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, 2012, 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 <linux/libcfs/libcfs.h>
49 # include <linux/module.h>
52 #include <linux/libcfs/libcfs_hash.h>
53 #include <obd_class.h>
54 #include <obd_support.h>
55 #include <lustre_disk.h>
56 #include <lustre_fid.h>
57 #include <lu_object.h>
59 #include <linux/list.h>
61 static void lu_object_free(const struct lu_env *env, struct lu_object *o);
64 * Decrease reference counter on object. If last reference is freed, return
65 * object to the cache, unless lu_object_is_dying(o) holds. In the latter
66 * case, free object immediately.
68 void lu_object_put(const struct lu_env *env, struct lu_object *o)
70 struct lu_site_bkt_data *bkt;
71 struct lu_object_header *top;
73 struct lu_object *orig;
75 const struct lu_fid *fid;
78 site = o->lo_dev->ld_site;
82 * till we have full fids-on-OST implemented anonymous objects
83 * are possible in OSP. such an object isn't listed in the site
84 * so we should not remove it from the site.
86 fid = lu_object_fid(o);
87 if (fid_is_zero(fid)) {
88 LASSERT(top->loh_hash.next == NULL
89 && top->loh_hash.pprev == NULL);
90 LASSERT(list_empty(&top->loh_lru));
91 if (!atomic_dec_and_test(&top->loh_ref))
93 list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
94 if (o->lo_ops->loo_object_release != NULL)
95 o->lo_ops->loo_object_release(env, o);
97 lu_object_free(env, orig);
101 cfs_hash_bd_get(site->ls_obj_hash, &top->loh_fid, &bd);
102 bkt = cfs_hash_bd_extra_get(site->ls_obj_hash, &bd);
104 if (!cfs_hash_bd_dec_and_lock(site->ls_obj_hash, &bd, &top->loh_ref)) {
105 if (lu_object_is_dying(top)) {
108 * somebody may be waiting for this, currently only
109 * used for cl_object, see cl_object_put_last().
111 wake_up_all(&bkt->lsb_marche_funebre);
116 LASSERT(bkt->lsb_busy > 0);
119 * When last reference is released, iterate over object
120 * layers, and notify them that object is no longer busy.
122 list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
123 if (o->lo_ops->loo_object_release != NULL)
124 o->lo_ops->loo_object_release(env, o);
127 if (!lu_object_is_dying(top)) {
128 LASSERT(list_empty(&top->loh_lru));
129 list_add_tail(&top->loh_lru, &bkt->lsb_lru);
130 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
135 * If object is dying (will not be cached), removed it
136 * from hash table and LRU.
138 * This is done with hash table and LRU lists locked. As the only
139 * way to acquire first reference to previously unreferenced
140 * object is through hash-table lookup (lu_object_find()),
141 * or LRU scanning (lu_site_purge()), that are done under hash-table
142 * and LRU lock, no race with concurrent object lookup is possible
143 * and we can safely destroy object below.
145 if (!test_and_set_bit(LU_OBJECT_UNHASHED, &top->loh_flags))
146 cfs_hash_bd_del_locked(site->ls_obj_hash, &bd, &top->loh_hash);
147 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
149 * Object was already removed from hash and lru above, can
152 lu_object_free(env, orig);
154 EXPORT_SYMBOL(lu_object_put);
157 * Put object and don't keep in cache. This is temporary solution for
158 * multi-site objects when its layering is not constant.
160 void lu_object_put_nocache(const struct lu_env *env, struct lu_object *o)
162 set_bit(LU_OBJECT_HEARD_BANSHEE, &o->lo_header->loh_flags);
163 return lu_object_put(env, o);
165 EXPORT_SYMBOL(lu_object_put_nocache);
168 * Kill the object and take it out of LRU cache.
169 * Currently used by client code for layout change.
171 void lu_object_unhash(const struct lu_env *env, struct lu_object *o)
173 struct lu_object_header *top;
176 set_bit(LU_OBJECT_HEARD_BANSHEE, &top->loh_flags);
177 if (!test_and_set_bit(LU_OBJECT_UNHASHED, &top->loh_flags)) {
178 cfs_hash_t *obj_hash = o->lo_dev->ld_site->ls_obj_hash;
181 cfs_hash_bd_get_and_lock(obj_hash, &top->loh_fid, &bd, 1);
182 list_del_init(&top->loh_lru);
183 cfs_hash_bd_del_locked(obj_hash, &bd, &top->loh_hash);
184 cfs_hash_bd_unlock(obj_hash, &bd, 1);
187 EXPORT_SYMBOL(lu_object_unhash);
190 * Allocate new object.
192 * This follows object creation protocol, described in the comment within
193 * struct lu_device_operations definition.
195 static struct lu_object *lu_object_alloc(const struct lu_env *env,
196 struct lu_device *dev,
197 const struct lu_fid *f,
198 const struct lu_object_conf *conf)
200 struct lu_object *scan;
201 struct lu_object *top;
202 struct list_head *layers;
208 * Create top-level object slice. This will also create
211 top = dev->ld_ops->ldo_object_alloc(env, NULL, dev);
213 RETURN(ERR_PTR(-ENOMEM));
217 * This is the only place where object fid is assigned. It's constant
220 top->lo_header->loh_fid = *f;
221 layers = &top->lo_header->loh_layers;
224 * Call ->loo_object_init() repeatedly, until no more new
225 * object slices are created.
228 list_for_each_entry(scan, layers, lo_linkage) {
229 if (scan->lo_flags & LU_OBJECT_ALLOCATED)
232 scan->lo_header = top->lo_header;
233 result = scan->lo_ops->loo_object_init(env, scan, conf);
235 lu_object_free(env, top);
236 RETURN(ERR_PTR(result));
238 scan->lo_flags |= LU_OBJECT_ALLOCATED;
242 list_for_each_entry_reverse(scan, layers, lo_linkage) {
243 if (scan->lo_ops->loo_object_start != NULL) {
244 result = scan->lo_ops->loo_object_start(env, scan);
246 lu_object_free(env, top);
247 RETURN(ERR_PTR(result));
252 lprocfs_counter_incr(dev->ld_site->ls_stats, LU_SS_CREATED);
259 static void lu_object_free(const struct lu_env *env, struct lu_object *o)
261 struct lu_site_bkt_data *bkt;
262 struct lu_site *site;
263 struct lu_object *scan;
264 struct list_head *layers;
265 struct list_head splice;
267 site = o->lo_dev->ld_site;
268 layers = &o->lo_header->loh_layers;
269 bkt = lu_site_bkt_from_fid(site, &o->lo_header->loh_fid);
271 * First call ->loo_object_delete() method to release all resources.
273 list_for_each_entry_reverse(scan, layers, lo_linkage) {
274 if (scan->lo_ops->loo_object_delete != NULL)
275 scan->lo_ops->loo_object_delete(env, scan);
279 * Then, splice object layers into stand-alone list, and call
280 * ->loo_object_free() on all layers to free memory. Splice is
281 * necessary, because lu_object_header is freed together with the
284 INIT_LIST_HEAD(&splice);
285 list_splice_init(layers, &splice);
286 while (!list_empty(&splice)) {
288 * Free layers in bottom-to-top order, so that object header
289 * lives as long as possible and ->loo_object_free() methods
290 * can look at its contents.
292 o = container_of0(splice.prev, struct lu_object, lo_linkage);
293 list_del_init(&o->lo_linkage);
294 LASSERT(o->lo_ops->loo_object_free != NULL);
295 o->lo_ops->loo_object_free(env, o);
298 if (waitqueue_active(&bkt->lsb_marche_funebre))
299 wake_up_all(&bkt->lsb_marche_funebre);
303 * Free \a nr objects from the cold end of the site LRU list.
