1 /******************************************************************************
2 *******************************************************************************
4 ** Copyright (C) 2005-2010 Red Hat, Inc. All rights reserved.
6 ** This copyrighted material is made available to anyone wishing to use,
7 ** modify, copy, or redistribute it subject to the terms and conditions
8 ** of the GNU General Public License v.2.
10 *******************************************************************************
11 ******************************************************************************/
13 /* Central locking logic has four stages:
33 Stage 1 (lock, unlock) is mainly about checking input args and
34 splitting into one of the four main operations:
36 dlm_lock = request_lock
37 dlm_lock+CONVERT = convert_lock
38 dlm_unlock = unlock_lock
39 dlm_unlock+CANCEL = cancel_lock
41 Stage 2, xxxx_lock(), just finds and locks the relevant rsb which is
42 provided to the next stage.
44 Stage 3, _xxxx_lock(), determines if the operation is local or remote.
45 When remote, it calls send_xxxx(), when local it calls do_xxxx().
47 Stage 4, do_xxxx(), is the guts of the operation. It manipulates the
48 given rsb and lkb and queues callbacks.
50 For remote operations, send_xxxx() results in the corresponding do_xxxx()
51 function being executed on the remote node. The connecting send/receive
52 calls on local (L) and remote (R) nodes:
54 L: send_xxxx() -> R: receive_xxxx()
56 L: receive_xxxx_reply() <- R: send_xxxx_reply()
58 #include <linux/types.h>
59 #include <linux/rbtree.h>
60 #include <linux/slab.h>
61 #include "dlm_internal.h"
62 #include <linux/dlm_device.h>
65 #include "requestqueue.h"
69 #include "lockspace.h"
74 #include "lvb_table.h"
78 static int send_request(struct dlm_rsb *r, struct dlm_lkb *lkb);
79 static int send_convert(struct dlm_rsb *r, struct dlm_lkb *lkb);
80 static int send_unlock(struct dlm_rsb *r, struct dlm_lkb *lkb);
81 static int send_cancel(struct dlm_rsb *r, struct dlm_lkb *lkb);
82 static int send_grant(struct dlm_rsb *r, struct dlm_lkb *lkb);
83 static int send_bast(struct dlm_rsb *r, struct dlm_lkb *lkb, int mode);
84 static int send_lookup(struct dlm_rsb *r, struct dlm_lkb *lkb);
85 static int send_remove(struct dlm_rsb *r);
86 static int _request_lock(struct dlm_rsb *r, struct dlm_lkb *lkb);
87 static int _cancel_lock(struct dlm_rsb *r, struct dlm_lkb *lkb);
88 static void __receive_convert_reply(struct dlm_rsb *r, struct dlm_lkb *lkb,
89 struct dlm_message *ms);
90 static int receive_extralen(struct dlm_message *ms);
91 static void do_purge(struct dlm_ls *ls, int nodeid, int pid);
92 static void del_timeout(struct dlm_lkb *lkb);
93 static void toss_rsb(struct kref *kref);
96 * Lock compatibilty matrix - thanks Steve
97 * UN = Unlocked state. Not really a state, used as a flag
98 * PD = Padding. Used to make the matrix a nice power of two in size
99 * Other states are the same as the VMS DLM.
100 * Usage: matrix[grmode+1][rqmode+1] (although m[rq+1][gr+1] is the same)
103 static const int __dlm_compat_matrix[8][8] = {
104 /* UN NL CR CW PR PW EX PD */
105 {1, 1, 1, 1, 1, 1, 1, 0}, /* UN */
106 {1, 1, 1, 1, 1, 1, 1, 0}, /* NL */
107 {1, 1, 1, 1, 1, 1, 0, 0}, /* CR */
108 {1, 1, 1, 1, 0, 0, 0, 0}, /* CW */
109 {1, 1, 1, 0, 1, 0, 0, 0}, /* PR */
110 {1, 1, 1, 0, 0, 0, 0, 0}, /* PW */
111 {1, 1, 0, 0, 0, 0, 0, 0}, /* EX */
112 {0, 0, 0, 0, 0, 0, 0, 0} /* PD */
116 * This defines the direction of transfer of LVB data.
117 * Granted mode is the row; requested mode is the column.
118 * Usage: matrix[grmode+1][rqmode+1]
119 * 1 = LVB is returned to the caller
120 * 0 = LVB is written to the resource
121 * -1 = nothing happens to the LVB
124 const int dlm_lvb_operations[8][8] = {
125 /* UN NL CR CW PR PW EX PD*/
126 { -1, 1, 1, 1, 1, 1, 1, -1 }, /* UN */
127 { -1, 1, 1, 1, 1, 1, 1, 0 }, /* NL */
128 { -1, -1, 1, 1, 1, 1, 1, 0 }, /* CR */
129 { -1, -1, -1, 1, 1, 1, 1, 0 }, /* CW */
130 { -1, -1, -1, -1, 1, 1, 1, 0 }, /* PR */
131 { -1, 0, 0, 0, 0, 0, 1, 0 }, /* PW */
132 { -1, 0, 0, 0, 0, 0, 0, 0 }, /* EX */
133 { -1, 0, 0, 0, 0, 0, 0, 0 } /* PD */
136 #define modes_compat(gr, rq) \
137 __dlm_compat_matrix[(gr)->lkb_grmode + 1][(rq)->lkb_rqmode + 1]
139 int dlm_modes_compat(int mode1, int mode2)
141 return __dlm_compat_matrix[mode1 + 1][mode2 + 1];
145 * Compatibility matrix for conversions with QUECVT set.
146 * Granted mode is the row; requested mode is the column.
147 * Usage: matrix[grmode+1][rqmode+1]
150 static const int __quecvt_compat_matrix[8][8] = {
151 /* UN NL CR CW PR PW EX PD */
152 {0, 0, 0, 0, 0, 0, 0, 0}, /* UN */
153 {0, 0, 1, 1, 1, 1, 1, 0}, /* NL */
154 {0, 0, 0, 1, 1, 1, 1, 0}, /* CR */
155 {0, 0, 0, 0, 1, 1, 1, 0}, /* CW */
156 {0, 0, 0, 1, 0, 1, 1, 0}, /* PR */
157 {0, 0, 0, 0, 0, 0, 1, 0}, /* PW */
158 {0, 0, 0, 0, 0, 0, 0, 0}, /* EX */
159 {0, 0, 0, 0, 0, 0, 0, 0} /* PD */
162 void dlm_print_lkb(struct dlm_lkb *lkb)
164 printk(KERN_ERR "lkb: nodeid %d id %x remid %x exflags %x flags %x "
165 "sts %d rq %d gr %d wait_type %d wait_nodeid %d seq %llu\n",
166 lkb->lkb_nodeid, lkb->lkb_id, lkb->lkb_remid, lkb->lkb_exflags,
167 lkb->lkb_flags, lkb->lkb_status, lkb->lkb_rqmode,
168 lkb->lkb_grmode, lkb->lkb_wait_type, lkb->lkb_wait_nodeid,
169 (unsigned long long)lkb->lkb_recover_seq);
172 static void dlm_print_rsb(struct dlm_rsb *r)
174 printk(KERN_ERR "rsb: nodeid %d master %d dir %d flags %lx first %x "
176 r->res_nodeid, r->res_master_nodeid, r->res_dir_nodeid,
177 r->res_flags, r->res_first_lkid, r->res_recover_locks_count,
181 void dlm_dump_rsb(struct dlm_rsb *r)
187 printk(KERN_ERR "rsb: root_list empty %d recover_list empty %d\n",
188 list_empty(&r->res_root_list), list_empty(&r->res_recover_list));
189 printk(KERN_ERR "rsb lookup list\n");
190 list_for_each_entry(lkb, &r->res_lookup, lkb_rsb_lookup)
192 printk(KERN_ERR "rsb grant queue:\n");
193 list_for_each_entry(lkb, &r->res_grantqueue, lkb_statequeue)
195 printk(KERN_ERR "rsb convert queue:\n");
196 list_for_each_entry(lkb, &r->res_convertqueue, lkb_statequeue)
198 printk(KERN_ERR "rsb wait queue:\n");
199 list_for_each_entry(lkb, &r->res_waitqueue, lkb_statequeue)
203 /* Threads cannot use the lockspace while it's being recovered */
205 static inline void dlm_lock_recovery(struct dlm_ls *ls)
207 down_read(&ls->ls_in_recovery);
210 void dlm_unlock_recovery(struct dlm_ls *ls)
212 up_read(&ls->ls_in_recovery);
215 int dlm_lock_recovery_try(struct dlm_ls *ls)
217 return down_read_trylock(&ls->ls_in_recovery);
220 static inline int can_be_queued(struct dlm_lkb *lkb)
222 return !(lkb->lkb_exflags & DLM_LKF_NOQUEUE);
225 static inline int force_blocking_asts(struct dlm_lkb *lkb)
227 return (lkb->lkb_exflags & DLM_LKF_NOQUEUEBAST);
230 static inline int is_demoted(struct dlm_lkb *lkb)
232 return (lkb->lkb_sbflags & DLM_SBF_DEMOTED);
235 static inline int is_altmode(struct dlm_lkb *lkb)
237 return (lkb->lkb_sbflags & DLM_SBF_ALTMODE);
240 static inline int is_granted(struct dlm_lkb *lkb)
242 return (lkb->lkb_status == DLM_LKSTS_GRANTED);
245 static inline int is_remote(struct dlm_rsb *r)
247 DLM_ASSERT(r->res_nodeid >= 0, dlm_print_rsb(r););
248 return !!r->res_nodeid;
251 static inline int is_process_copy(struct dlm_lkb *lkb)
253 return (lkb->lkb_nodeid && !(lkb->lkb_flags & DLM_IFL_MSTCPY));
256 static inline int is_master_copy(struct dlm_lkb *lkb)
258 return (lkb->lkb_flags & DLM_IFL_MSTCPY) ? 1 : 0;
261 static inline int middle_conversion(struct dlm_lkb *lkb)
263 if ((lkb->lkb_grmode==DLM_LOCK_PR && lkb->lkb_rqmode==DLM_LOCK_CW) ||
264 (lkb->lkb_rqmode==DLM_LOCK_PR && lkb->lkb_grmode==DLM_LOCK_CW))
269 static inline int down_conversion(struct dlm_lkb *lkb)
271 return (!middle_conversion(lkb) && lkb->lkb_rqmode < lkb->lkb_grmode);
274 static inline int is_overlap_unlock(struct dlm_lkb *lkb)
276 return lkb->lkb_flags & DLM_IFL_OVERLAP_UNLOCK;
279 static inline int is_overlap_cancel(struct dlm_lkb *lkb)
281 return lkb->lkb_flags & DLM_IFL_OVERLAP_CANCEL;
284 static inline int is_overlap(struct dlm_lkb *lkb)
286 return (lkb->lkb_flags & (DLM_IFL_OVERLAP_UNLOCK |
287 DLM_IFL_OVERLAP_CANCEL));
290 static void queue_cast(struct dlm_rsb *r, struct dlm_lkb *lkb, int rv)
292 if (is_master_copy(lkb))
297 DLM_ASSERT(lkb->lkb_lksb, dlm_print_lkb(lkb););
299 /* if the operation was a cancel, then return -DLM_ECANCEL, if a
300 timeout caused the cancel then return -ETIMEDOUT */
301 if (rv == -DLM_ECANCEL && (lkb->lkb_flags & DLM_IFL_TIMEOUT_CANCEL)) {
302 lkb->lkb_flags &= ~DLM_IFL_TIMEOUT_CANCEL;
306 if (rv == -DLM_ECANCEL && (lkb->lkb_flags & DLM_IFL_DEADLOCK_CANCEL)) {
307 lkb->lkb_flags &= ~DLM_IFL_DEADLOCK_CANCEL;
311 dlm_add_cb(lkb, DLM_CB_CAST, lkb->lkb_grmode, rv, lkb->lkb_sbflags);
314 static inline void queue_cast_overlap(struct dlm_rsb *r, struct dlm_lkb *lkb)
317 is_overlap_unlock(lkb) ? -DLM_EUNLOCK : -DLM_ECANCEL);
320 static void queue_bast(struct dlm_rsb *r, struct dlm_lkb *lkb, int rqmode)
322 if (is_master_copy(lkb)) {
323 send_bast(r, lkb, rqmode);
325 dlm_add_cb(lkb, DLM_CB_BAST, rqmode, 0, 0);
330 * Basic operations on rsb's and lkb's
333 /* This is only called to add a reference when the code already holds
334 a valid reference to the rsb, so there's no need for locking. */
336 static inline void hold_rsb(struct dlm_rsb *r)
338 kref_get(&r->res_ref);
341 void dlm_hold_rsb(struct dlm_rsb *r)
346 /* When all references to the rsb are gone it's transferred to
347 the tossed list for later disposal. */
349 static void put_rsb(struct dlm_rsb *r)
351 struct dlm_ls *ls = r->res_ls;
352 uint32_t bucket = r->res_bucket;
354 spin_lock(&ls->ls_rsbtbl[bucket].lock);
355 kref_put(&r->res_ref, toss_rsb);
356 spin_unlock(&ls->ls_rsbtbl[bucket].lock);
359 void dlm_put_rsb(struct dlm_rsb *r)
364 static int pre_rsb_struct(struct dlm_ls *ls)
366 struct dlm_rsb *r1, *r2;
369 spin_lock(&ls->ls_new_rsb_spin);
370 if (ls->ls_new_rsb_count > dlm_config.ci_new_rsb_count / 2) {
371 spin_unlock(&ls->ls_new_rsb_spin);
374 spin_unlock(&ls->ls_new_rsb_spin);
376 r1 = dlm_allocate_rsb(ls);
377 r2 = dlm_allocate_rsb(ls);
379 spin_lock(&ls->ls_new_rsb_spin);
381 list_add(&r1->res_hashchain, &ls->ls_new_rsb);
382 ls->ls_new_rsb_count++;
385 list_add(&r2->res_hashchain, &ls->ls_new_rsb);
386 ls->ls_new_rsb_count++;
388 count = ls->ls_new_rsb_count;
389 spin_unlock(&ls->ls_new_rsb_spin);
396 /* If ls->ls_new_rsb is empty, return -EAGAIN, so the caller can
397 unlock any spinlocks, go back and call pre_rsb_struct again.
398 Otherwise, take an rsb off the list and return it. */
400 static int get_rsb_struct(struct dlm_ls *ls, char *name, int len,
401 struct dlm_rsb **r_ret)
406 spin_lock(&ls->ls_new_rsb_spin);
407 if (list_empty(&ls->ls_new_rsb)) {
408 count = ls->ls_new_rsb_count;
409 spin_unlock(&ls->ls_new_rsb_spin);
410 log_debug(ls, "find_rsb retry %d %d %s",
411 count, dlm_config.ci_new_rsb_count, name);
415 r = list_first_entry(&ls->ls_new_rsb, struct dlm_rsb, res_hashchain);
416 list_del(&r->res_hashchain);
417 /* Convert the empty list_head to a NULL rb_node for tree usage: */
418 memset(&r->res_hashnode, 0, sizeof(struct rb_node));
419 ls->ls_new_rsb_count--;
420 spin_unlock(&ls->ls_new_rsb_spin);
424 memcpy(r->res_name, name, len);
425 mutex_init(&r->res_mutex);
427 INIT_LIST_HEAD(&r->res_lookup);
428 INIT_LIST_HEAD(&r->res_grantqueue);
429 INIT_LIST_HEAD(&r->res_convertqueue);
430 INIT_LIST_HEAD(&r->res_waitqueue);
431 INIT_LIST_HEAD(&r->res_root_list);
432 INIT_LIST_HEAD(&r->res_recover_list);
438 static int rsb_cmp(struct dlm_rsb *r, const char *name, int nlen)
440 char maxname[DLM_RESNAME_MAXLEN];
442 memset(maxname, 0, DLM_RESNAME_MAXLEN);
443 memcpy(maxname, name, nlen);
444 return memcmp(r->res_name, maxname, DLM_RESNAME_MAXLEN);
447 int dlm_search_rsb_tree(struct rb_root *tree, char *name, int len,
448 struct dlm_rsb **r_ret)
450 struct rb_node *node = tree->rb_node;
455 r = rb_entry(node, struct dlm_rsb, res_hashnode);
456 rc = rsb_cmp(r, name, len);
458 node = node->rb_left;
460 node = node->rb_right;
472 static int rsb_insert(struct dlm_rsb *rsb, struct rb_root *tree)
474 struct rb_node **newn = &tree->rb_node;
475 struct rb_node *parent = NULL;
479 struct dlm_rsb *cur = rb_entry(*newn, struct dlm_rsb,
483 rc = rsb_cmp(cur, rsb->res_name, rsb->res_length);
485 newn = &parent->rb_left;
487 newn = &parent->rb_right;
489 log_print("rsb_insert match");
496 rb_link_node(&rsb->res_hashnode, parent, newn);
497 rb_insert_color(&rsb->res_hashnode, tree);
502 * Find rsb in rsbtbl and potentially create/add one
504 * Delaying the release of rsb's has a similar benefit to applications keeping
505 * NL locks on an rsb, but without the guarantee that the cached master value
506 * will still be valid when the rsb is reused. Apps aren't always smart enough
507 * to keep NL locks on an rsb that they may lock again shortly; this can lead
508 * to excessive master lookups and removals if we don't delay the release.
510 * Searching for an rsb means looking through both the normal list and toss
511 * list. When found on the toss list the rsb is moved to the normal list with
512 * ref count of 1; when found on normal list the ref count is incremented.
514 * rsb's on the keep list are being used locally and refcounted.
515 * rsb's on the toss list are not being used locally, and are not refcounted.
517 * The toss list rsb's were either
518 * - previously used locally but not any more (were on keep list, then
519 * moved to toss list when last refcount dropped)
520 * - created and put on toss list as a directory record for a lookup
521 * (we are the dir node for the res, but are not using the res right now,
522 * but some other node is)
524 * The purpose of find_rsb() is to return a refcounted rsb for local use.
525 * So, if the given rsb is on the toss list, it is moved to the keep list
526 * before being returned.
528 * toss_rsb() happens when all local usage of the rsb is done, i.e. no
529 * more refcounts exist, so the rsb is moved from the keep list to the
532 * rsb's on both keep and toss lists are used for doing a name to master
533 * lookups. rsb's that are in use locally (and being refcounted) are on
534 * the keep list, rsb's that are not in use locally (not refcounted) and
535 * only exist for name/master lookups are on the toss list.
537 * rsb's on the toss list who's dir_nodeid is not local can have stale
538 * name/master mappings. So, remote requests on such rsb's can potentially
539 * return with an error, which means the mapping is stale and needs to
540 * be updated with a new lookup. (The idea behind MASTER UNCERTAIN and
541 * first_lkid is to keep only a single outstanding request on an rsb
542 * while that rsb has a potentially stale master.)
545 static int find_rsb_dir(struct dlm_ls *ls, char *name, int len,
546 uint32_t hash, uint32_t b,
547 int dir_nodeid, int from_nodeid,
548 unsigned int flags, struct dlm_rsb **r_ret)
550 struct dlm_rsb *r = NULL;
551 int our_nodeid = dlm_our_nodeid();
558 if (flags & R_RECEIVE_REQUEST) {
559 if (from_nodeid == dir_nodeid)
563 } else if (flags & R_REQUEST) {
568 * flags & R_RECEIVE_RECOVER is from dlm_recover_master_copy, so
569 * from_nodeid has sent us a lock in dlm_recover_locks, believing
570 * we're the new master. Our local recovery may not have set
571 * res_master_nodeid to our_nodeid yet, so allow either. Don't
572 * create the rsb; dlm_recover_process_copy() will handle EBADR
575 * If someone sends us a request, we are the dir node, and we do
576 * not find the rsb anywhere, then recreate it. This happens if
577 * someone sends us a request after we have removed/freed an rsb
578 * from our toss list. (They sent a request instead of lookup
579 * because they are using an rsb from their toss list.)
582 if (from_local || from_dir ||
583 (from_other && (dir_nodeid == our_nodeid))) {
589 error = pre_rsb_struct(ls);
594 spin_lock(&ls->ls_rsbtbl[b].lock);
596 error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].keep, name, len, &r);
601 * rsb is active, so we can't check master_nodeid without lock_rsb.
604 kref_get(&r->res_ref);
610 error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].toss, name, len, &r);
615 * rsb found inactive (master_nodeid may be out of date unless
616 * we are the dir_nodeid or were the master) No other thread
617 * is using this rsb because it's on the toss list, so we can
618 * look at or update res_master_nodeid without lock_rsb.
621 if ((r->res_master_nodeid != our_nodeid) && from_other) {
622 /* our rsb was not master, and another node (not the dir node)
623 has sent us a request */
624 log_debug(ls, "find_rsb toss from_other %d master %d dir %d %s",
625 from_nodeid, r->res_master_nodeid, dir_nodeid,
631 if ((r->res_master_nodeid != our_nodeid) && from_dir) {
632 /* don't think this should ever happen */
633 log_error(ls, "find_rsb toss from_dir %d master %d",
634 from_nodeid, r->res_master_nodeid);
636 /* fix it and go on */
637 r->res_master_nodeid = our_nodeid;
639 rsb_clear_flag(r, RSB_MASTER_UNCERTAIN);
640 r->res_first_lkid = 0;
643 if (from_local && (r->res_master_nodeid != our_nodeid)) {
644 /* Because we have held no locks on this rsb,
645 res_master_nodeid could have become stale. */
646 rsb_set_flag(r, RSB_MASTER_UNCERTAIN);
647 r->res_first_lkid = 0;
650 rb_erase(&r->res_hashnode, &ls->ls_rsbtbl[b].toss);
651 error = rsb_insert(r, &ls->ls_rsbtbl[b].keep);
660 if (error == -EBADR && !create)
663 error = get_rsb_struct(ls, name, len, &r);
664 if (error == -EAGAIN) {
665 spin_unlock(&ls->ls_rsbtbl[b].lock);
673 r->res_dir_nodeid = dir_nodeid;
674 kref_init(&r->res_ref);
677 /* want to see how often this happens */
678 log_debug(ls, "find_rsb new from_dir %d recreate %s",
679 from_nodeid, r->res_name);
680 r->res_master_nodeid = our_nodeid;
685 if (from_other && (dir_nodeid != our_nodeid)) {
686 /* should never happen */
687 log_error(ls, "find_rsb new from_other %d dir %d our %d %s",
688 from_nodeid, dir_nodeid, our_nodeid, r->res_name);
696 log_debug(ls, "find_rsb new from_other %d dir %d %s",
697 from_nodeid, dir_nodeid, r->res_name);
700 if (dir_nodeid == our_nodeid) {
701 /* When we are the dir nodeid, we can set the master
703 r->res_master_nodeid = our_nodeid;
706 /* set_master will send_lookup to dir_nodeid */
707 r->res_master_nodeid = 0;
712 error = rsb_insert(r, &ls->ls_rsbtbl[b].keep);
714 spin_unlock(&ls->ls_rsbtbl[b].lock);
720 /* During recovery, other nodes can send us new MSTCPY locks (from
721 dlm_recover_locks) before we've made ourself master (in
722 dlm_recover_masters). */
724 static int find_rsb_nodir(struct dlm_ls *ls, char *name, int len,
725 uint32_t hash, uint32_t b,
726 int dir_nodeid, int from_nodeid,
727 unsigned int flags, struct dlm_rsb **r_ret)
729 struct dlm_rsb *r = NULL;
730 int our_nodeid = dlm_our_nodeid();
731 int recover = (flags & R_RECEIVE_RECOVER);
735 error = pre_rsb_struct(ls);
739 spin_lock(&ls->ls_rsbtbl[b].lock);
741 error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].keep, name, len, &r);
746 * rsb is active, so we can't check master_nodeid without lock_rsb.
