Loading...
1/******************************************************************************
2*******************************************************************************
3**
4** Copyright (C) 2005-2011 Red Hat, Inc. All rights reserved.
5**
6** This copyrighted material is made available to anyone wishing to use,
7** modify, copy, or redistribute it subject to the terms and conditions
8** of the GNU General Public License v.2.
9**
10*******************************************************************************
11******************************************************************************/
12
13#include "dlm_internal.h"
14#include "lockspace.h"
15#include "member.h"
16#include "recoverd.h"
17#include "recover.h"
18#include "rcom.h"
19#include "config.h"
20#include "lowcomms.h"
21
22int dlm_slots_version(struct dlm_header *h)
23{
24 if ((h->h_version & 0x0000FFFF) < DLM_HEADER_SLOTS)
25 return 0;
26 return 1;
27}
28
29void dlm_slot_save(struct dlm_ls *ls, struct dlm_rcom *rc,
30 struct dlm_member *memb)
31{
32 struct rcom_config *rf = (struct rcom_config *)rc->rc_buf;
33
34 if (!dlm_slots_version(&rc->rc_header))
35 return;
36
37 memb->slot = le16_to_cpu(rf->rf_our_slot);
38 memb->generation = le32_to_cpu(rf->rf_generation);
39}
40
41void dlm_slots_copy_out(struct dlm_ls *ls, struct dlm_rcom *rc)
42{
43 struct dlm_slot *slot;
44 struct rcom_slot *ro;
45 int i;
46
47 ro = (struct rcom_slot *)(rc->rc_buf + sizeof(struct rcom_config));
48
49 /* ls_slots array is sparse, but not rcom_slots */
50
51 for (i = 0; i < ls->ls_slots_size; i++) {
52 slot = &ls->ls_slots[i];
53 if (!slot->nodeid)
54 continue;
55 ro->ro_nodeid = cpu_to_le32(slot->nodeid);
56 ro->ro_slot = cpu_to_le16(slot->slot);
57 ro++;
58 }
59}
60
61#define SLOT_DEBUG_LINE 128
62
63static void log_slots(struct dlm_ls *ls, uint32_t gen, int num_slots,
64 struct rcom_slot *ro0, struct dlm_slot *array,
65 int array_size)
66{
67 char line[SLOT_DEBUG_LINE];
68 int len = SLOT_DEBUG_LINE - 1;
69 int pos = 0;
70 int ret, i;
71
72 memset(line, 0, sizeof(line));
73
74 if (array) {
75 for (i = 0; i < array_size; i++) {
76 if (!array[i].nodeid)
77 continue;
78
79 ret = snprintf(line + pos, len - pos, " %d:%d",
80 array[i].slot, array[i].nodeid);
81 if (ret >= len - pos)
82 break;
83 pos += ret;
84 }
85 } else if (ro0) {
86 for (i = 0; i < num_slots; i++) {
87 ret = snprintf(line + pos, len - pos, " %d:%d",
88 ro0[i].ro_slot, ro0[i].ro_nodeid);
89 if (ret >= len - pos)
90 break;
91 pos += ret;
92 }
93 }
94
95 log_rinfo(ls, "generation %u slots %d%s", gen, num_slots, line);
96}
97
98int dlm_slots_copy_in(struct dlm_ls *ls)
99{
100 struct dlm_member *memb;
101 struct dlm_rcom *rc = ls->ls_recover_buf;
102 struct rcom_config *rf = (struct rcom_config *)rc->rc_buf;
103 struct rcom_slot *ro0, *ro;
104 int our_nodeid = dlm_our_nodeid();
105 int i, num_slots;
106 uint32_t gen;
107
108 if (!dlm_slots_version(&rc->rc_header))
109 return -1;
110
111 gen = le32_to_cpu(rf->rf_generation);
112 if (gen <= ls->ls_generation) {
113 log_error(ls, "dlm_slots_copy_in gen %u old %u",
114 gen, ls->ls_generation);
115 }
116 ls->ls_generation = gen;
117
118 num_slots = le16_to_cpu(rf->rf_num_slots);
119 if (!num_slots)
120 return -1;
121
122 ro0 = (struct rcom_slot *)(rc->rc_buf + sizeof(struct rcom_config));
123
124 for (i = 0, ro = ro0; i < num_slots; i++, ro++) {
125 ro->ro_nodeid = le32_to_cpu(ro->ro_nodeid);
126 ro->ro_slot = le16_to_cpu(ro->ro_slot);
127 }
128
129 log_slots(ls, gen, num_slots, ro0, NULL, 0);
130
131 list_for_each_entry(memb, &ls->ls_nodes, list) {
132 for (i = 0, ro = ro0; i < num_slots; i++, ro++) {
133 if (ro->ro_nodeid != memb->nodeid)
134 continue;
135 memb->slot = ro->ro_slot;
136 memb->slot_prev = memb->slot;
137 break;
138 }
139
140 if (memb->nodeid == our_nodeid) {
141 if (ls->ls_slot && ls->ls_slot != memb->slot) {
142 log_error(ls, "dlm_slots_copy_in our slot "
143 "changed %d %d", ls->ls_slot,
144 memb->slot);
145 return -1;
146 }
147
148 if (!ls->ls_slot)
149 ls->ls_slot = memb->slot;
150 }
151
152 if (!memb->slot) {
153 log_error(ls, "dlm_slots_copy_in nodeid %d no slot",
154 memb->nodeid);
155 return -1;
156 }
157 }
158
159 return 0;
160}
161
162/* for any nodes that do not support slots, we will not have set memb->slot
163 in wait_status_all(), so memb->slot will remain -1, and we will not
164 assign slots or set ls_num_slots here */
165
166int dlm_slots_assign(struct dlm_ls *ls, int *num_slots, int *slots_size,
167 struct dlm_slot **slots_out, uint32_t *gen_out)
168{
169 struct dlm_member *memb;
170 struct dlm_slot *array;
171 int our_nodeid = dlm_our_nodeid();
172 int array_size, max_slots, i;
173 int need = 0;
174 int max = 0;
175 int num = 0;
176 uint32_t gen = 0;
177
178 /* our own memb struct will have slot -1 gen 0 */
179
180 list_for_each_entry(memb, &ls->ls_nodes, list) {
181 if (memb->nodeid == our_nodeid) {
182 memb->slot = ls->ls_slot;
183 memb->generation = ls->ls_generation;
184 break;
185 }
186 }
187
188 list_for_each_entry(memb, &ls->ls_nodes, list) {
189 if (memb->generation > gen)
190 gen = memb->generation;
191
192 /* node doesn't support slots */
