4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/obdclass/lu_object.c
35 * These are the only exported functions, they provide some generic
36 * infrastructure for managing object devices
38 * Author: Nikita Danilov <nikita.danilov@sun.com>
41 #define DEBUG_SUBSYSTEM S_CLASS
43 #include <linux/module.h>
44 #include <linux/list.h>
45 #ifdef HAVE_PROCESSOR_H
46 #include <linux/processor.h>
48 #include <libcfs/linux/processor.h>
50 #include <linux/random.h>
52 #include <libcfs/libcfs.h>
53 #include <libcfs/libcfs_hash.h> /* hash_long() */
54 #include <libcfs/linux/linux-mem.h>
55 #include <obd_class.h>
56 #include <obd_support.h>
57 #include <lustre_disk.h>
58 #include <lustre_fid.h>
59 #include <lu_object.h>
62 struct lu_site_bkt_data {
64 * LRU list, updated on each access to object. Protected by
67 * "Cold" end of LRU is lu_site::ls_lru.next. Accessed object are
68 * moved to the lu_site::ls_lru.prev
70 struct list_head lsb_lru;
72 * Wait-queue signaled when an object in this site is ultimately
73 * destroyed (lu_object_free()) or initialized (lu_object_start()).
74 * It is used by lu_object_find() to wait before re-trying when
75 * object in the process of destruction is found in the hash table;
76 * or wait object to be initialized by the allocator.
78 * \see htable_lookup().
80 wait_queue_head_t lsb_waitq;
84 LU_CACHE_PERCENT_MAX = 50,
85 LU_CACHE_PERCENT_DEFAULT = 20
88 #define LU_CACHE_NR_MAX_ADJUST 512
89 #define LU_CACHE_NR_UNLIMITED -1
90 #define LU_CACHE_NR_DEFAULT LU_CACHE_NR_UNLIMITED
91 #define LU_CACHE_NR_LDISKFS_LIMIT LU_CACHE_NR_UNLIMITED
92 /** This is set to roughly (20 * OSS_NTHRS_MAX) to prevent thrashing */
93 #define LU_CACHE_NR_ZFS_LIMIT 10240
95 #define LU_SITE_BITS_MIN 12
96 #define LU_SITE_BITS_MAX 24
97 #define LU_SITE_BITS_MAX_CL 19
99 * Max 256 buckets, we don't want too many buckets because:
100 * - consume too much memory (currently max 16K)
101 * - avoid unbalanced LRU list
102 * With few cpus there is little gain from extra buckets, so
103 * we treat this as a maximum in lu_site_init().
105 #define LU_SITE_BKT_BITS 8
108 static unsigned int lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
109 module_param(lu_cache_percent, int, 0644);
110 MODULE_PARM_DESC(lu_cache_percent, "Percentage of memory to be used as lu_object cache");
112 static long lu_cache_nr = LU_CACHE_NR_DEFAULT;
113 module_param(lu_cache_nr, long, 0644);
114 MODULE_PARM_DESC(lu_cache_nr, "Maximum number of objects in lu_object cache");
116 static void lu_object_free(const struct lu_env *env, struct lu_object *o);
117 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx);
119 static u32 lu_fid_hash(const void *data, u32 seed)
121 const struct lu_fid *fid = data;
123 seed = cfs_hash_32(seed ^ fid->f_oid, 32);
124 seed ^= cfs_hash_64(fid->f_seq, 32);
128 static inline int lu_bkt_hash(struct lu_site *s, const struct lu_fid *fid)
130 return lu_fid_hash(fid, s->ls_bkt_seed) &
135 lu_site_wq_from_fid(struct lu_site *site, struct lu_fid *fid)
137 struct lu_site_bkt_data *bkt;
139 bkt = &site->ls_bkts[lu_bkt_hash(site, fid)];
140 return &bkt->lsb_waitq;
142 EXPORT_SYMBOL(lu_site_wq_from_fid);
145 * Decrease reference counter on object. If last reference is freed, return
146 * object to the cache, unless lu_object_is_dying(o) holds. In the latter
147 * case, free object immediately.
149 void lu_object_put(const struct lu_env *env, struct lu_object *o)
151 struct lu_site_bkt_data *bkt;
152 struct lu_object_header *top = o->lo_header;
153 struct lu_site *site = o->lo_dev->ld_site;
154 struct lu_object *orig = o;
155 struct cfs_hash_bd bd;
156 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(top->loh_hash.next == NULL
166 && top->loh_hash.pprev == NULL);
167 LASSERT(list_empty(&top->loh_lru));
168 if (!atomic_dec_and_test(&top->loh_ref))
170 list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
171 if (o->lo_ops->loo_object_release != NULL)
172 o->lo_ops->loo_object_release(env, o);
174 lu_object_free(env, orig);
178 cfs_hash_bd_get(site->ls_obj_hash, &top->loh_fid, &bd);
180 is_dying = lu_object_is_dying(top);
181 if (!cfs_hash_bd_dec_and_lock(site->ls_obj_hash, &bd, &top->loh_ref)) {
182 /* at this point the object reference is dropped and lock is
183 * not taken, so lu_object should not be touched because it
184 * can be freed by concurrent thread. Use local variable for
189 * somebody may be waiting for this, currently only
190 * used for cl_object, see cl_object_put_last().
192 bkt = &site->ls_bkts[lu_bkt_hash(site, &top->loh_fid)];
193 wake_up_all(&bkt->lsb_waitq);
199 * When last reference is released, iterate over object
200 * layers, and notify them that object is no longer busy.
202 list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
203 if (o->lo_ops->loo_object_release != NULL)
204 o->lo_ops->loo_object_release(env, o);
207 bkt = &site->ls_bkts[lu_bkt_hash(site, &top->loh_fid)];
208 spin_lock(&bkt->lsb_waitq.lock);
210 /* don't use local 'is_dying' here because if was taken without lock
211 * but here we need the latest actual value of it so check lu_object
214 if (!lu_object_is_dying(top) &&
215 (lu_object_exists(orig) || lu_object_is_cl(orig))) {
216 LASSERT(list_empty(&top->loh_lru));
217 list_add_tail(&top->loh_lru, &bkt->lsb_lru);
218 spin_unlock(&bkt->lsb_waitq.lock);
219 percpu_counter_inc(&site->ls_lru_len_counter);
220 CDEBUG(D_INODE, "Add %p/%p to site lru. hash: %p, bkt: %p\n",
221 orig, top, site->ls_obj_hash, bkt);
222 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
227 * If object is dying (will not be cached) then remove it
228 * from hash table (it is already not on the LRU).
230 * This is done with hash table lists locked. As the only
231 * way to acquire first reference to previously unreferenced
232 * object is through hash-table lookup (lu_object_find())
233 * which is done under hash-table, no race with concurrent
234 * object lookup is possible and we can safely destroy object below.
236 if (!test_and_set_bit(LU_OBJECT_UNHASHED, &top->loh_flags))
237 cfs_hash_bd_del_locked(site->ls_obj_hash, &bd, &top->loh_hash);
238 spin_unlock(&bkt->lsb_waitq.lock);
239 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
240 /* Object was already removed from hash above, can kill it. */
241 lu_object_free(env, orig);
243 EXPORT_SYMBOL(lu_object_put);
246 * Put object and don't keep in cache. This is temporary solution for
247 * multi-site objects when its layering is not constant.
249 void lu_object_put_nocache(const struct lu_env *env, struct lu_object *o)
251 set_bit(LU_OBJECT_HEARD_BANSHEE, &o->lo_header->loh_flags);
252 return lu_object_put(env, o);
254 EXPORT_SYMBOL(lu_object_put_nocache);
257 * Kill the object and take it out of LRU cache.
258 * Currently used by client code for layout change.
260 void lu_object_unhash(const struct lu_env *env, struct lu_object *o)
262 struct lu_object_header *top;
265 set_bit(LU_OBJECT_HEARD_BANSHEE, &top->loh_flags);
266 if (!test_and_set_bit(LU_OBJECT_UNHASHED, &top->loh_flags)) {
267 struct lu_site *site = o->lo_dev->ld_site;
268 struct cfs_hash *obj_hash = site->ls_obj_hash;
269 struct cfs_hash_bd bd;
271 cfs_hash_bd_get_and_lock(obj_hash, &top->loh_fid, &bd, 1);
272 if (!list_empty(&top->loh_lru)) {
273 struct lu_site_bkt_data *bkt;
275 bkt = &site->ls_bkts[lu_bkt_hash(site, &top->loh_fid)];
276 spin_lock(&bkt->lsb_waitq.lock);
277 list_del_init(&top->loh_lru);
278 spin_unlock(&bkt->lsb_waitq.lock);
279 percpu_counter_dec(&site->ls_lru_len_counter);
281 cfs_hash_bd_del_locked(obj_hash, &bd, &top->loh_hash);
282 cfs_hash_bd_unlock(obj_hash, &bd, 1);
285 EXPORT_SYMBOL(lu_object_unhash);
288 * Allocate new object.
290 * This follows object creation protocol, described in the comment within
291 * struct lu_device_operations definition.
