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 struct list_head waiting_locks_list;
45 static spinlock_t waiting_locks_spinlock;
46 static struct timer_list waiting_locks_timer;
47 static int ldlm_already_setup = 0;
49 static void waiting_locks_callback(unsigned long unused)
51 struct list_head *liter, *n;
53 spin_lock_bh(&waiting_locks_spinlock);
54 list_for_each_safe(liter, n, &waiting_locks_list) {
55 struct ldlm_lock *l = list_entry(liter, struct ldlm_lock,
57 if (l->l_callback_timeout > jiffies)
59 LDLM_DEBUG(l, "timer expired, recovering exp %p on conn %p",
60 l->l_export, l->l_export->exp_connection);
61 recovd_conn_fail(l->l_export->exp_connection);
63 spin_unlock_bh(&waiting_locks_spinlock);
67 * Indicate that we're waiting for a client to call us back cancelling a given
68 * lock. We add it to the pending-callback chain, and schedule the lock-timeout
69 * timer to fire appropriately. (We round up to the next second, to avoid
70 * floods of timer firings during periods of high lock contention and traffic).
72 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
74 unsigned long timeout_rounded;
77 LASSERT(list_empty(&lock->l_pending_chain));
79 spin_lock_bh(&waiting_locks_spinlock);
80 lock->l_callback_timeout = jiffies + (obd_timeout * HZ);
82 timeout_rounded = round_timeout(lock->l_callback_timeout);
84 if (timeout_rounded < waiting_locks_timer.expires ||
85 !timer_pending(&waiting_locks_timer)) {
86 mod_timer(&waiting_locks_timer, timeout_rounded);
88 list_add_tail(&lock->l_pending_chain, &waiting_locks_list); /* FIFO */
89 spin_unlock_bh(&waiting_locks_spinlock);
94 * Remove a lock from the pending list, likely because it had its cancellation
95 * callback arrive without incident. This adjusts the lock-timeout timer if
96 * needed. Returns 0 if the lock wasn't pending after all, 1 if it was.
98 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
100 struct list_head *list_next;
104 spin_lock_bh(&waiting_locks_spinlock);
106 if (list_empty(&lock->l_pending_chain)) {
107 spin_unlock_bh(&waiting_locks_spinlock);
111 list_next = lock->l_pending_chain.next;
112 if (lock->l_pending_chain.prev == &waiting_locks_list) {
113 /* Removing the head of the list, adjust timer. */
114 if (list_next == &waiting_locks_list) {
115 /* No more, just cancel. */
116 del_timer(&waiting_locks_timer);
118 struct ldlm_lock *next;
119 next = list_entry(list_next, struct ldlm_lock,
121 mod_timer(&waiting_locks_timer,
122 round_timeout(next->l_callback_timeout));
125 list_del_init(&lock->l_pending_chain);
126 spin_unlock_bh(&waiting_locks_spinlock);
130 static int ldlm_server_blocking_ast(struct ldlm_lock *lock,
131 struct ldlm_lock_desc *desc,
132 void *data, __u32 data_len, int flag)
134 struct ldlm_request *body;
135 struct ptlrpc_request *req;
136 int rc = 0, size = sizeof(*body);
139 if (flag == LDLM_CB_CANCELING) {
140 /* Don't need to do anything here. */
144 req = ptlrpc_prep_req(&lock->l_export->exp_ldlm_data.led_import,
145 LDLM_BL_CALLBACK, 1, &size, NULL);
149 body = lustre_msg_buf(req->rq_reqmsg, 0);
150 memcpy(&body->lock_handle1, &lock->l_remote_handle,
151 sizeof(body->lock_handle1));
152 memcpy(&body->lock_desc, desc, sizeof(*desc));
154 LDLM_DEBUG(lock, "server preparing blocking AST");
155 req->rq_replen = 0; /* no reply needed */
157 ldlm_add_waiting_lock(lock);
158 (void)ptl_send_rpc(req);
160 /* no commit, and no waiting for reply, so 2x decref now */
