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>
32 extern kmem_cache_t *ldlm_resource_slab;
33 extern kmem_cache_t *ldlm_lock_slab;
34 extern struct list_head ldlm_namespace_list;
35 extern int (*mds_reint_p)(int offset, struct ptlrpc_request *req);
36 extern int (*mds_getattr_name_p)(int offset, struct ptlrpc_request *req);
38 inline unsigned long round_timeout(unsigned long timeout)
40 return ((timeout / HZ) + 1) * HZ;
43 static void waiting_locks_callback(unsigned long unused)
45 CERROR("lock(s) expired! need to start recovery!\n");
48 static struct list_head waiting_locks_list;
49 static spinlock_t waiting_locks_spinlock;
50 static struct timer_list waiting_locks_timer;
52 * Indicate that we're waiting for a client to call us back cancelling a given
53 * lock. We add it to the pending-callback chain, and schedule the lock-timeout
54 * timer to fire appropriately. (We round up to the next second, to avoid floods
55 * of timer firings during periods of high lock contention and traffic.
57 int ldlm_add_waiting_lock(struct ldlm_lock *lock)
59 unsigned long timeout_rounded;
62 LASSERT(list_empty(&lock->l_pending_chain));
64 CERROR("waiting for lock %p\n", lock);
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 CERROR("adjusting timer! (%0lx/%0lx < %0lx)\n",
73 lock->l_callback_timeout, timeout_rounded,
74 waiting_locks_timer.expires);
75 mod_timer(&waiting_locks_timer, timeout_rounded);
77 list_add(&lock->l_pending_chain, waiting_locks_list.prev); /* FIFO */
78 spin_unlock_bh(&waiting_locks_spinlock);
83 * Remove a lock from the pending list, likely because it had its cancellation
84 * callback arrive without incident. This adjusts the lock-timeout timer if
85 * needed. Returns 0 if the lock wasn't pending after all, 1 if it was.
87 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
89 struct list_head *list_next;
93 spin_lock_bh(&waiting_locks_spinlock);
95 if (list_empty(&lock->l_pending_chain)) {
96 spin_unlock_bh(&waiting_locks_spinlock);
100 CERROR("no longer waiting for lock %p\n", lock);
102 list_next = lock->l_pending_chain.next;
103 if (lock->l_pending_chain.prev == &waiting_locks_list) {
104 /* Removing the head of the list, adjust timer. */
105 if (list_next == &waiting_locks_list) {
106 /* No more, just cancel. */
107 CERROR("no more locks waiting, stopping timer\n");
108 del_timer(&waiting_locks_timer);
110 struct ldlm_lock *next;
111 next = list_entry(list_next, struct ldlm_lock,
113 CERROR("adjusting timer (%0lx %0lx)\n",
114 waiting_locks_timer.expires,
115 round_timeout(next->l_callback_timeout));
116 mod_timer(&waiting_locks_timer,
117 round_timeout(next->l_callback_timeout));
120 list_del_init(&lock->l_pending_chain);
121 spin_unlock_bh(&waiting_locks_spinlock);
125 static int ldlm_server_blocking_ast(struct ldlm_lock *lock,
126 struct ldlm_lock_desc *desc,
127 void *data, __u32 data_len)
129 struct ldlm_request *body;
130 struct ptlrpc_request *req;
131 struct ptlrpc_client *cl;
132 int rc = 0, size = sizeof(*body);
135 cl = &lock->l_resource->lr_namespace->ns_rpc_client;
136 req = ptlrpc_prep_req(cl, lock->l_export->exp_connection,
137 LDLM_BL_CALLBACK, 1, &size, NULL);
141 body = lustre_msg_buf(req->rq_reqmsg, 0);
142 memcpy(&body->lock_handle1, &lock->l_remote_handle,
