1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
32 * Copyright (c) 2003 Hewlett-Packard Development Company LP.
33 * Developed under the sponsorship of the US Government under
34 * Subcontract No. B514193
37 * This file is part of Lustre, http://www.lustre.org/
38 * Lustre is a trademark of Sun Microsystems, Inc.
41 #define DEBUG_SUBSYSTEM S_LDLM
44 #include <lustre_dlm.h>
45 #include <obd_support.h>
46 #include <obd_class.h>
47 #include <lustre_lib.h>
48 #include <libcfs/list.h>
50 #include <liblustre.h>
51 #include <obd_class.h>
54 #include "ldlm_internal.h"
56 #define l_flock_waitq l_lru
59 * Wait queue for Posix lock deadlock detection, added with
60 * ldlm_lock::l_flock_waitq.
62 static CFS_LIST_HEAD(ldlm_flock_waitq);
64 * Lock protecting access to ldlm_flock_waitq.
66 spinlock_t ldlm_flock_waitq_lock = SPIN_LOCK_UNLOCKED;
68 int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
69 void *data, int flag);
72 * list_for_remaining_safe - iterate over the remaining entries in a list
73 * and safeguard against removal of a list entry.
74 * @pos: the &struct list_head to use as a loop counter. pos MUST
75 * have been initialized prior to using it in this macro.
76 * @n: another &struct list_head to use as temporary storage
77 * @head: the head for your list.
79 #define list_for_remaining_safe(pos, n, head) \
80 for (n = pos->next; pos != (head); pos = n, n = pos->next)
83 ldlm_same_flock_owner(struct ldlm_lock *lock, struct ldlm_lock *new)
85 return((new->l_policy_data.l_flock.pid ==
86 lock->l_policy_data.l_flock.pid) &&
87 (new->l_export == lock->l_export));
91 ldlm_flocks_overlap(struct ldlm_lock *lock, struct ldlm_lock *new)
93 return((new->l_policy_data.l_flock.start <=
94 lock->l_policy_data.l_flock.end) &&
95 (new->l_policy_data.l_flock.end >=
96 lock->l_policy_data.l_flock.start));
100 ldlm_flock_destroy(struct ldlm_lock *lock, ldlm_mode_t mode, int flags)
104 LDLM_DEBUG(lock, "ldlm_flock_destroy(mode: %d, flags: 0x%x)",
107 /* Safe to not lock here, since it should be empty anyway */
108 LASSERT(list_empty(&lock->l_flock_waitq));
110 list_del_init(&lock->l_res_link);
111 if (flags == LDLM_FL_WAIT_NOREPROC) {
112 /* client side - set a flag to prevent sending a CANCEL */
113 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_CBPENDING;
115 /* when reaching here, it is under lock_res_and_lock(). Thus,
116 need call the nolock version of ldlm_lock_decref_internal*/
117 ldlm_lock_decref_internal_nolock(lock, mode);
120 ldlm_lock_destroy_nolock(lock);
125 ldlm_flock_deadlock(struct ldlm_lock *req, struct ldlm_lock *blocking_lock)
127 struct obd_export *req_export = req->l_export;
128 struct obd_export *blocking_export = blocking_lock->l_export;
129 pid_t req_pid = req->l_policy_data.l_flock.pid;
130 pid_t blocking_pid = blocking_lock->l_policy_data.l_flock.pid;
131 struct ldlm_lock *lock;
133 spin_lock(&ldlm_flock_waitq_lock);
135 list_for_each_entry(lock, &ldlm_flock_waitq, l_flock_waitq) {
136 if ((lock->l_policy_data.l_flock.pid != blocking_pid) ||
137 (lock->l_export != blocking_export))
140 blocking_pid = lock->l_policy_data.l_flock.blocking_pid;
141 blocking_export = (struct obd_export *)(long)
142 lock->l_policy_data.l_flock.blocking_export;
143 if (blocking_pid == req_pid && blocking_export == req_export) {
144 spin_unlock(&ldlm_flock_waitq_lock);
150 spin_unlock(&ldlm_flock_waitq_lock);
