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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2013, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/obdclass/lu_object.c
39 * These are the only exported functions, they provide some generic
40 * infrastructure for managing object devices
42 * Author: Nikita Danilov <nikita.danilov@sun.com>
45 #define DEBUG_SUBSYSTEM S_CLASS
47 #include <libcfs/libcfs.h>
48 #include <linux/module.h>
49 #include <libcfs/libcfs_hash.h> /* hash_long() */
50 #include <obd_class.h>
51 #include <obd_support.h>
52 #include <lustre_disk.h>
53 #include <lustre_fid.h>
54 #include <lu_object.h>
56 #include <libcfs/list.h>
59 LU_CACHE_PERCENT_MAX = 50,
60 LU_CACHE_PERCENT_DEFAULT = 20
63 #define LU_CACHE_NR_MAX_ADJUST 128
64 #define LU_CACHE_NR_UNLIMITED -1
65 #define LU_CACHE_NR_DEFAULT LU_CACHE_NR_UNLIMITED
66 #define LU_CACHE_NR_LDISKFS_LIMIT LU_CACHE_NR_UNLIMITED
67 /** This is set to roughly (20 * OSS_NTHRS_MAX) to prevent thrashing */
68 #define LU_CACHE_NR_ZFS_LIMIT 10240
70 #define LU_SITE_BITS_MIN 12
71 #define LU_SITE_BITS_MAX 24
73 * total 256 buckets, we don't want too many buckets because:
74 * - consume too much memory
75 * - avoid unbalanced LRU list
77 #define LU_SITE_BKT_BITS 8
80 static unsigned int lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
81 CFS_MODULE_PARM(lu_cache_percent, "i", int, 0644,
82 "Percentage of memory to be used as lu_object cache");
84 static long lu_cache_nr = LU_CACHE_NR_DEFAULT;
85 CFS_MODULE_PARM(lu_cache_nr, "l", long, 0644,
86 "Maximum number of objects in lu_object cache");
88 static void lu_object_free(const struct lu_env *env, struct lu_object *o);
91 * Decrease reference counter on object. If last reference is freed, return
92 * object to the cache, unless lu_object_is_dying(o) holds. In the latter
93 * case, free object immediately.
95 void lu_object_put(const struct lu_env *env, struct lu_object *o)
97 struct lu_site_bkt_data *bkt;
98 struct lu_object_header *top;
100 struct lu_object *orig;
102 const struct lu_fid *fid;
105 site = o->lo_dev->ld_site;
109 * till we have full fids-on-OST implemented anonymous objects
110 * are possible in OSP. such an object isn't listed in the site
111 * so we should not remove it from the site.
113 fid = lu_object_fid(o);
114 if (fid_is_zero(fid)) {
115 LASSERT(top->loh_hash.next == NULL
116 && top->loh_hash.pprev == NULL);
117 LASSERT(list_empty(&top->loh_lru));
118 if (!atomic_dec_and_test(&top->loh_ref))
120 list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
121 if (o->lo_ops->loo_object_release != NULL)
122 o->lo_ops->loo_object_release(env, o);
124 lu_object_free(env, orig);
128 cfs_hash_bd_get(site->ls_obj_hash, &top->loh_fid, &bd);
129 bkt = cfs_hash_bd_extra_get(site->ls_obj_hash, &bd);
131 if (!cfs_hash_bd_dec_and_lock(site->ls_obj_hash, &bd, &top->loh_ref)) {
132 if (lu_object_is_dying(top)) {
135 * somebody may be waiting for this, currently only
136 * used for cl_object, see cl_object_put_last().
138 wake_up_all(&bkt->lsb_marche_funebre);
143 LASSERT(bkt->lsb_busy > 0);
146 * When last reference is released, iterate over object
147 * layers, and notify them that object is no longer busy.
149 list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
150 if (o->lo_ops->loo_object_release != NULL)
151 o->lo_ops->loo_object_release(env, o);
154 if (!lu_object_is_dying(top)) {
155 LASSERT(list_empty(&top->loh_lru));
156 list_add_tail(&top->loh_lru, &bkt->lsb_lru);
157 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
162 * If object is dying (will not be cached), removed it
163 * from hash table and LRU.
165 * This is done with hash table and LRU lists locked. As the only
166 * way to acquire first reference to previously unreferenced
167 * object is through hash-table lookup (lu_object_find()),
168 * or LRU scanning (lu_site_purge()), that are done under hash-table
169 * and LRU lock, no race with concurrent object lookup is possible
170 * and we can safely destroy object below.
172 if (!test_and_set_bit(LU_OBJECT_UNHASHED, &top->loh_flags))
173 cfs_hash_bd_del_locked(site->ls_obj_hash, &bd, &top->loh_hash);
174 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
176 * Object was already removed from hash and lru above, can
179 lu_object_free(env, orig);
181 EXPORT_SYMBOL(lu_object_put);
184 * Put object and don't keep in cache. This is temporary solution for
185 * multi-site objects when its layering is not constant.
187 void lu_object_put_nocache(const struct lu_env *env, struct lu_object *o)
189 set_bit(LU_OBJECT_HEARD_BANSHEE, &o->lo_header->loh_flags);
190 return lu_object_put(env, o);
192 EXPORT_SYMBOL(lu_object_put_nocache);
195 * Kill the object and take it out of LRU cache.
196 * Currently used by client code for layout change.
198 void lu_object_unhash(const struct lu_env *env, struct lu_object *o)
200 struct lu_object_header *top;
203 set_bit(LU_OBJECT_HEARD_BANSHEE, &top->loh_flags);
204 if (!test_and_set_bit(LU_OBJECT_UNHASHED, &top->loh_flags)) {
205 cfs_hash_t *obj_hash = o->lo_dev->ld_site->ls_obj_hash;
208 cfs_hash_bd_get_and_lock(obj_hash, &top->loh_fid, &bd, 1);
209 list_del_init(&top->loh_lru);
210 cfs_hash_bd_del_locked(obj_hash, &bd, &top->loh_hash);
211 cfs_hash_bd_unlock(obj_hash, &bd, 1);
214 EXPORT_SYMBOL(lu_object_unhash);
217 * Allocate new object.
219 * This follows object creation protocol, described in the comment within
220 * struct lu_device_operations definition.
222 static struct lu_object *lu_object_alloc(const struct lu_env *env,
223 struct lu_device *dev,
224 const struct lu_fid *f,
225 const struct lu_object_conf *conf)
227 struct lu_object *scan;
228 struct lu_object *top;
229 struct list_head *layers;
230 unsigned int init_mask = 0;
231 unsigned int init_flag;
237 * Create top-level object slice. This will also create
240 top = dev->ld_ops->ldo_object_alloc(env, NULL, dev);
242 RETURN(ERR_PTR(-ENOMEM));
246 * This is the only place where object fid is assigned. It's constant
249 top->lo_header->loh_fid = *f;
250 layers = &top->lo_header->loh_layers;
254 * Call ->loo_object_init() repeatedly, until no more new
255 * object slices are created.
259 list_for_each_entry(scan, layers, lo_linkage) {
260 if (init_mask & init_flag)
263 scan->lo_header = top->lo_header;
264 result = scan->lo_ops->loo_object_init(env, scan, conf);
266 lu_object_free(env, top);
267 RETURN(ERR_PTR(result));
269 init_mask |= init_flag;
275 list_for_each_entry_reverse(scan, layers, lo_linkage) {
276 if (scan->lo_ops->loo_object_start != NULL) {
277 result = scan->lo_ops->loo_object_start(env, scan);
279 lu_object_free(env, top);
280 RETURN(ERR_PTR(result));
285 lprocfs_counter_incr(dev->ld_site->ls_stats, LU_SS_CREATED);
292 static void lu_object_free(const struct lu_env *env, struct lu_object *o)
294 struct lu_site_bkt_data *bkt;
295 struct lu_site *site;
296 struct lu_object *scan;
297 struct list_head *layers;
298 struct list_head splice;
300 site = o->lo_dev->ld_site;
301 layers = &o->lo_header->loh_layers;
302 bkt = lu_site_bkt_from_fid(site, &o->lo_header->loh_fid);
304 * First call ->loo_object_delete() method to release all resources.
306 list_for_each_entry_reverse(scan, layers, lo_linkage) {
307 if (scan->lo_ops->loo_object_delete != NULL)
308 scan->lo_ops->loo_object_delete(env, scan);
312 * Then, splice object layers into stand-alone list, and call
313 * ->loo_object_free() on all layers to free memory. Splice is
314 * necessary, because lu_object_header is freed together with the
317 INIT_LIST_HEAD(&splice);
318 list_splice_init(layers, &splice);
319 while (!list_empty(&splice)) {
321 * Free layers in bottom-to-top order, so that object header
322 * lives as long as possible and ->loo_object_free() methods
323 * can look at its contents.
