1 /******************************************************************************
2 *******************************************************************************
4 ** Copyright (C) Sistina Software, Inc. 1997-2003 All rights reserved.
5 ** Copyright (C) 2004-2005 Red Hat, Inc. All rights reserved.
7 ** This copyrighted material is made available to anyone wishing to use,
8 ** modify, copy, or redistribute it subject to the terms and conditions
9 ** of the GNU General Public License v.2.
11 *******************************************************************************
12 ******************************************************************************/
14 #include "dlm_internal.h"
15 #include "lockspace.h"
28 * Recovery waiting routines: these functions wait for a particular reply from
29 * a remote node, or for the remote node to report a certain status. They need
30 * to abort if the lockspace is stopped indicating a node has failed (perhaps
31 * the one being waited for).
35 * Wait until given function returns non-zero or lockspace is stopped
36 * (LS_RECOVERY_STOP set due to failure of a node in ls_nodes). When another
37 * function thinks it could have completed the waited-on task, they should wake
38 * up ls_wait_general to get an immediate response rather than waiting for the
39 * timeout. This uses a timeout so it can check periodically if the wait
40 * should abort due to node failure (which doesn't cause a wake_up).
41 * This should only be called by the dlm_recoverd thread.
44 int dlm_wait_function(struct dlm_ls *ls, int (*testfn) (struct dlm_ls *ls))
50 rv = wait_event_timeout(ls->ls_wait_general,
51 testfn(ls) || dlm_recovery_stopped(ls),
52 dlm_config.ci_recover_timer * HZ);
57 if (dlm_recovery_stopped(ls)) {
58 log_debug(ls, "dlm_wait_function aborted");
65 * An efficient way for all nodes to wait for all others to have a certain
66 * status. The node with the lowest nodeid polls all the others for their
67 * status (wait_status_all) and all the others poll the node with the low id
68 * for its accumulated result (wait_status_low). When all nodes have set
69 * status flag X, then status flag X_ALL will be set on the low nodeid.
72 uint32_t dlm_recover_status(struct dlm_ls *ls)
75 spin_lock(&ls->ls_recover_lock);
76 status = ls->ls_recover_status;
77 spin_unlock(&ls->ls_recover_lock);
81 static void _set_recover_status(struct dlm_ls *ls, uint32_t status)
83 ls->ls_recover_status |= status;
86 void dlm_set_recover_status(struct dlm_ls *ls, uint32_t status)
88 spin_lock(&ls->ls_recover_lock);
89 _set_recover_status(ls, status);
90 spin_unlock(&ls->ls_recover_lock);
93 static int wait_status_all(struct dlm_ls *ls, uint32_t wait_status,
96 struct dlm_rcom *rc = ls->ls_recover_buf;
97 struct dlm_member *memb;
100 list_for_each_entry(memb, &ls->ls_nodes, list) {
103 if (dlm_recovery_stopped(ls)) {
108 error = dlm_rcom_status(ls, memb->nodeid, 0);
113 dlm_slot_save(ls, rc, memb);
115 if (rc->rc_result & wait_status)
126 static int wait_status_low(struct dlm_ls *ls, uint32_t wait_status,
127 uint32_t status_flags)
129 struct dlm_rcom *rc = ls->ls_recover_buf;
130 int error = 0, delay = 0, nodeid = ls->ls_low_nodeid;
133 if (dlm_recovery_stopped(ls)) {
138 error = dlm_rcom_status(ls, nodeid, status_flags);
142 if (rc->rc_result & wait_status)
152 static int wait_status(struct dlm_ls *ls, uint32_t status)
154 uint32_t status_all = status << 1;
157 if (ls->ls_low_nodeid == dlm_our_nodeid()) {
