xref: /openbmc/linux/fs/ocfs2/dlm/dlmmaster.c (revision b270f492)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * dlmmod.c
4  *
5  * standalone DLM module
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/spinlock.h>
23 #include <linux/delay.h>
24 
25 
26 #include "../cluster/heartbeat.h"
27 #include "../cluster/nodemanager.h"
28 #include "../cluster/tcp.h"
29 
30 #include "dlmapi.h"
31 #include "dlmcommon.h"
32 #include "dlmdomain.h"
33 #include "dlmdebug.h"
34 
35 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_MASTER)
36 #include "../cluster/masklog.h"
37 
38 static void dlm_mle_node_down(struct dlm_ctxt *dlm,
39 			      struct dlm_master_list_entry *mle,
40 			      struct o2nm_node *node,
41 			      int idx);
42 static void dlm_mle_node_up(struct dlm_ctxt *dlm,
43 			    struct dlm_master_list_entry *mle,
44 			    struct o2nm_node *node,
45 			    int idx);
46 
47 static void dlm_assert_master_worker(struct dlm_work_item *item, void *data);
48 static int dlm_do_assert_master(struct dlm_ctxt *dlm,
49 				struct dlm_lock_resource *res,
50 				void *nodemap, u32 flags);
51 static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data);
52 
dlm_mle_equal(struct dlm_ctxt * dlm,struct dlm_master_list_entry * mle,const char * name,unsigned int namelen)53 static inline int dlm_mle_equal(struct dlm_ctxt *dlm,
54 				struct dlm_master_list_entry *mle,
55 				const char *name,
56 				unsigned int namelen)
57 {
58 	if (dlm != mle->dlm)
59 		return 0;
60 
61 	if (namelen != mle->mnamelen ||
62 	    memcmp(name, mle->mname, namelen) != 0)
63 		return 0;
64 
65 	return 1;
66 }
67 
68 static struct kmem_cache *dlm_lockres_cache;
69 static struct kmem_cache *dlm_lockname_cache;
70 static struct kmem_cache *dlm_mle_cache;
71 
72 static void dlm_mle_release(struct kref *kref);
73 static void dlm_init_mle(struct dlm_master_list_entry *mle,
74 			enum dlm_mle_type type,
75 			struct dlm_ctxt *dlm,
76 			struct dlm_lock_resource *res,
77 			const char *name,
78 			unsigned int namelen);
79 static void dlm_put_mle(struct dlm_master_list_entry *mle);
80 static void __dlm_put_mle(struct dlm_master_list_entry *mle);
81 static int dlm_find_mle(struct dlm_ctxt *dlm,
82 			struct dlm_master_list_entry **mle,
83 			char *name, unsigned int namelen);
84 
85 static int dlm_do_master_request(struct dlm_lock_resource *res,
86 				 struct dlm_master_list_entry *mle, int to);
87 
88 
89 static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
90 				     struct dlm_lock_resource *res,
91 				     struct dlm_master_list_entry *mle,
92 				     int *blocked);
93 static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
94 				    struct dlm_lock_resource *res,
95 				    struct dlm_master_list_entry *mle,
96 				    int blocked);
97 static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
98 				 struct dlm_lock_resource *res,
99 				 struct dlm_master_list_entry *mle,
100 				 struct dlm_master_list_entry **oldmle,
101 				 const char *name, unsigned int namelen,
102 				 u8 new_master, u8 master);
103 
104 static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
105 				    struct dlm_lock_resource *res);
106 static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
107 				      struct dlm_lock_resource *res);
108 static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
109 				       struct dlm_lock_resource *res,
110 				       u8 target);
111 static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
112 				       struct dlm_lock_resource *res);
113 
114 
dlm_is_host_down(int errno)115 int dlm_is_host_down(int errno)
116 {
117 	switch (errno) {
118 		case -EBADF:
119 		case -ECONNREFUSED:
120 		case -ENOTCONN:
121 		case -ECONNRESET:
122 		case -EPIPE:
123 		case -EHOSTDOWN:
124 		case -EHOSTUNREACH:
125 		case -ETIMEDOUT:
126 		case -ECONNABORTED:
127 		case -ENETDOWN:
128 		case -ENETUNREACH:
129 		case -ENETRESET:
130 		case -ESHUTDOWN:
131 		case -ENOPROTOOPT:
132 		case -EINVAL:   /* if returned from our tcp code,
133 				   this means there is no socket */
134 			return 1;
135 	}
136 	return 0;
137 }
138 
139 
140 /*
141  * MASTER LIST FUNCTIONS
142  */
143 
144 
145 /*
146  * regarding master list entries and heartbeat callbacks:
147  *
148  * in order to avoid sleeping and allocation that occurs in
149  * heartbeat, master list entries are simply attached to the
150  * dlm's established heartbeat callbacks.  the mle is attached
151  * when it is created, and since the dlm->spinlock is held at
152  * that time, any heartbeat event will be properly discovered
153  * by the mle.  the mle needs to be detached from the
154  * dlm->mle_hb_events list as soon as heartbeat events are no
155  * longer useful to the mle, and before the mle is freed.
156  *
157  * as a general rule, heartbeat events are no longer needed by
158  * the mle once an "answer" regarding the lock master has been
159  * received.
160  */
__dlm_mle_attach_hb_events(struct dlm_ctxt * dlm,struct dlm_master_list_entry * mle)161 static inline void __dlm_mle_attach_hb_events(struct dlm_ctxt *dlm,
162 					      struct dlm_master_list_entry *mle)
163 {
164 	assert_spin_locked(&dlm->spinlock);
165 
166 	list_add_tail(&mle->hb_events, &dlm->mle_hb_events);
167 }
168 
169 
__dlm_mle_detach_hb_events(struct dlm_ctxt * dlm,struct dlm_master_list_entry * mle)170 static inline void __dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
171 					      struct dlm_master_list_entry *mle)
172 {
173 	if (!list_empty(&mle->hb_events))
174 		list_del_init(&mle->hb_events);
175 }
176 
177 
dlm_mle_detach_hb_events(struct dlm_ctxt * dlm,struct dlm_master_list_entry * mle)178 static inline void dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
179 					    struct dlm_master_list_entry *mle)
180 {
181 	spin_lock(&dlm->spinlock);
182 	__dlm_mle_detach_hb_events(dlm, mle);
183 	spin_unlock(&dlm->spinlock);
184 }
185 
dlm_get_mle_inuse(struct dlm_master_list_entry * mle)186 static void dlm_get_mle_inuse(struct dlm_master_list_entry *mle)
187 {
188 	struct dlm_ctxt *dlm;
189 	dlm = mle->dlm;
190 
191 	assert_spin_locked(&dlm->spinlock);
192 	assert_spin_locked(&dlm->master_lock);
193 	mle->inuse++;
194 	kref_get(&mle->mle_refs);
195 }
196 
dlm_put_mle_inuse(struct dlm_master_list_entry * mle)197 static void dlm_put_mle_inuse(struct dlm_master_list_entry *mle)
198 {
199 	struct dlm_ctxt *dlm;
200 	dlm = mle->dlm;
201 
202 	spin_lock(&dlm->spinlock);
203 	spin_lock(&dlm->master_lock);
204 	mle->inuse--;
205 	__dlm_put_mle(mle);
206 	spin_unlock(&dlm->master_lock);
207 	spin_unlock(&dlm->spinlock);
208 
209 }
210 
211 /* remove from list and free */
__dlm_put_mle(struct dlm_master_list_entry * mle)212 static void __dlm_put_mle(struct dlm_master_list_entry *mle)
213 {
214 	struct dlm_ctxt *dlm;
215 	dlm = mle->dlm;
216 
217 	assert_spin_locked(&dlm->spinlock);
218 	assert_spin_locked(&dlm->master_lock);
219 	if (!kref_read(&mle->mle_refs)) {
220 		/* this may or may not crash, but who cares.
221 		 * it's a BUG. */
222 		mlog(ML_ERROR, "bad mle: %p\n", mle);
223 		dlm_print_one_mle(mle);
224 		BUG();
225 	} else
226 		kref_put(&mle->mle_refs, dlm_mle_release);
227 }
228 
229 
230 /* must not have any spinlocks coming in */
dlm_put_mle(struct dlm_master_list_entry * mle)231 static void dlm_put_mle(struct dlm_master_list_entry *mle)
232 {
233 	struct dlm_ctxt *dlm;
234 	dlm = mle->dlm;
235 
236 	spin_lock(&dlm->spinlock);
237 	spin_lock(&dlm->master_lock);
238 	__dlm_put_mle(mle);
239 	spin_unlock(&dlm->master_lock);
240 	spin_unlock(&dlm->spinlock);
241 }
242 
dlm_get_mle(struct dlm_master_list_entry * mle)243 static inline void dlm_get_mle(struct dlm_master_list_entry *mle)
244 {
245 	kref_get(&mle->mle_refs);
246 }
247 
dlm_init_mle(struct dlm_master_list_entry * mle,enum dlm_mle_type type,struct dlm_ctxt * dlm,struct dlm_lock_resource * res,const char * name,unsigned int namelen)248 static void dlm_init_mle(struct dlm_master_list_entry *mle,
249 			enum dlm_mle_type type,
250 			struct dlm_ctxt *dlm,
251 			struct dlm_lock_resource *res,
252 			const char *name,
253 			unsigned int namelen)
254 {
255 	assert_spin_locked(&dlm->spinlock);
256 
257 	mle->dlm = dlm;
258 	mle->type = type;
259 	INIT_HLIST_NODE(&mle->master_hash_node);
260 	INIT_LIST_HEAD(&mle->hb_events);
261 	bitmap_zero(mle->maybe_map, O2NM_MAX_NODES);
262 	spin_lock_init(&mle->spinlock);
263 	init_waitqueue_head(&mle->wq);
264 	atomic_set(&mle->woken, 0);
265 	kref_init(&mle->mle_refs);
266 	bitmap_zero(mle->response_map, O2NM_MAX_NODES);
267 	mle->master = O2NM_MAX_NODES;
268 	mle->new_master = O2NM_MAX_NODES;
269 	mle->inuse = 0;
270 
271 	BUG_ON(mle->type != DLM_MLE_BLOCK &&
272 	       mle->type != DLM_MLE_MASTER &&
273 	       mle->type != DLM_MLE_MIGRATION);
274 
275 	if (mle->type == DLM_MLE_MASTER) {
276 		BUG_ON(!res);
277 		mle->mleres = res;
278 		memcpy(mle->mname, res->lockname.name, res->lockname.len);
279 		mle->mnamelen = res->lockname.len;
280 		mle->mnamehash = res->lockname.hash;
281 	} else {
282 		BUG_ON(!name);
283 		mle->mleres = NULL;
284 		memcpy(mle->mname, name, namelen);
285 		mle->mnamelen = namelen;
286 		mle->mnamehash = dlm_lockid_hash(name, namelen);
287 	}
288 
289 	atomic_inc(&dlm->mle_tot_count[mle->type]);
290 	atomic_inc(&dlm->mle_cur_count[mle->type]);
291 
292 	/* copy off the node_map and register hb callbacks on our copy */
293 	bitmap_copy(mle->node_map, dlm->domain_map, O2NM_MAX_NODES);
294 	bitmap_copy(mle->vote_map, dlm->domain_map, O2NM_MAX_NODES);
295 	clear_bit(dlm->node_num, mle->vote_map);
296 	clear_bit(dlm->node_num, mle->node_map);
297 
298 	/* attach the mle to the domain node up/down events */
299 	__dlm_mle_attach_hb_events(dlm, mle);
300 }
301 
__dlm_unlink_mle(struct dlm_ctxt * dlm,struct dlm_master_list_entry * mle)302 void __dlm_unlink_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle)
303 {
304 	assert_spin_locked(&dlm->spinlock);
305 	assert_spin_locked(&dlm->master_lock);
306 
307 	if (!hlist_unhashed(&mle->master_hash_node))
308 		hlist_del_init(&mle->master_hash_node);
309 }
310 
__dlm_insert_mle(struct dlm_ctxt * dlm,struct dlm_master_list_entry * mle)311 void __dlm_insert_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle)
312 {
313 	struct hlist_head *bucket;
314 
315 	assert_spin_locked(&dlm->master_lock);
316 
317 	bucket = dlm_master_hash(dlm, mle->mnamehash);
318 	hlist_add_head(&mle->master_hash_node, bucket);
319 }
320 
321 /* returns 1 if found, 0 if not */
dlm_find_mle(struct dlm_ctxt * dlm,struct dlm_master_list_entry ** mle,char * name,unsigned int namelen)322 static int dlm_find_mle(struct dlm_ctxt *dlm,
323 			struct dlm_master_list_entry **mle,
324 			char *name, unsigned int namelen)
325 {
326 	struct dlm_master_list_entry *tmpmle;
327 	struct hlist_head *bucket;
328 	unsigned int hash;
329 
330 	assert_spin_locked(&dlm->master_lock);
331 
332 	hash = dlm_lockid_hash(name, namelen);
333 	bucket = dlm_master_hash(dlm, hash);
334 	hlist_for_each_entry(tmpmle, bucket, master_hash_node) {
335 		if (!dlm_mle_equal(dlm, tmpmle, name, namelen))
336 			continue;
337 		dlm_get_mle(tmpmle);
338 		*mle = tmpmle;
339 		return 1;
340 	}
341 	return 0;
342 }
343 
dlm_hb_event_notify_attached(struct dlm_ctxt * dlm,int idx,int node_up)344 void dlm_hb_event_notify_attached(struct dlm_ctxt *dlm, int idx, int node_up)
345 {
346 	struct dlm_master_list_entry *mle;
347 
348 	assert_spin_locked(&dlm->spinlock);
349 
350 	list_for_each_entry(mle, &dlm->mle_hb_events, hb_events) {
351 		if (node_up)
352 			dlm_mle_node_up(dlm, mle, NULL, idx);
353 		else
354 			dlm_mle_node_down(dlm, mle, NULL, idx);
355 	}
356 }
357 
dlm_mle_node_down(struct dlm_ctxt * dlm,struct dlm_master_list_entry * mle,struct o2nm_node * node,int idx)358 static void dlm_mle_node_down(struct dlm_ctxt *dlm,
359 			      struct dlm_master_list_entry *mle,
360 			      struct o2nm_node *node, int idx)
361 {
362 	spin_lock(&mle->spinlock);
363 
364 	if (!test_bit(idx, mle->node_map))
365 		mlog(0, "node %u already removed from nodemap!\n", idx);
366 	else
367 		clear_bit(idx, mle->node_map);
368 
369 	spin_unlock(&mle->spinlock);
370 }
371 
dlm_mle_node_up(struct dlm_ctxt * dlm,struct dlm_master_list_entry * mle,struct o2nm_node * node,int idx)372 static void dlm_mle_node_up(struct dlm_ctxt *dlm,
373 			    struct dlm_master_list_entry *mle,
374 			    struct o2nm_node *node, int idx)
375 {
376 	spin_lock(&mle->spinlock);
377 
378 	if (test_bit(idx, mle->node_map))
379 		mlog(0, "node %u already in node map!\n", idx);
380 	else
381 		set_bit(idx, mle->node_map);
382 
383 	spin_unlock(&mle->spinlock);
384 }
385 
386 
dlm_init_mle_cache(void)387 int dlm_init_mle_cache(void)
388 {
389 	dlm_mle_cache = kmem_cache_create("o2dlm_mle",
390 					  sizeof(struct dlm_master_list_entry),
391 					  0, SLAB_HWCACHE_ALIGN,
392 					  NULL);
393 	if (dlm_mle_cache == NULL)
394 		return -ENOMEM;
395 	return 0;
396 }
397 
dlm_destroy_mle_cache(void)398 void dlm_destroy_mle_cache(void)
399 {
400 	kmem_cache_destroy(dlm_mle_cache);
401 }
402 
dlm_mle_release(struct kref * kref)403 static void dlm_mle_release(struct kref *kref)
404 {
405 	struct dlm_master_list_entry *mle;
406 	struct dlm_ctxt *dlm;
407 
408 	mle = container_of(kref, struct dlm_master_list_entry, mle_refs);
409 	dlm = mle->dlm;
410 
411 	assert_spin_locked(&dlm->spinlock);
412 	assert_spin_locked(&dlm->master_lock);
413 
414 	mlog(0, "Releasing mle for %.*s, type %d\n", mle->mnamelen, mle->mname,
415 	     mle->type);
416 
417 	/* remove from list if not already */
418 	__dlm_unlink_mle(dlm, mle);
419 
420 	/* detach the mle from the domain node up/down events */
421 	__dlm_mle_detach_hb_events(dlm, mle);
422 
423 	atomic_dec(&dlm->mle_cur_count[mle->type]);
424 
425 	/* NOTE: kfree under spinlock here.
426 	 * if this is bad, we can move this to a freelist. */
427 	kmem_cache_free(dlm_mle_cache, mle);
428 }
429 
430 
431 /*
432  * LOCK RESOURCE FUNCTIONS
433  */
434 
dlm_init_master_caches(void)435 int dlm_init_master_caches(void)
436 {
437 	dlm_lockres_cache = kmem_cache_create("o2dlm_lockres",
438 					      sizeof(struct dlm_lock_resource),
439 					      0, SLAB_HWCACHE_ALIGN, NULL);
440 	if (!dlm_lockres_cache)
441 		goto bail;
442 
443 	dlm_lockname_cache = kmem_cache_create("o2dlm_lockname",
444 					       DLM_LOCKID_NAME_MAX, 0,
445 					       SLAB_HWCACHE_ALIGN, NULL);
446 	if (!dlm_lockname_cache)
447 		goto bail;
448 
449 	return 0;
450 bail:
451 	dlm_destroy_master_caches();
452 	return -ENOMEM;
453 }
454 
dlm_destroy_master_caches(void)455 void dlm_destroy_master_caches(void)
456 {
457 	kmem_cache_destroy(dlm_lockname_cache);
458 	dlm_lockname_cache = NULL;
459 
460 	kmem_cache_destroy(dlm_lockres_cache);
461 	dlm_lockres_cache = NULL;
462 }
463 
dlm_lockres_release(struct kref * kref)464 static void dlm_lockres_release(struct kref *kref)
465 {
466 	struct dlm_lock_resource *res;
467 	struct dlm_ctxt *dlm;
468 
469 	res = container_of(kref, struct dlm_lock_resource, refs);
470 	dlm = res->dlm;
471 
472 	/* This should not happen -- all lockres' have a name
473 	 * associated with them at init time. */
474 	BUG_ON(!res->lockname.name);
475 
476 	mlog(0, "destroying lockres %.*s\n", res->lockname.len,
477 	     res->lockname.name);
478 
479 	atomic_dec(&dlm->res_cur_count);
480 
481 	if (!hlist_unhashed(&res->hash_node) ||
482 	    !list_empty(&res->granted) ||
483 	    !list_empty(&res->converting) ||
484 	    !list_empty(&res->blocked) ||
485 	    !list_empty(&res->dirty) ||
486 	    !list_empty(&res->recovering) ||
487 	    !list_empty(&res->purge)) {
488 		mlog(ML_ERROR,
489 		     "Going to BUG for resource %.*s."
