xref: /openbmc/linux/fs/ocfs2/dlm/dlmrecovery.c (revision 643d1f7f)
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 dlm_work_item *item, *next;
162 	dlm_workfunc_t *workfunc;
163 	int tot=0;
164 
165 	spin_lock(&dlm->work_lock);
166 	list_splice_init(&dlm->work_list, &tmp_list);
167 	spin_unlock(&dlm->work_lock);
168 
169 	list_for_each_entry(item, &tmp_list, list) {
170 		tot++;
171 	}
172 	mlog(0, "%s: work thread has %d work items\n", dlm->name, tot);
173 
174 	list_for_each_entry_safe(item, next, &tmp_list, list) {
175 		workfunc = item->func;
176 		list_del_init(&item->list);
177 
178 		/* already have ref on dlm to avoid having
179 		 * it disappear.  just double-check. */
180 		BUG_ON(item->dlm != dlm);
181 
182 		/* this is allowed to sleep and
183 		 * call network stuff */
184 		workfunc(item, item->data);
185 
186 		dlm_put(dlm);
187 		kfree(item);
188 	}
189 }
190 
191 /*
192  * RECOVERY THREAD
193  */
194 
195 void dlm_kick_recovery_thread(struct dlm_ctxt *dlm)
196 {
197 	/* wake the recovery thread
198 	 * this will wake the reco thread in one of three places
199 	 * 1) sleeping with no recovery happening
200 	 * 2) sleeping with recovery mastered elsewhere
201 	 * 3) recovery mastered here, waiting on reco data */
202 
203 	wake_up(&dlm->dlm_reco_thread_wq);
204 }
205 
206 /* Launch the recovery thread */
207 int dlm_launch_recovery_thread(struct dlm_ctxt *dlm)
208 {
209 	mlog(0, "starting dlm recovery thread...\n");
210 
211 	dlm->dlm_reco_thread_task = kthread_run(dlm_recovery_thread, dlm,
212 						"dlm_reco_thread");
213 	if (IS_ERR(dlm->dlm_reco_thread_task)) {
214 		mlog_errno(PTR_ERR(dlm->dlm_reco_thread_task));
215 		dlm->dlm_reco_thread_task = NULL;
216 		return -EINVAL;
217 	}
218 
219 	return 0;
220 }
221 
222 void dlm_complete_recovery_thread(struct dlm_ctxt *dlm)
223 {
224 	if (dlm->dlm_reco_thread_task) {
225 		mlog(0, "waiting for dlm recovery thread to exit\n");
226 		kthread_stop(dlm->dlm_reco_thread_task);
227 		dlm->dlm_reco_thread_task = NULL;
228 	}
229 }
230 
231 
232 
233 /*
234  * this is lame, but here's how recovery works...
235  * 1) all recovery threads cluster wide will work on recovering
236  *    ONE node at a time
237  * 2) negotiate who will take over all the locks for the dead node.
238  *    thats right... ALL the locks.
239  * 3) once a new master is chosen, everyone scans all locks
240  *    and moves aside those mastered by the dead guy
241  * 4) each of these locks should be locked until recovery is done
242  * 5) the new master collects up all of secondary lock queue info
243  *    one lock at a time, forcing each node to communicate back
244  *    before continuing
245  * 6) each secondary lock queue responds with the full known lock info
246  * 7) once the new master has run all its locks, it sends a ALLDONE!
247  *    message to everyone
248  * 8) upon receiving this message, the secondary queue node unlocks
249  *    and responds to the ALLDONE
250  * 9) once the new master gets responses from everyone, he unlocks
251  *    everything and recovery for this dead node is done
252  *10) go back to 2) while there are still dead nodes
253  *
254  */
255 
256 static void dlm_print_reco_node_status(struct dlm_ctxt *dlm)
257 {
258 	struct dlm_reco_node_data *ndata;
259 	struct dlm_lock_resource *res;
260 
261 	mlog(ML_NOTICE, "%s(%d): recovery info, state=%s, dead=%u, master=%u\n",
262 	     dlm->name, task_pid_nr(dlm->dlm_reco_thread_task),
263 	     dlm->reco.state & DLM_RECO_STATE_ACTIVE ? "ACTIVE" : "inactive",
264 	     dlm->reco.dead_node, dlm->reco.new_master);
265 
266 	list_for_each_entry(ndata, &dlm->reco.node_data, list) {
267 		char *st = "unknown";
268 		switch (ndata->state) {
269 			case DLM_RECO_NODE_DATA_INIT:
270 				st = "init";
271 				break;
272 			case DLM_RECO_NODE_DATA_REQUESTING:
273 				st = "requesting";
274 				break;
275 			case DLM_RECO_NODE_DATA_DEAD:
276 				st = "dead";
277 				break;
278 			case DLM_RECO_NODE_DATA_RECEIVING:
279 				st = "receiving";
280 				break;
281 			case DLM_RECO_NODE_DATA_REQUESTED:
282 				st = "requested";
283 				break;
284 			case DLM_RECO_NODE_DATA_DONE:
285 				st = "done";
286 				break;
287 			case DLM_RECO_NODE_DATA_FINALIZE_SENT:
288 				st = "finalize-sent";
289 				break;
290 			default:
291 				st = "bad";
292 				break;
293 		}
294 		mlog(ML_NOTICE, "%s: reco state, node %u, state=%s\n",
295 		     dlm->name, ndata->node_num, st);
296 	}
297 	list_for_each_entry(res, &dlm->reco.resources, recovering) {
298 		mlog(ML_NOTICE, "%s: lockres %.*s on recovering list\n",
299 		     dlm->name, res->lockname.len, res->lockname.name);
300 	}
301 }
302 
303 #define DLM_RECO_THREAD_TIMEOUT_MS (5 * 1000)
304 
305 static int dlm_recovery_thread(void *data)
306 {
307 	int status;
308 	struct dlm_ctxt *dlm = data;
309 	unsigned long timeout = msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS);
310 
311 	mlog(0, "dlm thread running for %s...\n", dlm->name);
312 
313 	while (!kthread_should_stop()) {
314 		if (dlm_joined(dlm)) {
315 			status = dlm_do_recovery(dlm);
316 			if (status == -EAGAIN) {
317 				/* do not sleep, recheck immediately. */
318 				continue;
319 			}
320 			if (status < 0)
321 				mlog_errno(status);
322 		}
323 
324 		wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq,
325 						 kthread_should_stop(),
326 						 timeout);
327 	}
328 
329 	mlog(0, "quitting DLM recovery thread\n");
330 	return 0;
331 }
332 
333 /* returns true when the recovery master has contacted us */
334 static int dlm_reco_master_ready(struct dlm_ctxt *dlm)
335 {
336 	int ready;
337 	spin_lock(&dlm->spinlock);
338 	ready = (dlm->reco.new_master != O2NM_INVALID_NODE_NUM);
339 	spin_unlock(&dlm->spinlock);
340 	return ready;
341 }
342 
343 /* returns true if node is no longer in the domain
344  * could be dead or just not joined */
345 int dlm_is_node_dead(struct dlm_ctxt *dlm, u8 node)
346 {
347 	int dead;
348 	spin_lock(&dlm->spinlock);
349 	dead = !test_bit(node, dlm->domain_map);
350 	spin_unlock(&dlm->spinlock);
351 	return dead;
352 }
353 
354 /* returns true if node is no longer in the domain
355  * could be dead or just not joined */
356 static int dlm_is_node_recovered(struct dlm_ctxt *dlm, u8 node)
357 {
358 	int recovered;
359 	spin_lock(&dlm->spinlock);
360 	recovered = !test_bit(node, dlm->recovery_map);
361 	spin_unlock(&dlm->spinlock);
362 	return recovered;
363 }
364 
365 
366 int dlm_wait_for_node_death(struct dlm_ctxt *dlm, u8 node, int timeout)
367 {
368 	if (timeout) {
369 		mlog(ML_NOTICE, "%s: waiting %dms for notification of "
370 		     "death of node %u\n", dlm->name, timeout, node);
371 		wait_event_timeout(dlm->dlm_reco_thread_wq,
372 			   dlm_is_node_dead(dlm, node),
373 			   msecs_to_jiffies(timeout));
374 	} else {
375 		mlog(ML_NOTICE, "%s: waiting indefinitely for notification "
376 		     "of death of node %u\n", dlm->name, node);
377 		wait_event(dlm->dlm_reco_thread_wq,
378 			   dlm_is_node_dead(dlm, node));
379 	}
380 	/* for now, return 0 */
381 	return 0;
382 }
383 
384 int dlm_wait_for_node_recovery(struct dlm_ctxt *dlm, u8 node, int timeout)
385 {
386 	if (timeout) {
387 		mlog(0, "%s: waiting %dms for notification of "
388 		     "recovery of node %u\n", dlm->name, timeout, node);
389 		wait_event_timeout(dlm->dlm_reco_thread_wq,
390 			   dlm_is_node_recovered(dlm, node),
391 			   msecs_to_jiffies(timeout));
392 	} else {
393 		mlog(0, "%s: waiting indefinitely for notification "
394 		     "of recovery of node %u\n", dlm->name, node);
395 		wait_event(dlm->dlm_reco_thread_wq,
396 			   dlm_is_node_recovered(dlm, node));
397 	}
398 	/* for now, return 0 */
399 	return 0;
400 }
401 
402 /* callers of the top-level api calls (dlmlock/dlmunlock) should
403  * block on the dlm->reco.event when recovery is in progress.
404  * the dlm recovery thread will set this state when it begins
405  * recovering a dead node (as the new master or not) and clear
406  * the state and wake as soon as all affected lock resources have
407  * been marked with the RECOVERY flag */
408 static int dlm_in_recovery(struct dlm_ctxt *dlm)
409 {
410 	int in_recovery;
411 	spin_lock(&dlm->spinlock);
412 	in_recovery = !!(dlm->reco.state & DLM_RECO_STATE_ACTIVE);
413 	spin_unlock(&dlm->spinlock);
414 	return in_recovery;
415 }
416 
417 
418 void dlm_wait_for_recovery(struct dlm_ctxt *dlm)
419 {
420 	if (dlm_in_recovery(dlm)) {
421 		mlog(0, "%s: reco thread %d in recovery: "
422 		     "state=%d, master=%u, dead=%u\n",
423 		     dlm->name, task_pid_nr(dlm->dlm_reco_thread_task),
424 		     dlm->reco.state, dlm->reco.new_master,
425 		     dlm->reco.dead_node);
426 	}
427 	wait_event(dlm->reco.event, !dlm_in_recovery(dlm));
428 }
429 
430 static void dlm_begin_recovery(struct dlm_ctxt *dlm)
431 {
432 	spin_lock(&dlm->spinlock);
433 	BUG_ON(dlm->reco.state & DLM_RECO_STATE_ACTIVE);
434 	dlm->reco.state |= DLM_RECO_STATE_ACTIVE;
435 	spin_unlock(&dlm->spinlock);
436 }
437 
438 static void dlm_end_recovery(struct dlm_ctxt *dlm)
439 {
440 	spin_lock(&dlm->spinlock);
441 	BUG_ON(!(dlm->reco.state & DLM_RECO_STATE_ACTIVE));
442 	dlm->reco.state &= ~DLM_RECO_STATE_ACTIVE;
443 	spin_unlock(&dlm->spinlock);
444 	wake_up(&dlm->reco.event);
445 }
446 
447 static int dlm_do_recovery(struct dlm_ctxt *dlm)
448 {
449 	int status = 0;
450 	int ret;
451 
452 	spin_lock(&dlm->spinlock);
453 
454 	/* check to see if the new master has died */
455 	if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM &&
456 	    test_bit(dlm->reco.new_master, dlm->recovery_map)) {
457 		mlog(0, "new master %u died while recovering %u!\n",
458 		     dlm->reco.new_master, dlm->reco.dead_node);
459 		/* unset the new_master, leave dead_node */
460 		dlm_set_reco_master(dlm, O2NM_INVALID_NODE_NUM);
461 	}
462 
463 	/* select a target to recover */
464 	if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
465 		int bit;
466 
467 		bit = find_next_bit (dlm->recovery_map, O2NM_MAX_NODES+1, 0);
468 		if (bit >= O2NM_MAX_NODES || bit < 0)
469 			dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
470 		else
471 			dlm_set_reco_dead_node(dlm, bit);
472 	} else if (!test_bit(dlm->reco.dead_node, dlm->recovery_map)) {
473 		/* BUG? */
474 		mlog(ML_ERROR, "dead_node %u no longer in recovery map!\n",
475 		     dlm->reco.dead_node);
476 		dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
477 	}
478 
479 	if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
480 		// mlog(0, "nothing to recover!  sleeping now!\n");
481 		spin_unlock(&dlm->spinlock);
482 		/* return to main thread loop and sleep. */
483 		return 0;
484 	}
485 	mlog(0, "%s(%d):recovery thread found node %u in the recovery map!\n",
486 	     dlm->name, task_pid_nr(dlm->dlm_reco_thread_task),
487 	     dlm->reco.dead_node);
488 	spin_unlock(&dlm->spinlock);
489 
490 	/* take write barrier */
491 	/* (stops the list reshuffling thread, proxy ast handling) */
492 	dlm_begin_recovery(dlm);
493 
494 	if (dlm->reco.new_master == dlm->node_num)
495 		goto master_here;
496 
497 	if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) {
498 		/* choose a new master, returns 0 if this node
499 		 * is the master, -EEXIST if it's another node.
