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 ((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 dlm_add_member(struct dlm_ls *ls, struct dlm_config_node *node)
312 struct dlm_member *memb;
315 memb = kzalloc(sizeof(*memb), GFP_NOFS);
319 error = dlm_lowcomms_connect_node(node->nodeid);
325 memb->nodeid = node->nodeid;
326 memb->weight = node->weight;
327 memb->comm_seq = node->comm_seq;
328 dlm_midcomms_add_member(node->nodeid);
329 add_ordered_member(ls, memb);
334 static struct dlm_member *find_memb(struct list_head *head, int nodeid)
336 struct dlm_member *memb;
338 list_for_each_entry(memb, head, list) {
339 if (memb->nodeid == nodeid)
345 int dlm_is_member(struct dlm_ls *ls, int nodeid)
347 if (find_memb(&ls->ls_nodes, nodeid))
352 int dlm_is_removed(struct dlm_ls *ls, int nodeid)
354 if (find_memb(&ls->ls_nodes_gone, nodeid))
359 static void clear_memb_list(struct list_head *head,
360 void (*after_del)(int nodeid))
362 struct dlm_member *memb;
364 while (!list_empty(head)) {
365 memb = list_entry(head->next, struct dlm_member, list);
366 list_del(&memb->list);
368 after_del(memb->nodeid);
373 static void clear_members_cb(int nodeid)
375 dlm_midcomms_remove_member(nodeid);
378 void dlm_clear_members(struct dlm_ls *ls)
380 clear_memb_list(&ls->ls_nodes, clear_members_cb);
381 ls->ls_num_nodes = 0;
384 void dlm_clear_members_gone(struct dlm_ls *ls)
386 clear_memb_list(&ls->ls_nodes_gone, NULL);
389 static void make_member_array(struct dlm_ls *ls)
391 struct dlm_member *memb;
392 int i, w, x = 0, total = 0, all_zero = 0, *array;
394 kfree(ls->ls_node_array);
395 ls->ls_node_array = NULL;
397 list_for_each_entry(memb, &ls->ls_nodes, list) {
399 total += memb->weight;
402 /* all nodes revert to weight of 1 if all have weight 0 */
405 total = ls->ls_num_nodes;
409 ls->ls_total_weight = total;
410 array = kmalloc_array(total, sizeof(*array), GFP_NOFS);
414 list_for_each_entry(memb, &ls->ls_nodes, list) {
415 if (!all_zero && !memb->weight)
423 DLM_ASSERT(x < total, printk("total %d x %d\n", total, x););
425 for (i = 0; i < w; i++)
426 array[x++] = memb->nodeid;
429 ls->ls_node_array = array;
432 /* send a status request to all members just to establish comms connections */
434 static int ping_members(struct dlm_ls *ls)
436 struct dlm_member *memb;
439 list_for_each_entry(memb, &ls->ls_nodes, list) {
440 if (dlm_recovery_stopped(ls)) {
444 error = dlm_rcom_status(ls, memb->nodeid, 0);
449 log_rinfo(ls, "ping_members aborted %d last nodeid %d",
450 error, ls->ls_recover_nodeid);
454 static void dlm_lsop_recover_prep(struct dlm_ls *ls)
456 if (!ls->ls_ops || !ls->ls_ops->recover_prep)
458 ls->ls_ops->recover_prep(ls->ls_ops_arg);
461 static void dlm_lsop_recover_slot(struct dlm_ls *ls, struct dlm_member *memb)
463 struct dlm_slot slot;
467 if (!ls->ls_ops || !ls->ls_ops->recover_slot)
470 /* if there is no comms connection with this node
471 or the present comms connection is newer
472 than the one when this member was added, then
473 we consider the node to have failed (versus
474 being removed due to dlm_release_lockspace) */
476 error = dlm_comm_seq(memb->nodeid, &seq);
478 if (!error && seq == memb->comm_seq)
481 slot.nodeid = memb->nodeid;
482 slot.slot = memb->slot;
484 ls->ls_ops->recover_slot(ls->ls_ops_arg, &slot);
487 void dlm_lsop_recover_done(struct dlm_ls *ls)
489 struct dlm_member *memb;
490 struct dlm_slot *slots;
493 if (!ls->ls_ops || !ls->ls_ops->recover_done)
496 num = ls->ls_num_nodes;
497 slots = kcalloc(num, sizeof(*slots), GFP_KERNEL);
502 list_for_each_entry(memb, &ls->ls_nodes, list) {
504 log_error(ls, "dlm_lsop_recover_done bad num %d", num);
507 slots[i].nodeid = memb->nodeid;
508 slots[i].slot = memb->slot;
512 ls->ls_ops->recover_done(ls->ls_ops_arg, slots, num,
513 ls->ls_slot, ls->ls_generation);
518 static struct dlm_config_node *find_config_node(struct dlm_recover *rv,
523 for (i = 0; i < rv->nodes_count; i++) {
524 if (rv->nodes[i].nodeid == nodeid)
525 return &rv->nodes[i];
530 int dlm_recover_members(struct dlm_ls *ls, struct dlm_recover *rv, int *neg_out)
532 struct dlm_member *memb, *safe;
533 struct dlm_config_node *node;
534 int i, error, neg = 0, low = -1;
536 /* previously removed members that we've not finished removing need to
537 count as a negative change so the "neg" recovery steps will happen */
539 list_for_each_entry(memb, &ls->ls_nodes_gone, list) {
540 log_rinfo(ls, "prev removed member %d", memb->nodeid);
544 /* move departed members from ls_nodes to ls_nodes_gone */
546 list_for_each_entry_safe(memb, safe, &ls->ls_nodes, list) {
547 node = find_config_node(rv, memb->nodeid);
548 if (node && !node->new)
552 log_rinfo(ls, "remove member %d", memb->nodeid);
554 /* removed and re-added */
555 log_rinfo(ls, "remove member %d comm_seq %u %u",
556 memb->nodeid, memb->comm_seq, node->comm_seq);
560 list_move(&memb->list, &ls->ls_nodes_gone);
561 dlm_midcomms_remove_member(memb->nodeid);
563 dlm_lsop_recover_slot(ls, memb);
566 /* add new members to ls_nodes */
568 for (i = 0; i < rv->nodes_count; i++) {
569 node = &rv->nodes[i];
570 if (dlm_is_member(ls, node->nodeid))
572 dlm_add_member(ls, node);
573 log_rinfo(ls, "add member %d", node->nodeid);
576 list_for_each_entry(memb, &ls->ls_nodes, list) {
577 if (low == -1 || memb->nodeid < low)
580 ls->ls_low_nodeid = low;
582 make_member_array(ls);
585 error = ping_members(ls);
586 /* error -EINTR means that a new recovery action is triggered.
