4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, Whamcloud, Inc.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/ptlrpc/recov_thread.c
38 * OST<->MDS recovery logging thread.
39 * Invariants in implementation:
40 * - we do not share logs among different OST<->MDS connections, so that
41 * if an OST or MDS fails it need only look at log(s) relevant to itself
43 * Author: Andreas Dilger <adilger@clusterfs.com>
44 * Yury Umanets <yury.umanets@sun.com>
45 * Alexey Lyashkov <alexey.lyashkov@sun.com>
48 #define DEBUG_SUBSYSTEM S_LOG
51 # define EXPORT_SYMTAB
55 # include <libcfs/libcfs.h>
57 # include <libcfs/list.h>
58 # include <liblustre.h>
61 #include <obd_class.h>
62 #include <obd_support.h>
63 #include <obd_class.h>
64 #include <lustre_net.h>
65 #include <lnet/types.h>
66 #include <libcfs/list.h>
67 #include <lustre_log.h>
68 #include "ptlrpc_internal.h"
70 static cfs_atomic_t llcd_count = CFS_ATOMIC_INIT(0);
71 static cfs_mem_cache_t *llcd_cache = NULL;
75 LLOG_LCM_FL_START = 1 << 0,
76 LLOG_LCM_FL_EXIT = 1 << 1
79 struct llcd_async_args {
80 struct llog_canceld_ctxt *la_ctxt;
83 static void llcd_print(struct llog_canceld_ctxt *llcd,
84 const char *func, int line)
86 CDEBUG(D_RPCTRACE, "Llcd (%p) at %s:%d:\n", llcd, func, line);
87 CDEBUG(D_RPCTRACE, " size: %d\n", llcd->llcd_size);
88 CDEBUG(D_RPCTRACE, " ctxt: %p\n", llcd->llcd_ctxt);
89 CDEBUG(D_RPCTRACE, " lcm : %p\n", llcd->llcd_lcm);
90 CDEBUG(D_RPCTRACE, " cookiebytes : %d\n", llcd->llcd_cookiebytes);
94 * Allocate new llcd from cache, init it and return to caller.
95 * Bumps number of objects allocated.
97 static struct llog_canceld_ctxt *llcd_alloc(struct llog_commit_master *lcm)
99 struct llog_canceld_ctxt *llcd;
102 LASSERT(lcm != NULL);
105 * We want to send one page of cookies with rpc header. This buffer
106 * will be assigned later to the rpc, this is why we preserve the
107 * space for rpc header.
109 size = CFS_PAGE_SIZE - lustre_msg_size(LUSTRE_MSG_MAGIC_V2, 1, NULL);
110 overhead = offsetof(struct llog_canceld_ctxt, llcd_cookies);
111 OBD_SLAB_ALLOC(llcd, llcd_cache, CFS_ALLOC_STD, size + overhead);
115 CFS_INIT_LIST_HEAD(&llcd->llcd_list);
116 llcd->llcd_cookiebytes = 0;
117 llcd->llcd_size = size;
119 cfs_spin_lock(&lcm->lcm_lock);
120 llcd->llcd_lcm = lcm;
121 cfs_atomic_inc(&lcm->lcm_count);
122 cfs_list_add_tail(&llcd->llcd_list, &lcm->lcm_llcds);
123 cfs_spin_unlock(&lcm->lcm_lock);
124 cfs_atomic_inc(&llcd_count);
126 CDEBUG(D_RPCTRACE, "Alloc llcd %p on lcm %p (%d)\n",
127 llcd, lcm, cfs_atomic_read(&lcm->lcm_count));
133 * Returns passed llcd to cache.