490 		     "  We're on a list! [%c%c%c%c%c%c%c]\n",
491 		     res->lockname.len, res->lockname.name,
492 		     !hlist_unhashed(&res->hash_node) ? 'H' : ' ',
493 		     !list_empty(&res->granted) ? 'G' : ' ',
494 		     !list_empty(&res->converting) ? 'C' : ' ',
495 		     !list_empty(&res->blocked) ? 'B' : ' ',
496 		     !list_empty(&res->dirty) ? 'D' : ' ',
497 		     !list_empty(&res->recovering) ? 'R' : ' ',
498 		     !list_empty(&res->purge) ? 'P' : ' ');
499 
500 		dlm_print_one_lock_resource(res);
501 	}
502 
503 	/* By the time we're ready to blow this guy away, we shouldn't
504 	 * be on any lists. */
505 	BUG_ON(!hlist_unhashed(&res->hash_node));
506 	BUG_ON(!list_empty(&res->granted));
507 	BUG_ON(!list_empty(&res->converting));
508 	BUG_ON(!list_empty(&res->blocked));
509 	BUG_ON(!list_empty(&res->dirty));
510 	BUG_ON(!list_empty(&res->recovering));
511 	BUG_ON(!list_empty(&res->purge));
512 
513 	kmem_cache_free(dlm_lockname_cache, (void *)res->lockname.name);
514 
515 	kmem_cache_free(dlm_lockres_cache, res);
516 }
517 
dlm_lockres_put(struct dlm_lock_resource * res)518 void dlm_lockres_put(struct dlm_lock_resource *res)
519 {
520 	kref_put(&res->refs, dlm_lockres_release);
521 }
522 
dlm_init_lockres(struct dlm_ctxt * dlm,struct dlm_lock_resource * res,const char * name,unsigned int namelen)523 static void dlm_init_lockres(struct dlm_ctxt *dlm,
524 			     struct dlm_lock_resource *res,
525 			     const char *name, unsigned int namelen)
526 {
527 	char *qname;
528 
529 	/* If we memset here, we lose our reference to the kmalloc'd
530 	 * res->lockname.name, so be sure to init every field
531 	 * correctly! */
532 
533 	qname = (char *) res->lockname.name;
534 	memcpy(qname, name, namelen);
535 
536 	res->lockname.len = namelen;
537 	res->lockname.hash = dlm_lockid_hash(name, namelen);
538 
539 	init_waitqueue_head(&res->wq);
540 	spin_lock_init(&res->spinlock);
541 	INIT_HLIST_NODE(&res->hash_node);
542 	INIT_LIST_HEAD(&res->granted);
543 	INIT_LIST_HEAD(&res->converting);
544 	INIT_LIST_HEAD(&res->blocked);
545 	INIT_LIST_HEAD(&res->dirty);
546 	INIT_LIST_HEAD(&res->recovering);
547 	INIT_LIST_HEAD(&res->purge);
548 	INIT_LIST_HEAD(&res->tracking);
549 	atomic_set(&res->asts_reserved, 0);
550 	res->migration_pending = 0;
551 	res->inflight_locks = 0;
552 	res->inflight_assert_workers = 0;
553 
554 	res->dlm = dlm;
555 
556 	kref_init(&res->refs);
557 
558 	atomic_inc(&dlm->res_tot_count);
559 	atomic_inc(&dlm->res_cur_count);
560 
561 	/* just for consistency */
562 	spin_lock(&res->spinlock);
563 	dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
564 	spin_unlock(&res->spinlock);
565 
566 	res->state = DLM_LOCK_RES_IN_PROGRESS;
567 
568 	res->last_used = 0;
569 
570 	spin_lock(&dlm->track_lock);
571 	list_add_tail(&res->tracking, &dlm->tracking_list);
572 	spin_unlock(&dlm->track_lock);
573 
574 	memset(res->lvb, 0, DLM_LVB_LEN);
575 	bitmap_zero(res->refmap, O2NM_MAX_NODES);
576 }
577 
dlm_new_lockres(struct dlm_ctxt * dlm,const char * name,unsigned int namelen)578 struct dlm_lock_resource *dlm_new_lockres(struct dlm_ctxt *dlm,
579 				   const char *name,
580 				   unsigned int namelen)
581 {
582 	struct dlm_lock_resource *res = NULL;
583 
584 	res = kmem_cache_zalloc(dlm_lockres_cache, GFP_NOFS);
585 	if (!res)
586 		goto error;
587 
588 	res->lockname.name = kmem_cache_zalloc(dlm_lockname_cache, GFP_NOFS);
589 	if (!res->lockname.name)
590 		goto error;
591 
592 	dlm_init_lockres(dlm, res, name, namelen);
593 	return res;
594 
595 error:
596 	if (res)
597 		kmem_cache_free(dlm_lockres_cache, res);
598 	return NULL;
599 }
600 
dlm_lockres_set_refmap_bit(struct dlm_ctxt * dlm,struct dlm_lock_resource * res,int bit)601 void dlm_lockres_set_refmap_bit(struct dlm_ctxt *dlm,
602 				struct dlm_lock_resource *res, int bit)
603 {
604 	assert_spin_locked(&res->spinlock);
605 
606 	mlog(0, "res %.*s, set node %u, %ps()\n", res->lockname.len,
607 	     res->lockname.name, bit, __builtin_return_address(0));
608 
609 	set_bit(bit, res->refmap);
610 }
611 
dlm_lockres_clear_refmap_bit(struct dlm_ctxt * dlm,struct dlm_lock_resource * res,int bit)612 void dlm_lockres_clear_refmap_bit(struct dlm_ctxt *dlm,
613 				  struct dlm_lock_resource *res, int bit)
614 {
615 	assert_spin_locked(&res->spinlock);
616 
617 	mlog(0, "res %.*s, clr node %u, %ps()\n", res->lockname.len,
618 	     res->lockname.name, bit, __builtin_return_address(0));
619 
620 	clear_bit(bit, res->refmap);
621 }
622 
__dlm_lockres_grab_inflight_ref(struct dlm_ctxt * dlm,struct dlm_lock_resource * res)623 static void __dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm,
624 				   struct dlm_lock_resource *res)
625 {
626 	res->inflight_locks++;
627 
628 	mlog(0, "%s: res %.*s, inflight++: now %u, %ps()\n", dlm->name,
629 	     res->lockname.len, res->lockname.name, res->inflight_locks,
630 	     __builtin_return_address(0));
631 }
632 
dlm_lockres_grab_inflight_ref(struct dlm_ctxt * dlm,struct dlm_lock_resource * res)633 void dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm,
634 				   struct dlm_lock_resource *res)
635 {
636 	assert_spin_locked(&res->spinlock);
637 	__dlm_lockres_grab_inflight_ref(dlm, res);
638 }
639 
dlm_lockres_drop_inflight_ref(struct dlm_ctxt * dlm,struct dlm_lock_resource * res)640 void dlm_lockres_drop_inflight_ref(struct dlm_ctxt *dlm,
641 				   struct dlm_lock_resource *res)
642 {
643 	assert_spin_locked(&res->spinlock);
644 
645 	BUG_ON(res->inflight_locks == 0);
646 
647 	res->inflight_locks--;
648 
649 	mlog(0, "%s: res %.*s, inflight--: now %u, %ps()\n", dlm->name,
650 	     res->lockname.len, res->lockname.name, res->inflight_locks,
651 	     __builtin_return_address(0));
652 
653 	wake_up(&res->wq);
654 }
655 
__dlm_lockres_grab_inflight_worker(struct dlm_ctxt * dlm,struct dlm_lock_resource * res)656 void __dlm_lockres_grab_inflight_worker(struct dlm_ctxt *dlm,
657 		struct dlm_lock_resource *res)
658 {
659 	assert_spin_locked(&res->spinlock);
660 	res->inflight_assert_workers++;
661 	mlog(0, "%s:%.*s: inflight assert worker++: now %u\n",
662 			dlm->name, res->lockname.len, res->lockname.name,
663 			res->inflight_assert_workers);
664 }
665 
__dlm_lockres_drop_inflight_worker(struct dlm_ctxt * dlm,struct dlm_lock_resource * res)666 static void __dlm_lockres_drop_inflight_worker(struct dlm_ctxt *dlm,
667 		struct dlm_lock_resource *res)
668 {
669 	assert_spin_locked(&res->spinlock);
670 	BUG_ON(res->inflight_assert_workers == 0);
671 	res->inflight_assert_workers--;
672 	mlog(0, "%s:%.*s: inflight assert worker--: now %u\n",
673 			dlm->name, res->lockname.len, res->lockname.name,
674 			res->inflight_assert_workers);
675 }
676 
dlm_lockres_drop_inflight_worker(struct dlm_ctxt * dlm,struct dlm_lock_resource * res)677 static void dlm_lockres_drop_inflight_worker(struct dlm_ctxt *dlm,
678 		struct dlm_lock_resource *res)
679 {
680 	spin_lock(&res->spinlock);
681 	__dlm_lockres_drop_inflight_worker(dlm, res);
682 	spin_unlock(&res->spinlock);
683 }
684 
685 /*
686  * lookup a lock resource by name.
687  * may already exist in the hashtable.
688  * lockid is null terminated
689  *
690  * if not, allocate enough for the lockres and for
691  * the temporary structure used in doing the mastering.
692  *
693  * also, do a lookup in the dlm->master_list to see
694  * if another node has begun mastering the same lock.
695  * if so, there should be a block entry in there
696  * for this name, and we should *not* attempt to master
697  * the lock here.   need to wait around for that node
698  * to assert_master (or die).
699  *
700  */
dlm_get_lock_resource(struct dlm_ctxt * dlm,const char * lockid,int namelen,int flags)701 struct dlm_lock_resource * dlm_get_lock_resource(struct dlm_ctxt *dlm,
702 					  const char *lockid,
703 					  int namelen,
704 					  int flags)
705 {
706 	struct dlm_lock_resource *tmpres=NULL, *res=NULL;
707 	struct dlm_master_list_entry *mle = NULL;
708 	struct dlm_master_list_entry *alloc_mle = NULL;
709 	int blocked = 0;
710 	int ret, nodenum;
711 	struct dlm_node_iter iter;
712 	unsigned int hash;
713 	int tries = 0;
714 	int bit, wait_on_recovery = 0;
715 
716 	BUG_ON(!lockid);
717 
718 	hash = dlm_lockid_hash(lockid, namelen);
719 
720 	mlog(0, "get lockres %s (len %d)\n", lockid, namelen);
721 
722 lookup:
723 	spin_lock(&dlm->spinlock);
724 	tmpres = __dlm_lookup_lockres_full(dlm, lockid, namelen, hash);
725 	if (tmpres) {
726 		spin_unlock(&dlm->spinlock);
727 		spin_lock(&tmpres->spinlock);
728 
729 		/*
730 		 * Right after dlm spinlock was released, dlm_thread could have
731 		 * purged the lockres. Check if lockres got unhashed. If so
732 		 * start over.
733 		 */
734 		if (hlist_unhashed(&tmpres->hash_node)) {
735 			spin_unlock(&tmpres->spinlock);
736 			dlm_lockres_put(tmpres);
737 			tmpres = NULL;
738 			goto lookup;
739 		}
740 
741 		/* Wait on the thread that is mastering the resource */
742 		if (tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
743 			__dlm_wait_on_lockres(tmpres);
744 			BUG_ON(tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN);
745 			spin_unlock(&tmpres->spinlock);
746 			dlm_lockres_put(tmpres);
747 			tmpres = NULL;
748 			goto lookup;
749 		}
750 
751 		/* Wait on the resource purge to complete before continuing */
752 		if (tmpres->state & DLM_LOCK_RES_DROPPING_REF) {
753 			BUG_ON(tmpres->owner == dlm->node_num);
754 			__dlm_wait_on_lockres_flags(tmpres,
755 						    DLM_LOCK_RES_DROPPING_REF);
756 			spin_unlock(&tmpres->spinlock);
757 			dlm_lockres_put(tmpres);
758 			tmpres = NULL;
759 			goto lookup;
760 		}
761 
762 		/* Grab inflight ref to pin the resource */
763 		dlm_lockres_grab_inflight_ref(dlm, tmpres);
764 
765 		spin_unlock(&tmpres->spinlock);
766 		if (res) {
767 			spin_lock(&dlm->track_lock);
768 			if (!list_empty(&res->tracking))
769 				list_del_init(&res->tracking);
770 			else
771 				mlog(ML_ERROR, "Resource %.*s not "
772 						"on the Tracking list\n",
773 						res->lockname.len,
774 						res->lockname.name);
775 			spin_unlock(&dlm->track_lock);
776 			dlm_lockres_put(res);
777 		}
778 		res = tmpres;
779 		goto leave;
780 	}
781 
782 	if (!res) {
783 		spin_unlock(&dlm->spinlock);
784 		mlog(0, "allocating a new resource\n");
785 		/* nothing found and we need to allocate one. */
786 		alloc_mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
787 		if (!alloc_mle)
788 			goto leave;
789 		res = dlm_new_lockres(dlm, lockid, namelen);
790 		if (!res)
791 			goto leave;
792 		goto lookup;
793 	}
794 
795 	mlog(0, "no lockres found, allocated our own: %p\n", res);
796 
797 	if (flags & LKM_LOCAL) {
798 		/* caller knows it's safe to assume it's not mastered elsewhere
799 		 * DONE!  return right away */
800 		spin_lock(&res->spinlock);
801 		dlm_change_lockres_owner(dlm, res, dlm->node_num);
802 		__dlm_insert_lockres(dlm, res);
803 		dlm_lockres_grab_inflight_ref(dlm, res);
804 		spin_unlock(&res->spinlock);
805 		spin_unlock(&dlm->spinlock);
806 		/* lockres still marked IN_PROGRESS */
807 		goto wake_waiters;
808 	}
809 
810 	/* check master list to see if another node has started mastering it */
811 	spin_lock(&dlm->master_lock);
812 
813 	/* if we found a block, wait for lock to be mastered by another node */
814 	blocked = dlm_find_mle(dlm, &mle, (char *)lockid, namelen);
815 	if (blocked) {
816 		int mig;
817 		if (mle->type == DLM_MLE_MASTER) {
818 			mlog(ML_ERROR, "master entry for nonexistent lock!\n");
819 			BUG();
820 		}
821 		mig = (mle->type == DLM_MLE_MIGRATION);
822 		/* if there is a migration in progress, let the migration
823 		 * finish before continuing.  we can wait for the absence
824 		 * of the MIGRATION mle: either the migrate finished or
825 		 * one of the nodes died and the mle was cleaned up.
826 		 * if there is a BLOCK here, but it already has a master
827 		 * set, we are too late.  the master does not have a ref
828 		 * for us in the refmap.  detach the mle and drop it.
