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, Whamcloud, Inc.
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>
60 #include <libcfs/list.h>
61 /* lu_time_global_{init,fini}() */
64 static void lu_object_free(const struct lu_env *env, struct lu_object *o);
67 * Decrease reference counter on object. If last reference is freed, return
68 * object to the cache, unless lu_object_is_dying(o) holds. In the latter
69 * case, free object immediately.
71 void lu_object_put(const struct lu_env *env, struct lu_object *o)
73 struct lu_site_bkt_data *bkt;
74 struct lu_object_header *top;
76 struct lu_object *orig;
78 const struct lu_fid *fid;
81 site = o->lo_dev->ld_site;
85 * till we have full fids-on-OST implemented anonymous objects
86 * are possible in OSP. such an object isn't listed in the site
87 * so we should not remove it from the site.
89 fid = lu_object_fid(o);
90 if (fid_is_zero(fid)) {
91 LASSERT(top->loh_hash.next == NULL
92 && top->loh_hash.pprev == NULL);
93 LASSERT(cfs_list_empty(&top->loh_lru));
94 if (!cfs_atomic_dec_and_test(&top->loh_ref))
96 cfs_list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
97 if (o->lo_ops->loo_object_release != NULL)
98 o->lo_ops->loo_object_release(env, o);
100 lu_object_free(env, orig);
104 cfs_hash_bd_get(site->ls_obj_hash, &top->loh_fid, &bd);
105 bkt = cfs_hash_bd_extra_get(site->ls_obj_hash, &bd);
107 if (!cfs_hash_bd_dec_and_lock(site->ls_obj_hash, &bd, &top->loh_ref)) {
108 if (lu_object_is_dying(top)) {
111 * somebody may be waiting for this, currently only
112 * used for cl_object, see cl_object_put_last().
114 cfs_waitq_broadcast(&bkt->lsb_marche_funebre);
119 LASSERT(bkt->lsb_busy > 0);
122 * When last reference is released, iterate over object
123 * layers, and notify them that object is no longer busy.
125 cfs_list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
126 if (o->lo_ops->loo_object_release != NULL)
127 o->lo_ops->loo_object_release(env, o);
130 if (!lu_object_is_dying(top)) {
131 LASSERT(cfs_list_empty(&top->loh_lru));
132 cfs_list_add_tail(&top->loh_lru, &bkt->lsb_lru);
133 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
138 * If object is dying (will not be cached), removed it
139 * from hash table and LRU.
141 * This is done with hash table and LRU lists locked. As the only
142 * way to acquire first reference to previously unreferenced
143 * object is through hash-table lookup (lu_object_find()),
144 * or LRU scanning (lu_site_purge()), that are done under hash-table
145 * and LRU lock, no race with concurrent object lookup is possible
146 * and we can safely destroy object below.
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,
165 &o->lo_header->loh_flags);
166 return lu_object_put(env, o);
168 EXPORT_SYMBOL(lu_object_put_nocache);
171 * Allocate new object.
173 * This follows object creation protocol, described in the comment within
174 * struct lu_device_operations definition.
176 static struct lu_object *lu_object_alloc(const struct lu_env *env,
177 struct lu_device *dev,
178 const struct lu_fid *f,
179 const struct lu_object_conf *conf)
181 struct lu_object *scan;
182 struct lu_object *top;
189 * Create top-level object slice. This will also create
192 top = dev->ld_ops->ldo_object_alloc(env, NULL, dev);
194 RETURN(ERR_PTR(-ENOMEM));
196 * This is the only place where object fid is assigned. It's constant
199 LASSERT(fid_is_igif(f) || fid_ver(f) == 0);
200 top->lo_header->loh_fid = *f;
201 layers = &top->lo_header->loh_layers;
204 * Call ->loo_object_init() repeatedly, until no more new
205 * object slices are created.
208 cfs_list_for_each_entry(scan, layers, lo_linkage) {
209 if (scan->lo_flags & LU_OBJECT_ALLOCATED)
212 scan->lo_header = top->lo_header;
213 result = scan->lo_ops->loo_object_init(env, scan, conf);
215 lu_object_free(env, top);
216 RETURN(ERR_PTR(result));
218 scan->lo_flags |= LU_OBJECT_ALLOCATED;
222 cfs_list_for_each_entry_reverse(scan, layers, lo_linkage) {
223 if (scan->lo_ops->loo_object_start != NULL) {
224 result = scan->lo_ops->loo_object_start(env, scan);
226 lu_object_free(env, top);
227 RETURN(ERR_PTR(result));
232 lprocfs_counter_incr(dev->ld_site->ls_stats, LU_SS_CREATED);
239 static void lu_object_free(const struct lu_env *env, struct lu_object *o)
241 struct lu_site_bkt_data *bkt;
242 struct lu_site *site;
243 struct lu_object *scan;
247 site = o->lo_dev->ld_site;
248 layers = &o->lo_header->loh_layers;
249 bkt = lu_site_bkt_from_fid(site, &o->lo_header->loh_fid);
251 * First call ->loo_object_delete() method to release all resources.
253 cfs_list_for_each_entry_reverse(scan, layers, lo_linkage) {
254 if (scan->lo_ops->loo_object_delete != NULL)
255 scan->lo_ops->loo_object_delete(env, scan);
259 * Then, splice object layers into stand-alone list, and call
260 * ->loo_object_free() on all layers to free memory. Splice is
261 * necessary, because lu_object_header is freed together with the
264 CFS_INIT_LIST_HEAD(&splice);
265 cfs_list_splice_init(layers, &splice);
266 while (!cfs_list_empty(&splice)) {
268 * Free layers in bottom-to-top order, so that object header
269 * lives as long as possible and ->loo_object_free() methods
270 * can look at its contents.
272 o = container_of0(splice.prev, struct lu_object, lo_linkage);
273 cfs_list_del_init(&o->lo_linkage);
274 LASSERT(o->lo_ops->loo_object_free != NULL);
275 o->lo_ops->loo_object_free(env, o);
278 if (cfs_waitq_active(&bkt->lsb_marche_funebre))
279 cfs_waitq_broadcast(&bkt->lsb_marche_funebre);
283 * Free \a nr objects from the cold end of the site LRU list.
285 int lu_site_purge(const struct lu_env *env, struct lu_site *s, int nr)
287 struct lu_object_header *h;
288 struct lu_object_header *temp;
289 struct lu_site_bkt_data *bkt;
299 CFS_INIT_LIST_HEAD(&dispose);
301 * Under LRU list lock, scan LRU list and move unreferenced objects to
302 * the dispose list, removing them from LRU and hash table.
304 start = s->ls_purge_start;
305 bnr = (nr == ~0) ? -1 : nr / CFS_HASH_NBKT(s->ls_obj_hash) + 1;
308 cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
312 cfs_hash_bd_lock(s->ls_obj_hash, &bd, 1);
313 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
315 cfs_list_for_each_entry_safe(h, temp, &bkt->lsb_lru, loh_lru) {
316 LASSERT(cfs_atomic_read(&h->loh_ref) == 0);
318 cfs_hash_bd_get(s->ls_obj_hash, &h->loh_fid, &bd2);
319 LASSERT(bd.bd_bucket == bd2.bd_bucket);
321 cfs_hash_bd_del_locked(s->ls_obj_hash,
323 cfs_list_move(&h->loh_lru, &dispose);
327 if (nr != ~0 && --nr == 0)
330 if (count > 0 && --count == 0)
334 cfs_hash_bd_unlock(s->ls_obj_hash, &bd, 1);
337 * Free everything on the dispose list. This is safe against
338 * races due to the reasons described in lu_object_put().
