Commit | Line | Data |
---|---|---|
d7e09d03 PT |
1 | /* |
2 | * GPL HEADER START | |
3 | * | |
4 | * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. | |
5 | * | |
6 | * This program is free software; you can redistribute it and/or modify | |
7 | * it under the terms of the GNU General Public License version 2 only, | |
8 | * as published by the Free Software Foundation. | |
9 | * | |
10 | * This program is distributed in the hope that it will be useful, but | |
11 | * WITHOUT ANY WARRANTY; without even the implied warranty of | |
12 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
13 | * General Public License version 2 for more details (a copy is included | |
14 | * in the LICENSE file that accompanied this code). | |
15 | * | |
16 | * You should have received a copy of the GNU General Public License | |
17 | * version 2 along with this program; If not, see | |
18 | * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf | |
19 | * | |
20 | * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, | |
21 | * CA 95054 USA or visit www.sun.com if you need additional information or | |
22 | * have any questions. | |
23 | * | |
24 | * GPL HEADER END | |
25 | */ | |
26 | /* | |
27 | * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved. | |
28 | * Use is subject to license terms. | |
29 | * | |
30 | * Copyright (c) 2010, 2012, Intel Corporation. | |
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 | # include <lustre_dlm.h> | |
44 | ||
45 | #include <lustre_fid.h> | |
46 | #include <obd_class.h> | |
47 | #include "ldlm_internal.h" | |
48 | ||
49 | struct kmem_cache *ldlm_resource_slab, *ldlm_lock_slab; | |
50 | ||
91a50030 OD |
51 | int ldlm_srv_namespace_nr = 0; |
52 | int ldlm_cli_namespace_nr = 0; | |
d7e09d03 PT |
53 | |
54 | struct mutex ldlm_srv_namespace_lock; | |
55 | LIST_HEAD(ldlm_srv_namespace_list); | |
56 | ||
57 | struct mutex ldlm_cli_namespace_lock; | |
91a50030 OD |
58 | /* Client Namespaces that have active resources in them. |
59 | * Once all resources go away, ldlm_poold moves such namespaces to the | |
60 | * inactive list */ | |
61 | LIST_HEAD(ldlm_cli_active_namespace_list); | |
62 | /* Client namespaces that don't have any locks in them */ | |
63 | LIST_HEAD(ldlm_cli_inactive_namespace_list); | |
d7e09d03 PT |
64 | |
65 | proc_dir_entry_t *ldlm_type_proc_dir = NULL; | |
66 | proc_dir_entry_t *ldlm_ns_proc_dir = NULL; | |
67 | proc_dir_entry_t *ldlm_svc_proc_dir = NULL; | |
68 | ||
69 | extern unsigned int ldlm_cancel_unused_locks_before_replay; | |
70 | ||
71 | /* during debug dump certain amount of granted locks for one resource to avoid | |
72 | * DDOS. */ | |
73 | unsigned int ldlm_dump_granted_max = 256; | |
74 | ||
75 | #ifdef LPROCFS | |
73bb1da6 PT |
76 | static ssize_t lprocfs_wr_dump_ns(struct file *file, const char *buffer, |
77 | size_t count, loff_t *off) | |
d7e09d03 PT |
78 | { |
79 | ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE); | |
80 | ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE); | |
0a3bdb00 | 81 | return count; |
d7e09d03 | 82 | } |
73bb1da6 PT |
83 | LPROC_SEQ_FOPS_WR_ONLY(ldlm, dump_ns); |
84 | ||
85 | LPROC_SEQ_FOPS_RW_TYPE(ldlm_rw, uint); | |
86 | LPROC_SEQ_FOPS_RO_TYPE(ldlm, uint); | |
d7e09d03 PT |
87 | |
88 | int ldlm_proc_setup(void) | |
89 | { | |
90 | int rc; | |
91 | struct lprocfs_vars list[] = { | |
73bb1da6 PT |
92 | { "dump_namespaces", &ldlm_dump_ns_fops, 0, 0222 }, |
93 | { "dump_granted_max", &ldlm_rw_uint_fops, | |
94 | &ldlm_dump_granted_max }, | |
95 | { "cancel_unused_locks_before_replay", &ldlm_rw_uint_fops, | |
96 | &ldlm_cancel_unused_locks_before_replay }, | |
d7e09d03 | 97 | { NULL }}; |
d7e09d03 PT |
98 | LASSERT(ldlm_ns_proc_dir == NULL); |
99 | ||
100 | ldlm_type_proc_dir = lprocfs_register(OBD_LDLM_DEVICENAME, | |
101 | proc_lustre_root, | |
102 | NULL, NULL); | |
103 | if (IS_ERR(ldlm_type_proc_dir)) { | |
104 | CERROR("LProcFS failed in ldlm-init\n"); | |
105 | rc = PTR_ERR(ldlm_type_proc_dir); | |
106 | GOTO(err, rc); | |
107 | } | |
108 | ||
109 | ldlm_ns_proc_dir = lprocfs_register("namespaces", | |
110 | ldlm_type_proc_dir, | |
111 | NULL, NULL); | |
112 | if (IS_ERR(ldlm_ns_proc_dir)) { | |
113 | CERROR("LProcFS failed in ldlm-init\n"); | |
114 | rc = PTR_ERR(ldlm_ns_proc_dir); | |
115 | GOTO(err_type, rc); | |
116 | } | |
117 | ||
118 | ldlm_svc_proc_dir = lprocfs_register("services", | |
119 | ldlm_type_proc_dir, | |
120 | NULL, NULL); | |
121 | if (IS_ERR(ldlm_svc_proc_dir)) { | |
122 | CERROR("LProcFS failed in ldlm-init\n"); | |
123 | rc = PTR_ERR(ldlm_svc_proc_dir); | |
124 | GOTO(err_ns, rc); | |
125 | } | |
126 | ||
127 | rc = lprocfs_add_vars(ldlm_type_proc_dir, list, NULL); | |
128 | ||
0a3bdb00 | 129 | return 0; |
d7e09d03 PT |
130 | |
131 | err_ns: | |
132 | lprocfs_remove(&ldlm_ns_proc_dir); | |
133 | err_type: | |
134 | lprocfs_remove(&ldlm_type_proc_dir); | |
135 | err: | |
136 | ldlm_svc_proc_dir = NULL; | |
1e4db2b3 PT |
137 | ldlm_type_proc_dir = NULL; |
138 | ldlm_ns_proc_dir = NULL; | |
0a3bdb00 | 139 | return rc; |
d7e09d03 PT |
140 | } |
141 | ||
142 | void ldlm_proc_cleanup(void) | |
143 | { | |
144 | if (ldlm_svc_proc_dir) | |
145 | lprocfs_remove(&ldlm_svc_proc_dir); | |
146 | ||
147 | if (ldlm_ns_proc_dir) | |
148 | lprocfs_remove(&ldlm_ns_proc_dir); | |
149 | ||
150 | if (ldlm_type_proc_dir) | |
151 | lprocfs_remove(&ldlm_type_proc_dir); | |
1e4db2b3 PT |
152 | |
153 | ldlm_svc_proc_dir = NULL; | |
154 | ldlm_type_proc_dir = NULL; | |
155 | ldlm_ns_proc_dir = NULL; | |
d7e09d03 PT |
156 | } |
157 | ||
73bb1da6 | 158 | static int lprocfs_ns_resources_seq_show(struct seq_file *m, void *v) |
d7e09d03 | 159 | { |
73bb1da6 | 160 | struct ldlm_namespace *ns = m->private; |
d7e09d03 PT |
161 | __u64 res = 0; |
162 | cfs_hash_bd_t bd; | |
163 | int i; | |
164 | ||
165 | /* result is not strictly consistant */ | |
166 | cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, i) | |
167 | res += cfs_hash_bd_count_get(&bd); | |
73bb1da6 | 168 | return lprocfs_rd_u64(m, &res); |
d7e09d03 | 169 | } |
73bb1da6 | 170 | LPROC_SEQ_FOPS_RO(lprocfs_ns_resources); |
d7e09d03 | 171 | |
73bb1da6 | 172 | static int lprocfs_ns_locks_seq_show(struct seq_file *m, void *v) |
d7e09d03 | 173 | { |
73bb1da6 | 174 | struct ldlm_namespace *ns = m->private; |
d7e09d03 PT |
175 | __u64 locks; |
176 | ||
177 | locks = lprocfs_stats_collector(ns->ns_stats, LDLM_NSS_LOCKS, | |
178 | LPROCFS_FIELDS_FLAGS_SUM); | |
73bb1da6 | 179 | return lprocfs_rd_u64(m, &locks); |
d7e09d03 | 180 | } |
73bb1da6 | 181 | LPROC_SEQ_FOPS_RO(lprocfs_ns_locks); |
d7e09d03 | 182 | |
