1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (C) 2002 Cluster File Systems, Inc.
5 * Author: Peter Braam <braam@clusterfs.com>
6 * Author: Phil Schwan <phil@clusterfs.com>
8 * This file is part of Lustre, http://www.lustre.org.
10 * Lustre is free software; you can redistribute it and/or
11 * modify it under the terms of version 2 of the GNU General Public
12 * License as published by the Free Software Foundation.
14 * Lustre is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 * GNU General Public License for more details.
19 * You should have received a copy of the GNU General Public License
20 * along with Lustre; if not, write to the Free Software
21 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
25 #define DEBUG_SUBSYSTEM S_LDLM
27 #include <linux/module.h>
28 #include <linux/slab.h>
29 #include <linux/lustre_dlm.h>
30 #include <linux/init.h>
31 #include <linux/obd_class.h>
33 extern kmem_cache_t *ldlm_resource_slab;
34 extern kmem_cache_t *ldlm_lock_slab;
35 extern struct list_head ldlm_namespace_list;
36 extern int (*mds_reint_p)(int offset, struct ptlrpc_request *req);
37 extern int (*mds_getattr_name_p)(int offset, struct ptlrpc_request *req);
39 inline unsigned long round_timeout(unsigned long timeout)
41 return ((timeout / HZ) + 1) * HZ;
44 static void waiting_locks_callback(unsigned long unused)
46 CERROR("lock(s) expired! need to start recovery!\n");
49 static struct list_head waiting_locks_list;
50 static spinlock_t waiting_locks_spinlock;
51 static struct timer_list waiting_locks_timer;
53 * Indicate that we're waiting for a client to call us back cancelling a given
54 * lock. We add it to the pending-callback chain, and schedule the lock-timeout
55 * timer to fire appropriately. (We round up to the next second, to avoid
56 * floods of timer firings during periods of high lock contention and traffic.
58 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
60 unsigned long timeout_rounded;
63 LASSERT(list_empty(&lock->l_pending_chain));
65 spin_lock_bh(&waiting_locks_spinlock);
66 lock->l_callback_timeout = jiffies + (obd_timeout * HZ);
68 timeout_rounded = round_timeout(lock->l_callback_timeout);
70 if (timeout_rounded < waiting_locks_timer.expires ||
71 !timer_pending(&waiting_locks_timer)) {
72 mod_timer(&waiting_locks_timer, timeout_rounded);
74 list_add_tail(&lock->l_pending_chain, &waiting_locks_list); /* FIFO */
75 spin_unlock_bh(&waiting_locks_spinlock);
80 * Remove a lock from the pending list, likely because it had its cancellation
81 * callback arrive without incident. This adjusts the lock-timeout timer if
82 * needed. Returns 0 if the lock wasn't pending after all, 1 if it was.
84 static int ldlm_del_waiting_lock(struct ldlm_lock *lock)
86 struct list_head *list_next;
90 spin_lock_bh(&waiting_locks_spinlock);
92 if (list_empty(&lock->l_pending_chain)) {
93 spin_unlock_bh(&waiting_locks_spinlock);
97 list_next = lock->l_pending_chain.next;
98 if (lock->l_pending_chain.prev == &waiting_locks_list) {
99 /* Removing the head of the list, adjust timer. */
100 if (list_next == &waiting_locks_list) {
101 /* No more, just cancel. */
102 del_timer(&waiting_locks_timer);
104 struct ldlm_lock *next;
105 next = list_entry(list_next, struct ldlm_lock,
107 mod_timer(&waiting_locks_timer,
108 round_timeout(next->l_callback_timeout));
111 list_del_init(&lock->l_pending_chain);
112 spin_unlock_bh(&waiting_locks_spinlock);
116 static int ldlm_server_blocking_ast(struct ldlm_lock *lock,
