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 floods
56 * of timer firings during periods of high lock contention and traffic.
58 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 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 struct ptlrpc_client *cl;
123 int rc = 0, size = sizeof(*body);
126 cl = &lock->l_resource->lr_namespace->ns_rpc_client;
127 req = ptlrpc_prep_req(cl, lock->l_export->exp_connection,
128 LDLM_BL_CALLBACK, 1, &size, NULL);
132 body = lustre_msg_buf(req->rq_reqmsg, 0);
133 memcpy(&body->lock_handle1, &lock->l_remote_handle,
134 sizeof(body->lock_handle1));
135 memcpy(&body->lock_desc, desc, sizeof(*desc));
137 LDLM_DEBUG(lock, "server preparing blocking AST");
138 req->rq_replen = lustre_msg_size(0, NULL);
140 ldlm_add_waiting_lock(lock);
141 rc = ptlrpc_queue_wait(req);
142 rc = ptlrpc_check_status(req, rc);
143 ptlrpc_free_req(req);
148 static int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags)
150 struct ldlm_request *body;
151 struct ptlrpc_request *req;
152 struct ptlrpc_client *cl;
153 int rc = 0, size = sizeof(*body);
161 cl = &lock->l_resource->lr_namespace->ns_rpc_client;
162 req = ptlrpc_prep_req(cl, lock->l_export->exp_connection,
163 LDLM_CP_CALLBACK, 1, &size, NULL);
167 body = lustre_msg_buf(req->rq_reqmsg, 0);
168 memcpy(&body->lock_handle1, &lock->l_remote_handle,
169 sizeof(body->lock_handle1));
170 body->lock_flags = flags;
171 ldlm_lock2desc(lock, &body->lock_desc);
173 LDLM_DEBUG(lock, "server preparing completion AST");
174 req->rq_replen = lustre_msg_size(0, NULL);
176 rc = ptlrpc_queue_wait(req);
177 rc = ptlrpc_check_status(req, rc);
178 ptlrpc_free_req(req);
182 int ldlm_handle_enqueue(struct ptlrpc_request *req)
184 struct obd_device *obddev = req->rq_export->exp_obd;
185 struct ldlm_reply *dlm_rep;
186 struct ldlm_request *dlm_req;
187 int rc, size = sizeof(*dlm_rep), cookielen = 0;
190 struct ldlm_lock *lock = NULL;
194 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
196 dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
197 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_MDSINTENT) {
198 /* In this case, the reply buffer is allocated deep in
199 * local_lock_enqueue by the policy function. */
201 cookielen = sizeof(*req);
203 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
206 CERROR("out of memory\n");
209 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT) {
210 cookie = &dlm_req->lock_desc.l_extent;
211 cookielen = sizeof(struct ldlm_extent);
215 lock = ldlm_lock_create(obddev->obd_namespace,
216 &dlm_req->lock_handle2,
217 dlm_req->lock_desc.l_resource.lr_name,
218 dlm_req->lock_desc.l_resource.lr_type,
219 dlm_req->lock_desc.l_req_mode, NULL, 0);
221 GOTO(out, err = -ENOMEM);
223 memcpy(&lock->l_remote_handle, &dlm_req->lock_handle1,
224 sizeof(lock->l_remote_handle));
225 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
227 flags = dlm_req->lock_flags;
228 err = ldlm_lock_enqueue(lock, cookie, cookielen, &flags,
229 ldlm_server_completion_ast,
230 ldlm_server_blocking_ast);
234 dlm_rep = lustre_msg_buf(req->rq_repmsg, 0);
235 dlm_rep->lock_flags = flags;
237 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
238 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
239 memcpy(&dlm_rep->lock_extent, &lock->l_extent,
240 sizeof(lock->l_extent));
241 if (dlm_rep->lock_flags & LDLM_FL_LOCK_CHANGED) {
242 memcpy(dlm_rep->lock_resource_name, lock->l_resource->lr_name,