143 sizeof(body->lock_handle1));
144 memcpy(&body->lock_desc, desc, sizeof(*desc));
146 LDLM_DEBUG(lock, "server preparing blocking AST");
147 req->rq_replen = lustre_msg_size(0, NULL);
149 ldlm_add_waiting_lock(lock);
150 rc = ptlrpc_queue_wait(req);
151 rc = ptlrpc_check_status(req, rc);
152 ptlrpc_free_req(req);
157 static int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags)
159 struct ldlm_request *body;
160 struct ptlrpc_request *req;
161 struct ptlrpc_client *cl;
162 int rc = 0, size = sizeof(*body);
170 cl = &lock->l_resource->lr_namespace->ns_rpc_client;
171 req = ptlrpc_prep_req(cl, lock->l_export->exp_connection,
172 LDLM_CP_CALLBACK, 1, &size, NULL);
176 body = lustre_msg_buf(req->rq_reqmsg, 0);
177 memcpy(&body->lock_handle1, &lock->l_remote_handle,
178 sizeof(body->lock_handle1));
179 body->lock_flags = flags;
180 ldlm_lock2desc(lock, &body->lock_desc);
182 LDLM_DEBUG(lock, "server preparing completion AST");
183 req->rq_replen = lustre_msg_size(0, NULL);
185 rc = ptlrpc_queue_wait(req);
186 rc = ptlrpc_check_status(req, rc);
187 ptlrpc_free_req(req);
191 int ldlm_handle_enqueue(struct ptlrpc_request *req)
193 struct obd_device *obddev = req->rq_export->exp_obd;
194 struct ldlm_reply *dlm_rep;
195 struct ldlm_request *dlm_req;
196 int rc, size = sizeof(*dlm_rep), cookielen = 0;
199 struct ldlm_lock *lock = NULL;
203 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
205 dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
206 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_MDSINTENT) {
207 /* In this case, the reply buffer is allocated deep in
208 * local_lock_enqueue by the policy function. */
210 cookielen = sizeof(*req);
212 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
215 CERROR("out of memory\n");
218 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT) {
219 cookie = &dlm_req->lock_desc.l_extent;
220 cookielen = sizeof(struct ldlm_extent);
224 lock = ldlm_lock_create(obddev->obd_namespace,
225 &dlm_req->lock_handle2,
226 dlm_req->lock_desc.l_resource.lr_name,
227 dlm_req->lock_desc.l_resource.lr_type,
228 dlm_req->lock_desc.l_req_mode, NULL, 0);
230 GOTO(out, err = -ENOMEM);
232 memcpy(&lock->l_remote_handle, &dlm_req->lock_handle1,
233 sizeof(lock->l_remote_handle));
234 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
236 flags = dlm_req->lock_flags;
237 err = ldlm_lock_enqueue(lock, cookie, cookielen, &flags,
238 ldlm_server_completion_ast,
239 ldlm_server_blocking_ast);
243 dlm_rep = lustre_msg_buf(req->rq_repmsg, 0);
244 dlm_rep->lock_flags = flags;
246 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
247 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
248 memcpy(&dlm_rep->lock_extent, &lock->l_extent,
249 sizeof(lock->l_extent));
250 if (dlm_rep->lock_flags & LDLM_FL_LOCK_CHANGED) {
251 memcpy(dlm_rep->lock_resource_name, lock->l_resource->lr_name,
252 sizeof(dlm_rep->lock_resource_name));
253 dlm_rep->lock_mode = lock->l_req_mode;
256 lock->l_export = req->rq_export;
257 ptlrpc_connection_addref(req->rq_connection);
261 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
263 req->rq_status = err;
265 if (ptlrpc_reply(req->rq_svc, req))
270 ldlm_reprocess_all(lock->l_resource);
273 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p)", lock);
