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()) or initialized (lu_object_start()).
74 * It is used by lu_object_find() to wait before re-trying when
75 * object in the process of destruction is found in the hash table;
76 * or wait object to be initialized by the allocator.
78 * \see htable_lookup().
80 wait_queue_head_t lsb_waitq;
84 LU_CACHE_PERCENT_MAX = 50,
85 LU_CACHE_PERCENT_DEFAULT = 20
88 #define LU_CACHE_NR_MAX_ADJUST 512
89 #define LU_CACHE_NR_UNLIMITED -1
90 #define LU_CACHE_NR_DEFAULT LU_CACHE_NR_UNLIMITED
91 #define LU_CACHE_NR_LDISKFS_LIMIT LU_CACHE_NR_UNLIMITED
92 /** This is set to roughly (20 * OSS_NTHRS_MAX) to prevent thrashing */
93 #define LU_CACHE_NR_ZFS_LIMIT 10240
95 #define LU_SITE_BITS_MIN 12
96 #define LU_SITE_BITS_MAX 24
97 #define LU_SITE_BITS_MAX_CL 19
99 * total 256 buckets, we don't want too many buckets because:
100 * - consume too much memory
101 * - avoid unbalanced LRU list
103 #define LU_SITE_BKT_BITS 8
106 static unsigned int lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
107 module_param(lu_cache_percent, int, 0644);
108 MODULE_PARM_DESC(lu_cache_percent, "Percentage of memory to be used as lu_object cache");
110 static long lu_cache_nr = LU_CACHE_NR_DEFAULT;
111 module_param(lu_cache_nr, long, 0644);
112 MODULE_PARM_DESC(lu_cache_nr, "Maximum number of objects in lu_object cache");
114 static void lu_object_free(const struct lu_env *env, struct lu_object *o);
115 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx);
118 lu_site_wq_from_fid(struct lu_site *site, struct lu_fid *fid)
120 struct cfs_hash_bd bd;
121 struct lu_site_bkt_data *bkt;
123 cfs_hash_bd_get(site->ls_obj_hash, fid, &bd);
124 bkt = cfs_hash_bd_extra_get(site->ls_obj_hash, &bd);
125 return &bkt->lsb_waitq;
127 EXPORT_SYMBOL(lu_site_wq_from_fid);
130 * Decrease reference counter on object. If last reference is freed, return
131 * object to the cache, unless lu_object_is_dying(o) holds. In the latter
132 * case, free object immediately.
134 void lu_object_put(const struct lu_env *env, struct lu_object *o)
136 struct lu_site_bkt_data *bkt;
137 struct lu_object_header *top = o->lo_header;
138 struct lu_site *site = o->lo_dev->ld_site;
139 struct lu_object *orig = o;
140 struct cfs_hash_bd bd;
141 const struct lu_fid *fid = lu_object_fid(o);
145 * till we have full fids-on-OST implemented anonymous objects
146 * are possible in OSP. such an object isn't listed in the site
147 * so we should not remove it from the site.
149 if (fid_is_zero(fid)) {
150 LASSERT(top->loh_hash.next == NULL
151 && top->loh_hash.pprev == NULL);
152 LASSERT(list_empty(&top->loh_lru));
153 if (!atomic_dec_and_test(&top->loh_ref))
155 list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
156 if (o->lo_ops->loo_object_release != NULL)
157 o->lo_ops->loo_object_release(env, o);
159 lu_object_free(env, orig);
163 cfs_hash_bd_get(site->ls_obj_hash, &top->loh_fid, &bd);
164 bkt = cfs_hash_bd_extra_get(site->ls_obj_hash, &bd);
166 is_dying = lu_object_is_dying(top);
167 if (!cfs_hash_bd_dec_and_lock(site->ls_obj_hash, &bd, &top->loh_ref)) {
168 /* at this point the object reference is dropped and lock is
169 * not taken, so lu_object should not be touched because it
170 * can be freed by concurrent thread. Use local variable for
175 * somebody may be waiting for this, currently only
176 * used for cl_object, see cl_object_put_last().
178 wake_up_all(&bkt->lsb_waitq);
184 * When last reference is released, iterate over object
185 * layers, and notify them that object is no longer busy.
187 list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
188 if (o->lo_ops->loo_object_release != NULL)
189 o->lo_ops->loo_object_release(env, o);
192 /* don't use local 'is_dying' here because if was taken without lock
193 * but here we need the latest actual value of it so check lu_object
196 if (!lu_object_is_dying(top) &&
197 (lu_object_exists(orig) || lu_object_is_cl(orig))) {
198 LASSERT(list_empty(&top->loh_lru));
199 list_add_tail(&top->loh_lru, &bkt->lsb_lru);
200 percpu_counter_inc(&site->ls_lru_len_counter);
201 CDEBUG(D_INODE, "Add %p/%p to site lru. hash: %p, bkt: %p\n",
202 orig, top, site->ls_obj_hash, bkt);
203 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
208 * If object is dying (will not be cached) then remove it
209 * from hash table and LRU.
211 * This is done with hash table and LRU lists locked. As the only
212 * way to acquire first reference to previously unreferenced
213 * object is through hash-table lookup (lu_object_find()),
214 * or LRU scanning (lu_site_purge()), that are done under hash-table
215 * and LRU lock, no race with concurrent object lookup is possible
216 * and we can safely destroy object below.
218 if (!test_and_set_bit(LU_OBJECT_UNHASHED, &top->loh_flags))
219 cfs_hash_bd_del_locked(site->ls_obj_hash, &bd, &top->loh_hash);
220 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
222 * Object was already removed from hash and lru above, can
225 lu_object_free(env, orig);
227 EXPORT_SYMBOL(lu_object_put);
230 * Put object and don't keep in cache. This is temporary solution for
231 * multi-site objects when its layering is not constant.
233 void lu_object_put_nocache(const struct lu_env *env, struct lu_object *o)
235 set_bit(LU_OBJECT_HEARD_BANSHEE, &o->lo_header->loh_flags);
236 return lu_object_put(env, o);
238 EXPORT_SYMBOL(lu_object_put_nocache);
241 * Kill the object and take it out of LRU cache.
242 * Currently used by client code for layout change.
244 void lu_object_unhash(const struct lu_env *env, struct lu_object *o)
246 struct lu_object_header *top;
249 set_bit(LU_OBJECT_HEARD_BANSHEE, &top->loh_flags);
250 if (!test_and_set_bit(LU_OBJECT_UNHASHED, &top->loh_flags)) {
251 struct lu_site *site = o->lo_dev->ld_site;
252 struct cfs_hash *obj_hash = site->ls_obj_hash;
253 struct cfs_hash_bd bd;
255 cfs_hash_bd_get_and_lock(obj_hash, &top->loh_fid, &bd, 1);
256 if (!list_empty(&top->loh_lru)) {
257 struct lu_site_bkt_data *bkt;
259 list_del_init(&top->loh_lru);
260 bkt = cfs_hash_bd_extra_get(obj_hash, &bd);
261 percpu_counter_dec(&site->ls_lru_len_counter);
263 cfs_hash_bd_del_locked(obj_hash, &bd, &top->loh_hash);
264 cfs_hash_bd_unlock(obj_hash, &bd, 1);
267 EXPORT_SYMBOL(lu_object_unhash);
270 * Allocate new object.
272 * This follows object creation protocol, described in the comment within
273 * struct lu_device_operations definition.
275 static struct lu_object *lu_object_alloc(const struct lu_env *env,
276 struct lu_device *dev,
277 const struct lu_fid *f)
279 struct lu_object *top;
282 * Create top-level object slice. This will also create
285 top = dev->ld_ops->ldo_object_alloc(env, NULL, dev);
287 return ERR_PTR(-ENOMEM);
291 * This is the only place where object fid is assigned. It's constant
294 top->lo_header->loh_fid = *f;
302 * This is called after object hash insertion to avoid returning an object with
305 static int lu_object_start(const struct lu_env *env, struct lu_device *dev,
306 struct lu_object *top,
307 const struct lu_object_conf *conf)
309 struct lu_object *scan;
310 struct list_head *layers;
311 unsigned int init_mask = 0;
312 unsigned int init_flag;
316 layers = &top->lo_header->loh_layers;
320 * Call ->loo_object_init() repeatedly, until no more new
321 * object slices are created.
325 list_for_each_entry(scan, layers, lo_linkage) {
326 if (init_mask & init_flag)
329 scan->lo_header = top->lo_header;
330 result = scan->lo_ops->loo_object_init(env, scan, conf);
334 init_mask |= init_flag;
340 list_for_each_entry_reverse(scan, layers, lo_linkage) {
341 if (scan->lo_ops->loo_object_start != NULL) {
342 result = scan->lo_ops->loo_object_start(env, scan);
348 lprocfs_counter_incr(dev->ld_site->ls_stats, LU_SS_CREATED);
350 set_bit(LU_OBJECT_INITED, &top->lo_header->loh_flags);
358 static void lu_object_free(const struct lu_env *env, struct lu_object *o)
360 wait_queue_head_t *wq;
361 struct lu_site *site;
362 struct lu_object *scan;
363 struct list_head *layers;
364 struct list_head splice;
366 site = o->lo_dev->ld_site;
367 layers = &o->lo_header->loh_layers;
368 wq = lu_site_wq_from_fid(site, &o->lo_header->loh_fid);
370 * First call ->loo_object_delete() method to release all resources.
