xref: /linux/fs/ocfs2/dlm/dlmrecovery.c (revision c537b994505099b7197e7d3125b942ecbcc51eb6)
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/utsname.h>
34 #include <linux/init.h>
35 #include <linux/sysctl.h>
36 #include <linux/random.h>
37 #include <linux/blkdev.h>
38 #include <linux/socket.h>
39 #include <linux/inet.h>
40 #include <linux/timer.h>
41 #include <linux/kthread.h>
42 #include <linux/delay.h>
43 
44 
45 #include "cluster/heartbeat.h"
46 #include "cluster/nodemanager.h"
47 #include "cluster/tcp.h"
48 
49 #include "dlmapi.h"
50 #include "dlmcommon.h"
51 #include "dlmdomain.h"
52 
53 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_RECOVERY)
54 #include "cluster/masklog.h"
55 
56 static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node);
57 
58 static int dlm_recovery_thread(void *data);
59 void dlm_complete_recovery_thread(struct dlm_ctxt *dlm);
60 int dlm_launch_recovery_thread(struct dlm_ctxt *dlm);
61 void dlm_kick_recovery_thread(struct dlm_ctxt *dlm);
62 static int dlm_do_recovery(struct dlm_ctxt *dlm);
63 
64 static int dlm_pick_recovery_master(struct dlm_ctxt *dlm);
65 static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node);
66 static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node);
67 static int dlm_request_all_locks(struct dlm_ctxt *dlm,
68 				 u8 request_from, u8 dead_node);
69 static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm, u8 dead_node);
70 
71 static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res);
72 static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres,
73 					const char *lockname, int namelen,
74 					int total_locks, u64 cookie,
75 					u8 flags, u8 master);
76 static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm,
77 				    struct dlm_migratable_lockres *mres,
78 				    u8 send_to,
79 				    struct dlm_lock_resource *res,
80 				    int total_locks);
81 static int dlm_process_recovery_data(struct dlm_ctxt *dlm,
82 				     struct dlm_lock_resource *res,
83 				     struct dlm_migratable_lockres *mres);
84 static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm);
85 static int dlm_send_all_done_msg(struct dlm_ctxt *dlm,
86 				 u8 dead_node, u8 send_to);
87 static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node);
88 static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm,
89 					struct list_head *list, u8 dead_node);
90 static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm,
91 					      u8 dead_node, u8 new_master);
92 static void dlm_reco_ast(void *astdata);
93 static void dlm_reco_bast(void *astdata, int blocked_type);
94 static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st);
95 static void dlm_request_all_locks_worker(struct dlm_work_item *item,
96 					 void *data);
97 static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data);
98 static int dlm_lockres_master_requery(struct dlm_ctxt *dlm,
99 				      struct dlm_lock_resource *res,
100 				      u8 *real_master);
101 
102 static u64 dlm_get_next_mig_cookie(void);
103 
104 static DEFINE_SPINLOCK(dlm_reco_state_lock);
105 static DEFINE_SPINLOCK(dlm_mig_cookie_lock);
106 static u64 dlm_mig_cookie = 1;
107 
108 static u64 dlm_get_next_mig_cookie(void)
109 {
110 	u64 c;
111 	spin_lock(&dlm_mig_cookie_lock);
112 	c = dlm_mig_cookie;
113 	if (dlm_mig_cookie == (~0ULL))
114 		dlm_mig_cookie = 1;
115 	else
116 		dlm_mig_cookie++;
117 	spin_unlock(&dlm_mig_cookie_lock);
118 	return c;
119 }
120 
121 static inline void dlm_set_reco_dead_node(struct dlm_ctxt *dlm,
122 					  u8 dead_node)
123 {
124 	assert_spin_locked(&dlm->spinlock);
125 	if (dlm->reco.dead_node != dead_node)
126 		mlog(0, "%s: changing dead_node from %u to %u\n",
127 		     dlm->name, dlm->reco.dead_node, dead_node);
128 	dlm->reco.dead_node = dead_node;
129 }
130 
131 static inline void dlm_set_reco_master(struct dlm_ctxt *dlm,
132 				       u8 master)
133 {
134 	assert_spin_locked(&dlm->spinlock);
135 	mlog(0, "%s: changing new_master from %u to %u\n",
136 	     dlm->name, dlm->reco.new_master, master);
137 	dlm->reco.new_master = master;
138 }
139 
140 static inline void __dlm_reset_recovery(struct dlm_ctxt *dlm)
141 {
142 	assert_spin_locked(&dlm->spinlock);
143 	clear_bit(dlm->reco.dead_node, dlm->recovery_map);
144 	dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
145 	dlm_set_reco_master(dlm, O2NM_INVALID_NODE_NUM);
146 }
147 
148 static inline void dlm_reset_recovery(struct dlm_ctxt *dlm)
149 {
150 	spin_lock(&dlm->spinlock);
151 	__dlm_reset_recovery(dlm);
152 	spin_unlock(&dlm->spinlock);
153 }
154 
155 /* Worker function used during recovery. */
156 void dlm_dispatch_work(struct work_struct *work)
157 {
158 	struct dlm_ctxt *dlm =
159 		container_of(work, struct dlm_ctxt, dispatched_work);
160 	LIST_HEAD(tmp_list);
161 	struct list_head *iter, *iter2;
162 	struct dlm_work_item *item;
163 	dlm_workfunc_t *workfunc;
164 	int tot=0;
165 
166 	spin_lock(&dlm->work_lock);
167 	list_splice_init(&dlm->work_list, &tmp_list);
168 	spin_unlock(&dlm->work_lock);
169 
170 	list_for_each_safe(iter, iter2, &tmp_list) {
171 		tot++;
172 	}
173 	mlog(0, "%s: work thread has %d work items\n", dlm->name, tot);
174 
175 	list_for_each_safe(iter, iter2, &tmp_list) {
176 		item = list_entry(iter, struct dlm_work_item, list);
177 		workfunc = item->func;
178 		list_del_init(&item->list);
179 
180 		/* already have ref on dlm to avoid having
181 		 * it disappear.  just double-check. */
182 		BUG_ON(item->dlm != dlm);
183 
184 		/* this is allowed to sleep and
185 		 * call network stuff */
186 		workfunc(item, item->data);
187 
188 		dlm_put(dlm);
189 		kfree(item);
190 	}
191 }
192 
193 /*
194  * RECOVERY THREAD
195  */
196 
197 void dlm_kick_recovery_thread(struct dlm_ctxt *dlm)
198 {
199 	/* wake the recovery thread
200 	 * this will wake the reco thread in one of three places
201 	 * 1) sleeping with no recovery happening
202 	 * 2) sleeping with recovery mastered elsewhere
203 	 * 3) recovery mastered here, waiting on reco data */
204 
205 	wake_up(&dlm->dlm_reco_thread_wq);
206 }
207 
208 /* Launch the recovery thread */
209 int dlm_launch_recovery_thread(struct dlm_ctxt *dlm)
210 {
211 	mlog(0, "starting dlm recovery thread...\n");
212 
213 	dlm->dlm_reco_thread_task = kthread_run(dlm_recovery_thread, dlm,
214 						"dlm_reco_thread");
215 	if (IS_ERR(dlm->dlm_reco_thread_task)) {
216 		mlog_errno(PTR_ERR(dlm->dlm_reco_thread_task));
217 		dlm->dlm_reco_thread_task = NULL;
218 		return -EINVAL;
219 	}
220 
221 	return 0;
222 }
223 
224 void dlm_complete_recovery_thread(struct dlm_ctxt *dlm)
225 {
226 	if (dlm->dlm_reco_thread_task) {
227 		mlog(0, "waiting for dlm recovery thread to exit\n");
228 		kthread_stop(dlm->dlm_reco_thread_task);
229 		dlm->dlm_reco_thread_task = NULL;
230 	}
231 }
232 
233 
234 
235 /*
236  * this is lame, but here's how recovery works...
237  * 1) all recovery threads cluster wide will work on recovering
238  *    ONE node at a time
239  * 2) negotiate who will take over all the locks for the dead node.
240  *    thats right... ALL the locks.
241  * 3) once a new master is chosen, everyone scans all locks
242  *    and moves aside those mastered by the dead guy
243  * 4) each of these locks should be locked until recovery is done
244  * 5) the new master collects up all of secondary lock queue info
245  *    one lock at a time, forcing each node to communicate back
246  *    before continuing
247  * 6) each secondary lock queue responds with the full known lock info
248  * 7) once the new master has run all its locks, it sends a ALLDONE!
249  *    message to everyone
250  * 8) upon receiving this message, the secondary queue node unlocks
251  *    and responds to the ALLDONE
252  * 9) once the new master gets responses from everyone, he unlocks
253  *    everything and recovery for this dead node is done
254  *10) go back to 2) while there are still dead nodes
255  *
256  */
257 
258 static void dlm_print_reco_node_status(struct dlm_ctxt *dlm)
259 {
260 	struct dlm_reco_node_data *ndata;
261 	struct dlm_lock_resource *res;
262 
263 	mlog(ML_NOTICE, "%s(%d): recovery info, state=%s, dead=%u, master=%u\n",
264 	     dlm->name, dlm->dlm_reco_thread_task->pid,
265 	     dlm->reco.state & DLM_RECO_STATE_ACTIVE ? "ACTIVE" : "inactive",
266 	     dlm->reco.dead_node, dlm->reco.new_master);
267 
268 	list_for_each_entry(ndata, &dlm->reco.node_data, list) {
269 		char *st = "unknown";
270 		switch (ndata->state) {
271 			case DLM_RECO_NODE_DATA_INIT:
272 				st = "init";
273 				break;
274 			case DLM_RECO_NODE_DATA_REQUESTING:
275 				st = "requesting";
276 				break;
277 			case DLM_RECO_NODE_DATA_DEAD:
278 				st = "dead";
279 				break;
280 			case DLM_RECO_NODE_DATA_RECEIVING:
281 				st = "receiving";
282 				break;
283 			case DLM_RECO_NODE_DATA_REQUESTED:
284 				st = "requested";
285 				break;
286 			case DLM_RECO_NODE_DATA_DONE:
287 				st = "done";
288 				break;
289 			case DLM_RECO_NODE_DATA_FINALIZE_SENT:
290 				st = "finalize-sent";
291 				break;
292 			default:
293 				st = "bad";
294 				break;
295 		}
296 		mlog(ML_NOTICE, "%s: reco state, node %u, state=%s\n",
297 		     dlm->name, ndata->node_num, st);
298 	}
299 	list_for_each_entry(res, &dlm->reco.resources, recovering) {
300 		mlog(ML_NOTICE, "%s: lockres %.*s on recovering list\n",
301 		     dlm->name, res->lockname.len, res->lockname.name);
302 	}
303 }
304 
305 #define DLM_RECO_THREAD_TIMEOUT_MS (5 * 1000)
306 
307 static int dlm_recovery_thread(void *data)
308 {
309 	int status;
310 	struct dlm_ctxt *dlm = data;
311 	unsigned long timeout = msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS);
312 
313 	mlog(0, "dlm thread running for %s...\n", dlm->name);
314 
315 	while (!kthread_should_stop()) {
316 		if (dlm_joined(dlm)) {
317 			status = dlm_do_recovery(dlm);
318 			if (status == -EAGAIN) {
319 				/* do not sleep, recheck immediately. */
320 				continue;
321 			}
322 			if (status < 0)
323 				mlog_errno(status);
324 		}
325 
326 		wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq,
327 						 kthread_should_stop(),
328 						 timeout);
329 	}
330 
331 	mlog(0, "quitting DLM recovery thread\n");
332 	return 0;
333 }
334 
335 /* returns true when the recovery master has contacted us */
336 static int dlm_reco_master_ready(struct dlm_ctxt *dlm)
337 {
338 	int ready;
339 	spin_lock(&dlm->spinlock);
340 	ready = (dlm->reco.new_master != O2NM_INVALID_NODE_NUM);
341 	spin_unlock(&dlm->spinlock);
342 	return ready;
343 }
344 
345 /* returns true if node is no longer in the domain
346  * could be dead or just not joined */
347 int dlm_is_node_dead(struct dlm_ctxt *dlm, u8 node)
348 {
349 	int dead;
350 	spin_lock(&dlm->spinlock);
351 	dead = !test_bit(node, dlm->domain_map);
352 	spin_unlock(&dlm->spinlock);
353 	return dead;
354 }
355 
356 /* returns true if node is no longer in the domain
357  * could be dead or just not joined */
358 static int dlm_is_node_recovered(struct dlm_ctxt *dlm, u8 node)
359 {
360 	int recovered;
361 	spin_lock(&dlm->spinlock);
362 	recovered = !test_bit(node, dlm->recovery_map);
363 	spin_unlock(&dlm->spinlock);
364 	return recovered;
365 }
366 
367 
368 int dlm_wait_for_node_death(struct dlm_ctxt *dlm, u8 node, int timeout)
369 {
370 	if (timeout) {
371 		mlog(ML_NOTICE, "%s: waiting %dms for notification of "
372 		     "death of node %u\n", dlm->name, timeout, node);
373 		wait_event_timeout(dlm->dlm_reco_thread_wq,
374 			   dlm_is_node_dead(dlm, node),
375 			   msecs_to_jiffies(timeout));
376 	} else {
377 		mlog(ML_NOTICE, "%s: waiting indefinitely for notification "
378 		     "of death of node %u\n", dlm->name, node);
379 		wait_event(dlm->dlm_reco_thread_wq,
380 			   dlm_is_node_dead(dlm, node));
381 	}
382 	/* for now, return 0 */
383 	return 0;
384 }
385 
386 int dlm_wait_for_node_recovery(struct dlm_ctxt *dlm, u8 node, int timeout)
387 {
388 	if (timeout) {
389 		mlog(0, "%s: waiting %dms for notification of "
390 		     "recovery of node %u\n", dlm->name, timeout, node);
391 		wait_event_timeout(dlm->dlm_reco_thread_wq,
392 			   dlm_is_node_recovered(dlm, node),
393 			   msecs_to_jiffies(timeout));
394 	} else {
395 		mlog(0, "%s: waiting indefinitely for notification "
396 		     "of recovery of node %u\n", dlm->name, node);
397 		wait_event(dlm->dlm_reco_thread_wq,
398 			   dlm_is_node_recovered(dlm, node));
399 	}
400 	/* for now, return 0 */
401 	return 0;
402 }
403 
404 /* callers of the top-level api calls (dlmlock/dlmunlock) should
405  * block on the dlm->reco.event when recovery is in progress.
