1 // SPDX-License-Identifier: GPL-2.0-only
2 /******************************************************************************
3 *******************************************************************************
5 ** Copyright (C) Sistina Software, Inc. 1997-2003 All rights reserved.
6 ** Copyright (C) 2004-2005 Red Hat, Inc. All rights reserved.
9 *******************************************************************************
10 ******************************************************************************/
12 #include "dlm_internal.h"
13 #include "lockspace.h"
26 * Recovery waiting routines: these functions wait for a particular reply from
27 * a remote node, or for the remote node to report a certain status. They need
28 * to abort if the lockspace is stopped indicating a node has failed (perhaps
29 * the one being waited for).
33 * Wait until given function returns non-zero or lockspace is stopped
34 * (LS_RECOVERY_STOP set due to failure of a node in ls_nodes). When another
35 * function thinks it could have completed the waited-on task, they should wake
36 * up ls_wait_general to get an immediate response rather than waiting for the
37 * timeout. This uses a timeout so it can check periodically if the wait
38 * should abort due to node failure (which doesn't cause a wake_up).
39 * This should only be called by the dlm_recoverd thread.
42 int dlm_wait_function(struct dlm_ls *ls, int (*testfn) (struct dlm_ls *ls))
48 rv = wait_event_timeout(ls->ls_wait_general,
49 testfn(ls) || dlm_recovery_stopped(ls),
50 dlm_config.ci_recover_timer * HZ);
53 if (test_bit(LSFL_RCOM_WAIT, &ls->ls_flags)) {
54 log_debug(ls, "dlm_wait_function timed out");
59 if (dlm_recovery_stopped(ls)) {
60 log_debug(ls, "dlm_wait_function aborted");
67 * An efficient way for all nodes to wait for all others to have a certain
68 * status. The node with the lowest nodeid polls all the others for their
69 * status (wait_status_all) and all the others poll the node with the low id
70 * for its accumulated result (wait_status_low). When all nodes have set
71 * status flag X, then status flag X_ALL will be set on the low nodeid.
74 uint32_t dlm_recover_status(struct dlm_ls *ls)
77 spin_lock(&ls->ls_recover_lock);
78 status = ls->ls_recover_status;
79 spin_unlock(&ls->ls_recover_lock);
83 static void _set_recover_status(struct dlm_ls *ls, uint32_t status)
85 ls->ls_recover_status |= status;
88 void dlm_set_recover_status(struct dlm_ls *ls, uint32_t status)
90 spin_lock(&ls->ls_recover_lock);
91 _set_recover_status(ls, status);
92 spin_unlock(&ls->ls_recover_lock);
95 static int wait_status_all(struct dlm_ls *ls, uint32_t wait_status,
98 struct dlm_rcom *rc = ls->ls_recover_buf;
99 struct dlm_member *memb;
100 int error = 0, delay;
102 list_for_each_entry(memb, &ls->ls_nodes, list) {
105 if (dlm_recovery_stopped(ls)) {
110 error = dlm_rcom_status(ls, memb->nodeid, 0);
115 dlm_slot_save(ls, rc, memb);
117 if (rc->rc_result & wait_status)
128 static int wait_status_low(struct dlm_ls *ls, uint32_t wait_status,
129 uint32_t status_flags)
131 struct dlm_rcom *rc = ls->ls_recover_buf;
132 int error = 0, delay = 0, nodeid = ls->ls_low_nodeid;
135 if (dlm_recovery_stopped(ls)) {
140 error = dlm_rcom_status(ls, nodeid, status_flags);
144 if (rc->rc_result & wait_status)
154 static int wait_status(struct dlm_ls *ls, uint32_t status)
156 uint32_t status_all = status << 1;