749 kref_get(&r->res_ref);
754 error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].toss, name, len, &r);
759 * rsb found inactive. No other thread is using this rsb because
760 * it's on the toss list, so we can look at or update
761 * res_master_nodeid without lock_rsb.
764 if (!recover && (r->res_master_nodeid != our_nodeid) && from_nodeid) {
765 /* our rsb is not master, and another node has sent us a
766 request; this should never happen */
767 log_error(ls, "find_rsb toss from_nodeid %d master %d dir %d",
768 from_nodeid, r->res_master_nodeid, dir_nodeid);
774 if (!recover && (r->res_master_nodeid != our_nodeid) &&
775 (dir_nodeid == our_nodeid)) {
776 /* our rsb is not master, and we are dir; may as well fix it;
777 this should never happen */
778 log_error(ls, "find_rsb toss our %d master %d dir %d",
779 our_nodeid, r->res_master_nodeid, dir_nodeid);
781 r->res_master_nodeid = our_nodeid;
785 rb_erase(&r->res_hashnode, &ls->ls_rsbtbl[b].toss);
786 error = rsb_insert(r, &ls->ls_rsbtbl[b].keep);
795 error = get_rsb_struct(ls, name, len, &r);
796 if (error == -EAGAIN) {
797 spin_unlock(&ls->ls_rsbtbl[b].lock);
805 r->res_dir_nodeid = dir_nodeid;
806 r->res_master_nodeid = dir_nodeid;
807 r->res_nodeid = (dir_nodeid == our_nodeid) ? 0 : dir_nodeid;
808 kref_init(&r->res_ref);
810 error = rsb_insert(r, &ls->ls_rsbtbl[b].keep);
812 spin_unlock(&ls->ls_rsbtbl[b].lock);
818 static int find_rsb(struct dlm_ls *ls, char *name, int len, int from_nodeid,
819 unsigned int flags, struct dlm_rsb **r_ret)
824 if (len > DLM_RESNAME_MAXLEN)
827 hash = jhash(name, len, 0);
828 b = hash & (ls->ls_rsbtbl_size - 1);
830 dir_nodeid = dlm_hash2nodeid(ls, hash);
832 if (dlm_no_directory(ls))
833 return find_rsb_nodir(ls, name, len, hash, b, dir_nodeid,
834 from_nodeid, flags, r_ret);
836 return find_rsb_dir(ls, name, len, hash, b, dir_nodeid,
837 from_nodeid, flags, r_ret);
840 /* we have received a request and found that res_master_nodeid != our_nodeid,
841 so we need to return an error or make ourself the master */
843 static int validate_master_nodeid(struct dlm_ls *ls, struct dlm_rsb *r,
846 if (dlm_no_directory(ls)) {
847 log_error(ls, "find_rsb keep from_nodeid %d master %d dir %d",
848 from_nodeid, r->res_master_nodeid,
854 if (from_nodeid != r->res_dir_nodeid) {
855 /* our rsb is not master, and another node (not the dir node)
856 has sent us a request. this is much more common when our
857 master_nodeid is zero, so limit debug to non-zero. */
859 if (r->res_master_nodeid) {
860 log_debug(ls, "validate master from_other %d master %d "
861 "dir %d first %x %s", from_nodeid,
862 r->res_master_nodeid, r->res_dir_nodeid,
863 r->res_first_lkid, r->res_name);
867 /* our rsb is not master, but the dir nodeid has sent us a
868 request; this could happen with master 0 / res_nodeid -1 */
870 if (r->res_master_nodeid) {
871 log_error(ls, "validate master from_dir %d master %d "
873 from_nodeid, r->res_master_nodeid,
874 r->res_first_lkid, r->res_name);
877 r->res_master_nodeid = dlm_our_nodeid();
884 * We're the dir node for this res and another node wants to know the
885 * master nodeid. During normal operation (non recovery) this is only
886 * called from receive_lookup(); master lookups when the local node is
887 * the dir node are done by find_rsb().
889 * normal operation, we are the dir node for a resource
894 * . dlm_master_lookup flags 0
896 * recover directory, we are rebuilding dir for all resources
897 * . dlm_recover_directory
899 * remote node sends back the rsb names it is master of and we are dir of
900 * . dlm_master_lookup RECOVER_DIR (fix_master 0, from_master 1)
901 * we either create new rsb setting remote node as master, or find existing
902 * rsb and set master to be the remote node.
904 * recover masters, we are finding the new master for resources
905 * . dlm_recover_masters
907 * . dlm_send_rcom_lookup
908 * . receive_rcom_lookup
909 * . dlm_master_lookup RECOVER_MASTER (fix_master 1, from_master 0)
912 int dlm_master_lookup(struct dlm_ls *ls, int from_nodeid, char *name, int len,
913 unsigned int flags, int *r_nodeid, int *result)
915 struct dlm_rsb *r = NULL;
917 int from_master = (flags & DLM_LU_RECOVER_DIR);
918 int fix_master = (flags & DLM_LU_RECOVER_MASTER);
919 int our_nodeid = dlm_our_nodeid();
920 int dir_nodeid, error, toss_list = 0;
922 if (len > DLM_RESNAME_MAXLEN)
925 if (from_nodeid == our_nodeid) {
926 log_error(ls, "dlm_master_lookup from our_nodeid %d flags %x",
931 hash = jhash(name, len, 0);
932 b = hash & (ls->ls_rsbtbl_size - 1);
934 dir_nodeid = dlm_hash2nodeid(ls, hash);
935 if (dir_nodeid != our_nodeid) {
936 log_error(ls, "dlm_master_lookup from %d dir %d our %d h %x %d",
937 from_nodeid, dir_nodeid, our_nodeid, hash,
944 error = pre_rsb_struct(ls);
948 spin_lock(&ls->ls_rsbtbl[b].lock);
949 error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].keep, name, len, &r);
951 /* because the rsb is active, we need to lock_rsb before
952 checking/changing re_master_nodeid */
955 spin_unlock(&ls->ls_rsbtbl[b].lock);
960 error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].toss, name, len, &r);
964 /* because the rsb is inactive (on toss list), it's not refcounted
965 and lock_rsb is not used, but is protected by the rsbtbl lock */
969 if (r->res_dir_nodeid != our_nodeid) {
970 /* should not happen, but may as well fix it and carry on */
971 log_error(ls, "dlm_master_lookup res_dir %d our %d %s",
972 r->res_dir_nodeid, our_nodeid, r->res_name);
973 r->res_dir_nodeid = our_nodeid;
976 if (fix_master && dlm_is_removed(ls, r->res_master_nodeid)) {
977 /* Recovery uses this function to set a new master when
978 the previous master failed. Setting NEW_MASTER will
979 force dlm_recover_masters to call recover_master on this
980 rsb even though the res_nodeid is no longer removed. */
982 r->res_master_nodeid = from_nodeid;
983 r->res_nodeid = from_nodeid;
984 rsb_set_flag(r, RSB_NEW_MASTER);
987 /* I don't think we should ever find it on toss list. */
988 log_error(ls, "dlm_master_lookup fix_master on toss");
993 if (from_master && (r->res_master_nodeid != from_nodeid)) {
994 /* this will happen if from_nodeid became master during
995 a previous recovery cycle, and we aborted the previous
996 cycle before recovering this master value */
998 log_limit(ls, "dlm_master_lookup from_master %d "
999 "master_nodeid %d res_nodeid %d first %x %s",
1000 from_nodeid, r->res_master_nodeid, r->res_nodeid,
1001 r->res_first_lkid, r->res_name);
1003 if (r->res_master_nodeid == our_nodeid) {
1004 log_error(ls, "from_master %d our_master", from_nodeid);
1009 r->res_master_nodeid = from_nodeid;
1010 r->res_nodeid = from_nodeid;
1011 rsb_set_flag(r, RSB_NEW_MASTER);
1014 if (!r->res_master_nodeid) {
1015 /* this will happen if recovery happens while we're looking
1016 up the master for this rsb */
1018 log_debug(ls, "dlm_master_lookup master 0 to %d first %x %s",
1019 from_nodeid, r->res_first_lkid, r->res_name);
1020 r->res_master_nodeid = from_nodeid;
1021 r->res_nodeid = from_nodeid;
1024 if (!from_master && !fix_master &&
1025 (r->res_master_nodeid == from_nodeid)) {
1026 /* this can happen when the master sends remove, the dir node
1027 finds the rsb on the keep list and ignores the remove,
1028 and the former master sends a lookup */
1030 log_limit(ls, "dlm_master_lookup from master %d flags %x "
1031 "first %x %s", from_nodeid, flags,
1032 r->res_first_lkid, r->res_name);
1036 *r_nodeid = r->res_master_nodeid;
1038 *result = DLM_LU_MATCH;
1041 r->res_toss_time = jiffies;
1042 /* the rsb was inactive (on toss list) */
1043 spin_unlock(&ls->ls_rsbtbl[b].lock);
1045 /* the rsb was active */
1052 error = get_rsb_struct(ls, name, len, &r);
1053 if (error == -EAGAIN) {
1054 spin_unlock(&ls->ls_rsbtbl[b].lock);
1062 r->res_dir_nodeid = our_nodeid;
1063 r->res_master_nodeid = from_nodeid;
1064 r->res_nodeid = from_nodeid;
1065 kref_init(&r->res_ref);
1066 r->res_toss_time = jiffies;
1068 error = rsb_insert(r, &ls->ls_rsbtbl[b].toss);
1070 /* should never happen */
1072 spin_unlock(&ls->ls_rsbtbl[b].lock);
1077 *result = DLM_LU_ADD;
1078 *r_nodeid = from_nodeid;
1081 spin_unlock(&ls->ls_rsbtbl[b].lock);
1085 static void dlm_dump_rsb_hash(struct dlm_ls *ls, uint32_t hash)
1091 for (i = 0; i < ls->ls_rsbtbl_size; i++) {
1092 spin_lock(&ls->ls_rsbtbl[i].lock);
1093 for (n = rb_first(&ls->ls_rsbtbl[i].keep); n; n = rb_next(n)) {
1094 r = rb_entry(n, struct dlm_rsb, res_hashnode);
1095 if (r->res_hash == hash)
1098 spin_unlock(&ls->ls_rsbtbl[i].lock);
1102 void dlm_dump_rsb_name(struct dlm_ls *ls, char *name, int len)
1104 struct dlm_rsb *r = NULL;
1108 hash = jhash(name, len, 0);
1109 b = hash & (ls->ls_rsbtbl_size - 1);
1111 spin_lock(&ls->ls_rsbtbl[b].lock);
1112 error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].keep, name, len, &r);
1116 error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].toss, name, len, &r);
1122 spin_unlock(&ls->ls_rsbtbl[b].lock);
1125 static void toss_rsb(struct kref *kref)
1127 struct dlm_rsb *r = container_of(kref, struct dlm_rsb, res_ref);
1128 struct dlm_ls *ls = r->res_ls;
1130 DLM_ASSERT(list_empty(&r->res_root_list), dlm_print_rsb(r););
1131 kref_init(&r->res_ref);
1132 rb_erase(&r->res_hashnode, &ls->ls_rsbtbl[r->res_bucket].keep);
1133 rsb_insert(r, &ls->ls_rsbtbl[r->res_bucket].toss);
1134 r->res_toss_time = jiffies;
1135 ls->ls_rsbtbl[r->res_bucket].flags |= DLM_RTF_SHRINK;
1136 if (r->res_lvbptr) {
1137 dlm_free_lvb(r->res_lvbptr);
1138 r->res_lvbptr = NULL;
1142 /* See comment for unhold_lkb */
1144 static void unhold_rsb(struct dlm_rsb *r)
1147 rv = kref_put(&r->res_ref, toss_rsb);
1148 DLM_ASSERT(!rv, dlm_dump_rsb(r););
1151 static void kill_rsb(struct kref *kref)
1153 struct dlm_rsb *r = container_of(kref, struct dlm_rsb, res_ref);
1155 /* All work is done after the return from kref_put() so we
1156 can release the write_lock before the remove and free. */
1158 DLM_ASSERT(list_empty(&r->res_lookup), dlm_dump_rsb(r););
1159 DLM_ASSERT(list_empty(&r->res_grantqueue), dlm_dump_rsb(r););
1160 DLM_ASSERT(list_empty(&r->res_convertqueue), dlm_dump_rsb(r););
1161 DLM_ASSERT(list_empty(&r->res_waitqueue), dlm_dump_rsb(r););
1162 DLM_ASSERT(list_empty(&r->res_root_list), dlm_dump_rsb(r););
1163 DLM_ASSERT(list_empty(&r->res_recover_list), dlm_dump_rsb(r););
1166 /* Attaching/detaching lkb's from rsb's is for rsb reference counting.
1167 The rsb must exist as long as any lkb's for it do. */
1169 static void attach_lkb(struct dlm_rsb *r, struct dlm_lkb *lkb)
1172 lkb->lkb_resource = r;
1175 static void detach_lkb(struct dlm_lkb *lkb)
1177 if (lkb->lkb_resource) {
1178 put_rsb(lkb->lkb_resource);
1179 lkb->lkb_resource = NULL;
1183 static int create_lkb(struct dlm_ls *ls, struct dlm_lkb **lkb_ret)
1185 struct dlm_lkb *lkb;
1188 lkb = dlm_allocate_lkb(ls);
1192 lkb->lkb_nodeid = -1;
1193 lkb->lkb_grmode = DLM_LOCK_IV;
1194 kref_init(&lkb->lkb_ref);
1195 INIT_LIST_HEAD(&lkb->lkb_ownqueue);
1196 INIT_LIST_HEAD(&lkb->lkb_rsb_lookup);
1197 INIT_LIST_HEAD(&lkb->lkb_time_list);
1198 INIT_LIST_HEAD(&lkb->lkb_cb_list);
1199 mutex_init(&lkb->lkb_cb_mutex);
1200 INIT_WORK(&lkb->lkb_cb_work, dlm_callback_work);
1202 idr_preload(GFP_NOFS);
1203 spin_lock(&ls->ls_lkbidr_spin);
1204 rv = idr_alloc(&ls->ls_lkbidr, lkb, 1, 0, GFP_NOWAIT);
1207 spin_unlock(&ls->ls_lkbidr_spin);
1211 log_error(ls, "create_lkb idr error %d", rv);
1220 static int find_lkb(struct dlm_ls *ls, uint32_t lkid, struct dlm_lkb **lkb_ret)
1222 struct dlm_lkb *lkb;
1224 spin_lock(&ls->ls_lkbidr_spin);
1225 lkb = idr_find(&ls->ls_lkbidr, lkid);
1227 kref_get(&lkb->lkb_ref);
1228 spin_unlock(&ls->ls_lkbidr_spin);
1231 return lkb ? 0 : -ENOENT;
1234 static void kill_lkb(struct kref *kref)
1236 struct dlm_lkb *lkb = container_of(kref, struct dlm_lkb, lkb_ref);
1238 /* All work is done after the return from kref_put() so we
1239 can release the write_lock before the detach_lkb */
1241 DLM_ASSERT(!lkb->lkb_status, dlm_print_lkb(lkb););
1244 /* __put_lkb() is used when an lkb may not have an rsb attached to
1245 it so we need to provide the lockspace explicitly */
1247 static int __put_lkb(struct dlm_ls *ls, struct dlm_lkb *lkb)
1249 uint32_t lkid = lkb->lkb_id;
1251 spin_lock(&ls->ls_lkbidr_spin);
1252 if (kref_put(&lkb->lkb_ref, kill_lkb)) {
1253 idr_remove(&ls->ls_lkbidr, lkid);
1254 spin_unlock(&ls->ls_lkbidr_spin);
1258 /* for local/process lkbs, lvbptr points to caller's lksb */
1259 if (lkb->lkb_lvbptr && is_master_copy(lkb))
1260 dlm_free_lvb(lkb->lkb_lvbptr);
1264 spin_unlock(&ls->ls_lkbidr_spin);
1269 int dlm_put_lkb(struct dlm_lkb *lkb)
1273 DLM_ASSERT(lkb->lkb_resource, dlm_print_lkb(lkb););
1274 DLM_ASSERT(lkb->lkb_resource->res_ls, dlm_print_lkb(lkb););
1276 ls = lkb->lkb_resource->res_ls;
1277 return __put_lkb(ls, lkb);
1280 /* This is only called to add a reference when the code already holds
1281 a valid reference to the lkb, so there's no need for locking. */
1283 static inline void hold_lkb(struct dlm_lkb *lkb)
1285 kref_get(&lkb->lkb_ref);
1288 /* This is called when we need to remove a reference and are certain
1289 it's not the last ref. e.g. del_lkb is always called between a
1290 find_lkb/put_lkb and is always the inverse of a previous add_lkb.
1291 put_lkb would work fine, but would involve unnecessary locking */
1293 static inline void unhold_lkb(struct dlm_lkb *lkb)
1296 rv = kref_put(&lkb->lkb_ref, kill_lkb);
1297 DLM_ASSERT(!rv, dlm_print_lkb(lkb););
1300 static void lkb_add_ordered(struct list_head *new, struct list_head *head,
1303 struct dlm_lkb *lkb = NULL;
1305 list_for_each_entry(lkb, head, lkb_statequeue)
1306 if (lkb->lkb_rqmode < mode)
1309 __list_add(new, lkb->lkb_statequeue.prev, &lkb->lkb_statequeue);
1312 /* add/remove lkb to rsb's grant/convert/wait queue */
1314 static void add_lkb(struct dlm_rsb *r, struct dlm_lkb *lkb, int status)
1316 kref_get(&lkb->lkb_ref);
1318 DLM_ASSERT(!lkb->lkb_status, dlm_print_lkb(lkb););
1320 lkb->lkb_timestamp = ktime_get();
1322 lkb->lkb_status = status;
1325 case DLM_LKSTS_WAITING:
1326 if (lkb->lkb_exflags & DLM_LKF_HEADQUE)
1327 list_add(&lkb->lkb_statequeue, &r->res_waitqueue);
1329 list_add_tail(&lkb->lkb_statequeue, &r->res_waitqueue);
1331 case DLM_LKSTS_GRANTED:
1332 /* convention says granted locks kept in order of grmode */
1333 lkb_add_ordered(&lkb->lkb_statequeue, &r->res_grantqueue,
1336 case DLM_LKSTS_CONVERT:
1337 if (lkb->lkb_exflags & DLM_LKF_HEADQUE)
1338 list_add(&lkb->lkb_statequeue, &r->res_convertqueue);
1340 list_add_tail(&lkb->lkb_statequeue,
1341 &r->res_convertqueue);
1344 DLM_ASSERT(0, dlm_print_lkb(lkb); printk("sts=%d\n", status););
1348 static void del_lkb(struct dlm_rsb *r, struct dlm_lkb *lkb)
1350 lkb->lkb_status = 0;
1351 list_del(&lkb->lkb_statequeue);
1355 static void move_lkb(struct dlm_rsb *r, struct dlm_lkb *lkb, int sts)
1359 add_lkb(r, lkb, sts);
1363 static int msg_reply_type(int mstype)
1366 case DLM_MSG_REQUEST:
1367 return DLM_MSG_REQUEST_REPLY;
1368 case DLM_MSG_CONVERT:
1369 return DLM_MSG_CONVERT_REPLY;
1370 case DLM_MSG_UNLOCK:
1371 return DLM_MSG_UNLOCK_REPLY;
1372 case DLM_MSG_CANCEL:
1373 return DLM_MSG_CANCEL_REPLY;
1374 case DLM_MSG_LOOKUP:
1375 return DLM_MSG_LOOKUP_REPLY;
1380 static int nodeid_warned(int nodeid, int num_nodes, int *warned)
1384 for (i = 0; i < num_nodes; i++) {
1389 if (warned[i] == nodeid)
1395 void dlm_scan_waiters(struct dlm_ls *ls)
1397 struct dlm_lkb *lkb;
1399 s64 debug_maxus = 0;
1400 u32 debug_scanned = 0;
1401 u32 debug_expired = 0;
1405 if (!dlm_config.ci_waitwarn_us)
1408 mutex_lock(&ls->ls_waiters_mutex);
1410 list_for_each_entry(lkb, &ls->ls_waiters, lkb_wait_reply) {
1411 if (!lkb->lkb_wait_time)
1416 us = ktime_to_us(ktime_sub(ktime_get(), lkb->lkb_wait_time));
1418 if (us < dlm_config.ci_waitwarn_us)
1421 lkb->lkb_wait_time = 0;
1424 if (us > debug_maxus)
1428 num_nodes = ls->ls_num_nodes;
1429 warned = kcalloc(num_nodes, sizeof(int), GFP_KERNEL);
1433 if (nodeid_warned(lkb->lkb_wait_nodeid, num_nodes, warned))
1436 log_error(ls, "waitwarn %x %lld %d us check connection to "
1437 "node %d", lkb->lkb_id, (long long)us,
1438 dlm_config.ci_waitwarn_us, lkb->lkb_wait_nodeid);
1440 mutex_unlock(&ls->ls_waiters_mutex);
1444 log_debug(ls, "scan_waiters %u warn %u over %d us max %lld us",
1445 debug_scanned, debug_expired,
1446 dlm_config.ci_waitwarn_us, (long long)debug_maxus);
1449 /* add/remove lkb from global waiters list of lkb's waiting for
1450 a reply from a remote node */
1452 static int add_to_waiters(struct dlm_lkb *lkb, int mstype, int to_nodeid)
1454 struct dlm_ls *ls = lkb->lkb_resource->res_ls;
1457 mutex_lock(&ls->ls_waiters_mutex);
1459 if (is_overlap_unlock(lkb) ||
1460 (is_overlap_cancel(lkb) && (mstype == DLM_MSG_CANCEL))) {
1465 if (lkb->lkb_wait_type || is_overlap_cancel(lkb)) {
1467 case DLM_MSG_UNLOCK:
1468 lkb->lkb_flags |= DLM_IFL_OVERLAP_UNLOCK;
1470 case DLM_MSG_CANCEL:
1471 lkb->lkb_flags |= DLM_IFL_OVERLAP_CANCEL;
1477 lkb->lkb_wait_count++;
1480 log_debug(ls, "addwait %x cur %d overlap %d count %d f %x",
1481 lkb->lkb_id, lkb->lkb_wait_type, mstype,
1482 lkb->lkb_wait_count, lkb->lkb_flags);
1486 DLM_ASSERT(!lkb->lkb_wait_count,
1488 printk("wait_count %d\n", lkb->lkb_wait_count););
1490 lkb->lkb_wait_count++;
1491 lkb->lkb_wait_type = mstype;
1492 lkb->lkb_wait_time = ktime_get();
1493 lkb->lkb_wait_nodeid = to_nodeid; /* for debugging */
1495 list_add(&lkb->lkb_wait_reply, &ls->ls_waiters);
1498 log_error(ls, "addwait error %x %d flags %x %d %d %s",
1499 lkb->lkb_id, error, lkb->lkb_flags, mstype,
1500 lkb->lkb_wait_type, lkb->lkb_resource->res_name);
1501 mutex_unlock(&ls->ls_waiters_mutex);
1505 /* We clear the RESEND flag because we might be taking an lkb off the waiters
1506 list as part of process_requestqueue (e.g. a lookup that has an optimized
1507 request reply on the requestqueue) between dlm_recover_waiters_pre() which
1508 set RESEND and dlm_recover_waiters_post() */
1510 static int _remove_from_waiters(struct dlm_lkb *lkb, int mstype,
1511 struct dlm_message *ms)
1513 struct dlm_ls *ls = lkb->lkb_resource->res_ls;
1514 int overlap_done = 0;
1516 if (is_overlap_unlock(lkb) && (mstype == DLM_MSG_UNLOCK_REPLY)) {
1517 log_debug(ls, "remwait %x unlock_reply overlap", lkb->lkb_id);
1518 lkb->lkb_flags &= ~DLM_IFL_OVERLAP_UNLOCK;
1523 if (is_overlap_cancel(lkb) && (mstype == DLM_MSG_CANCEL_REPLY)) {
1524 log_debug(ls, "remwait %x cancel_reply overlap", lkb->lkb_id);
1525 lkb->lkb_flags &= ~DLM_IFL_OVERLAP_CANCEL;
1530 /* Cancel state was preemptively cleared by a successful convert,
1531 see next comment, nothing to do. */
1533 if ((mstype == DLM_MSG_CANCEL_REPLY) &&
1534 (lkb->lkb_wait_type != DLM_MSG_CANCEL)) {
1535 log_debug(ls, "remwait %x cancel_reply wait_type %d",
1536 lkb->lkb_id, lkb->lkb_wait_type);
1540 /* Remove for the convert reply, and premptively remove for the
1541 cancel reply. A convert has been granted while there's still
1542 an outstanding cancel on it (the cancel is moot and the result
1543 in the cancel reply should be 0). We preempt the cancel reply
1544 because the app gets the convert result and then can follow up
1545 with another op, like convert. This subsequent op would see the
1546 lingering state of the cancel and fail with -EBUSY. */
1548 if ((mstype == DLM_MSG_CONVERT_REPLY) &&
1549 (lkb->lkb_wait_type == DLM_MSG_CONVERT) &&
1550 is_overlap_cancel(lkb) && ms && !ms->m_result) {
1551 log_debug(ls, "remwait %x convert_reply zap overlap_cancel",
1553 lkb->lkb_wait_type = 0;
1554 lkb->lkb_flags &= ~DLM_IFL_OVERLAP_CANCEL;
1555 lkb->lkb_wait_count--;
1559 /* N.B. type of reply may not always correspond to type of original
1560 msg due to lookup->request optimization, verify others? */
1562 if (lkb->lkb_wait_type) {
1563 lkb->lkb_wait_type = 0;
1567 log_error(ls, "remwait error %x remote %d %x msg %d flags %x no wait",
1568 lkb->lkb_id, ms ? ms->m_header.h_nodeid : 0, lkb->lkb_remid,
1569 mstype, lkb->lkb_flags);
1573 /* the force-unlock/cancel has completed and we haven't recvd a reply
1574 to the op that was in progress prior to the unlock/cancel; we
1575 give up on any reply to the earlier op. FIXME: not sure when/how
1576 this would happen */
1578 if (overlap_done && lkb->lkb_wait_type) {
1579 log_error(ls, "remwait error %x reply %d wait_type %d overlap",
1580 lkb->lkb_id, mstype, lkb->lkb_wait_type);
1581 lkb->lkb_wait_count--;
1582 lkb->lkb_wait_type = 0;
1585 DLM_ASSERT(lkb->lkb_wait_count, dlm_print_lkb(lkb););
1587 lkb->lkb_flags &= ~DLM_IFL_RESEND;
1588 lkb->lkb_wait_count--;
1589 if (!lkb->lkb_wait_count)
1590 list_del_init(&lkb->lkb_wait_reply);
1595 static int remove_from_waiters(struct dlm_lkb *lkb, int mstype)
1597 struct dlm_ls *ls = lkb->lkb_resource->res_ls;
1600 mutex_lock(&ls->ls_waiters_mutex);
1601 error = _remove_from_waiters(lkb, mstype, NULL);
1602 mutex_unlock(&ls->ls_waiters_mutex);
1606 /* Handles situations where we might be processing a "fake" or "stub" reply in
1607 which we can't try to take waiters_mutex again. */
1609 static int remove_from_waiters_ms(struct dlm_lkb *lkb, struct dlm_message *ms)
1611 struct dlm_ls *ls = lkb->lkb_resource->res_ls;
1614 if (ms->m_flags != DLM_IFL_STUB_MS)
1615 mutex_lock(&ls->ls_waiters_mutex);
1616 error = _remove_from_waiters(lkb, ms->m_type, ms);
1617 if (ms->m_flags != DLM_IFL_STUB_MS)
1618 mutex_unlock(&ls->ls_waiters_mutex);
1622 /* If there's an rsb for the same resource being removed, ensure
1623 that the remove message is sent before the new lookup message.
