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 # include <libcfs/libcfs.h>
53 # include <libcfs/list.h>
54 # include <liblustre.h>
57 #include <obd_class.h>
58 #include <obd_support.h>
59 #include <obd_class.h>
60 #include <lustre_net.h>
61 #include <lnet/types.h>
62 #include <libcfs/list.h>
63 #include <lustre_log.h>
64 #include "ptlrpc_internal.h"
66 static cfs_atomic_t llcd_count = CFS_ATOMIC_INIT(0);
67 static cfs_mem_cache_t *llcd_cache = NULL;
71 LLOG_LCM_FL_START = 1 << 0,
72 LLOG_LCM_FL_EXIT = 1 << 1
75 struct llcd_async_args {
76 struct llog_canceld_ctxt *la_ctxt;
79 static void llcd_print(struct llog_canceld_ctxt *llcd,
80 const char *func, int line)
82 CDEBUG(D_RPCTRACE, "Llcd (%p) at %s:%d:\n", llcd, func, line);
83 CDEBUG(D_RPCTRACE, " size: %d\n", llcd->llcd_size);
84 CDEBUG(D_RPCTRACE, " ctxt: %p\n", llcd->llcd_ctxt);
85 CDEBUG(D_RPCTRACE, " lcm : %p\n", llcd->llcd_lcm);
86 CDEBUG(D_RPCTRACE, " cookiebytes : %d\n", llcd->llcd_cookiebytes);
90 * Allocate new llcd from cache, init it and return to caller.
91 * Bumps number of objects allocated.
93 static struct llog_canceld_ctxt *llcd_alloc(struct llog_commit_master *lcm)
95 struct llog_canceld_ctxt *llcd;
101 * We want to send one page of cookies with rpc header. This buffer
102 * will be assigned later to the rpc, this is why we preserve the
103 * space for rpc header.
105 size = CFS_PAGE_SIZE - lustre_msg_size(LUSTRE_MSG_MAGIC_V2, 1, NULL);
106 overhead = offsetof(struct llog_canceld_ctxt, llcd_cookies);
107 OBD_SLAB_ALLOC_GFP(llcd, llcd_cache, size + overhead, CFS_ALLOC_STD);
111 CFS_INIT_LIST_HEAD(&llcd->llcd_list);
112 llcd->llcd_cookiebytes = 0;
113 llcd->llcd_size = size;
115 cfs_spin_lock(&lcm->lcm_lock);
116 llcd->llcd_lcm = lcm;
117 cfs_atomic_inc(&lcm->lcm_count);
118 cfs_list_add_tail(&llcd->llcd_list, &lcm->lcm_llcds);
119 cfs_spin_unlock(&lcm->lcm_lock);
120 cfs_atomic_inc(&llcd_count);
122 CDEBUG(D_RPCTRACE, "Alloc llcd %p on lcm %p (%d)\n",
123 llcd, lcm, cfs_atomic_read(&lcm->lcm_count));
129 * Returns passed llcd to cache.
131 static void llcd_free(struct llog_canceld_ctxt *llcd)
133 struct llog_commit_master *lcm = llcd->llcd_lcm;
137 if (cfs_atomic_read(&lcm->lcm_count) == 0) {
138 CERROR("Invalid llcd free %p\n", llcd);
139 llcd_print(llcd, __FUNCTION__, __LINE__);
142 cfs_spin_lock(&lcm->lcm_lock);
143 LASSERT(!cfs_list_empty(&llcd->llcd_list));
144 cfs_list_del_init(&llcd->llcd_list);
145 cfs_atomic_dec(&lcm->lcm_count);
146 cfs_spin_unlock(&lcm->lcm_lock);
148 CDEBUG(D_RPCTRACE, "Free llcd %p on lcm %p (%d)\n",
149 llcd, lcm, cfs_atomic_read(&lcm->lcm_count));
152 LASSERT(cfs_atomic_read(&llcd_count) > 0);
153 cfs_atomic_dec(&llcd_count);
155 size = offsetof(struct llog_canceld_ctxt, llcd_cookies) +
157 OBD_SLAB_FREE(llcd, llcd_cache, size);
161 * Checks if passed cookie fits into llcd free space buffer. Returns
162 * 1 if yes and 0 otherwise.