293 static struct lu_object *lu_object_alloc(const struct lu_env *env,
294 struct lu_device *dev,
295 const struct lu_fid *f)
297 struct lu_object *top;
300 * Create top-level object slice. This will also create
303 top = dev->ld_ops->ldo_object_alloc(env, NULL, dev);
305 return ERR_PTR(-ENOMEM);
309 * This is the only place where object fid is assigned. It's constant
312 top->lo_header->loh_fid = *f;
320 * This is called after object hash insertion to avoid returning an object with
323 static int lu_object_start(const struct lu_env *env, struct lu_device *dev,
324 struct lu_object *top,
325 const struct lu_object_conf *conf)
327 struct lu_object *scan;
328 struct list_head *layers;
329 unsigned int init_mask = 0;
330 unsigned int init_flag;
334 layers = &top->lo_header->loh_layers;
338 * Call ->loo_object_init() repeatedly, until no more new
339 * object slices are created.
343 list_for_each_entry(scan, layers, lo_linkage) {
344 if (init_mask & init_flag)
347 scan->lo_header = top->lo_header;
348 result = scan->lo_ops->loo_object_init(env, scan, conf);
352 init_mask |= init_flag;
358 list_for_each_entry_reverse(scan, layers, lo_linkage) {
359 if (scan->lo_ops->loo_object_start != NULL) {
360 result = scan->lo_ops->loo_object_start(env, scan);
366 lprocfs_counter_incr(dev->ld_site->ls_stats, LU_SS_CREATED);
368 set_bit(LU_OBJECT_INITED, &top->lo_header->loh_flags);
376 static void lu_object_free(const struct lu_env *env, struct lu_object *o)
378 wait_queue_head_t *wq;
379 struct lu_site *site;
380 struct lu_object *scan;
381 struct list_head *layers;
384 site = o->lo_dev->ld_site;
385 layers = &o->lo_header->loh_layers;
386 wq = lu_site_wq_from_fid(site, &o->lo_header->loh_fid);
388 * First call ->loo_object_delete() method to release all resources.
390 list_for_each_entry_reverse(scan, layers, lo_linkage) {
391 if (scan->lo_ops->loo_object_delete != NULL)
392 scan->lo_ops->loo_object_delete(env, scan);
396 * Then, splice object layers into stand-alone list, and call
397 * ->loo_object_free() on all layers to free memory. Splice is
398 * necessary, because lu_object_header is freed together with the
401 list_splice_init(layers, &splice);
402 while (!list_empty(&splice)) {
404 * Free layers in bottom-to-top order, so that object header
405 * lives as long as possible and ->loo_object_free() methods
406 * can look at its contents.
408 o = container_of0(splice.prev, struct lu_object, lo_linkage);
409 list_del_init(&o->lo_linkage);
410 LASSERT(o->lo_ops->loo_object_free != NULL);
411 o->lo_ops->loo_object_free(env, o);
414 if (waitqueue_active(wq))
419 * Free \a nr objects from the cold end of the site LRU list.
420 * if canblock is 0, then don't block awaiting for another
421 * instance of lu_site_purge() to complete
423 int lu_site_purge_objects(const struct lu_env *env, struct lu_site *s,
424 int nr, int canblock)
426 struct lu_object_header *h;
427 struct lu_object_header *temp;
428 struct lu_site_bkt_data *bkt;
431 unsigned int start = 0;
436 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_NO_LRU))
440 * Under LRU list lock, scan LRU list and move unreferenced objects to
441 * the dispose list, removing them from LRU and hash table.
444 start = s->ls_purge_start;
445 bnr = (nr == ~0) ? -1 : nr / s->ls_bkt_cnt + 1;
448 * It doesn't make any sense to make purge threads parallel, that can
449 * only bring troubles to us. See LU-5331.
452 mutex_lock(&s->ls_purge_mutex);
453 else if (mutex_trylock(&s->ls_purge_mutex) == 0)
457 for (i = start; i < s->ls_bkt_cnt ; i++) {
459 bkt = &s->ls_bkts[i];
460 spin_lock(&bkt->lsb_waitq.lock);
462 list_for_each_entry_safe(h, temp, &bkt->lsb_lru, loh_lru) {
463 LASSERT(atomic_read(&h->loh_ref) == 0);
465 LINVRNT(lu_bkt_hash(s, &h->loh_fid) == i);
467 /* Cannot remove from hash under current spinlock,
468 * so set flag to stop object from being found
469 * by htable_lookup().
471 set_bit(LU_OBJECT_PURGING, &h->loh_flags);
472 list_move(&h->loh_lru, &dispose);
473 percpu_counter_dec(&s->ls_lru_len_counter);
477 if (nr != ~0 && --nr == 0)
480 if (count > 0 && --count == 0)
484 spin_unlock(&bkt->lsb_waitq.lock);
487 * Free everything on the dispose list. This is safe against
488 * races due to the reasons described in lu_object_put().
490 while ((h = list_first_entry_or_null(&dispose,
491 struct lu_object_header,
493 cfs_hash_del(s->ls_obj_hash, &h->loh_fid, &h->loh_hash);
494 list_del_init(&h->loh_lru);
495 lu_object_free(env, lu_object_top(h));
496 lprocfs_counter_incr(s->ls_stats, LU_SS_LRU_PURGED);
502 mutex_unlock(&s->ls_purge_mutex);
504 if (nr != 0 && did_sth && start != 0) {
505 start = 0; /* restart from the first bucket */
508 /* race on s->ls_purge_start, but nobody cares */
509 s->ls_purge_start = i & (s->ls_bkt_cnt - 1);
513 EXPORT_SYMBOL(lu_site_purge_objects);
518 * Code below has to jump through certain loops to output object description
519 * into libcfs_debug_msg-based log. The problem is that lu_object_print()
520 * composes object description from strings that are parts of _lines_ of
521 * output (i.e., strings that are not terminated by newline). This doesn't fit
522 * very well into libcfs_debug_msg() interface that assumes that each message
523 * supplied to it is a self-contained output line.
525 * To work around this, strings are collected in a temporary buffer
526 * (implemented as a value of lu_cdebug_key key), until terminating newline
527 * character is detected.
535 * XXX overflow is not handled correctly.
540 struct lu_cdebug_data {
544 char lck_area[LU_CDEBUG_LINE];
547 /* context key constructor/destructor: lu_global_key_init, lu_global_key_fini */
548 LU_KEY_INIT_FINI(lu_global, struct lu_cdebug_data);
551 * Key, holding temporary buffer. This key is registered very early by
554 static struct lu_context_key lu_global_key = {
555 .lct_tags = LCT_MD_THREAD | LCT_DT_THREAD |
556 LCT_MG_THREAD | LCT_CL_THREAD | LCT_LOCAL,
557 .lct_init = lu_global_key_init,
558 .lct_fini = lu_global_key_fini
562 * Printer function emitting messages through libcfs_debug_msg().
564 int lu_cdebug_printer(const struct lu_env *env,
565 void *cookie, const char *format, ...)
567 struct libcfs_debug_msg_data *msgdata = cookie;
568 struct lu_cdebug_data *key;
573 va_start(args, format);
575 key = lu_context_key_get(&env->le_ctx, &lu_global_key);
576 LASSERT(key != NULL);
578 used = strlen(key->lck_area);
579 complete = format[strlen(format) - 1] == '\n';
581 * Append new chunk to the buffer.
583 vsnprintf(key->lck_area + used,
584 ARRAY_SIZE(key->lck_area) - used, format, args);
586 if (cfs_cdebug_show(msgdata->msg_mask, msgdata->msg_subsys))
587 libcfs_debug_msg(msgdata, "%s\n", key->lck_area);
588 key->lck_area[0] = 0;
593 EXPORT_SYMBOL(lu_cdebug_printer);
596 * Print object header.
598 void lu_object_header_print(const struct lu_env *env, void *cookie,
599 lu_printer_t printer,
600 const struct lu_object_header *hdr)
602 (*printer)(env, cookie, "header@%p[%#lx, %d, "DFID"%s%s%s]",
603 hdr, hdr->loh_flags, atomic_read(&hdr->loh_ref),
605 hlist_unhashed(&hdr->loh_hash) ? "" : " hash",
606 list_empty((struct list_head *)&hdr->loh_lru) ? \
608 hdr->loh_attr & LOHA_EXISTS ? " exist" : "");
610 EXPORT_SYMBOL(lu_object_header_print);
613 * Print human readable representation of the \a o to the \a printer.
615 void lu_object_print(const struct lu_env *env, void *cookie,
616 lu_printer_t printer, const struct lu_object *o)
618 static const char ruler[] = "........................................";
619 struct lu_object_header *top;
623 lu_object_header_print(env, cookie, printer, top);
624 (*printer)(env, cookie, "{\n");
626 list_for_each_entry(o, &top->loh_layers, lo_linkage) {
628 * print `.' \a depth times followed by type name and address
630 (*printer)(env, cookie, "%*.*s%s@%p", depth, depth, ruler,
631 o->lo_dev->ld_type->ldt_name, o);
633 if (o->lo_ops->loo_object_print != NULL)
634 (*o->lo_ops->loo_object_print)(env, cookie, printer, o);
636 (*printer)(env, cookie, "\n");
639 (*printer)(env, cookie, "} header@%p\n", top);
641 EXPORT_SYMBOL(lu_object_print);
644 * Check object consistency.