158 error = wait_status_all(ls, status, 0);
160 dlm_set_recover_status(ls, status_all);
162 error = wait_status_low(ls, status_all, 0);
167 int dlm_recover_members_wait(struct dlm_ls *ls)
169 struct dlm_member *memb;
170 struct dlm_slot *slots;
171 int num_slots, slots_size;
175 list_for_each_entry(memb, &ls->ls_nodes, list) {
177 memb->generation = 0;
180 if (ls->ls_low_nodeid == dlm_our_nodeid()) {
181 error = wait_status_all(ls, DLM_RS_NODES, 1);
185 /* slots array is sparse, slots_size may be > num_slots */
187 rv = dlm_slots_assign(ls, &num_slots, &slots_size, &slots, &gen);
189 spin_lock(&ls->ls_recover_lock);
190 _set_recover_status(ls, DLM_RS_NODES_ALL);
191 ls->ls_num_slots = num_slots;
192 ls->ls_slots_size = slots_size;
193 ls->ls_slots = slots;
194 ls->ls_generation = gen;
195 spin_unlock(&ls->ls_recover_lock);
197 dlm_set_recover_status(ls, DLM_RS_NODES_ALL);
200 error = wait_status_low(ls, DLM_RS_NODES_ALL, DLM_RSF_NEED_SLOTS);
204 dlm_slots_copy_in(ls);
210 int dlm_recover_directory_wait(struct dlm_ls *ls)
212 return wait_status(ls, DLM_RS_DIR);
215 int dlm_recover_locks_wait(struct dlm_ls *ls)
217 return wait_status(ls, DLM_RS_LOCKS);
220 int dlm_recover_done_wait(struct dlm_ls *ls)
222 return wait_status(ls, DLM_RS_DONE);
226 * The recover_list contains all the rsb's for which we've requested the new
227 * master nodeid. As replies are returned from the resource directories the
228 * rsb's are removed from the list. When the list is empty we're done.
230 * The recover_list is later similarly used for all rsb's for which we've sent
231 * new lkb's and need to receive new corresponding lkid's.
233 * We use the address of the rsb struct as a simple local identifier for the
234 * rsb so we can match an rcom reply with the rsb it was sent for.
237 static int recover_list_empty(struct dlm_ls *ls)
241 spin_lock(&ls->ls_recover_list_lock);
242 empty = list_empty(&ls->ls_recover_list);
243 spin_unlock(&ls->ls_recover_list_lock);
248 static void recover_list_add(struct dlm_rsb *r)
250 struct dlm_ls *ls = r->res_ls;
252 spin_lock(&ls->ls_recover_list_lock);
253 if (list_empty(&r->res_recover_list)) {
254 list_add_tail(&r->res_recover_list, &ls->ls_recover_list);
255 ls->ls_recover_list_count++;
258 spin_unlock(&ls->ls_recover_list_lock);
261 static void recover_list_del(struct dlm_rsb *r)
263 struct dlm_ls *ls = r->res_ls;
265 spin_lock(&ls->ls_recover_list_lock);
266 list_del_init(&r->res_recover_list);
267 ls->ls_recover_list_count--;
268 spin_unlock(&ls->ls_recover_list_lock);
273 static void recover_list_clear(struct dlm_ls *ls)
275 struct dlm_rsb *r, *s;
277 spin_lock(&ls->ls_recover_list_lock);
278 list_for_each_entry_safe(r, s, &ls->ls_recover_list, res_recover_list) {
279 list_del_init(&r->res_recover_list);
280 r->res_recover_locks_count = 0;
282 ls->ls_recover_list_count--;
285 if (ls->ls_recover_list_count != 0) {
286 log_error(ls, "warning: recover_list_count %d",
287 ls->ls_recover_list_count);
288 ls->ls_recover_list_count = 0;
290 spin_unlock(&ls->ls_recover_list_lock);
293 static int recover_idr_empty(struct dlm_ls *ls)
297 spin_lock(&ls->ls_recover_idr_lock);
298 if (ls->ls_recover_list_count)
300 spin_unlock(&ls->ls_recover_idr_lock);
305 static int recover_idr_add(struct dlm_rsb *r)
307 struct dlm_ls *ls = r->res_ls;
310 rv = idr_pre_get(&ls->ls_recover_idr, GFP_NOFS);
314 spin_lock(&ls->ls_recover_idr_lock);
316 spin_unlock(&ls->ls_recover_idr_lock);