340 while (!cfs_list_empty(&dispose)) {
341 h = container_of0(dispose.next,
342 struct lu_object_header, loh_lru);
343 cfs_list_del_init(&h->loh_lru);
344 lu_object_free(env, lu_object_top(h));
345 lprocfs_counter_incr(s->ls_stats, LU_SS_LRU_PURGED);
352 if (nr != 0 && did_sth && start != 0) {
353 start = 0; /* restart from the first bucket */
356 /* race on s->ls_purge_start, but nobody cares */
357 s->ls_purge_start = i % CFS_HASH_NBKT(s->ls_obj_hash);
361 EXPORT_SYMBOL(lu_site_purge);
366 * Code below has to jump through certain loops to output object description
367 * into libcfs_debug_msg-based log. The problem is that lu_object_print()
368 * composes object description from strings that are parts of _lines_ of
369 * output (i.e., strings that are not terminated by newline). This doesn't fit
370 * very well into libcfs_debug_msg() interface that assumes that each message
371 * supplied to it is a self-contained output line.
373 * To work around this, strings are collected in a temporary buffer
374 * (implemented as a value of lu_cdebug_key key), until terminating newline
375 * character is detected.
383 * XXX overflow is not handled correctly.
388 struct lu_cdebug_data {
392 char lck_area[LU_CDEBUG_LINE];
395 /* context key constructor/destructor: lu_global_key_init, lu_global_key_fini */
396 LU_KEY_INIT_FINI(lu_global, struct lu_cdebug_data);
399 * Key, holding temporary buffer. This key is registered very early by
402 struct lu_context_key lu_global_key = {
403 .lct_tags = LCT_MD_THREAD | LCT_DT_THREAD |
404 LCT_MG_THREAD | LCT_CL_THREAD,
405 .lct_init = lu_global_key_init,
406 .lct_fini = lu_global_key_fini
410 * Printer function emitting messages through libcfs_debug_msg().
412 int lu_cdebug_printer(const struct lu_env *env,
413 void *cookie, const char *format, ...)
415 struct libcfs_debug_msg_data *msgdata = cookie;
416 struct lu_cdebug_data *key;
421 va_start(args, format);
423 key = lu_context_key_get(&env->le_ctx, &lu_global_key);
424 LASSERT(key != NULL);
426 used = strlen(key->lck_area);
427 complete = format[strlen(format) - 1] == '\n';
429 * Append new chunk to the buffer.
431 vsnprintf(key->lck_area + used,
432 ARRAY_SIZE(key->lck_area) - used, format, args);
434 if (cfs_cdebug_show(msgdata->msg_mask, msgdata->msg_subsys))
435 libcfs_debug_msg(msgdata, "%s", key->lck_area);
436 key->lck_area[0] = 0;
441 EXPORT_SYMBOL(lu_cdebug_printer);
444 * Print object header.
446 void lu_object_header_print(const struct lu_env *env, void *cookie,
447 lu_printer_t printer,
448 const struct lu_object_header *hdr)
450 (*printer)(env, cookie, "header@%p[%#lx, %d, "DFID"%s%s%s]",
451 hdr, hdr->loh_flags, cfs_atomic_read(&hdr->loh_ref),
453 cfs_hlist_unhashed(&hdr->loh_hash) ? "" : " hash",
454 cfs_list_empty((cfs_list_t *)&hdr->loh_lru) ? \
456 hdr->loh_attr & LOHA_EXISTS ? " exist":"");
458 EXPORT_SYMBOL(lu_object_header_print);
461 * Print human readable representation of the \a o to the \a printer.
463 void lu_object_print(const struct lu_env *env, void *cookie,
464 lu_printer_t printer, const struct lu_object *o)
466 static const char ruler[] = "........................................";
467 struct lu_object_header *top;
471 lu_object_header_print(env, cookie, printer, top);
472 (*printer)(env, cookie, "{ \n");
473 cfs_list_for_each_entry(o, &top->loh_layers, lo_linkage) {
474 depth = o->lo_depth + 4;
477 * print `.' \a depth times followed by type name and address
479 (*printer)(env, cookie, "%*.*s%s@%p", depth, depth, ruler,
480 o->lo_dev->ld_type->ldt_name, o);
481 if (o->lo_ops->loo_object_print != NULL)
482 o->lo_ops->loo_object_print(env, cookie, printer, o);
483 (*printer)(env, cookie, "\n");
485 (*printer)(env, cookie, "} header@%p\n", top);
487 EXPORT_SYMBOL(lu_object_print);
490 * Check object consistency.
492 int lu_object_invariant(const struct lu_object *o)
494 struct lu_object_header *top;
497 cfs_list_for_each_entry(o, &top->loh_layers, lo_linkage) {
498 if (o->lo_ops->loo_object_invariant != NULL &&
499 !o->lo_ops->loo_object_invariant(o))
504 EXPORT_SYMBOL(lu_object_invariant);
506 static struct lu_object *htable_lookup(struct lu_site *s,
508 const struct lu_fid *f,
509 cfs_waitlink_t *waiter,
512 struct lu_site_bkt_data *bkt;
513 struct lu_object_header *h;
514 cfs_hlist_node_t *hnode;
515 __u64 ver = cfs_hash_bd_version_get(bd);
521 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, bd);
522 /* cfs_hash_bd_lookup_intent is a somehow "internal" function
523 * of cfs_hash, but we don't want refcount on object right now */
524 hnode = cfs_hash_bd_lookup_locked(s->ls_obj_hash, bd, (void *)f);
526 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS);
530 h = container_of0(hnode, struct lu_object_header, loh_hash);
531 if (likely(!lu_object_is_dying(h))) {
532 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_HIT);
533 cfs_list_del_init(&h->loh_lru);
534 return lu_object_top(h);
538 * Lookup found an object being destroyed this object cannot be
539 * returned (to assure that references to dying objects are eventually
540 * drained), and moreover, lookup has to wait until object is freed.
542 cfs_atomic_dec(&h->loh_ref);
544 cfs_waitlink_init(waiter);
545 cfs_waitq_add(&bkt->lsb_marche_funebre, waiter);
546 cfs_set_current_state(CFS_TASK_UNINT);
547 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_DEATH_RACE);
548 return ERR_PTR(-EAGAIN);
552 * Search cache for an object with the fid \a f. If such object is found,
553 * return it. Otherwise, create new object, insert it into cache and return
554 * it. In any case, additional reference is acquired on the returned object.
556 struct lu_object *lu_object_find(const struct lu_env *env,
557 struct lu_device *dev, const struct lu_fid *f,
558 const struct lu_object_conf *conf)
560 return lu_object_find_at(env, dev->ld_site->ls_top_dev, f, conf);
562 EXPORT_SYMBOL(lu_object_find);
564 static struct lu_object *lu_object_new(const struct lu_env *env,
565 struct lu_device *dev,
566 const struct lu_fid *f,
567 const struct lu_object_conf *conf)
572 struct lu_site_bkt_data *bkt;
574 o = lu_object_alloc(env, dev, f, conf);
575 if (unlikely(IS_ERR(o)))
578 hs = dev->ld_site->ls_obj_hash;
579 cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
580 bkt = cfs_hash_bd_extra_get(hs, &bd);
581 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
583 cfs_hash_bd_unlock(hs, &bd, 1);
588 * Core logic of lu_object_find*() functions.
590 static struct lu_object *lu_object_find_try(const struct lu_env *env,
591 struct lu_device *dev,
592 const struct lu_fid *f,
593 const struct lu_object_conf *conf,
594 cfs_waitlink_t *waiter)
597 struct lu_object *shadow;
604 * This uses standard index maintenance protocol:
606 * - search index under lock, and return object if found;
607 * - otherwise, unlock index, allocate new object;
608 * - lock index and search again;
609 * - if nothing is found (usual case), insert newly created
611 * - otherwise (race: other thread inserted object), free
612 * object just allocated.
