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;
75 OBD_ALLOC(rqbd, sizeof (*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) {
87 OBD_FREE(rqbd, sizeof (*rqbd));
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);
113 OBD_FREE (rqbd, sizeof (*rqbd));
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)
260 return ptlrpc_init_svc(c->psc_nbufs, c->psc_bufsize,
261 c->psc_max_req_size, c->psc_max_reply_size,
262 c->psc_req_portal, c->psc_rep_portal,
263 c->psc_watchdog_timeout,
265 prntfn, c->psc_num_threads, c->psc_ctx_tags);
267 EXPORT_SYMBOL(ptlrpc_init_svc_conf);
269 struct ptlrpc_service *
270 ptlrpc_init_svc(int nbufs, int bufsize, int max_req_size, int max_reply_size,
271 int req_portal, int rep_portal, int watchdog_timeout,
272 svc_handler_t handler, char *name,
273 cfs_proc_dir_entry_t *proc_entry,
274 svcreq_printfn_t svcreq_printfn, int num_threads,
278 struct ptlrpc_service *service;
282 LASSERT (bufsize >= max_req_size);
283 LASSERT (ctx_tags != 0);
285 OBD_ALLOC(service, sizeof(*service));
289 /* First initialise enough for early teardown */
291 service->srv_name = name;
292 spin_lock_init(&service->srv_lock);
293 CFS_INIT_LIST_HEAD(&service->srv_threads);
294 cfs_waitq_init(&service->srv_waitq);
296 service->srv_nbuf_per_group = test_req_buffer_pressure ? 1 : nbufs;
297 service->srv_max_req_size = max_req_size;
298 service->srv_buf_size = bufsize;
299 service->srv_rep_portal = rep_portal;
300 service->srv_req_portal = req_portal;
301 service->srv_watchdog_timeout = watchdog_timeout;
302 service->srv_handler = handler;
303 service->srv_request_history_print_fn = svcreq_printfn;
304 service->srv_request_seq = 1; /* valid seq #s start at 1 */
305 service->srv_request_max_cull_seq = 0;
306 service->srv_num_threads = num_threads;
307 service->srv_ctx_tags = ctx_tags;
309 rc = LNetSetLazyPortal(service->srv_req_portal);
312 CFS_INIT_LIST_HEAD(&service->srv_request_queue);
313 CFS_INIT_LIST_HEAD(&service->srv_idle_rqbds);
314 CFS_INIT_LIST_HEAD(&service->srv_active_rqbds);
315 CFS_INIT_LIST_HEAD(&service->srv_history_rqbds);
316 CFS_INIT_LIST_HEAD(&service->srv_request_history);
317 CFS_INIT_LIST_HEAD(&service->srv_active_replies);
318 CFS_INIT_LIST_HEAD(&service->srv_reply_queue);
319 CFS_INIT_LIST_HEAD(&service->srv_free_rs_list);
320 cfs_waitq_init(&service->srv_free_rs_waitq);
322 spin_lock (&ptlrpc_all_services_lock);
323 list_add (&service->srv_list, &ptlrpc_all_services);
324 spin_unlock (&ptlrpc_all_services_lock);
326 /* Now allocate the request buffers */
327 rc = ptlrpc_grow_req_bufs(service);
328 /* We shouldn't be under memory pressure at startup, so
329 * fail if we can't post all our buffers at this time. */
333 /* Now allocate pool of reply buffers */
334 /* Increase max reply size to next power of two */
335 service->srv_max_reply_size = 1;
336 while (service->srv_max_reply_size < max_reply_size)
337 service->srv_max_reply_size <<= 1;
339 if (proc_entry != NULL)
340 ptlrpc_lprocfs_register_service(proc_entry, service);
342 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
343 service->srv_name, service->srv_req_portal);
347 ptlrpc_unregister_service(service);
351 static void __ptlrpc_server_free_request(struct ptlrpc_request *req)
353 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
355 list_del(&req->rq_list);
357 if (req->rq_reply_state != NULL) {
358 ptlrpc_rs_decref(req->rq_reply_state);
359 req->rq_reply_state = NULL;
362 if (req != &rqbd->rqbd_req) {
363 /* NB request buffers use an embedded
364 * req if the incoming req unlinked the
365 * MD; this isn't one of them! */
366 OBD_FREE(req, sizeof(*req));
371 ptlrpc_server_free_request(struct ptlrpc_request *req)
373 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
374 struct ptlrpc_service *svc = rqbd->rqbd_service;
376 struct list_head *tmp;
377 struct list_head *nxt;
379 spin_lock(&svc->srv_lock);
381 svc->srv_n_active_reqs--;
382 list_add(&req->rq_list, &rqbd->rqbd_reqs);
384 refcount = --(rqbd->rqbd_refcount);
386 /* request buffer is now idle: add to history */
387 list_del(&rqbd->rqbd_list);
388 list_add_tail(&rqbd->rqbd_list, &svc->srv_history_rqbds);
389 svc->srv_n_history_rqbds++;
391 /* cull some history?
