1 // SPDX-License-Identifier: GPL-2.0-only
2 /******************************************************************************
3 *******************************************************************************
5 ** Copyright (C) 2005-2011 Red Hat, Inc. All rights reserved.
8 *******************************************************************************
9 ******************************************************************************/
11 #include "dlm_internal.h"
12 #include "lockspace.h"
21 int dlm_slots_version(const struct dlm_header *h)
23 if ((le32_to_cpu(h->h_version) & 0x0000FFFF) < DLM_HEADER_SLOTS)
28 void dlm_slot_save(struct dlm_ls *ls, struct dlm_rcom *rc,
29 struct dlm_member *memb)
31 struct rcom_config *rf = (struct rcom_config *)rc->rc_buf;
33 if (!dlm_slots_version(&rc->rc_header))
36 memb->slot = le16_to_cpu(rf->rf_our_slot);
37 memb->generation = le32_to_cpu(rf->rf_generation);
40 void dlm_slots_copy_out(struct dlm_ls *ls, struct dlm_rcom *rc)
42 struct dlm_slot *slot;
46 ro = (struct rcom_slot *)(rc->rc_buf + sizeof(struct rcom_config));
48 /* ls_slots array is sparse, but not rcom_slots */
50 for (i = 0; i < ls->ls_slots_size; i++) {
51 slot = &ls->ls_slots[i];
54 ro->ro_nodeid = cpu_to_le32(slot->nodeid);
55 ro->ro_slot = cpu_to_le16(slot->slot);
60 #define SLOT_DEBUG_LINE 128
62 static void log_slots(struct dlm_ls *ls, uint32_t gen, int num_slots,
63 struct rcom_slot *ro0, struct dlm_slot *array,
66 char line[SLOT_DEBUG_LINE];
67 int len = SLOT_DEBUG_LINE - 1;
71 memset(line, 0, sizeof(line));
74 for (i = 0; i < array_size; i++) {
78 ret = snprintf(line + pos, len - pos, " %d:%d",
79 array[i].slot, array[i].nodeid);
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);
94 log_rinfo(ls, "generation %u slots %d%s", gen, num_slots, line);
97 int dlm_slots_copy_in(struct dlm_ls *ls)
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();
107 if (!dlm_slots_version(&rc->rc_header))
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);
115 ls->ls_generation = gen;
117 num_slots = le16_to_cpu(rf->rf_num_slots);
121 ro0 = (struct rcom_slot *)(rc->rc_buf + sizeof(struct rcom_config));
123 log_slots(ls, gen, num_slots, ro0, NULL, 0);
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)
129 memb->slot = le16_to_cpu(ro->ro_slot);
130 memb->slot_prev = memb->slot;
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,
143 ls->ls_slot = memb->slot;
147 log_error(ls, "dlm_slots_copy_in nodeid %d no slot",
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 */
160 int dlm_slots_assign(struct dlm_ls *ls, int *num_slots, int *slots_size,
161 struct dlm_slot **slots_out, uint32_t *gen_out)
163 struct dlm_member *memb;
164 struct dlm_slot *array;
165 int our_nodeid = dlm_our_nodeid();
166 int array_size, max_slots, i;
172 /* our own memb struct will have slot -1 gen 0 */
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;
182 list_for_each_entry(memb, &ls->ls_nodes, list) {
183 if (memb->generation > gen)
184 gen = memb->generation;
186 /* node doesn't support slots */
188 if (memb->slot == -1)
191 /* node needs a slot assigned */
196 /* node has a slot assigned */
200 if (!max || max < memb->slot)
203 /* sanity check, once slot is assigned it shouldn't change */
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);
210 memb->slot_prev = memb->slot;
213 array_size = max + need;
214 array = kcalloc(array_size, sizeof(*array), GFP_NOFS);
220 /* fill in slots (offsets) that are used */
222 list_for_each_entry(memb, &ls->ls_nodes, list) {
226 if (memb->slot > array_size) {
227 log_error(ls, "invalid slot number %d", memb->slot);
232 array[memb->slot - 1].nodeid = memb->nodeid;
233 array[memb->slot - 1].slot = memb->slot;
