xref: /linux/fs/ocfs2/dlm/dlmrecovery.c (revision e26207a3819684e9b4450a2d30bdd065fa92d9c7)
1 /* -*- mode: c; c-basic-offset: 8; -*-
2  * vim: noexpandtab sw=8 ts=8 sts=0:
3  *
4  * dlmrecovery.c
5  *
6  * recovery stuff
7  *
8  * Copyright (C) 2004 Oracle.  All rights reserved.
9  *
10  * This program is free software; you can redistribute it and/or
11  * modify it under the terms of the GNU General Public
12  * License as published by the Free Software Foundation; either
13  * version 2 of the License, or (at your option) any later version.
14  *
15  * This program is distributed in the hope that it will be useful,
16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
18  * General Public License for more details.
19  *
20  * You should have received a copy of the GNU General Public
21  * License along with this program; if not, write to the
22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23  * Boston, MA 021110-1307, USA.
24  *
25  */
26 
27 
28 #include <linux/module.h>
29 #include <linux/fs.h>
30 #include <linux/types.h>
31 #include <linux/slab.h>
32 #include <linux/highmem.h>
33 #include <linux/init.h>
34 #include <linux/sysctl.h>
35 #include <linux/random.h>
36 #include <linux/blkdev.h>
37 #include <linux/socket.h>
38 #include <linux/inet.h>
39 #include <linux/timer.h>
40 #include <linux/kthread.h>
41 #include <linux/delay.h>
42 
43 
44 #include "cluster/heartbeat.h"
45 #include "cluster/nodemanager.h"
46 #include "cluster/tcp.h"
47 
48 #include "dlmapi.h"
49 #include "dlmcommon.h"
50 #include "dlmdomain.h"
51 
52 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_RECOVERY)
53 #include "cluster/masklog.h"
54 
55 static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node);
56 
57 static int dlm_recovery_thread(void *data);
58 void dlm_complete_recovery_thread(struct dlm_ctxt *dlm);
59 int dlm_launch_recovery_thread(struct dlm_ctxt *dlm);
60 void dlm_kick_recovery_thread(struct dlm_ctxt *dlm);
61 static int dlm_do_recovery(struct dlm_ctxt *dlm);
62 
63 static int dlm_pick_recovery_master(struct dlm_ctxt *dlm);
64 static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node);
65 static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node);
66 static int dlm_request_all_locks(struct dlm_ctxt *dlm,
67 				 u8 request_from, u8 dead_node);
68 static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm, u8 dead_node);
69 
70 static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res);
71 static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres,
72 					const char *lockname, int namelen,
73 					int total_locks, u64 cookie,
74 					u8 flags, u8 master);
75 static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm,
76 				    struct dlm_migratable_lockres *mres,
77 				    u8 send_to,
78 				    struct dlm_lock_resource *res,
79 				    int total_locks);
80 static int dlm_process_recovery_data(struct dlm_ctxt *dlm,
81 				     struct dlm_lock_resource *res,
82 				     struct dlm_migratable_lockres *mres);
83 static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm);
84 static int dlm_send_all_done_msg(struct dlm_ctxt *dlm,
85 				 u8 dead_node, u8 send_to);
86 static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node);
87 static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm,
88 					struct list_head *list, u8 dead_node);
89 static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm,
90 					      u8 dead_node, u8 new_master);
91 static void dlm_reco_ast(void *astdata);
92 static void dlm_reco_bast(void *astdata, int blocked_type);
93 static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st);
94 static void dlm_request_all_locks_worker(struct dlm_work_item *item,
95 					 void *data);
96 static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data);
97 static int dlm_lockres_master_requery(struct dlm_ctxt *dlm,
98 				      struct dlm_lock_resource *res,
99 				      u8 *real_master);
100 
101 static u64 dlm_get_next_mig_cookie(void);
102 
103 static DEFINE_SPINLOCK(dlm_reco_state_lock);
104 static DEFINE_SPINLOCK(dlm_mig_cookie_lock);
105 static u64 dlm_mig_cookie = 1;
106 
107 static u64 dlm_get_next_mig_cookie(void)
108 {
109 	u64 c;
110 	spin_lock(&dlm_mig_cookie_lock);
111 	c = dlm_mig_cookie;
112 	if (dlm_mig_cookie == (~0ULL))
113 		dlm_mig_cookie = 1;
114 	else
115 		dlm_mig_cookie++;
116 	spin_unlock(&dlm_mig_cookie_lock);
117 	return c;
118 }
119 
120 static inline void dlm_set_reco_dead_node(struct dlm_ctxt *dlm,
121 					  u8 dead_node)
122 {
123 	assert_spin_locked(&dlm->spinlock);
124 	if (dlm->reco.dead_node != dead_node)
125 		mlog(0, "%s: changing dead_node from %u to %u\n",
126 		     dlm->name, dlm->reco.dead_node, dead_node);
127 	dlm->reco.dead_node = dead_node;
128 }
129 
130 static inline void dlm_set_reco_master(struct dlm_ctxt *dlm,
131 				       u8 master)
132 {
133 	assert_spin_locked(&dlm->spinlock);
134 	mlog(0, "%s: changing new_master from %u to %u\n",
135 	     dlm->name, dlm->reco.new_master, master);
136 	dlm->reco.new_master = master;
137 }
138 
139 static inline void __dlm_reset_recovery(struct dlm_ctxt *dlm)
140 {
141 	assert_spin_locked(&dlm->spinlock);
142 	clear_bit(dlm->reco.dead_node, dlm->recovery_map);
143 	dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
144 	dlm_set_reco_master(dlm, O2NM_INVALID_NODE_NUM);
145 }
146 
147 static inline void dlm_reset_recovery(struct dlm_ctxt *dlm)
148 {
149 	spin_lock(&dlm->spinlock);
150 	__dlm_reset_recovery(dlm);
151 	spin_unlock(&dlm->spinlock);
152 }
153 
154 /* Worker function used during recovery. */
155 void dlm_dispatch_work(struct work_struct *work)
156 {
157 	struct dlm_ctxt *dlm =
158 		container_of(work, struct dlm_ctxt, dispatched_work);
159 	LIST_HEAD(tmp_list);
160 	struct dlm_work_item *item, *next;
161 	dlm_workfunc_t *workfunc;
162 	int tot=0;
163 
164 	spin_lock(&dlm->work_lock);
165 	list_splice_init(&dlm->work_list, &tmp_list);
166 	spin_unlock(&dlm->work_lock);
167 
168 	list_for_each_entry(item, &tmp_list, list) {
169 		tot++;
170 	}
171 	mlog(0, "%s: work thread has %d work items\n", dlm->name, tot);
172 
173 	list_for_each_entry_safe(item, next, &tmp_list, list) {
174 		workfunc = item->func;
175 		list_del_init(&item->list);
176 
177 		/* already have ref on dlm to avoid having
178 		 * it disappear.  just double-check. */
179 		BUG_ON(item->dlm != dlm);
180 
181 		/* this is allowed to sleep and
182 		 * call network stuff */
183 		workfunc(item, item->data);
184 
185 		dlm_put(dlm);
186 		kfree(item);
187 	}
188 }
189 
190 /*
191  * RECOVERY THREAD
192  */
193 
194 void dlm_kick_recovery_thread(struct dlm_ctxt *dlm)
195 {
196 	/* wake the recovery thread
197 	 * this will wake the reco thread in one of three places
198 	 * 1) sleeping with no recovery happening
199 	 * 2) sleeping with recovery mastered elsewhere
200 	 * 3) recovery mastered here, waiting on reco data */
201 
202 	wake_up(&dlm->dlm_reco_thread_wq);
203 }
204 
205 /* Launch the recovery thread */
206 int dlm_launch_recovery_thread(struct dlm_ctxt *dlm)
207 {
208 	mlog(0, "starting dlm recovery thread...\n");
209 
210 	dlm->dlm_reco_thread_task = kthread_run(dlm_recovery_thread, dlm,
211 						"dlm_reco_thread");
212 	if (IS_ERR(dlm->dlm_reco_thread_task)) {
213 		mlog_errno(PTR_ERR(dlm->dlm_reco_thread_task));
214 		dlm->dlm_reco_thread_task = NULL;
215 		return -EINVAL;
216 	}
217 
218 	return 0;
219 }
220 
221 void dlm_complete_recovery_thread(struct dlm_ctxt *dlm)
222 {
223 	if (dlm->dlm_reco_thread_task) {
224 		mlog(0, "waiting for dlm recovery thread to exit\n");
225 		kthread_stop(dlm->dlm_reco_thread_task);
226 		dlm->dlm_reco_thread_task = NULL;
227 	}
228 }
229 
230 
231 
232 /*
233  * this is lame, but here's how recovery works...
234  * 1) all recovery threads cluster wide will work on recovering
235  *    ONE node at a time
236  * 2) negotiate who will take over all the locks for the dead node.
237  *    thats right... ALL the locks.
238  * 3) once a new master is chosen, everyone scans all locks
239  *    and moves aside those mastered by the dead guy
240  * 4) each of these locks should be locked until recovery is done
241  * 5) the new master collects up all of secondary lock queue info
242  *    one lock at a time, forcing each node to communicate back
243  *    before continuing
244  * 6) each secondary lock queue responds with the full known lock info
245  * 7) once the new master has run all its locks, it sends a ALLDONE!
246  *    message to everyone
247  * 8) upon receiving this message, the secondary queue node unlocks
248  *    and responds to the ALLDONE
249  * 9) once the new master gets responses from everyone, he unlocks
250  *    everything and recovery for this dead node is done
251  *10) go back to 2) while there are still dead nodes
252  *
253  */
254 
255 static void dlm_print_reco_node_status(struct dlm_ctxt *dlm)
256 {
257 	struct dlm_reco_node_data *ndata;
258 	struct dlm_lock_resource *res;
259 
260 	mlog(ML_NOTICE, "%s(%d): recovery info, state=%s, dead=%u, master=%u\n",
261 	     dlm->name, task_pid_nr(dlm->dlm_reco_thread_task),
262 	     dlm->reco.state & DLM_RECO_STATE_ACTIVE ? "ACTIVE" : "inactive",
263 	     dlm->reco.dead_node, dlm->reco.new_master);
264 
265 	list_for_each_entry(ndata, &dlm->reco.node_data, list) {
266 		char *st = "unknown";
267 		switch (ndata->state) {
268 			case DLM_RECO_NODE_DATA_INIT:
269 				st = "init";
270 				break;
271 			case DLM_RECO_NODE_DATA_REQUESTING:
272 				st = "requesting";
273 				break;
274 			case DLM_RECO_NODE_DATA_DEAD:
275 				st = "dead";
276 				break;
277 			case DLM_RECO_NODE_DATA_RECEIVING:
278 				st = "receiving";
279 				break;
280 			case DLM_RECO_NODE_DATA_REQUESTED:
281 				st = "requested";
282 				break;
283 			case DLM_RECO_NODE_DATA_DONE:
284 				st = "done";
285 				break;
286 			case DLM_RECO_NODE_DATA_FINALIZE_SENT:
287 				st = "finalize-sent";
288 				break;
289 			default:
290 				st = "bad";
291 				break;
292 		}
293 		mlog(ML_NOTICE, "%s: reco state, node %u, state=%s\n",
294 		     dlm->name, ndata->node_num, st);
295 	}
296 	list_for_each_entry(res, &dlm->reco.resources, recovering) {
297 		mlog(ML_NOTICE, "%s: lockres %.*s on recovering list\n",
298 		     dlm->name, res->lockname.len, res->lockname.name);
299 	}
300 }
301 
302 #define DLM_RECO_THREAD_TIMEOUT_MS (5 * 1000)
303 
304 static int dlm_recovery_thread(void *data)
305 {
306 	int status;
307 	struct dlm_ctxt *dlm = data;
308 	unsigned long timeout = msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS);
309 
310 	mlog(0, "dlm thread running for %s...\n", dlm->name);
311 
312 	while (!kthread_should_stop()) {
313 		if (dlm_joined(dlm)) {
314 			status = dlm_do_recovery(dlm);
315 			if (status == -EAGAIN) {
316 				/* do not sleep, recheck immediately. */
317 				continue;
318 			}
319 			if (status < 0)
320 				mlog_errno(status);
321 		}
322 
323 		wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq,
324 						 kthread_should_stop(),
325 						 timeout);
326 	}
327 
328 	mlog(0, "quitting DLM recovery thread\n");
329 	return 0;
330 }
331 
332 /* returns true when the recovery master has contacted us */
333 static int dlm_reco_master_ready(struct dlm_ctxt *dlm)
334 {
335 	int ready;
336 	spin_lock(&dlm->spinlock);
337 	ready = (dlm->reco.new_master != O2NM_INVALID_NODE_NUM);
338 	spin_unlock(&dlm->spinlock);
339 	return ready;
340 }
341 
342 /* returns true if node is no longer in the domain
343  * could be dead or just not joined */
344 int dlm_is_node_dead(struct dlm_ctxt *dlm, u8 node)
345 {
346 	int dead;
347 	spin_lock(&dlm->spinlock);
348 	dead = !test_bit(node, dlm->domain_map);
349 	spin_unlock(&dlm->spinlock);
350 	return dead;
351 }
352 
353 /* returns true if node is no longer in the domain
354  * could be dead or just not joined */
355 static int dlm_is_node_recovered(struct dlm_ctxt *dlm, u8 node)
356 {
357 	int recovered;
358 	spin_lock(&dlm->spinlock);
359 	recovered = !test_bit(node, dlm->recovery_map);
360 	spin_unlock(&dlm->spinlock);
361 	return recovered;
362 }
363 
364 
365 int dlm_wait_for_node_death(struct dlm_ctxt *dlm, u8 node, int timeout)
366 {
367 	if (timeout) {
368 		mlog(ML_NOTICE, "%s: waiting %dms for notification of "
369 		     "death of node %u\n", dlm->name, timeout, node);
370 		wait_event_timeout(dlm->dlm_reco_thread_wq,
371 			   dlm_is_node_dead(dlm, node),
372 			   msecs_to_jiffies(timeout));
373 	} else {
374 		mlog(ML_NOTICE, "%s: waiting indefinitely for notification "
375 		     "of death of node %u\n", dlm->name, node);
376 		wait_event(dlm->dlm_reco_thread_wq,
377 			   dlm_is_node_dead(dlm, node));
378 	}
379 	/* for now, return 0 */
380 	return 0;
381 }
382 
383 int dlm_wait_for_node_recovery(struct dlm_ctxt *dlm, u8 node, int timeout)
384 {
385 	if (timeout) {
386 		mlog(0, "%s: waiting %dms for notification of "
387 		     "recovery of node %u\n", dlm->name, timeout, node);
388 		wait_event_timeout(dlm->dlm_reco_thread_wq,
389 			   dlm_is_node_recovered(dlm, node),
390 			   msecs_to_jiffies(timeout));
391 	} else {
392 		mlog(0, "%s: waiting indefinitely for notification "
393 		     "of recovery of node %u\n", dlm->name, node);
394 		wait_event(dlm->dlm_reco_thread_wq,
395 			   dlm_is_node_recovered(dlm, node));
396 	}
397 	/* for now, return 0 */
398 	return 0;
399 }
400 
401 /* callers of the top-level api calls (dlmlock/dlmunlock) should
402  * block on the dlm->reco.event when recovery is in progress.