372 list_for_each_entry_reverse(scan, layers, lo_linkage) {
373 if (scan->lo_ops->loo_object_delete != NULL)
374 scan->lo_ops->loo_object_delete(env, scan);
378 * Then, splice object layers into stand-alone list, and call
379 * ->loo_object_free() on all layers to free memory. Splice is
380 * necessary, because lu_object_header is freed together with the
383 INIT_LIST_HEAD(&splice);
384 list_splice_init(layers, &splice);
385 while (!list_empty(&splice)) {
387 * Free layers in bottom-to-top order, so that object header
388 * lives as long as possible and ->loo_object_free() methods
389 * can look at its contents.
391 o = container_of0(splice.prev, struct lu_object, lo_linkage);
392 list_del_init(&o->lo_linkage);
393 LASSERT(o->lo_ops->loo_object_free != NULL);
394 o->lo_ops->loo_object_free(env, o);
397 if (waitqueue_active(wq))
402 * Free \a nr objects from the cold end of the site LRU list.
403 * if canblock is 0, then don't block awaiting for another
404 * instance of lu_site_purge() to complete
406 int lu_site_purge_objects(const struct lu_env *env, struct lu_site *s,
407 int nr, int canblock)
409 struct lu_object_header *h;
410 struct lu_object_header *temp;
411 struct lu_site_bkt_data *bkt;
412 struct cfs_hash_bd bd;
413 struct cfs_hash_bd bd2;
414 struct list_head dispose;
416 unsigned int start = 0;
421 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_NO_LRU))
424 INIT_LIST_HEAD(&dispose);
426 * Under LRU list lock, scan LRU list and move unreferenced objects to
427 * the dispose list, removing them from LRU and hash table.
430 start = s->ls_purge_start;
431 bnr = (nr == ~0) ? -1 : nr / (int)CFS_HASH_NBKT(s->ls_obj_hash) + 1;
434 * It doesn't make any sense to make purge threads parallel, that can
435 * only bring troubles to us. See LU-5331.
438 mutex_lock(&s->ls_purge_mutex);
439 else if (mutex_trylock(&s->ls_purge_mutex) == 0)
443 cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
447 cfs_hash_bd_lock(s->ls_obj_hash, &bd, 1);
448 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
450 list_for_each_entry_safe(h, temp, &bkt->lsb_lru, loh_lru) {
451 LASSERT(atomic_read(&h->loh_ref) == 0);
453 cfs_hash_bd_get(s->ls_obj_hash, &h->loh_fid, &bd2);
454 LASSERT(bd.bd_bucket == bd2.bd_bucket);
456 cfs_hash_bd_del_locked(s->ls_obj_hash,
458 list_move(&h->loh_lru, &dispose);
459 percpu_counter_dec(&s->ls_lru_len_counter);
463 if (nr != ~0 && --nr == 0)
466 if (count > 0 && --count == 0)
470 cfs_hash_bd_unlock(s->ls_obj_hash, &bd, 1);
473 * Free everything on the dispose list. This is safe against
474 * races due to the reasons described in lu_object_put().
476 while (!list_empty(&dispose)) {
477 h = container_of0(dispose.next,
478 struct lu_object_header, loh_lru);
479 list_del_init(&h->loh_lru);
480 lu_object_free(env, lu_object_top(h));
481 lprocfs_counter_incr(s->ls_stats, LU_SS_LRU_PURGED);
487 mutex_unlock(&s->ls_purge_mutex);
489 if (nr != 0 && did_sth && start != 0) {
490 start = 0; /* restart from the first bucket */
493 /* race on s->ls_purge_start, but nobody cares */
494 s->ls_purge_start = i % CFS_HASH_NBKT(s->ls_obj_hash);
499 EXPORT_SYMBOL(lu_site_purge_objects);
504 * Code below has to jump through certain loops to output object description
505 * into libcfs_debug_msg-based log. The problem is that lu_object_print()
506 * composes object description from strings that are parts of _lines_ of
507 * output (i.e., strings that are not terminated by newline). This doesn't fit
508 * very well into libcfs_debug_msg() interface that assumes that each message
509 * supplied to it is a self-contained output line.
511 * To work around this, strings are collected in a temporary buffer
512 * (implemented as a value of lu_cdebug_key key), until terminating newline
513 * character is detected.
521 * XXX overflow is not handled correctly.
526 struct lu_cdebug_data {
530 char lck_area[LU_CDEBUG_LINE];
533 /* context key constructor/destructor: lu_global_key_init, lu_global_key_fini */
534 LU_KEY_INIT_FINI(lu_global, struct lu_cdebug_data);
537 * Key, holding temporary buffer. This key is registered very early by
540 static struct lu_context_key lu_global_key = {
541 .lct_tags = LCT_MD_THREAD | LCT_DT_THREAD |
542 LCT_MG_THREAD | LCT_CL_THREAD | LCT_LOCAL,
543 .lct_init = lu_global_key_init,
544 .lct_fini = lu_global_key_fini
548 * Printer function emitting messages through libcfs_debug_msg().
550 int lu_cdebug_printer(const struct lu_env *env,
551 void *cookie, const char *format, ...)
553 struct libcfs_debug_msg_data *msgdata = cookie;
554 struct lu_cdebug_data *key;
559 va_start(args, format);
561 key = lu_context_key_get(&env->le_ctx, &lu_global_key);
562 LASSERT(key != NULL);
564 used = strlen(key->lck_area);
565 complete = format[strlen(format) - 1] == '\n';
567 * Append new chunk to the buffer.
569 vsnprintf(key->lck_area + used,
570 ARRAY_SIZE(key->lck_area) - used, format, args);
572 if (cfs_cdebug_show(msgdata->msg_mask, msgdata->msg_subsys))
573 libcfs_debug_msg(msgdata, "%s\n", key->lck_area);
574 key->lck_area[0] = 0;
579 EXPORT_SYMBOL(lu_cdebug_printer);
582 * Print object header.
584 void lu_object_header_print(const struct lu_env *env, void *cookie,
585 lu_printer_t printer,
586 const struct lu_object_header *hdr)
588 (*printer)(env, cookie, "header@%p[%#lx, %d, "DFID"%s%s%s]",
589 hdr, hdr->loh_flags, atomic_read(&hdr->loh_ref),
591 hlist_unhashed(&hdr->loh_hash) ? "" : " hash",
592 list_empty((struct list_head *)&hdr->loh_lru) ? \
594 hdr->loh_attr & LOHA_EXISTS ? " exist" : "");
596 EXPORT_SYMBOL(lu_object_header_print);
599 * Print human readable representation of the \a o to the \a printer.
601 void lu_object_print(const struct lu_env *env, void *cookie,
602 lu_printer_t printer, const struct lu_object *o)
604 static const char ruler[] = "........................................";
605 struct lu_object_header *top;
609 lu_object_header_print(env, cookie, printer, top);
610 (*printer)(env, cookie, "{\n");
612 list_for_each_entry(o, &top->loh_layers, lo_linkage) {
614 * print `.' \a depth times followed by type name and address
616 (*printer)(env, cookie, "%*.*s%s@%p", depth, depth, ruler,
617 o->lo_dev->ld_type->ldt_name, o);
619 if (o->lo_ops->loo_object_print != NULL)
620 (*o->lo_ops->loo_object_print)(env, cookie, printer, o);
622 (*printer)(env, cookie, "\n");
625 (*printer)(env, cookie, "} header@%p\n", top);
627 EXPORT_SYMBOL(lu_object_print);
630 * Check object consistency.
632 int lu_object_invariant(const struct lu_object *o)
634 struct lu_object_header *top;
637 list_for_each_entry(o, &top->loh_layers, lo_linkage) {
638 if (o->lo_ops->loo_object_invariant != NULL &&
639 !o->lo_ops->loo_object_invariant(o))
645 static struct lu_object *htable_lookup(struct lu_site *s,
646 struct cfs_hash_bd *bd,
647 const struct lu_fid *f,
650 struct lu_object_header *h;
651 struct hlist_node *hnode;
652 __u64 ver = cfs_hash_bd_version_get(bd);
655 return ERR_PTR(-ENOENT);
658 /* cfs_hash_bd_peek_locked is a somehow "internal" function
659 * of cfs_hash, it doesn't add refcount on object. */
660 hnode = cfs_hash_bd_peek_locked(s->ls_obj_hash, bd, (void *)f);
662 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS);
663 return ERR_PTR(-ENOENT);
666 h = container_of0(hnode, struct lu_object_header, loh_hash);
667 cfs_hash_get(s->ls_obj_hash, hnode);
668 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_HIT);
669 if (!list_empty(&h->loh_lru)) {
670 list_del_init(&h->loh_lru);
671 percpu_counter_dec(&s->ls_lru_len_counter);
673 return lu_object_top(h);
677 * Search cache for an object with the fid \a f. If such object is found,
678 * return it. Otherwise, create new object, insert it into cache and return
679 * it. In any case, additional reference is acquired on the returned object.