73bb1da6 | 183 | static int lprocfs_lru_size_seq_show(struct seq_file *m, void *v) |
d7e09d03 | 184 | { |
73bb1da6 | 185 | struct ldlm_namespace *ns = m->private; |
d7e09d03 PT |
186 | __u32 *nr = &ns->ns_max_unused; |
187 | ||
188 | if (ns_connect_lru_resize(ns)) | |
189 | nr = &ns->ns_nr_unused; | |
73bb1da6 | 190 | return lprocfs_rd_uint(m, nr); |
d7e09d03 PT |
191 | } |
192 | ||
73bb1da6 PT |
193 | static ssize_t lprocfs_lru_size_seq_write(struct file *file, const char *buffer, |
194 | size_t count, loff_t *off) | |
d7e09d03 | 195 | { |
73bb1da6 | 196 | struct ldlm_namespace *ns = ((struct seq_file *)file->private_data)->private; |
d7e09d03 PT |
197 | char dummy[MAX_STRING_SIZE + 1], *end; |
198 | unsigned long tmp; | |
199 | int lru_resize; | |
200 | ||
201 | dummy[MAX_STRING_SIZE] = '\0'; | |
202 | if (copy_from_user(dummy, buffer, MAX_STRING_SIZE)) | |
203 | return -EFAULT; | |
204 | ||
205 | if (strncmp(dummy, "clear", 5) == 0) { | |
206 | CDEBUG(D_DLMTRACE, | |
207 | "dropping all unused locks from namespace %s\n", | |
208 | ldlm_ns_name(ns)); | |
209 | if (ns_connect_lru_resize(ns)) { | |
210 | int canceled, unused = ns->ns_nr_unused; | |
211 | ||
212 | /* Try to cancel all @ns_nr_unused locks. */ | |
213 | canceled = ldlm_cancel_lru(ns, unused, 0, | |
214 | LDLM_CANCEL_PASSED); | |
215 | if (canceled < unused) { | |
216 | CDEBUG(D_DLMTRACE, | |
217 | "not all requested locks are canceled, " | |
218 | "requested: %d, canceled: %d\n", unused, | |
219 | canceled); | |
220 | return -EINVAL; | |
221 | } | |
222 | } else { | |
223 | tmp = ns->ns_max_unused; | |
224 | ns->ns_max_unused = 0; | |
225 | ldlm_cancel_lru(ns, 0, 0, LDLM_CANCEL_PASSED); | |
226 | ns->ns_max_unused = tmp; | |
227 | } | |
228 | return count; | |
229 | } | |
230 | ||
231 | tmp = simple_strtoul(dummy, &end, 0); | |
232 | if (dummy == end) { | |
233 | CERROR("invalid value written\n"); | |
234 | return -EINVAL; | |
235 | } | |
236 | lru_resize = (tmp == 0); | |
237 | ||
238 | if (ns_connect_lru_resize(ns)) { | |
239 | if (!lru_resize) | |
240 | ns->ns_max_unused = (unsigned int)tmp; | |
241 | ||
242 | if (tmp > ns->ns_nr_unused) | |
243 | tmp = ns->ns_nr_unused; | |
244 | tmp = ns->ns_nr_unused - tmp; | |
245 | ||
246 | CDEBUG(D_DLMTRACE, | |
247 | "changing namespace %s unused locks from %u to %u\n", | |
248 | ldlm_ns_name(ns), ns->ns_nr_unused, | |
249 | (unsigned int)tmp); | |
250 | ldlm_cancel_lru(ns, tmp, LCF_ASYNC, LDLM_CANCEL_PASSED); | |
251 | ||
252 | if (!lru_resize) { | |
253 | CDEBUG(D_DLMTRACE, | |
254 | "disable lru_resize for namespace %s\n", | |
255 | ldlm_ns_name(ns)); | |
256 | ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE; | |
257 | } | |
258 | } else { | |
259 | CDEBUG(D_DLMTRACE, | |
260 | "changing namespace %s max_unused from %u to %u\n", | |
261 | ldlm_ns_name(ns), ns->ns_max_unused, | |
262 | (unsigned int)tmp); | |
263 | ns->ns_max_unused = (unsigned int)tmp; | |
264 | ldlm_cancel_lru(ns, 0, LCF_ASYNC, LDLM_CANCEL_PASSED); | |
265 | ||
266 | /* Make sure that LRU resize was originally supported before | |
267 | * turning it on here. */ | |
268 | if (lru_resize && | |
269 | (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) { | |
270 | CDEBUG(D_DLMTRACE, | |
271 | "enable lru_resize for namespace %s\n", | |
272 | ldlm_ns_name(ns)); | |
273 | ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE; | |
274 | } | |
275 | } | |
276 | ||
277 | return count; | |
278 | } | |
73bb1da6 | 279 | LPROC_SEQ_FOPS(lprocfs_lru_size); |
d7e09d03 | 280 | |
73bb1da6 | 281 | static int lprocfs_elc_seq_show(struct seq_file *m, void *v) |
d7e09d03 | 282 | { |
73bb1da6 | 283 | struct ldlm_namespace *ns = m->private; |
d7e09d03 PT |
284 | unsigned int supp = ns_connect_cancelset(ns); |
285 | ||
73bb1da6 | 286 | return lprocfs_rd_uint(m, &supp); |
d7e09d03 PT |
287 | } |
288 | ||
73bb1da6 PT |
289 | static ssize_t lprocfs_elc_seq_write(struct file *file, const char *buffer, |
290 | size_t count, loff_t *off) | |
d7e09d03 | 291 | { |
73bb1da6 | 292 | struct ldlm_namespace *ns = ((struct seq_file *)file->private_data)->private; |
d7e09d03 PT |
293 | unsigned int supp = -1; |
294 | int rc; | |
295 | ||
296 | rc = lprocfs_wr_uint(file, buffer, count, &supp); | |
297 | if (rc < 0) | |
298 | return rc; | |
299 | ||
300 | if (supp == 0) | |
301 | ns->ns_connect_flags &= ~OBD_CONNECT_CANCELSET; | |
302 | else if (ns->ns_orig_connect_flags & OBD_CONNECT_CANCELSET) | |
303 | ns->ns_connect_flags |= OBD_CONNECT_CANCELSET; | |
304 | return count; | |
305 | } | |
73bb1da6 | 306 | LPROC_SEQ_FOPS(lprocfs_elc); |
d7e09d03 PT |
307 | |
308 | void ldlm_namespace_proc_unregister(struct ldlm_namespace *ns) | |
309 | { | |
73bb1da6 | 310 | if (ns->ns_proc_dir_entry == NULL) |
d7e09d03 PT |
311 | CERROR("dlm namespace %s has no procfs dir?\n", |
312 | ldlm_ns_name(ns)); | |
73bb1da6 PT |
313 | else |
314 | lprocfs_remove(&ns->ns_proc_dir_entry); | |
d7e09d03 PT |
315 | |
316 | if (ns->ns_stats != NULL) | |
317 | lprocfs_free_stats(&ns->ns_stats); | |
318 | } | |
319 | ||
73bb1da6 PT |
320 | #define LDLM_NS_ADD_VAR(name, var, ops) \ |
321 | do { \ | |
322 | snprintf(lock_name, MAX_STRING_SIZE, name); \ | |
323 | lock_vars[0].data = var; \ | |
324 | lock_vars[0].fops = ops; \ | |
325 | lprocfs_add_vars(ns_pde, lock_vars, 0); \ | |
326 | } while (0) | |
327 | ||
d7e09d03 PT |
328 | int ldlm_namespace_proc_register(struct ldlm_namespace *ns) |
329 | { | |
330 | struct lprocfs_vars lock_vars[2]; | |
331 | char lock_name[MAX_STRING_SIZE + 1]; | |
73bb1da6 | 332 | proc_dir_entry_t *ns_pde; |
d7e09d03 PT |
333 | |
334 | LASSERT(ns != NULL); | |
335 | LASSERT(ns->ns_rs_hash != NULL); | |
336 | ||
73bb1da6 PT |
337 | if (ns->ns_proc_dir_entry != NULL) { |
338 | ns_pde = ns->ns_proc_dir_entry; | |
339 | } else { | |
340 | ns_pde = proc_mkdir(ldlm_ns_name(ns), ldlm_ns_proc_dir); | |
341 | if (ns_pde == NULL) | |
342 | return -ENOMEM; | |
343 | ns->ns_proc_dir_entry = ns_pde; | |
344 | } | |
345 | ||
d7e09d03 PT |
346 | ns->ns_stats = lprocfs_alloc_stats(LDLM_NSS_LAST, 0); |
347 | if (ns->ns_stats == NULL) | |
348 | return -ENOMEM; | |
349 | ||
350 | lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS, | |
351 | LPROCFS_CNTR_AVGMINMAX, "locks", "locks"); | |
352 | ||
353 | lock_name[MAX_STRING_SIZE] = '\0'; | |
354 | ||
355 | memset(lock_vars, 0, sizeof(lock_vars)); | |
356 | lock_vars[0].name = lock_name; | |
357 | ||
73bb1da6 PT |
358 | LDLM_NS_ADD_VAR("resource_count", ns, &lprocfs_ns_resources_fops); |
359 | LDLM_NS_ADD_VAR("lock_count", ns, &lprocfs_ns_locks_fops); | |
d7e09d03 PT |
360 | |
361 | if (ns_is_client(ns)) { | |
