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.
32 * Copyright (c) 2011, Whamcloud, Inc.
35 * This file is part of Lustre, http://www.lustre.org/
36 * Lustre is a trademark of Sun Microsystems, Inc.
38 * lustre/ptlrpc/recov_thread.c
40 * OST<->MDS recovery logging thread.
41 * Invariants in implementation:
42 * - we do not share logs among different OST<->MDS connections, so that
43 * if an OST or MDS fails it need only look at log(s) relevant to itself
45 * Author: Andreas Dilger <adilger@clusterfs.com>
46 * Yury Umanets <yury.umanets@sun.com>
47 * Alexey Lyashkov <alexey.lyashkov@sun.com>
50 #define DEBUG_SUBSYSTEM S_LOG
53 # define EXPORT_SYMTAB
57 # include <libcfs/libcfs.h>
59 # include <libcfs/list.h>
60 # include <liblustre.h>
63 #include <obd_class.h>
64 #include <obd_support.h>
65 #include <obd_class.h>
66 #include <lustre_net.h>
67 #include <lnet/types.h>
68 #include <libcfs/list.h>
69 #include <lustre_log.h>
70 #include "ptlrpc_internal.h"
72 static cfs_atomic_t llcd_count = CFS_ATOMIC_INIT(0);
73 static cfs_mem_cache_t *llcd_cache = NULL;
77 LLOG_LCM_FL_START = 1 << 0,
78 LLOG_LCM_FL_EXIT = 1 << 1
81 struct llcd_async_args {
82 struct llog_canceld_ctxt *la_ctxt;
85 static void llcd_print(struct llog_canceld_ctxt *llcd,
86 const char *func, int line)
88 CDEBUG(D_RPCTRACE, "Llcd (%p) at %s:%d:\n", llcd, func, line);
89 CDEBUG(D_RPCTRACE, " size: %d\n", llcd->llcd_size);
90 CDEBUG(D_RPCTRACE, " ctxt: %p\n", llcd->llcd_ctxt);
91 CDEBUG(D_RPCTRACE, " lcm : %p\n", llcd->llcd_lcm);
92 CDEBUG(D_RPCTRACE, " cookiebytes : %d\n", llcd->llcd_cookiebytes);
96 * Allocate new llcd from cache, init it and return to caller.
97 * Bumps number of objects allocated.
99 static struct llog_canceld_ctxt *llcd_alloc(struct llog_commit_master *lcm)
101 struct llog_canceld_ctxt *llcd;
104 LASSERT(lcm != NULL);
107 * We want to send one page of cookies with rpc header. This buffer
108 * will be assigned later to the rpc, this is why we preserve the
109 * space for rpc header.
111 size = CFS_PAGE_SIZE - lustre_msg_size(LUSTRE_MSG_MAGIC_V2, 1, NULL);
112 overhead = offsetof(struct llog_canceld_ctxt, llcd_cookies);
113 OBD_SLAB_ALLOC(llcd, llcd_cache, CFS_ALLOC_STD, size + overhead);
117 CFS_INIT_LIST_HEAD(&llcd->llcd_list);
118 llcd->llcd_cookiebytes = 0;
119 llcd->llcd_size = size;
121 cfs_spin_lock(&lcm->lcm_lock);
122 llcd->llcd_lcm = lcm;
123 cfs_atomic_inc(&lcm->lcm_count);
124 cfs_list_add_tail(&llcd->llcd_list, &lcm->lcm_llcds);
125 cfs_spin_unlock(&lcm->lcm_lock);
126 cfs_atomic_inc(&llcd_count);
128 CDEBUG(D_RPCTRACE, "Alloc llcd %p on lcm %p (%d)\n",
129 llcd, lcm, cfs_atomic_read(&lcm->lcm_count));
135 * Returns passed llcd to cache.