325 o = container_of0(splice.prev, struct lu_object, lo_linkage);
326 list_del_init(&o->lo_linkage);
327 LASSERT(o->lo_ops->loo_object_free != NULL);
328 o->lo_ops->loo_object_free(env, o);
331 if (waitqueue_active(&bkt->lsb_marche_funebre))
332 wake_up_all(&bkt->lsb_marche_funebre);
336 * Free \a nr objects from the cold end of the site LRU list.
338 int lu_site_purge(const struct lu_env *env, struct lu_site *s, int nr)
340 struct lu_object_header *h;
341 struct lu_object_header *temp;
342 struct lu_site_bkt_data *bkt;
345 struct list_head dispose;
352 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_NO_LRU))
355 INIT_LIST_HEAD(&dispose);
357 * Under LRU list lock, scan LRU list and move unreferenced objects to
358 * the dispose list, removing them from LRU and hash table.
360 start = s->ls_purge_start;
361 bnr = (nr == ~0) ? -1 : nr / (int)CFS_HASH_NBKT(s->ls_obj_hash) + 1;
364 * It doesn't make any sense to make purge threads parallel, that can
365 * only bring troubles to us. See LU-5331.
367 mutex_lock(&s->ls_purge_mutex);
369 cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
373 cfs_hash_bd_lock(s->ls_obj_hash, &bd, 1);
374 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
376 list_for_each_entry_safe(h, temp, &bkt->lsb_lru, loh_lru) {
377 LASSERT(atomic_read(&h->loh_ref) == 0);
379 cfs_hash_bd_get(s->ls_obj_hash, &h->loh_fid, &bd2);
380 LASSERT(bd.bd_bucket == bd2.bd_bucket);
382 cfs_hash_bd_del_locked(s->ls_obj_hash,
384 list_move(&h->loh_lru, &dispose);
388 if (nr != ~0 && --nr == 0)
391 if (count > 0 && --count == 0)
395 cfs_hash_bd_unlock(s->ls_obj_hash, &bd, 1);
398 * Free everything on the dispose list. This is safe against
399 * races due to the reasons described in lu_object_put().
401 while (!list_empty(&dispose)) {
402 h = container_of0(dispose.next,
403 struct lu_object_header, loh_lru);
404 list_del_init(&h->loh_lru);
405 lu_object_free(env, lu_object_top(h));
406 lprocfs_counter_incr(s->ls_stats, LU_SS_LRU_PURGED);
412 mutex_unlock(&s->ls_purge_mutex);
414 if (nr != 0 && did_sth && start != 0) {
415 start = 0; /* restart from the first bucket */
418 /* race on s->ls_purge_start, but nobody cares */
419 s->ls_purge_start = i % CFS_HASH_NBKT(s->ls_obj_hash);
423 EXPORT_SYMBOL(lu_site_purge);
428 * Code below has to jump through certain loops to output object description
429 * into libcfs_debug_msg-based log. The problem is that lu_object_print()
430 * composes object description from strings that are parts of _lines_ of
431 * output (i.e., strings that are not terminated by newline). This doesn't fit
432 * very well into libcfs_debug_msg() interface that assumes that each message
433 * supplied to it is a self-contained output line.
435 * To work around this, strings are collected in a temporary buffer
436 * (implemented as a value of lu_cdebug_key key), until terminating newline
437 * character is detected.
445 * XXX overflow is not handled correctly.
450 struct lu_cdebug_data {
454 char lck_area[LU_CDEBUG_LINE];
457 /* context key constructor/destructor: lu_global_key_init, lu_global_key_fini */
458 LU_KEY_INIT_FINI(lu_global, struct lu_cdebug_data);
461 * Key, holding temporary buffer. This key is registered very early by
464 struct lu_context_key lu_global_key = {
465 .lct_tags = LCT_MD_THREAD | LCT_DT_THREAD |
466 LCT_MG_THREAD | LCT_CL_THREAD | LCT_LOCAL,
467 .lct_init = lu_global_key_init,
468 .lct_fini = lu_global_key_fini
472 * Printer function emitting messages through libcfs_debug_msg().
474 int lu_cdebug_printer(const struct lu_env *env,
475 void *cookie, const char *format, ...)
477 struct libcfs_debug_msg_data *msgdata = cookie;
478 struct lu_cdebug_data *key;
483 va_start(args, format);
485 key = lu_context_key_get(&env->le_ctx, &lu_global_key);
486 LASSERT(key != NULL);
488 used = strlen(key->lck_area);
489 complete = format[strlen(format) - 1] == '\n';
491 * Append new chunk to the buffer.
493 vsnprintf(key->lck_area + used,
494 ARRAY_SIZE(key->lck_area) - used, format, args);
496 if (cfs_cdebug_show(msgdata->msg_mask, msgdata->msg_subsys))
497 libcfs_debug_msg(msgdata, "%s\n", key->lck_area);
498 key->lck_area[0] = 0;
503 EXPORT_SYMBOL(lu_cdebug_printer);
506 * Print object header.
508 void lu_object_header_print(const struct lu_env *env, void *cookie,
509 lu_printer_t printer,
510 const struct lu_object_header *hdr)
512 (*printer)(env, cookie, "header@%p[%#lx, %d, "DFID"%s%s%s]",
513 hdr, hdr->loh_flags, atomic_read(&hdr->loh_ref),
515 hlist_unhashed(&hdr->loh_hash) ? "" : " hash",
516 list_empty((struct list_head *)&hdr->loh_lru) ? \
518 hdr->loh_attr & LOHA_EXISTS ? " exist" : "");
520 EXPORT_SYMBOL(lu_object_header_print);
523 * Print human readable representation of the \a o to the \a printer.
525 void lu_object_print(const struct lu_env *env, void *cookie,
526 lu_printer_t printer, const struct lu_object *o)
528 static const char ruler[] = "........................................";
529 struct lu_object_header *top;
533 lu_object_header_print(env, cookie, printer, top);
534 (*printer)(env, cookie, "{\n");
536 list_for_each_entry(o, &top->loh_layers, lo_linkage) {
538 * print `.' \a depth times followed by type name and address
540 (*printer)(env, cookie, "%*.*s%s@%p", depth, depth, ruler,
541 o->lo_dev->ld_type->ldt_name, o);
543 if (o->lo_ops->loo_object_print != NULL)
544 (*o->lo_ops->loo_object_print)(env, cookie, printer, o);
546 (*printer)(env, cookie, "\n");
549 (*printer)(env, cookie, "} header@%p\n", top);
551 EXPORT_SYMBOL(lu_object_print);
554 * Check object consistency.
556 int lu_object_invariant(const struct lu_object *o)
558 struct lu_object_header *top;
561 list_for_each_entry(o, &top->loh_layers, lo_linkage) {
562 if (o->lo_ops->loo_object_invariant != NULL &&
563 !o->lo_ops->loo_object_invariant(o))
568 EXPORT_SYMBOL(lu_object_invariant);
570 static struct lu_object *htable_lookup(struct lu_site *s,
572 const struct lu_fid *f,
573 wait_queue_t *waiter,
576 struct lu_site_bkt_data *bkt;
577 struct lu_object_header *h;
578 struct hlist_node *hnode;
579 __u64 ver = cfs_hash_bd_version_get(bd);
582 return ERR_PTR(-ENOENT);
585 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, bd);
586 /* cfs_hash_bd_peek_locked is a somehow "internal" function
587 * of cfs_hash, it doesn't add refcount on object. */
588 hnode = cfs_hash_bd_peek_locked(s->ls_obj_hash, bd, (void *)f);
590 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS);
591 return ERR_PTR(-ENOENT);
594 h = container_of0(hnode, struct lu_object_header, loh_hash);
595 if (likely(!lu_object_is_dying(h))) {
596 cfs_hash_get(s->ls_obj_hash, hnode);
597 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_HIT);
598 list_del_init(&h->loh_lru);
599 return lu_object_top(h);
603 * Lookup found an object being destroyed this object cannot be
604 * returned (to assure that references to dying objects are eventually
605 * drained), and moreover, lookup has to wait until object is freed.
