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)
426 struct lu_object_header *h;
427 struct hlist_node *scan;
429 hlist_for_each_entry(h, scan, bucket, loh_hash) {
430 s->ls_stats.s_cache_check ++;
431 if (likely(lu_fid_eq(&h->loh_fid, f) &&
432 !lu_object_is_dying(h))) {
433 /* bump reference count... */
434 if (atomic_add_return(1, &h->loh_ref) == 1)
436 /* and move to the head of the LRU */
438 * XXX temporary disable this to measure effects of
439 * read-write locking.
441 /* list_move_tail(&h->loh_lru, &s->ls_lru); */
442 s->ls_stats.s_cache_hit ++;
443 return lu_object_top(h);
446 s->ls_stats.s_cache_miss ++;
450 static __u32 fid_hash(const struct lu_fid *f, int bits)
452 /* all objects with same id and different versions will belong to same
453 * collisions list. */
454 return hash_long(fid_flatten(f), bits);
458 * Search cache for an object with the fid @f. If such object is found, return
459 * it. Otherwise, create new object, insert it into cache and return it. In
460 * any case, additional reference is acquired on the returned object.
462 struct lu_object *lu_object_find(const struct lu_env *env,
463 struct lu_site *s, const struct lu_fid *f)
466 struct lu_object *shadow;
467 struct hlist_head *bucket;
470 * This uses standard index maintenance protocol:
472 * - search index under lock, and return object if found;
473 * - otherwise, unlock index, allocate new object;
474 * - lock index and search again;
475 * - if nothing is found (usual case), insert newly created
477 * - otherwise (race: other thread inserted object), free
478 * object just allocated.
483 bucket = s->ls_hash + fid_hash(f, s->ls_hash_bits);
485 read_lock(&s->ls_guard);
486 o = htable_lookup(s, bucket, f);
487 read_unlock(&s->ls_guard);
493 * Allocate new object. This may result in rather complicated
494 * operations, including fld queries, inode loading, etc.
496 o = lu_object_alloc(env, s, f);
497 if (unlikely(IS_ERR(o)))
500 LASSERT(lu_fid_eq(lu_object_fid(o), f));
502 write_lock(&s->ls_guard);
503 shadow = htable_lookup(s, bucket, f);
504 if (likely(shadow == NULL)) {
505 hlist_add_head(&o->lo_header->loh_hash, bucket);
506 list_add_tail(&o->lo_header->loh_lru, &s->ls_lru);
512 s->ls_stats.s_cache_race ++;
513 write_unlock(&s->ls_guard);
515 lu_object_free(env, o);
518 EXPORT_SYMBOL(lu_object_find);
522 * Global list of all device types.
524 static CFS_LIST_HEAD(lu_device_types);
526 int lu_device_type_init(struct lu_device_type *ldt)
530 CFS_INIT_LIST_HEAD(&ldt->ldt_linkage);
531 result = ldt->ldt_ops->ldto_init(ldt);
533 list_add(&ldt->ldt_linkage, &lu_device_types);
536 EXPORT_SYMBOL(lu_device_type_init);
538 void lu_device_type_fini(struct lu_device_type *ldt)
540 list_del_init(&ldt->ldt_linkage);
541 ldt->ldt_ops->ldto_fini(ldt);
543 EXPORT_SYMBOL(lu_device_type_fini);
545 void lu_types_stop(void)
547 struct lu_device_type *ldt;
549 list_for_each_entry(ldt, &lu_device_types, ldt_linkage) {
550 if (ldt->ldt_device_nr == 0)
551 ldt->ldt_ops->ldto_stop(ldt);
554 EXPORT_SYMBOL(lu_types_stop);
557 * Global list of all sites on this node
559 static CFS_LIST_HEAD(lu_sites);
560 static DECLARE_MUTEX(lu_sites_guard);
563 * Global environment used by site shrinker.
565 static struct lu_env lu_shrink_env;
568 * Print all objects in \a s.