392 * I expect only about 1 or 2 rqbds need to be recycled here */
393 while (svc->srv_n_history_rqbds > svc->srv_max_history_rqbds) {
394 rqbd = list_entry(svc->srv_history_rqbds.next,
395 struct ptlrpc_request_buffer_desc,
398 list_del(&rqbd->rqbd_list);
399 svc->srv_n_history_rqbds--;
401 /* remove rqbd's reqs from svc's req history while
402 * I've got the service lock */
403 list_for_each(tmp, &rqbd->rqbd_reqs) {
404 req = list_entry(tmp, struct ptlrpc_request,
406 /* Track the highest culled req seq */
407 if (req->rq_history_seq >
408 svc->srv_request_max_cull_seq)
409 svc->srv_request_max_cull_seq =
411 list_del(&req->rq_history_list);
414 spin_unlock(&svc->srv_lock);
416 list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
417 req = list_entry(rqbd->rqbd_reqs.next,
418 struct ptlrpc_request,
420 __ptlrpc_server_free_request(req);
423 spin_lock(&svc->srv_lock);
425 /* schedule request buffer for re-use.
426 * NB I can only do this after I've disposed of their
427 * reqs; particularly the embedded req */
428 list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
430 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
431 /* If we are low on memory, we are not interested in
433 list_del(&req->rq_history_list);
434 __ptlrpc_server_free_request(req);
437 spin_unlock(&svc->srv_lock);
441 /* This function makes sure dead exports are evicted in a timely manner.
442 This function is only called when some export receives a message (i.e.,
443 the network is up.) */
444 static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
446 struct obd_export *oldest_exp;
453 /* Compensate for slow machines, etc, by faking our request time
454 into the future. Although this can break the strict time-ordering
455 of the list, we can be really lazy here - we don't have to evict
456 at the exact right moment. Eventually, all silent exports
457 will make it to the top of the list. */
458 exp->exp_last_request_time = max(exp->exp_last_request_time,
459 (time_t)CURRENT_SECONDS + extra_delay);
461 CDEBUG(D_INFO, "updating export %s at %ld\n",
462 exp->exp_client_uuid.uuid,
463 exp->exp_last_request_time);
465 /* exports may get disconnected from the chain even though the
466 export has references, so we must keep the spin lock while
467 manipulating the lists */
468 spin_lock(&exp->exp_obd->obd_dev_lock);
470 if (list_empty(&exp->exp_obd_chain_timed)) {
471 /* this one is not timed */
472 spin_unlock(&exp->exp_obd->obd_dev_lock);
477 list_move_tail(&exp->exp_obd_chain_timed,
478 &exp->exp_obd->obd_exports_timed);
480 oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
481 struct obd_export, exp_obd_chain_timed);
482 oldest_time = oldest_exp->exp_last_request_time;
483 spin_unlock(&exp->exp_obd->obd_dev_lock);
485 if (exp->exp_obd->obd_recovering) {
486 /* be nice to everyone during recovery */
491 /* Note - racing to start/reset the obd_eviction timer is safe */
492 if (exp->exp_obd->obd_eviction_timer == 0) {
493 /* Check if the oldest entry is expired. */
494 if (CURRENT_SECONDS > (oldest_time +
495 (3 * obd_timeout / 2) + extra_delay)) {
496 /* We need a second timer, in case the net was down and
497 * it just came back. Since the pinger may skip every
498 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
499 * we better wait for 3. */
500 exp->exp_obd->obd_eviction_timer = CURRENT_SECONDS +
502 CDEBUG(D_HA, "%s: Think about evicting %s from %ld\n",
503 exp->exp_obd->obd_name, obd_export_nid2str(exp),
507 if (CURRENT_SECONDS > (exp->exp_obd->obd_eviction_timer +