73bb1da6 PT |
362 | LDLM_NS_ADD_VAR("lock_unused_count", &ns->ns_nr_unused, |
363 | &ldlm_uint_fops); | |
364 | LDLM_NS_ADD_VAR("lru_size", ns, &lprocfs_lru_size_fops); | |
365 | LDLM_NS_ADD_VAR("lru_max_age", &ns->ns_max_age, | |
366 | &ldlm_rw_uint_fops); | |
367 | LDLM_NS_ADD_VAR("early_lock_cancel", ns, &lprocfs_elc_fops); | |
d7e09d03 | 368 | } else { |
73bb1da6 PT |
369 | LDLM_NS_ADD_VAR("ctime_age_limit", &ns->ns_ctime_age_limit, |
370 | &ldlm_rw_uint_fops); | |
371 | LDLM_NS_ADD_VAR("lock_timeouts", &ns->ns_timeouts, | |
372 | &ldlm_uint_fops); | |
373 | LDLM_NS_ADD_VAR("max_nolock_bytes", &ns->ns_max_nolock_size, | |
374 | &ldlm_rw_uint_fops); | |
375 | LDLM_NS_ADD_VAR("contention_seconds", &ns->ns_contention_time, | |
376 | &ldlm_rw_uint_fops); | |
377 | LDLM_NS_ADD_VAR("contended_locks", &ns->ns_contended_locks, | |
378 | &ldlm_rw_uint_fops); | |
379 | LDLM_NS_ADD_VAR("max_parallel_ast", &ns->ns_max_parallel_ast, | |
380 | &ldlm_rw_uint_fops); | |
d7e09d03 PT |
381 | } |
382 | return 0; | |
383 | } | |
384 | #undef MAX_STRING_SIZE | |
385 | #else /* LPROCFS */ | |
386 | ||
387 | #define ldlm_namespace_proc_unregister(ns) ({;}) | |
388 | #define ldlm_namespace_proc_register(ns) ({0;}) | |
389 | ||
390 | #endif /* LPROCFS */ | |
391 | ||
392 | static unsigned ldlm_res_hop_hash(cfs_hash_t *hs, | |
393 | const void *key, unsigned mask) | |
394 | { | |
395 | const struct ldlm_res_id *id = key; | |
396 | unsigned val = 0; | |
397 | unsigned i; | |
398 | ||
399 | for (i = 0; i < RES_NAME_SIZE; i++) | |
400 | val += id->name[i]; | |
401 | return val & mask; | |
402 | } | |
403 | ||
404 | static unsigned ldlm_res_hop_fid_hash(cfs_hash_t *hs, | |
405 | const void *key, unsigned mask) | |
406 | { | |
407 | const struct ldlm_res_id *id = key; | |
408 | struct lu_fid fid; | |
409 | __u32 hash; | |
410 | __u32 val; | |
411 | ||
412 | fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF]; | |
413 | fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_VER_OID_OFF]; | |
414 | fid.f_ver = (__u32)(id->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32); | |
415 | ||
416 | hash = fid_flatten32(&fid); | |
417 | hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */ | |
418 | if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) { | |
419 | val = id->name[LUSTRE_RES_ID_HSH_OFF]; | |
420 | hash += (val >> 5) + (val << 11); | |
421 | } else { | |
422 | val = fid_oid(&fid); | |
423 | } | |
424 | hash = cfs_hash_long(hash, hs->hs_bkt_bits); | |
425 | /* give me another random factor */ | |
426 | hash -= cfs_hash_long((unsigned long)hs, val % 11 + 3); | |
427 | ||
428 | hash <<= hs->hs_cur_bits - hs->hs_bkt_bits; | |
429 | hash |= ldlm_res_hop_hash(hs, key, CFS_HASH_NBKT(hs) - 1); | |
430 | ||
431 | return hash & mask; | |
432 | } | |
433 | ||
434 | static void *ldlm_res_hop_key(struct hlist_node *hnode) | |
435 | { | |
436 | struct ldlm_resource *res; | |
437 | ||
438 | res = hlist_entry(hnode, struct ldlm_resource, lr_hash); | |
439 | return &res->lr_name; | |
440 | } | |
441 | ||
442 | static int ldlm_res_hop_keycmp(const void *key, struct hlist_node *hnode) | |
443 | { | |
444 | struct ldlm_resource *res; | |
445 | ||
446 | res = hlist_entry(hnode, struct ldlm_resource, lr_hash); | |
447 | return ldlm_res_eq((const struct ldlm_res_id *)key, | |
448 | (const struct ldlm_res_id *)&res->lr_name); | |
449 | } | |
450 | ||
451 | static void *ldlm_res_hop_object(struct hlist_node *hnode) | |
452 | { | |
453 | return hlist_entry(hnode, struct ldlm_resource, lr_hash); | |
454 | } | |
455 | ||
456 | static void ldlm_res_hop_get_locked(cfs_hash_t *hs, struct hlist_node *hnode) | |
457 | { | |
458 | struct ldlm_resource *res; | |
459 | ||
460 | res = hlist_entry(hnode, struct ldlm_resource, lr_hash); | |
461 | ldlm_resource_getref(res); | |
462 | } | |
463 | ||
464 | static void ldlm_res_hop_put_locked(cfs_hash_t *hs, struct hlist_node *hnode) | |
465 | { | |
466 | struct ldlm_resource *res; | |
467 | ||
468 | res = hlist_entry(hnode, struct ldlm_resource, lr_hash); | |
469 | /* cfs_hash_for_each_nolock is the only chance we call it */ | |
470 | ldlm_resource_putref_locked(res); | |
471 | } | |
472 | ||
473 | static void ldlm_res_hop_put(cfs_hash_t *hs, struct hlist_node *hnode) | |
474 | { | |
475 | struct ldlm_resource *res; | |
476 | ||
477 | res = hlist_entry(hnode, struct ldlm_resource, lr_hash); | |
478 | ldlm_resource_putref(res); | |
479 | } | |
480 | ||
481 | cfs_hash_ops_t ldlm_ns_hash_ops = { | |
482 | .hs_hash = ldlm_res_hop_hash, | |
483 | .hs_key = ldlm_res_hop_key, | |
484 | .hs_keycmp = ldlm_res_hop_keycmp, | |
485 | .hs_keycpy = NULL, | |
486 | .hs_object = ldlm_res_hop_object, | |
487 | .hs_get = ldlm_res_hop_get_locked, | |
488 | .hs_put_locked = ldlm_res_hop_put_locked, | |
489 | .hs_put = ldlm_res_hop_put | |
490 | }; | |
491 | ||
492 | cfs_hash_ops_t ldlm_ns_fid_hash_ops = { | |
493 | .hs_hash = ldlm_res_hop_fid_hash, | |
494 | .hs_key = ldlm_res_hop_key, | |
495 | .hs_keycmp = ldlm_res_hop_keycmp, | |
496 | .hs_keycpy = NULL, | |
497 | .hs_object = ldlm_res_hop_object, | |
498 | .hs_get = ldlm_res_hop_get_locked, | |
499 | .hs_put_locked = ldlm_res_hop_put_locked, | |
500 | .hs_put = ldlm_res_hop_put | |
501 | }; | |
502 | ||
503 | typedef struct { | |
504 | ldlm_ns_type_t nsd_type; | |
505 | /** hash bucket bits */ | |
506 | unsigned nsd_bkt_bits; | |
507 | /** hash bits */ | |
508 | unsigned nsd_all_bits; | |
509 | /** hash operations */ | |
510 | cfs_hash_ops_t *nsd_hops; | |
511 | } ldlm_ns_hash_def_t; | |
512 | ||
513 | ldlm_ns_hash_def_t ldlm_ns_hash_defs[] = | |
514 | { | |
515 | { | |
516 | .nsd_type = LDLM_NS_TYPE_MDC, | |
517 | .nsd_bkt_bits = 11, | |
518 | .nsd_all_bits = 16, | |
519 | .nsd_hops = &ldlm_ns_fid_hash_ops, | |
520 | }, | |
521 | { | |
522 | .nsd_type = LDLM_NS_TYPE_MDT, | |
523 | .nsd_bkt_bits = 14, | |
524 | .nsd_all_bits = 21, | |
525 | .nsd_hops = &ldlm_ns_fid_hash_ops, | |
526 | }, | |
527 | { | |
528 | .nsd_type = LDLM_NS_TYPE_OSC, | |
529 | .nsd_bkt_bits = 8, | |
530 | .nsd_all_bits = 12, | |
531 | .nsd_hops = &ldlm_ns_hash_ops, | |
532 | }, | |
533 | { | |
534 | .nsd_type = LDLM_NS_TYPE_OST, | |
535 | .nsd_bkt_bits = 11, | |
536 | .nsd_all_bits = 17, | |
537 | .nsd_hops = &ldlm_ns_hash_ops, | |
538 | }, | |
539 | { | |
540 | .nsd_type = LDLM_NS_TYPE_MGC, | |
541 | .nsd_bkt_bits = 4, | |
542 | .nsd_all_bits = 4, | |
543 | .nsd_hops = &ldlm_ns_hash_ops, | |
544 | }, | |
545 | { | |
546 | .nsd_type = LDLM_NS_TYPE_MGT, | |
547 | .nsd_bkt_bits = 4, | |
548 | .nsd_all_bits = 4, | |
549 | .nsd_hops = &ldlm_ns_hash_ops, | |
550 | }, | |
551 | { | |
