1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
33 * Copyright (c) 2011 Whamcloud, Inc.
36 * This file is part of Lustre, http://www.lustre.org/
37 * Lustre is a trademark of Sun Microsystems, Inc.
39 * lustre/ptlrpc/recov_thread.c
41 * OST<->MDS recovery logging thread.
42 * Invariants in implementation:
43 * - we do not share logs among different OST<->MDS connections, so that
44 * if an OST or MDS fails it need only look at log(s) relevant to itself
46 * Author: Andreas Dilger <adilger@clusterfs.com>
47 * Yury Umanets <yury.umanets@sun.com>
48 * Alexey Lyashkov <alexey.lyashkov@sun.com>
51 #define DEBUG_SUBSYSTEM S_LOG
54 # define EXPORT_SYMTAB
58 # include <libcfs/libcfs.h>
60 # include <libcfs/list.h>
61 # include <liblustre.h>
64 #include <obd_class.h>
65 #include <obd_support.h>
66 #include <obd_class.h>
67 #include <lustre_net.h>
68 #include <lnet/types.h>
69 #include <libcfs/list.h>
70 #include <lustre_log.h>
71 #include "ptlrpc_internal.h"
73 static cfs_atomic_t llcd_count = CFS_ATOMIC_INIT(0);
74 static cfs_mem_cache_t *llcd_cache = NULL;
78 LLOG_LCM_FL_START = 1 << 0,
79 LLOG_LCM_FL_EXIT = 1 << 1
82 struct llcd_async_args {
83 struct llog_canceld_ctxt *la_ctxt;
86 static void llcd_print(struct llog_canceld_ctxt *llcd,
87 const char *func, int line)
89 CDEBUG(D_RPCTRACE, "Llcd (%p) at %s:%d:\n", llcd, func, line);
90 CDEBUG(D_RPCTRACE, " size: %d\n", llcd->llcd_size);
91 CDEBUG(D_RPCTRACE, " ctxt: %p\n", llcd->llcd_ctxt);
92 CDEBUG(D_RPCTRACE, " lcm : %p\n", llcd->llcd_lcm);
93 CDEBUG(D_RPCTRACE, " cookiebytes : %d\n", llcd->llcd_cookiebytes);
97 * Allocate new llcd from cache, init it and return to caller.
98 * Bumps number of objects allocated.
100 static struct llog_canceld_ctxt *llcd_alloc(struct llog_commit_master *lcm)
102 struct llog_canceld_ctxt *llcd;
105 LASSERT(lcm != NULL);
108 * We want to send one page of cookies with rpc header. This buffer
109 * will be assigned later to the rpc, this is why we preserve the
110 * space for rpc header.
112 size = CFS_PAGE_SIZE - lustre_msg_size(LUSTRE_MSG_MAGIC_V2, 1, NULL);
113 overhead = offsetof(struct llog_canceld_ctxt, llcd_cookies);
114 OBD_SLAB_ALLOC(llcd, llcd_cache, CFS_ALLOC_STD, size + overhead);
118 CFS_INIT_LIST_HEAD(&llcd->llcd_list);
119 llcd->llcd_cookiebytes = 0;
120 llcd->llcd_size = size;
122 cfs_spin_lock(&lcm->lcm_lock);
123 llcd->llcd_lcm = lcm;
124 cfs_atomic_inc(&lcm->lcm_count);
125 cfs_list_add_tail(&llcd->llcd_list, &lcm->lcm_llcds);
126 cfs_spin_unlock(&lcm->lcm_lock);
127 cfs_atomic_inc(&llcd_count);
129 CDEBUG(D_RPCTRACE, "Alloc llcd %p on lcm %p (%d)\n",
130 llcd, lcm, cfs_atomic_read(&lcm->lcm_count));
136 * Returns passed llcd to cache.
