1 // SPDX-License-Identifier: GPL-2.0-or-later
5 * Code which implements the kernel side of a minimal userspace
6 * interface to our DLM.
8 * Many of the functions here are pared down versions of dlmglue.c
11 * Copyright (C) 2003, 2004 Oracle. All rights reserved.
14 #include <linux/signal.h>
15 #include <linux/sched/signal.h>
17 #include <linux/module.h>
19 #include <linux/types.h>
20 #include <linux/crc32.h>
22 #include "../ocfs2_lockingver.h"
23 #include "../stackglue.h"
26 #define MLOG_MASK_PREFIX ML_DLMFS
27 #include "../cluster/masklog.h"
30 static inline struct user_lock_res *user_lksb_to_lock_res(struct ocfs2_dlm_lksb *lksb)
32 return container_of(lksb, struct user_lock_res, l_lksb);
35 static inline int user_check_wait_flag(struct user_lock_res *lockres,
40 spin_lock(&lockres->l_lock);
41 ret = lockres->l_flags & flag;
42 spin_unlock(&lockres->l_lock);
47 static inline void user_wait_on_busy_lock(struct user_lock_res *lockres)
50 wait_event(lockres->l_event,
51 !user_check_wait_flag(lockres, USER_LOCK_BUSY));
54 static inline void user_wait_on_blocked_lock(struct user_lock_res *lockres)
57 wait_event(lockres->l_event,
58 !user_check_wait_flag(lockres, USER_LOCK_BLOCKED));
61 /* I heart container_of... */
62 static inline struct ocfs2_cluster_connection *
63 cluster_connection_from_user_lockres(struct user_lock_res *lockres)
65 struct dlmfs_inode_private *ip;
67 ip = container_of(lockres,
68 struct dlmfs_inode_private,
74 user_dlm_inode_from_user_lockres(struct user_lock_res *lockres)
76 struct dlmfs_inode_private *ip;
78 ip = container_of(lockres,
79 struct dlmfs_inode_private,
81 return &ip->ip_vfs_inode;
84 static inline void user_recover_from_dlm_error(struct user_lock_res *lockres)
86 spin_lock(&lockres->l_lock);
87 lockres->l_flags &= ~USER_LOCK_BUSY;
88 spin_unlock(&lockres->l_lock);
91 #define user_log_dlm_error(_func, _stat, _lockres) do { \
92 mlog(ML_ERROR, "Dlm error %d while calling %s on " \
93 "resource %.*s\n", _stat, _func, \
94 _lockres->l_namelen, _lockres->l_name); \
97 /* WARNING: This function lives in a world where the only three lock
98 * levels are EX, PR, and NL. It *will* have to be adjusted when more
99 * lock types are added. */
100 static inline int user_highest_compat_lock_level(int level)
102 int new_level = DLM_LOCK_EX;
104 if (level == DLM_LOCK_EX)
105 new_level = DLM_LOCK_NL;
106 else if (level == DLM_LOCK_PR)
107 new_level = DLM_LOCK_PR;
111 static void user_ast(struct ocfs2_dlm_lksb *lksb)
113 struct user_lock_res *lockres = user_lksb_to_lock_res(lksb);
116 mlog(ML_BASTS, "AST fired for lockres %.*s, level %d => %d\n",
117 lockres->l_namelen, lockres->l_name, lockres->l_level,
118 lockres->l_requested);
120 spin_lock(&lockres->l_lock);
122 status = ocfs2_dlm_lock_status(&lockres->l_lksb);
124 mlog(ML_ERROR, "lksb status value of %u on lockres %.*s\n",
125 status, lockres->l_namelen, lockres->l_name);
126 spin_unlock(&lockres->l_lock);
130 mlog_bug_on_msg(lockres->l_requested == DLM_LOCK_IV,
131 "Lockres %.*s, requested ivmode. flags 0x%x\n",
132 lockres->l_namelen, lockres->l_name, lockres->l_flags);
134 /* we're downconverting. */
135 if (lockres->l_requested < lockres->l_level) {
136 if (lockres->l_requested <=
