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 (c) 2007, 2010, Oracle and/or its affiliates. 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 cfs_write_lock(&site->ls_guard);
86 if (cfs_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 cfs_list_for_each_entry_reverse(o, &top->loh_layers,
93 if (o->lo_ops->loo_object_release != NULL)
94 o->lo_ops->loo_object_release(env, o);
97 if (lu_object_is_dying(top)) {
99 * If object is dying (will not be cached), removed it
100 * from hash table and LRU.
102 * This is done with hash table and LRU lists
103 * locked. As the only way to acquire first reference
104 * to previously unreferenced object is through
105 * hash-table lookup (lu_object_find()), or LRU
106 * scanning (lu_site_purge()), that are done under
107 * hash-table and LRU lock, no race with concurrent
108 * object lookup is possible and we can safely destroy
111 cfs_hlist_del_init(&top->loh_hash);
112 cfs_list_del_init(&top->loh_lru);
116 } else if (lu_object_is_dying(top)) {
118 * somebody may be waiting for this, currently only used
119 * for cl_object, see cl_object_put_last().
121 cfs_waitq_broadcast(&site->ls_marche_funebre);
123 cfs_write_unlock(&site->ls_guard);
126 * Object was already removed from hash and lru above, can
129 lu_object_free(env, orig);
131 EXPORT_SYMBOL(lu_object_put);
134 * Allocate new object.
136 * This follows object creation protocol, described in the comment within
137 * struct lu_device_operations definition.
139 static struct lu_object *lu_object_alloc(const struct lu_env *env,
140 struct lu_device *dev,
141 const struct lu_fid *f,
142 const struct lu_object_conf *conf)
144 struct lu_object *scan;
145 struct lu_object *top;
152 * Create top-level object slice. This will also create
155 top = dev->ld_ops->ldo_object_alloc(env, NULL, dev);
157 RETURN(ERR_PTR(-ENOMEM));
159 * This is the only place where object fid is assigned. It's constant
162 LASSERT(fid_is_igif(f) || fid_ver(f) == 0);
163 top->lo_header->loh_fid = *f;
164 layers = &top->lo_header->loh_layers;
167 * Call ->loo_object_init() repeatedly, until no more new
168 * object slices are created.
171 cfs_list_for_each_entry(scan, layers, lo_linkage) {
172 if (scan->lo_flags & LU_OBJECT_ALLOCATED)
175 scan->lo_header = top->lo_header;
176 result = scan->lo_ops->loo_object_init(env, scan, conf);
178 lu_object_free(env, top);
179 RETURN(ERR_PTR(result));
181 scan->lo_flags |= LU_OBJECT_ALLOCATED;
185 cfs_list_for_each_entry_reverse(scan, layers, lo_linkage) {
186 if (scan->lo_ops->loo_object_start != NULL) {
187 result = scan->lo_ops->loo_object_start(env, scan);
189 lu_object_free(env, top);
190 RETURN(ERR_PTR(result));
195 dev->ld_site->ls_stats.s_created ++;
202 static void lu_object_free(const struct lu_env *env, struct lu_object *o)
205 struct lu_object *scan;
206 struct lu_site *site;
209 site = o->lo_dev->ld_site;
210 layers = &o->lo_header->loh_layers;
212 * First call ->loo_object_delete() method to release all resources.
214 cfs_list_for_each_entry_reverse(scan, layers, lo_linkage) {
215 if (scan->lo_ops->loo_object_delete != NULL)
216 scan->lo_ops->loo_object_delete(env, scan);
220 * Then, splice object layers into stand-alone list, and call
221 * ->loo_object_free() on all layers to free memory. Splice is
222 * necessary, because lu_object_header is freed together with the
225 CFS_INIT_LIST_HEAD(&splice);
226 cfs_list_splice_init(layers, &splice);
227 while (!cfs_list_empty(&splice)) {
229 * Free layers in bottom-to-top order, so that object header
230 * lives as long as possible and ->loo_object_free() methods
231 * can look at its contents.
233 o = container_of0(splice.prev, struct lu_object, lo_linkage);
234 cfs_list_del_init(&o->lo_linkage);
235 LASSERT(o->lo_ops->loo_object_free != NULL);
236 o->lo_ops->loo_object_free(env, o);
238 cfs_waitq_broadcast(&site->ls_marche_funebre);
242 * Free \a nr objects from the cold end of the site LRU list.
244 int lu_site_purge(const struct lu_env *env, struct lu_site *s, int nr)
247 struct lu_object_header *h;
248 struct lu_object_header *temp;
250 CFS_INIT_LIST_HEAD(&dispose);
252 * Under LRU list lock, scan LRU list and move unreferenced objects to
253 * the dispose list, removing them from LRU and hash table.
255 cfs_write_lock(&s->ls_guard);
256 cfs_list_for_each_entry_safe(h, temp, &s->ls_lru, loh_lru) {
258 * Objects are sorted in lru order, and "busy" objects (ones
259 * with h->loh_ref > 0) naturally tend to live near hot end
260 * that we scan last. Unfortunately, sites usually have small
261 * (less then ten) number of busy yet rarely accessed objects
262 * (some global objects, accessed directly through pointers,
263 * bypassing hash table). Currently algorithm scans them over
264 * and over again. Probably we should move busy objects out of
265 * LRU, or we can live with that.
269 if (cfs_atomic_read(&h->loh_ref) > 0)
271 cfs_hlist_del_init(&h->loh_hash);
272 cfs_list_move(&h->loh_lru, &dispose);
275 cfs_write_unlock(&s->ls_guard);
277 * Free everything on the dispose list. This is safe against races due
278 * to the reasons described in lu_object_put().
