4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/obdclass/lu_object.c
39 * These are the only exported functions, they provide some generic
40 * infrastructure for managing object devices
42 * Author: Nikita Danilov <nikita.danilov@sun.com>
45 #define DEBUG_SUBSYSTEM S_CLASS
47 #include <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>
62 static void lu_object_free(const struct lu_env *env, struct lu_object *o);
65 * Decrease reference counter on object. If last reference is freed, return
66 * object to the cache, unless lu_object_is_dying(o) holds. In the latter
67 * case, free object immediately.
69 void lu_object_put(const struct lu_env *env, struct lu_object *o)
71 struct lu_site_bkt_data *bkt;
72 struct lu_object_header *top;
74 struct lu_object *orig;
76 const struct lu_fid *fid;
79 site = o->lo_dev->ld_site;
83 * till we have full fids-on-OST implemented anonymous objects
84 * are possible in OSP. such an object isn't listed in the site
85 * so we should not remove it from the site.
87 fid = lu_object_fid(o);
88 if (fid_is_zero(fid)) {
89 LASSERT(top->loh_hash.next == NULL
90 && top->loh_hash.pprev == NULL);
91 LASSERT(cfs_list_empty(&top->loh_lru));
92 if (!cfs_atomic_dec_and_test(&top->loh_ref))
94 cfs_list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
95 if (o->lo_ops->loo_object_release != NULL)
96 o->lo_ops->loo_object_release(env, o);
98 lu_object_free(env, orig);
102 cfs_hash_bd_get(site->ls_obj_hash, &top->loh_fid, &bd);
103 bkt = cfs_hash_bd_extra_get(site->ls_obj_hash, &bd);
105 if (!cfs_hash_bd_dec_and_lock(site->ls_obj_hash, &bd, &top->loh_ref)) {
106 if (lu_object_is_dying(top)) {
109 * somebody may be waiting for this, currently only
110 * used for cl_object, see cl_object_put_last().
112 cfs_waitq_broadcast(&bkt->lsb_marche_funebre);
117 LASSERT(bkt->lsb_busy > 0);
120 * When last reference is released, iterate over object
121 * layers, and notify them that object is no longer busy.
123 cfs_list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
124 if (o->lo_ops->loo_object_release != NULL)
125 o->lo_ops->loo_object_release(env, o);
128 if (!lu_object_is_dying(top)) {
129 LASSERT(cfs_list_empty(&top->loh_lru));
130 cfs_list_add_tail(&top->loh_lru, &bkt->lsb_lru);
131 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
136 * If object is dying (will not be cached), removed it
137 * from hash table and LRU.
139 * This is done with hash table and LRU lists locked. As the only
140 * way to acquire first reference to previously unreferenced
141 * object is through hash-table lookup (lu_object_find()),
142 * or LRU scanning (lu_site_purge()), that are done under hash-table
143 * and LRU lock, no race with concurrent object lookup is possible
144 * and we can safely destroy object below.
146 cfs_hash_bd_del_locked(site->ls_obj_hash, &bd, &top->loh_hash);
147 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
149 * Object was already removed from hash and lru above, can
152 lu_object_free(env, orig);
154 EXPORT_SYMBOL(lu_object_put);
157 * Put object and don't keep in cache. This is temporary solution for
158 * multi-site objects when its layering is not constant.
160 void lu_object_put_nocache(const struct lu_env *env, struct lu_object *o)
162 set_bit(LU_OBJECT_HEARD_BANSHEE,
163 &o->lo_header->loh_flags);
164 return lu_object_put(env, o);
166 EXPORT_SYMBOL(lu_object_put_nocache);
169 * Allocate new object.
171 * This follows object creation protocol, described in the comment within
172 * struct lu_device_operations definition.
174 static struct lu_object *lu_object_alloc(const struct lu_env *env,
175 struct lu_device *dev,
176 const struct lu_fid *f,
177 const struct lu_object_conf *conf)
179 struct lu_object *scan;
180 struct lu_object *top;
187 * Create top-level object slice. This will also create
190 top = dev->ld_ops->ldo_object_alloc(env, NULL, dev);
192 RETURN(ERR_PTR(-ENOMEM));
196 * This is the only place where object fid is assigned. It's constant
199 top->lo_header->loh_fid = *f;
200 layers = &top->lo_header->loh_layers;
203 * Call ->loo_object_init() repeatedly, until no more new
204 * object slices are created.
207 cfs_list_for_each_entry(scan, layers, lo_linkage) {
208 if (scan->lo_flags & LU_OBJECT_ALLOCATED)
211 scan->lo_header = top->lo_header;
212 result = scan->lo_ops->loo_object_init(env, scan, conf);
214 lu_object_free(env, top);
215 RETURN(ERR_PTR(result));
217 scan->lo_flags |= LU_OBJECT_ALLOCATED;
221 cfs_list_for_each_entry_reverse(scan, layers, lo_linkage) {
222 if (scan->lo_ops->loo_object_start != NULL) {
223 result = scan->lo_ops->loo_object_start(env, scan);
225 lu_object_free(env, top);
226 RETURN(ERR_PTR(result));
231 lprocfs_counter_incr(dev->ld_site->ls_stats, LU_SS_CREATED);
238 static void lu_object_free(const struct lu_env *env, struct lu_object *o)
240 struct lu_site_bkt_data *bkt;
241 struct lu_site *site;
242 struct lu_object *scan;
246 site = o->lo_dev->ld_site;
247 layers = &o->lo_header->loh_layers;
248 bkt = lu_site_bkt_from_fid(site, &o->lo_header->loh_fid);
250 * First call ->loo_object_delete() method to release all resources.
252 cfs_list_for_each_entry_reverse(scan, layers, lo_linkage) {
253 if (scan->lo_ops->loo_object_delete != NULL)
254 scan->lo_ops->loo_object_delete(env, scan);
258 * Then, splice object layers into stand-alone list, and call
259 * ->loo_object_free() on all layers to free memory. Splice is
260 * necessary, because lu_object_header is freed together with the
263 CFS_INIT_LIST_HEAD(&splice);
264 cfs_list_splice_init(layers, &splice);
265 while (!cfs_list_empty(&splice)) {
267 * Free layers in bottom-to-top order, so that object header
268 * lives as long as possible and ->loo_object_free() methods
269 * can look at its contents.
271 o = container_of0(splice.prev, struct lu_object, lo_linkage);
272 cfs_list_del_init(&o->lo_linkage);
273 LASSERT(o->lo_ops->loo_object_free != NULL);
274 o->lo_ops->loo_object_free(env, o);
277 if (cfs_waitq_active(&bkt->lsb_marche_funebre))
278 cfs_waitq_broadcast(&bkt->lsb_marche_funebre);
282 * Free \a nr objects from the cold end of the site LRU list.
284 int lu_site_purge(const struct lu_env *env, struct lu_site *s, int nr)
286 struct lu_object_header *h;
287 struct lu_object_header *temp;
288 struct lu_site_bkt_data *bkt;
298 CFS_INIT_LIST_HEAD(&dispose);
300 * Under LRU list lock, scan LRU list and move unreferenced objects to
301 * the dispose list, removing them from LRU and hash table.
303 start = s->ls_purge_start;
304 bnr = (nr == ~0) ? -1 : nr / CFS_HASH_NBKT(s->ls_obj_hash) + 1;
307 cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
311 cfs_hash_bd_lock(s->ls_obj_hash, &bd, 1);
312 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
314 cfs_list_for_each_entry_safe(h, temp, &bkt->lsb_lru, loh_lru) {
315 LASSERT(cfs_atomic_read(&h->loh_ref) == 0);
317 cfs_hash_bd_get(s->ls_obj_hash, &h->loh_fid, &bd2);
318 LASSERT(bd.bd_bucket == bd2.bd_bucket);
320 cfs_hash_bd_del_locked(s->ls_obj_hash,
322 cfs_list_move(&h->loh_lru, &dispose);
326 if (nr != ~0 && --nr == 0)
329 if (count > 0 && --count == 0)
333 cfs_hash_bd_unlock(s->ls_obj_hash, &bd, 1);
336 * Free everything on the dispose list. This is safe against
337 * races due to the reasons described in lu_object_put().