406  * the dlm recovery thread will set this state when it begins
407  * recovering a dead node (as the new master or not) and clear
408  * the state and wake as soon as all affected lock resources have
409  * been marked with the RECOVERY flag */
410 static int dlm_in_recovery(struct dlm_ctxt *dlm)
411 {
412 	int in_recovery;
413 	spin_lock(&dlm->spinlock);
414 	in_recovery = !!(dlm->reco.state & DLM_RECO_STATE_ACTIVE);
415 	spin_unlock(&dlm->spinlock);
416 	return in_recovery;
417 }
418 
419 
420 void dlm_wait_for_recovery(struct dlm_ctxt *dlm)
421 {
422 	if (dlm_in_recovery(dlm)) {
423 		mlog(0, "%s: reco thread %d in recovery: "
424 		     "state=%d, master=%u, dead=%u\n",
425 		     dlm->name, dlm->dlm_reco_thread_task->pid,
426 		     dlm->reco.state, dlm->reco.new_master,
427 		     dlm->reco.dead_node);
428 	}
429 	wait_event(dlm->reco.event, !dlm_in_recovery(dlm));
430 }
431 
432 static void dlm_begin_recovery(struct dlm_ctxt *dlm)
433 {
434 	spin_lock(&dlm->spinlock);
435 	BUG_ON(dlm->reco.state & DLM_RECO_STATE_ACTIVE);
436 	dlm->reco.state |= DLM_RECO_STATE_ACTIVE;
437 	spin_unlock(&dlm->spinlock);
438 }
439 
440 static void dlm_end_recovery(struct dlm_ctxt *dlm)
441 {
442 	spin_lock(&dlm->spinlock);
443 	BUG_ON(!(dlm->reco.state & DLM_RECO_STATE_ACTIVE));
444 	dlm->reco.state &= ~DLM_RECO_STATE_ACTIVE;
445 	spin_unlock(&dlm->spinlock);
446 	wake_up(&dlm->reco.event);
447 }
448 
449 static int dlm_do_recovery(struct dlm_ctxt *dlm)
450 {
451 	int status = 0;
452 	int ret;
453 
454 	spin_lock(&dlm->spinlock);
455 
456 	/* check to see if the new master has died */
457 	if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM &&
458 	    test_bit(dlm->reco.new_master, dlm->recovery_map)) {
459 		mlog(0, "new master %u died while recovering %u!\n",
460 		     dlm->reco.new_master, dlm->reco.dead_node);
461 		/* unset the new_master, leave dead_node */
462 		dlm_set_reco_master(dlm, O2NM_INVALID_NODE_NUM);
463 	}
464 
465 	/* select a target to recover */
466 	if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
467 		int bit;
468 
469 		bit = find_next_bit (dlm->recovery_map, O2NM_MAX_NODES+1, 0);
470 		if (bit >= O2NM_MAX_NODES || bit < 0)
471 			dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
472 		else
473 			dlm_set_reco_dead_node(dlm, bit);
474 	} else if (!test_bit(dlm->reco.dead_node, dlm->recovery_map)) {
475 		/* BUG? */
476 		mlog(ML_ERROR, "dead_node %u no longer in recovery map!\n",
477 		     dlm->reco.dead_node);
478 		dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
479 	}
480 
481 	if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
482 		// mlog(0, "nothing to recover!  sleeping now!\n");
483 		spin_unlock(&dlm->spinlock);
484 		/* return to main thread loop and sleep. */
485 		return 0;
486 	}
487 	mlog(0, "%s(%d):recovery thread found node %u in the recovery map!\n",
488 	     dlm->name, dlm->dlm_reco_thread_task->pid,
489 	     dlm->reco.dead_node);
490 	spin_unlock(&dlm->spinlock);
491 
492 	/* take write barrier */
493 	/* (stops the list reshuffling thread, proxy ast handling) */
494 	dlm_begin_recovery(dlm);
495 
496 	if (dlm->reco.new_master == dlm->node_num)
497 		goto master_here;
498 
499 	if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) {
500 		/* choose a new master, returns 0 if this node
501 		 * is the master, -EEXIST if it's another node.
502 		 * this does not return until a new master is chosen
503 		 * or recovery completes entirely. */
504 		ret = dlm_pick_recovery_master(dlm);
505 		if (!ret) {
506 			/* already notified everyone.  go. */
507 			goto master_here;
508 		}
509 		mlog(0, "another node will master this recovery session.\n");
510 	}
511 	mlog(0, "dlm=%s (%d), new_master=%u, this node=%u, dead_node=%u\n",
512 	     dlm->name, dlm->dlm_reco_thread_task->pid, dlm->reco.new_master,
513 	     dlm->node_num, dlm->reco.dead_node);
514 
515 	/* it is safe to start everything back up here
516 	 * because all of the dead node's lock resources
517 	 * have been marked as in-recovery */
518 	dlm_end_recovery(dlm);
519 
520 	/* sleep out in main dlm_recovery_thread loop. */
521 	return 0;
522 
523 master_here:
524 	mlog(0, "(%d) mastering recovery of %s:%u here(this=%u)!\n",
525 	     dlm->dlm_reco_thread_task->pid,
526 	     dlm->name, dlm->reco.dead_node, dlm->node_num);
527 
528 	status = dlm_remaster_locks(dlm, dlm->reco.dead_node);
529 	if (status < 0) {
530 		/* we should never hit this anymore */
531 		mlog(ML_ERROR, "error %d remastering locks for node %u, "
532 		     "retrying.\n", status, dlm->reco.dead_node);
533 		/* yield a bit to allow any final network messages
534 		 * to get handled on remaining nodes */
535 		msleep(100);
536 	} else {
537 		/* success!  see if any other nodes need recovery */
538 		mlog(0, "DONE mastering recovery of %s:%u here(this=%u)!\n",
539 		     dlm->name, dlm->reco.dead_node, dlm->node_num);
540 		dlm_reset_recovery(dlm);
541 	}
542 	dlm_end_recovery(dlm);
543 
544 	/* continue and look for another dead node */
545 	return -EAGAIN;
546 }
547 
548 static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node)
549 {
550 	int status = 0;
551 	struct dlm_reco_node_data *ndata;
552 	struct list_head *iter;
553 	int all_nodes_done;
554 	int destroy = 0;
555 	int pass = 0;
556 
557 	do {
558 		/* we have become recovery master.  there is no escaping
559 		 * this, so just keep trying until we get it. */
560 		status = dlm_init_recovery_area(dlm, dead_node);
561 		if (status < 0) {
562 			mlog(ML_ERROR, "%s: failed to alloc recovery area, "
563 			     "retrying\n", dlm->name);
564 			msleep(1000);
565 		}
566 	} while (status != 0);
567 
568 	/* safe to access the node data list without a lock, since this
569 	 * process is the only one to change the list */
570 	list_for_each(iter, &dlm->reco.node_data) {
571 		ndata = list_entry (iter, struct dlm_reco_node_data, list);
572 		BUG_ON(ndata->state != DLM_RECO_NODE_DATA_INIT);
573 		ndata->state = DLM_RECO_NODE_DATA_REQUESTING;
574 
575 		mlog(0, "requesting lock info from node %u\n",
576 		     ndata->node_num);
577 
578 		if (ndata->node_num == dlm->node_num) {
579 			ndata->state = DLM_RECO_NODE_DATA_DONE;
580 			continue;
581 		}
582 
583 		do {
584 			status = dlm_request_all_locks(dlm, ndata->node_num,
585 						       dead_node);
586 			if (status < 0) {
587 				mlog_errno(status);
588 				if (dlm_is_host_down(status)) {
589 					/* node died, ignore it for recovery */
590 					status = 0;
591 					ndata->state = DLM_RECO_NODE_DATA_DEAD;
592 					/* wait for the domain map to catch up
593 					 * with the network state. */
594 					wait_event_timeout(dlm->dlm_reco_thread_wq,
595 							   dlm_is_node_dead(dlm,
596 								ndata->node_num),
597 							   msecs_to_jiffies(1000));
598 					mlog(0, "waited 1 sec for %u, "
599 					     "dead? %s\n", ndata->node_num,
600 					     dlm_is_node_dead(dlm, ndata->node_num) ?
601 					     "yes" : "no");
602 				} else {
603 					/* -ENOMEM on the other node */
604 					mlog(0, "%s: node %u returned "
605 					     "%d during recovery, retrying "
606 					     "after a short wait\n",
607 					     dlm->name, ndata->node_num,
608 					     status);
609 					msleep(100);
610 				}
611 			}
612 		} while (status != 0);
613 
614 		switch (ndata->state) {
615 			case DLM_RECO_NODE_DATA_INIT:
616 			case DLM_RECO_NODE_DATA_FINALIZE_SENT:
617 			case DLM_RECO_NODE_DATA_REQUESTED:
618 				BUG();
619 				break;
620 			case DLM_RECO_NODE_DATA_DEAD:
621 				mlog(0, "node %u died after requesting "
622 				     "recovery info for node %u\n",
623 				     ndata->node_num, dead_node);
624 				/* fine.  don't need this node's info.
625 				 * continue without it. */
626 				break;
627 			case DLM_RECO_NODE_DATA_REQUESTING:
628 				ndata->state = DLM_RECO_NODE_DATA_REQUESTED;
629 				mlog(0, "now receiving recovery data from "
630 				     "node %u for dead node %u\n",
631 				     ndata->node_num, dead_node);
632 				break;
633 			case DLM_RECO_NODE_DATA_RECEIVING:
634 				mlog(0, "already receiving recovery data from "
635 				     "node %u for dead node %u\n",
636 				     ndata->node_num, dead_node);
637 				break;
638 			case DLM_RECO_NODE_DATA_DONE:
639 				mlog(0, "already DONE receiving recovery data "
640 				     "from node %u for dead node %u\n",
641 				     ndata->node_num, dead_node);
642 				break;
643 		}
644 	}
645 
646 	mlog(0, "done requesting all lock info\n");
647 
648 	/* nodes should be sending reco data now
649 	 * just need to wait */
650 
651 	while (1) {
652 		/* check all the nodes now to see if we are
653 		 * done, or if anyone died */
654 		all_nodes_done = 1;
655 		spin_lock(&dlm_reco_state_lock);
656 		list_for_each(iter, &dlm->reco.node_data) {
657 			ndata = list_entry (iter, struct dlm_reco_node_data, list);
658 
659 			mlog(0, "checking recovery state of node %u\n",
660 			     ndata->node_num);
661 			switch (ndata->state) {
662 				case DLM_RECO_NODE_DATA_INIT:
663 				case DLM_RECO_NODE_DATA_REQUESTING:
664 					mlog(ML_ERROR, "bad ndata state for "
665 					     "node %u: state=%d\n",
666 					     ndata->node_num, ndata->state);
667 					BUG();
668 					break;
669 				case DLM_RECO_NODE_DATA_DEAD:
670 					mlog(0, "node %u died after "
671 					     "requesting recovery info for "
672 					     "node %u\n", ndata->node_num,
673 					     dead_node);
674 					break;
675 				case DLM_RECO_NODE_DATA_RECEIVING:
676 				case DLM_RECO_NODE_DATA_REQUESTED:
677 					mlog(0, "%s: node %u still in state %s\n",
678 					     dlm->name, ndata->node_num,
679 					     ndata->state==DLM_RECO_NODE_DATA_RECEIVING ?