161 ptlrpc_req_finished(req);
162 ptlrpc_req_finished(req);
167 static int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags)
169 struct ldlm_request *body;
170 struct ptlrpc_request *req;
171 int rc = 0, size = sizeof(*body);
179 req = ptlrpc_prep_req(&lock->l_export->exp_ldlm_data.led_import,
180 LDLM_CP_CALLBACK, 1, &size, NULL);
184 body = lustre_msg_buf(req->rq_reqmsg, 0);
185 memcpy(&body->lock_handle1, &lock->l_remote_handle,
186 sizeof(body->lock_handle1));
187 body->lock_flags = flags;
188 ldlm_lock2desc(lock, &body->lock_desc);
190 LDLM_DEBUG(lock, "server preparing completion AST");
191 req->rq_replen = 0; /* no reply needed */
193 (void)ptl_send_rpc(req);
194 /* no commit, and no waiting for reply, so 2x decref now */
195 ptlrpc_req_finished(req);
196 ptlrpc_req_finished(req);
201 int ldlm_handle_enqueue(struct ptlrpc_request *req)
203 struct obd_device *obddev = req->rq_export->exp_obd;
204 struct ldlm_reply *dlm_rep;
205 struct ldlm_request *dlm_req;
206 int rc, size = sizeof(*dlm_rep), cookielen = 0;
209 struct ldlm_lock *lock = NULL;
213 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
215 dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
216 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_MDSINTENT) {
217 /* In this case, the reply buffer is allocated deep in
218 * local_lock_enqueue by the policy function. */
220 cookielen = sizeof(*req);
222 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
225 CERROR("out of memory\n");
228 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT) {
229 cookie = &dlm_req->lock_desc.l_extent;
230 cookielen = sizeof(struct ldlm_extent);
234 /* XXX notice that this lock has no callback data: of course the
235 export would be exactly what we may want to use here... */
236 lock = ldlm_lock_create(obddev->obd_namespace,
237 &dlm_req->lock_handle2,
238 dlm_req->lock_desc.l_resource.lr_name,
239 dlm_req->lock_desc.l_resource.lr_type,
240 dlm_req->lock_desc.l_req_mode, NULL, 0);
242 GOTO(out, err = -ENOMEM);
244 memcpy(&lock->l_remote_handle, &dlm_req->lock_handle1,
245 sizeof(lock->l_remote_handle));
246 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
248 flags = dlm_req->lock_flags;
249 err = ldlm_lock_enqueue(lock, cookie, cookielen, &flags,
250 ldlm_server_completion_ast,
251 ldlm_server_blocking_ast);
255 dlm_rep = lustre_msg_buf(req->rq_repmsg, 0);
256 dlm_rep->lock_flags = flags;
258 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
259 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
260 memcpy(&dlm_rep->lock_extent, &lock->l_extent,
261 sizeof(lock->l_extent));
262 if (dlm_rep->lock_flags & LDLM_FL_LOCK_CHANGED) {
263 memcpy(dlm_rep->lock_resource_name, lock->l_resource->lr_name,
264 sizeof(dlm_rep->lock_resource_name));
265 dlm_rep->lock_mode = lock->l_req_mode;
268 lock->l_export = req->rq_export;
269 if (lock->l_export) {
270 l_lock(&lock->l_resource->lr_namespace->ns_lock);
271 list_add(&lock->l_export_chain,
272 &lock->l_export->exp_ldlm_data.led_held_locks);
273 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
279 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
281 req->rq_status = err;
283 if (ptlrpc_reply(req->rq_svc, req))
288 ldlm_reprocess_all(lock->l_resource);
291 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p)", lock);
296 int ldlm_handle_convert(struct ptlrpc_request *req)
