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);
173 CDEBUG(D_DLMTRACE, "flags %#x pid %u mode %u start "LPU64" end "LPU64
174 "\n", *flags, new->l_policy_data.l_flock.pid, mode,
175 req->l_policy_data.l_flock.start,
176 req->l_policy_data.l_flock.end);
181 /* No blocking ASTs are sent to the clients for
182 * Posix file & record locks */
183 req->l_blocking_ast = NULL;
185 /* Called on the server for lock cancels. */
186 req->l_blocking_ast = ldlm_flock_blocking_ast;
190 if ((*flags == LDLM_FL_WAIT_NOREPROC) || (mode == LCK_NL)) {
191 /* This loop determines where this processes locks start
192 * in the resource lr_granted list. */
193 list_for_each(tmp, &res->lr_granted) {
194 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
195 if (ldlm_same_flock_owner(lock, req)) {
201 lockmode_verify(mode);
203 /* This loop determines if there are existing locks
204 * that conflict with the new lock request. */
205 list_for_each(tmp, &res->lr_granted) {
206 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
208 if (ldlm_same_flock_owner(lock, req)) {
214 /* locks are compatible, overlap doesn't matter */
215 if (lockmode_compat(lock->l_granted_mode, mode))
218 if (!ldlm_flocks_overlap(lock, req))
222 RETURN(LDLM_ITER_CONTINUE);
224 if (*flags & LDLM_FL_BLOCK_NOWAIT) {
225 ldlm_flock_destroy(req, mode, *flags);
227 RETURN(LDLM_ITER_STOP);
230 if (*flags & LDLM_FL_TEST_LOCK) {
231 ldlm_flock_destroy(req, mode, *flags);
232 req->l_req_mode = lock->l_granted_mode;
233 req->l_policy_data.l_flock.pid =
234 lock->l_policy_data.l_flock.pid;
235 req->l_policy_data.l_flock.start =
236 lock->l_policy_data.l_flock.start;
237 req->l_policy_data.l_flock.end =
238 lock->l_policy_data.l_flock.end;
239 *flags |= LDLM_FL_LOCK_CHANGED;
240 RETURN(LDLM_ITER_STOP);
243 if (ldlm_flock_deadlock(req, lock)) {
244 ldlm_flock_destroy(req, mode, *flags);
246 RETURN(LDLM_ITER_STOP);
249 req->l_policy_data.l_flock.blocking_pid =
250 lock->l_policy_data.l_flock.pid;
251 req->l_policy_data.l_flock.blocking_export =
252 (long)(void *)lock->l_export;
254 LASSERT(list_empty(&req->l_flock_waitq));
255 spin_lock(&ldlm_flock_waitq_lock);
256 list_add_tail(&req->l_flock_waitq, &ldlm_flock_waitq);
257 spin_unlock(&ldlm_flock_waitq_lock);
259 ldlm_resource_add_lock(res, &res->lr_waiting, req);
260 *flags |= LDLM_FL_BLOCK_GRANTED;
261 RETURN(LDLM_ITER_STOP);
265 if (*flags & LDLM_FL_TEST_LOCK) {
266 ldlm_flock_destroy(req, mode, *flags);
267 req->l_req_mode = LCK_NL;
268 *flags |= LDLM_FL_LOCK_CHANGED;
269 RETURN(LDLM_ITER_STOP);
272 /* In case we had slept on this lock request take it off of the
273 * deadlock detection waitq. */
274 spin_lock(&ldlm_flock_waitq_lock);
275 list_del_init(&req->l_flock_waitq);
276 spin_unlock(&ldlm_flock_waitq_lock);
278 /* Scan the locks owned by this process that overlap this request.
