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;
737 * This uses standard index maintenance protocol:
739 * - search index under lock, and return object if found;
740 * - otherwise, unlock index, allocate new object;
741 * - lock index and search again;
742 * - if nothing is found (usual case), insert newly created
744 * - otherwise (race: other thread inserted object), free
745 * object just allocated.
749 * For "LOC_F_NEW" case, we are sure the object is new established.
750 * It is unnecessary to perform lookup-alloc-lookup-insert, instead,
751 * just alloc and insert directly.
757 if (unlikely(OBD_FAIL_PRECHECK(OBD_FAIL_OBD_ZERO_NLINK_RACE)))
758 lu_site_purge(env, s, -1);
760 cfs_hash_bd_get(hs, f, &bd);
761 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
762 if (!(conf && conf->loc_flags & LOC_F_NEW)) {
763 cfs_hash_bd_lock(hs, &bd, 1);
764 o = htable_lookup(s, &bd, f, &version);
765 cfs_hash_bd_unlock(hs, &bd, 1);
768 if (likely(lu_object_is_inited(o->lo_header)))
771 wait_event_idle(bkt->lsb_waitq,
772 lu_object_is_inited(o->lo_header) ||
773 lu_object_is_dying(o->lo_header));
775 if (lu_object_is_dying(o->lo_header)) {
776 lu_object_put(env, o);
778 RETURN(ERR_PTR(-ENOENT));
784 if (PTR_ERR(o) != -ENOENT)
789 * Allocate new object, NB, object is unitialized in case object
790 * is changed between allocation and hash insertion, thus the object
791 * with stale attributes is returned.
793 o = lu_object_alloc(env, dev, f);
797 LASSERT(lu_fid_eq(lu_object_fid(o), f));
799 CFS_RACE_WAIT(OBD_FAIL_OBD_ZERO_NLINK_RACE);
801 cfs_hash_bd_lock(hs, &bd, 1);
803 if (conf && conf->loc_flags & LOC_F_NEW)
804 shadow = ERR_PTR(-ENOENT);
806 shadow = htable_lookup(s, &bd, f, &version);
807 if (likely(PTR_ERR(shadow) == -ENOENT)) {
808 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
809 cfs_hash_bd_unlock(hs, &bd, 1);
812 * This may result in rather complicated operations, including
813 * fld queries, inode loading, etc.
815 rc = lu_object_start(env, dev, o, conf);
817 lu_object_put_nocache(env, o);
821 wake_up_all(&bkt->lsb_waitq);
823 lu_object_limit(env, dev);
828 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_RACE);
829 cfs_hash_bd_unlock(hs, &bd, 1);
830 lu_object_free(env, o);
832 if (!(conf && conf->loc_flags & LOC_F_NEW) &&
833 !lu_object_is_inited(shadow->lo_header)) {
834 wait_event_idle(bkt->lsb_waitq,
835 lu_object_is_inited(shadow->lo_header) ||
836 lu_object_is_dying(shadow->lo_header));
838 if (lu_object_is_dying(shadow->lo_header)) {
839 lu_object_put(env, shadow);
841 RETURN(ERR_PTR(-ENOENT));
847 EXPORT_SYMBOL(lu_object_find_at);
850 * Find object with given fid, and return its slice belonging to given device.
852 struct lu_object *lu_object_find_slice(const struct lu_env *env,
853 struct lu_device *dev,
854 const struct lu_fid *f,
855 const struct lu_object_conf *conf)
857 struct lu_object *top;
858 struct lu_object *obj;
860 top = lu_object_find(env, dev, f, conf);
864 obj = lu_object_locate(top->lo_header, dev->ld_type);
865 if (unlikely(obj == NULL)) {
866 lu_object_put(env, top);
867 obj = ERR_PTR(-ENOENT);
872 EXPORT_SYMBOL(lu_object_find_slice);
874 int lu_device_type_init(struct lu_device_type *ldt)
878 atomic_set(&ldt->ldt_device_nr, 0);
879 if (ldt->ldt_ops->ldto_init)
880 result = ldt->ldt_ops->ldto_init(ldt);
884 EXPORT_SYMBOL(lu_device_type_init);
886 void lu_device_type_fini(struct lu_device_type *ldt)
888 if (ldt->ldt_ops->ldto_fini)
889 ldt->ldt_ops->ldto_fini(ldt);
891 EXPORT_SYMBOL(lu_device_type_fini);
894 * Global list of all sites on this node
896 static LIST_HEAD(lu_sites);
897 static DECLARE_RWSEM(lu_sites_guard);
900 * Global environment used by site shrinker.
902 static struct lu_env lu_shrink_env;
904 struct lu_site_print_arg {
905 struct lu_env *lsp_env;
907 lu_printer_t lsp_printer;
911 lu_site_obj_print(struct cfs_hash *hs, struct cfs_hash_bd *bd,
912 struct hlist_node *hnode, void *data)
914 struct lu_site_print_arg *arg = (struct lu_site_print_arg *)data;
915 struct lu_object_header *h;
917 h = hlist_entry(hnode, struct lu_object_header, loh_hash);
918 if (!list_empty(&h->loh_layers)) {
919 const struct lu_object *o;
921 o = lu_object_top(h);
922 lu_object_print(arg->lsp_env, arg->lsp_cookie,
923 arg->lsp_printer, o);
925 lu_object_header_print(arg->lsp_env, arg->lsp_cookie,
926 arg->lsp_printer, h);
932 * Print all objects in \a s.
934 void lu_site_print(const struct lu_env *env, struct lu_site *s, void *cookie,
935 lu_printer_t printer)
937 struct lu_site_print_arg arg = {
938 .lsp_env = (struct lu_env *)env,
939 .lsp_cookie = cookie,
940 .lsp_printer = printer,
943 cfs_hash_for_each(s->ls_obj_hash, lu_site_obj_print, &arg);
945 EXPORT_SYMBOL(lu_site_print);
948 * Return desired hash table order.
950 static unsigned long lu_htable_order(struct lu_device *top)
952 unsigned long cache_size;
954 unsigned long bits_max = LU_SITE_BITS_MAX;
957 * For ZFS based OSDs the cache should be disabled by default. This
958 * allows the ZFS ARC maximum flexibility in determining what buffers
959 * to cache. If Lustre has objects or buffer which it wants to ensure
960 * always stay cached it must maintain a hold on them.