500 		 * this does not return until a new master is chosen
501 		 * or recovery completes entirely. */
502 		ret = dlm_pick_recovery_master(dlm);
503 		if (!ret) {
504 			/* already notified everyone.  go. */
505 			goto master_here;
506 		}
507 		mlog(0, "another node will master this recovery session.\n");
508 	}
509 	mlog(0, "dlm=%s (%d), new_master=%u, this node=%u, dead_node=%u\n",
510 	     dlm->name, task_pid_nr(dlm->dlm_reco_thread_task), dlm->reco.new_master,
511 	     dlm->node_num, dlm->reco.dead_node);
512 
513 	/* it is safe to start everything back up here
514 	 * because all of the dead node's lock resources
515 	 * have been marked as in-recovery */
516 	dlm_end_recovery(dlm);
517 
518 	/* sleep out in main dlm_recovery_thread loop. */
519 	return 0;
520 
521 master_here:
522 	mlog(0, "(%d) mastering recovery of %s:%u here(this=%u)!\n",
523 	     task_pid_nr(dlm->dlm_reco_thread_task),
524 	     dlm->name, dlm->reco.dead_node, dlm->node_num);
525 
526 	status = dlm_remaster_locks(dlm, dlm->reco.dead_node);
527 	if (status < 0) {
528 		/* we should never hit this anymore */
529 		mlog(ML_ERROR, "error %d remastering locks for node %u, "
530 		     "retrying.\n", status, dlm->reco.dead_node);
531 		/* yield a bit to allow any final network messages
532 		 * to get handled on remaining nodes */
533 		msleep(100);
534 	} else {
535 		/* success!  see if any other nodes need recovery */
536 		mlog(0, "DONE mastering recovery of %s:%u here(this=%u)!\n",
537 		     dlm->name, dlm->reco.dead_node, dlm->node_num);
538 		dlm_reset_recovery(dlm);
539 	}
540 	dlm_end_recovery(dlm);
541 
542 	/* continue and look for another dead node */
543 	return -EAGAIN;
544 }
545 
546 static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node)
547 {
548 	int status = 0;
549 	struct dlm_reco_node_data *ndata;
550 	int all_nodes_done;
551 	int destroy = 0;
552 	int pass = 0;
553 
554 	do {
555 		/* we have become recovery master.  there is no escaping
556 		 * this, so just keep trying until we get it. */
557 		status = dlm_init_recovery_area(dlm, dead_node);
558 		if (status < 0) {
559 			mlog(ML_ERROR, "%s: failed to alloc recovery area, "
560 			     "retrying\n", dlm->name);
561 			msleep(1000);
562 		}
563 	} while (status != 0);
564 
565 	/* safe to access the node data list without a lock, since this
566 	 * process is the only one to change the list */
567 	list_for_each_entry(ndata, &dlm->reco.node_data, list) {
568 		BUG_ON(ndata->state != DLM_RECO_NODE_DATA_INIT);
569 		ndata->state = DLM_RECO_NODE_DATA_REQUESTING;
570 
571 		mlog(0, "requesting lock info from node %u\n",
572 		     ndata->node_num);
573 
574 		if (ndata->node_num == dlm->node_num) {
575 			ndata->state = DLM_RECO_NODE_DATA_DONE;
576 			continue;
577 		}
578 
579 		do {
580 			status = dlm_request_all_locks(dlm, ndata->node_num,
581 						       dead_node);
582 			if (status < 0) {
583 				mlog_errno(status);
584 				if (dlm_is_host_down(status)) {
585 					/* node died, ignore it for recovery */
586 					status = 0;
587 					ndata->state = DLM_RECO_NODE_DATA_DEAD;
588 					/* wait for the domain map to catch up
589 					 * with the network state. */
590 					wait_event_timeout(dlm->dlm_reco_thread_wq,
591 							   dlm_is_node_dead(dlm,
592 								ndata->node_num),
593 							   msecs_to_jiffies(1000));
594 					mlog(0, "waited 1 sec for %u, "
595 					     "dead? %s\n", ndata->node_num,
596 					     dlm_is_node_dead(dlm, ndata->node_num) ?
597 					     "yes" : "no");
598 				} else {
599 					/* -ENOMEM on the other node */
600 					mlog(0, "%s: node %u returned "
601 					     "%d during recovery, retrying "
602 					     "after a short wait\n",
603 					     dlm->name, ndata->node_num,
604 					     status);
605 					msleep(100);
606 				}
607 			}
608 		} while (status != 0);
609 
610 		spin_lock(&dlm_reco_state_lock);
611 		switch (ndata->state) {
612 			case DLM_RECO_NODE_DATA_INIT:
613 			case DLM_RECO_NODE_DATA_FINALIZE_SENT:
614 			case DLM_RECO_NODE_DATA_REQUESTED:
615 				BUG();
616 				break;
617 			case DLM_RECO_NODE_DATA_DEAD:
618 				mlog(0, "node %u died after requesting "
619 				     "recovery info for node %u\n",
620 				     ndata->node_num, dead_node);
621 				/* fine.  don't need this node's info.
622 				 * continue without it. */
623 				break;
624 			case DLM_RECO_NODE_DATA_REQUESTING:
625 				ndata->state = DLM_RECO_NODE_DATA_REQUESTED;
626 				mlog(0, "now receiving recovery data from "
627 				     "node %u for dead node %u\n",
628 				     ndata->node_num, dead_node);
629 				break;
630 			case DLM_RECO_NODE_DATA_RECEIVING:
631 				mlog(0, "already receiving recovery data from "
632 				     "node %u for dead node %u\n",
633 				     ndata->node_num, dead_node);
634 				break;
635 			case DLM_RECO_NODE_DATA_DONE:
636 				mlog(0, "already DONE receiving recovery data "
637 				     "from node %u for dead node %u\n",
638 				     ndata->node_num, dead_node);
639 				break;
640 		}
641 		spin_unlock(&dlm_reco_state_lock);
642 	}
643 
644 	mlog(0, "done requesting all lock info\n");
645 
646 	/* nodes should be sending reco data now
647 	 * just need to wait */
648 
649 	while (1) {
650 		/* check all the nodes now to see if we are
651 		 * done, or if anyone died */
652 		all_nodes_done = 1;
653 		spin_lock(&dlm_reco_state_lock);
654 		list_for_each_entry(ndata, &dlm->reco.node_data, list) {
655 			mlog(0, "checking recovery state of node %u\n",
656 			     ndata->node_num);
657 			switch (ndata->state) {
658 				case DLM_RECO_NODE_DATA_INIT:
659 				case DLM_RECO_NODE_DATA_REQUESTING:
660 					mlog(ML_ERROR, "bad ndata state for "
661 					     "node %u: state=%d\n",
662 					     ndata->node_num, ndata->state);
663 					BUG();
664 					break;
665 				case DLM_RECO_NODE_DATA_DEAD:
666 					mlog(0, "node %u died after "
667 					     "requesting recovery info for "
668 					     "node %u\n", ndata->node_num,
669 					     dead_node);
670 					break;
671 				case DLM_RECO_NODE_DATA_RECEIVING:
672 				case DLM_RECO_NODE_DATA_REQUESTED:
673 					mlog(0, "%s: node %u still in state %s\n",
674 					     dlm->name, ndata->node_num,
675 					     ndata->state==DLM_RECO_NODE_DATA_RECEIVING ?
676 					     "receiving" : "requested");
677 					all_nodes_done = 0;
678 					break;
679 				case DLM_RECO_NODE_DATA_DONE:
680 					mlog(0, "%s: node %u state is done\n",
681 					     dlm->name, ndata->node_num);
682 					break;
683 				case DLM_RECO_NODE_DATA_FINALIZE_SENT:
684 					mlog(0, "%s: node %u state is finalize\n",
685 					     dlm->name, ndata->node_num);
686 					break;
687 			}
688 		}
689 		spin_unlock(&dlm_reco_state_lock);
690 
691 		mlog(0, "pass #%d, all_nodes_done?: %s\n", ++pass,
692 		     all_nodes_done?"yes":"no");
693 		if (all_nodes_done) {
694 			int ret;
695 
696 			/* all nodes are now in DLM_RECO_NODE_DATA_DONE state
697 	 		 * just send a finalize message to everyone and
698 	 		 * clean up */
699 			mlog(0, "all nodes are done! send finalize\n");
700 			ret = dlm_send_finalize_reco_message(dlm);
701 			if (ret < 0)
702 				mlog_errno(ret);
703 
704 			spin_lock(&dlm->spinlock);
705 			dlm_finish_local_lockres_recovery(dlm, dead_node,
706 							  dlm->node_num);
707 			spin_unlock(&dlm->spinlock);
708 			mlog(0, "should be done with recovery!\n");
709 
710 			mlog(0, "finishing recovery of %s at %lu, "
711 			     "dead=%u, this=%u, new=%u\n", dlm->name,
712 			     jiffies, dlm->reco.dead_node,
713 			     dlm->node_num, dlm->reco.new_master);
714 			destroy = 1;
715 			status = 0;
716 			/* rescan everything marked dirty along the way */
717 			dlm_kick_thread(dlm, NULL);
718 			break;
719 		}
720 		/* wait to be signalled, with periodic timeout
721 		 * to check for node death */
722 		wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq,
723 					 kthread_should_stop(),
724 					 msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS));
725 
726 	}
727 
728 	if (destroy)
729 		dlm_destroy_recovery_area(dlm, dead_node);
730 
731 	mlog_exit(status);
732 	return status;
733 }
734 
735 static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node)
736 {
737 	int num=0;
738 	struct dlm_reco_node_data *ndata;
739 
740 	spin_lock(&dlm->spinlock);
741 	memcpy(dlm->reco.node_map, dlm->domain_map, sizeof(dlm->domain_map));
742 	/* nodes can only be removed (by dying) after dropping
743 	 * this lock, and death will be trapped later, so this should do */
744 	spin_unlock(&dlm->spinlock);
745 
746 	while (1) {
747 		num = find_next_bit (dlm->reco.node_map, O2NM_MAX_NODES, num);
748 		if (num >= O2NM_MAX_NODES) {
749 			break;
750 		}
751 		BUG_ON(num == dead_node);
752 
753 		ndata = kzalloc(sizeof(*ndata), GFP_NOFS);
754 		if (!ndata) {
755 			dlm_destroy_recovery_area(dlm, dead_node);
756 			return -ENOMEM;
757 		}
758 		ndata->node_num = num;
759 		ndata->state = DLM_RECO_NODE_DATA_INIT;
760 		spin_lock(&dlm_reco_state_lock);
761 		list_add_tail(&ndata->list, &dlm->reco.node_data);
762 		spin_unlock(&dlm_reco_state_lock);
763 		num++;
764 	}
765 
766 	return 0;
767 }
768 
769 static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm, u8 dead_node)
770 {
771 	struct dlm_reco_node_data *ndata, *next;
772 	LIST_HEAD(tmplist);
773 
774 	spin_lock(&dlm_reco_state_lock);
775 	list_splice_init(&dlm->reco.node_data, &tmplist);
776 	spin_unlock(&dlm_reco_state_lock);
777 
778 	list_for_each_entry_safe(ndata, next, &tmplist, list) {
779 		list_del_init(&ndata->list);
780 		kfree(ndata);
781 	}
782 }
783 
784 static int dlm_request_all_locks(struct dlm_ctxt *dlm, u8 request_from,
785 				 u8 dead_node)
786 {
787 	struct dlm_lock_request lr;
788 	enum dlm_status ret;
789 
790 	mlog(0, "\n");
791 
792 
793 	mlog(0, "dlm_request_all_locks: dead node is %u, sending request "
794 		  "to %u\n", dead_node, request_from);
795 
796 	memset(&lr, 0, sizeof(lr));
797 	lr.node_idx = dlm->node_num;
798 	lr.dead_node = dead_node;
799 
800 	// send message
801 	ret = DLM_NOLOCKMGR;
802 	ret = o2net_send_message(DLM_LOCK_REQUEST_MSG, dlm->key,
803 				 &lr, sizeof(lr), request_from, NULL);
804 
805 	/* negative status is handled by caller */
806 	if (ret < 0)
807 		mlog_errno(ret);
808 
809 	// return from here, then
810 	// sleep until all received or error
811 	return ret;
812 
813 }
814 
815 int dlm_request_all_locks_handler(struct o2net_msg *msg, u32 len, void *data,
816 				  void **ret_data)
817 {
818 	struct dlm_ctxt *dlm = data;
819 	struct dlm_lock_request *lr = (struct dlm_lock_request *)msg->buf;