137 user_highest_compat_lock_level(lockres->l_blocking)) {
138 lockres->l_blocking = DLM_LOCK_NL;
139 lockres->l_flags &= ~USER_LOCK_BLOCKED;
143 lockres->l_level = lockres->l_requested;
144 lockres->l_requested = DLM_LOCK_IV;
145 lockres->l_flags |= USER_LOCK_ATTACHED;
146 lockres->l_flags &= ~USER_LOCK_BUSY;
148 spin_unlock(&lockres->l_lock);
150 wake_up(&lockres->l_event);
153 static inline void user_dlm_grab_inode_ref(struct user_lock_res *lockres)
156 inode = user_dlm_inode_from_user_lockres(lockres);
161 static void user_dlm_unblock_lock(struct work_struct *work);
163 static void __user_dlm_queue_lockres(struct user_lock_res *lockres)
165 if (!(lockres->l_flags & USER_LOCK_QUEUED)) {
166 user_dlm_grab_inode_ref(lockres);
168 INIT_WORK(&lockres->l_work, user_dlm_unblock_lock);
170 queue_work(user_dlm_worker, &lockres->l_work);
171 lockres->l_flags |= USER_LOCK_QUEUED;
175 static void __user_dlm_cond_queue_lockres(struct user_lock_res *lockres)
179 if (!(lockres->l_flags & USER_LOCK_BLOCKED))
182 switch (lockres->l_blocking) {
184 if (!lockres->l_ex_holders && !lockres->l_ro_holders)
188 if (!lockres->l_ex_holders)
196 __user_dlm_queue_lockres(lockres);
199 static void user_bast(struct ocfs2_dlm_lksb *lksb, int level)
201 struct user_lock_res *lockres = user_lksb_to_lock_res(lksb);
203 mlog(ML_BASTS, "BAST fired for lockres %.*s, blocking %d, level %d\n",
204 lockres->l_namelen, lockres->l_name, level, lockres->l_level);
206 spin_lock(&lockres->l_lock);
207 lockres->l_flags |= USER_LOCK_BLOCKED;
208 if (level > lockres->l_blocking)
209 lockres->l_blocking = level;
211 __user_dlm_queue_lockres(lockres);
212 spin_unlock(&lockres->l_lock);
214 wake_up(&lockres->l_event);
217 static void user_unlock_ast(struct ocfs2_dlm_lksb *lksb, int status)
219 struct user_lock_res *lockres = user_lksb_to_lock_res(lksb);
221 mlog(ML_BASTS, "UNLOCK AST fired for lockres %.*s, flags 0x%x\n",
222 lockres->l_namelen, lockres->l_name, lockres->l_flags);
225 mlog(ML_ERROR, "dlm returns status %d\n", status);
227 spin_lock(&lockres->l_lock);
228 /* The teardown flag gets set early during the unlock process,
229 * so test the cancel flag to make sure that this ast isn't
230 * for a concurrent cancel. */
231 if (lockres->l_flags & USER_LOCK_IN_TEARDOWN
232 && !(lockres->l_flags & USER_LOCK_IN_CANCEL)) {
233 lockres->l_level = DLM_LOCK_IV;
234 } else if (status == DLM_CANCELGRANT) {
235 /* We tried to cancel a convert request, but it was
236 * already granted. Don't clear the busy flag - the
237 * ast should've done this already. */
238 BUG_ON(!(lockres->l_flags & USER_LOCK_IN_CANCEL));
239 lockres->l_flags &= ~USER_LOCK_IN_CANCEL;
242 BUG_ON(!(lockres->l_flags & USER_LOCK_IN_CANCEL));
243 /* Cancel succeeded, we want to re-queue */
244 lockres->l_requested = DLM_LOCK_IV; /* cancel an
247 lockres->l_flags &= ~USER_LOCK_IN_CANCEL;
248 /* we want the unblock thread to look at it again
250 if (lockres->l_flags & USER_LOCK_BLOCKED)
251 __user_dlm_queue_lockres(lockres);
254 lockres->l_flags &= ~USER_LOCK_BUSY;
256 spin_unlock(&lockres->l_lock);
258 wake_up(&lockres->l_event);
262 * This is the userdlmfs locking protocol version.
264 * See fs/ocfs2/dlmglue.c for more details on locking versions.