616 * For "LOC_F_NEW" case, we are sure the object is new established.
617 * It is unnecessary to perform lookup-alloc-lookup-insert, instead,
618 * just alloc and insert directly.
620 * If dying object is found during index search, add @waiter to the
621 * site wait-queue and return ERR_PTR(-EAGAIN).
623 if (conf != NULL && conf->loc_flags & LOC_F_NEW)
624 return lu_object_new(env, dev, f, conf);
628 cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
629 o = htable_lookup(s, &bd, f, waiter, &version);
630 cfs_hash_bd_unlock(hs, &bd, 1);
635 * Allocate new object. This may result in rather complicated
636 * operations, including fld queries, inode loading, etc.
638 o = lu_object_alloc(env, dev, f, conf);
639 if (unlikely(IS_ERR(o)))
642 LASSERT(lu_fid_eq(lu_object_fid(o), f));
644 cfs_hash_bd_lock(hs, &bd, 1);
646 shadow = htable_lookup(s, &bd, f, waiter, &version);
647 if (likely(shadow == NULL)) {
648 struct lu_site_bkt_data *bkt;
650 bkt = cfs_hash_bd_extra_get(hs, &bd);
651 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
653 cfs_hash_bd_unlock(hs, &bd, 1);
657 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_RACE);
658 cfs_hash_bd_unlock(hs, &bd, 1);
659 lu_object_free(env, o);
664 * Much like lu_object_find(), but top level device of object is specifically
665 * \a dev rather than top level device of the site. This interface allows
666 * objects of different "stacking" to be created within the same site.
668 struct lu_object *lu_object_find_at(const struct lu_env *env,
669 struct lu_device *dev,
670 const struct lu_fid *f,
671 const struct lu_object_conf *conf)
673 struct lu_site_bkt_data *bkt;
674 struct lu_object *obj;
678 obj = lu_object_find_try(env, dev, f, conf, &wait);
679 if (obj != ERR_PTR(-EAGAIN))
682 * lu_object_find_try() already added waiter into the
685 cfs_waitq_wait(&wait, CFS_TASK_UNINT);
686 bkt = lu_site_bkt_from_fid(dev->ld_site, (void *)f);
687 cfs_waitq_del(&bkt->lsb_marche_funebre, &wait);
690 EXPORT_SYMBOL(lu_object_find_at);
693 * Find object with given fid, and return its slice belonging to given device.
695 struct lu_object *lu_object_find_slice(const struct lu_env *env,
696 struct lu_device *dev,
697 const struct lu_fid *f,
698 const struct lu_object_conf *conf)
700 struct lu_object *top;
701 struct lu_object *obj;
703 top = lu_object_find(env, dev, f, conf);
705 obj = lu_object_locate(top->lo_header, dev->ld_type);
707 lu_object_put(env, top);
712 EXPORT_SYMBOL(lu_object_find_slice);
715 * Global list of all device types.
717 static CFS_LIST_HEAD(lu_device_types);
719 int lu_device_type_init(struct lu_device_type *ldt)
723 CFS_INIT_LIST_HEAD(&ldt->ldt_linkage);
724 if (ldt->ldt_ops->ldto_init)
725 result = ldt->ldt_ops->ldto_init(ldt);
727 cfs_list_add(&ldt->ldt_linkage, &lu_device_types);
730 EXPORT_SYMBOL(lu_device_type_init);
732 void lu_device_type_fini(struct lu_device_type *ldt)
734 cfs_list_del_init(&ldt->ldt_linkage);
735 if (ldt->ldt_ops->ldto_fini)
736 ldt->ldt_ops->ldto_fini(ldt);
738 EXPORT_SYMBOL(lu_device_type_fini);
740 void lu_types_stop(void)
742 struct lu_device_type *ldt;
744 cfs_list_for_each_entry(ldt, &lu_device_types, ldt_linkage) {
745 if (ldt->ldt_device_nr == 0 && ldt->ldt_ops->ldto_stop)
746 ldt->ldt_ops->ldto_stop(ldt);
749 EXPORT_SYMBOL(lu_types_stop);
752 * Global list of all sites on this node
754 static CFS_LIST_HEAD(lu_sites);
755 static DEFINE_MUTEX(lu_sites_guard);
758 * Global environment used by site shrinker.
760 static struct lu_env lu_shrink_env;
762 struct lu_site_print_arg {
763 struct lu_env *lsp_env;
765 lu_printer_t lsp_printer;
769 lu_site_obj_print(cfs_hash_t *hs, cfs_hash_bd_t *bd,
770 cfs_hlist_node_t *hnode, void *data)
772 struct lu_site_print_arg *arg = (struct lu_site_print_arg *)data;
773 struct lu_object_header *h;
775 h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
776 if (!cfs_list_empty(&h->loh_layers)) {
777 const struct lu_object *o;
779 o = lu_object_top(h);
780 lu_object_print(arg->lsp_env, arg->lsp_cookie,
781 arg->lsp_printer, o);
783 lu_object_header_print(arg->lsp_env, arg->lsp_cookie,
784 arg->lsp_printer, h);
790 * Print all objects in \a s.
792 void lu_site_print(const struct lu_env *env, struct lu_site *s, void *cookie,
793 lu_printer_t printer)
795 struct lu_site_print_arg arg = {
796 .lsp_env = (struct lu_env *)env,
797 .lsp_cookie = cookie,
798 .lsp_printer = printer,
801 cfs_hash_for_each(s->ls_obj_hash, lu_site_obj_print, &arg);
803 EXPORT_SYMBOL(lu_site_print);
806 LU_CACHE_PERCENT_MAX = 50,
807 LU_CACHE_PERCENT_DEFAULT = 20
810 static unsigned int lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
811 CFS_MODULE_PARM(lu_cache_percent, "i", int, 0644,
812 "Percentage of memory to be used as lu_object cache");
815 * Return desired hash table order.
817 static int lu_htable_order(void)
819 unsigned long cache_size;
823 * Calculate hash table size, assuming that we want reasonable
824 * performance when 20% of total memory is occupied by cache of
827 * Size of lu_object is (arbitrary) taken as 1K (together with inode).
