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); | |
81 | RETURN(count); | |
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 PT |
97 | { NULL }}; |
98 | ENTRY; | |
99 | LASSERT(ldlm_ns_proc_dir == NULL); | |
100 | ||
101 | ldlm_type_proc_dir = lprocfs_register(OBD_LDLM_DEVICENAME, | |
102 | proc_lustre_root, | |
103 | NULL, NULL); | |
104 | if (IS_ERR(ldlm_type_proc_dir)) { | |
105 | CERROR("LProcFS failed in ldlm-init\n"); | |
106 | rc = PTR_ERR(ldlm_type_proc_dir); | |
107 | GOTO(err, rc); | |
108 | } | |
109 | ||
110 | ldlm_ns_proc_dir = lprocfs_register("namespaces", | |
111 | ldlm_type_proc_dir, | |
112 | NULL, NULL); | |
113 | if (IS_ERR(ldlm_ns_proc_dir)) { | |
114 | CERROR("LProcFS failed in ldlm-init\n"); | |
115 | rc = PTR_ERR(ldlm_ns_proc_dir); | |
116 | GOTO(err_type, rc); | |
117 | } | |
118 | ||
119 | ldlm_svc_proc_dir = lprocfs_register("services", | |
120 | ldlm_type_proc_dir, | |
121 | NULL, NULL); | |
122 | if (IS_ERR(ldlm_svc_proc_dir)) { | |
123 | CERROR("LProcFS failed in ldlm-init\n"); | |
124 | rc = PTR_ERR(ldlm_svc_proc_dir); | |
125 | GOTO(err_ns, rc); | |
126 | } | |
127 | ||
128 | rc = lprocfs_add_vars(ldlm_type_proc_dir, list, NULL); | |
129 | ||
130 | RETURN(0); | |
131 | ||
132 | err_ns: | |
133 | lprocfs_remove(&ldlm_ns_proc_dir); | |
134 | err_type: | |
135 | lprocfs_remove(&ldlm_type_proc_dir); | |
136 | err: | |
137 | ldlm_svc_proc_dir = NULL; | |
1e4db2b3 PT |
138 | ldlm_type_proc_dir = NULL; |
139 | ldlm_ns_proc_dir = NULL; | |
d7e09d03 PT |
140 | RETURN(rc); |
141 | } | |
142 | ||
143 | void ldlm_proc_cleanup(void) | |
144 | { | |
145 | if (ldlm_svc_proc_dir) | |
146 | lprocfs_remove(&ldlm_svc_proc_dir); | |
147 | ||
148 | if (ldlm_ns_proc_dir) | |
149 | lprocfs_remove(&ldlm_ns_proc_dir); | |
150 | ||
151 | if (ldlm_type_proc_dir) | |
152 | lprocfs_remove(&ldlm_type_proc_dir); | |
1e4db2b3 PT |
153 | |
154 | ldlm_svc_proc_dir = NULL; | |
155 | ldlm_type_proc_dir = NULL; | |
156 | ldlm_ns_proc_dir = NULL; | |
d7e09d03 PT |
157 | } |
158 | ||
73bb1da6 | 159 | static int lprocfs_ns_resources_seq_show(struct seq_file *m, void *v) |
d7e09d03 | 160 | { |
73bb1da6 | 161 | struct ldlm_namespace *ns = m->private; |
d7e09d03 PT |
162 | __u64 res = 0; |
163 | cfs_hash_bd_t bd; | |
164 | int i; | |
165 | ||
166 | /* result is not strictly consistant */ | |
167 | cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, i) | |
168 | res += cfs_hash_bd_count_get(&bd); | |
73bb1da6 | 169 | return lprocfs_rd_u64(m, &res); |
d7e09d03 | 170 | } |
73bb1da6 | 171 | LPROC_SEQ_FOPS_RO(lprocfs_ns_resources); |
d7e09d03 | 172 | |
73bb1da6 | 173 | static int lprocfs_ns_locks_seq_show(struct seq_file *m, void *v) |
d7e09d03 | 174 | { |
73bb1da6 | 175 | struct ldlm_namespace *ns = m->private; |
d7e09d03 PT |
176 | __u64 locks; |
177 | ||
178 | locks = lprocfs_stats_collector(ns->ns_stats, LDLM_NSS_LOCKS, | |
179 | LPROCFS_FIELDS_FLAGS_SUM); | |
73bb1da6 | 180 | return lprocfs_rd_u64(m, &locks); |
d7e09d03 | 181 | } |
73bb1da6 | 182 | LPROC_SEQ_FOPS_RO(lprocfs_ns_locks); |
d7e09d03 | 183 | |
73bb1da6 | 184 | static int lprocfs_lru_size_seq_show(struct seq_file *m, void *v) |
d7e09d03 | 185 | { |
73bb1da6 | 186 | struct ldlm_namespace *ns = m->private; |
d7e09d03 PT |
187 | __u32 *nr = &ns->ns_max_unused; |
188 | ||
189 | if (ns_connect_lru_resize(ns)) | |
190 | nr = &ns->ns_nr_unused; | |
73bb1da6 | 191 | return lprocfs_rd_uint(m, nr); |
d7e09d03 PT |
192 | } |
193 | ||
73bb1da6 PT |
194 | static ssize_t lprocfs_lru_size_seq_write(struct file *file, const char *buffer, |
195 | size_t count, loff_t *off) | |
d7e09d03 | 196 | { |
73bb1da6 | 197 | struct ldlm_namespace *ns = ((struct seq_file *)file->private_data)->private; |
d7e09d03 PT |
198 | char dummy[MAX_STRING_SIZE + 1], *end; |
199 | unsigned long tmp; | |
200 | int lru_resize; | |
201 | ||
202 | dummy[MAX_STRING_SIZE] = '\0'; | |
203 | if (copy_from_user(dummy, buffer, MAX_STRING_SIZE)) | |
204 | return -EFAULT; | |
205 | ||
206 | if (strncmp(dummy, "clear", 5) == 0) { | |
207 | CDEBUG(D_DLMTRACE, | |
208 | "dropping all unused locks from namespace %s\n", | |
209 | ldlm_ns_name(ns)); | |
210 | if (ns_connect_lru_resize(ns)) { | |
211 | int canceled, unused = ns->ns_nr_unused; | |
212 | ||
213 | /* Try to cancel all @ns_nr_unused locks. */ | |
214 | canceled = ldlm_cancel_lru(ns, unused, 0, | |
215 | LDLM_CANCEL_PASSED); | |
216 | if (canceled < unused) { | |
217 | CDEBUG(D_DLMTRACE, | |
218 | "not all requested locks are canceled, " | |
219 | "requested: %d, canceled: %d\n", unused, | |
220 | canceled); | |
221 | return -EINVAL; | |
222 | } | |
223 | } else { | |
224 | tmp = ns->ns_max_unused; | |
225 | ns->ns_max_unused = 0; | |
226 | ldlm_cancel_lru(ns, 0, 0, LDLM_CANCEL_PASSED); | |
227 | ns->ns_max_unused = tmp; | |
228 | } | |
229 | return count; | |
230 | } | |
231 | ||
232 | tmp = simple_strtoul(dummy, &end, 0); | |
233 | if (dummy == end) { | |
234 | CERROR("invalid value written\n"); | |
235 | return -EINVAL; | |
236 | } | |
237 | lru_resize = (tmp == 0); | |
238 | ||
239 | if (ns_connect_lru_resize(ns)) { | |
240 | if (!lru_resize) | |
241 | ns->ns_max_unused = (unsigned int)tmp; | |
242 | ||
243 | if (tmp > ns->ns_nr_unused) | |
244 | tmp = ns->ns_nr_unused; | |
245 | tmp = ns->ns_nr_unused - tmp; | |
246 | ||
247 | CDEBUG(D_DLMTRACE, | |
248 | "changing namespace %s unused locks from %u to %u\n", | |
249 | ldlm_ns_name(ns), ns->ns_nr_unused, | |
250 | (unsigned int)tmp); | |
251 | ldlm_cancel_lru(ns, tmp, LCF_ASYNC, LDLM_CANCEL_PASSED); | |
252 | ||
253 | if (!lru_resize) { | |
254 | CDEBUG(D_DLMTRACE, | |
255 | "disable lru_resize for namespace %s\n", | |
256 | ldlm_ns_name(ns)); | |
257 | ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE; | |
258 | } | |
259 | } else { | |
260 | CDEBUG(D_DLMTRACE, | |
261 | "changing namespace %s max_unused from %u to %u\n", | |
262 | ldlm_ns_name(ns), ns->ns_max_unused, | |
263 | (unsigned int)tmp); | |
264 | ns->ns_max_unused = (unsigned int)tmp; | |
265 | ldlm_cancel_lru(ns, 0, LCF_ASYNC, LDLM_CANCEL_PASSED); | |
266 | ||
267 | /* Make sure that LRU resize was originally supported before | |
268 | * turning it on here. */ | |
