xref: /linux/fs/ocfs2/dlm/dlmrecovery.c (revision 7ae9fb1b7ecbb5d85d07857943f677fd1a559b18)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * dlmrecovery.c
4  *
5  * recovery stuff
6  *
7  * Copyright (C) 2004 Oracle.  All rights reserved.
8  */
9 
10 
11 #include <linux/module.h>
12 #include <linux/fs.h>
13 #include <linux/types.h>
14 #include <linux/slab.h>
15 #include <linux/highmem.h>
16 #include <linux/init.h>
17 #include <linux/sysctl.h>
18 #include <linux/random.h>
19 #include <linux/blkdev.h>
20 #include <linux/socket.h>
21 #include <linux/inet.h>
22 #include <linux/timer.h>
23 #include <linux/kthread.h>
24 #include <linux/delay.h>
25 
26 
27 #include "../cluster/heartbeat.h"
28 #include "../cluster/nodemanager.h"
29 #include "../cluster/tcp.h"
30 
31 #include "dlmapi.h"
32 #include "dlmcommon.h"
33 #include "dlmdomain.h"
34 
35 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_RECOVERY)
36 #include "../cluster/masklog.h"
37 
38 static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node);
39 
40 static int dlm_recovery_thread(void *data);
41 static int dlm_do_recovery(struct dlm_ctxt *dlm);
42 
43 static int dlm_pick_recovery_master(struct dlm_ctxt *dlm);
44 static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node);
45 static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node);
46 static int dlm_request_all_locks(struct dlm_ctxt *dlm,
47 				 u8 request_from, u8 dead_node);
48 static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm);
49 
50 static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res);
51 static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres,
52 					const char *lockname, int namelen,
53 					int total_locks, u64 cookie,
54 					u8 flags, u8 master);
55 static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm,
56 				    struct dlm_migratable_lockres *mres,
57 				    u8 send_to,
58 				    struct dlm_lock_resource *res,
59 				    int total_locks);
60 static int dlm_process_recovery_data(struct dlm_ctxt *dlm,
61 				     struct dlm_lock_resource *res,
62 				     struct dlm_migratable_lockres *mres);
63 static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm);
64 static int dlm_send_all_done_msg(struct dlm_ctxt *dlm,
65 				 u8 dead_node, u8 send_to);
66 static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node);
67 static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm,
68 					struct list_head *list, u8 dead_node);
69 static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm,
70 					      u8 dead_node, u8 new_master);
71 static void dlm_reco_ast(void *astdata);
72 static void dlm_reco_bast(void *astdata, int blocked_type);
73 static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st);
74 static void dlm_request_all_locks_worker(struct dlm_work_item *item,
75 					 void *data);
76 static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data);
77 static int dlm_lockres_master_requery(struct dlm_ctxt *dlm,
78 				      struct dlm_lock_resource *res,
79 				      u8 *real_master);
80 
81 static u64 dlm_get_next_mig_cookie(void);
82 
83 static DEFINE_SPINLOCK(dlm_reco_state_lock);
84 static DEFINE_SPINLOCK(dlm_mig_cookie_lock);
85 static u64 dlm_mig_cookie = 1;
86 
dlm_get_next_mig_cookie(void)87 static u64 dlm_get_next_mig_cookie(void)
88 {
89 	u64 c;
90 	spin_lock(&dlm_mig_cookie_lock);
91 	c = dlm_mig_cookie;
92 	if (dlm_mig_cookie == (~0ULL))
93 		dlm_mig_cookie = 1;
94 	else
95 		dlm_mig_cookie++;
96 	spin_unlock(&dlm_mig_cookie_lock);
97 	return c;
98 }
99 
dlm_set_reco_dead_node(struct dlm_ctxt * dlm,u8 dead_node)100 static inline void dlm_set_reco_dead_node(struct dlm_ctxt *dlm,
101 					  u8 dead_node)
102 {
103 	assert_spin_locked(&dlm->spinlock);
104 	if (dlm->reco.dead_node != dead_node)
105 		mlog(0, "%s: changing dead_node from %u to %u\n",
106 		     dlm->name, dlm->reco.dead_node, dead_node);
107 	dlm->reco.dead_node = dead_node;
108 }
109 
dlm_set_reco_master(struct dlm_ctxt * dlm,u8 master)110 static inline void dlm_set_reco_master(struct dlm_ctxt *dlm,
111 				       u8 master)
112 {
113 	assert_spin_locked(&dlm->spinlock);
114 	mlog(0, "%s: changing new_master from %u to %u\n",
115 	     dlm->name, dlm->reco.new_master, master);
116 	dlm->reco.new_master = master;
117 }
118 
__dlm_reset_recovery(struct dlm_ctxt * dlm)119 static inline void __dlm_reset_recovery(struct dlm_ctxt *dlm)
120 {
121 	assert_spin_locked(&dlm->spinlock);
122 	clear_bit(dlm->reco.dead_node, dlm->recovery_map);
123 	dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
124 	dlm_set_reco_master(dlm, O2NM_INVALID_NODE_NUM);
125 }
126 
127 /* Worker function used during recovery. */
dlm_dispatch_work(struct work_struct * work)128 void dlm_dispatch_work(struct work_struct *work)
129 {
130 	struct dlm_ctxt *dlm =
131 		container_of(work, struct dlm_ctxt, dispatched_work);
132 	LIST_HEAD(tmp_list);
133 	struct dlm_work_item *item, *next;
134 	dlm_workfunc_t *workfunc;
135 	int tot=0;
136 
137 	spin_lock(&dlm->work_lock);
138 	list_splice_init(&dlm->work_list, &tmp_list);
139 	spin_unlock(&dlm->work_lock);
140 
141 	list_for_each_entry(item, &tmp_list, list) {
142 		tot++;
143 	}
144 	mlog(0, "%s: work thread has %d work items\n", dlm->name, tot);
145 
146 	list_for_each_entry_safe(item, next, &tmp_list, list) {
147 		workfunc = item->func;
148 		list_del_init(&item->list);
149 
150 		/* already have ref on dlm to avoid having
151 		 * it disappear.  just double-check. */
152 		BUG_ON(item->dlm != dlm);
153 
154 		/* this is allowed to sleep and
155 		 * call network stuff */
156 		workfunc(item, item->data);
157 
158 		dlm_put(dlm);
159 		kfree(item);
160 	}
161 }
162 
163 /*
164  * RECOVERY THREAD
165  */
166 
dlm_kick_recovery_thread(struct dlm_ctxt * dlm)167 void dlm_kick_recovery_thread(struct dlm_ctxt *dlm)
168 {
169 	/* wake the recovery thread
170 	 * this will wake the reco thread in one of three places
171 	 * 1) sleeping with no recovery happening
172 	 * 2) sleeping with recovery mastered elsewhere
173 	 * 3) recovery mastered here, waiting on reco data */
174 
175 	wake_up(&dlm->dlm_reco_thread_wq);
176 }
177 
178 /* Launch the recovery thread */
dlm_launch_recovery_thread(struct dlm_ctxt * dlm)179 int dlm_launch_recovery_thread(struct dlm_ctxt *dlm)
180 {
181 	mlog(0, "starting dlm recovery thread...\n");
182 
183 	dlm->dlm_reco_thread_task = kthread_run(dlm_recovery_thread, dlm,
184 			"dlm_reco-%s", dlm->name);
185 	if (IS_ERR(dlm->dlm_reco_thread_task)) {
186 		mlog_errno(PTR_ERR(dlm->dlm_reco_thread_task));
187 		dlm->dlm_reco_thread_task = NULL;
188 		return -EINVAL;
189 	}
190 
191 	return 0;
192 }
193 
dlm_complete_recovery_thread(struct dlm_ctxt * dlm)194 void dlm_complete_recovery_thread(struct dlm_ctxt *dlm)
195 {
196 	if (dlm->dlm_reco_thread_task) {
197 		mlog(0, "waiting for dlm recovery thread to exit\n");
198 		kthread_stop(dlm->dlm_reco_thread_task);
199 		dlm->dlm_reco_thread_task = NULL;
200 	}
201 }
202 
203 
204 
205 /*
206  * this is lame, but here's how recovery works...
207  * 1) all recovery threads cluster wide will work on recovering
208  *    ONE node at a time
209  * 2) negotiate who will take over all the locks for the dead node.
210  *    thats right... ALL the locks.
211  * 3) once a new master is chosen, everyone scans all locks
212  *    and moves aside those mastered by the dead guy
213  * 4) each of these locks should be locked until recovery is done
214  * 5) the new master collects up all of secondary lock queue info
215  *    one lock at a time, forcing each node to communicate back
216  *    before continuing
217  * 6) each secondary lock queue responds with the full known lock info
218  * 7) once the new master has run all its locks, it sends a ALLDONE!
219  *    message to everyone
220  * 8) upon receiving this message, the secondary queue node unlocks
221  *    and responds to the ALLDONE
222  * 9) once the new master gets responses from everyone, he unlocks
223  *    everything and recovery for this dead node is done
224  *10) go back to 2) while there are still dead nodes
225  *
226  */
227 
dlm_print_reco_node_status(struct dlm_ctxt * dlm)228 static void dlm_print_reco_node_status(struct dlm_ctxt *dlm)
229 {
230 	struct dlm_reco_node_data *ndata;
231 	struct dlm_lock_resource *res;
232 
233 	mlog(ML_NOTICE, "%s(%d): recovery info, state=%s, dead=%u, master=%u\n",
234 	     dlm->name, task_pid_nr(dlm->dlm_reco_thread_task),
235 	     dlm->reco.state & DLM_RECO_STATE_ACTIVE ? "ACTIVE" : "inactive",
236 	     dlm->reco.dead_node, dlm->reco.new_master);
237 
238 	list_for_each_entry(ndata, &dlm->reco.node_data, list) {
239 		char *st = "unknown";
240 		switch (ndata->state) {
241 			case DLM_RECO_NODE_DATA_INIT:
242 				st = "init";
243 				break;
244 			case DLM_RECO_NODE_DATA_REQUESTING:
245 				st = "requesting";
246 				break;
247 			case DLM_RECO_NODE_DATA_DEAD:
248 				st = "dead";
249 				break;
250 			case DLM_RECO_NODE_DATA_RECEIVING:
251 				st = "receiving";
252 				break;
253 			case DLM_RECO_NODE_DATA_REQUESTED:
254 				st = "requested";
255 				break;
256 			case DLM_RECO_NODE_DATA_DONE:
257 				st = "done";
258 				break;
259 			case DLM_RECO_NODE_DATA_FINALIZE_SENT:
260 				st = "finalize-sent";
261 				break;
262 			default:
263 				st = "bad";
264 				break;
265 		}
266 		mlog(ML_NOTICE, "%s: reco state, node %u, state=%s\n",
267 		     dlm->name, ndata->node_num, st);
268 	}
269 	list_for_each_entry(res, &dlm->reco.resources, recovering) {
270 		mlog(ML_NOTICE, "%s: lockres %.*s on recovering list\n",
271 		     dlm->name, res->lockname.len, res->lockname.name);
272 	}
273 }
274 
275 #define DLM_RECO_THREAD_TIMEOUT_MS (5 * 1000)
276 
dlm_recovery_thread(void * data)277 static int dlm_recovery_thread(void *data)
278 {
279 	int status;
280 	struct dlm_ctxt *dlm = data;
281 	unsigned long timeout = msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS);
282 
283 	mlog(0, "dlm thread running for %s...\n", dlm->name);
284 
285 	while (!kthread_should_stop()) {
286 		if (dlm_domain_fully_joined(dlm)) {
287 			status = dlm_do_recovery(dlm);
288 			if (status == -EAGAIN) {
289 				/* do not sleep, recheck immediately. */
290 				continue;
291 			}
292 			if (status < 0)
293 				mlog_errno(status);
294 		}
295 
296 		wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq,
297 						 kthread_should_stop(),
298 						 timeout);
299 	}
300 
301 	mlog(0, "quitting DLM recovery thread\n");
302 	return 0;
303 }
304 
305 /* returns true when the recovery master has contacted us */
dlm_reco_master_ready(struct dlm_ctxt * dlm)306 static int dlm_reco_master_ready(struct dlm_ctxt *dlm)
307 {
308 	int ready;
309 	spin_lock(&dlm->spinlock);
310 	ready = (dlm->reco.new_master != O2NM_INVALID_NODE_NUM);
311 	spin_unlock(&dlm->spinlock);
312 	return ready;
313 }
314 
315 /* returns true if node is no longer in the domain
316  * could be dead or just not joined */
dlm_is_node_dead(struct dlm_ctxt * dlm,u8 node)317 int dlm_is_node_dead(struct dlm_ctxt *dlm, u8 node)
318 {
319 	int dead;
320 	spin_lock(&dlm->spinlock);
321 	dead = !test_bit(node, dlm->domain_map);
322 	spin_unlock(&dlm->spinlock);
323 	return dead;
324 }
325 
326 /* returns true if node is no longer in the domain
327  * could be dead or just not joined */
dlm_is_node_recovered(struct dlm_ctxt * dlm,u8 node)328 static int dlm_is_node_recovered(struct dlm_ctxt *dlm, u8 node)
329 {
330 	int recovered;
331 	spin_lock(&dlm->spinlock);
332 	recovered = !test_bit(node, dlm->recovery_map);
333 	spin_unlock(&dlm->spinlock);
334 	return recovered;
335 }
336 
337 
dlm_wait_for_node_death(struct dlm_ctxt * dlm,u8 node,int timeout)338 void dlm_wait_for_node_death(struct dlm_ctxt *dlm, u8 node, int timeout)
339 {
340 	if (dlm_is_node_dead(dlm, node))
341 		return;
342 
343 	printk(KERN_NOTICE "o2dlm: Waiting on the death of node %u in "
344 	       "domain %s\n", node, dlm->name);
345 
346 	if (timeout)
347 		wait_event_timeout(dlm->dlm_reco_thread_wq,
348 				   dlm_is_node_dead(dlm, node),
349 				   msecs_to_jiffies(timeout));
350 	else
351 		wait_event(dlm->dlm_reco_thread_wq,
352 			   dlm_is_node_dead(dlm, node));
353 }
354 
dlm_wait_for_node_recovery(struct dlm_ctxt * dlm,u8 node,int timeout)355 void dlm_wait_for_node_recovery(struct dlm_ctxt *dlm, u8 node, int timeout)
356 {
357 	if (dlm_is_node_recovered(dlm, node))
358 		return;
359 
360 	printk(KERN_NOTICE "o2dlm: Waiting on the recovery of node %u in "
361 	       "domain %s\n", node, dlm->name);
362 
363 	if (timeout)
364 		wait_event_timeout(dlm->dlm_reco_thread_wq,
365 				   dlm_is_node_recovered(dlm, node),
366 				   msecs_to_jiffies(timeout));
367 	else
368 		wait_event(dlm->dlm_reco_thread_wq,
369 			   dlm_is_node_recovered(dlm, node));
370 }
371 
372 /* callers of the top-level api calls (dlmlock/dlmunlock) should
373  * block on the dlm->reco.event when recovery is in progress.
