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>
50 # include <linux/module.h>
54 #include <libcfs/libcfs_hash.h>
55 #include <obd_class.h>
56 #include <obd_support.h>
57 #include <lustre_disk.h>
58 #include <lustre_fid.h>
59 #include <lu_object.h>
61 #include <libcfs/list.h>
63 static void lu_object_free(const struct lu_env *env, struct lu_object *o);
66 * Decrease reference counter on object. If last reference is freed, return
67 * object to the cache, unless lu_object_is_dying(o) holds. In the latter
68 * case, free object immediately.
70 void lu_object_put(const struct lu_env *env, struct lu_object *o)
72 struct lu_site_bkt_data *bkt;
73 struct lu_object_header *top;
75 struct lu_object *orig;
77 const struct lu_fid *fid;
80 site = o->lo_dev->ld_site;
84 * till we have full fids-on-OST implemented anonymous objects
85 * are possible in OSP. such an object isn't listed in the site
86 * so we should not remove it from the site.
88 fid = lu_object_fid(o);
89 if (fid_is_zero(fid)) {
90 LASSERT(top->loh_hash.next == NULL
91 && top->loh_hash.pprev == NULL);
92 LASSERT(cfs_list_empty(&top->loh_lru));
93 if (!cfs_atomic_dec_and_test(&top->loh_ref))
95 cfs_list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
96 if (o->lo_ops->loo_object_release != NULL)
97 o->lo_ops->loo_object_release(env, o);
99 lu_object_free(env, orig);
103 cfs_hash_bd_get(site->ls_obj_hash, &top->loh_fid, &bd);
104 bkt = cfs_hash_bd_extra_get(site->ls_obj_hash, &bd);
106 if (!cfs_hash_bd_dec_and_lock(site->ls_obj_hash, &bd, &top->loh_ref)) {
107 if (lu_object_is_dying(top)) {
110 * somebody may be waiting for this, currently only
111 * used for cl_object, see cl_object_put_last().
113 cfs_waitq_broadcast(&bkt->lsb_marche_funebre);
118 LASSERT(bkt->lsb_busy > 0);
121 * When last reference is released, iterate over object
122 * layers, and notify them that object is no longer busy.
124 cfs_list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
125 if (o->lo_ops->loo_object_release != NULL)
126 o->lo_ops->loo_object_release(env, o);
129 if (!lu_object_is_dying(top)) {
130 LASSERT(cfs_list_empty(&top->loh_lru));
131 cfs_list_add_tail(&top->loh_lru, &bkt->lsb_lru);
132 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
137 * If object is dying (will not be cached), removed it
138 * from hash table and LRU.
140 * This is done with hash table and LRU lists locked. As the only
141 * way to acquire first reference to previously unreferenced
142 * object is through hash-table lookup (lu_object_find()),
143 * or LRU scanning (lu_site_purge()), that are done under hash-table
144 * and LRU lock, no race with concurrent object lookup is possible
145 * and we can safely destroy object below.
147 if (!test_and_set_bit(LU_OBJECT_UNHASHED, &top->loh_flags))
148 cfs_hash_bd_del_locked(site->ls_obj_hash, &bd, &top->loh_hash);
149 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
151 * Object was already removed from hash and lru above, can
154 lu_object_free(env, orig);
156 EXPORT_SYMBOL(lu_object_put);
159 * Put object and don't keep in cache. This is temporary solution for
160 * multi-site objects when its layering is not constant.
162 void lu_object_put_nocache(const struct lu_env *env, struct lu_object *o)
164 set_bit(LU_OBJECT_HEARD_BANSHEE, &o->lo_header->loh_flags);
165 return lu_object_put(env, o);
167 EXPORT_SYMBOL(lu_object_put_nocache);
170 * Kill the object and take it out of LRU cache.
171 * Currently used by client code for layout change.
173 void lu_object_unhash(const struct lu_env *env, struct lu_object *o)
175 struct lu_object_header *top;
178 set_bit(LU_OBJECT_HEARD_BANSHEE, &top->loh_flags);
179 if (!test_and_set_bit(LU_OBJECT_UNHASHED, &top->loh_flags)) {
180 cfs_hash_t *obj_hash = o->lo_dev->ld_site->ls_obj_hash;
183 cfs_hash_bd_get_and_lock(obj_hash, &top->loh_fid, &bd, 1);
184 cfs_list_del_init(&top->loh_lru);
185 cfs_hash_bd_del_locked(obj_hash, &bd, &top->loh_hash);
186 cfs_hash_bd_unlock(obj_hash, &bd, 1);
189 EXPORT_SYMBOL(lu_object_unhash);
192 * Allocate new object.
194 * This follows object creation protocol, described in the comment within
195 * struct lu_device_operations definition.
197 static struct lu_object *lu_object_alloc(const struct lu_env *env,
198 struct lu_device *dev,
199 const struct lu_fid *f,
200 const struct lu_object_conf *conf)
202 struct lu_object *scan;
203 struct lu_object *top;
205 unsigned int init_mask = 0;
206 unsigned int init_flag;
212 * Create top-level object slice. This will also create
215 top = dev->ld_ops->ldo_object_alloc(env, NULL, dev);
217 RETURN(ERR_PTR(-ENOMEM));
221 * This is the only place where object fid is assigned. It's constant
224 top->lo_header->loh_fid = *f;
225 layers = &top->lo_header->loh_layers;
229 * Call ->loo_object_init() repeatedly, until no more new
230 * object slices are created.
234 cfs_list_for_each_entry(scan, layers, lo_linkage) {
235 if (init_mask & init_flag)
238 scan->lo_header = top->lo_header;
239 result = scan->lo_ops->loo_object_init(env, scan, conf);
241 lu_object_free(env, top);
242 RETURN(ERR_PTR(result));
244 init_mask |= init_flag;
250 cfs_list_for_each_entry_reverse(scan, layers, lo_linkage) {
251 if (scan->lo_ops->loo_object_start != NULL) {
252 result = scan->lo_ops->loo_object_start(env, scan);
254 lu_object_free(env, top);
255 RETURN(ERR_PTR(result));
260 lprocfs_counter_incr(dev->ld_site->ls_stats, LU_SS_CREATED);
267 static void lu_object_free(const struct lu_env *env, struct lu_object *o)
269 struct lu_site_bkt_data *bkt;
270 struct lu_site *site;
271 struct lu_object *scan;
275 site = o->lo_dev->ld_site;
276 layers = &o->lo_header->loh_layers;
277 bkt = lu_site_bkt_from_fid(site, &o->lo_header->loh_fid);
279 * First call ->loo_object_delete() method to release all resources.
281 cfs_list_for_each_entry_reverse(scan, layers, lo_linkage) {
282 if (scan->lo_ops->loo_object_delete != NULL)
283 scan->lo_ops->loo_object_delete(env, scan);
287 * Then, splice object layers into stand-alone list, and call
288 * ->loo_object_free() on all layers to free memory. Splice is
289 * necessary, because lu_object_header is freed together with the
292 CFS_INIT_LIST_HEAD(&splice);
293 cfs_list_splice_init(layers, &splice);
294 while (!cfs_list_empty(&splice)) {
296 * Free layers in bottom-to-top order, so that object header
297 * lives as long as possible and ->loo_object_free() methods
298 * can look at its contents.
300 o = container_of0(splice.prev, struct lu_object, lo_linkage);
301 cfs_list_del_init(&o->lo_linkage);
302 LASSERT(o->lo_ops->loo_object_free != NULL);
303 o->lo_ops->loo_object_free(env, o);
306 if (cfs_waitq_active(&bkt->lsb_marche_funebre))
307 cfs_waitq_broadcast(&bkt->lsb_marche_funebre);
311 * Free \a nr objects from the cold end of the site LRU list.
313 int lu_site_purge(const struct lu_env *env, struct lu_site *s, int nr)
315 struct lu_object_header *h;
316 struct lu_object_header *temp;
317 struct lu_site_bkt_data *bkt;
327 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_NO_LRU))
330 CFS_INIT_LIST_HEAD(&dispose);
332 * Under LRU list lock, scan LRU list and move unreferenced objects to
333 * the dispose list, removing them from LRU and hash table.