159 if (ls->ls_low_nodeid == dlm_our_nodeid()) {
160 error = wait_status_all(ls, status, 0);
162 dlm_set_recover_status(ls, status_all);
164 error = wait_status_low(ls, status_all, 0);
169 int dlm_recover_members_wait(struct dlm_ls *ls)
171 struct dlm_member *memb;
172 struct dlm_slot *slots;
173 int num_slots, slots_size;
177 list_for_each_entry(memb, &ls->ls_nodes, list) {
179 memb->generation = 0;
182 if (ls->ls_low_nodeid == dlm_our_nodeid()) {
183 error = wait_status_all(ls, DLM_RS_NODES, 1);
187 /* slots array is sparse, slots_size may be > num_slots */
189 rv = dlm_slots_assign(ls, &num_slots, &slots_size, &slots, &gen);
191 spin_lock(&ls->ls_recover_lock);
192 _set_recover_status(ls, DLM_RS_NODES_ALL);
193 ls->ls_num_slots = num_slots;
194 ls->ls_slots_size = slots_size;
195 ls->ls_slots = slots;
196 ls->ls_generation = gen;
197 spin_unlock(&ls->ls_recover_lock);
199 dlm_set_recover_status(ls, DLM_RS_NODES_ALL);
202 error = wait_status_low(ls, DLM_RS_NODES_ALL, DLM_RSF_NEED_SLOTS);
206 dlm_slots_copy_in(ls);
212 int dlm_recover_directory_wait(struct dlm_ls *ls)
214 return wait_status(ls, DLM_RS_DIR);
217 int dlm_recover_locks_wait(struct dlm_ls *ls)
219 return wait_status(ls, DLM_RS_LOCKS);
222 int dlm_recover_done_wait(struct dlm_ls *ls)
224 return wait_status(ls, DLM_RS_DONE);
228 * The recover_list contains all the rsb's for which we've requested the new
229 * master nodeid. As replies are returned from the resource directories the
230 * rsb's are removed from the list. When the list is empty we're done.
232 * The recover_list is later similarly used for all rsb's for which we've sent
233 * new lkb's and need to receive new corresponding lkid's.
235 * We use the address of the rsb struct as a simple local identifier for the
236 * rsb so we can match an rcom reply with the rsb it was sent for.
239 static int recover_list_empty(struct dlm_ls *ls)
243 spin_lock(&ls->ls_recover_list_lock);
244 empty = list_empty(&ls->ls_recover_list);
245 spin_unlock(&ls->ls_recover_list_lock);
250 static void recover_list_add(struct dlm_rsb *r)
252 struct dlm_ls *ls = r->res_ls;
254 spin_lock(&ls->ls_recover_list_lock);
255 if (list_empty(&r->res_recover_list)) {
256 list_add_tail(&r->res_recover_list, &ls->ls_recover_list);
257 ls->ls_recover_list_count++;
260 spin_unlock(&ls->ls_recover_list_lock);
263 static void recover_list_del(struct dlm_rsb *r)
265 struct dlm_ls *ls = r->res_ls;
267 spin_lock(&ls->ls_recover_list_lock);
268 list_del_init(&r->res_recover_list);
269 ls->ls_recover_list_count--;
270 spin_unlock(&ls->ls_recover_list_lock);
275 static void recover_list_clear(struct dlm_ls *ls)
277 struct dlm_rsb *r, *s;
279 spin_lock(&ls->ls_recover_list_lock);
280 list_for_each_entry_safe(r, s, &ls->ls_recover_list, res_recover_list) {
281 list_del_init(&r->res_recover_list);
282 r->res_recover_locks_count = 0;
284 ls->ls_recover_list_count--;
287 if (ls->ls_recover_list_count != 0) {
288 log_error(ls, "warning: recover_list_count %d",
289 ls->ls_recover_list_count);
290 ls->ls_recover_list_count = 0;
292 spin_unlock(&ls->ls_recover_list_lock);
295 static int recover_idr_empty(struct dlm_ls *ls)
299 spin_lock(&ls->ls_recover_idr_lock);
300 if (ls->ls_recover_list_count)
302 spin_unlock(&ls->ls_recover_idr_lock);
307 static int recover_idr_add(struct dlm_rsb *r)
309 struct dlm_ls *ls = r->res_ls;