374  * the dlm recovery thread will set this state when it begins
375  * recovering a dead node (as the new master or not) and clear
376  * the state and wake as soon as all affected lock resources have
377  * been marked with the RECOVERY flag */
dlm_in_recovery(struct dlm_ctxt * dlm)378 static int dlm_in_recovery(struct dlm_ctxt *dlm)
379 {
380 	int in_recovery;
381 	spin_lock(&dlm->spinlock);
382 	in_recovery = !!(dlm->reco.state & DLM_RECO_STATE_ACTIVE);
383 	spin_unlock(&dlm->spinlock);
384 	return in_recovery;
385 }
386 
387 
dlm_wait_for_recovery(struct dlm_ctxt * dlm)388 void dlm_wait_for_recovery(struct dlm_ctxt *dlm)
389 {
390 	if (dlm_in_recovery(dlm)) {
391 		mlog(0, "%s: reco thread %d in recovery: "
392 		     "state=%d, master=%u, dead=%u\n",
393 		     dlm->name, task_pid_nr(dlm->dlm_reco_thread_task),
394 		     dlm->reco.state, dlm->reco.new_master,
395 		     dlm->reco.dead_node);
396 	}
397 	wait_event(dlm->reco.event, !dlm_in_recovery(dlm));
398 }
399 
dlm_begin_recovery(struct dlm_ctxt * dlm)400 static void dlm_begin_recovery(struct dlm_ctxt *dlm)
401 {
402 	assert_spin_locked(&dlm->spinlock);
403 	BUG_ON(dlm->reco.state & DLM_RECO_STATE_ACTIVE);
404 	printk(KERN_NOTICE "o2dlm: Begin recovery on domain %s for node %u\n",
405 	       dlm->name, dlm->reco.dead_node);
406 	dlm->reco.state |= DLM_RECO_STATE_ACTIVE;
407 }
408 
dlm_end_recovery(struct dlm_ctxt * dlm)409 static void dlm_end_recovery(struct dlm_ctxt *dlm)
410 {
411 	spin_lock(&dlm->spinlock);
412 	BUG_ON(!(dlm->reco.state & DLM_RECO_STATE_ACTIVE));
413 	dlm->reco.state &= ~DLM_RECO_STATE_ACTIVE;
414 	spin_unlock(&dlm->spinlock);
415 	printk(KERN_NOTICE "o2dlm: End recovery on domain %s\n", dlm->name);
416 	wake_up(&dlm->reco.event);
417 }
418 
dlm_print_recovery_master(struct dlm_ctxt * dlm)419 static void dlm_print_recovery_master(struct dlm_ctxt *dlm)
420 {
421 	printk(KERN_NOTICE "o2dlm: Node %u (%s) is the Recovery Master for the "
422 	       "dead node %u in domain %s\n", dlm->reco.new_master,
423 	       (dlm->node_num == dlm->reco.new_master ? "me" : "he"),
424 	       dlm->reco.dead_node, dlm->name);
425 }
426 
dlm_do_recovery(struct dlm_ctxt * dlm)427 static int dlm_do_recovery(struct dlm_ctxt *dlm)
428 {
429 	int status = 0;
430 	int ret;
431 
432 	spin_lock(&dlm->spinlock);
433 
434 	if (dlm->migrate_done) {
435 		mlog(0, "%s: no need do recovery after migrating all "
436 		     "lock resources\n", dlm->name);
437 		spin_unlock(&dlm->spinlock);
438 		return 0;
439 	}
440 
441 	/* check to see if the new master has died */
442 	if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM &&
443 	    test_bit(dlm->reco.new_master, dlm->recovery_map)) {
444 		mlog(0, "new master %u died while recovering %u!\n",
445 		     dlm->reco.new_master, dlm->reco.dead_node);
446 		/* unset the new_master, leave dead_node */
447 		dlm_set_reco_master(dlm, O2NM_INVALID_NODE_NUM);
448 	}
449 
450 	/* select a target to recover */
451 	if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
452 		int bit;
453 
454 		bit = find_first_bit(dlm->recovery_map, O2NM_MAX_NODES);
455 		if (bit >= O2NM_MAX_NODES || bit < 0)
456 			dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
457 		else
458 			dlm_set_reco_dead_node(dlm, bit);
459 	} else if (!test_bit(dlm->reco.dead_node, dlm->recovery_map)) {
460 		/* BUG? */
461 		mlog(ML_ERROR, "dead_node %u no longer in recovery map!\n",
462 		     dlm->reco.dead_node);
463 		dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
464 	}
465 
466 	if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
467 		// mlog(0, "nothing to recover!  sleeping now!\n");
468 		spin_unlock(&dlm->spinlock);
469 		/* return to main thread loop and sleep. */
470 		return 0;
471 	}
472 	mlog(0, "%s(%d):recovery thread found node %u in the recovery map!\n",
473 	     dlm->name, task_pid_nr(dlm->dlm_reco_thread_task),
474 	     dlm->reco.dead_node);
475 
476 	/* take write barrier */
477 	/* (stops the list reshuffling thread, proxy ast handling) */
478 	dlm_begin_recovery(dlm);
479 
480 	spin_unlock(&dlm->spinlock);
481 
482 	if (dlm->reco.new_master == dlm->node_num)
483 		goto master_here;
484 
485 	if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) {
486 		/* choose a new master, returns 0 if this node
487 		 * is the master, -EEXIST if it's another node.
488 		 * this does not return until a new master is chosen
489 		 * or recovery completes entirely. */
490 		ret = dlm_pick_recovery_master(dlm);
491 		if (!ret) {
492 			/* already notified everyone.  go. */
493 			goto master_here;
494 		}
495 		mlog(0, "another node will master this recovery session.\n");
496 	}
497 
498 	dlm_print_recovery_master(dlm);
499 
500 	/* it is safe to start everything back up here
501 	 * because all of the dead node's lock resources
502 	 * have been marked as in-recovery */
503 	dlm_end_recovery(dlm);
504 
505 	/* sleep out in main dlm_recovery_thread loop. */
506 	return 0;
507 
508 master_here:
509 	dlm_print_recovery_master(dlm);
510 
511 	status = dlm_remaster_locks(dlm, dlm->reco.dead_node);
512 	if (status < 0) {
513 		/* we should never hit this anymore */
514 		mlog(ML_ERROR, "%s: Error %d remastering locks for node %u, "
515 		     "retrying.\n", dlm->name, status, dlm->reco.dead_node);
516 		/* yield a bit to allow any final network messages
517 		 * to get handled on remaining nodes */
518 		msleep(100);
519 	} else {
520 		/* success!  see if any other nodes need recovery */
521 		mlog(0, "DONE mastering recovery of %s:%u here(this=%u)!\n",
522 		     dlm->name, dlm->reco.dead_node, dlm->node_num);
523 		spin_lock(&dlm->spinlock);
524 		__dlm_reset_recovery(dlm);
525 		dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE;
526 		spin_unlock(&dlm->spinlock);
527 	}
528 	dlm_end_recovery(dlm);
529 
530 	/* continue and look for another dead node */
531 	return -EAGAIN;
532 }
533 
dlm_remaster_locks(struct dlm_ctxt * dlm,u8 dead_node)534 static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node)
535 {
536 	int status = 0;
537 	struct dlm_reco_node_data *ndata;
538 	int all_nodes_done;
539 	int destroy = 0;
540 	int pass = 0;
541 
542 	do {
543 		/* we have become recovery master.  there is no escaping
544 		 * this, so just keep trying until we get it. */
545 		status = dlm_init_recovery_area(dlm, dead_node);
546 		if (status < 0) {
547 			mlog(ML_ERROR, "%s: failed to alloc recovery area, "
548 			     "retrying\n", dlm->name);
549 			msleep(1000);
550 		}
551 	} while (status != 0);
552 
553 	/* safe to access the node data list without a lock, since this
554 	 * process is the only one to change the list */
555 	list_for_each_entry(ndata, &dlm->reco.node_data, list) {
556 		BUG_ON(ndata->state != DLM_RECO_NODE_DATA_INIT);
557 		ndata->state = DLM_RECO_NODE_DATA_REQUESTING;
558 
559 		mlog(0, "%s: Requesting lock info from node %u\n", dlm->name,
560 		     ndata->node_num);
561 
562 		if (ndata->node_num == dlm->node_num) {
563 			ndata->state = DLM_RECO_NODE_DATA_DONE;
564 			continue;
565 		}
566 
567 		do {
568 			status = dlm_request_all_locks(dlm, ndata->node_num,
569 						       dead_node);
570 			if (status < 0) {
571 				mlog_errno(status);
572 				if (dlm_is_host_down(status)) {
573 					/* node died, ignore it for recovery */
574 					status = 0;
575 					ndata->state = DLM_RECO_NODE_DATA_DEAD;
576 					/* wait for the domain map to catch up
577 					 * with the network state. */
578 					wait_event_timeout(dlm->dlm_reco_thread_wq,
579 							   dlm_is_node_dead(dlm,
580 								ndata->node_num),
581 							   msecs_to_jiffies(1000));
582 					mlog(0, "waited 1 sec for %u, "
583 					     "dead? %s\n", ndata->node_num,
584 					     dlm_is_node_dead(dlm, ndata->node_num) ?
585 					     "yes" : "no");
586 				} else {
587 					/* -ENOMEM on the other node */
588 					mlog(0, "%s: node %u returned "
589 					     "%d during recovery, retrying "
590 					     "after a short wait\n",
591 					     dlm->name, ndata->node_num,
592 					     status);
593 					msleep(100);
594 				}
595 			}
596 		} while (status != 0);
597 
598 		spin_lock(&dlm_reco_state_lock);
599 		switch (ndata->state) {
600 			case DLM_RECO_NODE_DATA_INIT:
601 			case DLM_RECO_NODE_DATA_FINALIZE_SENT:
602 			case DLM_RECO_NODE_DATA_REQUESTED:
603 				BUG();
604 				break;
605 			case DLM_RECO_NODE_DATA_DEAD:
606 				mlog(0, "node %u died after requesting "
607 				     "recovery info for node %u\n",
608 				     ndata->node_num, dead_node);
609 				/* fine.  don't need this node's info.
610 				 * continue without it. */
611 				break;
612 			case DLM_RECO_NODE_DATA_REQUESTING:
613 				ndata->state = DLM_RECO_NODE_DATA_REQUESTED;
614 				mlog(0, "now receiving recovery data from "
615 				     "node %u for dead node %u\n",
616 				     ndata->node_num, dead_node);
617 				break;
618 			case DLM_RECO_NODE_DATA_RECEIVING:
619 				mlog(0, "already receiving recovery data from "
620 				     "node %u for dead node %u\n",
621 				     ndata->node_num, dead_node);
622 				break;
623 			case DLM_RECO_NODE_DATA_DONE:
624 				mlog(0, "already DONE receiving recovery data "
625 				     "from node %u for dead node %u\n",
626 				     ndata->node_num, dead_node);
627 				break;
628 		}
629 		spin_unlock(&dlm_reco_state_lock);
630 	}
631 
632 	mlog(0, "%s: Done requesting all lock info\n", dlm->name);
633 
634 	/* nodes should be sending reco data now
635 	 * just need to wait */
636 
637 	while (1) {
638 		/* check all the nodes now to see if we are
639 		 * done, or if anyone died */
640 		all_nodes_done = 1;
641 		spin_lock(&dlm_reco_state_lock);
642 		list_for_each_entry(ndata, &dlm->reco.node_data, list) {
643 			mlog(0, "checking recovery state of node %u\n",
644 			     ndata->node_num);
645 			switch (ndata->state) {
646 				case DLM_RECO_NODE_DATA_INIT:
647 				case DLM_RECO_NODE_DATA_REQUESTING:
648 					mlog(ML_ERROR, "bad ndata state for "
649 					     "node %u: state=%d\n",
650 					     ndata->node_num, ndata->state);
651 					BUG();
652 					break;
653 				case DLM_RECO_NODE_DATA_DEAD:
654 					mlog(0, "node %u died after "
655 					     "requesting recovery info for "
656 					     "node %u\n", ndata->node_num,
657 					     dead_node);
658 					break;
659 				case DLM_RECO_NODE_DATA_RECEIVING:
660 				case DLM_RECO_NODE_DATA_REQUESTED:
661 					mlog(0, "%s: node %u still in state %s\n",
662 					     dlm->name, ndata->node_num,
663 					     ndata->state==DLM_RECO_NODE_DATA_RECEIVING ?
