Whamcloud - gitweb
1b4e46ed3beb200224128c31c98338cee993419f
[fs/lustre-release.git] / lustre / ldlm / ldlm_resource.c
1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2  * vim:expandtab:shiftwidth=8:tabstop=8:
3  *
4  * GPL HEADER START
5  *
6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
7  *
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.
11  *
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).
17  *
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
21  *
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
24  * have any questions.
25  *
26  * GPL HEADER END
27  */
28 /*
29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
30  * Use is subject to license terms.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  *
36  * lustre/ldlm/ldlm_resource.c
37  *
38  * Author: Phil Schwan <phil@clusterfs.com>
39  * Author: Peter Braam <braam@clusterfs.com>
40  */
41
42 #define DEBUG_SUBSYSTEM S_LDLM
43 #ifdef __KERNEL__
44 # include <lustre_dlm.h>
45 #else
46 # include <liblustre.h>
47 #endif
48
49 #include <obd_class.h>
50 #include "ldlm_internal.h"
51
52 cfs_mem_cache_t *ldlm_resource_slab, *ldlm_lock_slab;
53
54 atomic_t ldlm_srv_namespace_nr = ATOMIC_INIT(0);
55 atomic_t ldlm_cli_namespace_nr = ATOMIC_INIT(0);
56
57 struct semaphore ldlm_srv_namespace_lock;
58 CFS_LIST_HEAD(ldlm_srv_namespace_list);
59
60 struct semaphore ldlm_cli_namespace_lock;
61 CFS_LIST_HEAD(ldlm_cli_namespace_list);
62
63 cfs_proc_dir_entry_t *ldlm_type_proc_dir = NULL;
64 cfs_proc_dir_entry_t *ldlm_ns_proc_dir = NULL;
65 cfs_proc_dir_entry_t *ldlm_svc_proc_dir = NULL;
66
67 #ifdef LPROCFS
68 static int ldlm_proc_dump_ns(struct file *file, const char *buffer,
69                              unsigned long count, void *data)
70 {
71         ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
72         ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
73         RETURN(count);
74 }
75
76 int ldlm_proc_setup(void)
77 {
78         int rc;
79         struct lprocfs_vars list[] = {
80                 { "dump_namespaces", NULL, ldlm_proc_dump_ns, NULL },
81                 { NULL }};
82         ENTRY;
83         LASSERT(ldlm_ns_proc_dir == NULL);
84
85         ldlm_type_proc_dir = lprocfs_register(OBD_LDLM_DEVICENAME,
86                                               proc_lustre_root,
87                                               NULL, NULL);
88         if (IS_ERR(ldlm_type_proc_dir)) {
89                 CERROR("LProcFS failed in ldlm-init\n");
90                 rc = PTR_ERR(ldlm_type_proc_dir);
91                 GOTO(err, rc);
92         }
93
94         ldlm_ns_proc_dir = lprocfs_register("namespaces",
95                                             ldlm_type_proc_dir,
96                                             NULL, NULL);
97         if (IS_ERR(ldlm_ns_proc_dir)) {
98                 CERROR("LProcFS failed in ldlm-init\n");
99                 rc = PTR_ERR(ldlm_ns_proc_dir);
100                 GOTO(err_type, rc);
101         }
102
103         ldlm_svc_proc_dir = lprocfs_register("services",
104                                             ldlm_type_proc_dir,
105                                             NULL, NULL);
106         if (IS_ERR(ldlm_svc_proc_dir)) {
107                 CERROR("LProcFS failed in ldlm-init\n");
108                 rc = PTR_ERR(ldlm_svc_proc_dir);
109                 GOTO(err_ns, rc);
110         }
111
112         rc = lprocfs_add_vars(ldlm_type_proc_dir, list, NULL);
113
114         RETURN(0);
115
116 err_ns:
117         lprocfs_remove(&ldlm_ns_proc_dir);
118 err_type:
119         lprocfs_remove(&ldlm_type_proc_dir);
120 err:
121         ldlm_svc_proc_dir = NULL;
122         RETURN(rc);
123 }
124
125 void ldlm_proc_cleanup(void)
126 {
127         if (ldlm_svc_proc_dir)
128                 lprocfs_remove(&ldlm_svc_proc_dir);
129
130         if (ldlm_ns_proc_dir)
131                 lprocfs_remove(&ldlm_ns_proc_dir);
132
133         if (ldlm_type_proc_dir)
134                 lprocfs_remove(&ldlm_type_proc_dir);
135 }
136
137 static int lprocfs_rd_lru_size(char *page, char **start, off_t off,
138                                int count, int *eof, void *data)
139 {
140         struct ldlm_namespace *ns = data;
141         __u32 *nr = &ns->ns_max_unused;
142
143         if (ns_connect_lru_resize(ns))
144                 nr = &ns->ns_nr_unused;
145         return lprocfs_rd_uint(page, start, off, count, eof, nr);
146 }
147
148 static int lprocfs_wr_lru_size(struct file *file, const char *buffer,
149                                unsigned long count, void *data)
150 {
151         struct ldlm_namespace *ns = data;
152         char dummy[MAX_STRING_SIZE + 1], *end;
153         unsigned long tmp;
154         int lru_resize;
155
156         dummy[MAX_STRING_SIZE] = '\0';
157         if (copy_from_user(dummy, buffer, MAX_STRING_SIZE))
158                 return -EFAULT;
159
160         if (strncmp(dummy, "clear", 5) == 0) {
161                 CDEBUG(D_DLMTRACE,
162                        "dropping all unused locks from namespace %s\n",
163                        ns->ns_name);
164                 if (ns_connect_lru_resize(ns)) {