1624 It should be rare to need a delay here, but if not, then it may
1625 be worthwhile to add a proper wait mechanism rather than a delay. */
1627 static void wait_pending_remove(struct dlm_rsb *r)
1629 struct dlm_ls *ls = r->res_ls;
1631 spin_lock(&ls->ls_remove_spin);
1632 if (ls->ls_remove_len &&
1633 !rsb_cmp(r, ls->ls_remove_name, ls->ls_remove_len)) {
1634 log_debug(ls, "delay lookup for remove dir %d %s",
1635 r->res_dir_nodeid, r->res_name);
1636 spin_unlock(&ls->ls_remove_spin);
1640 spin_unlock(&ls->ls_remove_spin);
1644 * ls_remove_spin protects ls_remove_name and ls_remove_len which are
1645 * read by other threads in wait_pending_remove. ls_remove_names
1646 * and ls_remove_lens are only used by the scan thread, so they do
1647 * not need protection.
1650 static void shrink_bucket(struct dlm_ls *ls, int b)
1652 struct rb_node *n, *next;
1655 int our_nodeid = dlm_our_nodeid();
1656 int remote_count = 0;
1657 int need_shrink = 0;
1660 memset(&ls->ls_remove_lens, 0, sizeof(int) * DLM_REMOVE_NAMES_MAX);
1662 spin_lock(&ls->ls_rsbtbl[b].lock);
1664 if (!(ls->ls_rsbtbl[b].flags & DLM_RTF_SHRINK)) {
1665 spin_unlock(&ls->ls_rsbtbl[b].lock);
1669 for (n = rb_first(&ls->ls_rsbtbl[b].toss); n; n = next) {
1671 r = rb_entry(n, struct dlm_rsb, res_hashnode);
1673 /* If we're the directory record for this rsb, and
1674 we're not the master of it, then we need to wait
1675 for the master node to send us a dir remove for
1676 before removing the dir record. */
1678 if (!dlm_no_directory(ls) &&
1679 (r->res_master_nodeid != our_nodeid) &&
1680 (dlm_dir_nodeid(r) == our_nodeid)) {
1686 if (!time_after_eq(jiffies, r->res_toss_time +
1687 dlm_config.ci_toss_secs * HZ)) {
1691 if (!dlm_no_directory(ls) &&
1692 (r->res_master_nodeid == our_nodeid) &&
1693 (dlm_dir_nodeid(r) != our_nodeid)) {
1695 /* We're the master of this rsb but we're not
1696 the directory record, so we need to tell the
1697 dir node to remove the dir record. */
1699 ls->ls_remove_lens[remote_count] = r->res_length;
1700 memcpy(ls->ls_remove_names[remote_count], r->res_name,
1701 DLM_RESNAME_MAXLEN);
1704 if (remote_count >= DLM_REMOVE_NAMES_MAX)
1709 if (!kref_put(&r->res_ref, kill_rsb)) {
1710 log_error(ls, "tossed rsb in use %s", r->res_name);
1714 rb_erase(&r->res_hashnode, &ls->ls_rsbtbl[b].toss);
1719 ls->ls_rsbtbl[b].flags |= DLM_RTF_SHRINK;
1721 ls->ls_rsbtbl[b].flags &= ~DLM_RTF_SHRINK;
1722 spin_unlock(&ls->ls_rsbtbl[b].lock);
1725 * While searching for rsb's to free, we found some that require
1726 * remote removal. We leave them in place and find them again here
1727 * so there is a very small gap between removing them from the toss
1728 * list and sending the removal. Keeping this gap small is
1729 * important to keep us (the master node) from being out of sync
1730 * with the remote dir node for very long.
1732 * From the time the rsb is removed from toss until just after
1733 * send_remove, the rsb name is saved in ls_remove_name. A new
1734 * lookup checks this to ensure that a new lookup message for the
1735 * same resource name is not sent just before the remove message.
1738 for (i = 0; i < remote_count; i++) {
1739 name = ls->ls_remove_names[i];
1740 len = ls->ls_remove_lens[i];
1742 spin_lock(&ls->ls_rsbtbl[b].lock);
1743 rv = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].toss, name, len, &r);
1745 spin_unlock(&ls->ls_rsbtbl[b].lock);
1746 log_debug(ls, "remove_name not toss %s", name);
1750 if (r->res_master_nodeid != our_nodeid) {
1751 spin_unlock(&ls->ls_rsbtbl[b].lock);
1752 log_debug(ls, "remove_name master %d dir %d our %d %s",
1753 r->res_master_nodeid, r->res_dir_nodeid,
1758 if (r->res_dir_nodeid == our_nodeid) {
1759 /* should never happen */
1760 spin_unlock(&ls->ls_rsbtbl[b].lock);
1761 log_error(ls, "remove_name dir %d master %d our %d %s",
1762 r->res_dir_nodeid, r->res_master_nodeid,
1767 if (!time_after_eq(jiffies, r->res_toss_time +
1768 dlm_config.ci_toss_secs * HZ)) {
1769 spin_unlock(&ls->ls_rsbtbl[b].lock);
1770 log_debug(ls, "remove_name toss_time %lu now %lu %s",
1771 r->res_toss_time, jiffies, name);
1775 if (!kref_put(&r->res_ref, kill_rsb)) {
1776 spin_unlock(&ls->ls_rsbtbl[b].lock);
1777 log_error(ls, "remove_name in use %s", name);
1781 rb_erase(&r->res_hashnode, &ls->ls_rsbtbl[b].toss);
1783 /* block lookup of same name until we've sent remove */
1784 spin_lock(&ls->ls_remove_spin);
1785 ls->ls_remove_len = len;
1786 memcpy(ls->ls_remove_name, name, DLM_RESNAME_MAXLEN);
1787 spin_unlock(&ls->ls_remove_spin);
1788 spin_unlock(&ls->ls_rsbtbl[b].lock);
1792 /* allow lookup of name again */
1793 spin_lock(&ls->ls_remove_spin);
1794 ls->ls_remove_len = 0;
1795 memset(ls->ls_remove_name, 0, DLM_RESNAME_MAXLEN);
1796 spin_unlock(&ls->ls_remove_spin);
1802 void dlm_scan_rsbs(struct dlm_ls *ls)
1806 for (i = 0; i < ls->ls_rsbtbl_size; i++) {
1807 shrink_bucket(ls, i);
1808 if (dlm_locking_stopped(ls))
1814 static void add_timeout(struct dlm_lkb *lkb)
1816 struct dlm_ls *ls = lkb->lkb_resource->res_ls;
1818 if (is_master_copy(lkb))
1821 if (test_bit(LSFL_TIMEWARN, &ls->ls_flags) &&
1822 !(lkb->lkb_exflags & DLM_LKF_NODLCKWT)) {
1823 lkb->lkb_flags |= DLM_IFL_WATCH_TIMEWARN;
1826 if (lkb->lkb_exflags & DLM_LKF_TIMEOUT)
1831 DLM_ASSERT(list_empty(&lkb->lkb_time_list), dlm_print_lkb(lkb););
1832 mutex_lock(&ls->ls_timeout_mutex);
1834 list_add_tail(&lkb->lkb_time_list, &ls->ls_timeout);
1835 mutex_unlock(&ls->ls_timeout_mutex);
1838 static void del_timeout(struct dlm_lkb *lkb)
1840 struct dlm_ls *ls = lkb->lkb_resource->res_ls;
1842 mutex_lock(&ls->ls_timeout_mutex);
1843 if (!list_empty(&lkb->lkb_time_list)) {
1844 list_del_init(&lkb->lkb_time_list);
1847 mutex_unlock(&ls->ls_timeout_mutex);
1850 /* FIXME: is it safe to look at lkb_exflags, lkb_flags, lkb_timestamp, and
1851 lkb_lksb_timeout without lock_rsb? Note: we can't lock timeout_mutex
1852 and then lock rsb because of lock ordering in add_timeout. We may need
1853 to specify some special timeout-related bits in the lkb that are just to
1854 be accessed under the timeout_mutex. */
1856 void dlm_scan_timeout(struct dlm_ls *ls)
1859 struct dlm_lkb *lkb;
1860 int do_cancel, do_warn;
1864 if (dlm_locking_stopped(ls))
1869 mutex_lock(&ls->ls_timeout_mutex);
1870 list_for_each_entry(lkb, &ls->ls_timeout, lkb_time_list) {
1872 wait_us = ktime_to_us(ktime_sub(ktime_get(),
1873 lkb->lkb_timestamp));
1875 if ((lkb->lkb_exflags & DLM_LKF_TIMEOUT) &&
1876 wait_us >= (lkb->lkb_timeout_cs * 10000))
1879 if ((lkb->lkb_flags & DLM_IFL_WATCH_TIMEWARN) &&
1880 wait_us >= dlm_config.ci_timewarn_cs * 10000)
1883 if (!do_cancel && !do_warn)
1888 mutex_unlock(&ls->ls_timeout_mutex);
1890 if (!do_cancel && !do_warn)
1893 r = lkb->lkb_resource;
1898 /* clear flag so we only warn once */
1899 lkb->lkb_flags &= ~DLM_IFL_WATCH_TIMEWARN;
1900 if (!(lkb->lkb_exflags & DLM_LKF_TIMEOUT))
1902 dlm_timeout_warn(lkb);
1906 log_debug(ls, "timeout cancel %x node %d %s",
1907 lkb->lkb_id, lkb->lkb_nodeid, r->res_name);
1908 lkb->lkb_flags &= ~DLM_IFL_WATCH_TIMEWARN;
1909 lkb->lkb_flags |= DLM_IFL_TIMEOUT_CANCEL;
1911 _cancel_lock(r, lkb);
1920 /* This is only called by dlm_recoverd, and we rely on dlm_ls_stop() stopping
1921 dlm_recoverd before checking/setting ls_recover_begin. */
1923 void dlm_adjust_timeouts(struct dlm_ls *ls)
1925 struct dlm_lkb *lkb;
1926 u64 adj_us = jiffies_to_usecs(jiffies - ls->ls_recover_begin);
1928 ls->ls_recover_begin = 0;
1929 mutex_lock(&ls->ls_timeout_mutex);
1930 list_for_each_entry(lkb, &ls->ls_timeout, lkb_time_list)
1931 lkb->lkb_timestamp = ktime_add_us(lkb->lkb_timestamp, adj_us);
1932 mutex_unlock(&ls->ls_timeout_mutex);
1934 if (!dlm_config.ci_waitwarn_us)
1937 mutex_lock(&ls->ls_waiters_mutex);
1938 list_for_each_entry(lkb, &ls->ls_waiters, lkb_wait_reply) {
1939 if (ktime_to_us(lkb->lkb_wait_time))
1940 lkb->lkb_wait_time = ktime_get();
1942 mutex_unlock(&ls->ls_waiters_mutex);
1945 /* lkb is master or local copy */
1947 static void set_lvb_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
1949 int b, len = r->res_ls->ls_lvblen;
1951 /* b=1 lvb returned to caller
1952 b=0 lvb written to rsb or invalidated
1955 b = dlm_lvb_operations[lkb->lkb_grmode + 1][lkb->lkb_rqmode + 1];
1958 if (!lkb->lkb_lvbptr)
1961 if (!(lkb->lkb_exflags & DLM_LKF_VALBLK))
1967 memcpy(lkb->lkb_lvbptr, r->res_lvbptr, len);
1968 lkb->lkb_lvbseq = r->res_lvbseq;
1970 } else if (b == 0) {
1971 if (lkb->lkb_exflags & DLM_LKF_IVVALBLK) {
1972 rsb_set_flag(r, RSB_VALNOTVALID);
1976 if (!lkb->lkb_lvbptr)
1979 if (!(lkb->lkb_exflags & DLM_LKF_VALBLK))
1983 r->res_lvbptr = dlm_allocate_lvb(r->res_ls);
1988 memcpy(r->res_lvbptr, lkb->lkb_lvbptr, len);
1990 lkb->lkb_lvbseq = r->res_lvbseq;
1991 rsb_clear_flag(r, RSB_VALNOTVALID);
1994 if (rsb_flag(r, RSB_VALNOTVALID))
1995 lkb->lkb_sbflags |= DLM_SBF_VALNOTVALID;
1998 static void set_lvb_unlock(struct dlm_rsb *r, struct dlm_lkb *lkb)
2000 if (lkb->lkb_grmode < DLM_LOCK_PW)
2003 if (lkb->lkb_exflags & DLM_LKF_IVVALBLK) {
2004 rsb_set_flag(r, RSB_VALNOTVALID);
2008 if (!lkb->lkb_lvbptr)
2011 if (!(lkb->lkb_exflags & DLM_LKF_VALBLK))
2015 r->res_lvbptr = dlm_allocate_lvb(r->res_ls);
2020 memcpy(r->res_lvbptr, lkb->lkb_lvbptr, r->res_ls->ls_lvblen);
2022 rsb_clear_flag(r, RSB_VALNOTVALID);
2025 /* lkb is process copy (pc) */
2027 static void set_lvb_lock_pc(struct dlm_rsb *r, struct dlm_lkb *lkb,
2028 struct dlm_message *ms)
2032 if (!lkb->lkb_lvbptr)
2035 if (!(lkb->lkb_exflags & DLM_LKF_VALBLK))
2038 b = dlm_lvb_operations[lkb->lkb_grmode + 1][lkb->lkb_rqmode + 1];
2040 int len = receive_extralen(ms);
2041 if (len > r->res_ls->ls_lvblen)
2042 len = r->res_ls->ls_lvblen;
2043 memcpy(lkb->lkb_lvbptr, ms->m_extra, len);
2044 lkb->lkb_lvbseq = ms->m_lvbseq;
2048 /* Manipulate lkb's on rsb's convert/granted/waiting queues
2049 remove_lock -- used for unlock, removes lkb from granted
2050 revert_lock -- used for cancel, moves lkb from convert to granted
2051 grant_lock -- used for request and convert, adds lkb to granted or
2052 moves lkb from convert or waiting to granted
2054 Each of these is used for master or local copy lkb's. There is
2055 also a _pc() variation used to make the corresponding change on
2056 a process copy (pc) lkb. */
2058 static void _remove_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
2061 lkb->lkb_grmode = DLM_LOCK_IV;
2062 /* this unhold undoes the original ref from create_lkb()
2063 so this leads to the lkb being freed */
2067 static void remove_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
2069 set_lvb_unlock(r, lkb);
2070 _remove_lock(r, lkb);
2073 static void remove_lock_pc(struct dlm_rsb *r, struct dlm_lkb *lkb)
2075 _remove_lock(r, lkb);
2078 /* returns: 0 did nothing
2079 1 moved lock to granted
2082 static int revert_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
2086 lkb->lkb_rqmode = DLM_LOCK_IV;
2088 switch (lkb->lkb_status) {
2089 case DLM_LKSTS_GRANTED:
2091 case DLM_LKSTS_CONVERT:
2092 move_lkb(r, lkb, DLM_LKSTS_GRANTED);
2095 case DLM_LKSTS_WAITING:
2097 lkb->lkb_grmode = DLM_LOCK_IV;
2098 /* this unhold undoes the original ref from create_lkb()
2099 so this leads to the lkb being freed */
2104 log_print("invalid status for revert %d", lkb->lkb_status);
2109 static int revert_lock_pc(struct dlm_rsb *r, struct dlm_lkb *lkb)
2111 return revert_lock(r, lkb);
2114 static void _grant_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
2116 if (lkb->lkb_grmode != lkb->lkb_rqmode) {
2117 lkb->lkb_grmode = lkb->lkb_rqmode;
2118 if (lkb->lkb_status)
2119 move_lkb(r, lkb, DLM_LKSTS_GRANTED);
2121 add_lkb(r, lkb, DLM_LKSTS_GRANTED);
2124 lkb->lkb_rqmode = DLM_LOCK_IV;
2125 lkb->lkb_highbast = 0;
2128 static void grant_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
2130 set_lvb_lock(r, lkb);
2131 _grant_lock(r, lkb);
2134 static void grant_lock_pc(struct dlm_rsb *r, struct dlm_lkb *lkb,
2135 struct dlm_message *ms)
2137 set_lvb_lock_pc(r, lkb, ms);
2138 _grant_lock(r, lkb);
2141 /* called by grant_pending_locks() which means an async grant message must
2142 be sent to the requesting node in addition to granting the lock if the
2143 lkb belongs to a remote node. */
2145 static void grant_lock_pending(struct dlm_rsb *r, struct dlm_lkb *lkb)
2148 if (is_master_copy(lkb))
2151 queue_cast(r, lkb, 0);
2154 /* The special CONVDEADLK, ALTPR and ALTCW flags allow the master to
2155 change the granted/requested modes. We're munging things accordingly in
2157 CONVDEADLK: our grmode may have been forced down to NL to resolve a
2159 ALTPR/ALTCW: our rqmode may have been changed to PR or CW to become
2160 compatible with other granted locks */
2162 static void munge_demoted(struct dlm_lkb *lkb)
2164 if (lkb->lkb_rqmode == DLM_LOCK_IV || lkb->lkb_grmode == DLM_LOCK_IV) {
2165 log_print("munge_demoted %x invalid modes gr %d rq %d",
2166 lkb->lkb_id, lkb->lkb_grmode, lkb->lkb_rqmode);
2170 lkb->lkb_grmode = DLM_LOCK_NL;
2173 static void munge_altmode(struct dlm_lkb *lkb, struct dlm_message *ms)
2175 if (ms->m_type != DLM_MSG_REQUEST_REPLY &&
2176 ms->m_type != DLM_MSG_GRANT) {
2177 log_print("munge_altmode %x invalid reply type %d",
2178 lkb->lkb_id, ms->m_type);
2182 if (lkb->lkb_exflags & DLM_LKF_ALTPR)
2183 lkb->lkb_rqmode = DLM_LOCK_PR;
2184 else if (lkb->lkb_exflags & DLM_LKF_ALTCW)
2185 lkb->lkb_rqmode = DLM_LOCK_CW;
2187 log_print("munge_altmode invalid exflags %x", lkb->lkb_exflags);
2192 static inline int first_in_list(struct dlm_lkb *lkb, struct list_head *head)
2194 struct dlm_lkb *first = list_entry(head->next, struct dlm_lkb,
2196 if (lkb->lkb_id == first->lkb_id)
2202 /* Check if the given lkb conflicts with another lkb on the queue. */
2204 static int queue_conflict(struct list_head *head, struct dlm_lkb *lkb)
2206 struct dlm_lkb *this;
2208 list_for_each_entry(this, head, lkb_statequeue) {
2211 if (!modes_compat(this, lkb))
2218 * "A conversion deadlock arises with a pair of lock requests in the converting
2219 * queue for one resource. The granted mode of each lock blocks the requested
2220 * mode of the other lock."