664 					     "receiving" : "requested");
665 					all_nodes_done = 0;
666 					break;
667 				case DLM_RECO_NODE_DATA_DONE:
668 					mlog(0, "%s: node %u state is done\n",
669 					     dlm->name, ndata->node_num);
670 					break;
671 				case DLM_RECO_NODE_DATA_FINALIZE_SENT:
672 					mlog(0, "%s: node %u state is finalize\n",
673 					     dlm->name, ndata->node_num);
674 					break;
675 			}
676 		}
677 		spin_unlock(&dlm_reco_state_lock);
678 
679 		mlog(0, "pass #%d, all_nodes_done?: %s\n", ++pass,
680 		     all_nodes_done?"yes":"no");
681 		if (all_nodes_done) {
682 			int ret;
683 
684 			/* Set this flag on recovery master to avoid
685 			 * a new recovery for another dead node start
686 			 * before the recovery is not done. That may
687 			 * cause recovery hung.*/
688 			spin_lock(&dlm->spinlock);
689 			dlm->reco.state |= DLM_RECO_STATE_FINALIZE;
690 			spin_unlock(&dlm->spinlock);
691 
692 			/* all nodes are now in DLM_RECO_NODE_DATA_DONE state
693 	 		 * just send a finalize message to everyone and
694 	 		 * clean up */
695 			mlog(0, "all nodes are done! send finalize\n");
696 			ret = dlm_send_finalize_reco_message(dlm);
697 			if (ret < 0)
698 				mlog_errno(ret);
699 
700 			spin_lock(&dlm->spinlock);
701 			dlm_finish_local_lockres_recovery(dlm, dead_node,
702 							  dlm->node_num);
703 			spin_unlock(&dlm->spinlock);
704 			mlog(0, "should be done with recovery!\n");
705 
706 			mlog(0, "finishing recovery of %s at %lu, "
707 			     "dead=%u, this=%u, new=%u\n", dlm->name,
708 			     jiffies, dlm->reco.dead_node,
709 			     dlm->node_num, dlm->reco.new_master);
710 			destroy = 1;
711 			status = 0;
712 			/* rescan everything marked dirty along the way */
713 			dlm_kick_thread(dlm, NULL);
714 			break;
715 		}
716 		/* wait to be signalled, with periodic timeout
717 		 * to check for node death */
718 		wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq,
719 					 kthread_should_stop(),
720 					 msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS));
721 
722 	}
723 
724 	if (destroy)
725 		dlm_destroy_recovery_area(dlm);
726 
727 	return status;
728 }
729 
dlm_init_recovery_area(struct dlm_ctxt * dlm,u8 dead_node)730 static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node)
731 {
732 	int num=0;
733 	struct dlm_reco_node_data *ndata;
734 
735 	spin_lock(&dlm->spinlock);
736 	bitmap_copy(dlm->reco.node_map, dlm->domain_map, O2NM_MAX_NODES);
737 	/* nodes can only be removed (by dying) after dropping
738 	 * this lock, and death will be trapped later, so this should do */
739 	spin_unlock(&dlm->spinlock);
740 
741 	while (1) {
742 		num = find_next_bit (dlm->reco.node_map, O2NM_MAX_NODES, num);
743 		if (num >= O2NM_MAX_NODES) {
744 			break;
745 		}
746 		BUG_ON(num == dead_node);
747 
748 		ndata = kzalloc(sizeof(*ndata), GFP_NOFS);
749 		if (!ndata) {
750 			dlm_destroy_recovery_area(dlm);
751 			return -ENOMEM;
752 		}
753 		ndata->node_num = num;
754 		ndata->state = DLM_RECO_NODE_DATA_INIT;
755 		spin_lock(&dlm_reco_state_lock);
756 		list_add_tail(&ndata->list, &dlm->reco.node_data);
757 		spin_unlock(&dlm_reco_state_lock);
758 		num++;
759 	}
760 
761 	return 0;
762 }
763 
dlm_destroy_recovery_area(struct dlm_ctxt * dlm)764 static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm)
765 {
766 	struct dlm_reco_node_data *ndata, *next;
767 	LIST_HEAD(tmplist);
768 
769 	spin_lock(&dlm_reco_state_lock);
770 	list_splice_init(&dlm->reco.node_data, &tmplist);
771 	spin_unlock(&dlm_reco_state_lock);
772 
773 	list_for_each_entry_safe(ndata, next, &tmplist, list) {
774 		list_del_init(&ndata->list);
775 		kfree(ndata);
776 	}
777 }
778 
dlm_request_all_locks(struct dlm_ctxt * dlm,u8 request_from,u8 dead_node)779 static int dlm_request_all_locks(struct dlm_ctxt *dlm, u8 request_from,
780 				 u8 dead_node)
781 {
782 	struct dlm_lock_request lr;
783 	int ret;
784 	int status;
785 
786 	mlog(0, "\n");
787 
788 
789 	mlog(0, "dlm_request_all_locks: dead node is %u, sending request "
790 		  "to %u\n", dead_node, request_from);
791 
792 	memset(&lr, 0, sizeof(lr));
793 	lr.node_idx = dlm->node_num;
794 	lr.dead_node = dead_node;
795 
796 	// send message
797 	ret = o2net_send_message(DLM_LOCK_REQUEST_MSG, dlm->key,
798 				 &lr, sizeof(lr), request_from, &status);
799 
800 	/* negative status is handled by caller */
801 	if (ret < 0)
802 		mlog(ML_ERROR, "%s: Error %d send LOCK_REQUEST to node %u "
803 		     "to recover dead node %u\n", dlm->name, ret,
804 		     request_from, dead_node);
805 	else
806 		ret = status;
807 	// return from here, then
808 	// sleep until all received or error
809 	return ret;
810 
811 }
812 
dlm_request_all_locks_handler(struct o2net_msg * msg,u32 len,void * data,void ** ret_data)813 int dlm_request_all_locks_handler(struct o2net_msg *msg, u32 len, void *data,
814 				  void **ret_data)
815 {
816 	struct dlm_ctxt *dlm = data;
817 	struct dlm_lock_request *lr = (struct dlm_lock_request *)msg->buf;
818 	char *buf = NULL;
819 	struct dlm_work_item *item = NULL;
820 
821 	if (!dlm_grab(dlm))
822 		return -EINVAL;
823 
824 	if (lr->dead_node != dlm->reco.dead_node) {
825 		mlog(ML_ERROR, "%s: node %u sent dead_node=%u, but local "
826 		     "dead_node is %u\n", dlm->name, lr->node_idx,
827 		     lr->dead_node, dlm->reco.dead_node);
828 		dlm_print_reco_node_status(dlm);
829 		/* this is a hack */
830 		dlm_put(dlm);
831 		return -ENOMEM;
832 	}
833 	BUG_ON(lr->dead_node != dlm->reco.dead_node);
834 
835 	item = kzalloc(sizeof(*item), GFP_NOFS);
836 	if (!item) {
837 		dlm_put(dlm);
838 		return -ENOMEM;
839 	}
840 
841 	/* this will get freed by dlm_request_all_locks_worker */
842 	buf = (char *) __get_free_page(GFP_NOFS);
843 	if (!buf) {
844 		kfree(item);
845 		dlm_put(dlm);
846 		return -ENOMEM;
847 	}
848 
849 	/* queue up work for dlm_request_all_locks_worker */
850 	dlm_grab(dlm);  /* get an extra ref for the work item */
851 	dlm_init_work_item(dlm, item, dlm_request_all_locks_worker, buf);
852 	item->u.ral.reco_master = lr->node_idx;
853 	item->u.ral.dead_node = lr->dead_node;
854 	spin_lock(&dlm->work_lock);
855 	list_add_tail(&item->list, &dlm->work_list);
856 	spin_unlock(&dlm->work_lock);
857 	queue_work(dlm->dlm_worker, &dlm->dispatched_work);
858 
859 	dlm_put(dlm);
860 	return 0;
861 }
862 
dlm_request_all_locks_worker(struct dlm_work_item * item,void * data)863 static void dlm_request_all_locks_worker(struct dlm_work_item *item, void *data)
864 {
865 	struct dlm_migratable_lockres *mres;
866 	struct dlm_lock_resource *res;
867 	struct dlm_ctxt *dlm;
868 	LIST_HEAD(resources);
869 	int ret;
870 	u8 dead_node, reco_master;
871 	int skip_all_done = 0;
872 
873 	dlm = item->dlm;
874 	dead_node = item->u.ral.dead_node;
875 	reco_master = item->u.ral.reco_master;
876 	mres = (struct dlm_migratable_lockres *)data;
877 
878 	mlog(0, "%s: recovery worker started, dead=%u, master=%u\n",
879 	     dlm->name, dead_node, reco_master);
880 
881 	if (dead_node != dlm->reco.dead_node ||
882 	    reco_master != dlm->reco.new_master) {
883 		/* worker could have been created before the recovery master
884 		 * died.  if so, do not continue, but do not error. */
885 		if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) {
886 			mlog(ML_NOTICE, "%s: will not send recovery state, "
887 			     "recovery master %u died, thread=(dead=%u,mas=%u)"
888 			     " current=(dead=%u,mas=%u)\n", dlm->name,
889 			     reco_master, dead_node, reco_master,
890 			     dlm->reco.dead_node, dlm->reco.new_master);
891 		} else {
892 			mlog(ML_NOTICE, "%s: reco state invalid: reco(dead=%u, "
893 			     "master=%u), request(dead=%u, master=%u)\n",
894 			     dlm->name, dlm->reco.dead_node,
895 			     dlm->reco.new_master, dead_node, reco_master);
896 		}
897 		goto leave;
898 	}
899 
900 	/* lock resources should have already been moved to the
901  	 * dlm->reco.resources list.  now move items from that list
902  	 * to a temp list if the dead owner matches.  note that the
903 	 * whole cluster recovers only one node at a time, so we
904 	 * can safely move UNKNOWN lock resources for each recovery
905 	 * session. */
906 	dlm_move_reco_locks_to_list(dlm, &resources, dead_node);
907 
908 	/* now we can begin blasting lockreses without the dlm lock */
909 
910 	/* any errors returned will be due to the new_master dying,
911 	 * the dlm_reco_thread should detect this */
912 	list_for_each_entry(res, &resources, recovering) {
913 		ret = dlm_send_one_lockres(dlm, res, mres, reco_master,
914 				   	DLM_MRES_RECOVERY);
915 		if (ret < 0) {
916 			mlog(ML_ERROR, "%s: node %u went down while sending "
917 			     "recovery state for dead node %u, ret=%d\n", dlm->name,
918 			     reco_master, dead_node, ret);
919 			skip_all_done = 1;
920 			break;
921 		}
922 	}
923 
924 	/* move the resources back to the list */
925 	spin_lock(&dlm->spinlock);
926 	list_splice_init(&resources, &dlm->reco.resources);
927 	spin_unlock(&dlm->spinlock);
928 
929 	if (!skip_all_done) {
930 		ret = dlm_send_all_done_msg(dlm, dead_node, reco_master);
931 		if (ret < 0) {
932 			mlog(ML_ERROR, "%s: node %u went down while sending "
933 			     "recovery all-done for dead node %u, ret=%d\n",
934 			     dlm->name, reco_master, dead_node, ret);
935 		}
936 	}
937 leave:
938 	free_page((unsigned long)data);
939 }
940 
941 
dlm_send_all_done_msg(struct dlm_ctxt * dlm,u8 dead_node,u8 send_to)942 static int dlm_send_all_done_msg(struct dlm_ctxt *dlm, u8 dead_node, u8 send_to)
943 {
944 	int ret, tmpret;
945 	struct dlm_reco_data_done done_msg;
946 
947 	memset(&done_msg, 0, sizeof(done_msg));
948 	done_msg.node_idx = dlm->node_num;
949 	done_msg.dead_node = dead_node;
950 	mlog(0, "sending DATA DONE message to %u, "
951 	     "my node=%u, dead node=%u\n", send_to, done_msg.node_idx,
952 	     done_msg.dead_node);
953 
954 	ret = o2net_send_message(DLM_RECO_DATA_DONE_MSG, dlm->key, &done_msg,
955 				 sizeof(done_msg), send_to, &tmpret);
956 	if (ret < 0) {
957 		mlog(ML_ERROR, "%s: Error %d send RECO_DATA_DONE to node %u "
958 		     "to recover dead node %u\n", dlm->name, ret, send_to,
959 		     dead_node);
960 		if (!dlm_is_host_down(ret)) {
961 			BUG();
962 		}
963 	} else
964 		ret = tmpret;
965 	return ret;
966 }
967 
968 
dlm_reco_data_done_handler(struct o2net_msg * msg,u32 len,void * data,void ** ret_data)969 int dlm_reco_data_done_handler(struct o2net_msg *msg, u32 len, void *data,
970 			       void **ret_data)
971 {
972 	struct dlm_ctxt *dlm = data;
973 	struct dlm_reco_data_done *done = (struct dlm_reco_data_done *)msg->buf;
974 	struct dlm_reco_node_data *ndata = NULL;
975 	int ret = -EINVAL;
976 
977 	if (!dlm_grab(dlm))
978 		return -EINVAL;
979 
980 	mlog(0, "got DATA DONE: dead_node=%u, reco.dead_node=%u, "
981 	     "node_idx=%u, this node=%u\n", done->dead_node,
982 	     dlm->reco.dead_node, done->node_idx, dlm->node_num);
983 
984 	mlog_bug_on_msg((done->dead_node != dlm->reco.dead_node),
985 			"Got DATA DONE: dead_node=%u, reco.dead_node=%u, "
986 			"node_idx=%u, this node=%u\n", done->dead_node,
987 			dlm->reco.dead_node, done->node_idx, dlm->node_num);
988 
989 	spin_lock(&dlm_reco_state_lock);
990 	list_for_each_entry(ndata, &dlm->reco.node_data, list) {
991 		if (ndata->node_num != done->node_idx)
992 			continue;
993 
994 		switch (ndata->state) {
995 			/* should have moved beyond INIT but not to FINALIZE yet */
996 			case DLM_RECO_NODE_DATA_INIT:
997 			case DLM_RECO_NODE_DATA_DEAD:
998 			case DLM_RECO_NODE_DATA_FINALIZE_SENT:
999 				mlog(ML_ERROR, "bad ndata state for node %u:"
1000 				     " state=%d\n", ndata->node_num,
1001 				     ndata->state);
1002 				BUG();
1003 				break;
1004 			/* these states are possible at this point, anywhere along
1005 			 * the line of recovery */
1006 			case DLM_RECO_NODE_DATA_DONE:
1007 			case DLM_RECO_NODE_DATA_RECEIVING:
1008 			case DLM_RECO_NODE_DATA_REQUESTED:
1009 			case DLM_RECO_NODE_DATA_REQUESTING:
1010 				mlog(0, "node %u is DONE sending "
1011 					  "recovery data!\n",
1012 					  ndata->node_num);
1013 
1014 				ndata->state = DLM_RECO_NODE_DATA_DONE;
1015 				ret = 0;
1016 				break;
1017 		}
1018 	}
1019 	spin_unlock(&dlm_reco_state_lock);
1020 
1021 	/* wake the recovery thread, some node is done */
1022 	if (!ret)
1023 		dlm_kick_recovery_thread(dlm);
1024 
1025 	if (ret < 0)
1026 		mlog(ML_ERROR, "failed to find recovery node data for node "
1027 		     "%u\n", done->node_idx);
1028 	dlm_put(dlm);
1029 
1030 	mlog(0, "leaving reco data done handler, ret=%d\n", ret);
1031 	return ret;
1032 }
1033 
dlm_move_reco_locks_to_list(struct dlm_ctxt * dlm,struct list_head * list,u8 dead_node)1034 static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm,
1035 					struct list_head *list,
1036 				       	u8 dead_node)
1037 {
1038 	struct dlm_lock_resource *res, *next;
1039 	struct dlm_lock *lock;
1040 
1041 	spin_lock(&dlm->spinlock);
1042 	list_for_each_entry_safe(res, next, &dlm->reco.resources, recovering) {
1043 		/* always prune any $RECOVERY entries for dead nodes,
1044 		 * otherwise hangs can occur during later recovery */
1045 		if (dlm_is_recovery_lock(res->lockname.name,
1046 					 res->lockname.len)) {
1047 			spin_lock(&res->spinlock);
1048 			list_for_each_entry(lock, &res->granted, list) {
1049 				if (lock->ml.node == dead_node) {
1050 					mlog(0, "AHA! there was "
1051 					     "a $RECOVERY lock for dead "
1052 					     "node %u (%s)!\n",
1053 					     dead_node, dlm->name);
1054 					list_del_init(&lock->list);
1055 					dlm_lock_put(lock);
1056 					/* Can't schedule DLM_UNLOCK_FREE_LOCK
1057 					 * - do manually */
1058 					dlm_lock_put(lock);
1059 					break;
1060 				}
1061 			}
1062 			spin_unlock(&res->spinlock);
1063 			continue;
1064 		}
1065 
1066 		if (res->owner == dead_node) {
1067 			mlog(0, "found lockres owned by dead node while "
1068 				  "doing recovery for node %u. sending it.\n",
1069 				  dead_node);
1070 			list_move_tail(&res->recovering, list);
1071 		} else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
1072 			mlog(0, "found UNKNOWN owner while doing recovery "
1073 				  "for node %u. sending it.\n", dead_node);
1074 			list_move_tail(&res->recovering, list);
1075 		}
1076 	}
1077 	spin_unlock(&dlm->spinlock);
1078 }
1079 
dlm_num_locks_in_lockres(struct dlm_lock_resource * res)1080 static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res)
1081 {
1082 	int total_locks = 0;
1083 	struct list_head *iter, *queue = &res->granted;
1084 	int i;
1085 
1086 	for (i=0; i<3; i++) {
1087 		list_for_each(iter, queue)
1088 			total_locks++;
1089 		queue++;
1090 	}
1091 	return total_locks;
1092 }
1093 
1094 
dlm_send_mig_lockres_msg(struct dlm_ctxt * dlm,struct dlm_migratable_lockres * mres,u8 send_to,struct dlm_lock_resource * res,int total_locks)1095 static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm,
1096 				      struct dlm_migratable_lockres *mres,
1097 				      u8 send_to,
1098 				      struct dlm_lock_resource *res,
1099 				      int total_locks)
1100 {
1101 	u64 mig_cookie = be64_to_cpu(mres->mig_cookie);
1102 	int mres_total_locks = be32_to_cpu(mres->total_locks);
1103 	int ret = 0, status = 0;
1104 	u8 orig_flags = mres->flags,
1105 	   orig_master = mres->master;
1106 
1107 	BUG_ON(mres->num_locks > DLM_MAX_MIGRATABLE_LOCKS);
1108 	if (!mres->num_locks)
1109 		return 0;
1110 
1111 	/* add an all-done flag if we reached the last lock */
1112 	orig_flags = mres->flags;
1113 	BUG_ON(total_locks > mres_total_locks);
1114 	if (total_locks == mres_total_locks)
1115 		mres->flags |= DLM_MRES_ALL_DONE;
1116 
1117 	mlog(0, "%s:%.*s: sending mig lockres (%s) to %u\n",
1118 	     dlm->name, res->lockname.len, res->lockname.name,
1119 	     orig_flags & DLM_MRES_MIGRATION ? "migration" : "recovery",
1120 	     send_to);
1121 
1122 	/* send it */
1123 	ret = o2net_send_message(DLM_MIG_LOCKRES_MSG, dlm->key, mres,
1124 				 struct_size(mres, ml, mres->num_locks),
1125 				 send_to, &status);
1126 	if (ret < 0) {
1127 		/* XXX: negative status is not handled.
1128 		 * this will end up killing this node. */
1129 		mlog(ML_ERROR, "%s: res %.*s, Error %d send MIG_LOCKRES to "
1130 		     "node %u (%s)\n", dlm->name, mres->lockname_len,
1131 		     mres->lockname, ret, send_to,
1132 		     (orig_flags & DLM_MRES_MIGRATION ?
