1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (C) 2002 Cluster File Systems, Inc.
6 * This file is part of the Lustre file system, http://www.lustre.org
7 * Lustre is a trademark of Cluster File Systems, Inc.
9 * You may have signed or agreed to another license before downloading
10 * this software. If so, you are bound by the terms and conditions
11 * of that agreement, and the following does not apply to you. See the
12 * LICENSE file included with this distribution for more information.
14 * If you did not agree to a different license, then this copy of Lustre
15 * is open source software; you can redistribute it and/or modify it
16 * under the terms of version 2 of the GNU General Public License as
17 * published by the Free Software Foundation.
19 * In either case, Lustre is distributed in the hope that it will be
20 * useful, but WITHOUT ANY WARRANTY; without even the implied warranty
21 * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
22 * license text for more details.
26 #define DEBUG_SUBSYSTEM S_RPC
28 #include <liblustre.h>
29 #include <libcfs/kp30.h>
31 #include <obd_support.h>
32 #include <obd_class.h>
33 #include <lustre_net.h>
34 #include <lu_object.h>
35 #include <lnet/types.h>
36 #include "ptlrpc_internal.h"
38 int test_req_buffer_pressure = 0;
39 CFS_MODULE_PARM(test_req_buffer_pressure, "i", int, 0444,
40 "set non-zero to put pressure on request buffer pools");
43 static int ptlrpc_server_post_idle_rqbds (struct ptlrpc_service *svc);
45 static CFS_LIST_HEAD(ptlrpc_all_services);
46 spinlock_t ptlrpc_all_services_lock;
49 ptlrpc_alloc_request_buffer (int size)
53 if (size > SVC_BUF_VMALLOC_THRESHOLD)
54 OBD_VMALLOC(ptr, size);
62 ptlrpc_free_request_buffer (char *ptr, int size)
64 if (size > SVC_BUF_VMALLOC_THRESHOLD)
70 struct ptlrpc_request_buffer_desc *
71 ptlrpc_alloc_rqbd (struct ptlrpc_service *svc)
73 struct ptlrpc_request_buffer_desc *rqbd;
79 rqbd->rqbd_service = svc;
80 rqbd->rqbd_refcount = 0;
81 rqbd->rqbd_cbid.cbid_fn = request_in_callback;
82 rqbd->rqbd_cbid.cbid_arg = rqbd;
83 CFS_INIT_LIST_HEAD(&rqbd->rqbd_reqs);
84 rqbd->rqbd_buffer = ptlrpc_alloc_request_buffer(svc->srv_buf_size);
86 if (rqbd->rqbd_buffer == NULL) {
91 spin_lock(&svc->srv_lock);
92 list_add(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
94 spin_unlock(&svc->srv_lock);
100 ptlrpc_free_rqbd (struct ptlrpc_request_buffer_desc *rqbd)
102 struct ptlrpc_service *svc = rqbd->rqbd_service;
104 LASSERT (rqbd->rqbd_refcount == 0);
105 LASSERT (list_empty(&rqbd->rqbd_reqs));
107 spin_lock(&svc->srv_lock);
108 list_del(&rqbd->rqbd_list);
110 spin_unlock(&svc->srv_lock);
112 ptlrpc_free_request_buffer (rqbd->rqbd_buffer, svc->srv_buf_size);
117 ptlrpc_grow_req_bufs(struct ptlrpc_service *svc)
119 struct ptlrpc_request_buffer_desc *rqbd;
122 CDEBUG(D_RPCTRACE, "%s: allocate %d new %d-byte reqbufs (%d/%d left)\n",
123 svc->srv_name, svc->srv_nbuf_per_group, svc->srv_buf_size,
124 svc->srv_nrqbd_receiving, svc->srv_nbufs);
125 for (i = 0; i < svc->srv_nbuf_per_group; i++) {
126 rqbd = ptlrpc_alloc_rqbd(svc);
129 CERROR ("%s: Can't allocate request buffer\n",
134 if (ptlrpc_server_post_idle_rqbds(svc) < 0)
142 ptlrpc_save_lock (struct ptlrpc_request *req,
143 struct lustre_handle *lock, int mode)
145 struct ptlrpc_reply_state *rs = req->rq_reply_state;
149 LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
151 idx = rs->rs_nlocks++;
152 rs->rs_locks[idx] = *lock;
153 rs->rs_modes[idx] = mode;
154 rs->rs_difficult = 1;
158 ptlrpc_schedule_difficult_reply (struct ptlrpc_reply_state *rs)
160 struct ptlrpc_service *svc = rs->rs_service;
163 LASSERT (spin_is_locked (&svc->srv_lock));
165 LASSERT (rs->rs_difficult);
166 rs->rs_scheduled_ever = 1; /* flag any notification attempt */
168 if (rs->rs_scheduled) /* being set up or already notified */
171 rs->rs_scheduled = 1;
172 list_del (&rs->rs_list);
173 list_add (&rs->rs_list, &svc->srv_reply_queue);
174 cfs_waitq_signal (&svc->srv_waitq);
178 ptlrpc_commit_replies (struct obd_device *obd)
180 struct list_head *tmp;
181 struct list_head *nxt;
183 /* Find any replies that have been committed and get their service
184 * to attend to complete them. */
186 /* CAVEAT EMPTOR: spinlock ordering!!! */
187 spin_lock(&obd->obd_uncommitted_replies_lock);
189 list_for_each_safe (tmp, nxt, &obd->obd_uncommitted_replies) {
190 struct ptlrpc_reply_state *rs =
191 list_entry(tmp, struct ptlrpc_reply_state, rs_obd_list);
193 LASSERT (rs->rs_difficult);
195 if (rs->rs_transno <= obd->obd_last_committed) {
