4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
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.
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).
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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/obdclass/lu_object.c
35 * These are the only exported functions, they provide some generic
36 * infrastructure for managing object devices
38 * Author: Nikita Danilov <nikita.danilov@sun.com>
41 #define DEBUG_SUBSYSTEM S_CLASS
43 #include <linux/module.h>
44 #include <linux/list.h>
45 #ifdef HAVE_PROCESSOR_H
46 #include <linux/processor.h>
48 #include <libcfs/linux/processor.h>
51 #include <libcfs/libcfs.h>
52 #include <libcfs/libcfs_hash.h> /* hash_long() */
53 #include <libcfs/linux/linux-mem.h>
54 #include <obd_class.h>
55 #include <obd_support.h>
56 #include <lustre_disk.h>
57 #include <lustre_fid.h>
58 #include <lu_object.h>
61 struct lu_site_bkt_data {
63 * LRU list, updated on each access to object. Protected by
64 * bucket lock of lu_site::ls_obj_hash.
66 * "Cold" end of LRU is lu_site::ls_lru.next. Accessed object are
67 * moved to the lu_site::ls_lru.prev (this is due to the non-existence
68 * of list_for_each_entry_safe_reverse()).
70 struct list_head lsb_lru;
72 * Wait-queue signaled when an object in this site is ultimately
73 * destroyed (lu_object_free()). It is used by lu_object_find() to
74 * wait before re-trying when object in the process of destruction is
75 * found in the hash table.
77 * \see htable_lookup().
79 wait_queue_head_t lsb_marche_funebre;
83 LU_CACHE_PERCENT_MAX = 50,
84 LU_CACHE_PERCENT_DEFAULT = 20
87 #define LU_CACHE_NR_MAX_ADJUST 512
88 #define LU_CACHE_NR_UNLIMITED -1
89 #define LU_CACHE_NR_DEFAULT LU_CACHE_NR_UNLIMITED
90 #define LU_CACHE_NR_LDISKFS_LIMIT LU_CACHE_NR_UNLIMITED
91 /** This is set to roughly (20 * OSS_NTHRS_MAX) to prevent thrashing */
92 #define LU_CACHE_NR_ZFS_LIMIT 10240
94 #define LU_SITE_BITS_MIN 12
95 #define LU_SITE_BITS_MAX 24
96 #define LU_SITE_BITS_MAX_CL 19
98 * total 256 buckets, we don't want too many buckets because:
99 * - consume too much memory
100 * - avoid unbalanced LRU list
102 #define LU_SITE_BKT_BITS 8
105 static unsigned int lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
106 module_param(lu_cache_percent, int, 0644);
107 MODULE_PARM_DESC(lu_cache_percent, "Percentage of memory to be used as lu_object cache");
109 static long lu_cache_nr = LU_CACHE_NR_DEFAULT;
110 module_param(lu_cache_nr, long, 0644);
111 MODULE_PARM_DESC(lu_cache_nr, "Maximum number of objects in lu_object cache");
113 static void lu_object_free(const struct lu_env *env, struct lu_object *o);
114 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx);
117 lu_site_wq_from_fid(struct lu_site *site, struct lu_fid *fid)
119 struct cfs_hash_bd bd;
120 struct lu_site_bkt_data *bkt;
122 cfs_hash_bd_get(site->ls_obj_hash, fid, &bd);
123 bkt = cfs_hash_bd_extra_get(site->ls_obj_hash, &bd);
124 return &bkt->lsb_marche_funebre;
126 EXPORT_SYMBOL(lu_site_wq_from_fid);
129 * Decrease reference counter on object. If last reference is freed, return
130 * object to the cache, unless lu_object_is_dying(o) holds. In the latter
131 * case, free object immediately.
133 void lu_object_put(const struct lu_env *env, struct lu_object *o)
135 struct lu_site_bkt_data *bkt;
136 struct lu_object_header *top = o->lo_header;
137 struct lu_site *site = o->lo_dev->ld_site;
138 struct lu_object *orig = o;
139 struct cfs_hash_bd bd;
140 const struct lu_fid *fid = lu_object_fid(o);
144 * till we have full fids-on-OST implemented anonymous objects
145 * are possible in OSP. such an object isn't listed in the site
146 * so we should not remove it from the site.
148 if (fid_is_zero(fid)) {
149 LASSERT(top->loh_hash.next == NULL
150 && top->loh_hash.pprev == NULL);
151 LASSERT(list_empty(&top->loh_lru));
152 if (!atomic_dec_and_test(&top->loh_ref))
154 list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
155 if (o->lo_ops->loo_object_release != NULL)
156 o->lo_ops->loo_object_release(env, o);
158 lu_object_free(env, orig);
162 cfs_hash_bd_get(site->ls_obj_hash, &top->loh_fid, &bd);
163 bkt = cfs_hash_bd_extra_get(site->ls_obj_hash, &bd);
165 is_dying = lu_object_is_dying(top);
166 if (!cfs_hash_bd_dec_and_lock(site->ls_obj_hash, &bd, &top->loh_ref)) {
167 /* at this point the object reference is dropped and lock is
168 * not taken, so lu_object should not be touched because it
169 * can be freed by concurrent thread. Use local variable for
174 * somebody may be waiting for this, currently only
175 * used for cl_object, see cl_object_put_last().
177 wake_up_all(&bkt->lsb_marche_funebre);
183 * When last reference is released, iterate over object
184 * layers, and notify them that object is no longer busy.
186 list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
187 if (o->lo_ops->loo_object_release != NULL)
188 o->lo_ops->loo_object_release(env, o);
191 /* don't use local 'is_dying' here because if was taken without lock
192 * but here we need the latest actual value of it so check lu_object
195 if (!lu_object_is_dying(top) &&
196 (lu_object_exists(orig) || lu_object_is_cl(orig))) {
197 LASSERT(list_empty(&top->loh_lru));
198 list_add_tail(&top->loh_lru, &bkt->lsb_lru);
199 percpu_counter_inc(&site->ls_lru_len_counter);
200 CDEBUG(D_INODE, "Add %p/%p to site lru. hash: %p, bkt: %p\n",
201 orig, top, site->ls_obj_hash, bkt);
202 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
207 * If object is dying (will not be cached) then remove it
208 * from hash table and LRU.
210 * This is done with hash table and LRU lists locked. As the only
211 * way to acquire first reference to previously unreferenced
212 * object is through hash-table lookup (lu_object_find()),
213 * or LRU scanning (lu_site_purge()), that are done under hash-table
214 * and LRU lock, no race with concurrent object lookup is possible
215 * and we can safely destroy object below.
217 if (!test_and_set_bit(LU_OBJECT_UNHASHED, &top->loh_flags))
218 cfs_hash_bd_del_locked(site->ls_obj_hash, &bd, &top->loh_hash);
219 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
221 * Object was already removed from hash and lru above, can
224 lu_object_free(env, orig);
226 EXPORT_SYMBOL(lu_object_put);
229 * Put object and don't keep in cache. This is temporary solution for
230 * multi-site objects when its layering is not constant.
232 void lu_object_put_nocache(const struct lu_env *env, struct lu_object *o)
234 set_bit(LU_OBJECT_HEARD_BANSHEE, &o->lo_header->loh_flags);
235 return lu_object_put(env, o);
237 EXPORT_SYMBOL(lu_object_put_nocache);
240 * Kill the object and take it out of LRU cache.
241 * Currently used by client code for layout change.
243 void lu_object_unhash(const struct lu_env *env, struct lu_object *o)
245 struct lu_object_header *top;
248 set_bit(LU_OBJECT_HEARD_BANSHEE, &top->loh_flags);
249 if (!test_and_set_bit(LU_OBJECT_UNHASHED, &top->loh_flags)) {
250 struct lu_site *site = o->lo_dev->ld_site;
251 struct cfs_hash *obj_hash = site->ls_obj_hash;
252 struct cfs_hash_bd bd;
254 cfs_hash_bd_get_and_lock(obj_hash, &top->loh_fid, &bd, 1);
255 if (!list_empty(&top->loh_lru)) {
256 struct lu_site_bkt_data *bkt;
258 list_del_init(&top->loh_lru);
259 bkt = cfs_hash_bd_extra_get(obj_hash, &bd);
260 percpu_counter_dec(&site->ls_lru_len_counter);
262 cfs_hash_bd_del_locked(obj_hash, &bd, &top->loh_hash);
263 cfs_hash_bd_unlock(obj_hash, &bd, 1);
266 EXPORT_SYMBOL(lu_object_unhash);
269 * Allocate new object.
271 * This follows object creation protocol, described in the comment within
272 * struct lu_device_operations definition.
