1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
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 # define EXPORT_SYMTAB
50 #include <libcfs/libcfs.h>
53 # include <linux/module.h>
57 #include <libcfs/libcfs_hash.h>
58 #include <obd_class.h>
59 #include <obd_support.h>
60 #include <lustre_disk.h>
61 #include <lustre_fid.h>
62 #include <lu_object.h>
63 #include <libcfs/list.h>
64 /* lu_time_global_{init,fini}() */
67 static void lu_object_free(const struct lu_env *env, struct lu_object *o);
70 * Decrease reference counter on object. If last reference is freed, return
71 * object to the cache, unless lu_object_is_dying(o) holds. In the latter
72 * case, free object immediately.
74 void lu_object_put(const struct lu_env *env, struct lu_object *o)
76 struct lu_object_header *top;
78 struct lu_object *orig;
82 site = o->lo_dev->ld_site;
85 write_lock(&site->ls_guard);
86 if (atomic_dec_and_test(&top->loh_ref)) {
88 * When last reference is released, iterate over object
89 * layers, and notify them that object is no longer busy.
91 list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
92 if (o->lo_ops->loo_object_release != NULL)
93 o->lo_ops->loo_object_release(env, o);
96 if (lu_object_is_dying(top)) {
98 * If object is dying (will not be cached), removed it
99 * from hash table and LRU.
101 * This is done with hash table and LRU lists
102 * locked. As the only way to acquire first reference
103 * to previously unreferenced object is through
104 * hash-table lookup (lu_object_find()), or LRU
105 * scanning (lu_site_purge()), that are done under
106 * hash-table and LRU lock, no race with concurrent
107 * object lookup is possible and we can safely destroy
110 hlist_del_init(&top->loh_hash);
111 list_del_init(&top->loh_lru);
116 write_unlock(&site->ls_guard);
119 * Object was already removed from hash and lru above, can
122 lu_object_free(env, orig);
124 EXPORT_SYMBOL(lu_object_put);
127 * Allocate new object.
129 * This follows object creation protocol, described in the comment within
130 * struct lu_device_operations definition.
132 static struct lu_object *lu_object_alloc(const struct lu_env *env,
133 struct lu_device *dev,
134 const struct lu_fid *f,
135 const struct lu_object_conf *conf)
137 struct lu_object *scan;
138 struct lu_object *top;
139 struct list_head *layers;
145 * Create top-level object slice. This will also create
148 top = dev->ld_ops->ldo_object_alloc(env, NULL, dev);
150 RETURN(ERR_PTR(-ENOMEM));
152 * This is the only place where object fid is assigned. It's constant
155 LASSERT(fid_is_igif(f) || fid_ver(f) == 0);
156 top->lo_header->loh_fid = *f;
157 layers = &top->lo_header->loh_layers;
160 * Call ->loo_object_init() repeatedly, until no more new
161 * object slices are created.
164 list_for_each_entry(scan, layers, lo_linkage) {
165 if (scan->lo_flags & LU_OBJECT_ALLOCATED)
168 scan->lo_header = top->lo_header;
169 result = scan->lo_ops->loo_object_init(env, scan, conf);
171 lu_object_free(env, top);
172 RETURN(ERR_PTR(result));
174 scan->lo_flags |= LU_OBJECT_ALLOCATED;
178 list_for_each_entry_reverse(scan, layers, lo_linkage) {
179 if (scan->lo_ops->loo_object_start != NULL) {
180 result = scan->lo_ops->loo_object_start(env, scan);
182 lu_object_free(env, top);
183 RETURN(ERR_PTR(result));
188 dev->ld_site->ls_stats.s_created ++;
195 static void lu_object_free(const struct lu_env *env, struct lu_object *o)
197 struct list_head splice;
198 struct lu_object *scan;
199 struct lu_site *site;
200 struct list_head *layers;
202 site = o->lo_dev->ld_site;
203 layers = &o->lo_header->loh_layers;
205 * First call ->loo_object_delete() method to release all resources.
207 list_for_each_entry_reverse(scan, layers, lo_linkage) {
208 if (scan->lo_ops->loo_object_delete != NULL)
209 scan->lo_ops->loo_object_delete(env, scan);
213 * Then, splice object layers into stand-alone list, and call
214 * ->loo_object_free() on all layers to free memory. Splice is
215 * necessary, because lu_object_header is freed together with the
218 CFS_INIT_LIST_HEAD(&splice);
219 list_splice_init(layers, &splice);
220 while (!list_empty(&splice)) {
222 * Free layers in bottom-to-top order, so that object header
223 * lives as long as possible and ->loo_object_free() methods
224 * can look at its contents.
226 o = container_of0(splice.prev, struct lu_object, lo_linkage);
227 list_del_init(&o->lo_linkage);
228 LASSERT(o->lo_ops->loo_object_free != NULL);
229 o->lo_ops->loo_object_free(env, o);
231 cfs_waitq_broadcast(&site->ls_marche_funebre);
235 * Free \a nr objects from the cold end of the site LRU list.
237 int lu_site_purge(const struct lu_env *env, struct lu_site *s, int nr)
239 struct list_head dispose;
240 struct lu_object_header *h;
241 struct lu_object_header *temp;
243 CFS_INIT_LIST_HEAD(&dispose);
245 * Under LRU list lock, scan LRU list and move unreferenced objects to
246 * the dispose list, removing them from LRU and hash table.
