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 * Payload of lustre_msg V2 is bigger.
103 llcd_size = CFS_PAGE_SIZE -
104 lustre_msg_size(LUSTRE_MSG_MAGIC_V2, 1, NULL);
105 llcd_size += offsetof(struct llog_canceld_ctxt, llcd_cookies);
106 OBD_SLAB_ALLOC(llcd, llcd_cache, CFS_ALLOC_STD, llcd_size);
110 CFS_INIT_LIST_HEAD(&llcd->llcd_list);
111 llcd->llcd_size = llcd_size;
112 llcd->llcd_cookiebytes = 0;
114 spin_lock(&lcm->lcm_lock);
115 llcd->llcd_lcm = lcm;
116 atomic_inc(&lcm->lcm_count);
117 list_add_tail(&llcd->llcd_list, &lcm->lcm_llcds);
118 spin_unlock(&lcm->lcm_lock);
119 atomic_inc(&llcd_count);
121 CDEBUG(D_RPCTRACE, "Alloc llcd %p on lcm %p (%d)\n",
122 llcd, lcm, atomic_read(&lcm->lcm_count));
128 * Returns passed llcd to cache.
130 static void llcd_free(struct llog_canceld_ctxt *llcd)
132 struct llog_commit_master *lcm = llcd->llcd_lcm;
135 if (atomic_read(&lcm->lcm_count) == 0) {
136 CERROR("Invalid llcd free %p\n", llcd);
137 llcd_print(llcd, __FUNCTION__, __LINE__);
140 spin_lock(&lcm->lcm_lock);
141 LASSERT(!list_empty(&llcd->llcd_list));
142 list_del_init(&llcd->llcd_list);
143 atomic_dec(&lcm->lcm_count);
144 spin_unlock(&lcm->lcm_lock);
147 CDEBUG(D_RPCTRACE, "Free llcd %p on lcm %p (%d)\n",
148 llcd, lcm, atomic_read(&lcm->lcm_count));
150 LASSERT(atomic_read(&llcd_count) > 0);
151 atomic_dec(&llcd_count);
152 OBD_SLAB_FREE(llcd, llcd_cache, llcd->llcd_size);
156 * Copy passed @cookies to @llcd.
158 static void llcd_copy(struct llog_canceld_ctxt *llcd,
159 struct llog_cookie *cookies)
161 memcpy((char *)llcd->llcd_cookies + llcd->llcd_cookiebytes,
162 cookies, sizeof(*cookies));
163 llcd->llcd_cookiebytes += sizeof(*cookies);
167 * Checks if passed cookie fits into llcd free space buffer. Returns
168 * 1 if yes and 0 otherwise.
170 static int llcd_fit(struct llog_canceld_ctxt *llcd,
171 struct llog_cookie *cookies)
173 return (llcd->llcd_size -
174 llcd->llcd_cookiebytes) >= sizeof(*cookies);
178 * Llcd completion function. Called uppon llcd send finish regardless
179 * sending result. Error is passed in @rc. Note, that this will be called
180 * in cleanup time when all inflight rpcs aborted.
183 llcd_interpret(const struct lu_env *env,
184 struct ptlrpc_request *req, void *noused, int rc)
186 struct llog_canceld_ctxt *llcd = req->rq_async_args.pointer_arg[0];
187 CDEBUG(D_RPCTRACE, "Sent llcd %p (%d) - killing it\n", llcd, rc);
193 * Send @llcd to remote node. Free llcd uppon completion or error. Sending
194 * is performed in async style so this function will return asap without
197 static int llcd_send(struct llog_canceld_ctxt *llcd)
199 char *bufs[2] = { NULL, (char *)llcd->llcd_cookies };
200 struct obd_import *import = NULL;
201 struct llog_commit_master *lcm;
202 struct ptlrpc_request *req;
203 struct llog_ctxt *ctxt;
207 ctxt = llcd->llcd_ctxt;
209 CERROR("Invalid llcd with NULL ctxt found (%p)\n",
211 llcd_print(llcd, __FUNCTION__, __LINE__);
214 LASSERT_SEM_LOCKED(&ctxt->loc_sem);
216 if (llcd->llcd_cookiebytes == 0)
219 lcm = llcd->llcd_lcm;
222 * Check if we're in exit stage. Do not send llcd in
225 if (test_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags))
226 GOTO(exit, rc = -ENODEV);
228 CDEBUG(D_RPCTRACE, "Sending llcd %p\n", llcd);
230 import = llcd->llcd_ctxt->loc_imp;
231 if (!import || (import == LP_POISON) ||
232 (import->imp_client == LP_POISON)) {
233 CERROR("Invalid import %p for llcd %p\n",
235 GOTO(exit, rc = -ENODEV);
238 OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_RECOV, 10);
241 * No need to get import here as it is already done in
242 * llog_receptor_accept().