305 int lu_site_purge(const struct lu_env *env, struct lu_site *s, int nr)
307 struct lu_object_header *h;
308 struct lu_object_header *temp;
309 struct lu_site_bkt_data *bkt;
312 struct list_head dispose;
319 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_NO_LRU))
322 INIT_LIST_HEAD(&dispose);
324 * Under LRU list lock, scan LRU list and move unreferenced objects to
325 * the dispose list, removing them from LRU and hash table.
327 start = s->ls_purge_start;
328 bnr = (nr == ~0) ? -1 : nr / CFS_HASH_NBKT(s->ls_obj_hash) + 1;
331 cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
335 cfs_hash_bd_lock(s->ls_obj_hash, &bd, 1);
336 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
338 list_for_each_entry_safe(h, temp, &bkt->lsb_lru, loh_lru) {
339 LASSERT(atomic_read(&h->loh_ref) == 0);
341 cfs_hash_bd_get(s->ls_obj_hash, &h->loh_fid, &bd2);
342 LASSERT(bd.bd_bucket == bd2.bd_bucket);
344 cfs_hash_bd_del_locked(s->ls_obj_hash,
346 list_move(&h->loh_lru, &dispose);
350 if (nr != ~0 && --nr == 0)
353 if (count > 0 && --count == 0)
357 cfs_hash_bd_unlock(s->ls_obj_hash, &bd, 1);
360 * Free everything on the dispose list. This is safe against
361 * races due to the reasons described in lu_object_put().
363 while (!list_empty(&dispose)) {
364 h = container_of0(dispose.next,
365 struct lu_object_header, loh_lru);
366 list_del_init(&h->loh_lru);
367 lu_object_free(env, lu_object_top(h));
368 lprocfs_counter_incr(s->ls_stats, LU_SS_LRU_PURGED);
375 if (nr != 0 && did_sth && start != 0) {
376 start = 0; /* restart from the first bucket */
379 /* race on s->ls_purge_start, but nobody cares */
380 s->ls_purge_start = i % CFS_HASH_NBKT(s->ls_obj_hash);
384 EXPORT_SYMBOL(lu_site_purge);
389 * Code below has to jump through certain loops to output object description
390 * into libcfs_debug_msg-based log. The problem is that lu_object_print()
391 * composes object description from strings that are parts of _lines_ of
392 * output (i.e., strings that are not terminated by newline). This doesn't fit
393 * very well into libcfs_debug_msg() interface that assumes that each message
394 * supplied to it is a self-contained output line.
396 * To work around this, strings are collected in a temporary buffer
397 * (implemented as a value of lu_cdebug_key key), until terminating newline
398 * character is detected.
406 * XXX overflow is not handled correctly.
411 struct lu_cdebug_data {
415 char lck_area[LU_CDEBUG_LINE];
418 /* context key constructor/destructor: lu_global_key_init, lu_global_key_fini */
419 LU_KEY_INIT_FINI(lu_global, struct lu_cdebug_data);
422 * Key, holding temporary buffer. This key is registered very early by
425 struct lu_context_key lu_global_key = {
426 .lct_tags = LCT_MD_THREAD | LCT_DT_THREAD |
427 LCT_MG_THREAD | LCT_CL_THREAD,
428 .lct_init = lu_global_key_init,
429 .lct_fini = lu_global_key_fini
433 * Printer function emitting messages through libcfs_debug_msg().
435 int lu_cdebug_printer(const struct lu_env *env,
436 void *cookie, const char *format, ...)
438 struct libcfs_debug_msg_data *msgdata = cookie;
439 struct lu_cdebug_data *key;
444 va_start(args, format);
446 key = lu_context_key_get(&env->le_ctx, &lu_global_key);
447 LASSERT(key != NULL);
449 used = strlen(key->lck_area);
450 complete = format[strlen(format) - 1] == '\n';
452 * Append new chunk to the buffer.
454 vsnprintf(key->lck_area + used,
455 ARRAY_SIZE(key->lck_area) - used, format, args);
457 if (cfs_cdebug_show(msgdata->msg_mask, msgdata->msg_subsys))
458 libcfs_debug_msg(msgdata, "%s", key->lck_area);
459 key->lck_area[0] = 0;
464 EXPORT_SYMBOL(lu_cdebug_printer);
467 * Print object header.
469 void lu_object_header_print(const struct lu_env *env, void *cookie,
470 lu_printer_t printer,
471 const struct lu_object_header *hdr)
473 (*printer)(env, cookie, "header@%p[%#lx, %d, "DFID"%s%s%s]",
474 hdr, hdr->loh_flags, atomic_read(&hdr->loh_ref),
476 hlist_unhashed(&hdr->loh_hash) ? "" : " hash",
477 list_empty((struct list_head *)&hdr->loh_lru) ? \
479 hdr->loh_attr & LOHA_EXISTS ? " exist":"");
481 EXPORT_SYMBOL(lu_object_header_print);
484 * Print human readable representation of the \a o to the \a printer.
486 void lu_object_print(const struct lu_env *env, void *cookie,
487 lu_printer_t printer, const struct lu_object *o)
489 static const char ruler[] = "........................................";
490 struct lu_object_header *top;
494 lu_object_header_print(env, cookie, printer, top);
495 (*printer)(env, cookie, "{ \n");
496 list_for_each_entry(o, &top->loh_layers, lo_linkage) {
497 depth = o->lo_depth + 4;
500 * print `.' \a depth times followed by type name and address
502 (*printer)(env, cookie, "%*.*s%s@%p", depth, depth, ruler,
503 o->lo_dev->ld_type->ldt_name, o);
504 if (o->lo_ops->loo_object_print != NULL)
505 o->lo_ops->loo_object_print(env, cookie, printer, o);
506 (*printer)(env, cookie, "\n");
508 (*printer)(env, cookie, "} header@%p\n", top);
510 EXPORT_SYMBOL(lu_object_print);
513 * Check object consistency.
515 int lu_object_invariant(const struct lu_object *o)
517 struct lu_object_header *top;
520 list_for_each_entry(o, &top->loh_layers, lo_linkage) {
521 if (o->lo_ops->loo_object_invariant != NULL &&
522 !o->lo_ops->loo_object_invariant(o))
527 EXPORT_SYMBOL(lu_object_invariant);
529 static struct lu_object *htable_lookup(struct lu_site *s,
531 const struct lu_fid *f,
532 wait_queue_t *waiter,
535 struct lu_site_bkt_data *bkt;
536 struct lu_object_header *h;
537 struct hlist_node *hnode;
538 __u64 ver = cfs_hash_bd_version_get(bd);
544 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, bd);
545 /* cfs_hash_bd_peek_locked is a somehow "internal" function
546 * of cfs_hash, it doesn't add refcount on object. */
547 hnode = cfs_hash_bd_peek_locked(s->ls_obj_hash, bd, (void *)f);
549 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS);
553 h = container_of0(hnode, struct lu_object_header, loh_hash);
554 if (likely(!lu_object_is_dying(h))) {
555 cfs_hash_get(s->ls_obj_hash, hnode);
556 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_HIT);
557 list_del_init(&h->loh_lru);
558 return lu_object_top(h);
562 * Lookup found an object being destroyed this object cannot be
563 * returned (to assure that references to dying objects are eventually
564 * drained), and moreover, lookup has to wait until object is freed.
567 init_waitqueue_entry_current(waiter);
568 add_wait_queue(&bkt->lsb_marche_funebre, waiter);
569 set_current_state(TASK_UNINTERRUPTIBLE);
570 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_DEATH_RACE);
571 return ERR_PTR(-EAGAIN);
575 * Search cache for an object with the fid \a f. If such object is found,
576 * return it. Otherwise, create new object, insert it into cache and return
577 * it. In any case, additional reference is acquired on the returned object.