298 struct ldlm_request *dlm_req;
299 struct ldlm_reply *dlm_rep;
300 struct ldlm_lock *lock;
301 int rc, size = sizeof(*dlm_rep);
304 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
306 CERROR("out of memory\n");
309 dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
310 dlm_rep = lustre_msg_buf(req->rq_repmsg, 0);
311 dlm_rep->lock_flags = dlm_req->lock_flags;
313 lock = ldlm_handle2lock(&dlm_req->lock_handle1);
315 req->rq_status = EINVAL;
317 LDLM_DEBUG(lock, "server-side convert handler START");
318 ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
319 &dlm_rep->lock_flags);
320 if (ldlm_del_waiting_lock(lock))
321 CDEBUG(D_DLMTRACE, "converted waiting lock %p\n", lock);
324 if (ptlrpc_reply(req->rq_svc, req) != 0)
328 ldlm_reprocess_all(lock->l_resource);
329 LDLM_DEBUG(lock, "server-side convert handler END");
332 LDLM_DEBUG_NOLOCK("server-side convert handler END");
337 int ldlm_handle_cancel(struct ptlrpc_request *req)
339 struct ldlm_request *dlm_req;
340 struct ldlm_lock *lock;
344 rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
346 CERROR("out of memory\n");
349 dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
351 CERROR("bad request buffer for cancel\n");
355 lock = ldlm_handle2lock(&dlm_req->lock_handle1);
357 CERROR("bad lock handle\n");
358 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock (lock "
359 "%p)", (void *)(unsigned long)
360 dlm_req->lock_handle1.addr);
361 req->rq_status = ESTALE;
363 LDLM_DEBUG(lock, "server-side cancel handler START");
364 ldlm_lock_cancel(lock);
365 if (ldlm_del_waiting_lock(lock))
366 CDEBUG(D_DLMTRACE, "cancelled waiting lock %p\n", lock);
370 if (ptlrpc_reply(req->rq_svc, req) != 0)
374 ldlm_reprocess_all(lock->l_resource);
375 LDLM_DEBUG(lock, "server-side cancel handler END");
382 static int ldlm_handle_bl_callback(struct ptlrpc_request *req)
384 struct ldlm_request *dlm_req;
385 struct ldlm_lock *lock;
389 OBD_FAIL_RETURN(OBD_FAIL_OSC_LOCK_BL_AST, 0);
391 dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
393 lock = ldlm_handle2lock(&dlm_req->lock_handle1);
395 CERROR("blocking callback on lock "LPX64" - lock disappeared\n",
396 dlm_req->lock_handle1.addr);
400 LDLM_DEBUG(lock, "client blocking AST callback handler START");
402 l_lock(&lock->l_resource->lr_namespace->ns_lock);
403 lock->l_flags |= LDLM_FL_CBPENDING;
404 do_ast = (!lock->l_readers && !lock->l_writers);
405 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
408 LDLM_DEBUG(lock, "already unused, calling "
409 "callback (%p)", lock->l_blocking_ast);
410 if (lock->l_blocking_ast != NULL) {
411 lock->l_blocking_ast(lock, &dlm_req->lock_desc,
412 lock->l_data, lock->l_data_len,
416 LDLM_DEBUG(lock, "Lock still has references, will be"
419 LDLM_DEBUG(lock, "client blocking callback handler END");
424 static int ldlm_handle_cp_callback(struct ptlrpc_request *req)
426 struct list_head ast_list = LIST_HEAD_INIT(ast_list);
427 struct ldlm_request *dlm_req;
428 struct ldlm_lock *lock;
431 OBD_FAIL_RETURN(OBD_FAIL_OSC_LOCK_CP_AST, 0);
433 dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
435 lock = ldlm_handle2lock(&dlm_req->lock_handle1);
437 CERROR("completion callback on lock "LPX64" - lock "
438 "disappeared\n", dlm_req->lock_handle1.addr);
442 LDLM_DEBUG(lock, "client completion callback handler START");
444 l_lock(&lock->l_resource->lr_namespace->ns_lock);