829 		 * either way, go back to the top and start over. */
830 		if (mig || mle->master != O2NM_MAX_NODES) {
831 			BUG_ON(mig && mle->master == dlm->node_num);
832 			/* we arrived too late.  the master does not
833 			 * have a ref for us. retry. */
834 			mlog(0, "%s:%.*s: late on %s\n",
835 			     dlm->name, namelen, lockid,
836 			     mig ?  "MIGRATION" : "BLOCK");
837 			spin_unlock(&dlm->master_lock);
838 			spin_unlock(&dlm->spinlock);
839 
840 			/* master is known, detach */
841 			if (!mig)
842 				dlm_mle_detach_hb_events(dlm, mle);
843 			dlm_put_mle(mle);
844 			mle = NULL;
845 			/* this is lame, but we can't wait on either
846 			 * the mle or lockres waitqueue here */
847 			if (mig)
848 				msleep(100);
849 			goto lookup;
850 		}
851 	} else {
852 		/* go ahead and try to master lock on this node */
853 		mle = alloc_mle;
854 		/* make sure this does not get freed below */
855 		alloc_mle = NULL;
856 		dlm_init_mle(mle, DLM_MLE_MASTER, dlm, res, NULL, 0);
857 		set_bit(dlm->node_num, mle->maybe_map);
858 		__dlm_insert_mle(dlm, mle);
859 
860 		/* still holding the dlm spinlock, check the recovery map
861 		 * to see if there are any nodes that still need to be
862 		 * considered.  these will not appear in the mle nodemap
863 		 * but they might own this lockres.  wait on them. */
864 		bit = find_first_bit(dlm->recovery_map, O2NM_MAX_NODES);
865 		if (bit < O2NM_MAX_NODES) {
866 			mlog(0, "%s: res %.*s, At least one node (%d) "
867 			     "to recover before lock mastery can begin\n",
868 			     dlm->name, namelen, (char *)lockid, bit);
869 			wait_on_recovery = 1;
870 		}
871 	}
872 
873 	/* at this point there is either a DLM_MLE_BLOCK or a
874 	 * DLM_MLE_MASTER on the master list, so it's safe to add the
875 	 * lockres to the hashtable.  anyone who finds the lock will
876 	 * still have to wait on the IN_PROGRESS. */
877 
878 	/* finally add the lockres to its hash bucket */
879 	__dlm_insert_lockres(dlm, res);
880 
881 	/* since this lockres is new it doesn't not require the spinlock */
882 	__dlm_lockres_grab_inflight_ref(dlm, res);
883 
884 	/* get an extra ref on the mle in case this is a BLOCK
885 	 * if so, the creator of the BLOCK may try to put the last
886 	 * ref at this time in the assert master handler, so we
887 	 * need an extra one to keep from a bad ptr deref. */
888 	dlm_get_mle_inuse(mle);
889 	spin_unlock(&dlm->master_lock);
890 	spin_unlock(&dlm->spinlock);
891 
892 redo_request:
893 	while (wait_on_recovery) {
894 		/* any cluster changes that occurred after dropping the
895 		 * dlm spinlock would be detectable be a change on the mle,
896 		 * so we only need to clear out the recovery map once. */
897 		if (dlm_is_recovery_lock(lockid, namelen)) {
898 			mlog(0, "%s: Recovery map is not empty, but must "
899 			     "master $RECOVERY lock now\n", dlm->name);
900 			if (!dlm_pre_master_reco_lockres(dlm, res))
901 				wait_on_recovery = 0;
902 			else {
903 				mlog(0, "%s: waiting 500ms for heartbeat state "
904 				    "change\n", dlm->name);
905 				msleep(500);
906 			}
907 			continue;
908 		}
909 
910 		dlm_kick_recovery_thread(dlm);
911 		msleep(1000);
912 		dlm_wait_for_recovery(dlm);
913 
914 		spin_lock(&dlm->spinlock);
915 		bit = find_first_bit(dlm->recovery_map, O2NM_MAX_NODES);
916 		if (bit < O2NM_MAX_NODES) {
917 			mlog(0, "%s: res %.*s, At least one node (%d) "
918 			     "to recover before lock mastery can begin\n",
919 			     dlm->name, namelen, (char *)lockid, bit);
920 			wait_on_recovery = 1;
921 		} else
922 			wait_on_recovery = 0;
923 		spin_unlock(&dlm->spinlock);
924 
925 		if (wait_on_recovery)
926 			dlm_wait_for_node_recovery(dlm, bit, 10000);
927 	}
928 
929 	/* must wait for lock to be mastered elsewhere */
930 	if (blocked)
931 		goto wait;
932 
933 	ret = -EINVAL;
934 	dlm_node_iter_init(mle->vote_map, &iter);
935 	while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
936 		ret = dlm_do_master_request(res, mle, nodenum);
937 		if (ret < 0)
938 			mlog_errno(ret);
939 		if (mle->master != O2NM_MAX_NODES) {
940 			/* found a master ! */
941 			if (mle->master <= nodenum)
942 				break;
943 			/* if our master request has not reached the master
944 			 * yet, keep going until it does.  this is how the
945 			 * master will know that asserts are needed back to
946 			 * the lower nodes. */
947 			mlog(0, "%s: res %.*s, Requests only up to %u but "
948 			     "master is %u, keep going\n", dlm->name, namelen,
949 			     lockid, nodenum, mle->master);
950 		}
951 	}
952 
953 wait:
954 	/* keep going until the response map includes all nodes */
955 	ret = dlm_wait_for_lock_mastery(dlm, res, mle, &blocked);
956 	if (ret < 0) {
957 		wait_on_recovery = 1;
958 		mlog(0, "%s: res %.*s, Node map changed, redo the master "
959 		     "request now, blocked=%d\n", dlm->name, res->lockname.len,
960 		     res->lockname.name, blocked);
961 		if (++tries > 20) {
962 			mlog(ML_ERROR, "%s: res %.*s, Spinning on "
963 			     "dlm_wait_for_lock_mastery, blocked = %d\n",
964 			     dlm->name, res->lockname.len,
965 			     res->lockname.name, blocked);
966 			dlm_print_one_lock_resource(res);
967 			dlm_print_one_mle(mle);
968 			tries = 0;
969 		}
970 		goto redo_request;
971 	}
972 
973 	mlog(0, "%s: res %.*s, Mastered by %u\n", dlm->name, res->lockname.len,
974 	     res->lockname.name, res->owner);
975 	/* make sure we never continue without this */
976 	BUG_ON(res->owner == O2NM_MAX_NODES);
977 
978 	/* master is known, detach if not already detached */
979 	dlm_mle_detach_hb_events(dlm, mle);
980 	dlm_put_mle(mle);
981 	/* put the extra ref */
982 	dlm_put_mle_inuse(mle);
983 
984 wake_waiters:
985 	spin_lock(&res->spinlock);
986 	res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
987 	spin_unlock(&res->spinlock);
988 	wake_up(&res->wq);
989 
990 leave:
991 	/* need to free the unused mle */
992 	if (alloc_mle)
993 		kmem_cache_free(dlm_mle_cache, alloc_mle);
994 
995 	return res;
996 }
997 
998 
999 #define DLM_MASTERY_TIMEOUT_MS   5000
1000 
dlm_wait_for_lock_mastery(struct dlm_ctxt * dlm,struct dlm_lock_resource * res,struct dlm_master_list_entry * mle,int * blocked)1001 static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
1002 				     struct dlm_lock_resource *res,
1003 				     struct dlm_master_list_entry *mle,
1004 				     int *blocked)
1005 {
1006 	u8 m;
1007 	int ret, bit;
1008 	int map_changed, voting_done;
1009 	int assert, sleep;
1010 
1011 recheck:
1012 	ret = 0;
1013 	assert = 0;
1014 
1015 	/* check if another node has already become the owner */
1016 	spin_lock(&res->spinlock);
1017 	if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1018 		mlog(0, "%s:%.*s: owner is suddenly %u\n", dlm->name,
1019 		     res->lockname.len, res->lockname.name, res->owner);
1020 		spin_unlock(&res->spinlock);
1021 		/* this will cause the master to re-assert across
1022 		 * the whole cluster, freeing up mles */
1023 		if (res->owner != dlm->node_num) {
1024 			ret = dlm_do_master_request(res, mle, res->owner);
1025 			if (ret < 0) {
1026 				/* give recovery a chance to run */
1027 				mlog(ML_ERROR, "link to %u went down?: %d\n", res->owner, ret);
1028 				msleep(500);
1029 				goto recheck;
1030 			}
1031 		}
1032 		ret = 0;
1033 		goto leave;
1034 	}
1035 	spin_unlock(&res->spinlock);
1036 
1037 	spin_lock(&mle->spinlock);
1038 	m = mle->master;
1039 	map_changed = !bitmap_equal(mle->vote_map, mle->node_map,
1040 				    O2NM_MAX_NODES);
1041 	voting_done = bitmap_equal(mle->vote_map, mle->response_map,
1042 				   O2NM_MAX_NODES);
1043 
1044 	/* restart if we hit any errors */
1045 	if (map_changed) {
1046 		int b;
1047 		mlog(0, "%s: %.*s: node map changed, restarting\n",
1048 		     dlm->name, res->lockname.len, res->lockname.name);
1049 		ret = dlm_restart_lock_mastery(dlm, res, mle, *blocked);
1050 		b = (mle->type == DLM_MLE_BLOCK);
1051 		if ((*blocked && !b) || (!*blocked && b)) {
1052 			mlog(0, "%s:%.*s: status change: old=%d new=%d\n",
1053 			     dlm->name, res->lockname.len, res->lockname.name,
1054 			     *blocked, b);
1055 			*blocked = b;
1056 		}
1057 		spin_unlock(&mle->spinlock);
1058 		if (ret < 0) {
1059 			mlog_errno(ret);
1060 			goto leave;
1061 		}
1062 		mlog(0, "%s:%.*s: restart lock mastery succeeded, "
1063 		     "rechecking now\n", dlm->name, res->lockname.len,
1064 		     res->lockname.name);
1065 		goto recheck;
1066 	} else {
1067 		if (!voting_done) {
1068 			mlog(0, "map not changed and voting not done "
1069 			     "for %s:%.*s\n", dlm->name, res->lockname.len,
1070 			     res->lockname.name);
1071 		}
1072 	}
1073 
1074 	if (m != O2NM_MAX_NODES) {
1075 		/* another node has done an assert!
1076 		 * all done! */
1077 		sleep = 0;
1078 	} else {
1079 		sleep = 1;
1080 		/* have all nodes responded? */
1081 		if (voting_done && !*blocked) {
1082 			bit = find_first_bit(mle->maybe_map, O2NM_MAX_NODES);
1083 			if (dlm->node_num <= bit) {
1084 				/* my node number is lowest.
1085 			 	 * now tell other nodes that I am
1086 				 * mastering this. */
1087 				mle->master = dlm->node_num;
1088 				/* ref was grabbed in get_lock_resource
1089 				 * will be dropped in dlmlock_master */
1090 				assert = 1;
1091 				sleep = 0;
1092 			}
1093 			/* if voting is done, but we have not received
1094 			 * an assert master yet, we must sleep */
1095 		}
1096 	}
1097 
1098 	spin_unlock(&mle->spinlock);
1099 
1100 	/* sleep if we haven't finished voting yet */
1101 	if (sleep) {
1102 		unsigned long timeo = msecs_to_jiffies(DLM_MASTERY_TIMEOUT_MS);
1103 		atomic_set(&mle->woken, 0);
1104 		(void)wait_event_timeout(mle->wq,
1105 					 (atomic_read(&mle->woken) == 1),
1106 					 timeo);
1107 		if (res->owner == O2NM_MAX_NODES) {
1108 			mlog(0, "%s:%.*s: waiting again\n", dlm->name,
1109 			     res->lockname.len, res->lockname.name);
1110 			goto recheck;
1111 		}
1112 		mlog(0, "done waiting, master is %u\n", res->owner);
1113 		ret = 0;
1114 		goto leave;
1115 	}
1116 
1117 	ret = 0;   /* done */
1118 	if (assert) {
1119 		m = dlm->node_num;
1120 		mlog(0, "about to master %.*s here, this=%u\n",
1121 		     res->lockname.len, res->lockname.name, m);
1122 		ret = dlm_do_assert_master(dlm, res, mle->vote_map, 0);
1123 		if (ret) {
1124 			/* This is a failure in the network path,
1125 			 * not in the response to the assert_master
1126 			 * (any nonzero response is a BUG on this node).
1127 			 * Most likely a socket just got disconnected
1128 			 * due to node death. */
1129 			mlog_errno(ret);
1130 		}
1131 		/* no longer need to restart lock mastery.
1132 		 * all living nodes have been contacted. */
1133 		ret = 0;
1134 	}
1135 
1136 	/* set the lockres owner */
1137 	spin_lock(&res->spinlock);
1138 	/* mastery reference obtained either during
1139 	 * assert_master_handler or in get_lock_resource */
1140 	dlm_change_lockres_owner(dlm, res, m);
1141 	spin_unlock(&res->spinlock);
1142 
1143 leave:
1144 	return ret;
1145 }
1146 
1147 struct dlm_bitmap_diff_iter
1148 {
1149 	int curnode;
1150 	unsigned long *orig_bm;
1151 	unsigned long *cur_bm;
1152 	unsigned long diff_bm[BITS_TO_LONGS(O2NM_MAX_NODES)];
1153 };
1154 
1155 enum dlm_node_state_change
1156 {
1157 	NODE_DOWN = -1,
1158 	NODE_NO_CHANGE = 0,
1159 	NODE_UP
1160 };
1161 
dlm_bitmap_diff_iter_init(struct dlm_bitmap_diff_iter * iter,unsigned long * orig_bm,unsigned long * cur_bm)1162 static void dlm_bitmap_diff_iter_init(struct dlm_bitmap_diff_iter *iter,
1163 				      unsigned long *orig_bm,
1164 				      unsigned long *cur_bm)
1165 {
1166 	unsigned long p1, p2;
1167 	int i;
1168 
1169 	iter->curnode = -1;
1170 	iter->orig_bm = orig_bm;
1171 	iter->cur_bm = cur_bm;
1172 
1173 	for (i = 0; i < BITS_TO_LONGS(O2NM_MAX_NODES); i++) {
1174        		p1 = *(iter->orig_bm + i);
1175 	       	p2 = *(iter->cur_bm + i);
1176 		iter->diff_bm[i] = (p1 & ~p2) | (p2 & ~p1);
1177 	}
1178 }
1179 
dlm_bitmap_diff_iter_next(struct dlm_bitmap_diff_iter * iter,enum dlm_node_state_change * state)1180 static int dlm_bitmap_diff_iter_next(struct dlm_bitmap_diff_iter *iter,
1181 				     enum dlm_node_state_change *state)
1182 {
1183 	int bit;
1184 
1185 	if (iter->curnode >= O2NM_MAX_NODES)
1186 		return -ENOENT;
1187 
1188 	bit = find_next_bit(iter->diff_bm, O2NM_MAX_NODES,
1189 			    iter->curnode+1);
1190 	if (bit >= O2NM_MAX_NODES) {
1191 		iter->curnode = O2NM_MAX_NODES;
1192 		return -ENOENT;
1193 	}
1194 
1195 	/* if it was there in the original then this node died */
1196 	if (test_bit(bit, iter->orig_bm))
1197 		*state = NODE_DOWN;
1198 	else
1199 		*state = NODE_UP;
1200 
1201 	iter->curnode = bit;
1202 	return bit;
1203 }
1204 
1205 
dlm_restart_lock_mastery(struct dlm_ctxt * dlm,struct dlm_lock_resource * res,struct dlm_master_list_entry * mle,int blocked)1206 static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
1207 				    struct dlm_lock_resource *res,
1208 				    struct dlm_master_list_entry *mle,
1209 				    int blocked)
1210 {
1211 	struct dlm_bitmap_diff_iter bdi;
1212 	enum dlm_node_state_change sc;
1213 	int node;
1214 	int ret = 0;
1215 
1216 	mlog(0, "something happened such that the "
1217 	     "master process may need to be restarted!\n");
1218 
1219 	assert_spin_locked(&mle->spinlock);
1220 
1221 	dlm_bitmap_diff_iter_init(&bdi, mle->vote_map, mle->node_map);
1222 	node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1223 	while (node >= 0) {
1224 		if (sc == NODE_UP) {
1225 			/* a node came up.  clear any old vote from
1226 			 * the response map and set it in the vote map
1227 			 * then restart the mastery. */
1228 			mlog(ML_NOTICE, "node %d up while restarting\n", node);
1229 
1230 			/* redo the master request, but only for the new node */
1231 			mlog(0, "sending request to new node\n");
1232 			clear_bit(node, mle->response_map);
1233 			set_bit(node, mle->vote_map);
1234 		} else {
1235 			mlog(ML_ERROR, "node down! %d\n", node);
1236 			if (blocked) {
1237 				int lowest = find_first_bit(mle->maybe_map,
1238 						       O2NM_MAX_NODES);
1239 
1240 				/* act like it was never there */
1241 				clear_bit(node, mle->maybe_map);
1242 
1243 			       	if (node == lowest) {
1244 					mlog(0, "expected master %u died"
1245 					    " while this node was blocked "
1246 					    "waiting on it!\n", node);
1247 					lowest = find_next_bit(mle->maybe_map,
1248 						       	O2NM_MAX_NODES,
1249 						       	lowest+1);
1250 					if (lowest < O2NM_MAX_NODES) {
1251 						mlog(0, "%s:%.*s:still "
1252 						     "blocked. waiting on %u "
1253 						     "now\n", dlm->name,
1254 						     res->lockname.len,
1255 						     res->lockname.name,
1256 						     lowest);
1257 					} else {
1258 						/* mle is an MLE_BLOCK, but
1259 						 * there is now nothing left to
1260 						 * block on.  we need to return
1261 						 * all the way back out and try
1262 						 * again with an MLE_MASTER.
1263 						 * dlm_do_local_recovery_cleanup
1264 						 * has already run, so the mle
1265 						 * refcount is ok */
1266 						mlog(0, "%s:%.*s: no "
1267 						     "longer blocking. try to "
1268 						     "master this here\n",
1269 						     dlm->name,
1270 						     res->lockname.len,
1271 						     res->lockname.name);
1272 						mle->type = DLM_MLE_MASTER;
1273 						mle->mleres = res;
1274 					}
1275 				}
1276 			}
1277 
1278 			/* now blank out everything, as if we had never
1279 			 * contacted anyone */
1280 			bitmap_zero(mle->maybe_map, O2NM_MAX_NODES);
1281 			bitmap_zero(mle->response_map, O2NM_MAX_NODES);
1282 			/* reset the vote_map to the current node_map */
1283 			bitmap_copy(mle->vote_map, mle->node_map,
1284 				    O2NM_MAX_NODES);
1285 			/* put myself into the maybe map */
1286 			if (mle->type != DLM_MLE_BLOCK)
1287 				set_bit(dlm->node_num, mle->maybe_map);
1288 		}
1289 		ret = -EAGAIN;
1290 		node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1291 	}
1292 	return ret;
1293 }
1294 
1295 
1296 /*
1297  * DLM_MASTER_REQUEST_MSG
1298  *
1299  * returns: 0 on success,
1300  *          -errno on a network error
1301  *
1302  * on error, the caller should assume the target node is "dead"
1303  *
1304  */
1305 
dlm_do_master_request(struct dlm_lock_resource * res,struct dlm_master_list_entry * mle,int to)1306 static int dlm_do_master_request(struct dlm_lock_resource *res,
1307 				 struct dlm_master_list_entry *mle, int to)
1308 {
1309 	struct dlm_ctxt *dlm = mle->dlm;
1310 	struct dlm_master_request request;
1311 	int ret, response=0, resend;
1312 
1313 	memset(&request, 0, sizeof(request));
1314 	request.node_idx = dlm->node_num;
1315 
1316 	BUG_ON(mle->type == DLM_MLE_MIGRATION);
1317 
1318 	request.namelen = (u8)mle->mnamelen;
1319 	memcpy(request.name, mle->mname, request.namelen);
1320 
1321 again:
1322 	ret = o2net_send_message(DLM_MASTER_REQUEST_MSG, dlm->key, &request,
1323 				 sizeof(request), to, &response);
1324 	if (ret < 0)  {
1325 		if (ret == -ESRCH) {
1326 			/* should never happen */
1327 			mlog(ML_ERROR, "TCP stack not ready!\n");
1328 			BUG();
1329 		} else if (ret == -EINVAL) {
1330 			mlog(ML_ERROR, "bad args passed to o2net!\n");
1331 			BUG();
1332 		} else if (ret == -ENOMEM) {
1333 			mlog(ML_ERROR, "out of memory while trying to send "
1334 			     "network message!  retrying\n");
1335 			/* this is totally crude */
1336 			msleep(50);
1337 			goto again;
1338 		} else if (!dlm_is_host_down(ret)) {
1339 			/* not a network error. bad. */
1340 			mlog_errno(ret);
1341 			mlog(ML_ERROR, "unhandled error!");
1342 			BUG();
1343 		}
1344 		/* all other errors should be network errors,
1345 		 * and likely indicate node death */
1346 		mlog(ML_ERROR, "link to %d went down!\n", to);
1347 		goto out;
1348 	}
1349 
1350 	ret = 0;
1351 	resend = 0;
1352 	spin_lock(&mle->spinlock);
1353 	switch (response) {
1354 		case DLM_MASTER_RESP_YES:
1355 			set_bit(to, mle->response_map);
1356 			mlog(0, "node %u is the master, response=YES\n", to);
1357 			mlog(0, "%s:%.*s: master node %u now knows I have a "
1358 			     "reference\n", dlm->name, res->lockname.len,
1359 			     res->lockname.name, to);
1360 			mle->master = to;
1361 			break;
1362 		case DLM_MASTER_RESP_NO:
1363 			mlog(0, "node %u not master, response=NO\n", to);
1364 			set_bit(to, mle->response_map);
1365 			break;
1366 		case DLM_MASTER_RESP_MAYBE:
1367 			mlog(0, "node %u not master, response=MAYBE\n", to);
1368 			set_bit(to, mle->response_map);
1369 			set_bit(to, mle->maybe_map);
1370 			break;
1371 		case DLM_MASTER_RESP_ERROR:
1372 			mlog(0, "node %u hit an error, resending\n", to);
1373 			resend = 1;
1374 			response = 0;
1375 			break;
1376 		default:
1377 			mlog(ML_ERROR, "bad response! %u\n", response);
1378 			BUG();
1379 	}
1380 	spin_unlock(&mle->spinlock);
1381 	if (resend) {
1382 		/* this is also totally crude */
1383 		msleep(50);
1384 		goto again;
1385 	}
1386 
1387 out:
1388 	return ret;
1389 }
1390 
1391 /*
1392  * locks that can be taken here:
1393  * dlm->spinlock
1394  * res->spinlock
1395  * mle->spinlock
1396  * dlm->master_list
1397  *
1398  * if possible, TRIM THIS DOWN!!!