680 					     "receiving" : "requested");
681 					all_nodes_done = 0;
682 					break;
683 				case DLM_RECO_NODE_DATA_DONE:
684 					mlog(0, "%s: node %u state is done\n",
685 					     dlm->name, ndata->node_num);
686 					break;
687 				case DLM_RECO_NODE_DATA_FINALIZE_SENT:
688 					mlog(0, "%s: node %u state is finalize\n",
689 					     dlm->name, ndata->node_num);
690 					break;
691 			}
692 		}
693 		spin_unlock(&dlm_reco_state_lock);
694 
695 		mlog(0, "pass #%d, all_nodes_done?: %s\n", ++pass,
696 		     all_nodes_done?"yes":"no");
697 		if (all_nodes_done) {
698 			int ret;
699 
700 			/* all nodes are now in DLM_RECO_NODE_DATA_DONE state
701 	 		 * just send a finalize message to everyone and
702 	 		 * clean up */
703 			mlog(0, "all nodes are done! send finalize\n");
704 			ret = dlm_send_finalize_reco_message(dlm);
705 			if (ret < 0)
706 				mlog_errno(ret);
707 
708 			spin_lock(&dlm->spinlock);
709 			dlm_finish_local_lockres_recovery(dlm, dead_node,
710 							  dlm->node_num);
711 			spin_unlock(&dlm->spinlock);
712 			mlog(0, "should be done with recovery!\n");
713 
714 			mlog(0, "finishing recovery of %s at %lu, "
715 			     "dead=%u, this=%u, new=%u\n", dlm->name,
716 			     jiffies, dlm->reco.dead_node,
717 			     dlm->node_num, dlm->reco.new_master);
718 			destroy = 1;
719 			status = 0;
720 			/* rescan everything marked dirty along the way */
721 			dlm_kick_thread(dlm, NULL);
722 			break;
723 		}
724 		/* wait to be signalled, with periodic timeout
725 		 * to check for node death */
726 		wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq,
727 					 kthread_should_stop(),
728 					 msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS));
729 
730 	}
731 
732 	if (destroy)
733 		dlm_destroy_recovery_area(dlm, dead_node);
734 
735 	mlog_exit(status);
736 	return status;
737 }
738 
739 static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node)
740 {
741 	int num=0;
742 	struct dlm_reco_node_data *ndata;
743 
744 	spin_lock(&dlm->spinlock);
745 	memcpy(dlm->reco.node_map, dlm->domain_map, sizeof(dlm->domain_map));
746 	/* nodes can only be removed (by dying) after dropping
747 	 * this lock, and death will be trapped later, so this should do */
748 	spin_unlock(&dlm->spinlock);
749 
750 	while (1) {
751 		num = find_next_bit (dlm->reco.node_map, O2NM_MAX_NODES, num);
752 		if (num >= O2NM_MAX_NODES) {
753 			break;
754 		}
755 		BUG_ON(num == dead_node);
756 
757 		ndata = kzalloc(sizeof(*ndata), GFP_NOFS);
758 		if (!ndata) {
759 			dlm_destroy_recovery_area(dlm, dead_node);
760 			return -ENOMEM;
761 		}
762 		ndata->node_num = num;
763 		ndata->state = DLM_RECO_NODE_DATA_INIT;
764 		spin_lock(&dlm_reco_state_lock);
765 		list_add_tail(&ndata->list, &dlm->reco.node_data);
766 		spin_unlock(&dlm_reco_state_lock);
767 		num++;
768 	}
769 
770 	return 0;
771 }
772 
773 static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm, u8 dead_node)
774 {
775 	struct list_head *iter, *iter2;
776 	struct dlm_reco_node_data *ndata;
777 	LIST_HEAD(tmplist);
778 
779 	spin_lock(&dlm_reco_state_lock);
780 	list_splice_init(&dlm->reco.node_data, &tmplist);
781 	spin_unlock(&dlm_reco_state_lock);
782 
783 	list_for_each_safe(iter, iter2, &tmplist) {
784 		ndata = list_entry (iter, struct dlm_reco_node_data, list);
785 		list_del_init(&ndata->list);
786 		kfree(ndata);
787 	}
788 }
789 
790 static int dlm_request_all_locks(struct dlm_ctxt *dlm, u8 request_from,
791 				 u8 dead_node)
792 {
793 	struct dlm_lock_request lr;
794 	enum dlm_status ret;
795 
796 	mlog(0, "\n");
797 
798 
799 	mlog(0, "dlm_request_all_locks: dead node is %u, sending request "
800 		  "to %u\n", dead_node, request_from);
801 
802 	memset(&lr, 0, sizeof(lr));
803 	lr.node_idx = dlm->node_num;
804 	lr.dead_node = dead_node;
805 
806 	// send message
807 	ret = DLM_NOLOCKMGR;
808 	ret = o2net_send_message(DLM_LOCK_REQUEST_MSG, dlm->key,
809 				 &lr, sizeof(lr), request_from, NULL);
810 
811 	/* negative status is handled by caller */
812 	if (ret < 0)
813 		mlog_errno(ret);
814 
815 	// return from here, then
816 	// sleep until all received or error
817 	return ret;
818 
819 }
820 
821 int dlm_request_all_locks_handler(struct o2net_msg *msg, u32 len, void *data,
822 				  void **ret_data)
823 {
824 	struct dlm_ctxt *dlm = data;
825 	struct dlm_lock_request *lr = (struct dlm_lock_request *)msg->buf;
826 	char *buf = NULL;
827 	struct dlm_work_item *item = NULL;
828 
829 	if (!dlm_grab(dlm))
830 		return -EINVAL;
831 
832 	if (lr->dead_node != dlm->reco.dead_node) {
833 		mlog(ML_ERROR, "%s: node %u sent dead_node=%u, but local "
834 		     "dead_node is %u\n", dlm->name, lr->node_idx,
835 		     lr->dead_node, dlm->reco.dead_node);
836 		dlm_print_reco_node_status(dlm);
837 		/* this is a hack */
838 		dlm_put(dlm);
839 		return -ENOMEM;
840 	}
841 	BUG_ON(lr->dead_node != dlm->reco.dead_node);
842 
843 	item = kzalloc(sizeof(*item), GFP_NOFS);
844 	if (!item) {
845 		dlm_put(dlm);
846 		return -ENOMEM;
847 	}
848 
849 	/* this will get freed by dlm_request_all_locks_worker */
850 	buf = (char *) __get_free_page(GFP_NOFS);
851 	if (!buf) {
852 		kfree(item);
853 		dlm_put(dlm);
854 		return -ENOMEM;
855 	}
856 
857 	/* queue up work for dlm_request_all_locks_worker */
858 	dlm_grab(dlm);  /* get an extra ref for the work item */
859 	dlm_init_work_item(dlm, item, dlm_request_all_locks_worker, buf);
860 	item->u.ral.reco_master = lr->node_idx;
861 	item->u.ral.dead_node = lr->dead_node;
862 	spin_lock(&dlm->work_lock);
863 	list_add_tail(&item->list, &dlm->work_list);
864 	spin_unlock(&dlm->work_lock);
865 	queue_work(dlm->dlm_worker, &dlm->dispatched_work);
866 
867 	dlm_put(dlm);
868 	return 0;
869 }
870 
871 static void dlm_request_all_locks_worker(struct dlm_work_item *item, void *data)
872 {
873 	struct dlm_migratable_lockres *mres;
874 	struct dlm_lock_resource *res;
875 	struct dlm_ctxt *dlm;
876 	LIST_HEAD(resources);
877 	struct list_head *iter;
878 	int ret;
879 	u8 dead_node, reco_master;
880 	int skip_all_done = 0;
881 
882 	dlm = item->dlm;
883 	dead_node = item->u.ral.dead_node;
884 	reco_master = item->u.ral.reco_master;
885 	mres = (struct dlm_migratable_lockres *)data;
886 
887 	mlog(0, "%s: recovery worker started, dead=%u, master=%u\n",
888 	     dlm->name, dead_node, reco_master);
889 
890 	if (dead_node != dlm->reco.dead_node ||
891 	    reco_master != dlm->reco.new_master) {
892 		/* worker could have been created before the recovery master
893 		 * died.  if so, do not continue, but do not error. */
894 		if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) {
895 			mlog(ML_NOTICE, "%s: will not send recovery state, "
896 			     "recovery master %u died, thread=(dead=%u,mas=%u)"
897 			     " current=(dead=%u,mas=%u)\n", dlm->name,
898 			     reco_master, dead_node, reco_master,
899 			     dlm->reco.dead_node, dlm->reco.new_master);
900 		} else {
901 			mlog(ML_NOTICE, "%s: reco state invalid: reco(dead=%u, "
902 			     "master=%u), request(dead=%u, master=%u)\n",
903 			     dlm->name, dlm->reco.dead_node,
904 			     dlm->reco.new_master, dead_node, reco_master);
905 		}
906 		goto leave;
907 	}
908 
909 	/* lock resources should have already been moved to the
910  	 * dlm->reco.resources list.  now move items from that list
911  	 * to a temp list if the dead owner matches.  note that the
912 	 * whole cluster recovers only one node at a time, so we
913 	 * can safely move UNKNOWN lock resources for each recovery
914 	 * session. */
915 	dlm_move_reco_locks_to_list(dlm, &resources, dead_node);
916 
917 	/* now we can begin blasting lockreses without the dlm lock */
918 
919 	/* any errors returned will be due to the new_master dying,
920 	 * the dlm_reco_thread should detect this */
921 	list_for_each(iter, &resources) {
922 		res = list_entry (iter, struct dlm_lock_resource, recovering);
923 		ret = dlm_send_one_lockres(dlm, res, mres, reco_master,
924 				   	DLM_MRES_RECOVERY);
925 		if (ret < 0) {
926 			mlog(ML_ERROR, "%s: node %u went down while sending "
927 			     "recovery state for dead node %u, ret=%d\n", dlm->name,
928 			     reco_master, dead_node, ret);
929 			skip_all_done = 1;
930 			break;
931 		}
932 	}
933 
934 	/* move the resources back to the list */
935 	spin_lock(&dlm->spinlock);
936 	list_splice_init(&resources, &dlm->reco.resources);
937 	spin_unlock(&dlm->spinlock);
938 
939 	if (!skip_all_done) {
940 		ret = dlm_send_all_done_msg(dlm, dead_node, reco_master);
941 		if (ret < 0) {
942 			mlog(ML_ERROR, "%s: node %u went down while sending "
943 			     "recovery all-done for dead node %u, ret=%d\n",
944 			     dlm->name, reco_master, dead_node, ret);
945 		}
946 	}
947 leave:
948 	free_page((unsigned long)data);
949 }
950 
951 
952 static int dlm_send_all_done_msg(struct dlm_ctxt *dlm, u8 dead_node, u8 send_to)
953 {
954 	int ret, tmpret;
955 	struct dlm_reco_data_done done_msg;
956 
957 	memset(&done_msg, 0, sizeof(done_msg));
958 	done_msg.node_idx = dlm->node_num;
959 	done_msg.dead_node = dead_node;
960 	mlog(0, "sending DATA DONE message to %u, "
961 	     "my node=%u, dead node=%u\n", send_to, done_msg.node_idx,
962 	     done_msg.dead_node);
963 
964 	ret = o2net_send_message(DLM_RECO_DATA_DONE_MSG, dlm->key, &done_msg,
965 				 sizeof(done_msg), send_to, &tmpret);
966 	if (ret < 0) {
967 		if (!dlm_is_host_down(ret)) {
968 			mlog_errno(ret);
969 			mlog(ML_ERROR, "%s: unknown error sending data-done "
970 			     "to %u\n", dlm->name, send_to);
971 			BUG();
972 		}
973 	} else
974 		ret = tmpret;
975 	return ret;
976 }
977 
978 
979 int dlm_reco_data_done_handler(struct o2net_msg *msg, u32 len, void *data,
980 			       void **ret_data)
981 {
982 	struct dlm_ctxt *dlm = data;
983 	struct dlm_reco_data_done *done = (struct dlm_reco_data_done *)msg->buf;
984 	struct list_head *iter;
985 	struct dlm_reco_node_data *ndata = NULL;
986 	int ret = -EINVAL;
987 
988 	if (!dlm_grab(dlm))
989 		return -EINVAL;
990 
991 	mlog(0, "got DATA DONE: dead_node=%u, reco.dead_node=%u, "
992 	     "node_idx=%u, this node=%u\n", done->dead_node,
993 	     dlm->reco.dead_node, done->node_idx, dlm->node_num);
994 
995 	mlog_bug_on_msg((done->dead_node != dlm->reco.dead_node),
996 			"Got DATA DONE: dead_node=%u, reco.dead_node=%u, "
997 			"node_idx=%u, this node=%u\n", done->dead_node,
998 			dlm->reco.dead_node, done->node_idx, dlm->node_num);
999 
1000 	spin_lock(&dlm_reco_state_lock);
1001 	list_for_each(iter, &dlm->reco.node_data) {
1002 		ndata = list_entry (iter, struct dlm_reco_node_data, list);
1003 		if (ndata->node_num != done->node_idx)
1004 			continue;
1005 
1006 		switch (ndata->state) {
1007 			/* should have moved beyond INIT but not to FINALIZE yet */
1008 			case DLM_RECO_NODE_DATA_INIT:
1009 			case DLM_RECO_NODE_DATA_DEAD:
1010 			case DLM_RECO_NODE_DATA_FINALIZE_SENT:
1011 				mlog(ML_ERROR, "bad ndata state for node %u:"
1012 				     " state=%d\n", ndata->node_num,
1013 				     ndata->state);
1014 				BUG();
1015 				break;
1016 			/* these states are possible at this point, anywhere along
1017 			 * the line of recovery */
1018 			case DLM_RECO_NODE_DATA_DONE:
1019 			case DLM_RECO_NODE_DATA_RECEIVING:
1020 			case DLM_RECO_NODE_DATA_REQUESTED:
1021 			case DLM_RECO_NODE_DATA_REQUESTING:
1022 				mlog(0, "node %u is DONE sending "
1023 					  "recovery data!\n",
1024 					  ndata->node_num);
1025 
1026 				ndata->state = DLM_RECO_NODE_DATA_DONE;
1027 				ret = 0;
1028 				break;
1029 		}
1030 	}
1031 	spin_unlock(&dlm_reco_state_lock);
1032 
1033 	/* wake the recovery thread, some node is done */
1034 	if (!ret)
1035 		dlm_kick_recovery_thread(dlm);
1036 
1037 	if (ret < 0)
1038 		mlog(ML_ERROR, "failed to find recovery node data for node "
1039 		     "%u\n", done->node_idx);
1040 	dlm_put(dlm);
1041 
1042 	mlog(0, "leaving reco data done handler, ret=%d\n", ret);
1043 	return ret;
1044 }
1045 
1046 static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm,
1047 					struct list_head *list,
1048 				       	u8 dead_node)
1049 {
1050 	struct dlm_lock_resource *res;
1051 	struct list_head *iter, *iter2;
1052 	struct dlm_lock *lock;
1053 
1054 	spin_lock(&dlm->spinlock);
1055 	list_for_each_safe(iter, iter2, &dlm->reco.resources) {
1056 		res = list_entry (iter, struct dlm_lock_resource, recovering);
1057 		/* always prune any $RECOVERY entries for dead nodes,
1058 		 * otherwise hangs can occur during later recovery */
1059 		if (dlm_is_recovery_lock(res->lockname.name,
1060 					 res->lockname.len)) {
1061 			spin_lock(&res->spinlock);
1062 			list_for_each_entry(lock, &res->granted, list) {
1063 				if (lock->ml.node == dead_node) {
1064 					mlog(0, "AHA! there was "
1065 					     "a $RECOVERY lock for dead "
1066 					     "node %u (%s)!\n",
1067 					     dead_node, dlm->name);
1068 					list_del_init(&lock->list);
1069 					dlm_lock_put(lock);
1070 					break;
1071 				}
1072 			}
1073 			spin_unlock(&res->spinlock);
1074 			continue;
1075 		}
1076 
1077 		if (res->owner == dead_node) {
1078 			mlog(0, "found lockres owned by dead node while "
1079 				  "doing recovery for node %u. sending it.\n",
1080 				  dead_node);
1081 			list_move_tail(&res->recovering, list);
1082 		} else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
1083 			mlog(0, "found UNKNOWN owner while doing recovery "
1084 				  "for node %u. sending it.\n", dead_node);
1085 			list_move_tail(&res->recovering, list);
1086 		}
1087 	}
1088 	spin_unlock(&dlm->spinlock);
1089 }
1090 
1091 static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res)
1092 {
1093 	int total_locks = 0;
1094 	struct list_head *iter, *queue = &res->granted;
1095 	int i;
1096 
1097 	for (i=0; i<3; i++) {
1098 		list_for_each(iter, queue)
1099 			total_locks++;
1100 		queue++;
1101 	}
1102 	return total_locks;
1103 }
1104 
1105 
1106 static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm,
1107 				      struct dlm_migratable_lockres *mres,
1108 				      u8 send_to,
1109 				      struct dlm_lock_resource *res,
1110 				      int total_locks)
1111 {
1112 	u64 mig_cookie = be64_to_cpu(mres->mig_cookie);
1113 	int mres_total_locks = be32_to_cpu(mres->total_locks);
1114 	int sz, ret = 0, status = 0;
1115 	u8 orig_flags = mres->flags,
1116 	   orig_master = mres->master;
1117 
1118 	BUG_ON(mres->num_locks > DLM_MAX_MIGRATABLE_LOCKS);
1119 	if (!mres->num_locks)
1120 		return 0;
1121 
1122 	sz = sizeof(struct dlm_migratable_lockres) +
1123 		(mres->num_locks * sizeof(struct dlm_migratable_lock));
1124 
1125 	/* add an all-done flag if we reached the last lock */
1126 	orig_flags = mres->flags;
1127 	BUG_ON(total_locks > mres_total_locks);
1128 	if (total_locks == mres_total_locks)
1129 		mres->flags |= DLM_MRES_ALL_DONE;
1130 
1131 	mlog(0, "%s:%.*s: sending mig lockres (%s) to %u\n",
1132 	     dlm->name, res->lockname.len, res->lockname.name,
1133 	     orig_flags & DLM_MRES_MIGRATION ? "migrate" : "recovery",
1134 	     send_to);
1135 
1136 	/* send it */
1137 	ret = o2net_send_message(DLM_MIG_LOCKRES_MSG, dlm->key, mres,
1138 				 sz, send_to, &status);
1139 	if (ret < 0) {
1140 		/* XXX: negative status is not handled.