117 struct ldlm_lock_desc *desc,
118 void *data, __u32 data_len)
120 struct ldlm_request *body;
121 struct ptlrpc_request *req;
122 int rc = 0, size = sizeof(*body);
125 req = ptlrpc_prep_req(&lock->l_export->exp_ldlm_data.led_import,
126 LDLM_BL_CALLBACK, 1, &size, NULL);
130 body = lustre_msg_buf(req->rq_reqmsg, 0);
131 memcpy(&body->lock_handle1, &lock->l_remote_handle,
132 sizeof(body->lock_handle1));
133 memcpy(&body->lock_desc, desc, sizeof(*desc));
135 LDLM_DEBUG(lock, "server preparing blocking AST");
136 req->rq_replen = lustre_msg_size(0, NULL);
138 ldlm_add_waiting_lock(lock);
139 rc = ptlrpc_queue_wait(req);
140 rc = ptlrpc_check_status(req, rc);
141 ptlrpc_free_req(req);
146 static int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags)
148 struct ldlm_request *body;
149 struct ptlrpc_request *req;
150 int rc = 0, size = sizeof(*body);
158 req = ptlrpc_prep_req(&lock->l_export->exp_ldlm_data.led_import,
159 LDLM_CP_CALLBACK, 1, &size, NULL);
163 body = lustre_msg_buf(req->rq_reqmsg, 0);
164 memcpy(&body->lock_handle1, &lock->l_remote_handle,
165 sizeof(body->lock_handle1));
166 body->lock_flags = flags;
167 ldlm_lock2desc(lock, &body->lock_desc);
169 LDLM_DEBUG(lock, "server preparing completion AST");
170 req->rq_replen = lustre_msg_size(0, NULL);
172 rc = ptlrpc_queue_wait(req);
173 rc = ptlrpc_check_status(req, rc);
174 ptlrpc_free_req(req);
178 int ldlm_handle_enqueue(struct ptlrpc_request *req)
180 struct obd_device *obddev = req->rq_export->exp_obd;
181 struct ldlm_reply *dlm_rep;
182 struct ldlm_request *dlm_req;
183 int rc, size = sizeof(*dlm_rep), cookielen = 0;
186 struct ldlm_lock *lock = NULL;
190 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
192 dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
193 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_MDSINTENT) {
194 /* In this case, the reply buffer is allocated deep in
195 * local_lock_enqueue by the policy function. */
197 cookielen = sizeof(*req);
199 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
202 CERROR("out of memory\n");
205 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT) {
206 cookie = &dlm_req->lock_desc.l_extent;
207 cookielen = sizeof(struct ldlm_extent);
211 /* XXX notice that this lock has no callback data: of course the
212 export would be exactly what we may want to use here... */
213 lock = ldlm_lock_create(obddev->obd_namespace,
214 &dlm_req->lock_handle2,
215 dlm_req->lock_desc.l_resource.lr_name,
216 dlm_req->lock_desc.l_resource.lr_type,
217 dlm_req->lock_desc.l_req_mode, NULL, 0);
219 GOTO(out, err = -ENOMEM);
221 memcpy(&lock->l_remote_handle, &dlm_req->lock_handle1,
222 sizeof(lock->l_remote_handle));
223 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
225 flags = dlm_req->lock_flags;
226 err = ldlm_lock_enqueue(lock, cookie, cookielen, &flags,
227 ldlm_server_completion_ast,
228 ldlm_server_blocking_ast);
232 dlm_rep = lustre_msg_buf(req->rq_repmsg, 0);
233 dlm_rep->lock_flags = flags;
235 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
236 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
237 memcpy(&dlm_rep->lock_extent, &lock->l_extent,
238 sizeof(lock->l_extent));
239 if (dlm_rep->lock_flags & LDLM_FL_LOCK_CHANGED) {
240 memcpy(dlm_rep->lock_resource_name, lock->l_resource->lr_name,
241 sizeof(dlm_rep->lock_resource_name));
242 dlm_rep->lock_mode = lock->l_req_mode;
245 lock->l_export = req->rq_export;
246 if (lock->l_export) {