335 start = s->ls_purge_start;
336 bnr = (nr == ~0) ? -1 : nr / CFS_HASH_NBKT(s->ls_obj_hash) + 1;
339 cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
343 cfs_hash_bd_lock(s->ls_obj_hash, &bd, 1);
344 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
346 cfs_list_for_each_entry_safe(h, temp, &bkt->lsb_lru, loh_lru) {
347 LASSERT(cfs_atomic_read(&h->loh_ref) == 0);
349 cfs_hash_bd_get(s->ls_obj_hash, &h->loh_fid, &bd2);
350 LASSERT(bd.bd_bucket == bd2.bd_bucket);
352 cfs_hash_bd_del_locked(s->ls_obj_hash,
354 cfs_list_move(&h->loh_lru, &dispose);
358 if (nr != ~0 && --nr == 0)
361 if (count > 0 && --count == 0)
365 cfs_hash_bd_unlock(s->ls_obj_hash, &bd, 1);
368 * Free everything on the dispose list. This is safe against
369 * races due to the reasons described in lu_object_put().
371 while (!cfs_list_empty(&dispose)) {
372 h = container_of0(dispose.next,
373 struct lu_object_header, loh_lru);
374 cfs_list_del_init(&h->loh_lru);
375 lu_object_free(env, lu_object_top(h));
376 lprocfs_counter_incr(s->ls_stats, LU_SS_LRU_PURGED);
383 if (nr != 0 && did_sth && start != 0) {
384 start = 0; /* restart from the first bucket */
387 /* race on s->ls_purge_start, but nobody cares */
388 s->ls_purge_start = i % CFS_HASH_NBKT(s->ls_obj_hash);
392 EXPORT_SYMBOL(lu_site_purge);
397 * Code below has to jump through certain loops to output object description
398 * into libcfs_debug_msg-based log. The problem is that lu_object_print()
399 * composes object description from strings that are parts of _lines_ of
400 * output (i.e., strings that are not terminated by newline). This doesn't fit
401 * very well into libcfs_debug_msg() interface that assumes that each message
402 * supplied to it is a self-contained output line.
404 * To work around this, strings are collected in a temporary buffer
405 * (implemented as a value of lu_cdebug_key key), until terminating newline
406 * character is detected.
414 * XXX overflow is not handled correctly.
419 struct lu_cdebug_data {
423 char lck_area[LU_CDEBUG_LINE];
426 /* context key constructor/destructor: lu_global_key_init, lu_global_key_fini */
427 LU_KEY_INIT_FINI(lu_global, struct lu_cdebug_data);
430 * Key, holding temporary buffer. This key is registered very early by
433 struct lu_context_key lu_global_key = {
434 .lct_tags = LCT_MD_THREAD | LCT_DT_THREAD |
435 LCT_MG_THREAD | LCT_CL_THREAD | LCT_LOCAL,
436 .lct_init = lu_global_key_init,
437 .lct_fini = lu_global_key_fini
441 * Printer function emitting messages through libcfs_debug_msg().
443 int lu_cdebug_printer(const struct lu_env *env,
444 void *cookie, const char *format, ...)
446 struct libcfs_debug_msg_data *msgdata = cookie;
447 struct lu_cdebug_data *key;
452 va_start(args, format);
454 key = lu_context_key_get(&env->le_ctx, &lu_global_key);
455 LASSERT(key != NULL);
457 used = strlen(key->lck_area);
458 complete = format[strlen(format) - 1] == '\n';
460 * Append new chunk to the buffer.
462 vsnprintf(key->lck_area + used,
463 ARRAY_SIZE(key->lck_area) - used, format, args);
465 if (cfs_cdebug_show(msgdata->msg_mask, msgdata->msg_subsys))
466 libcfs_debug_msg(msgdata, "%s", key->lck_area);
467 key->lck_area[0] = 0;
472 EXPORT_SYMBOL(lu_cdebug_printer);
475 * Print object header.
477 void lu_object_header_print(const struct lu_env *env, void *cookie,
478 lu_printer_t printer,
479 const struct lu_object_header *hdr)
481 (*printer)(env, cookie, "header@%p[%#lx, %d, "DFID"%s%s%s]",
482 hdr, hdr->loh_flags, cfs_atomic_read(&hdr->loh_ref),
484 cfs_hlist_unhashed(&hdr->loh_hash) ? "" : " hash",
485 cfs_list_empty((cfs_list_t *)&hdr->loh_lru) ? \
487 hdr->loh_attr & LOHA_EXISTS ? " exist":"");
489 EXPORT_SYMBOL(lu_object_header_print);
492 * Print human readable representation of the \a o to the \a printer.
494 void lu_object_print(const struct lu_env *env, void *cookie,
495 lu_printer_t printer, const struct lu_object *o)
497 static const char ruler[] = "........................................";
498 struct lu_object_header *top;
502 lu_object_header_print(env, cookie, printer, top);
503 (*printer)(env, cookie, "{\n");
505 cfs_list_for_each_entry(o, &top->loh_layers, lo_linkage) {
507 * print `.' \a depth times followed by type name and address
509 (*printer)(env, cookie, "%*.*s%s@%p", depth, depth, ruler,
510 o->lo_dev->ld_type->ldt_name, o);
512 if (o->lo_ops->loo_object_print != NULL)
513 (*o->lo_ops->loo_object_print)(env, cookie, printer, o);
515 (*printer)(env, cookie, "\n");
518 (*printer)(env, cookie, "} header@%p\n", top);
520 EXPORT_SYMBOL(lu_object_print);
523 * Check object consistency.
525 int lu_object_invariant(const struct lu_object *o)
527 struct lu_object_header *top;
530 cfs_list_for_each_entry(o, &top->loh_layers, lo_linkage) {
531 if (o->lo_ops->loo_object_invariant != NULL &&
532 !o->lo_ops->loo_object_invariant(o))
537 EXPORT_SYMBOL(lu_object_invariant);
539 static struct lu_object *htable_lookup(struct lu_site *s,
541 const struct lu_fid *f,
542 cfs_waitlink_t *waiter,
545 struct lu_site_bkt_data *bkt;
546 struct lu_object_header *h;
547 cfs_hlist_node_t *hnode;
548 __u64 ver = cfs_hash_bd_version_get(bd);
551 return ERR_PTR(-ENOENT);
554 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, bd);
555 /* cfs_hash_bd_peek_locked is a somehow "internal" function
556 * of cfs_hash, it doesn't add refcount on object. */
557 hnode = cfs_hash_bd_peek_locked(s->ls_obj_hash, bd, (void *)f);
559 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS);
560 return ERR_PTR(-ENOENT);
563 h = container_of0(hnode, struct lu_object_header, loh_hash);
564 if (likely(!lu_object_is_dying(h))) {
565 cfs_hash_get(s->ls_obj_hash, hnode);
566 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_HIT);
567 cfs_list_del_init(&h->loh_lru);
568 return lu_object_top(h);
572 * Lookup found an object being destroyed this object cannot be
573 * returned (to assure that references to dying objects are eventually
574 * drained), and moreover, lookup has to wait until object is freed.
577 cfs_waitlink_init(waiter);
578 cfs_waitq_add(&bkt->lsb_marche_funebre, waiter);
579 cfs_set_current_state(CFS_TASK_UNINT);
580 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_DEATH_RACE);
581 return ERR_PTR(-EAGAIN);
584 static struct lu_object *htable_lookup_nowait(struct lu_site *s,
586 const struct lu_fid *f)
588 cfs_hlist_node_t *hnode;
589 struct lu_object_header *h;
591 /* cfs_hash_bd_peek_locked is a somehow "internal" function
592 * of cfs_hash, it doesn't add refcount on object. */
593 hnode = cfs_hash_bd_peek_locked(s->ls_obj_hash, bd, (void *)f);
595 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS);
596 return ERR_PTR(-ENOENT);
599 h = container_of0(hnode, struct lu_object_header, loh_hash);
600 if (unlikely(lu_object_is_dying(h)))
601 return ERR_PTR(-ENOENT);
603 cfs_hash_get(s->ls_obj_hash, hnode);
604 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_HIT);
605 cfs_list_del_init(&h->loh_lru);
606 return lu_object_top(h);
610 * Search cache for an object with the fid \a f. If such object is found,
611 * return it. Otherwise, create new object, insert it into cache and return
612 * it. In any case, additional reference is acquired on the returned object.