2222 * Part 2: if the granted mode of lkb is preventing an earlier lkb in the
2223 * convert queue from being granted, then deadlk/demote lkb.
2226 * Granted Queue: empty
2227 * Convert Queue: NL->EX (first lock)
2228 * PR->EX (second lock)
2230 * The first lock can't be granted because of the granted mode of the second
2231 * lock and the second lock can't be granted because it's not first in the
2232 * list. We either cancel lkb's conversion (PR->EX) and return EDEADLK, or we
2233 * demote the granted mode of lkb (from PR to NL) if it has the CONVDEADLK
2234 * flag set and return DEMOTED in the lksb flags.
2236 * Originally, this function detected conv-deadlk in a more limited scope:
2237 * - if !modes_compat(lkb1, lkb2) && !modes_compat(lkb2, lkb1), or
2238 * - if lkb1 was the first entry in the queue (not just earlier), and was
2239 * blocked by the granted mode of lkb2, and there was nothing on the
2240 * granted queue preventing lkb1 from being granted immediately, i.e.
2241 * lkb2 was the only thing preventing lkb1 from being granted.
2243 * That second condition meant we'd only say there was conv-deadlk if
2244 * resolving it (by demotion) would lead to the first lock on the convert
2245 * queue being granted right away. It allowed conversion deadlocks to exist
2246 * between locks on the convert queue while they couldn't be granted anyway.
2248 * Now, we detect and take action on conversion deadlocks immediately when
2249 * they're created, even if they may not be immediately consequential. If
2250 * lkb1 exists anywhere in the convert queue and lkb2 comes in with a granted
2251 * mode that would prevent lkb1's conversion from being granted, we do a
2252 * deadlk/demote on lkb2 right away and don't let it onto the convert queue.
2253 * I think this means that the lkb_is_ahead condition below should always
2254 * be zero, i.e. there will never be conv-deadlk between two locks that are
2255 * both already on the convert queue.
2258 static int conversion_deadlock_detect(struct dlm_rsb *r, struct dlm_lkb *lkb2)
2260 struct dlm_lkb *lkb1;
2261 int lkb_is_ahead = 0;
2263 list_for_each_entry(lkb1, &r->res_convertqueue, lkb_statequeue) {
2269 if (!lkb_is_ahead) {
2270 if (!modes_compat(lkb2, lkb1))
2273 if (!modes_compat(lkb2, lkb1) &&
2274 !modes_compat(lkb1, lkb2))
2282 * Return 1 if the lock can be granted, 0 otherwise.
2283 * Also detect and resolve conversion deadlocks.
2285 * lkb is the lock to be granted
2287 * now is 1 if the function is being called in the context of the
2288 * immediate request, it is 0 if called later, after the lock has been
2291 * recover is 1 if dlm_recover_grant() is trying to grant conversions
2294 * References are from chapter 6 of "VAXcluster Principles" by Roy Davis
2297 static int _can_be_granted(struct dlm_rsb *r, struct dlm_lkb *lkb, int now,
2300 int8_t conv = (lkb->lkb_grmode != DLM_LOCK_IV);
2303 * 6-10: Version 5.4 introduced an option to address the phenomenon of
2304 * a new request for a NL mode lock being blocked.
2306 * 6-11: If the optional EXPEDITE flag is used with the new NL mode
2307 * request, then it would be granted. In essence, the use of this flag
2308 * tells the Lock Manager to expedite theis request by not considering
2309 * what may be in the CONVERTING or WAITING queues... As of this
2310 * writing, the EXPEDITE flag can be used only with new requests for NL
2311 * mode locks. This flag is not valid for conversion requests.
2313 * A shortcut. Earlier checks return an error if EXPEDITE is used in a
2314 * conversion or used with a non-NL requested mode. We also know an
2315 * EXPEDITE request is always granted immediately, so now must always
2316 * be 1. The full condition to grant an expedite request: (now &&
2317 * !conv && lkb->rqmode == DLM_LOCK_NL && (flags & EXPEDITE)) can
2318 * therefore be shortened to just checking the flag.
2321 if (lkb->lkb_exflags & DLM_LKF_EXPEDITE)
2325 * A shortcut. Without this, !queue_conflict(grantqueue, lkb) would be
2326 * added to the remaining conditions.
2329 if (queue_conflict(&r->res_grantqueue, lkb))
2333 * 6-3: By default, a conversion request is immediately granted if the
2334 * requested mode is compatible with the modes of all other granted
2338 if (queue_conflict(&r->res_convertqueue, lkb))
2342 * The RECOVER_GRANT flag means dlm_recover_grant() is granting
2343 * locks for a recovered rsb, on which lkb's have been rebuilt.
2344 * The lkb's may have been rebuilt on the queues in a different
2345 * order than they were in on the previous master. So, granting
2346 * queued conversions in order after recovery doesn't make sense
2347 * since the order hasn't been preserved anyway. The new order
2348 * could also have created a new "in place" conversion deadlock.
2349 * (e.g. old, failed master held granted EX, with PR->EX, NL->EX.
2350 * After recovery, there would be no granted locks, and possibly
2351 * NL->EX, PR->EX, an in-place conversion deadlock.) So, after
2352 * recovery, grant conversions without considering order.
2355 if (conv && recover)
2359 * 6-5: But the default algorithm for deciding whether to grant or
2360 * queue conversion requests does not by itself guarantee that such
2361 * requests are serviced on a "first come first serve" basis. This, in
2362 * turn, can lead to a phenomenon known as "indefinate postponement".
2364 * 6-7: This issue is dealt with by using the optional QUECVT flag with
2365 * the system service employed to request a lock conversion. This flag
2366 * forces certain conversion requests to be queued, even if they are
2367 * compatible with the granted modes of other locks on the same
2368 * resource. Thus, the use of this flag results in conversion requests
2369 * being ordered on a "first come first servce" basis.
2371 * DCT: This condition is all about new conversions being able to occur
2372 * "in place" while the lock remains on the granted queue (assuming
2373 * nothing else conflicts.) IOW if QUECVT isn't set, a conversion
2374 * doesn't _have_ to go onto the convert queue where it's processed in
2375 * order. The "now" variable is necessary to distinguish converts
2376 * being received and processed for the first time now, because once a
2377 * convert is moved to the conversion queue the condition below applies
2378 * requiring fifo granting.
2381 if (now && conv && !(lkb->lkb_exflags & DLM_LKF_QUECVT))
2385 * Even if the convert is compat with all granted locks,
2386 * QUECVT forces it behind other locks on the convert queue.
2389 if (now && conv && (lkb->lkb_exflags & DLM_LKF_QUECVT)) {
2390 if (list_empty(&r->res_convertqueue))
2397 * The NOORDER flag is set to avoid the standard vms rules on grant
2401 if (lkb->lkb_exflags & DLM_LKF_NOORDER)
2405 * 6-3: Once in that queue [CONVERTING], a conversion request cannot be
2406 * granted until all other conversion requests ahead of it are granted
2410 if (!now && conv && first_in_list(lkb, &r->res_convertqueue))
2414 * 6-4: By default, a new request is immediately granted only if all
2415 * three of the following conditions are satisfied when the request is
2417 * - The queue of ungranted conversion requests for the resource is
2419 * - The queue of ungranted new requests for the resource is empty.
2420 * - The mode of the new request is compatible with the most
2421 * restrictive mode of all granted locks on the resource.
2424 if (now && !conv && list_empty(&r->res_convertqueue) &&
2425 list_empty(&r->res_waitqueue))
2429 * 6-4: Once a lock request is in the queue of ungranted new requests,
2430 * it cannot be granted until the queue of ungranted conversion
2431 * requests is empty, all ungranted new requests ahead of it are
2432 * granted and/or canceled, and it is compatible with the granted mode
2433 * of the most restrictive lock granted on the resource.
2436 if (!now && !conv && list_empty(&r->res_convertqueue) &&
2437 first_in_list(lkb, &r->res_waitqueue))
2443 static int can_be_granted(struct dlm_rsb *r, struct dlm_lkb *lkb, int now,
2444 int recover, int *err)
2447 int8_t alt = 0, rqmode = lkb->lkb_rqmode;
2448 int8_t is_convert = (lkb->lkb_grmode != DLM_LOCK_IV);
2453 rv = _can_be_granted(r, lkb, now, recover);
2458 * The CONVDEADLK flag is non-standard and tells the dlm to resolve
2459 * conversion deadlocks by demoting grmode to NL, otherwise the dlm
2460 * cancels one of the locks.
2463 if (is_convert && can_be_queued(lkb) &&
2464 conversion_deadlock_detect(r, lkb)) {
2465 if (lkb->lkb_exflags & DLM_LKF_CONVDEADLK) {
2466 lkb->lkb_grmode = DLM_LOCK_NL;
2467 lkb->lkb_sbflags |= DLM_SBF_DEMOTED;
2471 log_print("can_be_granted deadlock %x now %d",
2479 * The ALTPR and ALTCW flags are non-standard and tell the dlm to try
2480 * to grant a request in a mode other than the normal rqmode. It's a
2481 * simple way to provide a big optimization to applications that can
2485 if (rqmode != DLM_LOCK_PR && (lkb->lkb_exflags & DLM_LKF_ALTPR))
2487 else if (rqmode != DLM_LOCK_CW && (lkb->lkb_exflags & DLM_LKF_ALTCW))
2491 lkb->lkb_rqmode = alt;
2492 rv = _can_be_granted(r, lkb, now, 0);
2494 lkb->lkb_sbflags |= DLM_SBF_ALTMODE;
2496 lkb->lkb_rqmode = rqmode;
2502 /* Returns the highest requested mode of all blocked conversions; sets
2503 cw if there's a blocked conversion to DLM_LOCK_CW. */
2505 static int grant_pending_convert(struct dlm_rsb *r, int high, int *cw,
2506 unsigned int *count)
2508 struct dlm_lkb *lkb, *s;
2509 int recover = rsb_flag(r, RSB_RECOVER_GRANT);
2510 int hi, demoted, quit, grant_restart, demote_restart;
2519 list_for_each_entry_safe(lkb, s, &r->res_convertqueue, lkb_statequeue) {
2520 demoted = is_demoted(lkb);
2523 if (can_be_granted(r, lkb, 0, recover, &deadlk)) {
2524 grant_lock_pending(r, lkb);
2531 if (!demoted && is_demoted(lkb)) {
2532 log_print("WARN: pending demoted %x node %d %s",
2533 lkb->lkb_id, lkb->lkb_nodeid, r->res_name);
2540 * If DLM_LKB_NODLKWT flag is set and conversion
2541 * deadlock is detected, we request blocking AST and
2542 * down (or cancel) conversion.
2544 if (lkb->lkb_exflags & DLM_LKF_NODLCKWT) {
2545 if (lkb->lkb_highbast < lkb->lkb_rqmode) {
2546 queue_bast(r, lkb, lkb->lkb_rqmode);
2547 lkb->lkb_highbast = lkb->lkb_rqmode;
2550 log_print("WARN: pending deadlock %x node %d %s",
2551 lkb->lkb_id, lkb->lkb_nodeid,
2558 hi = max_t(int, lkb->lkb_rqmode, hi);
2560 if (cw && lkb->lkb_rqmode == DLM_LOCK_CW)
2566 if (demote_restart && !quit) {
2571 return max_t(int, high, hi);
2574 static int grant_pending_wait(struct dlm_rsb *r, int high, int *cw,
2575 unsigned int *count)
2577 struct dlm_lkb *lkb, *s;
2579 list_for_each_entry_safe(lkb, s, &r->res_waitqueue, lkb_statequeue) {
2580 if (can_be_granted(r, lkb, 0, 0, NULL)) {
2581 grant_lock_pending(r, lkb);
2585 high = max_t(int, lkb->lkb_rqmode, high);
2586 if (lkb->lkb_rqmode == DLM_LOCK_CW)
2594 /* cw of 1 means there's a lock with a rqmode of DLM_LOCK_CW that's blocked
2595 on either the convert or waiting queue.
2596 high is the largest rqmode of all locks blocked on the convert or
2599 static int lock_requires_bast(struct dlm_lkb *gr, int high, int cw)
2601 if (gr->lkb_grmode == DLM_LOCK_PR && cw) {
2602 if (gr->lkb_highbast < DLM_LOCK_EX)
2607 if (gr->lkb_highbast < high &&
2608 !__dlm_compat_matrix[gr->lkb_grmode+1][high+1])
2613 static void grant_pending_locks(struct dlm_rsb *r, unsigned int *count)
2615 struct dlm_lkb *lkb, *s;
2616 int high = DLM_LOCK_IV;
2619 if (!is_master(r)) {
2620 log_print("grant_pending_locks r nodeid %d", r->res_nodeid);
2625 high = grant_pending_convert(r, high, &cw, count);
2626 high = grant_pending_wait(r, high, &cw, count);
2628 if (high == DLM_LOCK_IV)
2632 * If there are locks left on the wait/convert queue then send blocking
2633 * ASTs to granted locks based on the largest requested mode (high)
2637 list_for_each_entry_safe(lkb, s, &r->res_grantqueue, lkb_statequeue) {
2638 if (lkb->lkb_bastfn && lock_requires_bast(lkb, high, cw)) {
2639 if (cw && high == DLM_LOCK_PR &&
2640 lkb->lkb_grmode == DLM_LOCK_PR)
2641 queue_bast(r, lkb, DLM_LOCK_CW);
2643 queue_bast(r, lkb, high);
2644 lkb->lkb_highbast = high;
2649 static int modes_require_bast(struct dlm_lkb *gr, struct dlm_lkb *rq)
2651 if ((gr->lkb_grmode == DLM_LOCK_PR && rq->lkb_rqmode == DLM_LOCK_CW) ||
2652 (gr->lkb_grmode == DLM_LOCK_CW && rq->lkb_rqmode == DLM_LOCK_PR)) {
2653 if (gr->lkb_highbast < DLM_LOCK_EX)
2658 if (gr->lkb_highbast < rq->lkb_rqmode && !modes_compat(gr, rq))
2663 static void send_bast_queue(struct dlm_rsb *r, struct list_head *head,
2664 struct dlm_lkb *lkb)
2668 list_for_each_entry(gr, head, lkb_statequeue) {
2669 /* skip self when sending basts to convertqueue */
2672 if (gr->lkb_bastfn && modes_require_bast(gr, lkb)) {
2673 queue_bast(r, gr, lkb->lkb_rqmode);
2674 gr->lkb_highbast = lkb->lkb_rqmode;
2679 static void send_blocking_asts(struct dlm_rsb *r, struct dlm_lkb *lkb)
2681 send_bast_queue(r, &r->res_grantqueue, lkb);
2684 static void send_blocking_asts_all(struct dlm_rsb *r, struct dlm_lkb *lkb)
2686 send_bast_queue(r, &r->res_grantqueue, lkb);
2687 send_bast_queue(r, &r->res_convertqueue, lkb);
2690 /* set_master(r, lkb) -- set the master nodeid of a resource
2692 The purpose of this function is to set the nodeid field in the given
2693 lkb using the nodeid field in the given rsb. If the rsb's nodeid is
2694 known, it can just be copied to the lkb and the function will return
2695 0. If the rsb's nodeid is _not_ known, it needs to be looked up
2696 before it can be copied to the lkb.
2698 When the rsb nodeid is being looked up remotely, the initial lkb
2699 causing the lookup is kept on the ls_waiters list waiting for the
2700 lookup reply. Other lkb's waiting for the same rsb lookup are kept
2701 on the rsb's res_lookup list until the master is verified.
2704 0: nodeid is set in rsb/lkb and the caller should go ahead and use it
2705 1: the rsb master is not available and the lkb has been placed on
2709 static int set_master(struct dlm_rsb *r, struct dlm_lkb *lkb)
2711 int our_nodeid = dlm_our_nodeid();
2713 if (rsb_flag(r, RSB_MASTER_UNCERTAIN)) {
2714 rsb_clear_flag(r, RSB_MASTER_UNCERTAIN);
2715 r->res_first_lkid = lkb->lkb_id;
2716 lkb->lkb_nodeid = r->res_nodeid;
2720 if (r->res_first_lkid && r->res_first_lkid != lkb->lkb_id) {
2721 list_add_tail(&lkb->lkb_rsb_lookup, &r->res_lookup);
2725 if (r->res_master_nodeid == our_nodeid) {
2726 lkb->lkb_nodeid = 0;
2730 if (r->res_master_nodeid) {
2731 lkb->lkb_nodeid = r->res_master_nodeid;
2735 if (dlm_dir_nodeid(r) == our_nodeid) {
2736 /* This is a somewhat unusual case; find_rsb will usually
2737 have set res_master_nodeid when dir nodeid is local, but
2738 there are cases where we become the dir node after we've
2739 past find_rsb and go through _request_lock again.
2740 confirm_master() or process_lookup_list() needs to be
2741 called after this. */
2742 log_debug(r->res_ls, "set_master %x self master %d dir %d %s",
2743 lkb->lkb_id, r->res_master_nodeid, r->res_dir_nodeid,
2745 r->res_master_nodeid = our_nodeid;
2747 lkb->lkb_nodeid = 0;
2751 wait_pending_remove(r);
2753 r->res_first_lkid = lkb->lkb_id;
2754 send_lookup(r, lkb);
2758 static void process_lookup_list(struct dlm_rsb *r)
2760 struct dlm_lkb *lkb, *safe;
2762 list_for_each_entry_safe(lkb, safe, &r->res_lookup, lkb_rsb_lookup) {
2763 list_del_init(&lkb->lkb_rsb_lookup);
2764 _request_lock(r, lkb);
2769 /* confirm_master -- confirm (or deny) an rsb's master nodeid */
2771 static void confirm_master(struct dlm_rsb *r, int error)
2773 struct dlm_lkb *lkb;
2775 if (!r->res_first_lkid)
2781 r->res_first_lkid = 0;
2782 process_lookup_list(r);
2788 /* the remote request failed and won't be retried (it was
2789 a NOQUEUE, or has been canceled/unlocked); make a waiting
2790 lkb the first_lkid */
2792 r->res_first_lkid = 0;
2794 if (!list_empty(&r->res_lookup)) {
2795 lkb = list_entry(r->res_lookup.next, struct dlm_lkb,
2797 list_del_init(&lkb->lkb_rsb_lookup);
2798 r->res_first_lkid = lkb->lkb_id;
2799 _request_lock(r, lkb);
2804 log_error(r->res_ls, "confirm_master unknown error %d", error);
2808 static int set_lock_args(int mode, struct dlm_lksb *lksb, uint32_t flags,
2809 int namelen, unsigned long timeout_cs,
2810 void (*ast) (void *astparam),
2812 void (*bast) (void *astparam, int mode),
2813 struct dlm_args *args)
2817 /* check for invalid arg usage */
2819 if (mode < 0 || mode > DLM_LOCK_EX)
2822 if (!(flags & DLM_LKF_CONVERT) && (namelen > DLM_RESNAME_MAXLEN))
2825 if (flags & DLM_LKF_CANCEL)
2828 if (flags & DLM_LKF_QUECVT && !(flags & DLM_LKF_CONVERT))
2831 if (flags & DLM_LKF_CONVDEADLK && !(flags & DLM_LKF_CONVERT))
2834 if (flags & DLM_LKF_CONVDEADLK && flags & DLM_LKF_NOQUEUE)
2837 if (flags & DLM_LKF_EXPEDITE && flags & DLM_LKF_CONVERT)
2840 if (flags & DLM_LKF_EXPEDITE && flags & DLM_LKF_QUECVT)
2843 if (flags & DLM_LKF_EXPEDITE && flags & DLM_LKF_NOQUEUE)
2846 if (flags & DLM_LKF_EXPEDITE && mode != DLM_LOCK_NL)
2852 if (flags & DLM_LKF_VALBLK && !lksb->sb_lvbptr)
2855 if (flags & DLM_LKF_CONVERT && !lksb->sb_lkid)
2858 /* these args will be copied to the lkb in validate_lock_args,
2859 it cannot be done now because when converting locks, fields in
2860 an active lkb cannot be modified before locking the rsb */
2862 args->flags = flags;
2864 args->astparam = astparam;
2865 args->bastfn = bast;
2866 args->timeout = timeout_cs;
2874 static int set_unlock_args(uint32_t flags, void *astarg, struct dlm_args *args)
2876 if (flags & ~(DLM_LKF_CANCEL | DLM_LKF_VALBLK | DLM_LKF_IVVALBLK |
2877 DLM_LKF_FORCEUNLOCK))
2880 if (flags & DLM_LKF_CANCEL && flags & DLM_LKF_FORCEUNLOCK)
2883 args->flags = flags;
2884 args->astparam = astarg;
2888 static int validate_lock_args(struct dlm_ls *ls, struct dlm_lkb *lkb,
2889 struct dlm_args *args)
2893 if (args->flags & DLM_LKF_CONVERT) {
2894 if (lkb->lkb_flags & DLM_IFL_MSTCPY)
2897 if (args->flags & DLM_LKF_QUECVT &&
2898 !__quecvt_compat_matrix[lkb->lkb_grmode+1][args->mode+1])
2902 if (lkb->lkb_status != DLM_LKSTS_GRANTED)
2905 if (lkb->lkb_wait_type)
2908 if (is_overlap(lkb))
2912 lkb->lkb_exflags = args->flags;
2913 lkb->lkb_sbflags = 0;
2914 lkb->lkb_astfn = args->astfn;
2915 lkb->lkb_astparam = args->astparam;
2916 lkb->lkb_bastfn = args->bastfn;
2917 lkb->lkb_rqmode = args->mode;
2918 lkb->lkb_lksb = args->lksb;
2919 lkb->lkb_lvbptr = args->lksb->sb_lvbptr;
2920 lkb->lkb_ownpid = (int) current->pid;
2921 lkb->lkb_timeout_cs = args->timeout;
2925 log_debug(ls, "validate_lock_args %d %x %x %x %d %d %s",
2926 rv, lkb->lkb_id, lkb->lkb_flags, args->flags,
2927 lkb->lkb_status, lkb->lkb_wait_type,
2928 lkb->lkb_resource->res_name);
2932 /* when dlm_unlock() sees -EBUSY with CANCEL/FORCEUNLOCK it returns 0
2935 /* note: it's valid for lkb_nodeid/res_nodeid to be -1 when we get here
2936 because there may be a lookup in progress and it's valid to do
2937 cancel/unlockf on it */
2939 static int validate_unlock_args(struct dlm_lkb *lkb, struct dlm_args *args)
2941 struct dlm_ls *ls = lkb->lkb_resource->res_ls;
2944 if (lkb->lkb_flags & DLM_IFL_MSTCPY) {
2945 log_error(ls, "unlock on MSTCPY %x", lkb->lkb_id);
2950 /* an lkb may still exist even though the lock is EOL'ed due to a
2951 cancel, unlock or failed noqueue request; an app can't use these
2952 locks; return same error as if the lkid had not been found at all */
2954 if (lkb->lkb_flags & DLM_IFL_ENDOFLIFE) {
2955 log_debug(ls, "unlock on ENDOFLIFE %x", lkb->lkb_id);
2960 /* an lkb may be waiting for an rsb lookup to complete where the
2961 lookup was initiated by another lock */
2963 if (!list_empty(&lkb->lkb_rsb_lookup)) {
2964 if (args->flags & (DLM_LKF_CANCEL | DLM_LKF_FORCEUNLOCK)) {
2965 log_debug(ls, "unlock on rsb_lookup %x", lkb->lkb_id);
2966 list_del_init(&lkb->lkb_rsb_lookup);
2967 queue_cast(lkb->lkb_resource, lkb,
2968 args->flags & DLM_LKF_CANCEL ?
