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) 2007, 2010, Oracle and/or its affiliates. All rights reserved. | |
28 | * Use is subject to license terms. | |
29 | * | |
1dc563a6 | 30 | * Copyright (c) 2011, 2015, Intel Corporation. |
d7e09d03 PT |
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/obdclass/lu_object.c | |
37 | * | |
38 | * Lustre Object. | |
39 | * These are the only exported functions, they provide some generic | |
40 | * infrastructure for managing object devices | |
41 | * | |
42 | * Author: Nikita Danilov <nikita.danilov@sun.com> | |
43 | */ | |
44 | ||
45 | #define DEBUG_SUBSYSTEM S_CLASS | |
46 | ||
9fdaf8c0 | 47 | #include "../../include/linux/libcfs/libcfs.h" |
d7e09d03 PT |
48 | |
49 | # include <linux/module.h> | |
50 | ||
51 | /* hash_long() */ | |
9fdaf8c0 | 52 | #include "../../include/linux/libcfs/libcfs_hash.h" |
610f7377 GKH |
53 | #include "../include/obd_class.h" |
54 | #include "../include/obd_support.h" | |
55 | #include "../include/lustre_disk.h" | |
56 | #include "../include/lustre_fid.h" | |
57 | #include "../include/lu_object.h" | |
58 | #include "../include/lu_ref.h" | |
d7e09d03 PT |
59 | #include <linux/list.h> |
60 | ||
61 | static void lu_object_free(const struct lu_env *env, struct lu_object *o); | |
a0b8803a | 62 | static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx); |
d7e09d03 PT |
63 | |
64 | /** | |
65 | * Decrease reference counter on object. If last reference is freed, return | |
66 | * object to the cache, unless lu_object_is_dying(o) holds. In the latter | |
67 | * case, free object immediately. | |
68 | */ | |
69 | void lu_object_put(const struct lu_env *env, struct lu_object *o) | |
70 | { | |
71 | struct lu_site_bkt_data *bkt; | |
72 | struct lu_object_header *top; | |
73 | struct lu_site *site; | |
74 | struct lu_object *orig; | |
6ea510c1 | 75 | struct cfs_hash_bd bd; |
d7e09d03 PT |
76 | const struct lu_fid *fid; |
77 | ||
78 | top = o->lo_header; | |
79 | site = o->lo_dev->ld_site; | |
80 | orig = o; | |
81 | ||
82 | /* | |
83 | * till we have full fids-on-OST implemented anonymous objects | |
84 | * are possible in OSP. such an object isn't listed in the site | |
85 | * so we should not remove it from the site. | |
86 | */ | |
87 | fid = lu_object_fid(o); | |
88 | if (fid_is_zero(fid)) { | |
cce3c2da | 89 | LASSERT(!top->loh_hash.next && !top->loh_hash.pprev); |
d7e09d03 PT |
90 | LASSERT(list_empty(&top->loh_lru)); |
91 | if (!atomic_dec_and_test(&top->loh_ref)) | |
92 | return; | |
93 | list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) { | |
cce3c2da | 94 | if (o->lo_ops->loo_object_release) |
d7e09d03 PT |
95 | o->lo_ops->loo_object_release(env, o); |
96 | } | |
97 | lu_object_free(env, orig); | |
98 | return; | |
99 | } | |
100 | ||
101 | cfs_hash_bd_get(site->ls_obj_hash, &top->loh_fid, &bd); | |
102 | bkt = cfs_hash_bd_extra_get(site->ls_obj_hash, &bd); | |
103 | ||
104 | if (!cfs_hash_bd_dec_and_lock(site->ls_obj_hash, &bd, &top->loh_ref)) { | |
105 | if (lu_object_is_dying(top)) { | |
106 | ||
107 | /* | |
108 | * somebody may be waiting for this, currently only | |
109 | * used for cl_object, see cl_object_put_last(). | |
110 | */ | |
111 | wake_up_all(&bkt->lsb_marche_funebre); | |
112 | } | |
113 | return; | |
114 | } | |
115 | ||
d7e09d03 PT |
116 | /* |
117 | * When last reference is released, iterate over object | |
118 | * layers, and notify them that object is no longer busy. | |
119 | */ | |
120 | list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) { | |
cce3c2da | 121 | if (o->lo_ops->loo_object_release) |
d7e09d03 PT |
122 | o->lo_ops->loo_object_release(env, o); |
123 | } | |
124 | ||
125 | if (!lu_object_is_dying(top)) { | |
126 | LASSERT(list_empty(&top->loh_lru)); | |
127 | list_add_tail(&top->loh_lru, &bkt->lsb_lru); | |
6e580ab5 | 128 | bkt->lsb_lru_len++; |
a0b8803a AK |
129 | lprocfs_counter_incr(site->ls_stats, LU_SS_LRU_LEN); |
130 | CDEBUG(D_INODE, "Add %p to site lru. hash: %p, bkt: %p, lru_len: %ld\n", | |
131 | o, site->ls_obj_hash, bkt, bkt->lsb_lru_len); | |
d7e09d03 PT |
132 | cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1); |
133 | return; | |
134 | } | |
135 | ||
136 | /* | |
9c379663 | 137 | * If object is dying (will not be cached), then removed it |
d7e09d03 PT |
138 | * from hash table and LRU. |
139 | * | |
140 | * This is done with hash table and LRU lists locked. As the only | |
141 | * way to acquire first reference to previously unreferenced | |
142 | * object is through hash-table lookup (lu_object_find()), | |
143 | * or LRU scanning (lu_site_purge()), that are done under hash-table | |
144 | * and LRU lock, no race with concurrent object lookup is possible | |
145 | * and we can safely destroy object below. | |
146 | */ | |
147 | if (!test_and_set_bit(LU_OBJECT_UNHASHED, &top->loh_flags)) | |
148 | cfs_hash_bd_del_locked(site->ls_obj_hash, &bd, &top->loh_hash); | |
149 | cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1); | |
150 | /* | |
151 | * Object was already removed from hash and lru above, can | |
152 | * kill it. | |
153 | */ | |
154 | lu_object_free(env, orig); | |
155 | } | |
156 | EXPORT_SYMBOL(lu_object_put); | |
157 | ||
d7e09d03 PT |
158 | /** |
159 | * Kill the object and take it out of LRU cache. | |
160 | * Currently used by client code for layout change. | |
161 | */ | |
162 | void lu_object_unhash(const struct lu_env *env, struct lu_object *o) | |
163 | { | |
164 | struct lu_object_header *top; | |
165 | ||
166 | top = o->lo_header; | |
167 | set_bit(LU_OBJECT_HEARD_BANSHEE, &top->loh_flags); | |
168 | if (!test_and_set_bit(LU_OBJECT_UNHASHED, &top->loh_flags)) { | |
a0b8803a AK |
169 | struct lu_site *site = o->lo_dev->ld_site; |
170 | struct cfs_hash *obj_hash = site->ls_obj_hash; | |
6ea510c1 | 171 | struct cfs_hash_bd bd; |
d7e09d03 PT |
172 | |
173 | cfs_hash_bd_get_and_lock(obj_hash, &top->loh_fid, &bd, 1); | |
6e580ab5 FZ |
174 | if (!list_empty(&top->loh_lru)) { |
175 | struct lu_site_bkt_data *bkt; | |
176 | ||
a0b8803a | 177 | list_del_init(&top->loh_lru); |
6e580ab5 FZ |
178 | bkt = cfs_hash_bd_extra_get(obj_hash, &bd); |
179 | bkt->lsb_lru_len--; | |
a0b8803a | 180 | lprocfs_counter_decr(site->ls_stats, LU_SS_LRU_LEN); |
6e580ab5 | 181 | } |
d7e09d03 PT |
182 | cfs_hash_bd_del_locked(obj_hash, &bd, &top->loh_hash); |
183 | cfs_hash_bd_unlock(obj_hash, &bd, 1); | |
184 | } | |
185 | } | |
186 | EXPORT_SYMBOL(lu_object_unhash); | |
187 | ||
188 | /** | |
189 | * Allocate new object. | |
190 | * | |
191 | * This follows object creation protocol, described in the comment within | |
192 | * struct lu_device_operations definition. | |
193 | */ | |
194 | static struct lu_object *lu_object_alloc(const struct lu_env *env, | |
195 | struct lu_device *dev, | |
196 | const struct lu_fid *f, | |
197 | const struct lu_object_conf *conf) | |
198 | { | |
199 | struct lu_object *scan; | |
200 | struct lu_object *top; | |
201 | struct list_head *layers; | |
7cd875d2 JH |
202 | unsigned int init_mask = 0; |
203 | unsigned int init_flag; | |
d7e09d03 PT |
204 | int clean; |
205 | int result; | |
d7e09d03 PT |
206 | |
207 | /* | |
208 | * Create top-level object slice. This will also create | |
209 | * lu_object_header. | |
210 | */ | |
211 | top = dev->ld_ops->ldo_object_alloc(env, NULL, dev); | |
cce3c2da | 212 | if (!top) |
0a3bdb00 | 213 | return ERR_PTR(-ENOMEM); |
d7e09d03 | 214 | if (IS_ERR(top)) |
0a3bdb00 | 215 | return top; |
d7e09d03 PT |
216 | /* |
217 | * This is the only place where object fid is assigned. It's constant | |
218 | * after this point. | |
219 | */ | |
220 | top->lo_header->loh_fid = *f; | |
221 | layers = &top->lo_header->loh_layers; | |
7cd875d2 | 222 | |
d7e09d03 PT |
223 | do { |
224 | /* | |
225 | * Call ->loo_object_init() repeatedly, until no more new | |
226 | * object slices are created. | |
227 | */ | |
228 | clean = 1; | |
7cd875d2 | 229 | init_flag = 1; |
d7e09d03 | 230 | list_for_each_entry(scan, layers, lo_linkage) { |
7cd875d2 JH |
231 | if (init_mask & init_flag) |
232 | goto next; | |
d7e09d03 PT |
233 | clean = 0; |
234 | scan->lo_header = top->lo_header; | |
235 | result = scan->lo_ops->loo_object_init(env, scan, conf); | |
236 | if (result != 0) { | |
237 | lu_object_free(env, top); | |
0a3bdb00 | 238 | return ERR_PTR(result); |
d7e09d03 | 239 | } |
7cd875d2 JH |
240 | init_mask |= init_flag; |
241 | next: | |
242 | init_flag <<= 1; | |
d7e09d03 PT |
243 | } |
244 | } while (!clean); | |
245 | ||
246 | list_for_each_entry_reverse(scan, layers, lo_linkage) { | |
cce3c2da | 247 | if (scan->lo_ops->loo_object_start) { |
d7e09d03 PT |
248 | result = scan->lo_ops->loo_object_start(env, scan); |
249 | if (result != 0) { | |
