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 * 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_SEM_LOCKED(&ctxt->loc_sem);
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 rc = ptlrpc_set_add_new_req(&lcm->lcm_pc, req);
291 ptlrpc_request_free(req);
296 CDEBUG(D_RPCTRACE, "Refused llcd %p\n", llcd);
302 * Attach @llcd to @ctxt. Establish llcd vs. ctxt reserve connection
303 * so hat they can refer each other.
306 llcd_attach(struct llog_ctxt *ctxt, struct llog_canceld_ctxt *llcd)
308 LASSERT(ctxt != NULL && llcd != NULL);
309 LASSERT_SEM_LOCKED(&ctxt->loc_sem);
310 LASSERT(ctxt->loc_llcd == NULL);
311 llcd->llcd_ctxt = llog_ctxt_get(ctxt);
312 ctxt->loc_llcd = llcd;
314 CDEBUG(D_RPCTRACE, "Attach llcd %p to ctxt %p\n",
321 * Opposite to llcd_attach(). Detaches llcd from its @ctxt. This makes
322 * sure that this llcd will not be found another time we try to cancel.
324 static struct llog_canceld_ctxt *llcd_detach(struct llog_ctxt *ctxt)
326 struct llog_canceld_ctxt *llcd;
328 LASSERT(ctxt != NULL);
329 LASSERT_SEM_LOCKED(&ctxt->loc_sem);
331 llcd = ctxt->loc_llcd;
335 CDEBUG(D_RPCTRACE, "Detach llcd %p from ctxt %p\n",
338 ctxt->loc_llcd = NULL;
344 * Return @llcd cached in @ctxt. Allocate new one if required. Attach it
345 * to ctxt so that it may be used for gathering cookies and sending.
347 static struct llog_canceld_ctxt *llcd_get(struct llog_ctxt *ctxt)
349 struct llog_canceld_ctxt *llcd;
351 llcd = llcd_alloc(ctxt->loc_lcm);
353 CERROR("Can't alloc an llcd for ctxt %p\n", ctxt);
356 llcd_attach(ctxt, llcd);
361 * Deatch llcd from its @ctxt. Free llcd.
363 static void llcd_put(struct llog_ctxt *ctxt)
365 struct llog_canceld_ctxt *llcd;
367 llcd = llcd_detach(ctxt);
373 * Detach llcd from its @ctxt so that nobody will find it with try to
374 * re-use. Send llcd to remote node.
376 static int llcd_push(struct llog_ctxt *ctxt)
378 struct llog_canceld_ctxt *llcd;
382 * Make sure that this llcd will not be sent again as we detach
385 llcd = llcd_detach(ctxt);
387 CERROR("Invalid detached llcd found %p\n", llcd);
388 llcd_print(llcd, __FUNCTION__, __LINE__);
392 rc = llcd_send(llcd);
394 CERROR("Couldn't send llcd %p (%d)\n", llcd, rc);
399 * Start recovery thread which actually deals llcd sending. This
400 * is all ptlrpc standard thread based so there is not much of work
403 int llog_recov_thread_start(struct llog_commit_master *lcm)
408 rc = ptlrpcd_start(lcm->lcm_name, &lcm->lcm_pc);
410 CERROR("Error %d while starting recovery thread %s\n",
416 EXPORT_SYMBOL(llog_recov_thread_start);
419 * Stop recovery thread. Complement to llog_recov_thread_start().
421 void llog_recov_thread_stop(struct llog_commit_master *lcm, int force)
426 * Let all know that we're stopping. This will also make
427 * llcd_send() refuse any new llcds.
429 cfs_set_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags);
432 * Stop processing thread. No new rpcs will be accepted for
433 * for processing now.