509 /* The evictor won't evict anyone who we've heard from
510 * recently, so we don't have to check before we start
512 if (!ping_evictor_wake(exp))
513 exp->exp_obd->obd_eviction_timer = 0;
521 ptlrpc_server_handle_request(struct ptlrpc_service *svc,
522 struct ptlrpc_thread *thread)
524 struct ptlrpc_request *request;
525 struct timeval work_start;
526 struct timeval work_end;
533 spin_lock(&svc->srv_lock);
534 if (list_empty (&svc->srv_request_queue) ||
535 (svc->srv_n_difficult_replies != 0 &&
536 svc->srv_n_active_reqs >= (svc->srv_nthreads - 1))) {
537 /* If all the other threads are handling requests, I must
538 * remain free to handle any 'difficult' reply that might
540 spin_unlock(&svc->srv_lock);
544 request = list_entry (svc->srv_request_queue.next,
545 struct ptlrpc_request, rq_list);
546 list_del_init (&request->rq_list);
547 svc->srv_n_queued_reqs--;
548 svc->srv_n_active_reqs++;
550 spin_unlock(&svc->srv_lock);
552 do_gettimeofday(&work_start);
553 timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
554 if (svc->srv_stats != NULL) {
555 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
557 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
558 svc->srv_n_queued_reqs);
559 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
560 svc->srv_n_active_reqs);
564 /* Clear request swab mask; this is a new request */
565 request->rq_req_swab_mask = 0;
567 rc = lustre_unpack_msg(request->rq_reqmsg, request->rq_reqlen);
569 CERROR ("error unpacking request: ptl %d from %s"
570 " xid "LPU64"\n", svc->srv_req_portal,
571 libcfs_id2str(request->rq_peer), request->rq_xid);
575 rc = lustre_unpack_ptlrpc_body(request->rq_reqmsg);
577 CERROR ("error unpacking ptlrpc body: ptl %d from %s"
578 " xid "LPU64"\n", svc->srv_req_portal,
579 libcfs_id2str(request->rq_peer), request->rq_xid);
584 if (lustre_msg_get_type(request->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
585 CERROR("wrong packet type received (type=%u) from %s\n",
586 lustre_msg_get_type(request->rq_reqmsg),
587 libcfs_id2str(request->rq_peer));
591 CDEBUG(D_NET, "got req "LPD64"\n", request->rq_xid);
593 request->rq_svc_thread = thread;
594 request->rq_export = class_conn2export(
595 lustre_msg_get_handle(request->rq_reqmsg));
597 if (request->rq_export) {
598 if (lustre_msg_get_conn_cnt(request->rq_reqmsg) <
599 request->rq_export->exp_conn_cnt) {
600 DEBUG_REQ(D_ERROR, request,
601 "DROPPING req from old connection %d < %d",
602 lustre_msg_get_conn_cnt(request->rq_reqmsg),
603 request->rq_export->exp_conn_cnt);
606 if (request->rq_export->exp_obd &&
607 request->rq_export->exp_obd->obd_fail) {
608 /* Failing over, don't handle any more reqs, send
609 error response instead. */
610 CDEBUG(D_HA, "Dropping req %p for failed obd %s\n",
611 request, request->rq_export->exp_obd->obd_name);
612 request->rq_status = -ENODEV;
613 ptlrpc_error(request);
617 ptlrpc_update_export_timer(request->rq_export, timediff/500000);
620 /* Discard requests queued for longer than my timeout. If the
621 * client's timeout is similar to mine, she'll be timing out this
622 * REQ anyway (bug 1502) */
623 if (timediff / 1000000 > (long)obd_timeout) {
624 CERROR("Dropping timed-out opc %d request from %s"
625 ": %ld seconds old\n",
626 lustre_msg_get_opc(request->rq_reqmsg),
627 libcfs_id2str(request->rq_peer),
632 request->rq_phase = RQ_PHASE_INTERPRET;
634 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
635 "%s:%s+%d:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
636 (request->rq_export ?