135 static void llcd_free(struct llog_canceld_ctxt *llcd)
137 struct llog_commit_master *lcm = llcd->llcd_lcm;
141 if (cfs_atomic_read(&lcm->lcm_count) == 0) {
142 CERROR("Invalid llcd free %p\n", llcd);
143 llcd_print(llcd, __FUNCTION__, __LINE__);
146 cfs_spin_lock(&lcm->lcm_lock);
147 LASSERT(!cfs_list_empty(&llcd->llcd_list));
148 cfs_list_del_init(&llcd->llcd_list);
149 cfs_atomic_dec(&lcm->lcm_count);
150 cfs_spin_unlock(&lcm->lcm_lock);
152 CDEBUG(D_RPCTRACE, "Free llcd %p on lcm %p (%d)\n",
153 llcd, lcm, cfs_atomic_read(&lcm->lcm_count));
156 LASSERT(cfs_atomic_read(&llcd_count) > 0);
157 cfs_atomic_dec(&llcd_count);
159 size = offsetof(struct llog_canceld_ctxt, llcd_cookies) +
161 OBD_SLAB_FREE(llcd, llcd_cache, size);
165 * Checks if passed cookie fits into llcd free space buffer. Returns
166 * 1 if yes and 0 otherwise.
169 llcd_fit(struct llog_canceld_ctxt *llcd, struct llog_cookie *cookies)
171 return (llcd->llcd_size - llcd->llcd_cookiebytes >= sizeof(*cookies));
175 * Copy passed @cookies to @llcd.
178 llcd_copy(struct llog_canceld_ctxt *llcd, struct llog_cookie *cookies)
180 LASSERT(llcd_fit(llcd, cookies));
181 memcpy((char *)llcd->llcd_cookies + llcd->llcd_cookiebytes,
182 cookies, sizeof(*cookies));
183 llcd->llcd_cookiebytes += sizeof(*cookies);
187 * Llcd completion function. Called uppon llcd send finish regardless
188 * sending result. Error is passed in @rc. Note, that this will be called
189 * in cleanup time when all inflight rpcs aborted.
192 llcd_interpret(const struct lu_env *env,
193 struct ptlrpc_request *req, void *args, int rc)
195 struct llcd_async_args *la = args;
196 struct llog_canceld_ctxt *llcd = la->la_ctxt;
198 CDEBUG(D_RPCTRACE, "Sent llcd %p (%d) - killing it\n", llcd, rc);
204 * Send @llcd to remote node. Free llcd uppon completion or error. Sending
205 * is performed in async style so this function will return asap without
208 static int llcd_send(struct llog_canceld_ctxt *llcd)
210 char *bufs[2] = { NULL, (char *)llcd->llcd_cookies };
211 struct obd_import *import = NULL;
212 struct llog_commit_master *lcm;
213 struct llcd_async_args *la;
214 struct ptlrpc_request *req;
215 struct llog_ctxt *ctxt;
219 ctxt = llcd->llcd_ctxt;
221 CERROR("Invalid llcd with NULL ctxt found (%p)\n",
223 llcd_print(llcd, __FUNCTION__, __LINE__);
226 LASSERT_MUTEX_LOCKED(&ctxt->loc_mutex);
228 if (llcd->llcd_cookiebytes == 0)
231 lcm = llcd->llcd_lcm;
234 * Check if we're in exit stage. Do not send llcd in
237 if (cfs_test_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags))
238 GOTO(exit, rc = -ENODEV);
240 CDEBUG(D_RPCTRACE, "Sending llcd %p\n", llcd);
242 import = llcd->llcd_ctxt->loc_imp;
243 if (!import || (import == LP_POISON) ||
244 (import->imp_client == LP_POISON)) {
245 CERROR("Invalid import %p for llcd %p\n",
247 GOTO(exit, rc = -ENODEV);
250 OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_RECOV, 10);
253 * No need to get import here as it is already done in
254 * llog_receptor_accept().
256 req = ptlrpc_request_alloc(import, &RQF_LOG_CANCEL);
258 CERROR("Can't allocate request for sending llcd %p\n",
260 GOTO(exit, rc = -ENOMEM);
262 req_capsule_set_size(&req->rq_pill, &RMF_LOGCOOKIES,
263 RCL_CLIENT, llcd->llcd_cookiebytes);
265 rc = ptlrpc_request_bufs_pack(req, LUSTRE_LOG_VERSION,
266 OBD_LOG_CANCEL, bufs, NULL);
268 ptlrpc_request_free(req);
272 ptlrpc_at_set_req_timeout(req);
273 ptlrpc_request_set_replen(req);
276 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
277 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
279 req->rq_interpret_reply = (ptlrpc_interpterer_t)llcd_interpret;
281 CLASSERT(sizeof(*la) <= sizeof(req->rq_async_args));
282 la = ptlrpc_req_async_args(req);
285 /* llog cancels will be replayed after reconnect so this will do twice
286 * first from replay llog, second for resended rpc */
287 req->rq_no_delay = req->rq_no_resend = 1;
289 ptlrpc_set_add_new_req(&lcm->lcm_pc, req);
292 CDEBUG(D_RPCTRACE, "Refused llcd %p\n", llcd);
298 * Attach @llcd to @ctxt. Establish llcd vs. ctxt reserve connection
299 * so hat they can refer each other.