962 if (strcmp(top->ld_type->ldt_name, LUSTRE_OSD_ZFS_NAME) == 0) {
963 lu_cache_percent = 1;
964 lu_cache_nr = LU_CACHE_NR_ZFS_LIMIT;
965 return LU_SITE_BITS_MIN;
968 if (strcmp(top->ld_type->ldt_name, LUSTRE_VVP_NAME) == 0)
969 bits_max = LU_SITE_BITS_MAX_CL;
972 * Calculate hash table size, assuming that we want reasonable
973 * performance when 20% of total memory is occupied by cache of
976 * Size of lu_object is (arbitrary) taken as 1K (together with inode).
978 cache_size = cfs_totalram_pages();
980 #if BITS_PER_LONG == 32
981 /* limit hashtable size for lowmem systems to low RAM */
982 if (cache_size > 1 << (30 - PAGE_SHIFT))
983 cache_size = 1 << (30 - PAGE_SHIFT) * 3 / 4;
986 /* clear off unreasonable cache setting. */
987 if (lu_cache_percent == 0 || lu_cache_percent > LU_CACHE_PERCENT_MAX) {
988 CWARN("obdclass: invalid lu_cache_percent: %u, it must be in"
989 " the range of (0, %u]. Will use default value: %u.\n",
990 lu_cache_percent, LU_CACHE_PERCENT_MAX,
991 LU_CACHE_PERCENT_DEFAULT);
993 lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
995 cache_size = cache_size / 100 * lu_cache_percent *
998 for (bits = 1; (1 << bits) < cache_size; ++bits) {
1002 return clamp_t(typeof(bits), bits, LU_SITE_BITS_MIN, bits_max);
1005 static unsigned lu_obj_hop_hash(struct cfs_hash *hs,
1006 const void *key, unsigned mask)
1008 struct lu_fid *fid = (struct lu_fid *)key;
1011 hash = fid_flatten32(fid);
1012 hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
1013 hash = hash_long(hash, hs->hs_bkt_bits);
1015 /* give me another random factor */
1016 hash -= hash_long((unsigned long)hs, fid_oid(fid) % 11 + 3);
1018 hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
1019 hash |= (fid_seq(fid) + fid_oid(fid)) & (CFS_HASH_NBKT(hs) - 1);
1024 static void *lu_obj_hop_object(struct hlist_node *hnode)
1026 return hlist_entry(hnode, struct lu_object_header, loh_hash);
1029 static void *lu_obj_hop_key(struct hlist_node *hnode)
1031 struct lu_object_header *h;
1033 h = hlist_entry(hnode, struct lu_object_header, loh_hash);
1037 static int lu_obj_hop_keycmp(const void *key, struct hlist_node *hnode)
1039 struct lu_object_header *h;
1041 h = hlist_entry(hnode, struct lu_object_header, loh_hash);
1042 return lu_fid_eq(&h->loh_fid, (struct lu_fid *)key);
1045 static void lu_obj_hop_get(struct cfs_hash *hs, struct hlist_node *hnode)
1047 struct lu_object_header *h;
1049 h = hlist_entry(hnode, struct lu_object_header, loh_hash);
1050 atomic_inc(&h->loh_ref);
1053 static void lu_obj_hop_put_locked(struct cfs_hash *hs, struct hlist_node *hnode)
1055 LBUG(); /* we should never called it */
1058 static struct cfs_hash_ops lu_site_hash_ops = {
1059 .hs_hash = lu_obj_hop_hash,
1060 .hs_key = lu_obj_hop_key,
1061 .hs_keycmp = lu_obj_hop_keycmp,
1062 .hs_object = lu_obj_hop_object,
1063 .hs_get = lu_obj_hop_get,
1064 .hs_put_locked = lu_obj_hop_put_locked,
1067 void lu_dev_add_linkage(struct lu_site *s, struct lu_device *d)
1069 spin_lock(&s->ls_ld_lock);
1070 if (list_empty(&d->ld_linkage))
1071 list_add(&d->ld_linkage, &s->ls_ld_linkage);
1072 spin_unlock(&s->ls_ld_lock);
1074 EXPORT_SYMBOL(lu_dev_add_linkage);
1076 void lu_dev_del_linkage(struct lu_site *s, struct lu_device *d)
1078 spin_lock(&s->ls_ld_lock);
1079 list_del_init(&d->ld_linkage);
1080 spin_unlock(&s->ls_ld_lock);
1082 EXPORT_SYMBOL(lu_dev_del_linkage);
1085 * Initialize site \a s, with \a d as the top level device.
1087 int lu_site_init(struct lu_site *s, struct lu_device *top)
1089 struct lu_site_bkt_data *bkt;
1090 struct cfs_hash_bd bd;
1097 memset(s, 0, sizeof *s);
1098 mutex_init(&s->ls_purge_mutex);
1100 #ifdef HAVE_PERCPU_COUNTER_INIT_GFP_FLAG
1101 rc = percpu_counter_init(&s->ls_lru_len_counter, 0, GFP_NOFS);
1103 rc = percpu_counter_init(&s->ls_lru_len_counter, 0);
1108 snprintf(name, sizeof(name), "lu_site_%s", top->ld_type->ldt_name);
1109 for (bits = lu_htable_order(top);
1110 bits >= LU_SITE_BITS_MIN; bits--) {
1111 s->ls_obj_hash = cfs_hash_create(name, bits, bits,
1112 bits - LU_SITE_BKT_BITS,
1115 CFS_HASH_SPIN_BKTLOCK |
1116 CFS_HASH_NO_ITEMREF |
1118 CFS_HASH_ASSERT_EMPTY |
1120 if (s->ls_obj_hash != NULL)
1124 if (s->ls_obj_hash == NULL) {
1125 CERROR("failed to create lu_site hash with bits: %lu\n", bits);
1129 cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
1130 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
1131 INIT_LIST_HEAD(&bkt->lsb_lru);
1132 init_waitqueue_head(&bkt->lsb_waitq);
1135 s->ls_stats = lprocfs_alloc_stats(LU_SS_LAST_STAT, 0);
1136 if (s->ls_stats == NULL) {
1137 cfs_hash_putref(s->ls_obj_hash);
1138 s->ls_obj_hash = NULL;
1142 lprocfs_counter_init(s->ls_stats, LU_SS_CREATED,
1143 0, "created", "created");
1144 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_HIT,
1145 0, "cache_hit", "cache_hit");
1146 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_MISS,
1147 0, "cache_miss", "cache_miss");
1148 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_RACE,
1149 0, "cache_race", "cache_race");
1150 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_DEATH_RACE,
1151 0, "cache_death_race", "cache_death_race");
1152 lprocfs_counter_init(s->ls_stats, LU_SS_LRU_PURGED,
1153 0, "lru_purged", "lru_purged");
1155 INIT_LIST_HEAD(&s->ls_linkage);
1156 s->ls_top_dev = top;
1159 lu_ref_add(&top->ld_reference, "site-top", s);
1161 INIT_LIST_HEAD(&s->ls_ld_linkage);
1162 spin_lock_init(&s->ls_ld_lock);
1164 lu_dev_add_linkage(s, top);
1168 EXPORT_SYMBOL(lu_site_init);
1171 * Finalize \a s and release its resources.