243 sizeof(dlm_rep->lock_resource_name));
244 dlm_rep->lock_mode = lock->l_req_mode;
247 lock->l_export = req->rq_export;
248 if (lock->l_export) {
249 l_lock(&lock->l_resource->lr_namespace->ns_lock);
250 list_add(&lock->l_export_chain,
251 &lock->l_export->exp_ldlm_data.led_held_locks);
252 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
255 ptlrpc_connection_addref(req->rq_connection);
259 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
261 req->rq_status = err;
263 if (ptlrpc_reply(req->rq_svc, req))
268 ldlm_reprocess_all(lock->l_resource);
271 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p)", lock);
276 int ldlm_handle_convert(struct ptlrpc_request *req)
278 struct ldlm_request *dlm_req;
279 struct ldlm_reply *dlm_rep;
280 struct ldlm_lock *lock;
281 int rc, size = sizeof(*dlm_rep);
284 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
286 CERROR("out of memory\n");
289 dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
290 dlm_rep = lustre_msg_buf(req->rq_repmsg, 0);
291 dlm_rep->lock_flags = dlm_req->lock_flags;
293 lock = ldlm_handle2lock(&dlm_req->lock_handle1);
295 req->rq_status = EINVAL;
297 LDLM_DEBUG(lock, "server-side convert handler START");
298 ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
299 &dlm_rep->lock_flags);
300 if (ldlm_del_waiting_lock(lock))
301 CDEBUG(D_DLMTRACE, "converted waiting lock %p\n", lock);
304 if (ptlrpc_reply(req->rq_svc, req) != 0)
308 ldlm_reprocess_all(lock->l_resource);
309 LDLM_DEBUG(lock, "server-side convert handler END");
312 LDLM_DEBUG_NOLOCK("server-side convert handler END");
317 int ldlm_handle_cancel(struct ptlrpc_request *req)
319 struct ldlm_request *dlm_req;
320 struct ldlm_lock *lock;
324 rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
326 CERROR("out of memory\n");
329 dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
331 lock = ldlm_handle2lock(&dlm_req->lock_handle1);
333 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock (lock "
334 "%p)", (void *)(unsigned long)
335 dlm_req->lock_handle1.addr);
336 req->rq_status = ESTALE;
338 LDLM_DEBUG(lock, "server-side cancel handler START");
339 ldlm_lock_cancel(lock);
340 if (ldlm_del_waiting_lock(lock))
341 CDEBUG(D_DLMTRACE, "cancelled waiting lock %p\n", lock);
345 if (ptlrpc_reply(req->rq_svc, req) != 0)
349 ldlm_reprocess_all(lock->l_resource);
350 LDLM_DEBUG(lock, "server-side cancel handler END");
357 static int ldlm_handle_bl_callback(struct ptlrpc_request *req)
359 struct ldlm_request *dlm_req;
360 struct ldlm_lock *lock;
364 OBD_FAIL_RETURN(OBD_FAIL_OSC_LOCK_BL_REPLY, 0);
366 rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
369 rc = ptlrpc_reply(req->rq_svc, req);
373 OBD_FAIL_RETURN(OBD_FAIL_OSC_LOCK_BL_AST, 0);
375 dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
377 lock = ldlm_handle2lock(&dlm_req->lock_handle1);
379 CERROR("blocking callback on lock %Lx - lock disappeared\n",
380 dlm_req->lock_handle1.addr);
384 LDLM_DEBUG(lock, "client blocking AST callback handler START");
386 l_lock(&lock->l_resource->lr_namespace->ns_lock);
387 lock->l_flags |= LDLM_FL_CBPENDING;
388 do_ast = (!lock->l_readers && !lock->l_writers);
389 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
392 LDLM_DEBUG(lock, "already unused, calling "
393 "callback (%p)", lock->l_blocking_ast);
394 if (lock->l_blocking_ast != NULL) {
395 lock->l_blocking_ast(lock, &dlm_req->lock_desc,
396 lock->l_data, lock->l_data_len);
399 LDLM_DEBUG(lock, "Lock still has references, will be"
