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(struct dlm_header *h)
23 if ((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 for (i = 0, ro = ro0; i < num_slots; i++, ro++) {
124 ro->ro_nodeid = le32_to_cpu(ro->ro_nodeid);
125 ro->ro_slot = le16_to_cpu(ro->ro_slot);
128 log_slots(ls, gen, num_slots, ro0, NULL, 0);
130 list_for_each_entry(memb, &ls->ls_nodes, list) {
131 for (i = 0, ro = ro0; i < num_slots; i++, ro++) {
132 if (ro->ro_nodeid != memb->nodeid)
134 memb->slot = ro->ro_slot;
135 memb->slot_prev = memb->slot;
139 if (memb->nodeid == our_nodeid) {
140 if (ls->ls_slot && ls->ls_slot != memb->slot) {
141 log_error(ls, "dlm_slots_copy_in our slot "
142 "changed %d %d", ls->ls_slot,
148 ls->ls_slot = memb->slot;
152 log_error(ls, "dlm_slots_copy_in nodeid %d no slot",
161 /* for any nodes that do not support slots, we will not have set memb->slot
162 in wait_status_all(), so memb->slot will remain -1, and we will not
163 assign slots or set ls_num_slots here */
165 int dlm_slots_assign(struct dlm_ls *ls, int *num_slots, int *slots_size,
166 struct dlm_slot **slots_out, uint32_t *gen_out)
168 struct dlm_member *memb;
169 struct dlm_slot *array;
170 int our_nodeid = dlm_our_nodeid();
171 int array_size, max_slots, i;
177 /* our own memb struct will have slot -1 gen 0 */
179 list_for_each_entry(memb, &ls->ls_nodes, list) {
180 if (memb->nodeid == our_nodeid) {
181 memb->slot = ls->ls_slot;
182 memb->generation = ls->ls_generation;
187 list_for_each_entry(memb, &ls->ls_nodes, list) {
188 if (memb->generation > gen)
189 gen = memb->generation;
191 /* node doesn't support slots */
193 if (memb->slot == -1)
196 /* node needs a slot assigned */
201 /* node has a slot assigned */
205 if (!max || max < memb->slot)
208 /* sanity check, once slot is assigned it shouldn't change */
210 if (memb->slot_prev && memb->slot && memb->slot_prev != memb->slot) {
211 log_error(ls, "nodeid %d slot changed %d %d",
212 memb->nodeid, memb->slot_prev, memb->slot);
215 memb->slot_prev = memb->slot;
218 array_size = max + need;
219 array = kcalloc(array_size, sizeof(*array), GFP_NOFS);
225 /* fill in slots (offsets) that are used */
227 list_for_each_entry(memb, &ls->ls_nodes, list) {
231 if (memb->slot > array_size) {
232 log_error(ls, "invalid slot number %d", memb->slot);
237 array[memb->slot - 1].nodeid = memb->nodeid;
238 array[memb->slot - 1].slot = memb->slot;
242 /* assign new slots from unused offsets */
244 list_for_each_entry(memb, &ls->ls_nodes, list) {
248 for (i = 0; i < array_size; i++) {
253 memb->slot_prev = memb->slot;
254 array[i].nodeid = memb->nodeid;
255 array[i].slot = memb->slot;
258 if (!ls->ls_slot && memb->nodeid == our_nodeid)
259 ls->ls_slot = memb->slot;
264 log_error(ls, "no free slot found");
272 log_slots(ls, gen, num, NULL, array, array_size);
274 max_slots = (LOWCOMMS_MAX_TX_BUFFER_LEN - sizeof(struct dlm_rcom) -
275 sizeof(struct rcom_config)) / sizeof(struct rcom_slot);
277 if (num > max_slots) {
278 log_error(ls, "num_slots %d exceeds max_slots %d",
286 *slots_size = array_size;
291 static void add_ordered_member(struct dlm_ls *ls, struct dlm_member *new)
293 struct dlm_member *memb = NULL;
294 struct list_head *tmp;
295 struct list_head *newlist = &new->list;
296 struct list_head *head = &ls->ls_nodes;
298 list_for_each(tmp, head) {
299 memb = list_entry(tmp, struct dlm_member, list);
300 if (new->nodeid < memb->nodeid)
305 list_add_tail(newlist, head);
307 /* FIXME: can use list macro here */
308 newlist->prev = tmp->prev;
310 tmp->prev->next = newlist;
315 static int dlm_add_member(struct dlm_ls *ls, struct dlm_config_node *node)
317 struct dlm_member *memb;
320 memb = kzalloc(sizeof(*memb), GFP_NOFS);
324 error = dlm_lowcomms_connect_node(node->nodeid);
330 memb->nodeid = node->nodeid;
331 memb->weight = node->weight;
332 memb->comm_seq = node->comm_seq;
333 dlm_midcomms_add_member(node->nodeid);
334 add_ordered_member(ls, memb);
339 static struct dlm_member *find_memb(struct list_head *head, int nodeid)
341 struct dlm_member *memb;
343 list_for_each_entry(memb, head, list) {
344 if (memb->nodeid == nodeid)
350 int dlm_is_member(struct dlm_ls *ls, int nodeid)
352 if (find_memb(&ls->ls_nodes, nodeid))
357 int dlm_is_removed(struct dlm_ls *ls, int nodeid)
359 if (find_memb(&ls->ls_nodes_gone, nodeid))
364 static void clear_memb_list(struct list_head *head,
365 void (*after_del)(int nodeid))
367 struct dlm_member *memb;
369 while (!list_empty(head)) {