1173 void lu_site_fini(struct lu_site *s)
1175 down_write(&lu_sites_guard);
1176 list_del_init(&s->ls_linkage);
1177 up_write(&lu_sites_guard);
1179 percpu_counter_destroy(&s->ls_lru_len_counter);
1181 if (s->ls_obj_hash != NULL) {
1182 cfs_hash_putref(s->ls_obj_hash);
1183 s->ls_obj_hash = NULL;
1186 if (s->ls_top_dev != NULL) {
1187 s->ls_top_dev->ld_site = NULL;
1188 lu_ref_del(&s->ls_top_dev->ld_reference, "site-top", s);
1189 lu_device_put(s->ls_top_dev);
1190 s->ls_top_dev = NULL;
1193 if (s->ls_stats != NULL)
1194 lprocfs_free_stats(&s->ls_stats);
1196 EXPORT_SYMBOL(lu_site_fini);
1199 * Called when initialization of stack for this site is completed.
1201 int lu_site_init_finish(struct lu_site *s)
1204 down_write(&lu_sites_guard);
1205 result = lu_context_refill(&lu_shrink_env.le_ctx);
1207 list_add(&s->ls_linkage, &lu_sites);
1208 up_write(&lu_sites_guard);
1211 EXPORT_SYMBOL(lu_site_init_finish);
1214 * Acquire additional reference on device \a d
1216 void lu_device_get(struct lu_device *d)
1218 atomic_inc(&d->ld_ref);
1220 EXPORT_SYMBOL(lu_device_get);
1223 * Release reference on device \a d.
1225 void lu_device_put(struct lu_device *d)
1227 LASSERT(atomic_read(&d->ld_ref) > 0);
1228 atomic_dec(&d->ld_ref);
1230 EXPORT_SYMBOL(lu_device_put);
1233 * Initialize device \a d of type \a t.
1235 int lu_device_init(struct lu_device *d, struct lu_device_type *t)
1237 if (atomic_inc_return(&t->ldt_device_nr) == 1 &&
1238 t->ldt_ops->ldto_start != NULL)
1239 t->ldt_ops->ldto_start(t);
1241 memset(d, 0, sizeof *d);
1243 lu_ref_init(&d->ld_reference);
1244 INIT_LIST_HEAD(&d->ld_linkage);
1248 EXPORT_SYMBOL(lu_device_init);
1251 * Finalize device \a d.
1253 void lu_device_fini(struct lu_device *d)
1255 struct lu_device_type *t = d->ld_type;
1257 if (d->ld_obd != NULL) {
1258 d->ld_obd->obd_lu_dev = NULL;
1262 lu_ref_fini(&d->ld_reference);
1263 LASSERTF(atomic_read(&d->ld_ref) == 0,
1264 "Refcount is %u\n", atomic_read(&d->ld_ref));
1265 LASSERT(atomic_read(&t->ldt_device_nr) > 0);
1267 if (atomic_dec_and_test(&t->ldt_device_nr) &&
1268 t->ldt_ops->ldto_stop != NULL)
1269 t->ldt_ops->ldto_stop(t);
1271 EXPORT_SYMBOL(lu_device_fini);
1274 * Initialize object \a o that is part of compound object \a h and was created
1277 int lu_object_init(struct lu_object *o, struct lu_object_header *h,
1278 struct lu_device *d)
1280 memset(o, 0, sizeof(*o));
1284 lu_ref_add_at(&d->ld_reference, &o->lo_dev_ref, "lu_object", o);
1285 INIT_LIST_HEAD(&o->lo_linkage);
1289 EXPORT_SYMBOL(lu_object_init);
1292 * Finalize object and release its resources.
1294 void lu_object_fini(struct lu_object *o)
1296 struct lu_device *dev = o->lo_dev;
1298 LASSERT(list_empty(&o->lo_linkage));
1301 lu_ref_del_at(&dev->ld_reference, &o->lo_dev_ref,
1307 EXPORT_SYMBOL(lu_object_fini);
1310 * Add object \a o as first layer of compound object \a h
1312 * This is typically called by the ->ldo_object_alloc() method of top-level
1315 void lu_object_add_top(struct lu_object_header *h, struct lu_object *o)
1317 list_move(&o->lo_linkage, &h->loh_layers);
1319 EXPORT_SYMBOL(lu_object_add_top);
1322 * Add object \a o as a layer of compound object, going after \a before.
1324 * This is typically called by the ->ldo_object_alloc() method of \a
1327 void lu_object_add(struct lu_object *before, struct lu_object *o)
1329 list_move(&o->lo_linkage, &before->lo_linkage);
1331 EXPORT_SYMBOL(lu_object_add);
1334 * Initialize compound object.
1336 int lu_object_header_init(struct lu_object_header *h)
1338 memset(h, 0, sizeof *h);
1339 atomic_set(&h->loh_ref, 1);
1340 INIT_HLIST_NODE(&h->loh_hash);
1341 INIT_LIST_HEAD(&h->loh_lru);
1342 INIT_LIST_HEAD(&h->loh_layers);
1343 lu_ref_init(&h->loh_reference);
1346 EXPORT_SYMBOL(lu_object_header_init);
1349 * Finalize compound object.
1351 void lu_object_header_fini(struct lu_object_header *h)
1353 LASSERT(list_empty(&h->loh_layers));
1354 LASSERT(list_empty(&h->loh_lru));
1355 LASSERT(hlist_unhashed(&h->loh_hash));
1356 lu_ref_fini(&h->loh_reference);
1358 EXPORT_SYMBOL(lu_object_header_fini);
1361 * Given a compound object, find its slice, corresponding to the device type
1364 struct lu_object *lu_object_locate(struct lu_object_header *h,
1365 const struct lu_device_type *dtype)
1367 struct lu_object *o;
1369 list_for_each_entry(o, &h->loh_layers, lo_linkage) {
1370 if (o->lo_dev->ld_type == dtype)
1375 EXPORT_SYMBOL(lu_object_locate);
1378 * Finalize and free devices in the device stack.
1380 * Finalize device stack by purging object cache, and calling
1381 * lu_device_type_operations::ldto_device_fini() and
1382 * lu_device_type_operations::ldto_device_free() on all devices in the stack.