138 static void llcd_free(struct llog_canceld_ctxt *llcd)
140 struct llog_commit_master *lcm = llcd->llcd_lcm;
144 if (cfs_atomic_read(&lcm->lcm_count) == 0) {
145 CERROR("Invalid llcd free %p\n", llcd);
146 llcd_print(llcd, __FUNCTION__, __LINE__);
149 cfs_spin_lock(&lcm->lcm_lock);
150 LASSERT(!cfs_list_empty(&llcd->llcd_list));
151 cfs_list_del_init(&llcd->llcd_list);
152 cfs_atomic_dec(&lcm->lcm_count);
153 cfs_spin_unlock(&lcm->lcm_lock);
155 CDEBUG(D_RPCTRACE, "Free llcd %p on lcm %p (%d)\n",
156 llcd, lcm, cfs_atomic_read(&lcm->lcm_count));
159 LASSERT(cfs_atomic_read(&llcd_count) > 0);
160 cfs_atomic_dec(&llcd_count);
162 size = offsetof(struct llog_canceld_ctxt, llcd_cookies) +
164 OBD_SLAB_FREE(llcd, llcd_cache, size);
168 * Checks if passed cookie fits into llcd free space buffer. Returns
169 * 1 if yes and 0 otherwise.
172 llcd_fit(struct llog_canceld_ctxt *llcd, struct llog_cookie *cookies)
174 return (llcd->llcd_size - llcd->llcd_cookiebytes >= sizeof(*cookies));
178 * Copy passed @cookies to @llcd.
181 llcd_copy(struct llog_canceld_ctxt *llcd, struct llog_cookie *cookies)
183 LASSERT(llcd_fit(llcd, cookies));
184 memcpy((char *)llcd->llcd_cookies + llcd->llcd_cookiebytes,
185 cookies, sizeof(*cookies));
186 llcd->llcd_cookiebytes += sizeof(*cookies);
190 * Llcd completion function. Called uppon llcd send finish regardless
191 * sending result. Error is passed in @rc. Note, that this will be called
192 * in cleanup time when all inflight rpcs aborted.
195 llcd_interpret(const struct lu_env *env,
196 struct ptlrpc_request *req, void *args, int rc)
198 struct llcd_async_args *la = args;
199 struct llog_canceld_ctxt *llcd = la->la_ctxt;
201 CDEBUG(D_RPCTRACE, "Sent llcd %p (%d) - killing it\n", llcd, rc);
207 * Send @llcd to remote node. Free llcd uppon completion or error. Sending
208 * is performed in async style so this function will return asap without
211 static int llcd_send(struct llog_canceld_ctxt *llcd)
213 char *bufs[2] = { NULL, (char *)llcd->llcd_cookies };
214 struct obd_import *import = NULL;
215 struct llog_commit_master *lcm;
216 struct llcd_async_args *la;
217 struct ptlrpc_request *req;
218 struct llog_ctxt *ctxt;
222 ctxt = llcd->llcd_ctxt;
224 CERROR("Invalid llcd with NULL ctxt found (%p)\n",
226 llcd_print(llcd, __FUNCTION__, __LINE__);
229 LASSERT_SEM_LOCKED(&ctxt->loc_sem);
231 if (llcd->llcd_cookiebytes == 0)
234 lcm = llcd->llcd_lcm;
237 * Check if we're in exit stage. Do not send llcd in
240 if (cfs_test_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags))
241 GOTO(exit, rc = -ENODEV);
243 CDEBUG(D_RPCTRACE, "Sending llcd %p\n", llcd);
245 import = llcd->llcd_ctxt->loc_imp;
246 if (!import || (import == LP_POISON) ||
247 (import->imp_client == LP_POISON)) {
248 CERROR("Invalid import %p for llcd %p\n",
250 GOTO(exit, rc = -ENODEV);
253 OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_RECOV, 10);
256 * No need to get import here as it is already done in
257 * llog_receptor_accept().
259 req = ptlrpc_request_alloc(import, &RQF_LOG_CANCEL);
261 CERROR("Can't allocate request for sending llcd %p\n",
263 GOTO(exit, rc = -ENOMEM);
265 req_capsule_set_size(&req->rq_pill, &RMF_LOGCOOKIES,
266 RCL_CLIENT, llcd->llcd_cookiebytes);
268 rc = ptlrpc_request_bufs_pack(req, LUSTRE_LOG_VERSION,
269 OBD_LOG_CANCEL, bufs, NULL);
271 ptlrpc_request_free(req);
275 ptlrpc_at_set_req_timeout(req);
276 ptlrpc_request_set_replen(req);
279 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
280 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
282 req->rq_interpret_reply = (ptlrpc_interpterer_t)llcd_interpret;
284 CLASSERT(sizeof(*la) <= sizeof(req->rq_async_args));
285 la = ptlrpc_req_async_args(req);
288 /* llog cancels will be replayed after reconnect so this will do twice
289 * first from replay llog, second for resended rpc */
290 req->rq_no_delay = req->rq_no_resend = 1;
292 ptlrpc_set_add_new_req(&lcm->lcm_pc, req);
295 CDEBUG(D_RPCTRACE, "Refused llcd %p\n", llcd);
301 * Attach @llcd to @ctxt. Establish llcd vs. ctxt reserve connection
302 * so hat they can refer each other.