637 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
638 (request->rq_export ?
639 atomic_read(&request->rq_export->exp_refcount) : -99),
640 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
641 libcfs_id2str(request->rq_peer),
642 lustre_msg_get_opc(request->rq_reqmsg));
644 rc = svc->srv_handler(request);
646 request->rq_phase = RQ_PHASE_COMPLETE;
648 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
649 "%s:%s+%d:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
650 (request->rq_export ?
651 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
652 (request->rq_export ?
653 atomic_read(&request->rq_export->exp_refcount) : -99),
654 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
655 libcfs_id2str(request->rq_peer),
656 lustre_msg_get_opc(request->rq_reqmsg));
659 if (request->rq_export != NULL)
660 class_export_put(request->rq_export);
663 do_gettimeofday(&work_end);
665 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
667 if (timediff / 1000000 > (long)obd_timeout)
668 CERROR("request "LPU64" opc %u from %s processed in %lds "
669 "trans "LPU64" rc %d/%d\n",
670 request->rq_xid, lustre_msg_get_opc(request->rq_reqmsg),
671 libcfs_id2str(request->rq_peer),
672 cfs_timeval_sub(&work_end, &request->rq_arrival_time,
675 lustre_msg_get_transno(request->rq_repmsg) :
676 request->rq_transno, request->rq_status,
678 lustre_msg_get_status(request->rq_repmsg) :
681 CDEBUG(D_HA, "request "LPU64" opc %u from %s processed in "
682 "%ldus (%ldus total) trans "LPU64" rc %d/%d\n",
683 request->rq_xid, lustre_msg_get_opc(request->rq_reqmsg),
684 libcfs_id2str(request->rq_peer), timediff,
685 cfs_timeval_sub(&work_end, &request->rq_arrival_time,
687 request->rq_transno, request->rq_status,
689 lustre_msg_get_status(request->rq_repmsg) :
692 if (svc->srv_stats != NULL) {
693 int opc = opcode_offset(lustre_msg_get_opc(request->rq_reqmsg));
695 LASSERT(opc < LUSTRE_MAX_OPCODES);
696 lprocfs_counter_add(svc->srv_stats,
697 opc + PTLRPC_LAST_CNTR,
702 ptlrpc_server_free_request(request);
708 ptlrpc_server_handle_reply (struct ptlrpc_service *svc)
710 struct ptlrpc_reply_state *rs;
711 struct obd_export *exp;
712 struct obd_device *obd;
717 spin_lock(&svc->srv_lock);
718 if (list_empty (&svc->srv_reply_queue)) {
719 spin_unlock(&svc->srv_lock);
723 rs = list_entry (svc->srv_reply_queue.next,
724 struct ptlrpc_reply_state, rs_list);
729 LASSERT (rs->rs_difficult);
730 LASSERT (rs->rs_scheduled);
732 list_del_init (&rs->rs_list);
734 /* Disengage from notifiers carefully (lock order - irqrestore below!)*/
735 spin_unlock(&svc->srv_lock);
737 spin_lock (&obd->obd_uncommitted_replies_lock);
738 /* Noop if removed already */
739 list_del_init (&rs->rs_obd_list);
740 spin_unlock (&obd->obd_uncommitted_replies_lock);
742 spin_lock (&exp->exp_lock);
743 /* Noop if removed already */
744 list_del_init (&rs->rs_exp_list);
745 spin_unlock (&exp->exp_lock);
747 spin_lock(&svc->srv_lock);
749 been_handled = rs->rs_handled;
752 nlocks = rs->rs_nlocks; /* atomic "steal", but */
753 rs->rs_nlocks = 0; /* locks still on rs_locks! */
755 if (nlocks == 0 && !been_handled) {
756 /* If we see this, we should already have seen the warning
757 * in mds_steal_ack_locks() */
758 CWARN("All locks stolen from rs %p x"LPD64".t"LPD64
761 rs->rs_xid, rs->rs_transno,
762 lustre_msg_get_opc(rs->rs_msg),
763 libcfs_nid2str(exp->exp_connection->c_peer.nid));
766 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
767 spin_unlock(&svc->srv_lock);
769 if (!been_handled && rs->rs_on_net) {
770 LNetMDUnlink(rs->rs_md_h);
771 /* Ignore return code; we're racing with
776 ldlm_lock_decref(&rs->rs_locks[nlocks],
777 rs->rs_modes[nlocks]);
779 spin_lock(&svc->srv_lock);
