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 CERROR("%s: connection lost to %s@%s\n",
104 imp->imp_obd->obd_name,
105 imp->imp_target_uuid.uuid,
106 imp->imp_connection->c_remote_uuid.uuid);
107 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
108 spin_unlock_irqrestore(&imp->imp_lock, flags);
109 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
112 spin_unlock_irqrestore(&imp->imp_lock, flags);
113 CDEBUG(D_HA, "%p %s: import already not connected: %s\n",
114 imp,imp->imp_client->cli_name,
115 ptlrpc_import_state_name(imp->imp_state));
122 * This acts as a barrier; all existing requests are rejected, and
123 * no new requests will be accepted until the import is valid again.
125 void ptlrpc_deactivate_import(struct obd_import *imp)
130 spin_lock_irqsave(&imp->imp_lock, flags);
131 CDEBUG(D_HA, "setting import %s INVALID\n",
132 imp->imp_target_uuid.uuid);
133 imp->imp_invalid = 1;
134 imp->imp_generation++;
135 spin_unlock_irqrestore(&imp->imp_lock, flags);
137 ptlrpc_abort_inflight(imp);
138 obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
142 * This function will invalidate the import, if necessary, then block
143 * for all the RPC completions, and finally notify the obd to
144 * invalidate its state (ie cancel locks, clear pending requests,
147 * in_rpc: true if this is called while processing an rpc, like
148 * CONNECT. It will allow for one RPC to be inflight while
149 * waiting for requests to complete. Ugly, yes, but I don't see an
150 * cleaner way right now.
152 void ptlrpc_invalidate_import(struct obd_import *imp, int in_rpc)
154 struct l_wait_info lwi;
155 unsigned long timeout;
159 if (!imp->imp_invalid)
160 ptlrpc_deactivate_import(imp);
162 LASSERT(imp->imp_invalid);
167 /* wait for all requests to error out and call completion
169 if (imp->imp_server_timeout)
170 timeout = obd_timeout / 2;
172 timeout = obd_timeout;
173 timeout = MAX(timeout * HZ, 1);
174 lwi = LWI_TIMEOUT_INTR(timeout, NULL, NULL, NULL);
175 rc = l_wait_event(imp->imp_recovery_waitq,
176 (atomic_read(&imp->imp_inflight) == inflight),
180 CERROR("%s: rc = %d waiting for callback (%d != %d)\n",
181 imp->imp_target_uuid.uuid, rc,
182 atomic_read(&imp->imp_inflight), inflight);
184 obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
187 void ptlrpc_activate_import(struct obd_import *imp)
189 struct obd_device *obd = imp->imp_obd;
192 spin_lock_irqsave(&imp->imp_lock, flags);
193 imp->imp_invalid = 0;
194 spin_unlock_irqrestore(&imp->imp_lock, flags);
196 obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
199 void ptlrpc_fail_import(struct obd_import *imp, int generation)
203 LASSERT (!imp->imp_dlm_fake);
205 if (ptlrpc_set_import_discon(imp)) {
208 if (!imp->imp_replayable) {
209 CDEBUG(D_HA, "import %s@%s for %s not replayable, "
210 "auto-deactivating\n",
211 imp->imp_target_uuid.uuid,
212 imp->imp_connection->c_remote_uuid.uuid,
213 imp->imp_obd->obd_name);
214 ptlrpc_deactivate_import(imp);
217 CDEBUG(D_HA, "%s: waking up pinger\n",
218 imp->imp_target_uuid.uuid);
220 spin_lock_irqsave(&imp->imp_lock, flags);
221 imp->imp_force_verify = 1;
222 spin_unlock_irqrestore(&imp->imp_lock, flags);
224 ptlrpc_pinger_wake_up();
229 int ptlrpc_connect_import(struct obd_import *imp, char * new_uuid)
231 struct obd_device *obd = imp->imp_obd;
232 int initial_connect = 0;
234 __u64 committed_before_reconnect = 0;
235 struct ptlrpc_request *request;
236 int size[] = {sizeof(imp->imp_target_uuid),
237 sizeof(obd->obd_uuid),
238 sizeof(imp->imp_dlm_handle),
239 sizeof(unsigned long)};
240 char *tmp[] = {imp->imp_target_uuid.uuid,
242 (char *)&imp->imp_dlm_handle,
243 (char *)&imp->imp_connect_flags}; /* XXX: make this portable! */