608 if (likely(waiter != NULL)) {
609 init_waitqueue_entry_current(waiter);
610 add_wait_queue(&bkt->lsb_marche_funebre, waiter);
611 set_current_state(TASK_UNINTERRUPTIBLE);
612 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_DEATH_RACE);
615 return ERR_PTR(-EAGAIN);
618 static struct lu_object *htable_lookup_nowait(struct lu_site *s,
620 const struct lu_fid *f)
622 struct hlist_node *hnode;
623 struct lu_object_header *h;
625 /* cfs_hash_bd_peek_locked is a somehow "internal" function
626 * of cfs_hash, it doesn't add refcount on object. */
627 hnode = cfs_hash_bd_peek_locked(s->ls_obj_hash, bd, (void *)f);
629 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS);
630 return ERR_PTR(-ENOENT);
633 h = container_of0(hnode, struct lu_object_header, loh_hash);
634 if (unlikely(lu_object_is_dying(h)))
635 return ERR_PTR(-ENOENT);
637 cfs_hash_get(s->ls_obj_hash, hnode);
638 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_HIT);
639 list_del_init(&h->loh_lru);
640 return lu_object_top(h);
644 * Search cache for an object with the fid \a f. If such object is found,
645 * return it. Otherwise, create new object, insert it into cache and return
646 * it. In any case, additional reference is acquired on the returned object.
648 struct lu_object *lu_object_find(const struct lu_env *env,
649 struct lu_device *dev, const struct lu_fid *f,
650 const struct lu_object_conf *conf)
652 return lu_object_find_at(env, dev->ld_site->ls_top_dev, f, conf);
654 EXPORT_SYMBOL(lu_object_find);
657 * Limit the lu_object cache to a maximum of lu_cache_nr objects. Because
658 * the calculation for the number of objects to reclaim is not covered by
659 * a lock the maximum number of objects is capped by LU_CACHE_MAX_ADJUST.
660 * This ensures that many concurrent threads will not accidentally purge
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 = cfs_hash_size_get(dev->ld_site->ls_obj_hash);
672 nr = (__u64)lu_cache_nr;
674 lu_site_purge(env, dev->ld_site,
675 MIN(size - nr, LU_CACHE_NR_MAX_ADJUST));
680 static struct lu_object *lu_object_new(const struct lu_env *env,
681 struct lu_device *dev,
682 const struct lu_fid *f,
683 const struct lu_object_conf *conf)
688 struct lu_site_bkt_data *bkt;
690 o = lu_object_alloc(env, dev, f, conf);
691 if (unlikely(IS_ERR(o)))
694 hs = dev->ld_site->ls_obj_hash;
695 cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
696 bkt = cfs_hash_bd_extra_get(hs, &bd);
697 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
699 cfs_hash_bd_unlock(hs, &bd, 1);
701 lu_object_limit(env, dev);
707 * Core logic of lu_object_find*() functions.
709 static struct lu_object *lu_object_find_try(const struct lu_env *env,
710 struct lu_device *dev,
711 const struct lu_fid *f,
712 const struct lu_object_conf *conf,
713 wait_queue_t *waiter)
716 struct lu_object *shadow;
723 * This uses standard index maintenance protocol:
725 * - search index under lock, and return object if found;
726 * - otherwise, unlock index, allocate new object;
727 * - lock index and search again;
728 * - if nothing is found (usual case), insert newly created
730 * - otherwise (race: other thread inserted object), free
731 * object just allocated.
735 * For "LOC_F_NEW" case, we are sure the object is new established.
736 * It is unnecessary to perform lookup-alloc-lookup-insert, instead,
737 * just alloc and insert directly.
739 * If dying object is found during index search, add @waiter to the
740 * site wait-queue and return ERR_PTR(-EAGAIN).
742 if (conf != NULL && conf->loc_flags & LOC_F_NEW)
743 return lu_object_new(env, dev, f, conf);
747 cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
748 o = htable_lookup(s, &bd, f, waiter, &version);
749 cfs_hash_bd_unlock(hs, &bd, 1);
750 if (!IS_ERR(o) || PTR_ERR(o) != -ENOENT)
754 * Allocate new object. This may result in rather complicated
755 * operations, including fld queries, inode loading, etc.
757 o = lu_object_alloc(env, dev, f, conf);
758 if (unlikely(IS_ERR(o)))
761 LASSERT(lu_fid_eq(lu_object_fid(o), f));
763 cfs_hash_bd_lock(hs, &bd, 1);
765 shadow = htable_lookup(s, &bd, f, waiter, &version);
766 if (likely(IS_ERR(shadow) && PTR_ERR(shadow) == -ENOENT)) {
767 struct lu_site_bkt_data *bkt;
769 bkt = cfs_hash_bd_extra_get(hs, &bd);
770 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
772 cfs_hash_bd_unlock(hs, &bd, 1);
774 lu_object_limit(env, dev);
779 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_RACE);
780 cfs_hash_bd_unlock(hs, &bd, 1);
781 lu_object_free(env, o);
786 * Much like lu_object_find(), but top level device of object is specifically
787 * \a dev rather than top level device of the site. This interface allows
788 * objects of different "stacking" to be created within the same site.
790 struct lu_object *lu_object_find_at(const struct lu_env *env,
791 struct lu_device *dev,
792 const struct lu_fid *f,
793 const struct lu_object_conf *conf)
795 struct lu_site_bkt_data *bkt;
796 struct lu_object *obj;
800 if (conf != NULL && conf->loc_flags & LOC_F_NOWAIT) {
801 obj = lu_object_find_try(env, dev, f, conf, NULL);
806 obj = lu_object_find_try(env, dev, f, conf, &wait);
807 if (obj != ERR_PTR(-EAGAIN))
810 * lu_object_find_try() already added waiter into the
813 waitq_wait(&wait, TASK_UNINTERRUPTIBLE);
814 bkt = lu_site_bkt_from_fid(dev->ld_site, (void *)f);
815 remove_wait_queue(&bkt->lsb_marche_funebre, &wait);
818 EXPORT_SYMBOL(lu_object_find_at);
821 * Try to find the object in cache without waiting for the dead object
822 * to be released nor allocating object if no cached one was found.
824 * The found object will be set as LU_OBJECT_HEARD_BANSHEE for purging.
826 void lu_object_purge(const struct lu_env *env, struct lu_device *dev,
827 const struct lu_fid *f)
829 struct lu_site *s = dev->ld_site;
830 cfs_hash_t *hs = s->ls_obj_hash;
834 cfs_hash_bd_get_and_lock(hs, f, &bd, 1);
835 o = htable_lookup_nowait(s, &bd, f);
836 cfs_hash_bd_unlock(hs, &bd, 1);
838 set_bit(LU_OBJECT_HEARD_BANSHEE, &o->lo_header->loh_flags);
839 lu_object_put(env, o);
842 EXPORT_SYMBOL(lu_object_purge);
845 * Find object with given fid, and return its slice belonging to given device.
847 struct lu_object *lu_object_find_slice(const struct lu_env *env,
848 struct lu_device *dev,
849 const struct lu_fid *f,
850 const struct lu_object_conf *conf)
852 struct lu_object *top;
853 struct lu_object *obj;
855 top = lu_object_find(env, dev, f, conf);
859 obj = lu_object_locate(top->lo_header, dev->ld_type);
860 if (unlikely(obj == NULL)) {
861 lu_object_put(env, top);
862 obj = ERR_PTR(-ENOENT);
867 EXPORT_SYMBOL(lu_object_find_slice);
870 * Global list of all device types.
872 static struct list_head lu_device_types;
874 int lu_device_type_init(struct lu_device_type *ldt)
878 atomic_set(&ldt->ldt_device_nr, 0);
879 INIT_LIST_HEAD(&ldt->ldt_linkage);
880 if (ldt->ldt_ops->ldto_init)
881 result = ldt->ldt_ops->ldto_init(ldt);
884 spin_lock(&obd_types_lock);
885 list_add(&ldt->ldt_linkage, &lu_device_types);
886 spin_unlock(&obd_types_lock);
891 EXPORT_SYMBOL(lu_device_type_init);
893 void lu_device_type_fini(struct lu_device_type *ldt)
895 spin_lock(&obd_types_lock);
896 list_del_init(&ldt->ldt_linkage);
897 spin_unlock(&obd_types_lock);
898 if (ldt->ldt_ops->ldto_fini)
899 ldt->ldt_ops->ldto_fini(ldt);
901 EXPORT_SYMBOL(lu_device_type_fini);
904 * Global list of all sites on this node
906 static struct list_head lu_sites;
907 static DEFINE_MUTEX(lu_sites_guard);
910 * Global environment used by site shrinker.