782 rs->rs_scheduled = 0;
784 if (!rs->rs_on_net) {
786 svc->srv_n_difficult_replies--;
787 spin_unlock(&svc->srv_lock);
789 class_export_put (exp);
790 rs->rs_export = NULL;
791 ptlrpc_rs_decref (rs);
792 atomic_dec (&svc->srv_outstanding_replies);
796 /* still on the net; callback will schedule */
797 spin_unlock(&svc->srv_lock);
802 /* FIXME make use of timeout later */
804 liblustre_check_services (void *arg)
806 int did_something = 0;
808 struct list_head *tmp, *nxt;
811 /* I'm relying on being single threaded, not to have to lock
812 * ptlrpc_all_services etc */
813 list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
814 struct ptlrpc_service *svc =
815 list_entry (tmp, struct ptlrpc_service, srv_list);
817 if (svc->srv_nthreads != 0) /* I've recursed */
820 /* service threads can block for bulk, so this limits us
821 * (arbitrarily) to recursing 1 stack frame per service.
822 * Note that the problem with recursion is that we have to
823 * unwind completely before our caller can resume. */
828 rc = ptlrpc_server_handle_reply(svc);
829 rc |= ptlrpc_server_handle_request(svc, NULL);
830 rc |= (ptlrpc_server_post_idle_rqbds(svc) > 0);
837 RETURN(did_something);
839 #define ptlrpc_stop_all_threads(s) do {} while (0)
841 #else /* __KERNEL__ */
843 /* Don't use daemonize, it removes fs struct from new thread (bug 418) */
844 void ptlrpc_daemonize(char *name)
846 struct fs_struct *fs = current->fs;
848 atomic_inc(&fs->count);
850 exit_fs(cfs_current());
852 ll_set_fs_pwd(current->fs, init_task.fs->pwdmnt, init_task.fs->pwd);
856 ptlrpc_check_rqbd_pool(struct ptlrpc_service *svc)
858 int avail = svc->srv_nrqbd_receiving;
859 int low_water = test_req_buffer_pressure ? 0 :
860 svc->srv_nbuf_per_group/2;
862 /* NB I'm not locking; just looking. */
864 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
865 * allowed the request history to grow out of control. We could put a
866 * sanity check on that here and cull some history if we need the
869 if (avail <= low_water)
870 ptlrpc_grow_req_bufs(svc);
873 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQBUF_AVAIL_CNTR,
878 ptlrpc_retry_rqbds(void *arg)
880 struct ptlrpc_service *svc = (struct ptlrpc_service *)arg;
882 svc->srv_rqbd_timeout = 0;
886 static int ptlrpc_main(void *arg)
888 struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
889 struct ptlrpc_service *svc = data->svc;
890 struct ptlrpc_thread *thread = data->thread;
891 struct ptlrpc_reply_state *rs;
892 struct lc_watchdog *watchdog;
893 #ifdef WITH_GROUP_INFO
894 struct group_info *ginfo = NULL;
896 struct lu_context ctx;
900 ptlrpc_daemonize(data->name);
902 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,9) && CONFIG_NUMA
903 /* we need to do this before any per-thread allocation is done so that
904 * we get the per-thread allocations on local node. bug 7342 */
905 if (svc->srv_cpu_affinity) {
908 for (cpu = 0, num_cpu = 0; cpu < NR_CPUS; cpu++) {
909 if (!cpu_online(cpu))
911 if (num_cpu == thread->t_id % num_online_cpus())
915 set_cpus_allowed(cfs_current(), node_to_cpumask(cpu_to_node(cpu)));
919 #ifdef WITH_GROUP_INFO
920 ginfo = groups_alloc(0);
926 set_current_groups(ginfo);
927 put_group_info(ginfo);
930 if (svc->srv_init != NULL) {
931 rc = svc->srv_init(thread);
936 rc = lu_context_init(&ctx, svc->srv_ctx_tags);
940 thread->t_ctx = &ctx;
941 ctx.lc_thread = thread;
943 /* Alloc reply state structure for this one */
944 OBD_ALLOC_GFP(rs, svc->srv_max_reply_size, CFS_ALLOC_STD);
950 /* Record that the thread is running */
951 thread->t_flags = SVC_RUNNING;
953 * wake up our creator. Note: @data is invalid after this point,
954 * because it's allocated on ptlrpc_start_thread() stack.