269 | if (lru_resize && | |
270 | (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) { | |
271 | CDEBUG(D_DLMTRACE, | |
272 | "enable lru_resize for namespace %s\n", | |
273 | ldlm_ns_name(ns)); | |
274 | ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE; | |
275 | } | |
276 | } | |
277 | ||
278 | return count; | |
279 | } | |
73bb1da6 | 280 | LPROC_SEQ_FOPS(lprocfs_lru_size); |
d7e09d03 | 281 | |
73bb1da6 | 282 | static int lprocfs_elc_seq_show(struct seq_file *m, void *v) |
d7e09d03 | 283 | { |
73bb1da6 | 284 | struct ldlm_namespace *ns = m->private; |
d7e09d03 PT |
285 | unsigned int supp = ns_connect_cancelset(ns); |
286 | ||
73bb1da6 | 287 | return lprocfs_rd_uint(m, &supp); |
d7e09d03 PT |
288 | } |
289 | ||
73bb1da6 PT |
290 | static ssize_t lprocfs_elc_seq_write(struct file *file, const char *buffer, |
291 | size_t count, loff_t *off) | |
d7e09d03 | 292 | { |
73bb1da6 | 293 | struct ldlm_namespace *ns = ((struct seq_file *)file->private_data)->private; |
d7e09d03 PT |
294 | unsigned int supp = -1; |
295 | int rc; | |
296 | ||
297 | rc = lprocfs_wr_uint(file, buffer, count, &supp); | |
298 | if (rc < 0) | |
299 | return rc; | |
300 | ||
301 | if (supp == 0) | |
302 | ns->ns_connect_flags &= ~OBD_CONNECT_CANCELSET; | |
303 | else if (ns->ns_orig_connect_flags & OBD_CONNECT_CANCELSET) | |
304 | ns->ns_connect_flags |= OBD_CONNECT_CANCELSET; | |
305 | return count; | |
306 | } | |
73bb1da6 | 307 | LPROC_SEQ_FOPS(lprocfs_elc); |
d7e09d03 PT |
308 | |
309 | void ldlm_namespace_proc_unregister(struct ldlm_namespace *ns) | |
310 | { | |
73bb1da6 | 311 | if (ns->ns_proc_dir_entry == NULL) |
d7e09d03 PT |
312 | CERROR("dlm namespace %s has no procfs dir?\n", |
313 | ldlm_ns_name(ns)); | |
73bb1da6 PT |
314 | else |
315 | lprocfs_remove(&ns->ns_proc_dir_entry); | |
d7e09d03 PT |
316 | |
317 | if (ns->ns_stats != NULL) | |
318 | lprocfs_free_stats(&ns->ns_stats); | |
319 | } | |
320 | ||
73bb1da6 PT |
321 | #define LDLM_NS_ADD_VAR(name, var, ops) \ |
322 | do { \ | |
323 | snprintf(lock_name, MAX_STRING_SIZE, name); \ | |
324 | lock_vars[0].data = var; \ | |
325 | lock_vars[0].fops = ops; \ | |
326 | lprocfs_add_vars(ns_pde, lock_vars, 0); \ | |
327 | } while (0) | |
328 | ||
d7e09d03 PT |
329 | int ldlm_namespace_proc_register(struct ldlm_namespace *ns) |
330 | { | |
331 | struct lprocfs_vars lock_vars[2]; | |
332 | char lock_name[MAX_STRING_SIZE + 1]; | |
73bb1da6 | 333 | proc_dir_entry_t *ns_pde; |
d7e09d03 PT |
334 | |
335 | LASSERT(ns != NULL); | |
336 | LASSERT(ns->ns_rs_hash != NULL); | |
337 | ||
73bb1da6 PT |
338 | if (ns->ns_proc_dir_entry != NULL) { |
339 | ns_pde = ns->ns_proc_dir_entry; | |
340 | } else { | |
341 | ns_pde = proc_mkdir(ldlm_ns_name(ns), ldlm_ns_proc_dir); | |
342 | if (ns_pde == NULL) | |
343 | return -ENOMEM; | |
344 | ns->ns_proc_dir_entry = ns_pde; | |
345 | } | |
346 | ||
d7e09d03 PT |
347 | ns->ns_stats = lprocfs_alloc_stats(LDLM_NSS_LAST, 0); |
348 | if (ns->ns_stats == NULL) | |
349 | return -ENOMEM; | |
350 | ||
351 | lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS, | |
352 | LPROCFS_CNTR_AVGMINMAX, "locks", "locks"); | |
353 | ||
354 | lock_name[MAX_STRING_SIZE] = '\0'; | |
355 | ||
356 | memset(lock_vars, 0, sizeof(lock_vars)); | |
357 | lock_vars[0].name = lock_name; | |
358 | ||
73bb1da6 PT |
359 | LDLM_NS_ADD_VAR("resource_count", ns, &lprocfs_ns_resources_fops); |
360 | LDLM_NS_ADD_VAR("lock_count", ns, &lprocfs_ns_locks_fops); | |
d7e09d03 PT |
361 | |
362 | if (ns_is_client(ns)) { | |
73bb1da6 PT |
363 | LDLM_NS_ADD_VAR("lock_unused_count", &ns->ns_nr_unused, |
364 | &ldlm_uint_fops); | |
365 | LDLM_NS_ADD_VAR("lru_size", ns, &lprocfs_lru_size_fops); | |
366 | LDLM_NS_ADD_VAR("lru_max_age", &ns->ns_max_age, | |
367 | &ldlm_rw_uint_fops); | |
368 | LDLM_NS_ADD_VAR("early_lock_cancel", ns, &lprocfs_elc_fops); | |
d7e09d03 | 369 | } else { |
73bb1da6 PT |
370 | LDLM_NS_ADD_VAR("ctime_age_limit", &ns->ns_ctime_age_limit, |
371 | &ldlm_rw_uint_fops); | |
372 | LDLM_NS_ADD_VAR("lock_timeouts", &ns->ns_timeouts, | |
373 | &ldlm_uint_fops); | |
374 | LDLM_NS_ADD_VAR("max_nolock_bytes", &ns->ns_max_nolock_size, | |
375 | &ldlm_rw_uint_fops); | |
376 | LDLM_NS_ADD_VAR("contention_seconds", &ns->ns_contention_time, | |
377 | &ldlm_rw_uint_fops); | |
378 | LDLM_NS_ADD_VAR("contended_locks", &ns->ns_contended_locks, | |
379 | &ldlm_rw_uint_fops); | |
380 | LDLM_NS_ADD_VAR("max_parallel_ast", &ns->ns_max_parallel_ast, | |
381 | &ldlm_rw_uint_fops); | |
d7e09d03 PT |
382 | } |
383 | return 0; | |
384 | } | |
385 | #undef MAX_STRING_SIZE | |
386 | #else /* LPROCFS */ | |
387 | ||
388 | #define ldlm_namespace_proc_unregister(ns) ({;}) | |
389 | #define ldlm_namespace_proc_register(ns) ({0;}) | |
390 | ||
391 | #endif /* LPROCFS */ | |
392 | ||
393 | static unsigned ldlm_res_hop_hash(cfs_hash_t *hs, | |
394 | const void *key, unsigned mask) | |
395 | { | |
396 | const struct ldlm_res_id *id = key; | |
397 | unsigned val = 0; | |
398 | unsigned i; | |
399 | ||
400 | for (i = 0; i < RES_NAME_SIZE; i++) | |
401 | val += id->name[i]; | |
402 | return val & mask; | |
403 | } | |
404 | ||
405 | static unsigned ldlm_res_hop_fid_hash(cfs_hash_t *hs, | |
406 | const void *key, unsigned mask) | |
407 | { | |
408 | const struct ldlm_res_id *id = key; | |
409 | struct lu_fid fid; | |
410 | __u32 hash; | |
411 | __u32 val; | |
412 | ||
413 | fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF]; | |
414 | fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_VER_OID_OFF]; | |
415 | fid.f_ver = (__u32)(id->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32); | |
416 | ||
417 | hash = fid_flatten32(&fid); | |
418 | hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */ | |
419 | if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) { | |
420 | val = id->name[LUSTRE_RES_ID_HSH_OFF]; | |
421 | hash += (val >> 5) + (val << 11); | |
422 | } else { | |
423 | val = fid_oid(&fid); | |
424 | } | |
425 | hash = cfs_hash_long(hash, hs->hs_bkt_bits); | |
426 | /* give me another random factor */ | |
427 | hash -= cfs_hash_long((unsigned long)hs, val % 11 + 3); | |
428 | ||
429 | hash <<= hs->hs_cur_bits - hs->hs_bkt_bits; | |
