4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, Whamcloud, Inc.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/obdclass/lu_object.c
39 * These are the only exported functions, they provide some generic
40 * infrastructure for managing object devices
42 * Author: Nikita Danilov <nikita.danilov@sun.com>
45 #define DEBUG_SUBSYSTEM S_CLASS
47 #include <libcfs/libcfs.h>
50 # include <linux/module.h>
54 #include <libcfs/libcfs_hash.h>
55 #include <obd_class.h>
56 #include <obd_support.h>
57 #include <lustre_disk.h>
58 #include <lustre_fid.h>
59 #include <lu_object.h>
60 #include <libcfs/list.h>
61 /* lu_time_global_{init,fini}() */
64 static void lu_object_free(const struct lu_env *env, struct lu_object *o);
67 * Decrease reference counter on object. If last reference is freed, return
68 * object to the cache, unless lu_object_is_dying(o) holds. In the latter
69 * case, free object immediately.
71 void lu_object_put(const struct lu_env *env, struct lu_object *o)
73 struct lu_site_bkt_data *bkt;
74 struct lu_object_header *top;
76 struct lu_object *orig;
80 site = o->lo_dev->ld_site;
83 cfs_hash_bd_get(site->ls_obj_hash, &top->loh_fid, &bd);
84 bkt = cfs_hash_bd_extra_get(site->ls_obj_hash, &bd);
86 if (!cfs_hash_bd_dec_and_lock(site->ls_obj_hash, &bd, &top->loh_ref)) {
87 if (lu_object_is_dying(top)) {
90 * somebody may be waiting for this, currently only
91 * used for cl_object, see cl_object_put_last().
93 cfs_waitq_broadcast(&bkt->lsb_marche_funebre);
98 LASSERT(bkt->lsb_busy > 0);
101 * When last reference is released, iterate over object
102 * layers, and notify them that object is no longer busy.
104 cfs_list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
105 if (o->lo_ops->loo_object_release != NULL)
106 o->lo_ops->loo_object_release(env, o);
109 if (!lu_object_is_dying(top)) {
110 LASSERT(cfs_list_empty(&top->loh_lru));
111 cfs_list_add_tail(&top->loh_lru, &bkt->lsb_lru);
112 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
117 * If object is dying (will not be cached), removed it
118 * from hash table and LRU.
120 * This is done with hash table and LRU lists locked. As the only
121 * way to acquire first reference to previously unreferenced
122 * object is through hash-table lookup (lu_object_find()),
123 * or LRU scanning (lu_site_purge()), that are done under hash-table
124 * and LRU lock, no race with concurrent object lookup is possible
125 * and we can safely destroy object below.
127 cfs_hash_bd_del_locked(site->ls_obj_hash, &bd, &top->loh_hash);
128 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
130 * Object was already removed from hash and lru above, can
133 lu_object_free(env, orig);
135 EXPORT_SYMBOL(lu_object_put);
138 * Put object and don't keep in cache. This is temporary solution for
139 * multi-site objects when its layering is not constant.
141 void lu_object_put_nocache(const struct lu_env *env, struct lu_object *o)
143 cfs_set_bit(LU_OBJECT_HEARD_BANSHEE,
144 &o->lo_header->loh_flags);
145 return lu_object_put(env, o);
147 EXPORT_SYMBOL(lu_object_put_nocache);
150 * Allocate new object.
152 * This follows object creation protocol, described in the comment within
153 * struct lu_device_operations definition.
155 static struct lu_object *lu_object_alloc(const struct lu_env *env,
156 struct lu_device *dev,
157 const struct lu_fid *f,
158 const struct lu_object_conf *conf)
160 struct lu_object *scan;
161 struct lu_object *top;
168 * Create top-level object slice. This will also create
171 top = dev->ld_ops->ldo_object_alloc(env, NULL, dev);
173 RETURN(ERR_PTR(-ENOMEM));
175 * This is the only place where object fid is assigned. It's constant
178 LASSERT(fid_is_igif(f) || fid_ver(f) == 0);
179 top->lo_header->loh_fid = *f;
180 layers = &top->lo_header->loh_layers;
183 * Call ->loo_object_init() repeatedly, until no more new
184 * object slices are created.
187 cfs_list_for_each_entry(scan, layers, lo_linkage) {
188 if (scan->lo_flags & LU_OBJECT_ALLOCATED)
191 scan->lo_header = top->lo_header;
192 result = scan->lo_ops->loo_object_init(env, scan, conf);
194 lu_object_free(env, top);
195 RETURN(ERR_PTR(result));
197 scan->lo_flags |= LU_OBJECT_ALLOCATED;
201 cfs_list_for_each_entry_reverse(scan, layers, lo_linkage) {
202 if (scan->lo_ops->loo_object_start != NULL) {
203 result = scan->lo_ops->loo_object_start(env, scan);
205 lu_object_free(env, top);
206 RETURN(ERR_PTR(result));
211 lprocfs_counter_incr(dev->ld_site->ls_stats, LU_SS_CREATED);
218 static void lu_object_free(const struct lu_env *env, struct lu_object *o)
220 struct lu_site_bkt_data *bkt;
221 struct lu_site *site;
222 struct lu_object *scan;
226 site = o->lo_dev->ld_site;
227 layers = &o->lo_header->loh_layers;
228 bkt = lu_site_bkt_from_fid(site, &o->lo_header->loh_fid);
230 * First call ->loo_object_delete() method to release all resources.
232 cfs_list_for_each_entry_reverse(scan, layers, lo_linkage) {
233 if (scan->lo_ops->loo_object_delete != NULL)
234 scan->lo_ops->loo_object_delete(env, scan);
238 * Then, splice object layers into stand-alone list, and call
239 * ->loo_object_free() on all layers to free memory. Splice is
240 * necessary, because lu_object_header is freed together with the
243 CFS_INIT_LIST_HEAD(&splice);
244 cfs_list_splice_init(layers, &splice);
245 while (!cfs_list_empty(&splice)) {
247 * Free layers in bottom-to-top order, so that object header
248 * lives as long as possible and ->loo_object_free() methods
249 * can look at its contents.
251 o = container_of0(splice.prev, struct lu_object, lo_linkage);
252 cfs_list_del_init(&o->lo_linkage);
253 LASSERT(o->lo_ops->loo_object_free != NULL);
254 o->lo_ops->loo_object_free(env, o);
257 if (cfs_waitq_active(&bkt->lsb_marche_funebre))
258 cfs_waitq_broadcast(&bkt->lsb_marche_funebre);
262 * Free \a nr objects from the cold end of the site LRU list.
264 int lu_site_purge(const struct lu_env *env, struct lu_site *s, int nr)
266 struct lu_object_header *h;
267 struct lu_object_header *temp;
268 struct lu_site_bkt_data *bkt;
278 CFS_INIT_LIST_HEAD(&dispose);
280 * Under LRU list lock, scan LRU list and move unreferenced objects to
281 * the dispose list, removing them from LRU and hash table.