250 | lu_object_free(env, top); | |
0a3bdb00 | 251 | return ERR_PTR(result); |
d7e09d03 PT |
252 | } |
253 | } | |
254 | } | |
255 | ||
256 | lprocfs_counter_incr(dev->ld_site->ls_stats, LU_SS_CREATED); | |
0a3bdb00 | 257 | return top; |
d7e09d03 PT |
258 | } |
259 | ||
260 | /** | |
261 | * Free an object. | |
262 | */ | |
263 | static void lu_object_free(const struct lu_env *env, struct lu_object *o) | |
264 | { | |
265 | struct lu_site_bkt_data *bkt; | |
266 | struct lu_site *site; | |
267 | struct lu_object *scan; | |
268 | struct list_head *layers; | |
269 | struct list_head splice; | |
270 | ||
271 | site = o->lo_dev->ld_site; | |
272 | layers = &o->lo_header->loh_layers; | |
273 | bkt = lu_site_bkt_from_fid(site, &o->lo_header->loh_fid); | |
274 | /* | |
275 | * First call ->loo_object_delete() method to release all resources. | |
276 | */ | |
277 | list_for_each_entry_reverse(scan, layers, lo_linkage) { | |
cce3c2da | 278 | if (scan->lo_ops->loo_object_delete) |
d7e09d03 PT |
279 | scan->lo_ops->loo_object_delete(env, scan); |
280 | } | |
281 | ||
282 | /* | |
283 | * Then, splice object layers into stand-alone list, and call | |
284 | * ->loo_object_free() on all layers to free memory. Splice is | |
285 | * necessary, because lu_object_header is freed together with the | |
286 | * top-level slice. | |
287 | */ | |
288 | INIT_LIST_HEAD(&splice); | |
289 | list_splice_init(layers, &splice); | |
290 | while (!list_empty(&splice)) { | |
291 | /* | |
292 | * Free layers in bottom-to-top order, so that object header | |
293 | * lives as long as possible and ->loo_object_free() methods | |
294 | * can look at its contents. | |
295 | */ | |
296 | o = container_of0(splice.prev, struct lu_object, lo_linkage); | |
297 | list_del_init(&o->lo_linkage); | |
d7e09d03 PT |
298 | o->lo_ops->loo_object_free(env, o); |
299 | } | |
300 | ||
301 | if (waitqueue_active(&bkt->lsb_marche_funebre)) | |
302 | wake_up_all(&bkt->lsb_marche_funebre); | |
303 | } | |
304 | ||
305 | /** | |
306 | * Free \a nr objects from the cold end of the site LRU list. | |
307 | */ | |
308 | int lu_site_purge(const struct lu_env *env, struct lu_site *s, int nr) | |
309 | { | |
310 | struct lu_object_header *h; | |
311 | struct lu_object_header *temp; | |
312 | struct lu_site_bkt_data *bkt; | |
6ea510c1 LN |
313 | struct cfs_hash_bd bd; |
314 | struct cfs_hash_bd bd2; | |
d7e09d03 PT |
315 | struct list_head dispose; |
316 | int did_sth; | |
317 | int start; | |
318 | int count; | |
319 | int bnr; | |
320 | int i; | |
321 | ||
322 | if (OBD_FAIL_CHECK(OBD_FAIL_OBD_NO_LRU)) | |
0a3bdb00 | 323 | return 0; |
d7e09d03 PT |
324 | |
325 | INIT_LIST_HEAD(&dispose); | |
326 | /* | |
327 | * Under LRU list lock, scan LRU list and move unreferenced objects to | |
328 | * the dispose list, removing them from LRU and hash table. | |
329 | */ | |
330 | start = s->ls_purge_start; | |
331 | bnr = (nr == ~0) ? -1 : nr / CFS_HASH_NBKT(s->ls_obj_hash) + 1; | |
332 | again: | |
333 | did_sth = 0; | |
334 | cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) { | |
335 | if (i < start) | |
336 | continue; | |
337 | count = bnr; | |
338 | cfs_hash_bd_lock(s->ls_obj_hash, &bd, 1); | |
339 | bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd); | |
340 | ||
341 | list_for_each_entry_safe(h, temp, &bkt->lsb_lru, loh_lru) { | |
342 | LASSERT(atomic_read(&h->loh_ref) == 0); | |
343 | ||
344 | cfs_hash_bd_get(s->ls_obj_hash, &h->loh_fid, &bd2); | |
345 | LASSERT(bd.bd_bucket == bd2.bd_bucket); | |
346 | ||
347 | cfs_hash_bd_del_locked(s->ls_obj_hash, | |
348 | &bd2, &h->loh_hash); | |
349 | list_move(&h->loh_lru, &dispose); | |
6e580ab5 | 350 | bkt->lsb_lru_len--; |
a0b8803a | 351 | lprocfs_counter_decr(s->ls_stats, LU_SS_LRU_LEN); |
d7e09d03 PT |
352 | if (did_sth == 0) |
353 | did_sth = 1; | |
354 | ||
355 | if (nr != ~0 && --nr == 0) | |
356 | break; | |
357 | ||
358 | if (count > 0 && --count == 0) | |
359 | break; | |
360 | ||
361 | } | |
362 | cfs_hash_bd_unlock(s->ls_obj_hash, &bd, 1); | |
363 | cond_resched(); | |
364 | /* | |
365 | * Free everything on the dispose list. This is safe against | |
366 | * races due to the reasons described in lu_object_put(). | |
367 | */ | |
368 | while (!list_empty(&dispose)) { | |
369 | h = container_of0(dispose.next, | |
370 | struct lu_object_header, loh_lru); | |
371 | list_del_init(&h->loh_lru); | |
372 | lu_object_free(env, lu_object_top(h)); | |
373 | lprocfs_counter_incr(s->ls_stats, LU_SS_LRU_PURGED); | |
374 | } | |
375 | ||
376 | if (nr == 0) | |
377 | break; | |
378 | } | |
379 | ||
380 | if (nr != 0 && did_sth && start != 0) { | |
381 | start = 0; /* restart from the first bucket */ | |
382 | goto again; | |
383 | } | |
384 | /* race on s->ls_purge_start, but nobody cares */ | |
385 | s->ls_purge_start = i % CFS_HASH_NBKT(s->ls_obj_hash); | |
386 | ||
387 | return nr; | |
388 | } | |
389 | EXPORT_SYMBOL(lu_site_purge); | |
390 | ||
391 | /* | |
392 | * Object printing. | |
393 | * | |
394 | * Code below has to jump through certain loops to output object description | |
395 | * into libcfs_debug_msg-based log. The problem is that lu_object_print() | |
396 | * composes object description from strings that are parts of _lines_ of | |
397 | * output (i.e., strings that are not terminated by newline). This doesn't fit | |
398 | * very well into libcfs_debug_msg() interface that assumes that each message | |
399 | * supplied to it is a self-contained output line. | |
400 | * | |
401 | * To work around this, strings are collected in a temporary buffer | |
402 | * (implemented as a value of lu_cdebug_key key), until terminating newline | |
403 | * character is detected. | |
404 | * | |
405 | */ | |
406 | ||
407 | enum { | |
408 | /** | |
409 | * Maximal line size. | |
410 | * | |
411 | * XXX overflow is not handled correctly. | |
412 | */ | |
413 | LU_CDEBUG_LINE = 512 | |
414 | }; | |
415 | ||
416 | struct lu_cdebug_data { | |
417 | /** | |
418 | * Temporary buffer. | |
419 | */ | |
420 | char lck_area[LU_CDEBUG_LINE]; | |
421 | }; | |
422 | ||
423 | /* context key constructor/destructor: lu_global_key_init, lu_global_key_fini */ | |
424 | LU_KEY_INIT_FINI(lu_global, struct lu_cdebug_data); | |
425 | ||
426 | /** | |
427 | * Key, holding temporary buffer. This key is registered very early by | |
428 | * lu_global_init(). | |
429 | */ | |
05311893 | 430 | static struct lu_context_key lu_global_key = { |
d7e09d03 | 431 | .lct_tags = LCT_MD_THREAD | LCT_DT_THREAD | |
aa4e3c8a | 432 | LCT_MG_THREAD | LCT_CL_THREAD | LCT_LOCAL, |
d7e09d03 PT |
433 | .lct_init = lu_global_key_init, |
434 | .lct_fini = lu_global_key_fini | |
435 | }; | |
436 | ||
437 | /** | |
438 | * Printer function emitting messages through libcfs_debug_msg(). | |
439 | */ | |
440 | int lu_cdebug_printer(const struct lu_env *env, | |
441 | void *cookie, const char *format, ...) | |
442 | { | |
443 | struct libcfs_debug_msg_data *msgdata = cookie; | |
444 | struct lu_cdebug_data *key; | |
445 | int used; | |
446 | int complete; | |
447 | va_list args; | |
448 | ||
449 | va_start(args, format); | |
450 | ||
451 | key = lu_context_key_get(&env->le_ctx, &lu_global_key); | |
d7e09d03 PT |
452 | |
453 | used = strlen(key->lck_area); | |
454 | complete = format[strlen(format) - 1] == '\n'; | |
455 | /* | |
456 | * Append new chunk to the buffer. | |
457 | */ | |
458 | vsnprintf(key->lck_area + used, | |
459 | ARRAY_SIZE(key->lck_area) - used, format, args); | |
460 | if (complete) { | |
461 | if (cfs_cdebug_show(msgdata->msg_mask, msgdata->msg_subsys)) | |
19b2056f | 462 | libcfs_debug_msg(msgdata, "%s\n", key->lck_area); |
d7e09d03 PT |
463 | key->lck_area[0] = 0; |
464 | } | |
465 | va_end(args); | |
466 | return 0; | |
467 | } | |
468 | EXPORT_SYMBOL(lu_cdebug_printer); | |
469 | ||
470 | /** | |
471 | * Print object header. | |
472 | */ | |
473 | void lu_object_header_print(const struct lu_env *env, void *cookie, | |
474 | lu_printer_t printer, | |
475 | const struct lu_object_header *hdr) | |
476 | { | |
477 | (*printer)(env, cookie, "header@%p[%#lx, %d, "DFID"%s%s%s]", | |
478 | hdr, hdr->loh_flags, atomic_read(&hdr->loh_ref), | |
479 | PFID(&hdr->loh_fid), | |
480 | hlist_unhashed(&hdr->loh_hash) ? "" : " hash", | |
481 | list_empty((struct list_head *)&hdr->loh_lru) ? \ | |
482 | "" : " lru", | |
483 | hdr->loh_attr & LOHA_EXISTS ? " exist":""); | |
484 | } | |
485 | EXPORT_SYMBOL(lu_object_header_print); | |
486 | ||
487 | /** | |
488 | * Print human readable representation of the \a o to the \a printer. | |
489 | */ | |
490 | void lu_object_print(const struct lu_env *env, void *cookie, | |
491 | lu_printer_t printer, const struct lu_object *o) | |
492 | { | |
493 | static const char ruler[] = "........................................"; | |