244 struct ptlrpc_connect_async_args *aa;
247 spin_lock_irqsave(&imp->imp_lock, flags);
248 if (imp->imp_state == LUSTRE_IMP_CLOSED) {
249 spin_unlock_irqrestore(&imp->imp_lock, flags);
250 CERROR("can't connect to a closed import\n");
252 } else if (imp->imp_state == LUSTRE_IMP_FULL) {
253 spin_unlock_irqrestore(&imp->imp_lock, flags);
254 CERROR("already connected\n");
256 } else if (imp->imp_state == LUSTRE_IMP_CONNECTING) {
257 spin_unlock_irqrestore(&imp->imp_lock, flags);
258 CERROR("already connecting\n");
262 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
264 imp->imp_resend_replay = 0;
266 if (imp->imp_remote_handle.cookie == 0) {
269 committed_before_reconnect = imp->imp_peer_committed_transno;;
274 spin_unlock_irqrestore(&imp->imp_lock, flags);
277 struct ptlrpc_connection *conn;
278 struct obd_uuid uuid;
279 struct obd_export *dlmexp;
281 obd_str2uuid(&uuid, new_uuid);
283 conn = ptlrpc_uuid_to_connection(&uuid);
285 GOTO(out, rc = -ENOENT);
287 CDEBUG(D_HA, "switching import %s/%s from %s to %s\n",
288 imp->imp_target_uuid.uuid, imp->imp_obd->obd_name,
289 imp->imp_connection->c_remote_uuid.uuid,
290 conn->c_remote_uuid.uuid);
292 /* Switch the import's connection and the DLM export's
293 * connection (which are almost certainly the same, but we
294 * keep distinct refs just to make things clearer. I think. */
295 if (imp->imp_connection)
296 ptlrpc_put_connection(imp->imp_connection);
297 /* We hand off the ref from ptlrpc_get_connection. */
298 imp->imp_connection = conn;
300 dlmexp = class_conn2export(&imp->imp_dlm_handle);
302 LASSERT(dlmexp != NULL);
304 if (dlmexp->exp_connection)
305 ptlrpc_put_connection(dlmexp->exp_connection);
306 dlmexp->exp_connection = ptlrpc_connection_addref(conn);
307 class_export_put(dlmexp);
311 request = ptlrpc_prep_req(imp, imp->imp_connect_op, 4, size, tmp);
313 GOTO(out, rc = -ENOMEM);
316 lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_LIBCLIENT);
319 request->rq_send_state = LUSTRE_IMP_CONNECTING;
320 request->rq_replen = lustre_msg_size(0, NULL);
321 request->rq_interpret_reply = ptlrpc_connect_interpret;
323 LASSERT (sizeof (*aa) <= sizeof (request->rq_async_args));
324 aa = (struct ptlrpc_connect_async_args *)&request->rq_async_args;
325 memset(aa, 0, sizeof *aa);
327 aa->pcaa_peer_committed = committed_before_reconnect;
328 aa->pcaa_initial_connect = initial_connect;
330 if (aa->pcaa_initial_connect) {
331 lustre_msg_add_op_flags(request->rq_reqmsg,
332 MSG_CONNECT_INITIAL);
333 imp->imp_replayable = 1;
336 ptlrpcd_add_req(request);
338 imp->imp_connect_start = jiffies;
341 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
347 static int ptlrpc_connect_interpret(struct ptlrpc_request *request,
350 struct ptlrpc_connect_async_args *aa = data;
351 struct obd_import *imp = request->rq_import;
352 struct lustre_handle old_hdl;
357 spin_lock_irqsave(&imp->imp_lock, flags);
358 if (imp->imp_state == LUSTRE_IMP_CLOSED) {
359 spin_unlock_irqrestore(&imp->imp_lock, flags);
362 spin_unlock_irqrestore(&imp->imp_lock, flags);
367 msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
369 if (aa->pcaa_initial_connect) {
370 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
371 CDEBUG(D_HA, "connected to replayable target: %s\n",
372 imp->imp_target_uuid.uuid);
373 imp->imp_pingable = imp->imp_replayable = 1;
375 imp->imp_replayable = 0;
377 LASSERTF(imp->imp_conn_cnt < request->rq_repmsg->conn_cnt,
378 "imp conn_cnt %d req conn_cnt %d",
379 imp->imp_conn_cnt, request->rq_repmsg->conn_cnt);
380 imp->imp_conn_cnt = request->rq_repmsg->conn_cnt;
381 imp->imp_remote_handle = request->rq_repmsg->handle;
382 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