302 llcd_attach(struct llog_ctxt *ctxt, struct llog_canceld_ctxt *llcd)
304 LASSERT(ctxt != NULL && llcd != NULL);
305 LASSERT_MUTEX_LOCKED(&ctxt->loc_mutex);
306 LASSERT(ctxt->loc_llcd == NULL);
307 llcd->llcd_ctxt = llog_ctxt_get(ctxt);
308 ctxt->loc_llcd = llcd;
310 CDEBUG(D_RPCTRACE, "Attach llcd %p to ctxt %p\n",
317 * Opposite to llcd_attach(). Detaches llcd from its @ctxt. This makes
318 * sure that this llcd will not be found another time we try to cancel.
320 static struct llog_canceld_ctxt *llcd_detach(struct llog_ctxt *ctxt)
322 struct llog_canceld_ctxt *llcd;
324 LASSERT(ctxt != NULL);
325 LASSERT_MUTEX_LOCKED(&ctxt->loc_mutex);
327 llcd = ctxt->loc_llcd;
331 CDEBUG(D_RPCTRACE, "Detach llcd %p from ctxt %p\n",
334 ctxt->loc_llcd = NULL;
340 * Return @llcd cached in @ctxt. Allocate new one if required. Attach it
341 * to ctxt so that it may be used for gathering cookies and sending.
343 static struct llog_canceld_ctxt *llcd_get(struct llog_ctxt *ctxt)
345 struct llog_canceld_ctxt *llcd;
347 llcd = llcd_alloc(ctxt->loc_lcm);
349 CERROR("Can't alloc an llcd for ctxt %p\n", ctxt);
352 llcd_attach(ctxt, llcd);
357 * Deatch llcd from its @ctxt. Free llcd.
359 static void llcd_put(struct llog_ctxt *ctxt)
361 struct llog_canceld_ctxt *llcd;
363 llcd = llcd_detach(ctxt);
369 * Detach llcd from its @ctxt so that nobody will find it with try to
370 * re-use. Send llcd to remote node.
372 static int llcd_push(struct llog_ctxt *ctxt)
374 struct llog_canceld_ctxt *llcd;
378 * Make sure that this llcd will not be sent again as we detach
381 llcd = llcd_detach(ctxt);
383 CERROR("Invalid detached llcd found %p\n", llcd);
384 llcd_print(llcd, __FUNCTION__, __LINE__);
388 rc = llcd_send(llcd);
390 CERROR("Couldn't send llcd %p (%d)\n", llcd, rc);
395 * Start recovery thread which actually deals llcd sending. This
396 * is all ptlrpc standard thread based so there is not much of work
399 int llog_recov_thread_start(struct llog_commit_master *lcm)
404 rc = ptlrpcd_start(-1, 1, lcm->lcm_name, &lcm->lcm_pc);
406 CERROR("Error %d while starting recovery thread %s\n",
412 EXPORT_SYMBOL(llog_recov_thread_start);
415 * Stop recovery thread. Complement to llog_recov_thread_start().
417 void llog_recov_thread_stop(struct llog_commit_master *lcm, int force)
422 * Let all know that we're stopping. This will also make
423 * llcd_send() refuse any new llcds.
425 cfs_set_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags);
428 * Stop processing thread. No new rpcs will be accepted for
429 * for processing now.