820 	char *buf = NULL;
821 	struct dlm_work_item *item = NULL;
822 
823 	if (!dlm_grab(dlm))
824 		return -EINVAL;
825 
826 	if (lr->dead_node != dlm->reco.dead_node) {
827 		mlog(ML_ERROR, "%s: node %u sent dead_node=%u, but local "
828 		     "dead_node is %u\n", dlm->name, lr->node_idx,
829 		     lr->dead_node, dlm->reco.dead_node);
830 		dlm_print_reco_node_status(dlm);
831 		/* this is a hack */
832 		dlm_put(dlm);
833 		return -ENOMEM;
834 	}
835 	BUG_ON(lr->dead_node != dlm->reco.dead_node);
836 
837 	item = kzalloc(sizeof(*item), GFP_NOFS);
838 	if (!item) {
839 		dlm_put(dlm);
840 		return -ENOMEM;
841 	}
842 
843 	/* this will get freed by dlm_request_all_locks_worker */
844 	buf = (char *) __get_free_page(GFP_NOFS);
845 	if (!buf) {
846 		kfree(item);
847 		dlm_put(dlm);
848 		return -ENOMEM;
849 	}
850 
851 	/* queue up work for dlm_request_all_locks_worker */
852 	dlm_grab(dlm);  /* get an extra ref for the work item */
853 	dlm_init_work_item(dlm, item, dlm_request_all_locks_worker, buf);
854 	item->u.ral.reco_master = lr->node_idx;
855 	item->u.ral.dead_node = lr->dead_node;
856 	spin_lock(&dlm->work_lock);
857 	list_add_tail(&item->list, &dlm->work_list);
858 	spin_unlock(&dlm->work_lock);
859 	queue_work(dlm->dlm_worker, &dlm->dispatched_work);
860 
861 	dlm_put(dlm);
862 	return 0;
863 }
864 
865 static void dlm_request_all_locks_worker(struct dlm_work_item *item, void *data)
866 {
867 	struct dlm_migratable_lockres *mres;
868 	struct dlm_lock_resource *res;
869 	struct dlm_ctxt *dlm;
870 	LIST_HEAD(resources);
871 	int ret;
872 	u8 dead_node, reco_master;
873 	int skip_all_done = 0;
874 
875 	dlm = item->dlm;
876 	dead_node = item->u.ral.dead_node;
877 	reco_master = item->u.ral.reco_master;
878 	mres = (struct dlm_migratable_lockres *)data;
879 
880 	mlog(0, "%s: recovery worker started, dead=%u, master=%u\n",
881 	     dlm->name, dead_node, reco_master);
882 
883 	if (dead_node != dlm->reco.dead_node ||
884 	    reco_master != dlm->reco.new_master) {
885 		/* worker could have been created before the recovery master
886 		 * died.  if so, do not continue, but do not error. */
887 		if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) {
888 			mlog(ML_NOTICE, "%s: will not send recovery state, "
889 			     "recovery master %u died, thread=(dead=%u,mas=%u)"
890 			     " current=(dead=%u,mas=%u)\n", dlm->name,
891 			     reco_master, dead_node, reco_master,
892 			     dlm->reco.dead_node, dlm->reco.new_master);
893 		} else {
894 			mlog(ML_NOTICE, "%s: reco state invalid: reco(dead=%u, "
895 			     "master=%u), request(dead=%u, master=%u)\n",
896 			     dlm->name, dlm->reco.dead_node,
897 			     dlm->reco.new_master, dead_node, reco_master);
898 		}
899 		goto leave;
900 	}
901 
902 	/* lock resources should have already been moved to the
903  	 * dlm->reco.resources list.  now move items from that list
904  	 * to a temp list if the dead owner matches.  note that the
905 	 * whole cluster recovers only one node at a time, so we
906 	 * can safely move UNKNOWN lock resources for each recovery
907 	 * session. */
908 	dlm_move_reco_locks_to_list(dlm, &resources, dead_node);
909 
910 	/* now we can begin blasting lockreses without the dlm lock */
911 
912 	/* any errors returned will be due to the new_master dying,
913 	 * the dlm_reco_thread should detect this */
914 	list_for_each_entry(res, &resources, recovering) {
915 		ret = dlm_send_one_lockres(dlm, res, mres, reco_master,
916 				   	DLM_MRES_RECOVERY);
917 		if (ret < 0) {
918 			mlog(ML_ERROR, "%s: node %u went down while sending "
919 			     "recovery state for dead node %u, ret=%d\n", dlm->name,
920 			     reco_master, dead_node, ret);
921 			skip_all_done = 1;
922 			break;
923 		}
924 	}
925 
926 	/* move the resources back to the list */
927 	spin_lock(&dlm->spinlock);
928 	list_splice_init(&resources, &dlm->reco.resources);
929 	spin_unlock(&dlm->spinlock);
930 
931 	if (!skip_all_done) {
932 		ret = dlm_send_all_done_msg(dlm, dead_node, reco_master);
933 		if (ret < 0) {
934 			mlog(ML_ERROR, "%s: node %u went down while sending "
935 			     "recovery all-done for dead node %u, ret=%d\n",
936 			     dlm->name, reco_master, dead_node, ret);
937 		}
938 	}
939 leave:
940 	free_page((unsigned long)data);
941 }
942 
943 
944 static int dlm_send_all_done_msg(struct dlm_ctxt *dlm, u8 dead_node, u8 send_to)
945 {
946 	int ret, tmpret;
947 	struct dlm_reco_data_done done_msg;
948 
949 	memset(&done_msg, 0, sizeof(done_msg));
950 	done_msg.node_idx = dlm->node_num;
951 	done_msg.dead_node = dead_node;
952 	mlog(0, "sending DATA DONE message to %u, "
953 	     "my node=%u, dead node=%u\n", send_to, done_msg.node_idx,
954 	     done_msg.dead_node);
955 
956 	ret = o2net_send_message(DLM_RECO_DATA_DONE_MSG, dlm->key, &done_msg,
957 				 sizeof(done_msg), send_to, &tmpret);
958 	if (ret < 0) {
959 		if (!dlm_is_host_down(ret)) {
960 			mlog_errno(ret);
961 			mlog(ML_ERROR, "%s: unknown error sending data-done "
962 			     "to %u\n", dlm->name, send_to);
963 			BUG();
964 		}
965 	} else
966 		ret = tmpret;
967 	return ret;
968 }
969 
970 
971 int dlm_reco_data_done_handler(struct o2net_msg *msg, u32 len, void *data,
972 			       void **ret_data)
973 {
974 	struct dlm_ctxt *dlm = data;
975 	struct dlm_reco_data_done *done = (struct dlm_reco_data_done *)msg->buf;
976 	struct dlm_reco_node_data *ndata = NULL;
977 	int ret = -EINVAL;
978 
979 	if (!dlm_grab(dlm))
980 		return -EINVAL;
981 
982 	mlog(0, "got DATA DONE: dead_node=%u, reco.dead_node=%u, "
983 	     "node_idx=%u, this node=%u\n", done->dead_node,
984 	     dlm->reco.dead_node, done->node_idx, dlm->node_num);
985 
986 	mlog_bug_on_msg((done->dead_node != dlm->reco.dead_node),
987 			"Got DATA DONE: dead_node=%u, reco.dead_node=%u, "
988 			"node_idx=%u, this node=%u\n", done->dead_node,
989 			dlm->reco.dead_node, done->node_idx, dlm->node_num);
990 
991 	spin_lock(&dlm_reco_state_lock);
992 	list_for_each_entry(ndata, &dlm->reco.node_data, list) {
993 		if (ndata->node_num != done->node_idx)
994 			continue;
995 
996 		switch (ndata->state) {
997 			/* should have moved beyond INIT but not to FINALIZE yet */
998 			case DLM_RECO_NODE_DATA_INIT:
999 			case DLM_RECO_NODE_DATA_DEAD:
1000 			case DLM_RECO_NODE_DATA_FINALIZE_SENT:
1001 				mlog(ML_ERROR, "bad ndata state for node %u:"
1002 				     " state=%d\n", ndata->node_num,
1003 				     ndata->state);
1004 				BUG();
1005 				break;
1006 			/* these states are possible at this point, anywhere along
1007 			 * the line of recovery */
1008 			case DLM_RECO_NODE_DATA_DONE:
1009 			case DLM_RECO_NODE_DATA_RECEIVING:
1010 			case DLM_RECO_NODE_DATA_REQUESTED:
1011 			case DLM_RECO_NODE_DATA_REQUESTING:
1012 				mlog(0, "node %u is DONE sending "
1013 					  "recovery data!\n",
1014 					  ndata->node_num);
1015 
1016 				ndata->state = DLM_RECO_NODE_DATA_DONE;
1017 				ret = 0;
1018 				break;
1019 		}
1020 	}
1021 	spin_unlock(&dlm_reco_state_lock);
1022 
1023 	/* wake the recovery thread, some node is done */
1024 	if (!ret)
1025 		dlm_kick_recovery_thread(dlm);
1026 
1027 	if (ret < 0)
1028 		mlog(ML_ERROR, "failed to find recovery node data for node "
1029 		     "%u\n", done->node_idx);
1030 	dlm_put(dlm);
1031 
1032 	mlog(0, "leaving reco data done handler, ret=%d\n", ret);
1033 	return ret;
1034 }
1035 
1036 static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm,
1037 					struct list_head *list,
1038 				       	u8 dead_node)
1039 {
1040 	struct dlm_lock_resource *res, *next;
1041 	struct dlm_lock *lock;
1042 
1043 	spin_lock(&dlm->spinlock);
1044 	list_for_each_entry_safe(res, next, &dlm->reco.resources, recovering) {
1045 		/* always prune any $RECOVERY entries for dead nodes,
1046 		 * otherwise hangs can occur during later recovery */
1047 		if (dlm_is_recovery_lock(res->lockname.name,
1048 					 res->lockname.len)) {
1049 			spin_lock(&res->spinlock);
1050 			list_for_each_entry(lock, &res->granted, list) {
1051 				if (lock->ml.node == dead_node) {
1052 					mlog(0, "AHA! there was "
1053 					     "a $RECOVERY lock for dead "
1054 					     "node %u (%s)!\n",
1055 					     dead_node, dlm->name);
1056 					list_del_init(&lock->list);
1057 					dlm_lock_put(lock);
1058 					break;
1059 				}
1060 			}
1061 			spin_unlock(&res->spinlock);
1062 			continue;
1063 		}
1064 
1065 		if (res->owner == dead_node) {
1066 			mlog(0, "found lockres owned by dead node while "
1067 				  "doing recovery for node %u. sending it.\n",
1068 				  dead_node);
1069 			list_move_tail(&res->recovering, list);
1070 		} else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
1071 			mlog(0, "found UNKNOWN owner while doing recovery "
1072 				  "for node %u. sending it.\n", dead_node);
1073 			list_move_tail(&res->recovering, list);
1074 		}
1075 	}
1076 	spin_unlock(&dlm->spinlock);
1077 }
1078 
1079 static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res)
1080 {
1081 	int total_locks = 0;
1082 	struct list_head *iter, *queue = &res->granted;
1083 	int i;
1084 
1085 	for (i=0; i<3; i++) {
1086 		list_for_each(iter, queue)
1087 			total_locks++;
1088 		queue++;
1089 	}
1090 	return total_locks;
1091 }
1092 
1093 
1094 static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm,
1095 				      struct dlm_migratable_lockres *mres,
1096 				      u8 send_to,
1097 				      struct dlm_lock_resource *res,
1098 				      int total_locks)
1099 {
1100 	u64 mig_cookie = be64_to_cpu(mres->mig_cookie);
1101 	int mres_total_locks = be32_to_cpu(mres->total_locks);
1102 	int sz, ret = 0, status = 0;
1103 	u8 orig_flags = mres->flags,
1104 	   orig_master = mres->master;
1105 
1106 	BUG_ON(mres->num_locks > DLM_MAX_MIGRATABLE_LOCKS);
1107 	if (!mres->num_locks)
1108 		return 0;
1109 
1110 	sz = sizeof(struct dlm_migratable_lockres) +
1111 		(mres->num_locks * sizeof(struct dlm_migratable_lock));
1112 
1113 	/* add an all-done flag if we reached the last lock */
1114 	orig_flags = mres->flags;
1115 	BUG_ON(total_locks > mres_total_locks);
1116 	if (total_locks == mres_total_locks)
1117 		mres->flags |= DLM_MRES_ALL_DONE;
1118 
1119 	mlog(0, "%s:%.*s: sending mig lockres (%s) to %u\n",
1120 	     dlm->name, res->lockname.len, res->lockname.name,
1121 	     orig_flags & DLM_MRES_MIGRATION ? "migrate" : "recovery",
1122 	     send_to);
1123 
1124 	/* send it */
1125 	ret = o2net_send_message(DLM_MIG_LOCKRES_MSG, dlm->key, mres,
1126 				 sz, send_to, &status);
1127 	if (ret < 0) {
1128 		/* XXX: negative status is not handled.