956 cfs_waitq_signal(&thread->t_ctl_waitq);
958 watchdog = lc_watchdog_add(svc->srv_watchdog_timeout, NULL, NULL);
960 spin_lock(&svc->srv_lock);
962 list_add(&rs->rs_list, &svc->srv_free_rs_list);
963 spin_unlock(&svc->srv_lock);
964 cfs_waitq_signal(&svc->srv_free_rs_waitq);
966 CDEBUG(D_NET, "service thread %d started\n", thread->t_id);
968 /* XXX maintain a list of all managed devices: insert here */
970 while ((thread->t_flags & SVC_STOPPING) == 0 ||
971 svc->srv_n_difficult_replies != 0) {
972 /* Don't exit while there are replies to be handled */
973 struct l_wait_info lwi = LWI_TIMEOUT(svc->srv_rqbd_timeout,
974 ptlrpc_retry_rqbds, svc);
976 lc_watchdog_disable(watchdog);
978 l_wait_event_exclusive (svc->srv_waitq,
979 ((thread->t_flags & SVC_STOPPING) != 0 &&
980 svc->srv_n_difficult_replies == 0) ||
981 (!list_empty(&svc->srv_idle_rqbds) &&
982 svc->srv_rqbd_timeout == 0) ||
983 !list_empty (&svc->srv_reply_queue) ||
984 (!list_empty (&svc->srv_request_queue) &&
985 (svc->srv_n_difficult_replies == 0 ||
986 svc->srv_n_active_reqs <
987 (svc->srv_nthreads - 1))),
990 lc_watchdog_touch(watchdog);
992 ptlrpc_check_rqbd_pool(svc);
994 if (!list_empty (&svc->srv_reply_queue))
995 ptlrpc_server_handle_reply (svc);
997 /* only handle requests if there are no difficult replies
998 * outstanding, or I'm not the last thread handling
1000 if (!list_empty (&svc->srv_request_queue) &&
1001 (svc->srv_n_difficult_replies == 0 ||
1002 svc->srv_n_active_reqs < (svc->srv_nthreads - 1))) {
1003 lu_context_enter(&ctx);
1004 ptlrpc_server_handle_request(svc, thread);
1005 lu_context_exit(&ctx);
1008 if (!list_empty(&svc->srv_idle_rqbds) &&
1009 ptlrpc_server_post_idle_rqbds(svc) < 0) {
1010 /* I just failed to repost request buffers. Wait
1011 * for a timeout (unless something else happens)
1012 * before I try again */
1013 svc->srv_rqbd_timeout = cfs_time_seconds(1)/10;
1014 CDEBUG(D_RPCTRACE,"Posted buffers: %d\n",
1015 svc->srv_nrqbd_receiving);
1019 lc_watchdog_delete(watchdog);
1023 * deconstruct service specific state created by ptlrpc_start_thread()
1025 if (svc->srv_done != NULL)
1026 svc->srv_done(thread);
1029 lu_context_fini(&ctx);
1031 CDEBUG(D_NET, "service thread %d exiting: rc %d\n", thread->t_id, rc);
1033 spin_lock(&svc->srv_lock);
1034 svc->srv_nthreads--; /* must know immediately */
1036 thread->t_flags = SVC_STOPPED;
1038 cfs_waitq_signal(&thread->t_ctl_waitq);
1039 spin_unlock(&svc->srv_lock);
1044 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
1045 struct ptlrpc_thread *thread)
1047 struct l_wait_info lwi = { 0 };
1049 spin_lock(&svc->srv_lock);
1050 thread->t_flags = SVC_STOPPING;
1051 spin_unlock(&svc->srv_lock);
1053 cfs_waitq_broadcast(&svc->srv_waitq);
1054 l_wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED),
1057 spin_lock(&svc->srv_lock);
1058 list_del(&thread->t_link);
1059 spin_unlock(&svc->srv_lock);
1061 OBD_FREE(thread, sizeof(*thread));
1064 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
1066 struct ptlrpc_thread *thread;