2969 -DLM_ECANCEL : -DLM_EUNLOCK);
2970 unhold_lkb(lkb); /* undoes create_lkb() */
2972 /* caller changes -EBUSY to 0 for CANCEL and FORCEUNLOCK */
2977 /* cancel not allowed with another cancel/unlock in progress */
2979 if (args->flags & DLM_LKF_CANCEL) {
2980 if (lkb->lkb_exflags & DLM_LKF_CANCEL)
2983 if (is_overlap(lkb))
2986 /* don't let scand try to do a cancel */
2989 if (lkb->lkb_flags & DLM_IFL_RESEND) {
2990 lkb->lkb_flags |= DLM_IFL_OVERLAP_CANCEL;
2995 /* there's nothing to cancel */
2996 if (lkb->lkb_status == DLM_LKSTS_GRANTED &&
2997 !lkb->lkb_wait_type) {
3002 switch (lkb->lkb_wait_type) {
3003 case DLM_MSG_LOOKUP:
3004 case DLM_MSG_REQUEST:
3005 lkb->lkb_flags |= DLM_IFL_OVERLAP_CANCEL;
3008 case DLM_MSG_UNLOCK:
3009 case DLM_MSG_CANCEL:
3012 /* add_to_waiters() will set OVERLAP_CANCEL */
3016 /* do we need to allow a force-unlock if there's a normal unlock
3017 already in progress? in what conditions could the normal unlock
3018 fail such that we'd want to send a force-unlock to be sure? */
3020 if (args->flags & DLM_LKF_FORCEUNLOCK) {
3021 if (lkb->lkb_exflags & DLM_LKF_FORCEUNLOCK)
3024 if (is_overlap_unlock(lkb))
3027 /* don't let scand try to do a cancel */
3030 if (lkb->lkb_flags & DLM_IFL_RESEND) {
3031 lkb->lkb_flags |= DLM_IFL_OVERLAP_UNLOCK;
3036 switch (lkb->lkb_wait_type) {
3037 case DLM_MSG_LOOKUP:
3038 case DLM_MSG_REQUEST:
3039 lkb->lkb_flags |= DLM_IFL_OVERLAP_UNLOCK;
3042 case DLM_MSG_UNLOCK:
3045 /* add_to_waiters() will set OVERLAP_UNLOCK */
3049 /* normal unlock not allowed if there's any op in progress */
3051 if (lkb->lkb_wait_type || lkb->lkb_wait_count)
3055 /* an overlapping op shouldn't blow away exflags from other op */
3056 lkb->lkb_exflags |= args->flags;
3057 lkb->lkb_sbflags = 0;
3058 lkb->lkb_astparam = args->astparam;
3062 log_debug(ls, "validate_unlock_args %d %x %x %x %x %d %s", rv,
3063 lkb->lkb_id, lkb->lkb_flags, lkb->lkb_exflags,
3064 args->flags, lkb->lkb_wait_type,
3065 lkb->lkb_resource->res_name);
3070 * Four stage 4 varieties:
3071 * do_request(), do_convert(), do_unlock(), do_cancel()
3072 * These are called on the master node for the given lock and
3073 * from the central locking logic.
3076 static int do_request(struct dlm_rsb *r, struct dlm_lkb *lkb)
3080 if (can_be_granted(r, lkb, 1, 0, NULL)) {
3082 queue_cast(r, lkb, 0);
3086 if (can_be_queued(lkb)) {
3087 error = -EINPROGRESS;
3088 add_lkb(r, lkb, DLM_LKSTS_WAITING);
3094 queue_cast(r, lkb, -EAGAIN);
3099 static void do_request_effects(struct dlm_rsb *r, struct dlm_lkb *lkb,
3104 if (force_blocking_asts(lkb))
3105 send_blocking_asts_all(r, lkb);
3108 send_blocking_asts(r, lkb);
3113 static int do_convert(struct dlm_rsb *r, struct dlm_lkb *lkb)
3118 /* changing an existing lock may allow others to be granted */
3120 if (can_be_granted(r, lkb, 1, 0, &deadlk)) {
3122 queue_cast(r, lkb, 0);
3126 /* can_be_granted() detected that this lock would block in a conversion
3127 deadlock, so we leave it on the granted queue and return EDEADLK in
3128 the ast for the convert. */
3130 if (deadlk && !(lkb->lkb_exflags & DLM_LKF_NODLCKWT)) {
3131 /* it's left on the granted queue */
3132 revert_lock(r, lkb);
3133 queue_cast(r, lkb, -EDEADLK);
3138 /* is_demoted() means the can_be_granted() above set the grmode
3139 to NL, and left us on the granted queue. This auto-demotion
3140 (due to CONVDEADLK) might mean other locks, and/or this lock, are
3141 now grantable. We have to try to grant other converting locks
3142 before we try again to grant this one. */
3144 if (is_demoted(lkb)) {
3145 grant_pending_convert(r, DLM_LOCK_IV, NULL, NULL);
3146 if (_can_be_granted(r, lkb, 1, 0)) {
3148 queue_cast(r, lkb, 0);
3151 /* else fall through and move to convert queue */
3154 if (can_be_queued(lkb)) {
3155 error = -EINPROGRESS;
3157 add_lkb(r, lkb, DLM_LKSTS_CONVERT);
3163 queue_cast(r, lkb, -EAGAIN);
3168 static void do_convert_effects(struct dlm_rsb *r, struct dlm_lkb *lkb,
3173 grant_pending_locks(r, NULL);
3174 /* grant_pending_locks also sends basts */
3177 if (force_blocking_asts(lkb))
3178 send_blocking_asts_all(r, lkb);
3181 send_blocking_asts(r, lkb);
3186 static int do_unlock(struct dlm_rsb *r, struct dlm_lkb *lkb)
3188 remove_lock(r, lkb);
3189 queue_cast(r, lkb, -DLM_EUNLOCK);
3190 return -DLM_EUNLOCK;
3193 static void do_unlock_effects(struct dlm_rsb *r, struct dlm_lkb *lkb,
3196 grant_pending_locks(r, NULL);
3199 /* returns: 0 did nothing, -DLM_ECANCEL canceled lock */
3201 static int do_cancel(struct dlm_rsb *r, struct dlm_lkb *lkb)
3205 error = revert_lock(r, lkb);
3207 queue_cast(r, lkb, -DLM_ECANCEL);
3208 return -DLM_ECANCEL;
3213 static void do_cancel_effects(struct dlm_rsb *r, struct dlm_lkb *lkb,
3217 grant_pending_locks(r, NULL);
3221 * Four stage 3 varieties:
3222 * _request_lock(), _convert_lock(), _unlock_lock(), _cancel_lock()
3225 /* add a new lkb to a possibly new rsb, called by requesting process */
3227 static int _request_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
3231 /* set_master: sets lkb nodeid from r */
3233 error = set_master(r, lkb);
3242 /* receive_request() calls do_request() on remote node */
3243 error = send_request(r, lkb);
3245 error = do_request(r, lkb);
3246 /* for remote locks the request_reply is sent
3247 between do_request and do_request_effects */
3248 do_request_effects(r, lkb, error);
3254 /* change some property of an existing lkb, e.g. mode */
3256 static int _convert_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
3261 /* receive_convert() calls do_convert() on remote node */
3262 error = send_convert(r, lkb);
3264 error = do_convert(r, lkb);
3265 /* for remote locks the convert_reply is sent
3266 between do_convert and do_convert_effects */
3267 do_convert_effects(r, lkb, error);
3273 /* remove an existing lkb from the granted queue */
3275 static int _unlock_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
3280 /* receive_unlock() calls do_unlock() on remote node */
3281 error = send_unlock(r, lkb);
3283 error = do_unlock(r, lkb);
3284 /* for remote locks the unlock_reply is sent
3285 between do_unlock and do_unlock_effects */
3286 do_unlock_effects(r, lkb, error);
3292 /* remove an existing lkb from the convert or wait queue */
3294 static int _cancel_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
3299 /* receive_cancel() calls do_cancel() on remote node */
3300 error = send_cancel(r, lkb);
3302 error = do_cancel(r, lkb);
3303 /* for remote locks the cancel_reply is sent
3304 between do_cancel and do_cancel_effects */
3305 do_cancel_effects(r, lkb, error);
3312 * Four stage 2 varieties:
3313 * request_lock(), convert_lock(), unlock_lock(), cancel_lock()
3316 static int request_lock(struct dlm_ls *ls, struct dlm_lkb *lkb, char *name,
3317 int len, struct dlm_args *args)
3322 error = validate_lock_args(ls, lkb, args);
3326 error = find_rsb(ls, name, len, 0, R_REQUEST, &r);
3333 lkb->lkb_lksb->sb_lkid = lkb->lkb_id;
3335 error = _request_lock(r, lkb);
3342 static int convert_lock(struct dlm_ls *ls, struct dlm_lkb *lkb,
3343 struct dlm_args *args)
3348 r = lkb->lkb_resource;
3353 error = validate_lock_args(ls, lkb, args);
3357 error = _convert_lock(r, lkb);
3364 static int unlock_lock(struct dlm_ls *ls, struct dlm_lkb *lkb,
3365 struct dlm_args *args)
3370 r = lkb->lkb_resource;
3375 error = validate_unlock_args(lkb, args);
3379 error = _unlock_lock(r, lkb);
3386 static int cancel_lock(struct dlm_ls *ls, struct dlm_lkb *lkb,
3387 struct dlm_args *args)
3392 r = lkb->lkb_resource;
3397 error = validate_unlock_args(lkb, args);
3401 error = _cancel_lock(r, lkb);
3409 * Two stage 1 varieties: dlm_lock() and dlm_unlock()
3412 int dlm_lock(dlm_lockspace_t *lockspace,
3414 struct dlm_lksb *lksb,
3417 unsigned int namelen,
3418 uint32_t parent_lkid,
3419 void (*ast) (void *astarg),
3421 void (*bast) (void *astarg, int mode))
3424 struct dlm_lkb *lkb;
3425 struct dlm_args args;
3426 int error, convert = flags & DLM_LKF_CONVERT;
3428 ls = dlm_find_lockspace_local(lockspace);
3432 dlm_lock_recovery(ls);
3435 error = find_lkb(ls, lksb->sb_lkid, &lkb);
3437 error = create_lkb(ls, &lkb);
3442 error = set_lock_args(mode, lksb, flags, namelen, 0, ast,
3443 astarg, bast, &args);
3448 error = convert_lock(ls, lkb, &args);
3450 error = request_lock(ls, lkb, name, namelen, &args);
3452 if (error == -EINPROGRESS)
3455 if (convert || error)
3457 if (error == -EAGAIN || error == -EDEADLK)
3460 dlm_unlock_recovery(ls);
3461 dlm_put_lockspace(ls);
3465 int dlm_unlock(dlm_lockspace_t *lockspace,
3468 struct dlm_lksb *lksb,
3472 struct dlm_lkb *lkb;
3473 struct dlm_args args;
3476 ls = dlm_find_lockspace_local(lockspace);
3480 dlm_lock_recovery(ls);
3482 error = find_lkb(ls, lkid, &lkb);
3486 error = set_unlock_args(flags, astarg, &args);
3490 if (flags & DLM_LKF_CANCEL)
3491 error = cancel_lock(ls, lkb, &args);
3493 error = unlock_lock(ls, lkb, &args);
3495 if (error == -DLM_EUNLOCK || error == -DLM_ECANCEL)
3497 if (error == -EBUSY && (flags & (DLM_LKF_CANCEL | DLM_LKF_FORCEUNLOCK)))
3502 dlm_unlock_recovery(ls);
3503 dlm_put_lockspace(ls);
3508 * send/receive routines for remote operations and replies
3512 * send_request receive_request
3513 * send_convert receive_convert
3514 * send_unlock receive_unlock
3515 * send_cancel receive_cancel
3516 * send_grant receive_grant
3517 * send_bast receive_bast
3518 * send_lookup receive_lookup
3519 * send_remove receive_remove
3522 * receive_request_reply send_request_reply
3523 * receive_convert_reply send_convert_reply
3524 * receive_unlock_reply send_unlock_reply
3525 * receive_cancel_reply send_cancel_reply
3526 * receive_lookup_reply send_lookup_reply
3529 static int _create_message(struct dlm_ls *ls, int mb_len,
3530 int to_nodeid, int mstype,
3531 struct dlm_message **ms_ret,
3532 struct dlm_mhandle **mh_ret)
3534 struct dlm_message *ms;
3535 struct dlm_mhandle *mh;
3538 /* get_buffer gives us a message handle (mh) that we need to
3539 pass into lowcomms_commit and a message buffer (mb) that we
3540 write our data into */
3542 mh = dlm_lowcomms_get_buffer(to_nodeid, mb_len, GFP_NOFS, &mb);
3546 memset(mb, 0, mb_len);
3548 ms = (struct dlm_message *) mb;
3550 ms->m_header.h_version = (DLM_HEADER_MAJOR | DLM_HEADER_MINOR);
3551 ms->m_header.h_lockspace = ls->ls_global_id;
3552 ms->m_header.h_nodeid = dlm_our_nodeid();
3553 ms->m_header.h_length = mb_len;
3554 ms->m_header.h_cmd = DLM_MSG;
3556 ms->m_type = mstype;
3563 static int create_message(struct dlm_rsb *r, struct dlm_lkb *lkb,
3564 int to_nodeid, int mstype,
3565 struct dlm_message **ms_ret,
3566 struct dlm_mhandle **mh_ret)
3568 int mb_len = sizeof(struct dlm_message);
3571 case DLM_MSG_REQUEST:
3572 case DLM_MSG_LOOKUP:
3573 case DLM_MSG_REMOVE:
3574 mb_len += r->res_length;
3576 case DLM_MSG_CONVERT:
3577 case DLM_MSG_UNLOCK:
3578 case DLM_MSG_REQUEST_REPLY:
3579 case DLM_MSG_CONVERT_REPLY:
3581 if (lkb && lkb->lkb_lvbptr)
3582 mb_len += r->res_ls->ls_lvblen;
3586 return _create_message(r->res_ls, mb_len, to_nodeid, mstype,
3590 /* further lowcomms enhancements or alternate implementations may make
3591 the return value from this function useful at some point */
3593 static int send_message(struct dlm_mhandle *mh, struct dlm_message *ms)
3595 dlm_message_out(ms);
3596 dlm_lowcomms_commit_buffer(mh);
3600 static void send_args(struct dlm_rsb *r, struct dlm_lkb *lkb,
3601 struct dlm_message *ms)
3603 ms->m_nodeid = lkb->lkb_nodeid;
3604 ms->m_pid = lkb->lkb_ownpid;
3605 ms->m_lkid = lkb->lkb_id;
3606 ms->m_remid = lkb->lkb_remid;
3607 ms->m_exflags = lkb->lkb_exflags;
3608 ms->m_sbflags = lkb->lkb_sbflags;
3609 ms->m_flags = lkb->lkb_flags;
3610 ms->m_lvbseq = lkb->lkb_lvbseq;
3611 ms->m_status = lkb->lkb_status;
3612 ms->m_grmode = lkb->lkb_grmode;
3613 ms->m_rqmode = lkb->lkb_rqmode;
3614 ms->m_hash = r->res_hash;
3616 /* m_result and m_bastmode are set from function args,
3617 not from lkb fields */
3619 if (lkb->lkb_bastfn)
3620 ms->m_asts |= DLM_CB_BAST;
3622 ms->m_asts |= DLM_CB_CAST;
3624 /* compare with switch in create_message; send_remove() doesn't
3627 switch (ms->m_type) {
3628 case DLM_MSG_REQUEST:
3629 case DLM_MSG_LOOKUP:
3630 memcpy(ms->m_extra, r->res_name, r->res_length);
3632 case DLM_MSG_CONVERT:
3633 case DLM_MSG_UNLOCK:
3634 case DLM_MSG_REQUEST_REPLY:
3635 case DLM_MSG_CONVERT_REPLY:
3637 if (!lkb->lkb_lvbptr)
3639 memcpy(ms->m_extra, lkb->lkb_lvbptr, r->res_ls->ls_lvblen);
3644 static int send_common(struct dlm_rsb *r, struct dlm_lkb *lkb, int mstype)
3646 struct dlm_message *ms;
3647 struct dlm_mhandle *mh;
3648 int to_nodeid, error;
3650 to_nodeid = r->res_nodeid;
3652 error = add_to_waiters(lkb, mstype, to_nodeid);
3656 error = create_message(r, lkb, to_nodeid, mstype, &ms, &mh);
3660 send_args(r, lkb, ms);
3662 error = send_message(mh, ms);
3668 remove_from_waiters(lkb, msg_reply_type(mstype));
3672 static int send_request(struct dlm_rsb *r, struct dlm_lkb *lkb)
3674 return send_common(r, lkb, DLM_MSG_REQUEST);
3677 static int send_convert(struct dlm_rsb *r, struct dlm_lkb *lkb)
3681 error = send_common(r, lkb, DLM_MSG_CONVERT);
3683 /* down conversions go without a reply from the master */
3684 if (!error && down_conversion(lkb)) {
3685 remove_from_waiters(lkb, DLM_MSG_CONVERT_REPLY);
3686 r->res_ls->ls_stub_ms.m_flags = DLM_IFL_STUB_MS;
3687 r->res_ls->ls_stub_ms.m_type = DLM_MSG_CONVERT_REPLY;
3688 r->res_ls->ls_stub_ms.m_result = 0;
3689 __receive_convert_reply(r, lkb, &r->res_ls->ls_stub_ms);
3695 /* FIXME: if this lkb is the only lock we hold on the rsb, then set
3696 MASTER_UNCERTAIN to force the next request on the rsb to confirm
3697 that the master is still correct. */
3699 static int send_unlock(struct dlm_rsb *r, struct dlm_lkb *lkb)
3701 return send_common(r, lkb, DLM_MSG_UNLOCK);
3704 static int send_cancel(struct dlm_rsb *r, struct dlm_lkb *lkb)
3706 return send_common(r, lkb, DLM_MSG_CANCEL);
3709 static int send_grant(struct dlm_rsb *r, struct dlm_lkb *lkb)
3711 struct dlm_message *ms;
3712 struct dlm_mhandle *mh;
3713 int to_nodeid, error;
3715 to_nodeid = lkb->lkb_nodeid;
3717 error = create_message(r, lkb, to_nodeid, DLM_MSG_GRANT, &ms, &mh);
3721 send_args(r, lkb, ms);
3725 error = send_message(mh, ms);
3730 static int send_bast(struct dlm_rsb *r, struct dlm_lkb *lkb, int mode)
3732 struct dlm_message *ms;
3733 struct dlm_mhandle *mh;
3734 int to_nodeid, error;
3736 to_nodeid = lkb->lkb_nodeid;
3738 error = create_message(r, NULL, to_nodeid, DLM_MSG_BAST, &ms, &mh);
3742 send_args(r, lkb, ms);
3744 ms->m_bastmode = mode;
3746 error = send_message(mh, ms);
3751 static int send_lookup(struct dlm_rsb *r, struct dlm_lkb *lkb)
3753 struct dlm_message *ms;
3754 struct dlm_mhandle *mh;
3755 int to_nodeid, error;
3757 to_nodeid = dlm_dir_nodeid(r);
3759 error = add_to_waiters(lkb, DLM_MSG_LOOKUP, to_nodeid);
3763 error = create_message(r, NULL, to_nodeid, DLM_MSG_LOOKUP, &ms, &mh);
3767 send_args(r, lkb, ms);
3769 error = send_message(mh, ms);
3775 remove_from_waiters(lkb, DLM_MSG_LOOKUP_REPLY);
3779 static int send_remove(struct dlm_rsb *r)
3781 struct dlm_message *ms;
3782 struct dlm_mhandle *mh;
3783 int to_nodeid, error;
3785 to_nodeid = dlm_dir_nodeid(r);
3787 error = create_message(r, NULL, to_nodeid, DLM_MSG_REMOVE, &ms, &mh);
3791 memcpy(ms->m_extra, r->res_name, r->res_length);
3792 ms->m_hash = r->res_hash;
3794 error = send_message(mh, ms);
3799 static int send_common_reply(struct dlm_rsb *r, struct dlm_lkb *lkb,
3802 struct dlm_message *ms;
3803 struct dlm_mhandle *mh;
3804 int to_nodeid, error;
3806 to_nodeid = lkb->lkb_nodeid;
3808 error = create_message(r, lkb, to_nodeid, mstype, &ms, &mh);
3812 send_args(r, lkb, ms);
3816 error = send_message(mh, ms);
3821 static int send_request_reply(struct dlm_rsb *r, struct dlm_lkb *lkb, int rv)
3823 return send_common_reply(r, lkb, DLM_MSG_REQUEST_REPLY, rv);
3826 static int send_convert_reply(struct dlm_rsb *r, struct dlm_lkb *lkb, int rv)
3828 return send_common_reply(r, lkb, DLM_MSG_CONVERT_REPLY, rv);
3831 static int send_unlock_reply(struct dlm_rsb *r, struct dlm_lkb *lkb, int rv)
3833 return send_common_reply(r, lkb, DLM_MSG_UNLOCK_REPLY, rv);
3836 static int send_cancel_reply(struct dlm_rsb *r, struct dlm_lkb *lkb, int rv)
3838 return send_common_reply(r, lkb, DLM_MSG_CANCEL_REPLY, rv);
3841 static int send_lookup_reply(struct dlm_ls *ls, struct dlm_message *ms_in,
3842 int ret_nodeid, int rv)
3844 struct dlm_rsb *r = &ls->ls_stub_rsb;
3845 struct dlm_message *ms;
3846 struct dlm_mhandle *mh;
3847 int error, nodeid = ms_in->m_header.h_nodeid;
3849 error = create_message(r, NULL, nodeid, DLM_MSG_LOOKUP_REPLY, &ms, &mh);
3853 ms->m_lkid = ms_in->m_lkid;
3855 ms->m_nodeid = ret_nodeid;
3857 error = send_message(mh, ms);
3862 /* which args we save from a received message depends heavily on the type
3863 of message, unlike the send side where we can safely send everything about
3864 the lkb for any type of message */
3866 static void receive_flags(struct dlm_lkb *lkb, struct dlm_message *ms)
3868 lkb->lkb_exflags = ms->m_exflags;
3869 lkb->lkb_sbflags = ms->m_sbflags;
3870 lkb->lkb_flags = (lkb->lkb_flags & 0xFFFF0000) |
3871 (ms->m_flags & 0x0000FFFF);
3874 static void receive_flags_reply(struct dlm_lkb *lkb, struct dlm_message *ms)
3876 if (ms->m_flags == DLM_IFL_STUB_MS)
3879 lkb->lkb_sbflags = ms->m_sbflags;
3880 lkb->lkb_flags = (lkb->lkb_flags & 0xFFFF0000) |
3881 (ms->m_flags & 0x0000FFFF);
3884 static int receive_extralen(struct dlm_message *ms)
3886 return (ms->m_header.h_length - sizeof(struct dlm_message));
3889 static int receive_lvb(struct dlm_ls *ls, struct dlm_lkb *lkb,
3890 struct dlm_message *ms)
3894 if (lkb->lkb_exflags & DLM_LKF_VALBLK) {
3895 if (!lkb->lkb_lvbptr)
3896 lkb->lkb_lvbptr = dlm_allocate_lvb(ls);
3897 if (!lkb->lkb_lvbptr)
3899 len = receive_extralen(ms);
3900 if (len > ls->ls_lvblen)
3901 len = ls->ls_lvblen;
3902 memcpy(lkb->lkb_lvbptr, ms->m_extra, len);
3907 static void fake_bastfn(void *astparam, int mode)