165                         int canceled, unused  = ns->ns_nr_unused;
166
167                         /* Try to cancel all @ns_nr_unused locks. */
168                         canceled = ldlm_cancel_lru(ns, unused, LDLM_SYNC,
169                                                    LDLM_CANCEL_PASSED);
170                         if (canceled < unused) {
171                                 CERROR("not all requested locks are canceled, "
172                                        "requested: %d, canceled: %d\n", unused,
173                                        canceled);
174                                 return -EINVAL;
175                         }
176                 } else {
177                         tmp = ns->ns_max_unused;
178                         ns->ns_max_unused = 0;
179                         ldlm_cancel_lru(ns, 0, LDLM_SYNC, LDLM_CANCEL_PASSED);
180                         ns->ns_max_unused = tmp;
181                 }
182                 return count;
183         }
184
185         tmp = simple_strtoul(dummy, &end, 0);
186         if (dummy == end) {
187                 CERROR("invalid value written\n");
188                 return -EINVAL;
189         }
190         lru_resize = (tmp == 0);
191
192         if (ns_connect_lru_resize(ns)) {
193                 if (!lru_resize)
194                         ns->ns_max_unused = (unsigned int)tmp;
195
196                 if (tmp > ns->ns_nr_unused)
197                         tmp = ns->ns_nr_unused;
198                 tmp = ns->ns_nr_unused - tmp;
199
200                 CDEBUG(D_DLMTRACE,
201                        "changing namespace %s unused locks from %u to %u\n",
202                        ns->ns_name, ns->ns_nr_unused, (unsigned int)tmp);
203                 ldlm_cancel_lru(ns, tmp, LDLM_ASYNC, LDLM_CANCEL_PASSED);
204
205                 if (!lru_resize) {
206                         CDEBUG(D_DLMTRACE,
207                                "disable lru_resize for namespace %s\n",
208                                ns->ns_name);
209                         ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
210                 }
211         } else {
212                 CDEBUG(D_DLMTRACE,
213                        "changing namespace %s max_unused from %u to %u\n",
214                        ns->ns_name, ns->ns_max_unused, (unsigned int)tmp);
215                 ns->ns_max_unused = (unsigned int)tmp;
216                 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, LDLM_CANCEL_PASSED);
217
218                 /* Make sure that originally lru resize was supported before
219                  * turning it on here. */
220                 if (lru_resize &&
221                     (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
222                         CDEBUG(D_DLMTRACE,
223                                "enable lru_resize for namespace %s\n",
224                                ns->ns_name);
225                         ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
226                 }
227         }
228
229         return count;
230 }
231
232 void ldlm_proc_namespace(struct ldlm_namespace *ns)
233 {
234         struct lprocfs_vars lock_vars[2];
235         char lock_name[MAX_STRING_SIZE + 1];
236
237         LASSERT(ns != NULL);
238         LASSERT(ns->ns_name != NULL);
239
240         lock_name[MAX_STRING_SIZE] = '\0';
241
242         memset(lock_vars, 0, sizeof(lock_vars));
243         lock_vars[0].name = lock_name;
244
245         snprintf(lock_name, MAX_STRING_SIZE, "%s/resource_count", ns->ns_name);
246         lock_vars[0].data = &ns->ns_refcount;
247         lock_vars[0].read_fptr = lprocfs_rd_atomic;
248         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
249
250         snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_count", ns->ns_name);
251         lock_vars[0].data = &ns->ns_locks;
252         lock_vars[0].read_fptr = lprocfs_rd_atomic;
253         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
254
255         if (ns_is_client(ns)) {
256                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_unused_count",
257                          ns->ns_name);
258                 lock_vars[0].data = &ns->ns_nr_unused;
259                 lock_vars[0].read_fptr = lprocfs_rd_uint;
260                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
261
262                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_size",
263                          ns->ns_name);
264                 lock_vars[0].data = ns;
265                 lock_vars[0].read_fptr = lprocfs_rd_lru_size;
266                 lock_vars[0].write_fptr = lprocfs_wr_lru_size;
267                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
268
269                 snprintf(lock_name, MAX_STRING_SIZE, "%s/shrink_thumb",
270                          ns->ns_name);
271                 lock_vars[0].data = ns;
272                 lock_vars[0].read_fptr = lprocfs_rd_uint;
273                 lock_vars[0].write_fptr = lprocfs_wr_uint;
274                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
275
