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 * Copyright (c) 2011 Whamcloud, Inc.
36 * This file is part of Lustre, http://www.lustre.org/
37 * Lustre is a trademark of Sun Microsystems, Inc.
39 * lustre/obdclass/lu_object.c
42 * These are the only exported functions, they provide some generic
43 * infrastructure for managing object devices
45 * Author: Nikita Danilov <nikita.danilov@sun.com>
48 #define DEBUG_SUBSYSTEM S_CLASS
50 # define EXPORT_SYMTAB
53 #include <libcfs/libcfs.h>
56 # include <linux/module.h>
60 #include <libcfs/libcfs_hash.h>
61 #include <obd_class.h>
62 #include <obd_support.h>
63 #include <lustre_disk.h>
64 #include <lustre_fid.h>
65 #include <lu_object.h>
66 #include <libcfs/list.h>
67 /* lu_time_global_{init,fini}() */
70 static void lu_object_free(const struct lu_env *env, struct lu_object *o);
73 * Decrease reference counter on object. If last reference is freed, return
74 * object to the cache, unless lu_object_is_dying(o) holds. In the latter
75 * case, free object immediately.
77 void lu_object_put(const struct lu_env *env, struct lu_object *o)
79 struct lu_site_bkt_data *bkt;
80 struct lu_object_header *top;
82 struct lu_object *orig;
86 site = o->lo_dev->ld_site;
89 cfs_hash_bd_get(site->ls_obj_hash, &top->loh_fid, &bd);
90 bkt = cfs_hash_bd_extra_get(site->ls_obj_hash, &bd);
92 if (!cfs_hash_bd_dec_and_lock(site->ls_obj_hash, &bd, &top->loh_ref)) {
93 if (lu_object_is_dying(top)) {
96 * somebody may be waiting for this, currently only
97 * used for cl_object, see cl_object_put_last().
99 cfs_waitq_broadcast(&bkt->lsb_marche_funebre);
104 LASSERT(bkt->lsb_busy > 0);
107 * When last reference is released, iterate over object
108 * layers, and notify them that object is no longer busy.
110 cfs_list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
111 if (o->lo_ops->loo_object_release != NULL)
112 o->lo_ops->loo_object_release(env, o);
115 if (!lu_object_is_dying(top)) {
116 LASSERT(cfs_list_empty(&top->loh_lru));
117 cfs_list_add_tail(&top->loh_lru, &bkt->lsb_lru);
118 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
123 * If object is dying (will not be cached), removed it
124 * from hash table and LRU.
126 * This is done with hash table and LRU lists locked. As the only
127 * way to acquire first reference to previously unreferenced
128 * object is through hash-table lookup (lu_object_find()),
129 * or LRU scanning (lu_site_purge()), that are done under hash-table
130 * and LRU lock, no race with concurrent object lookup is possible
131 * and we can safely destroy object below.
133 cfs_hash_bd_del_locked(site->ls_obj_hash, &bd, &top->loh_hash);
134 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
136 * Object was already removed from hash and lru above, can
139 lu_object_free(env, orig);
141 EXPORT_SYMBOL(lu_object_put);
144 * Allocate new object.
146 * This follows object creation protocol, described in the comment within
147 * struct lu_device_operations definition.
149 static struct lu_object *lu_object_alloc(const struct lu_env *env,
150 struct lu_device *dev,
151 const struct lu_fid *f,
152 const struct lu_object_conf *conf)
154 struct lu_object *scan;
155 struct lu_object *top;
162 * Create top-level object slice. This will also create
165 top = dev->ld_ops->ldo_object_alloc(env, NULL, dev);
167 RETURN(ERR_PTR(-ENOMEM));
169 * This is the only place where object fid is assigned. It's constant
172 LASSERT(fid_is_igif(f) || fid_ver(f) == 0);
173 top->lo_header->loh_fid = *f;
174 layers = &top->lo_header->loh_layers;
177 * Call ->loo_object_init() repeatedly, until no more new
178 * object slices are created.
181 cfs_list_for_each_entry(scan, layers, lo_linkage) {
182 if (scan->lo_flags & LU_OBJECT_ALLOCATED)
185 scan->lo_header = top->lo_header;
186 result = scan->lo_ops->loo_object_init(env, scan, conf);
188 lu_object_free(env, top);
189 RETURN(ERR_PTR(result));
191 scan->lo_flags |= LU_OBJECT_ALLOCATED;
195 cfs_list_for_each_entry_reverse(scan, layers, lo_linkage) {
196 if (scan->lo_ops->loo_object_start != NULL) {
197 result = scan->lo_ops->loo_object_start(env, scan);
199 lu_object_free(env, top);
200 RETURN(ERR_PTR(result));
205 lprocfs_counter_incr(dev->ld_site->ls_stats, LU_SS_CREATED);
212 static void lu_object_free(const struct lu_env *env, struct lu_object *o)
214 struct lu_site_bkt_data *bkt;
215 struct lu_site *site;
216 struct lu_object *scan;
220 site = o->lo_dev->ld_site;
221 layers = &o->lo_header->loh_layers;
222 bkt = lu_site_bkt_from_fid(site, &o->lo_header->loh_fid);
224 * First call ->loo_object_delete() method to release all resources.
226 cfs_list_for_each_entry_reverse(scan, layers, lo_linkage) {
227 if (scan->lo_ops->loo_object_delete != NULL)
228 scan->lo_ops->loo_object_delete(env, scan);
232 * Then, splice object layers into stand-alone list, and call
233 * ->loo_object_free() on all layers to free memory. Splice is
234 * necessary, because lu_object_header is freed together with the
237 CFS_INIT_LIST_HEAD(&splice);
238 cfs_list_splice_init(layers, &splice);
239 while (!cfs_list_empty(&splice)) {
241 * Free layers in bottom-to-top order, so that object header
242 * lives as long as possible and ->loo_object_free() methods
243 * can look at its contents.
245 o = container_of0(splice.prev, struct lu_object, lo_linkage);
246 cfs_list_del_init(&o->lo_linkage);
247 LASSERT(o->lo_ops->loo_object_free != NULL);
248 o->lo_ops->loo_object_free(env, o);
251 if (cfs_waitq_active(&bkt->lsb_marche_funebre))
252 cfs_waitq_broadcast(&bkt->lsb_marche_funebre);
256 * Free \a nr objects from the cold end of the site LRU list.
258 int lu_site_purge(const struct lu_env *env, struct lu_site *s, int nr)
260 struct lu_object_header *h;
261 struct lu_object_header *temp;
262 struct lu_site_bkt_data *bkt;
272 CFS_INIT_LIST_HEAD(&dispose);
274 * Under LRU list lock, scan LRU list and move unreferenced objects to
275 * the dispose list, removing them from LRU and hash table.