1068 spin_lock(&svc->srv_lock);
1069 while (!list_empty(&svc->srv_threads)) {
1070 thread = list_entry(svc->srv_threads.next,
1071 struct ptlrpc_thread, t_link);
1073 spin_unlock(&svc->srv_lock);
1074 ptlrpc_stop_thread(svc, thread);
1075 spin_lock(&svc->srv_lock);
1078 spin_unlock(&svc->srv_lock);
1081 /* @base_name should be 11 characters or less - 3 will be added on */
1082 int ptlrpc_start_threads(struct obd_device *dev, struct ptlrpc_service *svc,
1088 LASSERT(svc->srv_num_threads > 0);
1089 for (i = 0; i < svc->srv_num_threads; i++) {
1091 sprintf(name, "%s_%02d", base_name, i);
1092 rc = ptlrpc_start_thread(dev, svc, name, i);
1094 CERROR("cannot start %s thread #%d: rc %d\n", base_name,
1096 ptlrpc_stop_all_threads(svc);
1102 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc,
1105 struct l_wait_info lwi = { 0 };
1106 struct ptlrpc_svc_data d;
1107 struct ptlrpc_thread *thread;
1111 OBD_ALLOC(thread, sizeof(*thread));
1114 cfs_waitq_init(&thread->t_ctl_waitq);
1117 spin_lock(&svc->srv_lock);
1118 list_add(&thread->t_link, &svc->srv_threads);
1119 spin_unlock(&svc->srv_lock);
1126 /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
1127 * just drop the VM and FILES in ptlrpc_daemonize() right away.
1129 rc = cfs_kernel_thread(ptlrpc_main, &d, CLONE_VM | CLONE_FILES);
1131 CERROR("cannot start thread '%s': rc %d\n", name, rc);
1133 spin_lock(&svc->srv_lock);
1134 list_del(&thread->t_link);
1135 spin_unlock(&svc->srv_lock);
1137 OBD_FREE(thread, sizeof(*thread));
1140 l_wait_event(thread->t_ctl_waitq,
1141 thread->t_flags & (SVC_RUNNING | SVC_STOPPED), &lwi);
1143 rc = (thread->t_flags & SVC_STOPPED) ? thread->t_id : 0;
1148 int ptlrpc_unregister_service(struct ptlrpc_service *service)
1151 struct l_wait_info lwi;
1152 struct list_head *tmp;
1153 struct ptlrpc_reply_state *rs, *t;
1155 ptlrpc_stop_all_threads(service);
1156 LASSERT(list_empty(&service->srv_threads));
1158 spin_lock (&ptlrpc_all_services_lock);
1159 list_del_init (&service->srv_list);
1160 spin_unlock (&ptlrpc_all_services_lock);
1162 ptlrpc_lprocfs_unregister_service(service);
1164 /* All history will be culled when the next request buffer is
1166 service->srv_max_history_rqbds = 0;
1168 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
1170 rc = LNetClearLazyPortal(service->srv_req_portal);
1173 /* Unlink all the request buffers. This forces a 'final' event with
1174 * its 'unlink' flag set for each posted rqbd */
1175 list_for_each(tmp, &service->srv_active_rqbds) {
1176 struct ptlrpc_request_buffer_desc *rqbd =
1177 list_entry(tmp, struct ptlrpc_request_buffer_desc,
1180 rc = LNetMDUnlink(rqbd->rqbd_md_h);
1181 LASSERT (rc == 0 || rc == -ENOENT);
1184 /* Wait for the network to release any buffers it's currently
1187 spin_lock(&service->srv_lock);
1188 rc = service->srv_nrqbd_receiving;
1189 spin_unlock(&service->srv_lock);
1194 /* Network access will complete in finite time but the HUGE
1195 * timeout lets us CWARN for visibility of sluggish NALs */