552 | .nsd_type = LDLM_NS_TYPE_UNKNOWN, | |
553 | }, | |
554 | }; | |
555 | ||
556 | /** | |
557 | * Create and initialize new empty namespace. | |
558 | */ | |
559 | struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name, | |
560 | ldlm_side_t client, | |
561 | ldlm_appetite_t apt, | |
562 | ldlm_ns_type_t ns_type) | |
563 | { | |
564 | struct ldlm_namespace *ns = NULL; | |
565 | struct ldlm_ns_bucket *nsb; | |
566 | ldlm_ns_hash_def_t *nsd; | |
567 | cfs_hash_bd_t bd; | |
568 | int idx; | |
569 | int rc; | |
d7e09d03 PT |
570 | |
571 | LASSERT(obd != NULL); | |
572 | ||
573 | rc = ldlm_get_ref(); | |
574 | if (rc) { | |
575 | CERROR("ldlm_get_ref failed: %d\n", rc); | |
0a3bdb00 | 576 | return NULL; |
d7e09d03 PT |
577 | } |
578 | ||
579 | for (idx = 0;;idx++) { | |
580 | nsd = &ldlm_ns_hash_defs[idx]; | |
581 | if (nsd->nsd_type == LDLM_NS_TYPE_UNKNOWN) { | |
582 | CERROR("Unknown type %d for ns %s\n", ns_type, name); | |
583 | GOTO(out_ref, NULL); | |
584 | } | |
585 | ||
586 | if (nsd->nsd_type == ns_type) | |
587 | break; | |
588 | } | |
589 | ||
590 | OBD_ALLOC_PTR(ns); | |
591 | if (!ns) | |
592 | GOTO(out_ref, NULL); | |
593 | ||
594 | ns->ns_rs_hash = cfs_hash_create(name, | |
595 | nsd->nsd_all_bits, nsd->nsd_all_bits, | |
596 | nsd->nsd_bkt_bits, sizeof(*nsb), | |
597 | CFS_HASH_MIN_THETA, | |
598 | CFS_HASH_MAX_THETA, | |
599 | nsd->nsd_hops, | |
600 | CFS_HASH_DEPTH | | |
601 | CFS_HASH_BIGNAME | | |
602 | CFS_HASH_SPIN_BKTLOCK | | |
603 | CFS_HASH_NO_ITEMREF); | |
604 | if (ns->ns_rs_hash == NULL) | |
605 | GOTO(out_ns, NULL); | |
606 | ||
607 | cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, idx) { | |
608 | nsb = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd); | |
609 | at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0); | |
610 | nsb->nsb_namespace = ns; | |
611 | } | |
612 | ||
613 | ns->ns_obd = obd; | |
614 | ns->ns_appetite = apt; | |
615 | ns->ns_client = client; | |
616 | ||
617 | INIT_LIST_HEAD(&ns->ns_list_chain); | |
618 | INIT_LIST_HEAD(&ns->ns_unused_list); | |
619 | spin_lock_init(&ns->ns_lock); | |
620 | atomic_set(&ns->ns_bref, 0); | |
621 | init_waitqueue_head(&ns->ns_waitq); | |
622 | ||
623 | ns->ns_max_nolock_size = NS_DEFAULT_MAX_NOLOCK_BYTES; | |
624 | ns->ns_contention_time = NS_DEFAULT_CONTENTION_SECONDS; | |
625 | ns->ns_contended_locks = NS_DEFAULT_CONTENDED_LOCKS; | |
626 | ||
627 | ns->ns_max_parallel_ast = LDLM_DEFAULT_PARALLEL_AST_LIMIT; | |
628 | ns->ns_nr_unused = 0; | |
629 | ns->ns_max_unused = LDLM_DEFAULT_LRU_SIZE; | |
630 | ns->ns_max_age = LDLM_DEFAULT_MAX_ALIVE; | |
631 | ns->ns_ctime_age_limit = LDLM_CTIME_AGE_LIMIT; | |
632 | ns->ns_timeouts = 0; | |
633 | ns->ns_orig_connect_flags = 0; | |
634 | ns->ns_connect_flags = 0; | |
635 | ns->ns_stopping = 0; | |
636 | rc = ldlm_namespace_proc_register(ns); | |
637 | if (rc != 0) { | |
638 | CERROR("Can't initialize ns proc, rc %d\n", rc); | |
639 | GOTO(out_hash, rc); | |
640 | } | |
641 | ||
91a50030 | 642 | idx = ldlm_namespace_nr_read(client); |
d7e09d03 PT |
643 | rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client); |
644 | if (rc) { | |
645 | CERROR("Can't initialize lock pool, rc %d\n", rc); | |
646 | GOTO(out_proc, rc); | |
647 | } | |
648 | ||
649 | ldlm_namespace_register(ns, client); | |
0a3bdb00 | 650 | return ns; |
d7e09d03 PT |
651 | out_proc: |
652 | ldlm_namespace_proc_unregister(ns); | |
653 | ldlm_namespace_cleanup(ns, 0); | |
654 | out_hash: | |
655 | cfs_hash_putref(ns->ns_rs_hash); | |
656 | out_ns: | |
657 | OBD_FREE_PTR(ns); | |
658 | out_ref: | |
659 | ldlm_put_ref(); | |
0a3bdb00 | 660 | return NULL; |
d7e09d03 PT |
661 | } |
662 | EXPORT_SYMBOL(ldlm_namespace_new); | |
663 | ||
664 | extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock); | |
665 | ||
666 | /** | |
667 | * Cancel and destroy all locks on a resource. | |
668 | * | |
669 | * If flags contains FL_LOCAL_ONLY, don't try to tell the server, just | |
670 | * clean up. This is currently only used for recovery, and we make | |
671 | * certain assumptions as a result--notably, that we shouldn't cancel | |
672 | * locks with refs. | |
673 | */ | |
674 | static void cleanup_resource(struct ldlm_resource *res, struct list_head *q, | |
675 | __u64 flags) | |
676 | { | |
677 | struct list_head *tmp; | |
678 | int rc = 0, client = ns_is_client(ldlm_res_to_ns(res)); | |
679 | bool local_only = !!(flags & LDLM_FL_LOCAL_ONLY); | |
680 | ||
681 | do { | |
682 | struct ldlm_lock *lock = NULL; | |
683 | ||
684 | /* First, we look for non-cleaned-yet lock | |
685 | * all cleaned locks are marked by CLEANED flag. */ | |
686 | lock_res(res); | |
687 | list_for_each(tmp, q) { | |
688 | lock = list_entry(tmp, struct ldlm_lock, | |
689 | l_res_link); | |
690 | if (lock->l_flags & LDLM_FL_CLEANED) { | |
691 | lock = NULL; | |
692 | continue; | |
693 | } | |
694 | LDLM_LOCK_GET(lock); | |
695 | lock->l_flags |= LDLM_FL_CLEANED; | |
696 | break; | |
697 | } | |
698 | ||
699 | if (lock == NULL) { | |
700 | unlock_res(res); | |
701 | break; | |
702 | } | |
703 | ||
704 | /* Set CBPENDING so nothing in the cancellation path | |
705 | * can match this lock. */ | |
706 | lock->l_flags |= LDLM_FL_CBPENDING; | |
707 | lock->l_flags |= LDLM_FL_FAILED; | |
708 | lock->l_flags |= flags; | |
709 | ||
710 | /* ... without sending a CANCEL message for local_only. */ | |
711 | if (local_only) | |
712 | lock->l_flags |= LDLM_FL_LOCAL_ONLY; | |
713 | ||
714 | if (local_only && (lock->l_readers || lock->l_writers)) { | |
715 | /* This is a little bit gross, but much better than the | |
716 | * alternative: pretend that we got a blocking AST from | |
717 | * the server, so that when the lock is decref'd, it | |
718 | * will go away ... */ | |
719 | unlock_res(res); | |
720 | LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY"); | |
721 | if (lock->l_completion_ast) | |
722 | lock->l_completion_ast(lock, 0, NULL); | |
723 | LDLM_LOCK_RELEASE(lock); | |
724 | continue; | |
725 | } | |
726 | ||
727 | if (client) { | |
728 | struct lustre_handle lockh; | |
729 | ||
730 | unlock_res(res); | |
731 | ldlm_lock2handle(lock, &lockh); | |
732 | rc = ldlm_cli_cancel(&lockh, LCF_ASYNC); | |
733 | if (rc) | |
734 | CERROR("ldlm_cli_cancel: %d\n", rc); | |
735 | } else { | |
736 | ldlm_resource_unlink_lock(lock); | |
737 | unlock_res(res); | |
738 | LDLM_DEBUG(lock, "Freeing a lock still held by a " | |
739 | "client node"); | |
740 | ldlm_lock_destroy(lock); | |
741 | } | |
742 | LDLM_LOCK_RELEASE(lock); | |
743 | } while (1); | |
744 | } | |
745 | ||
746 | static int ldlm_resource_clean(cfs_hash_t *hs, cfs_hash_bd_t *bd, | |