248 write_lock(&s->ls_guard);
249 list_for_each_entry_safe(h, temp, &s->ls_lru, loh_lru) {
251 * Objects are sorted in lru order, and "busy" objects (ones
252 * with h->loh_ref > 0) naturally tend to live near hot end
253 * that we scan last. Unfortunately, sites usually have small
254 * (less then ten) number of busy yet rarely accessed objects
255 * (some global objects, accessed directly through pointers,
256 * bypassing hash table). Currently algorithm scans them over
257 * and over again. Probably we should move busy objects out of
258 * LRU, or we can live with that.
262 if (atomic_read(&h->loh_ref) > 0)
264 hlist_del_init(&h->loh_hash);
265 list_move(&h->loh_lru, &dispose);
268 write_unlock(&s->ls_guard);
270 * Free everything on the dispose list. This is safe against races due
271 * to the reasons described in lu_object_put().
273 while (!list_empty(&dispose)) {
274 h = container_of0(dispose.next,
275 struct lu_object_header, loh_lru);
276 list_del_init(&h->loh_lru);
277 lu_object_free(env, lu_object_top(h));
278 s->ls_stats.s_lru_purged ++;
282 EXPORT_SYMBOL(lu_site_purge);
287 * Code below has to jump through certain loops to output object description
288 * into libcfs_debug_msg-based log. The problem is that lu_object_print()
289 * composes object description from strings that are parts of _lines_ of
290 * output (i.e., strings that are not terminated by newline). This doesn't fit
291 * very well into libcfs_debug_msg() interface that assumes that each message
292 * supplied to it is a self-contained output line.
294 * To work around this, strings are collected in a temporary buffer
295 * (implemented as a value of lu_cdebug_key key), until terminating newline
296 * character is detected.
304 * XXX overflow is not handled correctly.
309 struct lu_cdebug_data {
313 char lck_area[LU_CDEBUG_LINE];
316 /* context key constructor/destructor: lu_global_key_init, lu_global_key_fini */
317 LU_KEY_INIT_FINI(lu_global, struct lu_cdebug_data);
320 * Key, holding temporary buffer. This key is registered very early by
323 struct lu_context_key lu_global_key = {
324 .lct_tags = LCT_MD_THREAD|LCT_DT_THREAD|LCT_CL_THREAD,
325 .lct_init = lu_global_key_init,
326 .lct_fini = lu_global_key_fini
330 * Printer function emitting messages through libcfs_debug_msg().
332 int lu_cdebug_printer(const struct lu_env *env,
333 void *cookie, const char *format, ...)
335 struct lu_cdebug_print_info *info = cookie;
336 struct lu_cdebug_data *key;
341 va_start(args, format);
343 key = lu_context_key_get(&env->le_ctx, &lu_global_key);
344 LASSERT(key != NULL);
346 used = strlen(key->lck_area);
347 complete = format[strlen(format) - 1] == '\n';
349 * Append new chunk to the buffer.
351 vsnprintf(key->lck_area + used,
352 ARRAY_SIZE(key->lck_area) - used, format, args);
354 libcfs_debug_msg(NULL, info->lpi_subsys, info->lpi_mask,
355 (char *)info->lpi_file, info->lpi_fn,
356 info->lpi_line, "%s", key->lck_area);
357 key->lck_area[0] = 0;
362 EXPORT_SYMBOL(lu_cdebug_printer);
365 * Print object header.
367 static void lu_object_header_print(const struct lu_env *env,
368 void *cookie, lu_printer_t printer,
369 const struct lu_object_header *hdr)
371 (*printer)(env, cookie, "header@%p[%#lx, %d, "DFID"%s%s%s]",
372 hdr, hdr->loh_flags, atomic_read(&hdr->loh_ref),
374 hlist_unhashed(&hdr->loh_hash) ? "" : " hash",
375 list_empty(&hdr->loh_lru) ? "" : " lru",
376 hdr->loh_attr & LOHA_EXISTS ? " exist":"");
380 * Print human readable representation of the @o to the @printer.
382 void lu_object_print(const struct lu_env *env, void *cookie,
383 lu_printer_t printer, const struct lu_object *o)
385 static const char ruler[] = "........................................";
386 struct lu_object_header *top;
390 lu_object_header_print(env, cookie, printer, top);
391 (*printer)(env, cookie, "\n");
392 list_for_each_entry(o, &top->loh_layers, lo_linkage) {
393 depth = o->lo_depth + 4;
394 LASSERT(o->lo_ops->loo_object_print != NULL);
396 * print `.' @depth times.
398 (*printer)(env, cookie, "%*.*s", depth, depth, ruler);
399 o->lo_ops->loo_object_print(env, cookie, printer, o);
400 (*printer)(env, cookie, "\n");
403 EXPORT_SYMBOL(lu_object_print);
406 * Check object consistency.
408 int lu_object_invariant(const struct lu_object *o)
410 struct lu_object_header *top;
413 list_for_each_entry(o, &top->loh_layers, lo_linkage) {
414 if (o->lo_ops->loo_object_invariant != NULL &&
415 !o->lo_ops->loo_object_invariant(o))
420 EXPORT_SYMBOL(lu_object_invariant);
422 static struct lu_object *htable_lookup(struct lu_site *s,
423 const struct hlist_head *bucket,
424 const struct lu_fid *f,
425 cfs_waitlink_t *waiter)
427 struct lu_object_header *h;
428 struct hlist_node *scan;
430 hlist_for_each_entry(h, scan, bucket, loh_hash) {
431 s->ls_stats.s_cache_check ++;
432 if (likely(lu_fid_eq(&h->loh_fid, f))) {
433 if (unlikely(lu_object_is_dying(h))) {
435 * Lookup found an object being destroyed;
436 * this object cannot be returned (to assure
437 * that references to dying objects are
438 * eventually drained), and moreover, lookup
439 * has to wait until object is freed.