196 struct ptlrpc_service *svc = rs->rs_service;
198 spin_lock (&svc->srv_lock);
199 list_del_init (&rs->rs_obd_list);
200 ptlrpc_schedule_difficult_reply (rs);
201 spin_unlock (&svc->srv_lock);
205 spin_unlock(&obd->obd_uncommitted_replies_lock);
209 ptlrpc_server_post_idle_rqbds (struct ptlrpc_service *svc)
211 struct ptlrpc_request_buffer_desc *rqbd;
216 spin_lock(&svc->srv_lock);
218 if (list_empty (&svc->srv_idle_rqbds)) {
219 spin_unlock(&svc->srv_lock);
223 rqbd = list_entry(svc->srv_idle_rqbds.next,
224 struct ptlrpc_request_buffer_desc,
226 list_del (&rqbd->rqbd_list);
228 /* assume we will post successfully */
229 svc->srv_nrqbd_receiving++;
230 list_add (&rqbd->rqbd_list, &svc->srv_active_rqbds);
232 spin_unlock(&svc->srv_lock);
234 rc = ptlrpc_register_rqbd(rqbd);
241 spin_lock(&svc->srv_lock);
243 svc->srv_nrqbd_receiving--;
244 list_del(&rqbd->rqbd_list);
245 list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
247 /* Don't complain if no request buffers are posted right now; LNET
248 * won't drop requests because we set the portal lazy! */
250 spin_unlock(&svc->srv_lock);
255 struct ptlrpc_service *ptlrpc_init_svc_conf(struct ptlrpc_service_conf *c,
256 svc_handler_t h, char *name,
257 struct proc_dir_entry *proc_entry,
258 svcreq_printfn_t prntfn,
261 return ptlrpc_init_svc(c->psc_nbufs, c->psc_bufsize,
262 c->psc_max_req_size, c->psc_max_reply_size,
263 c->psc_req_portal, c->psc_rep_portal,
264 c->psc_watchdog_timeout,
266 prntfn, c->psc_min_threads, c->psc_max_threads,
267 threadname, c->psc_ctx_tags);
269 EXPORT_SYMBOL(ptlrpc_init_svc_conf);
271 /* @threadname should be 11 characters or less - 3 will be added on */
272 struct ptlrpc_service *
273 ptlrpc_init_svc(int nbufs, int bufsize, int max_req_size, int max_reply_size,
274 int req_portal, int rep_portal, int watchdog_timeout,
275 svc_handler_t handler, char *name,
276 cfs_proc_dir_entry_t *proc_entry,
277 svcreq_printfn_t svcreq_printfn,
278 int min_threads, int max_threads,
279 char *threadname, __u32 ctx_tags)
282 struct ptlrpc_service *service;
286 LASSERT (bufsize >= max_req_size + SPTLRPC_MAX_PAYLOAD);
287 LASSERT (ctx_tags != 0);
289 OBD_ALLOC_PTR(service);
293 /* First initialise enough for early teardown */
295 service->srv_name = name;
296 spin_lock_init(&service->srv_lock);
297 CFS_INIT_LIST_HEAD(&service->srv_threads);
298 cfs_waitq_init(&service->srv_waitq);
300 service->srv_nbuf_per_group = test_req_buffer_pressure ? 1 : nbufs;
301 service->srv_max_req_size = max_req_size + SPTLRPC_MAX_PAYLOAD;
302 service->srv_buf_size = bufsize;
303 service->srv_rep_portal = rep_portal;
304 service->srv_req_portal = req_portal;
305 service->srv_watchdog_timeout = watchdog_timeout;
306 service->srv_handler = handler;
307 service->srv_request_history_print_fn = svcreq_printfn;
308 service->srv_request_seq = 1; /* valid seq #s start at 1 */
309 service->srv_request_max_cull_seq = 0;
310 service->srv_threads_min = min_threads;
311 service->srv_threads_max = max_threads;
312 service->srv_thread_name = threadname;
313 service->srv_ctx_tags = ctx_tags;
315 rc = LNetSetLazyPortal(service->srv_req_portal);
318 CFS_INIT_LIST_HEAD(&service->srv_request_queue);
319 CFS_INIT_LIST_HEAD(&service->srv_idle_rqbds);
320 CFS_INIT_LIST_HEAD(&service->srv_active_rqbds);
321 CFS_INIT_LIST_HEAD(&service->srv_history_rqbds);
322 CFS_INIT_LIST_HEAD(&service->srv_request_history);
323 CFS_INIT_LIST_HEAD(&service->srv_active_replies);
324 CFS_INIT_LIST_HEAD(&service->srv_reply_queue);
325 CFS_INIT_LIST_HEAD(&service->srv_free_rs_list);
326 cfs_waitq_init(&service->srv_free_rs_waitq);
328 spin_lock (&ptlrpc_all_services_lock);
329 list_add (&service->srv_list, &ptlrpc_all_services);
330 spin_unlock (&ptlrpc_all_services_lock);
332 /* Now allocate the request buffers */
333 rc = ptlrpc_grow_req_bufs(service);
334 /* We shouldn't be under memory pressure at startup, so
335 * fail if we can't post all our buffers at this time. */
339 /* Now allocate pool of reply buffers */
340 /* Increase max reply size to next power of two */
341 service->srv_max_reply_size = 1;
342 while (service->srv_max_reply_size <
343 max_reply_size + SPTLRPC_MAX_PAYLOAD)
344 service->srv_max_reply_size <<= 1;
346 if (proc_entry != NULL)
347 ptlrpc_lprocfs_register_service(proc_entry, service);
349 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
350 service->srv_name, service->srv_req_portal);
354 ptlrpc_unregister_service(service);
358 static void __ptlrpc_server_free_request(struct ptlrpc_request *req)
360 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
