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/
31 * lustre/obdclass/lu_object.c
34 * These are the only exported functions, they provide some generic
35 * infrastructure for managing object devices
37 * Author: Nikita Danilov <nikita.danilov@sun.com>
40 #define DEBUG_SUBSYSTEM S_CLASS
42 #include <linux/delay.h>
43 #include <linux/module.h>
44 #include <linux/list.h>
45 #include <linux/processor.h>
46 #include <linux/random.h>
48 #include <libcfs/libcfs.h>
49 #include <libcfs/linux/linux-mem.h>
50 #include <libcfs/linux/linux-hash.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;
818 /* FID is from disk or network, zero FID is meaningless, return error
819 * early to avoid assertion in lu_object_put. If a zero FID is wanted,
820 * it should be allocated via lu_object_anon().
823 RETURN(ERR_PTR(-EINVAL));
826 * This uses standard index maintenance protocol:
828 * - search index under lock, and return object if found;
829 * - otherwise, unlock index, allocate new object;
830 * - lock index and search again;
831 * - if nothing is found (usual case), insert newly created
833 * - otherwise (race: other thread inserted object), free
834 * object just allocated.
838 * For "LOC_F_NEW" case, we are sure the object is new established.
839 * It is unnecessary to perform lookup-alloc-lookup-insert, instead,
840 * just alloc and insert directly.
844 hs = &s->ls_obj_hash;
846 if (unlikely(OBD_FAIL_PRECHECK(OBD_FAIL_OBD_ZERO_NLINK_RACE)))
847 lu_site_purge(env, s, -1);
849 bkt = &s->ls_bkts[lu_bkt_hash(s, f)];
850 if (!(conf && conf->loc_flags & LOC_F_NEW)) {
851 o = htable_lookup(env, dev, bkt, f, NULL);
854 if (likely(lu_object_is_inited(o->lo_header)))
857 wait_event_idle(bkt->lsb_waitq,
858 lu_object_is_inited(o->lo_header) ||
859 lu_object_is_dying(o->lo_header));
861 if (lu_object_is_dying(o->lo_header)) {
862 lu_object_put(env, o);
864 RETURN(ERR_PTR(-ENOENT));
870 if (PTR_ERR(o) != -ENOENT)
875 * Allocate new object, NB, object is unitialized in case object
876 * is changed between allocation and hash insertion, thus the object
877 * with stale attributes is returned.
879 o = lu_object_alloc(env, dev, f);
883 LASSERT(lu_fid_eq(lu_object_fid(o), f));
885 CFS_RACE_WAIT(OBD_FAIL_OBD_ZERO_NLINK_RACE);
887 if (conf && conf->loc_flags & LOC_F_NEW) {
888 int status = rhashtable_insert_fast(hs, &o->lo_header->loh_hash,
891 /* Strange error - go the slow way */
892 shadow = htable_lookup(env, dev, bkt, f, o->lo_header);
894 shadow = ERR_PTR(-ENOENT);
896 shadow = htable_lookup(env, dev, bkt, f, o->lo_header);
898 if (likely(PTR_ERR(shadow) == -ENOENT)) {
900 * The new object has been successfully inserted.
902 * This may result in rather complicated operations, including
903 * fld queries, inode loading, etc.
905 rc = lu_object_start(env, dev, o, conf);
907 lu_object_put_nocache(env, o);
911 wake_up(&bkt->lsb_waitq);
913 lu_object_limit(env, dev);
918 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_RACE);
919 lu_object_free(env, o);
921 if (!(conf && conf->loc_flags & LOC_F_NEW) &&
923 !lu_object_is_inited(shadow->lo_header)) {
924 wait_event_idle(bkt->lsb_waitq,
925 lu_object_is_inited(shadow->lo_header) ||
926 lu_object_is_dying(shadow->lo_header));
928 if (lu_object_is_dying(shadow->lo_header)) {
929 lu_object_put(env, shadow);
931 RETURN(ERR_PTR(-ENOENT));
937 EXPORT_SYMBOL(lu_object_find_at);
940 * Find object with given fid, and return its slice belonging to given device.
942 struct lu_object *lu_object_find_slice(const struct lu_env *env,
943 struct lu_device *dev,
944 const struct lu_fid *f,
945 const struct lu_object_conf *conf)
947 struct lu_object *top;
948 struct lu_object *obj;
950 top = lu_object_find(env, dev, f, conf);
954 obj = lu_object_locate(top->lo_header, dev->ld_type);
955 if (unlikely(obj == NULL)) {
956 lu_object_put(env, top);
957 obj = ERR_PTR(-ENOENT);
962 EXPORT_SYMBOL(lu_object_find_slice);
964 int lu_device_type_init(struct lu_device_type *ldt)
968 atomic_set(&ldt->ldt_device_nr, 0);
969 if (ldt->ldt_ops->ldto_init)
970 result = ldt->ldt_ops->ldto_init(ldt);
974 EXPORT_SYMBOL(lu_device_type_init);
976 void lu_device_type_fini(struct lu_device_type *ldt)
978 if (ldt->ldt_ops->ldto_fini)
979 ldt->ldt_ops->ldto_fini(ldt);
981 EXPORT_SYMBOL(lu_device_type_fini);
984 * Global list of all sites on this node
986 static LIST_HEAD(lu_sites);
987 static DECLARE_RWSEM(lu_sites_guard);
990 * Global environment used by site shrinker.
992 static struct lu_env lu_shrink_env;
994 struct lu_site_print_arg {
995 struct lu_env *lsp_env;
997 lu_printer_t lsp_printer;
1001 lu_site_obj_print(struct lu_object_header *h, struct lu_site_print_arg *arg)
1003 if (!list_empty(&h->loh_layers)) {
1004 const struct lu_object *o;
1006 o = lu_object_top(h);
1007 lu_object_print(arg->lsp_env, arg->lsp_cookie,
1008 arg->lsp_printer, o);
1010 lu_object_header_print(arg->lsp_env, arg->lsp_cookie,
1011 arg->lsp_printer, h);
1016 * Print all objects in \a s.
