1 // SPDX-License-Identifier: GPL-2.0-only
2 /******************************************************************************
3 *******************************************************************************
5 ** Copyright (C) 2005-2010 Red Hat, Inc. All rights reserved.
8 *******************************************************************************
9 ******************************************************************************/
11 /* Central locking logic has four stages:
31 Stage 1 (lock, unlock) is mainly about checking input args and
32 splitting into one of the four main operations:
34 dlm_lock = request_lock
35 dlm_lock+CONVERT = convert_lock
36 dlm_unlock = unlock_lock
37 dlm_unlock+CANCEL = cancel_lock
39 Stage 2, xxxx_lock(), just finds and locks the relevant rsb which is
40 provided to the next stage.
42 Stage 3, _xxxx_lock(), determines if the operation is local or remote.
43 When remote, it calls send_xxxx(), when local it calls do_xxxx().
45 Stage 4, do_xxxx(), is the guts of the operation. It manipulates the
46 given rsb and lkb and queues callbacks.
48 For remote operations, send_xxxx() results in the corresponding do_xxxx()
49 function being executed on the remote node. The connecting send/receive
50 calls on local (L) and remote (R) nodes:
52 L: send_xxxx() -> R: receive_xxxx()
54 L: receive_xxxx_reply() <- R: send_xxxx_reply()
56 #include <trace/events/dlm.h>
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 const struct dlm_message *ms, bool local);
90 static int receive_extralen(const struct dlm_message *ms);
91 static void do_purge(struct dlm_ls *ls, int nodeid, int pid);
92 static void toss_rsb(struct kref *kref);
95 * Lock compatibilty matrix - thanks Steve
96 * UN = Unlocked state. Not really a state, used as a flag
97 * PD = Padding. Used to make the matrix a nice power of two in size
98 * Other states are the same as the VMS DLM.
99 * Usage: matrix[grmode+1][rqmode+1] (although m[rq+1][gr+1] is the same)
102 static const int __dlm_compat_matrix[8][8] = {
103 /* UN NL CR CW PR PW EX PD */
104 {1, 1, 1, 1, 1, 1, 1, 0}, /* UN */
105 {1, 1, 1, 1, 1, 1, 1, 0}, /* NL */
106 {1, 1, 1, 1, 1, 1, 0, 0}, /* CR */
107 {1, 1, 1, 1, 0, 0, 0, 0}, /* CW */
108 {1, 1, 1, 0, 1, 0, 0, 0}, /* PR */
109 {1, 1, 1, 0, 0, 0, 0, 0}, /* PW */
110 {1, 1, 0, 0, 0, 0, 0, 0}, /* EX */
111 {0, 0, 0, 0, 0, 0, 0, 0} /* PD */
115 * This defines the direction of transfer of LVB data.
116 * Granted mode is the row; requested mode is the column.
117 * Usage: matrix[grmode+1][rqmode+1]
118 * 1 = LVB is returned to the caller
119 * 0 = LVB is written to the resource
120 * -1 = nothing happens to the LVB
123 const int dlm_lvb_operations[8][8] = {
124 /* UN NL CR CW PR PW EX PD*/
125 { -1, 1, 1, 1, 1, 1, 1, -1 }, /* UN */
126 { -1, 1, 1, 1, 1, 1, 1, 0 }, /* NL */
127 { -1, -1, 1, 1, 1, 1, 1, 0 }, /* CR */
128 { -1, -1, -1, 1, 1, 1, 1, 0 }, /* CW */
129 { -1, -1, -1, -1, 1, 1, 1, 0 }, /* PR */
130 { -1, 0, 0, 0, 0, 0, 1, 0 }, /* PW */
131 { -1, 0, 0, 0, 0, 0, 0, 0 }, /* EX */
132 { -1, 0, 0, 0, 0, 0, 0, 0 } /* PD */
135 #define modes_compat(gr, rq) \
136 __dlm_compat_matrix[(gr)->lkb_grmode + 1][(rq)->lkb_rqmode + 1]
138 int dlm_modes_compat(int mode1, int mode2)
140 return __dlm_compat_matrix[mode1 + 1][mode2 + 1];
144 * Compatibility matrix for conversions with QUECVT set.
145 * Granted mode is the row; requested mode is the column.
146 * Usage: matrix[grmode+1][rqmode+1]
149 static const int __quecvt_compat_matrix[8][8] = {
150 /* UN NL CR CW PR PW EX PD */
151 {0, 0, 0, 0, 0, 0, 0, 0}, /* UN */
152 {0, 0, 1, 1, 1, 1, 1, 0}, /* NL */
153 {0, 0, 0, 1, 1, 1, 1, 0}, /* CR */
154 {0, 0, 0, 0, 1, 1, 1, 0}, /* CW */
155 {0, 0, 0, 1, 0, 1, 1, 0}, /* PR */
156 {0, 0, 0, 0, 0, 0, 1, 0}, /* PW */
157 {0, 0, 0, 0, 0, 0, 0, 0}, /* EX */
158 {0, 0, 0, 0, 0, 0, 0, 0} /* PD */
161 void dlm_print_lkb(struct dlm_lkb *lkb)
163 printk(KERN_ERR "lkb: nodeid %d id %x remid %x exflags %x flags %x "
164 "sts %d rq %d gr %d wait_type %d wait_nodeid %d seq %llu\n",
165 lkb->lkb_nodeid, lkb->lkb_id, lkb->lkb_remid, lkb->lkb_exflags,
166 dlm_iflags_val(lkb), lkb->lkb_status, lkb->lkb_rqmode,
167 lkb->lkb_grmode, lkb->lkb_wait_type, lkb->lkb_wait_nodeid,
168 (unsigned long long)lkb->lkb_recover_seq);
171 static void dlm_print_rsb(struct dlm_rsb *r)
173 printk(KERN_ERR "rsb: nodeid %d master %d dir %d flags %lx first %x "
175 r->res_nodeid, r->res_master_nodeid, r->res_dir_nodeid,
176 r->res_flags, r->res_first_lkid, r->res_recover_locks_count,
180 void dlm_dump_rsb(struct dlm_rsb *r)
186 printk(KERN_ERR "rsb: root_list empty %d recover_list empty %d\n",
187 list_empty(&r->res_root_list), list_empty(&r->res_recover_list));
188 printk(KERN_ERR "rsb lookup list\n");
189 list_for_each_entry(lkb, &r->res_lookup, lkb_rsb_lookup)
191 printk(KERN_ERR "rsb grant queue:\n");
192 list_for_each_entry(lkb, &r->res_grantqueue, lkb_statequeue)
194 printk(KERN_ERR "rsb convert queue:\n");
195 list_for_each_entry(lkb, &r->res_convertqueue, lkb_statequeue)
197 printk(KERN_ERR "rsb wait queue:\n");
198 list_for_each_entry(lkb, &r->res_waitqueue, lkb_statequeue)
202 /* Threads cannot use the lockspace while it's being recovered */
204 static inline void dlm_lock_recovery(struct dlm_ls *ls)
206 down_read(&ls->ls_in_recovery);
209 void dlm_unlock_recovery(struct dlm_ls *ls)
211 up_read(&ls->ls_in_recovery);
214 int dlm_lock_recovery_try(struct dlm_ls *ls)
216 return down_read_trylock(&ls->ls_in_recovery);
219 static inline int can_be_queued(struct dlm_lkb *lkb)
221 return !(lkb->lkb_exflags & DLM_LKF_NOQUEUE);
224 static inline int force_blocking_asts(struct dlm_lkb *lkb)
226 return (lkb->lkb_exflags & DLM_LKF_NOQUEUEBAST);
229 static inline int is_demoted(struct dlm_lkb *lkb)
231 return test_bit(DLM_SBF_DEMOTED_BIT, &lkb->lkb_sbflags);
234 static inline int is_altmode(struct dlm_lkb *lkb)
236 return test_bit(DLM_SBF_ALTMODE_BIT, &lkb->lkb_sbflags);
239 static inline int is_granted(struct dlm_lkb *lkb)
241 return (lkb->lkb_status == DLM_LKSTS_GRANTED);
244 static inline int is_remote(struct dlm_rsb *r)
246 DLM_ASSERT(r->res_nodeid >= 0, dlm_print_rsb(r););
247 return !!r->res_nodeid;
250 static inline int is_process_copy(struct dlm_lkb *lkb)
252 return lkb->lkb_nodeid &&
253 !test_bit(DLM_IFL_MSTCPY_BIT, &lkb->lkb_iflags);
256 static inline int is_master_copy(struct dlm_lkb *lkb)
258 return test_bit(DLM_IFL_MSTCPY_BIT, &lkb->lkb_iflags);
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 test_bit(DLM_IFL_OVERLAP_UNLOCK_BIT, &lkb->lkb_iflags);
279 static inline int is_overlap_cancel(struct dlm_lkb *lkb)
281 return test_bit(DLM_IFL_OVERLAP_CANCEL_BIT, &lkb->lkb_iflags);
284 static inline int is_overlap(struct dlm_lkb *lkb)
286 return test_bit(DLM_IFL_OVERLAP_UNLOCK_BIT, &lkb->lkb_iflags) ||
287 test_bit(DLM_IFL_OVERLAP_CANCEL_BIT, &lkb->lkb_iflags);
290 static void queue_cast(struct dlm_rsb *r, struct dlm_lkb *lkb, int rv)
292 if (is_master_copy(lkb))
295 DLM_ASSERT(lkb->lkb_lksb, dlm_print_lkb(lkb););
297 if (rv == -DLM_ECANCEL &&
298 test_and_clear_bit(DLM_IFL_DEADLOCK_CANCEL_BIT, &lkb->lkb_iflags))
301 dlm_add_cb(lkb, DLM_CB_CAST, lkb->lkb_grmode, rv, dlm_sbflags_val(lkb));
304 static inline void queue_cast_overlap(struct dlm_rsb *r, struct dlm_lkb *lkb)
307 is_overlap_unlock(lkb) ? -DLM_EUNLOCK : -DLM_ECANCEL);
310 static void queue_bast(struct dlm_rsb *r, struct dlm_lkb *lkb, int rqmode)
312 if (is_master_copy(lkb)) {
313 send_bast(r, lkb, rqmode);
315 dlm_add_cb(lkb, DLM_CB_BAST, rqmode, 0, 0);
320 * Basic operations on rsb's and lkb's
323 /* This is only called to add a reference when the code already holds
324 a valid reference to the rsb, so there's no need for locking. */
326 static inline void hold_rsb(struct dlm_rsb *r)
328 kref_get(&r->res_ref);
331 void dlm_hold_rsb(struct dlm_rsb *r)
336 /* When all references to the rsb are gone it's transferred to
337 the tossed list for later disposal. */
339 static void put_rsb(struct dlm_rsb *r)
341 struct dlm_ls *ls = r->res_ls;
342 uint32_t bucket = r->res_bucket;
345 rv = kref_put_lock(&r->res_ref, toss_rsb,
346 &ls->ls_rsbtbl[bucket].lock);
348 spin_unlock(&ls->ls_rsbtbl[bucket].lock);
351 void dlm_put_rsb(struct dlm_rsb *r)
356 static int pre_rsb_struct(struct dlm_ls *ls)
358 struct dlm_rsb *r1, *r2;
361 spin_lock(&ls->ls_new_rsb_spin);
362 if (ls->ls_new_rsb_count > dlm_config.ci_new_rsb_count / 2) {
363 spin_unlock(&ls->ls_new_rsb_spin);
366 spin_unlock(&ls->ls_new_rsb_spin);
368 r1 = dlm_allocate_rsb(ls);
369 r2 = dlm_allocate_rsb(ls);
371 spin_lock(&ls->ls_new_rsb_spin);
373 list_add(&r1->res_hashchain, &ls->ls_new_rsb);
374 ls->ls_new_rsb_count++;
377 list_add(&r2->res_hashchain, &ls->ls_new_rsb);
378 ls->ls_new_rsb_count++;
380 count = ls->ls_new_rsb_count;
381 spin_unlock(&ls->ls_new_rsb_spin);
388 /* If ls->ls_new_rsb is empty, return -EAGAIN, so the caller can
389 unlock any spinlocks, go back and call pre_rsb_struct again.
390 Otherwise, take an rsb off the list and return it. */
392 static int get_rsb_struct(struct dlm_ls *ls, const void *name, int len,
393 struct dlm_rsb **r_ret)
398 spin_lock(&ls->ls_new_rsb_spin);
399 if (list_empty(&ls->ls_new_rsb)) {
400 count = ls->ls_new_rsb_count;
401 spin_unlock(&ls->ls_new_rsb_spin);
402 log_debug(ls, "find_rsb retry %d %d %s",
403 count, dlm_config.ci_new_rsb_count,
408 r = list_first_entry(&ls->ls_new_rsb, struct dlm_rsb, res_hashchain);
409 list_del(&r->res_hashchain);
410 /* Convert the empty list_head to a NULL rb_node for tree usage: */
411 memset(&r->res_hashnode, 0, sizeof(struct rb_node));
412 ls->ls_new_rsb_count--;
413 spin_unlock(&ls->ls_new_rsb_spin);
417 memcpy(r->res_name, name, len);
418 mutex_init(&r->res_mutex);
420 INIT_LIST_HEAD(&r->res_lookup);
421 INIT_LIST_HEAD(&r->res_grantqueue);
422 INIT_LIST_HEAD(&r->res_convertqueue);
423 INIT_LIST_HEAD(&r->res_waitqueue);
424 INIT_LIST_HEAD(&r->res_root_list);
425 INIT_LIST_HEAD(&r->res_recover_list);
431 static int rsb_cmp(struct dlm_rsb *r, const char *name, int nlen)
433 char maxname[DLM_RESNAME_MAXLEN];
435 memset(maxname, 0, DLM_RESNAME_MAXLEN);
436 memcpy(maxname, name, nlen);
437 return memcmp(r->res_name, maxname, DLM_RESNAME_MAXLEN);
440 int dlm_search_rsb_tree(struct rb_root *tree, const void *name, int len,
441 struct dlm_rsb **r_ret)
443 struct rb_node *node = tree->rb_node;
448 r = rb_entry(node, struct dlm_rsb, res_hashnode);
449 rc = rsb_cmp(r, name, len);
451 node = node->rb_left;
453 node = node->rb_right;
465 static int rsb_insert(struct dlm_rsb *rsb, struct rb_root *tree)
467 struct rb_node **newn = &tree->rb_node;
468 struct rb_node *parent = NULL;
472 struct dlm_rsb *cur = rb_entry(*newn, struct dlm_rsb,
476 rc = rsb_cmp(cur, rsb->res_name, rsb->res_length);
478 newn = &parent->rb_left;
480 newn = &parent->rb_right;
482 log_print("rsb_insert match");
489 rb_link_node(&rsb->res_hashnode, parent, newn);
490 rb_insert_color(&rsb->res_hashnode, tree);
495 * Find rsb in rsbtbl and potentially create/add one
497 * Delaying the release of rsb's has a similar benefit to applications keeping
498 * NL locks on an rsb, but without the guarantee that the cached master value
499 * will still be valid when the rsb is reused. Apps aren't always smart enough
500 * to keep NL locks on an rsb that they may lock again shortly; this can lead
501 * to excessive master lookups and removals if we don't delay the release.
503 * Searching for an rsb means looking through both the normal list and toss
504 * list. When found on the toss list the rsb is moved to the normal list with
505 * ref count of 1; when found on normal list the ref count is incremented.
507 * rsb's on the keep list are being used locally and refcounted.
508 * rsb's on the toss list are not being used locally, and are not refcounted.
510 * The toss list rsb's were either
511 * - previously used locally but not any more (were on keep list, then
512 * moved to toss list when last refcount dropped)
513 * - created and put on toss list as a directory record for a lookup
514 * (we are the dir node for the res, but are not using the res right now,
515 * but some other node is)
517 * The purpose of find_rsb() is to return a refcounted rsb for local use.
518 * So, if the given rsb is on the toss list, it is moved to the keep list
519 * before being returned.
521 * toss_rsb() happens when all local usage of the rsb is done, i.e. no
522 * more refcounts exist, so the rsb is moved from the keep list to the
525 * rsb's on both keep and toss lists are used for doing a name to master
526 * lookups. rsb's that are in use locally (and being refcounted) are on
527 * the keep list, rsb's that are not in use locally (not refcounted) and
528 * only exist for name/master lookups are on the toss list.
530 * rsb's on the toss list who's dir_nodeid is not local can have stale
531 * name/master mappings. So, remote requests on such rsb's can potentially
532 * return with an error, which means the mapping is stale and needs to
533 * be updated with a new lookup. (The idea behind MASTER UNCERTAIN and
534 * first_lkid is to keep only a single outstanding request on an rsb
535 * while that rsb has a potentially stale master.)
538 static int find_rsb_dir(struct dlm_ls *ls, const void *name, int len,
539 uint32_t hash, uint32_t b,
540 int dir_nodeid, int from_nodeid,
541 unsigned int flags, struct dlm_rsb **r_ret)
543 struct dlm_rsb *r = NULL;
544 int our_nodeid = dlm_our_nodeid();
551 if (flags & R_RECEIVE_REQUEST) {
552 if (from_nodeid == dir_nodeid)
556 } else if (flags & R_REQUEST) {
561 * flags & R_RECEIVE_RECOVER is from dlm_recover_master_copy, so
562 * from_nodeid has sent us a lock in dlm_recover_locks, believing
563 * we're the new master. Our local recovery may not have set
564 * res_master_nodeid to our_nodeid yet, so allow either. Don't
565 * create the rsb; dlm_recover_process_copy() will handle EBADR
568 * If someone sends us a request, we are the dir node, and we do
569 * not find the rsb anywhere, then recreate it. This happens if
570 * someone sends us a request after we have removed/freed an rsb
571 * from our toss list. (They sent a request instead of lookup
572 * because they are using an rsb from their toss list.)
575 if (from_local || from_dir ||
576 (from_other && (dir_nodeid == our_nodeid))) {
582 error = pre_rsb_struct(ls);
587 spin_lock(&ls->ls_rsbtbl[b].lock);
589 error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].keep, name, len, &r);
594 * rsb is active, so we can't check master_nodeid without lock_rsb.
597 kref_get(&r->res_ref);
602 error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].toss, name, len, &r);
607 * rsb found inactive (master_nodeid may be out of date unless
608 * we are the dir_nodeid or were the master) No other thread
609 * is using this rsb because it's on the toss list, so we can
610 * look at or update res_master_nodeid without lock_rsb.
613 if ((r->res_master_nodeid != our_nodeid) && from_other) {
614 /* our rsb was not master, and another node (not the dir node)
615 has sent us a request */
616 log_debug(ls, "find_rsb toss from_other %d master %d dir %d %s",
617 from_nodeid, r->res_master_nodeid, dir_nodeid,
623 if ((r->res_master_nodeid != our_nodeid) && from_dir) {
624 /* don't think this should ever happen */
625 log_error(ls, "find_rsb toss from_dir %d master %d",
626 from_nodeid, r->res_master_nodeid);
628 /* fix it and go on */
629 r->res_master_nodeid = our_nodeid;
631 rsb_clear_flag(r, RSB_MASTER_UNCERTAIN);
632 r->res_first_lkid = 0;
635 if (from_local && (r->res_master_nodeid != our_nodeid)) {
636 /* Because we have held no locks on this rsb,
637 res_master_nodeid could have become stale. */
638 rsb_set_flag(r, RSB_MASTER_UNCERTAIN);
639 r->res_first_lkid = 0;
642 rb_erase(&r->res_hashnode, &ls->ls_rsbtbl[b].toss);
643 error = rsb_insert(r, &ls->ls_rsbtbl[b].keep);
652 if (error == -EBADR && !create)
655 error = get_rsb_struct(ls, name, len, &r);
656 if (error == -EAGAIN) {
657 spin_unlock(&ls->ls_rsbtbl[b].lock);
665 r->res_dir_nodeid = dir_nodeid;
666 kref_init(&r->res_ref);
669 /* want to see how often this happens */
670 log_debug(ls, "find_rsb new from_dir %d recreate %s",
671 from_nodeid, r->res_name);
672 r->res_master_nodeid = our_nodeid;
677 if (from_other && (dir_nodeid != our_nodeid)) {
678 /* should never happen */
679 log_error(ls, "find_rsb new from_other %d dir %d our %d %s",
680 from_nodeid, dir_nodeid, our_nodeid, r->res_name);
688 log_debug(ls, "find_rsb new from_other %d dir %d %s",
689 from_nodeid, dir_nodeid, r->res_name);
692 if (dir_nodeid == our_nodeid) {
693 /* When we are the dir nodeid, we can set the master
695 r->res_master_nodeid = our_nodeid;
698 /* set_master will send_lookup to dir_nodeid */
699 r->res_master_nodeid = 0;
704 error = rsb_insert(r, &ls->ls_rsbtbl[b].keep);
706 spin_unlock(&ls->ls_rsbtbl[b].lock);
712 /* During recovery, other nodes can send us new MSTCPY locks (from
713 dlm_recover_locks) before we've made ourself master (in
714 dlm_recover_masters). */
716 static int find_rsb_nodir(struct dlm_ls *ls, const void *name, int len,
717 uint32_t hash, uint32_t b,
718 int dir_nodeid, int from_nodeid,
719 unsigned int flags, struct dlm_rsb **r_ret)
721 struct dlm_rsb *r = NULL;
722 int our_nodeid = dlm_our_nodeid();
723 int recover = (flags & R_RECEIVE_RECOVER);
727 error = pre_rsb_struct(ls);
731 spin_lock(&ls->ls_rsbtbl[b].lock);
733 error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].keep, name, len, &r);
738 * rsb is active, so we can't check master_nodeid without lock_rsb.
741 kref_get(&r->res_ref);
746 error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].toss, name, len, &r);
751 * rsb found inactive. No other thread is using this rsb because
752 * it's on the toss list, so we can look at or update
753 * res_master_nodeid without lock_rsb.