614 struct lu_object *lu_object_find(const struct lu_env *env,
615 struct lu_device *dev, const struct lu_fid *f,
616 const struct lu_object_conf *conf)
618 return lu_object_find_at(env, dev->ld_site->ls_top_dev, f, conf);
620 EXPORT_SYMBOL(lu_object_find);
622 static struct lu_object *lu_object_new(const struct lu_env *env,
623 struct lu_device *dev,
624 const struct lu_fid *f,
625 const struct lu_object_conf *conf)
630 struct lu_site_bkt_data *bkt;
632 o = lu_object_alloc(env, dev, f, conf);
633 if (unlikely(IS_ERR(o)))
636 hs = dev->ld_site->ls_obj_hash;
637 cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
638 bkt = cfs_hash_bd_extra_get(hs, &bd);
639 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
641 cfs_hash_bd_unlock(hs, &bd, 1);
646 * Core logic of lu_object_find*() functions.
648 static struct lu_object *lu_object_find_try(const struct lu_env *env,
649 struct lu_device *dev,
650 const struct lu_fid *f,
651 const struct lu_object_conf *conf,
652 cfs_waitlink_t *waiter)
655 struct lu_object *shadow;
662 * This uses standard index maintenance protocol:
664 * - search index under lock, and return object if found;
665 * - otherwise, unlock index, allocate new object;
666 * - lock index and search again;
667 * - if nothing is found (usual case), insert newly created
669 * - otherwise (race: other thread inserted object), free
670 * object just allocated.
674 * For "LOC_F_NEW" case, we are sure the object is new established.
675 * It is unnecessary to perform lookup-alloc-lookup-insert, instead,
676 * just alloc and insert directly.
678 * If dying object is found during index search, add @waiter to the
679 * site wait-queue and return ERR_PTR(-EAGAIN).
681 if (conf != NULL && conf->loc_flags & LOC_F_NEW)
682 return lu_object_new(env, dev, f, conf);
686 cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
687 o = htable_lookup(s, &bd, f, waiter, &version);
688 cfs_hash_bd_unlock(hs, &bd, 1);
689 if (!IS_ERR(o) || PTR_ERR(o) != -ENOENT)
693 * Allocate new object. This may result in rather complicated
694 * operations, including fld queries, inode loading, etc.
696 o = lu_object_alloc(env, dev, f, conf);
697 if (unlikely(IS_ERR(o)))
700 LASSERT(lu_fid_eq(lu_object_fid(o), f));
702 cfs_hash_bd_lock(hs, &bd, 1);
704 shadow = htable_lookup(s, &bd, f, waiter, &version);
705 if (likely(IS_ERR(shadow) && PTR_ERR(shadow) == -ENOENT)) {
706 struct lu_site_bkt_data *bkt;
708 bkt = cfs_hash_bd_extra_get(hs, &bd);
709 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
711 cfs_hash_bd_unlock(hs, &bd, 1);
715 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_RACE);
716 cfs_hash_bd_unlock(hs, &bd, 1);
717 lu_object_free(env, o);
722 * Much like lu_object_find(), but top level device of object is specifically
723 * \a dev rather than top level device of the site. This interface allows
724 * objects of different "stacking" to be created within the same site.
726 struct lu_object *lu_object_find_at(const struct lu_env *env,
727 struct lu_device *dev,
728 const struct lu_fid *f,
729 const struct lu_object_conf *conf)
731 struct lu_site_bkt_data *bkt;
732 struct lu_object *obj;
736 obj = lu_object_find_try(env, dev, f, conf, &wait);
737 if (obj != ERR_PTR(-EAGAIN))
740 * lu_object_find_try() already added waiter into the
743 cfs_waitq_wait(&wait, CFS_TASK_UNINT);
744 bkt = lu_site_bkt_from_fid(dev->ld_site, (void *)f);
745 cfs_waitq_del(&bkt->lsb_marche_funebre, &wait);
748 EXPORT_SYMBOL(lu_object_find_at);
751 * Try to find the object in cache without waiting for the dead object
752 * to be released nor allocating object if no cached one was found.
754 * The found object will be set as LU_OBJECT_HEARD_BANSHEE for purging.
756 void lu_object_purge(const struct lu_env *env, struct lu_device *dev,
757 const struct lu_fid *f)
759 struct lu_site *s = dev->ld_site;
760 cfs_hash_t *hs = s->ls_obj_hash;
764 cfs_hash_bd_get_and_lock(hs, f, &bd, 1);
765 o = htable_lookup_nowait(s, &bd, f);
766 cfs_hash_bd_unlock(hs, &bd, 1);
768 set_bit(LU_OBJECT_HEARD_BANSHEE, &o->lo_header->loh_flags);
769 lu_object_put(env, o);
772 EXPORT_SYMBOL(lu_object_purge);
775 * Find object with given fid, and return its slice belonging to given device.
777 struct lu_object *lu_object_find_slice(const struct lu_env *env,
778 struct lu_device *dev,
779 const struct lu_fid *f,
780 const struct lu_object_conf *conf)
782 struct lu_object *top;
783 struct lu_object *obj;
785 top = lu_object_find(env, dev, f, conf);
787 obj = lu_object_locate(top->lo_header, dev->ld_type);
789 lu_object_put(env, top);
794 EXPORT_SYMBOL(lu_object_find_slice);
797 * Global list of all device types.
799 static CFS_LIST_HEAD(lu_device_types);
801 int lu_device_type_init(struct lu_device_type *ldt)
805 CFS_INIT_LIST_HEAD(&ldt->ldt_linkage);
806 if (ldt->ldt_ops->ldto_init)
807 result = ldt->ldt_ops->ldto_init(ldt);
809 cfs_list_add(&ldt->ldt_linkage, &lu_device_types);
812 EXPORT_SYMBOL(lu_device_type_init);
814 void lu_device_type_fini(struct lu_device_type *ldt)
816 cfs_list_del_init(&ldt->ldt_linkage);
817 if (ldt->ldt_ops->ldto_fini)
818 ldt->ldt_ops->ldto_fini(ldt);
820 EXPORT_SYMBOL(lu_device_type_fini);
822 void lu_types_stop(void)
824 struct lu_device_type *ldt;
826 cfs_list_for_each_entry(ldt, &lu_device_types, ldt_linkage) {
827 if (ldt->ldt_device_nr == 0 && ldt->ldt_ops->ldto_stop)
828 ldt->ldt_ops->ldto_stop(ldt);
831 EXPORT_SYMBOL(lu_types_stop);
834 * Global list of all sites on this node
836 static CFS_LIST_HEAD(lu_sites);
837 static DEFINE_MUTEX(lu_sites_guard);
840 * Global environment used by site shrinker.
842 static struct lu_env lu_shrink_env;
844 struct lu_site_print_arg {
845 struct lu_env *lsp_env;
847 lu_printer_t lsp_printer;
851 lu_site_obj_print(cfs_hash_t *hs, cfs_hash_bd_t *bd,
852 cfs_hlist_node_t *hnode, void *data)
854 struct lu_site_print_arg *arg = (struct lu_site_print_arg *)data;
855 struct lu_object_header *h;
857 h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
858 if (!cfs_list_empty(&h->loh_layers)) {
859 const struct lu_object *o;
861 o = lu_object_top(h);
862 lu_object_print(arg->lsp_env, arg->lsp_cookie,
863 arg->lsp_printer, o);
865 lu_object_header_print(arg->lsp_env, arg->lsp_cookie,
866 arg->lsp_printer, h);
872 * Print all objects in \a s.