1399  */
dlm_master_request_handler(struct o2net_msg * msg,u32 len,void * data,void ** ret_data)1400 int dlm_master_request_handler(struct o2net_msg *msg, u32 len, void *data,
1401 			       void **ret_data)
1402 {
1403 	u8 response = DLM_MASTER_RESP_MAYBE;
1404 	struct dlm_ctxt *dlm = data;
1405 	struct dlm_lock_resource *res = NULL;
1406 	struct dlm_master_request *request = (struct dlm_master_request *) msg->buf;
1407 	struct dlm_master_list_entry *mle = NULL, *tmpmle = NULL;
1408 	char *name;
1409 	unsigned int namelen, hash;
1410 	int found, ret;
1411 	int set_maybe;
1412 	int dispatch_assert = 0;
1413 	int dispatched = 0;
1414 
1415 	if (!dlm_grab(dlm))
1416 		return DLM_MASTER_RESP_NO;
1417 
1418 	if (!dlm_domain_fully_joined(dlm)) {
1419 		response = DLM_MASTER_RESP_NO;
1420 		goto send_response;
1421 	}
1422 
1423 	name = request->name;
1424 	namelen = request->namelen;
1425 	hash = dlm_lockid_hash(name, namelen);
1426 
1427 	if (namelen > DLM_LOCKID_NAME_MAX) {
1428 		response = DLM_IVBUFLEN;
1429 		goto send_response;
1430 	}
1431 
1432 way_up_top:
1433 	spin_lock(&dlm->spinlock);
1434 	res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1435 	if (res) {
1436 		spin_unlock(&dlm->spinlock);
1437 
1438 		/* take care of the easy cases up front */
1439 		spin_lock(&res->spinlock);
1440 
1441 		/*
1442 		 * Right after dlm spinlock was released, dlm_thread could have
1443 		 * purged the lockres. Check if lockres got unhashed. If so
1444 		 * start over.
1445 		 */
1446 		if (hlist_unhashed(&res->hash_node)) {
1447 			spin_unlock(&res->spinlock);
1448 			dlm_lockres_put(res);
1449 			goto way_up_top;
1450 		}
1451 
1452 		if (res->state & (DLM_LOCK_RES_RECOVERING|
1453 				  DLM_LOCK_RES_MIGRATING)) {
1454 			spin_unlock(&res->spinlock);
1455 			mlog(0, "returning DLM_MASTER_RESP_ERROR since res is "
1456 			     "being recovered/migrated\n");
1457 			response = DLM_MASTER_RESP_ERROR;
1458 			if (mle)
1459 				kmem_cache_free(dlm_mle_cache, mle);
1460 			goto send_response;
1461 		}
1462 
1463 		if (res->owner == dlm->node_num) {
1464 			dlm_lockres_set_refmap_bit(dlm, res, request->node_idx);
1465 			spin_unlock(&res->spinlock);
1466 			response = DLM_MASTER_RESP_YES;
1467 			if (mle)
1468 				kmem_cache_free(dlm_mle_cache, mle);
1469 
1470 			/* this node is the owner.
1471 			 * there is some extra work that needs to
1472 			 * happen now.  the requesting node has
1473 			 * caused all nodes up to this one to
1474 			 * create mles.  this node now needs to
1475 			 * go back and clean those up. */
1476 			dispatch_assert = 1;
1477 			goto send_response;
1478 		} else if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1479 			spin_unlock(&res->spinlock);
1480 			// mlog(0, "node %u is the master\n", res->owner);
1481 			response = DLM_MASTER_RESP_NO;
1482 			if (mle)
1483 				kmem_cache_free(dlm_mle_cache, mle);
1484 			goto send_response;
1485 		}
1486 
1487 		/* ok, there is no owner.  either this node is
1488 		 * being blocked, or it is actively trying to
1489 		 * master this lock. */
1490 		if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1491 			mlog(ML_ERROR, "lock with no owner should be "
1492 			     "in-progress!\n");
1493 			BUG();
1494 		}
1495 
1496 		// mlog(0, "lockres is in progress...\n");
1497 		spin_lock(&dlm->master_lock);
1498 		found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1499 		if (!found) {
1500 			mlog(ML_ERROR, "no mle found for this lock!\n");
1501 			BUG();
1502 		}
1503 		set_maybe = 1;
1504 		spin_lock(&tmpmle->spinlock);
1505 		if (tmpmle->type == DLM_MLE_BLOCK) {
1506 			// mlog(0, "this node is waiting for "
1507 			// "lockres to be mastered\n");
1508 			response = DLM_MASTER_RESP_NO;
1509 		} else if (tmpmle->type == DLM_MLE_MIGRATION) {
1510 			mlog(0, "node %u is master, but trying to migrate to "
1511 			     "node %u.\n", tmpmle->master, tmpmle->new_master);
1512 			if (tmpmle->master == dlm->node_num) {
1513 				mlog(ML_ERROR, "no owner on lockres, but this "
1514 				     "node is trying to migrate it to %u?!\n",
1515 				     tmpmle->new_master);
1516 				BUG();
1517 			} else {
1518 				/* the real master can respond on its own */
1519 				response = DLM_MASTER_RESP_NO;
1520 			}
1521 		} else if (tmpmle->master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1522 			set_maybe = 0;
1523 			if (tmpmle->master == dlm->node_num) {
1524 				response = DLM_MASTER_RESP_YES;
1525 				/* this node will be the owner.
1526 				 * go back and clean the mles on any
1527 				 * other nodes */
1528 				dispatch_assert = 1;
1529 				dlm_lockres_set_refmap_bit(dlm, res,
1530 							   request->node_idx);
1531 			} else
1532 				response = DLM_MASTER_RESP_NO;
1533 		} else {
1534 			// mlog(0, "this node is attempting to "
1535 			// "master lockres\n");
1536 			response = DLM_MASTER_RESP_MAYBE;
1537 		}
1538 		if (set_maybe)
1539 			set_bit(request->node_idx, tmpmle->maybe_map);
1540 		spin_unlock(&tmpmle->spinlock);
1541 
1542 		spin_unlock(&dlm->master_lock);
1543 		spin_unlock(&res->spinlock);
1544 
1545 		/* keep the mle attached to heartbeat events */
1546 		dlm_put_mle(tmpmle);
1547 		if (mle)
1548 			kmem_cache_free(dlm_mle_cache, mle);
1549 		goto send_response;
1550 	}
1551 
1552 	/*
1553 	 * lockres doesn't exist on this node
1554 	 * if there is an MLE_BLOCK, return NO
1555 	 * if there is an MLE_MASTER, return MAYBE
1556 	 * otherwise, add an MLE_BLOCK, return NO
1557 	 */
1558 	spin_lock(&dlm->master_lock);
1559 	found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1560 	if (!found) {
1561 		/* this lockid has never been seen on this node yet */
1562 		// mlog(0, "no mle found\n");
1563 		if (!mle) {
1564 			spin_unlock(&dlm->master_lock);
1565 			spin_unlock(&dlm->spinlock);
1566 
1567 			mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
1568 			if (!mle) {
1569 				response = DLM_MASTER_RESP_ERROR;
1570 				mlog_errno(-ENOMEM);
1571 				goto send_response;
1572 			}
1573 			goto way_up_top;
1574 		}
1575 
1576 		// mlog(0, "this is second time thru, already allocated, "
1577 		// "add the block.\n");
1578 		dlm_init_mle(mle, DLM_MLE_BLOCK, dlm, NULL, name, namelen);
1579 		set_bit(request->node_idx, mle->maybe_map);
1580 		__dlm_insert_mle(dlm, mle);
1581 		response = DLM_MASTER_RESP_NO;
1582 	} else {
1583 		spin_lock(&tmpmle->spinlock);
1584 		if (tmpmle->master == dlm->node_num) {
1585 			mlog(ML_ERROR, "no lockres, but an mle with this node as master!\n");
1586 			BUG();
1587 		}
1588 		if (tmpmle->type == DLM_MLE_BLOCK)
1589 			response = DLM_MASTER_RESP_NO;
1590 		else if (tmpmle->type == DLM_MLE_MIGRATION) {
1591 			mlog(0, "migration mle was found (%u->%u)\n",
1592 			     tmpmle->master, tmpmle->new_master);
1593 			/* real master can respond on its own */
1594 			response = DLM_MASTER_RESP_NO;
1595 		} else
1596 			response = DLM_MASTER_RESP_MAYBE;
1597 		set_bit(request->node_idx, tmpmle->maybe_map);
1598 		spin_unlock(&tmpmle->spinlock);
1599 	}
1600 	spin_unlock(&dlm->master_lock);
1601 	spin_unlock(&dlm->spinlock);
1602 
1603 	if (found) {
1604 		/* keep the mle attached to heartbeat events */
1605 		dlm_put_mle(tmpmle);
1606 	}
1607 send_response:
1608 	/*
1609 	 * __dlm_lookup_lockres() grabbed a reference to this lockres.
1610 	 * The reference is released by dlm_assert_master_worker() under
1611 	 * the call to dlm_dispatch_assert_master().  If
1612 	 * dlm_assert_master_worker() isn't called, we drop it here.
1613 	 */
1614 	if (dispatch_assert) {
1615 		mlog(0, "%u is the owner of %.*s, cleaning everyone else\n",
1616 			     dlm->node_num, res->lockname.len, res->lockname.name);
1617 		spin_lock(&res->spinlock);
1618 		ret = dlm_dispatch_assert_master(dlm, res, 0, request->node_idx,
1619 						 DLM_ASSERT_MASTER_MLE_CLEANUP);
1620 		if (ret < 0) {
1621 			mlog(ML_ERROR, "failed to dispatch assert master work\n");
1622 			response = DLM_MASTER_RESP_ERROR;
1623 			spin_unlock(&res->spinlock);
1624 			dlm_lockres_put(res);
1625 		} else {
1626 			dispatched = 1;
1627 			__dlm_lockres_grab_inflight_worker(dlm, res);
1628 			spin_unlock(&res->spinlock);
1629 		}
1630 	} else {
1631 		if (res)
1632 			dlm_lockres_put(res);
1633 	}
1634 
1635 	if (!dispatched)
1636 		dlm_put(dlm);
1637 	return response;
1638 }
1639 
1640 /*
1641  * DLM_ASSERT_MASTER_MSG
1642  */
1643 
1644 
1645 /*
1646  * NOTE: this can be used for debugging
1647  * can periodically run all locks owned by this node
1648  * and re-assert across the cluster...
1649  */
dlm_do_assert_master(struct dlm_ctxt * dlm,struct dlm_lock_resource * res,void * nodemap,u32 flags)1650 static int dlm_do_assert_master(struct dlm_ctxt *dlm,
1651 				struct dlm_lock_resource *res,
1652 				void *nodemap, u32 flags)
1653 {
1654 	struct dlm_assert_master assert;
1655 	int to, tmpret;
1656 	struct dlm_node_iter iter;
1657 	int ret = 0;
1658 	int reassert;
1659 	const char *lockname = res->lockname.name;
1660 	unsigned int namelen = res->lockname.len;
1661 
1662 	BUG_ON(namelen > O2NM_MAX_NAME_LEN);
1663 
1664 	spin_lock(&res->spinlock);
1665 	res->state |= DLM_LOCK_RES_SETREF_INPROG;
1666 	spin_unlock(&res->spinlock);
1667 
1668 again:
1669 	reassert = 0;
1670 
1671 	/* note that if this nodemap is empty, it returns 0 */
1672 	dlm_node_iter_init(nodemap, &iter);
1673 	while ((to = dlm_node_iter_next(&iter)) >= 0) {
1674 		int r = 0;
1675 		struct dlm_master_list_entry *mle = NULL;
1676 
1677 		mlog(0, "sending assert master to %d (%.*s)\n", to,
1678 		     namelen, lockname);
1679 		memset(&assert, 0, sizeof(assert));
1680 		assert.node_idx = dlm->node_num;
1681 		assert.namelen = namelen;
1682 		memcpy(assert.name, lockname, namelen);
1683 		assert.flags = cpu_to_be32(flags);
1684 
1685 		tmpret = o2net_send_message(DLM_ASSERT_MASTER_MSG, dlm->key,
1686 					    &assert, sizeof(assert), to, &r);
1687 		if (tmpret < 0) {
1688 			mlog(ML_ERROR, "Error %d when sending message %u (key "
1689 			     "0x%x) to node %u\n", tmpret,
1690 			     DLM_ASSERT_MASTER_MSG, dlm->key, to);
1691 			if (!dlm_is_host_down(tmpret)) {
1692 				mlog(ML_ERROR, "unhandled error=%d!\n", tmpret);
1693 				BUG();
1694 			}
1695 			/* a node died.  finish out the rest of the nodes. */
1696 			mlog(0, "link to %d went down!\n", to);
1697 			/* any nonzero status return will do */
1698 			ret = tmpret;
1699 			r = 0;
1700 		} else if (r < 0) {
1701 			/* ok, something horribly messed.  kill thyself. */
1702 			mlog(ML_ERROR,"during assert master of %.*s to %u, "
1703 			     "got %d.\n", namelen, lockname, to, r);
1704 			spin_lock(&dlm->spinlock);
1705 			spin_lock(&dlm->master_lock);
1706 			if (dlm_find_mle(dlm, &mle, (char *)lockname,
1707 					 namelen)) {
1708 				dlm_print_one_mle(mle);
1709 				__dlm_put_mle(mle);
1710 			}
1711 			spin_unlock(&dlm->master_lock);
1712 			spin_unlock(&dlm->spinlock);
1713 			BUG();
1714 		}
1715 
1716 		if (r & DLM_ASSERT_RESPONSE_REASSERT &&
1717 		    !(r & DLM_ASSERT_RESPONSE_MASTERY_REF)) {
1718 				mlog(ML_ERROR, "%.*s: very strange, "
1719 				     "master MLE but no lockres on %u\n",
1720 				     namelen, lockname, to);
1721 		}
1722 
1723 		if (r & DLM_ASSERT_RESPONSE_REASSERT) {
1724 			mlog(0, "%.*s: node %u create mles on other "
1725 			     "nodes and requests a re-assert\n",
1726 			     namelen, lockname, to);
1727 			reassert = 1;
1728 		}
1729 		if (r & DLM_ASSERT_RESPONSE_MASTERY_REF) {
1730 			mlog(0, "%.*s: node %u has a reference to this "
1731 			     "lockres, set the bit in the refmap\n",
1732 			     namelen, lockname, to);
1733 			spin_lock(&res->spinlock);
1734 			dlm_lockres_set_refmap_bit(dlm, res, to);
1735 			spin_unlock(&res->spinlock);
1736 		}
1737 	}
1738 
1739 	if (reassert)
1740 		goto again;
1741 
1742 	spin_lock(&res->spinlock);
1743 	res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
1744 	spin_unlock(&res->spinlock);
1745 	wake_up(&res->wq);
1746 
1747 	return ret;
1748 }
1749 
1750 /*
1751  * locks that can be taken here:
1752  * dlm->spinlock
1753  * res->spinlock
1754  * mle->spinlock
1755  * dlm->master_list
1756  *
1757  * if possible, TRIM THIS DOWN!!!
1758  */
dlm_assert_master_handler(struct o2net_msg * msg,u32 len,void * data,void ** ret_data)1759 int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data,
1760 			      void **ret_data)
1761 {
1762 	struct dlm_ctxt *dlm = data;
1763 	struct dlm_master_list_entry *mle = NULL;
1764 	struct dlm_assert_master *assert = (struct dlm_assert_master *)msg->buf;
1765 	struct dlm_lock_resource *res = NULL;
1766 	char *name;
1767 	unsigned int namelen, hash;
1768 	u32 flags;
1769 	int master_request = 0, have_lockres_ref = 0;
1770 	int ret = 0;
1771 
1772 	if (!dlm_grab(dlm))
1773 		return 0;
1774 
1775 	name = assert->name;
1776 	namelen = assert->namelen;
1777 	hash = dlm_lockid_hash(name, namelen);
1778 	flags = be32_to_cpu(assert->flags);
1779 
1780 	if (namelen > DLM_LOCKID_NAME_MAX) {
1781 		mlog(ML_ERROR, "Invalid name length!");
1782 		goto done;
1783 	}
1784 
1785 	spin_lock(&dlm->spinlock);
1786 
1787 	if (flags)
1788 		mlog(0, "assert_master with flags: %u\n", flags);
1789 
1790 	/* find the MLE */
1791 	spin_lock(&dlm->master_lock);
1792 	if (!dlm_find_mle(dlm, &mle, name, namelen)) {
1793 		/* not an error, could be master just re-asserting */
1794 		mlog(0, "just got an assert_master from %u, but no "
1795 		     "MLE for it! (%.*s)\n", assert->node_idx,
1796 		     namelen, name);
1797 	} else {
1798 		int bit = find_first_bit(mle->maybe_map, O2NM_MAX_NODES);
1799 		if (bit >= O2NM_MAX_NODES) {
1800 			/* not necessarily an error, though less likely.