1384 void lu_stack_fini(const struct lu_env *env, struct lu_device *top)
1386 struct lu_site *site = top->ld_site;
1387 struct lu_device *scan;
1388 struct lu_device *next;
1390 lu_site_purge(env, site, ~0);
1391 for (scan = top; scan != NULL; scan = next) {
1392 next = scan->ld_type->ldt_ops->ldto_device_fini(env, scan);
1393 lu_ref_del(&scan->ld_reference, "lu-stack", &lu_site_init);
1394 lu_device_put(scan);
1398 lu_site_purge(env, site, ~0);
1400 for (scan = top; scan != NULL; scan = next) {
1401 const struct lu_device_type *ldt = scan->ld_type;
1403 next = ldt->ldt_ops->ldto_device_free(env, scan);
1409 * Maximal number of tld slots.
1411 LU_CONTEXT_KEY_NR = 40
1414 static struct lu_context_key *lu_keys[LU_CONTEXT_KEY_NR] = { NULL, };
1416 static DECLARE_RWSEM(lu_key_initing);
1419 * Global counter incremented whenever key is registered, unregistered,
1420 * revived or quiesced. This is used to void unnecessary calls to
1421 * lu_context_refill(). No locking is provided, as initialization and shutdown
1422 * are supposed to be externally serialized.
1424 static atomic_t key_set_version = ATOMIC_INIT(0);
1429 int lu_context_key_register(struct lu_context_key *key)
1434 LASSERT(key->lct_init != NULL);
1435 LASSERT(key->lct_fini != NULL);
1436 LASSERT(key->lct_tags != 0);
1437 LASSERT(key->lct_owner != NULL);
1440 atomic_set(&key->lct_used, 1);
1441 lu_ref_init(&key->lct_reference);
1442 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1446 if (cmpxchg(&lu_keys[i], NULL, key) != NULL)
1450 atomic_inc(&key_set_version);
1454 lu_ref_fini(&key->lct_reference);
1455 atomic_set(&key->lct_used, 0);
1459 EXPORT_SYMBOL(lu_context_key_register);
1461 static void key_fini(struct lu_context *ctx, int index)
1463 if (ctx->lc_value != NULL && ctx->lc_value[index] != NULL) {
1464 struct lu_context_key *key;
1466 key = lu_keys[index];
1467 LASSERT(key != NULL);
1468 LASSERT(key->lct_fini != NULL);
1469 LASSERT(atomic_read(&key->lct_used) > 0);
1471 key->lct_fini(ctx, key, ctx->lc_value[index]);
1472 lu_ref_del(&key->lct_reference, "ctx", ctx);
1473 if (atomic_dec_and_test(&key->lct_used))
1474 wake_up_var(&key->lct_used);
1476 LASSERT(key->lct_owner != NULL);
1477 if ((ctx->lc_tags & LCT_NOREF) == 0) {
1478 LINVRNT(module_refcount(key->lct_owner) > 0);
1479 module_put(key->lct_owner);
1481 ctx->lc_value[index] = NULL;
1488 void lu_context_key_degister(struct lu_context_key *key)
1490 LASSERT(atomic_read(&key->lct_used) >= 1);
1491 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1493 lu_context_key_quiesce(key);
1495 key_fini(&lu_shrink_env.le_ctx, key->lct_index);
1498 * Wait until all transient contexts referencing this key have
1499 * run lu_context_key::lct_fini() method.
1501 atomic_dec(&key->lct_used);
1502 wait_var_event(&key->lct_used, atomic_read(&key->lct_used) == 0);
1504 if (!WARN_ON(lu_keys[key->lct_index] == NULL))
1505 lu_ref_fini(&key->lct_reference);
1507 smp_store_release(&lu_keys[key->lct_index], NULL);
1509 EXPORT_SYMBOL(lu_context_key_degister);
1512 * Register a number of keys. This has to be called after all keys have been
1513 * initialized by a call to LU_CONTEXT_KEY_INIT().
1515 int lu_context_key_register_many(struct lu_context_key *k, ...)
1517 struct lu_context_key *key = k;
1523 result = lu_context_key_register(key);
1526 key = va_arg(args, struct lu_context_key *);
1527 } while (key != NULL);
1533 lu_context_key_degister(k);
1534 k = va_arg(args, struct lu_context_key *);
1541 EXPORT_SYMBOL(lu_context_key_register_many);
1544 * De-register a number of keys. This is a dual to
1545 * lu_context_key_register_many().
1547 void lu_context_key_degister_many(struct lu_context_key *k, ...)
1553 lu_context_key_degister(k);
1554 k = va_arg(args, struct lu_context_key*);
1555 } while (k != NULL);
1558 EXPORT_SYMBOL(lu_context_key_degister_many);
1561 * Revive a number of keys.
1563 void lu_context_key_revive_many(struct lu_context_key *k, ...)
1569 lu_context_key_revive(k);
1570 k = va_arg(args, struct lu_context_key*);
1571 } while (k != NULL);
1574 EXPORT_SYMBOL(lu_context_key_revive_many);
1577 * Quiescent a number of keys.
1579 void lu_context_key_quiesce_many(struct lu_context_key *k, ...)
1585 lu_context_key_quiesce(k);
1586 k = va_arg(args, struct lu_context_key*);
1587 } while (k != NULL);
1590 EXPORT_SYMBOL(lu_context_key_quiesce_many);
1593 * Return value associated with key \a key in context \a ctx.
1595 void *lu_context_key_get(const struct lu_context *ctx,
1596 const struct lu_context_key *key)
1598 LINVRNT(ctx->lc_state == LCS_ENTERED);
1599 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1600 LASSERT(lu_keys[key->lct_index] == key);
1601 return ctx->lc_value[key->lct_index];
1603 EXPORT_SYMBOL(lu_context_key_get);
1606 * List of remembered contexts. XXX document me.
1608 static LIST_HEAD(lu_context_remembered);
1609 static DEFINE_SPINLOCK(lu_context_remembered_guard);
1612 * Destroy \a key in all remembered contexts. This is used to destroy key
1613 * values in "shared" contexts (like service threads), when a module owning
1614 * the key is about to be unloaded.
1616 void lu_context_key_quiesce(struct lu_context_key *key)
1618 struct lu_context *ctx;
1620 if (!(key->lct_tags & LCT_QUIESCENT)) {
1622 * The write-lock on lu_key_initing will ensure that any
1623 * keys_fill() which didn't see LCT_QUIESCENT will have
1624 * finished before we call key_fini().