403  * the dlm recovery thread will set this state when it begins
404  * recovering a dead node (as the new master or not) and clear
405  * the state and wake as soon as all affected lock resources have
406  * been marked with the RECOVERY flag */
407 static int dlm_in_recovery(struct dlm_ctxt *dlm)
408 {
409 	int in_recovery;
410 	spin_lock(&dlm->spinlock);
411 	in_recovery = !!(dlm->reco.state & DLM_RECO_STATE_ACTIVE);
412 	spin_unlock(&dlm->spinlock);
413 	return in_recovery;
414 }
415 
416 
417 void dlm_wait_for_recovery(struct dlm_ctxt *dlm)
418 {
419 	if (dlm_in_recovery(dlm)) {
420 		mlog(0, "%s: reco thread %d in recovery: "
421 		     "state=%d, master=%u, dead=%u\n",
422 		     dlm->name, task_pid_nr(dlm->dlm_reco_thread_task),
423 		     dlm->reco.state, dlm->reco.new_master,
424 		     dlm->reco.dead_node);
425 	}
426 	wait_event(dlm->reco.event, !dlm_in_recovery(dlm));
427 }
428 
429 static void dlm_begin_recovery(struct dlm_ctxt *dlm)
430 {
431 	spin_lock(&dlm->spinlock);
432 	BUG_ON(dlm->reco.state & DLM_RECO_STATE_ACTIVE);
433 	dlm->reco.state |= DLM_RECO_STATE_ACTIVE;
434 	spin_unlock(&dlm->spinlock);
435 }
436 
437 static void dlm_end_recovery(struct dlm_ctxt *dlm)
438 {
439 	spin_lock(&dlm->spinlock);
440 	BUG_ON(!(dlm->reco.state & DLM_RECO_STATE_ACTIVE));
441 	dlm->reco.state &= ~DLM_RECO_STATE_ACTIVE;
442 	spin_unlock(&dlm->spinlock);
443 	wake_up(&dlm->reco.event);
444 }
445 
446 static int dlm_do_recovery(struct dlm_ctxt *dlm)
447 {
448 	int status = 0;
449 	int ret;
450 
451 	spin_lock(&dlm->spinlock);
452 
453 	/* check to see if the new master has died */
454 	if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM &&
455 	    test_bit(dlm->reco.new_master, dlm->recovery_map)) {
456 		mlog(0, "new master %u died while recovering %u!\n",
457 		     dlm->reco.new_master, dlm->reco.dead_node);
458 		/* unset the new_master, leave dead_node */
459 		dlm_set_reco_master(dlm, O2NM_INVALID_NODE_NUM);
460 	}
461 
462 	/* select a target to recover */
463 	if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
464 		int bit;
465 
466 		bit = find_next_bit (dlm->recovery_map, O2NM_MAX_NODES+1, 0);
467 		if (bit >= O2NM_MAX_NODES || bit < 0)
468 			dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
469 		else
470 			dlm_set_reco_dead_node(dlm, bit);
471 	} else if (!test_bit(dlm->reco.dead_node, dlm->recovery_map)) {
472 		/* BUG? */
473 		mlog(ML_ERROR, "dead_node %u no longer in recovery map!\n",
474 		     dlm->reco.dead_node);
475 		dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
476 	}
477 
478 	if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
479 		// mlog(0, "nothing to recover!  sleeping now!\n");
480 		spin_unlock(&dlm->spinlock);
481 		/* return to main thread loop and sleep. */
482 		return 0;
483 	}
484 	mlog(0, "%s(%d):recovery thread found node %u in the recovery map!\n",
485 	     dlm->name, task_pid_nr(dlm->dlm_reco_thread_task),
486 	     dlm->reco.dead_node);
487 	spin_unlock(&dlm->spinlock);
488 
489 	/* take write barrier */
490 	/* (stops the list reshuffling thread, proxy ast handling) */
491 	dlm_begin_recovery(dlm);
492 
493 	if (dlm->reco.new_master == dlm->node_num)
494 		goto master_here;
495 
496 	if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) {
497 		/* choose a new master, returns 0 if this node
498 		 * is the master, -EEXIST if it's another node.
499 		 * this does not return until a new master is chosen
500 		 * or recovery completes entirely. */
501 		ret = dlm_pick_recovery_master(dlm);
502 		if (!ret) {
503 			/* already notified everyone.  go. */
504 			goto master_here;
505 		}
506 		mlog(0, "another node will master this recovery session.\n");
507 	}
508 	mlog(0, "dlm=%s (%d), new_master=%u, this node=%u, dead_node=%u\n",
509 	     dlm->name, task_pid_nr(dlm->dlm_reco_thread_task), dlm->reco.new_master,
510 	     dlm->node_num, dlm->reco.dead_node);
511 
512 	/* it is safe to start everything back up here
513 	 * because all of the dead node's lock resources
514 	 * have been marked as in-recovery */
515 	dlm_end_recovery(dlm);
516 
517 	/* sleep out in main dlm_recovery_thread loop. */
518 	return 0;
519 
520 master_here:
521 	mlog(ML_NOTICE, "(%d) Node %u is the Recovery Master for the Dead Node "
522 	     "%u for Domain %s\n", task_pid_nr(dlm->dlm_reco_thread_task),
523 	     dlm->node_num, dlm->reco.dead_node, dlm->name);
524 
525 	status = dlm_remaster_locks(dlm, dlm->reco.dead_node);
526 	if (status < 0) {
527 		/* we should never hit this anymore */
528 		mlog(ML_ERROR, "error %d remastering locks for node %u, "
529 		     "retrying.\n", status, dlm->reco.dead_node);
530 		/* yield a bit to allow any final network messages
531 		 * to get handled on remaining nodes */
532 		msleep(100);
533 	} else {
534 		/* success!  see if any other nodes need recovery */
535 		mlog(0, "DONE mastering recovery of %s:%u here(this=%u)!\n",
536 		     dlm->name, dlm->reco.dead_node, dlm->node_num);
537 		dlm_reset_recovery(dlm);
538 	}
539 	dlm_end_recovery(dlm);
540 
541 	/* continue and look for another dead node */
542 	return -EAGAIN;
543 }
544 
545 static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node)
546 {
547 	int status = 0;
548 	struct dlm_reco_node_data *ndata;
549 	int all_nodes_done;
550 	int destroy = 0;
551 	int pass = 0;
552 
553 	do {
554 		/* we have become recovery master.  there is no escaping
555 		 * this, so just keep trying until we get it. */
556 		status = dlm_init_recovery_area(dlm, dead_node);
557 		if (status < 0) {
558 			mlog(ML_ERROR, "%s: failed to alloc recovery area, "
559 			     "retrying\n", dlm->name);
560 			msleep(1000);
561 		}
562 	} while (status != 0);
563 
564 	/* safe to access the node data list without a lock, since this
565 	 * process is the only one to change the list */
566 	list_for_each_entry(ndata, &dlm->reco.node_data, list) {
567 		BUG_ON(ndata->state != DLM_RECO_NODE_DATA_INIT);
568 		ndata->state = DLM_RECO_NODE_DATA_REQUESTING;
569 
570 		mlog(0, "requesting lock info from node %u\n",
571 		     ndata->node_num);
572 
573 		if (ndata->node_num == dlm->node_num) {
574 			ndata->state = DLM_RECO_NODE_DATA_DONE;
575 			continue;
576 		}
577 
578 		do {
579 			status = dlm_request_all_locks(dlm, ndata->node_num,
580 						       dead_node);
581 			if (status < 0) {
582 				mlog_errno(status);
583 				if (dlm_is_host_down(status)) {
584 					/* node died, ignore it for recovery */
585 					status = 0;
586 					ndata->state = DLM_RECO_NODE_DATA_DEAD;
587 					/* wait for the domain map to catch up
588 					 * with the network state. */
589 					wait_event_timeout(dlm->dlm_reco_thread_wq,
590 							   dlm_is_node_dead(dlm,
591 								ndata->node_num),
592 							   msecs_to_jiffies(1000));
593 					mlog(0, "waited 1 sec for %u, "
594 					     "dead? %s\n", ndata->node_num,
595 					     dlm_is_node_dead(dlm, ndata->node_num) ?
596 					     "yes" : "no");
597 				} else {
598 					/* -ENOMEM on the other node */
599 					mlog(0, "%s: node %u returned "
600 					     "%d during recovery, retrying "
601 					     "after a short wait\n",
602 					     dlm->name, ndata->node_num,
603 					     status);
604 					msleep(100);
605 				}
606 			}
607 		} while (status != 0);
608 
609 		spin_lock(&dlm_reco_state_lock);
610 		switch (ndata->state) {
611 			case DLM_RECO_NODE_DATA_INIT:
612 			case DLM_RECO_NODE_DATA_FINALIZE_SENT:
613 			case DLM_RECO_NODE_DATA_REQUESTED:
614 				BUG();
615 				break;
616 			case DLM_RECO_NODE_DATA_DEAD:
617 				mlog(0, "node %u died after requesting "
618 				     "recovery info for node %u\n",
619 				     ndata->node_num, dead_node);
620 				/* fine.  don't need this node's info.
621 				 * continue without it. */
622 				break;
623 			case DLM_RECO_NODE_DATA_REQUESTING:
624 				ndata->state = DLM_RECO_NODE_DATA_REQUESTED;
625 				mlog(0, "now receiving recovery data from "
626 				     "node %u for dead node %u\n",
627 				     ndata->node_num, dead_node);
628 				break;
629 			case DLM_RECO_NODE_DATA_RECEIVING:
630 				mlog(0, "already receiving recovery data from "
631 				     "node %u for dead node %u\n",
632 				     ndata->node_num, dead_node);
633 				break;
634 			case DLM_RECO_NODE_DATA_DONE:
635 				mlog(0, "already DONE receiving recovery data "
636 				     "from node %u for dead node %u\n",
637 				     ndata->node_num, dead_node);
638 				break;
639 		}
640 		spin_unlock(&dlm_reco_state_lock);
641 	}
642 
643 	mlog(0, "done requesting all lock info\n");
644 
645 	/* nodes should be sending reco data now
646 	 * just need to wait */
647 
648 	while (1) {
649 		/* check all the nodes now to see if we are
650 		 * done, or if anyone died */
651 		all_nodes_done = 1;
652 		spin_lock(&dlm_reco_state_lock);
653 		list_for_each_entry(ndata, &dlm->reco.node_data, list) {
654 			mlog(0, "checking recovery state of node %u\n",
655 			     ndata->node_num);
656 			switch (ndata->state) {
657 				case DLM_RECO_NODE_DATA_INIT:
658 				case DLM_RECO_NODE_DATA_REQUESTING:
659 					mlog(ML_ERROR, "bad ndata state for "
660 					     "node %u: state=%d\n",
661 					     ndata->node_num, ndata->state);
662 					BUG();
663 					break;
664 				case DLM_RECO_NODE_DATA_DEAD:
665 					mlog(0, "node %u died after "
666 					     "requesting recovery info for "
667 					     "node %u\n", ndata->node_num,
668 					     dead_node);
669 					break;
670 				case DLM_RECO_NODE_DATA_RECEIVING:
671 				case DLM_RECO_NODE_DATA_REQUESTED:
672 					mlog(0, "%s: node %u still in state %s\n",
673 					     dlm->name, ndata->node_num,
674 					     ndata->state==DLM_RECO_NODE_DATA_RECEIVING ?