276                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_max_age",
277                          ns->ns_name);
278                 lock_vars[0].data = &ns->ns_max_age;
279                 lock_vars[0].read_fptr = lprocfs_rd_uint;
280                 lock_vars[0].write_fptr = lprocfs_wr_uint;
281                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
282         } else {
283                 snprintf(lock_name, MAX_STRING_SIZE, "%s/ctime_age_limit",
284                          ns->ns_name);
285                 lock_vars[0].data = &ns->ns_ctime_age_limit;
286                 lock_vars[0].read_fptr = lprocfs_rd_uint;
287                 lock_vars[0].write_fptr = lprocfs_wr_uint;
288                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
289
290                 snprintf(lock_name, MAX_STRING_SIZE, "%s/max_nolock_bytes",
291                          ns->ns_name);
292                 lock_vars[0].data = &ns->ns_max_nolock_size;
293                 lock_vars[0].read_fptr = lprocfs_rd_uint;
294                 lock_vars[0].write_fptr = lprocfs_wr_uint;
295                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
296
297                 snprintf(lock_name, MAX_STRING_SIZE, "%s/contention_seconds",
298                          ns->ns_name);
299                 lock_vars[0].data = &ns->ns_contention_time;
300                 lock_vars[0].read_fptr = lprocfs_rd_uint;
301                 lock_vars[0].write_fptr = lprocfs_wr_uint;
302                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
303
304                 snprintf(lock_name, MAX_STRING_SIZE, "%s/contended_locks",
305                          ns->ns_name);
306                 lock_vars[0].data = &ns->ns_contended_locks;
307                 lock_vars[0].read_fptr = lprocfs_rd_uint;
308                 lock_vars[0].write_fptr = lprocfs_wr_uint;
309                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
310         }
311 }
312 #undef MAX_STRING_SIZE
313 #else
314 #define ldlm_proc_namespace(ns) do {} while (0)
315 #endif /* LPROCFS */
316
317 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
318                                           ldlm_side_t client, ldlm_appetite_t apt)
319 {
320         struct ldlm_namespace *ns = NULL;
321         struct list_head *bucket;
322         int rc, idx, namelen;
323         ENTRY;
324
325         rc = ldlm_get_ref();
326         if (rc) {
327                 CERROR("ldlm_get_ref failed: %d\n", rc);
328                 RETURN(NULL);
329         }
330
331         OBD_ALLOC_PTR(ns);
332         if (!ns)
333                 GOTO(out_ref, NULL);
334
335         OBD_VMALLOC(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
336         if (!ns->ns_hash)
337                 GOTO(out_ns, NULL);
338
339         ns->ns_shrink_thumb = LDLM_LOCK_SHRINK_THUMB;
340         ns->ns_appetite = apt;
341
342         LASSERT(obd != NULL);
343         ns->ns_obd = obd;
344
345         namelen = strlen(name);
346         OBD_ALLOC(ns->ns_name, namelen + 1);
347         if (!ns->ns_name)
348                 GOTO(out_hash, NULL);
349
350         strcpy(ns->ns_name, name);
351
352         CFS_INIT_LIST_HEAD(&ns->ns_root_list);
353         CFS_INIT_LIST_HEAD(&ns->ns_list_chain);
354         ns->ns_refcount = 0;
355         ns->ns_client = client;
356         spin_lock_init(&ns->ns_hash_lock);
357         atomic_set(&ns->ns_locks, 0);
358         ns->ns_resources = 0;
359         cfs_waitq_init(&ns->ns_waitq);
360         ns->ns_max_nolock_size = NS_DEFAULT_MAX_NOLOCK_BYTES;
361         ns->ns_contention_time = NS_DEFAULT_CONTENTION_SECONDS;
362         ns->ns_contended_locks = NS_DEFAULT_CONTENDED_LOCKS;
363
364         for (bucket = ns->ns_hash + RES_HASH_SIZE - 1; bucket >= ns->ns_hash;
365              bucket--)
366                 CFS_INIT_LIST_HEAD(bucket);
367
368         CFS_INIT_LIST_HEAD(&ns->ns_unused_list);
369         ns->ns_nr_unused = 0;
370         ns->ns_max_unused = LDLM_DEFAULT_LRU_SIZE;
371         ns->ns_max_age = LDLM_DEFAULT_MAX_ALIVE;
372         ns->ns_ctime_age_limit = LDLM_CTIME_AGE_LIMIT;
373         spin_lock_init(&ns->ns_unused_lock);
374         ns->ns_orig_connect_flags = 0;
375         ns->ns_connect_flags = 0;
376         ldlm_proc_namespace(ns);
377
378         idx = atomic_read(ldlm_namespace_nr(client));
379         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
380         if (rc) {
381                 CERROR("Can't initialize lock pool, rc %d\n", rc);
382                 GOTO(out_proc, rc);
383         }
384
385         at_init(&ns->ns_at_estimate, ldlm_enqueue_min, 0);
386
387         ldlm_namespace_register(ns, client);
388         RETURN(ns);
389 out_proc:
390         ldlm_namespace_cleanup(ns, 0);
391         OBD_FREE(ns->ns_name, namelen + 1);
392 out_hash:
393         OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
394 out_ns:
395         OBD_FREE_PTR(ns);
396 out_ref:
397         ldlm_put_ref();
398         RETURN(NULL);
399 }
400
401 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
402
403 /* If flags contains FL_LOCAL_ONLY, don't try to tell the server, just cleanup.