1133 		      "migration" : "recovery"));
1134 	} else {
1135 		/* might get an -ENOMEM back here */
1136 		ret = status;
1137 		if (ret < 0) {
1138 			mlog_errno(ret);
1139 
1140 			if (ret == -EFAULT) {
1141 				mlog(ML_ERROR, "node %u told me to kill "
1142 				     "myself!\n", send_to);
1143 				BUG();
1144 			}
1145 		}
1146 	}
1147 
1148 	/* zero and reinit the message buffer */
1149 	dlm_init_migratable_lockres(mres, res->lockname.name,
1150 				    res->lockname.len, mres_total_locks,
1151 				    mig_cookie, orig_flags, orig_master);
1152 	return ret;
1153 }
1154 
dlm_init_migratable_lockres(struct dlm_migratable_lockres * mres,const char * lockname,int namelen,int total_locks,u64 cookie,u8 flags,u8 master)1155 static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres,
1156 					const char *lockname, int namelen,
1157 					int total_locks, u64 cookie,
1158 					u8 flags, u8 master)
1159 {
1160 	/* mres here is one full page */
1161 	clear_page(mres);
1162 	mres->lockname_len = namelen;
1163 	memcpy(mres->lockname, lockname, namelen);
1164 	mres->num_locks = 0;
1165 	mres->total_locks = cpu_to_be32(total_locks);
1166 	mres->mig_cookie = cpu_to_be64(cookie);
1167 	mres->flags = flags;
1168 	mres->master = master;
1169 }
1170 
dlm_prepare_lvb_for_migration(struct dlm_lock * lock,struct dlm_migratable_lockres * mres,int queue)1171 static void dlm_prepare_lvb_for_migration(struct dlm_lock *lock,
1172 					  struct dlm_migratable_lockres *mres,
1173 					  int queue)
1174 {
1175 	if (!lock->lksb)
1176 	       return;
1177 
1178 	/* Ignore lvb in all locks in the blocked list */
1179 	if (queue == DLM_BLOCKED_LIST)
1180 		return;
1181 
1182 	/* Only consider lvbs in locks with granted EX or PR lock levels */
1183 	if (lock->ml.type != LKM_EXMODE && lock->ml.type != LKM_PRMODE)
1184 		return;
1185 
1186 	if (dlm_lvb_is_empty(mres->lvb)) {
1187 		memcpy(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN);
1188 		return;
1189 	}
1190 
1191 	/* Ensure the lvb copied for migration matches in other valid locks */
1192 	if (!memcmp(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN))
1193 		return;
1194 
1195 	mlog(ML_ERROR, "Mismatched lvb in lock cookie=%u:%llu, name=%.*s, "
1196 	     "node=%u\n",
1197 	     dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)),
1198 	     dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)),
1199 	     lock->lockres->lockname.len, lock->lockres->lockname.name,
1200 	     lock->ml.node);
1201 	dlm_print_one_lock_resource(lock->lockres);
1202 	BUG();
1203 }
1204 
1205 /* returns 1 if this lock fills the network structure,
1206  * 0 otherwise */
dlm_add_lock_to_array(struct dlm_lock * lock,struct dlm_migratable_lockres * mres,int queue)1207 static int dlm_add_lock_to_array(struct dlm_lock *lock,
1208 				 struct dlm_migratable_lockres *mres, int queue)
1209 {
1210 	struct dlm_migratable_lock *ml;
1211 	int lock_num = mres->num_locks;
1212 
1213 	ml = &(mres->ml[lock_num]);
1214 	ml->cookie = lock->ml.cookie;
1215 	ml->type = lock->ml.type;
1216 	ml->convert_type = lock->ml.convert_type;
1217 	ml->highest_blocked = lock->ml.highest_blocked;
1218 	ml->list = queue;
1219 	if (lock->lksb) {
1220 		ml->flags = lock->lksb->flags;
1221 		dlm_prepare_lvb_for_migration(lock, mres, queue);
1222 	}
1223 	ml->node = lock->ml.node;
1224 	mres->num_locks++;
1225 	/* we reached the max, send this network message */
1226 	if (mres->num_locks == DLM_MAX_MIGRATABLE_LOCKS)
1227 		return 1;
1228 	return 0;
1229 }
1230 
dlm_add_dummy_lock(struct dlm_ctxt * dlm,struct dlm_migratable_lockres * mres)1231 static void dlm_add_dummy_lock(struct dlm_ctxt *dlm,
1232 			       struct dlm_migratable_lockres *mres)
1233 {
1234 	struct dlm_lock dummy;
1235 	memset(&dummy, 0, sizeof(dummy));
1236 	dummy.ml.cookie = 0;
1237 	dummy.ml.type = LKM_IVMODE;
1238 	dummy.ml.convert_type = LKM_IVMODE;
1239 	dummy.ml.highest_blocked = LKM_IVMODE;
1240 	dummy.lksb = NULL;
1241 	dummy.ml.node = dlm->node_num;
1242 	dlm_add_lock_to_array(&dummy, mres, DLM_BLOCKED_LIST);
1243 }
1244 
dlm_is_dummy_lock(struct dlm_ctxt * dlm,struct dlm_migratable_lock * ml,u8 * nodenum)1245 static inline int dlm_is_dummy_lock(struct dlm_ctxt *dlm,
1246 				    struct dlm_migratable_lock *ml,
1247 				    u8 *nodenum)
1248 {
1249 	if (unlikely(ml->cookie == 0 &&
1250 	    ml->type == LKM_IVMODE &&
1251 	    ml->convert_type == LKM_IVMODE &&
1252 	    ml->highest_blocked == LKM_IVMODE &&
1253 	    ml->list == DLM_BLOCKED_LIST)) {
1254 		*nodenum = ml->node;
1255 		return 1;
1256 	}
1257 	return 0;
1258 }
1259 
dlm_send_one_lockres(struct dlm_ctxt * dlm,struct dlm_lock_resource * res,struct dlm_migratable_lockres * mres,u8 send_to,u8 flags)1260 int dlm_send_one_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
1261 			 struct dlm_migratable_lockres *mres,
1262 			 u8 send_to, u8 flags)
1263 {
1264 	struct list_head *queue;
1265 	int total_locks, i;
1266 	u64 mig_cookie = 0;
1267 	struct dlm_lock *lock;
1268 	int ret = 0;
1269 
1270 	BUG_ON(!(flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION)));
1271 
1272 	mlog(0, "sending to %u\n", send_to);
1273 
1274 	total_locks = dlm_num_locks_in_lockres(res);
1275 	if (total_locks > DLM_MAX_MIGRATABLE_LOCKS) {
1276 		/* rare, but possible */
1277 		mlog(0, "argh.  lockres has %d locks.  this will "
1278 			  "require more than one network packet to "
1279 			  "migrate\n", total_locks);
1280 		mig_cookie = dlm_get_next_mig_cookie();
1281 	}
1282 
1283 	dlm_init_migratable_lockres(mres, res->lockname.name,
1284 				    res->lockname.len, total_locks,
1285 				    mig_cookie, flags, res->owner);
1286 
1287 	total_locks = 0;
1288 	for (i=DLM_GRANTED_LIST; i<=DLM_BLOCKED_LIST; i++) {
1289 		queue = dlm_list_idx_to_ptr(res, i);
1290 		list_for_each_entry(lock, queue, list) {
1291 			/* add another lock. */
1292 			total_locks++;
1293 			if (!dlm_add_lock_to_array(lock, mres, i))
1294 				continue;
1295 
1296 			/* this filled the lock message,
1297 			 * we must send it immediately. */
1298 			ret = dlm_send_mig_lockres_msg(dlm, mres, send_to,
1299 						       res, total_locks);
1300 			if (ret < 0)
1301 				goto error;
1302 		}
1303 	}
1304 	if (total_locks == 0) {
1305 		/* send a dummy lock to indicate a mastery reference only */
1306 		mlog(0, "%s:%.*s: sending dummy lock to %u, %s\n",
1307 		     dlm->name, res->lockname.len, res->lockname.name,
1308 		     send_to, flags & DLM_MRES_RECOVERY ? "recovery" :
1309 		     "migration");
1310 		dlm_add_dummy_lock(dlm, mres);
1311 	}
1312 	/* flush any remaining locks */
1313 	ret = dlm_send_mig_lockres_msg(dlm, mres, send_to, res, total_locks);
1314 	if (ret < 0)
1315 		goto error;
1316 	return ret;
1317 
1318 error:
1319 	mlog(ML_ERROR, "%s: dlm_send_mig_lockres_msg returned %d\n",
1320 	     dlm->name, ret);
1321 	if (!dlm_is_host_down(ret))
1322 		BUG();
1323 	mlog(0, "%s: node %u went down while sending %s "
1324 	     "lockres %.*s\n", dlm->name, send_to,
1325 	     flags & DLM_MRES_RECOVERY ?  "recovery" : "migration",
1326 	     res->lockname.len, res->lockname.name);
1327 	return ret;
1328 }
1329 
1330 
1331 
1332 /*
1333  * this message will contain no more than one page worth of
1334  * recovery data, and it will work on only one lockres.
1335  * there may be many locks in this page, and we may need to wait
1336  * for additional packets to complete all the locks (rare, but
1337  * possible).
1338  */
1339 /*
1340  * NOTE: the allocation error cases here are scary
1341  * we really cannot afford to fail an alloc in recovery
1342  * do we spin?  returning an error only delays the problem really
1343  */
1344 
dlm_mig_lockres_handler(struct o2net_msg * msg,u32 len,void * data,void ** ret_data)1345 int dlm_mig_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
1346 			    void **ret_data)
1347 {
1348 	struct dlm_ctxt *dlm = data;
1349 	struct dlm_migratable_lockres *mres =
1350 		(struct dlm_migratable_lockres *)msg->buf;
1351 	int ret = 0;
1352 	u8 real_master;
1353 	u8 extra_refs = 0;
1354 	char *buf = NULL;
1355 	struct dlm_work_item *item = NULL;
1356 	struct dlm_lock_resource *res = NULL;
1357 	unsigned int hash;
1358 
1359 	if (!dlm_grab(dlm))
1360 		return -EINVAL;
1361 
1362 	if (!dlm_joined(dlm)) {
1363 		mlog(ML_ERROR, "Domain %s not joined! "
1364 			  "lockres %.*s, master %u\n",
1365 			  dlm->name, mres->lockname_len,
1366 			  mres->lockname, mres->master);
1367 		dlm_put(dlm);
1368 		return -EINVAL;
1369 	}
1370 
1371 	BUG_ON(!(mres->flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION)));
1372 
1373 	real_master = mres->master;
1374 	if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1375 		/* cannot migrate a lockres with no master */
1376 		BUG_ON(!(mres->flags & DLM_MRES_RECOVERY));
1377 	}
1378 
1379 	mlog(0, "%s message received from node %u\n",
1380 		  (mres->flags & DLM_MRES_RECOVERY) ?
1381 		  "recovery" : "migration", mres->master);
1382 	if (mres->flags & DLM_MRES_ALL_DONE)
1383 		mlog(0, "all done flag.  all lockres data received!\n");
1384 
1385 	ret = -ENOMEM;
1386 	buf = kmalloc(be16_to_cpu(msg->data_len), GFP_NOFS);
1387 	item = kzalloc(sizeof(*item), GFP_NOFS);
1388 	if (!buf || !item)
1389 		goto leave;
1390 
1391 	/* lookup the lock to see if we have a secondary queue for this
1392 	 * already...  just add the locks in and this will have its owner
1393 	 * and RECOVERY flag changed when it completes. */
1394 	hash = dlm_lockid_hash(mres->lockname, mres->lockname_len);
1395 	spin_lock(&dlm->spinlock);
1396 	res = __dlm_lookup_lockres_full(dlm, mres->lockname, mres->lockname_len,
1397 			hash);
1398 	if (res) {
1399 	 	/* this will get a ref on res */
1400 		/* mark it as recovering/migrating and hash it */
1401 		spin_lock(&res->spinlock);
1402 		if (res->state & DLM_LOCK_RES_DROPPING_REF) {
1403 			mlog(0, "%s: node is attempting to migrate "
1404 				"lockres %.*s, but marked as dropping "
1405 				" ref!\n", dlm->name,
1406 				mres->lockname_len, mres->lockname);
1407 			ret = -EINVAL;
1408 			spin_unlock(&res->spinlock);
1409 			spin_unlock(&dlm->spinlock);
1410 			dlm_lockres_put(res);
1411 			goto leave;
1412 		}
1413 
1414 		if (mres->flags & DLM_MRES_RECOVERY) {
1415 			res->state |= DLM_LOCK_RES_RECOVERING;
1416 		} else {
1417 			if (res->state & DLM_LOCK_RES_MIGRATING) {
1418 				/* this is at least the second
1419 				 * lockres message */
1420 				mlog(0, "lock %.*s is already migrating\n",
1421 					  mres->lockname_len,
1422 					  mres->lockname);
1423 			} else if (res->state & DLM_LOCK_RES_RECOVERING) {
1424 				/* caller should BUG */
1425 				mlog(ML_ERROR, "node is attempting to migrate "
1426 				     "lock %.*s, but marked as recovering!\n",
1427 				     mres->lockname_len, mres->lockname);
1428 				ret = -EFAULT;
1429 				spin_unlock(&res->spinlock);
1430 				spin_unlock(&dlm->spinlock);
1431 				dlm_lockres_put(res);
1432 				goto leave;
1433 			}
1434 			res->state |= DLM_LOCK_RES_MIGRATING;
1435 		}
1436 		spin_unlock(&res->spinlock);
1437 		spin_unlock(&dlm->spinlock);
1438 	} else {
1439 		spin_unlock(&dlm->spinlock);
1440 		/* need to allocate, just like if it was
1441 		 * mastered here normally  */
1442 		res = dlm_new_lockres(dlm, mres->lockname, mres->lockname_len);
1443 		if (!res)
1444 			goto leave;
1445 
1446 		/* to match the ref that we would have gotten if
1447 		 * dlm_lookup_lockres had succeeded */
1448 		dlm_lockres_get(res);
1449 
1450 		/* mark it as recovering/migrating and hash it */
1451 		if (mres->flags & DLM_MRES_RECOVERY)
1452 			res->state |= DLM_LOCK_RES_RECOVERING;
1453 		else
1454 			res->state |= DLM_LOCK_RES_MIGRATING;
1455 
1456 		spin_lock(&dlm->spinlock);
1457 		__dlm_insert_lockres(dlm, res);
1458 		spin_unlock(&dlm->spinlock);
1459 
1460 		/* Add an extra ref for this lock-less lockres lest the
1461 		 * dlm_thread purges it before we get the chance to add
1462 		 * locks to it */
1463 		dlm_lockres_get(res);
1464 
1465 		/* There are three refs that need to be put.
1466 		 * 1. Taken above.
1467 		 * 2. kref_init in dlm_new_lockres()->dlm_init_lockres().