646 int lu_object_invariant(const struct lu_object *o)
648 struct lu_object_header *top;
651 list_for_each_entry(o, &top->loh_layers, lo_linkage) {
652 if (o->lo_ops->loo_object_invariant != NULL &&
653 !o->lo_ops->loo_object_invariant(o))
659 static struct lu_object *htable_lookup(struct lu_site *s,
660 struct cfs_hash_bd *bd,
661 const struct lu_fid *f,
664 struct lu_object_header *h;
665 struct hlist_node *hnode;
666 __u64 ver = cfs_hash_bd_version_get(bd);
669 return ERR_PTR(-ENOENT);
672 /* cfs_hash_bd_peek_locked is a somehow "internal" function
673 * of cfs_hash, it doesn't add refcount on object. */
674 hnode = cfs_hash_bd_peek_locked(s->ls_obj_hash, bd, (void *)f);
676 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS);
677 return ERR_PTR(-ENOENT);
680 h = container_of0(hnode, struct lu_object_header, loh_hash);
681 if (!list_empty(&h->loh_lru)) {
682 struct lu_site_bkt_data *bkt;
684 bkt = &s->ls_bkts[lu_bkt_hash(s, &h->loh_fid)];
685 spin_lock(&bkt->lsb_waitq.lock);
686 /* Might have just been moved to the dispose list, in which
687 * case LU_OBJECT_PURGING will be set. In that case,
688 * delete it from the hash table immediately.
689 * When lu_site_purge_objects() tried, it will find it
690 * isn't there, which is harmless.
692 if (test_bit(LU_OBJECT_PURGING, &h->loh_flags)) {
693 spin_unlock(&bkt->lsb_waitq.lock);
694 cfs_hash_bd_del_locked(s->ls_obj_hash, bd, hnode);
695 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS);
696 return ERR_PTR(-ENOENT);
698 list_del_init(&h->loh_lru);
699 spin_unlock(&bkt->lsb_waitq.lock);
700 percpu_counter_dec(&s->ls_lru_len_counter);
702 cfs_hash_get(s->ls_obj_hash, hnode);
703 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_HIT);
704 return lu_object_top(h);
708 * Search cache for an object with the fid \a f. If such object is found,
709 * return it. Otherwise, create new object, insert it into cache and return
710 * it. In any case, additional reference is acquired on the returned object.
712 struct lu_object *lu_object_find(const struct lu_env *env,
713 struct lu_device *dev, const struct lu_fid *f,
714 const struct lu_object_conf *conf)
716 return lu_object_find_at(env, dev->ld_site->ls_top_dev, f, conf);
718 EXPORT_SYMBOL(lu_object_find);
721 * Limit the lu_object cache to a maximum of lu_cache_nr objects. Because
722 * the calculation for the number of objects to reclaim is not covered by
723 * a lock the maximum number of objects is capped by LU_CACHE_MAX_ADJUST.
724 * This ensures that many concurrent threads will not accidentally purge
727 static void lu_object_limit(const struct lu_env *env,
728 struct lu_device *dev)
732 if (lu_cache_nr == LU_CACHE_NR_UNLIMITED)
735 size = cfs_hash_size_get(dev->ld_site->ls_obj_hash);
736 nr = (__u64)lu_cache_nr;
740 lu_site_purge_objects(env, dev->ld_site,
741 MIN(size - nr, LU_CACHE_NR_MAX_ADJUST), 0);
745 * Core logic of lu_object_find*() functions.
747 * Much like lu_object_find(), but top level device of object is specifically
748 * \a dev rather than top level device of the site. This interface allows
749 * objects of different "stacking" to be created within the same site.
751 struct lu_object *lu_object_find_at(const struct lu_env *env,
752 struct lu_device *dev,
753 const struct lu_fid *f,
754 const struct lu_object_conf *conf)
757 struct lu_object *shadow;
760 struct cfs_hash_bd bd;
761 struct lu_site_bkt_data *bkt;
768 * This uses standard index maintenance protocol:
770 * - search index under lock, and return object if found;
771 * - otherwise, unlock index, allocate new object;
772 * - lock index and search again;
773 * - if nothing is found (usual case), insert newly created
775 * - otherwise (race: other thread inserted object), free
776 * object just allocated.
780 * For "LOC_F_NEW" case, we are sure the object is new established.
781 * It is unnecessary to perform lookup-alloc-lookup-insert, instead,
782 * just alloc and insert directly.
788 if (unlikely(OBD_FAIL_PRECHECK(OBD_FAIL_OBD_ZERO_NLINK_RACE)))
789 lu_site_purge(env, s, -1);
791 bkt = &s->ls_bkts[lu_bkt_hash(s, f)];
792 cfs_hash_bd_get(hs, f, &bd);
793 if (!(conf && conf->loc_flags & LOC_F_NEW)) {
794 cfs_hash_bd_lock(hs, &bd, 1);
795 o = htable_lookup(s, &bd, f, &version);
796 cfs_hash_bd_unlock(hs, &bd, 1);
799 if (likely(lu_object_is_inited(o->lo_header)))
802 wait_event_idle(bkt->lsb_waitq,
803 lu_object_is_inited(o->lo_header) ||
804 lu_object_is_dying(o->lo_header));
806 if (lu_object_is_dying(o->lo_header)) {
807 lu_object_put(env, o);
809 RETURN(ERR_PTR(-ENOENT));
815 if (PTR_ERR(o) != -ENOENT)
820 * Allocate new object, NB, object is unitialized in case object
821 * is changed between allocation and hash insertion, thus the object
822 * with stale attributes is returned.
824 o = lu_object_alloc(env, dev, f);
828 LASSERT(lu_fid_eq(lu_object_fid(o), f));
830 CFS_RACE_WAIT(OBD_FAIL_OBD_ZERO_NLINK_RACE);
832 cfs_hash_bd_lock(hs, &bd, 1);
834 if (conf && conf->loc_flags & LOC_F_NEW)
835 shadow = ERR_PTR(-ENOENT);
837 shadow = htable_lookup(s, &bd, f, &version);
838 if (likely(PTR_ERR(shadow) == -ENOENT)) {
839 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
840 cfs_hash_bd_unlock(hs, &bd, 1);
843 * This may result in rather complicated operations, including
844 * fld queries, inode loading, etc.
846 rc = lu_object_start(env, dev, o, conf);
848 lu_object_put_nocache(env, o);
852 wake_up_all(&bkt->lsb_waitq);
854 lu_object_limit(env, dev);
859 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_RACE);
860 cfs_hash_bd_unlock(hs, &bd, 1);
861 lu_object_free(env, o);
863 if (!(conf && conf->loc_flags & LOC_F_NEW) &&
864 !lu_object_is_inited(shadow->lo_header)) {
865 wait_event_idle(bkt->lsb_waitq,
866 lu_object_is_inited(shadow->lo_header) ||
867 lu_object_is_dying(shadow->lo_header));
869 if (lu_object_is_dying(shadow->lo_header)) {
870 lu_object_put(env, shadow);
872 RETURN(ERR_PTR(-ENOENT));
878 EXPORT_SYMBOL(lu_object_find_at);
881 * Find object with given fid, and return its slice belonging to given device.
883 struct lu_object *lu_object_find_slice(const struct lu_env *env,
884 struct lu_device *dev,
885 const struct lu_fid *f,
886 const struct lu_object_conf *conf)
888 struct lu_object *top;
889 struct lu_object *obj;
891 top = lu_object_find(env, dev, f, conf);
895 obj = lu_object_locate(top->lo_header, dev->ld_type);
896 if (unlikely(obj == NULL)) {
897 lu_object_put(env, top);
898 obj = ERR_PTR(-ENOENT);
903 EXPORT_SYMBOL(lu_object_find_slice);
905 int lu_device_type_init(struct lu_device_type *ldt)
909 atomic_set(&ldt->ldt_device_nr, 0);
910 if (ldt->ldt_ops->ldto_init)
911 result = ldt->ldt_ops->ldto_init(ldt);
915 EXPORT_SYMBOL(lu_device_type_init);
917 void lu_device_type_fini(struct lu_device_type *ldt)
919 if (ldt->ldt_ops->ldto_fini)
920 ldt->ldt_ops->ldto_fini(ldt);
922 EXPORT_SYMBOL(lu_device_type_fini);
925 * Global list of all sites on this node
927 static LIST_HEAD(lu_sites);
928 static DECLARE_RWSEM(lu_sites_guard);
931 * Global environment used by site shrinker.
933 static struct lu_env lu_shrink_env;
935 struct lu_site_print_arg {
936 struct lu_env *lsp_env;
938 lu_printer_t lsp_printer;
942 lu_site_obj_print(struct cfs_hash *hs, struct cfs_hash_bd *bd,
943 struct hlist_node *hnode, void *data)
945 struct lu_site_print_arg *arg = (struct lu_site_print_arg *)data;
946 struct lu_object_header *h;
948 h = hlist_entry(hnode, struct lu_object_header, loh_hash);
949 if (!list_empty(&h->loh_layers)) {
950 const struct lu_object *o;
952 o = lu_object_top(h);
953 lu_object_print(arg->lsp_env, arg->lsp_cookie,
954 arg->lsp_printer, o);
956 lu_object_header_print(arg->lsp_env, arg->lsp_cookie,
957 arg->lsp_printer, h);
963 * Print all objects in \a s.