1196 lwi = LWI_TIMEOUT(cfs_time_seconds(300), NULL, NULL);
1197 rc = l_wait_event(service->srv_waitq,
1198 service->srv_nrqbd_receiving == 0,
1200 if (rc == -ETIMEDOUT)
1201 CWARN("Service %s waiting for request buffers\n",
1205 /* schedule all outstanding replies to terminate them */
1206 spin_lock(&service->srv_lock);
1207 while (!list_empty(&service->srv_active_replies)) {
1208 struct ptlrpc_reply_state *rs =
1209 list_entry(service->srv_active_replies.next,
1210 struct ptlrpc_reply_state, rs_list);
1211 ptlrpc_schedule_difficult_reply(rs);
1213 spin_unlock(&service->srv_lock);
1215 /* purge the request queue. NB No new replies (rqbds all unlinked)
1216 * and no service threads, so I'm the only thread noodling the
1217 * request queue now */
1218 while (!list_empty(&service->srv_request_queue)) {
1219 struct ptlrpc_request *req =
1220 list_entry(service->srv_request_queue.next,
1221 struct ptlrpc_request,
1224 list_del(&req->rq_list);
1225 service->srv_n_queued_reqs--;
1226 service->srv_n_active_reqs++;
1228 ptlrpc_server_free_request(req);
1230 LASSERT(service->srv_n_queued_reqs == 0);
1231 LASSERT(service->srv_n_active_reqs == 0);
1232 LASSERT(service->srv_n_history_rqbds == 0);
1233 LASSERT(list_empty(&service->srv_active_rqbds));
1235 /* Now free all the request buffers since nothing references them
1237 while (!list_empty(&service->srv_idle_rqbds)) {
1238 struct ptlrpc_request_buffer_desc *rqbd =
1239 list_entry(service->srv_idle_rqbds.next,
1240 struct ptlrpc_request_buffer_desc,
1243 ptlrpc_free_rqbd(rqbd);
1246 /* wait for all outstanding replies to complete (they were
1247 * scheduled having been flagged to abort above) */
1248 while (atomic_read(&service->srv_outstanding_replies) != 0) {
1249 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10), NULL, NULL);
1251 rc = l_wait_event(service->srv_waitq,
1252 !list_empty(&service->srv_reply_queue), &lwi);
1253 LASSERT(rc == 0 || rc == -ETIMEDOUT);
1256 ptlrpc_server_handle_reply(service);
1259 CWARN("Unexpectedly long timeout %p\n", service);
1262 list_for_each_entry_safe(rs, t, &service->srv_free_rs_list, rs_list) {
1263 list_del(&rs->rs_list);
1264 OBD_FREE(rs, service->srv_max_reply_size);
1267 OBD_FREE(service, sizeof(*service));
1271 /* Returns 0 if the service is healthy.
1273 * Right now, it just checks to make sure that requests aren't languishing
1274 * in the queue. We'll use this health check to govern whether a node needs
1275 * to be shot, so it's intentionally non-aggressive. */
1276 int ptlrpc_service_health_check(struct ptlrpc_service *svc)
1278 struct ptlrpc_request *request;
1279 struct timeval right_now;
1280 long timediff, cutoff;
1286 spin_lock(&svc->srv_lock);
1288 if (list_empty(&svc->srv_request_queue))
1291 request = list_entry(svc->srv_request_queue.next,
1292 struct ptlrpc_request, rq_list);
1294 do_gettimeofday(&right_now);
1295 timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
1297 cutoff = obd_health_check_timeout;
1299 if (timediff / 1000000 > cutoff) {
1305 spin_unlock(&svc->srv_lock);