756 if (!recover && (r->res_master_nodeid != our_nodeid) && from_nodeid) {
757 /* our rsb is not master, and another node has sent us a
758 request; this should never happen */
759 log_error(ls, "find_rsb toss from_nodeid %d master %d dir %d",
760 from_nodeid, r->res_master_nodeid, dir_nodeid);
766 if (!recover && (r->res_master_nodeid != our_nodeid) &&
767 (dir_nodeid == our_nodeid)) {
768 /* our rsb is not master, and we are dir; may as well fix it;
769 this should never happen */
770 log_error(ls, "find_rsb toss our %d master %d dir %d",
771 our_nodeid, r->res_master_nodeid, dir_nodeid);
773 r->res_master_nodeid = our_nodeid;
777 rb_erase(&r->res_hashnode, &ls->ls_rsbtbl[b].toss);
778 error = rsb_insert(r, &ls->ls_rsbtbl[b].keep);
787 error = get_rsb_struct(ls, name, len, &r);
788 if (error == -EAGAIN) {
789 spin_unlock(&ls->ls_rsbtbl[b].lock);
797 r->res_dir_nodeid = dir_nodeid;
798 r->res_master_nodeid = dir_nodeid;
799 r->res_nodeid = (dir_nodeid == our_nodeid) ? 0 : dir_nodeid;
800 kref_init(&r->res_ref);
802 error = rsb_insert(r, &ls->ls_rsbtbl[b].keep);
804 spin_unlock(&ls->ls_rsbtbl[b].lock);
810 static int find_rsb(struct dlm_ls *ls, const void *name, int len,
811 int from_nodeid, unsigned int flags,
812 struct dlm_rsb **r_ret)
817 if (len > DLM_RESNAME_MAXLEN)
820 hash = jhash(name, len, 0);
821 b = hash & (ls->ls_rsbtbl_size - 1);
823 dir_nodeid = dlm_hash2nodeid(ls, hash);
825 if (dlm_no_directory(ls))
826 return find_rsb_nodir(ls, name, len, hash, b, dir_nodeid,
827 from_nodeid, flags, r_ret);
829 return find_rsb_dir(ls, name, len, hash, b, dir_nodeid,
830 from_nodeid, flags, r_ret);
833 /* we have received a request and found that res_master_nodeid != our_nodeid,
834 so we need to return an error or make ourself the master */
836 static int validate_master_nodeid(struct dlm_ls *ls, struct dlm_rsb *r,
839 if (dlm_no_directory(ls)) {
840 log_error(ls, "find_rsb keep from_nodeid %d master %d dir %d",
841 from_nodeid, r->res_master_nodeid,
847 if (from_nodeid != r->res_dir_nodeid) {
848 /* our rsb is not master, and another node (not the dir node)
849 has sent us a request. this is much more common when our
850 master_nodeid is zero, so limit debug to non-zero. */
852 if (r->res_master_nodeid) {
853 log_debug(ls, "validate master from_other %d master %d "
854 "dir %d first %x %s", from_nodeid,
855 r->res_master_nodeid, r->res_dir_nodeid,
856 r->res_first_lkid, r->res_name);
860 /* our rsb is not master, but the dir nodeid has sent us a
861 request; this could happen with master 0 / res_nodeid -1 */
863 if (r->res_master_nodeid) {
864 log_error(ls, "validate master from_dir %d master %d "
866 from_nodeid, r->res_master_nodeid,
867 r->res_first_lkid, r->res_name);
870 r->res_master_nodeid = dlm_our_nodeid();
876 static void __dlm_master_lookup(struct dlm_ls *ls, struct dlm_rsb *r, int our_nodeid,
877 int from_nodeid, bool toss_list, unsigned int flags,
878 int *r_nodeid, int *result)
880 int fix_master = (flags & DLM_LU_RECOVER_MASTER);
881 int from_master = (flags & DLM_LU_RECOVER_DIR);
883 if (r->res_dir_nodeid != our_nodeid) {
884 /* should not happen, but may as well fix it and carry on */
885 log_error(ls, "%s res_dir %d our %d %s", __func__,
886 r->res_dir_nodeid, our_nodeid, r->res_name);
887 r->res_dir_nodeid = our_nodeid;
890 if (fix_master && dlm_is_removed(ls, r->res_master_nodeid)) {
891 /* Recovery uses this function to set a new master when
892 * the previous master failed. Setting NEW_MASTER will
893 * force dlm_recover_masters to call recover_master on this
894 * rsb even though the res_nodeid is no longer removed.
897 r->res_master_nodeid = from_nodeid;
898 r->res_nodeid = from_nodeid;
899 rsb_set_flag(r, RSB_NEW_MASTER);
902 /* I don't think we should ever find it on toss list. */
903 log_error(ls, "%s fix_master on toss", __func__);
908 if (from_master && (r->res_master_nodeid != from_nodeid)) {
909 /* this will happen if from_nodeid became master during
910 * a previous recovery cycle, and we aborted the previous
911 * cycle before recovering this master value
914 log_limit(ls, "%s from_master %d master_nodeid %d res_nodeid %d first %x %s",
915 __func__, from_nodeid, r->res_master_nodeid,
916 r->res_nodeid, r->res_first_lkid, r->res_name);
918 if (r->res_master_nodeid == our_nodeid) {
919 log_error(ls, "from_master %d our_master", from_nodeid);
924 r->res_master_nodeid = from_nodeid;
925 r->res_nodeid = from_nodeid;
926 rsb_set_flag(r, RSB_NEW_MASTER);
929 if (!r->res_master_nodeid) {
930 /* this will happen if recovery happens while we're looking
931 * up the master for this rsb
934 log_debug(ls, "%s master 0 to %d first %x %s", __func__,
935 from_nodeid, r->res_first_lkid, r->res_name);
936 r->res_master_nodeid = from_nodeid;
937 r->res_nodeid = from_nodeid;
940 if (!from_master && !fix_master &&
941 (r->res_master_nodeid == from_nodeid)) {
942 /* this can happen when the master sends remove, the dir node
943 * finds the rsb on the keep list and ignores the remove,
944 * and the former master sends a lookup
947 log_limit(ls, "%s from master %d flags %x first %x %s",
948 __func__, from_nodeid, flags, r->res_first_lkid,
953 *r_nodeid = r->res_master_nodeid;
955 *result = DLM_LU_MATCH;
959 * We're the dir node for this res and another node wants to know the
960 * master nodeid. During normal operation (non recovery) this is only
961 * called from receive_lookup(); master lookups when the local node is
962 * the dir node are done by find_rsb().
964 * normal operation, we are the dir node for a resource
969 * . dlm_master_lookup flags 0
971 * recover directory, we are rebuilding dir for all resources
972 * . dlm_recover_directory
974 * remote node sends back the rsb names it is master of and we are dir of
975 * . dlm_master_lookup RECOVER_DIR (fix_master 0, from_master 1)
976 * we either create new rsb setting remote node as master, or find existing
977 * rsb and set master to be the remote node.
979 * recover masters, we are finding the new master for resources
980 * . dlm_recover_masters
982 * . dlm_send_rcom_lookup
983 * . receive_rcom_lookup
984 * . dlm_master_lookup RECOVER_MASTER (fix_master 1, from_master 0)
987 int dlm_master_lookup(struct dlm_ls *ls, int from_nodeid, const char *name,
988 int len, unsigned int flags, int *r_nodeid, int *result)
990 struct dlm_rsb *r = NULL;
992 int our_nodeid = dlm_our_nodeid();
993 int dir_nodeid, error;
995 if (len > DLM_RESNAME_MAXLEN)
998 if (from_nodeid == our_nodeid) {
999 log_error(ls, "dlm_master_lookup from our_nodeid %d flags %x",
1004 hash = jhash(name, len, 0);
1005 b = hash & (ls->ls_rsbtbl_size - 1);
1007 dir_nodeid = dlm_hash2nodeid(ls, hash);
1008 if (dir_nodeid != our_nodeid) {
1009 log_error(ls, "dlm_master_lookup from %d dir %d our %d h %x %d",
1010 from_nodeid, dir_nodeid, our_nodeid, hash,
1017 error = pre_rsb_struct(ls);
1021 spin_lock(&ls->ls_rsbtbl[b].lock);
1022 error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].keep, name, len, &r);
1024 /* because the rsb is active, we need to lock_rsb before
1025 * checking/changing re_master_nodeid
1029 spin_unlock(&ls->ls_rsbtbl[b].lock);
1032 __dlm_master_lookup(ls, r, our_nodeid, from_nodeid, false,
1033 flags, r_nodeid, result);
1035 /* the rsb was active */
1042 error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].toss, name, len, &r);
1046 /* because the rsb is inactive (on toss list), it's not refcounted
1047 * and lock_rsb is not used, but is protected by the rsbtbl lock
1050 __dlm_master_lookup(ls, r, our_nodeid, from_nodeid, true, flags,
1053 r->res_toss_time = jiffies;
1054 /* the rsb was inactive (on toss list) */
1055 spin_unlock(&ls->ls_rsbtbl[b].lock);
1060 error = get_rsb_struct(ls, name, len, &r);
1061 if (error == -EAGAIN) {
1062 spin_unlock(&ls->ls_rsbtbl[b].lock);
1070 r->res_dir_nodeid = our_nodeid;
1071 r->res_master_nodeid = from_nodeid;
1072 r->res_nodeid = from_nodeid;
1073 kref_init(&r->res_ref);
1074 r->res_toss_time = jiffies;
1076 error = rsb_insert(r, &ls->ls_rsbtbl[b].toss);
1078 /* should never happen */
1080 spin_unlock(&ls->ls_rsbtbl[b].lock);
1085 *result = DLM_LU_ADD;
1086 *r_nodeid = from_nodeid;
1088 spin_unlock(&ls->ls_rsbtbl[b].lock);
1092 static void dlm_dump_rsb_hash(struct dlm_ls *ls, uint32_t hash)
1098 for (i = 0; i < ls->ls_rsbtbl_size; i++) {
1099 spin_lock(&ls->ls_rsbtbl[i].lock);
1100 for (n = rb_first(&ls->ls_rsbtbl[i].keep); n; n = rb_next(n)) {
1101 r = rb_entry(n, struct dlm_rsb, res_hashnode);
1102 if (r->res_hash == hash)
1105 spin_unlock(&ls->ls_rsbtbl[i].lock);
1109 void dlm_dump_rsb_name(struct dlm_ls *ls, const char *name, int len)
1111 struct dlm_rsb *r = NULL;
1115 hash = jhash(name, len, 0);
1116 b = hash & (ls->ls_rsbtbl_size - 1);
1118 spin_lock(&ls->ls_rsbtbl[b].lock);
1119 error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].keep, name, len, &r);
1123 error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].toss, name, len, &r);
1129 spin_unlock(&ls->ls_rsbtbl[b].lock);
1132 static void toss_rsb(struct kref *kref)
1134 struct dlm_rsb *r = container_of(kref, struct dlm_rsb, res_ref);
1135 struct dlm_ls *ls = r->res_ls;
1137 DLM_ASSERT(list_empty(&r->res_root_list), dlm_print_rsb(r););
1138 kref_init(&r->res_ref);
1139 rb_erase(&r->res_hashnode, &ls->ls_rsbtbl[r->res_bucket].keep);
1140 rsb_insert(r, &ls->ls_rsbtbl[r->res_bucket].toss);
1141 r->res_toss_time = jiffies;
1142 set_bit(DLM_RTF_SHRINK_BIT, &ls->ls_rsbtbl[r->res_bucket].flags);
1143 if (r->res_lvbptr) {
1144 dlm_free_lvb(r->res_lvbptr);
1145 r->res_lvbptr = NULL;
1149 /* See comment for unhold_lkb */
1151 static void unhold_rsb(struct dlm_rsb *r)
1154 rv = kref_put(&r->res_ref, toss_rsb);
1155 DLM_ASSERT(!rv, dlm_dump_rsb(r););
1158 static void kill_rsb(struct kref *kref)
1160 struct dlm_rsb *r = container_of(kref, struct dlm_rsb, res_ref);
1162 /* All work is done after the return from kref_put() so we
1163 can release the write_lock before the remove and free. */
1165 DLM_ASSERT(list_empty(&r->res_lookup), dlm_dump_rsb(r););
1166 DLM_ASSERT(list_empty(&r->res_grantqueue), dlm_dump_rsb(r););
1167 DLM_ASSERT(list_empty(&r->res_convertqueue), dlm_dump_rsb(r););
1168 DLM_ASSERT(list_empty(&r->res_waitqueue), dlm_dump_rsb(r););
1169 DLM_ASSERT(list_empty(&r->res_root_list), dlm_dump_rsb(r););
1170 DLM_ASSERT(list_empty(&r->res_recover_list), dlm_dump_rsb(r););
1173 /* Attaching/detaching lkb's from rsb's is for rsb reference counting.
1174 The rsb must exist as long as any lkb's for it do. */
1176 static void attach_lkb(struct dlm_rsb *r, struct dlm_lkb *lkb)
1179 lkb->lkb_resource = r;
1182 static void detach_lkb(struct dlm_lkb *lkb)
1184 if (lkb->lkb_resource) {
1185 put_rsb(lkb->lkb_resource);
1186 lkb->lkb_resource = NULL;
1190 static int _create_lkb(struct dlm_ls *ls, struct dlm_lkb **lkb_ret,
1193 struct dlm_lkb *lkb;
1196 lkb = dlm_allocate_lkb(ls);
1200 lkb->lkb_last_bast_mode = -1;
1201 lkb->lkb_nodeid = -1;
1202 lkb->lkb_grmode = DLM_LOCK_IV;
1203 kref_init(&lkb->lkb_ref);
1204 INIT_LIST_HEAD(&lkb->lkb_ownqueue);
1205 INIT_LIST_HEAD(&lkb->lkb_rsb_lookup);
1206 INIT_LIST_HEAD(&lkb->lkb_cb_list);
1207 INIT_LIST_HEAD(&lkb->lkb_callbacks);
1208 spin_lock_init(&lkb->lkb_cb_lock);
1209 INIT_WORK(&lkb->lkb_cb_work, dlm_callback_work);
1211 idr_preload(GFP_NOFS);
1212 spin_lock(&ls->ls_lkbidr_spin);
1213 rv = idr_alloc(&ls->ls_lkbidr, lkb, start, end, GFP_NOWAIT);
1216 spin_unlock(&ls->ls_lkbidr_spin);
1220 log_error(ls, "create_lkb idr error %d", rv);
1229 static int create_lkb(struct dlm_ls *ls, struct dlm_lkb **lkb_ret)
1231 return _create_lkb(ls, lkb_ret, 1, 0);
1234 static int find_lkb(struct dlm_ls *ls, uint32_t lkid, struct dlm_lkb **lkb_ret)
1236 struct dlm_lkb *lkb;
1238 spin_lock(&ls->ls_lkbidr_spin);
1239 lkb = idr_find(&ls->ls_lkbidr, lkid);
1241 kref_get(&lkb->lkb_ref);
1242 spin_unlock(&ls->ls_lkbidr_spin);
1245 return lkb ? 0 : -ENOENT;
1248 static void kill_lkb(struct kref *kref)
1250 struct dlm_lkb *lkb = container_of(kref, struct dlm_lkb, lkb_ref);
1252 /* All work is done after the return from kref_put() so we
1253 can release the write_lock before the detach_lkb */
1255 DLM_ASSERT(!lkb->lkb_status, dlm_print_lkb(lkb););
1258 /* __put_lkb() is used when an lkb may not have an rsb attached to
1259 it so we need to provide the lockspace explicitly */
1261 static int __put_lkb(struct dlm_ls *ls, struct dlm_lkb *lkb)
1263 uint32_t lkid = lkb->lkb_id;
1266 rv = kref_put_lock(&lkb->lkb_ref, kill_lkb,
1267 &ls->ls_lkbidr_spin);
1269 idr_remove(&ls->ls_lkbidr, lkid);
1270 spin_unlock(&ls->ls_lkbidr_spin);
1274 /* for local/process lkbs, lvbptr points to caller's lksb */
1275 if (lkb->lkb_lvbptr && is_master_copy(lkb))
1276 dlm_free_lvb(lkb->lkb_lvbptr);
1283 int dlm_put_lkb(struct dlm_lkb *lkb)
1287 DLM_ASSERT(lkb->lkb_resource, dlm_print_lkb(lkb););
1288 DLM_ASSERT(lkb->lkb_resource->res_ls, dlm_print_lkb(lkb););
1290 ls = lkb->lkb_resource->res_ls;
1291 return __put_lkb(ls, lkb);
1294 /* This is only called to add a reference when the code already holds
1295 a valid reference to the lkb, so there's no need for locking. */
1297 static inline void hold_lkb(struct dlm_lkb *lkb)
1299 kref_get(&lkb->lkb_ref);
1302 static void unhold_lkb_assert(struct kref *kref)
1304 struct dlm_lkb *lkb = container_of(kref, struct dlm_lkb, lkb_ref);
1306 DLM_ASSERT(false, dlm_print_lkb(lkb););
1309 /* This is called when we need to remove a reference and are certain
1310 it's not the last ref. e.g. del_lkb is always called between a
1311 find_lkb/put_lkb and is always the inverse of a previous add_lkb.
1312 put_lkb would work fine, but would involve unnecessary locking */
1314 static inline void unhold_lkb(struct dlm_lkb *lkb)
1316 kref_put(&lkb->lkb_ref, unhold_lkb_assert);
1319 static void lkb_add_ordered(struct list_head *new, struct list_head *head,
1322 struct dlm_lkb *lkb = NULL, *iter;
1324 list_for_each_entry(iter, head, lkb_statequeue)
1325 if (iter->lkb_rqmode < mode) {
1327 list_add_tail(new, &iter->lkb_statequeue);
1332 list_add_tail(new, head);
1335 /* add/remove lkb to rsb's grant/convert/wait queue */
1337 static void add_lkb(struct dlm_rsb *r, struct dlm_lkb *lkb, int status)
1339 kref_get(&lkb->lkb_ref);
1341 DLM_ASSERT(!lkb->lkb_status, dlm_print_lkb(lkb););
1343 lkb->lkb_timestamp = ktime_get();
1345 lkb->lkb_status = status;
1348 case DLM_LKSTS_WAITING:
1349 if (lkb->lkb_exflags & DLM_LKF_HEADQUE)
1350 list_add(&lkb->lkb_statequeue, &r->res_waitqueue);
1352 list_add_tail(&lkb->lkb_statequeue, &r->res_waitqueue);
1354 case DLM_LKSTS_GRANTED:
1355 /* convention says granted locks kept in order of grmode */
1356 lkb_add_ordered(&lkb->lkb_statequeue, &r->res_grantqueue,
1359 case DLM_LKSTS_CONVERT:
1360 if (lkb->lkb_exflags & DLM_LKF_HEADQUE)
1361 list_add(&lkb->lkb_statequeue, &r->res_convertqueue);
1363 list_add_tail(&lkb->lkb_statequeue,
1364 &r->res_convertqueue);
1367 DLM_ASSERT(0, dlm_print_lkb(lkb); printk("sts=%d\n", status););
1371 static void del_lkb(struct dlm_rsb *r, struct dlm_lkb *lkb)
1373 lkb->lkb_status = 0;
1374 list_del(&lkb->lkb_statequeue);
1378 static void move_lkb(struct dlm_rsb *r, struct dlm_lkb *lkb, int sts)
1382 add_lkb(r, lkb, sts);
1386 static int msg_reply_type(int mstype)
1389 case DLM_MSG_REQUEST:
1390 return DLM_MSG_REQUEST_REPLY;
1391 case DLM_MSG_CONVERT:
1392 return DLM_MSG_CONVERT_REPLY;
1393 case DLM_MSG_UNLOCK:
1394 return DLM_MSG_UNLOCK_REPLY;
1395 case DLM_MSG_CANCEL:
1396 return DLM_MSG_CANCEL_REPLY;
1397 case DLM_MSG_LOOKUP:
1398 return DLM_MSG_LOOKUP_REPLY;
1403 /* add/remove lkb from global waiters list of lkb's waiting for
1404 a reply from a remote node */
1406 static int add_to_waiters(struct dlm_lkb *lkb, int mstype, int to_nodeid)
1408 struct dlm_ls *ls = lkb->lkb_resource->res_ls;
1411 mutex_lock(&ls->ls_waiters_mutex);
1413 if (is_overlap_unlock(lkb) ||
1414 (is_overlap_cancel(lkb) && (mstype == DLM_MSG_CANCEL))) {
1419 if (lkb->lkb_wait_type || is_overlap_cancel(lkb)) {
1421 case DLM_MSG_UNLOCK:
1422 set_bit(DLM_IFL_OVERLAP_UNLOCK_BIT, &lkb->lkb_iflags);
1424 case DLM_MSG_CANCEL:
1425 set_bit(DLM_IFL_OVERLAP_CANCEL_BIT, &lkb->lkb_iflags);
1431 lkb->lkb_wait_count++;
1434 log_debug(ls, "addwait %x cur %d overlap %d count %d f %x",
1435 lkb->lkb_id, lkb->lkb_wait_type, mstype,
1436 lkb->lkb_wait_count, dlm_iflags_val(lkb));
1440 DLM_ASSERT(!lkb->lkb_wait_count,
1442 printk("wait_count %d\n", lkb->lkb_wait_count););
1444 lkb->lkb_wait_count++;
1445 lkb->lkb_wait_type = mstype;
1446 lkb->lkb_wait_nodeid = to_nodeid; /* for debugging */
1448 list_add(&lkb->lkb_wait_reply, &ls->ls_waiters);
1451 log_error(ls, "addwait error %x %d flags %x %d %d %s",
1452 lkb->lkb_id, error, dlm_iflags_val(lkb), mstype,
1453 lkb->lkb_wait_type, lkb->lkb_resource->res_name);
1454 mutex_unlock(&ls->ls_waiters_mutex);
1458 /* We clear the RESEND flag because we might be taking an lkb off the waiters
1459 list as part of process_requestqueue (e.g. a lookup that has an optimized
1460 request reply on the requestqueue) between dlm_recover_waiters_pre() which
1461 set RESEND and dlm_recover_waiters_post() */
1463 static int _remove_from_waiters(struct dlm_lkb *lkb, int mstype,
1464 const struct dlm_message *ms)
1466 struct dlm_ls *ls = lkb->lkb_resource->res_ls;
1467 int overlap_done = 0;
1469 if (mstype == DLM_MSG_UNLOCK_REPLY &&
1470 test_and_clear_bit(DLM_IFL_OVERLAP_UNLOCK_BIT, &lkb->lkb_iflags)) {
1471 log_debug(ls, "remwait %x unlock_reply overlap", lkb->lkb_id);
1476 if (mstype == DLM_MSG_CANCEL_REPLY &&
1477 test_and_clear_bit(DLM_IFL_OVERLAP_CANCEL_BIT, &lkb->lkb_iflags)) {
1478 log_debug(ls, "remwait %x cancel_reply overlap", lkb->lkb_id);
1483 /* Cancel state was preemptively cleared by a successful convert,
1484 see next comment, nothing to do. */
1486 if ((mstype == DLM_MSG_CANCEL_REPLY) &&
1487 (lkb->lkb_wait_type != DLM_MSG_CANCEL)) {
1488 log_debug(ls, "remwait %x cancel_reply wait_type %d",
1489 lkb->lkb_id, lkb->lkb_wait_type);
1493 /* Remove for the convert reply, and premptively remove for the
1494 cancel reply. A convert has been granted while there's still
1495 an outstanding cancel on it (the cancel is moot and the result
1496 in the cancel reply should be 0). We preempt the cancel reply
1497 because the app gets the convert result and then can follow up
1498 with another op, like convert. This subsequent op would see the
1499 lingering state of the cancel and fail with -EBUSY. */
1501 if ((mstype == DLM_MSG_CONVERT_REPLY) &&
1502 (lkb->lkb_wait_type == DLM_MSG_CONVERT) && ms && !ms->m_result &&
1503 test_and_clear_bit(DLM_IFL_OVERLAP_CANCEL_BIT, &lkb->lkb_iflags)) {
1504 log_debug(ls, "remwait %x convert_reply zap overlap_cancel",
1506 lkb->lkb_wait_type = 0;
1507 lkb->lkb_wait_count--;
1512 /* N.B. type of reply may not always correspond to type of original
1513 msg due to lookup->request optimization, verify others? */
1515 if (lkb->lkb_wait_type) {
1516 lkb->lkb_wait_type = 0;
1520 log_error(ls, "remwait error %x remote %d %x msg %d flags %x no wait",
1521 lkb->lkb_id, ms ? le32_to_cpu(ms->m_header.h_nodeid) : 0,
1522 lkb->lkb_remid, mstype, dlm_iflags_val(lkb));
1526 /* the force-unlock/cancel has completed and we haven't recvd a reply
1527 to the op that was in progress prior to the unlock/cancel; we
1528 give up on any reply to the earlier op. FIXME: not sure when/how
1529 this would happen */
1531 if (overlap_done && lkb->lkb_wait_type) {
1532 log_error(ls, "remwait error %x reply %d wait_type %d overlap",
1533 lkb->lkb_id, mstype, lkb->lkb_wait_type);
1534 lkb->lkb_wait_count--;
1536 lkb->lkb_wait_type = 0;
1539 DLM_ASSERT(lkb->lkb_wait_count, dlm_print_lkb(lkb););
1541 clear_bit(DLM_IFL_RESEND_BIT, &lkb->lkb_iflags);
1542 lkb->lkb_wait_count--;
1543 if (!lkb->lkb_wait_count)
1544 list_del_init(&lkb->lkb_wait_reply);
1549 static int remove_from_waiters(struct dlm_lkb *lkb, int mstype)
1551 struct dlm_ls *ls = lkb->lkb_resource->res_ls;
1554 mutex_lock(&ls->ls_waiters_mutex);
1555 error = _remove_from_waiters(lkb, mstype, NULL);
1556 mutex_unlock(&ls->ls_waiters_mutex);
1560 /* Handles situations where we might be processing a "fake" or "local" reply in
1561 which we can't try to take waiters_mutex again. */
1563 static int remove_from_waiters_ms(struct dlm_lkb *lkb,
1564 const struct dlm_message *ms, bool local)
1566 struct dlm_ls *ls = lkb->lkb_resource->res_ls;
1570 mutex_lock(&ls->ls_waiters_mutex);
1571 error = _remove_from_waiters(lkb, le32_to_cpu(ms->m_type), ms);
1573 mutex_unlock(&ls->ls_waiters_mutex);
1577 static void shrink_bucket(struct dlm_ls *ls, int b)
1579 struct rb_node *n, *next;
1582 int our_nodeid = dlm_our_nodeid();
1583 int remote_count = 0;
1584 int need_shrink = 0;
1587 memset(&ls->ls_remove_lens, 0, sizeof(int) * DLM_REMOVE_NAMES_MAX);
1589 spin_lock(&ls->ls_rsbtbl[b].lock);
1591 if (!test_bit(DLM_RTF_SHRINK_BIT, &ls->ls_rsbtbl[b].flags)) {
1592 spin_unlock(&ls->ls_rsbtbl[b].lock);
1596 for (n = rb_first(&ls->ls_rsbtbl[b].toss); n; n = next) {
1598 r = rb_entry(n, struct dlm_rsb, res_hashnode);
1600 /* If we're the directory record for this rsb, and
1601 we're not the master of it, then we need to wait
1602 for the master node to send us a dir remove for
1603 before removing the dir record. */
1605 if (!dlm_no_directory(ls) &&
1606 (r->res_master_nodeid != our_nodeid) &&
1607 (dlm_dir_nodeid(r) == our_nodeid)) {
1613 if (!time_after_eq(jiffies, r->res_toss_time +
1614 dlm_config.ci_toss_secs * HZ)) {
1618 if (!dlm_no_directory(ls) &&
1619 (r->res_master_nodeid == our_nodeid) &&
1620 (dlm_dir_nodeid(r) != our_nodeid)) {
1622 /* We're the master of this rsb but we're not
1623 the directory record, so we need to tell the
1624 dir node to remove the dir record. */
1626 ls->ls_remove_lens[remote_count] = r->res_length;
1627 memcpy(ls->ls_remove_names[remote_count], r->res_name,
1628 DLM_RESNAME_MAXLEN);
1631 if (remote_count >= DLM_REMOVE_NAMES_MAX)
1636 if (!kref_put(&r->res_ref, kill_rsb)) {
1637 log_error(ls, "tossed rsb in use %s", r->res_name);
1641 rb_erase(&r->res_hashnode, &ls->ls_rsbtbl[b].toss);
1646 set_bit(DLM_RTF_SHRINK_BIT, &ls->ls_rsbtbl[b].flags);
1648 clear_bit(DLM_RTF_SHRINK_BIT, &ls->ls_rsbtbl[b].flags);
1649 spin_unlock(&ls->ls_rsbtbl[b].lock);
1652 * While searching for rsb's to free, we found some that require
1653 * remote removal. We leave them in place and find them again here
1654 * so there is a very small gap between removing them from the toss
1655 * list and sending the removal. Keeping this gap small is
1656 * important to keep us (the master node) from being out of sync
1657 * with the remote dir node for very long.