494 | struct lu_object_header *top; | |
7cd875d2 | 495 | int depth = 4; |
d7e09d03 PT |
496 | |
497 | top = o->lo_header; | |
498 | lu_object_header_print(env, cookie, printer, top); | |
7cd875d2 | 499 | (*printer)(env, cookie, "{\n"); |
d7e09d03 | 500 | |
7cd875d2 | 501 | list_for_each_entry(o, &top->loh_layers, lo_linkage) { |
d7e09d03 PT |
502 | /* |
503 | * print `.' \a depth times followed by type name and address | |
504 | */ | |
505 | (*printer)(env, cookie, "%*.*s%s@%p", depth, depth, ruler, | |
506 | o->lo_dev->ld_type->ldt_name, o); | |
7cd875d2 | 507 | |
cce3c2da | 508 | if (o->lo_ops->loo_object_print) |
7cd875d2 JH |
509 | (*o->lo_ops->loo_object_print)(env, cookie, printer, o); |
510 | ||
d7e09d03 PT |
511 | (*printer)(env, cookie, "\n"); |
512 | } | |
7cd875d2 | 513 | |
d7e09d03 PT |
514 | (*printer)(env, cookie, "} header@%p\n", top); |
515 | } | |
516 | EXPORT_SYMBOL(lu_object_print); | |
517 | ||
d7e09d03 | 518 | static struct lu_object *htable_lookup(struct lu_site *s, |
6ea510c1 | 519 | struct cfs_hash_bd *bd, |
d7e09d03 PT |
520 | const struct lu_fid *f, |
521 | wait_queue_t *waiter, | |
522 | __u64 *version) | |
523 | { | |
524 | struct lu_site_bkt_data *bkt; | |
525 | struct lu_object_header *h; | |
526 | struct hlist_node *hnode; | |
527 | __u64 ver = cfs_hash_bd_version_get(bd); | |
528 | ||
529 | if (*version == ver) | |
70b749d4 | 530 | return ERR_PTR(-ENOENT); |
d7e09d03 PT |
531 | |
532 | *version = ver; | |
533 | bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, bd); | |
534 | /* cfs_hash_bd_peek_locked is a somehow "internal" function | |
6ba59179 OD |
535 | * of cfs_hash, it doesn't add refcount on object. |
536 | */ | |
d7e09d03 | 537 | hnode = cfs_hash_bd_peek_locked(s->ls_obj_hash, bd, (void *)f); |
cce3c2da | 538 | if (!hnode) { |
d7e09d03 | 539 | lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS); |
70b749d4 | 540 | return ERR_PTR(-ENOENT); |
d7e09d03 PT |
541 | } |
542 | ||
543 | h = container_of0(hnode, struct lu_object_header, loh_hash); | |
544 | if (likely(!lu_object_is_dying(h))) { | |
545 | cfs_hash_get(s->ls_obj_hash, hnode); | |
546 | lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_HIT); | |
6e580ab5 | 547 | if (!list_empty(&h->loh_lru)) { |
a0b8803a | 548 | list_del_init(&h->loh_lru); |
6e580ab5 | 549 | bkt->lsb_lru_len--; |
a0b8803a | 550 | lprocfs_counter_decr(s->ls_stats, LU_SS_LRU_LEN); |
6e580ab5 | 551 | } |
d7e09d03 PT |
552 | return lu_object_top(h); |
553 | } | |
554 | ||
555 | /* | |
556 | * Lookup found an object being destroyed this object cannot be | |
557 | * returned (to assure that references to dying objects are eventually | |
558 | * drained), and moreover, lookup has to wait until object is freed. | |
559 | */ | |
560 | ||
9e795d35 | 561 | init_waitqueue_entry(waiter, current); |
d7e09d03 PT |
562 | add_wait_queue(&bkt->lsb_marche_funebre, waiter); |
563 | set_current_state(TASK_UNINTERRUPTIBLE); | |
564 | lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_DEATH_RACE); | |
565 | return ERR_PTR(-EAGAIN); | |
566 | } | |
567 | ||
568 | /** | |
569 | * Search cache for an object with the fid \a f. If such object is found, | |
570 | * return it. Otherwise, create new object, insert it into cache and return | |
571 | * it. In any case, additional reference is acquired on the returned object. | |
572 | */ | |
5913ef5e SB |
573 | static struct lu_object *lu_object_find(const struct lu_env *env, |
574 | struct lu_device *dev, | |
575 | const struct lu_fid *f, | |
576 | const struct lu_object_conf *conf) | |
d7e09d03 PT |
577 | { |
578 | return lu_object_find_at(env, dev->ld_site->ls_top_dev, f, conf); | |
579 | } | |
d7e09d03 PT |
580 | |
581 | static struct lu_object *lu_object_new(const struct lu_env *env, | |
582 | struct lu_device *dev, | |
583 | const struct lu_fid *f, | |
584 | const struct lu_object_conf *conf) | |
585 | { | |
586 | struct lu_object *o; | |
6da6eabe | 587 | struct cfs_hash *hs; |
6ea510c1 | 588 | struct cfs_hash_bd bd; |
d7e09d03 PT |
589 | |
590 | o = lu_object_alloc(env, dev, f, conf); | |
7f44cb0b | 591 | if (IS_ERR(o)) |
d7e09d03 PT |
592 | return o; |
593 | ||
594 | hs = dev->ld_site->ls_obj_hash; | |
595 | cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1); | |
d7e09d03 | 596 | cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash); |
d7e09d03 PT |
597 | cfs_hash_bd_unlock(hs, &bd, 1); |
598 | return o; | |
599 | } | |
600 | ||
601 | /** | |
602 | * Core logic of lu_object_find*() functions. | |
603 | */ | |
604 | static struct lu_object *lu_object_find_try(const struct lu_env *env, | |
605 | struct lu_device *dev, | |
606 | const struct lu_fid *f, | |
607 | const struct lu_object_conf *conf, | |
608 | wait_queue_t *waiter) | |
609 | { | |
610 | struct lu_object *o; | |
611 | struct lu_object *shadow; | |
612 | struct lu_site *s; | |
6da6eabe | 613 | struct cfs_hash *hs; |
6ea510c1 | 614 | struct cfs_hash_bd bd; |
d7e09d03 PT |
615 | __u64 version = 0; |
616 | ||
617 | /* | |
618 | * This uses standard index maintenance protocol: | |
619 | * | |
620 | * - search index under lock, and return object if found; | |
621 | * - otherwise, unlock index, allocate new object; | |
622 | * - lock index and search again; | |
623 | * - if nothing is found (usual case), insert newly created | |
624 | * object into index; | |
625 | * - otherwise (race: other thread inserted object), free | |
626 | * object just allocated. | |
627 | * - unlock index; | |
628 | * - return object. | |
629 | * | |
630 | * For "LOC_F_NEW" case, we are sure the object is new established. | |
631 | * It is unnecessary to perform lookup-alloc-lookup-insert, instead, | |
632 | * just alloc and insert directly. | |
633 | * | |
634 | * If dying object is found during index search, add @waiter to the | |
635 | * site wait-queue and return ERR_PTR(-EAGAIN). | |
636 | */ | |
cce3c2da | 637 | if (conf && conf->loc_flags & LOC_F_NEW) |
d7e09d03 PT |
638 | return lu_object_new(env, dev, f, conf); |
639 | ||
640 | s = dev->ld_site; | |
641 | hs = s->ls_obj_hash; | |
642 | cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1); | |
643 | o = htable_lookup(s, &bd, f, waiter, &version); | |
644 | cfs_hash_bd_unlock(hs, &bd, 1); | |
70b749d4 | 645 | if (!IS_ERR(o) || PTR_ERR(o) != -ENOENT) |
d7e09d03 PT |
646 | return o; |
647 | ||
648 | /* | |
649 | * Allocate new object. This may result in rather complicated | |
650 | * operations, including fld queries, inode loading, etc. | |
651 | */ | |
652 | o = lu_object_alloc(env, dev, f, conf); | |
7f44cb0b | 653 | if (IS_ERR(o)) |
d7e09d03 PT |
654 | return o; |
655 | ||
656 | LASSERT(lu_fid_eq(lu_object_fid(o), f)); | |
657 | ||
658 | cfs_hash_bd_lock(hs, &bd, 1); | |
659 | ||
660 | shadow = htable_lookup(s, &bd, f, waiter, &version); | |
208bf770 | 661 | if (likely(PTR_ERR(shadow) == -ENOENT)) { |
d7e09d03 | 662 | cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash); |
d7e09d03 PT |
663 | cfs_hash_bd_unlock(hs, &bd, 1); |
664 | return o; | |
665 | } | |
666 | ||
667 | lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_RACE); | |
668 | cfs_hash_bd_unlock(hs, &bd, 1); | |
669 | lu_object_free(env, o); | |
670 | return shadow; | |
671 | } | |
672 | ||
673 | /** | |
674 | * Much like lu_object_find(), but top level device of object is specifically | |
675 | * \a dev rather than top level device of the site. This interface allows | |
676 | * objects of different "stacking" to be created within the same site. | |
677 | */ | |
678 | struct lu_object *lu_object_find_at(const struct lu_env *env, | |
679 | struct lu_device *dev, | |
680 | const struct lu_fid *f, | |
681 | const struct lu_object_conf *conf) | |
682 | { | |
683 | struct lu_site_bkt_data *bkt; | |
684 | struct lu_object *obj; | |
685 | wait_queue_t wait; | |
686 | ||
687 | while (1) { | |
688 | obj = lu_object_find_try(env, dev, f, conf, &wait); | |
689 | if (obj != ERR_PTR(-EAGAIN)) | |
690 | return obj; | |
691 | /* | |
692 | * lu_object_find_try() already added waiter into the | |
693 | * wait queue. | |
694 | */ | |
b3669a7f | 695 | schedule(); |
d7e09d03 PT |
696 | bkt = lu_site_bkt_from_fid(dev->ld_site, (void *)f); |
697 | remove_wait_queue(&bkt->lsb_marche_funebre, &wait); | |
698 | } | |
699 | } | |
700 | EXPORT_SYMBOL(lu_object_find_at); | |
701 | ||
702 | /** | |
703 | * Find object with given fid, and return its slice belonging to given device. | |
704 | */ | |
705 | struct lu_object *lu_object_find_slice(const struct lu_env *env, | |
706 | struct lu_device *dev, | |
707 | const struct lu_fid *f, | |
708 | const struct lu_object_conf *conf) | |
709 | { | |
710 | struct lu_object *top; | |
711 | struct lu_object *obj; | |
712 | ||
