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
163 lwi = LWI_TIMEOUT_INTR(MAX(obd_timeout * HZ, 1), NULL,
165 rc = l_wait_event(imp->imp_recovery_waitq,
166 (atomic_read(&imp->imp_inflight) == inflight),
170 CERROR("%s: rc = %d waiting for callback (%d != %d)\n",
171 imp->imp_target_uuid.uuid, rc,
172 atomic_read(&imp->imp_inflight), inflight);
174 obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
177 static void ptlrpc_activate_import(struct obd_import *imp)
179 struct obd_device *obd = imp->imp_obd;
182 spin_lock_irqsave(&imp->imp_lock, flags);
183 imp->imp_invalid = 0;
184 spin_unlock_irqrestore(&imp->imp_lock, flags);
186 obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
189 void ptlrpc_fail_import(struct obd_import *imp, int generation)
193 LASSERT (!imp->imp_dlm_fake);
195 if (ptlrpc_set_import_discon(imp)) {
198 if (!imp->imp_replayable) {
199 CDEBUG(D_HA, "import %s@%s for %s not replayable, "
200 "auto-deactivating\n",
201 imp->imp_target_uuid.uuid,
202 imp->imp_connection->c_remote_uuid.uuid,
203 imp->imp_obd->obd_name);
204 ptlrpc_deactivate_import(imp);
207 CDEBUG(D_HA, "%s: waking up pinger\n",
208 imp->imp_target_uuid.uuid);
210 spin_lock_irqsave(&imp->imp_lock, flags);
211 imp->imp_force_verify = 1;
212 spin_unlock_irqrestore(&imp->imp_lock, flags);
214 ptlrpc_pinger_wake_up();
220 int ptlrpc_connect_import(struct obd_import *imp, char * new_uuid)
222 struct obd_device *obd = imp->imp_obd;
223 int initial_connect = 0;
225 __u64 committed_before_reconnect = 0;
226 struct ptlrpc_request *request;
227 int size[] = {sizeof(imp->imp_target_uuid),
228 sizeof(obd->obd_uuid),
229 sizeof(imp->imp_dlm_handle)};
230 char *tmp[] = {imp->imp_target_uuid.uuid,
232 (char *)&imp->imp_dlm_handle};
233 struct ptlrpc_connect_async_args *aa;
236 spin_lock_irqsave(&imp->imp_lock, flags);
237 if (imp->imp_state == LUSTRE_IMP_CLOSED) {
238 spin_unlock_irqrestore(&imp->imp_lock, flags);
239 CERROR("can't connect to a closed import\n");
241 } else if (imp->imp_state == LUSTRE_IMP_FULL) {
242 spin_unlock_irqrestore(&imp->imp_lock, flags);
243 CERROR("already connected\n");
245 } else if (imp->imp_state == LUSTRE_IMP_CONNECTING) {
246 spin_unlock_irqrestore(&imp->imp_lock, flags);
247 CERROR("already connecting\n");
251 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
254 imp->imp_last_replay_transno = 0;
256 if (imp->imp_remote_handle.cookie == 0) {
259 committed_before_reconnect = imp->imp_peer_committed_transno;;
264 spin_unlock_irqrestore(&imp->imp_lock, flags);
267 struct ptlrpc_connection *conn;
268 struct obd_uuid uuid;
269 struct obd_export *dlmexp;
271 obd_str2uuid(&uuid, new_uuid);
273 conn = ptlrpc_uuid_to_connection(&uuid);
275 GOTO(out, rc = -ENOENT);
277 CDEBUG(D_HA, "switching import %s/%s from %s to %s\n",
278 imp->imp_target_uuid.uuid, imp->imp_obd->obd_name,
279 imp->imp_connection->c_remote_uuid.uuid,
280 conn->c_remote_uuid.uuid);
282 /* Switch the import's connection and the DLM export's
283 * connection (which are almost certainly the same, but we
284 * keep distinct refs just to make things clearer. I think. */
285 if (imp->imp_connection)
286 ptlrpc_put_connection(imp->imp_connection);
287 /* We hand off the ref from ptlrpc_get_connection. */
288 imp->imp_connection = conn;
290 dlmexp = class_conn2export(&imp->imp_dlm_handle);
292 LASSERT(dlmexp != NULL);
294 if (dlmexp->exp_connection)
295 ptlrpc_put_connection(dlmexp->exp_connection);
296 dlmexp->exp_connection = ptlrpc_connection_addref(conn);
297 class_export_put(dlmexp);
301 request = ptlrpc_prep_req(imp, imp->imp_connect_op, 3, size, tmp);
303 GOTO(out, rc = -ENOMEM);
306 lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_LIBCLIENT);
309 request->rq_send_state = LUSTRE_IMP_CONNECTING;
310 request->rq_replen = lustre_msg_size(0, NULL);
