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/delay.h>
44 #include <linux/module.h>
45 #include <linux/list.h>
46 #include <linux/processor.h>
47 #include <linux/random.h>
49 #include <libcfs/libcfs.h>
50 #include <libcfs/linux/linux-mem.h>
51 #include <obd_class.h>
52 #include <obd_support.h>
53 #include <lustre_disk.h>
54 #include <lustre_fid.h>
55 #include <lu_object.h>
58 struct lu_site_bkt_data {
60 * LRU list, updated on each access to object. Protected by
63 * "Cold" end of LRU is lu_site::ls_lru.next. Accessed object are
64 * moved to the lu_site::ls_lru.prev
66 struct list_head lsb_lru;
68 * Wait-queue signaled when an object in this site is ultimately
69 * destroyed (lu_object_free()) or initialized (lu_object_start()).
70 * It is used by lu_object_find() to wait before re-trying when
71 * object in the process of destruction is found in the hash table;
72 * or wait object to be initialized by the allocator.
74 * \see htable_lookup().
76 wait_queue_head_t lsb_waitq;
80 LU_CACHE_PERCENT_MAX = 50,
81 LU_CACHE_PERCENT_DEFAULT = 20
84 #define LU_CACHE_NR_MAX_ADJUST 512
85 #define LU_CACHE_NR_UNLIMITED -1
86 #define LU_CACHE_NR_DEFAULT LU_CACHE_NR_UNLIMITED
87 /** This is set to roughly (20 * OSS_NTHRS_MAX) to prevent thrashing */
88 #define LU_CACHE_NR_ZFS_LIMIT 10240
90 #define LU_CACHE_NR_MIN 4096
91 #define LU_CACHE_NR_MAX 0x80000000UL
94 * Max 256 buckets, we don't want too many buckets because:
95 * - consume too much memory (currently max 16K)
96 * - avoid unbalanced LRU list
97 * With few cpus there is little gain from extra buckets, so
98 * we treat this as a maximum in lu_site_init().
100 #define LU_SITE_BKT_BITS 8
102 static unsigned int lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
103 module_param(lu_cache_percent, int, 0644);
104 MODULE_PARM_DESC(lu_cache_percent, "Percentage of memory to be used as lu_object cache");
106 static long lu_cache_nr = LU_CACHE_NR_DEFAULT;
107 module_param(lu_cache_nr, long, 0644);
108 MODULE_PARM_DESC(lu_cache_nr, "Maximum number of objects in lu_object cache");
110 static void lu_object_free(const struct lu_env *env, struct lu_object *o);
111 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx);
113 static u32 lu_fid_hash(const void *data, u32 len, u32 seed)
115 const struct lu_fid *fid = data;
117 seed = cfs_hash_32(seed ^ fid->f_oid, 32);
118 seed ^= cfs_hash_64(fid->f_seq, 32);
122 static const struct rhashtable_params obj_hash_params = {
123 .key_len = sizeof(struct lu_fid),
124 .key_offset = offsetof(struct lu_object_header, loh_fid),
125 .head_offset = offsetof(struct lu_object_header, loh_hash),
126 .hashfn = lu_fid_hash,
127 .automatic_shrinking = true,
130 static inline int lu_bkt_hash(struct lu_site *s, const struct lu_fid *fid)
132 return lu_fid_hash(fid, sizeof(*fid), s->ls_bkt_seed) &
137 lu_site_wq_from_fid(struct lu_site *site, struct lu_fid *fid)
139 struct lu_site_bkt_data *bkt;
141 bkt = &site->ls_bkts[lu_bkt_hash(site, fid)];
142 return &bkt->lsb_waitq;
144 EXPORT_SYMBOL(lu_site_wq_from_fid);
147 * Decrease reference counter on object. If last reference is freed, return
148 * object to the cache, unless lu_object_is_dying(o) holds. In the latter
149 * case, free object immediately.
151 void lu_object_put(const struct lu_env *env, struct lu_object *o)
153 struct lu_site_bkt_data *bkt;
154 struct lu_object_header *top = o->lo_header;
155 struct lu_site *site = o->lo_dev->ld_site;
156 struct lu_object *orig = o;
157 const struct lu_fid *fid = lu_object_fid(o);
160 * till we have full fids-on-OST implemented anonymous objects
161 * are possible in OSP. such an object isn't listed in the site
162 * so we should not remove it from the site.
164 if (fid_is_zero(fid)) {
165 LASSERT(list_empty(&top->loh_lru));
166 if (!atomic_dec_and_test(&top->loh_ref))
168 list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
169 if (o->lo_ops->loo_object_release != NULL)
170 o->lo_ops->loo_object_release(env, o);
172 lu_object_free(env, orig);
176 bkt = &site->ls_bkts[lu_bkt_hash(site, &top->loh_fid)];
177 if (atomic_add_unless(&top->loh_ref, -1, 1)) {
180 * At this point the object reference is dropped and lock is
181 * not taken, so lu_object should not be touched because it
182 * can be freed by concurrent thread.
184 * Somebody may be waiting for this, currently only used for
185 * cl_object, see cl_object_put_last().
187 wake_up(&bkt->lsb_waitq);
192 spin_lock(&bkt->lsb_waitq.lock);
193 if (!atomic_dec_and_test(&top->loh_ref)) {
194 spin_unlock(&bkt->lsb_waitq.lock);
199 * Refcount is zero, and cannot be incremented without taking the bkt
200 * lock, so object is stable.
204 * When last reference is released, iterate over object layers, and
205 * notify them that object is no longer busy.
207 list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
208 if (o->lo_ops->loo_object_release != NULL)
209 o->lo_ops->loo_object_release(env, o);
213 * Don't use local 'is_dying' here because if was taken without lock but
214 * here we need the latest actual value of it so check lu_object
217 if (!lu_object_is_dying(top) &&
218 (lu_object_exists(orig) || lu_object_is_cl(orig))) {
219 LASSERT(list_empty(&top->loh_lru));
220 list_add_tail(&top->loh_lru, &bkt->lsb_lru);
221 spin_unlock(&bkt->lsb_waitq.lock);
222 percpu_counter_inc(&site->ls_lru_len_counter);
223 CDEBUG(D_INODE, "Add %p/%p to site lru. bkt: %p\n",
229 * If object is dying (will not be cached) then remove it from hash
230 * table (it is already not on the LRU).
232 * This is done with bucket lock held. As the only way to acquire first
233 * reference to previously unreferenced object is through hash-table
234 * lookup (lu_object_find()) which takes the lock for first reference,
235 * no race with concurrent object lookup is possible and we can safely
236 * destroy object below.
238 if (!test_and_set_bit(LU_OBJECT_UNHASHED, &top->loh_flags))
239 rhashtable_remove_fast(&site->ls_obj_hash, &top->loh_hash,
242 spin_unlock(&bkt->lsb_waitq.lock);
243 /* Object was already removed from hash above, can kill it. */
244 lu_object_free(env, orig);
246 EXPORT_SYMBOL(lu_object_put);
249 * Put object and don't keep in cache. This is temporary solution for
250 * multi-site objects when its layering is not constant.
252 void lu_object_put_nocache(const struct lu_env *env, struct lu_object *o)
254 set_bit(LU_OBJECT_HEARD_BANSHEE, &o->lo_header->loh_flags);
255 return lu_object_put(env, o);
257 EXPORT_SYMBOL(lu_object_put_nocache);
260 * Kill the object and take it out of LRU cache.
261 * Currently used by client code for layout change.
263 void lu_object_unhash(const struct lu_env *env, struct lu_object *o)
265 struct lu_object_header *top;
268 set_bit(LU_OBJECT_HEARD_BANSHEE, &top->loh_flags);
269 if (!test_and_set_bit(LU_OBJECT_UNHASHED, &top->loh_flags)) {
270 struct lu_site *site = o->lo_dev->ld_site;
271 struct rhashtable *obj_hash = &site->ls_obj_hash;
272 struct lu_site_bkt_data *bkt;
274 bkt = &site->ls_bkts[lu_bkt_hash(site, &top->loh_fid)];
275 spin_lock(&bkt->lsb_waitq.lock);
276 if (!list_empty(&top->loh_lru)) {
277 list_del_init(&top->loh_lru);
278 percpu_counter_dec(&site->ls_lru_len_counter);
280 spin_unlock(&bkt->lsb_waitq.lock);
282 rhashtable_remove_fast(obj_hash, &top->loh_hash,
286 EXPORT_SYMBOL(lu_object_unhash);
289 * Allocate new object.
291 * This follows object creation protocol, described in the comment within
292 * struct lu_device_operations definition.
294 static struct lu_object *lu_object_alloc(const struct lu_env *env,
295 struct lu_device *dev,
296 const struct lu_fid *f)
298 struct lu_object *top;
301 * Create top-level object slice. This will also create
304 top = dev->ld_ops->ldo_object_alloc(env, NULL, dev);
306 return ERR_PTR(-ENOMEM);
310 * This is the only place where object fid is assigned. It's constant
313 top->lo_header->loh_fid = *f;
321 * This is called after object hash insertion to avoid returning an object with
324 static int lu_object_start(const struct lu_env *env, struct lu_device *dev,
325 struct lu_object *top,
326 const struct lu_object_conf *conf)
328 struct lu_object *scan;
329 struct list_head *layers;
330 unsigned int init_mask = 0;
331 unsigned int init_flag;
335 layers = &top->lo_header->loh_layers;
339 * Call ->loo_object_init() repeatedly, until no more new
340 * object slices are created.