874 void lu_site_print(const struct lu_env *env, struct lu_site *s, void *cookie,
875 lu_printer_t printer)
877 struct lu_site_print_arg arg = {
878 .lsp_env = (struct lu_env *)env,
879 .lsp_cookie = cookie,
880 .lsp_printer = printer,
883 cfs_hash_for_each(s->ls_obj_hash, lu_site_obj_print, &arg);
885 EXPORT_SYMBOL(lu_site_print);
888 LU_CACHE_PERCENT_MAX = 50,
889 LU_CACHE_PERCENT_DEFAULT = 20
892 static unsigned int lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
893 CFS_MODULE_PARM(lu_cache_percent, "i", int, 0644,
894 "Percentage of memory to be used as lu_object cache");
897 * Return desired hash table order.
899 static int lu_htable_order(void)
901 unsigned long cache_size;
905 * Calculate hash table size, assuming that we want reasonable
906 * performance when 20% of total memory is occupied by cache of
909 * Size of lu_object is (arbitrary) taken as 1K (together with inode).
911 cache_size = num_physpages;
913 #if BITS_PER_LONG == 32
914 /* limit hashtable size for lowmem systems to low RAM */
915 if (cache_size > 1 << (30 - PAGE_CACHE_SHIFT))
916 cache_size = 1 << (30 - PAGE_CACHE_SHIFT) * 3 / 4;
919 /* clear off unreasonable cache setting. */
920 if (lu_cache_percent == 0 || lu_cache_percent > LU_CACHE_PERCENT_MAX) {
921 CWARN("obdclass: invalid lu_cache_percent: %u, it must be in"
922 " the range of (0, %u]. Will use default value: %u.\n",
923 lu_cache_percent, LU_CACHE_PERCENT_MAX,
924 LU_CACHE_PERCENT_DEFAULT);
926 lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
928 cache_size = cache_size / 100 * lu_cache_percent *
929 (PAGE_CACHE_SIZE / 1024);
931 for (bits = 1; (1 << bits) < cache_size; ++bits) {
937 static unsigned lu_obj_hop_hash(cfs_hash_t *hs,
938 const void *key, unsigned mask)
940 struct lu_fid *fid = (struct lu_fid *)key;
943 hash = fid_flatten32(fid);
944 hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
945 hash = cfs_hash_long(hash, hs->hs_bkt_bits);
947 /* give me another random factor */
948 hash -= cfs_hash_long((unsigned long)hs, fid_oid(fid) % 11 + 3);
950 hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
951 hash |= (fid_seq(fid) + fid_oid(fid)) & (CFS_HASH_NBKT(hs) - 1);
956 static void *lu_obj_hop_object(cfs_hlist_node_t *hnode)
958 return cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
961 static void *lu_obj_hop_key(cfs_hlist_node_t *hnode)
963 struct lu_object_header *h;
965 h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
969 static int lu_obj_hop_keycmp(const void *key, cfs_hlist_node_t *hnode)
971 struct lu_object_header *h;
973 h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
974 return lu_fid_eq(&h->loh_fid, (struct lu_fid *)key);
977 static void lu_obj_hop_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
979 struct lu_object_header *h;
981 h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
982 if (cfs_atomic_add_return(1, &h->loh_ref) == 1) {
983 struct lu_site_bkt_data *bkt;
986 cfs_hash_bd_get(hs, &h->loh_fid, &bd);
987 bkt = cfs_hash_bd_extra_get(hs, &bd);
992 static void lu_obj_hop_put_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
994 LBUG(); /* we should never called it */
997 cfs_hash_ops_t lu_site_hash_ops = {
998 .hs_hash = lu_obj_hop_hash,
999 .hs_key = lu_obj_hop_key,
1000 .hs_keycmp = lu_obj_hop_keycmp,
1001 .hs_object = lu_obj_hop_object,
1002 .hs_get = lu_obj_hop_get,
1003 .hs_put_locked = lu_obj_hop_put_locked,
1006 void lu_dev_add_linkage(struct lu_site *s, struct lu_device *d)
1008 spin_lock(&s->ls_ld_lock);
1009 if (cfs_list_empty(&d->ld_linkage))
1010 cfs_list_add(&d->ld_linkage, &s->ls_ld_linkage);
1011 spin_unlock(&s->ls_ld_lock);
1013 EXPORT_SYMBOL(lu_dev_add_linkage);
1015 void lu_dev_del_linkage(struct lu_site *s, struct lu_device *d)
1017 spin_lock(&s->ls_ld_lock);
1018 cfs_list_del_init(&d->ld_linkage);
1019 spin_unlock(&s->ls_ld_lock);
1021 EXPORT_SYMBOL(lu_dev_del_linkage);
1024 * Initialize site \a s, with \a d as the top level device.
1026 #define LU_SITE_BITS_MIN 12
1027 #define LU_SITE_BITS_MAX 24
1029 * total 256 buckets, we don't want too many buckets because:
1030 * - consume too much memory
1031 * - avoid unbalanced LRU list
1033 #define LU_SITE_BKT_BITS 8
1035 int lu_site_init(struct lu_site *s, struct lu_device *top)
1037 struct lu_site_bkt_data *bkt;
1044 memset(s, 0, sizeof *s);
1045 bits = lu_htable_order();
1046 snprintf(name, 16, "lu_site_%s", top->ld_type->ldt_name);
1047 for (bits = min(max(LU_SITE_BITS_MIN, bits), LU_SITE_BITS_MAX);
1048 bits >= LU_SITE_BITS_MIN; bits--) {
1049 s->ls_obj_hash = cfs_hash_create(name, bits, bits,
1050 bits - LU_SITE_BKT_BITS,
1053 CFS_HASH_SPIN_BKTLOCK |
1054 CFS_HASH_NO_ITEMREF |
1056 CFS_HASH_ASSERT_EMPTY);
1057 if (s->ls_obj_hash != NULL)
1061 if (s->ls_obj_hash == NULL) {
1062 CERROR("failed to create lu_site hash with bits: %d\n", bits);
1066 cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
1067 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
1068 CFS_INIT_LIST_HEAD(&bkt->lsb_lru);
1069 cfs_waitq_init(&bkt->lsb_marche_funebre);
1072 s->ls_stats = lprocfs_alloc_stats(LU_SS_LAST_STAT, 0);
1073 if (s->ls_stats == NULL) {
1074 cfs_hash_putref(s->ls_obj_hash);
1075 s->ls_obj_hash = NULL;
1079 lprocfs_counter_init(s->ls_stats, LU_SS_CREATED,
1080 0, "created", "created");
1081 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_HIT,
1082 0, "cache_hit", "cache_hit");
1083 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_MISS,
1084 0, "cache_miss", "cache_miss");
1085 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_RACE,
1086 0, "cache_race", "cache_race");
1087 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_DEATH_RACE,
1088 0, "cache_death_race", "cache_death_race");
1089 lprocfs_counter_init(s->ls_stats, LU_SS_LRU_PURGED,
1090 0, "lru_purged", "lru_purged");
1092 CFS_INIT_LIST_HEAD(&s->ls_linkage);
1093 s->ls_top_dev = top;
1096 lu_ref_add(&top->ld_reference, "site-top", s);
1098 CFS_INIT_LIST_HEAD(&s->ls_ld_linkage);
1099 spin_lock_init(&s->ls_ld_lock);
1101 lu_dev_add_linkage(s, top);
1105 EXPORT_SYMBOL(lu_site_init);
1108 * Finalize \a s and release its resources.
1110 void lu_site_fini(struct lu_site *s)
1112 mutex_lock(&lu_sites_guard);
1113 cfs_list_del_init(&s->ls_linkage);
1114 mutex_unlock(&lu_sites_guard);
1116 if (s->ls_obj_hash != NULL) {
1117 cfs_hash_putref(s->ls_obj_hash);
1118 s->ls_obj_hash = NULL;
1121 if (s->ls_top_dev != NULL) {
1122 s->ls_top_dev->ld_site = NULL;
1123 lu_ref_del(&s->ls_top_dev->ld_reference, "site-top", s);
1124 lu_device_put(s->ls_top_dev);
1125 s->ls_top_dev = NULL;
1128 if (s->ls_stats != NULL)
1129 lprocfs_free_stats(&s->ls_stats);
1131 EXPORT_SYMBOL(lu_site_fini);
1134 * Called when initialization of stack for this site is completed.