277 start = s->ls_purge_start;
278 bnr = (nr == ~0) ? -1 : nr / CFS_HASH_NBKT(s->ls_obj_hash) + 1;
281 cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
285 cfs_hash_bd_lock(s->ls_obj_hash, &bd, 1);
286 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
288 cfs_list_for_each_entry_safe(h, temp, &bkt->lsb_lru, loh_lru) {
289 LASSERT(cfs_atomic_read(&h->loh_ref) == 0);
291 cfs_hash_bd_get(s->ls_obj_hash, &h->loh_fid, &bd2);
292 LASSERT(bd.bd_bucket == bd2.bd_bucket);
294 cfs_hash_bd_del_locked(s->ls_obj_hash,
296 cfs_list_move(&h->loh_lru, &dispose);
300 if (nr != ~0 && --nr == 0)
303 if (count > 0 && --count == 0)
307 cfs_hash_bd_unlock(s->ls_obj_hash, &bd, 1);
310 * Free everything on the dispose list. This is safe against
311 * races due to the reasons described in lu_object_put().
313 while (!cfs_list_empty(&dispose)) {
314 h = container_of0(dispose.next,
315 struct lu_object_header, loh_lru);
316 cfs_list_del_init(&h->loh_lru);
317 lu_object_free(env, lu_object_top(h));
318 lprocfs_counter_incr(s->ls_stats, LU_SS_LRU_PURGED);
325 if (nr != 0 && did_sth && start != 0) {
326 start = 0; /* restart from the first bucket */
329 /* race on s->ls_purge_start, but nobody cares */
330 s->ls_purge_start = i % CFS_HASH_NBKT(s->ls_obj_hash);
334 EXPORT_SYMBOL(lu_site_purge);
339 * Code below has to jump through certain loops to output object description
340 * into libcfs_debug_msg-based log. The problem is that lu_object_print()
341 * composes object description from strings that are parts of _lines_ of
342 * output (i.e., strings that are not terminated by newline). This doesn't fit
343 * very well into libcfs_debug_msg() interface that assumes that each message
344 * supplied to it is a self-contained output line.
346 * To work around this, strings are collected in a temporary buffer
347 * (implemented as a value of lu_cdebug_key key), until terminating newline
348 * character is detected.
356 * XXX overflow is not handled correctly.
361 struct lu_cdebug_data {
365 char lck_area[LU_CDEBUG_LINE];
368 /* context key constructor/destructor: lu_global_key_init, lu_global_key_fini */
369 LU_KEY_INIT_FINI(lu_global, struct lu_cdebug_data);
372 * Key, holding temporary buffer. This key is registered very early by
375 struct lu_context_key lu_global_key = {
376 .lct_tags = LCT_MD_THREAD|LCT_DT_THREAD|LCT_CL_THREAD,
377 .lct_init = lu_global_key_init,
378 .lct_fini = lu_global_key_fini
382 * Printer function emitting messages through libcfs_debug_msg().
384 int lu_cdebug_printer(const struct lu_env *env,
385 void *cookie, const char *format, ...)
387 struct lu_cdebug_print_info *info = cookie;
388 struct lu_cdebug_data *key;
393 va_start(args, format);
395 key = lu_context_key_get(&env->le_ctx, &lu_global_key);
396 LASSERT(key != NULL);
398 used = strlen(key->lck_area);
399 complete = format[strlen(format) - 1] == '\n';
401 * Append new chunk to the buffer.
403 vsnprintf(key->lck_area + used,
404 ARRAY_SIZE(key->lck_area) - used, format, args);
406 if (cfs_cdebug_show(info->lpi_mask, info->lpi_subsys))
407 libcfs_debug_msg(NULL, info->lpi_subsys, info->lpi_mask,
408 (char *)info->lpi_file, info->lpi_fn,
409 info->lpi_line, "%s", key->lck_area);
410 key->lck_area[0] = 0;
415 EXPORT_SYMBOL(lu_cdebug_printer);
418 * Print object header.
420 void lu_object_header_print(const struct lu_env *env, void *cookie,
421 lu_printer_t printer,
422 const struct lu_object_header *hdr)
424 (*printer)(env, cookie, "header@%p[%#lx, %d, "DFID"%s%s%s]",
425 hdr, hdr->loh_flags, cfs_atomic_read(&hdr->loh_ref),
427 cfs_hlist_unhashed(&hdr->loh_hash) ? "" : " hash",
428 cfs_list_empty((cfs_list_t *)&hdr->loh_lru) ? \
430 hdr->loh_attr & LOHA_EXISTS ? " exist":"");
432 EXPORT_SYMBOL(lu_object_header_print);
435 * Print human readable representation of the \a o to the \a printer.
437 void lu_object_print(const struct lu_env *env, void *cookie,
438 lu_printer_t printer, const struct lu_object *o)
440 static const char ruler[] = "........................................";
441 struct lu_object_header *top;
445 lu_object_header_print(env, cookie, printer, top);
446 (*printer)(env, cookie, "{ \n");
447 cfs_list_for_each_entry(o, &top->loh_layers, lo_linkage) {
448 depth = o->lo_depth + 4;
451 * print `.' \a depth times followed by type name and address
453 (*printer)(env, cookie, "%*.*s%s@%p", depth, depth, ruler,
454 o->lo_dev->ld_type->ldt_name, o);
455 if (o->lo_ops->loo_object_print != NULL)
456 o->lo_ops->loo_object_print(env, cookie, printer, o);
457 (*printer)(env, cookie, "\n");
459 (*printer)(env, cookie, "} header@%p\n", top);
461 EXPORT_SYMBOL(lu_object_print);
464 * Check object consistency.
466 int lu_object_invariant(const struct lu_object *o)
468 struct lu_object_header *top;
471 cfs_list_for_each_entry(o, &top->loh_layers, lo_linkage) {
472 if (o->lo_ops->loo_object_invariant != NULL &&
473 !o->lo_ops->loo_object_invariant(o))
478 EXPORT_SYMBOL(lu_object_invariant);
480 static struct lu_object *htable_lookup(struct lu_site *s,
482 const struct lu_fid *f,
483 cfs_waitlink_t *waiter,
486 struct lu_site_bkt_data *bkt;
487 struct lu_object_header *h;
488 cfs_hlist_node_t *hnode;
489 __u64 ver = cfs_hash_bd_version_get(bd);
495 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, bd);
496 /* cfs_hash_bd_lookup_intent is a somehow "internal" function
497 * of cfs_hash, but we don't want refcount on object right now */
498 hnode = cfs_hash_bd_lookup_locked(s->ls_obj_hash, bd, (void *)f);
500 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS);
504 h = container_of0(hnode, struct lu_object_header, loh_hash);
505 if (likely(!lu_object_is_dying(h))) {
506 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_HIT);
507 return lu_object_top(h);
511 * Lookup found an object being destroyed this object cannot be
512 * returned (to assure that references to dying objects are eventually
513 * drained), and moreover, lookup has to wait until object is freed.