283 start = s->ls_purge_start;
284 bnr = (nr == ~0) ? -1 : nr / CFS_HASH_NBKT(s->ls_obj_hash) + 1;
287 cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
291 cfs_hash_bd_lock(s->ls_obj_hash, &bd, 1);
292 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
294 cfs_list_for_each_entry_safe(h, temp, &bkt->lsb_lru, loh_lru) {
295 LASSERT(cfs_atomic_read(&h->loh_ref) == 0);
297 cfs_hash_bd_get(s->ls_obj_hash, &h->loh_fid, &bd2);
298 LASSERT(bd.bd_bucket == bd2.bd_bucket);
300 cfs_hash_bd_del_locked(s->ls_obj_hash,
302 cfs_list_move(&h->loh_lru, &dispose);
306 if (nr != ~0 && --nr == 0)
309 if (count > 0 && --count == 0)
313 cfs_hash_bd_unlock(s->ls_obj_hash, &bd, 1);
316 * Free everything on the dispose list. This is safe against
317 * races due to the reasons described in lu_object_put().
319 while (!cfs_list_empty(&dispose)) {
320 h = container_of0(dispose.next,
321 struct lu_object_header, loh_lru);
322 cfs_list_del_init(&h->loh_lru);
323 lu_object_free(env, lu_object_top(h));
324 lprocfs_counter_incr(s->ls_stats, LU_SS_LRU_PURGED);
331 if (nr != 0 && did_sth && start != 0) {
332 start = 0; /* restart from the first bucket */
335 /* race on s->ls_purge_start, but nobody cares */
336 s->ls_purge_start = i % CFS_HASH_NBKT(s->ls_obj_hash);
340 EXPORT_SYMBOL(lu_site_purge);
345 * Code below has to jump through certain loops to output object description
346 * into libcfs_debug_msg-based log. The problem is that lu_object_print()
347 * composes object description from strings that are parts of _lines_ of
348 * output (i.e., strings that are not terminated by newline). This doesn't fit
349 * very well into libcfs_debug_msg() interface that assumes that each message
350 * supplied to it is a self-contained output line.
352 * To work around this, strings are collected in a temporary buffer
353 * (implemented as a value of lu_cdebug_key key), until terminating newline
354 * character is detected.
362 * XXX overflow is not handled correctly.
367 struct lu_cdebug_data {
371 char lck_area[LU_CDEBUG_LINE];
374 /* context key constructor/destructor: lu_global_key_init, lu_global_key_fini */
375 LU_KEY_INIT_FINI(lu_global, struct lu_cdebug_data);
378 * Key, holding temporary buffer. This key is registered very early by
381 struct lu_context_key lu_global_key = {
382 .lct_tags = LCT_MD_THREAD | LCT_DT_THREAD |
383 LCT_MG_THREAD | LCT_CL_THREAD,
384 .lct_init = lu_global_key_init,
385 .lct_fini = lu_global_key_fini
389 * Printer function emitting messages through libcfs_debug_msg().
391 int lu_cdebug_printer(const struct lu_env *env,
392 void *cookie, const char *format, ...)
394 struct libcfs_debug_msg_data *msgdata = cookie;
395 struct lu_cdebug_data *key;
400 va_start(args, format);
402 key = lu_context_key_get(&env->le_ctx, &lu_global_key);
403 LASSERT(key != NULL);
405 used = strlen(key->lck_area);
406 complete = format[strlen(format) - 1] == '\n';
408 * Append new chunk to the buffer.
410 vsnprintf(key->lck_area + used,
411 ARRAY_SIZE(key->lck_area) - used, format, args);
413 if (cfs_cdebug_show(msgdata->msg_mask, msgdata->msg_subsys))
414 libcfs_debug_msg(msgdata, "%s", key->lck_area);
415 key->lck_area[0] = 0;
420 EXPORT_SYMBOL(lu_cdebug_printer);
423 * Print object header.
425 void lu_object_header_print(const struct lu_env *env, void *cookie,
426 lu_printer_t printer,
427 const struct lu_object_header *hdr)
429 (*printer)(env, cookie, "header@%p[%#lx, %d, "DFID"%s%s%s]",
430 hdr, hdr->loh_flags, cfs_atomic_read(&hdr->loh_ref),
432 cfs_hlist_unhashed(&hdr->loh_hash) ? "" : " hash",
433 cfs_list_empty((cfs_list_t *)&hdr->loh_lru) ? \
435 hdr->loh_attr & LOHA_EXISTS ? " exist":"");
437 EXPORT_SYMBOL(lu_object_header_print);
440 * Print human readable representation of the \a o to the \a printer.
442 void lu_object_print(const struct lu_env *env, void *cookie,
443 lu_printer_t printer, const struct lu_object *o)
445 static const char ruler[] = "........................................";
446 struct lu_object_header *top;
450 lu_object_header_print(env, cookie, printer, top);
451 (*printer)(env, cookie, "{ \n");
452 cfs_list_for_each_entry(o, &top->loh_layers, lo_linkage) {
453 depth = o->lo_depth + 4;
456 * print `.' \a depth times followed by type name and address
458 (*printer)(env, cookie, "%*.*s%s@%p", depth, depth, ruler,
459 o->lo_dev->ld_type->ldt_name, o);
460 if (o->lo_ops->loo_object_print != NULL)
461 o->lo_ops->loo_object_print(env, cookie, printer, o);
462 (*printer)(env, cookie, "\n");
464 (*printer)(env, cookie, "} header@%p\n", top);
466 EXPORT_SYMBOL(lu_object_print);
469 * Check object consistency.
471 int lu_object_invariant(const struct lu_object *o)
473 struct lu_object_header *top;
476 cfs_list_for_each_entry(o, &top->loh_layers, lo_linkage) {
477 if (o->lo_ops->loo_object_invariant != NULL &&
478 !o->lo_ops->loo_object_invariant(o))
483 EXPORT_SYMBOL(lu_object_invariant);
485 static struct lu_object *htable_lookup(struct lu_site *s,
487 const struct lu_fid *f,
488 cfs_waitlink_t *waiter,
491 struct lu_site_bkt_data *bkt;
492 struct lu_object_header *h;
493 cfs_hlist_node_t *hnode;
494 __u64 ver = cfs_hash_bd_version_get(bd);
500 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, bd);
501 /* cfs_hash_bd_lookup_intent is a somehow "internal" function
502 * of cfs_hash, but we don't want refcount on object right now */
503 hnode = cfs_hash_bd_lookup_locked(s->ls_obj_hash, bd, (void *)f);
505 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS);
509 h = container_of0(hnode, struct lu_object_header, loh_hash);
510 if (likely(!lu_object_is_dying(h))) {
511 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_HIT);
512 cfs_list_del_init(&h->loh_lru);
513 return lu_object_top(h);
517 * Lookup found an object being destroyed this object cannot be
518 * returned (to assure that references to dying objects are eventually
519 * drained), and moreover, lookup has to wait until object is freed.
521 cfs_atomic_dec(&h->loh_ref);
523 cfs_waitlink_init(waiter);
524 cfs_waitq_add(&bkt->lsb_marche_funebre, waiter);
525 cfs_set_current_state(CFS_TASK_UNINT);
526 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_DEATH_RACE);
527 return ERR_PTR(-EAGAIN);
531 * Search cache for an object with the fid \a f. If such object is found,
532 * return it. Otherwise, create new object, insert it into cache and return
533 * it. In any case, additional reference is acquired on the returned object.