1136 int lu_site_init_finish(struct lu_site *s)
1139 mutex_lock(&lu_sites_guard);
1140 result = lu_context_refill(&lu_shrink_env.le_ctx);
1142 cfs_list_add(&s->ls_linkage, &lu_sites);
1143 mutex_unlock(&lu_sites_guard);
1146 EXPORT_SYMBOL(lu_site_init_finish);
1149 * Acquire additional reference on device \a d
1151 void lu_device_get(struct lu_device *d)
1153 cfs_atomic_inc(&d->ld_ref);
1155 EXPORT_SYMBOL(lu_device_get);
1158 * Release reference on device \a d.
1160 void lu_device_put(struct lu_device *d)
1162 LASSERT(cfs_atomic_read(&d->ld_ref) > 0);
1163 cfs_atomic_dec(&d->ld_ref);
1165 EXPORT_SYMBOL(lu_device_put);
1168 * Initialize device \a d of type \a t.
1170 int lu_device_init(struct lu_device *d, struct lu_device_type *t)
1172 if (t->ldt_device_nr++ == 0 && t->ldt_ops->ldto_start != NULL)
1173 t->ldt_ops->ldto_start(t);
1174 memset(d, 0, sizeof *d);
1175 cfs_atomic_set(&d->ld_ref, 0);
1177 lu_ref_init(&d->ld_reference);
1178 CFS_INIT_LIST_HEAD(&d->ld_linkage);
1181 EXPORT_SYMBOL(lu_device_init);
1184 * Finalize device \a d.
1186 void lu_device_fini(struct lu_device *d)
1188 struct lu_device_type *t;
1191 if (d->ld_obd != NULL) {
1192 d->ld_obd->obd_lu_dev = NULL;
1196 lu_ref_fini(&d->ld_reference);
1197 LASSERTF(cfs_atomic_read(&d->ld_ref) == 0,
1198 "Refcount is %u\n", cfs_atomic_read(&d->ld_ref));
1199 LASSERT(t->ldt_device_nr > 0);
1200 if (--t->ldt_device_nr == 0 && t->ldt_ops->ldto_stop != NULL)
1201 t->ldt_ops->ldto_stop(t);
1203 EXPORT_SYMBOL(lu_device_fini);
1206 * Initialize object \a o that is part of compound object \a h and was created
1209 int lu_object_init(struct lu_object *o, struct lu_object_header *h,
1210 struct lu_device *d)
1212 memset(o, 0, sizeof(*o));
1216 lu_ref_add_at(&d->ld_reference, &o->lo_dev_ref, "lu_object", o);
1217 CFS_INIT_LIST_HEAD(&o->lo_linkage);
1221 EXPORT_SYMBOL(lu_object_init);
1224 * Finalize object and release its resources.
1226 void lu_object_fini(struct lu_object *o)
1228 struct lu_device *dev = o->lo_dev;
1230 LASSERT(cfs_list_empty(&o->lo_linkage));
1233 lu_ref_del_at(&dev->ld_reference, &o->lo_dev_ref,
1239 EXPORT_SYMBOL(lu_object_fini);
1242 * Add object \a o as first layer of compound object \a h
1244 * This is typically called by the ->ldo_object_alloc() method of top-level
1247 void lu_object_add_top(struct lu_object_header *h, struct lu_object *o)
1249 cfs_list_move(&o->lo_linkage, &h->loh_layers);
1251 EXPORT_SYMBOL(lu_object_add_top);
1254 * Add object \a o as a layer of compound object, going after \a before.
1256 * This is typically called by the ->ldo_object_alloc() method of \a
1259 void lu_object_add(struct lu_object *before, struct lu_object *o)
1261 cfs_list_move(&o->lo_linkage, &before->lo_linkage);
1263 EXPORT_SYMBOL(lu_object_add);
1266 * Initialize compound object.
1268 int lu_object_header_init(struct lu_object_header *h)
1270 memset(h, 0, sizeof *h);
1271 cfs_atomic_set(&h->loh_ref, 1);
1272 CFS_INIT_HLIST_NODE(&h->loh_hash);
1273 CFS_INIT_LIST_HEAD(&h->loh_lru);
1274 CFS_INIT_LIST_HEAD(&h->loh_layers);
1275 lu_ref_init(&h->loh_reference);
1278 EXPORT_SYMBOL(lu_object_header_init);
1281 * Finalize compound object.
1283 void lu_object_header_fini(struct lu_object_header *h)
1285 LASSERT(cfs_list_empty(&h->loh_layers));
1286 LASSERT(cfs_list_empty(&h->loh_lru));
1287 LASSERT(cfs_hlist_unhashed(&h->loh_hash));
1288 lu_ref_fini(&h->loh_reference);
1290 EXPORT_SYMBOL(lu_object_header_fini);
1293 * Given a compound object, find its slice, corresponding to the device type
1296 struct lu_object *lu_object_locate(struct lu_object_header *h,
1297 const struct lu_device_type *dtype)
1299 struct lu_object *o;
1301 cfs_list_for_each_entry(o, &h->loh_layers, lo_linkage) {
1302 if (o->lo_dev->ld_type == dtype)
1307 EXPORT_SYMBOL(lu_object_locate);
1312 * Finalize and free devices in the device stack.
1314 * Finalize device stack by purging object cache, and calling
1315 * lu_device_type_operations::ldto_device_fini() and
1316 * lu_device_type_operations::ldto_device_free() on all devices in the stack.
1318 void lu_stack_fini(const struct lu_env *env, struct lu_device *top)
1320 struct lu_site *site = top->ld_site;
1321 struct lu_device *scan;
1322 struct lu_device *next;
1324 lu_site_purge(env, site, ~0);
1325 for (scan = top; scan != NULL; scan = next) {
1326 next = scan->ld_type->ldt_ops->ldto_device_fini(env, scan);
1327 lu_ref_del(&scan->ld_reference, "lu-stack", &lu_site_init);
1328 lu_device_put(scan);
1332 lu_site_purge(env, site, ~0);
1334 for (scan = top; scan != NULL; scan = next) {
1335 const struct lu_device_type *ldt = scan->ld_type;
1336 struct obd_type *type;
1338 next = ldt->ldt_ops->ldto_device_free(env, scan);
1339 type = ldt->ldt_obd_type;
1342 class_put_type(type);
1346 EXPORT_SYMBOL(lu_stack_fini);
1350 * Maximal number of tld slots.
1352 LU_CONTEXT_KEY_NR = 40
1355 static struct lu_context_key *lu_keys[LU_CONTEXT_KEY_NR] = { NULL, };
1357 static DEFINE_SPINLOCK(lu_keys_guard);
1360 * Global counter incremented whenever key is registered, unregistered,
1361 * revived or quiesced. This is used to void unnecessary calls to
1362 * lu_context_refill(). No locking is provided, as initialization and shutdown
1363 * are supposed to be externally serialized.