1801 			 * could be master just re-asserting. */
1802 			mlog(0, "no bits set in the maybe_map, but %u "
1803 			     "is asserting! (%.*s)\n", assert->node_idx,
1804 			     namelen, name);
1805 		} else if (bit != assert->node_idx) {
1806 			if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1807 				mlog(0, "master %u was found, %u should "
1808 				     "back off\n", assert->node_idx, bit);
1809 			} else {
1810 				/* with the fix for bug 569, a higher node
1811 				 * number winning the mastery will respond
1812 				 * YES to mastery requests, but this node
1813 				 * had no way of knowing.  let it pass. */
1814 				mlog(0, "%u is the lowest node, "
1815 				     "%u is asserting. (%.*s)  %u must "
1816 				     "have begun after %u won.\n", bit,
1817 				     assert->node_idx, namelen, name, bit,
1818 				     assert->node_idx);
1819 			}
1820 		}
1821 		if (mle->type == DLM_MLE_MIGRATION) {
1822 			if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1823 				mlog(0, "%s:%.*s: got cleanup assert"
1824 				     " from %u for migration\n",
1825 				     dlm->name, namelen, name,
1826 				     assert->node_idx);
1827 			} else if (!(flags & DLM_ASSERT_MASTER_FINISH_MIGRATION)) {
1828 				mlog(0, "%s:%.*s: got unrelated assert"
1829 				     " from %u for migration, ignoring\n",
1830 				     dlm->name, namelen, name,
1831 				     assert->node_idx);
1832 				__dlm_put_mle(mle);
1833 				spin_unlock(&dlm->master_lock);
1834 				spin_unlock(&dlm->spinlock);
1835 				goto done;
1836 			}
1837 		}
1838 	}
1839 	spin_unlock(&dlm->master_lock);
1840 
1841 	/* ok everything checks out with the MLE
1842 	 * now check to see if there is a lockres */
1843 	res = __dlm_lookup_lockres(dlm, name, namelen, hash);
1844 	if (res) {
1845 		spin_lock(&res->spinlock);
1846 		if (res->state & DLM_LOCK_RES_RECOVERING)  {
1847 			mlog(ML_ERROR, "%u asserting but %.*s is "
1848 			     "RECOVERING!\n", assert->node_idx, namelen, name);
1849 			goto kill;
1850 		}
1851 		if (!mle) {
1852 			if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN &&
1853 			    res->owner != assert->node_idx) {
1854 				mlog(ML_ERROR, "DIE! Mastery assert from %u, "
1855 				     "but current owner is %u! (%.*s)\n",
1856 				     assert->node_idx, res->owner, namelen,
1857 				     name);
1858 				__dlm_print_one_lock_resource(res);
1859 				BUG();
1860 			}
1861 		} else if (mle->type != DLM_MLE_MIGRATION) {
1862 			if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1863 				/* owner is just re-asserting */
1864 				if (res->owner == assert->node_idx) {
1865 					mlog(0, "owner %u re-asserting on "
1866 					     "lock %.*s\n", assert->node_idx,
1867 					     namelen, name);
1868 					goto ok;
1869 				}
1870 				mlog(ML_ERROR, "got assert_master from "
1871 				     "node %u, but %u is the owner! "
1872 				     "(%.*s)\n", assert->node_idx,
1873 				     res->owner, namelen, name);
1874 				goto kill;
1875 			}
1876 			if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1877 				mlog(ML_ERROR, "got assert from %u, but lock "
1878 				     "with no owner should be "
1879 				     "in-progress! (%.*s)\n",
1880 				     assert->node_idx,
1881 				     namelen, name);
1882 				goto kill;
1883 			}
1884 		} else /* mle->type == DLM_MLE_MIGRATION */ {
1885 			/* should only be getting an assert from new master */
1886 			if (assert->node_idx != mle->new_master) {
1887 				mlog(ML_ERROR, "got assert from %u, but "
1888 				     "new master is %u, and old master "
1889 				     "was %u (%.*s)\n",
1890 				     assert->node_idx, mle->new_master,
1891 				     mle->master, namelen, name);
1892 				goto kill;
1893 			}
1894 
1895 		}
1896 ok:
1897 		spin_unlock(&res->spinlock);
1898 	}
1899 
1900 	// mlog(0, "woo!  got an assert_master from node %u!\n",
1901 	// 	     assert->node_idx);
1902 	if (mle) {
1903 		int extra_ref = 0;
1904 		int nn = -1;
1905 		int rr, err = 0;
1906 
1907 		spin_lock(&mle->spinlock);
1908 		if (mle->type == DLM_MLE_BLOCK || mle->type == DLM_MLE_MIGRATION)
1909 			extra_ref = 1;
1910 		else {
1911 			/* MASTER mle: if any bits set in the response map
1912 			 * then the calling node needs to re-assert to clear
1913 			 * up nodes that this node contacted */
1914 			while ((nn = find_next_bit (mle->response_map, O2NM_MAX_NODES,
1915 						    nn+1)) < O2NM_MAX_NODES) {
1916 				if (nn != dlm->node_num && nn != assert->node_idx) {
1917 					master_request = 1;
1918 					break;
1919 				}
1920 			}
1921 		}
1922 		mle->master = assert->node_idx;
1923 		atomic_set(&mle->woken, 1);
1924 		wake_up(&mle->wq);
1925 		spin_unlock(&mle->spinlock);
1926 
1927 		if (res) {
1928 			int wake = 0;
1929 			spin_lock(&res->spinlock);
1930 			if (mle->type == DLM_MLE_MIGRATION) {
1931 				mlog(0, "finishing off migration of lockres %.*s, "
1932 			     		"from %u to %u\n",
1933 			       		res->lockname.len, res->lockname.name,
1934 			       		dlm->node_num, mle->new_master);
1935 				res->state &= ~DLM_LOCK_RES_MIGRATING;
1936 				wake = 1;
1937 				dlm_change_lockres_owner(dlm, res, mle->new_master);
1938 				BUG_ON(res->state & DLM_LOCK_RES_DIRTY);
1939 			} else {
1940 				dlm_change_lockres_owner(dlm, res, mle->master);
1941 			}
1942 			spin_unlock(&res->spinlock);
1943 			have_lockres_ref = 1;
1944 			if (wake)
1945 				wake_up(&res->wq);
1946 		}
1947 
1948 		/* master is known, detach if not already detached.
1949 		 * ensures that only one assert_master call will happen
1950 		 * on this mle. */
1951 		spin_lock(&dlm->master_lock);
1952 
1953 		rr = kref_read(&mle->mle_refs);
1954 		if (mle->inuse > 0) {
1955 			if (extra_ref && rr < 3)
1956 				err = 1;
1957 			else if (!extra_ref && rr < 2)
1958 				err = 1;
1959 		} else {
1960 			if (extra_ref && rr < 2)
1961 				err = 1;
1962 			else if (!extra_ref && rr < 1)
1963 				err = 1;
1964 		}
1965 		if (err) {
1966 			mlog(ML_ERROR, "%s:%.*s: got assert master from %u "
1967 			     "that will mess up this node, refs=%d, extra=%d, "
1968 			     "inuse=%d\n", dlm->name, namelen, name,
1969 			     assert->node_idx, rr, extra_ref, mle->inuse);
1970 			dlm_print_one_mle(mle);
1971 		}
1972 		__dlm_unlink_mle(dlm, mle);
1973 		__dlm_mle_detach_hb_events(dlm, mle);
1974 		__dlm_put_mle(mle);
1975 		if (extra_ref) {
1976 			/* the assert master message now balances the extra
1977 		 	 * ref given by the master / migration request message.
1978 		 	 * if this is the last put, it will be removed
1979 		 	 * from the list. */
1980 			__dlm_put_mle(mle);
1981 		}
1982 		spin_unlock(&dlm->master_lock);
1983 	} else if (res) {
1984 		if (res->owner != assert->node_idx) {
1985 			mlog(0, "assert_master from %u, but current "
1986 			     "owner is %u (%.*s), no mle\n", assert->node_idx,
1987 			     res->owner, namelen, name);
1988 		}
1989 	}
1990 	spin_unlock(&dlm->spinlock);
1991 
1992 done:
1993 	ret = 0;
1994 	if (res) {
1995 		spin_lock(&res->spinlock);
1996 		res->state |= DLM_LOCK_RES_SETREF_INPROG;
1997 		spin_unlock(&res->spinlock);
1998 		*ret_data = (void *)res;
1999 	}
2000 	dlm_put(dlm);
2001 	if (master_request) {
2002 		mlog(0, "need to tell master to reassert\n");
2003 		/* positive. negative would shoot down the node. */
2004 		ret |= DLM_ASSERT_RESPONSE_REASSERT;
2005 		if (!have_lockres_ref) {
2006 			mlog(ML_ERROR, "strange, got assert from %u, MASTER "
2007 			     "mle present here for %s:%.*s, but no lockres!\n",
2008 			     assert->node_idx, dlm->name, namelen, name);
2009 		}
2010 	}
2011 	if (have_lockres_ref) {
2012 		/* let the master know we have a reference to the lockres */
2013 		ret |= DLM_ASSERT_RESPONSE_MASTERY_REF;
2014 		mlog(0, "%s:%.*s: got assert from %u, need a ref\n",
2015 		     dlm->name, namelen, name, assert->node_idx);
2016 	}
2017 	return ret;
2018 
2019 kill:
2020 	/* kill the caller! */
2021 	mlog(ML_ERROR, "Bad message received from another node.  Dumping state "
2022 	     "and killing the other node now!  This node is OK and can continue.\n");
2023 	__dlm_print_one_lock_resource(res);
2024 	spin_unlock(&res->spinlock);
2025 	spin_lock(&dlm->master_lock);
2026 	if (mle)
2027 		__dlm_put_mle(mle);
2028 	spin_unlock(&dlm->master_lock);
2029 	spin_unlock(&dlm->spinlock);
2030 	*ret_data = (void *)res;
2031 	dlm_put(dlm);
2032 	return -EINVAL;
2033 }
2034 
dlm_assert_master_post_handler(int status,void * data,void * ret_data)2035 void dlm_assert_master_post_handler(int status, void *data, void *ret_data)
2036 {
2037 	struct dlm_lock_resource *res = (struct dlm_lock_resource *)ret_data;
2038 
2039 	if (ret_data) {
2040 		spin_lock(&res->spinlock);
2041 		res->state &= ~DLM_LOCK_RES_SETREF_INPROG;
2042 		spin_unlock(&res->spinlock);
2043 		wake_up(&res->wq);
2044 		dlm_lockres_put(res);
2045 	}
2046 	return;
2047 }
2048 
dlm_dispatch_assert_master(struct dlm_ctxt * dlm,struct dlm_lock_resource * res,int ignore_higher,u8 request_from,u32 flags)2049 int dlm_dispatch_assert_master(struct dlm_ctxt *dlm,
2050 			       struct dlm_lock_resource *res,
2051 			       int ignore_higher, u8 request_from, u32 flags)
2052 {
2053 	struct dlm_work_item *item;
2054 	item = kzalloc(sizeof(*item), GFP_ATOMIC);
2055 	if (!item)
2056 		return -ENOMEM;
2057 
2058 
2059 	/* queue up work for dlm_assert_master_worker */
2060 	dlm_init_work_item(dlm, item, dlm_assert_master_worker, NULL);
2061 	item->u.am.lockres = res; /* already have a ref */
2062 	/* can optionally ignore node numbers higher than this node */
2063 	item->u.am.ignore_higher = ignore_higher;
2064 	item->u.am.request_from = request_from;
2065 	item->u.am.flags = flags;
2066 
2067 	if (ignore_higher)
2068 		mlog(0, "IGNORE HIGHER: %.*s\n", res->lockname.len,
2069 		     res->lockname.name);
2070 
2071 	spin_lock(&dlm->work_lock);
2072 	list_add_tail(&item->list, &dlm->work_list);
2073 	spin_unlock(&dlm->work_lock);
2074 
2075 	queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2076 	return 0;
2077 }
2078 
dlm_assert_master_worker(struct dlm_work_item * item,void * data)2079 static void dlm_assert_master_worker(struct dlm_work_item *item, void *data)
2080 {
2081 	struct dlm_ctxt *dlm = data;
2082 	int ret = 0;
2083 	struct dlm_lock_resource *res;
2084 	unsigned long nodemap[BITS_TO_LONGS(O2NM_MAX_NODES)];
2085 	int ignore_higher;
2086 	int bit;
2087 	u8 request_from;
2088 	u32 flags;
2089 
2090 	dlm = item->dlm;
2091 	res = item->u.am.lockres;
2092 	ignore_higher = item->u.am.ignore_higher;
2093 	request_from = item->u.am.request_from;
2094 	flags = item->u.am.flags;
2095 
2096 	spin_lock(&dlm->spinlock);
2097 	bitmap_copy(nodemap, dlm->domain_map, O2NM_MAX_NODES);
2098 	spin_unlock(&dlm->spinlock);
2099 
2100 	clear_bit(dlm->node_num, nodemap);
2101 	if (ignore_higher) {
2102 		/* if is this just to clear up mles for nodes below
2103 		 * this node, do not send the message to the original
2104 		 * caller or any node number higher than this */
2105 		clear_bit(request_from, nodemap);
2106 		bit = dlm->node_num;
2107 		while (1) {
2108 			bit = find_next_bit(nodemap, O2NM_MAX_NODES,
2109 					    bit+1);
2110 		       	if (bit >= O2NM_MAX_NODES)
2111 				break;
2112 			clear_bit(bit, nodemap);
2113 		}
2114 	}
2115 
2116 	/*
2117 	 * If we're migrating this lock to someone else, we are no
2118 	 * longer allowed to assert out own mastery.  OTOH, we need to
2119 	 * prevent migration from starting while we're still asserting
2120 	 * our dominance.  The reserved ast delays migration.
2121 	 */
2122 	spin_lock(&res->spinlock);
2123 	if (res->state & DLM_LOCK_RES_MIGRATING) {
2124 		mlog(0, "Someone asked us to assert mastery, but we're "
2125 		     "in the middle of migration.  Skipping assert, "
2126 		     "the new master will handle that.\n");
2127 		spin_unlock(&res->spinlock);
2128 		goto put;
2129 	} else
2130 		__dlm_lockres_reserve_ast(res);
2131 	spin_unlock(&res->spinlock);
2132 
2133 	/* this call now finishes out the nodemap
2134 	 * even if one or more nodes die */
2135 	mlog(0, "worker about to master %.*s here, this=%u\n",
2136 		     res->lockname.len, res->lockname.name, dlm->node_num);
2137 	ret = dlm_do_assert_master(dlm, res, nodemap, flags);
2138 	if (ret < 0) {
2139 		/* no need to restart, we are done */
2140 		if (!dlm_is_host_down(ret))
2141 			mlog_errno(ret);
2142 	}
2143 
2144 	/* Ok, we've asserted ourselves.  Let's let migration start. */
2145 	dlm_lockres_release_ast(dlm, res);
2146 
2147 put:
2148 	dlm_lockres_drop_inflight_worker(dlm, res);
2149 
2150 	dlm_lockres_put(res);
2151 
2152 	mlog(0, "finished with dlm_assert_master_worker\n");
2153 }
2154 
2155 /* SPECIAL CASE for the $RECOVERY lock used by the recovery thread.