713 | top = lu_object_find(env, dev, f, conf); | |
714 | if (!IS_ERR(top)) { | |
715 | obj = lu_object_locate(top->lo_header, dev->ld_type); | |
cce3c2da | 716 | if (!obj) |
d7e09d03 PT |
717 | lu_object_put(env, top); |
718 | } else | |
719 | obj = top; | |
720 | return obj; | |
721 | } | |
722 | EXPORT_SYMBOL(lu_object_find_slice); | |
723 | ||
724 | /** | |
725 | * Global list of all device types. | |
726 | */ | |
727 | static LIST_HEAD(lu_device_types); | |
728 | ||
729 | int lu_device_type_init(struct lu_device_type *ldt) | |
730 | { | |
731 | int result = 0; | |
732 | ||
733 | INIT_LIST_HEAD(&ldt->ldt_linkage); | |
734 | if (ldt->ldt_ops->ldto_init) | |
735 | result = ldt->ldt_ops->ldto_init(ldt); | |
736 | if (result == 0) | |
737 | list_add(&ldt->ldt_linkage, &lu_device_types); | |
738 | return result; | |
739 | } | |
740 | EXPORT_SYMBOL(lu_device_type_init); | |
741 | ||
742 | void lu_device_type_fini(struct lu_device_type *ldt) | |
743 | { | |
744 | list_del_init(&ldt->ldt_linkage); | |
745 | if (ldt->ldt_ops->ldto_fini) | |
746 | ldt->ldt_ops->ldto_fini(ldt); | |
747 | } | |
748 | EXPORT_SYMBOL(lu_device_type_fini); | |
749 | ||
750 | void lu_types_stop(void) | |
751 | { | |
752 | struct lu_device_type *ldt; | |
753 | ||
754 | list_for_each_entry(ldt, &lu_device_types, ldt_linkage) { | |
755 | if (ldt->ldt_device_nr == 0 && ldt->ldt_ops->ldto_stop) | |
756 | ldt->ldt_ops->ldto_stop(ldt); | |
757 | } | |
758 | } | |
759 | EXPORT_SYMBOL(lu_types_stop); | |
760 | ||
761 | /** | |
762 | * Global list of all sites on this node | |
763 | */ | |
764 | static LIST_HEAD(lu_sites); | |
765 | static DEFINE_MUTEX(lu_sites_guard); | |
766 | ||
767 | /** | |
768 | * Global environment used by site shrinker. | |
769 | */ | |
770 | static struct lu_env lu_shrink_env; | |
771 | ||
772 | struct lu_site_print_arg { | |
773 | struct lu_env *lsp_env; | |
774 | void *lsp_cookie; | |
775 | lu_printer_t lsp_printer; | |
776 | }; | |
777 | ||
778 | static int | |
6da6eabe | 779 | lu_site_obj_print(struct cfs_hash *hs, struct cfs_hash_bd *bd, |
d7e09d03 PT |
780 | struct hlist_node *hnode, void *data) |
781 | { | |
782 | struct lu_site_print_arg *arg = (struct lu_site_print_arg *)data; | |
783 | struct lu_object_header *h; | |
784 | ||
785 | h = hlist_entry(hnode, struct lu_object_header, loh_hash); | |
786 | if (!list_empty(&h->loh_layers)) { | |
787 | const struct lu_object *o; | |
788 | ||
789 | o = lu_object_top(h); | |
790 | lu_object_print(arg->lsp_env, arg->lsp_cookie, | |
791 | arg->lsp_printer, o); | |
792 | } else { | |
793 | lu_object_header_print(arg->lsp_env, arg->lsp_cookie, | |
794 | arg->lsp_printer, h); | |
795 | } | |
796 | return 0; | |
797 | } | |
798 | ||
799 | /** | |
800 | * Print all objects in \a s. | |
801 | */ | |
802 | void lu_site_print(const struct lu_env *env, struct lu_site *s, void *cookie, | |
803 | lu_printer_t printer) | |
804 | { | |
805 | struct lu_site_print_arg arg = { | |
806 | .lsp_env = (struct lu_env *)env, | |
807 | .lsp_cookie = cookie, | |
808 | .lsp_printer = printer, | |
809 | }; | |
810 | ||
811 | cfs_hash_for_each(s->ls_obj_hash, lu_site_obj_print, &arg); | |
812 | } | |
813 | EXPORT_SYMBOL(lu_site_print); | |
814 | ||
815 | enum { | |
816 | LU_CACHE_PERCENT_MAX = 50, | |
817 | LU_CACHE_PERCENT_DEFAULT = 20 | |
818 | }; | |
819 | ||
820 | static unsigned int lu_cache_percent = LU_CACHE_PERCENT_DEFAULT; | |
8cc7b4b9 PT |
821 | module_param(lu_cache_percent, int, 0644); |
822 | MODULE_PARM_DESC(lu_cache_percent, "Percentage of memory to be used as lu_object cache"); | |
d7e09d03 PT |
823 | |
824 | /** | |
825 | * Return desired hash table order. | |
826 | */ | |
827 | static int lu_htable_order(void) | |
828 | { | |
829 | unsigned long cache_size; | |
830 | int bits; | |
831 | ||
832 | /* | |
833 | * Calculate hash table size, assuming that we want reasonable | |
834 | * performance when 20% of total memory is occupied by cache of | |
835 | * lu_objects. | |
836 | * | |
837 | * Size of lu_object is (arbitrary) taken as 1K (together with inode). | |
838 | */ | |
4f6cc9ab | 839 | cache_size = totalram_pages; |
d7e09d03 PT |
840 | |
841 | #if BITS_PER_LONG == 32 | |
842 | /* limit hashtable size for lowmem systems to low RAM */ | |
843 | if (cache_size > 1 << (30 - PAGE_CACHE_SHIFT)) | |
844 | cache_size = 1 << (30 - PAGE_CACHE_SHIFT) * 3 / 4; | |
845 | #endif | |
846 | ||
847 | /* clear off unreasonable cache setting. */ | |
848 | if (lu_cache_percent == 0 || lu_cache_percent > LU_CACHE_PERCENT_MAX) { | |
2d00bd17 | 849 | CWARN("obdclass: invalid lu_cache_percent: %u, it must be in the range of (0, %u]. Will use default value: %u.\n", |
d7e09d03 PT |
850 | lu_cache_percent, LU_CACHE_PERCENT_MAX, |
851 | LU_CACHE_PERCENT_DEFAULT); | |
852 | ||
853 | lu_cache_percent = LU_CACHE_PERCENT_DEFAULT; | |
854 | } | |
855 | cache_size = cache_size / 100 * lu_cache_percent * | |
856 | (PAGE_CACHE_SIZE / 1024); | |
857 | ||
858 | for (bits = 1; (1 << bits) < cache_size; ++bits) { | |
859 | ; | |
860 | } | |
861 | return bits; | |
862 | } | |
863 | ||
6da6eabe | 864 | static unsigned lu_obj_hop_hash(struct cfs_hash *hs, |
d7e09d03 PT |
865 | const void *key, unsigned mask) |
866 | { | |
867 | struct lu_fid *fid = (struct lu_fid *)key; | |
868 | __u32 hash; | |
869 | ||
870 | hash = fid_flatten32(fid); | |
871 | hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */ | |
72c0824a | 872 | hash = hash_long(hash, hs->hs_bkt_bits); |
d7e09d03 PT |
873 | |
874 | /* give me another random factor */ | |
72c0824a | 875 | hash -= hash_long((unsigned long)hs, fid_oid(fid) % 11 + 3); |
d7e09d03 PT |
876 | |
877 | hash <<= hs->hs_cur_bits - hs->hs_bkt_bits; | |
878 | hash |= (fid_seq(fid) + fid_oid(fid)) & (CFS_HASH_NBKT(hs) - 1); | |
879 | ||
880 | return hash & mask; | |
881 | } | |
882 | ||
883 | static void *lu_obj_hop_object(struct hlist_node *hnode) | |
884 | { | |
885 | return hlist_entry(hnode, struct lu_object_header, loh_hash); | |
886 | } | |
887 | ||
888 | static void *lu_obj_hop_key(struct hlist_node *hnode) | |
889 | { | |
890 | struct lu_object_header *h; | |
891 | ||
892 | h = hlist_entry(hnode, struct lu_object_header, loh_hash); | |
893 | return &h->loh_fid; | |
894 | } | |
895 | ||
896 | static int lu_obj_hop_keycmp(const void *key, struct hlist_node *hnode) | |
897 | { | |
898 | struct lu_object_header *h; | |
899 | ||
900 | h = hlist_entry(hnode, struct lu_object_header, loh_hash); | |
901 | return lu_fid_eq(&h->loh_fid, (struct lu_fid *)key); | |
902 | } | |
903 | ||
6da6eabe | 904 | static void lu_obj_hop_get(struct cfs_hash *hs, struct hlist_node *hnode) |
d7e09d03 PT |
905 | { |
906 | struct lu_object_header *h; | |
907 | ||
908 | h = hlist_entry(hnode, struct lu_object_header, loh_hash); | |
6e580ab5 | 909 | atomic_inc(&h->loh_ref); |
d7e09d03 PT |
910 | } |
911 | ||
6da6eabe | 912 | static void lu_obj_hop_put_locked(struct cfs_hash *hs, struct hlist_node *hnode) |
d7e09d03 PT |
913 | { |
914 | LBUG(); /* we should never called it */ | |
915 | } | |
916 | ||
fb7a0201 | 917 | static struct cfs_hash_ops lu_site_hash_ops = { |
d7e09d03 | 918 | .hs_hash = lu_obj_hop_hash, |
db9fc06b | 919 | .hs_key = lu_obj_hop_key, |
d7e09d03 PT |
920 | .hs_keycmp = lu_obj_hop_keycmp, |
921 | .hs_object = lu_obj_hop_object, | |
db9fc06b | 922 | .hs_get = lu_obj_hop_get, |
d7e09d03 PT |
923 | .hs_put_locked = lu_obj_hop_put_locked, |
924 | }; | |
925 | ||
5913ef5e | 926 | static void lu_dev_add_linkage(struct lu_site *s, struct lu_device *d) |
d7e09d03 PT |
927 | { |
928 | spin_lock(&s->ls_ld_lock); | |
929 | if (list_empty(&d->ld_linkage)) | |
930 | list_add(&d->ld_linkage, &s->ls_ld_linkage); | |
931 | spin_unlock(&s->ls_ld_lock); | |
932 | } | |
d7e09d03 | 933 | |
d7e09d03 PT |
934 | /** |
935 | * Initialize site \a s, with \a d as the top level device. | |
936 | */ | |
937 | #define LU_SITE_BITS_MIN 12 | |
938 | #define LU_SITE_BITS_MAX 24 | |
939 | /** | |
940 | * total 256 buckets, we don't want too many buckets because: | |
941 | * - consume too much memory | |
942 | * - avoid unbalanced LRU list | |
943 | */ | |
944 | #define LU_SITE_BKT_BITS 8 | |
945 | ||
946 | int lu_site_init(struct lu_site *s, struct lu_device *top) | |
947 | { | |
948 | struct lu_site_bkt_data *bkt; | |
6ea510c1 | 949 | struct cfs_hash_bd bd; |
d7e09d03 PT |
950 | char name[16]; |
951 | int bits; | |
952 | int i; | |
d7e09d03 | 953 | |
ec83e611 | 954 | memset(s, 0, sizeof(*s)); |
d7e09d03 PT |
955 | bits = lu_htable_order(); |
956 | snprintf(name, 16, "lu_site_%s", top->ld_type->ldt_name); | |
957 | for (bits = min(max(LU_SITE_BITS_MIN, bits), LU_SITE_BITS_MAX); | |
958 | bits >= LU_SITE_BITS_MIN; bits--) { | |
959 | s->ls_obj_hash = cfs_hash_create(name, bits, bits, | |