193
194 if (memb->slot == -1)
195 return -1;
196
197 /* node needs a slot assigned */
198
199 if (!memb->slot)
200 need++;
201
202 /* node has a slot assigned */
203
204 num++;
205
206 if (!max || max < memb->slot)
207 max = memb->slot;
208
209 /* sanity check, once slot is assigned it shouldn't change */
210
211 if (memb->slot_prev && memb->slot && memb->slot_prev != memb->slot) {
212 log_error(ls, "nodeid %d slot changed %d %d",
213 memb->nodeid, memb->slot_prev, memb->slot);
214 return -1;
215 }
216 memb->slot_prev = memb->slot;
217 }
218
219 array_size = max + need;
220
221 array = kzalloc(array_size * sizeof(struct dlm_slot), GFP_NOFS);
222 if (!array)
223 return -ENOMEM;
224
225 num = 0;
226
227 /* fill in slots (offsets) that are used */
228
229 list_for_each_entry(memb, &ls->ls_nodes, list) {
230 if (!memb->slot)
231 continue;
232
233 if (memb->slot > array_size) {
234 log_error(ls, "invalid slot number %d", memb->slot);
235 kfree(array);
236 return -1;
237 }
238
239 array[memb->slot - 1].nodeid = memb->nodeid;
240 array[memb->slot - 1].slot = memb->slot;
241 num++;
242 }
243
244 /* assign new slots from unused offsets */
245
246 list_for_each_entry(memb, &ls->ls_nodes, list) {
247 if (memb->slot)
248 continue;
249
250 for (i = 0; i < array_size; i++) {
251 if (array[i].nodeid)
252 continue;
253
254 memb->slot = i + 1;
255 memb->slot_prev = memb->slot;
256 array[i].nodeid = memb->nodeid;
257 array[i].slot = memb->slot;
258 num++;
259
260 if (!ls->ls_slot && memb->nodeid == our_nodeid)
261 ls->ls_slot = memb->slot;
262 break;
263 }
264
265 if (!memb->slot) {
266 log_error(ls, "no free slot found");
267 kfree(array);
268 return -1;
269 }
270 }
271
272 gen++;
273
274 log_slots(ls, gen, num, NULL, array, array_size);
275
276 max_slots = (dlm_config.ci_buffer_size - sizeof(struct dlm_rcom) -
277 sizeof(struct rcom_config)) / sizeof(struct rcom_slot);
278
279 if (num > max_slots) {
280 log_error(ls, "num_slots %d exceeds max_slots %d",
281 num, max_slots);
282 kfree(array);
283 return -1;
284 }
285
286 *gen_out = gen;
287 *slots_out = array;
288 *slots_size = array_size;
289 *num_slots = num;
290 return 0;
291}
292
293static void add_ordered_member(struct dlm_ls *ls, struct dlm_member *new)
294{
295 struct dlm_member *memb = NULL;
296 struct list_head *tmp;
297 struct list_head *newlist = &new->list;
298 struct list_head *head = &ls->ls_nodes;
299
300 list_for_each(tmp, head) {
301 memb = list_entry(tmp, struct dlm_member, list);
302 if (new->nodeid < memb->nodeid)
303 break;
304 }
305
306 if (!memb)
307 list_add_tail(newlist, head);
308 else {
309 /* FIXME: can use list macro here */
310 newlist->prev = tmp->prev;
311 newlist->next = tmp;
312 tmp->prev->next = newlist;
313 tmp->prev = newlist;
314 }
315}
316
317static int dlm_add_member(struct dlm_ls *ls, struct dlm_config_node *node)
318{
319 struct dlm_member *memb;
320 int error;
321
322 memb = kzalloc(sizeof(struct dlm_member), GFP_NOFS);
323 if (!memb)
324 return -ENOMEM;
325
326 error = dlm_lowcomms_connect_node(node->nodeid);
327 if (error < 0) {
328 kfree(memb);
329 return error;
330 }
331
332 memb->nodeid = node->nodeid;
333 memb->weight = node->weight;
334 memb->comm_seq = node->comm_seq;
335 add_ordered_member(ls, memb);
336 ls->ls_num_nodes++;
337 return 0;
338}
339
340static struct dlm_member *find_memb(struct list_head *head, int nodeid)
341{
342 struct dlm_member *memb;
343
344 list_for_each_entry(memb, head, list) {
345 if (memb->nodeid == nodeid)
346 return memb;
347 }
348 return NULL;
349}
350
351int dlm_is_member(struct dlm_ls *ls, int nodeid)
352{
353 if (find_memb(&ls->ls_nodes, nodeid))
354 return 1;
355 return 0;
356}
357
358int dlm_is_removed(struct dlm_ls *ls, int nodeid)
359{
360 if (find_memb(&ls->ls_nodes_gone, nodeid))
361 return 1;
362 return 0;
363}
364
365static void clear_memb_list(struct list_head *head)
366{
367 struct dlm_member *memb;
368
369 while (!list_empty(head)) {
370 memb = list_entry(head->next, struct dlm_member, list);
371 list_del(&memb->list);
372 kfree(memb);
373 }
374}
375
376void dlm_clear_members(struct dlm_ls *ls)
377{
378 clear_memb_list(&ls->ls_nodes);
379 ls->ls_num_nodes = 0;
380}
381
382void dlm_clear_members_gone(struct dlm_ls *ls)
383{
384 clear_memb_list(&ls->ls_nodes_gone);
385}
386
387static void make_member_array(struct dlm_ls *ls)
388{
389 struct dlm_member *memb;
390 int i, w, x = 0, total = 0, all_zero = 0, *array;
391
392 kfree(ls->ls_node_array);
393 ls->ls_node_array = NULL;
394
395 list_for_each_entry(memb, &ls->ls_nodes, list) {
396 if (memb->weight)
397 total += memb->weight;
398 }
399
400 /* all nodes revert to weight of 1 if all have weight 0 */
401
402 if (!total) {
403 total = ls->ls_num_nodes;
404 all_zero = 1;
405 }
406
407 ls->ls_total_weight = total;
408
409 array = kmalloc(sizeof(int) * total, GFP_NOFS);
410 if (!array)
411 return;
412
413 list_for_each_entry(memb, &ls->ls_nodes, list) {
414 if (!all_zero && !memb->weight)
415 continue;
416
417 if (all_zero)
418 w = 1;
419 else
420 w = memb->weight;
421
422 DLM_ASSERT(x < total, printk("total %d x %d\n", total, x););