339 while (!cfs_list_empty(&dispose)) {
340 h = container_of0(dispose.next,
341 struct lu_object_header, loh_lru);
342 cfs_list_del_init(&h->loh_lru);
343 lu_object_free(env, lu_object_top(h));
344 lprocfs_counter_incr(s->ls_stats, LU_SS_LRU_PURGED);
351 if (nr != 0 && did_sth && start != 0) {
352 start = 0; /* restart from the first bucket */
355 /* race on s->ls_purge_start, but nobody cares */
356 s->ls_purge_start = i % CFS_HASH_NBKT(s->ls_obj_hash);
360 EXPORT_SYMBOL(lu_site_purge);
365 * Code below has to jump through certain loops to output object description
366 * into libcfs_debug_msg-based log. The problem is that lu_object_print()
367 * composes object description from strings that are parts of _lines_ of
368 * output (i.e., strings that are not terminated by newline). This doesn't fit
369 * very well into libcfs_debug_msg() interface that assumes that each message
370 * supplied to it is a self-contained output line.
372 * To work around this, strings are collected in a temporary buffer
373 * (implemented as a value of lu_cdebug_key key), until terminating newline
374 * character is detected.
382 * XXX overflow is not handled correctly.
387 struct lu_cdebug_data {
391 char lck_area[LU_CDEBUG_LINE];
394 /* context key constructor/destructor: lu_global_key_init, lu_global_key_fini */
395 LU_KEY_INIT_FINI(lu_global, struct lu_cdebug_data);
398 * Key, holding temporary buffer. This key is registered very early by
401 struct lu_context_key lu_global_key = {
402 .lct_tags = LCT_MD_THREAD | LCT_DT_THREAD |
403 LCT_MG_THREAD | LCT_CL_THREAD,
404 .lct_init = lu_global_key_init,
405 .lct_fini = lu_global_key_fini
409 * Printer function emitting messages through libcfs_debug_msg().
411 int lu_cdebug_printer(const struct lu_env *env,
412 void *cookie, const char *format, ...)
414 struct libcfs_debug_msg_data *msgdata = cookie;
415 struct lu_cdebug_data *key;
420 va_start(args, format);
422 key = lu_context_key_get(&env->le_ctx, &lu_global_key);
423 LASSERT(key != NULL);
425 used = strlen(key->lck_area);
426 complete = format[strlen(format) - 1] == '\n';
428 * Append new chunk to the buffer.
430 vsnprintf(key->lck_area + used,
431 ARRAY_SIZE(key->lck_area) - used, format, args);
433 if (cfs_cdebug_show(msgdata->msg_mask, msgdata->msg_subsys))
434 libcfs_debug_msg(msgdata, "%s", key->lck_area);
435 key->lck_area[0] = 0;
440 EXPORT_SYMBOL(lu_cdebug_printer);
443 * Print object header.
445 void lu_object_header_print(const struct lu_env *env, void *cookie,
446 lu_printer_t printer,
447 const struct lu_object_header *hdr)
449 (*printer)(env, cookie, "header@%p[%#lx, %d, "DFID"%s%s%s]",
450 hdr, hdr->loh_flags, cfs_atomic_read(&hdr->loh_ref),
452 cfs_hlist_unhashed(&hdr->loh_hash) ? "" : " hash",
453 cfs_list_empty((cfs_list_t *)&hdr->loh_lru) ? \
455 hdr->loh_attr & LOHA_EXISTS ? " exist":"");
457 EXPORT_SYMBOL(lu_object_header_print);
460 * Print human readable representation of the \a o to the \a printer.
462 void lu_object_print(const struct lu_env *env, void *cookie,
463 lu_printer_t printer, const struct lu_object *o)
465 static const char ruler[] = "........................................";
466 struct lu_object_header *top;
470 lu_object_header_print(env, cookie, printer, top);
471 (*printer)(env, cookie, "{ \n");
472 cfs_list_for_each_entry(o, &top->loh_layers, lo_linkage) {
473 depth = o->lo_depth + 4;
476 * print `.' \a depth times followed by type name and address
478 (*printer)(env, cookie, "%*.*s%s@%p", depth, depth, ruler,
479 o->lo_dev->ld_type->ldt_name, o);
480 if (o->lo_ops->loo_object_print != NULL)
481 o->lo_ops->loo_object_print(env, cookie, printer, o);
482 (*printer)(env, cookie, "\n");
484 (*printer)(env, cookie, "} header@%p\n", top);
486 EXPORT_SYMBOL(lu_object_print);
489 * Check object consistency.
491 int lu_object_invariant(const struct lu_object *o)
493 struct lu_object_header *top;
496 cfs_list_for_each_entry(o, &top->loh_layers, lo_linkage) {
497 if (o->lo_ops->loo_object_invariant != NULL &&
498 !o->lo_ops->loo_object_invariant(o))
503 EXPORT_SYMBOL(lu_object_invariant);
505 static struct lu_object *htable_lookup(struct lu_site *s,
507 const struct lu_fid *f,
508 cfs_waitlink_t *waiter,
511 struct lu_site_bkt_data *bkt;
512 struct lu_object_header *h;
513 cfs_hlist_node_t *hnode;
514 __u64 ver = cfs_hash_bd_version_get(bd);
520 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, bd);
521 /* cfs_hash_bd_peek_locked is a somehow "internal" function
522 * of cfs_hash, it doesn't add refcount on object. */
523 hnode = cfs_hash_bd_peek_locked(s->ls_obj_hash, bd, (void *)f);
525 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS);
529 h = container_of0(hnode, struct lu_object_header, loh_hash);
530 if (likely(!lu_object_is_dying(h))) {
531 cfs_hash_get(s->ls_obj_hash, hnode);
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.
543 cfs_waitlink_init(waiter);
544 cfs_waitq_add(&bkt->lsb_marche_funebre, waiter);
545 cfs_set_current_state(CFS_TASK_UNINT);
546 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_DEATH_RACE);
547 return ERR_PTR(-EAGAIN);
551 * Search cache for an object with the fid \a f. If such object is found,
552 * return it. Otherwise, create new object, insert it into cache and return
553 * it. In any case, additional reference is acquired on the returned object.
555 struct lu_object *lu_object_find(const struct lu_env *env,
556 struct lu_device *dev, const struct lu_fid *f,
557 const struct lu_object_conf *conf)
559 return lu_object_find_at(env, dev->ld_site->ls_top_dev, f, conf);
561 EXPORT_SYMBOL(lu_object_find);
563 static struct lu_object *lu_object_new(const struct lu_env *env,
564 struct lu_device *dev,
565 const struct lu_fid *f,
566 const struct lu_object_conf *conf)
571 struct lu_site_bkt_data *bkt;
573 o = lu_object_alloc(env, dev, f, conf);
574 if (unlikely(IS_ERR(o)))
577 hs = dev->ld_site->ls_obj_hash;
578 cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
579 bkt = cfs_hash_bd_extra_get(hs, &bd);
580 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
582 cfs_hash_bd_unlock(hs, &bd, 1);
587 * Core logic of lu_object_find*() functions.