344 list_for_each_entry(scan, layers, lo_linkage) {
345 if (init_mask & init_flag)
348 scan->lo_header = top->lo_header;
349 result = scan->lo_ops->loo_object_init(env, scan, conf);
353 init_mask |= init_flag;
359 list_for_each_entry_reverse(scan, layers, lo_linkage) {
360 if (scan->lo_ops->loo_object_start != NULL) {
361 result = scan->lo_ops->loo_object_start(env, scan);
367 lprocfs_counter_incr(dev->ld_site->ls_stats, LU_SS_CREATED);
369 set_bit(LU_OBJECT_INITED, &top->lo_header->loh_flags);
377 static void lu_object_free(const struct lu_env *env, struct lu_object *o)
379 wait_queue_head_t *wq;
380 struct lu_site *site;
381 struct lu_object *scan;
382 struct list_head *layers;
385 site = o->lo_dev->ld_site;
386 layers = &o->lo_header->loh_layers;
387 wq = lu_site_wq_from_fid(site, &o->lo_header->loh_fid);
389 * First call ->loo_object_delete() method to release all resources.
391 list_for_each_entry_reverse(scan, layers, lo_linkage) {
392 if (scan->lo_ops->loo_object_delete != NULL)
393 scan->lo_ops->loo_object_delete(env, scan);
397 * Then, splice object layers into stand-alone list, and call
398 * ->loo_object_free() on all layers to free memory. Splice is
399 * necessary, because lu_object_header is freed together with the
402 list_splice_init(layers, &splice);
403 while (!list_empty(&splice)) {
405 * Free layers in bottom-to-top order, so that object header
406 * lives as long as possible and ->loo_object_free() methods
407 * can look at its contents.
409 o = container_of(splice.prev, struct lu_object, lo_linkage);
410 list_del_init(&o->lo_linkage);
411 LASSERT(o->lo_ops->loo_object_free != NULL);
412 o->lo_ops->loo_object_free(env, o);
415 if (waitqueue_active(wq))
420 * Free \a nr objects from the cold end of the site LRU list.
421 * if canblock is 0, then don't block awaiting for another
422 * instance of lu_site_purge() to complete
424 int lu_site_purge_objects(const struct lu_env *env, struct lu_site *s,
425 int nr, int canblock)
427 struct lu_object_header *h;
428 struct lu_object_header *temp;
429 struct lu_site_bkt_data *bkt;
432 unsigned int start = 0;
437 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_NO_LRU))
441 * Under LRU list lock, scan LRU list and move unreferenced objects to
442 * the dispose list, removing them from LRU and hash table.
445 start = s->ls_purge_start;
446 bnr = (nr == ~0) ? -1 : nr / s->ls_bkt_cnt + 1;
449 * It doesn't make any sense to make purge threads parallel, that can
450 * only bring troubles to us. See LU-5331.
453 mutex_lock(&s->ls_purge_mutex);
454 else if (mutex_trylock(&s->ls_purge_mutex) == 0)
458 for (i = start; i < s->ls_bkt_cnt ; i++) {
460 bkt = &s->ls_bkts[i];
461 spin_lock(&bkt->lsb_waitq.lock);
463 list_for_each_entry_safe(h, temp, &bkt->lsb_lru, loh_lru) {
464 LASSERT(atomic_read(&h->loh_ref) == 0);
466 LINVRNT(lu_bkt_hash(s, &h->loh_fid) == i);
468 set_bit(LU_OBJECT_UNHASHED, &h->loh_flags);
469 rhashtable_remove_fast(&s->ls_obj_hash, &h->loh_hash,
471 list_move(&h->loh_lru, &dispose);
472 percpu_counter_dec(&s->ls_lru_len_counter);
476 if (nr != ~0 && --nr == 0)
479 if (count > 0 && --count == 0)
483 spin_unlock(&bkt->lsb_waitq.lock);
486 * Free everything on the dispose list. This is safe against
487 * races due to the reasons described in lu_object_put().
489 while ((h = list_first_entry_or_null(&dispose,
490 struct lu_object_header,
492 list_del_init(&h->loh_lru);
493 lu_object_free(env, lu_object_top(h));
494 lprocfs_counter_incr(s->ls_stats, LU_SS_LRU_PURGED);
500 mutex_unlock(&s->ls_purge_mutex);
502 if (nr != 0 && did_sth && start != 0) {
503 start = 0; /* restart from the first bucket */
506 /* race on s->ls_purge_start, but nobody cares */
507 s->ls_purge_start = i & (s->ls_bkt_cnt - 1);
511 EXPORT_SYMBOL(lu_site_purge_objects);
516 * Code below has to jump through certain loops to output object description
517 * into libcfs_debug_msg-based log. The problem is that lu_object_print()
518 * composes object description from strings that are parts of _lines_ of
519 * output (i.e., strings that are not terminated by newline). This doesn't fit
520 * very well into libcfs_debug_msg() interface that assumes that each message
521 * supplied to it is a self-contained output line.
523 * To work around this, strings are collected in a temporary buffer
524 * (implemented as a value of lu_cdebug_key key), until terminating newline
525 * character is detected.
533 * XXX overflow is not handled correctly.
538 struct lu_cdebug_data {
542 char lck_area[LU_CDEBUG_LINE];
545 /* context key constructor/destructor: lu_global_key_init, lu_global_key_fini */
546 LU_KEY_INIT_FINI(lu_global, struct lu_cdebug_data);
549 * Key, holding temporary buffer. This key is registered very early by
552 static struct lu_context_key lu_global_key = {
553 .lct_tags = LCT_MD_THREAD | LCT_DT_THREAD |
554 LCT_MG_THREAD | LCT_CL_THREAD | LCT_LOCAL,
555 .lct_init = lu_global_key_init,
556 .lct_fini = lu_global_key_fini
560 * Printer function emitting messages through libcfs_debug_msg().
562 int lu_cdebug_printer(const struct lu_env *env,
563 void *cookie, const char *format, ...)
565 struct libcfs_debug_msg_data *msgdata = cookie;
566 struct lu_cdebug_data *key;
571 va_start(args, format);
573 key = lu_context_key_get(&env->le_ctx, &lu_global_key);
574 LASSERT(key != NULL);
576 used = strlen(key->lck_area);
577 complete = format[strlen(format) - 1] == '\n';
579 * Append new chunk to the buffer.
581 vsnprintf(key->lck_area + used,
582 ARRAY_SIZE(key->lck_area) - used, format, args);
584 if (cfs_cdebug_show(msgdata->msg_mask, msgdata->msg_subsys))
585 libcfs_debug_msg(msgdata, "%s\n", key->lck_area);
586 key->lck_area[0] = 0;
591 EXPORT_SYMBOL(lu_cdebug_printer);
594 * Print object header.
596 void lu_object_header_print(const struct lu_env *env, void *cookie,
597 lu_printer_t printer,
598 const struct lu_object_header *hdr)
600 (*printer)(env, cookie, "header@%p[%#lx, %d, "DFID"%s%s%s]",
601 hdr, hdr->loh_flags, atomic_read(&hdr->loh_ref),
603 test_bit(LU_OBJECT_UNHASHED,
604 &hdr->loh_flags) ? "" : " hash",
605 list_empty(&hdr->loh_lru) ? "" : " lru",
606 hdr->loh_attr & LOHA_EXISTS ? " exist" : "");
608 EXPORT_SYMBOL(lu_object_header_print);
611 * Print human readable representation of the \a o to the \a printer.
613 void lu_object_print(const struct lu_env *env, void *cookie,
614 lu_printer_t printer, const struct lu_object *o)
616 static const char ruler[] = "........................................";
617 struct lu_object_header *top;
621 lu_object_header_print(env, cookie, printer, top);
622 (*printer)(env, cookie, "{\n");
624 list_for_each_entry(o, &top->loh_layers, lo_linkage) {
626 * print `.' \a depth times followed by type name and address
628 (*printer)(env, cookie, "%*.*s%s@%p", depth, depth, ruler,
629 o->lo_dev->ld_type->ldt_name, o);
631 if (o->lo_ops->loo_object_print != NULL)
632 (*o->lo_ops->loo_object_print)(env, cookie, printer, o);
634 (*printer)(env, cookie, "\n");
637 (*printer)(env, cookie, "} header@%p\n", top);
639 EXPORT_SYMBOL(lu_object_print);
642 * Check object consistency.
644 int lu_object_invariant(const struct lu_object *o)
646 struct lu_object_header *top;
649 list_for_each_entry(o, &top->loh_layers, lo_linkage) {
650 if (o->lo_ops->loo_object_invariant != NULL &&
651 !o->lo_ops->loo_object_invariant(o))
658 * Limit the lu_object cache to a maximum of lu_cache_nr objects. Because the
659 * calculation for the number of objects to reclaim is not covered by a lock the
660 * maximum number of objects is capped by LU_CACHE_MAX_ADJUST. This ensures
661 * that many concurrent threads will not accidentally purge the entire cache.