137 static void llcd_free(struct llog_canceld_ctxt *llcd)
139 struct llog_commit_master *lcm = llcd->llcd_lcm;
143 if (cfs_atomic_read(&lcm->lcm_count) == 0) {
144 CERROR("Invalid llcd free %p\n", llcd);
145 llcd_print(llcd, __FUNCTION__, __LINE__);
148 cfs_spin_lock(&lcm->lcm_lock);
149 LASSERT(!cfs_list_empty(&llcd->llcd_list));
150 cfs_list_del_init(&llcd->llcd_list);
151 cfs_atomic_dec(&lcm->lcm_count);
152 cfs_spin_unlock(&lcm->lcm_lock);
154 CDEBUG(D_RPCTRACE, "Free llcd %p on lcm %p (%d)\n",
155 llcd, lcm, cfs_atomic_read(&lcm->lcm_count));
158 LASSERT(cfs_atomic_read(&llcd_count) > 0);
159 cfs_atomic_dec(&llcd_count);
161 size = offsetof(struct llog_canceld_ctxt, llcd_cookies) +
163 OBD_SLAB_FREE(llcd, llcd_cache, size);
167 * Checks if passed cookie fits into llcd free space buffer. Returns
168 * 1 if yes and 0 otherwise.
171 llcd_fit(struct llog_canceld_ctxt *llcd, struct llog_cookie *cookies)
173 return (llcd->llcd_size - llcd->llcd_cookiebytes >= sizeof(*cookies));
177 * Copy passed @cookies to @llcd.
180 llcd_copy(struct llog_canceld_ctxt *llcd, struct llog_cookie *cookies)
182 LASSERT(llcd_fit(llcd, cookies));
183 memcpy((char *)llcd->llcd_cookies + llcd->llcd_cookiebytes,
184 cookies, sizeof(*cookies));
185 llcd->llcd_cookiebytes += sizeof(*cookies);
189 * Llcd completion function. Called uppon llcd send finish regardless
190 * sending result. Error is passed in @rc. Note, that this will be called
191 * in cleanup time when all inflight rpcs aborted.
194 llcd_interpret(const struct lu_env *env,
195 struct ptlrpc_request *req, void *args, int rc)
197 struct llcd_async_args *la = args;
198 struct llog_canceld_ctxt *llcd = la->la_ctxt;
200 CDEBUG(D_RPCTRACE, "Sent llcd %p (%d) - killing it\n", llcd, rc);
206 * Send @llcd to remote node. Free llcd uppon completion or error. Sending
207 * is performed in async style so this function will return asap without
210 static int llcd_send(struct llog_canceld_ctxt *llcd)
212 char *bufs[2] = { NULL, (char *)llcd->llcd_cookies };
213 struct obd_import *import = NULL;
214 struct llog_commit_master *lcm;
215 struct llcd_async_args *la;
216 struct ptlrpc_request *req;
217 struct llog_ctxt *ctxt;
221 ctxt = llcd->llcd_ctxt;
223 CERROR("Invalid llcd with NULL ctxt found (%p)\n",
225 llcd_print(llcd, __FUNCTION__, __LINE__);
228 LASSERT_SEM_LOCKED(&ctxt->loc_sem);
230 if (llcd->llcd_cookiebytes == 0)
233 lcm = llcd->llcd_lcm;
236 * Check if we're in exit stage. Do not send llcd in
239 if (cfs_test_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags))
240 GOTO(exit, rc = -ENODEV);
242 CDEBUG(D_RPCTRACE, "Sending llcd %p\n", llcd);
244 import = llcd->llcd_ctxt->loc_imp;
245 if (!import || (import == LP_POISON) ||
246 (import->imp_client == LP_POISON)) {
247 CERROR("Invalid import %p for llcd %p\n",
249 GOTO(exit, rc = -ENODEV);
252 OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_RECOV, 10);
255 * No need to get import here as it is already done in
256 * llog_receptor_accept().
258 req = ptlrpc_request_alloc(import, &RQF_LOG_CANCEL);
260 CERROR("Can't allocate request for sending llcd %p\n",
262 GOTO(exit, rc = -ENOMEM);
264 req_capsule_set_size(&req->rq_pill, &RMF_LOGCOOKIES,
265 RCL_CLIENT, llcd->llcd_cookiebytes);
267 rc = ptlrpc_request_bufs_pack(req, LUSTRE_LOG_VERSION,
268 OBD_LOG_CANCEL, bufs, NULL);
270 ptlrpc_request_free(req);
274 ptlrpc_at_set_req_timeout(req);
275 ptlrpc_request_set_replen(req);
278 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
279 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
281 req->rq_interpret_reply = (ptlrpc_interpterer_t)llcd_interpret;
283 CLASSERT(sizeof(*la) <= sizeof(req->rq_async_args));
284 la = ptlrpc_req_async_args(req);
287 /* llog cancels will be replayed after reconnect so this will do twice
288 * first from replay llog, second for resended rpc */
289 req->rq_no_delay = req->rq_no_resend = 1;
291 ptlrpc_set_add_new_req(&lcm->lcm_pc, req);
294 CDEBUG(D_RPCTRACE, "Refused llcd %p\n", llcd);
300 * Attach @llcd to @ctxt. Establish llcd vs. ctxt reserve connection
301 * so hat they can refer each other.