912 static struct lu_env lu_shrink_env;
914 struct lu_site_print_arg {
915 struct lu_env *lsp_env;
917 lu_printer_t lsp_printer;
921 lu_site_obj_print(cfs_hash_t *hs, cfs_hash_bd_t *bd,
922 struct hlist_node *hnode, void *data)
924 struct lu_site_print_arg *arg = (struct lu_site_print_arg *)data;
925 struct lu_object_header *h;
927 h = hlist_entry(hnode, struct lu_object_header, loh_hash);
928 if (!list_empty(&h->loh_layers)) {
929 const struct lu_object *o;
931 o = lu_object_top(h);
932 lu_object_print(arg->lsp_env, arg->lsp_cookie,
933 arg->lsp_printer, o);
935 lu_object_header_print(arg->lsp_env, arg->lsp_cookie,
936 arg->lsp_printer, h);
942 * Print all objects in \a s.
944 void lu_site_print(const struct lu_env *env, struct lu_site *s, void *cookie,
945 lu_printer_t printer)
947 struct lu_site_print_arg arg = {
948 .lsp_env = (struct lu_env *)env,
949 .lsp_cookie = cookie,
950 .lsp_printer = printer,
953 cfs_hash_for_each(s->ls_obj_hash, lu_site_obj_print, &arg);
955 EXPORT_SYMBOL(lu_site_print);
958 * Return desired hash table order.
960 static unsigned long lu_htable_order(struct lu_device *top)
962 unsigned long cache_size;
966 * For ZFS based OSDs the cache should be disabled by default. This
967 * allows the ZFS ARC maximum flexibility in determining what buffers
968 * to cache. If Lustre has objects or buffer which it wants to ensure
969 * always stay cached it must maintain a hold on them.
971 if (strcmp(top->ld_type->ldt_name, LUSTRE_OSD_ZFS_NAME) == 0) {
972 lu_cache_percent = 1;
973 lu_cache_nr = LU_CACHE_NR_ZFS_LIMIT;
974 return LU_SITE_BITS_MIN;
978 * Calculate hash table size, assuming that we want reasonable
979 * performance when 20% of total memory is occupied by cache of
982 * Size of lu_object is (arbitrary) taken as 1K (together with inode).
984 cache_size = totalram_pages;
986 #if BITS_PER_LONG == 32
987 /* limit hashtable size for lowmem systems to low RAM */
988 if (cache_size > 1 << (30 - PAGE_CACHE_SHIFT))
989 cache_size = 1 << (30 - PAGE_CACHE_SHIFT) * 3 / 4;
992 /* clear off unreasonable cache setting. */
993 if (lu_cache_percent == 0 || lu_cache_percent > LU_CACHE_PERCENT_MAX) {
994 CWARN("obdclass: invalid lu_cache_percent: %u, it must be in"
995 " the range of (0, %u]. Will use default value: %u.\n",
996 lu_cache_percent, LU_CACHE_PERCENT_MAX,
997 LU_CACHE_PERCENT_DEFAULT);
999 lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
1001 cache_size = cache_size / 100 * lu_cache_percent *
1002 (PAGE_CACHE_SIZE / 1024);
1004 for (bits = 1; (1 << bits) < cache_size; ++bits) {
1010 static unsigned lu_obj_hop_hash(cfs_hash_t *hs,
1011 const void *key, unsigned mask)
1013 struct lu_fid *fid = (struct lu_fid *)key;
1016 hash = fid_flatten32(fid);
1017 hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
1018 hash = hash_long(hash, hs->hs_bkt_bits);
1020 /* give me another random factor */
1021 hash -= hash_long((unsigned long)hs, fid_oid(fid) % 11 + 3);
1023 hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
1024 hash |= (fid_seq(fid) + fid_oid(fid)) & (CFS_HASH_NBKT(hs) - 1);
1029 static void *lu_obj_hop_object(struct hlist_node *hnode)
1031 return hlist_entry(hnode, struct lu_object_header, loh_hash);
1034 static void *lu_obj_hop_key(struct hlist_node *hnode)
1036 struct lu_object_header *h;
1038 h = hlist_entry(hnode, struct lu_object_header, loh_hash);
1042 static int lu_obj_hop_keycmp(const void *key, struct hlist_node *hnode)
1044 struct lu_object_header *h;
1046 h = hlist_entry(hnode, struct lu_object_header, loh_hash);
1047 return lu_fid_eq(&h->loh_fid, (struct lu_fid *)key);
1050 static void lu_obj_hop_get(cfs_hash_t *hs, struct hlist_node *hnode)
1052 struct lu_object_header *h;
1054 h = hlist_entry(hnode, struct lu_object_header, loh_hash);
1055 if (atomic_add_return(1, &h->loh_ref) == 1) {
1056 struct lu_site_bkt_data *bkt;
1059 cfs_hash_bd_get(hs, &h->loh_fid, &bd);
1060 bkt = cfs_hash_bd_extra_get(hs, &bd);
1065 static void lu_obj_hop_put_locked(cfs_hash_t *hs, struct hlist_node *hnode)
1067 LBUG(); /* we should never called it */
1070 cfs_hash_ops_t lu_site_hash_ops = {
1071 .hs_hash = lu_obj_hop_hash,
1072 .hs_key = lu_obj_hop_key,
1073 .hs_keycmp = lu_obj_hop_keycmp,
1074 .hs_object = lu_obj_hop_object,
1075 .hs_get = lu_obj_hop_get,
1076 .hs_put_locked = lu_obj_hop_put_locked,
1079 void lu_dev_add_linkage(struct lu_site *s, struct lu_device *d)
1081 spin_lock(&s->ls_ld_lock);
1082 if (list_empty(&d->ld_linkage))
1083 list_add(&d->ld_linkage, &s->ls_ld_linkage);
1084 spin_unlock(&s->ls_ld_lock);
1086 EXPORT_SYMBOL(lu_dev_add_linkage);
1088 void lu_dev_del_linkage(struct lu_site *s, struct lu_device *d)
1090 spin_lock(&s->ls_ld_lock);
1091 list_del_init(&d->ld_linkage);
1092 spin_unlock(&s->ls_ld_lock);
1094 EXPORT_SYMBOL(lu_dev_del_linkage);
1097 * Initialize site \a s, with \a d as the top level device.
1099 int lu_site_init(struct lu_site *s, struct lu_device *top)
1101 struct lu_site_bkt_data *bkt;
1108 memset(s, 0, sizeof *s);
1109 mutex_init(&s->ls_purge_mutex);
1110 bits = lu_htable_order(top);
1111 snprintf(name, sizeof(name), "lu_site_%s", top->ld_type->ldt_name);
1112 for (bits = clamp_t(typeof(bits), bits,
1113 LU_SITE_BITS_MIN, LU_SITE_BITS_MAX);
1114 bits >= LU_SITE_BITS_MIN; bits--) {
1115 s->ls_obj_hash = cfs_hash_create(name, bits, bits,
1116 bits - LU_SITE_BKT_BITS,
1119 CFS_HASH_SPIN_BKTLOCK |
1120 CFS_HASH_NO_ITEMREF |
1122 CFS_HASH_ASSERT_EMPTY |
1124 if (s->ls_obj_hash != NULL)
1128 if (s->ls_obj_hash == NULL) {
1129 CERROR("failed to create lu_site hash with bits: %lu\n", bits);
1133 cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
1134 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
1135 INIT_LIST_HEAD(&bkt->lsb_lru);
1136 init_waitqueue_head(&bkt->lsb_marche_funebre);
1139 s->ls_stats = lprocfs_alloc_stats(LU_SS_LAST_STAT, 0);
1140 if (s->ls_stats == NULL) {
1141 cfs_hash_putref(s->ls_obj_hash);
1142 s->ls_obj_hash = NULL;
1146 lprocfs_counter_init(s->ls_stats, LU_SS_CREATED,
1147 0, "created", "created");
1148 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_HIT,
1149 0, "cache_hit", "cache_hit");
1150 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_MISS,
1151 0, "cache_miss", "cache_miss");
1152 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_RACE,
1153 0, "cache_race", "cache_race");
1154 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_DEATH_RACE,
1155 0, "cache_death_race", "cache_death_race");
1156 lprocfs_counter_init(s->ls_stats, LU_SS_LRU_PURGED,
1157 0, "lru_purged", "lru_purged");
1159 INIT_LIST_HEAD(&s->ls_linkage);
1160 s->ls_top_dev = top;
1163 lu_ref_add(&top->ld_reference, "site-top", s);
1165 INIT_LIST_HEAD(&s->ls_ld_linkage);
1166 spin_lock_init(&s->ls_ld_lock);
1168 lu_dev_add_linkage(s, top);
1172 EXPORT_SYMBOL(lu_site_init);
1175 * Finalize \a s and release its resources.