829 cache_size = cfs_num_physpages;
831 #if BITS_PER_LONG == 32
832 /* limit hashtable size for lowmem systems to low RAM */
833 if (cache_size > 1 << (30 - CFS_PAGE_SHIFT))
834 cache_size = 1 << (30 - CFS_PAGE_SHIFT) * 3 / 4;
837 /* clear off unreasonable cache setting. */
838 if (lu_cache_percent == 0 || lu_cache_percent > LU_CACHE_PERCENT_MAX) {
839 CWARN("obdclass: invalid lu_cache_percent: %u, it must be in"
840 " the range of (0, %u]. Will use default value: %u.\n",
841 lu_cache_percent, LU_CACHE_PERCENT_MAX,
842 LU_CACHE_PERCENT_DEFAULT);
844 lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
846 cache_size = cache_size / 100 * lu_cache_percent *
847 (CFS_PAGE_SIZE / 1024);
849 for (bits = 1; (1 << bits) < cache_size; ++bits) {
855 static unsigned lu_obj_hop_hash(cfs_hash_t *hs,
856 const void *key, unsigned mask)
858 struct lu_fid *fid = (struct lu_fid *)key;
861 hash = fid_flatten32(fid);
862 hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
863 hash = cfs_hash_long(hash, hs->hs_bkt_bits);
865 /* give me another random factor */
866 hash -= cfs_hash_long((unsigned long)hs, fid_oid(fid) % 11 + 3);
868 hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
869 hash |= (fid_seq(fid) + fid_oid(fid)) & (CFS_HASH_NBKT(hs) - 1);
874 static void *lu_obj_hop_object(cfs_hlist_node_t *hnode)
876 return cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
879 static void *lu_obj_hop_key(cfs_hlist_node_t *hnode)
881 struct lu_object_header *h;
883 h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
887 static int lu_obj_hop_keycmp(const void *key, cfs_hlist_node_t *hnode)
889 struct lu_object_header *h;
891 h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
892 return lu_fid_eq(&h->loh_fid, (struct lu_fid *)key);
895 static void lu_obj_hop_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
897 struct lu_object_header *h;
899 h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
900 if (cfs_atomic_add_return(1, &h->loh_ref) == 1) {
901 struct lu_site_bkt_data *bkt;
904 cfs_hash_bd_get(hs, &h->loh_fid, &bd);
905 bkt = cfs_hash_bd_extra_get(hs, &bd);
910 static void lu_obj_hop_put_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
912 LBUG(); /* we should never called it */
915 cfs_hash_ops_t lu_site_hash_ops = {
916 .hs_hash = lu_obj_hop_hash,
917 .hs_key = lu_obj_hop_key,
918 .hs_keycmp = lu_obj_hop_keycmp,
919 .hs_object = lu_obj_hop_object,
920 .hs_get = lu_obj_hop_get,
921 .hs_put_locked = lu_obj_hop_put_locked,
924 void lu_dev_add_linkage(struct lu_site *s, struct lu_device *d)
926 spin_lock(&s->ls_ld_lock);
927 if (cfs_list_empty(&d->ld_linkage))
928 cfs_list_add(&d->ld_linkage, &s->ls_ld_linkage);
929 spin_unlock(&s->ls_ld_lock);
931 EXPORT_SYMBOL(lu_dev_add_linkage);
933 void lu_dev_del_linkage(struct lu_site *s, struct lu_device *d)
935 spin_lock(&s->ls_ld_lock);
936 cfs_list_del_init(&d->ld_linkage);
937 spin_unlock(&s->ls_ld_lock);
939 EXPORT_SYMBOL(lu_dev_del_linkage);
942 * Initialize site \a s, with \a d as the top level device.
944 #define LU_SITE_BITS_MIN 12
945 #define LU_SITE_BITS_MAX 24
947 * total 256 buckets, we don't want too many buckets because:
948 * - consume too much memory
949 * - avoid unbalanced LRU list
951 #define LU_SITE_BKT_BITS 8
953 int lu_site_init(struct lu_site *s, struct lu_device *top)
955 struct lu_site_bkt_data *bkt;
962 memset(s, 0, sizeof *s);
963 bits = lu_htable_order();
964 snprintf(name, 16, "lu_site_%s", top->ld_type->ldt_name);
965 for (bits = min(max(LU_SITE_BITS_MIN, bits), LU_SITE_BITS_MAX);
966 bits >= LU_SITE_BITS_MIN; bits--) {
967 s->ls_obj_hash = cfs_hash_create(name, bits, bits,
968 bits - LU_SITE_BKT_BITS,
971 CFS_HASH_SPIN_BKTLOCK |
972 CFS_HASH_NO_ITEMREF |
974 CFS_HASH_ASSERT_EMPTY);
975 if (s->ls_obj_hash != NULL)
979 if (s->ls_obj_hash == NULL) {
980 CERROR("failed to create lu_site hash with bits: %d\n", bits);
984 cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
985 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
986 CFS_INIT_LIST_HEAD(&bkt->lsb_lru);
987 cfs_waitq_init(&bkt->lsb_marche_funebre);
990 s->ls_stats = lprocfs_alloc_stats(LU_SS_LAST_STAT, 0);
991 if (s->ls_stats == NULL) {
992 cfs_hash_putref(s->ls_obj_hash);
993 s->ls_obj_hash = NULL;
997 lprocfs_counter_init(s->ls_stats, LU_SS_CREATED,
998 0, "created", "created");
999 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_HIT,
1000 0, "cache_hit", "cache_hit");
1001 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_MISS,
1002 0, "cache_miss", "cache_miss");
1003 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_RACE,
1004 0, "cache_race", "cache_race");
1005 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_DEATH_RACE,
1006 0, "cache_death_race", "cache_death_race");
1007 lprocfs_counter_init(s->ls_stats, LU_SS_LRU_PURGED,
1008 0, "lru_purged", "lru_purged");
1010 CFS_INIT_LIST_HEAD(&s->ls_linkage);
1011 s->ls_top_dev = top;
1014 lu_ref_add(&top->ld_reference, "site-top", s);
1016 CFS_INIT_LIST_HEAD(&s->ls_ld_linkage);
1017 spin_lock_init(&s->ls_ld_lock);
1019 lu_dev_add_linkage(s, top);
1023 EXPORT_SYMBOL(lu_site_init);
1026 * Finalize \a s and release its resources.
1028 void lu_site_fini(struct lu_site *s)
1030 mutex_lock(&lu_sites_guard);
1031 cfs_list_del_init(&s->ls_linkage);
1032 mutex_unlock(&lu_sites_guard);
1034 if (s->ls_obj_hash != NULL) {
1035 cfs_hash_putref(s->ls_obj_hash);
1036 s->ls_obj_hash = NULL;
1039 if (s->ls_top_dev != NULL) {
1040 s->ls_top_dev->ld_site = NULL;
1041 lu_ref_del(&s->ls_top_dev->ld_reference, "site-top", s);
1042 lu_device_put(s->ls_top_dev);
1043 s->ls_top_dev = NULL;
1046 if (s->ls_stats != NULL)
1047 lprocfs_free_stats(&s->ls_stats);
1049 EXPORT_SYMBOL(lu_site_fini);
1052 * Called when initialization of stack for this site is completed.
1054 int lu_site_init_finish(struct lu_site *s)
1057 mutex_lock(&lu_sites_guard);
1058 result = lu_context_refill(&lu_shrink_env.le_ctx);
1060 cfs_list_add(&s->ls_linkage, &lu_sites);
1061 mutex_unlock(&lu_sites_guard);
1064 EXPORT_SYMBOL(lu_site_init_finish);
1067 * Acquire additional reference on device \a d
1069 void lu_device_get(struct lu_device *d)
1071 cfs_atomic_inc(&d->ld_ref);
1073 EXPORT_SYMBOL(lu_device_get);
1076 * Release reference on device \a d.
1078 void lu_device_put(struct lu_device *d)
1080 LASSERT(cfs_atomic_read(&d->ld_ref) > 0);
1081 cfs_atomic_dec(&d->ld_ref);
1083 EXPORT_SYMBOL(lu_device_put);
1086 * Initialize device \a d of type \a t.
1088 int lu_device_init(struct lu_device *d, struct lu_device_type *t)
1090 if (t->ldt_device_nr++ == 0 && t->ldt_ops->ldto_start != NULL)
1091 t->ldt_ops->ldto_start(t);
1092 memset(d, 0, sizeof *d);
1093 cfs_atomic_set(&d->ld_ref, 0);
1095 lu_ref_init(&d->ld_reference);
1096 CFS_INIT_LIST_HEAD(&d->ld_linkage);
1099 EXPORT_SYMBOL(lu_device_init);
1102 * Finalize device \a d.