1141 		 * this will end up killing this node. */
1142 		mlog_errno(ret);
1143 	} else {
1144 		/* might get an -ENOMEM back here */
1145 		ret = status;
1146 		if (ret < 0) {
1147 			mlog_errno(ret);
1148 
1149 			if (ret == -EFAULT) {
1150 				mlog(ML_ERROR, "node %u told me to kill "
1151 				     "myself!\n", send_to);
1152 				BUG();
1153 			}
1154 		}
1155 	}
1156 
1157 	/* zero and reinit the message buffer */
1158 	dlm_init_migratable_lockres(mres, res->lockname.name,
1159 				    res->lockname.len, mres_total_locks,
1160 				    mig_cookie, orig_flags, orig_master);
1161 	return ret;
1162 }
1163 
1164 static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres,
1165 					const char *lockname, int namelen,
1166 					int total_locks, u64 cookie,
1167 					u8 flags, u8 master)
1168 {
1169 	/* mres here is one full page */
1170 	memset(mres, 0, PAGE_SIZE);
1171 	mres->lockname_len = namelen;
1172 	memcpy(mres->lockname, lockname, namelen);
1173 	mres->num_locks = 0;
1174 	mres->total_locks = cpu_to_be32(total_locks);
1175 	mres->mig_cookie = cpu_to_be64(cookie);
1176 	mres->flags = flags;
1177 	mres->master = master;
1178 }
1179 
1180 
1181 /* returns 1 if this lock fills the network structure,
1182  * 0 otherwise */
1183 static int dlm_add_lock_to_array(struct dlm_lock *lock,
1184 				 struct dlm_migratable_lockres *mres, int queue)
1185 {
1186 	struct dlm_migratable_lock *ml;
1187 	int lock_num = mres->num_locks;
1188 
1189 	ml = &(mres->ml[lock_num]);
1190 	ml->cookie = lock->ml.cookie;
1191 	ml->type = lock->ml.type;
1192 	ml->convert_type = lock->ml.convert_type;
1193 	ml->highest_blocked = lock->ml.highest_blocked;
1194 	ml->list = queue;
1195 	if (lock->lksb) {
1196 		ml->flags = lock->lksb->flags;
1197 		/* send our current lvb */
1198 		if (ml->type == LKM_EXMODE ||
1199 		    ml->type == LKM_PRMODE) {
1200 			/* if it is already set, this had better be a PR
1201 			 * and it has to match */
1202 			if (!dlm_lvb_is_empty(mres->lvb) &&
1203 			    (ml->type == LKM_EXMODE ||
1204 			     memcmp(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN))) {
1205 				mlog(ML_ERROR, "mismatched lvbs!\n");
1206 				__dlm_print_one_lock_resource(lock->lockres);
1207 				BUG();
1208 			}
1209 			memcpy(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN);
1210 		}
1211 	}
1212 	ml->node = lock->ml.node;
1213 	mres->num_locks++;
1214 	/* we reached the max, send this network message */
1215 	if (mres->num_locks == DLM_MAX_MIGRATABLE_LOCKS)
1216 		return 1;
1217 	return 0;
1218 }
1219 
1220 static void dlm_add_dummy_lock(struct dlm_ctxt *dlm,
1221 			       struct dlm_migratable_lockres *mres)
1222 {
1223 	struct dlm_lock dummy;
1224 	memset(&dummy, 0, sizeof(dummy));
1225 	dummy.ml.cookie = 0;
1226 	dummy.ml.type = LKM_IVMODE;
1227 	dummy.ml.convert_type = LKM_IVMODE;
1228 	dummy.ml.highest_blocked = LKM_IVMODE;
1229 	dummy.lksb = NULL;
1230 	dummy.ml.node = dlm->node_num;
1231 	dlm_add_lock_to_array(&dummy, mres, DLM_BLOCKED_LIST);
1232 }
1233 
1234 static inline int dlm_is_dummy_lock(struct dlm_ctxt *dlm,
1235 				    struct dlm_migratable_lock *ml,
1236 				    u8 *nodenum)
1237 {
1238 	if (unlikely(ml->cookie == 0 &&
1239 	    ml->type == LKM_IVMODE &&
1240 	    ml->convert_type == LKM_IVMODE &&
1241 	    ml->highest_blocked == LKM_IVMODE &&
1242 	    ml->list == DLM_BLOCKED_LIST)) {
1243 		*nodenum = ml->node;
1244 		return 1;
1245 	}
1246 	return 0;
1247 }
1248 
1249 int dlm_send_one_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
1250 			 struct dlm_migratable_lockres *mres,
1251 			 u8 send_to, u8 flags)
1252 {
1253 	struct list_head *queue, *iter;
1254 	int total_locks, i;
1255 	u64 mig_cookie = 0;
1256 	struct dlm_lock *lock;
1257 	int ret = 0;
1258 
1259 	BUG_ON(!(flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION)));
1260 
1261 	mlog(0, "sending to %u\n", send_to);
1262 
1263 	total_locks = dlm_num_locks_in_lockres(res);
1264 	if (total_locks > DLM_MAX_MIGRATABLE_LOCKS) {
1265 		/* rare, but possible */
1266 		mlog(0, "argh.  lockres has %d locks.  this will "
1267 			  "require more than one network packet to "
1268 			  "migrate\n", total_locks);
1269 		mig_cookie = dlm_get_next_mig_cookie();
1270 	}
1271 
1272 	dlm_init_migratable_lockres(mres, res->lockname.name,
1273 				    res->lockname.len, total_locks,
1274 				    mig_cookie, flags, res->owner);
1275 
1276 	total_locks = 0;
1277 	for (i=DLM_GRANTED_LIST; i<=DLM_BLOCKED_LIST; i++) {
1278 		queue = dlm_list_idx_to_ptr(res, i);
1279 		list_for_each(iter, queue) {
1280 			lock = list_entry (iter, struct dlm_lock, list);
1281 
1282 			/* add another lock. */
1283 			total_locks++;
1284 			if (!dlm_add_lock_to_array(lock, mres, i))
1285 				continue;
1286 
1287 			/* this filled the lock message,
1288 			 * we must send it immediately. */
1289 			ret = dlm_send_mig_lockres_msg(dlm, mres, send_to,
1290 						       res, total_locks);
1291 			if (ret < 0)
1292 				goto error;
1293 		}
1294 	}
1295 	if (total_locks == 0) {
1296 		/* send a dummy lock to indicate a mastery reference only */
1297 		mlog(0, "%s:%.*s: sending dummy lock to %u, %s\n",
1298 		     dlm->name, res->lockname.len, res->lockname.name,
1299 		     send_to, flags & DLM_MRES_RECOVERY ? "recovery" :
1300 		     "migration");
1301 		dlm_add_dummy_lock(dlm, mres);
1302 	}
1303 	/* flush any remaining locks */
1304 	ret = dlm_send_mig_lockres_msg(dlm, mres, send_to, res, total_locks);
1305 	if (ret < 0)
1306 		goto error;
1307 	return ret;
1308 
1309 error:
1310 	mlog(ML_ERROR, "%s: dlm_send_mig_lockres_msg returned %d\n",
1311 	     dlm->name, ret);
1312 	if (!dlm_is_host_down(ret))
1313 		BUG();
1314 	mlog(0, "%s: node %u went down while sending %s "
1315 	     "lockres %.*s\n", dlm->name, send_to,
1316 	     flags & DLM_MRES_RECOVERY ?  "recovery" : "migration",
1317 	     res->lockname.len, res->lockname.name);
1318 	return ret;
1319 }
1320 
1321 
1322 
1323 /*
1324  * this message will contain no more than one page worth of
1325  * recovery data, and it will work on only one lockres.
1326  * there may be many locks in this page, and we may need to wait
1327  * for additional packets to complete all the locks (rare, but
1328  * possible).
1329  */
1330 /*
1331  * NOTE: the allocation error cases here are scary
1332  * we really cannot afford to fail an alloc in recovery
1333  * do we spin?  returning an error only delays the problem really
1334  */
1335 
1336 int dlm_mig_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
1337 			    void **ret_data)
1338 {
1339 	struct dlm_ctxt *dlm = data;
1340 	struct dlm_migratable_lockres *mres =
1341 		(struct dlm_migratable_lockres *)msg->buf;
1342 	int ret = 0;
1343 	u8 real_master;
1344 	char *buf = NULL;
1345 	struct dlm_work_item *item = NULL;
1346 	struct dlm_lock_resource *res = NULL;
1347 
1348 	if (!dlm_grab(dlm))
1349 		return -EINVAL;
1350 
1351 	BUG_ON(!(mres->flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION)));
1352 
1353 	real_master = mres->master;
1354 	if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1355 		/* cannot migrate a lockres with no master */
1356 		BUG_ON(!(mres->flags & DLM_MRES_RECOVERY));
1357 	}
1358 
1359 	mlog(0, "%s message received from node %u\n",
1360 		  (mres->flags & DLM_MRES_RECOVERY) ?
1361 		  "recovery" : "migration", mres->master);
1362 	if (mres->flags & DLM_MRES_ALL_DONE)
1363 		mlog(0, "all done flag.  all lockres data received!\n");
1364 
1365 	ret = -ENOMEM;
1366 	buf = kmalloc(be16_to_cpu(msg->data_len), GFP_NOFS);
1367 	item = kzalloc(sizeof(*item), GFP_NOFS);
1368 	if (!buf || !item)
1369 		goto leave;
1370 
1371 	/* lookup the lock to see if we have a secondary queue for this
1372 	 * already...  just add the locks in and this will have its owner
1373 	 * and RECOVERY flag changed when it completes. */
1374 	res = dlm_lookup_lockres(dlm, mres->lockname, mres->lockname_len);
1375 	if (res) {
1376 	 	/* this will get a ref on res */
1377 		/* mark it as recovering/migrating and hash it */
1378 		spin_lock(&res->spinlock);
1379 		if (mres->flags & DLM_MRES_RECOVERY) {
1380 			res->state |= DLM_LOCK_RES_RECOVERING;
1381 		} else {
1382 			if (res->state & DLM_LOCK_RES_MIGRATING) {
1383 				/* this is at least the second
1384 				 * lockres message */
1385 				mlog(0, "lock %.*s is already migrating\n",
1386 					  mres->lockname_len,
1387 					  mres->lockname);
1388 			} else if (res->state & DLM_LOCK_RES_RECOVERING) {
1389 				/* caller should BUG */
1390 				mlog(ML_ERROR, "node is attempting to migrate "
1391 				     "lock %.*s, but marked as recovering!\n",
1392 				     mres->lockname_len, mres->lockname);
1393 				ret = -EFAULT;
1394 				spin_unlock(&res->spinlock);
1395 				goto leave;
1396 			}
1397 			res->state |= DLM_LOCK_RES_MIGRATING;
1398 		}
1399 		spin_unlock(&res->spinlock);
1400 	} else {
1401 		/* need to allocate, just like if it was
1402 		 * mastered here normally  */
1403 		res = dlm_new_lockres(dlm, mres->lockname, mres->lockname_len);
1404 		if (!res)
1405 			goto leave;
1406 
1407 		/* to match the ref that we would have gotten if
1408 		 * dlm_lookup_lockres had succeeded */
1409 		dlm_lockres_get(res);
1410 
1411 		/* mark it as recovering/migrating and hash it */
1412 		if (mres->flags & DLM_MRES_RECOVERY)
1413 			res->state |= DLM_LOCK_RES_RECOVERING;
1414 		else
1415 			res->state |= DLM_LOCK_RES_MIGRATING;
1416 
1417 		spin_lock(&dlm->spinlock);
1418 		__dlm_insert_lockres(dlm, res);
1419 		spin_unlock(&dlm->spinlock);
1420 
1421 		/* now that the new lockres is inserted,
1422 		 * make it usable by other processes */
1423 		spin_lock(&res->spinlock);
1424 		res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
1425 		spin_unlock(&res->spinlock);
1426 		wake_up(&res->wq);
1427 
1428 		/* add an extra ref for just-allocated lockres
1429 		 * otherwise the lockres will be purged immediately */
1430 		dlm_lockres_get(res);
1431 	}
1432 
1433 	/* at this point we have allocated everything we need,
1434 	 * and we have a hashed lockres with an extra ref and
1435 	 * the proper res->state flags. */
1436 	ret = 0;
1437 	spin_lock(&res->spinlock);
1438 	/* drop this either when master requery finds a different master
1439 	 * or when a lock is added by the recovery worker */
1440 	dlm_lockres_grab_inflight_ref(dlm, res);
1441 	if (mres->master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1442 		/* migration cannot have an unknown master */
1443 		BUG_ON(!(mres->flags & DLM_MRES_RECOVERY));
1444 		mlog(0, "recovery has passed me a lockres with an "
1445 			  "unknown owner.. will need to requery: "
1446 			  "%.*s\n", mres->lockname_len, mres->lockname);
1447 	} else {
1448 		/* take a reference now to pin the lockres, drop it
1449 		 * when locks are added in the worker */
1450 		dlm_change_lockres_owner(dlm, res, dlm->node_num);
1451 	}
1452 	spin_unlock(&res->spinlock);
1453 
1454 	/* queue up work for dlm_mig_lockres_worker */
1455 	dlm_grab(dlm);  /* get an extra ref for the work item */
1456 	memcpy(buf, msg->buf, be16_to_cpu(msg->data_len));  /* copy the whole message */
1457 	dlm_init_work_item(dlm, item, dlm_mig_lockres_worker, buf);
1458 	item->u.ml.lockres = res; /* already have a ref */
1459 	item->u.ml.real_master = real_master;
1460 	spin_lock(&dlm->work_lock);
1461 	list_add_tail(&item->list, &dlm->work_list);
1462 	spin_unlock(&dlm->work_lock);
1463 	queue_work(dlm->dlm_worker, &dlm->dispatched_work);
1464 
1465 leave:
1466 	dlm_put(dlm);
1467 	if (ret < 0) {
1468 		if (buf)
1469 			kfree(buf);
1470 		if (item)
1471 			kfree(item);
1472 	}
1473 
1474 	mlog_exit(ret);
1475 	return ret;
1476 }
1477 
1478 
1479 static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data)
1480 {
1481 	struct dlm_ctxt *dlm = data;
1482 	struct dlm_migratable_lockres *mres;
1483 	int ret = 0;
1484 	struct dlm_lock_resource *res;
1485 	u8 real_master;
1486 
1487 	dlm = item->dlm;
1488 	mres = (struct dlm_migratable_lockres *)data;
1489 
1490 	res = item->u.ml.lockres;
1491 	real_master = item->u.ml.real_master;
1492 
1493 	if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1494 		/* this case is super-rare. only occurs if
1495 		 * node death happens during migration. */
1496 again:
1497 		ret = dlm_lockres_master_requery(dlm, res, &real_master);
1498 		if (ret < 0) {
1499 			mlog(0, "dlm_lockres_master_requery ret=%d\n",
1500 				  ret);
1501 			goto again;
1502 		}
1503 		if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1504 			mlog(0, "lockres %.*s not claimed.  "
1505 				   "this node will take it.\n",
1506 				   res->lockname.len, res->lockname.name);
1507 		} else {
1508 			spin_lock(&res->spinlock);
1509 			dlm_lockres_drop_inflight_ref(dlm, res);
1510 			spin_unlock(&res->spinlock);
1511 			mlog(0, "master needs to respond to sender "
1512 				  "that node %u still owns %.*s\n",
1513 				  real_master, res->lockname.len,
1514 				  res->lockname.name);
1515 			/* cannot touch this lockres */
1516 			goto leave;
1517 		}
1518 	}
1519 
1520 	ret = dlm_process_recovery_data(dlm, res, mres);
1521 	if (ret < 0)
1522 		mlog(0, "dlm_process_recovery_data returned  %d\n", ret);
1523 	else
1524 		mlog(0, "dlm_process_recovery_data succeeded\n");
1525 
1526 	if ((mres->flags & (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) ==
1527 	                   (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) {
1528 		ret = dlm_finish_migration(dlm, res, mres->master);
1529 		if (ret < 0)
1530 			mlog_errno(ret);
1531 	}
1532 
1533 leave:
1534 	kfree(data);
1535 	mlog_exit(ret);
1536 }
1537 
1538 
1539 
1540 static int dlm_lockres_master_requery(struct dlm_ctxt *dlm,
1541 				      struct dlm_lock_resource *res,
1542 				      u8 *real_master)
1543 {
1544 	struct dlm_node_iter iter;
1545 	int nodenum;
1546 	int ret = 0;
1547 
1548 	*real_master = DLM_LOCK_RES_OWNER_UNKNOWN;
1549 
1550 	/* we only reach here if one of the two nodes in a
1551 	 * migration died while the migration was in progress.