1177 void lu_site_fini(struct lu_site *s)
1179 mutex_lock(&lu_sites_guard);
1180 list_del_init(&s->ls_linkage);
1181 mutex_unlock(&lu_sites_guard);
1183 if (s->ls_obj_hash != NULL) {
1184 cfs_hash_putref(s->ls_obj_hash);
1185 s->ls_obj_hash = NULL;
1188 if (s->ls_top_dev != NULL) {
1189 s->ls_top_dev->ld_site = NULL;
1190 lu_ref_del(&s->ls_top_dev->ld_reference, "site-top", s);
1191 lu_device_put(s->ls_top_dev);
1192 s->ls_top_dev = NULL;
1195 if (s->ls_stats != NULL)
1196 lprocfs_free_stats(&s->ls_stats);
1198 EXPORT_SYMBOL(lu_site_fini);
1201 * Called when initialization of stack for this site is completed.
1203 int lu_site_init_finish(struct lu_site *s)
1206 mutex_lock(&lu_sites_guard);
1207 result = lu_context_refill(&lu_shrink_env.le_ctx);
1209 list_add(&s->ls_linkage, &lu_sites);
1210 mutex_unlock(&lu_sites_guard);
1213 EXPORT_SYMBOL(lu_site_init_finish);
1216 * Acquire additional reference on device \a d
1218 void lu_device_get(struct lu_device *d)
1220 atomic_inc(&d->ld_ref);
1222 EXPORT_SYMBOL(lu_device_get);
1225 * Release reference on device \a d.
1227 void lu_device_put(struct lu_device *d)
1229 LASSERT(atomic_read(&d->ld_ref) > 0);
1230 atomic_dec(&d->ld_ref);
1232 EXPORT_SYMBOL(lu_device_put);
1235 * Initialize device \a d of type \a t.
1237 int lu_device_init(struct lu_device *d, struct lu_device_type *t)
1239 if (atomic_inc_return(&t->ldt_device_nr) == 1 &&
1240 t->ldt_ops->ldto_start != NULL)
1241 t->ldt_ops->ldto_start(t);
1243 memset(d, 0, sizeof *d);
1245 lu_ref_init(&d->ld_reference);
1246 INIT_LIST_HEAD(&d->ld_linkage);
1250 EXPORT_SYMBOL(lu_device_init);
1253 * Finalize device \a d.
1255 void lu_device_fini(struct lu_device *d)
1257 struct lu_device_type *t = d->ld_type;
1259 if (d->ld_obd != NULL) {
1260 d->ld_obd->obd_lu_dev = NULL;
1264 lu_ref_fini(&d->ld_reference);
1265 LASSERTF(atomic_read(&d->ld_ref) == 0,
1266 "Refcount is %u\n", atomic_read(&d->ld_ref));
1267 LASSERT(atomic_read(&t->ldt_device_nr) > 0);
1269 if (atomic_dec_and_test(&t->ldt_device_nr) &&
1270 t->ldt_ops->ldto_stop != NULL)
1271 t->ldt_ops->ldto_stop(t);
1273 EXPORT_SYMBOL(lu_device_fini);
1276 * Initialize object \a o that is part of compound object \a h and was created
1279 int lu_object_init(struct lu_object *o, struct lu_object_header *h,
1280 struct lu_device *d)
1282 memset(o, 0, sizeof(*o));
1286 lu_ref_add_at(&d->ld_reference, &o->lo_dev_ref, "lu_object", o);
1287 INIT_LIST_HEAD(&o->lo_linkage);
1291 EXPORT_SYMBOL(lu_object_init);
1294 * Finalize object and release its resources.
1296 void lu_object_fini(struct lu_object *o)
1298 struct lu_device *dev = o->lo_dev;
1300 LASSERT(list_empty(&o->lo_linkage));
1303 lu_ref_del_at(&dev->ld_reference, &o->lo_dev_ref,
1309 EXPORT_SYMBOL(lu_object_fini);
1312 * Add object \a o as first layer of compound object \a h
1314 * This is typically called by the ->ldo_object_alloc() method of top-level
1317 void lu_object_add_top(struct lu_object_header *h, struct lu_object *o)
1319 list_move(&o->lo_linkage, &h->loh_layers);
1321 EXPORT_SYMBOL(lu_object_add_top);
1324 * Add object \a o as a layer of compound object, going after \a before.
1326 * This is typically called by the ->ldo_object_alloc() method of \a
1329 void lu_object_add(struct lu_object *before, struct lu_object *o)
1331 list_move(&o->lo_linkage, &before->lo_linkage);
1333 EXPORT_SYMBOL(lu_object_add);
1336 * Initialize compound object.
1338 int lu_object_header_init(struct lu_object_header *h)
1340 memset(h, 0, sizeof *h);
1341 atomic_set(&h->loh_ref, 1);
1342 INIT_HLIST_NODE(&h->loh_hash);
1343 INIT_LIST_HEAD(&h->loh_lru);
1344 INIT_LIST_HEAD(&h->loh_layers);
1345 lu_ref_init(&h->loh_reference);
1348 EXPORT_SYMBOL(lu_object_header_init);
1351 * Finalize compound object.
1353 void lu_object_header_fini(struct lu_object_header *h)
1355 LASSERT(list_empty(&h->loh_layers));
1356 LASSERT(list_empty(&h->loh_lru));
1357 LASSERT(hlist_unhashed(&h->loh_hash));
1358 lu_ref_fini(&h->loh_reference);
1360 EXPORT_SYMBOL(lu_object_header_fini);
1363 * Given a compound object, find its slice, corresponding to the device type
1366 struct lu_object *lu_object_locate(struct lu_object_header *h,
1367 const struct lu_device_type *dtype)
1369 struct lu_object *o;
1371 list_for_each_entry(o, &h->loh_layers, lo_linkage) {
1372 if (o->lo_dev->ld_type == dtype)
1377 EXPORT_SYMBOL(lu_object_locate);
1380 * Finalize and free devices in the device stack.
1382 * Finalize device stack by purging object cache, and calling
1383 * lu_device_type_operations::ldto_device_fini() and
1384 * lu_device_type_operations::ldto_device_free() on all devices in the stack.
1386 void lu_stack_fini(const struct lu_env *env, struct lu_device *top)
1388 struct lu_site *site = top->ld_site;
1389 struct lu_device *scan;
1390 struct lu_device *next;
1392 lu_site_purge(env, site, ~0);
1393 for (scan = top; scan != NULL; scan = next) {
1394 next = scan->ld_type->ldt_ops->ldto_device_fini(env, scan);
1395 lu_ref_del(&scan->ld_reference, "lu-stack", &lu_site_init);
1396 lu_device_put(scan);
1400 lu_site_purge(env, site, ~0);
1402 for (scan = top; scan != NULL; scan = next) {
1403 const struct lu_device_type *ldt = scan->ld_type;
1404 struct obd_type *type;
1406 next = ldt->ldt_ops->ldto_device_free(env, scan);
1407 type = ldt->ldt_obd_type;
1410 class_put_type(type);
1414 EXPORT_SYMBOL(lu_stack_fini);
1418 * Maximal number of tld slots.
1420 LU_CONTEXT_KEY_NR = 40
1423 static struct lu_context_key *lu_keys[LU_CONTEXT_KEY_NR] = { NULL, };
1425 static DEFINE_SPINLOCK(lu_keys_guard);
1428 * Global counter incremented whenever key is registered, unregistered,
1429 * revived or quiesced. This is used to void unnecessary calls to
1430 * lu_context_refill(). No locking is provided, as initialization and shutdown
1431 * are supposed to be externally serialized.