1626 down_write(&lu_key_initing);
1627 key->lct_tags |= LCT_QUIESCENT;
1628 up_write(&lu_key_initing);
1630 spin_lock(&lu_context_remembered_guard);
1631 list_for_each_entry(ctx, &lu_context_remembered, lc_remember) {
1632 spin_until_cond(READ_ONCE(ctx->lc_state) != LCS_LEAVING);
1633 key_fini(ctx, key->lct_index);
1636 spin_unlock(&lu_context_remembered_guard);
1640 void lu_context_key_revive(struct lu_context_key *key)
1642 key->lct_tags &= ~LCT_QUIESCENT;
1643 atomic_inc(&key_set_version);
1646 static void keys_fini(struct lu_context *ctx)
1650 if (ctx->lc_value == NULL)
1653 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i)
1656 OBD_FREE(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1657 ctx->lc_value = NULL;
1660 static int keys_fill(struct lu_context *ctx)
1666 * A serialisation with lu_context_key_quiesce() is needed, to
1667 * ensure we see LCT_QUIESCENT and don't allocate a new value
1668 * after it freed one. The rwsem provides this. As down_read()
1669 * does optimistic spinning while the writer is active, this is
1670 * unlikely to ever sleep.
1672 down_read(&lu_key_initing);
1673 ctx->lc_version = atomic_read(&key_set_version);
1675 LINVRNT(ctx->lc_value);
1676 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1677 struct lu_context_key *key;
1680 if (!ctx->lc_value[i] && key &&
1681 (key->lct_tags & ctx->lc_tags) &&
1683 * Don't create values for a LCT_QUIESCENT key, as this
1684 * will pin module owning a key.
1686 !(key->lct_tags & LCT_QUIESCENT)) {
1689 LINVRNT(key->lct_init != NULL);
1690 LINVRNT(key->lct_index == i);
1692 LASSERT(key->lct_owner != NULL);
1693 if (!(ctx->lc_tags & LCT_NOREF) &&
1694 try_module_get(key->lct_owner) == 0) {
1695 /* module is unloading, skip this key */
1699 value = key->lct_init(ctx, key);
1700 if (unlikely(IS_ERR(value))) {
1701 rc = PTR_ERR(value);
1705 lu_ref_add_atomic(&key->lct_reference, "ctx", ctx);
1706 atomic_inc(&key->lct_used);
1708 * This is the only place in the code, where an
1709 * element of ctx->lc_value[] array is set to non-NULL
1712 ctx->lc_value[i] = value;
1713 if (key->lct_exit != NULL)
1714 ctx->lc_tags |= LCT_HAS_EXIT;
1718 up_read(&lu_key_initing);
1722 static int keys_init(struct lu_context *ctx)
1724 OBD_ALLOC(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1725 if (likely(ctx->lc_value != NULL))
1726 return keys_fill(ctx);
1732 * Initialize context data-structure. Create values for all keys.
1734 int lu_context_init(struct lu_context *ctx, __u32 tags)
1738 memset(ctx, 0, sizeof *ctx);
1739 ctx->lc_state = LCS_INITIALIZED;
1740 ctx->lc_tags = tags;
1741 if (tags & LCT_REMEMBER) {
1742 spin_lock(&lu_context_remembered_guard);
1743 list_add(&ctx->lc_remember, &lu_context_remembered);
1744 spin_unlock(&lu_context_remembered_guard);
1746 INIT_LIST_HEAD(&ctx->lc_remember);
1749 rc = keys_init(ctx);
1751 lu_context_fini(ctx);
1755 EXPORT_SYMBOL(lu_context_init);
1758 * Finalize context data-structure. Destroy key values.
1760 void lu_context_fini(struct lu_context *ctx)
1762 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1763 ctx->lc_state = LCS_FINALIZED;
1765 if ((ctx->lc_tags & LCT_REMEMBER) == 0) {
1766 LASSERT(list_empty(&ctx->lc_remember));
1768 /* could race with key degister */
1769 spin_lock(&lu_context_remembered_guard);
1770 list_del_init(&ctx->lc_remember);
1771 spin_unlock(&lu_context_remembered_guard);
1775 EXPORT_SYMBOL(lu_context_fini);
1778 * Called before entering context.
1780 void lu_context_enter(struct lu_context *ctx)
1782 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1783 ctx->lc_state = LCS_ENTERED;
1785 EXPORT_SYMBOL(lu_context_enter);
1788 * Called after exiting from \a ctx
1790 void lu_context_exit(struct lu_context *ctx)
1794 LINVRNT(ctx->lc_state == LCS_ENTERED);
1796 * Disable preempt to ensure we get a warning if
1797 * any lct_exit ever tries to sleep. That would hurt
1798 * lu_context_key_quiesce() which spins waiting for us.
1799 * This also ensure we aren't preempted while the state
1800 * is LCS_LEAVING, as that too would cause problems for
1801 * lu_context_key_quiesce().
1805 * Ensure lu_context_key_quiesce() sees LCS_LEAVING
1806 * or we see LCT_QUIESCENT
1808 smp_store_mb(ctx->lc_state, LCS_LEAVING);
1809 if (ctx->lc_tags & LCT_HAS_EXIT && ctx->lc_value) {
1810 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1811 struct lu_context_key *key;
1814 if (ctx->lc_value[i] &&
1815 !(key->lct_tags & LCT_QUIESCENT) &&
1817 key->lct_exit(ctx, key, ctx->lc_value[i]);
1821 smp_store_release(&ctx->lc_state, LCS_LEFT);
1824 EXPORT_SYMBOL(lu_context_exit);
1827 * Allocate for context all missing keys that were registered after context
1828 * creation. key_set_version is only changed in rare cases when modules
1829 * are loaded and removed.