579 struct lu_object *lu_object_find(const struct lu_env *env,
580 struct lu_device *dev, const struct lu_fid *f,
581 const struct lu_object_conf *conf)
583 return lu_object_find_at(env, dev->ld_site->ls_top_dev, f, conf);
585 EXPORT_SYMBOL(lu_object_find);
587 static struct lu_object *lu_object_new(const struct lu_env *env,
588 struct lu_device *dev,
589 const struct lu_fid *f,
590 const struct lu_object_conf *conf)
595 struct lu_site_bkt_data *bkt;
597 o = lu_object_alloc(env, dev, f, conf);
598 if (unlikely(IS_ERR(o)))
601 hs = dev->ld_site->ls_obj_hash;
602 cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
603 bkt = cfs_hash_bd_extra_get(hs, &bd);
604 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
606 cfs_hash_bd_unlock(hs, &bd, 1);
611 * Core logic of lu_object_find*() functions.
613 static struct lu_object *lu_object_find_try(const struct lu_env *env,
614 struct lu_device *dev,
615 const struct lu_fid *f,
616 const struct lu_object_conf *conf,
617 wait_queue_t *waiter)
620 struct lu_object *shadow;
627 * This uses standard index maintenance protocol:
629 * - search index under lock, and return object if found;
630 * - otherwise, unlock index, allocate new object;
631 * - lock index and search again;
632 * - if nothing is found (usual case), insert newly created
634 * - otherwise (race: other thread inserted object), free
635 * object just allocated.
639 * For "LOC_F_NEW" case, we are sure the object is new established.
640 * It is unnecessary to perform lookup-alloc-lookup-insert, instead,
641 * just alloc and insert directly.
643 * If dying object is found during index search, add @waiter to the
644 * site wait-queue and return ERR_PTR(-EAGAIN).
646 if (conf != NULL && conf->loc_flags & LOC_F_NEW)
647 return lu_object_new(env, dev, f, conf);
651 cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
652 o = htable_lookup(s, &bd, f, waiter, &version);
653 cfs_hash_bd_unlock(hs, &bd, 1);
658 * Allocate new object. This may result in rather complicated
659 * operations, including fld queries, inode loading, etc.
661 o = lu_object_alloc(env, dev, f, conf);
662 if (unlikely(IS_ERR(o)))
665 LASSERT(lu_fid_eq(lu_object_fid(o), f));
667 cfs_hash_bd_lock(hs, &bd, 1);
669 shadow = htable_lookup(s, &bd, f, waiter, &version);
670 if (likely(shadow == NULL)) {
671 struct lu_site_bkt_data *bkt;
673 bkt = cfs_hash_bd_extra_get(hs, &bd);
674 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
676 cfs_hash_bd_unlock(hs, &bd, 1);
680 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_RACE);
681 cfs_hash_bd_unlock(hs, &bd, 1);
682 lu_object_free(env, o);
687 * Much like lu_object_find(), but top level device of object is specifically
688 * \a dev rather than top level device of the site. This interface allows
689 * objects of different "stacking" to be created within the same site.
691 struct lu_object *lu_object_find_at(const struct lu_env *env,
692 struct lu_device *dev,
693 const struct lu_fid *f,
694 const struct lu_object_conf *conf)
696 struct lu_site_bkt_data *bkt;
697 struct lu_object *obj;
701 obj = lu_object_find_try(env, dev, f, conf, &wait);
702 if (obj != ERR_PTR(-EAGAIN))
705 * lu_object_find_try() already added waiter into the
708 waitq_wait(&wait, TASK_UNINTERRUPTIBLE);
709 bkt = lu_site_bkt_from_fid(dev->ld_site, (void *)f);
710 remove_wait_queue(&bkt->lsb_marche_funebre, &wait);
713 EXPORT_SYMBOL(lu_object_find_at);
716 * Find object with given fid, and return its slice belonging to given device.
718 struct lu_object *lu_object_find_slice(const struct lu_env *env,
719 struct lu_device *dev,
720 const struct lu_fid *f,
721 const struct lu_object_conf *conf)
723 struct lu_object *top;
724 struct lu_object *obj;
726 top = lu_object_find(env, dev, f, conf);
728 obj = lu_object_locate(top->lo_header, dev->ld_type);
730 lu_object_put(env, top);
735 EXPORT_SYMBOL(lu_object_find_slice);
738 * Global list of all device types.
740 static LIST_HEAD(lu_device_types);
742 int lu_device_type_init(struct lu_device_type *ldt)
746 INIT_LIST_HEAD(&ldt->ldt_linkage);
747 if (ldt->ldt_ops->ldto_init)
748 result = ldt->ldt_ops->ldto_init(ldt);
750 list_add(&ldt->ldt_linkage, &lu_device_types);
753 EXPORT_SYMBOL(lu_device_type_init);
755 void lu_device_type_fini(struct lu_device_type *ldt)
757 list_del_init(&ldt->ldt_linkage);
758 if (ldt->ldt_ops->ldto_fini)
759 ldt->ldt_ops->ldto_fini(ldt);
761 EXPORT_SYMBOL(lu_device_type_fini);
763 void lu_types_stop(void)
765 struct lu_device_type *ldt;
767 list_for_each_entry(ldt, &lu_device_types, ldt_linkage) {
768 if (ldt->ldt_device_nr == 0 && ldt->ldt_ops->ldto_stop)
769 ldt->ldt_ops->ldto_stop(ldt);
772 EXPORT_SYMBOL(lu_types_stop);
775 * Global list of all sites on this node
777 static LIST_HEAD(lu_sites);
778 static DEFINE_MUTEX(lu_sites_guard);
781 * Global environment used by site shrinker.
783 static struct lu_env lu_shrink_env;
785 struct lu_site_print_arg {
786 struct lu_env *lsp_env;
788 lu_printer_t lsp_printer;
792 lu_site_obj_print(cfs_hash_t *hs, cfs_hash_bd_t *bd,
793 struct hlist_node *hnode, void *data)
795 struct lu_site_print_arg *arg = (struct lu_site_print_arg *)data;
796 struct lu_object_header *h;
798 h = hlist_entry(hnode, struct lu_object_header, loh_hash);
799 if (!list_empty(&h->loh_layers)) {
800 const struct lu_object *o;
802 o = lu_object_top(h);
803 lu_object_print(arg->lsp_env, arg->lsp_cookie,
804 arg->lsp_printer, o);
806 lu_object_header_print(arg->lsp_env, arg->lsp_cookie,
807 arg->lsp_printer, h);
813 * Print all objects in \a s.
815 void lu_site_print(const struct lu_env *env, struct lu_site *s, void *cookie,
816 lu_printer_t printer)
818 struct lu_site_print_arg arg = {
819 .lsp_env = (struct lu_env *)env,
820 .lsp_cookie = cookie,
821 .lsp_printer = printer,
824 cfs_hash_for_each(s->ls_obj_hash, lu_site_obj_print, &arg);
826 EXPORT_SYMBOL(lu_site_print);
829 LU_CACHE_PERCENT_MAX = 50,
830 LU_CACHE_PERCENT_DEFAULT = 20
833 static unsigned int lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
834 CFS_MODULE_PARM(lu_cache_percent, "i", int, 0644,
835 "Percentage of memory to be used as lu_object cache");
838 * Return desired hash table order.