1660 for (i = 0; i < remote_count; i++) {
1661 name = ls->ls_remove_names[i];
1662 len = ls->ls_remove_lens[i];
1664 spin_lock(&ls->ls_rsbtbl[b].lock);
1665 rv = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].toss, name, len, &r);
1667 spin_unlock(&ls->ls_rsbtbl[b].lock);
1668 log_debug(ls, "remove_name not toss %s", name);
1672 if (r->res_master_nodeid != our_nodeid) {
1673 spin_unlock(&ls->ls_rsbtbl[b].lock);
1674 log_debug(ls, "remove_name master %d dir %d our %d %s",
1675 r->res_master_nodeid, r->res_dir_nodeid,
1680 if (r->res_dir_nodeid == our_nodeid) {
1681 /* should never happen */
1682 spin_unlock(&ls->ls_rsbtbl[b].lock);
1683 log_error(ls, "remove_name dir %d master %d our %d %s",
1684 r->res_dir_nodeid, r->res_master_nodeid,
1689 if (!time_after_eq(jiffies, r->res_toss_time +
1690 dlm_config.ci_toss_secs * HZ)) {
1691 spin_unlock(&ls->ls_rsbtbl[b].lock);
1692 log_debug(ls, "remove_name toss_time %lu now %lu %s",
1693 r->res_toss_time, jiffies, name);
1697 if (!kref_put(&r->res_ref, kill_rsb)) {
1698 spin_unlock(&ls->ls_rsbtbl[b].lock);
1699 log_error(ls, "remove_name in use %s", name);
1703 rb_erase(&r->res_hashnode, &ls->ls_rsbtbl[b].toss);
1705 spin_unlock(&ls->ls_rsbtbl[b].lock);
1711 void dlm_scan_rsbs(struct dlm_ls *ls)
1715 for (i = 0; i < ls->ls_rsbtbl_size; i++) {
1716 shrink_bucket(ls, i);
1717 if (dlm_locking_stopped(ls))
1723 /* lkb is master or local copy */
1725 static void set_lvb_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
1727 int b, len = r->res_ls->ls_lvblen;
1729 /* b=1 lvb returned to caller
1730 b=0 lvb written to rsb or invalidated
1733 b = dlm_lvb_operations[lkb->lkb_grmode + 1][lkb->lkb_rqmode + 1];
1736 if (!lkb->lkb_lvbptr)
1739 if (!(lkb->lkb_exflags & DLM_LKF_VALBLK))
1745 memcpy(lkb->lkb_lvbptr, r->res_lvbptr, len);
1746 lkb->lkb_lvbseq = r->res_lvbseq;
1748 } else if (b == 0) {
1749 if (lkb->lkb_exflags & DLM_LKF_IVVALBLK) {
1750 rsb_set_flag(r, RSB_VALNOTVALID);
1754 if (!lkb->lkb_lvbptr)
1757 if (!(lkb->lkb_exflags & DLM_LKF_VALBLK))
1761 r->res_lvbptr = dlm_allocate_lvb(r->res_ls);
1766 memcpy(r->res_lvbptr, lkb->lkb_lvbptr, len);
1768 lkb->lkb_lvbseq = r->res_lvbseq;
1769 rsb_clear_flag(r, RSB_VALNOTVALID);
1772 if (rsb_flag(r, RSB_VALNOTVALID))
1773 set_bit(DLM_SBF_VALNOTVALID_BIT, &lkb->lkb_sbflags);
1776 static void set_lvb_unlock(struct dlm_rsb *r, struct dlm_lkb *lkb)
1778 if (lkb->lkb_grmode < DLM_LOCK_PW)
1781 if (lkb->lkb_exflags & DLM_LKF_IVVALBLK) {
1782 rsb_set_flag(r, RSB_VALNOTVALID);
1786 if (!lkb->lkb_lvbptr)
1789 if (!(lkb->lkb_exflags & DLM_LKF_VALBLK))
1793 r->res_lvbptr = dlm_allocate_lvb(r->res_ls);
1798 memcpy(r->res_lvbptr, lkb->lkb_lvbptr, r->res_ls->ls_lvblen);
1800 rsb_clear_flag(r, RSB_VALNOTVALID);
1803 /* lkb is process copy (pc) */
1805 static void set_lvb_lock_pc(struct dlm_rsb *r, struct dlm_lkb *lkb,
1806 const struct dlm_message *ms)
1810 if (!lkb->lkb_lvbptr)
1813 if (!(lkb->lkb_exflags & DLM_LKF_VALBLK))
1816 b = dlm_lvb_operations[lkb->lkb_grmode + 1][lkb->lkb_rqmode + 1];
1818 int len = receive_extralen(ms);
1819 if (len > r->res_ls->ls_lvblen)
1820 len = r->res_ls->ls_lvblen;
1821 memcpy(lkb->lkb_lvbptr, ms->m_extra, len);
1822 lkb->lkb_lvbseq = le32_to_cpu(ms->m_lvbseq);
1826 /* Manipulate lkb's on rsb's convert/granted/waiting queues
1827 remove_lock -- used for unlock, removes lkb from granted
1828 revert_lock -- used for cancel, moves lkb from convert to granted
1829 grant_lock -- used for request and convert, adds lkb to granted or
1830 moves lkb from convert or waiting to granted
1832 Each of these is used for master or local copy lkb's. There is
1833 also a _pc() variation used to make the corresponding change on
1834 a process copy (pc) lkb. */
1836 static void _remove_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
1839 lkb->lkb_grmode = DLM_LOCK_IV;
1840 /* this unhold undoes the original ref from create_lkb()
1841 so this leads to the lkb being freed */
1845 static void remove_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
1847 set_lvb_unlock(r, lkb);
1848 _remove_lock(r, lkb);
1851 static void remove_lock_pc(struct dlm_rsb *r, struct dlm_lkb *lkb)
1853 _remove_lock(r, lkb);
1856 /* returns: 0 did nothing
1857 1 moved lock to granted
1860 static int revert_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
1864 lkb->lkb_rqmode = DLM_LOCK_IV;
1866 switch (lkb->lkb_status) {
1867 case DLM_LKSTS_GRANTED:
1869 case DLM_LKSTS_CONVERT:
1870 move_lkb(r, lkb, DLM_LKSTS_GRANTED);
1873 case DLM_LKSTS_WAITING:
1875 lkb->lkb_grmode = DLM_LOCK_IV;
1876 /* this unhold undoes the original ref from create_lkb()
1877 so this leads to the lkb being freed */
1882 log_print("invalid status for revert %d", lkb->lkb_status);
1887 static int revert_lock_pc(struct dlm_rsb *r, struct dlm_lkb *lkb)
1889 return revert_lock(r, lkb);
1892 static void _grant_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
1894 if (lkb->lkb_grmode != lkb->lkb_rqmode) {
1895 lkb->lkb_grmode = lkb->lkb_rqmode;
1896 if (lkb->lkb_status)
1897 move_lkb(r, lkb, DLM_LKSTS_GRANTED);
1899 add_lkb(r, lkb, DLM_LKSTS_GRANTED);
1902 lkb->lkb_rqmode = DLM_LOCK_IV;
1903 lkb->lkb_highbast = 0;
1906 static void grant_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
1908 set_lvb_lock(r, lkb);
1909 _grant_lock(r, lkb);
1912 static void grant_lock_pc(struct dlm_rsb *r, struct dlm_lkb *lkb,
1913 const struct dlm_message *ms)
1915 set_lvb_lock_pc(r, lkb, ms);
1916 _grant_lock(r, lkb);
1919 /* called by grant_pending_locks() which means an async grant message must
1920 be sent to the requesting node in addition to granting the lock if the
1921 lkb belongs to a remote node. */
1923 static void grant_lock_pending(struct dlm_rsb *r, struct dlm_lkb *lkb)
1926 if (is_master_copy(lkb))
1929 queue_cast(r, lkb, 0);
1932 /* The special CONVDEADLK, ALTPR and ALTCW flags allow the master to
1933 change the granted/requested modes. We're munging things accordingly in
1935 CONVDEADLK: our grmode may have been forced down to NL to resolve a
1937 ALTPR/ALTCW: our rqmode may have been changed to PR or CW to become
1938 compatible with other granted locks */
1940 static void munge_demoted(struct dlm_lkb *lkb)
1942 if (lkb->lkb_rqmode == DLM_LOCK_IV || lkb->lkb_grmode == DLM_LOCK_IV) {
1943 log_print("munge_demoted %x invalid modes gr %d rq %d",
1944 lkb->lkb_id, lkb->lkb_grmode, lkb->lkb_rqmode);
1948 lkb->lkb_grmode = DLM_LOCK_NL;
1951 static void munge_altmode(struct dlm_lkb *lkb, const struct dlm_message *ms)
1953 if (ms->m_type != cpu_to_le32(DLM_MSG_REQUEST_REPLY) &&
1954 ms->m_type != cpu_to_le32(DLM_MSG_GRANT)) {
1955 log_print("munge_altmode %x invalid reply type %d",
1956 lkb->lkb_id, le32_to_cpu(ms->m_type));
1960 if (lkb->lkb_exflags & DLM_LKF_ALTPR)
1961 lkb->lkb_rqmode = DLM_LOCK_PR;
1962 else if (lkb->lkb_exflags & DLM_LKF_ALTCW)
1963 lkb->lkb_rqmode = DLM_LOCK_CW;
1965 log_print("munge_altmode invalid exflags %x", lkb->lkb_exflags);
1970 static inline int first_in_list(struct dlm_lkb *lkb, struct list_head *head)
1972 struct dlm_lkb *first = list_entry(head->next, struct dlm_lkb,
1974 if (lkb->lkb_id == first->lkb_id)
1980 /* Check if the given lkb conflicts with another lkb on the queue. */
1982 static int queue_conflict(struct list_head *head, struct dlm_lkb *lkb)
1984 struct dlm_lkb *this;
1986 list_for_each_entry(this, head, lkb_statequeue) {
1989 if (!modes_compat(this, lkb))
1996 * "A conversion deadlock arises with a pair of lock requests in the converting
1997 * queue for one resource. The granted mode of each lock blocks the requested
1998 * mode of the other lock."
2000 * Part 2: if the granted mode of lkb is preventing an earlier lkb in the
2001 * convert queue from being granted, then deadlk/demote lkb.
2004 * Granted Queue: empty
2005 * Convert Queue: NL->EX (first lock)
2006 * PR->EX (second lock)
2008 * The first lock can't be granted because of the granted mode of the second
2009 * lock and the second lock can't be granted because it's not first in the
2010 * list. We either cancel lkb's conversion (PR->EX) and return EDEADLK, or we
2011 * demote the granted mode of lkb (from PR to NL) if it has the CONVDEADLK
2012 * flag set and return DEMOTED in the lksb flags.
2014 * Originally, this function detected conv-deadlk in a more limited scope:
2015 * - if !modes_compat(lkb1, lkb2) && !modes_compat(lkb2, lkb1), or
2016 * - if lkb1 was the first entry in the queue (not just earlier), and was
2017 * blocked by the granted mode of lkb2, and there was nothing on the
2018 * granted queue preventing lkb1 from being granted immediately, i.e.
2019 * lkb2 was the only thing preventing lkb1 from being granted.
2021 * That second condition meant we'd only say there was conv-deadlk if
2022 * resolving it (by demotion) would lead to the first lock on the convert
2023 * queue being granted right away. It allowed conversion deadlocks to exist
2024 * between locks on the convert queue while they couldn't be granted anyway.
2026 * Now, we detect and take action on conversion deadlocks immediately when
2027 * they're created, even if they may not be immediately consequential. If
2028 * lkb1 exists anywhere in the convert queue and lkb2 comes in with a granted
2029 * mode that would prevent lkb1's conversion from being granted, we do a
2030 * deadlk/demote on lkb2 right away and don't let it onto the convert queue.
2031 * I think this means that the lkb_is_ahead condition below should always
2032 * be zero, i.e. there will never be conv-deadlk between two locks that are
2033 * both already on the convert queue.
2036 static int conversion_deadlock_detect(struct dlm_rsb *r, struct dlm_lkb *lkb2)
2038 struct dlm_lkb *lkb1;
2039 int lkb_is_ahead = 0;
2041 list_for_each_entry(lkb1, &r->res_convertqueue, lkb_statequeue) {
2047 if (!lkb_is_ahead) {
2048 if (!modes_compat(lkb2, lkb1))
2051 if (!modes_compat(lkb2, lkb1) &&
2052 !modes_compat(lkb1, lkb2))
2060 * Return 1 if the lock can be granted, 0 otherwise.
2061 * Also detect and resolve conversion deadlocks.
2063 * lkb is the lock to be granted
2065 * now is 1 if the function is being called in the context of the
2066 * immediate request, it is 0 if called later, after the lock has been
2069 * recover is 1 if dlm_recover_grant() is trying to grant conversions
2072 * References are from chapter 6 of "VAXcluster Principles" by Roy Davis
2075 static int _can_be_granted(struct dlm_rsb *r, struct dlm_lkb *lkb, int now,
2078 int8_t conv = (lkb->lkb_grmode != DLM_LOCK_IV);
2081 * 6-10: Version 5.4 introduced an option to address the phenomenon of
2082 * a new request for a NL mode lock being blocked.
2084 * 6-11: If the optional EXPEDITE flag is used with the new NL mode
2085 * request, then it would be granted. In essence, the use of this flag
2086 * tells the Lock Manager to expedite theis request by not considering
2087 * what may be in the CONVERTING or WAITING queues... As of this
2088 * writing, the EXPEDITE flag can be used only with new requests for NL
2089 * mode locks. This flag is not valid for conversion requests.
2091 * A shortcut. Earlier checks return an error if EXPEDITE is used in a
2092 * conversion or used with a non-NL requested mode. We also know an
2093 * EXPEDITE request is always granted immediately, so now must always
2094 * be 1. The full condition to grant an expedite request: (now &&
2095 * !conv && lkb->rqmode == DLM_LOCK_NL && (flags & EXPEDITE)) can
2096 * therefore be shortened to just checking the flag.
2099 if (lkb->lkb_exflags & DLM_LKF_EXPEDITE)
2103 * A shortcut. Without this, !queue_conflict(grantqueue, lkb) would be
2104 * added to the remaining conditions.
2107 if (queue_conflict(&r->res_grantqueue, lkb))
2111 * 6-3: By default, a conversion request is immediately granted if the
2112 * requested mode is compatible with the modes of all other granted
2116 if (queue_conflict(&r->res_convertqueue, lkb))
2120 * The RECOVER_GRANT flag means dlm_recover_grant() is granting
2121 * locks for a recovered rsb, on which lkb's have been rebuilt.
2122 * The lkb's may have been rebuilt on the queues in a different
2123 * order than they were in on the previous master. So, granting
2124 * queued conversions in order after recovery doesn't make sense
2125 * since the order hasn't been preserved anyway. The new order
2126 * could also have created a new "in place" conversion deadlock.
2127 * (e.g. old, failed master held granted EX, with PR->EX, NL->EX.
2128 * After recovery, there would be no granted locks, and possibly
2129 * NL->EX, PR->EX, an in-place conversion deadlock.) So, after
2130 * recovery, grant conversions without considering order.
2133 if (conv && recover)
2137 * 6-5: But the default algorithm for deciding whether to grant or
2138 * queue conversion requests does not by itself guarantee that such
2139 * requests are serviced on a "first come first serve" basis. This, in
2140 * turn, can lead to a phenomenon known as "indefinate postponement".
2142 * 6-7: This issue is dealt with by using the optional QUECVT flag with
2143 * the system service employed to request a lock conversion. This flag
2144 * forces certain conversion requests to be queued, even if they are
2145 * compatible with the granted modes of other locks on the same
2146 * resource. Thus, the use of this flag results in conversion requests
2147 * being ordered on a "first come first servce" basis.
2149 * DCT: This condition is all about new conversions being able to occur
2150 * "in place" while the lock remains on the granted queue (assuming
2151 * nothing else conflicts.) IOW if QUECVT isn't set, a conversion
2152 * doesn't _have_ to go onto the convert queue where it's processed in
2153 * order. The "now" variable is necessary to distinguish converts
2154 * being received and processed for the first time now, because once a
2155 * convert is moved to the conversion queue the condition below applies
2156 * requiring fifo granting.
2159 if (now && conv && !(lkb->lkb_exflags & DLM_LKF_QUECVT))
2163 * Even if the convert is compat with all granted locks,
2164 * QUECVT forces it behind other locks on the convert queue.
2167 if (now && conv && (lkb->lkb_exflags & DLM_LKF_QUECVT)) {
2168 if (list_empty(&r->res_convertqueue))
2175 * The NOORDER flag is set to avoid the standard vms rules on grant
2179 if (lkb->lkb_exflags & DLM_LKF_NOORDER)
2183 * 6-3: Once in that queue [CONVERTING], a conversion request cannot be
2184 * granted until all other conversion requests ahead of it are granted
2188 if (!now && conv && first_in_list(lkb, &r->res_convertqueue))
2192 * 6-4: By default, a new request is immediately granted only if all
2193 * three of the following conditions are satisfied when the request is
2195 * - The queue of ungranted conversion requests for the resource is
2197 * - The queue of ungranted new requests for the resource is empty.
2198 * - The mode of the new request is compatible with the most
2199 * restrictive mode of all granted locks on the resource.
2202 if (now && !conv && list_empty(&r->res_convertqueue) &&
2203 list_empty(&r->res_waitqueue))
2207 * 6-4: Once a lock request is in the queue of ungranted new requests,
2208 * it cannot be granted until the queue of ungranted conversion
2209 * requests is empty, all ungranted new requests ahead of it are
2210 * granted and/or canceled, and it is compatible with the granted mode
2211 * of the most restrictive lock granted on the resource.
2214 if (!now && !conv && list_empty(&r->res_convertqueue) &&
2215 first_in_list(lkb, &r->res_waitqueue))
2221 static int can_be_granted(struct dlm_rsb *r, struct dlm_lkb *lkb, int now,
2222 int recover, int *err)
2225 int8_t alt = 0, rqmode = lkb->lkb_rqmode;
2226 int8_t is_convert = (lkb->lkb_grmode != DLM_LOCK_IV);
2231 rv = _can_be_granted(r, lkb, now, recover);
2236 * The CONVDEADLK flag is non-standard and tells the dlm to resolve
2237 * conversion deadlocks by demoting grmode to NL, otherwise the dlm
2238 * cancels one of the locks.