165 llcd_fit(struct llog_canceld_ctxt *llcd, struct llog_cookie *cookies)
167 return (llcd->llcd_size - llcd->llcd_cookiebytes >= sizeof(*cookies));
171 * Copy passed @cookies to @llcd.
174 llcd_copy(struct llog_canceld_ctxt *llcd, struct llog_cookie *cookies)
176 LASSERT(llcd_fit(llcd, cookies));
177 memcpy((char *)llcd->llcd_cookies + llcd->llcd_cookiebytes,
178 cookies, sizeof(*cookies));
179 llcd->llcd_cookiebytes += sizeof(*cookies);
183 * Llcd completion function. Called uppon llcd send finish regardless
184 * sending result. Error is passed in @rc. Note, that this will be called
185 * in cleanup time when all inflight rpcs aborted.
188 llcd_interpret(const struct lu_env *env,
189 struct ptlrpc_request *req, void *args, int rc)
191 struct llcd_async_args *la = args;
192 struct llog_canceld_ctxt *llcd = la->la_ctxt;
194 CDEBUG(D_RPCTRACE, "Sent llcd %p (%d) - killing it\n", llcd, rc);
200 * Send @llcd to remote node. Free llcd uppon completion or error. Sending
201 * is performed in async style so this function will return asap without
204 static int llcd_send(struct llog_canceld_ctxt *llcd)
206 char *bufs[2] = { NULL, (char *)llcd->llcd_cookies };
207 struct obd_import *import = NULL;
208 struct llog_commit_master *lcm;
209 struct llcd_async_args *la;
210 struct ptlrpc_request *req;
211 struct llog_ctxt *ctxt;
215 ctxt = llcd->llcd_ctxt;
217 CERROR("Invalid llcd with NULL ctxt found (%p)\n",
219 llcd_print(llcd, __FUNCTION__, __LINE__);
222 LASSERT_MUTEX_LOCKED(&ctxt->loc_mutex);
224 if (llcd->llcd_cookiebytes == 0)
227 lcm = llcd->llcd_lcm;
230 * Check if we're in exit stage. Do not send llcd in
233 if (cfs_test_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags))
234 GOTO(exit, rc = -ENODEV);
236 CDEBUG(D_RPCTRACE, "Sending llcd %p\n", llcd);
238 import = llcd->llcd_ctxt->loc_imp;
239 if (!import || (import == LP_POISON) ||
240 (import->imp_client == LP_POISON)) {
241 CERROR("Invalid import %p for llcd %p\n",
243 GOTO(exit, rc = -ENODEV);
246 OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_RECOV, 10);
249 * No need to get import here as it is already done in
250 * llog_receptor_accept().
252 req = ptlrpc_request_alloc(import, &RQF_LOG_CANCEL);
254 CERROR("Can't allocate request for sending llcd %p\n",
256 GOTO(exit, rc = -ENOMEM);
258 req_capsule_set_size(&req->rq_pill, &RMF_LOGCOOKIES,
259 RCL_CLIENT, llcd->llcd_cookiebytes);
261 rc = ptlrpc_request_bufs_pack(req, LUSTRE_LOG_VERSION,
262 OBD_LOG_CANCEL, bufs, NULL);
264 ptlrpc_request_free(req);
268 ptlrpc_at_set_req_timeout(req);
269 ptlrpc_request_set_replen(req);
272 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
273 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
275 req->rq_interpret_reply = (ptlrpc_interpterer_t)llcd_interpret;
277 CLASSERT(sizeof(*la) <= sizeof(req->rq_async_args));
278 la = ptlrpc_req_async_args(req);
281 /* llog cancels will be replayed after reconnect so this will do twice
282 * first from replay llog, second for resended rpc */
283 req->rq_no_delay = req->rq_no_resend = 1;
285 ptlrpc_set_add_new_req(&lcm->lcm_pc, req);
288 CDEBUG(D_RPCTRACE, "Refused llcd %p\n", llcd);
294 * Attach @llcd to @ctxt. Establish llcd vs. ctxt reserve connection
295 * so hat they can refer each other.