960 | bits - LU_SITE_BKT_BITS, | |
961 | sizeof(*bkt), 0, 0, | |
962 | &lu_site_hash_ops, | |
963 | CFS_HASH_SPIN_BKTLOCK | | |
964 | CFS_HASH_NO_ITEMREF | | |
965 | CFS_HASH_DEPTH | | |
966 | CFS_HASH_ASSERT_EMPTY); | |
cce3c2da | 967 | if (s->ls_obj_hash) |
d7e09d03 PT |
968 | break; |
969 | } | |
970 | ||
cce3c2da | 971 | if (!s->ls_obj_hash) { |
d7e09d03 PT |
972 | CERROR("failed to create lu_site hash with bits: %d\n", bits); |
973 | return -ENOMEM; | |
974 | } | |
975 | ||
976 | cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) { | |
977 | bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd); | |
978 | INIT_LIST_HEAD(&bkt->lsb_lru); | |
979 | init_waitqueue_head(&bkt->lsb_marche_funebre); | |
980 | } | |
981 | ||
982 | s->ls_stats = lprocfs_alloc_stats(LU_SS_LAST_STAT, 0); | |
cce3c2da | 983 | if (!s->ls_stats) { |
d7e09d03 PT |
984 | cfs_hash_putref(s->ls_obj_hash); |
985 | s->ls_obj_hash = NULL; | |
986 | return -ENOMEM; | |
987 | } | |
988 | ||
989 | lprocfs_counter_init(s->ls_stats, LU_SS_CREATED, | |
990 | 0, "created", "created"); | |
991 | lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_HIT, | |
992 | 0, "cache_hit", "cache_hit"); | |
993 | lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_MISS, | |
994 | 0, "cache_miss", "cache_miss"); | |
995 | lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_RACE, | |
996 | 0, "cache_race", "cache_race"); | |
997 | lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_DEATH_RACE, | |
998 | 0, "cache_death_race", "cache_death_race"); | |
999 | lprocfs_counter_init(s->ls_stats, LU_SS_LRU_PURGED, | |
1000 | 0, "lru_purged", "lru_purged"); | |
a0b8803a AK |
1001 | /* |
1002 | * Unlike other counters, lru_len can be decremented so | |
1003 | * need lc_sum instead of just lc_count | |
1004 | */ | |
1005 | lprocfs_counter_init(s->ls_stats, LU_SS_LRU_LEN, | |
1006 | LPROCFS_CNTR_AVGMINMAX, "lru_len", "lru_len"); | |
d7e09d03 PT |
1007 | |
1008 | INIT_LIST_HEAD(&s->ls_linkage); | |
1009 | s->ls_top_dev = top; | |
1010 | top->ld_site = s; | |
1011 | lu_device_get(top); | |
1012 | lu_ref_add(&top->ld_reference, "site-top", s); | |
1013 | ||
1014 | INIT_LIST_HEAD(&s->ls_ld_linkage); | |
1015 | spin_lock_init(&s->ls_ld_lock); | |
1016 | ||
1017 | lu_dev_add_linkage(s, top); | |
1018 | ||
0a3bdb00 | 1019 | return 0; |
d7e09d03 PT |
1020 | } |
1021 | EXPORT_SYMBOL(lu_site_init); | |
1022 | ||
1023 | /** | |
1024 | * Finalize \a s and release its resources. | |
1025 | */ | |
1026 | void lu_site_fini(struct lu_site *s) | |
1027 | { | |
1028 | mutex_lock(&lu_sites_guard); | |
1029 | list_del_init(&s->ls_linkage); | |
1030 | mutex_unlock(&lu_sites_guard); | |
1031 | ||
cce3c2da | 1032 | if (s->ls_obj_hash) { |
d7e09d03 PT |
1033 | cfs_hash_putref(s->ls_obj_hash); |
1034 | s->ls_obj_hash = NULL; | |
1035 | } | |
1036 | ||
cce3c2da | 1037 | if (s->ls_top_dev) { |
d7e09d03 PT |
1038 | s->ls_top_dev->ld_site = NULL; |
1039 | lu_ref_del(&s->ls_top_dev->ld_reference, "site-top", s); | |
1040 | lu_device_put(s->ls_top_dev); | |
1041 | s->ls_top_dev = NULL; | |
1042 | } | |
1043 | ||
cce3c2da | 1044 | if (s->ls_stats) |
d7e09d03 PT |
1045 | lprocfs_free_stats(&s->ls_stats); |
1046 | } | |
1047 | EXPORT_SYMBOL(lu_site_fini); | |
1048 | ||
1049 | /** | |
1050 | * Called when initialization of stack for this site is completed. | |
1051 | */ | |
1052 | int lu_site_init_finish(struct lu_site *s) | |
1053 | { | |
1054 | int result; | |
50ffcb7e | 1055 | |
d7e09d03 PT |
1056 | mutex_lock(&lu_sites_guard); |
1057 | result = lu_context_refill(&lu_shrink_env.le_ctx); | |
1058 | if (result == 0) | |
1059 | list_add(&s->ls_linkage, &lu_sites); | |
1060 | mutex_unlock(&lu_sites_guard); | |
1061 | return result; | |
1062 | } | |
1063 | EXPORT_SYMBOL(lu_site_init_finish); | |
1064 | ||
1065 | /** | |
1066 | * Acquire additional reference on device \a d | |
1067 | */ | |
1068 | void lu_device_get(struct lu_device *d) | |
1069 | { | |
1070 | atomic_inc(&d->ld_ref); | |
1071 | } | |
1072 | EXPORT_SYMBOL(lu_device_get); | |
1073 | ||
1074 | /** | |
1075 | * Release reference on device \a d. | |
1076 | */ | |
1077 | void lu_device_put(struct lu_device *d) | |
1078 | { | |
1079 | LASSERT(atomic_read(&d->ld_ref) > 0); | |
1080 | atomic_dec(&d->ld_ref); | |
1081 | } | |
1082 | EXPORT_SYMBOL(lu_device_put); | |
1083 | ||
1084 | /** | |
1085 | * Initialize device \a d of type \a t. | |
1086 | */ | |
1087 | int lu_device_init(struct lu_device *d, struct lu_device_type *t) | |
1088 | { | |
cce3c2da | 1089 | if (t->ldt_device_nr++ == 0 && t->ldt_ops->ldto_start) |
d7e09d03 | 1090 | t->ldt_ops->ldto_start(t); |
ec83e611 | 1091 | memset(d, 0, sizeof(*d)); |
d7e09d03 PT |
1092 | atomic_set(&d->ld_ref, 0); |
1093 | d->ld_type = t; | |
1094 | lu_ref_init(&d->ld_reference); | |
1095 | INIT_LIST_HEAD(&d->ld_linkage); | |
1096 | return 0; | |
1097 | } | |
1098 | EXPORT_SYMBOL(lu_device_init); | |
1099 | ||
1100 | /** | |
1101 | * Finalize device \a d. | |
1102 | */ | |
1103 | void lu_device_fini(struct lu_device *d) | |
1104 | { | |
1105 | struct lu_device_type *t; | |
1106 | ||
1107 | t = d->ld_type; | |
cce3c2da | 1108 | if (d->ld_obd) { |
d7e09d03 PT |
1109 | d->ld_obd->obd_lu_dev = NULL; |
1110 | d->ld_obd = NULL; | |
1111 | } | |
1112 | ||
1113 | lu_ref_fini(&d->ld_reference); | |
1114 | LASSERTF(atomic_read(&d->ld_ref) == 0, | |
1115 | "Refcount is %u\n", atomic_read(&d->ld_ref)); | |
1116 | LASSERT(t->ldt_device_nr > 0); | |
cce3c2da | 1117 | if (--t->ldt_device_nr == 0 && t->ldt_ops->ldto_stop) |
d7e09d03 PT |
1118 | t->ldt_ops->ldto_stop(t); |
1119 | } | |
1120 | EXPORT_SYMBOL(lu_device_fini); | |
1121 | ||
1122 | /** | |
1123 | * Initialize object \a o that is part of compound object \a h and was created | |
1124 | * by device \a d. | |
1125 | */ | |
631abc6e JH |
1126 | int lu_object_init(struct lu_object *o, struct lu_object_header *h, |
1127 | struct lu_device *d) | |
d7e09d03 | 1128 | { |
631abc6e | 1129 | memset(o, 0, sizeof(*o)); |
d7e09d03 | 1130 | o->lo_header = h; |
631abc6e | 1131 | o->lo_dev = d; |
d7e09d03 | 1132 | lu_device_get(d); |
631abc6e | 1133 | lu_ref_add_at(&d->ld_reference, &o->lo_dev_ref, "lu_object", o); |
d7e09d03 | 1134 | INIT_LIST_HEAD(&o->lo_linkage); |
631abc6e | 1135 | |
d7e09d03 PT |
1136 | return 0; |
1137 | } | |
1138 | EXPORT_SYMBOL(lu_object_init); | |
1139 | ||
1140 | /** | |
1141 | * Finalize object and release its resources. | |
1142 | */ | |
1143 | void lu_object_fini(struct lu_object *o) | |
1144 | { | |
1145 | struct lu_device *dev = o->lo_dev; | |
1146 | ||
1147 | LASSERT(list_empty(&o->lo_linkage)); | |
1148 | ||
cce3c2da | 1149 | if (dev) { |
631abc6e JH |
1150 | lu_ref_del_at(&dev->ld_reference, &o->lo_dev_ref, |
1151 | "lu_object", o); | |
d7e09d03 PT |
1152 | lu_device_put(dev); |
1153 | o->lo_dev = NULL; | |
1154 | } | |
1155 | } | |
1156 | EXPORT_SYMBOL(lu_object_fini); | |
1157 | ||
1158 | /** | |
1159 | * Add object \a o as first layer of compound object \a h | |
1160 | * | |
1161 | * This is typically called by the ->ldo_object_alloc() method of top-level | |
1162 | * device. | |
1163 | */ | |
1164 | void lu_object_add_top(struct lu_object_header *h, struct lu_object *o) | |
1165 | { | |
1166 | list_move(&o->lo_linkage, &h->loh_layers); | |
1167 | } | |
1168 | EXPORT_SYMBOL(lu_object_add_top); | |
1169 | ||
1170 | /** | |
1171 | * Add object \a o as a layer of compound object, going after \a before. | |
1172 | * | |
1173 | * This is typically called by the ->ldo_object_alloc() method of \a | |
1174 | * before->lo_dev. | |
1175 | */ | |
1176 | void lu_object_add(struct lu_object *before, struct lu_object *o) | |
1177 | { | |
1178 | list_move(&o->lo_linkage, &before->lo_linkage); | |
1179 | } | |
1180 | EXPORT_SYMBOL(lu_object_add); | |
1181 | ||
1182 | /** | |
1183 | * Initialize compound object. | |
1184 | */ | |
1185 | int lu_object_header_init(struct lu_object_header *h) | |
1186 | { | |
ec83e611 | 1187 | memset(h, 0, sizeof(*h)); |
d7e09d03 PT |
1188 | atomic_set(&h->loh_ref, 1); |
1189 | INIT_HLIST_NODE(&h->loh_hash); | |
1190 | INIT_LIST_HEAD(&h->loh_lru); | |
1191 | INIT_LIST_HEAD(&h->loh_layers); | |
1192 | lu_ref_init(&h->loh_reference); | |
1193 | return 0; | |
1194 | } | |
1195 | EXPORT_SYMBOL(lu_object_header_init); | |
1196 | ||
1197 | /** | |
1198 | * Finalize compound object. | |
1199 | */ | |
1200 | void lu_object_header_fini(struct lu_object_header *h) | |
1201 | { | |
1202 | LASSERT(list_empty(&h->loh_layers)); | |
1203 | LASSERT(list_empty(&h->loh_lru)); | |
1204 | LASSERT(hlist_unhashed(&h->loh_hash)); | |
1205 | lu_ref_fini(&h->loh_reference); | |
1206 | } | |
1207 | EXPORT_SYMBOL(lu_object_header_fini); | |
1208 | ||
1209 | /** | |