2241 if (is_convert && can_be_queued(lkb) &&
2242 conversion_deadlock_detect(r, lkb)) {
2243 if (lkb->lkb_exflags & DLM_LKF_CONVDEADLK) {
2244 lkb->lkb_grmode = DLM_LOCK_NL;
2245 set_bit(DLM_SBF_DEMOTED_BIT, &lkb->lkb_sbflags);
2249 log_print("can_be_granted deadlock %x now %d",
2257 * The ALTPR and ALTCW flags are non-standard and tell the dlm to try
2258 * to grant a request in a mode other than the normal rqmode. It's a
2259 * simple way to provide a big optimization to applications that can
2263 if (rqmode != DLM_LOCK_PR && (lkb->lkb_exflags & DLM_LKF_ALTPR))
2265 else if (rqmode != DLM_LOCK_CW && (lkb->lkb_exflags & DLM_LKF_ALTCW))
2269 lkb->lkb_rqmode = alt;
2270 rv = _can_be_granted(r, lkb, now, 0);
2272 set_bit(DLM_SBF_ALTMODE_BIT, &lkb->lkb_sbflags);
2274 lkb->lkb_rqmode = rqmode;
2280 /* Returns the highest requested mode of all blocked conversions; sets
2281 cw if there's a blocked conversion to DLM_LOCK_CW. */
2283 static int grant_pending_convert(struct dlm_rsb *r, int high, int *cw,
2284 unsigned int *count)
2286 struct dlm_lkb *lkb, *s;
2287 int recover = rsb_flag(r, RSB_RECOVER_GRANT);
2288 int hi, demoted, quit, grant_restart, demote_restart;
2297 list_for_each_entry_safe(lkb, s, &r->res_convertqueue, lkb_statequeue) {
2298 demoted = is_demoted(lkb);
2301 if (can_be_granted(r, lkb, 0, recover, &deadlk)) {
2302 grant_lock_pending(r, lkb);
2309 if (!demoted && is_demoted(lkb)) {
2310 log_print("WARN: pending demoted %x node %d %s",
2311 lkb->lkb_id, lkb->lkb_nodeid, r->res_name);
2318 * If DLM_LKB_NODLKWT flag is set and conversion
2319 * deadlock is detected, we request blocking AST and
2320 * down (or cancel) conversion.
2322 if (lkb->lkb_exflags & DLM_LKF_NODLCKWT) {
2323 if (lkb->lkb_highbast < lkb->lkb_rqmode) {
2324 queue_bast(r, lkb, lkb->lkb_rqmode);
2325 lkb->lkb_highbast = lkb->lkb_rqmode;
2328 log_print("WARN: pending deadlock %x node %d %s",
2329 lkb->lkb_id, lkb->lkb_nodeid,
2336 hi = max_t(int, lkb->lkb_rqmode, hi);
2338 if (cw && lkb->lkb_rqmode == DLM_LOCK_CW)
2344 if (demote_restart && !quit) {
2349 return max_t(int, high, hi);
2352 static int grant_pending_wait(struct dlm_rsb *r, int high, int *cw,
2353 unsigned int *count)
2355 struct dlm_lkb *lkb, *s;
2357 list_for_each_entry_safe(lkb, s, &r->res_waitqueue, lkb_statequeue) {
2358 if (can_be_granted(r, lkb, 0, 0, NULL)) {
2359 grant_lock_pending(r, lkb);
2363 high = max_t(int, lkb->lkb_rqmode, high);
2364 if (lkb->lkb_rqmode == DLM_LOCK_CW)
2372 /* cw of 1 means there's a lock with a rqmode of DLM_LOCK_CW that's blocked
2373 on either the convert or waiting queue.
2374 high is the largest rqmode of all locks blocked on the convert or
2377 static int lock_requires_bast(struct dlm_lkb *gr, int high, int cw)
2379 if (gr->lkb_grmode == DLM_LOCK_PR && cw) {
2380 if (gr->lkb_highbast < DLM_LOCK_EX)
2385 if (gr->lkb_highbast < high &&
2386 !__dlm_compat_matrix[gr->lkb_grmode+1][high+1])
2391 static void grant_pending_locks(struct dlm_rsb *r, unsigned int *count)
2393 struct dlm_lkb *lkb, *s;
2394 int high = DLM_LOCK_IV;
2397 if (!is_master(r)) {
2398 log_print("grant_pending_locks r nodeid %d", r->res_nodeid);
2403 high = grant_pending_convert(r, high, &cw, count);
2404 high = grant_pending_wait(r, high, &cw, count);
2406 if (high == DLM_LOCK_IV)
2410 * If there are locks left on the wait/convert queue then send blocking
2411 * ASTs to granted locks based on the largest requested mode (high)
2415 list_for_each_entry_safe(lkb, s, &r->res_grantqueue, lkb_statequeue) {
2416 if (lkb->lkb_bastfn && lock_requires_bast(lkb, high, cw)) {
2417 if (cw && high == DLM_LOCK_PR &&
2418 lkb->lkb_grmode == DLM_LOCK_PR)
2419 queue_bast(r, lkb, DLM_LOCK_CW);
2421 queue_bast(r, lkb, high);
2422 lkb->lkb_highbast = high;
2427 static int modes_require_bast(struct dlm_lkb *gr, struct dlm_lkb *rq)
2429 if ((gr->lkb_grmode == DLM_LOCK_PR && rq->lkb_rqmode == DLM_LOCK_CW) ||
2430 (gr->lkb_grmode == DLM_LOCK_CW && rq->lkb_rqmode == DLM_LOCK_PR)) {
2431 if (gr->lkb_highbast < DLM_LOCK_EX)
2436 if (gr->lkb_highbast < rq->lkb_rqmode && !modes_compat(gr, rq))
2441 static void send_bast_queue(struct dlm_rsb *r, struct list_head *head,
2442 struct dlm_lkb *lkb)
2446 list_for_each_entry(gr, head, lkb_statequeue) {
2447 /* skip self when sending basts to convertqueue */
2450 if (gr->lkb_bastfn && modes_require_bast(gr, lkb)) {
2451 queue_bast(r, gr, lkb->lkb_rqmode);
2452 gr->lkb_highbast = lkb->lkb_rqmode;
2457 static void send_blocking_asts(struct dlm_rsb *r, struct dlm_lkb *lkb)
2459 send_bast_queue(r, &r->res_grantqueue, lkb);
2462 static void send_blocking_asts_all(struct dlm_rsb *r, struct dlm_lkb *lkb)
2464 send_bast_queue(r, &r->res_grantqueue, lkb);
2465 send_bast_queue(r, &r->res_convertqueue, lkb);
2468 /* set_master(r, lkb) -- set the master nodeid of a resource
2470 The purpose of this function is to set the nodeid field in the given
2471 lkb using the nodeid field in the given rsb. If the rsb's nodeid is
2472 known, it can just be copied to the lkb and the function will return
2473 0. If the rsb's nodeid is _not_ known, it needs to be looked up
2474 before it can be copied to the lkb.
2476 When the rsb nodeid is being looked up remotely, the initial lkb
2477 causing the lookup is kept on the ls_waiters list waiting for the
2478 lookup reply. Other lkb's waiting for the same rsb lookup are kept
2479 on the rsb's res_lookup list until the master is verified.
2482 0: nodeid is set in rsb/lkb and the caller should go ahead and use it
2483 1: the rsb master is not available and the lkb has been placed on
2487 static int set_master(struct dlm_rsb *r, struct dlm_lkb *lkb)
2489 int our_nodeid = dlm_our_nodeid();
2491 if (rsb_flag(r, RSB_MASTER_UNCERTAIN)) {
2492 rsb_clear_flag(r, RSB_MASTER_UNCERTAIN);
2493 r->res_first_lkid = lkb->lkb_id;
2494 lkb->lkb_nodeid = r->res_nodeid;
2498 if (r->res_first_lkid && r->res_first_lkid != lkb->lkb_id) {
2499 list_add_tail(&lkb->lkb_rsb_lookup, &r->res_lookup);
2503 if (r->res_master_nodeid == our_nodeid) {
2504 lkb->lkb_nodeid = 0;
2508 if (r->res_master_nodeid) {
2509 lkb->lkb_nodeid = r->res_master_nodeid;
2513 if (dlm_dir_nodeid(r) == our_nodeid) {
2514 /* This is a somewhat unusual case; find_rsb will usually
2515 have set res_master_nodeid when dir nodeid is local, but
2516 there are cases where we become the dir node after we've
2517 past find_rsb and go through _request_lock again.
2518 confirm_master() or process_lookup_list() needs to be
2519 called after this. */
2520 log_debug(r->res_ls, "set_master %x self master %d dir %d %s",
2521 lkb->lkb_id, r->res_master_nodeid, r->res_dir_nodeid,
2523 r->res_master_nodeid = our_nodeid;
2525 lkb->lkb_nodeid = 0;
2529 r->res_first_lkid = lkb->lkb_id;
2530 send_lookup(r, lkb);
2534 static void process_lookup_list(struct dlm_rsb *r)
2536 struct dlm_lkb *lkb, *safe;
2538 list_for_each_entry_safe(lkb, safe, &r->res_lookup, lkb_rsb_lookup) {
2539 list_del_init(&lkb->lkb_rsb_lookup);
2540 _request_lock(r, lkb);
2545 /* confirm_master -- confirm (or deny) an rsb's master nodeid */
2547 static void confirm_master(struct dlm_rsb *r, int error)
2549 struct dlm_lkb *lkb;
2551 if (!r->res_first_lkid)
2557 r->res_first_lkid = 0;
2558 process_lookup_list(r);
2564 /* the remote request failed and won't be retried (it was
2565 a NOQUEUE, or has been canceled/unlocked); make a waiting
2566 lkb the first_lkid */
2568 r->res_first_lkid = 0;
2570 if (!list_empty(&r->res_lookup)) {
2571 lkb = list_entry(r->res_lookup.next, struct dlm_lkb,
2573 list_del_init(&lkb->lkb_rsb_lookup);
2574 r->res_first_lkid = lkb->lkb_id;
2575 _request_lock(r, lkb);
2580 log_error(r->res_ls, "confirm_master unknown error %d", error);
2584 static int set_lock_args(int mode, struct dlm_lksb *lksb, uint32_t flags,
2585 int namelen, void (*ast)(void *astparam),
2587 void (*bast)(void *astparam, int mode),
2588 struct dlm_args *args)
2592 /* check for invalid arg usage */
2594 if (mode < 0 || mode > DLM_LOCK_EX)
2597 if (!(flags & DLM_LKF_CONVERT) && (namelen > DLM_RESNAME_MAXLEN))
2600 if (flags & DLM_LKF_CANCEL)
2603 if (flags & DLM_LKF_QUECVT && !(flags & DLM_LKF_CONVERT))
2606 if (flags & DLM_LKF_CONVDEADLK && !(flags & DLM_LKF_CONVERT))
2609 if (flags & DLM_LKF_CONVDEADLK && flags & DLM_LKF_NOQUEUE)
2612 if (flags & DLM_LKF_EXPEDITE && flags & DLM_LKF_CONVERT)
2615 if (flags & DLM_LKF_EXPEDITE && flags & DLM_LKF_QUECVT)
2618 if (flags & DLM_LKF_EXPEDITE && flags & DLM_LKF_NOQUEUE)
2621 if (flags & DLM_LKF_EXPEDITE && mode != DLM_LOCK_NL)
2627 if (flags & DLM_LKF_VALBLK && !lksb->sb_lvbptr)
2630 if (flags & DLM_LKF_CONVERT && !lksb->sb_lkid)
2633 /* these args will be copied to the lkb in validate_lock_args,
2634 it cannot be done now because when converting locks, fields in
2635 an active lkb cannot be modified before locking the rsb */
2637 args->flags = flags;
2639 args->astparam = astparam;
2640 args->bastfn = bast;
2648 static int set_unlock_args(uint32_t flags, void *astarg, struct dlm_args *args)
2650 if (flags & ~(DLM_LKF_CANCEL | DLM_LKF_VALBLK | DLM_LKF_IVVALBLK |
2651 DLM_LKF_FORCEUNLOCK))
2654 if (flags & DLM_LKF_CANCEL && flags & DLM_LKF_FORCEUNLOCK)
2657 args->flags = flags;
2658 args->astparam = astarg;
2662 static int validate_lock_args(struct dlm_ls *ls, struct dlm_lkb *lkb,
2663 struct dlm_args *args)
2667 if (args->flags & DLM_LKF_CONVERT) {
2668 if (lkb->lkb_status != DLM_LKSTS_GRANTED)
2671 /* lock not allowed if there's any op in progress */
2672 if (lkb->lkb_wait_type || lkb->lkb_wait_count)
2675 if (is_overlap(lkb))
2679 if (test_bit(DLM_IFL_MSTCPY_BIT, &lkb->lkb_iflags))
2682 if (args->flags & DLM_LKF_QUECVT &&
2683 !__quecvt_compat_matrix[lkb->lkb_grmode+1][args->mode+1])
2687 lkb->lkb_exflags = args->flags;
2688 dlm_set_sbflags_val(lkb, 0);
2689 lkb->lkb_astfn = args->astfn;
2690 lkb->lkb_astparam = args->astparam;
2691 lkb->lkb_bastfn = args->bastfn;
2692 lkb->lkb_rqmode = args->mode;
2693 lkb->lkb_lksb = args->lksb;
2694 lkb->lkb_lvbptr = args->lksb->sb_lvbptr;
2695 lkb->lkb_ownpid = (int) current->pid;
2702 /* annoy the user because dlm usage is wrong */
2704 log_error(ls, "%s %d %x %x %x %d %d %s", __func__,
2705 rv, lkb->lkb_id, dlm_iflags_val(lkb), args->flags,
2706 lkb->lkb_status, lkb->lkb_wait_type,
2707 lkb->lkb_resource->res_name);
2710 log_debug(ls, "%s %d %x %x %x %d %d %s", __func__,
2711 rv, lkb->lkb_id, dlm_iflags_val(lkb), args->flags,
2712 lkb->lkb_status, lkb->lkb_wait_type,
2713 lkb->lkb_resource->res_name);
2720 /* when dlm_unlock() sees -EBUSY with CANCEL/FORCEUNLOCK it returns 0
2723 /* note: it's valid for lkb_nodeid/res_nodeid to be -1 when we get here
2724 because there may be a lookup in progress and it's valid to do
2725 cancel/unlockf on it */
2727 static int validate_unlock_args(struct dlm_lkb *lkb, struct dlm_args *args)
2729 struct dlm_ls *ls = lkb->lkb_resource->res_ls;
2732 /* normal unlock not allowed if there's any op in progress */
2733 if (!(args->flags & (DLM_LKF_CANCEL | DLM_LKF_FORCEUNLOCK)) &&
2734 (lkb->lkb_wait_type || lkb->lkb_wait_count))
2737 /* an lkb may be waiting for an rsb lookup to complete where the
2738 lookup was initiated by another lock */
2740 if (!list_empty(&lkb->lkb_rsb_lookup)) {
2741 if (args->flags & (DLM_LKF_CANCEL | DLM_LKF_FORCEUNLOCK)) {
2742 log_debug(ls, "unlock on rsb_lookup %x", lkb->lkb_id);
2743 list_del_init(&lkb->lkb_rsb_lookup);
2744 queue_cast(lkb->lkb_resource, lkb,
2745 args->flags & DLM_LKF_CANCEL ?
2746 -DLM_ECANCEL : -DLM_EUNLOCK);
2747 unhold_lkb(lkb); /* undoes create_lkb() */
2749 /* caller changes -EBUSY to 0 for CANCEL and FORCEUNLOCK */
2754 if (test_bit(DLM_IFL_MSTCPY_BIT, &lkb->lkb_iflags)) {
2755 log_error(ls, "unlock on MSTCPY %x", lkb->lkb_id);
2760 /* an lkb may still exist even though the lock is EOL'ed due to a
2761 * cancel, unlock or failed noqueue request; an app can't use these
2762 * locks; return same error as if the lkid had not been found at all
2765 if (test_bit(DLM_IFL_ENDOFLIFE_BIT, &lkb->lkb_iflags)) {
2766 log_debug(ls, "unlock on ENDOFLIFE %x", lkb->lkb_id);
2771 /* cancel not allowed with another cancel/unlock in progress */
2773 if (args->flags & DLM_LKF_CANCEL) {
2774 if (lkb->lkb_exflags & DLM_LKF_CANCEL)
2777 if (is_overlap(lkb))
2780 if (test_bit(DLM_IFL_RESEND_BIT, &lkb->lkb_iflags)) {
2781 set_bit(DLM_IFL_OVERLAP_CANCEL_BIT, &lkb->lkb_iflags);
2786 /* there's nothing to cancel */
2787 if (lkb->lkb_status == DLM_LKSTS_GRANTED &&
2788 !lkb->lkb_wait_type) {
2793 switch (lkb->lkb_wait_type) {
2794 case DLM_MSG_LOOKUP:
2795 case DLM_MSG_REQUEST:
2796 set_bit(DLM_IFL_OVERLAP_CANCEL_BIT, &lkb->lkb_iflags);
2799 case DLM_MSG_UNLOCK:
2800 case DLM_MSG_CANCEL:
2803 /* add_to_waiters() will set OVERLAP_CANCEL */
2807 /* do we need to allow a force-unlock if there's a normal unlock
2808 already in progress? in what conditions could the normal unlock
2809 fail such that we'd want to send a force-unlock to be sure? */
2811 if (args->flags & DLM_LKF_FORCEUNLOCK) {
2812 if (lkb->lkb_exflags & DLM_LKF_FORCEUNLOCK)
2815 if (is_overlap_unlock(lkb))
2818 if (test_bit(DLM_IFL_RESEND_BIT, &lkb->lkb_iflags)) {
2819 set_bit(DLM_IFL_OVERLAP_UNLOCK_BIT, &lkb->lkb_iflags);
2824 switch (lkb->lkb_wait_type) {
2825 case DLM_MSG_LOOKUP:
2826 case DLM_MSG_REQUEST:
2827 set_bit(DLM_IFL_OVERLAP_UNLOCK_BIT, &lkb->lkb_iflags);
2830 case DLM_MSG_UNLOCK:
2833 /* add_to_waiters() will set OVERLAP_UNLOCK */
2837 /* an overlapping op shouldn't blow away exflags from other op */
2838 lkb->lkb_exflags |= args->flags;
2839 dlm_set_sbflags_val(lkb, 0);
2840 lkb->lkb_astparam = args->astparam;
2847 /* annoy the user because dlm usage is wrong */
2849 log_error(ls, "%s %d %x %x %x %x %d %s", __func__, rv,
2850 lkb->lkb_id, dlm_iflags_val(lkb), lkb->lkb_exflags,
2851 args->flags, lkb->lkb_wait_type,
2852 lkb->lkb_resource->res_name);
2855 log_debug(ls, "%s %d %x %x %x %x %d %s", __func__, rv,
2856 lkb->lkb_id, dlm_iflags_val(lkb), lkb->lkb_exflags,
2857 args->flags, lkb->lkb_wait_type,
2858 lkb->lkb_resource->res_name);
2866 * Four stage 4 varieties:
2867 * do_request(), do_convert(), do_unlock(), do_cancel()
2868 * These are called on the master node for the given lock and
2869 * from the central locking logic.