274 static struct lu_object *lu_object_alloc(const struct lu_env *env,
275 struct lu_device *dev,
276 const struct lu_fid *f,
277 const struct lu_object_conf *conf)
279 struct lu_object *scan;
280 struct lu_object *top;
281 struct list_head *layers;
282 unsigned int init_mask = 0;
283 unsigned int init_flag;
289 * Create top-level object slice. This will also create
292 top = dev->ld_ops->ldo_object_alloc(env, NULL, dev);
294 RETURN(ERR_PTR(-ENOMEM));
298 * This is the only place where object fid is assigned. It's constant
301 top->lo_header->loh_fid = *f;
302 layers = &top->lo_header->loh_layers;
306 * Call ->loo_object_init() repeatedly, until no more new
307 * object slices are created.
311 list_for_each_entry(scan, layers, lo_linkage) {
312 if (init_mask & init_flag)
315 scan->lo_header = top->lo_header;
316 result = scan->lo_ops->loo_object_init(env, scan, conf);
318 lu_object_free(env, top);
319 RETURN(ERR_PTR(result));
321 init_mask |= init_flag;
327 list_for_each_entry_reverse(scan, layers, lo_linkage) {
328 if (scan->lo_ops->loo_object_start != NULL) {
329 result = scan->lo_ops->loo_object_start(env, scan);
331 lu_object_free(env, top);
332 RETURN(ERR_PTR(result));
337 lprocfs_counter_incr(dev->ld_site->ls_stats, LU_SS_CREATED);
344 static void lu_object_free(const struct lu_env *env, struct lu_object *o)
346 wait_queue_head_t *wq;
347 struct lu_site *site;
348 struct lu_object *scan;
349 struct list_head *layers;
350 struct list_head splice;
352 site = o->lo_dev->ld_site;
353 layers = &o->lo_header->loh_layers;
354 wq = lu_site_wq_from_fid(site, &o->lo_header->loh_fid);
356 * First call ->loo_object_delete() method to release all resources.
358 list_for_each_entry_reverse(scan, layers, lo_linkage) {
359 if (scan->lo_ops->loo_object_delete != NULL)
360 scan->lo_ops->loo_object_delete(env, scan);
364 * Then, splice object layers into stand-alone list, and call
365 * ->loo_object_free() on all layers to free memory. Splice is
366 * necessary, because lu_object_header is freed together with the
369 INIT_LIST_HEAD(&splice);
370 list_splice_init(layers, &splice);
371 while (!list_empty(&splice)) {
373 * Free layers in bottom-to-top order, so that object header
374 * lives as long as possible and ->loo_object_free() methods
375 * can look at its contents.
377 o = container_of0(splice.prev, struct lu_object, lo_linkage);
378 list_del_init(&o->lo_linkage);
379 LASSERT(o->lo_ops->loo_object_free != NULL);
380 o->lo_ops->loo_object_free(env, o);
383 if (waitqueue_active(wq))
388 * Free \a nr objects from the cold end of the site LRU list.
389 * if canblock is 0, then don't block awaiting for another
390 * instance of lu_site_purge() to complete
392 int lu_site_purge_objects(const struct lu_env *env, struct lu_site *s,
393 int nr, int canblock)
395 struct lu_object_header *h;
396 struct lu_object_header *temp;
397 struct lu_site_bkt_data *bkt;
398 struct cfs_hash_bd bd;
399 struct cfs_hash_bd bd2;
400 struct list_head dispose;
402 unsigned int start = 0;
407 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_NO_LRU))
410 INIT_LIST_HEAD(&dispose);
412 * Under LRU list lock, scan LRU list and move unreferenced objects to
413 * the dispose list, removing them from LRU and hash table.
416 start = s->ls_purge_start;
417 bnr = (nr == ~0) ? -1 : nr / (int)CFS_HASH_NBKT(s->ls_obj_hash) + 1;
420 * It doesn't make any sense to make purge threads parallel, that can
421 * only bring troubles to us. See LU-5331.
424 mutex_lock(&s->ls_purge_mutex);
425 else if (mutex_trylock(&s->ls_purge_mutex) == 0)
429 cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
433 cfs_hash_bd_lock(s->ls_obj_hash, &bd, 1);
434 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
436 list_for_each_entry_safe(h, temp, &bkt->lsb_lru, loh_lru) {
437 LASSERT(atomic_read(&h->loh_ref) == 0);
439 cfs_hash_bd_get(s->ls_obj_hash, &h->loh_fid, &bd2);
440 LASSERT(bd.bd_bucket == bd2.bd_bucket);
442 cfs_hash_bd_del_locked(s->ls_obj_hash,
444 list_move(&h->loh_lru, &dispose);
445 percpu_counter_dec(&s->ls_lru_len_counter);
449 if (nr != ~0 && --nr == 0)
452 if (count > 0 && --count == 0)
456 cfs_hash_bd_unlock(s->ls_obj_hash, &bd, 1);
459 * Free everything on the dispose list. This is safe against
460 * races due to the reasons described in lu_object_put().
462 while (!list_empty(&dispose)) {
463 h = container_of0(dispose.next,
464 struct lu_object_header, loh_lru);
465 list_del_init(&h->loh_lru);
466 lu_object_free(env, lu_object_top(h));
467 lprocfs_counter_incr(s->ls_stats, LU_SS_LRU_PURGED);
473 mutex_unlock(&s->ls_purge_mutex);
475 if (nr != 0 && did_sth && start != 0) {
476 start = 0; /* restart from the first bucket */
479 /* race on s->ls_purge_start, but nobody cares */
480 s->ls_purge_start = i % CFS_HASH_NBKT(s->ls_obj_hash);
485 EXPORT_SYMBOL(lu_site_purge_objects);
490 * Code below has to jump through certain loops to output object description
491 * into libcfs_debug_msg-based log. The problem is that lu_object_print()
492 * composes object description from strings that are parts of _lines_ of
493 * output (i.e., strings that are not terminated by newline). This doesn't fit
494 * very well into libcfs_debug_msg() interface that assumes that each message
495 * supplied to it is a self-contained output line.
497 * To work around this, strings are collected in a temporary buffer
498 * (implemented as a value of lu_cdebug_key key), until terminating newline
499 * character is detected.
507 * XXX overflow is not handled correctly.
512 struct lu_cdebug_data {
516 char lck_area[LU_CDEBUG_LINE];
519 /* context key constructor/destructor: lu_global_key_init, lu_global_key_fini */
520 LU_KEY_INIT_FINI(lu_global, struct lu_cdebug_data);
523 * Key, holding temporary buffer. This key is registered very early by
526 static struct lu_context_key lu_global_key = {
527 .lct_tags = LCT_MD_THREAD | LCT_DT_THREAD |
528 LCT_MG_THREAD | LCT_CL_THREAD | LCT_LOCAL,
529 .lct_init = lu_global_key_init,
530 .lct_fini = lu_global_key_fini
534 * Printer function emitting messages through libcfs_debug_msg().
536 int lu_cdebug_printer(const struct lu_env *env,
537 void *cookie, const char *format, ...)
539 struct libcfs_debug_msg_data *msgdata = cookie;
540 struct lu_cdebug_data *key;
545 va_start(args, format);
547 key = lu_context_key_get(&env->le_ctx, &lu_global_key);
548 LASSERT(key != NULL);
550 used = strlen(key->lck_area);
551 complete = format[strlen(format) - 1] == '\n';
553 * Append new chunk to the buffer.
555 vsnprintf(key->lck_area + used,
556 ARRAY_SIZE(key->lck_area) - used, format, args);
558 if (cfs_cdebug_show(msgdata->msg_mask, msgdata->msg_subsys))
559 libcfs_debug_msg(msgdata, "%s\n", key->lck_area);
560 key->lck_area[0] = 0;
565 EXPORT_SYMBOL(lu_cdebug_printer);
568 * Print object header.
570 void lu_object_header_print(const struct lu_env *env, void *cookie,
571 lu_printer_t printer,
572 const struct lu_object_header *hdr)
574 (*printer)(env, cookie, "header@%p[%#lx, %d, "DFID"%s%s%s]",
575 hdr, hdr->loh_flags, atomic_read(&hdr->loh_ref),
577 hlist_unhashed(&hdr->loh_hash) ? "" : " hash",
578 list_empty((struct list_head *)&hdr->loh_lru) ? \
580 hdr->loh_attr & LOHA_EXISTS ? " exist" : "");
582 EXPORT_SYMBOL(lu_object_header_print);
585 * Print human readable representation of the \a o to the \a printer.
587 void lu_object_print(const struct lu_env *env, void *cookie,
588 lu_printer_t printer, const struct lu_object *o)
590 static const char ruler[] = "........................................";
591 struct lu_object_header *top;
595 lu_object_header_print(env, cookie, printer, top);
596 (*printer)(env, cookie, "{\n");
598 list_for_each_entry(o, &top->loh_layers, lo_linkage) {
600 * print `.' \a depth times followed by type name and address
602 (*printer)(env, cookie, "%*.*s%s@%p", depth, depth, ruler,
603 o->lo_dev->ld_type->ldt_name, o);
605 if (o->lo_ops->loo_object_print != NULL)
606 (*o->lo_ops->loo_object_print)(env, cookie, printer, o);
608 (*printer)(env, cookie, "\n");
611 (*printer)(env, cookie, "} header@%p\n", top);
613 EXPORT_SYMBOL(lu_object_print);
616 * Check object consistency.