570 void lu_site_print(const struct lu_env *env, struct lu_site *s, void *cookie,
571 lu_printer_t printer)
575 for (i = 0; i < s->ls_hash_size; ++i) {
576 struct lu_object_header *h;
577 struct hlist_node *scan;
579 read_lock(&s->ls_guard);
580 hlist_for_each_entry(h, scan, &s->ls_hash[i], loh_hash) {
582 if (!list_empty(&h->loh_layers)) {
583 const struct lu_object *obj;
585 obj = lu_object_top(h);
586 lu_object_print(env, cookie, printer, obj);
588 lu_object_header_print(env, cookie, printer, h);
590 read_unlock(&s->ls_guard);
593 EXPORT_SYMBOL(lu_site_print);
596 LU_CACHE_PERCENT = 20,
600 * Return desired hash table order.
602 static int lu_htable_order(void)
604 unsigned long cache_size;
608 * Calculate hash table size, assuming that we want reasonable
609 * performance when 20% of total memory is occupied by cache of
612 * Size of lu_object is (arbitrary) taken as 1K (together with inode).
614 cache_size = num_physpages;
616 #if BITS_PER_LONG == 32
617 /* limit hashtable size for lowmem systems to low RAM */
618 if (cache_size > 1 << (30 - CFS_PAGE_SHIFT))
619 cache_size = 1 << (30 - CFS_PAGE_SHIFT) * 3 / 4;
622 cache_size = cache_size / 100 * LU_CACHE_PERCENT *
623 (CFS_PAGE_SIZE / 1024);
625 for (bits = 1; (1 << bits) < cache_size; ++bits) {
632 * Initialize site @s, with @d as the top level device.
634 int lu_site_init(struct lu_site *s, struct lu_device *top)
641 memset(s, 0, sizeof *s);
642 rwlock_init(&s->ls_guard);
643 CFS_INIT_LIST_HEAD(&s->ls_lru);
644 CFS_INIT_LIST_HEAD(&s->ls_linkage);
649 for (bits = lu_htable_order(), size = 1 << bits;
651 cfs_alloc_large(size * sizeof s->ls_hash[0])) == NULL;
652 --bits, size >>= 1) {
654 * Scale hash table down, until allocation succeeds.
659 s->ls_hash_size = size;
660 s->ls_hash_bits = bits;
661 s->ls_hash_mask = size - 1;
663 for (i = 0; i < size; i++)
664 INIT_HLIST_HEAD(&s->ls_hash[i]);
668 EXPORT_SYMBOL(lu_site_init);
671 * Finalize @s and release its resources.
673 void lu_site_fini(struct lu_site *s)
675 LASSERT(list_empty(&s->ls_lru));
676 LASSERT(s->ls_total == 0);
678 down(&lu_sites_guard);
679 list_del_init(&s->ls_linkage);
682 if (s->ls_hash != NULL) {
684 for (i = 0; i < s->ls_hash_size; i++)
685 LASSERT(hlist_empty(&s->ls_hash[i]));
686 cfs_free_large(s->ls_hash);
689 if (s->ls_top_dev != NULL) {
690 s->ls_top_dev->ld_site = NULL;
691 lu_device_put(s->ls_top_dev);
692 s->ls_top_dev = NULL;
695 EXPORT_SYMBOL(lu_site_fini);
698 * Called when initialization of stack for this site is completed.
700 int lu_site_init_finish(struct lu_site *s)
703 down(&lu_sites_guard);
704 result = lu_context_refill(&lu_shrink_env.le_ctx);
706 list_add(&s->ls_linkage, &lu_sites);
710 EXPORT_SYMBOL(lu_site_init_finish);
713 * Acquire additional reference on device \a d
715 void lu_device_get(struct lu_device *d)
717 atomic_inc(&d->ld_ref);
719 EXPORT_SYMBOL(lu_device_get);
722 * Release reference on device \a d.
724 void lu_device_put(struct lu_device *d)
726 LASSERT(atomic_read(&d->ld_ref) > 0);
727 atomic_dec(&d->ld_ref);
729 EXPORT_SYMBOL(lu_device_put);
732 * Initialize device \a d of type \a t.