675 					     "receiving" : "requested");
676 					all_nodes_done = 0;
677 					break;
678 				case DLM_RECO_NODE_DATA_DONE:
679 					mlog(0, "%s: node %u state is done\n",
680 					     dlm->name, ndata->node_num);
681 					break;
682 				case DLM_RECO_NODE_DATA_FINALIZE_SENT:
683 					mlog(0, "%s: node %u state is finalize\n",
684 					     dlm->name, ndata->node_num);
685 					break;
686 			}
687 		}
688 		spin_unlock(&dlm_reco_state_lock);
689 
690 		mlog(0, "pass #%d, all_nodes_done?: %s\n", ++pass,
691 		     all_nodes_done?"yes":"no");
692 		if (all_nodes_done) {
693 			int ret;
694 
695 			/* all nodes are now in DLM_RECO_NODE_DATA_DONE state
696 	 		 * just send a finalize message to everyone and
697 	 		 * clean up */
698 			mlog(0, "all nodes are done! send finalize\n");
699 			ret = dlm_send_finalize_reco_message(dlm);
700 			if (ret < 0)
701 				mlog_errno(ret);
702 
703 			spin_lock(&dlm->spinlock);
704 			dlm_finish_local_lockres_recovery(dlm, dead_node,
705 							  dlm->node_num);
706 			spin_unlock(&dlm->spinlock);
707 			mlog(0, "should be done with recovery!\n");
708 
709 			mlog(0, "finishing recovery of %s at %lu, "
710 			     "dead=%u, this=%u, new=%u\n", dlm->name,
711 			     jiffies, dlm->reco.dead_node,
712 			     dlm->node_num, dlm->reco.new_master);
713 			destroy = 1;
714 			status = 0;
715 			/* rescan everything marked dirty along the way */
716 			dlm_kick_thread(dlm, NULL);
717 			break;
718 		}
719 		/* wait to be signalled, with periodic timeout
720 		 * to check for node death */
721 		wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq,
722 					 kthread_should_stop(),
723 					 msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS));
724 
725 	}
726 
727 	if (destroy)
728 		dlm_destroy_recovery_area(dlm, dead_node);
729 
730 	mlog_exit(status);
731 	return status;
732 }
733 
734 static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node)
735 {
736 	int num=0;
737 	struct dlm_reco_node_data *ndata;
738 
739 	spin_lock(&dlm->spinlock);
740 	memcpy(dlm->reco.node_map, dlm->domain_map, sizeof(dlm->domain_map));
741 	/* nodes can only be removed (by dying) after dropping
742 	 * this lock, and death will be trapped later, so this should do */
743 	spin_unlock(&dlm->spinlock);
744 
745 	while (1) {
746 		num = find_next_bit (dlm->reco.node_map, O2NM_MAX_NODES, num);
747 		if (num >= O2NM_MAX_NODES) {
748 			break;
749 		}
750 		BUG_ON(num == dead_node);
751 
752 		ndata = kzalloc(sizeof(*ndata), GFP_NOFS);
753 		if (!ndata) {
754 			dlm_destroy_recovery_area(dlm, dead_node);
755 			return -ENOMEM;
756 		}
757 		ndata->node_num = num;
758 		ndata->state = DLM_RECO_NODE_DATA_INIT;
759 		spin_lock(&dlm_reco_state_lock);
760 		list_add_tail(&ndata->list, &dlm->reco.node_data);
761 		spin_unlock(&dlm_reco_state_lock);
762 		num++;
763 	}
764 
765 	return 0;
766 }
767 
768 static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm, u8 dead_node)
769 {
770 	struct dlm_reco_node_data *ndata, *next;
771 	LIST_HEAD(tmplist);
772 
773 	spin_lock(&dlm_reco_state_lock);
774 	list_splice_init(&dlm->reco.node_data, &tmplist);
775 	spin_unlock(&dlm_reco_state_lock);
776 
777 	list_for_each_entry_safe(ndata, next, &tmplist, list) {
778 		list_del_init(&ndata->list);
779 		kfree(ndata);
780 	}
781 }
782 
783 static int dlm_request_all_locks(struct dlm_ctxt *dlm, u8 request_from,
784 				 u8 dead_node)
785 {
786 	struct dlm_lock_request lr;
787 	enum dlm_status ret;
788 
789 	mlog(0, "\n");
790 
791 
792 	mlog(0, "dlm_request_all_locks: dead node is %u, sending request "
793 		  "to %u\n", dead_node, request_from);
794 
795 	memset(&lr, 0, sizeof(lr));
796 	lr.node_idx = dlm->node_num;
797 	lr.dead_node = dead_node;
798 
799 	// send message
800 	ret = DLM_NOLOCKMGR;
801 	ret = o2net_send_message(DLM_LOCK_REQUEST_MSG, dlm->key,
802 				 &lr, sizeof(lr), request_from, NULL);
803 
804 	/* negative status is handled by caller */
805 	if (ret < 0)
806 		mlog_errno(ret);
807 
808 	// return from here, then
809 	// sleep until all received or error
810 	return ret;
811 
812 }
813 
814 int dlm_request_all_locks_handler(struct o2net_msg *msg, u32 len, void *data,
815 				  void **ret_data)
816 {
817 	struct dlm_ctxt *dlm = data;
818 	struct dlm_lock_request *lr = (struct dlm_lock_request *)msg->buf;
819 	char *buf = NULL;
820 	struct dlm_work_item *item = NULL;
821 
822 	if (!dlm_grab(dlm))
823 		return -EINVAL;
824 
825 	if (lr->dead_node != dlm->reco.dead_node) {
826 		mlog(ML_ERROR, "%s: node %u sent dead_node=%u, but local "
827 		     "dead_node is %u\n", dlm->name, lr->node_idx,
828 		     lr->dead_node, dlm->reco.dead_node);
829 		dlm_print_reco_node_status(dlm);
830 		/* this is a hack */
831 		dlm_put(dlm);
832 		return -ENOMEM;
833 	}
834 	BUG_ON(lr->dead_node != dlm->reco.dead_node);
835 
836 	item = kzalloc(sizeof(*item), GFP_NOFS);
837 	if (!item) {
838 		dlm_put(dlm);
839 		return -ENOMEM;
840 	}
841 
842 	/* this will get freed by dlm_request_all_locks_worker */
843 	buf = (char *) __get_free_page(GFP_NOFS);
844 	if (!buf) {
845 		kfree(item);
846 		dlm_put(dlm);
847 		return -ENOMEM;
848 	}
849 
850 	/* queue up work for dlm_request_all_locks_worker */
851 	dlm_grab(dlm);  /* get an extra ref for the work item */
852 	dlm_init_work_item(dlm, item, dlm_request_all_locks_worker, buf);
853 	item->u.ral.reco_master = lr->node_idx;
854 	item->u.ral.dead_node = lr->dead_node;
855 	spin_lock(&dlm->work_lock);
856 	list_add_tail(&item->list, &dlm->work_list);
857 	spin_unlock(&dlm->work_lock);
858 	queue_work(dlm->dlm_worker, &dlm->dispatched_work);
859 
860 	dlm_put(dlm);
861 	return 0;
862 }
863 
864 static void dlm_request_all_locks_worker(struct dlm_work_item *item, void *data)
865 {
866 	struct dlm_migratable_lockres *mres;
867 	struct dlm_lock_resource *res;
868 	struct dlm_ctxt *dlm;
869 	LIST_HEAD(resources);
870 	int ret;
871 	u8 dead_node, reco_master;
872 	int skip_all_done = 0;
873 
874 	dlm = item->dlm;
875 	dead_node = item->u.ral.dead_node;
876 	reco_master = item->u.ral.reco_master;
877 	mres = (struct dlm_migratable_lockres *)data;
878 
879 	mlog(0, "%s: recovery worker started, dead=%u, master=%u\n",
880 	     dlm->name, dead_node, reco_master);
881 
882 	if (dead_node != dlm->reco.dead_node ||
883 	    reco_master != dlm->reco.new_master) {
884 		/* worker could have been created before the recovery master
885 		 * died.  if so, do not continue, but do not error. */
886 		if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) {
887 			mlog(ML_NOTICE, "%s: will not send recovery state, "
888 			     "recovery master %u died, thread=(dead=%u,mas=%u)"
889 			     " current=(dead=%u,mas=%u)\n", dlm->name,
890 			     reco_master, dead_node, reco_master,
891 			     dlm->reco.dead_node, dlm->reco.new_master);
892 		} else {
893 			mlog(ML_NOTICE, "%s: reco state invalid: reco(dead=%u, "
894 			     "master=%u), request(dead=%u, master=%u)\n",
895 			     dlm->name, dlm->reco.dead_node,
896 			     dlm->reco.new_master, dead_node, reco_master);
897 		}
898 		goto leave;
899 	}
900 
901 	/* lock resources should have already been moved to the
902  	 * dlm->reco.resources list.  now move items from that list
903  	 * to a temp list if the dead owner matches.  note that the
904 	 * whole cluster recovers only one node at a time, so we
905 	 * can safely move UNKNOWN lock resources for each recovery
906 	 * session. */
907 	dlm_move_reco_locks_to_list(dlm, &resources, dead_node);
908 
909 	/* now we can begin blasting lockreses without the dlm lock */
910 
911 	/* any errors returned will be due to the new_master dying,
912 	 * the dlm_reco_thread should detect this */
913 	list_for_each_entry(res, &resources, recovering) {
914 		ret = dlm_send_one_lockres(dlm, res, mres, reco_master,
915 				   	DLM_MRES_RECOVERY);
916 		if (ret < 0) {
917 			mlog(ML_ERROR, "%s: node %u went down while sending "
918 			     "recovery state for dead node %u, ret=%d\n", dlm->name,
919 			     reco_master, dead_node, ret);
920 			skip_all_done = 1;
921 			break;
922 		}
923 	}
924 
925 	/* move the resources back to the list */
926 	spin_lock(&dlm->spinlock);
927 	list_splice_init(&resources, &dlm->reco.resources);
928 	spin_unlock(&dlm->spinlock);
929 
930 	if (!skip_all_done) {
931 		ret = dlm_send_all_done_msg(dlm, dead_node, reco_master);
932 		if (ret < 0) {
933 			mlog(ML_ERROR, "%s: node %u went down while sending "
934 			     "recovery all-done for dead node %u, ret=%d\n",
935 			     dlm->name, reco_master, dead_node, ret);
936 		}
937 	}
938 leave:
939 	free_page((unsigned long)data);
940 }
941 
942 
943 static int dlm_send_all_done_msg(struct dlm_ctxt *dlm, u8 dead_node, u8 send_to)
944 {
945 	int ret, tmpret;
946 	struct dlm_reco_data_done done_msg;
947 
948 	memset(&done_msg, 0, sizeof(done_msg));
949 	done_msg.node_idx = dlm->node_num;
950 	done_msg.dead_node = dead_node;
951 	mlog(0, "sending DATA DONE message to %u, "
952 	     "my node=%u, dead node=%u\n", send_to, done_msg.node_idx,
953 	     done_msg.dead_node);
954 
955 	ret = o2net_send_message(DLM_RECO_DATA_DONE_MSG, dlm->key, &done_msg,
956 				 sizeof(done_msg), send_to, &tmpret);
957 	if (ret < 0) {
958 		if (!dlm_is_host_down(ret)) {
959 			mlog_errno(ret);
960 			mlog(ML_ERROR, "%s: unknown error sending data-done "
961 			     "to %u\n", dlm->name, send_to);
962 			BUG();
963 		}
964 	} else
965 		ret = tmpret;
966 	return ret;
967 }
968 
969 
970 int dlm_reco_data_done_handler(struct o2net_msg *msg, u32 len, void *data,
971 			       void **ret_data)
972 {
973 	struct dlm_ctxt *dlm = data;
974 	struct dlm_reco_data_done *done = (struct dlm_reco_data_done *)msg->buf;
975 	struct dlm_reco_node_data *ndata = NULL;
976 	int ret = -EINVAL;
977 
978 	if (!dlm_grab(dlm))
979 		return -EINVAL;
980 
981 	mlog(0, "got DATA DONE: dead_node=%u, reco.dead_node=%u, "
982 	     "node_idx=%u, this node=%u\n", done->dead_node,
983 	     dlm->reco.dead_node, done->node_idx, dlm->node_num);
984 
985 	mlog_bug_on_msg((done->dead_node != dlm->reco.dead_node),
986 			"Got DATA DONE: dead_node=%u, reco.dead_node=%u, "
987 			"node_idx=%u, this node=%u\n", done->dead_node,
988 			dlm->reco.dead_node, done->node_idx, dlm->node_num);
989 
990 	spin_lock(&dlm_reco_state_lock);
991 	list_for_each_entry(ndata, &dlm->reco.node_data, list) {
992 		if (ndata->node_num != done->node_idx)
993 			continue;
994 
995 		switch (ndata->state) {
996 			/* should have moved beyond INIT but not to FINALIZE yet */
997 			case DLM_RECO_NODE_DATA_INIT:
998 			case DLM_RECO_NODE_DATA_DEAD:
999 			case DLM_RECO_NODE_DATA_FINALIZE_SENT:
1000 				mlog(ML_ERROR, "bad ndata state for node %u:"
1001 				     " state=%d\n", ndata->node_num,
1002 				     ndata->state);
1003 				BUG();
1004 				break;
1005 			/* these states are possible at this point, anywhere along
1006 			 * the line of recovery */
1007 			case DLM_RECO_NODE_DATA_DONE:
1008 			case DLM_RECO_NODE_DATA_RECEIVING:
1009 			case DLM_RECO_NODE_DATA_REQUESTED:
1010 			case DLM_RECO_NODE_DATA_REQUESTING:
1011 				mlog(0, "node %u is DONE sending "
1012 					  "recovery data!\n",
1013 					  ndata->node_num);
1014 
1015 				ndata->state = DLM_RECO_NODE_DATA_DONE;
1016 				ret = 0;
1017 				break;
1018 		}
1019 	}
1020 	spin_unlock(&dlm_reco_state_lock);
1021 
1022 	/* wake the recovery thread, some node is done */
1023 	if (!ret)
1024 		dlm_kick_recovery_thread(dlm);
1025 
1026 	if (ret < 0)
1027 		mlog(ML_ERROR, "failed to find recovery node data for node "
1028 		     "%u\n", done->node_idx);
1029 	dlm_put(dlm);
1030 
1031 	mlog(0, "leaving reco data done handler, ret=%d\n", ret);
1032 	return ret;
1033 }
1034 
1035 static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm,
1036 					struct list_head *list,
1037 				       	u8 dead_node)
1038 {
1039 	struct dlm_lock_resource *res, *next;
1040 	struct dlm_lock *lock;
1041 
1042 	spin_lock(&dlm->spinlock);
1043 	list_for_each_entry_safe(res, next, &dlm->reco.resources, recovering) {
1044 		/* always prune any $RECOVERY entries for dead nodes,
1045 		 * otherwise hangs can occur during later recovery */
1046 		if (dlm_is_recovery_lock(res->lockname.name,
1047 					 res->lockname.len)) {
1048 			spin_lock(&res->spinlock);
1049 			list_for_each_entry(lock, &res->granted, list) {
1050 				if (lock->ml.node == dead_node) {
1051 					mlog(0, "AHA! there was "
1052 					     "a $RECOVERY lock for dead "
1053 					     "node %u (%s)!\n",
1054 					     dead_node, dlm->name);
1055 					list_del_init(&lock->list);
1056 					dlm_lock_put(lock);
1057 					break;
1058 				}
1059 			}
1060 			spin_unlock(&res->spinlock);
1061 			continue;
1062 		}
1063 
1064 		if (res->owner == dead_node) {
1065 			mlog(0, "found lockres owned by dead node while "
1066 				  "doing recovery for node %u. sending it.\n",
1067 				  dead_node);
1068 			list_move_tail(&res->recovering, list);
1069 		} else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
1070 			mlog(0, "found UNKNOWN owner while doing recovery "
1071 				  "for node %u. sending it.\n", dead_node);
1072 			list_move_tail(&res->recovering, list);
1073 		}
1074 	}
1075 	spin_unlock(&dlm->spinlock);
1076 }
1077 
1078 static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res)
1079 {
1080 	int total_locks = 0;
1081 	struct list_head *iter, *queue = &res->granted;
1082 	int i;
1083 
1084 	for (i=0; i<3; i++) {
1085 		list_for_each(iter, queue)
1086 			total_locks++;
1087 		queue++;
1088 	}
1089 	return total_locks;
1090 }
1091 
1092 
1093 static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm,
1094 				      struct dlm_migratable_lockres *mres,
1095 				      u8 send_to,
1096 				      struct dlm_lock_resource *res,
1097 				      int total_locks)
1098 {
1099 	u64 mig_cookie = be64_to_cpu(mres->mig_cookie);
1100 	int mres_total_locks = be32_to_cpu(mres->total_locks);
1101 	int sz, ret = 0, status = 0;
1102 	u8 orig_flags = mres->flags,
1103 	   orig_master = mres->master;
1104 
1105 	BUG_ON(mres->num_locks > DLM_MAX_MIGRATABLE_LOCKS);
1106 	if (!mres->num_locks)
1107 		return 0;
1108 
1109 	sz = sizeof(struct dlm_migratable_lockres) +
1110 		(mres->num_locks * sizeof(struct dlm_migratable_lock));
1111 
1112 	/* add an all-done flag if we reached the last lock */
1113 	orig_flags = mres->flags;
1114 	BUG_ON(total_locks > mres_total_locks);
1115 	if (total_locks == mres_total_locks)
1116 		mres->flags |= DLM_MRES_ALL_DONE;
1117 
1118 	mlog(0, "%s:%.*s: sending mig lockres (%s) to %u\n",
1119 	     dlm->name, res->lockname.len, res->lockname.name,
1120 	     orig_flags & DLM_MRES_MIGRATION ? "migration" : "recovery",
1121 	     send_to);
1122 
1123 	/* send it */
1124 	ret = o2net_send_message(DLM_MIG_LOCKRES_MSG, dlm->key, mres,
1125 				 sz, send_to, &status);
1126 	if (ret < 0) {
1127 		/* XXX: negative status is not handled.
