1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
6 * Code which implements the kernel side of a minimal userspace
7 * interface to our DLM.
9 * Many of the functions here are pared down versions of dlmglue.c
12 * Copyright (C) 2003, 2004 Oracle. All rights reserved.
14 * This program is free software; you can redistribute it and/or
15 * modify it under the terms of the GNU General Public
16 * License as published by the Free Software Foundation; either
17 * version 2 of the License, or (at your option) any later version.
19 * This program is distributed in the hope that it will be useful,
20 * but WITHOUT ANY WARRANTY; without even the implied warranty of
21 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
22 * General Public License for more details.
24 * You should have received a copy of the GNU General Public
25 * License along with this program; if not, write to the
26 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
27 * Boston, MA 021110-1307, USA.
30 #include <linux/signal.h>
32 #include <linux/module.h>
34 #include <linux/types.h>
35 #include <linux/crc32.h>
38 #include "cluster/nodemanager.h"
39 #include "cluster/heartbeat.h"
40 #include "cluster/tcp.h"
46 #define MLOG_MASK_PREFIX ML_DLMFS
47 #include "cluster/masklog.h"
49 static inline int user_check_wait_flag(struct user_lock_res *lockres,
54 spin_lock(&lockres->l_lock);
55 ret = lockres->l_flags & flag;
56 spin_unlock(&lockres->l_lock);
61 static inline void user_wait_on_busy_lock(struct user_lock_res *lockres)
64 wait_event(lockres->l_event,
65 !user_check_wait_flag(lockres, USER_LOCK_BUSY));
68 static inline void user_wait_on_blocked_lock(struct user_lock_res *lockres)
71 wait_event(lockres->l_event,
72 !user_check_wait_flag(lockres, USER_LOCK_BLOCKED));
75 /* I heart container_of... */
76 static inline struct dlm_ctxt *
77 dlm_ctxt_from_user_lockres(struct user_lock_res *lockres)
79 struct dlmfs_inode_private *ip;
81 ip = container_of(lockres,
82 struct dlmfs_inode_private,
88 user_dlm_inode_from_user_lockres(struct user_lock_res *lockres)
90 struct dlmfs_inode_private *ip;
92 ip = container_of(lockres,
93 struct dlmfs_inode_private,
95 return &ip->ip_vfs_inode;
98 static inline void user_recover_from_dlm_error(struct user_lock_res *lockres)
100 spin_lock(&lockres->l_lock);
101 lockres->l_flags &= ~USER_LOCK_BUSY;
102 spin_unlock(&lockres->l_lock);
105 #define user_log_dlm_error(_func, _stat, _lockres) do { \
106 mlog(ML_ERROR, "Dlm error \"%s\" while calling %s on " \
107 "resource %s: %s\n", dlm_errname(_stat), _func, \
108 _lockres->l_name, dlm_errmsg(_stat)); \
111 /* WARNING: This function lives in a world where the only three lock
112 * levels are EX, PR, and NL. It *will* have to be adjusted when more
113 * lock types are added. */
114 static inline int user_highest_compat_lock_level(int level)
116 int new_level = LKM_EXMODE;
118 if (level == LKM_EXMODE)
119 new_level = LKM_NLMODE;
120 else if (level == LKM_PRMODE)
121 new_level = LKM_PRMODE;
125 static void user_ast(void *opaque)
127 struct user_lock_res *lockres = opaque;
128 struct dlm_lockstatus *lksb;
130 mlog(0, "AST fired for lockres %s\n", lockres->l_name);
132 spin_lock(&lockres->l_lock);
134 lksb = &(lockres->l_lksb);
135 if (lksb->status != DLM_NORMAL) {
136 mlog(ML_ERROR, "lksb status value of %u on lockres %s\n",
137 lksb->status, lockres->l_name);
138 spin_unlock(&lockres->l_lock);
142 /* we're downconverting. */
143 if (lockres->l_requested < lockres->l_level) {
144 if (lockres->l_requested <=
145 user_highest_compat_lock_level(lockres->l_blocking)) {
146 lockres->l_blocking = LKM_NLMODE;
147 lockres->l_flags &= ~USER_LOCK_BLOCKED;
151 lockres->l_level = lockres->l_requested;
152 lockres->l_requested = LKM_IVMODE;