404  * This is currently only used for recovery, and we make certain assumptions
405  * as a result--notably, that we shouldn't cancel locks with refs. -phil
406  *
407  * Called with the ns_lock held. */
408 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
409                              int flags)
410 {
411         struct list_head *tmp;
412         int rc = 0, client = ns_is_client(res->lr_namespace);
413         int local_only = (flags & LDLM_FL_LOCAL_ONLY);
414         ENTRY;
415
416
417         do {
418                 struct ldlm_lock *lock = NULL;
419
420                 /* first, we look for non-cleaned-yet lock
421                  * all cleaned locks are marked by CLEANED flag */
422                 lock_res(res);
423                 list_for_each(tmp, q) {
424                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
425                         if (lock->l_flags & LDLM_FL_CLEANED) {
426                                 lock = NULL;
427                                 continue;
428                         }
429                         LDLM_LOCK_GET(lock);
430                         lock->l_flags |= LDLM_FL_CLEANED;
431                         break;
432                 }
433
434                 if (lock == NULL) {
435                         unlock_res(res);
436                         break;
437                 }
438
439                 /* Set CBPENDING so nothing in the cancellation path
440                  * can match this lock */
441                 lock->l_flags |= LDLM_FL_CBPENDING;
442                 lock->l_flags |= LDLM_FL_FAILED;
443                 lock->l_flags |= flags;
444
445                 /* ... without sending a CANCEL message for local_only. */
446                 if (local_only)
447                         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
448
449                 if (local_only && (lock->l_readers || lock->l_writers)) {
450                         /* This is a little bit gross, but much better than the
451                          * alternative: pretend that we got a blocking AST from
452                          * the server, so that when the lock is decref'd, it
453                          * will go away ... */
454                         unlock_res(res);
455                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
456                         if (lock->l_completion_ast)
457                                 lock->l_completion_ast(lock, 0, NULL);
458                         LDLM_LOCK_RELEASE(lock);
459                         continue;
460                 }
461
462                 if (client) {
463                         struct lustre_handle lockh;
464
465                         unlock_res(res);
466                         ldlm_lock2handle(lock, &lockh);
467                         rc = ldlm_cli_cancel(&lockh);
468                         if (rc)
469                                 CERROR("ldlm_cli_cancel: %d\n", rc);
470                 } else {
471                         ldlm_resource_unlink_lock(lock);
472                         unlock_res(res);
473                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
474                                    "client node");
475                         ldlm_lock_destroy(lock);
476                 }
477                 LDLM_LOCK_RELEASE(lock);
478         } while (1);
479
480         EXIT;
481 }
482
483 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int flags)
484 {
485         struct list_head *tmp;
486         int i;
487
488         if (ns == NULL) {
489                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
490                 return ELDLM_OK;
491         }
492
493         for (i = 0; i < RES_HASH_SIZE; i++) {
494                 spin_lock(&ns->ns_hash_lock);
495                 tmp = ns->ns_hash[i].next;
496                 while (tmp != &(ns->ns_hash[i])) {
497                         struct ldlm_resource *res;
498                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
499                         ldlm_resource_getref(res);
500                         spin_unlock(&ns->ns_hash_lock);
501                         LDLM_RESOURCE_ADDREF(res);
502
503                         cleanup_resource(res, &res->lr_granted, flags);
504                         cleanup_resource(res, &res->lr_converting, flags);
505                         cleanup_resource(res, &res->lr_waiting, flags);
506
507                         spin_lock(&ns->ns_hash_lock);
508                         tmp  = tmp->next;
509
510                         /* XXX: former stuff caused issues in case of race
511                          * between ldlm_namespace_cleanup() and lockd() when
512                          * client gets blocking ast when lock gets distracted by
513                          * server. This is 1_4 branch solution, let's see how
514                          * will it behave. */
515                         LDLM_RESOURCE_DELREF(res);
516                         if (!ldlm_resource_putref_locked(res))
517                                 CDEBUG(D_INFO,
518                                        "Namespace %s resource refcount nonzero "
519                                        "(%d) after lock cleanup; forcing cleanup.\n",
520                                        ns->ns_name, atomic_read(&res->lr_refcount));
521                 }
522                 spin_unlock(&ns->ns_hash_lock);
523         }
524
525         return ELDLM_OK;
526 }
527
528 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
529 {
530         ENTRY;
531
532         /* At shutdown time, don't call the cancellation callback */
533         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
534
535         if (ns->ns_refcount > 0) {
536                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
537                 int rc;
538                 CDEBUG(D_DLMTRACE,
539                        "dlm namespace %s free waiting on refcount %d\n",
540                        ns->ns_name, ns->ns_refcount);
541 force_wait:
542                 if (force)
543                         lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
544
545                 rc = l_wait_event(ns->ns_waitq,
546                                   ns->ns_refcount == 0, &lwi);
547
548                 /* Forced cleanups should be able to reclaim all references,
549                  * so it's safe to wait forever... we can't leak locks... */
550                 if (force && rc == -ETIMEDOUT) {
551                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
552                                        "namespace with %d resources in use, "
553                                        "(rc=%d)\n", ns->ns_name,
554                                        ns->ns_refcount, rc);
555                         GOTO(force_wait, rc);
556                 }
557
558                 if (ns->ns_refcount) {
559                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
560                                        "with %d resources in use, (rc=%d)\n",
561                                        ns->ns_name,
562                                        ns->ns_refcount, rc);
563                         RETURN(ELDLM_NAMESPACE_EXISTS);
564                 }
565                 CDEBUG(D_DLMTRACE,
566                        "dlm namespace %s free done waiting\n", ns->ns_name);
567         }
568
569         RETURN(ELDLM_OK);
570 }
571
572 /**
573  * Performs various cleanups for passed \a ns to make it drop refc and be ready
574  * for freeing. Waits for refc == 0.