840 static int lu_htable_order(void)
842 unsigned long cache_size;
846 * Calculate hash table size, assuming that we want reasonable
847 * performance when 20% of total memory is occupied by cache of
850 * Size of lu_object is (arbitrary) taken as 1K (together with inode).
852 cache_size = num_physpages;
854 #if BITS_PER_LONG == 32
855 /* limit hashtable size for lowmem systems to low RAM */
856 if (cache_size > 1 << (30 - PAGE_CACHE_SHIFT))
857 cache_size = 1 << (30 - PAGE_CACHE_SHIFT) * 3 / 4;
860 /* clear off unreasonable cache setting. */
861 if (lu_cache_percent == 0 || lu_cache_percent > LU_CACHE_PERCENT_MAX) {
862 CWARN("obdclass: invalid lu_cache_percent: %u, it must be in"
863 " the range of (0, %u]. Will use default value: %u.\n",
864 lu_cache_percent, LU_CACHE_PERCENT_MAX,
865 LU_CACHE_PERCENT_DEFAULT);
867 lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
869 cache_size = cache_size / 100 * lu_cache_percent *
870 (PAGE_CACHE_SIZE / 1024);
872 for (bits = 1; (1 << bits) < cache_size; ++bits) {
878 static unsigned lu_obj_hop_hash(cfs_hash_t *hs,
879 const void *key, unsigned mask)
881 struct lu_fid *fid = (struct lu_fid *)key;
884 hash = fid_flatten32(fid);
885 hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
886 hash = cfs_hash_long(hash, hs->hs_bkt_bits);
888 /* give me another random factor */
889 hash -= cfs_hash_long((unsigned long)hs, fid_oid(fid) % 11 + 3);
891 hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
892 hash |= (fid_seq(fid) + fid_oid(fid)) & (CFS_HASH_NBKT(hs) - 1);
897 static void *lu_obj_hop_object(struct hlist_node *hnode)
899 return hlist_entry(hnode, struct lu_object_header, loh_hash);
902 static void *lu_obj_hop_key(struct hlist_node *hnode)
904 struct lu_object_header *h;
906 h = hlist_entry(hnode, struct lu_object_header, loh_hash);
910 static int lu_obj_hop_keycmp(const void *key, struct hlist_node *hnode)
912 struct lu_object_header *h;
914 h = hlist_entry(hnode, struct lu_object_header, loh_hash);
915 return lu_fid_eq(&h->loh_fid, (struct lu_fid *)key);
918 static void lu_obj_hop_get(cfs_hash_t *hs, struct hlist_node *hnode)
920 struct lu_object_header *h;
922 h = hlist_entry(hnode, struct lu_object_header, loh_hash);
923 if (atomic_add_return(1, &h->loh_ref) == 1) {
924 struct lu_site_bkt_data *bkt;
927 cfs_hash_bd_get(hs, &h->loh_fid, &bd);
928 bkt = cfs_hash_bd_extra_get(hs, &bd);
933 static void lu_obj_hop_put_locked(cfs_hash_t *hs, struct hlist_node *hnode)
935 LBUG(); /* we should never called it */
938 cfs_hash_ops_t lu_site_hash_ops = {
939 .hs_hash = lu_obj_hop_hash,
940 .hs_key = lu_obj_hop_key,
941 .hs_keycmp = lu_obj_hop_keycmp,
942 .hs_object = lu_obj_hop_object,
943 .hs_get = lu_obj_hop_get,
944 .hs_put_locked = lu_obj_hop_put_locked,
947 void lu_dev_add_linkage(struct lu_site *s, struct lu_device *d)
949 spin_lock(&s->ls_ld_lock);
950 if (list_empty(&d->ld_linkage))
951 list_add(&d->ld_linkage, &s->ls_ld_linkage);
952 spin_unlock(&s->ls_ld_lock);
954 EXPORT_SYMBOL(lu_dev_add_linkage);
956 void lu_dev_del_linkage(struct lu_site *s, struct lu_device *d)
958 spin_lock(&s->ls_ld_lock);
959 list_del_init(&d->ld_linkage);
960 spin_unlock(&s->ls_ld_lock);
962 EXPORT_SYMBOL(lu_dev_del_linkage);
965 * Initialize site \a s, with \a d as the top level device.
967 #define LU_SITE_BITS_MIN 12
968 #define LU_SITE_BITS_MAX 24
970 * total 256 buckets, we don't want too many buckets because:
971 * - consume too much memory
972 * - avoid unbalanced LRU list
974 #define LU_SITE_BKT_BITS 8
976 int lu_site_init(struct lu_site *s, struct lu_device *top)
978 struct lu_site_bkt_data *bkt;
985 memset(s, 0, sizeof *s);
986 bits = lu_htable_order();
987 snprintf(name, 16, "lu_site_%s", top->ld_type->ldt_name);
988 for (bits = min(max(LU_SITE_BITS_MIN, bits), LU_SITE_BITS_MAX);
989 bits >= LU_SITE_BITS_MIN; bits--) {
990 s->ls_obj_hash = cfs_hash_create(name, bits, bits,
991 bits - LU_SITE_BKT_BITS,
994 CFS_HASH_SPIN_BKTLOCK |
995 CFS_HASH_NO_ITEMREF |
997 CFS_HASH_ASSERT_EMPTY);
998 if (s->ls_obj_hash != NULL)
1002 if (s->ls_obj_hash == NULL) {
1003 CERROR("failed to create lu_site hash with bits: %d\n", bits);
1007 cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
1008 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
1009 INIT_LIST_HEAD(&bkt->lsb_lru);
1010 init_waitqueue_head(&bkt->lsb_marche_funebre);
1013 s->ls_stats = lprocfs_alloc_stats(LU_SS_LAST_STAT, 0);
1014 if (s->ls_stats == NULL) {
1015 cfs_hash_putref(s->ls_obj_hash);
1016 s->ls_obj_hash = NULL;
1020 lprocfs_counter_init(s->ls_stats, LU_SS_CREATED,
1021 0, "created", "created");
1022 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_HIT,
1023 0, "cache_hit", "cache_hit");
1024 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_MISS,
1025 0, "cache_miss", "cache_miss");
1026 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_RACE,
1027 0, "cache_race", "cache_race");
1028 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_DEATH_RACE,
1029 0, "cache_death_race", "cache_death_race");
1030 lprocfs_counter_init(s->ls_stats, LU_SS_LRU_PURGED,
1031 0, "lru_purged", "lru_purged");
1033 INIT_LIST_HEAD(&s->ls_linkage);
1034 s->ls_top_dev = top;
1037 lu_ref_add(&top->ld_reference, "site-top", s);
1039 INIT_LIST_HEAD(&s->ls_ld_linkage);
1040 spin_lock_init(&s->ls_ld_lock);
1042 lu_dev_add_linkage(s, top);
1046 EXPORT_SYMBOL(lu_site_init);
1049 * Finalize \a s and release its resources.
1051 void lu_site_fini(struct lu_site *s)
1053 mutex_lock(&lu_sites_guard);
1054 list_del_init(&s->ls_linkage);
1055 mutex_unlock(&lu_sites_guard);
1057 if (s->ls_obj_hash != NULL) {
1058 cfs_hash_putref(s->ls_obj_hash);
1059 s->ls_obj_hash = NULL;
1062 if (s->ls_top_dev != NULL) {
1063 s->ls_top_dev->ld_site = NULL;
1064 lu_ref_del(&s->ls_top_dev->ld_reference, "site-top", s);
1065 lu_device_put(s->ls_top_dev);
1066 s->ls_top_dev = NULL;
1069 if (s->ls_stats != NULL)
1070 lprocfs_free_stats(&s->ls_stats);
1072 EXPORT_SYMBOL(lu_site_fini);
1075 * Called when initialization of stack for this site is completed.