1433 static unsigned key_set_version = 0;
1438 int lu_context_key_register(struct lu_context_key *key)
1443 LASSERT(key->lct_init != NULL);
1444 LASSERT(key->lct_fini != NULL);
1445 LASSERT(key->lct_tags != 0);
1446 LASSERT(key->lct_owner != NULL);
1449 spin_lock(&lu_keys_guard);
1450 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1451 if (lu_keys[i] == NULL) {
1453 atomic_set(&key->lct_used, 1);
1455 lu_ref_init(&key->lct_reference);
1461 spin_unlock(&lu_keys_guard);
1464 EXPORT_SYMBOL(lu_context_key_register);
1466 static void key_fini(struct lu_context *ctx, int index)
1468 if (ctx->lc_value != NULL && ctx->lc_value[index] != NULL) {
1469 struct lu_context_key *key;
1471 key = lu_keys[index];
1472 LASSERT(key != NULL);
1473 LASSERT(key->lct_fini != NULL);
1474 LASSERT(atomic_read(&key->lct_used) > 1);
1476 key->lct_fini(ctx, key, ctx->lc_value[index]);
1477 lu_ref_del(&key->lct_reference, "ctx", ctx);
1478 atomic_dec(&key->lct_used);
1480 LASSERT(key->lct_owner != NULL);
1481 if ((ctx->lc_tags & LCT_NOREF) == 0) {
1482 LINVRNT(module_refcount(key->lct_owner) > 0);
1483 module_put(key->lct_owner);
1485 ctx->lc_value[index] = NULL;
1492 void lu_context_key_degister(struct lu_context_key *key)
1494 LASSERT(atomic_read(&key->lct_used) >= 1);
1495 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1497 lu_context_key_quiesce(key);
1500 spin_lock(&lu_keys_guard);
1501 key_fini(&lu_shrink_env.le_ctx, key->lct_index);
1502 if (lu_keys[key->lct_index]) {
1503 lu_keys[key->lct_index] = NULL;
1504 lu_ref_fini(&key->lct_reference);
1506 spin_unlock(&lu_keys_guard);
1508 LASSERTF(atomic_read(&key->lct_used) == 1,
1509 "key has instances: %d\n",
1510 atomic_read(&key->lct_used));
1512 EXPORT_SYMBOL(lu_context_key_degister);
1515 * Register a number of keys. This has to be called after all keys have been
1516 * initialized by a call to LU_CONTEXT_KEY_INIT().
1518 int lu_context_key_register_many(struct lu_context_key *k, ...)
1520 struct lu_context_key *key = k;
1526 result = lu_context_key_register(key);
1529 key = va_arg(args, struct lu_context_key *);
1530 } while (key != NULL);
1536 lu_context_key_degister(k);
1537 k = va_arg(args, struct lu_context_key *);
1544 EXPORT_SYMBOL(lu_context_key_register_many);
1547 * De-register a number of keys. This is a dual to
1548 * lu_context_key_register_many().
1550 void lu_context_key_degister_many(struct lu_context_key *k, ...)
1556 lu_context_key_degister(k);
1557 k = va_arg(args, struct lu_context_key*);
1558 } while (k != NULL);
1561 EXPORT_SYMBOL(lu_context_key_degister_many);
1564 * Revive a number of keys.
1566 void lu_context_key_revive_many(struct lu_context_key *k, ...)
1572 lu_context_key_revive(k);
1573 k = va_arg(args, struct lu_context_key*);
1574 } while (k != NULL);
1577 EXPORT_SYMBOL(lu_context_key_revive_many);
1580 * Quiescent a number of keys.
1582 void lu_context_key_quiesce_many(struct lu_context_key *k, ...)
1588 lu_context_key_quiesce(k);
1589 k = va_arg(args, struct lu_context_key*);
1590 } while (k != NULL);
1593 EXPORT_SYMBOL(lu_context_key_quiesce_many);
1596 * Return value associated with key \a key in context \a ctx.
1598 void *lu_context_key_get(const struct lu_context *ctx,
1599 const struct lu_context_key *key)
1601 LINVRNT(ctx->lc_state == LCS_ENTERED);
1602 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1603 LASSERT(lu_keys[key->lct_index] == key);
1604 return ctx->lc_value[key->lct_index];
1606 EXPORT_SYMBOL(lu_context_key_get);
1609 * List of remembered contexts. XXX document me.
1611 static struct list_head lu_context_remembered;
1614 * Destroy \a key in all remembered contexts. This is used to destroy key
1615 * values in "shared" contexts (like service threads), when a module owning
1616 * the key is about to be unloaded.
1618 void lu_context_key_quiesce(struct lu_context_key *key)
1620 struct lu_context *ctx;
1621 extern unsigned cl_env_cache_purge(unsigned nr);
1623 if (!(key->lct_tags & LCT_QUIESCENT)) {
1625 * XXX layering violation.
1627 cl_env_cache_purge(~0);
1628 key->lct_tags |= LCT_QUIESCENT;
1630 * XXX memory barrier has to go here.
1632 spin_lock(&lu_keys_guard);
1633 list_for_each_entry(ctx, &lu_context_remembered,
1635 key_fini(ctx, key->lct_index);
1636 spin_unlock(&lu_keys_guard);
1640 EXPORT_SYMBOL(lu_context_key_quiesce);
1642 void lu_context_key_revive(struct lu_context_key *key)
1644 key->lct_tags &= ~LCT_QUIESCENT;
1647 EXPORT_SYMBOL(lu_context_key_revive);
1649 static void keys_fini(struct lu_context *ctx)
1653 if (ctx->lc_value == NULL)
1656 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i)
1659 OBD_FREE(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1660 ctx->lc_value = NULL;
1663 static int keys_fill(struct lu_context *ctx)
1667 LINVRNT(ctx->lc_value != NULL);
1668 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1669 struct lu_context_key *key;
1672 if (ctx->lc_value[i] == NULL && key != NULL &&
1673 (key->lct_tags & ctx->lc_tags) &&
1675 * Don't create values for a LCT_QUIESCENT key, as this
1676 * will pin module owning a key.
1678 !(key->lct_tags & LCT_QUIESCENT)) {
1681 LINVRNT(key->lct_init != NULL);
1682 LINVRNT(key->lct_index == i);
1684 value = key->lct_init(ctx, key);
1685 if (unlikely(IS_ERR(value)))
1686 return PTR_ERR(value);
1688 LASSERT(key->lct_owner != NULL);
1689 if (!(ctx->lc_tags & LCT_NOREF))
1690 try_module_get(key->lct_owner);
1691 lu_ref_add_atomic(&key->lct_reference, "ctx", ctx);
1692 atomic_inc(&key->lct_used);
1694 * This is the only place in the code, where an
1695 * element of ctx->lc_value[] array is set to non-NULL
1698 ctx->lc_value[i] = value;
1699 if (key->lct_exit != NULL)
1700 ctx->lc_tags |= LCT_HAS_EXIT;
1702 ctx->lc_version = key_set_version;
1707 static int keys_init(struct lu_context *ctx)
1709 OBD_ALLOC(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1710 if (likely(ctx->lc_value != NULL))
1711 return keys_fill(ctx);
1717 * Initialize context data-structure. Create values for all keys.
1719 int lu_context_init(struct lu_context *ctx, __u32 tags)
1723 memset(ctx, 0, sizeof *ctx);
1724 ctx->lc_state = LCS_INITIALIZED;
1725 ctx->lc_tags = tags;
1726 if (tags & LCT_REMEMBER) {
1727 spin_lock(&lu_keys_guard);
1728 list_add(&ctx->lc_remember, &lu_context_remembered);
1729 spin_unlock(&lu_keys_guard);
1731 INIT_LIST_HEAD(&ctx->lc_remember);
1734 rc = keys_init(ctx);
1736 lu_context_fini(ctx);
1740 EXPORT_SYMBOL(lu_context_init);
1743 * Finalize context data-structure. Destroy key values.
1745 void lu_context_fini(struct lu_context *ctx)
1747 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1748 ctx->lc_state = LCS_FINALIZED;
1750 if ((ctx->lc_tags & LCT_REMEMBER) == 0) {
1751 LASSERT(list_empty(&ctx->lc_remember));
1754 } else { /* could race with key degister */
1755 spin_lock(&lu_keys_guard);
1757 list_del_init(&ctx->lc_remember);
1758 spin_unlock(&lu_keys_guard);
1761 EXPORT_SYMBOL(lu_context_fini);
1764 * Called before entering context.
1766 void lu_context_enter(struct lu_context *ctx)
1768 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1769 ctx->lc_state = LCS_ENTERED;
1771 EXPORT_SYMBOL(lu_context_enter);
1774 * Called after exiting from \a ctx
1776 void lu_context_exit(struct lu_context *ctx)
1780 LINVRNT(ctx->lc_state == LCS_ENTERED);
1781 ctx->lc_state = LCS_LEFT;
1782 if (ctx->lc_tags & LCT_HAS_EXIT && ctx->lc_value != NULL) {
1783 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1784 if (ctx->lc_value[i] != NULL) {
1785 struct lu_context_key *key;
1788 LASSERT(key != NULL);
1789 if (key->lct_exit != NULL)
1791 key, ctx->lc_value[i]);
1796 EXPORT_SYMBOL(lu_context_exit);
1799 * Allocate for context all missing keys that were registered after context
1800 * creation. key_set_version is only changed in rare cases when modules
1801 * are loaded and removed.