237 /* assign new slots from unused offsets */
239 list_for_each_entry(memb, &ls->ls_nodes, list) {
243 for (i = 0; i < array_size; i++) {
248 memb->slot_prev = memb->slot;
249 array[i].nodeid = memb->nodeid;
250 array[i].slot = memb->slot;
253 if (!ls->ls_slot && memb->nodeid == our_nodeid)
254 ls->ls_slot = memb->slot;
259 log_error(ls, "no free slot found");
267 log_slots(ls, gen, num, NULL, array, array_size);
269 max_slots = (DLM_MAX_APP_BUFSIZE - sizeof(struct dlm_rcom) -
270 sizeof(struct rcom_config)) / sizeof(struct rcom_slot);
272 if (num > max_slots) {
273 log_error(ls, "num_slots %d exceeds max_slots %d",
281 *slots_size = array_size;
286 static void add_ordered_member(struct dlm_ls *ls, struct dlm_member *new)
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;
293 list_for_each(tmp, head) {
294 memb = list_entry(tmp, struct dlm_member, list);
295 if (new->nodeid < memb->nodeid)
300 list_add_tail(newlist, head);
302 /* FIXME: can use list macro here */
303 newlist->prev = tmp->prev;
305 tmp->prev->next = newlist;
310 static int add_remote_member(int nodeid)
314 if (nodeid == dlm_our_nodeid())
317 error = dlm_lowcomms_connect_node(nodeid);
321 dlm_midcomms_add_member(nodeid);
325 static int dlm_add_member(struct dlm_ls *ls, struct dlm_config_node *node)
327 struct dlm_member *memb;
330 memb = kzalloc(sizeof(*memb), GFP_NOFS);
334 memb->nodeid = node->nodeid;
335 memb->weight = node->weight;
336 memb->comm_seq = node->comm_seq;
338 error = add_remote_member(node->nodeid);
344 add_ordered_member(ls, memb);
349 static struct dlm_member *find_memb(struct list_head *head, int nodeid)
351 struct dlm_member *memb;
353 list_for_each_entry(memb, head, list) {
354 if (memb->nodeid == nodeid)
360 int dlm_is_member(struct dlm_ls *ls, int nodeid)
362 if (find_memb(&ls->ls_nodes, nodeid))
367 int dlm_is_removed(struct dlm_ls *ls, int nodeid)
369 if (find_memb(&ls->ls_nodes_gone, nodeid))
374 static void clear_memb_list(struct list_head *head,
375 void (*after_del)(int nodeid))
377 struct dlm_member *memb;
379 while (!list_empty(head)) {
380 memb = list_entry(head->next, struct dlm_member, list);
381 list_del(&memb->list);
383 after_del(memb->nodeid);
388 static void remove_remote_member(int nodeid)
390 if (nodeid == dlm_our_nodeid())
393 dlm_midcomms_remove_member(nodeid);
396 void dlm_clear_members(struct dlm_ls *ls)
398 clear_memb_list(&ls->ls_nodes, remove_remote_member);
399 ls->ls_num_nodes = 0;
402 void dlm_clear_members_gone(struct dlm_ls *ls)
404 clear_memb_list(&ls->ls_nodes_gone, NULL);
407 static void make_member_array(struct dlm_ls *ls)
409 struct dlm_member *memb;
410 int i, w, x = 0, total = 0, all_zero = 0, *array;
412 kfree(ls->ls_node_array);
413 ls->ls_node_array = NULL;
415 list_for_each_entry(memb, &ls->ls_nodes, list) {
417 total += memb->weight;
420 /* all nodes revert to weight of 1 if all have weight 0 */
423 total = ls->ls_num_nodes;
427 ls->ls_total_weight = total;
428 array = kmalloc_array(total, sizeof(*array), GFP_NOFS);
432 list_for_each_entry(memb, &ls->ls_nodes, list) {
433 if (!all_zero && !memb->weight)
441 DLM_ASSERT(x < total, printk("total %d x %d\n", total, x););
443 for (i = 0; i < w; i++)
444 array[x++] = memb->nodeid;
447 ls->ls_node_array = array;
450 /* send a status request to all members just to establish comms connections */
452 static int ping_members(struct dlm_ls *ls, uint64_t seq)
454 struct dlm_member *memb;
457 list_for_each_entry(memb, &ls->ls_nodes, list) {
458 if (dlm_recovery_stopped(ls)) {
462 error = dlm_rcom_status(ls, memb->nodeid, 0, seq);
467 log_rinfo(ls, "ping_members aborted %d last nodeid %d",
468 error, ls->ls_recover_nodeid);
472 static void dlm_lsop_recover_prep(struct dlm_ls *ls)