2872 static int do_request(struct dlm_rsb *r, struct dlm_lkb *lkb)
2876 if (can_be_granted(r, lkb, 1, 0, NULL)) {
2878 queue_cast(r, lkb, 0);
2882 if (can_be_queued(lkb)) {
2883 error = -EINPROGRESS;
2884 add_lkb(r, lkb, DLM_LKSTS_WAITING);
2889 queue_cast(r, lkb, -EAGAIN);
2894 static void do_request_effects(struct dlm_rsb *r, struct dlm_lkb *lkb,
2899 if (force_blocking_asts(lkb))
2900 send_blocking_asts_all(r, lkb);
2903 send_blocking_asts(r, lkb);
2908 static int do_convert(struct dlm_rsb *r, struct dlm_lkb *lkb)
2913 /* changing an existing lock may allow others to be granted */
2915 if (can_be_granted(r, lkb, 1, 0, &deadlk)) {
2917 queue_cast(r, lkb, 0);
2921 /* can_be_granted() detected that this lock would block in a conversion
2922 deadlock, so we leave it on the granted queue and return EDEADLK in
2923 the ast for the convert. */
2925 if (deadlk && !(lkb->lkb_exflags & DLM_LKF_NODLCKWT)) {
2926 /* it's left on the granted queue */
2927 revert_lock(r, lkb);
2928 queue_cast(r, lkb, -EDEADLK);
2933 /* is_demoted() means the can_be_granted() above set the grmode
2934 to NL, and left us on the granted queue. This auto-demotion
2935 (due to CONVDEADLK) might mean other locks, and/or this lock, are
2936 now grantable. We have to try to grant other converting locks
2937 before we try again to grant this one. */
2939 if (is_demoted(lkb)) {
2940 grant_pending_convert(r, DLM_LOCK_IV, NULL, NULL);
2941 if (_can_be_granted(r, lkb, 1, 0)) {
2943 queue_cast(r, lkb, 0);
2946 /* else fall through and move to convert queue */
2949 if (can_be_queued(lkb)) {
2950 error = -EINPROGRESS;
2952 add_lkb(r, lkb, DLM_LKSTS_CONVERT);
2957 queue_cast(r, lkb, -EAGAIN);
2962 static void do_convert_effects(struct dlm_rsb *r, struct dlm_lkb *lkb,
2967 grant_pending_locks(r, NULL);
2968 /* grant_pending_locks also sends basts */
2971 if (force_blocking_asts(lkb))
2972 send_blocking_asts_all(r, lkb);
2975 send_blocking_asts(r, lkb);
2980 static int do_unlock(struct dlm_rsb *r, struct dlm_lkb *lkb)
2982 remove_lock(r, lkb);
2983 queue_cast(r, lkb, -DLM_EUNLOCK);
2984 return -DLM_EUNLOCK;
2987 static void do_unlock_effects(struct dlm_rsb *r, struct dlm_lkb *lkb,
2990 grant_pending_locks(r, NULL);
2993 /* returns: 0 did nothing, -DLM_ECANCEL canceled lock */
2995 static int do_cancel(struct dlm_rsb *r, struct dlm_lkb *lkb)
2999 error = revert_lock(r, lkb);
3001 queue_cast(r, lkb, -DLM_ECANCEL);
3002 return -DLM_ECANCEL;
3007 static void do_cancel_effects(struct dlm_rsb *r, struct dlm_lkb *lkb,
3011 grant_pending_locks(r, NULL);
3015 * Four stage 3 varieties:
3016 * _request_lock(), _convert_lock(), _unlock_lock(), _cancel_lock()
3019 /* add a new lkb to a possibly new rsb, called by requesting process */
3021 static int _request_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
3025 /* set_master: sets lkb nodeid from r */
3027 error = set_master(r, lkb);
3036 /* receive_request() calls do_request() on remote node */
3037 error = send_request(r, lkb);
3039 error = do_request(r, lkb);
3040 /* for remote locks the request_reply is sent
3041 between do_request and do_request_effects */
3042 do_request_effects(r, lkb, error);
3048 /* change some property of an existing lkb, e.g. mode */
3050 static int _convert_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
3055 /* receive_convert() calls do_convert() on remote node */
3056 error = send_convert(r, lkb);
3058 error = do_convert(r, lkb);
3059 /* for remote locks the convert_reply is sent
3060 between do_convert and do_convert_effects */
3061 do_convert_effects(r, lkb, error);
3067 /* remove an existing lkb from the granted queue */
3069 static int _unlock_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
3074 /* receive_unlock() calls do_unlock() on remote node */
3075 error = send_unlock(r, lkb);
3077 error = do_unlock(r, lkb);
3078 /* for remote locks the unlock_reply is sent
3079 between do_unlock and do_unlock_effects */
3080 do_unlock_effects(r, lkb, error);
3086 /* remove an existing lkb from the convert or wait queue */
3088 static int _cancel_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
3093 /* receive_cancel() calls do_cancel() on remote node */
3094 error = send_cancel(r, lkb);
3096 error = do_cancel(r, lkb);
3097 /* for remote locks the cancel_reply is sent
3098 between do_cancel and do_cancel_effects */
3099 do_cancel_effects(r, lkb, error);
3106 * Four stage 2 varieties:
3107 * request_lock(), convert_lock(), unlock_lock(), cancel_lock()
3110 static int request_lock(struct dlm_ls *ls, struct dlm_lkb *lkb,
3111 const void *name, int len,
3112 struct dlm_args *args)
3117 error = validate_lock_args(ls, lkb, args);
3121 error = find_rsb(ls, name, len, 0, R_REQUEST, &r);
3128 lkb->lkb_lksb->sb_lkid = lkb->lkb_id;
3130 error = _request_lock(r, lkb);
3137 static int convert_lock(struct dlm_ls *ls, struct dlm_lkb *lkb,
3138 struct dlm_args *args)
3143 r = lkb->lkb_resource;
3148 error = validate_lock_args(ls, lkb, args);
3152 error = _convert_lock(r, lkb);
3159 static int unlock_lock(struct dlm_ls *ls, struct dlm_lkb *lkb,
3160 struct dlm_args *args)
3165 r = lkb->lkb_resource;
3170 error = validate_unlock_args(lkb, args);
3174 error = _unlock_lock(r, lkb);
3181 static int cancel_lock(struct dlm_ls *ls, struct dlm_lkb *lkb,
3182 struct dlm_args *args)
3187 r = lkb->lkb_resource;
3192 error = validate_unlock_args(lkb, args);
3196 error = _cancel_lock(r, lkb);
3204 * Two stage 1 varieties: dlm_lock() and dlm_unlock()
3207 int dlm_lock(dlm_lockspace_t *lockspace,
3209 struct dlm_lksb *lksb,
3212 unsigned int namelen,
3213 uint32_t parent_lkid,
3214 void (*ast) (void *astarg),
3216 void (*bast) (void *astarg, int mode))
3219 struct dlm_lkb *lkb;
3220 struct dlm_args args;
3221 int error, convert = flags & DLM_LKF_CONVERT;
3223 ls = dlm_find_lockspace_local(lockspace);
3227 dlm_lock_recovery(ls);
3230 error = find_lkb(ls, lksb->sb_lkid, &lkb);
3232 error = create_lkb(ls, &lkb);
3237 trace_dlm_lock_start(ls, lkb, name, namelen, mode, flags);
3239 error = set_lock_args(mode, lksb, flags, namelen, ast, astarg, bast,
3245 error = convert_lock(ls, lkb, &args);
3247 error = request_lock(ls, lkb, name, namelen, &args);
3249 if (error == -EINPROGRESS)
3252 trace_dlm_lock_end(ls, lkb, name, namelen, mode, flags, error, true);
3254 if (convert || error)
3256 if (error == -EAGAIN || error == -EDEADLK)
3259 dlm_unlock_recovery(ls);
3260 dlm_put_lockspace(ls);
3264 int dlm_unlock(dlm_lockspace_t *lockspace,
3267 struct dlm_lksb *lksb,
3271 struct dlm_lkb *lkb;
3272 struct dlm_args args;
3275 ls = dlm_find_lockspace_local(lockspace);
3279 dlm_lock_recovery(ls);
3281 error = find_lkb(ls, lkid, &lkb);
3285 trace_dlm_unlock_start(ls, lkb, flags);
3287 error = set_unlock_args(flags, astarg, &args);
3291 if (flags & DLM_LKF_CANCEL)
3292 error = cancel_lock(ls, lkb, &args);
3294 error = unlock_lock(ls, lkb, &args);
3296 if (error == -DLM_EUNLOCK || error == -DLM_ECANCEL)
3298 if (error == -EBUSY && (flags & (DLM_LKF_CANCEL | DLM_LKF_FORCEUNLOCK)))
3301 trace_dlm_unlock_end(ls, lkb, flags, error);
3305 dlm_unlock_recovery(ls);
3306 dlm_put_lockspace(ls);
3311 * send/receive routines for remote operations and replies
3315 * send_request receive_request
3316 * send_convert receive_convert
3317 * send_unlock receive_unlock
3318 * send_cancel receive_cancel
3319 * send_grant receive_grant
3320 * send_bast receive_bast
3321 * send_lookup receive_lookup
3322 * send_remove receive_remove
3325 * receive_request_reply send_request_reply
3326 * receive_convert_reply send_convert_reply
3327 * receive_unlock_reply send_unlock_reply
3328 * receive_cancel_reply send_cancel_reply
3329 * receive_lookup_reply send_lookup_reply
3332 static int _create_message(struct dlm_ls *ls, int mb_len,
3333 int to_nodeid, int mstype,
3334 struct dlm_message **ms_ret,
3335 struct dlm_mhandle **mh_ret,
3338 struct dlm_message *ms;
3339 struct dlm_mhandle *mh;
3342 /* get_buffer gives us a message handle (mh) that we need to
3343 pass into midcomms_commit and a message buffer (mb) that we
3344 write our data into */
3346 mh = dlm_midcomms_get_mhandle(to_nodeid, mb_len, allocation, &mb);
3350 ms = (struct dlm_message *) mb;
3352 ms->m_header.h_version = cpu_to_le32(DLM_HEADER_MAJOR | DLM_HEADER_MINOR);
3353 ms->m_header.u.h_lockspace = cpu_to_le32(ls->ls_global_id);
3354 ms->m_header.h_nodeid = cpu_to_le32(dlm_our_nodeid());
3355 ms->m_header.h_length = cpu_to_le16(mb_len);
3356 ms->m_header.h_cmd = DLM_MSG;
3358 ms->m_type = cpu_to_le32(mstype);
3365 static int create_message(struct dlm_rsb *r, struct dlm_lkb *lkb,
3366 int to_nodeid, int mstype,
3367 struct dlm_message **ms_ret,
3368 struct dlm_mhandle **mh_ret,
3371 int mb_len = sizeof(struct dlm_message);
3374 case DLM_MSG_REQUEST:
3375 case DLM_MSG_LOOKUP:
3376 case DLM_MSG_REMOVE:
3377 mb_len += r->res_length;
3379 case DLM_MSG_CONVERT:
3380 case DLM_MSG_UNLOCK:
3381 case DLM_MSG_REQUEST_REPLY:
3382 case DLM_MSG_CONVERT_REPLY:
3384 if (lkb && lkb->lkb_lvbptr && (lkb->lkb_exflags & DLM_LKF_VALBLK))
3385 mb_len += r->res_ls->ls_lvblen;
3389 return _create_message(r->res_ls, mb_len, to_nodeid, mstype,
3390 ms_ret, mh_ret, allocation);
3393 /* further lowcomms enhancements or alternate implementations may make
3394 the return value from this function useful at some point */
3396 static int send_message(struct dlm_mhandle *mh, struct dlm_message *ms,
3397 const void *name, int namelen)
3399 dlm_midcomms_commit_mhandle(mh, name, namelen);
3403 static void send_args(struct dlm_rsb *r, struct dlm_lkb *lkb,
3404 struct dlm_message *ms)
3406 ms->m_nodeid = cpu_to_le32(lkb->lkb_nodeid);
3407 ms->m_pid = cpu_to_le32(lkb->lkb_ownpid);
3408 ms->m_lkid = cpu_to_le32(lkb->lkb_id);
3409 ms->m_remid = cpu_to_le32(lkb->lkb_remid);
3410 ms->m_exflags = cpu_to_le32(lkb->lkb_exflags);
3411 ms->m_sbflags = cpu_to_le32(dlm_sbflags_val(lkb));
3412 ms->m_flags = cpu_to_le32(dlm_dflags_val(lkb));
3413 ms->m_lvbseq = cpu_to_le32(lkb->lkb_lvbseq);
3414 ms->m_status = cpu_to_le32(lkb->lkb_status);
3415 ms->m_grmode = cpu_to_le32(lkb->lkb_grmode);
3416 ms->m_rqmode = cpu_to_le32(lkb->lkb_rqmode);
3417 ms->m_hash = cpu_to_le32(r->res_hash);
3419 /* m_result and m_bastmode are set from function args,
3420 not from lkb fields */
3422 if (lkb->lkb_bastfn)
3423 ms->m_asts |= cpu_to_le32(DLM_CB_BAST);
3425 ms->m_asts |= cpu_to_le32(DLM_CB_CAST);
3427 /* compare with switch in create_message; send_remove() doesn't
3430 switch (ms->m_type) {
3431 case cpu_to_le32(DLM_MSG_REQUEST):
3432 case cpu_to_le32(DLM_MSG_LOOKUP):
3433 memcpy(ms->m_extra, r->res_name, r->res_length);
3435 case cpu_to_le32(DLM_MSG_CONVERT):
3436 case cpu_to_le32(DLM_MSG_UNLOCK):
3437 case cpu_to_le32(DLM_MSG_REQUEST_REPLY):
3438 case cpu_to_le32(DLM_MSG_CONVERT_REPLY):
3439 case cpu_to_le32(DLM_MSG_GRANT):
3440 if (!lkb->lkb_lvbptr || !(lkb->lkb_exflags & DLM_LKF_VALBLK))
3442 memcpy(ms->m_extra, lkb->lkb_lvbptr, r->res_ls->ls_lvblen);
3447 static int send_common(struct dlm_rsb *r, struct dlm_lkb *lkb, int mstype)
3449 struct dlm_message *ms;
3450 struct dlm_mhandle *mh;
3451 int to_nodeid, error;
3453 to_nodeid = r->res_nodeid;
3455 error = add_to_waiters(lkb, mstype, to_nodeid);
3459 error = create_message(r, lkb, to_nodeid, mstype, &ms, &mh, GFP_NOFS);
3463 send_args(r, lkb, ms);
3465 error = send_message(mh, ms, r->res_name, r->res_length);
3471 remove_from_waiters(lkb, msg_reply_type(mstype));
3475 static int send_request(struct dlm_rsb *r, struct dlm_lkb *lkb)
3477 return send_common(r, lkb, DLM_MSG_REQUEST);
3480 static int send_convert(struct dlm_rsb *r, struct dlm_lkb *lkb)
3484 error = send_common(r, lkb, DLM_MSG_CONVERT);
3486 /* down conversions go without a reply from the master */
3487 if (!error && down_conversion(lkb)) {
3488 remove_from_waiters(lkb, DLM_MSG_CONVERT_REPLY);
3489 r->res_ls->ls_local_ms.m_type = cpu_to_le32(DLM_MSG_CONVERT_REPLY);
3490 r->res_ls->ls_local_ms.m_result = 0;
3491 __receive_convert_reply(r, lkb, &r->res_ls->ls_local_ms, true);
3497 /* FIXME: if this lkb is the only lock we hold on the rsb, then set
3498 MASTER_UNCERTAIN to force the next request on the rsb to confirm
3499 that the master is still correct. */
3501 static int send_unlock(struct dlm_rsb *r, struct dlm_lkb *lkb)
3503 return send_common(r, lkb, DLM_MSG_UNLOCK);
3506 static int send_cancel(struct dlm_rsb *r, struct dlm_lkb *lkb)
3508 return send_common(r, lkb, DLM_MSG_CANCEL);
3511 static int send_grant(struct dlm_rsb *r, struct dlm_lkb *lkb)
3513 struct dlm_message *ms;
3514 struct dlm_mhandle *mh;
3515 int to_nodeid, error;
3517 to_nodeid = lkb->lkb_nodeid;
3519 error = create_message(r, lkb, to_nodeid, DLM_MSG_GRANT, &ms, &mh,
3524 send_args(r, lkb, ms);
3528 error = send_message(mh, ms, r->res_name, r->res_length);
3533 static int send_bast(struct dlm_rsb *r, struct dlm_lkb *lkb, int mode)
3535 struct dlm_message *ms;
3536 struct dlm_mhandle *mh;
3537 int to_nodeid, error;
3539 to_nodeid = lkb->lkb_nodeid;
3541 error = create_message(r, NULL, to_nodeid, DLM_MSG_BAST, &ms, &mh,
3546 send_args(r, lkb, ms);
3548 ms->m_bastmode = cpu_to_le32(mode);
3550 error = send_message(mh, ms, r->res_name, r->res_length);
3555 static int send_lookup(struct dlm_rsb *r, struct dlm_lkb *lkb)
3557 struct dlm_message *ms;
3558 struct dlm_mhandle *mh;
3559 int to_nodeid, error;
3561 to_nodeid = dlm_dir_nodeid(r);
3563 error = add_to_waiters(lkb, DLM_MSG_LOOKUP, to_nodeid);
3567 error = create_message(r, NULL, to_nodeid, DLM_MSG_LOOKUP, &ms, &mh,
3572 send_args(r, lkb, ms);
3574 error = send_message(mh, ms, r->res_name, r->res_length);
3580 remove_from_waiters(lkb, DLM_MSG_LOOKUP_REPLY);
3584 static int send_remove(struct dlm_rsb *r)
3586 struct dlm_message *ms;
3587 struct dlm_mhandle *mh;
3588 int to_nodeid, error;
3590 to_nodeid = dlm_dir_nodeid(r);
3592 error = create_message(r, NULL, to_nodeid, DLM_MSG_REMOVE, &ms, &mh,
3597 memcpy(ms->m_extra, r->res_name, r->res_length);
3598 ms->m_hash = cpu_to_le32(r->res_hash);
3600 error = send_message(mh, ms, r->res_name, r->res_length);
3605 static int send_common_reply(struct dlm_rsb *r, struct dlm_lkb *lkb,
3608 struct dlm_message *ms;
3609 struct dlm_mhandle *mh;
3610 int to_nodeid, error;
3612 to_nodeid = lkb->lkb_nodeid;
3614 error = create_message(r, lkb, to_nodeid, mstype, &ms, &mh, GFP_NOFS);
3618 send_args(r, lkb, ms);
3620 ms->m_result = cpu_to_le32(to_dlm_errno(rv));
3622 error = send_message(mh, ms, r->res_name, r->res_length);
3627 static int send_request_reply(struct dlm_rsb *r, struct dlm_lkb *lkb, int rv)
3629 return send_common_reply(r, lkb, DLM_MSG_REQUEST_REPLY, rv);
3632 static int send_convert_reply(struct dlm_rsb *r, struct dlm_lkb *lkb, int rv)
3634 return send_common_reply(r, lkb, DLM_MSG_CONVERT_REPLY, rv);
3637 static int send_unlock_reply(struct dlm_rsb *r, struct dlm_lkb *lkb, int rv)
3639 return send_common_reply(r, lkb, DLM_MSG_UNLOCK_REPLY, rv);
3642 static int send_cancel_reply(struct dlm_rsb *r, struct dlm_lkb *lkb, int rv)
3644 return send_common_reply(r, lkb, DLM_MSG_CANCEL_REPLY, rv);
3647 static int send_lookup_reply(struct dlm_ls *ls,
3648 const struct dlm_message *ms_in, int ret_nodeid,
3651 struct dlm_rsb *r = &ls->ls_local_rsb;
3652 struct dlm_message *ms;
3653 struct dlm_mhandle *mh;
3654 int error, nodeid = le32_to_cpu(ms_in->m_header.h_nodeid);
3656 error = create_message(r, NULL, nodeid, DLM_MSG_LOOKUP_REPLY, &ms, &mh,
3661 ms->m_lkid = ms_in->m_lkid;
3662 ms->m_result = cpu_to_le32(to_dlm_errno(rv));
3663 ms->m_nodeid = cpu_to_le32(ret_nodeid);
3665 error = send_message(mh, ms, ms_in->m_extra, receive_extralen(ms_in));
3670 /* which args we save from a received message depends heavily on the type
3671 of message, unlike the send side where we can safely send everything about
3672 the lkb for any type of message */
3674 static void receive_flags(struct dlm_lkb *lkb, const struct dlm_message *ms)
3676 lkb->lkb_exflags = le32_to_cpu(ms->m_exflags);
3677 dlm_set_sbflags_val(lkb, le32_to_cpu(ms->m_sbflags));
3678 dlm_set_dflags_val(lkb, le32_to_cpu(ms->m_flags));
3681 static void receive_flags_reply(struct dlm_lkb *lkb,
3682 const struct dlm_message *ms,
3688 dlm_set_sbflags_val(lkb, le32_to_cpu(ms->m_sbflags));
3689 dlm_set_dflags_val(lkb, le32_to_cpu(ms->m_flags));
3692 static int receive_extralen(const struct dlm_message *ms)
3694 return (le16_to_cpu(ms->m_header.h_length) -
3695 sizeof(struct dlm_message));
3698 static int receive_lvb(struct dlm_ls *ls, struct dlm_lkb *lkb,
3699 const struct dlm_message *ms)
3703 if (lkb->lkb_exflags & DLM_LKF_VALBLK) {
3704 if (!lkb->lkb_lvbptr)
3705 lkb->lkb_lvbptr = dlm_allocate_lvb(ls);
3706 if (!lkb->lkb_lvbptr)
3708 len = receive_extralen(ms);
3709 if (len > ls->ls_lvblen)
3710 len = ls->ls_lvblen;
3711 memcpy(lkb->lkb_lvbptr, ms->m_extra, len);
3716 static void fake_bastfn(void *astparam, int mode)
3718 log_print("fake_bastfn should not be called");
3721 static void fake_astfn(void *astparam)
3723 log_print("fake_astfn should not be called");
3726 static int receive_request_args(struct dlm_ls *ls, struct dlm_lkb *lkb,
3727 const struct dlm_message *ms)
3729 lkb->lkb_nodeid = le32_to_cpu(ms->m_header.h_nodeid);
3730 lkb->lkb_ownpid = le32_to_cpu(ms->m_pid);
3731 lkb->lkb_remid = le32_to_cpu(ms->m_lkid);
3732 lkb->lkb_grmode = DLM_LOCK_IV;
3733 lkb->lkb_rqmode = le32_to_cpu(ms->m_rqmode);
3735 lkb->lkb_bastfn = (ms->m_asts & cpu_to_le32(DLM_CB_BAST)) ? &fake_bastfn : NULL;
3736 lkb->lkb_astfn = (ms->m_asts & cpu_to_le32(DLM_CB_CAST)) ? &fake_astfn : NULL;
3738 if (lkb->lkb_exflags & DLM_LKF_VALBLK) {
3739 /* lkb was just created so there won't be an lvb yet */
3740 lkb->lkb_lvbptr = dlm_allocate_lvb(ls);
3741 if (!lkb->lkb_lvbptr)
3748 static int receive_convert_args(struct dlm_ls *ls, struct dlm_lkb *lkb,
3749 const struct dlm_message *ms)
3751 if (lkb->lkb_status != DLM_LKSTS_GRANTED)
3754 if (receive_lvb(ls, lkb, ms))
3757 lkb->lkb_rqmode = le32_to_cpu(ms->m_rqmode);
3758 lkb->lkb_lvbseq = le32_to_cpu(ms->m_lvbseq);
3763 static int receive_unlock_args(struct dlm_ls *ls, struct dlm_lkb *lkb,
3764 const struct dlm_message *ms)
3766 if (receive_lvb(ls, lkb, ms))
3771 /* We fill in the local-lkb fields with the info that send_xxxx_reply()
3772 uses to send a reply and that the remote end uses to process the reply. */
3774 static void setup_local_lkb(struct dlm_ls *ls, const struct dlm_message *ms)
3776 struct dlm_lkb *lkb = &ls->ls_local_lkb;
3777 lkb->lkb_nodeid = le32_to_cpu(ms->m_header.h_nodeid);
3778 lkb->lkb_remid = le32_to_cpu(ms->m_lkid);
3781 /* This is called after the rsb is locked so that we can safely inspect
3782 fields in the lkb. */
3784 static int validate_message(struct dlm_lkb *lkb, const struct dlm_message *ms)
3786 int from = le32_to_cpu(ms->m_header.h_nodeid);
3789 /* currently mixing of user/kernel locks are not supported */
3790 if (ms->m_flags & cpu_to_le32(BIT(DLM_DFL_USER_BIT)) &&
3791 !test_bit(DLM_DFL_USER_BIT, &lkb->lkb_dflags)) {
3792 log_error(lkb->lkb_resource->res_ls,
3793 "got user dlm message for a kernel lock");
3798 switch (ms->m_type) {
3799 case cpu_to_le32(DLM_MSG_CONVERT):
3800 case cpu_to_le32(DLM_MSG_UNLOCK):
3801 case cpu_to_le32(DLM_MSG_CANCEL):
3802 if (!is_master_copy(lkb) || lkb->lkb_nodeid != from)
3806 case cpu_to_le32(DLM_MSG_CONVERT_REPLY):
3807 case cpu_to_le32(DLM_MSG_UNLOCK_REPLY):
3808 case cpu_to_le32(DLM_MSG_CANCEL_REPLY):
3809 case cpu_to_le32(DLM_MSG_GRANT):
3810 case cpu_to_le32(DLM_MSG_BAST):
3811 if (!is_process_copy(lkb) || lkb->lkb_nodeid != from)
3815 case cpu_to_le32(DLM_MSG_REQUEST_REPLY):
3816 if (!is_process_copy(lkb))
3818 else if (lkb->lkb_nodeid != -1 && lkb->lkb_nodeid != from)
3828 log_error(lkb->lkb_resource->res_ls,
3829 "ignore invalid message %d from %d %x %x %x %d",
3830 le32_to_cpu(ms->m_type), from, lkb->lkb_id,
3831 lkb->lkb_remid, dlm_iflags_val(lkb),
3836 static int receive_request(struct dlm_ls *ls, const struct dlm_message *ms)
3838 struct dlm_lkb *lkb;
3841 int error, namelen = 0;
3843 from_nodeid = le32_to_cpu(ms->m_header.h_nodeid);
3845 error = create_lkb(ls, &lkb);
3849 receive_flags(lkb, ms);
3850 set_bit(DLM_IFL_MSTCPY_BIT, &lkb->lkb_iflags);
3851 error = receive_request_args(ls, lkb, ms);
3857 /* The dir node is the authority on whether we are the master
3858 for this rsb or not, so if the master sends us a request, we should
3859 recreate the rsb if we've destroyed it. This race happens when we
3860 send a remove message to the dir node at the same time that the dir
3861 node sends us a request for the rsb. */
3863 namelen = receive_extralen(ms);
3865 error = find_rsb(ls, ms->m_extra, namelen, from_nodeid,
3866 R_RECEIVE_REQUEST, &r);
3874 if (r->res_master_nodeid != dlm_our_nodeid()) {
3875 error = validate_master_nodeid(ls, r, from_nodeid);
3885 error = do_request(r, lkb);
3886 send_request_reply(r, lkb, error);
3887 do_request_effects(r, lkb, error);
3892 if (error == -EINPROGRESS)
3899 /* TODO: instead of returning ENOTBLK, add the lkb to res_lookup
3900 and do this receive_request again from process_lookup_list once
3901 we get the lookup reply. This would avoid a many repeated
3902 ENOTBLK request failures when the lookup reply designating us
3903 as master is delayed. */
3905 if (error != -ENOTBLK) {
3906 log_limit(ls, "receive_request %x from %d %d",
3907 le32_to_cpu(ms->m_lkid), from_nodeid, error);
3910 setup_local_lkb(ls, ms);
3911 send_request_reply(&ls->ls_local_rsb, &ls->ls_local_lkb, error);
3915 static int receive_convert(struct dlm_ls *ls, const struct dlm_message *ms)
3917 struct dlm_lkb *lkb;
3919 int error, reply = 1;
3921 error = find_lkb(ls, le32_to_cpu(ms->m_remid), &lkb);
3925 if (lkb->lkb_remid != le32_to_cpu(ms->m_lkid)) {
3926 log_error(ls, "receive_convert %x remid %x recover_seq %llu "
3927 "remote %d %x", lkb->lkb_id, lkb->lkb_remid,
3928 (unsigned long long)lkb->lkb_recover_seq,
3929 le32_to_cpu(ms->m_header.h_nodeid),
3930 le32_to_cpu(ms->m_lkid));
3936 r = lkb->lkb_resource;
3941 error = validate_message(lkb, ms);
3945 receive_flags(lkb, ms);
3947 error = receive_convert_args(ls, lkb, ms);
3949 send_convert_reply(r, lkb, error);
3953 reply = !down_conversion(lkb);
3955 error = do_convert(r, lkb);
3957 send_convert_reply(r, lkb, error);
3958 do_convert_effects(r, lkb, error);
3966 setup_local_lkb(ls, ms);
3967 send_convert_reply(&ls->ls_local_rsb, &ls->ls_local_lkb, error);
3971 static int receive_unlock(struct dlm_ls *ls, const struct dlm_message *ms)
3973 struct dlm_lkb *lkb;
3977 error = find_lkb(ls, le32_to_cpu(ms->m_remid), &lkb);
3981 if (lkb->lkb_remid != le32_to_cpu(ms->m_lkid)) {
3982 log_error(ls, "receive_unlock %x remid %x remote %d %x",
3983 lkb->lkb_id, lkb->lkb_remid,
3984 le32_to_cpu(ms->m_header.h_nodeid),
3985 le32_to_cpu(ms->m_lkid));
3991 r = lkb->lkb_resource;
3996 error = validate_message(lkb, ms);
4000 receive_flags(lkb, ms);
4002 error = receive_unlock_args(ls, lkb, ms);
4004 send_unlock_reply(r, lkb, error);
4008 error = do_unlock(r, lkb);
4009 send_unlock_reply(r, lkb, error);
4010 do_unlock_effects(r, lkb, error);
4018 setup_local_lkb(ls, ms);
4019 send_unlock_reply(&ls->ls_local_rsb, &ls->ls_local_lkb, error);
4023 static int receive_cancel(struct dlm_ls *ls, const struct dlm_message *ms)
4025 struct dlm_lkb *lkb;
4029 error = find_lkb(ls, le32_to_cpu(ms->m_remid), &lkb);
4033 receive_flags(lkb, ms);
4035 r = lkb->lkb_resource;
4040 error = validate_message(lkb, ms);
4044 error = do_cancel(r, lkb);
4045 send_cancel_reply(r, lkb, error);
4046 do_cancel_effects(r, lkb, error);
4054 setup_local_lkb(ls, ms);
4055 send_cancel_reply(&ls->ls_local_rsb, &ls->ls_local_lkb, error);
4059 static int receive_grant(struct dlm_ls *ls, const struct dlm_message *ms)
4061 struct dlm_lkb *lkb;
4065 error = find_lkb(ls, le32_to_cpu(ms->m_remid), &lkb);
4069 r = lkb->lkb_resource;
4074 error = validate_message(lkb, ms);
4078 receive_flags_reply(lkb, ms, false);
4079 if (is_altmode(lkb))
4080 munge_altmode(lkb, ms);
4081 grant_lock_pc(r, lkb, ms);
4082 queue_cast(r, lkb, 0);
4090 static int receive_bast(struct dlm_ls *ls, const struct dlm_message *ms)
4092 struct dlm_lkb *lkb;
4096 error = find_lkb(ls, le32_to_cpu(ms->m_remid), &lkb);
4100 r = lkb->lkb_resource;
4105 error = validate_message(lkb, ms);
4109 queue_bast(r, lkb, le32_to_cpu(ms->m_bastmode));
4110 lkb->lkb_highbast = le32_to_cpu(ms->m_bastmode);
4118 static void receive_lookup(struct dlm_ls *ls, const struct dlm_message *ms)
4120 int len, error, ret_nodeid, from_nodeid, our_nodeid;
4122 from_nodeid = le32_to_cpu(ms->m_header.h_nodeid);
4123 our_nodeid = dlm_our_nodeid();
4125 len = receive_extralen(ms);
4127 error = dlm_master_lookup(ls, from_nodeid, ms->m_extra, len, 0,
4130 /* Optimization: we're master so treat lookup as a request */
4131 if (!error && ret_nodeid == our_nodeid) {
4132 receive_request(ls, ms);
4135 send_lookup_reply(ls, ms, ret_nodeid, error);
4138 static void receive_remove(struct dlm_ls *ls, const struct dlm_message *ms)
4140 char name[DLM_RESNAME_MAXLEN+1];
4143 int rv, len, dir_nodeid, from_nodeid;
4145 from_nodeid = le32_to_cpu(ms->m_header.h_nodeid);
4147 len = receive_extralen(ms);
4149 if (len > DLM_RESNAME_MAXLEN) {
4150 log_error(ls, "receive_remove from %d bad len %d",
4155 dir_nodeid = dlm_hash2nodeid(ls, le32_to_cpu(ms->m_hash));
4156 if (dir_nodeid != dlm_our_nodeid()) {
4157 log_error(ls, "receive_remove from %d bad nodeid %d",
4158 from_nodeid, dir_nodeid);
4162 /* Look for name on rsbtbl.toss, if it's there, kill it.
