1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (c) 2002, 2003 Cluster File Systems, Inc.
5 * Author: Mike Shaver <shaver@clusterfs.com>
7 * This file is part of Lustre, http://www.lustre.org.
9 * Lustre is free software; you can redistribute it and/or
10 * modify it under the terms of version 2 of the GNU General Public
11 * License as published by the Free Software Foundation.
13 * Lustre is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 * GNU General Public License for more details.
18 * You should have received a copy of the GNU General Public License
19 * along with Lustre; if not, write to the Free Software
20 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
23 #define DEBUG_SUBSYSTEM S_RPC
25 # include <linux/config.h>
26 # include <linux/module.h>
27 # include <linux/kmod.h>
29 # include <liblustre.h>
32 #include <linux/obd_support.h>
33 #include <linux/lustre_ha.h>
34 #include <linux/lustre_net.h>
35 #include <linux/lustre_import.h>
36 #include <linux/lustre_export.h>
37 #include <linux/obd.h>
38 #include <linux/obd_class.h>
40 #include "ptlrpc_internal.h"
42 struct ptlrpc_connect_async_args {
43 __u64 pcaa_peer_committed;
44 int pcaa_initial_connect;
47 /* A CLOSED import should remain so. */
48 #define IMPORT_SET_STATE_NOLOCK(imp, state) \
50 if (imp->imp_state != LUSTRE_IMP_CLOSED) { \
51 CDEBUG(D_HA, "%p %s: changing import state from %s to %s\n", \
52 imp, imp->imp_target_uuid.uuid, \
53 ptlrpc_import_state_name(imp->imp_state), \
54 ptlrpc_import_state_name(state)); \
55 imp->imp_state = state; \
59 #define IMPORT_SET_STATE(imp, state) \
61 unsigned long flags; \
63 spin_lock_irqsave(&imp->imp_lock, flags); \
64 IMPORT_SET_STATE_NOLOCK(imp, state); \
65 spin_unlock_irqrestore(&imp->imp_lock, flags); \
69 static int ptlrpc_connect_interpret(struct ptlrpc_request *request,
71 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
73 /* Only this function is allowed to change the import state when it is
74 * CLOSED. I would rather refcount the import and free it after
75 * disconnection like we do with exports. To do that, the client_obd
76 * will need to save the peer info somewhere other than in the import,
78 int ptlrpc_init_import(struct obd_import *imp)
82 spin_lock_irqsave(&imp->imp_lock, flags);
84 imp->imp_generation++;
85 imp->imp_state = LUSTRE_IMP_NEW;
87 spin_unlock_irqrestore(&imp->imp_lock, flags);
92 /* Returns true if import was FULL, false if import was already not
95 int ptlrpc_set_import_discon(struct obd_import *imp)
100 spin_lock_irqsave(&imp->imp_lock, flags);
102 if (imp->imp_state == LUSTRE_IMP_FULL) {
103 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
104 spin_unlock_irqrestore(&imp->imp_lock, flags);
105 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
108 spin_unlock_irqrestore(&imp->imp_lock, flags);
109 CDEBUG(D_HA, "%p %s: import already not connected: %s\n",
110 imp,imp->imp_client->cli_name,
111 ptlrpc_import_state_name(imp->imp_state));
118 * This acts as a barrier; all existing requests are rejected, and
119 * no new requests will be accepted until the import is valid again.
121 void ptlrpc_deactivate_import(struct obd_import *imp)
126 spin_lock_irqsave(&imp->imp_lock, flags);
127 CDEBUG(D_HA, "setting import %s INVALID\n",
128 imp->imp_target_uuid.uuid);
129 imp->imp_invalid = 1;
130 imp->imp_generation++;
131 spin_unlock_irqrestore(&imp->imp_lock, flags);
133 ptlrpc_abort_inflight(imp);
134 obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
138 * This function will invalidate the import, if necessary, then block
139 * for all the RPC completions, and finally notify the obd to
140 * invalidate its state (ie cancel locks, clear pending requests,
143 * in_rpc: true if this is called while processing an rpc, like
144 * CONNECT. It will allow for one RPC to be inflight while
145 * waiting for requests to complete. Ugly, yes, but I don't see an
146 * cleaner way right now.