1803 int lu_context_refill(struct lu_context *ctx)
1805 return likely(ctx->lc_version == key_set_version) ? 0 : keys_fill(ctx);
1807 EXPORT_SYMBOL(lu_context_refill);
1810 * lu_ctx_tags/lu_ses_tags will be updated if there are new types of
1811 * obd being added. Currently, this is only used on client side, specifically
1812 * for echo device client, for other stack (like ptlrpc threads), context are
1813 * predefined when the lu_device type are registered, during the module probe
1816 __u32 lu_context_tags_default = 0;
1817 __u32 lu_session_tags_default = 0;
1819 void lu_context_tags_update(__u32 tags)
1821 spin_lock(&lu_keys_guard);
1822 lu_context_tags_default |= tags;
1824 spin_unlock(&lu_keys_guard);
1826 EXPORT_SYMBOL(lu_context_tags_update);
1828 void lu_context_tags_clear(__u32 tags)
1830 spin_lock(&lu_keys_guard);
1831 lu_context_tags_default &= ~tags;
1833 spin_unlock(&lu_keys_guard);
1835 EXPORT_SYMBOL(lu_context_tags_clear);
1837 void lu_session_tags_update(__u32 tags)
1839 spin_lock(&lu_keys_guard);
1840 lu_session_tags_default |= tags;
1842 spin_unlock(&lu_keys_guard);
1844 EXPORT_SYMBOL(lu_session_tags_update);
1846 void lu_session_tags_clear(__u32 tags)
1848 spin_lock(&lu_keys_guard);
1849 lu_session_tags_default &= ~tags;
1851 spin_unlock(&lu_keys_guard);
1853 EXPORT_SYMBOL(lu_session_tags_clear);
1855 int lu_env_init(struct lu_env *env, __u32 tags)
1860 result = lu_context_init(&env->le_ctx, tags);
1861 if (likely(result == 0))
1862 lu_context_enter(&env->le_ctx);
1865 EXPORT_SYMBOL(lu_env_init);
1867 void lu_env_fini(struct lu_env *env)
1869 lu_context_exit(&env->le_ctx);
1870 lu_context_fini(&env->le_ctx);
1873 EXPORT_SYMBOL(lu_env_fini);
1875 int lu_env_refill(struct lu_env *env)
1879 result = lu_context_refill(&env->le_ctx);
1880 if (result == 0 && env->le_ses != NULL)
1881 result = lu_context_refill(env->le_ses);
1884 EXPORT_SYMBOL(lu_env_refill);
1887 * Currently, this API will only be used by echo client.
1888 * Because echo client and normal lustre client will share
1889 * same cl_env cache. So echo client needs to refresh
1890 * the env context after it get one from the cache, especially
1891 * when normal client and echo client co-exist in the same client.
1893 int lu_env_refill_by_tags(struct lu_env *env, __u32 ctags,
1898 if ((env->le_ctx.lc_tags & ctags) != ctags) {
1899 env->le_ctx.lc_version = 0;
1900 env->le_ctx.lc_tags |= ctags;
1903 if (env->le_ses && (env->le_ses->lc_tags & stags) != stags) {
1904 env->le_ses->lc_version = 0;
1905 env->le_ses->lc_tags |= stags;
1908 result = lu_env_refill(env);
1912 EXPORT_SYMBOL(lu_env_refill_by_tags);
1914 static struct shrinker *lu_site_shrinker;
1916 typedef struct lu_site_stats{
1917 unsigned lss_populated;
1918 unsigned lss_max_search;
1923 static void lu_site_stats_get(cfs_hash_t *hs,
1924 lu_site_stats_t *stats, int populated)
1929 cfs_hash_for_each_bucket(hs, &bd, i) {
1930 struct lu_site_bkt_data *bkt = cfs_hash_bd_extra_get(hs, &bd);
1931 struct hlist_head *hhead;
1933 cfs_hash_bd_lock(hs, &bd, 1);
1934 stats->lss_busy += bkt->lsb_busy;
1935 stats->lss_total += cfs_hash_bd_count_get(&bd);
1936 stats->lss_max_search = max((int)stats->lss_max_search,
1937 cfs_hash_bd_depmax_get(&bd));
1939 cfs_hash_bd_unlock(hs, &bd, 1);
1943 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
1944 if (!hlist_empty(hhead))
1945 stats->lss_populated++;
1947 cfs_hash_bd_unlock(hs, &bd, 1);
1952 static unsigned long lu_cache_shrink_count(struct shrinker *sk,
1953 struct shrink_control *sc)
1955 lu_site_stats_t stats;
1957 struct lu_site *tmp;
1958 unsigned long cached = 0;
1960 if (!(sc->gfp_mask & __GFP_FS))
1963 mutex_lock(&lu_sites_guard);
1964 list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
1965 memset(&stats, 0, sizeof(stats));
1966 lu_site_stats_get(s->ls_obj_hash, &stats, 0);
1967 cached += stats.lss_total - stats.lss_busy;
1969 mutex_unlock(&lu_sites_guard);
1971 cached = (cached / 100) * sysctl_vfs_cache_pressure;
1972 CDEBUG(D_INODE, "%ld objects cached\n", cached);
1976 static unsigned long lu_cache_shrink_scan(struct shrinker *sk,
1977 struct shrink_control *sc)
1980 struct lu_site *tmp;
1981 unsigned long remain = sc->nr_to_scan;
1984 if (!(sc->gfp_mask & __GFP_FS))
1985 /* We must not take the lu_sites_guard lock when
1986 * __GFP_FS is *not* set because of the deadlock
1987 * possibility detailed above. Additionally,
1988 * since we cannot determine the number of
1989 * objects in the cache without taking this
1990 * lock, we're in a particularly tough spot. As
1991 * a result, we'll just lie and say our cache is
1992 * empty. This _should_ be ok, as we can't
1993 * reclaim objects when __GFP_FS is *not* set
1998 mutex_lock(&lu_sites_guard);
1999 list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
2000 remain = lu_site_purge(&lu_shrink_env, s, remain);
2002 * Move just shrunk site to the tail of site list to
2003 * assure shrinking fairness.
2005 list_move_tail(&s->ls_linkage, &splice);
2007 list_splice(&splice, lu_sites.prev);
2008 mutex_unlock(&lu_sites_guard);
2010 return sc->nr_to_scan - remain;
2013 #ifndef HAVE_SHRINKER_COUNT
2015 * There exists a potential lock inversion deadlock scenario when using
2016 * Lustre on top of ZFS. This occurs between one of ZFS's
2017 * buf_hash_table.ht_lock's, and Lustre's lu_sites_guard lock. Essentially,
2018 * thread A will take the lu_sites_guard lock and sleep on the ht_lock,
2019 * while thread B will take the ht_lock and sleep on the lu_sites_guard
2020 * lock. Obviously neither thread will wake and drop their respective hold
2023 * To prevent this from happening we must ensure the lu_sites_guard lock is
2024 * not taken while down this code path. ZFS reliably does not set the
2025 * __GFP_FS bit in its code paths, so this can be used to determine if it
2026 * is safe to take the lu_sites_guard lock.
2028 * Ideally we should accurately return the remaining number of cached
2029 * objects without taking the lu_sites_guard lock, but this is not
2030 * possible in the current implementation.
2032 static int lu_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
2035 struct shrink_control scv = {
2036 .nr_to_scan = shrink_param(sc, nr_to_scan),
2037 .gfp_mask = shrink_param(sc, gfp_mask)
2039 #if !defined(HAVE_SHRINKER_WANT_SHRINK_PTR) && !defined(HAVE_SHRINK_CONTROL)
2040 struct shrinker* shrinker = NULL;
2044 CDEBUG(D_INODE, "Shrink %lu objects\n", scv.nr_to_scan);
2046 lu_cache_shrink_scan(shrinker, &scv);
2048 cached = lu_cache_shrink_count(shrinker, &scv);
2049 if (scv.nr_to_scan == 0)
2050 CDEBUG(D_INODE, "%d objects cached\n", cached);
2054 #endif /* HAVE_SHRINKER_COUNT */
2062 * Environment to be used in debugger, contains all tags.
2064 struct lu_env lu_debugging_env;
2067 * Debugging printer function using printk().
2069 int lu_printk_printer(const struct lu_env *env,
2070 void *unused, const char *format, ...)
2074 va_start(args, format);
2075 vprintk(format, args);
2080 int lu_debugging_setup(void)
2082 return lu_env_init(&lu_debugging_env, ~0);
2085 void lu_context_keys_dump(void)
2089 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
2090 struct lu_context_key *key;
2094 CERROR("[%d]: %p %x (%p,%p,%p) %d %d \"%s\"@%p\n",
2095 i, key, key->lct_tags,
2096 key->lct_init, key->lct_fini, key->lct_exit,
2097 key->lct_index, atomic_read(&key->lct_used),
2098 key->lct_owner ? key->lct_owner->name : "",
2100 lu_ref_print(&key->lct_reference);
2104 EXPORT_SYMBOL(lu_context_keys_dump);
2107 * Initialization of global lu_* data.