681 struct lu_object *lu_object_find(const struct lu_env *env,
682 struct lu_device *dev, const struct lu_fid *f,
683 const struct lu_object_conf *conf)
685 return lu_object_find_at(env, dev->ld_site->ls_top_dev, f, conf);
687 EXPORT_SYMBOL(lu_object_find);
690 * Limit the lu_object cache to a maximum of lu_cache_nr objects. Because
691 * the calculation for the number of objects to reclaim is not covered by
692 * a lock the maximum number of objects is capped by LU_CACHE_MAX_ADJUST.
693 * This ensures that many concurrent threads will not accidentally purge
696 static void lu_object_limit(const struct lu_env *env,
697 struct lu_device *dev)
701 if (lu_cache_nr == LU_CACHE_NR_UNLIMITED)
704 size = cfs_hash_size_get(dev->ld_site->ls_obj_hash);
705 nr = (__u64)lu_cache_nr;
709 lu_site_purge_objects(env, dev->ld_site,
710 MIN(size - nr, LU_CACHE_NR_MAX_ADJUST), 0);
714 * Core logic of lu_object_find*() functions.
716 * Much like lu_object_find(), but top level device of object is specifically
717 * \a dev rather than top level device of the site. This interface allows
718 * objects of different "stacking" to be created within the same site.
720 struct lu_object *lu_object_find_at(const struct lu_env *env,
721 struct lu_device *dev,
722 const struct lu_fid *f,
723 const struct lu_object_conf *conf)
726 struct lu_object *shadow;
729 struct cfs_hash_bd bd;
730 struct lu_site_bkt_data *bkt;
731 struct l_wait_info lwi = { 0 };
738 * This uses standard index maintenance protocol:
740 * - search index under lock, and return object if found;
741 * - otherwise, unlock index, allocate new object;
742 * - lock index and search again;
743 * - if nothing is found (usual case), insert newly created
745 * - otherwise (race: other thread inserted object), free
746 * object just allocated.
750 * For "LOC_F_NEW" case, we are sure the object is new established.
751 * It is unnecessary to perform lookup-alloc-lookup-insert, instead,
752 * just alloc and insert directly.
758 if (unlikely(OBD_FAIL_PRECHECK(OBD_FAIL_OBD_ZERO_NLINK_RACE)))
759 lu_site_purge(env, s, -1);
761 cfs_hash_bd_get(hs, f, &bd);
762 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
763 if (!(conf && conf->loc_flags & LOC_F_NEW)) {
764 cfs_hash_bd_lock(hs, &bd, 1);
765 o = htable_lookup(s, &bd, f, &version);
766 cfs_hash_bd_unlock(hs, &bd, 1);
769 if (likely(lu_object_is_inited(o->lo_header)))
772 l_wait_event(bkt->lsb_waitq,
773 lu_object_is_inited(o->lo_header) ||
774 lu_object_is_dying(o->lo_header), &lwi);
776 if (lu_object_is_dying(o->lo_header)) {
777 lu_object_put(env, o);
779 RETURN(ERR_PTR(-ENOENT));
785 if (PTR_ERR(o) != -ENOENT)
790 * Allocate new object, NB, object is unitialized in case object
791 * is changed between allocation and hash insertion, thus the object
792 * with stale attributes is returned.
794 o = lu_object_alloc(env, dev, f);
798 LASSERT(lu_fid_eq(lu_object_fid(o), f));
800 CFS_RACE_WAIT(OBD_FAIL_OBD_ZERO_NLINK_RACE);
802 cfs_hash_bd_lock(hs, &bd, 1);
804 if (conf && conf->loc_flags & LOC_F_NEW)
805 shadow = ERR_PTR(-ENOENT);
807 shadow = htable_lookup(s, &bd, f, &version);
808 if (likely(PTR_ERR(shadow) == -ENOENT)) {
809 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
810 cfs_hash_bd_unlock(hs, &bd, 1);
813 * This may result in rather complicated operations, including
814 * fld queries, inode loading, etc.
816 rc = lu_object_start(env, dev, o, conf);
818 lu_object_put_nocache(env, o);
822 wake_up_all(&bkt->lsb_waitq);
824 lu_object_limit(env, dev);
829 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_RACE);
830 cfs_hash_bd_unlock(hs, &bd, 1);
831 lu_object_free(env, o);
833 if (!(conf && conf->loc_flags & LOC_F_NEW) &&
834 !lu_object_is_inited(shadow->lo_header)) {
835 l_wait_event(bkt->lsb_waitq,
836 lu_object_is_inited(shadow->lo_header) ||
837 lu_object_is_dying(shadow->lo_header), &lwi);
839 if (lu_object_is_dying(shadow->lo_header)) {
840 lu_object_put(env, shadow);
842 RETURN(ERR_PTR(-ENOENT));
848 EXPORT_SYMBOL(lu_object_find_at);
851 * Find object with given fid, and return its slice belonging to given device.
853 struct lu_object *lu_object_find_slice(const struct lu_env *env,
854 struct lu_device *dev,
855 const struct lu_fid *f,
856 const struct lu_object_conf *conf)
858 struct lu_object *top;
859 struct lu_object *obj;
861 top = lu_object_find(env, dev, f, conf);
865 obj = lu_object_locate(top->lo_header, dev->ld_type);
866 if (unlikely(obj == NULL)) {
867 lu_object_put(env, top);
868 obj = ERR_PTR(-ENOENT);
873 EXPORT_SYMBOL(lu_object_find_slice);
875 int lu_device_type_init(struct lu_device_type *ldt)
879 atomic_set(&ldt->ldt_device_nr, 0);
880 if (ldt->ldt_ops->ldto_init)
881 result = ldt->ldt_ops->ldto_init(ldt);
885 EXPORT_SYMBOL(lu_device_type_init);
887 void lu_device_type_fini(struct lu_device_type *ldt)
889 if (ldt->ldt_ops->ldto_fini)
890 ldt->ldt_ops->ldto_fini(ldt);
892 EXPORT_SYMBOL(lu_device_type_fini);
895 * Global list of all sites on this node
897 static LIST_HEAD(lu_sites);
898 static DECLARE_RWSEM(lu_sites_guard);
901 * Global environment used by site shrinker.
903 static struct lu_env lu_shrink_env;
905 struct lu_site_print_arg {
906 struct lu_env *lsp_env;
908 lu_printer_t lsp_printer;
912 lu_site_obj_print(struct cfs_hash *hs, struct cfs_hash_bd *bd,
913 struct hlist_node *hnode, void *data)
915 struct lu_site_print_arg *arg = (struct lu_site_print_arg *)data;
916 struct lu_object_header *h;
918 h = hlist_entry(hnode, struct lu_object_header, loh_hash);
919 if (!list_empty(&h->loh_layers)) {
920 const struct lu_object *o;
922 o = lu_object_top(h);
923 lu_object_print(arg->lsp_env, arg->lsp_cookie,
924 arg->lsp_printer, o);
926 lu_object_header_print(arg->lsp_env, arg->lsp_cookie,
927 arg->lsp_printer, h);
933 * Print all objects in \a s.
935 void lu_site_print(const struct lu_env *env, struct lu_site *s, void *cookie,
936 lu_printer_t printer)
938 struct lu_site_print_arg arg = {
939 .lsp_env = (struct lu_env *)env,
940 .lsp_cookie = cookie,
941 .lsp_printer = printer,
944 cfs_hash_for_each(s->ls_obj_hash, lu_site_obj_print, &arg);
946 EXPORT_SYMBOL(lu_site_print);
949 * Return desired hash table order.
951 static unsigned long lu_htable_order(struct lu_device *top)
953 unsigned long cache_size;
955 unsigned long bits_max = LU_SITE_BITS_MAX;
958 * For ZFS based OSDs the cache should be disabled by default. This
959 * allows the ZFS ARC maximum flexibility in determining what buffers
960 * to cache. If Lustre has objects or buffer which it wants to ensure
961 * always stay cached it must maintain a hold on them.
963 if (strcmp(top->ld_type->ldt_name, LUSTRE_OSD_ZFS_NAME) == 0) {
964 lu_cache_percent = 1;
965 lu_cache_nr = LU_CACHE_NR_ZFS_LIMIT;
966 return LU_SITE_BITS_MIN;
969 if (strcmp(top->ld_type->ldt_name, LUSTRE_VVP_NAME) == 0)
970 bits_max = LU_SITE_BITS_MAX_CL;
973 * Calculate hash table size, assuming that we want reasonable
974 * performance when 20% of total memory is occupied by cache of
977 * Size of lu_object is (arbitrary) taken as 1K (together with inode).