2156  * We cannot wait for node recovery to complete to begin mastering this
2157  * lockres because this lockres is used to kick off recovery! ;-)
2158  * So, do a pre-check on all living nodes to see if any of those nodes
2159  * think that $RECOVERY is currently mastered by a dead node.  If so,
2160  * we wait a short time to allow that node to get notified by its own
2161  * heartbeat stack, then check again.  All $RECOVERY lock resources
2162  * mastered by dead nodes are purged when the heartbeat callback is
2163  * fired, so we can know for sure that it is safe to continue once
2164  * the node returns a live node or no node.  */
dlm_pre_master_reco_lockres(struct dlm_ctxt * dlm,struct dlm_lock_resource * res)2165 static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
2166 				       struct dlm_lock_resource *res)
2167 {
2168 	struct dlm_node_iter iter;
2169 	int nodenum;
2170 	int ret = 0;
2171 	u8 master = DLM_LOCK_RES_OWNER_UNKNOWN;
2172 
2173 	spin_lock(&dlm->spinlock);
2174 	dlm_node_iter_init(dlm->domain_map, &iter);
2175 	spin_unlock(&dlm->spinlock);
2176 
2177 	while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2178 		/* do not send to self */
2179 		if (nodenum == dlm->node_num)
2180 			continue;
2181 		ret = dlm_do_master_requery(dlm, res, nodenum, &master);
2182 		if (ret < 0) {
2183 			mlog_errno(ret);
2184 			if (!dlm_is_host_down(ret))
2185 				BUG();
2186 			/* host is down, so answer for that node would be
2187 			 * DLM_LOCK_RES_OWNER_UNKNOWN.  continue. */
2188 			ret = 0;
2189 		}
2190 
2191 		if (master != DLM_LOCK_RES_OWNER_UNKNOWN) {
2192 			/* check to see if this master is in the recovery map */
2193 			spin_lock(&dlm->spinlock);
2194 			if (test_bit(master, dlm->recovery_map)) {
2195 				mlog(ML_NOTICE, "%s: node %u has not seen "
2196 				     "node %u go down yet, and thinks the "
2197 				     "dead node is mastering the recovery "
2198 				     "lock.  must wait.\n", dlm->name,
2199 				     nodenum, master);
2200 				ret = -EAGAIN;
2201 			}
2202 			spin_unlock(&dlm->spinlock);
2203 			mlog(0, "%s: reco lock master is %u\n", dlm->name,
2204 			     master);
2205 			break;
2206 		}
2207 	}
2208 	return ret;
2209 }
2210 
2211 /*
2212  * DLM_DEREF_LOCKRES_MSG
2213  */
2214 
dlm_drop_lockres_ref(struct dlm_ctxt * dlm,struct dlm_lock_resource * res)2215 int dlm_drop_lockres_ref(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2216 {
2217 	struct dlm_deref_lockres deref;
2218 	int ret = 0, r;
2219 	const char *lockname;
2220 	unsigned int namelen;
2221 
2222 	lockname = res->lockname.name;
2223 	namelen = res->lockname.len;
2224 	BUG_ON(namelen > O2NM_MAX_NAME_LEN);
2225 
2226 	memset(&deref, 0, sizeof(deref));
2227 	deref.node_idx = dlm->node_num;
2228 	deref.namelen = namelen;
2229 	memcpy(deref.name, lockname, namelen);
2230 
2231 	ret = o2net_send_message(DLM_DEREF_LOCKRES_MSG, dlm->key,
2232 				 &deref, sizeof(deref), res->owner, &r);
2233 	if (ret < 0)
2234 		mlog(ML_ERROR, "%s: res %.*s, error %d send DEREF to node %u\n",
2235 		     dlm->name, namelen, lockname, ret, res->owner);
2236 	else if (r < 0) {
2237 		/* BAD.  other node says I did not have a ref. */
2238 		mlog(ML_ERROR, "%s: res %.*s, DEREF to node %u got %d\n",
2239 		     dlm->name, namelen, lockname, res->owner, r);
2240 		dlm_print_one_lock_resource(res);
2241 		if (r == -ENOMEM)
2242 			BUG();
2243 	} else
2244 		ret = r;
2245 
2246 	return ret;
2247 }
2248 
dlm_deref_lockres_handler(struct o2net_msg * msg,u32 len,void * data,void ** ret_data)2249 int dlm_deref_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
2250 			      void **ret_data)
2251 {
2252 	struct dlm_ctxt *dlm = data;
2253 	struct dlm_deref_lockres *deref = (struct dlm_deref_lockres *)msg->buf;
2254 	struct dlm_lock_resource *res = NULL;
2255 	char *name;
2256 	unsigned int namelen;
2257 	int ret = -EINVAL;
2258 	u8 node;
2259 	unsigned int hash;
2260 	struct dlm_work_item *item;
2261 	int cleared = 0;
2262 	int dispatch = 0;
2263 
2264 	if (!dlm_grab(dlm))
2265 		return 0;
2266 
2267 	name = deref->name;
2268 	namelen = deref->namelen;
2269 	node = deref->node_idx;
2270 
2271 	if (namelen > DLM_LOCKID_NAME_MAX) {
2272 		mlog(ML_ERROR, "Invalid name length!");
2273 		goto done;
2274 	}
2275 	if (deref->node_idx >= O2NM_MAX_NODES) {
2276 		mlog(ML_ERROR, "Invalid node number: %u\n", node);
2277 		goto done;
2278 	}
2279 
2280 	hash = dlm_lockid_hash(name, namelen);
2281 
2282 	spin_lock(&dlm->spinlock);
2283 	res = __dlm_lookup_lockres_full(dlm, name, namelen, hash);
2284 	if (!res) {
2285 		spin_unlock(&dlm->spinlock);
2286 		mlog(ML_ERROR, "%s:%.*s: bad lockres name\n",
2287 		     dlm->name, namelen, name);
2288 		goto done;
2289 	}
2290 	spin_unlock(&dlm->spinlock);
2291 
2292 	spin_lock(&res->spinlock);
2293 	if (res->state & DLM_LOCK_RES_SETREF_INPROG)
2294 		dispatch = 1;
2295 	else {
2296 		BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2297 		if (test_bit(node, res->refmap)) {
2298 			dlm_lockres_clear_refmap_bit(dlm, res, node);
2299 			cleared = 1;
2300 		}
2301 	}
2302 	spin_unlock(&res->spinlock);
2303 
2304 	if (!dispatch) {
2305 		if (cleared)
2306 			dlm_lockres_calc_usage(dlm, res);
2307 		else {
2308 			mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2309 		     	"but it is already dropped!\n", dlm->name,
2310 		     	res->lockname.len, res->lockname.name, node);
2311 			dlm_print_one_lock_resource(res);
2312 		}
2313 		ret = DLM_DEREF_RESPONSE_DONE;
2314 		goto done;
2315 	}
2316 
2317 	item = kzalloc(sizeof(*item), GFP_NOFS);
2318 	if (!item) {
2319 		ret = -ENOMEM;
2320 		mlog_errno(ret);
2321 		goto done;
2322 	}
2323 
2324 	dlm_init_work_item(dlm, item, dlm_deref_lockres_worker, NULL);
2325 	item->u.dl.deref_res = res;
2326 	item->u.dl.deref_node = node;
2327 
2328 	spin_lock(&dlm->work_lock);
2329 	list_add_tail(&item->list, &dlm->work_list);
2330 	spin_unlock(&dlm->work_lock);
2331 
2332 	queue_work(dlm->dlm_worker, &dlm->dispatched_work);
2333 	return DLM_DEREF_RESPONSE_INPROG;
2334 
2335 done:
2336 	if (res)
2337 		dlm_lockres_put(res);
2338 	dlm_put(dlm);
2339 
2340 	return ret;
2341 }
2342 
dlm_deref_lockres_done_handler(struct o2net_msg * msg,u32 len,void * data,void ** ret_data)2343 int dlm_deref_lockres_done_handler(struct o2net_msg *msg, u32 len, void *data,
2344 			      void **ret_data)
2345 {
2346 	struct dlm_ctxt *dlm = data;
2347 	struct dlm_deref_lockres_done *deref
2348 			= (struct dlm_deref_lockres_done *)msg->buf;
2349 	struct dlm_lock_resource *res = NULL;
2350 	char *name;
2351 	unsigned int namelen;
2352 	int ret = -EINVAL;
2353 	u8 node;
2354 	unsigned int hash;
2355 
2356 	if (!dlm_grab(dlm))
2357 		return 0;
2358 
2359 	name = deref->name;
2360 	namelen = deref->namelen;
2361 	node = deref->node_idx;
2362 
2363 	if (namelen > DLM_LOCKID_NAME_MAX) {
2364 		mlog(ML_ERROR, "Invalid name length!");
2365 		goto done;
2366 	}
2367 	if (deref->node_idx >= O2NM_MAX_NODES) {
2368 		mlog(ML_ERROR, "Invalid node number: %u\n", node);
2369 		goto done;
2370 	}
2371 
2372 	hash = dlm_lockid_hash(name, namelen);
2373 
2374 	spin_lock(&dlm->spinlock);
2375 	res = __dlm_lookup_lockres_full(dlm, name, namelen, hash);
2376 	if (!res) {
2377 		spin_unlock(&dlm->spinlock);
2378 		mlog(ML_ERROR, "%s:%.*s: bad lockres name\n",
2379 		     dlm->name, namelen, name);
2380 		goto done;
2381 	}
2382 
2383 	spin_lock(&res->spinlock);
2384 	if (!(res->state & DLM_LOCK_RES_DROPPING_REF)) {
2385 		spin_unlock(&res->spinlock);
2386 		spin_unlock(&dlm->spinlock);
2387 		mlog(ML_NOTICE, "%s:%.*s: node %u sends deref done "
2388 			"but it is already derefed!\n", dlm->name,
2389 			res->lockname.len, res->lockname.name, node);
2390 		ret = 0;
2391 		goto done;
2392 	}
2393 
2394 	__dlm_do_purge_lockres(dlm, res);
2395 	spin_unlock(&res->spinlock);
2396 	wake_up(&res->wq);
2397 
2398 	spin_unlock(&dlm->spinlock);
2399 
2400 	ret = 0;
2401 done:
2402 	if (res)
2403 		dlm_lockres_put(res);
2404 	dlm_put(dlm);
2405 	return ret;
2406 }
2407 
dlm_drop_lockres_ref_done(struct dlm_ctxt * dlm,struct dlm_lock_resource * res,u8 node)2408 static void dlm_drop_lockres_ref_done(struct dlm_ctxt *dlm,
2409 		struct dlm_lock_resource *res, u8 node)
2410 {
2411 	struct dlm_deref_lockres_done deref;
2412 	int ret = 0, r;
2413 	const char *lockname;
2414 	unsigned int namelen;
2415 
2416 	lockname = res->lockname.name;
2417 	namelen = res->lockname.len;
2418 	BUG_ON(namelen > O2NM_MAX_NAME_LEN);
2419 
2420 	memset(&deref, 0, sizeof(deref));
2421 	deref.node_idx = dlm->node_num;
2422 	deref.namelen = namelen;
2423 	memcpy(deref.name, lockname, namelen);
2424 
2425 	ret = o2net_send_message(DLM_DEREF_LOCKRES_DONE, dlm->key,
2426 				 &deref, sizeof(deref), node, &r);
2427 	if (ret < 0) {
2428 		mlog(ML_ERROR, "%s: res %.*s, error %d send DEREF DONE "
2429 				" to node %u\n", dlm->name, namelen,
2430 				lockname, ret, node);
2431 	} else if (r < 0) {
2432 		/* ignore the error */
2433 		mlog(ML_ERROR, "%s: res %.*s, DEREF to node %u got %d\n",
2434 		     dlm->name, namelen, lockname, node, r);
2435 		dlm_print_one_lock_resource(res);
2436 	}
2437 }
2438 
dlm_deref_lockres_worker(struct dlm_work_item * item,void * data)2439 static void dlm_deref_lockres_worker(struct dlm_work_item *item, void *data)
2440 {
2441 	struct dlm_ctxt *dlm;
2442 	struct dlm_lock_resource *res;
2443 	u8 node;
2444 	u8 cleared = 0;
2445 
2446 	dlm = item->dlm;
2447 	res = item->u.dl.deref_res;
2448 	node = item->u.dl.deref_node;
2449 
2450 	spin_lock(&res->spinlock);
2451 	BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
2452 	__dlm_wait_on_lockres_flags(res, DLM_LOCK_RES_SETREF_INPROG);
2453 	if (test_bit(node, res->refmap)) {
2454 		dlm_lockres_clear_refmap_bit(dlm, res, node);
2455 		cleared = 1;
2456 	}
2457 	spin_unlock(&res->spinlock);
2458 
2459 	dlm_drop_lockres_ref_done(dlm, res, node);
2460 
2461 	if (cleared) {
2462 		mlog(0, "%s:%.*s node %u ref dropped in dispatch\n",
2463 		     dlm->name, res->lockname.len, res->lockname.name, node);
2464 		dlm_lockres_calc_usage(dlm, res);
2465 	} else {
2466 		mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
2467 		     "but it is already dropped!\n", dlm->name,
2468 		     res->lockname.len, res->lockname.name, node);
2469 		dlm_print_one_lock_resource(res);
2470 	}
2471 
2472 	dlm_lockres_put(res);
2473 }
2474 
2475 /*
2476  * A migratable resource is one that is :
2477  * 1. locally mastered, and,
2478  * 2. zero local locks, and,
2479  * 3. one or more non-local locks, or, one or more references
2480  * Returns 1 if yes, 0 if not.
2481  */
dlm_is_lockres_migratable(struct dlm_ctxt * dlm,struct dlm_lock_resource * res)2482 static int dlm_is_lockres_migratable(struct dlm_ctxt *dlm,
2483 				      struct dlm_lock_resource *res)
2484 {
2485 	enum dlm_lockres_list idx;
2486 	int nonlocal = 0, node_ref;
2487 	struct list_head *queue;
2488 	struct dlm_lock *lock;
2489 	u64 cookie;
2490 
2491 	assert_spin_locked(&res->spinlock);
2492 
2493 	/* delay migration when the lockres is in MIGRATING state */
2494 	if (res->state & DLM_LOCK_RES_MIGRATING)
2495 		return 0;
2496 
2497 	/* delay migration when the lockres is in RECOCERING state */
2498 	if (res->state & (DLM_LOCK_RES_RECOVERING|
2499 			DLM_LOCK_RES_RECOVERY_WAITING))
2500 		return 0;
2501 
2502 	if (res->owner != dlm->node_num)
2503 		return 0;
2504 
2505         for (idx = DLM_GRANTED_LIST; idx <= DLM_BLOCKED_LIST; idx++) {
2506 		queue = dlm_list_idx_to_ptr(res, idx);
2507 		list_for_each_entry(lock, queue, list) {
2508 			if (lock->ml.node != dlm->node_num) {
2509 				nonlocal++;
2510 				continue;
2511 			}
2512 			cookie = be64_to_cpu(lock->ml.cookie);
2513 			mlog(0, "%s: Not migratable res %.*s, lock %u:%llu on "
2514 			     "%s list\n", dlm->name, res->lockname.len,
2515 			     res->lockname.name,
2516 			     dlm_get_lock_cookie_node(cookie),
2517 			     dlm_get_lock_cookie_seq(cookie),
2518 			     dlm_list_in_text(idx));
2519 			return 0;
2520 		}
2521 	}
2522 
2523 	if (!nonlocal) {
2524 		node_ref = find_first_bit(res->refmap, O2NM_MAX_NODES);
2525 		if (node_ref >= O2NM_MAX_NODES)
2526 			return 0;
2527 	}
2528 
2529 	mlog(0, "%s: res %.*s, Migratable\n", dlm->name, res->lockname.len,
2530 	     res->lockname.name);
2531 
2532 	return 1;
2533 }
2534 
2535 /*
2536  * DLM_MIGRATE_LOCKRES
2537  */
2538 
2539 
dlm_migrate_lockres(struct dlm_ctxt * dlm,struct dlm_lock_resource * res,u8 target)2540 static int dlm_migrate_lockres(struct dlm_ctxt *dlm,
2541 			       struct dlm_lock_resource *res, u8 target)
2542 {
2543 	struct dlm_master_list_entry *mle = NULL;
2544 	struct dlm_master_list_entry *oldmle = NULL;
2545  	struct dlm_migratable_lockres *mres = NULL;
2546 	int ret = 0;
2547 	const char *name;
2548 	unsigned int namelen;
2549 	int mle_added = 0;
2550 	int wake = 0;
2551 
2552 	if (!dlm_grab(dlm))
2553 		return -EINVAL;
2554 
2555 	name = res->lockname.name;
2556 	namelen = res->lockname.len;
2557 
2558 	mlog(0, "%s: Migrating %.*s to node %u\n", dlm->name, namelen, name,
2559 	     target);
2560 
2561 	/* preallocate up front. if this fails, abort */
2562 	ret = -ENOMEM;
2563 	mres = (struct dlm_migratable_lockres *) __get_free_page(GFP_NOFS);
2564 	if (!mres) {
2565 		mlog_errno(ret);
2566 		goto leave;
2567 	}
2568 
2569 	mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
2570 	if (!mle) {
2571 		mlog_errno(ret);
2572 		goto leave;
2573 	}
2574 	ret = 0;
2575 
2576 	/*
2577 	 * clear any existing master requests and
2578 	 * add the migration mle to the list
2579 	 */
2580 	spin_lock(&dlm->spinlock);
2581 	spin_lock(&dlm->master_lock);
2582 	ret = dlm_add_migration_mle(dlm, res, mle, &oldmle, name,
2583 				    namelen, target, dlm->node_num);
2584 	/* get an extra reference on the mle.
2585 	 * otherwise the assert_master from the new
2586 	 * master will destroy this.
2587 	 */
2588 	if (ret != -EEXIST)
2589 		dlm_get_mle_inuse(mle);
2590 
2591 	spin_unlock(&dlm->master_lock);
2592 	spin_unlock(&dlm->spinlock);
2593 
2594 	if (ret == -EEXIST) {
2595 		mlog(0, "another process is already migrating it\n");
2596 		goto fail;
2597 	}
2598 	mle_added = 1;
2599 
2600 	/*
2601 	 * set the MIGRATING flag and flush asts
2602 	 * if we fail after this we need to re-dirty the lockres
2603 	 */
2604 	if (dlm_mark_lockres_migrating(dlm, res, target) < 0) {
2605 		mlog(ML_ERROR, "tried to migrate %.*s to %u, but "
2606 		     "the target went down.\n", res->lockname.len,
2607 		     res->lockname.name, target);
2608 		spin_lock(&res->spinlock);
2609 		res->state &= ~DLM_LOCK_RES_MIGRATING;
2610 		wake = 1;
2611 		spin_unlock(&res->spinlock);
2612 		ret = -EINVAL;
2613 	}
2614 
2615 fail:
2616 	if (ret != -EEXIST && oldmle) {
2617 		/* master is known, detach if not already detached */
2618 		dlm_mle_detach_hb_events(dlm, oldmle);
2619 		dlm_put_mle(oldmle);
2620 	}
2621 
2622 	if (ret < 0) {
2623 		if (mle_added) {
2624 			dlm_mle_detach_hb_events(dlm, mle);
2625 			dlm_put_mle(mle);
2626 			dlm_put_mle_inuse(mle);
2627 		} else if (mle) {
2628 			kmem_cache_free(dlm_mle_cache, mle);
2629 			mle = NULL;
2630 		}
2631 		goto leave;
2632 	}
2633 
2634 	/*
2635 	 * at this point, we have a migration target, an mle
2636 	 * in the master list, and the MIGRATING flag set on
2637 	 * the lockres
2638 	 */
2639 
2640 	/* now that remote nodes are spinning on the MIGRATING flag,
2641 	 * ensure that all assert_master work is flushed. */
2642 	flush_workqueue(dlm->dlm_worker);
2643 
2644 	/* notify new node and send all lock state */
2645 	/* call send_one_lockres with migration flag.
2646 	 * this serves as notice to the target node that a
2647 	 * migration is starting. */
2648 	ret = dlm_send_one_lockres(dlm, res, mres, target,
2649 				   DLM_MRES_MIGRATION);
2650 
2651 	if (ret < 0) {
2652 		mlog(0, "migration to node %u failed with %d\n",
2653 		     target, ret);
2654 		/* migration failed, detach and clean up mle */
2655 		dlm_mle_detach_hb_events(dlm, mle);
2656 		dlm_put_mle(mle);
2657 		dlm_put_mle_inuse(mle);
2658 		spin_lock(&res->spinlock);
2659 		res->state &= ~DLM_LOCK_RES_MIGRATING;
2660 		wake = 1;
2661 		spin_unlock(&res->spinlock);
2662 		if (dlm_is_host_down(ret))
2663 			dlm_wait_for_node_death(dlm, target,
2664 						DLM_NODE_DEATH_WAIT_MAX);
2665 		goto leave;
2666 	}
2667 
2668 	/* at this point, the target sends a message to all nodes,
2669 	 * (using dlm_do_migrate_request).  this node is skipped since
2670 	 * we had to put an mle in the list to begin the process.  this
2671 	 * node now waits for target to do an assert master.  this node
2672 	 * will be the last one notified, ensuring that the migration
2673 	 * is complete everywhere.  if the target dies while this is
2674 	 * going on, some nodes could potentially see the target as the
2675 	 * master, so it is important that my recovery finds the migration
2676 	 * mle and sets the master to UNKNOWN. */
2677 
2678 
2679 	/* wait for new node to assert master */
2680 	while (1) {
2681 		ret = wait_event_interruptible_timeout(mle->wq,
2682 					(atomic_read(&mle->woken) == 1),
2683 					msecs_to_jiffies(5000));
2684 
2685 		if (ret >= 0) {
2686 		       	if (atomic_read(&mle->woken) == 1 ||
2687 			    res->owner == target)
2688 				break;
2689 
2690 			mlog(0, "%s:%.*s: timed out during migration\n",
2691 			     dlm->name, res->lockname.len, res->lockname.name);
2692 			/* avoid hang during shutdown when migrating lockres
2693 			 * to a node which also goes down */
2694 			if (dlm_is_node_dead(dlm, target)) {
2695 				mlog(0, "%s:%.*s: expected migration "
2696 				     "target %u is no longer up, restarting\n",
2697 				     dlm->name, res->lockname.len,
2698 				     res->lockname.name, target);
2699 				ret = -EINVAL;
2700 				/* migration failed, detach and clean up mle */
2701 				dlm_mle_detach_hb_events(dlm, mle);
2702 				dlm_put_mle(mle);
2703 				dlm_put_mle_inuse(mle);
2704 				spin_lock(&res->spinlock);
2705 				res->state &= ~DLM_LOCK_RES_MIGRATING;
2706 				wake = 1;
2707 				spin_unlock(&res->spinlock);
2708 				goto leave;
2709 			}
2710 		} else
2711 			mlog(0, "%s:%.*s: caught signal during migration\n",
2712 			     dlm->name, res->lockname.len, res->lockname.name);
2713 	}
2714 
2715 	/* all done, set the owner, clear the flag */
2716 	spin_lock(&res->spinlock);
2717 	dlm_set_lockres_owner(dlm, res, target);
2718 	res->state &= ~DLM_LOCK_RES_MIGRATING;
2719 	dlm_remove_nonlocal_locks(dlm, res);
2720 	spin_unlock(&res->spinlock);
2721 	wake_up(&res->wq);
2722 
2723 	/* master is known, detach if not already detached */
2724 	dlm_mle_detach_hb_events(dlm, mle);
2725 	dlm_put_mle_inuse(mle);
2726 	ret = 0;
2727 
2728 	dlm_lockres_calc_usage(dlm, res);
2729 
2730 leave:
2731 	/* re-dirty the lockres if we failed */
2732 	if (ret < 0)
2733 		dlm_kick_thread(dlm, res);
2734 
2735 	/* wake up waiters if the MIGRATING flag got set
2736 	 * but migration failed */
2737 	if (wake)
2738 		wake_up(&res->wq);
2739 
2740 	if (mres)
2741 		free_page((unsigned long)mres);
2742 
2743 	dlm_put(dlm);
2744 
2745 	mlog(0, "%s: Migrating %.*s to %u, returns %d\n", dlm->name, namelen,
2746 	     name, target, ret);
2747 	return ret;
2748 }
2749 
2750 /*
2751  * Should be called only after beginning the domain leave process.