148 void ptlrpc_invalidate_import(struct obd_import *imp, int in_rpc)
150 struct l_wait_info lwi;
154 if (!imp->imp_invalid)
155 ptlrpc_deactivate_import(imp);
157 LASSERT(imp->imp_invalid);
161 /* wait for all requests to error out and call completion callbacks */
162 lwi = LWI_TIMEOUT_INTR(MAX(obd_timeout * HZ, 1), NULL,
164 rc = l_wait_event(imp->imp_recovery_waitq,
165 (atomic_read(&imp->imp_inflight) == inflight),
169 CERROR("%s: rc = %d waiting for callback (%d != %d)\n",
170 imp->imp_target_uuid.uuid, rc,
171 atomic_read(&imp->imp_inflight), inflight);
173 obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
176 static void ptlrpc_activate_import(struct obd_import *imp)
178 struct obd_device *obd = imp->imp_obd;
181 spin_lock_irqsave(&imp->imp_lock, flags);
182 imp->imp_invalid = 0;
183 spin_unlock_irqrestore(&imp->imp_lock, flags);
185 obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
188 void ptlrpc_fail_import(struct obd_import *imp, int generation)
192 LASSERT (!imp->imp_dlm_fake);
194 if (ptlrpc_set_import_discon(imp)) {
197 if (!imp->imp_replayable) {
198 CDEBUG(D_HA, "import %s@%s for %s not replayable, "
199 "auto-deactivating\n",
200 imp->imp_target_uuid.uuid,
201 imp->imp_connection->c_remote_uuid.uuid,
202 imp->imp_obd->obd_name);
203 ptlrpc_deactivate_import(imp);
206 CDEBUG(D_HA, "%s: waking up pinger\n",
207 imp->imp_target_uuid.uuid);
209 spin_lock_irqsave(&imp->imp_lock, flags);
210 imp->imp_force_verify = 1;
211 spin_unlock_irqrestore(&imp->imp_lock, flags);
213 ptlrpc_pinger_wake_up();
218 int ptlrpc_connect_import(struct obd_import *imp, char * new_uuid)
220 struct obd_device *obd = imp->imp_obd;
221 int initial_connect = 0;
223 __u64 committed_before_reconnect = 0;
224 struct ptlrpc_request *request;
225 int size[] = {sizeof(imp->imp_target_uuid),
226 sizeof(obd->obd_uuid),
227 sizeof(imp->imp_dlm_handle)};
228 char *tmp[] = {imp->imp_target_uuid.uuid,
230 (char *)&imp->imp_dlm_handle};
231 struct ptlrpc_connect_async_args *aa;
234 spin_lock_irqsave(&imp->imp_lock, flags);
235 if (imp->imp_state == LUSTRE_IMP_CLOSED) {
236 spin_unlock_irqrestore(&imp->imp_lock, flags);
237 CERROR("can't connect to a closed import\n");
239 } else if (imp->imp_state == LUSTRE_IMP_FULL) {
240 spin_unlock_irqrestore(&imp->imp_lock, flags);
241 CERROR("already connected\n");
243 } else if (imp->imp_state == LUSTRE_IMP_CONNECTING) {
244 spin_unlock_irqrestore(&imp->imp_lock, flags);
245 CERROR("already connecting\n");
249 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
252 imp->imp_last_replay_transno = 0;
254 if (imp->imp_remote_handle.cookie == 0) {
257 committed_before_reconnect = imp->imp_peer_committed_transno;;
262 spin_unlock_irqrestore(&imp->imp_lock, flags);
265 struct ptlrpc_connection *conn;
266 struct obd_uuid uuid;
267 struct obd_export *dlmexp;
269 obd_str2uuid(&uuid, new_uuid);
271 conn = ptlrpc_uuid_to_connection(&uuid);
273 GOTO(out, rc = -ENOENT);
275 CDEBUG(D_HA, "switching import %s/%s from %s to %s\n",
276 imp->imp_target_uuid.uuid, imp->imp_obd->obd_name,
277 imp->imp_connection->c_remote_uuid.uuid,
278 conn->c_remote_uuid.uuid);
280 /* Switch the import's connection and the DLM export's
281 * connection (which are almost certainly the same, but we
282 * keep distinct refs just to make things clearer. I think. */
283 if (imp->imp_connection)
284 ptlrpc_put_connection(imp->imp_connection);
285 /* We hand off the ref from ptlrpc_get_connection. */
286 imp->imp_connection = conn;
288 dlmexp = class_conn2export(&imp->imp_dlm_handle);
290 LASSERT(dlmexp != NULL);
292 if (dlmexp->exp_connection)
293 ptlrpc_put_connection(dlmexp->exp_connection);
294 dlmexp->exp_connection = ptlrpc_connection_addref(conn);
295 class_export_put(dlmexp);
299 request = ptlrpc_prep_req(imp, imp->imp_connect_op, 3, size, tmp);
301 GOTO(out, rc = -ENOMEM);