153 lockres->l_flags |= USER_LOCK_ATTACHED;
154 lockres->l_flags &= ~USER_LOCK_BUSY;
156 spin_unlock(&lockres->l_lock);
158 wake_up(&lockres->l_event);
161 static inline void user_dlm_grab_inode_ref(struct user_lock_res *lockres)
164 inode = user_dlm_inode_from_user_lockres(lockres);
169 static void user_dlm_unblock_lock(void *opaque);
171 static void __user_dlm_queue_lockres(struct user_lock_res *lockres)
173 if (!(lockres->l_flags & USER_LOCK_QUEUED)) {
174 user_dlm_grab_inode_ref(lockres);
176 INIT_WORK(&lockres->l_work, user_dlm_unblock_lock,
179 queue_work(user_dlm_worker, &lockres->l_work);
180 lockres->l_flags |= USER_LOCK_QUEUED;
184 static void __user_dlm_cond_queue_lockres(struct user_lock_res *lockres)
188 if (!(lockres->l_flags & USER_LOCK_BLOCKED))
191 switch (lockres->l_blocking) {
193 if (!lockres->l_ex_holders && !lockres->l_ro_holders)
197 if (!lockres->l_ex_holders)
205 __user_dlm_queue_lockres(lockres);
208 static void user_bast(void *opaque, int level)
210 struct user_lock_res *lockres = opaque;
212 mlog(0, "Blocking AST fired for lockres %s. Blocking level %d\n",
213 lockres->l_name, level);
215 spin_lock(&lockres->l_lock);
216 lockres->l_flags |= USER_LOCK_BLOCKED;
217 if (level > lockres->l_blocking)
218 lockres->l_blocking = level;
220 __user_dlm_queue_lockres(lockres);
221 spin_unlock(&lockres->l_lock);
223 wake_up(&lockres->l_event);
226 static void user_unlock_ast(void *opaque, enum dlm_status status)
228 struct user_lock_res *lockres = opaque;
230 mlog(0, "UNLOCK AST called on lock %s\n", lockres->l_name);
232 if (status != DLM_NORMAL)
233 mlog(ML_ERROR, "Dlm returns status %d\n", status);
235 spin_lock(&lockres->l_lock);
236 if (lockres->l_flags & USER_LOCK_IN_TEARDOWN)
237 lockres->l_level = LKM_IVMODE;
239 lockres->l_requested = LKM_IVMODE; /* cancel an
242 lockres->l_flags &= ~USER_LOCK_IN_CANCEL;
243 /* we want the unblock thread to look at it again
245 __user_dlm_queue_lockres(lockres);
248 lockres->l_flags &= ~USER_LOCK_BUSY;
249 spin_unlock(&lockres->l_lock);
251 wake_up(&lockres->l_event);
254 static inline void user_dlm_drop_inode_ref(struct user_lock_res *lockres)
257 inode = user_dlm_inode_from_user_lockres(lockres);
261 static void user_dlm_unblock_lock(void *opaque)
263 int new_level, status;
264 struct user_lock_res *lockres = (struct user_lock_res *) opaque;
265 struct dlm_ctxt *dlm = dlm_ctxt_from_user_lockres(lockres);
267 mlog(0, "processing lockres %s\n", lockres->l_name);
269 spin_lock(&lockres->l_lock);
271 mlog_bug_on_msg(!(lockres->l_flags & USER_LOCK_QUEUED),
272 "Lockres %s, flags 0x%x\n",
273 lockres->l_name, lockres->l_flags);
275 /* notice that we don't clear USER_LOCK_BLOCKED here. If it's
276 * set, we want user_ast clear it. */
277 lockres->l_flags &= ~USER_LOCK_QUEUED;
279 /* It's valid to get here and no longer be blocked - if we get
280 * several basts in a row, we might be queued by the first
281 * one, the unblock thread might run and clear the queued
282 * flag, and finally we might get another bast which re-queues
283 * us before our ast for the downconvert is called. */
284 if (!(lockres->l_flags & USER_LOCK_BLOCKED)) {
285 mlog(0, "Lockres %s, flags 0x%x: queued but not blocking\n",
286 lockres->l_name, lockres->l_flags);
287 spin_unlock(&lockres->l_lock);
291 if (lockres->l_flags & USER_LOCK_IN_TEARDOWN) {
292 mlog(0, "lock is in teardown so we do nothing\n");
293 spin_unlock(&lockres->l_lock);
297 if (lockres->l_flags & USER_LOCK_BUSY) {
298 mlog(0, "BUSY flag detected...\n");
299 if (lockres->l_flags & USER_LOCK_IN_CANCEL) {
300 spin_unlock(&lockres->l_lock);
304 lockres->l_flags |= USER_LOCK_IN_CANCEL;
305 spin_unlock(&lockres->l_lock);
307 status = dlmunlock(dlm,