535 struct lu_object *lu_object_find(const struct lu_env *env,
536 struct lu_device *dev, const struct lu_fid *f,
537 const struct lu_object_conf *conf)
539 return lu_object_find_at(env, dev->ld_site->ls_top_dev, f, conf);
541 EXPORT_SYMBOL(lu_object_find);
543 static struct lu_object *lu_object_new(const struct lu_env *env,
544 struct lu_device *dev,
545 const struct lu_fid *f,
546 const struct lu_object_conf *conf)
551 struct lu_site_bkt_data *bkt;
553 o = lu_object_alloc(env, dev, f, conf);
554 if (unlikely(IS_ERR(o)))
557 hs = dev->ld_site->ls_obj_hash;
558 cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
559 bkt = cfs_hash_bd_extra_get(hs, &bd);
560 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
562 cfs_hash_bd_unlock(hs, &bd, 1);
567 * Core logic of lu_object_find*() functions.
569 static struct lu_object *lu_object_find_try(const struct lu_env *env,
570 struct lu_device *dev,
571 const struct lu_fid *f,
572 const struct lu_object_conf *conf,
573 cfs_waitlink_t *waiter)
576 struct lu_object *shadow;
583 * This uses standard index maintenance protocol:
585 * - search index under lock, and return object if found;
586 * - otherwise, unlock index, allocate new object;
587 * - lock index and search again;
588 * - if nothing is found (usual case), insert newly created
590 * - otherwise (race: other thread inserted object), free
591 * object just allocated.
595 * For "LOC_F_NEW" case, we are sure the object is new established.
596 * It is unnecessary to perform lookup-alloc-lookup-insert, instead,
597 * just alloc and insert directly.
599 * If dying object is found during index search, add @waiter to the
600 * site wait-queue and return ERR_PTR(-EAGAIN).
602 if (conf != NULL && conf->loc_flags & LOC_F_NEW)
603 return lu_object_new(env, dev, f, conf);
607 cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
608 o = htable_lookup(s, &bd, f, waiter, &version);
609 cfs_hash_bd_unlock(hs, &bd, 1);
614 * Allocate new object. This may result in rather complicated
615 * operations, including fld queries, inode loading, etc.
617 o = lu_object_alloc(env, dev, f, conf);
618 if (unlikely(IS_ERR(o)))
621 LASSERT(lu_fid_eq(lu_object_fid(o), f));
623 cfs_hash_bd_lock(hs, &bd, 1);
625 shadow = htable_lookup(s, &bd, f, waiter, &version);
626 if (likely(shadow == NULL)) {
627 struct lu_site_bkt_data *bkt;
629 bkt = cfs_hash_bd_extra_get(hs, &bd);
630 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
632 cfs_hash_bd_unlock(hs, &bd, 1);
636 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_RACE);
637 cfs_hash_bd_unlock(hs, &bd, 1);
638 lu_object_free(env, o);
643 * Much like lu_object_find(), but top level device of object is specifically
644 * \a dev rather than top level device of the site. This interface allows
645 * objects of different "stacking" to be created within the same site.
647 struct lu_object *lu_object_find_at(const struct lu_env *env,
648 struct lu_device *dev,
649 const struct lu_fid *f,
650 const struct lu_object_conf *conf)
652 struct lu_site_bkt_data *bkt;
653 struct lu_object *obj;
657 obj = lu_object_find_try(env, dev, f, conf, &wait);
658 if (obj != ERR_PTR(-EAGAIN))
661 * lu_object_find_try() already added waiter into the
664 cfs_waitq_wait(&wait, CFS_TASK_UNINT);
665 bkt = lu_site_bkt_from_fid(dev->ld_site, (void *)f);
666 cfs_waitq_del(&bkt->lsb_marche_funebre, &wait);
669 EXPORT_SYMBOL(lu_object_find_at);
672 * Find object with given fid, and return its slice belonging to given device.
674 struct lu_object *lu_object_find_slice(const struct lu_env *env,
675 struct lu_device *dev,
676 const struct lu_fid *f,
677 const struct lu_object_conf *conf)
679 struct lu_object *top;
680 struct lu_object *obj;
682 top = lu_object_find(env, dev, f, conf);
684 obj = lu_object_locate(top->lo_header, dev->ld_type);
686 lu_object_put(env, top);
691 EXPORT_SYMBOL(lu_object_find_slice);
694 * Global list of all device types.
696 static CFS_LIST_HEAD(lu_device_types);
698 int lu_device_type_init(struct lu_device_type *ldt)
702 CFS_INIT_LIST_HEAD(&ldt->ldt_linkage);
703 result = ldt->ldt_ops->ldto_init(ldt);
705 cfs_list_add(&ldt->ldt_linkage, &lu_device_types);
708 EXPORT_SYMBOL(lu_device_type_init);
710 void lu_device_type_fini(struct lu_device_type *ldt)
712 cfs_list_del_init(&ldt->ldt_linkage);
713 ldt->ldt_ops->ldto_fini(ldt);
715 EXPORT_SYMBOL(lu_device_type_fini);
717 void lu_types_stop(void)
719 struct lu_device_type *ldt;
721 cfs_list_for_each_entry(ldt, &lu_device_types, ldt_linkage) {
722 if (ldt->ldt_device_nr == 0)
723 ldt->ldt_ops->ldto_stop(ldt);
726 EXPORT_SYMBOL(lu_types_stop);
729 * Global list of all sites on this node
731 static CFS_LIST_HEAD(lu_sites);
732 static CFS_DEFINE_MUTEX(lu_sites_guard);
735 * Global environment used by site shrinker.
737 static struct lu_env lu_shrink_env;
739 struct lu_site_print_arg {
740 struct lu_env *lsp_env;
742 lu_printer_t lsp_printer;
746 lu_site_obj_print(cfs_hash_t *hs, cfs_hash_bd_t *bd,
747 cfs_hlist_node_t *hnode, void *data)
749 struct lu_site_print_arg *arg = (struct lu_site_print_arg *)data;
750 struct lu_object_header *h;
752 h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
753 if (!cfs_list_empty(&h->loh_layers)) {
754 const struct lu_object *o;
756 o = lu_object_top(h);
757 lu_object_print(arg->lsp_env, arg->lsp_cookie,
758 arg->lsp_printer, o);
760 lu_object_header_print(arg->lsp_env, arg->lsp_cookie,
761 arg->lsp_printer, h);
767 * Print all objects in \a s.
769 void lu_site_print(const struct lu_env *env, struct lu_site *s, void *cookie,
770 lu_printer_t printer)
772 struct lu_site_print_arg arg = {
773 .lsp_env = (struct lu_env *)env,
774 .lsp_cookie = cookie,
775 .lsp_printer = printer,
778 cfs_hash_for_each(s->ls_obj_hash, lu_site_obj_print, &arg);
780 EXPORT_SYMBOL(lu_site_print);
783 LU_CACHE_PERCENT_MAX = 50,
784 LU_CACHE_PERCENT_DEFAULT = 20
787 static unsigned int lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
788 CFS_MODULE_PARM(lu_cache_percent, "i", int, 0644,
789 "Percentage of memory to be used as lu_object cache");
792 * Return desired hash table order.
794 static int lu_htable_order(void)
796 unsigned long cache_size;
800 * Calculate hash table size, assuming that we want reasonable
801 * performance when 20% of total memory is occupied by cache of
804 * Size of lu_object is (arbitrary) taken as 1K (together with inode).