1077 int lu_site_init_finish(struct lu_site *s)
1080 mutex_lock(&lu_sites_guard);
1081 result = lu_context_refill(&lu_shrink_env.le_ctx);
1083 list_add(&s->ls_linkage, &lu_sites);
1084 mutex_unlock(&lu_sites_guard);
1087 EXPORT_SYMBOL(lu_site_init_finish);
1090 * Acquire additional reference on device \a d
1092 void lu_device_get(struct lu_device *d)
1094 atomic_inc(&d->ld_ref);
1096 EXPORT_SYMBOL(lu_device_get);
1099 * Release reference on device \a d.
1101 void lu_device_put(struct lu_device *d)
1103 LASSERT(atomic_read(&d->ld_ref) > 0);
1104 atomic_dec(&d->ld_ref);
1106 EXPORT_SYMBOL(lu_device_put);
1109 * Initialize device \a d of type \a t.
1111 int lu_device_init(struct lu_device *d, struct lu_device_type *t)
1113 if (t->ldt_device_nr++ == 0 && t->ldt_ops->ldto_start != NULL)
1114 t->ldt_ops->ldto_start(t);
1115 memset(d, 0, sizeof *d);
1116 atomic_set(&d->ld_ref, 0);
1118 lu_ref_init(&d->ld_reference);
1119 INIT_LIST_HEAD(&d->ld_linkage);
1122 EXPORT_SYMBOL(lu_device_init);
1125 * Finalize device \a d.
1127 void lu_device_fini(struct lu_device *d)
1129 struct lu_device_type *t;
1132 if (d->ld_obd != NULL) {
1133 d->ld_obd->obd_lu_dev = NULL;
1137 lu_ref_fini(&d->ld_reference);
1138 LASSERTF(atomic_read(&d->ld_ref) == 0,
1139 "Refcount is %u\n", atomic_read(&d->ld_ref));
1140 LASSERT(t->ldt_device_nr > 0);
1141 if (--t->ldt_device_nr == 0 && t->ldt_ops->ldto_stop != NULL)
1142 t->ldt_ops->ldto_stop(t);
1144 EXPORT_SYMBOL(lu_device_fini);
1147 * Initialize object \a o that is part of compound object \a h and was created
1150 int lu_object_init(struct lu_object *o,
1151 struct lu_object_header *h, struct lu_device *d)
1153 memset(o, 0, sizeof *o);
1157 o->lo_dev_ref = lu_ref_add(&d->ld_reference, "lu_object", o);
1158 INIT_LIST_HEAD(&o->lo_linkage);
1161 EXPORT_SYMBOL(lu_object_init);
1164 * Finalize object and release its resources.
1166 void lu_object_fini(struct lu_object *o)
1168 struct lu_device *dev = o->lo_dev;
1170 LASSERT(list_empty(&o->lo_linkage));
1173 lu_ref_del_at(&dev->ld_reference,
1174 o->lo_dev_ref , "lu_object", o);
1179 EXPORT_SYMBOL(lu_object_fini);
1182 * Add object \a o as first layer of compound object \a h
1184 * This is typically called by the ->ldo_object_alloc() method of top-level
1187 void lu_object_add_top(struct lu_object_header *h, struct lu_object *o)
1189 list_move(&o->lo_linkage, &h->loh_layers);
1191 EXPORT_SYMBOL(lu_object_add_top);
1194 * Add object \a o as a layer of compound object, going after \a before.
1196 * This is typically called by the ->ldo_object_alloc() method of \a
1199 void lu_object_add(struct lu_object *before, struct lu_object *o)
1201 list_move(&o->lo_linkage, &before->lo_linkage);
1203 EXPORT_SYMBOL(lu_object_add);
1206 * Initialize compound object.
1208 int lu_object_header_init(struct lu_object_header *h)
1210 memset(h, 0, sizeof *h);
1211 atomic_set(&h->loh_ref, 1);
1212 INIT_HLIST_NODE(&h->loh_hash);
1213 INIT_LIST_HEAD(&h->loh_lru);
1214 INIT_LIST_HEAD(&h->loh_layers);
1215 lu_ref_init(&h->loh_reference);
1218 EXPORT_SYMBOL(lu_object_header_init);
1221 * Finalize compound object.
1223 void lu_object_header_fini(struct lu_object_header *h)
1225 LASSERT(list_empty(&h->loh_layers));
1226 LASSERT(list_empty(&h->loh_lru));
1227 LASSERT(hlist_unhashed(&h->loh_hash));
1228 lu_ref_fini(&h->loh_reference);
1230 EXPORT_SYMBOL(lu_object_header_fini);
1233 * Given a compound object, find its slice, corresponding to the device type
1236 struct lu_object *lu_object_locate(struct lu_object_header *h,
1237 const struct lu_device_type *dtype)
1239 struct lu_object *o;
1241 list_for_each_entry(o, &h->loh_layers, lo_linkage) {
1242 if (o->lo_dev->ld_type == dtype)
1247 EXPORT_SYMBOL(lu_object_locate);
1252 * Finalize and free devices in the device stack.
1254 * Finalize device stack by purging object cache, and calling
1255 * lu_device_type_operations::ldto_device_fini() and
1256 * lu_device_type_operations::ldto_device_free() on all devices in the stack.
1258 void lu_stack_fini(const struct lu_env *env, struct lu_device *top)
1260 struct lu_site *site = top->ld_site;
1261 struct lu_device *scan;
1262 struct lu_device *next;
1264 lu_site_purge(env, site, ~0);
1265 for (scan = top; scan != NULL; scan = next) {
1266 next = scan->ld_type->ldt_ops->ldto_device_fini(env, scan);
1267 lu_ref_del(&scan->ld_reference, "lu-stack", &lu_site_init);
1268 lu_device_put(scan);
1272 lu_site_purge(env, site, ~0);
1274 for (scan = top; scan != NULL; scan = next) {
1275 const struct lu_device_type *ldt = scan->ld_type;
1276 struct obd_type *type;
1278 next = ldt->ldt_ops->ldto_device_free(env, scan);
1279 type = ldt->ldt_obd_type;
1282 class_put_type(type);
1286 EXPORT_SYMBOL(lu_stack_fini);
1290 * Maximal number of tld slots.
1292 LU_CONTEXT_KEY_NR = 40
1295 static struct lu_context_key *lu_keys[LU_CONTEXT_KEY_NR] = { NULL, };
1297 static DEFINE_SPINLOCK(lu_keys_guard);
1300 * Global counter incremented whenever key is registered, unregistered,
1301 * revived or quiesced. This is used to void unnecessary calls to
1302 * lu_context_refill(). No locking is provided, as initialization and shutdown
1303 * are supposed to be externally serialized.