423
424 for (i = 0; i < w; i++)
425 array[x++] = memb->nodeid;
426 }
427
428 ls->ls_node_array = array;
429}
430
431/* send a status request to all members just to establish comms connections */
432
433static int ping_members(struct dlm_ls *ls)
434{
435 struct dlm_member *memb;
436 int error = 0;
437
438 list_for_each_entry(memb, &ls->ls_nodes, list) {
439 error = dlm_recovery_stopped(ls);
440 if (error)
441 break;
442 error = dlm_rcom_status(ls, memb->nodeid, 0);
443 if (error)
444 break;
445 }
446 if (error)
447 log_rinfo(ls, "ping_members aborted %d last nodeid %d",
448 error, ls->ls_recover_nodeid);
449 return error;
450}
451
452static void dlm_lsop_recover_prep(struct dlm_ls *ls)
453{
454 if (!ls->ls_ops || !ls->ls_ops->recover_prep)
455 return;
456 ls->ls_ops->recover_prep(ls->ls_ops_arg);
457}
458
459static void dlm_lsop_recover_slot(struct dlm_ls *ls, struct dlm_member *memb)
460{
461 struct dlm_slot slot;
462 uint32_t seq;
463 int error;
464
465 if (!ls->ls_ops || !ls->ls_ops->recover_slot)
466 return;
467
468 /* if there is no comms connection with this node
469 or the present comms connection is newer
470 than the one when this member was added, then
471 we consider the node to have failed (versus
472 being removed due to dlm_release_lockspace) */
473
474 error = dlm_comm_seq(memb->nodeid, &seq);
475
476 if (!error && seq == memb->comm_seq)
477 return;
478
479 slot.nodeid = memb->nodeid;
480 slot.slot = memb->slot;
481
482 ls->ls_ops->recover_slot(ls->ls_ops_arg, &slot);
483}
484
485void dlm_lsop_recover_done(struct dlm_ls *ls)
486{
487 struct dlm_member *memb;
488 struct dlm_slot *slots;
489 int i, num;
490
491 if (!ls->ls_ops || !ls->ls_ops->recover_done)
492 return;
493
494 num = ls->ls_num_nodes;
495
496 slots = kzalloc(num * sizeof(struct dlm_slot), GFP_KERNEL);
497 if (!slots)
498 return;
499
500 i = 0;
501 list_for_each_entry(memb, &ls->ls_nodes, list) {
502 if (i == num) {
503 log_error(ls, "dlm_lsop_recover_done bad num %d", num);
504 goto out;
505 }
506 slots[i].nodeid = memb->nodeid;
507 slots[i].slot = memb->slot;
508 i++;
509 }
510
511 ls->ls_ops->recover_done(ls->ls_ops_arg, slots, num,
512 ls->ls_slot, ls->ls_generation);
513 out:
514 kfree(slots);
515}
516
517static struct dlm_config_node *find_config_node(struct dlm_recover *rv,
518 int nodeid)
519{
520 int i;
521
522 for (i = 0; i < rv->nodes_count; i++) {
523 if (rv->nodes[i].nodeid == nodeid)
524 return &rv->nodes[i];
525 }
526 return NULL;
527}
528
529int dlm_recover_members(struct dlm_ls *ls, struct dlm_recover *rv, int *neg_out)
530{
531 struct dlm_member *memb, *safe;
532 struct dlm_config_node *node;
533 int i, error, neg = 0, low = -1;
534
535 /* previously removed members that we've not finished removing need to
536 count as a negative change so the "neg" recovery steps will happen */
537
538 list_for_each_entry(memb, &ls->ls_nodes_gone, list) {
539 log_rinfo(ls, "prev removed member %d", memb->nodeid);
540 neg++;
541 }
542
543 /* move departed members from ls_nodes to ls_nodes_gone */
544
545 list_for_each_entry_safe(memb, safe, &ls->ls_nodes, list) {
546 node = find_config_node(rv, memb->nodeid);
547 if (node && !node->new)
548 continue;
549
550 if (!node) {
551 log_rinfo(ls, "remove member %d", memb->nodeid);
552 } else {
553 /* removed and re-added */
554 log_rinfo(ls, "remove member %d comm_seq %u %u",
555 memb->nodeid, memb->comm_seq, node->comm_seq);
556 }
557
558 neg++;
559 list_move(&memb->list, &ls->ls_nodes_gone);
560 ls->ls_num_nodes--;
561 dlm_lsop_recover_slot(ls, memb);
562 }
563
564 /* add new members to ls_nodes */
565
566 for (i = 0; i < rv->nodes_count; i++) {
567 node = &rv->nodes[i];
568 if (dlm_is_member(ls, node->nodeid))
569 continue;
570 dlm_add_member(ls, node);
571 log_rinfo(ls, "add member %d", node->nodeid);
572 }
573
574 list_for_each_entry(memb, &ls->ls_nodes, list) {
575 if (low == -1 || memb->nodeid < low)
576 low = memb->nodeid;
577 }
578 ls->ls_low_nodeid = low;
579
580 make_member_array(ls);
581 *neg_out = neg;
582
583 error = ping_members(ls);
584 if (!error || error == -EPROTO) {
585 /* new_lockspace() may be waiting to know if the config
586 is good or bad */
587 ls->ls_members_result = error;
588 complete(&ls->ls_members_done);
589 }
590
591 log_rinfo(ls, "dlm_recover_members %d nodes", ls->ls_num_nodes);
592 return error;
593}
594
595/* Userspace guarantees that dlm_ls_stop() has completed on all nodes before
596 dlm_ls_start() is called on any of them to start the new recovery. */
597
598int dlm_ls_stop(struct dlm_ls *ls)
599{
600 int new;
601
602 /*
603 * Prevent dlm_recv from being in the middle of something when we do
604 * the stop. This includes ensuring dlm_recv isn't processing a
605 * recovery message (rcom), while dlm_recoverd is aborting and
606 * resetting things from an in-progress recovery. i.e. we want
607 * dlm_recoverd to abort its recovery without worrying about dlm_recv
608 * processing an rcom at the same time. Stopping dlm_recv also makes
609 * it easy for dlm_receive_message() to check locking stopped and add a
610 * message to the requestqueue without races.