304 llcd_attach(struct llog_ctxt *ctxt, struct llog_canceld_ctxt *llcd)
306 LASSERT(ctxt != NULL && llcd != NULL);
307 LASSERT_SEM_LOCKED(&ctxt->loc_sem);
308 LASSERT(ctxt->loc_llcd == NULL);
309 llcd->llcd_ctxt = llog_ctxt_get(ctxt);
310 ctxt->loc_llcd = llcd;
312 CDEBUG(D_RPCTRACE, "Attach llcd %p to ctxt %p\n",
319 * Opposite to llcd_attach(). Detaches llcd from its @ctxt. This makes
320 * sure that this llcd will not be found another time we try to cancel.
322 static struct llog_canceld_ctxt *llcd_detach(struct llog_ctxt *ctxt)
324 struct llog_canceld_ctxt *llcd;
326 LASSERT(ctxt != NULL);
327 LASSERT_SEM_LOCKED(&ctxt->loc_sem);
329 llcd = ctxt->loc_llcd;
333 CDEBUG(D_RPCTRACE, "Detach llcd %p from ctxt %p\n",
336 ctxt->loc_llcd = NULL;
342 * Return @llcd cached in @ctxt. Allocate new one if required. Attach it
343 * to ctxt so that it may be used for gathering cookies and sending.
345 static struct llog_canceld_ctxt *llcd_get(struct llog_ctxt *ctxt)
347 struct llog_canceld_ctxt *llcd;
349 llcd = llcd_alloc(ctxt->loc_lcm);
351 CERROR("Can't alloc an llcd for ctxt %p\n", ctxt);
354 llcd_attach(ctxt, llcd);
359 * Deatch llcd from its @ctxt. Free llcd.
361 static void llcd_put(struct llog_ctxt *ctxt)
363 struct llog_canceld_ctxt *llcd;
365 llcd = llcd_detach(ctxt);
371 * Detach llcd from its @ctxt so that nobody will find it with try to
372 * re-use. Send llcd to remote node.
374 static int llcd_push(struct llog_ctxt *ctxt)
376 struct llog_canceld_ctxt *llcd;
380 * Make sure that this llcd will not be sent again as we detach
383 llcd = llcd_detach(ctxt);
385 CERROR("Invalid detached llcd found %p\n", llcd);
386 llcd_print(llcd, __FUNCTION__, __LINE__);
390 rc = llcd_send(llcd);
392 CERROR("Couldn't send llcd %p (%d)\n", llcd, rc);
397 * Start recovery thread which actually deals llcd sending. This
398 * is all ptlrpc standard thread based so there is not much of work
401 int llog_recov_thread_start(struct llog_commit_master *lcm)
406 rc = ptlrpcd_start(-1, 1, lcm->lcm_name, &lcm->lcm_pc);
408 CERROR("Error %d while starting recovery thread %s\n",
414 EXPORT_SYMBOL(llog_recov_thread_start);
417 * Stop recovery thread. Complement to llog_recov_thread_start().
419 void llog_recov_thread_stop(struct llog_commit_master *lcm, int force)
424 * Let all know that we're stopping. This will also make
425 * llcd_send() refuse any new llcds.
427 cfs_set_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags);
430 * Stop processing thread. No new rpcs will be accepted for
431 * for processing now.