430 | hash |= ldlm_res_hop_hash(hs, key, CFS_HASH_NBKT(hs) - 1); | |
431 | ||
432 | return hash & mask; | |
433 | } | |
434 | ||
435 | static void *ldlm_res_hop_key(struct hlist_node *hnode) | |
436 | { | |
437 | struct ldlm_resource *res; | |
438 | ||
439 | res = hlist_entry(hnode, struct ldlm_resource, lr_hash); | |
440 | return &res->lr_name; | |
441 | } | |
442 | ||
443 | static int ldlm_res_hop_keycmp(const void *key, struct hlist_node *hnode) | |
444 | { | |
445 | struct ldlm_resource *res; | |
446 | ||
447 | res = hlist_entry(hnode, struct ldlm_resource, lr_hash); | |
448 | return ldlm_res_eq((const struct ldlm_res_id *)key, | |
449 | (const struct ldlm_res_id *)&res->lr_name); | |
450 | } | |
451 | ||
452 | static void *ldlm_res_hop_object(struct hlist_node *hnode) | |
453 | { | |
454 | return hlist_entry(hnode, struct ldlm_resource, lr_hash); | |
455 | } | |
456 | ||
457 | static void ldlm_res_hop_get_locked(cfs_hash_t *hs, struct hlist_node *hnode) | |
458 | { | |
459 | struct ldlm_resource *res; | |
460 | ||
461 | res = hlist_entry(hnode, struct ldlm_resource, lr_hash); | |
462 | ldlm_resource_getref(res); | |
463 | } | |
464 | ||
465 | static void ldlm_res_hop_put_locked(cfs_hash_t *hs, struct hlist_node *hnode) | |
466 | { | |
467 | struct ldlm_resource *res; | |
468 | ||
469 | res = hlist_entry(hnode, struct ldlm_resource, lr_hash); | |
470 | /* cfs_hash_for_each_nolock is the only chance we call it */ | |
471 | ldlm_resource_putref_locked(res); | |
472 | } | |
473 | ||
474 | static void ldlm_res_hop_put(cfs_hash_t *hs, struct hlist_node *hnode) | |
475 | { | |
476 | struct ldlm_resource *res; | |
477 | ||
478 | res = hlist_entry(hnode, struct ldlm_resource, lr_hash); | |
479 | ldlm_resource_putref(res); | |
480 | } | |
481 | ||
482 | cfs_hash_ops_t ldlm_ns_hash_ops = { | |
483 | .hs_hash = ldlm_res_hop_hash, | |
484 | .hs_key = ldlm_res_hop_key, | |
485 | .hs_keycmp = ldlm_res_hop_keycmp, | |
486 | .hs_keycpy = NULL, | |
487 | .hs_object = ldlm_res_hop_object, | |
488 | .hs_get = ldlm_res_hop_get_locked, | |
489 | .hs_put_locked = ldlm_res_hop_put_locked, | |
490 | .hs_put = ldlm_res_hop_put | |
491 | }; | |
492 | ||
493 | cfs_hash_ops_t ldlm_ns_fid_hash_ops = { | |
494 | .hs_hash = ldlm_res_hop_fid_hash, | |
495 | .hs_key = ldlm_res_hop_key, | |
496 | .hs_keycmp = ldlm_res_hop_keycmp, | |
497 | .hs_keycpy = NULL, | |
498 | .hs_object = ldlm_res_hop_object, | |
499 | .hs_get = ldlm_res_hop_get_locked, | |
500 | .hs_put_locked = ldlm_res_hop_put_locked, | |
501 | .hs_put = ldlm_res_hop_put | |
502 | }; | |
503 | ||
504 | typedef struct { | |
505 | ldlm_ns_type_t nsd_type; | |
506 | /** hash bucket bits */ | |
507 | unsigned nsd_bkt_bits; | |
508 | /** hash bits */ | |
509 | unsigned nsd_all_bits; | |
510 | /** hash operations */ | |
511 | cfs_hash_ops_t *nsd_hops; | |
512 | } ldlm_ns_hash_def_t; | |
513 | ||
514 | ldlm_ns_hash_def_t ldlm_ns_hash_defs[] = | |
515 | { | |
516 | { | |
517 | .nsd_type = LDLM_NS_TYPE_MDC, | |
518 | .nsd_bkt_bits = 11, | |
519 | .nsd_all_bits = 16, | |
520 | .nsd_hops = &ldlm_ns_fid_hash_ops, | |
521 | }, | |
522 | { | |
523 | .nsd_type = LDLM_NS_TYPE_MDT, | |
524 | .nsd_bkt_bits = 14, | |
525 | .nsd_all_bits = 21, | |
526 | .nsd_hops = &ldlm_ns_fid_hash_ops, | |
527 | }, | |
528 | { | |
529 | .nsd_type = LDLM_NS_TYPE_OSC, | |
530 | .nsd_bkt_bits = 8, | |
531 | .nsd_all_bits = 12, | |
532 | .nsd_hops = &ldlm_ns_hash_ops, | |
533 | }, | |
534 | { | |
535 | .nsd_type = LDLM_NS_TYPE_OST, | |
536 | .nsd_bkt_bits = 11, | |
537 | .nsd_all_bits = 17, | |
538 | .nsd_hops = &ldlm_ns_hash_ops, | |
539 | }, | |
540 | { | |
541 | .nsd_type = LDLM_NS_TYPE_MGC, | |
542 | .nsd_bkt_bits = 4, | |
543 | .nsd_all_bits = 4, | |
544 | .nsd_hops = &ldlm_ns_hash_ops, | |
545 | }, | |
546 | { | |
547 | .nsd_type = LDLM_NS_TYPE_MGT, | |
548 | .nsd_bkt_bits = 4, | |
549 | .nsd_all_bits = 4, | |
550 | .nsd_hops = &ldlm_ns_hash_ops, | |
551 | }, | |
552 | { | |
553 | .nsd_type = LDLM_NS_TYPE_UNKNOWN, | |
554 | }, | |
555 | }; | |
556 | ||
557 | /** | |
558 | * Create and initialize new empty namespace. | |
559 | */ | |
560 | struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name, | |
561 | ldlm_side_t client, | |
562 | ldlm_appetite_t apt, | |
563 | ldlm_ns_type_t ns_type) | |
564 | { | |
565 | struct ldlm_namespace *ns = NULL; | |
566 | struct ldlm_ns_bucket *nsb; | |
567 | ldlm_ns_hash_def_t *nsd; | |
568 | cfs_hash_bd_t bd; | |
569 | int idx; | |
570 | int rc; | |
571 | ENTRY; | |
572 | ||
573 | LASSERT(obd != NULL); | |
574 | ||
575 | rc = ldlm_get_ref(); | |
576 | if (rc) { | |
577 | CERROR("ldlm_get_ref failed: %d\n", rc); | |
578 | RETURN(NULL); | |
579 | } | |
580 | ||
581 | for (idx = 0;;idx++) { | |
582 | nsd = &ldlm_ns_hash_defs[idx]; | |
583 | if (nsd->nsd_type == LDLM_NS_TYPE_UNKNOWN) { | |
584 | CERROR("Unknown type %d for ns %s\n", ns_type, name); | |
585 | GOTO(out_ref, NULL); | |
586 | } | |
587 | ||
588 | if (nsd->nsd_type == ns_type) | |
589 | break; | |
590 | } | |
591 | ||
592 | OBD_ALLOC_PTR(ns); | |
593 | if (!ns) | |
594 | GOTO(out_ref, NULL); | |
595 | ||
596 | ns->ns_rs_hash = cfs_hash_create(name, | |
597 | nsd->nsd_all_bits, nsd->nsd_all_bits, | |
598 | nsd->nsd_bkt_bits, sizeof(*nsb), | |
599 | CFS_HASH_MIN_THETA, | |
600 | CFS_HASH_MAX_THETA, | |
601 | nsd->nsd_hops, | |
602 | CFS_HASH_DEPTH | | |
603 | CFS_HASH_BIGNAME | | |
604 | CFS_HASH_SPIN_BKTLOCK | | |
605 | CFS_HASH_NO_ITEMREF); | |
606 | if (ns->ns_rs_hash == NULL) | |
607 | GOTO(out_ns, NULL); | |
608 | ||
609 | cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, idx) { | |
610 | nsb = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd); | |
611 | at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0); | |
612 | nsb->nsb_namespace = ns; | |
613 | } | |
614 | ||
615 | ns->ns_obd = obd; | |
616 | ns->ns_appetite = apt; | |
617 | ns->ns_client = client; | |
618 | ||
619 | INIT_LIST_HEAD(&ns->ns_list_chain); | |
620 | INIT_LIST_HEAD(&ns->ns_unused_list); | |
621 | spin_lock_init(&ns->ns_lock); | |
622 | atomic_set(&ns->ns_bref, 0); | |
623 | init_waitqueue_head(&ns->ns_waitq); | |
624 | ||