319 rv = idr_get_new_above(&ls->ls_recover_idr, r, 1, &id);
321 spin_unlock(&ls->ls_recover_idr_lock);
325 ls->ls_recover_list_count++;
327 spin_unlock(&ls->ls_recover_idr_lock);
331 static void recover_idr_del(struct dlm_rsb *r)
333 struct dlm_ls *ls = r->res_ls;
335 spin_lock(&ls->ls_recover_idr_lock);
336 idr_remove(&ls->ls_recover_idr, r->res_id);
338 ls->ls_recover_list_count--;
339 spin_unlock(&ls->ls_recover_idr_lock);
344 static struct dlm_rsb *recover_idr_find(struct dlm_ls *ls, uint64_t id)
348 spin_lock(&ls->ls_recover_idr_lock);
349 r = idr_find(&ls->ls_recover_idr, (int)id);
350 spin_unlock(&ls->ls_recover_idr_lock);
354 static int recover_idr_clear_rsb(int id, void *p, void *data)
356 struct dlm_ls *ls = data;
357 struct dlm_rsb *r = p;
360 r->res_recover_locks_count = 0;
361 ls->ls_recover_list_count--;
367 static void recover_idr_clear(struct dlm_ls *ls)
369 spin_lock(&ls->ls_recover_idr_lock);
370 idr_for_each(&ls->ls_recover_idr, recover_idr_clear_rsb, ls);
371 idr_remove_all(&ls->ls_recover_idr);
373 if (ls->ls_recover_list_count != 0) {
374 log_error(ls, "warning: recover_list_count %d",
375 ls->ls_recover_list_count);
376 ls->ls_recover_list_count = 0;
378 spin_unlock(&ls->ls_recover_idr_lock);
382 /* Master recovery: find new master node for rsb's that were
383 mastered on nodes that have been removed.
387 dlm_send_rcom_lookup -> receive_rcom_lookup
389 receive_rcom_lookup_reply <-
390 dlm_recover_master_reply
397 * Set the lock master for all LKBs in a lock queue
398 * If we are the new master of the rsb, we may have received new
399 * MSTCPY locks from other nodes already which we need to ignore
400 * when setting the new nodeid.
403 static void set_lock_master(struct list_head *queue, int nodeid)
407 list_for_each_entry(lkb, queue, lkb_statequeue) {
408 if (!(lkb->lkb_flags & DLM_IFL_MSTCPY)) {
409 lkb->lkb_nodeid = nodeid;
415 static void set_master_lkbs(struct dlm_rsb *r)
417 set_lock_master(&r->res_grantqueue, r->res_nodeid);
418 set_lock_master(&r->res_convertqueue, r->res_nodeid);
419 set_lock_master(&r->res_waitqueue, r->res_nodeid);
423 * Propagate the new master nodeid to locks
424 * The NEW_MASTER flag tells dlm_recover_locks() which rsb's to consider.
425 * The NEW_MASTER2 flag tells recover_lvb() and recover_grant() which
429 static void set_new_master(struct dlm_rsb *r)
432 rsb_set_flag(r, RSB_NEW_MASTER);
433 rsb_set_flag(r, RSB_NEW_MASTER2);
437 * We do async lookups on rsb's that need new masters. The rsb's
438 * waiting for a lookup reply are kept on the recover_list.
440 * Another node recovering the master may have sent us a rcom lookup,
441 * and our dlm_master_lookup() set it as the new master, along with
442 * NEW_MASTER so that we'll recover it here (this implies dir_nodeid
443 * equals our_nodeid below).
446 static int recover_master(struct dlm_rsb *r, unsigned int *count)
448 struct dlm_ls *ls = r->res_ls;
449 int our_nodeid, dir_nodeid;
456 is_removed = dlm_is_removed(ls, r->res_nodeid);
458 if (!is_removed && !rsb_flag(r, RSB_NEW_MASTER))
461 our_nodeid = dlm_our_nodeid();
462 dir_nodeid = dlm_dir_nodeid(r);
464 if (dir_nodeid == our_nodeid) {
466 r->res_master_nodeid = our_nodeid;
470 /* set master of lkbs to ourself when is_removed, or to
471 another new master which we set along with NEW_MASTER
472 in dlm_master_lookup */
477 error = dlm_send_rcom_lookup(r, dir_nodeid);
485 * All MSTCPY locks are purged and rebuilt, even if the master stayed the same.
