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 2008 Sun Microsystems, Inc. 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 atomic_t llcd_count = 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 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(llcd, llcd_cache, CFS_ALLOC_STD, size + overhead);
111 CFS_INIT_LIST_HEAD(&llcd->llcd_list);
112 llcd->llcd_cookiebytes = 0;
113 llcd->llcd_size = size;
115 spin_lock(&lcm->lcm_lock);
116 llcd->llcd_lcm = lcm;
117 atomic_inc(&lcm->lcm_count);
118 list_add_tail(&llcd->llcd_list, &lcm->lcm_llcds);
119 spin_unlock(&lcm->lcm_lock);
120 atomic_inc(&llcd_count);
122 CDEBUG(D_RPCTRACE, "Alloc llcd %p on lcm %p (%d)\n",
123 llcd, lcm, 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 (atomic_read(&lcm->lcm_count) == 0) {
138 CERROR("Invalid llcd free %p\n", llcd);
139 llcd_print(llcd, __FUNCTION__, __LINE__);
142 spin_lock(&lcm->lcm_lock);
143 LASSERT(!list_empty(&llcd->llcd_list));
144 list_del_init(&llcd->llcd_list);
145 atomic_dec(&lcm->lcm_count);
146 spin_unlock(&lcm->lcm_lock);
148 CDEBUG(D_RPCTRACE, "Free llcd %p on lcm %p (%d)\n",
149 llcd, lcm, atomic_read(&lcm->lcm_count));
152 LASSERT(atomic_read(&llcd_count) > 0);
153 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(struct ptlrpc_request *req, void *noused, int rc)
190 struct llog_canceld_ctxt *llcd = req->rq_async_args.pointer_arg[0];
191 CDEBUG(D_RPCTRACE, "Sent llcd %p (%d) - killing it\n", llcd, rc);
197 * Send @llcd to remote node. Free llcd uppon completion or error. Sending
198 * is performed in async style so this function will return asap without
201 static int llcd_send(struct llog_canceld_ctxt *llcd)
203 int size[2] = { sizeof(struct ptlrpc_body),
204 llcd->llcd_cookiebytes };
205 char *bufs[2] = { NULL, (char *)llcd->llcd_cookies };
206 struct obd_import *import = NULL;
207 struct llog_commit_master *lcm;
208 struct ptlrpc_request *req;
209 struct llog_ctxt *ctxt;
213 ctxt = llcd->llcd_ctxt;
215 CERROR("Invalid llcd with NULL ctxt found (%p)\n",
217 llcd_print(llcd, __FUNCTION__, __LINE__);
220 LASSERT_SEM_LOCKED(&ctxt->loc_sem);
222 if (llcd->llcd_cookiebytes == 0)
225 lcm = llcd->llcd_lcm;
228 * Check if we're in exit stage. Do not send llcd in
231 if (test_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags))
232 GOTO(exit, rc = -ENODEV);
234 CDEBUG(D_RPCTRACE, "Sending llcd %p\n", llcd);
236 import = llcd->llcd_ctxt->loc_imp;
237 if (!import || (import == LP_POISON) ||
238 (import->imp_client == LP_POISON)) {
239 CERROR("Invalid import %p for llcd %p\n",
241 GOTO(exit, rc = -ENODEV);
244 OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_RECOV, 10);
247 * No need to get import here as it is already done in
248 * llog_receptor_accept().
250 req = ptlrpc_prep_req(import, LUSTRE_LOG_VERSION,
251 OBD_LOG_CANCEL, 2, size, bufs);
253 CERROR("Can't allocate request for sending llcd %p\n",
255 GOTO(exit, rc = -ENOMEM);
259 * Check if we're in exit stage again. Do not send llcd in
262 if (test_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags)) {
263 ptlrpc_req_finished(req);
264 GOTO(exit, rc = -ENODEV);
268 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
269 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
270 ptlrpc_req_set_repsize(req, 1, NULL);
271 ptlrpc_at_set_req_timeout(req);
272 req->rq_interpret_reply = llcd_interpret;
273 req->rq_async_args.pointer_arg[0] = llcd;
275 /* llog cancels will be replayed after reconnect so this will do twice
276 * first from replay llog, second for resended rpc */
277 req->rq_no_delay = req->rq_no_resend = 1;
279 rc = ptlrpc_set_add_new_req(&lcm->lcm_pc, req);
281 ptlrpc_req_finished(req);
286 CDEBUG(D_RPCTRACE, "Refused llcd %p\n", llcd);
292 * Attach @llcd to @ctxt. Establish llcd vs. ctxt reserve connection
293 * so hat they can refer each other.