575  *
576  * The following is done:
577  * (0) Unregister \a ns from its list to make inaccessible for potential users
578  * like pools thread and others;
579  * (1) Clear all locks in \a ns.
580  */
581 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
582                                struct obd_import *imp,
583                                int force)
584 {
585         int rc;
586         ENTRY;
587         if (!ns) {
588                 EXIT;
589                 return;
590         }
591
592         /*
593          * Make sure that nobody can find this ns in its list.
594          */
595         ldlm_namespace_unregister(ns, ns->ns_client);
596
597         /*
598          * Can fail with -EINTR when force == 0 in which case try harder.
599          */
600         rc = __ldlm_namespace_free(ns, force);
601         if (rc != ELDLM_OK) {
602                 if (imp) {
603                         ptlrpc_disconnect_import(imp, 0);
604                         ptlrpc_invalidate_import(imp);
605                 }
606
607                 /*
608                  * With all requests dropped and the import inactive
609                  * we are gaurenteed all reference will be dropped.
610                  */
611                 rc = __ldlm_namespace_free(ns, 1);
612                 LASSERT(rc == 0);
613         }
614         EXIT;
615 }
616
617 /**
618  * Performs freeing memory structures related to \a ns. This is only done when
619  * ldlm_namespce_free_prior() successfully removed all resources referencing
620  * \a ns and its refc == 0.
621  */
622 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
623 {
624         ENTRY;
625         if (!ns) {
626                 EXIT;
627                 return;
628         }
629
630         /*
631          * Fini pool _before_ parent proc dir is removed. This is important as
632          * ldlm_pool_fini() removes own proc dir which is child to @dir. Removing
633          * it after @dir may cause oops.
634          */
635         ldlm_pool_fini(&ns->ns_pool);
636
637 #ifdef LPROCFS
638         {
639                 struct proc_dir_entry *dir;
640                 dir = lprocfs_srch(ldlm_ns_proc_dir, ns->ns_name);
641                 if (dir == NULL) {
642                         CERROR("dlm namespace %s has no procfs dir?\n",
643                                ns->ns_name);
644                 } else {
645                         lprocfs_remove(&dir);
646                 }
647         }
648 #endif
649
650         OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
651         OBD_FREE(ns->ns_name, strlen(ns->ns_name) + 1);
652
653         /*
654          * Namespace \a ns should be not on list in this time, otherwise this
655          * will cause issues realted to using freed \a ns in pools thread.
656          */
657         LASSERT(list_empty(&ns->ns_list_chain));
658         OBD_FREE_PTR(ns);
659         ldlm_put_ref();
660         EXIT;
661 }
662
663
664 /* Cleanup the resource, and free namespace.
665  * bug 12864:
666  * Deadlock issue:
667  * proc1: destroy import
668  *        class_disconnect_export(grab cl_sem) ->
669  *              -> ldlm_namespace_free ->
670  *              -> lprocfs_remove(grab _lprocfs_lock).
671  * proc2: read proc info
672  *        lprocfs_fops_read(grab _lprocfs_lock) ->
673  *              -> osc_rd_active, etc(grab cl_sem).
674  *
675  * So that I have to split the ldlm_namespace_free into two parts - the first
676  * part ldlm_namespace_free_prior is used to cleanup the resource which is
677  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
678  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
679  * held.
680  */
681 void ldlm_namespace_free(struct ldlm_namespace *ns,
682                          struct obd_import *imp,
683                          int force)
684 {
685         ldlm_namespace_free_prior(ns, imp, force);
686         ldlm_namespace_free_post(ns);
687 }
688
689
690 void ldlm_namespace_get_locked(struct ldlm_namespace *ns)
691 {
692         LASSERT(ns->ns_refcount >= 0);
693         ns->ns_refcount++;
694 }
695
696 void ldlm_namespace_get(struct ldlm_namespace *ns)
697 {
698         spin_lock(&ns->ns_hash_lock);
699         ldlm_namespace_get_locked(ns);
700         spin_unlock(&ns->ns_hash_lock);
701 }
702
703 void ldlm_namespace_put_locked(struct ldlm_namespace *ns, int wakeup)
704 {
705         LASSERT(ns->ns_refcount > 0);
706         ns->ns_refcount--;
707         if (ns->ns_refcount == 0 && wakeup)
708                 wake_up(&ns->ns_waitq);
709 }
710
711 void ldlm_namespace_put(struct ldlm_namespace *ns, int wakeup)
712 {
713         spin_lock(&ns->ns_hash_lock);
714         ldlm_namespace_put_locked(ns, wakeup);
715         spin_unlock(&ns->ns_hash_lock);
716 }
717
718 /* Register @ns in the list of namespaces */
719 void ldlm_namespace_register(struct ldlm_namespace *ns, ldlm_side_t client)
720 {
721         mutex_down(ldlm_namespace_lock(client));
722         LASSERT(list_empty(&ns->ns_list_chain));
723         list_add(&ns->ns_list_chain, ldlm_namespace_list(client));
724         atomic_inc(ldlm_namespace_nr(client));
725         mutex_up(ldlm_namespace_lock(client));
726 }
727
728 /* Unregister @ns from the list of namespaces */
729 void ldlm_namespace_unregister(struct ldlm_namespace *ns, ldlm_side_t client)
730 {
731         mutex_down(ldlm_namespace_lock(client));
732         LASSERT(!list_empty(&ns->ns_list_chain));
733         /*
734          * Some asserts and possibly other parts of code still using
735          * list_empty(&ns->ns_list_chain). This is why it is important
736          * to use list_del_init() here.