1129 		 * this will end up killing this node. */
1130 		mlog_errno(ret);
1131 	} else {
1132 		/* might get an -ENOMEM back here */
1133 		ret = status;
1134 		if (ret < 0) {
1135 			mlog_errno(ret);
1136 
1137 			if (ret == -EFAULT) {
1138 				mlog(ML_ERROR, "node %u told me to kill "
1139 				     "myself!\n", send_to);
1140 				BUG();
1141 			}
1142 		}
1143 	}
1144 
1145 	/* zero and reinit the message buffer */
1146 	dlm_init_migratable_lockres(mres, res->lockname.name,
1147 				    res->lockname.len, mres_total_locks,
1148 				    mig_cookie, orig_flags, orig_master);
1149 	return ret;
1150 }
1151 
1152 static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres,
1153 					const char *lockname, int namelen,
1154 					int total_locks, u64 cookie,
1155 					u8 flags, u8 master)
1156 {
1157 	/* mres here is one full page */
1158 	clear_page(mres);
1159 	mres->lockname_len = namelen;
1160 	memcpy(mres->lockname, lockname, namelen);
1161 	mres->num_locks = 0;
1162 	mres->total_locks = cpu_to_be32(total_locks);
1163 	mres->mig_cookie = cpu_to_be64(cookie);
1164 	mres->flags = flags;
1165 	mres->master = master;
1166 }
1167 
1168 
1169 /* returns 1 if this lock fills the network structure,
1170  * 0 otherwise */
1171 static int dlm_add_lock_to_array(struct dlm_lock *lock,
1172 				 struct dlm_migratable_lockres *mres, int queue)
1173 {
1174 	struct dlm_migratable_lock *ml;
1175 	int lock_num = mres->num_locks;
1176 
1177 	ml = &(mres->ml[lock_num]);
1178 	ml->cookie = lock->ml.cookie;
1179 	ml->type = lock->ml.type;
1180 	ml->convert_type = lock->ml.convert_type;
1181 	ml->highest_blocked = lock->ml.highest_blocked;
1182 	ml->list = queue;
1183 	if (lock->lksb) {
1184 		ml->flags = lock->lksb->flags;
1185 		/* send our current lvb */
1186 		if (ml->type == LKM_EXMODE ||
1187 		    ml->type == LKM_PRMODE) {
1188 			/* if it is already set, this had better be a PR
1189 			 * and it has to match */
1190 			if (!dlm_lvb_is_empty(mres->lvb) &&
1191 			    (ml->type == LKM_EXMODE ||
1192 			     memcmp(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN))) {
1193 				mlog(ML_ERROR, "mismatched lvbs!\n");
1194 				__dlm_print_one_lock_resource(lock->lockres);
1195 				BUG();
1196 			}
1197 			memcpy(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN);
1198 		}
1199 	}
1200 	ml->node = lock->ml.node;
1201 	mres->num_locks++;
1202 	/* we reached the max, send this network message */
1203 	if (mres->num_locks == DLM_MAX_MIGRATABLE_LOCKS)
1204 		return 1;
1205 	return 0;
1206 }
1207 
1208 static void dlm_add_dummy_lock(struct dlm_ctxt *dlm,
1209 			       struct dlm_migratable_lockres *mres)
1210 {
1211 	struct dlm_lock dummy;
1212 	memset(&dummy, 0, sizeof(dummy));
1213 	dummy.ml.cookie = 0;
1214 	dummy.ml.type = LKM_IVMODE;
1215 	dummy.ml.convert_type = LKM_IVMODE;
1216 	dummy.ml.highest_blocked = LKM_IVMODE;
1217 	dummy.lksb = NULL;
1218 	dummy.ml.node = dlm->node_num;
1219 	dlm_add_lock_to_array(&dummy, mres, DLM_BLOCKED_LIST);
1220 }
1221 
1222 static inline int dlm_is_dummy_lock(struct dlm_ctxt *dlm,
1223 				    struct dlm_migratable_lock *ml,
1224 				    u8 *nodenum)
1225 {
1226 	if (unlikely(ml->cookie == 0 &&
1227 	    ml->type == LKM_IVMODE &&
1228 	    ml->convert_type == LKM_IVMODE &&
1229 	    ml->highest_blocked == LKM_IVMODE &&
1230 	    ml->list == DLM_BLOCKED_LIST)) {
1231 		*nodenum = ml->node;
1232 		return 1;
1233 	}
1234 	return 0;
1235 }
1236 
1237 int dlm_send_one_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
1238 			 struct dlm_migratable_lockres *mres,
1239 			 u8 send_to, u8 flags)
1240 {
1241 	struct list_head *queue;
1242 	int total_locks, i;
1243 	u64 mig_cookie = 0;
1244 	struct dlm_lock *lock;
1245 	int ret = 0;
1246 
1247 	BUG_ON(!(flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION)));
1248 
1249 	mlog(0, "sending to %u\n", send_to);
1250 
1251 	total_locks = dlm_num_locks_in_lockres(res);
1252 	if (total_locks > DLM_MAX_MIGRATABLE_LOCKS) {
1253 		/* rare, but possible */
1254 		mlog(0, "argh.  lockres has %d locks.  this will "
1255 			  "require more than one network packet to "
1256 			  "migrate\n", total_locks);
1257 		mig_cookie = dlm_get_next_mig_cookie();
1258 	}
1259 
1260 	dlm_init_migratable_lockres(mres, res->lockname.name,
1261 				    res->lockname.len, total_locks,
1262 				    mig_cookie, flags, res->owner);
1263 
1264 	total_locks = 0;
1265 	for (i=DLM_GRANTED_LIST; i<=DLM_BLOCKED_LIST; i++) {
1266 		queue = dlm_list_idx_to_ptr(res, i);
1267 		list_for_each_entry(lock, queue, list) {
1268 			/* add another lock. */
1269 			total_locks++;
1270 			if (!dlm_add_lock_to_array(lock, mres, i))
1271 				continue;
1272 
1273 			/* this filled the lock message,
1274 			 * we must send it immediately. */
1275 			ret = dlm_send_mig_lockres_msg(dlm, mres, send_to,
1276 						       res, total_locks);
1277 			if (ret < 0)
1278 				goto error;
1279 		}
1280 	}
1281 	if (total_locks == 0) {
1282 		/* send a dummy lock to indicate a mastery reference only */
1283 		mlog(0, "%s:%.*s: sending dummy lock to %u, %s\n",
1284 		     dlm->name, res->lockname.len, res->lockname.name,
1285 		     send_to, flags & DLM_MRES_RECOVERY ? "recovery" :
1286 		     "migration");
1287 		dlm_add_dummy_lock(dlm, mres);
1288 	}
1289 	/* flush any remaining locks */
1290 	ret = dlm_send_mig_lockres_msg(dlm, mres, send_to, res, total_locks);
1291 	if (ret < 0)
1292 		goto error;
1293 	return ret;
1294 
1295 error:
1296 	mlog(ML_ERROR, "%s: dlm_send_mig_lockres_msg returned %d\n",
1297 	     dlm->name, ret);
1298 	if (!dlm_is_host_down(ret))
1299 		BUG();
1300 	mlog(0, "%s: node %u went down while sending %s "
1301 	     "lockres %.*s\n", dlm->name, send_to,
1302 	     flags & DLM_MRES_RECOVERY ?  "recovery" : "migration",
1303 	     res->lockname.len, res->lockname.name);
1304 	return ret;
1305 }
1306 
1307 
1308 
1309 /*
1310  * this message will contain no more than one page worth of
1311  * recovery data, and it will work on only one lockres.
1312  * there may be many locks in this page, and we may need to wait
1313  * for additional packets to complete all the locks (rare, but
1314  * possible).
1315  */
1316 /*
1317  * NOTE: the allocation error cases here are scary
1318  * we really cannot afford to fail an alloc in recovery
1319  * do we spin?  returning an error only delays the problem really
1320  */
1321 
1322 int dlm_mig_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
1323 			    void **ret_data)
1324 {
1325 	struct dlm_ctxt *dlm = data;
1326 	struct dlm_migratable_lockres *mres =
1327 		(struct dlm_migratable_lockres *)msg->buf;
1328 	int ret = 0;
1329 	u8 real_master;
1330 	char *buf = NULL;
1331 	struct dlm_work_item *item = NULL;
1332 	struct dlm_lock_resource *res = NULL;
1333 
1334 	if (!dlm_grab(dlm))
1335 		return -EINVAL;
1336 
1337 	BUG_ON(!(mres->flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION)));
1338 
1339 	real_master = mres->master;
1340 	if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1341 		/* cannot migrate a lockres with no master */
1342 		BUG_ON(!(mres->flags & DLM_MRES_RECOVERY));
1343 	}
1344 
1345 	mlog(0, "%s message received from node %u\n",
1346 		  (mres->flags & DLM_MRES_RECOVERY) ?
1347 		  "recovery" : "migration", mres->master);
1348 	if (mres->flags & DLM_MRES_ALL_DONE)
1349 		mlog(0, "all done flag.  all lockres data received!\n");
1350 
1351 	ret = -ENOMEM;
1352 	buf = kmalloc(be16_to_cpu(msg->data_len), GFP_NOFS);
1353 	item = kzalloc(sizeof(*item), GFP_NOFS);
1354 	if (!buf || !item)
1355 		goto leave;
1356 
1357 	/* lookup the lock to see if we have a secondary queue for this
1358 	 * already...  just add the locks in and this will have its owner
1359 	 * and RECOVERY flag changed when it completes. */
1360 	res = dlm_lookup_lockres(dlm, mres->lockname, mres->lockname_len);
1361 	if (res) {
1362 	 	/* this will get a ref on res */
1363 		/* mark it as recovering/migrating and hash it */
1364 		spin_lock(&res->spinlock);
1365 		if (mres->flags & DLM_MRES_RECOVERY) {
1366 			res->state |= DLM_LOCK_RES_RECOVERING;
1367 		} else {
1368 			if (res->state & DLM_LOCK_RES_MIGRATING) {
1369 				/* this is at least the second
1370 				 * lockres message */
1371 				mlog(0, "lock %.*s is already migrating\n",
1372 					  mres->lockname_len,
1373 					  mres->lockname);
1374 			} else if (res->state & DLM_LOCK_RES_RECOVERING) {
1375 				/* caller should BUG */
1376 				mlog(ML_ERROR, "node is attempting to migrate "
1377 				     "lock %.*s, but marked as recovering!\n",
1378 				     mres->lockname_len, mres->lockname);
1379 				ret = -EFAULT;
1380 				spin_unlock(&res->spinlock);
1381 				goto leave;
1382 			}
1383 			res->state |= DLM_LOCK_RES_MIGRATING;
1384 		}
1385 		spin_unlock(&res->spinlock);
1386 	} else {
1387 		/* need to allocate, just like if it was
1388 		 * mastered here normally  */
1389 		res = dlm_new_lockres(dlm, mres->lockname, mres->lockname_len);
1390 		if (!res)
1391 			goto leave;
1392 
1393 		/* to match the ref that we would have gotten if
1394 		 * dlm_lookup_lockres had succeeded */
1395 		dlm_lockres_get(res);
1396 
1397 		/* mark it as recovering/migrating and hash it */
1398 		if (mres->flags & DLM_MRES_RECOVERY)
1399 			res->state |= DLM_LOCK_RES_RECOVERING;
1400 		else
1401 			res->state |= DLM_LOCK_RES_MIGRATING;
1402 
1403 		spin_lock(&dlm->spinlock);
1404 		__dlm_insert_lockres(dlm, res);
1405 		spin_unlock(&dlm->spinlock);
1406 
1407 		/* now that the new lockres is inserted,
1408 		 * make it usable by other processes */
1409 		spin_lock(&res->spinlock);
1410 		res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
1411 		spin_unlock(&res->spinlock);
1412 		wake_up(&res->wq);
1413 
1414 		/* add an extra ref for just-allocated lockres
1415 		 * otherwise the lockres will be purged immediately */
1416 		dlm_lockres_get(res);
1417 	}
1418 
1419 	/* at this point we have allocated everything we need,
1420 	 * and we have a hashed lockres with an extra ref and
1421 	 * the proper res->state flags. */
1422 	ret = 0;
1423 	spin_lock(&res->spinlock);
1424 	/* drop this either when master requery finds a different master
1425 	 * or when a lock is added by the recovery worker */
1426 	dlm_lockres_grab_inflight_ref(dlm, res);
1427 	if (mres->master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1428 		/* migration cannot have an unknown master */
1429 		BUG_ON(!(mres->flags & DLM_MRES_RECOVERY));
1430 		mlog(0, "recovery has passed me a lockres with an "
1431 			  "unknown owner.. will need to requery: "
1432 			  "%.*s\n", mres->lockname_len, mres->lockname);
1433 	} else {
1434 		/* take a reference now to pin the lockres, drop it
1435 		 * when locks are added in the worker */
1436 		dlm_change_lockres_owner(dlm, res, dlm->node_num);
1437 	}
1438 	spin_unlock(&res->spinlock);
1439 
1440 	/* queue up work for dlm_mig_lockres_worker */
1441 	dlm_grab(dlm);  /* get an extra ref for the work item */
1442 	memcpy(buf, msg->buf, be16_to_cpu(msg->data_len));  /* copy the whole message */
1443 	dlm_init_work_item(dlm, item, dlm_mig_lockres_worker, buf);
1444 	item->u.ml.lockres = res; /* already have a ref */
1445 	item->u.ml.real_master = real_master;
1446 	spin_lock(&dlm->work_lock);
1447 	list_add_tail(&item->list, &dlm->work_list);
1448 	spin_unlock(&dlm->work_lock);
1449 	queue_work(dlm->dlm_worker, &dlm->dispatched_work);
1450 
1451 leave:
1452 	dlm_put(dlm);
1453 	if (ret < 0) {
1454 		if (buf)
1455 			kfree(buf);
1456 		if (item)
1457 			kfree(item);
1458 	}
1459 
1460 	mlog_exit(ret);
1461 	return ret;
1462 }
1463 
1464 
1465 static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data)
1466 {
1467 	struct dlm_ctxt *dlm = data;
1468 	struct dlm_migratable_lockres *mres;
1469 	int ret = 0;
1470 	struct dlm_lock_resource *res;
1471 	u8 real_master;
1472 
1473 	dlm = item->dlm;
1474 	mres = (struct dlm_migratable_lockres *)data;
1475 
1476 	res = item->u.ml.lockres;
1477 	real_master = item->u.ml.real_master;
1478 
1479 	if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1480 		/* this case is super-rare. only occurs if
1481 		 * node death happens during migration. */
1482 again:
1483 		ret = dlm_lockres_master_requery(dlm, res, &real_master);
1484 		if (ret < 0) {
1485 			mlog(0, "dlm_lockres_master_requery ret=%d\n",
1486 				  ret);
1487 			goto again;
1488 		}
1489 		if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1490 			mlog(0, "lockres %.*s not claimed.  "
1491 				   "this node will take it.\n",
1492 				   res->lockname.len, res->lockname.name);
1493 		} else {
1494 			spin_lock(&res->spinlock);
1495 			dlm_lockres_drop_inflight_ref(dlm, res);
1496 			spin_unlock(&res->spinlock);
1497 			mlog(0, "master needs to respond to sender "
1498 				  "that node %u still owns %.*s\n",
1499 				  real_master, res->lockname.len,
1500 				  res->lockname.name);
1501 			/* cannot touch this lockres */
1502 			goto leave;
1503 		}
1504 	}
1505 
1506 	ret = dlm_process_recovery_data(dlm, res, mres);
1507 	if (ret < 0)
1508 		mlog(0, "dlm_process_recovery_data returned  %d\n", ret);
1509 	else
1510 		mlog(0, "dlm_process_recovery_data succeeded\n");
1511 
1512 	if ((mres->flags & (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) ==
1513 	                   (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) {
1514 		ret = dlm_finish_migration(dlm, res, mres->master);
1515 		if (ret < 0)
1516 			mlog_errno(ret);
1517 	}
1518 
1519 leave:
1520 	kfree(data);
1521 	mlog_exit(ret);
1522 }
1523 
1524 
1525 
1526 static int dlm_lockres_master_requery(struct dlm_ctxt *dlm,
1527 				      struct dlm_lock_resource *res,
1528 				      u8 *real_master)
1529 {
1530 	struct dlm_node_iter iter;
1531 	int nodenum;
1532 	int ret = 0;
1533 
1534 	*real_master = DLM_LOCK_RES_OWNER_UNKNOWN;
1535 
1536 	/* we only reach here if one of the two nodes in a
1537 	 * migration died while the migration was in progress.