1552 	 * at this point we need to requery the master.  we
1553 	 * know that the new_master got as far as creating
1554 	 * an mle on at least one node, but we do not know
1555 	 * if any nodes had actually cleared the mle and set
1556 	 * the master to the new_master.  the old master
1557 	 * is supposed to set the owner to UNKNOWN in the
1558 	 * event of a new_master death, so the only possible
1559 	 * responses that we can get from nodes here are
1560 	 * that the master is new_master, or that the master
1561 	 * is UNKNOWN.
1562 	 * if all nodes come back with UNKNOWN then we know
1563 	 * the lock needs remastering here.
1564 	 * if any node comes back with a valid master, check
1565 	 * to see if that master is the one that we are
1566 	 * recovering.  if so, then the new_master died and
1567 	 * we need to remaster this lock.  if not, then the
1568 	 * new_master survived and that node will respond to
1569 	 * other nodes about the owner.
1570 	 * if there is an owner, this node needs to dump this
1571 	 * lockres and alert the sender that this lockres
1572 	 * was rejected. */
1573 	spin_lock(&dlm->spinlock);
1574 	dlm_node_iter_init(dlm->domain_map, &iter);
1575 	spin_unlock(&dlm->spinlock);
1576 
1577 	while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
1578 		/* do not send to self */
1579 		if (nodenum == dlm->node_num)
1580 			continue;
1581 		ret = dlm_do_master_requery(dlm, res, nodenum, real_master);
1582 		if (ret < 0) {
1583 			mlog_errno(ret);
1584 			if (!dlm_is_host_down(ret))
1585 				BUG();
1586 			/* host is down, so answer for that node would be
1587 			 * DLM_LOCK_RES_OWNER_UNKNOWN.  continue. */
1588 		}
1589 		if (*real_master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1590 			mlog(0, "lock master is %u\n", *real_master);
1591 			break;
1592 		}
1593 	}
1594 	return ret;
1595 }
1596 
1597 
1598 int dlm_do_master_requery(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
1599 			  u8 nodenum, u8 *real_master)
1600 {
1601 	int ret = -EINVAL;
1602 	struct dlm_master_requery req;
1603 	int status = DLM_LOCK_RES_OWNER_UNKNOWN;
1604 
1605 	memset(&req, 0, sizeof(req));
1606 	req.node_idx = dlm->node_num;
1607 	req.namelen = res->lockname.len;
1608 	memcpy(req.name, res->lockname.name, res->lockname.len);
1609 
1610 	ret = o2net_send_message(DLM_MASTER_REQUERY_MSG, dlm->key,
1611 				 &req, sizeof(req), nodenum, &status);
1612 	/* XXX: negative status not handled properly here. */
1613 	if (ret < 0)
1614 		mlog_errno(ret);
1615 	else {
1616 		BUG_ON(status < 0);
1617 		BUG_ON(status > DLM_LOCK_RES_OWNER_UNKNOWN);
1618 		*real_master = (u8) (status & 0xff);
1619 		mlog(0, "node %u responded to master requery with %u\n",
1620 			  nodenum, *real_master);
1621 		ret = 0;
1622 	}
1623 	return ret;
1624 }
1625 
1626 
1627 /* this function cannot error, so unless the sending
1628  * or receiving of the message failed, the owner can
1629  * be trusted */
1630 int dlm_master_requery_handler(struct o2net_msg *msg, u32 len, void *data,
1631 			       void **ret_data)
1632 {
1633 	struct dlm_ctxt *dlm = data;
1634 	struct dlm_master_requery *req = (struct dlm_master_requery *)msg->buf;
1635 	struct dlm_lock_resource *res = NULL;
1636 	unsigned int hash;
1637 	int master = DLM_LOCK_RES_OWNER_UNKNOWN;
1638 	u32 flags = DLM_ASSERT_MASTER_REQUERY;
1639 
1640 	if (!dlm_grab(dlm)) {
1641 		/* since the domain has gone away on this
1642 		 * node, the proper response is UNKNOWN */
1643 		return master;
1644 	}
1645 
1646 	hash = dlm_lockid_hash(req->name, req->namelen);
1647 
1648 	spin_lock(&dlm->spinlock);
1649 	res = __dlm_lookup_lockres(dlm, req->name, req->namelen, hash);
1650 	if (res) {
1651 		spin_lock(&res->spinlock);
1652 		master = res->owner;
1653 		if (master == dlm->node_num) {
1654 			int ret = dlm_dispatch_assert_master(dlm, res,
1655 							     0, 0, flags);
1656 			if (ret < 0) {
1657 				mlog_errno(-ENOMEM);
1658 				/* retry!? */
1659 				BUG();
1660 			}
1661 		}
1662 		spin_unlock(&res->spinlock);
1663 	}
1664 	spin_unlock(&dlm->spinlock);
1665 
1666 	dlm_put(dlm);
1667 	return master;
1668 }
1669 
1670 static inline struct list_head *
1671 dlm_list_num_to_pointer(struct dlm_lock_resource *res, int list_num)
1672 {
1673 	struct list_head *ret;
1674 	BUG_ON(list_num < 0);
1675 	BUG_ON(list_num > 2);
1676 	ret = &(res->granted);
1677 	ret += list_num;
1678 	return ret;
1679 }
1680 /* TODO: do ast flush business
1681  * TODO: do MIGRATING and RECOVERING spinning
1682  */
1683 
1684 /*
1685 * NOTE about in-flight requests during migration:
1686 *
1687 * Before attempting the migrate, the master has marked the lockres as
1688 * MIGRATING and then flushed all of its pending ASTS.  So any in-flight
1689 * requests either got queued before the MIGRATING flag got set, in which
1690 * case the lock data will reflect the change and a return message is on
1691 * the way, or the request failed to get in before MIGRATING got set.  In
1692 * this case, the caller will be told to spin and wait for the MIGRATING
1693 * flag to be dropped, then recheck the master.
1694 * This holds true for the convert, cancel and unlock cases, and since lvb
1695 * updates are tied to these same messages, it applies to lvb updates as
1696 * well.  For the lock case, there is no way a lock can be on the master
1697 * queue and not be on the secondary queue since the lock is always added
1698 * locally first.  This means that the new target node will never be sent
1699 * a lock that he doesn't already have on the list.
1700 * In total, this means that the local lock is correct and should not be
1701 * updated to match the one sent by the master.  Any messages sent back
1702 * from the master before the MIGRATING flag will bring the lock properly
1703 * up-to-date, and the change will be ordered properly for the waiter.
1704 * We will *not* attempt to modify the lock underneath the waiter.
1705 */
1706 
1707 static int dlm_process_recovery_data(struct dlm_ctxt *dlm,
1708 				     struct dlm_lock_resource *res,
1709 				     struct dlm_migratable_lockres *mres)
1710 {
1711 	struct dlm_migratable_lock *ml;
1712 	struct list_head *queue;
1713 	struct list_head *tmpq = NULL;
1714 	struct dlm_lock *newlock = NULL;
1715 	struct dlm_lockstatus *lksb = NULL;
1716 	int ret = 0;
1717 	int i, j, bad;
1718 	struct list_head *iter;
1719 	struct dlm_lock *lock = NULL;
1720 	u8 from = O2NM_MAX_NODES;
1721 	unsigned int added = 0;
1722 
1723 	mlog(0, "running %d locks for this lockres\n", mres->num_locks);
1724 	for (i=0; i<mres->num_locks; i++) {
1725 		ml = &(mres->ml[i]);
1726 
1727 		if (dlm_is_dummy_lock(dlm, ml, &from)) {
1728 			/* placeholder, just need to set the refmap bit */
1729 			BUG_ON(mres->num_locks != 1);
1730 			mlog(0, "%s:%.*s: dummy lock for %u\n",
1731 			     dlm->name, mres->lockname_len, mres->lockname,
1732 			     from);
1733 			spin_lock(&res->spinlock);
1734 			dlm_lockres_set_refmap_bit(from, res);
1735 			spin_unlock(&res->spinlock);
1736 			added++;
1737 			break;
1738 		}
1739 		BUG_ON(ml->highest_blocked != LKM_IVMODE);
1740 		newlock = NULL;
1741 		lksb = NULL;
1742 
1743 		queue = dlm_list_num_to_pointer(res, ml->list);
1744 		tmpq = NULL;
1745 
1746 		/* if the lock is for the local node it needs to
1747 		 * be moved to the proper location within the queue.
1748 		 * do not allocate a new lock structure. */
1749 		if (ml->node == dlm->node_num) {
1750 			/* MIGRATION ONLY! */
1751 			BUG_ON(!(mres->flags & DLM_MRES_MIGRATION));
1752 
1753 			spin_lock(&res->spinlock);
1754 			for (j = DLM_GRANTED_LIST; j <= DLM_BLOCKED_LIST; j++) {
1755 				tmpq = dlm_list_idx_to_ptr(res, j);
1756 				list_for_each(iter, tmpq) {
1757 					lock = list_entry (iter, struct dlm_lock, list);
1758 					if (lock->ml.cookie != ml->cookie)
1759 						lock = NULL;
1760 					else
1761 						break;
1762 				}
1763 				if (lock)
1764 					break;
1765 			}
1766 
1767 			/* lock is always created locally first, and
1768 			 * destroyed locally last.  it must be on the list */
1769 			if (!lock) {
1770 				u64 c = ml->cookie;
1771 				mlog(ML_ERROR, "could not find local lock "
1772 					       "with cookie %u:%llu!\n",
1773 				     dlm_get_lock_cookie_node(be64_to_cpu(c)),
1774 				     dlm_get_lock_cookie_seq(be64_to_cpu(c)));
1775 				__dlm_print_one_lock_resource(res);
1776 				BUG();
1777 			}
1778 			BUG_ON(lock->ml.node != ml->node);
1779 
1780 			if (tmpq != queue) {
1781 				mlog(0, "lock was on %u instead of %u for %.*s\n",
1782 				     j, ml->list, res->lockname.len, res->lockname.name);
1783 				spin_unlock(&res->spinlock);
1784 				continue;
1785 			}
1786 
1787 			/* see NOTE above about why we do not update
1788 			 * to match the master here */
1789 
1790 			/* move the lock to its proper place */
1791 			/* do not alter lock refcount.  switching lists. */
1792 			list_move_tail(&lock->list, queue);
1793 			spin_unlock(&res->spinlock);
1794 			added++;
1795 
1796 			mlog(0, "just reordered a local lock!\n");
1797 			continue;
1798 		}
1799 
1800 		/* lock is for another node. */
1801 		newlock = dlm_new_lock(ml->type, ml->node,
1802 				       be64_to_cpu(ml->cookie), NULL);
1803 		if (!newlock) {
1804 			ret = -ENOMEM;
1805 			goto leave;
1806 		}
1807 		lksb = newlock->lksb;
1808 		dlm_lock_attach_lockres(newlock, res);
1809 
1810 		if (ml->convert_type != LKM_IVMODE) {
1811 			BUG_ON(queue != &res->converting);
1812 			newlock->ml.convert_type = ml->convert_type;
1813 		}
1814 		lksb->flags |= (ml->flags &
1815 				(DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB));
1816 
1817 		if (ml->type == LKM_NLMODE)
1818 			goto skip_lvb;
1819 
1820 		if (!dlm_lvb_is_empty(mres->lvb)) {
1821 			if (lksb->flags & DLM_LKSB_PUT_LVB) {
1822 				/* other node was trying to update
1823 				 * lvb when node died.  recreate the
1824 				 * lksb with the updated lvb. */
1825 				memcpy(lksb->lvb, mres->lvb, DLM_LVB_LEN);
1826 				/* the lock resource lvb update must happen
1827 				 * NOW, before the spinlock is dropped.