312 idr_preload(GFP_NOFS);
313 spin_lock(&ls->ls_recover_idr_lock);
318 rv = idr_alloc(&ls->ls_recover_idr, r, 1, 0, GFP_NOWAIT);
323 ls->ls_recover_list_count++;
327 spin_unlock(&ls->ls_recover_idr_lock);
332 static void recover_idr_del(struct dlm_rsb *r)
334 struct dlm_ls *ls = r->res_ls;
336 spin_lock(&ls->ls_recover_idr_lock);
337 idr_remove(&ls->ls_recover_idr, r->res_id);
339 ls->ls_recover_list_count--;
340 spin_unlock(&ls->ls_recover_idr_lock);
345 static struct dlm_rsb *recover_idr_find(struct dlm_ls *ls, uint64_t id)
349 spin_lock(&ls->ls_recover_idr_lock);
350 r = idr_find(&ls->ls_recover_idr, (int)id);
351 spin_unlock(&ls->ls_recover_idr_lock);
355 static void recover_idr_clear(struct dlm_ls *ls)
360 spin_lock(&ls->ls_recover_idr_lock);
362 idr_for_each_entry(&ls->ls_recover_idr, r, id) {
363 idr_remove(&ls->ls_recover_idr, id);
365 r->res_recover_locks_count = 0;
366 ls->ls_recover_list_count--;
371 if (ls->ls_recover_list_count != 0) {
372 log_error(ls, "warning: recover_list_count %d",
373 ls->ls_recover_list_count);
374 ls->ls_recover_list_count = 0;
376 spin_unlock(&ls->ls_recover_idr_lock);
380 /* Master recovery: find new master node for rsb's that were
381 mastered on nodes that have been removed.
385 dlm_send_rcom_lookup -> receive_rcom_lookup
387 receive_rcom_lookup_reply <-
388 dlm_recover_master_reply
395 * Set the lock master for all LKBs in a lock queue
396 * If we are the new master of the rsb, we may have received new
397 * MSTCPY locks from other nodes already which we need to ignore
398 * when setting the new nodeid.
401 static void set_lock_master(struct list_head *queue, int nodeid)
405 list_for_each_entry(lkb, queue, lkb_statequeue) {
406 if (!(lkb->lkb_flags & DLM_IFL_MSTCPY)) {
407 lkb->lkb_nodeid = nodeid;
413 static void set_master_lkbs(struct dlm_rsb *r)
415 set_lock_master(&r->res_grantqueue, r->res_nodeid);
416 set_lock_master(&r->res_convertqueue, r->res_nodeid);
417 set_lock_master(&r->res_waitqueue, r->res_nodeid);
421 * Propagate the new master nodeid to locks
422 * The NEW_MASTER flag tells dlm_recover_locks() which rsb's to consider.
423 * The NEW_MASTER2 flag tells recover_lvb() and recover_grant() which
427 static void set_new_master(struct dlm_rsb *r)
430 rsb_set_flag(r, RSB_NEW_MASTER);
431 rsb_set_flag(r, RSB_NEW_MASTER2);
435 * We do async lookups on rsb's that need new masters. The rsb's
436 * waiting for a lookup reply are kept on the recover_list.
438 * Another node recovering the master may have sent us a rcom lookup,
439 * and our dlm_master_lookup() set it as the new master, along with
440 * NEW_MASTER so that we'll recover it here (this implies dir_nodeid
441 * equals our_nodeid below).
444 static int recover_master(struct dlm_rsb *r, unsigned int *count)
446 struct dlm_ls *ls = r->res_ls;
447 int our_nodeid, dir_nodeid;
454 is_removed = dlm_is_removed(ls, r->res_nodeid);
456 if (!is_removed && !rsb_flag(r, RSB_NEW_MASTER))
459 our_nodeid = dlm_our_nodeid();
460 dir_nodeid = dlm_dir_nodeid(r);
462 if (dir_nodeid == our_nodeid) {
464 r->res_master_nodeid = our_nodeid;
468 /* set master of lkbs to ourself when is_removed, or to
469 another new master which we set along with NEW_MASTER
470 in dlm_master_lookup */
475 error = dlm_send_rcom_lookup(r, dir_nodeid);
483 * All MSTCPY locks are purged and rebuilt, even if the master stayed the same.