433 ptlrpcd_stop(&lcm->lcm_pc, force);
436 * By this point no alive inflight llcds should be left. Only
437 * those forgotten in sync may still be attached to ctxt. Let's
440 if (cfs_atomic_read(&lcm->lcm_count) != 0) {
441 struct llog_canceld_ctxt *llcd;
444 CERROR("Busy llcds found (%d) on lcm %p\n",
445 cfs_atomic_read(&lcm->lcm_count), lcm);
447 cfs_spin_lock(&lcm->lcm_lock);
448 cfs_list_for_each(tmp, &lcm->lcm_llcds) {
449 llcd = cfs_list_entry(tmp, struct llog_canceld_ctxt,
451 llcd_print(llcd, __FUNCTION__, __LINE__);
453 cfs_spin_unlock(&lcm->lcm_lock);
456 * No point to go further with busy llcds at this point
457 * as this is clear bug. It might mean we got hanging
458 * rpc which holds import ref and this means we will not
459 * be able to cleanup anyways.
461 * Or we just missed to kill them when they were not
462 * attached to ctxt. In this case our slab will remind
463 * us about this a bit later.
469 EXPORT_SYMBOL(llog_recov_thread_stop);
472 * Initialize commit master structure and start recovery thread on it.
474 struct llog_commit_master *llog_recov_thread_init(char *name)
476 struct llog_commit_master *lcm;
485 * Try to create threads with unique names.
487 snprintf(lcm->lcm_name, sizeof(lcm->lcm_name),
490 cfs_atomic_set(&lcm->lcm_count, 0);
491 cfs_atomic_set(&lcm->lcm_refcount, 1);
492 cfs_spin_lock_init(&lcm->lcm_lock);
493 CFS_INIT_LIST_HEAD(&lcm->lcm_llcds);
494 rc = llog_recov_thread_start(lcm);
496 CERROR("Can't start commit thread, rc %d\n", rc);
504 EXPORT_SYMBOL(llog_recov_thread_init);
507 * Finalize commit master and its recovery thread.
509 void llog_recov_thread_fini(struct llog_commit_master *lcm, int force)
512 llog_recov_thread_stop(lcm, force);
516 EXPORT_SYMBOL(llog_recov_thread_fini);
518 static int llog_recov_thread_replay(struct llog_ctxt *ctxt,
521 struct obd_device *obd = ctxt->loc_obd;
522 struct llog_process_cat_args *lpca;
526 if (obd->obd_stopping)
530 * This will be balanced in llog_cat_process_thread()
537 lpca->lpca_arg = arg;
540 * This will be balanced in llog_cat_process_thread()
542 lpca->lpca_ctxt = llog_ctxt_get(ctxt);
543 if (!lpca->lpca_ctxt) {
547 rc = cfs_create_thread(llog_cat_process_thread, lpca, CFS_DAEMON_FLAGS);
549 CERROR("Error starting llog_cat_process_thread(): %d\n", rc);
553 CDEBUG(D_HA, "Started llog_cat_process_thread(): %d\n", rc);
560 int llog_obd_repl_connect(struct llog_ctxt *ctxt,
561 struct llog_logid *logid, struct llog_gen *gen,
562 struct obd_uuid *uuid)
568 * Send back cached llcd from llog before recovery if we have any.
569 * This is void is nothing cached is found there.
571 llog_sync(ctxt, NULL);
574 * Start recovery in separate thread.
576 cfs_mutex_down(&ctxt->loc_sem);
577 ctxt->loc_gen = *gen;
578 rc = llog_recov_thread_replay(ctxt, ctxt->llog_proc_cb, logid);
579 cfs_mutex_up(&ctxt->loc_sem);
583 EXPORT_SYMBOL(llog_obd_repl_connect);
586 * Deleted objects have a commit callback that cancels the MDS
587 * log record for the deletion. The commit callback calls this
590 int llog_obd_repl_cancel(struct llog_ctxt *ctxt,
591 struct lov_stripe_md *lsm, int count,
592 struct llog_cookie *cookies, int flags)
594 struct llog_commit_master *lcm;
595 struct llog_canceld_ctxt *llcd;
599 LASSERT(ctxt != NULL);
601 cfs_mutex_down(&ctxt->loc_sem);
602 if (!ctxt->loc_lcm) {
603 CDEBUG(D_RPCTRACE, "No lcm for ctxt %p\n", ctxt);
604 GOTO(out, rc = -ENODEV);
607 CDEBUG(D_INFO, "cancel on lsm %p\n", lcm);
610 * Let's check if we have all structures alive. We also check for
611 * possible shutdown. Do nothing if we're stopping.