3909 log_print("fake_bastfn should not be called");
3912 static void fake_astfn(void *astparam)
3914 log_print("fake_astfn should not be called");
3917 static int receive_request_args(struct dlm_ls *ls, struct dlm_lkb *lkb,
3918 struct dlm_message *ms)
3920 lkb->lkb_nodeid = ms->m_header.h_nodeid;
3921 lkb->lkb_ownpid = ms->m_pid;
3922 lkb->lkb_remid = ms->m_lkid;
3923 lkb->lkb_grmode = DLM_LOCK_IV;
3924 lkb->lkb_rqmode = ms->m_rqmode;
3926 lkb->lkb_bastfn = (ms->m_asts & DLM_CB_BAST) ? &fake_bastfn : NULL;
3927 lkb->lkb_astfn = (ms->m_asts & DLM_CB_CAST) ? &fake_astfn : NULL;
3929 if (lkb->lkb_exflags & DLM_LKF_VALBLK) {
3930 /* lkb was just created so there won't be an lvb yet */
3931 lkb->lkb_lvbptr = dlm_allocate_lvb(ls);
3932 if (!lkb->lkb_lvbptr)
3939 static int receive_convert_args(struct dlm_ls *ls, struct dlm_lkb *lkb,
3940 struct dlm_message *ms)
3942 if (lkb->lkb_status != DLM_LKSTS_GRANTED)
3945 if (receive_lvb(ls, lkb, ms))
3948 lkb->lkb_rqmode = ms->m_rqmode;
3949 lkb->lkb_lvbseq = ms->m_lvbseq;
3954 static int receive_unlock_args(struct dlm_ls *ls, struct dlm_lkb *lkb,
3955 struct dlm_message *ms)
3957 if (receive_lvb(ls, lkb, ms))
3962 /* We fill in the stub-lkb fields with the info that send_xxxx_reply()
3963 uses to send a reply and that the remote end uses to process the reply. */
3965 static void setup_stub_lkb(struct dlm_ls *ls, struct dlm_message *ms)
3967 struct dlm_lkb *lkb = &ls->ls_stub_lkb;
3968 lkb->lkb_nodeid = ms->m_header.h_nodeid;
3969 lkb->lkb_remid = ms->m_lkid;
3972 /* This is called after the rsb is locked so that we can safely inspect
3973 fields in the lkb. */
3975 static int validate_message(struct dlm_lkb *lkb, struct dlm_message *ms)
3977 int from = ms->m_header.h_nodeid;
3980 switch (ms->m_type) {
3981 case DLM_MSG_CONVERT:
3982 case DLM_MSG_UNLOCK:
3983 case DLM_MSG_CANCEL:
3984 if (!is_master_copy(lkb) || lkb->lkb_nodeid != from)
3988 case DLM_MSG_CONVERT_REPLY:
3989 case DLM_MSG_UNLOCK_REPLY:
3990 case DLM_MSG_CANCEL_REPLY:
3993 if (!is_process_copy(lkb) || lkb->lkb_nodeid != from)
3997 case DLM_MSG_REQUEST_REPLY:
3998 if (!is_process_copy(lkb))
4000 else if (lkb->lkb_nodeid != -1 && lkb->lkb_nodeid != from)
4009 log_error(lkb->lkb_resource->res_ls,
4010 "ignore invalid message %d from %d %x %x %x %d",
4011 ms->m_type, from, lkb->lkb_id, lkb->lkb_remid,
4012 lkb->lkb_flags, lkb->lkb_nodeid);
4016 static void send_repeat_remove(struct dlm_ls *ls, char *ms_name, int len)
4018 char name[DLM_RESNAME_MAXLEN + 1];
4019 struct dlm_message *ms;
4020 struct dlm_mhandle *mh;
4025 memset(name, 0, sizeof(name));
4026 memcpy(name, ms_name, len);
4028 hash = jhash(name, len, 0);
4029 b = hash & (ls->ls_rsbtbl_size - 1);
4031 dir_nodeid = dlm_hash2nodeid(ls, hash);
4033 log_error(ls, "send_repeat_remove dir %d %s", dir_nodeid, name);
4035 spin_lock(&ls->ls_rsbtbl[b].lock);
4036 rv = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].keep, name, len, &r);
4038 spin_unlock(&ls->ls_rsbtbl[b].lock);
4039 log_error(ls, "repeat_remove on keep %s", name);
4043 rv = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].toss, name, len, &r);
4045 spin_unlock(&ls->ls_rsbtbl[b].lock);
4046 log_error(ls, "repeat_remove on toss %s", name);
4050 /* use ls->remove_name2 to avoid conflict with shrink? */
4052 spin_lock(&ls->ls_remove_spin);
4053 ls->ls_remove_len = len;
4054 memcpy(ls->ls_remove_name, name, DLM_RESNAME_MAXLEN);
4055 spin_unlock(&ls->ls_remove_spin);
4056 spin_unlock(&ls->ls_rsbtbl[b].lock);
4058 rv = _create_message(ls, sizeof(struct dlm_message) + len,
4059 dir_nodeid, DLM_MSG_REMOVE, &ms, &mh);
4063 memcpy(ms->m_extra, name, len);
4066 send_message(mh, ms);
4068 spin_lock(&ls->ls_remove_spin);
4069 ls->ls_remove_len = 0;
4070 memset(ls->ls_remove_name, 0, DLM_RESNAME_MAXLEN);
4071 spin_unlock(&ls->ls_remove_spin);
4074 static int receive_request(struct dlm_ls *ls, struct dlm_message *ms)
4076 struct dlm_lkb *lkb;
4079 int error, namelen = 0;
4081 from_nodeid = ms->m_header.h_nodeid;
4083 error = create_lkb(ls, &lkb);
4087 receive_flags(lkb, ms);
4088 lkb->lkb_flags |= DLM_IFL_MSTCPY;
4089 error = receive_request_args(ls, lkb, ms);
4095 /* The dir node is the authority on whether we are the master
4096 for this rsb or not, so if the master sends us a request, we should
4097 recreate the rsb if we've destroyed it. This race happens when we
4098 send a remove message to the dir node at the same time that the dir
4099 node sends us a request for the rsb. */
4101 namelen = receive_extralen(ms);
4103 error = find_rsb(ls, ms->m_extra, namelen, from_nodeid,
4104 R_RECEIVE_REQUEST, &r);
4112 if (r->res_master_nodeid != dlm_our_nodeid()) {
4113 error = validate_master_nodeid(ls, r, from_nodeid);
4123 error = do_request(r, lkb);
4124 send_request_reply(r, lkb, error);
4125 do_request_effects(r, lkb, error);
4130 if (error == -EINPROGRESS)
4137 /* TODO: instead of returning ENOTBLK, add the lkb to res_lookup
4138 and do this receive_request again from process_lookup_list once
4139 we get the lookup reply. This would avoid a many repeated
4140 ENOTBLK request failures when the lookup reply designating us
4141 as master is delayed. */
4143 /* We could repeatedly return -EBADR here if our send_remove() is
4144 delayed in being sent/arriving/being processed on the dir node.
4145 Another node would repeatedly lookup up the master, and the dir
4146 node would continue returning our nodeid until our send_remove
4149 We send another remove message in case our previous send_remove
4150 was lost/ignored/missed somehow. */
4152 if (error != -ENOTBLK) {
4153 log_limit(ls, "receive_request %x from %d %d",
4154 ms->m_lkid, from_nodeid, error);
4157 if (namelen && error == -EBADR) {
4158 send_repeat_remove(ls, ms->m_extra, namelen);
4162 setup_stub_lkb(ls, ms);
4163 send_request_reply(&ls->ls_stub_rsb, &ls->ls_stub_lkb, error);
4167 static int receive_convert(struct dlm_ls *ls, struct dlm_message *ms)
4169 struct dlm_lkb *lkb;
4171 int error, reply = 1;
4173 error = find_lkb(ls, ms->m_remid, &lkb);
4177 if (lkb->lkb_remid != ms->m_lkid) {
4178 log_error(ls, "receive_convert %x remid %x recover_seq %llu "
4179 "remote %d %x", lkb->lkb_id, lkb->lkb_remid,
4180 (unsigned long long)lkb->lkb_recover_seq,
4181 ms->m_header.h_nodeid, ms->m_lkid);
4186 r = lkb->lkb_resource;
4191 error = validate_message(lkb, ms);
4195 receive_flags(lkb, ms);
4197 error = receive_convert_args(ls, lkb, ms);
4199 send_convert_reply(r, lkb, error);
4203 reply = !down_conversion(lkb);
4205 error = do_convert(r, lkb);
4207 send_convert_reply(r, lkb, error);
4208 do_convert_effects(r, lkb, error);
4216 setup_stub_lkb(ls, ms);
4217 send_convert_reply(&ls->ls_stub_rsb, &ls->ls_stub_lkb, error);
4221 static int receive_unlock(struct dlm_ls *ls, struct dlm_message *ms)
4223 struct dlm_lkb *lkb;
4227 error = find_lkb(ls, ms->m_remid, &lkb);
4231 if (lkb->lkb_remid != ms->m_lkid) {
4232 log_error(ls, "receive_unlock %x remid %x remote %d %x",
4233 lkb->lkb_id, lkb->lkb_remid,
4234 ms->m_header.h_nodeid, ms->m_lkid);
4239 r = lkb->lkb_resource;
4244 error = validate_message(lkb, ms);
4248 receive_flags(lkb, ms);
4250 error = receive_unlock_args(ls, lkb, ms);
4252 send_unlock_reply(r, lkb, error);
4256 error = do_unlock(r, lkb);
4257 send_unlock_reply(r, lkb, error);
4258 do_unlock_effects(r, lkb, error);
4266 setup_stub_lkb(ls, ms);
4267 send_unlock_reply(&ls->ls_stub_rsb, &ls->ls_stub_lkb, error);
4271 static int receive_cancel(struct dlm_ls *ls, struct dlm_message *ms)
4273 struct dlm_lkb *lkb;
4277 error = find_lkb(ls, ms->m_remid, &lkb);
4281 receive_flags(lkb, ms);
4283 r = lkb->lkb_resource;
4288 error = validate_message(lkb, ms);
4292 error = do_cancel(r, lkb);
4293 send_cancel_reply(r, lkb, error);
4294 do_cancel_effects(r, lkb, error);
4302 setup_stub_lkb(ls, ms);
4303 send_cancel_reply(&ls->ls_stub_rsb, &ls->ls_stub_lkb, error);
4307 static int receive_grant(struct dlm_ls *ls, struct dlm_message *ms)
4309 struct dlm_lkb *lkb;
4313 error = find_lkb(ls, ms->m_remid, &lkb);
4317 r = lkb->lkb_resource;
4322 error = validate_message(lkb, ms);
4326 receive_flags_reply(lkb, ms);
4327 if (is_altmode(lkb))
4328 munge_altmode(lkb, ms);
4329 grant_lock_pc(r, lkb, ms);
4330 queue_cast(r, lkb, 0);
4338 static int receive_bast(struct dlm_ls *ls, struct dlm_message *ms)
4340 struct dlm_lkb *lkb;
4344 error = find_lkb(ls, ms->m_remid, &lkb);
4348 r = lkb->lkb_resource;
4353 error = validate_message(lkb, ms);
4357 queue_bast(r, lkb, ms->m_bastmode);
4358 lkb->lkb_highbast = ms->m_bastmode;
4366 static void receive_lookup(struct dlm_ls *ls, struct dlm_message *ms)
4368 int len, error, ret_nodeid, from_nodeid, our_nodeid;
4370 from_nodeid = ms->m_header.h_nodeid;
4371 our_nodeid = dlm_our_nodeid();
4373 len = receive_extralen(ms);
4375 error = dlm_master_lookup(ls, from_nodeid, ms->m_extra, len, 0,
4378 /* Optimization: we're master so treat lookup as a request */
4379 if (!error && ret_nodeid == our_nodeid) {
4380 receive_request(ls, ms);
4383 send_lookup_reply(ls, ms, ret_nodeid, error);
4386 static void receive_remove(struct dlm_ls *ls, struct dlm_message *ms)
4388 char name[DLM_RESNAME_MAXLEN+1];
4391 int rv, len, dir_nodeid, from_nodeid;
4393 from_nodeid = ms->m_header.h_nodeid;
4395 len = receive_extralen(ms);
4397 if (len > DLM_RESNAME_MAXLEN) {
4398 log_error(ls, "receive_remove from %d bad len %d",
4403 dir_nodeid = dlm_hash2nodeid(ls, ms->m_hash);
4404 if (dir_nodeid != dlm_our_nodeid()) {
4405 log_error(ls, "receive_remove from %d bad nodeid %d",
4406 from_nodeid, dir_nodeid);
4410 /* Look for name on rsbtbl.toss, if it's there, kill it.
4411 If it's on rsbtbl.keep, it's being used, and we should ignore this
4412 message. This is an expected race between the dir node sending a
4413 request to the master node at the same time as the master node sends
4414 a remove to the dir node. The resolution to that race is for the
4415 dir node to ignore the remove message, and the master node to
4416 recreate the master rsb when it gets a request from the dir node for
4417 an rsb it doesn't have. */
4419 memset(name, 0, sizeof(name));
4420 memcpy(name, ms->m_extra, len);
4422 hash = jhash(name, len, 0);
4423 b = hash & (ls->ls_rsbtbl_size - 1);
4425 spin_lock(&ls->ls_rsbtbl[b].lock);
4427 rv = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].toss, name, len, &r);
4429 /* verify the rsb is on keep list per comment above */
4430 rv = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].keep, name, len, &r);
4432 /* should not happen */
4433 log_error(ls, "receive_remove from %d not found %s",
4435 spin_unlock(&ls->ls_rsbtbl[b].lock);
4438 if (r->res_master_nodeid != from_nodeid) {
4439 /* should not happen */
4440 log_error(ls, "receive_remove keep from %d master %d",
4441 from_nodeid, r->res_master_nodeid);
4443 spin_unlock(&ls->ls_rsbtbl[b].lock);
4447 log_debug(ls, "receive_remove from %d master %d first %x %s",
4448 from_nodeid, r->res_master_nodeid, r->res_first_lkid,
4450 spin_unlock(&ls->ls_rsbtbl[b].lock);
4454 if (r->res_master_nodeid != from_nodeid) {
4455 log_error(ls, "receive_remove toss from %d master %d",
4456 from_nodeid, r->res_master_nodeid);
4458 spin_unlock(&ls->ls_rsbtbl[b].lock);
4462 if (kref_put(&r->res_ref, kill_rsb)) {
4463 rb_erase(&r->res_hashnode, &ls->ls_rsbtbl[b].toss);
4464 spin_unlock(&ls->ls_rsbtbl[b].lock);
4467 log_error(ls, "receive_remove from %d rsb ref error",
4470 spin_unlock(&ls->ls_rsbtbl[b].lock);
4474 static void receive_purge(struct dlm_ls *ls, struct dlm_message *ms)
4476 do_purge(ls, ms->m_nodeid, ms->m_pid);
4479 static int receive_request_reply(struct dlm_ls *ls, struct dlm_message *ms)
4481 struct dlm_lkb *lkb;
4483 int error, mstype, result;
4484 int from_nodeid = ms->m_header.h_nodeid;
4486 error = find_lkb(ls, ms->m_remid, &lkb);
4490 r = lkb->lkb_resource;
4494 error = validate_message(lkb, ms);
4498 mstype = lkb->lkb_wait_type;
4499 error = remove_from_waiters(lkb, DLM_MSG_REQUEST_REPLY);
4501 log_error(ls, "receive_request_reply %x remote %d %x result %d",
4502 lkb->lkb_id, from_nodeid, ms->m_lkid, ms->m_result);
4507 /* Optimization: the dir node was also the master, so it took our
4508 lookup as a request and sent request reply instead of lookup reply */
4509 if (mstype == DLM_MSG_LOOKUP) {
4510 r->res_master_nodeid = from_nodeid;
4511 r->res_nodeid = from_nodeid;
4512 lkb->lkb_nodeid = from_nodeid;
4515 /* this is the value returned from do_request() on the master */
4516 result = ms->m_result;
4520 /* request would block (be queued) on remote master */
4521 queue_cast(r, lkb, -EAGAIN);
4522 confirm_master(r, -EAGAIN);
4523 unhold_lkb(lkb); /* undoes create_lkb() */
4528 /* request was queued or granted on remote master */
4529 receive_flags_reply(lkb, ms);
4530 lkb->lkb_remid = ms->m_lkid;
4531 if (is_altmode(lkb))
4532 munge_altmode(lkb, ms);
4534 add_lkb(r, lkb, DLM_LKSTS_WAITING);
4537 grant_lock_pc(r, lkb, ms);
4538 queue_cast(r, lkb, 0);
4540 confirm_master(r, result);
4545 /* find_rsb failed to find rsb or rsb wasn't master */
4546 log_limit(ls, "receive_request_reply %x from %d %d "
4547 "master %d dir %d first %x %s", lkb->lkb_id,
4548 from_nodeid, result, r->res_master_nodeid,
4549 r->res_dir_nodeid, r->res_first_lkid, r->res_name);
4551 if (r->res_dir_nodeid != dlm_our_nodeid() &&
4552 r->res_master_nodeid != dlm_our_nodeid()) {
4553 /* cause _request_lock->set_master->send_lookup */
4554 r->res_master_nodeid = 0;
4556 lkb->lkb_nodeid = -1;
4559 if (is_overlap(lkb)) {
4560 /* we'll ignore error in cancel/unlock reply */
4561 queue_cast_overlap(r, lkb);
4562 confirm_master(r, result);
4563 unhold_lkb(lkb); /* undoes create_lkb() */
4565 _request_lock(r, lkb);
4567 if (r->res_master_nodeid == dlm_our_nodeid())
4568 confirm_master(r, 0);
4573 log_error(ls, "receive_request_reply %x error %d",
4574 lkb->lkb_id, result);
4577 if (is_overlap_unlock(lkb) && (result == 0 || result == -EINPROGRESS)) {
4578 log_debug(ls, "receive_request_reply %x result %d unlock",
4579 lkb->lkb_id, result);
4580 lkb->lkb_flags &= ~DLM_IFL_OVERLAP_UNLOCK;
4581 lkb->lkb_flags &= ~DLM_IFL_OVERLAP_CANCEL;
4582 send_unlock(r, lkb);
4583 } else if (is_overlap_cancel(lkb) && (result == -EINPROGRESS)) {
4584 log_debug(ls, "receive_request_reply %x cancel", lkb->lkb_id);
4585 lkb->lkb_flags &= ~DLM_IFL_OVERLAP_UNLOCK;
4586 lkb->lkb_flags &= ~DLM_IFL_OVERLAP_CANCEL;
4587 send_cancel(r, lkb);
4589 lkb->lkb_flags &= ~DLM_IFL_OVERLAP_CANCEL;
4590 lkb->lkb_flags &= ~DLM_IFL_OVERLAP_UNLOCK;
4599 static void __receive_convert_reply(struct dlm_rsb *r, struct dlm_lkb *lkb,
4600 struct dlm_message *ms)
4602 /* this is the value returned from do_convert() on the master */
4603 switch (ms->m_result) {
4605 /* convert would block (be queued) on remote master */
4606 queue_cast(r, lkb, -EAGAIN);
4610 receive_flags_reply(lkb, ms);
4611 revert_lock_pc(r, lkb);
4612 queue_cast(r, lkb, -EDEADLK);
4616 /* convert was queued on remote master */
4617 receive_flags_reply(lkb, ms);
4618 if (is_demoted(lkb))
4621 add_lkb(r, lkb, DLM_LKSTS_CONVERT);
4626 /* convert was granted on remote master */
4627 receive_flags_reply(lkb, ms);
4628 if (is_demoted(lkb))
4630 grant_lock_pc(r, lkb, ms);
4631 queue_cast(r, lkb, 0);
4635 log_error(r->res_ls, "receive_convert_reply %x remote %d %x %d",
4636 lkb->lkb_id, ms->m_header.h_nodeid, ms->m_lkid,
4643 static void _receive_convert_reply(struct dlm_lkb *lkb, struct dlm_message *ms)
4645 struct dlm_rsb *r = lkb->lkb_resource;
4651 error = validate_message(lkb, ms);
4655 /* stub reply can happen with waiters_mutex held */
4656 error = remove_from_waiters_ms(lkb, ms);
4660 __receive_convert_reply(r, lkb, ms);
4666 static int receive_convert_reply(struct dlm_ls *ls, struct dlm_message *ms)
4668 struct dlm_lkb *lkb;
4671 error = find_lkb(ls, ms->m_remid, &lkb);
4675 _receive_convert_reply(lkb, ms);
4680 static void _receive_unlock_reply(struct dlm_lkb *lkb, struct dlm_message *ms)
4682 struct dlm_rsb *r = lkb->lkb_resource;
4688 error = validate_message(lkb, ms);
4692 /* stub reply can happen with waiters_mutex held */
4693 error = remove_from_waiters_ms(lkb, ms);
4697 /* this is the value returned from do_unlock() on the master */
4699 switch (ms->m_result) {
4701 receive_flags_reply(lkb, ms);
4702 remove_lock_pc(r, lkb);
4703 queue_cast(r, lkb, -DLM_EUNLOCK);
4708 log_error(r->res_ls, "receive_unlock_reply %x error %d",
4709 lkb->lkb_id, ms->m_result);
4716 static int receive_unlock_reply(struct dlm_ls *ls, struct dlm_message *ms)
4718 struct dlm_lkb *lkb;
4721 error = find_lkb(ls, ms->m_remid, &lkb);
4725 _receive_unlock_reply(lkb, ms);
4730 static void _receive_cancel_reply(struct dlm_lkb *lkb, struct dlm_message *ms)
4732 struct dlm_rsb *r = lkb->lkb_resource;
4738 error = validate_message(lkb, ms);
4742 /* stub reply can happen with waiters_mutex held */
4743 error = remove_from_waiters_ms(lkb, ms);
4747 /* this is the value returned from do_cancel() on the master */
4749 switch (ms->m_result) {
4751 receive_flags_reply(lkb, ms);
4752 revert_lock_pc(r, lkb);
4753 queue_cast(r, lkb, -DLM_ECANCEL);
4758 log_error(r->res_ls, "receive_cancel_reply %x error %d",
4759 lkb->lkb_id, ms->m_result);
4766 static int receive_cancel_reply(struct dlm_ls *ls, struct dlm_message *ms)
4768 struct dlm_lkb *lkb;
4771 error = find_lkb(ls, ms->m_remid, &lkb);
4775 _receive_cancel_reply(lkb, ms);
4780 static void receive_lookup_reply(struct dlm_ls *ls, struct dlm_message *ms)
4782 struct dlm_lkb *lkb;
4784 int error, ret_nodeid;
4785 int do_lookup_list = 0;
4787 error = find_lkb(ls, ms->m_lkid, &lkb);
4789 log_error(ls, "receive_lookup_reply no lkid %x", ms->m_lkid);
4793 /* ms->m_result is the value returned by dlm_master_lookup on dir node
4794 FIXME: will a non-zero error ever be returned? */
4796 r = lkb->lkb_resource;
4800 error = remove_from_waiters(lkb, DLM_MSG_LOOKUP_REPLY);
4804 ret_nodeid = ms->m_nodeid;
4806 /* We sometimes receive a request from the dir node for this
4807 rsb before we've received the dir node's loookup_reply for it.