589 static struct lu_object *lu_object_find_try(const struct lu_env *env,
590 struct lu_device *dev,
591 const struct lu_fid *f,
592 const struct lu_object_conf *conf,
593 cfs_waitlink_t *waiter)
596 struct lu_object *shadow;
603 * This uses standard index maintenance protocol:
605 * - search index under lock, and return object if found;
606 * - otherwise, unlock index, allocate new object;
607 * - lock index and search again;
608 * - if nothing is found (usual case), insert newly created
610 * - otherwise (race: other thread inserted object), free
611 * object just allocated.
615 * For "LOC_F_NEW" case, we are sure the object is new established.
616 * It is unnecessary to perform lookup-alloc-lookup-insert, instead,
617 * just alloc and insert directly.
619 * If dying object is found during index search, add @waiter to the
620 * site wait-queue and return ERR_PTR(-EAGAIN).
622 if (conf != NULL && conf->loc_flags & LOC_F_NEW)
623 return lu_object_new(env, dev, f, conf);
627 cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
628 o = htable_lookup(s, &bd, f, waiter, &version);
629 cfs_hash_bd_unlock(hs, &bd, 1);
634 * Allocate new object. This may result in rather complicated
635 * operations, including fld queries, inode loading, etc.
637 o = lu_object_alloc(env, dev, f, conf);
638 if (unlikely(IS_ERR(o)))
641 LASSERT(lu_fid_eq(lu_object_fid(o), f));
643 cfs_hash_bd_lock(hs, &bd, 1);
645 shadow = htable_lookup(s, &bd, f, waiter, &version);
646 if (likely(shadow == NULL)) {
647 struct lu_site_bkt_data *bkt;
649 bkt = cfs_hash_bd_extra_get(hs, &bd);
650 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
652 cfs_hash_bd_unlock(hs, &bd, 1);
656 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_RACE);
657 cfs_hash_bd_unlock(hs, &bd, 1);
658 lu_object_free(env, o);
663 * Much like lu_object_find(), but top level device of object is specifically
664 * \a dev rather than top level device of the site. This interface allows
665 * objects of different "stacking" to be created within the same site.
667 struct lu_object *lu_object_find_at(const struct lu_env *env,
668 struct lu_device *dev,
669 const struct lu_fid *f,
670 const struct lu_object_conf *conf)
672 struct lu_site_bkt_data *bkt;
673 struct lu_object *obj;
677 obj = lu_object_find_try(env, dev, f, conf, &wait);
678 if (obj != ERR_PTR(-EAGAIN))
681 * lu_object_find_try() already added waiter into the
684 cfs_waitq_wait(&wait, CFS_TASK_UNINT);
685 bkt = lu_site_bkt_from_fid(dev->ld_site, (void *)f);
686 cfs_waitq_del(&bkt->lsb_marche_funebre, &wait);
689 EXPORT_SYMBOL(lu_object_find_at);
692 * Find object with given fid, and return its slice belonging to given device.
694 struct lu_object *lu_object_find_slice(const struct lu_env *env,
695 struct lu_device *dev,
696 const struct lu_fid *f,
697 const struct lu_object_conf *conf)
699 struct lu_object *top;
700 struct lu_object *obj;
702 top = lu_object_find(env, dev, f, conf);
704 obj = lu_object_locate(top->lo_header, dev->ld_type);
706 lu_object_put(env, top);
711 EXPORT_SYMBOL(lu_object_find_slice);
714 * Global list of all device types.
716 static CFS_LIST_HEAD(lu_device_types);
718 int lu_device_type_init(struct lu_device_type *ldt)
722 CFS_INIT_LIST_HEAD(&ldt->ldt_linkage);
723 if (ldt->ldt_ops->ldto_init)
724 result = ldt->ldt_ops->ldto_init(ldt);
726 cfs_list_add(&ldt->ldt_linkage, &lu_device_types);
729 EXPORT_SYMBOL(lu_device_type_init);
731 void lu_device_type_fini(struct lu_device_type *ldt)
733 cfs_list_del_init(&ldt->ldt_linkage);
734 if (ldt->ldt_ops->ldto_fini)
735 ldt->ldt_ops->ldto_fini(ldt);
737 EXPORT_SYMBOL(lu_device_type_fini);
739 void lu_types_stop(void)
741 struct lu_device_type *ldt;
743 cfs_list_for_each_entry(ldt, &lu_device_types, ldt_linkage) {
744 if (ldt->ldt_device_nr == 0 && ldt->ldt_ops->ldto_stop)
745 ldt->ldt_ops->ldto_stop(ldt);
748 EXPORT_SYMBOL(lu_types_stop);
751 * Global list of all sites on this node
753 static CFS_LIST_HEAD(lu_sites);
754 static DEFINE_MUTEX(lu_sites_guard);
757 * Global environment used by site shrinker.
759 static struct lu_env lu_shrink_env;
761 struct lu_site_print_arg {
762 struct lu_env *lsp_env;
764 lu_printer_t lsp_printer;
768 lu_site_obj_print(cfs_hash_t *hs, cfs_hash_bd_t *bd,
769 cfs_hlist_node_t *hnode, void *data)
771 struct lu_site_print_arg *arg = (struct lu_site_print_arg *)data;
772 struct lu_object_header *h;
774 h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
775 if (!cfs_list_empty(&h->loh_layers)) {
776 const struct lu_object *o;
778 o = lu_object_top(h);
779 lu_object_print(arg->lsp_env, arg->lsp_cookie,
780 arg->lsp_printer, o);
782 lu_object_header_print(arg->lsp_env, arg->lsp_cookie,
783 arg->lsp_printer, h);
789 * Print all objects in \a s.
791 void lu_site_print(const struct lu_env *env, struct lu_site *s, void *cookie,
792 lu_printer_t printer)
794 struct lu_site_print_arg arg = {
795 .lsp_env = (struct lu_env *)env,
796 .lsp_cookie = cookie,
797 .lsp_printer = printer,
800 cfs_hash_for_each(s->ls_obj_hash, lu_site_obj_print, &arg);
802 EXPORT_SYMBOL(lu_site_print);
805 LU_CACHE_PERCENT_MAX = 50,
806 LU_CACHE_PERCENT_DEFAULT = 20
809 static unsigned int lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
810 CFS_MODULE_PARM(lu_cache_percent, "i", int, 0644,
811 "Percentage of memory to be used as lu_object cache");
814 * Return desired hash table order.
816 static int lu_htable_order(void)
818 unsigned long cache_size;
822 * Calculate hash table size, assuming that we want reasonable
823 * performance when 20% of total memory is occupied by cache of
826 * Size of lu_object is (arbitrary) taken as 1K (together with inode).