979 cache_size = cfs_totalram_pages();
981 #if BITS_PER_LONG == 32
982 /* limit hashtable size for lowmem systems to low RAM */
983 if (cache_size > 1 << (30 - PAGE_SHIFT))
984 cache_size = 1 << (30 - PAGE_SHIFT) * 3 / 4;
987 /* clear off unreasonable cache setting. */
988 if (lu_cache_percent == 0 || lu_cache_percent > LU_CACHE_PERCENT_MAX) {
989 CWARN("obdclass: invalid lu_cache_percent: %u, it must be in"
990 " the range of (0, %u]. Will use default value: %u.\n",
991 lu_cache_percent, LU_CACHE_PERCENT_MAX,
992 LU_CACHE_PERCENT_DEFAULT);
994 lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
996 cache_size = cache_size / 100 * lu_cache_percent *
999 for (bits = 1; (1 << bits) < cache_size; ++bits) {
1003 return clamp_t(typeof(bits), bits, LU_SITE_BITS_MIN, bits_max);
1006 static unsigned lu_obj_hop_hash(struct cfs_hash *hs,
1007 const void *key, unsigned mask)
1009 struct lu_fid *fid = (struct lu_fid *)key;
1012 hash = fid_flatten32(fid);
1013 hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
1014 hash = hash_long(hash, hs->hs_bkt_bits);
1016 /* give me another random factor */
1017 hash -= hash_long((unsigned long)hs, fid_oid(fid) % 11 + 3);
1019 hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
1020 hash |= (fid_seq(fid) + fid_oid(fid)) & (CFS_HASH_NBKT(hs) - 1);
1025 static void *lu_obj_hop_object(struct hlist_node *hnode)
1027 return hlist_entry(hnode, struct lu_object_header, loh_hash);
1030 static void *lu_obj_hop_key(struct hlist_node *hnode)
1032 struct lu_object_header *h;
1034 h = hlist_entry(hnode, struct lu_object_header, loh_hash);
1038 static int lu_obj_hop_keycmp(const void *key, struct hlist_node *hnode)
1040 struct lu_object_header *h;
1042 h = hlist_entry(hnode, struct lu_object_header, loh_hash);
1043 return lu_fid_eq(&h->loh_fid, (struct lu_fid *)key);
1046 static void lu_obj_hop_get(struct cfs_hash *hs, struct hlist_node *hnode)
1048 struct lu_object_header *h;
1050 h = hlist_entry(hnode, struct lu_object_header, loh_hash);
1051 atomic_inc(&h->loh_ref);
1054 static void lu_obj_hop_put_locked(struct cfs_hash *hs, struct hlist_node *hnode)
1056 LBUG(); /* we should never called it */
1059 static struct cfs_hash_ops lu_site_hash_ops = {
1060 .hs_hash = lu_obj_hop_hash,
1061 .hs_key = lu_obj_hop_key,
1062 .hs_keycmp = lu_obj_hop_keycmp,
1063 .hs_object = lu_obj_hop_object,
1064 .hs_get = lu_obj_hop_get,
1065 .hs_put_locked = lu_obj_hop_put_locked,
1068 void lu_dev_add_linkage(struct lu_site *s, struct lu_device *d)
1070 spin_lock(&s->ls_ld_lock);
1071 if (list_empty(&d->ld_linkage))
1072 list_add(&d->ld_linkage, &s->ls_ld_linkage);
1073 spin_unlock(&s->ls_ld_lock);
1075 EXPORT_SYMBOL(lu_dev_add_linkage);
1077 void lu_dev_del_linkage(struct lu_site *s, struct lu_device *d)
1079 spin_lock(&s->ls_ld_lock);
1080 list_del_init(&d->ld_linkage);
1081 spin_unlock(&s->ls_ld_lock);
1083 EXPORT_SYMBOL(lu_dev_del_linkage);
1086 * Initialize site \a s, with \a d as the top level device.
1088 int lu_site_init(struct lu_site *s, struct lu_device *top)
1090 struct lu_site_bkt_data *bkt;
1091 struct cfs_hash_bd bd;
1098 memset(s, 0, sizeof *s);
1099 mutex_init(&s->ls_purge_mutex);
1101 #ifdef HAVE_PERCPU_COUNTER_INIT_GFP_FLAG
1102 rc = percpu_counter_init(&s->ls_lru_len_counter, 0, GFP_NOFS);
1104 rc = percpu_counter_init(&s->ls_lru_len_counter, 0);
1109 snprintf(name, sizeof(name), "lu_site_%s", top->ld_type->ldt_name);
1110 for (bits = lu_htable_order(top);
1111 bits >= LU_SITE_BITS_MIN; bits--) {
1112 s->ls_obj_hash = cfs_hash_create(name, bits, bits,
1113 bits - LU_SITE_BKT_BITS,
1116 CFS_HASH_SPIN_BKTLOCK |
1117 CFS_HASH_NO_ITEMREF |
1119 CFS_HASH_ASSERT_EMPTY |
1121 if (s->ls_obj_hash != NULL)
1125 if (s->ls_obj_hash == NULL) {
1126 CERROR("failed to create lu_site hash with bits: %lu\n", bits);
1130 cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
1131 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
1132 INIT_LIST_HEAD(&bkt->lsb_lru);
1133 init_waitqueue_head(&bkt->lsb_waitq);
1136 s->ls_stats = lprocfs_alloc_stats(LU_SS_LAST_STAT, 0);
1137 if (s->ls_stats == NULL) {
1138 cfs_hash_putref(s->ls_obj_hash);
1139 s->ls_obj_hash = NULL;
1143 lprocfs_counter_init(s->ls_stats, LU_SS_CREATED,
1144 0, "created", "created");
1145 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_HIT,
1146 0, "cache_hit", "cache_hit");
1147 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_MISS,
1148 0, "cache_miss", "cache_miss");
1149 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_RACE,
1150 0, "cache_race", "cache_race");
1151 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_DEATH_RACE,
1152 0, "cache_death_race", "cache_death_race");
1153 lprocfs_counter_init(s->ls_stats, LU_SS_LRU_PURGED,
1154 0, "lru_purged", "lru_purged");
1156 INIT_LIST_HEAD(&s->ls_linkage);
1157 s->ls_top_dev = top;
1160 lu_ref_add(&top->ld_reference, "site-top", s);
1162 INIT_LIST_HEAD(&s->ls_ld_linkage);
1163 spin_lock_init(&s->ls_ld_lock);
1165 lu_dev_add_linkage(s, top);
1169 EXPORT_SYMBOL(lu_site_init);
1172 * Finalize \a s and release its resources.
1174 void lu_site_fini(struct lu_site *s)
1176 down_write(&lu_sites_guard);
1177 list_del_init(&s->ls_linkage);
1178 up_write(&lu_sites_guard);
1180 percpu_counter_destroy(&s->ls_lru_len_counter);
1182 if (s->ls_obj_hash != NULL) {
1183 cfs_hash_putref(s->ls_obj_hash);
1184 s->ls_obj_hash = NULL;
1187 if (s->ls_top_dev != NULL) {
1188 s->ls_top_dev->ld_site = NULL;
1189 lu_ref_del(&s->ls_top_dev->ld_reference, "site-top", s);
1190 lu_device_put(s->ls_top_dev);
1191 s->ls_top_dev = NULL;
1194 if (s->ls_stats != NULL)
1195 lprocfs_free_stats(&s->ls_stats);
1197 EXPORT_SYMBOL(lu_site_fini);
1200 * Called when initialization of stack for this site is completed.
1202 int lu_site_init_finish(struct lu_site *s)
1205 down_write(&lu_sites_guard);
1206 result = lu_context_refill(&lu_shrink_env.le_ctx);
1208 list_add(&s->ls_linkage, &lu_sites);
1209 up_write(&lu_sites_guard);
1212 EXPORT_SYMBOL(lu_site_init_finish);
1215 * Acquire additional reference on device \a d
1217 void lu_device_get(struct lu_device *d)
1219 atomic_inc(&d->ld_ref);
1221 EXPORT_SYMBOL(lu_device_get);
1224 * Release reference on device \a d.
1226 void lu_device_put(struct lu_device *d)
1228 LASSERT(atomic_read(&d->ld_ref) > 0);
1229 atomic_dec(&d->ld_ref);
1231 EXPORT_SYMBOL(lu_device_put);
1234 * Initialize device \a d of type \a t.
1236 int lu_device_init(struct lu_device *d, struct lu_device_type *t)
1238 if (atomic_inc_return(&t->ldt_device_nr) == 1 &&
1239 t->ldt_ops->ldto_start != NULL)
1240 t->ldt_ops->ldto_start(t);
1242 memset(d, 0, sizeof *d);
1244 lu_ref_init(&d->ld_reference);
1245 INIT_LIST_HEAD(&d->ld_linkage);
1249 EXPORT_SYMBOL(lu_device_init);
1252 * Finalize device \a d.