663 static void lu_object_limit(const struct lu_env *env,
664 struct lu_device *dev)
668 if (lu_cache_nr == LU_CACHE_NR_UNLIMITED)
671 size = atomic_read(&dev->ld_site->ls_obj_hash.nelems);
672 nr = (u64)lu_cache_nr;
676 lu_site_purge_objects(env, dev->ld_site,
677 min_t(u64, size - nr, LU_CACHE_NR_MAX_ADJUST),
681 static struct lu_object *htable_lookup(const struct lu_env *env,
682 struct lu_device *dev,
683 struct lu_site_bkt_data *bkt,
684 const struct lu_fid *f,
685 struct lu_object_header *new)
687 struct lu_site *s = dev->ld_site;
688 struct lu_object_header *h;
693 h = rhashtable_lookup_get_insert_fast(&s->ls_obj_hash,
697 h = rhashtable_lookup(&s->ls_obj_hash, f, obj_hash_params);
699 if (IS_ERR_OR_NULL(h)) {
702 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS);
704 if (PTR_ERR(h) == -ENOMEM) {
708 lu_object_limit(env, dev);
709 if (PTR_ERR(h) == -E2BIG)
712 return ERR_PTR(-ENOENT);
715 if (atomic_inc_not_zero(&h->loh_ref)) {
717 return lu_object_top(h);
720 spin_lock(&bkt->lsb_waitq.lock);
721 if (lu_object_is_dying(h) ||
722 test_bit(LU_OBJECT_UNHASHED, &h->loh_flags)) {
723 spin_unlock(&bkt->lsb_waitq.lock);
727 * Old object might have already been removed, or will
728 * be soon. We need to insert our new object, so
729 * remove the old one just in case it is still there.
731 rhashtable_remove_fast(&s->ls_obj_hash, &h->loh_hash,
735 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS);
736 return ERR_PTR(-ENOENT);
738 /* Now protected by spinlock */
741 if (!list_empty(&h->loh_lru)) {
742 list_del_init(&h->loh_lru);
743 percpu_counter_dec(&s->ls_lru_len_counter);
745 atomic_inc(&h->loh_ref);
746 spin_unlock(&bkt->lsb_waitq.lock);
747 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_HIT);
748 return lu_object_top(h);
752 * Search cache for an object with the fid \a f. If such object is found,
753 * return it. Otherwise, create new object, insert it into cache and return
754 * it. In any case, additional reference is acquired on the returned object.
756 struct lu_object *lu_object_find(const struct lu_env *env,
757 struct lu_device *dev, const struct lu_fid *f,
758 const struct lu_object_conf *conf)
760 return lu_object_find_at(env, dev->ld_site->ls_top_dev, f, conf);
762 EXPORT_SYMBOL(lu_object_find);
765 * Get a 'first' reference to an object that was found while looking through the
768 struct lu_object *lu_object_get_first(struct lu_object_header *h,
769 struct lu_device *dev)
771 struct lu_site *s = dev->ld_site;
772 struct lu_object *ret;
774 if (IS_ERR_OR_NULL(h) || lu_object_is_dying(h))
777 ret = lu_object_locate(h, dev->ld_type);
781 if (!atomic_inc_not_zero(&h->loh_ref)) {
782 struct lu_site_bkt_data *bkt;
784 bkt = &s->ls_bkts[lu_bkt_hash(s, &h->loh_fid)];
785 spin_lock(&bkt->lsb_waitq.lock);
786 if (!lu_object_is_dying(h) &&
787 !test_bit(LU_OBJECT_UNHASHED, &h->loh_flags))
788 atomic_inc(&h->loh_ref);
791 spin_unlock(&bkt->lsb_waitq.lock);
795 EXPORT_SYMBOL(lu_object_get_first);
798 * Core logic of lu_object_find*() functions.
800 * Much like lu_object_find(), but top level device of object is specifically
801 * \a dev rather than top level device of the site. This interface allows
802 * objects of different "stacking" to be created within the same site.
804 struct lu_object *lu_object_find_at(const struct lu_env *env,
805 struct lu_device *dev,
806 const struct lu_fid *f,
807 const struct lu_object_conf *conf)
810 struct lu_object *shadow;
812 struct lu_site_bkt_data *bkt;
813 struct rhashtable *hs;
819 * This uses standard index maintenance protocol:
821 * - search index under lock, and return object if found;
822 * - otherwise, unlock index, allocate new object;
823 * - lock index and search again;
824 * - if nothing is found (usual case), insert newly created
826 * - otherwise (race: other thread inserted object), free
827 * object just allocated.
831 * For "LOC_F_NEW" case, we are sure the object is new established.
832 * It is unnecessary to perform lookup-alloc-lookup-insert, instead,
833 * just alloc and insert directly.
837 hs = &s->ls_obj_hash;
839 if (unlikely(OBD_FAIL_PRECHECK(OBD_FAIL_OBD_ZERO_NLINK_RACE)))
840 lu_site_purge(env, s, -1);
842 bkt = &s->ls_bkts[lu_bkt_hash(s, f)];
843 if (!(conf && conf->loc_flags & LOC_F_NEW)) {
844 o = htable_lookup(env, dev, bkt, f, NULL);
847 if (likely(lu_object_is_inited(o->lo_header)))
850 wait_event_idle(bkt->lsb_waitq,
851 lu_object_is_inited(o->lo_header) ||
852 lu_object_is_dying(o->lo_header));
854 if (lu_object_is_dying(o->lo_header)) {
855 lu_object_put(env, o);
857 RETURN(ERR_PTR(-ENOENT));
863 if (PTR_ERR(o) != -ENOENT)
868 * Allocate new object, NB, object is unitialized in case object
869 * is changed between allocation and hash insertion, thus the object
870 * with stale attributes is returned.
872 o = lu_object_alloc(env, dev, f);
876 LASSERT(lu_fid_eq(lu_object_fid(o), f));
878 CFS_RACE_WAIT(OBD_FAIL_OBD_ZERO_NLINK_RACE);
880 if (conf && conf->loc_flags & LOC_F_NEW) {
881 int status = rhashtable_insert_fast(hs, &o->lo_header->loh_hash,
884 /* Strange error - go the slow way */
885 shadow = htable_lookup(env, dev, bkt, f, o->lo_header);
887 shadow = ERR_PTR(-ENOENT);
889 shadow = htable_lookup(env, dev, bkt, f, o->lo_header);
891 if (likely(PTR_ERR(shadow) == -ENOENT)) {
893 * The new object has been successfully inserted.
895 * This may result in rather complicated operations, including
896 * fld queries, inode loading, etc.
898 rc = lu_object_start(env, dev, o, conf);
900 lu_object_put_nocache(env, o);
904 wake_up(&bkt->lsb_waitq);
906 lu_object_limit(env, dev);
911 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_RACE);
912 lu_object_free(env, o);
914 if (!(conf && conf->loc_flags & LOC_F_NEW) &&
916 !lu_object_is_inited(shadow->lo_header)) {
917 wait_event_idle(bkt->lsb_waitq,
918 lu_object_is_inited(shadow->lo_header) ||
919 lu_object_is_dying(shadow->lo_header));
921 if (lu_object_is_dying(shadow->lo_header)) {
922 lu_object_put(env, shadow);
924 RETURN(ERR_PTR(-ENOENT));
930 EXPORT_SYMBOL(lu_object_find_at);
933 * Find object with given fid, and return its slice belonging to given device.
935 struct lu_object *lu_object_find_slice(const struct lu_env *env,
936 struct lu_device *dev,
937 const struct lu_fid *f,
938 const struct lu_object_conf *conf)
940 struct lu_object *top;
941 struct lu_object *obj;
943 top = lu_object_find(env, dev, f, conf);
947 obj = lu_object_locate(top->lo_header, dev->ld_type);
948 if (unlikely(obj == NULL)) {
949 lu_object_put(env, top);
950 obj = ERR_PTR(-ENOENT);
955 EXPORT_SYMBOL(lu_object_find_slice);
957 int lu_device_type_init(struct lu_device_type *ldt)
961 atomic_set(&ldt->ldt_device_nr, 0);
962 if (ldt->ldt_ops->ldto_init)
963 result = ldt->ldt_ops->ldto_init(ldt);
967 EXPORT_SYMBOL(lu_device_type_init);
969 void lu_device_type_fini(struct lu_device_type *ldt)
971 if (ldt->ldt_ops->ldto_fini)
972 ldt->ldt_ops->ldto_fini(ldt);
974 EXPORT_SYMBOL(lu_device_type_fini);
977 * Global list of all sites on this node
979 static LIST_HEAD(lu_sites);
980 static DECLARE_RWSEM(lu_sites_guard);
983 * Global environment used by site shrinker.