625 | ns->ns_max_nolock_size = NS_DEFAULT_MAX_NOLOCK_BYTES; | |
626 | ns->ns_contention_time = NS_DEFAULT_CONTENTION_SECONDS; | |
627 | ns->ns_contended_locks = NS_DEFAULT_CONTENDED_LOCKS; | |
628 | ||
629 | ns->ns_max_parallel_ast = LDLM_DEFAULT_PARALLEL_AST_LIMIT; | |
630 | ns->ns_nr_unused = 0; | |
631 | ns->ns_max_unused = LDLM_DEFAULT_LRU_SIZE; | |
632 | ns->ns_max_age = LDLM_DEFAULT_MAX_ALIVE; | |
633 | ns->ns_ctime_age_limit = LDLM_CTIME_AGE_LIMIT; | |
634 | ns->ns_timeouts = 0; | |
635 | ns->ns_orig_connect_flags = 0; | |
636 | ns->ns_connect_flags = 0; | |
637 | ns->ns_stopping = 0; | |
638 | rc = ldlm_namespace_proc_register(ns); | |
639 | if (rc != 0) { | |
640 | CERROR("Can't initialize ns proc, rc %d\n", rc); | |
641 | GOTO(out_hash, rc); | |
642 | } | |
643 | ||
91a50030 | 644 | idx = ldlm_namespace_nr_read(client); |
d7e09d03 PT |
645 | rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client); |
646 | if (rc) { | |
647 | CERROR("Can't initialize lock pool, rc %d\n", rc); | |
648 | GOTO(out_proc, rc); | |
649 | } | |
650 | ||
651 | ldlm_namespace_register(ns, client); | |
652 | RETURN(ns); | |
653 | out_proc: | |
654 | ldlm_namespace_proc_unregister(ns); | |
655 | ldlm_namespace_cleanup(ns, 0); | |
656 | out_hash: | |
657 | cfs_hash_putref(ns->ns_rs_hash); | |
658 | out_ns: | |
659 | OBD_FREE_PTR(ns); | |
660 | out_ref: | |
661 | ldlm_put_ref(); | |
662 | RETURN(NULL); | |
663 | } | |
664 | EXPORT_SYMBOL(ldlm_namespace_new); | |
665 | ||
666 | extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock); | |
667 | ||
668 | /** | |
669 | * Cancel and destroy all locks on a resource. | |
670 | * | |
671 | * If flags contains FL_LOCAL_ONLY, don't try to tell the server, just | |
672 | * clean up. This is currently only used for recovery, and we make | |
673 | * certain assumptions as a result--notably, that we shouldn't cancel | |
674 | * locks with refs. | |
675 | */ | |
676 | static void cleanup_resource(struct ldlm_resource *res, struct list_head *q, | |
677 | __u64 flags) | |
678 | { | |
679 | struct list_head *tmp; | |
680 | int rc = 0, client = ns_is_client(ldlm_res_to_ns(res)); | |
681 | bool local_only = !!(flags & LDLM_FL_LOCAL_ONLY); | |
682 | ||
683 | do { | |
684 | struct ldlm_lock *lock = NULL; | |
685 | ||
686 | /* First, we look for non-cleaned-yet lock | |
687 | * all cleaned locks are marked by CLEANED flag. */ | |
688 | lock_res(res); | |
689 | list_for_each(tmp, q) { | |
690 | lock = list_entry(tmp, struct ldlm_lock, | |
691 | l_res_link); | |
692 | if (lock->l_flags & LDLM_FL_CLEANED) { | |
693 | lock = NULL; | |
694 | continue; | |
695 | } | |
696 | LDLM_LOCK_GET(lock); | |
697 | lock->l_flags |= LDLM_FL_CLEANED; | |
698 | break; | |
699 | } | |
700 | ||
701 | if (lock == NULL) { | |
702 | unlock_res(res); | |
703 | break; | |
704 | } | |
705 | ||
706 | /* Set CBPENDING so nothing in the cancellation path | |
707 | * can match this lock. */ | |
708 | lock->l_flags |= LDLM_FL_CBPENDING; | |
709 | lock->l_flags |= LDLM_FL_FAILED; | |
710 | lock->l_flags |= flags; | |
711 | ||
712 | /* ... without sending a CANCEL message for local_only. */ | |
713 | if (local_only) | |
714 | lock->l_flags |= LDLM_FL_LOCAL_ONLY; | |
715 | ||
716 | if (local_only && (lock->l_readers || lock->l_writers)) { | |
717 | /* This is a little bit gross, but much better than the | |
718 | * alternative: pretend that we got a blocking AST from | |
719 | * the server, so that when the lock is decref'd, it | |
720 | * will go away ... */ | |
721 | unlock_res(res); | |
722 | LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY"); | |
723 | if (lock->l_completion_ast) | |
724 | lock->l_completion_ast(lock, 0, NULL); | |
725 | LDLM_LOCK_RELEASE(lock); | |
726 | continue; | |
727 | } | |
728 | ||
729 | if (client) { | |
730 | struct lustre_handle lockh; | |
731 | ||
732 | unlock_res(res); | |
733 | ldlm_lock2handle(lock, &lockh); | |
734 | rc = ldlm_cli_cancel(&lockh, LCF_ASYNC); | |
735 | if (rc) | |
736 | CERROR("ldlm_cli_cancel: %d\n", rc); | |
737 | } else { | |
738 | ldlm_resource_unlink_lock(lock); | |
739 | unlock_res(res); | |
740 | LDLM_DEBUG(lock, "Freeing a lock still held by a " | |
741 | "client node"); | |
742 | ldlm_lock_destroy(lock); | |
743 | } | |
744 | LDLM_LOCK_RELEASE(lock); | |
745 | } while (1); | |
746 | } | |
747 | ||
748 | static int ldlm_resource_clean(cfs_hash_t *hs, cfs_hash_bd_t *bd, | |
749 | struct hlist_node *hnode, void *arg) | |
750 | { | |
751 | struct ldlm_resource *res = cfs_hash_object(hs, hnode); | |
752 | __u64 flags = *(__u64 *)arg; | |
753 | ||
754 | cleanup_resource(res, &res->lr_granted, flags); | |
755 | cleanup_resource(res, &res->lr_converting, flags); | |
756 | cleanup_resource(res, &res->lr_waiting, flags); | |
757 | ||
758 | return 0; | |
759 | } | |
760 | ||
761 | static int ldlm_resource_complain(cfs_hash_t *hs, cfs_hash_bd_t *bd, | |
762 | struct hlist_node *hnode, void *arg) | |
763 | { | |
764 | struct ldlm_resource *res = cfs_hash_object(hs, hnode); | |
765 | ||
766 | lock_res(res); | |
767 | CERROR("Namespace %s resource refcount nonzero " | |
768 | "(%d) after lock cleanup; forcing " | |
769 | "cleanup.\n", | |
770 | ldlm_ns_name(ldlm_res_to_ns(res)), | |
771 | atomic_read(&res->lr_refcount) - 1); | |
772 | ||
773 | CERROR("Resource: %p ("LPU64"/"LPU64"/"LPU64"/" | |
774 | LPU64") (rc: %d)\n", res, | |
775 | res->lr_name.name[0], res->lr_name.name[1], | |
776 | res->lr_name.name[2], res->lr_name.name[3], | |
777 | atomic_read(&res->lr_refcount) - 1); | |
778 | ||
779 | ldlm_resource_dump(D_ERROR, res); | |
780 | unlock_res(res); | |
781 | return 0; | |
782 | } | |
783 | ||
784 | /** | |
785 | * Cancel and destroy all locks in the namespace. | |
786 | * | |
787 | * Typically used during evictions when server notified client that it was | |
788 | * evicted and all of its state needs to be destroyed. | |
789 | * Also used during shutdown. | |
790 | */ | |
791 | int ldlm_namespace_cleanup(struct ldlm_namespace *ns, __u64 flags) | |
792 | { | |
793 | if (ns == NULL) { | |
794 | CDEBUG(D_INFO, "NULL ns, skipping cleanup\n"); | |
795 | return ELDLM_OK; | |
796 | } | |
797 | ||