266 static struct ocfs2_locking_protocol user_dlm_lproto = {
268 .pv_major = OCFS2_LOCKING_PROTOCOL_MAJOR,
269 .pv_minor = OCFS2_LOCKING_PROTOCOL_MINOR,
271 .lp_lock_ast = user_ast,
272 .lp_blocking_ast = user_bast,
273 .lp_unlock_ast = user_unlock_ast,
276 static inline void user_dlm_drop_inode_ref(struct user_lock_res *lockres)
279 inode = user_dlm_inode_from_user_lockres(lockres);
283 static void user_dlm_unblock_lock(struct work_struct *work)
285 int new_level, status;
286 struct user_lock_res *lockres =
287 container_of(work, struct user_lock_res, l_work);
288 struct ocfs2_cluster_connection *conn =
289 cluster_connection_from_user_lockres(lockres);
291 mlog(0, "lockres %.*s\n", lockres->l_namelen, lockres->l_name);
293 spin_lock(&lockres->l_lock);
295 mlog_bug_on_msg(!(lockres->l_flags & USER_LOCK_QUEUED),
296 "Lockres %.*s, flags 0x%x\n",
297 lockres->l_namelen, lockres->l_name, lockres->l_flags);
299 /* notice that we don't clear USER_LOCK_BLOCKED here. If it's
300 * set, we want user_ast clear it. */
301 lockres->l_flags &= ~USER_LOCK_QUEUED;
303 /* It's valid to get here and no longer be blocked - if we get
304 * several basts in a row, we might be queued by the first
305 * one, the unblock thread might run and clear the queued
306 * flag, and finally we might get another bast which re-queues
307 * us before our ast for the downconvert is called. */
308 if (!(lockres->l_flags & USER_LOCK_BLOCKED)) {
309 mlog(ML_BASTS, "lockres %.*s USER_LOCK_BLOCKED\n",
310 lockres->l_namelen, lockres->l_name);
311 spin_unlock(&lockres->l_lock);
315 if (lockres->l_flags & USER_LOCK_IN_TEARDOWN) {
316 mlog(ML_BASTS, "lockres %.*s USER_LOCK_IN_TEARDOWN\n",
317 lockres->l_namelen, lockres->l_name);
318 spin_unlock(&lockres->l_lock);
322 if (lockres->l_flags & USER_LOCK_BUSY) {
323 if (lockres->l_flags & USER_LOCK_IN_CANCEL) {
324 mlog(ML_BASTS, "lockres %.*s USER_LOCK_IN_CANCEL\n",
325 lockres->l_namelen, lockres->l_name);
326 spin_unlock(&lockres->l_lock);
330 lockres->l_flags |= USER_LOCK_IN_CANCEL;
331 spin_unlock(&lockres->l_lock);
333 status = ocfs2_dlm_unlock(conn, &lockres->l_lksb,
336 user_log_dlm_error("ocfs2_dlm_unlock", status, lockres);
340 /* If there are still incompat holders, we can exit safely
341 * without worrying about re-queueing this lock as that will
342 * happen on the last call to user_cluster_unlock. */
343 if ((lockres->l_blocking == DLM_LOCK_EX)
344 && (lockres->l_ex_holders || lockres->l_ro_holders)) {
345 spin_unlock(&lockres->l_lock);
346 mlog(ML_BASTS, "lockres %.*s, EX/PR Holders %u,%u\n",
347 lockres->l_namelen, lockres->l_name,
348 lockres->l_ex_holders, lockres->l_ro_holders);
352 if ((lockres->l_blocking == DLM_LOCK_PR)
353 && lockres->l_ex_holders) {
354 spin_unlock(&lockres->l_lock);
355 mlog(ML_BASTS, "lockres %.*s, EX Holders %u\n",
356 lockres->l_namelen, lockres->l_name,
357 lockres->l_ex_holders);
361 /* yay, we can downconvert now. */
362 new_level = user_highest_compat_lock_level(lockres->l_blocking);
363 lockres->l_requested = new_level;
364 lockres->l_flags |= USER_LOCK_BUSY;
365 mlog(ML_BASTS, "lockres %.*s, downconvert %d => %d\n",
366 lockres->l_namelen, lockres->l_name, lockres->l_level, new_level);
367 spin_unlock(&lockres->l_lock);