1538 	 * at this point we need to requery the master.  we
1539 	 * know that the new_master got as far as creating
1540 	 * an mle on at least one node, but we do not know
1541 	 * if any nodes had actually cleared the mle and set
1542 	 * the master to the new_master.  the old master
1543 	 * is supposed to set the owner to UNKNOWN in the
1544 	 * event of a new_master death, so the only possible
1545 	 * responses that we can get from nodes here are
1546 	 * that the master is new_master, or that the master
1547 	 * is UNKNOWN.
1548 	 * if all nodes come back with UNKNOWN then we know
1549 	 * the lock needs remastering here.
1550 	 * if any node comes back with a valid master, check
1551 	 * to see if that master is the one that we are
1552 	 * recovering.  if so, then the new_master died and
1553 	 * we need to remaster this lock.  if not, then the
1554 	 * new_master survived and that node will respond to
1555 	 * other nodes about the owner.
1556 	 * if there is an owner, this node needs to dump this
1557 	 * lockres and alert the sender that this lockres
1558 	 * was rejected. */
1559 	spin_lock(&dlm->spinlock);
1560 	dlm_node_iter_init(dlm->domain_map, &iter);
1561 	spin_unlock(&dlm->spinlock);
1562 
1563 	while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
1564 		/* do not send to self */
1565 		if (nodenum == dlm->node_num)
1566 			continue;
1567 		ret = dlm_do_master_requery(dlm, res, nodenum, real_master);
1568 		if (ret < 0) {
1569 			mlog_errno(ret);
1570 			if (!dlm_is_host_down(ret))
1571 				BUG();
1572 			/* host is down, so answer for that node would be
1573 			 * DLM_LOCK_RES_OWNER_UNKNOWN.  continue. */
1574 		}
1575 		if (*real_master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1576 			mlog(0, "lock master is %u\n", *real_master);
1577 			break;
1578 		}
1579 	}
1580 	return ret;
1581 }
1582 
1583 
1584 int dlm_do_master_requery(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
1585 			  u8 nodenum, u8 *real_master)
1586 {
1587 	int ret = -EINVAL;
1588 	struct dlm_master_requery req;
1589 	int status = DLM_LOCK_RES_OWNER_UNKNOWN;
1590 
1591 	memset(&req, 0, sizeof(req));
1592 	req.node_idx = dlm->node_num;
1593 	req.namelen = res->lockname.len;
1594 	memcpy(req.name, res->lockname.name, res->lockname.len);
1595 
1596 	ret = o2net_send_message(DLM_MASTER_REQUERY_MSG, dlm->key,
1597 				 &req, sizeof(req), nodenum, &status);
1598 	/* XXX: negative status not handled properly here. */
1599 	if (ret < 0)
1600 		mlog_errno(ret);
1601 	else {
1602 		BUG_ON(status < 0);
1603 		BUG_ON(status > DLM_LOCK_RES_OWNER_UNKNOWN);
1604 		*real_master = (u8) (status & 0xff);
1605 		mlog(0, "node %u responded to master requery with %u\n",
1606 			  nodenum, *real_master);
1607 		ret = 0;
1608 	}
1609 	return ret;
1610 }
1611 
1612 
1613 /* this function cannot error, so unless the sending
1614  * or receiving of the message failed, the owner can
1615  * be trusted */
1616 int dlm_master_requery_handler(struct o2net_msg *msg, u32 len, void *data,
1617 			       void **ret_data)
1618 {
1619 	struct dlm_ctxt *dlm = data;
1620 	struct dlm_master_requery *req = (struct dlm_master_requery *)msg->buf;
1621 	struct dlm_lock_resource *res = NULL;
1622 	unsigned int hash;
1623 	int master = DLM_LOCK_RES_OWNER_UNKNOWN;
1624 	u32 flags = DLM_ASSERT_MASTER_REQUERY;
1625 
1626 	if (!dlm_grab(dlm)) {
1627 		/* since the domain has gone away on this
1628 		 * node, the proper response is UNKNOWN */
1629 		return master;
1630 	}
1631 
1632 	hash = dlm_lockid_hash(req->name, req->namelen);
1633 
1634 	spin_lock(&dlm->spinlock);
1635 	res = __dlm_lookup_lockres(dlm, req->name, req->namelen, hash);
1636 	if (res) {
1637 		spin_lock(&res->spinlock);
1638 		master = res->owner;
1639 		if (master == dlm->node_num) {
1640 			int ret = dlm_dispatch_assert_master(dlm, res,
1641 							     0, 0, flags);
1642 			if (ret < 0) {
1643 				mlog_errno(-ENOMEM);
1644 				/* retry!? */
1645 				BUG();
1646 			}
1647 		}
1648 		spin_unlock(&res->spinlock);
1649 	}
1650 	spin_unlock(&dlm->spinlock);
1651 
1652 	dlm_put(dlm);
1653 	return master;
1654 }
1655 
1656 static inline struct list_head *
1657 dlm_list_num_to_pointer(struct dlm_lock_resource *res, int list_num)
1658 {
1659 	struct list_head *ret;
1660 	BUG_ON(list_num < 0);
1661 	BUG_ON(list_num > 2);
1662 	ret = &(res->granted);
1663 	ret += list_num;
1664 	return ret;
1665 }
1666 /* TODO: do ast flush business
1667  * TODO: do MIGRATING and RECOVERING spinning
1668  */
1669 
1670 /*
1671 * NOTE about in-flight requests during migration:
1672 *
1673 * Before attempting the migrate, the master has marked the lockres as
1674 * MIGRATING and then flushed all of its pending ASTS.  So any in-flight
1675 * requests either got queued before the MIGRATING flag got set, in which
1676 * case the lock data will reflect the change and a return message is on
1677 * the way, or the request failed to get in before MIGRATING got set.  In
1678 * this case, the caller will be told to spin and wait for the MIGRATING
1679 * flag to be dropped, then recheck the master.
1680 * This holds true for the convert, cancel and unlock cases, and since lvb
1681 * updates are tied to these same messages, it applies to lvb updates as
1682 * well.  For the lock case, there is no way a lock can be on the master
1683 * queue and not be on the secondary queue since the lock is always added
1684 * locally first.  This means that the new target node will never be sent
1685 * a lock that he doesn't already have on the list.
1686 * In total, this means that the local lock is correct and should not be
1687 * updated to match the one sent by the master.  Any messages sent back
1688 * from the master before the MIGRATING flag will bring the lock properly
1689 * up-to-date, and the change will be ordered properly for the waiter.
1690 * We will *not* attempt to modify the lock underneath the waiter.
1691 */
1692 
1693 static int dlm_process_recovery_data(struct dlm_ctxt *dlm,
1694 				     struct dlm_lock_resource *res,
1695 				     struct dlm_migratable_lockres *mres)
1696 {
1697 	struct dlm_migratable_lock *ml;
1698 	struct list_head *queue;
1699 	struct list_head *tmpq = NULL;
1700 	struct dlm_lock *newlock = NULL;
1701 	struct dlm_lockstatus *lksb = NULL;
1702 	int ret = 0;
1703 	int i, j, bad;
1704 	struct dlm_lock *lock = NULL;
1705 	u8 from = O2NM_MAX_NODES;
1706 	unsigned int added = 0;
1707 
1708 	mlog(0, "running %d locks for this lockres\n", mres->num_locks);
1709 	for (i=0; i<mres->num_locks; i++) {
1710 		ml = &(mres->ml[i]);
1711 
1712 		if (dlm_is_dummy_lock(dlm, ml, &from)) {
1713 			/* placeholder, just need to set the refmap bit */
1714 			BUG_ON(mres->num_locks != 1);
1715 			mlog(0, "%s:%.*s: dummy lock for %u\n",
1716 			     dlm->name, mres->lockname_len, mres->lockname,
1717 			     from);
1718 			spin_lock(&res->spinlock);
1719 			dlm_lockres_set_refmap_bit(from, res);
1720 			spin_unlock(&res->spinlock);
1721 			added++;
1722 			break;
1723 		}
1724 		BUG_ON(ml->highest_blocked != LKM_IVMODE);
1725 		newlock = NULL;
1726 		lksb = NULL;
1727 
1728 		queue = dlm_list_num_to_pointer(res, ml->list);
1729 		tmpq = NULL;
1730 
1731 		/* if the lock is for the local node it needs to
1732 		 * be moved to the proper location within the queue.
1733 		 * do not allocate a new lock structure. */
1734 		if (ml->node == dlm->node_num) {
1735 			/* MIGRATION ONLY! */
1736 			BUG_ON(!(mres->flags & DLM_MRES_MIGRATION));
1737 
1738 			spin_lock(&res->spinlock);
1739 			for (j = DLM_GRANTED_LIST; j <= DLM_BLOCKED_LIST; j++) {
1740 				tmpq = dlm_list_idx_to_ptr(res, j);
1741 				list_for_each_entry(lock, tmpq, list) {
1742 					if (lock->ml.cookie != ml->cookie)
1743 						lock = NULL;
1744 					else
1745 						break;
1746 				}
1747 				if (lock)
1748 					break;
1749 			}
1750 
1751 			/* lock is always created locally first, and
1752 			 * destroyed locally last.  it must be on the list */
1753 			if (!lock) {
1754 				__be64 c = ml->cookie;
1755 				mlog(ML_ERROR, "could not find local lock "
1756 					       "with cookie %u:%llu!\n",
1757 				     dlm_get_lock_cookie_node(be64_to_cpu(c)),
1758 				     dlm_get_lock_cookie_seq(be64_to_cpu(c)));
1759 				__dlm_print_one_lock_resource(res);
1760 				BUG();
1761 			}
1762 			BUG_ON(lock->ml.node != ml->node);
1763 
1764 			if (tmpq != queue) {
1765 				mlog(0, "lock was on %u instead of %u for %.*s\n",
1766 				     j, ml->list, res->lockname.len, res->lockname.name);
1767 				spin_unlock(&res->spinlock);
1768 				continue;
1769 			}
1770 
1771 			/* see NOTE above about why we do not update
1772 			 * to match the master here */
1773 
1774 			/* move the lock to its proper place */
1775 			/* do not alter lock refcount.  switching lists. */
1776 			list_move_tail(&lock->list, queue);
1777 			spin_unlock(&res->spinlock);
1778 			added++;
1779 
1780 			mlog(0, "just reordered a local lock!\n");
1781 			continue;
1782 		}
1783 
1784 		/* lock is for another node. */
1785 		newlock = dlm_new_lock(ml->type, ml->node,
1786 				       be64_to_cpu(ml->cookie), NULL);
1787 		if (!newlock) {
1788 			ret = -ENOMEM;
1789 			goto leave;
1790 		}
1791 		lksb = newlock->lksb;
1792 		dlm_lock_attach_lockres(newlock, res);
1793 
1794 		if (ml->convert_type != LKM_IVMODE) {
1795 			BUG_ON(queue != &res->converting);
1796 			newlock->ml.convert_type = ml->convert_type;
1797 		}
1798 		lksb->flags |= (ml->flags &
1799 				(DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB));
1800 
1801 		if (ml->type == LKM_NLMODE)
1802 			goto skip_lvb;
1803 
1804 		if (!dlm_lvb_is_empty(mres->lvb)) {
1805 			if (lksb->flags & DLM_LKSB_PUT_LVB) {
1806 				/* other node was trying to update
1807 				 * lvb when node died.  recreate the
1808 				 * lksb with the updated lvb. */
1809 				memcpy(lksb->lvb, mres->lvb, DLM_LVB_LEN);
1810 				/* the lock resource lvb update must happen
1811 				 * NOW, before the spinlock is dropped.
1812 				 * we no longer wait for the AST to update
1813 				 * the lvb. */
1814 				memcpy(res->lvb, mres->lvb, DLM_LVB_LEN);
1815 			} else {
1816 				/* otherwise, the node is sending its
1817 				 * most recent valid lvb info */
1818 				BUG_ON(ml->type != LKM_EXMODE &&
1819 				       ml->type != LKM_PRMODE);
1820 				if (!dlm_lvb_is_empty(res->lvb) &&
1821  				    (ml->type == LKM_EXMODE ||
1822  				     memcmp(res->lvb, mres->lvb, DLM_LVB_LEN))) {
1823  					int i;
1824  					mlog(ML_ERROR, "%s:%.*s: received bad "
1825  					     "lvb! type=%d\n", dlm->name,
1826  					     res->lockname.len,
1827  					     res->lockname.name, ml->type);
1828  					printk("lockres lvb=[");
1829  					for (i=0; i<DLM_LVB_LEN; i++)
1830  						printk("%02x", res->lvb[i]);
1831  					printk("]\nmigrated lvb=[");
1832  					for (i=0; i<DLM_LVB_LEN; i++)
1833  						printk("%02x", mres->lvb[i]);
1834  					printk("]\n");
1835  					dlm_print_one_lock_resource(res);
1836  					BUG();
1837 				}
1838 				memcpy(res->lvb, mres->lvb, DLM_LVB_LEN);
1839 			}
1840 		}
1841 skip_lvb:
1842 
1843 		/* NOTE:
1844 		 * wrt lock queue ordering and recovery:
1845 		 *    1. order of locks on granted queue is
1846 		 *       meaningless.