486 * This is necessary because recovery can be started, aborted and restarted,
487 * causing the master nodeid to briefly change during the aborted recovery, and
488 * change back to the original value in the second recovery. The MSTCPY locks
489 * may or may not have been purged during the aborted recovery. Another node
490 * with an outstanding request in waiters list and a request reply saved in the
491 * requestqueue, cannot know whether it should ignore the reply and resend the
492 * request, or accept the reply and complete the request. It must do the
493 * former if the remote node purged MSTCPY locks, and it must do the later if
494 * the remote node did not. This is solved by always purging MSTCPY locks, in
495 * which case, the request reply would always be ignored and the request
499 static int recover_master_static(struct dlm_rsb *r, unsigned int *count)
501 int dir_nodeid = dlm_dir_nodeid(r);
502 int new_master = dir_nodeid;
504 if (dir_nodeid == dlm_our_nodeid())
507 dlm_purge_mstcpy_locks(r);
508 r->res_master_nodeid = dir_nodeid;
509 r->res_nodeid = new_master;
516 * Go through local root resources and for each rsb which has a master which
517 * has departed, get the new master nodeid from the directory. The dir will
518 * assign mastery to the first node to look up the new master. That means
519 * we'll discover in this lookup if we're the new master of any rsb's.
521 * We fire off all the dir lookup requests individually and asynchronously to
522 * the correct dir node.
525 int dlm_recover_masters(struct dlm_ls *ls)
528 unsigned int total = 0;
529 unsigned int count = 0;
530 int nodir = dlm_no_directory(ls);
533 log_debug(ls, "dlm_recover_masters");
535 down_read(&ls->ls_root_sem);
536 list_for_each_entry(r, &ls->ls_root_list, res_root_list) {
537 if (dlm_recovery_stopped(ls)) {
538 up_read(&ls->ls_root_sem);
545 error = recover_master_static(r, &count);
547 error = recover_master(r, &count);
553 up_read(&ls->ls_root_sem);
557 up_read(&ls->ls_root_sem);
559 log_debug(ls, "dlm_recover_masters %u of %u", count, total);
561 error = dlm_wait_function(ls, &recover_idr_empty);
564 recover_idr_clear(ls);
568 int dlm_recover_master_reply(struct dlm_ls *ls, struct dlm_rcom *rc)
571 int ret_nodeid, new_master;
573 r = recover_idr_find(ls, rc->rc_id);
575 log_error(ls, "dlm_recover_master_reply no id %llx",
576 (unsigned long long)rc->rc_id);
580 ret_nodeid = rc->rc_result;
582 if (ret_nodeid == dlm_our_nodeid())
585 new_master = ret_nodeid;
588 r->res_master_nodeid = ret_nodeid;
589 r->res_nodeid = new_master;
594 if (recover_idr_empty(ls))
595 wake_up(&ls->ls_wait_general);
601 /* Lock recovery: rebuild the process-copy locks we hold on a
602 remastered rsb on the new rsb master.