431 ptlrpcd_stop(&lcm->lcm_pc, force);
434 * By this point no alive inflight llcds should be left. Only
435 * those forgotten in sync may still be attached to ctxt. Let's
438 if (cfs_atomic_read(&lcm->lcm_count) != 0) {
439 struct llog_canceld_ctxt *llcd;
442 CERROR("Busy llcds found (%d) on lcm %p\n",
443 cfs_atomic_read(&lcm->lcm_count), lcm);
445 cfs_spin_lock(&lcm->lcm_lock);
446 cfs_list_for_each(tmp, &lcm->lcm_llcds) {
447 llcd = cfs_list_entry(tmp, struct llog_canceld_ctxt,
449 llcd_print(llcd, __FUNCTION__, __LINE__);
451 cfs_spin_unlock(&lcm->lcm_lock);
454 * No point to go further with busy llcds at this point
455 * as this is clear bug. It might mean we got hanging
456 * rpc which holds import ref and this means we will not
457 * be able to cleanup anyways.
459 * Or we just missed to kill them when they were not
460 * attached to ctxt. In this case our slab will remind
461 * us about this a bit later.
467 EXPORT_SYMBOL(llog_recov_thread_stop);
470 * Initialize commit master structure and start recovery thread on it.
472 struct llog_commit_master *llog_recov_thread_init(char *name)
474 struct llog_commit_master *lcm;
483 * Try to create threads with unique names.
485 snprintf(lcm->lcm_name, sizeof(lcm->lcm_name),
488 cfs_atomic_set(&lcm->lcm_count, 0);
489 cfs_atomic_set(&lcm->lcm_refcount, 1);
490 cfs_spin_lock_init(&lcm->lcm_lock);
491 CFS_INIT_LIST_HEAD(&lcm->lcm_llcds);
492 rc = llog_recov_thread_start(lcm);
494 CERROR("Can't start commit thread, rc %d\n", rc);
502 EXPORT_SYMBOL(llog_recov_thread_init);
505 * Finalize commit master and its recovery thread.
507 void llog_recov_thread_fini(struct llog_commit_master *lcm, int force)
510 llog_recov_thread_stop(lcm, force);
514 EXPORT_SYMBOL(llog_recov_thread_fini);
516 static int llog_recov_thread_replay(struct llog_ctxt *ctxt,
519 struct obd_device *obd = ctxt->loc_obd;
520 struct llog_process_cat_args *lpca;
524 if (obd->obd_stopping)
528 * This will be balanced in llog_cat_process_thread()
535 lpca->lpca_arg = arg;
538 * This will be balanced in llog_cat_process_thread()
540 lpca->lpca_ctxt = llog_ctxt_get(ctxt);
541 if (!lpca->lpca_ctxt) {
545 rc = cfs_create_thread(llog_cat_process_thread, lpca, CFS_DAEMON_FLAGS);
547 CERROR("Error starting llog_cat_process_thread(): %d\n", rc);
551 CDEBUG(D_HA, "Started llog_cat_process_thread(): %d\n", rc);
558 int llog_obd_repl_connect(struct llog_ctxt *ctxt,
559 struct llog_logid *logid, struct llog_gen *gen,
560 struct obd_uuid *uuid)
566 * Send back cached llcd from llog before recovery if we have any.
567 * This is void is nothing cached is found there.
569 llog_sync(ctxt, NULL);
572 * Start recovery in separate thread.
574 cfs_mutex_lock(&ctxt->loc_mutex);
575 ctxt->loc_gen = *gen;
576 rc = llog_recov_thread_replay(ctxt, ctxt->llog_proc_cb, logid);
577 cfs_mutex_unlock(&ctxt->loc_mutex);
581 EXPORT_SYMBOL(llog_obd_repl_connect);
584 * Deleted objects have a commit callback that cancels the MDS
585 * log record for the deletion. The commit callback calls this
588 int llog_obd_repl_cancel(struct llog_ctxt *ctxt,
589 struct lov_stripe_md *lsm, int count,
590 struct llog_cookie *cookies, int flags)
592 struct llog_commit_master *lcm;
593 struct llog_canceld_ctxt *llcd;
597 LASSERT(ctxt != NULL);
599 cfs_mutex_lock(&ctxt->loc_mutex);
600 if (!ctxt->loc_lcm) {
601 CDEBUG(D_RPCTRACE, "No lcm for ctxt %p\n", ctxt);
602 GOTO(out, rc = -ENODEV);
605 CDEBUG(D_INFO, "cancel on lsm %p\n", lcm);
608 * Let's check if we have all structures alive. We also check for
609 * possible shutdown. Do nothing if we're stopping.