611 */
612
613 down_write(&ls->ls_recv_active);
614
615 /*
616 * Abort any recovery that's in progress (see RECOVER_STOP,
617 * dlm_recovery_stopped()) and tell any other threads running in the
618 * dlm to quit any processing (see RUNNING, dlm_locking_stopped()).
619 */
620
621 spin_lock(&ls->ls_recover_lock);
622 set_bit(LSFL_RECOVER_STOP, &ls->ls_flags);
623 new = test_and_clear_bit(LSFL_RUNNING, &ls->ls_flags);
624 ls->ls_recover_seq++;
625 spin_unlock(&ls->ls_recover_lock);
626
627 /*
628 * Let dlm_recv run again, now any normal messages will be saved on the
629 * requestqueue for later.
630 */
631
632 up_write(&ls->ls_recv_active);
633
634 /*
635 * This in_recovery lock does two things:
636 * 1) Keeps this function from returning until all threads are out
637 * of locking routines and locking is truly stopped.
638 * 2) Keeps any new requests from being processed until it's unlocked
639 * when recovery is complete.
640 */
641
642 if (new) {
643 set_bit(LSFL_RECOVER_DOWN, &ls->ls_flags);
644 wake_up_process(ls->ls_recoverd_task);
645 wait_event(ls->ls_recover_lock_wait,
646 test_bit(LSFL_RECOVER_LOCK, &ls->ls_flags));
647 }
648
649 /*
650 * The recoverd suspend/resume makes sure that dlm_recoverd (if
651 * running) has noticed RECOVER_STOP above and quit processing the
652 * previous recovery.
653 */
654
655 dlm_recoverd_suspend(ls);
656
657 spin_lock(&ls->ls_recover_lock);
658 kfree(ls->ls_slots);
659 ls->ls_slots = NULL;
660 ls->ls_num_slots = 0;
661 ls->ls_slots_size = 0;
662 ls->ls_recover_status = 0;
663 spin_unlock(&ls->ls_recover_lock);
664
665 dlm_recoverd_resume(ls);
666
667 if (!ls->ls_recover_begin)
668 ls->ls_recover_begin = jiffies;
669
670 dlm_lsop_recover_prep(ls);
671 return 0;
672}
673
674int dlm_ls_start(struct dlm_ls *ls)
675{
676 struct dlm_recover *rv = NULL, *rv_old;
677 struct dlm_config_node *nodes;
678 int error, count;
679
680 rv = kzalloc(sizeof(struct dlm_recover), GFP_NOFS);
681 if (!rv)
682 return -ENOMEM;
683
684 error = dlm_config_nodes(ls->ls_name, &nodes, &count);
685 if (error < 0)
686 goto fail;
687
688 spin_lock(&ls->ls_recover_lock);
689
690 /* the lockspace needs to be stopped before it can be started */
691
692 if (!dlm_locking_stopped(ls)) {
693 spin_unlock(&ls->ls_recover_lock);
694 log_error(ls, "start ignored: lockspace running");
695 error = -EINVAL;
696 goto fail;
697 }
698
699 rv->nodes = nodes;
700 rv->nodes_count = count;
701 rv->seq = ++ls->ls_recover_seq;
702 rv_old = ls->ls_recover_args;
703 ls->ls_recover_args = rv;
704 spin_unlock(&ls->ls_recover_lock);
705
706 if (rv_old) {
707 log_error(ls, "unused recovery %llx %d",
708 (unsigned long long)rv_old->seq, rv_old->nodes_count);
709 kfree(rv_old->nodes);
710 kfree(rv_old);
711 }
712
713 set_bit(LSFL_RECOVER_WORK, &ls->ls_flags);
714 wake_up_process(ls->ls_recoverd_task);
715 return 0;
716
717 fail:
718 kfree(rv);
719 kfree(nodes);
720 return error;
721}
722
1// SPDX-License-Identifier: GPL-2.0-only
2/******************************************************************************
3*******************************************************************************
4**
5** Copyright (C) 2005-2011 Red Hat, Inc. All rights reserved.