965 void lu_site_print(const struct lu_env *env, struct lu_site *s, void *cookie,
966 lu_printer_t printer)
968 struct lu_site_print_arg arg = {
969 .lsp_env = (struct lu_env *)env,
970 .lsp_cookie = cookie,
971 .lsp_printer = printer,
974 cfs_hash_for_each(s->ls_obj_hash, lu_site_obj_print, &arg);
976 EXPORT_SYMBOL(lu_site_print);
979 * Return desired hash table order.
981 static unsigned long lu_htable_order(struct lu_device *top)
983 unsigned long cache_size;
985 unsigned long bits_max = LU_SITE_BITS_MAX;
988 * For ZFS based OSDs the cache should be disabled by default. This
989 * allows the ZFS ARC maximum flexibility in determining what buffers
990 * to cache. If Lustre has objects or buffer which it wants to ensure
991 * always stay cached it must maintain a hold on them.
993 if (strcmp(top->ld_type->ldt_name, LUSTRE_OSD_ZFS_NAME) == 0) {
994 lu_cache_percent = 1;
995 lu_cache_nr = LU_CACHE_NR_ZFS_LIMIT;
996 return LU_SITE_BITS_MIN;
999 if (strcmp(top->ld_type->ldt_name, LUSTRE_VVP_NAME) == 0)
1000 bits_max = LU_SITE_BITS_MAX_CL;
1003 * Calculate hash table size, assuming that we want reasonable
1004 * performance when 20% of total memory is occupied by cache of
1007 * Size of lu_object is (arbitrary) taken as 1K (together with inode).
1009 cache_size = cfs_totalram_pages();
1011 #if BITS_PER_LONG == 32
1012 /* limit hashtable size for lowmem systems to low RAM */
1013 if (cache_size > 1 << (30 - PAGE_SHIFT))
1014 cache_size = 1 << (30 - PAGE_SHIFT) * 3 / 4;
1017 /* clear off unreasonable cache setting. */
1018 if (lu_cache_percent == 0 || lu_cache_percent > LU_CACHE_PERCENT_MAX) {
1019 CWARN("obdclass: invalid lu_cache_percent: %u, it must be in"
1020 " the range of (0, %u]. Will use default value: %u.\n",
1021 lu_cache_percent, LU_CACHE_PERCENT_MAX,
1022 LU_CACHE_PERCENT_DEFAULT);
1024 lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
1026 cache_size = cache_size / 100 * lu_cache_percent *
1029 for (bits = 1; (1 << bits) < cache_size; ++bits) {
1033 return clamp_t(typeof(bits), bits, LU_SITE_BITS_MIN, bits_max);
1036 static unsigned lu_obj_hop_hash(struct cfs_hash *hs,
1037 const void *key, unsigned mask)
1039 struct lu_fid *fid = (struct lu_fid *)key;
1042 hash = fid_flatten32(fid);
1043 hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
1044 hash = hash_long(hash, hs->hs_bkt_bits);
1046 /* give me another random factor */
1047 hash -= hash_long((unsigned long)hs, fid_oid(fid) % 11 + 3);
1049 hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
1050 hash |= (fid_seq(fid) + fid_oid(fid)) & (CFS_HASH_NBKT(hs) - 1);
1055 static void *lu_obj_hop_object(struct hlist_node *hnode)
1057 return hlist_entry(hnode, struct lu_object_header, loh_hash);
1060 static void *lu_obj_hop_key(struct hlist_node *hnode)
1062 struct lu_object_header *h;
1064 h = hlist_entry(hnode, struct lu_object_header, loh_hash);
1068 static int lu_obj_hop_keycmp(const void *key, struct hlist_node *hnode)
1070 struct lu_object_header *h;
1072 h = hlist_entry(hnode, struct lu_object_header, loh_hash);
1073 return lu_fid_eq(&h->loh_fid, (struct lu_fid *)key);
1076 static void lu_obj_hop_get(struct cfs_hash *hs, struct hlist_node *hnode)
1078 struct lu_object_header *h;
1080 h = hlist_entry(hnode, struct lu_object_header, loh_hash);
1081 atomic_inc(&h->loh_ref);
1084 static void lu_obj_hop_put_locked(struct cfs_hash *hs, struct hlist_node *hnode)
1086 LBUG(); /* we should never called it */
1089 static struct cfs_hash_ops lu_site_hash_ops = {
1090 .hs_hash = lu_obj_hop_hash,
1091 .hs_key = lu_obj_hop_key,
1092 .hs_keycmp = lu_obj_hop_keycmp,
1093 .hs_object = lu_obj_hop_object,
1094 .hs_get = lu_obj_hop_get,
1095 .hs_put_locked = lu_obj_hop_put_locked,
1098 void lu_dev_add_linkage(struct lu_site *s, struct lu_device *d)
1100 spin_lock(&s->ls_ld_lock);
1101 if (list_empty(&d->ld_linkage))
1102 list_add(&d->ld_linkage, &s->ls_ld_linkage);
1103 spin_unlock(&s->ls_ld_lock);
1105 EXPORT_SYMBOL(lu_dev_add_linkage);
1107 void lu_dev_del_linkage(struct lu_site *s, struct lu_device *d)
1109 spin_lock(&s->ls_ld_lock);
1110 list_del_init(&d->ld_linkage);
1111 spin_unlock(&s->ls_ld_lock);
1113 EXPORT_SYMBOL(lu_dev_del_linkage);
1116 * Initialize site \a s, with \a d as the top level device.
1118 int lu_site_init(struct lu_site *s, struct lu_device *top)
1120 struct lu_site_bkt_data *bkt;
1127 memset(s, 0, sizeof *s);
1128 mutex_init(&s->ls_purge_mutex);
1130 #ifdef HAVE_PERCPU_COUNTER_INIT_GFP_FLAG
1131 rc = percpu_counter_init(&s->ls_lru_len_counter, 0, GFP_NOFS);
1133 rc = percpu_counter_init(&s->ls_lru_len_counter, 0);
1138 snprintf(name, sizeof(name), "lu_site_%s", top->ld_type->ldt_name);
1139 for (bits = lu_htable_order(top);
1140 bits >= LU_SITE_BITS_MIN; bits--) {
1141 s->ls_obj_hash = cfs_hash_create(name, bits, bits,
1142 bits - LU_SITE_BKT_BITS,
1145 CFS_HASH_SPIN_BKTLOCK |
1146 CFS_HASH_NO_ITEMREF |
1148 CFS_HASH_ASSERT_EMPTY |
1150 if (s->ls_obj_hash != NULL)
1154 if (s->ls_obj_hash == NULL) {
1155 CERROR("failed to create lu_site hash with bits: %lu\n", bits);
1159 s->ls_bkt_seed = prandom_u32();
1160 s->ls_bkt_cnt = max_t(long, 1 << LU_SITE_BKT_BITS,
1161 2 * num_possible_cpus());
1162 s->ls_bkt_cnt = roundup_pow_of_two(s->ls_bkt_cnt);
1163 OBD_ALLOC_LARGE(s->ls_bkts, s->ls_bkt_cnt * sizeof(*bkt));
1165 cfs_hash_putref(s->ls_obj_hash);
1166 s->ls_obj_hash = NULL;
1171 for (i = 0; i < s->ls_bkt_cnt; i++) {
1172 bkt = &s->ls_bkts[i];
1173 INIT_LIST_HEAD(&bkt->lsb_lru);
1174 init_waitqueue_head(&bkt->lsb_waitq);
1177 s->ls_stats = lprocfs_alloc_stats(LU_SS_LAST_STAT, 0);
1178 if (s->ls_stats == NULL) {
1179 OBD_FREE_LARGE(s->ls_bkts, s->ls_bkt_cnt * sizeof(*bkt));
1180 cfs_hash_putref(s->ls_obj_hash);
1181 s->ls_obj_hash = NULL;
1186 lprocfs_counter_init(s->ls_stats, LU_SS_CREATED,
1187 0, "created", "created");
1188 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_HIT,
1189 0, "cache_hit", "cache_hit");
1190 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_MISS,
1191 0, "cache_miss", "cache_miss");
1192 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_RACE,
1193 0, "cache_race", "cache_race");
1194 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_DEATH_RACE,
1195 0, "cache_death_race", "cache_death_race");
1196 lprocfs_counter_init(s->ls_stats, LU_SS_LRU_PURGED,
1197 0, "lru_purged", "lru_purged");
1199 INIT_LIST_HEAD(&s->ls_linkage);
1200 s->ls_top_dev = top;
1203 lu_ref_add(&top->ld_reference, "site-top", s);
1205 INIT_LIST_HEAD(&s->ls_ld_linkage);
1206 spin_lock_init(&s->ls_ld_lock);
1208 lu_dev_add_linkage(s, top);
1212 EXPORT_SYMBOL(lu_site_init);
1215 * Finalize \a s and release its resources.
1217 void lu_site_fini(struct lu_site *s)
1219 down_write(&lu_sites_guard);
1220 list_del_init(&s->ls_linkage);
1221 up_write(&lu_sites_guard);
1223 percpu_counter_destroy(&s->ls_lru_len_counter);
1225 if (s->ls_obj_hash != NULL) {
1226 cfs_hash_putref(s->ls_obj_hash);
1227 s->ls_obj_hash = NULL;
1230 OBD_FREE_LARGE(s->ls_bkts, s->ls_bkt_cnt * sizeof(*s->ls_bkts));
1232 if (s->ls_top_dev != NULL) {
1233 s->ls_top_dev->ld_site = NULL;
1234 lu_ref_del(&s->ls_top_dev->ld_reference, "site-top", s);
1235 lu_device_put(s->ls_top_dev);
1236 s->ls_top_dev = NULL;
1239 if (s->ls_stats != NULL)
1240 lprocfs_free_stats(&s->ls_stats);
1242 EXPORT_SYMBOL(lu_site_fini);
1245 * Called when initialization of stack for this site is completed.