156 ldlm_process_flock_lock(struct ldlm_lock *req, int *flags, int first_enq,
157 ldlm_error_t *err, struct list_head *work_list)
159 struct ldlm_resource *res = req->l_resource;
160 struct ldlm_namespace *ns = res->lr_namespace;
161 struct list_head *tmp;
162 struct list_head *ownlocks = NULL;
163 struct ldlm_lock *lock = NULL;
164 struct ldlm_lock *new = req;
165 struct ldlm_lock *new2 = NULL;
166 ldlm_mode_t mode = req->l_req_mode;
167 int local = ns_is_client(ns);
168 int added = (mode == LCK_NL);
171 const struct ldlm_callback_suite null_cbs = { NULL };
174 CDEBUG(D_DLMTRACE, "flags %#x pid %u mode %u start "LPU64" end "LPU64
175 "\n", *flags, new->l_policy_data.l_flock.pid, mode,
176 req->l_policy_data.l_flock.start,
177 req->l_policy_data.l_flock.end);
182 /* No blocking ASTs are sent to the clients for
183 * Posix file & record locks */
184 req->l_blocking_ast = NULL;
186 /* Called on the server for lock cancels. */
187 req->l_blocking_ast = ldlm_flock_blocking_ast;
191 if ((*flags == LDLM_FL_WAIT_NOREPROC) || (mode == LCK_NL)) {
192 /* This loop determines where this processes locks start
193 * in the resource lr_granted list. */
194 list_for_each(tmp, &res->lr_granted) {
195 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
196 if (ldlm_same_flock_owner(lock, req)) {
202 lockmode_verify(mode);
204 /* This loop determines if there are existing locks
205 * that conflict with the new lock request. */
206 list_for_each(tmp, &res->lr_granted) {
207 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
209 if (ldlm_same_flock_owner(lock, req)) {
215 /* locks are compatible, overlap doesn't matter */
216 if (lockmode_compat(lock->l_granted_mode, mode))
219 if (!ldlm_flocks_overlap(lock, req))
223 RETURN(LDLM_ITER_CONTINUE);
225 if (*flags & LDLM_FL_BLOCK_NOWAIT) {
226 ldlm_flock_destroy(req, mode, *flags);
228 RETURN(LDLM_ITER_STOP);
231 if (*flags & LDLM_FL_TEST_LOCK) {
232 ldlm_flock_destroy(req, mode, *flags);
233 req->l_req_mode = lock->l_granted_mode;
234 req->l_policy_data.l_flock.pid =
235 lock->l_policy_data.l_flock.pid;
236 req->l_policy_data.l_flock.start =
237 lock->l_policy_data.l_flock.start;
238 req->l_policy_data.l_flock.end =
239 lock->l_policy_data.l_flock.end;
240 *flags |= LDLM_FL_LOCK_CHANGED;
241 RETURN(LDLM_ITER_STOP);
244 if (ldlm_flock_deadlock(req, lock)) {
245 ldlm_flock_destroy(req, mode, *flags);
247 RETURN(LDLM_ITER_STOP);
250 req->l_policy_data.l_flock.blocking_pid =
251 lock->l_policy_data.l_flock.pid;
252 req->l_policy_data.l_flock.blocking_export =
253 (long)(void *)lock->l_export;
255 LASSERT(list_empty(&req->l_flock_waitq));
256 spin_lock(&ldlm_flock_waitq_lock);
257 list_add_tail(&req->l_flock_waitq, &ldlm_flock_waitq);
258 spin_unlock(&ldlm_flock_waitq_lock);
260 ldlm_resource_add_lock(res, &res->lr_waiting, req);
261 *flags |= LDLM_FL_BLOCK_GRANTED;
262 RETURN(LDLM_ITER_STOP);
266 if (*flags & LDLM_FL_TEST_LOCK) {
267 ldlm_flock_destroy(req, mode, *flags);
268 req->l_req_mode = LCK_NL;
269 *flags |= LDLM_FL_LOCK_CHANGED;
270 RETURN(LDLM_ITER_STOP);
273 /* In case we had slept on this lock request take it off of the
274 * deadlock detection waitq. */
275 spin_lock(&ldlm_flock_waitq_lock);
276 list_del_init(&req->l_flock_waitq);
277 spin_unlock(&ldlm_flock_waitq_lock);
279 /* Scan the locks owned by this process that overlap this request.