1847 		 *    2. order of locks on converting queue is
1848 		 *       LOST with the node death.  sorry charlie.
1849 		 *    3. order of locks on the blocked queue is
1850 		 *       also LOST.
1851 		 * order of locks does not affect integrity, it
1852 		 * just means that a lock request may get pushed
1853 		 * back in line as a result of the node death.
1854 		 * also note that for a given node the lock order
1855 		 * for its secondary queue locks is preserved
1856 		 * relative to each other, but clearly *not*
1857 		 * preserved relative to locks from other nodes.
1858 		 */
1859 		bad = 0;
1860 		spin_lock(&res->spinlock);
1861 		list_for_each_entry(lock, queue, list) {
1862 			if (lock->ml.cookie == ml->cookie) {
1863 				__be64 c = lock->ml.cookie;
1864 				mlog(ML_ERROR, "%s:%.*s: %u:%llu: lock already "
1865 				     "exists on this lockres!\n", dlm->name,
1866 				     res->lockname.len, res->lockname.name,
1867 				     dlm_get_lock_cookie_node(be64_to_cpu(c)),
1868 				     dlm_get_lock_cookie_seq(be64_to_cpu(c)));
1869 
1870 				mlog(ML_NOTICE, "sent lock: type=%d, conv=%d, "
1871 				     "node=%u, cookie=%u:%llu, queue=%d\n",
1872 	      			     ml->type, ml->convert_type, ml->node,
1873 				     dlm_get_lock_cookie_node(be64_to_cpu(ml->cookie)),
1874 				     dlm_get_lock_cookie_seq(be64_to_cpu(ml->cookie)),
1875 				     ml->list);
1876 
1877 				__dlm_print_one_lock_resource(res);
1878 				bad = 1;
1879 				break;
1880 			}
1881 		}
1882 		if (!bad) {
1883 			dlm_lock_get(newlock);
1884 			list_add_tail(&newlock->list, queue);
1885 			mlog(0, "%s:%.*s: added lock for node %u, "
1886 			     "setting refmap bit\n", dlm->name,
1887 			     res->lockname.len, res->lockname.name, ml->node);
1888 			dlm_lockres_set_refmap_bit(ml->node, res);
1889 			added++;
1890 		}
1891 		spin_unlock(&res->spinlock);
1892 	}
1893 	mlog(0, "done running all the locks\n");
1894 
1895 leave:
1896 	/* balance the ref taken when the work was queued */
1897 	spin_lock(&res->spinlock);
1898 	dlm_lockres_drop_inflight_ref(dlm, res);
1899 	spin_unlock(&res->spinlock);
1900 
1901 	if (ret < 0) {
1902 		mlog_errno(ret);
1903 		if (newlock)
1904 			dlm_lock_put(newlock);
1905 	}
1906 
1907 	mlog_exit(ret);
1908 	return ret;
1909 }
1910 
1911 void dlm_move_lockres_to_recovery_list(struct dlm_ctxt *dlm,
1912 				       struct dlm_lock_resource *res)
1913 {
1914 	int i;
1915 	struct list_head *queue;
1916 	struct dlm_lock *lock, *next;
1917 
1918 	res->state |= DLM_LOCK_RES_RECOVERING;
1919 	if (!list_empty(&res->recovering)) {
1920 		mlog(0,
1921 		     "Recovering res %s:%.*s, is already on recovery list!\n",
1922 		     dlm->name, res->lockname.len, res->lockname.name);
1923 		list_del_init(&res->recovering);
1924 	}
1925 	/* We need to hold a reference while on the recovery list */
1926 	dlm_lockres_get(res);
1927 	list_add_tail(&res->recovering, &dlm->reco.resources);
1928 
1929 	/* find any pending locks and put them back on proper list */
1930 	for (i=DLM_BLOCKED_LIST; i>=DLM_GRANTED_LIST; i--) {
1931 		queue = dlm_list_idx_to_ptr(res, i);
1932 		list_for_each_entry_safe(lock, next, queue, list) {
1933 			dlm_lock_get(lock);
1934 			if (lock->convert_pending) {
1935 				/* move converting lock back to granted */
1936 				BUG_ON(i != DLM_CONVERTING_LIST);
1937 				mlog(0, "node died with convert pending "
1938 				     "on %.*s. move back to granted list.\n",
1939 				     res->lockname.len, res->lockname.name);
1940 				dlm_revert_pending_convert(res, lock);
1941 				lock->convert_pending = 0;
1942 			} else if (lock->lock_pending) {
1943 				/* remove pending lock requests completely */
1944 				BUG_ON(i != DLM_BLOCKED_LIST);
1945 				mlog(0, "node died with lock pending "
1946 				     "on %.*s. remove from blocked list and skip.\n",
1947 				     res->lockname.len, res->lockname.name);
1948 				/* lock will be floating until ref in
1949 				 * dlmlock_remote is freed after the network
1950 				 * call returns.  ok for it to not be on any
1951 				 * list since no ast can be called
1952 				 * (the master is dead). */
1953 				dlm_revert_pending_lock(res, lock);
1954 				lock->lock_pending = 0;
1955 			} else if (lock->unlock_pending) {
1956 				/* if an unlock was in progress, treat as
1957 				 * if this had completed successfully
1958 				 * before sending this lock state to the
1959 				 * new master.  note that the dlm_unlock
1960 				 * call is still responsible for calling
1961 				 * the unlockast.  that will happen after
1962 				 * the network call times out.  for now,
1963 				 * just move lists to prepare the new
1964 				 * recovery master.  */
1965 				BUG_ON(i != DLM_GRANTED_LIST);
1966 				mlog(0, "node died with unlock pending "
1967 				     "on %.*s. remove from blocked list and skip.\n",
1968 				     res->lockname.len, res->lockname.name);
1969 				dlm_commit_pending_unlock(res, lock);
1970 				lock->unlock_pending = 0;
1971 			} else if (lock->cancel_pending) {
1972 				/* if a cancel was in progress, treat as
1973 				 * if this had completed successfully
1974 				 * before sending this lock state to the
1975 				 * new master */
1976 				BUG_ON(i != DLM_CONVERTING_LIST);
1977 				mlog(0, "node died with cancel pending "
1978 				     "on %.*s. move back to granted list.\n",
1979 				     res->lockname.len, res->lockname.name);
1980 				dlm_commit_pending_cancel(res, lock);
1981 				lock->cancel_pending = 0;
1982 			}
1983 			dlm_lock_put(lock);
1984 		}
1985 	}
1986 }
1987 
1988 
1989 
1990 /* removes all recovered locks from the recovery list.
1991  * sets the res->owner to the new master.
1992  * unsets the RECOVERY flag and wakes waiters. */
1993 static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm,
1994 					      u8 dead_node, u8 new_master)
1995 {
1996 	int i;
1997 	struct hlist_node *hash_iter;
1998 	struct hlist_head *bucket;
1999 	struct dlm_lock_resource *res, *next;
2000 
2001 	mlog_entry_void();
2002 
2003 	assert_spin_locked(&dlm->spinlock);
2004 
2005 	list_for_each_entry_safe(res, next, &dlm->reco.resources, recovering) {
2006 		if (res->owner == dead_node) {
2007 			list_del_init(&res->recovering);
2008 			spin_lock(&res->spinlock);
2009 			/* new_master has our reference from
2010 			 * the lock state sent during recovery */
2011 			dlm_change_lockres_owner(dlm, res, new_master);
2012 			res->state &= ~DLM_LOCK_RES_RECOVERING;
2013 			if (__dlm_lockres_has_locks(res))
2014 				__dlm_dirty_lockres(dlm, res);
2015 			spin_unlock(&res->spinlock);
2016 			wake_up(&res->wq);
2017 			dlm_lockres_put(res);
2018 		}
2019 	}
2020 
2021 	/* this will become unnecessary eventually, but
2022 	 * for now we need to run the whole hash, clear
2023 	 * the RECOVERING state and set the owner
2024 	 * if necessary */
2025 	for (i = 0; i < DLM_HASH_BUCKETS; i++) {
2026 		bucket = dlm_lockres_hash(dlm, i);
2027 		hlist_for_each_entry(res, hash_iter, bucket, hash_node) {
2028 			if (res->state & DLM_LOCK_RES_RECOVERING) {
2029 				if (res->owner == dead_node) {
2030 					mlog(0, "(this=%u) res %.*s owner=%u "
2031 					     "was not on recovering list, but "
2032 					     "clearing state anyway\n",
2033 					     dlm->node_num, res->lockname.len,
2034 					     res->lockname.name, new_master);
2035 				} else if (res->owner == dlm->node_num) {
2036 					mlog(0, "(this=%u) res %.*s owner=%u "
2037 					     "was not on recovering list, "
2038 					     "owner is THIS node, clearing\n",
2039 					     dlm->node_num, res->lockname.len,
2040 					     res->lockname.name, new_master);
2041 				} else
2042 					continue;
2043 
2044 				if (!list_empty(&res->recovering)) {
2045 					mlog(0, "%s:%.*s: lockres was "
2046 					     "marked RECOVERING, owner=%u\n",
2047 					     dlm->name, res->lockname.len,
2048 					     res->lockname.name, res->owner);
2049 					list_del_init(&res->recovering);
2050 					dlm_lockres_put(res);
2051 				}
2052 				spin_lock(&res->spinlock);
2053 				/* new_master has our reference from
2054 				 * the lock state sent during recovery */
2055 				dlm_change_lockres_owner(dlm, res, new_master);
2056 				res->state &= ~DLM_LOCK_RES_RECOVERING;
2057 				if (__dlm_lockres_has_locks(res))
2058 					__dlm_dirty_lockres(dlm, res);
2059 				spin_unlock(&res->spinlock);
2060 				wake_up(&res->wq);
2061 			}
2062 		}
2063 	}
2064 }
2065 
2066 static inline int dlm_lvb_needs_invalidation(struct dlm_lock *lock, int local)
2067 {
2068 	if (local) {
2069 		if (lock->ml.type != LKM_EXMODE &&
2070 		    lock->ml.type != LKM_PRMODE)
2071 			return 1;
2072 	} else if (lock->ml.type == LKM_EXMODE)
2073 		return 1;
2074 	return 0;
2075 }
2076 
2077 static void dlm_revalidate_lvb(struct dlm_ctxt *dlm,
2078 			       struct dlm_lock_resource *res, u8 dead_node)
2079 {
2080 	struct list_head *queue;
2081 	struct dlm_lock *lock;
2082 	int blank_lvb = 0, local = 0;
2083 	int i;
2084 	u8 search_node;
2085 
2086 	assert_spin_locked(&dlm->spinlock);
2087 	assert_spin_locked(&res->spinlock);
2088 
2089 	if (res->owner == dlm->node_num)
2090 		/* if this node owned the lockres, and if the dead node
2091 		 * had an EX when he died, blank out the lvb */
2092 		search_node = dead_node;
2093 	else {
2094 		/* if this is a secondary lockres, and we had no EX or PR
2095 		 * locks granted, we can no longer trust the lvb */
2096 		search_node = dlm->node_num;
2097 		local = 1;  /* check local state for valid lvb */
2098 	}
2099 
2100 	for (i=DLM_GRANTED_LIST; i<=DLM_CONVERTING_LIST; i++) {
2101 		queue = dlm_list_idx_to_ptr(res, i);
2102 		list_for_each_entry(lock, queue, list) {
2103 			if (lock->ml.node == search_node) {
2104 				if (dlm_lvb_needs_invalidation(lock, local)) {
2105 					/* zero the lksb lvb and lockres lvb */
2106 					blank_lvb = 1;
2107 					memset(lock->lksb->lvb, 0, DLM_LVB_LEN);
2108 				}
2109 			}
2110 		}
2111 	}
2112 
2113 	if (blank_lvb) {
2114 		mlog(0, "clearing %.*s lvb, dead node %u had EX\n",
2115 		     res->lockname.len, res->lockname.name, dead_node);
2116 		memset(res->lvb, 0, DLM_LVB_LEN);
2117 	}
2118 }
2119 
2120 static void dlm_free_dead_locks(struct dlm_ctxt *dlm,
2121 				struct dlm_lock_resource *res, u8 dead_node)
2122 {
2123 	struct dlm_lock *lock, *next;
2124 	unsigned int freed = 0;
2125 
2126 	/* this node is the lockres master:
2127 	 * 1) remove any stale locks for the dead node
2128 	 * 2) if the dead node had an EX when he died, blank out the lvb
2129 	 */
2130 	assert_spin_locked(&dlm->spinlock);
2131 	assert_spin_locked(&res->spinlock);
2132 
2133 	/* TODO: check pending_asts, pending_basts here */
2134 	list_for_each_entry_safe(lock, next, &res->granted, list) {
2135 		if (lock->ml.node == dead_node) {
2136 			list_del_init(&lock->list);
2137 			dlm_lock_put(lock);
2138 			freed++;
2139 		}
2140 	}
2141 	list_for_each_entry_safe(lock, next, &res->converting, list) {
2142 		if (lock->ml.node == dead_node) {
2143 			list_del_init(&lock->list);
2144 			dlm_lock_put(lock);
2145 			freed++;
2146 		}
2147 	}
2148 	list_for_each_entry_safe(lock, next, &res->blocked, list) {
2149 		if (lock->ml.node == dead_node) {
2150 			list_del_init(&lock->list);
2151 			dlm_lock_put(lock);
2152 			freed++;
2153 		}
2154 	}
2155 
2156 	if (freed) {
2157 		mlog(0, "%s:%.*s: freed %u locks for dead node %u, "
2158 		     "dropping ref from lockres\n", dlm->name,
2159 		     res->lockname.len, res->lockname.name, freed, dead_node);
2160 		BUG_ON(!test_bit(dead_node, res->refmap));
2161 		dlm_lockres_clear_refmap_bit(dead_node, res);
2162 	} else if (test_bit(dead_node, res->refmap)) {
2163 		mlog(0, "%s:%.*s: dead node %u had a ref, but had "
2164 		     "no locks and had not purged before dying\n", dlm->name,
2165 		     res->lockname.len, res->lockname.name, dead_node);
2166 		dlm_lockres_clear_refmap_bit(dead_node, res);
2167 	}
2168 
2169 	/* do not kick thread yet */
2170 	__dlm_dirty_lockres(dlm, res);
2171 }
2172 
2173 /* if this node is the recovery master, and there are no
2174  * locks for a given lockres owned by this node that are in
2175  * either PR or EX mode, zero out the lvb before requesting.