747 | struct hlist_node *hnode, void *arg) | |
748 | { | |
749 | struct ldlm_resource *res = cfs_hash_object(hs, hnode); | |
750 | __u64 flags = *(__u64 *)arg; | |
751 | ||
752 | cleanup_resource(res, &res->lr_granted, flags); | |
753 | cleanup_resource(res, &res->lr_converting, flags); | |
754 | cleanup_resource(res, &res->lr_waiting, flags); | |
755 | ||
756 | return 0; | |
757 | } | |
758 | ||
759 | static int ldlm_resource_complain(cfs_hash_t *hs, cfs_hash_bd_t *bd, | |
760 | struct hlist_node *hnode, void *arg) | |
761 | { | |
762 | struct ldlm_resource *res = cfs_hash_object(hs, hnode); | |
763 | ||
764 | lock_res(res); | |
765 | CERROR("Namespace %s resource refcount nonzero " | |
766 | "(%d) after lock cleanup; forcing " | |
767 | "cleanup.\n", | |
768 | ldlm_ns_name(ldlm_res_to_ns(res)), | |
769 | atomic_read(&res->lr_refcount) - 1); | |
770 | ||
771 | CERROR("Resource: %p ("LPU64"/"LPU64"/"LPU64"/" | |
772 | LPU64") (rc: %d)\n", res, | |
773 | res->lr_name.name[0], res->lr_name.name[1], | |
774 | res->lr_name.name[2], res->lr_name.name[3], | |
775 | atomic_read(&res->lr_refcount) - 1); | |
776 | ||
777 | ldlm_resource_dump(D_ERROR, res); | |
778 | unlock_res(res); | |
779 | return 0; | |
780 | } | |
781 | ||
782 | /** | |
783 | * Cancel and destroy all locks in the namespace. | |
784 | * | |
785 | * Typically used during evictions when server notified client that it was | |
786 | * evicted and all of its state needs to be destroyed. | |
787 | * Also used during shutdown. | |
788 | */ | |
789 | int ldlm_namespace_cleanup(struct ldlm_namespace *ns, __u64 flags) | |
790 | { | |
791 | if (ns == NULL) { | |
792 | CDEBUG(D_INFO, "NULL ns, skipping cleanup\n"); | |
793 | return ELDLM_OK; | |
794 | } | |
795 | ||
796 | cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean, &flags); | |
797 | cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain, NULL); | |
798 | return ELDLM_OK; | |
799 | } | |
800 | EXPORT_SYMBOL(ldlm_namespace_cleanup); | |
801 | ||
802 | /** | |
803 | * Attempts to free namespace. | |
804 | * | |
805 | * Only used when namespace goes away, like during an unmount. | |
806 | */ | |
807 | static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force) | |
808 | { | |
d7e09d03 PT |
809 | /* At shutdown time, don't call the cancellation callback */ |
810 | ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0); | |
811 | ||
812 | if (atomic_read(&ns->ns_bref) > 0) { | |
813 | struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL); | |
814 | int rc; | |
815 | CDEBUG(D_DLMTRACE, | |
816 | "dlm namespace %s free waiting on refcount %d\n", | |
817 | ldlm_ns_name(ns), atomic_read(&ns->ns_bref)); | |
818 | force_wait: | |
819 | if (force) | |
820 | lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL); | |
821 | ||
822 | rc = l_wait_event(ns->ns_waitq, | |
823 | atomic_read(&ns->ns_bref) == 0, &lwi); | |
824 | ||
825 | /* Forced cleanups should be able to reclaim all references, | |
826 | * so it's safe to wait forever... we can't leak locks... */ | |
827 | if (force && rc == -ETIMEDOUT) { | |
828 | LCONSOLE_ERROR("Forced cleanup waiting for %s " | |
829 | "namespace with %d resources in use, " | |
830 | "(rc=%d)\n", ldlm_ns_name(ns), | |
831 | atomic_read(&ns->ns_bref), rc); | |
832 | GOTO(force_wait, rc); | |
833 | } | |
834 | ||
835 | if (atomic_read(&ns->ns_bref)) { | |
836 | LCONSOLE_ERROR("Cleanup waiting for %s namespace " | |
837 | "with %d resources in use, (rc=%d)\n", | |
838 | ldlm_ns_name(ns), | |
839 | atomic_read(&ns->ns_bref), rc); | |
0a3bdb00 | 840 | return ELDLM_NAMESPACE_EXISTS; |
d7e09d03 PT |
841 | } |
842 | CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n", | |
843 | ldlm_ns_name(ns)); | |
844 | } | |
845 | ||
0a3bdb00 | 846 | return ELDLM_OK; |
d7e09d03 PT |
847 | } |
848 | ||
849 | /** | |
850 | * Performs various cleanups for passed \a ns to make it drop refc and be | |
851 | * ready for freeing. Waits for refc == 0. | |
852 | * | |
853 | * The following is done: | |
854 | * (0) Unregister \a ns from its list to make inaccessible for potential | |
855 | * users like pools thread and others; | |
856 | * (1) Clear all locks in \a ns. | |
857 | */ | |
858 | void ldlm_namespace_free_prior(struct ldlm_namespace *ns, | |
859 | struct obd_import *imp, | |
860 | int force) | |
861 | { | |
862 | int rc; | |
29aaf496 | 863 | |
d7e09d03 | 864 | if (!ns) { |
d7e09d03 PT |
865 | return; |
866 | } | |
867 | ||
868 | spin_lock(&ns->ns_lock); | |
869 | ns->ns_stopping = 1; | |
870 | spin_unlock(&ns->ns_lock); | |
871 | ||
872 | /* | |
873 | * Can fail with -EINTR when force == 0 in which case try harder. | |
874 | */ | |
875 | rc = __ldlm_namespace_free(ns, force); | |
876 | if (rc != ELDLM_OK) { | |
877 | if (imp) { | |
878 | ptlrpc_disconnect_import(imp, 0); | |
879 | ptlrpc_invalidate_import(imp); | |
880 | } | |
881 | ||
882 | /* | |
883 | * With all requests dropped and the import inactive | |
884 | * we are gaurenteed all reference will be dropped. | |
885 | */ | |
886 | rc = __ldlm_namespace_free(ns, 1); | |
887 | LASSERT(rc == 0); | |
888 | } | |
d7e09d03 PT |
889 | } |
890 | ||
891 | /** | |
892 | * Performs freeing memory structures related to \a ns. This is only done | |
893 | * when ldlm_namespce_free_prior() successfully removed all resources | |
894 | * referencing \a ns and its refc == 0. | |
895 | */ | |
896 | void ldlm_namespace_free_post(struct ldlm_namespace *ns) | |
897 | { | |
d7e09d03 | 898 | if (!ns) { |
d7e09d03 PT |
899 | return; |
900 | } | |
901 | ||
902 | /* Make sure that nobody can find this ns in its list. */ | |
903 | ldlm_namespace_unregister(ns, ns->ns_client); | |
904 | /* Fini pool _before_ parent proc dir is removed. This is important as | |
905 | * ldlm_pool_fini() removes own proc dir which is child to @dir. | |
906 | * Removing it after @dir may cause oops. */ | |
907 | ldlm_pool_fini(&ns->ns_pool); | |
908 | ||
909 | ldlm_namespace_proc_unregister(ns); | |
910 | cfs_hash_putref(ns->ns_rs_hash); | |
911 | /* Namespace \a ns should be not on list at this time, otherwise | |
912 | * this will cause issues related to using freed \a ns in poold | |
913 | * thread. */ | |
914 | LASSERT(list_empty(&ns->ns_list_chain)); | |
915 | OBD_FREE_PTR(ns); | |
916 | ldlm_put_ref(); | |
d7e09d03 PT |
917 | } |
918 | ||
919 | /** | |
920 | * Cleanup the resource, and free namespace. | |
921 | * bug 12864: | |
922 | * Deadlock issue: | |
923 | * proc1: destroy import | |
924 | * class_disconnect_export(grab cl_sem) -> | |
925 | * -> ldlm_namespace_free -> | |
926 | * -> lprocfs_remove(grab _lprocfs_lock). | |