280 * We may have to merge or split existing locks. */
283 ownlocks = &res->lr_granted;
285 list_for_remaining_safe(ownlocks, tmp, &res->lr_granted) {
286 lock = list_entry(ownlocks, struct ldlm_lock, l_res_link);
288 if (!ldlm_same_flock_owner(lock, new))
291 if (lock->l_granted_mode == mode) {
292 /* If the modes are the same then we need to process
293 * locks that overlap OR adjoin the new lock. The extra
294 * logic condition is necessary to deal with arithmetic
295 * overflow and underflow. */
296 if ((new->l_policy_data.l_flock.start >
297 (lock->l_policy_data.l_flock.end + 1))
298 && (lock->l_policy_data.l_flock.end !=
302 if ((new->l_policy_data.l_flock.end <
303 (lock->l_policy_data.l_flock.start - 1))
304 && (lock->l_policy_data.l_flock.start != 0))
307 if (new->l_policy_data.l_flock.start <
308 lock->l_policy_data.l_flock.start) {
309 lock->l_policy_data.l_flock.start =
310 new->l_policy_data.l_flock.start;
312 new->l_policy_data.l_flock.start =
313 lock->l_policy_data.l_flock.start;
316 if (new->l_policy_data.l_flock.end >
317 lock->l_policy_data.l_flock.end) {
318 lock->l_policy_data.l_flock.end =
319 new->l_policy_data.l_flock.end;
321 new->l_policy_data.l_flock.end =
322 lock->l_policy_data.l_flock.end;
326 ldlm_flock_destroy(lock, mode, *flags);
334 if (new->l_policy_data.l_flock.start >
335 lock->l_policy_data.l_flock.end)
338 if (new->l_policy_data.l_flock.end <
339 lock->l_policy_data.l_flock.start)
344 if (new->l_policy_data.l_flock.start <=
345 lock->l_policy_data.l_flock.start) {
346 if (new->l_policy_data.l_flock.end <
347 lock->l_policy_data.l_flock.end) {
348 lock->l_policy_data.l_flock.start =
349 new->l_policy_data.l_flock.end + 1;
352 ldlm_flock_destroy(lock, lock->l_req_mode, *flags);
355 if (new->l_policy_data.l_flock.end >=
356 lock->l_policy_data.l_flock.end) {
357 lock->l_policy_data.l_flock.end =
358 new->l_policy_data.l_flock.start - 1;
362 /* split the existing lock into two locks */
364 /* if this is an F_UNLCK operation then we could avoid
365 * allocating a new lock and use the req lock passed in
366 * with the request but this would complicate the reply
367 * processing since updates to req get reflected in the
368 * reply. The client side replays the lock request so
369 * it must see the original lock data in the reply. */
371 /* XXX - if ldlm_lock_new() can sleep we should
372 * release the ns_lock, allocate the new lock,
373 * and restart processing this lock. */
375 unlock_res_and_lock(req);
376 new2 = ldlm_lock_create(ns, &res->lr_name, LDLM_FLOCK,
377 lock->l_granted_mode, &null_cbs,
379 lock_res_and_lock(req);
381 ldlm_flock_destroy(req, lock->l_granted_mode, *flags);
383 RETURN(LDLM_ITER_STOP);
390 new2->l_granted_mode = lock->l_granted_mode;
391 new2->l_policy_data.l_flock.pid =
392 new->l_policy_data.l_flock.pid;
393 new2->l_policy_data.l_flock.start =
394 lock->l_policy_data.l_flock.start;
395 new2->l_policy_data.l_flock.end =
396 new->l_policy_data.l_flock.start - 1;
397 lock->l_policy_data.l_flock.start =
398 new->l_policy_data.l_flock.end + 1;
399 new2->l_conn_export = lock->l_conn_export;
400 if (lock->l_export != NULL) {
401 new2->l_export = class_export_get(lock->l_export);
402 if (new2->l_export->exp_lock_hash &&