484 * This is necessary because recovery can be started, aborted and restarted,
485 * causing the master nodeid to briefly change during the aborted recovery, and
486 * change back to the original value in the second recovery. The MSTCPY locks
487 * may or may not have been purged during the aborted recovery. Another node
488 * with an outstanding request in waiters list and a request reply saved in the
489 * requestqueue, cannot know whether it should ignore the reply and resend the
490 * request, or accept the reply and complete the request. It must do the
491 * former if the remote node purged MSTCPY locks, and it must do the later if
492 * the remote node did not. This is solved by always purging MSTCPY locks, in
493 * which case, the request reply would always be ignored and the request
497 static int recover_master_static(struct dlm_rsb *r, unsigned int *count)
499 int dir_nodeid = dlm_dir_nodeid(r);
500 int new_master = dir_nodeid;
502 if (dir_nodeid == dlm_our_nodeid())
505 dlm_purge_mstcpy_locks(r);
506 r->res_master_nodeid = dir_nodeid;
507 r->res_nodeid = new_master;
514 * Go through local root resources and for each rsb which has a master which
515 * has departed, get the new master nodeid from the directory. The dir will
516 * assign mastery to the first node to look up the new master. That means
517 * we'll discover in this lookup if we're the new master of any rsb's.
519 * We fire off all the dir lookup requests individually and asynchronously to
520 * the correct dir node.
523 int dlm_recover_masters(struct dlm_ls *ls)
526 unsigned int total = 0;
527 unsigned int count = 0;
528 int nodir = dlm_no_directory(ls);
531 log_rinfo(ls, "dlm_recover_masters");
533 down_read(&ls->ls_root_sem);
534 list_for_each_entry(r, &ls->ls_root_list, res_root_list) {
535 if (dlm_recovery_stopped(ls)) {
536 up_read(&ls->ls_root_sem);
543 error = recover_master_static(r, &count);
545 error = recover_master(r, &count);
551 up_read(&ls->ls_root_sem);
555 up_read(&ls->ls_root_sem);
557 log_rinfo(ls, "dlm_recover_masters %u of %u", count, total);
559 error = dlm_wait_function(ls, &recover_idr_empty);
562 recover_idr_clear(ls);
566 int dlm_recover_master_reply(struct dlm_ls *ls, struct dlm_rcom *rc)
569 int ret_nodeid, new_master;
571 r = recover_idr_find(ls, rc->rc_id);
573 log_error(ls, "dlm_recover_master_reply no id %llx",
574 (unsigned long long)rc->rc_id);
578 ret_nodeid = rc->rc_result;
580 if (ret_nodeid == dlm_our_nodeid())
583 new_master = ret_nodeid;
586 r->res_master_nodeid = ret_nodeid;
587 r->res_nodeid = new_master;
592 if (recover_idr_empty(ls))
593 wake_up(&ls->ls_wait_general);
599 /* Lock recovery: rebuild the process-copy locks we hold on a
600 remastered rsb on the new rsb master.
605 dlm_send_rcom_lock -> receive_rcom_lock
606 dlm_recover_master_copy
607 receive_rcom_lock_reply <-
608 dlm_recover_process_copy
613 * keep a count of the number of lkb's we send to the new master; when we get
614 * an equal number of replies then recovery for the rsb is done
617 static int recover_locks_queue(struct dlm_rsb *r, struct list_head *head)
622 list_for_each_entry(lkb, head, lkb_statequeue) {
623 error = dlm_send_rcom_lock(r, lkb);
626 r->res_recover_locks_count++;
632 static int recover_locks(struct dlm_rsb *r)
638 DLM_ASSERT(!r->res_recover_locks_count, dlm_dump_rsb(r););
640 error = recover_locks_queue(r, &r->res_grantqueue);
643 error = recover_locks_queue(r, &r->res_convertqueue);
646 error = recover_locks_queue(r, &r->res_waitqueue);
650 if (r->res_recover_locks_count)
653 rsb_clear_flag(r, RSB_NEW_MASTER);
659 int dlm_recover_locks(struct dlm_ls *ls)
662 int error, count = 0;
664 down_read(&ls->ls_root_sem);
665 list_for_each_entry(r, &ls->ls_root_list, res_root_list) {
667 rsb_clear_flag(r, RSB_NEW_MASTER);
671 if (!rsb_flag(r, RSB_NEW_MASTER))
674 if (dlm_recovery_stopped(ls)) {
676 up_read(&ls->ls_root_sem);
680 error = recover_locks(r);
682 up_read(&ls->ls_root_sem);
686 count += r->res_recover_locks_count;
688 up_read(&ls->ls_root_sem);
690 log_rinfo(ls, "dlm_recover_locks %d out", count);
692 error = dlm_wait_function(ls, &recover_list_empty);
695 recover_list_clear(ls);
699 void dlm_recovered_lock(struct dlm_rsb *r)
701 DLM_ASSERT(rsb_flag(r, RSB_NEW_MASTER), dlm_dump_rsb(r););
703 r->res_recover_locks_count--;
704 if (!r->res_recover_locks_count) {
705 rsb_clear_flag(r, RSB_NEW_MASTER);
709 if (recover_list_empty(r->res_ls))
710 wake_up(&r->res_ls->ls_wait_general);
714 * The lvb needs to be recovered on all master rsb's. This includes setting
715 * the VALNOTVALID flag if necessary, and determining the correct lvb contents
716 * based on the lvb's of the locks held on the rsb.