1828 				 * we no longer wait for the AST to update
1829 				 * the lvb. */
1830 				memcpy(res->lvb, mres->lvb, DLM_LVB_LEN);
1831 			} else {
1832 				/* otherwise, the node is sending its
1833 				 * most recent valid lvb info */
1834 				BUG_ON(ml->type != LKM_EXMODE &&
1835 				       ml->type != LKM_PRMODE);
1836 				if (!dlm_lvb_is_empty(res->lvb) &&
1837  				    (ml->type == LKM_EXMODE ||
1838  				     memcmp(res->lvb, mres->lvb, DLM_LVB_LEN))) {
1839  					int i;
1840  					mlog(ML_ERROR, "%s:%.*s: received bad "
1841  					     "lvb! type=%d\n", dlm->name,
1842  					     res->lockname.len,
1843  					     res->lockname.name, ml->type);
1844  					printk("lockres lvb=[");
1845  					for (i=0; i<DLM_LVB_LEN; i++)
1846  						printk("%02x", res->lvb[i]);
1847  					printk("]\nmigrated lvb=[");
1848  					for (i=0; i<DLM_LVB_LEN; i++)
1849  						printk("%02x", mres->lvb[i]);
1850  					printk("]\n");
1851  					dlm_print_one_lock_resource(res);
1852  					BUG();
1853 				}
1854 				memcpy(res->lvb, mres->lvb, DLM_LVB_LEN);
1855 			}
1856 		}
1857 skip_lvb:
1858 
1859 		/* NOTE:
1860 		 * wrt lock queue ordering and recovery:
1861 		 *    1. order of locks on granted queue is
1862 		 *       meaningless.
1863 		 *    2. order of locks on converting queue is
1864 		 *       LOST with the node death.  sorry charlie.
1865 		 *    3. order of locks on the blocked queue is
1866 		 *       also LOST.
1867 		 * order of locks does not affect integrity, it
1868 		 * just means that a lock request may get pushed
1869 		 * back in line as a result of the node death.
1870 		 * also note that for a given node the lock order
1871 		 * for its secondary queue locks is preserved
1872 		 * relative to each other, but clearly *not*
1873 		 * preserved relative to locks from other nodes.
1874 		 */
1875 		bad = 0;
1876 		spin_lock(&res->spinlock);
1877 		list_for_each_entry(lock, queue, list) {
1878 			if (lock->ml.cookie == ml->cookie) {
1879 				u64 c = lock->ml.cookie;
1880 				mlog(ML_ERROR, "%s:%.*s: %u:%llu: lock already "
1881 				     "exists on this lockres!\n", dlm->name,
1882 				     res->lockname.len, res->lockname.name,
1883 				     dlm_get_lock_cookie_node(be64_to_cpu(c)),
1884 				     dlm_get_lock_cookie_seq(be64_to_cpu(c)));
1885 
1886 				mlog(ML_NOTICE, "sent lock: type=%d, conv=%d, "
1887 				     "node=%u, cookie=%u:%llu, queue=%d\n",
1888 	      			     ml->type, ml->convert_type, ml->node,
1889 				     dlm_get_lock_cookie_node(be64_to_cpu(ml->cookie)),
1890 				     dlm_get_lock_cookie_seq(be64_to_cpu(ml->cookie)),
1891 				     ml->list);
1892 
1893 				__dlm_print_one_lock_resource(res);
1894 				bad = 1;
1895 				break;
1896 			}
1897 		}
1898 		if (!bad) {
1899 			dlm_lock_get(newlock);
1900 			list_add_tail(&newlock->list, queue);
1901 			mlog(0, "%s:%.*s: added lock for node %u, "
1902 			     "setting refmap bit\n", dlm->name,
1903 			     res->lockname.len, res->lockname.name, ml->node);
1904 			dlm_lockres_set_refmap_bit(ml->node, res);
1905 			added++;
1906 		}
1907 		spin_unlock(&res->spinlock);
1908 	}
1909 	mlog(0, "done running all the locks\n");
1910 
1911 leave:
1912 	/* balance the ref taken when the work was queued */
1913 	spin_lock(&res->spinlock);
1914 	dlm_lockres_drop_inflight_ref(dlm, res);
1915 	spin_unlock(&res->spinlock);
1916 
1917 	if (ret < 0) {
1918 		mlog_errno(ret);
1919 		if (newlock)
1920 			dlm_lock_put(newlock);
1921 	}
1922 
1923 	mlog_exit(ret);
1924 	return ret;
1925 }
1926 
1927 void dlm_move_lockres_to_recovery_list(struct dlm_ctxt *dlm,
1928 				       struct dlm_lock_resource *res)
1929 {
1930 	int i;
1931 	struct list_head *queue, *iter, *iter2;
1932 	struct dlm_lock *lock;
1933 
1934 	res->state |= DLM_LOCK_RES_RECOVERING;
1935 	if (!list_empty(&res->recovering)) {
1936 		mlog(0,
1937 		     "Recovering res %s:%.*s, is already on recovery list!\n",
1938 		     dlm->name, res->lockname.len, res->lockname.name);
1939 		list_del_init(&res->recovering);
1940 	}
1941 	/* We need to hold a reference while on the recovery list */
1942 	dlm_lockres_get(res);
1943 	list_add_tail(&res->recovering, &dlm->reco.resources);
1944 
1945 	/* find any pending locks and put them back on proper list */
1946 	for (i=DLM_BLOCKED_LIST; i>=DLM_GRANTED_LIST; i--) {
1947 		queue = dlm_list_idx_to_ptr(res, i);
1948 		list_for_each_safe(iter, iter2, queue) {
1949 			lock = list_entry (iter, struct dlm_lock, list);
1950 			dlm_lock_get(lock);
1951 			if (lock->convert_pending) {
1952 				/* move converting lock back to granted */
1953 				BUG_ON(i != DLM_CONVERTING_LIST);
1954 				mlog(0, "node died with convert pending "
1955 				     "on %.*s. move back to granted list.\n",
1956 				     res->lockname.len, res->lockname.name);
1957 				dlm_revert_pending_convert(res, lock);
1958 				lock->convert_pending = 0;
1959 			} else if (lock->lock_pending) {
1960 				/* remove pending lock requests completely */
1961 				BUG_ON(i != DLM_BLOCKED_LIST);
1962 				mlog(0, "node died with lock pending "
1963 				     "on %.*s. remove from blocked list and skip.\n",
1964 				     res->lockname.len, res->lockname.name);
1965 				/* lock will be floating until ref in
1966 				 * dlmlock_remote is freed after the network
1967 				 * call returns.  ok for it to not be on any
1968 				 * list since no ast can be called
1969 				 * (the master is dead). */
1970 				dlm_revert_pending_lock(res, lock);
1971 				lock->lock_pending = 0;
1972 			} else if (lock->unlock_pending) {
1973 				/* if an unlock was in progress, treat as
1974 				 * if this had completed successfully
1975 				 * before sending this lock state to the
1976 				 * new master.  note that the dlm_unlock
1977 				 * call is still responsible for calling
1978 				 * the unlockast.  that will happen after
1979 				 * the network call times out.  for now,
1980 				 * just move lists to prepare the new
1981 				 * recovery master.  */
1982 				BUG_ON(i != DLM_GRANTED_LIST);
1983 				mlog(0, "node died with unlock pending "
1984 				     "on %.*s. remove from blocked list and skip.\n",
1985 				     res->lockname.len, res->lockname.name);
1986 				dlm_commit_pending_unlock(res, lock);
1987 				lock->unlock_pending = 0;
1988 			} else if (lock->cancel_pending) {
1989 				/* if a cancel was in progress, treat as
1990 				 * if this had completed successfully
1991 				 * before sending this lock state to the
1992 				 * new master */
1993 				BUG_ON(i != DLM_CONVERTING_LIST);
1994 				mlog(0, "node died with cancel pending "
1995 				     "on %.*s. move back to granted list.\n",
1996 				     res->lockname.len, res->lockname.name);
1997 				dlm_commit_pending_cancel(res, lock);
1998 				lock->cancel_pending = 0;
1999 			}
2000 			dlm_lock_put(lock);
2001 		}
2002 	}
2003 }
2004 
2005 
2006 
2007 /* removes all recovered locks from the recovery list.
2008  * sets the res->owner to the new master.
2009  * unsets the RECOVERY flag and wakes waiters. */
2010 static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm,
2011 					      u8 dead_node, u8 new_master)
2012 {
2013 	int i;
2014 	struct list_head *iter, *iter2;
2015 	struct hlist_node *hash_iter;
2016 	struct hlist_head *bucket;
2017 
2018 	struct dlm_lock_resource *res;
2019 
2020 	mlog_entry_void();
2021 
2022 	assert_spin_locked(&dlm->spinlock);
2023 
2024 	list_for_each_safe(iter, iter2, &dlm->reco.resources) {
2025 		res = list_entry (iter, struct dlm_lock_resource, recovering);
2026 		if (res->owner == dead_node) {
2027 			list_del_init(&res->recovering);
2028 			spin_lock(&res->spinlock);
2029 			/* new_master has our reference from
2030 			 * the lock state sent during recovery */
2031 			dlm_change_lockres_owner(dlm, res, new_master);
2032 			res->state &= ~DLM_LOCK_RES_RECOVERING;
2033 			if (__dlm_lockres_has_locks(res))
2034 				__dlm_dirty_lockres(dlm, res);
2035 			spin_unlock(&res->spinlock);
2036 			wake_up(&res->wq);
2037 			dlm_lockres_put(res);
2038 		}
2039 	}
2040 
2041 	/* this will become unnecessary eventually, but
2042 	 * for now we need to run the whole hash, clear
2043 	 * the RECOVERING state and set the owner
2044 	 * if necessary */
2045 	for (i = 0; i < DLM_HASH_BUCKETS; i++) {
2046 		bucket = dlm_lockres_hash(dlm, i);
2047 		hlist_for_each_entry(res, hash_iter, bucket, hash_node) {
2048 			if (res->state & DLM_LOCK_RES_RECOVERING) {
2049 				if (res->owner == dead_node) {
2050 					mlog(0, "(this=%u) res %.*s owner=%u "
2051 					     "was not on recovering list, but "
2052 					     "clearing state anyway\n",
2053 					     dlm->node_num, res->lockname.len,
2054 					     res->lockname.name, new_master);
2055 				} else if (res->owner == dlm->node_num) {
2056 					mlog(0, "(this=%u) res %.*s owner=%u "
2057 					     "was not on recovering list, "
2058 					     "owner is THIS node, clearing\n",
2059 					     dlm->node_num, res->lockname.len,
2060 					     res->lockname.name, new_master);
2061 				} else
2062 					continue;
2063 
2064 				if (!list_empty(&res->recovering)) {
2065 					mlog(0, "%s:%.*s: lockres was "
2066 					     "marked RECOVERING, owner=%u\n",
2067 					     dlm->name, res->lockname.len,
2068 					     res->lockname.name, res->owner);
2069 					list_del_init(&res->recovering);
2070 					dlm_lockres_put(res);
2071 				}
2072 				spin_lock(&res->spinlock);
2073 				/* new_master has our reference from
2074 				 * the lock state sent during recovery */
2075 				dlm_change_lockres_owner(dlm, res, new_master);
2076 				res->state &= ~DLM_LOCK_RES_RECOVERING;
2077 				if (__dlm_lockres_has_locks(res))
2078 					__dlm_dirty_lockres(dlm, res);
2079 				spin_unlock(&res->spinlock);
2080 				wake_up(&res->wq);
2081 			}
2082 		}
2083 	}
2084 }
2085 
2086 static inline int dlm_lvb_needs_invalidation(struct dlm_lock *lock, int local)
2087 {
2088 	if (local) {
2089 		if (lock->ml.type != LKM_EXMODE &&
2090 		    lock->ml.type != LKM_PRMODE)
2091 			return 1;
2092 	} else if (lock->ml.type == LKM_EXMODE)
2093 		return 1;
2094 	return 0;
2095 }
2096 
2097 static void dlm_revalidate_lvb(struct dlm_ctxt *dlm,
2098 			       struct dlm_lock_resource *res, u8 dead_node)
2099 {
2100 	struct list_head *iter, *queue;
2101 	struct dlm_lock *lock;
2102 	int blank_lvb = 0, local = 0;
2103 	int i;
2104 	u8 search_node;
2105 
2106 	assert_spin_locked(&dlm->spinlock);
2107 	assert_spin_locked(&res->spinlock);
2108 
2109 	if (res->owner == dlm->node_num)
2110 		/* if this node owned the lockres, and if the dead node
2111 		 * had an EX when he died, blank out the lvb */
2112 		search_node = dead_node;
2113 	else {
2114 		/* if this is a secondary lockres, and we had no EX or PR
2115 		 * locks granted, we can no longer trust the lvb */
2116 		search_node = dlm->node_num;
2117 		local = 1;  /* check local state for valid lvb */
2118 	}
2119 
2120 	for (i=DLM_GRANTED_LIST; i<=DLM_CONVERTING_LIST; i++) {
2121 		queue = dlm_list_idx_to_ptr(res, i);
2122 		list_for_each(iter, queue) {
2123 			lock = list_entry (iter, struct dlm_lock, list);
2124 			if (lock->ml.node == search_node) {
2125 				if (dlm_lvb_needs_invalidation(lock, local)) {
2126 					/* zero the lksb lvb and lockres lvb */
2127 					blank_lvb = 1;
2128 					memset(lock->lksb->lvb, 0, DLM_LVB_LEN);
2129 				}
2130 			}
2131 		}
2132 	}
2133 
2134 	if (blank_lvb) {
2135 		mlog(0, "clearing %.*s lvb, dead node %u had EX\n",
2136 		     res->lockname.len, res->lockname.name, dead_node);
2137 		memset(res->lvb, 0, DLM_LVB_LEN);
2138 	}
2139 }
2140 
2141 static void dlm_free_dead_locks(struct dlm_ctxt *dlm,
2142 				struct dlm_lock_resource *res, u8 dead_node)
2143 {
2144 	struct list_head *iter, *tmpiter;
2145 	struct dlm_lock *lock;
2146 	unsigned int freed = 0;
2147 
2148 	/* this node is the lockres master:
2149 	 * 1) remove any stale locks for the dead node
2150 	 * 2) if the dead node had an EX when he died, blank out the lvb
2151 	 */
2152 	assert_spin_locked(&dlm->spinlock);
2153 	assert_spin_locked(&res->spinlock);
2154 
2155 	/* TODO: check pending_asts, pending_basts here */
2156 	list_for_each_safe(iter, tmpiter, &res->granted) {
2157 		lock = list_entry (iter, struct dlm_lock, list);
2158 		if (lock->ml.node == dead_node) {
2159 			list_del_init(&lock->list);
2160 			dlm_lock_put(lock);
2161 			freed++;
2162 		}
2163 	}
2164 	list_for_each_safe(iter, tmpiter, &res->converting) {
2165 		lock = list_entry (iter, struct dlm_lock, list);
2166 		if (lock->ml.node == dead_node) {
2167 			list_del_init(&lock->list);
2168 			dlm_lock_put(lock);
2169 			freed++;
2170 		}
2171 	}
2172 	list_for_each_safe(iter, tmpiter, &res->blocked) {
2173 		lock = list_entry (iter, struct dlm_lock, list);
2174 		if (lock->ml.node == dead_node) {
2175 			list_del_init(&lock->list);
2176 			dlm_lock_put(lock);
2177 			freed++;
2178 		}
2179 	}
2180 
2181 	if (freed) {
2182 		mlog(0, "%s:%.*s: freed %u locks for dead node %u, "
2183 		     "dropping ref from lockres\n", dlm->name,
2184 		     res->lockname.len, res->lockname.name, freed, dead_node);
2185 		BUG_ON(!test_bit(dead_node, res->refmap));
2186 		dlm_lockres_clear_refmap_bit(dead_node, res);
2187 	} else if (test_bit(dead_node, res->refmap)) {
2188 		mlog(0, "%s:%.*s: dead node %u had a ref, but had "
2189 		     "no locks and had not purged before dying\n", dlm->name,
2190 		     res->lockname.len, res->lockname.name, dead_node);
2191 		dlm_lockres_clear_refmap_bit(dead_node, res);
2192 	}
2193 
2194 	/* do not kick thread yet */
2195 	__dlm_dirty_lockres(dlm, res);
2196 }
2197 
2198 /* if this node is the recovery master, and there are no
2199  * locks for a given lockres owned by this node that are in
2200  * either PR or EX mode, zero out the lvb before requesting.