515 cfs_atomic_dec(&h->loh_ref);
517 cfs_waitlink_init(waiter);
518 cfs_waitq_add(&bkt->lsb_marche_funebre, waiter);
519 cfs_set_current_state(CFS_TASK_UNINT);
520 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_DEATH_RACE);
521 return ERR_PTR(-EAGAIN);
525 * Search cache for an object with the fid \a f. If such object is found,
526 * return it. Otherwise, create new object, insert it into cache and return
527 * it. In any case, additional reference is acquired on the returned object.
529 struct lu_object *lu_object_find(const struct lu_env *env,
530 struct lu_device *dev, const struct lu_fid *f,
531 const struct lu_object_conf *conf)
533 return lu_object_find_at(env, dev->ld_site->ls_top_dev, f, conf);
535 EXPORT_SYMBOL(lu_object_find);
537 static struct lu_object *lu_object_new(const struct lu_env *env,
538 struct lu_device *dev,
539 const struct lu_fid *f,
540 const struct lu_object_conf *conf)
545 struct lu_site_bkt_data *bkt;
547 o = lu_object_alloc(env, dev, f, conf);
548 if (unlikely(IS_ERR(o)))
551 hs = dev->ld_site->ls_obj_hash;
552 cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
553 bkt = cfs_hash_bd_extra_get(hs, &bd);
554 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
556 cfs_hash_bd_unlock(hs, &bd, 1);
561 * Core logic of lu_object_find*() functions.
563 static struct lu_object *lu_object_find_try(const struct lu_env *env,
564 struct lu_device *dev,
565 const struct lu_fid *f,
566 const struct lu_object_conf *conf,
567 cfs_waitlink_t *waiter)
570 struct lu_object *shadow;
577 * This uses standard index maintenance protocol:
579 * - search index under lock, and return object if found;
580 * - otherwise, unlock index, allocate new object;
581 * - lock index and search again;
582 * - if nothing is found (usual case), insert newly created
584 * - otherwise (race: other thread inserted object), free
585 * object just allocated.
589 * For "LOC_F_NEW" case, we are sure the object is new established.
590 * It is unnecessary to perform lookup-alloc-lookup-insert, instead,
591 * just alloc and insert directly.
593 * If dying object is found during index search, add @waiter to the
594 * site wait-queue and return ERR_PTR(-EAGAIN).
596 if (conf != NULL && conf->loc_flags & LOC_F_NEW)
597 return lu_object_new(env, dev, f, conf);
601 cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
602 o = htable_lookup(s, &bd, f, waiter, &version);
603 if (o != NULL && !cfs_list_empty(&o->lo_header->loh_lru))
604 cfs_list_del_init(&o->lo_header->loh_lru);
605 cfs_hash_bd_unlock(hs, &bd, 1);
610 * Allocate new object. This may result in rather complicated
611 * operations, including fld queries, inode loading, etc.
613 o = lu_object_alloc(env, dev, f, conf);
614 if (unlikely(IS_ERR(o)))
617 LASSERT(lu_fid_eq(lu_object_fid(o), f));
619 cfs_hash_bd_lock(hs, &bd, 1);
621 shadow = htable_lookup(s, &bd, f, waiter, &version);
622 if (likely(shadow == NULL)) {
623 struct lu_site_bkt_data *bkt;
625 bkt = cfs_hash_bd_extra_get(hs, &bd);
626 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
628 cfs_hash_bd_unlock(hs, &bd, 1);
632 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_RACE);
633 cfs_hash_bd_unlock(hs, &bd, 1);
634 lu_object_free(env, o);
639 * Much like lu_object_find(), but top level device of object is specifically
640 * \a dev rather than top level device of the site. This interface allows
641 * objects of different "stacking" to be created within the same site.
643 struct lu_object *lu_object_find_at(const struct lu_env *env,
644 struct lu_device *dev,
645 const struct lu_fid *f,
646 const struct lu_object_conf *conf)
648 struct lu_site_bkt_data *bkt;
649 struct lu_object *obj;
653 obj = lu_object_find_try(env, dev, f, conf, &wait);
654 if (obj != ERR_PTR(-EAGAIN))
657 * lu_object_find_try() already added waiter into the
660 cfs_waitq_wait(&wait, CFS_TASK_UNINT);
661 bkt = lu_site_bkt_from_fid(dev->ld_site, (void *)f);
662 cfs_waitq_del(&bkt->lsb_marche_funebre, &wait);
665 EXPORT_SYMBOL(lu_object_find_at);
668 * Find object with given fid, and return its slice belonging to given device.
670 struct lu_object *lu_object_find_slice(const struct lu_env *env,
671 struct lu_device *dev,
672 const struct lu_fid *f,
673 const struct lu_object_conf *conf)
675 struct lu_object *top;
676 struct lu_object *obj;
678 top = lu_object_find(env, dev, f, conf);
680 obj = lu_object_locate(top->lo_header, dev->ld_type);
682 lu_object_put(env, top);
687 EXPORT_SYMBOL(lu_object_find_slice);
690 * Global list of all device types.
692 static CFS_LIST_HEAD(lu_device_types);
694 int lu_device_type_init(struct lu_device_type *ldt)
698 CFS_INIT_LIST_HEAD(&ldt->ldt_linkage);
699 result = ldt->ldt_ops->ldto_init(ldt);
701 cfs_list_add(&ldt->ldt_linkage, &lu_device_types);
704 EXPORT_SYMBOL(lu_device_type_init);
706 void lu_device_type_fini(struct lu_device_type *ldt)
708 cfs_list_del_init(&ldt->ldt_linkage);
709 ldt->ldt_ops->ldto_fini(ldt);
711 EXPORT_SYMBOL(lu_device_type_fini);
713 void lu_types_stop(void)
715 struct lu_device_type *ldt;
717 cfs_list_for_each_entry(ldt, &lu_device_types, ldt_linkage) {
718 if (ldt->ldt_device_nr == 0)
719 ldt->ldt_ops->ldto_stop(ldt);
722 EXPORT_SYMBOL(lu_types_stop);
725 * Global list of all sites on this node
727 static CFS_LIST_HEAD(lu_sites);
728 static CFS_DECLARE_MUTEX(lu_sites_guard);
731 * Global environment used by site shrinker.