280 while (!cfs_list_empty(&dispose)) {
281 h = container_of0(dispose.next,
282 struct lu_object_header, loh_lru);
283 cfs_list_del_init(&h->loh_lru);
284 lu_object_free(env, lu_object_top(h));
285 s->ls_stats.s_lru_purged ++;
289 EXPORT_SYMBOL(lu_site_purge);
294 * Code below has to jump through certain loops to output object description
295 * into libcfs_debug_msg-based log. The problem is that lu_object_print()
296 * composes object description from strings that are parts of _lines_ of
297 * output (i.e., strings that are not terminated by newline). This doesn't fit
298 * very well into libcfs_debug_msg() interface that assumes that each message
299 * supplied to it is a self-contained output line.
301 * To work around this, strings are collected in a temporary buffer
302 * (implemented as a value of lu_cdebug_key key), until terminating newline
303 * character is detected.
311 * XXX overflow is not handled correctly.
316 struct lu_cdebug_data {
320 char lck_area[LU_CDEBUG_LINE];
323 /* context key constructor/destructor: lu_global_key_init, lu_global_key_fini */
324 LU_KEY_INIT_FINI(lu_global, struct lu_cdebug_data);
327 * Key, holding temporary buffer. This key is registered very early by
330 struct lu_context_key lu_global_key = {
331 .lct_tags = LCT_MD_THREAD|LCT_DT_THREAD|LCT_CL_THREAD,
332 .lct_init = lu_global_key_init,
333 .lct_fini = lu_global_key_fini
337 * Printer function emitting messages through libcfs_debug_msg().
339 int lu_cdebug_printer(const struct lu_env *env,
340 void *cookie, const char *format, ...)
342 struct lu_cdebug_print_info *info = cookie;
343 struct lu_cdebug_data *key;
348 va_start(args, format);
350 key = lu_context_key_get(&env->le_ctx, &lu_global_key);
351 LASSERT(key != NULL);
353 used = strlen(key->lck_area);
354 complete = format[strlen(format) - 1] == '\n';
356 * Append new chunk to the buffer.
358 vsnprintf(key->lck_area + used,
359 ARRAY_SIZE(key->lck_area) - used, format, args);
361 if (cfs_cdebug_show(info->lpi_mask, info->lpi_subsys))
362 libcfs_debug_msg(NULL, info->lpi_subsys, info->lpi_mask,
363 (char *)info->lpi_file, info->lpi_fn,
364 info->lpi_line, "%s", key->lck_area);
365 key->lck_area[0] = 0;
370 EXPORT_SYMBOL(lu_cdebug_printer);
373 * Print object header.
375 void lu_object_header_print(const struct lu_env *env, void *cookie,
376 lu_printer_t printer,
377 const struct lu_object_header *hdr)
379 (*printer)(env, cookie, "header@%p[%#lx, %d, "DFID"%s%s%s]",
380 hdr, hdr->loh_flags, cfs_atomic_read(&hdr->loh_ref),
382 cfs_hlist_unhashed(&hdr->loh_hash) ? "" : " hash",
383 cfs_list_empty((cfs_list_t *)&hdr->loh_lru) ? \
385 hdr->loh_attr & LOHA_EXISTS ? " exist":"");
387 EXPORT_SYMBOL(lu_object_header_print);
390 * Print human readable representation of the \a o to the \a printer.
392 void lu_object_print(const struct lu_env *env, void *cookie,
393 lu_printer_t printer, const struct lu_object *o)
395 static const char ruler[] = "........................................";
396 struct lu_object_header *top;
400 lu_object_header_print(env, cookie, printer, top);
401 (*printer)(env, cookie, "{ \n");
402 cfs_list_for_each_entry(o, &top->loh_layers, lo_linkage) {
403 depth = o->lo_depth + 4;
406 * print `.' \a depth times followed by type name and address
408 (*printer)(env, cookie, "%*.*s%s@%p", depth, depth, ruler,
409 o->lo_dev->ld_type->ldt_name, o);
410 if (o->lo_ops->loo_object_print != NULL)
411 o->lo_ops->loo_object_print(env, cookie, printer, o);
412 (*printer)(env, cookie, "\n");
414 (*printer)(env, cookie, "} header@%p\n", top);
416 EXPORT_SYMBOL(lu_object_print);
419 * Check object consistency.
421 int lu_object_invariant(const struct lu_object *o)
423 struct lu_object_header *top;
426 cfs_list_for_each_entry(o, &top->loh_layers, lo_linkage) {
427 if (o->lo_ops->loo_object_invariant != NULL &&
428 !o->lo_ops->loo_object_invariant(o))
433 EXPORT_SYMBOL(lu_object_invariant);
435 static struct lu_object *htable_lookup(struct lu_site *s,
436 const cfs_hlist_head_t *bucket,
437 const struct lu_fid *f,
438 cfs_waitlink_t *waiter)
440 struct lu_object_header *h;
441 cfs_hlist_node_t *scan;
443 cfs_hlist_for_each_entry(h, scan, bucket, loh_hash) {
444 s->ls_stats.s_cache_check ++;
445 if (likely(lu_fid_eq(&h->loh_fid, f))) {
446 if (unlikely(lu_object_is_dying(h))) {
448 * Lookup found an object being destroyed;
449 * this object cannot be returned (to assure
450 * that references to dying objects are
451 * eventually drained), and moreover, lookup
452 * has to wait until object is freed.