441 cfs_waitlink_init(waiter);
442 cfs_waitq_add(&s->ls_marche_funebre, waiter);
443 set_current_state(CFS_TASK_UNINT);
444 s->ls_stats.s_cache_death_race ++;
445 return ERR_PTR(-EAGAIN);
447 /* bump reference count... */
448 if (atomic_add_return(1, &h->loh_ref) == 1)
450 /* and move to the head of the LRU */
452 * XXX temporary disable this to measure effects of
453 * read-write locking.
455 /* list_move_tail(&h->loh_lru, &s->ls_lru); */
456 s->ls_stats.s_cache_hit ++;
457 return lu_object_top(h);
460 s->ls_stats.s_cache_miss ++;
464 static __u32 fid_hash(const struct lu_fid *f, int bits)
466 /* all objects with same id and different versions will belong to same
467 * collisions list. */
468 return hash_long(fid_flatten(f), bits);
472 * Search cache for an object with the fid \a f. If such object is found,
473 * return it. Otherwise, create new object, insert it into cache and return
474 * it. In any case, additional reference is acquired on the returned object.
476 struct lu_object *lu_object_find(const struct lu_env *env,
477 struct lu_device *dev, const struct lu_fid *f,
478 const struct lu_object_conf *conf)
480 return lu_object_find_at(env, dev->ld_site->ls_top_dev, f, conf);
482 EXPORT_SYMBOL(lu_object_find);
485 * Core logic of lu_object_find*() functions.
487 static struct lu_object *lu_object_find_try(const struct lu_env *env,
488 struct lu_device *dev,
489 const struct lu_fid *f,
490 const struct lu_object_conf *conf,
491 cfs_waitlink_t *waiter)
495 struct lu_object *shadow;
496 struct hlist_head *bucket;
499 * This uses standard index maintenance protocol:
501 * - search index under lock, and return object if found;
502 * - otherwise, unlock index, allocate new object;
503 * - lock index and search again;
504 * - if nothing is found (usual case), insert newly created
506 * - otherwise (race: other thread inserted object), free
507 * object just allocated.
511 * If dying object is found during index search, add @waiter to the
512 * site wait-queue and return ERR_PTR(-EAGAIN).
516 bucket = s->ls_hash + fid_hash(f, s->ls_hash_bits);
518 read_lock(&s->ls_guard);
519 o = htable_lookup(s, bucket, f, waiter);
520 read_unlock(&s->ls_guard);
526 * Allocate new object. This may result in rather complicated
527 * operations, including fld queries, inode loading, etc.
529 o = lu_object_alloc(env, dev, f, conf);
530 if (unlikely(IS_ERR(o)))
533 LASSERT(lu_fid_eq(lu_object_fid(o), f));
535 write_lock(&s->ls_guard);
536 shadow = htable_lookup(s, bucket, f, waiter);
537 if (likely(shadow == NULL)) {
538 hlist_add_head(&o->lo_header->loh_hash, bucket);
539 list_add_tail(&o->lo_header->loh_lru, &s->ls_lru);
545 s->ls_stats.s_cache_race ++;
546 write_unlock(&s->ls_guard);
548 lu_object_free(env, o);
553 * Much like lu_object_find(), but top level device of object is specifically
554 * \a dev rather than top level device of the site. This interface allows
555 * objects of different "stacking" to be created within the same site.
557 struct lu_object *lu_object_find_at(const struct lu_env *env,
558 struct lu_device *dev,
559 const struct lu_fid *f,
560 const struct lu_object_conf *conf)
562 struct lu_object *obj;
566 obj = lu_object_find_try(env, dev, f, conf, &wait);
567 if (obj == ERR_PTR(-EAGAIN)) {
569 * lu_object_find_try() already added waiter into the
572 cfs_waitq_wait(&wait, CFS_TASK_UNINT);
573 cfs_waitq_del(&dev->ld_site->ls_marche_funebre, &wait);
579 EXPORT_SYMBOL(lu_object_find_at);
582 * Find object with given fid, and return its slice belonging to given device.
584 struct lu_object *lu_object_find_slice(const struct lu_env *env,
585 struct lu_device *dev,
586 const struct lu_fid *f,
587 const struct lu_object_conf *conf)
589 struct lu_object *top;
590 struct lu_object *obj;
592 top = lu_object_find(env, dev, f, conf);
594 obj = lu_object_locate(top->lo_header, dev->ld_type);
596 lu_object_put(env, top);
601 EXPORT_SYMBOL(lu_object_find_slice);
604 * Global list of all device types.
606 static CFS_LIST_HEAD(lu_device_types);
608 int lu_device_type_init(struct lu_device_type *ldt)
612 CFS_INIT_LIST_HEAD(&ldt->ldt_linkage);
613 result = ldt->ldt_ops->ldto_init(ldt);
615 list_add(&ldt->ldt_linkage, &lu_device_types);
618 EXPORT_SYMBOL(lu_device_type_init);
620 void lu_device_type_fini(struct lu_device_type *ldt)
622 list_del_init(&ldt->ldt_linkage);
623 ldt->ldt_ops->ldto_fini(ldt);
625 EXPORT_SYMBOL(lu_device_type_fini);
627 void lu_types_stop(void)
629 struct lu_device_type *ldt;
631 list_for_each_entry(ldt, &lu_device_types, ldt_linkage) {
632 if (ldt->ldt_device_nr == 0)
633 ldt->ldt_ops->ldto_stop(ldt);
636 EXPORT_SYMBOL(lu_types_stop);
639 * Global list of all sites on this node
641 static CFS_LIST_HEAD(lu_sites);
642 static DECLARE_MUTEX(lu_sites_guard);
645 * Global environment used by site shrinker.