403 hlist_unhashed(&new2->l_exp_hash))
404 lustre_hash_add(new2->l_export->exp_lock_hash,
405 &new2->l_remote_handle,
408 if (*flags == LDLM_FL_WAIT_NOREPROC) {
409 ldlm_lock_addref_internal_nolock(new2, lock->l_granted_mode);
412 /* insert new2 at lock */
413 ldlm_resource_add_lock(res, ownlocks, new2);
414 LDLM_LOCK_RELEASE(new2);
418 /* if new2 is created but never used, destroy it*/
419 if (splitted == 0 && new2 != NULL)
420 ldlm_lock_destroy_nolock(new2);
422 /* At this point we're granting the lock request. */
423 req->l_granted_mode = req->l_req_mode;
425 /* Add req to the granted queue before calling ldlm_reprocess_all(). */
427 list_del_init(&req->l_res_link);
428 /* insert new lock before ownlocks in list. */
429 ldlm_resource_add_lock(res, ownlocks, req);
432 if (*flags != LDLM_FL_WAIT_NOREPROC) {
434 /* If this is an unlock, reprocess the waitq and
435 * send completions ASTs for locks that can now be
436 * granted. The only problem with doing this
437 * reprocessing here is that the completion ASTs for
438 * newly granted locks will be sent before the unlock
439 * completion is sent. It shouldn't be an issue. Also
440 * note that ldlm_process_flock_lock() will recurse,
441 * but only once because first_enq will be false from
442 * ldlm_reprocess_queue. */
443 if ((mode == LCK_NL) && overlaps) {
444 CFS_LIST_HEAD(rpc_list);
447 ldlm_reprocess_queue(res, &res->lr_waiting,
450 unlock_res_and_lock(req);
451 rc = ldlm_run_ast_work(&rpc_list,
453 lock_res_and_lock(req);
455 GOTO(restart, -ERESTART);
458 LASSERT(req->l_completion_ast);
459 ldlm_add_ast_work_item(req, NULL, work_list);
463 /* In case we're reprocessing the requested lock we can't destroy
464 * it until after calling ldlm_ast_work_item() above so that lawi()
465 * can bump the reference count on req. Otherwise req could be freed
466 * before the completion AST can be sent. */
468 ldlm_flock_destroy(req, mode, *flags);
470 ldlm_resource_dump(D_INFO, res);
471 RETURN(LDLM_ITER_CONTINUE);
474 struct ldlm_flock_wait_data {
475 struct ldlm_lock *fwd_lock;
480 ldlm_flock_interrupted_wait(void *data)
482 struct ldlm_lock *lock;
483 struct lustre_handle lockh;
487 lock = ((struct ldlm_flock_wait_data *)data)->fwd_lock;
489 /* take lock off the deadlock detection waitq. */
490 spin_lock(&ldlm_flock_waitq_lock);
491 list_del_init(&lock->l_flock_waitq);
492 spin_unlock(&ldlm_flock_waitq_lock);
494 /* client side - set flag to prevent lock from being put on lru list */
495 lock->l_flags |= LDLM_FL_CBPENDING;
497 ldlm_lock_decref_internal(lock, lock->l_req_mode);
498 ldlm_lock2handle(lock, &lockh);
499 rc = ldlm_cli_cancel(&lockh);
501 CERROR("ldlm_cli_cancel: %d\n", rc);
507 ldlm_flock_completion_ast(struct ldlm_lock *lock, int flags, void *data)
509 struct ldlm_namespace *ns;
510 cfs_flock_t *getlk = lock->l_ast_data;
511 struct ldlm_flock_wait_data fwd;
512 struct obd_device *obd;
513 struct obd_import *imp = NULL;
516 struct l_wait_info lwi;
519 CDEBUG(D_DLMTRACE, "flags: 0x%x data: %p getlk: %p\n",
522 /* Import invalidation. We need to actually release the lock
523 * references being held, so that it can go away. No point in
524 * holding the lock even if app still believes it has it, since