1018 void lu_site_print(const struct lu_env *env, struct lu_site *s, atomic_t *ref,
1019 int msg_flag, lu_printer_t printer)
1021 struct lu_site_print_arg arg = {
1022 .lsp_env = (struct lu_env *)env,
1023 .lsp_printer = printer,
1025 struct rhashtable_iter iter;
1026 struct lu_object_header *h;
1027 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, msg_flag, NULL);
1029 if (!s || !atomic_read(ref))
1032 arg.lsp_cookie = (void *)&msgdata;
1034 rhashtable_walk_enter(&s->ls_obj_hash, &iter);
1035 rhashtable_walk_start(&iter);
1036 while ((h = rhashtable_walk_next(&iter)) != NULL) {
1039 lu_site_obj_print(h, &arg);
1041 rhashtable_walk_stop(&iter);
1042 rhashtable_walk_exit(&iter);
1044 EXPORT_SYMBOL(lu_site_print);
1047 * Return desired hash table order.
1049 static void lu_htable_limits(struct lu_device *top)
1051 unsigned long cache_size;
1054 * For ZFS based OSDs the cache should be disabled by default. This
1055 * allows the ZFS ARC maximum flexibility in determining what buffers
1056 * to cache. If Lustre has objects or buffer which it wants to ensure
1057 * always stay cached it must maintain a hold on them.
1059 if (strcmp(top->ld_type->ldt_name, LUSTRE_OSD_ZFS_NAME) == 0) {
1060 lu_cache_nr = LU_CACHE_NR_ZFS_LIMIT;
1065 * Calculate hash table size, assuming that we want reasonable
1066 * performance when 20% of total memory is occupied by cache of
1069 * Size of lu_object is (arbitrary) taken as 1K (together with inode).
1071 cache_size = cfs_totalram_pages();
1073 #if BITS_PER_LONG == 32
1074 /* limit hashtable size for lowmem systems to low RAM */
1075 if (cache_size > 1 << (30 - PAGE_SHIFT))
1076 cache_size = 1 << (30 - PAGE_SHIFT) * 3 / 4;
1079 /* clear off unreasonable cache setting. */
1080 if (lu_cache_percent == 0 || lu_cache_percent > LU_CACHE_PERCENT_MAX) {
1081 CWARN("obdclass: invalid lu_cache_percent: %u, it must be in the range of (0, %u]. Will use default value: %u.\n",
1082 lu_cache_percent, LU_CACHE_PERCENT_MAX,
1083 LU_CACHE_PERCENT_DEFAULT);
1085 lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
1087 cache_size = cache_size / 100 * lu_cache_percent *
1090 lu_cache_nr = clamp_t(typeof(cache_size), cache_size,
1091 LU_CACHE_NR_MIN, LU_CACHE_NR_MAX);
1094 void lu_dev_add_linkage(struct lu_site *s, struct lu_device *d)
1096 spin_lock(&s->ls_ld_lock);
1097 if (list_empty(&d->ld_linkage))
1098 list_add(&d->ld_linkage, &s->ls_ld_linkage);
1099 spin_unlock(&s->ls_ld_lock);
1101 EXPORT_SYMBOL(lu_dev_add_linkage);
1103 void lu_dev_del_linkage(struct lu_site *s, struct lu_device *d)
1105 spin_lock(&s->ls_ld_lock);
1106 list_del_init(&d->ld_linkage);
1107 spin_unlock(&s->ls_ld_lock);
1109 EXPORT_SYMBOL(lu_dev_del_linkage);
1112 * Initialize site \a s, with \a d as the top level device.
1114 int lu_site_init(struct lu_site *s, struct lu_device *top)
1116 struct lu_site_bkt_data *bkt;
1121 memset(s, 0, sizeof *s);
1122 mutex_init(&s->ls_purge_mutex);
1123 lu_htable_limits(top);
1125 #ifdef HAVE_PERCPU_COUNTER_INIT_GFP_FLAG
1126 rc = percpu_counter_init(&s->ls_lru_len_counter, 0, GFP_NOFS);
1128 rc = percpu_counter_init(&s->ls_lru_len_counter, 0);
1133 if (rhashtable_init(&s->ls_obj_hash, &obj_hash_params) != 0) {
1134 CERROR("failed to create lu_site hash\n");
1138 s->ls_bkt_seed = get_random_u32();
1139 s->ls_bkt_cnt = max_t(long, 1 << LU_SITE_BKT_BITS,
1140 2 * num_possible_cpus());
1141 s->ls_bkt_cnt = roundup_pow_of_two(s->ls_bkt_cnt);
1142 OBD_ALLOC_PTR_ARRAY_LARGE(s->ls_bkts, s->ls_bkt_cnt);
1144 rhashtable_destroy(&s->ls_obj_hash);
1149 for (i = 0; i < s->ls_bkt_cnt; i++) {
1150 bkt = &s->ls_bkts[i];
1151 INIT_LIST_HEAD(&bkt->lsb_lru);
1152 init_waitqueue_head(&bkt->lsb_waitq);
1155 s->ls_stats = lprocfs_stats_alloc(LU_SS_LAST_STAT, 0);
1156 if (s->ls_stats == NULL) {
1157 OBD_FREE_PTR_ARRAY_LARGE(s->ls_bkts, s->ls_bkt_cnt);
1159 rhashtable_destroy(&s->ls_obj_hash);
1163 lprocfs_counter_init(s->ls_stats, LU_SS_CREATED, 0, "created");
1164 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_HIT, 0, "cache_hit");
1165 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_MISS, 0, "cache_miss");
1166 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_RACE, 0, "cache_race");
1167 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_DEATH_RACE,
1168 0, "cache_death_race");
1169 lprocfs_counter_init(s->ls_stats, LU_SS_LRU_PURGED, 0, "lru_purged");
1171 INIT_LIST_HEAD(&s->ls_linkage);
1172 s->ls_top_dev = top;
1175 lu_ref_add(&top->ld_reference, "site-top", s);
1177 INIT_LIST_HEAD(&s->ls_ld_linkage);
1178 spin_lock_init(&s->ls_ld_lock);
1180 lu_dev_add_linkage(s, top);
1184 EXPORT_SYMBOL(lu_site_init);
1187 * Finalize \a s and release its resources.
1189 void lu_site_fini(struct lu_site *s)
1191 down_write(&lu_sites_guard);
1192 list_del_init(&s->ls_linkage);
1193 up_write(&lu_sites_guard);
1195 percpu_counter_destroy(&s->ls_lru_len_counter);
1198 rhashtable_destroy(&s->ls_obj_hash);
1199 OBD_FREE_PTR_ARRAY_LARGE(s->ls_bkts, s->ls_bkt_cnt);
1203 if (s->ls_top_dev != NULL) {
1204 s->ls_top_dev->ld_site = NULL;
1205 lu_ref_del(&s->ls_top_dev->ld_reference, "site-top", s);
1206 lu_device_put(s->ls_top_dev);
1207 s->ls_top_dev = NULL;
1210 if (s->ls_stats != NULL)
1211 lprocfs_stats_free(&s->ls_stats);
1213 EXPORT_SYMBOL(lu_site_fini);
1216 * Called when initialization of stack for this site is completed.