806 cache_size = cfs_num_physpages;
808 #if BITS_PER_LONG == 32
809 /* limit hashtable size for lowmem systems to low RAM */
810 if (cache_size > 1 << (30 - CFS_PAGE_SHIFT))
811 cache_size = 1 << (30 - CFS_PAGE_SHIFT) * 3 / 4;
814 /* clear off unreasonable cache setting. */
815 if (lu_cache_percent == 0 || lu_cache_percent > LU_CACHE_PERCENT_MAX) {
816 CWARN("obdclass: invalid lu_cache_percent: %u, it must be in"
817 " the range of (0, %u]. Will use default value: %u.\n",
818 lu_cache_percent, LU_CACHE_PERCENT_MAX,
819 LU_CACHE_PERCENT_DEFAULT);
821 lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
823 cache_size = cache_size / 100 * lu_cache_percent *
824 (CFS_PAGE_SIZE / 1024);
826 for (bits = 1; (1 << bits) < cache_size; ++bits) {
832 static unsigned lu_obj_hop_hash(cfs_hash_t *hs,
833 const void *key, unsigned mask)
835 struct lu_fid *fid = (struct lu_fid *)key;
838 hash = fid_flatten32(fid);
839 hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
840 hash = cfs_hash_long(hash, hs->hs_bkt_bits);
842 /* give me another random factor */
843 hash -= cfs_hash_long((unsigned long)hs, fid_oid(fid) % 11 + 3);
845 hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
846 hash |= (fid_seq(fid) + fid_oid(fid)) & (CFS_HASH_NBKT(hs) - 1);
851 static void *lu_obj_hop_object(cfs_hlist_node_t *hnode)
853 return cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
856 static void *lu_obj_hop_key(cfs_hlist_node_t *hnode)
858 struct lu_object_header *h;
860 h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
864 static int lu_obj_hop_keycmp(const void *key, cfs_hlist_node_t *hnode)
866 struct lu_object_header *h;
868 h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
869 return lu_fid_eq(&h->loh_fid, (struct lu_fid *)key);
872 static void lu_obj_hop_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
874 struct lu_object_header *h;
876 h = cfs_hlist_entry(hnode, struct lu_object_header, loh_hash);
877 if (cfs_atomic_add_return(1, &h->loh_ref) == 1) {
878 struct lu_site_bkt_data *bkt;
881 cfs_hash_bd_get(hs, &h->loh_fid, &bd);
882 bkt = cfs_hash_bd_extra_get(hs, &bd);
887 static void lu_obj_hop_put_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
889 LBUG(); /* we should never called it */
892 cfs_hash_ops_t lu_site_hash_ops = {
893 .hs_hash = lu_obj_hop_hash,
894 .hs_key = lu_obj_hop_key,
895 .hs_keycmp = lu_obj_hop_keycmp,
896 .hs_object = lu_obj_hop_object,
897 .hs_get = lu_obj_hop_get,
898 .hs_put_locked = lu_obj_hop_put_locked,
901 void lu_dev_add_linkage(struct lu_site *s, struct lu_device *d)
903 cfs_spin_lock(&s->ls_ld_lock);
904 if (cfs_list_empty(&d->ld_linkage))
905 cfs_list_add(&d->ld_linkage, &s->ls_ld_linkage);
906 cfs_spin_unlock(&s->ls_ld_lock);
908 EXPORT_SYMBOL(lu_dev_add_linkage);
910 void lu_dev_del_linkage(struct lu_site *s, struct lu_device *d)
912 cfs_spin_lock(&s->ls_ld_lock);
913 cfs_list_del_init(&d->ld_linkage);
914 cfs_spin_unlock(&s->ls_ld_lock);
916 EXPORT_SYMBOL(lu_dev_del_linkage);
919 * Initialize site \a s, with \a d as the top level device.
921 #define LU_SITE_BITS_MIN 12
922 #define LU_SITE_BITS_MAX 24
924 * total 256 buckets, we don't want too many buckets because:
925 * - consume too much memory
926 * - avoid unbalanced LRU list
928 #define LU_SITE_BKT_BITS 8
930 int lu_site_init(struct lu_site *s, struct lu_device *top)
932 struct lu_site_bkt_data *bkt;
939 memset(s, 0, sizeof *s);
940 bits = lu_htable_order();
941 snprintf(name, 16, "lu_site_%s", top->ld_type->ldt_name);
942 for (bits = min(max(LU_SITE_BITS_MIN, bits), LU_SITE_BITS_MAX);
943 bits >= LU_SITE_BITS_MIN; bits--) {
944 s->ls_obj_hash = cfs_hash_create(name, bits, bits,
945 bits - LU_SITE_BKT_BITS,
948 CFS_HASH_SPIN_BKTLOCK |
949 CFS_HASH_NO_ITEMREF |
951 CFS_HASH_ASSERT_EMPTY);
952 if (s->ls_obj_hash != NULL)
956 if (s->ls_obj_hash == NULL) {
957 CERROR("failed to create lu_site hash with bits: %d\n", bits);
961 cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
962 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
963 CFS_INIT_LIST_HEAD(&bkt->lsb_lru);
964 cfs_waitq_init(&bkt->lsb_marche_funebre);
967 s->ls_stats = lprocfs_alloc_stats(LU_SS_LAST_STAT, 0);
968 if (s->ls_stats == NULL) {
969 cfs_hash_putref(s->ls_obj_hash);
970 s->ls_obj_hash = NULL;
974 lprocfs_counter_init(s->ls_stats, LU_SS_CREATED,
975 0, "created", "created");
976 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_HIT,
977 0, "cache_hit", "cache_hit");
978 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_MISS,
979 0, "cache_miss", "cache_miss");
980 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_RACE,
981 0, "cache_race", "cache_race");
982 lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_DEATH_RACE,
983 0, "cache_death_race", "cache_death_race");
984 lprocfs_counter_init(s->ls_stats, LU_SS_LRU_PURGED,
985 0, "lru_purged", "lru_purged");
987 CFS_INIT_LIST_HEAD(&s->ls_linkage);
991 lu_ref_add(&top->ld_reference, "site-top", s);
993 CFS_INIT_LIST_HEAD(&s->ls_ld_linkage);
994 cfs_spin_lock_init(&s->ls_ld_lock);
996 lu_dev_add_linkage(s, top);
1000 EXPORT_SYMBOL(lu_site_init);
1003 * Finalize \a s and release its resources.
1005 void lu_site_fini(struct lu_site *s)
1007 cfs_mutex_lock(&lu_sites_guard);
1008 cfs_list_del_init(&s->ls_linkage);
1009 cfs_mutex_unlock(&lu_sites_guard);
1011 if (s->ls_obj_hash != NULL) {
1012 cfs_hash_putref(s->ls_obj_hash);
1013 s->ls_obj_hash = NULL;
1016 if (s->ls_top_dev != NULL) {
1017 s->ls_top_dev->ld_site = NULL;
1018 lu_ref_del(&s->ls_top_dev->ld_reference, "site-top", s);
1019 lu_device_put(s->ls_top_dev);
1020 s->ls_top_dev = NULL;
1023 if (s->ls_stats != NULL)
1024 lprocfs_free_stats(&s->ls_stats);
1026 EXPORT_SYMBOL(lu_site_fini);
1029 * Called when initialization of stack for this site is completed.