618 int lu_object_invariant(const struct lu_object *o)
620 struct lu_object_header *top;
623 list_for_each_entry(o, &top->loh_layers, lo_linkage) {
624 if (o->lo_ops->loo_object_invariant != NULL &&
625 !o->lo_ops->loo_object_invariant(o))
631 static struct lu_object *htable_lookup(struct lu_site *s,
632 struct cfs_hash_bd *bd,
633 const struct lu_fid *f,
636 struct lu_site_bkt_data *bkt;
637 struct lu_object_header *h;
638 struct hlist_node *hnode;
639 __u64 ver = cfs_hash_bd_version_get(bd);
642 return ERR_PTR(-ENOENT);
645 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, bd);
646 /* cfs_hash_bd_peek_locked is a somehow "internal" function
647 * of cfs_hash, it doesn't add refcount on object. */
648 hnode = cfs_hash_bd_peek_locked(s->ls_obj_hash, bd, (void *)f);
650 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS);
651 return ERR_PTR(-ENOENT);
654 h = container_of0(hnode, struct lu_object_header, loh_hash);
655 cfs_hash_get(s->ls_obj_hash, hnode);
656 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_HIT);
657 if (!list_empty(&h->loh_lru)) {
658 list_del_init(&h->loh_lru);
659 percpu_counter_dec(&s->ls_lru_len_counter);
661 return lu_object_top(h);
665 * Search cache for an object with the fid \a f. If such object is found,
666 * return it. Otherwise, create new object, insert it into cache and return
667 * it. In any case, additional reference is acquired on the returned object.
669 struct lu_object *lu_object_find(const struct lu_env *env,
670 struct lu_device *dev, const struct lu_fid *f,
671 const struct lu_object_conf *conf)
673 return lu_object_find_at(env, dev->ld_site->ls_top_dev, f, conf);
675 EXPORT_SYMBOL(lu_object_find);
678 * Limit the lu_object cache to a maximum of lu_cache_nr objects. Because
679 * the calculation for the number of objects to reclaim is not covered by
680 * a lock the maximum number of objects is capped by LU_CACHE_MAX_ADJUST.
681 * This ensures that many concurrent threads will not accidentally purge
684 static void lu_object_limit(const struct lu_env *env,
685 struct lu_device *dev)
689 if (lu_cache_nr == LU_CACHE_NR_UNLIMITED)
692 size = cfs_hash_size_get(dev->ld_site->ls_obj_hash);
693 nr = (__u64)lu_cache_nr;
697 lu_site_purge_objects(env, dev->ld_site,
698 MIN(size - nr, LU_CACHE_NR_MAX_ADJUST), 0);
702 * Core logic of lu_object_find*() functions.
704 * Much like lu_object_find(), but top level device of object is specifically
705 * \a dev rather than top level device of the site. This interface allows
706 * objects of different "stacking" to be created within the same site.
708 struct lu_object *lu_object_find_at(const struct lu_env *env,
709 struct lu_device *dev,
710 const struct lu_fid *f,
711 const struct lu_object_conf *conf)
714 struct lu_object *shadow;
717 struct cfs_hash_bd bd;
721 * This uses standard index maintenance protocol:
723 * - search index under lock, and return object if found;
724 * - otherwise, unlock index, allocate new object;
725 * - lock index and search again;
726 * - if nothing is found (usual case), insert newly created
728 * - otherwise (race: other thread inserted object), free
729 * object just allocated.
733 * For "LOC_F_NEW" case, we are sure the object is new established.
734 * It is unnecessary to perform lookup-alloc-lookup-insert, instead,
735 * just alloc and insert directly.
740 cfs_hash_bd_get(hs, f, &bd);
741 if (!(conf && conf->loc_flags & LOC_F_NEW)) {
742 cfs_hash_bd_lock(hs, &bd, 1);
743 o = htable_lookup(s, &bd, f, &version);
744 cfs_hash_bd_unlock(hs, &bd, 1);
746 if (!IS_ERR(o) || PTR_ERR(o) != -ENOENT)
750 * Allocate new object. This may result in rather complicated
751 * operations, including fld queries, inode loading, etc.
753 o = lu_object_alloc(env, dev, f, conf);
757 LASSERT(lu_fid_eq(lu_object_fid(o), f));
759 cfs_hash_bd_lock(hs, &bd, 1);
761 if (conf && conf->loc_flags & LOC_F_NEW)
762 shadow = ERR_PTR(-ENOENT);
764 shadow = htable_lookup(s, &bd, f, &version);
765 if (likely(PTR_ERR(shadow) == -ENOENT)) {
766 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
767 cfs_hash_bd_unlock(hs, &bd, 1);
769 lu_object_limit(env, dev);
774 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_RACE);
775 cfs_hash_bd_unlock(hs, &bd, 1);
776 lu_object_free(env, o);
779 EXPORT_SYMBOL(lu_object_find_at);
782 * Find object with given fid, and return its slice belonging to given device.
784 struct lu_object *lu_object_find_slice(const struct lu_env *env,
785 struct lu_device *dev,
786 const struct lu_fid *f,
787 const struct lu_object_conf *conf)
789 struct lu_object *top;
790 struct lu_object *obj;
792 top = lu_object_find(env, dev, f, conf);
796 obj = lu_object_locate(top->lo_header, dev->ld_type);
797 if (unlikely(obj == NULL)) {
798 lu_object_put(env, top);
799 obj = ERR_PTR(-ENOENT);
804 EXPORT_SYMBOL(lu_object_find_slice);
806 int lu_device_type_init(struct lu_device_type *ldt)
810 atomic_set(&ldt->ldt_device_nr, 0);
811 if (ldt->ldt_ops->ldto_init)
812 result = ldt->ldt_ops->ldto_init(ldt);
816 EXPORT_SYMBOL(lu_device_type_init);
818 void lu_device_type_fini(struct lu_device_type *ldt)
820 if (ldt->ldt_ops->ldto_fini)
821 ldt->ldt_ops->ldto_fini(ldt);
823 EXPORT_SYMBOL(lu_device_type_fini);
826 * Global list of all sites on this node
828 static LIST_HEAD(lu_sites);
829 static DECLARE_RWSEM(lu_sites_guard);
832 * Global environment used by site shrinker.
834 static struct lu_env lu_shrink_env;
836 struct lu_site_print_arg {
837 struct lu_env *lsp_env;
839 lu_printer_t lsp_printer;
843 lu_site_obj_print(struct cfs_hash *hs, struct cfs_hash_bd *bd,
844 struct hlist_node *hnode, void *data)
846 struct lu_site_print_arg *arg = (struct lu_site_print_arg *)data;
847 struct lu_object_header *h;
849 h = hlist_entry(hnode, struct lu_object_header, loh_hash);
850 if (!list_empty(&h->loh_layers)) {
851 const struct lu_object *o;
853 o = lu_object_top(h);
854 lu_object_print(arg->lsp_env, arg->lsp_cookie,
855 arg->lsp_printer, o);
857 lu_object_header_print(arg->lsp_env, arg->lsp_cookie,
858 arg->lsp_printer, h);
864 * Print all objects in \a s.
866 void lu_site_print(const struct lu_env *env, struct lu_site *s, void *cookie,
867 lu_printer_t printer)
869 struct lu_site_print_arg arg = {
870 .lsp_env = (struct lu_env *)env,
871 .lsp_cookie = cookie,
872 .lsp_printer = printer,
875 cfs_hash_for_each(s->ls_obj_hash, lu_site_obj_print, &arg);
877 EXPORT_SYMBOL(lu_site_print);
880 * Return desired hash table order.
882 static unsigned long lu_htable_order(struct lu_device *top)
884 unsigned long cache_size;
886 unsigned long bits_max = LU_SITE_BITS_MAX;
889 * For ZFS based OSDs the cache should be disabled by default. This
890 * allows the ZFS ARC maximum flexibility in determining what buffers
891 * to cache. If Lustre has objects or buffer which it wants to ensure
892 * always stay cached it must maintain a hold on them.
894 if (strcmp(top->ld_type->ldt_name, LUSTRE_OSD_ZFS_NAME) == 0) {
895 lu_cache_percent = 1;
896 lu_cache_nr = LU_CACHE_NR_ZFS_LIMIT;
897 return LU_SITE_BITS_MIN;
900 if (strcmp(top->ld_type->ldt_name, LUSTRE_VVP_NAME) == 0)
901 bits_max = LU_SITE_BITS_MAX_CL;
904 * Calculate hash table size, assuming that we want reasonable
905 * performance when 20% of total memory is occupied by cache of
908 * Size of lu_object is (arbitrary) taken as 1K (together with inode).
