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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2015, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/ptlrpc/recover.c
34 * Author: Mike Shaver <shaver@clusterfs.com>
37 #define DEBUG_SUBSYSTEM S_RPC
38 #include <linux/list.h>
39 #include <libcfs/libcfs.h>
40 #include <obd_support.h>
41 #include <lustre_ha.h>
42 #include <lustre_net.h>
43 #include <lustre_import.h>
44 #include <lustre_export.h>
46 #include <obd_class.h>
48 #include "ptlrpc_internal.h"
51 * Start recovery on disconnected import.
52 * This is done by just attempting a connect
54 void ptlrpc_initiate_recovery(struct obd_import *imp)
58 CDEBUG(D_HA, "%s: starting recovery\n", obd2cli_tgt(imp->imp_obd));
59 ptlrpc_connect_import(imp);
65 * Identify what request from replay list needs to be replayed next
66 * (based on what we have already replayed) and send it to server.
68 int ptlrpc_replay_next(struct obd_import *imp, int *inflight)
71 struct list_head *tmp, *pos;
72 struct ptlrpc_request *req = NULL;
78 /* It might have committed some after we last spoke, so make sure we
79 * get rid of them now.
81 spin_lock(&imp->imp_lock);
82 imp->imp_last_transno_checked = 0;
83 ptlrpc_free_committed(imp);
84 last_transno = imp->imp_last_replay_transno;
85 spin_unlock(&imp->imp_lock);
87 CDEBUG(D_HA, "import %p from %s committed %llu last %llu\n",
88 imp, obd2cli_tgt(imp->imp_obd),
89 imp->imp_peer_committed_transno, last_transno);
91 /* Do I need to hold a lock across this iteration? We shouldn't be
92 * racing with any additions to the list, because we're in recovery
93 * and are therefore not processing additional requests to add. Calls
94 * to ptlrpc_free_committed might commit requests, but nothing "newer"
95 * than the one we're replaying (it can't be committed until it's
96 * replayed, and we're doing that here). l_f_e_safe protects against
97 * problems with the current request being committed, in the unlikely
98 * event of that race. So, in conclusion, I think that it's safe to
99 * perform this list-walk without the imp_lock held.
101 * But, the {mdc,osc}_replay_open callbacks both iterate
102 * request lists, and have comments saying they assume the
103 * imp_lock is being held by ptlrpc_replay, but it's not. it's
104 * just a little race...
107 /* Replay all the committed open requests on committed_list first */
108 if (!list_empty(&imp->imp_committed_list)) {
109 tmp = imp->imp_committed_list.prev;
110 req = list_entry(tmp, struct ptlrpc_request,
113 /* The last request on committed_list hasn't been replayed */
114 if (req->rq_transno > last_transno) {
115 /* Since the imp_committed_list is immutable before
116 * all of it's requests being replayed, it's safe to
117 * use a cursor to accelerate the search */
118 if (!imp->imp_resend_replay ||
119 imp->imp_replay_cursor == &imp->imp_committed_list)
120 imp->imp_replay_cursor =
121 imp->imp_replay_cursor->next;
123 while (imp->imp_replay_cursor !=
124 &imp->imp_committed_list) {
125 req = list_entry(imp->imp_replay_cursor,
126 struct ptlrpc_request,
128 if (req->rq_transno > last_transno)
132 imp->imp_replay_cursor =
133 imp->imp_replay_cursor->next;
136 /* All requests on committed_list have been replayed */
137 imp->imp_replay_cursor = &imp->imp_committed_list;
142 /* All the requests in committed list have been replayed, let's replay
143 * the imp_replay_list */
145 list_for_each_safe(tmp, pos, &imp->imp_replay_list) {
146 req = list_entry(tmp, struct ptlrpc_request,
149 if (req->rq_transno > last_transno)
155 /* If need to resend the last sent transno (because a reconnect
156 * has occurred), then stop on the matching req and send it again.
157 * If, however, the last sent transno has been committed then we
158 * continue replay from the next request. */
159 if (req != NULL && imp->imp_resend_replay)
160 lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
162 spin_lock(&imp->imp_lock);
163 /* The resend replay request may have been removed from the
165 if (req != NULL && imp->imp_resend_replay &&
166 list_empty(&req->rq_unreplied_list)) {
167 ptlrpc_add_unreplied(req);
168 imp->imp_known_replied_xid = ptlrpc_known_replied_xid(imp);
171 imp->imp_resend_replay = 0;
172 spin_unlock(&imp->imp_lock);
175 /* The request should have been added back in unreplied list
176 * by ptlrpc_prepare_replay(). */
177 LASSERT(!list_empty(&req->rq_unreplied_list));
179 rc = ptlrpc_replay_req(req);
181 CERROR("recovery replay error %d for req "
182 "%llu\n", rc, req->rq_xid);
191 * Schedule resending of request on sending_list. This is done after
192 * we completed replaying of requests and locks.
194 int ptlrpc_resend(struct obd_import *imp)
196 struct ptlrpc_request *req, *next;
200 /* As long as we're in recovery, nothing should be added to the sending
201 * list, so we don't need to hold the lock during this iteration and
204 /* Well... what if lctl recover is called twice at the same time?