435 ptlrpcd_stop(&lcm->lcm_pc, force);
438 * By this point no alive inflight llcds should be left. Only
439 * those forgotten in sync may still be attached to ctxt. Let's
442 if (cfs_atomic_read(&lcm->lcm_count) != 0) {
443 struct llog_canceld_ctxt *llcd;
446 CERROR("Busy llcds found (%d) on lcm %p\n",
447 cfs_atomic_read(&lcm->lcm_count), lcm);
449 cfs_spin_lock(&lcm->lcm_lock);
450 cfs_list_for_each(tmp, &lcm->lcm_llcds) {
451 llcd = cfs_list_entry(tmp, struct llog_canceld_ctxt,
453 llcd_print(llcd, __FUNCTION__, __LINE__);
455 cfs_spin_unlock(&lcm->lcm_lock);
458 * No point to go further with busy llcds at this point
459 * as this is clear bug. It might mean we got hanging
460 * rpc which holds import ref and this means we will not
461 * be able to cleanup anyways.
463 * Or we just missed to kill them when they were not
464 * attached to ctxt. In this case our slab will remind
465 * us about this a bit later.
471 EXPORT_SYMBOL(llog_recov_thread_stop);
474 * Initialize commit master structure and start recovery thread on it.
476 struct llog_commit_master *llog_recov_thread_init(char *name)
478 struct llog_commit_master *lcm;
487 * Try to create threads with unique names.
489 snprintf(lcm->lcm_name, sizeof(lcm->lcm_name),
492 cfs_atomic_set(&lcm->lcm_count, 0);
493 cfs_atomic_set(&lcm->lcm_refcount, 1);
494 cfs_spin_lock_init(&lcm->lcm_lock);
495 CFS_INIT_LIST_HEAD(&lcm->lcm_llcds);
496 rc = llog_recov_thread_start(lcm);
498 CERROR("Can't start commit thread, rc %d\n", rc);
506 EXPORT_SYMBOL(llog_recov_thread_init);
509 * Finalize commit master and its recovery thread.
511 void llog_recov_thread_fini(struct llog_commit_master *lcm, int force)
514 llog_recov_thread_stop(lcm, force);
518 EXPORT_SYMBOL(llog_recov_thread_fini);
520 static int llog_recov_thread_replay(struct llog_ctxt *ctxt,
523 struct obd_device *obd = ctxt->loc_obd;
524 struct llog_process_cat_args *lpca;
528 if (obd->obd_stopping)
532 * This will be balanced in llog_cat_process_thread()
539 lpca->lpca_arg = arg;
542 * This will be balanced in llog_cat_process_thread()
544 lpca->lpca_ctxt = llog_ctxt_get(ctxt);
545 if (!lpca->lpca_ctxt) {
549 rc = cfs_kernel_thread(llog_cat_process_thread, lpca,
550 CLONE_VM | CLONE_FILES);
552 CERROR("Error starting llog_cat_process_thread(): %d\n", rc);
556 CDEBUG(D_HA, "Started llog_cat_process_thread(): %d\n", rc);
563 int llog_obd_repl_connect(struct llog_ctxt *ctxt,
564 struct llog_logid *logid, struct llog_gen *gen,
565 struct obd_uuid *uuid)
571 * Send back cached llcd from llog before recovery if we have any.
572 * This is void is nothing cached is found there.
574 llog_sync(ctxt, NULL);
577 * Start recovery in separate thread.
579 cfs_mutex_down(&ctxt->loc_sem);
580 ctxt->loc_gen = *gen;
581 rc = llog_recov_thread_replay(ctxt, ctxt->llog_proc_cb, logid);
582 cfs_mutex_up(&ctxt->loc_sem);
586 EXPORT_SYMBOL(llog_obd_repl_connect);
589 * Deleted objects have a commit callback that cancels the MDS
590 * log record for the deletion. The commit callback calls this
593 int llog_obd_repl_cancel(struct llog_ctxt *ctxt,
594 struct lov_stripe_md *lsm, int count,
595 struct llog_cookie *cookies, int flags)
597 struct llog_commit_master *lcm;
598 struct llog_canceld_ctxt *llcd;
602 LASSERT(ctxt != NULL);
604 cfs_mutex_down(&ctxt->loc_sem);
605 if (!ctxt->loc_lcm) {
606 CDEBUG(D_RPCTRACE, "No lcm for ctxt %p\n", ctxt);
607 GOTO(out, rc = -ENODEV);
610 CDEBUG(D_INFO, "cancel on lsm %p\n", lcm);
613 * Let's check if we have all structures alive. We also check for
614 * possible shutdown. Do nothing if we're stopping.