298 llcd_attach(struct llog_ctxt *ctxt, struct llog_canceld_ctxt *llcd)
300 LASSERT(ctxt != NULL && llcd != NULL);
301 LASSERT_MUTEX_LOCKED(&ctxt->loc_mutex);
302 LASSERT(ctxt->loc_llcd == NULL);
303 llcd->llcd_ctxt = llog_ctxt_get(ctxt);
304 ctxt->loc_llcd = llcd;
306 CDEBUG(D_RPCTRACE, "Attach llcd %p to ctxt %p\n",
313 * Opposite to llcd_attach(). Detaches llcd from its @ctxt. This makes
314 * sure that this llcd will not be found another time we try to cancel.
316 static struct llog_canceld_ctxt *llcd_detach(struct llog_ctxt *ctxt)
318 struct llog_canceld_ctxt *llcd;
320 LASSERT(ctxt != NULL);
321 LASSERT_MUTEX_LOCKED(&ctxt->loc_mutex);
323 llcd = ctxt->loc_llcd;
327 CDEBUG(D_RPCTRACE, "Detach llcd %p from ctxt %p\n",
330 ctxt->loc_llcd = NULL;
336 * Return @llcd cached in @ctxt. Allocate new one if required. Attach it
337 * to ctxt so that it may be used for gathering cookies and sending.
339 static struct llog_canceld_ctxt *llcd_get(struct llog_ctxt *ctxt)
341 struct llog_canceld_ctxt *llcd;
343 llcd = llcd_alloc(ctxt->loc_lcm);
345 CERROR("Can't alloc an llcd for ctxt %p\n", ctxt);
348 llcd_attach(ctxt, llcd);
353 * Deatch llcd from its @ctxt. Free llcd.
355 static void llcd_put(struct llog_ctxt *ctxt)
357 struct llog_canceld_ctxt *llcd;
359 llcd = llcd_detach(ctxt);
365 * Detach llcd from its @ctxt so that nobody will find it with try to
366 * re-use. Send llcd to remote node.
368 static int llcd_push(struct llog_ctxt *ctxt)
370 struct llog_canceld_ctxt *llcd;
374 * Make sure that this llcd will not be sent again as we detach
377 llcd = llcd_detach(ctxt);
379 CERROR("Invalid detached llcd found %p\n", llcd);
380 llcd_print(llcd, __FUNCTION__, __LINE__);
384 rc = llcd_send(llcd);
386 CERROR("Couldn't send llcd %p (%d)\n", llcd, rc);
391 * Start recovery thread which actually deals llcd sending. This
392 * is all ptlrpc standard thread based so there is not much of work
395 int llog_recov_thread_start(struct llog_commit_master *lcm)
400 rc = ptlrpcd_start(-1, 1, lcm->lcm_name, &lcm->lcm_pc);
402 CERROR("Error %d while starting recovery thread %s\n",
408 EXPORT_SYMBOL(llog_recov_thread_start);
411 * Stop recovery thread. Complement to llog_recov_thread_start().
413 void llog_recov_thread_stop(struct llog_commit_master *lcm, int force)
418 * Let all know that we're stopping. This will also make
419 * llcd_send() refuse any new llcds.
421 cfs_set_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags);
424 * Stop processing thread. No new rpcs will be accepted for
425 * for processing now.
427 ptlrpcd_stop(&lcm->lcm_pc, force);
430 * By this point no alive inflight llcds should be left. Only
431 * those forgotten in sync may still be attached to ctxt. Let's
434 if (cfs_atomic_read(&lcm->lcm_count) != 0) {
435 struct llog_canceld_ctxt *llcd;
438 CERROR("Busy llcds found (%d) on lcm %p\n",
439 cfs_atomic_read(&lcm->lcm_count), lcm);
441 cfs_spin_lock(&lcm->lcm_lock);
442 cfs_list_for_each(tmp, &lcm->lcm_llcds) {
443 llcd = cfs_list_entry(tmp, struct llog_canceld_ctxt,
445 llcd_print(llcd, __FUNCTION__, __LINE__);
447 cfs_spin_unlock(&lcm->lcm_lock);
450 * No point to go further with busy llcds at this point
451 * as this is clear bug. It might mean we got hanging
452 * rpc which holds import ref and this means we will not
453 * be able to cleanup anyways.