647 static struct lu_env lu_shrink_env;
650 * Print all objects in \a s.
652 void lu_site_print(const struct lu_env *env, struct lu_site *s, void *cookie,
653 lu_printer_t printer)
657 for (i = 0; i < s->ls_hash_size; ++i) {
658 struct lu_object_header *h;
659 struct hlist_node *scan;
661 read_lock(&s->ls_guard);
662 hlist_for_each_entry(h, scan, &s->ls_hash[i], loh_hash) {
664 if (!list_empty(&h->loh_layers)) {
665 const struct lu_object *obj;
667 obj = lu_object_top(h);
668 lu_object_print(env, cookie, printer, obj);
670 lu_object_header_print(env, cookie, printer, h);
672 read_unlock(&s->ls_guard);
675 EXPORT_SYMBOL(lu_site_print);
678 LU_CACHE_PERCENT = 20,
682 * Return desired hash table order.
684 static int lu_htable_order(void)
686 unsigned long cache_size;
690 * Calculate hash table size, assuming that we want reasonable
691 * performance when 20% of total memory is occupied by cache of
694 * Size of lu_object is (arbitrary) taken as 1K (together with inode).
696 cache_size = num_physpages;
698 #if BITS_PER_LONG == 32
699 /* limit hashtable size for lowmem systems to low RAM */
700 if (cache_size > 1 << (30 - CFS_PAGE_SHIFT))
701 cache_size = 1 << (30 - CFS_PAGE_SHIFT) * 3 / 4;
704 cache_size = cache_size / 100 * LU_CACHE_PERCENT *
705 (CFS_PAGE_SIZE / 1024);
707 for (bits = 1; (1 << bits) < cache_size; ++bits) {
713 static struct lock_class_key lu_site_guard_class;
716 * Initialize site \a s, with \a d as the top level device.
718 int lu_site_init(struct lu_site *s, struct lu_device *top)
725 memset(s, 0, sizeof *s);
726 rwlock_init(&s->ls_guard);
727 lockdep_set_class(&s->ls_guard, &lu_site_guard_class);
728 CFS_INIT_LIST_HEAD(&s->ls_lru);
729 CFS_INIT_LIST_HEAD(&s->ls_linkage);
730 cfs_waitq_init(&s->ls_marche_funebre);
734 lu_ref_add(&top->ld_reference, "site-top", s);
736 for (bits = lu_htable_order(), size = 1 << bits;
738 cfs_alloc_large(size * sizeof s->ls_hash[0])) == NULL;
739 --bits, size >>= 1) {
741 * Scale hash table down, until allocation succeeds.
746 s->ls_hash_size = size;
747 s->ls_hash_bits = bits;
748 s->ls_hash_mask = size - 1;
750 for (i = 0; i < size; i++)
751 INIT_HLIST_HEAD(&s->ls_hash[i]);
755 EXPORT_SYMBOL(lu_site_init);
758 * Finalize \a s and release its resources.
760 void lu_site_fini(struct lu_site *s)
762 LASSERT(list_empty(&s->ls_lru));
763 LASSERT(s->ls_total == 0);
765 down(&lu_sites_guard);
766 list_del_init(&s->ls_linkage);
769 if (s->ls_hash != NULL) {
771 for (i = 0; i < s->ls_hash_size; i++)
772 LASSERT(hlist_empty(&s->ls_hash[i]));
773 cfs_free_large(s->ls_hash);
776 if (s->ls_top_dev != NULL) {
777 s->ls_top_dev->ld_site = NULL;
778 lu_ref_del(&s->ls_top_dev->ld_reference, "site-top", s);
779 lu_device_put(s->ls_top_dev);
780 s->ls_top_dev = NULL;
783 EXPORT_SYMBOL(lu_site_fini);
786 * Called when initialization of stack for this site is completed.
788 int lu_site_init_finish(struct lu_site *s)
791 down(&lu_sites_guard);
792 result = lu_context_refill(&lu_shrink_env.le_ctx);
794 list_add(&s->ls_linkage, &lu_sites);
798 EXPORT_SYMBOL(lu_site_init_finish);
801 * Acquire additional reference on device \a d
803 void lu_device_get(struct lu_device *d)
805 atomic_inc(&d->ld_ref);
807 EXPORT_SYMBOL(lu_device_get);
810 * Release reference on device \a d.
812 void lu_device_put(struct lu_device *d)
814 LASSERT(atomic_read(&d->ld_ref) > 0);
815 atomic_dec(&d->ld_ref);
817 EXPORT_SYMBOL(lu_device_put);
820 * Initialize device \a d of type \a t.
822 int lu_device_init(struct lu_device *d, struct lu_device_type *t)
824 if (t->ldt_device_nr++ == 0 && t->ldt_ops->ldto_start != NULL)
825 t->ldt_ops->ldto_start(t);
826 memset(d, 0, sizeof *d);
827 atomic_set(&d->ld_ref, 0);
829 lu_ref_init(&d->ld_reference);
832 EXPORT_SYMBOL(lu_device_init);
835 * Finalize device \a d.