370 memb = list_entry(head->next, struct dlm_member, list);
371 list_del(&memb->list);
373 after_del(memb->nodeid);
378 static void clear_members_cb(int nodeid)
380 dlm_midcomms_remove_member(nodeid);
383 void dlm_clear_members(struct dlm_ls *ls)
385 clear_memb_list(&ls->ls_nodes, clear_members_cb);
386 ls->ls_num_nodes = 0;
389 void dlm_clear_members_gone(struct dlm_ls *ls)
391 clear_memb_list(&ls->ls_nodes_gone, NULL);
394 static void make_member_array(struct dlm_ls *ls)
396 struct dlm_member *memb;
397 int i, w, x = 0, total = 0, all_zero = 0, *array;
399 kfree(ls->ls_node_array);
400 ls->ls_node_array = NULL;
402 list_for_each_entry(memb, &ls->ls_nodes, list) {
404 total += memb->weight;
407 /* all nodes revert to weight of 1 if all have weight 0 */
410 total = ls->ls_num_nodes;
414 ls->ls_total_weight = total;
415 array = kmalloc_array(total, sizeof(*array), GFP_NOFS);
419 list_for_each_entry(memb, &ls->ls_nodes, list) {
420 if (!all_zero && !memb->weight)
428 DLM_ASSERT(x < total, printk("total %d x %d\n", total, x););
430 for (i = 0; i < w; i++)
431 array[x++] = memb->nodeid;
434 ls->ls_node_array = array;
437 /* send a status request to all members just to establish comms connections */
439 static int ping_members(struct dlm_ls *ls)
441 struct dlm_member *memb;
444 list_for_each_entry(memb, &ls->ls_nodes, list) {
445 error = dlm_recovery_stopped(ls);
448 error = dlm_rcom_status(ls, memb->nodeid, 0);
453 log_rinfo(ls, "ping_members aborted %d last nodeid %d",
454 error, ls->ls_recover_nodeid);
458 static void dlm_lsop_recover_prep(struct dlm_ls *ls)
460 if (!ls->ls_ops || !ls->ls_ops->recover_prep)
462 ls->ls_ops->recover_prep(ls->ls_ops_arg);
465 static void dlm_lsop_recover_slot(struct dlm_ls *ls, struct dlm_member *memb)
467 struct dlm_slot slot;
471 if (!ls->ls_ops || !ls->ls_ops->recover_slot)
474 /* if there is no comms connection with this node
475 or the present comms connection is newer
476 than the one when this member was added, then
477 we consider the node to have failed (versus
478 being removed due to dlm_release_lockspace) */
480 error = dlm_comm_seq(memb->nodeid, &seq);
482 if (!error && seq == memb->comm_seq)
485 slot.nodeid = memb->nodeid;
486 slot.slot = memb->slot;
488 ls->ls_ops->recover_slot(ls->ls_ops_arg, &slot);
491 void dlm_lsop_recover_done(struct dlm_ls *ls)
493 struct dlm_member *memb;
494 struct dlm_slot *slots;
497 if (!ls->ls_ops || !ls->ls_ops->recover_done)
500 num = ls->ls_num_nodes;
501 slots = kcalloc(num, sizeof(*slots), GFP_KERNEL);
506 list_for_each_entry(memb, &ls->ls_nodes, list) {
508 log_error(ls, "dlm_lsop_recover_done bad num %d", num);
511 slots[i].nodeid = memb->nodeid;
512 slots[i].slot = memb->slot;
516 ls->ls_ops->recover_done(ls->ls_ops_arg, slots, num,
517 ls->ls_slot, ls->ls_generation);
522 static struct dlm_config_node *find_config_node(struct dlm_recover *rv,
527 for (i = 0; i < rv->nodes_count; i++) {
528 if (rv->nodes[i].nodeid == nodeid)
529 return &rv->nodes[i];
534 int dlm_recover_members(struct dlm_ls *ls, struct dlm_recover *rv, int *neg_out)
536 struct dlm_member *memb, *safe;
537 struct dlm_config_node *node;
538 int i, error, neg = 0, low = -1;
540 /* previously removed members that we've not finished removing need to
541 count as a negative change so the "neg" recovery steps will happen */
543 list_for_each_entry(memb, &ls->ls_nodes_gone, list) {
544 log_rinfo(ls, "prev removed member %d", memb->nodeid);
548 /* move departed members from ls_nodes to ls_nodes_gone */
550 list_for_each_entry_safe(memb, safe, &ls->ls_nodes, list) {
551 node = find_config_node(rv, memb->nodeid);
552 if (node && !node->new)
556 log_rinfo(ls, "remove member %d", memb->nodeid);
558 /* removed and re-added */
559 log_rinfo(ls, "remove member %d comm_seq %u %u",
560 memb->nodeid, memb->comm_seq, node->comm_seq);
564 list_move(&memb->list, &ls->ls_nodes_gone);
565 dlm_midcomms_remove_member(memb->nodeid);
567 dlm_lsop_recover_slot(ls, memb);
570 /* add new members to ls_nodes */
572 for (i = 0; i < rv->nodes_count; i++) {
573 node = &rv->nodes[i];
574 if (dlm_is_member(ls, node->nodeid))
576 dlm_add_member(ls, node);
577 log_rinfo(ls, "add member %d", node->nodeid);
580 list_for_each_entry(memb, &ls->ls_nodes, list) {
581 if (low == -1 || memb->nodeid < low)
584 ls->ls_low_nodeid = low;
586 make_member_array(ls);
589 error = ping_members(ls);
590 /* error -EINTR means that a new recovery action is triggered.