305 llcd_attach(struct llog_ctxt *ctxt, struct llog_canceld_ctxt *llcd)
307 LASSERT(ctxt != NULL && llcd != NULL);
308 LASSERT_SEM_LOCKED(&ctxt->loc_sem);
309 LASSERT(ctxt->loc_llcd == NULL);
310 llcd->llcd_ctxt = llog_ctxt_get(ctxt);
311 ctxt->loc_llcd = llcd;
313 CDEBUG(D_RPCTRACE, "Attach llcd %p to ctxt %p\n",
320 * Opposite to llcd_attach(). Detaches llcd from its @ctxt. This makes
321 * sure that this llcd will not be found another time we try to cancel.
323 static struct llog_canceld_ctxt *llcd_detach(struct llog_ctxt *ctxt)
325 struct llog_canceld_ctxt *llcd;
327 LASSERT(ctxt != NULL);
328 LASSERT_SEM_LOCKED(&ctxt->loc_sem);
330 llcd = ctxt->loc_llcd;
334 CDEBUG(D_RPCTRACE, "Detach llcd %p from ctxt %p\n",
337 ctxt->loc_llcd = NULL;
343 * Return @llcd cached in @ctxt. Allocate new one if required. Attach it
344 * to ctxt so that it may be used for gathering cookies and sending.
346 static struct llog_canceld_ctxt *llcd_get(struct llog_ctxt *ctxt)
348 struct llog_canceld_ctxt *llcd;
350 llcd = llcd_alloc(ctxt->loc_lcm);
352 CERROR("Can't alloc an llcd for ctxt %p\n", ctxt);
355 llcd_attach(ctxt, llcd);
360 * Deatch llcd from its @ctxt. Free llcd.
362 static void llcd_put(struct llog_ctxt *ctxt)
364 struct llog_canceld_ctxt *llcd;
366 llcd = llcd_detach(ctxt);
372 * Detach llcd from its @ctxt so that nobody will find it with try to
373 * re-use. Send llcd to remote node.
375 static int llcd_push(struct llog_ctxt *ctxt)
377 struct llog_canceld_ctxt *llcd;
381 * Make sure that this llcd will not be sent again as we detach
384 llcd = llcd_detach(ctxt);
386 CERROR("Invalid detached llcd found %p\n", llcd);
387 llcd_print(llcd, __FUNCTION__, __LINE__);
391 rc = llcd_send(llcd);
393 CERROR("Couldn't send llcd %p (%d)\n", llcd, rc);
398 * Start recovery thread which actually deals llcd sending. This
399 * is all ptlrpc standard thread based so there is not much of work
402 int llog_recov_thread_start(struct llog_commit_master *lcm)
407 rc = ptlrpcd_start(-1, 1, lcm->lcm_name, &lcm->lcm_pc);
409 CERROR("Error %d while starting recovery thread %s\n",
415 EXPORT_SYMBOL(llog_recov_thread_start);
418 * Stop recovery thread. Complement to llog_recov_thread_start().
420 void llog_recov_thread_stop(struct llog_commit_master *lcm, int force)
425 * Let all know that we're stopping. This will also make
426 * llcd_send() refuse any new llcds.
428 cfs_set_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags);
431 * Stop processing thread. No new rpcs will be accepted for
432 * for processing now.