910 cache_size = totalram_pages;
912 #if BITS_PER_LONG == 32
913 /* limit hashtable size for lowmem systems to low RAM */
914 if (cache_size > 1 << (30 - PAGE_SHIFT))
915 cache_size = 1 << (30 - PAGE_SHIFT) * 3 / 4;
918 /* clear off unreasonable cache setting. */
919 if (lu_cache_percent == 0 || lu_cache_percent > LU_CACHE_PERCENT_MAX) {
920 CWARN("obdclass: invalid lu_cache_percent: %u, it must be in"
921 " the range of (0, %u]. Will use default value: %u.\n",
922 lu_cache_percent, LU_CACHE_PERCENT_MAX,
923 LU_CACHE_PERCENT_DEFAULT);
925 lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
927 cache_size = cache_size / 100 * lu_cache_percent *
930 for (bits = 1; (1 << bits) < cache_size; ++bits) {
934 return clamp_t(typeof(bits), bits, LU_SITE_BITS_MIN, bits_max);
937 static unsigned lu_obj_hop_hash(struct cfs_hash *hs,
938 const void *key, unsigned mask)
940 struct lu_fid *fid = (struct lu_fid *)key;
943 hash = fid_flatten32(fid);
944 hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
945 hash = hash_long(hash, hs->hs_bkt_bits);
947 /* give me another random factor */
948 hash -= hash_long((unsigned long)hs, fid_oid(fid) % 11 + 3);
950 hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
951 hash |= (fid_seq(fid) + fid_oid(fid)) & (CFS_HASH_NBKT(hs) - 1);
956 static void *lu_obj_hop_object(struct hlist_node *hnode)
958 return hlist_entry(hnode, struct lu_object_header, loh_hash);
961 static void *lu_obj_hop_key(struct hlist_node *hnode)
963 struct lu_object_header *h;
965 h = hlist_entry(hnode, struct lu_object_header, loh_hash);
969 static int lu_obj_hop_keycmp(const void *key, struct hlist_node *hnode)
971 struct lu_object_header *h;
973 h = hlist_entry(hnode, struct lu_object_header, loh_hash);
974 return lu_fid_eq(&h->loh_fid, (struct lu_fid *)key);
977 static void lu_obj_hop_get(struct cfs_hash *hs, struct hlist_node *hnode)
979 struct lu_object_header *h;
981 h = hlist_entry(hnode, struct lu_object_header, loh_hash);
982 atomic_inc(&h->loh_ref);
985 static void lu_obj_hop_put_locked(struct cfs_hash *hs, struct hlist_node *hnode)
987 LBUG(); /* we should never called it */
990 static struct cfs_hash_ops lu_site_hash_ops = {
991 .hs_hash = lu_obj_hop_hash,
992 .hs_key = lu_obj_hop_key,
993 .hs_keycmp = lu_obj_hop_keycmp,
994 .hs_object = lu_obj_hop_object,
995 .hs_get = lu_obj_hop_get,
996 .hs_put_locked = lu_obj_hop_put_locked,
999 void lu_dev_add_linkage(struct lu_site *s, struct lu_device *d)
1001 spin_lock(&s->ls_ld_lock);
1002 if (list_empty(&d->ld_linkage))
1003 list_add(&d->ld_linkage, &s->ls_ld_linkage);
1004 spin_unlock(&s->ls_ld_lock);
1006 EXPORT_SYMBOL(lu_dev_add_linkage);
1008 void lu_dev_del_linkage(struct lu_site *s, struct lu_device *d)
1010 spin_lock(&s->ls_ld_lock);
1011 list_del_init(&d->ld_linkage);
1012 spin_unlock(&s->ls_ld_lock);
1014 EXPORT_SYMBOL(lu_dev_del_linkage);
1017 * Initialize site \a s, with \a d as the top level device.
1019 int lu_site_init(struct lu_site *s, struct lu_device *top)
1021 struct lu_site_bkt_data *bkt;
1022 struct cfs_hash_bd bd;
1029 memset(s, 0, sizeof *s);
1030 mutex_init(&s->ls_purge_mutex);
1032 #ifdef HAVE_PERCPU_COUNTER_INIT_GFP_FLAG
1033 rc = percpu_counter_init(&s->ls_lru_len_counter, 0, GFP_NOFS);
1035 rc = percpu_counter_init(&s->ls_lru_len_counter, 0);
1040 snprintf(name, sizeof(name), "lu_site_%s", top->ld_type->ldt_name);
1041 for (bits = lu_htable_order(top);
1042 bits >= LU_SITE_BITS_MIN; bits--) {
1043 s->ls_obj_hash = cfs_hash_create(name, bits, bits,
1044 bits - LU_SITE_BKT_BITS,
1047 CFS_HASH_SPIN_BKTLOCK |
1048 CFS_HASH_NO_ITEMREF |
1050 CFS_HASH_ASSERT_EMPTY |
1052 if (s->ls_obj_hash != NULL)
1056 if (s->ls_obj_hash == NULL) {
1057 CERROR("failed to create lu_site hash with bits: %lu\n", bits);
1061 cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
1062 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
1063 INIT_LIST_HEAD(&bkt->lsb_lru);
1064 init_waitqueue_head(&bkt->lsb_marche_funebre);
1067 s->ls_stats = lprocfs_alloc_stats(LU_SS_LAST_STAT, 0);
1068 if (s->ls_stats == NULL) {
1069 cfs_hash_putref(s->ls_obj_hash);
1070 s->ls_obj_hash = NULL;
1074 lprocfs_counter_init(s->ls_stats, LU_SS_CREATED,
1075 0, "created", "created");
1076 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_HIT,
1077 0, "cache_hit", "cache_hit");
1078 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_MISS,
1079 0, "cache_miss", "cache_miss");
1080 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_RACE,
1081 0, "cache_race", "cache_race");
1082 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_DEATH_RACE,
1083 0, "cache_death_race", "cache_death_race");
1084 lprocfs_counter_init(s->ls_stats, LU_SS_LRU_PURGED,
1085 0, "lru_purged", "lru_purged");
1087 INIT_LIST_HEAD(&s->ls_linkage);
1088 s->ls_top_dev = top;
1091 lu_ref_add(&top->ld_reference, "site-top", s);
1093 INIT_LIST_HEAD(&s->ls_ld_linkage);
1094 spin_lock_init(&s->ls_ld_lock);
1096 lu_dev_add_linkage(s, top);
1100 EXPORT_SYMBOL(lu_site_init);
1103 * Finalize \a s and release its resources.
1105 void lu_site_fini(struct lu_site *s)
1107 down_write(&lu_sites_guard);
1108 list_del_init(&s->ls_linkage);
1109 up_write(&lu_sites_guard);
1111 percpu_counter_destroy(&s->ls_lru_len_counter);
1113 if (s->ls_obj_hash != NULL) {
1114 cfs_hash_putref(s->ls_obj_hash);
1115 s->ls_obj_hash = NULL;
1118 if (s->ls_top_dev != NULL) {
1119 s->ls_top_dev->ld_site = NULL;
1120 lu_ref_del(&s->ls_top_dev->ld_reference, "site-top", s);
1121 lu_device_put(s->ls_top_dev);
1122 s->ls_top_dev = NULL;
1125 if (s->ls_stats != NULL)
1126 lprocfs_free_stats(&s->ls_stats);
1128 EXPORT_SYMBOL(lu_site_fini);
1131 * Called when initialization of stack for this site is completed.
1133 int lu_site_init_finish(struct lu_site *s)
1136 down_write(&lu_sites_guard);
1137 result = lu_context_refill(&lu_shrink_env.le_ctx);
1139 list_add(&s->ls_linkage, &lu_sites);
1140 up_write(&lu_sites_guard);
1143 EXPORT_SYMBOL(lu_site_init_finish);
1146 * Acquire additional reference on device \a d
1148 void lu_device_get(struct lu_device *d)
1150 atomic_inc(&d->ld_ref);
1152 EXPORT_SYMBOL(lu_device_get);
1155 * Release reference on device \a d.
1157 void lu_device_put(struct lu_device *d)
1159 LASSERT(atomic_read(&d->ld_ref) > 0);
1160 atomic_dec(&d->ld_ref);
1162 EXPORT_SYMBOL(lu_device_put);
1165 * Initialize device \a d of type \a t.
1167 int lu_device_init(struct lu_device *d, struct lu_device_type *t)
1169 if (atomic_inc_return(&t->ldt_device_nr) == 1 &&
1170 t->ldt_ops->ldto_start != NULL)
1171 t->ldt_ops->ldto_start(t);
1173 memset(d, 0, sizeof *d);
1175 lu_ref_init(&d->ld_reference);
1176 INIT_LIST_HEAD(&d->ld_linkage);
1180 EXPORT_SYMBOL(lu_device_init);
1183 * Finalize device \a d.