1254 void lu_device_fini(struct lu_device *d)
1256 struct lu_device_type *t = d->ld_type;
1258 if (d->ld_obd != NULL) {
1259 d->ld_obd->obd_lu_dev = NULL;
1263 lu_ref_fini(&d->ld_reference);
1264 LASSERTF(atomic_read(&d->ld_ref) == 0,
1265 "Refcount is %u\n", atomic_read(&d->ld_ref));
1266 LASSERT(atomic_read(&t->ldt_device_nr) > 0);
1268 if (atomic_dec_and_test(&t->ldt_device_nr) &&
1269 t->ldt_ops->ldto_stop != NULL)
1270 t->ldt_ops->ldto_stop(t);
1272 EXPORT_SYMBOL(lu_device_fini);
1275 * Initialize object \a o that is part of compound object \a h and was created
1278 int lu_object_init(struct lu_object *o, struct lu_object_header *h,
1279 struct lu_device *d)
1281 memset(o, 0, sizeof(*o));
1285 lu_ref_add_at(&d->ld_reference, &o->lo_dev_ref, "lu_object", o);
1286 INIT_LIST_HEAD(&o->lo_linkage);
1290 EXPORT_SYMBOL(lu_object_init);
1293 * Finalize object and release its resources.
1295 void lu_object_fini(struct lu_object *o)
1297 struct lu_device *dev = o->lo_dev;
1299 LASSERT(list_empty(&o->lo_linkage));
1302 lu_ref_del_at(&dev->ld_reference, &o->lo_dev_ref,
1308 EXPORT_SYMBOL(lu_object_fini);
1311 * Add object \a o as first layer of compound object \a h
1313 * This is typically called by the ->ldo_object_alloc() method of top-level
1316 void lu_object_add_top(struct lu_object_header *h, struct lu_object *o)
1318 list_move(&o->lo_linkage, &h->loh_layers);
1320 EXPORT_SYMBOL(lu_object_add_top);
1323 * Add object \a o as a layer of compound object, going after \a before.
1325 * This is typically called by the ->ldo_object_alloc() method of \a
1328 void lu_object_add(struct lu_object *before, struct lu_object *o)
1330 list_move(&o->lo_linkage, &before->lo_linkage);
1332 EXPORT_SYMBOL(lu_object_add);
1335 * Initialize compound object.
1337 int lu_object_header_init(struct lu_object_header *h)
1339 memset(h, 0, sizeof *h);
1340 atomic_set(&h->loh_ref, 1);
1341 INIT_HLIST_NODE(&h->loh_hash);
1342 INIT_LIST_HEAD(&h->loh_lru);
1343 INIT_LIST_HEAD(&h->loh_layers);
1344 lu_ref_init(&h->loh_reference);
1347 EXPORT_SYMBOL(lu_object_header_init);
1350 * Finalize compound object.
1352 void lu_object_header_fini(struct lu_object_header *h)
1354 LASSERT(list_empty(&h->loh_layers));
1355 LASSERT(list_empty(&h->loh_lru));
1356 LASSERT(hlist_unhashed(&h->loh_hash));
1357 lu_ref_fini(&h->loh_reference);
1359 EXPORT_SYMBOL(lu_object_header_fini);
1362 * Given a compound object, find its slice, corresponding to the device type
1365 struct lu_object *lu_object_locate(struct lu_object_header *h,
1366 const struct lu_device_type *dtype)
1368 struct lu_object *o;
1370 list_for_each_entry(o, &h->loh_layers, lo_linkage) {
1371 if (o->lo_dev->ld_type == dtype)
1376 EXPORT_SYMBOL(lu_object_locate);
1379 * Finalize and free devices in the device stack.
1381 * Finalize device stack by purging object cache, and calling
1382 * lu_device_type_operations::ldto_device_fini() and
1383 * lu_device_type_operations::ldto_device_free() on all devices in the stack.
1385 void lu_stack_fini(const struct lu_env *env, struct lu_device *top)
1387 struct lu_site *site = top->ld_site;
1388 struct lu_device *scan;
1389 struct lu_device *next;
1391 lu_site_purge(env, site, ~0);
1392 for (scan = top; scan != NULL; scan = next) {
1393 next = scan->ld_type->ldt_ops->ldto_device_fini(env, scan);
1394 lu_ref_del(&scan->ld_reference, "lu-stack", &lu_site_init);
1395 lu_device_put(scan);
1399 lu_site_purge(env, site, ~0);
1401 for (scan = top; scan != NULL; scan = next) {
1402 const struct lu_device_type *ldt = scan->ld_type;
1403 struct obd_type *type;
1405 next = ldt->ldt_ops->ldto_device_free(env, scan);
1406 type = ldt->ldt_obd_type;
1409 class_put_type(type);
1416 * Maximal number of tld slots.
1418 LU_CONTEXT_KEY_NR = 40
1421 static struct lu_context_key *lu_keys[LU_CONTEXT_KEY_NR] = { NULL, };
1423 static DECLARE_RWSEM(lu_key_initing);
1426 * Global counter incremented whenever key is registered, unregistered,
1427 * revived or quiesced. This is used to void unnecessary calls to
1428 * lu_context_refill(). No locking is provided, as initialization and shutdown
1429 * are supposed to be externally serialized.
1431 static atomic_t key_set_version = ATOMIC_INIT(0);
1436 int lu_context_key_register(struct lu_context_key *key)
1441 LASSERT(key->lct_init != NULL);
1442 LASSERT(key->lct_fini != NULL);
1443 LASSERT(key->lct_tags != 0);
1444 LASSERT(key->lct_owner != NULL);
1447 atomic_set(&key->lct_used, 1);
1448 lu_ref_init(&key->lct_reference);
1449 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1453 if (cmpxchg(&lu_keys[i], NULL, key) != NULL)
1457 atomic_inc(&key_set_version);
1461 lu_ref_fini(&key->lct_reference);
1462 atomic_set(&key->lct_used, 0);
1466 EXPORT_SYMBOL(lu_context_key_register);
1468 static void key_fini(struct lu_context *ctx, int index)
1470 if (ctx->lc_value != NULL && ctx->lc_value[index] != NULL) {
1471 struct lu_context_key *key;
1473 key = lu_keys[index];
1474 LASSERT(key != NULL);
1475 LASSERT(key->lct_fini != NULL);
1476 LASSERT(atomic_read(&key->lct_used) > 0);
1478 key->lct_fini(ctx, key, ctx->lc_value[index]);
1479 lu_ref_del(&key->lct_reference, "ctx", ctx);
1480 if (atomic_dec_and_test(&key->lct_used))
1481 wake_up_var(&key->lct_used);
1483 LASSERT(key->lct_owner != NULL);
1484 if ((ctx->lc_tags & LCT_NOREF) == 0) {
1485 LINVRNT(module_refcount(key->lct_owner) > 0);
1486 module_put(key->lct_owner);
1488 ctx->lc_value[index] = NULL;
1495 void lu_context_key_degister(struct lu_context_key *key)
1497 LASSERT(atomic_read(&key->lct_used) >= 1);
1498 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1500 lu_context_key_quiesce(key);
1502 key_fini(&lu_shrink_env.le_ctx, key->lct_index);
1505 * Wait until all transient contexts referencing this key have
1506 * run lu_context_key::lct_fini() method.
1508 atomic_dec(&key->lct_used);
1509 wait_var_event(&key->lct_used, atomic_read(&key->lct_used) == 0);
1511 if (!WARN_ON(lu_keys[key->lct_index] == NULL))
1512 lu_ref_fini(&key->lct_reference);
1514 smp_store_release(&lu_keys[key->lct_index], NULL);
1516 EXPORT_SYMBOL(lu_context_key_degister);
1519 * Register a number of keys. This has to be called after all keys have been
1520 * initialized by a call to LU_CONTEXT_KEY_INIT().
1522 int lu_context_key_register_many(struct lu_context_key *k, ...)
1524 struct lu_context_key *key = k;
1530 result = lu_context_key_register(key);
1533 key = va_arg(args, struct lu_context_key *);
1534 } while (key != NULL);
1540 lu_context_key_degister(k);
1541 k = va_arg(args, struct lu_context_key *);
1548 EXPORT_SYMBOL(lu_context_key_register_many);
1551 * De-register a number of keys. This is a dual to
1552 * lu_context_key_register_many().
1554 void lu_context_key_degister_many(struct lu_context_key *k, ...)
1560 lu_context_key_degister(k);
1561 k = va_arg(args, struct lu_context_key*);
1562 } while (k != NULL);
1565 EXPORT_SYMBOL(lu_context_key_degister_many);
1568 * Revive a number of keys.
1570 void lu_context_key_revive_many(struct lu_context_key *k, ...)
1576 lu_context_key_revive(k);
1577 k = va_arg(args, struct lu_context_key*);
1578 } while (k != NULL);
1581 EXPORT_SYMBOL(lu_context_key_revive_many);
1584 * Quiescent a number of keys.
1586 void lu_context_key_quiesce_many(struct lu_context_key *k, ...)
1592 lu_context_key_quiesce(k);
1593 k = va_arg(args, struct lu_context_key*);
1594 } while (k != NULL);
1597 EXPORT_SYMBOL(lu_context_key_quiesce_many);
1600 * Return value associated with key \a key in context \a ctx.