1305 static unsigned key_set_version = 0;
1310 int lu_context_key_register(struct lu_context_key *key)
1315 LASSERT(key->lct_init != NULL);
1316 LASSERT(key->lct_fini != NULL);
1317 LASSERT(key->lct_tags != 0);
1318 LASSERT(key->lct_owner != NULL);
1321 spin_lock(&lu_keys_guard);
1322 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1323 if (lu_keys[i] == NULL) {
1325 atomic_set(&key->lct_used, 1);
1327 lu_ref_init(&key->lct_reference);
1333 spin_unlock(&lu_keys_guard);
1336 EXPORT_SYMBOL(lu_context_key_register);
1338 static void key_fini(struct lu_context *ctx, int index)
1340 if (ctx->lc_value != NULL && ctx->lc_value[index] != NULL) {
1341 struct lu_context_key *key;
1343 key = lu_keys[index];
1344 LASSERT(key != NULL);
1345 LASSERT(key->lct_fini != NULL);
1346 LASSERT(atomic_read(&key->lct_used) > 1);
1348 key->lct_fini(ctx, key, ctx->lc_value[index]);
1349 lu_ref_del(&key->lct_reference, "ctx", ctx);
1350 atomic_dec(&key->lct_used);
1352 LASSERT(key->lct_owner != NULL);
1353 if ((ctx->lc_tags & LCT_NOREF) == 0) {
1354 #ifdef CONFIG_MODULE_UNLOAD
1355 LINVRNT(module_refcount(key->lct_owner) > 0);
1357 module_put(key->lct_owner);
1359 ctx->lc_value[index] = NULL;
1366 void lu_context_key_degister(struct lu_context_key *key)
1368 LASSERT(atomic_read(&key->lct_used) >= 1);
1369 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1371 lu_context_key_quiesce(key);
1374 spin_lock(&lu_keys_guard);
1375 key_fini(&lu_shrink_env.le_ctx, key->lct_index);
1376 if (lu_keys[key->lct_index]) {
1377 lu_keys[key->lct_index] = NULL;
1378 lu_ref_fini(&key->lct_reference);
1380 spin_unlock(&lu_keys_guard);
1382 LASSERTF(atomic_read(&key->lct_used) == 1,
1383 "key has instances: %d\n",
1384 atomic_read(&key->lct_used));
1386 EXPORT_SYMBOL(lu_context_key_degister);
1389 * Register a number of keys. This has to be called after all keys have been
1390 * initialized by a call to LU_CONTEXT_KEY_INIT().
1392 int lu_context_key_register_many(struct lu_context_key *k, ...)
1394 struct lu_context_key *key = k;
1400 result = lu_context_key_register(key);
1403 key = va_arg(args, struct lu_context_key *);
1404 } while (key != NULL);
1410 lu_context_key_degister(k);
1411 k = va_arg(args, struct lu_context_key *);
1418 EXPORT_SYMBOL(lu_context_key_register_many);
1421 * De-register a number of keys. This is a dual to
1422 * lu_context_key_register_many().
1424 void lu_context_key_degister_many(struct lu_context_key *k, ...)
1430 lu_context_key_degister(k);
1431 k = va_arg(args, struct lu_context_key*);
1432 } while (k != NULL);
1435 EXPORT_SYMBOL(lu_context_key_degister_many);
1438 * Revive a number of keys.
1440 void lu_context_key_revive_many(struct lu_context_key *k, ...)
1446 lu_context_key_revive(k);
1447 k = va_arg(args, struct lu_context_key*);
1448 } while (k != NULL);
1451 EXPORT_SYMBOL(lu_context_key_revive_many);
1454 * Quiescent a number of keys.
1456 void lu_context_key_quiesce_many(struct lu_context_key *k, ...)
1462 lu_context_key_quiesce(k);
1463 k = va_arg(args, struct lu_context_key*);
1464 } while (k != NULL);
1467 EXPORT_SYMBOL(lu_context_key_quiesce_many);
1470 * Return value associated with key \a key in context \a ctx.
1472 void *lu_context_key_get(const struct lu_context *ctx,
1473 const struct lu_context_key *key)
1475 LINVRNT(ctx->lc_state == LCS_ENTERED);
1476 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1477 LASSERT(lu_keys[key->lct_index] == key);
1478 return ctx->lc_value[key->lct_index];
1480 EXPORT_SYMBOL(lu_context_key_get);
1483 * List of remembered contexts. XXX document me.
1485 static LIST_HEAD(lu_context_remembered);
1488 * Destroy \a key in all remembered contexts. This is used to destroy key
1489 * values in "shared" contexts (like service threads), when a module owning
1490 * the key is about to be unloaded.
1492 void lu_context_key_quiesce(struct lu_context_key *key)
1494 struct lu_context *ctx;
1496 if (!(key->lct_tags & LCT_QUIESCENT)) {
1498 * XXX layering violation.
1500 key->lct_tags |= LCT_QUIESCENT;
1502 * XXX memory barrier has to go here.
1504 spin_lock(&lu_keys_guard);
1505 list_for_each_entry(ctx, &lu_context_remembered,
1507 key_fini(ctx, key->lct_index);
1508 spin_unlock(&lu_keys_guard);
1512 EXPORT_SYMBOL(lu_context_key_quiesce);
1514 void lu_context_key_revive(struct lu_context_key *key)
1516 key->lct_tags &= ~LCT_QUIESCENT;
1519 EXPORT_SYMBOL(lu_context_key_revive);
1521 static void keys_fini(struct lu_context *ctx)
1525 if (ctx->lc_value == NULL)
1528 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i)
1531 OBD_FREE(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1532 ctx->lc_value = NULL;
1535 static int keys_fill(struct lu_context *ctx)
1539 LINVRNT(ctx->lc_value != NULL);
1540 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1541 struct lu_context_key *key;
1544 if (ctx->lc_value[i] == NULL && key != NULL &&
1545 (key->lct_tags & ctx->lc_tags) &&
1547 * Don't create values for a LCT_QUIESCENT key, as this
1548 * will pin module owning a key.
1550 !(key->lct_tags & LCT_QUIESCENT)) {
1553 LINVRNT(key->lct_init != NULL);
1554 LINVRNT(key->lct_index == i);
1556 value = key->lct_init(ctx, key);
1557 if (unlikely(IS_ERR(value)))
1558 return PTR_ERR(value);
1560 LASSERT(key->lct_owner != NULL);
1561 if (!(ctx->lc_tags & LCT_NOREF))
1562 try_module_get(key->lct_owner);
1563 lu_ref_add_atomic(&key->lct_reference, "ctx", ctx);
1564 atomic_inc(&key->lct_used);
1566 * This is the only place in the code, where an
1567 * element of ctx->lc_value[] array is set to non-NULL
1570 ctx->lc_value[i] = value;
1571 if (key->lct_exit != NULL)
1572 ctx->lc_tags |= LCT_HAS_EXIT;
1574 ctx->lc_version = key_set_version;
1579 static int keys_init(struct lu_context *ctx)
1581 OBD_ALLOC(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1582 if (likely(ctx->lc_value != NULL))
1583 return keys_fill(ctx);
1589 * Initialize context data-structure. Create values for all keys.
1591 int lu_context_init(struct lu_context *ctx, __u32 tags)
1595 memset(ctx, 0, sizeof *ctx);
1596 ctx->lc_state = LCS_INITIALIZED;
1597 ctx->lc_tags = tags;
1598 if (tags & LCT_REMEMBER) {
1599 spin_lock(&lu_keys_guard);
1600 list_add(&ctx->lc_remember, &lu_context_remembered);
1601 spin_unlock(&lu_keys_guard);
1603 INIT_LIST_HEAD(&ctx->lc_remember);
1606 rc = keys_init(ctx);
1608 lu_context_fini(ctx);
1612 EXPORT_SYMBOL(lu_context_init);
1615 * Finalize context data-structure. Destroy key values.
1617 void lu_context_fini(struct lu_context *ctx)
1619 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1620 ctx->lc_state = LCS_FINALIZED;
1622 if ((ctx->lc_tags & LCT_REMEMBER) == 0) {
1623 LASSERT(list_empty(&ctx->lc_remember));
1626 } else { /* could race with key degister */
1627 spin_lock(&lu_keys_guard);
1629 list_del_init(&ctx->lc_remember);
1630 spin_unlock(&lu_keys_guard);
1633 EXPORT_SYMBOL(lu_context_fini);
1636 * Called before entering context.