402 LDLM_DEBUG(lock, "client blocking callback handler END");
407 static int ldlm_handle_cp_callback(struct ptlrpc_request *req)
409 struct list_head ast_list = LIST_HEAD_INIT(ast_list);
410 struct ldlm_request *dlm_req;
411 struct ldlm_lock *lock;
415 rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
418 rc = ptlrpc_reply(req->rq_svc, req);
422 dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
424 lock = ldlm_handle2lock(&dlm_req->lock_handle1);
426 CERROR("completion callback on lock %Lx - lock disappeared\n",
427 dlm_req->lock_handle1.addr);
431 LDLM_DEBUG(lock, "client completion callback handler START");
433 l_lock(&lock->l_resource->lr_namespace->ns_lock);
435 /* If we receive the completion AST before the actual enqueue returned,
436 * then we might need to switch resources or lock modes. */
437 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
438 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
439 LDLM_DEBUG(lock, "completion AST, new lock mode");
441 ldlm_resource_unlink_lock(lock);
442 if (memcmp(dlm_req->lock_desc.l_resource.lr_name,
443 lock->l_resource->lr_name,
444 sizeof(__u64) * RES_NAME_SIZE) != 0) {
445 ldlm_lock_change_resource(lock, dlm_req->lock_desc.l_resource.lr_name);
446 LDLM_DEBUG(lock, "completion AST, new resource");
448 lock->l_resource->lr_tmp = &ast_list;
449 ldlm_grant_lock(lock);
450 lock->l_resource->lr_tmp = NULL;
451 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
452 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
455 ldlm_run_ast_work(&ast_list);
457 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
462 static int ldlm_callback_handler(struct ptlrpc_request *req)
467 rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
469 CERROR("lustre_ldlm: Invalid request\n");
473 if (req->rq_reqmsg->type != PTL_RPC_MSG_REQUEST) {
474 CERROR("lustre_ldlm: wrong packet type sent %d\n",
475 req->rq_reqmsg->type);
476 GOTO(out, rc = -EINVAL);
478 switch (req->rq_reqmsg->opc) {
479 case LDLM_BL_CALLBACK:
480 CDEBUG(D_INODE, "blocking ast\n");
481 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
482 rc = ldlm_handle_bl_callback(req);
484 case LDLM_CP_CALLBACK:
485 CDEBUG(D_INODE, "completion ast\n");
486 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CP_CALLBACK, 0);
487 rc = ldlm_handle_cp_callback(req);
491 rc = ptlrpc_error(req->rq_svc, req);
498 RETURN(ptlrpc_error(req->rq_svc, req));
503 static int ldlm_iocontrol(long cmd, struct lustre_handle *conn, int len,
504 void *karg, void *uarg)
506 struct obd_device *obddev = class_conn2obd(conn);
507 struct ptlrpc_connection *connection;
511 if (_IOC_TYPE(cmd) != IOC_LDLM_TYPE || _IOC_NR(cmd) < IOC_LDLM_MIN_NR ||
512 _IOC_NR(cmd) > IOC_LDLM_MAX_NR) {
513 CDEBUG(D_IOCTL, "invalid ioctl (type %ld, nr %ld, size %ld)\n",
514 _IOC_TYPE(cmd), _IOC_NR(cmd), _IOC_SIZE(cmd));
518 OBD_ALLOC(obddev->u.ldlm.ldlm_client,
519 sizeof(*obddev->u.ldlm.ldlm_client));
520 ptlrpc_init_client(NULL, NULL,
521 LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
522 obddev->u.ldlm.ldlm_client);
523 connection = ptlrpc_uuid_to_connection("ldlm");
525 CERROR("No LDLM UUID found: assuming ldlm is local.\n");
529 err = ldlm_test(obddev, conn);
530 CERROR("-- done err %d\n", err);
533 ldlm_dump_all_namespaces();
536 GOTO(out, err = -EINVAL);
541 ptlrpc_put_connection(connection);
542 OBD_FREE(obddev->u.ldlm.ldlm_client,
543 sizeof(*obddev->u.ldlm.ldlm_client));
547 #define LDLM_NUM_THREADS 8