434 ptlrpcd_stop(&lcm->lcm_pc, force);
437 * By this point no alive inflight llcds should be left. Only
438 * those forgotten in sync may still be attached to ctxt. Let's
441 if (cfs_atomic_read(&lcm->lcm_count) != 0) {
442 struct llog_canceld_ctxt *llcd;
445 CERROR("Busy llcds found (%d) on lcm %p\n",
446 cfs_atomic_read(&lcm->lcm_count), lcm);
448 cfs_spin_lock(&lcm->lcm_lock);
449 cfs_list_for_each(tmp, &lcm->lcm_llcds) {
450 llcd = cfs_list_entry(tmp, struct llog_canceld_ctxt,
452 llcd_print(llcd, __FUNCTION__, __LINE__);
454 cfs_spin_unlock(&lcm->lcm_lock);
457 * No point to go further with busy llcds at this point
458 * as this is clear bug. It might mean we got hanging
459 * rpc which holds import ref and this means we will not
460 * be able to cleanup anyways.
462 * Or we just missed to kill them when they were not
463 * attached to ctxt. In this case our slab will remind
464 * us about this a bit later.
470 EXPORT_SYMBOL(llog_recov_thread_stop);
473 * Initialize commit master structure and start recovery thread on it.
475 struct llog_commit_master *llog_recov_thread_init(char *name)
477 struct llog_commit_master *lcm;
486 * Try to create threads with unique names.
488 snprintf(lcm->lcm_name, sizeof(lcm->lcm_name),
491 cfs_atomic_set(&lcm->lcm_count, 0);
492 cfs_atomic_set(&lcm->lcm_refcount, 1);
493 cfs_spin_lock_init(&lcm->lcm_lock);
494 CFS_INIT_LIST_HEAD(&lcm->lcm_llcds);
495 rc = llog_recov_thread_start(lcm);
497 CERROR("Can't start commit thread, rc %d\n", rc);
505 EXPORT_SYMBOL(llog_recov_thread_init);
508 * Finalize commit master and its recovery thread.
510 void llog_recov_thread_fini(struct llog_commit_master *lcm, int force)
513 llog_recov_thread_stop(lcm, force);
517 EXPORT_SYMBOL(llog_recov_thread_fini);
519 static int llog_recov_thread_replay(struct llog_ctxt *ctxt,
522 struct obd_device *obd = ctxt->loc_obd;
523 struct llog_process_cat_args *lpca;
527 if (obd->obd_stopping)
531 * This will be balanced in llog_cat_process_thread()
538 lpca->lpca_arg = arg;
541 * This will be balanced in llog_cat_process_thread()
543 lpca->lpca_ctxt = llog_ctxt_get(ctxt);
544 if (!lpca->lpca_ctxt) {
548 rc = cfs_create_thread(llog_cat_process_thread, lpca, CFS_DAEMON_FLAGS);
550 CERROR("Error starting llog_cat_process_thread(): %d\n", rc);
554 CDEBUG(D_HA, "Started llog_cat_process_thread(): %d\n", rc);
561 int llog_obd_repl_connect(struct llog_ctxt *ctxt,
562 struct llog_logid *logid, struct llog_gen *gen,
563 struct obd_uuid *uuid)
569 * Send back cached llcd from llog before recovery if we have any.
570 * This is void is nothing cached is found there.
572 llog_sync(ctxt, NULL);
575 * Start recovery in separate thread.
577 cfs_mutex_down(&ctxt->loc_sem);
578 ctxt->loc_gen = *gen;
579 rc = llog_recov_thread_replay(ctxt, ctxt->llog_proc_cb, logid);
580 cfs_mutex_up(&ctxt->loc_sem);
584 EXPORT_SYMBOL(llog_obd_repl_connect);
587 * Deleted objects have a commit callback that cancels the MDS
588 * log record for the deletion. The commit callback calls this
591 int llog_obd_repl_cancel(struct llog_ctxt *ctxt,
592 struct lov_stripe_md *lsm, int count,
593 struct llog_cookie *cookies, int flags)
595 struct llog_commit_master *lcm;
596 struct llog_canceld_ctxt *llcd;
600 LASSERT(ctxt != NULL);
602 cfs_mutex_down(&ctxt->loc_sem);
603 if (!ctxt->loc_lcm) {
604 CDEBUG(D_RPCTRACE, "No lcm for ctxt %p\n", ctxt);
605 GOTO(out, rc = -ENODEV);
608 CDEBUG(D_INFO, "cancel on lsm %p\n", lcm);
611 * Let's check if we have all structures alive. We also check for
612 * possible shutdown. Do nothing if we're stopping.