369 /* need lock downconvert request now... */
370 status = ocfs2_dlm_lock(conn, new_level, &lockres->l_lksb,
371 DLM_LKF_CONVERT|DLM_LKF_VALBLK,
375 user_log_dlm_error("ocfs2_dlm_lock", status, lockres);
376 user_recover_from_dlm_error(lockres);
380 user_dlm_drop_inode_ref(lockres);
383 static inline void user_dlm_inc_holders(struct user_lock_res *lockres,
388 lockres->l_ex_holders++;
391 lockres->l_ro_holders++;
398 /* predict what lock level we'll be dropping down to on behalf
399 * of another node, and return true if the currently wanted
400 * level will be compatible with it. */
402 user_may_continue_on_blocked_lock(struct user_lock_res *lockres,
405 BUG_ON(!(lockres->l_flags & USER_LOCK_BLOCKED));
407 return wanted <= user_highest_compat_lock_level(lockres->l_blocking);
410 int user_dlm_cluster_lock(struct user_lock_res *lockres,
414 int status, local_flags;
415 struct ocfs2_cluster_connection *conn =
416 cluster_connection_from_user_lockres(lockres);
418 if (level != DLM_LOCK_EX &&
419 level != DLM_LOCK_PR) {
420 mlog(ML_ERROR, "lockres %.*s: invalid request!\n",
421 lockres->l_namelen, lockres->l_name);
426 mlog(ML_BASTS, "lockres %.*s, level %d, flags = 0x%x\n",
427 lockres->l_namelen, lockres->l_name, level, lkm_flags);
430 if (signal_pending(current)) {
431 status = -ERESTARTSYS;
435 spin_lock(&lockres->l_lock);
436 if (lockres->l_flags & USER_LOCK_IN_TEARDOWN) {
437 spin_unlock(&lockres->l_lock);
442 /* We only compare against the currently granted level
443 * here. If the lock is blocked waiting on a downconvert,
444 * we'll get caught below. */
445 if ((lockres->l_flags & USER_LOCK_BUSY) &&
446 (level > lockres->l_level)) {
447 /* is someone sitting in dlm_lock? If so, wait on
449 spin_unlock(&lockres->l_lock);
451 user_wait_on_busy_lock(lockres);
455 if ((lockres->l_flags & USER_LOCK_BLOCKED) &&
456 (!user_may_continue_on_blocked_lock(lockres, level))) {
457 /* is the lock is currently blocked on behalf of
459 spin_unlock(&lockres->l_lock);
461 user_wait_on_blocked_lock(lockres);
465 if (level > lockres->l_level) {
466 local_flags = lkm_flags | DLM_LKF_VALBLK;
467 if (lockres->l_level != DLM_LOCK_IV)
468 local_flags |= DLM_LKF_CONVERT;
470 lockres->l_requested = level;
471 lockres->l_flags |= USER_LOCK_BUSY;
472 spin_unlock(&lockres->l_lock);
474 BUG_ON(level == DLM_LOCK_IV);
475 BUG_ON(level == DLM_LOCK_NL);
477 /* call dlm_lock to upgrade lock now */
478 status = ocfs2_dlm_lock(conn, level, &lockres->l_lksb,
479 local_flags, lockres->l_name,
482 if ((lkm_flags & DLM_LKF_NOQUEUE) &&
484 user_log_dlm_error("ocfs2_dlm_lock",
486 user_recover_from_dlm_error(lockres);
490 user_wait_on_busy_lock(lockres);
494 user_dlm_inc_holders(lockres, level);
495 spin_unlock(&lockres->l_lock);
502 static inline void user_dlm_dec_holders(struct user_lock_res *lockres,
507 BUG_ON(!lockres->l_ex_holders);
508 lockres->l_ex_holders--;
511 BUG_ON(!lockres->l_ro_holders);
512 lockres->l_ro_holders--;
519 void user_dlm_cluster_unlock(struct user_lock_res *lockres,
522 if (level != DLM_LOCK_EX &&
523 level != DLM_LOCK_PR) {
524 mlog(ML_ERROR, "lockres %.*s: invalid request!\n",
525 lockres->l_namelen, lockres->l_name);
529 spin_lock(&lockres->l_lock);