733 static struct lu_env lu_shrink_env;
735 struct lu_site_print_arg {
736 struct lu_env *lsp_env;
738 lu_printer_t lsp_printer;
742 lu_site_obj_print(cfs_hash_t *hs, cfs_hash_bd_t *bd,
743 cfs_hlist_node_t *hnode, void *data)
745 struct lu_site_print_arg *arg = (struct lu_site_print_arg *)data;
746 struct lu_object_header *h;
748 h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
749 if (!cfs_list_empty(&h->loh_layers)) {
750 const struct lu_object *o;
752 o = lu_object_top(h);
753 lu_object_print(arg->lsp_env, arg->lsp_cookie,
754 arg->lsp_printer, o);
756 lu_object_header_print(arg->lsp_env, arg->lsp_cookie,
757 arg->lsp_printer, h);
763 * Print all objects in \a s.
765 void lu_site_print(const struct lu_env *env, struct lu_site *s, void *cookie,
766 lu_printer_t printer)
768 struct lu_site_print_arg arg = {
769 .lsp_env = (struct lu_env *)env,
770 .lsp_cookie = cookie,
771 .lsp_printer = printer,
774 cfs_hash_for_each(s->ls_obj_hash, lu_site_obj_print, &arg);
776 EXPORT_SYMBOL(lu_site_print);
779 LU_CACHE_PERCENT_MAX = 50,
780 LU_CACHE_PERCENT_DEFAULT = 20
783 static unsigned int lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
784 CFS_MODULE_PARM(lu_cache_percent, "i", int, 0644,
785 "Percentage of memory to be used as lu_object cache");
788 * Return desired hash table order.
790 static int lu_htable_order(void)
792 unsigned long cache_size;
796 * Calculate hash table size, assuming that we want reasonable
797 * performance when 20% of total memory is occupied by cache of
800 * Size of lu_object is (arbitrary) taken as 1K (together with inode).
802 cache_size = cfs_num_physpages;
804 #if BITS_PER_LONG == 32
805 /* limit hashtable size for lowmem systems to low RAM */
806 if (cache_size > 1 << (30 - CFS_PAGE_SHIFT))
807 cache_size = 1 << (30 - CFS_PAGE_SHIFT) * 3 / 4;
810 /* clear off unreasonable cache setting. */
811 if (lu_cache_percent == 0 || lu_cache_percent > LU_CACHE_PERCENT_MAX) {
812 CWARN("obdclass: invalid lu_cache_percent: %u, it must be in"
813 " the range of (0, %u]. Will use default value: %u.\n",
814 lu_cache_percent, LU_CACHE_PERCENT_MAX,
815 LU_CACHE_PERCENT_DEFAULT);
817 lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
819 cache_size = cache_size / 100 * lu_cache_percent *
820 (CFS_PAGE_SIZE / 1024);
822 for (bits = 1; (1 << bits) < cache_size; ++bits) {
828 static unsigned lu_obj_hop_hash(cfs_hash_t *hs,
829 const void *key, unsigned mask)
831 struct lu_fid *fid = (struct lu_fid *)key;
834 hash = (fid_seq(fid) + fid_oid(fid)) & (CFS_HASH_NBKT(hs) - 1);
835 hash += fid_hash(fid, hs->hs_bkt_bits) << hs->hs_bkt_bits;
839 static void *lu_obj_hop_object(cfs_hlist_node_t *hnode)
841 return cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
844 static void *lu_obj_hop_key(cfs_hlist_node_t *hnode)
846 struct lu_object_header *h;
848 h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
852 static int lu_obj_hop_keycmp(const void *key, cfs_hlist_node_t *hnode)
854 struct lu_object_header *h;
856 h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
857 return lu_fid_eq(&h->loh_fid, (struct lu_fid *)key);
860 static void lu_obj_hop_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
862 struct lu_object_header *h;
864 h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
865 if (cfs_atomic_add_return(1, &h->loh_ref) == 1) {
866 struct lu_site_bkt_data *bkt;
869 cfs_hash_bd_get(hs, &h->loh_fid, &bd);
870 bkt = cfs_hash_bd_extra_get(hs, &bd);
875 static void lu_obj_hop_put_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
877 LBUG(); /* we should never called it */
880 cfs_hash_ops_t lu_site_hash_ops = {
881 .hs_hash = lu_obj_hop_hash,
882 .hs_key = lu_obj_hop_key,
883 .hs_keycmp = lu_obj_hop_keycmp,
884 .hs_object = lu_obj_hop_object,
885 .hs_get = lu_obj_hop_get,
886 .hs_put_locked = lu_obj_hop_put_locked,
890 * Initialize site \a s, with \a d as the top level device.
892 #define LU_SITE_BITS_MIN 12
893 #define LU_SITE_BITS_MAX 23
895 * total 128 buckets, we don't want too many buckets because:
896 * - consume too much memory
897 * - avoid unbalanced LRU list
899 #define LU_SITE_BKT_BITS 7
901 int lu_site_init(struct lu_site *s, struct lu_device *top)
903 struct lu_site_bkt_data *bkt;
909 memset(s, 0, sizeof *s);
910 bits = lu_htable_order();
911 for (bits = min(max(LU_SITE_BITS_MIN, bits), LU_SITE_BITS_MAX);
912 bits >= LU_SITE_BITS_MIN; bits--) {
913 s->ls_obj_hash = cfs_hash_create("lu_site", bits, bits,
914 bits - LU_SITE_BKT_BITS,
917 CFS_HASH_SPIN_BKTLOCK |
918 CFS_HASH_NO_ITEMREF |
920 CFS_HASH_ASSERT_EMPTY);
921 if (s->ls_obj_hash != NULL)
925 if (s->ls_obj_hash == NULL) {
926 CERROR("failed to create lu_site hash with bits: %d\n", bits);
930 cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
931 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
932 CFS_INIT_LIST_HEAD(&bkt->lsb_lru);
933 cfs_waitq_init(&bkt->lsb_marche_funebre);
936 s->ls_stats = lprocfs_alloc_stats(LU_SS_LAST_STAT, 0);
937 if (s->ls_stats == NULL) {
938 cfs_hash_putref(s->ls_obj_hash);
939 s->ls_obj_hash = NULL;
943 lprocfs_counter_init(s->ls_stats, LU_SS_CREATED,
944 0, "created", "created");
945 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_HIT,
946 0, "cache_hit", "cache_hit");
947 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_MISS,
948 0, "cache_miss", "cache_miss");
949 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_RACE,
950 0, "cache_race", "cache_race");
951 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_DEATH_RACE,
952 0, "cache_death_race", "cache_death_race");
953 lprocfs_counter_init(s->ls_stats, LU_SS_LRU_PURGED,
954 0, "lru_purged", "lru_purged");
956 CFS_INIT_LIST_HEAD(&s->ls_linkage);
960 lu_ref_add(&top->ld_reference, "site-top", s);
964 EXPORT_SYMBOL(lu_site_init);
967 * Finalize \a s and release its resources.