1210 | * Given a compound object, find its slice, corresponding to the device type | |
1211 | * \a dtype. | |
1212 | */ | |
1213 | struct lu_object *lu_object_locate(struct lu_object_header *h, | |
1214 | const struct lu_device_type *dtype) | |
1215 | { | |
1216 | struct lu_object *o; | |
1217 | ||
1218 | list_for_each_entry(o, &h->loh_layers, lo_linkage) { | |
1219 | if (o->lo_dev->ld_type == dtype) | |
1220 | return o; | |
1221 | } | |
1222 | return NULL; | |
1223 | } | |
1224 | EXPORT_SYMBOL(lu_object_locate); | |
1225 | ||
d7e09d03 PT |
1226 | /** |
1227 | * Finalize and free devices in the device stack. | |
1228 | * | |
1229 | * Finalize device stack by purging object cache, and calling | |
1230 | * lu_device_type_operations::ldto_device_fini() and | |
1231 | * lu_device_type_operations::ldto_device_free() on all devices in the stack. | |
1232 | */ | |
1233 | void lu_stack_fini(const struct lu_env *env, struct lu_device *top) | |
1234 | { | |
1235 | struct lu_site *site = top->ld_site; | |
1236 | struct lu_device *scan; | |
1237 | struct lu_device *next; | |
1238 | ||
1239 | lu_site_purge(env, site, ~0); | |
cce3c2da | 1240 | for (scan = top; scan; scan = next) { |
d7e09d03 PT |
1241 | next = scan->ld_type->ldt_ops->ldto_device_fini(env, scan); |
1242 | lu_ref_del(&scan->ld_reference, "lu-stack", &lu_site_init); | |
1243 | lu_device_put(scan); | |
1244 | } | |
1245 | ||
1246 | /* purge again. */ | |
1247 | lu_site_purge(env, site, ~0); | |
1248 | ||
cce3c2da | 1249 | for (scan = top; scan; scan = next) { |
d7e09d03 PT |
1250 | const struct lu_device_type *ldt = scan->ld_type; |
1251 | struct obd_type *type; | |
1252 | ||
1253 | next = ldt->ldt_ops->ldto_device_free(env, scan); | |
1254 | type = ldt->ldt_obd_type; | |
cce3c2da | 1255 | if (type) { |
d7e09d03 PT |
1256 | type->typ_refcnt--; |
1257 | class_put_type(type); | |
1258 | } | |
1259 | } | |
1260 | } | |
1261 | EXPORT_SYMBOL(lu_stack_fini); | |
1262 | ||
1263 | enum { | |
1264 | /** | |
1265 | * Maximal number of tld slots. | |
1266 | */ | |
1267 | LU_CONTEXT_KEY_NR = 40 | |
1268 | }; | |
1269 | ||
1270 | static struct lu_context_key *lu_keys[LU_CONTEXT_KEY_NR] = { NULL, }; | |
1271 | ||
1272 | static DEFINE_SPINLOCK(lu_keys_guard); | |
1273 | ||
1274 | /** | |
1275 | * Global counter incremented whenever key is registered, unregistered, | |
1276 | * revived or quiesced. This is used to void unnecessary calls to | |
1277 | * lu_context_refill(). No locking is provided, as initialization and shutdown | |
1278 | * are supposed to be externally serialized. | |
1279 | */ | |
225f597c | 1280 | static unsigned key_set_version; |
d7e09d03 PT |
1281 | |
1282 | /** | |
1283 | * Register new key. | |
1284 | */ | |
1285 | int lu_context_key_register(struct lu_context_key *key) | |
1286 | { | |
1287 | int result; | |
1288 | int i; | |
1289 | ||
cce3c2da OD |
1290 | LASSERT(key->lct_init); |
1291 | LASSERT(key->lct_fini); | |
d7e09d03 | 1292 | LASSERT(key->lct_tags != 0); |
d7e09d03 PT |
1293 | |
1294 | result = -ENFILE; | |
1295 | spin_lock(&lu_keys_guard); | |
1296 | for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) { | |
cce3c2da | 1297 | if (!lu_keys[i]) { |
d7e09d03 PT |
1298 | key->lct_index = i; |
1299 | atomic_set(&key->lct_used, 1); | |
1300 | lu_keys[i] = key; | |
1301 | lu_ref_init(&key->lct_reference); | |
1302 | result = 0; | |
1303 | ++key_set_version; | |
1304 | break; | |
1305 | } | |
1306 | } | |
1307 | spin_unlock(&lu_keys_guard); | |
1308 | return result; | |
1309 | } | |
1310 | EXPORT_SYMBOL(lu_context_key_register); | |
1311 | ||
1312 | static void key_fini(struct lu_context *ctx, int index) | |
1313 | { | |
cce3c2da | 1314 | if (ctx->lc_value && ctx->lc_value[index]) { |
d7e09d03 PT |
1315 | struct lu_context_key *key; |
1316 | ||
1317 | key = lu_keys[index]; | |
d7e09d03 PT |
1318 | LASSERT(atomic_read(&key->lct_used) > 1); |
1319 | ||
1320 | key->lct_fini(ctx, key, ctx->lc_value[index]); | |
1321 | lu_ref_del(&key->lct_reference, "ctx", ctx); | |
1322 | atomic_dec(&key->lct_used); | |
1323 | ||
d7e09d03 | 1324 | if ((ctx->lc_tags & LCT_NOREF) == 0) { |
4a1a01ea | 1325 | #ifdef CONFIG_MODULE_UNLOAD |
d7e09d03 | 1326 | LINVRNT(module_refcount(key->lct_owner) > 0); |
4a1a01ea | 1327 | #endif |
d7e09d03 PT |
1328 | module_put(key->lct_owner); |
1329 | } | |
1330 | ctx->lc_value[index] = NULL; | |
1331 | } | |
1332 | } | |
1333 | ||
1334 | /** | |
1335 | * Deregister key. | |
1336 | */ | |
1337 | void lu_context_key_degister(struct lu_context_key *key) | |
1338 | { | |
1339 | LASSERT(atomic_read(&key->lct_used) >= 1); | |
1340 | LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys)); | |
1341 | ||
1342 | lu_context_key_quiesce(key); | |
1343 | ||
1344 | ++key_set_version; | |
1345 | spin_lock(&lu_keys_guard); | |
1346 | key_fini(&lu_shrink_env.le_ctx, key->lct_index); | |
1347 | if (lu_keys[key->lct_index]) { | |
1348 | lu_keys[key->lct_index] = NULL; | |
1349 | lu_ref_fini(&key->lct_reference); | |
1350 | } | |
1351 | spin_unlock(&lu_keys_guard); | |
1352 | ||
1353 | LASSERTF(atomic_read(&key->lct_used) == 1, | |
1354 | "key has instances: %d\n", | |
1355 | atomic_read(&key->lct_used)); | |
1356 | } | |
1357 | EXPORT_SYMBOL(lu_context_key_degister); | |
1358 | ||
1359 | /** | |
1360 | * Register a number of keys. This has to be called after all keys have been | |
1361 | * initialized by a call to LU_CONTEXT_KEY_INIT(). | |
1362 | */ | |
1363 | int lu_context_key_register_many(struct lu_context_key *k, ...) | |
1364 | { | |
1365 | struct lu_context_key *key = k; | |
1366 | va_list args; | |
1367 | int result; | |
1368 | ||
1369 | va_start(args, k); | |
1370 | do { | |
1371 | result = lu_context_key_register(key); | |
1372 | if (result) | |
1373 | break; | |
1374 | key = va_arg(args, struct lu_context_key *); | |
cce3c2da | 1375 | } while (key); |
d7e09d03 PT |
1376 | va_end(args); |
1377 | ||
1378 | if (result != 0) { | |
1379 | va_start(args, k); | |
1380 | while (k != key) { | |
1381 | lu_context_key_degister(k); | |
1382 | k = va_arg(args, struct lu_context_key *); | |
1383 | } | |
1384 | va_end(args); | |
1385 | } | |
1386 | ||
1387 | return result; | |
1388 | } | |
1389 | EXPORT_SYMBOL(lu_context_key_register_many); | |
1390 | ||
1391 | /** | |
1392 | * De-register a number of keys. This is a dual to | |
1393 | * lu_context_key_register_many(). | |
1394 | */ | |
1395 | void lu_context_key_degister_many(struct lu_context_key *k, ...) | |
1396 | { | |
1397 | va_list args; | |
1398 | ||
1399 | va_start(args, k); | |
1400 | do { | |
1401 | lu_context_key_degister(k); | |
1402 | k = va_arg(args, struct lu_context_key*); | |
cce3c2da | 1403 | } while (k); |
d7e09d03 PT |
1404 | va_end(args); |
1405 | } | |
1406 | EXPORT_SYMBOL(lu_context_key_degister_many); | |
1407 | ||
1408 | /** | |
1409 | * Revive a number of keys. | |
1410 | */ | |
1411 | void lu_context_key_revive_many(struct lu_context_key *k, ...) | |
1412 | { | |
1413 | va_list args; | |
1414 | ||
1415 | va_start(args, k); | |
1416 | do { | |
1417 | lu_context_key_revive(k); | |
1418 | k = va_arg(args, struct lu_context_key*); | |
cce3c2da | 1419 | } while (k); |
d7e09d03 PT |
1420 | va_end(args); |
1421 | } | |
1422 | EXPORT_SYMBOL(lu_context_key_revive_many); | |
1423 | ||
1424 | /** | |
1425 | * Quiescent a number of keys. | |
1426 | */ | |
1427 | void lu_context_key_quiesce_many(struct lu_context_key *k, ...) | |
1428 | { | |
1429 | va_list args; | |
1430 | ||
1431 | va_start(args, k); | |
1432 | do { | |
1433 | lu_context_key_quiesce(k); | |
1434 | k = va_arg(args, struct lu_context_key*); | |
cce3c2da | 1435 | } while (k); |
d7e09d03 PT |
1436 | va_end(args); |
1437 | } | |
1438 | EXPORT_SYMBOL(lu_context_key_quiesce_many); | |
1439 | ||
1440 | /** | |
1441 | * Return value associated with key \a key in context \a ctx. | |
1442 | */ | |
1443 | void *lu_context_key_get(const struct lu_context *ctx, | |
1444 | const struct lu_context_key *key) | |
1445 | { | |
1446 | LINVRNT(ctx->lc_state == LCS_ENTERED); | |
1447 | LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys)); | |
1448 | LASSERT(lu_keys[key->lct_index] == key); | |
1449 | return ctx->lc_value[key->lct_index]; | |
1450 | } | |
1451 | EXPORT_SYMBOL(lu_context_key_get); | |
1452 | ||
1453 | /** | |
1454 | * List of remembered contexts. XXX document me. | |
1455 | */ | |
1456 | static LIST_HEAD(lu_context_remembered); | |
1457 | ||
1458 | /** | |
1459 | * Destroy \a key in all remembered contexts. This is used to destroy key | |
1460 | * values in "shared" contexts (like service threads), when a module owning | |
1461 | * the key is about to be unloaded. | |
1462 | */ | |
1463 | void lu_context_key_quiesce(struct lu_context_key *key) | |
1464 | { | |