1218 int lu_site_init_finish(struct lu_site *s)
1222 down_write(&lu_sites_guard);
1223 result = lu_context_refill(&lu_shrink_env.le_ctx);
1225 list_add(&s->ls_linkage, &lu_sites);
1226 up_write(&lu_sites_guard);
1230 EXPORT_SYMBOL(lu_site_init_finish);
1233 * Acquire additional reference on device \a d
1235 void lu_device_get(struct lu_device *d)
1237 atomic_inc(&d->ld_ref);
1239 EXPORT_SYMBOL(lu_device_get);
1242 * Release reference on device \a d.
1244 void lu_device_put(struct lu_device *d)
1246 LASSERT(atomic_read(&d->ld_ref) > 0);
1247 atomic_dec(&d->ld_ref);
1249 EXPORT_SYMBOL(lu_device_put);
1251 enum { /* Maximal number of tld slots. */
1252 LU_CONTEXT_KEY_NR = 40
1254 static struct lu_context_key *lu_keys[LU_CONTEXT_KEY_NR] = { NULL, };
1255 static DECLARE_RWSEM(lu_key_initing);
1258 * Initialize device \a d of type \a t.
1260 int lu_device_init(struct lu_device *d, struct lu_device_type *t)
1262 if (atomic_add_unless(&t->ldt_device_nr, 1, 0) == 0) {
1263 down_write(&lu_key_initing);
1264 if (t->ldt_ops->ldto_start &&
1265 atomic_read(&t->ldt_device_nr) == 0)
1266 t->ldt_ops->ldto_start(t);
1267 atomic_inc(&t->ldt_device_nr);
1268 up_write(&lu_key_initing);
1271 memset(d, 0, sizeof *d);
1273 lu_ref_init(&d->ld_reference);
1274 INIT_LIST_HEAD(&d->ld_linkage);
1278 EXPORT_SYMBOL(lu_device_init);
1281 * Finalize device \a d.
1283 void lu_device_fini(struct lu_device *d)
1285 struct lu_device_type *t = d->ld_type;
1287 if (d->ld_obd != NULL) {
1288 d->ld_obd->obd_lu_dev = NULL;
1292 lu_ref_fini(&d->ld_reference);
1293 LASSERTF(atomic_read(&d->ld_ref) == 0,
1294 "Refcount is %u\n", atomic_read(&d->ld_ref));
1295 LASSERT(atomic_read(&t->ldt_device_nr) > 0);
1297 if (atomic_dec_and_test(&t->ldt_device_nr) &&
1298 t->ldt_ops->ldto_stop != NULL)
1299 t->ldt_ops->ldto_stop(t);
1301 EXPORT_SYMBOL(lu_device_fini);
1304 * Initialize object \a o that is part of compound object \a h and was created
1307 int lu_object_init(struct lu_object *o, struct lu_object_header *h,
1308 struct lu_device *d)
1310 memset(o, 0, sizeof(*o));
1314 lu_ref_add_at(&d->ld_reference, &o->lo_dev_ref, "lu_object", o);
1315 INIT_LIST_HEAD(&o->lo_linkage);
1319 EXPORT_SYMBOL(lu_object_init);
1322 * Finalize object and release its resources.
1324 void lu_object_fini(struct lu_object *o)
1326 struct lu_device *dev = o->lo_dev;
1328 LASSERT(list_empty(&o->lo_linkage));
1331 lu_ref_del_at(&dev->ld_reference, &o->lo_dev_ref,
1337 EXPORT_SYMBOL(lu_object_fini);
1340 * Add object \a o as first layer of compound object \a h
1342 * This is typically called by the ->ldo_object_alloc() method of top-level
1345 void lu_object_add_top(struct lu_object_header *h, struct lu_object *o)
1347 list_move(&o->lo_linkage, &h->loh_layers);
1349 EXPORT_SYMBOL(lu_object_add_top);
1352 * Add object \a o as a layer of compound object, going after \a before.
1354 * This is typically called by the ->ldo_object_alloc() method of \a
1357 void lu_object_add(struct lu_object *before, struct lu_object *o)
1359 list_move(&o->lo_linkage, &before->lo_linkage);
1361 EXPORT_SYMBOL(lu_object_add);
1364 * Initialize compound object.
1366 int lu_object_header_init(struct lu_object_header *h)
1368 memset(h, 0, sizeof *h);
1369 atomic_set(&h->loh_ref, 1);
1370 INIT_LIST_HEAD(&h->loh_lru);
1371 INIT_LIST_HEAD(&h->loh_layers);
1372 lu_ref_init(&h->loh_reference);
1375 EXPORT_SYMBOL(lu_object_header_init);
1378 * Finalize compound object.
1380 void lu_object_header_fini(struct lu_object_header *h)
1382 LASSERT(list_empty(&h->loh_layers));
1383 LASSERT(list_empty(&h->loh_lru));
1384 lu_ref_fini(&h->loh_reference);
1386 EXPORT_SYMBOL(lu_object_header_fini);
1389 * Free lu_object_header with proper RCU handling
1391 void lu_object_header_free(struct lu_object_header *h)
1393 lu_object_header_fini(h);
1394 OBD_FREE_PRE(h, sizeof(*h), "kfreed");
1395 kfree_rcu(h, loh_rcu);
1397 EXPORT_SYMBOL(lu_object_header_free);
1400 * Given a compound object, find its slice, corresponding to the device type
1403 struct lu_object *lu_object_locate(struct lu_object_header *h,
1404 const struct lu_device_type *dtype)
1406 struct lu_object *o;
1408 list_for_each_entry(o, &h->loh_layers, lo_linkage) {
1409 if (o->lo_dev->ld_type == dtype)
1414 EXPORT_SYMBOL(lu_object_locate);
1417 * Finalize and free devices in the device stack.
1419 * Finalize device stack by purging object cache, and calling
1420 * lu_device_type_operations::ldto_device_fini() and
1421 * lu_device_type_operations::ldto_device_free() on all devices in the stack.