1468 		 * 3. dlm_lookup_lockres()
1469 		 * The first one is handled at the end of this function. The
1470 		 * other two are handled in the worker thread after locks have
1471 		 * been attached. Yes, we don't wait for purge time to match
1472 		 * kref_init. The lockres will still have atleast one ref
1473 		 * added because it is in the hash __dlm_insert_lockres() */
1474 		extra_refs++;
1475 
1476 		/* now that the new lockres is inserted,
1477 		 * make it usable by other processes */
1478 		spin_lock(&res->spinlock);
1479 		res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
1480 		spin_unlock(&res->spinlock);
1481 		wake_up(&res->wq);
1482 	}
1483 
1484 	/* at this point we have allocated everything we need,
1485 	 * and we have a hashed lockres with an extra ref and
1486 	 * the proper res->state flags. */
1487 	ret = 0;
1488 	spin_lock(&res->spinlock);
1489 	/* drop this either when master requery finds a different master
1490 	 * or when a lock is added by the recovery worker */
1491 	dlm_lockres_grab_inflight_ref(dlm, res);
1492 	if (mres->master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1493 		/* migration cannot have an unknown master */
1494 		BUG_ON(!(mres->flags & DLM_MRES_RECOVERY));
1495 		mlog(0, "recovery has passed me a lockres with an "
1496 			  "unknown owner.. will need to requery: "
1497 			  "%.*s\n", mres->lockname_len, mres->lockname);
1498 	} else {
1499 		/* take a reference now to pin the lockres, drop it
1500 		 * when locks are added in the worker */
1501 		dlm_change_lockres_owner(dlm, res, dlm->node_num);
1502 	}
1503 	spin_unlock(&res->spinlock);
1504 
1505 	/* queue up work for dlm_mig_lockres_worker */
1506 	dlm_grab(dlm);  /* get an extra ref for the work item */
1507 	memcpy(buf, msg->buf, be16_to_cpu(msg->data_len));  /* copy the whole message */
1508 	dlm_init_work_item(dlm, item, dlm_mig_lockres_worker, buf);
1509 	item->u.ml.lockres = res; /* already have a ref */
1510 	item->u.ml.real_master = real_master;
1511 	item->u.ml.extra_ref = extra_refs;
1512 	spin_lock(&dlm->work_lock);
1513 	list_add_tail(&item->list, &dlm->work_list);
1514 	spin_unlock(&dlm->work_lock);
1515 	queue_work(dlm->dlm_worker, &dlm->dispatched_work);
1516 
1517 leave:
1518 	/* One extra ref taken needs to be put here */
1519 	if (extra_refs)
1520 		dlm_lockres_put(res);
1521 
1522 	dlm_put(dlm);
1523 	if (ret < 0) {
1524 		kfree(buf);
1525 		kfree(item);
1526 		mlog_errno(ret);
1527 	}
1528 
1529 	return ret;
1530 }
1531 
1532 
dlm_mig_lockres_worker(struct dlm_work_item * item,void * data)1533 static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data)
1534 {
1535 	struct dlm_ctxt *dlm;
1536 	struct dlm_migratable_lockres *mres;
1537 	int ret = 0;
1538 	struct dlm_lock_resource *res;
1539 	u8 real_master;
1540 	u8 extra_ref;
1541 
1542 	dlm = item->dlm;
1543 	mres = (struct dlm_migratable_lockres *)data;
1544 
1545 	res = item->u.ml.lockres;
1546 	real_master = item->u.ml.real_master;
1547 	extra_ref = item->u.ml.extra_ref;
1548 
1549 	if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1550 		/* this case is super-rare. only occurs if
1551 		 * node death happens during migration. */
1552 again:
1553 		ret = dlm_lockres_master_requery(dlm, res, &real_master);
1554 		if (ret < 0) {
1555 			mlog(0, "dlm_lockres_master_requery ret=%d\n",
1556 				  ret);
1557 			goto again;
1558 		}
1559 		if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1560 			mlog(0, "lockres %.*s not claimed.  "
1561 				   "this node will take it.\n",
1562 				   res->lockname.len, res->lockname.name);
1563 		} else {
1564 			spin_lock(&res->spinlock);
1565 			dlm_lockres_drop_inflight_ref(dlm, res);
1566 			spin_unlock(&res->spinlock);
1567 			mlog(0, "master needs to respond to sender "
1568 				  "that node %u still owns %.*s\n",
1569 				  real_master, res->lockname.len,
1570 				  res->lockname.name);
1571 			/* cannot touch this lockres */
1572 			goto leave;
1573 		}
1574 	}
1575 
1576 	ret = dlm_process_recovery_data(dlm, res, mres);
1577 	if (ret < 0)
1578 		mlog(0, "dlm_process_recovery_data returned  %d\n", ret);
1579 	else
1580 		mlog(0, "dlm_process_recovery_data succeeded\n");
1581 
1582 	if ((mres->flags & (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) ==
1583 	                   (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) {
1584 		ret = dlm_finish_migration(dlm, res, mres->master);
1585 		if (ret < 0)
1586 			mlog_errno(ret);
1587 	}
1588 
1589 leave:
1590 	/* See comment in dlm_mig_lockres_handler() */
1591 	if (res) {
1592 		if (extra_ref)
1593 			dlm_lockres_put(res);
1594 		dlm_lockres_put(res);
1595 	}
1596 	kfree(data);
1597 }
1598 
1599 
1600 
dlm_lockres_master_requery(struct dlm_ctxt * dlm,struct dlm_lock_resource * res,u8 * real_master)1601 static int dlm_lockres_master_requery(struct dlm_ctxt *dlm,
1602 				      struct dlm_lock_resource *res,
1603 				      u8 *real_master)
1604 {
1605 	struct dlm_node_iter iter;
1606 	int nodenum;
1607 	int ret = 0;
1608 
1609 	*real_master = DLM_LOCK_RES_OWNER_UNKNOWN;
1610 
1611 	/* we only reach here if one of the two nodes in a
1612 	 * migration died while the migration was in progress.
1613 	 * at this point we need to requery the master.  we
1614 	 * know that the new_master got as far as creating
1615 	 * an mle on at least one node, but we do not know
1616 	 * if any nodes had actually cleared the mle and set
1617 	 * the master to the new_master.  the old master
1618 	 * is supposed to set the owner to UNKNOWN in the
1619 	 * event of a new_master death, so the only possible
1620 	 * responses that we can get from nodes here are
1621 	 * that the master is new_master, or that the master
1622 	 * is UNKNOWN.
1623 	 * if all nodes come back with UNKNOWN then we know
1624 	 * the lock needs remastering here.
1625 	 * if any node comes back with a valid master, check
1626 	 * to see if that master is the one that we are
1627 	 * recovering.  if so, then the new_master died and
1628 	 * we need to remaster this lock.  if not, then the
1629 	 * new_master survived and that node will respond to
1630 	 * other nodes about the owner.
1631 	 * if there is an owner, this node needs to dump this
1632 	 * lockres and alert the sender that this lockres
1633 	 * was rejected. */
1634 	spin_lock(&dlm->spinlock);
1635 	dlm_node_iter_init(dlm->domain_map, &iter);
1636 	spin_unlock(&dlm->spinlock);
1637 
1638 	while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
1639 		/* do not send to self */
1640 		if (nodenum == dlm->node_num)
1641 			continue;
1642 		ret = dlm_do_master_requery(dlm, res, nodenum, real_master);
1643 		if (ret < 0) {
1644 			mlog_errno(ret);
1645 			if (!dlm_is_host_down(ret))
1646 				BUG();
1647 			/* host is down, so answer for that node would be
1648 			 * DLM_LOCK_RES_OWNER_UNKNOWN.  continue. */
1649 		}
1650 		if (*real_master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1651 			mlog(0, "lock master is %u\n", *real_master);
1652 			break;
1653 		}
1654 	}
1655 	return ret;
1656 }
1657 
1658 
dlm_do_master_requery(struct dlm_ctxt * dlm,struct dlm_lock_resource * res,u8 nodenum,u8 * real_master)1659 int dlm_do_master_requery(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
1660 			  u8 nodenum, u8 *real_master)
1661 {
1662 	int ret;
1663 	struct dlm_master_requery req;
1664 	int status = DLM_LOCK_RES_OWNER_UNKNOWN;
1665 
1666 	memset(&req, 0, sizeof(req));
1667 	req.node_idx = dlm->node_num;
1668 	req.namelen = res->lockname.len;
1669 	memcpy(req.name, res->lockname.name, res->lockname.len);
1670 
1671 resend:
1672 	ret = o2net_send_message(DLM_MASTER_REQUERY_MSG, dlm->key,
1673 				 &req, sizeof(req), nodenum, &status);
1674 	if (ret < 0)
1675 		mlog(ML_ERROR, "Error %d when sending message %u (key "
1676 		     "0x%x) to node %u\n", ret, DLM_MASTER_REQUERY_MSG,
1677 		     dlm->key, nodenum);
1678 	else if (status == -ENOMEM) {
1679 		mlog_errno(status);
1680 		msleep(50);
1681 		goto resend;
1682 	} else {
1683 		BUG_ON(status < 0);
1684 		BUG_ON(status > DLM_LOCK_RES_OWNER_UNKNOWN);
1685 		*real_master = (u8) (status & 0xff);
1686 		mlog(0, "node %u responded to master requery with %u\n",
1687 			  nodenum, *real_master);
1688 		ret = 0;
1689 	}
1690 	return ret;
1691 }
1692 
1693 
1694 /* this function cannot error, so unless the sending
1695  * or receiving of the message failed, the owner can
1696  * be trusted */
dlm_master_requery_handler(struct o2net_msg * msg,u32 len,void * data,void ** ret_data)1697 int dlm_master_requery_handler(struct o2net_msg *msg, u32 len, void *data,
1698 			       void **ret_data)
1699 {
1700 	struct dlm_ctxt *dlm = data;
1701 	struct dlm_master_requery *req = (struct dlm_master_requery *)msg->buf;
1702 	struct dlm_lock_resource *res = NULL;
1703 	unsigned int hash;
1704 	int master = DLM_LOCK_RES_OWNER_UNKNOWN;
1705 	u32 flags = DLM_ASSERT_MASTER_REQUERY;
1706 	int dispatched = 0;
1707 
1708 	if (!dlm_grab(dlm)) {
1709 		/* since the domain has gone away on this
1710 		 * node, the proper response is UNKNOWN */
1711 		return master;
1712 	}
1713 
1714 	hash = dlm_lockid_hash(req->name, req->namelen);
1715 
1716 	spin_lock(&dlm->spinlock);
1717 	res = __dlm_lookup_lockres(dlm, req->name, req->namelen, hash);
1718 	if (res) {
1719 		spin_lock(&res->spinlock);
1720 		master = res->owner;
1721 		if (master == dlm->node_num) {
1722 			int ret = dlm_dispatch_assert_master(dlm, res,
1723 							     0, 0, flags);
1724 			if (ret < 0) {
1725 				mlog_errno(ret);
1726 				spin_unlock(&res->spinlock);
1727 				dlm_lockres_put(res);
1728 				spin_unlock(&dlm->spinlock);
1729 				dlm_put(dlm);
1730 				/* sender will take care of this and retry */
1731 				return ret;
1732 			} else {
1733 				dispatched = 1;
1734 				__dlm_lockres_grab_inflight_worker(dlm, res);
1735 				spin_unlock(&res->spinlock);
1736 			}
1737 		} else {
1738 			/* put.. incase we are not the master */
1739 			spin_unlock(&res->spinlock);
1740 			dlm_lockres_put(res);
1741 		}
1742 	}
1743 	spin_unlock(&dlm->spinlock);
1744 
1745 	if (!dispatched)
1746 		dlm_put(dlm);
1747 	return master;
1748 }
1749 
1750 static inline struct list_head *
dlm_list_num_to_pointer(struct dlm_lock_resource * res,int list_num)1751 dlm_list_num_to_pointer(struct dlm_lock_resource *res, int list_num)
1752 {
1753 	struct list_head *ret;
1754 	BUG_ON(list_num < 0);
1755 	BUG_ON(list_num > 2);
1756 	ret = &(res->granted);
1757 	ret += list_num;
1758 	return ret;
1759 }
1760 /* TODO: do ast flush business
1761  * TODO: do MIGRATING and RECOVERING spinning
1762  */
1763 
1764 /*
1765 * NOTE about in-flight requests during migration:
1766 *
1767 * Before attempting the migrate, the master has marked the lockres as
1768 * MIGRATING and then flushed all of its pending ASTS.  So any in-flight
1769 * requests either got queued before the MIGRATING flag got set, in which
1770 * case the lock data will reflect the change and a return message is on
1771 * the way, or the request failed to get in before MIGRATING got set.  In
1772 * this case, the caller will be told to spin and wait for the MIGRATING
1773 * flag to be dropped, then recheck the master.
1774 * This holds true for the convert, cancel and unlock cases, and since lvb
1775 * updates are tied to these same messages, it applies to lvb updates as
1776 * well.  For the lock case, there is no way a lock can be on the master
1777 * queue and not be on the secondary queue since the lock is always added
1778 * locally first.  This means that the new target node will never be sent
1779 * a lock that he doesn't already have on the list.
1780 * In total, this means that the local lock is correct and should not be
1781 * updated to match the one sent by the master.  Any messages sent back
1782 * from the master before the MIGRATING flag will bring the lock properly
1783 * up-to-date, and the change will be ordered properly for the waiter.
1784 * We will *not* attempt to modify the lock underneath the waiter.
1785 */
1786 
dlm_process_recovery_data(struct dlm_ctxt * dlm,struct dlm_lock_resource * res,struct dlm_migratable_lockres * mres)1787 static int dlm_process_recovery_data(struct dlm_ctxt *dlm,
1788 				     struct dlm_lock_resource *res,
1789 				     struct dlm_migratable_lockres *mres)
1790 {
1791 	struct dlm_migratable_lock *ml;
1792 	struct list_head *queue, *iter;
1793 	struct list_head *tmpq = NULL;
1794 	struct dlm_lock *newlock = NULL;
1795 	struct dlm_lockstatus *lksb = NULL;
1796 	int ret = 0;
1797 	int i, j, bad;
1798 	struct dlm_lock *lock;
1799 	u8 from = O2NM_MAX_NODES;
1800 	__be64 c;
1801 
1802 	mlog(0, "running %d locks for this lockres\n", mres->num_locks);
1803 	for (i=0; i<mres->num_locks; i++) {
1804 		ml = &(mres->ml[i]);
1805 
1806 		if (dlm_is_dummy_lock(dlm, ml, &from)) {
1807 			/* placeholder, just need to set the refmap bit */
1808 			BUG_ON(mres->num_locks != 1);
1809 			mlog(0, "%s:%.*s: dummy lock for %u\n",
1810 			     dlm->name, mres->lockname_len, mres->lockname,
1811 			     from);
1812 			spin_lock(&res->spinlock);
1813 			dlm_lockres_set_refmap_bit(dlm, res, from);
1814 			spin_unlock(&res->spinlock);
1815 			break;
1816 		}
1817 		BUG_ON(ml->highest_blocked != LKM_IVMODE);
1818 		newlock = NULL;
1819 		lksb = NULL;
1820 
1821 		queue = dlm_list_num_to_pointer(res, ml->list);
1822 		tmpq = NULL;
1823 
1824 		/* if the lock is for the local node it needs to
1825 		 * be moved to the proper location within the queue.
1826 		 * do not allocate a new lock structure. */
1827 		if (ml->node == dlm->node_num) {
1828 			/* MIGRATION ONLY! */
1829 			BUG_ON(!(mres->flags & DLM_MRES_MIGRATION));
1830 
1831 			lock = NULL;
1832 			spin_lock(&res->spinlock);
1833 			for (j = DLM_GRANTED_LIST; j <= DLM_BLOCKED_LIST; j++) {
1834 				tmpq = dlm_list_idx_to_ptr(res, j);
1835 				list_for_each(iter, tmpq) {
1836 					lock = list_entry(iter,
1837 						  struct dlm_lock, list);
1838 					if (lock->ml.cookie == ml->cookie)
1839 						break;
1840 					lock = NULL;
1841 				}
1842 				if (lock)
1843 					break;
1844 			}
1845 
1846 			/* lock is always created locally first, and
1847 			 * destroyed locally last.  it must be on the list */
1848 			if (!lock) {
1849 				c = ml->cookie;
1850 				mlog(ML_ERROR, "Could not find local lock "
1851 					       "with cookie %u:%llu, node %u, "
1852 					       "list %u, flags 0x%x, type %d, "
1853 					       "conv %d, highest blocked %d\n",
1854 				     dlm_get_lock_cookie_node(be64_to_cpu(c)),
1855 				     dlm_get_lock_cookie_seq(be64_to_cpu(c)),
1856 				     ml->node, ml->list, ml->flags, ml->type,
1857 				     ml->convert_type, ml->highest_blocked);
1858 				__dlm_print_one_lock_resource(res);
1859 				BUG();
1860 			}
1861 
1862 			if (lock->ml.node != ml->node) {
1863 				c = lock->ml.cookie;
1864 				mlog(ML_ERROR, "Mismatched node# in lock "
1865 				     "cookie %u:%llu, name %.*s, node %u\n",
1866 				     dlm_get_lock_cookie_node(be64_to_cpu(c)),
1867 				     dlm_get_lock_cookie_seq(be64_to_cpu(c)),
1868 				     res->lockname.len, res->lockname.name,
1869 				     lock->ml.node);
1870 				c = ml->cookie;
1871 				mlog(ML_ERROR, "Migrate lock cookie %u:%llu, "
1872 				     "node %u, list %u, flags 0x%x, type %d, "
1873 				     "conv %d, highest blocked %d\n",
1874 				     dlm_get_lock_cookie_node(be64_to_cpu(c)),
1875 				     dlm_get_lock_cookie_seq(be64_to_cpu(c)),
1876 				     ml->node, ml->list, ml->flags, ml->type,
1877 				     ml->convert_type, ml->highest_blocked);
1878 				__dlm_print_one_lock_resource(res);
1879 				BUG();
1880 			}
1881 
1882 			if (tmpq != queue) {
1883 				c = ml->cookie;
1884 				mlog(0, "Lock cookie %u:%llu was on list %u "
1885 				     "instead of list %u for %.*s\n",
1886 				     dlm_get_lock_cookie_node(be64_to_cpu(c)),
1887 				     dlm_get_lock_cookie_seq(be64_to_cpu(c)),
1888 				     j, ml->list, res->lockname.len,
1889 				     res->lockname.name);
1890 				__dlm_print_one_lock_resource(res);
1891 				spin_unlock(&res->spinlock);
1892 				continue;
1893 			}
1894 
1895 			/* see NOTE above about why we do not update
1896 			 * to match the master here */
1897 
1898 			/* move the lock to its proper place */
1899 			/* do not alter lock refcount.  switching lists. */
1900 			list_move_tail(&lock->list, queue);
1901 			spin_unlock(&res->spinlock);
1902 
1903 			mlog(0, "just reordered a local lock!\n");
1904 			continue;
1905 		}
1906 
1907 		/* lock is for another node. */
1908 		newlock = dlm_new_lock(ml->type, ml->node,
1909 				       be64_to_cpu(ml->cookie), NULL);
1910 		if (!newlock) {
1911 			ret = -ENOMEM;
1912 			goto leave;
1913 		}
1914 		lksb = newlock->lksb;
1915 		dlm_lock_attach_lockres(newlock, res);
1916 
1917 		if (ml->convert_type != LKM_IVMODE) {
1918 			BUG_ON(queue != &res->converting);
1919 			newlock->ml.convert_type = ml->convert_type;
1920 		}
1921 		lksb->flags |= (ml->flags &
1922 				(DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB));
1923 
1924 		if (ml->type == LKM_NLMODE)
1925 			goto skip_lvb;
1926 
1927 		/*
1928 		 * If the lock is in the blocked list it can't have a valid lvb,
1929 		 * so skip it
1930 		 */
1931 		if (ml->list == DLM_BLOCKED_LIST)
1932 			goto skip_lvb;
1933 
1934 		if (!dlm_lvb_is_empty(mres->lvb)) {
1935 			if (lksb->flags & DLM_LKSB_PUT_LVB) {
1936 				/* other node was trying to update
1937 				 * lvb when node died.  recreate the
1938 				 * lksb with the updated lvb. */
1939 				memcpy(lksb->lvb, mres->lvb, DLM_LVB_LEN);
1940 				/* the lock resource lvb update must happen
1941 				 * NOW, before the spinlock is dropped.