611 if (ctxt->loc_imp == NULL) {
612 CDEBUG(D_RPCTRACE, "No import for ctxt %p\n", ctxt);
613 GOTO(out, rc = -ENODEV);
616 if (cfs_test_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags)) {
617 CDEBUG(D_RPCTRACE, "Commit thread is stopping for ctxt %p\n",
619 GOTO(out, rc = -ENODEV);
622 llcd = ctxt->loc_llcd;
624 if (count > 0 && cookies != NULL) {
626 * Get new llcd from ctxt if required.
629 llcd = llcd_get(ctxt);
631 GOTO(out, rc = -ENOMEM);
633 * Allocation is successful, let's check for stop
634 * flag again to fall back as soon as possible.
636 if (cfs_test_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags))
637 GOTO(out, rc = -ENODEV);
641 * Llcd does not have enough room for @cookies. Let's push
642 * it out and allocate new one.
644 if (!llcd_fit(llcd, cookies)) {
645 rc = llcd_push(ctxt);
648 llcd = llcd_get(ctxt);
650 GOTO(out, rc = -ENOMEM);
652 * Allocation is successful, let's check for stop
653 * flag again to fall back as soon as possible.
655 if (cfs_test_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags))
656 GOTO(out, rc = -ENODEV);
660 * Copy cookies to @llcd, no matter old or new allocated
663 llcd_copy(llcd, cookies);
667 * Let's check if we need to send copied @cookies asap. If yes
670 if (llcd && (flags & OBD_LLOG_FL_SENDNOW)) {
671 CDEBUG(D_RPCTRACE, "Sync llcd %p\n", llcd);
672 rc = llcd_push(ctxt);
680 cfs_mutex_unlock(&ctxt->loc_mutex);
683 EXPORT_SYMBOL(llog_obd_repl_cancel);
685 int llog_obd_repl_sync(struct llog_ctxt *ctxt, struct obd_export *exp)
691 * Flush any remaining llcd.
693 cfs_mutex_lock(&ctxt->loc_mutex);
694 if (exp && (ctxt->loc_imp == exp->exp_imp_reverse)) {
696 * This is ost->mds connection, we can't be sure that mds
697 * can still receive cookies, let's killed the cached llcd.
699 CDEBUG(D_RPCTRACE, "Kill cached llcd\n");
701 cfs_mutex_unlock(&ctxt->loc_mutex);
704 * This is either llog_sync() from generic llog code or sync
705 * on client disconnect. In either way let's do it and send
706 * llcds to the target with waiting for completion.
708 CDEBUG(D_RPCTRACE, "Sync cached llcd\n");
709 cfs_mutex_unlock(&ctxt->loc_mutex);
710 rc = llog_cancel(ctxt, NULL, 0, NULL, OBD_LLOG_FL_SENDNOW);
714 EXPORT_SYMBOL(llog_obd_repl_sync);
716 #else /* !__KERNEL__ */
718 int llog_obd_repl_cancel(struct llog_ctxt *ctxt,
719 struct lov_stripe_md *lsm, int count,
720 struct llog_cookie *cookies, int flags)
727 * Module init time fucntion. Initializes slab for llcd objects.
729 int llog_recov_init(void)
733 llcd_size = CFS_PAGE_SIZE -
734 lustre_msg_size(LUSTRE_MSG_MAGIC_V2, 1, NULL);
735 llcd_size += offsetof(struct llog_canceld_ctxt, llcd_cookies);
736 llcd_cache = cfs_mem_cache_create("llcd_cache", llcd_size, 0, 0);
738 CERROR("Error allocating llcd cache\n");
745 * Module fini time fucntion. Releases slab for llcd objects.
747 void llog_recov_fini(void)
750 * Kill llcd cache when thread is stopped and we're sure no
755 * In 2.6.22 cfs_mem_cache_destroy() will not return error
756 * for busy resources. Let's check it another way.
758 LASSERTF(cfs_atomic_read(&llcd_count) == 0,
759 "Can't destroy llcd cache! Number of "
760 "busy llcds: %d\n", cfs_atomic_read(&llcd_count));
761 cfs_mem_cache_destroy(llcd_cache);