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) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2014, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/ptlrpc/recover.c
38 * Author: Mike Shaver <shaver@clusterfs.com>
41 #define DEBUG_SUBSYSTEM S_RPC
42 #include <libcfs/libcfs.h>
43 #include <obd_support.h>
44 #include <lustre_ha.h>
45 #include <lustre_net.h>
46 #include <lustre_import.h>
47 #include <lustre_export.h>
49 #include <obd_class.h>
50 #include <libcfs/list.h>
52 #include "ptlrpc_internal.h"
55 * Start recovery on disconnected import.
56 * This is done by just attempting a connect
58 void ptlrpc_initiate_recovery(struct obd_import *imp)
62 CDEBUG(D_HA, "%s: starting recovery\n", obd2cli_tgt(imp->imp_obd));
63 ptlrpc_connect_import(imp);
69 * Identify what request from replay list needs to be replayed next
70 * (based on what we have already replayed) and send it to server.
72 int ptlrpc_replay_next(struct obd_import *imp, int *inflight)
75 struct list_head *tmp, *pos;
76 struct ptlrpc_request *req = NULL;
82 /* It might have committed some after we last spoke, so make sure we
83 * get rid of them now.
85 spin_lock(&imp->imp_lock);
86 imp->imp_last_transno_checked = 0;
87 ptlrpc_free_committed(imp);
88 last_transno = imp->imp_last_replay_transno;
89 spin_unlock(&imp->imp_lock);
91 CDEBUG(D_HA, "import %p from %s committed "LPU64" last "LPU64"\n",
92 imp, obd2cli_tgt(imp->imp_obd),
93 imp->imp_peer_committed_transno, last_transno);
95 /* Do I need to hold a lock across this iteration? We shouldn't be
96 * racing with any additions to the list, because we're in recovery
97 * and are therefore not processing additional requests to add. Calls
98 * to ptlrpc_free_committed might commit requests, but nothing "newer"
99 * than the one we're replaying (it can't be committed until it's
100 * replayed, and we're doing that here). l_f_e_safe protects against
101 * problems with the current request being committed, in the unlikely
102 * event of that race. So, in conclusion, I think that it's safe to
103 * perform this list-walk without the imp_lock held.
105 * But, the {mdc,osc}_replay_open callbacks both iterate
106 * request lists, and have comments saying they assume the
107 * imp_lock is being held by ptlrpc_replay, but it's not. it's
108 * just a little race...
111 /* Replay all the committed open requests on committed_list first */
112 if (!list_empty(&imp->imp_committed_list)) {
113 tmp = imp->imp_committed_list.prev;
114 req = list_entry(tmp, struct ptlrpc_request,
117 /* The last request on committed_list hasn't been replayed */
118 if (req->rq_transno > last_transno) {
119 /* Since the imp_committed_list is immutable before
120 * all of it's requests being replayed, it's safe to
121 * use a cursor to accelerate the search */
122 imp->imp_replay_cursor = imp->imp_replay_cursor->next;
124 while (imp->imp_replay_cursor !=
125 &imp->imp_committed_list) {
126 req = list_entry(imp->imp_replay_cursor,
127 struct ptlrpc_request,
129 if (req->rq_transno > last_transno)
133 imp->imp_replay_cursor =
134 imp->imp_replay_cursor->next;
137 /* All requests on committed_list have been replayed */
138 imp->imp_replay_cursor = &imp->imp_committed_list;
143 /* All the requests in committed list have been replayed, let's replay
144 * the imp_replay_list */
146 list_for_each_safe(tmp, pos, &imp->imp_replay_list) {
147 req = list_entry(tmp, struct ptlrpc_request,
150 if (req->rq_transno > last_transno)
156 /* If need to resend the last sent transno (because a reconnect
157 * has occurred), then stop on the matching req and send it again.
158 * If, however, the last sent transno has been committed then we
159 * continue replay from the next request. */
160 if (req != NULL && imp->imp_resend_replay)
161 lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
163 spin_lock(&imp->imp_lock);
164 imp->imp_resend_replay = 0;
165 spin_unlock(&imp->imp_lock);
168 rc = ptlrpc_replay_req(req);
170 CERROR("recovery replay error %d for req "
171 LPU64"\n", rc, req->rq_xid);
180 * Schedule resending of request on sending_list. This is done after
181 * we completed replaying of requests and locks.
183 int ptlrpc_resend(struct obd_import *imp)
185 struct ptlrpc_request *req, *next;
189 /* As long as we're in recovery, nothing should be added to the sending
190 * list, so we don't need to hold the lock during this iteration and
193 /* Well... what if lctl recover is called twice at the same time?