279 * We may have to merge or split existing locks. */
282 ownlocks = &res->lr_granted;
284 list_for_remaining_safe(ownlocks, tmp, &res->lr_granted) {
285 lock = list_entry(ownlocks, struct ldlm_lock, l_res_link);
287 if (!ldlm_same_flock_owner(lock, new))
290 if (lock->l_granted_mode == mode) {
291 /* If the modes are the same then we need to process
292 * locks that overlap OR adjoin the new lock. The extra
293 * logic condition is necessary to deal with arithmetic
294 * overflow and underflow. */
295 if ((new->l_policy_data.l_flock.start >
296 (lock->l_policy_data.l_flock.end + 1))
297 && (lock->l_policy_data.l_flock.end !=
301 if ((new->l_policy_data.l_flock.end <
302 (lock->l_policy_data.l_flock.start - 1))
303 && (lock->l_policy_data.l_flock.start != 0))
306 if (new->l_policy_data.l_flock.start <
307 lock->l_policy_data.l_flock.start) {
308 lock->l_policy_data.l_flock.start =
309 new->l_policy_data.l_flock.start;
311 new->l_policy_data.l_flock.start =
312 lock->l_policy_data.l_flock.start;
315 if (new->l_policy_data.l_flock.end >
316 lock->l_policy_data.l_flock.end) {
317 lock->l_policy_data.l_flock.end =
318 new->l_policy_data.l_flock.end;
320 new->l_policy_data.l_flock.end =
321 lock->l_policy_data.l_flock.end;
325 ldlm_flock_destroy(lock, mode, *flags);
333 if (new->l_policy_data.l_flock.start >
334 lock->l_policy_data.l_flock.end)
337 if (new->l_policy_data.l_flock.end <
338 lock->l_policy_data.l_flock.start)
343 if (new->l_policy_data.l_flock.start <=
344 lock->l_policy_data.l_flock.start) {
345 if (new->l_policy_data.l_flock.end <
346 lock->l_policy_data.l_flock.end) {
347 lock->l_policy_data.l_flock.start =
348 new->l_policy_data.l_flock.end + 1;
351 ldlm_flock_destroy(lock, lock->l_req_mode, *flags);
354 if (new->l_policy_data.l_flock.end >=
355 lock->l_policy_data.l_flock.end) {
356 lock->l_policy_data.l_flock.end =
357 new->l_policy_data.l_flock.start - 1;
361 /* split the existing lock into two locks */
363 /* if this is an F_UNLCK operation then we could avoid
364 * allocating a new lock and use the req lock passed in
365 * with the request but this would complicate the reply
366 * processing since updates to req get reflected in the
367 * reply. The client side replays the lock request so
368 * it must see the original lock data in the reply. */
370 /* XXX - if ldlm_lock_new() can sleep we should
371 * release the ns_lock, allocate the new lock,
372 * and restart processing this lock. */
374 unlock_res_and_lock(req);
375 new2 = ldlm_lock_create(ns, &res->lr_name, LDLM_FLOCK,
376 lock->l_granted_mode, NULL, NULL, NULL,
378 lock_res_and_lock(req);
380 ldlm_flock_destroy(req, lock->l_granted_mode, *flags);
382 RETURN(LDLM_ITER_STOP);
389 new2->l_granted_mode = lock->l_granted_mode;
390 new2->l_policy_data.l_flock.pid =
391 new->l_policy_data.l_flock.pid;
392 new2->l_policy_data.l_flock.start =
393 lock->l_policy_data.l_flock.start;
394 new2->l_policy_data.l_flock.end =
395 new->l_policy_data.l_flock.start - 1;
396 lock->l_policy_data.l_flock.start =
397 new->l_policy_data.l_flock.end + 1;
398 new2->l_conn_export = lock->l_conn_export;
399 if (lock->l_export != NULL) {
400 new2->l_export = class_export_get(lock->l_export);
401 spin_lock(&new2->l_export->exp_ldlm_data.led_lock);