1465 | struct lu_context *ctx; | |
1466 | ||
1467 | if (!(key->lct_tags & LCT_QUIESCENT)) { | |
1468 | /* | |
1469 | * XXX layering violation. | |
1470 | */ | |
1471 | key->lct_tags |= LCT_QUIESCENT; | |
1472 | /* | |
1473 | * XXX memory barrier has to go here. | |
1474 | */ | |
1475 | spin_lock(&lu_keys_guard); | |
926d6fb2 | 1476 | list_for_each_entry(ctx, &lu_context_remembered, lc_remember) |
d7e09d03 PT |
1477 | key_fini(ctx, key->lct_index); |
1478 | spin_unlock(&lu_keys_guard); | |
1479 | ++key_set_version; | |
1480 | } | |
1481 | } | |
1482 | EXPORT_SYMBOL(lu_context_key_quiesce); | |
1483 | ||
1484 | void lu_context_key_revive(struct lu_context_key *key) | |
1485 | { | |
1486 | key->lct_tags &= ~LCT_QUIESCENT; | |
1487 | ++key_set_version; | |
1488 | } | |
1489 | EXPORT_SYMBOL(lu_context_key_revive); | |
1490 | ||
1491 | static void keys_fini(struct lu_context *ctx) | |
1492 | { | |
1493 | int i; | |
1494 | ||
cce3c2da | 1495 | if (!ctx->lc_value) |
d7e09d03 PT |
1496 | return; |
1497 | ||
1498 | for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) | |
1499 | key_fini(ctx, i); | |
1500 | ||
d7279044 | 1501 | kfree(ctx->lc_value); |
d7e09d03 PT |
1502 | ctx->lc_value = NULL; |
1503 | } | |
1504 | ||
1505 | static int keys_fill(struct lu_context *ctx) | |
1506 | { | |
1507 | int i; | |
1508 | ||
cce3c2da | 1509 | LINVRNT(ctx->lc_value); |
d7e09d03 PT |
1510 | for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) { |
1511 | struct lu_context_key *key; | |
1512 | ||
1513 | key = lu_keys[i]; | |
cce3c2da | 1514 | if (!ctx->lc_value[i] && key && |
d7e09d03 PT |
1515 | (key->lct_tags & ctx->lc_tags) && |
1516 | /* | |
1517 | * Don't create values for a LCT_QUIESCENT key, as this | |
1518 | * will pin module owning a key. | |
1519 | */ | |
1520 | !(key->lct_tags & LCT_QUIESCENT)) { | |
1521 | void *value; | |
1522 | ||
cce3c2da | 1523 | LINVRNT(key->lct_init); |
d7e09d03 PT |
1524 | LINVRNT(key->lct_index == i); |
1525 | ||
1526 | value = key->lct_init(ctx, key); | |
7f44cb0b | 1527 | if (IS_ERR(value)) |
d7e09d03 PT |
1528 | return PTR_ERR(value); |
1529 | ||
d7e09d03 PT |
1530 | if (!(ctx->lc_tags & LCT_NOREF)) |
1531 | try_module_get(key->lct_owner); | |
1532 | lu_ref_add_atomic(&key->lct_reference, "ctx", ctx); | |
1533 | atomic_inc(&key->lct_used); | |
1534 | /* | |
1535 | * This is the only place in the code, where an | |
1536 | * element of ctx->lc_value[] array is set to non-NULL | |
1537 | * value. | |
1538 | */ | |
1539 | ctx->lc_value[i] = value; | |
cce3c2da | 1540 | if (key->lct_exit) |
d7e09d03 PT |
1541 | ctx->lc_tags |= LCT_HAS_EXIT; |
1542 | } | |
1543 | ctx->lc_version = key_set_version; | |
1544 | } | |
1545 | return 0; | |
1546 | } | |
1547 | ||
1548 | static int keys_init(struct lu_context *ctx) | |
1549 | { | |
d7279044 JL |
1550 | ctx->lc_value = kcalloc(ARRAY_SIZE(lu_keys), sizeof(ctx->lc_value[0]), |
1551 | GFP_NOFS); | |
cce3c2da | 1552 | if (likely(ctx->lc_value)) |
d7e09d03 PT |
1553 | return keys_fill(ctx); |
1554 | ||
1555 | return -ENOMEM; | |
1556 | } | |
1557 | ||
1558 | /** | |
1559 | * Initialize context data-structure. Create values for all keys. | |
1560 | */ | |
1561 | int lu_context_init(struct lu_context *ctx, __u32 tags) | |
1562 | { | |
1563 | int rc; | |
1564 | ||
ec83e611 | 1565 | memset(ctx, 0, sizeof(*ctx)); |
d7e09d03 PT |
1566 | ctx->lc_state = LCS_INITIALIZED; |
1567 | ctx->lc_tags = tags; | |
1568 | if (tags & LCT_REMEMBER) { | |
1569 | spin_lock(&lu_keys_guard); | |
1570 | list_add(&ctx->lc_remember, &lu_context_remembered); | |
1571 | spin_unlock(&lu_keys_guard); | |
1572 | } else { | |
1573 | INIT_LIST_HEAD(&ctx->lc_remember); | |
1574 | } | |
1575 | ||
1576 | rc = keys_init(ctx); | |
1577 | if (rc != 0) | |
1578 | lu_context_fini(ctx); | |
1579 | ||
1580 | return rc; | |
1581 | } | |
1582 | EXPORT_SYMBOL(lu_context_init); | |
1583 | ||
1584 | /** | |
1585 | * Finalize context data-structure. Destroy key values. | |
1586 | */ | |
1587 | void lu_context_fini(struct lu_context *ctx) | |
1588 | { | |
1589 | LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT); | |
1590 | ctx->lc_state = LCS_FINALIZED; | |
1591 | ||
1592 | if ((ctx->lc_tags & LCT_REMEMBER) == 0) { | |
1593 | LASSERT(list_empty(&ctx->lc_remember)); | |
1594 | keys_fini(ctx); | |
1595 | ||
1596 | } else { /* could race with key degister */ | |
1597 | spin_lock(&lu_keys_guard); | |
1598 | keys_fini(ctx); | |
1599 | list_del_init(&ctx->lc_remember); | |
1600 | spin_unlock(&lu_keys_guard); | |
1601 | } | |
1602 | } | |
1603 | EXPORT_SYMBOL(lu_context_fini); | |
1604 | ||
1605 | /** | |
1606 | * Called before entering context. | |
1607 | */ | |
1608 | void lu_context_enter(struct lu_context *ctx) | |
1609 | { | |
1610 | LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT); | |
1611 | ctx->lc_state = LCS_ENTERED; | |
1612 | } | |
1613 | EXPORT_SYMBOL(lu_context_enter); | |
1614 | ||
1615 | /** | |
1616 | * Called after exiting from \a ctx | |
1617 | */ | |
1618 | void lu_context_exit(struct lu_context *ctx) | |
1619 | { | |
1620 | int i; | |
1621 | ||
1622 | LINVRNT(ctx->lc_state == LCS_ENTERED); | |
1623 | ctx->lc_state = LCS_LEFT; | |
cce3c2da | 1624 | if (ctx->lc_tags & LCT_HAS_EXIT && ctx->lc_value) { |
d7e09d03 | 1625 | for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) { |
cce3c2da | 1626 | if (ctx->lc_value[i]) { |
d7e09d03 PT |
1627 | struct lu_context_key *key; |
1628 | ||
1629 | key = lu_keys[i]; | |
cce3c2da | 1630 | if (key->lct_exit) |
d7e09d03 PT |
1631 | key->lct_exit(ctx, |
1632 | key, ctx->lc_value[i]); | |
1633 | } | |
1634 | } | |
1635 | } | |
1636 | } | |
1637 | EXPORT_SYMBOL(lu_context_exit); | |
1638 | ||
1639 | /** | |
1640 | * Allocate for context all missing keys that were registered after context | |
1641 | * creation. key_set_version is only changed in rare cases when modules | |
1642 | * are loaded and removed. | |
1643 | */ | |
1644 | int lu_context_refill(struct lu_context *ctx) | |
1645 | { | |
1646 | return likely(ctx->lc_version == key_set_version) ? 0 : keys_fill(ctx); | |
1647 | } | |
1648 | EXPORT_SYMBOL(lu_context_refill); | |
1649 | ||
1650 | /** | |
1651 | * lu_ctx_tags/lu_ses_tags will be updated if there are new types of | |
1652 | * obd being added. Currently, this is only used on client side, specifically | |
1653 | * for echo device client, for other stack (like ptlrpc threads), context are | |
1654 | * predefined when the lu_device type are registered, during the module probe | |
1655 | * phase. | |
1656 | */ | |
ea28d21a MR |
1657 | __u32 lu_context_tags_default; |
1658 | __u32 lu_session_tags_default; | |
d7e09d03 | 1659 | |
d7e09d03 PT |
1660 | int lu_env_init(struct lu_env *env, __u32 tags) |
1661 | { | |
1662 | int result; | |
1663 | ||
1664 | env->le_ses = NULL; | |
1665 | result = lu_context_init(&env->le_ctx, tags); | |
1666 | if (likely(result == 0)) | |
1667 | lu_context_enter(&env->le_ctx); | |
1668 | return result; | |
1669 | } | |
1670 | EXPORT_SYMBOL(lu_env_init); | |
1671 | ||
1672 | void lu_env_fini(struct lu_env *env) | |
1673 | { | |
1674 | lu_context_exit(&env->le_ctx); | |
1675 | lu_context_fini(&env->le_ctx); | |
1676 | env->le_ses = NULL; | |
1677 | } | |
1678 | EXPORT_SYMBOL(lu_env_fini); | |
1679 | ||
1680 | int lu_env_refill(struct lu_env *env) | |
1681 | { | |
1682 | int result; | |
1683 | ||
1684 | result = lu_context_refill(&env->le_ctx); | |
cce3c2da | 1685 | if (result == 0 && env->le_ses) |
d7e09d03 PT |
1686 | result = lu_context_refill(env->le_ses); |
1687 | return result; | |
1688 | } | |
1689 | EXPORT_SYMBOL(lu_env_refill); | |
1690 | ||
2de5855c | 1691 | struct lu_site_stats { |
d7e09d03 PT |
1692 | unsigned lss_populated; |
1693 | unsigned lss_max_search; | |
1694 | unsigned lss_total; | |
1695 | unsigned lss_busy; | |
2de5855c | 1696 | }; |
d7e09d03 | 1697 | |
6da6eabe | 1698 | static void lu_site_stats_get(struct cfs_hash *hs, |
2de5855c | 1699 | struct lu_site_stats *stats, int populated) |
d7e09d03 | 1700 | { |
6ea510c1 | 1701 | struct cfs_hash_bd bd; |
d7e09d03 PT |
1702 | int i; |
1703 | ||
1704 | cfs_hash_for_each_bucket(hs, &bd, i) { | |
1705 | struct lu_site_bkt_data *bkt = cfs_hash_bd_extra_get(hs, &bd); | |
1706 | struct hlist_head *hhead; | |
1707 | ||
1708 | cfs_hash_bd_lock(hs, &bd, 1); | |
6e580ab5 FZ |
1709 | stats->lss_busy += |
1710 | cfs_hash_bd_count_get(&bd) - bkt->lsb_lru_len; | |
d7e09d03 PT |
1711 | stats->lss_total += cfs_hash_bd_count_get(&bd); |
1712 | stats->lss_max_search = max((int)stats->lss_max_search, | |
1713 | cfs_hash_bd_depmax_get(&bd)); | |
1714 | if (!populated) { | |
1715 | cfs_hash_bd_unlock(hs, &bd, 1); | |
1716 | continue; | |
1717 | } | |
1718 | ||
1719 | cfs_hash_bd_for_each_hlist(hs, &bd, hhead) { | |