454 cfs_waitlink_init(waiter);
455 cfs_waitq_add(&s->ls_marche_funebre, waiter);
456 cfs_set_current_state(CFS_TASK_UNINT);
457 s->ls_stats.s_cache_death_race ++;
458 return ERR_PTR(-EAGAIN);
460 /* bump reference count... */
461 if (cfs_atomic_add_return(1, &h->loh_ref) == 1)
463 /* and move to the head of the LRU */
465 * XXX temporary disable this to measure effects of
466 * read-write locking.
468 /* list_move_tail(&h->loh_lru, &s->ls_lru); */
469 s->ls_stats.s_cache_hit ++;
470 return lu_object_top(h);
473 s->ls_stats.s_cache_miss ++;
477 static __u32 fid_hash(const struct lu_fid *f, int bits)
479 /* all objects with same id and different versions will belong to same
480 * collisions list. */
481 return cfs_hash_long(fid_flatten(f), bits);
485 * Search cache for an object with the fid \a f. If such object is found,
486 * return it. Otherwise, create new object, insert it into cache and return
487 * it. In any case, additional reference is acquired on the returned object.
489 struct lu_object *lu_object_find(const struct lu_env *env,
490 struct lu_device *dev, const struct lu_fid *f,
491 const struct lu_object_conf *conf)
493 return lu_object_find_at(env, dev->ld_site->ls_top_dev, f, conf);
495 EXPORT_SYMBOL(lu_object_find);
498 * Core logic of lu_object_find*() functions.
500 static struct lu_object *lu_object_find_try(const struct lu_env *env,
501 struct lu_device *dev,
502 const struct lu_fid *f,
503 const struct lu_object_conf *conf,
504 cfs_waitlink_t *waiter)
508 struct lu_object *shadow;
509 cfs_hlist_head_t *bucket;
512 * This uses standard index maintenance protocol:
514 * - search index under lock, and return object if found;
515 * - otherwise, unlock index, allocate new object;
516 * - lock index and search again;
517 * - if nothing is found (usual case), insert newly created
519 * - otherwise (race: other thread inserted object), free
520 * object just allocated.
524 * If dying object is found during index search, add @waiter to the
525 * site wait-queue and return ERR_PTR(-EAGAIN).
529 bucket = s->ls_hash + fid_hash(f, s->ls_hash_bits);
531 cfs_read_lock(&s->ls_guard);
532 o = htable_lookup(s, bucket, f, waiter);
533 cfs_read_unlock(&s->ls_guard);
539 * Allocate new object. This may result in rather complicated
540 * operations, including fld queries, inode loading, etc.
542 o = lu_object_alloc(env, dev, f, conf);
543 if (unlikely(IS_ERR(o)))
546 LASSERT(lu_fid_eq(lu_object_fid(o), f));
548 cfs_write_lock(&s->ls_guard);
549 shadow = htable_lookup(s, bucket, f, waiter);
550 if (likely(shadow == NULL)) {
551 cfs_hlist_add_head(&o->lo_header->loh_hash, bucket);
552 cfs_list_add_tail(&o->lo_header->loh_lru, &s->ls_lru);
558 s->ls_stats.s_cache_race ++;
559 cfs_write_unlock(&s->ls_guard);
561 lu_object_free(env, o);
566 * Much like lu_object_find(), but top level device of object is specifically
567 * \a dev rather than top level device of the site. This interface allows
568 * objects of different "stacking" to be created within the same site.
570 struct lu_object *lu_object_find_at(const struct lu_env *env,
571 struct lu_device *dev,
572 const struct lu_fid *f,
573 const struct lu_object_conf *conf)
575 struct lu_object *obj;
579 obj = lu_object_find_try(env, dev, f, conf, &wait);
580 if (obj == ERR_PTR(-EAGAIN)) {
582 * lu_object_find_try() already added waiter into the
585 cfs_waitq_wait(&wait, CFS_TASK_UNINT);
586 cfs_waitq_del(&dev->ld_site->ls_marche_funebre, &wait);
592 EXPORT_SYMBOL(lu_object_find_at);
595 * Find object with given fid, and return its slice belonging to given device.
597 struct lu_object *lu_object_find_slice(const struct lu_env *env,
598 struct lu_device *dev,
599 const struct lu_fid *f,
600 const struct lu_object_conf *conf)
602 struct lu_object *top;
603 struct lu_object *obj;
605 top = lu_object_find(env, dev, f, conf);
607 obj = lu_object_locate(top->lo_header, dev->ld_type);
609 lu_object_put(env, top);
614 EXPORT_SYMBOL(lu_object_find_slice);
617 * Global list of all device types.
619 static CFS_LIST_HEAD(lu_device_types);
621 int lu_device_type_init(struct lu_device_type *ldt)
625 CFS_INIT_LIST_HEAD(&ldt->ldt_linkage);
626 result = ldt->ldt_ops->ldto_init(ldt);
628 cfs_list_add(&ldt->ldt_linkage, &lu_device_types);
631 EXPORT_SYMBOL(lu_device_type_init);
633 void lu_device_type_fini(struct lu_device_type *ldt)
635 cfs_list_del_init(&ldt->ldt_linkage);
636 ldt->ldt_ops->ldto_fini(ldt);
638 EXPORT_SYMBOL(lu_device_type_fini);
640 void lu_types_stop(void)
642 struct lu_device_type *ldt;
644 cfs_list_for_each_entry(ldt, &lu_device_types, ldt_linkage) {
645 if (ldt->ldt_device_nr == 0)
646 ldt->ldt_ops->ldto_stop(ldt);
649 EXPORT_SYMBOL(lu_types_stop);
652 * Global list of all sites on this node
654 static CFS_LIST_HEAD(lu_sites);
655 static CFS_DECLARE_MUTEX(lu_sites_guard);
658 * Global environment used by site shrinker.
660 static struct lu_env lu_shrink_env;
663 * Print all objects in \a s.