1128 		 * this will end up killing this node. */
1129 		mlog_errno(ret);
1130 	} else {
1131 		/* might get an -ENOMEM back here */
1132 		ret = status;
1133 		if (ret < 0) {
1134 			mlog_errno(ret);
1135 
1136 			if (ret == -EFAULT) {
1137 				mlog(ML_ERROR, "node %u told me to kill "
1138 				     "myself!\n", send_to);
1139 				BUG();
1140 			}
1141 		}
1142 	}
1143 
1144 	/* zero and reinit the message buffer */
1145 	dlm_init_migratable_lockres(mres, res->lockname.name,
1146 				    res->lockname.len, mres_total_locks,
1147 				    mig_cookie, orig_flags, orig_master);
1148 	return ret;
1149 }
1150 
1151 static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres,
1152 					const char *lockname, int namelen,
1153 					int total_locks, u64 cookie,
1154 					u8 flags, u8 master)
1155 {
1156 	/* mres here is one full page */
1157 	clear_page(mres);
1158 	mres->lockname_len = namelen;
1159 	memcpy(mres->lockname, lockname, namelen);
1160 	mres->num_locks = 0;
1161 	mres->total_locks = cpu_to_be32(total_locks);
1162 	mres->mig_cookie = cpu_to_be64(cookie);
1163 	mres->flags = flags;
1164 	mres->master = master;
1165 }
1166 
1167 
1168 /* returns 1 if this lock fills the network structure,
1169  * 0 otherwise */
1170 static int dlm_add_lock_to_array(struct dlm_lock *lock,
1171 				 struct dlm_migratable_lockres *mres, int queue)
1172 {
1173 	struct dlm_migratable_lock *ml;
1174 	int lock_num = mres->num_locks;
1175 
1176 	ml = &(mres->ml[lock_num]);
1177 	ml->cookie = lock->ml.cookie;
1178 	ml->type = lock->ml.type;
1179 	ml->convert_type = lock->ml.convert_type;
1180 	ml->highest_blocked = lock->ml.highest_blocked;
1181 	ml->list = queue;
1182 	if (lock->lksb) {
1183 		ml->flags = lock->lksb->flags;
1184 		/* send our current lvb */
1185 		if (ml->type == LKM_EXMODE ||
1186 		    ml->type == LKM_PRMODE) {
1187 			/* if it is already set, this had better be a PR
1188 			 * and it has to match */
1189 			if (!dlm_lvb_is_empty(mres->lvb) &&
1190 			    (ml->type == LKM_EXMODE ||
1191 			     memcmp(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN))) {
1192 				mlog(ML_ERROR, "mismatched lvbs!\n");
1193 				dlm_print_one_lock_resource(lock->lockres);
1194 				BUG();
1195 			}
1196 			memcpy(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN);
1197 		}
1198 	}
1199 	ml->node = lock->ml.node;
1200 	mres->num_locks++;
1201 	/* we reached the max, send this network message */
1202 	if (mres->num_locks == DLM_MAX_MIGRATABLE_LOCKS)
1203 		return 1;
1204 	return 0;
1205 }
1206 
1207 static void dlm_add_dummy_lock(struct dlm_ctxt *dlm,
1208 			       struct dlm_migratable_lockres *mres)
1209 {
1210 	struct dlm_lock dummy;
1211 	memset(&dummy, 0, sizeof(dummy));
1212 	dummy.ml.cookie = 0;
1213 	dummy.ml.type = LKM_IVMODE;
1214 	dummy.ml.convert_type = LKM_IVMODE;
1215 	dummy.ml.highest_blocked = LKM_IVMODE;
1216 	dummy.lksb = NULL;
1217 	dummy.ml.node = dlm->node_num;
1218 	dlm_add_lock_to_array(&dummy, mres, DLM_BLOCKED_LIST);
1219 }
1220 
1221 static inline int dlm_is_dummy_lock(struct dlm_ctxt *dlm,
1222 				    struct dlm_migratable_lock *ml,
1223 				    u8 *nodenum)
1224 {
1225 	if (unlikely(ml->cookie == 0 &&
1226 	    ml->type == LKM_IVMODE &&
1227 	    ml->convert_type == LKM_IVMODE &&
1228 	    ml->highest_blocked == LKM_IVMODE &&
1229 	    ml->list == DLM_BLOCKED_LIST)) {
1230 		*nodenum = ml->node;
1231 		return 1;
1232 	}
1233 	return 0;
1234 }
1235 
1236 int dlm_send_one_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
1237 			 struct dlm_migratable_lockres *mres,
1238 			 u8 send_to, u8 flags)
1239 {
1240 	struct list_head *queue;
1241 	int total_locks, i;
1242 	u64 mig_cookie = 0;
1243 	struct dlm_lock *lock;
1244 	int ret = 0;
1245 
1246 	BUG_ON(!(flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION)));
1247 
1248 	mlog(0, "sending to %u\n", send_to);
1249 
1250 	total_locks = dlm_num_locks_in_lockres(res);
1251 	if (total_locks > DLM_MAX_MIGRATABLE_LOCKS) {
1252 		/* rare, but possible */
1253 		mlog(0, "argh.  lockres has %d locks.  this will "
1254 			  "require more than one network packet to "
1255 			  "migrate\n", total_locks);
1256 		mig_cookie = dlm_get_next_mig_cookie();
1257 	}
1258 
1259 	dlm_init_migratable_lockres(mres, res->lockname.name,
1260 				    res->lockname.len, total_locks,
1261 				    mig_cookie, flags, res->owner);
1262 
1263 	total_locks = 0;
1264 	for (i=DLM_GRANTED_LIST; i<=DLM_BLOCKED_LIST; i++) {
1265 		queue = dlm_list_idx_to_ptr(res, i);
1266 		list_for_each_entry(lock, queue, list) {
1267 			/* add another lock. */
1268 			total_locks++;
1269 			if (!dlm_add_lock_to_array(lock, mres, i))
1270 				continue;
1271 
1272 			/* this filled the lock message,
1273 			 * we must send it immediately. */
1274 			ret = dlm_send_mig_lockres_msg(dlm, mres, send_to,
1275 						       res, total_locks);
1276 			if (ret < 0)
1277 				goto error;
1278 		}
1279 	}
1280 	if (total_locks == 0) {
1281 		/* send a dummy lock to indicate a mastery reference only */
1282 		mlog(0, "%s:%.*s: sending dummy lock to %u, %s\n",
1283 		     dlm->name, res->lockname.len, res->lockname.name,
1284 		     send_to, flags & DLM_MRES_RECOVERY ? "recovery" :
1285 		     "migration");
1286 		dlm_add_dummy_lock(dlm, mres);
1287 	}
1288 	/* flush any remaining locks */
1289 	ret = dlm_send_mig_lockres_msg(dlm, mres, send_to, res, total_locks);
1290 	if (ret < 0)
1291 		goto error;
1292 	return ret;
1293 
1294 error:
1295 	mlog(ML_ERROR, "%s: dlm_send_mig_lockres_msg returned %d\n",
1296 	     dlm->name, ret);
1297 	if (!dlm_is_host_down(ret))
1298 		BUG();
1299 	mlog(0, "%s: node %u went down while sending %s "
1300 	     "lockres %.*s\n", dlm->name, send_to,
1301 	     flags & DLM_MRES_RECOVERY ?  "recovery" : "migration",
1302 	     res->lockname.len, res->lockname.name);
1303 	return ret;
1304 }
1305 
1306 
1307 
1308 /*
1309  * this message will contain no more than one page worth of
1310  * recovery data, and it will work on only one lockres.
1311  * there may be many locks in this page, and we may need to wait
1312  * for additional packets to complete all the locks (rare, but
1313  * possible).
1314  */
1315 /*
1316  * NOTE: the allocation error cases here are scary
1317  * we really cannot afford to fail an alloc in recovery
1318  * do we spin?  returning an error only delays the problem really
1319  */
1320 
1321 int dlm_mig_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
1322 			    void **ret_data)
1323 {
1324 	struct dlm_ctxt *dlm = data;
1325 	struct dlm_migratable_lockres *mres =
1326 		(struct dlm_migratable_lockres *)msg->buf;
1327 	int ret = 0;
1328 	u8 real_master;
1329 	u8 extra_refs = 0;
1330 	char *buf = NULL;
1331 	struct dlm_work_item *item = NULL;
1332 	struct dlm_lock_resource *res = NULL;
1333 
1334 	if (!dlm_grab(dlm))
1335 		return -EINVAL;
1336 
1337 	BUG_ON(!(mres->flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION)));
1338 
1339 	real_master = mres->master;
1340 	if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1341 		/* cannot migrate a lockres with no master */
1342 		BUG_ON(!(mres->flags & DLM_MRES_RECOVERY));
1343 	}
1344 
1345 	mlog(0, "%s message received from node %u\n",
1346 		  (mres->flags & DLM_MRES_RECOVERY) ?
1347 		  "recovery" : "migration", mres->master);
1348 	if (mres->flags & DLM_MRES_ALL_DONE)
1349 		mlog(0, "all done flag.  all lockres data received!\n");
1350 
1351 	ret = -ENOMEM;
1352 	buf = kmalloc(be16_to_cpu(msg->data_len), GFP_NOFS);
1353 	item = kzalloc(sizeof(*item), GFP_NOFS);
1354 	if (!buf || !item)
1355 		goto leave;
1356 
1357 	/* lookup the lock to see if we have a secondary queue for this
1358 	 * already...  just add the locks in and this will have its owner
1359 	 * and RECOVERY flag changed when it completes. */
1360 	res = dlm_lookup_lockres(dlm, mres->lockname, mres->lockname_len);
1361 	if (res) {
1362 	 	/* this will get a ref on res */
1363 		/* mark it as recovering/migrating and hash it */
1364 		spin_lock(&res->spinlock);
1365 		if (mres->flags & DLM_MRES_RECOVERY) {
1366 			res->state |= DLM_LOCK_RES_RECOVERING;
1367 		} else {
1368 			if (res->state & DLM_LOCK_RES_MIGRATING) {
1369 				/* this is at least the second
1370 				 * lockres message */
1371 				mlog(0, "lock %.*s is already migrating\n",
1372 					  mres->lockname_len,
1373 					  mres->lockname);
1374 			} else if (res->state & DLM_LOCK_RES_RECOVERING) {
1375 				/* caller should BUG */
1376 				mlog(ML_ERROR, "node is attempting to migrate "
1377 				     "lock %.*s, but marked as recovering!\n",
1378 				     mres->lockname_len, mres->lockname);
1379 				ret = -EFAULT;
1380 				spin_unlock(&res->spinlock);
1381 				goto leave;
1382 			}
1383 			res->state |= DLM_LOCK_RES_MIGRATING;
1384 		}
1385 		spin_unlock(&res->spinlock);
1386 	} else {
1387 		/* need to allocate, just like if it was
1388 		 * mastered here normally  */
1389 		res = dlm_new_lockres(dlm, mres->lockname, mres->lockname_len);
1390 		if (!res)
1391 			goto leave;
1392 
1393 		/* to match the ref that we would have gotten if
1394 		 * dlm_lookup_lockres had succeeded */
1395 		dlm_lockres_get(res);
1396 
1397 		/* mark it as recovering/migrating and hash it */
1398 		if (mres->flags & DLM_MRES_RECOVERY)
1399 			res->state |= DLM_LOCK_RES_RECOVERING;
1400 		else
1401 			res->state |= DLM_LOCK_RES_MIGRATING;
1402 
1403 		spin_lock(&dlm->spinlock);
1404 		__dlm_insert_lockres(dlm, res);
1405 		spin_unlock(&dlm->spinlock);
1406 
1407 		/* Add an extra ref for this lock-less lockres lest the
1408 		 * dlm_thread purges it before we get the chance to add
1409 		 * locks to it */
1410 		dlm_lockres_get(res);
1411 
1412 		/* There are three refs that need to be put.
1413 		 * 1. Taken above.
1414 		 * 2. kref_init in dlm_new_lockres()->dlm_init_lockres().