455 * Or we just missed to kill them when they were not
456 * attached to ctxt. In this case our slab will remind
457 * us about this a bit later.
463 EXPORT_SYMBOL(llog_recov_thread_stop);
466 * Initialize commit master structure and start recovery thread on it.
468 struct llog_commit_master *llog_recov_thread_init(char *name)
470 struct llog_commit_master *lcm;
479 * Try to create threads with unique names.
481 snprintf(lcm->lcm_name, sizeof(lcm->lcm_name),
484 cfs_atomic_set(&lcm->lcm_count, 0);
485 cfs_atomic_set(&lcm->lcm_refcount, 1);
486 cfs_spin_lock_init(&lcm->lcm_lock);
487 CFS_INIT_LIST_HEAD(&lcm->lcm_llcds);
488 rc = llog_recov_thread_start(lcm);
490 CERROR("Can't start commit thread, rc %d\n", rc);
498 EXPORT_SYMBOL(llog_recov_thread_init);
501 * Finalize commit master and its recovery thread.
503 void llog_recov_thread_fini(struct llog_commit_master *lcm, int force)
506 llog_recov_thread_stop(lcm, force);
510 EXPORT_SYMBOL(llog_recov_thread_fini);
512 static int llog_recov_thread_replay(struct llog_ctxt *ctxt,
515 struct obd_device *obd = ctxt->loc_obd;
516 struct llog_process_cat_args *lpca;
520 if (obd->obd_stopping)
524 * This will be balanced in llog_cat_process_thread()
531 lpca->lpca_arg = arg;
534 * This will be balanced in llog_cat_process_thread()
536 lpca->lpca_ctxt = llog_ctxt_get(ctxt);
537 if (!lpca->lpca_ctxt) {
541 rc = cfs_create_thread(llog_cat_process_thread, lpca, CFS_DAEMON_FLAGS);
543 CERROR("Error starting llog_cat_process_thread(): %d\n", rc);
547 CDEBUG(D_HA, "Started llog_cat_process_thread(): %d\n", rc);
554 int llog_obd_repl_connect(struct llog_ctxt *ctxt,
555 struct llog_logid *logid, struct llog_gen *gen,
556 struct obd_uuid *uuid)
562 * Send back cached llcd from llog before recovery if we have any.
563 * This is void is nothing cached is found there.
565 llog_sync(ctxt, NULL, 0);
568 * Start recovery in separate thread.
570 cfs_mutex_lock(&ctxt->loc_mutex);
571 ctxt->loc_gen = *gen;
572 rc = llog_recov_thread_replay(ctxt, ctxt->llog_proc_cb, logid);
573 cfs_mutex_unlock(&ctxt->loc_mutex);
577 EXPORT_SYMBOL(llog_obd_repl_connect);
580 * Deleted objects have a commit callback that cancels the MDS
581 * log record for the deletion. The commit callback calls this
584 int llog_obd_repl_cancel(struct llog_ctxt *ctxt,
585 struct lov_stripe_md *lsm, int count,
586 struct llog_cookie *cookies, int flags)
588 struct llog_commit_master *lcm;
589 struct llog_canceld_ctxt *llcd;
593 LASSERT(ctxt != NULL);
595 cfs_mutex_lock(&ctxt->loc_mutex);
596 if (!ctxt->loc_lcm) {
597 CDEBUG(D_RPCTRACE, "No lcm for ctxt %p\n", ctxt);
598 GOTO(out, rc = -ENODEV);
601 CDEBUG(D_INFO, "cancel on lsm %p\n", lcm);
604 * Let's check if we have all structures alive. We also check for
605 * possible shutdown. Do nothing if we're stopping.
607 if (ctxt->loc_flags & LLOG_CTXT_FLAG_STOP) {
608 CDEBUG(D_RPCTRACE, "Last sync was done for ctxt %p\n", ctxt);
609 GOTO(out, rc = -ENODEV);
612 if (ctxt->loc_imp == NULL) {
613 CDEBUG(D_RPCTRACE, "No import for ctxt %p\n", ctxt);
614 GOTO(out, rc = -ENODEV);
617 if (cfs_test_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags)) {
618 CDEBUG(D_RPCTRACE, "Commit thread is stopping for ctxt %p\n",
620 GOTO(out, rc = -ENODEV);
623 llcd = ctxt->loc_llcd;
625 if (count > 0 && cookies != NULL) {
627 * Get new llcd from ctxt if required.