798 | cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean, &flags); | |
799 | cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain, NULL); | |
800 | return ELDLM_OK; | |
801 | } | |
802 | EXPORT_SYMBOL(ldlm_namespace_cleanup); | |
803 | ||
804 | /** | |
805 | * Attempts to free namespace. | |
806 | * | |
807 | * Only used when namespace goes away, like during an unmount. | |
808 | */ | |
809 | static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force) | |
810 | { | |
811 | ENTRY; | |
812 | ||
813 | /* At shutdown time, don't call the cancellation callback */ | |
814 | ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0); | |
815 | ||
816 | if (atomic_read(&ns->ns_bref) > 0) { | |
817 | struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL); | |
818 | int rc; | |
819 | CDEBUG(D_DLMTRACE, | |
820 | "dlm namespace %s free waiting on refcount %d\n", | |
821 | ldlm_ns_name(ns), atomic_read(&ns->ns_bref)); | |
822 | force_wait: | |
823 | if (force) | |
824 | lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL); | |
825 | ||
826 | rc = l_wait_event(ns->ns_waitq, | |
827 | atomic_read(&ns->ns_bref) == 0, &lwi); | |
828 | ||
829 | /* Forced cleanups should be able to reclaim all references, | |
830 | * so it's safe to wait forever... we can't leak locks... */ | |
831 | if (force && rc == -ETIMEDOUT) { | |
832 | LCONSOLE_ERROR("Forced cleanup waiting for %s " | |
833 | "namespace with %d resources in use, " | |
834 | "(rc=%d)\n", ldlm_ns_name(ns), | |
835 | atomic_read(&ns->ns_bref), rc); | |
836 | GOTO(force_wait, rc); | |
837 | } | |
838 | ||
839 | if (atomic_read(&ns->ns_bref)) { | |
840 | LCONSOLE_ERROR("Cleanup waiting for %s namespace " | |
841 | "with %d resources in use, (rc=%d)\n", | |
842 | ldlm_ns_name(ns), | |
843 | atomic_read(&ns->ns_bref), rc); | |
844 | RETURN(ELDLM_NAMESPACE_EXISTS); | |
845 | } | |
846 | CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n", | |
847 | ldlm_ns_name(ns)); | |
848 | } | |
849 | ||
850 | RETURN(ELDLM_OK); | |
851 | } | |
852 | ||
853 | /** | |
854 | * Performs various cleanups for passed \a ns to make it drop refc and be | |
855 | * ready for freeing. Waits for refc == 0. | |
856 | * | |
857 | * The following is done: | |
858 | * (0) Unregister \a ns from its list to make inaccessible for potential | |
859 | * users like pools thread and others; | |
860 | * (1) Clear all locks in \a ns. | |
861 | */ | |
862 | void ldlm_namespace_free_prior(struct ldlm_namespace *ns, | |
863 | struct obd_import *imp, | |
864 | int force) | |
865 | { | |
866 | int rc; | |
867 | ENTRY; | |
868 | if (!ns) { | |
869 | EXIT; | |
870 | return; | |
871 | } | |
872 | ||
873 | spin_lock(&ns->ns_lock); | |
874 | ns->ns_stopping = 1; | |
875 | spin_unlock(&ns->ns_lock); | |
876 | ||
877 | /* | |
878 | * Can fail with -EINTR when force == 0 in which case try harder. | |
879 | */ | |
880 | rc = __ldlm_namespace_free(ns, force); | |
881 | if (rc != ELDLM_OK) { | |
882 | if (imp) { | |
883 | ptlrpc_disconnect_import(imp, 0); | |
884 | ptlrpc_invalidate_import(imp); | |
885 | } | |
886 | ||
887 | /* | |
888 | * With all requests dropped and the import inactive | |
889 | * we are gaurenteed all reference will be dropped. | |
890 | */ | |
891 | rc = __ldlm_namespace_free(ns, 1); | |
892 | LASSERT(rc == 0); | |
893 | } | |
894 | EXIT; | |
895 | } | |
896 | ||
897 | /** | |
898 | * Performs freeing memory structures related to \a ns. This is only done | |
899 | * when ldlm_namespce_free_prior() successfully removed all resources | |
900 | * referencing \a ns and its refc == 0. | |
901 | */ | |
902 | void ldlm_namespace_free_post(struct ldlm_namespace *ns) | |
903 | { | |
904 | ENTRY; | |
905 | if (!ns) { | |
906 | EXIT; | |
907 | return; | |
908 | } | |
909 | ||
910 | /* Make sure that nobody can find this ns in its list. */ | |
911 | ldlm_namespace_unregister(ns, ns->ns_client); | |
912 | /* Fini pool _before_ parent proc dir is removed. This is important as | |
913 | * ldlm_pool_fini() removes own proc dir which is child to @dir. | |
914 | * Removing it after @dir may cause oops. */ | |
915 | ldlm_pool_fini(&ns->ns_pool); | |
916 | ||
917 | ldlm_namespace_proc_unregister(ns); | |
918 | cfs_hash_putref(ns->ns_rs_hash); | |
919 | /* Namespace \a ns should be not on list at this time, otherwise | |
920 | * this will cause issues related to using freed \a ns in poold | |
921 | * thread. */ | |
922 | LASSERT(list_empty(&ns->ns_list_chain)); | |
923 | OBD_FREE_PTR(ns); | |
924 | ldlm_put_ref(); | |
925 | EXIT; | |
926 | } | |
927 | ||
928 | /** | |
929 | * Cleanup the resource, and free namespace. | |
930 | * bug 12864: | |
931 | * Deadlock issue: | |
932 | * proc1: destroy import | |
933 | * class_disconnect_export(grab cl_sem) -> | |
934 | * -> ldlm_namespace_free -> | |
935 | * -> lprocfs_remove(grab _lprocfs_lock). | |
936 | * proc2: read proc info | |
937 | * lprocfs_fops_read(grab _lprocfs_lock) -> | |
938 | * -> osc_rd_active, etc(grab cl_sem). | |
939 | * | |
940 | * So that I have to split the ldlm_namespace_free into two parts - the first | |
941 | * part ldlm_namespace_free_prior is used to cleanup the resource which is | |
942 | * being used; the 2nd part ldlm_namespace_free_post is used to unregister the | |
943 | * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem | |
944 | * held. | |
945 | */ | |
946 | void ldlm_namespace_free(struct ldlm_namespace *ns, | |
947 | struct obd_import *imp, | |
948 | int force) | |
949 | { | |
950 | ldlm_namespace_free_prior(ns, imp, force); | |
951 | ldlm_namespace_free_post(ns); | |
952 | } | |
953 | EXPORT_SYMBOL(ldlm_namespace_free); | |
954 | ||
955 | void ldlm_namespace_get(struct ldlm_namespace *ns) | |
956 | { | |
957 | atomic_inc(&ns->ns_bref); | |
958 | } | |
959 | EXPORT_SYMBOL(ldlm_namespace_get); | |
960 | ||
91a50030 OD |
961 | /* This is only for callers that care about refcount */ |
962 | int ldlm_namespace_get_return(struct ldlm_namespace *ns) | |
963 | { | |
964 | return atomic_inc_return(&ns->ns_bref); | |
965 | } | |
966 | ||
d7e09d03 PT |
967 | void ldlm_namespace_put(struct ldlm_namespace *ns) |
968 | { | |
969 | if (atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) { | |