1602 void *lu_context_key_get(const struct lu_context *ctx,
1603 const struct lu_context_key *key)
1605 LINVRNT(ctx->lc_state == LCS_ENTERED);
1606 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1607 LASSERT(lu_keys[key->lct_index] == key);
1608 return ctx->lc_value[key->lct_index];
1610 EXPORT_SYMBOL(lu_context_key_get);
1613 * List of remembered contexts. XXX document me.
1615 static LIST_HEAD(lu_context_remembered);
1616 static DEFINE_SPINLOCK(lu_context_remembered_guard);
1619 * Destroy \a key in all remembered contexts. This is used to destroy key
1620 * values in "shared" contexts (like service threads), when a module owning
1621 * the key is about to be unloaded.
1623 void lu_context_key_quiesce(struct lu_context_key *key)
1625 struct lu_context *ctx;
1627 if (!(key->lct_tags & LCT_QUIESCENT)) {
1629 * The write-lock on lu_key_initing will ensure that any
1630 * keys_fill() which didn't see LCT_QUIESCENT will have
1631 * finished before we call key_fini().
1633 down_write(&lu_key_initing);
1634 key->lct_tags |= LCT_QUIESCENT;
1635 up_write(&lu_key_initing);
1637 spin_lock(&lu_context_remembered_guard);
1638 list_for_each_entry(ctx, &lu_context_remembered, lc_remember) {
1639 spin_until_cond(READ_ONCE(ctx->lc_state) != LCS_LEAVING);
1640 key_fini(ctx, key->lct_index);
1643 spin_unlock(&lu_context_remembered_guard);
1647 void lu_context_key_revive(struct lu_context_key *key)
1649 key->lct_tags &= ~LCT_QUIESCENT;
1650 atomic_inc(&key_set_version);
1653 static void keys_fini(struct lu_context *ctx)
1657 if (ctx->lc_value == NULL)
1660 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i)
1663 OBD_FREE(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1664 ctx->lc_value = NULL;
1667 static int keys_fill(struct lu_context *ctx)
1673 * A serialisation with lu_context_key_quiesce() is needed, to
1674 * ensure we see LCT_QUIESCENT and don't allocate a new value
1675 * after it freed one. The rwsem provides this. As down_read()
1676 * does optimistic spinning while the writer is active, this is
1677 * unlikely to ever sleep.
1679 down_read(&lu_key_initing);
1680 ctx->lc_version = atomic_read(&key_set_version);
1682 LINVRNT(ctx->lc_value);
1683 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1684 struct lu_context_key *key;
1687 if (!ctx->lc_value[i] && key &&
1688 (key->lct_tags & ctx->lc_tags) &&
1690 * Don't create values for a LCT_QUIESCENT key, as this
1691 * will pin module owning a key.
1693 !(key->lct_tags & LCT_QUIESCENT)) {
1696 LINVRNT(key->lct_init != NULL);
1697 LINVRNT(key->lct_index == i);
1699 LASSERT(key->lct_owner != NULL);
1700 if (!(ctx->lc_tags & LCT_NOREF) &&
1701 try_module_get(key->lct_owner) == 0) {
1702 /* module is unloading, skip this key */
1706 value = key->lct_init(ctx, key);
1707 if (unlikely(IS_ERR(value))) {
1708 rc = PTR_ERR(value);
1712 lu_ref_add_atomic(&key->lct_reference, "ctx", ctx);
1713 atomic_inc(&key->lct_used);
1715 * This is the only place in the code, where an
1716 * element of ctx->lc_value[] array is set to non-NULL
1719 ctx->lc_value[i] = value;
1720 if (key->lct_exit != NULL)
1721 ctx->lc_tags |= LCT_HAS_EXIT;
1725 up_read(&lu_key_initing);
1729 static int keys_init(struct lu_context *ctx)
1731 OBD_ALLOC(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1732 if (likely(ctx->lc_value != NULL))
1733 return keys_fill(ctx);
1739 * Initialize context data-structure. Create values for all keys.
1741 int lu_context_init(struct lu_context *ctx, __u32 tags)
1745 memset(ctx, 0, sizeof *ctx);
1746 ctx->lc_state = LCS_INITIALIZED;
1747 ctx->lc_tags = tags;
1748 if (tags & LCT_REMEMBER) {
1749 spin_lock(&lu_context_remembered_guard);
1750 list_add(&ctx->lc_remember, &lu_context_remembered);
1751 spin_unlock(&lu_context_remembered_guard);
1753 INIT_LIST_HEAD(&ctx->lc_remember);
1756 rc = keys_init(ctx);
1758 lu_context_fini(ctx);
1762 EXPORT_SYMBOL(lu_context_init);
1765 * Finalize context data-structure. Destroy key values.
1767 void lu_context_fini(struct lu_context *ctx)
1769 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1770 ctx->lc_state = LCS_FINALIZED;
1772 if ((ctx->lc_tags & LCT_REMEMBER) == 0) {
1773 LASSERT(list_empty(&ctx->lc_remember));
1775 /* could race with key degister */
1776 spin_lock(&lu_context_remembered_guard);
1777 list_del_init(&ctx->lc_remember);
1778 spin_unlock(&lu_context_remembered_guard);
1782 EXPORT_SYMBOL(lu_context_fini);
1785 * Called before entering context.
1787 void lu_context_enter(struct lu_context *ctx)
1789 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1790 ctx->lc_state = LCS_ENTERED;
1792 EXPORT_SYMBOL(lu_context_enter);
1795 * Called after exiting from \a ctx
1797 void lu_context_exit(struct lu_context *ctx)
1801 LINVRNT(ctx->lc_state == LCS_ENTERED);
1803 * Disable preempt to ensure we get a warning if
1804 * any lct_exit ever tries to sleep. That would hurt
1805 * lu_context_key_quiesce() which spins waiting for us.
1806 * This also ensure we aren't preempted while the state
1807 * is LCS_LEAVING, as that too would cause problems for
1808 * lu_context_key_quiesce().
1812 * Ensure lu_context_key_quiesce() sees LCS_LEAVING
1813 * or we see LCT_QUIESCENT
1815 smp_store_mb(ctx->lc_state, LCS_LEAVING);
1816 if (ctx->lc_tags & LCT_HAS_EXIT && ctx->lc_value) {
1817 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1818 struct lu_context_key *key;
1821 if (ctx->lc_value[i] &&
1822 !(key->lct_tags & LCT_QUIESCENT) &&
1824 key->lct_exit(ctx, key, ctx->lc_value[i]);
1828 smp_store_release(&ctx->lc_state, LCS_LEFT);
1831 EXPORT_SYMBOL(lu_context_exit);
1834 * Allocate for context all missing keys that were registered after context
1835 * creation. key_set_version is only changed in rare cases when modules
1836 * are loaded and removed.
1838 int lu_context_refill(struct lu_context *ctx)
1840 if (likely(ctx->lc_version == atomic_read(&key_set_version)))
1843 return keys_fill(ctx);
1847 * lu_ctx_tags/lu_ses_tags will be updated if there are new types of
1848 * obd being added. Currently, this is only used on client side, specifically
1849 * for echo device client, for other stack (like ptlrpc threads), context are
1850 * predefined when the lu_device type are registered, during the module probe
1853 u32 lu_context_tags_default = LCT_CL_THREAD;
1854 u32 lu_session_tags_default = LCT_SESSION;
1856 void lu_context_tags_update(__u32 tags)
1858 spin_lock(&lu_context_remembered_guard);
1859 lu_context_tags_default |= tags;
1860 atomic_inc(&key_set_version);
1861 spin_unlock(&lu_context_remembered_guard);
1863 EXPORT_SYMBOL(lu_context_tags_update);
1865 void lu_context_tags_clear(__u32 tags)
1867 spin_lock(&lu_context_remembered_guard);
1868 lu_context_tags_default &= ~tags;
1869 atomic_inc(&key_set_version);
1870 spin_unlock(&lu_context_remembered_guard);
1872 EXPORT_SYMBOL(lu_context_tags_clear);
1874 void lu_session_tags_update(__u32 tags)
1876 spin_lock(&lu_context_remembered_guard);
1877 lu_session_tags_default |= tags;
1878 atomic_inc(&key_set_version);
1879 spin_unlock(&lu_context_remembered_guard);
1881 EXPORT_SYMBOL(lu_session_tags_update);
1883 void lu_session_tags_clear(__u32 tags)
1885 spin_lock(&lu_context_remembered_guard);
1886 lu_session_tags_default &= ~tags;
1887 atomic_inc(&key_set_version);
1888 spin_unlock(&lu_context_remembered_guard);
1890 EXPORT_SYMBOL(lu_session_tags_clear);
1892 int lu_env_init(struct lu_env *env, __u32 tags)
1897 result = lu_context_init(&env->le_ctx, tags);
1898 if (likely(result == 0))
1899 lu_context_enter(&env->le_ctx);
1902 EXPORT_SYMBOL(lu_env_init);
1904 void lu_env_fini(struct lu_env *env)
1906 lu_context_exit(&env->le_ctx);
1907 lu_context_fini(&env->le_ctx);
1910 EXPORT_SYMBOL(lu_env_fini);
1912 int lu_env_refill(struct lu_env *env)
1916 result = lu_context_refill(&env->le_ctx);
1917 if (result == 0 && env->le_ses != NULL)
1918 result = lu_context_refill(env->le_ses);
1921 EXPORT_SYMBOL(lu_env_refill);
1924 * Currently, this API will only be used by echo client.