304 lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_LIBCLIENT);
307 request->rq_send_state = LUSTRE_IMP_CONNECTING;
308 request->rq_replen = lustre_msg_size(0, NULL);
309 request->rq_interpret_reply = ptlrpc_connect_interpret;
311 LASSERT (sizeof (*aa) <= sizeof (request->rq_async_args));
312 aa = (struct ptlrpc_connect_async_args *)&request->rq_async_args;
313 memset(aa, 0, sizeof *aa);
315 aa->pcaa_peer_committed = committed_before_reconnect;
316 aa->pcaa_initial_connect = initial_connect;
318 if (aa->pcaa_initial_connect)
319 imp->imp_replayable = 1;
321 ptlrpcd_add_req(request);
325 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
331 static int ptlrpc_connect_interpret(struct ptlrpc_request *request,
334 struct ptlrpc_connect_async_args *aa = data;
335 struct obd_import *imp = request->rq_import;
336 struct lustre_handle old_hdl;
341 spin_lock_irqsave(&imp->imp_lock, flags);
342 if (imp->imp_state == LUSTRE_IMP_CLOSED) {
343 spin_unlock_irqrestore(&imp->imp_lock, flags);
346 spin_unlock_irqrestore(&imp->imp_lock, flags);
351 msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
353 if (aa->pcaa_initial_connect) {
354 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
355 CDEBUG(D_HA, "connected to replayable target: %s\n",
356 imp->imp_target_uuid.uuid);
357 imp->imp_pingable = imp->imp_replayable = 1;
359 imp->imp_replayable = 0;
361 imp->imp_remote_handle = request->rq_repmsg->handle;
362 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
363 GOTO(finish, rc = 0);
366 /* Determine what recovery state to move the import to. */
367 if (MSG_CONNECT_RECONNECT & msg_flags) {
368 memset(&old_hdl, 0, sizeof(old_hdl));
369 if (!memcmp(&old_hdl, &request->rq_repmsg->handle,
371 CERROR("%s@%s didn't like our handle "LPX64
372 ", failed\n", imp->imp_target_uuid.uuid,
373 imp->imp_connection->c_remote_uuid.uuid,
374 imp->imp_dlm_handle.cookie);
375 GOTO(out, rc = -ENOTCONN);
378 if (memcmp(&imp->imp_remote_handle, &request->rq_repmsg->handle,
379 sizeof(imp->imp_remote_handle))) {
380 CERROR("%s@%s changed handle from "LPX64" to "LPX64
381 "; copying, but this may foreshadow disaster\n",
382 imp->imp_target_uuid.uuid,
383 imp->imp_connection->c_remote_uuid.uuid,
384 imp->imp_remote_handle.cookie,
385 request->rq_repmsg->handle.cookie);
386 imp->imp_remote_handle = request->rq_repmsg->handle;
388 CERROR("reconnected to %s@%s after partition\n",
389 imp->imp_target_uuid.uuid,
390 imp->imp_connection->c_remote_uuid.uuid);
393 if (imp->imp_invalid)
394 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
396 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
397 } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
398 LASSERT(imp->imp_replayable);
399 imp->imp_remote_handle = request->rq_repmsg->handle;
400 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
402 imp->imp_remote_handle = request->rq_repmsg->handle;
403 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
406 /* Sanity checks for a reconnected import. */
407 if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
408 CERROR("imp_replayable flag does not match server "
409 "after reconnect. We should LBUG right here.\n");
412 if (request->rq_repmsg->last_committed < aa->pcaa_peer_committed) {
413 CERROR("%s went back in time (transno "LPD64
414 " was previously committed, server now claims "LPD64
415 ")! is shared storage not coherent?\n",
416 imp->imp_target_uuid.uuid,
417 aa->pcaa_peer_committed,
418 request->rq_repmsg->last_committed);
422 rc = ptlrpc_import_recovery_state_machine(imp);
424 if (rc == -ENOTCONN) {
425 CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
426 "invalidating and reconnecting\n",
427 imp->imp_target_uuid.uuid,
428 imp->imp_connection->c_remote_uuid.uuid);
429 ptlrpc_connect_import(imp, NULL);
435 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
436 if (aa->pcaa_initial_connect && !imp->imp_initial_recov) {