4163 If it's on rsbtbl.keep, it's being used, and we should ignore this
4164 message. This is an expected race between the dir node sending a
4165 request to the master node at the same time as the master node sends
4166 a remove to the dir node. The resolution to that race is for the
4167 dir node to ignore the remove message, and the master node to
4168 recreate the master rsb when it gets a request from the dir node for
4169 an rsb it doesn't have. */
4171 memset(name, 0, sizeof(name));
4172 memcpy(name, ms->m_extra, len);
4174 hash = jhash(name, len, 0);
4175 b = hash & (ls->ls_rsbtbl_size - 1);
4177 spin_lock(&ls->ls_rsbtbl[b].lock);
4179 rv = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].toss, name, len, &r);
4181 /* verify the rsb is on keep list per comment above */
4182 rv = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].keep, name, len, &r);
4184 /* should not happen */
4185 log_error(ls, "receive_remove from %d not found %s",
4187 spin_unlock(&ls->ls_rsbtbl[b].lock);
4190 if (r->res_master_nodeid != from_nodeid) {
4191 /* should not happen */
4192 log_error(ls, "receive_remove keep from %d master %d",
4193 from_nodeid, r->res_master_nodeid);
4195 spin_unlock(&ls->ls_rsbtbl[b].lock);
4199 log_debug(ls, "receive_remove from %d master %d first %x %s",
4200 from_nodeid, r->res_master_nodeid, r->res_first_lkid,
4202 spin_unlock(&ls->ls_rsbtbl[b].lock);
4206 if (r->res_master_nodeid != from_nodeid) {
4207 log_error(ls, "receive_remove toss from %d master %d",
4208 from_nodeid, r->res_master_nodeid);
4210 spin_unlock(&ls->ls_rsbtbl[b].lock);
4214 if (kref_put(&r->res_ref, kill_rsb)) {
4215 rb_erase(&r->res_hashnode, &ls->ls_rsbtbl[b].toss);
4216 spin_unlock(&ls->ls_rsbtbl[b].lock);
4219 log_error(ls, "receive_remove from %d rsb ref error",
4222 spin_unlock(&ls->ls_rsbtbl[b].lock);
4226 static void receive_purge(struct dlm_ls *ls, const struct dlm_message *ms)
4228 do_purge(ls, le32_to_cpu(ms->m_nodeid), le32_to_cpu(ms->m_pid));
4231 static int receive_request_reply(struct dlm_ls *ls,
4232 const struct dlm_message *ms)
4234 struct dlm_lkb *lkb;
4236 int error, mstype, result;
4237 int from_nodeid = le32_to_cpu(ms->m_header.h_nodeid);
4239 error = find_lkb(ls, le32_to_cpu(ms->m_remid), &lkb);
4243 r = lkb->lkb_resource;
4247 error = validate_message(lkb, ms);
4251 mstype = lkb->lkb_wait_type;
4252 error = remove_from_waiters(lkb, DLM_MSG_REQUEST_REPLY);
4254 log_error(ls, "receive_request_reply %x remote %d %x result %d",
4255 lkb->lkb_id, from_nodeid, le32_to_cpu(ms->m_lkid),
4256 from_dlm_errno(le32_to_cpu(ms->m_result)));
4261 /* Optimization: the dir node was also the master, so it took our
4262 lookup as a request and sent request reply instead of lookup reply */
4263 if (mstype == DLM_MSG_LOOKUP) {
4264 r->res_master_nodeid = from_nodeid;
4265 r->res_nodeid = from_nodeid;
4266 lkb->lkb_nodeid = from_nodeid;
4269 /* this is the value returned from do_request() on the master */
4270 result = from_dlm_errno(le32_to_cpu(ms->m_result));
4274 /* request would block (be queued) on remote master */
4275 queue_cast(r, lkb, -EAGAIN);
4276 confirm_master(r, -EAGAIN);
4277 unhold_lkb(lkb); /* undoes create_lkb() */
4282 /* request was queued or granted on remote master */
4283 receive_flags_reply(lkb, ms, false);
4284 lkb->lkb_remid = le32_to_cpu(ms->m_lkid);
4285 if (is_altmode(lkb))
4286 munge_altmode(lkb, ms);
4288 add_lkb(r, lkb, DLM_LKSTS_WAITING);
4290 grant_lock_pc(r, lkb, ms);
4291 queue_cast(r, lkb, 0);
4293 confirm_master(r, result);
4298 /* find_rsb failed to find rsb or rsb wasn't master */
4299 log_limit(ls, "receive_request_reply %x from %d %d "
4300 "master %d dir %d first %x %s", lkb->lkb_id,
4301 from_nodeid, result, r->res_master_nodeid,
4302 r->res_dir_nodeid, r->res_first_lkid, r->res_name);
4304 if (r->res_dir_nodeid != dlm_our_nodeid() &&
4305 r->res_master_nodeid != dlm_our_nodeid()) {
4306 /* cause _request_lock->set_master->send_lookup */
4307 r->res_master_nodeid = 0;
4309 lkb->lkb_nodeid = -1;
4312 if (is_overlap(lkb)) {
4313 /* we'll ignore error in cancel/unlock reply */
4314 queue_cast_overlap(r, lkb);
4315 confirm_master(r, result);
4316 unhold_lkb(lkb); /* undoes create_lkb() */
4318 _request_lock(r, lkb);
4320 if (r->res_master_nodeid == dlm_our_nodeid())
4321 confirm_master(r, 0);
4326 log_error(ls, "receive_request_reply %x error %d",
4327 lkb->lkb_id, result);
4330 if ((result == 0 || result == -EINPROGRESS) &&
4331 test_and_clear_bit(DLM_IFL_OVERLAP_UNLOCK_BIT, &lkb->lkb_iflags)) {
4332 log_debug(ls, "receive_request_reply %x result %d unlock",
4333 lkb->lkb_id, result);
4334 clear_bit(DLM_IFL_OVERLAP_CANCEL_BIT, &lkb->lkb_iflags);
4335 send_unlock(r, lkb);
4336 } else if ((result == -EINPROGRESS) &&
4337 test_and_clear_bit(DLM_IFL_OVERLAP_CANCEL_BIT,
4338 &lkb->lkb_iflags)) {
4339 log_debug(ls, "receive_request_reply %x cancel", lkb->lkb_id);
4340 clear_bit(DLM_IFL_OVERLAP_UNLOCK_BIT, &lkb->lkb_iflags);
4341 send_cancel(r, lkb);
4343 clear_bit(DLM_IFL_OVERLAP_CANCEL_BIT, &lkb->lkb_iflags);
4344 clear_bit(DLM_IFL_OVERLAP_UNLOCK_BIT, &lkb->lkb_iflags);
4353 static void __receive_convert_reply(struct dlm_rsb *r, struct dlm_lkb *lkb,
4354 const struct dlm_message *ms, bool local)
4356 /* this is the value returned from do_convert() on the master */
4357 switch (from_dlm_errno(le32_to_cpu(ms->m_result))) {
4359 /* convert would block (be queued) on remote master */
4360 queue_cast(r, lkb, -EAGAIN);
4364 receive_flags_reply(lkb, ms, local);
4365 revert_lock_pc(r, lkb);
4366 queue_cast(r, lkb, -EDEADLK);
4370 /* convert was queued on remote master */
4371 receive_flags_reply(lkb, ms, local);
4372 if (is_demoted(lkb))
4375 add_lkb(r, lkb, DLM_LKSTS_CONVERT);
4379 /* convert was granted on remote master */
4380 receive_flags_reply(lkb, ms, local);
4381 if (is_demoted(lkb))
4383 grant_lock_pc(r, lkb, ms);
4384 queue_cast(r, lkb, 0);
4388 log_error(r->res_ls, "receive_convert_reply %x remote %d %x %d",
4389 lkb->lkb_id, le32_to_cpu(ms->m_header.h_nodeid),
4390 le32_to_cpu(ms->m_lkid),
4391 from_dlm_errno(le32_to_cpu(ms->m_result)));
4397 static void _receive_convert_reply(struct dlm_lkb *lkb,
4398 const struct dlm_message *ms, bool local)
4400 struct dlm_rsb *r = lkb->lkb_resource;
4406 error = validate_message(lkb, ms);
4410 /* local reply can happen with waiters_mutex held */
4411 error = remove_from_waiters_ms(lkb, ms, local);
4415 __receive_convert_reply(r, lkb, ms, local);
4421 static int receive_convert_reply(struct dlm_ls *ls,
4422 const struct dlm_message *ms)
4424 struct dlm_lkb *lkb;
4427 error = find_lkb(ls, le32_to_cpu(ms->m_remid), &lkb);
4431 _receive_convert_reply(lkb, ms, false);
4436 static void _receive_unlock_reply(struct dlm_lkb *lkb,
4437 const struct dlm_message *ms, bool local)
4439 struct dlm_rsb *r = lkb->lkb_resource;
4445 error = validate_message(lkb, ms);
4449 /* local reply can happen with waiters_mutex held */
4450 error = remove_from_waiters_ms(lkb, ms, local);
4454 /* this is the value returned from do_unlock() on the master */
4456 switch (from_dlm_errno(le32_to_cpu(ms->m_result))) {
4458 receive_flags_reply(lkb, ms, local);
4459 remove_lock_pc(r, lkb);
4460 queue_cast(r, lkb, -DLM_EUNLOCK);
4465 log_error(r->res_ls, "receive_unlock_reply %x error %d",
4466 lkb->lkb_id, from_dlm_errno(le32_to_cpu(ms->m_result)));
4473 static int receive_unlock_reply(struct dlm_ls *ls,
4474 const struct dlm_message *ms)
4476 struct dlm_lkb *lkb;
4479 error = find_lkb(ls, le32_to_cpu(ms->m_remid), &lkb);
4483 _receive_unlock_reply(lkb, ms, false);
4488 static void _receive_cancel_reply(struct dlm_lkb *lkb,
4489 const struct dlm_message *ms, bool local)
4491 struct dlm_rsb *r = lkb->lkb_resource;
4497 error = validate_message(lkb, ms);
4501 /* local reply can happen with waiters_mutex held */
4502 error = remove_from_waiters_ms(lkb, ms, local);
4506 /* this is the value returned from do_cancel() on the master */
4508 switch (from_dlm_errno(le32_to_cpu(ms->m_result))) {
4510 receive_flags_reply(lkb, ms, local);
4511 revert_lock_pc(r, lkb);
4512 queue_cast(r, lkb, -DLM_ECANCEL);
4517 log_error(r->res_ls, "receive_cancel_reply %x error %d",
4519 from_dlm_errno(le32_to_cpu(ms->m_result)));
4526 static int receive_cancel_reply(struct dlm_ls *ls,
4527 const struct dlm_message *ms)
4529 struct dlm_lkb *lkb;
4532 error = find_lkb(ls, le32_to_cpu(ms->m_remid), &lkb);
4536 _receive_cancel_reply(lkb, ms, false);
4541 static void receive_lookup_reply(struct dlm_ls *ls,
4542 const struct dlm_message *ms)
4544 struct dlm_lkb *lkb;
4546 int error, ret_nodeid;
4547 int do_lookup_list = 0;
4549 error = find_lkb(ls, le32_to_cpu(ms->m_lkid), &lkb);
4551 log_error(ls, "%s no lkid %x", __func__,
4552 le32_to_cpu(ms->m_lkid));
4556 /* ms->m_result is the value returned by dlm_master_lookup on dir node
4557 FIXME: will a non-zero error ever be returned? */
4559 r = lkb->lkb_resource;
4563 error = remove_from_waiters(lkb, DLM_MSG_LOOKUP_REPLY);
4567 ret_nodeid = le32_to_cpu(ms->m_nodeid);
4569 /* We sometimes receive a request from the dir node for this
4570 rsb before we've received the dir node's loookup_reply for it.
4571 The request from the dir node implies we're the master, so we set
4572 ourself as master in receive_request_reply, and verify here that
4573 we are indeed the master. */
4575 if (r->res_master_nodeid && (r->res_master_nodeid != ret_nodeid)) {
4576 /* This should never happen */
4577 log_error(ls, "receive_lookup_reply %x from %d ret %d "
4578 "master %d dir %d our %d first %x %s",
4579 lkb->lkb_id, le32_to_cpu(ms->m_header.h_nodeid),
4580 ret_nodeid, r->res_master_nodeid, r->res_dir_nodeid,
4581 dlm_our_nodeid(), r->res_first_lkid, r->res_name);
4584 if (ret_nodeid == dlm_our_nodeid()) {
4585 r->res_master_nodeid = ret_nodeid;
4588 r->res_first_lkid = 0;
4589 } else if (ret_nodeid == -1) {
4590 /* the remote node doesn't believe it's the dir node */
4591 log_error(ls, "receive_lookup_reply %x from %d bad ret_nodeid",
4592 lkb->lkb_id, le32_to_cpu(ms->m_header.h_nodeid));
4593 r->res_master_nodeid = 0;
4595 lkb->lkb_nodeid = -1;
4597 /* set_master() will set lkb_nodeid from r */
4598 r->res_master_nodeid = ret_nodeid;
4599 r->res_nodeid = ret_nodeid;
4602 if (is_overlap(lkb)) {
4603 log_debug(ls, "receive_lookup_reply %x unlock %x",
4604 lkb->lkb_id, dlm_iflags_val(lkb));
4605 queue_cast_overlap(r, lkb);
4606 unhold_lkb(lkb); /* undoes create_lkb() */
4610 _request_lock(r, lkb);
4614 process_lookup_list(r);
4621 static void _receive_message(struct dlm_ls *ls, const struct dlm_message *ms,
4624 int error = 0, noent = 0;
4626 if (WARN_ON_ONCE(!dlm_is_member(ls, le32_to_cpu(ms->m_header.h_nodeid)))) {
4627 log_limit(ls, "receive %d from non-member %d %x %x %d",
4628 le32_to_cpu(ms->m_type),
4629 le32_to_cpu(ms->m_header.h_nodeid),
4630 le32_to_cpu(ms->m_lkid), le32_to_cpu(ms->m_remid),
4631 from_dlm_errno(le32_to_cpu(ms->m_result)));
4635 switch (ms->m_type) {
4637 /* messages sent to a master node */
4639 case cpu_to_le32(DLM_MSG_REQUEST):
4640 error = receive_request(ls, ms);
4643 case cpu_to_le32(DLM_MSG_CONVERT):
4644 error = receive_convert(ls, ms);
4647 case cpu_to_le32(DLM_MSG_UNLOCK):
4648 error = receive_unlock(ls, ms);
4651 case cpu_to_le32(DLM_MSG_CANCEL):
4653 error = receive_cancel(ls, ms);
4656 /* messages sent from a master node (replies to above) */
4658 case cpu_to_le32(DLM_MSG_REQUEST_REPLY):
4659 error = receive_request_reply(ls, ms);
4662 case cpu_to_le32(DLM_MSG_CONVERT_REPLY):
4663 error = receive_convert_reply(ls, ms);
4666 case cpu_to_le32(DLM_MSG_UNLOCK_REPLY):
4667 error = receive_unlock_reply(ls, ms);
4670 case cpu_to_le32(DLM_MSG_CANCEL_REPLY):
4671 error = receive_cancel_reply(ls, ms);
4674 /* messages sent from a master node (only two types of async msg) */
4676 case cpu_to_le32(DLM_MSG_GRANT):
4678 error = receive_grant(ls, ms);
4681 case cpu_to_le32(DLM_MSG_BAST):
4683 error = receive_bast(ls, ms);
4686 /* messages sent to a dir node */
4688 case cpu_to_le32(DLM_MSG_LOOKUP):
4689 receive_lookup(ls, ms);
4692 case cpu_to_le32(DLM_MSG_REMOVE):
4693 receive_remove(ls, ms);
4696 /* messages sent from a dir node (remove has no reply) */
4698 case cpu_to_le32(DLM_MSG_LOOKUP_REPLY):
4699 receive_lookup_reply(ls, ms);
4702 /* other messages */
4704 case cpu_to_le32(DLM_MSG_PURGE):
4705 receive_purge(ls, ms);
4709 log_error(ls, "unknown message type %d",
4710 le32_to_cpu(ms->m_type));
4714 * When checking for ENOENT, we're checking the result of
4715 * find_lkb(m_remid):
4717 * The lock id referenced in the message wasn't found. This may
4718 * happen in normal usage for the async messages and cancel, so
4719 * only use log_debug for them.