6**
7**
8*******************************************************************************
9******************************************************************************/
10
11#include "dlm_internal.h"
12#include "lockspace.h"
13#include "member.h"
14#include "recoverd.h"
15#include "recover.h"
16#include "rcom.h"
17#include "config.h"
18#include "midcomms.h"
19#include "lowcomms.h"
20
21int dlm_slots_version(const struct dlm_header *h)
22{
23 if ((le32_to_cpu(h->h_version) & 0x0000FFFF) < DLM_HEADER_SLOTS)
24 return 0;
25 return 1;
26}
27
28void dlm_slot_save(struct dlm_ls *ls, struct dlm_rcom *rc,
29 struct dlm_member *memb)
30{
31 struct rcom_config *rf = (struct rcom_config *)rc->rc_buf;
32
33 if (!dlm_slots_version(&rc->rc_header))
34 return;
35
36 memb->slot = le16_to_cpu(rf->rf_our_slot);
37 memb->generation = le32_to_cpu(rf->rf_generation);
38}
39
40void dlm_slots_copy_out(struct dlm_ls *ls, struct dlm_rcom *rc)
41{
42 struct dlm_slot *slot;
43 struct rcom_slot *ro;
44 int i;
45
46 ro = (struct rcom_slot *)(rc->rc_buf + sizeof(struct rcom_config));
47
48 /* ls_slots array is sparse, but not rcom_slots */
49
50 for (i = 0; i < ls->ls_slots_size; i++) {
51 slot = &ls->ls_slots[i];
52 if (!slot->nodeid)
53 continue;
54 ro->ro_nodeid = cpu_to_le32(slot->nodeid);
55 ro->ro_slot = cpu_to_le16(slot->slot);
56 ro++;
57 }
58}
59
60#define SLOT_DEBUG_LINE 128
61
62static void log_slots(struct dlm_ls *ls, uint32_t gen, int num_slots,
63 struct rcom_slot *ro0, struct dlm_slot *array,
64 int array_size)
65{
66 char line[SLOT_DEBUG_LINE];
67 int len = SLOT_DEBUG_LINE - 1;
68 int pos = 0;
69 int ret, i;
70
71 memset(line, 0, sizeof(line));
72
73 if (array) {
74 for (i = 0; i < array_size; i++) {
75 if (!array[i].nodeid)
76 continue;
77
78 ret = snprintf(line + pos, len - pos, " %d:%d",
79 array[i].slot, array[i].nodeid);
80 if (ret >= len - pos)
81 break;
82 pos += ret;
83 }
84 } else if (ro0) {
85 for (i = 0; i < num_slots; i++) {
86 ret = snprintf(line + pos, len - pos, " %d:%d",
87 ro0[i].ro_slot, ro0[i].ro_nodeid);
88 if (ret >= len - pos)
89 break;
90 pos += ret;
91 }
92 }
93
94 log_rinfo(ls, "generation %u slots %d%s", gen, num_slots, line);
95}
96
97int dlm_slots_copy_in(struct dlm_ls *ls)
98{
99 struct dlm_member *memb;
100 struct dlm_rcom *rc = ls->ls_recover_buf;
101 struct rcom_config *rf = (struct rcom_config *)rc->rc_buf;
102 struct rcom_slot *ro0, *ro;
103 int our_nodeid = dlm_our_nodeid();
104 int i, num_slots;
105 uint32_t gen;
106
107 if (!dlm_slots_version(&rc->rc_header))
108 return -1;
109
110 gen = le32_to_cpu(rf->rf_generation);
111 if (gen <= ls->ls_generation) {
112 log_error(ls, "dlm_slots_copy_in gen %u old %u",
113 gen, ls->ls_generation);
114 }
115 ls->ls_generation = gen;
116
117 num_slots = le16_to_cpu(rf->rf_num_slots);
118 if (!num_slots)
119 return -1;
120
121 ro0 = (struct rcom_slot *)(rc->rc_buf + sizeof(struct rcom_config));
122
123 log_slots(ls, gen, num_slots, ro0, NULL, 0);
124
125 list_for_each_entry(memb, &ls->ls_nodes, list) {
126 for (i = 0, ro = ro0; i < num_slots; i++, ro++) {
127 if (le32_to_cpu(ro->ro_nodeid) != memb->nodeid)
128 continue;
129 memb->slot = le16_to_cpu(ro->ro_slot);
130 memb->slot_prev = memb->slot;
131 break;
132 }
133
134 if (memb->nodeid == our_nodeid) {
135 if (ls->ls_slot && ls->ls_slot != memb->slot) {
136 log_error(ls, "dlm_slots_copy_in our slot "
137 "changed %d %d", ls->ls_slot,
138 memb->slot);
139 return -1;
140 }
141
142 if (!ls->ls_slot)
143 ls->ls_slot = memb->slot;
144 }
145
146 if (!memb->slot) {
147 log_error(ls, "dlm_slots_copy_in nodeid %d no slot",
148 memb->nodeid);
149 return -1;
150 }
151 }
152
153 return 0;
154}
155
156/* for any nodes that do not support slots, we will not have set memb->slot
157 in wait_status_all(), so memb->slot will remain -1, and we will not
158 assign slots or set ls_num_slots here */
159
160int dlm_slots_assign(struct dlm_ls *ls, int *num_slots, int *slots_size,
161 struct dlm_slot **slots_out, uint32_t *gen_out)
162{
163 struct dlm_member *memb;
164 struct dlm_slot *array;
165 int our_nodeid = dlm_our_nodeid();
166 int array_size, max_slots, i;
167 int need = 0;
168 int max = 0;
169 int num = 0;
170 uint32_t gen = 0;
171
172 /* our own memb struct will have slot -1 gen 0 */
173
174 list_for_each_entry(memb, &ls->ls_nodes, list) {
175 if (memb->nodeid == our_nodeid) {
176 memb->slot = ls->ls_slot;
177 memb->generation = ls->ls_generation;