837 void lu_device_fini(struct lu_device *d)
839 struct lu_device_type *t;
842 if (d->ld_obd != NULL)
844 lprocfs_obd_cleanup(d->ld_obd);
846 lu_ref_fini(&d->ld_reference);
847 LASSERTF(atomic_read(&d->ld_ref) == 0,
848 "Refcount is %u\n", atomic_read(&d->ld_ref));
849 LASSERT(t->ldt_device_nr > 0);
850 if (--t->ldt_device_nr == 0 && t->ldt_ops->ldto_stop != NULL)
851 t->ldt_ops->ldto_stop(t);
853 EXPORT_SYMBOL(lu_device_fini);
856 * Initialize object \a o that is part of compound object \a h and was created
859 int lu_object_init(struct lu_object *o,
860 struct lu_object_header *h, struct lu_device *d)
862 memset(o, 0, sizeof *o);
866 o->lo_dev_ref = lu_ref_add(&d->ld_reference, "lu_object", o);
867 CFS_INIT_LIST_HEAD(&o->lo_linkage);
870 EXPORT_SYMBOL(lu_object_init);
873 * Finalize object and release its resources.
875 void lu_object_fini(struct lu_object *o)
877 struct lu_device *dev = o->lo_dev;
879 LASSERT(list_empty(&o->lo_linkage));
882 lu_ref_del_at(&dev->ld_reference,
883 o->lo_dev_ref , "lu_object", o);
888 EXPORT_SYMBOL(lu_object_fini);
891 * Add object \a o as first layer of compound object \a h
893 * This is typically called by the ->ldo_object_alloc() method of top-level
896 void lu_object_add_top(struct lu_object_header *h, struct lu_object *o)
898 list_move(&o->lo_linkage, &h->loh_layers);
900 EXPORT_SYMBOL(lu_object_add_top);
903 * Add object \a o as a layer of compound object, going after \a before.
905 * This is typically called by the ->ldo_object_alloc() method of \a
908 void lu_object_add(struct lu_object *before, struct lu_object *o)
910 list_move(&o->lo_linkage, &before->lo_linkage);
912 EXPORT_SYMBOL(lu_object_add);
915 * Initialize compound object.
917 int lu_object_header_init(struct lu_object_header *h)
919 memset(h, 0, sizeof *h);
920 atomic_set(&h->loh_ref, 1);
921 INIT_HLIST_NODE(&h->loh_hash);
922 CFS_INIT_LIST_HEAD(&h->loh_lru);
923 CFS_INIT_LIST_HEAD(&h->loh_layers);
924 lu_ref_init(&h->loh_reference);
927 EXPORT_SYMBOL(lu_object_header_init);
930 * Finalize compound object.
932 void lu_object_header_fini(struct lu_object_header *h)
934 LASSERT(list_empty(&h->loh_layers));
935 LASSERT(list_empty(&h->loh_lru));
936 LASSERT(hlist_unhashed(&h->loh_hash));
937 lu_ref_fini(&h->loh_reference);
939 EXPORT_SYMBOL(lu_object_header_fini);
942 * Given a compound object, find its slice, corresponding to the device type
945 struct lu_object *lu_object_locate(struct lu_object_header *h,
946 const struct lu_device_type *dtype)
950 list_for_each_entry(o, &h->loh_layers, lo_linkage) {
951 if (o->lo_dev->ld_type == dtype)
956 EXPORT_SYMBOL(lu_object_locate);
961 * Finalize and free devices in the device stack.
963 * Finalize device stack by purging object cache, and calling
964 * lu_device_type_operations::ldto_device_fini() and
965 * lu_device_type_operations::ldto_device_free() on all devices in the stack.
967 void lu_stack_fini(const struct lu_env *env, struct lu_device *top)
969 struct lu_site *site = top->ld_site;
970 struct lu_device *scan;
971 struct lu_device *next;
973 lu_site_purge(env, site, ~0);
974 for (scan = top; scan != NULL; scan = next) {
975 next = scan->ld_type->ldt_ops->ldto_device_fini(env, scan);
976 lu_ref_del(&scan->ld_reference, "lu-stack", &lu_site_init);
981 lu_site_purge(env, site, ~0);
983 if (!list_empty(&site->ls_lru) || site->ls_total != 0) {
985 * Uh-oh, objects still exist.
987 static DECLARE_LU_CDEBUG_PRINT_INFO(cookie, D_ERROR);
989 lu_site_print(env, site, &cookie, lu_cdebug_printer);
992 for (scan = top; scan != NULL; scan = next) {
993 const struct lu_device_type *ldt = scan->ld_type;
994 struct obd_type *type;
996 next = ldt->ldt_ops->ldto_device_free(env, scan);
997 type = ldt->ldt_obd_type;
1000 class_put_type(type);
1004 EXPORT_SYMBOL(lu_stack_fini);
1008 * Maximal number of tld slots.
1010 LU_CONTEXT_KEY_NR = 32
1013 static struct lu_context_key *lu_keys[LU_CONTEXT_KEY_NR] = { NULL, };
1015 static spinlock_t lu_keys_guard = SPIN_LOCK_UNLOCKED;
1018 * Global counter incremented whenever key is registered, unregistered,
1019 * revived or quiesced. This is used to void unnecessary calls to
1020 * lu_context_refill(). No locking is provided, as initialization and shutdown
1021 * are supposed to be externally serialized.