1104 void lu_device_fini(struct lu_device *d)
1106 struct lu_device_type *t;
1109 if (d->ld_obd != NULL) {
1110 d->ld_obd->obd_lu_dev = NULL;
1114 lu_ref_fini(&d->ld_reference);
1115 LASSERTF(cfs_atomic_read(&d->ld_ref) == 0,
1116 "Refcount is %u\n", cfs_atomic_read(&d->ld_ref));
1117 LASSERT(t->ldt_device_nr > 0);
1118 if (--t->ldt_device_nr == 0 && t->ldt_ops->ldto_stop != NULL)
1119 t->ldt_ops->ldto_stop(t);
1121 EXPORT_SYMBOL(lu_device_fini);
1124 * Initialize object \a o that is part of compound object \a h and was created
1127 int lu_object_init(struct lu_object *o,
1128 struct lu_object_header *h, struct lu_device *d)
1130 memset(o, 0, sizeof *o);
1134 o->lo_dev_ref = lu_ref_add(&d->ld_reference, "lu_object", o);
1135 CFS_INIT_LIST_HEAD(&o->lo_linkage);
1138 EXPORT_SYMBOL(lu_object_init);
1141 * Finalize object and release its resources.
1143 void lu_object_fini(struct lu_object *o)
1145 struct lu_device *dev = o->lo_dev;
1147 LASSERT(cfs_list_empty(&o->lo_linkage));
1150 lu_ref_del_at(&dev->ld_reference,
1151 o->lo_dev_ref , "lu_object", o);
1156 EXPORT_SYMBOL(lu_object_fini);
1159 * Add object \a o as first layer of compound object \a h
1161 * This is typically called by the ->ldo_object_alloc() method of top-level
1164 void lu_object_add_top(struct lu_object_header *h, struct lu_object *o)
1166 cfs_list_move(&o->lo_linkage, &h->loh_layers);
1168 EXPORT_SYMBOL(lu_object_add_top);
1171 * Add object \a o as a layer of compound object, going after \a before.
1173 * This is typically called by the ->ldo_object_alloc() method of \a
1176 void lu_object_add(struct lu_object *before, struct lu_object *o)
1178 cfs_list_move(&o->lo_linkage, &before->lo_linkage);
1180 EXPORT_SYMBOL(lu_object_add);
1183 * Initialize compound object.
1185 int lu_object_header_init(struct lu_object_header *h)
1187 memset(h, 0, sizeof *h);
1188 cfs_atomic_set(&h->loh_ref, 1);
1189 CFS_INIT_HLIST_NODE(&h->loh_hash);
1190 CFS_INIT_LIST_HEAD(&h->loh_lru);
1191 CFS_INIT_LIST_HEAD(&h->loh_layers);
1192 lu_ref_init(&h->loh_reference);
1195 EXPORT_SYMBOL(lu_object_header_init);
1198 * Finalize compound object.
1200 void lu_object_header_fini(struct lu_object_header *h)
1202 LASSERT(cfs_list_empty(&h->loh_layers));
1203 LASSERT(cfs_list_empty(&h->loh_lru));
1204 LASSERT(cfs_hlist_unhashed(&h->loh_hash));
1205 lu_ref_fini(&h->loh_reference);
1207 EXPORT_SYMBOL(lu_object_header_fini);
1210 * Given a compound object, find its slice, corresponding to the device type
1213 struct lu_object *lu_object_locate(struct lu_object_header *h,
1214 const struct lu_device_type *dtype)
1216 struct lu_object *o;
1218 cfs_list_for_each_entry(o, &h->loh_layers, lo_linkage) {
1219 if (o->lo_dev->ld_type == dtype)
1224 EXPORT_SYMBOL(lu_object_locate);
1229 * Finalize and free devices in the device stack.
1231 * Finalize device stack by purging object cache, and calling
1232 * lu_device_type_operations::ldto_device_fini() and
1233 * lu_device_type_operations::ldto_device_free() on all devices in the stack.
1235 void lu_stack_fini(const struct lu_env *env, struct lu_device *top)
1237 struct lu_site *site = top->ld_site;
1238 struct lu_device *scan;
1239 struct lu_device *next;
1241 lu_site_purge(env, site, ~0);
1242 for (scan = top; scan != NULL; scan = next) {
1243 next = scan->ld_type->ldt_ops->ldto_device_fini(env, scan);
1244 lu_ref_del(&scan->ld_reference, "lu-stack", &lu_site_init);
1245 lu_device_put(scan);
1249 lu_site_purge(env, site, ~0);
1251 for (scan = top; scan != NULL; scan = next) {
1252 const struct lu_device_type *ldt = scan->ld_type;
1253 struct obd_type *type;
1255 next = ldt->ldt_ops->ldto_device_free(env, scan);
1256 type = ldt->ldt_obd_type;
1259 class_put_type(type);
1263 EXPORT_SYMBOL(lu_stack_fini);
1267 * Maximal number of tld slots.
1269 LU_CONTEXT_KEY_NR = 40
1272 static struct lu_context_key *lu_keys[LU_CONTEXT_KEY_NR] = { NULL, };
1274 static DEFINE_SPINLOCK(lu_keys_guard);
1277 * Global counter incremented whenever key is registered, unregistered,
1278 * revived or quiesced. This is used to void unnecessary calls to
1279 * lu_context_refill(). No locking is provided, as initialization and shutdown
1280 * are supposed to be externally serialized.
1282 static unsigned key_set_version = 0;
1287 int lu_context_key_register(struct lu_context_key *key)
1292 LASSERT(key->lct_init != NULL);
1293 LASSERT(key->lct_fini != NULL);
1294 LASSERT(key->lct_tags != 0);
1295 LASSERT(key->lct_owner != NULL);
1298 spin_lock(&lu_keys_guard);
1299 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1300 if (lu_keys[i] == NULL) {
1302 cfs_atomic_set(&key->lct_used, 1);
1304 lu_ref_init(&key->lct_reference);
1310 spin_unlock(&lu_keys_guard);
1313 EXPORT_SYMBOL(lu_context_key_register);
1315 static void key_fini(struct lu_context *ctx, int index)
1317 if (ctx->lc_value != NULL && ctx->lc_value[index] != NULL) {
1318 struct lu_context_key *key;
1320 key = lu_keys[index];
1321 LASSERT(key != NULL);
1322 LASSERT(key->lct_fini != NULL);
1323 LASSERT(cfs_atomic_read(&key->lct_used) > 1);
1325 key->lct_fini(ctx, key, ctx->lc_value[index]);
1326 lu_ref_del(&key->lct_reference, "ctx", ctx);
1327 cfs_atomic_dec(&key->lct_used);
1329 LASSERT(key->lct_owner != NULL);
1330 if ((ctx->lc_tags & LCT_NOREF) == 0) {
1331 LINVRNT(cfs_module_refcount(key->lct_owner) > 0);
1332 cfs_module_put(key->lct_owner);
1334 ctx->lc_value[index] = NULL;
1341 void lu_context_key_degister(struct lu_context_key *key)
1343 LASSERT(cfs_atomic_read(&key->lct_used) >= 1);
1344 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1346 lu_context_key_quiesce(key);
1349 spin_lock(&lu_keys_guard);
1350 key_fini(&lu_shrink_env.le_ctx, key->lct_index);
1351 if (lu_keys[key->lct_index]) {
1352 lu_keys[key->lct_index] = NULL;
1353 lu_ref_fini(&key->lct_reference);
1355 spin_unlock(&lu_keys_guard);
1357 LASSERTF(cfs_atomic_read(&key->lct_used) == 1,
1358 "key has instances: %d\n",
1359 cfs_atomic_read(&key->lct_used));
1361 EXPORT_SYMBOL(lu_context_key_degister);
1364 * Register a number of keys. This has to be called after all keys have been
1365 * initialized by a call to LU_CONTEXT_KEY_INIT().
1367 int lu_context_key_register_many(struct lu_context_key *k, ...)