1031 int lu_site_init_finish(struct lu_site *s)
1034 cfs_mutex_lock(&lu_sites_guard);
1035 result = lu_context_refill(&lu_shrink_env.le_ctx);
1037 cfs_list_add(&s->ls_linkage, &lu_sites);
1038 cfs_mutex_unlock(&lu_sites_guard);
1041 EXPORT_SYMBOL(lu_site_init_finish);
1044 * Acquire additional reference on device \a d
1046 void lu_device_get(struct lu_device *d)
1048 cfs_atomic_inc(&d->ld_ref);
1050 EXPORT_SYMBOL(lu_device_get);
1053 * Release reference on device \a d.
1055 void lu_device_put(struct lu_device *d)
1057 LASSERT(cfs_atomic_read(&d->ld_ref) > 0);
1058 cfs_atomic_dec(&d->ld_ref);
1060 EXPORT_SYMBOL(lu_device_put);
1063 * Initialize device \a d of type \a t.
1065 int lu_device_init(struct lu_device *d, struct lu_device_type *t)
1067 if (t->ldt_device_nr++ == 0 && t->ldt_ops->ldto_start != NULL)
1068 t->ldt_ops->ldto_start(t);
1069 memset(d, 0, sizeof *d);
1070 cfs_atomic_set(&d->ld_ref, 0);
1072 lu_ref_init(&d->ld_reference);
1073 CFS_INIT_LIST_HEAD(&d->ld_linkage);
1076 EXPORT_SYMBOL(lu_device_init);
1079 * Finalize device \a d.
1081 void lu_device_fini(struct lu_device *d)
1083 struct lu_device_type *t;
1086 if (d->ld_obd != NULL) {
1087 d->ld_obd->obd_lu_dev = NULL;
1091 lu_ref_fini(&d->ld_reference);
1092 LASSERTF(cfs_atomic_read(&d->ld_ref) == 0,
1093 "Refcount is %u\n", cfs_atomic_read(&d->ld_ref));
1094 LASSERT(t->ldt_device_nr > 0);
1095 if (--t->ldt_device_nr == 0 && t->ldt_ops->ldto_stop != NULL)
1096 t->ldt_ops->ldto_stop(t);
1098 EXPORT_SYMBOL(lu_device_fini);
1101 * Initialize object \a o that is part of compound object \a h and was created
1104 int lu_object_init(struct lu_object *o,
1105 struct lu_object_header *h, struct lu_device *d)
1107 memset(o, 0, sizeof *o);
1111 o->lo_dev_ref = lu_ref_add(&d->ld_reference, "lu_object", o);
1112 CFS_INIT_LIST_HEAD(&o->lo_linkage);
1115 EXPORT_SYMBOL(lu_object_init);
1118 * Finalize object and release its resources.
1120 void lu_object_fini(struct lu_object *o)
1122 struct lu_device *dev = o->lo_dev;
1124 LASSERT(cfs_list_empty(&o->lo_linkage));
1127 lu_ref_del_at(&dev->ld_reference,
1128 o->lo_dev_ref , "lu_object", o);
1133 EXPORT_SYMBOL(lu_object_fini);
1136 * Add object \a o as first layer of compound object \a h
1138 * This is typically called by the ->ldo_object_alloc() method of top-level
1141 void lu_object_add_top(struct lu_object_header *h, struct lu_object *o)
1143 cfs_list_move(&o->lo_linkage, &h->loh_layers);
1145 EXPORT_SYMBOL(lu_object_add_top);
1148 * Add object \a o as a layer of compound object, going after \a before.
1150 * This is typically called by the ->ldo_object_alloc() method of \a
1153 void lu_object_add(struct lu_object *before, struct lu_object *o)
1155 cfs_list_move(&o->lo_linkage, &before->lo_linkage);
1157 EXPORT_SYMBOL(lu_object_add);
1160 * Initialize compound object.
1162 int lu_object_header_init(struct lu_object_header *h)
1164 memset(h, 0, sizeof *h);
1165 cfs_atomic_set(&h->loh_ref, 1);
1166 CFS_INIT_HLIST_NODE(&h->loh_hash);
1167 CFS_INIT_LIST_HEAD(&h->loh_lru);
1168 CFS_INIT_LIST_HEAD(&h->loh_layers);
1169 lu_ref_init(&h->loh_reference);
1172 EXPORT_SYMBOL(lu_object_header_init);
1175 * Finalize compound object.
1177 void lu_object_header_fini(struct lu_object_header *h)
1179 LASSERT(cfs_list_empty(&h->loh_layers));
1180 LASSERT(cfs_list_empty(&h->loh_lru));
1181 LASSERT(cfs_hlist_unhashed(&h->loh_hash));
1182 lu_ref_fini(&h->loh_reference);
1184 EXPORT_SYMBOL(lu_object_header_fini);
1187 * Given a compound object, find its slice, corresponding to the device type
1190 struct lu_object *lu_object_locate(struct lu_object_header *h,
1191 const struct lu_device_type *dtype)
1193 struct lu_object *o;
1195 cfs_list_for_each_entry(o, &h->loh_layers, lo_linkage) {
1196 if (o->lo_dev->ld_type == dtype)
1201 EXPORT_SYMBOL(lu_object_locate);
1206 * Finalize and free devices in the device stack.
1208 * Finalize device stack by purging object cache, and calling
1209 * lu_device_type_operations::ldto_device_fini() and
1210 * lu_device_type_operations::ldto_device_free() on all devices in the stack.
1212 void lu_stack_fini(const struct lu_env *env, struct lu_device *top)
1214 struct lu_site *site = top->ld_site;
1215 struct lu_device *scan;
1216 struct lu_device *next;
1218 lu_site_purge(env, site, ~0);
1219 for (scan = top; scan != NULL; scan = next) {
1220 next = scan->ld_type->ldt_ops->ldto_device_fini(env, scan);
1221 lu_ref_del(&scan->ld_reference, "lu-stack", &lu_site_init);
1222 lu_device_put(scan);
1226 lu_site_purge(env, site, ~0);
1228 for (scan = top; scan != NULL; scan = next) {
1229 const struct lu_device_type *ldt = scan->ld_type;
1230 struct obd_type *type;
1232 next = ldt->ldt_ops->ldto_device_free(env, scan);
1233 type = ldt->ldt_obd_type;
1236 class_put_type(type);
1240 EXPORT_SYMBOL(lu_stack_fini);
1244 * Maximal number of tld slots.
1246 LU_CONTEXT_KEY_NR = 32
1249 static struct lu_context_key *lu_keys[LU_CONTEXT_KEY_NR] = { NULL, };
1251 static DEFINE_SPINLOCK(lu_keys_guard);
1254 * Global counter incremented whenever key is registered, unregistered,
1255 * revived or quiesced. This is used to void unnecessary calls to
1256 * lu_context_refill(). No locking is provided, as initialization and shutdown
1257 * are supposed to be externally serialized.