969 void lu_site_fini(struct lu_site *s)
971 cfs_down(&lu_sites_guard);
972 cfs_list_del_init(&s->ls_linkage);
973 cfs_up(&lu_sites_guard);
975 if (s->ls_obj_hash != NULL) {
976 cfs_hash_putref(s->ls_obj_hash);
977 s->ls_obj_hash = NULL;
980 if (s->ls_top_dev != NULL) {
981 s->ls_top_dev->ld_site = NULL;
982 lu_ref_del(&s->ls_top_dev->ld_reference, "site-top", s);
983 lu_device_put(s->ls_top_dev);
984 s->ls_top_dev = NULL;
987 if (s->ls_stats != NULL)
988 lprocfs_free_stats(&s->ls_stats);
990 EXPORT_SYMBOL(lu_site_fini);
993 * Called when initialization of stack for this site is completed.
995 int lu_site_init_finish(struct lu_site *s)
998 cfs_down(&lu_sites_guard);
999 result = lu_context_refill(&lu_shrink_env.le_ctx);
1001 cfs_list_add(&s->ls_linkage, &lu_sites);
1002 cfs_up(&lu_sites_guard);
1005 EXPORT_SYMBOL(lu_site_init_finish);
1008 * Acquire additional reference on device \a d
1010 void lu_device_get(struct lu_device *d)
1012 cfs_atomic_inc(&d->ld_ref);
1014 EXPORT_SYMBOL(lu_device_get);
1017 * Release reference on device \a d.
1019 void lu_device_put(struct lu_device *d)
1021 LASSERT(cfs_atomic_read(&d->ld_ref) > 0);
1022 cfs_atomic_dec(&d->ld_ref);
1024 EXPORT_SYMBOL(lu_device_put);
1027 * Initialize device \a d of type \a t.
1029 int lu_device_init(struct lu_device *d, struct lu_device_type *t)
1031 if (t->ldt_device_nr++ == 0 && t->ldt_ops->ldto_start != NULL)
1032 t->ldt_ops->ldto_start(t);
1033 memset(d, 0, sizeof *d);
1034 cfs_atomic_set(&d->ld_ref, 0);
1036 lu_ref_init(&d->ld_reference);
1039 EXPORT_SYMBOL(lu_device_init);
1042 * Finalize device \a d.
1044 void lu_device_fini(struct lu_device *d)
1046 struct lu_device_type *t;
1049 if (d->ld_obd != NULL) {
1050 d->ld_obd->obd_lu_dev = NULL;
1054 lu_ref_fini(&d->ld_reference);
1055 LASSERTF(cfs_atomic_read(&d->ld_ref) == 0,
1056 "Refcount is %u\n", cfs_atomic_read(&d->ld_ref));
1057 LASSERT(t->ldt_device_nr > 0);
1058 if (--t->ldt_device_nr == 0 && t->ldt_ops->ldto_stop != NULL)
1059 t->ldt_ops->ldto_stop(t);
1061 EXPORT_SYMBOL(lu_device_fini);
1064 * Initialize object \a o that is part of compound object \a h and was created
1067 int lu_object_init(struct lu_object *o,
1068 struct lu_object_header *h, struct lu_device *d)
1070 memset(o, 0, sizeof *o);
1074 o->lo_dev_ref = lu_ref_add(&d->ld_reference, "lu_object", o);
1075 CFS_INIT_LIST_HEAD(&o->lo_linkage);
1078 EXPORT_SYMBOL(lu_object_init);
1081 * Finalize object and release its resources.
1083 void lu_object_fini(struct lu_object *o)
1085 struct lu_device *dev = o->lo_dev;
1087 LASSERT(cfs_list_empty(&o->lo_linkage));
1090 lu_ref_del_at(&dev->ld_reference,
1091 o->lo_dev_ref , "lu_object", o);
1096 EXPORT_SYMBOL(lu_object_fini);
1099 * Add object \a o as first layer of compound object \a h
1101 * This is typically called by the ->ldo_object_alloc() method of top-level
1104 void lu_object_add_top(struct lu_object_header *h, struct lu_object *o)
1106 cfs_list_move(&o->lo_linkage, &h->loh_layers);
1108 EXPORT_SYMBOL(lu_object_add_top);
1111 * Add object \a o as a layer of compound object, going after \a before.
1113 * This is typically called by the ->ldo_object_alloc() method of \a
1116 void lu_object_add(struct lu_object *before, struct lu_object *o)
1118 cfs_list_move(&o->lo_linkage, &before->lo_linkage);
1120 EXPORT_SYMBOL(lu_object_add);
1123 * Initialize compound object.
1125 int lu_object_header_init(struct lu_object_header *h)
1127 memset(h, 0, sizeof *h);
1128 cfs_atomic_set(&h->loh_ref, 1);
1129 CFS_INIT_HLIST_NODE(&h->loh_hash);
1130 CFS_INIT_LIST_HEAD(&h->loh_lru);
1131 CFS_INIT_LIST_HEAD(&h->loh_layers);
1132 lu_ref_init(&h->loh_reference);
1135 EXPORT_SYMBOL(lu_object_header_init);
1138 * Finalize compound object.
1140 void lu_object_header_fini(struct lu_object_header *h)
1142 LASSERT(cfs_list_empty(&h->loh_layers));
1143 LASSERT(cfs_list_empty(&h->loh_lru));
1144 LASSERT(cfs_hlist_unhashed(&h->loh_hash));
1145 lu_ref_fini(&h->loh_reference);
1147 EXPORT_SYMBOL(lu_object_header_fini);
1150 * Given a compound object, find its slice, corresponding to the device type
1153 struct lu_object *lu_object_locate(struct lu_object_header *h,
1154 const struct lu_device_type *dtype)
1156 struct lu_object *o;
1158 cfs_list_for_each_entry(o, &h->loh_layers, lo_linkage) {
1159 if (o->lo_dev->ld_type == dtype)
1164 EXPORT_SYMBOL(lu_object_locate);
1169 * Finalize and free devices in the device stack.
1171 * Finalize device stack by purging object cache, and calling
1172 * lu_device_type_operations::ldto_device_fini() and
1173 * lu_device_type_operations::ldto_device_free() on all devices in the stack.
1175 void lu_stack_fini(const struct lu_env *env, struct lu_device *top)
1177 struct lu_site *site = top->ld_site;
1178 struct lu_device *scan;
1179 struct lu_device *next;
1181 lu_site_purge(env, site, ~0);
1182 for (scan = top; scan != NULL; scan = next) {
1183 next = scan->ld_type->ldt_ops->ldto_device_fini(env, scan);
1184 lu_ref_del(&scan->ld_reference, "lu-stack", &lu_site_init);
1185 lu_device_put(scan);
1189 lu_site_purge(env, site, ~0);
1191 if (!cfs_hash_is_empty(site->ls_obj_hash)) {
1193 * Uh-oh, objects still exist.