718 * RSB_VALNOTVALID is set in two cases:
720 * 1. we are master, but not new, and we purged an EX/PW lock held by a
721 * failed node (in dlm_recover_purge which set RSB_RECOVER_LVB_INVAL)
723 * 2. we are a new master, and there are only NL/CR locks left.
724 * (We could probably improve this by only invaliding in this way when
725 * the previous master left uncleanly. VMS docs mention that.)
727 * The LVB contents are only considered for changing when this is a new master
728 * of the rsb (NEW_MASTER2). Then, the rsb's lvb is taken from any lkb with
729 * mode > CR. If no lkb's exist with mode above CR, the lvb contents are taken
730 * from the lkb with the largest lvb sequence number.
733 static void recover_lvb(struct dlm_rsb *r)
735 struct dlm_lkb *lkb, *high_lkb = NULL;
736 uint32_t high_seq = 0;
737 int lock_lvb_exists = 0;
738 int big_lock_exists = 0;
739 int lvblen = r->res_ls->ls_lvblen;
741 if (!rsb_flag(r, RSB_NEW_MASTER2) &&
742 rsb_flag(r, RSB_RECOVER_LVB_INVAL)) {
744 rsb_set_flag(r, RSB_VALNOTVALID);
748 if (!rsb_flag(r, RSB_NEW_MASTER2))
751 /* we are the new master, so figure out if VALNOTVALID should
752 be set, and set the rsb lvb from the best lkb available. */
754 list_for_each_entry(lkb, &r->res_grantqueue, lkb_statequeue) {
755 if (!(lkb->lkb_exflags & DLM_LKF_VALBLK))
760 if (lkb->lkb_grmode > DLM_LOCK_CR) {
765 if (((int)lkb->lkb_lvbseq - (int)high_seq) >= 0) {
767 high_seq = lkb->lkb_lvbseq;
771 list_for_each_entry(lkb, &r->res_convertqueue, lkb_statequeue) {
772 if (!(lkb->lkb_exflags & DLM_LKF_VALBLK))
777 if (lkb->lkb_grmode > DLM_LOCK_CR) {
782 if (((int)lkb->lkb_lvbseq - (int)high_seq) >= 0) {
784 high_seq = lkb->lkb_lvbseq;
789 if (!lock_lvb_exists)
792 /* lvb is invalidated if only NL/CR locks remain */
793 if (!big_lock_exists)
794 rsb_set_flag(r, RSB_VALNOTVALID);
796 if (!r->res_lvbptr) {
797 r->res_lvbptr = dlm_allocate_lvb(r->res_ls);
802 if (big_lock_exists) {
803 r->res_lvbseq = lkb->lkb_lvbseq;
804 memcpy(r->res_lvbptr, lkb->lkb_lvbptr, lvblen);
805 } else if (high_lkb) {
806 r->res_lvbseq = high_lkb->lkb_lvbseq;
807 memcpy(r->res_lvbptr, high_lkb->lkb_lvbptr, lvblen);
810 memset(r->res_lvbptr, 0, lvblen);
816 /* All master rsb's flagged RECOVER_CONVERT need to be looked at. The locks
817 converting PR->CW or CW->PR need to have their lkb_grmode set. */
819 static void recover_conversion(struct dlm_rsb *r)
821 struct dlm_ls *ls = r->res_ls;
825 list_for_each_entry(lkb, &r->res_grantqueue, lkb_statequeue) {
826 if (lkb->lkb_grmode == DLM_LOCK_PR ||
827 lkb->lkb_grmode == DLM_LOCK_CW) {
828 grmode = lkb->lkb_grmode;
833 list_for_each_entry(lkb, &r->res_convertqueue, lkb_statequeue) {