1369 struct lu_context_key *key = k;
1375 result = lu_context_key_register(key);
1378 key = va_arg(args, struct lu_context_key *);
1379 } while (key != NULL);
1385 lu_context_key_degister(k);
1386 k = va_arg(args, struct lu_context_key *);
1393 EXPORT_SYMBOL(lu_context_key_register_many);
1396 * De-register a number of keys. This is a dual to
1397 * lu_context_key_register_many().
1399 void lu_context_key_degister_many(struct lu_context_key *k, ...)
1405 lu_context_key_degister(k);
1406 k = va_arg(args, struct lu_context_key*);
1407 } while (k != NULL);
1410 EXPORT_SYMBOL(lu_context_key_degister_many);
1413 * Revive a number of keys.
1415 void lu_context_key_revive_many(struct lu_context_key *k, ...)
1421 lu_context_key_revive(k);
1422 k = va_arg(args, struct lu_context_key*);
1423 } while (k != NULL);
1426 EXPORT_SYMBOL(lu_context_key_revive_many);
1429 * Quiescent a number of keys.
1431 void lu_context_key_quiesce_many(struct lu_context_key *k, ...)
1437 lu_context_key_quiesce(k);
1438 k = va_arg(args, struct lu_context_key*);
1439 } while (k != NULL);
1442 EXPORT_SYMBOL(lu_context_key_quiesce_many);
1445 * Return value associated with key \a key in context \a ctx.
1447 void *lu_context_key_get(const struct lu_context *ctx,
1448 const struct lu_context_key *key)
1450 LINVRNT(ctx->lc_state == LCS_ENTERED);
1451 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1452 LASSERT(lu_keys[key->lct_index] == key);
1453 return ctx->lc_value[key->lct_index];
1455 EXPORT_SYMBOL(lu_context_key_get);
1458 * List of remembered contexts. XXX document me.
1460 static CFS_LIST_HEAD(lu_context_remembered);
1463 * Destroy \a key in all remembered contexts. This is used to destroy key
1464 * values in "shared" contexts (like service threads), when a module owning
1465 * the key is about to be unloaded.
1467 void lu_context_key_quiesce(struct lu_context_key *key)
1469 struct lu_context *ctx;
1470 extern unsigned cl_env_cache_purge(unsigned nr);
1472 if (!(key->lct_tags & LCT_QUIESCENT)) {
1474 * XXX layering violation.
1476 cl_env_cache_purge(~0);
1477 key->lct_tags |= LCT_QUIESCENT;
1479 * XXX memory barrier has to go here.
1481 spin_lock(&lu_keys_guard);
1482 cfs_list_for_each_entry(ctx, &lu_context_remembered,
1484 key_fini(ctx, key->lct_index);
1485 spin_unlock(&lu_keys_guard);
1489 EXPORT_SYMBOL(lu_context_key_quiesce);
1491 void lu_context_key_revive(struct lu_context_key *key)
1493 key->lct_tags &= ~LCT_QUIESCENT;
1496 EXPORT_SYMBOL(lu_context_key_revive);
1498 static void keys_fini(struct lu_context *ctx)
1502 if (ctx->lc_value == NULL)
1505 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i)
1508 OBD_FREE(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1509 ctx->lc_value = NULL;
1512 static int keys_fill(struct lu_context *ctx)
1516 LINVRNT(ctx->lc_value != NULL);
1517 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1518 struct lu_context_key *key;
1521 if (ctx->lc_value[i] == NULL && key != NULL &&
1522 (key->lct_tags & ctx->lc_tags) &&
1524 * Don't create values for a LCT_QUIESCENT key, as this
1525 * will pin module owning a key.
1527 !(key->lct_tags & LCT_QUIESCENT)) {
1530 LINVRNT(key->lct_init != NULL);
1531 LINVRNT(key->lct_index == i);
1533 value = key->lct_init(ctx, key);
1534 if (unlikely(IS_ERR(value)))
1535 return PTR_ERR(value);
1537 LASSERT(key->lct_owner != NULL);
1538 if (!(ctx->lc_tags & LCT_NOREF))
1539 cfs_try_module_get(key->lct_owner);
1540 lu_ref_add_atomic(&key->lct_reference, "ctx", ctx);
1541 cfs_atomic_inc(&key->lct_used);
1543 * This is the only place in the code, where an
1544 * element of ctx->lc_value[] array is set to non-NULL
1547 ctx->lc_value[i] = value;
1548 if (key->lct_exit != NULL)
1549 ctx->lc_tags |= LCT_HAS_EXIT;
1551 ctx->lc_version = key_set_version;
1556 static int keys_init(struct lu_context *ctx)
1558 OBD_ALLOC(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1559 if (likely(ctx->lc_value != NULL))
1560 return keys_fill(ctx);
1566 * Initialize context data-structure. Create values for all keys.
1568 int lu_context_init(struct lu_context *ctx, __u32 tags)
1572 memset(ctx, 0, sizeof *ctx);
1573 ctx->lc_state = LCS_INITIALIZED;
1574 ctx->lc_tags = tags;
1575 if (tags & LCT_REMEMBER) {
1576 spin_lock(&lu_keys_guard);
1577 cfs_list_add(&ctx->lc_remember, &lu_context_remembered);
1578 spin_unlock(&lu_keys_guard);
1580 CFS_INIT_LIST_HEAD(&ctx->lc_remember);
1583 rc = keys_init(ctx);
1585 lu_context_fini(ctx);
1589 EXPORT_SYMBOL(lu_context_init);
1592 * Finalize context data-structure. Destroy key values.
1594 void lu_context_fini(struct lu_context *ctx)
1596 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1597 ctx->lc_state = LCS_FINALIZED;
1599 if ((ctx->lc_tags & LCT_REMEMBER) == 0) {
1600 LASSERT(cfs_list_empty(&ctx->lc_remember));
1603 } else { /* could race with key degister */
1604 spin_lock(&lu_keys_guard);
1606 cfs_list_del_init(&ctx->lc_remember);
1607 spin_unlock(&lu_keys_guard);
1610 EXPORT_SYMBOL(lu_context_fini);
1613 * Called before entering context.
1615 void lu_context_enter(struct lu_context *ctx)
1617 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1618 ctx->lc_state = LCS_ENTERED;
1620 EXPORT_SYMBOL(lu_context_enter);
1623 * Called after exiting from \a ctx
1625 void lu_context_exit(struct lu_context *ctx)
1629 LINVRNT(ctx->lc_state == LCS_ENTERED);
1630 ctx->lc_state = LCS_LEFT;
1631 if (ctx->lc_tags & LCT_HAS_EXIT && ctx->lc_value != NULL) {
1632 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1633 if (ctx->lc_value[i] != NULL) {
1634 struct lu_context_key *key;
1637 LASSERT(key != NULL);
1638 if (key->lct_exit != NULL)
1640 key, ctx->lc_value[i]);
1645 EXPORT_SYMBOL(lu_context_exit);
1648 * Allocate for context all missing keys that were registered after context
1649 * creation. key_set_version is only changed in rare cases when modules
1650 * are loaded and removed.