1925 * Because echo client and normal lustre client will share
1926 * same cl_env cache. So echo client needs to refresh
1927 * the env context after it get one from the cache, especially
1928 * when normal client and echo client co-exist in the same client.
1930 int lu_env_refill_by_tags(struct lu_env *env, __u32 ctags,
1935 if ((env->le_ctx.lc_tags & ctags) != ctags) {
1936 env->le_ctx.lc_version = 0;
1937 env->le_ctx.lc_tags |= ctags;
1940 if (env->le_ses && (env->le_ses->lc_tags & stags) != stags) {
1941 env->le_ses->lc_version = 0;
1942 env->le_ses->lc_tags |= stags;
1945 result = lu_env_refill(env);
1949 EXPORT_SYMBOL(lu_env_refill_by_tags);
1952 struct lu_env_item {
1953 struct task_struct *lei_task; /* rhashtable key */
1954 struct rhash_head lei_linkage;
1955 struct lu_env *lei_env;
1956 struct rcu_head lei_rcu_head;
1959 static const struct rhashtable_params lu_env_rhash_params = {
1960 .key_len = sizeof(struct task_struct *),
1961 .key_offset = offsetof(struct lu_env_item, lei_task),
1962 .head_offset = offsetof(struct lu_env_item, lei_linkage),
1965 struct rhashtable lu_env_rhash;
1967 struct lu_env_percpu {
1968 struct task_struct *lep_task;
1969 struct lu_env *lep_env ____cacheline_aligned_in_smp;
1972 static struct lu_env_percpu lu_env_percpu[NR_CPUS];
1974 int lu_env_add(struct lu_env *env)
1976 struct lu_env_item *lei, *old;
1984 lei->lei_task = current;
1987 old = rhashtable_lookup_get_insert_fast(&lu_env_rhash,
1989 lu_env_rhash_params);
1994 EXPORT_SYMBOL(lu_env_add);
1996 static void lu_env_item_free(struct rcu_head *head)
1998 struct lu_env_item *lei;
2000 lei = container_of(head, struct lu_env_item, lei_rcu_head);
2004 void lu_env_remove(struct lu_env *env)
2006 struct lu_env_item *lei;
2007 const void *task = current;
2010 for_each_possible_cpu(i) {
2011 if (lu_env_percpu[i].lep_env == env) {
2012 LASSERT(lu_env_percpu[i].lep_task == task);
2013 lu_env_percpu[i].lep_task = NULL;
2014 lu_env_percpu[i].lep_env = NULL;
2018 /* The rcu_lock is not taking in this case since the key
2019 * used is the actual task_struct. This implies that each
2020 * object is only removed by the owning thread, so there
2021 * can never be a race on a particular object.
2023 lei = rhashtable_lookup_fast(&lu_env_rhash, &task,
2024 lu_env_rhash_params);
2025 if (lei && rhashtable_remove_fast(&lu_env_rhash, &lei->lei_linkage,
2026 lu_env_rhash_params) == 0)
2027 call_rcu(&lei->lei_rcu_head, lu_env_item_free);
2029 EXPORT_SYMBOL(lu_env_remove);
2031 struct lu_env *lu_env_find(void)
2033 struct lu_env *env = NULL;
2034 struct lu_env_item *lei;
2035 const void *task = current;
2038 if (lu_env_percpu[i].lep_task == current) {
2039 env = lu_env_percpu[i].lep_env;
2045 lei = rhashtable_lookup_fast(&lu_env_rhash, &task,
2046 lu_env_rhash_params);
2049 lu_env_percpu[i].lep_task = current;
2050 lu_env_percpu[i].lep_env = env;
2056 EXPORT_SYMBOL(lu_env_find);
2058 static struct shrinker *lu_site_shrinker;
2060 typedef struct lu_site_stats{
2061 unsigned lss_populated;
2062 unsigned lss_max_search;
2067 static void lu_site_stats_get(const struct lu_site *s,
2068 lu_site_stats_t *stats, int populated)
2070 struct cfs_hash *hs = s->ls_obj_hash;
2071 struct cfs_hash_bd bd;
2074 * percpu_counter_sum_positive() won't accept a const pointer
2075 * as it does modify the struct by taking a spinlock
2077 struct lu_site *s2 = (struct lu_site *)s;
2079 stats->lss_busy += cfs_hash_size_get(hs) -
2080 percpu_counter_sum_positive(&s2->ls_lru_len_counter);
2081 cfs_hash_for_each_bucket(hs, &bd, i) {
2082 struct hlist_head *hhead;
2084 cfs_hash_bd_lock(hs, &bd, 1);
2085 stats->lss_total += cfs_hash_bd_count_get(&bd);
2086 stats->lss_max_search = max((int)stats->lss_max_search,
2087 cfs_hash_bd_depmax_get(&bd));
2089 cfs_hash_bd_unlock(hs, &bd, 1);
2093 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
2094 if (!hlist_empty(hhead))
2095 stats->lss_populated++;
2097 cfs_hash_bd_unlock(hs, &bd, 1);
2103 * lu_cache_shrink_count() returns an approximate number of cached objects
2104 * that can be freed by shrink_slab(). A counter, which tracks the
2105 * number of items in the site's lru, is maintained in a percpu_counter
2106 * for each site. The percpu values are incremented and decremented as
2107 * objects are added or removed from the lru. The percpu values are summed
2108 * and saved whenever a percpu value exceeds a threshold. Thus the saved,
2109 * summed value at any given time may not accurately reflect the current
2110 * lru length. But this value is sufficiently accurate for the needs of
2113 * Using a per cpu counter is a compromise solution to concurrent access:
2114 * lu_object_put() can update the counter without locking the site and
2115 * lu_cache_shrink_count can sum the counters without locking each
2116 * ls_obj_hash bucket.
2118 static unsigned long lu_cache_shrink_count(struct shrinker *sk,
2119 struct shrink_control *sc)
2122 struct lu_site *tmp;
2123 unsigned long cached = 0;
2125 if (!(sc->gfp_mask & __GFP_FS))
2128 down_read(&lu_sites_guard);
2129 list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage)
2130 cached += percpu_counter_read_positive(&s->ls_lru_len_counter);
2131 up_read(&lu_sites_guard);
2133 cached = (cached / 100) * sysctl_vfs_cache_pressure;
2134 CDEBUG(D_INODE, "%ld objects cached, cache pressure %d\n",
2135 cached, sysctl_vfs_cache_pressure);
2140 static unsigned long lu_cache_shrink_scan(struct shrinker *sk,
2141 struct shrink_control *sc)
2144 struct lu_site *tmp;
2145 unsigned long remain = sc->nr_to_scan;
2148 if (!(sc->gfp_mask & __GFP_FS))
2149 /* We must not take the lu_sites_guard lock when
2150 * __GFP_FS is *not* set because of the deadlock
2151 * possibility detailed above. Additionally,
2152 * since we cannot determine the number of
2153 * objects in the cache without taking this
2154 * lock, we're in a particularly tough spot. As
2155 * a result, we'll just lie and say our cache is
2156 * empty. This _should_ be ok, as we can't
2157 * reclaim objects when __GFP_FS is *not* set
2162 down_write(&lu_sites_guard);
2163 list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
2164 remain = lu_site_purge(&lu_shrink_env, s, remain);
2166 * Move just shrunk site to the tail of site list to
2167 * assure shrinking fairness.
2169 list_move_tail(&s->ls_linkage, &splice);
2171 list_splice(&splice, lu_sites.prev);
2172 up_write(&lu_sites_guard);
2174 return sc->nr_to_scan - remain;
2177 #ifndef HAVE_SHRINKER_COUNT
2179 * There exists a potential lock inversion deadlock scenario when using
2180 * Lustre on top of ZFS. This occurs between one of ZFS's
2181 * buf_hash_table.ht_lock's, and Lustre's lu_sites_guard lock. Essentially,
2182 * thread A will take the lu_sites_guard lock and sleep on the ht_lock,
2183 * while thread B will take the ht_lock and sleep on the lu_sites_guard
2184 * lock. Obviously neither thread will wake and drop their respective hold
2187 * To prevent this from happening we must ensure the lu_sites_guard lock is
2188 * not taken while down this code path. ZFS reliably does not set the
2189 * __GFP_FS bit in its code paths, so this can be used to determine if it
2190 * is safe to take the lu_sites_guard lock.
2192 * Ideally we should accurately return the remaining number of cached
2193 * objects without taking the lu_sites_guard lock, but this is not
2194 * possible in the current implementation.