1415 		 * 3. dlm_lookup_lockres()
1416 		 * The first one is handled at the end of this function. The
1417 		 * other two are handled in the worker thread after locks have
1418 		 * been attached. Yes, we don't wait for purge time to match
1419 		 * kref_init. The lockres will still have atleast one ref
1420 		 * added because it is in the hash __dlm_insert_lockres() */
1421 		extra_refs++;
1422 
1423 		/* now that the new lockres is inserted,
1424 		 * make it usable by other processes */
1425 		spin_lock(&res->spinlock);
1426 		res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
1427 		spin_unlock(&res->spinlock);
1428 		wake_up(&res->wq);
1429 	}
1430 
1431 	/* at this point we have allocated everything we need,
1432 	 * and we have a hashed lockres with an extra ref and
1433 	 * the proper res->state flags. */
1434 	ret = 0;
1435 	spin_lock(&res->spinlock);
1436 	/* drop this either when master requery finds a different master
1437 	 * or when a lock is added by the recovery worker */
1438 	dlm_lockres_grab_inflight_ref(dlm, res);
1439 	if (mres->master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1440 		/* migration cannot have an unknown master */
1441 		BUG_ON(!(mres->flags & DLM_MRES_RECOVERY));
1442 		mlog(0, "recovery has passed me a lockres with an "
1443 			  "unknown owner.. will need to requery: "
1444 			  "%.*s\n", mres->lockname_len, mres->lockname);
1445 	} else {
1446 		/* take a reference now to pin the lockres, drop it
1447 		 * when locks are added in the worker */
1448 		dlm_change_lockres_owner(dlm, res, dlm->node_num);
1449 	}
1450 	spin_unlock(&res->spinlock);
1451 
1452 	/* queue up work for dlm_mig_lockres_worker */
1453 	dlm_grab(dlm);  /* get an extra ref for the work item */
1454 	memcpy(buf, msg->buf, be16_to_cpu(msg->data_len));  /* copy the whole message */
1455 	dlm_init_work_item(dlm, item, dlm_mig_lockres_worker, buf);
1456 	item->u.ml.lockres = res; /* already have a ref */
1457 	item->u.ml.real_master = real_master;
1458 	item->u.ml.extra_ref = extra_refs;
1459 	spin_lock(&dlm->work_lock);
1460 	list_add_tail(&item->list, &dlm->work_list);
1461 	spin_unlock(&dlm->work_lock);
1462 	queue_work(dlm->dlm_worker, &dlm->dispatched_work);
1463 
1464 leave:
1465 	/* One extra ref taken needs to be put here */
1466 	if (extra_refs)
1467 		dlm_lockres_put(res);
1468 
1469 	dlm_put(dlm);
1470 	if (ret < 0) {
1471 		if (buf)
1472 			kfree(buf);
1473 		if (item)
1474 			kfree(item);
1475 	}
1476 
1477 	mlog_exit(ret);
1478 	return ret;
1479 }
1480 
1481 
1482 static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data)
1483 {
1484 	struct dlm_ctxt *dlm;
1485 	struct dlm_migratable_lockres *mres;
1486 	int ret = 0;
1487 	struct dlm_lock_resource *res;
1488 	u8 real_master;
1489 	u8 extra_ref;
1490 
1491 	dlm = item->dlm;
1492 	mres = (struct dlm_migratable_lockres *)data;
1493 
1494 	res = item->u.ml.lockres;
1495 	real_master = item->u.ml.real_master;
1496 	extra_ref = item->u.ml.extra_ref;
1497 
1498 	if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1499 		/* this case is super-rare. only occurs if
1500 		 * node death happens during migration. */
1501 again:
1502 		ret = dlm_lockres_master_requery(dlm, res, &real_master);
1503 		if (ret < 0) {
1504 			mlog(0, "dlm_lockres_master_requery ret=%d\n",
1505 				  ret);
1506 			goto again;
1507 		}
1508 		if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1509 			mlog(0, "lockres %.*s not claimed.  "
1510 				   "this node will take it.\n",
1511 				   res->lockname.len, res->lockname.name);
1512 		} else {
1513 			spin_lock(&res->spinlock);
1514 			dlm_lockres_drop_inflight_ref(dlm, res);
1515 			spin_unlock(&res->spinlock);
1516 			mlog(0, "master needs to respond to sender "
1517 				  "that node %u still owns %.*s\n",
1518 				  real_master, res->lockname.len,
1519 				  res->lockname.name);
1520 			/* cannot touch this lockres */
1521 			goto leave;
1522 		}
1523 	}
1524 
1525 	ret = dlm_process_recovery_data(dlm, res, mres);
1526 	if (ret < 0)
1527 		mlog(0, "dlm_process_recovery_data returned  %d\n", ret);
1528 	else
1529 		mlog(0, "dlm_process_recovery_data succeeded\n");
1530 
1531 	if ((mres->flags & (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) ==
1532 	                   (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) {
1533 		ret = dlm_finish_migration(dlm, res, mres->master);
1534 		if (ret < 0)
1535 			mlog_errno(ret);
1536 	}
1537 
1538 leave:
1539 	/* See comment in dlm_mig_lockres_handler() */
1540 	if (res) {
1541 		if (extra_ref)
1542 			dlm_lockres_put(res);
1543 		dlm_lockres_put(res);
1544 	}
1545 	kfree(data);
1546 	mlog_exit(ret);
1547 }
1548 
1549 
1550 
1551 static int dlm_lockres_master_requery(struct dlm_ctxt *dlm,
1552 				      struct dlm_lock_resource *res,
1553 				      u8 *real_master)
1554 {
1555 	struct dlm_node_iter iter;
1556 	int nodenum;
1557 	int ret = 0;
1558 
1559 	*real_master = DLM_LOCK_RES_OWNER_UNKNOWN;
1560 
1561 	/* we only reach here if one of the two nodes in a
1562 	 * migration died while the migration was in progress.
1563 	 * at this point we need to requery the master.  we
1564 	 * know that the new_master got as far as creating
1565 	 * an mle on at least one node, but we do not know
1566 	 * if any nodes had actually cleared the mle and set
1567 	 * the master to the new_master.  the old master
1568 	 * is supposed to set the owner to UNKNOWN in the
1569 	 * event of a new_master death, so the only possible
1570 	 * responses that we can get from nodes here are
1571 	 * that the master is new_master, or that the master
1572 	 * is UNKNOWN.
1573 	 * if all nodes come back with UNKNOWN then we know
1574 	 * the lock needs remastering here.
1575 	 * if any node comes back with a valid master, check
1576 	 * to see if that master is the one that we are
1577 	 * recovering.  if so, then the new_master died and
1578 	 * we need to remaster this lock.  if not, then the
1579 	 * new_master survived and that node will respond to
1580 	 * other nodes about the owner.
1581 	 * if there is an owner, this node needs to dump this
1582 	 * lockres and alert the sender that this lockres
1583 	 * was rejected. */
1584 	spin_lock(&dlm->spinlock);
1585 	dlm_node_iter_init(dlm->domain_map, &iter);
1586 	spin_unlock(&dlm->spinlock);
1587 
1588 	while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
1589 		/* do not send to self */
1590 		if (nodenum == dlm->node_num)
1591 			continue;
1592 		ret = dlm_do_master_requery(dlm, res, nodenum, real_master);
1593 		if (ret < 0) {
1594 			mlog_errno(ret);
1595 			if (!dlm_is_host_down(ret))
1596 				BUG();
1597 			/* host is down, so answer for that node would be
1598 			 * DLM_LOCK_RES_OWNER_UNKNOWN.  continue. */
1599 		}
1600 		if (*real_master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1601 			mlog(0, "lock master is %u\n", *real_master);
1602 			break;
1603 		}
1604 	}
1605 	return ret;
1606 }
1607 
1608 
1609 int dlm_do_master_requery(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
1610 			  u8 nodenum, u8 *real_master)
1611 {
1612 	int ret = -EINVAL;
1613 	struct dlm_master_requery req;
1614 	int status = DLM_LOCK_RES_OWNER_UNKNOWN;
1615 
1616 	memset(&req, 0, sizeof(req));
1617 	req.node_idx = dlm->node_num;
1618 	req.namelen = res->lockname.len;
1619 	memcpy(req.name, res->lockname.name, res->lockname.len);
1620 
1621 	ret = o2net_send_message(DLM_MASTER_REQUERY_MSG, dlm->key,
1622 				 &req, sizeof(req), nodenum, &status);
1623 	/* XXX: negative status not handled properly here. */
1624 	if (ret < 0)
1625 		mlog_errno(ret);
1626 	else {
1627 		BUG_ON(status < 0);
1628 		BUG_ON(status > DLM_LOCK_RES_OWNER_UNKNOWN);
1629 		*real_master = (u8) (status & 0xff);
1630 		mlog(0, "node %u responded to master requery with %u\n",
1631 			  nodenum, *real_master);
1632 		ret = 0;
1633 	}
1634 	return ret;
1635 }
1636 
1637 
1638 /* this function cannot error, so unless the sending
1639  * or receiving of the message failed, the owner can
1640  * be trusted */
1641 int dlm_master_requery_handler(struct o2net_msg *msg, u32 len, void *data,
1642 			       void **ret_data)
1643 {
1644 	struct dlm_ctxt *dlm = data;
1645 	struct dlm_master_requery *req = (struct dlm_master_requery *)msg->buf;
1646 	struct dlm_lock_resource *res = NULL;
1647 	unsigned int hash;
1648 	int master = DLM_LOCK_RES_OWNER_UNKNOWN;
1649 	u32 flags = DLM_ASSERT_MASTER_REQUERY;
1650 
1651 	if (!dlm_grab(dlm)) {
1652 		/* since the domain has gone away on this
1653 		 * node, the proper response is UNKNOWN */
1654 		return master;
1655 	}
1656 
1657 	hash = dlm_lockid_hash(req->name, req->namelen);
1658 
1659 	spin_lock(&dlm->spinlock);
1660 	res = __dlm_lookup_lockres(dlm, req->name, req->namelen, hash);
1661 	if (res) {
1662 		spin_lock(&res->spinlock);
1663 		master = res->owner;
1664 		if (master == dlm->node_num) {
1665 			int ret = dlm_dispatch_assert_master(dlm, res,
1666 							     0, 0, flags);
1667 			if (ret < 0) {
1668 				mlog_errno(-ENOMEM);
1669 				/* retry!? */
1670 				BUG();
1671 			}
1672 		} else /* put.. incase we are not the master */
1673 			dlm_lockres_put(res);
1674 		spin_unlock(&res->spinlock);
1675 	}
1676 	spin_unlock(&dlm->spinlock);
1677 
1678 	dlm_put(dlm);
1679 	return master;
1680 }
1681 
1682 static inline struct list_head *
1683 dlm_list_num_to_pointer(struct dlm_lock_resource *res, int list_num)
1684 {
1685 	struct list_head *ret;
1686 	BUG_ON(list_num < 0);
1687 	BUG_ON(list_num > 2);
1688 	ret = &(res->granted);
1689 	ret += list_num;
1690 	return ret;
1691 }
1692 /* TODO: do ast flush business
1693  * TODO: do MIGRATING and RECOVERING spinning
1694  */
1695 
1696 /*
1697 * NOTE about in-flight requests during migration:
1698 *
1699 * Before attempting the migrate, the master has marked the lockres as
1700 * MIGRATING and then flushed all of its pending ASTS.  So any in-flight
1701 * requests either got queued before the MIGRATING flag got set, in which
1702 * case the lock data will reflect the change and a return message is on
1703 * the way, or the request failed to get in before MIGRATING got set.  In
1704 * this case, the caller will be told to spin and wait for the MIGRATING
1705 * flag to be dropped, then recheck the master.
1706 * This holds true for the convert, cancel and unlock cases, and since lvb
1707 * updates are tied to these same messages, it applies to lvb updates as
1708 * well.  For the lock case, there is no way a lock can be on the master
1709 * queue and not be on the secondary queue since the lock is always added
1710 * locally first.  This means that the new target node will never be sent
1711 * a lock that he doesn't already have on the list.
1712 * In total, this means that the local lock is correct and should not be
1713 * updated to match the one sent by the master.  Any messages sent back
1714 * from the master before the MIGRATING flag will bring the lock properly
1715 * up-to-date, and the change will be ordered properly for the waiter.
1716 * We will *not* attempt to modify the lock underneath the waiter.
1717 */
1718 
1719 static int dlm_process_recovery_data(struct dlm_ctxt *dlm,
1720 				     struct dlm_lock_resource *res,
1721 				     struct dlm_migratable_lockres *mres)
1722 {
1723 	struct dlm_migratable_lock *ml;
1724 	struct list_head *queue;
1725 	struct list_head *tmpq = NULL;
1726 	struct dlm_lock *newlock = NULL;
1727 	struct dlm_lockstatus *lksb = NULL;
1728 	int ret = 0;
1729 	int i, j, bad;
1730 	struct dlm_lock *lock = NULL;
1731 	u8 from = O2NM_MAX_NODES;
1732 	unsigned int added = 0;
1733 
1734 	mlog(0, "running %d locks for this lockres\n", mres->num_locks);
1735 	for (i=0; i<mres->num_locks; i++) {
1736 		ml = &(mres->ml[i]);
1737 
1738 		if (dlm_is_dummy_lock(dlm, ml, &from)) {
1739 			/* placeholder, just need to set the refmap bit */
1740 			BUG_ON(mres->num_locks != 1);
1741 			mlog(0, "%s:%.*s: dummy lock for %u\n",
1742 			     dlm->name, mres->lockname_len, mres->lockname,
1743 			     from);
1744 			spin_lock(&res->spinlock);
1745 			dlm_lockres_set_refmap_bit(from, res);
1746 			spin_unlock(&res->spinlock);
1747 			added++;
1748 			break;
1749 		}
1750 		BUG_ON(ml->highest_blocked != LKM_IVMODE);
1751 		newlock = NULL;
1752 		lksb = NULL;
1753 
1754 		queue = dlm_list_num_to_pointer(res, ml->list);
1755 		tmpq = NULL;
1756 
1757 		/* if the lock is for the local node it needs to
1758 		 * be moved to the proper location within the queue.
1759 		 * do not allocate a new lock structure. */
1760 		if (ml->node == dlm->node_num) {
1761 			/* MIGRATION ONLY! */
1762 			BUG_ON(!(mres->flags & DLM_MRES_MIGRATION));
1763 
1764 			spin_lock(&res->spinlock);
1765 			for (j = DLM_GRANTED_LIST; j <= DLM_BLOCKED_LIST; j++) {
1766 				tmpq = dlm_list_idx_to_ptr(res, j);
1767 				list_for_each_entry(lock, tmpq, list) {
1768 					if (lock->ml.cookie != ml->cookie)
1769 						lock = NULL;
1770 					else
1771 						break;
1772 				}
1773 				if (lock)
1774 					break;
1775 			}
1776 
1777 			/* lock is always created locally first, and
1778 			 * destroyed locally last.  it must be on the list */
1779 			if (!lock) {
1780 				__be64 c = ml->cookie;
1781 				mlog(ML_ERROR, "could not find local lock "
1782 					       "with cookie %u:%llu!\n",
1783 				     dlm_get_lock_cookie_node(be64_to_cpu(c)),
1784 				     dlm_get_lock_cookie_seq(be64_to_cpu(c)));
1785 				__dlm_print_one_lock_resource(res);
1786 				BUG();
1787 			}
1788 			BUG_ON(lock->ml.node != ml->node);
1789 
1790 			if (tmpq != queue) {
1791 				mlog(0, "lock was on %u instead of %u for %.*s\n",
1792 				     j, ml->list, res->lockname.len, res->lockname.name);
1793 				spin_unlock(&res->spinlock);
1794 				continue;
1795 			}
1796 
1797 			/* see NOTE above about why we do not update
1798 			 * to match the master here */
1799 
1800 			/* move the lock to its proper place */
1801 			/* do not alter lock refcount.  switching lists. */
1802 			list_move_tail(&lock->list, queue);
1803 			spin_unlock(&res->spinlock);
1804 			added++;
1805 
1806 			mlog(0, "just reordered a local lock!\n");
1807 			continue;
1808 		}
1809 
1810 		/* lock is for another node. */
1811 		newlock = dlm_new_lock(ml->type, ml->node,
1812 				       be64_to_cpu(ml->cookie), NULL);
1813 		if (!newlock) {
1814 			ret = -ENOMEM;
1815 			goto leave;
1816 		}
1817 		lksb = newlock->lksb;
1818 		dlm_lock_attach_lockres(newlock, res);
1819 
1820 		if (ml->convert_type != LKM_IVMODE) {
1821 			BUG_ON(queue != &res->converting);
1822 			newlock->ml.convert_type = ml->convert_type;
1823 		}
1824 		lksb->flags |= (ml->flags &
1825 				(DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB));
1826 
1827 		if (ml->type == LKM_NLMODE)
1828 			goto skip_lvb;
1829 
1830 		if (!dlm_lvb_is_empty(mres->lvb)) {
1831 			if (lksb->flags & DLM_LKSB_PUT_LVB) {
1832 				/* other node was trying to update
1833 				 * lvb when node died.  recreate the
1834 				 * lksb with the updated lvb. */
1835 				memcpy(lksb->lvb, mres->lvb, DLM_LVB_LEN);
1836 				/* the lock resource lvb update must happen
1837 				 * NOW, before the spinlock is dropped.