525 * server already dropped it anyway. Only for granted locks too. */
526 lock_res_and_lock(lock);
527 if ((lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_LOCAL_ONLY)) ==
528 (LDLM_FL_FAILED|LDLM_FL_LOCAL_ONLY)) {
529 unlock_res_and_lock(lock);
530 if (lock->l_req_mode == lock->l_granted_mode &&
531 lock->l_granted_mode != LCK_NL)
532 ldlm_lock_decref_internal(lock, lock->l_req_mode);
535 unlock_res_and_lock(lock);
537 LASSERT(flags != LDLM_FL_WAIT_NOREPROC);
539 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
540 LDLM_FL_BLOCK_CONV)))
543 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
547 obd = class_exp2obd(lock->l_conn_export);
549 /* if this is a local lock, then there is no import */
551 imp = obd->u.cli.cl_import;
554 spin_lock(&imp->imp_lock);
555 fwd.fwd_generation = imp->imp_generation;
556 spin_unlock(&imp->imp_lock);
559 lwi = LWI_TIMEOUT_INTR(0, NULL, ldlm_flock_interrupted_wait, &fwd);
561 /* Go to sleep until the lock is granted. */
562 rc = l_wait_event(lock->l_waitq,
563 ((lock->l_req_mode == lock->l_granted_mode) ||
564 lock->l_destroyed), &lwi);
566 LDLM_DEBUG(lock, "client-side enqueue waking up: rc = %d", rc);
570 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT, 10);
571 LDLM_DEBUG(lock, "client-side enqueue granted");
572 ns = lock->l_resource->lr_namespace;
573 lock_res_and_lock(lock);
575 /* before flock's complete ast gets here, the flock
576 * can possibly be freed by another thread
578 if (lock->l_destroyed) {
579 LDLM_DEBUG(lock, "already destroyed by another thread");
580 unlock_res(lock->l_resource);
584 /* take lock off the deadlock detection waitq. */
585 spin_lock(&ldlm_flock_waitq_lock);
586 list_del_init(&lock->l_flock_waitq);
587 spin_unlock(&ldlm_flock_waitq_lock);
589 /* ldlm_lock_enqueue() has already placed lock on the granted list. */
590 list_del_init(&lock->l_res_link);
592 if (flags & LDLM_FL_TEST_LOCK) {
593 /* fcntl(F_GETLK) request */
594 /* The old mode was saved in getlk->fl_type so that if the mode
595 * in the lock changes we can decref the approprate refcount. */
596 ldlm_flock_destroy(lock, cfs_flock_type(getlk), LDLM_FL_WAIT_NOREPROC);
597 switch (lock->l_granted_mode) {
599 cfs_flock_set_type(getlk, F_RDLCK);
602 cfs_flock_set_type(getlk, F_WRLCK);
605 cfs_flock_set_type(getlk, F_UNLCK);
607 cfs_flock_set_pid(getlk, (pid_t)lock->l_policy_data.l_flock.pid);
608 cfs_flock_set_start(getlk, (loff_t)lock->l_policy_data.l_flock.start);
609 cfs_flock_set_end(getlk, (loff_t)lock->l_policy_data.l_flock.end);
611 int noreproc = LDLM_FL_WAIT_NOREPROC;
613 /* We need to reprocess the lock to do merges or splits
614 * with existing locks owned by this process. */
615 ldlm_process_flock_lock(lock, &noreproc, 1, &err, NULL);
617 cfs_waitq_signal(&lock->l_waitq);
619 unlock_res_and_lock(lock);
622 EXPORT_SYMBOL(ldlm_flock_completion_ast);
624 int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
625 void *data, int flag)
627 struct ldlm_namespace *ns;
631 LASSERT(flag == LDLM_CB_CANCELING);
633 ns = lock->l_resource->lr_namespace;
635 /* take lock off the deadlock detection waitq. */
636 spin_lock(&ldlm_flock_waitq_lock);
637 list_del_init(&lock->l_flock_waitq);
638 spin_unlock(&ldlm_flock_waitq_lock);