1023 static unsigned key_set_version = 0;
1028 int lu_context_key_register(struct lu_context_key *key)
1033 LASSERT(key->lct_init != NULL);
1034 LASSERT(key->lct_fini != NULL);
1035 LASSERT(key->lct_tags != 0);
1036 LASSERT(key->lct_owner != NULL);
1039 spin_lock(&lu_keys_guard);
1040 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1041 if (lu_keys[i] == NULL) {
1043 atomic_set(&key->lct_used, 1);
1045 lu_ref_init(&key->lct_reference);
1051 spin_unlock(&lu_keys_guard);
1054 EXPORT_SYMBOL(lu_context_key_register);
1056 static void key_fini(struct lu_context *ctx, int index)
1058 if (ctx->lc_value[index] != NULL) {
1059 struct lu_context_key *key;
1061 key = lu_keys[index];
1062 LASSERT(key != NULL);
1063 LASSERT(key->lct_fini != NULL);
1064 LASSERT(atomic_read(&key->lct_used) > 1);
1066 key->lct_fini(ctx, key, ctx->lc_value[index]);
1067 lu_ref_del(&key->lct_reference, "ctx", ctx);
1068 atomic_dec(&key->lct_used);
1069 LASSERT(key->lct_owner != NULL);
1070 if (!(ctx->lc_tags & LCT_NOREF)) {
1071 LASSERT(module_refcount(key->lct_owner) > 0);
1072 module_put(key->lct_owner);
1074 ctx->lc_value[index] = NULL;
1081 void lu_context_key_degister(struct lu_context_key *key)
1083 LASSERT(atomic_read(&key->lct_used) >= 1);
1084 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1087 key_fini(&lu_shrink_env.le_ctx, key->lct_index);
1089 if (atomic_read(&key->lct_used) > 1)
1090 CERROR("key has instances.\n");
1091 spin_lock(&lu_keys_guard);
1092 lu_keys[key->lct_index] = NULL;
1093 spin_unlock(&lu_keys_guard);
1095 EXPORT_SYMBOL(lu_context_key_degister);
1098 * Register a number of keys. This has to be called after all keys have been
1099 * initialized by a call to LU_CONTEXT_KEY_INIT().
1101 int lu_context_key_register_many(struct lu_context_key *k, ...)
1103 struct lu_context_key *key = k;
1109 result = lu_context_key_register(key);
1112 key = va_arg(args, struct lu_context_key *);
1113 } while (key != NULL);
1119 lu_context_key_degister(k);
1120 k = va_arg(args, struct lu_context_key *);
1127 EXPORT_SYMBOL(lu_context_key_register_many);
1130 * De-register a number of keys. This is a dual to
1131 * lu_context_key_register_many().
1133 void lu_context_key_degister_many(struct lu_context_key *k, ...)
1139 lu_context_key_degister(k);
1140 k = va_arg(args, struct lu_context_key*);
1141 } while (k != NULL);
1144 EXPORT_SYMBOL(lu_context_key_degister_many);
1147 * Revive a number of keys.
1149 void lu_context_key_revive_many(struct lu_context_key *k, ...)
1155 lu_context_key_revive(k);
1156 k = va_arg(args, struct lu_context_key*);
1157 } while (k != NULL);
1160 EXPORT_SYMBOL(lu_context_key_revive_many);
1163 * Quiescent a number of keys.
1165 void lu_context_key_quiesce_many(struct lu_context_key *k, ...)
1171 lu_context_key_quiesce(k);
1172 k = va_arg(args, struct lu_context_key*);
1173 } while (k != NULL);
1176 EXPORT_SYMBOL(lu_context_key_quiesce_many);
1179 * Return value associated with key \a key in context \a ctx.
1181 void *lu_context_key_get(const struct lu_context *ctx,
1182 const struct lu_context_key *key)
1184 LINVRNT(ctx->lc_state == LCS_ENTERED);
1185 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1186 return ctx->lc_value[key->lct_index];
1188 EXPORT_SYMBOL(lu_context_key_get);
1191 * List of remembered contexts. XXX document me.
1193 static CFS_LIST_HEAD(lu_context_remembered);
1196 * Destroy \a key in all remembered contexts. This is used to destroy key
1197 * values in "shared" contexts (like service threads), when a module owning
1198 * the key is about to be unloaded.
1200 void lu_context_key_quiesce(struct lu_context_key *key)
1202 struct lu_context *ctx;
1204 if (!(key->lct_tags & LCT_QUIESCENT)) {
1205 key->lct_tags |= LCT_QUIESCENT;
1207 * XXX memory barrier has to go here.
1209 spin_lock(&lu_keys_guard);
1210 list_for_each_entry(ctx, &lu_context_remembered, lc_remember)
1211 key_fini(ctx, key->lct_index);
1212 spin_unlock(&lu_keys_guard);
1216 EXPORT_SYMBOL(lu_context_key_quiesce);
1218 void lu_context_key_revive(struct lu_context_key *key)
1220 key->lct_tags &= ~LCT_QUIESCENT;
1223 EXPORT_SYMBOL(lu_context_key_revive);
1225 static void keys_fini(struct lu_context *ctx)
1229 if (ctx->lc_value != NULL) {
1230 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i)
1232 OBD_FREE(ctx->lc_value,
1233 ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1234 ctx->lc_value = NULL;
1238 static int keys_fill(struct lu_context *ctx)
1242 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1243 struct lu_context_key *key;
1246 if (ctx->lc_value[i] == NULL && key != NULL &&
1247 (key->lct_tags & ctx->lc_tags) &&
1249 * Don't create values for a LCT_QUIESCENT key, as this
1250 * will pin module owning a key.