1247 int lu_site_init_finish(struct lu_site *s)
1250 down_write(&lu_sites_guard);
1251 result = lu_context_refill(&lu_shrink_env.le_ctx);
1253 list_add(&s->ls_linkage, &lu_sites);
1254 up_write(&lu_sites_guard);
1257 EXPORT_SYMBOL(lu_site_init_finish);
1260 * Acquire additional reference on device \a d
1262 void lu_device_get(struct lu_device *d)
1264 atomic_inc(&d->ld_ref);
1266 EXPORT_SYMBOL(lu_device_get);
1269 * Release reference on device \a d.
1271 void lu_device_put(struct lu_device *d)
1273 LASSERT(atomic_read(&d->ld_ref) > 0);
1274 atomic_dec(&d->ld_ref);
1276 EXPORT_SYMBOL(lu_device_put);
1279 * Initialize device \a d of type \a t.
1281 int lu_device_init(struct lu_device *d, struct lu_device_type *t)
1283 if (atomic_inc_return(&t->ldt_device_nr) == 1 &&
1284 t->ldt_ops->ldto_start != NULL)
1285 t->ldt_ops->ldto_start(t);
1287 memset(d, 0, sizeof *d);
1289 lu_ref_init(&d->ld_reference);
1290 INIT_LIST_HEAD(&d->ld_linkage);
1294 EXPORT_SYMBOL(lu_device_init);
1297 * Finalize device \a d.
1299 void lu_device_fini(struct lu_device *d)
1301 struct lu_device_type *t = d->ld_type;
1303 if (d->ld_obd != NULL) {
1304 d->ld_obd->obd_lu_dev = NULL;
1308 lu_ref_fini(&d->ld_reference);
1309 LASSERTF(atomic_read(&d->ld_ref) == 0,
1310 "Refcount is %u\n", atomic_read(&d->ld_ref));
1311 LASSERT(atomic_read(&t->ldt_device_nr) > 0);
1313 if (atomic_dec_and_test(&t->ldt_device_nr) &&
1314 t->ldt_ops->ldto_stop != NULL)
1315 t->ldt_ops->ldto_stop(t);
1317 EXPORT_SYMBOL(lu_device_fini);
1320 * Initialize object \a o that is part of compound object \a h and was created
1323 int lu_object_init(struct lu_object *o, struct lu_object_header *h,
1324 struct lu_device *d)
1326 memset(o, 0, sizeof(*o));
1330 lu_ref_add_at(&d->ld_reference, &o->lo_dev_ref, "lu_object", o);
1331 INIT_LIST_HEAD(&o->lo_linkage);
1335 EXPORT_SYMBOL(lu_object_init);
1338 * Finalize object and release its resources.
1340 void lu_object_fini(struct lu_object *o)
1342 struct lu_device *dev = o->lo_dev;
1344 LASSERT(list_empty(&o->lo_linkage));
1347 lu_ref_del_at(&dev->ld_reference, &o->lo_dev_ref,
1353 EXPORT_SYMBOL(lu_object_fini);
1356 * Add object \a o as first layer of compound object \a h
1358 * This is typically called by the ->ldo_object_alloc() method of top-level
1361 void lu_object_add_top(struct lu_object_header *h, struct lu_object *o)
1363 list_move(&o->lo_linkage, &h->loh_layers);
1365 EXPORT_SYMBOL(lu_object_add_top);
1368 * Add object \a o as a layer of compound object, going after \a before.
1370 * This is typically called by the ->ldo_object_alloc() method of \a
1373 void lu_object_add(struct lu_object *before, struct lu_object *o)
1375 list_move(&o->lo_linkage, &before->lo_linkage);
1377 EXPORT_SYMBOL(lu_object_add);
1380 * Initialize compound object.
1382 int lu_object_header_init(struct lu_object_header *h)
1384 memset(h, 0, sizeof *h);
1385 atomic_set(&h->loh_ref, 1);
1386 INIT_HLIST_NODE(&h->loh_hash);
1387 INIT_LIST_HEAD(&h->loh_lru);
1388 INIT_LIST_HEAD(&h->loh_layers);
1389 lu_ref_init(&h->loh_reference);
1392 EXPORT_SYMBOL(lu_object_header_init);
1395 * Finalize compound object.
1397 void lu_object_header_fini(struct lu_object_header *h)
1399 LASSERT(list_empty(&h->loh_layers));
1400 LASSERT(list_empty(&h->loh_lru));
1401 LASSERT(hlist_unhashed(&h->loh_hash));
1402 lu_ref_fini(&h->loh_reference);
1404 EXPORT_SYMBOL(lu_object_header_fini);
1407 * Given a compound object, find its slice, corresponding to the device type
1410 struct lu_object *lu_object_locate(struct lu_object_header *h,
1411 const struct lu_device_type *dtype)
1413 struct lu_object *o;
1415 list_for_each_entry(o, &h->loh_layers, lo_linkage) {
1416 if (o->lo_dev->ld_type == dtype)
1421 EXPORT_SYMBOL(lu_object_locate);
1424 * Finalize and free devices in the device stack.
1426 * Finalize device stack by purging object cache, and calling
1427 * lu_device_type_operations::ldto_device_fini() and
1428 * lu_device_type_operations::ldto_device_free() on all devices in the stack.
1430 void lu_stack_fini(const struct lu_env *env, struct lu_device *top)
1432 struct lu_site *site = top->ld_site;
1433 struct lu_device *scan;
1434 struct lu_device *next;
1436 lu_site_purge(env, site, ~0);
1437 for (scan = top; scan != NULL; scan = next) {
1438 next = scan->ld_type->ldt_ops->ldto_device_fini(env, scan);
1439 lu_ref_del(&scan->ld_reference, "lu-stack", &lu_site_init);
1440 lu_device_put(scan);
1444 lu_site_purge(env, site, ~0);
1446 for (scan = top; scan != NULL; scan = next) {
1447 const struct lu_device_type *ldt = scan->ld_type;
1449 next = ldt->ldt_ops->ldto_device_free(env, scan);
1455 * Maximal number of tld slots.
1457 LU_CONTEXT_KEY_NR = 40
1460 static struct lu_context_key *lu_keys[LU_CONTEXT_KEY_NR] = { NULL, };
1462 static DECLARE_RWSEM(lu_key_initing);
1465 * Global counter incremented whenever key is registered, unregistered,
1466 * revived or quiesced. This is used to void unnecessary calls to
1467 * lu_context_refill(). No locking is provided, as initialization and shutdown
1468 * are supposed to be externally serialized.
1470 static atomic_t key_set_version = ATOMIC_INIT(0);
1475 int lu_context_key_register(struct lu_context_key *key)
1480 LASSERT(key->lct_init != NULL);
1481 LASSERT(key->lct_fini != NULL);
1482 LASSERT(key->lct_tags != 0);
1483 LASSERT(key->lct_owner != NULL);
1486 atomic_set(&key->lct_used, 1);
1487 lu_ref_init(&key->lct_reference);
1488 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1492 if (cmpxchg(&lu_keys[i], NULL, key) != NULL)
1496 atomic_inc(&key_set_version);
1500 lu_ref_fini(&key->lct_reference);
1501 atomic_set(&key->lct_used, 0);
1505 EXPORT_SYMBOL(lu_context_key_register);
1507 static void key_fini(struct lu_context *ctx, int index)
1509 if (ctx->lc_value != NULL && ctx->lc_value[index] != NULL) {
1510 struct lu_context_key *key;
1512 key = lu_keys[index];
1513 LASSERT(key != NULL);
1514 LASSERT(key->lct_fini != NULL);
1515 LASSERT(atomic_read(&key->lct_used) > 0);
1517 key->lct_fini(ctx, key, ctx->lc_value[index]);
1518 lu_ref_del(&key->lct_reference, "ctx", ctx);
1519 if (atomic_dec_and_test(&key->lct_used))
1520 wake_up_var(&key->lct_used);
1522 LASSERT(key->lct_owner != NULL);
1523 if ((ctx->lc_tags & LCT_NOREF) == 0) {
1524 LINVRNT(module_refcount(key->lct_owner) > 0);
1525 module_put(key->lct_owner);
1527 ctx->lc_value[index] = NULL;
1534 void lu_context_key_degister(struct lu_context_key *key)
1536 LASSERT(atomic_read(&key->lct_used) >= 1);
1537 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1539 lu_context_key_quiesce(key);
1541 key_fini(&lu_shrink_env.le_ctx, key->lct_index);
1544 * Wait until all transient contexts referencing this key have
1545 * run lu_context_key::lct_fini() method.
1547 atomic_dec(&key->lct_used);
1548 wait_var_event(&key->lct_used, atomic_read(&key->lct_used) == 0);
1550 if (!WARN_ON(lu_keys[key->lct_index] == NULL))
1551 lu_ref_fini(&key->lct_reference);
1553 smp_store_release(&lu_keys[key->lct_index], NULL);
1555 EXPORT_SYMBOL(lu_context_key_degister);
1558 * Register a number of keys. This has to be called after all keys have been
1559 * initialized by a call to LU_CONTEXT_KEY_INIT().