828 cache_size = cfs_num_physpages;
830 #if BITS_PER_LONG == 32
831 /* limit hashtable size for lowmem systems to low RAM */
832 if (cache_size > 1 << (30 - CFS_PAGE_SHIFT))
833 cache_size = 1 << (30 - CFS_PAGE_SHIFT) * 3 / 4;
836 /* clear off unreasonable cache setting. */
837 if (lu_cache_percent == 0 || lu_cache_percent > LU_CACHE_PERCENT_MAX) {
838 CWARN("obdclass: invalid lu_cache_percent: %u, it must be in"
839 " the range of (0, %u]. Will use default value: %u.\n",
840 lu_cache_percent, LU_CACHE_PERCENT_MAX,
841 LU_CACHE_PERCENT_DEFAULT);
843 lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
845 cache_size = cache_size / 100 * lu_cache_percent *
846 (CFS_PAGE_SIZE / 1024);
848 for (bits = 1; (1 << bits) < cache_size; ++bits) {
854 static unsigned lu_obj_hop_hash(cfs_hash_t *hs,
855 const void *key, unsigned mask)
857 struct lu_fid *fid = (struct lu_fid *)key;
860 hash = fid_flatten32(fid);
861 hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
862 hash = cfs_hash_long(hash, hs->hs_bkt_bits);
864 /* give me another random factor */
865 hash -= cfs_hash_long((unsigned long)hs, fid_oid(fid) % 11 + 3);
867 hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
868 hash |= (fid_seq(fid) + fid_oid(fid)) & (CFS_HASH_NBKT(hs) - 1);
873 static void *lu_obj_hop_object(cfs_hlist_node_t *hnode)
875 return cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
878 static void *lu_obj_hop_key(cfs_hlist_node_t *hnode)
880 struct lu_object_header *h;
882 h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
886 static int lu_obj_hop_keycmp(const void *key, cfs_hlist_node_t *hnode)
888 struct lu_object_header *h;
890 h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
891 return lu_fid_eq(&h->loh_fid, (struct lu_fid *)key);
894 static void lu_obj_hop_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
896 struct lu_object_header *h;
898 h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
899 if (cfs_atomic_add_return(1, &h->loh_ref) == 1) {
900 struct lu_site_bkt_data *bkt;
903 cfs_hash_bd_get(hs, &h->loh_fid, &bd);
904 bkt = cfs_hash_bd_extra_get(hs, &bd);
909 static void lu_obj_hop_put_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
911 LBUG(); /* we should never called it */
914 cfs_hash_ops_t lu_site_hash_ops = {
915 .hs_hash = lu_obj_hop_hash,
916 .hs_key = lu_obj_hop_key,
917 .hs_keycmp = lu_obj_hop_keycmp,
918 .hs_object = lu_obj_hop_object,
919 .hs_get = lu_obj_hop_get,
920 .hs_put_locked = lu_obj_hop_put_locked,
923 void lu_dev_add_linkage(struct lu_site *s, struct lu_device *d)
925 spin_lock(&s->ls_ld_lock);
926 if (cfs_list_empty(&d->ld_linkage))
927 cfs_list_add(&d->ld_linkage, &s->ls_ld_linkage);
928 spin_unlock(&s->ls_ld_lock);
930 EXPORT_SYMBOL(lu_dev_add_linkage);
932 void lu_dev_del_linkage(struct lu_site *s, struct lu_device *d)
934 spin_lock(&s->ls_ld_lock);
935 cfs_list_del_init(&d->ld_linkage);
936 spin_unlock(&s->ls_ld_lock);
938 EXPORT_SYMBOL(lu_dev_del_linkage);
941 * Initialize site \a s, with \a d as the top level device.
943 #define LU_SITE_BITS_MIN 12
944 #define LU_SITE_BITS_MAX 24
946 * total 256 buckets, we don't want too many buckets because:
947 * - consume too much memory
948 * - avoid unbalanced LRU list
950 #define LU_SITE_BKT_BITS 8
952 int lu_site_init(struct lu_site *s, struct lu_device *top)
954 struct lu_site_bkt_data *bkt;
961 memset(s, 0, sizeof *s);
962 bits = lu_htable_order();
963 snprintf(name, 16, "lu_site_%s", top->ld_type->ldt_name);
964 for (bits = min(max(LU_SITE_BITS_MIN, bits), LU_SITE_BITS_MAX);
965 bits >= LU_SITE_BITS_MIN; bits--) {
966 s->ls_obj_hash = cfs_hash_create(name, bits, bits,
967 bits - LU_SITE_BKT_BITS,
970 CFS_HASH_SPIN_BKTLOCK |
971 CFS_HASH_NO_ITEMREF |
973 CFS_HASH_ASSERT_EMPTY);
974 if (s->ls_obj_hash != NULL)
978 if (s->ls_obj_hash == NULL) {
979 CERROR("failed to create lu_site hash with bits: %d\n", bits);
983 cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
984 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
985 CFS_INIT_LIST_HEAD(&bkt->lsb_lru);
986 cfs_waitq_init(&bkt->lsb_marche_funebre);
989 s->ls_stats = lprocfs_alloc_stats(LU_SS_LAST_STAT, 0);
990 if (s->ls_stats == NULL) {
991 cfs_hash_putref(s->ls_obj_hash);
992 s->ls_obj_hash = NULL;
996 lprocfs_counter_init(s->ls_stats, LU_SS_CREATED,
997 0, "created", "created");
998 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_HIT,
999 0, "cache_hit", "cache_hit");
1000 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_MISS,
1001 0, "cache_miss", "cache_miss");
1002 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_RACE,
1003 0, "cache_race", "cache_race");
1004 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_DEATH_RACE,
1005 0, "cache_death_race", "cache_death_race");
1006 lprocfs_counter_init(s->ls_stats, LU_SS_LRU_PURGED,
1007 0, "lru_purged", "lru_purged");
1009 CFS_INIT_LIST_HEAD(&s->ls_linkage);
1010 s->ls_top_dev = top;
1013 lu_ref_add(&top->ld_reference, "site-top", s);
1015 CFS_INIT_LIST_HEAD(&s->ls_ld_linkage);
1016 spin_lock_init(&s->ls_ld_lock);
1018 lu_dev_add_linkage(s, top);
1022 EXPORT_SYMBOL(lu_site_init);
1025 * Finalize \a s and release its resources.
1027 void lu_site_fini(struct lu_site *s)
1029 mutex_lock(&lu_sites_guard);
1030 cfs_list_del_init(&s->ls_linkage);
1031 mutex_unlock(&lu_sites_guard);
1033 if (s->ls_obj_hash != NULL) {
1034 cfs_hash_putref(s->ls_obj_hash);
1035 s->ls_obj_hash = NULL;
1038 if (s->ls_top_dev != NULL) {
1039 s->ls_top_dev->ld_site = NULL;
1040 lu_ref_del(&s->ls_top_dev->ld_reference, "site-top", s);
1041 lu_device_put(s->ls_top_dev);
1042 s->ls_top_dev = NULL;
1045 if (s->ls_stats != NULL)
1046 lprocfs_free_stats(&s->ls_stats);
1048 EXPORT_SYMBOL(lu_site_fini);
1051 * Called when initialization of stack for this site is completed.
1053 int lu_site_init_finish(struct lu_site *s)
1056 mutex_lock(&lu_sites_guard);
1057 result = lu_context_refill(&lu_shrink_env.le_ctx);
1059 cfs_list_add(&s->ls_linkage, &lu_sites);
1060 mutex_unlock(&lu_sites_guard);
1063 EXPORT_SYMBOL(lu_site_init_finish);
1066 * Acquire additional reference on device \a d
1068 void lu_device_get(struct lu_device *d)
1070 cfs_atomic_inc(&d->ld_ref);
1072 EXPORT_SYMBOL(lu_device_get);
1075 * Release reference on device \a d.
1077 void lu_device_put(struct lu_device *d)
1079 LASSERT(cfs_atomic_read(&d->ld_ref) > 0);
1080 cfs_atomic_dec(&d->ld_ref);
1082 EXPORT_SYMBOL(lu_device_put);
1085 * Initialize device \a d of type \a t.
1087 int lu_device_init(struct lu_device *d, struct lu_device_type *t)
1089 if (t->ldt_device_nr++ == 0 && t->ldt_ops->ldto_start != NULL)
1090 t->ldt_ops->ldto_start(t);
1091 memset(d, 0, sizeof *d);
1092 cfs_atomic_set(&d->ld_ref, 0);
1094 lu_ref_init(&d->ld_reference);
1095 CFS_INIT_LIST_HEAD(&d->ld_linkage);
1098 EXPORT_SYMBOL(lu_device_init);
1101 * Finalize device \a d.