734 int lu_device_init(struct lu_device *d, struct lu_device_type *t)
736 if (t->ldt_device_nr++ == 0 && t->ldt_ops->ldto_start != NULL)
737 t->ldt_ops->ldto_start(t);
738 memset(d, 0, sizeof *d);
739 atomic_set(&d->ld_ref, 0);
741 lu_ref_init(&d->ld_reference);
744 EXPORT_SYMBOL(lu_device_init);
747 * Finalize device \a d.
749 void lu_device_fini(struct lu_device *d)
751 struct lu_device_type *t;
754 if (d->ld_obd != NULL)
756 lprocfs_obd_cleanup(d->ld_obd);
758 lu_ref_fini(&d->ld_reference);
759 LASSERTF(atomic_read(&d->ld_ref) == 0,
760 "Refcount is %u\n", atomic_read(&d->ld_ref));
761 LASSERT(t->ldt_device_nr > 0);
762 if (--t->ldt_device_nr == 0 && t->ldt_ops->ldto_stop != NULL)
763 t->ldt_ops->ldto_stop(t);
765 EXPORT_SYMBOL(lu_device_fini);
768 * Initialize object \a o that is part of compound object \a h and was created
771 int lu_object_init(struct lu_object *o,
772 struct lu_object_header *h, struct lu_device *d)
774 memset(o, 0, sizeof *o);
778 CFS_INIT_LIST_HEAD(&o->lo_linkage);
781 EXPORT_SYMBOL(lu_object_init);
784 * Finalize object and release its resources.
786 void lu_object_fini(struct lu_object *o)
788 LASSERT(list_empty(&o->lo_linkage));
790 if (o->lo_dev != NULL) {
791 lu_device_put(o->lo_dev);
795 EXPORT_SYMBOL(lu_object_fini);
798 * Add object @o as first layer of compound object @h
800 * This is typically called by the ->ldo_object_alloc() method of top-level
803 void lu_object_add_top(struct lu_object_header *h, struct lu_object *o)
805 list_move(&o->lo_linkage, &h->loh_layers);
807 EXPORT_SYMBOL(lu_object_add_top);
810 * Add object \a o as a layer of compound object, going after \a before.
812 * This is typically called by the ->ldo_object_alloc() method of \a
815 void lu_object_add(struct lu_object *before, struct lu_object *o)
817 list_move(&o->lo_linkage, &before->lo_linkage);
819 EXPORT_SYMBOL(lu_object_add);
822 * Initialize compound object.
824 int lu_object_header_init(struct lu_object_header *h)
826 memset(h, 0, sizeof *h);
827 atomic_set(&h->loh_ref, 1);
828 INIT_HLIST_NODE(&h->loh_hash);
829 CFS_INIT_LIST_HEAD(&h->loh_lru);
830 CFS_INIT_LIST_HEAD(&h->loh_layers);
833 EXPORT_SYMBOL(lu_object_header_init);
836 * Finalize compound object.
838 void lu_object_header_fini(struct lu_object_header *h)
840 LASSERT(list_empty(&h->loh_layers));
841 LASSERT(list_empty(&h->loh_lru));
842 LASSERT(hlist_unhashed(&h->loh_hash));
844 EXPORT_SYMBOL(lu_object_header_fini);
847 * Given a compound object, find its slice, corresponding to the device type
850 struct lu_object *lu_object_locate(struct lu_object_header *h,
851 struct lu_device_type *dtype)
855 list_for_each_entry(o, &h->loh_layers, lo_linkage) {
856 if (o->lo_dev->ld_type == dtype)
861 EXPORT_SYMBOL(lu_object_locate);
866 * Finalize and free devices in the device stack.
868 * Finalize device stack by purging object cache, and calling
869 * lu_device_type_operations::ldto_device_fini() and
870 * lu_device_type_operations::ldto_device_free() on all devices in the stack.
872 void lu_stack_fini(const struct lu_env *env, struct lu_device *top)
874 struct lu_site *site = top->ld_site;
875 struct lu_device *scan;
876 struct lu_device *next;
878 lu_site_purge(env, site, ~0);
879 for (scan = top; scan != NULL; scan = next) {
880 next = scan->ld_type->ldt_ops->ldto_device_fini(env, scan);
885 lu_site_purge(env, site, ~0);
887 if (!list_empty(&site->ls_lru) || site->ls_total != 0) {
889 * Uh-oh, objects still exist.