4808 The request from the dir node implies we're the master, so we set
4809 ourself as master in receive_request_reply, and verify here that
4810 we are indeed the master. */
4812 if (r->res_master_nodeid && (r->res_master_nodeid != ret_nodeid)) {
4813 /* This should never happen */
4814 log_error(ls, "receive_lookup_reply %x from %d ret %d "
4815 "master %d dir %d our %d first %x %s",
4816 lkb->lkb_id, ms->m_header.h_nodeid, ret_nodeid,
4817 r->res_master_nodeid, r->res_dir_nodeid,
4818 dlm_our_nodeid(), r->res_first_lkid, r->res_name);
4821 if (ret_nodeid == dlm_our_nodeid()) {
4822 r->res_master_nodeid = ret_nodeid;
4825 r->res_first_lkid = 0;
4826 } else if (ret_nodeid == -1) {
4827 /* the remote node doesn't believe it's the dir node */
4828 log_error(ls, "receive_lookup_reply %x from %d bad ret_nodeid",
4829 lkb->lkb_id, ms->m_header.h_nodeid);
4830 r->res_master_nodeid = 0;
4832 lkb->lkb_nodeid = -1;
4834 /* set_master() will set lkb_nodeid from r */
4835 r->res_master_nodeid = ret_nodeid;
4836 r->res_nodeid = ret_nodeid;
4839 if (is_overlap(lkb)) {
4840 log_debug(ls, "receive_lookup_reply %x unlock %x",
4841 lkb->lkb_id, lkb->lkb_flags);
4842 queue_cast_overlap(r, lkb);
4843 unhold_lkb(lkb); /* undoes create_lkb() */
4847 _request_lock(r, lkb);
4851 process_lookup_list(r);
4858 static void _receive_message(struct dlm_ls *ls, struct dlm_message *ms,
4861 int error = 0, noent = 0;
4863 if (!dlm_is_member(ls, ms->m_header.h_nodeid)) {
4864 log_limit(ls, "receive %d from non-member %d %x %x %d",
4865 ms->m_type, ms->m_header.h_nodeid, ms->m_lkid,
4866 ms->m_remid, ms->m_result);
4870 switch (ms->m_type) {
4872 /* messages sent to a master node */
4874 case DLM_MSG_REQUEST:
4875 error = receive_request(ls, ms);
4878 case DLM_MSG_CONVERT:
4879 error = receive_convert(ls, ms);
4882 case DLM_MSG_UNLOCK:
4883 error = receive_unlock(ls, ms);
4886 case DLM_MSG_CANCEL:
4888 error = receive_cancel(ls, ms);
4891 /* messages sent from a master node (replies to above) */
4893 case DLM_MSG_REQUEST_REPLY:
4894 error = receive_request_reply(ls, ms);
4897 case DLM_MSG_CONVERT_REPLY:
4898 error = receive_convert_reply(ls, ms);
4901 case DLM_MSG_UNLOCK_REPLY:
4902 error = receive_unlock_reply(ls, ms);
4905 case DLM_MSG_CANCEL_REPLY:
4906 error = receive_cancel_reply(ls, ms);
4909 /* messages sent from a master node (only two types of async msg) */
4913 error = receive_grant(ls, ms);
4918 error = receive_bast(ls, ms);
4921 /* messages sent to a dir node */
4923 case DLM_MSG_LOOKUP:
4924 receive_lookup(ls, ms);
4927 case DLM_MSG_REMOVE:
4928 receive_remove(ls, ms);
4931 /* messages sent from a dir node (remove has no reply) */
4933 case DLM_MSG_LOOKUP_REPLY:
4934 receive_lookup_reply(ls, ms);
4937 /* other messages */
4940 receive_purge(ls, ms);
4944 log_error(ls, "unknown message type %d", ms->m_type);
4948 * When checking for ENOENT, we're checking the result of
4949 * find_lkb(m_remid):
4951 * The lock id referenced in the message wasn't found. This may
4952 * happen in normal usage for the async messages and cancel, so
4953 * only use log_debug for them.
4955 * Some errors are expected and normal.
4958 if (error == -ENOENT && noent) {
4959 log_debug(ls, "receive %d no %x remote %d %x saved_seq %u",
4960 ms->m_type, ms->m_remid, ms->m_header.h_nodeid,
4961 ms->m_lkid, saved_seq);
4962 } else if (error == -ENOENT) {
4963 log_error(ls, "receive %d no %x remote %d %x saved_seq %u",
4964 ms->m_type, ms->m_remid, ms->m_header.h_nodeid,
4965 ms->m_lkid, saved_seq);
4967 if (ms->m_type == DLM_MSG_CONVERT)
4968 dlm_dump_rsb_hash(ls, ms->m_hash);
4971 if (error == -EINVAL) {
4972 log_error(ls, "receive %d inval from %d lkid %x remid %x "
4974 ms->m_type, ms->m_header.h_nodeid,
4975 ms->m_lkid, ms->m_remid, saved_seq);
4979 /* If the lockspace is in recovery mode (locking stopped), then normal
4980 messages are saved on the requestqueue for processing after recovery is
4981 done. When not in recovery mode, we wait for dlm_recoverd to drain saved
4982 messages off the requestqueue before we process new ones. This occurs right
4983 after recovery completes when we transition from saving all messages on
4984 requestqueue, to processing all the saved messages, to processing new
4985 messages as they arrive. */
4987 static void dlm_receive_message(struct dlm_ls *ls, struct dlm_message *ms,
4990 if (dlm_locking_stopped(ls)) {
4991 /* If we were a member of this lockspace, left, and rejoined,
4992 other nodes may still be sending us messages from the
4993 lockspace generation before we left. */
4994 if (!ls->ls_generation) {
4995 log_limit(ls, "receive %d from %d ignore old gen",
4996 ms->m_type, nodeid);
5000 dlm_add_requestqueue(ls, nodeid, ms);
5002 dlm_wait_requestqueue(ls);
5003 _receive_message(ls, ms, 0);
5007 /* This is called by dlm_recoverd to process messages that were saved on
5008 the requestqueue. */
5010 void dlm_receive_message_saved(struct dlm_ls *ls, struct dlm_message *ms,
5013 _receive_message(ls, ms, saved_seq);
5016 /* This is called by the midcomms layer when something is received for
5017 the lockspace. It could be either a MSG (normal message sent as part of
5018 standard locking activity) or an RCOM (recovery message sent as part of
5019 lockspace recovery). */
5021 void dlm_receive_buffer(union dlm_packet *p, int nodeid)
5023 struct dlm_header *hd = &p->header;
5027 switch (hd->h_cmd) {
5029 dlm_message_in(&p->message);
5030 type = p->message.m_type;
5033 dlm_rcom_in(&p->rcom);
5034 type = p->rcom.rc_type;
5037 log_print("invalid h_cmd %d from %u", hd->h_cmd, nodeid);
5041 if (hd->h_nodeid != nodeid) {
5042 log_print("invalid h_nodeid %d from %d lockspace %x",
5043 hd->h_nodeid, nodeid, hd->h_lockspace);
5047 ls = dlm_find_lockspace_global(hd->h_lockspace);
5049 if (dlm_config.ci_log_debug) {
5050 printk_ratelimited(KERN_DEBUG "dlm: invalid lockspace "
5051 "%u from %d cmd %d type %d\n",
5052 hd->h_lockspace, nodeid, hd->h_cmd, type);
5055 if (hd->h_cmd == DLM_RCOM && type == DLM_RCOM_STATUS)
5056 dlm_send_ls_not_ready(nodeid, &p->rcom);
5060 /* this rwsem allows dlm_ls_stop() to wait for all dlm_recv threads to
5061 be inactive (in this ls) before transitioning to recovery mode */
5063 down_read(&ls->ls_recv_active);
5064 if (hd->h_cmd == DLM_MSG)
5065 dlm_receive_message(ls, &p->message, nodeid);
5067 dlm_receive_rcom(ls, &p->rcom, nodeid);
5068 up_read(&ls->ls_recv_active);
5070 dlm_put_lockspace(ls);
5073 static void recover_convert_waiter(struct dlm_ls *ls, struct dlm_lkb *lkb,
5074 struct dlm_message *ms_stub)
5076 if (middle_conversion(lkb)) {
5078 memset(ms_stub, 0, sizeof(struct dlm_message));
5079 ms_stub->m_flags = DLM_IFL_STUB_MS;
5080 ms_stub->m_type = DLM_MSG_CONVERT_REPLY;
5081 ms_stub->m_result = -EINPROGRESS;
5082 ms_stub->m_header.h_nodeid = lkb->lkb_nodeid;
5083 _receive_convert_reply(lkb, ms_stub);
5085 /* Same special case as in receive_rcom_lock_args() */
5086 lkb->lkb_grmode = DLM_LOCK_IV;
5087 rsb_set_flag(lkb->lkb_resource, RSB_RECOVER_CONVERT);
5090 } else if (lkb->lkb_rqmode >= lkb->lkb_grmode) {
5091 lkb->lkb_flags |= DLM_IFL_RESEND;
5094 /* lkb->lkb_rqmode < lkb->lkb_grmode shouldn't happen since down
5095 conversions are async; there's no reply from the remote master */
5098 /* A waiting lkb needs recovery if the master node has failed, or
5099 the master node is changing (only when no directory is used) */
5101 static int waiter_needs_recovery(struct dlm_ls *ls, struct dlm_lkb *lkb,
5104 if (dlm_no_directory(ls))
5107 if (dlm_is_removed(ls, lkb->lkb_wait_nodeid))
5113 /* Recovery for locks that are waiting for replies from nodes that are now
5114 gone. We can just complete unlocks and cancels by faking a reply from the
5115 dead node. Requests and up-conversions we flag to be resent after
5116 recovery. Down-conversions can just be completed with a fake reply like
5117 unlocks. Conversions between PR and CW need special attention. */
5119 void dlm_recover_waiters_pre(struct dlm_ls *ls)
5121 struct dlm_lkb *lkb, *safe;
5122 struct dlm_message *ms_stub;
5123 int wait_type, stub_unlock_result, stub_cancel_result;
5126 ms_stub = kmalloc(sizeof(*ms_stub), GFP_KERNEL);
5130 mutex_lock(&ls->ls_waiters_mutex);
5132 list_for_each_entry_safe(lkb, safe, &ls->ls_waiters, lkb_wait_reply) {
5134 dir_nodeid = dlm_dir_nodeid(lkb->lkb_resource);
5136 /* exclude debug messages about unlocks because there can be so
5137 many and they aren't very interesting */
5139 if (lkb->lkb_wait_type != DLM_MSG_UNLOCK) {
5140 log_debug(ls, "waiter %x remote %x msg %d r_nodeid %d "
5141 "lkb_nodeid %d wait_nodeid %d dir_nodeid %d",
5145 lkb->lkb_resource->res_nodeid,
5147 lkb->lkb_wait_nodeid,
5151 /* all outstanding lookups, regardless of destination will be
5152 resent after recovery is done */
5154 if (lkb->lkb_wait_type == DLM_MSG_LOOKUP) {
5155 lkb->lkb_flags |= DLM_IFL_RESEND;
5159 if (!waiter_needs_recovery(ls, lkb, dir_nodeid))
5162 wait_type = lkb->lkb_wait_type;
5163 stub_unlock_result = -DLM_EUNLOCK;
5164 stub_cancel_result = -DLM_ECANCEL;
5166 /* Main reply may have been received leaving a zero wait_type,
5167 but a reply for the overlapping op may not have been
5168 received. In that case we need to fake the appropriate
5169 reply for the overlap op. */
5172 if (is_overlap_cancel(lkb)) {
5173 wait_type = DLM_MSG_CANCEL;
5174 if (lkb->lkb_grmode == DLM_LOCK_IV)
5175 stub_cancel_result = 0;
5177 if (is_overlap_unlock(lkb)) {
5178 wait_type = DLM_MSG_UNLOCK;
5179 if (lkb->lkb_grmode == DLM_LOCK_IV)
5180 stub_unlock_result = -ENOENT;
5183 log_debug(ls, "rwpre overlap %x %x %d %d %d",
5184 lkb->lkb_id, lkb->lkb_flags, wait_type,
5185 stub_cancel_result, stub_unlock_result);
5188 switch (wait_type) {
5190 case DLM_MSG_REQUEST:
5191 lkb->lkb_flags |= DLM_IFL_RESEND;
5194 case DLM_MSG_CONVERT:
5195 recover_convert_waiter(ls, lkb, ms_stub);
5198 case DLM_MSG_UNLOCK:
5200 memset(ms_stub, 0, sizeof(struct dlm_message));
5201 ms_stub->m_flags = DLM_IFL_STUB_MS;
5202 ms_stub->m_type = DLM_MSG_UNLOCK_REPLY;
5203 ms_stub->m_result = stub_unlock_result;
5204 ms_stub->m_header.h_nodeid = lkb->lkb_nodeid;
5205 _receive_unlock_reply(lkb, ms_stub);
5209 case DLM_MSG_CANCEL:
5211 memset(ms_stub, 0, sizeof(struct dlm_message));
5212 ms_stub->m_flags = DLM_IFL_STUB_MS;
5213 ms_stub->m_type = DLM_MSG_CANCEL_REPLY;
5214 ms_stub->m_result = stub_cancel_result;
5215 ms_stub->m_header.h_nodeid = lkb->lkb_nodeid;
5216 _receive_cancel_reply(lkb, ms_stub);
5221 log_error(ls, "invalid lkb wait_type %d %d",
5222 lkb->lkb_wait_type, wait_type);
5226 mutex_unlock(&ls->ls_waiters_mutex);
5230 static struct dlm_lkb *find_resend_waiter(struct dlm_ls *ls)
5232 struct dlm_lkb *lkb;
5235 mutex_lock(&ls->ls_waiters_mutex);
5236 list_for_each_entry(lkb, &ls->ls_waiters, lkb_wait_reply) {
5237 if (lkb->lkb_flags & DLM_IFL_RESEND) {
5243 mutex_unlock(&ls->ls_waiters_mutex);
5250 /* Deal with lookups and lkb's marked RESEND from _pre. We may now be the
5251 master or dir-node for r. Processing the lkb may result in it being placed
5254 /* We do this after normal locking has been enabled and any saved messages
5255 (in requestqueue) have been processed. We should be confident that at
5256 this point we won't get or process a reply to any of these waiting
5257 operations. But, new ops may be coming in on the rsbs/locks here from
5258 userspace or remotely. */
5260 /* there may have been an overlap unlock/cancel prior to recovery or after
5261 recovery. if before, the lkb may still have a pos wait_count; if after, the
5262 overlap flag would just have been set and nothing new sent. we can be
5263 confident here than any replies to either the initial op or overlap ops
5264 prior to recovery have been received. */
5266 int dlm_recover_waiters_post(struct dlm_ls *ls)
5268 struct dlm_lkb *lkb;
5270 int error = 0, mstype, err, oc, ou;
5273 if (dlm_locking_stopped(ls)) {
5274 log_debug(ls, "recover_waiters_post aborted");
5279 lkb = find_resend_waiter(ls);
5283 r = lkb->lkb_resource;
5287 mstype = lkb->lkb_wait_type;
5288 oc = is_overlap_cancel(lkb);
5289 ou = is_overlap_unlock(lkb);
5292 log_debug(ls, "waiter %x remote %x msg %d r_nodeid %d "
5293 "lkb_nodeid %d wait_nodeid %d dir_nodeid %d "
5294 "overlap %d %d", lkb->lkb_id, lkb->lkb_remid, mstype,
5295 r->res_nodeid, lkb->lkb_nodeid, lkb->lkb_wait_nodeid,
5296 dlm_dir_nodeid(r), oc, ou);
5298 /* At this point we assume that we won't get a reply to any
5299 previous op or overlap op on this lock. First, do a big
5300 remove_from_waiters() for all previous ops. */
5302 lkb->lkb_flags &= ~DLM_IFL_RESEND;
5303 lkb->lkb_flags &= ~DLM_IFL_OVERLAP_UNLOCK;
5304 lkb->lkb_flags &= ~DLM_IFL_OVERLAP_CANCEL;
5305 lkb->lkb_wait_type = 0;
5306 lkb->lkb_wait_count = 0;
5307 mutex_lock(&ls->ls_waiters_mutex);
5308 list_del_init(&lkb->lkb_wait_reply);
5309 mutex_unlock(&ls->ls_waiters_mutex);
5310 unhold_lkb(lkb); /* for waiters list */
5313 /* do an unlock or cancel instead of resending */
5315 case DLM_MSG_LOOKUP:
5316 case DLM_MSG_REQUEST:
5317 queue_cast(r, lkb, ou ? -DLM_EUNLOCK :
5319 unhold_lkb(lkb); /* undoes create_lkb() */
5321 case DLM_MSG_CONVERT:
5323 queue_cast(r, lkb, -DLM_ECANCEL);
5325 lkb->lkb_exflags |= DLM_LKF_FORCEUNLOCK;
5326 _unlock_lock(r, lkb);
5334 case DLM_MSG_LOOKUP:
5335 case DLM_MSG_REQUEST:
5336 _request_lock(r, lkb);
5338 confirm_master(r, 0);
5340 case DLM_MSG_CONVERT:
5341 _convert_lock(r, lkb);
5349 log_error(ls, "waiter %x msg %d r_nodeid %d "
5350 "dir_nodeid %d overlap %d %d",
5351 lkb->lkb_id, mstype, r->res_nodeid,
5352 dlm_dir_nodeid(r), oc, ou);
5362 static void purge_mstcpy_list(struct dlm_ls *ls, struct dlm_rsb *r,
5363 struct list_head *list)
5365 struct dlm_lkb *lkb, *safe;
5367 list_for_each_entry_safe(lkb, safe, list, lkb_statequeue) {
5368 if (!is_master_copy(lkb))
5371 /* don't purge lkbs we've added in recover_master_copy for
5372 the current recovery seq */
5374 if (lkb->lkb_recover_seq == ls->ls_recover_seq)
5379 /* this put should free the lkb */
5380 if (!dlm_put_lkb(lkb))
5381 log_error(ls, "purged mstcpy lkb not released");
5385 void dlm_purge_mstcpy_locks(struct dlm_rsb *r)
5387 struct dlm_ls *ls = r->res_ls;
5389 purge_mstcpy_list(ls, r, &r->res_grantqueue);
5390 purge_mstcpy_list(ls, r, &r->res_convertqueue);
5391 purge_mstcpy_list(ls, r, &r->res_waitqueue);
5394 static void purge_dead_list(struct dlm_ls *ls, struct dlm_rsb *r,
5395 struct list_head *list,
5396 int nodeid_gone, unsigned int *count)
5398 struct dlm_lkb *lkb, *safe;
5400 list_for_each_entry_safe(lkb, safe, list, lkb_statequeue) {
5401 if (!is_master_copy(lkb))
5404 if ((lkb->lkb_nodeid == nodeid_gone) ||
5405 dlm_is_removed(ls, lkb->lkb_nodeid)) {
5407 /* tell recover_lvb to invalidate the lvb
5408 because a node holding EX/PW failed */
5409 if ((lkb->lkb_exflags & DLM_LKF_VALBLK) &&
5410 (lkb->lkb_grmode >= DLM_LOCK_PW)) {
5411 rsb_set_flag(r, RSB_RECOVER_LVB_INVAL);
5416 /* this put should free the lkb */
5417 if (!dlm_put_lkb(lkb))
5418 log_error(ls, "purged dead lkb not released");
5420 rsb_set_flag(r, RSB_RECOVER_GRANT);
5427 /* Get rid of locks held by nodes that are gone. */
5429 void dlm_recover_purge(struct dlm_ls *ls)
5432 struct dlm_member *memb;
5433 int nodes_count = 0;
5434 int nodeid_gone = 0;
5435 unsigned int lkb_count = 0;
5437 /* cache one removed nodeid to optimize the common
5438 case of a single node removed */
5440 list_for_each_entry(memb, &ls->ls_nodes_gone, list) {
5442 nodeid_gone = memb->nodeid;
5448 down_write(&ls->ls_root_sem);
5449 list_for_each_entry(r, &ls->ls_root_list, res_root_list) {
5453 purge_dead_list(ls, r, &r->res_grantqueue,
5454 nodeid_gone, &lkb_count);
5455 purge_dead_list(ls, r, &r->res_convertqueue,
5456 nodeid_gone, &lkb_count);
5457 purge_dead_list(ls, r, &r->res_waitqueue,
5458 nodeid_gone, &lkb_count);
5464 up_write(&ls->ls_root_sem);
5467 log_rinfo(ls, "dlm_recover_purge %u locks for %u nodes",
5468 lkb_count, nodes_count);
5471 static struct dlm_rsb *find_grant_rsb(struct dlm_ls *ls, int bucket)
5476 spin_lock(&ls->ls_rsbtbl[bucket].lock);
5477 for (n = rb_first(&ls->ls_rsbtbl[bucket].keep); n; n = rb_next(n)) {
5478 r = rb_entry(n, struct dlm_rsb, res_hashnode);
5480 if (!rsb_flag(r, RSB_RECOVER_GRANT))
5482 if (!is_master(r)) {
5483 rsb_clear_flag(r, RSB_RECOVER_GRANT);
5487 spin_unlock(&ls->ls_rsbtbl[bucket].lock);
5490 spin_unlock(&ls->ls_rsbtbl[bucket].lock);
5495 * Attempt to grant locks on resources that we are the master of.
5496 * Locks may have become grantable during recovery because locks
5497 * from departed nodes have been purged (or not rebuilt), allowing
5498 * previously blocked locks to now be granted. The subset of rsb's
5499 * we are interested in are those with lkb's on either the convert or
5502 * Simplest would be to go through each master rsb and check for non-empty
5503 * convert or waiting queues, and attempt to grant on those rsbs.