1831 int lu_context_refill(struct lu_context *ctx)
1833 if (likely(ctx->lc_version == atomic_read(&key_set_version)))
1836 return keys_fill(ctx);
1840 * lu_ctx_tags/lu_ses_tags will be updated if there are new types of
1841 * obd being added. Currently, this is only used on client side, specifically
1842 * for echo device client, for other stack (like ptlrpc threads), context are
1843 * predefined when the lu_device type are registered, during the module probe
1846 u32 lu_context_tags_default = LCT_CL_THREAD;
1847 u32 lu_session_tags_default = LCT_SESSION;
1849 void lu_context_tags_update(__u32 tags)
1851 spin_lock(&lu_context_remembered_guard);
1852 lu_context_tags_default |= tags;
1853 atomic_inc(&key_set_version);
1854 spin_unlock(&lu_context_remembered_guard);
1856 EXPORT_SYMBOL(lu_context_tags_update);
1858 void lu_context_tags_clear(__u32 tags)
1860 spin_lock(&lu_context_remembered_guard);
1861 lu_context_tags_default &= ~tags;
1862 atomic_inc(&key_set_version);
1863 spin_unlock(&lu_context_remembered_guard);
1865 EXPORT_SYMBOL(lu_context_tags_clear);
1867 void lu_session_tags_update(__u32 tags)
1869 spin_lock(&lu_context_remembered_guard);
1870 lu_session_tags_default |= tags;
1871 atomic_inc(&key_set_version);
1872 spin_unlock(&lu_context_remembered_guard);
1874 EXPORT_SYMBOL(lu_session_tags_update);
1876 void lu_session_tags_clear(__u32 tags)
1878 spin_lock(&lu_context_remembered_guard);
1879 lu_session_tags_default &= ~tags;
1880 atomic_inc(&key_set_version);
1881 spin_unlock(&lu_context_remembered_guard);
1883 EXPORT_SYMBOL(lu_session_tags_clear);
1885 int lu_env_init(struct lu_env *env, __u32 tags)
1890 result = lu_context_init(&env->le_ctx, tags);
1891 if (likely(result == 0))
1892 lu_context_enter(&env->le_ctx);
1895 EXPORT_SYMBOL(lu_env_init);
1897 void lu_env_fini(struct lu_env *env)
1899 lu_context_exit(&env->le_ctx);
1900 lu_context_fini(&env->le_ctx);
1903 EXPORT_SYMBOL(lu_env_fini);
1905 int lu_env_refill(struct lu_env *env)
1909 result = lu_context_refill(&env->le_ctx);
1910 if (result == 0 && env->le_ses != NULL)
1911 result = lu_context_refill(env->le_ses);
1914 EXPORT_SYMBOL(lu_env_refill);
1917 * Currently, this API will only be used by echo client.
1918 * Because echo client and normal lustre client will share
1919 * same cl_env cache. So echo client needs to refresh
1920 * the env context after it get one from the cache, especially
1921 * when normal client and echo client co-exist in the same client.
1923 int lu_env_refill_by_tags(struct lu_env *env, __u32 ctags,
1928 if ((env->le_ctx.lc_tags & ctags) != ctags) {
1929 env->le_ctx.lc_version = 0;
1930 env->le_ctx.lc_tags |= ctags;
1933 if (env->le_ses && (env->le_ses->lc_tags & stags) != stags) {
1934 env->le_ses->lc_version = 0;
1935 env->le_ses->lc_tags |= stags;
1938 result = lu_env_refill(env);
1942 EXPORT_SYMBOL(lu_env_refill_by_tags);
1945 struct lu_env_item {
1946 struct task_struct *lei_task; /* rhashtable key */
1947 struct rhash_head lei_linkage;
1948 struct lu_env *lei_env;
1949 struct rcu_head lei_rcu_head;
1952 static const struct rhashtable_params lu_env_rhash_params = {
1953 .key_len = sizeof(struct task_struct *),
1954 .key_offset = offsetof(struct lu_env_item, lei_task),
1955 .head_offset = offsetof(struct lu_env_item, lei_linkage),
1958 struct rhashtable lu_env_rhash;
1960 struct lu_env_percpu {
1961 struct task_struct *lep_task;
1962 struct lu_env *lep_env ____cacheline_aligned_in_smp;
1965 static struct lu_env_percpu lu_env_percpu[NR_CPUS];
1967 int lu_env_add(struct lu_env *env)
1969 struct lu_env_item *lei, *old;
1977 lei->lei_task = current;
1980 old = rhashtable_lookup_get_insert_fast(&lu_env_rhash,
1982 lu_env_rhash_params);
1987 EXPORT_SYMBOL(lu_env_add);
1989 static void lu_env_item_free(struct rcu_head *head)
1991 struct lu_env_item *lei;
1993 lei = container_of(head, struct lu_env_item, lei_rcu_head);
1997 void lu_env_remove(struct lu_env *env)
1999 struct lu_env_item *lei;
2000 const void *task = current;
2003 for_each_possible_cpu(i) {
2004 if (lu_env_percpu[i].lep_env == env) {
2005 LASSERT(lu_env_percpu[i].lep_task == task);
2006 lu_env_percpu[i].lep_task = NULL;
2007 lu_env_percpu[i].lep_env = NULL;
2011 /* The rcu_lock is not taking in this case since the key
2012 * used is the actual task_struct. This implies that each
2013 * object is only removed by the owning thread, so there
2014 * can never be a race on a particular object.
2016 lei = rhashtable_lookup_fast(&lu_env_rhash, &task,
2017 lu_env_rhash_params);
2018 if (lei && rhashtable_remove_fast(&lu_env_rhash, &lei->lei_linkage,
2019 lu_env_rhash_params) == 0)
2020 call_rcu(&lei->lei_rcu_head, lu_env_item_free);
2022 EXPORT_SYMBOL(lu_env_remove);
2024 struct lu_env *lu_env_find(void)
2026 struct lu_env *env = NULL;
2027 struct lu_env_item *lei;
2028 const void *task = current;
2031 if (lu_env_percpu[i].lep_task == current) {
2032 env = lu_env_percpu[i].lep_env;
2038 lei = rhashtable_lookup_fast(&lu_env_rhash, &task,
2039 lu_env_rhash_params);
2042 lu_env_percpu[i].lep_task = current;
2043 lu_env_percpu[i].lep_env = env;
2049 EXPORT_SYMBOL(lu_env_find);
2051 static struct shrinker *lu_site_shrinker;
2053 typedef struct lu_site_stats{
2054 unsigned lss_populated;
2055 unsigned lss_max_search;
2060 static void lu_site_stats_get(const struct lu_site *s,
2061 lu_site_stats_t *stats, int populated)
2063 struct cfs_hash *hs = s->ls_obj_hash;
2064 struct cfs_hash_bd bd;
2067 * percpu_counter_sum_positive() won't accept a const pointer
2068 * as it does modify the struct by taking a spinlock
2070 struct lu_site *s2 = (struct lu_site *)s;
2072 stats->lss_busy += cfs_hash_size_get(hs) -
2073 percpu_counter_sum_positive(&s2->ls_lru_len_counter);
2074 cfs_hash_for_each_bucket(hs, &bd, i) {
2075 struct hlist_head *hhead;
2077 cfs_hash_bd_lock(hs, &bd, 1);
2078 stats->lss_total += cfs_hash_bd_count_get(&bd);
2079 stats->lss_max_search = max((int)stats->lss_max_search,
2080 cfs_hash_bd_depmax_get(&bd));
2082 cfs_hash_bd_unlock(hs, &bd, 1);
2086 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
2087 if (!hlist_empty(hhead))
2088 stats->lss_populated++;
2090 cfs_hash_bd_unlock(hs, &bd, 1);
2096 * lu_cache_shrink_count() returns an approximate number of cached objects
2097 * that can be freed by shrink_slab(). A counter, which tracks the
2098 * number of items in the site's lru, is maintained in a percpu_counter
2099 * for each site. The percpu values are incremented and decremented as
2100 * objects are added or removed from the lru. The percpu values are summed
2101 * and saved whenever a percpu value exceeds a threshold. Thus the saved,
2102 * summed value at any given time may not accurately reflect the current
2103 * lru length. But this value is sufficiently accurate for the needs of
2106 * Using a per cpu counter is a compromise solution to concurrent access:
2107 * lu_object_put() can update the counter without locking the site and
2108 * lu_cache_shrink_count can sum the counters without locking each
2109 * ls_obj_hash bucket.