985 static struct lu_env lu_shrink_env;
987 struct lu_site_print_arg {
988 struct lu_env *lsp_env;
990 lu_printer_t lsp_printer;
994 lu_site_obj_print(struct lu_object_header *h, struct lu_site_print_arg *arg)
996 if (!list_empty(&h->loh_layers)) {
997 const struct lu_object *o;
999 o = lu_object_top(h);
1000 lu_object_print(arg->lsp_env, arg->lsp_cookie,
1001 arg->lsp_printer, o);
1003 lu_object_header_print(arg->lsp_env, arg->lsp_cookie,
1004 arg->lsp_printer, h);
1009 * Print all objects in \a s.
1011 void lu_site_print(const struct lu_env *env, struct lu_site *s, atomic_t *ref,
1012 int msg_flag, lu_printer_t printer)
1014 struct lu_site_print_arg arg = {
1015 .lsp_env = (struct lu_env *)env,
1016 .lsp_printer = printer,
1018 struct rhashtable_iter iter;
1019 struct lu_object_header *h;
1020 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, msg_flag, NULL);
1022 if (!s || !atomic_read(ref))
1025 arg.lsp_cookie = (void *)&msgdata;
1027 rhashtable_walk_enter(&s->ls_obj_hash, &iter);
1028 rhashtable_walk_start(&iter);
1029 while ((h = rhashtable_walk_next(&iter)) != NULL) {
1032 lu_site_obj_print(h, &arg);
1034 rhashtable_walk_stop(&iter);
1035 rhashtable_walk_exit(&iter);
1037 EXPORT_SYMBOL(lu_site_print);
1040 * Return desired hash table order.
1042 static void lu_htable_limits(struct lu_device *top)
1044 unsigned long cache_size;
1047 * For ZFS based OSDs the cache should be disabled by default. This
1048 * allows the ZFS ARC maximum flexibility in determining what buffers
1049 * to cache. If Lustre has objects or buffer which it wants to ensure
1050 * always stay cached it must maintain a hold on them.
1052 if (strcmp(top->ld_type->ldt_name, LUSTRE_OSD_ZFS_NAME) == 0) {
1053 lu_cache_nr = LU_CACHE_NR_ZFS_LIMIT;
1058 * Calculate hash table size, assuming that we want reasonable
1059 * performance when 20% of total memory is occupied by cache of
1062 * Size of lu_object is (arbitrary) taken as 1K (together with inode).
1064 cache_size = cfs_totalram_pages();
1066 #if BITS_PER_LONG == 32
1067 /* limit hashtable size for lowmem systems to low RAM */
1068 if (cache_size > 1 << (30 - PAGE_SHIFT))
1069 cache_size = 1 << (30 - PAGE_SHIFT) * 3 / 4;
1072 /* clear off unreasonable cache setting. */
1073 if (lu_cache_percent == 0 || lu_cache_percent > LU_CACHE_PERCENT_MAX) {
1074 CWARN("obdclass: invalid lu_cache_percent: %u, it must be in the range of (0, %u]. Will use default value: %u.\n",
1075 lu_cache_percent, LU_CACHE_PERCENT_MAX,
1076 LU_CACHE_PERCENT_DEFAULT);
1078 lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
1080 cache_size = cache_size / 100 * lu_cache_percent *
1083 lu_cache_nr = clamp_t(typeof(cache_size), cache_size,
1084 LU_CACHE_NR_MIN, LU_CACHE_NR_MAX);
1087 void lu_dev_add_linkage(struct lu_site *s, struct lu_device *d)
1089 spin_lock(&s->ls_ld_lock);
1090 if (list_empty(&d->ld_linkage))
1091 list_add(&d->ld_linkage, &s->ls_ld_linkage);
1092 spin_unlock(&s->ls_ld_lock);
1094 EXPORT_SYMBOL(lu_dev_add_linkage);
1096 void lu_dev_del_linkage(struct lu_site *s, struct lu_device *d)
1098 spin_lock(&s->ls_ld_lock);
1099 list_del_init(&d->ld_linkage);
1100 spin_unlock(&s->ls_ld_lock);
1102 EXPORT_SYMBOL(lu_dev_del_linkage);
1105 * Initialize site \a s, with \a d as the top level device.
1107 int lu_site_init(struct lu_site *s, struct lu_device *top)
1109 struct lu_site_bkt_data *bkt;
1114 memset(s, 0, sizeof *s);
1115 mutex_init(&s->ls_purge_mutex);
1116 lu_htable_limits(top);
1118 #ifdef HAVE_PERCPU_COUNTER_INIT_GFP_FLAG
1119 rc = percpu_counter_init(&s->ls_lru_len_counter, 0, GFP_NOFS);
1121 rc = percpu_counter_init(&s->ls_lru_len_counter, 0);
1126 if (rhashtable_init(&s->ls_obj_hash, &obj_hash_params) != 0) {
1127 CERROR("failed to create lu_site hash\n");
1131 s->ls_bkt_seed = prandom_u32();
1132 s->ls_bkt_cnt = max_t(long, 1 << LU_SITE_BKT_BITS,
1133 2 * num_possible_cpus());
1134 s->ls_bkt_cnt = roundup_pow_of_two(s->ls_bkt_cnt);
1135 OBD_ALLOC_PTR_ARRAY_LARGE(s->ls_bkts, s->ls_bkt_cnt);
1137 rhashtable_destroy(&s->ls_obj_hash);
1142 for (i = 0; i < s->ls_bkt_cnt; i++) {
1143 bkt = &s->ls_bkts[i];
1144 INIT_LIST_HEAD(&bkt->lsb_lru);
1145 init_waitqueue_head(&bkt->lsb_waitq);
1148 s->ls_stats = lprocfs_alloc_stats(LU_SS_LAST_STAT, 0);
1149 if (s->ls_stats == NULL) {
1150 OBD_FREE_PTR_ARRAY_LARGE(s->ls_bkts, s->ls_bkt_cnt);
1152 rhashtable_destroy(&s->ls_obj_hash);
1156 lprocfs_counter_init(s->ls_stats, LU_SS_CREATED,
1157 0, "created", "created");
1158 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_HIT,
1159 0, "cache_hit", "cache_hit");
1160 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_MISS,
1161 0, "cache_miss", "cache_miss");
1162 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_RACE,
1163 0, "cache_race", "cache_race");
1164 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_DEATH_RACE,
1165 0, "cache_death_race", "cache_death_race");
1166 lprocfs_counter_init(s->ls_stats, LU_SS_LRU_PURGED,
1167 0, "lru_purged", "lru_purged");
1169 INIT_LIST_HEAD(&s->ls_linkage);
1170 s->ls_top_dev = top;
1173 lu_ref_add(&top->ld_reference, "site-top", s);
1175 INIT_LIST_HEAD(&s->ls_ld_linkage);
1176 spin_lock_init(&s->ls_ld_lock);
1178 lu_dev_add_linkage(s, top);
1182 EXPORT_SYMBOL(lu_site_init);
1185 * Finalize \a s and release its resources.
1187 void lu_site_fini(struct lu_site *s)
1189 down_write(&lu_sites_guard);
1190 list_del_init(&s->ls_linkage);
1191 up_write(&lu_sites_guard);
1193 percpu_counter_destroy(&s->ls_lru_len_counter);
1196 rhashtable_destroy(&s->ls_obj_hash);
1197 OBD_FREE_PTR_ARRAY_LARGE(s->ls_bkts, s->ls_bkt_cnt);
1201 if (s->ls_top_dev != NULL) {
1202 s->ls_top_dev->ld_site = NULL;
1203 lu_ref_del(&s->ls_top_dev->ld_reference, "site-top", s);
1204 lu_device_put(s->ls_top_dev);
1205 s->ls_top_dev = NULL;
1208 if (s->ls_stats != NULL)
1209 lprocfs_free_stats(&s->ls_stats);
1211 EXPORT_SYMBOL(lu_site_fini);
1214 * Called when initialization of stack for this site is completed.
1216 int lu_site_init_finish(struct lu_site *s)
1219 down_write(&lu_sites_guard);
1220 result = lu_context_refill(&lu_shrink_env.le_ctx);
1222 list_add(&s->ls_linkage, &lu_sites);
1223 up_write(&lu_sites_guard);
1226 EXPORT_SYMBOL(lu_site_init_finish);
1229 * Acquire additional reference on device \a d
1231 void lu_device_get(struct lu_device *d)
1233 atomic_inc(&d->ld_ref);
1235 EXPORT_SYMBOL(lu_device_get);
1238 * Release reference on device \a d.
1240 void lu_device_put(struct lu_device *d)
1242 LASSERT(atomic_read(&d->ld_ref) > 0);
1243 atomic_dec(&d->ld_ref);
1245 EXPORT_SYMBOL(lu_device_put);
1247 enum { /* Maximal number of tld slots. */
1248 LU_CONTEXT_KEY_NR = 40
1250 static struct lu_context_key *lu_keys[LU_CONTEXT_KEY_NR] = { NULL, };
1251 static DECLARE_RWSEM(lu_key_initing);
1254 * Initialize device \a d of type \a t.