891 static DECLARE_LU_CDEBUG_PRINT_INFO(cookie, D_ERROR);
893 lu_site_print(env, site, &cookie, lu_cdebug_printer);
896 for (scan = top; scan != NULL; scan = next) {
897 const struct lu_device_type *ldt = scan->ld_type;
898 struct obd_type *type;
900 next = ldt->ldt_ops->ldto_device_free(env, scan);
901 type = ldt->ldt_obd_type;
904 class_put_type(type);
908 EXPORT_SYMBOL(lu_stack_fini);
912 * Maximal number of tld slots.
914 LU_CONTEXT_KEY_NR = 32
917 static struct lu_context_key *lu_keys[LU_CONTEXT_KEY_NR] = { NULL, };
919 static spinlock_t lu_keys_guard = SPIN_LOCK_UNLOCKED;
922 * Global counter incremented whenever key is registered, unregistered,
923 * revived or quiesced. This is used to void unnecessary calls to
924 * lu_context_refill(). No locking is provided, as initialization and shutdown
925 * are supposed to be externally serialized.
927 static unsigned key_set_version = 0;
932 int lu_context_key_register(struct lu_context_key *key)
937 LASSERT(key->lct_init != NULL);
938 LASSERT(key->lct_fini != NULL);
939 LASSERT(key->lct_tags != 0);
940 LASSERT(key->lct_owner != NULL);
943 spin_lock(&lu_keys_guard);
944 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
945 if (lu_keys[i] == NULL) {
947 atomic_set(&key->lct_used, 1);
949 lu_ref_init(&key->lct_reference);
955 spin_unlock(&lu_keys_guard);
958 EXPORT_SYMBOL(lu_context_key_register);
960 static void key_fini(struct lu_context *ctx, int index)
962 if (ctx->lc_value[index] != NULL) {
963 struct lu_context_key *key;
965 key = lu_keys[index];
966 LASSERT(key != NULL);
967 LASSERT(key->lct_fini != NULL);
968 LASSERT(atomic_read(&key->lct_used) > 1);
970 key->lct_fini(ctx, key, ctx->lc_value[index]);
971 lu_ref_del(&key->lct_reference, "ctx", ctx);
972 atomic_dec(&key->lct_used);
973 LASSERT(key->lct_owner != NULL);
974 if (!(ctx->lc_tags & LCT_NOREF)) {
975 LASSERT(module_refcount(key->lct_owner) > 0);
976 module_put(key->lct_owner);
978 ctx->lc_value[index] = NULL;
985 void lu_context_key_degister(struct lu_context_key *key)
987 LASSERT(atomic_read(&key->lct_used) >= 1);
988 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
991 key_fini(&lu_shrink_env.le_ctx, key->lct_index);
993 if (atomic_read(&key->lct_used) > 1)
994 CERROR("key has instances.\n");
995 spin_lock(&lu_keys_guard);
996 lu_keys[key->lct_index] = NULL;
997 spin_unlock(&lu_keys_guard);
999 EXPORT_SYMBOL(lu_context_key_degister);
1002 * Register a number of keys. This has to be called after all keys have been
1003 * initialized by a call to LU_CONTEXT_KEY_INIT().
1005 int lu_context_key_register_many(struct lu_context_key *k, ...)
1007 struct lu_context_key *key = k;
1013 result = lu_context_key_register(key);
1016 key = va_arg(args, struct lu_context_key *);
1017 } while (key != NULL);
1023 lu_context_key_degister(k);
1024 k = va_arg(args, struct lu_context_key *);
1031 EXPORT_SYMBOL(lu_context_key_register_many);
1034 * De-register a number of keys. This is a dual to
1035 * lu_context_key_register_many().
1037 void lu_context_key_degister_many(struct lu_context_key *k, ...)
1043 lu_context_key_degister(k);
1044 k = va_arg(args, struct lu_context_key*);
1045 } while (k != NULL);
1048 EXPORT_SYMBOL(lu_context_key_degister_many);
1051 * Revive a number of keys.
1053 void lu_context_key_revive_many(struct lu_context_key *k, ...)