1838 				 * we no longer wait for the AST to update
1839 				 * the lvb. */
1840 				memcpy(res->lvb, mres->lvb, DLM_LVB_LEN);
1841 			} else {
1842 				/* otherwise, the node is sending its
1843 				 * most recent valid lvb info */
1844 				BUG_ON(ml->type != LKM_EXMODE &&
1845 				       ml->type != LKM_PRMODE);
1846 				if (!dlm_lvb_is_empty(res->lvb) &&
1847  				    (ml->type == LKM_EXMODE ||
1848  				     memcmp(res->lvb, mres->lvb, DLM_LVB_LEN))) {
1849  					int i;
1850  					mlog(ML_ERROR, "%s:%.*s: received bad "
1851  					     "lvb! type=%d\n", dlm->name,
1852  					     res->lockname.len,
1853  					     res->lockname.name, ml->type);
1854  					printk("lockres lvb=[");
1855  					for (i=0; i<DLM_LVB_LEN; i++)
1856  						printk("%02x", res->lvb[i]);
1857  					printk("]\nmigrated lvb=[");
1858  					for (i=0; i<DLM_LVB_LEN; i++)
1859  						printk("%02x", mres->lvb[i]);
1860  					printk("]\n");
1861  					dlm_print_one_lock_resource(res);
1862  					BUG();
1863 				}
1864 				memcpy(res->lvb, mres->lvb, DLM_LVB_LEN);
1865 			}
1866 		}
1867 skip_lvb:
1868 
1869 		/* NOTE:
1870 		 * wrt lock queue ordering and recovery:
1871 		 *    1. order of locks on granted queue is
1872 		 *       meaningless.
1873 		 *    2. order of locks on converting queue is
1874 		 *       LOST with the node death.  sorry charlie.
1875 		 *    3. order of locks on the blocked queue is
1876 		 *       also LOST.
1877 		 * order of locks does not affect integrity, it
1878 		 * just means that a lock request may get pushed
1879 		 * back in line as a result of the node death.
1880 		 * also note that for a given node the lock order
1881 		 * for its secondary queue locks is preserved
1882 		 * relative to each other, but clearly *not*
1883 		 * preserved relative to locks from other nodes.
1884 		 */
1885 		bad = 0;
1886 		spin_lock(&res->spinlock);
1887 		list_for_each_entry(lock, queue, list) {
1888 			if (lock->ml.cookie == ml->cookie) {
1889 				__be64 c = lock->ml.cookie;
1890 				mlog(ML_ERROR, "%s:%.*s: %u:%llu: lock already "
1891 				     "exists on this lockres!\n", dlm->name,
1892 				     res->lockname.len, res->lockname.name,
1893 				     dlm_get_lock_cookie_node(be64_to_cpu(c)),
1894 				     dlm_get_lock_cookie_seq(be64_to_cpu(c)));
1895 
1896 				mlog(ML_NOTICE, "sent lock: type=%d, conv=%d, "
1897 				     "node=%u, cookie=%u:%llu, queue=%d\n",
1898 	      			     ml->type, ml->convert_type, ml->node,
1899 				     dlm_get_lock_cookie_node(be64_to_cpu(ml->cookie)),
1900 				     dlm_get_lock_cookie_seq(be64_to_cpu(ml->cookie)),
1901 				     ml->list);
1902 
1903 				__dlm_print_one_lock_resource(res);
1904 				bad = 1;
1905 				break;
1906 			}
1907 		}
1908 		if (!bad) {
1909 			dlm_lock_get(newlock);
1910 			list_add_tail(&newlock->list, queue);
1911 			mlog(0, "%s:%.*s: added lock for node %u, "
1912 			     "setting refmap bit\n", dlm->name,
1913 			     res->lockname.len, res->lockname.name, ml->node);
1914 			dlm_lockres_set_refmap_bit(ml->node, res);
1915 			added++;
1916 		}
1917 		spin_unlock(&res->spinlock);
1918 	}
1919 	mlog(0, "done running all the locks\n");
1920 
1921 leave:
1922 	/* balance the ref taken when the work was queued */
1923 	spin_lock(&res->spinlock);
1924 	dlm_lockres_drop_inflight_ref(dlm, res);
1925 	spin_unlock(&res->spinlock);
1926 
1927 	if (ret < 0) {
1928 		mlog_errno(ret);
1929 		if (newlock)
1930 			dlm_lock_put(newlock);
1931 	}
1932 
1933 	mlog_exit(ret);
1934 	return ret;
1935 }
1936 
1937 void dlm_move_lockres_to_recovery_list(struct dlm_ctxt *dlm,
1938 				       struct dlm_lock_resource *res)
1939 {
1940 	int i;
1941 	struct list_head *queue;
1942 	struct dlm_lock *lock, *next;
1943 
1944 	res->state |= DLM_LOCK_RES_RECOVERING;
1945 	if (!list_empty(&res->recovering)) {
1946 		mlog(0,
1947 		     "Recovering res %s:%.*s, is already on recovery list!\n",
1948 		     dlm->name, res->lockname.len, res->lockname.name);
1949 		list_del_init(&res->recovering);
1950 		dlm_lockres_put(res);
1951 	}
1952 	/* We need to hold a reference while on the recovery list */
1953 	dlm_lockres_get(res);
1954 	list_add_tail(&res->recovering, &dlm->reco.resources);
1955 
1956 	/* find any pending locks and put them back on proper list */
1957 	for (i=DLM_BLOCKED_LIST; i>=DLM_GRANTED_LIST; i--) {
1958 		queue = dlm_list_idx_to_ptr(res, i);
1959 		list_for_each_entry_safe(lock, next, queue, list) {
1960 			dlm_lock_get(lock);
1961 			if (lock->convert_pending) {
1962 				/* move converting lock back to granted */
1963 				BUG_ON(i != DLM_CONVERTING_LIST);
1964 				mlog(0, "node died with convert pending "
1965 				     "on %.*s. move back to granted list.\n",
1966 				     res->lockname.len, res->lockname.name);
1967 				dlm_revert_pending_convert(res, lock);
1968 				lock->convert_pending = 0;
1969 			} else if (lock->lock_pending) {
1970 				/* remove pending lock requests completely */
1971 				BUG_ON(i != DLM_BLOCKED_LIST);
1972 				mlog(0, "node died with lock pending "
1973 				     "on %.*s. remove from blocked list and skip.\n",
1974 				     res->lockname.len, res->lockname.name);
1975 				/* lock will be floating until ref in
1976 				 * dlmlock_remote is freed after the network
1977 				 * call returns.  ok for it to not be on any
1978 				 * list since no ast can be called
1979 				 * (the master is dead). */
1980 				dlm_revert_pending_lock(res, lock);
1981 				lock->lock_pending = 0;
1982 			} else if (lock->unlock_pending) {
1983 				/* if an unlock was in progress, treat as
1984 				 * if this had completed successfully
1985 				 * before sending this lock state to the
1986 				 * new master.  note that the dlm_unlock
1987 				 * call is still responsible for calling
1988 				 * the unlockast.  that will happen after
1989 				 * the network call times out.  for now,
1990 				 * just move lists to prepare the new
1991 				 * recovery master.  */
1992 				BUG_ON(i != DLM_GRANTED_LIST);
1993 				mlog(0, "node died with unlock pending "
1994 				     "on %.*s. remove from blocked list and skip.\n",
1995 				     res->lockname.len, res->lockname.name);
1996 				dlm_commit_pending_unlock(res, lock);
1997 				lock->unlock_pending = 0;
1998 			} else if (lock->cancel_pending) {
1999 				/* if a cancel was in progress, treat as
2000 				 * if this had completed successfully
2001 				 * before sending this lock state to the
2002 				 * new master */
2003 				BUG_ON(i != DLM_CONVERTING_LIST);
2004 				mlog(0, "node died with cancel pending "
2005 				     "on %.*s. move back to granted list.\n",
2006 				     res->lockname.len, res->lockname.name);
2007 				dlm_commit_pending_cancel(res, lock);
2008 				lock->cancel_pending = 0;
2009 			}
2010 			dlm_lock_put(lock);
2011 		}
2012 	}
2013 }
2014 
2015 
2016 
2017 /* removes all recovered locks from the recovery list.
2018  * sets the res->owner to the new master.
2019  * unsets the RECOVERY flag and wakes waiters. */
2020 static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm,
2021 					      u8 dead_node, u8 new_master)
2022 {
2023 	int i;
2024 	struct hlist_node *hash_iter;
2025 	struct hlist_head *bucket;
2026 	struct dlm_lock_resource *res, *next;
2027 
2028 	mlog_entry_void();
2029 
2030 	assert_spin_locked(&dlm->spinlock);
2031 
2032 	list_for_each_entry_safe(res, next, &dlm->reco.resources, recovering) {
2033 		if (res->owner == dead_node) {
2034 			list_del_init(&res->recovering);
2035 			spin_lock(&res->spinlock);
2036 			/* new_master has our reference from
2037 			 * the lock state sent during recovery */
2038 			dlm_change_lockres_owner(dlm, res, new_master);
2039 			res->state &= ~DLM_LOCK_RES_RECOVERING;
2040 			if (__dlm_lockres_has_locks(res))
2041 				__dlm_dirty_lockres(dlm, res);
2042 			spin_unlock(&res->spinlock);
2043 			wake_up(&res->wq);
2044 			dlm_lockres_put(res);
2045 		}
2046 	}
2047 
2048 	/* this will become unnecessary eventually, but
2049 	 * for now we need to run the whole hash, clear
2050 	 * the RECOVERING state and set the owner
2051 	 * if necessary */
2052 	for (i = 0; i < DLM_HASH_BUCKETS; i++) {
2053 		bucket = dlm_lockres_hash(dlm, i);
2054 		hlist_for_each_entry(res, hash_iter, bucket, hash_node) {
2055 			if (res->state & DLM_LOCK_RES_RECOVERING) {
2056 				if (res->owner == dead_node) {
2057 					mlog(0, "(this=%u) res %.*s owner=%u "
2058 					     "was not on recovering list, but "
2059 					     "clearing state anyway\n",
2060 					     dlm->node_num, res->lockname.len,
2061 					     res->lockname.name, new_master);
2062 				} else if (res->owner == dlm->node_num) {
2063 					mlog(0, "(this=%u) res %.*s owner=%u "
2064 					     "was not on recovering list, "
2065 					     "owner is THIS node, clearing\n",
2066 					     dlm->node_num, res->lockname.len,
2067 					     res->lockname.name, new_master);
2068 				} else
2069 					continue;
2070 
2071 				if (!list_empty(&res->recovering)) {
2072 					mlog(0, "%s:%.*s: lockres was "
2073 					     "marked RECOVERING, owner=%u\n",
2074 					     dlm->name, res->lockname.len,
2075 					     res->lockname.name, res->owner);
2076 					list_del_init(&res->recovering);
2077 					dlm_lockres_put(res);
2078 				}
2079 				spin_lock(&res->spinlock);
2080 				/* new_master has our reference from
2081 				 * the lock state sent during recovery */
2082 				dlm_change_lockres_owner(dlm, res, new_master);
2083 				res->state &= ~DLM_LOCK_RES_RECOVERING;
2084 				if (__dlm_lockres_has_locks(res))
2085 					__dlm_dirty_lockres(dlm, res);
2086 				spin_unlock(&res->spinlock);
2087 				wake_up(&res->wq);
2088 			}
2089 		}
2090 	}
2091 }
2092 
2093 static inline int dlm_lvb_needs_invalidation(struct dlm_lock *lock, int local)
2094 {
2095 	if (local) {
2096 		if (lock->ml.type != LKM_EXMODE &&
2097 		    lock->ml.type != LKM_PRMODE)
2098 			return 1;
2099 	} else if (lock->ml.type == LKM_EXMODE)
2100 		return 1;
2101 	return 0;
2102 }
2103 
2104 static void dlm_revalidate_lvb(struct dlm_ctxt *dlm,
2105 			       struct dlm_lock_resource *res, u8 dead_node)
2106 {
2107 	struct list_head *queue;
2108 	struct dlm_lock *lock;
2109 	int blank_lvb = 0, local = 0;
2110 	int i;
2111 	u8 search_node;
2112 
2113 	assert_spin_locked(&dlm->spinlock);
2114 	assert_spin_locked(&res->spinlock);
2115 
2116 	if (res->owner == dlm->node_num)
2117 		/* if this node owned the lockres, and if the dead node
2118 		 * had an EX when he died, blank out the lvb */
2119 		search_node = dead_node;
2120 	else {
2121 		/* if this is a secondary lockres, and we had no EX or PR
2122 		 * locks granted, we can no longer trust the lvb */
2123 		search_node = dlm->node_num;
2124 		local = 1;  /* check local state for valid lvb */
2125 	}
2126 
2127 	for (i=DLM_GRANTED_LIST; i<=DLM_CONVERTING_LIST; i++) {
2128 		queue = dlm_list_idx_to_ptr(res, i);
2129 		list_for_each_entry(lock, queue, list) {
2130 			if (lock->ml.node == search_node) {
2131 				if (dlm_lvb_needs_invalidation(lock, local)) {
2132 					/* zero the lksb lvb and lockres lvb */
2133 					blank_lvb = 1;
2134 					memset(lock->lksb->lvb, 0, DLM_LVB_LEN);
2135 				}
2136 			}
2137 		}
2138 	}
2139 
2140 	if (blank_lvb) {
2141 		mlog(0, "clearing %.*s lvb, dead node %u had EX\n",
2142 		     res->lockname.len, res->lockname.name, dead_node);
2143 		memset(res->lvb, 0, DLM_LVB_LEN);
2144 	}
2145 }
2146 
2147 static void dlm_free_dead_locks(struct dlm_ctxt *dlm,
2148 				struct dlm_lock_resource *res, u8 dead_node)
2149 {
2150 	struct dlm_lock *lock, *next;
2151 	unsigned int freed = 0;
2152 
2153 	/* this node is the lockres master:
2154 	 * 1) remove any stale locks for the dead node
2155 	 * 2) if the dead node had an EX when he died, blank out the lvb
2156 	 */
2157 	assert_spin_locked(&dlm->spinlock);
2158 	assert_spin_locked(&res->spinlock);
2159 
2160 	/* We do two dlm_lock_put(). One for removing from list and the other is
2161 	 * to force the DLM_UNLOCK_FREE_LOCK action so as to free the locks */
2162 
2163 	/* TODO: check pending_asts, pending_basts here */
2164 	list_for_each_entry_safe(lock, next, &res->granted, list) {
2165 		if (lock->ml.node == dead_node) {
2166 			list_del_init(&lock->list);
2167 			dlm_lock_put(lock);
2168 			/* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */
2169 			dlm_lock_put(lock);
2170 			freed++;
2171 		}
2172 	}
2173 	list_for_each_entry_safe(lock, next, &res->converting, list) {
2174 		if (lock->ml.node == dead_node) {
2175 			list_del_init(&lock->list);
2176 			dlm_lock_put(lock);
2177 			/* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */
2178 			dlm_lock_put(lock);
2179 			freed++;
2180 		}
2181 	}
2182 	list_for_each_entry_safe(lock, next, &res->blocked, list) {
2183 		if (lock->ml.node == dead_node) {
2184 			list_del_init(&lock->list);
2185 			dlm_lock_put(lock);
2186 			/* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */
2187 			dlm_lock_put(lock);
2188 			freed++;
2189 		}
2190 	}
2191 
2192 	if (freed) {
2193 		mlog(0, "%s:%.*s: freed %u locks for dead node %u, "
2194 		     "dropping ref from lockres\n", dlm->name,
2195 		     res->lockname.len, res->lockname.name, freed, dead_node);
2196 		BUG_ON(!test_bit(dead_node, res->refmap));
2197 		dlm_lockres_clear_refmap_bit(dead_node, res);
2198 	} else if (test_bit(dead_node, res->refmap)) {
2199 		mlog(0, "%s:%.*s: dead node %u had a ref, but had "
2200 		     "no locks and had not purged before dying\n", dlm->name,
2201 		     res->lockname.len, res->lockname.name, dead_node);
2202 		dlm_lockres_clear_refmap_bit(dead_node, res);
2203 	}
2204 
2205 	/* do not kick thread yet */
2206 	__dlm_dirty_lockres(dlm, res);
2207 }
2208 
2209 /* if this node is the recovery master, and there are no
2210  * locks for a given lockres owned by this node that are in
2211  * either PR or EX mode, zero out the lvb before requesting.