665 void lu_site_print(const struct lu_env *env, struct lu_site *s, void *cookie,
666 lu_printer_t printer)
670 for (i = 0; i < s->ls_hash_size; ++i) {
671 struct lu_object_header *h;
672 cfs_hlist_node_t *scan;
674 cfs_read_lock(&s->ls_guard);
675 cfs_hlist_for_each_entry(h, scan, &s->ls_hash[i], loh_hash) {
677 if (!cfs_list_empty(&h->loh_layers)) {
678 const struct lu_object *obj;
680 obj = lu_object_top(h);
681 lu_object_print(env, cookie, printer, obj);
683 lu_object_header_print(env, cookie, printer, h);
685 cfs_read_unlock(&s->ls_guard);
688 EXPORT_SYMBOL(lu_site_print);
691 LU_CACHE_PERCENT = 20,
695 * Return desired hash table order.
697 static int lu_htable_order(void)
699 unsigned long cache_size;
703 * Calculate hash table size, assuming that we want reasonable
704 * performance when 20% of total memory is occupied by cache of
707 * Size of lu_object is (arbitrary) taken as 1K (together with inode).
709 cache_size = cfs_num_physpages;
711 #if BITS_PER_LONG == 32
712 /* limit hashtable size for lowmem systems to low RAM */
713 if (cache_size > 1 << (30 - CFS_PAGE_SHIFT))
714 cache_size = 1 << (30 - CFS_PAGE_SHIFT) * 3 / 4;
717 cache_size = cache_size / 100 * LU_CACHE_PERCENT *
718 (CFS_PAGE_SIZE / 1024);
720 for (bits = 1; (1 << bits) < cache_size; ++bits) {
726 static cfs_lock_class_key_t lu_site_guard_class;
729 * Initialize site \a s, with \a d as the top level device.
731 int lu_site_init(struct lu_site *s, struct lu_device *top)
738 memset(s, 0, sizeof *s);
739 cfs_rwlock_init(&s->ls_guard);
740 cfs_lockdep_set_class(&s->ls_guard, &lu_site_guard_class);
741 CFS_INIT_LIST_HEAD(&s->ls_lru);
742 CFS_INIT_LIST_HEAD(&s->ls_linkage);
743 cfs_waitq_init(&s->ls_marche_funebre);
747 lu_ref_add(&top->ld_reference, "site-top", s);
749 for (bits = lu_htable_order(), size = 1 << bits;
751 cfs_alloc_large(size * sizeof s->ls_hash[0])) == NULL;
752 --bits, size >>= 1) {
754 * Scale hash table down, until allocation succeeds.
759 s->ls_hash_size = size;
760 s->ls_hash_bits = bits;
761 s->ls_hash_mask = size - 1;
763 for (i = 0; i < size; i++)
764 CFS_INIT_HLIST_HEAD(&s->ls_hash[i]);
768 EXPORT_SYMBOL(lu_site_init);
771 * Finalize \a s and release its resources.
773 void lu_site_fini(struct lu_site *s)
775 LASSERT(cfs_list_empty(&s->ls_lru));
776 LASSERT(s->ls_total == 0);
778 cfs_down(&lu_sites_guard);
779 cfs_list_del_init(&s->ls_linkage);
780 cfs_up(&lu_sites_guard);
782 if (s->ls_hash != NULL) {
784 for (i = 0; i < s->ls_hash_size; i++)
785 LASSERT(cfs_hlist_empty(&s->ls_hash[i]));
786 cfs_free_large(s->ls_hash);
789 if (s->ls_top_dev != NULL) {
790 s->ls_top_dev->ld_site = NULL;
791 lu_ref_del(&s->ls_top_dev->ld_reference, "site-top", s);
792 lu_device_put(s->ls_top_dev);
793 s->ls_top_dev = NULL;
796 EXPORT_SYMBOL(lu_site_fini);
799 * Called when initialization of stack for this site is completed.
801 int lu_site_init_finish(struct lu_site *s)
804 cfs_down(&lu_sites_guard);
805 result = lu_context_refill(&lu_shrink_env.le_ctx);
807 cfs_list_add(&s->ls_linkage, &lu_sites);
808 cfs_up(&lu_sites_guard);
811 EXPORT_SYMBOL(lu_site_init_finish);
814 * Acquire additional reference on device \a d
816 void lu_device_get(struct lu_device *d)
818 cfs_atomic_inc(&d->ld_ref);
820 EXPORT_SYMBOL(lu_device_get);
823 * Release reference on device \a d.
825 void lu_device_put(struct lu_device *d)
827 LASSERT(cfs_atomic_read(&d->ld_ref) > 0);
828 cfs_atomic_dec(&d->ld_ref);
830 EXPORT_SYMBOL(lu_device_put);
833 * Initialize device \a d of type \a t.
835 int lu_device_init(struct lu_device *d, struct lu_device_type *t)
837 if (t->ldt_device_nr++ == 0 && t->ldt_ops->ldto_start != NULL)
838 t->ldt_ops->ldto_start(t);
839 memset(d, 0, sizeof *d);
840 cfs_atomic_set(&d->ld_ref, 0);
842 lu_ref_init(&d->ld_reference);
845 EXPORT_SYMBOL(lu_device_init);
848 * Finalize device \a d.