311 request->rq_interpret_reply = ptlrpc_connect_interpret;
313 LASSERT (sizeof (*aa) <= sizeof (request->rq_async_args));
314 aa = (struct ptlrpc_connect_async_args *)&request->rq_async_args;
315 memset(aa, 0, sizeof *aa);
317 aa->pcaa_peer_committed = committed_before_reconnect;
318 aa->pcaa_initial_connect = initial_connect;
320 if (aa->pcaa_initial_connect)
321 imp->imp_replayable = 1;
323 ptlrpcd_add_req(request);
327 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
333 static int ptlrpc_connect_interpret(struct ptlrpc_request *request,
336 struct ptlrpc_connect_async_args *aa = data;
337 struct obd_import *imp = request->rq_import;
338 struct lustre_handle old_hdl;
343 spin_lock_irqsave(&imp->imp_lock, flags);
344 if (imp->imp_state == LUSTRE_IMP_CLOSED) {
345 spin_unlock_irqrestore(&imp->imp_lock, flags);
348 spin_unlock_irqrestore(&imp->imp_lock, flags);
353 msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
355 if (aa->pcaa_initial_connect) {
356 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
357 CDEBUG(D_HA, "connected to replayable target: %s\n",
358 imp->imp_target_uuid.uuid);
359 imp->imp_pingable = imp->imp_replayable = 1;
361 imp->imp_replayable = 0;
363 imp->imp_remote_handle = request->rq_repmsg->handle;
364 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
365 GOTO(finish, rc = 0);
368 /* Determine what recovery state to move the import to. */
369 if (MSG_CONNECT_RECONNECT & msg_flags) {
370 memset(&old_hdl, 0, sizeof(old_hdl));
371 if (!memcmp(&old_hdl, &request->rq_repmsg->handle,
373 CERROR("%s@%s didn't like our handle "LPX64
374 ", failed\n", imp->imp_target_uuid.uuid,
375 imp->imp_connection->c_remote_uuid.uuid,
376 imp->imp_dlm_handle.cookie);
377 GOTO(out, rc = -ENOTCONN);
380 if (memcmp(&imp->imp_remote_handle, &request->rq_repmsg->handle,
381 sizeof(imp->imp_remote_handle))) {
382 CERROR("%s@%s changed handle from "LPX64" to "LPX64
383 "; copying, but this may foreshadow disaster\n",
384 imp->imp_target_uuid.uuid,
385 imp->imp_connection->c_remote_uuid.uuid,
386 imp->imp_remote_handle.cookie,
387 request->rq_repmsg->handle.cookie);
388 imp->imp_remote_handle = request->rq_repmsg->handle;
390 CERROR("reconnected to %s@%s after partition\n",
391 imp->imp_target_uuid.uuid,
392 imp->imp_connection->c_remote_uuid.uuid);
395 if (imp->imp_invalid)
396 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
398 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
400 else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
401 LASSERT(imp->imp_replayable);
402 imp->imp_remote_handle = request->rq_repmsg->handle;
403 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
406 imp->imp_remote_handle = request->rq_repmsg->handle;
407 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
410 /* Sanity checks for a reconnected import. */
411 if (!(imp->imp_replayable) !=
412 !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
413 CERROR("imp_replayable flag does not match server "
414 "after reconnect. We should LBUG right here.\n");
417 if (request->rq_repmsg->last_committed < aa->pcaa_peer_committed) {
418 CERROR("%s went back in time (transno "LPD64
419 " was previously committed, server now claims "LPD64
420 ")! is shared storage not coherent?\n",
421 imp->imp_target_uuid.uuid,
422 aa->pcaa_peer_committed,
423 request->rq_repmsg->last_committed);
427 rc = ptlrpc_import_recovery_state_machine(imp);
429 if (rc == -ENOTCONN) {
430 CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
431 "invalidating and reconnecting\n",
432 imp->imp_target_uuid.uuid,
433 imp->imp_connection->c_remote_uuid.uuid);
434 ptlrpc_connect_import(imp, NULL);
440 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
441 if (aa->pcaa_initial_connect && !imp->imp_initial_recov) {
442 ptlrpc_deactivate_import(imp);