927 | * proc2: read proc info | |
928 | * lprocfs_fops_read(grab _lprocfs_lock) -> | |
929 | * -> osc_rd_active, etc(grab cl_sem). | |
930 | * | |
931 | * So that I have to split the ldlm_namespace_free into two parts - the first | |
932 | * part ldlm_namespace_free_prior is used to cleanup the resource which is | |
933 | * being used; the 2nd part ldlm_namespace_free_post is used to unregister the | |
934 | * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem | |
935 | * held. | |
936 | */ | |
937 | void ldlm_namespace_free(struct ldlm_namespace *ns, | |
938 | struct obd_import *imp, | |
939 | int force) | |
940 | { | |
941 | ldlm_namespace_free_prior(ns, imp, force); | |
942 | ldlm_namespace_free_post(ns); | |
943 | } | |
944 | EXPORT_SYMBOL(ldlm_namespace_free); | |
945 | ||
946 | void ldlm_namespace_get(struct ldlm_namespace *ns) | |
947 | { | |
948 | atomic_inc(&ns->ns_bref); | |
949 | } | |
950 | EXPORT_SYMBOL(ldlm_namespace_get); | |
951 | ||
91a50030 OD |
952 | /* This is only for callers that care about refcount */ |
953 | int ldlm_namespace_get_return(struct ldlm_namespace *ns) | |
954 | { | |
955 | return atomic_inc_return(&ns->ns_bref); | |
956 | } | |
957 | ||
d7e09d03 PT |
958 | void ldlm_namespace_put(struct ldlm_namespace *ns) |
959 | { | |
960 | if (atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) { | |
961 | wake_up(&ns->ns_waitq); | |
962 | spin_unlock(&ns->ns_lock); | |
963 | } | |
964 | } | |
965 | EXPORT_SYMBOL(ldlm_namespace_put); | |
966 | ||
967 | /** Register \a ns in the list of namespaces */ | |
968 | void ldlm_namespace_register(struct ldlm_namespace *ns, ldlm_side_t client) | |
969 | { | |
970 | mutex_lock(ldlm_namespace_lock(client)); | |
971 | LASSERT(list_empty(&ns->ns_list_chain)); | |
91a50030 OD |
972 | list_add(&ns->ns_list_chain, ldlm_namespace_inactive_list(client)); |
973 | ldlm_namespace_nr_inc(client); | |
d7e09d03 PT |
974 | mutex_unlock(ldlm_namespace_lock(client)); |
975 | } | |
976 | ||
977 | /** Unregister \a ns from the list of namespaces. */ | |
978 | void ldlm_namespace_unregister(struct ldlm_namespace *ns, ldlm_side_t client) | |
979 | { | |
980 | mutex_lock(ldlm_namespace_lock(client)); | |
981 | LASSERT(!list_empty(&ns->ns_list_chain)); | |
982 | /* Some asserts and possibly other parts of the code are still | |
983 | * using list_empty(&ns->ns_list_chain). This is why it is | |
984 | * important to use list_del_init() here. */ | |
985 | list_del_init(&ns->ns_list_chain); | |
91a50030 | 986 | ldlm_namespace_nr_dec(client); |
d7e09d03 PT |
987 | mutex_unlock(ldlm_namespace_lock(client)); |
988 | } | |
989 | ||
990 | /** Should be called with ldlm_namespace_lock(client) taken. */ | |
91a50030 OD |
991 | void ldlm_namespace_move_to_active_locked(struct ldlm_namespace *ns, |
992 | ldlm_side_t client) | |
d7e09d03 PT |
993 | { |
994 | LASSERT(!list_empty(&ns->ns_list_chain)); | |
995 | LASSERT(mutex_is_locked(ldlm_namespace_lock(client))); | |
996 | list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client)); | |
997 | } | |
998 | ||
91a50030 OD |
999 | /** Should be called with ldlm_namespace_lock(client) taken. */ |
1000 | void ldlm_namespace_move_to_inactive_locked(struct ldlm_namespace *ns, | |
1001 | ldlm_side_t client) | |
1002 | { | |
1003 | LASSERT(!list_empty(&ns->ns_list_chain)); | |
1004 | LASSERT(mutex_is_locked(ldlm_namespace_lock(client))); | |
1005 | list_move_tail(&ns->ns_list_chain, | |
1006 | ldlm_namespace_inactive_list(client)); | |
1007 | } | |
1008 | ||
d7e09d03 PT |
1009 | /** Should be called with ldlm_namespace_lock(client) taken. */ |
1010 | struct ldlm_namespace *ldlm_namespace_first_locked(ldlm_side_t client) | |
1011 | { | |
1012 | LASSERT(mutex_is_locked(ldlm_namespace_lock(client))); | |
1013 | LASSERT(!list_empty(ldlm_namespace_list(client))); | |
1014 | return container_of(ldlm_namespace_list(client)->next, | |
1015 | struct ldlm_namespace, ns_list_chain); | |
1016 | } | |
1017 | ||
1018 | /** Create and initialize new resource. */ | |
1019 | static struct ldlm_resource *ldlm_resource_new(void) | |
1020 | { | |
1021 | struct ldlm_resource *res; | |
1022 | int idx; | |
1023 | ||
1024 | OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, __GFP_IO); | |
1025 | if (res == NULL) | |
1026 | return NULL; | |
1027 | ||
1028 | INIT_LIST_HEAD(&res->lr_granted); | |
1029 | INIT_LIST_HEAD(&res->lr_converting); | |
1030 | INIT_LIST_HEAD(&res->lr_waiting); | |
1031 | ||
1032 | /* Initialize interval trees for each lock mode. */ | |
1033 | for (idx = 0; idx < LCK_MODE_NUM; idx++) { | |
1034 | res->lr_itree[idx].lit_size = 0; | |
1035 | res->lr_itree[idx].lit_mode = 1 << idx; | |
1036 | res->lr_itree[idx].lit_root = NULL; | |
1037 | } | |
1038 | ||
1039 | atomic_set(&res->lr_refcount, 1); | |
1040 | spin_lock_init(&res->lr_lock); | |
1041 | lu_ref_init(&res->lr_reference); | |
1042 | ||
1043 | /* The creator of the resource must unlock the mutex after LVB | |
1044 | * initialization. */ | |
1045 | mutex_init(&res->lr_lvb_mutex); | |
1046 | mutex_lock(&res->lr_lvb_mutex); | |
1047 | ||
1048 | return res; | |
1049 | } | |
1050 | ||
1051 | /** | |
1052 | * Return a reference to resource with given name, creating it if necessary. | |
1053 | * Args: namespace with ns_lock unlocked | |
1054 | * Locks: takes and releases NS hash-lock and res->lr_lock | |
1055 | * Returns: referenced, unlocked ldlm_resource or NULL | |
1056 | */ | |
1057 | struct ldlm_resource * | |
1058 | ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent, | |
1059 | const struct ldlm_res_id *name, ldlm_type_t type, int create) | |
1060 | { | |
1061 | struct hlist_node *hnode; | |
1062 | struct ldlm_resource *res; | |
1063 | cfs_hash_bd_t bd; | |
1064 | __u64 version; | |
91a50030 | 1065 | int ns_refcount = 0; |
d7e09d03 PT |
1066 | |
1067 | LASSERT(ns != NULL); | |
1068 | LASSERT(parent == NULL); | |
1069 | LASSERT(ns->ns_rs_hash != NULL); | |
1070 | LASSERT(name->name[0] != 0); | |
1071 | ||
1072 | cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0); | |
1073 | hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name); | |
1074 | if (hnode != NULL) { | |
1075 | cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0); | |
1076 | res = hlist_entry(hnode, struct ldlm_resource, lr_hash); | |
1077 | /* Synchronize with regard to resource creation. */ | |
1078 | if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) { | |
1079 | mutex_lock(&res->lr_lvb_mutex); | |
1080 | mutex_unlock(&res->lr_lvb_mutex); | |
1081 | } | |
1082 | ||
1083 | if (unlikely(res->lr_lvb_len < 0)) { | |
1084 | ldlm_resource_putref(res); | |
