projects/linux·blob master · refs

777 lines · 17.4 KiB

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 = kzalloc_objs(*array, array_size, 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 list_add_tail(newlist, tmp);
303 }
304}
305
306static int add_remote_member(int nodeid)
307{
308 int error;
309
310 if (nodeid == dlm_our_nodeid())
311 return 0;
312
313 error = dlm_lowcomms_connect_node(nodeid);
314 if (error < 0)
315 return error;
316
317 dlm_midcomms_add_member(nodeid);
318 return 0;
319}
320
321static int dlm_add_member(struct dlm_ls *ls, struct dlm_config_node *node)
322{
323 struct dlm_member *memb;
324 int error;
325
326 memb = kzalloc_obj(*memb, GFP_NOFS);
327 if (!memb)
328 return -ENOMEM;
329
330 memb->nodeid = node->nodeid;
331 memb->weight = node->weight;
332 memb->comm_seq = node->comm_seq;
333
334 error = add_remote_member(node->nodeid);
335 if (error < 0) {
336 kfree(memb);
337 return error;
338 }
339
340 add_ordered_member(ls, memb);
341 ls->ls_num_nodes++;
342 return 0;
343}
344
345static struct dlm_member *find_memb(struct list_head *head, int nodeid)
346{
347 struct dlm_member *memb;
348
349 list_for_each_entry(memb, head, list) {
350 if (memb->nodeid == nodeid)
351 return memb;
352 }
353 return NULL;
354}
355
356int dlm_is_member(struct dlm_ls *ls, int nodeid)
357{
358 if (find_memb(&ls->ls_nodes, nodeid))
359 return 1;
360 return 0;
361}
362
363int dlm_is_removed(struct dlm_ls *ls, int nodeid)
364{
365 WARN_ON_ONCE(!nodeid || nodeid == -1);
366
367 if (find_memb(&ls->ls_nodes_gone, nodeid))
368 return 1;
369 return 0;
370}
371
372static void clear_memb_list(struct list_head *head,
373 void (*after_del)(int nodeid))
374{
375 struct dlm_member *memb;
376
377 while (!list_empty(head)) {
378 memb = list_entry(head->next, struct dlm_member, list);
379 list_del(&memb->list);
380 if (after_del)
381 after_del(memb->nodeid);
382 kfree(memb);
383 }
384}
385
386static void remove_remote_member(int nodeid)
387{
388 if (nodeid == dlm_our_nodeid())
389 return;
390
391 dlm_midcomms_remove_member(nodeid);
392}
393
394void dlm_clear_members(struct dlm_ls *ls)
395{
396 clear_memb_list(&ls->ls_nodes, remove_remote_member);
397 ls->ls_num_nodes = 0;
398}
399
400void dlm_clear_members_gone(struct dlm_ls *ls)
401{
402 clear_memb_list(&ls->ls_nodes_gone, NULL);
403}
404
405static int make_member_array(struct dlm_ls *ls)
406{
407 struct dlm_member *memb;
408 int i, w, x = 0, total = 0, all_zero = 0, *array;
409
410 kfree(ls->ls_node_array);
411 ls->ls_node_array = NULL;
412
413 list_for_each_entry(memb, &ls->ls_nodes, list) {
414 if (check_add_overflow(total, memb->weight, &total))
415 return -EOVERFLOW;
416 }
417
418 /* all nodes revert to weight of 1 if all have weight 0 */
419
420 if (!total) {
421 total = ls->ls_num_nodes;
422 all_zero = 1;
423 }
424
425 array = kmalloc_objs(*array, total, GFP_NOFS);
426 if (!array)
427 return -ENOMEM;
428
429 list_for_each_entry(memb, &ls->ls_nodes, list) {
430 if (!all_zero && !memb->weight)
431 continue;
432
433 if (all_zero)
434 w = 1;
435 else
436 w = memb->weight;
437
438 DLM_ASSERT(x < total, printk("total %d x %d\n", total, x););
439
440 for (i = 0; i < w; i++)
441 array[x++] = memb->nodeid;
442 }
443
444 ls->ls_total_weight = total;
445 ls->ls_node_array = array;
446
447 return 0;
448}
449
450/* send a status request to all members just to establish comms connections */
451
452static int ping_members(struct dlm_ls *ls, uint64_t seq)
453{
454 struct dlm_member *memb;
455 int error = 0;
456
457 list_for_each_entry(memb, &ls->ls_nodes, list) {
458 if (dlm_recovery_stopped(ls)) {
459 error = -EINTR;
460 break;
461 }
462 error = dlm_rcom_status(ls, memb->nodeid, 0, seq);
463 if (error)
464 break;
465 }
466 if (error)
467 log_rinfo(ls, "ping_members aborted %d last nodeid %d",
468 error, ls->ls_recover_nodeid);
469 return error;
470}
471
472static void dlm_lsop_recover_prep(struct dlm_ls *ls)
473{
474 if (!ls->ls_ops || !ls->ls_ops->recover_prep)
475 return;
476 ls->ls_ops->recover_prep(ls->ls_ops_arg);
477}
478
479static void dlm_lsop_recover_slot(struct dlm_ls *ls, struct dlm_member *memb,