1059 lu_context_key_revive(k);
1060 k = va_arg(args, struct lu_context_key*);
1061 } while (k != NULL);
1064 EXPORT_SYMBOL(lu_context_key_revive_many);
1067 * Quiescent a number of keys.
1069 void lu_context_key_quiesce_many(struct lu_context_key *k, ...)
1075 lu_context_key_quiesce(k);
1076 k = va_arg(args, struct lu_context_key*);
1077 } while (k != NULL);
1080 EXPORT_SYMBOL(lu_context_key_quiesce_many);
1083 * Return value associated with key \a key in context \a ctx.
1085 void *lu_context_key_get(const struct lu_context *ctx,
1086 const struct lu_context_key *key)
1088 LINVRNT(ctx->lc_state == LCS_ENTERED);
1089 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1090 return ctx->lc_value[key->lct_index];
1092 EXPORT_SYMBOL(lu_context_key_get);
1095 * List of remembered contexts. XXX document me.
1097 static CFS_LIST_HEAD(lu_context_remembered);
1100 * Destroy \a key in all remembered contexts. This is used to destroy key
1101 * values in "shared" contexts (like service threads), when a module owning
1102 * the key is about to be unloaded.
1104 void lu_context_key_quiesce(struct lu_context_key *key)
1106 struct lu_context *ctx;
1108 if (!(key->lct_tags & LCT_QUIESCENT)) {
1109 key->lct_tags |= LCT_QUIESCENT;
1111 * XXX memory barrier has to go here.
1113 spin_lock(&lu_keys_guard);
1114 list_for_each_entry(ctx, &lu_context_remembered, lc_remember)
1115 key_fini(ctx, key->lct_index);
1116 spin_unlock(&lu_keys_guard);
1120 EXPORT_SYMBOL(lu_context_key_quiesce);
1122 void lu_context_key_revive(struct lu_context_key *key)
1124 key->lct_tags &= ~LCT_QUIESCENT;
1127 EXPORT_SYMBOL(lu_context_key_revive);
1129 static void keys_fini(struct lu_context *ctx)
1133 if (ctx->lc_value != NULL) {
1134 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i)
1136 OBD_FREE(ctx->lc_value,
1137 ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1138 ctx->lc_value = NULL;
1142 static int keys_fill(struct lu_context *ctx)
1146 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1147 struct lu_context_key *key;
1150 if (ctx->lc_value[i] == NULL && key != NULL &&
1151 (key->lct_tags & ctx->lc_tags) &&
1153 * Don't create values for a LCT_QUIESCENT key, as this
1154 * will pin module owning a key.
1156 !(key->lct_tags & LCT_QUIESCENT)) {
1159 LINVRNT(key->lct_init != NULL);
1160 LINVRNT(key->lct_index == i);
1162 value = key->lct_init(ctx, key);
1163 if (unlikely(IS_ERR(value)))
1164 return PTR_ERR(value);
1165 LASSERT(key->lct_owner != NULL);
1166 if (!(ctx->lc_tags & LCT_NOREF))
1167 try_module_get(key->lct_owner);
1168 lu_ref_add_atomic(&key->lct_reference, "ctx", ctx);
1169 atomic_inc(&key->lct_used);
1171 * This is the only place in the code, where an
1172 * element of ctx->lc_value[] array is set to non-NULL
1175 ctx->lc_value[i] = value;
1176 if (key->lct_exit != NULL)
1177 ctx->lc_tags |= LCT_HAS_EXIT;
1179 ctx->lc_version = key_set_version;
1184 static int keys_init(struct lu_context *ctx)
1188 OBD_ALLOC(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1189 if (likely(ctx->lc_value != NULL))
1190 result = keys_fill(ctx);
1200 * Initialize context data-structure. Create values for all keys.