244 req = ptlrpc_request_alloc(import, &RQF_LOG_CANCEL);
246 CERROR("Can't allocate request for sending llcd %p\n",
248 GOTO(exit, rc = -ENOMEM);
250 req_capsule_set_size(&req->rq_pill, &RMF_LOGCOOKIES,
251 RCL_CLIENT, llcd->llcd_cookiebytes);
253 rc = ptlrpc_request_bufs_pack(req, LUSTRE_LOG_VERSION,
254 OBD_LOG_CANCEL, bufs, NULL);
256 ptlrpc_request_free(req);
260 ptlrpc_at_set_req_timeout(req);
261 ptlrpc_request_set_replen(req);
264 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
265 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
266 req->rq_interpret_reply = (ptlrpc_interpterer_t)llcd_interpret;
267 req->rq_async_args.pointer_arg[0] = llcd;
268 rc = ptlrpc_set_add_new_req(&lcm->lcm_pc, req);
270 ptlrpc_request_free(req);
275 CDEBUG(D_RPCTRACE, "Refused llcd %p\n", llcd);
281 * Attach @llcd to @ctxt. Establish llcd vs. ctxt reserve connection
282 * so hat they can refer each other.
285 llcd_attach(struct llog_ctxt *ctxt, struct llog_canceld_ctxt *llcd)
287 LASSERT(ctxt != NULL && llcd != NULL);
288 LASSERT_SEM_LOCKED(&ctxt->loc_sem);
289 LASSERT(ctxt->loc_llcd == NULL);
290 llcd->llcd_ctxt = llog_ctxt_get(ctxt);
291 ctxt->loc_llcd = llcd;
293 CDEBUG(D_RPCTRACE, "Attach llcd %p to ctxt %p\n",
300 * Opposite to llcd_attach(). Detaches llcd from its @ctxt. This makes
301 * sure that this llcd will not be found another time we try to cancel.
303 static struct llog_canceld_ctxt *llcd_detach(struct llog_ctxt *ctxt)
305 struct llog_canceld_ctxt *llcd;
307 LASSERT(ctxt != NULL);
308 LASSERT_SEM_LOCKED(&ctxt->loc_sem);
310 llcd = ctxt->loc_llcd;
314 CDEBUG(D_RPCTRACE, "Detach llcd %p from ctxt %p\n",
317 ctxt->loc_llcd = NULL;
323 * Return @llcd cached in @ctxt. Allocate new one if required. Attach it
324 * to ctxt so that it may be used for gathering cookies and sending.
326 static struct llog_canceld_ctxt *llcd_get(struct llog_ctxt *ctxt)
328 struct llog_canceld_ctxt *llcd;
330 llcd = llcd_alloc(ctxt->loc_lcm);
332 CERROR("Can't alloc an llcd for ctxt %p\n", ctxt);
335 llcd_attach(ctxt, llcd);
340 * Deatch llcd from its @ctxt. Free llcd.
342 static void llcd_put(struct llog_ctxt *ctxt)
344 struct llog_canceld_ctxt *llcd;
346 llcd = llcd_detach(ctxt);
352 * Detach llcd from its @ctxt so that nobody will find it with try to
353 * re-use. Send llcd to remote node.
355 static int llcd_push(struct llog_ctxt *ctxt)
357 struct llog_canceld_ctxt *llcd;
361 * Make sure that this llcd will not be sent again as we detach
364 llcd = llcd_detach(ctxt);
366 CERROR("Invalid detached llcd found %p\n", llcd);
367 llcd_print(llcd, __FUNCTION__, __LINE__);
371 rc = llcd_send(llcd);
373 CERROR("Couldn't send llcd %p (%d)\n", llcd, rc);
378 * Start recovery thread which actually deals llcd sending. This
379 * is all ptlrpc standard thread based so there is not much of work
382 int llog_recov_thread_start(struct llog_commit_master *lcm)
387 rc = ptlrpcd_start(lcm->lcm_name, &lcm->lcm_pc);
389 CERROR("Error %d while starting recovery thread %s\n",
395 EXPORT_SYMBOL(llog_recov_thread_start);
398 * Stop recovery thread. Complement to llog_recov_thread_start().