970 | wake_up(&ns->ns_waitq); | |
971 | spin_unlock(&ns->ns_lock); | |
972 | } | |
973 | } | |
974 | EXPORT_SYMBOL(ldlm_namespace_put); | |
975 | ||
976 | /** Register \a ns in the list of namespaces */ | |
977 | void ldlm_namespace_register(struct ldlm_namespace *ns, ldlm_side_t client) | |
978 | { | |
979 | mutex_lock(ldlm_namespace_lock(client)); | |
980 | LASSERT(list_empty(&ns->ns_list_chain)); | |
91a50030 OD |
981 | list_add(&ns->ns_list_chain, ldlm_namespace_inactive_list(client)); |
982 | ldlm_namespace_nr_inc(client); | |
d7e09d03 PT |
983 | mutex_unlock(ldlm_namespace_lock(client)); |
984 | } | |
985 | ||
986 | /** Unregister \a ns from the list of namespaces. */ | |
987 | void ldlm_namespace_unregister(struct ldlm_namespace *ns, ldlm_side_t client) | |
988 | { | |
989 | mutex_lock(ldlm_namespace_lock(client)); | |
990 | LASSERT(!list_empty(&ns->ns_list_chain)); | |
991 | /* Some asserts and possibly other parts of the code are still | |
992 | * using list_empty(&ns->ns_list_chain). This is why it is | |
993 | * important to use list_del_init() here. */ | |
994 | list_del_init(&ns->ns_list_chain); | |
91a50030 | 995 | ldlm_namespace_nr_dec(client); |
d7e09d03 PT |
996 | mutex_unlock(ldlm_namespace_lock(client)); |
997 | } | |
998 | ||
999 | /** Should be called with ldlm_namespace_lock(client) taken. */ | |
91a50030 OD |
1000 | void ldlm_namespace_move_to_active_locked(struct ldlm_namespace *ns, |
1001 | ldlm_side_t client) | |
d7e09d03 PT |
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, ldlm_namespace_list(client)); | |
1006 | } | |
1007 | ||
91a50030 OD |
1008 | /** Should be called with ldlm_namespace_lock(client) taken. */ |
1009 | void ldlm_namespace_move_to_inactive_locked(struct ldlm_namespace *ns, | |
1010 | ldlm_side_t client) | |
1011 | { | |
1012 | LASSERT(!list_empty(&ns->ns_list_chain)); | |
1013 | LASSERT(mutex_is_locked(ldlm_namespace_lock(client))); | |
1014 | list_move_tail(&ns->ns_list_chain, | |
1015 | ldlm_namespace_inactive_list(client)); | |
1016 | } | |
1017 | ||
d7e09d03 PT |
1018 | /** Should be called with ldlm_namespace_lock(client) taken. */ |
1019 | struct ldlm_namespace *ldlm_namespace_first_locked(ldlm_side_t client) | |
1020 | { | |
1021 | LASSERT(mutex_is_locked(ldlm_namespace_lock(client))); | |
1022 | LASSERT(!list_empty(ldlm_namespace_list(client))); | |
1023 | return container_of(ldlm_namespace_list(client)->next, | |
1024 | struct ldlm_namespace, ns_list_chain); | |
1025 | } | |
1026 | ||
1027 | /** Create and initialize new resource. */ | |
1028 | static struct ldlm_resource *ldlm_resource_new(void) | |
1029 | { | |
1030 | struct ldlm_resource *res; | |
1031 | int idx; | |
1032 | ||
1033 | OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, __GFP_IO); | |
1034 | if (res == NULL) | |
1035 | return NULL; | |
1036 | ||
1037 | INIT_LIST_HEAD(&res->lr_granted); | |
1038 | INIT_LIST_HEAD(&res->lr_converting); | |
1039 | INIT_LIST_HEAD(&res->lr_waiting); | |
1040 | ||
1041 | /* Initialize interval trees for each lock mode. */ | |
1042 | for (idx = 0; idx < LCK_MODE_NUM; idx++) { | |
1043 | res->lr_itree[idx].lit_size = 0; | |
1044 | res->lr_itree[idx].lit_mode = 1 << idx; | |
1045 | res->lr_itree[idx].lit_root = NULL; | |
1046 | } | |
1047 | ||
1048 | atomic_set(&res->lr_refcount, 1); | |
1049 | spin_lock_init(&res->lr_lock); | |
1050 | lu_ref_init(&res->lr_reference); | |
1051 | ||
1052 | /* The creator of the resource must unlock the mutex after LVB | |
1053 | * initialization. */ | |
1054 | mutex_init(&res->lr_lvb_mutex); | |
1055 | mutex_lock(&res->lr_lvb_mutex); | |
1056 | ||
1057 | return res; | |
1058 | } | |
1059 | ||
1060 | /** | |
1061 | * Return a reference to resource with given name, creating it if necessary. | |
1062 | * Args: namespace with ns_lock unlocked | |
1063 | * Locks: takes and releases NS hash-lock and res->lr_lock | |
1064 | * Returns: referenced, unlocked ldlm_resource or NULL | |
1065 | */ | |
1066 | struct ldlm_resource * | |
1067 | ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent, | |
1068 | const struct ldlm_res_id *name, ldlm_type_t type, int create) | |
1069 | { | |
1070 | struct hlist_node *hnode; | |
1071 | struct ldlm_resource *res; | |
1072 | cfs_hash_bd_t bd; | |
1073 | __u64 version; | |
91a50030 | 1074 | int ns_refcount = 0; |
d7e09d03 PT |
1075 | |
1076 | LASSERT(ns != NULL); | |
1077 | LASSERT(parent == NULL); | |
1078 | LASSERT(ns->ns_rs_hash != NULL); | |
1079 | LASSERT(name->name[0] != 0); | |
1080 | ||
1081 | cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0); | |
1082 | hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name); | |
1083 | if (hnode != NULL) { | |
1084 | cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0); | |
1085 | res = hlist_entry(hnode, struct ldlm_resource, lr_hash); | |
1086 | /* Synchronize with regard to resource creation. */ | |
1087 | if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) { | |
1088 | mutex_lock(&res->lr_lvb_mutex); | |
1089 | mutex_unlock(&res->lr_lvb_mutex); | |
1090 | } | |
1091 | ||
1092 | if (unlikely(res->lr_lvb_len < 0)) { | |
1093 | ldlm_resource_putref(res); | |
1094 | res = NULL; | |
1095 | } | |
1096 | return res; | |
1097 | } | |
1098 | ||
1099 | version = cfs_hash_bd_version_get(&bd); | |
1100 | cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0); | |
1101 | ||
1102 | if (create == 0) | |
1103 | return NULL; | |
1104 | ||
1105 | LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE, | |
1106 | "type: %d\n", type); | |
1107 | res = ldlm_resource_new(); | |
1108 | if (!res) | |
1109 | return NULL; | |
1110 | ||
1111 | res->lr_ns_bucket = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd); | |
1112 | res->lr_name = *name; | |
1113 | res->lr_type = type; | |
1114 | res->lr_most_restr = LCK_NL; | |
1115 | ||
1116 | cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1); | |
1117 | hnode = (version == cfs_hash_bd_version_get(&bd)) ? NULL : | |
1118 | cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name); | |
1119 | ||
1120 | if (hnode != NULL) { | |
1121 | /* Someone won the race and already added the resource. */ | |
1122 | cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1); | |