1202 int lu_context_init(struct lu_context *ctx, __u32 tags)
1204 memset(ctx, 0, sizeof *ctx);
1205 ctx->lc_state = LCS_INITIALIZED;
1206 ctx->lc_tags = tags;
1207 if (tags & LCT_REMEMBER) {
1208 spin_lock(&lu_keys_guard);
1209 list_add(&ctx->lc_remember, &lu_context_remembered);
1210 spin_unlock(&lu_keys_guard);
1212 CFS_INIT_LIST_HEAD(&ctx->lc_remember);
1213 return keys_init(ctx);
1215 EXPORT_SYMBOL(lu_context_init);
1218 * Finalize context data-structure. Destroy key values.
1220 void lu_context_fini(struct lu_context *ctx)
1222 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1223 ctx->lc_state = LCS_FINALIZED;
1225 spin_lock(&lu_keys_guard);
1226 list_del_init(&ctx->lc_remember);
1227 spin_unlock(&lu_keys_guard);
1229 EXPORT_SYMBOL(lu_context_fini);
1232 * Called before entering context.
1234 void lu_context_enter(struct lu_context *ctx)
1236 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1237 ctx->lc_state = LCS_ENTERED;
1239 EXPORT_SYMBOL(lu_context_enter);
1242 * Called after exiting from \a ctx
1244 void lu_context_exit(struct lu_context *ctx)
1248 LINVRNT(ctx->lc_state == LCS_ENTERED);
1249 ctx->lc_state = LCS_LEFT;
1250 if (ctx->lc_tags & LCT_HAS_EXIT && ctx->lc_value != NULL) {
1251 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1252 if (ctx->lc_value[i] != NULL) {
1253 struct lu_context_key *key;
1256 LASSERT(key != NULL);
1257 if (key->lct_exit != NULL)
1259 key, ctx->lc_value[i]);
1264 EXPORT_SYMBOL(lu_context_exit);
1267 * Allocate for context all missing keys that were registered after context
1270 int lu_context_refill(struct lu_context *ctx)
1272 LINVRNT(ctx->lc_value != NULL);
1273 return ctx->lc_version == key_set_version ? 0 : keys_fill(ctx);
1275 EXPORT_SYMBOL(lu_context_refill);
1277 static int lu_env_setup(struct lu_env *env, struct lu_context *ses,
1278 __u32 tags, int noref)
1282 LINVRNT(ergo(!noref, !(tags & LCT_NOREF)));
1285 result = lu_context_init(&env->le_ctx, tags);
1286 if (likely(result == 0))
1287 lu_context_enter(&env->le_ctx);
1291 static int lu_env_init_noref(struct lu_env *env, struct lu_context *ses,
1294 return lu_env_setup(env, ses, tags, 1);
1297 int lu_env_init(struct lu_env *env, struct lu_context *ses, __u32 tags)
1299 return lu_env_setup(env, ses, tags, 0);
1301 EXPORT_SYMBOL(lu_env_init);
1303 void lu_env_fini(struct lu_env *env)
1305 lu_context_exit(&env->le_ctx);
1306 lu_context_fini(&env->le_ctx);
1309 EXPORT_SYMBOL(lu_env_fini);
1311 int lu_env_refill(struct lu_env *env)
1315 result = lu_context_refill(&env->le_ctx);
1316 if (result == 0 && env->le_ses != NULL)
1317 result = lu_context_refill(env->le_ses);
1320 EXPORT_SYMBOL(lu_env_refill);
1322 static struct shrinker *lu_site_shrinker = NULL;
1325 static int lu_cache_shrink(int nr, unsigned int gfp_mask)
1328 struct lu_site *tmp;
1331 CFS_LIST_HEAD(splice);
1333 if (nr != 0 && !(gfp_mask & __GFP_FS))
1336 down(&lu_sites_guard);
1337 list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
1339 remain = lu_site_purge(&lu_shrink_env, s, remain);
1341 * Move just shrunk site to the tail of site list to
1342 * assure shrinking fairness.
1344 list_move_tail(&s->ls_linkage, &splice);
1346 read_lock(&s->ls_guard);
1347 cached += s->ls_total - s->ls_busy;
1348 read_unlock(&s->ls_guard);
1352 list_splice(&splice, lu_sites.prev);
1353 up(&lu_sites_guard);
1357 #else /* !__KERNEL__ */
1358 static int lu_cache_shrink(int nr, unsigned int gfp_mask)
1362 #endif /* __KERNEL__ */
1364 int lu_ref_global_init(void);
1365 void lu_ref_global_fini(void);
1368 * Initialization of global lu_* data.
