max_reply_size);
1999 /* In case somebody rearmed this in the meantime */
2000 cfs_timer_disarm(&service->srv_at_timer);
2002 if (array->paa_reqs_array != NULL) {
2003 OBD_FREE(array->paa_reqs_array,
2004 sizeof(struct list_head) * array->paa_size);
2005 array->paa_reqs_array = NULL;
2008 if (array->paa_reqs_count != NULL) {
2009 OBD_FREE(array->paa_reqs_count,
2010 sizeof(__u32) * array->paa_size);
2011 array->paa_reqs_count= NULL;
2014 OBD_FREE(service, sizeof(*service));
2018 /* Returns 0 if the service is healthy.
2020 * Right now, it just checks to make sure that requests aren't languishing
2021 * in the queue. We'll use this health check to govern whether a node needs
2022 * to be shot, so it's intentionally non-aggressive. */
2023 int ptlrpc_service_health_check(struct ptlrpc_service *svc)
2025 struct ptlrpc_request *request;
2026 struct timeval right_now;
2032 do_gettimeofday(&right_now);
2034 spin_lock(&svc->srv_lock);
2035 if (!ptlrpc_server_request_pending(svc, 1)) {
2036 spin_unlock(&svc->srv_lock);
2040 /* How long has the next entry been waiting? */
2041 if (list_empty(&svc->srv_request_queue))
2042 request = list_entry(svc->srv_request_hpq.next,
2043 struct ptlrpc_request, rq_list);
2045 request = list_entry(svc->srv_request_queue.next,
2046 struct ptlrpc_request, rq_list);
2047 timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
2048 spin_unlock(&svc->srv_lock);
2050 if ((timediff / ONE_MILLION) > (AT_OFF ? obd_timeout * 3/2 :
2052 CERROR("%s: unhealthy - request has been waiting %lds\n",
2053 svc->srv_name, timediff / ONE_MILLION);