1423 void lu_stack_fini(const struct lu_env *env, struct lu_device *top)
1425 struct lu_site *site = top->ld_site;
1426 struct lu_device *scan;
1427 struct lu_device *next;
1429 lu_site_purge(env, site, ~0);
1430 for (scan = top; scan != NULL; scan = next) {
1431 next = scan->ld_type->ldt_ops->ldto_device_fini(env, scan);
1432 lu_ref_del(&scan->ld_reference, "lu-stack", &lu_site_init);
1433 lu_device_put(scan);
1437 lu_site_purge(env, site, ~0);
1439 for (scan = top; scan != NULL; scan = next) {
1440 const struct lu_device_type *ldt = scan->ld_type;
1442 next = ldt->ldt_ops->ldto_device_free(env, scan);
1447 * Global counter incremented whenever key is registered, unregistered,
1448 * revived or quiesced. This is used to void unnecessary calls to
1449 * lu_context_refill(). No locking is provided, as initialization and shutdown
1450 * are supposed to be externally serialized.
1452 static atomic_t key_set_version = ATOMIC_INIT(0);
1457 int lu_context_key_register(struct lu_context_key *key)
1462 LASSERT(key->lct_init != NULL);
1463 LASSERT(key->lct_fini != NULL);
1464 LASSERT(key->lct_tags != 0);
1465 LASSERT(key->lct_owner != NULL);
1468 atomic_set(&key->lct_used, 1);
1469 lu_ref_init(&key->lct_reference);
1470 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1475 if (strncmp("osd_", module_name(key->lct_owner), 4) == 0)
1476 CFS_RACE_WAIT(OBD_FAIL_OBD_SETUP);
1478 if (cmpxchg(&lu_keys[i], NULL, key) != NULL)
1482 atomic_inc(&key_set_version);
1486 lu_ref_fini(&key->lct_reference);
1487 atomic_set(&key->lct_used, 0);
1491 EXPORT_SYMBOL(lu_context_key_register);
1493 static void key_fini(struct lu_context *ctx, int index)
1495 if (ctx->lc_value != NULL && ctx->lc_value[index] != NULL) {
1496 struct lu_context_key *key;
1498 key = lu_keys[index];
1499 LASSERT(key != NULL);
1500 LASSERT(key->lct_fini != NULL);
1501 LASSERT(atomic_read(&key->lct_used) > 0);
1503 key->lct_fini(ctx, key, ctx->lc_value[index]);
1504 lu_ref_del(&key->lct_reference, "ctx", ctx);
1505 if (atomic_dec_and_test(&key->lct_used))
1506 wake_up_var(&key->lct_used);
1508 LASSERT(key->lct_owner != NULL);
1509 if ((ctx->lc_tags & LCT_NOREF) == 0) {
1510 LINVRNT(module_refcount(key->lct_owner) > 0);
1511 module_put(key->lct_owner);
1513 ctx->lc_value[index] = NULL;
1520 void lu_context_key_degister(struct lu_context_key *key)
1522 LASSERT(atomic_read(&key->lct_used) >= 1);
1523 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1525 lu_context_key_quiesce(NULL, key);
1527 key_fini(&lu_shrink_env.le_ctx, key->lct_index);
1530 * Wait until all transient contexts referencing this key have
1531 * run lu_context_key::lct_fini() method.
1533 atomic_dec(&key->lct_used);
1534 wait_var_event(&key->lct_used, atomic_read(&key->lct_used) == 0);
1536 if (!WARN_ON(lu_keys[key->lct_index] == NULL))
1537 lu_ref_fini(&key->lct_reference);
1539 smp_store_release(&lu_keys[key->lct_index], NULL);
1541 EXPORT_SYMBOL(lu_context_key_degister);
1544 * Register a number of keys. This has to be called after all keys have been
1545 * initialized by a call to LU_CONTEXT_KEY_INIT().
1547 int lu_context_key_register_many(struct lu_context_key *k, ...)
1549 struct lu_context_key *key = k;
1555 result = lu_context_key_register(key);
1558 key = va_arg(args, struct lu_context_key *);
1559 } while (key != NULL);
1565 lu_context_key_degister(k);
1566 k = va_arg(args, struct lu_context_key *);
1573 EXPORT_SYMBOL(lu_context_key_register_many);
1576 * De-register a number of keys. This is a dual to
1577 * lu_context_key_register_many().
1579 void lu_context_key_degister_many(struct lu_context_key *k, ...)
1585 lu_context_key_degister(k);
1586 k = va_arg(args, struct lu_context_key*);
1587 } while (k != NULL);
1590 EXPORT_SYMBOL(lu_context_key_degister_many);
1593 * Revive a number of keys.
1595 void lu_context_key_revive_many(struct lu_context_key *k, ...)
1601 lu_context_key_revive(k);
1602 k = va_arg(args, struct lu_context_key*);
1603 } while (k != NULL);
1606 EXPORT_SYMBOL(lu_context_key_revive_many);
1609 * Quiescent a number of keys.
1611 void lu_context_key_quiesce_many(struct lu_device_type *t,
1612 struct lu_context_key *k, ...)
1618 lu_context_key_quiesce(t, k);
1619 k = va_arg(args, struct lu_context_key*);
1620 } while (k != NULL);
1623 EXPORT_SYMBOL(lu_context_key_quiesce_many);
1626 * Return value associated with key \a key in context \a ctx.
1628 void *lu_context_key_get(const struct lu_context *ctx,
1629 const struct lu_context_key *key)
1631 LINVRNT(ctx->lc_state == LCS_ENTERED);
1632 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1633 LASSERT(lu_keys[key->lct_index] == key);
1634 return ctx->lc_value[key->lct_index];
1636 EXPORT_SYMBOL(lu_context_key_get);
1639 * List of remembered contexts. XXX document me.
1641 static LIST_HEAD(lu_context_remembered);
1642 static DEFINE_SPINLOCK(lu_context_remembered_guard);
1645 * Destroy \a key in all remembered contexts. This is used to destroy key
1646 * values in "shared" contexts (like service threads), when a module owning
1647 * the key is about to be unloaded.
1649 void lu_context_key_quiesce(struct lu_device_type *t,
1650 struct lu_context_key *key)
1652 struct lu_context *ctx;
1654 if (key->lct_tags & LCT_QUIESCENT)
1657 * The write-lock on lu_key_initing will ensure that any
1658 * keys_fill() which didn't see LCT_QUIESCENT will have
1659 * finished before we call key_fini().