480 unsigned int release_recover)
481{
482 struct dlm_slot slot;
483 uint32_t seq;
484 int error;
485
486 if (!ls->ls_ops || !ls->ls_ops->recover_slot)
487 return;
488
489 /* if there is no comms connection with this node
490 or the present comms connection is newer
491 than the one when this member was added, then
492 we consider the node to have failed (versus
493 being removed due to dlm_release_lockspace) */
494
495 error = dlm_comm_seq(memb->nodeid, &seq, false);
496
497 if (!release_recover && !error && seq == memb->comm_seq)
498 return;
499
500 slot.nodeid = memb->nodeid;
501 slot.slot = memb->slot;
502
503 ls->ls_ops->recover_slot(ls->ls_ops_arg, &slot);
504}
505
506void dlm_lsop_recover_done(struct dlm_ls *ls)
507{
508 struct dlm_member *memb;
509 struct dlm_slot *slots;
510 int i, num;
511
512 if (!ls->ls_ops || !ls->ls_ops->recover_done)
513 return;
514
515 num = ls->ls_num_nodes;
516 slots = kzalloc_objs(*slots, num);
517 if (!slots)
518 return;
519
520 i = 0;
521 list_for_each_entry(memb, &ls->ls_nodes, list) {
522 if (i == num) {
523 log_error(ls, "dlm_lsop_recover_done bad num %d", num);
524 goto out;
525 }
526 slots[i].nodeid = memb->nodeid;
527 slots[i].slot = memb->slot;
528 i++;
529 }
530
531 ls->ls_ops->recover_done(ls->ls_ops_arg, slots, num,
532 ls->ls_slot, ls->ls_generation);
533 out:
534 kfree(slots);
535}
536
537static struct dlm_config_node *find_config_node(struct dlm_recover *rv,
538 int nodeid)
539{
540 int i;
541
542 for (i = 0; i < rv->nodes_count; i++) {
543 if (rv->nodes[i].nodeid == nodeid)
544 return &rv->nodes[i];
545 }
546 return NULL;
547}
548
549int dlm_recover_members(struct dlm_ls *ls, struct dlm_recover *rv, int *neg_out)
550{
551 struct dlm_member *memb, *safe;
552 struct dlm_config_node *node;
553 int i, error, neg = 0, low = -1;
554 unsigned int release_recover;
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) {
573 log_error(ls, "remove member %d invalid",
574 memb->nodeid);
575 return -EFAULT;
576 }
577
578 if (!node->new && !node->gone)
579 continue;
580
581 release_recover = 0;
582
583 if (node->gone) {
584 release_recover = node->release_recover;
585 log_rinfo(ls, "remove member %d%s", memb->nodeid,
586 release_recover ? " (release_recover)" : "");
587 } else {
588 /* removed and re-added */
589 log_rinfo(ls, "remove member %d comm_seq %u %u",
590 memb->nodeid, memb->comm_seq, node->comm_seq);
591 }
592
593 neg++;
594 list_move(&memb->list, &ls->ls_nodes_gone);
595 remove_remote_member(memb->nodeid);
596 ls->ls_num_nodes--;
597 dlm_lsop_recover_slot(ls, memb, release_recover);
598 }
599
600 /* add new members to ls_nodes */
601
602 for (i = 0; i < rv->nodes_count; i++) {
603 node = &rv->nodes[i];
604 if (node->gone)
605 continue;
606
607 if (dlm_is_member(ls, node->nodeid))
608 continue;
609 error = dlm_add_member(ls, node);
610 if (error)
611 return error;
612
613 log_rinfo(ls, "add member %d", node->nodeid);
614 }
615
616 list_for_each_entry(memb, &ls->ls_nodes, list) {
617 if (low == -1 || memb->nodeid < low)
618 low = memb->nodeid;
619 }
620 ls->ls_low_nodeid = low;
621
622 error = make_member_array(ls);
623 if (error)
624 return error;
625
626 *neg_out = neg;
627
628 error = ping_members(ls, rv->seq);
629 log_rinfo(ls, "dlm_recover_members %d nodes", ls->ls_num_nodes);
630 return error;
631}
632
633/* Userspace guarantees that dlm_ls_stop() has completed on all nodes before
634 dlm_ls_start() is called on any of them to start the new recovery. */
635
636int dlm_ls_stop(struct dlm_ls *ls)
637{
638 int new;
639
640 /*
641 * Prevent dlm_recv from being in the middle of something when we do
642 * the stop. This includes ensuring dlm_recv isn't processing a
643 * recovery message (rcom), while dlm_recoverd is aborting and
644 * resetting things from an in-progress recovery. i.e. we want
645 * dlm_recoverd to abort its recovery without worrying about dlm_recv
646 * processing an rcom at the same time. Stopping dlm_recv also makes
647 * it easy for dlm_receive_message() to check locking stopped and add a
648 * message to the requestqueue without races.