400 void llog_recov_thread_stop(struct llog_commit_master *lcm, int force)
405 * Let all know that we're stopping. This will also make
406 * llcd_send() refuse any new llcds.
408 set_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags);
411 * Stop processing thread. No new rpcs will be accepted for
412 * for processing now.
414 ptlrpcd_stop(&lcm->lcm_pc, force);
417 * By this point no alive inflight llcds should be left. Only
418 * those forgotten in sync may still be attached to ctxt. Let's
421 if (atomic_read(&lcm->lcm_count) != 0) {
422 struct llog_canceld_ctxt *llcd;
423 struct list_head *tmp;
425 CERROR("Busy llcds found (%d) on lcm %p\n",
426 atomic_read(&lcm->lcm_count) == 0, lcm);
428 spin_lock(&lcm->lcm_lock);
429 list_for_each(tmp, &lcm->lcm_llcds) {
430 llcd = list_entry(tmp, struct llog_canceld_ctxt,
432 llcd_print(llcd, __FUNCTION__, __LINE__);
434 spin_unlock(&lcm->lcm_lock);
437 * No point to go further with busy llcds at this point
438 * as this is clear bug. It might mean we got hanging
439 * rpc which holds import ref and this means we will not
440 * be able to cleanup anyways.
442 * Or we just missed to kill them when they were not
443 * attached to ctxt. In this case our slab will remind
444 * us about this a bit later.
450 EXPORT_SYMBOL(llog_recov_thread_stop);
453 * Initialize commit master structure and start recovery thread on it.
455 struct llog_commit_master *llog_recov_thread_init(char *name)
457 struct llog_commit_master *lcm;
466 * Try to create threads with unique names.
468 snprintf(lcm->lcm_name, sizeof(lcm->lcm_name),
469 "ll_log_commit_%s", name);
471 strncpy(lcm->lcm_name, name, sizeof(lcm->lcm_name));
472 atomic_set(&lcm->lcm_count, 0);
473 spin_lock_init(&lcm->lcm_lock);
474 CFS_INIT_LIST_HEAD(&lcm->lcm_llcds);
475 rc = llog_recov_thread_start(lcm);
477 CERROR("Can't start commit thread, rc %d\n", rc);
485 EXPORT_SYMBOL(llog_recov_thread_init);
488 * Finalize commit master and its recovery thread.
490 void llog_recov_thread_fini(struct llog_commit_master *lcm, int force)
493 llog_recov_thread_stop(lcm, force);
497 EXPORT_SYMBOL(llog_recov_thread_fini);
499 static int llog_recov_thread_replay(struct llog_ctxt *ctxt,
502 struct obd_device *obd = ctxt->loc_obd;
503 struct llog_process_cat_args *lpca;
507 if (obd->obd_stopping)
511 * This will be balanced in llog_cat_process_thread()
518 lpca->lpca_arg = arg;
521 * This will be balanced in llog_cat_process_thread()
523 lpca->lpca_ctxt = llog_ctxt_get(ctxt);
524 if (!lpca->lpca_ctxt) {
528 rc = cfs_kernel_thread(llog_cat_process_thread, lpca,
529 CLONE_VM | CLONE_FILES);
531 CERROR("Error starting llog_cat_process_thread(): %d\n", rc);
535 CDEBUG(D_HA, "Started llog_cat_process_thread(): %d\n", rc);
542 int llog_obd_repl_connect(struct llog_ctxt *ctxt,
543 struct llog_logid *logid, struct llog_gen *gen,
544 struct obd_uuid *uuid)
550 * Send back cached llcd from llog before recovery if we have any.
551 * This is void is nothing cached is found there.
553 llog_sync(ctxt, NULL);
556 * Start recovery in separate thread.
558 mutex_down(&ctxt->loc_sem);
559 ctxt->loc_gen = *gen;
560 rc = llog_recov_thread_replay(ctxt, ctxt->llog_proc_cb, logid);
561 mutex_up(&ctxt->loc_sem);
565 EXPORT_SYMBOL(llog_obd_repl_connect);
568 * Deleted objects have a commit callback that cancels the MDS
569 * log record for the deletion. The commit callback calls this
572 int llog_obd_repl_cancel(struct llog_ctxt *ctxt,
573 struct lov_stripe_md *lsm, int count,
574 struct llog_cookie *cookies, int flags)
576 struct llog_commit_master *lcm;
577 struct llog_canceld_ctxt *llcd;
581 LASSERT(ctxt != NULL);
583 mutex_down(&ctxt->loc_sem);
587 * Let's check if we have all structures alive. We also check for
588 * possible shutdown. Do nothing if we're stopping.