1259 static unsigned key_set_version = 0;
1264 int lu_context_key_register(struct lu_context_key *key)
1269 LASSERT(key->lct_init != NULL);
1270 LASSERT(key->lct_fini != NULL);
1271 LASSERT(key->lct_tags != 0);
1272 LASSERT(key->lct_owner != NULL);
1275 cfs_spin_lock(&lu_keys_guard);
1276 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1277 if (lu_keys[i] == NULL) {
1279 cfs_atomic_set(&key->lct_used, 1);
1281 lu_ref_init(&key->lct_reference);
1287 cfs_spin_unlock(&lu_keys_guard);
1290 EXPORT_SYMBOL(lu_context_key_register);
1292 static void key_fini(struct lu_context *ctx, int index)
1294 if (ctx->lc_value != NULL && ctx->lc_value[index] != NULL) {
1295 struct lu_context_key *key;
1297 key = lu_keys[index];
1298 LASSERT(key != NULL);
1299 LASSERT(key->lct_fini != NULL);
1300 LASSERT(cfs_atomic_read(&key->lct_used) > 1);
1302 key->lct_fini(ctx, key, ctx->lc_value[index]);
1303 lu_ref_del(&key->lct_reference, "ctx", ctx);
1304 cfs_atomic_dec(&key->lct_used);
1306 LASSERT(key->lct_owner != NULL);
1307 if ((ctx->lc_tags & LCT_NOREF) == 0) {
1308 LINVRNT(cfs_module_refcount(key->lct_owner) > 0);
1309 cfs_module_put(key->lct_owner);
1311 ctx->lc_value[index] = NULL;
1318 void lu_context_key_degister(struct lu_context_key *key)
1320 LASSERT(cfs_atomic_read(&key->lct_used) >= 1);
1321 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1323 lu_context_key_quiesce(key);
1326 cfs_spin_lock(&lu_keys_guard);
1327 key_fini(&lu_shrink_env.le_ctx, key->lct_index);
1328 if (lu_keys[key->lct_index]) {
1329 lu_keys[key->lct_index] = NULL;
1330 lu_ref_fini(&key->lct_reference);
1332 cfs_spin_unlock(&lu_keys_guard);
1334 LASSERTF(cfs_atomic_read(&key->lct_used) == 1,
1335 "key has instances: %d\n",
1336 cfs_atomic_read(&key->lct_used));
1338 EXPORT_SYMBOL(lu_context_key_degister);
1341 * Register a number of keys. This has to be called after all keys have been
1342 * initialized by a call to LU_CONTEXT_KEY_INIT().
1344 int lu_context_key_register_many(struct lu_context_key *k, ...)
1346 struct lu_context_key *key = k;
1352 result = lu_context_key_register(key);
1355 key = va_arg(args, struct lu_context_key *);
1356 } while (key != NULL);
1362 lu_context_key_degister(k);
1363 k = va_arg(args, struct lu_context_key *);
1370 EXPORT_SYMBOL(lu_context_key_register_many);
1373 * De-register a number of keys. This is a dual to
1374 * lu_context_key_register_many().
1376 void lu_context_key_degister_many(struct lu_context_key *k, ...)
1382 lu_context_key_degister(k);
1383 k = va_arg(args, struct lu_context_key*);
1384 } while (k != NULL);
1387 EXPORT_SYMBOL(lu_context_key_degister_many);
1390 * Revive a number of keys.
1392 void lu_context_key_revive_many(struct lu_context_key *k, ...)
1398 lu_context_key_revive(k);
1399 k = va_arg(args, struct lu_context_key*);
1400 } while (k != NULL);
1403 EXPORT_SYMBOL(lu_context_key_revive_many);
1406 * Quiescent a number of keys.
1408 void lu_context_key_quiesce_many(struct lu_context_key *k, ...)
1414 lu_context_key_quiesce(k);
1415 k = va_arg(args, struct lu_context_key*);
1416 } while (k != NULL);
1419 EXPORT_SYMBOL(lu_context_key_quiesce_many);
1422 * Return value associated with key \a key in context \a ctx.
1424 void *lu_context_key_get(const struct lu_context *ctx,
1425 const struct lu_context_key *key)
1427 LINVRNT(ctx->lc_state == LCS_ENTERED);
1428 LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1429 LASSERT(lu_keys[key->lct_index] == key);
1430 return ctx->lc_value[key->lct_index];
1432 EXPORT_SYMBOL(lu_context_key_get);
1435 * List of remembered contexts. XXX document me.
1437 static CFS_LIST_HEAD(lu_context_remembered);
1440 * Destroy \a key in all remembered contexts. This is used to destroy key
1441 * values in "shared" contexts (like service threads), when a module owning
1442 * the key is about to be unloaded.
1444 void lu_context_key_quiesce(struct lu_context_key *key)
1446 struct lu_context *ctx;
1447 extern unsigned cl_env_cache_purge(unsigned nr);
1449 if (!(key->lct_tags & LCT_QUIESCENT)) {
1451 * XXX layering violation.
1453 cl_env_cache_purge(~0);
1454 key->lct_tags |= LCT_QUIESCENT;
1456 * XXX memory barrier has to go here.
1458 cfs_spin_lock(&lu_keys_guard);
1459 cfs_list_for_each_entry(ctx, &lu_context_remembered,
1461 key_fini(ctx, key->lct_index);
1462 cfs_spin_unlock(&lu_keys_guard);
1466 EXPORT_SYMBOL(lu_context_key_quiesce);
1468 void lu_context_key_revive(struct lu_context_key *key)
1470 key->lct_tags &= ~LCT_QUIESCENT;
1473 EXPORT_SYMBOL(lu_context_key_revive);
1475 static void keys_fini(struct lu_context *ctx)
1479 if (ctx->lc_value == NULL)
1482 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i)
1485 OBD_FREE(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1486 ctx->lc_value = NULL;
1489 static int keys_fill(struct lu_context *ctx)
1493 LINVRNT(ctx->lc_value != NULL);
1494 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1495 struct lu_context_key *key;
1498 if (ctx->lc_value[i] == NULL && key != NULL &&
1499 (key->lct_tags & ctx->lc_tags) &&
1501 * Don't create values for a LCT_QUIESCENT key, as this
1502 * will pin module owning a key.
1504 !(key->lct_tags & LCT_QUIESCENT)) {
1507 LINVRNT(key->lct_init != NULL);
1508 LINVRNT(key->lct_index == i);
1510 value = key->lct_init(ctx, key);
1511 if (unlikely(IS_ERR(value)))
1512 return PTR_ERR(value);
1514 LASSERT(key->lct_owner != NULL);
1515 if (!(ctx->lc_tags & LCT_NOREF))
1516 cfs_try_module_get(key->lct_owner);
1517 lu_ref_add_atomic(&key->lct_reference, "ctx", ctx);
1518 cfs_atomic_inc(&key->lct_used);
1520 * This is the only place in the code, where an
1521 * element of ctx->lc_value[] array is set to non-NULL
1524 ctx->lc_value[i] = value;
1525 if (key->lct_exit != NULL)
1526 ctx->lc_tags |= LCT_HAS_EXIT;
1528 ctx->lc_version = key_set_version;
1533 static int keys_init(struct lu_context *ctx)
1535 OBD_ALLOC(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1536 if (likely(ctx->lc_value != NULL))
1537 return keys_fill(ctx);
1543 * Initialize context data-structure. Create values for all keys.
1545 int lu_context_init(struct lu_context *ctx, __u32 tags)
1549 memset(ctx, 0, sizeof *ctx);
1550 ctx->lc_state = LCS_INITIALIZED;
1551 ctx->lc_tags = tags;
1552 if (tags & LCT_REMEMBER) {
1553 cfs_spin_lock(&lu_keys_guard);
1554 cfs_list_add(&ctx->lc_remember, &lu_context_remembered);
1555 cfs_spin_unlock(&lu_keys_guard);
1557 CFS_INIT_LIST_HEAD(&ctx->lc_remember);
1560 rc = keys_init(ctx);
1562 lu_context_fini(ctx);
1566 EXPORT_SYMBOL(lu_context_init);
1569 * Finalize context data-structure. Destroy key values.