312 if (status == DLM_CANCELGRANT) {
313 /* If we got this, then the ast was fired
314 * before we could cancel. We cleanup our
315 * state, and restart the function. */
316 spin_lock(&lockres->l_lock);
317 lockres->l_flags &= ~USER_LOCK_IN_CANCEL;
318 spin_unlock(&lockres->l_lock);
319 } else if (status != DLM_NORMAL)
320 user_log_dlm_error("dlmunlock", status, lockres);
324 /* If there are still incompat holders, we can exit safely
325 * without worrying about re-queueing this lock as that will
326 * happen on the last call to user_cluster_unlock. */
327 if ((lockres->l_blocking == LKM_EXMODE)
328 && (lockres->l_ex_holders || lockres->l_ro_holders)) {
329 spin_unlock(&lockres->l_lock);
330 mlog(0, "can't downconvert for ex: ro = %u, ex = %u\n",
331 lockres->l_ro_holders, lockres->l_ex_holders);
335 if ((lockres->l_blocking == LKM_PRMODE)
336 && lockres->l_ex_holders) {
337 spin_unlock(&lockres->l_lock);
338 mlog(0, "can't downconvert for pr: ex = %u\n",
339 lockres->l_ex_holders);
343 /* yay, we can downconvert now. */
344 new_level = user_highest_compat_lock_level(lockres->l_blocking);
345 lockres->l_requested = new_level;
346 lockres->l_flags |= USER_LOCK_BUSY;
347 mlog(0, "Downconvert lock from %d to %d\n",
348 lockres->l_level, new_level);
349 spin_unlock(&lockres->l_lock);
351 /* need lock downconvert request now... */
352 status = dlmlock(dlm,
355 LKM_CONVERT|LKM_VALBLK,
360 if (status != DLM_NORMAL) {
361 user_log_dlm_error("dlmlock", status, lockres);
362 user_recover_from_dlm_error(lockres);
366 user_dlm_drop_inode_ref(lockres);
369 static inline void user_dlm_inc_holders(struct user_lock_res *lockres,
374 lockres->l_ex_holders++;
377 lockres->l_ro_holders++;
384 /* predict what lock level we'll be dropping down to on behalf
385 * of another node, and return true if the currently wanted
386 * level will be compatible with it. */
388 user_may_continue_on_blocked_lock(struct user_lock_res *lockres,
391 BUG_ON(!(lockres->l_flags & USER_LOCK_BLOCKED));
393 return wanted <= user_highest_compat_lock_level(lockres->l_blocking);
396 int user_dlm_cluster_lock(struct user_lock_res *lockres,
400 int status, local_flags;
401 struct dlm_ctxt *dlm = dlm_ctxt_from_user_lockres(lockres);
403 if (level != LKM_EXMODE &&
404 level != LKM_PRMODE) {
405 mlog(ML_ERROR, "lockres %s: invalid request!\n",
411 mlog(0, "lockres %s: asking for %s lock, passed flags = 0x%x\n",
413 (level == LKM_EXMODE) ? "LKM_EXMODE" : "LKM_PRMODE",
417 if (signal_pending(current)) {
418 status = -ERESTARTSYS;
422 spin_lock(&lockres->l_lock);
424 /* We only compare against the currently granted level
425 * here. If the lock is blocked waiting on a downconvert,
426 * we'll get caught below. */
427 if ((lockres->l_flags & USER_LOCK_BUSY) &&
428 (level > lockres->l_level)) {
429 /* is someone sitting in dlm_lock? If so, wait on
431 spin_unlock(&lockres->l_lock);
433 user_wait_on_busy_lock(lockres);
437 if ((lockres->l_flags & USER_LOCK_BLOCKED) &&
438 (!user_may_continue_on_blocked_lock(lockres, level))) {
439 /* is the lock is currently blocked on behalf of
441 spin_unlock(&lockres->l_lock);
443 user_wait_on_blocked_lock(lockres);
447 if (level > lockres->l_level) {
448 local_flags = lkm_flags | LKM_VALBLK;
449 if (lockres->l_level != LKM_IVMODE)
450 local_flags |= LKM_CONVERT;
452 lockres->l_requested = level;
453 lockres->l_flags |= USER_LOCK_BUSY;
454 spin_unlock(&lockres->l_lock);
456 BUG_ON(level == LKM_IVMODE);
457 BUG_ON(level == LKM_NLMODE);
459 mlog(0, "lock %s, get lock from %d to level = %d\n",
460 lockres->l_name, lockres->l_level, level);
462 /* call dlm_lock to upgrade lock now */
463 status = dlmlock(dlm,
471 if (status != DLM_NORMAL) {
472 if ((lkm_flags & LKM_NOQUEUE) &&