2111 static unsigned long lu_cache_shrink_count(struct shrinker *sk,
2112 struct shrink_control *sc)
2115 struct lu_site *tmp;
2116 unsigned long cached = 0;
2118 if (!(sc->gfp_mask & __GFP_FS))
2121 down_read(&lu_sites_guard);
2122 list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage)
2123 cached += percpu_counter_read_positive(&s->ls_lru_len_counter);
2124 up_read(&lu_sites_guard);
2126 cached = (cached / 100) * sysctl_vfs_cache_pressure;
2127 CDEBUG(D_INODE, "%ld objects cached, cache pressure %d\n",
2128 cached, sysctl_vfs_cache_pressure);
2133 static unsigned long lu_cache_shrink_scan(struct shrinker *sk,
2134 struct shrink_control *sc)
2137 struct lu_site *tmp;
2138 unsigned long remain = sc->nr_to_scan;
2141 if (!(sc->gfp_mask & __GFP_FS))
2142 /* We must not take the lu_sites_guard lock when
2143 * __GFP_FS is *not* set because of the deadlock
2144 * possibility detailed above. Additionally,
2145 * since we cannot determine the number of
2146 * objects in the cache without taking this
2147 * lock, we're in a particularly tough spot. As
2148 * a result, we'll just lie and say our cache is
2149 * empty. This _should_ be ok, as we can't
2150 * reclaim objects when __GFP_FS is *not* set
2155 down_write(&lu_sites_guard);
2156 list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
2157 remain = lu_site_purge(&lu_shrink_env, s, remain);
2159 * Move just shrunk site to the tail of site list to
2160 * assure shrinking fairness.
2162 list_move_tail(&s->ls_linkage, &splice);
2164 list_splice(&splice, lu_sites.prev);
2165 up_write(&lu_sites_guard);
2167 return sc->nr_to_scan - remain;
2170 #ifndef HAVE_SHRINKER_COUNT
2172 * There exists a potential lock inversion deadlock scenario when using
2173 * Lustre on top of ZFS. This occurs between one of ZFS's
2174 * buf_hash_table.ht_lock's, and Lustre's lu_sites_guard lock. Essentially,
2175 * thread A will take the lu_sites_guard lock and sleep on the ht_lock,
2176 * while thread B will take the ht_lock and sleep on the lu_sites_guard
2177 * lock. Obviously neither thread will wake and drop their respective hold
2180 * To prevent this from happening we must ensure the lu_sites_guard lock is
2181 * not taken while down this code path. ZFS reliably does not set the
2182 * __GFP_FS bit in its code paths, so this can be used to determine if it
2183 * is safe to take the lu_sites_guard lock.
2185 * Ideally we should accurately return the remaining number of cached
2186 * objects without taking the lu_sites_guard lock, but this is not
2187 * possible in the current implementation.
2189 static int lu_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
2192 struct shrink_control scv = {
2193 .nr_to_scan = shrink_param(sc, nr_to_scan),
2194 .gfp_mask = shrink_param(sc, gfp_mask)
2196 #if !defined(HAVE_SHRINKER_WANT_SHRINK_PTR) && !defined(HAVE_SHRINK_CONTROL)
2197 struct shrinker* shrinker = NULL;
2201 CDEBUG(D_INODE, "Shrink %lu objects\n", scv.nr_to_scan);
2203 if (scv.nr_to_scan != 0)
2204 lu_cache_shrink_scan(shrinker, &scv);
2206 cached = lu_cache_shrink_count(shrinker, &scv);
2210 #endif /* HAVE_SHRINKER_COUNT */
2218 * Environment to be used in debugger, contains all tags.
2220 static struct lu_env lu_debugging_env;
2223 * Debugging printer function using printk().
2225 int lu_printk_printer(const struct lu_env *env,
2226 void *unused, const char *format, ...)
2230 va_start(args, format);
2231 vprintk(format, args);
2236 int lu_debugging_setup(void)
2238 return lu_env_init(&lu_debugging_env, ~0);
2241 void lu_context_keys_dump(void)
2245 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
2246 struct lu_context_key *key;
2250 CERROR("[%d]: %p %x (%p,%p,%p) %d %d \"%s\"@%p\n",
2251 i, key, key->lct_tags,
2252 key->lct_init, key->lct_fini, key->lct_exit,
2253 key->lct_index, atomic_read(&key->lct_used),
2254 key->lct_owner ? key->lct_owner->name : "",
2256 lu_ref_print(&key->lct_reference);
2262 * Initialization of global lu_* data.
2264 int lu_global_init(void)
2267 DEF_SHRINKER_VAR(shvar, lu_cache_shrink,
2268 lu_cache_shrink_count, lu_cache_shrink_scan);
2270 CDEBUG(D_INFO, "Lustre LU module (%p).\n", &lu_keys);
2272 result = lu_ref_global_init();
2276 LU_CONTEXT_KEY_INIT(&lu_global_key);
2277 result = lu_context_key_register(&lu_global_key);
2282 * At this level, we don't know what tags are needed, so allocate them
2283 * conservatively. This should not be too bad, because this
2284 * environment is global.
2286 down_write(&lu_sites_guard);
2287 result = lu_env_init(&lu_shrink_env, LCT_SHRINKER);
2288 up_write(&lu_sites_guard);
2293 * seeks estimation: 3 seeks to read a record from oi, one to read
2294 * inode, one for ea. Unfortunately setting this high value results in
2295 * lu_object/inode cache consuming all the memory.