1103 void lu_device_fini(struct lu_device *d)
1105 struct lu_device_type *t;
1108 if (d->ld_obd != NULL) {
1109 d->ld_obd->obd_lu_dev = NULL;
1113 lu_ref_fini(&d->ld_reference);
1114 LASSERTF(cfs_atomic_read(&d->ld_ref) == 0,
1115 "Refcount is %u\n", cfs_atomic_read(&d->ld_ref));
1116 LASSERT(t->ldt_device_nr > 0);
1117 if (--t->ldt_device_nr == 0 && t->ldt_ops->ldto_stop != NULL)
1118 t->ldt_ops->ldto_stop(t);
1120 EXPORT_SYMBOL(lu_device_fini);
1123 * Initialize object \a o that is part of compound object \a h and was created
1126 int lu_object_init(struct lu_object *o,
1127 struct lu_object_header *h, struct lu_device *d)
1129 memset(o, 0, sizeof *o);
1133 o->lo_dev_ref = lu_ref_add(&d->ld_reference, "lu_object", o);
1134 CFS_INIT_LIST_HEAD(&o->lo_linkage);
1137 EXPORT_SYMBOL(lu_object_init);
1140 * Finalize object and release its resources.
1142 void lu_object_fini(struct lu_object *o)
1144 struct lu_device *dev = o->lo_dev;
1146 LASSERT(cfs_list_empty(&o->lo_linkage));
1149 lu_ref_del_at(&dev->ld_reference,
1150 o->lo_dev_ref , "lu_object", o);
1155 EXPORT_SYMBOL(lu_object_fini);
1158 * Add object \a o as first layer of compound object \a h
1160 * This is typically called by the ->ldo_object_alloc() method of top-level
1163 void lu_object_add_top(struct lu_object_header *h, struct lu_object *o)
1165 cfs_list_move(&o->lo_linkage, &h->loh_layers);
1167 EXPORT_SYMBOL(lu_object_add_top);
1170 * Add object \a o as a layer of compound object, going after \a before.
1172 * This is typically called by the ->ldo_object_alloc() method of \a
1175 void lu_object_add(struct lu_object *before, struct lu_object *o)
1177 cfs_list_move(&o->lo_linkage, &before->lo_linkage);
1179 EXPORT_SYMBOL(lu_object_add);
1182 * Initialize compound object.
1184 int lu_object_header_init(struct lu_object_header *h)
1186 memset(h, 0, sizeof *h);
1187 cfs_atomic_set(&h->loh_ref, 1);
1188 CFS_INIT_HLIST_NODE(&h->loh_hash);
1189 CFS_INIT_LIST_HEAD(&h->loh_lru);
1190 CFS_INIT_LIST_HEAD(&h->loh_layers);
1191 lu_ref_init(&h->loh_reference);
1194 EXPORT_SYMBOL(lu_object_header_init);
1197 * Finalize compound object.
1199 void lu_object_header_fini(struct lu_object_header *h)
1201 LASSERT(cfs_list_empty(&h->loh_layers));
1202 LASSERT(cfs_list_empty(&h->loh_lru));
1203 LASSERT(cfs_hlist_unhashed(&h->loh_hash));
1204 lu_ref_fini(&h->loh_reference);
1206 EXPORT_SYMBOL(lu_object_header_fini);
1209 * Given a compound object, find its slice, corresponding to the device type
1212 struct lu_object *lu_object_locate(struct lu_object_header *h,
1213 const struct lu_device_type *dtype)
1215 struct lu_object *o;
1217 cfs_list_for_each_entry(o, &h->loh_layers, lo_linkage) {
1218 if (o->lo_dev->ld_type == dtype)
1223 EXPORT_SYMBOL(lu_object_locate);
1228 * Finalize and free devices in the device stack.
1230 * Finalize device stack by purging object cache, and calling
1231 * lu_device_type_operations::ldto_device_fini() and
1232 * lu_device_type_operations::ldto_device_free() on all devices in the stack.
1234 void lu_stack_fini(const struct lu_env *env, struct lu_device *top)
1236 struct lu_site *site = top->ld_site;
1237 struct lu_device *scan;
1238 struct lu_device *next;
1240 lu_site_purge(env, site, ~0);
1241 for (scan = top; scan != NULL; scan = next) {
1242 next = scan->ld_type->ldt_ops->ldto_device_fini(env, scan);
1243 lu_ref_del(&scan->ld_reference, "lu-stack", &lu_site_init);
1244 lu_device_put(scan);
1248 lu_site_purge(env, site, ~0);
1250 for (scan = top; scan != NULL; scan = next) {
1251 const struct lu_device_type *ldt = scan->ld_type;
1252 struct obd_type *type;
1254 next = ldt->ldt_ops->ldto_device_free(env, scan);
1255 type = ldt->ldt_obd_type;
1258 class_put_type(type);
1262 EXPORT_SYMBOL(lu_stack_fini);
1266 * Maximal number of tld slots.
1268 LU_CONTEXT_KEY_NR = 40
1271 static struct lu_context_key *lu_keys[LU_CONTEXT_KEY_NR] = { NULL, };
1273 static DEFINE_SPINLOCK(lu_keys_guard);
1276 * Global counter incremented whenever key is registered, unregistered,
1277 * revived or quiesced. This is used to void unnecessary calls to
1278 * lu_context_refill(). No locking is provided, as initialization and shutdown
1279 * are supposed to be externally serialized.
1281 static unsigned key_set_version = 0;
1286 int lu_context_key_register(struct lu_context_key *key)
1291 LASSERT(key->lct_init != NULL);
1292 LASSERT(key->lct_fini != NULL);
1293 LASSERT(key->lct_tags != 0);
1294 LASSERT(key->lct_owner != NULL);
1297 spin_lock(&lu_keys_guard);
1298 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1299 if (lu_keys[i] == NULL) {
1301 cfs_atomic_set(&key->lct_used, 1);
1303 lu_ref_init(&key->lct_reference);
1309 spin_unlock(&lu_keys_guard);
1312 EXPORT_SYMBOL(lu_context_key_register);
1314 static void key_fini(struct lu_context *ctx, int index)
1316 if (ctx->lc_value != NULL && ctx->lc_value[index] != NULL) {
1317 struct lu_context_key *key;
1319 key = lu_keys[index];
1320 LASSERT(key != NULL);
1321 LASSERT(key->lct_fini != NULL);
1322 LASSERT(cfs_atomic_read(&key->lct_used) > 1);
1324 key->lct_fini(ctx, key, ctx->lc_value[index]);
1325 lu_ref_del(&key->lct_reference, "ctx", ctx);
1326 cfs_atomic_dec(&key->lct_used);
1328 LASSERT(key->lct_owner != NULL);
1329 if ((ctx->lc_tags & LCT_NOREF) == 0) {
1330 LINVRNT(cfs_module_refcount(key->lct_owner) > 0);
1331 cfs_module_put(key->lct_owner);
1333 ctx->lc_value[index] = NULL;
1340 void lu_context_key_degister(struct lu_context_key *key)
1342 LASSERT(cfs_atomic_read(&key->lct_used) >= 1);
1343 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1345 lu_context_key_quiesce(key);
1348 spin_lock(&lu_keys_guard);
1349 key_fini(&lu_shrink_env.le_ctx, key->lct_index);
1350 if (lu_keys[key->lct_index]) {
1351 lu_keys[key->lct_index] = NULL;
1352 lu_ref_fini(&key->lct_reference);
1354 spin_unlock(&lu_keys_guard);
1356 LASSERTF(cfs_atomic_read(&key->lct_used) == 1,
1357 "key has instances: %d\n",
1358 cfs_atomic_read(&key->lct_used));
1360 EXPORT_SYMBOL(lu_context_key_degister);
1363 * Register a number of keys. This has to be called after all keys have been
1364 * initialized by a call to LU_CONTEXT_KEY_INIT().