296 llcd_attach(struct llog_ctxt *ctxt, struct llog_canceld_ctxt *llcd)
298 LASSERT(ctxt != NULL && llcd != NULL);
299 LASSERT_SEM_LOCKED(&ctxt->loc_sem);
300 LASSERT(ctxt->loc_llcd == NULL);
301 llcd->llcd_ctxt = llog_ctxt_get(ctxt);
302 ctxt->loc_llcd = llcd;
304 CDEBUG(D_RPCTRACE, "Attach llcd %p to ctxt %p\n",
311 * Opposite to llcd_attach(). Detaches llcd from its @ctxt. This makes
312 * sure that this llcd will not be found another time we try to cancel.
314 static struct llog_canceld_ctxt *llcd_detach(struct llog_ctxt *ctxt)
316 struct llog_canceld_ctxt *llcd;
318 LASSERT(ctxt != NULL);
319 LASSERT_SEM_LOCKED(&ctxt->loc_sem);
321 llcd = ctxt->loc_llcd;
325 CDEBUG(D_RPCTRACE, "Detach llcd %p from ctxt %p\n",
328 ctxt->loc_llcd = NULL;
334 * Return @llcd cached in @ctxt. Allocate new one if required. Attach it
335 * to ctxt so that it may be used for gathering cookies and sending.
337 static struct llog_canceld_ctxt *llcd_get(struct llog_ctxt *ctxt)
339 struct llog_canceld_ctxt *llcd;
341 llcd = llcd_alloc(ctxt->loc_lcm);
343 CERROR("Can't alloc an llcd for ctxt %p\n", ctxt);
346 llcd_attach(ctxt, llcd);
351 * Deatch llcd from its @ctxt. Free llcd.
353 static void llcd_put(struct llog_ctxt *ctxt)
355 struct llog_canceld_ctxt *llcd;
357 llcd = llcd_detach(ctxt);
363 * Detach llcd from its @ctxt so that nobody will find it with try to
364 * re-use. Send llcd to remote node.
366 static int llcd_push(struct llog_ctxt *ctxt)
368 struct llog_canceld_ctxt *llcd;
372 * Make sure that this llcd will not be sent again as we detach
375 llcd = llcd_detach(ctxt);
377 CERROR("Invalid detached llcd found %p\n", llcd);
378 llcd_print(llcd, __FUNCTION__, __LINE__);
382 rc = llcd_send(llcd);
384 CERROR("Couldn't send llcd %p (%d)\n", llcd, rc);
389 * Start recovery thread which actually deals llcd sending. This
390 * is all ptlrpc standard thread based so there is not much of work
393 int llog_recov_thread_start(struct llog_commit_master *lcm)
398 rc = ptlrpcd_start(lcm->lcm_name, &lcm->lcm_pc);
400 CERROR("Error %d while starting recovery thread %s\n",
406 EXPORT_SYMBOL(llog_recov_thread_start);
409 * Stop recovery thread. Complement to llog_recov_thread_start().
411 void llog_recov_thread_stop(struct llog_commit_master *lcm, int force)
416 * Let all know that we're stopping. This will also make
417 * llcd_send() refuse any new llcds.
419 set_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags);
422 * Stop processing thread. No new rpcs will be accepted for
423 * for processing now.