1185 void lu_device_fini(struct lu_device *d)
1187 struct lu_device_type *t = d->ld_type;
1189 if (d->ld_obd != NULL) {
1190 d->ld_obd->obd_lu_dev = NULL;
1194 lu_ref_fini(&d->ld_reference);
1195 LASSERTF(atomic_read(&d->ld_ref) == 0,
1196 "Refcount is %u\n", atomic_read(&d->ld_ref));
1197 LASSERT(atomic_read(&t->ldt_device_nr) > 0);
1199 if (atomic_dec_and_test(&t->ldt_device_nr) &&
1200 t->ldt_ops->ldto_stop != NULL)
1201 t->ldt_ops->ldto_stop(t);
1203 EXPORT_SYMBOL(lu_device_fini);
1206 * Initialize object \a o that is part of compound object \a h and was created
1209 int lu_object_init(struct lu_object *o, struct lu_object_header *h,
1210 struct lu_device *d)
1212 memset(o, 0, sizeof(*o));
1216 lu_ref_add_at(&d->ld_reference, &o->lo_dev_ref, "lu_object", o);
1217 INIT_LIST_HEAD(&o->lo_linkage);
1221 EXPORT_SYMBOL(lu_object_init);
1224 * Finalize object and release its resources.
1226 void lu_object_fini(struct lu_object *o)
1228 struct lu_device *dev = o->lo_dev;
1230 LASSERT(list_empty(&o->lo_linkage));
1233 lu_ref_del_at(&dev->ld_reference, &o->lo_dev_ref,
1239 EXPORT_SYMBOL(lu_object_fini);
1242 * Add object \a o as first layer of compound object \a h
1244 * This is typically called by the ->ldo_object_alloc() method of top-level
1247 void lu_object_add_top(struct lu_object_header *h, struct lu_object *o)
1249 list_move(&o->lo_linkage, &h->loh_layers);
1251 EXPORT_SYMBOL(lu_object_add_top);
1254 * Add object \a o as a layer of compound object, going after \a before.
1256 * This is typically called by the ->ldo_object_alloc() method of \a
1259 void lu_object_add(struct lu_object *before, struct lu_object *o)
1261 list_move(&o->lo_linkage, &before->lo_linkage);
1263 EXPORT_SYMBOL(lu_object_add);
1266 * Initialize compound object.
1268 int lu_object_header_init(struct lu_object_header *h)
1270 memset(h, 0, sizeof *h);
1271 atomic_set(&h->loh_ref, 1);
1272 INIT_HLIST_NODE(&h->loh_hash);
1273 INIT_LIST_HEAD(&h->loh_lru);
1274 INIT_LIST_HEAD(&h->loh_layers);
1275 lu_ref_init(&h->loh_reference);
1278 EXPORT_SYMBOL(lu_object_header_init);
1281 * Finalize compound object.
1283 void lu_object_header_fini(struct lu_object_header *h)
1285 LASSERT(list_empty(&h->loh_layers));
1286 LASSERT(list_empty(&h->loh_lru));
1287 LASSERT(hlist_unhashed(&h->loh_hash));
1288 lu_ref_fini(&h->loh_reference);
1290 EXPORT_SYMBOL(lu_object_header_fini);
1293 * Given a compound object, find its slice, corresponding to the device type
1296 struct lu_object *lu_object_locate(struct lu_object_header *h,
1297 const struct lu_device_type *dtype)
1299 struct lu_object *o;
1301 list_for_each_entry(o, &h->loh_layers, lo_linkage) {
1302 if (o->lo_dev->ld_type == dtype)
1307 EXPORT_SYMBOL(lu_object_locate);
1310 * Finalize and free devices in the device stack.
1312 * Finalize device stack by purging object cache, and calling
1313 * lu_device_type_operations::ldto_device_fini() and
1314 * lu_device_type_operations::ldto_device_free() on all devices in the stack.
1316 void lu_stack_fini(const struct lu_env *env, struct lu_device *top)
1318 struct lu_site *site = top->ld_site;
1319 struct lu_device *scan;
1320 struct lu_device *next;
1322 lu_site_purge(env, site, ~0);
1323 for (scan = top; scan != NULL; scan = next) {
1324 next = scan->ld_type->ldt_ops->ldto_device_fini(env, scan);
1325 lu_ref_del(&scan->ld_reference, "lu-stack", &lu_site_init);
1326 lu_device_put(scan);
1330 lu_site_purge(env, site, ~0);
1332 for (scan = top; scan != NULL; scan = next) {
1333 const struct lu_device_type *ldt = scan->ld_type;
1334 struct obd_type *type;
1336 next = ldt->ldt_ops->ldto_device_free(env, scan);
1337 type = ldt->ldt_obd_type;
1340 class_put_type(type);
1347 * Maximal number of tld slots.
1349 LU_CONTEXT_KEY_NR = 40
1352 static struct lu_context_key *lu_keys[LU_CONTEXT_KEY_NR] = { NULL, };
1354 static DECLARE_RWSEM(lu_key_initing);
1357 * Global counter incremented whenever key is registered, unregistered,
1358 * revived or quiesced. This is used to void unnecessary calls to
1359 * lu_context_refill(). No locking is provided, as initialization and shutdown
1360 * are supposed to be externally serialized.
1362 static atomic_t key_set_version = ATOMIC_INIT(0);
1367 int lu_context_key_register(struct lu_context_key *key)
1372 LASSERT(key->lct_init != NULL);
1373 LASSERT(key->lct_fini != NULL);
1374 LASSERT(key->lct_tags != 0);
1375 LASSERT(key->lct_owner != NULL);
1378 atomic_set(&key->lct_used, 1);
1379 lu_ref_init(&key->lct_reference);
1380 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1384 if (cmpxchg(&lu_keys[i], NULL, key) != NULL)
1388 atomic_inc(&key_set_version);
1392 lu_ref_fini(&key->lct_reference);
1393 atomic_set(&key->lct_used, 0);
1397 EXPORT_SYMBOL(lu_context_key_register);
1399 static void key_fini(struct lu_context *ctx, int index)
1401 if (ctx->lc_value != NULL && ctx->lc_value[index] != NULL) {
1402 struct lu_context_key *key;
1404 key = lu_keys[index];
1405 LASSERT(key != NULL);
1406 LASSERT(key->lct_fini != NULL);
1407 LASSERT(atomic_read(&key->lct_used) > 0);
1409 key->lct_fini(ctx, key, ctx->lc_value[index]);
1410 lu_ref_del(&key->lct_reference, "ctx", ctx);
1411 if (atomic_dec_and_test(&key->lct_used))
1412 wake_up_var(&key->lct_used);
1414 LASSERT(key->lct_owner != NULL);
1415 if ((ctx->lc_tags & LCT_NOREF) == 0) {
1416 LINVRNT(module_refcount(key->lct_owner) > 0);
1417 module_put(key->lct_owner);
1419 ctx->lc_value[index] = NULL;
1426 void lu_context_key_degister(struct lu_context_key *key)
1428 LASSERT(atomic_read(&key->lct_used) >= 1);
1429 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1431 lu_context_key_quiesce(key);
1433 key_fini(&lu_shrink_env.le_ctx, key->lct_index);
1436 * Wait until all transient contexts referencing this key have
1437 * run lu_context_key::lct_fini() method.
1439 atomic_dec(&key->lct_used);
1440 wait_var_event(&key->lct_used, atomic_read(&key->lct_used) == 0);
1442 if (!WARN_ON(lu_keys[key->lct_index] == NULL))
1443 lu_ref_fini(&key->lct_reference);
1445 smp_store_release(&lu_keys[key->lct_index], NULL);
1447 EXPORT_SYMBOL(lu_context_key_degister);
1450 * Register a number of keys. This has to be called after all keys have been
1451 * initialized by a call to LU_CONTEXT_KEY_INIT().
1453 int lu_context_key_register_many(struct lu_context_key *k, ...)
1455 struct lu_context_key *key = k;
1461 result = lu_context_key_register(key);
1464 key = va_arg(args, struct lu_context_key *);
1465 } while (key != NULL);
1471 lu_context_key_degister(k);
1472 k = va_arg(args, struct lu_context_key *);
1479 EXPORT_SYMBOL(lu_context_key_register_many);
1482 * De-register a number of keys. This is a dual to
1483 * lu_context_key_register_many().
1485 void lu_context_key_degister_many(struct lu_context_key *k, ...)
1491 lu_context_key_degister(k);
1492 k = va_arg(args, struct lu_context_key*);
1493 } while (k != NULL);
1496 EXPORT_SYMBOL(lu_context_key_degister_many);
1499 * Revive a number of keys.
1501 void lu_context_key_revive_many(struct lu_context_key *k, ...)
1507 lu_context_key_revive(k);
1508 k = va_arg(args, struct lu_context_key*);
1509 } while (k != NULL);
1512 EXPORT_SYMBOL(lu_context_key_revive_many);
1515 * Quiescent a number of keys.
1517 void lu_context_key_quiesce_many(struct lu_context_key *k, ...)