1661 down_write(&lu_key_initing);
1662 if (!(key->lct_tags & LCT_QUIESCENT)) {
1663 if (t == NULL || atomic_read(&t->ldt_device_nr) == 0)
1664 key->lct_tags |= LCT_QUIESCENT;
1665 up_write(&lu_key_initing);
1667 spin_lock(&lu_context_remembered_guard);
1668 list_for_each_entry(ctx, &lu_context_remembered, lc_remember) {
1669 spin_until_cond(READ_ONCE(ctx->lc_state) != LCS_LEAVING);
1670 key_fini(ctx, key->lct_index);
1672 spin_unlock(&lu_context_remembered_guard);
1676 up_write(&lu_key_initing);
1679 void lu_context_key_revive(struct lu_context_key *key)
1681 key->lct_tags &= ~LCT_QUIESCENT;
1682 atomic_inc(&key_set_version);
1685 static void keys_fini(struct lu_context *ctx)
1689 if (ctx->lc_value == NULL)
1692 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i)
1695 OBD_FREE_PTR_ARRAY(ctx->lc_value, ARRAY_SIZE(lu_keys));
1696 ctx->lc_value = NULL;
1699 static int keys_fill(struct lu_context *ctx)
1705 * A serialisation with lu_context_key_quiesce() is needed, to
1706 * ensure we see LCT_QUIESCENT and don't allocate a new value
1707 * after it freed one. The rwsem provides this. As down_read()
1708 * does optimistic spinning while the writer is active, this is
1709 * unlikely to ever sleep.
1711 down_read(&lu_key_initing);
1712 ctx->lc_version = atomic_read(&key_set_version);
1714 LINVRNT(ctx->lc_value);
1715 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1716 struct lu_context_key *key;
1719 if (!ctx->lc_value[i] && key &&
1720 (key->lct_tags & ctx->lc_tags) &&
1722 * Don't create values for a LCT_QUIESCENT key, as this
1723 * will pin module owning a key.
1725 !(key->lct_tags & LCT_QUIESCENT)) {
1728 LINVRNT(key->lct_init != NULL);
1729 LINVRNT(key->lct_index == i);
1731 LASSERT(key->lct_owner != NULL);
1732 if (!(ctx->lc_tags & LCT_NOREF) &&
1733 try_module_get(key->lct_owner) == 0) {
1734 /* module is unloading, skip this key */
1738 value = key->lct_init(ctx, key);
1739 if (unlikely(IS_ERR(value))) {
1740 rc = PTR_ERR(value);
1744 lu_ref_add_atomic(&key->lct_reference, "ctx", ctx);
1745 atomic_inc(&key->lct_used);
1747 * This is the only place in the code, where an
1748 * element of ctx->lc_value[] array is set to non-NULL
1751 ctx->lc_value[i] = value;
1752 if (key->lct_exit != NULL)
1753 ctx->lc_tags |= LCT_HAS_EXIT;
1757 up_read(&lu_key_initing);
1761 static int keys_init(struct lu_context *ctx)
1763 OBD_ALLOC_PTR_ARRAY(ctx->lc_value, ARRAY_SIZE(lu_keys));
1764 if (likely(ctx->lc_value != NULL))
1765 return keys_fill(ctx);
1771 * Initialize context data-structure. Create values for all keys.
1773 int lu_context_init(struct lu_context *ctx, __u32 tags)
1777 memset(ctx, 0, sizeof *ctx);
1778 ctx->lc_state = LCS_INITIALIZED;
1779 ctx->lc_tags = tags;
1780 if (tags & LCT_REMEMBER) {
1781 spin_lock(&lu_context_remembered_guard);
1782 list_add(&ctx->lc_remember, &lu_context_remembered);
1783 spin_unlock(&lu_context_remembered_guard);
1785 INIT_LIST_HEAD(&ctx->lc_remember);
1788 rc = keys_init(ctx);
1790 lu_context_fini(ctx);
1794 EXPORT_SYMBOL(lu_context_init);
1797 * Finalize context data-structure. Destroy key values.
1799 void lu_context_fini(struct lu_context *ctx)
1801 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1802 ctx->lc_state = LCS_FINALIZED;
1804 if ((ctx->lc_tags & LCT_REMEMBER) == 0) {
1805 LASSERT(list_empty(&ctx->lc_remember));
1807 /* could race with key degister */
1808 spin_lock(&lu_context_remembered_guard);
1809 list_del_init(&ctx->lc_remember);
1810 spin_unlock(&lu_context_remembered_guard);
1814 EXPORT_SYMBOL(lu_context_fini);
1817 * Called before entering context.
1819 void lu_context_enter(struct lu_context *ctx)
1821 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1822 ctx->lc_state = LCS_ENTERED;
1824 EXPORT_SYMBOL(lu_context_enter);
1827 * Called after exiting from \a ctx
1829 void lu_context_exit(struct lu_context *ctx)
1833 LINVRNT(ctx->lc_state == LCS_ENTERED);
1835 * Disable preempt to ensure we get a warning if
1836 * any lct_exit ever tries to sleep. That would hurt
1837 * lu_context_key_quiesce() which spins waiting for us.
1838 * This also ensure we aren't preempted while the state
1839 * is LCS_LEAVING, as that too would cause problems for
1840 * lu_context_key_quiesce().
1844 * Ensure lu_context_key_quiesce() sees LCS_LEAVING
1845 * or we see LCT_QUIESCENT
1847 smp_store_mb(ctx->lc_state, LCS_LEAVING);
1848 if (ctx->lc_tags & LCT_HAS_EXIT && ctx->lc_value) {
1849 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1850 struct lu_context_key *key;
1853 if (ctx->lc_value[i] &&
1854 !(key->lct_tags & LCT_QUIESCENT) &&
1856 key->lct_exit(ctx, key, ctx->lc_value[i]);
1860 smp_store_release(&ctx->lc_state, LCS_LEFT);
1863 EXPORT_SYMBOL(lu_context_exit);
1866 * Allocate for context all missing keys that were registered after context
1867 * creation. key_set_version is only changed in rare cases when modules
1868 * are loaded and removed.