1638 void lu_context_enter(struct lu_context *ctx)
1640 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1641 ctx->lc_state = LCS_ENTERED;
1643 EXPORT_SYMBOL(lu_context_enter);
1646 * Called after exiting from \a ctx
1648 void lu_context_exit(struct lu_context *ctx)
1652 LINVRNT(ctx->lc_state == LCS_ENTERED);
1653 ctx->lc_state = LCS_LEFT;
1654 if (ctx->lc_tags & LCT_HAS_EXIT && ctx->lc_value != NULL) {
1655 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1656 if (ctx->lc_value[i] != NULL) {
1657 struct lu_context_key *key;
1660 LASSERT(key != NULL);
1661 if (key->lct_exit != NULL)
1663 key, ctx->lc_value[i]);
1668 EXPORT_SYMBOL(lu_context_exit);
1671 * Allocate for context all missing keys that were registered after context
1672 * creation. key_set_version is only changed in rare cases when modules
1673 * are loaded and removed.
1675 int lu_context_refill(struct lu_context *ctx)
1677 return likely(ctx->lc_version == key_set_version) ? 0 : keys_fill(ctx);
1679 EXPORT_SYMBOL(lu_context_refill);
1682 * lu_ctx_tags/lu_ses_tags will be updated if there are new types of
1683 * obd being added. Currently, this is only used on client side, specifically
1684 * for echo device client, for other stack (like ptlrpc threads), context are
1685 * predefined when the lu_device type are registered, during the module probe
1688 __u32 lu_context_tags_default = 0;
1689 __u32 lu_session_tags_default = 0;
1691 void lu_context_tags_update(__u32 tags)
1693 spin_lock(&lu_keys_guard);
1694 lu_context_tags_default |= tags;
1696 spin_unlock(&lu_keys_guard);
1698 EXPORT_SYMBOL(lu_context_tags_update);
1700 void lu_context_tags_clear(__u32 tags)
1702 spin_lock(&lu_keys_guard);
1703 lu_context_tags_default &= ~tags;
1705 spin_unlock(&lu_keys_guard);
1707 EXPORT_SYMBOL(lu_context_tags_clear);
1709 void lu_session_tags_update(__u32 tags)
1711 spin_lock(&lu_keys_guard);
1712 lu_session_tags_default |= tags;
1714 spin_unlock(&lu_keys_guard);
1716 EXPORT_SYMBOL(lu_session_tags_update);
1718 void lu_session_tags_clear(__u32 tags)
1720 spin_lock(&lu_keys_guard);
1721 lu_session_tags_default &= ~tags;
1723 spin_unlock(&lu_keys_guard);
1725 EXPORT_SYMBOL(lu_session_tags_clear);
1727 int lu_env_init(struct lu_env *env, __u32 tags)
1732 result = lu_context_init(&env->le_ctx, tags);
1733 if (likely(result == 0))
1734 lu_context_enter(&env->le_ctx);
1737 EXPORT_SYMBOL(lu_env_init);
1739 void lu_env_fini(struct lu_env *env)
1741 lu_context_exit(&env->le_ctx);
1742 lu_context_fini(&env->le_ctx);
1745 EXPORT_SYMBOL(lu_env_fini);
1747 int lu_env_refill(struct lu_env *env)
1751 result = lu_context_refill(&env->le_ctx);
1752 if (result == 0 && env->le_ses != NULL)
1753 result = lu_context_refill(env->le_ses);
1756 EXPORT_SYMBOL(lu_env_refill);
1759 * Currently, this API will only be used by echo client.
1760 * Because echo client and normal lustre client will share
1761 * same cl_env cache. So echo client needs to refresh
1762 * the env context after it get one from the cache, especially
1763 * when normal client and echo client co-exist in the same client.
1765 int lu_env_refill_by_tags(struct lu_env *env, __u32 ctags,
1770 if ((env->le_ctx.lc_tags & ctags) != ctags) {
1771 env->le_ctx.lc_version = 0;
1772 env->le_ctx.lc_tags |= ctags;
1775 if (env->le_ses && (env->le_ses->lc_tags & stags) != stags) {
1776 env->le_ses->lc_version = 0;
1777 env->le_ses->lc_tags |= stags;
1780 result = lu_env_refill(env);
1784 EXPORT_SYMBOL(lu_env_refill_by_tags);
1786 static struct shrinker *lu_site_shrinker = NULL;
1788 typedef struct lu_site_stats{
1789 unsigned lss_populated;
1790 unsigned lss_max_search;
1795 static void lu_site_stats_get(cfs_hash_t *hs,
1796 lu_site_stats_t *stats, int populated)
1801 cfs_hash_for_each_bucket(hs, &bd, i) {
1802 struct lu_site_bkt_data *bkt = cfs_hash_bd_extra_get(hs, &bd);
1803 struct hlist_head *hhead;
1805 cfs_hash_bd_lock(hs, &bd, 1);
1806 stats->lss_busy += bkt->lsb_busy;
1807 stats->lss_total += cfs_hash_bd_count_get(&bd);
1808 stats->lss_max_search = max((int)stats->lss_max_search,
1809 cfs_hash_bd_depmax_get(&bd));
1811 cfs_hash_bd_unlock(hs, &bd, 1);
1815 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
1816 if (!hlist_empty(hhead))
1817 stats->lss_populated++;
1819 cfs_hash_bd_unlock(hs, &bd, 1);
1825 * There exists a potential lock inversion deadlock scenario when using
1826 * Lustre on top of ZFS. This occurs between one of ZFS's
1827 * buf_hash_table.ht_lock's, and Lustre's lu_sites_guard lock. Essentially,
1828 * thread A will take the lu_sites_guard lock and sleep on the ht_lock,
1829 * while thread B will take the ht_lock and sleep on the lu_sites_guard
1830 * lock. Obviously neither thread will wake and drop their respective hold
1833 * To prevent this from happening we must ensure the lu_sites_guard lock is
1834 * not taken while down this code path. ZFS reliably does not set the
1835 * __GFP_FS bit in its code paths, so this can be used to determine if it
1836 * is safe to take the lu_sites_guard lock.
1838 * Ideally we should accurately return the remaining number of cached
1839 * objects without taking the lu_sites_guard lock, but this is not
1840 * possible in the current implementation.
1842 static int lu_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
1844 lu_site_stats_t stats;
1846 struct lu_site *tmp;
1848 int remain = shrink_param(sc, nr_to_scan);
1851 if (!(shrink_param(sc, gfp_mask) & __GFP_FS)) {
1855 /* We must not take the lu_sites_guard lock when
1856 * __GFP_FS is *not* set because of the deadlock
1857 * possibility detailed above. Additionally,
1858 * since we cannot determine the number of
1859 * objects in the cache without taking this
1860 * lock, we're in a particularly tough spot. As
1861 * a result, we'll just lie and say our cache is
1862 * empty. This _should_ be ok, as we can't
1863 * reclaim objects when __GFP_FS is *not* set
1869 CDEBUG(D_INODE, "Shrink %d objects\n", remain);
1871 mutex_lock(&lu_sites_guard);
1872 list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
1873 if (shrink_param(sc, nr_to_scan) != 0) {
1874 remain = lu_site_purge(&lu_shrink_env, s, remain);
1876 * Move just shrunk site to the tail of site list to
1877 * assure shrinking fairness.
1879 list_move_tail(&s->ls_linkage, &splice);
1882 memset(&stats, 0, sizeof(stats));
1883 lu_site_stats_get(s->ls_obj_hash, &stats, 0);
1884 cached += stats.lss_total - stats.lss_busy;
1885 if (shrink_param(sc, nr_to_scan) && remain <= 0)
1888 list_splice(&splice, lu_sites.prev);
1889 mutex_unlock(&lu_sites_guard);
1891 cached = (cached / 100) * sysctl_vfs_cache_pressure;
1892 if (shrink_param(sc, nr_to_scan) == 0)
1893 CDEBUG(D_INODE, "%d objects cached\n", cached);
1902 * Environment to be used in debugger, contains all tags.