247 l_lock(&lock->l_resource->lr_namespace->ns_lock);
248 list_add(&lock->l_export_chain,
249 &lock->l_export->exp_ldlm_data.led_held_locks);
250 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
253 ptlrpc_connection_addref(req->rq_connection);
257 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
259 req->rq_status = err;
261 if (ptlrpc_reply(req->rq_svc, req))
266 ldlm_reprocess_all(lock->l_resource);
269 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p)", lock);
274 int ldlm_handle_convert(struct ptlrpc_request *req)
276 struct ldlm_request *dlm_req;
277 struct ldlm_reply *dlm_rep;
278 struct ldlm_lock *lock;
279 int rc, size = sizeof(*dlm_rep);
282 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
284 CERROR("out of memory\n");
287 dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
288 dlm_rep = lustre_msg_buf(req->rq_repmsg, 0);
289 dlm_rep->lock_flags = dlm_req->lock_flags;
291 lock = ldlm_handle2lock(&dlm_req->lock_handle1);
293 req->rq_status = EINVAL;
295 LDLM_DEBUG(lock, "server-side convert handler START");
296 ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
297 &dlm_rep->lock_flags);
298 if (ldlm_del_waiting_lock(lock))
299 CDEBUG(D_DLMTRACE, "converted waiting lock %p\n", lock);
302 if (ptlrpc_reply(req->rq_svc, req) != 0)
306 ldlm_reprocess_all(lock->l_resource);
307 LDLM_DEBUG(lock, "server-side convert handler END");
310 LDLM_DEBUG_NOLOCK("server-side convert handler END");
315 int ldlm_handle_cancel(struct ptlrpc_request *req)
317 struct ldlm_request *dlm_req;
318 struct ldlm_lock *lock;
322 rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
324 CERROR("out of memory\n");
327 dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
329 lock = ldlm_handle2lock(&dlm_req->lock_handle1);
331 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock (lock "
332 "%p)", (void *)(unsigned long)
333 dlm_req->lock_handle1.addr);
334 req->rq_status = ESTALE;
336 LDLM_DEBUG(lock, "server-side cancel handler START");
337 ldlm_lock_cancel(lock);
338 if (ldlm_del_waiting_lock(lock))
339 CDEBUG(D_DLMTRACE, "cancelled waiting lock %p\n", lock);
343 if (ptlrpc_reply(req->rq_svc, req) != 0)
347 ldlm_reprocess_all(lock->l_resource);
348 LDLM_DEBUG(lock, "server-side cancel handler END");
355 static int ldlm_handle_bl_callback(struct ptlrpc_request *req)
357 struct ldlm_request *dlm_req;
358 struct ldlm_lock *lock;
362 OBD_FAIL_RETURN(OBD_FAIL_OSC_LOCK_BL_REPLY, 0);
364 rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
367 rc = ptlrpc_reply(req->rq_svc, req);
371 OBD_FAIL_RETURN(OBD_FAIL_OSC_LOCK_BL_AST, 0);
373 dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
375 lock = ldlm_handle2lock(&dlm_req->lock_handle1);
377 CERROR("blocking callback on lock "LPX64" - lock disappeared\n",
378 dlm_req->lock_handle1.addr);
382 LDLM_DEBUG(lock, "client blocking AST callback handler START");
384 l_lock(&lock->l_resource->lr_namespace->ns_lock);
385 lock->l_flags |= LDLM_FL_CBPENDING;
386 do_ast = (!lock->l_readers && !lock->l_writers);
387 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
390 LDLM_DEBUG(lock, "already unused, calling "
391 "callback (%p)", lock->l_blocking_ast);
392 if (lock->l_blocking_ast != NULL) {
393 lock->l_blocking_ast(lock, &dlm_req->lock_desc,
394 lock->l_data, lock->l_data_len);
397 LDLM_DEBUG(lock, "Lock still has references, will be"
400 LDLM_DEBUG(lock, "client blocking callback handler END");
405 static int ldlm_handle_cp_callback(struct ptlrpc_request *req)