206 spin_lock(&imp->imp_lock);
207 if (imp->imp_state != LUSTRE_IMP_RECOVER) {
208 spin_unlock(&imp->imp_lock);
212 list_for_each_entry_safe(req, next, &imp->imp_sending_list, rq_list) {
213 LASSERTF((long)req > PAGE_SIZE && req != LP_POISON,
214 "req %p bad\n", req);
215 LASSERTF(req->rq_type != LI_POISON, "req %p freed\n", req);
217 /* If the request is allowed to be sent during replay and it
218 * is not timeout yet, then it does not need to be resent. */
219 if (!ptlrpc_no_resend(req) &&
220 (req->rq_timedout || !req->rq_allow_replay))
221 ptlrpc_resend_req(req);
223 spin_unlock(&imp->imp_lock);
229 * Go through all requests in delayed list and wake their threads
232 void ptlrpc_wake_delayed(struct obd_import *imp)
234 struct list_head *tmp, *pos;
235 struct ptlrpc_request *req;
237 spin_lock(&imp->imp_lock);
238 list_for_each_safe(tmp, pos, &imp->imp_delayed_list) {
239 req = list_entry(tmp, struct ptlrpc_request, rq_list);
241 DEBUG_REQ(D_HA, req, "waking (set %p):", req->rq_set);
242 ptlrpc_client_wake_req(req);
244 spin_unlock(&imp->imp_lock);
247 void ptlrpc_request_handle_notconn(struct ptlrpc_request *failed_req)
249 struct obd_import *imp = failed_req->rq_import;
252 CDEBUG(D_HA, "import %s of %s@%s abruptly disconnected: reconnecting\n",
253 imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
254 imp->imp_connection->c_remote_uuid.uuid);
256 if (ptlrpc_set_import_discon(imp,
257 lustre_msg_get_conn_cnt(failed_req->rq_reqmsg))) {
258 if (!imp->imp_replayable) {
259 CDEBUG(D_HA, "import %s@%s for %s not replayable, "
260 "auto-deactivating\n",
261 obd2cli_tgt(imp->imp_obd),
262 imp->imp_connection->c_remote_uuid.uuid,
263 imp->imp_obd->obd_name);
264 ptlrpc_deactivate_import(imp);
266 /* to control recovery via lctl {disable|enable}_recovery */
267 if (imp->imp_deactive == 0)
268 ptlrpc_connect_import(imp);
271 /* Wait for recovery to complete and resend. If evicted, then
272 this request will be errored out later.*/
273 spin_lock(&failed_req->rq_lock);
274 if (!failed_req->rq_no_resend)
275 failed_req->rq_resend = 1;
276 spin_unlock(&failed_req->rq_lock);
282 * Administratively active/deactive a client.
283 * This should only be called by the ioctl interface, currently
284 * - the lctl deactivate and activate commands
285 * - echo 0/1 >> /proc/osc/XXX/active
286 * - client umount -f (ll_umount_begin)
288 int ptlrpc_set_import_active(struct obd_import *imp, int active)
290 struct obd_device *obd = imp->imp_obd;
296 /* When deactivating, mark import invalid, and abort in-flight
299 LCONSOLE_WARN("setting import %s INACTIVE by administrator "
300 "request\n", obd2cli_tgt(imp->imp_obd));
302 /* set before invalidate to avoid messages about imp_inval
303 * set without imp_deactive in ptlrpc_import_delay_req */
304 spin_lock(&imp->imp_lock);
305 imp->imp_deactive = 1;
306 spin_unlock(&imp->imp_lock);
308 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DEACTIVATE);
310 ptlrpc_invalidate_import(imp);
313 /* When activating, mark import valid, and attempt recovery */
315 CDEBUG(D_HA, "setting import %s VALID\n",
316 obd2cli_tgt(imp->imp_obd));
318 spin_lock(&imp->imp_lock);
319 imp->imp_deactive = 0;
320 spin_unlock(&imp->imp_lock);
321 obd_import_event(imp->imp_obd, imp, IMP_EVENT_ACTIVATE);
323 rc = ptlrpc_recover_import(imp, NULL, 0);
328 EXPORT_SYMBOL(ptlrpc_set_import_active);
330 /* Attempt to reconnect an import */
331 int ptlrpc_recover_import(struct obd_import *imp, char *new_uuid, int async)
336 spin_lock(&imp->imp_lock);
337 if (imp->imp_state == LUSTRE_IMP_NEW || imp->imp_deactive ||
338 atomic_read(&imp->imp_inval_count))
340 spin_unlock(&imp->imp_lock);
344 /* force import to be disconnected. */
345 ptlrpc_set_import_discon(imp, 0);
348 struct obd_uuid uuid;
350 /* intruct import to use new uuid */
351 obd_str2uuid(&uuid, new_uuid);
352 rc = import_set_conn_priority(imp, &uuid);
357 /* Check if reconnect is already in progress */
358 spin_lock(&imp->imp_lock);
359 if (imp->imp_state != LUSTRE_IMP_DISCON) {
360 imp->imp_force_verify = 1;
363 spin_unlock(&imp->imp_lock);
367 rc = ptlrpc_connect_import(imp);
372 struct l_wait_info lwi;
373 int secs = cfs_time_seconds(obd_timeout);
375 CDEBUG(D_HA, "%s: recovery started, waiting %u seconds\n",
376 obd2cli_tgt(imp->imp_obd), secs);
378 lwi = LWI_TIMEOUT(secs, NULL, NULL);
379 rc = l_wait_event(imp->imp_recovery_waitq,
380 !ptlrpc_import_in_recovery(imp), &lwi);
381 CDEBUG(D_HA, "%s: recovery finished\n",
382 obd2cli_tgt(imp->imp_obd));
389 EXPORT_SYMBOL(ptlrpc_recover_import);
391 int ptlrpc_import_in_recovery(struct obd_import *imp)
395 spin_lock(&imp->imp_lock);
396 if (imp->imp_state == LUSTRE_IMP_FULL ||
397 imp->imp_state == LUSTRE_IMP_CLOSED ||
398 imp->imp_state == LUSTRE_IMP_DISCON ||
399 imp->imp_obd->obd_no_recov)
401 spin_unlock(&imp->imp_lock);