1256 int lu_device_init(struct lu_device *d, struct lu_device_type *t)
1258 if (atomic_add_unless(&t->ldt_device_nr, 1, 0) == 0) {
1259 down_write(&lu_key_initing);
1260 if (t->ldt_ops->ldto_start &&
1261 atomic_read(&t->ldt_device_nr) == 0)
1262 t->ldt_ops->ldto_start(t);
1263 atomic_inc(&t->ldt_device_nr);
1264 up_write(&lu_key_initing);
1267 memset(d, 0, sizeof *d);
1269 lu_ref_init(&d->ld_reference);
1270 INIT_LIST_HEAD(&d->ld_linkage);
1274 EXPORT_SYMBOL(lu_device_init);
1277 * Finalize device \a d.
1279 void lu_device_fini(struct lu_device *d)
1281 struct lu_device_type *t = d->ld_type;
1283 if (d->ld_obd != NULL) {
1284 d->ld_obd->obd_lu_dev = NULL;
1288 lu_ref_fini(&d->ld_reference);
1289 LASSERTF(atomic_read(&d->ld_ref) == 0,
1290 "Refcount is %u\n", atomic_read(&d->ld_ref));
1291 LASSERT(atomic_read(&t->ldt_device_nr) > 0);
1293 if (atomic_dec_and_test(&t->ldt_device_nr) &&
1294 t->ldt_ops->ldto_stop != NULL)
1295 t->ldt_ops->ldto_stop(t);
1297 EXPORT_SYMBOL(lu_device_fini);
1300 * Initialize object \a o that is part of compound object \a h and was created
1303 int lu_object_init(struct lu_object *o, struct lu_object_header *h,
1304 struct lu_device *d)
1306 memset(o, 0, sizeof(*o));
1310 lu_ref_add_at(&d->ld_reference, &o->lo_dev_ref, "lu_object", o);
1311 INIT_LIST_HEAD(&o->lo_linkage);
1315 EXPORT_SYMBOL(lu_object_init);
1318 * Finalize object and release its resources.
1320 void lu_object_fini(struct lu_object *o)
1322 struct lu_device *dev = o->lo_dev;
1324 LASSERT(list_empty(&o->lo_linkage));
1327 lu_ref_del_at(&dev->ld_reference, &o->lo_dev_ref,
1333 EXPORT_SYMBOL(lu_object_fini);
1336 * Add object \a o as first layer of compound object \a h
1338 * This is typically called by the ->ldo_object_alloc() method of top-level
1341 void lu_object_add_top(struct lu_object_header *h, struct lu_object *o)
1343 list_move(&o->lo_linkage, &h->loh_layers);
1345 EXPORT_SYMBOL(lu_object_add_top);
1348 * Add object \a o as a layer of compound object, going after \a before.
1350 * This is typically called by the ->ldo_object_alloc() method of \a
1353 void lu_object_add(struct lu_object *before, struct lu_object *o)
1355 list_move(&o->lo_linkage, &before->lo_linkage);
1357 EXPORT_SYMBOL(lu_object_add);
1360 * Initialize compound object.
1362 int lu_object_header_init(struct lu_object_header *h)
1364 memset(h, 0, sizeof *h);
1365 atomic_set(&h->loh_ref, 1);
1366 INIT_LIST_HEAD(&h->loh_lru);
1367 INIT_LIST_HEAD(&h->loh_layers);
1368 lu_ref_init(&h->loh_reference);
1371 EXPORT_SYMBOL(lu_object_header_init);
1374 * Finalize compound object.
1376 void lu_object_header_fini(struct lu_object_header *h)
1378 LASSERT(list_empty(&h->loh_layers));
1379 LASSERT(list_empty(&h->loh_lru));
1380 lu_ref_fini(&h->loh_reference);
1382 EXPORT_SYMBOL(lu_object_header_fini);
1385 * Given a compound object, find its slice, corresponding to the device type
1388 struct lu_object *lu_object_locate(struct lu_object_header *h,
1389 const struct lu_device_type *dtype)
1391 struct lu_object *o;
1393 list_for_each_entry(o, &h->loh_layers, lo_linkage) {
1394 if (o->lo_dev->ld_type == dtype)
1399 EXPORT_SYMBOL(lu_object_locate);
1402 * Finalize and free devices in the device stack.
1404 * Finalize device stack by purging object cache, and calling
1405 * lu_device_type_operations::ldto_device_fini() and
1406 * lu_device_type_operations::ldto_device_free() on all devices in the stack.
1408 void lu_stack_fini(const struct lu_env *env, struct lu_device *top)
1410 struct lu_site *site = top->ld_site;
1411 struct lu_device *scan;
1412 struct lu_device *next;
1414 lu_site_purge(env, site, ~0);
1415 for (scan = top; scan != NULL; scan = next) {
1416 next = scan->ld_type->ldt_ops->ldto_device_fini(env, scan);
1417 lu_ref_del(&scan->ld_reference, "lu-stack", &lu_site_init);
1418 lu_device_put(scan);
1422 lu_site_purge(env, site, ~0);
1424 for (scan = top; scan != NULL; scan = next) {
1425 const struct lu_device_type *ldt = scan->ld_type;
1427 next = ldt->ldt_ops->ldto_device_free(env, scan);
1432 * Global counter incremented whenever key is registered, unregistered,
1433 * revived or quiesced. This is used to void unnecessary calls to
1434 * lu_context_refill(). No locking is provided, as initialization and shutdown
1435 * are supposed to be externally serialized.
1437 static atomic_t key_set_version = ATOMIC_INIT(0);
1442 int lu_context_key_register(struct lu_context_key *key)
1447 LASSERT(key->lct_init != NULL);
1448 LASSERT(key->lct_fini != NULL);
1449 LASSERT(key->lct_tags != 0);
1450 LASSERT(key->lct_owner != NULL);
1453 atomic_set(&key->lct_used, 1);
1454 lu_ref_init(&key->lct_reference);
1455 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1460 if (strncmp("osd_", module_name(key->lct_owner), 4) == 0)
1461 CFS_RACE_WAIT(OBD_FAIL_OBD_SETUP);
1463 if (cmpxchg(&lu_keys[i], NULL, key) != NULL)
1467 atomic_inc(&key_set_version);
1471 lu_ref_fini(&key->lct_reference);
1472 atomic_set(&key->lct_used, 0);
1476 EXPORT_SYMBOL(lu_context_key_register);
1478 static void key_fini(struct lu_context *ctx, int index)
1480 if (ctx->lc_value != NULL && ctx->lc_value[index] != NULL) {
1481 struct lu_context_key *key;
1483 key = lu_keys[index];
1484 LASSERT(key != NULL);
1485 LASSERT(key->lct_fini != NULL);
1486 LASSERT(atomic_read(&key->lct_used) > 0);
1488 key->lct_fini(ctx, key, ctx->lc_value[index]);
1489 lu_ref_del(&key->lct_reference, "ctx", ctx);
1490 if (atomic_dec_and_test(&key->lct_used))
1491 wake_up_var(&key->lct_used);
1493 LASSERT(key->lct_owner != NULL);
1494 if ((ctx->lc_tags & LCT_NOREF) == 0) {
1495 LINVRNT(module_refcount(key->lct_owner) > 0);
1496 module_put(key->lct_owner);
1498 ctx->lc_value[index] = NULL;
1505 void lu_context_key_degister(struct lu_context_key *key)
1507 LASSERT(atomic_read(&key->lct_used) >= 1);
1508 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1510 lu_context_key_quiesce(NULL, key);
1512 key_fini(&lu_shrink_env.le_ctx, key->lct_index);
1515 * Wait until all transient contexts referencing this key have
1516 * run lu_context_key::lct_fini() method.
1518 atomic_dec(&key->lct_used);
1519 wait_var_event(&key->lct_used, atomic_read(&key->lct_used) == 0);
1521 if (!WARN_ON(lu_keys[key->lct_index] == NULL))
1522 lu_ref_fini(&key->lct_reference);
1524 smp_store_release(&lu_keys[key->lct_index], NULL);
1526 EXPORT_SYMBOL(lu_context_key_degister);
1529 * Register a number of keys. This has to be called after all keys have been
1530 * initialized by a call to LU_CONTEXT_KEY_INIT().
1532 int lu_context_key_register_many(struct lu_context_key *k, ...)
1534 struct lu_context_key *key = k;
1540 result = lu_context_key_register(key);
1543 key = va_arg(args, struct lu_context_key *);
1544 } while (key != NULL);
1550 lu_context_key_degister(k);
1551 k = va_arg(args, struct lu_context_key *);
1558 EXPORT_SYMBOL(lu_context_key_register_many);
1561 * De-register a number of keys. This is a dual to
1562 * lu_context_key_register_many().
1564 void lu_context_key_degister_many(struct lu_context_key *k, ...)
1570 lu_context_key_degister(k);
1571 k = va_arg(args, struct lu_context_key*);
1572 } while (k != NULL);
1575 EXPORT_SYMBOL(lu_context_key_degister_many);
1578 * Revive a number of keys.
1580 void lu_context_key_revive_many(struct lu_context_key *k, ...)
1586 lu_context_key_revive(k);
1587 k = va_arg(args, struct lu_context_key*);
1588 } while (k != NULL);
1591 EXPORT_SYMBOL(lu_context_key_revive_many);
1594 * Quiescent a number of keys.