1571 void lu_context_fini(struct lu_context *ctx)
1573 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1574 ctx->lc_state = LCS_FINALIZED;
1576 if ((ctx->lc_tags & LCT_REMEMBER) == 0) {
1577 LASSERT(cfs_list_empty(&ctx->lc_remember));
1580 } else { /* could race with key degister */
1581 cfs_spin_lock(&lu_keys_guard);
1583 cfs_list_del_init(&ctx->lc_remember);
1584 cfs_spin_unlock(&lu_keys_guard);
1587 EXPORT_SYMBOL(lu_context_fini);
1590 * Called before entering context.
1592 void lu_context_enter(struct lu_context *ctx)
1594 LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1595 ctx->lc_state = LCS_ENTERED;
1597 EXPORT_SYMBOL(lu_context_enter);
1600 * Called after exiting from \a ctx
1602 void lu_context_exit(struct lu_context *ctx)
1606 LINVRNT(ctx->lc_state == LCS_ENTERED);
1607 ctx->lc_state = LCS_LEFT;
1608 if (ctx->lc_tags & LCT_HAS_EXIT && ctx->lc_value != NULL) {
1609 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1610 if (ctx->lc_value[i] != NULL) {
1611 struct lu_context_key *key;
1614 LASSERT(key != NULL);
1615 if (key->lct_exit != NULL)
1617 key, ctx->lc_value[i]);
1622 EXPORT_SYMBOL(lu_context_exit);
1625 * Allocate for context all missing keys that were registered after context
1626 * creation. key_set_version is only changed in rare cases when modules
1627 * are loaded and removed.
1629 int lu_context_refill(struct lu_context *ctx)
1631 return likely(ctx->lc_version == key_set_version) ? 0 : keys_fill(ctx);
1633 EXPORT_SYMBOL(lu_context_refill);
1636 * lu_ctx_tags/lu_ses_tags will be updated if there are new types of
1637 * obd being added. Currently, this is only used on client side, specifically
1638 * for echo device client, for other stack (like ptlrpc threads), context are
1639 * predefined when the lu_device type are registered, during the module probe
1642 __u32 lu_context_tags_default = 0;
1643 __u32 lu_session_tags_default = 0;
1645 void lu_context_tags_update(__u32 tags)
1647 cfs_spin_lock(&lu_keys_guard);
1648 lu_context_tags_default |= tags;
1650 cfs_spin_unlock(&lu_keys_guard);
1652 EXPORT_SYMBOL(lu_context_tags_update);
1654 void lu_context_tags_clear(__u32 tags)
1656 cfs_spin_lock(&lu_keys_guard);
1657 lu_context_tags_default &= ~tags;
1659 cfs_spin_unlock(&lu_keys_guard);
1661 EXPORT_SYMBOL(lu_context_tags_clear);
1663 void lu_session_tags_update(__u32 tags)
1665 cfs_spin_lock(&lu_keys_guard);
1666 lu_session_tags_default |= tags;
1668 cfs_spin_unlock(&lu_keys_guard);
1670 EXPORT_SYMBOL(lu_session_tags_update);
1672 void lu_session_tags_clear(__u32 tags)
1674 cfs_spin_lock(&lu_keys_guard);
1675 lu_session_tags_default &= ~tags;
1677 cfs_spin_unlock(&lu_keys_guard);
1679 EXPORT_SYMBOL(lu_session_tags_clear);
1681 int lu_env_init(struct lu_env *env, __u32 tags)
1686 result = lu_context_init(&env->le_ctx, tags);
1687 if (likely(result == 0))
1688 lu_context_enter(&env->le_ctx);
1691 EXPORT_SYMBOL(lu_env_init);
1693 void lu_env_fini(struct lu_env *env)
1695 lu_context_exit(&env->le_ctx);
1696 lu_context_fini(&env->le_ctx);
1699 EXPORT_SYMBOL(lu_env_fini);
1701 int lu_env_refill(struct lu_env *env)
1705 result = lu_context_refill(&env->le_ctx);
1706 if (result == 0 && env->le_ses != NULL)
1707 result = lu_context_refill(env->le_ses);
1710 EXPORT_SYMBOL(lu_env_refill);
1713 * Currently, this API will only be used by echo client.
1714 * Because echo client and normal lustre client will share
1715 * same cl_env cache. So echo client needs to refresh
1716 * the env context after it get one from the cache, especially
1717 * when normal client and echo client co-exist in the same client.
1719 int lu_env_refill_by_tags(struct lu_env *env, __u32 ctags,
1724 if ((env->le_ctx.lc_tags & ctags) != ctags) {
1725 env->le_ctx.lc_version = 0;
1726 env->le_ctx.lc_tags |= ctags;
1729 if (env->le_ses && (env->le_ses->lc_tags & stags) != stags) {
1730 env->le_ses->lc_version = 0;
1731 env->le_ses->lc_tags |= stags;
1734 result = lu_env_refill(env);
1738 EXPORT_SYMBOL(lu_env_refill_by_tags);
1740 static struct cfs_shrinker *lu_site_shrinker = NULL;
1742 typedef struct lu_site_stats{
1743 unsigned lss_populated;
1744 unsigned lss_max_search;
1749 static void lu_site_stats_get(cfs_hash_t *hs,
1750 lu_site_stats_t *stats, int populated)
1755 cfs_hash_for_each_bucket(hs, &bd, i) {
1756 struct lu_site_bkt_data *bkt = cfs_hash_bd_extra_get(hs, &bd);
1757 cfs_hlist_head_t *hhead;
1759 cfs_hash_bd_lock(hs, &bd, 1);
1760 stats->lss_busy += bkt->lsb_busy;
1761 stats->lss_total += cfs_hash_bd_count_get(&bd);
1762 stats->lss_max_search = max((int)stats->lss_max_search,
1763 cfs_hash_bd_depmax_get(&bd));
1765 cfs_hash_bd_unlock(hs, &bd, 1);
1769 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
1770 if (!cfs_hlist_empty(hhead))
1771 stats->lss_populated++;
1773 cfs_hash_bd_unlock(hs, &bd, 1);
1779 static int lu_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
1781 lu_site_stats_t stats;
1783 struct lu_site *tmp;
1785 int remain = shrink_param(sc, nr_to_scan);
1786 CFS_LIST_HEAD(splice);
1789 if (!(shrink_param(sc, gfp_mask) & __GFP_FS))
1791 CDEBUG(D_INODE, "Shrink %d objects\n", remain);
1794 cfs_mutex_lock(&lu_sites_guard);
1795 cfs_list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
1796 if (shrink_param(sc, nr_to_scan) != 0) {
1797 remain = lu_site_purge(&lu_shrink_env, s, remain);
1799 * Move just shrunk site to the tail of site list to
1800 * assure shrinking fairness.