1366 int lu_context_key_register_many(struct lu_context_key *k, ...)
1368 struct lu_context_key *key = k;
1374 result = lu_context_key_register(key);
1377 key = va_arg(args, struct lu_context_key *);
1378 } while (key != NULL);
1384 lu_context_key_degister(k);
1385 k = va_arg(args, struct lu_context_key *);
1392 EXPORT_SYMBOL(lu_context_key_register_many);
1395 * De-register a number of keys. This is a dual to
1396 * lu_context_key_register_many().
1398 void lu_context_key_degister_many(struct lu_context_key *k, ...)
1404 lu_context_key_degister(k);
1405 k = va_arg(args, struct lu_context_key*);
1406 } while (k != NULL);
1409 EXPORT_SYMBOL(lu_context_key_degister_many);
1412 * Revive a number of keys.
1414 void lu_context_key_revive_many(struct lu_context_key *k, ...)
1420 lu_context_key_revive(k);
1421 k = va_arg(args, struct lu_context_key*);
1422 } while (k != NULL);
1425 EXPORT_SYMBOL(lu_context_key_revive_many);
1428 * Quiescent a number of keys.
1430 void lu_context_key_quiesce_many(struct lu_context_key *k, ...)
1436 lu_context_key_quiesce(k);
1437 k = va_arg(args, struct lu_context_key*);
1438 } while (k != NULL);
1441 EXPORT_SYMBOL(lu_context_key_quiesce_many);
1444 * Return value associated with key \a key in context \a ctx.
1446 void *lu_context_key_get(const struct lu_context *ctx,
1447 const struct lu_context_key *key)
1449 LINVRNT(ctx->lc_state == LCS_ENTERED);
1450 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1451 LASSERT(lu_keys[key->lct_index] == key);
1452 return ctx->lc_value[key->lct_index];
1454 EXPORT_SYMBOL(lu_context_key_get);
1457 * List of remembered contexts. XXX document me.
1459 static CFS_LIST_HEAD(lu_context_remembered);
1462 * Destroy \a key in all remembered contexts. This is used to destroy key
1463 * values in "shared" contexts (like service threads), when a module owning
1464 * the key is about to be unloaded.
1466 void lu_context_key_quiesce(struct lu_context_key *key)
1468 struct lu_context *ctx;
1469 extern unsigned cl_env_cache_purge(unsigned nr);
1471 if (!(key->lct_tags & LCT_QUIESCENT)) {
1473 * XXX layering violation.
1475 cl_env_cache_purge(~0);
1476 key->lct_tags |= LCT_QUIESCENT;
1478 * XXX memory barrier has to go here.
1480 spin_lock(&lu_keys_guard);
1481 cfs_list_for_each_entry(ctx, &lu_context_remembered,
1483 key_fini(ctx, key->lct_index);
1484 spin_unlock(&lu_keys_guard);
1488 EXPORT_SYMBOL(lu_context_key_quiesce);
1490 void lu_context_key_revive(struct lu_context_key *key)
1492 key->lct_tags &= ~LCT_QUIESCENT;
1495 EXPORT_SYMBOL(lu_context_key_revive);
1497 static void keys_fini(struct lu_context *ctx)
1501 if (ctx->lc_value == NULL)
1504 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i)
1507 OBD_FREE(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1508 ctx->lc_value = NULL;
1511 static int keys_fill(struct lu_context *ctx)
1515 LINVRNT(ctx->lc_value != NULL);
1516 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1517 struct lu_context_key *key;
1520 if (ctx->lc_value[i] == NULL && key != NULL &&
1521 (key->lct_tags & ctx->lc_tags) &&
1523 * Don't create values for a LCT_QUIESCENT key, as this
1524 * will pin module owning a key.
1526 !(key->lct_tags & LCT_QUIESCENT)) {
1529 LINVRNT(key->lct_init != NULL);
1530 LINVRNT(key->lct_index == i);
1532 value = key->lct_init(ctx, key);
1533 if (unlikely(IS_ERR(value)))
1534 return PTR_ERR(value);
1536 LASSERT(key->lct_owner != NULL);
1537 if (!(ctx->lc_tags & LCT_NOREF))
1538 cfs_try_module_get(key->lct_owner);
1539 lu_ref_add_atomic(&key->lct_reference, "ctx", ctx);
1540 cfs_atomic_inc(&key->lct_used);
1542 * This is the only place in the code, where an
1543 * element of ctx->lc_value[] array is set to non-NULL
1546 ctx->lc_value[i] = value;
1547 if (key->lct_exit != NULL)
1548 ctx->lc_tags |= LCT_HAS_EXIT;
1550 ctx->lc_version = key_set_version;
1555 static int keys_init(struct lu_context *ctx)
1557 OBD_ALLOC(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1558 if (likely(ctx->lc_value != NULL))
1559 return keys_fill(ctx);
1565 * Initialize context data-structure. Create values for all keys.
1567 int lu_context_init(struct lu_context *ctx, __u32 tags)
1571 memset(ctx, 0, sizeof *ctx);
1572 ctx->lc_state = LCS_INITIALIZED;
1573 ctx->lc_tags = tags;
1574 if (tags & LCT_REMEMBER) {
1575 spin_lock(&lu_keys_guard);
1576 cfs_list_add(&ctx->lc_remember, &lu_context_remembered);
1577 spin_unlock(&lu_keys_guard);
1579 CFS_INIT_LIST_HEAD(&ctx->lc_remember);
1582 rc = keys_init(ctx);
1584 lu_context_fini(ctx);
1588 EXPORT_SYMBOL(lu_context_init);
1591 * Finalize context data-structure. Destroy key values.
1593 void lu_context_fini(struct lu_context *ctx)
1595 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1596 ctx->lc_state = LCS_FINALIZED;
1598 if ((ctx->lc_tags & LCT_REMEMBER) == 0) {
1599 LASSERT(cfs_list_empty(&ctx->lc_remember));
1602 } else { /* could race with key degister */
1603 spin_lock(&lu_keys_guard);
1605 cfs_list_del_init(&ctx->lc_remember);
1606 spin_unlock(&lu_keys_guard);
1609 EXPORT_SYMBOL(lu_context_fini);
1612 * Called before entering context.
1614 void lu_context_enter(struct lu_context *ctx)
1616 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1617 ctx->lc_state = LCS_ENTERED;
1619 EXPORT_SYMBOL(lu_context_enter);
1622 * Called after exiting from \a ctx
1624 void lu_context_exit(struct lu_context *ctx)
1628 LINVRNT(ctx->lc_state == LCS_ENTERED);
1629 ctx->lc_state = LCS_LEFT;
1630 if (ctx->lc_tags & LCT_HAS_EXIT && ctx->lc_value != NULL) {
1631 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1632 if (ctx->lc_value[i] != NULL) {
1633 struct lu_context_key *key;
1636 LASSERT(key != NULL);
1637 if (key->lct_exit != NULL)
1639 key, ctx->lc_value[i]);
1644 EXPORT_SYMBOL(lu_context_exit);
1647 * Allocate for context all missing keys that were registered after context
1648 * creation. key_set_version is only changed in rare cases when modules
1649 * are loaded and removed.