278 int ldlm_handle_convert(struct ptlrpc_request *req)
280 struct ldlm_request *dlm_req;
281 struct ldlm_reply *dlm_rep;
282 struct ldlm_lock *lock;
283 int rc, size = sizeof(*dlm_rep);
286 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
288 CERROR("out of memory\n");
291 dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
292 dlm_rep = lustre_msg_buf(req->rq_repmsg, 0);
293 dlm_rep->lock_flags = dlm_req->lock_flags;
295 lock = ldlm_handle2lock(&dlm_req->lock_handle1);
297 req->rq_status = EINVAL;
299 LDLM_DEBUG(lock, "server-side convert handler START");
300 ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
301 &dlm_rep->lock_flags);
302 if (ldlm_del_waiting_lock(lock))
303 CDEBUG(D_DLMTRACE, "converted waiting lock %p\n", lock);
306 if (ptlrpc_reply(req->rq_svc, req) != 0)
310 ldlm_reprocess_all(lock->l_resource);
311 LDLM_DEBUG(lock, "server-side convert handler END");
314 LDLM_DEBUG_NOLOCK("server-side convert handler END");
319 int ldlm_handle_cancel(struct ptlrpc_request *req)
321 struct ldlm_request *dlm_req;
322 struct ldlm_lock *lock;
326 rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
328 CERROR("out of memory\n");
331 dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
333 lock = ldlm_handle2lock(&dlm_req->lock_handle1);
335 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock (lock "
336 "%p)", (void *)(unsigned long)
337 dlm_req->lock_handle1.addr);
338 req->rq_status = ESTALE;
340 LDLM_DEBUG(lock, "server-side cancel handler START");
341 ldlm_lock_cancel(lock);
342 if (ldlm_del_waiting_lock(lock))
343 CDEBUG(D_DLMTRACE, "cancelled waiting lock %p\n", lock);
347 if (ptlrpc_reply(req->rq_svc, req) != 0)
351 ldlm_reprocess_all(lock->l_resource);
352 LDLM_DEBUG(lock, "server-side cancel handler END");
359 static int ldlm_handle_bl_callback(struct ptlrpc_request *req)
361 struct ldlm_request *dlm_req;
362 struct ldlm_lock *lock;
366 OBD_FAIL_RETURN(OBD_FAIL_OSC_LOCK_BL_REPLY, 0);
368 rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
371 rc = ptlrpc_reply(req->rq_svc, req);
375 OBD_FAIL_RETURN(OBD_FAIL_OSC_LOCK_BL_AST, 0);
377 dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
379 lock = ldlm_handle2lock(&dlm_req->lock_handle1);
381 CERROR("blocking callback on lock %Lx - lock disappeared\n",
382 dlm_req->lock_handle1.addr);
386 LDLM_DEBUG(lock, "client blocking AST callback handler START");
388 l_lock(&lock->l_resource->lr_namespace->ns_lock);
389 lock->l_flags |= LDLM_FL_CBPENDING;
390 do_ast = (!lock->l_readers && !lock->l_writers);
391 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
394 LDLM_DEBUG(lock, "already unused, calling "
395 "callback (%p)", lock->l_blocking_ast);
396 if (lock->l_blocking_ast != NULL) {
397 lock->l_blocking_ast(lock, &dlm_req->lock_desc,
398 lock->l_data, lock->l_data_len);
401 LDLM_DEBUG(lock, "Lock still has references, will be"
404 LDLM_DEBUG(lock, "client blocking callback handler END");
409 static int ldlm_handle_cp_callback(struct ptlrpc_request *req)
411 struct list_head ast_list = LIST_HEAD_INIT(ast_list);
412 struct ldlm_request *dlm_req;
413 struct ldlm_lock *lock;
417 rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
420 rc = ptlrpc_reply(req->rq_svc, req);
424 dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