614 if (ctxt->loc_imp == NULL) {
615 CDEBUG(D_RPCTRACE, "No import for ctxt %p\n", ctxt);
616 GOTO(out, rc = -ENODEV);
619 if (cfs_test_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags)) {
620 CDEBUG(D_RPCTRACE, "Commit thread is stopping for ctxt %p\n",
622 GOTO(out, rc = -ENODEV);
625 llcd = ctxt->loc_llcd;
627 if (count > 0 && cookies != NULL) {
629 * Get new llcd from ctxt if required.
632 llcd = llcd_get(ctxt);
634 GOTO(out, rc = -ENOMEM);
636 * Allocation is successful, let's check for stop
637 * flag again to fall back as soon as possible.
639 if (cfs_test_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags))
640 GOTO(out, rc = -ENODEV);
644 * Llcd does not have enough room for @cookies. Let's push
645 * it out and allocate new one.
647 if (!llcd_fit(llcd, cookies)) {
648 rc = llcd_push(ctxt);
651 llcd = llcd_get(ctxt);
653 GOTO(out, rc = -ENOMEM);
655 * Allocation is successful, let's check for stop
656 * flag again to fall back as soon as possible.
658 if (cfs_test_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags))
659 GOTO(out, rc = -ENODEV);
663 * Copy cookies to @llcd, no matter old or new allocated
666 llcd_copy(llcd, cookies);
670 * Let's check if we need to send copied @cookies asap. If yes
673 if (llcd && (flags & OBD_LLOG_FL_SENDNOW)) {
674 CDEBUG(D_RPCTRACE, "Sync llcd %p\n", llcd);
675 rc = llcd_push(ctxt);
683 cfs_mutex_up(&ctxt->loc_sem);
686 EXPORT_SYMBOL(llog_obd_repl_cancel);
688 int llog_obd_repl_sync(struct llog_ctxt *ctxt, struct obd_export *exp)
694 * Flush any remaining llcd.
696 cfs_mutex_down(&ctxt->loc_sem);
697 if (exp && (ctxt->loc_imp == exp->exp_imp_reverse)) {
699 * This is ost->mds connection, we can't be sure that mds
700 * can still receive cookies, let's killed the cached llcd.
702 CDEBUG(D_RPCTRACE, "Kill cached llcd\n");
704 cfs_mutex_up(&ctxt->loc_sem);
707 * This is either llog_sync() from generic llog code or sync
708 * on client disconnect. In either way let's do it and send
709 * llcds to the target with waiting for completion.
711 CDEBUG(D_RPCTRACE, "Sync cached llcd\n");
712 cfs_mutex_up(&ctxt->loc_sem);
713 rc = llog_cancel(ctxt, NULL, 0, NULL, OBD_LLOG_FL_SENDNOW);
717 EXPORT_SYMBOL(llog_obd_repl_sync);
719 #else /* !__KERNEL__ */
721 int llog_obd_repl_cancel(struct llog_ctxt *ctxt,
722 struct lov_stripe_md *lsm, int count,
723 struct llog_cookie *cookies, int flags)
730 * Module init time fucntion. Initializes slab for llcd objects.
732 int llog_recov_init(void)
736 llcd_size = CFS_PAGE_SIZE -
737 lustre_msg_size(LUSTRE_MSG_MAGIC_V2, 1, NULL);
738 llcd_size += offsetof(struct llog_canceld_ctxt, llcd_cookies);
739 llcd_cache = cfs_mem_cache_create("llcd_cache", llcd_size, 0, 0);
741 CERROR("Error allocating llcd cache\n");
748 * Module fini time fucntion. Releases slab for llcd objects.
750 void llog_recov_fini(void)
753 * Kill llcd cache when thread is stopped and we're sure no
758 * In 2.6.22 cfs_mem_cache_destroy() will not return error
759 * for busy resources. Let's check it another way.
761 LASSERTF(cfs_atomic_read(&llcd_count) == 0,
762 "Can't destroy llcd cache! Number of "
763 "busy llcds: %d\n", cfs_atomic_read(&llcd_count));
764 cfs_mem_cache_destroy(llcd_cache);