613 if (ctxt->loc_imp == NULL) {
614 CDEBUG(D_RPCTRACE, "No import for ctxt %p\n", ctxt);
615 GOTO(out, rc = -ENODEV);
618 if (cfs_test_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags)) {
619 CDEBUG(D_RPCTRACE, "Commit thread is stopping for ctxt %p\n",
621 GOTO(out, rc = -ENODEV);
624 llcd = ctxt->loc_llcd;
626 if (count > 0 && cookies != NULL) {
628 * Get new llcd from ctxt if required.
631 llcd = llcd_get(ctxt);
633 GOTO(out, rc = -ENOMEM);
635 * Allocation is successful, let's check for stop
636 * flag again to fall back as soon as possible.
638 if (cfs_test_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags))
639 GOTO(out, rc = -ENODEV);
643 * Llcd does not have enough room for @cookies. Let's push
644 * it out and allocate new one.
646 if (!llcd_fit(llcd, cookies)) {
647 rc = llcd_push(ctxt);
650 llcd = llcd_get(ctxt);
652 GOTO(out, rc = -ENOMEM);
654 * Allocation is successful, let's check for stop
655 * flag again to fall back as soon as possible.
657 if (cfs_test_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags))
658 GOTO(out, rc = -ENODEV);
662 * Copy cookies to @llcd, no matter old or new allocated
665 llcd_copy(llcd, cookies);
669 * Let's check if we need to send copied @cookies asap. If yes
672 if (llcd && (flags & OBD_LLOG_FL_SENDNOW)) {
673 CDEBUG(D_RPCTRACE, "Sync llcd %p\n", llcd);
674 rc = llcd_push(ctxt);
682 cfs_mutex_up(&ctxt->loc_sem);
685 EXPORT_SYMBOL(llog_obd_repl_cancel);
687 int llog_obd_repl_sync(struct llog_ctxt *ctxt, struct obd_export *exp)
693 * Flush any remaining llcd.
695 cfs_mutex_down(&ctxt->loc_sem);
696 if (exp && (ctxt->loc_imp == exp->exp_imp_reverse)) {
698 * This is ost->mds connection, we can't be sure that mds
699 * can still receive cookies, let's killed the cached llcd.
701 CDEBUG(D_RPCTRACE, "Kill cached llcd\n");
703 cfs_mutex_up(&ctxt->loc_sem);
706 * This is either llog_sync() from generic llog code or sync
707 * on client disconnect. In either way let's do it and send
708 * llcds to the target with waiting for completion.
710 CDEBUG(D_RPCTRACE, "Sync cached llcd\n");
711 cfs_mutex_up(&ctxt->loc_sem);
712 rc = llog_cancel(ctxt, NULL, 0, NULL, OBD_LLOG_FL_SENDNOW);
716 EXPORT_SYMBOL(llog_obd_repl_sync);
718 #else /* !__KERNEL__ */
720 int llog_obd_repl_cancel(struct llog_ctxt *ctxt,
721 struct lov_stripe_md *lsm, int count,
722 struct llog_cookie *cookies, int flags)
729 * Module init time fucntion. Initializes slab for llcd objects.
731 int llog_recov_init(void)
735 llcd_size = CFS_PAGE_SIZE -
736 lustre_msg_size(LUSTRE_MSG_MAGIC_V2, 1, NULL);
737 llcd_size += offsetof(struct llog_canceld_ctxt, llcd_cookies);
738 llcd_cache = cfs_mem_cache_create("llcd_cache", llcd_size, 0, 0);
740 CERROR("Error allocating llcd cache\n");
747 * Module fini time fucntion. Releases slab for llcd objects.
749 void llog_recov_fini(void)
752 * Kill llcd cache when thread is stopped and we're sure no
757 * In 2.6.22 cfs_mem_cache_destroy() will not return error
758 * for busy resources. Let's check it another way.
760 LASSERTF(cfs_atomic_read(&llcd_count) == 0,
761 "Can't destroy llcd cache! Number of "
762 "busy llcds: %d\n", cfs_atomic_read(&llcd_count));
763 cfs_mem_cache_destroy(llcd_cache);