446 /* If we receive the completion AST before the actual enqueue returned,
447 * then we might need to switch lock modes, resources, or extents. */
448 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
449 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
450 LDLM_DEBUG(lock, "completion AST, new lock mode");
452 if (lock->l_resource->lr_type == LDLM_EXTENT)
453 memcpy(&lock->l_extent, &dlm_req->lock_desc.l_extent,
454 sizeof(lock->l_extent));
455 ldlm_resource_unlink_lock(lock);
456 if (memcmp(dlm_req->lock_desc.l_resource.lr_name,
457 lock->l_resource->lr_name,
458 sizeof(__u64) * RES_NAME_SIZE) != 0) {
459 ldlm_lock_change_resource(lock, dlm_req->lock_desc.l_resource.lr_name);
460 LDLM_DEBUG(lock, "completion AST, new resource");
462 lock->l_resource->lr_tmp = &ast_list;
463 ldlm_grant_lock(lock);
464 lock->l_resource->lr_tmp = NULL;
465 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
466 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
469 ldlm_run_ast_work(&ast_list);
471 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
476 static int ldlm_callback_handler(struct ptlrpc_request *req)
481 rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
483 CERROR("lustre_ldlm: Invalid request: %d\n", rc);
487 if (req->rq_export == NULL) {
488 CERROR("lustre_dlm: operation %d with missing/invalid export\n",
489 req->rq_reqmsg->opc);
493 switch (req->rq_reqmsg->opc) {
494 case LDLM_BL_CALLBACK:
495 CDEBUG(D_INODE, "blocking ast\n");
496 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
497 rc = ldlm_handle_bl_callback(req);
499 case LDLM_CP_CALLBACK:
500 CDEBUG(D_INODE, "completion ast\n");
501 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CP_CALLBACK, 0);
502 rc = ldlm_handle_cp_callback(req);
505 CERROR("invalid opcode %d\n", req->rq_reqmsg->opc);
513 static int ldlm_iocontrol(long cmd, struct lustre_handle *conn, int len,
514 void *karg, void *uarg)
516 struct obd_device *obddev = class_conn2obd(conn);
517 struct ptlrpc_connection *connection;
521 if (_IOC_TYPE(cmd) != IOC_LDLM_TYPE || _IOC_NR(cmd) < IOC_LDLM_MIN_NR ||
522 _IOC_NR(cmd) > IOC_LDLM_MAX_NR) {
523 CDEBUG(D_IOCTL, "invalid ioctl (type %ld, nr %ld, size %ld)\n",
524 _IOC_TYPE(cmd), _IOC_NR(cmd), _IOC_SIZE(cmd));
528 OBD_ALLOC(obddev->u.ldlm.ldlm_client,
529 sizeof(*obddev->u.ldlm.ldlm_client));
530 connection = ptlrpc_uuid_to_connection("ldlm");
532 CERROR("No LDLM UUID found: assuming ldlm is local.\n");
535 ptlrpc_init_client(LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
536 obddev->u.ldlm.ldlm_client, connection);
541 err = ldlm_test(obddev, conn);
542 CERROR("-- done err %d\n", err);
545 ldlm_dump_all_namespaces();
548 GOTO(out, err = -EINVAL);
553 ptlrpc_put_connection(connection);
554 OBD_FREE(obddev->u.ldlm.ldlm_client,
555 sizeof(*obddev->u.ldlm.ldlm_client));
559 #define LDLM_NUM_THREADS 8
561 static int ldlm_setup(struct obd_device *obddev, obd_count len, void *buf)
563 struct ldlm_obd *ldlm = &obddev->u.ldlm;
567 if (ldlm_already_setup)
572 rc = ldlm_proc_setup(obddev);
577 ptlrpc_init_svc(LDLM_NEVENTS, LDLM_NBUFS, LDLM_BUFSIZE,
578 LDLM_MAXREQSIZE, LDLM_REQUEST_PORTAL,
579 LDLM_REPLY_PORTAL, "self",
580 ldlm_callback_handler, "ldlm");
581 if (!ldlm->ldlm_service)
582 GOTO(out_dec, rc = -ENOMEM);
584 for (i = 0; i < LDLM_NUM_THREADS; i++) {
586 sprintf(name, "lustre_dlm_%02d", i);
587 rc = ptlrpc_start_thread(obddev, ldlm->ldlm_service, name);