549 static int ldlm_setup(struct obd_device *obddev, obd_count len, void *buf)
551 struct ldlm_obd *ldlm = &obddev->u.ldlm;
556 rc = ldlm_proc_setup(obddev);
560 ldlm->ldlm_service = ptlrpc_init_svc(64 * 1024, LDLM_REQUEST_PORTAL,
561 LDLM_REPLY_PORTAL, "self",
562 ldlm_callback_handler, "ldlm");
563 if (!ldlm->ldlm_service)
564 GOTO(out_proc, rc = -ENOMEM);
566 for (i = 0; i < LDLM_NUM_THREADS; i++) {
568 sprintf(name, "lustre_dlm_%02d", i);
569 rc = ptlrpc_start_thread(obddev, ldlm->ldlm_service, name);
571 CERROR("cannot start LDLM thread #%d: rc %d\n", i, rc);
573 GOTO(out_thread, rc);
577 INIT_LIST_HEAD(&waiting_locks_list);
578 spin_lock_init(&waiting_locks_spinlock);
579 waiting_locks_timer.function = waiting_locks_callback;
580 waiting_locks_timer.data = 0;
581 init_timer(&waiting_locks_timer);
586 ptlrpc_stop_all_threads(ldlm->ldlm_service);
587 ptlrpc_unregister_service(ldlm->ldlm_service);
589 ldlm_proc_cleanup(obddev);
595 static int ldlm_cleanup(struct obd_device *obddev)
597 struct ldlm_obd *ldlm = &obddev->u.ldlm;
600 if (!list_empty(&ldlm_namespace_list)) {
601 CERROR("ldlm still has namespaces; clean these up first.\n");
605 ptlrpc_stop_all_threads(ldlm->ldlm_service);
606 ptlrpc_unregister_service(ldlm->ldlm_service);
607 ldlm_proc_cleanup(obddev);
613 struct obd_ops ldlm_obd_ops = {
614 o_iocontrol: ldlm_iocontrol,
616 o_cleanup: ldlm_cleanup,
617 o_connect: class_connect,
618 o_disconnect: class_disconnect
621 static int __init ldlm_init(void)
623 int rc = class_register_type(&ldlm_obd_ops, OBD_LDLM_DEVICENAME);
627 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
628 sizeof(struct ldlm_resource), 0,
629 SLAB_HWCACHE_ALIGN, NULL, NULL);
630 if (ldlm_resource_slab == NULL)
633 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
634 sizeof(struct ldlm_lock), 0,
635 SLAB_HWCACHE_ALIGN, NULL, NULL);
636 if (ldlm_lock_slab == NULL) {
637 kmem_cache_destroy(ldlm_resource_slab);
644 static void __exit ldlm_exit(void)
646 class_unregister_type(OBD_LDLM_DEVICENAME);
647 if (kmem_cache_destroy(ldlm_resource_slab) != 0)
648 CERROR("couldn't free ldlm resource slab\n");
649 if (kmem_cache_destroy(ldlm_lock_slab) != 0)
650 CERROR("couldn't free ldlm lock slab\n");
653 EXPORT_SYMBOL(ldlm_completion_ast);
654 EXPORT_SYMBOL(ldlm_handle_enqueue);
655 EXPORT_SYMBOL(ldlm_handle_cancel);
656 EXPORT_SYMBOL(ldlm_handle_convert);
657 EXPORT_SYMBOL(ldlm_register_intent);
658 EXPORT_SYMBOL(ldlm_unregister_intent);
659 EXPORT_SYMBOL(ldlm_lockname);
660 EXPORT_SYMBOL(ldlm_typename);
661 EXPORT_SYMBOL(ldlm_handle2lock);
662 EXPORT_SYMBOL(ldlm_lock2handle);
663 EXPORT_SYMBOL(ldlm_lock_put);
664 EXPORT_SYMBOL(ldlm_lock_match);
665 EXPORT_SYMBOL(ldlm_lock_addref);
666 EXPORT_SYMBOL(ldlm_lock_decref);
667 EXPORT_SYMBOL(ldlm_lock_change_resource);
668 EXPORT_SYMBOL(ldlm_cli_convert);
669 EXPORT_SYMBOL(ldlm_cli_enqueue);
670 EXPORT_SYMBOL(ldlm_cli_cancel);
671 EXPORT_SYMBOL(ldlm_match_or_enqueue);
672 EXPORT_SYMBOL(ldlm_it2str);
673 EXPORT_SYMBOL(ldlm_test);
674 EXPORT_SYMBOL(ldlm_regression_start);
675 EXPORT_SYMBOL(ldlm_regression_stop);
676 EXPORT_SYMBOL(ldlm_lock_dump);
677 EXPORT_SYMBOL(ldlm_namespace_new);
678 EXPORT_SYMBOL(ldlm_namespace_free);
679 EXPORT_SYMBOL(l_lock);
680 EXPORT_SYMBOL(l_unlock);
682 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
683 MODULE_DESCRIPTION("Lustre Lock Management Module v0.1");
684 MODULE_LICENSE("GPL");
686 module_init(ldlm_init);
687 module_exit(ldlm_exit);