1651 int lu_context_refill(struct lu_context *ctx)
1653 return likely(ctx->lc_version == key_set_version) ? 0 : keys_fill(ctx);
1655 EXPORT_SYMBOL(lu_context_refill);
1658 * lu_ctx_tags/lu_ses_tags will be updated if there are new types of
1659 * obd being added. Currently, this is only used on client side, specifically
1660 * for echo device client, for other stack (like ptlrpc threads), context are
1661 * predefined when the lu_device type are registered, during the module probe
1664 __u32 lu_context_tags_default = 0;
1665 __u32 lu_session_tags_default = 0;
1667 void lu_context_tags_update(__u32 tags)
1669 spin_lock(&lu_keys_guard);
1670 lu_context_tags_default |= tags;
1672 spin_unlock(&lu_keys_guard);
1674 EXPORT_SYMBOL(lu_context_tags_update);
1676 void lu_context_tags_clear(__u32 tags)
1678 spin_lock(&lu_keys_guard);
1679 lu_context_tags_default &= ~tags;
1681 spin_unlock(&lu_keys_guard);
1683 EXPORT_SYMBOL(lu_context_tags_clear);
1685 void lu_session_tags_update(__u32 tags)
1687 spin_lock(&lu_keys_guard);
1688 lu_session_tags_default |= tags;
1690 spin_unlock(&lu_keys_guard);
1692 EXPORT_SYMBOL(lu_session_tags_update);
1694 void lu_session_tags_clear(__u32 tags)
1696 spin_lock(&lu_keys_guard);
1697 lu_session_tags_default &= ~tags;
1699 spin_unlock(&lu_keys_guard);
1701 EXPORT_SYMBOL(lu_session_tags_clear);
1703 int lu_env_init(struct lu_env *env, __u32 tags)
1708 result = lu_context_init(&env->le_ctx, tags);
1709 if (likely(result == 0))
1710 lu_context_enter(&env->le_ctx);
1713 EXPORT_SYMBOL(lu_env_init);
1715 void lu_env_fini(struct lu_env *env)
1717 lu_context_exit(&env->le_ctx);
1718 lu_context_fini(&env->le_ctx);
1721 EXPORT_SYMBOL(lu_env_fini);
1723 int lu_env_refill(struct lu_env *env)
1727 result = lu_context_refill(&env->le_ctx);
1728 if (result == 0 && env->le_ses != NULL)
1729 result = lu_context_refill(env->le_ses);
1732 EXPORT_SYMBOL(lu_env_refill);
1735 * Currently, this API will only be used by echo client.
1736 * Because echo client and normal lustre client will share
1737 * same cl_env cache. So echo client needs to refresh
1738 * the env context after it get one from the cache, especially
1739 * when normal client and echo client co-exist in the same client.
1741 int lu_env_refill_by_tags(struct lu_env *env, __u32 ctags,
1746 if ((env->le_ctx.lc_tags & ctags) != ctags) {
1747 env->le_ctx.lc_version = 0;
1748 env->le_ctx.lc_tags |= ctags;
1751 if (env->le_ses && (env->le_ses->lc_tags & stags) != stags) {
1752 env->le_ses->lc_version = 0;
1753 env->le_ses->lc_tags |= stags;
1756 result = lu_env_refill(env);
1760 EXPORT_SYMBOL(lu_env_refill_by_tags);
1762 static struct cfs_shrinker *lu_site_shrinker = NULL;
1764 typedef struct lu_site_stats{
1765 unsigned lss_populated;
1766 unsigned lss_max_search;
1771 static void lu_site_stats_get(cfs_hash_t *hs,
1772 lu_site_stats_t *stats, int populated)
1777 cfs_hash_for_each_bucket(hs, &bd, i) {
1778 struct lu_site_bkt_data *bkt = cfs_hash_bd_extra_get(hs, &bd);
1779 cfs_hlist_head_t *hhead;
1781 cfs_hash_bd_lock(hs, &bd, 1);
1782 stats->lss_busy += bkt->lsb_busy;
1783 stats->lss_total += cfs_hash_bd_count_get(&bd);
1784 stats->lss_max_search = max((int)stats->lss_max_search,
1785 cfs_hash_bd_depmax_get(&bd));
1787 cfs_hash_bd_unlock(hs, &bd, 1);
1791 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
1792 if (!cfs_hlist_empty(hhead))
1793 stats->lss_populated++;
1795 cfs_hash_bd_unlock(hs, &bd, 1);
1802 * There exists a potential lock inversion deadlock scenario when using
1803 * Lustre on top of ZFS. This occurs between one of ZFS's
1804 * buf_hash_table.ht_lock's, and Lustre's lu_sites_guard lock. Essentially,
1805 * thread A will take the lu_sites_guard lock and sleep on the ht_lock,
1806 * while thread B will take the ht_lock and sleep on the lu_sites_guard
1807 * lock. Obviously neither thread will wake and drop their respective hold
1810 * To prevent this from happening we must ensure the lu_sites_guard lock is
1811 * not taken while down this code path. ZFS reliably does not set the
1812 * __GFP_FS bit in its code paths, so this can be used to determine if it
1813 * is safe to take the lu_sites_guard lock.
1815 * Ideally we should accurately return the remaining number of cached
1816 * objects without taking the lu_sites_guard lock, but this is not
1817 * possible in the current implementation.
1819 static int lu_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
1821 lu_site_stats_t stats;
1823 struct lu_site *tmp;
1825 int remain = shrink_param(sc, nr_to_scan);
1826 CFS_LIST_HEAD(splice);
1828 if (!(shrink_param(sc, gfp_mask) & __GFP_FS)) {
1832 /* We must not take the lu_sites_guard lock when
1833 * __GFP_FS is *not* set because of the deadlock
1834 * possibility detailed above. Additionally,
1835 * since we cannot determine the number of
1836 * objects in the cache without taking this
1837 * lock, we're in a particularly tough spot. As
1838 * a result, we'll just lie and say our cache is
1839 * empty. This _should_ be ok, as we can't
1840 * reclaim objects when __GFP_FS is *not* set
1846 CDEBUG(D_INODE, "Shrink %d objects\n", remain);
1848 mutex_lock(&lu_sites_guard);
1849 cfs_list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
1850 if (shrink_param(sc, nr_to_scan) != 0) {
1851 remain = lu_site_purge(&lu_shrink_env, s, remain);
1853 * Move just shrunk site to the tail of site list to
1854 * assure shrinking fairness.
1856 cfs_list_move_tail(&s->ls_linkage, &splice);
1859 memset(&stats, 0, sizeof(stats));
1860 lu_site_stats_get(s->ls_obj_hash, &stats, 0);
1861 cached += stats.lss_total - stats.lss_busy;
1862 if (shrink_param(sc, nr_to_scan) && remain <= 0)
1865 cfs_list_splice(&splice, lu_sites.prev);
1866 mutex_unlock(&lu_sites_guard);
1868 cached = (cached / 100) * sysctl_vfs_cache_pressure;
1869 if (shrink_param(sc, nr_to_scan) == 0)
1870 CDEBUG(D_INODE, "%d objects cached\n", cached);
1879 * Environment to be used in debugger, contains all tags.
1881 struct lu_env lu_debugging_env;
1884 * Debugging printer function using printk().
1886 int lu_printk_printer(const struct lu_env *env,
1887 void *unused, const char *format, ...)