1370 int lu_global_init(void)
1374 CDEBUG(D_CONSOLE, "Lustre LU module (%p).\n", &lu_keys);
1376 LU_CONTEXT_KEY_INIT(&lu_global_key);
1377 result = lu_context_key_register(&lu_global_key);
1381 * At this level, we don't know what tags are needed, so allocate them
1382 * conservatively. This should not be too bad, because this
1383 * environment is global.
1385 down(&lu_sites_guard);
1386 result = lu_env_init_noref(&lu_shrink_env, NULL, LCT_SHRINKER);
1387 up(&lu_sites_guard);
1391 result = lu_ref_global_init();
1395 * seeks estimation: 3 seeks to read a record from oi, one to read
1396 * inode, one for ea. Unfortunately setting this high value results in
1397 * lu_object/inode cache consuming all the memory.
1399 lu_site_shrinker = set_shrinker(DEFAULT_SEEKS, lu_cache_shrink);
1400 if (lu_site_shrinker == NULL)
1403 result = lu_time_global_init();
1408 * Dual to lu_global_init().
1410 void lu_global_fini(void)
1412 lu_time_global_fini();
1413 if (lu_site_shrinker != NULL) {
1414 remove_shrinker(lu_site_shrinker);
1415 lu_site_shrinker = NULL;
1418 lu_context_key_degister(&lu_global_key);
1421 * Tear shrinker environment down _after_ de-registering
1422 * lu_global_key, because the latter has a value in the former.
1424 down(&lu_sites_guard);
1425 lu_env_fini(&lu_shrink_env);
1426 up(&lu_sites_guard);
1428 lu_ref_global_fini();
1431 struct lu_buf LU_BUF_NULL = {
1435 EXPORT_SYMBOL(LU_BUF_NULL);
1438 * XXX: Functions below logically belong to fid module, but they are used by
1439 * dt_store_open(). Put them here until better place is found.
1442 void fid_pack(struct lu_fid_pack *pack, const struct lu_fid *fid,
1443 struct lu_fid *befider)
1453 * Two cases: compact 6 bytes representation for a common case, and
1454 * full 17 byte representation for "unusual" fid.
1458 * Check that usual case is really usual.
1460 CLASSERT(LUSTRE_SEQ_MAX_WIDTH < 0xffffull);
1462 if (fid_is_igif(fid) ||
1463 seq > 0xffffffull || oid > 0xffff || fid_ver(fid) != 0) {
1464 fid_cpu_to_be(befider, fid);
1465 recsize = sizeof *befider;
1467 unsigned char *small_befider;
1469 small_befider = (char *)befider;
1471 small_befider[0] = seq >> 16;
1472 small_befider[1] = seq >> 8;
1473 small_befider[2] = seq;
1475 small_befider[3] = oid >> 8;
1476 small_befider[4] = oid;
1480 memcpy(pack->fp_area, befider, recsize);
1481 pack->fp_len = recsize + 1;
1483 EXPORT_SYMBOL(fid_pack);
1485 int fid_unpack(const struct lu_fid_pack *pack, struct lu_fid *fid)
1490 switch (pack->fp_len) {
1491 case sizeof *fid + 1:
1492 memcpy(fid, pack->fp_area, sizeof *fid);
1493 fid_be_to_cpu(fid, fid);
1496 const unsigned char *area;
1498 area = pack->fp_area;
1499 fid->f_seq = (area[0] << 16) | (area[1] << 8) | area[2];
1500 fid->f_oid = (area[3] << 8) | area[4];
1505 CERROR("Unexpected packed fid size: %d\n", pack->fp_len);
1510 EXPORT_SYMBOL(fid_unpack);
1512 const char *lu_time_names[LU_TIME_NR] = {
1513 [LU_TIME_FIND_LOOKUP] = "find_lookup",
1514 [LU_TIME_FIND_ALLOC] = "find_alloc",
1515 [LU_TIME_FIND_INSERT] = "find_insert"
1517 EXPORT_SYMBOL(lu_time_names);