1596 void lu_context_key_quiesce_many(struct lu_device_type *t,
1597 struct lu_context_key *k, ...)
1603 lu_context_key_quiesce(t, k);
1604 k = va_arg(args, struct lu_context_key*);
1605 } while (k != NULL);
1608 EXPORT_SYMBOL(lu_context_key_quiesce_many);
1611 * Return value associated with key \a key in context \a ctx.
1613 void *lu_context_key_get(const struct lu_context *ctx,
1614 const struct lu_context_key *key)
1616 LINVRNT(ctx->lc_state == LCS_ENTERED);
1617 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1618 LASSERT(lu_keys[key->lct_index] == key);
1619 return ctx->lc_value[key->lct_index];
1621 EXPORT_SYMBOL(lu_context_key_get);
1624 * List of remembered contexts. XXX document me.
1626 static LIST_HEAD(lu_context_remembered);
1627 static DEFINE_SPINLOCK(lu_context_remembered_guard);
1630 * Destroy \a key in all remembered contexts. This is used to destroy key
1631 * values in "shared" contexts (like service threads), when a module owning
1632 * the key is about to be unloaded.
1634 void lu_context_key_quiesce(struct lu_device_type *t,
1635 struct lu_context_key *key)
1637 struct lu_context *ctx;
1639 if (key->lct_tags & LCT_QUIESCENT)
1642 * The write-lock on lu_key_initing will ensure that any
1643 * keys_fill() which didn't see LCT_QUIESCENT will have
1644 * finished before we call key_fini().
1646 down_write(&lu_key_initing);
1647 if (!(key->lct_tags & LCT_QUIESCENT)) {
1648 if (t == NULL || atomic_read(&t->ldt_device_nr) == 0)
1649 key->lct_tags |= LCT_QUIESCENT;
1650 up_write(&lu_key_initing);
1652 spin_lock(&lu_context_remembered_guard);
1653 list_for_each_entry(ctx, &lu_context_remembered, lc_remember) {
1654 spin_until_cond(READ_ONCE(ctx->lc_state) != LCS_LEAVING);
1655 key_fini(ctx, key->lct_index);
1657 spin_unlock(&lu_context_remembered_guard);
1661 up_write(&lu_key_initing);
1664 void lu_context_key_revive(struct lu_context_key *key)
1666 key->lct_tags &= ~LCT_QUIESCENT;
1667 atomic_inc(&key_set_version);
1670 static void keys_fini(struct lu_context *ctx)
1674 if (ctx->lc_value == NULL)
1677 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i)
1680 OBD_FREE_PTR_ARRAY(ctx->lc_value, ARRAY_SIZE(lu_keys));
1681 ctx->lc_value = NULL;
1684 static int keys_fill(struct lu_context *ctx)
1690 * A serialisation with lu_context_key_quiesce() is needed, to
1691 * ensure we see LCT_QUIESCENT and don't allocate a new value
1692 * after it freed one. The rwsem provides this. As down_read()
1693 * does optimistic spinning while the writer is active, this is
1694 * unlikely to ever sleep.
1696 down_read(&lu_key_initing);
1697 ctx->lc_version = atomic_read(&key_set_version);
1699 LINVRNT(ctx->lc_value);
1700 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1701 struct lu_context_key *key;
1704 if (!ctx->lc_value[i] && key &&
1705 (key->lct_tags & ctx->lc_tags) &&
1707 * Don't create values for a LCT_QUIESCENT key, as this
1708 * will pin module owning a key.
1710 !(key->lct_tags & LCT_QUIESCENT)) {
1713 LINVRNT(key->lct_init != NULL);
1714 LINVRNT(key->lct_index == i);
1716 LASSERT(key->lct_owner != NULL);
1717 if (!(ctx->lc_tags & LCT_NOREF) &&
1718 try_module_get(key->lct_owner) == 0) {
1719 /* module is unloading, skip this key */
1723 value = key->lct_init(ctx, key);
1724 if (unlikely(IS_ERR(value))) {
1725 rc = PTR_ERR(value);
1729 lu_ref_add_atomic(&key->lct_reference, "ctx", ctx);
1730 atomic_inc(&key->lct_used);
1732 * This is the only place in the code, where an
1733 * element of ctx->lc_value[] array is set to non-NULL
1736 ctx->lc_value[i] = value;
1737 if (key->lct_exit != NULL)
1738 ctx->lc_tags |= LCT_HAS_EXIT;
1742 up_read(&lu_key_initing);
1746 static int keys_init(struct lu_context *ctx)
1748 OBD_ALLOC_PTR_ARRAY(ctx->lc_value, ARRAY_SIZE(lu_keys));
1749 if (likely(ctx->lc_value != NULL))
1750 return keys_fill(ctx);
1756 * Initialize context data-structure. Create values for all keys.
1758 int lu_context_init(struct lu_context *ctx, __u32 tags)
1762 memset(ctx, 0, sizeof *ctx);
1763 ctx->lc_state = LCS_INITIALIZED;
1764 ctx->lc_tags = tags;
1765 if (tags & LCT_REMEMBER) {
1766 spin_lock(&lu_context_remembered_guard);
1767 list_add(&ctx->lc_remember, &lu_context_remembered);
1768 spin_unlock(&lu_context_remembered_guard);
1770 INIT_LIST_HEAD(&ctx->lc_remember);
1773 rc = keys_init(ctx);
1775 lu_context_fini(ctx);
1779 EXPORT_SYMBOL(lu_context_init);
1782 * Finalize context data-structure. Destroy key values.
1784 void lu_context_fini(struct lu_context *ctx)
1786 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1787 ctx->lc_state = LCS_FINALIZED;
1789 if ((ctx->lc_tags & LCT_REMEMBER) == 0) {
1790 LASSERT(list_empty(&ctx->lc_remember));
1792 /* could race with key degister */
1793 spin_lock(&lu_context_remembered_guard);
1794 list_del_init(&ctx->lc_remember);
1795 spin_unlock(&lu_context_remembered_guard);
1799 EXPORT_SYMBOL(lu_context_fini);
1802 * Called before entering context.
1804 void lu_context_enter(struct lu_context *ctx)
1806 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1807 ctx->lc_state = LCS_ENTERED;
1809 EXPORT_SYMBOL(lu_context_enter);
1812 * Called after exiting from \a ctx
1814 void lu_context_exit(struct lu_context *ctx)
1818 LINVRNT(ctx->lc_state == LCS_ENTERED);
1820 * Disable preempt to ensure we get a warning if
1821 * any lct_exit ever tries to sleep. That would hurt
1822 * lu_context_key_quiesce() which spins waiting for us.
1823 * This also ensure we aren't preempted while the state
1824 * is LCS_LEAVING, as that too would cause problems for
1825 * lu_context_key_quiesce().
1829 * Ensure lu_context_key_quiesce() sees LCS_LEAVING
1830 * or we see LCT_QUIESCENT
1832 smp_store_mb(ctx->lc_state, LCS_LEAVING);
1833 if (ctx->lc_tags & LCT_HAS_EXIT && ctx->lc_value) {
1834 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1835 struct lu_context_key *key;
1838 if (ctx->lc_value[i] &&
1839 !(key->lct_tags & LCT_QUIESCENT) &&
1841 key->lct_exit(ctx, key, ctx->lc_value[i]);
1845 smp_store_release(&ctx->lc_state, LCS_LEFT);
1848 EXPORT_SYMBOL(lu_context_exit);
1851 * Allocate for context all missing keys that were registered after context
1852 * creation. key_set_version is only changed in rare cases when modules
1853 * are loaded and removed.
1855 int lu_context_refill(struct lu_context *ctx)
1857 if (likely(ctx->lc_version == atomic_read(&key_set_version)))
1860 return keys_fill(ctx);
1864 * lu_ctx_tags/lu_ses_tags will be updated if there are new types of
1865 * obd being added. Currently, this is only used on client side, specifically
1866 * for echo device client, for other stack (like ptlrpc threads), context are
1867 * predefined when the lu_device type are registered, during the module probe
1870 u32 lu_context_tags_default = LCT_CL_THREAD;
1871 u32 lu_session_tags_default = LCT_SESSION;
1873 void lu_context_tags_update(__u32 tags)
1875 spin_lock(&lu_context_remembered_guard);
1876 lu_context_tags_default |= tags;
1877 atomic_inc(&key_set_version);
1878 spin_unlock(&lu_context_remembered_guard);
1880 EXPORT_SYMBOL(lu_context_tags_update);
1882 void lu_context_tags_clear(__u32 tags)
1884 spin_lock(&lu_context_remembered_guard);
1885 lu_context_tags_default &= ~tags;
1886 atomic_inc(&key_set_version);
1887 spin_unlock(&lu_context_remembered_guard);
1889 EXPORT_SYMBOL(lu_context_tags_clear);
1891 void lu_session_tags_update(__u32 tags)
1893 spin_lock(&lu_context_remembered_guard);
1894 lu_session_tags_default |= tags;
1895 atomic_inc(&key_set_version);
1896 spin_unlock(&lu_context_remembered_guard);
1898 EXPORT_SYMBOL(lu_session_tags_update);
1900 void lu_session_tags_clear(__u32 tags)
1902 spin_lock(&lu_context_remembered_guard);
1903 lu_session_tags_default &= ~tags;
1904 atomic_inc(&key_set_version);
1905 spin_unlock(&lu_context_remembered_guard);
1907 EXPORT_SYMBOL(lu_session_tags_clear);
1909 int lu_env_init(struct lu_env *env, __u32 tags)
1914 result = lu_context_init(&env->le_ctx, tags);
1915 if (likely(result == 0))
1916 lu_context_enter(&env->le_ctx);
1919 EXPORT_SYMBOL(lu_env_init);
1921 void lu_env_fini(struct lu_env *env)
1923 lu_context_exit(&env->le_ctx);
1924 lu_context_fini(&env->le_ctx);
1927 EXPORT_SYMBOL(lu_env_fini);
1929 int lu_env_refill(struct lu_env *env)
1933 result = lu_context_refill(&env->le_ctx);
1934 if (result == 0 && env->le_ses != NULL)
1935 result = lu_context_refill(env->le_ses);
1938 EXPORT_SYMBOL(lu_env_refill);
1941 * Currently, this API will only be used by echo client.