178 break;
179 }
180 }
181
182 list_for_each_entry(memb, &ls->ls_nodes, list) {
183 if (memb->generation > gen)
184 gen = memb->generation;
185
186 /* node doesn't support slots */
187
188 if (memb->slot == -1)
189 return -1;
190
191 /* node needs a slot assigned */
192
193 if (!memb->slot)
194 need++;
195
196 /* node has a slot assigned */
197
198 num++;
199
200 if (!max || max < memb->slot)
201 max = memb->slot;
202
203 /* sanity check, once slot is assigned it shouldn't change */
204
205 if (memb->slot_prev && memb->slot && memb->slot_prev != memb->slot) {
206 log_error(ls, "nodeid %d slot changed %d %d",
207 memb->nodeid, memb->slot_prev, memb->slot);
208 return -1;
209 }
210 memb->slot_prev = memb->slot;
211 }
212
213 array_size = max + need;
214 array = kcalloc(array_size, sizeof(*array), GFP_NOFS);
215 if (!array)
216 return -ENOMEM;
217
218 num = 0;
219
220 /* fill in slots (offsets) that are used */
221
222 list_for_each_entry(memb, &ls->ls_nodes, list) {
223 if (!memb->slot)
224 continue;
225
226 if (memb->slot > array_size) {
227 log_error(ls, "invalid slot number %d", memb->slot);
228 kfree(array);
229 return -1;
230 }
231
232 array[memb->slot - 1].nodeid = memb->nodeid;
233 array[memb->slot - 1].slot = memb->slot;
234 num++;
235 }
236
237 /* assign new slots from unused offsets */
238
239 list_for_each_entry(memb, &ls->ls_nodes, list) {
240 if (memb->slot)
241 continue;
242
243 for (i = 0; i < array_size; i++) {
244 if (array[i].nodeid)
245 continue;
246
247 memb->slot = i + 1;
248 memb->slot_prev = memb->slot;
249 array[i].nodeid = memb->nodeid;
250 array[i].slot = memb->slot;
251 num++;
252
253 if (!ls->ls_slot && memb->nodeid == our_nodeid)
254 ls->ls_slot = memb->slot;
255 break;
256 }
257
258 if (!memb->slot) {
259 log_error(ls, "no free slot found");
260 kfree(array);
261 return -1;
262 }
263 }
264
265 gen++;
266
267 log_slots(ls, gen, num, NULL, array, array_size);
268
269 max_slots = (DLM_MAX_APP_BUFSIZE - sizeof(struct dlm_rcom) -
270 sizeof(struct rcom_config)) / sizeof(struct rcom_slot);
271
272 if (num > max_slots) {
273 log_error(ls, "num_slots %d exceeds max_slots %d",
274 num, max_slots);
275 kfree(array);
276 return -1;
277 }
278
279 *gen_out = gen;
280 *slots_out = array;
281 *slots_size = array_size;
282 *num_slots = num;
283 return 0;
284}
285
286static void add_ordered_member(struct dlm_ls *ls, struct dlm_member *new)
287{
288 struct dlm_member *memb = NULL;
289 struct list_head *tmp;
290 struct list_head *newlist = &new->list;
291 struct list_head *head = &ls->ls_nodes;
292
293 list_for_each(tmp, head) {
294 memb = list_entry(tmp, struct dlm_member, list);
295 if (new->nodeid < memb->nodeid)
296 break;
297 }
298
299 if (!memb)
300 list_add_tail(newlist, head);
301 else {
302 /* FIXME: can use list macro here */
303 newlist->prev = tmp->prev;
304 newlist->next = tmp;
305 tmp->prev->next = newlist;
306 tmp->prev = newlist;
307 }
308}
309
310static int add_remote_member(int nodeid)
311{
312 int error;
313
314 if (nodeid == dlm_our_nodeid())
315 return 0;
316
317 error = dlm_lowcomms_connect_node(nodeid);
318 if (error < 0)
319 return error;
320
321 dlm_midcomms_add_member(nodeid);
322 return 0;
323}
324
325static int dlm_add_member(struct dlm_ls *ls, struct dlm_config_node *node)
326{
327 struct dlm_member *memb;
328 int error;
329
330 memb = kzalloc(sizeof(*memb), GFP_NOFS);
331 if (!memb)
332 return -ENOMEM;
333
334 memb->nodeid = node->nodeid;
335 memb->weight = node->weight;
336 memb->comm_seq = node->comm_seq;
337
338 error = add_remote_member(node->nodeid);
339 if (error < 0) {
340 kfree(memb);
341 return error;
342 }
343
344 add_ordered_member(ls, memb);
345 ls->ls_num_nodes++;
346 return 0;
347}
348
349static struct dlm_member *find_memb(struct list_head *head, int nodeid)
350{
351 struct dlm_member *memb;
352
353 list_for_each_entry(memb, head, list) {
354 if (memb->nodeid == nodeid)
355 return memb;
356 }
357 return NULL;
358}
359
360int dlm_is_member(struct dlm_ls *ls, int nodeid)
361{
362 if (find_memb(&ls->ls_nodes, nodeid))
363 return 1;
364 return 0;
365}
366
367int dlm_is_removed(struct dlm_ls *ls, int nodeid)
368{
369 WARN_ON_ONCE(!nodeid || nodeid == -1);
370
371 if (find_memb(&ls->ls_nodes_gone, nodeid))
372 return 1;
373 return 0;
374}
375
376static void clear_memb_list(struct list_head *head,
377 void (*after_del)(int nodeid))
378{
379 struct dlm_member *memb;