737          */
738         list_del_init(&ns->ns_list_chain);
739         atomic_dec(ldlm_namespace_nr(client));
740         mutex_up(ldlm_namespace_lock(client));
741 }
742
743 /* Should be called under ldlm_namespace_lock(client) taken */
744 void ldlm_namespace_move_locked(struct ldlm_namespace *ns, ldlm_side_t client)
745 {
746         LASSERT(!list_empty(&ns->ns_list_chain));
747         LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
748         list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
749 }
750
751 /* Should be called under ldlm_namespace_lock(client) taken */
752 struct ldlm_namespace *ldlm_namespace_first_locked(ldlm_side_t client)
753 {
754         LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
755         LASSERT(!list_empty(ldlm_namespace_list(client)));
756         return container_of(ldlm_namespace_list(client)->next,
757                 struct ldlm_namespace, ns_list_chain);
758 }
759 static __u32 ldlm_hash_fn(struct ldlm_resource *parent,
760                           const struct ldlm_res_id *name)
761 {
762         __u32 hash = 0;
763         int i;
764
765         for (i = 0; i < RES_NAME_SIZE; i++)
766                 hash += name->name[i];
767
768         hash += (__u32)((unsigned long)parent >> 4);
769
770         return (hash & RES_HASH_MASK);
771 }
772
773 static struct ldlm_resource *ldlm_resource_new(void)
774 {
775         struct ldlm_resource *res;
776         int idx;
777
778         OBD_SLAB_ALLOC(res, ldlm_resource_slab, CFS_ALLOC_IO, sizeof *res);
779         if (res == NULL)
780                 return NULL;
781
782         memset(res, 0, sizeof(*res));
783
784         CFS_INIT_LIST_HEAD(&res->lr_children);
785         CFS_INIT_LIST_HEAD(&res->lr_childof);
786         CFS_INIT_LIST_HEAD(&res->lr_granted);
787         CFS_INIT_LIST_HEAD(&res->lr_converting);
788         CFS_INIT_LIST_HEAD(&res->lr_waiting);
789
790         /* initialize interval trees for each lock mode*/
791         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
792                 res->lr_itree[idx].lit_size = 0;
793                 res->lr_itree[idx].lit_mode = 1 << idx;
794                 res->lr_itree[idx].lit_root = NULL;
795         }
796
797         atomic_set(&res->lr_refcount, 1);
798         spin_lock_init(&res->lr_lock);
799         lu_ref_init(&res->lr_reference);
800
801         /* one who creates the resource must unlock
802          * the semaphore after lvb initialization */
803         init_MUTEX_LOCKED(&res->lr_lvb_sem);
804
805         return res;
806 }
807
808 /* must be called with hash lock held */
809 static struct ldlm_resource *
810 ldlm_resource_find(struct ldlm_namespace *ns, const struct ldlm_res_id *name,
811                    __u32 hash)
812 {
813         struct list_head *bucket, *tmp;
814         struct ldlm_resource *res;
815
816         LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
817         bucket = ns->ns_hash + hash;
818
819         list_for_each(tmp, bucket) {
820                 res = list_entry(tmp, struct ldlm_resource, lr_hash);
821                 if (memcmp(&res->lr_name, name, sizeof(res->lr_name)) == 0)
822                         return res;
823         }
824
825         return NULL;
826 }
827
828 /* Args: locked namespace
829  * Returns: newly-allocated, referenced, unlocked resource */
830 static struct ldlm_resource *
831 ldlm_resource_add(struct ldlm_namespace *ns, struct ldlm_resource *parent,
832                   const struct ldlm_res_id *name, __u32 hash, ldlm_type_t type)
833 {
834         struct list_head *bucket;
835         struct ldlm_resource *res, *old_res;
836         ENTRY;
837
838         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
839                  "type: %d\n", type);
840
841         res = ldlm_resource_new();
842         if (!res)
843                 RETURN(NULL);
844
845         res->lr_name = *name;
846         res->lr_namespace = ns;
847         res->lr_type = type;
848         res->lr_most_restr = LCK_NL;
849
850         spin_lock(&ns->ns_hash_lock);
851         old_res = ldlm_resource_find(ns, name, hash);
852         if (old_res) {
853                 /* someone won the race and added the resource before */
854                 ldlm_resource_getref(old_res);
855                 spin_unlock(&ns->ns_hash_lock);
856                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
857                 /* synchronize WRT resource creation */
858                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
859                         down(&old_res->lr_lvb_sem);
860                         up(&old_res->lr_lvb_sem);
861                 }
862                 RETURN(old_res);
863         }
864
865         /* we won! let's add the resource */
866         bucket = ns->ns_hash + hash;
867         list_add(&res->lr_hash, bucket);
868         ns->ns_resources++;
869         ldlm_namespace_get_locked(ns);
870
871         if (parent == NULL) {
872                 list_add(&res->lr_childof, &ns->ns_root_list);
873         } else {
874                 res->lr_parent = parent;
875                 list_add(&res->lr_childof, &parent->lr_children);
876         }
877         spin_unlock(&ns->ns_hash_lock);
878
879         if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
880                 int rc;
881
882                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
883                 rc = ns->ns_lvbo->lvbo_init(res);