630 llcd = llcd_get(ctxt);
632 GOTO(out, rc = -ENOMEM);
634 * Allocation is successful, let's check for stop
635 * flag again to fall back as soon as possible.
637 if (cfs_test_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags))
638 GOTO(out, rc = -ENODEV);
642 * Llcd does not have enough room for @cookies. Let's push
643 * it out and allocate new one.
645 if (!llcd_fit(llcd, cookies)) {
646 rc = llcd_push(ctxt);
649 llcd = llcd_get(ctxt);
651 GOTO(out, rc = -ENOMEM);
653 * Allocation is successful, let's check for stop
654 * flag again to fall back as soon as possible.
656 if (cfs_test_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags))
657 GOTO(out, rc = -ENODEV);
661 * Copy cookies to @llcd, no matter old or new allocated
664 llcd_copy(llcd, cookies);
668 * Let's check if we need to send copied @cookies asap. If yes
671 if (llcd && (flags & OBD_LLOG_FL_SENDNOW)) {
672 CDEBUG(D_RPCTRACE, "Sync llcd %p\n", llcd);
673 rc = llcd_push(ctxt);
682 if (flags & OBD_LLOG_FL_EXIT)
683 ctxt->loc_flags = LLOG_CTXT_FLAG_STOP;
685 cfs_mutex_unlock(&ctxt->loc_mutex);
688 EXPORT_SYMBOL(llog_obd_repl_cancel);
690 int llog_obd_repl_sync(struct llog_ctxt *ctxt, struct obd_export *exp,
697 * Flush any remaining llcd.
699 cfs_mutex_lock(&ctxt->loc_mutex);
700 if (exp && (ctxt->loc_imp == exp->exp_imp_reverse)) {
702 * This is ost->mds connection, we can't be sure that mds
703 * can still receive cookies, let's killed the cached llcd.
705 CDEBUG(D_RPCTRACE, "Kill cached llcd\n");
708 if (flags & OBD_LLOG_FL_EXIT)
709 ctxt->loc_flags = LLOG_CTXT_FLAG_STOP;
711 cfs_mutex_unlock(&ctxt->loc_mutex);
714 * This is either llog_sync() from generic llog code or sync
715 * on client disconnect. In either way let's do it and send
716 * llcds to the target with waiting for completion.
718 CDEBUG(D_RPCTRACE, "Sync cached llcd\n");
719 cfs_mutex_unlock(&ctxt->loc_mutex);
720 rc = llog_cancel(ctxt, NULL, 0, NULL, OBD_LLOG_FL_SENDNOW |
725 EXPORT_SYMBOL(llog_obd_repl_sync);
727 #else /* !__KERNEL__ */
729 int llog_obd_repl_cancel(struct llog_ctxt *ctxt,
730 struct lov_stripe_md *lsm, int count,
731 struct llog_cookie *cookies, int flags)
738 * Module init time fucntion. Initializes slab for llcd objects.
740 int llog_recov_init(void)
744 llcd_size = CFS_PAGE_SIZE -
745 lustre_msg_size(LUSTRE_MSG_MAGIC_V2, 1, NULL);
746 llcd_size += offsetof(struct llog_canceld_ctxt, llcd_cookies);
747 llcd_cache = cfs_mem_cache_create("llcd_cache", llcd_size, 0, 0);
749 CERROR("Error allocating llcd cache\n");
756 * Module fini time fucntion. Releases slab for llcd objects.
758 void llog_recov_fini(void)
761 * Kill llcd cache when thread is stopped and we're sure no
766 * In 2.6.22 cfs_mem_cache_destroy() will not return error
767 * for busy resources. Let's check it another way.
769 LASSERTF(cfs_atomic_read(&llcd_count) == 0,
770 "Can't destroy llcd cache! Number of "
771 "busy llcds: %d\n", cfs_atomic_read(&llcd_count));
772 cfs_mem_cache_destroy(llcd_cache);