1561 int lu_context_key_register_many(struct lu_context_key *k, ...)
1563 struct lu_context_key *key = k;
1569 result = lu_context_key_register(key);
1572 key = va_arg(args, struct lu_context_key *);
1573 } while (key != NULL);
1579 lu_context_key_degister(k);
1580 k = va_arg(args, struct lu_context_key *);
1587 EXPORT_SYMBOL(lu_context_key_register_many);
1590 * De-register a number of keys. This is a dual to
1591 * lu_context_key_register_many().
1593 void lu_context_key_degister_many(struct lu_context_key *k, ...)
1599 lu_context_key_degister(k);
1600 k = va_arg(args, struct lu_context_key*);
1601 } while (k != NULL);
1604 EXPORT_SYMBOL(lu_context_key_degister_many);
1607 * Revive a number of keys.
1609 void lu_context_key_revive_many(struct lu_context_key *k, ...)
1615 lu_context_key_revive(k);
1616 k = va_arg(args, struct lu_context_key*);
1617 } while (k != NULL);
1620 EXPORT_SYMBOL(lu_context_key_revive_many);
1623 * Quiescent a number of keys.
1625 void lu_context_key_quiesce_many(struct lu_context_key *k, ...)
1631 lu_context_key_quiesce(k);
1632 k = va_arg(args, struct lu_context_key*);
1633 } while (k != NULL);
1636 EXPORT_SYMBOL(lu_context_key_quiesce_many);
1639 * Return value associated with key \a key in context \a ctx.
1641 void *lu_context_key_get(const struct lu_context *ctx,
1642 const struct lu_context_key *key)
1644 LINVRNT(ctx->lc_state == LCS_ENTERED);
1645 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1646 LASSERT(lu_keys[key->lct_index] == key);
1647 return ctx->lc_value[key->lct_index];
1649 EXPORT_SYMBOL(lu_context_key_get);
1652 * List of remembered contexts. XXX document me.
1654 static LIST_HEAD(lu_context_remembered);
1655 static DEFINE_SPINLOCK(lu_context_remembered_guard);
1658 * Destroy \a key in all remembered contexts. This is used to destroy key
1659 * values in "shared" contexts (like service threads), when a module owning
1660 * the key is about to be unloaded.
1662 void lu_context_key_quiesce(struct lu_context_key *key)
1664 struct lu_context *ctx;
1666 if (!(key->lct_tags & LCT_QUIESCENT)) {
1668 * The write-lock on lu_key_initing will ensure that any
1669 * keys_fill() which didn't see LCT_QUIESCENT will have
1670 * finished before we call key_fini().
1672 down_write(&lu_key_initing);
1673 key->lct_tags |= LCT_QUIESCENT;
1674 up_write(&lu_key_initing);
1676 spin_lock(&lu_context_remembered_guard);
1677 list_for_each_entry(ctx, &lu_context_remembered, lc_remember) {
1678 spin_until_cond(READ_ONCE(ctx->lc_state) != LCS_LEAVING);
1679 key_fini(ctx, key->lct_index);
1682 spin_unlock(&lu_context_remembered_guard);
1686 void lu_context_key_revive(struct lu_context_key *key)
1688 key->lct_tags &= ~LCT_QUIESCENT;
1689 atomic_inc(&key_set_version);
1692 static void keys_fini(struct lu_context *ctx)
1696 if (ctx->lc_value == NULL)
1699 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i)
1702 OBD_FREE(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1703 ctx->lc_value = NULL;
1706 static int keys_fill(struct lu_context *ctx)
1712 * A serialisation with lu_context_key_quiesce() is needed, to
1713 * ensure we see LCT_QUIESCENT and don't allocate a new value
1714 * after it freed one. The rwsem provides this. As down_read()
1715 * does optimistic spinning while the writer is active, this is
1716 * unlikely to ever sleep.
1718 down_read(&lu_key_initing);
1719 ctx->lc_version = atomic_read(&key_set_version);
1721 LINVRNT(ctx->lc_value);
1722 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1723 struct lu_context_key *key;
1726 if (!ctx->lc_value[i] && key &&
1727 (key->lct_tags & ctx->lc_tags) &&
1729 * Don't create values for a LCT_QUIESCENT key, as this
1730 * will pin module owning a key.
1732 !(key->lct_tags & LCT_QUIESCENT)) {
1735 LINVRNT(key->lct_init != NULL);
1736 LINVRNT(key->lct_index == i);
1738 LASSERT(key->lct_owner != NULL);
1739 if (!(ctx->lc_tags & LCT_NOREF) &&
1740 try_module_get(key->lct_owner) == 0) {
1741 /* module is unloading, skip this key */
1745 value = key->lct_init(ctx, key);
1746 if (unlikely(IS_ERR(value))) {
1747 rc = PTR_ERR(value);
1751 lu_ref_add_atomic(&key->lct_reference, "ctx", ctx);
1752 atomic_inc(&key->lct_used);
1754 * This is the only place in the code, where an
1755 * element of ctx->lc_value[] array is set to non-NULL
1758 ctx->lc_value[i] = value;
1759 if (key->lct_exit != NULL)
1760 ctx->lc_tags |= LCT_HAS_EXIT;
1764 up_read(&lu_key_initing);
1768 static int keys_init(struct lu_context *ctx)
1770 OBD_ALLOC(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1771 if (likely(ctx->lc_value != NULL))
1772 return keys_fill(ctx);
1778 * Initialize context data-structure. Create values for all keys.
1780 int lu_context_init(struct lu_context *ctx, __u32 tags)
1784 memset(ctx, 0, sizeof *ctx);
1785 ctx->lc_state = LCS_INITIALIZED;
1786 ctx->lc_tags = tags;
1787 if (tags & LCT_REMEMBER) {
1788 spin_lock(&lu_context_remembered_guard);
1789 list_add(&ctx->lc_remember, &lu_context_remembered);
1790 spin_unlock(&lu_context_remembered_guard);
1792 INIT_LIST_HEAD(&ctx->lc_remember);
1795 rc = keys_init(ctx);
1797 lu_context_fini(ctx);
1801 EXPORT_SYMBOL(lu_context_init);
1804 * Finalize context data-structure. Destroy key values.
1806 void lu_context_fini(struct lu_context *ctx)
1808 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1809 ctx->lc_state = LCS_FINALIZED;
1811 if ((ctx->lc_tags & LCT_REMEMBER) == 0) {
1812 LASSERT(list_empty(&ctx->lc_remember));
1814 /* could race with key degister */
1815 spin_lock(&lu_context_remembered_guard);
1816 list_del_init(&ctx->lc_remember);
1817 spin_unlock(&lu_context_remembered_guard);
1821 EXPORT_SYMBOL(lu_context_fini);
1824 * Called before entering context.
1826 void lu_context_enter(struct lu_context *ctx)
1828 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1829 ctx->lc_state = LCS_ENTERED;
1831 EXPORT_SYMBOL(lu_context_enter);
1834 * Called after exiting from \a ctx
1836 void lu_context_exit(struct lu_context *ctx)
1840 LINVRNT(ctx->lc_state == LCS_ENTERED);
1842 * Disable preempt to ensure we get a warning if
1843 * any lct_exit ever tries to sleep. That would hurt
1844 * lu_context_key_quiesce() which spins waiting for us.
1845 * This also ensure we aren't preempted while the state
1846 * is LCS_LEAVING, as that too would cause problems for
1847 * lu_context_key_quiesce().
1851 * Ensure lu_context_key_quiesce() sees LCS_LEAVING
1852 * or we see LCT_QUIESCENT
1854 smp_store_mb(ctx->lc_state, LCS_LEAVING);
1855 if (ctx->lc_tags & LCT_HAS_EXIT && ctx->lc_value) {
1856 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1857 struct lu_context_key *key;
1860 if (ctx->lc_value[i] &&
1861 !(key->lct_tags & LCT_QUIESCENT) &&
1863 key->lct_exit(ctx, key, ctx->lc_value[i]);
1867 smp_store_release(&ctx->lc_state, LCS_LEFT);
1870 EXPORT_SYMBOL(lu_context_exit);
1873 * Allocate for context all missing keys that were registered after context
1874 * creation. key_set_version is only changed in rare cases when modules
1875 * are loaded and removed.