2752  * There should not be any remaining locks on nonlocal lock resources,
2753  * and there should be no local locks left on locally mastered resources.
2754  *
2755  * Called with the dlm spinlock held, may drop it to do migration, but
2756  * will re-acquire before exit.
2757  *
2758  * Returns: 1 if dlm->spinlock was dropped/retaken, 0 if never dropped
2759  */
dlm_empty_lockres(struct dlm_ctxt * dlm,struct dlm_lock_resource * res)2760 int dlm_empty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2761 	__must_hold(&dlm->spinlock)
2762 {
2763 	int ret;
2764 	int lock_dropped = 0;
2765 	u8 target = O2NM_MAX_NODES;
2766 
2767 	assert_spin_locked(&dlm->spinlock);
2768 
2769 	spin_lock(&res->spinlock);
2770 	if (dlm_is_lockres_migratable(dlm, res))
2771 		target = dlm_pick_migration_target(dlm, res);
2772 	spin_unlock(&res->spinlock);
2773 
2774 	if (target == O2NM_MAX_NODES)
2775 		goto leave;
2776 
2777 	/* Wheee! Migrate lockres here! Will sleep so drop spinlock. */
2778 	spin_unlock(&dlm->spinlock);
2779 	lock_dropped = 1;
2780 	ret = dlm_migrate_lockres(dlm, res, target);
2781 	if (ret)
2782 		mlog(0, "%s: res %.*s, Migrate to node %u failed with %d\n",
2783 		     dlm->name, res->lockname.len, res->lockname.name,
2784 		     target, ret);
2785 	spin_lock(&dlm->spinlock);
2786 leave:
2787 	return lock_dropped;
2788 }
2789 
dlm_lock_basts_flushed(struct dlm_ctxt * dlm,struct dlm_lock * lock)2790 int dlm_lock_basts_flushed(struct dlm_ctxt *dlm, struct dlm_lock *lock)
2791 {
2792 	int ret;
2793 	spin_lock(&dlm->ast_lock);
2794 	spin_lock(&lock->spinlock);
2795 	ret = (list_empty(&lock->bast_list) && !lock->bast_pending);
2796 	spin_unlock(&lock->spinlock);
2797 	spin_unlock(&dlm->ast_lock);
2798 	return ret;
2799 }
2800 
dlm_migration_can_proceed(struct dlm_ctxt * dlm,struct dlm_lock_resource * res,u8 mig_target)2801 static int dlm_migration_can_proceed(struct dlm_ctxt *dlm,
2802 				     struct dlm_lock_resource *res,
2803 				     u8 mig_target)
2804 {
2805 	int can_proceed;
2806 	spin_lock(&res->spinlock);
2807 	can_proceed = !!(res->state & DLM_LOCK_RES_MIGRATING);
2808 	spin_unlock(&res->spinlock);
2809 
2810 	/* target has died, so make the caller break out of the
2811 	 * wait_event, but caller must recheck the domain_map */
2812 	spin_lock(&dlm->spinlock);
2813 	if (!test_bit(mig_target, dlm->domain_map))
2814 		can_proceed = 1;
2815 	spin_unlock(&dlm->spinlock);
2816 	return can_proceed;
2817 }
2818 
dlm_lockres_is_dirty(struct dlm_ctxt * dlm,struct dlm_lock_resource * res)2819 static int dlm_lockres_is_dirty(struct dlm_ctxt *dlm,
2820 				struct dlm_lock_resource *res)
2821 {
2822 	int ret;
2823 	spin_lock(&res->spinlock);
2824 	ret = !!(res->state & DLM_LOCK_RES_DIRTY);
2825 	spin_unlock(&res->spinlock);
2826 	return ret;
2827 }
2828 
2829 
dlm_mark_lockres_migrating(struct dlm_ctxt * dlm,struct dlm_lock_resource * res,u8 target)2830 static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
2831 				       struct dlm_lock_resource *res,
2832 				       u8 target)
2833 {
2834 	int ret = 0;
2835 
2836 	mlog(0, "dlm_mark_lockres_migrating: %.*s, from %u to %u\n",
2837 	       res->lockname.len, res->lockname.name, dlm->node_num,
2838 	       target);
2839 	/* need to set MIGRATING flag on lockres.  this is done by
2840 	 * ensuring that all asts have been flushed for this lockres. */
2841 	spin_lock(&res->spinlock);
2842 	BUG_ON(res->migration_pending);
2843 	res->migration_pending = 1;
2844 	/* strategy is to reserve an extra ast then release
2845 	 * it below, letting the release do all of the work */
2846 	__dlm_lockres_reserve_ast(res);
2847 	spin_unlock(&res->spinlock);
2848 
2849 	/* now flush all the pending asts */
2850 	dlm_kick_thread(dlm, res);
2851 	/* before waiting on DIRTY, block processes which may
2852 	 * try to dirty the lockres before MIGRATING is set */
2853 	spin_lock(&res->spinlock);
2854 	BUG_ON(res->state & DLM_LOCK_RES_BLOCK_DIRTY);
2855 	res->state |= DLM_LOCK_RES_BLOCK_DIRTY;
2856 	spin_unlock(&res->spinlock);
2857 	/* now wait on any pending asts and the DIRTY state */
2858 	wait_event(dlm->ast_wq, !dlm_lockres_is_dirty(dlm, res));
2859 	dlm_lockres_release_ast(dlm, res);
2860 
2861 	mlog(0, "about to wait on migration_wq, dirty=%s\n",
2862 	       res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
2863 	/* if the extra ref we just put was the final one, this
2864 	 * will pass thru immediately.  otherwise, we need to wait
2865 	 * for the last ast to finish. */
2866 again:
2867 	ret = wait_event_interruptible_timeout(dlm->migration_wq,
2868 		   dlm_migration_can_proceed(dlm, res, target),
2869 		   msecs_to_jiffies(1000));
2870 	if (ret < 0) {
2871 		mlog(0, "woken again: migrating? %s, dead? %s\n",
2872 		       res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2873 		       test_bit(target, dlm->domain_map) ? "no":"yes");
2874 	} else {
2875 		mlog(0, "all is well: migrating? %s, dead? %s\n",
2876 		       res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2877 		       test_bit(target, dlm->domain_map) ? "no":"yes");
2878 	}
2879 	if (!dlm_migration_can_proceed(dlm, res, target)) {
2880 		mlog(0, "trying again...\n");
2881 		goto again;
2882 	}
2883 
2884 	ret = 0;
2885 	/* did the target go down or die? */
2886 	spin_lock(&dlm->spinlock);
2887 	if (!test_bit(target, dlm->domain_map)) {
2888 		mlog(ML_ERROR, "aha. migration target %u just went down\n",
2889 		     target);
2890 		ret = -EHOSTDOWN;
2891 	}
2892 	spin_unlock(&dlm->spinlock);
2893 
2894 	/*
2895 	 * if target is down, we need to clear DLM_LOCK_RES_BLOCK_DIRTY for
2896 	 * another try; otherwise, we are sure the MIGRATING state is there,
2897 	 * drop the unneeded state which blocked threads trying to DIRTY
2898 	 */
2899 	spin_lock(&res->spinlock);
2900 	BUG_ON(!(res->state & DLM_LOCK_RES_BLOCK_DIRTY));
2901 	res->state &= ~DLM_LOCK_RES_BLOCK_DIRTY;
2902 	if (!ret)
2903 		BUG_ON(!(res->state & DLM_LOCK_RES_MIGRATING));
2904 	else
2905 		res->migration_pending = 0;
2906 	spin_unlock(&res->spinlock);
2907 
2908 	/*
2909 	 * at this point:
2910 	 *
2911 	 *   o the DLM_LOCK_RES_MIGRATING flag is set if target not down
2912 	 *   o there are no pending asts on this lockres
2913 	 *   o all processes trying to reserve an ast on this
2914 	 *     lockres must wait for the MIGRATING flag to clear
2915 	 */
2916 	return ret;
2917 }
2918 
2919 /* last step in the migration process.
2920  * original master calls this to free all of the dlm_lock
2921  * structures that used to be for other nodes. */
dlm_remove_nonlocal_locks(struct dlm_ctxt * dlm,struct dlm_lock_resource * res)2922 static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
2923 				      struct dlm_lock_resource *res)
2924 {
2925 	struct list_head *queue = &res->granted;
2926 	int i, bit;
2927 	struct dlm_lock *lock, *next;
2928 
2929 	assert_spin_locked(&res->spinlock);
2930 
2931 	BUG_ON(res->owner == dlm->node_num);
2932 
2933 	for (i=0; i<3; i++) {
2934 		list_for_each_entry_safe(lock, next, queue, list) {
2935 			if (lock->ml.node != dlm->node_num) {
2936 				mlog(0, "putting lock for node %u\n",
2937 				     lock->ml.node);
2938 				/* be extra careful */
2939 				BUG_ON(!list_empty(&lock->ast_list));
2940 				BUG_ON(!list_empty(&lock->bast_list));
2941 				BUG_ON(lock->ast_pending);
2942 				BUG_ON(lock->bast_pending);
2943 				dlm_lockres_clear_refmap_bit(dlm, res,
2944 							     lock->ml.node);
2945 				list_del_init(&lock->list);
2946 				dlm_lock_put(lock);
2947 				/* In a normal unlock, we would have added a
2948 				 * DLM_UNLOCK_FREE_LOCK action. Force it. */
2949 				dlm_lock_put(lock);
2950 			}
2951 		}
2952 		queue++;
2953 	}
2954 	bit = 0;
2955 	while (1) {
2956 		bit = find_next_bit(res->refmap, O2NM_MAX_NODES, bit);
2957 		if (bit >= O2NM_MAX_NODES)
2958 			break;
2959 		/* do not clear the local node reference, if there is a
2960 		 * process holding this, let it drop the ref itself */
2961 		if (bit != dlm->node_num) {
2962 			mlog(0, "%s:%.*s: node %u had a ref to this "
2963 			     "migrating lockres, clearing\n", dlm->name,
2964 			     res->lockname.len, res->lockname.name, bit);
2965 			dlm_lockres_clear_refmap_bit(dlm, res, bit);
2966 		}
2967 		bit++;
2968 	}
2969 }
2970 
2971 /*
2972  * Pick a node to migrate the lock resource to. This function selects a
2973  * potential target based first on the locks and then on refmap. It skips
2974  * nodes that are in the process of exiting the domain.
2975  */
dlm_pick_migration_target(struct dlm_ctxt * dlm,struct dlm_lock_resource * res)2976 static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
2977 				    struct dlm_lock_resource *res)
2978 {
2979 	enum dlm_lockres_list idx;
2980 	struct list_head *queue;
2981 	struct dlm_lock *lock;
2982 	int noderef;
2983 	u8 nodenum = O2NM_MAX_NODES;
2984 
2985 	assert_spin_locked(&dlm->spinlock);
2986 	assert_spin_locked(&res->spinlock);
2987 
2988 	/* Go through all the locks */
2989 	for (idx = DLM_GRANTED_LIST; idx <= DLM_BLOCKED_LIST; idx++) {
2990 		queue = dlm_list_idx_to_ptr(res, idx);
2991 		list_for_each_entry(lock, queue, list) {
2992 			if (lock->ml.node == dlm->node_num)
2993 				continue;
2994 			if (test_bit(lock->ml.node, dlm->exit_domain_map))
2995 				continue;
2996 			nodenum = lock->ml.node;
2997 			goto bail;
2998 		}
2999 	}
3000 
3001 	/* Go thru the refmap */
3002 	noderef = -1;
3003 	while (1) {
3004 		noderef = find_next_bit(res->refmap, O2NM_MAX_NODES,
3005 					noderef + 1);
3006 		if (noderef >= O2NM_MAX_NODES)
3007 			break;
3008 		if (noderef == dlm->node_num)
3009 			continue;
3010 		if (test_bit(noderef, dlm->exit_domain_map))
3011 			continue;
3012 		nodenum = noderef;
3013 		goto bail;
3014 	}
3015 
3016 bail:
3017 	return nodenum;
3018 }
3019 
3020 /* this is called by the new master once all lockres
3021  * data has been received */
dlm_do_migrate_request(struct dlm_ctxt * dlm,struct dlm_lock_resource * res,u8 master,u8 new_master,struct dlm_node_iter * iter)3022 static int dlm_do_migrate_request(struct dlm_ctxt *dlm,
3023 				  struct dlm_lock_resource *res,
3024 				  u8 master, u8 new_master,
3025 				  struct dlm_node_iter *iter)
3026 {
3027 	struct dlm_migrate_request migrate;
3028 	int ret, skip, status = 0;
3029 	int nodenum;
3030 
3031 	memset(&migrate, 0, sizeof(migrate));
3032 	migrate.namelen = res->lockname.len;
3033 	memcpy(migrate.name, res->lockname.name, migrate.namelen);
3034 	migrate.new_master = new_master;
3035 	migrate.master = master;
3036 
3037 	ret = 0;
3038 
3039 	/* send message to all nodes, except the master and myself */
3040 	while ((nodenum = dlm_node_iter_next(iter)) >= 0) {
3041 		if (nodenum == master ||
3042 		    nodenum == new_master)
3043 			continue;
3044 
3045 		/* We could race exit domain. If exited, skip. */
3046 		spin_lock(&dlm->spinlock);
3047 		skip = (!test_bit(nodenum, dlm->domain_map));
3048 		spin_unlock(&dlm->spinlock);
3049 		if (skip) {
3050 			clear_bit(nodenum, iter->node_map);
3051 			continue;
3052 		}
3053 
3054 		ret = o2net_send_message(DLM_MIGRATE_REQUEST_MSG, dlm->key,
3055 					 &migrate, sizeof(migrate), nodenum,
3056 					 &status);
3057 		if (ret < 0) {
3058 			mlog(ML_ERROR, "%s: res %.*s, Error %d send "
3059 			     "MIGRATE_REQUEST to node %u\n", dlm->name,
3060 			     migrate.namelen, migrate.name, ret, nodenum);
3061 			if (!dlm_is_host_down(ret)) {
3062 				mlog(ML_ERROR, "unhandled error=%d!\n", ret);
3063 				BUG();
3064 			}
3065 			clear_bit(nodenum, iter->node_map);
3066 			ret = 0;
3067 		} else if (status < 0) {
3068 			mlog(0, "migrate request (node %u) returned %d!\n",
3069 			     nodenum, status);
3070 			ret = status;
3071 		} else if (status == DLM_MIGRATE_RESPONSE_MASTERY_REF) {
3072 			/* during the migration request we short-circuited
3073 			 * the mastery of the lockres.  make sure we have
3074 			 * a mastery ref for nodenum */
3075 			mlog(0, "%s:%.*s: need ref for node %u\n",
3076 			     dlm->name, res->lockname.len, res->lockname.name,
3077 			     nodenum);
3078 			spin_lock(&res->spinlock);
3079 			dlm_lockres_set_refmap_bit(dlm, res, nodenum);
3080 			spin_unlock(&res->spinlock);
3081 		}
3082 	}
3083 
3084 	if (ret < 0)
3085 		mlog_errno(ret);
3086 
3087 	mlog(0, "returning ret=%d\n", ret);
3088 	return ret;
3089 }
3090 
3091 
3092 /* if there is an existing mle for this lockres, we now know who the master is.
3093  * (the one who sent us *this* message) we can clear it up right away.