1942 * Because echo client and normal lustre client will share
1943 * same cl_env cache. So echo client needs to refresh
1944 * the env context after it get one from the cache, especially
1945 * when normal client and echo client co-exist in the same client.
1947 int lu_env_refill_by_tags(struct lu_env *env, __u32 ctags,
1952 if ((env->le_ctx.lc_tags & ctags) != ctags) {
1953 env->le_ctx.lc_version = 0;
1954 env->le_ctx.lc_tags |= ctags;
1957 if (env->le_ses && (env->le_ses->lc_tags & stags) != stags) {
1958 env->le_ses->lc_version = 0;
1959 env->le_ses->lc_tags |= stags;
1962 result = lu_env_refill(env);
1966 EXPORT_SYMBOL(lu_env_refill_by_tags);
1969 struct lu_env_item {
1970 struct task_struct *lei_task; /* rhashtable key */
1971 struct rhash_head lei_linkage;
1972 struct lu_env *lei_env;
1973 struct rcu_head lei_rcu_head;
1976 static const struct rhashtable_params lu_env_rhash_params = {
1977 .key_len = sizeof(struct task_struct *),
1978 .key_offset = offsetof(struct lu_env_item, lei_task),
1979 .head_offset = offsetof(struct lu_env_item, lei_linkage),
1982 struct rhashtable lu_env_rhash;
1984 struct lu_env_percpu {
1985 struct task_struct *lep_task;
1986 struct lu_env *lep_env ____cacheline_aligned_in_smp;
1989 static struct lu_env_percpu lu_env_percpu[NR_CPUS];
1991 int lu_env_add_task(struct lu_env *env, struct task_struct *task)
1993 struct lu_env_item *lei, *old;
2001 lei->lei_task = task;
2004 old = rhashtable_lookup_get_insert_fast(&lu_env_rhash,
2006 lu_env_rhash_params);
2011 EXPORT_SYMBOL(lu_env_add_task);
2013 int lu_env_add(struct lu_env *env)
2015 return lu_env_add_task(env, current);
2017 EXPORT_SYMBOL(lu_env_add);
2019 static void lu_env_item_free(struct rcu_head *head)
2021 struct lu_env_item *lei;
2023 lei = container_of(head, struct lu_env_item, lei_rcu_head);
2027 void lu_env_remove(struct lu_env *env)
2029 struct lu_env_item *lei;
2030 const void *task = current;
2033 for_each_possible_cpu(i) {
2034 if (lu_env_percpu[i].lep_env == env) {
2035 LASSERT(lu_env_percpu[i].lep_task == task);
2036 lu_env_percpu[i].lep_task = NULL;
2037 lu_env_percpu[i].lep_env = NULL;
2041 /* The rcu_lock is not taking in this case since the key
2042 * used is the actual task_struct. This implies that each
2043 * object is only removed by the owning thread, so there
2044 * can never be a race on a particular object.
2046 lei = rhashtable_lookup_fast(&lu_env_rhash, &task,
2047 lu_env_rhash_params);
2048 if (lei && rhashtable_remove_fast(&lu_env_rhash, &lei->lei_linkage,
2049 lu_env_rhash_params) == 0)
2050 call_rcu(&lei->lei_rcu_head, lu_env_item_free);
2052 EXPORT_SYMBOL(lu_env_remove);
2054 struct lu_env *lu_env_find(void)
2056 struct lu_env *env = NULL;
2057 struct lu_env_item *lei;
2058 const void *task = current;
2061 if (lu_env_percpu[i].lep_task == current) {
2062 env = lu_env_percpu[i].lep_env;
2068 lei = rhashtable_lookup_fast(&lu_env_rhash, &task,
2069 lu_env_rhash_params);
2072 lu_env_percpu[i].lep_task = current;
2073 lu_env_percpu[i].lep_env = env;
2079 EXPORT_SYMBOL(lu_env_find);
2081 static struct shrinker *lu_site_shrinker;
2083 typedef struct lu_site_stats{
2084 unsigned lss_populated;
2085 unsigned lss_max_search;
2090 static void lu_site_stats_get(const struct lu_site *s,
2091 lu_site_stats_t *stats)
2093 int cnt = atomic_read(&s->ls_obj_hash.nelems);
2095 * percpu_counter_sum_positive() won't accept a const pointer
2096 * as it does modify the struct by taking a spinlock
2098 struct lu_site *s2 = (struct lu_site *)s;
2100 stats->lss_busy += cnt -
2101 percpu_counter_sum_positive(&s2->ls_lru_len_counter);
2103 stats->lss_total += cnt;
2104 stats->lss_max_search = 0;
2105 stats->lss_populated = 0;
2110 * lu_cache_shrink_count() returns an approximate number of cached objects
2111 * that can be freed by shrink_slab(). A counter, which tracks the
2112 * number of items in the site's lru, is maintained in a percpu_counter
2113 * for each site. The percpu values are incremented and decremented as
2114 * objects are added or removed from the lru. The percpu values are summed
2115 * and saved whenever a percpu value exceeds a threshold. Thus the saved,
2116 * summed value at any given time may not accurately reflect the current
2117 * lru length. But this value is sufficiently accurate for the needs of
2120 * Using a per cpu counter is a compromise solution to concurrent access:
2121 * lu_object_put() can update the counter without locking the site and
2122 * lu_cache_shrink_count can sum the counters without locking each
2123 * ls_obj_hash bucket.
2125 static unsigned long lu_cache_shrink_count(struct shrinker *sk,
2126 struct shrink_control *sc)
2129 struct lu_site *tmp;
2130 unsigned long cached = 0;
2132 if (!(sc->gfp_mask & __GFP_FS))
2135 down_read(&lu_sites_guard);
2136 list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage)
2137 cached += percpu_counter_read_positive(&s->ls_lru_len_counter);
2138 up_read(&lu_sites_guard);
2140 cached = (cached / 100) * sysctl_vfs_cache_pressure;
2141 CDEBUG(D_INODE, "%ld objects cached, cache pressure %d\n",
2142 cached, sysctl_vfs_cache_pressure);
2147 static unsigned long lu_cache_shrink_scan(struct shrinker *sk,
2148 struct shrink_control *sc)
2151 struct lu_site *tmp;
2152 unsigned long remain = sc->nr_to_scan;
2155 if (!(sc->gfp_mask & __GFP_FS))
2156 /* We must not take the lu_sites_guard lock when
2157 * __GFP_FS is *not* set because of the deadlock
2158 * possibility detailed above. Additionally,
2159 * since we cannot determine the number of
2160 * objects in the cache without taking this
2161 * lock, we're in a particularly tough spot. As
2162 * a result, we'll just lie and say our cache is
2163 * empty. This _should_ be ok, as we can't
2164 * reclaim objects when __GFP_FS is *not* set
2169 down_write(&lu_sites_guard);
2170 list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
2171 remain = lu_site_purge(&lu_shrink_env, s, remain);
2173 * Move just shrunk site to the tail of site list to
2174 * assure shrinking fairness.
2176 list_move_tail(&s->ls_linkage, &splice);
2178 list_splice(&splice, lu_sites.prev);
2179 up_write(&lu_sites_guard);
2181 return sc->nr_to_scan - remain;
2184 #ifndef HAVE_SHRINKER_COUNT
2186 * There exists a potential lock inversion deadlock scenario when using
2187 * Lustre on top of ZFS. This occurs between one of ZFS's
2188 * buf_hash_table.ht_lock's, and Lustre's lu_sites_guard lock. Essentially,
2189 * thread A will take the lu_sites_guard lock and sleep on the ht_lock,
2190 * while thread B will take the ht_lock and sleep on the lu_sites_guard
2191 * lock. Obviously neither thread will wake and drop their respective hold
2194 * To prevent this from happening we must ensure the lu_sites_guard lock is
2195 * not taken while down this code path. ZFS reliably does not set the
2196 * __GFP_FS bit in its code paths, so this can be used to determine if it
2197 * is safe to take the lu_sites_guard lock.