2201  *
2202  */
2203 
2204 
2205 static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node)
2206 {
2207 	struct hlist_node *iter;
2208 	struct dlm_lock_resource *res;
2209 	int i;
2210 	struct hlist_head *bucket;
2211 	struct dlm_lock *lock;
2212 
2213 
2214 	/* purge any stale mles */
2215 	dlm_clean_master_list(dlm, dead_node);
2216 
2217 	/*
2218 	 * now clean up all lock resources.  there are two rules:
2219 	 *
2220 	 * 1) if the dead node was the master, move the lockres
2221 	 *    to the recovering list.  set the RECOVERING flag.
2222 	 *    this lockres needs to be cleaned up before it can
2223 	 *    be used further.
2224 	 *
2225 	 * 2) if this node was the master, remove all locks from
2226 	 *    each of the lockres queues that were owned by the
2227 	 *    dead node.  once recovery finishes, the dlm thread
2228 	 *    can be kicked again to see if any ASTs or BASTs
2229 	 *    need to be fired as a result.
2230 	 */
2231 	for (i = 0; i < DLM_HASH_BUCKETS; i++) {
2232 		bucket = dlm_lockres_hash(dlm, i);
2233 		hlist_for_each_entry(res, iter, bucket, hash_node) {
2234  			/* always prune any $RECOVERY entries for dead nodes,
2235  			 * otherwise hangs can occur during later recovery */
2236 			if (dlm_is_recovery_lock(res->lockname.name,
2237 						 res->lockname.len)) {
2238 				spin_lock(&res->spinlock);
2239 				list_for_each_entry(lock, &res->granted, list) {
2240 					if (lock->ml.node == dead_node) {
2241 						mlog(0, "AHA! there was "
2242 						     "a $RECOVERY lock for dead "
2243 						     "node %u (%s)!\n",
2244 						     dead_node, dlm->name);
2245 						list_del_init(&lock->list);
2246 						dlm_lock_put(lock);
2247 						break;
2248 					}
2249 				}
2250 				spin_unlock(&res->spinlock);
2251 				continue;
2252 			}
2253 			spin_lock(&res->spinlock);
2254 			/* zero the lvb if necessary */
2255 			dlm_revalidate_lvb(dlm, res, dead_node);
2256 			if (res->owner == dead_node) {
2257 				if (res->state & DLM_LOCK_RES_DROPPING_REF)
2258 					mlog(0, "%s:%.*s: owned by "
2259 					     "dead node %u, this node was "
2260 					     "dropping its ref when it died. "
2261 					     "continue, dropping the flag.\n",
2262 					     dlm->name, res->lockname.len,
2263 					     res->lockname.name, dead_node);
2264 
2265 				/* the wake_up for this will happen when the
2266 				 * RECOVERING flag is dropped later */
2267 				res->state &= ~DLM_LOCK_RES_DROPPING_REF;
2268 
2269 				dlm_move_lockres_to_recovery_list(dlm, res);
2270 			} else if (res->owner == dlm->node_num) {
2271 				dlm_free_dead_locks(dlm, res, dead_node);
2272 				__dlm_lockres_calc_usage(dlm, res);
2273 			}
2274 			spin_unlock(&res->spinlock);
2275 		}
2276 	}
2277 
2278 }
2279 
2280 static void __dlm_hb_node_down(struct dlm_ctxt *dlm, int idx)
2281 {
2282 	assert_spin_locked(&dlm->spinlock);
2283 
2284 	if (dlm->reco.new_master == idx) {
2285 		mlog(0, "%s: recovery master %d just died\n",
2286 		     dlm->name, idx);
2287 		if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2288 			/* finalize1 was reached, so it is safe to clear
2289 			 * the new_master and dead_node.  that recovery
2290 			 * is complete. */
2291 			mlog(0, "%s: dead master %d had reached "
2292 			     "finalize1 state, clearing\n", dlm->name, idx);
2293 			dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE;
2294 			__dlm_reset_recovery(dlm);
2295 		}
2296 	}
2297 
2298 	/* check to see if the node is already considered dead */
2299 	if (!test_bit(idx, dlm->live_nodes_map)) {
2300 		mlog(0, "for domain %s, node %d is already dead. "
2301 		     "another node likely did recovery already.\n",
2302 		     dlm->name, idx);
2303 		return;
2304 	}
2305 
2306 	/* check to see if we do not care about this node */
2307 	if (!test_bit(idx, dlm->domain_map)) {
2308 		/* This also catches the case that we get a node down
2309 		 * but haven't joined the domain yet. */
2310 		mlog(0, "node %u already removed from domain!\n", idx);
2311 		return;
2312 	}
2313 
2314 	clear_bit(idx, dlm->live_nodes_map);
2315 
2316 	/* Clean up join state on node death. */
2317 	if (dlm->joining_node == idx) {
2318 		mlog(0, "Clearing join state for node %u\n", idx);
2319 		__dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
2320 	}
2321 
2322 	/* make sure local cleanup occurs before the heartbeat events */
2323 	if (!test_bit(idx, dlm->recovery_map))
2324 		dlm_do_local_recovery_cleanup(dlm, idx);
2325 
2326 	/* notify anything attached to the heartbeat events */
2327 	dlm_hb_event_notify_attached(dlm, idx, 0);
2328 
2329 	mlog(0, "node %u being removed from domain map!\n", idx);
2330 	clear_bit(idx, dlm->domain_map);
2331 	/* wake up migration waiters if a node goes down.
2332 	 * perhaps later we can genericize this for other waiters. */
2333 	wake_up(&dlm->migration_wq);
2334 
2335 	if (test_bit(idx, dlm->recovery_map))
2336 		mlog(0, "domain %s, node %u already added "
2337 		     "to recovery map!\n", dlm->name, idx);
2338 	else
2339 		set_bit(idx, dlm->recovery_map);
2340 }
2341 
2342 void dlm_hb_node_down_cb(struct o2nm_node *node, int idx, void *data)
2343 {
2344 	struct dlm_ctxt *dlm = data;
2345 
2346 	if (!dlm_grab(dlm))
2347 		return;
2348 
2349 	spin_lock(&dlm->spinlock);
2350 	__dlm_hb_node_down(dlm, idx);
2351 	spin_unlock(&dlm->spinlock);
2352 
2353 	dlm_put(dlm);
2354 }
2355 
2356 void dlm_hb_node_up_cb(struct o2nm_node *node, int idx, void *data)
2357 {
2358 	struct dlm_ctxt *dlm = data;
2359 
2360 	if (!dlm_grab(dlm))
2361 		return;
2362 
2363 	spin_lock(&dlm->spinlock);
2364 	set_bit(idx, dlm->live_nodes_map);
2365 	/* do NOT notify mle attached to the heartbeat events.
2366 	 * new nodes are not interesting in mastery until joined. */
2367 	spin_unlock(&dlm->spinlock);
2368 
2369 	dlm_put(dlm);
2370 }
2371 
2372 static void dlm_reco_ast(void *astdata)
2373 {
2374 	struct dlm_ctxt *dlm = astdata;
2375 	mlog(0, "ast for recovery lock fired!, this=%u, dlm=%s\n",
2376 	     dlm->node_num, dlm->name);
2377 }
2378 static void dlm_reco_bast(void *astdata, int blocked_type)
2379 {
2380 	struct dlm_ctxt *dlm = astdata;
2381 	mlog(0, "bast for recovery lock fired!, this=%u, dlm=%s\n",
2382 	     dlm->node_num, dlm->name);
2383 }
2384 static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st)
2385 {
2386 	mlog(0, "unlockast for recovery lock fired!\n");
2387 }
2388 
2389 /*
2390  * dlm_pick_recovery_master will continually attempt to use
2391  * dlmlock() on the special "$RECOVERY" lockres with the
2392  * LKM_NOQUEUE flag to get an EX.  every thread that enters
2393  * this function on each node racing to become the recovery
2394  * master will not stop attempting this until either:
2395  * a) this node gets the EX (and becomes the recovery master),
2396  * or b) dlm->reco.new_master gets set to some nodenum
2397  * != O2NM_INVALID_NODE_NUM (another node will do the reco).