1252 !(key->lct_tags & LCT_QUIESCENT)) {
1255 LINVRNT(key->lct_init != NULL);
1256 LINVRNT(key->lct_index == i);
1258 value = key->lct_init(ctx, key);
1259 if (unlikely(IS_ERR(value)))
1260 return PTR_ERR(value);
1261 LASSERT(key->lct_owner != NULL);
1262 if (!(ctx->lc_tags & LCT_NOREF))
1263 try_module_get(key->lct_owner);
1264 lu_ref_add_atomic(&key->lct_reference, "ctx", ctx);
1265 atomic_inc(&key->lct_used);
1267 * This is the only place in the code, where an
1268 * element of ctx->lc_value[] array is set to non-NULL
1271 ctx->lc_value[i] = value;
1272 if (key->lct_exit != NULL)
1273 ctx->lc_tags |= LCT_HAS_EXIT;
1275 ctx->lc_version = key_set_version;
1280 static int keys_init(struct lu_context *ctx)
1284 OBD_ALLOC(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1285 if (likely(ctx->lc_value != NULL))
1286 result = keys_fill(ctx);
1296 * Initialize context data-structure. Create values for all keys.
1298 int lu_context_init(struct lu_context *ctx, __u32 tags)
1300 memset(ctx, 0, sizeof *ctx);
1301 ctx->lc_state = LCS_INITIALIZED;
1302 ctx->lc_tags = tags;
1303 if (tags & LCT_REMEMBER) {
1304 spin_lock(&lu_keys_guard);
1305 list_add(&ctx->lc_remember, &lu_context_remembered);
1306 spin_unlock(&lu_keys_guard);
1308 CFS_INIT_LIST_HEAD(&ctx->lc_remember);
1309 return keys_init(ctx);
1311 EXPORT_SYMBOL(lu_context_init);
1314 * Finalize context data-structure. Destroy key values.
1316 void lu_context_fini(struct lu_context *ctx)
1318 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1319 ctx->lc_state = LCS_FINALIZED;
1321 spin_lock(&lu_keys_guard);
1322 list_del_init(&ctx->lc_remember);
1323 spin_unlock(&lu_keys_guard);
1325 EXPORT_SYMBOL(lu_context_fini);
1328 * Called before entering context.
1330 void lu_context_enter(struct lu_context *ctx)
1332 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1333 ctx->lc_state = LCS_ENTERED;
1335 EXPORT_SYMBOL(lu_context_enter);
1338 * Called after exiting from \a ctx
1340 void lu_context_exit(struct lu_context *ctx)
1344 LINVRNT(ctx->lc_state == LCS_ENTERED);
1345 ctx->lc_state = LCS_LEFT;
1346 if (ctx->lc_tags & LCT_HAS_EXIT && ctx->lc_value != NULL) {
1347 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1348 if (ctx->lc_value[i] != NULL) {
1349 struct lu_context_key *key;
1352 LASSERT(key != NULL);
1353 if (key->lct_exit != NULL)
1355 key, ctx->lc_value[i]);
1360 EXPORT_SYMBOL(lu_context_exit);
1363 * Allocate for context all missing keys that were registered after context
1366 int lu_context_refill(struct lu_context *ctx)
1368 LINVRNT(ctx->lc_value != NULL);
1369 return ctx->lc_version == key_set_version ? 0 : keys_fill(ctx);
1371 EXPORT_SYMBOL(lu_context_refill);
1373 static int lu_env_setup(struct lu_env *env, struct lu_context *ses,
1374 __u32 tags, int noref)
1378 LINVRNT(ergo(!noref, !(tags & LCT_NOREF)));
1381 result = lu_context_init(&env->le_ctx, tags);
1382 if (likely(result == 0))
1383 lu_context_enter(&env->le_ctx);
1387 static int lu_env_init_noref(struct lu_env *env, struct lu_context *ses,
1390 return lu_env_setup(env, ses, tags, 1);
1393 int lu_env_init(struct lu_env *env, struct lu_context *ses, __u32 tags)
1395 return lu_env_setup(env, ses, tags, 0);
1397 EXPORT_SYMBOL(lu_env_init);
1399 void lu_env_fini(struct lu_env *env)
1401 lu_context_exit(&env->le_ctx);
1402 lu_context_fini(&env->le_ctx);
1405 EXPORT_SYMBOL(lu_env_fini);
1407 int lu_env_refill(struct lu_env *env)
1411 result = lu_context_refill(&env->le_ctx);
1412 if (result == 0 && env->le_ses != NULL)
1413 result = lu_context_refill(env->le_ses);
1416 EXPORT_SYMBOL(lu_env_refill);
1418 static struct shrinker *lu_site_shrinker = NULL;
1421 static int lu_cache_shrink(int nr, unsigned int gfp_mask)
1424 struct lu_site *tmp;
1427 CFS_LIST_HEAD(splice);
1429 if (nr != 0 && !(gfp_mask & __GFP_FS))
1432 down(&lu_sites_guard);
1433 list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
1435 remain = lu_site_purge(&lu_shrink_env, s, remain);
1437 * Move just shrunk site to the tail of site list to
1438 * assure shrinking fairness.