1195 static DECLARE_LU_CDEBUG_PRINT_INFO(cookie, D_ERROR);
1197 lu_site_print(env, site, &cookie, lu_cdebug_printer);
1200 for (scan = top; scan != NULL; scan = next) {
1201 const struct lu_device_type *ldt = scan->ld_type;
1202 struct obd_type *type;
1204 next = ldt->ldt_ops->ldto_device_free(env, scan);
1205 type = ldt->ldt_obd_type;
1208 class_put_type(type);
1212 EXPORT_SYMBOL(lu_stack_fini);
1216 * Maximal number of tld slots.
1218 LU_CONTEXT_KEY_NR = 32
1221 static struct lu_context_key *lu_keys[LU_CONTEXT_KEY_NR] = { NULL, };
1223 static cfs_spinlock_t lu_keys_guard = CFS_SPIN_LOCK_UNLOCKED;
1226 * Global counter incremented whenever key is registered, unregistered,
1227 * revived or quiesced. This is used to void unnecessary calls to
1228 * lu_context_refill(). No locking is provided, as initialization and shutdown
1229 * are supposed to be externally serialized.
1231 static unsigned key_set_version = 0;
1236 int lu_context_key_register(struct lu_context_key *key)
1241 LASSERT(key->lct_init != NULL);
1242 LASSERT(key->lct_fini != NULL);
1243 LASSERT(key->lct_tags != 0);
1244 LASSERT(key->lct_owner != NULL);
1247 cfs_spin_lock(&lu_keys_guard);
1248 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1249 if (lu_keys[i] == NULL) {
1251 cfs_atomic_set(&key->lct_used, 1);
1253 lu_ref_init(&key->lct_reference);
1259 cfs_spin_unlock(&lu_keys_guard);
1262 EXPORT_SYMBOL(lu_context_key_register);
1264 static void key_fini(struct lu_context *ctx, int index)
1266 if (ctx->lc_value != NULL && ctx->lc_value[index] != NULL) {
1267 struct lu_context_key *key;
1269 key = lu_keys[index];
1270 LASSERT(key != NULL);
1271 LASSERT(key->lct_fini != NULL);
1272 LASSERT(cfs_atomic_read(&key->lct_used) > 1);
1274 key->lct_fini(ctx, key, ctx->lc_value[index]);
1275 lu_ref_del(&key->lct_reference, "ctx", ctx);
1276 cfs_atomic_dec(&key->lct_used);
1277 LASSERT(key->lct_owner != NULL);
1278 if (!(ctx->lc_tags & LCT_NOREF)) {
1279 LASSERT(cfs_module_refcount(key->lct_owner) > 0);
1280 cfs_module_put(key->lct_owner);
1282 ctx->lc_value[index] = NULL;
1289 void lu_context_key_degister(struct lu_context_key *key)
1291 LASSERT(cfs_atomic_read(&key->lct_used) >= 1);
1292 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1294 lu_context_key_quiesce(key);
1297 cfs_spin_lock(&lu_keys_guard);
1298 key_fini(&lu_shrink_env.le_ctx, key->lct_index);
1299 if (lu_keys[key->lct_index]) {
1300 lu_keys[key->lct_index] = NULL;
1301 lu_ref_fini(&key->lct_reference);
1303 cfs_spin_unlock(&lu_keys_guard);
1305 LASSERTF(cfs_atomic_read(&key->lct_used) == 1,
1306 "key has instances: %d\n",
1307 cfs_atomic_read(&key->lct_used));
1309 EXPORT_SYMBOL(lu_context_key_degister);
1312 * Register a number of keys. This has to be called after all keys have been
1313 * initialized by a call to LU_CONTEXT_KEY_INIT().
1315 int lu_context_key_register_many(struct lu_context_key *k, ...)
1317 struct lu_context_key *key = k;
1323 result = lu_context_key_register(key);
1326 key = va_arg(args, struct lu_context_key *);
1327 } while (key != NULL);
1333 lu_context_key_degister(k);
1334 k = va_arg(args, struct lu_context_key *);
1341 EXPORT_SYMBOL(lu_context_key_register_many);
1344 * De-register a number of keys. This is a dual to
1345 * lu_context_key_register_many().
1347 void lu_context_key_degister_many(struct lu_context_key *k, ...)
1353 lu_context_key_degister(k);
1354 k = va_arg(args, struct lu_context_key*);
1355 } while (k != NULL);
1358 EXPORT_SYMBOL(lu_context_key_degister_many);
1361 * Revive a number of keys.
1363 void lu_context_key_revive_many(struct lu_context_key *k, ...)
1369 lu_context_key_revive(k);
1370 k = va_arg(args, struct lu_context_key*);
1371 } while (k != NULL);
1374 EXPORT_SYMBOL(lu_context_key_revive_many);
1377 * Quiescent a number of keys.
1379 void lu_context_key_quiesce_many(struct lu_context_key *k, ...)
1385 lu_context_key_quiesce(k);
1386 k = va_arg(args, struct lu_context_key*);
1387 } while (k != NULL);
1390 EXPORT_SYMBOL(lu_context_key_quiesce_many);
1393 * Return value associated with key \a key in context \a ctx.
1395 void *lu_context_key_get(const struct lu_context *ctx,
1396 const struct lu_context_key *key)
1398 LINVRNT(ctx->lc_state == LCS_ENTERED);
1399 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1400 LASSERT(lu_keys[key->lct_index] == key);
1401 return ctx->lc_value[key->lct_index];
1403 EXPORT_SYMBOL(lu_context_key_get);
1406 * List of remembered contexts. XXX document me.
1408 static CFS_LIST_HEAD(lu_context_remembered);
1411 * Destroy \a key in all remembered contexts. This is used to destroy key
1412 * values in "shared" contexts (like service threads), when a module owning
1413 * the key is about to be unloaded.
1415 void lu_context_key_quiesce(struct lu_context_key *key)
1417 struct lu_context *ctx;
1418 extern unsigned cl_env_cache_purge(unsigned nr);
1420 if (!(key->lct_tags & LCT_QUIESCENT)) {
1422 * XXX layering violation.
1424 cl_env_cache_purge(~0);
1425 key->lct_tags |= LCT_QUIESCENT;
1427 * XXX memory barrier has to go here.