850 void lu_device_fini(struct lu_device *d)
852 struct lu_device_type *t;
855 if (d->ld_obd != NULL) {
856 d->ld_obd->obd_lu_dev = NULL;
860 lu_ref_fini(&d->ld_reference);
861 LASSERTF(cfs_atomic_read(&d->ld_ref) == 0,
862 "Refcount is %u\n", cfs_atomic_read(&d->ld_ref));
863 LASSERT(t->ldt_device_nr > 0);
864 if (--t->ldt_device_nr == 0 && t->ldt_ops->ldto_stop != NULL)
865 t->ldt_ops->ldto_stop(t);
867 EXPORT_SYMBOL(lu_device_fini);
870 * Initialize object \a o that is part of compound object \a h and was created
873 int lu_object_init(struct lu_object *o,
874 struct lu_object_header *h, struct lu_device *d)
876 memset(o, 0, sizeof *o);
880 o->lo_dev_ref = lu_ref_add(&d->ld_reference, "lu_object", o);
881 CFS_INIT_LIST_HEAD(&o->lo_linkage);
884 EXPORT_SYMBOL(lu_object_init);
887 * Finalize object and release its resources.
889 void lu_object_fini(struct lu_object *o)
891 struct lu_device *dev = o->lo_dev;
893 LASSERT(cfs_list_empty(&o->lo_linkage));
896 lu_ref_del_at(&dev->ld_reference,
897 o->lo_dev_ref , "lu_object", o);
902 EXPORT_SYMBOL(lu_object_fini);
905 * Add object \a o as first layer of compound object \a h
907 * This is typically called by the ->ldo_object_alloc() method of top-level
910 void lu_object_add_top(struct lu_object_header *h, struct lu_object *o)
912 cfs_list_move(&o->lo_linkage, &h->loh_layers);
914 EXPORT_SYMBOL(lu_object_add_top);
917 * Add object \a o as a layer of compound object, going after \a before.
919 * This is typically called by the ->ldo_object_alloc() method of \a
922 void lu_object_add(struct lu_object *before, struct lu_object *o)
924 cfs_list_move(&o->lo_linkage, &before->lo_linkage);
926 EXPORT_SYMBOL(lu_object_add);
929 * Initialize compound object.
931 int lu_object_header_init(struct lu_object_header *h)
933 memset(h, 0, sizeof *h);
934 cfs_atomic_set(&h->loh_ref, 1);
935 CFS_INIT_HLIST_NODE(&h->loh_hash);
936 CFS_INIT_LIST_HEAD(&h->loh_lru);
937 CFS_INIT_LIST_HEAD(&h->loh_layers);
938 lu_ref_init(&h->loh_reference);
941 EXPORT_SYMBOL(lu_object_header_init);
944 * Finalize compound object.
946 void lu_object_header_fini(struct lu_object_header *h)
948 LASSERT(cfs_list_empty(&h->loh_layers));
949 LASSERT(cfs_list_empty(&h->loh_lru));
950 LASSERT(cfs_hlist_unhashed(&h->loh_hash));
951 lu_ref_fini(&h->loh_reference);
953 EXPORT_SYMBOL(lu_object_header_fini);
956 * Given a compound object, find its slice, corresponding to the device type
959 struct lu_object *lu_object_locate(struct lu_object_header *h,
960 const struct lu_device_type *dtype)
964 cfs_list_for_each_entry(o, &h->loh_layers, lo_linkage) {
965 if (o->lo_dev->ld_type == dtype)
970 EXPORT_SYMBOL(lu_object_locate);
975 * Finalize and free devices in the device stack.
977 * Finalize device stack by purging object cache, and calling
978 * lu_device_type_operations::ldto_device_fini() and
979 * lu_device_type_operations::ldto_device_free() on all devices in the stack.
981 void lu_stack_fini(const struct lu_env *env, struct lu_device *top)
983 struct lu_site *site = top->ld_site;
984 struct lu_device *scan;
985 struct lu_device *next;
987 lu_site_purge(env, site, ~0);
988 for (scan = top; scan != NULL; scan = next) {
989 next = scan->ld_type->ldt_ops->ldto_device_fini(env, scan);
990 lu_ref_del(&scan->ld_reference, "lu-stack", &lu_site_init);
995 lu_site_purge(env, site, ~0);
997 if (!cfs_list_empty(&site->ls_lru) || site->ls_total != 0) {
999 * Uh-oh, objects still exist.
1001 static DECLARE_LU_CDEBUG_PRINT_INFO(cookie, D_ERROR);
1003 lu_site_print(env, site, &cookie, lu_cdebug_printer);
1006 for (scan = top; scan != NULL; scan = next) {
1007 const struct lu_device_type *ldt = scan->ld_type;
1008 struct obd_type *type;
1010 next = ldt->ldt_ops->ldto_device_free(env, scan);
1011 type = ldt->ldt_obd_type;
1014 class_put_type(type);
1018 EXPORT_SYMBOL(lu_stack_fini);
1022 * Maximal number of tld slots.
1024 LU_CONTEXT_KEY_NR = 32
1027 static struct lu_context_key *lu_keys[LU_CONTEXT_KEY_NR] = { NULL, };
1029 static cfs_spinlock_t lu_keys_guard = CFS_SPIN_LOCK_UNLOCKED;
1032 * Global counter incremented whenever key is registered, unregistered,
1033 * revived or quiesced. This is used to void unnecessary calls to
1034 * lu_context_refill(). No locking is provided, as initialization and shutdown
1035 * are supposed to be externally serialized.