444 CDEBUG(D_ERROR, "recovery of %s on %s failed (%d)\n",
445 imp->imp_target_uuid.uuid,
446 (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
449 wake_up(&imp->imp_recovery_waitq);
453 static int completed_replay_interpret(struct ptlrpc_request *req,
456 atomic_dec(&req->rq_import->imp_replay_inflight);
457 ptlrpc_import_recovery_state_machine(req->rq_import);
461 static int signal_completed_replay(struct obd_import *imp)
463 struct ptlrpc_request *req;
466 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
467 atomic_inc(&imp->imp_replay_inflight);
469 req = ptlrpc_prep_req(imp, OBD_PING, 0, NULL, NULL);
473 req->rq_replen = lustre_msg_size(0, NULL);
474 req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
475 req->rq_reqmsg->flags |= MSG_LAST_REPLAY;
476 req->rq_timeout *= 3;
477 req->rq_interpret_reply = completed_replay_interpret;
479 ptlrpcd_add_req(req);
483 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
488 if (imp->imp_state == LUSTRE_IMP_EVICTED) {
489 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
490 imp->imp_target_uuid.uuid,
491 imp->imp_connection->c_remote_uuid.uuid);
493 ptlrpc_invalidate_import(imp, 1);
495 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
498 if (imp->imp_state == LUSTRE_IMP_REPLAY) {
499 CDEBUG(D_HA, "replay requested by %s\n",
500 imp->imp_target_uuid.uuid);
501 rc = ptlrpc_replay_next(imp, &inflight);
503 atomic_read(&imp->imp_replay_inflight) == 0) {
504 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
505 rc = ldlm_replay_locks(imp);
512 if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
513 if (atomic_read(&imp->imp_replay_inflight) == 0) {
514 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
515 rc = signal_completed_replay(imp);
522 if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
523 if (atomic_read(&imp->imp_replay_inflight) == 0) {
524 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
528 if (imp->imp_state == LUSTRE_IMP_RECOVER) {
529 CDEBUG(D_HA, "reconnected to %s@%s\n",
530 imp->imp_target_uuid.uuid,
531 imp->imp_connection->c_remote_uuid.uuid);
533 rc = ptlrpc_resend(imp);
536 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
537 ptlrpc_activate_import(imp);
540 if (imp->imp_state == LUSTRE_IMP_FULL) {
541 wake_up(&imp->imp_recovery_waitq);
542 ptlrpc_wake_delayed(imp);
549 static int back_to_sleep(void *unused)
554 int ptlrpc_disconnect_import(struct obd_import *imp)
556 struct ptlrpc_request *request;
562 switch (imp->imp_connect_op) {
563 case OST_CONNECT: rq_opc = OST_DISCONNECT; break;
564 case MDS_CONNECT: rq_opc = MDS_DISCONNECT; break;
565 case MGMT_CONNECT:rq_opc = MGMT_DISCONNECT;break;
567 CERROR("don't know how to disconnect from %s (connect_op %d)\n",
568 imp->imp_target_uuid.uuid, imp->imp_connect_op);
573 if (ptlrpc_import_in_recovery(imp)) {
574 struct l_wait_info lwi;
575 lwi = LWI_TIMEOUT_INTR(MAX(obd_timeout * HZ, 1), back_to_sleep,
577 rc = l_wait_event(imp->imp_recovery_waitq,
578 !ptlrpc_import_in_recovery(imp), &lwi);
582 spin_lock_irqsave(&imp->imp_lock, flags);
583 if (imp->imp_state != LUSTRE_IMP_FULL) {
586 spin_unlock_irqrestore(&imp->imp_lock, flags);
588 request = ptlrpc_prep_req(imp, rq_opc, 0, NULL, NULL);
590 /* For non-replayable connections, don't attempt
591 reconnect if this fails */
592 if (!imp->imp_replayable) {
593 request->rq_no_resend = 1;
594 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
595 request->rq_send_state = LUSTRE_IMP_CONNECTING;
597 request->rq_replen = lustre_msg_size(0, NULL);
598 rc = ptlrpc_queue_wait(request);
599 ptlrpc_req_finished(request);
602 spin_lock_irqsave(&imp->imp_lock, flags);
604 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
605 memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
606 spin_unlock_irqrestore(&imp->imp_lock, flags);