426 lock = ldlm_handle2lock(&dlm_req->lock_handle1);
428 CERROR("completion callback on lock %Lx - lock disappeared\n",
429 dlm_req->lock_handle1.addr);
433 LDLM_DEBUG(lock, "client completion callback handler START");
435 l_lock(&lock->l_resource->lr_namespace->ns_lock);
437 /* If we receive the completion AST before the actual enqueue returned,
438 * then we might need to switch resources or lock modes. */
439 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
440 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
441 LDLM_DEBUG(lock, "completion AST, new lock mode");
443 ldlm_resource_unlink_lock(lock);
444 if (memcmp(dlm_req->lock_desc.l_resource.lr_name,
445 lock->l_resource->lr_name,
446 sizeof(__u64) * RES_NAME_SIZE) != 0) {
447 ldlm_lock_change_resource(lock, dlm_req->lock_desc.l_resource.lr_name);
448 LDLM_DEBUG(lock, "completion AST, new resource");
450 lock->l_resource->lr_tmp = &ast_list;
451 ldlm_grant_lock(lock);
452 lock->l_resource->lr_tmp = NULL;
453 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
454 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
457 ldlm_run_ast_work(&ast_list);
459 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
464 static int ldlm_callback_handler(struct ptlrpc_request *req)
469 rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
471 CERROR("lustre_ldlm: Invalid request\n");
475 if (req->rq_reqmsg->type != PTL_RPC_MSG_REQUEST) {
476 CERROR("lustre_ldlm: wrong packet type sent %d\n",
477 req->rq_reqmsg->type);
478 GOTO(out, rc = -EINVAL);
480 switch (req->rq_reqmsg->opc) {
481 case LDLM_BL_CALLBACK:
482 CDEBUG(D_INODE, "blocking ast\n");
483 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
484 rc = ldlm_handle_bl_callback(req);
486 case LDLM_CP_CALLBACK:
487 CDEBUG(D_INODE, "completion ast\n");
488 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CP_CALLBACK, 0);
489 rc = ldlm_handle_cp_callback(req);
493 rc = ptlrpc_error(req->rq_svc, req);
500 RETURN(ptlrpc_error(req->rq_svc, req));
505 static int ldlm_iocontrol(long cmd, struct lustre_handle *conn, int len,
506 void *karg, void *uarg)
508 struct obd_device *obddev = class_conn2obd(conn);
509 struct ptlrpc_connection *connection;
513 if (_IOC_TYPE(cmd) != IOC_LDLM_TYPE || _IOC_NR(cmd) < IOC_LDLM_MIN_NR ||
514 _IOC_NR(cmd) > IOC_LDLM_MAX_NR) {
515 CDEBUG(D_IOCTL, "invalid ioctl (type %ld, nr %ld, size %ld)\n",
516 _IOC_TYPE(cmd), _IOC_NR(cmd), _IOC_SIZE(cmd));
520 OBD_ALLOC(obddev->u.ldlm.ldlm_client,
521 sizeof(*obddev->u.ldlm.ldlm_client));
522 ptlrpc_init_client(NULL, NULL,
523 LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
524 obddev->u.ldlm.ldlm_client);
525 connection = ptlrpc_uuid_to_connection("ldlm");
527 CERROR("No LDLM UUID found: assuming ldlm is local.\n");
531 err = ldlm_test(obddev, conn);
532 CERROR("-- done err %d\n", err);
535 ldlm_dump_all_namespaces();
538 GOTO(out, err = -EINVAL);
543 ptlrpc_put_connection(connection);
544 OBD_FREE(obddev->u.ldlm.ldlm_client,
545 sizeof(*obddev->u.ldlm.ldlm_client));
549 #define LDLM_NUM_THREADS 8
551 static int ldlm_setup(struct obd_device *obddev, obd_count len, void *buf)
553 struct ldlm_obd *ldlm = &obddev->u.ldlm;
558 rc = ldlm_proc_setup(obddev);
562 ldlm->ldlm_service = ptlrpc_init_svc(64 * 1024, LDLM_REQUEST_PORTAL,