362 list_del(&req->rq_list);
364 if (req->rq_reply_state != NULL) {
365 ptlrpc_rs_decref(req->rq_reply_state);
366 req->rq_reply_state = NULL;
369 sptlrpc_svc_ctx_decref(req);
371 if (req != &rqbd->rqbd_req) {
372 /* NB request buffers use an embedded
373 * req if the incoming req unlinked the
374 * MD; this isn't one of them! */
375 OBD_FREE(req, sizeof(*req));
380 ptlrpc_server_free_request(struct ptlrpc_request *req)
382 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
383 struct ptlrpc_service *svc = rqbd->rqbd_service;
385 struct list_head *tmp;
386 struct list_head *nxt;
388 spin_lock(&svc->srv_lock);
390 svc->srv_n_active_reqs--;
391 list_add(&req->rq_list, &rqbd->rqbd_reqs);
393 refcount = --(rqbd->rqbd_refcount);
395 /* request buffer is now idle: add to history */
396 list_del(&rqbd->rqbd_list);
397 list_add_tail(&rqbd->rqbd_list, &svc->srv_history_rqbds);
398 svc->srv_n_history_rqbds++;
400 /* cull some history?
401 * I expect only about 1 or 2 rqbds need to be recycled here */
402 while (svc->srv_n_history_rqbds > svc->srv_max_history_rqbds) {
403 rqbd = list_entry(svc->srv_history_rqbds.next,
404 struct ptlrpc_request_buffer_desc,
407 list_del(&rqbd->rqbd_list);
408 svc->srv_n_history_rqbds--;
410 /* remove rqbd's reqs from svc's req history while
411 * I've got the service lock */
412 list_for_each(tmp, &rqbd->rqbd_reqs) {
413 req = list_entry(tmp, struct ptlrpc_request,
415 /* Track the highest culled req seq */
416 if (req->rq_history_seq >
417 svc->srv_request_max_cull_seq)
418 svc->srv_request_max_cull_seq =
420 list_del(&req->rq_history_list);
423 spin_unlock(&svc->srv_lock);
425 list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
426 req = list_entry(rqbd->rqbd_reqs.next,
427 struct ptlrpc_request,
429 __ptlrpc_server_free_request(req);
432 spin_lock(&svc->srv_lock);
434 /* schedule request buffer for re-use.
435 * NB I can only do this after I've disposed of their
436 * reqs; particularly the embedded req */
437 list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
439 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
440 /* If we are low on memory, we are not interested in
442 list_del(&req->rq_history_list);
443 __ptlrpc_server_free_request(req);
446 spin_unlock(&svc->srv_lock);
450 /* This function makes sure dead exports are evicted in a timely manner.
451 This function is only called when some export receives a message (i.e.,
452 the network is up.) */
453 static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
455 struct obd_export *oldest_exp;
462 /* Compensate for slow machines, etc, by faking our request time
463 into the future. Although this can break the strict time-ordering
464 of the list, we can be really lazy here - we don't have to evict
465 at the exact right moment. Eventually, all silent exports
466 will make it to the top of the list. */
467 exp->exp_last_request_time = max(exp->exp_last_request_time,
468 cfs_time_current_sec() + extra_delay);
470 CDEBUG(D_HA, "updating export %s at "CFS_TIME_T" exp %p\n",
471 exp->exp_client_uuid.uuid,
472 exp->exp_last_request_time, exp);
474 /* exports may get disconnected from the chain even though the
475 export has references, so we must keep the spin lock while
476 manipulating the lists */
477 spin_lock(&exp->exp_obd->obd_dev_lock);
479 if (list_empty(&exp->exp_obd_chain_timed)) {
480 /* this one is not timed */
481 spin_unlock(&exp->exp_obd->obd_dev_lock);
486 list_move_tail(&exp->exp_obd_chain_timed,
487 &exp->exp_obd->obd_exports_timed);
489 oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
490 struct obd_export, exp_obd_chain_timed);
491 oldest_time = oldest_exp->exp_last_request_time;
492 spin_unlock(&exp->exp_obd->obd_dev_lock);
494 if (exp->exp_obd->obd_recovering) {
495 /* be nice to everyone during recovery */
500 /* Note - racing to start/reset the obd_eviction timer is safe */
501 if (exp->exp_obd->obd_eviction_timer == 0) {
502 /* Check if the oldest entry is expired. */
503 if (cfs_time_current_sec() > (oldest_time +
504 (3 * obd_timeout / 2) + extra_delay)) {
505 /* We need a second timer, in case the net was down and
506 * it just came back. Since the pinger may skip every
507 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
508 * we better wait for 3. */
509 exp->exp_obd->obd_eviction_timer = cfs_time_current_sec() +
511 CDEBUG(D_HA, "%s: Think about evicting %s from "CFS_TIME_T"\n",
512 exp->exp_obd->obd_name, obd_export_nid2str(exp),
516 if (cfs_time_current_sec() > (exp->exp_obd->obd_eviction_timer +