5504 * Checking the queues requires lock_rsb, though, for which we'd need
5505 * to release the rsbtbl lock. This would make iterating through all
5506 * rsb's very inefficient. So, we rely on earlier recovery routines
5507 * to set RECOVER_GRANT on any rsb's that we should attempt to grant
5511 void dlm_recover_grant(struct dlm_ls *ls)
5515 unsigned int count = 0;
5516 unsigned int rsb_count = 0;
5517 unsigned int lkb_count = 0;
5520 r = find_grant_rsb(ls, bucket);
5522 if (bucket == ls->ls_rsbtbl_size - 1)
5530 /* the RECOVER_GRANT flag is checked in the grant path */
5531 grant_pending_locks(r, &count);
5532 rsb_clear_flag(r, RSB_RECOVER_GRANT);
5534 confirm_master(r, 0);
5541 log_rinfo(ls, "dlm_recover_grant %u locks on %u resources",
5542 lkb_count, rsb_count);
5545 static struct dlm_lkb *search_remid_list(struct list_head *head, int nodeid,
5548 struct dlm_lkb *lkb;
5550 list_for_each_entry(lkb, head, lkb_statequeue) {
5551 if (lkb->lkb_nodeid == nodeid && lkb->lkb_remid == remid)
5557 static struct dlm_lkb *search_remid(struct dlm_rsb *r, int nodeid,
5560 struct dlm_lkb *lkb;
5562 lkb = search_remid_list(&r->res_grantqueue, nodeid, remid);
5565 lkb = search_remid_list(&r->res_convertqueue, nodeid, remid);
5568 lkb = search_remid_list(&r->res_waitqueue, nodeid, remid);
5574 /* needs at least dlm_rcom + rcom_lock */
5575 static int receive_rcom_lock_args(struct dlm_ls *ls, struct dlm_lkb *lkb,
5576 struct dlm_rsb *r, struct dlm_rcom *rc)
5578 struct rcom_lock *rl = (struct rcom_lock *) rc->rc_buf;
5580 lkb->lkb_nodeid = rc->rc_header.h_nodeid;
5581 lkb->lkb_ownpid = le32_to_cpu(rl->rl_ownpid);
5582 lkb->lkb_remid = le32_to_cpu(rl->rl_lkid);
5583 lkb->lkb_exflags = le32_to_cpu(rl->rl_exflags);
5584 lkb->lkb_flags = le32_to_cpu(rl->rl_flags) & 0x0000FFFF;
5585 lkb->lkb_flags |= DLM_IFL_MSTCPY;
5586 lkb->lkb_lvbseq = le32_to_cpu(rl->rl_lvbseq);
5587 lkb->lkb_rqmode = rl->rl_rqmode;
5588 lkb->lkb_grmode = rl->rl_grmode;
5589 /* don't set lkb_status because add_lkb wants to itself */
5591 lkb->lkb_bastfn = (rl->rl_asts & DLM_CB_BAST) ? &fake_bastfn : NULL;
5592 lkb->lkb_astfn = (rl->rl_asts & DLM_CB_CAST) ? &fake_astfn : NULL;
5594 if (lkb->lkb_exflags & DLM_LKF_VALBLK) {
5595 int lvblen = rc->rc_header.h_length - sizeof(struct dlm_rcom) -
5596 sizeof(struct rcom_lock);
5597 if (lvblen > ls->ls_lvblen)
5599 lkb->lkb_lvbptr = dlm_allocate_lvb(ls);
5600 if (!lkb->lkb_lvbptr)
5602 memcpy(lkb->lkb_lvbptr, rl->rl_lvb, lvblen);
5605 /* Conversions between PR and CW (middle modes) need special handling.
5606 The real granted mode of these converting locks cannot be determined
5607 until all locks have been rebuilt on the rsb (recover_conversion) */
5609 if (rl->rl_wait_type == cpu_to_le16(DLM_MSG_CONVERT) &&
5610 middle_conversion(lkb)) {
5611 rl->rl_status = DLM_LKSTS_CONVERT;
5612 lkb->lkb_grmode = DLM_LOCK_IV;
5613 rsb_set_flag(r, RSB_RECOVER_CONVERT);
5619 /* This lkb may have been recovered in a previous aborted recovery so we need
5620 to check if the rsb already has an lkb with the given remote nodeid/lkid.
5621 If so we just send back a standard reply. If not, we create a new lkb with
5622 the given values and send back our lkid. We send back our lkid by sending
5623 back the rcom_lock struct we got but with the remid field filled in. */
5625 /* needs at least dlm_rcom + rcom_lock */
5626 int dlm_recover_master_copy(struct dlm_ls *ls, struct dlm_rcom *rc)
5628 struct rcom_lock *rl = (struct rcom_lock *) rc->rc_buf;
5630 struct dlm_lkb *lkb;
5632 int from_nodeid = rc->rc_header.h_nodeid;
5635 if (rl->rl_parent_lkid) {
5636 error = -EOPNOTSUPP;
5640 remid = le32_to_cpu(rl->rl_lkid);
5642 /* In general we expect the rsb returned to be R_MASTER, but we don't
5643 have to require it. Recovery of masters on one node can overlap
5644 recovery of locks on another node, so one node can send us MSTCPY
5645 locks before we've made ourselves master of this rsb. We can still
5646 add new MSTCPY locks that we receive here without any harm; when
5647 we make ourselves master, dlm_recover_masters() won't touch the
5648 MSTCPY locks we've received early. */
5650 error = find_rsb(ls, rl->rl_name, le16_to_cpu(rl->rl_namelen),
5651 from_nodeid, R_RECEIVE_RECOVER, &r);
5657 if (dlm_no_directory(ls) && (dlm_dir_nodeid(r) != dlm_our_nodeid())) {
5658 log_error(ls, "dlm_recover_master_copy remote %d %x not dir",
5659 from_nodeid, remid);
5664 lkb = search_remid(r, from_nodeid, remid);
5670 error = create_lkb(ls, &lkb);
5674 error = receive_rcom_lock_args(ls, lkb, r, rc);
5681 add_lkb(r, lkb, rl->rl_status);
5683 ls->ls_recover_locks_in++;
5685 if (!list_empty(&r->res_waitqueue) || !list_empty(&r->res_convertqueue))
5686 rsb_set_flag(r, RSB_RECOVER_GRANT);
5689 /* this is the new value returned to the lock holder for
5690 saving in its process-copy lkb */
5691 rl->rl_remid = cpu_to_le32(lkb->lkb_id);
5693 lkb->lkb_recover_seq = ls->ls_recover_seq;
5699 if (error && error != -EEXIST)
5700 log_rinfo(ls, "dlm_recover_master_copy remote %d %x error %d",
5701 from_nodeid, remid, error);
5702 rl->rl_result = cpu_to_le32(error);
5706 /* needs at least dlm_rcom + rcom_lock */
5707 int dlm_recover_process_copy(struct dlm_ls *ls, struct dlm_rcom *rc)
5709 struct rcom_lock *rl = (struct rcom_lock *) rc->rc_buf;
5711 struct dlm_lkb *lkb;
5712 uint32_t lkid, remid;
5715 lkid = le32_to_cpu(rl->rl_lkid);
5716 remid = le32_to_cpu(rl->rl_remid);
5717 result = le32_to_cpu(rl->rl_result);
5719 error = find_lkb(ls, lkid, &lkb);
5721 log_error(ls, "dlm_recover_process_copy no %x remote %d %x %d",
5722 lkid, rc->rc_header.h_nodeid, remid, result);
5726 r = lkb->lkb_resource;
5730 if (!is_process_copy(lkb)) {
5731 log_error(ls, "dlm_recover_process_copy bad %x remote %d %x %d",
5732 lkid, rc->rc_header.h_nodeid, remid, result);
5742 /* There's a chance the new master received our lock before
5743 dlm_recover_master_reply(), this wouldn't happen if we did
5744 a barrier between recover_masters and recover_locks. */
5746 log_debug(ls, "dlm_recover_process_copy %x remote %d %x %d",
5747 lkid, rc->rc_header.h_nodeid, remid, result);
5749 dlm_send_rcom_lock(r, lkb);
5753 lkb->lkb_remid = remid;
5756 log_error(ls, "dlm_recover_process_copy %x remote %d %x %d unk",
5757 lkid, rc->rc_header.h_nodeid, remid, result);
5760 /* an ack for dlm_recover_locks() which waits for replies from
5761 all the locks it sends to new masters */
5762 dlm_recovered_lock(r);
5771 int dlm_user_request(struct dlm_ls *ls, struct dlm_user_args *ua,
5772 int mode, uint32_t flags, void *name, unsigned int namelen,
5773 unsigned long timeout_cs)
5775 struct dlm_lkb *lkb;
5776 struct dlm_args args;
5779 dlm_lock_recovery(ls);
5781 error = create_lkb(ls, &lkb);
5787 if (flags & DLM_LKF_VALBLK) {
5788 ua->lksb.sb_lvbptr = kzalloc(DLM_USER_LVB_LEN, GFP_NOFS);
5789 if (!ua->lksb.sb_lvbptr) {
5797 /* After ua is attached to lkb it will be freed by dlm_free_lkb().
5798 When DLM_IFL_USER is set, the dlm knows that this is a userspace
5799 lock and that lkb_astparam is the dlm_user_args structure. */
5801 error = set_lock_args(mode, &ua->lksb, flags, namelen, timeout_cs,
5802 fake_astfn, ua, fake_bastfn, &args);
5803 lkb->lkb_flags |= DLM_IFL_USER;
5810 error = request_lock(ls, lkb, name, namelen, &args);
5826 /* add this new lkb to the per-process list of locks */
5827 spin_lock(&ua->proc->locks_spin);
5829 list_add_tail(&lkb->lkb_ownqueue, &ua->proc->locks);
5830 spin_unlock(&ua->proc->locks_spin);
5832 dlm_unlock_recovery(ls);
5836 int dlm_user_convert(struct dlm_ls *ls, struct dlm_user_args *ua_tmp,
5837 int mode, uint32_t flags, uint32_t lkid, char *lvb_in,
5838 unsigned long timeout_cs)
5840 struct dlm_lkb *lkb;
5841 struct dlm_args args;
5842 struct dlm_user_args *ua;
5845 dlm_lock_recovery(ls);
5847 error = find_lkb(ls, lkid, &lkb);
5851 /* user can change the params on its lock when it converts it, or
5852 add an lvb that didn't exist before */
5856 if (flags & DLM_LKF_VALBLK && !ua->lksb.sb_lvbptr) {
5857 ua->lksb.sb_lvbptr = kzalloc(DLM_USER_LVB_LEN, GFP_NOFS);
5858 if (!ua->lksb.sb_lvbptr) {
5863 if (lvb_in && ua->lksb.sb_lvbptr)
5864 memcpy(ua->lksb.sb_lvbptr, lvb_in, DLM_USER_LVB_LEN);
5866 ua->xid = ua_tmp->xid;
5867 ua->castparam = ua_tmp->castparam;
5868 ua->castaddr = ua_tmp->castaddr;
5869 ua->bastparam = ua_tmp->bastparam;
5870 ua->bastaddr = ua_tmp->bastaddr;
5871 ua->user_lksb = ua_tmp->user_lksb;
5873 error = set_lock_args(mode, &ua->lksb, flags, 0, timeout_cs,
5874 fake_astfn, ua, fake_bastfn, &args);
5878 error = convert_lock(ls, lkb, &args);
5880 if (error == -EINPROGRESS || error == -EAGAIN || error == -EDEADLK)
5885 dlm_unlock_recovery(ls);
5891 * The caller asks for an orphan lock on a given resource with a given mode.
5892 * If a matching lock exists, it's moved to the owner's list of locks and
5893 * the lkid is returned.
5896 int dlm_user_adopt_orphan(struct dlm_ls *ls, struct dlm_user_args *ua_tmp,
5897 int mode, uint32_t flags, void *name, unsigned int namelen,
5898 unsigned long timeout_cs, uint32_t *lkid)
5900 struct dlm_lkb *lkb;
5901 struct dlm_user_args *ua;
5902 int found_other_mode = 0;
5906 mutex_lock(&ls->ls_orphans_mutex);
5907 list_for_each_entry(lkb, &ls->ls_orphans, lkb_ownqueue) {
5908 if (lkb->lkb_resource->res_length != namelen)
5910 if (memcmp(lkb->lkb_resource->res_name, name, namelen))
5912 if (lkb->lkb_grmode != mode) {
5913 found_other_mode = 1;
5918 list_del_init(&lkb->lkb_ownqueue);
5919 lkb->lkb_flags &= ~DLM_IFL_ORPHAN;
5920 *lkid = lkb->lkb_id;
5923 mutex_unlock(&ls->ls_orphans_mutex);
5925 if (!found && found_other_mode) {
5935 lkb->lkb_exflags = flags;
5936 lkb->lkb_ownpid = (int) current->pid;
5940 ua->proc = ua_tmp->proc;
5941 ua->xid = ua_tmp->xid;
5942 ua->castparam = ua_tmp->castparam;
5943 ua->castaddr = ua_tmp->castaddr;
5944 ua->bastparam = ua_tmp->bastparam;
5945 ua->bastaddr = ua_tmp->bastaddr;
5946 ua->user_lksb = ua_tmp->user_lksb;
5949 * The lkb reference from the ls_orphans list was not
5950 * removed above, and is now considered the reference
5951 * for the proc locks list.
5954 spin_lock(&ua->proc->locks_spin);
5955 list_add_tail(&lkb->lkb_ownqueue, &ua->proc->locks);
5956 spin_unlock(&ua->proc->locks_spin);
5962 int dlm_user_unlock(struct dlm_ls *ls, struct dlm_user_args *ua_tmp,
5963 uint32_t flags, uint32_t lkid, char *lvb_in)
5965 struct dlm_lkb *lkb;
5966 struct dlm_args args;
5967 struct dlm_user_args *ua;
5970 dlm_lock_recovery(ls);
5972 error = find_lkb(ls, lkid, &lkb);
5978 if (lvb_in && ua->lksb.sb_lvbptr)
5979 memcpy(ua->lksb.sb_lvbptr, lvb_in, DLM_USER_LVB_LEN);
5980 if (ua_tmp->castparam)
5981 ua->castparam = ua_tmp->castparam;
5982 ua->user_lksb = ua_tmp->user_lksb;
5984 error = set_unlock_args(flags, ua, &args);
5988 error = unlock_lock(ls, lkb, &args);
5990 if (error == -DLM_EUNLOCK)
5992 /* from validate_unlock_args() */
5993 if (error == -EBUSY && (flags & DLM_LKF_FORCEUNLOCK))
5998 spin_lock(&ua->proc->locks_spin);
5999 /* dlm_user_add_cb() may have already taken lkb off the proc list */
6000 if (!list_empty(&lkb->lkb_ownqueue))
6001 list_move(&lkb->lkb_ownqueue, &ua->proc->unlocking);
6002 spin_unlock(&ua->proc->locks_spin);
6006 dlm_unlock_recovery(ls);
6011 int dlm_user_cancel(struct dlm_ls *ls, struct dlm_user_args *ua_tmp,
6012 uint32_t flags, uint32_t lkid)
6014 struct dlm_lkb *lkb;
6015 struct dlm_args args;
6016 struct dlm_user_args *ua;
6019 dlm_lock_recovery(ls);
6021 error = find_lkb(ls, lkid, &lkb);
6026 if (ua_tmp->castparam)
6027 ua->castparam = ua_tmp->castparam;
6028 ua->user_lksb = ua_tmp->user_lksb;
6030 error = set_unlock_args(flags, ua, &args);
6034 error = cancel_lock(ls, lkb, &args);
6036 if (error == -DLM_ECANCEL)
6038 /* from validate_unlock_args() */
6039 if (error == -EBUSY)
6044 dlm_unlock_recovery(ls);
6049 int dlm_user_deadlock(struct dlm_ls *ls, uint32_t flags, uint32_t lkid)
6051 struct dlm_lkb *lkb;
6052 struct dlm_args args;
6053 struct dlm_user_args *ua;
6057 dlm_lock_recovery(ls);
6059 error = find_lkb(ls, lkid, &lkb);
6065 error = set_unlock_args(flags, ua, &args);
6069 /* same as cancel_lock(), but set DEADLOCK_CANCEL after lock_rsb */
6071 r = lkb->lkb_resource;
6075 error = validate_unlock_args(lkb, &args);
6078 lkb->lkb_flags |= DLM_IFL_DEADLOCK_CANCEL;
6080 error = _cancel_lock(r, lkb);
6085 if (error == -DLM_ECANCEL)
6087 /* from validate_unlock_args() */
6088 if (error == -EBUSY)
6093 dlm_unlock_recovery(ls);
6097 /* lkb's that are removed from the waiters list by revert are just left on the
6098 orphans list with the granted orphan locks, to be freed by purge */
6100 static int orphan_proc_lock(struct dlm_ls *ls, struct dlm_lkb *lkb)
6102 struct dlm_args args;
6105 hold_lkb(lkb); /* reference for the ls_orphans list */
6106 mutex_lock(&ls->ls_orphans_mutex);
6107 list_add_tail(&lkb->lkb_ownqueue, &ls->ls_orphans);
6108 mutex_unlock(&ls->ls_orphans_mutex);
6110 set_unlock_args(0, lkb->lkb_ua, &args);
6112 error = cancel_lock(ls, lkb, &args);
6113 if (error == -DLM_ECANCEL)
6118 /* The FORCEUNLOCK flag allows the unlock to go ahead even if the lkb isn't
6119 granted. Regardless of what rsb queue the lock is on, it's removed and
6120 freed. The IVVALBLK flag causes the lvb on the resource to be invalidated
6121 if our lock is PW/EX (it's ignored if our granted mode is smaller.) */
6123 static int unlock_proc_lock(struct dlm_ls *ls, struct dlm_lkb *lkb)
6125 struct dlm_args args;
6128 set_unlock_args(DLM_LKF_FORCEUNLOCK | DLM_LKF_IVVALBLK,
6129 lkb->lkb_ua, &args);
6131 error = unlock_lock(ls, lkb, &args);
6132 if (error == -DLM_EUNLOCK)
6137 /* We have to release clear_proc_locks mutex before calling unlock_proc_lock()
6138 (which does lock_rsb) due to deadlock with receiving a message that does
6139 lock_rsb followed by dlm_user_add_cb() */
6141 static struct dlm_lkb *del_proc_lock(struct dlm_ls *ls,
6142 struct dlm_user_proc *proc)
6144 struct dlm_lkb *lkb = NULL;
6146 mutex_lock(&ls->ls_clear_proc_locks);
6147 if (list_empty(&proc->locks))
6150 lkb = list_entry(proc->locks.next, struct dlm_lkb, lkb_ownqueue);
6151 list_del_init(&lkb->lkb_ownqueue);
6153 if (lkb->lkb_exflags & DLM_LKF_PERSISTENT)
6154 lkb->lkb_flags |= DLM_IFL_ORPHAN;
6156 lkb->lkb_flags |= DLM_IFL_DEAD;
6158 mutex_unlock(&ls->ls_clear_proc_locks);
6162 /* The ls_clear_proc_locks mutex protects against dlm_user_add_cb() which
6163 1) references lkb->ua which we free here and 2) adds lkbs to proc->asts,
6164 which we clear here. */
6166 /* proc CLOSING flag is set so no more device_reads should look at proc->asts
6167 list, and no more device_writes should add lkb's to proc->locks list; so we
6168 shouldn't need to take asts_spin or locks_spin here. this assumes that
6169 device reads/writes/closes are serialized -- FIXME: we may need to serialize
6172 void dlm_clear_proc_locks(struct dlm_ls *ls, struct dlm_user_proc *proc)
6174 struct dlm_lkb *lkb, *safe;
6176 dlm_lock_recovery(ls);
6179 lkb = del_proc_lock(ls, proc);
6183 if (lkb->lkb_exflags & DLM_LKF_PERSISTENT)
6184 orphan_proc_lock(ls, lkb);
6186 unlock_proc_lock(ls, lkb);
6188 /* this removes the reference for the proc->locks list
6189 added by dlm_user_request, it may result in the lkb
6195 mutex_lock(&ls->ls_clear_proc_locks);
6197 /* in-progress unlocks */
6198 list_for_each_entry_safe(lkb, safe, &proc->unlocking, lkb_ownqueue) {
6199 list_del_init(&lkb->lkb_ownqueue);
6200 lkb->lkb_flags |= DLM_IFL_DEAD;
6204 list_for_each_entry_safe(lkb, safe, &proc->asts, lkb_cb_list) {
6205 memset(&lkb->lkb_callbacks, 0,
6206 sizeof(struct dlm_callback) * DLM_CALLBACKS_SIZE);
6207 list_del_init(&lkb->lkb_cb_list);
6211 mutex_unlock(&ls->ls_clear_proc_locks);
6212 dlm_unlock_recovery(ls);
6215 static void purge_proc_locks(struct dlm_ls *ls, struct dlm_user_proc *proc)
6217 struct dlm_lkb *lkb, *safe;
6221 spin_lock(&proc->locks_spin);
6222 if (!list_empty(&proc->locks)) {
6223 lkb = list_entry(proc->locks.next, struct dlm_lkb,
6225 list_del_init(&lkb->lkb_ownqueue);
6227 spin_unlock(&proc->locks_spin);
6232 lkb->lkb_flags |= DLM_IFL_DEAD;
6233 unlock_proc_lock(ls, lkb);
6234 dlm_put_lkb(lkb); /* ref from proc->locks list */
6237 spin_lock(&proc->locks_spin);
6238 list_for_each_entry_safe(lkb, safe, &proc->unlocking, lkb_ownqueue) {
6239 list_del_init(&lkb->lkb_ownqueue);
6240 lkb->lkb_flags |= DLM_IFL_DEAD;
6243 spin_unlock(&proc->locks_spin);
6245 spin_lock(&proc->asts_spin);
6246 list_for_each_entry_safe(lkb, safe, &proc->asts, lkb_cb_list) {
6247 memset(&lkb->lkb_callbacks, 0,
6248 sizeof(struct dlm_callback) * DLM_CALLBACKS_SIZE);
6249 list_del_init(&lkb->lkb_cb_list);
6252 spin_unlock(&proc->asts_spin);
6255 /* pid of 0 means purge all orphans */
6257 static void do_purge(struct dlm_ls *ls, int nodeid, int pid)
6259 struct dlm_lkb *lkb, *safe;
6261 mutex_lock(&ls->ls_orphans_mutex);
6262 list_for_each_entry_safe(lkb, safe, &ls->ls_orphans, lkb_ownqueue) {
6263 if (pid && lkb->lkb_ownpid != pid)
6265 unlock_proc_lock(ls, lkb);
6266 list_del_init(&lkb->lkb_ownqueue);
6269 mutex_unlock(&ls->ls_orphans_mutex);
6272 static int send_purge(struct dlm_ls *ls, int nodeid, int pid)
6274 struct dlm_message *ms;
6275 struct dlm_mhandle *mh;
6278 error = _create_message(ls, sizeof(struct dlm_message), nodeid,
6279 DLM_MSG_PURGE, &ms, &mh);
6282 ms->m_nodeid = nodeid;
6285 return send_message(mh, ms);
6288 int dlm_user_purge(struct dlm_ls *ls, struct dlm_user_proc *proc,
6289 int nodeid, int pid)
6293 if (nodeid && (nodeid != dlm_our_nodeid())) {
6294 error = send_purge(ls, nodeid, pid);
6296 dlm_lock_recovery(ls);
6297 if (pid == current->pid)
6298 purge_proc_locks(ls, proc);
6300 do_purge(ls, nodeid, pid);
6301 dlm_unlock_recovery(ls);