2212  *
2213  */
2214 
2215 
2216 static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node)
2217 {
2218 	struct hlist_node *iter;
2219 	struct dlm_lock_resource *res;
2220 	int i;
2221 	struct hlist_head *bucket;
2222 	struct dlm_lock *lock;
2223 
2224 
2225 	/* purge any stale mles */
2226 	dlm_clean_master_list(dlm, dead_node);
2227 
2228 	/*
2229 	 * now clean up all lock resources.  there are two rules:
2230 	 *
2231 	 * 1) if the dead node was the master, move the lockres
2232 	 *    to the recovering list.  set the RECOVERING flag.
2233 	 *    this lockres needs to be cleaned up before it can
2234 	 *    be used further.
2235 	 *
2236 	 * 2) if this node was the master, remove all locks from
2237 	 *    each of the lockres queues that were owned by the
2238 	 *    dead node.  once recovery finishes, the dlm thread
2239 	 *    can be kicked again to see if any ASTs or BASTs
2240 	 *    need to be fired as a result.
2241 	 */
2242 	for (i = 0; i < DLM_HASH_BUCKETS; i++) {
2243 		bucket = dlm_lockres_hash(dlm, i);
2244 		hlist_for_each_entry(res, iter, bucket, hash_node) {
2245  			/* always prune any $RECOVERY entries for dead nodes,
2246  			 * otherwise hangs can occur during later recovery */
2247 			if (dlm_is_recovery_lock(res->lockname.name,
2248 						 res->lockname.len)) {
2249 				spin_lock(&res->spinlock);
2250 				list_for_each_entry(lock, &res->granted, list) {
2251 					if (lock->ml.node == dead_node) {
2252 						mlog(0, "AHA! there was "
2253 						     "a $RECOVERY lock for dead "
2254 						     "node %u (%s)!\n",
2255 						     dead_node, dlm->name);
2256 						list_del_init(&lock->list);
2257 						dlm_lock_put(lock);
2258 						break;
2259 					}
2260 				}
2261 				spin_unlock(&res->spinlock);
2262 				continue;
2263 			}
2264 			spin_lock(&res->spinlock);
2265 			/* zero the lvb if necessary */
2266 			dlm_revalidate_lvb(dlm, res, dead_node);
2267 			if (res->owner == dead_node) {
2268 				if (res->state & DLM_LOCK_RES_DROPPING_REF)
2269 					mlog(0, "%s:%.*s: owned by "
2270 					     "dead node %u, this node was "
2271 					     "dropping its ref when it died. "
2272 					     "continue, dropping the flag.\n",
2273 					     dlm->name, res->lockname.len,
2274 					     res->lockname.name, dead_node);
2275 
2276 				/* the wake_up for this will happen when the
2277 				 * RECOVERING flag is dropped later */
2278 				res->state &= ~DLM_LOCK_RES_DROPPING_REF;
2279 
2280 				dlm_move_lockres_to_recovery_list(dlm, res);
2281 			} else if (res->owner == dlm->node_num) {
2282 				dlm_free_dead_locks(dlm, res, dead_node);
2283 				__dlm_lockres_calc_usage(dlm, res);
2284 			}
2285 			spin_unlock(&res->spinlock);
2286 		}
2287 	}
2288 
2289 }
2290 
2291 static void __dlm_hb_node_down(struct dlm_ctxt *dlm, int idx)
2292 {
2293 	assert_spin_locked(&dlm->spinlock);
2294 
2295 	if (dlm->reco.new_master == idx) {
2296 		mlog(0, "%s: recovery master %d just died\n",
2297 		     dlm->name, idx);
2298 		if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2299 			/* finalize1 was reached, so it is safe to clear
2300 			 * the new_master and dead_node.  that recovery
2301 			 * is complete. */
2302 			mlog(0, "%s: dead master %d had reached "
2303 			     "finalize1 state, clearing\n", dlm->name, idx);
2304 			dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE;
2305 			__dlm_reset_recovery(dlm);
2306 		}
2307 	}
2308 
2309 	/* Clean up join state on node death. */
2310 	if (dlm->joining_node == idx) {
2311 		mlog(0, "Clearing join state for node %u\n", idx);
2312 		__dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
2313 	}
2314 
2315 	/* check to see if the node is already considered dead */
2316 	if (!test_bit(idx, dlm->live_nodes_map)) {
2317 		mlog(0, "for domain %s, node %d is already dead. "
2318 		     "another node likely did recovery already.\n",
2319 		     dlm->name, idx);
2320 		return;
2321 	}
2322 
2323 	/* check to see if we do not care about this node */
2324 	if (!test_bit(idx, dlm->domain_map)) {
2325 		/* This also catches the case that we get a node down
2326 		 * but haven't joined the domain yet. */
2327 		mlog(0, "node %u already removed from domain!\n", idx);
2328 		return;
2329 	}
2330 
2331 	clear_bit(idx, dlm->live_nodes_map);
2332 
2333 	/* make sure local cleanup occurs before the heartbeat events */
2334 	if (!test_bit(idx, dlm->recovery_map))
2335 		dlm_do_local_recovery_cleanup(dlm, idx);
2336 
2337 	/* notify anything attached to the heartbeat events */
2338 	dlm_hb_event_notify_attached(dlm, idx, 0);
2339 
2340 	mlog(0, "node %u being removed from domain map!\n", idx);
2341 	clear_bit(idx, dlm->domain_map);
2342 	/* wake up migration waiters if a node goes down.
2343 	 * perhaps later we can genericize this for other waiters. */
2344 	wake_up(&dlm->migration_wq);
2345 
2346 	if (test_bit(idx, dlm->recovery_map))
2347 		mlog(0, "domain %s, node %u already added "
2348 		     "to recovery map!\n", dlm->name, idx);
2349 	else
2350 		set_bit(idx, dlm->recovery_map);
2351 }
2352 
2353 void dlm_hb_node_down_cb(struct o2nm_node *node, int idx, void *data)
2354 {
2355 	struct dlm_ctxt *dlm = data;
2356 
2357 	if (!dlm_grab(dlm))
2358 		return;
2359 
2360 	/*
2361 	 * This will notify any dlm users that a node in our domain
2362 	 * went away without notifying us first.
2363 	 */
2364 	if (test_bit(idx, dlm->domain_map))
2365 		dlm_fire_domain_eviction_callbacks(dlm, idx);
2366 
2367 	spin_lock(&dlm->spinlock);
2368 	__dlm_hb_node_down(dlm, idx);
2369 	spin_unlock(&dlm->spinlock);
2370 
2371 	dlm_put(dlm);
2372 }
2373 
2374 void dlm_hb_node_up_cb(struct o2nm_node *node, int idx, void *data)
2375 {
2376 	struct dlm_ctxt *dlm = data;
2377 
2378 	if (!dlm_grab(dlm))
2379 		return;
2380 
2381 	spin_lock(&dlm->spinlock);
2382 	set_bit(idx, dlm->live_nodes_map);
2383 	/* do NOT notify mle attached to the heartbeat events.
2384 	 * new nodes are not interesting in mastery until joined. */
2385 	spin_unlock(&dlm->spinlock);
2386 
2387 	dlm_put(dlm);
2388 }
2389 
2390 static void dlm_reco_ast(void *astdata)
2391 {
2392 	struct dlm_ctxt *dlm = astdata;
2393 	mlog(0, "ast for recovery lock fired!, this=%u, dlm=%s\n",
2394 	     dlm->node_num, dlm->name);
2395 }
2396 static void dlm_reco_bast(void *astdata, int blocked_type)
2397 {
2398 	struct dlm_ctxt *dlm = astdata;
2399 	mlog(0, "bast for recovery lock fired!, this=%u, dlm=%s\n",
2400 	     dlm->node_num, dlm->name);
2401 }
2402 static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st)
2403 {
2404 	mlog(0, "unlockast for recovery lock fired!\n");
2405 }
2406 
2407 /*
2408  * dlm_pick_recovery_master will continually attempt to use
2409  * dlmlock() on the special "$RECOVERY" lockres with the
2410  * LKM_NOQUEUE flag to get an EX.  every thread that enters
2411  * this function on each node racing to become the recovery
2412  * master will not stop attempting this until either:
2413  * a) this node gets the EX (and becomes the recovery master),
2414  * or b) dlm->reco.new_master gets set to some nodenum
2415  * != O2NM_INVALID_NODE_NUM (another node will do the reco).