1942 				 * we no longer wait for the AST to update
1943 				 * the lvb. */
1944 				memcpy(res->lvb, mres->lvb, DLM_LVB_LEN);
1945 			} else {
1946 				/* otherwise, the node is sending its
1947 				 * most recent valid lvb info */
1948 				BUG_ON(ml->type != LKM_EXMODE &&
1949 				       ml->type != LKM_PRMODE);
1950 				if (!dlm_lvb_is_empty(res->lvb) &&
1951  				    (ml->type == LKM_EXMODE ||
1952  				     memcmp(res->lvb, mres->lvb, DLM_LVB_LEN))) {
1953  					int i;
1954  					mlog(ML_ERROR, "%s:%.*s: received bad "
1955  					     "lvb! type=%d\n", dlm->name,
1956  					     res->lockname.len,
1957  					     res->lockname.name, ml->type);
1958  					printk("lockres lvb=[");
1959  					for (i=0; i<DLM_LVB_LEN; i++)
1960  						printk("%02x", res->lvb[i]);
1961  					printk("]\nmigrated lvb=[");
1962  					for (i=0; i<DLM_LVB_LEN; i++)
1963  						printk("%02x", mres->lvb[i]);
1964  					printk("]\n");
1965  					dlm_print_one_lock_resource(res);
1966  					BUG();
1967 				}
1968 				memcpy(res->lvb, mres->lvb, DLM_LVB_LEN);
1969 			}
1970 		}
1971 skip_lvb:
1972 
1973 		/* NOTE:
1974 		 * wrt lock queue ordering and recovery:
1975 		 *    1. order of locks on granted queue is
1976 		 *       meaningless.
1977 		 *    2. order of locks on converting queue is
1978 		 *       LOST with the node death.  sorry charlie.
1979 		 *    3. order of locks on the blocked queue is
1980 		 *       also LOST.
1981 		 * order of locks does not affect integrity, it
1982 		 * just means that a lock request may get pushed
1983 		 * back in line as a result of the node death.
1984 		 * also note that for a given node the lock order
1985 		 * for its secondary queue locks is preserved
1986 		 * relative to each other, but clearly *not*
1987 		 * preserved relative to locks from other nodes.
1988 		 */
1989 		bad = 0;
1990 		spin_lock(&res->spinlock);
1991 		list_for_each_entry(lock, queue, list) {
1992 			if (lock->ml.cookie == ml->cookie) {
1993 				c = lock->ml.cookie;
1994 				mlog(ML_ERROR, "%s:%.*s: %u:%llu: lock already "
1995 				     "exists on this lockres!\n", dlm->name,
1996 				     res->lockname.len, res->lockname.name,
1997 				     dlm_get_lock_cookie_node(be64_to_cpu(c)),
1998 				     dlm_get_lock_cookie_seq(be64_to_cpu(c)));
1999 
2000 				mlog(ML_NOTICE, "sent lock: type=%d, conv=%d, "
2001 				     "node=%u, cookie=%u:%llu, queue=%d\n",
2002 	      			     ml->type, ml->convert_type, ml->node,
2003 				     dlm_get_lock_cookie_node(be64_to_cpu(ml->cookie)),
2004 				     dlm_get_lock_cookie_seq(be64_to_cpu(ml->cookie)),
2005 				     ml->list);
2006 
2007 				__dlm_print_one_lock_resource(res);
2008 				bad = 1;
2009 				break;
2010 			}
2011 		}
2012 		if (!bad) {
2013 			dlm_lock_get(newlock);
2014 			if (mres->flags & DLM_MRES_RECOVERY &&
2015 					ml->list == DLM_CONVERTING_LIST &&
2016 					newlock->ml.type >
2017 					newlock->ml.convert_type) {
2018 				/* newlock is doing downconvert, add it to the
2019 				 * head of converting list */
2020 				list_add(&newlock->list, queue);
2021 			} else
2022 				list_add_tail(&newlock->list, queue);
2023 			mlog(0, "%s:%.*s: added lock for node %u, "
2024 			     "setting refmap bit\n", dlm->name,
2025 			     res->lockname.len, res->lockname.name, ml->node);
2026 			dlm_lockres_set_refmap_bit(dlm, res, ml->node);
2027 		}
2028 		spin_unlock(&res->spinlock);
2029 	}
2030 	mlog(0, "done running all the locks\n");
2031 
2032 leave:
2033 	/* balance the ref taken when the work was queued */
2034 	spin_lock(&res->spinlock);
2035 	dlm_lockres_drop_inflight_ref(dlm, res);
2036 	spin_unlock(&res->spinlock);
2037 
2038 	if (ret < 0)
2039 		mlog_errno(ret);
2040 
2041 	return ret;
2042 }
2043 
dlm_move_lockres_to_recovery_list(struct dlm_ctxt * dlm,struct dlm_lock_resource * res)2044 void dlm_move_lockres_to_recovery_list(struct dlm_ctxt *dlm,
2045 				       struct dlm_lock_resource *res)
2046 {
2047 	int i;
2048 	struct list_head *queue;
2049 	struct dlm_lock *lock, *next;
2050 
2051 	assert_spin_locked(&dlm->spinlock);
2052 	assert_spin_locked(&res->spinlock);
2053 	res->state |= DLM_LOCK_RES_RECOVERING;
2054 	if (!list_empty(&res->recovering)) {
2055 		mlog(0,
2056 		     "Recovering res %s:%.*s, is already on recovery list!\n",
2057 		     dlm->name, res->lockname.len, res->lockname.name);
2058 		list_del_init(&res->recovering);
2059 		dlm_lockres_put(res);
2060 	}
2061 	/* We need to hold a reference while on the recovery list */
2062 	dlm_lockres_get(res);
2063 	list_add_tail(&res->recovering, &dlm->reco.resources);
2064 
2065 	/* find any pending locks and put them back on proper list */
2066 	for (i=DLM_BLOCKED_LIST; i>=DLM_GRANTED_LIST; i--) {
2067 		queue = dlm_list_idx_to_ptr(res, i);
2068 		list_for_each_entry_safe(lock, next, queue, list) {
2069 			dlm_lock_get(lock);
2070 			if (lock->convert_pending) {
2071 				/* move converting lock back to granted */
2072 				mlog(0, "node died with convert pending "
2073 				     "on %.*s. move back to granted list.\n",
2074 				     res->lockname.len, res->lockname.name);
2075 				dlm_revert_pending_convert(res, lock);
2076 				lock->convert_pending = 0;
2077 			} else if (lock->lock_pending) {
2078 				/* remove pending lock requests completely */
2079 				BUG_ON(i != DLM_BLOCKED_LIST);
2080 				mlog(0, "node died with lock pending "
2081 				     "on %.*s. remove from blocked list and skip.\n",
2082 				     res->lockname.len, res->lockname.name);
2083 				/* lock will be floating until ref in
2084 				 * dlmlock_remote is freed after the network
2085 				 * call returns.  ok for it to not be on any
2086 				 * list since no ast can be called
2087 				 * (the master is dead). */
2088 				dlm_revert_pending_lock(res, lock);
2089 				lock->lock_pending = 0;
2090 			} else if (lock->unlock_pending) {
2091 				/* if an unlock was in progress, treat as
2092 				 * if this had completed successfully
2093 				 * before sending this lock state to the
2094 				 * new master.  note that the dlm_unlock
2095 				 * call is still responsible for calling
2096 				 * the unlockast.  that will happen after
2097 				 * the network call times out.  for now,
2098 				 * just move lists to prepare the new
2099 				 * recovery master.  */
2100 				BUG_ON(i != DLM_GRANTED_LIST);
2101 				mlog(0, "node died with unlock pending "
2102 				     "on %.*s. remove from blocked list and skip.\n",
2103 				     res->lockname.len, res->lockname.name);
2104 				dlm_commit_pending_unlock(res, lock);
2105 				lock->unlock_pending = 0;
2106 			} else if (lock->cancel_pending) {
2107 				/* if a cancel was in progress, treat as
2108 				 * if this had completed successfully
2109 				 * before sending this lock state to the
2110 				 * new master */
2111 				BUG_ON(i != DLM_CONVERTING_LIST);
2112 				mlog(0, "node died with cancel pending "
2113 				     "on %.*s. move back to granted list.\n",
2114 				     res->lockname.len, res->lockname.name);
2115 				dlm_commit_pending_cancel(res, lock);
2116 				lock->cancel_pending = 0;
2117 			}
2118 			dlm_lock_put(lock);
2119 		}
2120 	}
2121 }
2122 
2123 
2124 
2125 /* removes all recovered locks from the recovery list.
2126  * sets the res->owner to the new master.
2127  * unsets the RECOVERY flag and wakes waiters. */
dlm_finish_local_lockres_recovery(struct dlm_ctxt * dlm,u8 dead_node,u8 new_master)2128 static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm,
2129 					      u8 dead_node, u8 new_master)
2130 {
2131 	int i;
2132 	struct hlist_head *bucket;
2133 	struct dlm_lock_resource *res, *next;
2134 
2135 	assert_spin_locked(&dlm->spinlock);
2136 
2137 	list_for_each_entry_safe(res, next, &dlm->reco.resources, recovering) {
2138 		if (res->owner == dead_node) {
2139 			mlog(0, "%s: res %.*s, Changing owner from %u to %u\n",
2140 			     dlm->name, res->lockname.len, res->lockname.name,
2141 			     res->owner, new_master);
2142 			list_del_init(&res->recovering);
2143 			spin_lock(&res->spinlock);
2144 			/* new_master has our reference from
2145 			 * the lock state sent during recovery */
2146 			dlm_change_lockres_owner(dlm, res, new_master);
2147 			res->state &= ~DLM_LOCK_RES_RECOVERING;
2148 			if (__dlm_lockres_has_locks(res))
2149 				__dlm_dirty_lockres(dlm, res);
2150 			spin_unlock(&res->spinlock);
2151 			wake_up(&res->wq);
2152 			dlm_lockres_put(res);
2153 		}
2154 	}
2155 
2156 	/* this will become unnecessary eventually, but
2157 	 * for now we need to run the whole hash, clear
2158 	 * the RECOVERING state and set the owner
2159 	 * if necessary */
2160 	for (i = 0; i < DLM_HASH_BUCKETS; i++) {
2161 		bucket = dlm_lockres_hash(dlm, i);
2162 		hlist_for_each_entry(res, bucket, hash_node) {
2163 			if (res->state & DLM_LOCK_RES_RECOVERY_WAITING) {
2164 				spin_lock(&res->spinlock);
2165 				res->state &= ~DLM_LOCK_RES_RECOVERY_WAITING;
2166 				spin_unlock(&res->spinlock);
2167 				wake_up(&res->wq);
2168 			}
2169 
2170 			if (!(res->state & DLM_LOCK_RES_RECOVERING))
2171 				continue;
2172 
2173 			if (res->owner != dead_node &&
2174 			    res->owner != dlm->node_num)
2175 				continue;
2176 
2177 			if (!list_empty(&res->recovering)) {
2178 				list_del_init(&res->recovering);
2179 				dlm_lockres_put(res);
2180 			}
2181 
2182 			/* new_master has our reference from
2183 			 * the lock state sent during recovery */
2184 			mlog(0, "%s: res %.*s, Changing owner from %u to %u\n",
2185 			     dlm->name, res->lockname.len, res->lockname.name,
2186 			     res->owner, new_master);
2187 			spin_lock(&res->spinlock);
2188 			dlm_change_lockres_owner(dlm, res, new_master);
2189 			res->state &= ~DLM_LOCK_RES_RECOVERING;
2190 			if (__dlm_lockres_has_locks(res))
2191 				__dlm_dirty_lockres(dlm, res);
2192 			spin_unlock(&res->spinlock);
2193 			wake_up(&res->wq);
2194 		}
2195 	}
2196 }
2197 
dlm_lvb_needs_invalidation(struct dlm_lock * lock,int local)2198 static inline int dlm_lvb_needs_invalidation(struct dlm_lock *lock, int local)
2199 {
2200 	if (local) {
2201 		if (lock->ml.type != LKM_EXMODE &&
2202 		    lock->ml.type != LKM_PRMODE)
2203 			return 1;
2204 	} else if (lock->ml.type == LKM_EXMODE)
2205 		return 1;
2206 	return 0;
2207 }
2208 
dlm_revalidate_lvb(struct dlm_ctxt * dlm,struct dlm_lock_resource * res,u8 dead_node)2209 static void dlm_revalidate_lvb(struct dlm_ctxt *dlm,
2210 			       struct dlm_lock_resource *res, u8 dead_node)
2211 {
2212 	struct list_head *queue;
2213 	struct dlm_lock *lock;
2214 	int blank_lvb = 0, local = 0;
2215 	int i;
2216 	u8 search_node;
2217 
2218 	assert_spin_locked(&dlm->spinlock);
2219 	assert_spin_locked(&res->spinlock);
2220 
2221 	if (res->owner == dlm->node_num)
2222 		/* if this node owned the lockres, and if the dead node
2223 		 * had an EX when he died, blank out the lvb */
2224 		search_node = dead_node;
2225 	else {
2226 		/* if this is a secondary lockres, and we had no EX or PR
2227 		 * locks granted, we can no longer trust the lvb */
2228 		search_node = dlm->node_num;
2229 		local = 1;  /* check local state for valid lvb */
2230 	}
2231 
2232 	for (i=DLM_GRANTED_LIST; i<=DLM_CONVERTING_LIST; i++) {
2233 		queue = dlm_list_idx_to_ptr(res, i);
2234 		list_for_each_entry(lock, queue, list) {
2235 			if (lock->ml.node == search_node) {
2236 				if (dlm_lvb_needs_invalidation(lock, local)) {
2237 					/* zero the lksb lvb and lockres lvb */
2238 					blank_lvb = 1;
2239 					memset(lock->lksb->lvb, 0, DLM_LVB_LEN);
2240 				}
2241 			}
2242 		}
2243 	}
2244 
2245 	if (blank_lvb) {
2246 		mlog(0, "clearing %.*s lvb, dead node %u had EX\n",
2247 		     res->lockname.len, res->lockname.name, dead_node);
2248 		memset(res->lvb, 0, DLM_LVB_LEN);
2249 	}
2250 }
2251 
dlm_free_dead_locks(struct dlm_ctxt * dlm,struct dlm_lock_resource * res,u8 dead_node)2252 static void dlm_free_dead_locks(struct dlm_ctxt *dlm,
2253 				struct dlm_lock_resource *res, u8 dead_node)
2254 {
2255 	struct dlm_lock *lock, *next;
2256 	unsigned int freed = 0;
2257 
2258 	/* this node is the lockres master:
2259 	 * 1) remove any stale locks for the dead node
2260 	 * 2) if the dead node had an EX when he died, blank out the lvb
2261 	 */
2262 	assert_spin_locked(&dlm->spinlock);
2263 	assert_spin_locked(&res->spinlock);
2264 
2265 	/* We do two dlm_lock_put(). One for removing from list and the other is
2266 	 * to force the DLM_UNLOCK_FREE_LOCK action so as to free the locks */
2267 
2268 	/* TODO: check pending_asts, pending_basts here */
2269 	list_for_each_entry_safe(lock, next, &res->granted, list) {
2270 		if (lock->ml.node == dead_node) {
2271 			list_del_init(&lock->list);
2272 			dlm_lock_put(lock);
2273 			/* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */
2274 			dlm_lock_put(lock);
2275 			freed++;
2276 		}
2277 	}
2278 	list_for_each_entry_safe(lock, next, &res->converting, list) {
2279 		if (lock->ml.node == dead_node) {
2280 			list_del_init(&lock->list);
2281 			dlm_lock_put(lock);
2282 			/* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */
2283 			dlm_lock_put(lock);
2284 			freed++;
2285 		}
2286 	}
2287 	list_for_each_entry_safe(lock, next, &res->blocked, list) {
2288 		if (lock->ml.node == dead_node) {
2289 			list_del_init(&lock->list);
2290 			dlm_lock_put(lock);
2291 			/* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */
2292 			dlm_lock_put(lock);
2293 			freed++;
2294 		}
2295 	}
2296 
2297 	if (freed) {
2298 		mlog(0, "%s:%.*s: freed %u locks for dead node %u, "
2299 		     "dropping ref from lockres\n", dlm->name,
2300 		     res->lockname.len, res->lockname.name, freed, dead_node);
2301 		if(!test_bit(dead_node, res->refmap)) {
2302 			mlog(ML_ERROR, "%s:%.*s: freed %u locks for dead node %u, "
2303 			     "but ref was not set\n", dlm->name,
2304 			     res->lockname.len, res->lockname.name, freed, dead_node);
2305 			__dlm_print_one_lock_resource(res);
2306 		}
2307 		res->state |= DLM_LOCK_RES_RECOVERY_WAITING;
2308 		dlm_lockres_clear_refmap_bit(dlm, res, dead_node);
2309 	} else if (test_bit(dead_node, res->refmap)) {
2310 		mlog(0, "%s:%.*s: dead node %u had a ref, but had "
2311 		     "no locks and had not purged before dying\n", dlm->name,
2312 		     res->lockname.len, res->lockname.name, dead_node);
2313 		dlm_lockres_clear_refmap_bit(dlm, res, dead_node);
2314 	}
2315 
2316 	/* do not kick thread yet */
2317 	__dlm_dirty_lockres(dlm, res);
2318 }
2319 
dlm_do_local_recovery_cleanup(struct dlm_ctxt * dlm,u8 dead_node)2320 static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node)
2321 {
2322 	struct dlm_lock_resource *res;
2323 	int i;
2324 	struct hlist_head *bucket;
2325 	struct hlist_node *tmp;
2326 	struct dlm_lock *lock;
2327 
2328 
2329 	/* purge any stale mles */
2330 	dlm_clean_master_list(dlm, dead_node);
2331 
2332 	/*
2333 	 * now clean up all lock resources.  there are two rules:
2334 	 *
2335 	 * 1) if the dead node was the master, move the lockres
2336 	 *    to the recovering list.  set the RECOVERING flag.