380
381 while (!list_empty(head)) {
382 memb = list_entry(head->next, struct dlm_member, list);
383 list_del(&memb->list);
384 if (after_del)
385 after_del(memb->nodeid);
386 kfree(memb);
387 }
388}
389
390static void remove_remote_member(int nodeid)
391{
392 if (nodeid == dlm_our_nodeid())
393 return;
394
395 dlm_midcomms_remove_member(nodeid);
396}
397
398void dlm_clear_members(struct dlm_ls *ls)
399{
400 clear_memb_list(&ls->ls_nodes, remove_remote_member);
401 ls->ls_num_nodes = 0;
402}
403
404void dlm_clear_members_gone(struct dlm_ls *ls)
405{
406 clear_memb_list(&ls->ls_nodes_gone, NULL);
407}
408
409static void make_member_array(struct dlm_ls *ls)
410{
411 struct dlm_member *memb;
412 int i, w, x = 0, total = 0, all_zero = 0, *array;
413
414 kfree(ls->ls_node_array);
415 ls->ls_node_array = NULL;
416
417 list_for_each_entry(memb, &ls->ls_nodes, list) {
418 if (memb->weight)
419 total += memb->weight;
420 }
421
422 /* all nodes revert to weight of 1 if all have weight 0 */
423
424 if (!total) {
425 total = ls->ls_num_nodes;
426 all_zero = 1;
427 }
428
429 ls->ls_total_weight = total;
430 array = kmalloc_array(total, sizeof(*array), GFP_NOFS);
431 if (!array)
432 return;
433
434 list_for_each_entry(memb, &ls->ls_nodes, list) {
435 if (!all_zero && !memb->weight)
436 continue;
437
438 if (all_zero)
439 w = 1;
440 else
441 w = memb->weight;
442
443 DLM_ASSERT(x < total, printk("total %d x %d\n", total, x););
444
445 for (i = 0; i < w; i++)
446 array[x++] = memb->nodeid;
447 }
448
449 ls->ls_node_array = array;
450}
451
452/* send a status request to all members just to establish comms connections */
453
454static int ping_members(struct dlm_ls *ls, uint64_t seq)
455{
456 struct dlm_member *memb;
457 int error = 0;
458
459 list_for_each_entry(memb, &ls->ls_nodes, list) {
460 if (dlm_recovery_stopped(ls)) {
461 error = -EINTR;
462 break;
463 }
464 error = dlm_rcom_status(ls, memb->nodeid, 0, seq);
465 if (error)
466 break;
467 }
468 if (error)
469 log_rinfo(ls, "ping_members aborted %d last nodeid %d",
470 error, ls->ls_recover_nodeid);
471 return error;
472}
473
474static void dlm_lsop_recover_prep(struct dlm_ls *ls)
475{
476 if (!ls->ls_ops || !ls->ls_ops->recover_prep)
477 return;
478 ls->ls_ops->recover_prep(ls->ls_ops_arg);
479}
480
481static void dlm_lsop_recover_slot(struct dlm_ls *ls, struct dlm_member *memb)
482{
483 struct dlm_slot slot;
484 uint32_t seq;
485 int error;
486
487 if (!ls->ls_ops || !ls->ls_ops->recover_slot)
488 return;
489
490 /* if there is no comms connection with this node
491 or the present comms connection is newer
492 than the one when this member was added, then
493 we consider the node to have failed (versus
494 being removed due to dlm_release_lockspace) */
495
496 error = dlm_comm_seq(memb->nodeid, &seq, false);
497
498 if (!error && seq == memb->comm_seq)
499 return;
500
501 slot.nodeid = memb->nodeid;
502 slot.slot = memb->slot;
503
504 ls->ls_ops->recover_slot(ls->ls_ops_arg, &slot);
505}
506
507void dlm_lsop_recover_done(struct dlm_ls *ls)
508{
509 struct dlm_member *memb;
510 struct dlm_slot *slots;
511 int i, num;
512
513 if (!ls->ls_ops || !ls->ls_ops->recover_done)
514 return;
515
516 num = ls->ls_num_nodes;
517 slots = kcalloc(num, sizeof(*slots), GFP_KERNEL);
518 if (!slots)
519 return;
520
521 i = 0;
522 list_for_each_entry(memb, &ls->ls_nodes, list) {
523 if (i == num) {
524 log_error(ls, "dlm_lsop_recover_done bad num %d", num);
525 goto out;
526 }
527 slots[i].nodeid = memb->nodeid;
528 slots[i].slot = memb->slot;
529 i++;
530 }
531
532 ls->ls_ops->recover_done(ls->ls_ops_arg, slots, num,
533 ls->ls_slot, ls->ls_generation);
534 out:
535 kfree(slots);
536}
537
538static struct dlm_config_node *find_config_node(struct dlm_recover *rv,
539 int nodeid)
540{
541 int i;
542
543 for (i = 0; i < rv->nodes_count; i++) {
544 if (rv->nodes[i].nodeid == nodeid)
545 return &rv->nodes[i];
546 }
547 return NULL;
548}
549
550int dlm_recover_members(struct dlm_ls *ls, struct dlm_recover *rv, int *neg_out)
551{
552 struct dlm_member *memb, *safe;
553 struct dlm_config_node *node;
554 int i, error, neg = 0, low = -1;
555
556 /* previously removed members that we've not finished removing need to
557 * count as a negative change so the "neg" recovery steps will happen
558 *
559 * This functionality must report all member changes to lsops or
560 * midcomms layer and must never return before.