518 /* The evictor won't evict anyone who we've heard from
519 * recently, so we don't have to check before we start
521 if (!ping_evictor_wake(exp))
522 exp->exp_obd->obd_eviction_timer = 0;
530 int lu_context_init(struct lu_context *ctx, __u32 tags)
535 void lu_context_fini(struct lu_context *ctx)
539 void lu_context_enter(struct lu_context *ctx)
543 void lu_context_exit(struct lu_context *ctx)
550 ptlrpc_server_handle_request(struct ptlrpc_service *svc,
551 struct ptlrpc_thread *thread)
553 struct obd_export *export = NULL;
554 struct ptlrpc_request *request;
555 struct timeval work_start;
556 struct timeval work_end;
563 spin_lock(&svc->srv_lock);
564 if (unlikely(list_empty (&svc->srv_request_queue) ||
565 (svc->srv_n_difficult_replies != 0 &&
566 svc->srv_n_active_reqs >= (svc->srv_threads_running - 1)))) {
567 /* If all the other threads are handling requests, I must
568 * remain free to handle any 'difficult' reply that might
570 spin_unlock(&svc->srv_lock);
574 request = list_entry (svc->srv_request_queue.next,
575 struct ptlrpc_request, rq_list);
576 list_del_init (&request->rq_list);
577 svc->srv_n_queued_reqs--;
578 svc->srv_n_active_reqs++;
580 spin_unlock(&svc->srv_lock);
582 do_gettimeofday(&work_start);
583 timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
584 if (likely(svc->srv_stats != NULL)) {
585 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
587 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
588 svc->srv_n_queued_reqs);
589 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
590 svc->srv_n_active_reqs);
593 /* go through security check/transform */
594 rc = sptlrpc_svc_unwrap_request(request);
598 case SECSVC_COMPLETE:
599 target_send_reply(request, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
607 /* Clear request swab mask; this is a new request */
608 request->rq_req_swab_mask = 0;
610 rc = lustre_unpack_msg(request->rq_reqmsg, request->rq_reqlen);
612 CERROR ("error unpacking request: ptl %d from %s"
613 " xid "LPU64"\n", svc->srv_req_portal,
614 libcfs_id2str(request->rq_peer), request->rq_xid);
618 rc = lustre_unpack_req_ptlrpc_body(request, MSG_PTLRPC_BODY_OFF);
620 CERROR ("error unpacking ptlrpc body: ptl %d from %s"
621 " xid "LPU64"\n", svc->srv_req_portal,
622 libcfs_id2str(request->rq_peer), request->rq_xid);
627 if (lustre_msg_get_type(request->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
628 CERROR("wrong packet type received (type=%u) from %s\n",
629 lustre_msg_get_type(request->rq_reqmsg),
630 libcfs_id2str(request->rq_peer));
634 rc = lu_context_init(&request->rq_session, LCT_SESSION);
636 CERROR("Failure to initialize session: %d\n", rc);
639 request->rq_session.lc_thread = thread;
640 lu_context_enter(&request->rq_session);
642 CDEBUG(D_NET, "got req "LPD64"\n", request->rq_xid);
644 request->rq_svc_thread = thread;
646 request->rq_svc_thread->t_env->le_ses = &request->rq_session;
648 request->rq_export = class_conn2export(
649 lustre_msg_get_handle(request->rq_reqmsg));
651 if (likely(request->rq_export)) {
652 if (unlikely(lustre_msg_get_conn_cnt(request->rq_reqmsg) <
653 request->rq_export->exp_conn_cnt)) {
654 DEBUG_REQ(D_ERROR, request,
655 "DROPPING req from old connection %d < %d",
656 lustre_msg_get_conn_cnt(request->rq_reqmsg),
657 request->rq_export->exp_conn_cnt);
660 if (unlikely(request->rq_export->exp_obd &&
661 request->rq_export->exp_obd->obd_fail)) {
662 /* Failing over, don't handle any more reqs, send
663 error response instead. */
664 CDEBUG(D_RPCTRACE,"Dropping req %p for failed obd %s\n",
665 request, request->rq_export->exp_obd->obd_name);
666 request->rq_status = -ENODEV;
667 ptlrpc_error(request);
671 rc = sptlrpc_target_export_check(request->rq_export, request);
673 DEBUG_REQ(D_ERROR, request,
674 "DROPPING req with illegal security flavor,");
678 ptlrpc_update_export_timer(request->rq_export, timediff/500000);
679 export = class_export_rpc_get(request->rq_export);
682 /* Discard requests queued for longer than my timeout. If the
683 * client's timeout is similar to mine, she'll be timing out this
684 * REQ anyway (bug 1502) */
685 if (unlikely(timediff / 1000000 > (long)obd_timeout)) {
686 CERROR("Dropping timed-out opc %d request from %s"
687 ": %ld seconds old\n",
688 lustre_msg_get_opc(request->rq_reqmsg),
689 libcfs_id2str(request->rq_peer),
694 request->rq_phase = RQ_PHASE_INTERPRET;
696 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
697 "%s:%s+%d:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
698 (request->rq_export ?