530 user_dlm_dec_holders(lockres, level);
531 __user_dlm_cond_queue_lockres(lockres);
532 spin_unlock(&lockres->l_lock);
535 void user_dlm_write_lvb(struct inode *inode,
539 struct user_lock_res *lockres = &DLMFS_I(inode)->ip_lockres;
542 BUG_ON(len > DLM_LVB_LEN);
544 spin_lock(&lockres->l_lock);
546 BUG_ON(lockres->l_level < DLM_LOCK_EX);
547 lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
548 memcpy(lvb, val, len);
550 spin_unlock(&lockres->l_lock);
553 bool user_dlm_read_lvb(struct inode *inode, char *val)
555 struct user_lock_res *lockres = &DLMFS_I(inode)->ip_lockres;
559 spin_lock(&lockres->l_lock);
561 BUG_ON(lockres->l_level < DLM_LOCK_PR);
562 if (ocfs2_dlm_lvb_valid(&lockres->l_lksb)) {
563 lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
564 memcpy(val, lvb, DLM_LVB_LEN);
568 spin_unlock(&lockres->l_lock);
572 void user_dlm_lock_res_init(struct user_lock_res *lockres,
573 struct dentry *dentry)
575 memset(lockres, 0, sizeof(*lockres));
577 spin_lock_init(&lockres->l_lock);
578 init_waitqueue_head(&lockres->l_event);
579 lockres->l_level = DLM_LOCK_IV;
580 lockres->l_requested = DLM_LOCK_IV;
581 lockres->l_blocking = DLM_LOCK_IV;
583 /* should have been checked before getting here. */
584 BUG_ON(dentry->d_name.len >= USER_DLM_LOCK_ID_MAX_LEN);
586 memcpy(lockres->l_name,
589 lockres->l_namelen = dentry->d_name.len;
592 int user_dlm_destroy_lock(struct user_lock_res *lockres)
595 struct ocfs2_cluster_connection *conn =
596 cluster_connection_from_user_lockres(lockres);
598 mlog(ML_BASTS, "lockres %.*s\n", lockres->l_namelen, lockres->l_name);
600 spin_lock(&lockres->l_lock);
601 if (lockres->l_flags & USER_LOCK_IN_TEARDOWN) {
602 spin_unlock(&lockres->l_lock);
606 lockres->l_flags |= USER_LOCK_IN_TEARDOWN;
608 while (lockres->l_flags & USER_LOCK_BUSY) {
609 spin_unlock(&lockres->l_lock);
611 user_wait_on_busy_lock(lockres);
613 spin_lock(&lockres->l_lock);
616 if (lockres->l_ro_holders || lockres->l_ex_holders) {
617 lockres->l_flags &= ~USER_LOCK_IN_TEARDOWN;
618 spin_unlock(&lockres->l_lock);
623 if (!(lockres->l_flags & USER_LOCK_ATTACHED)) {
625 * lock is never requested, leave USER_LOCK_IN_TEARDOWN set
626 * to avoid new lock request coming in.
628 spin_unlock(&lockres->l_lock);
632 lockres->l_flags |= USER_LOCK_BUSY;
633 spin_unlock(&lockres->l_lock);
635 status = ocfs2_dlm_unlock(conn, &lockres->l_lksb, DLM_LKF_VALBLK);
637 spin_lock(&lockres->l_lock);
638 lockres->l_flags &= ~USER_LOCK_IN_TEARDOWN;
639 lockres->l_flags &= ~USER_LOCK_BUSY;
640 spin_unlock(&lockres->l_lock);
641 user_log_dlm_error("ocfs2_dlm_unlock", status, lockres);
645 user_wait_on_busy_lock(lockres);
652 static void user_dlm_recovery_handler_noop(int node_num,
655 /* We ignore recovery events */
659 void user_dlm_set_locking_protocol(void)
661 ocfs2_stack_glue_set_max_proto_version(&user_dlm_lproto.lp_max_version);
664 struct ocfs2_cluster_connection *user_dlm_register(const struct qstr *name)
667 struct ocfs2_cluster_connection *conn;
669 rc = ocfs2_cluster_connect_agnostic(name->name, name->len,
671 user_dlm_recovery_handler_noop,
676 return rc ? ERR_PTR(rc) : conn;
679 void user_dlm_unregister(struct ocfs2_cluster_connection *conn)
681 ocfs2_cluster_disconnect(conn, 0);