425 ptlrpcd_stop(&lcm->lcm_pc, force);
428 * By this point no alive inflight llcds should be left. Only
429 * those forgotten in sync may still be attached to ctxt. Let's
432 if (atomic_read(&lcm->lcm_count) != 0) {
433 struct llog_canceld_ctxt *llcd;
434 struct list_head *tmp;
436 CERROR("Busy llcds found (%d) on lcm %p\n",
437 atomic_read(&lcm->lcm_count) == 0, lcm);
439 spin_lock(&lcm->lcm_lock);
440 list_for_each(tmp, &lcm->lcm_llcds) {
441 llcd = list_entry(tmp, struct llog_canceld_ctxt,
443 llcd_print(llcd, __FUNCTION__, __LINE__);
445 spin_unlock(&lcm->lcm_lock);
448 * No point to go further with busy llcds at this point
449 * as this is clear bug. It might mean we got hanging
450 * rpc which holds import ref and this means we will not
451 * be able to cleanup anyways.
453 * Or we just missed to kill them when they were not
454 * attached to ctxt. In this case our slab will remind
455 * us about this a bit later.
461 EXPORT_SYMBOL(llog_recov_thread_stop);
464 * Initialize commit master structure and start recovery thread on it.
466 struct llog_commit_master *llog_recov_thread_init(char *name)
468 struct llog_commit_master *lcm;
477 * Try to create threads with unique names.
479 snprintf(lcm->lcm_name, sizeof(lcm->lcm_name),
482 atomic_set(&lcm->lcm_count, 0);
483 spin_lock_init(&lcm->lcm_lock);
484 CFS_INIT_LIST_HEAD(&lcm->lcm_llcds);
485 rc = llog_recov_thread_start(lcm);
487 CERROR("Can't start commit thread, rc %d\n", rc);
495 EXPORT_SYMBOL(llog_recov_thread_init);
498 * Finalize commit master and its recovery thread.
500 void llog_recov_thread_fini(struct llog_commit_master *lcm, int force)
503 llog_recov_thread_stop(lcm, force);
507 EXPORT_SYMBOL(llog_recov_thread_fini);
509 static int llog_recov_thread_replay(struct llog_ctxt *ctxt,
512 struct obd_device *obd = ctxt->loc_obd;
513 struct llog_process_cat_args *lpca;
517 if (obd->obd_stopping)
521 * This will be balanced in llog_cat_process_thread()
528 lpca->lpca_arg = arg;
531 * This will be balanced in llog_cat_process_thread()
533 lpca->lpca_ctxt = llog_ctxt_get(ctxt);
534 if (!lpca->lpca_ctxt) {
538 rc = cfs_kernel_thread(llog_cat_process_thread, lpca,
539 CLONE_VM | CLONE_FILES);
541 CERROR("Error starting llog_cat_process_thread(): %d\n", rc);
545 CDEBUG(D_HA, "Started llog_cat_process_thread(): %d\n", rc);
552 int llog_obd_repl_connect(struct llog_ctxt *ctxt,
553 struct llog_logid *logid, struct llog_gen *gen,
554 struct obd_uuid *uuid)
560 * Send back cached llcd from llog before recovery if we have any.
561 * This is void is nothing cached is found there.
563 llog_sync(ctxt, NULL);
566 * Start recovery in separate thread.
568 mutex_down(&ctxt->loc_sem);
569 ctxt->loc_gen = *gen;
570 rc = llog_recov_thread_replay(ctxt, ctxt->llog_proc_cb, logid);
571 mutex_up(&ctxt->loc_sem);
575 EXPORT_SYMBOL(llog_obd_repl_connect);
578 * Deleted objects have a commit callback that cancels the MDS
579 * log record for the deletion. The commit callback calls this
582 int llog_obd_repl_cancel(struct llog_ctxt *ctxt,
583 struct lov_stripe_md *lsm, int count,
584 struct llog_cookie *cookies, int flags)
586 struct llog_commit_master *lcm;
587 struct llog_canceld_ctxt *llcd;
591 LASSERT(ctxt != NULL);
593 mutex_down(&ctxt->loc_sem);
595 CDEBUG(D_INFO, "cancel on lsm %p\n", lcm);
598 * Let's check if we have all structures alive. We also check for
599 * possible shutdown. Do nothing if we're stopping.