4721 * Some errors are expected and normal.
4724 if (error == -ENOENT && noent) {
4725 log_debug(ls, "receive %d no %x remote %d %x saved_seq %u",
4726 le32_to_cpu(ms->m_type), le32_to_cpu(ms->m_remid),
4727 le32_to_cpu(ms->m_header.h_nodeid),
4728 le32_to_cpu(ms->m_lkid), saved_seq);
4729 } else if (error == -ENOENT) {
4730 log_error(ls, "receive %d no %x remote %d %x saved_seq %u",
4731 le32_to_cpu(ms->m_type), le32_to_cpu(ms->m_remid),
4732 le32_to_cpu(ms->m_header.h_nodeid),
4733 le32_to_cpu(ms->m_lkid), saved_seq);
4735 if (ms->m_type == cpu_to_le32(DLM_MSG_CONVERT))
4736 dlm_dump_rsb_hash(ls, le32_to_cpu(ms->m_hash));
4739 if (error == -EINVAL) {
4740 log_error(ls, "receive %d inval from %d lkid %x remid %x "
4742 le32_to_cpu(ms->m_type),
4743 le32_to_cpu(ms->m_header.h_nodeid),
4744 le32_to_cpu(ms->m_lkid), le32_to_cpu(ms->m_remid),
4749 /* If the lockspace is in recovery mode (locking stopped), then normal
4750 messages are saved on the requestqueue for processing after recovery is
4751 done. When not in recovery mode, we wait for dlm_recoverd to drain saved
4752 messages off the requestqueue before we process new ones. This occurs right
4753 after recovery completes when we transition from saving all messages on
4754 requestqueue, to processing all the saved messages, to processing new
4755 messages as they arrive. */
4757 static void dlm_receive_message(struct dlm_ls *ls, const struct dlm_message *ms,
4760 if (dlm_locking_stopped(ls)) {
4761 /* If we were a member of this lockspace, left, and rejoined,
4762 other nodes may still be sending us messages from the
4763 lockspace generation before we left. */
4764 if (WARN_ON_ONCE(!ls->ls_generation)) {
4765 log_limit(ls, "receive %d from %d ignore old gen",
4766 le32_to_cpu(ms->m_type), nodeid);
4770 dlm_add_requestqueue(ls, nodeid, ms);
4772 dlm_wait_requestqueue(ls);
4773 _receive_message(ls, ms, 0);
4777 /* This is called by dlm_recoverd to process messages that were saved on
4778 the requestqueue. */
4780 void dlm_receive_message_saved(struct dlm_ls *ls, const struct dlm_message *ms,
4783 _receive_message(ls, ms, saved_seq);
4786 /* This is called by the midcomms layer when something is received for
4787 the lockspace. It could be either a MSG (normal message sent as part of
4788 standard locking activity) or an RCOM (recovery message sent as part of
4789 lockspace recovery). */
4791 void dlm_receive_buffer(const union dlm_packet *p, int nodeid)
4793 const struct dlm_header *hd = &p->header;
4797 switch (hd->h_cmd) {
4799 type = le32_to_cpu(p->message.m_type);
4802 type = le32_to_cpu(p->rcom.rc_type);
4805 log_print("invalid h_cmd %d from %u", hd->h_cmd, nodeid);
4809 if (le32_to_cpu(hd->h_nodeid) != nodeid) {
4810 log_print("invalid h_nodeid %d from %d lockspace %x",
4811 le32_to_cpu(hd->h_nodeid), nodeid,
4812 le32_to_cpu(hd->u.h_lockspace));
4816 ls = dlm_find_lockspace_global(le32_to_cpu(hd->u.h_lockspace));
4818 if (dlm_config.ci_log_debug) {
4819 printk_ratelimited(KERN_DEBUG "dlm: invalid lockspace "
4820 "%u from %d cmd %d type %d\n",
4821 le32_to_cpu(hd->u.h_lockspace), nodeid,
4825 if (hd->h_cmd == DLM_RCOM && type == DLM_RCOM_STATUS)
4826 dlm_send_ls_not_ready(nodeid, &p->rcom);
4830 /* this rwsem allows dlm_ls_stop() to wait for all dlm_recv threads to
4831 be inactive (in this ls) before transitioning to recovery mode */
4833 down_read(&ls->ls_recv_active);
4834 if (hd->h_cmd == DLM_MSG)
4835 dlm_receive_message(ls, &p->message, nodeid);
4836 else if (hd->h_cmd == DLM_RCOM)
4837 dlm_receive_rcom(ls, &p->rcom, nodeid);
4839 log_error(ls, "invalid h_cmd %d from %d lockspace %x",
4840 hd->h_cmd, nodeid, le32_to_cpu(hd->u.h_lockspace));
4841 up_read(&ls->ls_recv_active);
4843 dlm_put_lockspace(ls);
4846 static void recover_convert_waiter(struct dlm_ls *ls, struct dlm_lkb *lkb,
4847 struct dlm_message *ms_local)
4849 if (middle_conversion(lkb)) {
4851 memset(ms_local, 0, sizeof(struct dlm_message));
4852 ms_local->m_type = cpu_to_le32(DLM_MSG_CONVERT_REPLY);
4853 ms_local->m_result = cpu_to_le32(to_dlm_errno(-EINPROGRESS));
4854 ms_local->m_header.h_nodeid = cpu_to_le32(lkb->lkb_nodeid);
4855 _receive_convert_reply(lkb, ms_local, true);
4857 /* Same special case as in receive_rcom_lock_args() */
4858 lkb->lkb_grmode = DLM_LOCK_IV;
4859 rsb_set_flag(lkb->lkb_resource, RSB_RECOVER_CONVERT);
4862 } else if (lkb->lkb_rqmode >= lkb->lkb_grmode) {
4863 set_bit(DLM_IFL_RESEND_BIT, &lkb->lkb_iflags);
4866 /* lkb->lkb_rqmode < lkb->lkb_grmode shouldn't happen since down
4867 conversions are async; there's no reply from the remote master */
4870 /* A waiting lkb needs recovery if the master node has failed, or
4871 the master node is changing (only when no directory is used) */
4873 static int waiter_needs_recovery(struct dlm_ls *ls, struct dlm_lkb *lkb,
4876 if (dlm_no_directory(ls))
4879 if (dlm_is_removed(ls, lkb->lkb_wait_nodeid))
4885 /* Recovery for locks that are waiting for replies from nodes that are now
4886 gone. We can just complete unlocks and cancels by faking a reply from the
4887 dead node. Requests and up-conversions we flag to be resent after
4888 recovery. Down-conversions can just be completed with a fake reply like
4889 unlocks. Conversions between PR and CW need special attention. */
4891 void dlm_recover_waiters_pre(struct dlm_ls *ls)
4893 struct dlm_lkb *lkb, *safe;
4894 struct dlm_message *ms_local;
4895 int wait_type, local_unlock_result, local_cancel_result;
4898 ms_local = kmalloc(sizeof(*ms_local), GFP_KERNEL);
4902 mutex_lock(&ls->ls_waiters_mutex);
4904 list_for_each_entry_safe(lkb, safe, &ls->ls_waiters, lkb_wait_reply) {
4906 dir_nodeid = dlm_dir_nodeid(lkb->lkb_resource);
4908 /* exclude debug messages about unlocks because there can be so
4909 many and they aren't very interesting */
4911 if (lkb->lkb_wait_type != DLM_MSG_UNLOCK) {
4912 log_debug(ls, "waiter %x remote %x msg %d r_nodeid %d "
4913 "lkb_nodeid %d wait_nodeid %d dir_nodeid %d",
4917 lkb->lkb_resource->res_nodeid,
4919 lkb->lkb_wait_nodeid,
4923 /* all outstanding lookups, regardless of destination will be
4924 resent after recovery is done */
4926 if (lkb->lkb_wait_type == DLM_MSG_LOOKUP) {
4927 set_bit(DLM_IFL_RESEND_BIT, &lkb->lkb_iflags);
4931 if (!waiter_needs_recovery(ls, lkb, dir_nodeid))
4934 wait_type = lkb->lkb_wait_type;
4935 local_unlock_result = -DLM_EUNLOCK;
4936 local_cancel_result = -DLM_ECANCEL;
4938 /* Main reply may have been received leaving a zero wait_type,
4939 but a reply for the overlapping op may not have been
4940 received. In that case we need to fake the appropriate
4941 reply for the overlap op. */
4944 if (is_overlap_cancel(lkb)) {
4945 wait_type = DLM_MSG_CANCEL;
4946 if (lkb->lkb_grmode == DLM_LOCK_IV)
4947 local_cancel_result = 0;
4949 if (is_overlap_unlock(lkb)) {
4950 wait_type = DLM_MSG_UNLOCK;
4951 if (lkb->lkb_grmode == DLM_LOCK_IV)
4952 local_unlock_result = -ENOENT;
4955 log_debug(ls, "rwpre overlap %x %x %d %d %d",
4956 lkb->lkb_id, dlm_iflags_val(lkb), wait_type,
4957 local_cancel_result, local_unlock_result);
4960 switch (wait_type) {
4962 case DLM_MSG_REQUEST:
4963 set_bit(DLM_IFL_RESEND_BIT, &lkb->lkb_iflags);
4966 case DLM_MSG_CONVERT:
4967 recover_convert_waiter(ls, lkb, ms_local);
4970 case DLM_MSG_UNLOCK:
4972 memset(ms_local, 0, sizeof(struct dlm_message));
4973 ms_local->m_type = cpu_to_le32(DLM_MSG_UNLOCK_REPLY);
4974 ms_local->m_result = cpu_to_le32(to_dlm_errno(local_unlock_result));
4975 ms_local->m_header.h_nodeid = cpu_to_le32(lkb->lkb_nodeid);
4976 _receive_unlock_reply(lkb, ms_local, true);
4980 case DLM_MSG_CANCEL:
4982 memset(ms_local, 0, sizeof(struct dlm_message));
4983 ms_local->m_type = cpu_to_le32(DLM_MSG_CANCEL_REPLY);
4984 ms_local->m_result = cpu_to_le32(to_dlm_errno(local_cancel_result));
4985 ms_local->m_header.h_nodeid = cpu_to_le32(lkb->lkb_nodeid);
4986 _receive_cancel_reply(lkb, ms_local, true);
4991 log_error(ls, "invalid lkb wait_type %d %d",
4992 lkb->lkb_wait_type, wait_type);
4996 mutex_unlock(&ls->ls_waiters_mutex);
5000 static struct dlm_lkb *find_resend_waiter(struct dlm_ls *ls)
5002 struct dlm_lkb *lkb = NULL, *iter;
5004 mutex_lock(&ls->ls_waiters_mutex);
5005 list_for_each_entry(iter, &ls->ls_waiters, lkb_wait_reply) {
5006 if (test_bit(DLM_IFL_RESEND_BIT, &iter->lkb_iflags)) {
5012 mutex_unlock(&ls->ls_waiters_mutex);
5018 * Forced state reset for locks that were in the middle of remote operations
5019 * when recovery happened (i.e. lkbs that were on the waiters list, waiting
5020 * for a reply from a remote operation.) The lkbs remaining on the waiters
5021 * list need to be reevaluated; some may need resending to a different node
5022 * than previously, and some may now need local handling rather than remote.
5024 * First, the lkb state for the voided remote operation is forcibly reset,
5025 * equivalent to what remove_from_waiters() would normally do:
5026 * . lkb removed from ls_waiters list
5027 * . lkb wait_type cleared
5028 * . lkb waiters_count cleared
5029 * . lkb ref count decremented for each waiters_count (almost always 1,
5030 * but possibly 2 in case of cancel/unlock overlapping, which means
5031 * two remote replies were being expected for the lkb.)
5033 * Second, the lkb is reprocessed like an original operation would be,
5034 * by passing it to _request_lock or _convert_lock, which will either
5035 * process the lkb operation locally, or send it to a remote node again
5036 * and put the lkb back onto the waiters list.
5038 * When reprocessing the lkb, we may find that it's flagged for an overlapping
5039 * force-unlock or cancel, either from before recovery began, or after recovery
5040 * finished. If this is the case, the unlock/cancel is done directly, and the
5041 * original operation is not initiated again (no _request_lock/_convert_lock.)
5044 int dlm_recover_waiters_post(struct dlm_ls *ls)
5046 struct dlm_lkb *lkb;
5048 int error = 0, mstype, err, oc, ou;
5051 if (dlm_locking_stopped(ls)) {
5052 log_debug(ls, "recover_waiters_post aborted");
5058 * Find an lkb from the waiters list that's been affected by
5059 * recovery node changes, and needs to be reprocessed. Does
5060 * hold_lkb(), adding a refcount.
5062 lkb = find_resend_waiter(ls);
5066 r = lkb->lkb_resource;
5071 * If the lkb has been flagged for a force unlock or cancel,
5072 * then the reprocessing below will be replaced by just doing
5073 * the unlock/cancel directly.
5075 mstype = lkb->lkb_wait_type;
5076 oc = test_and_clear_bit(DLM_IFL_OVERLAP_CANCEL_BIT,
5078 ou = test_and_clear_bit(DLM_IFL_OVERLAP_UNLOCK_BIT,
5082 log_debug(ls, "waiter %x remote %x msg %d r_nodeid %d "
5083 "lkb_nodeid %d wait_nodeid %d dir_nodeid %d "
5084 "overlap %d %d", lkb->lkb_id, lkb->lkb_remid, mstype,
5085 r->res_nodeid, lkb->lkb_nodeid, lkb->lkb_wait_nodeid,
5086 dlm_dir_nodeid(r), oc, ou);
5089 * No reply to the pre-recovery operation will now be received,
5090 * so a forced equivalent of remove_from_waiters() is needed to
5091 * reset the waiters state that was in place before recovery.
5094 clear_bit(DLM_IFL_RESEND_BIT, &lkb->lkb_iflags);
5096 /* Forcibly clear wait_type */
5097 lkb->lkb_wait_type = 0;
5100 * Forcibly reset wait_count and associated refcount. The
5101 * wait_count will almost always be 1, but in case of an
5102 * overlapping unlock/cancel it could be 2: see where
5103 * add_to_waiters() finds the lkb is already on the waiters
5104 * list and does lkb_wait_count++; hold_lkb().
5106 while (lkb->lkb_wait_count) {
5107 lkb->lkb_wait_count--;
5111 /* Forcibly remove from waiters list */
5112 mutex_lock(&ls->ls_waiters_mutex);
5113 list_del_init(&lkb->lkb_wait_reply);
5114 mutex_unlock(&ls->ls_waiters_mutex);
5117 * The lkb is now clear of all prior waiters state and can be
5118 * processed locally, or sent to remote node again, or directly
5119 * cancelled/unlocked.
5123 /* do an unlock or cancel instead of resending */
5125 case DLM_MSG_LOOKUP:
5126 case DLM_MSG_REQUEST:
5127 queue_cast(r, lkb, ou ? -DLM_EUNLOCK :
5129 unhold_lkb(lkb); /* undoes create_lkb() */
5131 case DLM_MSG_CONVERT:
5133 queue_cast(r, lkb, -DLM_ECANCEL);
5135 lkb->lkb_exflags |= DLM_LKF_FORCEUNLOCK;
5136 _unlock_lock(r, lkb);
5144 case DLM_MSG_LOOKUP:
5145 case DLM_MSG_REQUEST:
5146 _request_lock(r, lkb);
5148 confirm_master(r, 0);
5150 case DLM_MSG_CONVERT:
5151 _convert_lock(r, lkb);
5159 log_error(ls, "waiter %x msg %d r_nodeid %d "
5160 "dir_nodeid %d overlap %d %d",
5161 lkb->lkb_id, mstype, r->res_nodeid,
5162 dlm_dir_nodeid(r), oc, ou);
5172 static void purge_mstcpy_list(struct dlm_ls *ls, struct dlm_rsb *r,
5173 struct list_head *list)
5175 struct dlm_lkb *lkb, *safe;
5177 list_for_each_entry_safe(lkb, safe, list, lkb_statequeue) {
5178 if (!is_master_copy(lkb))
5181 /* don't purge lkbs we've added in recover_master_copy for
5182 the current recovery seq */
5184 if (lkb->lkb_recover_seq == ls->ls_recover_seq)
5189 /* this put should free the lkb */
5190 if (!dlm_put_lkb(lkb))
5191 log_error(ls, "purged mstcpy lkb not released");
5195 void dlm_purge_mstcpy_locks(struct dlm_rsb *r)
5197 struct dlm_ls *ls = r->res_ls;
5199 purge_mstcpy_list(ls, r, &r->res_grantqueue);
5200 purge_mstcpy_list(ls, r, &r->res_convertqueue);
5201 purge_mstcpy_list(ls, r, &r->res_waitqueue);
5204 static void purge_dead_list(struct dlm_ls *ls, struct dlm_rsb *r,
5205 struct list_head *list,
5206 int nodeid_gone, unsigned int *count)
5208 struct dlm_lkb *lkb, *safe;
5210 list_for_each_entry_safe(lkb, safe, list, lkb_statequeue) {
5211 if (!is_master_copy(lkb))
5214 if ((lkb->lkb_nodeid == nodeid_gone) ||
5215 dlm_is_removed(ls, lkb->lkb_nodeid)) {
5217 /* tell recover_lvb to invalidate the lvb
5218 because a node holding EX/PW failed */
5219 if ((lkb->lkb_exflags & DLM_LKF_VALBLK) &&
5220 (lkb->lkb_grmode >= DLM_LOCK_PW)) {
5221 rsb_set_flag(r, RSB_RECOVER_LVB_INVAL);
5226 /* this put should free the lkb */
5227 if (!dlm_put_lkb(lkb))
5228 log_error(ls, "purged dead lkb not released");
5230 rsb_set_flag(r, RSB_RECOVER_GRANT);
5237 /* Get rid of locks held by nodes that are gone. */
5239 void dlm_recover_purge(struct dlm_ls *ls)
5242 struct dlm_member *memb;
5243 int nodes_count = 0;
5244 int nodeid_gone = 0;
5245 unsigned int lkb_count = 0;
5247 /* cache one removed nodeid to optimize the common
5248 case of a single node removed */
5250 list_for_each_entry(memb, &ls->ls_nodes_gone, list) {
5252 nodeid_gone = memb->nodeid;
5258 down_write(&ls->ls_root_sem);
5259 list_for_each_entry(r, &ls->ls_root_list, res_root_list) {
5263 purge_dead_list(ls, r, &r->res_grantqueue,
5264 nodeid_gone, &lkb_count);
5265 purge_dead_list(ls, r, &r->res_convertqueue,
5266 nodeid_gone, &lkb_count);
5267 purge_dead_list(ls, r, &r->res_waitqueue,
5268 nodeid_gone, &lkb_count);
5274 up_write(&ls->ls_root_sem);
5277 log_rinfo(ls, "dlm_recover_purge %u locks for %u nodes",
5278 lkb_count, nodes_count);
5281 static struct dlm_rsb *find_grant_rsb(struct dlm_ls *ls, int bucket)
5286 spin_lock(&ls->ls_rsbtbl[bucket].lock);
5287 for (n = rb_first(&ls->ls_rsbtbl[bucket].keep); n; n = rb_next(n)) {
5288 r = rb_entry(n, struct dlm_rsb, res_hashnode);
5290 if (!rsb_flag(r, RSB_RECOVER_GRANT))
5292 if (!is_master(r)) {
5293 rsb_clear_flag(r, RSB_RECOVER_GRANT);
5297 spin_unlock(&ls->ls_rsbtbl[bucket].lock);
5300 spin_unlock(&ls->ls_rsbtbl[bucket].lock);
5305 * Attempt to grant locks on resources that we are the master of.
5306 * Locks may have become grantable during recovery because locks
5307 * from departed nodes have been purged (or not rebuilt), allowing
5308 * previously blocked locks to now be granted. The subset of rsb's
5309 * we are interested in are those with lkb's on either the convert or
5312 * Simplest would be to go through each master rsb and check for non-empty
5313 * convert or waiting queues, and attempt to grant on those rsbs.
5314 * Checking the queues requires lock_rsb, though, for which we'd need
5315 * to release the rsbtbl lock. This would make iterating through all
5316 * rsb's very inefficient. So, we rely on earlier recovery routines
5317 * to set RECOVER_GRANT on any rsb's that we should attempt to grant
5321 void dlm_recover_grant(struct dlm_ls *ls)
5325 unsigned int count = 0;
5326 unsigned int rsb_count = 0;
5327 unsigned int lkb_count = 0;
5330 r = find_grant_rsb(ls, bucket);
5332 if (bucket == ls->ls_rsbtbl_size - 1)
5340 /* the RECOVER_GRANT flag is checked in the grant path */
5341 grant_pending_locks(r, &count);
5342 rsb_clear_flag(r, RSB_RECOVER_GRANT);
5344 confirm_master(r, 0);
5351 log_rinfo(ls, "dlm_recover_grant %u locks on %u resources",
5352 lkb_count, rsb_count);
5355 static struct dlm_lkb *search_remid_list(struct list_head *head, int nodeid,
5358 struct dlm_lkb *lkb;
5360 list_for_each_entry(lkb, head, lkb_statequeue) {
5361 if (lkb->lkb_nodeid == nodeid && lkb->lkb_remid == remid)
5367 static struct dlm_lkb *search_remid(struct dlm_rsb *r, int nodeid,
5370 struct dlm_lkb *lkb;
5372 lkb = search_remid_list(&r->res_grantqueue, nodeid, remid);
5375 lkb = search_remid_list(&r->res_convertqueue, nodeid, remid);
5378 lkb = search_remid_list(&r->res_waitqueue, nodeid, remid);
5384 /* needs at least dlm_rcom + rcom_lock */
5385 static int receive_rcom_lock_args(struct dlm_ls *ls, struct dlm_lkb *lkb,
5386 struct dlm_rsb *r, const struct dlm_rcom *rc)
5388 struct rcom_lock *rl = (struct rcom_lock *) rc->rc_buf;
5390 lkb->lkb_nodeid = le32_to_cpu(rc->rc_header.h_nodeid);
5391 lkb->lkb_ownpid = le32_to_cpu(rl->rl_ownpid);
5392 lkb->lkb_remid = le32_to_cpu(rl->rl_lkid);
5393 lkb->lkb_exflags = le32_to_cpu(rl->rl_exflags);
5394 dlm_set_dflags_val(lkb, le32_to_cpu(rl->rl_flags));
5395 set_bit(DLM_IFL_MSTCPY_BIT, &lkb->lkb_iflags);
5396 lkb->lkb_lvbseq = le32_to_cpu(rl->rl_lvbseq);
5397 lkb->lkb_rqmode = rl->rl_rqmode;
5398 lkb->lkb_grmode = rl->rl_grmode;
5399 /* don't set lkb_status because add_lkb wants to itself */
5401 lkb->lkb_bastfn = (rl->rl_asts & DLM_CB_BAST) ? &fake_bastfn : NULL;
5402 lkb->lkb_astfn = (rl->rl_asts & DLM_CB_CAST) ? &fake_astfn : NULL;
5404 if (lkb->lkb_exflags & DLM_LKF_VALBLK) {
5405 int lvblen = le16_to_cpu(rc->rc_header.h_length) -
5406 sizeof(struct dlm_rcom) - sizeof(struct rcom_lock);
5407 if (lvblen > ls->ls_lvblen)
5409 lkb->lkb_lvbptr = dlm_allocate_lvb(ls);
5410 if (!lkb->lkb_lvbptr)
5412 memcpy(lkb->lkb_lvbptr, rl->rl_lvb, lvblen);
5415 /* Conversions between PR and CW (middle modes) need special handling.