383 GOTO(finish, rc = 0);
386 /* Determine what recovery state to move the import to. */
387 if (MSG_CONNECT_RECONNECT & msg_flags) {
388 memset(&old_hdl, 0, sizeof(old_hdl));
389 if (!memcmp(&old_hdl, &request->rq_repmsg->handle,
391 CERROR("%s@%s didn't like our handle "LPX64
392 ", failed\n", imp->imp_target_uuid.uuid,
393 imp->imp_connection->c_remote_uuid.uuid,
394 imp->imp_dlm_handle.cookie);
395 GOTO(out, rc = -ENOTCONN);
398 if (memcmp(&imp->imp_remote_handle, &request->rq_repmsg->handle,
399 sizeof(imp->imp_remote_handle))) {
400 CERROR("%s@%s changed handle from "LPX64" to "LPX64
401 "; copying, but this may foreshadow disaster\n",
402 imp->imp_target_uuid.uuid,
403 imp->imp_connection->c_remote_uuid.uuid,
404 imp->imp_remote_handle.cookie,
405 request->rq_repmsg->handle.cookie);
406 imp->imp_remote_handle = request->rq_repmsg->handle;
408 CERROR("reconnected to %s@%s after partition\n",
409 imp->imp_target_uuid.uuid,
410 imp->imp_connection->c_remote_uuid.uuid);
413 if (imp->imp_invalid) {
414 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
415 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
416 CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
417 imp->imp_obd->obd_name,
418 imp->imp_target_uuid.uuid);
419 imp->imp_resend_replay = 1;
420 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
422 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
424 } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
425 LASSERT(imp->imp_replayable);
426 imp->imp_remote_handle = request->rq_repmsg->handle;
427 imp->imp_last_replay_transno = 0;
428 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
430 CWARN("oops! we get evicted from %s\n", imp->imp_target_uuid.uuid);
431 imp->imp_remote_handle = request->rq_repmsg->handle;
432 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
435 /* Sanity checks for a reconnected import. */
436 if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
437 CERROR("imp_replayable flag does not match server "
438 "after reconnect. We should LBUG right here.\n");
441 if (request->rq_repmsg->last_committed < aa->pcaa_peer_committed) {
442 CERROR("%s went back in time (transno "LPD64
443 " was previously committed, server now claims "LPD64
444 ")! is shared storage not coherent?\n",
445 imp->imp_target_uuid.uuid,
446 aa->pcaa_peer_committed,
447 request->rq_repmsg->last_committed);
451 rc = ptlrpc_import_recovery_state_machine(imp);
453 if (rc == -ENOTCONN) {
454 CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
455 "invalidating and reconnecting\n",
456 imp->imp_target_uuid.uuid,
457 imp->imp_connection->c_remote_uuid.uuid);
458 ptlrpc_connect_import(imp, NULL);
464 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
465 if (aa->pcaa_initial_connect && !imp->imp_initial_recov) {
466 ptlrpc_deactivate_import(imp);
468 if (rc == -ETIMEDOUT && (jiffies - imp->imp_connect_start) > HZ) {
469 CDEBUG(D_ERROR, "recovery of %s on %s failed (timeout)\n",
470 imp->imp_target_uuid.uuid,
471 (char *)imp->imp_connection->c_remote_uuid.uuid);
472 ptlrpc_connect_import(imp, NULL);
475 CDEBUG(D_ERROR, "recovery of %s on %s failed (%d)\n",
476 imp->imp_target_uuid.uuid,
477 (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
480 wake_up(&imp->imp_recovery_waitq);
484 static int completed_replay_interpret(struct ptlrpc_request *req,
487 atomic_dec(&req->rq_import->imp_replay_inflight);
488 if (req->rq_status == 0) {
489 ptlrpc_import_recovery_state_machine(req->rq_import);
491 CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
493 req->rq_import->imp_obd->obd_name, req->rq_status);
494 ptlrpc_connect_import(req->rq_import, NULL);
500 static int signal_completed_replay(struct obd_import *imp)