2337 	 *    this lockres needs to be cleaned up before it can
2338 	 *    be used further.
2339 	 *
2340 	 * 2) if this node was the master, remove all locks from
2341 	 *    each of the lockres queues that were owned by the
2342 	 *    dead node.  once recovery finishes, the dlm thread
2343 	 *    can be kicked again to see if any ASTs or BASTs
2344 	 *    need to be fired as a result.
2345 	 */
2346 	for (i = 0; i < DLM_HASH_BUCKETS; i++) {
2347 		bucket = dlm_lockres_hash(dlm, i);
2348 		hlist_for_each_entry_safe(res, tmp, bucket, hash_node) {
2349  			/* always prune any $RECOVERY entries for dead nodes,
2350  			 * otherwise hangs can occur during later recovery */
2351 			if (dlm_is_recovery_lock(res->lockname.name,
2352 						 res->lockname.len)) {
2353 				spin_lock(&res->spinlock);
2354 				list_for_each_entry(lock, &res->granted, list) {
2355 					if (lock->ml.node == dead_node) {
2356 						mlog(0, "AHA! there was "
2357 						     "a $RECOVERY lock for dead "
2358 						     "node %u (%s)!\n",
2359 						     dead_node, dlm->name);
2360 						list_del_init(&lock->list);
2361 						dlm_lock_put(lock);
2362 						/* Can't schedule
2363 						 * DLM_UNLOCK_FREE_LOCK
2364 						 * - do manually */
2365 						dlm_lock_put(lock);
2366 						break;
2367 					}
2368 				}
2369 
2370 				if ((res->owner == dead_node) &&
2371 							(res->state & DLM_LOCK_RES_DROPPING_REF)) {
2372 					dlm_lockres_get(res);
2373 					__dlm_do_purge_lockres(dlm, res);
2374 					spin_unlock(&res->spinlock);
2375 					wake_up(&res->wq);
2376 					dlm_lockres_put(res);
2377 					continue;
2378 				} else if (res->owner == dlm->node_num)
2379 					dlm_lockres_clear_refmap_bit(dlm, res, dead_node);
2380 				spin_unlock(&res->spinlock);
2381 				continue;
2382 			}
2383 			spin_lock(&res->spinlock);
2384 			/* zero the lvb if necessary */
2385 			dlm_revalidate_lvb(dlm, res, dead_node);
2386 			if (res->owner == dead_node) {
2387 				if (res->state & DLM_LOCK_RES_DROPPING_REF) {
2388 					mlog(0, "%s:%.*s: owned by "
2389 						"dead node %u, this node was "
2390 						"dropping its ref when master died. "
2391 						"continue, purging the lockres.\n",
2392 						dlm->name, res->lockname.len,
2393 						res->lockname.name, dead_node);
2394 					dlm_lockres_get(res);
2395 					__dlm_do_purge_lockres(dlm, res);
2396 					spin_unlock(&res->spinlock);
2397 					wake_up(&res->wq);
2398 					dlm_lockres_put(res);
2399 					continue;
2400 				}
2401 				dlm_move_lockres_to_recovery_list(dlm, res);
2402 			} else if (res->owner == dlm->node_num) {
2403 				dlm_free_dead_locks(dlm, res, dead_node);
2404 				__dlm_lockres_calc_usage(dlm, res);
2405 			} else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
2406 				if (test_bit(dead_node, res->refmap)) {
2407 					mlog(0, "%s:%.*s: dead node %u had a ref, but had "
2408 						"no locks and had not purged before dying\n",
2409 						dlm->name, res->lockname.len,
2410 						res->lockname.name, dead_node);
2411 					dlm_lockres_clear_refmap_bit(dlm, res, dead_node);
2412 				}
2413 			}
2414 			spin_unlock(&res->spinlock);
2415 		}
2416 	}
2417 
2418 }
2419 
__dlm_hb_node_down(struct dlm_ctxt * dlm,int idx)2420 static void __dlm_hb_node_down(struct dlm_ctxt *dlm, int idx)
2421 {
2422 	assert_spin_locked(&dlm->spinlock);
2423 
2424 	if (dlm->reco.new_master == idx) {
2425 		mlog(0, "%s: recovery master %d just died\n",
2426 		     dlm->name, idx);
2427 		if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2428 			/* finalize1 was reached, so it is safe to clear
2429 			 * the new_master and dead_node.  that recovery
2430 			 * is complete. */
2431 			mlog(0, "%s: dead master %d had reached "
2432 			     "finalize1 state, clearing\n", dlm->name, idx);
2433 			dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE;
2434 			__dlm_reset_recovery(dlm);
2435 		}
2436 	}
2437 
2438 	/* Clean up join state on node death. */
2439 	if (dlm->joining_node == idx) {
2440 		mlog(0, "Clearing join state for node %u\n", idx);
2441 		__dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
2442 	}
2443 
2444 	/* check to see if the node is already considered dead */
2445 	if (!test_bit(idx, dlm->live_nodes_map)) {
2446 		mlog(0, "for domain %s, node %d is already dead. "
2447 		     "another node likely did recovery already.\n",
2448 		     dlm->name, idx);
2449 		return;
2450 	}
2451 
2452 	/* check to see if we do not care about this node */
2453 	if (!test_bit(idx, dlm->domain_map)) {
2454 		/* This also catches the case that we get a node down
2455 		 * but haven't joined the domain yet. */
2456 		mlog(0, "node %u already removed from domain!\n", idx);
2457 		return;
2458 	}
2459 
2460 	clear_bit(idx, dlm->live_nodes_map);
2461 
2462 	/* make sure local cleanup occurs before the heartbeat events */
2463 	if (!test_bit(idx, dlm->recovery_map))
2464 		dlm_do_local_recovery_cleanup(dlm, idx);
2465 
2466 	/* notify anything attached to the heartbeat events */
2467 	dlm_hb_event_notify_attached(dlm, idx, 0);
2468 
2469 	mlog(0, "node %u being removed from domain map!\n", idx);
2470 	clear_bit(idx, dlm->domain_map);
2471 	clear_bit(idx, dlm->exit_domain_map);
2472 	/* wake up migration waiters if a node goes down.
2473 	 * perhaps later we can genericize this for other waiters. */
2474 	wake_up(&dlm->migration_wq);
2475 
2476 	set_bit(idx, dlm->recovery_map);
2477 }
2478 
dlm_hb_node_down_cb(struct o2nm_node * node,int idx,void * data)2479 void dlm_hb_node_down_cb(struct o2nm_node *node, int idx, void *data)
2480 {
2481 	struct dlm_ctxt *dlm = data;
2482 
2483 	if (!dlm_grab(dlm))
2484 		return;
2485 
2486 	/*
2487 	 * This will notify any dlm users that a node in our domain
2488 	 * went away without notifying us first.
2489 	 */
2490 	if (test_bit(idx, dlm->domain_map))
2491 		dlm_fire_domain_eviction_callbacks(dlm, idx);
2492 
2493 	spin_lock(&dlm->spinlock);
2494 	__dlm_hb_node_down(dlm, idx);
2495 	spin_unlock(&dlm->spinlock);
2496 
2497 	dlm_put(dlm);
2498 }
2499 
dlm_hb_node_up_cb(struct o2nm_node * node,int idx,void * data)2500 void dlm_hb_node_up_cb(struct o2nm_node *node, int idx, void *data)
2501 {
2502 	struct dlm_ctxt *dlm = data;
2503 
2504 	if (!dlm_grab(dlm))
2505 		return;
2506 
2507 	spin_lock(&dlm->spinlock);
2508 	set_bit(idx, dlm->live_nodes_map);
2509 	/* do NOT notify mle attached to the heartbeat events.
2510 	 * new nodes are not interesting in mastery until joined. */
2511 	spin_unlock(&dlm->spinlock);
2512 
2513 	dlm_put(dlm);
2514 }
2515 
dlm_reco_ast(void * astdata)2516 static void dlm_reco_ast(void *astdata)
2517 {
2518 	struct dlm_ctxt *dlm = astdata;
2519 	mlog(0, "ast for recovery lock fired!, this=%u, dlm=%s\n",
2520 	     dlm->node_num, dlm->name);
2521 }
dlm_reco_bast(void * astdata,int blocked_type)2522 static void dlm_reco_bast(void *astdata, int blocked_type)
2523 {
2524 	struct dlm_ctxt *dlm = astdata;
2525 	mlog(0, "bast for recovery lock fired!, this=%u, dlm=%s\n",
2526 	     dlm->node_num, dlm->name);
2527 }
dlm_reco_unlock_ast(void * astdata,enum dlm_status st)2528 static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st)
2529 {
2530 	mlog(0, "unlockast for recovery lock fired!\n");
2531 }
2532 
2533 /*
2534  * dlm_pick_recovery_master will continually attempt to use
2535  * dlmlock() on the special "$RECOVERY" lockres with the
2536  * LKM_NOQUEUE flag to get an EX.  every thread that enters
2537  * this function on each node racing to become the recovery
2538  * master will not stop attempting this until either:
2539  * a) this node gets the EX (and becomes the recovery master),
2540  * or b) dlm->reco.new_master gets set to some nodenum
2541  * != O2NM_INVALID_NODE_NUM (another node will do the reco).