474 if (!ls->ls_ops || !ls->ls_ops->recover_prep)
476 ls->ls_ops->recover_prep(ls->ls_ops_arg);
479 static void dlm_lsop_recover_slot(struct dlm_ls *ls, struct dlm_member *memb)
481 struct dlm_slot slot;
485 if (!ls->ls_ops || !ls->ls_ops->recover_slot)
488 /* if there is no comms connection with this node
489 or the present comms connection is newer
490 than the one when this member was added, then
491 we consider the node to have failed (versus
492 being removed due to dlm_release_lockspace) */
494 error = dlm_comm_seq(memb->nodeid, &seq);
496 if (!error && seq == memb->comm_seq)
499 slot.nodeid = memb->nodeid;
500 slot.slot = memb->slot;
502 ls->ls_ops->recover_slot(ls->ls_ops_arg, &slot);
505 void dlm_lsop_recover_done(struct dlm_ls *ls)
507 struct dlm_member *memb;
508 struct dlm_slot *slots;
511 if (!ls->ls_ops || !ls->ls_ops->recover_done)
514 num = ls->ls_num_nodes;
515 slots = kcalloc(num, sizeof(*slots), GFP_KERNEL);
520 list_for_each_entry(memb, &ls->ls_nodes, list) {
522 log_error(ls, "dlm_lsop_recover_done bad num %d", num);
525 slots[i].nodeid = memb->nodeid;
526 slots[i].slot = memb->slot;
530 ls->ls_ops->recover_done(ls->ls_ops_arg, slots, num,
531 ls->ls_slot, ls->ls_generation);
536 static struct dlm_config_node *find_config_node(struct dlm_recover *rv,
541 for (i = 0; i < rv->nodes_count; i++) {
542 if (rv->nodes[i].nodeid == nodeid)
543 return &rv->nodes[i];
548 int dlm_recover_members(struct dlm_ls *ls, struct dlm_recover *rv, int *neg_out)
550 struct dlm_member *memb, *safe;
551 struct dlm_config_node *node;
552 int i, error, neg = 0, low = -1;
554 /* previously removed members that we've not finished removing need to
555 * count as a negative change so the "neg" recovery steps will happen
557 * This functionality must report all member changes to lsops or
558 * midcomms layer and must never return before.
561 list_for_each_entry(memb, &ls->ls_nodes_gone, list) {
562 log_rinfo(ls, "prev removed member %d", memb->nodeid);
566 /* move departed members from ls_nodes to ls_nodes_gone */
568 list_for_each_entry_safe(memb, safe, &ls->ls_nodes, list) {
569 node = find_config_node(rv, memb->nodeid);
570 if (node && !node->new)
574 log_rinfo(ls, "remove member %d", memb->nodeid);
576 /* removed and re-added */
577 log_rinfo(ls, "remove member %d comm_seq %u %u",
578 memb->nodeid, memb->comm_seq, node->comm_seq);
582 list_move(&memb->list, &ls->ls_nodes_gone);
583 remove_remote_member(memb->nodeid);
585 dlm_lsop_recover_slot(ls, memb);
588 /* add new members to ls_nodes */
590 for (i = 0; i < rv->nodes_count; i++) {
591 node = &rv->nodes[i];
592 if (dlm_is_member(ls, node->nodeid))
594 error = dlm_add_member(ls, node);
598 log_rinfo(ls, "add member %d", node->nodeid);
601 list_for_each_entry(memb, &ls->ls_nodes, list) {
602 if (low == -1 || memb->nodeid < low)
605 ls->ls_low_nodeid = low;
607 make_member_array(ls);
610 error = ping_members(ls, rv->seq);
611 log_rinfo(ls, "dlm_recover_members %d nodes", ls->ls_num_nodes);
615 /* Userspace guarantees that dlm_ls_stop() has completed on all nodes before
616 dlm_ls_start() is called on any of them to start the new recovery. */
618 int dlm_ls_stop(struct dlm_ls *ls)
623 * Prevent dlm_recv from being in the middle of something when we do
624 * the stop. This includes ensuring dlm_recv isn't processing a
625 * recovery message (rcom), while dlm_recoverd is aborting and
626 * resetting things from an in-progress recovery. i.e. we want
627 * dlm_recoverd to abort its recovery without worrying about dlm_recv
628 * processing an rcom at the same time. Stopping dlm_recv also makes
629 * it easy for dlm_receive_message() to check locking stopped and add a
630 * message to the requestqueue without races.