1429 cfs_spin_lock(&lu_keys_guard);
1430 cfs_list_for_each_entry(ctx, &lu_context_remembered,
1432 key_fini(ctx, key->lct_index);
1433 cfs_spin_unlock(&lu_keys_guard);
1437 EXPORT_SYMBOL(lu_context_key_quiesce);
1439 void lu_context_key_revive(struct lu_context_key *key)
1441 key->lct_tags &= ~LCT_QUIESCENT;
1444 EXPORT_SYMBOL(lu_context_key_revive);
1446 static void keys_fini(struct lu_context *ctx)
1450 cfs_spin_lock(&lu_keys_guard);
1451 if (ctx->lc_value != NULL) {
1452 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i)
1454 OBD_FREE(ctx->lc_value,
1455 ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1456 ctx->lc_value = NULL;
1458 cfs_spin_unlock(&lu_keys_guard);
1461 static int keys_fill(struct lu_context *ctx)
1465 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1466 struct lu_context_key *key;
1469 if (ctx->lc_value[i] == NULL && key != NULL &&
1470 (key->lct_tags & ctx->lc_tags) &&
1472 * Don't create values for a LCT_QUIESCENT key, as this
1473 * will pin module owning a key.
1475 !(key->lct_tags & LCT_QUIESCENT)) {
1478 LINVRNT(key->lct_init != NULL);
1479 LINVRNT(key->lct_index == i);
1481 value = key->lct_init(ctx, key);
1482 if (unlikely(IS_ERR(value)))
1483 return PTR_ERR(value);
1485 LASSERT(key->lct_owner != NULL);
1486 if (!(ctx->lc_tags & LCT_NOREF))
1487 cfs_try_module_get(key->lct_owner);
1488 lu_ref_add_atomic(&key->lct_reference, "ctx", ctx);
1489 cfs_atomic_inc(&key->lct_used);
1491 * This is the only place in the code, where an
1492 * element of ctx->lc_value[] array is set to non-NULL
1495 ctx->lc_value[i] = value;
1496 if (key->lct_exit != NULL)
1497 ctx->lc_tags |= LCT_HAS_EXIT;
1499 ctx->lc_version = key_set_version;
1504 static int keys_init(struct lu_context *ctx)
1508 OBD_ALLOC(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1509 if (likely(ctx->lc_value != NULL))
1510 result = keys_fill(ctx);
1520 * Initialize context data-structure. Create values for all keys.
1522 int lu_context_init(struct lu_context *ctx, __u32 tags)
1524 memset(ctx, 0, sizeof *ctx);
1525 ctx->lc_state = LCS_INITIALIZED;
1526 ctx->lc_tags = tags;
1527 if (tags & LCT_REMEMBER) {
1528 cfs_spin_lock(&lu_keys_guard);
1529 cfs_list_add(&ctx->lc_remember, &lu_context_remembered);
1530 cfs_spin_unlock(&lu_keys_guard);
1532 CFS_INIT_LIST_HEAD(&ctx->lc_remember);
1533 return keys_init(ctx);
1535 EXPORT_SYMBOL(lu_context_init);
1538 * Finalize context data-structure. Destroy key values.
1540 void lu_context_fini(struct lu_context *ctx)
1542 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1543 ctx->lc_state = LCS_FINALIZED;
1545 cfs_spin_lock(&lu_keys_guard);
1546 cfs_list_del_init(&ctx->lc_remember);
1547 cfs_spin_unlock(&lu_keys_guard);
1549 EXPORT_SYMBOL(lu_context_fini);
1552 * Called before entering context.
1554 void lu_context_enter(struct lu_context *ctx)
1556 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1557 ctx->lc_state = LCS_ENTERED;
1559 EXPORT_SYMBOL(lu_context_enter);
1562 * Called after exiting from \a ctx
1564 void lu_context_exit(struct lu_context *ctx)
1568 LINVRNT(ctx->lc_state == LCS_ENTERED);
1569 ctx->lc_state = LCS_LEFT;
1570 if (ctx->lc_tags & LCT_HAS_EXIT && ctx->lc_value != NULL) {
1571 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1572 if (ctx->lc_value[i] != NULL) {
1573 struct lu_context_key *key;
1576 LASSERT(key != NULL);
1577 if (key->lct_exit != NULL)
1579 key, ctx->lc_value[i]);
1584 EXPORT_SYMBOL(lu_context_exit);
1587 * Allocate for context all missing keys that were registered after context
1590 int lu_context_refill(struct lu_context *ctx)
1592 LINVRNT(ctx->lc_value != NULL);
1593 return ctx->lc_version == key_set_version ? 0 : keys_fill(ctx);
1595 EXPORT_SYMBOL(lu_context_refill);
1597 int lu_env_init(struct lu_env *env, __u32 tags)
1602 result = lu_context_init(&env->le_ctx, tags);
1603 if (likely(result == 0))
1604 lu_context_enter(&env->le_ctx);
1607 EXPORT_SYMBOL(lu_env_init);
1609 void lu_env_fini(struct lu_env *env)
1611 lu_context_exit(&env->le_ctx);
1612 lu_context_fini(&env->le_ctx);
1615 EXPORT_SYMBOL(lu_env_fini);
1617 int lu_env_refill(struct lu_env *env)
1621 result = lu_context_refill(&env->le_ctx);
1622 if (result == 0 && env->le_ses != NULL)
1623 result = lu_context_refill(env->le_ses);
1626 EXPORT_SYMBOL(lu_env_refill);
1628 static struct cfs_shrinker *lu_site_shrinker = NULL;
1630 typedef struct lu_site_stats{
1631 unsigned lss_populated;
1632 unsigned lss_max_search;
1637 static void lu_site_stats_get(cfs_hash_t *hs,
1638 lu_site_stats_t *stats, int populated)
1643 cfs_hash_for_each_bucket(hs, &bd, i) {
1644 struct lu_site_bkt_data *bkt = cfs_hash_bd_extra_get(hs, &bd);
1645 cfs_hlist_head_t *hhead;
1647 cfs_hash_bd_lock(hs, &bd, 1);
1648 stats->lss_busy += bkt->lsb_busy;
1649 stats->lss_total += cfs_hash_bd_count_get(&bd);
1650 stats->lss_max_search = max((int)stats->lss_max_search,
1651 cfs_hash_bd_depmax_get(&bd));
1653 cfs_hash_bd_unlock(hs, &bd, 1);
1657 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
1658 if (!cfs_hlist_empty(hhead))
1659 stats->lss_populated++;
1661 cfs_hash_bd_unlock(hs, &bd, 1);
1667 static int lu_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
1669 lu_site_stats_t stats;
1671 struct lu_site *tmp;
1673 int remain = shrink_param(sc, nr_to_scan);
1674 CFS_LIST_HEAD(splice);
1677 if (!(shrink_param(sc, gfp_mask) & __GFP_FS))
1679 CDEBUG(D_INODE, "Shrink %d objects\n", remain);
1682 cfs_down(&lu_sites_guard);
1683 cfs_list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
1684 if (shrink_param(sc, nr_to_scan) != 0) {
1685 remain = lu_site_purge(&lu_shrink_env, s, remain);
1687 * Move just shrunk site to the tail of site list to
1688 * assure shrinking fairness.