1870 int lu_context_refill(struct lu_context *ctx)
1872 if (likely(ctx->lc_version == atomic_read(&key_set_version)))
1875 return keys_fill(ctx);
1879 * lu_ctx_tags/lu_ses_tags will be updated if there are new types of
1880 * obd being added. Currently, this is only used on client side, specifically
1881 * for echo device client, for other stack (like ptlrpc threads), context are
1882 * predefined when the lu_device type are registered, during the module probe
1885 u32 lu_context_tags_default = LCT_CL_THREAD;
1886 u32 lu_session_tags_default = LCT_SESSION;
1888 void lu_context_tags_update(__u32 tags)
1890 spin_lock(&lu_context_remembered_guard);
1891 lu_context_tags_default |= tags;
1892 atomic_inc(&key_set_version);
1893 spin_unlock(&lu_context_remembered_guard);
1895 EXPORT_SYMBOL(lu_context_tags_update);
1897 void lu_context_tags_clear(__u32 tags)
1899 spin_lock(&lu_context_remembered_guard);
1900 lu_context_tags_default &= ~tags;
1901 atomic_inc(&key_set_version);
1902 spin_unlock(&lu_context_remembered_guard);
1904 EXPORT_SYMBOL(lu_context_tags_clear);
1906 void lu_session_tags_update(__u32 tags)
1908 spin_lock(&lu_context_remembered_guard);
1909 lu_session_tags_default |= tags;
1910 atomic_inc(&key_set_version);
1911 spin_unlock(&lu_context_remembered_guard);
1913 EXPORT_SYMBOL(lu_session_tags_update);
1915 void lu_session_tags_clear(__u32 tags)
1917 spin_lock(&lu_context_remembered_guard);
1918 lu_session_tags_default &= ~tags;
1919 atomic_inc(&key_set_version);
1920 spin_unlock(&lu_context_remembered_guard);
1922 EXPORT_SYMBOL(lu_session_tags_clear);
1924 int lu_env_init(struct lu_env *env, __u32 tags)
1929 result = lu_context_init(&env->le_ctx, tags);
1930 if (likely(result == 0))
1931 lu_context_enter(&env->le_ctx);
1934 EXPORT_SYMBOL(lu_env_init);
1936 void lu_env_fini(struct lu_env *env)
1938 lu_context_exit(&env->le_ctx);
1939 lu_context_fini(&env->le_ctx);
1942 EXPORT_SYMBOL(lu_env_fini);
1944 int lu_env_refill(struct lu_env *env)
1948 result = lu_context_refill(&env->le_ctx);
1949 if (result == 0 && env->le_ses != NULL)
1950 result = lu_context_refill(env->le_ses);
1953 EXPORT_SYMBOL(lu_env_refill);
1956 * Currently, this API will only be used by echo client.
1957 * Because echo client and normal lustre client will share
1958 * same cl_env cache. So echo client needs to refresh
1959 * the env context after it get one from the cache, especially
1960 * when normal client and echo client co-exist in the same client.
1962 int lu_env_refill_by_tags(struct lu_env *env, __u32 ctags,
1967 if ((env->le_ctx.lc_tags & ctags) != ctags) {
1968 env->le_ctx.lc_version = 0;
1969 env->le_ctx.lc_tags |= ctags;
1972 if (env->le_ses && (env->le_ses->lc_tags & stags) != stags) {
1973 env->le_ses->lc_version = 0;
1974 env->le_ses->lc_tags |= stags;
1977 result = lu_env_refill(env);
1981 EXPORT_SYMBOL(lu_env_refill_by_tags);
1984 struct lu_env_item {
1985 struct task_struct *lei_task; /* rhashtable key */
1986 struct rhash_head lei_linkage;
1987 struct lu_env *lei_env;
1988 struct rcu_head lei_rcu_head;
1991 static const struct rhashtable_params lu_env_rhash_params = {
1992 .key_len = sizeof(struct task_struct *),
1993 .key_offset = offsetof(struct lu_env_item, lei_task),
1994 .head_offset = offsetof(struct lu_env_item, lei_linkage),
1997 struct rhashtable lu_env_rhash;
1999 struct lu_env_percpu {
2000 struct task_struct *lep_task;
2001 struct lu_env *lep_env ____cacheline_aligned_in_smp;
2004 static struct lu_env_percpu lu_env_percpu[NR_CPUS];
2006 int lu_env_add_task(struct lu_env *env, struct task_struct *task)
2008 struct lu_env_item *lei, *old;
2016 lei->lei_task = task;
2019 old = rhashtable_lookup_get_insert_fast(&lu_env_rhash,
2021 lu_env_rhash_params);
2026 EXPORT_SYMBOL(lu_env_add_task);
2028 int lu_env_add(struct lu_env *env)
2030 return lu_env_add_task(env, current);
2032 EXPORT_SYMBOL(lu_env_add);
2034 static void lu_env_item_free(struct rcu_head *head)
2036 struct lu_env_item *lei;
2038 lei = container_of(head, struct lu_env_item, lei_rcu_head);
2042 void lu_env_remove(struct lu_env *env)
2044 struct lu_env_item *lei;
2045 const void *task = current;
2048 for_each_possible_cpu(i) {
2049 if (lu_env_percpu[i].lep_env == env) {
2050 LASSERT(lu_env_percpu[i].lep_task == task);
2051 lu_env_percpu[i].lep_task = NULL;
2052 lu_env_percpu[i].lep_env = NULL;
2056 /* The rcu_lock is not taking in this case since the key
2057 * used is the actual task_struct. This implies that each
2058 * object is only removed by the owning thread, so there
2059 * can never be a race on a particular object.
2061 lei = rhashtable_lookup_fast(&lu_env_rhash, &task,
2062 lu_env_rhash_params);
2063 if (lei && rhashtable_remove_fast(&lu_env_rhash, &lei->lei_linkage,
2064 lu_env_rhash_params) == 0)
2065 call_rcu(&lei->lei_rcu_head, lu_env_item_free);
2067 EXPORT_SYMBOL(lu_env_remove);
2069 struct lu_env *lu_env_find(void)
2071 struct lu_env *env = NULL;
2072 struct lu_env_item *lei;
2073 const void *task = current;
2076 if (lu_env_percpu[i].lep_task == current) {
2077 env = lu_env_percpu[i].lep_env;
2083 lei = rhashtable_lookup_fast(&lu_env_rhash, &task,
2084 lu_env_rhash_params);
2087 lu_env_percpu[i].lep_task = current;
2088 lu_env_percpu[i].lep_env = env;
2094 EXPORT_SYMBOL(lu_env_find);
2096 typedef struct lu_site_stats{
2097 unsigned lss_populated;
2098 unsigned lss_max_search;
2103 static void lu_site_stats_get(const struct lu_site *s,
2104 lu_site_stats_t *stats)
2106 int cnt = atomic_read(&s->ls_obj_hash.nelems);
2108 * percpu_counter_sum_positive() won't accept a const pointer
2109 * as it does modify the struct by taking a spinlock
2111 struct lu_site *s2 = (struct lu_site *)s;
2113 stats->lss_busy += cnt -
2114 percpu_counter_sum_positive(&s2->ls_lru_len_counter);
2116 stats->lss_total += cnt;
2117 stats->lss_max_search = 0;
2118 stats->lss_populated = 0;
2123 * lu_cache_shrink_count() returns an approximate number of cached objects
2124 * that can be freed by shrink_slab(). A counter, which tracks the
2125 * number of items in the site's lru, is maintained in a percpu_counter
2126 * for each site. The percpu values are incremented and decremented as
2127 * objects are added or removed from the lru. The percpu values are summed
2128 * and saved whenever a percpu value exceeds a threshold. Thus the saved,
2129 * summed value at any given time may not accurately reflect the current
2130 * lru length. But this value is sufficiently accurate for the needs of
2133 * Using a per cpu counter is a compromise solution to concurrent access:
2134 * lu_object_put() can update the counter without locking the site and
2135 * lu_cache_shrink_count can sum the counters without locking each
2136 * ls_obj_hash bucket.