834 if (lkb->lkb_grmode != DLM_LOCK_IV)
837 log_debug(ls, "recover_conversion %x set gr to rq %d",
838 lkb->lkb_id, lkb->lkb_rqmode);
839 lkb->lkb_grmode = lkb->lkb_rqmode;
841 log_debug(ls, "recover_conversion %x set gr %d",
842 lkb->lkb_id, grmode);
843 lkb->lkb_grmode = grmode;
848 /* We've become the new master for this rsb and waiting/converting locks may
849 need to be granted in dlm_recover_grant() due to locks that may have
850 existed from a removed node. */
852 static void recover_grant(struct dlm_rsb *r)
854 if (!list_empty(&r->res_waitqueue) || !list_empty(&r->res_convertqueue))
855 rsb_set_flag(r, RSB_RECOVER_GRANT);
858 void dlm_recover_rsbs(struct dlm_ls *ls)
861 unsigned int count = 0;
863 down_read(&ls->ls_root_sem);
864 list_for_each_entry(r, &ls->ls_root_list, res_root_list) {
867 if (rsb_flag(r, RSB_RECOVER_CONVERT))
868 recover_conversion(r);
870 /* recover lvb before granting locks so the updated
871 lvb/VALNOTVALID is presented in the completion */
874 if (rsb_flag(r, RSB_NEW_MASTER2))
878 rsb_clear_flag(r, RSB_VALNOTVALID);
880 rsb_clear_flag(r, RSB_RECOVER_CONVERT);
881 rsb_clear_flag(r, RSB_RECOVER_LVB_INVAL);
882 rsb_clear_flag(r, RSB_NEW_MASTER2);
885 up_read(&ls->ls_root_sem);
888 log_rinfo(ls, "dlm_recover_rsbs %d done", count);
891 /* Create a single list of all root rsb's to be used during recovery */
893 int dlm_create_root_list(struct dlm_ls *ls)
899 down_write(&ls->ls_root_sem);
900 if (!list_empty(&ls->ls_root_list)) {
901 log_error(ls, "root list not empty");
906 for (i = 0; i < ls->ls_rsbtbl_size; i++) {
907 spin_lock(&ls->ls_rsbtbl[i].lock);
908 for (n = rb_first(&ls->ls_rsbtbl[i].keep); n; n = rb_next(n)) {
909 r = rb_entry(n, struct dlm_rsb, res_hashnode);
910 list_add(&r->res_root_list, &ls->ls_root_list);
914 if (!RB_EMPTY_ROOT(&ls->ls_rsbtbl[i].toss))
915 log_error(ls, "dlm_create_root_list toss not empty");
916 spin_unlock(&ls->ls_rsbtbl[i].lock);
919 up_write(&ls->ls_root_sem);
923 void dlm_release_root_list(struct dlm_ls *ls)
925 struct dlm_rsb *r, *safe;
927 down_write(&ls->ls_root_sem);
928 list_for_each_entry_safe(r, safe, &ls->ls_root_list, res_root_list) {
929 list_del_init(&r->res_root_list);
932 up_write(&ls->ls_root_sem);
935 void dlm_clear_toss(struct dlm_ls *ls)
937 struct rb_node *n, *next;
939 unsigned int count = 0;
942 for (i = 0; i < ls->ls_rsbtbl_size; i++) {
943 spin_lock(&ls->ls_rsbtbl[i].lock);
944 for (n = rb_first(&ls->ls_rsbtbl[i].toss); n; n = next) {
946 r = rb_entry(n, struct dlm_rsb, res_hashnode);
947 rb_erase(n, &ls->ls_rsbtbl[i].toss);
951 spin_unlock(&ls->ls_rsbtbl[i].lock);
955 log_rinfo(ls, "dlm_clear_toss %u done", count);