5416 The real granted mode of these converting locks cannot be determined
5417 until all locks have been rebuilt on the rsb (recover_conversion) */
5419 if (rl->rl_wait_type == cpu_to_le16(DLM_MSG_CONVERT) &&
5420 middle_conversion(lkb)) {
5421 rl->rl_status = DLM_LKSTS_CONVERT;
5422 lkb->lkb_grmode = DLM_LOCK_IV;
5423 rsb_set_flag(r, RSB_RECOVER_CONVERT);
5429 /* This lkb may have been recovered in a previous aborted recovery so we need
5430 to check if the rsb already has an lkb with the given remote nodeid/lkid.
5431 If so we just send back a standard reply. If not, we create a new lkb with
5432 the given values and send back our lkid. We send back our lkid by sending
5433 back the rcom_lock struct we got but with the remid field filled in. */
5435 /* needs at least dlm_rcom + rcom_lock */
5436 int dlm_recover_master_copy(struct dlm_ls *ls, const struct dlm_rcom *rc,
5437 __le32 *rl_remid, __le32 *rl_result)
5439 struct rcom_lock *rl = (struct rcom_lock *) rc->rc_buf;
5441 struct dlm_lkb *lkb;
5443 int from_nodeid = le32_to_cpu(rc->rc_header.h_nodeid);
5446 /* init rl_remid with rcom lock rl_remid */
5447 *rl_remid = rl->rl_remid;
5449 if (rl->rl_parent_lkid) {
5450 error = -EOPNOTSUPP;
5454 remid = le32_to_cpu(rl->rl_lkid);
5456 /* In general we expect the rsb returned to be R_MASTER, but we don't
5457 have to require it. Recovery of masters on one node can overlap
5458 recovery of locks on another node, so one node can send us MSTCPY
5459 locks before we've made ourselves master of this rsb. We can still
5460 add new MSTCPY locks that we receive here without any harm; when
5461 we make ourselves master, dlm_recover_masters() won't touch the
5462 MSTCPY locks we've received early. */
5464 error = find_rsb(ls, rl->rl_name, le16_to_cpu(rl->rl_namelen),
5465 from_nodeid, R_RECEIVE_RECOVER, &r);
5471 if (dlm_no_directory(ls) && (dlm_dir_nodeid(r) != dlm_our_nodeid())) {
5472 log_error(ls, "dlm_recover_master_copy remote %d %x not dir",
5473 from_nodeid, remid);
5478 lkb = search_remid(r, from_nodeid, remid);
5484 error = create_lkb(ls, &lkb);
5488 error = receive_rcom_lock_args(ls, lkb, r, rc);
5495 add_lkb(r, lkb, rl->rl_status);
5496 ls->ls_recover_locks_in++;
5498 if (!list_empty(&r->res_waitqueue) || !list_empty(&r->res_convertqueue))
5499 rsb_set_flag(r, RSB_RECOVER_GRANT);
5502 /* this is the new value returned to the lock holder for
5503 saving in its process-copy lkb */
5504 *rl_remid = cpu_to_le32(lkb->lkb_id);
5506 lkb->lkb_recover_seq = ls->ls_recover_seq;
5512 if (error && error != -EEXIST)
5513 log_rinfo(ls, "dlm_recover_master_copy remote %d %x error %d",
5514 from_nodeid, remid, error);
5515 *rl_result = cpu_to_le32(error);
5519 /* needs at least dlm_rcom + rcom_lock */
5520 int dlm_recover_process_copy(struct dlm_ls *ls, const struct dlm_rcom *rc,
5523 struct rcom_lock *rl = (struct rcom_lock *) rc->rc_buf;
5525 struct dlm_lkb *lkb;
5526 uint32_t lkid, remid;
5529 lkid = le32_to_cpu(rl->rl_lkid);
5530 remid = le32_to_cpu(rl->rl_remid);
5531 result = le32_to_cpu(rl->rl_result);
5533 error = find_lkb(ls, lkid, &lkb);
5535 log_error(ls, "dlm_recover_process_copy no %x remote %d %x %d",
5536 lkid, le32_to_cpu(rc->rc_header.h_nodeid), remid,
5541 r = lkb->lkb_resource;
5545 if (!is_process_copy(lkb)) {
5546 log_error(ls, "dlm_recover_process_copy bad %x remote %d %x %d",
5547 lkid, le32_to_cpu(rc->rc_header.h_nodeid), remid,
5558 /* There's a chance the new master received our lock before
5559 dlm_recover_master_reply(), this wouldn't happen if we did
5560 a barrier between recover_masters and recover_locks. */
5562 log_debug(ls, "dlm_recover_process_copy %x remote %d %x %d",
5563 lkid, le32_to_cpu(rc->rc_header.h_nodeid), remid,
5566 dlm_send_rcom_lock(r, lkb, seq);
5570 lkb->lkb_remid = remid;
5573 log_error(ls, "dlm_recover_process_copy %x remote %d %x %d unk",
5574 lkid, le32_to_cpu(rc->rc_header.h_nodeid), remid,
5578 /* an ack for dlm_recover_locks() which waits for replies from
5579 all the locks it sends to new masters */
5580 dlm_recovered_lock(r);
5589 int dlm_user_request(struct dlm_ls *ls, struct dlm_user_args *ua,
5590 int mode, uint32_t flags, void *name, unsigned int namelen)
5592 struct dlm_lkb *lkb;
5593 struct dlm_args args;
5597 dlm_lock_recovery(ls);
5599 error = create_lkb(ls, &lkb);
5605 trace_dlm_lock_start(ls, lkb, name, namelen, mode, flags);
5607 if (flags & DLM_LKF_VALBLK) {
5608 ua->lksb.sb_lvbptr = kzalloc(DLM_USER_LVB_LEN, GFP_NOFS);
5609 if (!ua->lksb.sb_lvbptr) {
5615 error = set_lock_args(mode, &ua->lksb, flags, namelen, fake_astfn, ua,
5616 fake_bastfn, &args);
5618 kfree(ua->lksb.sb_lvbptr);
5619 ua->lksb.sb_lvbptr = NULL;
5624 /* After ua is attached to lkb it will be freed by dlm_free_lkb().
5625 When DLM_DFL_USER_BIT is set, the dlm knows that this is a userspace
5626 lock and that lkb_astparam is the dlm_user_args structure. */
5627 set_bit(DLM_DFL_USER_BIT, &lkb->lkb_dflags);
5628 error = request_lock(ls, lkb, name, namelen, &args);
5643 /* add this new lkb to the per-process list of locks */
5644 spin_lock(&ua->proc->locks_spin);
5646 list_add_tail(&lkb->lkb_ownqueue, &ua->proc->locks);
5647 spin_unlock(&ua->proc->locks_spin);
5650 trace_dlm_lock_end(ls, lkb, name, namelen, mode, flags, error, false);
5654 dlm_unlock_recovery(ls);
5658 int dlm_user_convert(struct dlm_ls *ls, struct dlm_user_args *ua_tmp,
5659 int mode, uint32_t flags, uint32_t lkid, char *lvb_in)
5661 struct dlm_lkb *lkb;
5662 struct dlm_args args;
5663 struct dlm_user_args *ua;
5666 dlm_lock_recovery(ls);
5668 error = find_lkb(ls, lkid, &lkb);
5672 trace_dlm_lock_start(ls, lkb, NULL, 0, mode, flags);
5674 /* user can change the params on its lock when it converts it, or
5675 add an lvb that didn't exist before */
5679 if (flags & DLM_LKF_VALBLK && !ua->lksb.sb_lvbptr) {
5680 ua->lksb.sb_lvbptr = kzalloc(DLM_USER_LVB_LEN, GFP_NOFS);
5681 if (!ua->lksb.sb_lvbptr) {
5686 if (lvb_in && ua->lksb.sb_lvbptr)
5687 memcpy(ua->lksb.sb_lvbptr, lvb_in, DLM_USER_LVB_LEN);
5689 ua->xid = ua_tmp->xid;
5690 ua->castparam = ua_tmp->castparam;
5691 ua->castaddr = ua_tmp->castaddr;
5692 ua->bastparam = ua_tmp->bastparam;
5693 ua->bastaddr = ua_tmp->bastaddr;
5694 ua->user_lksb = ua_tmp->user_lksb;
5696 error = set_lock_args(mode, &ua->lksb, flags, 0, fake_astfn, ua,
5697 fake_bastfn, &args);
5701 error = convert_lock(ls, lkb, &args);
5703 if (error == -EINPROGRESS || error == -EAGAIN || error == -EDEADLK)
5706 trace_dlm_lock_end(ls, lkb, NULL, 0, mode, flags, error, false);
5709 dlm_unlock_recovery(ls);
5715 * The caller asks for an orphan lock on a given resource with a given mode.
5716 * If a matching lock exists, it's moved to the owner's list of locks and
5717 * the lkid is returned.
5720 int dlm_user_adopt_orphan(struct dlm_ls *ls, struct dlm_user_args *ua_tmp,
5721 int mode, uint32_t flags, void *name, unsigned int namelen,
5724 struct dlm_lkb *lkb = NULL, *iter;
5725 struct dlm_user_args *ua;
5726 int found_other_mode = 0;
5729 mutex_lock(&ls->ls_orphans_mutex);
5730 list_for_each_entry(iter, &ls->ls_orphans, lkb_ownqueue) {
5731 if (iter->lkb_resource->res_length != namelen)
5733 if (memcmp(iter->lkb_resource->res_name, name, namelen))
5735 if (iter->lkb_grmode != mode) {
5736 found_other_mode = 1;
5741 list_del_init(&iter->lkb_ownqueue);
5742 clear_bit(DLM_DFL_ORPHAN_BIT, &iter->lkb_dflags);
5743 *lkid = iter->lkb_id;
5746 mutex_unlock(&ls->ls_orphans_mutex);
5748 if (!lkb && found_other_mode) {
5758 lkb->lkb_exflags = flags;
5759 lkb->lkb_ownpid = (int) current->pid;
5763 ua->proc = ua_tmp->proc;
5764 ua->xid = ua_tmp->xid;
5765 ua->castparam = ua_tmp->castparam;
5766 ua->castaddr = ua_tmp->castaddr;
5767 ua->bastparam = ua_tmp->bastparam;
5768 ua->bastaddr = ua_tmp->bastaddr;
5769 ua->user_lksb = ua_tmp->user_lksb;
5772 * The lkb reference from the ls_orphans list was not
5773 * removed above, and is now considered the reference
5774 * for the proc locks list.
5777 spin_lock(&ua->proc->locks_spin);
5778 list_add_tail(&lkb->lkb_ownqueue, &ua->proc->locks);
5779 spin_unlock(&ua->proc->locks_spin);
5785 int dlm_user_unlock(struct dlm_ls *ls, struct dlm_user_args *ua_tmp,
5786 uint32_t flags, uint32_t lkid, char *lvb_in)
5788 struct dlm_lkb *lkb;
5789 struct dlm_args args;
5790 struct dlm_user_args *ua;
5793 dlm_lock_recovery(ls);
5795 error = find_lkb(ls, lkid, &lkb);
5799 trace_dlm_unlock_start(ls, lkb, flags);
5803 if (lvb_in && ua->lksb.sb_lvbptr)
5804 memcpy(ua->lksb.sb_lvbptr, lvb_in, DLM_USER_LVB_LEN);
5805 if (ua_tmp->castparam)
5806 ua->castparam = ua_tmp->castparam;
5807 ua->user_lksb = ua_tmp->user_lksb;
5809 error = set_unlock_args(flags, ua, &args);
5813 error = unlock_lock(ls, lkb, &args);
5815 if (error == -DLM_EUNLOCK)
5817 /* from validate_unlock_args() */
5818 if (error == -EBUSY && (flags & DLM_LKF_FORCEUNLOCK))
5823 spin_lock(&ua->proc->locks_spin);
5824 /* dlm_user_add_cb() may have already taken lkb off the proc list */
5825 if (!list_empty(&lkb->lkb_ownqueue))
5826 list_move(&lkb->lkb_ownqueue, &ua->proc->unlocking);
5827 spin_unlock(&ua->proc->locks_spin);
5829 trace_dlm_unlock_end(ls, lkb, flags, error);
5832 dlm_unlock_recovery(ls);
5837 int dlm_user_cancel(struct dlm_ls *ls, struct dlm_user_args *ua_tmp,
5838 uint32_t flags, uint32_t lkid)
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 trace_dlm_unlock_start(ls, lkb, flags);
5854 if (ua_tmp->castparam)
5855 ua->castparam = ua_tmp->castparam;
5856 ua->user_lksb = ua_tmp->user_lksb;
5858 error = set_unlock_args(flags, ua, &args);
5862 error = cancel_lock(ls, lkb, &args);
5864 if (error == -DLM_ECANCEL)
5866 /* from validate_unlock_args() */
5867 if (error == -EBUSY)
5870 trace_dlm_unlock_end(ls, lkb, flags, error);
5873 dlm_unlock_recovery(ls);
5878 int dlm_user_deadlock(struct dlm_ls *ls, uint32_t flags, uint32_t lkid)
5880 struct dlm_lkb *lkb;
5881 struct dlm_args args;
5882 struct dlm_user_args *ua;
5886 dlm_lock_recovery(ls);
5888 error = find_lkb(ls, lkid, &lkb);
5892 trace_dlm_unlock_start(ls, lkb, flags);
5896 error = set_unlock_args(flags, ua, &args);
5900 /* same as cancel_lock(), but set DEADLOCK_CANCEL after lock_rsb */
5902 r = lkb->lkb_resource;
5906 error = validate_unlock_args(lkb, &args);
5909 set_bit(DLM_IFL_DEADLOCK_CANCEL_BIT, &lkb->lkb_iflags);
5911 error = _cancel_lock(r, lkb);
5916 if (error == -DLM_ECANCEL)
5918 /* from validate_unlock_args() */
5919 if (error == -EBUSY)
5922 trace_dlm_unlock_end(ls, lkb, flags, error);
5925 dlm_unlock_recovery(ls);
5929 /* lkb's that are removed from the waiters list by revert are just left on the
5930 orphans list with the granted orphan locks, to be freed by purge */
5932 static int orphan_proc_lock(struct dlm_ls *ls, struct dlm_lkb *lkb)
5934 struct dlm_args args;
5937 hold_lkb(lkb); /* reference for the ls_orphans list */
5938 mutex_lock(&ls->ls_orphans_mutex);
5939 list_add_tail(&lkb->lkb_ownqueue, &ls->ls_orphans);
5940 mutex_unlock(&ls->ls_orphans_mutex);
5942 set_unlock_args(0, lkb->lkb_ua, &args);
5944 error = cancel_lock(ls, lkb, &args);
5945 if (error == -DLM_ECANCEL)
5950 /* The FORCEUNLOCK flag allows the unlock to go ahead even if the lkb isn't
5951 granted. Regardless of what rsb queue the lock is on, it's removed and
5952 freed. The IVVALBLK flag causes the lvb on the resource to be invalidated
5953 if our lock is PW/EX (it's ignored if our granted mode is smaller.) */
5955 static int unlock_proc_lock(struct dlm_ls *ls, struct dlm_lkb *lkb)
5957 struct dlm_args args;
5960 set_unlock_args(DLM_LKF_FORCEUNLOCK | DLM_LKF_IVVALBLK,
5961 lkb->lkb_ua, &args);
5963 error = unlock_lock(ls, lkb, &args);
5964 if (error == -DLM_EUNLOCK)
5969 /* We have to release clear_proc_locks mutex before calling unlock_proc_lock()
5970 (which does lock_rsb) due to deadlock with receiving a message that does
5971 lock_rsb followed by dlm_user_add_cb() */
5973 static struct dlm_lkb *del_proc_lock(struct dlm_ls *ls,
5974 struct dlm_user_proc *proc)
5976 struct dlm_lkb *lkb = NULL;
5978 spin_lock(&ls->ls_clear_proc_locks);
5979 if (list_empty(&proc->locks))
5982 lkb = list_entry(proc->locks.next, struct dlm_lkb, lkb_ownqueue);
5983 list_del_init(&lkb->lkb_ownqueue);
5985 if (lkb->lkb_exflags & DLM_LKF_PERSISTENT)
5986 set_bit(DLM_DFL_ORPHAN_BIT, &lkb->lkb_dflags);
5988 set_bit(DLM_IFL_DEAD_BIT, &lkb->lkb_iflags);
5990 spin_unlock(&ls->ls_clear_proc_locks);
5994 /* The ls_clear_proc_locks mutex protects against dlm_user_add_cb() which
5995 1) references lkb->ua which we free here and 2) adds lkbs to proc->asts,
5996 which we clear here. */
5998 /* proc CLOSING flag is set so no more device_reads should look at proc->asts
5999 list, and no more device_writes should add lkb's to proc->locks list; so we
6000 shouldn't need to take asts_spin or locks_spin here. this assumes that
6001 device reads/writes/closes are serialized -- FIXME: we may need to serialize
6004 void dlm_clear_proc_locks(struct dlm_ls *ls, struct dlm_user_proc *proc)
6006 struct dlm_lkb *lkb, *safe;
6008 dlm_lock_recovery(ls);
6011 lkb = del_proc_lock(ls, proc);
6014 if (lkb->lkb_exflags & DLM_LKF_PERSISTENT)
6015 orphan_proc_lock(ls, lkb);
6017 unlock_proc_lock(ls, lkb);
6019 /* this removes the reference for the proc->locks list
6020 added by dlm_user_request, it may result in the lkb
6026 spin_lock(&ls->ls_clear_proc_locks);
6028 /* in-progress unlocks */
6029 list_for_each_entry_safe(lkb, safe, &proc->unlocking, lkb_ownqueue) {
6030 list_del_init(&lkb->lkb_ownqueue);
6031 set_bit(DLM_IFL_DEAD_BIT, &lkb->lkb_iflags);
6035 list_for_each_entry_safe(lkb, safe, &proc->asts, lkb_cb_list) {
6036 dlm_purge_lkb_callbacks(lkb);
6037 list_del_init(&lkb->lkb_cb_list);
6041 spin_unlock(&ls->ls_clear_proc_locks);
6042 dlm_unlock_recovery(ls);
6045 static void purge_proc_locks(struct dlm_ls *ls, struct dlm_user_proc *proc)
6047 struct dlm_lkb *lkb, *safe;
6051 spin_lock(&proc->locks_spin);
6052 if (!list_empty(&proc->locks)) {
6053 lkb = list_entry(proc->locks.next, struct dlm_lkb,
6055 list_del_init(&lkb->lkb_ownqueue);
6057 spin_unlock(&proc->locks_spin);
6062 set_bit(DLM_IFL_DEAD_BIT, &lkb->lkb_iflags);
6063 unlock_proc_lock(ls, lkb);
6064 dlm_put_lkb(lkb); /* ref from proc->locks list */
6067 spin_lock(&proc->locks_spin);
6068 list_for_each_entry_safe(lkb, safe, &proc->unlocking, lkb_ownqueue) {
6069 list_del_init(&lkb->lkb_ownqueue);
6070 set_bit(DLM_IFL_DEAD_BIT, &lkb->lkb_iflags);
6073 spin_unlock(&proc->locks_spin);
6075 spin_lock(&proc->asts_spin);
6076 list_for_each_entry_safe(lkb, safe, &proc->asts, lkb_cb_list) {
6077 dlm_purge_lkb_callbacks(lkb);
6078 list_del_init(&lkb->lkb_cb_list);
6081 spin_unlock(&proc->asts_spin);
6084 /* pid of 0 means purge all orphans */
6086 static void do_purge(struct dlm_ls *ls, int nodeid, int pid)
6088 struct dlm_lkb *lkb, *safe;
6090 mutex_lock(&ls->ls_orphans_mutex);
6091 list_for_each_entry_safe(lkb, safe, &ls->ls_orphans, lkb_ownqueue) {
6092 if (pid && lkb->lkb_ownpid != pid)
6094 unlock_proc_lock(ls, lkb);
6095 list_del_init(&lkb->lkb_ownqueue);
6098 mutex_unlock(&ls->ls_orphans_mutex);
6101 static int send_purge(struct dlm_ls *ls, int nodeid, int pid)
6103 struct dlm_message *ms;
6104 struct dlm_mhandle *mh;
6107 error = _create_message(ls, sizeof(struct dlm_message), nodeid,
6108 DLM_MSG_PURGE, &ms, &mh, GFP_NOFS);
6111 ms->m_nodeid = cpu_to_le32(nodeid);
6112 ms->m_pid = cpu_to_le32(pid);
6114 return send_message(mh, ms, NULL, 0);
6117 int dlm_user_purge(struct dlm_ls *ls, struct dlm_user_proc *proc,
6118 int nodeid, int pid)
6122 if (nodeid && (nodeid != dlm_our_nodeid())) {
6123 error = send_purge(ls, nodeid, pid);
6125 dlm_lock_recovery(ls);
6126 if (pid == current->pid)
6127 purge_proc_locks(ls, proc);
6129 do_purge(ls, nodeid, pid);
6130 dlm_unlock_recovery(ls);
6135 /* debug functionality */
6136 int dlm_debug_add_lkb(struct dlm_ls *ls, uint32_t lkb_id, char *name, int len,
6137 int lkb_nodeid, unsigned int lkb_dflags, int lkb_status)
6139 struct dlm_lksb *lksb;
6140 struct dlm_lkb *lkb;
6144 /* we currently can't set a valid user lock */
6145 if (lkb_dflags & BIT(DLM_DFL_USER_BIT))
6148 lksb = kzalloc(sizeof(*lksb), GFP_NOFS);
6152 error = _create_lkb(ls, &lkb, lkb_id, lkb_id + 1);
6158 dlm_set_dflags_val(lkb, lkb_dflags);
6159 lkb->lkb_nodeid = lkb_nodeid;
6160 lkb->lkb_lksb = lksb;
6161 /* user specific pointer, just don't have it NULL for kernel locks */
6162 if (~lkb_dflags & BIT(DLM_DFL_USER_BIT))
6163 lkb->lkb_astparam = (void *)0xDEADBEEF;
6165 error = find_rsb(ls, name, len, 0, R_REQUEST, &r);
6174 add_lkb(r, lkb, lkb_status);
6181 int dlm_debug_add_lkb_to_waiters(struct dlm_ls *ls, uint32_t lkb_id,
6182 int mstype, int to_nodeid)
6184 struct dlm_lkb *lkb;
6187 error = find_lkb(ls, lkb_id, &lkb);
6191 error = add_to_waiters(lkb, mstype, to_nodeid);