2138 static unsigned long lu_cache_shrink_count(struct shrinker *sk,
2139 struct shrink_control *sc)
2142 struct lu_site *tmp;
2143 unsigned long cached = 0;
2145 if (!(sc->gfp_mask & __GFP_FS))
2148 down_read(&lu_sites_guard);
2149 list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage)
2150 cached += percpu_counter_read_positive(&s->ls_lru_len_counter);
2151 up_read(&lu_sites_guard);
2153 cached = (cached / 100) * sysctl_vfs_cache_pressure;
2154 CDEBUG(D_INODE, "%ld objects cached, cache pressure %d\n",
2155 cached, sysctl_vfs_cache_pressure);
2160 static unsigned long lu_cache_shrink_scan(struct shrinker *sk,
2161 struct shrink_control *sc)
2164 struct lu_site *tmp;
2165 unsigned long remain = sc->nr_to_scan;
2168 if (!(sc->gfp_mask & __GFP_FS))
2169 /* We must not take the lu_sites_guard lock when
2170 * __GFP_FS is *not* set because of the deadlock
2171 * possibility detailed above. Additionally,
2172 * since we cannot determine the number of
2173 * objects in the cache without taking this
2174 * lock, we're in a particularly tough spot. As
2175 * a result, we'll just lie and say our cache is
2176 * empty. This _should_ be ok, as we can't
2177 * reclaim objects when __GFP_FS is *not* set
2182 down_write(&lu_sites_guard);
2183 list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
2184 remain = lu_site_purge(&lu_shrink_env, s, remain);
2186 * Move just shrunk site to the tail of site list to
2187 * assure shrinking fairness.
2189 list_move_tail(&s->ls_linkage, &splice);
2191 list_splice(&splice, lu_sites.prev);
2192 up_write(&lu_sites_guard);
2194 return sc->nr_to_scan - remain;
2197 #ifdef HAVE_SHRINKER_COUNT
2198 static struct shrinker lu_site_shrinker = {
2199 .count_objects = lu_cache_shrink_count,
2200 .scan_objects = lu_cache_shrink_scan,
2201 .seeks = DEFAULT_SEEKS,
2206 * There exists a potential lock inversion deadlock scenario when using
2207 * Lustre on top of ZFS. This occurs between one of ZFS's
2208 * buf_hash_table.ht_lock's, and Lustre's lu_sites_guard lock. Essentially,
2209 * thread A will take the lu_sites_guard lock and sleep on the ht_lock,
2210 * while thread B will take the ht_lock and sleep on the lu_sites_guard
2211 * lock. Obviously neither thread will wake and drop their respective hold
2214 * To prevent this from happening we must ensure the lu_sites_guard lock is
2215 * not taken while down this code path. ZFS reliably does not set the
2216 * __GFP_FS bit in its code paths, so this can be used to determine if it
2217 * is safe to take the lu_sites_guard lock.
2219 * Ideally we should accurately return the remaining number of cached
2220 * objects without taking the lu_sites_guard lock, but this is not
2221 * possible in the current implementation.
2223 static int lu_cache_shrink(struct shrinker *shrinker,
2224 struct shrink_control *sc)
2228 CDEBUG(D_INODE, "Shrink %lu objects\n", sc->nr_to_scan);
2230 if (sc->nr_to_scan != 0)
2231 lu_cache_shrink_scan(shrinker, sc);
2233 cached = lu_cache_shrink_count(shrinker, sc);
2237 static struct shrinker lu_site_shrinker = {
2238 .shrink = lu_cache_shrink,
2239 .seeks = DEFAULT_SEEKS,
2242 #endif /* HAVE_SHRINKER_COUNT */
2250 * Environment to be used in debugger, contains all tags.
2252 static struct lu_env lu_debugging_env;
2255 * Debugging printer function using printk().
2257 int lu_printk_printer(const struct lu_env *env,
2258 void *unused, const char *format, ...)
2262 va_start(args, format);
2263 vprintk(format, args);
2268 int lu_debugging_setup(void)
2270 return lu_env_init(&lu_debugging_env, ~0);
2273 void lu_context_keys_dump(void)
2277 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
2278 struct lu_context_key *key;
2282 CERROR("LU context keys [%d]: %p %x (%p,%p,%p) %d %d \"%s\"@%p\n",
2283 i, key, key->lct_tags,
2284 key->lct_init, key->lct_fini, key->lct_exit,
2285 key->lct_index, atomic_read(&key->lct_used),
2286 key->lct_owner ? key->lct_owner->name : "",
2288 lu_ref_print(&key->lct_reference);
2294 * Initialization of global lu_* data.
2296 int lu_global_init(void)
2300 CDEBUG(D_INFO, "Lustre LU module (%p).\n", &lu_keys);
2302 result = lu_ref_global_init();
2306 LU_CONTEXT_KEY_INIT(&lu_global_key);
2307 result = lu_context_key_register(&lu_global_key);
2312 * At this level, we don't know what tags are needed, so allocate them
2313 * conservatively. This should not be too bad, because this
2314 * environment is global.
2316 down_write(&lu_sites_guard);
2317 result = lu_env_init(&lu_shrink_env, LCT_SHRINKER);
2318 up_write(&lu_sites_guard);
2320 lu_context_key_degister(&lu_global_key);
2325 * seeks estimation: 3 seeks to read a record from oi, one to read
2326 * inode, one for ea. Unfortunately setting this high value results in
2327 * lu_object/inode cache consuming all the memory.
2329 result = register_shrinker(&lu_site_shrinker);
2333 result = rhashtable_init(&lu_env_rhash, &lu_env_rhash_params);
2341 unregister_shrinker(&lu_site_shrinker);
2343 /* ordering here is explained in lu_global_fini() */
2344 lu_context_key_degister(&lu_global_key);
2345 down_write(&lu_sites_guard);
2346 lu_env_fini(&lu_shrink_env);
2347 up_write(&lu_sites_guard);
2349 lu_ref_global_fini();
2354 * Dual to lu_global_init().