1523 lu_context_key_quiesce(k);
1524 k = va_arg(args, struct lu_context_key*);
1525 } while (k != NULL);
1528 EXPORT_SYMBOL(lu_context_key_quiesce_many);
1531 * Return value associated with key \a key in context \a ctx.
1533 void *lu_context_key_get(const struct lu_context *ctx,
1534 const struct lu_context_key *key)
1536 LINVRNT(ctx->lc_state == LCS_ENTERED);
1537 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1538 LASSERT(lu_keys[key->lct_index] == key);
1539 return ctx->lc_value[key->lct_index];
1541 EXPORT_SYMBOL(lu_context_key_get);
1544 * List of remembered contexts. XXX document me.
1546 static LIST_HEAD(lu_context_remembered);
1547 static DEFINE_SPINLOCK(lu_context_remembered_guard);
1550 * Destroy \a key in all remembered contexts. This is used to destroy key
1551 * values in "shared" contexts (like service threads), when a module owning
1552 * the key is about to be unloaded.
1554 void lu_context_key_quiesce(struct lu_context_key *key)
1556 struct lu_context *ctx;
1558 if (!(key->lct_tags & LCT_QUIESCENT)) {
1560 * The write-lock on lu_key_initing will ensure that any
1561 * keys_fill() which didn't see LCT_QUIESCENT will have
1562 * finished before we call key_fini().
1564 down_write(&lu_key_initing);
1565 key->lct_tags |= LCT_QUIESCENT;
1566 up_write(&lu_key_initing);
1568 spin_lock(&lu_context_remembered_guard);
1569 list_for_each_entry(ctx, &lu_context_remembered, lc_remember) {
1570 spin_until_cond(READ_ONCE(ctx->lc_state) != LCS_LEAVING);
1571 key_fini(ctx, key->lct_index);
1574 spin_unlock(&lu_context_remembered_guard);
1578 void lu_context_key_revive(struct lu_context_key *key)
1580 key->lct_tags &= ~LCT_QUIESCENT;
1581 atomic_inc(&key_set_version);
1584 static void keys_fini(struct lu_context *ctx)
1588 if (ctx->lc_value == NULL)
1591 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i)
1594 OBD_FREE(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1595 ctx->lc_value = NULL;
1598 static int keys_fill(struct lu_context *ctx)
1604 * A serialisation with lu_context_key_quiesce() is needed, to
1605 * ensure we see LCT_QUIESCENT and don't allocate a new value
1606 * after it freed one. The rwsem provides this. As down_read()
1607 * does optimistic spinning while the writer is active, this is
1608 * unlikely to ever sleep.
1610 down_read(&lu_key_initing);
1611 ctx->lc_version = atomic_read(&key_set_version);
1613 LINVRNT(ctx->lc_value);
1614 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1615 struct lu_context_key *key;
1618 if (!ctx->lc_value[i] && key &&
1619 (key->lct_tags & ctx->lc_tags) &&
1621 * Don't create values for a LCT_QUIESCENT key, as this
1622 * will pin module owning a key.
1624 !(key->lct_tags & LCT_QUIESCENT)) {
1627 LINVRNT(key->lct_init != NULL);
1628 LINVRNT(key->lct_index == i);
1630 LASSERT(key->lct_owner != NULL);
1631 if (!(ctx->lc_tags & LCT_NOREF) &&
1632 try_module_get(key->lct_owner) == 0) {
1633 /* module is unloading, skip this key */
1637 value = key->lct_init(ctx, key);
1638 if (unlikely(IS_ERR(value))) {
1639 rc = PTR_ERR(value);
1643 lu_ref_add_atomic(&key->lct_reference, "ctx", ctx);
1644 atomic_inc(&key->lct_used);
1646 * This is the only place in the code, where an
1647 * element of ctx->lc_value[] array is set to non-NULL
1650 ctx->lc_value[i] = value;
1651 if (key->lct_exit != NULL)
1652 ctx->lc_tags |= LCT_HAS_EXIT;
1656 up_read(&lu_key_initing);
1660 static int keys_init(struct lu_context *ctx)
1662 OBD_ALLOC(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1663 if (likely(ctx->lc_value != NULL))
1664 return keys_fill(ctx);
1670 * Initialize context data-structure. Create values for all keys.
1672 int lu_context_init(struct lu_context *ctx, __u32 tags)
1676 memset(ctx, 0, sizeof *ctx);
1677 ctx->lc_state = LCS_INITIALIZED;
1678 ctx->lc_tags = tags;
1679 if (tags & LCT_REMEMBER) {
1680 spin_lock(&lu_context_remembered_guard);
1681 list_add(&ctx->lc_remember, &lu_context_remembered);
1682 spin_unlock(&lu_context_remembered_guard);
1684 INIT_LIST_HEAD(&ctx->lc_remember);
1687 rc = keys_init(ctx);
1689 lu_context_fini(ctx);
1693 EXPORT_SYMBOL(lu_context_init);
1696 * Finalize context data-structure. Destroy key values.
1698 void lu_context_fini(struct lu_context *ctx)
1700 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1701 ctx->lc_state = LCS_FINALIZED;
1703 if ((ctx->lc_tags & LCT_REMEMBER) == 0) {
1704 LASSERT(list_empty(&ctx->lc_remember));
1706 /* could race with key degister */
1707 spin_lock(&lu_context_remembered_guard);
1708 list_del_init(&ctx->lc_remember);
1709 spin_unlock(&lu_context_remembered_guard);
1713 EXPORT_SYMBOL(lu_context_fini);
1716 * Called before entering context.
1718 void lu_context_enter(struct lu_context *ctx)
1720 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1721 ctx->lc_state = LCS_ENTERED;
1723 EXPORT_SYMBOL(lu_context_enter);
1726 * Called after exiting from \a ctx
1728 void lu_context_exit(struct lu_context *ctx)
1732 LINVRNT(ctx->lc_state == LCS_ENTERED);
1734 * Disable preempt to ensure we get a warning if
1735 * any lct_exit ever tries to sleep. That would hurt
1736 * lu_context_key_quiesce() which spins waiting for us.
1737 * This also ensure we aren't preempted while the state
1738 * is LCS_LEAVING, as that too would cause problems for
1739 * lu_context_key_quiesce().
1743 * Ensure lu_context_key_quiesce() sees LCS_LEAVING
1744 * or we see LCT_QUIESCENT
1746 smp_store_mb(ctx->lc_state, LCS_LEAVING);
1747 if (ctx->lc_tags & LCT_HAS_EXIT && ctx->lc_value) {
1748 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1749 struct lu_context_key *key;
1752 if (ctx->lc_value[i] &&
1753 !(key->lct_tags & LCT_QUIESCENT) &&
1755 key->lct_exit(ctx, key, ctx->lc_value[i]);
1759 smp_store_release(&ctx->lc_state, LCS_LEFT);
1762 EXPORT_SYMBOL(lu_context_exit);
1765 * Allocate for context all missing keys that were registered after context
1766 * creation. key_set_version is only changed in rare cases when modules
1767 * are loaded and removed.
1769 int lu_context_refill(struct lu_context *ctx)
1771 if (likely(ctx->lc_version == atomic_read(&key_set_version)))
1774 return keys_fill(ctx);
1778 * lu_ctx_tags/lu_ses_tags will be updated if there are new types of
1779 * obd being added. Currently, this is only used on client side, specifically
1780 * for echo device client, for other stack (like ptlrpc threads), context are
1781 * predefined when the lu_device type are registered, during the module probe
1784 u32 lu_context_tags_default;
1785 u32 lu_session_tags_default;
1787 #ifdef HAVE_SERVER_SUPPORT
1788 void lu_context_tags_update(__u32 tags)
1790 spin_lock(&lu_context_remembered_guard);
1791 lu_context_tags_default |= tags;
1792 atomic_inc(&key_set_version);
1793 spin_unlock(&lu_context_remembered_guard);
1795 EXPORT_SYMBOL(lu_context_tags_update);
1797 void lu_context_tags_clear(__u32 tags)
1799 spin_lock(&lu_context_remembered_guard);
1800 lu_context_tags_default &= ~tags;
1801 atomic_inc(&key_set_version);
1802 spin_unlock(&lu_context_remembered_guard);
1804 EXPORT_SYMBOL(lu_context_tags_clear);
1806 void lu_session_tags_update(__u32 tags)
1808 spin_lock(&lu_context_remembered_guard);
1809 lu_session_tags_default |= tags;
1810 atomic_inc(&key_set_version);
1811 spin_unlock(&lu_context_remembered_guard);
1813 EXPORT_SYMBOL(lu_session_tags_update);
1815 void lu_session_tags_clear(__u32 tags)
1817 spin_lock(&lu_context_remembered_guard);
1818 lu_session_tags_default &= ~tags;
1819 atomic_inc(&key_set_version);
1820 spin_unlock(&lu_context_remembered_guard);
1822 EXPORT_SYMBOL(lu_session_tags_clear);
1823 #endif /* HAVE_SERVER_SUPPORT */
1825 int lu_env_init(struct lu_env *env, __u32 tags)
1830 result = lu_context_init(&env->le_ctx, tags);
1831 if (likely(result == 0))
1832 lu_context_enter(&env->le_ctx);
1835 EXPORT_SYMBOL(lu_env_init);
1837 void lu_env_fini(struct lu_env *env)
1839 lu_context_exit(&env->le_ctx);
1840 lu_context_fini(&env->le_ctx);
1843 EXPORT_SYMBOL(lu_env_fini);
1845 int lu_env_refill(struct lu_env *env)
1849 result = lu_context_refill(&env->le_ctx);
1850 if (result == 0 && env->le_ses != NULL)
1851 result = lu_context_refill(env->le_ses);
1854 EXPORT_SYMBOL(lu_env_refill);
1857 * Currently, this API will only be used by echo client.