699 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
700 (request->rq_export ?
701 atomic_read(&request->rq_export->exp_refcount) : -99),
702 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
703 libcfs_id2str(request->rq_peer),
704 lustre_msg_get_opc(request->rq_reqmsg));
706 rc = svc->srv_handler(request);
708 request->rq_phase = RQ_PHASE_COMPLETE;
710 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
711 "%s:%s+%d:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
712 (request->rq_export ?
713 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
714 (request->rq_export ?
715 atomic_read(&request->rq_export->exp_refcount) : -99),
716 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
717 libcfs_id2str(request->rq_peer),
718 lustre_msg_get_opc(request->rq_reqmsg));
722 class_export_rpc_put(export);
724 if (likely(request->rq_export != NULL))
725 class_export_put(request->rq_export);
727 lu_context_exit(&request->rq_session);
728 lu_context_fini(&request->rq_session);
730 reply = request->rq_reply_state && request->rq_repmsg; /* bug 11169 */
732 do_gettimeofday(&work_end);
734 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
736 if (unlikely(timediff / 1000000 > (long)obd_timeout))
737 CERROR("request "LPU64" opc %u from %s processed in %lds "
738 "trans "LPU64" rc %d/%d\n",
741 lustre_msg_get_opc(request->rq_reqmsg) : 0,
742 libcfs_id2str(request->rq_peer),
743 cfs_timeval_sub(&work_end, &request->rq_arrival_time,
745 reply ? lustre_msg_get_transno(request->rq_repmsg) :
746 request->rq_transno, request->rq_status,
747 reply ? lustre_msg_get_status(request->rq_repmsg) : -999);
749 CDEBUG(D_RPCTRACE,"request "LPU64" opc %u from %s processed in "
750 "%ldus (%ldus total) trans "LPU64" rc %d/%d\n",
753 lustre_msg_get_opc(request->rq_reqmsg) : 0,
754 libcfs_id2str(request->rq_peer), timediff,
755 cfs_timeval_sub(&work_end, &request->rq_arrival_time,
757 request->rq_transno, request->rq_status,
758 reply ? lustre_msg_get_status(request->rq_repmsg) : -999);
760 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
761 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
762 int opc = opcode_offset(op);
763 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
764 LASSERT(opc < LUSTRE_MAX_OPCODES);
765 lprocfs_counter_add(svc->srv_stats,
766 opc + EXTRA_MAX_OPCODES,
772 ptlrpc_server_free_request(request);
778 ptlrpc_server_handle_reply (struct ptlrpc_service *svc)
780 struct ptlrpc_reply_state *rs;
781 struct obd_export *exp;
782 struct obd_device *obd;
787 spin_lock(&svc->srv_lock);
788 if (list_empty (&svc->srv_reply_queue)) {
789 spin_unlock(&svc->srv_lock);
793 rs = list_entry (svc->srv_reply_queue.next,
794 struct ptlrpc_reply_state, rs_list);
799 LASSERT (rs->rs_difficult);
800 LASSERT (rs->rs_scheduled);
802 list_del_init (&rs->rs_list);
804 /* Disengage from notifiers carefully (lock order - irqrestore below!)*/
805 spin_unlock(&svc->srv_lock);
807 spin_lock (&obd->obd_uncommitted_replies_lock);
808 /* Noop if removed already */
809 list_del_init (&rs->rs_obd_list);
810 spin_unlock (&obd->obd_uncommitted_replies_lock);
812 spin_lock (&exp->exp_lock);
813 /* Noop if removed already */
814 list_del_init (&rs->rs_exp_list);
815 spin_unlock (&exp->exp_lock);
817 spin_lock(&svc->srv_lock);
819 been_handled = rs->rs_handled;
822 nlocks = rs->rs_nlocks; /* atomic "steal", but */
823 rs->rs_nlocks = 0; /* locks still on rs_locks! */
825 if (nlocks == 0 && !been_handled) {
826 /* If we see this, we should already have seen the warning
827 * in mds_steal_ack_locks() */
828 CWARN("All locks stolen from rs %p x"LPD64".t"LPD64
831 rs->rs_xid, rs->rs_transno,
832 lustre_msg_get_opc(rs->rs_msg),
833 libcfs_nid2str(exp->exp_connection->c_peer.nid));
836 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
837 spin_unlock(&svc->srv_lock);
839 if (!been_handled && rs->rs_on_net) {
840 LNetMDUnlink(rs->rs_md_h);
841 /* Ignore return code; we're racing with
846 ldlm_lock_decref(&rs->rs_locks[nlocks],
847 rs->rs_modes[nlocks]);
849 spin_lock(&svc->srv_lock);
852 rs->rs_scheduled = 0;
854 if (!rs->rs_on_net) {
856 svc->srv_n_difficult_replies--;
857 spin_unlock(&svc->srv_lock);
859 class_export_put (exp);
860 rs->rs_export = NULL;
861 ptlrpc_rs_decref (rs);
862 atomic_dec (&svc->srv_outstanding_replies);
866 /* still on the net; callback will schedule */
867 spin_unlock(&svc->srv_lock);
872 /* FIXME make use of timeout later */
874 liblustre_check_services (void *arg)
876 int did_something = 0;
878 struct list_head *tmp, *nxt;
881 /* I'm relying on being single threaded, not to have to lock
882 * ptlrpc_all_services etc */
883 list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
884 struct ptlrpc_service *svc =
885 list_entry (tmp, struct ptlrpc_service, srv_list);
887 if (svc->srv_threads_running != 0) /* I've recursed */
890 /* service threads can block for bulk, so this limits us
891 * (arbitrarily) to recursing 1 stack frame per service.