607 dlm_send_rcom_lock -> receive_rcom_lock
608 dlm_recover_master_copy
609 receive_rcom_lock_reply <-
610 dlm_recover_process_copy
615 * keep a count of the number of lkb's we send to the new master; when we get
616 * an equal number of replies then recovery for the rsb is done
619 static int recover_locks_queue(struct dlm_rsb *r, struct list_head *head)
624 list_for_each_entry(lkb, head, lkb_statequeue) {
625 error = dlm_send_rcom_lock(r, lkb);
628 r->res_recover_locks_count++;
634 static int recover_locks(struct dlm_rsb *r)
640 DLM_ASSERT(!r->res_recover_locks_count, dlm_dump_rsb(r););
642 error = recover_locks_queue(r, &r->res_grantqueue);
645 error = recover_locks_queue(r, &r->res_convertqueue);
648 error = recover_locks_queue(r, &r->res_waitqueue);
652 if (r->res_recover_locks_count)
655 rsb_clear_flag(r, RSB_NEW_MASTER);
661 int dlm_recover_locks(struct dlm_ls *ls)
664 int error, count = 0;
666 down_read(&ls->ls_root_sem);
667 list_for_each_entry(r, &ls->ls_root_list, res_root_list) {
669 rsb_clear_flag(r, RSB_NEW_MASTER);
673 if (!rsb_flag(r, RSB_NEW_MASTER))
676 if (dlm_recovery_stopped(ls)) {
678 up_read(&ls->ls_root_sem);
682 error = recover_locks(r);
684 up_read(&ls->ls_root_sem);
688 count += r->res_recover_locks_count;
690 up_read(&ls->ls_root_sem);
692 log_debug(ls, "dlm_recover_locks %d out", count);
694 error = dlm_wait_function(ls, &recover_list_empty);
697 recover_list_clear(ls);
701 void dlm_recovered_lock(struct dlm_rsb *r)
703 DLM_ASSERT(rsb_flag(r, RSB_NEW_MASTER), dlm_dump_rsb(r););
705 r->res_recover_locks_count--;
706 if (!r->res_recover_locks_count) {
707 rsb_clear_flag(r, RSB_NEW_MASTER);
711 if (recover_list_empty(r->res_ls))
712 wake_up(&r->res_ls->ls_wait_general);
716 * The lvb needs to be recovered on all master rsb's. This includes setting
717 * the VALNOTVALID flag if necessary, and determining the correct lvb contents
718 * based on the lvb's of the locks held on the rsb.
720 * RSB_VALNOTVALID is set if there are only NL/CR locks on the rsb. If it
721 * was already set prior to recovery, it's not cleared, regardless of locks.
723 * The LVB contents are only considered for changing when this is a new master
724 * of the rsb (NEW_MASTER2). Then, the rsb's lvb is taken from any lkb with
725 * mode > CR. If no lkb's exist with mode above CR, the lvb contents are taken
726 * from the lkb with the largest lvb sequence number.
729 static void recover_lvb(struct dlm_rsb *r)
731 struct dlm_lkb *lkb, *high_lkb = NULL;
732 uint32_t high_seq = 0;
733 int lock_lvb_exists = 0;
734 int big_lock_exists = 0;
735 int lvblen = r->res_ls->ls_lvblen;
737 list_for_each_entry(lkb, &r->res_grantqueue, lkb_statequeue) {
738 if (!(lkb->lkb_exflags & DLM_LKF_VALBLK))
743 if (lkb->lkb_grmode > DLM_LOCK_CR) {
748 if (((int)lkb->lkb_lvbseq - (int)high_seq) >= 0) {
750 high_seq = lkb->lkb_lvbseq;
754 list_for_each_entry(lkb, &r->res_convertqueue, 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;
772 if (!lock_lvb_exists)
775 if (!big_lock_exists)
776 rsb_set_flag(r, RSB_VALNOTVALID);
778 /* don't mess with the lvb unless we're the new master */
779 if (!rsb_flag(r, RSB_NEW_MASTER2))
782 if (!r->res_lvbptr) {
783 r->res_lvbptr = dlm_allocate_lvb(r->res_ls);