2416  * so each time a recovery master is needed, the entire cluster
2417  * will sync at this point.  if the new master dies, that will
2418  * be detected in dlm_do_recovery */
2419 static int dlm_pick_recovery_master(struct dlm_ctxt *dlm)
2420 {
2421 	enum dlm_status ret;
2422 	struct dlm_lockstatus lksb;
2423 	int status = -EINVAL;
2424 
2425 	mlog(0, "starting recovery of %s at %lu, dead=%u, this=%u\n",
2426 	     dlm->name, jiffies, dlm->reco.dead_node, dlm->node_num);
2427 again:
2428 	memset(&lksb, 0, sizeof(lksb));
2429 
2430 	ret = dlmlock(dlm, LKM_EXMODE, &lksb, LKM_NOQUEUE|LKM_RECOVERY,
2431 		      DLM_RECOVERY_LOCK_NAME, DLM_RECOVERY_LOCK_NAME_LEN,
2432 		      dlm_reco_ast, dlm, dlm_reco_bast);
2433 
2434 	mlog(0, "%s: dlmlock($RECOVERY) returned %d, lksb=%d\n",
2435 	     dlm->name, ret, lksb.status);
2436 
2437 	if (ret == DLM_NORMAL) {
2438 		mlog(0, "dlm=%s dlmlock says I got it (this=%u)\n",
2439 		     dlm->name, dlm->node_num);
2440 
2441 		/* got the EX lock.  check to see if another node
2442 		 * just became the reco master */
2443 		if (dlm_reco_master_ready(dlm)) {
2444 			mlog(0, "%s: got reco EX lock, but %u will "
2445 			     "do the recovery\n", dlm->name,
2446 			     dlm->reco.new_master);
2447 			status = -EEXIST;
2448 		} else {
2449 			status = 0;
2450 
2451 			/* see if recovery was already finished elsewhere */
2452 			spin_lock(&dlm->spinlock);
2453 			if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
2454 				status = -EINVAL;
2455 				mlog(0, "%s: got reco EX lock, but "
2456 				     "node got recovered already\n", dlm->name);
2457 				if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) {
2458 					mlog(ML_ERROR, "%s: new master is %u "
2459 					     "but no dead node!\n",
2460 					     dlm->name, dlm->reco.new_master);
2461 					BUG();
2462 				}
2463 			}
2464 			spin_unlock(&dlm->spinlock);
2465 		}
2466 
2467 		/* if this node has actually become the recovery master,
2468 		 * set the master and send the messages to begin recovery */
2469 		if (!status) {
2470 			mlog(0, "%s: dead=%u, this=%u, sending "
2471 			     "begin_reco now\n", dlm->name,
2472 			     dlm->reco.dead_node, dlm->node_num);
2473 			status = dlm_send_begin_reco_message(dlm,
2474 				      dlm->reco.dead_node);
2475 			/* this always succeeds */
2476 			BUG_ON(status);
2477 
2478 			/* set the new_master to this node */
2479 			spin_lock(&dlm->spinlock);
2480 			dlm_set_reco_master(dlm, dlm->node_num);
2481 			spin_unlock(&dlm->spinlock);
2482 		}
2483 
2484 		/* recovery lock is a special case.  ast will not get fired,
2485 		 * so just go ahead and unlock it. */
2486 		ret = dlmunlock(dlm, &lksb, 0, dlm_reco_unlock_ast, dlm);
2487 		if (ret == DLM_DENIED) {
2488 			mlog(0, "got DLM_DENIED, trying LKM_CANCEL\n");
2489 			ret = dlmunlock(dlm, &lksb, LKM_CANCEL, dlm_reco_unlock_ast, dlm);
2490 		}
2491 		if (ret != DLM_NORMAL) {
2492 			/* this would really suck. this could only happen
2493 			 * if there was a network error during the unlock
2494 			 * because of node death.  this means the unlock
2495 			 * is actually "done" and the lock structure is
2496 			 * even freed.  we can continue, but only
2497 			 * because this specific lock name is special. */
2498 			mlog(ML_ERROR, "dlmunlock returned %d\n", ret);
2499 		}
2500 	} else if (ret == DLM_NOTQUEUED) {
2501 		mlog(0, "dlm=%s dlmlock says another node got it (this=%u)\n",
2502 		     dlm->name, dlm->node_num);
2503 		/* another node is master. wait on
2504 		 * reco.new_master != O2NM_INVALID_NODE_NUM
2505 		 * for at most one second */
2506 		wait_event_timeout(dlm->dlm_reco_thread_wq,
2507 					 dlm_reco_master_ready(dlm),
2508 					 msecs_to_jiffies(1000));
2509 		if (!dlm_reco_master_ready(dlm)) {
2510 			mlog(0, "%s: reco master taking awhile\n",
2511 			     dlm->name);
2512 			goto again;
2513 		}
2514 		/* another node has informed this one that it is reco master */
2515 		mlog(0, "%s: reco master %u is ready to recover %u\n",
2516 		     dlm->name, dlm->reco.new_master, dlm->reco.dead_node);
2517 		status = -EEXIST;
2518 	} else if (ret == DLM_RECOVERING) {
2519 		mlog(0, "dlm=%s dlmlock says master node died (this=%u)\n",
2520 		     dlm->name, dlm->node_num);
2521 		goto again;
2522 	} else {
2523 		struct dlm_lock_resource *res;
2524 
2525 		/* dlmlock returned something other than NOTQUEUED or NORMAL */
2526 		mlog(ML_ERROR, "%s: got %s from dlmlock($RECOVERY), "
2527 		     "lksb.status=%s\n", dlm->name, dlm_errname(ret),
2528 		     dlm_errname(lksb.status));
2529 		res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME,
2530 					 DLM_RECOVERY_LOCK_NAME_LEN);
2531 		if (res) {
2532 			dlm_print_one_lock_resource(res);
2533 			dlm_lockres_put(res);
2534 		} else {
2535 			mlog(ML_ERROR, "recovery lock not found\n");
2536 		}
2537 		BUG();
2538 	}
2539 
2540 	return status;
2541 }
2542 
2543 static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node)
2544 {
2545 	struct dlm_begin_reco br;
2546 	int ret = 0;
2547 	struct dlm_node_iter iter;
2548 	int nodenum;
2549 	int status;
2550 
2551 	mlog_entry("%u\n", dead_node);
2552 
2553 	mlog(0, "%s: dead node is %u\n", dlm->name, dead_node);
2554 
2555 	spin_lock(&dlm->spinlock);
2556 	dlm_node_iter_init(dlm->domain_map, &iter);
2557 	spin_unlock(&dlm->spinlock);
2558 
2559 	clear_bit(dead_node, iter.node_map);
2560 
2561 	memset(&br, 0, sizeof(br));
2562 	br.node_idx = dlm->node_num;
2563 	br.dead_node = dead_node;
2564 
2565 	while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2566 		ret = 0;
2567 		if (nodenum == dead_node) {
2568 			mlog(0, "not sending begin reco to dead node "
2569 				  "%u\n", dead_node);
2570 			continue;
2571 		}
2572 		if (nodenum == dlm->node_num) {
2573 			mlog(0, "not sending begin reco to self\n");
2574 			continue;
2575 		}
2576 retry:
2577 		ret = -EINVAL;
2578 		mlog(0, "attempting to send begin reco msg to %d\n",
2579 			  nodenum);
2580 		ret = o2net_send_message(DLM_BEGIN_RECO_MSG, dlm->key,
2581 					 &br, sizeof(br), nodenum, &status);
2582 		/* negative status is handled ok by caller here */
2583 		if (ret >= 0)
2584 			ret = status;
2585 		if (dlm_is_host_down(ret)) {
2586 			/* node is down.  not involved in recovery
2587 			 * so just keep going */
2588 			mlog(0, "%s: node %u was down when sending "
2589 			     "begin reco msg (%d)\n", dlm->name, nodenum, ret);
2590 			ret = 0;
2591 		}
2592 		if (ret == -EAGAIN) {
2593 			mlog(0, "%s: trying to start recovery of node "
2594 			     "%u, but node %u is waiting for last recovery "
2595 			     "to complete, backoff for a bit\n", dlm->name,
2596 			     dead_node, nodenum);
2597 			msleep(100);
2598 			goto retry;
2599 		}
2600 		if (ret < 0) {
2601 			struct dlm_lock_resource *res;
2602 			/* this is now a serious problem, possibly ENOMEM
2603 			 * in the network stack.  must retry */
2604 			mlog_errno(ret);
2605 			mlog(ML_ERROR, "begin reco of dlm %s to node %u "
2606 			    " returned %d\n", dlm->name, nodenum, ret);
2607 			res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME,
2608 						 DLM_RECOVERY_LOCK_NAME_LEN);
2609 			if (res) {
2610 				dlm_print_one_lock_resource(res);
2611 				dlm_lockres_put(res);
2612 			} else {
2613 				mlog(ML_ERROR, "recovery lock not found\n");
2614 			}
2615 			/* sleep for a bit in hopes that we can avoid
2616 			 * another ENOMEM */
2617 			msleep(100);
2618 			goto retry;
2619 		}
2620 	}
2621 
2622 	return ret;
2623 }
2624 
2625 int dlm_begin_reco_handler(struct o2net_msg *msg, u32 len, void *data,
2626 			   void **ret_data)
2627 {
2628 	struct dlm_ctxt *dlm = data;
2629 	struct dlm_begin_reco *br = (struct dlm_begin_reco *)msg->buf;
2630 
2631 	/* ok to return 0, domain has gone away */
2632 	if (!dlm_grab(dlm))
2633 		return 0;
2634 
2635 	spin_lock(&dlm->spinlock);
2636 	if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2637 		mlog(0, "%s: node %u wants to recover node %u (%u:%u) "
2638 		     "but this node is in finalize state, waiting on finalize2\n",
2639 		     dlm->name, br->node_idx, br->dead_node,
2640 		     dlm->reco.dead_node, dlm->reco.new_master);
2641 		spin_unlock(&dlm->spinlock);
2642 		return -EAGAIN;
2643 	}
2644 	spin_unlock(&dlm->spinlock);
2645 
2646 	mlog(0, "%s: node %u wants to recover node %u (%u:%u)\n",
2647 	     dlm->name, br->node_idx, br->dead_node,
2648 	     dlm->reco.dead_node, dlm->reco.new_master);
2649 
2650 	dlm_fire_domain_eviction_callbacks(dlm, br->dead_node);
2651 
2652 	spin_lock(&dlm->spinlock);
2653 	if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) {
2654 		if (test_bit(dlm->reco.new_master, dlm->recovery_map)) {
2655 			mlog(0, "%s: new_master %u died, changing "
2656 			     "to %u\n", dlm->name, dlm->reco.new_master,
2657 			     br->node_idx);
2658 		} else {
2659 			mlog(0, "%s: new_master %u NOT DEAD, changing "
2660 			     "to %u\n", dlm->name, dlm->reco.new_master,
2661 			     br->node_idx);
2662 			/* may not have seen the new master as dead yet */
2663 		}
2664 	}
2665 	if (dlm->reco.dead_node != O2NM_INVALID_NODE_NUM) {
2666 		mlog(ML_NOTICE, "%s: dead_node previously set to %u, "
2667 		     "node %u changing it to %u\n", dlm->name,
2668 		     dlm->reco.dead_node, br->node_idx, br->dead_node);
2669 	}
2670 	dlm_set_reco_master(dlm, br->node_idx);
2671 	dlm_set_reco_dead_node(dlm, br->dead_node);
2672 	if (!test_bit(br->dead_node, dlm->recovery_map)) {
2673 		mlog(0, "recovery master %u sees %u as dead, but this "
2674 		     "node has not yet.  marking %u as dead\n",
2675 		     br->node_idx, br->dead_node, br->dead_node);
2676 		if (!test_bit(br->dead_node, dlm->domain_map) ||
2677 		    !test_bit(br->dead_node, dlm->live_nodes_map))
2678 			mlog(0, "%u not in domain/live_nodes map "
2679 			     "so setting it in reco map manually\n",
2680 			     br->dead_node);
2681 		/* force the recovery cleanup in __dlm_hb_node_down
2682 		 * both of these will be cleared in a moment */
2683 		set_bit(br->dead_node, dlm->domain_map);
2684 		set_bit(br->dead_node, dlm->live_nodes_map);
2685 		__dlm_hb_node_down(dlm, br->dead_node);
2686 	}
2687 	spin_unlock(&dlm->spinlock);
2688 
2689 	dlm_kick_recovery_thread(dlm);
2690 
2691 	mlog(0, "%s: recovery started by node %u, for %u (%u:%u)\n",
2692 	     dlm->name, br->node_idx, br->dead_node,
2693 	     dlm->reco.dead_node, dlm->reco.new_master);
2694 
2695 	dlm_put(dlm);
2696 	return 0;
2697 }
2698 
2699 #define DLM_FINALIZE_STAGE2  0x01
2700 static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm)
2701 {
2702 	int ret = 0;
2703 	struct dlm_finalize_reco fr;
2704 	struct dlm_node_iter iter;
2705 	int nodenum;
2706 	int status;
2707 	int stage = 1;
2708 
2709 	mlog(0, "finishing recovery for node %s:%u, "
2710 	     "stage %d\n", dlm->name, dlm->reco.dead_node, stage);
2711 
2712 	spin_lock(&dlm->spinlock);
2713 	dlm_node_iter_init(dlm->domain_map, &iter);
2714 	spin_unlock(&dlm->spinlock);
2715 
2716 stage2:
2717 	memset(&fr, 0, sizeof(fr));
2718 	fr.node_idx = dlm->node_num;
2719 	fr.dead_node = dlm->reco.dead_node;
2720 	if (stage == 2)
2721 		fr.flags |= DLM_FINALIZE_STAGE2;
2722 
2723 	while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2724 		if (nodenum == dlm->node_num)
2725 			continue;
2726 		ret = o2net_send_message(DLM_FINALIZE_RECO_MSG, dlm->key,
2727 					 &fr, sizeof(fr), nodenum, &status);
2728 		if (ret >= 0)
2729 			ret = status;
2730 		if (ret < 0) {
2731 			mlog_errno(ret);
2732 			if (dlm_is_host_down(ret)) {
2733 				/* this has no effect on this recovery
2734 				 * session, so set the status to zero to
2735 				 * finish out the last recovery */
2736 				mlog(ML_ERROR, "node %u went down after this "
2737 				     "node finished recovery.\n", nodenum);
2738 				ret = 0;
2739 				continue;
2740 			}
2741 			break;
2742 		}
2743 	}
2744 	if (stage == 1) {
2745 		/* reset the node_iter back to the top and send finalize2 */
2746 		iter.curnode = -1;
2747 		stage = 2;
2748 		goto stage2;
2749 	}
2750 
2751 	return ret;
2752 }
2753 
2754 int dlm_finalize_reco_handler(struct o2net_msg *msg, u32 len, void *data,
2755 			      void **ret_data)
2756 {
2757 	struct dlm_ctxt *dlm = data;
2758 	struct dlm_finalize_reco *fr = (struct dlm_finalize_reco *)msg->buf;
2759 	int stage = 1;
2760 
2761 	/* ok to return 0, domain has gone away */
2762 	if (!dlm_grab(dlm))
2763 		return 0;
2764 
2765 	if (fr->flags & DLM_FINALIZE_STAGE2)
2766 		stage = 2;
2767 
2768 	mlog(0, "%s: node %u finalizing recovery stage%d of "
2769 	     "node %u (%u:%u)\n", dlm->name, fr->node_idx, stage,
2770 	     fr->dead_node, dlm->reco.dead_node, dlm->reco.new_master);
2771 
2772 	spin_lock(&dlm->spinlock);
2773 
2774 	if (dlm->reco.new_master != fr->node_idx) {
2775 		mlog(ML_ERROR, "node %u sent recovery finalize msg, but node "
2776 		     "%u is supposed to be the new master, dead=%u\n",
2777 		     fr->node_idx, dlm->reco.new_master, fr->dead_node);
2778 		BUG();
2779 	}
2780 	if (dlm->reco.dead_node != fr->dead_node) {
2781 		mlog(ML_ERROR, "node %u sent recovery finalize msg for dead "
2782 		     "node %u, but node %u is supposed to be dead\n",
2783 		     fr->node_idx, fr->dead_node, dlm->reco.dead_node);
2784 		BUG();
2785 	}
2786 
2787 	switch (stage) {
2788 		case 1:
2789 			dlm_finish_local_lockres_recovery(dlm, fr->dead_node, fr->node_idx);
2790 			if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2791 				mlog(ML_ERROR, "%s: received finalize1 from "
2792 				     "new master %u for dead node %u, but "
2793 				     "this node has already received it!\n",
2794 				     dlm->name, fr->node_idx, fr->dead_node);
2795 				dlm_print_reco_node_status(dlm);
2796 				BUG();
2797 			}
2798 			dlm->reco.state |= DLM_RECO_STATE_FINALIZE;
2799 			spin_unlock(&dlm->spinlock);
2800 			break;
2801 		case 2:
2802 			if (!(dlm->reco.state & DLM_RECO_STATE_FINALIZE)) {
2803 				mlog(ML_ERROR, "%s: received finalize2 from "
2804 				     "new master %u for dead node %u, but "
2805 				     "this node did not have finalize1!\n",
2806 				     dlm->name, fr->node_idx, fr->dead_node);
2807 				dlm_print_reco_node_status(dlm);
2808 				BUG();
2809 			}
2810 			dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE;
2811 			spin_unlock(&dlm->spinlock);
2812 			dlm_reset_recovery(dlm);
2813 			dlm_kick_recovery_thread(dlm);
2814 			break;
2815 		default:
2816 			BUG();
2817 	}
2818 
2819 	mlog(0, "%s: recovery done, reco master was %u, dead now %u, master now %u\n",
2820 	     dlm->name, fr->node_idx, dlm->reco.dead_node, dlm->reco.new_master);
2821 
2822 	dlm_put(dlm);
2823 	return 0;
2824 }
2825