589 CERROR("cannot start LDLM thread #%d: rc %d\n", i, rc);
591 GOTO(out_thread, rc);
595 INIT_LIST_HEAD(&waiting_locks_list);
596 spin_lock_init(&waiting_locks_spinlock);
597 waiting_locks_timer.function = waiting_locks_callback;
598 waiting_locks_timer.data = 0;
599 init_timer(&waiting_locks_timer);
601 ldlm_already_setup = 1;
606 ptlrpc_stop_all_threads(ldlm->ldlm_service);
607 ptlrpc_unregister_service(ldlm->ldlm_service);
611 ldlm_proc_cleanup(obddev);
618 static int ldlm_cleanup(struct obd_device *obddev)
620 struct ldlm_obd *ldlm = &obddev->u.ldlm;
623 if (!list_empty(&ldlm_namespace_list)) {
624 CERROR("ldlm still has namespaces; clean these up first.\n");
628 ptlrpc_stop_all_threads(ldlm->ldlm_service);
629 ptlrpc_unregister_service(ldlm->ldlm_service);
630 /* ldlm_proc_cleanup(obddev); */
632 ldlm_already_setup = 0;
637 static int ldlm_connect(struct lustre_handle *conn, struct obd_device *src,
638 obd_uuid_t cluuid, struct recovd_obd *recovd,
639 ptlrpc_recovery_cb_t recover)
641 return class_connect(conn, src, cluuid);
644 struct obd_ops ldlm_obd_ops = {
645 o_iocontrol: ldlm_iocontrol,
647 o_cleanup: ldlm_cleanup,
648 o_connect: ldlm_connect,
649 o_disconnect: class_disconnect
652 static int __init ldlm_init(void)
654 int rc = class_register_type(&ldlm_obd_ops, OBD_LDLM_DEVICENAME);
658 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
659 sizeof(struct ldlm_resource), 0,
660 SLAB_HWCACHE_ALIGN, NULL, NULL);
661 if (ldlm_resource_slab == NULL)
664 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
665 sizeof(struct ldlm_lock), 0,
666 SLAB_HWCACHE_ALIGN, NULL, NULL);
667 if (ldlm_lock_slab == NULL) {
668 kmem_cache_destroy(ldlm_resource_slab);
675 static void __exit ldlm_exit(void)
677 class_unregister_type(OBD_LDLM_DEVICENAME);
678 if (kmem_cache_destroy(ldlm_resource_slab) != 0)
679 CERROR("couldn't free ldlm resource slab\n");
680 if (kmem_cache_destroy(ldlm_lock_slab) != 0)
681 CERROR("couldn't free ldlm lock slab\n");
684 EXPORT_SYMBOL(ldlm_completion_ast);
685 EXPORT_SYMBOL(ldlm_handle_enqueue);
686 EXPORT_SYMBOL(ldlm_handle_cancel);
687 EXPORT_SYMBOL(ldlm_handle_convert);
688 EXPORT_SYMBOL(ldlm_register_intent);
689 EXPORT_SYMBOL(ldlm_unregister_intent);
690 EXPORT_SYMBOL(ldlm_lockname);
691 EXPORT_SYMBOL(ldlm_typename);
692 EXPORT_SYMBOL(ldlm_handle2lock);
693 EXPORT_SYMBOL(ldlm_lock2handle);
694 EXPORT_SYMBOL(ldlm_lock_put);
695 EXPORT_SYMBOL(ldlm_lock_match);
696 EXPORT_SYMBOL(ldlm_lock_addref);
697 EXPORT_SYMBOL(ldlm_lock_decref);
698 EXPORT_SYMBOL(ldlm_lock_change_resource);
699 EXPORT_SYMBOL(ldlm_cli_convert);
700 EXPORT_SYMBOL(ldlm_cli_enqueue);
701 EXPORT_SYMBOL(ldlm_cli_cancel);
702 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
703 EXPORT_SYMBOL(ldlm_match_or_enqueue);
704 EXPORT_SYMBOL(ldlm_it2str);
705 EXPORT_SYMBOL(ldlm_test);
706 EXPORT_SYMBOL(ldlm_regression_start);
707 EXPORT_SYMBOL(ldlm_regression_stop);
708 EXPORT_SYMBOL(ldlm_lock_dump);
709 EXPORT_SYMBOL(ldlm_namespace_new);
710 EXPORT_SYMBOL(ldlm_namespace_cleanup);
711 EXPORT_SYMBOL(ldlm_namespace_free);
712 EXPORT_SYMBOL(ldlm_namespace_dump);
713 EXPORT_SYMBOL(ldlm_cancel_locks_for_export);
714 EXPORT_SYMBOL(l_lock);
715 EXPORT_SYMBOL(l_unlock);
717 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
718 MODULE_DESCRIPTION("Lustre Lock Management Module v0.1");
719 MODULE_LICENSE("GPL");
721 module_init(ldlm_init);
722 module_exit(ldlm_exit);