407 struct list_head ast_list = LIST_HEAD_INIT(ast_list);
408 struct ldlm_request *dlm_req;
409 struct ldlm_lock *lock;
413 rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
416 rc = ptlrpc_reply(req->rq_svc, req);
420 dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
422 lock = ldlm_handle2lock(&dlm_req->lock_handle1);
424 CERROR("completion callback on lock "LPX64" - lock "
425 "disappeared\n", dlm_req->lock_handle1.addr);
429 LDLM_DEBUG(lock, "client completion callback handler START");
431 l_lock(&lock->l_resource->lr_namespace->ns_lock);
433 /* If we receive the completion AST before the actual enqueue returned,
434 * then we might need to switch resources or lock modes. */
435 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
436 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
437 LDLM_DEBUG(lock, "completion AST, new lock mode");
439 ldlm_resource_unlink_lock(lock);
440 if (memcmp(dlm_req->lock_desc.l_resource.lr_name,
441 lock->l_resource->lr_name,
442 sizeof(__u64) * RES_NAME_SIZE) != 0) {
443 ldlm_lock_change_resource(lock, dlm_req->lock_desc.l_resource.lr_name);
444 LDLM_DEBUG(lock, "completion AST, new resource");
446 lock->l_resource->lr_tmp = &ast_list;
447 ldlm_grant_lock(lock);
448 lock->l_resource->lr_tmp = NULL;
449 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
450 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
453 ldlm_run_ast_work(&ast_list);
455 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
460 static int ldlm_callback_handler(struct ptlrpc_request *req)
465 rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
467 CERROR("lustre_ldlm: Invalid request\n");
471 if (req->rq_export == NULL) {
472 CERROR("lustre_dlm: operation %d with missing/invalid export\n",
473 req->rq_reqmsg->opc);
474 GOTO(out, rc = -ENOTCONN);
477 switch (req->rq_reqmsg->opc) {
478 case LDLM_BL_CALLBACK:
479 CDEBUG(D_INODE, "blocking ast\n");
480 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
481 rc = ldlm_handle_bl_callback(req);
483 case LDLM_CP_CALLBACK:
484 CDEBUG(D_INODE, "completion ast\n");
485 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CP_CALLBACK, 0);
486 rc = ldlm_handle_cp_callback(req);
490 rc = ptlrpc_error(req->rq_svc, req);
497 RETURN(ptlrpc_error(req->rq_svc, req));
502 static int ldlm_iocontrol(long cmd, struct lustre_handle *conn, int len,
503 void *karg, void *uarg)
505 struct obd_device *obddev = class_conn2obd(conn);
506 struct ptlrpc_connection *connection;
510 if (_IOC_TYPE(cmd) != IOC_LDLM_TYPE || _IOC_NR(cmd) < IOC_LDLM_MIN_NR ||
511 _IOC_NR(cmd) > IOC_LDLM_MAX_NR) {
512 CDEBUG(D_IOCTL, "invalid ioctl (type %ld, nr %ld, size %ld)\n",
513 _IOC_TYPE(cmd), _IOC_NR(cmd), _IOC_SIZE(cmd));
517 OBD_ALLOC(obddev->u.ldlm.ldlm_client,
518 sizeof(*obddev->u.ldlm.ldlm_client));
519 connection = ptlrpc_uuid_to_connection("ldlm");
521 CERROR("No LDLM UUID found: assuming ldlm is local.\n");
524 ptlrpc_init_client(LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
525 obddev->u.ldlm.ldlm_client, connection);
530 err = ldlm_test(obddev, conn);
531 CERROR("-- done err %d\n", err);
534 ldlm_dump_all_namespaces();
537 GOTO(out, err = -EINVAL);
542 ptlrpc_put_connection(connection);
543 OBD_FREE(obddev->u.ldlm.ldlm_client,
544 sizeof(*obddev->u.ldlm.ldlm_client));
548 #define LDLM_NUM_THREADS 8
550 static int ldlm_setup(struct obd_device *obddev, obd_count len, void *buf)
552 struct ldlm_obd *ldlm = &obddev->u.ldlm;