633 down_write(&ls->ls_recv_active);
636 * Abort any recovery that's in progress (see RECOVER_STOP,
637 * dlm_recovery_stopped()) and tell any other threads running in the
638 * dlm to quit any processing (see RUNNING, dlm_locking_stopped()).
641 spin_lock(&ls->ls_recover_lock);
642 set_bit(LSFL_RECOVER_STOP, &ls->ls_flags);
643 new = test_and_clear_bit(LSFL_RUNNING, &ls->ls_flags);
644 ls->ls_recover_seq++;
645 spin_unlock(&ls->ls_recover_lock);
648 * Let dlm_recv run again, now any normal messages will be saved on the
649 * requestqueue for later.
652 up_write(&ls->ls_recv_active);
655 * This in_recovery lock does two things:
656 * 1) Keeps this function from returning until all threads are out
657 * of locking routines and locking is truly stopped.
658 * 2) Keeps any new requests from being processed until it's unlocked
659 * when recovery is complete.
663 set_bit(LSFL_RECOVER_DOWN, &ls->ls_flags);
664 wake_up_process(ls->ls_recoverd_task);
665 wait_event(ls->ls_recover_lock_wait,
666 test_bit(LSFL_RECOVER_LOCK, &ls->ls_flags));
670 * The recoverd suspend/resume makes sure that dlm_recoverd (if
671 * running) has noticed RECOVER_STOP above and quit processing the
675 dlm_recoverd_suspend(ls);
677 spin_lock(&ls->ls_recover_lock);
680 ls->ls_num_slots = 0;
681 ls->ls_slots_size = 0;
682 ls->ls_recover_status = 0;
683 spin_unlock(&ls->ls_recover_lock);
685 dlm_recoverd_resume(ls);
687 if (!ls->ls_recover_begin)
688 ls->ls_recover_begin = jiffies;
690 /* call recover_prep ops only once and not multiple times
691 * for each possible dlm_ls_stop() when recovery is already
694 * If we successful was able to clear LSFL_RUNNING bit and
695 * it was set we know it is the first dlm_ls_stop() call.
698 dlm_lsop_recover_prep(ls);
703 int dlm_ls_start(struct dlm_ls *ls)
705 struct dlm_recover *rv, *rv_old;
706 struct dlm_config_node *nodes = NULL;
709 rv = kzalloc(sizeof(*rv), GFP_NOFS);
713 error = dlm_config_nodes(ls->ls_name, &nodes, &count);
717 spin_lock(&ls->ls_recover_lock);
719 /* the lockspace needs to be stopped before it can be started */
721 if (!dlm_locking_stopped(ls)) {
722 spin_unlock(&ls->ls_recover_lock);
723 log_error(ls, "start ignored: lockspace running");
729 rv->nodes_count = count;
730 rv->seq = ++ls->ls_recover_seq;
731 rv_old = ls->ls_recover_args;
732 ls->ls_recover_args = rv;
733 spin_unlock(&ls->ls_recover_lock);
736 log_error(ls, "unused recovery %llx %d",
737 (unsigned long long)rv_old->seq, rv_old->nodes_count);
738 kfree(rv_old->nodes);
742 set_bit(LSFL_RECOVER_WORK, &ls->ls_flags);
743 wake_up_process(ls->ls_recoverd_task);