1652 int lu_context_refill(struct lu_context *ctx)
1654 return likely(ctx->lc_version == key_set_version) ? 0 : keys_fill(ctx);
1656 EXPORT_SYMBOL(lu_context_refill);
1659 * lu_ctx_tags/lu_ses_tags will be updated if there are new types of
1660 * obd being added. Currently, this is only used on client side, specifically
1661 * for echo device client, for other stack (like ptlrpc threads), context are
1662 * predefined when the lu_device type are registered, during the module probe
1665 __u32 lu_context_tags_default = 0;
1666 __u32 lu_session_tags_default = 0;
1668 void lu_context_tags_update(__u32 tags)
1670 spin_lock(&lu_keys_guard);
1671 lu_context_tags_default |= tags;
1673 spin_unlock(&lu_keys_guard);
1675 EXPORT_SYMBOL(lu_context_tags_update);
1677 void lu_context_tags_clear(__u32 tags)
1679 spin_lock(&lu_keys_guard);
1680 lu_context_tags_default &= ~tags;
1682 spin_unlock(&lu_keys_guard);
1684 EXPORT_SYMBOL(lu_context_tags_clear);
1686 void lu_session_tags_update(__u32 tags)
1688 spin_lock(&lu_keys_guard);
1689 lu_session_tags_default |= tags;
1691 spin_unlock(&lu_keys_guard);
1693 EXPORT_SYMBOL(lu_session_tags_update);
1695 void lu_session_tags_clear(__u32 tags)
1697 spin_lock(&lu_keys_guard);
1698 lu_session_tags_default &= ~tags;
1700 spin_unlock(&lu_keys_guard);
1702 EXPORT_SYMBOL(lu_session_tags_clear);
1704 int lu_env_init(struct lu_env *env, __u32 tags)
1709 result = lu_context_init(&env->le_ctx, tags);
1710 if (likely(result == 0))
1711 lu_context_enter(&env->le_ctx);
1714 EXPORT_SYMBOL(lu_env_init);
1716 void lu_env_fini(struct lu_env *env)
1718 lu_context_exit(&env->le_ctx);
1719 lu_context_fini(&env->le_ctx);
1722 EXPORT_SYMBOL(lu_env_fini);
1724 int lu_env_refill(struct lu_env *env)
1728 result = lu_context_refill(&env->le_ctx);
1729 if (result == 0 && env->le_ses != NULL)
1730 result = lu_context_refill(env->le_ses);
1733 EXPORT_SYMBOL(lu_env_refill);
1736 * Currently, this API will only be used by echo client.
1737 * Because echo client and normal lustre client will share
1738 * same cl_env cache. So echo client needs to refresh
1739 * the env context after it get one from the cache, especially
1740 * when normal client and echo client co-exist in the same client.
1742 int lu_env_refill_by_tags(struct lu_env *env, __u32 ctags,
1747 if ((env->le_ctx.lc_tags & ctags) != ctags) {
1748 env->le_ctx.lc_version = 0;
1749 env->le_ctx.lc_tags |= ctags;
1752 if (env->le_ses && (env->le_ses->lc_tags & stags) != stags) {
1753 env->le_ses->lc_version = 0;
1754 env->le_ses->lc_tags |= stags;
1757 result = lu_env_refill(env);
1761 EXPORT_SYMBOL(lu_env_refill_by_tags);
1763 static struct cfs_shrinker *lu_site_shrinker = NULL;
1765 typedef struct lu_site_stats{
1766 unsigned lss_populated;
1767 unsigned lss_max_search;
1772 static void lu_site_stats_get(cfs_hash_t *hs,
1773 lu_site_stats_t *stats, int populated)
1778 cfs_hash_for_each_bucket(hs, &bd, i) {
1779 struct lu_site_bkt_data *bkt = cfs_hash_bd_extra_get(hs, &bd);
1780 cfs_hlist_head_t *hhead;
1782 cfs_hash_bd_lock(hs, &bd, 1);
1783 stats->lss_busy += bkt->lsb_busy;
1784 stats->lss_total += cfs_hash_bd_count_get(&bd);
1785 stats->lss_max_search = max((int)stats->lss_max_search,
1786 cfs_hash_bd_depmax_get(&bd));
1788 cfs_hash_bd_unlock(hs, &bd, 1);
1792 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
1793 if (!cfs_hlist_empty(hhead))
1794 stats->lss_populated++;
1796 cfs_hash_bd_unlock(hs, &bd, 1);
1803 * There exists a potential lock inversion deadlock scenario when using
1804 * Lustre on top of ZFS. This occurs between one of ZFS's
1805 * buf_hash_table.ht_lock's, and Lustre's lu_sites_guard lock. Essentially,
1806 * thread A will take the lu_sites_guard lock and sleep on the ht_lock,
1807 * while thread B will take the ht_lock and sleep on the lu_sites_guard
1808 * lock. Obviously neither thread will wake and drop their respective hold
1811 * To prevent this from happening we must ensure the lu_sites_guard lock is
1812 * not taken while down this code path. ZFS reliably does not set the
1813 * __GFP_FS bit in its code paths, so this can be used to determine if it
1814 * is safe to take the lu_sites_guard lock.
1816 * Ideally we should accurately return the remaining number of cached
1817 * objects without taking the lu_sites_guard lock, but this is not
1818 * possible in the current implementation.
1820 static int lu_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
1822 lu_site_stats_t stats;
1824 struct lu_site *tmp;
1826 int remain = shrink_param(sc, nr_to_scan);
1827 CFS_LIST_HEAD(splice);
1829 if (!(shrink_param(sc, gfp_mask) & __GFP_FS)) {
1833 /* We must not take the lu_sites_guard lock when
1834 * __GFP_FS is *not* set because of the deadlock
1835 * possibility detailed above. Additionally,
1836 * since we cannot determine the number of
1837 * objects in the cache without taking this
1838 * lock, we're in a particularly tough spot. As
1839 * a result, we'll just lie and say our cache is
1840 * empty. This _should_ be ok, as we can't
1841 * reclaim objects when __GFP_FS is *not* set
1847 CDEBUG(D_INODE, "Shrink %d objects\n", remain);
1849 mutex_lock(&lu_sites_guard);
1850 cfs_list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
1851 if (shrink_param(sc, nr_to_scan) != 0) {
1852 remain = lu_site_purge(&lu_shrink_env, s, remain);
1854 * Move just shrunk site to the tail of site list to
1855 * assure shrinking fairness.
1857 cfs_list_move_tail(&s->ls_linkage, &splice);
1860 memset(&stats, 0, sizeof(stats));
1861 lu_site_stats_get(s->ls_obj_hash, &stats, 0);
1862 cached += stats.lss_total - stats.lss_busy;
1863 if (shrink_param(sc, nr_to_scan) && remain <= 0)
1866 cfs_list_splice(&splice, lu_sites.prev);
1867 mutex_unlock(&lu_sites_guard);
1869 cached = (cached / 100) * sysctl_vfs_cache_pressure;
1870 if (shrink_param(sc, nr_to_scan) == 0)
1871 CDEBUG(D_INODE, "%d objects cached\n", cached);
1880 * Environment to be used in debugger, contains all tags.
1882 struct lu_env lu_debugging_env;
1885 * Debugging printer function using printk().
1887 int lu_printk_printer(const struct lu_env *env,
1888 void *unused, const char *format, ...)
1892 va_start(args, format);
1893 vprintk(format, args);
1898 void lu_debugging_setup(void)
1900 lu_env_init(&lu_debugging_env, ~0);
1903 void lu_context_keys_dump(void)
1907 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1908 struct lu_context_key *key;
1912 CERROR("[%d]: %p %x (%p,%p,%p) %d %d \"%s\"@%p\n",
1913 i, key, key->lct_tags,
1914 key->lct_init, key->lct_fini, key->lct_exit,
1915 key->lct_index, cfs_atomic_read(&key->lct_used),
1916 key->lct_owner ? key->lct_owner->name : "",
1918 lu_ref_print(&key->lct_reference);
1922 EXPORT_SYMBOL(lu_context_keys_dump);
1923 #else /* !__KERNEL__ */
1924 static int lu_cache_shrink(int nr, unsigned int gfp_mask)
1928 #endif /* __KERNEL__ */
1930 int cl_global_init(void);
1931 void cl_global_fini(void);
1932 int lu_ref_global_init(void);
1933 void lu_ref_global_fini(void);
1935 int dt_global_init(void);
1936 void dt_global_fini(void);
1938 int llo_global_init(void);
1939 void llo_global_fini(void);
1941 /* context key constructor/destructor: lu_ucred_key_init, lu_ucred_key_fini */
1942 LU_KEY_INIT_FINI(lu_ucred, struct lu_ucred);
1944 static struct lu_context_key lu_ucred_key = {
1945 .lct_tags = LCT_SESSION,
1946 .lct_init = lu_ucred_key_init,
1947 .lct_fini = lu_ucred_key_fini
1951 * Get ucred key if session exists and ucred key is allocated on it.