2297 lu_site_shrinker = set_shrinker(DEFAULT_SEEKS, &shvar);
2298 if (lu_site_shrinker == NULL)
2301 result = rhashtable_init(&lu_env_rhash, &lu_env_rhash_params);
2307 * Dual to lu_global_init().
2309 void lu_global_fini(void)
2311 if (lu_site_shrinker != NULL) {
2312 remove_shrinker(lu_site_shrinker);
2313 lu_site_shrinker = NULL;
2316 lu_context_key_degister(&lu_global_key);
2319 * Tear shrinker environment down _after_ de-registering
2320 * lu_global_key, because the latter has a value in the former.
2322 down_write(&lu_sites_guard);
2323 lu_env_fini(&lu_shrink_env);
2324 up_write(&lu_sites_guard);
2326 rhashtable_destroy(&lu_env_rhash);
2328 lu_ref_global_fini();
2331 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx)
2333 #ifdef CONFIG_PROC_FS
2334 struct lprocfs_counter ret;
2336 lprocfs_stats_collect(stats, idx, &ret);
2337 return (__u32)ret.lc_count;
2344 * Output site statistical counters into a buffer. Suitable for
2345 * lprocfs_rd_*()-style functions.
2347 int lu_site_stats_seq_print(const struct lu_site *s, struct seq_file *m)
2349 lu_site_stats_t stats;
2351 memset(&stats, 0, sizeof(stats));
2352 lu_site_stats_get(s, &stats, 1);
2354 seq_printf(m, "%d/%d %d/%d %d %d %d %d %d %d %d\n",
2357 stats.lss_populated,
2358 CFS_HASH_NHLIST(s->ls_obj_hash),
2359 stats.lss_max_search,
2360 ls_stats_read(s->ls_stats, LU_SS_CREATED),
2361 ls_stats_read(s->ls_stats, LU_SS_CACHE_HIT),
2362 ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
2363 ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
2364 ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
2365 ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED));
2368 EXPORT_SYMBOL(lu_site_stats_seq_print);
2371 * Helper function to initialize a number of kmem slab caches at once.
2373 int lu_kmem_init(struct lu_kmem_descr *caches)
2376 struct lu_kmem_descr *iter = caches;
2378 for (result = 0; iter->ckd_cache != NULL; ++iter) {
2379 *iter->ckd_cache = kmem_cache_create(iter->ckd_name,
2382 if (*iter->ckd_cache == NULL) {
2384 /* free all previously allocated caches */
2385 lu_kmem_fini(caches);
2391 EXPORT_SYMBOL(lu_kmem_init);
2394 * Helper function to finalize a number of kmem slab cached at once. Dual to
2397 void lu_kmem_fini(struct lu_kmem_descr *caches)
2399 for (; caches->ckd_cache != NULL; ++caches) {
2400 if (*caches->ckd_cache != NULL) {
2401 kmem_cache_destroy(*caches->ckd_cache);
2402 *caches->ckd_cache = NULL;
2406 EXPORT_SYMBOL(lu_kmem_fini);
2409 * Temporary solution to be able to assign fid in ->do_create()
2410 * till we have fully-functional OST fids
2412 void lu_object_assign_fid(const struct lu_env *env, struct lu_object *o,
2413 const struct lu_fid *fid)
2415 struct lu_site *s = o->lo_dev->ld_site;
2416 struct lu_fid *old = &o->lo_header->loh_fid;
2417 struct cfs_hash *hs;
2418 struct cfs_hash_bd bd;
2420 LASSERT(fid_is_zero(old));
2422 /* supposed to be unique */
2423 hs = s->ls_obj_hash;
2424 cfs_hash_bd_get_and_lock(hs, (void *)fid, &bd, 1);
2425 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
2428 struct lu_object *shadow;
2430 shadow = htable_lookup(s, &bd, fid, &version);
2431 /* supposed to be unique */
2432 LASSERT(IS_ERR(shadow) && PTR_ERR(shadow) == -ENOENT);
2436 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
2437 cfs_hash_bd_unlock(hs, &bd, 1);
2439 EXPORT_SYMBOL(lu_object_assign_fid);
2442 * allocates object with 0 (non-assiged) fid
2443 * XXX: temporary solution to be able to assign fid in ->do_create()
2444 * till we have fully-functional OST fids
2446 struct lu_object *lu_object_anon(const struct lu_env *env,
2447 struct lu_device *dev,
2448 const struct lu_object_conf *conf)
2451 struct lu_object *o;
2455 o = lu_object_alloc(env, dev, &fid);
2457 rc = lu_object_start(env, dev, o, conf);
2459 lu_object_free(env, o);
2466 EXPORT_SYMBOL(lu_object_anon);
2468 struct lu_buf LU_BUF_NULL = {
2472 EXPORT_SYMBOL(LU_BUF_NULL);
2474 void lu_buf_free(struct lu_buf *buf)
2478 LASSERT(buf->lb_len > 0);
2479 OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);
2484 EXPORT_SYMBOL(lu_buf_free);
2486 void lu_buf_alloc(struct lu_buf *buf, size_t size)
2489 LASSERT(buf->lb_buf == NULL);
2490 LASSERT(buf->lb_len == 0);
2491 OBD_ALLOC_LARGE(buf->lb_buf, size);
2492 if (likely(buf->lb_buf))
2495 EXPORT_SYMBOL(lu_buf_alloc);
2497 void lu_buf_realloc(struct lu_buf *buf, size_t size)
2500 lu_buf_alloc(buf, size);
2502 EXPORT_SYMBOL(lu_buf_realloc);
2504 struct lu_buf *lu_buf_check_and_alloc(struct lu_buf *buf, size_t len)
2506 if (buf->lb_buf == NULL && buf->lb_len == 0)
2507 lu_buf_alloc(buf, len);
2509 if ((len > buf->lb_len) && (buf->lb_buf != NULL))
2510 lu_buf_realloc(buf, len);
2514 EXPORT_SYMBOL(lu_buf_check_and_alloc);
2517 * Increase the size of the \a buf.
2518 * preserves old data in buffer
2519 * old buffer remains unchanged on error
2520 * \retval 0 or -ENOMEM
2522 int lu_buf_check_and_grow(struct lu_buf *buf, size_t len)
2526 if (len <= buf->lb_len)
2529 OBD_ALLOC_LARGE(ptr, len);
2533 /* Free the old buf */
2534 if (buf->lb_buf != NULL) {
2535 memcpy(ptr, buf->lb_buf, buf->lb_len);
2536 OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);
2543 EXPORT_SYMBOL(lu_buf_check_and_grow);