1365 static unsigned key_set_version = 0;
1370 int lu_context_key_register(struct lu_context_key *key)
1375 LASSERT(key->lct_init != NULL);
1376 LASSERT(key->lct_fini != NULL);
1377 LASSERT(key->lct_tags != 0);
1378 LASSERT(key->lct_owner != NULL);
1381 spin_lock(&lu_keys_guard);
1382 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1383 if (lu_keys[i] == NULL) {
1385 cfs_atomic_set(&key->lct_used, 1);
1387 lu_ref_init(&key->lct_reference);
1393 spin_unlock(&lu_keys_guard);
1396 EXPORT_SYMBOL(lu_context_key_register);
1398 static void key_fini(struct lu_context *ctx, int index)
1400 if (ctx->lc_value != NULL && ctx->lc_value[index] != NULL) {
1401 struct lu_context_key *key;
1403 key = lu_keys[index];
1404 LASSERT(key != NULL);
1405 LASSERT(key->lct_fini != NULL);
1406 LASSERT(cfs_atomic_read(&key->lct_used) > 1);
1408 key->lct_fini(ctx, key, ctx->lc_value[index]);
1409 lu_ref_del(&key->lct_reference, "ctx", ctx);
1410 cfs_atomic_dec(&key->lct_used);
1412 LASSERT(key->lct_owner != NULL);
1413 if ((ctx->lc_tags & LCT_NOREF) == 0) {
1414 LINVRNT(cfs_module_refcount(key->lct_owner) > 0);
1415 cfs_module_put(key->lct_owner);
1417 ctx->lc_value[index] = NULL;
1424 void lu_context_key_degister(struct lu_context_key *key)
1426 LASSERT(cfs_atomic_read(&key->lct_used) >= 1);
1427 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1429 lu_context_key_quiesce(key);
1432 spin_lock(&lu_keys_guard);
1433 key_fini(&lu_shrink_env.le_ctx, key->lct_index);
1434 if (lu_keys[key->lct_index]) {
1435 lu_keys[key->lct_index] = NULL;
1436 lu_ref_fini(&key->lct_reference);
1438 spin_unlock(&lu_keys_guard);
1440 LASSERTF(cfs_atomic_read(&key->lct_used) == 1,
1441 "key has instances: %d\n",
1442 cfs_atomic_read(&key->lct_used));
1444 EXPORT_SYMBOL(lu_context_key_degister);
1447 * Register a number of keys. This has to be called after all keys have been
1448 * initialized by a call to LU_CONTEXT_KEY_INIT().
1450 int lu_context_key_register_many(struct lu_context_key *k, ...)
1452 struct lu_context_key *key = k;
1458 result = lu_context_key_register(key);
1461 key = va_arg(args, struct lu_context_key *);
1462 } while (key != NULL);
1468 lu_context_key_degister(k);
1469 k = va_arg(args, struct lu_context_key *);
1476 EXPORT_SYMBOL(lu_context_key_register_many);
1479 * De-register a number of keys. This is a dual to
1480 * lu_context_key_register_many().
1482 void lu_context_key_degister_many(struct lu_context_key *k, ...)
1488 lu_context_key_degister(k);
1489 k = va_arg(args, struct lu_context_key*);
1490 } while (k != NULL);
1493 EXPORT_SYMBOL(lu_context_key_degister_many);
1496 * Revive a number of keys.
1498 void lu_context_key_revive_many(struct lu_context_key *k, ...)
1504 lu_context_key_revive(k);
1505 k = va_arg(args, struct lu_context_key*);
1506 } while (k != NULL);
1509 EXPORT_SYMBOL(lu_context_key_revive_many);
1512 * Quiescent a number of keys.
1514 void lu_context_key_quiesce_many(struct lu_context_key *k, ...)
1520 lu_context_key_quiesce(k);
1521 k = va_arg(args, struct lu_context_key*);
1522 } while (k != NULL);
1525 EXPORT_SYMBOL(lu_context_key_quiesce_many);
1528 * Return value associated with key \a key in context \a ctx.
1530 void *lu_context_key_get(const struct lu_context *ctx,
1531 const struct lu_context_key *key)
1533 LINVRNT(ctx->lc_state == LCS_ENTERED);
1534 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1535 LASSERT(lu_keys[key->lct_index] == key);
1536 return ctx->lc_value[key->lct_index];
1538 EXPORT_SYMBOL(lu_context_key_get);
1541 * List of remembered contexts. XXX document me.
1543 static CFS_LIST_HEAD(lu_context_remembered);
1546 * Destroy \a key in all remembered contexts. This is used to destroy key
1547 * values in "shared" contexts (like service threads), when a module owning
1548 * the key is about to be unloaded.
1550 void lu_context_key_quiesce(struct lu_context_key *key)
1552 struct lu_context *ctx;
1554 if (!(key->lct_tags & LCT_QUIESCENT)) {
1556 * XXX layering violation.
1558 key->lct_tags |= LCT_QUIESCENT;
1560 * XXX memory barrier has to go here.
1562 spin_lock(&lu_keys_guard);
1563 cfs_list_for_each_entry(ctx, &lu_context_remembered,
1565 key_fini(ctx, key->lct_index);
1566 spin_unlock(&lu_keys_guard);
1570 EXPORT_SYMBOL(lu_context_key_quiesce);
1572 void lu_context_key_revive(struct lu_context_key *key)
1574 key->lct_tags &= ~LCT_QUIESCENT;
1577 EXPORT_SYMBOL(lu_context_key_revive);
1579 static void keys_fini(struct lu_context *ctx)
1583 if (ctx->lc_value == NULL)
1586 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i)
1589 OBD_FREE(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1590 ctx->lc_value = NULL;
1593 static int keys_fill(struct lu_context *ctx)
1597 LINVRNT(ctx->lc_value != NULL);
1598 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1599 struct lu_context_key *key;
1602 if (ctx->lc_value[i] == NULL && key != NULL &&
1603 (key->lct_tags & ctx->lc_tags) &&
1605 * Don't create values for a LCT_QUIESCENT key, as this
1606 * will pin module owning a key.
1608 !(key->lct_tags & LCT_QUIESCENT)) {
1611 LINVRNT(key->lct_init != NULL);
1612 LINVRNT(key->lct_index == i);
1614 value = key->lct_init(ctx, key);
1615 if (unlikely(IS_ERR(value)))
1616 return PTR_ERR(value);
1618 LASSERT(key->lct_owner != NULL);
1619 if (!(ctx->lc_tags & LCT_NOREF))
1620 cfs_try_module_get(key->lct_owner);
1621 lu_ref_add_atomic(&key->lct_reference, "ctx", ctx);
1622 cfs_atomic_inc(&key->lct_used);
1624 * This is the only place in the code, where an
1625 * element of ctx->lc_value[] array is set to non-NULL
1628 ctx->lc_value[i] = value;
1629 if (key->lct_exit != NULL)
1630 ctx->lc_tags |= LCT_HAS_EXIT;
1632 ctx->lc_version = key_set_version;
1637 static int keys_init(struct lu_context *ctx)
1639 OBD_ALLOC(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1640 if (likely(ctx->lc_value != NULL))
1641 return keys_fill(ctx);
1647 * Initialize context data-structure. Create values for all keys.
1649 int lu_context_init(struct lu_context *ctx, __u32 tags)
1653 memset(ctx, 0, sizeof *ctx);
1654 ctx->lc_state = LCS_INITIALIZED;
1655 ctx->lc_tags = tags;
1656 if (tags & LCT_REMEMBER) {
1657 spin_lock(&lu_keys_guard);
1658 cfs_list_add(&ctx->lc_remember, &lu_context_remembered);
1659 spin_unlock(&lu_keys_guard);
1661 CFS_INIT_LIST_HEAD(&ctx->lc_remember);
1664 rc = keys_init(ctx);
1666 lu_context_fini(ctx);
1670 EXPORT_SYMBOL(lu_context_init);
1673 * Finalize context data-structure. Destroy key values.
1675 void lu_context_fini(struct lu_context *ctx)
1677 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1678 ctx->lc_state = LCS_FINALIZED;
1680 if ((ctx->lc_tags & LCT_REMEMBER) == 0) {
1681 LASSERT(cfs_list_empty(&ctx->lc_remember));
1684 } else { /* could race with key degister */
1685 spin_lock(&lu_keys_guard);
1687 cfs_list_del_init(&ctx->lc_remember);
1688 spin_unlock(&lu_keys_guard);
1691 EXPORT_SYMBOL(lu_context_fini);
1694 * Called before entering context.