1877 int lu_context_refill(struct lu_context *ctx)
1879 if (likely(ctx->lc_version == atomic_read(&key_set_version)))
1882 return keys_fill(ctx);
1886 * lu_ctx_tags/lu_ses_tags will be updated if there are new types of
1887 * obd being added. Currently, this is only used on client side, specifically
1888 * for echo device client, for other stack (like ptlrpc threads), context are
1889 * predefined when the lu_device type are registered, during the module probe
1892 u32 lu_context_tags_default = LCT_CL_THREAD;
1893 u32 lu_session_tags_default = LCT_SESSION;
1895 void lu_context_tags_update(__u32 tags)
1897 spin_lock(&lu_context_remembered_guard);
1898 lu_context_tags_default |= tags;
1899 atomic_inc(&key_set_version);
1900 spin_unlock(&lu_context_remembered_guard);
1902 EXPORT_SYMBOL(lu_context_tags_update);
1904 void lu_context_tags_clear(__u32 tags)
1906 spin_lock(&lu_context_remembered_guard);
1907 lu_context_tags_default &= ~tags;
1908 atomic_inc(&key_set_version);
1909 spin_unlock(&lu_context_remembered_guard);
1911 EXPORT_SYMBOL(lu_context_tags_clear);
1913 void lu_session_tags_update(__u32 tags)
1915 spin_lock(&lu_context_remembered_guard);
1916 lu_session_tags_default |= tags;
1917 atomic_inc(&key_set_version);
1918 spin_unlock(&lu_context_remembered_guard);
1920 EXPORT_SYMBOL(lu_session_tags_update);
1922 void lu_session_tags_clear(__u32 tags)
1924 spin_lock(&lu_context_remembered_guard);
1925 lu_session_tags_default &= ~tags;
1926 atomic_inc(&key_set_version);
1927 spin_unlock(&lu_context_remembered_guard);
1929 EXPORT_SYMBOL(lu_session_tags_clear);
1931 int lu_env_init(struct lu_env *env, __u32 tags)
1936 result = lu_context_init(&env->le_ctx, tags);
1937 if (likely(result == 0))
1938 lu_context_enter(&env->le_ctx);
1941 EXPORT_SYMBOL(lu_env_init);
1943 void lu_env_fini(struct lu_env *env)
1945 lu_context_exit(&env->le_ctx);
1946 lu_context_fini(&env->le_ctx);
1949 EXPORT_SYMBOL(lu_env_fini);
1951 int lu_env_refill(struct lu_env *env)
1955 result = lu_context_refill(&env->le_ctx);
1956 if (result == 0 && env->le_ses != NULL)
1957 result = lu_context_refill(env->le_ses);
1960 EXPORT_SYMBOL(lu_env_refill);
1963 * Currently, this API will only be used by echo client.
1964 * Because echo client and normal lustre client will share
1965 * same cl_env cache. So echo client needs to refresh
1966 * the env context after it get one from the cache, especially
1967 * when normal client and echo client co-exist in the same client.
1969 int lu_env_refill_by_tags(struct lu_env *env, __u32 ctags,
1974 if ((env->le_ctx.lc_tags & ctags) != ctags) {
1975 env->le_ctx.lc_version = 0;
1976 env->le_ctx.lc_tags |= ctags;
1979 if (env->le_ses && (env->le_ses->lc_tags & stags) != stags) {
1980 env->le_ses->lc_version = 0;
1981 env->le_ses->lc_tags |= stags;
1984 result = lu_env_refill(env);
1988 EXPORT_SYMBOL(lu_env_refill_by_tags);
1991 struct lu_env_item {
1992 struct task_struct *lei_task; /* rhashtable key */
1993 struct rhash_head lei_linkage;
1994 struct lu_env *lei_env;
1995 struct rcu_head lei_rcu_head;
1998 static const struct rhashtable_params lu_env_rhash_params = {
1999 .key_len = sizeof(struct task_struct *),
2000 .key_offset = offsetof(struct lu_env_item, lei_task),
2001 .head_offset = offsetof(struct lu_env_item, lei_linkage),
2004 struct rhashtable lu_env_rhash;
2006 struct lu_env_percpu {
2007 struct task_struct *lep_task;
2008 struct lu_env *lep_env ____cacheline_aligned_in_smp;
2011 static struct lu_env_percpu lu_env_percpu[NR_CPUS];
2013 int lu_env_add(struct lu_env *env)
2015 struct lu_env_item *lei, *old;
2023 lei->lei_task = current;
2026 old = rhashtable_lookup_get_insert_fast(&lu_env_rhash,
2028 lu_env_rhash_params);
2033 EXPORT_SYMBOL(lu_env_add);
2035 static void lu_env_item_free(struct rcu_head *head)
2037 struct lu_env_item *lei;
2039 lei = container_of(head, struct lu_env_item, lei_rcu_head);
2043 void lu_env_remove(struct lu_env *env)
2045 struct lu_env_item *lei;
2046 const void *task = current;
2049 for_each_possible_cpu(i) {
2050 if (lu_env_percpu[i].lep_env == env) {
2051 LASSERT(lu_env_percpu[i].lep_task == task);
2052 lu_env_percpu[i].lep_task = NULL;
2053 lu_env_percpu[i].lep_env = NULL;
2057 /* The rcu_lock is not taking in this case since the key
2058 * used is the actual task_struct. This implies that each
2059 * object is only removed by the owning thread, so there
2060 * can never be a race on a particular object.
2062 lei = rhashtable_lookup_fast(&lu_env_rhash, &task,
2063 lu_env_rhash_params);
2064 if (lei && rhashtable_remove_fast(&lu_env_rhash, &lei->lei_linkage,
2065 lu_env_rhash_params) == 0)
2066 call_rcu(&lei->lei_rcu_head, lu_env_item_free);
2068 EXPORT_SYMBOL(lu_env_remove);
2070 struct lu_env *lu_env_find(void)
2072 struct lu_env *env = NULL;
2073 struct lu_env_item *lei;
2074 const void *task = current;
2077 if (lu_env_percpu[i].lep_task == current) {
2078 env = lu_env_percpu[i].lep_env;
2084 lei = rhashtable_lookup_fast(&lu_env_rhash, &task,
2085 lu_env_rhash_params);
2088 lu_env_percpu[i].lep_task = current;
2089 lu_env_percpu[i].lep_env = env;
2095 EXPORT_SYMBOL(lu_env_find);
2097 static struct shrinker *lu_site_shrinker;
2099 typedef struct lu_site_stats{
2100 unsigned lss_populated;
2101 unsigned lss_max_search;
2106 static void lu_site_stats_get(const struct lu_site *s,
2107 lu_site_stats_t *stats)
2109 int cnt = cfs_hash_size_get(s->ls_obj_hash);
2111 * percpu_counter_sum_positive() won't accept a const pointer
2112 * as it does modify the struct by taking a spinlock
2114 struct lu_site *s2 = (struct lu_site *)s;
2116 stats->lss_busy += cnt -
2117 percpu_counter_sum_positive(&s2->ls_lru_len_counter);
2119 stats->lss_total += cnt;
2120 stats->lss_max_search = 0;
2121 stats->lss_populated = 0;
2126 * lu_cache_shrink_count() returns an approximate number of cached objects
2127 * that can be freed by shrink_slab(). A counter, which tracks the
2128 * number of items in the site's lru, is maintained in a percpu_counter
2129 * for each site. The percpu values are incremented and decremented as
2130 * objects are added or removed from the lru. The percpu values are summed
2131 * and saved whenever a percpu value exceeds a threshold. Thus the saved,
2132 * summed value at any given time may not accurately reflect the current
2133 * lru length. But this value is sufficiently accurate for the needs of
2136 * Using a per cpu counter is a compromise solution to concurrent access:
2137 * lu_object_put() can update the counter without locking the site and
2138 * lu_cache_shrink_count can sum the counters without locking each
2139 * ls_obj_hash bucket.
2141 static unsigned long lu_cache_shrink_count(struct shrinker *sk,
2142 struct shrink_control *sc)
2145 struct lu_site *tmp;
2146 unsigned long cached = 0;
2148 if (!(sc->gfp_mask & __GFP_FS))
2151 down_read(&lu_sites_guard);
2152 list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage)
2153 cached += percpu_counter_read_positive(&s->ls_lru_len_counter);
2154 up_read(&lu_sites_guard);
2156 cached = (cached / 100) * sysctl_vfs_cache_pressure;
2157 CDEBUG(D_INODE, "%ld objects cached, cache pressure %d\n",
2158 cached, sysctl_vfs_cache_pressure);
2163 static unsigned long lu_cache_shrink_scan(struct shrinker *sk,
2164 struct shrink_control *sc)
2167 struct lu_site *tmp;
2168 unsigned long remain = sc->nr_to_scan;
2171 if (!(sc->gfp_mask & __GFP_FS))
2172 /* We must not take the lu_sites_guard lock when
2173 * __GFP_FS is *not* set because of the deadlock
2174 * possibility detailed above. Additionally,
2175 * since we cannot determine the number of
2176 * objects in the cache without taking this
2177 * lock, we're in a particularly tough spot. As
2178 * a result, we'll just lie and say our cache is
2179 * empty. This _should_ be ok, as we can't
2180 * reclaim objects when __GFP_FS is *not* set
2185 down_write(&lu_sites_guard);
2186 list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
2187 remain = lu_site_purge(&lu_shrink_env, s, remain);
2189 * Move just shrunk site to the tail of site list to
2190 * assure shrinking fairness.
2192 list_move_tail(&s->ls_linkage, &splice);
2194 list_splice(&splice, lu_sites.prev);
2195 up_write(&lu_sites_guard);
2197 return sc->nr_to_scan - remain;
2200 #ifndef HAVE_SHRINKER_COUNT
2202 * There exists a potential lock inversion deadlock scenario when using
2203 * Lustre on top of ZFS. This occurs between one of ZFS's
2204 * buf_hash_table.ht_lock's, and Lustre's lu_sites_guard lock. Essentially,
2205 * thread A will take the lu_sites_guard lock and sleep on the ht_lock,
2206 * while thread B will take the ht_lock and sleep on the lu_sites_guard
2207 * lock. Obviously neither thread will wake and drop their respective hold
2210 * To prevent this from happening we must ensure the lu_sites_guard lock is
2211 * not taken while down this code path. ZFS reliably does not set the
2212 * __GFP_FS bit in its code paths, so this can be used to determine if it
2213 * is safe to take the lu_sites_guard lock.