1440 list_move_tail(&s->ls_linkage, &splice);
1442 read_lock(&s->ls_guard);
1443 cached += s->ls_total - s->ls_busy;
1444 read_unlock(&s->ls_guard);
1448 list_splice(&splice, lu_sites.prev);
1449 up(&lu_sites_guard);
1453 #else /* !__KERNEL__ */
1454 static int lu_cache_shrink(int nr, unsigned int gfp_mask)
1458 #endif /* __KERNEL__ */
1460 int lu_ref_global_init(void);
1461 void lu_ref_global_fini(void);
1464 * Initialization of global lu_* data.
1466 int lu_global_init(void)
1470 CDEBUG(D_CONSOLE, "Lustre LU module (%p).\n", &lu_keys);
1472 LU_CONTEXT_KEY_INIT(&lu_global_key);
1473 result = lu_context_key_register(&lu_global_key);
1477 * At this level, we don't know what tags are needed, so allocate them
1478 * conservatively. This should not be too bad, because this
1479 * environment is global.
1481 down(&lu_sites_guard);
1482 result = lu_env_init_noref(&lu_shrink_env, NULL, LCT_SHRINKER);
1483 up(&lu_sites_guard);
1487 result = lu_ref_global_init();
1491 * seeks estimation: 3 seeks to read a record from oi, one to read
1492 * inode, one for ea. Unfortunately setting this high value results in
1493 * lu_object/inode cache consuming all the memory.
1495 lu_site_shrinker = set_shrinker(DEFAULT_SEEKS, lu_cache_shrink);
1496 if (lu_site_shrinker == NULL)
1499 result = lu_time_global_init();
1504 * Dual to lu_global_init().
1506 void lu_global_fini(void)
1508 lu_time_global_fini();
1509 if (lu_site_shrinker != NULL) {
1510 remove_shrinker(lu_site_shrinker);
1511 lu_site_shrinker = NULL;
1514 lu_context_key_degister(&lu_global_key);
1517 * Tear shrinker environment down _after_ de-registering
1518 * lu_global_key, because the latter has a value in the former.
1520 down(&lu_sites_guard);
1521 lu_env_fini(&lu_shrink_env);
1522 up(&lu_sites_guard);
1524 lu_ref_global_fini();
1527 struct lu_buf LU_BUF_NULL = {
1531 EXPORT_SYMBOL(LU_BUF_NULL);
1534 * XXX: Functions below logically belong to fid module, but they are used by
1535 * dt_store_open(). Put them here until better place is found.
1538 void fid_pack(struct lu_fid_pack *pack, const struct lu_fid *fid,
1539 struct lu_fid *befider)
1549 * Two cases: compact 6 bytes representation for a common case, and
1550 * full 17 byte representation for "unusual" fid.
1554 * Check that usual case is really usual.
1556 CLASSERT(LUSTRE_SEQ_MAX_WIDTH < 0xffffull);
1558 if (fid_is_igif(fid) ||
1559 seq > 0xffffffull || oid > 0xffff || fid_ver(fid) != 0) {
1560 fid_cpu_to_be(befider, fid);
1561 recsize = sizeof *befider;
1563 unsigned char *small_befider;
1565 small_befider = (char *)befider;
1567 small_befider[0] = seq >> 16;
1568 small_befider[1] = seq >> 8;
1569 small_befider[2] = seq;
1571 small_befider[3] = oid >> 8;
1572 small_befider[4] = oid;
1576 memcpy(pack->fp_area, befider, recsize);
1577 pack->fp_len = recsize + 1;
1579 EXPORT_SYMBOL(fid_pack);
1581 int fid_unpack(const struct lu_fid_pack *pack, struct lu_fid *fid)
1586 switch (pack->fp_len) {
1587 case sizeof *fid + 1:
1588 memcpy(fid, pack->fp_area, sizeof *fid);
1589 fid_be_to_cpu(fid, fid);
1592 const unsigned char *area;
1594 area = pack->fp_area;
1595 fid->f_seq = (area[0] << 16) | (area[1] << 8) | area[2];
1596 fid->f_oid = (area[3] << 8) | area[4];
1601 CERROR("Unexpected packed fid size: %d\n", pack->fp_len);
1606 EXPORT_SYMBOL(fid_unpack);
1608 const char *lu_time_names[LU_TIME_NR] = {
1609 [LU_TIME_FIND_LOOKUP] = "find_lookup",
1610 [LU_TIME_FIND_ALLOC] = "find_alloc",
1611 [LU_TIME_FIND_INSERT] = "find_insert"
1613 EXPORT_SYMBOL(lu_time_names);
1616 * Helper function to initialize a number of kmem slab caches at once.
1618 int lu_kmem_init(struct lu_kmem_descr *caches)
1622 for (result = 0; caches->ckd_cache != NULL; ++caches) {
1623 *caches->ckd_cache = cfs_mem_cache_create(caches->ckd_name,
1626 if (*caches->ckd_cache == NULL) {
1633 EXPORT_SYMBOL(lu_kmem_init);
1636 * Helper function to finalize a number of kmem slab cached at once. Dual to
1639 void lu_kmem_fini(struct lu_kmem_descr *caches)
1643 for (; caches->ckd_cache != NULL; ++caches) {
1644 if (*caches->ckd_cache != NULL) {
1645 rc = cfs_mem_cache_destroy(*caches->ckd_cache);
1646 LASSERTF(rc == 0, "couldn't destroy %s slab\n",
1648 *caches->ckd_cache = NULL;
1652 EXPORT_SYMBOL(lu_kmem_fini);