1696 void lu_context_enter(struct lu_context *ctx)
1698 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1699 ctx->lc_state = LCS_ENTERED;
1701 EXPORT_SYMBOL(lu_context_enter);
1704 * Called after exiting from \a ctx
1706 void lu_context_exit(struct lu_context *ctx)
1710 LINVRNT(ctx->lc_state == LCS_ENTERED);
1711 ctx->lc_state = LCS_LEFT;
1712 if (ctx->lc_tags & LCT_HAS_EXIT && ctx->lc_value != NULL) {
1713 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1714 if (ctx->lc_value[i] != NULL) {
1715 struct lu_context_key *key;
1718 LASSERT(key != NULL);
1719 if (key->lct_exit != NULL)
1721 key, ctx->lc_value[i]);
1726 EXPORT_SYMBOL(lu_context_exit);
1729 * Allocate for context all missing keys that were registered after context
1730 * creation. key_set_version is only changed in rare cases when modules
1731 * are loaded and removed.
1733 int lu_context_refill(struct lu_context *ctx)
1735 return likely(ctx->lc_version == key_set_version) ? 0 : keys_fill(ctx);
1737 EXPORT_SYMBOL(lu_context_refill);
1740 * lu_ctx_tags/lu_ses_tags will be updated if there are new types of
1741 * obd being added. Currently, this is only used on client side, specifically
1742 * for echo device client, for other stack (like ptlrpc threads), context are
1743 * predefined when the lu_device type are registered, during the module probe
1746 __u32 lu_context_tags_default = 0;
1747 __u32 lu_session_tags_default = 0;
1749 void lu_context_tags_update(__u32 tags)
1751 spin_lock(&lu_keys_guard);
1752 lu_context_tags_default |= tags;
1754 spin_unlock(&lu_keys_guard);
1756 EXPORT_SYMBOL(lu_context_tags_update);
1758 void lu_context_tags_clear(__u32 tags)
1760 spin_lock(&lu_keys_guard);
1761 lu_context_tags_default &= ~tags;
1763 spin_unlock(&lu_keys_guard);
1765 EXPORT_SYMBOL(lu_context_tags_clear);
1767 void lu_session_tags_update(__u32 tags)
1769 spin_lock(&lu_keys_guard);
1770 lu_session_tags_default |= tags;
1772 spin_unlock(&lu_keys_guard);
1774 EXPORT_SYMBOL(lu_session_tags_update);
1776 void lu_session_tags_clear(__u32 tags)
1778 spin_lock(&lu_keys_guard);
1779 lu_session_tags_default &= ~tags;
1781 spin_unlock(&lu_keys_guard);
1783 EXPORT_SYMBOL(lu_session_tags_clear);
1785 int lu_env_init(struct lu_env *env, __u32 tags)
1790 result = lu_context_init(&env->le_ctx, tags);
1791 if (likely(result == 0))
1792 lu_context_enter(&env->le_ctx);
1795 EXPORT_SYMBOL(lu_env_init);
1797 void lu_env_fini(struct lu_env *env)
1799 lu_context_exit(&env->le_ctx);
1800 lu_context_fini(&env->le_ctx);
1803 EXPORT_SYMBOL(lu_env_fini);
1805 int lu_env_refill(struct lu_env *env)
1809 result = lu_context_refill(&env->le_ctx);
1810 if (result == 0 && env->le_ses != NULL)
1811 result = lu_context_refill(env->le_ses);
1814 EXPORT_SYMBOL(lu_env_refill);
1817 * Currently, this API will only be used by echo client.
1818 * Because echo client and normal lustre client will share
1819 * same cl_env cache. So echo client needs to refresh
1820 * the env context after it get one from the cache, especially
1821 * when normal client and echo client co-exist in the same client.
1823 int lu_env_refill_by_tags(struct lu_env *env, __u32 ctags,
1828 if ((env->le_ctx.lc_tags & ctags) != ctags) {
1829 env->le_ctx.lc_version = 0;
1830 env->le_ctx.lc_tags |= ctags;
1833 if (env->le_ses && (env->le_ses->lc_tags & stags) != stags) {
1834 env->le_ses->lc_version = 0;
1835 env->le_ses->lc_tags |= stags;
1838 result = lu_env_refill(env);
1842 EXPORT_SYMBOL(lu_env_refill_by_tags);
1844 static struct shrinker *lu_site_shrinker;
1846 typedef struct lu_site_stats{
1847 unsigned lss_populated;
1848 unsigned lss_max_search;
1853 static void lu_site_stats_get(cfs_hash_t *hs,
1854 lu_site_stats_t *stats, int populated)
1859 cfs_hash_for_each_bucket(hs, &bd, i) {
1860 struct lu_site_bkt_data *bkt = cfs_hash_bd_extra_get(hs, &bd);
1861 cfs_hlist_head_t *hhead;
1863 cfs_hash_bd_lock(hs, &bd, 1);
1864 stats->lss_busy += bkt->lsb_busy;
1865 stats->lss_total += cfs_hash_bd_count_get(&bd);
1866 stats->lss_max_search = max((int)stats->lss_max_search,
1867 cfs_hash_bd_depmax_get(&bd));
1869 cfs_hash_bd_unlock(hs, &bd, 1);
1873 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
1874 if (!cfs_hlist_empty(hhead))
1875 stats->lss_populated++;
1877 cfs_hash_bd_unlock(hs, &bd, 1);
1884 * There exists a potential lock inversion deadlock scenario when using
1885 * Lustre on top of ZFS. This occurs between one of ZFS's
1886 * buf_hash_table.ht_lock's, and Lustre's lu_sites_guard lock. Essentially,
1887 * thread A will take the lu_sites_guard lock and sleep on the ht_lock,
1888 * while thread B will take the ht_lock and sleep on the lu_sites_guard
1889 * lock. Obviously neither thread will wake and drop their respective hold
1892 * To prevent this from happening we must ensure the lu_sites_guard lock is
1893 * not taken while down this code path. ZFS reliably does not set the
1894 * __GFP_FS bit in its code paths, so this can be used to determine if it
1895 * is safe to take the lu_sites_guard lock.
1897 * Ideally we should accurately return the remaining number of cached
1898 * objects without taking the lu_sites_guard lock, but this is not
1899 * possible in the current implementation.
1901 static int lu_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
1903 lu_site_stats_t stats;
1905 struct lu_site *tmp;
1907 int remain = shrink_param(sc, nr_to_scan);
1908 CFS_LIST_HEAD(splice);
1910 if (!(shrink_param(sc, gfp_mask) & __GFP_FS)) {
1914 /* We must not take the lu_sites_guard lock when
1915 * __GFP_FS is *not* set because of the deadlock
1916 * possibility detailed above. Additionally,
1917 * since we cannot determine the number of
1918 * objects in the cache without taking this
1919 * lock, we're in a particularly tough spot. As
1920 * a result, we'll just lie and say our cache is
1921 * empty. This _should_ be ok, as we can't
1922 * reclaim objects when __GFP_FS is *not* set
1928 CDEBUG(D_INODE, "Shrink %d objects\n", remain);
1930 mutex_lock(&lu_sites_guard);
1931 cfs_list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
1932 if (shrink_param(sc, nr_to_scan) != 0) {
1933 remain = lu_site_purge(&lu_shrink_env, s, remain);
1935 * Move just shrunk site to the tail of site list to
1936 * assure shrinking fairness.
1938 cfs_list_move_tail(&s->ls_linkage, &splice);
1941 memset(&stats, 0, sizeof(stats));
1942 lu_site_stats_get(s->ls_obj_hash, &stats, 0);
1943 cached += stats.lss_total - stats.lss_busy;
1944 if (shrink_param(sc, nr_to_scan) && remain <= 0)
1947 cfs_list_splice(&splice, lu_sites.prev);
1948 mutex_unlock(&lu_sites_guard);
1950 cached = (cached / 100) * sysctl_vfs_cache_pressure;
1951 if (shrink_param(sc, nr_to_scan) == 0)
1952 CDEBUG(D_INODE, "%d objects cached\n", cached);
1961 * Environment to be used in debugger, contains all tags.
1963 struct lu_env lu_debugging_env;
1966 * Debugging printer function using printk().
1968 int lu_printk_printer(const struct lu_env *env,
1969 void *unused, const char *format, ...)