892 * Note that the problem with recursion is that we have to
893 * unwind completely before our caller can resume. */
895 svc->srv_threads_running++;
898 rc = ptlrpc_server_handle_reply(svc);
899 rc |= ptlrpc_server_handle_request(svc, NULL);
900 rc |= (ptlrpc_server_post_idle_rqbds(svc) > 0);
904 svc->srv_threads_running--;
907 RETURN(did_something);
909 #define ptlrpc_stop_all_threads(s) do {} while (0)
911 #else /* __KERNEL__ */
913 /* Don't use daemonize, it removes fs struct from new thread (bug 418) */
914 void ptlrpc_daemonize(char *name)
916 struct fs_struct *fs = current->fs;
918 atomic_inc(&fs->count);
920 exit_fs(cfs_current());
922 ll_set_fs_pwd(current->fs, init_task.fs->pwdmnt, init_task.fs->pwd);
926 ptlrpc_check_rqbd_pool(struct ptlrpc_service *svc)
928 int avail = svc->srv_nrqbd_receiving;
929 int low_water = test_req_buffer_pressure ? 0 :
930 svc->srv_nbuf_per_group/2;
932 /* NB I'm not locking; just looking. */
934 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
935 * allowed the request history to grow out of control. We could put a
936 * sanity check on that here and cull some history if we need the
939 if (avail <= low_water)
940 ptlrpc_grow_req_bufs(svc);
943 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQBUF_AVAIL_CNTR,
948 ptlrpc_retry_rqbds(void *arg)
950 struct ptlrpc_service *svc = (struct ptlrpc_service *)arg;
952 svc->srv_rqbd_timeout = 0;
956 static int ptlrpc_main(void *arg)
958 struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
959 struct ptlrpc_service *svc = data->svc;
960 struct ptlrpc_thread *thread = data->thread;
961 struct obd_device *dev = data->dev;
962 struct ptlrpc_reply_state *rs;
963 struct lc_watchdog *watchdog;
964 #ifdef WITH_GROUP_INFO
965 struct group_info *ginfo = NULL;
971 ptlrpc_daemonize(data->name);
973 #if defined(HAVE_NODE_TO_CPUMASK) && defined(CONFIG_NUMA)
974 /* we need to do this before any per-thread allocation is done so that
975 * we get the per-thread allocations on local node. bug 7342 */
976 if (svc->srv_cpu_affinity) {
979 for (cpu = 0, num_cpu = 0; cpu < num_possible_cpus(); cpu++) {
980 if (!cpu_online(cpu))
982 if (num_cpu == thread->t_id % num_online_cpus())
986 set_cpus_allowed(cfs_current(), node_to_cpumask(cpu_to_node(cpu)));
990 #ifdef WITH_GROUP_INFO
991 ginfo = groups_alloc(0);
997 set_current_groups(ginfo);
998 put_group_info(ginfo);
1001 if (svc->srv_init != NULL) {
1002 rc = svc->srv_init(thread);
1007 rc = lu_context_init(&env.le_ctx, svc->srv_ctx_tags);
1011 thread->t_env = &env;
1012 env.le_ctx.lc_thread = thread;
1014 /* Alloc reply state structure for this one */
1015 OBD_ALLOC_GFP(rs, svc->srv_max_reply_size, CFS_ALLOC_STD);
1021 /* Record that the thread is running */
1022 thread->t_flags = SVC_RUNNING;
1024 * wake up our creator. Note: @data is invalid after this point,
1025 * because it's allocated on ptlrpc_start_thread() stack.