2199 * Ideally we should accurately return the remaining number of cached
2200 * objects without taking the lu_sites_guard lock, but this is not
2201 * possible in the current implementation.
2203 static int lu_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
2206 struct shrink_control scv = {
2207 .nr_to_scan = shrink_param(sc, nr_to_scan),
2208 .gfp_mask = shrink_param(sc, gfp_mask)
2211 CDEBUG(D_INODE, "Shrink %lu objects\n", scv.nr_to_scan);
2213 if (scv.nr_to_scan != 0)
2214 lu_cache_shrink_scan(shrinker, &scv);
2216 cached = lu_cache_shrink_count(shrinker, &scv);
2220 #endif /* HAVE_SHRINKER_COUNT */
2228 * Environment to be used in debugger, contains all tags.
2230 static struct lu_env lu_debugging_env;
2233 * Debugging printer function using printk().
2235 int lu_printk_printer(const struct lu_env *env,
2236 void *unused, const char *format, ...)
2240 va_start(args, format);
2241 vprintk(format, args);
2246 int lu_debugging_setup(void)
2248 return lu_env_init(&lu_debugging_env, ~0);
2251 void lu_context_keys_dump(void)
2255 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
2256 struct lu_context_key *key;
2260 CERROR("[%d]: %p %x (%p,%p,%p) %d %d \"%s\"@%p\n",
2261 i, key, key->lct_tags,
2262 key->lct_init, key->lct_fini, key->lct_exit,
2263 key->lct_index, atomic_read(&key->lct_used),
2264 key->lct_owner ? key->lct_owner->name : "",
2266 lu_ref_print(&key->lct_reference);
2272 * Initialization of global lu_* data.
2274 int lu_global_init(void)
2277 DEF_SHRINKER_VAR(shvar, lu_cache_shrink,
2278 lu_cache_shrink_count, lu_cache_shrink_scan);
2280 CDEBUG(D_INFO, "Lustre LU module (%p).\n", &lu_keys);
2282 result = lu_ref_global_init();
2286 LU_CONTEXT_KEY_INIT(&lu_global_key);
2287 result = lu_context_key_register(&lu_global_key);
2292 * At this level, we don't know what tags are needed, so allocate them
2293 * conservatively. This should not be too bad, because this
2294 * environment is global.
2296 down_write(&lu_sites_guard);
2297 result = lu_env_init(&lu_shrink_env, LCT_SHRINKER);
2298 up_write(&lu_sites_guard);
2303 * seeks estimation: 3 seeks to read a record from oi, one to read
2304 * inode, one for ea. Unfortunately setting this high value results in
2305 * lu_object/inode cache consuming all the memory.
2307 lu_site_shrinker = set_shrinker(DEFAULT_SEEKS, &shvar);
2308 if (lu_site_shrinker == NULL)
2311 result = rhashtable_init(&lu_env_rhash, &lu_env_rhash_params);
2317 * Dual to lu_global_init().
2319 void lu_global_fini(void)
2321 if (lu_site_shrinker != NULL) {
2322 remove_shrinker(lu_site_shrinker);
2323 lu_site_shrinker = NULL;
2326 lu_context_key_degister(&lu_global_key);
2329 * Tear shrinker environment down _after_ de-registering
2330 * lu_global_key, because the latter has a value in the former.
2332 down_write(&lu_sites_guard);
2333 lu_env_fini(&lu_shrink_env);
2334 up_write(&lu_sites_guard);
2336 rhashtable_destroy(&lu_env_rhash);
2338 lu_ref_global_fini();
2341 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx)
2343 #ifdef CONFIG_PROC_FS
2344 struct lprocfs_counter ret;
2346 lprocfs_stats_collect(stats, idx, &ret);
2347 return (__u32)ret.lc_count;
2354 * Output site statistical counters into a buffer. Suitable for
2355 * lprocfs_rd_*()-style functions.
2357 int lu_site_stats_seq_print(const struct lu_site *s, struct seq_file *m)
2359 const struct bucket_table *tbl;
2360 lu_site_stats_t stats;
2361 unsigned int chains;
2363 memset(&stats, 0, sizeof(stats));
2364 lu_site_stats_get(s, &stats);
2367 tbl = rht_dereference_rcu(s->ls_obj_hash.tbl,
2368 &((struct lu_site *)s)->ls_obj_hash);
2371 seq_printf(m, "%d/%d %d/%u %d %d %d %d %d %d %d\n",
2374 stats.lss_populated,
2376 stats.lss_max_search,
2377 ls_stats_read(s->ls_stats, LU_SS_CREATED),
2378 ls_stats_read(s->ls_stats, LU_SS_CACHE_HIT),
2379 ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
2380 ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
2381 ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
2382 ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED));
2385 EXPORT_SYMBOL(lu_site_stats_seq_print);
2388 * Helper function to initialize a number of kmem slab caches at once.
2390 int lu_kmem_init(struct lu_kmem_descr *caches)
2393 struct lu_kmem_descr *iter = caches;
2395 for (result = 0; iter->ckd_cache != NULL; ++iter) {
2396 *iter->ckd_cache = kmem_cache_create(iter->ckd_name,
2399 if (*iter->ckd_cache == NULL) {
2401 /* free all previously allocated caches */
2402 lu_kmem_fini(caches);
2408 EXPORT_SYMBOL(lu_kmem_init);
2411 * Helper function to finalize a number of kmem slab cached at once. Dual to
2414 void lu_kmem_fini(struct lu_kmem_descr *caches)
2416 for (; caches->ckd_cache != NULL; ++caches) {
2417 if (*caches->ckd_cache != NULL) {
2418 kmem_cache_destroy(*caches->ckd_cache);
2419 *caches->ckd_cache = NULL;
2423 EXPORT_SYMBOL(lu_kmem_fini);
2426 * Temporary solution to be able to assign fid in ->do_create()
2427 * till we have fully-functional OST fids
2429 void lu_object_assign_fid(const struct lu_env *env, struct lu_object *o,
2430 const struct lu_fid *fid)
2432 struct lu_site *s = o->lo_dev->ld_site;
2433 struct lu_fid *old = &o->lo_header->loh_fid;
2436 LASSERT(fid_is_zero(old));
2439 rc = rhashtable_lookup_insert_fast(&s->ls_obj_hash,
2440 &o->lo_header->loh_hash,
2442 /* supposed to be unique */
2443 LASSERT(rc != -EEXIST);
2444 /* handle hash table resizing */
2445 if (rc == -ENOMEM) {
2449 /* trim the hash if its growing to big */
2450 lu_object_limit(env, o->lo_dev);
2454 LASSERTF(rc == 0, "failed hashtable insertion: rc = %d\n", rc);
2456 EXPORT_SYMBOL(lu_object_assign_fid);
2459 * allocates object with 0 (non-assiged) fid
2460 * XXX: temporary solution to be able to assign fid in ->do_create()
2461 * till we have fully-functional OST fids
2463 struct lu_object *lu_object_anon(const struct lu_env *env,
2464 struct lu_device *dev,
2465 const struct lu_object_conf *conf)
2468 struct lu_object *o;
2472 o = lu_object_alloc(env, dev, &fid);
2474 rc = lu_object_start(env, dev, o, conf);
2476 lu_object_free(env, o);
2483 EXPORT_SYMBOL(lu_object_anon);
2485 struct lu_buf LU_BUF_NULL = {
2489 EXPORT_SYMBOL(LU_BUF_NULL);
2491 void lu_buf_free(struct lu_buf *buf)
2495 LASSERT(buf->lb_len > 0);
2496 OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);
2501 EXPORT_SYMBOL(lu_buf_free);
2503 void lu_buf_alloc(struct lu_buf *buf, size_t size)
2506 LASSERT(buf->lb_buf == NULL);
2507 LASSERT(buf->lb_len == 0);
2508 OBD_ALLOC_LARGE(buf->lb_buf, size);
2509 if (likely(buf->lb_buf))
2512 EXPORT_SYMBOL(lu_buf_alloc);
2514 void lu_buf_realloc(struct lu_buf *buf, size_t size)
2517 lu_buf_alloc(buf, size);
2519 EXPORT_SYMBOL(lu_buf_realloc);
2521 struct lu_buf *lu_buf_check_and_alloc(struct lu_buf *buf, size_t len)
2523 if (buf->lb_buf == NULL && buf->lb_len == 0)
2524 lu_buf_alloc(buf, len);
2526 if ((len > buf->lb_len) && (buf->lb_buf != NULL))
2527 lu_buf_realloc(buf, len);
2531 EXPORT_SYMBOL(lu_buf_check_and_alloc);
2534 * Increase the size of the \a buf.
2535 * preserves old data in buffer
2536 * old buffer remains unchanged on error
2537 * \retval 0 or -ENOMEM
2539 int lu_buf_check_and_grow(struct lu_buf *buf, size_t len)
2543 if (len <= buf->lb_len)
2546 OBD_ALLOC_LARGE(ptr, len);
2550 /* Free the old buf */
2551 if (buf->lb_buf != NULL) {
2552 memcpy(ptr, buf->lb_buf, buf->lb_len);
2553 OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);
2560 EXPORT_SYMBOL(lu_buf_check_and_grow);