1037 static unsigned key_set_version = 0;
1042 int lu_context_key_register(struct lu_context_key *key)
1047 LASSERT(key->lct_init != NULL);
1048 LASSERT(key->lct_fini != NULL);
1049 LASSERT(key->lct_tags != 0);
1050 LASSERT(key->lct_owner != NULL);
1053 cfs_spin_lock(&lu_keys_guard);
1054 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1055 if (lu_keys[i] == NULL) {
1057 cfs_atomic_set(&key->lct_used, 1);
1059 lu_ref_init(&key->lct_reference);
1065 cfs_spin_unlock(&lu_keys_guard);
1068 EXPORT_SYMBOL(lu_context_key_register);
1070 static void key_fini(struct lu_context *ctx, int index)
1072 if (ctx->lc_value != NULL && ctx->lc_value[index] != NULL) {
1073 struct lu_context_key *key;
1075 key = lu_keys[index];
1076 LASSERT(key != NULL);
1077 LASSERT(key->lct_fini != NULL);
1078 LASSERT(cfs_atomic_read(&key->lct_used) > 1);
1080 key->lct_fini(ctx, key, ctx->lc_value[index]);
1081 lu_ref_del(&key->lct_reference, "ctx", ctx);
1082 cfs_atomic_dec(&key->lct_used);
1083 LASSERT(key->lct_owner != NULL);
1084 if (!(ctx->lc_tags & LCT_NOREF)) {
1085 LASSERT(cfs_module_refcount(key->lct_owner) > 0);
1086 cfs_module_put(key->lct_owner);
1088 ctx->lc_value[index] = NULL;
1095 void lu_context_key_degister(struct lu_context_key *key)
1097 LASSERT(cfs_atomic_read(&key->lct_used) >= 1);
1098 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1100 lu_context_key_quiesce(key);
1103 cfs_spin_lock(&lu_keys_guard);
1104 key_fini(&lu_shrink_env.le_ctx, key->lct_index);
1105 if (lu_keys[key->lct_index]) {
1106 lu_keys[key->lct_index] = NULL;
1107 lu_ref_fini(&key->lct_reference);
1109 cfs_spin_unlock(&lu_keys_guard);
1111 LASSERTF(cfs_atomic_read(&key->lct_used) == 1,
1112 "key has instances: %d\n",
1113 cfs_atomic_read(&key->lct_used));
1115 EXPORT_SYMBOL(lu_context_key_degister);
1118 * Register a number of keys. This has to be called after all keys have been
1119 * initialized by a call to LU_CONTEXT_KEY_INIT().
1121 int lu_context_key_register_many(struct lu_context_key *k, ...)
1123 struct lu_context_key *key = k;
1129 result = lu_context_key_register(key);
1132 key = va_arg(args, struct lu_context_key *);
1133 } while (key != NULL);
1139 lu_context_key_degister(k);
1140 k = va_arg(args, struct lu_context_key *);
1147 EXPORT_SYMBOL(lu_context_key_register_many);
1150 * De-register a number of keys. This is a dual to
1151 * lu_context_key_register_many().
1153 void lu_context_key_degister_many(struct lu_context_key *k, ...)
1159 lu_context_key_degister(k);
1160 k = va_arg(args, struct lu_context_key*);
1161 } while (k != NULL);
1164 EXPORT_SYMBOL(lu_context_key_degister_many);
1167 * Revive a number of keys.
1169 void lu_context_key_revive_many(struct lu_context_key *k, ...)
1175 lu_context_key_revive(k);
1176 k = va_arg(args, struct lu_context_key*);
1177 } while (k != NULL);
1180 EXPORT_SYMBOL(lu_context_key_revive_many);
1183 * Quiescent a number of keys.
1185 void lu_context_key_quiesce_many(struct lu_context_key *k, ...)
1191 lu_context_key_quiesce(k);
1192 k = va_arg(args, struct lu_context_key*);
1193 } while (k != NULL);
1196 EXPORT_SYMBOL(lu_context_key_quiesce_many);
1199 * Return value associated with key \a key in context \a ctx.
1201 void *lu_context_key_get(const struct lu_context *ctx,
1202 const struct lu_context_key *key)
1204 LINVRNT(ctx->lc_state == LCS_ENTERED);
1205 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1206 LASSERT(lu_keys[key->lct_index] == key);
1207 return ctx->lc_value[key->lct_index];
1209 EXPORT_SYMBOL(lu_context_key_get);
1212 * List of remembered contexts. XXX document me.
1214 static CFS_LIST_HEAD(lu_context_remembered);
1217 * Destroy \a key in all remembered contexts. This is used to destroy key
1218 * values in "shared" contexts (like service threads), when a module owning
1219 * the key is about to be unloaded.
1221 void lu_context_key_quiesce(struct lu_context_key *key)
1223 struct lu_context *ctx;
1224 extern unsigned cl_env_cache_purge(unsigned nr);
1226 if (!(key->lct_tags & LCT_QUIESCENT)) {
1228 * XXX layering violation.
1230 cl_env_cache_purge(~0);
1231 key->lct_tags |= LCT_QUIESCENT;
1233 * XXX memory barrier has to go here.
1235 cfs_spin_lock(&lu_keys_guard);
1236 cfs_list_for_each_entry(ctx, &lu_context_remembered,
1238 key_fini(ctx, key->lct_index);
1239 cfs_spin_unlock(&lu_keys_guard);
1243 EXPORT_SYMBOL(lu_context_key_quiesce);
1245 void lu_context_key_revive(struct lu_context_key *key)
1247 key->lct_tags &= ~LCT_QUIESCENT;
1250 EXPORT_SYMBOL(lu_context_key_revive);
1252 static void keys_fini(struct lu_context *ctx)
1256 cfs_spin_lock(&lu_keys_guard);
1257 if (ctx->lc_value != NULL) {
1258 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i)
1260 OBD_FREE(ctx->lc_value,
1261 ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1262 ctx->lc_value = NULL;
1264 cfs_spin_unlock(&lu_keys_guard);
1267 static int keys_fill(struct lu_context *ctx)
1271 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1272 struct lu_context_key *key;
1275 if (ctx->lc_value[i] == NULL && key != NULL &&
1276 (key->lct_tags & ctx->lc_tags) &&
1278 * Don't create values for a LCT_QUIESCENT key, as this
1279 * will pin module owning a key.