601 if (ctxt->loc_imp == NULL) {
602 CDEBUG(D_RPCTRACE, "No import for ctxt %p\n", ctxt);
603 GOTO(out, rc = -ENODEV);
606 if (test_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags)) {
607 CDEBUG(D_RPCTRACE, "Commit thread is stopping for ctxt %p\n",
609 GOTO(out, rc = -ENODEV);
612 llcd = ctxt->loc_llcd;
614 if (count > 0 && cookies != NULL) {
616 * Get new llcd from ctxt if required.
619 llcd = llcd_get(ctxt);
621 GOTO(out, rc = -ENOMEM);
623 * Allocation is successful, let's check for stop
624 * flag again to fall back as soon as possible.
626 if (test_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags))
627 GOTO(out, rc = -ENODEV);
631 * Llcd does not have enough room for @cookies. Let's push
632 * it out and allocate new one.
634 if (!llcd_fit(llcd, cookies)) {
635 rc = llcd_push(ctxt);
638 llcd = llcd_get(ctxt);
640 GOTO(out, rc = -ENOMEM);
642 * Allocation is successful, let's check for stop
643 * flag again to fall back as soon as possible.
645 if (test_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags))
646 GOTO(out, rc = -ENODEV);
650 * Copy cookies to @llcd, no matter old or new allocated
653 llcd_copy(llcd, cookies);
657 * Let's check if we need to send copied @cookies asap. If yes
660 if (llcd && (flags & OBD_LLOG_FL_SENDNOW)) {
661 CDEBUG(D_RPCTRACE, "Sync llcd %p\n", llcd);
662 rc = llcd_push(ctxt);
670 mutex_up(&ctxt->loc_sem);
673 EXPORT_SYMBOL(llog_obd_repl_cancel);
675 int llog_obd_repl_sync(struct llog_ctxt *ctxt, struct obd_export *exp)
681 * Flush any remaining llcd.
683 mutex_down(&ctxt->loc_sem);
684 if (exp && (ctxt->loc_imp == exp->exp_imp_reverse)) {
686 * This is ost->mds connection, we can't be sure that mds
687 * can still receive cookies, let's killed the cached llcd.
689 CDEBUG(D_RPCTRACE, "Kill cached llcd\n");
691 mutex_up(&ctxt->loc_sem);
694 * This is either llog_sync() from generic llog code or sync
695 * on client disconnect. In either way let's do it and send
696 * llcds to the target with waiting for completion.
698 CDEBUG(D_RPCTRACE, "Sync cached llcd\n");
699 mutex_up(&ctxt->loc_sem);
700 rc = llog_cancel(ctxt, NULL, 0, NULL, OBD_LLOG_FL_SENDNOW);
704 EXPORT_SYMBOL(llog_obd_repl_sync);
706 #else /* !__KERNEL__ */
708 int llog_obd_repl_cancel(struct llog_ctxt *ctxt,
709 struct lov_stripe_md *lsm, int count,
710 struct llog_cookie *cookies, int flags)
717 * Module init time fucntion. Initializes slab for llcd objects.
719 int llog_recov_init(void)
723 llcd_size = CFS_PAGE_SIZE -
724 lustre_msg_size(LUSTRE_MSG_MAGIC_V2, 1, NULL);
725 llcd_size += offsetof(struct llog_canceld_ctxt, llcd_cookies);
726 llcd_cache = cfs_mem_cache_create("llcd_cache", llcd_size, 0, 0);
728 CERROR("Error allocating llcd cache\n");
735 * Module fini time fucntion. Releases slab for llcd objects.
737 void llog_recov_fini(void)
740 * Kill llcd cache when thread is stopped and we're sure no
745 * In 2.6.22 cfs_mem_cache_destroy() will not return error
746 * for busy resources. Let's check it another way.
748 LASSERTF(atomic_read(&llcd_count) == 0,
749 "Can't destroy llcd cache! Number of "
750 "busy llcds: %d\n", atomic_read(&llcd_count));
751 cfs_mem_cache_destroy(llcd_cache);