1123 | /* Clean lu_ref for failed resource. */ | |
1124 | lu_ref_fini(&res->lr_reference); | |
1125 | /* We have taken lr_lvb_mutex. Drop it. */ | |
1126 | mutex_unlock(&res->lr_lvb_mutex); | |
1127 | OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res); | |
1128 | ||
1129 | res = hlist_entry(hnode, struct ldlm_resource, lr_hash); | |
1130 | /* Synchronize with regard to resource creation. */ | |
1131 | if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) { | |
1132 | mutex_lock(&res->lr_lvb_mutex); | |
1133 | mutex_unlock(&res->lr_lvb_mutex); | |
1134 | } | |
1135 | ||
1136 | if (unlikely(res->lr_lvb_len < 0)) { | |
1137 | ldlm_resource_putref(res); | |
1138 | res = NULL; | |
1139 | } | |
1140 | return res; | |
1141 | } | |
1142 | /* We won! Let's add the resource. */ | |
1143 | cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash); | |
1144 | if (cfs_hash_bd_count_get(&bd) == 1) | |
91a50030 | 1145 | ns_refcount = ldlm_namespace_get_return(ns); |
d7e09d03 PT |
1146 | |
1147 | cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1); | |
1148 | if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) { | |
1149 | int rc; | |
1150 | ||
1151 | OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2); | |
1152 | rc = ns->ns_lvbo->lvbo_init(res); | |
1153 | if (rc < 0) { | |
c5b60ba7 AD |
1154 | CERROR("%s: lvbo_init failed for resource "LPX64":" |
1155 | LPX64": rc = %d\n", ns->ns_obd->obd_name, | |
1156 | name->name[0], name->name[1], rc); | |
d7e09d03 PT |
1157 | if (res->lr_lvb_data) { |
1158 | OBD_FREE(res->lr_lvb_data, res->lr_lvb_len); | |
1159 | res->lr_lvb_data = NULL; | |
1160 | } | |
1161 | res->lr_lvb_len = rc; | |
1162 | mutex_unlock(&res->lr_lvb_mutex); | |
1163 | ldlm_resource_putref(res); | |
1164 | return NULL; | |
1165 | } | |
1166 | } | |
1167 | ||
1168 | /* We create resource with locked lr_lvb_mutex. */ | |
1169 | mutex_unlock(&res->lr_lvb_mutex); | |
1170 | ||
91a50030 OD |
1171 | /* Let's see if we happened to be the very first resource in this |
1172 | * namespace. If so, and this is a client namespace, we need to move | |
1173 | * the namespace into the active namespaces list to be patrolled by | |
1174 | * the ldlm_poold. | |
1175 | * A notable exception, for quota namespaces qsd_lib.c already took a | |
1176 | * namespace reference, so it won't be participating in all of this, | |
1177 | * but I guess that's ok since we have no business cancelling quota | |
1178 | * locks anyway */ | |
1179 | if (ns_is_client(ns) && ns_refcount == 1) { | |
1180 | mutex_lock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT)); | |
1181 | ldlm_namespace_move_to_active_locked(ns, LDLM_NAMESPACE_CLIENT); | |
1182 | mutex_unlock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT)); | |
1183 | } | |
1184 | ||
d7e09d03 PT |
1185 | return res; |
1186 | } | |
1187 | EXPORT_SYMBOL(ldlm_resource_get); | |
1188 | ||
1189 | struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res) | |
1190 | { | |
1191 | LASSERT(res != NULL); | |
1192 | LASSERT(res != LP_POISON); | |
1193 | atomic_inc(&res->lr_refcount); | |
1194 | CDEBUG(D_INFO, "getref res: %p count: %d\n", res, | |
1195 | atomic_read(&res->lr_refcount)); | |
1196 | return res; | |
1197 | } | |
1198 | ||
1199 | static void __ldlm_resource_putref_final(cfs_hash_bd_t *bd, | |
1200 | struct ldlm_resource *res) | |
1201 | { | |
1202 | struct ldlm_ns_bucket *nsb = res->lr_ns_bucket; | |
1203 | ||
1204 | if (!list_empty(&res->lr_granted)) { | |
1205 | ldlm_resource_dump(D_ERROR, res); | |
1206 | LBUG(); | |
1207 | } | |
1208 | ||
1209 | if (!list_empty(&res->lr_converting)) { | |
1210 | ldlm_resource_dump(D_ERROR, res); | |
1211 | LBUG(); | |
1212 | } | |
1213 | ||
1214 | if (!list_empty(&res->lr_waiting)) { | |
1215 | ldlm_resource_dump(D_ERROR, res); | |
1216 | LBUG(); | |
1217 | } | |
1218 | ||
1219 | cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash, | |
1220 | bd, &res->lr_hash); | |
1221 | lu_ref_fini(&res->lr_reference); | |
1222 | if (cfs_hash_bd_count_get(bd) == 0) | |
1223 | ldlm_namespace_put(nsb->nsb_namespace); | |
1224 | } | |
1225 | ||
1226 | /* Returns 1 if the resource was freed, 0 if it remains. */ | |
1227 | int ldlm_resource_putref(struct ldlm_resource *res) | |
1228 | { | |
1229 | struct ldlm_namespace *ns = ldlm_res_to_ns(res); | |
1230 | cfs_hash_bd_t bd; | |
1231 | ||
1232 | LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON); | |
1233 | CDEBUG(D_INFO, "putref res: %p count: %d\n", | |
1234 | res, atomic_read(&res->lr_refcount) - 1); | |
1235 | ||
1236 | cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd); | |
1237 | if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) { | |
1238 | __ldlm_resource_putref_final(&bd, res); | |
1239 | cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1); | |
1240 | if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free) | |
1241 | ns->ns_lvbo->lvbo_free(res); | |
1242 | OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res); | |
1243 | return 1; | |
1244 | } | |
1245 | return 0; | |
1246 | } | |
1247 | EXPORT_SYMBOL(ldlm_resource_putref); | |
1248 | ||
1249 | /* Returns 1 if the resource was freed, 0 if it remains. */ | |
1250 | int ldlm_resource_putref_locked(struct ldlm_resource *res) | |
1251 | { | |
1252 | struct ldlm_namespace *ns = ldlm_res_to_ns(res); | |
1253 | ||
1254 | LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON); | |
1255 | CDEBUG(D_INFO, "putref res: %p count: %d\n", | |
1256 | res, atomic_read(&res->lr_refcount) - 1); | |
1257 | ||
1258 | if (atomic_dec_and_test(&res->lr_refcount)) { | |
1259 | cfs_hash_bd_t bd; | |
1260 | ||
1261 | cfs_hash_bd_get(ldlm_res_to_ns(res)->ns_rs_hash, | |
1262 | &res->lr_name, &bd); | |
1263 | __ldlm_resource_putref_final(&bd, res); | |
1264 | cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1); | |
1265 | /* NB: ns_rs_hash is created with CFS_HASH_NO_ITEMREF, | |
1266 | * so we should never be here while calling cfs_hash_del, | |
1267 | * cfs_hash_for_each_nolock is the only case we can get | |
1268 | * here, which is safe to release cfs_hash_bd_lock. | |
1269 | */ | |
1270 | if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free) | |