563 LDLM_REPLY_PORTAL, "self",
564 ldlm_callback_handler, "ldlm");
565 if (!ldlm->ldlm_service)
566 GOTO(out_proc, rc = -ENOMEM);
568 for (i = 0; i < LDLM_NUM_THREADS; i++) {
570 sprintf(name, "lustre_dlm_%02d", i);
571 rc = ptlrpc_start_thread(obddev, ldlm->ldlm_service, name);
573 CERROR("cannot start LDLM thread #%d: rc %d\n", i, rc);
575 GOTO(out_thread, rc);
579 INIT_LIST_HEAD(&waiting_locks_list);
580 spin_lock_init(&waiting_locks_spinlock);
581 waiting_locks_timer.function = waiting_locks_callback;
582 waiting_locks_timer.data = 0;
583 init_timer(&waiting_locks_timer);
588 ptlrpc_stop_all_threads(ldlm->ldlm_service);
589 ptlrpc_unregister_service(ldlm->ldlm_service);
591 ldlm_proc_cleanup(obddev);
597 static int ldlm_cleanup(struct obd_device *obddev)
599 struct ldlm_obd *ldlm = &obddev->u.ldlm;
602 if (!list_empty(&ldlm_namespace_list)) {
603 CERROR("ldlm still has namespaces; clean these up first.\n");
607 ptlrpc_stop_all_threads(ldlm->ldlm_service);
608 ptlrpc_unregister_service(ldlm->ldlm_service);
609 ldlm_proc_cleanup(obddev);
615 struct obd_ops ldlm_obd_ops = {
616 o_iocontrol: ldlm_iocontrol,
618 o_cleanup: ldlm_cleanup,
619 o_connect: class_connect,
620 o_disconnect: class_disconnect
623 static int __init ldlm_init(void)
625 int rc = class_register_type(&ldlm_obd_ops, OBD_LDLM_DEVICENAME);
629 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
630 sizeof(struct ldlm_resource), 0,
631 SLAB_HWCACHE_ALIGN, NULL, NULL);
632 if (ldlm_resource_slab == NULL)
635 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
636 sizeof(struct ldlm_lock), 0,
637 SLAB_HWCACHE_ALIGN, NULL, NULL);
638 if (ldlm_lock_slab == NULL) {
639 kmem_cache_destroy(ldlm_resource_slab);
646 static void __exit ldlm_exit(void)
648 class_unregister_type(OBD_LDLM_DEVICENAME);
649 if (kmem_cache_destroy(ldlm_resource_slab) != 0)
650 CERROR("couldn't free ldlm resource slab\n");
651 if (kmem_cache_destroy(ldlm_lock_slab) != 0)
652 CERROR("couldn't free ldlm lock slab\n");
655 EXPORT_SYMBOL(ldlm_completion_ast);
656 EXPORT_SYMBOL(ldlm_handle_enqueue);
657 EXPORT_SYMBOL(ldlm_handle_cancel);
658 EXPORT_SYMBOL(ldlm_handle_convert);
659 EXPORT_SYMBOL(ldlm_register_intent);
660 EXPORT_SYMBOL(ldlm_unregister_intent);
661 EXPORT_SYMBOL(ldlm_lockname);
662 EXPORT_SYMBOL(ldlm_typename);
663 EXPORT_SYMBOL(ldlm_handle2lock);
664 EXPORT_SYMBOL(ldlm_lock2handle);
665 EXPORT_SYMBOL(ldlm_lock_put);
666 EXPORT_SYMBOL(ldlm_lock_match);
667 EXPORT_SYMBOL(ldlm_lock_addref);
668 EXPORT_SYMBOL(ldlm_lock_decref);
669 EXPORT_SYMBOL(ldlm_lock_change_resource);
670 EXPORT_SYMBOL(ldlm_cli_convert);
671 EXPORT_SYMBOL(ldlm_cli_enqueue);
672 EXPORT_SYMBOL(ldlm_cli_cancel);
673 EXPORT_SYMBOL(ldlm_match_or_enqueue);
674 EXPORT_SYMBOL(ldlm_it2str);
675 EXPORT_SYMBOL(ldlm_test);
676 EXPORT_SYMBOL(ldlm_regression_start);
677 EXPORT_SYMBOL(ldlm_regression_stop);
678 EXPORT_SYMBOL(ldlm_lock_dump);
679 EXPORT_SYMBOL(ldlm_namespace_new);
680 EXPORT_SYMBOL(ldlm_namespace_free);
681 EXPORT_SYMBOL(l_lock);
682 EXPORT_SYMBOL(l_unlock);
684 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
685 MODULE_DESCRIPTION("Lustre Lock Management Module v0.1");
686 MODULE_LICENSE("GPL");
688 module_init(ldlm_init);
689 module_exit(ldlm_exit);