1281 !(key->lct_tags & LCT_QUIESCENT)) {
1284 LINVRNT(key->lct_init != NULL);
1285 LINVRNT(key->lct_index == i);
1287 value = key->lct_init(ctx, key);
1288 if (unlikely(IS_ERR(value)))
1289 return PTR_ERR(value);
1291 LASSERT(key->lct_owner != NULL);
1292 if (!(ctx->lc_tags & LCT_NOREF))
1293 cfs_try_module_get(key->lct_owner);
1294 lu_ref_add_atomic(&key->lct_reference, "ctx", ctx);
1295 cfs_atomic_inc(&key->lct_used);
1297 * This is the only place in the code, where an
1298 * element of ctx->lc_value[] array is set to non-NULL
1301 ctx->lc_value[i] = value;
1302 if (key->lct_exit != NULL)
1303 ctx->lc_tags |= LCT_HAS_EXIT;
1305 ctx->lc_version = key_set_version;
1310 static int keys_init(struct lu_context *ctx)
1314 OBD_ALLOC(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1315 if (likely(ctx->lc_value != NULL))
1316 result = keys_fill(ctx);
1326 * Initialize context data-structure. Create values for all keys.
1328 int lu_context_init(struct lu_context *ctx, __u32 tags)
1330 memset(ctx, 0, sizeof *ctx);
1331 ctx->lc_state = LCS_INITIALIZED;
1332 ctx->lc_tags = tags;
1333 if (tags & LCT_REMEMBER) {
1334 cfs_spin_lock(&lu_keys_guard);
1335 cfs_list_add(&ctx->lc_remember, &lu_context_remembered);
1336 cfs_spin_unlock(&lu_keys_guard);
1338 CFS_INIT_LIST_HEAD(&ctx->lc_remember);
1339 return keys_init(ctx);
1341 EXPORT_SYMBOL(lu_context_init);
1344 * Finalize context data-structure. Destroy key values.
1346 void lu_context_fini(struct lu_context *ctx)
1348 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1349 ctx->lc_state = LCS_FINALIZED;
1351 cfs_spin_lock(&lu_keys_guard);
1352 cfs_list_del_init(&ctx->lc_remember);
1353 cfs_spin_unlock(&lu_keys_guard);
1355 EXPORT_SYMBOL(lu_context_fini);
1358 * Called before entering context.
1360 void lu_context_enter(struct lu_context *ctx)
1362 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1363 ctx->lc_state = LCS_ENTERED;
1365 EXPORT_SYMBOL(lu_context_enter);
1368 * Called after exiting from \a ctx
1370 void lu_context_exit(struct lu_context *ctx)
1374 LINVRNT(ctx->lc_state == LCS_ENTERED);
1375 ctx->lc_state = LCS_LEFT;
1376 if (ctx->lc_tags & LCT_HAS_EXIT && ctx->lc_value != NULL) {
1377 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1378 if (ctx->lc_value[i] != NULL) {
1379 struct lu_context_key *key;
1382 LASSERT(key != NULL);
1383 if (key->lct_exit != NULL)
1385 key, ctx->lc_value[i]);
1390 EXPORT_SYMBOL(lu_context_exit);
1393 * Allocate for context all missing keys that were registered after context
1396 int lu_context_refill(struct lu_context *ctx)
1398 LINVRNT(ctx->lc_value != NULL);
1399 return ctx->lc_version == key_set_version ? 0 : keys_fill(ctx);
1401 EXPORT_SYMBOL(lu_context_refill);
1403 int lu_env_init(struct lu_env *env, __u32 tags)
1408 result = lu_context_init(&env->le_ctx, tags);
1409 if (likely(result == 0))
1410 lu_context_enter(&env->le_ctx);
1413 EXPORT_SYMBOL(lu_env_init);
1415 void lu_env_fini(struct lu_env *env)
1417 lu_context_exit(&env->le_ctx);
1418 lu_context_fini(&env->le_ctx);
1421 EXPORT_SYMBOL(lu_env_fini);
1423 int lu_env_refill(struct lu_env *env)
1427 result = lu_context_refill(&env->le_ctx);
1428 if (result == 0 && env->le_ses != NULL)
1429 result = lu_context_refill(env->le_ses);
1432 EXPORT_SYMBOL(lu_env_refill);
1434 static struct cfs_shrinker *lu_site_shrinker = NULL;
1437 static int lu_cache_shrink(int nr, unsigned int gfp_mask)
1440 struct lu_site *tmp;
1443 CFS_LIST_HEAD(splice);
1446 if (!(gfp_mask & __GFP_FS))
1448 CDEBUG(D_INODE, "Shrink %d objects\n", nr);
1451 cfs_down(&lu_sites_guard);
1452 cfs_list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
1454 remain = lu_site_purge(&lu_shrink_env, s, remain);
1456 * Move just shrunk site to the tail of site list to
1457 * assure shrinking fairness.
1459 cfs_list_move_tail(&s->ls_linkage, &splice);
1461 cfs_read_lock(&s->ls_guard);
1462 cached += s->ls_total - s->ls_busy;
1463 cfs_read_unlock(&s->ls_guard);
1464 if (nr && remain <= 0)
1467 cfs_list_splice(&splice, lu_sites.prev);
1468 cfs_up(&lu_sites_guard);
1470 cached = (cached / 100) * sysctl_vfs_cache_pressure;
1472 CDEBUG(D_INODE, "%d objects cached\n", cached);
1481 * Environment to be used in debugger, contains all tags.