1891 va_start(args, format);
1892 vprintk(format, args);
1897 void lu_debugging_setup(void)
1899 lu_env_init(&lu_debugging_env, ~0);
1902 void lu_context_keys_dump(void)
1906 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1907 struct lu_context_key *key;
1911 CERROR("[%d]: %p %x (%p,%p,%p) %d %d \"%s\"@%p\n",
1912 i, key, key->lct_tags,
1913 key->lct_init, key->lct_fini, key->lct_exit,
1914 key->lct_index, cfs_atomic_read(&key->lct_used),
1915 key->lct_owner ? key->lct_owner->name : "",
1917 lu_ref_print(&key->lct_reference);
1921 EXPORT_SYMBOL(lu_context_keys_dump);
1922 #else /* !__KERNEL__ */
1923 static int lu_cache_shrink(int nr, unsigned int gfp_mask)
1927 #endif /* __KERNEL__ */
1929 int cl_global_init(void);
1930 void cl_global_fini(void);
1931 int lu_ref_global_init(void);
1932 void lu_ref_global_fini(void);
1934 int dt_global_init(void);
1935 void dt_global_fini(void);
1937 int llo_global_init(void);
1938 void llo_global_fini(void);
1940 /* context key constructor/destructor: lu_ucred_key_init, lu_ucred_key_fini */
1941 LU_KEY_INIT_FINI(lu_ucred, struct lu_ucred);
1943 static struct lu_context_key lu_ucred_key = {
1944 .lct_tags = LCT_SESSION,
1945 .lct_init = lu_ucred_key_init,
1946 .lct_fini = lu_ucred_key_fini
1950 * Get ucred key if session exists and ucred key is allocated on it.
1951 * Return NULL otherwise.
1953 struct lu_ucred *lu_ucred(const struct lu_env *env)
1957 return lu_context_key_get(env->le_ses, &lu_ucred_key);
1959 EXPORT_SYMBOL(lu_ucred);
1962 * Get ucred key and check if it is properly initialized.
1963 * Return NULL otherwise.
1965 struct lu_ucred *lu_ucred_check(const struct lu_env *env)
1967 struct lu_ucred *uc = lu_ucred(env);
1968 if (uc && uc->uc_valid != UCRED_OLD && uc->uc_valid != UCRED_NEW)
1972 EXPORT_SYMBOL(lu_ucred_check);
1975 * Get ucred key, which must exist and must be properly initialized.
1978 struct lu_ucred *lu_ucred_assert(const struct lu_env *env)
1980 struct lu_ucred *uc = lu_ucred_check(env);
1981 LASSERT(uc != NULL);
1984 EXPORT_SYMBOL(lu_ucred_assert);
1987 * Initialization of global lu_* data.
1989 int lu_global_init(void)
1993 CDEBUG(D_INFO, "Lustre LU module (%p).\n", &lu_keys);
1995 result = lu_ref_global_init();
1999 LU_CONTEXT_KEY_INIT(&lu_global_key);
2000 result = lu_context_key_register(&lu_global_key);
2004 LU_CONTEXT_KEY_INIT(&lu_ucred_key);
2005 result = lu_context_key_register(&lu_ucred_key);
2010 * At this level, we don't know what tags are needed, so allocate them
2011 * conservatively. This should not be too bad, because this
2012 * environment is global.
2014 mutex_lock(&lu_sites_guard);
2015 result = lu_env_init(&lu_shrink_env, LCT_SHRINKER);
2016 mutex_unlock(&lu_sites_guard);
2021 * seeks estimation: 3 seeks to read a record from oi, one to read
2022 * inode, one for ea. Unfortunately setting this high value results in
2023 * lu_object/inode cache consuming all the memory.
2025 lu_site_shrinker = cfs_set_shrinker(CFS_DEFAULT_SEEKS, lu_cache_shrink);
2026 if (lu_site_shrinker == NULL)
2030 result = dt_global_init();
2034 result = llo_global_init();
2038 result = cl_global_init();
2044 * Dual to lu_global_init().
2046 void lu_global_fini(void)
2053 if (lu_site_shrinker != NULL) {
2054 cfs_remove_shrinker(lu_site_shrinker);
2055 lu_site_shrinker = NULL;
2058 lu_context_key_degister(&lu_global_key);
2059 lu_context_key_degister(&lu_ucred_key);
2062 * Tear shrinker environment down _after_ de-registering
2063 * lu_global_key, because the latter has a value in the former.
2065 mutex_lock(&lu_sites_guard);
2066 lu_env_fini(&lu_shrink_env);
2067 mutex_unlock(&lu_sites_guard);
2069 lu_ref_global_fini();
2072 struct lu_buf LU_BUF_NULL = {
2076 EXPORT_SYMBOL(LU_BUF_NULL);
2078 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx)
2081 struct lprocfs_counter ret;
2083 lprocfs_stats_collect(stats, idx, &ret);
2084 return (__u32)ret.lc_count;
2091 * Output site statistical counters into a buffer. Suitable for
2092 * lprocfs_rd_*()-style functions.
2094 int lu_site_stats_print(const struct lu_site *s, char *page, int count)
2096 lu_site_stats_t stats;
2098 memset(&stats, 0, sizeof(stats));
2099 lu_site_stats_get(s->ls_obj_hash, &stats, 1);
2101 return snprintf(page, count, "%d/%d %d/%d %d %d %d %d %d %d %d\n",
2104 stats.lss_populated,
2105 CFS_HASH_NHLIST(s->ls_obj_hash),
2106 stats.lss_max_search,
2107 ls_stats_read(s->ls_stats, LU_SS_CREATED),
2108 ls_stats_read(s->ls_stats, LU_SS_CACHE_HIT),
2109 ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
2110 ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
2111 ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
2112 ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED));
2114 EXPORT_SYMBOL(lu_site_stats_print);
2117 * Helper function to initialize a number of kmem slab caches at once.
2119 int lu_kmem_init(struct lu_kmem_descr *caches)
2122 struct lu_kmem_descr *iter = caches;
2124 for (result = 0; iter->ckd_cache != NULL; ++iter) {
2125 *iter->ckd_cache = cfs_mem_cache_create(iter->ckd_name,
2128 if (*iter->ckd_cache == NULL) {
2130 /* free all previously allocated caches */
2131 lu_kmem_fini(caches);
2137 EXPORT_SYMBOL(lu_kmem_init);
2140 * Helper function to finalize a number of kmem slab cached at once. Dual to
2143 void lu_kmem_fini(struct lu_kmem_descr *caches)
2147 for (; caches->ckd_cache != NULL; ++caches) {
2148 if (*caches->ckd_cache != NULL) {
2149 rc = cfs_mem_cache_destroy(*caches->ckd_cache);
2150 LASSERTF(rc == 0, "couldn't destroy %s slab\n",
2152 *caches->ckd_cache = NULL;
2156 EXPORT_SYMBOL(lu_kmem_fini);
2159 * Temporary solution to be able to assign fid in ->do_create()
2160 * till we have fully-functional OST fids
2162 void lu_object_assign_fid(const struct lu_env *env, struct lu_object *o,
2163 const struct lu_fid *fid)
2165 struct lu_site *s = o->lo_dev->ld_site;
2166 struct lu_fid *old = &o->lo_header->loh_fid;
2167 struct lu_site_bkt_data *bkt;
2168 struct lu_object *shadow;
2169 cfs_waitlink_t waiter;
2174 LASSERT(fid_is_zero(old));
2176 hs = s->ls_obj_hash;
2177 cfs_hash_bd_get_and_lock(hs, (void *)fid, &bd, 1);
2178 shadow = htable_lookup(s, &bd, fid, &waiter, &version);
2179 /* supposed to be unique */
2180 LASSERT(shadow == NULL);
2182 bkt = cfs_hash_bd_extra_get(hs, &bd);
2183 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
2185 cfs_hash_bd_unlock(hs, &bd, 1);
2187 EXPORT_SYMBOL(lu_object_assign_fid);
2190 * allocates object with 0 (non-assiged) fid
2191 * XXX: temporary solution to be able to assign fid in ->do_create()
2192 * till we have fully-functional OST fids
2194 struct lu_object *lu_object_anon(const struct lu_env *env,
2195 struct lu_device *dev,
2196 const struct lu_object_conf *conf)
2199 struct lu_object *o;
2202 o = lu_object_alloc(env, dev, &fid, conf);
2206 EXPORT_SYMBOL(lu_object_anon);