590 if (ctxt->loc_imp == NULL) {
591 CDEBUG(D_RPCTRACE, "No import for ctxt %p\n", ctxt);
592 GOTO(out, rc = -ENODEV);
595 if (test_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags)) {
596 CDEBUG(D_RPCTRACE, "Commit thread is stopping for ctxt %p\n",
598 GOTO(out, rc = -ENODEV);
601 llcd = ctxt->loc_llcd;
603 if (count > 0 && cookies != NULL) {
605 * Get new llcd from ctxt if required.
608 llcd = llcd_get(ctxt);
610 GOTO(out, rc = -ENOMEM);
612 * Allocation is successful, let's check for stop
613 * flag again to fall back as soon as possible.
615 if (test_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags))
616 GOTO(out, rc = -ENODEV);
620 * Llcd does not have enough room for @cookies. Let's push
621 * it out and allocate new one.
623 if (!llcd_fit(llcd, cookies)) {
624 rc = llcd_push(ctxt);
627 llcd = llcd_get(ctxt);
629 GOTO(out, rc = -ENOMEM);
631 * Allocation is successful, let's check for stop
632 * flag again to fall back as soon as possible.
634 if (test_bit(LLOG_LCM_FL_EXIT, &lcm->lcm_flags))
635 GOTO(out, rc = -ENODEV);
639 * Copy cookies to @llcd, no matter old or new allocated
642 llcd_copy(llcd, cookies);
646 * Let's check if we need to send copied @cookies asap. If yes
649 if (llcd && (flags & OBD_LLOG_FL_SENDNOW)) {
650 CDEBUG(D_RPCTRACE, "Sync llcd %p\n", llcd);
651 rc = llcd_push(ctxt);
659 mutex_up(&ctxt->loc_sem);
662 EXPORT_SYMBOL(llog_obd_repl_cancel);
664 int llog_obd_repl_sync(struct llog_ctxt *ctxt, struct obd_export *exp)
670 * Flush any remaining llcd.
672 mutex_down(&ctxt->loc_sem);
673 if (exp && (ctxt->loc_imp == exp->exp_imp_reverse)) {
675 * This is ost->mds connection, we can't be sure that mds
676 * can still receive cookies, let's killed the cached llcd.
678 CDEBUG(D_RPCTRACE, "Kill cached llcd\n");
680 mutex_up(&ctxt->loc_sem);
683 * This is either llog_sync() from generic llog code or sync
684 * on client disconnect. In either way let's do it and send
685 * llcds to the target with waiting for completion.
687 CDEBUG(D_RPCTRACE, "Sync cached llcd\n");
688 mutex_up(&ctxt->loc_sem);
689 rc = llog_cancel(ctxt, NULL, 0, NULL, OBD_LLOG_FL_SENDNOW);
693 EXPORT_SYMBOL(llog_obd_repl_sync);
695 #else /* !__KERNEL__ */
697 int llog_obd_repl_cancel(struct llog_ctxt *ctxt,
698 struct lov_stripe_md *lsm, int count,
699 struct llog_cookie *cookies, int flags)
706 * Module init time fucntion. Initializes slab for llcd objects.
708 int llog_recov_init(void)
712 llcd_size = CFS_PAGE_SIZE -
713 lustre_msg_size(LUSTRE_MSG_MAGIC_V2, 1, NULL);
714 llcd_size += offsetof(struct llog_canceld_ctxt, llcd_cookies);
715 llcd_cache = cfs_mem_cache_create("llcd_cache", llcd_size, 0, 0);
717 CERROR("Error allocating llcd cache\n");
724 * Module fini time fucntion. Releases slab for llcd objects.
726 void llog_recov_fini(void)
729 * Kill llcd cache when thread is stopped and we're sure no
734 * In 2.6.22 cfs_mem_cache_destroy() will not return error
735 * for busy resources. Let's check it another way.
737 LASSERTF(atomic_read(&llcd_count) == 0,
738 "Can't destroy llcd cache! Number of "
739 "busy llcds: %d\n", atomic_read(&llcd_count));
740 cfs_mem_cache_destroy(llcd_cache);