195 spin_lock(&imp->imp_lock);
196 if (imp->imp_state != LUSTRE_IMP_RECOVER) {
197 spin_unlock(&imp->imp_lock);
201 list_for_each_entry_safe(req, next, &imp->imp_sending_list,
203 LASSERTF((long)req > PAGE_CACHE_SIZE && req != LP_POISON,
204 "req %p bad\n", req);
205 LASSERTF(req->rq_type != LI_POISON, "req %p freed\n", req);
206 if (!ptlrpc_no_resend(req))
207 ptlrpc_resend_req(req);
209 spin_unlock(&imp->imp_lock);
215 * Go through all requests in delayed list and wake their threads
218 void ptlrpc_wake_delayed(struct obd_import *imp)
220 struct list_head *tmp, *pos;
221 struct ptlrpc_request *req;
223 spin_lock(&imp->imp_lock);
224 list_for_each_safe(tmp, pos, &imp->imp_delayed_list) {
225 req = list_entry(tmp, struct ptlrpc_request, rq_list);
227 DEBUG_REQ(D_HA, req, "waking (set %p):", req->rq_set);
228 ptlrpc_client_wake_req(req);
230 spin_unlock(&imp->imp_lock);
233 void ptlrpc_request_handle_notconn(struct ptlrpc_request *failed_req)
235 struct obd_import *imp = failed_req->rq_import;
238 CDEBUG(D_HA, "import %s of %s@%s abruptly disconnected: reconnecting\n",
239 imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
240 imp->imp_connection->c_remote_uuid.uuid);
242 if (ptlrpc_set_import_discon(imp,
243 lustre_msg_get_conn_cnt(failed_req->rq_reqmsg))) {
244 if (!imp->imp_replayable) {
245 CDEBUG(D_HA, "import %s@%s for %s not replayable, "
246 "auto-deactivating\n",
247 obd2cli_tgt(imp->imp_obd),
248 imp->imp_connection->c_remote_uuid.uuid,
249 imp->imp_obd->obd_name);
250 ptlrpc_deactivate_import(imp);
252 /* to control recovery via lctl {disable|enable}_recovery */
253 if (imp->imp_deactive == 0)
254 ptlrpc_connect_import(imp);
257 /* Wait for recovery to complete and resend. If evicted, then
258 this request will be errored out later.*/
259 spin_lock(&failed_req->rq_lock);
260 if (!failed_req->rq_no_resend)
261 failed_req->rq_resend = 1;
262 spin_unlock(&failed_req->rq_lock);
268 * Administratively active/deactive a client.
269 * This should only be called by the ioctl interface, currently
270 * - the lctl deactivate and activate commands
271 * - echo 0/1 >> /proc/osc/XXX/active
272 * - client umount -f (ll_umount_begin)
274 int ptlrpc_set_import_active(struct obd_import *imp, int active)
276 struct obd_device *obd = imp->imp_obd;
282 /* When deactivating, mark import invalid, and abort in-flight
285 LCONSOLE_WARN("setting import %s INACTIVE by administrator "
286 "request\n", obd2cli_tgt(imp->imp_obd));
288 /* set before invalidate to avoid messages about imp_inval
289 * set without imp_deactive in ptlrpc_import_delay_req */
290 spin_lock(&imp->imp_lock);
291 imp->imp_deactive = 1;
292 spin_unlock(&imp->imp_lock);
294 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DEACTIVATE);
296 ptlrpc_invalidate_import(imp);
299 /* When activating, mark import valid, and attempt recovery */
301 CDEBUG(D_HA, "setting import %s VALID\n",
302 obd2cli_tgt(imp->imp_obd));
304 spin_lock(&imp->imp_lock);
305 imp->imp_deactive = 0;
306 spin_unlock(&imp->imp_lock);
307 obd_import_event(imp->imp_obd, imp, IMP_EVENT_ACTIVATE);
309 rc = ptlrpc_recover_import(imp, NULL, 0);
314 EXPORT_SYMBOL(ptlrpc_set_import_active);
316 /* Attempt to reconnect an import */
317 int ptlrpc_recover_import(struct obd_import *imp, char *new_uuid, int async)
322 spin_lock(&imp->imp_lock);
323 if (imp->imp_state == LUSTRE_IMP_NEW || imp->imp_deactive ||
324 atomic_read(&imp->imp_inval_count))
326 spin_unlock(&imp->imp_lock);
330 /* force import to be disconnected. */
331 ptlrpc_set_import_discon(imp, 0);
334 struct obd_uuid uuid;
336 /* intruct import to use new uuid */
337 obd_str2uuid(&uuid, new_uuid);
338 rc = import_set_conn_priority(imp, &uuid);
343 /* Check if reconnect is already in progress */
344 spin_lock(&imp->imp_lock);
345 if (imp->imp_state != LUSTRE_IMP_DISCON) {
346 imp->imp_force_verify = 1;
349 spin_unlock(&imp->imp_lock);
353 rc = ptlrpc_connect_import(imp);
358 struct l_wait_info lwi;
359 int secs = cfs_time_seconds(obd_timeout);
361 CDEBUG(D_HA, "%s: recovery started, waiting %u seconds\n",
362 obd2cli_tgt(imp->imp_obd), secs);
364 lwi = LWI_TIMEOUT(secs, NULL, NULL);
365 rc = l_wait_event(imp->imp_recovery_waitq,
366 !ptlrpc_import_in_recovery(imp), &lwi);
367 CDEBUG(D_HA, "%s: recovery finished\n",
368 obd2cli_tgt(imp->imp_obd));
375 EXPORT_SYMBOL(ptlrpc_recover_import);
377 int ptlrpc_import_in_recovery(struct obd_import *imp)
381 spin_lock(&imp->imp_lock);
382 if (imp->imp_state == LUSTRE_IMP_FULL ||
383 imp->imp_state == LUSTRE_IMP_CLOSED ||
384 imp->imp_state == LUSTRE_IMP_DISCON ||
385 imp->imp_obd->obd_no_recov)
387 spin_unlock(&imp->imp_lock);