1027 cfs_waitq_signal(&thread->t_ctl_waitq);
1029 watchdog = lc_watchdog_add(svc->srv_watchdog_timeout, NULL, NULL);
1031 spin_lock(&svc->srv_lock);
1032 svc->srv_threads_running++;
1033 list_add(&rs->rs_list, &svc->srv_free_rs_list);
1034 spin_unlock(&svc->srv_lock);
1035 cfs_waitq_signal(&svc->srv_free_rs_waitq);
1037 CDEBUG(D_NET, "service thread %d (#%d)started\n", thread->t_id,
1038 svc->srv_threads_running);
1040 /* XXX maintain a list of all managed devices: insert here */
1042 while ((thread->t_flags & SVC_STOPPING) == 0 ||
1043 svc->srv_n_difficult_replies != 0) {
1044 /* Don't exit while there are replies to be handled */
1045 struct l_wait_info lwi = LWI_TIMEOUT(svc->srv_rqbd_timeout,
1046 ptlrpc_retry_rqbds, svc);
1048 lc_watchdog_disable(watchdog);
1052 l_wait_event_exclusive (svc->srv_waitq,
1053 ((thread->t_flags & SVC_STOPPING) != 0 &&
1054 svc->srv_n_difficult_replies == 0) ||
1055 (!list_empty(&svc->srv_idle_rqbds) &&
1056 svc->srv_rqbd_timeout == 0) ||
1057 !list_empty (&svc->srv_reply_queue) ||
1058 (!list_empty (&svc->srv_request_queue) &&
1059 (svc->srv_n_difficult_replies == 0 ||
1060 svc->srv_n_active_reqs <
1061 (svc->srv_threads_running - 1))),
1064 lc_watchdog_touch(watchdog);
1066 ptlrpc_check_rqbd_pool(svc);
1068 if ((svc->srv_threads_started < svc->srv_threads_max) &&
1069 (svc->srv_n_active_reqs >= (svc->srv_threads_started - 1))){
1070 /* Ignore return code - we tried... */
1071 ptlrpc_start_thread(dev, svc);
1074 if (!list_empty (&svc->srv_reply_queue))
1075 ptlrpc_server_handle_reply (svc);
1077 /* only handle requests if there are no difficult replies
1078 * outstanding, or I'm not the last thread handling
1080 if (!list_empty (&svc->srv_request_queue) &&
1081 (svc->srv_n_difficult_replies == 0 ||
1082 svc->srv_n_active_reqs < (svc->srv_threads_running - 1))) {
1083 lu_context_enter(&env.le_ctx);
1084 ptlrpc_server_handle_request(svc, thread);
1085 lu_context_exit(&env.le_ctx);
1088 if (!list_empty(&svc->srv_idle_rqbds) &&
1089 ptlrpc_server_post_idle_rqbds(svc) < 0) {
1090 /* I just failed to repost request buffers. Wait
1091 * for a timeout (unless something else happens)
1092 * before I try again */
1093 svc->srv_rqbd_timeout = cfs_time_seconds(1)/10;
1094 CDEBUG(D_RPCTRACE,"Posted buffers: %d\n",
1095 svc->srv_nrqbd_receiving);
1099 lc_watchdog_delete(watchdog);
1103 * deconstruct service specific state created by ptlrpc_start_thread()
1105 if (svc->srv_done != NULL)
1106 svc->srv_done(thread);
1108 lu_context_fini(&env.le_ctx);
1110 CDEBUG(D_NET, "service thread %d exiting: rc %d\n", thread->t_id, rc);
1112 spin_lock(&svc->srv_lock);
1113 svc->srv_threads_running--; /* must know immediately */
1115 thread->t_flags = SVC_STOPPED;
1117 cfs_waitq_signal(&thread->t_ctl_waitq);
1118 spin_unlock(&svc->srv_lock);
1123 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
1124 struct ptlrpc_thread *thread)
1126 struct l_wait_info lwi = { 0 };
1128 spin_lock(&svc->srv_lock);
1129 thread->t_flags = SVC_STOPPING;
1130 spin_unlock(&svc->srv_lock);
1132 cfs_waitq_broadcast(&svc->srv_waitq);
1133 l_wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED),
1136 spin_lock(&svc->srv_lock);
1137 list_del(&thread->t_link);
1138 spin_unlock(&svc->srv_lock);
1140 OBD_FREE_PTR(thread);
1143 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
1145 struct ptlrpc_thread *thread;
1147 spin_lock(&svc->srv_lock);
1148 while (!list_empty(&svc->srv_threads)) {
1149 thread = list_entry(svc->srv_threads.next,
1150 struct ptlrpc_thread, t_link);
1152 spin_unlock(&svc->srv_lock);
1153 ptlrpc_stop_thread(svc, thread);
1154 spin_lock(&svc->srv_lock);
1157 spin_unlock(&svc->srv_lock);
1160 int ptlrpc_start_threads(struct obd_device *dev, struct ptlrpc_service *svc)
1165 LASSERT(svc->srv_threads_min > 0);
1166 for (i = 0; i < svc->srv_threads_min; i++) {
1167 rc = ptlrpc_start_thread(dev, svc);
1169 CERROR("cannot start %s thread #%d: rc %d\n",
1170 svc->srv_thread_name, i, rc);
1171 ptlrpc_stop_all_threads(svc);
1177 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc)
1179 struct l_wait_info lwi = { 0 };
1180 struct ptlrpc_svc_data d;
1181 struct ptlrpc_thread *thread;
1186 CDEBUG(D_RPCTRACE, "%s started %d min %d max %d running %d\n",
1187 svc->srv_name, svc->srv_threads_started, svc->srv_threads_min,
1188 svc->srv_threads_max, svc->srv_threads_running);
1189 if (svc->srv_threads_started >= svc->srv_threads_max)
1192 OBD_ALLOC_PTR(thread);
1195 cfs_waitq_init(&thread->t_ctl_waitq);
1197 spin_lock(&svc->srv_lock);
1198 if (svc->srv_threads_started >= svc->srv_threads_max) {
1199 spin_unlock(&svc->srv_lock);
1200 OBD_FREE_PTR(thread);
1203 list_add(&thread->t_link, &svc->srv_threads);
1204 id = svc->srv_threads_started++;
1205 spin_unlock(&svc->srv_lock);
1208 sprintf(name, "%s_%02d", svc->srv_thread_name, id);
1214 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", name);
1216 /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
1217 * just drop the VM and FILES in ptlrpc_daemonize() right away.