1904 struct lu_env lu_debugging_env;
1907 * Debugging printer function using printk().
1909 int lu_printk_printer(const struct lu_env *env,
1910 void *unused, const char *format, ...)
1914 va_start(args, format);
1915 vprintk(format, args);
1921 * Initialization of global lu_* data.
1923 int lu_global_init(void)
1927 CDEBUG(D_INFO, "Lustre LU module (%p).\n", &lu_keys);
1929 result = lu_ref_global_init();
1933 LU_CONTEXT_KEY_INIT(&lu_global_key);
1934 result = lu_context_key_register(&lu_global_key);
1939 * At this level, we don't know what tags are needed, so allocate them
1940 * conservatively. This should not be too bad, because this
1941 * environment is global.
1943 mutex_lock(&lu_sites_guard);
1944 result = lu_env_init(&lu_shrink_env, LCT_SHRINKER);
1945 mutex_unlock(&lu_sites_guard);
1950 * seeks estimation: 3 seeks to read a record from oi, one to read
1951 * inode, one for ea. Unfortunately setting this high value results in
1952 * lu_object/inode cache consuming all the memory.
1954 lu_site_shrinker = set_shrinker(DEFAULT_SEEKS, lu_cache_shrink);
1955 if (lu_site_shrinker == NULL)
1962 * Dual to lu_global_init().
1964 void lu_global_fini(void)
1966 if (lu_site_shrinker != NULL) {
1967 remove_shrinker(lu_site_shrinker);
1968 lu_site_shrinker = NULL;
1971 lu_context_key_degister(&lu_global_key);
1974 * Tear shrinker environment down _after_ de-registering
1975 * lu_global_key, because the latter has a value in the former.
1977 mutex_lock(&lu_sites_guard);
1978 lu_env_fini(&lu_shrink_env);
1979 mutex_unlock(&lu_sites_guard);
1981 lu_ref_global_fini();
1984 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx)
1987 struct lprocfs_counter ret;
1989 lprocfs_stats_collect(stats, idx, &ret);
1990 return (__u32)ret.lc_count;
1997 * Output site statistical counters into a buffer. Suitable for
1998 * lprocfs_rd_*()-style functions.
2000 int lu_site_stats_print(const struct lu_site *s, struct seq_file *m)
2002 lu_site_stats_t stats;
2004 memset(&stats, 0, sizeof(stats));
2005 lu_site_stats_get(s->ls_obj_hash, &stats, 1);
2007 return seq_printf(m, "%d/%d %d/%d %d %d %d %d %d %d %d\n",
2010 stats.lss_populated,
2011 CFS_HASH_NHLIST(s->ls_obj_hash),
2012 stats.lss_max_search,
2013 ls_stats_read(s->ls_stats, LU_SS_CREATED),
2014 ls_stats_read(s->ls_stats, LU_SS_CACHE_HIT),
2015 ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
2016 ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
2017 ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
2018 ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED));
2020 EXPORT_SYMBOL(lu_site_stats_print);
2023 * Helper function to initialize a number of kmem slab caches at once.
2025 int lu_kmem_init(struct lu_kmem_descr *caches)
2028 struct lu_kmem_descr *iter = caches;
2030 for (result = 0; iter->ckd_cache != NULL; ++iter) {
2031 *iter->ckd_cache = kmem_cache_create(iter->ckd_name,
2034 if (*iter->ckd_cache == NULL) {
2036 /* free all previously allocated caches */
2037 lu_kmem_fini(caches);
2043 EXPORT_SYMBOL(lu_kmem_init);
2046 * Helper function to finalize a number of kmem slab cached at once. Dual to
2049 void lu_kmem_fini(struct lu_kmem_descr *caches)
2051 for (; caches->ckd_cache != NULL; ++caches) {
2052 if (*caches->ckd_cache != NULL) {
2053 kmem_cache_destroy(*caches->ckd_cache);
2054 *caches->ckd_cache = NULL;
2058 EXPORT_SYMBOL(lu_kmem_fini);
2061 * Temporary solution to be able to assign fid in ->do_create()
2062 * till we have fully-functional OST fids
2064 void lu_object_assign_fid(const struct lu_env *env, struct lu_object *o,
2065 const struct lu_fid *fid)
2067 struct lu_site *s = o->lo_dev->ld_site;
2068 struct lu_fid *old = &o->lo_header->loh_fid;
2069 struct lu_site_bkt_data *bkt;
2070 struct lu_object *shadow;
2071 wait_queue_t waiter;
2076 LASSERT(fid_is_zero(old));
2078 hs = s->ls_obj_hash;
2079 cfs_hash_bd_get_and_lock(hs, (void *)fid, &bd, 1);
2080 shadow = htable_lookup(s, &bd, fid, &waiter, &version);
2081 /* supposed to be unique */
2082 LASSERT(shadow == NULL);
2084 bkt = cfs_hash_bd_extra_get(hs, &bd);
2085 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
2087 cfs_hash_bd_unlock(hs, &bd, 1);
2089 EXPORT_SYMBOL(lu_object_assign_fid);
2092 * allocates object with 0 (non-assiged) fid
2093 * XXX: temporary solution to be able to assign fid in ->do_create()
2094 * till we have fully-functional OST fids
2096 struct lu_object *lu_object_anon(const struct lu_env *env,
2097 struct lu_device *dev,
2098 const struct lu_object_conf *conf)
2101 struct lu_object *o;
2104 o = lu_object_alloc(env, dev, &fid, conf);
2108 EXPORT_SYMBOL(lu_object_anon);
2110 struct lu_buf LU_BUF_NULL = {
2114 EXPORT_SYMBOL(LU_BUF_NULL);
2116 void lu_buf_free(struct lu_buf *buf)
2120 LASSERT(buf->lb_len > 0);
2121 OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);
2126 EXPORT_SYMBOL(lu_buf_free);
2128 void lu_buf_alloc(struct lu_buf *buf, int size)
2131 LASSERT(buf->lb_buf == NULL);
2132 LASSERT(buf->lb_len == 0);
2133 OBD_ALLOC_LARGE(buf->lb_buf, size);
2134 if (likely(buf->lb_buf))
2137 EXPORT_SYMBOL(lu_buf_alloc);
2139 void lu_buf_realloc(struct lu_buf *buf, int size)
2142 lu_buf_alloc(buf, size);
2144 EXPORT_SYMBOL(lu_buf_realloc);
2146 struct lu_buf *lu_buf_check_and_alloc(struct lu_buf *buf, int len)
2148 if (buf->lb_buf == NULL && buf->lb_len == 0)
2149 lu_buf_alloc(buf, len);
2151 if ((len > buf->lb_len) && (buf->lb_buf != NULL))
2152 lu_buf_realloc(buf, len);
2156 EXPORT_SYMBOL(lu_buf_check_and_alloc);
2159 * Increase the size of the \a buf.
2160 * preserves old data in buffer
2161 * old buffer remains unchanged on error
2162 * \retval 0 or -ENOMEM
2164 int lu_buf_check_and_grow(struct lu_buf *buf, int len)
2168 if (len <= buf->lb_len)
2171 OBD_ALLOC_LARGE(ptr, len);
2175 /* Free the old buf */
2176 if (buf->lb_buf != NULL) {
2177 memcpy(ptr, buf->lb_buf, buf->lb_len);
2178 OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);
2185 EXPORT_SYMBOL(lu_buf_check_and_grow);