2196 static int lu_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
2199 struct shrink_control scv = {
2200 .nr_to_scan = shrink_param(sc, nr_to_scan),
2201 .gfp_mask = shrink_param(sc, gfp_mask)
2203 #if !defined(HAVE_SHRINKER_WANT_SHRINK_PTR) && !defined(HAVE_SHRINK_CONTROL)
2204 struct shrinker* shrinker = NULL;
2208 CDEBUG(D_INODE, "Shrink %lu objects\n", scv.nr_to_scan);
2210 if (scv.nr_to_scan != 0)
2211 lu_cache_shrink_scan(shrinker, &scv);
2213 cached = lu_cache_shrink_count(shrinker, &scv);
2217 #endif /* HAVE_SHRINKER_COUNT */
2225 * Environment to be used in debugger, contains all tags.
2227 static struct lu_env lu_debugging_env;
2230 * Debugging printer function using printk().
2232 int lu_printk_printer(const struct lu_env *env,
2233 void *unused, const char *format, ...)
2237 va_start(args, format);
2238 vprintk(format, args);
2243 int lu_debugging_setup(void)
2245 return lu_env_init(&lu_debugging_env, ~0);
2248 void lu_context_keys_dump(void)
2252 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
2253 struct lu_context_key *key;
2257 CERROR("[%d]: %p %x (%p,%p,%p) %d %d \"%s\"@%p\n",
2258 i, key, key->lct_tags,
2259 key->lct_init, key->lct_fini, key->lct_exit,
2260 key->lct_index, atomic_read(&key->lct_used),
2261 key->lct_owner ? key->lct_owner->name : "",
2263 lu_ref_print(&key->lct_reference);
2269 * Initialization of global lu_* data.
2271 int lu_global_init(void)
2274 DEF_SHRINKER_VAR(shvar, lu_cache_shrink,
2275 lu_cache_shrink_count, lu_cache_shrink_scan);
2277 CDEBUG(D_INFO, "Lustre LU module (%p).\n", &lu_keys);
2279 result = lu_ref_global_init();
2283 LU_CONTEXT_KEY_INIT(&lu_global_key);
2284 result = lu_context_key_register(&lu_global_key);
2289 * At this level, we don't know what tags are needed, so allocate them
2290 * conservatively. This should not be too bad, because this
2291 * environment is global.
2293 down_write(&lu_sites_guard);
2294 result = lu_env_init(&lu_shrink_env, LCT_SHRINKER);
2295 up_write(&lu_sites_guard);
2300 * seeks estimation: 3 seeks to read a record from oi, one to read
2301 * inode, one for ea. Unfortunately setting this high value results in
2302 * lu_object/inode cache consuming all the memory.
2304 lu_site_shrinker = set_shrinker(DEFAULT_SEEKS, &shvar);
2305 if (lu_site_shrinker == NULL)
2308 result = rhashtable_init(&lu_env_rhash, &lu_env_rhash_params);
2314 * Dual to lu_global_init().
2316 void lu_global_fini(void)
2318 if (lu_site_shrinker != NULL) {
2319 remove_shrinker(lu_site_shrinker);
2320 lu_site_shrinker = NULL;
2323 lu_context_key_degister(&lu_global_key);
2326 * Tear shrinker environment down _after_ de-registering
2327 * lu_global_key, because the latter has a value in the former.
2329 down_write(&lu_sites_guard);
2330 lu_env_fini(&lu_shrink_env);
2331 up_write(&lu_sites_guard);
2333 rhashtable_destroy(&lu_env_rhash);
2335 lu_ref_global_fini();
2338 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx)
2340 #ifdef CONFIG_PROC_FS
2341 struct lprocfs_counter ret;
2343 lprocfs_stats_collect(stats, idx, &ret);
2344 return (__u32)ret.lc_count;
2351 * Output site statistical counters into a buffer. Suitable for
2352 * lprocfs_rd_*()-style functions.
2354 int lu_site_stats_seq_print(const struct lu_site *s, struct seq_file *m)
2356 lu_site_stats_t stats;
2358 memset(&stats, 0, sizeof(stats));
2359 lu_site_stats_get(s, &stats, 1);
2361 seq_printf(m, "%d/%d %d/%d %d %d %d %d %d %d %d\n",
2364 stats.lss_populated,
2365 CFS_HASH_NHLIST(s->ls_obj_hash),
2366 stats.lss_max_search,
2367 ls_stats_read(s->ls_stats, LU_SS_CREATED),
2368 ls_stats_read(s->ls_stats, LU_SS_CACHE_HIT),
2369 ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
2370 ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
2371 ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
2372 ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED));
2375 EXPORT_SYMBOL(lu_site_stats_seq_print);
2378 * Helper function to initialize a number of kmem slab caches at once.
2380 int lu_kmem_init(struct lu_kmem_descr *caches)
2383 struct lu_kmem_descr *iter = caches;
2385 for (result = 0; iter->ckd_cache != NULL; ++iter) {
2386 *iter->ckd_cache = kmem_cache_create(iter->ckd_name,
2389 if (*iter->ckd_cache == NULL) {
2391 /* free all previously allocated caches */
2392 lu_kmem_fini(caches);
2398 EXPORT_SYMBOL(lu_kmem_init);
2401 * Helper function to finalize a number of kmem slab cached at once. Dual to
2404 void lu_kmem_fini(struct lu_kmem_descr *caches)
2406 for (; caches->ckd_cache != NULL; ++caches) {
2407 if (*caches->ckd_cache != NULL) {
2408 kmem_cache_destroy(*caches->ckd_cache);
2409 *caches->ckd_cache = NULL;
2413 EXPORT_SYMBOL(lu_kmem_fini);
2416 * Temporary solution to be able to assign fid in ->do_create()
2417 * till we have fully-functional OST fids
2419 void lu_object_assign_fid(const struct lu_env *env, struct lu_object *o,
2420 const struct lu_fid *fid)
2422 struct lu_site *s = o->lo_dev->ld_site;
2423 struct lu_fid *old = &o->lo_header->loh_fid;
2424 struct cfs_hash *hs;
2425 struct cfs_hash_bd bd;
2427 LASSERT(fid_is_zero(old));
2429 /* supposed to be unique */
2430 hs = s->ls_obj_hash;
2431 cfs_hash_bd_get_and_lock(hs, (void *)fid, &bd, 1);
2432 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
2435 struct lu_object *shadow;
2437 shadow = htable_lookup(s, &bd, fid, &version);
2438 /* supposed to be unique */
2439 LASSERT(IS_ERR(shadow) && PTR_ERR(shadow) == -ENOENT);
2443 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
2444 cfs_hash_bd_unlock(hs, &bd, 1);
2446 EXPORT_SYMBOL(lu_object_assign_fid);
2449 * allocates object with 0 (non-assiged) fid
2450 * XXX: temporary solution to be able to assign fid in ->do_create()
2451 * till we have fully-functional OST fids
2453 struct lu_object *lu_object_anon(const struct lu_env *env,
2454 struct lu_device *dev,
2455 const struct lu_object_conf *conf)
2458 struct lu_object *o;
2462 o = lu_object_alloc(env, dev, &fid);
2464 rc = lu_object_start(env, dev, o, conf);
2466 lu_object_free(env, o);
2473 EXPORT_SYMBOL(lu_object_anon);
2475 struct lu_buf LU_BUF_NULL = {
2479 EXPORT_SYMBOL(LU_BUF_NULL);
2481 void lu_buf_free(struct lu_buf *buf)
2485 LASSERT(buf->lb_len > 0);
2486 OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);
2491 EXPORT_SYMBOL(lu_buf_free);
2493 void lu_buf_alloc(struct lu_buf *buf, size_t size)
2496 LASSERT(buf->lb_buf == NULL);
2497 LASSERT(buf->lb_len == 0);
2498 OBD_ALLOC_LARGE(buf->lb_buf, size);
2499 if (likely(buf->lb_buf))
2502 EXPORT_SYMBOL(lu_buf_alloc);
2504 void lu_buf_realloc(struct lu_buf *buf, size_t size)
2507 lu_buf_alloc(buf, size);
2509 EXPORT_SYMBOL(lu_buf_realloc);
2511 struct lu_buf *lu_buf_check_and_alloc(struct lu_buf *buf, size_t len)
2513 if (buf->lb_buf == NULL && buf->lb_len == 0)
2514 lu_buf_alloc(buf, len);
2516 if ((len > buf->lb_len) && (buf->lb_buf != NULL))
2517 lu_buf_realloc(buf, len);
2521 EXPORT_SYMBOL(lu_buf_check_and_alloc);
2524 * Increase the size of the \a buf.
2525 * preserves old data in buffer
2526 * old buffer remains unchanged on error
2527 * \retval 0 or -ENOMEM
2529 int lu_buf_check_and_grow(struct lu_buf *buf, size_t len)
2533 if (len <= buf->lb_len)
2536 OBD_ALLOC_LARGE(ptr, len);
2540 /* Free the old buf */
2541 if (buf->lb_buf != NULL) {
2542 memcpy(ptr, buf->lb_buf, buf->lb_len);
2543 OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);
2550 EXPORT_SYMBOL(lu_buf_check_and_grow);