1690 cfs_list_move_tail(&s->ls_linkage, &splice);
1693 memset(&stats, 0, sizeof(stats));
1694 lu_site_stats_get(s->ls_obj_hash, &stats, 0);
1695 cached += stats.lss_total - stats.lss_busy;
1696 if (shrink_param(sc, nr_to_scan) && remain <= 0)
1699 cfs_list_splice(&splice, lu_sites.prev);
1700 cfs_up(&lu_sites_guard);
1702 cached = (cached / 100) * sysctl_vfs_cache_pressure;
1703 if (shrink_param(sc, nr_to_scan) == 0)
1704 CDEBUG(D_INODE, "%d objects cached\n", cached);
1713 * Environment to be used in debugger, contains all tags.
1715 struct lu_env lu_debugging_env;
1718 * Debugging printer function using printk().
1720 int lu_printk_printer(const struct lu_env *env,
1721 void *unused, const char *format, ...)
1725 va_start(args, format);
1726 vprintk(format, args);
1731 void lu_debugging_setup(void)
1733 lu_env_init(&lu_debugging_env, ~0);
1736 void lu_context_keys_dump(void)
1740 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1741 struct lu_context_key *key;
1745 CERROR("[%d]: %p %x (%p,%p,%p) %d %d \"%s\"@%p\n",
1746 i, key, key->lct_tags,
1747 key->lct_init, key->lct_fini, key->lct_exit,
1748 key->lct_index, cfs_atomic_read(&key->lct_used),
1749 key->lct_owner ? key->lct_owner->name : "",
1751 lu_ref_print(&key->lct_reference);
1755 EXPORT_SYMBOL(lu_context_keys_dump);
1756 #else /* !__KERNEL__ */
1757 static int lu_cache_shrink(int nr, unsigned int gfp_mask)
1761 #endif /* __KERNEL__ */
1763 int cl_global_init(void);
1764 void cl_global_fini(void);
1765 int lu_ref_global_init(void);
1766 void lu_ref_global_fini(void);
1768 int dt_global_init(void);
1769 void dt_global_fini(void);
1771 int llo_global_init(void);
1772 void llo_global_fini(void);
1775 * Initialization of global lu_* data.
1777 int lu_global_init(void)
1781 CDEBUG(D_INFO, "Lustre LU module (%p).\n", &lu_keys);
1783 result = lu_ref_global_init();
1787 LU_CONTEXT_KEY_INIT(&lu_global_key);
1788 result = lu_context_key_register(&lu_global_key);
1792 * At this level, we don't know what tags are needed, so allocate them
1793 * conservatively. This should not be too bad, because this
1794 * environment is global.
1796 cfs_down(&lu_sites_guard);
1797 result = lu_env_init(&lu_shrink_env, LCT_SHRINKER);
1798 cfs_up(&lu_sites_guard);
1803 * seeks estimation: 3 seeks to read a record from oi, one to read
1804 * inode, one for ea. Unfortunately setting this high value results in
1805 * lu_object/inode cache consuming all the memory.
1807 lu_site_shrinker = cfs_set_shrinker(CFS_DEFAULT_SEEKS, lu_cache_shrink);
1808 if (lu_site_shrinker == NULL)
1811 result = lu_time_global_init();
1816 result = dt_global_init();
1820 result = llo_global_init();
1824 result = cl_global_init();
1831 * Dual to lu_global_init().
1833 void lu_global_fini(void)
1840 lu_time_global_fini();
1841 if (lu_site_shrinker != NULL) {
1842 cfs_remove_shrinker(lu_site_shrinker);
1843 lu_site_shrinker = NULL;
1846 lu_context_key_degister(&lu_global_key);
1849 * Tear shrinker environment down _after_ de-registering
1850 * lu_global_key, because the latter has a value in the former.
1852 cfs_down(&lu_sites_guard);
1853 lu_env_fini(&lu_shrink_env);
1854 cfs_up(&lu_sites_guard);
1856 lu_ref_global_fini();
1859 struct lu_buf LU_BUF_NULL = {
1863 EXPORT_SYMBOL(LU_BUF_NULL);
1865 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx)
1868 struct lprocfs_counter ret;
1870 lprocfs_stats_collect(stats, idx, &ret);
1871 return (__u32)ret.lc_count;
1878 * Output site statistical counters into a buffer. Suitable for
1879 * lprocfs_rd_*()-style functions.
1881 int lu_site_stats_print(const struct lu_site *s, char *page, int count)
1883 lu_site_stats_t stats;
1885 memset(&stats, 0, sizeof(stats));
1886 lu_site_stats_get(s->ls_obj_hash, &stats, 1);
1888 return snprintf(page, count, "%d/%d %d/%d %d %d %d %d %d %d %d\n",
1891 stats.lss_populated,
1892 CFS_HASH_NHLIST(s->ls_obj_hash),
1893 stats.lss_max_search,
1894 ls_stats_read(s->ls_stats, LU_SS_CREATED),
1895 ls_stats_read(s->ls_stats, LU_SS_CACHE_HIT),
1896 ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
1897 ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
1898 ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
1899 ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED));
1901 EXPORT_SYMBOL(lu_site_stats_print);
1903 const char *lu_time_names[LU_TIME_NR] = {
1904 [LU_TIME_FIND_LOOKUP] = "find_lookup",
1905 [LU_TIME_FIND_ALLOC] = "find_alloc",
1906 [LU_TIME_FIND_INSERT] = "find_insert"
1908 EXPORT_SYMBOL(lu_time_names);
1911 * Helper function to initialize a number of kmem slab caches at once.
1913 int lu_kmem_init(struct lu_kmem_descr *caches)
1916 struct lu_kmem_descr *iter = caches;
1918 for (result = 0; iter->ckd_cache != NULL; ++iter) {
1919 *iter->ckd_cache = cfs_mem_cache_create(iter->ckd_name,
1922 if (*iter->ckd_cache == NULL) {
1924 /* free all previously allocated caches */
1925 lu_kmem_fini(caches);
1931 EXPORT_SYMBOL(lu_kmem_init);
1934 * Helper function to finalize a number of kmem slab cached at once. Dual to
1937 void lu_kmem_fini(struct lu_kmem_descr *caches)
1941 for (; caches->ckd_cache != NULL; ++caches) {
1942 if (*caches->ckd_cache != NULL) {
1943 rc = cfs_mem_cache_destroy(*caches->ckd_cache);
1944 LASSERTF(rc == 0, "couldn't destroy %s slab\n",
1946 *caches->ckd_cache = NULL;
1950 EXPORT_SYMBOL(lu_kmem_fini);