2542  * so each time a recovery master is needed, the entire cluster
2543  * will sync at this point.  if the new master dies, that will
2544  * be detected in dlm_do_recovery */
dlm_pick_recovery_master(struct dlm_ctxt * dlm)2545 static int dlm_pick_recovery_master(struct dlm_ctxt *dlm)
2546 {
2547 	enum dlm_status ret;
2548 	struct dlm_lockstatus lksb;
2549 	int status = -EINVAL;
2550 
2551 	mlog(0, "starting recovery of %s at %lu, dead=%u, this=%u\n",
2552 	     dlm->name, jiffies, dlm->reco.dead_node, dlm->node_num);
2553 again:
2554 	memset(&lksb, 0, sizeof(lksb));
2555 
2556 	ret = dlmlock(dlm, LKM_EXMODE, &lksb, LKM_NOQUEUE|LKM_RECOVERY,
2557 		      DLM_RECOVERY_LOCK_NAME, DLM_RECOVERY_LOCK_NAME_LEN,
2558 		      dlm_reco_ast, dlm, dlm_reco_bast);
2559 
2560 	mlog(0, "%s: dlmlock($RECOVERY) returned %d, lksb=%d\n",
2561 	     dlm->name, ret, lksb.status);
2562 
2563 	if (ret == DLM_NORMAL) {
2564 		mlog(0, "dlm=%s dlmlock says I got it (this=%u)\n",
2565 		     dlm->name, dlm->node_num);
2566 
2567 		/* got the EX lock.  check to see if another node
2568 		 * just became the reco master */
2569 		if (dlm_reco_master_ready(dlm)) {
2570 			mlog(0, "%s: got reco EX lock, but %u will "
2571 			     "do the recovery\n", dlm->name,
2572 			     dlm->reco.new_master);
2573 			status = -EEXIST;
2574 		} else {
2575 			status = 0;
2576 
2577 			/* see if recovery was already finished elsewhere */
2578 			spin_lock(&dlm->spinlock);
2579 			if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
2580 				status = -EINVAL;
2581 				mlog(0, "%s: got reco EX lock, but "
2582 				     "node got recovered already\n", dlm->name);
2583 				if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) {
2584 					mlog(ML_ERROR, "%s: new master is %u "
2585 					     "but no dead node!\n",
2586 					     dlm->name, dlm->reco.new_master);
2587 					BUG();
2588 				}
2589 			}
2590 			spin_unlock(&dlm->spinlock);
2591 		}
2592 
2593 		/* if this node has actually become the recovery master,
2594 		 * set the master and send the messages to begin recovery */
2595 		if (!status) {
2596 			mlog(0, "%s: dead=%u, this=%u, sending "
2597 			     "begin_reco now\n", dlm->name,
2598 			     dlm->reco.dead_node, dlm->node_num);
2599 			status = dlm_send_begin_reco_message(dlm,
2600 				      dlm->reco.dead_node);
2601 			/* this always succeeds */
2602 			BUG_ON(status);
2603 
2604 			/* set the new_master to this node */
2605 			spin_lock(&dlm->spinlock);
2606 			dlm_set_reco_master(dlm, dlm->node_num);
2607 			spin_unlock(&dlm->spinlock);
2608 		}
2609 
2610 		/* recovery lock is a special case.  ast will not get fired,
2611 		 * so just go ahead and unlock it. */
2612 		ret = dlmunlock(dlm, &lksb, 0, dlm_reco_unlock_ast, dlm);
2613 		if (ret == DLM_DENIED) {
2614 			mlog(0, "got DLM_DENIED, trying LKM_CANCEL\n");
2615 			ret = dlmunlock(dlm, &lksb, LKM_CANCEL, dlm_reco_unlock_ast, dlm);
2616 		}
2617 		if (ret != DLM_NORMAL) {
2618 			/* this would really suck. this could only happen
2619 			 * if there was a network error during the unlock
2620 			 * because of node death.  this means the unlock
2621 			 * is actually "done" and the lock structure is
2622 			 * even freed.  we can continue, but only
2623 			 * because this specific lock name is special. */
2624 			mlog(ML_ERROR, "dlmunlock returned %d\n", ret);
2625 		}
2626 	} else if (ret == DLM_NOTQUEUED) {
2627 		mlog(0, "dlm=%s dlmlock says another node got it (this=%u)\n",
2628 		     dlm->name, dlm->node_num);
2629 		/* another node is master. wait on
2630 		 * reco.new_master != O2NM_INVALID_NODE_NUM
2631 		 * for at most one second */
2632 		wait_event_timeout(dlm->dlm_reco_thread_wq,
2633 					 dlm_reco_master_ready(dlm),
2634 					 msecs_to_jiffies(1000));
2635 		if (!dlm_reco_master_ready(dlm)) {
2636 			mlog(0, "%s: reco master taking awhile\n",
2637 			     dlm->name);
2638 			goto again;
2639 		}
2640 		/* another node has informed this one that it is reco master */
2641 		mlog(0, "%s: reco master %u is ready to recover %u\n",
2642 		     dlm->name, dlm->reco.new_master, dlm->reco.dead_node);
2643 		status = -EEXIST;
2644 	} else if (ret == DLM_RECOVERING) {
2645 		mlog(0, "dlm=%s dlmlock says master node died (this=%u)\n",
2646 		     dlm->name, dlm->node_num);
2647 		goto again;
2648 	} else {
2649 		struct dlm_lock_resource *res;
2650 
2651 		/* dlmlock returned something other than NOTQUEUED or NORMAL */
2652 		mlog(ML_ERROR, "%s: got %s from dlmlock($RECOVERY), "
2653 		     "lksb.status=%s\n", dlm->name, dlm_errname(ret),
2654 		     dlm_errname(lksb.status));
2655 		res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME,
2656 					 DLM_RECOVERY_LOCK_NAME_LEN);
2657 		if (res) {
2658 			dlm_print_one_lock_resource(res);
2659 			dlm_lockres_put(res);
2660 		} else {
2661 			mlog(ML_ERROR, "recovery lock not found\n");
2662 		}
2663 		BUG();
2664 	}
2665 
2666 	return status;
2667 }
2668 
dlm_send_begin_reco_message(struct dlm_ctxt * dlm,u8 dead_node)2669 static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node)
2670 {
2671 	struct dlm_begin_reco br;
2672 	int ret = 0;
2673 	struct dlm_node_iter iter;
2674 	int nodenum;
2675 	int status;
2676 
2677 	mlog(0, "%s: dead node is %u\n", dlm->name, dead_node);
2678 
2679 	spin_lock(&dlm->spinlock);
2680 	dlm_node_iter_init(dlm->domain_map, &iter);
2681 	spin_unlock(&dlm->spinlock);
2682 
2683 	clear_bit(dead_node, iter.node_map);
2684 
2685 	memset(&br, 0, sizeof(br));
2686 	br.node_idx = dlm->node_num;
2687 	br.dead_node = dead_node;
2688 
2689 	while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2690 		ret = 0;
2691 		if (nodenum == dead_node) {
2692 			mlog(0, "not sending begin reco to dead node "
2693 				  "%u\n", dead_node);
2694 			continue;
2695 		}
2696 		if (nodenum == dlm->node_num) {
2697 			mlog(0, "not sending begin reco to self\n");
2698 			continue;
2699 		}
2700 retry:
2701 		mlog(0, "attempting to send begin reco msg to %d\n",
2702 			  nodenum);
2703 		ret = o2net_send_message(DLM_BEGIN_RECO_MSG, dlm->key,
2704 					 &br, sizeof(br), nodenum, &status);
2705 		/* negative status is handled ok by caller here */
2706 		if (ret >= 0)
2707 			ret = status;
2708 		if (dlm_is_host_down(ret)) {
2709 			/* node is down.  not involved in recovery
2710 			 * so just keep going */
2711 			mlog(ML_NOTICE, "%s: node %u was down when sending "
2712 			     "begin reco msg (%d)\n", dlm->name, nodenum, ret);
2713 			ret = 0;
2714 		}
2715 
2716 		/*
2717 		 * Prior to commit aad1b15310b9bcd59fa81ab8f2b1513b59553ea8,
2718 		 * dlm_begin_reco_handler() returned EAGAIN and not -EAGAIN.
2719 		 * We are handling both for compatibility reasons.
2720 		 */
2721 		if (ret == -EAGAIN || ret == EAGAIN) {
2722 			mlog(0, "%s: trying to start recovery of node "
2723 			     "%u, but node %u is waiting for last recovery "
2724 			     "to complete, backoff for a bit\n", dlm->name,
2725 			     dead_node, nodenum);
2726 			msleep(100);
2727 			goto retry;
2728 		}
2729 		if (ret < 0) {
2730 			struct dlm_lock_resource *res;
2731 
2732 			/* this is now a serious problem, possibly ENOMEM
2733 			 * in the network stack.  must retry */
2734 			mlog_errno(ret);
2735 			mlog(ML_ERROR, "begin reco of dlm %s to node %u "
2736 			     "returned %d\n", dlm->name, nodenum, ret);
2737 			res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME,
2738 						 DLM_RECOVERY_LOCK_NAME_LEN);
2739 			if (res) {
2740 				dlm_print_one_lock_resource(res);
2741 				dlm_lockres_put(res);
2742 			} else {
2743 				mlog(ML_ERROR, "recovery lock not found\n");
2744 			}
2745 			/* sleep for a bit in hopes that we can avoid
2746 			 * another ENOMEM */
2747 			msleep(100);
2748 			goto retry;
2749 		}
2750 	}
2751 
2752 	return ret;
2753 }
2754 
dlm_begin_reco_handler(struct o2net_msg * msg,u32 len,void * data,void ** ret_data)2755 int dlm_begin_reco_handler(struct o2net_msg *msg, u32 len, void *data,
2756 			   void **ret_data)
2757 {
2758 	struct dlm_ctxt *dlm = data;
2759 	struct dlm_begin_reco *br = (struct dlm_begin_reco *)msg->buf;
2760 
2761 	/* ok to return 0, domain has gone away */
2762 	if (!dlm_grab(dlm))
2763 		return 0;
2764 
2765 	spin_lock(&dlm->spinlock);
2766 	if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2767 		mlog(0, "%s: node %u wants to recover node %u (%u:%u) "
2768 		     "but this node is in finalize state, waiting on finalize2\n",
2769 		     dlm->name, br->node_idx, br->dead_node,
2770 		     dlm->reco.dead_node, dlm->reco.new_master);
2771 		spin_unlock(&dlm->spinlock);
2772 		dlm_put(dlm);
2773 		return -EAGAIN;
2774 	}
2775 	spin_unlock(&dlm->spinlock);
2776 
2777 	mlog(0, "%s: node %u wants to recover node %u (%u:%u)\n",
2778 	     dlm->name, br->node_idx, br->dead_node,
2779 	     dlm->reco.dead_node, dlm->reco.new_master);
2780 
2781 	dlm_fire_domain_eviction_callbacks(dlm, br->dead_node);
2782 
2783 	spin_lock(&dlm->spinlock);
2784 	if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) {
2785 		if (test_bit(dlm->reco.new_master, dlm->recovery_map)) {
2786 			mlog(0, "%s: new_master %u died, changing "
2787 			     "to %u\n", dlm->name, dlm->reco.new_master,
2788 			     br->node_idx);
2789 		} else {
2790 			mlog(0, "%s: new_master %u NOT DEAD, changing "
2791 			     "to %u\n", dlm->name, dlm->reco.new_master,
2792 			     br->node_idx);
2793 			/* may not have seen the new master as dead yet */
2794 		}
2795 	}
2796 	if (dlm->reco.dead_node != O2NM_INVALID_NODE_NUM) {
2797 		mlog(ML_NOTICE, "%s: dead_node previously set to %u, "
2798 		     "node %u changing it to %u\n", dlm->name,
2799 		     dlm->reco.dead_node, br->node_idx, br->dead_node);
2800 	}
2801 	dlm_set_reco_master(dlm, br->node_idx);
2802 	dlm_set_reco_dead_node(dlm, br->dead_node);
2803 	if (!test_bit(br->dead_node, dlm->recovery_map)) {
2804 		mlog(0, "recovery master %u sees %u as dead, but this "
2805 		     "node has not yet.  marking %u as dead\n",
2806 		     br->node_idx, br->dead_node, br->dead_node);
2807 		if (!test_bit(br->dead_node, dlm->domain_map) ||
2808 		    !test_bit(br->dead_node, dlm->live_nodes_map))
2809 			mlog(0, "%u not in domain/live_nodes map "
2810 			     "so setting it in reco map manually\n",
2811 			     br->dead_node);
2812 		/* force the recovery cleanup in __dlm_hb_node_down
2813 		 * both of these will be cleared in a moment */
2814 		set_bit(br->dead_node, dlm->domain_map);
2815 		set_bit(br->dead_node, dlm->live_nodes_map);
2816 		__dlm_hb_node_down(dlm, br->dead_node);
2817 	}
2818 	spin_unlock(&dlm->spinlock);
2819 
2820 	dlm_kick_recovery_thread(dlm);
2821 
2822 	mlog(0, "%s: recovery started by node %u, for %u (%u:%u)\n",
2823 	     dlm->name, br->node_idx, br->dead_node,
2824 	     dlm->reco.dead_node, dlm->reco.new_master);
2825 
2826 	dlm_put(dlm);
2827 	return 0;
2828 }
2829 
2830 #define DLM_FINALIZE_STAGE2  0x01
dlm_send_finalize_reco_message(struct dlm_ctxt * dlm)2831 static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm)
2832 {
2833 	int ret = 0;
2834 	struct dlm_finalize_reco fr;
2835 	struct dlm_node_iter iter;
2836 	int nodenum;
2837 	int status;
2838 	int stage = 1;
2839 
2840 	mlog(0, "finishing recovery for node %s:%u, "
2841 	     "stage %d\n", dlm->name, dlm->reco.dead_node, stage);
2842 
2843 	spin_lock(&dlm->spinlock);
2844 	dlm_node_iter_init(dlm->domain_map, &iter);
2845 	spin_unlock(&dlm->spinlock);
2846 
2847 stage2:
2848 	memset(&fr, 0, sizeof(fr));
2849 	fr.node_idx = dlm->node_num;
2850 	fr.dead_node = dlm->reco.dead_node;
2851 	if (stage == 2)
2852 		fr.flags |= DLM_FINALIZE_STAGE2;
2853 
2854 	while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2855 		if (nodenum == dlm->node_num)
2856 			continue;
2857 		ret = o2net_send_message(DLM_FINALIZE_RECO_MSG, dlm->key,
2858 					 &fr, sizeof(fr), nodenum, &status);
2859 		if (ret >= 0)
2860 			ret = status;
2861 		if (ret < 0) {
2862 			mlog(ML_ERROR, "Error %d when sending message %u (key "
2863 			     "0x%x) to node %u\n", ret, DLM_FINALIZE_RECO_MSG,
2864 			     dlm->key, nodenum);
2865 			if (dlm_is_host_down(ret)) {
2866 				/* this has no effect on this recovery
2867 				 * session, so set the status to zero to
2868 				 * finish out the last recovery */
2869 				mlog(ML_ERROR, "node %u went down after this "
2870 				     "node finished recovery.\n", nodenum);
2871 				ret = 0;
2872 				continue;
2873 			}
2874 			break;
2875 		}
2876 	}
2877 	if (stage == 1) {
2878 		/* reset the node_iter back to the top and send finalize2 */
2879 		iter.curnode = -1;
2880 		stage = 2;
2881 		goto stage2;
2882 	}
2883 
2884 	return ret;
2885 }
2886 
dlm_finalize_reco_handler(struct o2net_msg * msg,u32 len,void * data,void ** ret_data)2887 int dlm_finalize_reco_handler(struct o2net_msg *msg, u32 len, void *data,
2888 			      void **ret_data)
2889 {
2890 	struct dlm_ctxt *dlm = data;
2891 	struct dlm_finalize_reco *fr = (struct dlm_finalize_reco *)msg->buf;
2892 	int stage = 1;
2893 
2894 	/* ok to return 0, domain has gone away */
2895 	if (!dlm_grab(dlm))
2896 		return 0;
2897 
2898 	if (fr->flags & DLM_FINALIZE_STAGE2)
2899 		stage = 2;
2900 
2901 	mlog(0, "%s: node %u finalizing recovery stage%d of "
2902 	     "node %u (%u:%u)\n", dlm->name, fr->node_idx, stage,
2903 	     fr->dead_node, dlm->reco.dead_node, dlm->reco.new_master);
2904 
2905 	spin_lock(&dlm->spinlock);
2906 
2907 	if (dlm->reco.new_master != fr->node_idx) {
2908 		mlog(ML_ERROR, "node %u sent recovery finalize msg, but node "
2909 		     "%u is supposed to be the new master, dead=%u\n",
2910 		     fr->node_idx, dlm->reco.new_master, fr->dead_node);
2911 		BUG();
2912 	}
2913 	if (dlm->reco.dead_node != fr->dead_node) {
2914 		mlog(ML_ERROR, "node %u sent recovery finalize msg for dead "
2915 		     "node %u, but node %u is supposed to be dead\n",
2916 		     fr->node_idx, fr->dead_node, dlm->reco.dead_node);
2917 		BUG();
2918 	}
2919 
2920 	switch (stage) {
2921 		case 1:
2922 			dlm_finish_local_lockres_recovery(dlm, fr->dead_node, fr->node_idx);
2923 			if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2924 				mlog(ML_ERROR, "%s: received finalize1 from "
2925 				     "new master %u for dead node %u, but "
2926 				     "this node has already received it!\n",
2927 				     dlm->name, fr->node_idx, fr->dead_node);
2928 				dlm_print_reco_node_status(dlm);
2929 				BUG();
2930 			}
2931 			dlm->reco.state |= DLM_RECO_STATE_FINALIZE;
2932 			spin_unlock(&dlm->spinlock);
2933 			break;
2934 		case 2:
2935 			if (!(dlm->reco.state & DLM_RECO_STATE_FINALIZE)) {
2936 				mlog(ML_ERROR, "%s: received finalize2 from "
2937 				     "new master %u for dead node %u, but "
2938 				     "this node did not have finalize1!\n",
2939 				     dlm->name, fr->node_idx, fr->dead_node);
2940 				dlm_print_reco_node_status(dlm);
2941 				BUG();
2942 			}
2943 			dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE;
2944 			__dlm_reset_recovery(dlm);
2945 			spin_unlock(&dlm->spinlock);
2946 			dlm_kick_recovery_thread(dlm);
2947 			break;
2948 	}
2949 
2950 	mlog(0, "%s: recovery done, reco master was %u, dead now %u, master now %u\n",
2951 	     dlm->name, fr->node_idx, dlm->reco.dead_node, dlm->reco.new_master);
2952 
2953 	dlm_put(dlm);
2954 	return 0;
2955 }
2956