2215 * Ideally we should accurately return the remaining number of cached
2216 * objects without taking the lu_sites_guard lock, but this is not
2217 * possible in the current implementation.
2219 static int lu_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
2222 struct shrink_control scv = {
2223 .nr_to_scan = shrink_param(sc, nr_to_scan),
2224 .gfp_mask = shrink_param(sc, gfp_mask)
2227 CDEBUG(D_INODE, "Shrink %lu objects\n", scv.nr_to_scan);
2229 if (scv.nr_to_scan != 0)
2230 lu_cache_shrink_scan(shrinker, &scv);
2232 cached = lu_cache_shrink_count(shrinker, &scv);
2236 #endif /* HAVE_SHRINKER_COUNT */
2244 * Environment to be used in debugger, contains all tags.
2246 static struct lu_env lu_debugging_env;
2249 * Debugging printer function using printk().
2251 int lu_printk_printer(const struct lu_env *env,
2252 void *unused, const char *format, ...)
2256 va_start(args, format);
2257 vprintk(format, args);
2262 int lu_debugging_setup(void)
2264 return lu_env_init(&lu_debugging_env, ~0);
2267 void lu_context_keys_dump(void)
2271 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
2272 struct lu_context_key *key;
2276 CERROR("[%d]: %p %x (%p,%p,%p) %d %d \"%s\"@%p\n",
2277 i, key, key->lct_tags,
2278 key->lct_init, key->lct_fini, key->lct_exit,
2279 key->lct_index, atomic_read(&key->lct_used),
2280 key->lct_owner ? key->lct_owner->name : "",
2282 lu_ref_print(&key->lct_reference);
2288 * Initialization of global lu_* data.
2290 int lu_global_init(void)
2293 DEF_SHRINKER_VAR(shvar, lu_cache_shrink,
2294 lu_cache_shrink_count, lu_cache_shrink_scan);
2296 CDEBUG(D_INFO, "Lustre LU module (%p).\n", &lu_keys);
2298 result = lu_ref_global_init();
2302 LU_CONTEXT_KEY_INIT(&lu_global_key);
2303 result = lu_context_key_register(&lu_global_key);
2308 * At this level, we don't know what tags are needed, so allocate them
2309 * conservatively. This should not be too bad, because this
2310 * environment is global.
2312 down_write(&lu_sites_guard);
2313 result = lu_env_init(&lu_shrink_env, LCT_SHRINKER);
2314 up_write(&lu_sites_guard);
2319 * seeks estimation: 3 seeks to read a record from oi, one to read
2320 * inode, one for ea. Unfortunately setting this high value results in
2321 * lu_object/inode cache consuming all the memory.
2323 lu_site_shrinker = set_shrinker(DEFAULT_SEEKS, &shvar);
2324 if (lu_site_shrinker == NULL)
2327 result = rhashtable_init(&lu_env_rhash, &lu_env_rhash_params);
2333 * Dual to lu_global_init().
2335 void lu_global_fini(void)
2337 if (lu_site_shrinker != NULL) {
2338 remove_shrinker(lu_site_shrinker);
2339 lu_site_shrinker = NULL;
2342 lu_context_key_degister(&lu_global_key);
2345 * Tear shrinker environment down _after_ de-registering
2346 * lu_global_key, because the latter has a value in the former.
2348 down_write(&lu_sites_guard);
2349 lu_env_fini(&lu_shrink_env);
2350 up_write(&lu_sites_guard);
2352 rhashtable_destroy(&lu_env_rhash);
2354 lu_ref_global_fini();
2357 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx)
2359 #ifdef CONFIG_PROC_FS
2360 struct lprocfs_counter ret;
2362 lprocfs_stats_collect(stats, idx, &ret);
2363 return (__u32)ret.lc_count;
2370 * Output site statistical counters into a buffer. Suitable for
2371 * lprocfs_rd_*()-style functions.
2373 int lu_site_stats_seq_print(const struct lu_site *s, struct seq_file *m)
2375 lu_site_stats_t stats;
2377 memset(&stats, 0, sizeof(stats));
2378 lu_site_stats_get(s, &stats);
2380 seq_printf(m, "%d/%d %d/%d %d %d %d %d %d %d %d\n",
2383 stats.lss_populated,
2384 CFS_HASH_NHLIST(s->ls_obj_hash),
2385 stats.lss_max_search,
2386 ls_stats_read(s->ls_stats, LU_SS_CREATED),
2387 ls_stats_read(s->ls_stats, LU_SS_CACHE_HIT),
2388 ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
2389 ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
2390 ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
2391 ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED));
2394 EXPORT_SYMBOL(lu_site_stats_seq_print);
2397 * Helper function to initialize a number of kmem slab caches at once.
2399 int lu_kmem_init(struct lu_kmem_descr *caches)
2402 struct lu_kmem_descr *iter = caches;
2404 for (result = 0; iter->ckd_cache != NULL; ++iter) {
2405 *iter->ckd_cache = kmem_cache_create(iter->ckd_name,
2408 if (*iter->ckd_cache == NULL) {
2410 /* free all previously allocated caches */
2411 lu_kmem_fini(caches);
2417 EXPORT_SYMBOL(lu_kmem_init);
2420 * Helper function to finalize a number of kmem slab cached at once. Dual to
2423 void lu_kmem_fini(struct lu_kmem_descr *caches)
2425 for (; caches->ckd_cache != NULL; ++caches) {
2426 if (*caches->ckd_cache != NULL) {
2427 kmem_cache_destroy(*caches->ckd_cache);
2428 *caches->ckd_cache = NULL;
2432 EXPORT_SYMBOL(lu_kmem_fini);
2435 * Temporary solution to be able to assign fid in ->do_create()
2436 * till we have fully-functional OST fids
2438 void lu_object_assign_fid(const struct lu_env *env, struct lu_object *o,
2439 const struct lu_fid *fid)
2441 struct lu_site *s = o->lo_dev->ld_site;
2442 struct lu_fid *old = &o->lo_header->loh_fid;
2443 struct cfs_hash *hs;
2444 struct cfs_hash_bd bd;
2446 LASSERT(fid_is_zero(old));
2448 /* supposed to be unique */
2449 hs = s->ls_obj_hash;
2450 cfs_hash_bd_get_and_lock(hs, (void *)fid, &bd, 1);
2451 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
2454 struct lu_object *shadow;
2456 shadow = htable_lookup(s, &bd, fid, &version);
2457 /* supposed to be unique */
2458 LASSERT(IS_ERR(shadow) && PTR_ERR(shadow) == -ENOENT);
2462 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
2463 cfs_hash_bd_unlock(hs, &bd, 1);
2465 EXPORT_SYMBOL(lu_object_assign_fid);
2468 * allocates object with 0 (non-assiged) fid
2469 * XXX: temporary solution to be able to assign fid in ->do_create()
2470 * till we have fully-functional OST fids
2472 struct lu_object *lu_object_anon(const struct lu_env *env,
2473 struct lu_device *dev,
2474 const struct lu_object_conf *conf)
2477 struct lu_object *o;
2481 o = lu_object_alloc(env, dev, &fid);
2483 rc = lu_object_start(env, dev, o, conf);
2485 lu_object_free(env, o);
2492 EXPORT_SYMBOL(lu_object_anon);
2494 struct lu_buf LU_BUF_NULL = {
2498 EXPORT_SYMBOL(LU_BUF_NULL);
2500 void lu_buf_free(struct lu_buf *buf)
2504 LASSERT(buf->lb_len > 0);
2505 OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);
2510 EXPORT_SYMBOL(lu_buf_free);
2512 void lu_buf_alloc(struct lu_buf *buf, size_t size)
2515 LASSERT(buf->lb_buf == NULL);
2516 LASSERT(buf->lb_len == 0);
2517 OBD_ALLOC_LARGE(buf->lb_buf, size);
2518 if (likely(buf->lb_buf))
2521 EXPORT_SYMBOL(lu_buf_alloc);
2523 void lu_buf_realloc(struct lu_buf *buf, size_t size)
2526 lu_buf_alloc(buf, size);
2528 EXPORT_SYMBOL(lu_buf_realloc);
2530 struct lu_buf *lu_buf_check_and_alloc(struct lu_buf *buf, size_t len)
2532 if (buf->lb_buf == NULL && buf->lb_len == 0)
2533 lu_buf_alloc(buf, len);
2535 if ((len > buf->lb_len) && (buf->lb_buf != NULL))
2536 lu_buf_realloc(buf, len);
2540 EXPORT_SYMBOL(lu_buf_check_and_alloc);
2543 * Increase the size of the \a buf.
2544 * preserves old data in buffer
2545 * old buffer remains unchanged on error
2546 * \retval 0 or -ENOMEM
2548 int lu_buf_check_and_grow(struct lu_buf *buf, size_t len)
2552 if (len <= buf->lb_len)
2555 OBD_ALLOC_LARGE(ptr, len);
2559 /* Free the old buf */
2560 if (buf->lb_buf != NULL) {
2561 memcpy(ptr, buf->lb_buf, buf->lb_len);
2562 OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);
2569 EXPORT_SYMBOL(lu_buf_check_and_grow);