1483 struct lu_env lu_debugging_env;
1486 * Debugging printer function using printk().
1488 int lu_printk_printer(const struct lu_env *env,
1489 void *unused, const char *format, ...)
1493 va_start(args, format);
1494 vprintk(format, args);
1499 void lu_debugging_setup(void)
1501 lu_env_init(&lu_debugging_env, ~0);
1504 void lu_context_keys_dump(void)
1508 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1509 struct lu_context_key *key;
1513 CERROR("[%i]: %p %x (%p,%p,%p) %i %i \"%s\"@%p\n",
1514 i, key, key->lct_tags,
1515 key->lct_init, key->lct_fini, key->lct_exit,
1516 key->lct_index, cfs_atomic_read(&key->lct_used),
1517 key->lct_owner ? key->lct_owner->name : "",
1519 lu_ref_print(&key->lct_reference);
1523 EXPORT_SYMBOL(lu_context_keys_dump);
1524 #else /* !__KERNEL__ */
1525 static int lu_cache_shrink(int nr, unsigned int gfp_mask)
1529 #endif /* __KERNEL__ */
1531 int cl_global_init(void);
1532 void cl_global_fini(void);
1533 int lu_ref_global_init(void);
1534 void lu_ref_global_fini(void);
1536 int dt_global_init(void);
1537 void dt_global_fini(void);
1539 int llo_global_init(void);
1540 void llo_global_fini(void);
1543 * Initialization of global lu_* data.
1545 int lu_global_init(void)
1549 CDEBUG(D_CONSOLE, "Lustre LU module (%p).\n", &lu_keys);
1551 result = lu_ref_global_init();
1555 LU_CONTEXT_KEY_INIT(&lu_global_key);
1556 result = lu_context_key_register(&lu_global_key);
1560 * At this level, we don't know what tags are needed, so allocate them
1561 * conservatively. This should not be too bad, because this
1562 * environment is global.
1564 cfs_down(&lu_sites_guard);
1565 result = lu_env_init(&lu_shrink_env, LCT_SHRINKER);
1566 cfs_up(&lu_sites_guard);
1571 * seeks estimation: 3 seeks to read a record from oi, one to read
1572 * inode, one for ea. Unfortunately setting this high value results in
1573 * lu_object/inode cache consuming all the memory.
1575 lu_site_shrinker = cfs_set_shrinker(CFS_DEFAULT_SEEKS, lu_cache_shrink);
1576 if (lu_site_shrinker == NULL)
1579 result = lu_time_global_init();
1584 result = dt_global_init();
1588 result = llo_global_init();
1592 result = cl_global_init();
1599 * Dual to lu_global_init().
1601 void lu_global_fini(void)
1608 lu_time_global_fini();
1609 if (lu_site_shrinker != NULL) {
1610 cfs_remove_shrinker(lu_site_shrinker);
1611 lu_site_shrinker = NULL;
1614 lu_context_key_degister(&lu_global_key);
1617 * Tear shrinker environment down _after_ de-registering
1618 * lu_global_key, because the latter has a value in the former.
1620 cfs_down(&lu_sites_guard);
1621 lu_env_fini(&lu_shrink_env);
1622 cfs_up(&lu_sites_guard);
1624 lu_ref_global_fini();
1627 struct lu_buf LU_BUF_NULL = {
1631 EXPORT_SYMBOL(LU_BUF_NULL);
1634 * Output site statistical counters into a buffer. Suitable for
1635 * lprocfs_rd_*()-style functions.
1637 int lu_site_stats_print(const struct lu_site *s, char *page, int count)
1643 * How many hash buckets are not-empty? Don't bother with locks: it's
1644 * an estimation anyway.
1646 for (i = 0, populated = 0; i < s->ls_hash_size; i++)
1647 populated += !cfs_hlist_empty(&s->ls_hash[i]);
1649 return snprintf(page, count, "%d %d %d/%d %d %d %d %d %d %d %d\n",
1654 s->ls_stats.s_created,
1655 s->ls_stats.s_cache_hit,
1656 s->ls_stats.s_cache_miss,
1657 s->ls_stats.s_cache_check,
1658 s->ls_stats.s_cache_race,
1659 s->ls_stats.s_cache_death_race,
1660 s->ls_stats.s_lru_purged);
1662 EXPORT_SYMBOL(lu_site_stats_print);
1664 const char *lu_time_names[LU_TIME_NR] = {
1665 [LU_TIME_FIND_LOOKUP] = "find_lookup",
1666 [LU_TIME_FIND_ALLOC] = "find_alloc",
1667 [LU_TIME_FIND_INSERT] = "find_insert"
1669 EXPORT_SYMBOL(lu_time_names);
1672 * Helper function to initialize a number of kmem slab caches at once.
1674 int lu_kmem_init(struct lu_kmem_descr *caches)
1678 for (result = 0; caches->ckd_cache != NULL; ++caches) {
1679 *caches->ckd_cache = cfs_mem_cache_create(caches->ckd_name,
1682 if (*caches->ckd_cache == NULL) {
1689 EXPORT_SYMBOL(lu_kmem_init);
1692 * Helper function to finalize a number of kmem slab cached at once. Dual to
1695 void lu_kmem_fini(struct lu_kmem_descr *caches)
1699 for (; caches->ckd_cache != NULL; ++caches) {
1700 if (*caches->ckd_cache != NULL) {
1701 rc = cfs_mem_cache_destroy(*caches->ckd_cache);
1702 LASSERTF(rc == 0, "couldn't destroy %s slab\n",
1704 *caches->ckd_cache = NULL;
1708 EXPORT_SYMBOL(lu_kmem_fini);