591 * We ignore this recovery action and let run the new one which might
592 * have new member configuration.
597 /* new_lockspace() may be waiting to know if the config
600 ls->ls_members_result = error;
601 complete(&ls->ls_members_done);
603 log_rinfo(ls, "dlm_recover_members %d nodes", ls->ls_num_nodes);
607 /* Userspace guarantees that dlm_ls_stop() has completed on all nodes before
608 dlm_ls_start() is called on any of them to start the new recovery. */
610 int dlm_ls_stop(struct dlm_ls *ls)
615 * Prevent dlm_recv from being in the middle of something when we do
616 * the stop. This includes ensuring dlm_recv isn't processing a
617 * recovery message (rcom), while dlm_recoverd is aborting and
618 * resetting things from an in-progress recovery. i.e. we want
619 * dlm_recoverd to abort its recovery without worrying about dlm_recv
620 * processing an rcom at the same time. Stopping dlm_recv also makes
621 * it easy for dlm_receive_message() to check locking stopped and add a
622 * message to the requestqueue without races.
625 down_write(&ls->ls_recv_active);
628 * Abort any recovery that's in progress (see RECOVER_STOP,
629 * dlm_recovery_stopped()) and tell any other threads running in the
630 * dlm to quit any processing (see RUNNING, dlm_locking_stopped()).
633 spin_lock(&ls->ls_recover_lock);
634 set_bit(LSFL_RECOVER_STOP, &ls->ls_flags);
635 new = test_and_clear_bit(LSFL_RUNNING, &ls->ls_flags);
636 ls->ls_recover_seq++;
637 spin_unlock(&ls->ls_recover_lock);
640 * Let dlm_recv run again, now any normal messages will be saved on the
641 * requestqueue for later.
644 up_write(&ls->ls_recv_active);
647 * This in_recovery lock does two things:
648 * 1) Keeps this function from returning until all threads are out
649 * of locking routines and locking is truly stopped.
650 * 2) Keeps any new requests from being processed until it's unlocked
651 * when recovery is complete.
655 set_bit(LSFL_RECOVER_DOWN, &ls->ls_flags);
656 wake_up_process(ls->ls_recoverd_task);
657 wait_event(ls->ls_recover_lock_wait,
658 test_bit(LSFL_RECOVER_LOCK, &ls->ls_flags));
662 * The recoverd suspend/resume makes sure that dlm_recoverd (if
663 * running) has noticed RECOVER_STOP above and quit processing the
667 dlm_recoverd_suspend(ls);
669 spin_lock(&ls->ls_recover_lock);
672 ls->ls_num_slots = 0;
673 ls->ls_slots_size = 0;
674 ls->ls_recover_status = 0;
675 spin_unlock(&ls->ls_recover_lock);
677 dlm_recoverd_resume(ls);
679 if (!ls->ls_recover_begin)
680 ls->ls_recover_begin = jiffies;
682 dlm_lsop_recover_prep(ls);
686 int dlm_ls_start(struct dlm_ls *ls)
688 struct dlm_recover *rv, *rv_old;
689 struct dlm_config_node *nodes = NULL;
692 rv = kzalloc(sizeof(*rv), GFP_NOFS);
696 error = dlm_config_nodes(ls->ls_name, &nodes, &count);
700 spin_lock(&ls->ls_recover_lock);
702 /* the lockspace needs to be stopped before it can be started */
704 if (!dlm_locking_stopped(ls)) {
705 spin_unlock(&ls->ls_recover_lock);
706 log_error(ls, "start ignored: lockspace running");
712 rv->nodes_count = count;
713 rv->seq = ++ls->ls_recover_seq;
714 rv_old = ls->ls_recover_args;
715 ls->ls_recover_args = rv;
716 spin_unlock(&ls->ls_recover_lock);
719 log_error(ls, "unused recovery %llx %d",
720 (unsigned long long)rv_old->seq, rv_old->nodes_count);
721 kfree(rv_old->nodes);
725 set_bit(LSFL_RECOVER_WORK, &ls->ls_flags);
726 wake_up_process(ls->ls_recoverd_task);