884                 if (rc)
885                         CERROR("lvbo_init failed for resource "
886                                LPU64": rc %d\n", name->name[0], rc);
887                 /* we create resource with locked lr_lvb_sem */
888                 up(&res->lr_lvb_sem);
889         }
890
891         RETURN(res);
892 }
893
894 /* Args: unlocked namespace
895  * Locks: takes and releases ns->ns_lock and res->lr_lock
896  * Returns: referenced, unlocked ldlm_resource or NULL */
897 struct ldlm_resource *
898 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
899                   const struct ldlm_res_id *name, ldlm_type_t type, int create)
900 {
901         __u32 hash = ldlm_hash_fn(parent, name);
902         struct ldlm_resource *res = NULL;
903         ENTRY;
904
905         LASSERT(ns != NULL);
906         LASSERT(ns->ns_hash != NULL);
907         LASSERT(name->name[0] != 0);
908
909         spin_lock(&ns->ns_hash_lock);
910         res = ldlm_resource_find(ns, name, hash);
911         if (res) {
912                 ldlm_resource_getref(res);
913                 spin_unlock(&ns->ns_hash_lock);
914                 /* synchronize WRT resource creation */
915                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
916                         down(&res->lr_lvb_sem);
917                         up(&res->lr_lvb_sem);
918                 }
919                 RETURN(res);
920         }
921         spin_unlock(&ns->ns_hash_lock);
922
923         if (create == 0)
924                 RETURN(NULL);
925
926         res = ldlm_resource_add(ns, parent, name, hash, type);
927         RETURN(res);
928 }
929
930 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
931 {
932         LASSERT(res != NULL);
933         LASSERT(res != LP_POISON);
934         atomic_inc(&res->lr_refcount);
935         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
936                atomic_read(&res->lr_refcount));
937         return res;
938 }
939
940 void __ldlm_resource_putref_final(struct ldlm_resource *res)
941 {
942         struct ldlm_namespace *ns = res->lr_namespace;
943
944         LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
945
946         if (!list_empty(&res->lr_granted)) {
947                 ldlm_resource_dump(D_ERROR, res);
948                 LBUG();
949         }
950
951         if (!list_empty(&res->lr_converting)) {
952                 ldlm_resource_dump(D_ERROR, res);
953                 LBUG();
954         }
955
956         if (!list_empty(&res->lr_waiting)) {
957                 ldlm_resource_dump(D_ERROR, res);
958                 LBUG();
959         }
960
961         if (!list_empty(&res->lr_children)) {
962                 ldlm_resource_dump(D_ERROR, res);
963                 LBUG();
964         }
965
966         /* Pass 0 here to not wake ->ns_waitq up yet, we will do it few
967          * lines below when all children are freed. */
968         ldlm_namespace_put_locked(ns, 0);
969         list_del_init(&res->lr_hash);
970         list_del_init(&res->lr_childof);
971         lu_ref_fini(&res->lr_reference);
972
973         ns->ns_resources--;
974         if (ns->ns_resources == 0)
975                 wake_up(&ns->ns_waitq);
976 }
977
978 /* Returns 1 if the resource was freed, 0 if it remains. */
979 int ldlm_resource_putref(struct ldlm_resource *res)
980 {
981         struct ldlm_namespace *ns = res->lr_namespace;
982         int rc = 0;
983         ENTRY;
984
985         CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
986                atomic_read(&res->lr_refcount) - 1);
987         LASSERTF(atomic_read(&res->lr_refcount) > 0, "%d",
988                  atomic_read(&res->lr_refcount));
989         LASSERTF(atomic_read(&res->lr_refcount) < LI_POISON, "%d",
990                  atomic_read(&res->lr_refcount));
991
992         if (atomic_dec_and_lock(&res->lr_refcount, &ns->ns_hash_lock)) {
993                 __ldlm_resource_putref_final(res);
994                 spin_unlock(&ns->ns_hash_lock);
995                 if (res->lr_lvb_data)
996                         OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
997                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
998                 rc = 1;
999         }
1000
1001         RETURN(rc);
1002 }
1003
1004 /* Returns 1 if the resource was freed, 0 if it remains. */
1005 int ldlm_resource_putref_locked(struct ldlm_resource *res)
1006 {
1007         int rc = 0;
1008         ENTRY;
1009
1010         CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
1011                atomic_read(&res->lr_refcount) - 1);
1012         LASSERT(atomic_read(&res->lr_refcount) > 0);
1013         LASSERT(atomic_read(&res->lr_refcount) < LI_POISON);
1014
1015         LASSERT(atomic_read(&res->lr_refcount) >= 0);
1016         if (atomic_dec_and_test(&res->lr_refcount)) {
1017                 __ldlm_resource_putref_final(res);
1018                 if (res->lr_lvb_data)
1019                         OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
1020                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1021                 rc = 1;
1022         }
1023
1024         RETURN(rc);
1025 }
1026
1027 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
1028                             struct ldlm_lock *lock)
1029 {
1030         check_res_locked(res);
1031
1032         ldlm_resource_dump(D_INFO, res);