558 rc = ldlm_proc_setup(obddev);
562 ldlm->ldlm_service = ptlrpc_init_svc(LDLM_NEVENTS, LDLM_NBUFS,
563 LDLM_BUFSIZE, LDLM_MAXREQSIZE,
564 LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
565 "self", ldlm_callback_handler, "ldlm");
566 if (!ldlm->ldlm_service)
567 GOTO(out_dec, rc = -ENOMEM);
569 for (i = 0; i < LDLM_NUM_THREADS; i++) {
571 sprintf(name, "lustre_dlm_%02d", i);
572 rc = ptlrpc_start_thread(obddev, ldlm->ldlm_service, name);
574 CERROR("cannot start LDLM thread #%d: rc %d\n", i, rc);
576 GOTO(out_thread, rc);
580 INIT_LIST_HEAD(&waiting_locks_list);
581 spin_lock_init(&waiting_locks_spinlock);
582 waiting_locks_timer.function = waiting_locks_callback;
583 waiting_locks_timer.data = 0;
584 init_timer(&waiting_locks_timer);
589 ptlrpc_stop_all_threads(ldlm->ldlm_service);
590 ptlrpc_unregister_service(ldlm->ldlm_service);
594 ldlm_proc_cleanup(obddev);
601 static int ldlm_cleanup(struct obd_device *obddev)
603 struct ldlm_obd *ldlm = &obddev->u.ldlm;
606 if (!list_empty(&ldlm_namespace_list)) {
607 CERROR("ldlm still has namespaces; clean these up first.\n");
611 ptlrpc_stop_all_threads(ldlm->ldlm_service);
612 ptlrpc_unregister_service(ldlm->ldlm_service);
613 /* ldlm_proc_cleanup(obddev); */
619 struct obd_ops ldlm_obd_ops = {
620 o_iocontrol: ldlm_iocontrol,
622 o_cleanup: ldlm_cleanup,
623 o_connect: class_connect,
624 o_disconnect: class_disconnect
627 static int __init ldlm_init(void)
629 int rc = class_register_type(&ldlm_obd_ops, OBD_LDLM_DEVICENAME);
633 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
634 sizeof(struct ldlm_resource), 0,
635 SLAB_HWCACHE_ALIGN, NULL, NULL);
636 if (ldlm_resource_slab == NULL)
639 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
640 sizeof(struct ldlm_lock), 0,
641 SLAB_HWCACHE_ALIGN, NULL, NULL);
642 if (ldlm_lock_slab == NULL) {
643 kmem_cache_destroy(ldlm_resource_slab);
650 static void __exit ldlm_exit(void)
652 class_unregister_type(OBD_LDLM_DEVICENAME);
653 if (kmem_cache_destroy(ldlm_resource_slab) != 0)
654 CERROR("couldn't free ldlm resource slab\n");
655 if (kmem_cache_destroy(ldlm_lock_slab) != 0)
656 CERROR("couldn't free ldlm lock slab\n");
659 EXPORT_SYMBOL(ldlm_completion_ast);
660 EXPORT_SYMBOL(ldlm_handle_enqueue);
661 EXPORT_SYMBOL(ldlm_handle_cancel);
662 EXPORT_SYMBOL(ldlm_handle_convert);
663 EXPORT_SYMBOL(ldlm_register_intent);
664 EXPORT_SYMBOL(ldlm_unregister_intent);
665 EXPORT_SYMBOL(ldlm_lockname);
666 EXPORT_SYMBOL(ldlm_typename);
667 EXPORT_SYMBOL(ldlm_handle2lock);
668 EXPORT_SYMBOL(ldlm_lock2handle);
669 EXPORT_SYMBOL(ldlm_lock_put);
670 EXPORT_SYMBOL(ldlm_lock_match);
671 EXPORT_SYMBOL(ldlm_lock_addref);
672 EXPORT_SYMBOL(ldlm_lock_decref);
673 EXPORT_SYMBOL(ldlm_lock_change_resource);
674 EXPORT_SYMBOL(ldlm_cli_convert);
675 EXPORT_SYMBOL(ldlm_cli_enqueue);
676 EXPORT_SYMBOL(ldlm_cli_cancel);
677 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
678 EXPORT_SYMBOL(ldlm_match_or_enqueue);
679 EXPORT_SYMBOL(ldlm_it2str);
680 EXPORT_SYMBOL(ldlm_test);
681 EXPORT_SYMBOL(ldlm_regression_start);
682 EXPORT_SYMBOL(ldlm_regression_stop);
683 EXPORT_SYMBOL(ldlm_lock_dump);
684 EXPORT_SYMBOL(ldlm_namespace_new);
685 EXPORT_SYMBOL(ldlm_namespace_free);
686 EXPORT_SYMBOL(l_lock);
687 EXPORT_SYMBOL(l_unlock);
689 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
690 MODULE_DESCRIPTION("Lustre Lock Management Module v0.1");
691 MODULE_LICENSE("GPL");
693 module_init(ldlm_init);
694 module_exit(ldlm_exit);