402 list_add(&new2->l_export_chain,
403 &new2->l_export->exp_ldlm_data.led_held_locks);
404 spin_unlock(&new2->l_export->exp_ldlm_data.led_lock);
406 if (*flags == LDLM_FL_WAIT_NOREPROC) {
407 ldlm_lock_addref_internal_nolock(new2, lock->l_granted_mode);
410 /* insert new2 at lock */
411 ldlm_resource_add_lock(res, ownlocks, new2);
416 /* if new2 is created but never used, destroy it*/
417 if (splitted == 0 && new2 != NULL)
418 ldlm_lock_destroy_nolock(new2);
420 /* At this point we're granting the lock request. */
421 req->l_granted_mode = req->l_req_mode;
423 /* Add req to the granted queue before calling ldlm_reprocess_all(). */
425 list_del_init(&req->l_res_link);
426 /* insert new lock before ownlocks in list. */
427 ldlm_resource_add_lock(res, ownlocks, req);
430 if (*flags != LDLM_FL_WAIT_NOREPROC) {
432 /* If this is an unlock, reprocess the waitq and
433 * send completions ASTs for locks that can now be
434 * granted. The only problem with doing this
435 * reprocessing here is that the completion ASTs for
436 * newly granted locks will be sent before the unlock
437 * completion is sent. It shouldn't be an issue. Also
438 * note that ldlm_process_flock_lock() will recurse,
439 * but only once because first_enq will be false from
440 * ldlm_reprocess_queue. */
441 if ((mode == LCK_NL) && overlaps) {
442 CFS_LIST_HEAD(rpc_list);
445 ldlm_reprocess_queue(res, &res->lr_waiting,
448 unlock_res_and_lock(req);
449 rc = ldlm_run_ast_work(&rpc_list,
451 lock_res_and_lock(req);
453 GOTO(restart, -ERESTART);
456 LASSERT(req->l_completion_ast);
457 ldlm_add_ast_work_item(req, NULL, work_list);
461 /* In case we're reprocessing the requested lock we can't destroy
462 * it until after calling ldlm_ast_work_item() above so that lawi()
463 * can bump the reference count on req. Otherwise req could be freed
464 * before the completion AST can be sent. */
466 ldlm_flock_destroy(req, mode, *flags);
468 ldlm_resource_dump(D_INFO, res);
469 RETURN(LDLM_ITER_CONTINUE);
472 struct ldlm_flock_wait_data {
473 struct ldlm_lock *fwd_lock;
478 ldlm_flock_interrupted_wait(void *data)
480 struct ldlm_lock *lock;
481 struct lustre_handle lockh;
485 lock = ((struct ldlm_flock_wait_data *)data)->fwd_lock;
487 /* take lock off the deadlock detection waitq. */
488 spin_lock(&ldlm_flock_waitq_lock);
489 list_del_init(&lock->l_flock_waitq);
490 spin_unlock(&ldlm_flock_waitq_lock);
492 /* client side - set flag to prevent lock from being put on lru list */
493 lock->l_flags |= LDLM_FL_CBPENDING;
495 ldlm_lock_decref_internal(lock, lock->l_req_mode);
496 ldlm_lock2handle(lock, &lockh);
497 rc = ldlm_cli_cancel(&lockh);
499 CERROR("ldlm_cli_cancel: %d\n", rc);
505 ldlm_flock_completion_ast(struct ldlm_lock *lock, int flags, void *data)
507 struct ldlm_namespace *ns;
508 cfs_flock_t *getlk = lock->l_ast_data;
509 struct ldlm_flock_wait_data fwd;
510 struct obd_device *obd;
511 struct obd_import *imp = NULL;
514 struct l_wait_info lwi;
517 CDEBUG(D_DLMTRACE, "flags: 0x%x data: %p getlk: %p\n",
520 /* Import invalidation. We need to actually release the lock
521 * references being held, so that it can go away. No point in
522 * holding the lock even if app still believes it has it, since