1219 rc = cfs_kernel_thread(ptlrpc_main, &d, CLONE_VM | CLONE_FILES);
1221 CERROR("cannot start thread '%s': rc %d\n", name, rc);
1223 spin_lock(&svc->srv_lock);
1224 list_del(&thread->t_link);
1225 --svc->srv_threads_started;
1226 spin_unlock(&svc->srv_lock);
1228 OBD_FREE(thread, sizeof(*thread));
1231 l_wait_event(thread->t_ctl_waitq,
1232 thread->t_flags & (SVC_RUNNING | SVC_STOPPED), &lwi);
1234 rc = (thread->t_flags & SVC_STOPPED) ? thread->t_id : 0;
1239 int ptlrpc_unregister_service(struct ptlrpc_service *service)
1242 struct l_wait_info lwi;
1243 struct list_head *tmp;
1244 struct ptlrpc_reply_state *rs, *t;
1246 ptlrpc_stop_all_threads(service);
1247 LASSERT(list_empty(&service->srv_threads));
1249 spin_lock (&ptlrpc_all_services_lock);
1250 list_del_init (&service->srv_list);
1251 spin_unlock (&ptlrpc_all_services_lock);
1253 ptlrpc_lprocfs_unregister_service(service);
1255 /* All history will be culled when the next request buffer is
1257 service->srv_max_history_rqbds = 0;
1259 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
1261 rc = LNetClearLazyPortal(service->srv_req_portal);
1264 /* Unlink all the request buffers. This forces a 'final' event with
1265 * its 'unlink' flag set for each posted rqbd */
1266 list_for_each(tmp, &service->srv_active_rqbds) {
1267 struct ptlrpc_request_buffer_desc *rqbd =
1268 list_entry(tmp, struct ptlrpc_request_buffer_desc,
1271 rc = LNetMDUnlink(rqbd->rqbd_md_h);
1272 LASSERT (rc == 0 || rc == -ENOENT);
1275 /* Wait for the network to release any buffers it's currently
1278 spin_lock(&service->srv_lock);
1279 rc = service->srv_nrqbd_receiving;
1280 spin_unlock(&service->srv_lock);
1285 /* Network access will complete in finite time but the HUGE
1286 * timeout lets us CWARN for visibility of sluggish NALs */
1287 lwi = LWI_TIMEOUT(cfs_time_seconds(300), NULL, NULL);
1288 rc = l_wait_event(service->srv_waitq,
1289 service->srv_nrqbd_receiving == 0,
1291 if (rc == -ETIMEDOUT)
1292 CWARN("Service %s waiting for request buffers\n",
1296 /* schedule all outstanding replies to terminate them */
1297 spin_lock(&service->srv_lock);
1298 while (!list_empty(&service->srv_active_replies)) {
1299 struct ptlrpc_reply_state *rs =
1300 list_entry(service->srv_active_replies.next,
1301 struct ptlrpc_reply_state, rs_list);
1302 ptlrpc_schedule_difficult_reply(rs);
1304 spin_unlock(&service->srv_lock);
1306 /* purge the request queue. NB No new replies (rqbds all unlinked)
1307 * and no service threads, so I'm the only thread noodling the
1308 * request queue now */
1309 while (!list_empty(&service->srv_request_queue)) {
1310 struct ptlrpc_request *req =
1311 list_entry(service->srv_request_queue.next,
1312 struct ptlrpc_request,
1315 list_del(&req->rq_list);
1316 service->srv_n_queued_reqs--;
1317 service->srv_n_active_reqs++;
1319 ptlrpc_server_free_request(req);
1321 LASSERT(service->srv_n_queued_reqs == 0);
1322 LASSERT(service->srv_n_active_reqs == 0);
1323 LASSERT(service->srv_n_history_rqbds == 0);
1324 LASSERT(list_empty(&service->srv_active_rqbds));
1326 /* Now free all the request buffers since nothing references them
1328 while (!list_empty(&service->srv_idle_rqbds)) {
1329 struct ptlrpc_request_buffer_desc *rqbd =
1330 list_entry(service->srv_idle_rqbds.next,
1331 struct ptlrpc_request_buffer_desc,
1334 ptlrpc_free_rqbd(rqbd);
1337 /* wait for all outstanding replies to complete (they were
1338 * scheduled having been flagged to abort above) */
1339 while (atomic_read(&service->srv_outstanding_replies) != 0) {
1340 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10), NULL, NULL);
1342 rc = l_wait_event(service->srv_waitq,
1343 !list_empty(&service->srv_reply_queue), &lwi);
1344 LASSERT(rc == 0 || rc == -ETIMEDOUT);
1347 ptlrpc_server_handle_reply(service);
1350 CWARN("Unexpectedly long timeout %p\n", service);
1353 list_for_each_entry_safe(rs, t, &service->srv_free_rs_list, rs_list) {
1354 list_del(&rs->rs_list);
1355 OBD_FREE(rs, service->srv_max_reply_size);
1358 OBD_FREE_PTR(service);
1362 /* Returns 0 if the service is healthy.
1364 * Right now, it just checks to make sure that requests aren't languishing
1365 * in the queue. We'll use this health check to govern whether a node needs
1366 * to be shot, so it's intentionally non-aggressive. */
1367 int ptlrpc_service_health_check(struct ptlrpc_service *svc)
1369 struct ptlrpc_request *request;
1370 struct timeval right_now;
1371 long timediff, cutoff;
1377 spin_lock(&svc->srv_lock);
1379 if (list_empty(&svc->srv_request_queue))
1382 request = list_entry(svc->srv_request_queue.next,
1383 struct ptlrpc_request, rq_list);
1385 do_gettimeofday(&right_now);
1386 timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
1388 cutoff = obd_health_check_timeout;
1390 if (timediff / 1000000 > cutoff) {
1396 spin_unlock(&svc->srv_lock);