1858 * Because echo client and normal lustre client will share
1859 * same cl_env cache. So echo client needs to refresh
1860 * the env context after it get one from the cache, especially
1861 * when normal client and echo client co-exist in the same client.
1863 int lu_env_refill_by_tags(struct lu_env *env, __u32 ctags,
1868 if ((env->le_ctx.lc_tags & ctags) != ctags) {
1869 env->le_ctx.lc_version = 0;
1870 env->le_ctx.lc_tags |= ctags;
1873 if (env->le_ses && (env->le_ses->lc_tags & stags) != stags) {
1874 env->le_ses->lc_version = 0;
1875 env->le_ses->lc_tags |= stags;
1878 result = lu_env_refill(env);
1882 EXPORT_SYMBOL(lu_env_refill_by_tags);
1885 struct lu_env_item {
1886 struct task_struct *lei_task; /* rhashtable key */
1887 struct rhash_head lei_linkage;
1888 struct lu_env *lei_env;
1891 static const struct rhashtable_params lu_env_rhash_params = {
1892 .key_len = sizeof(struct task_struct *),
1893 .key_offset = offsetof(struct lu_env_item, lei_task),
1894 .head_offset = offsetof(struct lu_env_item, lei_linkage),
1897 struct rhashtable lu_env_rhash;
1899 struct lu_env_percpu {
1900 struct task_struct *lep_task;
1901 struct lu_env *lep_env ____cacheline_aligned_in_smp;
1904 static struct lu_env_percpu lu_env_percpu[NR_CPUS];
1906 int lu_env_add(struct lu_env *env)
1908 struct lu_env_item *lei, *old;
1916 lei->lei_task = current;
1919 old = rhashtable_lookup_get_insert_fast(&lu_env_rhash,
1921 lu_env_rhash_params);
1926 EXPORT_SYMBOL(lu_env_add);
1928 void lu_env_remove(struct lu_env *env)
1930 struct lu_env_item *lei;
1931 const void *task = current;
1934 for_each_possible_cpu(i) {
1935 if (lu_env_percpu[i].lep_env == env) {
1936 LASSERT(lu_env_percpu[i].lep_task == task);
1937 lu_env_percpu[i].lep_task = NULL;
1938 lu_env_percpu[i].lep_env = NULL;
1943 lei = rhashtable_lookup_fast(&lu_env_rhash, &task,
1944 lu_env_rhash_params);
1945 if (lei && rhashtable_remove_fast(&lu_env_rhash, &lei->lei_linkage,
1946 lu_env_rhash_params) == 0)
1950 EXPORT_SYMBOL(lu_env_remove);
1952 struct lu_env *lu_env_find(void)
1954 struct lu_env *env = NULL;
1955 struct lu_env_item *lei;
1956 const void *task = current;
1959 if (lu_env_percpu[i].lep_task == current) {
1960 env = lu_env_percpu[i].lep_env;
1966 lei = rhashtable_lookup_fast(&lu_env_rhash, &task,
1967 lu_env_rhash_params);
1970 lu_env_percpu[i].lep_task = current;
1971 lu_env_percpu[i].lep_env = env;
1977 EXPORT_SYMBOL(lu_env_find);
1979 static struct shrinker *lu_site_shrinker;
1981 typedef struct lu_site_stats{
1982 unsigned lss_populated;
1983 unsigned lss_max_search;
1988 static void lu_site_stats_get(const struct lu_site *s,
1989 lu_site_stats_t *stats, int populated)
1991 struct cfs_hash *hs = s->ls_obj_hash;
1992 struct cfs_hash_bd bd;
1995 * percpu_counter_sum_positive() won't accept a const pointer
1996 * as it does modify the struct by taking a spinlock
1998 struct lu_site *s2 = (struct lu_site *)s;
2000 stats->lss_busy += cfs_hash_size_get(hs) -
2001 percpu_counter_sum_positive(&s2->ls_lru_len_counter);
2002 cfs_hash_for_each_bucket(hs, &bd, i) {
2003 struct hlist_head *hhead;
2005 cfs_hash_bd_lock(hs, &bd, 1);
2006 stats->lss_total += cfs_hash_bd_count_get(&bd);
2007 stats->lss_max_search = max((int)stats->lss_max_search,
2008 cfs_hash_bd_depmax_get(&bd));
2010 cfs_hash_bd_unlock(hs, &bd, 1);
2014 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
2015 if (!hlist_empty(hhead))
2016 stats->lss_populated++;
2018 cfs_hash_bd_unlock(hs, &bd, 1);
2024 * lu_cache_shrink_count() returns an approximate number of cached objects
2025 * that can be freed by shrink_slab(). A counter, which tracks the
2026 * number of items in the site's lru, is maintained in a percpu_counter
2027 * for each site. The percpu values are incremented and decremented as
2028 * objects are added or removed from the lru. The percpu values are summed
2029 * and saved whenever a percpu value exceeds a threshold. Thus the saved,
2030 * summed value at any given time may not accurately reflect the current
2031 * lru length. But this value is sufficiently accurate for the needs of
2034 * Using a per cpu counter is a compromise solution to concurrent access:
2035 * lu_object_put() can update the counter without locking the site and
2036 * lu_cache_shrink_count can sum the counters without locking each
2037 * ls_obj_hash bucket.
2039 static unsigned long lu_cache_shrink_count(struct shrinker *sk,
2040 struct shrink_control *sc)
2043 struct lu_site *tmp;
2044 unsigned long cached = 0;
2046 if (!(sc->gfp_mask & __GFP_FS))
2049 down_read(&lu_sites_guard);
2050 list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage)
2051 cached += percpu_counter_read_positive(&s->ls_lru_len_counter);
2052 up_read(&lu_sites_guard);
2054 cached = (cached / 100) * sysctl_vfs_cache_pressure;
2055 CDEBUG(D_INODE, "%ld objects cached, cache pressure %d\n",
2056 cached, sysctl_vfs_cache_pressure);
2061 static unsigned long lu_cache_shrink_scan(struct shrinker *sk,
2062 struct shrink_control *sc)
2065 struct lu_site *tmp;
2066 unsigned long remain = sc->nr_to_scan;
2069 if (!(sc->gfp_mask & __GFP_FS))
2070 /* We must not take the lu_sites_guard lock when
2071 * __GFP_FS is *not* set because of the deadlock
2072 * possibility detailed above. Additionally,
2073 * since we cannot determine the number of
2074 * objects in the cache without taking this
2075 * lock, we're in a particularly tough spot. As
2076 * a result, we'll just lie and say our cache is
2077 * empty. This _should_ be ok, as we can't
2078 * reclaim objects when __GFP_FS is *not* set
2083 down_write(&lu_sites_guard);
2084 list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
2085 remain = lu_site_purge(&lu_shrink_env, s, remain);
2087 * Move just shrunk site to the tail of site list to
2088 * assure shrinking fairness.
2090 list_move_tail(&s->ls_linkage, &splice);
2092 list_splice(&splice, lu_sites.prev);
2093 up_write(&lu_sites_guard);
2095 return sc->nr_to_scan - remain;
2098 #ifndef HAVE_SHRINKER_COUNT
2100 * There exists a potential lock inversion deadlock scenario when using
2101 * Lustre on top of ZFS. This occurs between one of ZFS's
2102 * buf_hash_table.ht_lock's, and Lustre's lu_sites_guard lock. Essentially,
2103 * thread A will take the lu_sites_guard lock and sleep on the ht_lock,
2104 * while thread B will take the ht_lock and sleep on the lu_sites_guard
2105 * lock. Obviously neither thread will wake and drop their respective hold
2108 * To prevent this from happening we must ensure the lu_sites_guard lock is
2109 * not taken while down this code path. ZFS reliably does not set the
2110 * __GFP_FS bit in its code paths, so this can be used to determine if it
2111 * is safe to take the lu_sites_guard lock.
2113 * Ideally we should accurately return the remaining number of cached
2114 * objects without taking the lu_sites_guard lock, but this is not
2115 * possible in the current implementation.