788 if (big_lock_exists) {
789 r->res_lvbseq = lkb->lkb_lvbseq;
790 memcpy(r->res_lvbptr, lkb->lkb_lvbptr, lvblen);
791 } else if (high_lkb) {
792 r->res_lvbseq = high_lkb->lkb_lvbseq;
793 memcpy(r->res_lvbptr, high_lkb->lkb_lvbptr, lvblen);
796 memset(r->res_lvbptr, 0, lvblen);
802 /* All master rsb's flagged RECOVER_CONVERT need to be looked at. The locks
803 converting PR->CW or CW->PR need to have their lkb_grmode set. */
805 static void recover_conversion(struct dlm_rsb *r)
807 struct dlm_ls *ls = r->res_ls;
811 list_for_each_entry(lkb, &r->res_grantqueue, lkb_statequeue) {
812 if (lkb->lkb_grmode == DLM_LOCK_PR ||
813 lkb->lkb_grmode == DLM_LOCK_CW) {
814 grmode = lkb->lkb_grmode;
819 list_for_each_entry(lkb, &r->res_convertqueue, lkb_statequeue) {
820 if (lkb->lkb_grmode != DLM_LOCK_IV)
823 log_debug(ls, "recover_conversion %x set gr to rq %d",
824 lkb->lkb_id, lkb->lkb_rqmode);
825 lkb->lkb_grmode = lkb->lkb_rqmode;
827 log_debug(ls, "recover_conversion %x set gr %d",
828 lkb->lkb_id, grmode);
829 lkb->lkb_grmode = grmode;
834 /* We've become the new master for this rsb and waiting/converting locks may
835 need to be granted in dlm_recover_grant() due to locks that may have
836 existed from a removed node. */
838 static void recover_grant(struct dlm_rsb *r)
840 if (!list_empty(&r->res_waitqueue) || !list_empty(&r->res_convertqueue))
841 rsb_set_flag(r, RSB_RECOVER_GRANT);
844 void dlm_recover_rsbs(struct dlm_ls *ls)
847 unsigned int count = 0;
849 down_read(&ls->ls_root_sem);
850 list_for_each_entry(r, &ls->ls_root_list, res_root_list) {
853 if (rsb_flag(r, RSB_RECOVER_CONVERT))
854 recover_conversion(r);
855 if (rsb_flag(r, RSB_NEW_MASTER2))
860 rsb_clear_flag(r, RSB_RECOVER_CONVERT);
861 rsb_clear_flag(r, RSB_NEW_MASTER2);
864 up_read(&ls->ls_root_sem);
867 log_debug(ls, "dlm_recover_rsbs %d done", count);
870 /* Create a single list of all root rsb's to be used during recovery */
872 int dlm_create_root_list(struct dlm_ls *ls)
878 down_write(&ls->ls_root_sem);
879 if (!list_empty(&ls->ls_root_list)) {
880 log_error(ls, "root list not empty");
885 for (i = 0; i < ls->ls_rsbtbl_size; i++) {
886 spin_lock(&ls->ls_rsbtbl[i].lock);
887 for (n = rb_first(&ls->ls_rsbtbl[i].keep); n; n = rb_next(n)) {
888 r = rb_entry(n, struct dlm_rsb, res_hashnode);
889 list_add(&r->res_root_list, &ls->ls_root_list);
893 if (!RB_EMPTY_ROOT(&ls->ls_rsbtbl[i].toss))
894 log_error(ls, "dlm_create_root_list toss not empty");
895 spin_unlock(&ls->ls_rsbtbl[i].lock);
898 up_write(&ls->ls_root_sem);
902 void dlm_release_root_list(struct dlm_ls *ls)
904 struct dlm_rsb *r, *safe;
906 down_write(&ls->ls_root_sem);
907 list_for_each_entry_safe(r, safe, &ls->ls_root_list, res_root_list) {
908 list_del_init(&r->res_root_list);
911 up_write(&ls->ls_root_sem);
914 void dlm_clear_toss(struct dlm_ls *ls)
916 struct rb_node *n, *next;
918 unsigned int count = 0;
921 for (i = 0; i < ls->ls_rsbtbl_size; i++) {
922 spin_lock(&ls->ls_rsbtbl[i].lock);
923 for (n = rb_first(&ls->ls_rsbtbl[i].toss); n; n = next) {
925 r = rb_entry(n, struct dlm_rsb, res_hashnode);
926 rb_erase(n, &ls->ls_rsbtbl[i].toss);
930 spin_unlock(&ls->ls_rsbtbl[i].lock);
934 log_debug(ls, "dlm_clear_toss %u done", count);