473 (status == DLM_NOTQUEUED))
476 user_log_dlm_error("dlmlock", status, lockres);
479 user_recover_from_dlm_error(lockres);
483 mlog(0, "lock %s, successfull return from dlmlock\n",
486 user_wait_on_busy_lock(lockres);
490 user_dlm_inc_holders(lockres, level);
491 spin_unlock(&lockres->l_lock);
493 mlog(0, "lockres %s: Got %s lock!\n", lockres->l_name,
494 (level == LKM_EXMODE) ? "LKM_EXMODE" : "LKM_PRMODE");
501 static inline void user_dlm_dec_holders(struct user_lock_res *lockres,
506 BUG_ON(!lockres->l_ex_holders);
507 lockres->l_ex_holders--;
510 BUG_ON(!lockres->l_ro_holders);
511 lockres->l_ro_holders--;
518 void user_dlm_cluster_unlock(struct user_lock_res *lockres,
521 if (level != LKM_EXMODE &&
522 level != LKM_PRMODE) {
523 mlog(ML_ERROR, "lockres %s: invalid request!\n", lockres->l_name);
527 mlog(0, "lockres %s: dropping %s lock\n", lockres->l_name,
528 (level == LKM_EXMODE) ? "LKM_EXMODE" : "LKM_PRMODE");
530 spin_lock(&lockres->l_lock);
531 user_dlm_dec_holders(lockres, level);
532 __user_dlm_cond_queue_lockres(lockres);
533 spin_unlock(&lockres->l_lock);
536 void user_dlm_write_lvb(struct inode *inode,
540 struct user_lock_res *lockres = &DLMFS_I(inode)->ip_lockres;
541 char *lvb = lockres->l_lksb.lvb;
543 BUG_ON(len > DLM_LVB_LEN);
545 spin_lock(&lockres->l_lock);
547 BUG_ON(lockres->l_level < LKM_EXMODE);
548 memcpy(lvb, val, len);
550 spin_unlock(&lockres->l_lock);
553 void user_dlm_read_lvb(struct inode *inode,
557 struct user_lock_res *lockres = &DLMFS_I(inode)->ip_lockres;
558 char *lvb = lockres->l_lksb.lvb;
560 BUG_ON(len > DLM_LVB_LEN);
562 spin_lock(&lockres->l_lock);
564 BUG_ON(lockres->l_level < LKM_PRMODE);
565 memcpy(val, lvb, len);
567 spin_unlock(&lockres->l_lock);
570 void user_dlm_lock_res_init(struct user_lock_res *lockres,
571 struct dentry *dentry)
573 memset(lockres, 0, sizeof(*lockres));
575 spin_lock_init(&lockres->l_lock);
576 init_waitqueue_head(&lockres->l_event);
577 lockres->l_level = LKM_IVMODE;
578 lockres->l_requested = LKM_IVMODE;
579 lockres->l_blocking = LKM_IVMODE;
581 /* should have been checked before getting here. */
582 BUG_ON(dentry->d_name.len >= USER_DLM_LOCK_ID_MAX_LEN);
584 memcpy(lockres->l_name,
589 int user_dlm_destroy_lock(struct user_lock_res *lockres)
592 struct dlm_ctxt *dlm = dlm_ctxt_from_user_lockres(lockres);
594 mlog(0, "asked to destroy %s\n", lockres->l_name);
596 spin_lock(&lockres->l_lock);
597 while (lockres->l_flags & USER_LOCK_BUSY) {
598 spin_unlock(&lockres->l_lock);
600 mlog(0, "lock %s is busy\n", lockres->l_name);
602 user_wait_on_busy_lock(lockres);
604 spin_lock(&lockres->l_lock);
607 if (lockres->l_ro_holders || lockres->l_ex_holders) {
608 spin_unlock(&lockres->l_lock);
609 mlog(0, "lock %s has holders\n", lockres->l_name);
614 if (!(lockres->l_flags & USER_LOCK_ATTACHED)) {
615 spin_unlock(&lockres->l_lock);
616 mlog(0, "lock %s is not attached\n", lockres->l_name);
620 lockres->l_flags &= ~USER_LOCK_ATTACHED;
621 lockres->l_flags |= USER_LOCK_BUSY;
622 lockres->l_flags |= USER_LOCK_IN_TEARDOWN;
623 spin_unlock(&lockres->l_lock);
625 mlog(0, "unlocking lockres %s\n", lockres->l_name);
626 status = dlmunlock(dlm,
631 if (status != DLM_NORMAL) {
632 user_log_dlm_error("dlmunlock", status, lockres);
637 user_wait_on_busy_lock(lockres);
644 struct dlm_ctxt *user_dlm_register_context(struct qstr *name)
646 struct dlm_ctxt *dlm;
650 domain = kmalloc(name->len + 1, GFP_KERNEL);
653 return ERR_PTR(-ENOMEM);
656 dlm_key = crc32_le(0, name->name, name->len);
658 snprintf(domain, name->len + 1, "%.*s", name->len, name->name);
660 dlm = dlm_register_domain(domain, dlm_key);
662 mlog_errno(PTR_ERR(dlm));
668 void user_dlm_unregister_context(struct dlm_ctxt *dlm)
670 dlm_unregister_domain(dlm);