1973 va_start(args, format);
1974 vprintk(format, args);
1979 int lu_debugging_setup(void)
1981 return lu_env_init(&lu_debugging_env, ~0);
1984 void lu_context_keys_dump(void)
1988 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1989 struct lu_context_key *key;
1993 CERROR("[%d]: %p %x (%p,%p,%p) %d %d \"%s\"@%p\n",
1994 i, key, key->lct_tags,
1995 key->lct_init, key->lct_fini, key->lct_exit,
1996 key->lct_index, cfs_atomic_read(&key->lct_used),
1997 key->lct_owner ? key->lct_owner->name : "",
1999 lu_ref_print(&key->lct_reference);
2003 EXPORT_SYMBOL(lu_context_keys_dump);
2004 #else /* !__KERNEL__ */
2005 static int lu_cache_shrink(int nr, unsigned int gfp_mask)
2009 #endif /* __KERNEL__ */
2012 * Initialization of global lu_* data.
2014 int lu_global_init(void)
2018 CDEBUG(D_INFO, "Lustre LU module (%p).\n", &lu_keys);
2020 result = lu_ref_global_init();
2024 LU_CONTEXT_KEY_INIT(&lu_global_key);
2025 result = lu_context_key_register(&lu_global_key);
2030 * At this level, we don't know what tags are needed, so allocate them
2031 * conservatively. This should not be too bad, because this
2032 * environment is global.
2034 mutex_lock(&lu_sites_guard);
2035 result = lu_env_init(&lu_shrink_env, LCT_SHRINKER);
2036 mutex_unlock(&lu_sites_guard);
2041 * seeks estimation: 3 seeks to read a record from oi, one to read
2042 * inode, one for ea. Unfortunately setting this high value results in
2043 * lu_object/inode cache consuming all the memory.
2045 lu_site_shrinker = set_shrinker(DEFAULT_SEEKS, lu_cache_shrink);
2046 if (lu_site_shrinker == NULL)
2053 * Dual to lu_global_init().
2055 void lu_global_fini(void)
2057 if (lu_site_shrinker != NULL) {
2058 remove_shrinker(lu_site_shrinker);
2059 lu_site_shrinker = NULL;
2062 lu_context_key_degister(&lu_global_key);
2065 * Tear shrinker environment down _after_ de-registering
2066 * lu_global_key, because the latter has a value in the former.
2068 mutex_lock(&lu_sites_guard);
2069 lu_env_fini(&lu_shrink_env);
2070 mutex_unlock(&lu_sites_guard);
2072 lu_ref_global_fini();
2075 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx)
2078 struct lprocfs_counter ret;
2080 lprocfs_stats_collect(stats, idx, &ret);
2081 return (__u32)ret.lc_count;
2088 * Output site statistical counters into a buffer. Suitable for
2089 * lprocfs_rd_*()-style functions.
2091 int lu_site_stats_print(const struct lu_site *s, char *page, int count)
2093 lu_site_stats_t stats;
2095 memset(&stats, 0, sizeof(stats));
2096 lu_site_stats_get(s->ls_obj_hash, &stats, 1);
2098 return snprintf(page, count, "%d/%d %d/%d %d %d %d %d %d %d %d\n",
2101 stats.lss_populated,
2102 CFS_HASH_NHLIST(s->ls_obj_hash),
2103 stats.lss_max_search,
2104 ls_stats_read(s->ls_stats, LU_SS_CREATED),
2105 ls_stats_read(s->ls_stats, LU_SS_CACHE_HIT),
2106 ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
2107 ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
2108 ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
2109 ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED));
2111 EXPORT_SYMBOL(lu_site_stats_print);
2114 * Helper function to initialize a number of kmem slab caches at once.
2116 int lu_kmem_init(struct lu_kmem_descr *caches)
2119 struct lu_kmem_descr *iter = caches;
2121 for (result = 0; iter->ckd_cache != NULL; ++iter) {
2122 *iter->ckd_cache = kmem_cache_create(iter->ckd_name,
2125 if (*iter->ckd_cache == NULL) {
2127 /* free all previously allocated caches */
2128 lu_kmem_fini(caches);
2134 EXPORT_SYMBOL(lu_kmem_init);
2137 * Helper function to finalize a number of kmem slab cached at once. Dual to
2140 void lu_kmem_fini(struct lu_kmem_descr *caches)
2142 for (; caches->ckd_cache != NULL; ++caches) {
2143 if (*caches->ckd_cache != NULL) {
2144 kmem_cache_destroy(*caches->ckd_cache);
2145 *caches->ckd_cache = NULL;
2149 EXPORT_SYMBOL(lu_kmem_fini);
2152 * Temporary solution to be able to assign fid in ->do_create()
2153 * till we have fully-functional OST fids
2155 void lu_object_assign_fid(const struct lu_env *env, struct lu_object *o,
2156 const struct lu_fid *fid)
2158 struct lu_site *s = o->lo_dev->ld_site;
2159 struct lu_fid *old = &o->lo_header->loh_fid;
2160 struct lu_site_bkt_data *bkt;
2161 struct lu_object *shadow;
2162 cfs_waitlink_t waiter;
2167 LASSERT(fid_is_zero(old));
2169 hs = s->ls_obj_hash;
2170 cfs_hash_bd_get_and_lock(hs, (void *)fid, &bd, 1);
2171 shadow = htable_lookup(s, &bd, fid, &waiter, &version);
2172 /* supposed to be unique */
2173 LASSERT(IS_ERR(shadow) && PTR_ERR(shadow) == -ENOENT);
2175 bkt = cfs_hash_bd_extra_get(hs, &bd);
2176 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
2178 cfs_hash_bd_unlock(hs, &bd, 1);
2180 EXPORT_SYMBOL(lu_object_assign_fid);
2183 * allocates object with 0 (non-assiged) fid
2184 * XXX: temporary solution to be able to assign fid in ->do_create()
2185 * till we have fully-functional OST fids
2187 struct lu_object *lu_object_anon(const struct lu_env *env,
2188 struct lu_device *dev,
2189 const struct lu_object_conf *conf)
2192 struct lu_object *o;
2195 o = lu_object_alloc(env, dev, &fid, conf);
2199 EXPORT_SYMBOL(lu_object_anon);
2201 struct lu_buf LU_BUF_NULL = {
2205 EXPORT_SYMBOL(LU_BUF_NULL);
2207 void lu_buf_free(struct lu_buf *buf)
2211 LASSERT(buf->lb_len > 0);
2212 OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);
2217 EXPORT_SYMBOL(lu_buf_free);
2219 void lu_buf_alloc(struct lu_buf *buf, int size)
2222 LASSERT(buf->lb_buf == NULL);
2223 LASSERT(buf->lb_len == 0);
2224 OBD_ALLOC_LARGE(buf->lb_buf, size);
2225 if (likely(buf->lb_buf))
2228 EXPORT_SYMBOL(lu_buf_alloc);
2230 void lu_buf_realloc(struct lu_buf *buf, int size)
2233 lu_buf_alloc(buf, size);
2235 EXPORT_SYMBOL(lu_buf_realloc);
2237 struct lu_buf *lu_buf_check_and_alloc(struct lu_buf *buf, int len)
2239 if (buf->lb_buf == NULL && buf->lb_len == 0)
2240 lu_buf_alloc(buf, len);
2242 if ((len > buf->lb_len) && (buf->lb_buf != NULL))
2243 lu_buf_realloc(buf, len);
2247 EXPORT_SYMBOL(lu_buf_check_and_alloc);
2250 * Increase the size of the \a buf.
2251 * preserves old data in buffer
2252 * old buffer remains unchanged on error
2253 * \retval 0 or -ENOMEM
2255 int lu_buf_check_and_grow(struct lu_buf *buf, int len)
2259 if (len <= buf->lb_len)
2262 OBD_ALLOC_LARGE(ptr, len);
2266 /* Free the old buf */
2267 if (buf->lb_buf != NULL) {
2268 memcpy(ptr, buf->lb_buf, buf->lb_len);
2269 OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);
2276 EXPORT_SYMBOL(lu_buf_check_and_grow);