2398  * so each time a recovery master is needed, the entire cluster
2399  * will sync at this point.  if the new master dies, that will
2400  * be detected in dlm_do_recovery */
2401 static int dlm_pick_recovery_master(struct dlm_ctxt *dlm)
2402 {
2403 	enum dlm_status ret;
2404 	struct dlm_lockstatus lksb;
2405 	int status = -EINVAL;
2406 
2407 	mlog(0, "starting recovery of %s at %lu, dead=%u, this=%u\n",
2408 	     dlm->name, jiffies, dlm->reco.dead_node, dlm->node_num);
2409 again:
2410 	memset(&lksb, 0, sizeof(lksb));
2411 
2412 	ret = dlmlock(dlm, LKM_EXMODE, &lksb, LKM_NOQUEUE|LKM_RECOVERY,
2413 		      DLM_RECOVERY_LOCK_NAME, DLM_RECOVERY_LOCK_NAME_LEN,
2414 		      dlm_reco_ast, dlm, dlm_reco_bast);
2415 
2416 	mlog(0, "%s: dlmlock($RECOVERY) returned %d, lksb=%d\n",
2417 	     dlm->name, ret, lksb.status);
2418 
2419 	if (ret == DLM_NORMAL) {
2420 		mlog(0, "dlm=%s dlmlock says I got it (this=%u)\n",
2421 		     dlm->name, dlm->node_num);
2422 
2423 		/* got the EX lock.  check to see if another node
2424 		 * just became the reco master */
2425 		if (dlm_reco_master_ready(dlm)) {
2426 			mlog(0, "%s: got reco EX lock, but %u will "
2427 			     "do the recovery\n", dlm->name,
2428 			     dlm->reco.new_master);
2429 			status = -EEXIST;
2430 		} else {
2431 			status = 0;
2432 
2433 			/* see if recovery was already finished elsewhere */
2434 			spin_lock(&dlm->spinlock);
2435 			if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
2436 				status = -EINVAL;
2437 				mlog(0, "%s: got reco EX lock, but "
2438 				     "node got recovered already\n", dlm->name);
2439 				if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) {
2440 					mlog(ML_ERROR, "%s: new master is %u "
2441 					     "but no dead node!\n",
2442 					     dlm->name, dlm->reco.new_master);
2443 					BUG();
2444 				}
2445 			}
2446 			spin_unlock(&dlm->spinlock);
2447 		}
2448 
2449 		/* if this node has actually become the recovery master,
2450 		 * set the master and send the messages to begin recovery */
2451 		if (!status) {
2452 			mlog(0, "%s: dead=%u, this=%u, sending "
2453 			     "begin_reco now\n", dlm->name,
2454 			     dlm->reco.dead_node, dlm->node_num);
2455 			status = dlm_send_begin_reco_message(dlm,
2456 				      dlm->reco.dead_node);
2457 			/* this always succeeds */
2458 			BUG_ON(status);
2459 
2460 			/* set the new_master to this node */
2461 			spin_lock(&dlm->spinlock);
2462 			dlm_set_reco_master(dlm, dlm->node_num);
2463 			spin_unlock(&dlm->spinlock);
2464 		}
2465 
2466 		/* recovery lock is a special case.  ast will not get fired,
2467 		 * so just go ahead and unlock it. */
2468 		ret = dlmunlock(dlm, &lksb, 0, dlm_reco_unlock_ast, dlm);
2469 		if (ret == DLM_DENIED) {
2470 			mlog(0, "got DLM_DENIED, trying LKM_CANCEL\n");
2471 			ret = dlmunlock(dlm, &lksb, LKM_CANCEL, dlm_reco_unlock_ast, dlm);
2472 		}
2473 		if (ret != DLM_NORMAL) {
2474 			/* this would really suck. this could only happen
2475 			 * if there was a network error during the unlock
2476 			 * because of node death.  this means the unlock
2477 			 * is actually "done" and the lock structure is
2478 			 * even freed.  we can continue, but only
2479 			 * because this specific lock name is special. */
2480 			mlog(ML_ERROR, "dlmunlock returned %d\n", ret);
2481 		}
2482 	} else if (ret == DLM_NOTQUEUED) {
2483 		mlog(0, "dlm=%s dlmlock says another node got it (this=%u)\n",
2484 		     dlm->name, dlm->node_num);
2485 		/* another node is master. wait on
2486 		 * reco.new_master != O2NM_INVALID_NODE_NUM
2487 		 * for at most one second */
2488 		wait_event_timeout(dlm->dlm_reco_thread_wq,
2489 					 dlm_reco_master_ready(dlm),
2490 					 msecs_to_jiffies(1000));
2491 		if (!dlm_reco_master_ready(dlm)) {
2492 			mlog(0, "%s: reco master taking awhile\n",
2493 			     dlm->name);
2494 			goto again;
2495 		}
2496 		/* another node has informed this one that it is reco master */
2497 		mlog(0, "%s: reco master %u is ready to recover %u\n",
2498 		     dlm->name, dlm->reco.new_master, dlm->reco.dead_node);
2499 		status = -EEXIST;
2500 	} else if (ret == DLM_RECOVERING) {
2501 		mlog(0, "dlm=%s dlmlock says master node died (this=%u)\n",
2502 		     dlm->name, dlm->node_num);
2503 		goto again;
2504 	} else {
2505 		struct dlm_lock_resource *res;
2506 
2507 		/* dlmlock returned something other than NOTQUEUED or NORMAL */
2508 		mlog(ML_ERROR, "%s: got %s from dlmlock($RECOVERY), "
2509 		     "lksb.status=%s\n", dlm->name, dlm_errname(ret),
2510 		     dlm_errname(lksb.status));
2511 		res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME,
2512 					 DLM_RECOVERY_LOCK_NAME_LEN);
2513 		if (res) {
2514 			dlm_print_one_lock_resource(res);
2515 			dlm_lockres_put(res);
2516 		} else {
2517 			mlog(ML_ERROR, "recovery lock not found\n");
2518 		}
2519 		BUG();
2520 	}
2521 
2522 	return status;
2523 }
2524 
2525 static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node)
2526 {
2527 	struct dlm_begin_reco br;
2528 	int ret = 0;
2529 	struct dlm_node_iter iter;
2530 	int nodenum;
2531 	int status;
2532 
2533 	mlog_entry("%u\n", dead_node);
2534 
2535 	mlog(0, "%s: dead node is %u\n", dlm->name, dead_node);
2536 
2537 	spin_lock(&dlm->spinlock);
2538 	dlm_node_iter_init(dlm->domain_map, &iter);
2539 	spin_unlock(&dlm->spinlock);
2540 
2541 	clear_bit(dead_node, iter.node_map);
2542 
2543 	memset(&br, 0, sizeof(br));
2544 	br.node_idx = dlm->node_num;
2545 	br.dead_node = dead_node;
2546 
2547 	while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2548 		ret = 0;
2549 		if (nodenum == dead_node) {
2550 			mlog(0, "not sending begin reco to dead node "
2551 				  "%u\n", dead_node);
2552 			continue;
2553 		}
2554 		if (nodenum == dlm->node_num) {
2555 			mlog(0, "not sending begin reco to self\n");
2556 			continue;
2557 		}
2558 retry:
2559 		ret = -EINVAL;
2560 		mlog(0, "attempting to send begin reco msg to %d\n",
2561 			  nodenum);
2562 		ret = o2net_send_message(DLM_BEGIN_RECO_MSG, dlm->key,
2563 					 &br, sizeof(br), nodenum, &status);
2564 		/* negative status is handled ok by caller here */
2565 		if (ret >= 0)
2566 			ret = status;
2567 		if (dlm_is_host_down(ret)) {
2568 			/* node is down.  not involved in recovery
2569 			 * so just keep going */
2570 			mlog(0, "%s: node %u was down when sending "
2571 			     "begin reco msg (%d)\n", dlm->name, nodenum, ret);
2572 			ret = 0;
2573 		}
2574 		if (ret < 0) {
2575 			struct dlm_lock_resource *res;
2576 			/* this is now a serious problem, possibly ENOMEM
2577 			 * in the network stack.  must retry */
2578 			mlog_errno(ret);
2579 			mlog(ML_ERROR, "begin reco of dlm %s to node %u "
2580 			    " returned %d\n", dlm->name, nodenum, ret);
2581 			res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME,
2582 						 DLM_RECOVERY_LOCK_NAME_LEN);
2583 			if (res) {
2584 				dlm_print_one_lock_resource(res);
2585 				dlm_lockres_put(res);
2586 			} else {
2587 				mlog(ML_ERROR, "recovery lock not found\n");
2588 			}
2589 			/* sleep for a bit in hopes that we can avoid
2590 			 * another ENOMEM */
2591 			msleep(100);
2592 			goto retry;
2593 		} else if (ret == EAGAIN) {
2594 			mlog(0, "%s: trying to start recovery of node "
2595 			     "%u, but node %u is waiting for last recovery "
2596 			     "to complete, backoff for a bit\n", dlm->name,
2597 			     dead_node, nodenum);
2598 			/* TODO Look into replacing msleep with cond_resched() */
2599 			msleep(100);
2600 			goto retry;
2601 		}
2602 	}
2603 
2604 	return ret;
2605 }
2606 
2607 int dlm_begin_reco_handler(struct o2net_msg *msg, u32 len, void *data,
2608 			   void **ret_data)
2609 {
2610 	struct dlm_ctxt *dlm = data;
2611 	struct dlm_begin_reco *br = (struct dlm_begin_reco *)msg->buf;
2612 
2613 	/* ok to return 0, domain has gone away */
2614 	if (!dlm_grab(dlm))
2615 		return 0;
2616 
2617 	spin_lock(&dlm->spinlock);
2618 	if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2619 		mlog(0, "%s: node %u wants to recover node %u (%u:%u) "
2620 		     "but this node is in finalize state, waiting on finalize2\n",
2621 		     dlm->name, br->node_idx, br->dead_node,
2622 		     dlm->reco.dead_node, dlm->reco.new_master);
2623 		spin_unlock(&dlm->spinlock);
2624 		return EAGAIN;
2625 	}
2626 	spin_unlock(&dlm->spinlock);
2627 
2628 	mlog(0, "%s: node %u wants to recover node %u (%u:%u)\n",
2629 	     dlm->name, br->node_idx, br->dead_node,
2630 	     dlm->reco.dead_node, dlm->reco.new_master);
2631 
2632 	dlm_fire_domain_eviction_callbacks(dlm, br->dead_node);
2633 
2634 	spin_lock(&dlm->spinlock);
2635 	if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) {
2636 		if (test_bit(dlm->reco.new_master, dlm->recovery_map)) {
2637 			mlog(0, "%s: new_master %u died, changing "
2638 			     "to %u\n", dlm->name, dlm->reco.new_master,
2639 			     br->node_idx);
2640 		} else {
2641 			mlog(0, "%s: new_master %u NOT DEAD, changing "
2642 			     "to %u\n", dlm->name, dlm->reco.new_master,
2643 			     br->node_idx);
2644 			/* may not have seen the new master as dead yet */
2645 		}
2646 	}
2647 	if (dlm->reco.dead_node != O2NM_INVALID_NODE_NUM) {
2648 		mlog(ML_NOTICE, "%s: dead_node previously set to %u, "
2649 		     "node %u changing it to %u\n", dlm->name,
2650 		     dlm->reco.dead_node, br->node_idx, br->dead_node);
2651 	}
2652 	dlm_set_reco_master(dlm, br->node_idx);
2653 	dlm_set_reco_dead_node(dlm, br->dead_node);
2654 	if (!test_bit(br->dead_node, dlm->recovery_map)) {
2655 		mlog(0, "recovery master %u sees %u as dead, but this "
2656 		     "node has not yet.  marking %u as dead\n",
2657 		     br->node_idx, br->dead_node, br->dead_node);
2658 		if (!test_bit(br->dead_node, dlm->domain_map) ||
2659 		    !test_bit(br->dead_node, dlm->live_nodes_map))
2660 			mlog(0, "%u not in domain/live_nodes map "
2661 			     "so setting it in reco map manually\n",
2662 			     br->dead_node);
2663 		/* force the recovery cleanup in __dlm_hb_node_down
2664 		 * both of these will be cleared in a moment */
2665 		set_bit(br->dead_node, dlm->domain_map);
2666 		set_bit(br->dead_node, dlm->live_nodes_map);
2667 		__dlm_hb_node_down(dlm, br->dead_node);
2668 	}
2669 	spin_unlock(&dlm->spinlock);
2670 
2671 	dlm_kick_recovery_thread(dlm);
2672 
2673 	mlog(0, "%s: recovery started by node %u, for %u (%u:%u)\n",
2674 	     dlm->name, br->node_idx, br->dead_node,
2675 	     dlm->reco.dead_node, dlm->reco.new_master);
2676 
2677 	dlm_put(dlm);
2678 	return 0;
2679 }
2680 
2681 #define DLM_FINALIZE_STAGE2  0x01
2682 static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm)
2683 {
2684 	int ret = 0;
2685 	struct dlm_finalize_reco fr;
2686 	struct dlm_node_iter iter;
2687 	int nodenum;
2688 	int status;
2689 	int stage = 1;
2690 
2691 	mlog(0, "finishing recovery for node %s:%u, "
2692 	     "stage %d\n", dlm->name, dlm->reco.dead_node, stage);
2693 
2694 	spin_lock(&dlm->spinlock);
2695 	dlm_node_iter_init(dlm->domain_map, &iter);
2696 	spin_unlock(&dlm->spinlock);
2697 
2698 stage2:
2699 	memset(&fr, 0, sizeof(fr));
2700 	fr.node_idx = dlm->node_num;
2701 	fr.dead_node = dlm->reco.dead_node;
2702 	if (stage == 2)
2703 		fr.flags |= DLM_FINALIZE_STAGE2;
2704 
2705 	while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2706 		if (nodenum == dlm->node_num)
2707 			continue;
2708 		ret = o2net_send_message(DLM_FINALIZE_RECO_MSG, dlm->key,
2709 					 &fr, sizeof(fr), nodenum, &status);
2710 		if (ret >= 0)
2711 			ret = status;
2712 		if (ret < 0) {
2713 			mlog_errno(ret);
2714 			if (dlm_is_host_down(ret)) {
2715 				/* this has no effect on this recovery
2716 				 * session, so set the status to zero to
2717 				 * finish out the last recovery */
2718 				mlog(ML_ERROR, "node %u went down after this "
2719 				     "node finished recovery.\n", nodenum);
2720 				ret = 0;
2721 				continue;
2722 			}
2723 			break;
2724 		}
2725 	}
2726 	if (stage == 1) {
2727 		/* reset the node_iter back to the top and send finalize2 */
2728 		iter.curnode = -1;
2729 		stage = 2;
2730 		goto stage2;
2731 	}
2732 
2733 	return ret;
2734 }
2735 
2736 int dlm_finalize_reco_handler(struct o2net_msg *msg, u32 len, void *data,
2737 			      void **ret_data)
2738 {
2739 	struct dlm_ctxt *dlm = data;
2740 	struct dlm_finalize_reco *fr = (struct dlm_finalize_reco *)msg->buf;
2741 	int stage = 1;
2742 
2743 	/* ok to return 0, domain has gone away */
2744 	if (!dlm_grab(dlm))
2745 		return 0;
2746 
2747 	if (fr->flags & DLM_FINALIZE_STAGE2)
2748 		stage = 2;
2749 
2750 	mlog(0, "%s: node %u finalizing recovery stage%d of "
2751 	     "node %u (%u:%u)\n", dlm->name, fr->node_idx, stage,
2752 	     fr->dead_node, dlm->reco.dead_node, dlm->reco.new_master);
2753 
2754 	spin_lock(&dlm->spinlock);
2755 
2756 	if (dlm->reco.new_master != fr->node_idx) {
2757 		mlog(ML_ERROR, "node %u sent recovery finalize msg, but node "
2758 		     "%u is supposed to be the new master, dead=%u\n",
2759 		     fr->node_idx, dlm->reco.new_master, fr->dead_node);
2760 		BUG();
2761 	}
2762 	if (dlm->reco.dead_node != fr->dead_node) {
2763 		mlog(ML_ERROR, "node %u sent recovery finalize msg for dead "
2764 		     "node %u, but node %u is supposed to be dead\n",
2765 		     fr->node_idx, fr->dead_node, dlm->reco.dead_node);
2766 		BUG();
2767 	}
2768 
2769 	switch (stage) {
2770 		case 1:
2771 			dlm_finish_local_lockres_recovery(dlm, fr->dead_node, fr->node_idx);
2772 			if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2773 				mlog(ML_ERROR, "%s: received finalize1 from "
2774 				     "new master %u for dead node %u, but "
2775 				     "this node has already received it!\n",
2776 				     dlm->name, fr->node_idx, fr->dead_node);
2777 				dlm_print_reco_node_status(dlm);
2778 				BUG();
2779 			}
2780 			dlm->reco.state |= DLM_RECO_STATE_FINALIZE;
2781 			spin_unlock(&dlm->spinlock);
2782 			break;
2783 		case 2:
2784 			if (!(dlm->reco.state & DLM_RECO_STATE_FINALIZE)) {
2785 				mlog(ML_ERROR, "%s: received finalize2 from "
2786 				     "new master %u for dead node %u, but "
2787 				     "this node did not have finalize1!\n",
2788 				     dlm->name, fr->node_idx, fr->dead_node);
2789 				dlm_print_reco_node_status(dlm);
2790 				BUG();
2791 			}
2792 			dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE;
2793 			spin_unlock(&dlm->spinlock);
2794 			dlm_reset_recovery(dlm);
2795 			dlm_kick_recovery_thread(dlm);
2796 			break;
2797 		default:
2798 			BUG();
2799 	}
2800 
2801 	mlog(0, "%s: recovery done, reco master was %u, dead now %u, master now %u\n",
2802 	     dlm->name, fr->node_idx, dlm->reco.dead_node, dlm->reco.new_master);
2803 
2804 	dlm_put(dlm);
2805 	return 0;
2806 }
2807