2356 void lu_global_fini(void)
2358 unregister_shrinker(&lu_site_shrinker);
2360 lu_context_key_degister(&lu_global_key);
2363 * Tear shrinker environment down _after_ de-registering
2364 * lu_global_key, because the latter has a value in the former.
2366 down_write(&lu_sites_guard);
2367 lu_env_fini(&lu_shrink_env);
2368 up_write(&lu_sites_guard);
2370 rhashtable_destroy(&lu_env_rhash);
2372 lu_ref_global_fini();
2375 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx)
2377 #ifdef CONFIG_PROC_FS
2378 struct lprocfs_counter ret;
2380 lprocfs_stats_collect(stats, idx, &ret);
2381 return (__u32)ret.lc_count;
2388 * Output site statistical counters into a buffer. Suitable for
2389 * lprocfs_rd_*()-style functions.
2391 int lu_site_stats_seq_print(const struct lu_site *s, struct seq_file *m)
2393 const struct bucket_table *tbl;
2394 lu_site_stats_t stats;
2395 unsigned int chains;
2397 memset(&stats, 0, sizeof(stats));
2398 lu_site_stats_get(s, &stats);
2401 tbl = rht_dereference_rcu(s->ls_obj_hash.tbl,
2402 &((struct lu_site *)s)->ls_obj_hash);
2405 seq_printf(m, "%d/%d %d/%u %d %d %d %d %d %d %d\n",
2408 stats.lss_populated,
2410 stats.lss_max_search,
2411 ls_stats_read(s->ls_stats, LU_SS_CREATED),
2412 ls_stats_read(s->ls_stats, LU_SS_CACHE_HIT),
2413 ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
2414 ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
2415 ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
2416 ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED));
2419 EXPORT_SYMBOL(lu_site_stats_seq_print);
2422 * Helper function to initialize a number of kmem slab caches at once.
2424 int lu_kmem_init(struct lu_kmem_descr *caches)
2427 struct lu_kmem_descr *iter = caches;
2429 for (result = 0; iter->ckd_cache != NULL; ++iter) {
2430 *iter->ckd_cache = kmem_cache_create(iter->ckd_name,
2433 if (*iter->ckd_cache == NULL) {
2435 /* free all previously allocated caches */
2436 lu_kmem_fini(caches);
2442 EXPORT_SYMBOL(lu_kmem_init);
2445 * Helper function to finalize a number of kmem slab cached at once. Dual to
2448 void lu_kmem_fini(struct lu_kmem_descr *caches)
2450 for (; caches->ckd_cache != NULL; ++caches) {
2451 if (*caches->ckd_cache != NULL) {
2452 kmem_cache_destroy(*caches->ckd_cache);
2453 *caches->ckd_cache = NULL;
2457 EXPORT_SYMBOL(lu_kmem_fini);
2460 * Temporary solution to be able to assign fid in ->do_create()
2461 * till we have fully-functional OST fids
2463 void lu_object_assign_fid(const struct lu_env *env, struct lu_object *o,
2464 const struct lu_fid *fid)
2466 struct lu_site *s = o->lo_dev->ld_site;
2467 struct lu_fid *old = &o->lo_header->loh_fid;
2470 LASSERT(fid_is_zero(old));
2473 rc = rhashtable_lookup_insert_fast(&s->ls_obj_hash,
2474 &o->lo_header->loh_hash,
2476 /* supposed to be unique */
2477 LASSERT(rc != -EEXIST);
2478 /* handle hash table resizing */
2479 if (rc == -ENOMEM || rc == -EBUSY) {
2483 /* trim the hash if its growing to big */
2484 lu_object_limit(env, o->lo_dev);
2488 LASSERTF(rc == 0, "failed hashtable insertion: rc = %d\n", rc);
2490 EXPORT_SYMBOL(lu_object_assign_fid);
2493 * allocates object with 0 (non-assiged) fid
2494 * XXX: temporary solution to be able to assign fid in ->do_create()
2495 * till we have fully-functional OST fids
2497 struct lu_object *lu_object_anon(const struct lu_env *env,
2498 struct lu_device *dev,
2499 const struct lu_object_conf *conf)
2502 struct lu_object *o;
2506 o = lu_object_alloc(env, dev, &fid);
2508 rc = lu_object_start(env, dev, o, conf);
2510 lu_object_free(env, o);
2517 EXPORT_SYMBOL(lu_object_anon);
2519 struct lu_buf LU_BUF_NULL = {
2523 EXPORT_SYMBOL(LU_BUF_NULL);
2525 void lu_buf_free(struct lu_buf *buf)
2529 LASSERT(buf->lb_len > 0);
2530 OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);
2535 EXPORT_SYMBOL(lu_buf_free);
2537 void lu_buf_alloc(struct lu_buf *buf, size_t size)
2540 LASSERT(buf->lb_buf == NULL);
2541 LASSERT(buf->lb_len == 0);
2542 OBD_ALLOC_LARGE(buf->lb_buf, size);
2543 if (likely(buf->lb_buf))
2546 EXPORT_SYMBOL(lu_buf_alloc);
2548 void lu_buf_realloc(struct lu_buf *buf, size_t size)
2551 lu_buf_alloc(buf, size);
2553 EXPORT_SYMBOL(lu_buf_realloc);
2555 struct lu_buf *lu_buf_check_and_alloc(struct lu_buf *buf, size_t len)
2557 if (buf->lb_buf == NULL && buf->lb_len == 0)
2558 lu_buf_alloc(buf, len);
2560 if ((len > buf->lb_len) && (buf->lb_buf != NULL))
2561 lu_buf_realloc(buf, len);
2565 EXPORT_SYMBOL(lu_buf_check_and_alloc);
2568 * Increase the size of the \a buf.
2569 * preserves old data in buffer
2570 * old buffer remains unchanged on error
2571 * \retval 0 or -ENOMEM
2573 int lu_buf_check_and_grow(struct lu_buf *buf, size_t len)
2577 if (len <= buf->lb_len)
2580 OBD_ALLOC_LARGE(ptr, len);
2584 /* Free the old buf */
2585 if (buf->lb_buf != NULL) {
2586 memcpy(ptr, buf->lb_buf, buf->lb_len);
2587 OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);
2594 EXPORT_SYMBOL(lu_buf_check_and_grow);