1085 | res = NULL; | |
1086 | } | |
1087 | return res; | |
1088 | } | |
1089 | ||
1090 | version = cfs_hash_bd_version_get(&bd); | |
1091 | cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0); | |
1092 | ||
1093 | if (create == 0) | |
1094 | return NULL; | |
1095 | ||
1096 | LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE, | |
1097 | "type: %d\n", type); | |
1098 | res = ldlm_resource_new(); | |
1099 | if (!res) | |
1100 | return NULL; | |
1101 | ||
1102 | res->lr_ns_bucket = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd); | |
1103 | res->lr_name = *name; | |
1104 | res->lr_type = type; | |
1105 | res->lr_most_restr = LCK_NL; | |
1106 | ||
1107 | cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1); | |
1108 | hnode = (version == cfs_hash_bd_version_get(&bd)) ? NULL : | |
1109 | cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name); | |
1110 | ||
1111 | if (hnode != NULL) { | |
1112 | /* Someone won the race and already added the resource. */ | |
1113 | cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1); | |
1114 | /* Clean lu_ref for failed resource. */ | |
1115 | lu_ref_fini(&res->lr_reference); | |
1116 | /* We have taken lr_lvb_mutex. Drop it. */ | |
1117 | mutex_unlock(&res->lr_lvb_mutex); | |
1118 | OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res); | |
1119 | ||
1120 | res = hlist_entry(hnode, struct ldlm_resource, lr_hash); | |
1121 | /* Synchronize with regard to resource creation. */ | |
1122 | if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) { | |
1123 | mutex_lock(&res->lr_lvb_mutex); | |
1124 | mutex_unlock(&res->lr_lvb_mutex); | |
1125 | } | |
1126 | ||
1127 | if (unlikely(res->lr_lvb_len < 0)) { | |
1128 | ldlm_resource_putref(res); | |
1129 | res = NULL; | |
1130 | } | |
1131 | return res; | |
1132 | } | |
1133 | /* We won! Let's add the resource. */ | |
1134 | cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash); | |
1135 | if (cfs_hash_bd_count_get(&bd) == 1) | |
91a50030 | 1136 | ns_refcount = ldlm_namespace_get_return(ns); |
d7e09d03 PT |
1137 | |
1138 | cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1); | |
1139 | if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) { | |
1140 | int rc; | |
1141 | ||
1142 | OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2); | |
1143 | rc = ns->ns_lvbo->lvbo_init(res); | |
1144 | if (rc < 0) { | |
c5b60ba7 AD |
1145 | CERROR("%s: lvbo_init failed for resource "LPX64":" |
1146 | LPX64": rc = %d\n", ns->ns_obd->obd_name, | |
1147 | name->name[0], name->name[1], rc); | |
d7e09d03 PT |
1148 | if (res->lr_lvb_data) { |
1149 | OBD_FREE(res->lr_lvb_data, res->lr_lvb_len); | |
1150 | res->lr_lvb_data = NULL; | |
1151 | } | |
1152 | res->lr_lvb_len = rc; | |
1153 | mutex_unlock(&res->lr_lvb_mutex); | |
1154 | ldlm_resource_putref(res); | |
1155 | return NULL; | |
1156 | } | |
1157 | } | |
1158 | ||
1159 | /* We create resource with locked lr_lvb_mutex. */ | |
1160 | mutex_unlock(&res->lr_lvb_mutex); | |
1161 | ||
91a50030 OD |
1162 | /* Let's see if we happened to be the very first resource in this |
1163 | * namespace. If so, and this is a client namespace, we need to move | |
1164 | * the namespace into the active namespaces list to be patrolled by | |
df9fcbeb | 1165 | * the ldlm_poold. */ |
91a50030 OD |
1166 | if (ns_is_client(ns) && ns_refcount == 1) { |
1167 | mutex_lock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT)); | |
1168 | ldlm_namespace_move_to_active_locked(ns, LDLM_NAMESPACE_CLIENT); | |
1169 | mutex_unlock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT)); | |
1170 | } | |
1171 | ||
d7e09d03 PT |
1172 | return res; |
1173 | } | |
1174 | EXPORT_SYMBOL(ldlm_resource_get); | |
1175 | ||
1176 | struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res) | |
1177 | { | |
1178 | LASSERT(res != NULL); | |
1179 | LASSERT(res != LP_POISON); | |
1180 | atomic_inc(&res->lr_refcount); | |
1181 | CDEBUG(D_INFO, "getref res: %p count: %d\n", res, | |
1182 | atomic_read(&res->lr_refcount)); | |
1183 | return res; | |
1184 | } | |
1185 | ||
1186 | static void __ldlm_resource_putref_final(cfs_hash_bd_t *bd, | |
1187 | struct ldlm_resource *res) | |
1188 | { | |
1189 | struct ldlm_ns_bucket *nsb = res->lr_ns_bucket; | |
1190 | ||
1191 | if (!list_empty(&res->lr_granted)) { | |
1192 | ldlm_resource_dump(D_ERROR, res); | |
1193 | LBUG(); | |
1194 | } | |
1195 | ||
1196 | if (!list_empty(&res->lr_converting)) { | |
1197 | ldlm_resource_dump(D_ERROR, res); | |
1198 | LBUG(); | |
1199 | } | |
1200 | ||
1201 | if (!list_empty(&res->lr_waiting)) { | |
1202 | ldlm_resource_dump(D_ERROR, res); | |
1203 | LBUG(); | |
1204 | } | |
1205 | ||
1206 | cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash, | |
1207 | bd, &res->lr_hash); | |
1208 | lu_ref_fini(&res->lr_reference); | |
1209 | if (cfs_hash_bd_count_get(bd) == 0) | |
1210 | ldlm_namespace_put(nsb->nsb_namespace); | |
1211 | } | |
1212 | ||
1213 | /* Returns 1 if the resource was freed, 0 if it remains. */ | |
1214 | int ldlm_resource_putref(struct ldlm_resource *res) | |
1215 | { | |
1216 | struct ldlm_namespace *ns = ldlm_res_to_ns(res); | |
1217 | cfs_hash_bd_t bd; | |
1218 | ||
1219 | LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON); | |
1220 | CDEBUG(D_INFO, "putref res: %p count: %d\n", | |
1221 | res, atomic_read(&res->lr_refcount) - 1); | |
1222 | ||
1223 | cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd); | |
1224 | if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) { | |
1225 | __ldlm_resource_putref_final(&bd, res); | |
1226 | cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1); | |
1227 | if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free) | |
1228 | ns->ns_lvbo->lvbo_free(res); | |
1229 | OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res); | |
1230 | return 1; | |
1231 | } | |
1232 | return 0; | |
1233 | } | |
1234 | EXPORT_SYMBOL(ldlm_resource_putref); | |
1235 | ||
1236 | /* Returns 1 if the resource was freed, 0 if it remains. */ | |
1237 | int ldlm_resource_putref_locked(struct ldlm_resource *res) | |
1238 | { | |
1239 | struct ldlm_namespace *ns = ldlm_res_to_ns(res); | |
1240 | ||
1241 | LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON); | |
1242 | CDEBUG(D_INFO, "putref res: %p count: %d\n", | |
1243 | res, atomic_read(&res->lr_refcount) - 1); | |
1244 | ||
1245 | if (atomic_dec_and_test(&res->lr_refcount)) { | |
1246 | cfs_hash_bd_t bd; | |
1247 | ||
1248 | cfs_hash_bd_get(ldlm_res_to_ns(res)->ns_rs_hash, | |
1249 | &res->lr_name, &bd); | |
1250 | __ldlm_resource_putref_final(&bd, res); | |