1033         CDEBUG(D_OTHER, "About to add this lock:\n");
1034         ldlm_lock_dump(D_OTHER, lock, 0);
1035
1036         if (lock->l_destroyed) {
1037                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1038                 return;
1039         }
1040
1041         LASSERT(list_empty(&lock->l_res_link));
1042
1043         list_add_tail(&lock->l_res_link, head);
1044 }
1045
1046 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1047                                      struct ldlm_lock *new)
1048 {
1049         struct ldlm_resource *res = original->l_resource;
1050
1051         check_res_locked(res);
1052
1053         ldlm_resource_dump(D_INFO, res);
1054         CDEBUG(D_OTHER, "About to insert this lock after %p:\n", original);
1055         ldlm_lock_dump(D_OTHER, new, 0);
1056
1057         if (new->l_destroyed) {
1058                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1059                 goto out;
1060         }
1061
1062         LASSERT(list_empty(&new->l_res_link));
1063
1064         list_add(&new->l_res_link, &original->l_res_link);
1065  out:;
1066 }
1067
1068 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1069 {
1070         int type = lock->l_resource->lr_type;
1071
1072         check_res_locked(lock->l_resource);
1073         if (type == LDLM_IBITS || type == LDLM_PLAIN)
1074                 ldlm_unlink_lock_skiplist(lock);
1075         else if (type == LDLM_EXTENT)
1076                 ldlm_extent_unlink_lock(lock);
1077         list_del_init(&lock->l_res_link);
1078 }
1079
1080 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1081 {
1082         desc->lr_type = res->lr_type;
1083         desc->lr_name = res->lr_name;
1084 }
1085
1086 void ldlm_dump_all_namespaces(ldlm_side_t client, int level)
1087 {
1088         struct list_head *tmp;
1089
1090         if (!((libcfs_debug | D_ERROR) & level))
1091                 return;
1092
1093         mutex_down(ldlm_namespace_lock(client));
1094
1095         list_for_each(tmp, ldlm_namespace_list(client)) {
1096                 struct ldlm_namespace *ns;
1097                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1098                 ldlm_namespace_dump(level, ns);
1099         }
1100
1101         mutex_up(ldlm_namespace_lock(client));
1102 }
1103
1104 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1105 {
1106         struct list_head *tmp;
1107
1108         if (!((libcfs_debug | D_ERROR) & level))
1109                 return;
1110
1111         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1112                ns->ns_name, ns->ns_refcount,
1113                ns_is_client(ns) ? "client" : "server");
1114
1115         if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1116                 return;
1117
1118         spin_lock(&ns->ns_hash_lock);
1119         tmp = ns->ns_root_list.next;
1120         while (tmp != &ns->ns_root_list) {
1121                 struct ldlm_resource *res;
1122                 res = list_entry(tmp, struct ldlm_resource, lr_childof);
1123
1124                 ldlm_resource_getref(res);
1125                 spin_unlock(&ns->ns_hash_lock);
1126                 LDLM_RESOURCE_ADDREF(res);
1127
1128                 lock_res(res);
1129                 ldlm_resource_dump(level, res);
1130                 unlock_res(res);
1131
1132                 LDLM_RESOURCE_DELREF(res);
1133                 spin_lock(&ns->ns_hash_lock);
1134                 tmp = tmp->next;
1135                 ldlm_resource_putref_locked(res);
1136         }
1137         ns->ns_next_dump = cfs_time_shift(10);
1138         spin_unlock(&ns->ns_hash_lock);
1139 }
1140
1141 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1142 {
1143         struct list_head *tmp;
1144         int pos;
1145
1146         CLASSERT(RES_NAME_SIZE == 4);
1147
1148         if (!((libcfs_debug | D_ERROR) & level))
1149                 return;
1150
1151         CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64
1152                ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1],
1153                res->lr_name.name[2], res->lr_name.name[3],
1154                atomic_read(&res->lr_refcount));
1155
1156         if (!list_empty(&res->lr_granted)) {
1157                 pos = 0;
1158                 CDEBUG(level, "Granted locks:\n");
1159                 list_for_each(tmp, &res->lr_granted) {
1160                         struct ldlm_lock *lock;
1161                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1162                         ldlm_lock_dump(level, lock, ++pos);
1163                 }
1164         }
1165         if (!list_empty(&res->lr_converting)) {
1166                 pos = 0;
1167                 CDEBUG(level, "Converting locks:\n");
1168                 list_for_each(tmp, &res->lr_converting) {
1169                         struct ldlm_lock *lock;
1170                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1171                         ldlm_lock_dump(level, lock, ++pos);
1172                 }
1173         }
1174         if (!list_empty(&res->lr_waiting)) {
1175                 pos = 0;
1176                 CDEBUG(level, "Waiting locks:\n");
1177                 list_for_each(tmp, &res->lr_waiting) {
1178                         struct ldlm_lock *lock;
1179                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1180                         ldlm_lock_dump(level, lock, ++pos);
1181                 }
1182         }
1183 }