587 * We ignore this recovery action and let run the new one which might
588 * have new member configuration.
593 /* new_lockspace() may be waiting to know if the config
596 ls->ls_members_result = error;
597 complete(&ls->ls_members_done);
599 log_rinfo(ls, "dlm_recover_members %d nodes", ls->ls_num_nodes);
603 /* Userspace guarantees that dlm_ls_stop() has completed on all nodes before
604 dlm_ls_start() is called on any of them to start the new recovery. */
606 int dlm_ls_stop(struct dlm_ls *ls)
611 * Prevent dlm_recv from being in the middle of something when we do
612 * the stop. This includes ensuring dlm_recv isn't processing a
613 * recovery message (rcom), while dlm_recoverd is aborting and
614 * resetting things from an in-progress recovery. i.e. we want
615 * dlm_recoverd to abort its recovery without worrying about dlm_recv
616 * processing an rcom at the same time. Stopping dlm_recv also makes
617 * it easy for dlm_receive_message() to check locking stopped and add a
618 * message to the requestqueue without races.
621 down_write(&ls->ls_recv_active);
624 * Abort any recovery that's in progress (see RECOVER_STOP,
625 * dlm_recovery_stopped()) and tell any other threads running in the
626 * dlm to quit any processing (see RUNNING, dlm_locking_stopped()).
629 spin_lock(&ls->ls_recover_lock);
630 set_bit(LSFL_RECOVER_STOP, &ls->ls_flags);
631 new = test_and_clear_bit(LSFL_RUNNING, &ls->ls_flags);
632 ls->ls_recover_seq++;
633 spin_unlock(&ls->ls_recover_lock);
636 * Let dlm_recv run again, now any normal messages will be saved on the
637 * requestqueue for later.
640 up_write(&ls->ls_recv_active);
643 * This in_recovery lock does two things:
644 * 1) Keeps this function from returning until all threads are out
645 * of locking routines and locking is truly stopped.
646 * 2) Keeps any new requests from being processed until it's unlocked
647 * when recovery is complete.
651 set_bit(LSFL_RECOVER_DOWN, &ls->ls_flags);
652 wake_up_process(ls->ls_recoverd_task);
653 wait_event(ls->ls_recover_lock_wait,
654 test_bit(LSFL_RECOVER_LOCK, &ls->ls_flags));
658 * The recoverd suspend/resume makes sure that dlm_recoverd (if
659 * running) has noticed RECOVER_STOP above and quit processing the
663 dlm_recoverd_suspend(ls);
665 spin_lock(&ls->ls_recover_lock);
668 ls->ls_num_slots = 0;
669 ls->ls_slots_size = 0;
670 ls->ls_recover_status = 0;
671 spin_unlock(&ls->ls_recover_lock);
673 dlm_recoverd_resume(ls);
675 if (!ls->ls_recover_begin)
676 ls->ls_recover_begin = jiffies;
678 dlm_lsop_recover_prep(ls);
682 int dlm_ls_start(struct dlm_ls *ls)
684 struct dlm_recover *rv, *rv_old;
685 struct dlm_config_node *nodes = NULL;
688 rv = kzalloc(sizeof(*rv), GFP_NOFS);
692 error = dlm_config_nodes(ls->ls_name, &nodes, &count);
696 spin_lock(&ls->ls_recover_lock);
698 /* the lockspace needs to be stopped before it can be started */
700 if (!dlm_locking_stopped(ls)) {
701 spin_unlock(&ls->ls_recover_lock);
702 log_error(ls, "start ignored: lockspace running");
708 rv->nodes_count = count;
709 rv->seq = ++ls->ls_recover_seq;
710 rv_old = ls->ls_recover_args;
711 ls->ls_recover_args = rv;
712 spin_unlock(&ls->ls_recover_lock);
715 log_error(ls, "unused recovery %llx %d",
716 (unsigned long long)rv_old->seq, rv_old->nodes_count);
717 kfree(rv_old->nodes);
721 set_bit(LSFL_RECOVER_WORK, &ls->ls_flags);
722 wake_up_process(ls->ls_recoverd_task);