649 */
650
651 write_lock_bh(&ls->ls_recv_active);
652
653 /*
654 * Abort any recovery that's in progress (see RECOVER_STOP,
655 * dlm_recovery_stopped()) and tell any other threads running in the
656 * dlm to quit any processing (see RUNNING, dlm_locking_stopped()).
657 */
658
659 spin_lock_bh(&ls->ls_recover_lock);
660 set_bit(LSFL_RECOVER_STOP, &ls->ls_flags);
661 new = test_and_clear_bit(LSFL_RUNNING, &ls->ls_flags);
662 if (new)
663 timer_delete_sync(&ls->ls_scan_timer);
664 ls->ls_recover_seq++;
665
666 /* activate requestqueue and stop processing */
667 write_lock_bh(&ls->ls_requestqueue_lock);
668 set_bit(LSFL_RECV_MSG_BLOCKED, &ls->ls_flags);
669 write_unlock_bh(&ls->ls_requestqueue_lock);
670 spin_unlock_bh(&ls->ls_recover_lock);
671
672 /*
673 * Let dlm_recv run again, now any normal messages will be saved on the
674 * requestqueue for later.
675 */
676
677 write_unlock_bh(&ls->ls_recv_active);
678
679 /*
680 * This in_recovery lock does two things:
681 * 1) Keeps this function from returning until all threads are out
682 * of locking routines and locking is truly stopped.
683 * 2) Keeps any new requests from being processed until it's unlocked
684 * when recovery is complete.
685 */
686
687 if (new) {
688 set_bit(LSFL_RECOVER_DOWN, &ls->ls_flags);
689 wake_up_process(ls->ls_recoverd_task);
690 wait_event(ls->ls_recover_lock_wait,
691 test_bit(LSFL_RECOVER_LOCK, &ls->ls_flags));
692 }
693
694 /*
695 * The recoverd suspend/resume makes sure that dlm_recoverd (if
696 * running) has noticed RECOVER_STOP above and quit processing the
697 * previous recovery.
698 */
699
700 dlm_recoverd_suspend(ls);
701
702 spin_lock_bh(&ls->ls_recover_lock);
703 kfree(ls->ls_slots);
704 ls->ls_slots = NULL;
705 ls->ls_num_slots = 0;
706 ls->ls_slots_size = 0;
707 ls->ls_recover_status = 0;
708 spin_unlock_bh(&ls->ls_recover_lock);
709
710 dlm_recoverd_resume(ls);
711
712 if (!ls->ls_recover_begin)
713 ls->ls_recover_begin = jiffies;
714
715 /* call recover_prep ops only once and not multiple times
716 * for each possible dlm_ls_stop() when recovery is already
717 * stopped.
718 *
719 * If we successful was able to clear LSFL_RUNNING bit and
720 * it was set we know it is the first dlm_ls_stop() call.
721 */
722 if (new)
723 dlm_lsop_recover_prep(ls);
724
725 return 0;
726}
727
728int dlm_ls_start(struct dlm_ls *ls)
729{
730 struct dlm_recover *rv, *rv_old;
731 struct dlm_config_node *nodes = NULL;
732 int error, count;
733
734 rv = kzalloc_obj(*rv, GFP_NOFS);
735 if (!rv)
736 return -ENOMEM;
737
738 error = dlm_config_nodes(ls->ls_name, &nodes, &count);
739 if (error < 0)
740 goto fail_rv;
741
742 spin_lock_bh(&ls->ls_recover_lock);
743
744 /* the lockspace needs to be stopped before it can be started */
745
746 if (!dlm_locking_stopped(ls)) {
747 spin_unlock_bh(&ls->ls_recover_lock);
748 log_error(ls, "start ignored: lockspace running");
749 error = -EINVAL;
750 goto fail;
751 }
752
753 rv->nodes = nodes;
754 rv->nodes_count = count;
755 rv->seq = ++ls->ls_recover_seq;
756 rv_old = ls->ls_recover_args;
757 ls->ls_recover_args = rv;
758 spin_unlock_bh(&ls->ls_recover_lock);
759
760 if (rv_old) {
761 log_error(ls, "unused recovery %llx %d",
762 (unsigned long long)rv_old->seq, rv_old->nodes_count);
763 kfree(rv_old->nodes);
764 kfree(rv_old);
765 }
766
767 set_bit(LSFL_RECOVER_WORK, &ls->ls_flags);
768 wake_up_process(ls->ls_recoverd_task);
769 return 0;
770
771 fail:
772 kfree(nodes);
773 fail_rv:
774 kfree(rv);
775 return error;
776}
777