561 */
562
563 list_for_each_entry(memb, &ls->ls_nodes_gone, list) {
564 log_rinfo(ls, "prev removed member %d", memb->nodeid);
565 neg++;
566 }
567
568 /* move departed members from ls_nodes to ls_nodes_gone */
569
570 list_for_each_entry_safe(memb, safe, &ls->ls_nodes, list) {
571 node = find_config_node(rv, memb->nodeid);
572 if (node && !node->new)
573 continue;
574
575 if (!node) {
576 log_rinfo(ls, "remove member %d", memb->nodeid);
577 } else {
578 /* removed and re-added */
579 log_rinfo(ls, "remove member %d comm_seq %u %u",
580 memb->nodeid, memb->comm_seq, node->comm_seq);
581 }
582
583 neg++;
584 list_move(&memb->list, &ls->ls_nodes_gone);
585 remove_remote_member(memb->nodeid);
586 ls->ls_num_nodes--;
587 dlm_lsop_recover_slot(ls, memb);
588 }
589
590 /* add new members to ls_nodes */
591
592 for (i = 0; i < rv->nodes_count; i++) {
593 node = &rv->nodes[i];
594 if (dlm_is_member(ls, node->nodeid))
595 continue;
596 error = dlm_add_member(ls, node);
597 if (error)
598 return error;
599
600 log_rinfo(ls, "add member %d", node->nodeid);
601 }
602
603 list_for_each_entry(memb, &ls->ls_nodes, list) {
604 if (low == -1 || memb->nodeid < low)
605 low = memb->nodeid;
606 }
607 ls->ls_low_nodeid = low;
608
609 make_member_array(ls);
610 *neg_out = neg;
611
612 error = ping_members(ls, rv->seq);
613 log_rinfo(ls, "dlm_recover_members %d nodes", ls->ls_num_nodes);
614 return error;
615}
616
617/* Userspace guarantees that dlm_ls_stop() has completed on all nodes before
618 dlm_ls_start() is called on any of them to start the new recovery. */
619
620int dlm_ls_stop(struct dlm_ls *ls)
621{
622 int new;
623
624 /*
625 * Prevent dlm_recv from being in the middle of something when we do
626 * the stop. This includes ensuring dlm_recv isn't processing a
627 * recovery message (rcom), while dlm_recoverd is aborting and
628 * resetting things from an in-progress recovery. i.e. we want
629 * dlm_recoverd to abort its recovery without worrying about dlm_recv
630 * processing an rcom at the same time. Stopping dlm_recv also makes
631 * it easy for dlm_receive_message() to check locking stopped and add a
632 * message to the requestqueue without races.
633 */
634
635 write_lock_bh(&ls->ls_recv_active);
636
637 /*
638 * Abort any recovery that's in progress (see RECOVER_STOP,
639 * dlm_recovery_stopped()) and tell any other threads running in the
640 * dlm to quit any processing (see RUNNING, dlm_locking_stopped()).
641 */
642
643 spin_lock_bh(&ls->ls_recover_lock);
644 set_bit(LSFL_RECOVER_STOP, &ls->ls_flags);
645 new = test_and_clear_bit(LSFL_RUNNING, &ls->ls_flags);
646 if (new)
647 timer_delete_sync(&ls->ls_scan_timer);
648 ls->ls_recover_seq++;
649
650 /* activate requestqueue and stop processing */
651 write_lock_bh(&ls->ls_requestqueue_lock);
652 set_bit(LSFL_RECV_MSG_BLOCKED, &ls->ls_flags);
653 write_unlock_bh(&ls->ls_requestqueue_lock);
654 spin_unlock_bh(&ls->ls_recover_lock);
655
656 /*
657 * Let dlm_recv run again, now any normal messages will be saved on the
658 * requestqueue for later.
659 */
660
661 write_unlock_bh(&ls->ls_recv_active);
662
663 /*
664 * This in_recovery lock does two things:
665 * 1) Keeps this function from returning until all threads are out
666 * of locking routines and locking is truly stopped.
667 * 2) Keeps any new requests from being processed until it's unlocked
668 * when recovery is complete.
669 */
670
671 if (new) {
672 set_bit(LSFL_RECOVER_DOWN, &ls->ls_flags);
673 wake_up_process(ls->ls_recoverd_task);
674 wait_event(ls->ls_recover_lock_wait,
675 test_bit(LSFL_RECOVER_LOCK, &ls->ls_flags));
676 }
677
678 /*
679 * The recoverd suspend/resume makes sure that dlm_recoverd (if
680 * running) has noticed RECOVER_STOP above and quit processing the
681 * previous recovery.
682 */
683
684 dlm_recoverd_suspend(ls);
685
686 spin_lock_bh(&ls->ls_recover_lock);
687 kfree(ls->ls_slots);
688 ls->ls_slots = NULL;
689 ls->ls_num_slots = 0;
690 ls->ls_slots_size = 0;
691 ls->ls_recover_status = 0;
692 spin_unlock_bh(&ls->ls_recover_lock);
693
694 dlm_recoverd_resume(ls);
695
696 if (!ls->ls_recover_begin)
697 ls->ls_recover_begin = jiffies;
698
699 /* call recover_prep ops only once and not multiple times
700 * for each possible dlm_ls_stop() when recovery is already
701 * stopped.
702 *
703 * If we successful was able to clear LSFL_RUNNING bit and
704 * it was set we know it is the first dlm_ls_stop() call.
705 */
706 if (new)
707 dlm_lsop_recover_prep(ls);
708
709 return 0;
710}
711
712int dlm_ls_start(struct dlm_ls *ls)
713{
714 struct dlm_recover *rv, *rv_old;
715 struct dlm_config_node *nodes = NULL;
716 int error, count;
717
718 rv = kzalloc(sizeof(*rv), GFP_NOFS);
719 if (!rv)
720 return -ENOMEM;
721
722 error = dlm_config_nodes(ls->ls_name, &nodes, &count);
723 if (error < 0)
724 goto fail_rv;
725
726 spin_lock_bh(&ls->ls_recover_lock);
727
728 /* the lockspace needs to be stopped before it can be started */
729
730 if (!dlm_locking_stopped(ls)) {
731 spin_unlock_bh(&ls->ls_recover_lock);
732 log_error(ls, "start ignored: lockspace running");
733 error = -EINVAL;
734 goto fail;
735 }
736
737 rv->nodes = nodes;
738 rv->nodes_count = count;
739 rv->seq = ++ls->ls_recover_seq;
740 rv_old = ls->ls_recover_args;
741 ls->ls_recover_args = rv;
742 spin_unlock_bh(&ls->ls_recover_lock);
743
744 if (rv_old) {
745 log_error(ls, "unused recovery %llx %d",
746 (unsigned long long)rv_old->seq, rv_old->nodes_count);
747 kfree(rv_old->nodes);
748 kfree(rv_old);
749 }
750
751 set_bit(LSFL_RECOVER_WORK, &ls->ls_flags);
752 wake_up_process(ls->ls_recoverd_task);
753 return 0;
754
755 fail:
756 kfree(nodes);
757 fail_rv:
758 kfree(rv);
759 return error;
760}
761