3094  * since the process that put the mle on the list still has a reference to it,
3095  * we can unhash it now, set the master and wake the process.  as a result,
3096  * we will have no mle in the list to start with.  now we can add an mle for
3097  * the migration and this should be the only one found for those scanning the
3098  * list.  */
dlm_migrate_request_handler(struct o2net_msg * msg,u32 len,void * data,void ** ret_data)3099 int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data,
3100 				void **ret_data)
3101 {
3102 	struct dlm_ctxt *dlm = data;
3103 	struct dlm_lock_resource *res = NULL;
3104 	struct dlm_migrate_request *migrate = (struct dlm_migrate_request *) msg->buf;
3105 	struct dlm_master_list_entry *mle = NULL, *oldmle = NULL;
3106 	const char *name;
3107 	unsigned int namelen, hash;
3108 	int ret = 0;
3109 
3110 	if (!dlm_grab(dlm))
3111 		return 0;
3112 
3113 	name = migrate->name;
3114 	namelen = migrate->namelen;
3115 	hash = dlm_lockid_hash(name, namelen);
3116 
3117 	/* preallocate.. if this fails, abort */
3118 	mle = kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
3119 
3120 	if (!mle) {
3121 		ret = -ENOMEM;
3122 		goto leave;
3123 	}
3124 
3125 	/* check for pre-existing lock */
3126 	spin_lock(&dlm->spinlock);
3127 	res = __dlm_lookup_lockres(dlm, name, namelen, hash);
3128 	if (res) {
3129 		spin_lock(&res->spinlock);
3130 		if (res->state & DLM_LOCK_RES_RECOVERING) {
3131 			/* if all is working ok, this can only mean that we got
3132 		 	* a migrate request from a node that we now see as
3133 		 	* dead.  what can we do here?  drop it to the floor? */
3134 			spin_unlock(&res->spinlock);
3135 			mlog(ML_ERROR, "Got a migrate request, but the "
3136 			     "lockres is marked as recovering!");
3137 			kmem_cache_free(dlm_mle_cache, mle);
3138 			ret = -EINVAL; /* need a better solution */
3139 			goto unlock;
3140 		}
3141 		res->state |= DLM_LOCK_RES_MIGRATING;
3142 		spin_unlock(&res->spinlock);
3143 	}
3144 
3145 	spin_lock(&dlm->master_lock);
3146 	/* ignore status.  only nonzero status would BUG. */
3147 	ret = dlm_add_migration_mle(dlm, res, mle, &oldmle,
3148 				    name, namelen,
3149 				    migrate->new_master,
3150 				    migrate->master);
3151 
3152 	if (ret < 0)
3153 		kmem_cache_free(dlm_mle_cache, mle);
3154 
3155 	spin_unlock(&dlm->master_lock);
3156 unlock:
3157 	spin_unlock(&dlm->spinlock);
3158 
3159 	if (oldmle) {
3160 		/* master is known, detach if not already detached */
3161 		dlm_mle_detach_hb_events(dlm, oldmle);
3162 		dlm_put_mle(oldmle);
3163 	}
3164 
3165 	if (res)
3166 		dlm_lockres_put(res);
3167 leave:
3168 	dlm_put(dlm);
3169 	return ret;
3170 }
3171 
3172 /* must be holding dlm->spinlock and dlm->master_lock
3173  * when adding a migration mle, we can clear any other mles
3174  * in the master list because we know with certainty that
3175  * the master is "master".  so we remove any old mle from
3176  * the list after setting it's master field, and then add
3177  * the new migration mle.  this way we can hold with the rule
3178  * of having only one mle for a given lock name at all times. */
dlm_add_migration_mle(struct dlm_ctxt * dlm,struct dlm_lock_resource * res,struct dlm_master_list_entry * mle,struct dlm_master_list_entry ** oldmle,const char * name,unsigned int namelen,u8 new_master,u8 master)3179 static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
3180 				 struct dlm_lock_resource *res,
3181 				 struct dlm_master_list_entry *mle,
3182 				 struct dlm_master_list_entry **oldmle,
3183 				 const char *name, unsigned int namelen,
3184 				 u8 new_master, u8 master)
3185 {
3186 	int found;
3187 	int ret = 0;
3188 
3189 	*oldmle = NULL;
3190 
3191 	assert_spin_locked(&dlm->spinlock);
3192 	assert_spin_locked(&dlm->master_lock);
3193 
3194 	/* caller is responsible for any ref taken here on oldmle */
3195 	found = dlm_find_mle(dlm, oldmle, (char *)name, namelen);
3196 	if (found) {
3197 		struct dlm_master_list_entry *tmp = *oldmle;
3198 		spin_lock(&tmp->spinlock);
3199 		if (tmp->type == DLM_MLE_MIGRATION) {
3200 			if (master == dlm->node_num) {
3201 				/* ah another process raced me to it */
3202 				mlog(0, "tried to migrate %.*s, but some "
3203 				     "process beat me to it\n",
3204 				     namelen, name);
3205 				spin_unlock(&tmp->spinlock);
3206 				return -EEXIST;
3207 			} else {
3208 				/* bad.  2 NODES are trying to migrate! */
3209 				mlog(ML_ERROR, "migration error  mle: "
3210 				     "master=%u new_master=%u // request: "
3211 				     "master=%u new_master=%u // "
3212 				     "lockres=%.*s\n",
3213 				     tmp->master, tmp->new_master,
3214 				     master, new_master,
3215 				     namelen, name);
3216 				BUG();
3217 			}
3218 		} else {
3219 			/* this is essentially what assert_master does */
3220 			tmp->master = master;
3221 			atomic_set(&tmp->woken, 1);
3222 			wake_up(&tmp->wq);
3223 			/* remove it so that only one mle will be found */
3224 			__dlm_unlink_mle(dlm, tmp);
3225 			__dlm_mle_detach_hb_events(dlm, tmp);
3226 			if (tmp->type == DLM_MLE_MASTER) {
3227 				ret = DLM_MIGRATE_RESPONSE_MASTERY_REF;
3228 				mlog(0, "%s:%.*s: master=%u, newmaster=%u, "
3229 						"telling master to get ref "
3230 						"for cleared out mle during "
3231 						"migration\n", dlm->name,
3232 						namelen, name, master,
3233 						new_master);
3234 			}
3235 		}
3236 		spin_unlock(&tmp->spinlock);
3237 	}
3238 
3239 	/* now add a migration mle to the tail of the list */
3240 	dlm_init_mle(mle, DLM_MLE_MIGRATION, dlm, res, name, namelen);
3241 	mle->new_master = new_master;
3242 	/* the new master will be sending an assert master for this.
3243 	 * at that point we will get the refmap reference */
3244 	mle->master = master;
3245 	/* do this for consistency with other mle types */
3246 	set_bit(new_master, mle->maybe_map);
3247 	__dlm_insert_mle(dlm, mle);
3248 
3249 	return ret;
3250 }
3251 
3252 /*
3253  * Sets the owner of the lockres, associated to the mle, to UNKNOWN
3254  */
dlm_reset_mleres_owner(struct dlm_ctxt * dlm,struct dlm_master_list_entry * mle)3255 static struct dlm_lock_resource *dlm_reset_mleres_owner(struct dlm_ctxt *dlm,
3256 					struct dlm_master_list_entry *mle)
3257 {
3258 	struct dlm_lock_resource *res;
3259 
3260 	/* Find the lockres associated to the mle and set its owner to UNK */
3261 	res = __dlm_lookup_lockres(dlm, mle->mname, mle->mnamelen,
3262 				   mle->mnamehash);
3263 	if (res) {
3264 		spin_unlock(&dlm->master_lock);
3265 
3266 		/* move lockres onto recovery list */
3267 		spin_lock(&res->spinlock);
3268 		dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
3269 		dlm_move_lockres_to_recovery_list(dlm, res);
3270 		spin_unlock(&res->spinlock);
3271 		dlm_lockres_put(res);
3272 
3273 		/* about to get rid of mle, detach from heartbeat */
3274 		__dlm_mle_detach_hb_events(dlm, mle);
3275 
3276 		/* dump the mle */
3277 		spin_lock(&dlm->master_lock);
3278 		__dlm_put_mle(mle);
3279 		spin_unlock(&dlm->master_lock);
3280 	}
3281 
3282 	return res;
3283 }
3284 
dlm_clean_migration_mle(struct dlm_ctxt * dlm,struct dlm_master_list_entry * mle)3285 static void dlm_clean_migration_mle(struct dlm_ctxt *dlm,
3286 				    struct dlm_master_list_entry *mle)
3287 {
3288 	__dlm_mle_detach_hb_events(dlm, mle);
3289 
3290 	spin_lock(&mle->spinlock);
3291 	__dlm_unlink_mle(dlm, mle);
3292 	atomic_set(&mle->woken, 1);
3293 	spin_unlock(&mle->spinlock);
3294 
3295 	wake_up(&mle->wq);
3296 }
3297 
dlm_clean_block_mle(struct dlm_ctxt * dlm,struct dlm_master_list_entry * mle,u8 dead_node)3298 static void dlm_clean_block_mle(struct dlm_ctxt *dlm,
3299 				struct dlm_master_list_entry *mle, u8 dead_node)
3300 {
3301 	int bit;
3302 
3303 	BUG_ON(mle->type != DLM_MLE_BLOCK);
3304 
3305 	spin_lock(&mle->spinlock);
3306 	bit = find_first_bit(mle->maybe_map, O2NM_MAX_NODES);
3307 	if (bit != dead_node) {
3308 		mlog(0, "mle found, but dead node %u would not have been "
3309 		     "master\n", dead_node);
3310 		spin_unlock(&mle->spinlock);
3311 	} else {
3312 		/* Must drop the refcount by one since the assert_master will
3313 		 * never arrive. This may result in the mle being unlinked and
3314 		 * freed, but there may still be a process waiting in the
3315 		 * dlmlock path which is fine. */
3316 		mlog(0, "node %u was expected master\n", dead_node);
3317 		atomic_set(&mle->woken, 1);
3318 		spin_unlock(&mle->spinlock);
3319 		wake_up(&mle->wq);
3320 
3321 		/* Do not need events any longer, so detach from heartbeat */
3322 		__dlm_mle_detach_hb_events(dlm, mle);
3323 		__dlm_put_mle(mle);
3324 	}
3325 }
3326 
dlm_clean_master_list(struct dlm_ctxt * dlm,u8 dead_node)3327 void dlm_clean_master_list(struct dlm_ctxt *dlm, u8 dead_node)
3328 {
3329 	struct dlm_master_list_entry *mle;
3330 	struct dlm_lock_resource *res;
3331 	struct hlist_head *bucket;
3332 	struct hlist_node *tmp;
3333 	unsigned int i;
3334 
3335 	mlog(0, "dlm=%s, dead node=%u\n", dlm->name, dead_node);
3336 top:
3337 	assert_spin_locked(&dlm->spinlock);
3338 
3339 	/* clean the master list */
3340 	spin_lock(&dlm->master_lock);
3341 	for (i = 0; i < DLM_HASH_BUCKETS; i++) {
3342 		bucket = dlm_master_hash(dlm, i);
3343 		hlist_for_each_entry_safe(mle, tmp, bucket, master_hash_node) {
3344 			BUG_ON(mle->type != DLM_MLE_BLOCK &&
3345 			       mle->type != DLM_MLE_MASTER &&
3346 			       mle->type != DLM_MLE_MIGRATION);
3347 
3348 			/* MASTER mles are initiated locally. The waiting
3349 			 * process will notice the node map change shortly.
3350 			 * Let that happen as normal. */
3351 			if (mle->type == DLM_MLE_MASTER)
3352 				continue;
3353 
3354 			/* BLOCK mles are initiated by other nodes. Need to
3355 			 * clean up if the dead node would have been the
3356 			 * master. */
3357 			if (mle->type == DLM_MLE_BLOCK) {
3358 				dlm_clean_block_mle(dlm, mle, dead_node);
3359 				continue;
3360 			}
3361 
3362 			/* Everything else is a MIGRATION mle */
3363 
3364 			/* The rule for MIGRATION mles is that the master
3365 			 * becomes UNKNOWN if *either* the original or the new
3366 			 * master dies. All UNKNOWN lockres' are sent to
3367 			 * whichever node becomes the recovery master. The new
3368 			 * master is responsible for determining if there is
3369 			 * still a master for this lockres, or if he needs to
3370 			 * take over mastery. Either way, this node should
3371 			 * expect another message to resolve this. */
3372 
3373 			if (mle->master != dead_node &&
3374 			    mle->new_master != dead_node)
3375 				continue;
3376 
3377 			if (mle->new_master == dead_node && mle->inuse) {
3378 				mlog(ML_NOTICE, "%s: target %u died during "
3379 						"migration from %u, the MLE is "
3380 						"still keep used, ignore it!\n",
3381 						dlm->name, dead_node,
3382 						mle->master);
3383 				continue;
3384 			}
3385 
3386 			/* If we have reached this point, this mle needs to be
3387 			 * removed from the list and freed. */
3388 			dlm_clean_migration_mle(dlm, mle);
3389 
3390 			mlog(0, "%s: node %u died during migration from "
3391 			     "%u to %u!\n", dlm->name, dead_node, mle->master,
3392 			     mle->new_master);
3393 
3394 			/* If we find a lockres associated with the mle, we've
3395 			 * hit this rare case that messes up our lock ordering.
3396 			 * If so, we need to drop the master lock so that we can
3397 			 * take the lockres lock, meaning that we will have to
3398 			 * restart from the head of list. */
3399 			res = dlm_reset_mleres_owner(dlm, mle);
3400 			if (res)
3401 				/* restart */
3402 				goto top;
3403 
3404 			/* This may be the last reference */
3405 			__dlm_put_mle(mle);
3406 		}
3407 	}
3408 	spin_unlock(&dlm->master_lock);
3409 }
3410 
dlm_finish_migration(struct dlm_ctxt * dlm,struct dlm_lock_resource * res,u8 old_master)3411 int dlm_finish_migration(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
3412 			 u8 old_master)
3413 {
3414 	struct dlm_node_iter iter;
3415 	int ret = 0;
3416 
3417 	spin_lock(&dlm->spinlock);
3418 	dlm_node_iter_init(dlm->domain_map, &iter);
3419 	clear_bit(old_master, iter.node_map);
3420 	clear_bit(dlm->node_num, iter.node_map);
3421 	spin_unlock(&dlm->spinlock);
3422 
3423 	/* ownership of the lockres is changing.  account for the
3424 	 * mastery reference here since old_master will briefly have
3425 	 * a reference after the migration completes */
3426 	spin_lock(&res->spinlock);
3427 	dlm_lockres_set_refmap_bit(dlm, res, old_master);
3428 	spin_unlock(&res->spinlock);
3429 
3430 	mlog(0, "now time to do a migrate request to other nodes\n");
3431 	ret = dlm_do_migrate_request(dlm, res, old_master,
3432 				     dlm->node_num, &iter);
3433 	if (ret < 0) {
3434 		mlog_errno(ret);
3435 		goto leave;
3436 	}
3437 
3438 	mlog(0, "doing assert master of %.*s to all except the original node\n",
3439 	     res->lockname.len, res->lockname.name);
3440 	/* this call now finishes out the nodemap
3441 	 * even if one or more nodes die */
3442 	ret = dlm_do_assert_master(dlm, res, iter.node_map,
3443 				   DLM_ASSERT_MASTER_FINISH_MIGRATION);
3444 	if (ret < 0) {
3445 		/* no longer need to retry.  all living nodes contacted. */
3446 		mlog_errno(ret);
3447 		ret = 0;
3448 	}
3449 
3450 	bitmap_zero(iter.node_map, O2NM_MAX_NODES);
3451 	set_bit(old_master, iter.node_map);
3452 	mlog(0, "doing assert master of %.*s back to %u\n",
3453 	     res->lockname.len, res->lockname.name, old_master);
3454 	ret = dlm_do_assert_master(dlm, res, iter.node_map,
3455 				   DLM_ASSERT_MASTER_FINISH_MIGRATION);
3456 	if (ret < 0) {
3457 		mlog(0, "assert master to original master failed "
3458 		     "with %d.\n", ret);
3459 		/* the only nonzero status here would be because of
3460 		 * a dead original node.  we're done. */
3461 		ret = 0;
3462 	}
3463 
3464 	/* all done, set the owner, clear the flag */
3465 	spin_lock(&res->spinlock);
3466 	dlm_set_lockres_owner(dlm, res, dlm->node_num);
3467 	res->state &= ~DLM_LOCK_RES_MIGRATING;
3468 	spin_unlock(&res->spinlock);
3469 	/* re-dirty it on the new master */
3470 	dlm_kick_thread(dlm, res);
3471 	wake_up(&res->wq);
3472 leave:
3473 	return ret;
3474 }
3475 
3476 /*
3477  * LOCKRES AST REFCOUNT
3478  * this is integral to migration
3479  */
3480 
3481 /* for future intent to call an ast, reserve one ahead of time.
3482  * this should be called only after waiting on the lockres
3483  * with dlm_wait_on_lockres, and while still holding the
3484  * spinlock after the call. */
__dlm_lockres_reserve_ast(struct dlm_lock_resource * res)3485 void __dlm_lockres_reserve_ast(struct dlm_lock_resource *res)
3486 {
3487 	assert_spin_locked(&res->spinlock);
3488 	if (res->state & DLM_LOCK_RES_MIGRATING) {
3489 		__dlm_print_one_lock_resource(res);
3490 	}
3491 	BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3492 
3493 	atomic_inc(&res->asts_reserved);
3494 }
3495 
3496 /*
3497  * used to drop the reserved ast, either because it went unused,
3498  * or because the ast/bast was actually called.
3499  *
3500  * also, if there is a pending migration on this lockres,
3501  * and this was the last pending ast on the lockres,
3502  * atomically set the MIGRATING flag before we drop the lock.
3503  * this is how we ensure that migration can proceed with no
3504  * asts in progress.  note that it is ok if the state of the
3505  * queues is such that a lock should be granted in the future
3506  * or that a bast should be fired, because the new master will
3507  * shuffle the lists on this lockres as soon as it is migrated.
3508  */
dlm_lockres_release_ast(struct dlm_ctxt * dlm,struct dlm_lock_resource * res)3509 void dlm_lockres_release_ast(struct dlm_ctxt *dlm,
3510 			     struct dlm_lock_resource *res)
3511 {
3512 	if (!atomic_dec_and_lock(&res->asts_reserved, &res->spinlock))
3513 		return;
3514 
3515 	if (!res->migration_pending) {
3516 		spin_unlock(&res->spinlock);
3517 		return;
3518 	}
3519 
3520 	BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
3521 	res->migration_pending = 0;
3522 	res->state |= DLM_LOCK_RES_MIGRATING;
3523 	spin_unlock(&res->spinlock);
3524 	wake_up(&res->wq);
3525 	wake_up(&dlm->migration_wq);
3526 }
3527 
dlm_force_free_mles(struct dlm_ctxt * dlm)3528 void dlm_force_free_mles(struct dlm_ctxt *dlm)
3529 {
3530 	int i;
3531 	struct hlist_head *bucket;
3532 	struct dlm_master_list_entry *mle;
3533 	struct hlist_node *tmp;
3534 
3535 	/*
3536 	 * We notified all other nodes that we are exiting the domain and
3537 	 * marked the dlm state to DLM_CTXT_LEAVING. If any mles are still
3538 	 * around we force free them and wake any processes that are waiting
3539 	 * on the mles
3540 	 */
3541 	spin_lock(&dlm->spinlock);
3542 	spin_lock(&dlm->master_lock);
3543 
3544 	BUG_ON(dlm->dlm_state != DLM_CTXT_LEAVING);
3545 	BUG_ON((find_first_bit(dlm->domain_map, O2NM_MAX_NODES) < O2NM_MAX_NODES));
3546 
3547 	for (i = 0; i < DLM_HASH_BUCKETS; i++) {
3548 		bucket = dlm_master_hash(dlm, i);
3549 		hlist_for_each_entry_safe(mle, tmp, bucket, master_hash_node) {
3550 			if (mle->type != DLM_MLE_BLOCK) {
3551 				mlog(ML_ERROR, "bad mle: %p\n", mle);
3552 				dlm_print_one_mle(mle);
3553 			}
3554 			atomic_set(&mle->woken, 1);
3555 			wake_up(&mle->wq);
3556 
3557 			__dlm_unlink_mle(dlm, mle);
3558 			__dlm_mle_detach_hb_events(dlm, mle);
3559 			__dlm_put_mle(mle);
3560 		}
3561 	}
3562 	spin_unlock(&dlm->master_lock);
3563 	spin_unlock(&dlm->spinlock);
3564 }
3565