1802 cfs_list_move_tail(&s->ls_linkage, &splice);
1805 memset(&stats, 0, sizeof(stats));
1806 lu_site_stats_get(s->ls_obj_hash, &stats, 0);
1807 cached += stats.lss_total - stats.lss_busy;
1808 if (shrink_param(sc, nr_to_scan) && remain <= 0)
1811 cfs_list_splice(&splice, lu_sites.prev);
1812 cfs_mutex_unlock(&lu_sites_guard);
1814 cached = (cached / 100) * sysctl_vfs_cache_pressure;
1815 if (shrink_param(sc, nr_to_scan) == 0)
1816 CDEBUG(D_INODE, "%d objects cached\n", cached);
1825 * Environment to be used in debugger, contains all tags.
1827 struct lu_env lu_debugging_env;
1830 * Debugging printer function using printk().
1832 int lu_printk_printer(const struct lu_env *env,
1833 void *unused, const char *format, ...)
1837 va_start(args, format);
1838 vprintk(format, args);
1843 void lu_debugging_setup(void)
1845 lu_env_init(&lu_debugging_env, ~0);
1848 void lu_context_keys_dump(void)
1852 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1853 struct lu_context_key *key;
1857 CERROR("[%d]: %p %x (%p,%p,%p) %d %d \"%s\"@%p\n",
1858 i, key, key->lct_tags,
1859 key->lct_init, key->lct_fini, key->lct_exit,
1860 key->lct_index, cfs_atomic_read(&key->lct_used),
1861 key->lct_owner ? key->lct_owner->name : "",
1863 lu_ref_print(&key->lct_reference);
1867 EXPORT_SYMBOL(lu_context_keys_dump);
1868 #else /* !__KERNEL__ */
1869 static int lu_cache_shrink(int nr, unsigned int gfp_mask)
1873 #endif /* __KERNEL__ */
1875 int cl_global_init(void);
1876 void cl_global_fini(void);
1877 int lu_ref_global_init(void);
1878 void lu_ref_global_fini(void);
1880 int dt_global_init(void);
1881 void dt_global_fini(void);
1883 int llo_global_init(void);
1884 void llo_global_fini(void);
1887 * Initialization of global lu_* data.
1889 int lu_global_init(void)
1893 CDEBUG(D_INFO, "Lustre LU module (%p).\n", &lu_keys);
1895 result = lu_ref_global_init();
1899 LU_CONTEXT_KEY_INIT(&lu_global_key);
1900 result = lu_context_key_register(&lu_global_key);
1904 * At this level, we don't know what tags are needed, so allocate them
1905 * conservatively. This should not be too bad, because this
1906 * environment is global.
1908 cfs_mutex_lock(&lu_sites_guard);
1909 result = lu_env_init(&lu_shrink_env, LCT_SHRINKER);
1910 cfs_mutex_unlock(&lu_sites_guard);
1915 * seeks estimation: 3 seeks to read a record from oi, one to read
1916 * inode, one for ea. Unfortunately setting this high value results in
1917 * lu_object/inode cache consuming all the memory.
1919 lu_site_shrinker = cfs_set_shrinker(CFS_DEFAULT_SEEKS, lu_cache_shrink);
1920 if (lu_site_shrinker == NULL)
1923 result = lu_time_global_init();
1928 result = dt_global_init();
1932 result = llo_global_init();
1936 result = cl_global_init();
1943 * Dual to lu_global_init().
1945 void lu_global_fini(void)
1952 lu_time_global_fini();
1953 if (lu_site_shrinker != NULL) {
1954 cfs_remove_shrinker(lu_site_shrinker);
1955 lu_site_shrinker = NULL;
1958 lu_context_key_degister(&lu_global_key);
1961 * Tear shrinker environment down _after_ de-registering
1962 * lu_global_key, because the latter has a value in the former.
1964 cfs_mutex_lock(&lu_sites_guard);
1965 lu_env_fini(&lu_shrink_env);
1966 cfs_mutex_unlock(&lu_sites_guard);
1968 lu_ref_global_fini();
1971 struct lu_buf LU_BUF_NULL = {
1975 EXPORT_SYMBOL(LU_BUF_NULL);
1977 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx)
1980 struct lprocfs_counter ret;
1982 lprocfs_stats_collect(stats, idx, &ret);
1983 return (__u32)ret.lc_count;
1990 * Output site statistical counters into a buffer. Suitable for
1991 * lprocfs_rd_*()-style functions.
1993 int lu_site_stats_print(const struct lu_site *s, char *page, int count)
1995 lu_site_stats_t stats;
1997 memset(&stats, 0, sizeof(stats));
1998 lu_site_stats_get(s->ls_obj_hash, &stats, 1);
2000 return snprintf(page, count, "%d/%d %d/%d %d %d %d %d %d %d %d\n",
2003 stats.lss_populated,
2004 CFS_HASH_NHLIST(s->ls_obj_hash),
2005 stats.lss_max_search,
2006 ls_stats_read(s->ls_stats, LU_SS_CREATED),
2007 ls_stats_read(s->ls_stats, LU_SS_CACHE_HIT),
2008 ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
2009 ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
2010 ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
2011 ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED));
2013 EXPORT_SYMBOL(lu_site_stats_print);
2015 const char *lu_time_names[LU_TIME_NR] = {
2016 [LU_TIME_FIND_LOOKUP] = "find_lookup",
2017 [LU_TIME_FIND_ALLOC] = "find_alloc",
2018 [LU_TIME_FIND_INSERT] = "find_insert"
2020 EXPORT_SYMBOL(lu_time_names);
2023 * Helper function to initialize a number of kmem slab caches at once.
2025 int lu_kmem_init(struct lu_kmem_descr *caches)
2028 struct lu_kmem_descr *iter = caches;
2030 for (result = 0; iter->ckd_cache != NULL; ++iter) {
2031 *iter->ckd_cache = cfs_mem_cache_create(iter->ckd_name,
2034 if (*iter->ckd_cache == NULL) {
2036 /* free all previously allocated caches */
2037 lu_kmem_fini(caches);
2043 EXPORT_SYMBOL(lu_kmem_init);
2046 * Helper function to finalize a number of kmem slab cached at once. Dual to
2049 void lu_kmem_fini(struct lu_kmem_descr *caches)
2053 for (; caches->ckd_cache != NULL; ++caches) {
2054 if (*caches->ckd_cache != NULL) {
2055 rc = cfs_mem_cache_destroy(*caches->ckd_cache);
2056 LASSERTF(rc == 0, "couldn't destroy %s slab\n",
2058 *caches->ckd_cache = NULL;
2062 EXPORT_SYMBOL(lu_kmem_fini);
2065 * Temporary solution to be able to assign fid in ->do_create()
2066 * till we have fully-functional OST fids
2068 void lu_object_assign_fid(const struct lu_env *env, struct lu_object *o,
2069 const struct lu_fid *fid)
2071 struct lu_site *s = o->lo_dev->ld_site;
2072 struct lu_fid *old = &o->lo_header->loh_fid;
2073 struct lu_site_bkt_data *bkt;
2074 struct lu_object *shadow;
2075 cfs_waitlink_t waiter;
2080 LASSERT(fid_is_zero(old));
2082 hs = s->ls_obj_hash;
2083 cfs_hash_bd_get_and_lock(hs, (void *)fid, &bd, 1);
2084 shadow = htable_lookup(s, &bd, fid, &waiter, &version);
2085 /* supposed to be unique */
2086 LASSERT(shadow == NULL);
2088 bkt = cfs_hash_bd_extra_get(hs, &bd);
2089 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
2091 cfs_hash_bd_unlock(hs, &bd, 1);
2093 EXPORT_SYMBOL(lu_object_assign_fid);