1720 | if (!hlist_empty(hhead)) | |
1721 | stats->lss_populated++; | |
1722 | } | |
1723 | cfs_hash_bd_unlock(hs, &bd, 1); | |
1724 | } | |
1725 | } | |
1726 | ||
d7e09d03 | 1727 | /* |
a0b8803a AK |
1728 | * lu_cache_shrink_count returns the number of cached objects that are |
1729 | * candidates to be freed by shrink_slab(). A counter, which tracks | |
1730 | * the number of items in the site's lru, is maintained in the per cpu | |
1731 | * stats of each site. The counter is incremented when an object is added | |
1732 | * to a site's lru and decremented when one is removed. The number of | |
1733 | * free-able objects is the sum of all per cpu counters for all sites. | |
d7e09d03 | 1734 | * |
a0b8803a AK |
1735 | * Using a per cpu counter is a compromise solution to concurrent access: |
1736 | * lu_object_put() can update the counter without locking the site and | |
1737 | * lu_cache_shrink_count can sum the counters without locking each | |
1738 | * ls_obj_hash bucket. | |
d7e09d03 | 1739 | */ |
fe92a055 PT |
1740 | static unsigned long lu_cache_shrink_count(struct shrinker *sk, |
1741 | struct shrink_control *sc) | |
d7e09d03 | 1742 | { |
d7e09d03 PT |
1743 | struct lu_site *s; |
1744 | struct lu_site *tmp; | |
fe92a055 | 1745 | unsigned long cached = 0; |
d7e09d03 | 1746 | |
fe92a055 PT |
1747 | if (!(sc->gfp_mask & __GFP_FS)) |
1748 | return 0; | |
d7e09d03 PT |
1749 | |
1750 | mutex_lock(&lu_sites_guard); | |
1751 | list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) { | |
a0b8803a | 1752 | cached += ls_stats_read(s->ls_stats, LU_SS_LRU_LEN); |
d7e09d03 | 1753 | } |
d7e09d03 PT |
1754 | mutex_unlock(&lu_sites_guard); |
1755 | ||
1756 | cached = (cached / 100) * sysctl_vfs_cache_pressure; | |
a0b8803a AK |
1757 | CDEBUG(D_INODE, "%ld objects cached, cache pressure %d\n", |
1758 | cached, sysctl_vfs_cache_pressure); | |
1759 | ||
d7e09d03 PT |
1760 | return cached; |
1761 | } | |
1762 | ||
fe92a055 PT |
1763 | static unsigned long lu_cache_shrink_scan(struct shrinker *sk, |
1764 | struct shrink_control *sc) | |
1765 | { | |
1766 | struct lu_site *s; | |
1767 | struct lu_site *tmp; | |
1768 | unsigned long remain = sc->nr_to_scan, freed = 0; | |
1769 | LIST_HEAD(splice); | |
1770 | ||
1771 | if (!(sc->gfp_mask & __GFP_FS)) | |
1772 | /* We must not take the lu_sites_guard lock when | |
1773 | * __GFP_FS is *not* set because of the deadlock | |
1774 | * possibility detailed above. Additionally, | |
1775 | * since we cannot determine the number of | |
1776 | * objects in the cache without taking this | |
1777 | * lock, we're in a particularly tough spot. As | |
1778 | * a result, we'll just lie and say our cache is | |
1779 | * empty. This _should_ be ok, as we can't | |
1780 | * reclaim objects when __GFP_FS is *not* set | |
1781 | * anyways. | |
1782 | */ | |
1783 | return SHRINK_STOP; | |
1784 | ||
1785 | mutex_lock(&lu_sites_guard); | |
1786 | list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) { | |
1787 | freed = lu_site_purge(&lu_shrink_env, s, remain); | |
1788 | remain -= freed; | |
1789 | /* | |
1790 | * Move just shrunk site to the tail of site list to | |
1791 | * assure shrinking fairness. | |
1792 | */ | |
1793 | list_move_tail(&s->ls_linkage, &splice); | |
1794 | } | |
1795 | list_splice(&splice, lu_sites.prev); | |
1796 | mutex_unlock(&lu_sites_guard); | |
1797 | ||
1798 | return sc->nr_to_scan - remain; | |
1799 | } | |
1800 | ||
d7e09d03 PT |
1801 | /** |
1802 | * Debugging printer function using printk(). | |
1803 | */ | |
fe92a055 PT |
1804 | static struct shrinker lu_site_shrinker = { |
1805 | .count_objects = lu_cache_shrink_count, | |
1806 | .scan_objects = lu_cache_shrink_scan, | |
1807 | .seeks = DEFAULT_SEEKS, | |
1808 | }; | |
1809 | ||
d7e09d03 PT |
1810 | /** |
1811 | * Initialization of global lu_* data. | |
1812 | */ | |
1813 | int lu_global_init(void) | |
1814 | { | |
1815 | int result; | |
1816 | ||
1817 | CDEBUG(D_INFO, "Lustre LU module (%p).\n", &lu_keys); | |
1818 | ||
1819 | result = lu_ref_global_init(); | |
1820 | if (result != 0) | |
1821 | return result; | |
1822 | ||
1823 | LU_CONTEXT_KEY_INIT(&lu_global_key); | |
1824 | result = lu_context_key_register(&lu_global_key); | |
1825 | if (result != 0) | |
1826 | return result; | |
1827 | ||
1828 | /* | |
1829 | * At this level, we don't know what tags are needed, so allocate them | |
1830 | * conservatively. This should not be too bad, because this | |
1831 | * environment is global. | |
1832 | */ | |
1833 | mutex_lock(&lu_sites_guard); | |
1834 | result = lu_env_init(&lu_shrink_env, LCT_SHRINKER); | |
1835 | mutex_unlock(&lu_sites_guard); | |
1836 | if (result != 0) | |
1837 | return result; | |
1838 | ||
1839 | /* | |
1840 | * seeks estimation: 3 seeks to read a record from oi, one to read | |
1841 | * inode, one for ea. Unfortunately setting this high value results in | |
1842 | * lu_object/inode cache consuming all the memory. | |
1843 | */ | |
fe92a055 | 1844 | register_shrinker(&lu_site_shrinker); |
d7e09d03 PT |
1845 | |
1846 | return result; | |
1847 | } | |
1848 | ||
1849 | /** | |
1850 | * Dual to lu_global_init(). | |
1851 | */ | |
1852 | void lu_global_fini(void) | |
1853 | { | |
fe92a055 | 1854 | unregister_shrinker(&lu_site_shrinker); |
d7e09d03 PT |
1855 | lu_context_key_degister(&lu_global_key); |
1856 | ||
1857 | /* | |
1858 | * Tear shrinker environment down _after_ de-registering | |
1859 | * lu_global_key, because the latter has a value in the former. | |
1860 | */ | |
1861 | mutex_lock(&lu_sites_guard); | |
1862 | lu_env_fini(&lu_shrink_env); | |
1863 | mutex_unlock(&lu_sites_guard); | |
1864 | ||
1865 | lu_ref_global_fini(); | |
1866 | } | |
1867 | ||
1868 | static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx) | |
1869 | { | |
d7e09d03 PT |
1870 | struct lprocfs_counter ret; |
1871 | ||
1872 | lprocfs_stats_collect(stats, idx, &ret); | |
a0b8803a AK |
1873 | if (idx == LU_SS_LRU_LEN) |
1874 | /* | |
1875 | * protect against counter on cpu A being decremented | |
1876 | * before counter is incremented on cpu B; unlikely | |
1877 | */ | |
1878 | return (__u32)((ret.lc_sum > 0) ? ret.lc_sum : 0); | |
1879 | ||
d7e09d03 | 1880 | return (__u32)ret.lc_count; |
d7e09d03 PT |
1881 | } |
1882 | ||
1883 | /** | |
1884 | * Output site statistical counters into a buffer. Suitable for | |
1885 | * lprocfs_rd_*()-style functions. | |
1886 | */ | |
73bb1da6 | 1887 | int lu_site_stats_print(const struct lu_site *s, struct seq_file *m) |
d7e09d03 | 1888 | { |
2de5855c | 1889 | struct lu_site_stats stats; |
d7e09d03 PT |
1890 | |
1891 | memset(&stats, 0, sizeof(stats)); | |
1892 | lu_site_stats_get(s->ls_obj_hash, &stats, 1); | |
1893 | ||
a0b8803a | 1894 | seq_printf(m, "%d/%d %d/%d %d %d %d %d %d %d %d %d\n", |
8faeebdf JP |
1895 | stats.lss_busy, |
1896 | stats.lss_total, | |
1897 | stats.lss_populated, | |
1898 | CFS_HASH_NHLIST(s->ls_obj_hash), | |
1899 | stats.lss_max_search, | |
1900 | ls_stats_read(s->ls_stats, LU_SS_CREATED), | |
1901 | ls_stats_read(s->ls_stats, LU_SS_CACHE_HIT), | |
1902 | ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS), | |
1903 | ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE), | |
1904 | ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE), | |
a0b8803a AK |
1905 | ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED), |
1906 | ls_stats_read(s->ls_stats, LU_SS_LRU_LEN)); | |
8faeebdf | 1907 | return 0; |
d7e09d03 PT |
1908 | } |
1909 | EXPORT_SYMBOL(lu_site_stats_print); | |
1910 | ||
1911 | /** | |
1912 | * Helper function to initialize a number of kmem slab caches at once. | |
1913 | */ | |
1914 | int lu_kmem_init(struct lu_kmem_descr *caches) | |
1915 | { | |
1916 | int result; | |
1917 | struct lu_kmem_descr *iter = caches; | |
1918 | ||
cce3c2da | 1919 | for (result = 0; iter->ckd_cache; ++iter) { |
d7e09d03 PT |
1920 | *iter->ckd_cache = kmem_cache_create(iter->ckd_name, |
1921 | iter->ckd_size, | |
1922 | 0, 0, NULL); | |
cce3c2da | 1923 | if (!*iter->ckd_cache) { |
d7e09d03 PT |
1924 | result = -ENOMEM; |
1925 | /* free all previously allocated caches */ | |
1926 | lu_kmem_fini(caches); | |
1927 | break; | |
1928 | } | |
1929 | } | |
1930 | return result; | |
1931 | } | |
1932 | EXPORT_SYMBOL(lu_kmem_init); | |
1933 | ||
1934 | /** | |
1935 | * Helper function to finalize a number of kmem slab cached at once. Dual to | |
1936 | * lu_kmem_init(). | |
1937 | */ | |
1938 | void lu_kmem_fini(struct lu_kmem_descr *caches) | |
1939 | { | |
cce3c2da | 1940 | for (; caches->ckd_cache; ++caches) { |
ce85ed4d JL |
1941 | kmem_cache_destroy(*caches->ckd_cache); |
1942 | *caches->ckd_cache = NULL; | |
d7e09d03 PT |
1943 | } |
1944 | } | |
1945 | EXPORT_SYMBOL(lu_kmem_fini); |