1271 | ns->ns_lvbo->lvbo_free(res); | |
1272 | OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res); | |
1273 | ||
1274 | cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1); | |
1275 | return 1; | |
1276 | } | |
1277 | return 0; | |
1278 | } | |
1279 | ||
1280 | /** | |
1281 | * Add a lock into a given resource into specified lock list. | |
1282 | */ | |
1283 | void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head, | |
1284 | struct ldlm_lock *lock) | |
1285 | { | |
1286 | check_res_locked(res); | |
1287 | ||
1288 | LDLM_DEBUG(lock, "About to add this lock:\n"); | |
1289 | ||
1290 | if (lock->l_destroyed) { | |
1291 | CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n"); | |
1292 | return; | |
1293 | } | |
1294 | ||
1295 | LASSERT(list_empty(&lock->l_res_link)); | |
1296 | ||
1297 | list_add_tail(&lock->l_res_link, head); | |
1298 | } | |
1299 | ||
1300 | /** | |
1301 | * Insert a lock into resource after specified lock. | |
1302 | * | |
1303 | * Obtain resource description from the lock we are inserting after. | |
1304 | */ | |
1305 | void ldlm_resource_insert_lock_after(struct ldlm_lock *original, | |
1306 | struct ldlm_lock *new) | |
1307 | { | |
1308 | struct ldlm_resource *res = original->l_resource; | |
1309 | ||
1310 | check_res_locked(res); | |
1311 | ||
1312 | ldlm_resource_dump(D_INFO, res); | |
1313 | LDLM_DEBUG(new, "About to insert this lock after %p:\n", original); | |
1314 | ||
1315 | if (new->l_destroyed) { | |
1316 | CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n"); | |
1317 | goto out; | |
1318 | } | |
1319 | ||
1320 | LASSERT(list_empty(&new->l_res_link)); | |
1321 | ||
1322 | list_add(&new->l_res_link, &original->l_res_link); | |
1323 | out:; | |
1324 | } | |
1325 | ||
1326 | void ldlm_resource_unlink_lock(struct ldlm_lock *lock) | |
1327 | { | |
1328 | int type = lock->l_resource->lr_type; | |
1329 | ||
1330 | check_res_locked(lock->l_resource); | |
1331 | if (type == LDLM_IBITS || type == LDLM_PLAIN) | |
1332 | ldlm_unlink_lock_skiplist(lock); | |
1333 | else if (type == LDLM_EXTENT) | |
1334 | ldlm_extent_unlink_lock(lock); | |
1335 | list_del_init(&lock->l_res_link); | |
1336 | } | |
1337 | EXPORT_SYMBOL(ldlm_resource_unlink_lock); | |
1338 | ||
1339 | void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc) | |
1340 | { | |
1341 | desc->lr_type = res->lr_type; | |
1342 | desc->lr_name = res->lr_name; | |
1343 | } | |
1344 | ||
1345 | /** | |
1346 | * Print information about all locks in all namespaces on this node to debug | |
1347 | * log. | |
1348 | */ | |
1349 | void ldlm_dump_all_namespaces(ldlm_side_t client, int level) | |
1350 | { | |
1351 | struct list_head *tmp; | |
1352 | ||
1353 | if (!((libcfs_debug | D_ERROR) & level)) | |
1354 | return; | |
1355 | ||
1356 | mutex_lock(ldlm_namespace_lock(client)); | |
1357 | ||
1358 | list_for_each(tmp, ldlm_namespace_list(client)) { | |
1359 | struct ldlm_namespace *ns; | |
1360 | ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain); | |
1361 | ldlm_namespace_dump(level, ns); | |
1362 | } | |
1363 | ||
1364 | mutex_unlock(ldlm_namespace_lock(client)); | |
1365 | } | |
1366 | EXPORT_SYMBOL(ldlm_dump_all_namespaces); | |
1367 | ||
1368 | static int ldlm_res_hash_dump(cfs_hash_t *hs, cfs_hash_bd_t *bd, | |
1369 | struct hlist_node *hnode, void *arg) | |
1370 | { | |
1371 | struct ldlm_resource *res = cfs_hash_object(hs, hnode); | |
1372 | int level = (int)(unsigned long)arg; | |
1373 | ||
1374 | lock_res(res); | |
1375 | ldlm_resource_dump(level, res); | |
1376 | unlock_res(res); | |
1377 | ||
1378 | return 0; | |
1379 | } | |
1380 | ||
1381 | /** | |
1382 | * Print information about all locks in this namespace on this node to debug | |
1383 | * log. | |
1384 | */ | |
1385 | void ldlm_namespace_dump(int level, struct ldlm_namespace *ns) | |
1386 | { | |
1387 | if (!((libcfs_debug | D_ERROR) & level)) | |
1388 | return; | |
1389 | ||
1390 | CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n", | |
1391 | ldlm_ns_name(ns), atomic_read(&ns->ns_bref), | |
1392 | ns_is_client(ns) ? "client" : "server"); | |
1393 | ||
1394 | if (cfs_time_before(cfs_time_current(), ns->ns_next_dump)) | |
1395 | return; | |
1396 | ||
1397 | cfs_hash_for_each_nolock(ns->ns_rs_hash, | |
1398 | ldlm_res_hash_dump, | |
1399 | (void *)(unsigned long)level); | |
1400 | spin_lock(&ns->ns_lock); | |
1401 | ns->ns_next_dump = cfs_time_shift(10); | |
1402 | spin_unlock(&ns->ns_lock); | |
1403 | } | |
1404 | EXPORT_SYMBOL(ldlm_namespace_dump); | |
1405 | ||
1406 | /** | |
1407 | * Print information about all locks in this resource to debug log. | |
1408 | */ | |
1409 | void ldlm_resource_dump(int level, struct ldlm_resource *res) | |
1410 | { | |
1411 | struct ldlm_lock *lock; | |
1412 | unsigned int granted = 0; | |
1413 | ||
1414 | CLASSERT(RES_NAME_SIZE == 4); | |
1415 | ||
1416 | if (!((libcfs_debug | D_ERROR) & level)) | |
1417 | return; | |
1418 | ||
1419 | CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64 | |
1420 | ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1], | |
1421 | res->lr_name.name[2], res->lr_name.name[3], | |
1422 | atomic_read(&res->lr_refcount)); | |
1423 | ||
1424 | if (!list_empty(&res->lr_granted)) { | |
1425 | CDEBUG(level, "Granted locks (in reverse order):\n"); | |
1426 | list_for_each_entry_reverse(lock, &res->lr_granted, | |
1427 | l_res_link) { | |
1428 | LDLM_DEBUG_LIMIT(level, lock, "###"); | |
1429 | if (!(level & D_CANTMASK) && | |
1430 | ++granted > ldlm_dump_granted_max) { | |
1431 | CDEBUG(level, "only dump %d granted locks to " | |
1432 | "avoid DDOS.\n", granted); | |
1433 | break; | |
1434 | } | |
1435 | } | |
1436 | } | |
1437 | if (!list_empty(&res->lr_converting)) { | |
1438 | CDEBUG(level, "Converting locks:\n"); | |
1439 | list_for_each_entry(lock, &res->lr_converting, l_res_link) | |
1440 | LDLM_DEBUG_LIMIT(level, lock, "###"); | |
1441 | } | |
1442 | if (!list_empty(&res->lr_waiting)) { | |
1443 | CDEBUG(level, "Waiting locks:\n"); | |
1444 | list_for_each_entry(lock, &res->lr_waiting, l_res_link) | |
1445 | LDLM_DEBUG_LIMIT(level, lock, "###"); | |
1446 | } | |
1447 | } |