2176  *
2177  */
2178 
2179 
2180 static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node)
2181 {
2182 	struct hlist_node *iter;
2183 	struct dlm_lock_resource *res;
2184 	int i;
2185 	struct hlist_head *bucket;
2186 	struct dlm_lock *lock;
2187 
2188 
2189 	/* purge any stale mles */
2190 	dlm_clean_master_list(dlm, dead_node);
2191 
2192 	/*
2193 	 * now clean up all lock resources.  there are two rules:
2194 	 *
2195 	 * 1) if the dead node was the master, move the lockres
2196 	 *    to the recovering list.  set the RECOVERING flag.
2197 	 *    this lockres needs to be cleaned up before it can
2198 	 *    be used further.
2199 	 *
2200 	 * 2) if this node was the master, remove all locks from
2201 	 *    each of the lockres queues that were owned by the
2202 	 *    dead node.  once recovery finishes, the dlm thread
2203 	 *    can be kicked again to see if any ASTs or BASTs
2204 	 *    need to be fired as a result.
2205 	 */
2206 	for (i = 0; i < DLM_HASH_BUCKETS; i++) {
2207 		bucket = dlm_lockres_hash(dlm, i);
2208 		hlist_for_each_entry(res, iter, bucket, hash_node) {
2209  			/* always prune any $RECOVERY entries for dead nodes,
2210  			 * otherwise hangs can occur during later recovery */
2211 			if (dlm_is_recovery_lock(res->lockname.name,
2212 						 res->lockname.len)) {
2213 				spin_lock(&res->spinlock);
2214 				list_for_each_entry(lock, &res->granted, list) {
2215 					if (lock->ml.node == dead_node) {
2216 						mlog(0, "AHA! there was "
2217 						     "a $RECOVERY lock for dead "
2218 						     "node %u (%s)!\n",
2219 						     dead_node, dlm->name);
2220 						list_del_init(&lock->list);
2221 						dlm_lock_put(lock);
2222 						break;
2223 					}
2224 				}
2225 				spin_unlock(&res->spinlock);
2226 				continue;
2227 			}
2228 			spin_lock(&res->spinlock);
2229 			/* zero the lvb if necessary */
2230 			dlm_revalidate_lvb(dlm, res, dead_node);
2231 			if (res->owner == dead_node) {
2232 				if (res->state & DLM_LOCK_RES_DROPPING_REF)
2233 					mlog(0, "%s:%.*s: owned by "
2234 					     "dead node %u, this node was "
2235 					     "dropping its ref when it died. "
2236 					     "continue, dropping the flag.\n",
2237 					     dlm->name, res->lockname.len,
2238 					     res->lockname.name, dead_node);
2239 
2240 				/* the wake_up for this will happen when the
2241 				 * RECOVERING flag is dropped later */
2242 				res->state &= ~DLM_LOCK_RES_DROPPING_REF;
2243 
2244 				dlm_move_lockres_to_recovery_list(dlm, res);
2245 			} else if (res->owner == dlm->node_num) {
2246 				dlm_free_dead_locks(dlm, res, dead_node);
2247 				__dlm_lockres_calc_usage(dlm, res);
2248 			}
2249 			spin_unlock(&res->spinlock);
2250 		}
2251 	}
2252 
2253 }
2254 
2255 static void __dlm_hb_node_down(struct dlm_ctxt *dlm, int idx)
2256 {
2257 	assert_spin_locked(&dlm->spinlock);
2258 
2259 	if (dlm->reco.new_master == idx) {
2260 		mlog(0, "%s: recovery master %d just died\n",
2261 		     dlm->name, idx);
2262 		if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2263 			/* finalize1 was reached, so it is safe to clear
2264 			 * the new_master and dead_node.  that recovery
2265 			 * is complete. */
2266 			mlog(0, "%s: dead master %d had reached "
2267 			     "finalize1 state, clearing\n", dlm->name, idx);
2268 			dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE;
2269 			__dlm_reset_recovery(dlm);
2270 		}
2271 	}
2272 
2273 	/* Clean up join state on node death. */
2274 	if (dlm->joining_node == idx) {
2275 		mlog(0, "Clearing join state for node %u\n", idx);
2276 		__dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
2277 	}
2278 
2279 	/* check to see if the node is already considered dead */
2280 	if (!test_bit(idx, dlm->live_nodes_map)) {
2281 		mlog(0, "for domain %s, node %d is already dead. "
2282 		     "another node likely did recovery already.\n",
2283 		     dlm->name, idx);
2284 		return;
2285 	}
2286 
2287 	/* check to see if we do not care about this node */
2288 	if (!test_bit(idx, dlm->domain_map)) {
2289 		/* This also catches the case that we get a node down
2290 		 * but haven't joined the domain yet. */
2291 		mlog(0, "node %u already removed from domain!\n", idx);
2292 		return;
2293 	}
2294 
2295 	clear_bit(idx, dlm->live_nodes_map);
2296 
2297 	/* make sure local cleanup occurs before the heartbeat events */
2298 	if (!test_bit(idx, dlm->recovery_map))
2299 		dlm_do_local_recovery_cleanup(dlm, idx);
2300 
2301 	/* notify anything attached to the heartbeat events */
2302 	dlm_hb_event_notify_attached(dlm, idx, 0);
2303 
2304 	mlog(0, "node %u being removed from domain map!\n", idx);
2305 	clear_bit(idx, dlm->domain_map);
2306 	/* wake up migration waiters if a node goes down.
2307 	 * perhaps later we can genericize this for other waiters. */
2308 	wake_up(&dlm->migration_wq);
2309 
2310 	if (test_bit(idx, dlm->recovery_map))
2311 		mlog(0, "domain %s, node %u already added "
2312 		     "to recovery map!\n", dlm->name, idx);
2313 	else
2314 		set_bit(idx, dlm->recovery_map);
2315 }
2316 
2317 void dlm_hb_node_down_cb(struct o2nm_node *node, int idx, void *data)
2318 {
2319 	struct dlm_ctxt *dlm = data;
2320 
2321 	if (!dlm_grab(dlm))
2322 		return;
2323 
2324 	/*
2325 	 * This will notify any dlm users that a node in our domain
2326 	 * went away without notifying us first.
2327 	 */
2328 	if (test_bit(idx, dlm->domain_map))
2329 		dlm_fire_domain_eviction_callbacks(dlm, idx);
2330 
2331 	spin_lock(&dlm->spinlock);
2332 	__dlm_hb_node_down(dlm, idx);
2333 	spin_unlock(&dlm->spinlock);
2334 
2335 	dlm_put(dlm);
2336 }
2337 
2338 void dlm_hb_node_up_cb(struct o2nm_node *node, int idx, void *data)
2339 {
2340 	struct dlm_ctxt *dlm = data;
2341 
2342 	if (!dlm_grab(dlm))
2343 		return;
2344 
2345 	spin_lock(&dlm->spinlock);
2346 	set_bit(idx, dlm->live_nodes_map);
2347 	/* do NOT notify mle attached to the heartbeat events.
2348 	 * new nodes are not interesting in mastery until joined. */
2349 	spin_unlock(&dlm->spinlock);
2350 
2351 	dlm_put(dlm);
2352 }
2353 
2354 static void dlm_reco_ast(void *astdata)
2355 {
2356 	struct dlm_ctxt *dlm = astdata;
2357 	mlog(0, "ast for recovery lock fired!, this=%u, dlm=%s\n",
2358 	     dlm->node_num, dlm->name);
2359 }
2360 static void dlm_reco_bast(void *astdata, int blocked_type)
2361 {
2362 	struct dlm_ctxt *dlm = astdata;
2363 	mlog(0, "bast for recovery lock fired!, this=%u, dlm=%s\n",
2364 	     dlm->node_num, dlm->name);
2365 }
2366 static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st)
2367 {
2368 	mlog(0, "unlockast for recovery lock fired!\n");
2369 }
2370 
2371 /*
2372  * dlm_pick_recovery_master will continually attempt to use
2373  * dlmlock() on the special "$RECOVERY" lockres with the
2374  * LKM_NOQUEUE flag to get an EX.  every thread that enters
2375  * this function on each node racing to become the recovery
2376  * master will not stop attempting this until either:
2377  * a) this node gets the EX (and becomes the recovery master),
2378  * or b) dlm->reco.new_master gets set to some nodenum
2379  * != O2NM_INVALID_NODE_NUM (another node will do the reco).
2380  * so each time a recovery master is needed, the entire cluster
2381  * will sync at this point.  if the new master dies, that will
2382  * be detected in dlm_do_recovery */
2383 static int dlm_pick_recovery_master(struct dlm_ctxt *dlm)
2384 {
2385 	enum dlm_status ret;
2386 	struct dlm_lockstatus lksb;
2387 	int status = -EINVAL;
2388 
2389 	mlog(0, "starting recovery of %s at %lu, dead=%u, this=%u\n",
2390 	     dlm->name, jiffies, dlm->reco.dead_node, dlm->node_num);
2391 again:
2392 	memset(&lksb, 0, sizeof(lksb));
2393 
2394 	ret = dlmlock(dlm, LKM_EXMODE, &lksb, LKM_NOQUEUE|LKM_RECOVERY,
2395 		      DLM_RECOVERY_LOCK_NAME, DLM_RECOVERY_LOCK_NAME_LEN,
2396 		      dlm_reco_ast, dlm, dlm_reco_bast);
2397 
2398 	mlog(0, "%s: dlmlock($RECOVERY) returned %d, lksb=%d\n",
2399 	     dlm->name, ret, lksb.status);
2400 
2401 	if (ret == DLM_NORMAL) {
2402 		mlog(0, "dlm=%s dlmlock says I got it (this=%u)\n",
2403 		     dlm->name, dlm->node_num);
2404 
2405 		/* got the EX lock.  check to see if another node
2406 		 * just became the reco master */
2407 		if (dlm_reco_master_ready(dlm)) {
2408 			mlog(0, "%s: got reco EX lock, but %u will "
2409 			     "do the recovery\n", dlm->name,
2410 			     dlm->reco.new_master);
2411 			status = -EEXIST;
2412 		} else {
2413 			status = 0;
2414 
2415 			/* see if recovery was already finished elsewhere */
2416 			spin_lock(&dlm->spinlock);
2417 			if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
2418 				status = -EINVAL;
2419 				mlog(0, "%s: got reco EX lock, but "
2420 				     "node got recovered already\n", dlm->name);
2421 				if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) {
2422 					mlog(ML_ERROR, "%s: new master is %u "
2423 					     "but no dead node!\n",
2424 					     dlm->name, dlm->reco.new_master);
2425 					BUG();
2426 				}
2427 			}
2428 			spin_unlock(&dlm->spinlock);
2429 		}
2430 
2431 		/* if this node has actually become the recovery master,
2432 		 * set the master and send the messages to begin recovery */
2433 		if (!status) {
2434 			mlog(0, "%s: dead=%u, this=%u, sending "
2435 			     "begin_reco now\n", dlm->name,
2436 			     dlm->reco.dead_node, dlm->node_num);
2437 			status = dlm_send_begin_reco_message(dlm,
2438 				      dlm->reco.dead_node);
2439 			/* this always succeeds */
2440 			BUG_ON(status);
2441 
2442 			/* set the new_master to this node */
2443 			spin_lock(&dlm->spinlock);
2444 			dlm_set_reco_master(dlm, dlm->node_num);
2445 			spin_unlock(&dlm->spinlock);
2446 		}
2447 
2448 		/* recovery lock is a special case.  ast will not get fired,
2449 		 * so just go ahead and unlock it. */
2450 		ret = dlmunlock(dlm, &lksb, 0, dlm_reco_unlock_ast, dlm);
2451 		if (ret == DLM_DENIED) {
2452 			mlog(0, "got DLM_DENIED, trying LKM_CANCEL\n");
2453 			ret = dlmunlock(dlm, &lksb, LKM_CANCEL, dlm_reco_unlock_ast, dlm);
2454 		}
2455 		if (ret != DLM_NORMAL) {
2456 			/* this would really suck. this could only happen
2457 			 * if there was a network error during the unlock
2458 			 * because of node death.  this means the unlock
2459 			 * is actually "done" and the lock structure is
2460 			 * even freed.  we can continue, but only
2461 			 * because this specific lock name is special. */
2462 			mlog(ML_ERROR, "dlmunlock returned %d\n", ret);
2463 		}
2464 	} else if (ret == DLM_NOTQUEUED) {
2465 		mlog(0, "dlm=%s dlmlock says another node got it (this=%u)\n",
2466 		     dlm->name, dlm->node_num);
2467 		/* another node is master. wait on
2468 		 * reco.new_master != O2NM_INVALID_NODE_NUM
2469 		 * for at most one second */
2470 		wait_event_timeout(dlm->dlm_reco_thread_wq,
2471 					 dlm_reco_master_ready(dlm),
2472 					 msecs_to_jiffies(1000));
2473 		if (!dlm_reco_master_ready(dlm)) {