437 ptlrpc_deactivate_import(imp);
439 CDEBUG(D_ERROR, "recovery of %s on %s failed (%d)\n",
440 imp->imp_target_uuid.uuid,
441 (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
444 wake_up(&imp->imp_recovery_waitq);
448 static int completed_replay_interpret(struct ptlrpc_request *req,
451 atomic_dec(&req->rq_import->imp_replay_inflight);
452 ptlrpc_import_recovery_state_machine(req->rq_import);
456 static int signal_completed_replay(struct obd_import *imp)
458 struct ptlrpc_request *req;
461 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
462 atomic_inc(&imp->imp_replay_inflight);
464 req = ptlrpc_prep_req(imp, OBD_PING, 0, NULL, NULL);
468 req->rq_replen = lustre_msg_size(0, NULL);
469 req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
470 req->rq_reqmsg->flags |= MSG_LAST_REPLAY;
471 req->rq_timeout *= 3;
472 req->rq_interpret_reply = completed_replay_interpret;
474 ptlrpcd_add_req(req);
478 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
483 if (imp->imp_state == LUSTRE_IMP_EVICTED) {
484 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
485 imp->imp_target_uuid.uuid,
486 imp->imp_connection->c_remote_uuid.uuid);
488 ptlrpc_invalidate_import(imp, 1);
490 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
493 if (imp->imp_state == LUSTRE_IMP_REPLAY) {
494 CDEBUG(D_HA, "replay requested by %s\n",
495 imp->imp_target_uuid.uuid);
496 rc = ptlrpc_replay_next(imp, &inflight);
498 atomic_read(&imp->imp_replay_inflight) == 0) {
499 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
500 rc = ldlm_replay_locks(imp);
507 if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
508 if (atomic_read(&imp->imp_replay_inflight) == 0) {
509 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
510 rc = signal_completed_replay(imp);
517 if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
518 if (atomic_read(&imp->imp_replay_inflight) == 0) {
519 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
523 if (imp->imp_state == LUSTRE_IMP_RECOVER) {
524 CDEBUG(D_HA, "reconnected to %s@%s\n",
525 imp->imp_target_uuid.uuid,
526 imp->imp_connection->c_remote_uuid.uuid);
528 rc = ptlrpc_resend(imp);
531 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
532 ptlrpc_activate_import(imp);
535 if (imp->imp_state == LUSTRE_IMP_FULL) {
536 wake_up(&imp->imp_recovery_waitq);
537 ptlrpc_wake_delayed(imp);
544 static int back_to_sleep(void *unused)
549 int ptlrpc_disconnect_import(struct obd_import *imp)
551 struct ptlrpc_request *request;
557 switch (imp->imp_connect_op) {
558 case OST_CONNECT: rq_opc = OST_DISCONNECT; break;
559 case MDS_CONNECT: rq_opc = MDS_DISCONNECT; break;
560 case MGMT_CONNECT:rq_opc = MGMT_DISCONNECT;break;
562 CERROR("don't know how to disconnect from %s (connect_op %d)\n",
563 imp->imp_target_uuid.uuid, imp->imp_connect_op);
568 if (ptlrpc_import_in_recovery(imp)) {
569 struct l_wait_info lwi;
570 lwi = LWI_TIMEOUT_INTR(MAX(obd_timeout * HZ, 1), back_to_sleep,
572 rc = l_wait_event(imp->imp_recovery_waitq,
573 !ptlrpc_import_in_recovery(imp), &lwi);
577 spin_lock_irqsave(&imp->imp_lock, flags);
578 if (imp->imp_state != LUSTRE_IMP_FULL) {
581 spin_unlock_irqrestore(&imp->imp_lock, flags);
583 request = ptlrpc_prep_req(imp, rq_opc, 0, NULL, NULL);
585 /* For non-replayable connections, don't attempt
586 reconnect if this fails */
587 if (!imp->imp_replayable) {
588 request->rq_no_resend = 1;
589 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
590 request->rq_send_state = LUSTRE_IMP_CONNECTING;
592 request->rq_replen = lustre_msg_size(0, NULL);
593 rc = ptlrpc_queue_wait(request);
594 ptlrpc_req_finished(request);
597 spin_lock_irqsave(&imp->imp_lock, flags);
599 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
600 memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
601 spin_unlock_irqrestore(&imp->imp_lock, flags);