2117 static int lu_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
2120 struct shrink_control scv = {
2121 .nr_to_scan = shrink_param(sc, nr_to_scan),
2122 .gfp_mask = shrink_param(sc, gfp_mask)
2124 #if !defined(HAVE_SHRINKER_WANT_SHRINK_PTR) && !defined(HAVE_SHRINK_CONTROL)
2125 struct shrinker* shrinker = NULL;
2129 CDEBUG(D_INODE, "Shrink %lu objects\n", scv.nr_to_scan);
2131 if (scv.nr_to_scan != 0)
2132 lu_cache_shrink_scan(shrinker, &scv);
2134 cached = lu_cache_shrink_count(shrinker, &scv);
2138 #endif /* HAVE_SHRINKER_COUNT */
2146 * Environment to be used in debugger, contains all tags.
2148 static struct lu_env lu_debugging_env;
2151 * Debugging printer function using printk().
2153 int lu_printk_printer(const struct lu_env *env,
2154 void *unused, const char *format, ...)
2158 va_start(args, format);
2159 vprintk(format, args);
2164 int lu_debugging_setup(void)
2166 return lu_env_init(&lu_debugging_env, ~0);
2169 void lu_context_keys_dump(void)
2173 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
2174 struct lu_context_key *key;
2178 CERROR("[%d]: %p %x (%p,%p,%p) %d %d \"%s\"@%p\n",
2179 i, key, key->lct_tags,
2180 key->lct_init, key->lct_fini, key->lct_exit,
2181 key->lct_index, atomic_read(&key->lct_used),
2182 key->lct_owner ? key->lct_owner->name : "",
2184 lu_ref_print(&key->lct_reference);
2190 * Initialization of global lu_* data.
2192 int lu_global_init(void)
2195 DEF_SHRINKER_VAR(shvar, lu_cache_shrink,
2196 lu_cache_shrink_count, lu_cache_shrink_scan);
2198 CDEBUG(D_INFO, "Lustre LU module (%p).\n", &lu_keys);
2200 result = lu_ref_global_init();
2204 LU_CONTEXT_KEY_INIT(&lu_global_key);
2205 result = lu_context_key_register(&lu_global_key);
2210 * At this level, we don't know what tags are needed, so allocate them
2211 * conservatively. This should not be too bad, because this
2212 * environment is global.
2214 down_write(&lu_sites_guard);
2215 result = lu_env_init(&lu_shrink_env, LCT_SHRINKER);
2216 up_write(&lu_sites_guard);
2221 * seeks estimation: 3 seeks to read a record from oi, one to read
2222 * inode, one for ea. Unfortunately setting this high value results in
2223 * lu_object/inode cache consuming all the memory.
2225 lu_site_shrinker = set_shrinker(DEFAULT_SEEKS, &shvar);
2226 if (lu_site_shrinker == NULL)
2229 result = rhashtable_init(&lu_env_rhash, &lu_env_rhash_params);
2235 * Dual to lu_global_init().
2237 void lu_global_fini(void)
2239 if (lu_site_shrinker != NULL) {
2240 remove_shrinker(lu_site_shrinker);
2241 lu_site_shrinker = NULL;
2244 lu_context_key_degister(&lu_global_key);
2247 * Tear shrinker environment down _after_ de-registering
2248 * lu_global_key, because the latter has a value in the former.
2250 down_write(&lu_sites_guard);
2251 lu_env_fini(&lu_shrink_env);
2252 up_write(&lu_sites_guard);
2254 rhashtable_destroy(&lu_env_rhash);
2256 lu_ref_global_fini();
2259 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx)
2261 #ifdef CONFIG_PROC_FS
2262 struct lprocfs_counter ret;
2264 lprocfs_stats_collect(stats, idx, &ret);
2265 return (__u32)ret.lc_count;
2272 * Output site statistical counters into a buffer. Suitable for
2273 * lprocfs_rd_*()-style functions.
2275 int lu_site_stats_seq_print(const struct lu_site *s, struct seq_file *m)
2277 lu_site_stats_t stats;
2279 memset(&stats, 0, sizeof(stats));
2280 lu_site_stats_get(s, &stats, 1);
2282 seq_printf(m, "%d/%d %d/%d %d %d %d %d %d %d %d\n",
2285 stats.lss_populated,
2286 CFS_HASH_NHLIST(s->ls_obj_hash),
2287 stats.lss_max_search,
2288 ls_stats_read(s->ls_stats, LU_SS_CREATED),
2289 ls_stats_read(s->ls_stats, LU_SS_CACHE_HIT),
2290 ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
2291 ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
2292 ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
2293 ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED));
2296 EXPORT_SYMBOL(lu_site_stats_seq_print);
2299 * Helper function to initialize a number of kmem slab caches at once.
2301 int lu_kmem_init(struct lu_kmem_descr *caches)
2304 struct lu_kmem_descr *iter = caches;
2306 for (result = 0; iter->ckd_cache != NULL; ++iter) {
2307 *iter->ckd_cache = kmem_cache_create(iter->ckd_name,
2310 if (*iter->ckd_cache == NULL) {
2312 /* free all previously allocated caches */
2313 lu_kmem_fini(caches);
2319 EXPORT_SYMBOL(lu_kmem_init);
2322 * Helper function to finalize a number of kmem slab cached at once. Dual to
2325 void lu_kmem_fini(struct lu_kmem_descr *caches)
2327 for (; caches->ckd_cache != NULL; ++caches) {
2328 if (*caches->ckd_cache != NULL) {
2329 kmem_cache_destroy(*caches->ckd_cache);
2330 *caches->ckd_cache = NULL;
2334 EXPORT_SYMBOL(lu_kmem_fini);
2337 * Temporary solution to be able to assign fid in ->do_create()
2338 * till we have fully-functional OST fids
2340 void lu_object_assign_fid(const struct lu_env *env, struct lu_object *o,
2341 const struct lu_fid *fid)
2343 struct lu_site *s = o->lo_dev->ld_site;
2344 struct lu_fid *old = &o->lo_header->loh_fid;
2345 struct cfs_hash *hs;
2346 struct cfs_hash_bd bd;
2348 LASSERT(fid_is_zero(old));
2350 /* supposed to be unique */
2351 hs = s->ls_obj_hash;
2352 cfs_hash_bd_get_and_lock(hs, (void *)fid, &bd, 1);
2353 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
2356 struct lu_object *shadow;
2358 shadow = htable_lookup(s, &bd, fid, &version);
2359 /* supposed to be unique */
2360 LASSERT(IS_ERR(shadow) && PTR_ERR(shadow) == -ENOENT);
2364 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
2365 cfs_hash_bd_unlock(hs, &bd, 1);
2367 EXPORT_SYMBOL(lu_object_assign_fid);
2370 * allocates object with 0 (non-assiged) fid
2371 * XXX: temporary solution to be able to assign fid in ->do_create()
2372 * till we have fully-functional OST fids
2374 struct lu_object *lu_object_anon(const struct lu_env *env,
2375 struct lu_device *dev,
2376 const struct lu_object_conf *conf)
2379 struct lu_object *o;
2382 o = lu_object_alloc(env, dev, &fid, conf);
2386 EXPORT_SYMBOL(lu_object_anon);
2388 struct lu_buf LU_BUF_NULL = {
2392 EXPORT_SYMBOL(LU_BUF_NULL);
2394 void lu_buf_free(struct lu_buf *buf)
2398 LASSERT(buf->lb_len > 0);
2399 OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);
2404 EXPORT_SYMBOL(lu_buf_free);
2406 void lu_buf_alloc(struct lu_buf *buf, size_t size)
2409 LASSERT(buf->lb_buf == NULL);
2410 LASSERT(buf->lb_len == 0);
2411 OBD_ALLOC_LARGE(buf->lb_buf, size);
2412 if (likely(buf->lb_buf))
2415 EXPORT_SYMBOL(lu_buf_alloc);
2417 void lu_buf_realloc(struct lu_buf *buf, size_t size)
2420 lu_buf_alloc(buf, size);
2422 EXPORT_SYMBOL(lu_buf_realloc);
2424 struct lu_buf *lu_buf_check_and_alloc(struct lu_buf *buf, size_t len)
2426 if (buf->lb_buf == NULL && buf->lb_len == 0)
2427 lu_buf_alloc(buf, len);
2429 if ((len > buf->lb_len) && (buf->lb_buf != NULL))
2430 lu_buf_realloc(buf, len);
2434 EXPORT_SYMBOL(lu_buf_check_and_alloc);
2437 * Increase the size of the \a buf.
2438 * preserves old data in buffer
2439 * old buffer remains unchanged on error
2440 * \retval 0 or -ENOMEM
2442 int lu_buf_check_and_grow(struct lu_buf *buf, size_t len)
2446 if (len <= buf->lb_len)
2449 OBD_ALLOC_LARGE(ptr, len);
2453 /* Free the old buf */
2454 if (buf->lb_buf != NULL) {
2455 memcpy(ptr, buf->lb_buf, buf->lb_len);
2456 OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);
2463 EXPORT_SYMBOL(lu_buf_check_and_grow);