502 struct ptlrpc_request *req;
505 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
506 atomic_inc(&imp->imp_replay_inflight);
508 req = ptlrpc_prep_req(imp, OBD_PING, 0, NULL, NULL);
512 req->rq_replen = lustre_msg_size(0, NULL);
513 req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
514 req->rq_reqmsg->flags |= MSG_LAST_REPLAY;
515 req->rq_timeout *= 3;
516 req->rq_interpret_reply = completed_replay_interpret;
518 ptlrpcd_add_req(req);
522 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
527 if (imp->imp_state == LUSTRE_IMP_EVICTED) {
528 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
529 imp->imp_target_uuid.uuid,
530 imp->imp_connection->c_remote_uuid.uuid);
532 ptlrpc_invalidate_import(imp, 1);
534 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
537 if (imp->imp_state == LUSTRE_IMP_REPLAY) {
538 CDEBUG(D_HA, "replay requested by %s\n",
539 imp->imp_target_uuid.uuid);
540 rc = ptlrpc_replay_next(imp, &inflight);
542 atomic_read(&imp->imp_replay_inflight) == 0) {
543 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
544 rc = ldlm_replay_locks(imp);
551 if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
552 if (atomic_read(&imp->imp_replay_inflight) == 0) {
553 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
554 rc = signal_completed_replay(imp);
561 if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
562 if (atomic_read(&imp->imp_replay_inflight) == 0) {
563 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
567 if (imp->imp_state == LUSTRE_IMP_RECOVER) {
568 CDEBUG(D_HA, "reconnected to %s@%s\n",
569 imp->imp_target_uuid.uuid,
570 imp->imp_connection->c_remote_uuid.uuid);
572 rc = ptlrpc_resend(imp);
575 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
576 ptlrpc_activate_import(imp);
577 CERROR("%s: connection restored to %s@%s\n",
578 imp->imp_obd->obd_name,
579 imp->imp_target_uuid.uuid,
580 imp->imp_connection->c_remote_uuid.uuid);
583 if (imp->imp_state == LUSTRE_IMP_FULL) {
584 wake_up(&imp->imp_recovery_waitq);
585 ptlrpc_wake_delayed(imp);
592 static int back_to_sleep(void *unused)
597 int ptlrpc_disconnect_import(struct obd_import *imp)
599 struct ptlrpc_request *request;
605 switch (imp->imp_connect_op) {
606 case OST_CONNECT: rq_opc = OST_DISCONNECT; break;
607 case MDS_CONNECT: rq_opc = MDS_DISCONNECT; break;
608 case MGMT_CONNECT:rq_opc = MGMT_DISCONNECT;break;
610 CERROR("don't know how to disconnect from %s (connect_op %d)\n",
611 imp->imp_target_uuid.uuid, imp->imp_connect_op);
616 if (ptlrpc_import_in_recovery(imp)) {
617 struct l_wait_info lwi;
618 unsigned long timeout;
619 if (imp->imp_server_timeout)
620 timeout = obd_timeout / 2;
622 timeout = obd_timeout;
623 timeout = MAX(timeout * HZ, 1);
624 lwi = LWI_TIMEOUT_INTR(obd_timeout, back_to_sleep, NULL, NULL);
625 rc = l_wait_event(imp->imp_recovery_waitq,
626 !ptlrpc_import_in_recovery(imp), &lwi);
630 spin_lock_irqsave(&imp->imp_lock, flags);
631 if (imp->imp_state != LUSTRE_IMP_FULL) {
634 spin_unlock_irqrestore(&imp->imp_lock, flags);
636 request = ptlrpc_prep_req(imp, rq_opc, 0, NULL, NULL);
638 /* For non-replayable connections, don't attempt
639 reconnect if this fails */
640 if (!imp->imp_replayable) {
641 request->rq_no_resend = 1;
642 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
643 request->rq_send_state = LUSTRE_IMP_CONNECTING;
645 request->rq_replen = lustre_msg_size(0, NULL);
646 rc = ptlrpc_queue_wait(request);
647 ptlrpc_req_finished(request);
650 spin_lock_irqsave(&imp->imp_lock, flags);
652 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
653 memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
654 imp->imp_conn_cnt = 0;
655 spin_unlock_irqrestore(&imp->imp_lock, flags);