2109 int lu_global_init(void)
2112 DEF_SHRINKER_VAR(shvar, lu_cache_shrink,
2113 lu_cache_shrink_count, lu_cache_shrink_scan);
2115 CDEBUG(D_INFO, "Lustre LU module (%p).\n", &lu_keys);
2117 INIT_LIST_HEAD(&lu_device_types);
2118 INIT_LIST_HEAD(&lu_context_remembered);
2119 INIT_LIST_HEAD(&lu_sites);
2121 result = lu_ref_global_init();
2125 LU_CONTEXT_KEY_INIT(&lu_global_key);
2126 result = lu_context_key_register(&lu_global_key);
2131 * At this level, we don't know what tags are needed, so allocate them
2132 * conservatively. This should not be too bad, because this
2133 * environment is global.
2135 mutex_lock(&lu_sites_guard);
2136 result = lu_env_init(&lu_shrink_env, LCT_SHRINKER);
2137 mutex_unlock(&lu_sites_guard);
2142 * seeks estimation: 3 seeks to read a record from oi, one to read
2143 * inode, one for ea. Unfortunately setting this high value results in
2144 * lu_object/inode cache consuming all the memory.
2146 lu_site_shrinker = set_shrinker(DEFAULT_SEEKS, &shvar);
2147 if (lu_site_shrinker == NULL)
2154 * Dual to lu_global_init().
2156 void lu_global_fini(void)
2158 if (lu_site_shrinker != NULL) {
2159 remove_shrinker(lu_site_shrinker);
2160 lu_site_shrinker = NULL;
2163 lu_context_key_degister(&lu_global_key);
2166 * Tear shrinker environment down _after_ de-registering
2167 * lu_global_key, because the latter has a value in the former.
2169 mutex_lock(&lu_sites_guard);
2170 lu_env_fini(&lu_shrink_env);
2171 mutex_unlock(&lu_sites_guard);
2173 lu_ref_global_fini();
2176 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx)
2179 struct lprocfs_counter ret;
2181 lprocfs_stats_collect(stats, idx, &ret);
2182 return (__u32)ret.lc_count;
2189 * Output site statistical counters into a buffer. Suitable for
2190 * lprocfs_rd_*()-style functions.
2192 int lu_site_stats_seq_print(const struct lu_site *s, struct seq_file *m)
2194 lu_site_stats_t stats;
2196 memset(&stats, 0, sizeof(stats));
2197 lu_site_stats_get(s->ls_obj_hash, &stats, 1);
2199 return seq_printf(m, "%d/%d %d/%d %d %d %d %d %d %d %d\n",
2202 stats.lss_populated,
2203 CFS_HASH_NHLIST(s->ls_obj_hash),
2204 stats.lss_max_search,
2205 ls_stats_read(s->ls_stats, LU_SS_CREATED),
2206 ls_stats_read(s->ls_stats, LU_SS_CACHE_HIT),
2207 ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
2208 ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
2209 ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
2210 ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED));
2212 EXPORT_SYMBOL(lu_site_stats_seq_print);
2214 int lu_site_stats_print(const struct lu_site *s, char *page, int count)
2216 lu_site_stats_t stats;
2218 memset(&stats, 0, sizeof(stats));
2219 lu_site_stats_get(s->ls_obj_hash, &stats, 1);
2221 return snprintf(page, count, "%d/%d %d/%d %d %d %d %d %d %d %d\n",
2224 stats.lss_populated,
2225 CFS_HASH_NHLIST(s->ls_obj_hash),
2226 stats.lss_max_search,
2227 ls_stats_read(s->ls_stats, LU_SS_CREATED),
2228 ls_stats_read(s->ls_stats, LU_SS_CACHE_HIT),
2229 ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
2230 ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
2231 ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
2232 ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED));
2234 EXPORT_SYMBOL(lu_site_stats_print);
2237 * Helper function to initialize a number of kmem slab caches at once.
2239 int lu_kmem_init(struct lu_kmem_descr *caches)
2242 struct lu_kmem_descr *iter = caches;
2244 for (result = 0; iter->ckd_cache != NULL; ++iter) {
2245 *iter->ckd_cache = kmem_cache_create(iter->ckd_name,
2248 if (*iter->ckd_cache == NULL) {
2250 /* free all previously allocated caches */
2251 lu_kmem_fini(caches);
2257 EXPORT_SYMBOL(lu_kmem_init);
2260 * Helper function to finalize a number of kmem slab cached at once. Dual to
2263 void lu_kmem_fini(struct lu_kmem_descr *caches)
2265 for (; caches->ckd_cache != NULL; ++caches) {
2266 if (*caches->ckd_cache != NULL) {
2267 kmem_cache_destroy(*caches->ckd_cache);
2268 *caches->ckd_cache = NULL;
2272 EXPORT_SYMBOL(lu_kmem_fini);
2275 * Temporary solution to be able to assign fid in ->do_create()
2276 * till we have fully-functional OST fids
2278 void lu_object_assign_fid(const struct lu_env *env, struct lu_object *o,
2279 const struct lu_fid *fid)
2281 struct lu_site *s = o->lo_dev->ld_site;
2282 struct lu_fid *old = &o->lo_header->loh_fid;
2283 struct lu_site_bkt_data *bkt;
2284 struct lu_object *shadow;
2285 wait_queue_t waiter;
2290 LASSERT(fid_is_zero(old));
2292 hs = s->ls_obj_hash;
2293 cfs_hash_bd_get_and_lock(hs, (void *)fid, &bd, 1);
2294 shadow = htable_lookup(s, &bd, fid, &waiter, &version);
2295 /* supposed to be unique */
2296 LASSERT(IS_ERR(shadow) && PTR_ERR(shadow) == -ENOENT);
2298 bkt = cfs_hash_bd_extra_get(hs, &bd);
2299 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
2301 cfs_hash_bd_unlock(hs, &bd, 1);
2303 EXPORT_SYMBOL(lu_object_assign_fid);
2306 * allocates object with 0 (non-assiged) fid
2307 * XXX: temporary solution to be able to assign fid in ->do_create()
2308 * till we have fully-functional OST fids
2310 struct lu_object *lu_object_anon(const struct lu_env *env,
2311 struct lu_device *dev,
2312 const struct lu_object_conf *conf)
2315 struct lu_object *o;
2318 o = lu_object_alloc(env, dev, &fid, conf);
2322 EXPORT_SYMBOL(lu_object_anon);
2324 struct lu_buf LU_BUF_NULL = {
2328 EXPORT_SYMBOL(LU_BUF_NULL);
2330 void lu_buf_free(struct lu_buf *buf)
2334 LASSERT(buf->lb_len > 0);
2335 OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);
2340 EXPORT_SYMBOL(lu_buf_free);
2342 void lu_buf_alloc(struct lu_buf *buf, size_t size)
2345 LASSERT(buf->lb_buf == NULL);
2346 LASSERT(buf->lb_len == 0);
2347 OBD_ALLOC_LARGE(buf->lb_buf, size);
2348 if (likely(buf->lb_buf))
2351 EXPORT_SYMBOL(lu_buf_alloc);
2353 void lu_buf_realloc(struct lu_buf *buf, size_t size)
2356 lu_buf_alloc(buf, size);
2358 EXPORT_SYMBOL(lu_buf_realloc);
2360 struct lu_buf *lu_buf_check_and_alloc(struct lu_buf *buf, size_t len)
2362 if (buf->lb_buf == NULL && buf->lb_len == 0)
2363 lu_buf_alloc(buf, len);
2365 if ((len > buf->lb_len) && (buf->lb_buf != NULL))
2366 lu_buf_realloc(buf, len);
2370 EXPORT_SYMBOL(lu_buf_check_and_alloc);
2373 * Increase the size of the \a buf.
2374 * preserves old data in buffer
2375 * old buffer remains unchanged on error
2376 * \retval 0 or -ENOMEM
2378 int lu_buf_check_and_grow(struct lu_buf *buf, size_t len)
2382 if (len <= buf->lb_len)
2385 OBD_ALLOC_LARGE(ptr, len);
2389 /* Free the old buf */
2390 if (buf->lb_buf != NULL) {
2391 memcpy(ptr, buf->lb_buf, buf->lb_len);
2392 OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);
2399 EXPORT_SYMBOL(lu_buf_check_and_grow);