1952 * Return NULL otherwise.
1954 struct lu_ucred *lu_ucred(const struct lu_env *env)
1958 return lu_context_key_get(env->le_ses, &lu_ucred_key);
1960 EXPORT_SYMBOL(lu_ucred);
1963 * Get ucred key and check if it is properly initialized.
1964 * Return NULL otherwise.
1966 struct lu_ucred *lu_ucred_check(const struct lu_env *env)
1968 struct lu_ucred *uc = lu_ucred(env);
1969 if (uc && uc->uc_valid != UCRED_OLD && uc->uc_valid != UCRED_NEW)
1973 EXPORT_SYMBOL(lu_ucred_check);
1976 * Get ucred key, which must exist and must be properly initialized.
1979 struct lu_ucred *lu_ucred_assert(const struct lu_env *env)
1981 struct lu_ucred *uc = lu_ucred_check(env);
1982 LASSERT(uc != NULL);
1985 EXPORT_SYMBOL(lu_ucred_assert);
1988 * Initialization of global lu_* data.
1990 int lu_global_init(void)
1994 CDEBUG(D_INFO, "Lustre LU module (%p).\n", &lu_keys);
1996 result = lu_ref_global_init();
2000 LU_CONTEXT_KEY_INIT(&lu_global_key);
2001 result = lu_context_key_register(&lu_global_key);
2005 LU_CONTEXT_KEY_INIT(&lu_ucred_key);
2006 result = lu_context_key_register(&lu_ucred_key);
2011 * At this level, we don't know what tags are needed, so allocate them
2012 * conservatively. This should not be too bad, because this
2013 * environment is global.
2015 mutex_lock(&lu_sites_guard);
2016 result = lu_env_init(&lu_shrink_env, LCT_SHRINKER);
2017 mutex_unlock(&lu_sites_guard);
2022 * seeks estimation: 3 seeks to read a record from oi, one to read
2023 * inode, one for ea. Unfortunately setting this high value results in
2024 * lu_object/inode cache consuming all the memory.
2026 lu_site_shrinker = cfs_set_shrinker(CFS_DEFAULT_SEEKS, lu_cache_shrink);
2027 if (lu_site_shrinker == NULL)
2030 result = lu_time_global_init();
2035 result = dt_global_init();
2039 result = llo_global_init();
2043 result = cl_global_init();
2050 * Dual to lu_global_init().
2052 void lu_global_fini(void)
2059 lu_time_global_fini();
2060 if (lu_site_shrinker != NULL) {
2061 cfs_remove_shrinker(lu_site_shrinker);
2062 lu_site_shrinker = NULL;
2065 lu_context_key_degister(&lu_global_key);
2066 lu_context_key_degister(&lu_ucred_key);
2069 * Tear shrinker environment down _after_ de-registering
2070 * lu_global_key, because the latter has a value in the former.
2072 mutex_lock(&lu_sites_guard);
2073 lu_env_fini(&lu_shrink_env);
2074 mutex_unlock(&lu_sites_guard);
2076 lu_ref_global_fini();
2079 struct lu_buf LU_BUF_NULL = {
2083 EXPORT_SYMBOL(LU_BUF_NULL);
2085 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx)
2088 struct lprocfs_counter ret;
2090 lprocfs_stats_collect(stats, idx, &ret);
2091 return (__u32)ret.lc_count;
2098 * Output site statistical counters into a buffer. Suitable for
2099 * lprocfs_rd_*()-style functions.
2101 int lu_site_stats_print(const struct lu_site *s, char *page, int count)
2103 lu_site_stats_t stats;
2105 memset(&stats, 0, sizeof(stats));
2106 lu_site_stats_get(s->ls_obj_hash, &stats, 1);
2108 return snprintf(page, count, "%d/%d %d/%d %d %d %d %d %d %d %d\n",
2111 stats.lss_populated,
2112 CFS_HASH_NHLIST(s->ls_obj_hash),
2113 stats.lss_max_search,
2114 ls_stats_read(s->ls_stats, LU_SS_CREATED),
2115 ls_stats_read(s->ls_stats, LU_SS_CACHE_HIT),
2116 ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
2117 ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
2118 ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
2119 ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED));
2121 EXPORT_SYMBOL(lu_site_stats_print);
2123 const char *lu_time_names[LU_TIME_NR] = {
2124 [LU_TIME_FIND_LOOKUP] = "find_lookup",
2125 [LU_TIME_FIND_ALLOC] = "find_alloc",
2126 [LU_TIME_FIND_INSERT] = "find_insert"
2128 EXPORT_SYMBOL(lu_time_names);
2131 * Helper function to initialize a number of kmem slab caches at once.
2133 int lu_kmem_init(struct lu_kmem_descr *caches)
2136 struct lu_kmem_descr *iter = caches;
2138 for (result = 0; iter->ckd_cache != NULL; ++iter) {
2139 *iter->ckd_cache = cfs_mem_cache_create(iter->ckd_name,
2142 if (*iter->ckd_cache == NULL) {
2144 /* free all previously allocated caches */
2145 lu_kmem_fini(caches);
2151 EXPORT_SYMBOL(lu_kmem_init);
2154 * Helper function to finalize a number of kmem slab cached at once. Dual to
2157 void lu_kmem_fini(struct lu_kmem_descr *caches)
2161 for (; caches->ckd_cache != NULL; ++caches) {
2162 if (*caches->ckd_cache != NULL) {
2163 rc = cfs_mem_cache_destroy(*caches->ckd_cache);
2164 LASSERTF(rc == 0, "couldn't destroy %s slab\n",
2166 *caches->ckd_cache = NULL;
2170 EXPORT_SYMBOL(lu_kmem_fini);
2173 * Temporary solution to be able to assign fid in ->do_create()
2174 * till we have fully-functional OST fids
2176 void lu_object_assign_fid(const struct lu_env *env, struct lu_object *o,
2177 const struct lu_fid *fid)
2179 struct lu_site *s = o->lo_dev->ld_site;
2180 struct lu_fid *old = &o->lo_header->loh_fid;
2181 struct lu_site_bkt_data *bkt;
2182 struct lu_object *shadow;
2183 cfs_waitlink_t waiter;
2188 LASSERT(fid_is_zero(old));
2190 hs = s->ls_obj_hash;
2191 cfs_hash_bd_get_and_lock(hs, (void *)fid, &bd, 1);
2192 shadow = htable_lookup(s, &bd, fid, &waiter, &version);
2193 /* supposed to be unique */
2194 LASSERT(shadow == NULL);
2196 bkt = cfs_hash_bd_extra_get(hs, &bd);
2197 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
2199 cfs_hash_bd_unlock(hs, &bd, 1);
2201 EXPORT_SYMBOL(lu_object_assign_fid);
2204 * allocates object with 0 (non-assiged) fid
2205 * XXX: temporary solution to be able to assign fid in ->do_create()
2206 * till we have fully-functional OST fids
2208 struct lu_object *lu_object_anon(const struct lu_env *env,
2209 struct lu_device *dev,
2210 const struct lu_object_conf *conf)
2213 struct lu_object *o;
2216 o = lu_object_alloc(env, dev, &fid, conf);
2220 EXPORT_SYMBOL(lu_object_anon);