616 if (ctxt->loc_imp == NULL) {
617 CDEBUG(D_RPCTRACE, "No import for ctxt %p\n", ctxt);
618 GOTO(out, rc = -ENODEV);
621 if (cfs_test_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags)) {
622 CDEBUG(D_RPCTRACE, "Commit thread is stopping for ctxt %p\n",
624 GOTO(out, rc = -ENODEV);
627 llcd = ctxt->loc_llcd;
629 if (count > 0 && cookies != NULL) {
631 * Get new llcd from ctxt if required.
634 llcd = llcd_get(ctxt);
636 GOTO(out, rc = -ENOMEM);
638 * Allocation is successful, let's check for stop
639 * flag again to fall back as soon as possible.
641 if (cfs_test_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags))
642 GOTO(out, rc = -ENODEV);
646 * Llcd does not have enough room for @cookies. Let's push
647 * it out and allocate new one.
649 if (!llcd_fit(llcd, cookies)) {
650 rc = llcd_push(ctxt);
653 llcd = llcd_get(ctxt);
655 GOTO(out, rc = -ENOMEM);
657 * Allocation is successful, let's check for stop
658 * flag again to fall back as soon as possible.
660 if (cfs_test_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags))
661 GOTO(out, rc = -ENODEV);
665 * Copy cookies to @llcd, no matter old or new allocated
668 llcd_copy(llcd, cookies);
672 * Let's check if we need to send copied @cookies asap. If yes
675 if (llcd && (flags & OBD_LLOG_FL_SENDNOW)) {
676 CDEBUG(D_RPCTRACE, "Sync llcd %p\n", llcd);
677 rc = llcd_push(ctxt);
685 cfs_mutex_up(&ctxt->loc_sem);
688 EXPORT_SYMBOL(llog_obd_repl_cancel);
690 int llog_obd_repl_sync(struct llog_ctxt *ctxt, struct obd_export *exp)
696 * Flush any remaining llcd.
698 cfs_mutex_down(&ctxt->loc_sem);
699 if (exp && (ctxt->loc_imp == exp->exp_imp_reverse)) {
701 * This is ost->mds connection, we can't be sure that mds
702 * can still receive cookies, let's killed the cached llcd.
704 CDEBUG(D_RPCTRACE, "Kill cached llcd\n");
706 cfs_mutex_up(&ctxt->loc_sem);
709 * This is either llog_sync() from generic llog code or sync
710 * on client disconnect. In either way let's do it and send
711 * llcds to the target with waiting for completion.
713 CDEBUG(D_RPCTRACE, "Sync cached llcd\n");
714 cfs_mutex_up(&ctxt->loc_sem);
715 rc = llog_cancel(ctxt, NULL, 0, NULL, OBD_LLOG_FL_SENDNOW);
719 EXPORT_SYMBOL(llog_obd_repl_sync);
721 #else /* !__KERNEL__ */
723 int llog_obd_repl_cancel(struct llog_ctxt *ctxt,
724 struct lov_stripe_md *lsm, int count,
725 struct llog_cookie *cookies, int flags)
732 * Module init time fucntion. Initializes slab for llcd objects.
734 int llog_recov_init(void)
738 llcd_size = CFS_PAGE_SIZE -
739 lustre_msg_size(LUSTRE_MSG_MAGIC_V2, 1, NULL);
740 llcd_size += offsetof(struct llog_canceld_ctxt, llcd_cookies);
741 llcd_cache = cfs_mem_cache_create("llcd_cache", llcd_size, 0, 0);
743 CERROR("Error allocating llcd cache\n");
750 * Module fini time fucntion. Releases slab for llcd objects.
752 void llog_recov_fini(void)
755 * Kill llcd cache when thread is stopped and we're sure no
760 * In 2.6.22 cfs_mem_cache_destroy() will not return error
761 * for busy resources. Let's check it another way.
763 LASSERTF(cfs_atomic_read(&llcd_count) == 0,
764 "Can't destroy llcd cache! Number of "
765 "busy llcds: %d\n", cfs_atomic_read(&llcd_count));
766 cfs_mem_cache_destroy(llcd_cache);