1251 | cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1); | |
1252 | /* NB: ns_rs_hash is created with CFS_HASH_NO_ITEMREF, | |
1253 | * so we should never be here while calling cfs_hash_del, | |
1254 | * cfs_hash_for_each_nolock is the only case we can get | |
1255 | * here, which is safe to release cfs_hash_bd_lock. | |
1256 | */ | |
1257 | if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free) | |
1258 | ns->ns_lvbo->lvbo_free(res); | |
1259 | OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res); | |
1260 | ||
1261 | cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1); | |
1262 | return 1; | |
1263 | } | |
1264 | return 0; | |
1265 | } | |
1266 | ||
1267 | /** | |
1268 | * Add a lock into a given resource into specified lock list. | |
1269 | */ | |
1270 | void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head, | |
1271 | struct ldlm_lock *lock) | |
1272 | { | |
1273 | check_res_locked(res); | |
1274 | ||
1275 | LDLM_DEBUG(lock, "About to add this lock:\n"); | |
1276 | ||
f2145eae | 1277 | if (lock->l_flags & LDLM_FL_DESTROYED) { |
d7e09d03 PT |
1278 | CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n"); |
1279 | return; | |
1280 | } | |
1281 | ||
1282 | LASSERT(list_empty(&lock->l_res_link)); | |
1283 | ||
1284 | list_add_tail(&lock->l_res_link, head); | |
1285 | } | |
1286 | ||
1287 | /** | |
1288 | * Insert a lock into resource after specified lock. | |
1289 | * | |
1290 | * Obtain resource description from the lock we are inserting after. | |
1291 | */ | |
1292 | void ldlm_resource_insert_lock_after(struct ldlm_lock *original, | |
1293 | struct ldlm_lock *new) | |
1294 | { | |
1295 | struct ldlm_resource *res = original->l_resource; | |
1296 | ||
1297 | check_res_locked(res); | |
1298 | ||
1299 | ldlm_resource_dump(D_INFO, res); | |
1300 | LDLM_DEBUG(new, "About to insert this lock after %p:\n", original); | |
1301 | ||
f2145eae | 1302 | if (new->l_flags & LDLM_FL_DESTROYED) { |
d7e09d03 PT |
1303 | CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n"); |
1304 | goto out; | |
1305 | } | |
1306 | ||
1307 | LASSERT(list_empty(&new->l_res_link)); | |
1308 | ||
1309 | list_add(&new->l_res_link, &original->l_res_link); | |
1310 | out:; | |
1311 | } | |
1312 | ||
1313 | void ldlm_resource_unlink_lock(struct ldlm_lock *lock) | |
1314 | { | |
1315 | int type = lock->l_resource->lr_type; | |
1316 | ||
1317 | check_res_locked(lock->l_resource); | |
1318 | if (type == LDLM_IBITS || type == LDLM_PLAIN) | |
1319 | ldlm_unlink_lock_skiplist(lock); | |
1320 | else if (type == LDLM_EXTENT) | |
1321 | ldlm_extent_unlink_lock(lock); | |
1322 | list_del_init(&lock->l_res_link); | |
1323 | } | |
1324 | EXPORT_SYMBOL(ldlm_resource_unlink_lock); | |
1325 | ||
1326 | void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc) | |
1327 | { | |
1328 | desc->lr_type = res->lr_type; | |
1329 | desc->lr_name = res->lr_name; | |
1330 | } | |
1331 | ||
1332 | /** | |
1333 | * Print information about all locks in all namespaces on this node to debug | |
1334 | * log. | |
1335 | */ | |
1336 | void ldlm_dump_all_namespaces(ldlm_side_t client, int level) | |
1337 | { | |
1338 | struct list_head *tmp; | |
1339 | ||
1340 | if (!((libcfs_debug | D_ERROR) & level)) | |
1341 | return; | |
1342 | ||
1343 | mutex_lock(ldlm_namespace_lock(client)); | |
1344 | ||
1345 | list_for_each(tmp, ldlm_namespace_list(client)) { | |
1346 | struct ldlm_namespace *ns; | |
1347 | ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain); | |
1348 | ldlm_namespace_dump(level, ns); | |
1349 | } | |
1350 | ||
1351 | mutex_unlock(ldlm_namespace_lock(client)); | |
1352 | } | |
1353 | EXPORT_SYMBOL(ldlm_dump_all_namespaces); | |
1354 | ||
1355 | static int ldlm_res_hash_dump(cfs_hash_t *hs, cfs_hash_bd_t *bd, | |
1356 | struct hlist_node *hnode, void *arg) | |
1357 | { | |
1358 | struct ldlm_resource *res = cfs_hash_object(hs, hnode); | |
1359 | int level = (int)(unsigned long)arg; | |
1360 | ||
1361 | lock_res(res); | |
1362 | ldlm_resource_dump(level, res); | |
1363 | unlock_res(res); | |
1364 | ||
1365 | return 0; | |
1366 | } | |
1367 | ||
1368 | /** | |
1369 | * Print information about all locks in this namespace on this node to debug | |
1370 | * log. | |
1371 | */ | |
1372 | void ldlm_namespace_dump(int level, struct ldlm_namespace *ns) | |
1373 | { | |
1374 | if (!((libcfs_debug | D_ERROR) & level)) | |
1375 | return; | |
1376 | ||
1377 | CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n", | |
1378 | ldlm_ns_name(ns), atomic_read(&ns->ns_bref), | |
1379 | ns_is_client(ns) ? "client" : "server"); | |
1380 | ||
1381 | if (cfs_time_before(cfs_time_current(), ns->ns_next_dump)) | |
1382 | return; | |
1383 | ||
1384 | cfs_hash_for_each_nolock(ns->ns_rs_hash, | |
1385 | ldlm_res_hash_dump, | |
1386 | (void *)(unsigned long)level); | |
1387 | spin_lock(&ns->ns_lock); | |
1388 | ns->ns_next_dump = cfs_time_shift(10); | |
1389 | spin_unlock(&ns->ns_lock); | |
1390 | } | |
1391 | EXPORT_SYMBOL(ldlm_namespace_dump); | |
1392 | ||
1393 | /** | |
1394 | * Print information about all locks in this resource to debug log. | |
1395 | */ | |
1396 | void ldlm_resource_dump(int level, struct ldlm_resource *res) | |
1397 | { | |
1398 | struct ldlm_lock *lock; | |
1399 | unsigned int granted = 0; | |
1400 | ||
1401 | CLASSERT(RES_NAME_SIZE == 4); | |
1402 | ||
1403 | if (!((libcfs_debug | D_ERROR) & level)) | |
1404 | return; | |
1405 | ||
1406 | CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64 | |
1407 | ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1], | |
1408 | res->lr_name.name[2], res->lr_name.name[3], | |
1409 | atomic_read(&res->lr_refcount)); | |
1410 | ||
1411 | if (!list_empty(&res->lr_granted)) { | |
1412 | CDEBUG(level, "Granted locks (in reverse order):\n"); | |
1413 | list_for_each_entry_reverse(lock, &res->lr_granted, | |
1414 | l_res_link) { | |
1415 | LDLM_DEBUG_LIMIT(level, lock, "###"); | |
1416 | if (!(level & D_CANTMASK) && | |
1417 | ++granted > ldlm_dump_granted_max) { | |
1418 | CDEBUG(level, "only dump %d granted locks to " | |
1419 | "avoid DDOS.\n", granted); | |
1420 | break; | |
1421 | } | |
1422 | } | |
1423 | } | |
1424 | if (!list_empty(&res->lr_converting)) { | |
1425 | CDEBUG(level, "Converting locks:\n"); | |
1426 | list_for_each_entry(lock, &res->lr_converting, l_res_link) | |
1427 | LDLM_DEBUG_LIMIT(level, lock, "###"); | |
1428 | } | |
1429 | if (!list_empty(&res->lr_waiting)) { | |
1430 | CDEBUG(level, "Waiting locks:\n"); | |
1431 | list_for_each_entry(lock, &res->lr_waiting, l_res_link) | |
1432 | LDLM_DEBUG_LIMIT(level, lock, "###"); | |
1433 | } | |
1434 | } |