2474 			mlog(0, "%s: reco master taking awhile\n",
2475 			     dlm->name);
2476 			goto again;
2477 		}
2478 		/* another node has informed this one that it is reco master */
2479 		mlog(0, "%s: reco master %u is ready to recover %u\n",
2480 		     dlm->name, dlm->reco.new_master, dlm->reco.dead_node);
2481 		status = -EEXIST;
2482 	} else if (ret == DLM_RECOVERING) {
2483 		mlog(0, "dlm=%s dlmlock says master node died (this=%u)\n",
2484 		     dlm->name, dlm->node_num);
2485 		goto again;
2486 	} else {
2487 		struct dlm_lock_resource *res;
2488 
2489 		/* dlmlock returned something other than NOTQUEUED or NORMAL */
2490 		mlog(ML_ERROR, "%s: got %s from dlmlock($RECOVERY), "
2491 		     "lksb.status=%s\n", dlm->name, dlm_errname(ret),
2492 		     dlm_errname(lksb.status));
2493 		res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME,
2494 					 DLM_RECOVERY_LOCK_NAME_LEN);
2495 		if (res) {
2496 			dlm_print_one_lock_resource(res);
2497 			dlm_lockres_put(res);
2498 		} else {
2499 			mlog(ML_ERROR, "recovery lock not found\n");
2500 		}
2501 		BUG();
2502 	}
2503 
2504 	return status;
2505 }
2506 
2507 static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node)
2508 {
2509 	struct dlm_begin_reco br;
2510 	int ret = 0;
2511 	struct dlm_node_iter iter;
2512 	int nodenum;
2513 	int status;
2514 
2515 	mlog_entry("%u\n", dead_node);
2516 
2517 	mlog(0, "%s: dead node is %u\n", dlm->name, dead_node);
2518 
2519 	spin_lock(&dlm->spinlock);
2520 	dlm_node_iter_init(dlm->domain_map, &iter);
2521 	spin_unlock(&dlm->spinlock);
2522 
2523 	clear_bit(dead_node, iter.node_map);
2524 
2525 	memset(&br, 0, sizeof(br));
2526 	br.node_idx = dlm->node_num;
2527 	br.dead_node = dead_node;
2528 
2529 	while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2530 		ret = 0;
2531 		if (nodenum == dead_node) {
2532 			mlog(0, "not sending begin reco to dead node "
2533 				  "%u\n", dead_node);
2534 			continue;
2535 		}
2536 		if (nodenum == dlm->node_num) {
2537 			mlog(0, "not sending begin reco to self\n");
2538 			continue;
2539 		}
2540 retry:
2541 		ret = -EINVAL;
2542 		mlog(0, "attempting to send begin reco msg to %d\n",
2543 			  nodenum);
2544 		ret = o2net_send_message(DLM_BEGIN_RECO_MSG, dlm->key,
2545 					 &br, sizeof(br), nodenum, &status);
2546 		/* negative status is handled ok by caller here */
2547 		if (ret >= 0)
2548 			ret = status;
2549 		if (dlm_is_host_down(ret)) {
2550 			/* node is down.  not involved in recovery
2551 			 * so just keep going */
2552 			mlog(0, "%s: node %u was down when sending "
2553 			     "begin reco msg (%d)\n", dlm->name, nodenum, ret);
2554 			ret = 0;
2555 		}
2556 		if (ret < 0) {
2557 			struct dlm_lock_resource *res;
2558 			/* this is now a serious problem, possibly ENOMEM
2559 			 * in the network stack.  must retry */
2560 			mlog_errno(ret);
2561 			mlog(ML_ERROR, "begin reco of dlm %s to node %u "
2562 			    " returned %d\n", dlm->name, nodenum, ret);
2563 			res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME,
2564 						 DLM_RECOVERY_LOCK_NAME_LEN);
2565 			if (res) {
2566 				dlm_print_one_lock_resource(res);
2567 				dlm_lockres_put(res);
2568 			} else {
2569 				mlog(ML_ERROR, "recovery lock not found\n");
2570 			}
2571 			/* sleep for a bit in hopes that we can avoid
2572 			 * another ENOMEM */
2573 			msleep(100);
2574 			goto retry;
2575 		} else if (ret == EAGAIN) {
2576 			mlog(0, "%s: trying to start recovery of node "
2577 			     "%u, but node %u is waiting for last recovery "
2578 			     "to complete, backoff for a bit\n", dlm->name,
2579 			     dead_node, nodenum);
2580 			/* TODO Look into replacing msleep with cond_resched() */
2581 			msleep(100);
2582 			goto retry;
2583 		}
2584 	}
2585 
2586 	return ret;
2587 }
2588 
2589 int dlm_begin_reco_handler(struct o2net_msg *msg, u32 len, void *data,
2590 			   void **ret_data)
2591 {
2592 	struct dlm_ctxt *dlm = data;
2593 	struct dlm_begin_reco *br = (struct dlm_begin_reco *)msg->buf;
2594 
2595 	/* ok to return 0, domain has gone away */
2596 	if (!dlm_grab(dlm))
2597 		return 0;
2598 
2599 	spin_lock(&dlm->spinlock);
2600 	if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2601 		mlog(0, "%s: node %u wants to recover node %u (%u:%u) "
2602 		     "but this node is in finalize state, waiting on finalize2\n",
2603 		     dlm->name, br->node_idx, br->dead_node,
2604 		     dlm->reco.dead_node, dlm->reco.new_master);
2605 		spin_unlock(&dlm->spinlock);
2606 		return EAGAIN;
2607 	}
2608 	spin_unlock(&dlm->spinlock);
2609 
2610 	mlog(0, "%s: node %u wants to recover node %u (%u:%u)\n",
2611 	     dlm->name, br->node_idx, br->dead_node,
2612 	     dlm->reco.dead_node, dlm->reco.new_master);
2613 
2614 	dlm_fire_domain_eviction_callbacks(dlm, br->dead_node);
2615 
2616 	spin_lock(&dlm->spinlock);
2617 	if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) {
2618 		if (test_bit(dlm->reco.new_master, dlm->recovery_map)) {
2619 			mlog(0, "%s: new_master %u died, changing "
2620 			     "to %u\n", dlm->name, dlm->reco.new_master,
2621 			     br->node_idx);
2622 		} else {
2623 			mlog(0, "%s: new_master %u NOT DEAD, changing "
2624 			     "to %u\n", dlm->name, dlm->reco.new_master,
2625 			     br->node_idx);
2626 			/* may not have seen the new master as dead yet */
2627 		}
2628 	}
2629 	if (dlm->reco.dead_node != O2NM_INVALID_NODE_NUM) {
2630 		mlog(ML_NOTICE, "%s: dead_node previously set to %u, "
2631 		     "node %u changing it to %u\n", dlm->name,
2632 		     dlm->reco.dead_node, br->node_idx, br->dead_node);
2633 	}
2634 	dlm_set_reco_master(dlm, br->node_idx);
2635 	dlm_set_reco_dead_node(dlm, br->dead_node);
2636 	if (!test_bit(br->dead_node, dlm->recovery_map)) {
2637 		mlog(0, "recovery master %u sees %u as dead, but this "
2638 		     "node has not yet.  marking %u as dead\n",
2639 		     br->node_idx, br->dead_node, br->dead_node);
2640 		if (!test_bit(br->dead_node, dlm->domain_map) ||
2641 		    !test_bit(br->dead_node, dlm->live_nodes_map))
2642 			mlog(0, "%u not in domain/live_nodes map "
2643 			     "so setting it in reco map manually\n",
2644 			     br->dead_node);
2645 		/* force the recovery cleanup in __dlm_hb_node_down
2646 		 * both of these will be cleared in a moment */
2647 		set_bit(br->dead_node, dlm->domain_map);
2648 		set_bit(br->dead_node, dlm->live_nodes_map);
2649 		__dlm_hb_node_down(dlm, br->dead_node);
2650 	}
2651 	spin_unlock(&dlm->spinlock);
2652 
2653 	dlm_kick_recovery_thread(dlm);
2654 
2655 	mlog(0, "%s: recovery started by node %u, for %u (%u:%u)\n",
2656 	     dlm->name, br->node_idx, br->dead_node,
2657 	     dlm->reco.dead_node, dlm->reco.new_master);
2658 
2659 	dlm_put(dlm);
2660 	return 0;
2661 }
2662 
2663 #define DLM_FINALIZE_STAGE2  0x01
2664 static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm)
2665 {
2666 	int ret = 0;
2667 	struct dlm_finalize_reco fr;
2668 	struct dlm_node_iter iter;
2669 	int nodenum;
2670 	int status;
2671 	int stage = 1;
2672 
2673 	mlog(0, "finishing recovery for node %s:%u, "
2674 	     "stage %d\n", dlm->name, dlm->reco.dead_node, stage);
2675 
2676 	spin_lock(&dlm->spinlock);
2677 	dlm_node_iter_init(dlm->domain_map, &iter);
2678 	spin_unlock(&dlm->spinlock);
2679 
2680 stage2:
2681 	memset(&fr, 0, sizeof(fr));
2682 	fr.node_idx = dlm->node_num;
2683 	fr.dead_node = dlm->reco.dead_node;
2684 	if (stage == 2)
2685 		fr.flags |= DLM_FINALIZE_STAGE2;
2686 
2687 	while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2688 		if (nodenum == dlm->node_num)
2689 			continue;
2690 		ret = o2net_send_message(DLM_FINALIZE_RECO_MSG, dlm->key,
2691 					 &fr, sizeof(fr), nodenum, &status);
2692 		if (ret >= 0)
2693 			ret = status;
2694 		if (ret < 0) {
2695 			mlog_errno(ret);
2696 			if (dlm_is_host_down(ret)) {
2697 				/* this has no effect on this recovery
2698 				 * session, so set the status to zero to
2699 				 * finish out the last recovery */
2700 				mlog(ML_ERROR, "node %u went down after this "
2701 				     "node finished recovery.\n", nodenum);
2702 				ret = 0;
2703 				continue;
2704 			}
2705 			break;
2706 		}
2707 	}
2708 	if (stage == 1) {
2709 		/* reset the node_iter back to the top and send finalize2 */
2710 		iter.curnode = -1;
2711 		stage = 2;
2712 		goto stage2;
2713 	}
2714 
2715 	return ret;
2716 }
2717 
2718 int dlm_finalize_reco_handler(struct o2net_msg *msg, u32 len, void *data,
2719 			      void **ret_data)
2720 {
2721 	struct dlm_ctxt *dlm = data;
2722 	struct dlm_finalize_reco *fr = (struct dlm_finalize_reco *)msg->buf;
2723 	int stage = 1;
2724 
2725 	/* ok to return 0, domain has gone away */
2726 	if (!dlm_grab(dlm))
2727 		return 0;
2728 
2729 	if (fr->flags & DLM_FINALIZE_STAGE2)
2730 		stage = 2;
2731 
2732 	mlog(0, "%s: node %u finalizing recovery stage%d of "
2733 	     "node %u (%u:%u)\n", dlm->name, fr->node_idx, stage,
2734 	     fr->dead_node, dlm->reco.dead_node, dlm->reco.new_master);
2735 
2736 	spin_lock(&dlm->spinlock);
2737 
2738 	if (dlm->reco.new_master != fr->node_idx) {
2739 		mlog(ML_ERROR, "node %u sent recovery finalize msg, but node "
2740 		     "%u is supposed to be the new master, dead=%u\n",
2741 		     fr->node_idx, dlm->reco.new_master, fr->dead_node);
2742 		BUG();
2743 	}
2744 	if (dlm->reco.dead_node != fr->dead_node) {
2745 		mlog(ML_ERROR, "node %u sent recovery finalize msg for dead "
2746 		     "node %u, but node %u is supposed to be dead\n",
2747 		     fr->node_idx, fr->dead_node, dlm->reco.dead_node);
2748 		BUG();
2749 	}
2750 
2751 	switch (stage) {
2752 		case 1:
2753 			dlm_finish_local_lockres_recovery(dlm, fr->dead_node, fr->node_idx);
2754 			if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2755 				mlog(ML_ERROR, "%s: received finalize1 from "
2756 				     "new master %u for dead node %u, but "
2757 				     "this node has already received it!\n",
2758 				     dlm->name, fr->node_idx, fr->dead_node);
2759 				dlm_print_reco_node_status(dlm);
2760 				BUG();
2761 			}
2762 			dlm->reco.state |= DLM_RECO_STATE_FINALIZE;
2763 			spin_unlock(&dlm->spinlock);
2764 			break;
2765 		case 2:
2766 			if (!(dlm->reco.state & DLM_RECO_STATE_FINALIZE)) {
2767 				mlog(ML_ERROR, "%s: received finalize2 from "
2768 				     "new master %u for dead node %u, but "
2769 				     "this node did not have finalize1!\n",
2770 				     dlm->name, fr->node_idx, fr->dead_node);
2771 				dlm_print_reco_node_status(dlm);
2772 				BUG();
2773 			}
2774 			dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE;
2775 			spin_unlock(&dlm->spinlock);
2776 			dlm_reset_recovery(dlm);
2777 			dlm_kick_recovery_thread(dlm);
2778 			break;
2779 		default:
2780 			BUG();
2781 	}
2782 
2783 	mlog(0, "%s: recovery done, reco master was %u, dead now %u, master now %u\n",
2784 	     dlm->name, fr->node_idx, dlm->reco.dead_node, dlm->reco.new_master);
2785 
2786 	dlm_put(dlm);
2787 	return 0;
2788 }
2789