523 * server already dropped it anyway. Only for granted locks too. */
524 lock_res_and_lock(lock);
525 if ((lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_LOCAL_ONLY)) ==
526 (LDLM_FL_FAILED|LDLM_FL_LOCAL_ONLY)) {
527 unlock_res_and_lock(lock);
528 if (lock->l_req_mode == lock->l_granted_mode &&
529 lock->l_granted_mode != LCK_NL)
530 ldlm_lock_decref_internal(lock, lock->l_req_mode);
533 unlock_res_and_lock(lock);
535 LASSERT(flags != LDLM_FL_WAIT_NOREPROC);
537 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
538 LDLM_FL_BLOCK_CONV)))
541 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
545 obd = class_exp2obd(lock->l_conn_export);
547 /* if this is a local lock, then there is no import */
549 imp = obd->u.cli.cl_import;
552 spin_lock(&imp->imp_lock);
553 fwd.fwd_generation = imp->imp_generation;
554 spin_unlock(&imp->imp_lock);
557 lwi = LWI_TIMEOUT_INTR(0, NULL, ldlm_flock_interrupted_wait, &fwd);
559 /* Go to sleep until the lock is granted. */
560 rc = l_wait_event(lock->l_waitq,
561 ((lock->l_req_mode == lock->l_granted_mode) ||
562 lock->l_destroyed), &lwi);
564 LDLM_DEBUG(lock, "client-side enqueue waking up: rc = %d", rc);
568 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT, 10);
569 LDLM_DEBUG(lock, "client-side enqueue granted");
570 ns = lock->l_resource->lr_namespace;
571 lock_res_and_lock(lock);
573 /* before flock's complete ast gets here, the flock
574 * can possibly be freed by another thread
576 if (lock->l_destroyed) {
577 LDLM_DEBUG(lock, "already destroyed by another thread");
578 unlock_res(lock->l_resource);
582 /* take lock off the deadlock detection waitq. */
583 spin_lock(&ldlm_flock_waitq_lock);
584 list_del_init(&lock->l_flock_waitq);
585 spin_unlock(&ldlm_flock_waitq_lock);
587 /* ldlm_lock_enqueue() has already placed lock on the granted list. */
588 list_del_init(&lock->l_res_link);
590 if (flags & LDLM_FL_TEST_LOCK) {
591 /* fcntl(F_GETLK) request */
592 /* The old mode was saved in getlk->fl_type so that if the mode
593 * in the lock changes we can decref the approprate refcount. */
594 ldlm_flock_destroy(lock, cfs_flock_type(getlk), LDLM_FL_WAIT_NOREPROC);
595 switch (lock->l_granted_mode) {
597 cfs_flock_set_type(getlk, F_RDLCK);
600 cfs_flock_set_type(getlk, F_WRLCK);
603 cfs_flock_set_type(getlk, F_UNLCK);
605 cfs_flock_set_pid(getlk, (pid_t)lock->l_policy_data.l_flock.pid);
606 cfs_flock_set_start(getlk, (loff_t)lock->l_policy_data.l_flock.start);
607 cfs_flock_set_end(getlk, (loff_t)lock->l_policy_data.l_flock.end);
609 int noreproc = LDLM_FL_WAIT_NOREPROC;
611 /* We need to reprocess the lock to do merges or splits
612 * with existing locks owned by this process. */
613 ldlm_process_flock_lock(lock, &noreproc, 1, &err, NULL);
615 cfs_waitq_signal(&lock->l_waitq);
617 unlock_res_and_lock(lock);
620 EXPORT_SYMBOL(ldlm_flock_completion_ast);
622 int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
623 void *data, int flag)
625 struct ldlm_namespace *ns;
629 LASSERT(flag == LDLM_CB_CANCELING);
631 ns = lock->l_resource->lr_namespace;
633 /* take lock off the deadlock detection waitq. */
634 spin_lock(&ldlm_flock_waitq_lock);
635 list_del_init(&lock->l_flock_waitq);
636 spin_unlock(&ldlm_flock_waitq_lock);